NONLINEAR SYSTEMS ANALYSIS SECOND EDITION M. VIDYASAGAR Centre for Al and Robotics, India a PRENTICE HALL, Englewood Cliffs, New Jersey 07632Liveary of congress cata ar systens analysts / #. Vigyasagar. == 2/e anchiges bib} so9tat ce) references ang inaex. ‘hece:v8a 1983 20a" Ponae20 s2-5200 ‘Acquisitions editor: Pete Janzow Prepress buyer: Linda Behrens Production editor: Jennifer Wencel Manufatuting buyer: David Dickey Copy editor: Cami Gofft Supplements editor: Alice Dworkin Cover designer: Joe DiDomenico Editorial assistant Phyllis Morgan SSB 01993, 1978 ty Prentice Hal, le ‘A Simon & Schuster Company Englewood Clifls, New lersey 07632 “The suthor and pbliser ofthis book have used their best efforts in preparing this book. These efforts inchde the ‘developmen, research, and testing of the theories and programs to determine ther effectiveness. The author sad publisher make no warranty of any kind, expressed or implied, with regard to these programs or the documentation ‘contained inthis book. The author and publisher shall not be liable in any event for incidental or consequential damage Jn connecton with, or atsing out of, the furnishing, performance, oF use ofthese programs. Allright served, No part of this book may be reproduced, in any form or by any means, without permission in writing from the publisher Printed inthe United States of America mo87 654321 ISBN 0-13-b234b3-1 HM 3 > ISBN 0-13-b234b3-1 | 9 ¥7801561254650 Prentice-Hall International (UK) Limited, London Prentice-Hall of Australia Pry Limited, Sydney Prentice-Hall Canada Ie. romto Prentice-Hall ispancamericana, S.A., Mexico Prentice-Hall of India Private Limited, New Delhi Prentice Hall of fpan Ie, Tayo Simon & Sebucor Asia Pre. Lid, Singapore ditora Prentice-Hall Brasil, ida, Rio de JaneiroTo Charlie DesoerSpa ar ser srefd APRA AagatAraiy ara via ere Asall water falling from the sky Eventually reaches the sea Sodo salutations to various deities Reach the same almighty From Sandhyavandanam (A salute to the wilight) Si¥owea sow SS en BALE Bd 458 Sour Bo AcS@ Sho Bajo arya Deors 4698 SdH Beyond the worlds ‘Their rulers and their denizens Beyond the unwordly void ‘The one Who shines alone Him I worship From Andhra Maha Bhagavatam by Bammera Potana, (€.1400A.D,)24 22 23 24 25 34 32 33 34 4d 42 43 sa 32 33 sa 55 56 37 CONTENT: PREFACE ix NOTETOTHEREADER xi INTRODUCTION 1 NONLINEAR DIFFERENTIAL EQUATIONS 6 Mathematical Preliminaries 6 Induced Norms and Matrix Measures. 19 Contraction Mapping Theorem 27 Nonlinear Differential Equations 33 Solution Estimates 46 SECOND-ORDERSYSTEMS 53 Preliminaries 53 Linearization Method 57 Periodic Solutions 67 ‘Two Analytical Appro APPROXIMATE ANALYSIS METHODS 88 Describing Functions 88 Periodic Solutions: Rigorous Arguments 109 Singular Perturbations 127 LYAPUNOYSTABILITY 135 Stability Definitions 135 Some Preliminaries 147 Lyapunov’s Direct Method 157 Stability of Linear Systems 193 Lyapunov’s Linearization Method 209 ‘The Lure Problem 2/9 Converse Theorems 2355.8 Applications of Converse Theorems 246 5.9 Discrete-Time Systems 264 6. INPUT-OUTPUTSTABILITY 270 6.1 LySpaces and their Extensions 271 6.2. Definitions of Input-Output Stability 277 6.3 Relationships Between UO and Lyapunov Stability 284 64. Open-Loop Stability of Linear Systems 292 65. Linear Time-Invariant Feedback Systems 309 6.6. Time-Varying and/or Nonlinear Systems 337 617 Discrete-Time Systems 365 7. DIFFERENTIAL GEOMETRIC METHODS 376 7.1 Basics of Differential Geometry 377 7.2. Distributions, Frobenius Theorem 392 7.3. Reachability and Observability 399 7.4 Feedback Linearization: Single-Input Case 427 7.5. Feedback Linearization: Multi-InputCase 438 7.6 Input-Output Linearization 456 7.7 Stabilizationof Linearizable Systems 464 A. PREVALENCE OF DIFFERENTIAL EQUATIONS WITH UNIQUE SOLU- TIONS 469 B. PROOF OF THEKALMAN-YACUBOVITCH LEMMA 474 C. PROOF OF THE FROBENIUS THEOREM 476 REFERENCES 486 INDEX 493PREFACE Itis now more than a decade since I wrote the book Nonlinear Systems Analysis. Since that time, several developments have taken place in this area which have made it desirable to update the contents of the book. Accordingly, virtually the entire book has been rewritten. ‘The most notable changes are the following: 1) During the past decade, there have been some significant advances in the arca of ‘nonlinear control system design based on the use of differential geometric methods. Thus it isimperative thatanyone interested in nonlinear system theory should have at least passing acquaintance with these methods. In this second edition, I have included a new chapter which discusses the differential geometric approach (Chapter 7). For ease of exposition, all systems are considered to evolve over an open subset of R*; thus the analysis is only local. ‘Topics covered include reachability, observability. and feedback linearization (in both the input-state and input-output settings), zero dynamics, and the stabilization of linearizable systems. In addition to presenting the theory, I have also included some applications of the theory to problems in robotics. Motivated by this chapter, an interested and diligent student could pursue a more rigorous course of study with an advanced text. 2) Several significant results have been obtained in the “traditional” areas of Lyapunov stability and input-output stability since the writing ofthe first edition. Some ofthese results ae included in the present edition, such as: observer-controller stabilization of nonlinear systems, and the stability of hierarchical systems (Section 5.8); relationships between Lyapanov stability and input-output stability (Section 6.3), and a useful class of transfer functions of distributed systems (Section 6.5). In addition othe above, Section 4.2, contain- ing.a rigorous analy sis of the describing function method, is also new. 3) Various standard texts in stability theory have gone out of print, making their con- tents all but inaccessible tothe student. Two examples of such books are: Stability of Motion by W. Hahn and Feedback Systems: Input-Output Properties by C. A. Desoer and myself. At the same time some of the techniques presented in these books are finding new and pre ‘ously unsuspected applications. With this in mind, in the present edition I have included some relevant material from these and other classic books, such as the converse Lyapunov theory (Section 5.7), and the feedback stability of time-varying and/or nonlinear systems (Section6.6). 4) In view of the increasing importance of digital computers, I have included a discus- sion of discrete-time systems in the chapters dealing with Lyapunov stability and input- ‘output stability. 5) Three new appendices have been added. Appendix A describes a sixty year-old theorem due to Witold Orlicz, on the prevalence of differential equations with unique solu- tions. This paper is quite inaccessible, but its contents deserve wide dissemination. Appen- dix B gives a proof of the Kalman-Yacubovitch lemma, while Appendix C contains a proof Of the Frobenius theorem. The contents of the last two appendices are of course readily available elsewhere, but their inclusion inthe present text makes it more self-contained. ix6) The original edition of this book contained examples which were mostly drill prob- lems or exercises. During the recent years I have come to feel that nonlinear system theory is most useful in studying the behavior of an entire class of systems rather than a given specific system. Accordingly, several applications of nonlinear system theory have been included throughout the book. Mostof them have to do with robotics in some form or other. With these changes, the book is somewhat bigger than the first edition. It would be dificult to cover the entire book during a single semester. However, [hope its value as a reference has been enhanced by the changes. Chapter 2 contains basic material which should be covered in order to appreciate the remainder of the text. But sincere attempt has been made to ensure that Chapters 3 through 7 are independent, so that an instructor can pick and choose material to suit hisher needs. Even within a chapter, itis possible to cover cer- tain sections and omit others. A perusal of the Contents reveals the amount of flexibility available in puting together a suitable course from the contents of the text. Inspte ofthe enlargement inthe size of the book, some topics which deserve the atten- tion of system theorists are not included. Examples of such topics are chaotic motions, averaging analysis, Volterra series, bifurcation theory, and catastrophe theory. [have made ‘a conscious decision to omit these topics, mainly to keep the length of the book within rea- sonable limits. But no study of nonlinear systems is complete without at least an introduc- tion to these topics. Moreover, there are several excellent texts available addressing each of the above topics. In the preface to the first edition, I wrote fancifully that the book could be used by “engineers, mathematicians, biologists etcetera.” Judging by the Science Citation Index, no biologists appear to have read the book (though two social scientists have, amazingly ‘enough). More realistically, would expect the present edition to be of interest primarily to engineers interested in a rigorous treatment of nonlinear systems, and to mathematicians interested in system theory. ‘Though some aspects of control are covered in the book (espe- cially in Chapter 7), the focus is till on analysis rather than synthesis. Hence I have retained the original title. Ido expect that the book can be used not just in Electrical Engineering, departments, but also in Mechanical Engineering departments, and perhaps in some depart- ments of Applied Mathematics. Above all, I hope it will continue to serve as a reference source for standard results in nonlinear system analysis. | would like to thank Toshiharu Sugie for his careful reading of early versions of Chapters 5 and 6, I would also like to thank those who reviewed the text, particularly Brian Anderson, Aristotle Araposthasis, Ragu Balakrishnan, Joseph Bentsman, Alan Desrochers, Brad Dickinson, Ashok Iyer, Bob Newcomb, Charles L. Phillips, nd Irwin Sandberg. Itis my pleasure and honor to dedicate to this book to Professor Charles A. Desoer of the University of California at Berkeley. Though I was not privileged to be one of his Ph.D. students, I was fortunate enough to have come under his influence while still ata formative stage in my career, Any instances of originality, creativity and clarity in my research and ‘exposition are but pale imitations of his shining example.NOTE TO THE READER Al items within each section are numbered consecutively, be they equations, theorems, definitions, or something else. A reference such as "(17)" refers to the 17-th item within the same section. When itis necessary to refer to an item from another section, the full citation is given, e.g., "Theorem (5.1.16)." All theorems, lemmas, and definitions ae stated in ital- ics. In a definition, the concept being defined is displayed in bold face. The same conven- ‘tion is used in the running text as well. The use of italics in the running text is reserved for ‘emphasis. The box symbol Mis used to denote the end of a proof. In cases where there right be some ambiguity, the same symbol is also used to denote the end of an example. Lower-case bold letters such as x denote vectors, upper-case bold letters such as A denote matrices, and italic letters denote scalars; however, there are a few exceptions to this con- vention, Forexample, the identity matrix isdenoted by J. Finally, the readeris urged to attempt all the problems, since they are an integral part of thetext, Happy reading!1. INTRODUCTION ‘The topic of this book i the analysis of nonlinear systems. Tae adjective "nonlinear" can be interpreted in one of two ways, namely: "nor linear" or "not necessarily linear.” The latter ‘meaning is intended here. Why should one study nonlinear systems? The fact is that virwally all physical sys- tems are nonlinear in nature. Sometimes itis possible to describe the operation of a physical system by a linear model, such as a set of ordinary linear differential equations. This is the case, for example, ifthe mode of operation ofthe physical system does not deviate too much from tive “nominal” set of operating conditions. Thus the analysis of linear systems occupies an important place in system theory. But in analyzing the behaviour of any physical syste ‘one often encounters situations where the linearized model is inadequate or inaccurate; that isthe time when the contents of this book may prove useful ‘There are several important differences between linear systems and nonlinear systems: 1) In the case of linear systems described by a set of linear ordinary differential equations, it isoften possible to derive closed-form expressions forthe solutions ofthe system equations. In general, tis is not possible in the case of nonlinear systems described by a set of non- linear ordinary differential equations. As a consequence, itis desirable to be able to make some predictions about the behaviour of a nonlinear system even in the absence of closed- form expressions for the solutions of the system equations. This type of analysis, called qualitative analysis or approximate analysis, is much less relevant to linear systems. 2) ‘The analysis of nonlinear systems makes use of @ wider variety of approaches and ‘mathematical tools than does the analysis of linear systems. The main reason for this variety is that no tool or methoxology in nonlinear systems analysis is universally applicable (in a fruitful manner). Hence the nonlinear systems analyst needs @ wide variety oftoolsin his or her arsenal. 3) In general, the evel of mathematics needed to master the basic ideas of non- linear systems analysis is higher than that forthe linear case. Wriereas matrix algebra usu- ally occupies center stage ina first cours i linear systems analysis, here we us ideas from more advanced topics such as functional analysis and differential geometry. A.commonly used mode! for a nanlinear system is. 1 i(t)=Me, x10), wif. WF20, where # denotes time; x(t) denotes the value of the function x() at time ¢ and is an n- dimensional vector, u(t) is similarly defined and is an m-dimensional vector; and the func~ tion f associates, with each value ofr, x(0), and u(t) a corresponding n-dimensional vector. Following common convention, this is denoted as: 1¢R,, x(QER", u(QeR™, and FR, xR'«R" — R", Note that (I) isa firstorder vector differential equation, The quantity (0) is generally referred to asthe state ofthe system at time f, while u(t) is called the inputIntroduction or the control function. It is clear that (1) represents a continuous-time system. Its discrete-time counterpart is 2 xe fH Os which isa first-order vector difference equation. There is no loss of generality in assuming thatthe system at hand is described by a first-order (differential or difference) equation. To see this, suppose the system is described by the n-th order scalar differential equation ayo) ae dy) 3 an SAL VOI. SORES This equation can be recast inthe form (1) by defining the n-dimensional state vector x(t) in the familiar way, namely ‘yo 4 ADEN KAI AAO = a Then (3) isequivalentto 4(=x(0), FAD=xx(0), =x(0 q(t SAL, 2100), 220.0 an HCO Now (5) is of the form (1) with 6 x)= [1 (--x, (OI. TRU x WEL x2 oa AL More generally, even coupled nonlinear differential equations can be put into the form (I) ‘Analogous remarks apply also to difference equations. In fact, much of the power of modern’ control theory derives from the generality and versatility of the state-space descriptions (1) and 2). ‘studying the system (1), one can make a distinction between two aspeets,! generally referred to as analysis and synthesis, respectively. Suppose the input function u() in (1) is § Henceforth attention i focused on the continous-time system (1), withthe understanding that al emaths apply, mutatis mundi, othe discrete system 2),Introduction 3 specified (i... fixed), and one would like to study the behaviour of the corresponding func- tion x("); this is usually referred to as analysis. Now suppose the problem is tumed around: the system description (1) is given, as well as the desired behaviour of the function x(-), and the problem is to find a suitable input function u(-) that would cause x(-).to behave in this desired fashion; this is usually referred to as synthesis. Most of this book is devoted to the analysis of nonlinear systems, The rest of this chapter is devoted to introducing several commonly used terms. The system (1) is said to be forced, or to have an input; in contrast, a system described by an equation ofthe form 8k) =Alr, xO), Ve20, is said tobe unforced. Note thatthe distinction is not too precise. Inthe system (1), if u() is specified, then itis possible to define a function fy: R,xR" -> R" by 9 £6 0=Me x, uO) Inthis case (1) becomes 10 i) =fylt, xO}, WFO. Moreover, if) is clear from the context, the subscript u on f, is often omitted. In this case there is no distinction between (10) and (8). Thus it is safer to think of (8) as describing one ‘of two possible cases: (i) there is no external input tothe system, or (i) there is an external input, which is kept fixed throughout the study. 11 Definition The system (1) or (8) is said to be autonomous ifthe function f does not ‘explicitly depend on its frstargument; itis said 10 be nonautonomous otherwise. Note that some authors use “time-invariant” instead of “autonomous” and “time- varying" instead of "nonautonomous.” Consider the system (1), and suppose itis autonomous, i¢., fis independent of t. Now suppose a non-constant input function u(-) is applied. Then the corresponding function f, , then W-Il, approaches the norm function of Example (©). {This is the reason for the subscripts in Examples (9) and (11),} The function I-11, clearly satisfies the conditions (N1) and (N2), and can be shown to satisfy (N3) whenever Sp See. Thus the pair (R", I-Il,) is a normed linear space for each value of p in the inter- val [1, eI; of course, for distinct values of p we have distinct normed linear spaces. The norm tt, is called the /,-normon R”. In particular, if p =2, then Sty, 1S tixlla= which is generally called the Euclidean norm on R°. Itis also called the [y-norm on R ‘The Euclidean norm is a particular example of a so-called inner product norm, which is defined in Section 2.1.3. ‘The norm ttl, can also be defined on the set C” in an entirely analogous fashion, si ply by interpreting the qaantity 11,1 in (14) as the magnitude of the complex number 1,1. ‘Thus the pair(C", Hl, isalso a normed linear space for each pe[1, =. Both R* and C* are examples of finite-dimensional linear vector spaces. As a conse- ‘quence, itcan be shown that, given any two norms UI, and Ill, on RY, there exist con- stants and k such that Ay lINI gS Ul Ske Illa, WERE (or C Forinstance, xt, sixty Sn ilxll., WxeR", and. Uxtl, sx; sm? Uxil,., Wxe R® A similar relationship exists between any two normson R" and C*. ‘Suppose (X, I'll) is anormed linear space, and that x, yeX. Then one can think of the quantity Ilx—y'l as the distance between x and y. With the aid of this notion of distance (or proximity), itis possible to define the notion of convergence in a normed linear space set- ting, 16 Definition A sequence {x,}729 in a normed linear space (X, lll) is said 10 converge 1oxy€X if, forevery# >0, there exists an integer N=N(e) such that 17 Wx;-x9ll 0, there exists an integer N=N (€) such that 2 We. cil 0 isgiven; then pick an integer N/such that 2x, -x9l 0, there exists a= 8(€, x9) such that14 Nonlinear Differential Equations Ch. 2B Nf ()-S(eo)lly , there exists a8=6(€) such that 29 IF ()-F Olly R, as follows: Ifx()e C"la, b jsthen SH HxOlle= max ix(OI Toshow that Il-lc isanorm on Ca, b J.one proceeds exactly as in Example (25). Axioms (N1) and (N2) are readily verified. To verify (N3), suppose x(-) and y(-)belong to C"a, b). Then HxC)+ yO) Me =maxlta(+¥( ‘Smax{ Ix( 1+ lly(0!) fromthe triangle inequality on RY ‘Smax tx) +max ly(0) = lx()lle+ Hy Mley ‘where all maxima are taken as f varies over the interval [a, b. Thus (N3) is satisfied and Ile isa normon C*[a, 6]. By the same reasoning as in Example (25), one can see that the pair(C"{a, 6}, lc)isa Banach space. In this example, itis essential to note the difference between Wl and l-tl¢; Well isa norm on R®, while I-llc isanorm on the space C*a, b]. The former has an n-vector as its argument, while the latter has a vector-valued function as its argument. When we study nonlinear differential equations in Section 2.4, this difference becomes crucialSec.2.1 Mathematical Preliminaries 15 2.1.3 Inner Product Spaces ‘An inner product space is a special type of normed linear space in which itis possible to define geometrically appealing concepts such as orthogonality and Fourier series. An inner product space can be defined axiomatically as follows: 32. Definition An inner product space isa linear vector space X with associated field F, together witha function <:,->:X»X — F such thatthe following axioms are satisfied: AD = PPAR = Sy, > IF aC, We yeX, (2) = + x.y, 2€X. (B) =a. Woe F, Vx. yeX, (4) 20, VxeX: =O ifand only ifx =O ‘The quantity is an abstraction of the familiar scalar product or dot product on R’ or’ An inner product space can be made intoa normed linear space in a natural way. 33. Theorem Given an inner product space (X, <->) define the function \I:XR by rs Mooixle” Then W-ll isa normon X, so thatthe pair(X, ill) is anormed linear space. ‘The proof of Theorem (33) depends on the following extremely useful inequality, known as Schwarz’ inequality. 38 Lemma (Schwarz’ Inequality) Let x, y belong to the inner product space (X, <>). Then 36 I thet ityll, and 37 =a? lie? +208 +BPlly I? By Axiom (14), we have that f(a, B)20 for al scalars, B. Since fis a quadratic form in these two scalars, it follows that f(a, B)20 Va, B if and only ifthe discriminant of the qua- dratic form is nonpositive, i.e. 39 0 whenever either tor isnonzero. Thisis true if and only if the discriminant ofthe quadratic form in (38)is negative, i.e. if 40 ' el 4 lly 242 5 Wx 4 lly + 21x ity Il (by Schwarz” inequality) (Ui + Hy? Taking square roots of both sides establishes the triangle inequality. Bl ‘Theorem (33) shows that every inner product space can be made into a normed linear space in a natural way. Hence it makes sense to ask whether an inner product space is com- plete (in the norm defined by the inner product). 41 Definition An inner product space which is complete in the norm defined by the inner productis called a Hilbert space. 43° Example Consider the linear vector space R', together with the function <1. >:RYXR' 5 R defined by HB = Dx Its routine to verify that the function in (44) satisfies all four axioms ofan inner product. ‘The norm on R® defined by the inner product isSee. 2.1 Mathematical Preliminaries "7 ix =[E x74, which is recognized as the [,-norm defined in Example (13). Note that R" together with the inner product defined in 44) isin facta Hilbert space. i 45 Example Let C"[a, bbe the linear space of Example (30), and define the inner pro- duct <:,-> on this space as fellows: Ifx(), ye C*La, b jy then let 46 at, ‘where the inner product inside the integral is that on R" defined in Example (43). Once again the function defined in (46) satisfies all the axioms of the inner product. However, with this inner product, C*[a, 6] is not a Hilbert space; contrast this with the fact that Ca, 6} is a Banach space with the norm W-iIc defined in Example (30). To see that C*la, b] is not a Hilbert space with the inner product in (46), pick a time T such that 4a 0, let 6(€)=€. To show that the definition i satisfied with this choice, suppose.x, ye Xand that18 Nonlinear Differential Equations ch 48 |ilx-yll| ) is an inner product space. Then, for each yeX, the function mapping x into :X + Risuniformly continuous. Proof Ify =0, then =0Vxe X, which is clearly a uniformly continuous fune- tion, soit is only necessary to study the case where y #0. Use Definition (27) of uniform continuity, and givene > 0, define 5(e) =e/Iy i. Now suppose € zeX, andilx~zil <8= = iy Then 53 -1=1I Slx=zll-lyll,by Schwarz’ inequality e <2 -tyl= Tyne Thiscompletesthe proof. i Problem 2.1 Show that the zero element of a linear vector space is unique. [Hint: ‘Assume that the linear vector space V has two zero elements 0, and 03, and use Axiom (v3) Problem 2.2 Show that, ina linear vector space, the additive inverse of an element is. unique. Problem 2.3 Give an example of a set which is nota linear vector space. Problem 2.4 Let S be the sequence space of Example (5), and define a subset 5, of S as the set ofall sequences converging tor. For what values of ris, asubspace of 5? Problem 2.5 Consider the normed linear space R?, with the norm lI-Il, defined in unple (13), Sketch the unit spheres, ic. the setsSee.2.2 Induced Norms and Matrix Measures 19 (ee RE: xt, =1) forthe values p= 1,2, 5, Problem 2.6 (a) Let I! be any norm on RY and 1X), °°°,%q be any collection of vectors in R". Using the triangle inequality, show that Ex, F tha & (&) Let C*{a, b] be as in Example (30). Using the Riemannian approximation to the integral, show that * 5 Uf xeaen'sf xia Problem 2.7 Prove Schwarz’ inequality for complex inner product spaces. Problem 2.8 Suppose (X, <:.">) isan inner product space. Show that the inner pro- ) converges to - [Hint: Write Sp, YP — <0, Yor = and use Schwarz’ inequality. Problem 2.9 Suppose X and Y are normed linear spaces and that f :X fous at x9¢X. Suppose {x;} is a sequence in X converging to xo. Show that the sequence {F (4) in ¥oonverges tof (xo). 2.2 INDUCED NORMS AND MATRIX MEASURES, Inthis section the concepts of the induced norm of a matrix and the measure of a matrix are introsuced. These concepts are used in Section 2.5 to derive estimates for the solutions ‘of nonlinear differential equations, without actually solving them. 2.2.1 Induced Norms Let C™ (respectively, R™") denote the set of all nxn matrices with complex (respec- tively, real) elements. ‘Then C"™* can be made into a complex linear Vector space if addition and scalar multiplication are done componentwise. Moreover, for each matrix AcC™ there is a corresponding linear mapping ct fromC” intoitself, defined by 1 alx)=Ax, xe", ‘Conversely, for every linear mapping ot from C" into itself, chere is a corresponding matrix20 ‘Nonlinear Differential Equations cn2 ‘Ae C™ such that(1) holds. Thus there isa one-to-one correspondence between matrices in (Cand linesr mappings mapping C” into itself. (Actually, this correspondence is one-to- ‘one only after the basis on C" has been chosen, However, in this book such subtleties of linear sigebra are not explored.) We do not in general distinguish between a matrix in C™* ‘andthe corresponding linear mapping on C*. However, this correspondence is the motiva- tion behind the concept of the induced norm of a matrix. 2° Definition Let I: bea given norm on C*. Then for each matrix Ne C"™, the quan- tty WAM, defined by cup UAE gy aearect xi ast {called the induced (tmatrix) norm of A corresponding tothe vector norm I should be noted that there are two distinet functions involved in Definition (2); one is the norm funetion WII mapping C” into R. and the other isthe induced norm function I-I, mapping C”™ intoR. 3 WAN= Waxu sup, Axil, ‘The induced norm of a matrix can be given a simple geometric interpretation. Equation (3) shows that IA I, isthe least upper bound of the ratio HAxI/IIxil asx varies over C*. In this sense, IIA Il, can be thought of as the “gain” of the linear mapping corresponding to A. Alternatively, let B denote the closed unit ballin C*; i.e, et 4 B=(xeC*: Ixtts1). Now suppose We distort B by replacing each x in B by Ax, i, its image under the mapping ‘A. Then what reslts is the image of the set B under the mapping A. In this setting, the induced norm XA I, of A can be thought of as the radius ofthe smallest ballin C” that com- pletely covers the image of B under A. Lemma (5) shows thatthe function Ill, isa valid norm on C™. 5 Lemma For each norm W-ll on C*, the induced norm function NN, maps C™* into (0, ), satisfies Axioms (NI) through (N3), and is therefore anormon C™*. Proof It is clear that Al, 20VAeC™, and Axioms (N1) and (N2) can be verified by inspection. To verify (N3), suppose A. Be C°™*, Then 6 WA+BI,= sup MA+BYxI= sup, WAx+ Ball S sup [Axil + HBxil) by the triangle inequality on C” < + sup WBxll= WAU, + BI. sup, WAxil+ sup, il 1+ UBM, Hence (N3) is also satisfied, and thus W-Il,isanormonC™",See.2.2 Induced Norms and Matrix Measures 21 In view of Lemma (5), is clear that, for each norm on C”, there is @ corresponding induced norm on C™*. However, the converse is not tue. Consider the function Usll,: C™* > R defined by 7 WAN, =max tay Ms ‘Then one can verify that I-11, sa norm on C”™™. Indeed, 1A Ii, is simply the /.. norm of the ‘n2x1 vector consisting ofall the components of the matrix A. However, there is no normon CC" such that Ill, is the corresponding induced matrix norm. This is a consequence of the next result. 8 Lemma Let lll, bean induced normon C™*. Then 9 WABI, < At; IBI,, VA, Bec”, Proof By definition, 10 WABI,=_ sup, HABxIL However, itfollows from (3) that 11 WAYI STAN llyll, VyeC” Soin particular, 12 WABKIS HAI IBxil, VEC". Similarly, 13° Bxil < UBIIlIxil, WxeC*, Combining (12) and (13) gives 14 WABxILS WAI, MBI xi, We C* Now (9) follows immediately from (14). i Thus induced norms have the special feature that they are submultiplicative: ie. the induced norm ofthe product of two matrices A and B is less than or equal to the product of the induced norms of A and B. It can be readily verified by example thatthe norm Ill, of (7) does not have this property (and hence cannot bean induced norm). In general, given a specific norm on C" [say, forinstance, the [,-norm defined in Exam- ple (2.1.13)], itis not always easy to find an explicit expression for the corresponding induced norm on C”*"—the equations in (3) serve more as definitions than as computable expressions. However, the induced matrix norms corresponding tothe vector norms I-lay U-il,, and H-Hly (as defined in Examples (2.1.9), (2.1.11) and (2.1.13) respectively] are2 Nonlinear Differential Equations Ch.2 known and are displayed in Table 2.1. Note that A* denotes the conjugate transpose of the ‘matrix A, and yyy (M) denotes the largest eigenvalue ofthe Hermitian matrix M. Table2.1 Normonc* Induced NormonC™* Wxile=maxix,l WAI, =max¥ lay! Wxil, = D141 WAN, =max¥ lay l Ela12)!—WAly2 =paA*A)]? lixily = 2.2.2. Matrix Measures Let I+, be an induced matrix norm on C"™*. Then the comesponding matrix meas- ureis the function u(.): C™ — R defined by 15 yay= tim UteAMs Dn ma Note that some authors use the termi logarithmic derivative instead, The measure of a matrix H(A) can be thought of as the directional derivative of the induced norm function lI-I, as evaluated at the identity matrix / in the direction A. The measure function has several useful properties, as shown next. 16 Theorem Let li-l, be an induced matrix norm on C™* and let W) be the corresponding matrix measure. Then) has the following properties: (M1) Foreach Ae C™, the limit indicated in(15) exists and is well-defined. (M2) -HAIl,Sw(A)S IAI, VAeC™. H(A), Wai20, Vac C™ (M3) (aA) (M4) max(y(A)-H(-B), H(B)—p(-A)} SpA +B) S H(A) +B), VA, BeC™, (MS) uC) isaconvex function: pla + (1a)B] 0*, and is bounded 18 iis =: fim f= H(A) iswell-defined. Towards this end, itis first shown that 9 0/ Of, Suppose 0 < 8 0, 23° -11AI, $f (€)S WAI, Ve>0. Hence f (€) is bounded below. By previous discussion, this shows that f (€) has a well- defined limit as e— 0". Therefore (A) is well-defined [Property (M1)] and satisfies Pro- perty (M2). To prove (M3), observe that4 Nonlinear Differential Equations Ch.2 Wea Weal, 24 WA) = tm MEANT fgg g MPC ee ee eas ere aa Toprove (M4), we begin by showing that 25 HA+B)sHA)+H®). Aslightrearrangementof (5) gives 26 WA+B)= tim W474 A+Bu,— 2. aoe = But foreache >0, wehave 14 A eae repre eee n dosas disp, 2-4 1 1 |- tyrone] Letting ¢ 0° in (27) proves (25). Now replace A by A+B and B by -B in the right side of (25). Then in the left side of (25) A + Bis replaced by A +B~B=A, which gives 28 WA)SW(A+B)+p(-B), or 29 WA)-W-B)SWA+B), Bysymmetry, 30 W(B)-W-A)S H(A +B). This establishes (M4). Now (MS) is a ready consequence of (M3) and (25). Finally, to prove (M6), let be an eigenvalue of A, and let v be a corresponding eigenvector. Assume ‘without loss of generality that IIvil = 1, where ll is the norm on C* which induces the ‘matrix norm tll, on C"™*. Foreache > 0, we have 31 +eAI,= sup i +eAyxit Bl +eAyvil S11 +eAL-vil= 11 +eat, Similarly follows thatSec.2.2 Induced Norms and Matrix Measures 25 32 N-eAL Ss IM/-eAlll,, Ve>0. Now, itiseasy to verify that 33° Reds lim HtAI=1 ane 34° ReA=— tim EAL ne Combining (31) 10(34) establishes (M6). i ‘Comparing the properties of the matrix measure and the induced matrix norm, we see that, although both functions are convex, the similarity’almost ends there, The measure can have positive as well 2s negative values, whereas a norm can assume only nonnegative values, The measure is "sign-sensitive” in that u(-A)#u(A) in general, whereas ll-All,= AI. Because of these special properties, the measure function, is useful in ‘obtaining tight upper bounds on the norms of solutions of vector differential equations. ‘Theorem (16) lists only some of the many interesting properties of the measure func- tion, A more complete discussion can be found in Desoer and Vidyasagar (1975) and Desoer and Haneda (1972). In defining the measure of a matrix in C™™, we have assumed that the norm used in (15) isan induced norm. Its possible, given any norm on C"™*, to define a corresponding measure function U() mapping C™" into R. Inthis case, Properties (Ml) through (MS) sill hold, but (M6) does not. Such a measure function is of no use in estimating the norm of a solution to vector differential equation; for such a purpose, only measures corresponding to induced matrix norms are useful In most applications, such as those involving differential equations, the linear vector space in question is R", and the matrices of interest belong to R"™*. Suppose ill isa norm con Rand let ll; denote the corresponding induced matrix norm defined on R™"; sup- pose we define the corresponding matrix measure M(-) as in (15), except that now Ae R™" and tI is only defined on R™", What properties does such a measure function have? An examination of the proofs of Properties (MI) through (MS) of Theorem (16) reveals that they carry over without modification to the case of real matrices. However, in proving Pro- perty (M6), essential use was made of the fact thatthe space in question is C” and not R", since in general the both the eigenvalue & and eigenvector v could be complex. To get around this difficulty, one can “extend” the given norm on IR" toaorm on C". The details are not given here, but it can be shown that even Property (96) is true for such a measure (see Problem 2.12). This can be summarized as follows: 35 Theorem Let I'll be a norm defined on R", and let Wl,:R™" Rand O):R™* > R be defined in a manner analogous to (2) and (1), respectively. Then w(-) satisfies Properties (MI) through (M6) of Theorem (16)6 ‘Nonlinear Differential Equations Ch.2 Given a particular vector non II-ll on C* (or R®), itis in general a difficult task to ‘obtain an explicit expression forthe corresponding induced matrix norm (as mentioned ear- lies), and it is therefore still more difficult to obtain an explicit expression for the coresponding matrix measure. Nevertheless, the measure functions corresponding to the ‘norms itl, lll, and Ill. can be calculated, and are displayed in Table 2.2 below. Table 22 Norm on" Matrix Measure on C™** Wxtl=maxix,! —— w(A)=max lay +Z lay! wyAdamarley+ Bley!) ally =(S 1)! 112(A) =Ama(A* + AV 36 Example Let . 621 Az| 0-12). 130 Using the formulas given in Table 2.2, one obtains by inspection that Using Property (M6) of Theorem (35) toestimate the real parts of the eigenvalues of A. one obtains -7 eo. Since the norm function is continuous, it follows that 9 tax Tim aq ~ xq = fm tags Hays ‘where we have used the fact the right side of (6) is independent of m. i [Note that in general itis not possibie w replace (2) by the weaker condition 10 WTe~Tyll Rby MW f(jax+ 5 ~tan and define T= f(x). Then 2 fe) <1, vreR, By the mean-value theorem,30 Nonlinear Differential Equations ch 13 f@)-f0)=f@)(«-y) forsome ze(x,y). Hence T satisfies (10). However, it follows from (11) that f(x)=x if and only if tan"!(x)=1/2. Clearly no such xexists, Hence Thas no fixed point in R. 14 Example Let /:R— Rbeacontinuously differentiable function, and suppose sup leol=pX. Suppose there exists aconstantp < I such that 16 IITx-Tyll Rbe the function defined by 2, iflxisi 20 T=} 0, ifleior If Mis chosen asthe interval [~1, 1], then Tis a contractiomrover M. However, itis ‘not possible to find an x9€ M such that the ball B defined in (17) is contained in M. ‘Accordingly, Thasno fixed pointin M. Suppose we do succeed in finding an x € M such that the hypotheses of Theorem (5) hold. Then the particular sequence defined in (18) converges to the unique fixed point x* of T in M. However, if we choose another starting point for the iteration, there is no guarantee thatthe resulting sequence will converge to.x*. In contrast, if Tis global contraction, then the sequence defined in (3) converges to x* converges for every starting point. There is one small consolation: If the sequence of iterations remains in M, then it must in fact converge to x*; see Theorem (22) below. Proof First, itis clear from (16) that Thas at most one fixed point in M. If roe M is chosen in such a way that the ball B defined if (17) is contained in M, then it follows that the sequence (1, } defined in (18) stays in B fr all n; to see this, apply the inequality (6) with ‘n=0, Because the contraction condition holds in B, one can show, just as in the proof of ‘Theorem (1), that (%,] is a Cauchy sequence in X and therefore converges to an element of X. Denote this limit by x*; then a routine application ofthe continuity of the norm function shows thatthe limit must also belong to B and hence to M. The rest ofthe proof exactly fol- lows thatof Theorem (1).22 Nonlinear Differential Equations Ch.2 21 Example Consider once again the case where X =R, and let f:R +R be continu- ously differentiable. Suppose sup, Ist)! seh <1 and that there existsan xo¢ [~1,1] such that _ filso)~ao , £0)-¥0 fo tot | Sb ‘Then Theorem (15) tells us that there is a unique x*¢ (~1,1] such that f (x*)=.x*, and that x* isthe limit of the sequence (x9. (of Lf (xo)}*" }. The situation is depicted in Figure 22, ‘This section is concluded with another theorem whose hypotheses and conclusions lie ‘between those of Theorems (1) and (15). This theorem is convenient for later applications, » yee Fig.22 vara 22 Theorem Let(X, til) bea Banach space, and et B bea closed bal in X, :e,,asetof the form 23° Ba{x:llx-zilsr) forsomezeX andr 0, LetT:X —+X be an operator satisfying the following conditions: (i) Tmaps Binto itself, i.e., Txe B whenever xe B, (ii) There exists aconstant p< | such that WA i1Tx-Tyll spilx-yil, Wx, yeB. Under these conditions, Thas exactly one fixed point in B. Ifx* denotes the fixed point of TSee.2.4 Nonlinear Differential Equations 3 inB, then foreachxo€ B, the sequence (x, defined by 2S yap = Tq 820, cowergestox*. Moreover 26 tix,—x0ll sf UTx9—xoll, Wn 20. ‘The proof is obvious from Theorem (15). ‘The difference between Theorems (15) and (22)is thatin te latter case Tis assumed 10 map the entre ball B into itself, whereas in the former case itis only assumed that for apar- ticular point x9€ B the sequence of iterations is contained in B. As a consequence, in the latter case one can start from an arbitrary starting point in B o compute x*. Troblem 2.13 Give a detailed proof of Theorem (22) 2.4 NONLINEAR DIFFERENTIAL EQUATIONS, In this section, we derive some general and very useful conditions which guarantee the existence and uniqueness of solutions tothe nonlinear differential equation A R0)=MF, x(0)], 1205210) =p, where x(1)e R” and £R,xR" > R". As shown in Chapter 1, the existence and uniqueness of solutions to (1} is not guaranteed unless some restrictions are placed on the nature of f. By solution of (1) over an interval [0, 7), we mean an element x(-) of C*(0, T] such that (i) x(-)is differentiable everywhere, and (i) Equation (1)holds at al ‘We firstestablish some conditions under which (1) has exactly one solution over every finite interval (0, 8] for sufficiently small 8, i... conditions for local existence and unique- ness. Then we present stronger results which guarantee global existence and uniqueness, i.e.,conditions under which (1) has exactly one solution over [0, =). ‘One small point is to be cleared up before we proceed to the theorems. First, ifx() isa solution of (1) over (0, T]and fis continuous, then x(- also satisfies the integral equation 2 x)= +] ft, xO] dt, 1€ 10,7) (On the other hand, if x(}€C*[0, T] satisties (2), then clearly x() is actually differentiable ‘everywhere and satisfies (1). Thus (1) and (2) are equivalent inthe sense that every solution of (I) isalsoa solution of (2) and vice versa,34 Nonlinear Differential Equations ch? 24.1 Local Existence and Uniqueness 3 ‘Theorem (Local Existence and Uniqueness) Suppose the function € in (1) is con- tinuous int and x and satisfies the following conditions: There exist finite constants T, , h, ‘and k such that 4 5 IU, x)-fG, y)lI Sklx-yll, Vx, yeB, Wee (0, 7], UR, x0) SA, Vee (0, Th. where Bisa ball in R" of the form 6 B= (xeR":IIx—x9ll Sr}. Then (1) has exactly one solution over [0, 8] whenever the number 8 is sufficiently small to satisfy the inequalities 7 Bexp(k6)sr, forsome constantp <1. 9 Remarks 1, While following the proof of Theorem (3), itis important to keep in mind the dis- tinction between II-Il (which is a norm on R®), and t-llc, (which is @ norm on C"(0, 8}. Also, it should be noted that B is a ball in R", while S defined in (10) below isa ball in C"(0, 8], 2. The condition (4) is known as a Lipschitz condition, and the constant k is known as a Lipschitz constant. Notice that we say a Lipschitz constant, because if kis a Lipschitz constant forthe functioaf, then so is any constant larger than k. Some authors reserve the term Lipschitz constant for the smallest number ksuch that (4) is satisfied. A function that satisfies a Lipschitz condition is said to be Lipschitz- continuous. Note that a Lipschitz-continuous function is also absolutely continu- ‘ous [see Royden (1963)] ands therefore differentiable almosteverywhere. 3. Equation (4) is known as a local Lipschitz condition, because it holds only for all x.y in some ball around Xp, for re [0, T]. Accordingly, Theorem (3) is a local existence and uniqueness theorem, because it guarantees existence and unigue- ness of solutions over a sufficiently small interval (0, 5}. Note that, given any finite constants k, r, T and h, (7) and (8) can always be satisfied by choosing & sufficiently small.Sec. 2.4 Nonlinear Differential Equations 35 Proof By a slight abuse of notation, we use Xx9() to denote the function in C*{0, 6] whose value is xo for all € [0, 8]. Suppose 5 satisfies (7) and (8), and let S be the ball in C*0, 8] defined by 10 S={xCJEC"0, 3): XC) =X lleSr} Let P denote the mapping of C"(0 8 into itself defined by 11 (PRY) x9 +f ft, x] dt, Vee (0, 8}. Clearly x(-)is a solution of (2) over the interval [0, 8] if and only if (Px)() = fixed point of the map P. Jyie.x()isa Itis first shown that P is a contraction on S. Let x() and y(-) be arbitrary elements of $; then x(¢) and y(t) lie in the ball B, forall r= (0, 8]. Thus 12 (Pxy{e)—(Pye)= ft, x(2)]~ At WO) at, 13 WPHKO-CPpKON [HE x) —fIe. yO dt <{klxts)-y(oii dr SktIIx)-¥OMMe SplxC)-yOIhe. ‘where in the last line we have used the fact that kr . Since lix(t)~x9ll is a continuous function of t and since x(0)—Xo Il =O, there is a unique first time ot < to <8 with the pro- perty that 17 lix(t)—xpIl <7, Vre{0, 0), and fix(o)—Xo l= Now, since x() satisfies (2), we have 18 x(x =[ Mt x(D] dt 3 = { (fle, x(4)] -£06, Xo) +14, Xo)} 47, Vie [0, at, 19 Ux(—xpl'g| kllx()—xpll dehSec.2.4 Nonlinear Differential Equations 37 sha+] Klx(t) —x9 Il dt, Vee 10, a). ‘Equation (19) gives an implicit bound for the quantity {Ix(2)—xy ll. This implicit bound can be replaced by an explicit bound, using a result known as the Gronwall inequality [see Lemma (5.7.1). Applying this inequality to (19) gives 20 lix(t)—x9ll Shotexp (kad), Wee [0, a Inparticular, then, 21 Ix(@)~xp ll Shocexp (ka) < ABexp (K8)5r, by (7). But (21) contradicts (17). This shows that, if any function x()e C"(0, 6 satisfies (2), and 5 is sufficiently small that (7) holds, then x(-) must necessarily belong to S. Thus we have shown that any fixed pointof Pin C"{0, 8] must in fact be in S. Since P has exactly one fixed pointin S, it follows that P has exactly one fixed point in C"[0, 8]. By the manner in which P is defined, we conclude that (2) has exactly one solution over 0, 5). ‘The following result is actually a corollary to Theorem (3), but isin a form that can be readily applied. 22 Corollary Consider the differential equation (1). Suppose that in some neighbor- ‘hood of (0, X9) the function f(, x) is continuously differentiable. Then (I) has exactly one solution over (0, 8] provided Bis sufficiently smal Proof The differentiability properties assumed on f ensure that f satisfies (4) and (5) for some Setof finite constants r,T,kand h. Ml ‘Thus far we have studied the existence and uniqueness of solutions to (1) over closed iniervals of the form [0,8]. The reasons for this are primarily technical. For example, C*[0, 3]isa Banach space, but C*[0, 6) isa much more tricky object. But now consider the following question: Suppose f(, x) is continuously differentiable everywhere. What isthe largest interval over which (1) has @ unique solution? Looking back over the proof of ‘Theorem (3), one can see thatthe proof is equally valid ifthe initial time is changed from Oto an arbitrary time fo, and all hypotheses are adjusted accordingly. Thus, if (2) has a unique solution over some interval [0, 8] {which it will, by Corollary (22), then one can again apply Corollary (22) with 8 as the initial time and x(6) aS the initial state, and conclude that there is unique solution to (1) over some interval (8, 8’). This solution can be concatenated with the earlier solution over (0, 8] to construct a ui (0, 8’), But the process can be repeated yet again with 8 as th initial state. Since this process can be repeated indefinitely, we see that there is no largest closed interval over which (1) has a unique solution. Instead, there is a numbet Bax (hich ‘may equal infinity) such that (1) has a unique solution over every closed interval (0, 8] inthe half-open interval [0, yar) this solution is called the maximal solution, Now, what can happen a8 1 8pax? If Bmax is finite and if x(t) remains well-behaved as ¢—> 8p and38 ‘Nonlinear Differential Equations Ch.2 approaches some finite Vector Xmay, then one can again apply Corollary (22) with Bg as the initial time and Xpux 8 the initial state, and thereby extend the solution still further in time, ‘which contradicts the definition of Bas. Thus, if Spay is finite, then Ux(t)ll must approach infinity a5 + Bay. This discussion can be summarized as follows: 23. Corollary Consider the differential equation (1). and suppose that ft, x) is continu- ‘ously differentiable everywhere. Then there exists a unique number Byyx = Bpax(%), Which could equal infinity, such that (1) has a unique solution over (0, Byun) and over no larger interval. If8qux is finite, then NX(@)lI 2a Byae 24 Example Consider the scalar differential equation X= 1427, x(0)=0. Then 8,2, = 1/2, and the maximal solution is x fant, Predictably, x(t) —>c0as¢—> 1/2. A solution x(1) with the property that IIx(t) + as 1 approaches some finite time is said to exhibit finite escape time. Another question one can ask about the differential equation (1)is this: Ist possible to solve (1) for negative values ofr? The answer, under the hypotheses of Theorem (3), is yes. In act, if one is interested in solutions of (1) for both negative as well as positive values of , then Corollary (23) should be modified o speak of a maximal interval (~Byig, Spay) Father than a half-open interval. Generally speaking, in control theory one is usually not interested in solving for the past behsvior of a system, only its future. Thus the topic is not pursued further in this book. However, inthe theory of dynamical systems, one is often interested in both the pastas well asthe future of a system. The interested reader i referred to Hirsch and Smale (1974) for further details 2.4.2 Global Existence and Uniqueness Inthis subsection, we show that (loosely speaking) iff satisfies a global Lipschitz con- dition, then (1) ha a unique solution over alltime. 25 Theorem (Global Existence and Uniqueness) Suppose thatfor each T¢|0, ©) there exist finite constants k and hy such that 26 WEG, x)~fU, y)II Skrllx—yil, Vx, yeR”, Wre[0, T], 27 Wf,x9)lI Shy, Vie (0, 7} Then (1) has exactly one solution over (0, >).Sec.2.4 Nonlinear Differential Equations 39 Remark Recall that by a solution over (0, =) we mean a solution over [0, 7] for each finiteT. Proof We givetwo alternate proofs. Proof No. 1 Let T 8. Now (1) has a unique solution over {0, 8]. Denote this solution by x,(>), and consider the "shifted" dif- ferential equation 29D =MLe, xO), XO) =%18), where Wf x)=M0+8,%). ‘Then f, also satisfies (26) and (27); therefore once again by Theorem (3), (29) has unique solution over [0, 6], where 8 is the same as before. Denote this solution by y2(). Itiseasy to verify that the function x2(-) defined by x(t), Ostsb =| 90-8), 551525 is the unique solution of (1) over the interval (0, 28}. Proceeding by induction, let xq() denote the unique solution of (1) over the interval (0, m5], and consider the differential equation 32 (1) =Fa Lt, x(¢)], X(0) =X), where 33 f(t, x) =f + m6, x). Let ¥ms1 denote the unique solution of (32) over the interval [0,8] (the same 6 as before). ‘Then the function x4; (°) defined by40 ‘Nonlinear Differential Equations ch. Xn(Es Osrsmé BM msi OM) y, (t= m8), mBStS (m-+1)8 isthe unique solution of (1) over the interval (0, (m+1)8}. In this manner, the unique solu- tioncan be extended toall of (0, 7). Proof No.2 Let T Oitfollows that 39 Wm p(t) M(t) Bp gi 01 Hm it) & nent (mit! ret Sheep ®See. 24 Nonlinear Differential Equations 4 Now consider the sequence of sums 7 {Sra r } am -s% this sequence comerges thy) ep T). Morcover, the astern i trou taconite earl wean andes cones ear Th by choosing sulin large is sarecan be made acai sal Tis shows that (x,()} is a Cauchy sequence in C*[0,T}. Since C"(0,) is a Banach space, the sequence converges to alimitin C*(0, T]. Denote this limit by x*(>) ‘Whenever 2, (-)and 23(-)are two elements in C"[0, T)], we have a (PaO PaaMo = [Ue 21] MUe, 2200) 4, 43 Pz (Px MONI Sf IM, 240) —M1t, 220)]N de SkyT lt, —E, Mle 4 WPAIO-~Pa VOM c=, sup MPH NO —(P2VO 0, there exists a8(€, T) > 0 such that the following is true: Suppose xy and Yoare vectors inR® that satisfy $8 xy -yoll <8, 7). Suppose x¢) and y(-)are the corresponding solutions tothe differential equations 59 HU)=Alt, x), x(0)=Ho, 0 HW=ME YO}, YO=¥o. Then 61 AxG)-yOllese. Proof The functions x-) and y(-) also satisfy 62 x(0)=% +f Alt, x(t] dt, 63 y=yo+ | ft, yO] at. Subtracting, we get 64 x) -yO= =¥o+ f flr, x(O]—A1t, yeo)}} dt,“4 Nonlinear Differential Equations ch.2 6s Hx(e)— YON Slag —yoll +r [x(e)—y(OM ae Applying Gronwall’ inequality [Lemma (5.7.1)] to (65) gives 66 Uix(2)-Y(OUE Ixy —yoll exp pT). Hence 67 UxC)—YOIle Ixy Yo ll exp (kT). Thus, givene> 0, (61) is satsfiedif we choose &(€. 7)= exp (rT). ML Remarks 1. The results contained in Theorems (50) and (57) can be given a simple interpreta- tion in terms of certain mappings being continuous. Let ¢:R" —> C"(0, T] be the ‘mapping that associates, with each initial condition xy¢R", the corresponding ‘unique solution of (2). Then Theorem (57) states that @ is uniformly continuous con R®. In the same vein, let yy: R" +R" be the mapping that associates, with cach initial condition x €'R". the value at time T of the corresponding unique solution of 2). Then Theorem (50) states that wy is one-to-one fi. given Wr(x), ‘one can uniquely determine x], and onto (i... the range of Wy isall of R"). Furth- cermore, Theorem (57) shows that both yy and its inverse map 7! are continuous. Itis important to note that Theorem (57) is strictly limited tothe case where the interval [0,7 is finite. Theorem (57) does nor say that the solution aver the infinite interval (0, =) depends continuously on the inital condition xp. tn fact, We shall see in Chapter 5 that one possible interpretation of so-called Lyapunov stability is precisely that the solution over the infinite interval depends continu- ‘ously on the initial condition 68 Example Consider the scalar differential equation 69 i(1)=tanh[x()]=:f [(9], (O)=xp, Since the function tanh(x) is everywhere continuously differentiable, and since this deriva- tive is everywhere bounded (in magnitude) by I, itis easy to verify that (() satisfies global Lipschitz condition of the form (26) with kr= 1 for all T (see also Problem 2.13 below). Also, for every xo, there exist a finite constant hy such that (27) holds. Hence, by Theorem (25), it follows that (69) has a unique solution over (0, =) corresponding to each xo; more- over, for every finite number 7, the map taking x9 into the corresponding solution function in€{0, Tis continuous, by Theorem (57). 70 Example Consider the linear vector differential equationSee.2. Nonlinear Differential Equations 45 TL (t)= AC) XC, X00) =%o, where AC) is continuous. Let Ill be a given norm oa R®. Since AC) is continuous. for every finite Tthere exists finite constant ky such that WAC, Sky, Yee (0, 7) Hence it follows that WACx-A(Hyll Sky Ilx~yil, Wx, yeR®, We (0, T], WAN I Sky hte 0, T]. So (26) is satisfied with ky as above, and (27) is satisfied with hy =ky. Therefore, by ‘Theorem (25), (71) has a unique solution over [0, °) corresponding to each initial condition X. Moreover, over each finite interval [0, 7] this solution depends continuously on x. Ml In conclusion, in this section we have presented some conditions that are sufficient to censure that a given nonlinear vector differential equation has a unique solution over some interval, or over al intervals. [tis easy teconstruct counterexamples to show that the condi- tions presented here are by no means necessary for the existence and uniqueness of solu- tions. For instance, consider the scalar differentia equation 2 iQ =-2,x()=1. This equation has a unique solution over (0, ), namely x(1)= IAC), even though the function f (x) =x? isnot globally Lipschitz-continuous ‘Ata first glance the condition of Lipschitz-continuity appears to be extremely restric- tive, since it is known that “almost all” continuous functions are not differentiable and thus, not Lipschitz-continyous. Nevertheless, it can be shown that differential equations with ‘nique solutions are prevalent inthe sense that almost all” differential equations with con- tinuous functions Phave unique solutions. The arguments used to make this statement pre- cise and to prove itare quite advanced; therefore, they are presented separately in Appendix. ‘A. The contents f this appendix show that itis quite reasonable to assume that a given dif- ferential equation has a unique solution. This isa useful fact to know, especially when we study the stability of differential equations in Chapter 5. Problem 2.14 Show that Lipschitz-continuity is independent of which norm on R® is used, Precisely, let l-, and Illi, betwo given norms on R". Show that for each finite 7 there exists finite constant kr such that WR, X) L,Y) lle SKor IXY lay VX YER, Vee (0, 7]. ifand only if, foreach finite T there exists a finite constant kyy such that46 ‘Nonlinear Differential Equations ch. WE, x) £6 lly Shor IX-yily, VX, YER", Vte(0, T}. Problem 2.15 (a) Let f:R,xR —> R be continuously differentiable in the second argu- ‘ment. Show that fsatisties (26) ifand only if, for each finite T there exists a finite constant kr such that 12f.2)} rey Sky, VreR, Vie 0, 71, ive, 18f (1, xx | is bounded independently of x over each finite interval 0, T]. (Hint: Use the mean-valve theorem.) (b) Let: R, xR” +R" be continuously differentiable in the second argument. Show that satisfies (26) if and only if, foreach finite T there exists finite constant kr such that lagu, x)! A ck Wi js Ware RE, Wie (0,7) boy i se the results of Problem 2.14 above.) Problem 2.16 Determine whether or not the following functions satisfy a global Lipschitz condition: (@) fO=UF=xyx2 ey —xdT () £0) =Ee, exp) xzexp xy!’ 2.5 SOLUTIONESTIMATES In this section, we give a method for obtaining both upper and lower bounds on the norm of a solution of a given differential equation. The Gronwall inequality (Lemma (5.7.1)] does give an easily applicable upper bound on the norm of the solution of a linear differential equation, and a similar inequality known as Langenhop’s inequality provides a Tower bound. However, both of these bounds suffer from the deficiency of being sign- insensitive; ie., they give exactly the same estimates for 1 =x) asfor 2 id AW XC, This is because both Gronwall's inequality and Langenhop’s inequality (not presented in this book) utilize 1/A(0)1I, which i of course sign-insensitive. In contrast, the method given here is based on the concept of the matrix measure, which i sign-sensitive. As a result, the ‘bounds derived in this section are always “tighter” than (or the same as) those given by theSee. 2.5 Solution Estimates 47 Gronwall and Langenhop inequal 3. Theorem Consider the differential equation 4 K=AOX), 120, where x(1)eR" and A(1) is a continuous nxn matrix-valued function. Let WII be a norm on RY, and let Ill, and w(-) denote respectively the corresponding induced matrix norm and the corresponding matrix measure on R™". Then, whenever 1920, we have that ) 5 txta)texp] f-ml-Aco}at < Hatt sxe) exp: func Proof From Example (2.4.70), we know that the differential equation (4) has a unique solution over {0, e). To prove the inequalities (5), observe first that, from the integral form of (4), it Follows that 6 x{1+8)=x11)+5A(1)x(1) +066), VB>0, ‘where 06) denotes an error term with the property that 7 tim Lot 0 Rearranging (6) gives, successively, 8 x(+8)=[1+ BAD] x10) +06), 9 tix +8)l SUL + BAC IIxG +008), 10 Ux(e+8)ll ~ Mx(apll SU +8AC)I, 1) x +0), Ux(e +3) = thx IL 5 SHAG) IX, a a Fo lxet= im sa where d*/dr denotes the right-hand derivative. Multiplying both sides of (11) by the integrating factor 12 exp} -futA@)}de + (or, equivalently, applying the Gronwall inequality) gives the right-hand inequality in (5). “The proof ofthe left-hand inequality in (5) is entirely similar, starting with48 Nonlinear Differential Equations Ch.2 13 x(1~8)=x(1)-8A() x(1) +068). ‘The completion of the proof isleftasan exercise. Ml ‘Theorem (3) provides both upper and lower bounds for the norm of the solution of the unforced linear equation (4). In applying the bounds (5), tis important to remember tha the norm being used and the measure must correspond to one another. Also, using different norms in Theorem (3) will give rise to distinct bounds. This is illustrated by the following examples, 14 Example Consider the equation (4) with n =2and 4 First Ietuscalculate the measures Hi, Hp Ps ofthe matrix A((). This gives Aw= 1 1] x= WLAOL=B[AW] =—e41, Hi-A@] =A] = 241, mIAOL 1 pal-A(O]= 2 ‘Thus, applying the inequalities (5) with each of the above measures gives expt) s Ixy(0)1 + Le2(0)1 Sexpr-72), exp(-f-F)S Ixy, Lxg(0! Sexp-r-72), exp (—1?)<[ 1x, (1)1? + 1x2(t)17]' Sexp (-72). ‘Thus the same two inequalities (5), when applied with different vector norms and corresponding matrix measures, yield different estimates for the vector x(t). By way of illustrating the bounds obtained above, the regions of R? to which the vector x(1) is confined by each of the above bounds are shown in Figures 2.3, 2.4, and 2.5, respectively. 15 Example Consider the equation (4) with n=2and a ‘Then the actual solution for x(t) is “3 of AW=| 354 —4e| + X(1)=(4B)exp (-17)—(LB)exp(-S72)(4/)exp(-1?) + (2B)exp9-Sr72)I. However, if we calculate the various measures of A(1), we getSeo.2.5 Solution Estimates 9 ip23 2 1 iG A Oa Lil 3 Fig24 WAC] =—1 wl-ACOI=5t, HolA()]=-2.97t, wyl-ACO1=5.031, HALA(]=-26, wal-AD] = 6, Thus the corresponding estimates for x(t) are as follows: Bexp(-2.517)S Ixy (1 + Lx2()1 $3exp(-0.577), NB exp(~2.5217)s [Lx (1? + 1xea(0)17]? V5 exp (~ 1.4857), 2exp-31*)¢ Ix (I, Ix S2exp(-7). ‘The bounds are depicted forthe case 1=0.5in Figures 2.6,2.7 and 2.8, respectively.Sec. 2.5 Solution Estimates 31 Fig.28 To extend the above estimation technique to nonlinear differential equations of the form 16 i= Alt, x(0)], x(0) = preliminary resultis needed. 17 Lemma Suppose: R.xR" — R" is continuously differentiable. Then there exists a ‘continuous function A: RXR" — R°™ such that 1B, x)=Mlr, 0)+ AG, x)x, V0, VaER", Proof Fix and.x,and consider f(, Ax)asa function ofthe scalar parameter. Then 19 t,x) 4 - \ 4 a os] Gul en dh = fle, O)+ if Ville, Doda) x. Hence (18)holds with32 Nonlinear Differential Equations Ch.2 20 AG, x= Vyflt, Ax) da Note that there is nothing special about the origin in the above formula. Indeed, given any fixed xo¢R", wecan write 21 fle, x)=fl6, x9) + BU x.) XX) for suitably chosen matrix-valued function B(t, x, Xo) 22 Theorem Consider the differential equation (16), and suppose (i) fis continuously differentiable, and ii) (t, 0)=0 Vt 20. Define A(t, x) as in(18). Let Ill be a norm on R°, ‘and let \-'), and w(-) denote the corresponding induced norm and matrix measure on R™"" Suppose there exist continuous functions a-) and B() such that 23 WIAG, 01 Sot), BO) SHI-AG, x3], ¥F20, VxER’. Then 24 Ixollexp| -[ B(x) dz] 0 (but not if r<0). Atan(0,0) is undefined, 8 Definition A function: R? -> R? iscalled a veetor field if both ofits components are smooth functions. A vector xe R? is called an equilibrium of a vector field tif f(x) =0. If XE R? is not an equilibrium of f, then the direction of the vector field f at the point x is denoted by _(x) ands defined as 9 Ox) =Atanl f(x), f200)) Figure 3.3 depicts the quantity (x) iis fix) A a) Fig.33 To see the utility of these concepts, suppose f:R? > R? isa vector field, and consider the associated differential equation WO ky =fy(x 1,42), 42 = fale, 2). Note that here and in the remainder of the chapter we follow the standard practice of not ‘explicitly displaying the time variable Suppose x= (x1. 2) is point in R?; then itis easy to see from (10) that Cis solu- tion trajectory of (10) passing through x, then the vector f(x) is tangent to Cat. Hence, in principle a¢ least, itis possible to construct graphically the solution trajectories of (10) by plotting the vector field f(x). Actually, the concept is very deep and has many applications,56 Second-Order Systems ch only a few of which are touched upon in this book. Furthermore, the concept of a vector field is applicable to (autonomous) systems of any order. ‘The reader interested in a deeper knowledge of the application of the vector field concept to differential equations may con- sult Amold (1973). Vector fields are encountered again in this book in Chapter 7. Note that itis quite common to refer to f(x) as the velocity vector field associated with the system of equations (10). ‘The objective of the present chapter is to present some ways of analyzing the system (10) by either finding the state-plane trajectory of the system to a reasonably high degree of accuracy or determining some qualitative features of the state-plane trajectory without doing too much work. ‘Throughout the chapter, the study is confined to autonomous sys- tems, because even though the concept of a state-plane trajectory is valid for nonauto- ‘nomous systems, most ofthe significant results are applicable only to autonomous systems, For example, the autonomous system (10) has a periodic solution x() ifthe corresponding solution trajectory is aclosed curve in R®, An analogous statement for nonautonomous sys- temsis false in general. Finally, a word about the existence and uniqueness of solutions o the system of equa- tions (10). Since fis smooth, it follows from Corollary (2.4.22) that (10) has a unique solu- tion at least locally; thats, given the system (10) together with an initial condition 1 x)= x10. 220)= 229, there exists a number 8 such that (J0-11) has exactly one solution over [0, 8). Additional conditions on f ensure that (10-11) has a unique solution over all of (9, ); see Theorem (2.4.25). Problem 3.1 Prove the relationship (6). Hint: Use (5) to write yE+AN =x (0) +Arx3(0) +0 (AN). Problem 3.2 Show that if C is a solution trajectory of (10) passing through x, then the ‘ecto field f(x) is tangent to Cat x. Hint: Express (10) in difference formas, A +AN =x (0+ Atfilx1(0.x20]+0(40, ag{t+An)=22(0) +44 falzy(, 7210] +0(08), andeeliminate Aras Ar —+0, Problem 3.3 Does the function f:R? —> R? defined by Siley #2)= 32401 - OF +33)"7), xy + [1 = (xf +03)" Ale X2) constitute a vector field? Justify your answer. Hint: Consider the behavior of f near theSee, 3.2 Linearization Method 7 origin. 3.2 LINEARIZATION METHOD ‘We begin by studying linear systems, which are simpler to analyze than nonlinear sys- tems and yet provide much insight into the behavior of nonlinear systems. The general form of a second-order autonomous linear system is 1 Ry=ayxy taararty anes taxa, together withthe inital conditions 2 x1) =x 19,220) = x29 In matrix notation (1) and (2) can be expressed as 3) =AxU), x(O)= ‘To.understand better the behavior of solutions to (3), its helpful to make a transforma- tion of variables. Accordingly, let 4 Mx, where M is a constant nonsingular 2x2 matrix with real coefficients. In terms of the transformed variable 2, (3) becomes Sift) = MAM" 2(0),2(0)=Mxp. Itisknown (see, for example, Bellman (1970)] that by appropriately choosing the matrix M,, the matrix MAM" can be made to have one of the following forms: 1. Diagonal form: inthis case, a, 0 [>a where A, and Ay are the real (and not necessarily distinct) eigenvalues of the matrix A. 6 MAM" 2. Jordan form: In thiscase, [) where isthe real repeated eigenvalue ofthe matrix A. 7 MAM"58 ‘Second-Order Systems ch.3 3. Complex conjugate form: Inthis case, a-B 8 MAM {5 |: a where cj are the complex conjugate eigenvalues of A, and we choose B>0 10 be definite, Each ofthese cases is studied in detail ‘Case 1 Diagonal Form: In this case (5) assumes the form 9 dt) 121(0), 22) =Anz2(0), 24(0)= 240, 220) =239 The sol nof (9) is WW z1()=ze*22()=209e™ At this point it can be assumed that not both A, and 2g are zero, because if both Ay, Re are zero then A =0 and 2(1) =29 for all consequently the state-plane plot consists of just asin gle point. Thus suppose A, #0. Then the parameter can be eliminated from (10) to give hey a 20'| 20 Equation (11) describes the state-plane trajectory of (9) in the z1~z2 plane. If, and Ay are of the same sign, then the trajectories have the characteristic shape shown in Figure 3.4, but if} and 2 have opposite signs then the trajectories have the characteristic shape shown in Figure3.5. The arrowheads in Figure 3.4 correspond to the case where hy < Ay <0sifAy and ‘are both positive then the direction of the arrowheads is reversed, and the trajectories go ‘away from the origin as 1 increases instead of going towards the origin as in Figure 3:4 Similarly the arrowheads in Figure 3.5 correspond to the case where A, <0-< hz. It should be emphasized that the trajectories depicted in Figures 3.4 and 3.5 are in the z;~z2 coordi- nate system; the corresponding trajectories in the x2 coordinate system, although they will have the same general appearance as those in the z,~z> coordinate system, will be alit- tle distorted. This can be seen in Figures 3.6 and 3.7, where the trajectories in the x ~x2 coordinate system are illustrated for the cases where 2 and My are of the same sign, and where A; and Ap are of opposite signs, respectively. If A; and 2 are ofthe same sign, then, the equilibrium at the origin is referred to as a node. It is called a stable node if both 2, and 1 are negative, andan unstable node i, and 2 are positive. In the case where 2 and axe of opposite sign, the equilibrium atthe origin is called a saddle. The rationale for this nomenclature is that if one were to make a three-dimensional plot of [x (t),x2(1), 1), then the resulting surface would resemble a saddle. MizSec.3.2 Linearization Method 59 Fig.3.4 Fig.3.50 ‘Second-Order Systems Ch.3 Fig.3.7 (Case 2 Jordan Form: In this case (5) assumes the form 12 2 =Az +200 E =Aza(0, 21(0)= 240, 22(0)= 22, The solution of (12)is 13 2y(t)=z19e¥ +2a08 e™, 22(t)=20e™. ‘Once again, ¢can be eliminated from (13); the resulting expression describing the trajectory is somewhat messy and its derivation is left as a problem (see Problem 3.4). The trajectories in the z,~z2 coordinate system, which can be obtained from (13), are shown in Figure 3.8 for the case &.<0: if 4 > 0, then the direction ofthe arrows is reversed. The corresponding trajectories in the x ~x2 coordinate system are shown in Figure 3.9. In this case also, the ‘equilibrium (0, 0)is called a stable node if 2. < Oand an unstable node if A > 0. Fig.38Sec.3.2 Linearization Method 6 Fig.3.9 Case 3 Complex Conjugate Form: In his case (5)6ecomes 14 24) =024(0)~ Beal), 22(0)=Be (0) +0201), ‘To simplify the equation further, introduce the polar coordinates 1S r=(e} +23)", = Atan(z,, 22). ‘Then (14) is transformed into art), 0 which has the solution 17 r(t)=r(O)e™, )=40) + Be In the z,~z2 coordinate system, (17) represents an exponential spiral. If o> 0, then the spiral expands as t increases, whereas if «<0, then the spiral shrinks ast increases; and if = 0 the trajectory isa circle. The equilibrium (0,0) is referred to as an unstable focus if >0, stable focusiif «<0, and a center if =0. The trajectories inthe z,~z2 coordinate system corresponding to each of these cases are depicted in Figures 3.10,3.11 and3.12. Table 3.1 summarizes the various kinds of equilibria for second-order linear systems. Note that 4, 2 are the eigenvalues of the matrix A. ‘Now consider an autonomous nonlinear system described by62 ‘Second-Order Systems 2 a>o Fig.3.10 a a0 Unstable focus dy. Ao imaginary Center 1B Ey =f ey x2) ¥2= foley 42). ‘The linearization method, as the name implies, consists of linearizing the given system in the neighborhood of an equilibrium and determining the behavior of the nonlinear system by studying the resulting inear system. The power of the method lies in the fact that, except for special cases tobe specified later, the method yields definitive results that are valid in some neighborhood of the equilibrium. “The method can be summarized as follows: Suppose (0,0) is an equilibrium of the sys- tem (1) and that both f and f, are continuously differentiable in some neighborhood of (0,0). Define 2 Aza gl ax an] ‘Then, by Taylor's theorem, itis possible toexpand f, and fin the form 2A fulzxad=f(0, Ota nix, +aiae tr (x12) Fayx) tayrytrGrz2) fale xs)=aaity tanks tale). + Fy (60) where r, and r are the remainder terms, and we have used the fact that /(0, 0)=0 since (6, 0) is an equilibrium. Ifthe equilibrium is not at (0, 0) but at some other point in R?, then ‘one can always translate the coordinates in such a way that the equilibrium is at the origin of the new coordinate system, Now, associated with the nonlinear system (18), define the TinearsystemoF ‘Second-Order Systems ch 22 Er=aybi +ainbs, bab) +022 ‘The linearization method is based on the fact (proved in Section 5.5) that if the matrix A does not have any eigenvalues with zero real parts, then the trajectories ofthe nonlinear sys- tem (18) in the vicinity of the equilibrivm x. =0, x2 =0 have the same characteristic shape as the trajectories of the Linear system (22) in the vicinity of the equilibrium &, =0, Table 3.2 summarizes the situation. Table3.2 Equilibriumofthe Equilibrium of the Linear System (22) Nonlinear System (18) ‘Stablenode Stable node Unstablenode Unstable node Saddle Saddle Stable focus Stable focus Unstable focus Unstable focus Center a ‘The lastentry in the table can be explained as follows: If the equilibrium (0, 0) ofthe system (22) isa center, then the linearized system exhibits perfect oscillations which neither grow nor decay with time. In such acase, the behavior of the trajectories of the original nonlinear system is determined by the remainder terms 7 and rs, which are neglected inthe lineariza- tion, Studying the linearized system alone does not provide a definitive answer about the behavior ofthe nonlinear system. 23 _ Example Consider the following second-order equation, commonly known as Van der Pol's equation: A F-wl-y?pity= ‘where 1 > Ois a constant. By defining the natural state variables xeyaeeh (24) is transformed ito the pair of first-order equations 28 i =eqige—ay +ml-a})a2 The linearization of (25) around the equilibrium (0, 0)is 26: =Es,Eo=-E, +. ‘The cigenvalues of the associated matrix A satisfy the characteristic equationSec. 3.2 Linearization Method 65 27 B-pAsl I positive values of H, the roots of (27) are complex with positive real parts, so that the ‘of (26) isan unstable focus. Referring to Table 3.2, we see that the Oof the original system (8) is also an unstable focus. Figure 3.13 shows the phase-plane trajectories of the Van der Pol oscillator. A notable feature of this system is that all solution trajectories starting from an initial state other than (0, 0) approach alimit cycle. This systemis further analyzed in Section 3.4, Problem 3.4 Eliminate from (13) and obtain an expression for the state-plane trajec- tory involving only z,, 225 210, and Zao 32 19 Fig 3.14 Problem 3.5 Consider the electrical circuit shown in Figure 3.14. (a) Select the capacitor voltage x , and the inductor current x as the state variables, and show thatthe network is described by the equations66 ‘Second-Order Systems ch3 2-Day ay #2y, by =a) (©) Suppose v(t)=0. Determine the nature of the equilibrium at (0,0) and find the matrix M that iransforms the above equations into the appropriate canonical form. Problem 3.6 Suppose the \/2 Ohm esistorin Figure 3.14 isreplaced by a general resis- tor. (a) Write the state equations for the network with v (1) =0. (b) For what values of & isthe equilibrium (0, 0) (i) node, Problem 3.7 Foreach of the matrices A given below: (a) Determine the matrix M that transforms A into the appropriate canonical form, (b) Sketch the state-plane trajectories in both the z1-22 coordinates and the x -x2 (©) Classify the equilibrium at (0, 0)as to its type. ‘coordinate system, 0 1) fo-t] fa o-1 2-3/1 2/-/0 2-2 Problem 3.8 Find all equilibria ofthe Volterra predator-prey equations A Gy Sn) tty, dye hp Linearize the system around each of the equilibria and determine, if possible, the nature of ‘he equilibrium. (Answer: One center, one saddle) 12 Fig.3.15 Problem 3.9 Consider the nonlinear circuit shown in Figure 3.15. Suppose the voltage-current relationship of the nonlinear resistor is given by 3H? +34, =f (i).See.3.3 Periodic Solutions 67 (a) Select the capacitor voltage x, and the inductor x; as the state variables, and show thatthe system equations are f(x). (b) With » =0, calculate the equilibria ofthe system. (©) Linearize the system around each of the equilibria and determine the nature of each equilibrium, Ry evexyony ine. 3.3 PERIODICSOLUTIONS 3.3.1 Introduction ‘Some autonomous systems exhibit periodic solutions. For example, consider a simple harmonic oscillator, which is described by the linear equations 1 yeas: x ‘The solution of (1) subject to the initial conditions 2 x1(= 240.2200 = 220 isgivenby 3 x(=rycos(-1+09),x2(0)=r9 sim(—14+ 0), where 4 19=(tig +240)", 9 = Atan(x 10, X20)- ‘Thus the solution of (1) is periodic irrespective of the initial conditions. Furthermore, the cetire phase-plane is covered with periodic solutions of (1): Given any point (xo, X29), one ‘can always find aperiodic solution passing throught In contrast, consider the system of nonlinear equations 8 exp tory Goa] 9), =n, torn Pad) ‘These equations can be expressed as 6 k=f(x) + B(x), where8 ‘Second-Order Systems ch.3 1 Note that fis exactly the velocity vector field of the system (1), while gis a so-called radial vector field, ie., g(x) is always aligned with the vector x. Now introduce the polar coordi- nates Si) eeu -x)) 8 rata} +14)", o= Atan(x 2). ‘Then the equations (5) are transformed to 9 B-7?).6 Itcan be easly verified thatthe solution of (9) is 0 r@=——_f__. 41) = 900), OO" Trccexp (apron? * where Thus the system (5) has only one nontrivial periodic solution, namely r= -xfp +149 =B°. (Note thatthe equilibrium solution x, =0, x3 =0 can also be considered a trivial periodic solution.) Furthermore, if > O, any solution of (5) with r(0)#0 approaches this periodie solution as t 2. This example differs from the earlier example of a simple hharmonic oscillator in that there is only one nontrivial periodic solution, and moreover, this periodic solution is isolated, ie. there exists a neighborhood of it that does not consain any other periodic solution, cis common to refer toa nontrivial periodic solution as a limit eyele, Note that some authors reserve this phrase only for an isolated periodic solution. By convention, an equili- brium is not regarded as aperiodic solution. Inthe remainder ofthis section, some results are presented pertaining tothe existence ‘orabsence of periodic solutions in nonlinear systems. 3.3.2 Bendixson’s Theorem Bendixson’s theorem presents a simple sufficient condition to guarantee that a given simply connected domain in the plane does not contain a periodic solution. Before stating the theorem, the terms “domain” and "simply connected” are defined. A domain in R? is just an open set. A subset Sc R? is simply connected if itcan be continuously shrunk toa single point in $, ie. if ere exists a point x9 S and a continuous function h0, 1]xS > such thatSec.3.3 Periodic Solutions 0 12 h(0,x)=4,4(1, x) =x, VES. For example, aclosed disk is simply connected, whereas an annular region isnot. 13° Theorem Consider the second-order system 16 X= file tata =fyOnx9). Suppose Disa simply connected domain in R° such thatthe quantity Vf(x) defined by af fr 1S Vi)= Sen) Stet) a arp is not identically zero over any subdomain of D and does not change sign over D. Under these conditions, D doesnot contain any nontrivial periodic solutions of(14). Proof Suppose J is a closed trajectory of (14). Then at each point x=(r),.x3)€J, the velocity vector field f(x) is tangent to J. Let n(x) denote the outward normal toJ at x. Then {(x)'n(x) =0 forallxeJ. Therefore 16 f foxx) di =0. d Butby the divergence theorem, 17 f f(xym@x)al = ff Veo ax ; > where S's the area enclosed by J. Therefore, in order for (17) to hold, either Vflx) must be identically zero over S, or else Vflx) must change sign over S. But if Sis a subset of D, then the hypotheses of the theorem rule aut both possibilities. Hence D contains no nontrivial periodic solutions of (14). i 18 Example Consider the application of Theorem (13) tothe linear system of equations Fy sary tayx2 ky =ane tanra From Section 3.2 we know that a necessary and sufficient condition for the system to have periodic solutions is that the matrix au an Az 4 an have two nonzero imaginary eigenvalues. Since the eigenvalues of A are the roots of the characteristic equation0 Second-Order Systems ch.3 tla tan)A+(ayax.-412421)=0, itis clear tha the system has periodic solutions if and only if, 44, +42 =0, 4), - 412021 >0. Equivalently, a necessary and sufficient condition for the absence of periodic solutions is, that either of the above conditions be violated. Applying Theorem (13) tothe present case gives Vatx)= titan, WxeR?. Hence Bendixson’s theorem states that if a, +2 #0, then the system has no periodic solu- tions, which is consistent with the previous discussion, 19 Example Consider the system of nonlinear equations Sy aap tauhigeany tabey ‘The linearization of this equation around the equilibrium at the origin is isan ts a continuum of periodic solutions. However, for the nonlinear system we Vix)=x} +2] > Ova. ‘Thus VE never changes sign, and is zero only atthe origin (which is not a subdomain since it is not an open set). Hence Bendixson’s theorem leads to the conclusion that the system under study has no nontrivial periodic solutions. 20 Example In applying Theorem (13), the assumption that D is a simply connected domain is crucial — itis not enough for D to be just connected. (A subset D of R? is said to bbe connected if every two points in D are connected by a path lying entirely in D. Thus an annular region is connected but not simply connected.) To see this, consider the system (5), and letD be the annular region D=(r,42): 283 ee and x(t,) >. The set ofall limit points ofa trajectory is called the limit set ofthe trajectory and is denoted byL Remarks Basically, a limit point ofa trajectory isa point to which the trajectory passes arbitrarily close infinitely many times as time progresses. We shall encounter fimit points and limit sets again in Section 5.2. 22 Theorem (Poincaré-Bendixson) Let 23 S={x(),120) denote a trajectory in R° ofthe system (14), and let L denote its limit set. IfLis contained in ‘closed bounded region M in R? and ifM contains no equilibria of(14),then either (@_Sisaperiodic solution of 14), or Gi) Lisaperiodic solution of 14). ‘The proof is omitted as itis beyond the scope ofthe book. Remarks Roughly speaking, Theorem (22) states the following: Suppose we can find a closed bounded region M in R? such that M does not contain any equilibria of (14) and such that all limit points of some trajectory S are contained in M. Then M contains atleast ‘one periodic solution of (14). In practice, an easy way to verify that M contains all the limit, points ofa trajectory Sis to verify that S eventually lis entirely in M, ie. to show that there exists atime T such that x(¢}¢ M Vr 27. Thus the theorem reduces to this: If we can find a closed bounded region M containing no equilibria such that some trajectory is eventually confined to M, then Mf contains a least one periodic solution. Now, asufficient condition for trajectory tobe eventually confined to Mis that, atevery point along the boundary of M, the ‘velocity vector field always points into M. If thisis true, then any trajectory originating from within M must semain in M, and hence M contains at least one periodic solution of the sys- temathand. (This is depicted in Figure 3.16.) 24 Example Consider once again the system (5), and let M be the annular region defined by M= ((4.42):0.98 Saf +43 $1.18") ‘Then M contains no equilibria ofthe system. Moreover, a sketch of the velocity vector field reveals that all along the boundary of M, the vector field always points into M, as depicted inn ‘Second-Order Systems Ch.3 rs Kio n>o Fig.3.16 Figure 3.16. Hence we can apply Theorem (22) and conclude that M contains a periodic solution Fig.3.17 25 Example In applying Theorem (22), the condition that M does not contain any equili- briais indispensable. To see this, consider the system. Sy soxy tay so xy ae ‘The velocity vector field for this system is sketched in Figure 3.17. If M is chosen to be the closed unit disk centered atthe origin, then all along the boundary of M the velocity vector field points into M. Hence all trajectories originating in M remain within M. The same con- clusion can be reached by analytical reasoning because, in polar coordinates, the system equations become noe which has the solutionSee. 3.3 Periodic Solutions B r(Q=rOexp 1), 0) =9(0)—t -M, M does not con- this case because M However, even though all trajectories starting within M remain wit ‘ain any nontrivial periodic solutions. Theorem (22) does not appl contains the equilibrium 0. 3.3.4 Index Theorems ‘The concept of index is a very powerful one, and the results given below only scratch the surface of the many results that are available. Unfortunately, the arguments involved in index theory are well beyond the scope of this book. Hence almost all of the results presented in this subsection are stated without proof. For further discussion, see Nemytskii and Stepanov (1960), ‘The definition below introduces the concept of index. Recall that a point xe R? is called an equilibrium of a vector field Ff f(x)=0, and recall also Definition (3.1.8) of the direction of a vector field ata point (other than an equilibrium). 26 Definition Suppose D is an open, simply connected subset of R, and suppose £:R? 5 R? isavector field on R®. Suppose D contains only isolated equilibria ofthe vector field f. Let J be a simple, closed, positively oriented Jordan curve in D that does not pass ‘through any equilibria of f, and let (x) denote the direction ofthe vector field at x. Then the index ofthe curve J with respect to the vector field fis denoted by Iy(J) and is defined as 27 id= x J 46,(x). Remarks A positively oriented curve is one which is traversed in the counter- clockwise direction, ie.,a curve with the property that the area enclosed by it always lies 0 the left of it, Since itis assumed that J does not pass through any equilibria of f the direction vector (x) is well-defined at all xe J. The index of J with respect tof is just the net change in the direction of f(x) as x traverses around J, divided by 2x, Clearly /,(J) is always an integer. 28 Definition Letpbe an isolated equilibrium of the vector field. Then the index of pis denoted by Iy(p) and is defined as I,(J) where J is any suitable Jordan curve such that (i) pis enclosed by J, and (i) J does not enclose any other equilibria off. Note that the same symbol /,() is used for both the index of a closed curve and of an equilibrium: [Now some facts are stated without proof. 29 Fact Suppose Jdoes not enclose any equilibria of. Then I¢(J)=0.“ Second-Order Systems ch 30. Fact The indices ofa center, focus, and node are each equal to |, while the index of a saddleis-1. This fact can be verified by sketching the vector field near each of the above types of equilibria. 31 Fact Suppose J encloses afinite number of equilibria off, say py, °**.Pa» Then 32 Id) = Dll, a 33 Fact Let fand g be two vector fields on R®. Let J be a simple, closed, positively oriented Jordan curve, and suppose that and g.are never in opposition along the boundary of Iie. suppose that \8q(x)~0,(x)| < rat all x along the boundary ofJ. Suppose in addi- tion that J does not pass through any equilibria of either forg. Under these conditions, 3 WD)=1,U). This fact follows from Definition (28) and the fact that both J¢(J) and 1,(J) are integers. 35. Fact Let bea simple, closed, positively oriented trajectory ofthe system Then 37 dah ‘This can be seen from the fact that the vector field fis always tangent to J. (Onthe basis ofthese facts, we can state the following general theorem. 38 Theorem Suppose the system (36) has only isolated equilibria. Then every closed trajectory of(36) ifany) encleses atleast one equilibrium. Moreover, the sumof the indices ofthe equilibriaenclosed by the closed trajectory is equal t0 | 39 Example As an illustration of Theorem (38), consider the Volterra predator-prey equations (introduced earlier in Problem 3.8) M0 ieee teats aka Let us digress briefly to discuss the rationale behind the above model. Let x, denote the number of predators (foxes, let us say), and let.x2 denote the nomber of prey (rabbits). If x2 =0, then the first equation in (40) reduces to x, =—.x), which states that in the absence of prey the number of predators will dwindle exponentially to zero. If x20, then the same {equation shows that , contains an exponential growth term proportional tox. The sitia- tion in the case of x2 is just the opposite. Ifx =0, then x will grow exponentially, while ifSee.3.3 Periodic Solutions 18 x) #0, then. contains an exponential decay term proportional tox Fig.3.18 ‘The velocity vector field for the predator-prey system is shown in Figure 3.18. Clearly there are two equilibria, namely (0, 0) and (1, 1). By linearizing the system (40) around each of these equilibria one can readily determine that (0, 0) is a saddle while (1,1) is a center. Hence the index of (0, 0) is~I while the index of (I, 1) is 1. Now, by Theorem (38), any closed trajectory of the system (40) must enclose (1, 1), and it must not enclose (0, 0). ‘Thus, by examining the index alone, one can derive a great deal of qualitative information about the possible closed trajectories ofa system, 33.5 An Analytical Method In this subsection, a technique is presented for obtaining analytical expressions for the closed trajectories of some nonlinear systems that exhibit a continuum of periodic trajec- tories. Rather than presenting a general theorem, which would have to be rather weak ‘because of all the possible pathological cases, we illustrate the method by means of a few examples. ‘The basic idea of the method is as follows: Given the system (14) and a continuously differentiable function V: R? > R, define the function V:R? —> Rby av av A Ver aad= SF Fiza) + 3 fale 2) The function Vis known as the derivative of V along the trajectories ofthe system (14), because if [x (.),¥2(0] is atrajectory ofthe system (14), then the derivative with respect tot of the function V [x (t), x2(d)]is precisely V[x (1), x2(t)]. We shall encounter this concept ‘again in Section 5.2. Now suppose we are able to find a domain D in R® such that V(x, x2) #0 for all xe D. Let (x 9, X29)€ D, and let C denote the solution trajectory of (14)16 ‘Second-Order Systems Ch.3 ‘originating from (xo, x29). The hypothesis that Vis identically zero implies that V(x, x2) isconstantalongC. In other words, 42 Vixi(0,x2(0]= VO gx), VI20. Letusnow consider the set 43 S= 1604.52): VO, 42) = VOR I0.X0)): ‘Then C is a subset of S. In particular, if 5 is itself a closed curve, then we can conclude (under reasonably mild additional assumptions) that C is itself a closed trajectory and is ‘equal to. Of course a great deal depends on the choice of the function V. If we choose V(x,,2)=1 for all (x;,.r2), then naturally V=0; but the set S in (43) is all of R?, and as.a res no insight has ben gained into the nature ofthe trajectories. However, nsome cases, by property choosing V, we can show thatthe family of sets 44 (xy rn: Vet an) =c) asc varies over an appropriate subset of real numbers, defines a continuum of closed trajec- toes of the system (14). 45 Example A very simple application which illustrates this procedure is the harmonic oscillator x Latutchooee Veriodext ead ‘Then Wa2e, (x2) +2e2-x))=0. Since the equation V (x), x2)=c defines a closed curve for each c > 0, we conclude that the system above exhibits a continuum of closed trajectories, described by 2 exdext, +x, aq ta} axfy tae. 46 Example Consider again the predator-prey equations (40), and try a function Vof the form 47 VG, 2)=hy (ay) +h ole) where h and h are tobe selected so that Vis identically zero. NowSec.3.3 Periodic Solutions 1 Very 2) = hi (ay) ey He) +S Ce) ea D)- In order that Vbe identically zero, itis necessary that Hy Cey)x4 G2 = A549) a (121) =0, which can be rearranged as ‘The left side ofthis equation is independent of x2, while the right side is independent of x Hence, in fact both must equal a constant, say c. In other words, 4 Ken pee Med Tae The solution of (48) is hy(xy)=e(Inx, —x), h(a) =c(Inx2 42) Hence an appropriate choice inthis example is V(x.x2)= (nx, x; +Inx2—x2), where the arbitrary constant c has been dropped without loss of generality. For the above choice of ¥, any set of the form (43) is actually a closed curve. Hence the family of curves defined by Inxy =x) +Inx2~x2=const constitutes a set of closed trajectories for the predator-prey system. Note that V is defined only in the first quadrant, ie.,if-x, >0, x2 >0. 49 Example Consider the pendulum equation 6+ fi sin@=0, ‘where @ denotes the angle of the pendulum from the vertical axis, g is the acceleration due to gravity, and fis the length of the pendulum. With the natural choice of state variables x1 =0,x2=6, the pendulum equation can be rewriven asB ‘Second-Order Systems ch. Letusonce again choose V to be of the form (47). Then, in order for Vto be identically zero, wemusthave WiC dxa~ FRG) sin, =0, which implies, asin Example (46), that ile) _ shoe) sing, Lg sonst. =¢ ‘The solution of these equations is & h(x =e 00824, hale Hence the family of curves 2 c0s.x} =const, constitute a set of closed trajectories of the pendulum equation. Remarks Examples (46) and (49) illustrate how the method presented here can some- times yield good results. However, it should be clear that (i) a function V of the form (47) does not always work, and (ii) even fit does, there is no guarantee that all closed trajectories are of the form (44). Despite these limitations, however, the method nevertheless has some value, as indicated by these two examples. Problem 3.10 Consider a mechanical system consisting of a unit mass, a nonlinear spring, and a nonlinear damper. Such a system can be modelled by the set of equations Breanna =-g(e)-hle where x isthe postion ofthe mass, () isthe restoring force of the spring, and h() i the damping force exerted by the damper. Assume that both g(-) and h() are continuously dif- ferentiable. Using Bendixson's theorem, show that ths system has no periodic solutions if hi G40 for all £40, ie, there is always some amount of damping when the mass is in motion Problem 3.11 Sketch a vector field with exactly one node and one saddle. Show that it is not possible to deform this vector field continuously in such a way that there i a periodic solution enclosing both the node and the saddle.See.3.4 Analytical Approximation 9 Problem 3.12 Using the method of Section 3.3.5, show that the undamped unit mass- nonlinear spring system described by yan tn =-g (ny) ‘always has a continuum of periodic solutions if, 21 g(r) 00x, 20. Derive an expression for these closed trajectories, 3.4 TWO ANALYTICAL APPROXIMATION METHODS In this section, we describe two techniques for obtaining analytical expressions that approximate the periodic solution of second-order nonlinear differential equations. In con- trast with the method presented in Section 3.3.5, which gives exact expressions if it works, the two methods presented here are only approximate. However, they have the advantage of ‘having a wide range of applicability and of enabling one to study the so-called “slowly vary- ing” oscillations. t should be emphasized that, depending on the particular problem to which they are applied, one technique might work better than the other. Moreover. the two ‘methods presented here are only a small part of the numerous techniques that are available for analyzing slowly varying oscillations. 3.4.1 Krylov-Boguliuboy Method. ‘The Krylov-Boguliubov method is an example of a so-called "avei applicable to differential equations of the form /* method. Itis 1 F+y=ef OH. ‘where jis a "small" parameter. The class of equations of the form (1) include several com- ‘monly encountered ones, such as the Van der Pol equation and the pendulum equation. Note that, in (1), the angular velocity of the oscillations corresponding to = 0 has been normal- ized to 1. This presents no loss of generality, and can always be achieved by scaling the time variable t. If =0, the solution of (1)is ofthe form 2 y= sin(¢+9), here a and 6 are constants determined by the intial conditions. With this in mind, let us assume that the solution of (1) when 4 # Oisof the form, 3 y@=a@sin +O),80 Second-Order Systems h3 4 jW=aeos{r+9). where a(-) and @(-) are “slowly varying,” Le., (1) and 6(0) are "small." Actually, if y() is given by (3), then SS) =a(a)sin [1+ 9(0)] +4(0)c0s [1+ OKI +4(0) Hence, inorder for (4) tobe valid, we must have 6 asin(r+9)+adcos(e+9)=0, where the dependence of a and 6 on 1 has been suppressed in the interests of clarity. Substi- luting for yand j from (3) and (4) into(1) gives 7 acos(r+9)—adsin(r+9)=uf [a sin(t +4), a cos(¢+9)]. Equations (6) and (7) represent two linear equations in the two unknowns é and 6. Solving for these quantities gives 8 a=peos(r+9)f la sin(t +9), acos(r+9)], H sin(+4)f la sin(1 +9), acos(t+4)) To find solutions to (1) of the form (3) where a(-) i periodic, an extra condition is imposed, namely a) © 10 T or.equivalently, un yy ay dr=0, Th : ‘where Tis the period ofthe function a(-). Unfortunately, (11) cannot be applied directly, since the period T is in general dependent on jt and hence unknown. To get around this difficulty, we observe that a’) goes through one complete period as the phase 0=1+6(0) goes from Oto 2x. Thus the variable of integration in (11) san be changed from rt0@. Then the limits ofthe integration become O and 2x, and the integranda(t) becomes, in view of (8), 12 pcos Of (a sin®, a cos 8) Finally, we make the approximationSec.3.4 Analytical Approximation 81 48 BO Equation (13) expresses the fact that as rvaries over one period, @ varies over 2. Thus (11) becomes 4 14 LJ pcos8/(asin®, acose)d0=0. Similarly, if is also required tobe periodic with period T, this leads to the relationship 15 2g | fin (asin0,acos6y a0 Equations (14) and (15) can be used inthe following way: Suppose we are interested in approximating the periodic solutions of (1) by functions ofthe form 16 y()=asin{(1 +8), ‘where a and 8 are now unknown constants, Inthis case, (14) and (15) simplify to = 17 om07/ Ce sin0,ac000)400 18 J sin f(a sind, acos@)d0=0. Since a is a constant, the function f (a sin, a cos 8) is periodic in @ with period 2r and thence can be expanded in a Fourier series. Now (17) and (18) state that in order for (16) to approximate (10 the first order in 1) aperiodic solution of (1), itis necessary forthe first har- ‘monic ofthe periodic function f (a sin, a cos ) tobe zero. This requirement is sometimes called the "principle of harmonic balance.” We shal encounter the same reasoning in CChapter4 in connection with the so-called describing function method Note that the parameter 1 does not appear in (17) and (18), because when we study the ‘periodic solutions of (1), we are in effect examining the steady-state oscillations of (1), and does not affect the steady-state solutions. However, } is prominently present when the so-called “slowly varying” or transient solutions of (1) are studied. For this purpose, we smake the approximations a(D)-a(0) w a 782 ‘Second-Order Systems Ch.3 (T)-90) the period of the steady-state oscillations. However, as in studying the steady- state oscillations, we have a(1)-a(0) a T x 1 on Hcos0 f(a sin, a cos®) 40, nm XD-¥0) H sin@ F(asin8, acos®)20 Hence the approximate equations describing the slowly varying oscillations of (1) are 23 [ucososta sin®, a cos@)d8, al On 2g | sin0s(asin8,o costae 25 Example Let us apply the Krylov-Boguliubov method to the Van der Pol equation, which can be rewritten in the form Styewi(l-y), ‘Thisis of the form (1) with F, H=j(-y?). Hence Fa sin®, a cos @) =a cos @(1 a? sin? @) cos 36. [-st] me ‘The integrals in (23) and (24) are just the Fourier coefficients of this function, multiplied by some constants. Thus the approximate equations (23) and (24) governing the slowly vary- ing oscillations of (1) are given by 6 Bea = 2 4),Sec. 34 Analytical Approximation 83 7 To find a steady-state periodic solution of Van der Pol’s equation, we set 4 =0 and 6=0, which gives a=2. Hence, to first order in i the limit cycle of the Van der Pol oscillator is, described by y()=2sin (14). To getthe slowly varying solution, we solve (26) and (27) which results in 1 ram Oe aol where c is a constant determined by the initial conditions. Hence the slowly varying solu- tion of Van der Pol’s equation is , ja Trepcan | Smtttoe vo] ‘Thus we see that, even though the parameter 11 does not affect the steady-state solution, it does affect the rate at which the transient solution approaches the steady-state solution, 3.4.2 Power Series Method ‘The power series method is applicable to autonomous second-order differential equa- ions containing a "small" parameter 4 and consists of attempting to expand the solution of the given equation as a power series in 1. Mathematically the method is full of pitfalls, butit sometimes works reasonably well. The method is illustrated by an example, Consider the differential equation 28 F+y+ny"=0, together withthe special initial condition 29 yO)=a, j(0)=0. ‘This equation can represent, for example, the motion of a unit mass constrained by a non- linear spring. If 1 > 0, the spring is said tobe "hard," whereas if < 0, the spring is said to be sof." ), the solution of (28) satis Clearly if ing the initial condition (29) is, 30 yo(t)=acost. If #0 butis "small," then we can attempt to express the solution of (28) ~ (29) as a power the form seriesinp,4 ‘Second-Order Systems ch.3 31 y(Q=yolt) + Hy (+H yaD+ ‘The idea isto substitute (31) into (28) and equate the coefficients of all powers ofp to zero. However, if this is done blindly, some of the y(?) may contain secular terms, ic, functions Which are unbounded. To see this phenomenon, let us substitute (31) into (28) and set the coefficients ofall powers of u equal to zero. This gives 32 Fo+¥o=0:yo(0)=a, jo(0)=0. 3B Vity +¥9=0:y=0. 510)=0. Solving first for yo() gives 34 yolt)=acost. This is as expected, since yo(-) isthe solution of (28) corresponding to =. Now the equa- tion for y,() becomes, . aa? a 38 Fy ty, =-yB =-c08! =~ AS cos cos. The solution ofthis equation is 36 y= 22 tint F cons + 2 cos3 8 32 2 The 1 sins term on the right side isthe secular term, which arises because the forcing func- tion of the nonhomogeneous equation (35) for y; contains a component of angular fre quency 1, which is also the resonance frequency of the unforced system corresponding to (35). Combining the above expressions for yc and y gives an approximate solution to (28) which is good tothe first order in 37 y= yolt) +Hy (0) 2 pa202) e081 3B” ines BA cos (1 -nas32) cos “HE rsinr+ BE cos 3 Itisclear that the above approximation is unacceptable because itis an unbounded function oft. ‘The presence of the secular terms can be rationalized as follows: If =O, the solution of (28) is periodic with period 2x. However, if 10, the period of the resulting solution need not necessarily equal 27, though it will be close. Onthe other hand, since yo(”), the s0- called "generating solution” of the sequence of functions y('), y2(), °°" has period 27, so will all functions y(-). This attempt to express a function whose period is not 2x asa series using functions whose period is 2m leads to secular terms. As an example, suppose 5 is "small," and let us expand cos [(I + 8)1] as a power series in. This leads toSee. 3.4 Analytical Approximation 85 2,2 38 cos[(1+8)1) =cos1—Brsint - = cost ‘This power series converges uniformly in over any finite interval, and can therefore be con- sidered as a valid expression. However, if the series is truncated after a finite number of terms, the resulting finite summation contains secular terms. Moreover, the periodicity and boundedness properties of the function cos {(1 +) are not at all apparent from the above power series expansion. Toalleviate tis difficulty, suppose thatthe solution () of (28) ~ (29) is periodic with angular frequency ¢o, which i itself expressed as a power series in In other words, sup- pose 39 ota 1+yE\(a)+wE(a)+ This canbe rewritten as, 40 P —w5s(a)— HE 2(a)— ‘Note that in (39) and (40) the dependence of the frequency on the initial condition ais expli- cilly identified. This is a purely nonlinear phenomenon which has no analog in linear sys- ‘tems. Substituting (40) and (31) into (28) and displaying only the constant and the first order termsiny yields AL Jo +H +0%y0—HE yo HHOry, + HYG + Collecting terms gives a Yo(0)=a, 3o(0)=0, ¥B +E, Yo. ¥1(0)=0, 3 ,(0)=1 and soon, Solving these equations gives 44 yo(t)=acosax, 45 5,()+ay1(1)=—a" cos? wr +8,ac0s oF 3 1 2+ fa cos.ar = 7a? cos 3am + 8a cos a Now, in order that the solution for y (-) does not contain any secular term, itis necessary (and sufficient) that the coefficient of cos earon the righ side of (45) be equal to zero. Thus86 ‘Second-Order Systems ch.3 303 46 &=4 With this condition, the solution for y(-) is obtained as 2 ap 47 l= 525 cosas <5 cos 3a, where a 43 oteit dua Hence the overall solution of (28)-(29), accurate tofirst order in is given by 3 3 49 y(t)=acosan ~~" cos ae + yo, or Sgt MHS cos 3ax, 50 Example Consider the simple pendulum equation SL ¥+sin: Equation (51) can be approximated by of the form (28) within =—1/6. Using the foregoing analysis, we conclude thatthe frequency of oscillation of the simple pendulum is related to the intial amplitude by ‘This i a refinement of the analysis tused on the linearization of (51), which states that the frequency of oscillation is independent of the intial amplitude. That conclusion is indeed valid to firstorderina, as can beseen from (52) Problem 3.13 Apply the Krylov-Boguliubov method to Rayleigh’s equation sorea[s-] Solve the same equation using the perturbation method, and show that both methods give the same solutica to the first order in. Problem 3.14 Apply the perturbation method to the Van der Pol equationSec.3.4 Analytical Approximation 87 Sty spd -y?). Problem 3.15 Apply the Krylov-Boguliubov method tothe pendulum equation 3 oe Sty 7G 20. ‘Show that the expression derived for the frequency of oscillation is the same as (52). Problem 3.16 Consider the second-order equation J+Y= MAOH. ¥O=0, 50) =. ‘Assuming thatthe function fis continuously differentiable with respect to both of its argu ‘ments, show that both the Krylov-Boguliubov method and the perturbation method give the same results. Notes and References Most of the material in this section is historic, and much of it can be generalized to higher-order systems. Discussions of nontinear oscillations can be found in many classical texts, including Nemytskii and Stepanov (1960), The method of averaging, briefly intro- duced in Section 3.5, can be made rigorous; see Boguliuboff and Mitropolsky (1961) or ‘Sanders and Verhulst (1985).4. APPROXIMATE ANALYSIS METHODS In this chapter, we present two methods for approximately analyzing a given nonlinear sys- tem. Since a closed-form analytical solution of a nonlinear differential equation is usually impossible to obtain (except in some special examples, which are often contrived) itis wse- ful in practice to have some methods for carrying out an approximate analysis. Two ‘methods are presented here, The Describing Function Method consists of replacing a non- linear element within a system by an “equivalent” linear time-invariant system which isin some sense the best possible linear approximation of the given nonlinear system. This method is often used to predict the existence of periodic solutions in feedback systems Singular Perturbation Methods, just touched upon here, are well-suited for the analysis of, systems where the inclusion or exclusion of a particular component changes the order of the differential equation describing the system. It should be emphasized that these are just two of the many methods that are available for approximate analysis. Moreover, even with regard to these methods, the presentation here is merely an introduction, especially in the case of singular perturbations. 4.1 DESCRIBING FUNCTIONS Inthis section, the concept of describing functions is introduced, and itis demonstrated thatthey can be used to predict the existence of periodic solutions in feedback systems. 4.11 Optimal Quasi-Linearization ‘The problem studied in this subsection is thet of approximating a given nonlinear sys- temby @ linear time-invariant system. Let C [0, o) denote the set of continuous real-valued functions over (0, =), and suppose iV is a given operator mapping C0, =) into itself. In other words, given any continuous function x€C[0, =), the operator V associates with it another function Nx€ € (0, e). One can think of Nx as the output of 2 nonlinear system in response to the input <. By alight abuse of notation, the nontinear system isalso denoted by the symbol V. ‘The problem at hand isto approximate the given nonlinear system N by a linear time- invariant system H in an optimal fashion. More precisely, suppose a function re C0, =), called the reference input, is specified. If H is a linear time-invariant system with the impulse response h(),! then the output of H in response tothe input ris given by "he concept ofthe impulse response is formalized in Section 6.4 88See. 4.1 Describing Functions 89 1 ine: he-tredt A measure of how well the linear system H approximates the nonlinear system Nis provided by the error criterion Wy 2 2 EuH)= jim rf [Nr)(1) -HNOF at, assuming of course thatthe indicated limit exists.? Thus the objective is to choose the linear system Hin such a way that the error criterion £ (H) is minimized. A linear time-invariant system H that minimizes the criterion E (H) is called an optimal quasi-linearization of the nonlinear system N, and the problem of finding such an H is called the optimal quasi- linearization problem. ‘The solution to this problem is provided in Theorem (12) below. But first a couple of technical questions are laid to rest. A function x¢ C[0, «)issaid to have finite average power if 3 fing] Roda Pe m=t[feosnde 3 Iis straight-forward to verify that <-,->rp satisfies all the axioms of an inner product space 2 This issu is cleared up ater; sce Lemma (4.90. Approximate Analysis Methods ch (see Section 21.3). Thus, if we define r i . a 7 aft. z[Poe H rel Sipe glen Next, observe that a function fe C (0, «) also belongs to Fif andonly if 9 im lf lle <2. Now supposes, ¥¢F. Then 10x 4y hyp = Ux 3p My p+2 pp, from (4) and (7) Ux te + Hy pF, from(s) Letting 7 20 0 tim J « Det dr= tim >] (He (t+1)dt=6,<(), VERO, Finally, note that the cross-correlation function is bilinear; hence (20) becomes 2 0= 6,44 wD= 0,4) ~O,.w(Ds VEZO. Equi (21) isthe same as(13). i ‘Theorem (12) represents an important principle, namely: A is a best possible linear approximation tothe nonlinear system N if and only ifthe linear system faithfully repro- C[0, «) isa memoryless time-invariant nonlinear- ity of the form (Nxt) =n fx), Ve20, where n: R—> Ris continuous, and suppose the reference input ris anonzero constant, i.e., rik, Wt20. ‘Thus r() is adc. signal. It is easy tosee that both rand Nrhave finite average power. Now an easy calculation shows that nail) = kk), WO. Hence the optimality condition (13) is satisfied if nik) =m) 4 BOs HD = AE x00), 120, AO where 8() denotes the unit impulse distribution. Thus an optimal quasi-linearization of N 1 to the chosen reference input is a constant gain of n (kW, sometimes called the of N. ‘The above quasi-linearization is not unique. In fact one can show that if His a stable linear time-invariant system with the transfer function f(s), then H is an optimal quasi- linearization of Nif and only if h(0)=n (kk. (See Problem 4.1.) In the preceding discussion it is assumed that the reference input r is deterministic. However, itis possible to define the notion of an optimal quasi-linearization of a nonlinear ‘operator with respect to a random input. Also, the development can be extended with no essential changes to multi-input, multi-output systems. For a more detailed discussion, see Gelb and Vander Velde (1968). 4.1.2 Describing Functions By far the most commonly used reference input in optimal quasi-linearization is the sinusoidal function, 23° r()=asinor, Itis easily verified that reF, ie., rhas finite average power. Ifthe operator N is bounded- input, bounded-output (BIBO) stable in the sense defined in Chapter 6, the output ris the sum of two functions: (i) the steady-state response z,, which is periodic, and the (i) the transient response z, which decays to zero after some time. The same situation prevails ifSec. 4.1 Describing Functions 93 rris applied toa BIBO stable linear time-invariant system of the form (1). The determination ‘ofan optimal linear approximation (o Nis greatly facilitated by the next two results, 24 Lemma Suppose re, and fe C10, ) satisfies 28 [ Peden a Then 26 6,720. Proof Note that (25) implies that “t 27 him pI Poa=o By definition (ef.(11)}, r 1 2B 6,40) fin r(tyf (edad. By analogy with the inequality (8), we have 26] {femal |4prae 29 le, Ris continuous. Then (a, @) is independent of. Proof Since ris given by (23), itfollows that96 Approximate Analysis Methods cha 48 2(1)=(NND=n(asinar) is also periodic with period 21, ie., 20. Let z; be the first harmonic of z. Then n(a, ©) is given by (45). Now letthe reference input be 49 x()=asinax ‘Thus rand xhave the same amplitude, but different frequencies. Then 80 x()=r(ar/o), i.e.,xcan be obtained from r by time-scaling. Now, since N is memoryless, it follows from (47) that 51 WN)= WAKO), Hence the first harmonics of Nix and Nr are the same, allowing for the time scaling. There- foren(a, @)=1(a, 0), & 52 Lemma Suppose N is ofthe form (47), and in addition, n(:)is an odd function. Then m(a)is realforalla. Proof Observe first that one can write 1(a) instead of n(a, ©) since 1 is independent of w by Lemma (46). If n(-) is odd, then (1NP)(e) is an odd function off, and there are no cosine terms in the Fourier series expansion of nr. Hence gw in (45) is zero and 7a) is real. a 53° Lemma Suppose N is a memoryless time-invariant operator of the form (47), and suppose in addition that n() is odd. Finally, suppose there exist constants k, and k such that 54 kyo" Son(o) passing through the origin (see Figure 4.1), Note that k <2, but one or both of the con- stants could be negative, Proof By Lemma (52),n(a) is eal forall a. Moreover,See. 4.1 Describing Functions 7 ‘The proof that n(a) 3, the output ofthe nonlinearity isa “clipped” sine wave as shown in Fig- ure 4.4. In this cas, itcan be verified through laborious but routine calculations that Wenn Fig 44 oms—m) | n(a)= = {sin 8), a sf 4)" +m, iflal>d. This can be expressed more compactly. Define the function 1 forx21 89 f=) 9 alsin x (1 =27)'), for Osx A sketch of the graph of the function fis shown in Figure 4.5. Itis easy to verify thatSee. 4.1 Describing Functions 99 Fig. 45 m(a)=(m —m;)f @a)+ mz, ‘Though Figure 4.3 depicts the case where m > ma, the above expression is valid for any choice of m, and m2. By choosing various values for these two constants, one can obtain the describing functions of several common nonlinear characteristics. For example, ifm, =0, then the nonlinearity becomes the "dead zone" characteristic shown in Figure 4.6. Im, >Obutm; =0, then we get the "limiter" characteristic of Figure4.7. nyo) ‘Slope m, Fig.4.6 ns(o) Slope ms Fig.4.7100 Approximate Analysis Methods ch Fig 48 ‘Thus far all the examples have involved only memoryless nonlinearities. The next wo examples discuss nonlinearities with memory. 60 Example Consider the hysteresis nonlinear operator N shown in Figure 4.8. Inthe steady-state, the output (Nxt) follows the upper straight line when the input is increasing {ie..2(#)>0] and the tower straight line when the inpat is decreasing. The number ain Fig- ure 4.8 depends on the amplitude of the input and is not a characteristic of N itself. Thus mx(t) + dif (1) >0 oeoxo =) mx(t)—b it i() <0 and "jumps" when i(#) goes through zero. ‘Suppose a sinusoidal input rof the form (23) is applied oN. The resulting steady-state output is shown in Figure 4.9. One can express Nr(:) as the sum of to signals as shown in Figure 4.10. From this tis clear that the firstharmonic of the steady-state part of Nris 4b 2) (0) =ma sin or+ cos on. Hence 4b na, ayem+j2. ‘Note that 7 is once again independent of «, because time scaling does not affect the output of N. Hence, even though Nis rot memoryless, the arguments in the proof of Lemma (46) apply. 61 Example Consider the hysteresis nonlinear operator N shown in Figure 4.11. In this ‘case if the input amplitude is ess than 8, the outputis simply equal tozero.See. 4.1 Describing Functions 101 worn Fig.4.9 con ee Jae Fig. 4.10 nia)=0ifa sd. Ifthe input amplitude a exceeds 6, then, in she steady-state, the output signal zis as shown inFigure 4.12. Let Be (0, 2/2)>be the unique number such that 28 sinB= Forconvenience let ® denote a. Then the steady-state outputis described by102 Approximate Analysis Methods ch Slope m {ma sin-m8, for0<@0as 0 >=. Thus o(@) > 0as@->=. andas aconsequence all suficiently‘arge belong to. Note also thatthe point (ja) lies inthe disk B(q) forall we Q’, B(@) contains the Nyquist plot of (je), and can be thought of as a "broadening" of the,122 Approximate Analysis Methods ch4 Imation Regtiu), -1/ata) Fig 421 Nyquist pot to reflect te high frequency behavior off i. the effets of neglecting the higher hamonics of Gu 78 Example Consider the feedback systemof Figure 4.13, where 100 80 Ce + SNe 1)” and Nis the odd piecewise-linear characteristic shown in Figure 4.22. This nonlinearity is of the form studied in Example (4.1.58) with sm =15, m: Hence its describing functionis given by n(a)=10+5/ (Va), where /()is the function defined in (4.1.59) Let us first use the heuristic arguments of Section 4.1.3. The Nyquist plot of 80) is shown in Figure 4.23. The plot intersects the negative real axis when @y = 80 = 8.9443 radisee, and Hie) =~ 0.0794, Thus nao which corresponds toSec. 4.2 Periodic Solutions 123 ‘Slope 10 Fig. 4.22 Imai Rego 010794, = 8.9443 Fig.4.23 4g =2.3742. Hence we ‘predict’ a periodic solution with an angular frequency of 8.9443 rad/sec and an amptivude of 2.3742, Now letus apply Corollary (72). Itis easy to see that N belongs to the incremental sec~ tor{10, 15}, sone canchoose c=125,r1 5. First let us determine the sets © and Mp of (63) and (64) respectively. The shaded region shown in Figure 4.23 is called the critical disk, and passes through the points124 Approximate Analysis Methods ch.4 0.0667, OL ‘The critical disk is denoted by C. It is an easy exercise in algebra to show that, if is any complex number, then ler iffzee. In any case, aderivation of the above relationship is found in the proof of Theorem (6.6.40), Now one can see from Figure 4.23 that the Nyquist plot enters the critical disk when @=8.14 and leaves itwhen @=9.61. Thus y= (@: kof 8.14, 9.61] VE2 1), = (a: koe{8.14, 9.61] Vk22), Hence, for simplicity, one can take Soby (i) of Corollary (72), we can state our first precise conclusion: The system has no periodic solution with a frequency > 9.61 radi/sec. Next, let us determine the frequencies ©, and @,. ‘These are the frequencies at which the disk B(@) is tangent to the plot of -1/y(a), which in this case is just the interval (0.1, -0.6671. Now itis easy to see that B () is tangent tothe real axis fand only if Info po, where 6(«) and p(«) are given by (74) and (75) respectively. This allows one to determine that ©,=8873, 0, =9.011. If@<@, or o> ,, then B(«) does not intersect the real axis. Based on (i) of Corollary (72), we can now state ou second precise conclusion: The system has no periodic solution with frequency in the intervals (4.805, 8.873) and (9.011, 9.61]. Finally, the disk B («is tangent torealaxisat ReG(o,) =-0.0808. SoSee. 4.2 Periodic Solutions 125 na) Similarly, L a)=~ Reto Soby (ii) of Corollary (72) we have our third precise conclusion: There is a periodic solution with angular frequency in the interval (8.873, 9.011} and Jirst harmonic amplitude in the interval (2.1821, 2.6028]. 79 Example Consider again the feedback system of Figure 4.13, with soye 000) 4" GDS and 1 is the limiter nonlinearity showa in Figure 4.24. From Example (4.1.58), it follows that the describing function of Nis (a) =2f (10/a), where fis defined in (4.1.59). ne) Fig. 4.24 Since the arguments here closely parallel those in Example (78), most of the details are ‘omitted, and only the final conclusions are given. “The Nyquist plot of §(/@) is shown in Figure 4.25, Itcrosses the negative real axis at 71283, corresponding tou =VI7 =4,1231 radisec. Now =14, orag= 17.0911 ‘Thus the classical describing function analysis predicts a periodic solution of amplitude 17.0911 and an angular frequency of 4.1231 rad/sec.126 Approximate Analysis Methods ch Imation Regia) (ot to scale) Fig. 4.25 Now let us apply Corollary (72). The nonlinear element N belongs tc the incremental sector (0, 2},so we can take 1 Now, ifzisany complex number, then ‘ z 1 12 |< 1iffRez>-05, [Tee | From the Nyquist plot, one can see that Re 8(i@) > -0.5 Vao> 4.9026. Hence we can choose 2 =4.9026, =), =(2.45 Next, the tangency conditions are satisfied when oy=341,.0,2442 Now (@) 1.0260 — j0.5296, C(0,) 0.6366 + j0.0783, n(a)=- ZA = 1.5707, oFa)= 14.894, Reto)Sec.4.3 ‘Singular Perturbations 127 1 ReCw,) nia, 9747, a,= 25.439. ‘So we can draw the following conclusions, based on Corollary (72) (There isno periodic solution witha frequency > 4.9026, There isno periodic solution witha frequency in (2.4513, 3.41) or (4.42, 4.9026]. There is a periodic solution with a frequency wel3.A1, 442) and a frst harmonic amplitude ae (14.894, 25.439}, Compared to Example (78), we see that the "spread” in both « and a is considerably larger in the present instance. This is because the describing function of the nonlinearity in the present example varies over a much larger range than in Example (78). That nonlinear- ity is much “closer” to being linear. One would naturally expect that there is much less uncertainty inthe results obtained using quasi-linearization methods when the nonlinearity isclose to being linear. This is reffected inthe results of Examples (78) and (79). Problem 4.9 Consider again the transfer function @(s) of Example (78), and suppose the nonlinear element Vis the form shown in Figure 4.3 (oF 4.22), where the initial slope m, ‘or small values ofthe input equals 13.the final slope m> equals 12, and the width 3 equals 1 Using Corollary (72), find upper and lower bounds on the amplitude and frequency of the periodic solution. Show tha, inthis cae, the “spreads” between the upper and lower bounds for both the amplitude and the frequency of the periodic solution are less than they are in Example (78). How do you explain this? Problem 4.10 Consider again the transfer function §(s) of Example (79), and suppose the nonlinear element Wis a dead-zone limiter of the form shown in Figure 4.7, with slope ‘my =20 and adead-zone width 8 =0,01. Using Corollary (72), find upper and lower bounds ‘on the amplitude and frequency ofthe periodic solution, if any. Compare with the results of Example (79) 4.3 SINGULAR PERTURBATIONS In this section, a brief introduction is given to the method of singular perturbations. ‘This method is valuable in analyzing systems whose dynamic order changes asa result of neglecting some elements, or making some simplifying assumptions [see Example (40) below]. Consider a system of nonlinear differential equations 1 x=fixy), eF=B%y). where x6 R*, YER", RXR" — R", and g:R"R" > R", Note that for any value of € ‘other than ero, the system (1) consists of n +m differential equations. However, ife=0, the128 Approximate Analysis Methods ch4 system (1) becomes a set of n differential equations, and malgebraic equations, namely 2 k=fxy), x.y) Suppose itis possible to solve the m algebraic equations above to obtain an explicit expres- sion for yin terms of x,of the form 3S y=hoo, where h: R" > R™. Then one can substitute (3) into the irst equation of (2) to obtain the set of differential equations 4 x=ffx, ho] Setting €=0 in (1) is called a singular perturbation since it changes the order of the system. This is to be contrasted with a so-called "regular" perturbation, described next, ‘Suppose we are given a systemof n differential equations 5 Mt fIx(0, pl. where x(1)e R", pe R" and f: RR" —> R" is continuous. One can think of p as a vector of physical parameters appearing in the system description. If the vector p is perturbed slightly, then the orderof the system isnot affected. In contrast, suppose (1) isrewritten as « ffm] \ 5) | ates | Then obviously the right side of (6) is nor continuous with respect to € at é=0. This is why setting ¢=0isa singular” perturbation. ‘The system (|) is called the full-order, unsimplified, or original system, while (4) is called the reduced-order or simplified system. In road terms, the basic objective of singu- Jar perturbation theory isto draw conclusions about the behavior ofthe original system (1) bbased upon a study of the simplified system (4). Atte moment we do not have the tals to study the stability of nonlinear systems of the form (1) or (4); these are presented in Chapter 5. So in the present section the scope of the study is limited to linear singularly perturbed systems, ofthe form GHC Ay Aoa| Ly}? where Xe’, YER, andthe matrices A have compatible dimensions. If i set equal toSec.4.3 Singular Perturbations 19 ‘zero in (7), the second equation becomes ‘Theorem (12) below presents the main result of this section. To make the theorem statement concise, aitle notation isintroduced first. Define MW Ag= An ~ApAAn, Let A= (24,""*,2q} denote the spectrum ofthe matrix Ap, ie, the set of eigenvalues of ‘Ap, where repeated eigenvalues are listed as many times as theirmultiplicity. Similarly, let TE [yi.°** Yq) demote the spectrum of Az> 12 Theorem Consider the sysiem (7). Suppose Azz is nonsingular and define Ao as in (11). Then given any 5 > 0, there exists an €q > 0 such that, whenever 0< \€1 <£q, then +m eigenvalues {0t),***, Oy 4m} ofthe matrix Au An ] Ault Any BoA satisfy the bounds 14 1h,-ay1 <8 fori=1, 1S lyq~€0,1 <8, fori=n4l,--\n+m. Remarks Clearly the eigenvalues of the matrix A, are the natural modes of the unsimplified system: (7). The inequalities (4) and (15) imply that, a5 € +0, exactly ofthe eigenvalues of A, converge to the eigenvalues of Ag, while the remaining m eigenvalues ‘approach” infinity, asymptotically like ¥/e. Moreover, if, isan times repeated eigen- value of Ap, then exactly ; eigenvalues of Ay converge to 2; similarly for y. [fall eigen- values of A, have negative real parts, then the solution of (7) approaches 0 as» efor each initial condition; inthis case, the system (7) is aid tobe asymptotically stable, On the other ‘hand, if some eigenvalue of A has a positive real part, then the norm of the solution of (7) Approaches infinity as 1 ~> for almost all initial conditions; inthis case the system (7) is said o be unstable. See Chapter 5 for precise definitions ofthese concepts.130 Approximate Analysis Methods cht Proof To compute the eigenvalues of A, let us carry outa similarity transformation of ‘Ac such thatthe resulting matrix is in a block-triangular form. More precisely, let us find a matrix Mee R’ such that 7 -M) {An An 18 Vee 1 | [Ave Arve Fe Oem GH, [.."]- Expanding the triple matrix product on the left side of (16) shows that, in order for the 1,2 block ofthe product to equal zero, the matrix M, must satisfy the following equation: 17 AuM.+An-M.—M.-M.—— #0. ‘Ase—+0,some matrices in (17) approach infinity. To get around this difficulty, suppose M, hhas the form, 18 M,=eP, Substituting for M, in 17) and clearing fractions gives the following equation for Py: 19 CA Pe + Aya ~€PeAnPe~PeAr2 =0. ‘This equation is quite well-behaved as ->0. In fact, substituting €=0in (19) gives 20 An—Pods Which has the unique solution 2 Py=AAGL. .¢0, then (19) is a quadratic equation in P,. Nevertheless, since the coefficients in (19) are continuous in, one can conclude tha for sufficiently small shere exists solution, of (19) which is close to Pp: but this need not be the unique solution of (19). Choose such a solution P,,and define M, asin (18). Now expanding (16) gives Fe nun An ~MAne Om 2 lem, [=| Aye AuMAt+Anee|” Now we know that 23° PL=AAzl +O(@). Hence, from (22),Sec. 4.3 ‘Singular Perturbations 131 24 Fee Ay -PeAn= 11 A zABAg, +O(€)= Aq +O), Aj2A3} +(e). Since the matrix in (22) is block-triangular, it follows that the spectrum of A. is the union of, the spectrum of F, and the spectrum of He. As €-+0, itis clear from (24) that Fe > Ag: hhence the spectrum of F, approaches that of Ap. ic. (14) is satisfied. Finally, as € 0, the ‘Any/e term on the right side of (25) swamps the other terms (recall that Az. is nonsingular), and (15) is satisfied. To state the next result concisely, a couple of terms are first introduced. A square matrix is sai to be hyperbolic if it has no eigenvalues with zero real part and itis said to be Hurwitz if all ofits eigenvalues have negative teal pats. 26 Corollary Consider the matrix A of 12), and suppose Az. is hyperbolic. Then the {following two statements are equivalent: 1. There exists any >Osuch that Ay is Hurwitz whenever0<€<€o 2. The matrices Ao and Azs are both Hurwitz. In the parlance of singular perturbation theory, the matrix A22 is said to epresent the fast dynamics, while Ao is said to represent the slow dynamics. Without geting too deeply into the technicalities ofthe subject (and there are plenty of them), the reasoning behind this terminology can be briefly explained as follows: Suppose both Ao and A> are Hurwitz. For each € sufficiently small and positive, define M, as the solution of (17) such that eM, =P, A,,Az} as 0. Now define a new state variable vectorby a] [1-M, y|7 07 | [] renee ‘Now (16) makes it clear that the dynamics of the new state vector are governed by [Fe 0] [ze [eam] [> i 28 y or, in expanded form, 29 eRe, J=Get, +Hey. Now define the "fast" time variable132 Approximate Analysis Methods ch. 30 rath, and define the functions¥, %, by 3 (e), RD =2(ED. ‘Thus ¥(-) is just te function y(-) with the time scaled by the factor ¢; similarly for Z. Itis easy tosee that AM) dye With this change of independent variable, the system equations (29) can be rewritten as, dat) ad Fez), “4 so. = 0G, Ze(7) +EHY(0. ‘These equations enable us to understand mote clearly che time behavior of the func- tions 2, and y. First, (33) shows that the time response of, is independent of the initial con- dition y(0) and depends only on the initial condition 2,(0). Since M, —>0 as €-»0, we see that z (0) becomes closer and closer to x(0) as € + O(ef.(27)}. Second, (34) shows that Z,() acts like a forcing function to ¥(). If ze(0) = 0, thea Z, =0, and (34) reduces to 35 FO e310) dt Now nt from 25) that e+ Axa £-¥0, Hence, as €-¥0. 36) looks approximately ike 3 F)=exp(And HO). 1f2,(0)20, then from (33), 37 a(t) exp (Ft) 24 (0)=exp (Apt) %(0), since F, -» Ay as€ 0, Now, ifeis very small then forthe purposes of analyzing (34) one can treat %(2) as @ constant vector 2,(0), and replace eG, by Azz [ef. (20)}. Then the approximate solution of (34) is 38 F(t)=exp (Ant) [¥(0) +AZA2:2(0)], orSee. 4.3 Singular Perturbations 133 39 KE)=exp(AncA) ly(0)+ABAn tO) Equation (39) shows why y is often referred to as the fast state variable, and also why he fast dynamics are determined by the matrix Az2. Now, as €-+0, the matrix M, —>8, and the ‘vector 2 x lef. (27)}. Thus the vector x is referred to as the slow state variable, and its time evolution is governed by the matrix Ao as demonstrated by (37). 40 Example In practice, singularly perturbed differential equations arise when some dynamical elements (often called "parasitics") are neglected during the modelling proces. ‘This s illustrated in this example. Consider the circuit shown in Figure 4.26, and suppose the operating point ofthe tunnel diode is so selected that its small-signal resistance is negative. Ifthe stray capacitance of the diode is included in the network, then one obtains the linearized model shown in Figure 4.27. Following the common practice in network theory, let us choose the capacitor vol- tages and the inductor current as the state variables. Let us suppose also that the diode capa- citance ¢is very small. Then the dynamics of the network are described by the third-order equation ae Ry L Fig.4.26 Ry fe GH Fig. 4.27134 Approximate Analysis Methods ch R. | a) PE oa el] afe|-b-pb ob |fn Cy RC) RC, . et 1 Re-k [LJ ae Settinge =O leads to Re Rk? and othe simplified model The coefficient matrix above is Ag. In this instance the matrix Az? isjusta scalar and equals RyRy re ee RaRy If Ry > Ry, then Az, >0, which means that Ap, is not Hurwitz, On the other hand, one can easily verify that Ap is Hurwitz if Ry > Ry. Using Corollary (26), we conclude that if, 0 R" is continuous. Itis further assumed that the equation (has a unique solution corresponding to each initial condition. Thisis the case, for exam- Ple, iff satisfies a global Lipschitz condition (see Theorem (2.4.25)]. Its shown in Appen- dix A that, roughly speaking, the preceding assumption is true for almost all continuous, functions f. Let s(, fo, X9) denote the solution of (1) corresponding to the initial condition %q,evaluated at time. Inother words, s satisfies the equation Ast ty X0)= Ml 4 fo, XI VERSE f. RG ‘The solution maps maps R, x R" into R", and satisfies he following properties: 3 Slo. 0,%) =Xos W%0ER", 4 she th. 8, fo. X0)]=8lt, foro) VE24, 2920, WayER" Recall thata vector xye R" isan equilibrium ofthe system (1) if St, x)=0,¥120. ‘Clearly, if (5) istrue, then 6 se fo, %9)=%, WE2920. nother words, ifthe system starts at an equilibrium, it stays there. The converse isalso true, sis easily shown. Throughout this chapter itis assumed that 0 is an equilibrium of the sys- tem (1). Ifthe equilibrium under study is not the origin, one can always redefine the coordi- nates on R" in such a way that the equilibrium of interest becomes the new origin. Thus, without loss of generality. itisassumed that 7 £(i,0)=0, 120. Thisis equivalent o the statement Bs, 10,0)=0, ¥F2t0, Lyapunov theory is concemed with the behavior ofthe function s(,fo.%) when X #0 but is “close” to it. Occasionally, however, the case where xy is "far" from 0 is also of intrest 9 Definition The equilibrium 0 is stable if, for each € > 0 and each tye R,., there exists aB=8(e, tp) such thatSee. 5.1 ‘Stability Definitions 137 10 thx ll <5(e, 10) liste 9, X9) Il 0, there exists a: €) such that 1 txp lh <8(€), 920% list, fo, Xo) <€, WEA. The equilibrium is unstable ifit isnot stable. According to Definition (9), the equilibrium 0is stable if, given that we do not want the norm Ilx(t)Il ofthe solution of (1) to exceed a prespecified positive number, we are able to determine an a priori bound (tq, ) on the norm of the initial condition tx(¢q)ll in such a ‘way that any solution trajectory of (1) starting at time fo from an initial state inside the ball of radius 8(¢o, €) always stays inside the ball of radius € at all future times ¢>¢9. In other ‘words: arbitrarily small perturbations of the initial state x(zq) about the initial state O result in arbitrarily small perturbations in the corresponding solution trajectories of (1). Itisalso possible to interpret stability asa form of continuity ofthe solution trajectories with respect to the initial conditions. We have seen {Theorem (2.4.57)] that, undes eason- able hypotheses such as Lipschitz continuity off, the solution of (1) is@ continuous function of the intial condition. This means that, given any fo 20 and any finite T, the map s( fo. Xa) ‘which takes the intial condition x into the corresponding solution trajectory in C"[tp, Tis continuous. This property is true whether or not the equilibrium 0 is stable. However, sta- bility requires something more. To state what itis, let C*Ito. =) denote the linear space of ‘continuous n-vector valued functions on [fo,), and let BC"[to, =) denote the subset of C"lto. =) consisting of bounded continuous functions. If we define the norm, 12 UxCI,= sup. lx, relinm) then BC"[19, -)isa Banach space, Now stability isequivalent to the following statements: 1) Foreach ¢920, there is a number d (tg) such that s(-, 19. ¥p)€ BC" tp, =) whenever xp Bay)» Where By is the ball 13 B= (xe R": IIx O and each 1920, a corresponding 6 can be found such that (10) holds. In general, this 5 depends on both ¢ and fo. However, if a 8 can be found that depends only on € and not on fo, then the equilibrium 0 is uniformly stable. If the system (1) is autonomous (f does not depend expli- cilly on ), then thereis no distinction between stability and uniform stability, since changing the initial time merely translates the resulting solution trajectories in time by a like amount. In terms of the map s, uniform stability is roughly equivalent to uniform continuity with respect 10 19, More precisely, uniform stability is equivalent to the following two state- ‘ments: 1") There is a number d>0 such that s(-,19,%9)€BC"[to,%) whenever xy¢ By, toe. 2°) The map S(, to, %9): By BC"[fo, =) is uniformly continuous in xp at © with respect toto. 14 Example Consider the motion of a simple pendulum. If Js the length of the pendu- Jum, @ is the angle of the pendulum measured from a vertical line, and g is the acceleration, ddue to gravity, then the motion of the pendulum is governed by 1S 6 +(g/)sino=0, By introducing the standard state variables x, =®, x = 8, (15) becomes £, 16 iy=m.in=-4 ‘As shown in Example (3.4.49), the trajectories ofthis system are described by 48 xo 17 FL cosey =P -£ cosy 1 ‘One can now show that 0 is a stable equilibrium by verifying the condition of Definition (9) directly. Suppose e > 0 is givem: then its possible to choose a numer a > 0 such that the curve described by (17) lies entirely within the ball B,. Now choose a 8>0 such that the ball Bg lies entirely within this curve (see Figure 5.1 for the construction). ‘Then (10) is satisfied. Since this procedure can be carried out for any €> 0, 0 isa stable equilibrium, ‘As mentioned above, there is no distinction between stability and uniform stability for autonomous systems, The next example illustrates that for nonautonomous systems the two concepts are indeed distinct, 18 Example (Massera 1949) Consider the scalar differential equationSee. 5.1 Stability Definitions 139 Fig.5.1 wD i@ 64 sint ~21) x(t). ‘The sofution (19) is given by 20 x(t) =x (to) exp (6sins 6 cos 1? ~6sintg +619 05%0 +88) To show thatthe origin isa stable equilibrium let 20 be any fixed initial time. Then a 6sinty +619 costo +13). Now, if tt >6, then the function on the right side of (21) is bounded above by exp[12+T(6~T)], where T=r—ro. Since this function is continuous in 1, itis bounded over [fo, T]as well. Hence if we define 22 e(tg)=superp (6sin!~64cos1~7?~6sinty +619 60819 #13) then c (19) is finite for each fixed t, Thus, given any €>0, the condition (10) is satisfied if we choose 5=e/c (to). This shows that 0 is a stable equilibrium. On the other hand, if f= 2nr, then it follows from (20) that 23_xf(2n4 1m] =x(2nn) exp ((4n+ 1N6—n)R. This shows that 24 c(2nn)2expl(4n + IN6—2)r]. Hence ¢ (to) is unbounded as a function of ro. Thus, given €>0, itis not possible to find a single 6€), independent of fo, such that (11) holds. Therefore the equilibrium 0 i not uni- formly stable.140 Lyapunov Stability chs ‘There is nothing particularly special about the preceding example, Problem 5.1 shows how one may go about constructing a class of systems for which 0 is a stable but not uni- formly stable equilibrium, Finally, letus tum toa discussion of instability. According to Definition (9), instability is merely the absence of stability. Its unfortunate that the term “instability” leads some to visualize a situation where some trajectory of the system "blows up" in the sense that Ix(0)Il east —>ee. While thisis one way in which instability can occur, itis by no means, the only way. Stability of the equilibrium 0 means that, given any e>0, one can find a corresponding 5>0 such that (10) holds. Therefore, 0 is an unstable equilibrium if, for some € > 0, no & > 0.can be found such that (10) holds: equivalently, there is a ball B. such that for every 8 > 0, no matter how small, there is a nonzero initial statex(t9) in Bg such that the corresponding trajectory eventually leaves Be. This. and only this, is the definition of instability. It may happen that some trajectories stating in Bg actually "blow up,” but this is, notnecessary for instability. Thisis distinction illustrated next. 25 Example Considerthe Van der Pol oscillator, described by 26 iy=xp i222) 4—aP)ED. ‘The origin is an equilibrium of this system. However, solution trajectories starting from every nonzero initial state, no matter how close to the origin, will eventually approach the imit cycle as shown in Figure 5.2. Now let us study the stability of the equilibrium 0 using Definition (9). By choosing € > 0 sufficiently small, we can ensure thatthe ball B, is con- tained entirely within the limit cycle (see Figure 5.2). Therefore all trajectories starting from a nonzero initial state within B, will eventually leave B , and so no 3 > Ocan be found such that (10) is satisfied. Accordingly, the origin is an unstable equilibrium. Note that all trajectories ofthe system are bounded, and none blows up. Sothe systems well-behaved in this sense 27 Definition The equilibrium 0 is attractive if for each tg R,., there is an M(t) >0 such that BB lag ll Ml) > 8Ulo +4, fo, %q) 2Oast oe, The equilibrium 0 is uniformly attractive ifthere isa numbern > Osuch that 29 Uxyll <1, 1920 Slt +4, fo, Xo) 9 0as1 ee, uniformly in xp, fo. ‘Thus attractivity simply means that, at each initial time ro¢R, every solution trajec- tory starting sufficiently close to @ actually approaches 0 as ¢9 +1 —> 2, Note that there is no requirement of uniformity at all in two ways: First, the size of the "ball of attraction” 1(¢) ‘can depend on fp. Second, even fora fixed £9, the solution trajectories starting inside the ball Bry.) but at different initial states can approach 0 at different rates. In contrast, uniform attractivity requires first that there be a ball of attraction By, whose size is independent of to, and second that the solution trajectories starting inside B,, all approach 0 at a uniform rate.See. 5.1 Stability Definitions 141 Fig 5.2 Note that (29) is equivalent to the following statement: For each e > O there exists a such that Te) 30 xp ll . However this does not mean thatthe origin s uniformly tractive, since the closer the intial state is to 0, the more slowly the restlting trajectory converges to 0. Now the origin is an tnstabl equilibrium. This canbe established, asin Example (25), by choosing e o small that S does not lie inside B. This shows that itis possible for an equilibrium to be attractive yetunsable Fig.5.3 Itcan be shown {see Theorem (49)] that ifthe origin of an autonomous system is stable and attractive, then itis also uniformly attractive. To the best ofthe author's knowledge, it has not been settled whether itis possible for an equilibrium to be uniformly attractive yet unstable, 36 Definition The equilibrium 0 is exponentially stable if there exist constants 1.4, b>Osuch that 37 Ms(to #2, to, Xo) ll Sa lhl Exp (bt), Wh, tp 20, V9 B,. Clearly exponential stability isa stronger property than uniform asymptotic stability. Allof the concepts of stability introduced thus far are local in nature, in the sense that they pertain only «o the behavior of solution trajectories starting from initial states near the ‘equilibrium. The final definition pertains, in contrast, to the global behavior of solution tra- jectories.See. 5.1 ‘Stability Definitions 143 38 Definition The equilibrium 0 is globally uniformly asymptotically stable (g.2.s.) if (i) it is uniformly stable, and (ii) for each pair of positive numbers M, € with M arbitrarily large and arbitrarily small, there exists finite number T = T(M,€)such that 39 ixgll 0 such that 40 s(t +1, Fo. 9) Sallxy fl exp(-br), Wt, f9 20, Wage RY Note that in order for an equilibrium to be either globally uniformly asymptotically stable or globally exponentially stable, a necessary condition is that it be the only equili- brium, Next we discuss the special case of autonomous and periodic systems. The system (1) is periodic with period Tif, AL fr +T, x)=f, x), ¥r20, VER", For a periodic system its clear that AZ S(C#T, to +T, Xo) =Mt, to, Xo), E2120, WHER" If the system (1) is autonomous, ie., if fdoes not depend explicitly on , then we can think of it as a periodic system with an arbitrary period. Hence all the results presented below for periodic systems apply equally well to autonomous systems. 43 Theorem Suppose the system (1) is periodic. Then the equilibrium 0 is uniformly stable ifand only ifitis stable. Remarks Ithe sysem (1) is autonomous, the theorem is obvious, since (42) holds for every T > 0. Butthe proof for periodic systems requires abit of work. -s stability, and only the reverse implication the period, and define Proof Clearly uniform stability imp needs to be proved. Suppose £9¢ [0, T]] where T 44 Wl, to) =sup Uso +4 foo)! Since 0 is a stable equilibrium, there is a number d (tp) such that (Xo. fo) is finite for all Xe Bye, [ef. the remark preceding (13)]. In general, it may happen that as to varies over (0, e), the number d(¢9), though nonzero for each fixed fo, cannot be bounded away front 0. However, due to the periodicity ofthe system, one can safely assume that d (fo +T)=4 (to). and it is only necessary to consider the number d(tq) as fo varies over (0, T]. Thus we can finda number d > 0, independent of f9,such that144 Lyapunov Stability chs 45 U(X, fo) <=, VE By. Veo (0, 71 Since u(a. fo) is clearly a continuous function of fo, the function 46 10%) Sif Hoo is finite forall xy¢ By, and is continuous atx) To show that 0 isa uniformly stable equilibrium, suppose e> 0 is given; we must find 5 >Dsuch that (11) holds. Due o periodicity, itis only necessary to show that the same 3 ‘works forall t9¢ [0, 7 instead f for all rp 20. Since 1¢) is continuous, we can find a >O such that 47 Iixyll nlmq) llslto +1, fo, Xo)ll Se, VE20. ‘This completes the proof. 49 Theorem Suppose the system (1) is periodic. Then the equilibrium 0 is uniformly asymptotically stable ifand only ifit isasymprotically stable. The proof of this theorem is quite involved, even for autonomous systems, in contrast to Theorem (43), which'is quite obvious for autonomous systems. The source of difficulty can be explained as follows: Suppose 0 is asymptotically stable, and that the system is auto- ‘nomous. Then, by Definitions (31) and (27), there isa number r =r(0) such that 30 lixpll 566, 0, %9) 90ast oo. By autonomy, it follows that SL Ux ll <7 1920 S(t9 +4, fo, Xo) 90 as 12>, uniformly into, But uniform attractivity requires something more, namely thatthe convergence in (51) be uniform with respect to x) as well. Proving this properly requires some work. As can be imagined, the periodic case is still more complex. For this reason, the proof is omitted, and the reader is referred to Hahn (1967); the relevant theorems are 38.3 and 38.5 Finally let us recast the various stability definitions in terms of so-called functions of class K and class L, so as to simplify considerably the proofs of subsequent stability theorems $2 Definition A function $:R, R. is of class K ifit is continuous, strictly increasing, ‘and (0) =0; itis of eass L ifitis continuous on (0, =), strictly decreasing, @(0) <=, and 7) asa,See, 5.1 Stability Definitions 145 53 Theorem The equilibrium ofthe system (1) is stable ifand only if, for each 1g Ry, there existanumberd (tg) > Oand afunction , of class K such that 54 ist t0,%9) 1S, (IM), VEZ. VRE Bay The equilibrium is uniformly stable ifand only ifthere exist a number d > Oand a function 9 ofclass K such hat 5 s(t, to. Xo) SO( III), VIZLo 20, Wape By. Remarks We can thus use ($4) and (55) as the definitions of stability and uniform sta- bility, if we wish Proof First, if (54) holds, then 0 is a stable equilibrium. To see this, given any €>0 and any fo€ Ry ,choose 56 8(€,10)=min{d (19), 0,08). To prove the converse, fix 7p, and suppose e > 0 is given. Then by definition there exists a number 8>0 such that (10) holds. Define 8,(€, fo) to be the supremum of all possible choices of & such that (10) holds. The function V,:€+8,(€, fo) is nondecreasing, satisfies ,,(0)=0, and y,,(€) > 0'Ve > 0; of course, it need not be continuous nor strictly increasing, However itis possible to finda function @, of class K such that 8, (€) 0, and foreach xy By) afunctionO,,, ofciassLsuch that 5B Ms(C0 +t £0. X9) Spay (, WEZ0, VRE Br) The equilibrium 0 is uniformly attractive ifand only ifthere exist a number r > O.and afunc- tion Gof class L such that $9 S(t +1, fo. Xo)ISOU), Vi f920, VQe B,. Proof If (58) holds then 0s attractive, since 6,, .,(f) -+0as <2. Conversely, sup- pose is attractive. For each € > 0, define T(e) tobe the smallest number 7 with the property that 60 US(to+4 fo. %o)ll 0, a function @ of class K. and a function 6 of class L, such that 62 IIslto +1, £0, %9)!1 SOCK 11) 60), Vb, fo 20, VXOEB,. ‘The proof can be found in Hahn (1967), Chap. V. But notice the similarity between (62) and Equation (37) defining exponential stability. Problem 5.1 The purpose of this problem is to generalize Example (18) by developing an entire class of linear systems with an equilibrium at = 0 which i stable but not uniformly stable, Consider the linear scalar differential equation Ht) =a(x(0, WE20, where a(’) is continuous function. (a) Verify thatthe general solution of this equation is, xQ=xUtgverp | fa(trdr (b) Show, using Definition (9), thatthe equilibrium Ois stable and only if supero| faceyde] -im(tgy0Wr>0. Then there exists a function oof class K such that o(r)<0(r) Vr. Moreover, if$(r) as r ~¥e, then o.can be chosen to have the same property. Proof Pick a strictly increasing sequence (4)} of positive numbers approaching infinity, and a strictly increasing sequence (k; of Positive numbers approaching 1. Define kon), if0srsq), Fi 0 " 2 alr)= kig(q)+ [ki 9) KO qi 0 and function kof class K such that 4 aflixiS V(t, x), 120, VxeB,. Vis decrescent ifthere exista constant r > Oand a function of class K such that148 Lyapunov Stability h.5 SVG, x)$BCIIxII, ¥420, VxE B,. Vis a positive definite function (pal if(4) holds forall xe R® (ie. ifr ==). Vis radially unbounded if (4) is satisfied for all xe R" and for some continuous function (not neces- sarily of class K) with the additional property that a(r) © as r 2. Vis a locally nega- tive definite function if—V is an Ipdf, and is a negative definite function if—V is apdf. Given a continuous function V:R, xR" R, it is rather difficult to determine ‘whether or not Visa pdf or an Ipdf using Definition (3). The main source of difficulty is the need to exhibit the function o(-). Lemmas (6) and (9) give equivalent characterizations of Ipaf’s and péf’s, and have the advantage that the conditions given therein are more readily verifiable than those in Definition (3). 6 Lemma A continuous function W: R” — R isan Ipdf if and only if it satisfies the fol- lowing two conditions wo (ii) there exists aconstant r > O such that W(x)>0, VxeB, - {0}. Wis.a pdf only ifit satisfies the following three conditions: (ii) W@)=0, (iv) W(X) > 0, Vxe R" - (0). (v) There exists aconstant r > O such that int, WOR) >0. Wis radially unbounded if and only if (vi) W(X) > ova Ux ll 0, uniformly in. Proof Consider first the case of Ipdf's. Suppose W is an Ipdf in the sense of Definition (3);then clearly (and (i) above hold. To prove the converse, suppose) and (i)above are true, and define 7 =, inf | WO). ‘Then§(0)=0, is continuous, and is nondecreasing because zp increases, the infimum is taken overa smaller region. Further, (p) > 0 whenever p > Oo see ths, note thatthe annu- lar region over which the infimum in (7) is taken is compact. Hence, if @(p)=0 for some positive p, then there would exist a nonzero x such that W(X) =0, which contradicts (i). Now by Lemma (1), there exists an oof class K such that (p)$6(p) We (0, r]. By the definition of, itnow follows thatSee.5.2 Some Preliminaries 149 8 a(lixil) spcixil) s W(x), VxeB,. Hence Wis an Ipdfiin the sense of Definition (3). Inthe case of pdf's, the necessity of conditions (ii) to (v)is immediate from Definition (@). The remainder ofthe proofisleftasanexercise. M Remark Note that conditions (iii) and (iv) alone, without condition (v), are not sufficient for Wto be a pdf consider the function W: R—> R defined by W(x) =x7A1 +24), 9 Lemma A continuous function V:R,XR"—>R is an Ipdf if and only if (i) V(1, 0)=0'V1,and (i) there exists an ipdf W: R° —+ Randaconstant r > 0 such that 10 Vit, 92 W(x), Vt20, Vxe B,. Visapdfifand only if(i) V(t, 0) =0Vt, and (i) there exists a pd/W: R —> R such that 11 Vee, 2 W0x), ¥e20, Vaxe Re Vis radially unbounded ifand only ifthere exists a radially unbounded function W:R° > R such that (11)is satisfied. Proof ‘The proof is given only for Ipdf's, since the other proofs are entirely similar. ‘Suppose Wis an Ipdf and that (10) holds; then it is easy to verify that Vis an Ipdf inthe sense of Definition (3). Conversely, suppose Vis an Ipdf in the sense of Definition (3), and let o(-) be the function of class K such that (4) holds; then W (x) = o( Ix) is an Ipdf such that (10) holds. ‘The completion of the proof is left as an exercise (see Problem 5.8). Ml Remarks 1. The conditions given in Lemma (6) are easier to verify than those in Definition 0). 2. If W(x) =x’ Mx, where Misa real symmetric mn matrix, then itis easy to show that W isa postive definite function if and only if M is a positive definite matrix (see Problem 5.6). Thus the two common usages ofthe term "positive definite” are consistent. 3. If W(x) is @ polynomial in the components of x, then one can systematically check, ina finite number of operations, whether or not Wis positive definite; see Bose (1982) for details. 4, Lemma (9) shows that a continuous function of # and x is an Ipdf if and only if it dominates, at each instant of time and over some ball in R", an Ipdf of x alone Similarly, a continuous function of rand x is a pdf if and only if it dominates, for all and x, a pdf of x alone.150 Lyapunov Stability chs 5. A function V:RR, xR" > R is decrescent if and only if for each p in some inter- val(0, ), we have 12 sup sup Vinx) <= 13. Examples ‘The function 1 Wirandext ed isa simple example of a radially unbounded pf. Clearly W (0)=0 and W s(x) >0 Vx+0. ‘Also W (x)= IIx? if we take Ill to be the Euclidean norm on R®; hence W, is radially unbounded ‘The function IS Vi( x12 Ded +) isa pdf because it dominates the time-invariant pdf W . For the same reason. ¥, is radially unbounded. However, it is not decrescent, because for each x20, the function V (t,x) is, unbounded asa function of ‘The function 16 Vat, 1,22) =¢xp(-1) +43) is nota pdf because no pdf W: R" > Rexists such that (1) holds. This ean be seen from the fact that, for each x, V(t, x) —>0 as 1, This example shows that itis not possible to ‘weaken the condition (11) tothe statement 17 Vit,x)>0, 0120, ¥x40. ‘The present function V2 isdecrescent. ‘The function 18 Wa(x1,22)=2] +5in? x2 is an Ipdf but is not a pdf. Note that W(0)=0, and that W(x) >0 whenever x#0 and 1 x21 <1, This is enough to ensure that W, is an Ipdf. However, Wis not a pdf, since it vanishes at points other than 0, for example, at(0, ‘The function 19 Waxy x2) =a4 tanh? x2 is pdf since W(O)=0 and W(x) > 0Vx#0. However, itis not radially unbounded, since tanh? x29 Las ry! o0,See.5.2 ‘Some Preliminaries 151 Next we introduce the concept of the derivative of a function along the trajectories of a 4ifferential equation. Suppose V: R, xR" > R has continuous partial derivatives, and sup- pose x(-) satisfies the differential equation 20 i) =Mt, x()), We20. ‘Then the function Vt, x(1)is differentiable with respect tot, and d a Vt, x00) Mi COL+ VUE MOI, XO} We use the symbol Vir, x(¢)] to denote the right-hand side of (21). This choice of symbols is ‘motivated by the fact that 22 Vit, x()=VI to, atta) + f Vie. x(a) ae ‘whenever x(-) is solution of (20). This leads tothe following definition, 23° Definition Let V:R, xR" —>R be continuously differentiable with respect to all of ltsarguments, and let VV denote the gradient of Vwith respect 10x (written asa row vector). Then the function V:R,, xR" > Risdefined by 4 Vax= xu x)+VV(s, Xft, x), andis called the derivative of Valong the trajectories of (20). Remarks 1. Note that V depends not only on the function V but also on the system (20). If we keep the same V but change the system (20), the resulting V will in general be dif- ferent 2. The quantity V(r, x) can be interpreted as follows: Suppose a solution trajectory of (20) passes through xq at time r. Then, atthe instant f, the rate of change of the quantity V{t, x]is Wo, x0) 3. Note that if Vis independent of rand the system (20) is autonomous, then Vis also independent of. ‘This section concludes with a discussion of invariance and of domains of atraction. 25 Definition A ser M CR® is called an invariant set ofthe differential equation (20) if foreachxgeM there exists atye R, such that152 Lyapunov Stability chs 26 St, fo. NoJEM, VEZ Eo. nother words, a setis invariant if, for every inital state in the set, a suitable inital time can be found such that thé resulting trajectory stays in the set at al future times. Note that, in the dynamical systems literature, oe often views a differential equation as being defined for all real r, rather than just all nonnegative f; in such a case, a set M satisfying (26) would be called positively invariant. ‘A few simple examples of invariant sets can be given. First, let ype R", £9¢R, be arbi- trary, and define S (tp, x) to be the resulting trajectory viewed as @ subset of R’, in other words, let S(¢0, %9) = (6 fo, Xo). Then S (fo, Xo) is invariant (Problem 5.9). An equil- ‘ium isan invariant set; more generally, soi any periodic solution. 27 Definition Suppose xj¢R", 1g¢R.. Thena point pe R" is called alimit point of the trajectory s(t, 9. %) if there exists a sequence {t,} of real numbers in {to,) such that t)pooand 28 im ip~stt, fo, Xo)! =0. The set ofall limit points ofS, fo, ¥o) is called the Vimit set of s(, to, X9). and is denoted by Ato, Xo). An equivalent definition is as follows: p is a limit point of the trajectory 8(-, fo, Xo) if, givenany € > OandT <=», there existsar 2 Tsuch that 29 IIp~S(t, to, %0)Il 0 and T p. Next,choose 2 Tsuch thatSec. 5.2 ‘Some Preliminaries 153 32 Ip, —s(t, 10, %9) <2. Such a exists because pj€ Q(t, Xo). Combining (31) and (32) gives (29). Letus define the distance d(x, £2) between a point x and a nonempty closed set as 33 d(x, Q)=mip Ixy ‘Then we have a further result, 34 Lemma LetxoeR", 19€ R,, and suppose s¢. (9, %9) isbounded. Then 35 dist, 1, X9), Ato, Xo) 9 0aSt >=, Proof If (35)is false, then there exists an € > 0 and a sequence of times {1)} approach- ings such that 36 Als, t0, Xo) Wlto, Xo] 2E, Vi However, the sequence (s(; f0.%9)) is bounded. Hence it contains a convergent subse- ‘quence, By the definition of 2(o, x), the limitof this convergent subsequence must belong, to (tq, Xo), Which contradicts (36). Hence (35) is true. Ml ‘The results stated thus far apply to arbitrary systems. The next lemma states a property that is special to periodic (and hence also to autonomous) systems. 37 Lemma Suppose the system (20) is periodic, and let xy€R", to€Ry. Uf 0. %9) is bounded, then Q(to, Xo) isan invariant set of (20). Proof Let Tbe the period of (20), so that 38st, fo, 0) (FAT, to +KT, xp), forall imtegers k>0. Let pe (ro, xp); it must be shown that there exists an initial time te R. such that 39st, %, ple MFo, Xo), WZ. Since pe Q(¢o, x) there exists a sequence {*;) approaching infinity such that (28) holds. Now, for each i, find an integer k, such that r,—k;T€0, 7). Then the sequence {1,—k,T bounded, and therefore contains a convergent subsequence. Choose such a subsequence, renumber it once again as (1,), let te [0, T] denote its limit, and note that (28) continues to hold. Now, since solutions depend in a continuous fashion on the initial conditions and on the time, we have 40 sits, (to X0)]154 Lyapunov Stability chs lim sle-+k,T, + +4)T, 8 to, Xo) by (38) = fim sft-+k 7, £1 86) £0, X0)] = lim (6 +KT, to. %0). where we have used the fact that t= lim (i,~K,T). This shows that (39) holds. i Now letus restrict attention to autonomous systems of the form AL (= fx). Note that, forsuch systems, AZ s(t, t,x) =5(¢ +1, 0, xg), V2 120, Vaye R™ ‘Suppose 0 is an attractive equilibrium of the system (41). By Definition (5.1.27), this implies that there exists aball B, such that every trajectory starting inside B, approaches O as tae, 43 Definition Suppose 0 is an attractive equilibrium of the system (41). The domain of attraction D(0) is defined as 44 DO)= (xER":5(¢, 0,%) > 0as +), Definition (5.1.27) implies that 0 is an interior point of D(0). Note thatthe terms “region of attraction” and "basin" are also used by some authors instead of "domain of attraction.” Also, we identify the region D (0) with the equilibrium 0, since (41) may have more than one attractive equilibrium, in which case each equilibrium will have its own domain of attraction, ‘Lemma (45) below states, among other things, that D(0) is a connected set. AS a prelude to this lemma, the notion of connectedness is defined. Let RT be a given set. ‘Two points x and y are said to be connected in Sif there isa path between x and y lying entirely in $; more precisely, x and y are connected in Sif there is a continuous function ‘h:{0, 1]->S such that h(0)=x, h(1) = y. Obviously, the property of being connected in Sis symmetric and transitive; thus, if x,y are connected in S and y, are connected in S, then so are xand2, The entre set Sis said to be eonnected ifevery pair of points in Sis connected in $. In R",aset Sis connected if and only ft cannot be contained inthe union of two disjoint opensets. 45 Lemma Suppose isan attractive equilibrium ofthe system (41). Then D(O)is open, connected, and invariant eSee.5.2 ‘Some Preliminaries 155 Proof To show that D (0) is invariant, suppose x9¢ D (0); itis enough to show that 46 57,0, xp)€D(O), ¥E20, By definition, xp¢ D (0) implies that AT 546, 0,%) 700519, Now fix t20andnote that 48 5{t, 0, 5(¢, 0, xo) =8(¢+ 1, 0, Xo) 90.81 9 =, by (47). Hence (46) follows. To show that D (0) is open, observe first that 0 is an interior point of D (0), and choose +> Osuch that B, is contained in D (0). Now let xe D (0) be arbitrary; it must be shown that there exists a ball 49° By, = LygER": IIKo~Yoll Osuch that SL Ms, 0, %9)—S(, 0, Yo) ll < 1/2, Vee 0, TI. VICE Baya ‘Such a d exists since solutions of (41) depend continuously on the initial conditions. Now (50) and (51) together imply that 52 Is(T, 0, ¥o)ll 0 as in the preceding para- ‘graph, i., such that B,CD(0). Let Xo, Yor D (O) be arbitrary, and select times Ty,, Ty, such that156 Lyapunov Stability chs 5S S(T. 0, Xp )€ Bes 8(T 0, Yo)e Br By the invariance of D(0), which has already been established, we know that 56 (6, 0,x9)€D(O), We (0, Ty}, ands(t, 0, Yo)e D0), ¥re [0, T,,} Hence, % is connected (0 $(Ty,,0.%)=:Po in DO), and yo is connected to S(Ty,.0, Yo) =:4o in DW). Also, since B, is convex and contained in D(0), Pp and qo are connected in D (0) (see Figure 5.5). Using the transitivity of connectedness, we can finally conclude that xp and yp are connected in D (0). i My, 0.30) Hr, ,0.30)7 4 Fig 55 Foranextension of Lemma (45) tononautonomous systems, see Problem 5.10. Problem 5.5 Determine whether of not each of the following functions is (i) locally positive definite (ii) positive definite, (ii) decrescent, and (iv) radially unbonded: (@) xf ted (b) xj +4 © +P, @ tate). ©) @j+xdMe+)), () sin? (xy +x.) +sin® (x, —x2). Problem 5.6Suppose V:R, xR” > Ris defined by VG, =X Mx, where Mis a continuous function oft, and M() is ral and symmetric foreach t. Determine necessary and sufficient conditions for V'o be (i) positive definite, and (i) decrescent. Problem 5.7 Complete the proof of Lemma (6) Problem 5.8 Complete the proof of Lemma (9).See. 5.3 Lyapunov's Direct Method 137 Problem 8.9 Suppose x¢R", fo€R,, and let S(¢o, X9) denote the resulting trajectory viewed as a subset of R" [see the paragraph after Definition (25)]. Show that 5 (to, xo) isan invariant set. Problem $.10 Consider the nonautonomous system (20), and suppose the origins uni- formly attractive in the sense of Definition (5.1.27). Define the domain of attraction D(O) as D= [xyER":F1gER, SA.st, f9,%) P0as 1 >). Is Lemma (45) stil true? Justify your answer. 5.3 LYAPUNOV’S DIRECT METHOD ‘The idea behind the various Lyapunov theorems on stability. asymptotic stability, and instability is as follows: Consider a system which is “isolated” in the sense that there are no ‘external forces acting on the system. Equation (5.1.1) isa suitable model for such a system ‘because no input is explicitly identified on the right side of this equation. Suppose that one ‘can identify the various equilibrium states of the system, and that @ is one of the equilibria (possibly the only equilibrium). Now suppose itis possible to define, in some sense, the total energy of the system, which isa function having the property that itis zero atthe oi and positive everywhere else. (In other words, the energy function has either a global or a local minimum at 0.) Ifthe system, which was originally in the equilibrium state 0, is per- turbed toa new nonzero intial state (where the energy level is positive, by assumption), then there are several possibilities. Ifthe system dynamics are such tha the energy of the system is nonincreasing with time, then the energy level ofthe system never increases beyond the initial positive value. Depending on the nature of the energy function, this may be sufficient to conclude that the equilibrium 0 is stable. If the dynamics are such that the energy of the system is monotonically decreasing with time and the energy eventually reduces to zero, this may be sufficient to conclude that the equilibrium 0 is asymptotically stable. Finally, if the energy function continues to increase beyond its initial value, then one may be able to conclude that the equilibrium 0 is unstable. Such an approach to analyzing the qualitative ‘behavior of mechanical systems was pionecred by Lagrange, who showed that an equili- brium of a conservative mechanical system is stable if it corresponds to local minimum of the potential energy function, and that it is unstable if it corresponds to local maximum of the potential energy function. The genius of Lyapunov lay in his ability to extract from this, type of reasoning a general theory that is applicable to any differential equation. This theory requires one to search for a function which satisfies some prespecified properties. This function is now commonly known as a Lyapunov function, and is a generalization of the ‘energy of a mechanical system. Subsequent researchers have of course refined the theory considerably. ‘This section deals with the so-called direct method of Lyapunov, sometimes also called the second method. The first, or indirect, method is based on power series expansions and. does not find much favor today. Three basic types of theorems are presented in this section, ‘namely: stability theorems, asymptotic stability theorems, and instability theorems. The various theorems are illustrated by several examples.158 Lyapunov Stability chs ‘Throughout this section, the following three abbreviations are eriployed to make the theorem statements more compact: C': continuously differentiable Ipdf: locally positive definite function pdf: positive definite function 8.3.1 Theorems on Stability ‘Theorem (1) isthe basic stability theorem of Lyapunov’ direct method. 1 Theorem The equilibrium 0 of the system (5.1.1) is stable if there exist a C' Ipdf V:R, xR" Randaconstartr > Osuch that 2 VG,x)<0, Wr2%9, WEB, where V is evaluated along the trajectories of 5.1.1). Proof that ince Vis an Ipdf, there exist a function cof class K and a constant s > 0 such 3° a(iIxil) SVG, x), Wr20, VxeB,. To show that 0 isa stable equilibrium, we must show that given any € > Oand any fo 20, we can find a §=8(e, 19) such that (5.1.10) is satisfied. Accordingly, given © and fo, let €,=min{e, r, s),and pick 8 > Osuch that sup, (to. =: Beto, 3) < ae). vat Such a8 can always be found because ae) > Oand Bitp, 8)» 0s 5—+0. To show that the above choice of 8 satisfies (5.1.10), suppose IIx II <8. Then V(to, x0) Osuch that160 chs 15 Vit, x) $0, ¥20, Vxe B,. Proof Siace Vis decrescent, the function 16 B= HB BME is finite for all sufficiently small 8, and is nondecreasing in 5. Now ete, =min(e, r, 5} and pick 5 > Osuch that 17 Be6)< ore) Now proceed exactly as in the proof of Theorem (1) to show that (5.1.11) holds with this choice of 8. The details are left as an exercise. i Remarks 1, ‘Theorem (1) states that if we can find aC Ipdf V such that its derivative along the trajectories of (5.1.1) is always nonpositive, then the equilibrium 0 is stable. Theorem (14) shows that in order to conclude the uniform stability of the equili- brium 6, itis enough to add the assumption that Vis also decrescent. It should be noted that Theorems (1) and (14) provide only sufficient conditions for stability and uniform stability espectively. [But the converses of these theorems are also true; see Hahn (1967).] 2. The definitions of stability given in Section $.1 are qualitative, inthe sense that given an €> 0, one is only required to find some 6 > 0 satisfying (5.1.10); to put another way, one is only required to demonstrate the existence of a suitable 3. In the same way, Theorems (1) and (14) are also qualitative in the sense that they provide conditions under which the existence of a suitable 8 can be concluded. However, in principle atleast, the conditions (4) if Theorem (1) is being applied, (17) if Theorem (14) is being applied, can be used to determine asuitable 8. But in practice this procedure israther messy, and often gives too conservative an esti- mate for 3. 3. The function Vis commonly known as a Lyapunov function of a Lyapunov func- tion candidate. The term Lyapunov function isa source of great confusion, In an ‘attempt to avoid confusion, the following convention is adopted: Suppose, for example, that we ae attempting to show that ® isa stable equilibrium by applying ‘Theorem (1). Then a function Vis referred to as a Lyapunov function candidate ifit satisfies the requirements imposed on V in the hypotheses of Theorem (1), i.e. if Vis" and isan Ipdf. If, fora particular system (5.1.1), the conditions imposed con V are also satisfied, then V is referred (0 as a Lyapunov function. The rationale behind this convention is as follows: Theorems (1) and (14) are sufficient conditions for certain stability properties. To apply them toa particular system, it isa fairly simple matter to find a function V satisfying the requirements on V. At this stage, Vis a Lyapunov function candidate. Now, for the particularSee. 5.3 Lyapunov's Direct Method 161 system under study and for the particular choice of V, the conditions on V may or ‘may not be met, If the requirements on Vare also met, then definite conclusions ccan be drawn, and V then becomes a Lyapunov function. On the other hand, if the requirements on V are not met, since these theorems are only sufficient condi- tions, no definite conclusions can be drawn, and one has to start again with another Lyapunov function candidate. The examples that follow illustrate this usage. 18 Example Consider again the simple pendulum, which is described by 6+sine=0 after suitable normalization. The state variable representation of this systems in Now the total energy of the pendulum is the sum of the potential and kinetic energies, which is Viena) = ate, 08.1): ye where the first term represents the potential energy and the second term represents the kinetic energy. One can readily verify that V is C’ and an Ipdf, so that V is a suitable Lyapunov function candidate for applying Theorem (1). Computing Vgives Veer x2) =sina iy 4x9 Fp =Sinxy(x2)-+x9(-sinz))=0. ‘Therefore Valso satisfies the requirements of Theorem (1). Hence Vis actually a Lyapunov function, and the equilibrium 0 is stable by Theorem (1). Further, because the system is autonomous, 0 isa uniformly stable equilibrium: see Theorem (5.1.43). 19 Example Consider the rotational motion of a rigid body in three-dimensional space. If denotes the angular velocity of the body and I the 3x3 inertia matrix of the body (both measured ina coordinate frame rigidly attached to the body), then in the absence of external torques the motion is described by 2 Ia+axlo=0, where x denotes the vector cross product. Equation (20) can be simplified considerably if the coordinate axes are chosen to be the principal axes of the body, i., a set of axes with respect to which Iisa diagonal matrix. Accordingly, let @=[0, 0, @,)',1=Diag (ly. by, ). ‘Then (20) reduces to162 Lyapunov Stabi chs A 1 O.=-(-h) 0, 0, Fy @y = (UL, ~1,) @, 1,0,=~(y~1) 0,0. Again, suppose without oss of generality that , > /, 2, >0. For notational simplicity, let usteplace ©, 0, by x y,z.respectively, and define kok lk Note thata, b, ¢20, Then (21) finally assumes the form 22 it=ayz, j=—bez, cy. [At this stage, assume for simplicity that the principal axes are unique; ths is equivalent to assuming that /, > J, > [,, or that a, b, c >0. This assumption excludes bodies with some symmetry, stich asa spinning top for example. Then the system (22) isin equilibrium if and only ifat least rwo of the quantities x, y, zare equal to zero. Hence the set of equilibria con- sists of the union ofthe x, y, and the z axes. Physically this corresponds to rotation around fone of the principal axes at a constant angular velocity. Note that none of the equilibria is, isolated. Consider first the equilibrium atthe origin, and try the obvious Lyapunov function can- didate VO y, 2)= px? + gy? +23, where p, q, 7 >0. Then Vis an Ipdf. (Actually, Visa pdf and is radially unbounded, but this fact isnot needed.) Computing V gives V=2 (pu + gy) + ei) =2ay2 (ap —bq +07). Clearly itis possible to choose p, q, r > Osuch that ap bg +er=0. For such a choice, V=0, which satisfies the hypotheses of Theorem (1). Hence 0 is a (uni- formly) stable equilibrium. Next, consider an equilibrium of the form (r9,0,0) where x9 #0. At this stage we can do one of two things: () We can translate the coordinates such that (x9, 0, 0) becomes the origin ofthe new coordinate system. This would enable us to apply Theorem (1) directly, ‘but would have the effect of making the system equations (22) more complicated. (ii) Alter natively, we can onstruct a Lyapunov function candidate V such that V is C', V¢xp,0,0)=0, and Vix, y, 2)> O for al (x,y,z) (xo, 0,0) and sufficiently near (x9, 0,0)See.5.3 Lyapunov’s Direct Method 163 For the sake of variety, the second approach is followed here; the first approach is discussed further in Problem 5.11. For the new equilibrium, let us try the Lyapunov function candi- date W(x, y, 2) =cy? +b2? + [2acy? +abz? + be(x? —33)F ‘Then W(*0, 0,0)=0,and W(x, y, 2) > Dif (, y, 2)#(Ex9, 0, 0). Hence Wis an Ipdf with respect tothe equilibrium (xo, 0,0), buts nota pdf since it vanishes at another point as well. Now routine computations show that W 20. Hence (9, 0, 0) isa stable equilibrium, By entirely analogous reasoning, one can show that every equilibrium of the form (0, 0,20) is also stable (Problem 5.11). However, it turns out that every equilibrium of the form (0, y0,0) with yo #0 is unstable. This is shown later in this section {see Example (105)]. Physically, this means that an object can be made to spin about its major axis and its ‘minor axis, but not about its "intermediate" axis. 23 Example Consider the system described by HU)+HO]+2+sing) y()=0, or, in state variable form, Fy say ty =o 9g 2+ sin, Note that the system is linear and periodic. This is an example of a so-called damped Mathieu equation. (If the j term were not there it would be an undamped Mathieu equation.) In this case there is no physical intuition readily available to guide us in the choice of V. Thus (after possibly a great deal of trial and error), we might be led to try the Lyapunov func- tioncandidate Veuxuxpaat+ il 7 2+sint ‘Note that V is periodic with the same period as the system, Now Vis C' it dominates the time-invariant pdf Wi, x2)=37 +298, and is dominated by the time-invariant function Wax), X2)= a7 +83. Hence, by Lemma (5.2.9), V is a pdf and decrescent, and is thus a suitable Lyapunov164 Lyapunov Stability h.5 function candidate for applying Theorem (14). Now a 2 VO, 12) =- 9 2 i de sin esing 2__ cost 2 F asing? *M2* Fesing 222 tind cost +212 4sin) (Q4sint? 442sint +eost 9 Qesinn? $0, VF20, Vry.x2. ‘Thus the requirements on Vin Theorem (14) are also met. Hence Vis a Lyapunov function for this system, and 0 isa uniformly stable equilibrium. 24 Example One of the main applications of Lyapunov theory is in obtaining stability conditions involving the parameters of the system under study. As an illustration, consider the system JO +POIO+EVO=0, ‘which can be represented in state variable form as Sysxata=—pxg-e tn. ‘The objective is vo find some conditions on the function p(-) that ensure the stability of the equilibrium 0, For this purpose, let uschaose Ves mmaad t's} Since Vis C! and dominates the pdf Wer agdaxt +43, Vis a suitable Lyapunov function candidate for applying Theorem (1). However, since Vis not decrescent, itis nora suitable Lyapunov function candidate for applying Theorem (14). Hence, using this particular V-function, we cannot hope to establish uniform stabi applying Theorem (14), Differentiation of VgivesSee.5.3 Lyapunov’s Direct Method 165 V(t. x),X2) =x] + 2x (2) + 2e'xal-P(e, ex] =e'x}I-2p +11 Hence we see that is always nonpositive provided 1 p> >. W120. ‘Thus the equilibrium Ois stable provided the above condition holds. It should be emphasized that by employing a different Lyapunov function candidate, ‘we mightbe able to obtain entirely different stability conditions involving p(). 5.3.2 Theoremson Asymptotic Stability In this section we present some theorems that give sufficient conditions for uniform asymptotic stability, exponential stability, and global versions of the same. 25 Theorem The equilibrium 0 of (5.1.1) is uniformly asymptotically stable if there existsaC! decrescentipdfV such that -V isan Ipdf. Proof If-Vis an pdf, then clearly V satisfies the hypothesis of Theorem (14), so that isa uniformly stable equilibrium. Thus, according to Definition (5.1.31), itis only necessary to prove that 0 is uniformly attractive. Precisely, itis necessary to show the existence of a 5, > Osuch that, for each e > Othere exists a 7(€) O such that 27 aflixil)S V(r, x) SP(IIxt), Vt2 to, VxEB,, 2B Vit, x)S—Y(IIxll), Ve2 to, VxEB,. Now, given > 0, define positive constants 8, 5,,and by 29 8 Hs(t3, to. Xo) ll

0, Vox) <0, Vx#0inS, Then Theorem (25) applies and Osan asymptotically table equilibrium. Now one can ask: ‘Does (38) imply that Sis contained in the domain of attraction D (0) defined in (5.2.43)? In other words, does (38) imply that s(, 0, x) 0 as 1-90 whenever xo $? One might be tempted to think that the answer is yes because (38) implies that VIs(, 0, x)] gradually decays 100 as to», but this reasoning is false. Ife S and ifs(t, 0, X9)€ S V1 20, then (38) would imply that Vist, 0, x)}—>0, and hence that s(/, 0, x»)—>0. However, (38) alone doesnot guarantee that every solution that startsin S stays in S. The valid conclusion is this: If (38) holds, then every invariant set of (5.1.1) contained in Sis also contained in D (0), but S itself need not be contained in D(0). But how does one go about finding such invariant sets? As easy way isto use so-called level sets of the Lyapunov function V. Let ce R, , and consider the set 39 Mulc)=[xeR": Vix) Sc}. Note that, depending on the nature of the function V, the set My(c) need not be connected (ee Figure 5.6). However, @always belongs to My(c). Now the level set Ly(c)is defined as the connected component of My(c) containing 0. Another equivalent definition is the fol- lowing: Ly(c) is the set of allxe R” withthe property that there exists a continuous function b:(0, 1] +R" such that h(O)=x, b(1)=0, and V[b(r)]-R and a domain 5 containing 0 such that V(0) =0 and (38) is satisfied. Let c be any positive constant such tha the level set Ly(c) is contained in Sand is bounded. Then Ly(c)isasubset of D (0). Proof First itis shown thatthe level set Ly(c) is invariant for the system (41). Suppose Xo€Ly(c). Then168 Lyapunov Stability chs 42 Vist, 0,x9)] $V (xp) since V(x) <0 VxeLy(c) CS, which implies that s(t, 0, xp)e Ly(c) Vt20. Since Ly(c) is bounded, the solution trajectory does not escape to infinity. Next, since ~Vis an Ipdf in L,(c), one can now proceed as in the proof of Theorem (25) toshow that s(, 0, X) —>as¢—> =, Ml Fig.5.7 43 Example As an illustration of Theorem (25) and Lemma (40), consider the nonlinear RC network shown in Figure 5.7, where R is a nonlinear resistive network, terminated in a bank of m linear capacitors. Denote these capacitances by Cy,"-",Cq. and assume that all ‘capacitances are positive. Let x; denote the voltage across the i-th capacitor, and et x denote the vector [xy °*+ xq]. Then the current vector through the capacitors is just Cx, where C=Diag (C1,°°°, Cah Let ix) denote the current vector that results when a voltage vector x is applied across the terminals of the resistive network. Suppose the network is unbiased, in the sense that §(0)=0. IfiisaC' function of x, then there exists a continuous function G: R” > R™ such that [see Lemma (2.5.17)], i) =GQ)x. (One can think of G(-) as the nonlinear version of the conductance matrix. Hence the overall systemis described by CX+G(x)x=0, or K=— CGE). ‘Now Os an equilibrium of this network, by assumption. ‘Tostudy the stability ofthis equilibrium, let us try the obvious Lyapunov function can- didate, namely the total energy stored in the capacitors. This equals, ie Vony= 5x'Cx ThereforeSee.5.3 Lyapunov’s Direct Method 169 Ve) 4ecxe¥ C8) 16H +60 where G’(x) isa shorthand for [G(x)J’. Define M(x)=G'(x) + G0). If there isa constant r > O such that M(x) is positive definite matrix forall xe B,, then -Vis an Ipdf, and the equilibrium 0 is asymptotically stable. Actually, the condition can be simplified further: Since M(-) is continuous, the equilibrium 0 is asymptotically stable if M()is positive definite. "Next, we focus atention on the particular network shown in Figure 5.8, and show how Lemma (40) can be used to estimate the domain of attraction D(0). The element; in Fig ure 5.8isa conventional diode with he i-v characteristic shown in Figure 5.9, while the ele- ‘ment 6 is a tunnel diode with the iv characteristic shown in Figure 5.10; the element denoted by 3 isa linear resistor wth positive conductance g3.Itis easy to see that [repate ig O20) +as0e2 x1) dE az Fig. 5.8 ainds6 Fig.5.9170 Lyapunov Stability chs Ifwedefine 4 4: arte Tegal thep the nonlinear conductance matrix G (x)can be written as awe[t? 2] 83 82483 where the arguments of g; and g2 have been omitted for clarity. Clearly G is positive definite ifand onl 81 +85 >0,detG (x)= 8182 +85(81 +82)>0. Since g (x1)>0 Vx and g > 0, the first condition is always satisfied, and only the second condition needs tobe verified. Rewrite this as The right sie of (44), without the minus sign isthe equivalent conductance of g and gs connected in series. At x; =x2 =0, (44) is satisfied since g» >0. Hence G (0) is positive definite, and by earlier reasoning, tis is enough to show that 0s an asymptotically stable equilibrium. ‘Next, letus determine a region Sin the x x2 plane such that (4) is satisfied, Note that wherever (44) is satisfied, G (x) isa positive definite matrix and thus V(x) < 0 (except at the origin of course). However, even if G(x) fails to be a positive definite matrix at a particular point x, itis nevertheless possible that V(x) is negative. ‘The reason is that the positive definiteness of G(x) implies that y’G oy >0Vy#0, whereas all we really need is that XG (x > 0. Ifx2 <0, then g (+2) > 0, sothat (44) is automatically satisfied. Now suppose +220, If is largeand positive, then g ~>e and g,, —> 3. Hence g(%2)>—g provided 2 does not belong to [b, c }, where b and c are identified in Figure 5.10. If x, is large and negative, then g | ->0, and gz, —r0. In this case g3(x2) > Bey provided x2 does not belong, tola, d}, where a and dare identified in Figure 5.10. In summary, the region where G (x) is not positive definite isthe shaded region shown in Figure 5.11. Hence, if we define Sto be the complement ofthe shaded region, then (38) holds. Now we know that every bounded invariant set contained in Sis also contained in the domain of attraction (0), Lemma (40) tells us that every bounded level set of Vcontained in Sis invariant, and is thus a scbset of, ‘DQ. Inthe present example, the level ets Lyld)= (Cry, x2): Cx + Caxd Sd) are ellipses centered at the origin, Hence an estimate for D(0) based on Lemma (40) is, ‘given by the ellipse shown in Figure 5.11See. 5.3 Lyapunov’s Direct Method. m ales) *6 Fig.5.10 Further analysis of this network is suggested in Problem 5.12. i ‘The next theorem gives a sufficient condition for exponential stability 45 Theorem Suppose there exist constants a,b, c,r>0, p21, and a C! function VR, xR" > R such that 46 allxl” SVG, x) Sb IixIl?, Ve20, WxEB,. 47 Vit, x)S—clIxlI?, ¥120, VxeB,, Then the equilibrium Vis exponentially stable. and suppose xy By, fo 20. Then, letting x() denote the solutions(, 0, x3), we have72 Lyapunov Stability chs dy © LV, ay} s-c txt <- EV, x 49-2 V, x) se batty” S— EVE XC) Hence 50 Vitg+t, alto +1)1< VI to, xolexp ley], W120. But since SL V(t, 9) $b xy IP, and 52. allatty +1)" SV [tg +t, xltg +9) itfollows that 53° alin(tg +1) Sb xg” expl-{c/)t], WFO. Finally, fo|* 54 Ialig+ i's) 2] xg llexpt-(e/py}, W120. Hence 1.37) is satisfied with (/a)" playing the rote ofa and cop playing the role of b. Thus Ois an exponentially stable equilibrium. i 58 Example Consider again the nonlinear circuit of Example (43). Let Cumin C;, Cy =max C, ‘Then Lqiixi®s vias Ley x, VxeR". a ae If the matrix M(O) is positive definite, then by continuity M(x) is also positive definite for each x belonging to some ball B,. Let d= inf Ryo MOO). where Anin(’) denotes the smallest eigenvalue of a symmetric matrix, and choose r>0 sufficiently small thatd >0. Then V0) =—xMOoxs—dilx?, Vxe B,. ‘Thus all hypotheses of Theorem (45) are satisfied, and we conclude that the equilibrium 0 is, in fact exponentially table. iSee.5.3 Lyapunov’s Direct Method 173 ‘The theorems for global uniform asymptotic stability and global exponential stability are straight-forward generalizations of Theorems (25) and (45) respectively. 86 Theorem The equilibrium 0 of (5.1.1) is globally uniformly asymptotically stable if there exists a C! function V:R, xR" —R such that (i) Vis a pdf. decrescent and radially unbounded, and (ii)=V isa pdf Proof The hypotheses include those of Theorem (25); hence Ois uniformly asymptoti- cally stable. Thus, to prove the theorem, it only remains to prove that 9s globally uniformly attractive, ie. that, given any M 0, there existsa7™=T(M, e) such that $7 Uxpl < Mit 20> listo +1, f0.%0)Il Osuch that 59 BUM) < air), This is possible since a(r) ->e0 as r >». Then, following the reasoning of Equations (9) — (23), it can be shown that every trajectory that starts in the ball By stays inside the ball B,, ice all trajectories of the system are bounded. Now choose 60 & 0,p2l,andaC'! function V:R, xR" -» R such that 63 allxll? sVi, x) Oisa constant, The solution ofthis equation isSec.5.3 Lyapunov’s Direct Method 175 Finite xine Fig.5.12 (0) O= Tar Ont Hence, if r(0) > 0, itfollows that HM) east T= 3G Inother words, the trajectory exhibits finite escape time, Now consider the analogous equa- tion HO=2B)F7O, where Byza>0, ve If r(0) >, then the solution of this equation increases at least as rapidly as the solution of the previous equation, since the present /(t) is at least as large as its inthe previous equa- tion. Therefore the solution of this equation also exhibits finite escape time, and the escape timeisno larger than 1 2ar(0)" ‘Returning now to the system at hand, observe that176 Lyapunov Stability ch.5 ota | 11 25105, Ve, 22 Viet 1 te | Hence, from(66), 422-15, orxg(t)2x2(0)~ 1.51, W120. Now let r=x?. Then from (66), 0 (x2 = 1)? Now itis claimed that, whenever the initial condition X =(x 9, x20) of (66) lies in the first or second quadrant and saiisfies 6B (x29— 1): 1401 >Y3, the resulting solution trajectory exhibits finite escape time. (Note: Itisnot claimed that these are the only initial conditions that lead tonite escape time.) Suppose (68) holds. Then (1) 2x29~ LSt, VI20. Define xm-l 3 ‘Then x71 2 22(1)= 12x29 - 151-12 + VE 10, Trl In view of the earlier discussion, the differential equation (67) exhibits finite escape time ‘hich isno larger than 1 “Gn Dxto | [eax View) T woe Meo) ‘This “closes the loop" on the circular reasoning and establishes the claim. Figure 5.12 ‘shows the initial conditions which are guaranteed to lead to trajectories having finite escape time. The final two theorems on asymptotic stability are applicable only to autonomous or periodic systems. This isin contrast to all ofthe preceding theorems in this section, which ccan be applied to arbitrary nonautonomous systems. The main feature of these theorems is that they enable one to claim (uniform) asymptotic stability if there exists C Ipdf V whose derivative V is nonpositive along trajectories, even if V is not locally negative definite,See,5.3 Lyapunov’s Direct Method 7 However, it should be noted that these theorems do not allow one to conclude exponential stability, ‘These theorerss were first proved in the Soviet Union by Barbashin and Krasovskii (1952) in a special case, and by Krasovskii (1959) in the general case. Later they were independently rediscovered in the West by LaSalle (1960). In Western literature these theorems are often referred to as LaSalle’s theorems, but itis more accurate to call them Krasovskii-LaSalle theorems. Asa prelude to stating these theorems, let us extend the notion of level sets to functions. of both r and x. Suppose V:R, xR” — R is continuous, and suppose V(r,0)=0 V1 20. Define 69 My(c)=(xeR":3r20 such that V(t, x) Se}, Note that if V is independent of 1, then (69) reduces to the earlier definition (39). Now Oc My(c) whenever c20. The level set Ly(c) is now defined as before, namely, the con- ected component of My(c) containing 0. Next we define another set, namely 10 Ay(c)=(xeLy(e): V(t, x) Se, Vt20}. Note the difference between the quantifiers in (69) and (70); also note that if V is indepen- dent oft, then A y(c) isthe same as Ly(c). ‘The following lemma is of independent interest, even if 0 is not asymptotically stable. 71 Lemma Suppose the system (5.1.1) is periodic. Suppose there exists a C' function V:R, xR" — R such that (i) V is periodic with the same period as the system, (ti) Vis an Ipdf, (iti) there exists an open neighborhood N of such that 72 Vit, x) <0, ¥120, WHEN. Choose a constant. > Osuch that the level set Ly(c) is bounded and contained in N. Finally, et TB. $=(xeLy(c):3120such that V(r, x)=0), and let M denote the largest invariant set of(5.1.1) Contained in S. Then TA xEAy(c), 10202 dIstt, t0,%9), M1 0as1 >, where d(y,M) denotes the distance from the pointy tothe set M (cf. (5.2.33). Proof Since Vis periodic, itis decrescent, so that Oi an interior point of Ay(c). Using the methods of Equations (9) ~ (13), it is easy to show that if xeAy(c) then tt, to, %)ELy(c) Wt to. Since Ly(c) is bounded, the limit set Oro, Xp) [see Definition (5.2.27)] isnonempty. Further, by Lemma(5.2.34),178 Lyapunov Stability chs 75 dIX(0), Rtg, Xo) F0ast—>e, ‘where x(t) isa shorthand for s(, 0, xy). Further, by Lemma (5.2.37), Q(to, X9) isan invari- ant set. Consider now what happens to the function V{t, x()}. Since Vt, x(7)] SOE 210, V(t, x()] is monotonic and has a definite limit as ¢ >»; also VUt, x(¢)] -¥0 as t p22. Sup- pose ye (to, Xo). Then, by definition, there exists a sequence {t;] approaching e such that X()—>y. Select integers &; such that t,~k;Te[0, 7) where T is the period. Then the sequence {t,k:T} is bounded and therefore contains a convergent subsequence. Renumber the subsequences again as {,} and (&,), and let t denote the limit of the sequence (kT). Then G6 Veey)= him Vee ~ im Veen y) = him 114, x01 Inother words, ye S. Since this is true of every ye 5, it follows that {(r9, x) 5, and since ‘to, Xo) i an invariant set, it follows that Af, x4) ME. The desired conclusion (74) now follows from (75). i 77 Theorem (Krasovskii-LaSalle) Suppose the system (5.1.1) is periodic. Suppose there exists a C' IpdfV:R, xR" +R having the same period as the system, and an open neighborhood N of O such that (72) holds. Choose a constant ¢ >0 such that the level set Ly(c) is bounded and contained in N, and define S as in (73). Under these conditions, if S contains no trajectories of the system other than the trivial trajectory x(t) =0 te 20, then the equilibrium is uniformly asymptotically stable. Proof This theorem is essentially a corollary of Lemma (71). Let M be the largest invariant set contained in S. Itis claimed that M = (0). To see this, et ye M. Then, by the definition of invariance, there exists a fo20 such that the corresponding trajectory sit, to, YJeM Wr2 to. However, by assumption, $ does not contain any trajectories other than the trivial trajectory, and Misa subsetof S. Hence y =0, ic.,M = (0). Next, note that since M-= (0), the distance d(g, M) is ust Wzll. Hence, by Lemma (71), and in particular (74), TB xyEAv(c)> Ist, . %9)!l 90, a5 t=. Hence the origin is attractive. It is also stable, by Theorem (1), and is thus asymptotically stable. Finally, by Theorem (5.1.49), 0isa uniformly asymptotically stable equilibrium. iSee,5.3 Lyapunov’s Direct Method 179 Remark The proof of Theorem (77) makes it clear thatthe set A y(c) defined in (70) is contained in the domain of attraction, in the sense that every trajectory starting in Ac), at whatever initial time, approaches 0 as > 2». 79 Theorem (Krasovskii-LaSalle) Suppose the system (5.1.1) is periodic. Suppose there exists a C' function V:R, xR" +R having the same period as the system such that (i) Visa pdfand is radially unbounded, and (ii) 80 V(r, x) <0, Vr20, Vee R". Define 81 R= (xe R":3120such that VG, x)=0), and suppose R does not contain any trajectories of the system other than the trivial trajec- tory. Then the equilibrium 0 is globally uniformly asymptotically stable. Proof Since V is radially unbounded, each set My(c) defined in (69) is bounded, whence Ly{c) is also bounded for each ¢ > 0. Thus, proceeding as inthe proof of Theorem (77), one can show that (78) holds for every ¢ >0. Thus the equilibrium Ois globally attrac- tive; itis also uniformly stable, by Theorem (1). Thus itonly remains to show that the attrac- tion to 0 is uniform with respect fo and lixy I. This part of the proof is omitted, and the reader is referred to Hahn (1967), Theorems 38.3 and 38.5, or Krasovskii (1959), Theorem 141. ‘The application of Theorems (77) and (79)isillustrated through several examples. 82 Example Consider a unit mass constrained by a nonlinear spring and nonlinear ftic- tion, Sucha system can be represented in state variable formby Sista 8(1)-S@2), ‘where g(°) isthe restoring force of the spring and fi) is the force due to friction. Suppose that), g() are continuous, and that f(0)= g (0) =0; then 0 isan equilibrium. In addition, suppose there exists aconstant > Osuch that rf (1) 90, ¥r#0,re(-He th re ()>0,Vr#0,re(-H. Wh Finally define te function (7)=| e(oras, and suppose there is a constant c >0 such that the level set L.g(c) is bounded. Under these conditions, itis claimed that 0 is an asymptotically stable equilibrium.180 Lyapuny Stability chs ‘Beforeestabishing the claim, let us reflect on what it means. The claim, simply, is that if the restoring spring force and the friction force are first and third quadrant functions in some neighborhood of the origin, then the origin is asymptotically stable. To prove the claim, select the total energy of the system as an obvious Lyapunov func- tion candidate, This is the sum of the potential energy stored in the spring and the kinetic energy of the mass. Thus Vexy,43)= O(n) + 4a =f s@do+ Then Vor, 2) =8 Oi 42. F2 =f (HD). Now various properties of Vand V are demonstrated, and finally asymptotic stability is con- cluded on the basis of Theorem (77), 1) Visan Ipdf. To show this, suppose that x0 and that Lx, |, Liz! Sq. Then V(x)>0 by virtue ofthe conditions on f(-)and g(-). (2)¥-0, rg(r)>0,Wr20, (7) e048 Ir 20, then Theorem (79) would apply and we can conclude that the origin is globally uniformly asymprotically stable. 83 Example A phase-locked loop in communication networks can be described by the equation FO+la+bapeosy HO +eMsinyo where b(-, c() ate periodic functions with the same period. In this example, the stability of this system is analyzed. One of the objectives ofthis example is to illustrate the difference between the sets Ly and Ay in (70). ‘We begin by rewriting the system equation in the form 52x20, 50 =-[a +b (Neos x (OL xxl) ~c() sinx (0. Suppose a > Oand that the following conditions hold. (-)iscontinuous and c(-)is C!, Ib@|Sby 0, VE, by assumption. Hence V(t, x) $0\Vt,x. To apply Theorem (77), the only remaining condi- tion i that the set $ defined in (73) does not contain any nontrivial trajectories. Since the fanction d(>) defined above is continuous, periodic, and positive-valued, i follows that dt) isbounded away from zero. Hence 3r20 such that Vor, x)=Oiff.x, Inother words, 8 = (xe Ly(d):x2=0) = ((44, 0): 1x41 Seosd), Suppose now that x(-)isa trajectory of the system lying entirely in S. Then 27(1) 20> 42(1) 20 c(0) sinx (20> x (20. Hence x(-)is the trivial trajectory. Therefore, by Theorem (77), is a uniformly asymptoti- cally stable equilibrium. Moreover, by the remark following the proof of Theorem (77), the setAy(d) for each d < 2s in the domain of attraction; but the same is not necessarily true of Lyd). 84 Example (Stabilization of a Rigid Robot) As an application of the Krasovskii- LaSalle theorem, consider the problem of stabilizing a rigid robot which is operating in a gravity-free environment. The absence of gravity can come about because the robot is ‘operating in outer space. Even in more “down to earth" applications, this assumption is valid if the robot is constrained to operate in a plane which is perpendicular to gravity, for example, a table-top robot operating on an air cushion, The assumption of rigidity ensures that the number of degrees of freedom equals the number of control actuators. Let q=[q1"--qql denote the vector of generalized coordinates of the robot, and let 1u=[u, *uq]" denote the vector of generalized forces, The assumption of rigidity means, in effect, that the vector u can be chosen arbitrarily and is thus a suitable control input. Now the dynamies ofthe robot are described by the Euler-Lagrange equations cll dr au) _ ae aq where Lis the Lagrangian of the system, Since its assumed that there is no gravity, the potential energy of the robot isa constant, which can be taken tobe zero, Hence the Lagran- sian equals the kinetic energy K. Asiscustomary, assume that K=7 4d D@a, where the matrix D(q) is called the inertia matrix. This matrix is configuration dependent184 Lyapunov Stability ch.s ‘but always positive definite. It is reasonable to assume that there exist positive constants o. and Bsuch that 0 0. Now combining all of these inequalities shows thatSec.5.3 Lyapunov’s Direct Method 187 9 BOE)ZV[, MOL=VIt x +f VU MOL drZe += 11)d, Vez However, the inequality (90) is absurd, since the right side is an unbounded function of 1 ‘while the left side isa fixed constant. This contradiction shows that the assumption is false, ive. it is nor true that lIx(1)ll @ and an r>0 such that every trajectory starting in B, other than the trivial trajectory eventually leaves the ball By While such instability occurs sometimes (e.g..the Van der Pol oscillator), this particular ins- tability theorem is much less useful than Theorem (88). Alternate sufficient conditions for instability are given by Theorem (92) below and ‘Theorem (99) following, 92 Theorem The equilibrium 0 of (5.1.1) is unstable if there exist a C' function V:R, XBR" R and a constant r>0 such that (i) V is decrescent, (ii) V(0,0)=0 and V(O,") assumes positive values arbitrarily close tothe origin, (iii) there exist a positive con- stant and a function W: R, XR" ~> R such that188 Lyapunov Stabi chs 93 VU, x)=AVet, x) + WU, x), and 94 Wit, 920, W120, We B,. Proof It is shown that if we choose =, then (5.1.10) cannot be satisfied for any choice of 5> 0. Given 8> 0, choose xy #0 in Bs such that V(O, %9)> O, and let x(7) denote the resulting solution trajectory s(, 0, %). Then, whenever x((e B,, we have 95 vn, XOI=AV Ie x(0)] + W Ut, OIZAV Ls, xO). and therefore 96 L lexpe-ADVIe-x(ON) 20, Hence 9 Vie, x12 VO, xpexp (A). Since the function on the right side is unbounded, x(1) must eventually leave B,. Therefore 0 isan unstable equilibrium. ml 98 Example Consider the system of equations iy oxy 42ey tryxh fy =2e, ta ade Let Very. x)=af =} ‘Then Vis a suitable Lyapunov function candidate for applying Theorem (92), Now Vt, x) =2e 1 Deny =e} 2x9 + Arba =2V0), x2) + 4edd Since W(x), x2) = 4x33 20 Vx, all conditions of Theorem (92) are satisfied, and 0 is an unstable equilibrium. i In Theorems (88) and (92), the function Vis required to satisfy certain conditions at all points belonging to some neighborhood of the origin. In contrast, the various conditions in ‘Theorem (99) below are only required to hold in a region for which the origin is a boundary point, This theorem is generally known as Chetaev’s theorem. 99 Theorem (Chetaev) The equilibrium 0 is unstable if there exist a C! function VIR, xIR" +R, a ball B,, an open set Qc B, anda function of class K such thatSec.5.3 Lyapunov’s Direct Method 189 Fig.5.14 100 0 0 to be specific. First Jetus translate the coordinates so that the equilibrium under study becomes the origin. This isachieved by rewriting the system equations as = a(y-y0)+ ave j—jo=—bxe, E=ex(y—Yo)+e¥¥0 Ify—yo isdenoted by y,, then the equations become190 Lyapunov Stability ch.5 Faysz +ayge, jy =~ bx, T= cay, Hexy9. Now apply Theorem (99) with Vox ys 2 a2, By= (x yy thx? 493 +22 < 98), Q=((4, yp. 2)€B,p:4>0 andz >} Then isopen,and BQOB, = (Cr, 0), where B, denotes the closed ball of radius r/2. Hence conditions (100) to (103) are satisfied. Finally, Vaxi+iz=2(y, +yoN(ex? +42"), IF (x, y,, 2)€, then y, +0 > 0, so that (104) is also satisfied. Hence, by Theorem (99), the origin (in the new coordinate system) is an unstable equilibrium. 5.3.4 Concluding Remarks In this section, several theorems in Lyapunov stability theory have been presented. ‘The favorable aspects of these theorems ar: I. They enable one to draw conclusions about the stability status of an equilibrium without solving the system equations 2. Especially in the theorems on stability and asymptotic stability, the Lyapunov function Vhasan intuitive appeal as the total energy of the system. The unfavorable aspects of these theorems are: 3. They represent only sufficient conditions forthe various forms of stability. Thus, iffa panicular Lyapunov function candidate V fails to satisfy the hypotheses on V, then no conclusions can be drawn, and one has to begin anew with another Lyapunov function candidate. 4, Ina generat system of nonlinear equations, which do not have the structure of Hamiltonian equations of motion or some other such structure, there is no sys- tematic procedure for generating Lyapunov function candidates. ‘Thus the reader is justified in asking what the role of Lyapunov theory is today. Two remarks may be made in response.See. 5.3 Lyapunov's Direct Method 191 (1) Originally Lyapunov stability theory was advanced as a means of testing the stabil- ity status of a given system. Nowadays, however, itis increasingly being used to guarantee stability. For example, in adaptive control or PD stabilization of robots [see Example (84)], ‘one frst chooses a Lyapunov function candidate, and then chooses the adaptation law or the control law to ensure thatthe hypotheses of a particular stability theorem are satisfied. In this way, the problem of searching for a Lyapunov function is alleviated. (2) Though the various theorems given here are only sutficient conditions. itis possible to prove so-called converse theorems, which state that ifthe equilibrium has a particular property, then there ‘exists a stable Lyapunov function that would enable us to deduce this property. Usually this Lyapunov function is specified in terms of the solution trajectories of the system, and ‘can be used in perturbational analysis. Roughly speaking, the line of reasoning goes like this: Begin with a system which is easy to analyze (such a linear system: cf. the next sec- tion). Construct Lyapunov function for the same. Now see under what conditions the same Lyapunov function candidate continues to work for a modified system. We shall see several examples of such an approach in this chapter and the next. Problem 5.11 For the system of Example (19), suppose it is desired to analyze the sta~ bility of an equilibrium of the form (19, 0, 0) where x #0. Set up a new set of coordinates such that the equilibrium under study is the origin of the new set, Define a suitable Lyapunov function such thatthe stability of the equilibrium can be established by applying, ‘Theorem (14). Repeat for an equilibrium of the form (0, 0, 29) where zo #0. Problem 5.12 Analyze the circuit of Example (48) (Figure 5.7) when the capacitances are also nonlinear. Let q; denote the charge across the i-th capacitor, and suppose qj is & (possibly nonlinear) function of the voltage x;. Assume that 4,(x,)=0, and define CiGx))=04.C4¥/84,. Suppose there exists constants a, and B, such that 0< $C (x) $B, for’ on Using the total energy stored in the capacitors as a Lyapunov function candidate, analyze the stability of the equilibrium x=0. Problem 5.13 Suppose a particle of mass m is moving in a smooth potential field. To simplify the problem, suppose the motion is one-dimensional. Let x denote the position coordinate of the particle, and let (x) denote the potential energy atx. If the only force act- ing on the particle is due tothe potential, then the motion ofthe particle is described by PW=S), where the prime denotes differentiation with respect to x. Show that every local minimum ofthe function @ isa stable equilibrium. Problem §.14 Consider the autonomous differential equation =flx(O) and suppose fis a C' function such that f(0)=0. Then there exists aC! matrix-valued192 Lyapunov Stability chs function A such that [ef Lemma (2.5.17)] f(x)= AQ, VER" (2) Show that if the matrix A’(0)+A(0) is negative definite, then the origin is an exponentially stable equilibrium. More generally, show that if there exists a positive definite matrix P such that A’(0)P+P A(0) is negative definite, then the origin is an exponentially stable equilibrium. (Hint: Consider the Lyapunov function candidate V(x) = Ixil?) (b) Extend the results in (a) to global stability. Problem 5.15 Consider the differential equation FO+F DOO +H Ol=0. ‘Transform this equation into state variable form by choosing iy =). (a) Suppose the functions f and g are continuous and satisfy the following conditions forsome positive number 8: 6g (0)>0, Vae(-8, 8), F(0)20, Voe(-8,8). Show that the equilibrium is stable. (b) Suppose that the condition on fis strengthened to £(0)>0, ¥oe(-8,8) ‘Show that the equilibrium O is asymptotically stable. (©) Show that if, in addition to the conditions in (b), both f and g are continuously dif- ferentiable, then the equilibrium 0 is exponentially stable. (@ Find suitable conditions on fand g to ensure global asymptotic stability and global exponential stability. Problem $.16 Consider the system Span, $208 82 =e tad. Using the Lyapunov function candidate Vow) show that O isan unstable equilibrium.Sec.5.4 ‘Stability of Linear Systems 193 Problem 5.17 Consider the system iy ext ayaa. 2-8] eye. Using the Lyapunov function candidate V(x) =x) (01-42) and Theorem (92), show that Ois an unstable equilibrium. 4 STABILITY OF LINEARSYSTEMS In this section we study the Lyapunov stability of systems described by linear vector differential equations. The results presented here notonly enable us to obtain necessary and sufficient conditions forthe stability of linear systems, but also pave the way to deriving Lyapunov’ linearization method, which is presented in thenext section. 5.4.1 Stability and the State Transition Matrix ‘Consider a system described by the linear vector differential equation 1 R@)=AMD,120. ‘The system (1) is autonomous if AC) is constant asa function of time: otherwise it isnonau- tonomous. It is clear that 0 is always an equilibrium of the system (1). Further, 0 is an iso- lated equilibrium if A(@) is nonsingular for some r>0. The general solution of (1) is given by 2 x(1)= OL 10) xlt0), where ¢, ) is the state transition matrix associated with A(-) and is the unique solution of theequation 3 eter) eatnou.son¥21620 4 Olle. fo)=h Veo 20. With the aid of this explicit characterization of the solutions of (1), it is possible to derive some useful conditions for the stability of the equilibrium 0. Since these conditions involve the state transition matrix ©, they are not of much computational value, because in general it isimpossible to derive an analytical expression for @, Nevertheless, they are of conceptual value, enabling one to understand the mechanisms of stability and instability in linear sys- tems.194 Lyapunov Stability h.5 5 Theorem The equilibrium Vis stable ifand only iffor each to 20 itis true that 6 sup lee, £0) = (60) <2, where II, denotes the induced norm of a matrix. Proof “If” Suppose (6) istrue, and let > 0, to 20 be specified. If we define &¢€, f0) as, e/m (to), then 7 ANnlto)II <8 (Il = MBE, fo) X(Fo) IS NOU, x(Eo) 0 be any positive number, and Jet be an arbitrary positive number. Itis shown that one can choose an x(¢q) inthe ball Bs such that the resulting solution x(t) satisfies lIx(¢)ll 2¢ for some 12 ro. Select a 6 in the open interval (0, 8). Since H(t, i9)ll, is unbounded as a function oft, there exists @ 2 fp such that e 8 HOG, tot > 5 Next,select a vector vof norm one such that 9 WOE, to) Vl = HOE, fo). ‘This is possible in view of the definition of the induced matrix norm. Finally, let x(fq)=3,¥. ‘Then xe Bs. Moreover, 10 Ux(a)ll = NOE, Fo) x(Fq) ll = 5, OL, fo) Vl =8, UO, to) ll, >. Hence the equilibrium 0is unstable, i Remark: Note that, in the case of linear systems, the instability of the equilibrium 0 does indeed imply that some solution trajectories actually "blow up.” This is in contrast 10 the case of nonlinear systems, where the instability of O can be accompanied by the bound- ‘edness of all solutions, as in the Van der Pol oscillator [see Example (5.1.25)] Necessary and sufficient conditions for uniform stability are given next. 11 Theorem The equilibrium 0 is uniformly stable ifand only ifSee.5.4 Stability of Linear Systems 195 12 mo:=sup m(o)= sup sup NOL, fo), < Proof “If Suppose mg is finite; then, for any e > O and any t 20, (5.1.11) is satisfied with = e/mg ‘Only if” Suppose mtg) is unbounded as a function oft. Then at least one component ‘of (,), say the ij-th component, has the property that 13 sup 1, (fo)! isunbounded as a function of Let xp =, the elementary vector with a 1 in the j-h row and zeros elsewhere. Then (13) implies that the quantity lix(t) 1/lIxo ll = M¢®(t, £9) x 1/lxo I cannot be bounded indepen- dently of fo. Hence 0 is nota uniformly stable equilibrium. ‘The next theorem characterizes uniform asymptotic stability 14 Theorem The equilibrium 0 is (globally) uniformly asymptotically stable ifand only it 1S sup sup (1, fo)ll, <=, bk 16 (eo +1, fo) ll, > 0ast 2», uniformly in tp, Remark: The condition (16) can be expressed equivalently as follows: Foreach € > 0, there existsa 7 = 7(e) such that 17 Wto4t, fo)ll, Osuch that 19 1O(E, ll, Sm exp [-AE=Ho)}. VEZ to 20. Proof “If” Suppose (19) is satisfied. Then clearly (15) and (16) are also true, whence 0 is uniformly asymptotically stable by Theorem (14).196 Lyapunov Stability chs “Only i” Suppoxe (15) and (16) are true. Then there exist fnite constants Wand Tsuch that 20 HOU 10), SH VIZt020, A NKto +4, 19), S12, WEZT, Vio20. Inpartcular, (21) implies that 22 to+T, 19), S12, Wto20. Now, givenany to and any > ro, pick animegerk such that fg +AT$1 (2)" Obvious. “@)>(1)" Suppose (2) is true for some particular matrix Q. Then we can apply Theorem (5.3.25) with the Lyapunov function candidate V(x) =X'Px. Then V(x) =—x’Qx, and one can conclude that 0 is asymptotically stable equilibrium. By Theorem (29), this, implies that A is Hurwitz200 Lyapunov Stabi chs “(1) > GY" Suppose A is Hurwitz and let QeR"™ be positive definite but otherwise arbitrary. By Lemma (38), Equation (34) has a corresponding unique solution P given by (39). It only remains to show that P is positive definite. For this purpose, factor Q in the formM’M where M is nonsingular. Now itis claimed that P is positive definite because 43° x’Px>0,Vxz0, With Q=M’M, P becomes Bern Thus, for any xe R", 45 XPxe fv’e*MMe™xdi= [ IMeMWH320, ‘where Il-Il> denotes the Euclidean norm, Next, ifx’Px=0, then 46 Me*x=0, vr20. ‘Substituting 1 =0 in (46) gives Mx =0, which in turn implies that x= 0 since M is nonsingu- lar. Hence P is positive definite and (1) implies (3). i Remarks 1. Theorem (42) is very important in that it enables one to determine the stability status of the equilibrium @ unambiguously, in the following manner: Given AcR™, pick Qe R"™ tobe any positive definite matrix. (A logical choice is the identity matrix or some other diagonal matrix.) Attempt to solve (34) for P. IF (34) has no solution or has a nonunique solution, then 0 is not asymptotically stable. If Pis unique but not positive definite, then once again 0 is not asymptoti cally stable. On the other hand, if Pis uniquely determined and positive definite, then 0isan asymptotically stable equilibrium, 2, Theorem (42) states that if A is a Hurwitz matrix, then whenever Q is positive definite, the corresponding P given by (34) is also positive definite. It does not say that, whenever P is positive definite, the corresponding Q is positive definite ‘Thisstatementis false in general see Problem 5.19) ‘Theorem (42) shows that, i Ais Hurwitz and Q is positive definite, then the solution P ‘of 34) is positive definite. The next result shows that, under certain conditions, Pis positive definite even when Q is only positive semidefinite, 47° Lemma Suppose Ae R™ and satisfies (36). Suppose CeR™, and that‘Stability of Linear Systems 201 jean Under these conditions, the equation 49 A’P+PA=-C’C hhas a unique solution for; moreover, Pis positive definite. Proof The uniqueness of P follows from Lemma (35). To show that P is positive definite, observe that (49) is of the form (34) with Q=~C’C. Now suppose x’Px=0. Then, ‘one can repeat the reasoning of (44) to(46) with M replaced by C. This shows that 50 x'Px=02Ce*x=0, ¥120. Lettiv) =Ce™x. Thenf) as wellasall ofits derivatives are identically zero. In particular, £00) c £0) ca st 0= 7 x a car £0 aro) | ‘But from (48), this shows that: ). Hence Pis positivedefinite. ‘Theorem (42) shows that if the equilibrium 0 of the system (28) is exponentially stable, then this fact can be ascertained by choosing a quadratic Lyapunov function and applying Theorem (5.3.45). The following result, stated without proof, allows one to prove that, i the equilibrium 0 js anstable because some-eigenvaluc of A has a positive real part then this, fact can also be ascertained by choosing a quadratic Lyapunov function and applying ‘Theorem (5.3.88). "Note that he equilibrium 0 can be unstable in another way as well namely thatthe minimal polynomial ofA has & multiple zero onthe imaginary axis,202 Lyapunov Stability chs 52 Lemma Consider (34), and suppose the condition (36) is satisfied, so that (34) has a ‘unique solution for P corresponding to each Qe R"*. Under these conditions, ifQ is posi- tive definite, then P has as many negative eigenvalues as there are eigenvalues of A wich positive real pan, A proof of this lemma can be found in Taussky (1961). Note that, since (36) is assumed. tohold, it follows that A does not have any eigenvalues with ero real part; consequently all cigenvalues of A have either positive real part ora negative real part To see how Lemma (52) can be applied, suppose A satisfies the hypotheses of this Jemma, and choose the Lyapunov function candidate (32). Then Vis given by (33). Now. if ‘AisHurwitz, then Vis positive definite and Vis negative definite, and the exponential tabil- ity of@ follows by Theorem (5.3.45). On the other hand, if A has one or more eigenvalues with positive real part, then is negative definite and V assumes negative values arbitrarily close tothe origin; thus che instability ofthe origincan be deduced using Theorem (5.3.8). 5.4.3 Nonautonomous Systems In the case of linear time-varying systems described by (1), the stability status of the equilibrium 0 can be ascertained, in principle at least, by studying the state transition matrix. ‘This is detailed in Section 5.4.1. The purpose of the present subsection is three-fold: (I) to prove the existence of quadratic Lyapunov functions for uniformly asymptotically stable linear systems; 2) to present some simple sufficient conditions for stability, asymptotic sta- based on the matrix measure: (3) to present necessary and sufficient conditions forthe stability of periodic systems. ‘The Existence of Quadratic Lyapunov Funetions ‘Theorem (42) shows that ifthe equilibrium 0 ofthe system (28) is exponentially stable, then.a quadratic Lyapunov function exists for this system. A similar result is now proved for onautonomous systems, under the assumption that 0 is exponentially stable [or equivalently, uniformly asymptotically stable; see Theorem (18)]. The relevant result is Theorem (64) below. This theorem is based on two preliminary lemmas, $3 Lemma Suppose Q:R, -> R™ is continuous and bounded, and that the equilibrium Gofit)isuniformly asymprotically stable. Then, for each 20, the matrix O(N QMOE, Ndt iswell-defined; moreover, P(t) is bounded as a function oft Proof The hypothesis of uniform asymptotic stability implies that Oisin fact exponen- tially stable, by Theorem (18). Thus there exist constants m, 2 > Osuch thatSec.5.4 Stability of Linear Systems 203, 538 1O(t, Olt, Smexp[-Me—n], Vez 120. ‘The bound (55), together with the boundedness of Q(), proves the lemma. Wi 56 Lemma Suppose that, in addition to the hypotheses of Lemma (53), the following conditions also hold: (1) Qd)is symmetric and positive definite for each t 20; moreover, there exists a.constant > Osuch that 87 ox'xSx'Qinx, ¥E20, VER" (2). The matrix A() is bounded; 58 mo sup WAC)I2 Osuch that 59 Bx'xsx'P(ox, V120. YxER" Proof Let xe R”, and consider the triple product x’ P(z)x. Then, from (54), 60 xX PEIK= JO, NQH OE, Nxde= (A NQMST 1 XYdT, where s(t f ») denotes (as before) the solution of (1) evaluated at time 7, corresponding to the initial condition xattime, Now (57) and (60) together imply that 61 vPexzaf tse. ma. ‘Next, by Theorem (2.5.1), wehave { 62 Uist, 1,x)llp2 xl exp | —joscxoneal Ue | { i. U : }204 Lyapunov Stability chs 2 linlly expl-mo(t—0), by (58). ‘Substituting from (62) into (61) gives ox'x 2m 63 x'POxza | x’xexp[-2mo(t—H]dt= ‘The inequality (59) now follows by taking B= a/2mo. ll 64 Theorem Suppose Q(-) and A(-) satisfy the hypotheses of Lemmas (53) and (56). Then for each function Q() satisfying the hypotheses, the function 65 Vi. =x PU)X isa Lyapunov function in the sense of Theorem (5.3.45) for establishing the exponential sta-~ bility ofthe equilibrium 0. Proof With V(t, x) defined as above, we have 66 Ve, x)=¥' [PO +A’(OPU) + PINAL. tis easy to verify by differentiating ($4) that 0 ACP) -PEOAM- QO, 8 Ve. x)=-x'Qn)x. ‘Thusthe functions Vand V satisfy all the conditions of Theorem (5.3.45). Conditions Based on the Matrix Measure Nextwe present some conditions for stability and instability based onthe matrix meas- ure. The following result proves useful fortis purpose. 69 Lemma With regard o the system (1), the following inequalities hold: 70 exp} f-nl-AG@)]dt} < Noe, onsen fare : where Ill isany normonR* and Ui the corresponding matrix measure on R™.Sec.5.4 Stability of Linear Systems 205 ‘The proof is immediate from Theorem (2.5.1). Many simple sufficient conditions for various forms of stability are ready conse- quences of Lemma (69). The proofs are straight-forward applications of results Section 5.4.1 andre leas exercises. 71 Lemma The equilibrium 0 of (1) is stable if, for each to, there exists a finite constant m(to) such that 2 JulAcoldesm(to), ¥e2 1020. The equilibrium is uniformly stable ifthere exists a fniteconstantma such that 73 Jalal dt smo, ¥i2t020. 74 Lemma The equilibrium of(1) is asymptotically stable if 15 f wlACdt>—eeass so, Vin20. The equilibrium is uniformly asymptoicaly stable ifthe convergence in (75) is uniform in toi if. foreverym > Othereexists.a T such that 76 J WIA@]de<—m, Vi2T, Wig20. 77 Lemma The equilibrium is unstable ifthere exists a time tg 20 such that 78 ful-AC@)}dt >was oe, Remarks 1. Asshown in Section 2.2, the measure of a matrix is strongly dependent on the vec tor norm on R® that is used to define it. fn Lemmas (71), (74), and (77), the con- clusions follow if the indicated conditions are satisfied for some norm on R". ‘Thus there isa great deal of flexibility in applying these lemmas. 2. The three lemmas provide only sufficient conditions and are by no means neces- sary. {But in this connection, see Vidyasagar (1978a).} However, they do have the advantage that one doesnot have to compute the state transition matrix.206 Lyapunov Stability chs 3. Lemma (77) is of rather dubious value, since its hypothesis actually assures com- plete instability of the system (1); in other words, every nontrivial trajectory of (1) blows up.” Periodic Systems ‘This subsection concludes with adiscussion of periodic systems. ‘Suppose that the matrix A(®) in (1) is periodic. In this case, we know by Theorems (5.1.43) and (5.1.49) that the stability of the equilibrium 0 is equivalent to its uniform stabil- ity, and that the asymptotic stability ofthe equilibrium 0 is equivalent to its uniform asymp- ‘otic stability, Theorem (89) below shows that inthe case of linear periodic systems, further simplifications are possible, 79 Lemma Suppose the matrix A() in (1) is periodic, and select a constant T >0 such that 80 A(E+T)=A(D), We20. Then the corresponding state transition matrix Xt, £9) has the form 81 Ot, £0) = "PC, to) exp [MU—to)]. where M is aconstant matrix, and'¥ is periodic inthe sense that 82 WEFT, 19) = HC, to), WE20. Proof Define 83 R=O(tg+T, t0), and choose a matrix M such that 84 Reexp(M7). Thisis possible since Ris nonsingular. Now itisclaimed that the matrix defined by 85 Ws, £9) = OC, fo)expl-MU~F)} satisfies (82). To see this, proceed as follows: 86 YCHT, fo) =OU+T, toexpl-MU+T-t0)] = O47, fo +T)Hltg+T, fo)exp(-MT)exp [-M(—to)}. However, by the periodicity of AC), we haveSee, 5.4 Stability of Linear Systems 207 87 @E+T, t+ T)= OE, 10). ‘Thus (86) simplifies to BB HEFT, 40) = HE, fo) epI-MUe—to)]= Kl e. fo). This establishes (82). i ‘Once the representation (81) for the state transition matrix is obtained, the results of Section 5.4.1 can be used to obtain necessary and sufficient conditions for stability. For this purpose, note that Ais an eigenvalue of Mif and only if exp (AT) isan eigenvalue of (T, 0). 89 Theorem Consider the system (1) and suppose the matrix AC) is periodic. Then the equilibrium 0 of (1) is uniformly asymptotically stable if and only if all eigenvalues of (7. 0) have magnitude less than one. The equilibrium is uniformly stable fand ontyifall eigenvalues of ®(7, 0) have magnitude no larger than one, and all eigenvalues of ®(T, 0) witha magnitude of one are simple zeros ofthe minimal polynomial of ®(T, 0). ‘The proofis elementary ands left as an exercise. 90 Example The purpose of this example is to show that the stability of a nonauto- ‘nomous system cannot be deduced by studying only the eigenvalues ofthe matrix A(j) for each fixed t. Consider the periodic system (1) with Aw= -I+acos?s 1=asinrcosr , thesystemisunsable. Tose hs nate tha the peiod Tequalinthiscase, and tha sono su ‘The eigenvalues of this matrix are obviously e@-"* and e™®, Ifa > 1, the first eigenvalue thas a magnitude greater han one, and the equilibrium O is unstable by Theorem (89). i208 Lyapunov Stability ch. In Section 5.82 itis shown that, for so-called slowly varying systems, itis possible to deduce the stability of the nonautonomous system by studying only the "frozen” systems. Problem 5.18 Complete the proof of Theorem (14). Probiem 5.19 Construct an example of a Hurwitz matrix A and a positive definite matrix Psuch that A’P + PA isnot negative definite Problem 5.20 Consider an RLC network that does not contain any capacitive loops or inductive cutsets. Such a network can be described in state variable form by choosing the capacitor voltages and the inductor currents asthe states. Specifically, etx denote the vec- tor of capacitor voltages, and etx, denote the vector of inductor cutsets. Then the system: equations are ofthe form Xe co a oL where Cis the diagonal positive definite matrix of capacitor values; Lis the positive definite matrix of inductor values, including mutual inductances; Ri, are the matrices arising from the resistive subnetwork, with Ry, and Ray symmetric, and R,, =—R’ if the resistive net- work is reciprocal. Using the total energy stored in the capacitors and the inductors as a Lyapunov function candidate, show that the equilibrium 0 is stable if both Rj, and Ra» are nonnegative definite, and is asymptotically stable if both R,, and Rzp are positive definite, Ry Ria] [xc] Ra Ry |x|: Problem 5.21 Using the results of Section 5.4.3, determine whether 0 isa uniformly stable or uniformly asymptotically stable equilibrium for the system X(t) =A(e) x(t). for each ofthe situations below: 2esin® 1] (a) AM=| cosy 1 [ar &) A@=| FP -P 0 2 5 4-1) Problem 5.22 The equilibrium 0 of the system (5.1.1)is said to be Boundedif,forevery 5 >Oandevery 1920, there existsane=e(6, such that Ix Ist, t5, 9) O there exists ane =(8) such that Ux II <8, f20= Iistt, fo, Xo) Il <€, VF toSec. 5.4 Stability of Linear Systems 209 (a) Show that a linear system is bounded if and only if itis stable, and that itis uni- formly bounded if and only ifit is uniformly stable. (b) Construct examples of nonlinear systems where the equilibrium 0 is stable but not bounded, and where the equilibrium 0 is bounded but not stable. Problem 5.23 Generalize Theorem (18) to a class of nonlinear systems. Specifically, suppose there exists a number d > 0 such thatthe solution trajectories ofthe system (5.1.1) satisfy the bound, Uis(t, £0, Xo)! Sp lx HOU =o), WAGE By. VEZ 1920, ‘where [1s some finite constant and o(-) isa function of class L. Show that the equilibrium 0 isexponentially stable. ~ Problem 5.24 Given a finite collection of nxn matrices A, hull Sas + Au define their convex ‘ A DAAsA20Wi, 52, a (a) Suppose there exists a positive definite matrix P such that A/P+PA, is negative definite foreach ibetween 1 andk. Show that every matrix inthe set Sis Hurwitz (b)Consider the differential equation HD =AOX(), where AMES W120, ‘Show that 0 isan exponentially stable equilibrium ofthis system. 5.5 LYAPUNOV’S LINEARIZATION METHOD In this section, the results of the preceding two sections are combined to obtain one of the most useful results in Lyapunov stability theory, namely the linearization method. The great value ofthis method lies in dhe fact that, under certain conditions, it enables one to draw conclusions about nonlinear system by studying the behavior ofa linear system. We begin by defining precisely the concept of linearizing a nonlinear system around an equilibrium. Consider first the autonomous system 1 R)=fIxX(). Suppose f(0)=0, 50 that O isan equilibrium ofthe system (2), and suppose also that fi con- tinuously differentiable, Define210 Lyapunov Stability chs i.e, let A denote the Jacobian matrix of f evaluated at x=0. By the definition of the Jaco- bian, itfollows thatif we define 3 f\00:=f)-Ax, then : goon “where, to be specific, all norms are taken as Euclidean norms. Alternatively, one can think of $ fO)=Ax+f,00 as the Taylor's series expansion of f() around the point x= 0 [recall that f(0) =0]. With this notation, the system 6 x)=Ax) is referred toas the linearization or the linearized system of (1) around the equilibrium 0. ‘The development for nonautonomous systems is similar but for an additional technical requirement. Given the nonautonomous sysiem 7 k=M xO), 8 fG,0) and thatfisaC* function. Define [*s 4 10 ,(,)=f0, )- Ax. 9 AW) ‘Then, by the definition of the Jacobian, i follows that for each fixed t 20, itis true thatSec.5.5 Lyapunov’s Linearization Method 2 Goll xt However, it may or may not be true that wht 2AM Eo In other words, the convergence in (11) may or may not be uniform in t. Provided (12) holds, the system 13° H)=ACztt) is called the linearization or linearized system of (7) around the equilibrium, 14 Example Considerthe system 542-2 (+03, 409 =4)() 1200. Inthiscase,f()is C! and -10 However, the remainderterm f(t, x)isgiven by f(.x)=[04 OF. Hence the uniformity condition (12) does nothold. Accordingly, the system f=-2. 22200-2210) is nota linearization of the original system. Ml ‘Theorem (15) is the main stability theorem of the linearization method. Since there is nothing to be gained by assuming that the system under study is autonomous, the resul stated for a general nonautonomous system. 15 Theorem Consider the system (7). Suppose that (8) holds and that fi) is continu- ‘ously differentiable. Define A(t). (GX) as in (9), (10), respectively, and assume that (i) (12) holds; and (i) AG) is bounded. Under these conditions, if is an exponentially stable equilibrium of the linear systert 16 AQ, then itisalso an exponentially stable equilibrium of the system (7).212 Lyapunov Stability chs Proof Since A()is bounded and the equilibrium 0 is uniformly asymptotically stable, itfollows from Lemma (5.4.56) that the matrix 17 P=] ONO, Ndt js well-defined forall 20; moreover, there exist constants > Osuch that 1B ox'xsxP(OxSBx’x, Vxe RY, VE20. Hence the function 19 V@H=xPOX is a decrescent pdf, and is thus a suitable Lyapunov function candidate for applying. ‘Theorem (5.3.45). Calculating V for the system (7) gives 20 Vir, xy=x’ Pay + er, x) P(r) x+x’ PCO), 0) =x’ [PO)+A PU) + PC) ACI X+ 2X PlO)N (t, 0). However, from (17) itcan be easily shown that 21 PU) +A) PL) +P) ACD: Hence 22 Vit.x)=-x’x 2x’ PUD». Inview of (12), one can picka number > Oandap <0.5 suchthat 23 “Itt, vist Ixil, ¥r20, VxeB,. Then (23) and (18) together imply that a (2x PEN wis Bw .WF20, Ve B, Therefore 25 Vit, x)S—(1~29)x'x, W120, Vae,. This shows that —Vis an Ipdf. Thus all the hypotheses of Theorem (5.3.45) are satisfied, and ‘we conclude that isan exponentially stable equilibrium. iSee. 5.5 Lyapunov’s Linearization Method 213 26 Corollary Consider the autonomous system (1). Suppose that £(0)=0, that fis con- tinuously differentiable, and define & as in (2). Under these conditions, 0 is an exponen- tially stable equilibrium of (1) fall eigenvalues of A have negative real parts. In the instability counterpart to Theorem (15), itis assumed that the linearized system is autonomous, even if the original nonlinear system is not. 27 Theorem Consider the system (7). Suppose that (8) holds, that € is C', and suppose inaddition that 28 ao=[ = Ap (aconstant matrix), Vr20, attr.» ex «and that (12) holds. Under these conditions, the equilibrium 0 is unstable if Ay has atleast one eigenvalue witha positive real part Proof The proof is given only for the case where Ag satisfies the condition (5.4.36). ‘The proof of the general case can be obtained from the one given below by using continuity arguments, Since Ag is assumed to satisfy (6.4.36) and has atleast one eigenvalue with positive real part, itfollows from Lemma (5.4.52) that the equation 29 AGP+PAy=I hhas a unique solution for P and that this matrix P has at least one positive eigenvalue. Now, bby arguments entirely analogous to those used in the proof of Theorem (15), itcan be shown thatif we choose 30 V(x)=x'Px, then V assumes positive values arbitrarily close to 0, and V is an Ipdf. Hence, by Theorem (5.3.88), it follows that O is an unstable equilibrium of the system (7). The details are left as anexercise. I Remarks: Theorems (15) and (27) are very useful because they enable one to draw conclusions about the stability status of the equilibrium 0 of a given nonlinear system by examining a linear system. The advantages of these results are self-evident. Some of the limitations of these results are the following: (i) The conclusions based on linearization are purely local in nature; to study global asymptotic stability, itis still necessary to resort 10 Lyapunov's direct method. (ii) In the case where the linearized system is 2utonomous, if some eigenvalues of A have zero real parts and the remainder have negative real pars, then linearization techniques are inconclusive, because this case falls outside the scope of both Corollary (26) and Theorem (27). In such a case, the stability of the equilibrium is deter- ‘mined by the higher order terms that are neglected inthe linearization process. [In this con- nection, see Theorem (5.8.1). (ii) In the case where the linearized system is nonauto- nomous, if the equilibrium 0 is asymptotically stable but not uniformly asymptotically214 Lyapunov Stability chs stable, then linearization is once again inconclusive. Itean be shown by means of examples that the assumption of uniform asymptotic stability in Theorem (15) is indispensable, Let us now return to the autonomous system (1). Suppose f(0)=0, and define Af) as in(2) and (3) respectively. Finally, suppose A is a Hurwitz matrix. Then Corollary (26) tells, us that @is an asymptotically stable equilibrium. Now let Q be an arbitrary positive definite matrix, and solve the corresponding Lyapunov matrix equation Q ‘Then P is also positive definite. Moreover, it can be seen from the proof of Theorem (15) that, foreach Q, the corresponding quadratic form 3 A'P+PA 32 VQp=xPx is a suitable Lyapunov function for applying Theorem (5.3.45). However, different Lyapunov functions will give rise to different estimates for the domain of attraction of 0. The question an thus be asked: What isthe "best" choice for Q? Let us restate the question in another form. Define V by (32). Then, along the trajec- tories of the system (1), we have 33 V(x)=x’(A’P + PA)x + 2x’ Pf (x). =—x'Qx + 2x’ PE (x). Now B= X'QKS—Apin( QIU, 351 xP E(x)! SAmpan(P) Hx, Ox) I Hence 36 V6) SP QUI? +2 gg (PATE = Ux [2 Aman CP) HE) (x) ll — Agnin(Q) Hx]. Nowletuschoose r>Osuch that WCDI Arial Q) Ix) ~ 2Ama(P)* Vxe B, ‘Then V(x) <0 whenever xe B, and x#0. By Lemma (5.3.40) and the discussion preceding it, every bounded level set of V.contained in B, is also contained in the domain of attraction DO). Now (37) makes itclear that the larger the ratio Arnin(QVAmax(P), the larger the possi- ble choice of r. Hence the “best” choice of Qs one that maximizes the ratioSec.5.5 Lyapunov’s Linearization Method 2s where P of course satisfies (31). Clearly HC) is not affected by scaling Q, ic. (aQ)=11(Q) Va > 0. Hence one can pose the question at hand as follows: Among all posi tive definite matrices Q with Apin(Q)=1, which one results in the smailest value for Aoyux(P)? The answer turns out tobe: the identity matrix. 39 Lemma Suppose Ais Hurwitz, and let M be the solutionof 40 A’M+MA= Suppose Qis positive definite, hsn(Q)= 1,and let P satisfy (31). Then 41 Arpax(M)Shmon(P). Proof Since Anig(Q)= 1, the matrix Q~/ has all nonnegative eigenvalues, and is therefore nonnegative definite, Subtracting (40) from (31) gives 42 A(@P-M)+(P-M)A=-(Q-. From Lemma (54.38), the solution of (42)is given by 43 P—M=[f exp(A’1)(Q—Dexp (Ande Hence P—M isnonnegative definite, which implies (41). i ‘The preceding discussion shows that, in some sense, the "best" choice for Q is the iden- tity matrix. 44 Example Consider the system Jy sca, try tap ieee td Thelinearztion of thisystemaround is a] pan © tal [10 Let A denote the matrix in (45). Then the eigenvalues of A are (-1+j¥3\2. Hence, by Corollary (26), is an asymptotically table equilibrium, 2 “To obtain an estimate of the domain of attraction, we first solve the equation216 Lyapunov Stability ch.5 A'P+PA=-/ for P. Itcan be easily verified that [ i -12 [-¥2 32] ‘Hence B in (18) can be chosen as the largest singular value of P, which is (5+V5 y/4,or approximately 1.81. To satisfy 23), we mustchoose r > 0 insucha way that 46 f(xy < © xt, vxeB,, 8 where p <0.5is some number. Now MOON _ xix} +49)'7 Wx GP axdy? 2h S thx, Hence, to satisfy (46), we can choose r as close as possible to 1/8, or approximately 0.54, Hence every level set in the ball of radius 0.54 isin the domain of attraction, 47 Example Consider the Van der Pol oscillator described by Sy igeig wl —aDa—ay. The linearization of this system around the equilibrium 0is i) for} fa in| 7|-1H} Jen IfU>0,then both eigenvalues of A have positive rea part. Hence, by Theorem (27), is an unstable equilibrium. Tn fact, since all eigenvalues of A have positive real parts, one can show, by applying Corollary (26) and reversing the direction of time, that 0 is a completely unstable equilibrium. 48 Example Consider again the spinning object of Example (5.3.105), and focus aten- tion on an equilibrium ofthe form (0, yo, 0) where yo #0. If we define y,=y yo, then the equations of motion can be written as Feayye tay oz, 5, =— biz, Es00y, +009. ‘The linearization of this system around the origin (in the new coordinates!) is obtained by neglecting all the higher order terms, and isSee. 5.5 Lyapunov’s Linearization Method 217 0 ayo] | x l=] 000 ]}y, 2} [ooo 0} Le ‘The eigenvalues of the A matrix above are 0, +Vac | yo!. If | yo! #0, then A has a positive real eigenvalue, Hence, by Theorem (27), the equilibrium 0, or (0, ¥9, 0) in the original coordinates, is unstable. 49 Application (Feedback Stabilization of Nonlinear Control Systems) Given an autonomous system described by 50 X()=fX(N), MO), where f:1R" xR" — R" the objective is to find a feedback control law of the form. st insuch a way that the equilibrium 0 of the resulting closed-loopsystem 582 X= FLX, gix(0)}} is asymptotically stable. A solution to this problem is given by the next result, which is a direct consequence of Corollary (26). 53. Theorem Consider the autonomous system (50) where £:R"XIR" > R". Suppose f isC! and hat{(0,0)=0, and define { ed ee Suppose there exists a matrix Ke R"™ such that all eigenvalues of A-BK have negative real parts. Under these conditions, ifwe apply the control law 55 us Kxtt) {n(50), then is an asymptotically stable equilibrium ofthe resulting system fIx(),~ Kx). Proof Define the function h: R* —> R" by 56 i 37 hx) =f(x,- Kyo. ‘Then the closed-loop system (56) can be represented as218 58 x(1)=hix(). Next, observe that which isHurwitz by assumption. The desired conclusion now follows from Corollary (26). 60 Application (Singularly Perturbed Systems) In Section 4.3 some results are derived concerning the stability of singularly perturbed linear systems. In this application, Lyapunov's indirect method is combined with these earlier results of Section 4.3 to derive some results for singularly perturbed nonlinear systems. The proof of Theorem (61) is mitted, because itfollows readily from earlier results. 61 Theorem Consider the system 62 RH) =NXY. YO). HO =ala), YI, where f: R" XR" — R®, g: RXR" — R" are continuously differentiable and satisfy © 10.000.40.540 bee #1) aval] canal a ) af] nel]. and suppose Ax. is nonsingular. Under these conditions, there exists a C' function bh: RY R" such that, insome neighborhood of 0,0), 6 ven Y ys fee isthe unique solution of 66 g(x, y)=0. Moreover, 1. Ufboth Ang and Ay, ~A,2Ai4An are Hurwitz then there is an Ey such that (0,0) isan asymptotically stable equilibrium ofthe system (62) whenever 0 <€ <9Sec. 5.5 LLyapunov’s Linearization Method 219 2. Ifatleastone eigenvalue of either Azz or Ayy~ArzAH4An, has a positive real Dart, then there i ant such that (0,0) is an unstable equilibrium of he system (62) whenever 0<€ R is continuously differentiable. Such systems are studied further in Section 5.6. Suppose that the equilibrium 0 of this system is asymptotically stable whenever 6 is set equal toa linear function of the form (0)=ko, ke 10, (a) Using Corollary (26), show that the equilibrium 0 continues tobe an asymptotically stable equilibrium if @() is any C' function whose derivative lies inthe interval (0, yin some neighborhood ofthe origin, (b) Generalize the results of part (a) to time-varying systems, using Theorem (15). 5.6 THELUR'E PROBLEM In this section, we study the stability of an important class of control systems, namely feedback systems whose forward path contains a linea time-invariant subsystem and whose feedback path contains a memoryless (possibly time-varying) nonlinearity. The type of ys- temis shown in Figure 5.15. The forward-path subsystemis described by Bt = Ax(t) + Bu), yi =Cx()+Dute), while the feedback subsystem is described by220 Lyapunov Stability chs 2 x)=Olr, yO). Of course, the feedback interconnection is described by 3 wt: at), In the above, it is assumed that x(2)eR", while u(r), y(1},20¢)€R" with m R", and a, BER with a R" belonging tothe sector{a, b In contrast with the systems studied thus far in this chapter, we are concerned at present not with a particular system but an entire family of systems, since © can be any nonlinearity in the sector [a,b]. The idea is that no detailed information about the nonlinearity is assumed — all that is known is that ® satisfies (10). For this reason, the problem under22 Lyapunov Stability ch. study is referred to as an absolute stability problem. Itis also known as the Lur'e problem, after he Russian scientist A.1. Lure ‘The Aizerman and Kalman Conjectures Suppose it were possible to deduce the stability of a family of nonlinear time-varying systems by examining only all the linear time-invariant systems within that family. Then the absolute stability problem would be very easy to solve. With this in mind, in 1949 the Russian mathematician M. A. Aizerman made a conjecture regarding the absolute stability problem in the case of strictly proper, single-input, single-output systems (ie., D=0 and ‘m=1), In thiscase, the only linear time-invariant maps 6 satisfying (the scalar version of) (10) are M1 60, =k, Vey, hela, 6. Aizerman's conjecture was that if the system (1) ~ (3) (with D=0 and m =1) is globally asymptotically stable forall linear time-invariant maps 6 of the form (11) as the constant k varies over the interval [a,b ],then the same is true forall time-invariant nonlinear elements @ inthe sector fa, 6. Unfortunately, while itis a tempting conjecture, itis false in general [But a modified version of itis true; see Theorem (6.6.126).] In 1957,R. E. Kalman made another conjecture, Suppose 9: R-> R is a memoryless time-invariant nonlinearity, and is continuously differentiable. Then $ is said to belong to the incremental sector (a, 6 if (0)=0, and in addition, 12 as¢y)sb, WER Kalman’s conjecture was that if the system (1) ~ (3) (with D=0 and m=1) is globally asymptotically stable forall @ ofthe form (11), then the same is true for all time-invariant nonlinear elements @ belonging tothe incremental sector [, b | Itis easy to see that if: R-»R belongs to the incremental sector a,b], then it also belongs to the sector {a, b this isa ready consequence of the mean-value theorem. But the converse is not true in general, Thus the family of nonlinearities @ covered by Kalman’s conjecture is strictly smaller thar that covered by Aizerman's conjecture. So Kalman’s con- jecture "had a better chance" of being true than Aizerman’s conjecture. Moreover, using Lyapunov’ linearization method [Corollary (5.5.26)], it can be shown that the following, statement is true (see Problem 5.26): Ifthe system (1) ~ (3) is globally asymptotically stable forall of the form (11), or equivalently, if A~BRC is a Hurwitz matrix forall ke[a, 6}, then x=0 is an asymptotically stable equilibrium of the system (1) ~ (3) for all time- invariant 6 belonging to the incremental sector (a, b]. Thus the essence of Kalman’s con- Jecture lies in replacing “asymptotically stable" by “globally asymptotically stable.” Nevertheless, Kalman’s conjecture is also false in general,Sec.5.6 ‘The Lur’e Problem 23 5.6.2 The Circle Criterion In this subsection, we present a sufficient condition for absolute stability, known as the circle criterion. The contents of this subsection as well as the next depend in an essential way on the following result of independent interest, known as the Kalman-Yacubovitch Iemma, 13. Theorem (Kalman-Yacubovitch) Consider the system (J), where x(Q)€R", and Y¥(0), wre R" with m 0, where * denotes the conjugate transpose, and nin denotes the smallest eigenvalue ofa Her- mmitian matrix. Under these conditions, there exist a symmetric positive definite matrix PeR™, matrices Qe R™", We R"™", andane > Osuch that 15 A’P+PA P-QQ, 16 BIP+W'Q=C, 1 WW=D4+D) The proof of Theorem (13) is given in Appendix B. Note that a transfer matrix H() satisfying (14) is said tobe strictly positive real. Now the simplest version of the circle criterion, called the passivity theorem, is presented. It will be seen later that more general versions of the criterion can be derived using a technique known as loop transformation, 18 Theorem (Passivity) Consider the system (1)~ (3). and suppose (i) the matrix A is Hurwitz, (it) the pair (A, B) is controllable, (iti) the pair (C, A) is observable, (iv) HC) satisfies (14), and(v) ® belongs to the sector (0, ),i.e., @Xt, 0)=0V!20, and 19 y' OQ, y)20, 120, VyeR™ Under hese conditions the system (1) ~(3)is globally exponentially stable. Proof Conditions (i) to (iv) of the hypotheses imply that Theorem (13) can be applied. (Choose P, Q, and W such that(15)~ (I7)hold. Define the Lyapunov function candidate 28 V@y=x'Px224 Lyapunov Stability chs =[Ax-BOy' Px+/P[Ax—BO] =x’ (A’P+PA)x-@'B'Px—’PBO, after substituting for u, and letting ® denote Of, ¥(1)]. Now, from (16) it follows that 2 B =C-W'Q. Hence 23° @'B'PK=0'Cx-O'WOx = (y-Du)-O WQx =O (y+D®)-0'WOx, Next, substituting from (23) into (21) gives 2A Vax’ (A'P+PA)x—@ (D+D')@-O'WQx-x QV WO-O'y-y¥'. Now substitute from (15) and (17) into (24), and observe that ®’y 2.0. This leads to 28 VS-ex'Px-¥’'Q'Qx-O' WWO-O'W'Ox-x WO ex Px [Qu + WO’ [Qx+ WO] <-exvPx, ‘The global exponential stability of the system now follows from Theorem (5.3.62). i ‘Theorem (18) only applies to the case where @ belongs to the sector [0, =). However, using a technique known as "loop transformation,” the theorem can be modified to cover the case where © belongs to a general sector [a, 6. The idea is that, if @ belongs to the sector [a, b}, then © af belongs to the sector (0, b~ a. Consequently, for each 8 > 0, the non- linearity $s qu 26 &,=(@—al){1+[1Ab-a +8(@-aly)"" belongs to the sector (0, ©). See Figure 5.17 for an interpretation of the nonlinearity ®,. In the process of modifying the feedback element from to ®,, the forward path element gets transformed from H()10 27H (s)=H(s)(+aH(s)"' + [146 -a +8)) 1. ‘This can be stated formally as follows:Sec. 5.6 ‘The Lur’e Problem 225 Tie ap oF Tee FIT Fig.5.17 28 Corollary Consider the system (1) ~ (3). Suppose (i) the pair (A, B) is controllable, ‘and the pair (C, A)is observable, (ii) ® belongs othe sector (a, Define 29 HAs) =Hts)/ +aH(s))"!. Suppose 1 ba 30 inf Amin (Ha) + Hy *GOv/2) + 20 and all poles of H,(-) have negative real parts. Under these conditions, the system (1) ~(3) is exponentially stable. Proof The idea is to show that the transformed system satisfies the hypotheses of ‘Theorem (18). Since (30) holds, itfollows that 31 inf Apinl Hao) +H, *G0)) + >0, 2 b-atd for sufficiently small 8 > 0. Now define H, by (27). Then (31) is equivalent to226 Lyapunov Stability Ch.5 32 inf Ayinl Hi) +H, *(j0)] > 0. Thus H,(> satisfies hypothesis (iv) of Theorem (18). As mentioned above, ®, satisfies hypothesis (v) of Theorem (18). As for the remaining conditions, a routine calculation shows that a realization for H,(-)is given by 33. A=A-aBU+aD)"C, B,=BU+aDy', C.=U+aDy'C, D,=DU+aDy" +[1Ab ~a +8)] Moreover, its easy to show that the pair (A,, B,)is controllable, and that the pair (C,, A,)is observable, which are respectively hypotheses (ii) and (iii) of Theorem (18). Thus, if all poles of H,() have negative real parts, then A, is Hurwitz, which is the last remaining hypothesis needed to apply Theorem (18). The desired conclusion now follows from ‘Theorem (18). i Corollary (28) applies equally well to multi-input, multi-output systems (m > 1) and to single-input, single-output systems (m = 1). However, inthe latter case, itis possible to give an elegant graphical interpretation of the condition (30). This leads to a result commonly known as thecitcle criterion, Toestablish the result, itis useful to make the following obser- vation. Suppose z=x + jy isa complex number, and a, be R witha < band a#0and b #0. Then ifba <0. ‘This can be established by high school algebra. In fact, both statements are equivalent to 36 bal? +y2)+baye+1>0. Let D(a, b) denote the closed disk in the complex plane centered at (b +a)/2ba and with radius (6—a)/21ba |. Then the observation is that (34) holds if and only if the complex ‘number z lies outside the disk D(a, b) in case ba >0, and lies in the interior of the disk D(a, by incase ba <0.Sec. 5.6 ‘The Lur'e Problem 227 In Theorem (37) below, reference is made to the (Nyquist) plot of A(je). This is the plot of f ja) as @ increases from —eo to «If h() has a pole on the a-axis, the plot is ‘obtained by "indenting" the jo-axis in such a way that the pole lies to the eft ofthe indented Jio-axis; see Figure 5.18 No Fig.5.18 37 Theorem (Circle Criterion) Consider the system (1) ~(3), and suppose? m= 1, (ii) the quadruplet (A,b.e,d) is a minimal realization of h(), and (ii) @ belongs to the sector [a, 6. Define thedisk D(a, b)as above. Under these conditions, the system (1)~(3)is glo- bbally exponentially stable fone of the following conditions, as appropriate, holds. Case (i) 0 0. The hypotheses on h(:) imply that (34) holds with z replaced by (ja). Moreover, since h(i) is bounded away from the disk D(a, 6), it fol- lowsthat 2 Since the single-input, single-output case is being considered, matrices are replaced by tow oF column vectors, or calar, a appropriate28 Lyapunov Stability chs inf Re Uo) 9 atk Teanuar which is (30) specialized to the scalar case. Next, the encirclement condition implies that the plot of h Gia) encircles the point -I/aexactly v times in the counter-clockwise direction. Hence by the well-known Nyquist stability criterion [see e.g., Theorem (6.5.35) for a very general version}, it follows that all poles of fg(-) have negative real parts. Since all hypotheses of Corollary (28) hold, the desired conclusion follows. Case (ii): In this case h, = hand (38) isthe scalar version of (30). Since A is Hurwitz, the desired conclusion follows from Corollary (28). Case (it): In this case ba <0. Hence the fact that # (je) lies inthe interior ofthe disk ‘D(a, b) implies that (4) holds with z replaced by hia). Moreover, since h (je) is bounded away from the circumference of D(a, b), it follows that (39) holds. Finally, since AisHurwitz, the desired conclusion follows from Corollary (28) Case iv}: Obvious. i An appealing aspect ofthe circle criterion is its geometric nature, which is reminiscent of the Nyquist criterion. Indeed, if ba ->0, then the “critical disk” D(a, 6) in Case (i) shrinks to the "critical point” ~1/a of the Nyquist criterion; in this case the circle criterion reduces to the sufficiency part of the Nyquist criterion. On the other hand, the circle cri- terion is applicable to time-varying and/or nonlinear systems, whereas the Nyquist criterion is only applicable to linear time-invariant systems. Another appealing feature of the circle criterion is that it depends only on the transfer function h() of the forward path, and not on the particular realization of h(-). This means that if We think of the forward path element asa "black box,” then in order to apply the circle criterion itis only necessary to determine the transfer function of this black box, which can be achieved through relatively straight-forward experiments; it is nor necessary to construct arealization of A(). 40 Example As an illustration ofthe circle criterion, suppose the transfer function of the forward-path subsystem in Figure,5.15 is (= Fyne + 2K + 346 #200) “The plotofh Ga) shown in Figure 5.19, witha porion oft shown in enlargement in Fig- ures 0, Suppose first that the feedback nonlinear element 6 belongs tothe sector [=5/3, 51 “The corresponding isk D(a, 6) passes through the points—0.2 and 0.6, as shown in Figure 5.19, Moreover, the plot of h les inside D(a, b). Hence, by Case (i) of the cirelecr-See.5.6 ‘The Lur’e Problem 229 02 ~06! 2 ° 02 Oa 08 Fig. 5.19 ‘erion, we conclude that the feedback system is globally exponentially stable for all @ belonging to the sector[~5/3, 5} ‘Now suppose ¢ belongs to the sector (0, 10}. In this case, (38) is Hence, by Case (i) ofthe circle criterion, we can conclude that the feedback system is glo- bally exponentially stable forall 6 belonging tothe sector {0, 10} At this stage, one might be tempted to combine the above two conclusions, and state that the feedback system is globally exponentially stable for all @ belonging to the sector [-54, 10}, onthe basis that [-58, 5110, 10]=[-54, 10} But the statement does not follow. Let N(a, 6) denote the set of nonlinear elements belong- ing to the sector (a, 6). Then one can see that N(-5B, SSIN(O, 10) #N(-5A, 10). ‘Asa final application, suppose @ belongs tothe sector (4000, 7000]. The correspond- ing disk D(a, b) is shown in Figure 5.20, Now itfollows from Case (i) ofthe circle criterion that the feedback system is globally exponentially stable forall @ belonging to the sector £4000, 7000}230 Lyapunov Stability ch.s Fig. 5.20 5.6.3 The Popov Criterion In this section, we derive another criterion for absolute stability, known as the Popov criterion, after the Roumanian scientist V.M. Popov. Unlike the circle criterion, the Popov criterion is applicable only to autonomous systems. The class of systems studied by Popov is described by a = Ax+bu, where xe R", 6, u, are scalars; and A, b, ¢,d have commensurate dimensions. The non- linear element 6: R-> R isa time-invariant nonlinearity belonging to the open sector 0, =). ‘This means that 43 9(0)=0,¥ 60) >0. ¥y 40.Sec.5.6 ‘The Lur’e Problem Zi Notice te differences between the system descriptions (1) (3) and (41) ~ (42). Inthe latter system, there isa pole atthe origin, and there is no throughput from the input to the output Moreover, the nonlinearity belongs to an open sector, and not a closed sector as ia the former system, The system description (41) can be rewritten as fi-(ea-f1- x vate ally Hence its transfer function is 45 h(s)= +e(s1-Ay'b. 46 Theorem (Popov Criterion) Consider the system (41) ~ (42) and suppose (i) the matrix A is Hurwitz, (i) the pair (A. b) is controtlable, (il) the pair (c, A)is observable, (iv) 4 > 0, and (v) the nonlinear element 6 belongs to the sector (0,2). Under these conditions, the system (41) ~ (42) is globally asymptotically stable if there exits a number r > such that 47 inf Re{(1 + jar) hV@)]>0. Remarks The quantity 1 + jax is called a "multiplier." Note that if r=0, then (47) reduces to (38) with b =». la this sense, and only inthis sense, the Popov criterion isa gen- eralization of the circle criterion. One should not carry the comparison too far, since the two criteria apply to distinct classes of systems. Case (i) of the cele criterion applies to open- loop stabie systems, while the Popov criterion applies to systems whose forward-path teansfer function ha a simple pole ats =O but ae otherwise stable. Moreover, the circlecri- terion guarantees global exponential stability, whereas the Popov criterion only guarantees, ‘plobal asymptotic stability Proof Note that 48 5(sI-AY' =(5I-A+A)(61-AY" =1+AGI-AY, Hence 49 (Lersyh(e)=(4r) 2 4e661-Ay" BL = Sard-te(st-Ay' b+reb+ red (sI-AY b. a 1's =o, then the term d/jeis purely imaginary, sothat232 Lyapunov Stability chs 50 Re[(1+Jwr)h jw) =Relr (d+ eb) +e(+rA) Gal A)" bl. Define the transfer function St g(s)=r(d+eb)+e(I-+rA) (sI- Ay" b, and observe that the quadruple (A, b, €(/ +A), r(d + b)) is a minimal realization of g. “Moreover, in view of (50), (47)is equivalent to 52 inf Reg (ja)>0. Hence, by Theorem (13), there exist a symmetric positive definite matrix Pe R™, a (row) vector qe R™", we R, and € > 0, suchthat 53. A’P+PA=-eP-q'a, 84 b’P+wq=e(/+rA), $5 wrar(d+eb) To establish the global asymptotic stability ef the system (41) ~ (42), choose the Lyapunov function candidate 56 V(x, B)=x'Px+d8? +27 Yo), where $7 yQ)= { (0) do. Since @ belongs tothe sector (0, =), it follows that yiy) 20 Wy. Hence Vis positive definite and radially unbounded, and is thus a suitable Lyapunov function candidate for applying Theorem (5.3.56). Now $8 Vak P+ x'PR+2dbi+2r Hyd =(Ax~ boy’ Px+-x'P(Ax- bo) ~ 2460 + 2r0 [e(Ax- bo) ~d9). Now nate thatd& = y— ex. Substituting this relationship in (58) and rearranging gives 9 (A P+ PA) x— 2b Px 29e (1 + rA) x-2r (d+eb) ¢*—2y0. ‘Now substitute from (53)~(55) into (59). This gives, by familiar arguments,Sec. 5.6 ‘The Lur'e Problem 233, 60 V=—ex’Px—(qx—wo)?—1 (d-+eb) 6? -2y6. Next, since g (je) > (d + eb) as «9, [cf (51)], (52) implies that r(d+eb) >0. Hence fronz (60), it follows that 61 Vs-ex'Px-29650, ¥x.& Itisnow shown that 62 V(x, 8) 0. Ifx=0 but 0, then y =d& #0, and y#> Osince 6 belongs to the open sector (0, =) {cf. (43)]. Hence once again V <0. Now tke global asymptotic sta- bility of the system follows from Theorem (5.3.56). Ml 63 Corollary Consider the system (41) (42). Letall ypotheses be asin Theorem (46), except that 6 belongs to the sector (0,k) where k > Ois some fnite number. Under these conditions, the system (41) ~(42) is globally asymptoicaly stable ifthere exists a number > Osuchthat o inf Re[(1+ jor) holt >0, Fig. 5.21234 Lyapunov Stability cas Proof Perform a loop transformation as shown in Figure 5.21, by placing a negative feedback of -/k around and a positive feedforward of {/k around k. This leads to 65 o=9l1-CA91', which belongs tothe sector (0, =), and 1 6 his)=h(s) + Now suppose (64) holds. Then, since 67 Re[(1 + jar) h,ja)] =Re[(1 + jor) kU] + i itfollows that 68 inf Re[(1 + jar)hja)]>6. ‘The global asymptotic stability ofthe system now follows from Theorem (46). i Like the circle criterion, the Popov criterion can also be given a graphical interpreta tion. Suppose we plot Re h jo) vs. «alm k je) as @ varies from 0 10 e2, Note that, since both Reh (je) and lm (ja) are even functions of «, itis rat necessary to plot negative values of @. The resulting plot is known as the Popov plot, in contrast with the Nyquist plot, which isa plot of Re h (ja) vs. Im # ja). The inequality (64) states that there exists a non- negative number r such that the Popov plot of h lies tothe right of a straight line of slope I/r passing through the point ~I/&. If =0, the straight line is vertical 69 Example Consider a system of the form (41) ~ (42), with a@=2— s@+i) ‘The Popov plotof his shown in Figure 5.22, tis clear from this igure that, if < 2, then itis always possible to draw a straight line through the point ~1/k such thatthe plot lis tothe rightof the straight line. Hence, by Corollary (63), we conclude that the feedback system is slobally asymptotically stable forall me-invariant néblinearites inthe sector (0, 2). Problem 5.27 Consider a feedback system of the form (1) ~(3) with MO GD E+DEH Using the circle criterion, determine several intervals [a,b] such that the feedback system is globally exponentially stable forall belonging to the sector (a, b].Sec.5.7 Converse Theorems 235 winger Fig. 5.22 Problem 5.28 Considera feedback system of the form (1) ~(3) with 1 hs)= —__. O° Gas Using the circle criterion, determine numbers 0 < a < such thatthe feedback system is glo- bally exponentially stable forall nonlinearities inthe sector (a, b Problem 5.29 Consider a system of the form (41) —(42) with A(s)= OTe Using the Popo criterion, show that the feedback system i globally asymptotically stable forall time-invariant nonlinearities @ belonging tothe sector (0k) where kis any fnite number. 5.7 CONVERSE THEOREMS Inthis section, the so-called converse theorems of Lyapunov stability theory are stated and proved. In the next section these theorems are applied to four problems in control theory, and itis shown that converse theorems lead to elegant solutions to each of these problems. ‘Though there are several converse theorems [see e.g. Hahn (1967), Sections 48 t0 51], only three are stated here, namely those for uniform asymptotic stability, exponential stabil- ity, and global exponential stability. In essence, these theorems state thatthe conditions of, Theorems (5.3.25), (5.3.45) and (5.3.62) are necessary as well as sufficient. Ifthe equil- bbrium 0 has the stability property mentioned in each of these theorems, then there exists a Lyapunov function V satisfying the conditions of the theorem. Since the function V is236 Lyapunov Stability chs constructed in terms of the solution trajectories of the system, the converse theorems cannot really be used to construct an explicit formula for the Lyapunov function, except in special cases (e.g, linear systems: see Section 5.4). However, they can be used in the same way as the Lyapunov functions for stable linear systems are used in Section 5.5: Knowing some- thing about the stability status of System A allows us o conclude something about the stabil- ity status of arelated System B. We begin by presenting two preliminary results. 1 Lemma (Gronwall 1919) Suppose a:R,—>R, is a continuous function, and b, © 2 Oare given constants. Under these conditions, if 2 ans be[eacarve 20, then 3 alt) < bexp(cr), vr 20. Remarks The point ofthe ema is to convert the implicit bound in (2) othe explicit bound in (3). If band care not constants but are themselves functions of, then the bound (3) needs to be replaced by a more complicated expression. This is known as Bellman's ine- ‘quality, see Bellman (1953). Proof Define 4 den=b+fe a(nde ‘Then (2) states that 5 a(t) sd@,We20 Further from (4) and (5), 6 d=calt) Sed(),r20. Now using the integrating factor exp (~ct)in(6), one can show that(6) implies 7d EdOexp(et)=bexp (ct) “The conclusion (3) follows from (7)and (8). Mt 8 — Lemma (Massera 1949) Suppose ot) isa given function of class L, and h>0 is a given constant. Then there exists a C™ function'¢:]R, —> R, such that (i) and Y' are both functions of class K, and (ii)Sec.5.7 Converse Theorems 237 9 Sricm|dtR, by O(7)=7[7/0)]. Then clearly 918(2)] =y0(¢)], and 10 {otecnar=, for each integer n 2 O there exists a unique time , such that G(,)= In +1). [OF course, ‘6(0)= 1,} Now define acontinyous function 7: (0, =») (0, )as follows: (i) A(t all n > 1. (ii) In the interval (t,t 1), 1(t) is an affine (i... linear plus a constant) function of t. (ii) In the interval (0, 1,). 1(2)= 14”, where O

ote) Vr > 0; (ii) for each number 7, we have u [roan ‘This is because, as t+, the Function 7(7) "blows up” quite slowly since p <1. Now iis continuously differentiable up to all orders except ata countable numberof values off. By rounding out the comers, one can replace 7 by another function n which is C™ and which continues to have the same three properties. Now define u,7:R, —>R, by 12 wr)=exp[-+ I(r) ifr >0, m0) =0, 1B y=[ wera 8 1p is claimed that this y() is the desired function, To see this, rote first that 1':(O, =)—+(0, 2) is well-defined, that 17 "(r)—¥0 as r yee, and 17'(r) 2° as r+0.238 Lyapunov Stability chs Fig 5.23 Hence is continuous and isin class K, (0) =0, and y(r) 1 as r—see, Since y’(r)= mtr), the second integral in (9) becomes, 14 flocs) exp (Ax) dt= 2. ‘Toshow that J is finite, note that 15 pl o(1)} < p[n(0), since w belongs toclass K, =expi-(+ Dn) exp [-(h+ It) Now itis clear that /2 is init. Showing that the first integral in (9) is finite requires a bit of work. Since (¢) < n(t) V4, itis enough to show that 16 YIN dt 9such that 6A Usle, t,x) S pila exp[-8(e—)}, Ve f 2 0, Va, Finally, suppose that, for some finite constant, 65 Daft, wl <2, Vt 2 0, VxE By. Under these conditions, there exist a C* function V:R,xR"—>R and constants a,b, ¢, m>0,p > I, such that 66 alli” < VUz,x) 0, VxeB,, 67 1D, V(t, x) < milxil, Ve > 0, VxeB,, Proof Choose the constant > I such that 68 (P-1)8>A, andlet 69 Yr) =r, o()=prexp(~30. ‘Then, as can be verified readily, the condition (39) is satisfied. Hence, by Theorem (24), the function TW VeeH=J Hse 4 x)ll]dt=J ste, 4, x) dt isaLyapunov function, Showing that V satisfies (66) is quite straight-forward. Firs, we uxu? exp p-p8(e-nyae= He xu, mn VG vsfu Wx exp l-p&le—H)de= 5 tixi?. ‘Next, from (64) und (65),See, 5.7 Converse Theorems 245; 72 Mle sc, 2] $ Alls 1, 2) < Agllal, ¥e> £20 Hence, from(31), 73 ast. 2 EE vets e+ 2m) ‘Therefore (70) implies that i108 m4 vena” fat a xii As shown in{62), 75 Veen) = yin) =x, Finally, using (53), and noting that 7(r) = pr’, gives 76D :V(t, 20M f py! xil?™ exp[—(p-DBtl exp (At) dt =constant IIx?! This completes the proof. i ‘The next corollary shows that, for an exponentially stable equilibrium, itis possible to construct a “quadratic type” Lyapunov function. 71 Corollary Suppose all hypotheses of Theorem (63) are satisfied. Then there exist a C* function W:R, x1R° —> Rand constants 0, B,Y, W > O such that 78 alxil? < W(, x) < Bilxil?, Wie, x) s ~ylIxIl?, Vr 2 0, Vue B,, 79 ID,WG, x) S pllxll, Ve 2 0, VxeB,, Proof Construct Vasin Theorem (63), and let 80 Wo,x)=1VG, OP, ‘The details areleftasan exercise. Theorem (63) and Corollary (77) can be extended in a totally straight-forward manner to global exponential stability,246 Lyapunov Stability Ch. 81 Theorem Consider the system (25). Suppose fis C*, and that f(t, 0)=0¥t > 0. Sup- pose there exist constants H, 8, 4> O such that 82° s(t, 1, x)II S plixilexp{-8(c—0), Vt > £2 0, VxeR", 83 WDaflt, II 0, VaR" Under these conditions, there exist a C* function W and constants ty, B, Y, > Osuch that 84 alin? < Wo, x) < Bilxll?, WO, x) Sy IhKI?, Ve 2 0, VxeR', 85 1D,WU,x)H < pllxll, Vr > 0, Vxe R™ Proof Simply follow the proof of Theorem (63) and replace B, and By, by R" throughout. In this way, it follows that the function V defined in (70) satisfies (66) and (67) forall xe R". Now define Wby (80). i Note that the condition (83) requires the function fto be globally Lipschitz continuous; thisis a much mote restrictive condition than (65), especially for autonomous systems. If we attempt to extend Theorem (24) to global uniform asymptotic stability, then we run into the difficulty that the function V defined in (41) may not be decrescent. The problem arises in (58), where the upper bound on T approaches ras r ->e. In the case of exponential stability, the difficulty does not arise, because the function (Ix) is inearin IIx; see (72) t0(74). Hence Theorem (24) can be extended to cover global stability provided the function OR, OR, defined by 86 (r)=sup sup sup Ist, x) ccan be bounded by a linear function of 5.8 APPLICATIONS OF CONVERSE THEOREMS: In this section, the converse theorems derived in the preceding section are used to solve four problems in control theory. 5.28.1 Exponential Stability of Nonlinear Systems 1 Theorem Consider the system 2 X=fxO), where fis andf0)=0. DefineSee:5.8 Applications of Converse Theorems 247 Then isan exponentially stable equilibrium of{2)ifand only ifthe linearized system 4 a=AU is (globally) exponentially stable. ‘Theorem (1) resolves one of the issues left hanging in Corollary (5.5.26). If A is defined as above and if all eigenvalues of A have negative real parts, then 0 is an exponen- tially stable equilibrium. If some eigenvalues of A have a positive real part, then 0 is an unstable equilibrium, But what if all eigenvalues of A have nonpositive real parts, but some havea zer0 real part? In such a case, depending on the nature ofthe neglected higher order terms, itis possible forthe origin to be asymptotically stable. But Theorem (1) shows that, cevenif the originis asymptotically stable, itcannot be exponentially stable. Proof "If" Thisis just Corollary (5.5.26), “Only if" Suppose 0 is an exponentially stable equilibrium of (25). Then, by Corollary (5.1.77), there existsa C? function V:R" ~» R andconstants B.7.}, and r > 0 such that Sal? < Vit, x) < Blal?, Vee, x) < URI, Vue B,, 6 VG, x)ll Osuch that 17 Ais,(t,1,%9) lI $ pln exp[-B(t—0)}, Vt 2 # 2 0, VxER®, VreR,. Finally, suppose there isa constant > O such hat 18 ID, fit, xi < ellxil, Vr 2 0, xe R". ‘Then the nonautonomous system (13) is globally exponentially stable provided 19 e< Me=VB-M PH where p> 1 isany number such that (p—1)8 >i.Sec. 5.8 Applications of Converse Theorems 249 Remarks To putthe conditions of the theorem in better perspective, consider the linear time-varying system 20 ke = A(X), and suppose A(1) is a Hurwitz matrix for each fixed ¢. Then 2 is the maximum of WA()IL with respect to ~ isthe largest (ie, the least negative) of the real pats ofthe eigenvalues of A(O), as varies; and is the maximum ofthe condition number of A(t) as varies. Proof We begin by estimating the rate of variation of the function s,(z, 0, x) with respect tor. From (14), it follows that 2 5,(t,0,x)=x+f fr, 5,(6,0, x)] do. Differentiating with respect to rgives 2, 22 5/8/(1,0,% =| Patt 56.0.2 d0+ Df 5(6,0.%) 20.0040 Forconciseness, define 23 g(D= 15,4, 0,x00r, and note from (18) that 2A ID )Mlr,$,66,0, n]ll Sells,(6, 0, x) < ep. lxl exp (8a). Using (24) and (16) in (22) gives 25 80) [eullatlexp(-80)d0+{ Ag(o)do < BPE [records Applying Gronvall’s lemma to(25) gives 26 85,0, 0,xV0rl=9(a) < HO" expan), ve20 Next, foreach re R,, define a Lyapunov function V,:.R" > R for the system (14) as in ‘Theorem (5.7.63). Select > 1+ lie., (p— 1)8 > A], and define250 Lyapunov Stability Ch. a von=f lis,(1, 0, x) I" dt. This is the same function as in (5.7.70), since the system (14) is autonomous. At this stage, replace rby 1,and define V:R, xR" = Rby 2 Von=| s(t 0.01" dt ‘Then, as shown in the proof of Theorem (5.7.63), it follows in analogy with (5.7.71), (5.7.74), and (5.7.75) that 29 uxil? s VG, x) < Fair, pd 30 DV(t, x)flt, x) =—IIxil, Letus compute the derivative V(r, x) along the trajectories of (13). By definition, BL Vit, =D, VG, x) +D2V UE, Ue, ¥)=D VCE, x)— XI, only remains toestimate D V(t, x). Levy :=p/2. Then, from (28), 32 Dvn 2 [s(x 0, x)5,(t, 0, x)" dt Ay Isi(t,0.998(6, 0,017" 81(0,0,3) 25,6, 0,004, 38 1D.ViI s f2¥ B56,0.90" 12504,0,01 a Now use the bounds (17) for tls,(z, 0, x) and (26) for 115,(t, 0, xV@ril, and note that 2y=p. This gives 34 1D\vecn spent BE expt -pbeedeide a 8@-D3-A] Let m denote the constant multiplying iIx!I? on the right side, and note that m < 1 by (19). Finally, from (31),See.5.8 Applications of Converse Theorems 251 35° Vit, x) $—(—m)Itxl, Now (29) and (35) show that V is a suitable Lyapunov function for applying Theorem (5.3.62) to conclude global exponential stability. 5.8.3 Odserver-Controller Stabilization in this subsection, itis shown that a well-known strategy for stabilizing linear time- invariant systems also works for nonlinear systems. As a moti described by tion for studying the problem, consider a linear time-invariant system 36 (0) = Ax(e) + Bus), y(t) =Cx(s). ‘Suppose the system is stabilizable and detectable [see Chen (1986) or Kailath (1980) for definitions of these terms]. By the assumption of stabilizability, there existsa matrix K such that A -BKis Hurwitz. Hence, if we could apply the feedback control 37 wl Kx(0), then the resulting closed-loop system would be stable, However, x(t) cannot be measured directly, and only y(1) is available for control purposes. To overcome this difficulty, one can setupadetector, which isa system of the form 38 H(0)=Ax(e)+ Butt) + Flys) ~ Ca), where F is called the filter gain. By the assumption of detectability, there exists a matrix F such that A FC is Hurwitz, Forsuch a choice of Fit follows that a(t) —x(0) > 0as r=; in other words, asymptotically 2(¢) becomes an accurate estimate of x(t). With this in mind, suppose one implements the control law 39 w(t) =—Ka(r) Then the closed-loop system is described by A -BK }[x Fo A-BK-Fe| [x LetM denote the square matrix in (40), and define 10) 41 T= 40 W ‘Then252 Lyapunov Stability chs This shows thatthe spectrum of M (ie., the set of eigenvalues of M) is just the union of the spectra of A~BK and A~FC. Since both matrices are Hurwitz, it follows that M is also Hurwitz. The conclusion is that the stabilizing control law (39) continues to do the job even if the true state x(t is replaced by the estimated state 2(1). For this reason, the strategy is known as observer-controller stabilization. This is sometimes called the (deterministic) separation theorem, The preceding proof is very much a "linear time-invariant” proof, being based on eigenvalue arguments. Thus itis perhaps surprising that a similar result also holds for non- linear nonautonomous systems. Suppose the system to be stabilized is described by 43 XN)=fr, x), WD], YOD= Ble. XC), where x(Q)eR", wi)eR", y(NeR!, ER, xR'XR" 9R", and gR,xR'R!, Its assumed that fis Cand that f(t, 0,0) =0, Vt 2 0, Now suppose h: R, xR" > R is con- tinuous, and define the control law 44 uy=ht, x) This control law is said to stabilize the system (43) if h(t, 0)=0'Vr 2 0, and isa uniformly asymptotically stable equilibrium of the closed-loop system 43 iC) =Hle, xl), hls, x(0)}) Now a nontinear analog of detectability is defined. The system (43) is said to be weakly detectable if one can find afunction r: R, xR" xR" xR! — R® such that 46 (0) =H, 260), 0), YON = FL 20, WO), BLE MOD) acts as a "weak detector" for the system (43). This means that (i) r(t, 0, 0,0)=0, and i) there exist a C! function W:R, xR" xR" RR, class K functions a, B,y and a number p> Osuch that 47 alix-zil) s WO, x.) S BlUIX-all), Ve 2 0, WO%, ZEB, XBp, 48_D,WO,x,2)+D,WUt, x, ef, x, w+ D5WU, x, 2) elt, Xu, BCE) 4S -1(lIK=zII), Vi 20, VOX, 2, WEB, XB, XBy These assumptions mean that if initially x(t) 0.as >», As the name implies, weak detectability isa weaker pro- perty than detectability, though for linear systems the two concepts are equivalent. This isSec.5.8 Applications of Converse Theorems 253 discussed further in Vidyasagar (1980b). Now itis possible to state the main result: 49 Theorem Suppose the system (43) is weakly detectable, and that with the control law (44), the equilibrium x=0 of the system (45) is uniformly asymptotically stable. Suppose there exist finite constants r, h, such that 50 sup sup, max{ IDafle x, wl, Daf, xw)l} SA, St sup sup WD 2h(, HS Then the origin in R" xR" is auniformly asymptotically stable equilibrium ofthe system 52 iG =r{t, x0), ble, 2001), 53 AN=TKE, XU), ble, 20], Blt XO} Proof Itis rst shown that (0, 0e R" x R® isuniformly stable; then itis shown that itis uniformly atractive, To prove that the origin is uniformly stable, suppose an € > 0 is specified; the objective istoconstructad > Osuch that 54 Ixy ll <8, llzp ll llx(o)ll , C7 functions 8, 6, yof class K, anda finiteconstant > 0, such that $6 WIXI) SVG x) $ O(IIXIl), Vi 20, VER, 57D, V(x) +D2V(t, x)Mlt, x. 9] S—YUIXID, Vt 2 0, VxEB,, 58° ID:V(, 91 SL, Wr 2 0, VER, ‘Note thatthe same B, appears in (51) as well as (56) ~ (58). This cuts down on the prolifera- tion of symbols, without any loss of generality. By the same token, it can be assumed that =r, where p appears in (47) ~ (48), and that € Sr. The construction of the quantity 8 is achievedin several stages. First, select254 Lyapunov Stability chs 59 5)=min{e.p. ph). ‘where jLappearsin(51). Next, select; such that 6(¢,) < 0(6,/2),and select 8; such that 6 &=min{e,,r, We YLA), where L, 2, 1 are defined in(58), (50), and (51) respectively. Next, choose 8; > 0 such that BBs) < a(8,), and define 61 8=min(p, 8/2). To show thatthe above choice of satisfies (54), itis first shown that 62 Ix(t)~2(t)l1 <8,, Vt 2 toy 63 LW, 0.20915 0,¥1 2 to a “4 vt, x(1)] <0, whenever e; < Ix(t)ll <€2, Vi 2 to. To follow the proof, itis helpful to observethat & a ce % 8s Ff. h fo, asclaimed. With the aid of (62) — (64), itis easy to establish (54), which is uniform stability. Using (62) — (64), one can prove (66) and (69) as before, with the interval [to, 7] replaced by {¢o,). Finally, since llz(0)ll < Ux(il+ ila(t)=x(O)l_-< (G/2+82) S 8; Se, (54) fol- Jows, and the origin in R" x R" isa uniformly stable equilibrium. To conclude the proof, itis shown that the origin in R* x R® is uniformly attractive. Pick any € > 0, and construct a corresponding 5>0 as above. It is shown that Bs Bs is region of uniform attraction. Suppose Xo, 29¢B3. Then, since lla(t)Il <8, Vt 2 fo, it fol~ lows from (50) and (59) that lh{t, 2@)]II 0 as t—>ee, Unfortunately, this step requires rather advanced concepts; hence only a sketch of the proof is given here. First, a so-called "comparison equation” is set up, namely TI e@)=nle]+00. Itcan be shown that TB €(0) > vO)Helt) 2 vi), Vt 20. ‘This is called the comparison principle; see e.g., Walter (1970). Then, since o(t) +0, it ccan be shown, using a generalization of the invariance arguments of Lemma (5.3.71), that ‘€(0) 90 as te»; sce Levin (1960). Now (78) implies that v(r) 0. Hence there exists @ function) of class L such that 79 Wx(tg +0) < H(t), Vt 20. Finally, (72) and (79) together show thatthe originis uniformly attractive. i 80 Example Consider the system iys-ay tury. nex} ayers ‘The linearization of this system around: 29 ef e{o] «2 af] i Note thatthe linearized system isnot stabilizable, but detectable. Hence the feasibility of ‘he observer-controller strategy cannot be established by linearization arguments.Sec.5.8 Applications of Converse Theorems 251 Suppose we could apply the control aw ay. ‘The resulting closed-loop system is described by Sys an td ip =a] ad Totest the stability of this system, choose the Lyapunov function candidate Vv: f+ Then 2x} =x 1x} +x) =— (x 0.523)? +0.75x$), ‘Thus x=0is an asymptotically stable equilibrium (but norexponentially stable). However, the afore-mentioned control law cannot be implemented because x cannot be measured directly. To get around this difficulty, let us set up the system ays-2) tu da=ch-2ty—y To test whether this system is a weak detector, choose the function WOe2)=() 24) Hera Then Wor, 2) (21 — 1) (E18) 422-22) 2-H) De, =a) #22) Gh =X} 422 —ADd] 20 4x] fern Ferman 22-2 aay 2} epee ‘The coefficient matrix [2 en] Jets 2} is positive definite if tIxll, lz1l are sufficiently small. Therefore it follows that (48) is satisfied, and thatthe system above isa weak detector. ‘Now, by Theorem (49), it follows that the implementable control law258 Lyapunov Stability chs stabilizes the system. Before leaving the example, two comments are in order. First, itis worth noting that the coefficient matrix M(x,2) is positive definite only when both tix! and lizik are small — it is not enough that the quantity lIx—2zl be small. Thus the system above is only a weak detector, and not a true detector. Second, since x2 is available directly, its somewhat extra- ‘agant to set up a detector for it. Instead one might think of setting up a "reduced-order” detector forx, alone. The theory tocover this situation is not available at present. 5.84 Stability of Hierarchical Systems Inthis section, we study the stability of systems of the form KH O=Nl, x1), HD =Hle 40, 2200), at HO =Gle, x40, x4) ‘Such a system is said to be in hierarchical or triangular form, since the differential equa- tion governing x,(¢) depends only on xy(t), ‘++, %(0), but not on x,(1) for j >. Given an arbitrary differential equation of the form (13), there exist systematic procedures for renumbering and regrouping the variables x,~°- x, im such a way that the system equa- tionsassume the hierarchical form (81); see Vidyasagar (1980c). The objective is to deduce the stability properties ofthe system (81) by studying only the simplified systems 82 =H I, 0-0 x40], +L. Comparing (82) with te i-th equation in (81), one sees that the equations in '81) have been decoupled by setting x, =0 for j i-1 in the i-th equation. For this reason, (82) isreferred toas the -th isolated subsystem. ‘Now the main result of this subsection is stated. To streamline the presentation, two notational conventions are employed, namely:Sec.5.8 Applications of Converse Theorems 259 xeK=| is} Ls Inother words, isthe state vector of the frst equations in ($1), whereas x vector ofthe overall system, is the state 84 Theorem Consider the system (81). Suppose each function f, is C', and that the fol- lowing conditions are satisfied for each ig (1,°*>,1): 85 £(1, 0," 0)=0, Vr 20, ‘and in addition, there exist constants i <2 andr > Osuch that 86 sup sup IDaf(r, XI 120 seb, Under these conditions, x = is a uniformly asymptotically stable equilibrium of the system (81) ifand onty ifx,=0 is a uniformly asymptotically stable equilibrium ofthe system (82) foreachie (1,-"°,I} Remarks Note that, if the system (81) is autonomous and each function fis auto- rnomous, then (86) is automatically satisfied. Hence the theorem requires only very mild conditions inorder to be applicable. Proof Itis helpful to distinguish between the solution trajectories of (81) and of (82) through appropriate notation. Let x/"(e) denote the solutions of (82), where the superscript ()is supposed to suggest “isolated.” x, (1), without the superscript, denotes the correspond {ng component ofthe solition of (81). Also, while the concept of uniform asymptotic stabil- ity is independent of the particular norm used, itis convenient to take MI to be the norm 1st). defined in Example (2.1.9). With this choice of norm, it follows that 87 x= max (xy Hoo, tI “Only if” Suppose x= (isa uniformly asymptotically stable equilibrium of (1). Then, by Theorem (5.1.61), there exist functions n of class K and o of class L, and a constant p > 0 such that 88 Ux(s)Il SAL Ix(¢0)I} GG —t0), We 2 fo, Wx(toe Bp. Inpanticular, suppose260 Lyapunov Stability chs 89 x(f0)=[0" 0" xjq OO’ x9 By, ‘where xj occurs in the i-th block. Then, in view of (85), it follows that the first i ~ 1 blocks of the solution x(-) equal zero, while the i-th block equals x{?(:). Now (88) implies that 9 ix!()l < HLIIxoll) ar—1), Vt > fo, This shows that x, =0 is a uniformly asymptotically stable equilibrium of (82). The argu- ‘meat can be repeated for each i. “If” The proof is given for the case ! =2; the general case follows by induction, By assumption, there exist functions n, of class K, 6; of class L, and a constant r > 0, such that HL UXPCU S MyLI x9 Nou(t=t0). ¥E 2 fo, VXi0e By, where x{(-) denotes the solution of (82) with i=1, and with the initial condition X(t) =o. Similarly, the hypothesis on the second isolated subsystem, combined with ‘Theorem (5.7.24), allow one to conclude the existence of aC function V, class K functions @, 8,7, andconstants r >Oand L <=, such that 92 allixy l) < Vit, x2) < Blllxg I), Wr 2 0, WEB, 9 Vm) 1 V(t, X2)-+D 2 Vt, x2) f(t, 0, x2) S —Y(IIxg II), VE 0, Uae By, 94 D2Vit,xa)ll SL, Wt 2 0, VRE B,, Now consider the system 95 NAL HO) HO =Salt, 21, 1210] Its to be shown that x=0 is a uniformly asymptotically stable equilibrium of this system. ‘The proof is divided into two parts, namely establishing (i) uniform stability, and (ii) uni- form attractivity. To establish uniform stability, suppose e > O is given; itis necessary 10 determine 5 > Osuch that 96 xo <8, Ixspll <8 lixy(t-+o)Il 0, first choose €; > Osuch that B(e,) < ate), and then choose 8, > O such thatSee. 5.8 Applications of Converse Theorems 261 97 5, 0 such that 98 B 0, and select 8 > 0 in accor- dance with 98). Suppose lix(to)ll <8. Modify (100) to is a uniformly stable equilibrium. 102 2 fe, x0] $78 (V Ue (OI) +L NV OE—Fo)- This inequality is very similar to (73). Mimicking those arguments shows that VUe+tq, xa(t +15)] 90 as ¢—> 22, uniformly in fo. The details are left as an exercise. This shows that B ‘sa region of uniform attractivity, and completes the proof. Using appropriate converse theorems, it is possible to establish theorems regarding other forms of stability 103 Theorem Consider the system (81). Suppose each function; is C', and satisfies (85) and (86). Under these conditions, x=0 is an exponentially stable equilibrium of the system (81) ifand only ifx, =O is an exponentially stable equilibrium of(82) or each ie (1,262 Lyapunov Stability ch.5 Proof “Only if" This part of the proof closely follows the corresponding part of the proof of Theorem (84). ati left as an exercise. “If” Since exponential stability implies uniform asymptotic stability, Theorem (84) implies that x=0 is a uniformly asymptotically stable equilibrium. It only remains to show that solution trajectories converge to x= O exponentially fast. Now x, ()={). Hence, by the hypothesis of exponential stability, there exist constants a, b, r>0 such that (ef. 6.13) 104 Hx,(0)11 $ allaioll exp [-B(¢— Fo) VE 2 fo, VXI0E By. Next, since x; =0 is assumed to be an exponentially stable equilibrium of the system (82) with = 2, it follows from Corollary (5.7.77) that there exista C' function Vand constants 6, But.r>OandL 0, Va3€B,, 106 VU, x) $ ~yllxIl, Ve 2 0, VE B,, 107 WD2V(t, x2)I1 $ Lilxgll, We 2 0, axe B, Without loss of generality, itis assumed that the same constant r appears in (104) as well as in(105) ~ (107). Now (102) becomes 108 uu, x2(01 $ CPA) Ut, x3 (0] + La Ux xg (0) exp [-b-ta)he $-GBWIt (01 + (Lda NG) Ixy MLV (6, x2(01) exp [-B Oto} WE 2 bo Define W(t) =V [t+10, X(t +19)]'@. Then it readily follows from (108) that 109 2W(1) < ~(~P)W(+ (Laat) x30 exp [b(t ~to)]. WE 2 to. Note that V(to.X20)1" SVB txzo From (109) and (110), it follows that there exist constants n,n such that 11 W(0) $ nL IIxyoll + xg lt] exp (~ 0), VF 2 0. Inturn, (111) implies thatSec. 5.8 Applications of Converse Theorems 263 HZ xsl ¢ Nar) Wee to) $ NG) [Xo I + Xo IH exp [WC —t)], Ve 2 fo. Now (104) and (112) together establish that the equilibrium x=Oisexponentialy stable, 113 Theorem Consider the system (81). Suppose each function f, is C', £(¢, 0)=0, Ve 2 0, and 114 sup sup D6, DI <=, for ie (1,--,1}. Under these conditions, x=0 is a globally exponentially stable equili- ‘rium of the system (81) if and only ifx,=0 isa globally exponentially stable equilibrium of the system (82) foreachi¢ {1,---, 1} Proof “Only if" This part ofthe proof is left asan exercise. "If" Let r =sein (104) 10 (107) and proceed asin the proof of Theorem (103). Ml Problem $.30 Consider he system XU) =£x(), WE), Ve 2 0, where x(:)eR", u(eR”. Suppose fis C?, and that f(t, 0, 0)=0, Ve > 0. Define at alae Recall that the pair (A, B)is said tobe stabilizable f there exists a matrix Ke R™™" such that A~BK isa Hurwitz matrix. (a) Prove the following statement: There exists a C? function r:R"—9R™ such that x=(isan exponentially stable equilibrium of the system 3d) =fIx(0), Flex], ifand only ifthe pair (A, B)is stabilizable. [Hint; Use Theorem (1),] (b) Construct an example to show that (a) is false if "exponentially stable" is replaced by “asymptotically stable.” (Hint: See Example (80) Problem 5:31 Using Corollary (5.7.77) and Lemma (5.4.53), state and prove an exten- sion of Theorem (1) to time-varying systems. Problem 5.32 Using the results of Problem 5.31, extend the results of Problem 5:30 to time-varying systems. Problem 5.33 Consider a modification of the linear system studied in Example (54.90), namely264 Lyapunov Stability chs RO=AOXO), where r -1+acos i Amel asin) cos At) ~asinkrcoskt —I+asin?ht | (a) Find the state transition matrix. (Hint: see the form given in ¢he Example and modify itsuitably.) (b) Using Lemma (5.4.79) and Theorem (5.4.89), find a range of values of the pair (a, A) for which the above system is asymptotically stable. () Use the result of Theorem (113) regarding slowly varying systems to construct suit- able bounds on a and 2. which assure that the system i asymptotically stable. Compare with the results obtained in part (b). Problem 5.34 Prove the "only if parts of Theorems (103)and (113). Problem 5.35 Give a proof of Theorem (103) based on Problem 5.31 in the case where each function fis C2, 5.9 DISCRETE-TIME SYSTEMS, Until now the emphasis inthis chapter has been on continuous-time systems described by differential equations. In the present section the focus is on discrete-time systems, described by arecursive relationship of the form 1 =hows where x¢R", and f,:R" > R" for all k 2 0, Note that (1) always has exactly one solution corresponding to an initial condition of the form x(K 9) = xo; this solution, evaluated at the k- th instant of time (k 2 ko), is denoted by s(k, k, Xo). If, in addition, itis assumed that fis a continuous function for all k, then s(k, kg, ) is also continuous for each pair (k, kg) with k 2 ko. Thus existence, uniqueness, and continuous dependence of solutions of recursion relations is really notan issue, in contrast with the case of differential equations. ‘The objective of the present section is to define various concepts of stability for the equilibria of the system (1), and to present various stability theorems. Since the details of. the definitions, theorems, and proofs very closely parallel those of their continuous-time counterparts, very few details are given, and all proofs are leftas exercises. A pointxye R" iscalled an equilibrium of the system (1) if 2 f(x) =x, VE 2 0, i.e. iffxy is a fixed point of the map f, for each k 2 0. Clearly, if (2) holds, thenSec.5.9 Discrete-Time Systems 265 3 stk, ko, Mo)=Xo, Wk 2 ko 2 0. One can assume, without loss of generality, that the equilibrium of interest isthe origin, ie, that 4 £,(0)=0, vk 20. ‘Suppose V:Z, xR" +R, where Z, denotes the set of nonnegative integers. Then, along the solution trajectories of (1), define SV =VIk Stk ko. Xo): ‘The forward difference of the sequence {V*,) is 6 AVE = Via VAS VIRHL, SH, ke x4) “VK, mi) With this in mind, we define the forward difference function AV:Z, xR" > R as fol- lows: 7 AVR, D=VIR+, 941, k O1- VOW). Obviously, AV depends on both the function Vand the system (1). Note that, along the tra- Jectories of (1), we have + DAVE x). 8 OM, 8 Definitions The equilibrium 0 ofthe system (1) is stable if for each €>0 and each ko 2 O.there existsa3=8€, ko) such that 9 Ux Ilstk, ko, xo) 0 there exists a= 6(@) such that 10 ko 20, Hxgll <5(€)> lis(k, ko, Xo)Il <€, Vk 2 ko. 11 Definitions ‘The equilibrium 0 is attractive if for each ko 2 0, there exists an Tu, such that 12 Ixpll (K, ko, Xq) 90 as kre, Itisuniformly attractive ifthere exists any > O such that 13 xpll 0ask 92, uniformly inks, X Equivalently, Ois uniformly attractive if there exists an 1, > 0, such hat for each €>0 there ‘exists an m =m (e) such that266 Lyapunov Stability chs 1d Ixy il me. The equilibrium 0 is asymptotically stable if it is stable and antractive; it is uniformly asymptotically stable fits uniformly stable and uniformly attractive. 1S Definition The equilibrium 0 of (1) is exponentially stable if there exist constants na >Oandp < I such that 16 xp 0, there existsanm =m(n, €) Such that 18 thxgll rm. tis globally exponentially stable ifthere exist constants a > O.and0p < I such that 19 listkg +s ky, X)ll Salix lip", Wh, ko 2 0, Vxe Rt So much for the definitions of various forms of stability. The stability theorems for discrete-time systems are also reminiscent of their continuous-time counterparts. To state these one needs the concepts of positive definiteness, etc. 20. Definitions A function V:Z., x R" +R isa locally positive definite function (Ipdf) ‘fli) V(k, 0)=0'Vk 2 0, and ii) there exists a constant r > Oand afunction ccofclass K such that 21 a(Iixil) $ Vik, x), Vk 2 0, Vxe 8,. Vis decrescent ifthere isa function B of class Kand aconstant r > O such that 2 Vik, x) $ BiIixll), Vk 2 0, xe B,, Visa positive definite function (pdf) if(i) V(k, 0)=0k > 0, and (i) there isa function o ofclass K such that 2B al ixIl) < Vék, x), Vk 2 0, Vxe R" V is radially unbounded if V(k, x) as IIxIl 500, uniformly in k, V is radially unbounded if V(k, x) —> as IIx ll >, uniformly ink. 24 Theorem (Stability) The equilibrium 0 of (I) is stable if there exist a function V:2, xR" > Randa constant r > Ouch that (i) Vis an ipdf, and (i) 25 AV(k, x) $0, Vk 2 0, Vxe8,. If,imaddition, Visdecrescent, then 0is uniformly stable.See.5.9 Discrete-Time Systems 267 26 ‘Theorem (Asymptotic Stability) The equilibrium 0 of (1) is uniformly asymproti- cally stable fthere isa decrescent [pdf V:Z, xR" > R such that-AV isan pdf. 27 Theorem (Global Asymptotic Stability) The equilibrium @ of (1) is globally uni- formly asymptotically stable if there is a function V:Z, xIR° > such that (i) Vis a pdf, decrescent, and radially unbounded, and (ii)~AV isa pdf. 28 ‘Theorem (Exponential Stability) Suppose there exist a function V:Z, xR" 9R, and constants a, b, ,r >Oandp > such that 29 allxil” < V(k, x) < blINII?, AV(K, x) S ~cllxil”, Wk 2 0, VxE B, Then is an exponentially stable equilibrium. If'ic is possible to replace B, by R in (29), then is globally exponentially stable. ‘Now consider a linear shift-invariant system described by 30X44) =AN, 20, If we try the obvious Lyapunov function candidate 31 V(x)=¥'Px, then 32 AV(x)=x’A’PAX-x'Px=x’(A’PA-P)x, Hence the discret ime Lyapunov matrix equation is. 33. A'PA-P=-Q. 34° Theorem The equilibrium 8 of (20) is (globally) asymptotically stable if and only if all eigenvalues of A have magnitude less than one. ‘A matrix A whose eigenvalues all lie inside the open unit disk is called a contractive matrix. 35 Theorem Wish respect t0 Equation (33), the following three statements are equivalent: 1. Alleigenvalues of A have magnitude less than one. 2. There exists a positive definite matrix Q such that(33) has a unique solution forP, and this P is positive definite. 3. Foreach positive definite matrix Q, (33) has a unique solution for, and this solu- tions positive definite.268 Lyapunov Stability ch.5 ‘The proof of Theorem (35) is facilitated by the following lemma, 36 Lemma Suppose A is contractive and Q is a given matrix. Then (33) has a unique solution, given by ” Eayea! a Now consider the notion of linearizing a nonlinear system of the form (1) around the equilibrium 0 (assuming of course that (4) is true]. Suppose f; is C! for each k 2 0, and define fan, ] x 38 AS + fin) = fy) — Agx Suppose itis true that foo ae eT ‘Thenthe system 40 =A iscalled the linearization of (1) around 0. 41 Theorem 1/0 is a uniformly asymptotically stable equilibrium of the linear system (40), then itis also.a uniformly asymptotically stable equilibrium of (1). 42 Theorem Consider the autonomous system Bx =f. where is C and{(0)=0. Define far] 4A | Se xo If-Ais contractive, then 0 is an asymptotically stable equilibrium of (43). If A has at least one eigenvalue with magnitude greater than one, then Vis an unstable equilibrium of (43). Problem 5.36 State and prove the discrete-time analogs of the converse theorems of Section 5.7.See.5.9 Discrete-Time Systems 269 Notes and References ‘The concept of Lyapunov stability was fst formulated by the Russian mathematician ‘A.M. Lyapunov in a famous memoir published in Russian in 1892. This memoir was translated into French in 1907 and is available inthe Western world from Princeton Univer- sity Press; see Lyapunov (1892). Lyapunoy stated and proved the basic stability theorem, and also proved the linearization method (Section 5.5) the results concerning the Lyapunov ‘matrix equation for analyzing linear systems (Section 5.4), as well as several other results which are no longer popular. The symbol V for the function which today we call the Lyapunov function was introduced by Lyapunov himself, IC is a good guess that he was thinking of potential energy when he used this symbol, since in problems of particles mov- ing in potential fields, the potential energy is a logical Lyapunov function candidate (see Problem 5.13, Section 5.3). Whatever his motivation, the symbol V is now so firmly ‘entrenched that no one would seriously think of using anything else, though some intrepid soulshave been known touse v Originally Lyapunov did not distinguish between stability and uniform stability, ee. The awareness that the'Wistinction was important came gradually through the efforts of several researchers, the most notable of whom were Malkin and Massera. The asymptotic stability theorems that are based on the idea of invariant sets are due to Barbashin and Krasovskii (1952) in the Soviet Union and LaSalle (1960) in the United States. The contents of Section 5.6 on the Lur’e problem again represent the collective efforts of many individuals. The Popov criterion was originally proved by Popov (1961) using a ‘method quite different from what is given here. This method is now called the hyperstability ‘method; see Popov (1973). The proof given here, which demonstrates the existence of a Lyapunov function if certain frequency-domain conditions are satisfied, requires the so- called Kalman-Yacubovitch lemma, discovered independently by Kalman (1962) and Yacubovitch (1964). The circle criterion, based on the Kalman-Yacubovitch lemma, was proved by Narendra and Goldwyn (1964). See Narendra and Taylor (1973) ior a book- length treatment of this topic. Malkin and Massera were again instrumental in proving that some of the Lyapunov theorems were “reversible” and proved most of the contents of Section 5.7. The applica- tions of the converse theorems found in Section 5.8 are due to various authors. The analysis of slowly-varying systems is found in the thesis of Barman (1973), with earlier efforts for linear systems due to Desoer (1969) and (1970). The proof that the observer-controller stra- tegy works for nonlinear systems is due to Vidyasagar (1980b), The result on the stability of hierarchical systems is found in Vidyasagar (1980c), and is a generalization of an earlier result due to Michel etal, (1978).6. INPUT-OUTPUT STABILITY In this chapter, we present the basic results of input-output stability theory. ‘Tis theory ‘much more recent in origin than Lyapunov theory, having been pioneered by Sandberg and ames in the 1960"s [see Sandberg (1964, 1965a, 1965b), Zames (1966a, 1966b)]. While this chapter contains mostof the principal results ofthe subject, the treatment is by no means encyclopaedic. The reader is referred to Desoer and Vidyasagar (1975) for a thorough dis- cussion ofthe subject Before proceeding tothe study of input-output stability it is necessary to reconcile the input-output approach to system analysis and the state Variable methods employed in the preceding chapters. The methods of the preceding four chapters are predicated on the sys- tem under study being governed by a set of differential equations which describe the time evolution ofthe system state variables. In contrast, the systems encountered in this chapter are assumed t0 be described by an input-ouput mapping that assigns, to cach input, a corresponding output. In view of this seeming dichotomy, itis important to realize that an input-output representation and a state variable representation are two different ways of looking at the same system-—the two types of representation are used because they each give different kind of insight into how the system works. It is now known that, not only does there exist a close relationship between the input-output representation and the state representation of a given system, but that there also exists avery close relationship between the kinds of stability results that oe can obtain using the two approaches. At this stage one may well ask: Why not simply use one of the two approaches—why use both? The answer is that while the two approaches are related, they are not equivalent. Since, in analyzing a system, we would like to have as many answers as we can, itis desir- able to have both the approaches at our disposal, each yielding its own set of insights and information. Finally, it should be mentioned that many of the arguments and proofs in input-output theory are conceptually clearer than their Lyapunov stability counterparts. Compare the proofs of the circle criterion and the Popov criterion in the two approaches, for example. Also, analyzing distributed systems (&g, systems containing delays) in an input-output set- ting is no more complicated than analyzing lumped systems. In contrast, in the case of Lyapunov stability, analyzing time-delay systems, for example, is substantially more com- plicated than analyzing ordinary differential equations [see e.., Hale (1977)]. On the other hand, ata first glance at least, understanding input-output theory would appear to require a greater background in mathematics than understanding Lyapunov theory, sinee input- ‘output theory makes reference to advanced concepts such as Lebesgue spaces and so on. In many ways, this impression is misleading. As shown subsequently, there are only a few places where the full power of Lebesgue theory is needed, and almost everywhere one canSec.6.1 1, Spaces and Their Extensions m get by with the more familiar notion of Riemann integration. Part of the objective of this ‘chapters to make input-output theory as accessible tothe student as Lyapunov theory. G1 L,-SPACES AND THEIR EXTENSIONS, In this section, a brief introduction is given tothe Lebesgue spaces L,, and their exten- sions, and to the concepts of truncations and causality. Much of input-oweput stability theory, including the problem statements and the stability theorems, is couched in these terms, so that a certain degree of familiarity with these concepts is necessary 10 appreciate input-output theory. On the other hand, most of the input-output results given here do not require any deep results from Lebesgue theory other than the completeness of the Ly spaces, so thatthe pedestrian treatment given below is sufficien’ forthe present purposes. 6A Ly ‘The reader is undoubtedly famitiar with the idea of Riemann integration, as taught in undergraduate calculus. While this fine idea, thee are many "reasonable" functions that are not Riemann integrable. For example, suppose f:R, —>R, isdefined by f (x)= ifxis rational, and f (x)=0 if xis rational, Then every interval contains an x such that f (x)=0 and a y such that f()= |. Hence fis not Riemann integrable. This shows the néed for a more general concep of integration ‘There is another argument as well: Consider the set C[0, 1] of continuous real-valued functions defined on the interval [0, 1}, together with the norm, wes mgs 0 Vena ve9 (The reason for the subscript » becomes apparent later.) Then convergence in the norm Il, is just uniform convergence over [0,1]. Now itis well-known that if a sequence of continuous functions converges uniformly toa function f-), then f) is also continuous; see eg., Royden (1963). Hence the normed space {C[0, 1}, Vil.) is a Banach space in the sense of Definition (2.1.24) In contrast, suppose we define the norm, r 2 2 tel Ife de Then convergence in the norm I-Ilz is just mean-squared convergence, Now there exist discontinuous functions /() which are the limit, in the norm I: II, of a sequence of continu- ous functions. Forexample, let 0, t€[0, 0.5), 3 O74, 16105, 11 Then the Fourier series expansion of /(-) converges to f{-) in the norm I: liz. Thus the2m Input-Output Stability ch ‘normed space (C0, 1), lz) isnorcomplete, This leads us to ask how one may enlarge the set C[0, 1] such thatthe resulting set is in fact complete under the norm ll. This too is another motivation for studying L,-spaces. Unfortunately even a modest discussion of the Lebesgue theory of measure and integration would either be very lengthy, orif itis too shor, potentially misleading. We mention that a function f : R, — R is measurable if and only if L(t) is the limit of a sequence of staircase (or piecewise-constant) functions at all r except those belonging toa set of measure zero. A reader who has no prior acquaintance with the Lebesgue theory of measure and integration can simply think of all functions encountered below as piecewise-continuous functions, and of all integrals as Riemann integrals. This ‘would lead to no conceptual difficulties and no loss of insight, except that occasionally some results from Lebesgue theory would have to be accepted on faith 4 Definition For each real pe, ), the set Ly[0, =) =p consists of all measurable functions f():R, > R such that 5 [is@iarce The set L..{0, ©) bounded! on |0, ©). Thus, for pel, =), [0 =) =Ly denotes the set ofall measurable functions whose p- th powers are absolutely integrable over (0, 2), while L..[0, -°)=L.. denotes the set of essentially bounded measurable functions. 6 — Example The function f (1) =e", o> 0, belongs to L, for all pe[1, tion g (= IAt+1)belongstoZ, foreach p > I butnotto L. The function » = consists of all measurable functions f(): Ry —R that are essentially ]. The func ut does nor belongto L, for any qp. Ly —>R. isdefined by where pe [1, «), belongs to the set L, 7 Definition Forpe (1, -), the function II %® fyora 8 fe" The function \ qi Ln», is defined by ' "Essentially bounded” means “bounded except on a set of measure zero.” The reader need not worry unduly abou the distinction between “bounded” and “essentially bounded 2 The nvsation “ess. sup” stands for “essential supremum,” ie, supremum except on 2 set ofSec. 6.1 L, -Spaces and Their Extensions 223 9 fora =ess.sup. IFO. Definition (7) introduces the functions M-II, for pe 1, 2}, which map the set L, into the half-line R,.. Note that, by virtue of Definition (4), the right sides of (8) and (9) are well- defined and finite for each f)E Ly. 10 Fact (Minkowski’s Inequality) Let pe{1,ol, and suppose ft).g()€L». Then S480) Ly and 1 fC) +90lp S MPODIly + Hay For adetailed proof, see Royden 1963). 12 Fact Foreach pe[1, «the pair (Ly, W-ll,) is a normed linear space in the sense of Definition (2.1.8). This fact follows readily from Minkowski’ inequality 13 Fact Foreach pelt, =, the normed linear space (Ly, I ly) iscomplete and is hence ‘Banach space. Forp =2. the norm Ix corresponds tothe inner product 4 =| £08 de ‘Thus L isa Hilbert space under the inner product of (14). + Fact (13) brings out one of the main reasons for dealing with Ly-spaces in studying input-output stability. We could instead work in the spaces C,0, =) of all continuous func- tions fin L,, and of course the pair (C,, t-ll,) is also a normed space. However, itis not complete except forthe special case of p =. Now some results in input-output theory, such as those on the well-posedness of feedback systems, require that the problem be set up in a Banach space so that one can apply results such as the contraction mapping principle. In such @ case, it is preferrable to work with the L,-spaces. Note that, for each pe {1, e), the space Lis precisely the completion of the space C,. in other words, each function in , can be approximated arbitrarily closely by a continvous function, provided p R. Then for each TeR,, the function fri, -R is defined by S(0, OStST A HO=9 Ter andis called the truncation of fto the interval {0, T) 22 Definition The set Ly, consists of all measurable functions f : R, > R with the pro- erty that fr€ Ly forall finite T, and is called the extension of L, or the extended L,-space. Thus the set Lp, consists of all measurable functions f which have the property that every truncation of fbelongs to Lp, although fitself may or may not belong to Ly. Itiseasy to see that Ly isa subset of Lp 23 Example Let f() be defined by fi ‘Then, for each finite value of T, the function f; belongs to all the spaces Lp, for each pel, =}. Hence the original function fbelongs to Ly foreach pe 1, =]. However, fitself does not belong toany of the unextended spaces Ly. Ml ‘The relationship between extended and unextended spaces is brought out in the next Jemma 24 Lemma Foreachpe(1,=), the set Ly isa linear vector space over the real numbers. Foreach fixed p and fe Lye (i) \frllp is a nondecreasing function of T, and (i) fe Ly ifand ‘only ifthere exists afnite constant m Such that 25 ilfylly sm, VT20. Inthiscase,Sec. 6.1 L, Spaces and Their Extensions 205 26 fly = fim ‘fry “The proofis almost obvious and is left as an exercise. Thus, in summary, the extended space L, is alinear space that contains the unextended space L,, as asubset. Notice however that L, isa normed space while Le is not. In order to deal with mult-input multi-output systems, we introduce the set L3 which consists ofall n-tuples f=(fy +" fl’ where fie L, foreach i, The normon Lis defined by ci ” 2 ust] Susu In other words, the norm of a vector-valued function is the square root of the sum of the ‘norms of the components ofthe vector. Thisdefinition is to some extent arbitrary but has the advantage that L4 is a Hilbert space with this definition. The symbol L', is defined analo- gously. Note that hereafter the symbols if;ll, and ilfll7, are used interchangeably. Also, the same symbol I), is used to denote the norm on L forall integers. 6.1.3 Causality ‘This section is concluded with an introduction tothe concept of causality. If we think ‘ofa mapping Aas representing a system and of Afas the ouput ofthe system corresponding to the input fthen a causal system is one where the value of te output at time r depends only ‘onthe values ofthe input up to time Tis ismade precise nex.? 28 ' Definition A mapping A:1, +L, is said to be causal i 29 (Af) =Afyn, VT20, VEL. ‘Analternative formulation of causality is provided by the following. 30 Lemma LetA:Lt, >Lf.. Then Aiscausal in the sense of Definition (28) ifand only Bf, BEL fe, fr=8r> (Af r=(Ag)r. VT2O Proof "If" Suppose A satisfies (31). Then it must be shown that A satisfies (29). Accordingly, let fe L?, and T20 be arbitrary. Then clearly fy = (fr)r, so that by (31) we have 5 Throughout this chapter, forthe most part bol-faced symbols are not use, since the various ‘quate can be either Scala or vectors276 Input-Output Stability ch. 32 A= (Afr Since the above holds for all fe Lf andaall 720,A is causal. "Only if” Suppose A is causal in the sense of Definition (28). Suppose f, ge L. and that f= gr for some T 20; it must be shown that (Af)}r=(Ag)r. For this purpose, note that, by 29), wehave 33° ANr= (Afr. (Ag)r=(ABrr. Moreover, since f= g7;(33) implies that 34 Afr= (Ag). which was the thing tobe proved. i Definition (28) and Lemma (30) provide two altemative but entirely equivalent imerpretations of causality. Definition (28) states that a mapping A is causal ifany trunca- tion of Af to an interval (0, 7} depends only on the corresponding truncation fr. To put it another way, the values of (Af \t) over (0, 7 depend only on the values of f(t) over (0, 7) Lemma (30) states that A is causal if, whenever two inputs are equal over an interval [0, T], the corresponding outputs are also equal over the same interval. Problem 6.1 Determine whether each of the following functions below belongs to any of the unextended spaces L, and to any of the extended spaces Le! (@ f@=texp-o & F@=tane, ©) f@=exp(). 0, ift=0, @ sO=4We, if0 Lye has the form, capyo= [ae os (@de Show thatthe operator H is causal if and only ifSec. 6.2 Definitions of Input-Output Stability m A(t, )=Owhenevers lly ll, SYp lly Hy R is L,-stable with finite gain and zero bias (wb) if i is L,-stable, and in addition there exists afinite constant fy such that 5 Gye R xeLy > lly lly SY lle, Since the abbreviation "wfgazb" is unpronounceable, let us form for "with finite gain and zero bias." One might interpret "wb" as "without bias.” Remarks 1. Clearly L,-stability wb implies implies Lp-stability wig, which in turn implies L,-stability. Example (7) below shows that these three concepts are indeed independent. 2. Suppose that, fora particular xe L,, no y€ Ly exists such chat (x, y)eR. Then the Conditions (3), (4)and (S) are deemed tobe satisfied vacuously.278 Input-Output Stability ch.6 6 Definition Suppose A: Lf, ->L. Then the map A is said tbe Ly-stable (wf, wb) if and only ifthe corresponding binary relation Ry on Li defined by (1) is stable (wf, wo) 7 Example Consider the functions f, g, hR ~> R defined by so) g()=r+1, hr) =log (1 +7?) Define an operator F: Ling ~> Lng bY (Fay=f[e(), Wr20, and define G, HL. ~ Lg analogously. Then F is L_-stable but not L..-stable with finite gain. Note that xe L., implies that Fe L., but no constants yb, can be found such that (4) holds. This is because the function f(r) cannot be bounded by any straight line of the form yr +6; see Figure 6.1. Gis L..-stable wig, but not L_-stable wb, since G(0)+0. H is L..-stable wb. Fig.6.1 8 — Example Consider the mapping A defined by (Af =f exp-att-D1f (edt, Wr20, a ‘where a> Oisa given constant, and suppose it is desired to study the L_-stability ofthis sys- tem. Firstit is established that A maps the extended space L... into itself, so that A isin the class of mappings covered by Definition (6). Accordingly, suppose fe Log. Then IIfr!l. is finite forall finite 7. Hence, for each finite T, there exists a finite constant my such that 171 Smrac., Wre[0 TH Here the term "a.e.” stands for "almost everywhere," i. hereafter this phrase is omitted, it being implicitly understood. To show that g suppose Tis some finite number. Then except on a set of measure zero; {fe LinesSe0.6.2 Definitions of Input-Output Stability 279 Igiol sfew Loe-a) if (@yidt SJ mpexp{-ar—1)] dt se Hence g(-) is bounded over (0, 7]. Since the argument can be repeated for each 7, i follows that ge Lge Ling > Lng [Next, itis shown that Ais L.,-stable wb in the sense of Definition (6). Suppose fe L... ‘Then there exists. finite constant m such that If (@)I Sm, We20. Using exactly the same reasoning as before, it can be shown that laf (nls, v2 laf SZ. WF20. Hence (5) is satisfied with y,, = 1/a, which shows that is L..-stable wb. i 9 Example Considerthe system whose input-output relationship is APO =f expos (at This mapping A also maps L, into itself. To see this, et fe Ln. Then, for each finite 7, there exists afinite constant m such that If (1 Smp, VtE[0, T). ‘Thus, whenever fe (0, 7}, we have 10 Asya)! 1%, is causal and Ly-stable wfg, and choose constants Yp ‘and by such that 12 AKI, Spilxll, +p, WE LI. Then 13 WAT My, Sfp Ilxllzp +Bp, VT2O, Vie LI, Proof Given arbitrary x¢Z7,,7€R,,n0te that xr< Lp. Hence AryeL?,and 14 WAxy lly Sp har, + bp =p IX ly +p However, since Ais causal, (Ar) =(Ax)r. Hence IS WAL My, = MAx)p y= Arp lly, S WAXp ly Now (13) follows from (14) and (15). i ‘The analog of Lemma (11) for Ly-stable operators wb follows simply by setting b= 0. Note that one can also define a notion of causality for elations, and prove a result similar to Lemma (11); the details are left tothe reader. 16 Lemma Suppose A:Lf,—>L7, is linear, Then Ais Ly-stable wf if and only if itis Ly-stable wb. Proof "If" Obvious.Sec. 6.2 Definitions of Input-Output Stability 281 “Only if" Suppose A is Ly-stable wg, and select constants Yp, By such that (12) holds. Letkbe any number. Then, by (12) and linearity, 17 WAC, S7p Ike llp +p. 18 Ax ly SYpllxlly + (0,/K). Letting k +00 shows that Ais [stable wb. Ml Next, the notion of Z,-gain is made precise. Upto now, ifa relation Ris L,-stable, then ‘Ypcan be any constant such that (4) holds. Now this arbitrariness is removed. 19 Definition Suppose Ris a binary relation on Li. If is Lp-stable wfg, then the Ly gainof Ris defined as 20 wR) inf¥p:1b, 20 such that (4) holds). IfRisLy-stable, then the Ly-gain with zero bias of Ris defined as 2 ¥,(R) inf ¥p:(5) holds). Note that both quantities are denoted by %»(R), the context making clear which is ‘meant. IfA:L%, LR, then the quantity 7,(A) is defined tobe (Ry). ‘Thus far we have discussed what might be called “open-loop” stability. But one of the main applications of input-output theory is the stability analysis of feedback systems, ofthe form shown in Figure 6.2. In this system, it is assumed that u,¢,,y2 are vector-valued functions with n components each, and that u2, €2. are vector-valued functions with na components each, To be specific, suppose pe [1, =] Fig.6.2 22 wyeryreLpls 23 ur er, ¥sELpe ‘Then the overall feedback system is described by the equations282 Input-Output Stability cho WA ey sus yn ep m2 bY = Ger yr ‘These equations can be written in a more compact form, Let n=n,+n2, and define [:be-b re Define maps G, FL}, —>L5, by 28 “he G0 % G=(9 g, and note that F is a constant matrix, known as the interconnection matrix. Then the equa- tions ofthe feedback system can be expressed as Me ~Fy, ‘On the basis of (27), we can define two binary relations Rye and Ryy on Le. To define Rugs Climinate y from (27) to obtain 28 e=u-FGe, or(1 + FG)e =u Todefine Ry, eliminate e from (27)to obtain 29 y=G(u-Fy). Note that we cannot in general expand G (u ~ Fy) as Gu~GFy, since G need not be linear. Now the binary relations Ry, and R,y are defined as follows: 30 Re={(u, ee Le: BL Ry =H, yeLP: 32 Definition The feedback system (27) is Ly-stable (u-e) ifthe relation Ry is Ly-stable; itis Ly-stable (w-y) ifthe relation Ry is Ly-stable, Finally, itis L,-stable (u-e-y) ifbth Rye and Ruy are Ly-stable, ‘The next lemma shows that we do not need quite so many concepts of stability. 33 Lemma The following three statements are equivalent: () Reeisly-stable.See.62 Definitions of Input-Output Stability 283 Gi) Reis Lp-stable. Gil) BothR,, and Ry, are Ly-stable. Proof Clearly itis enough to show that (i) and (ii) are equivalent. ()> Gi): Suppose (i) is true, ie., that Rye is L,-stable. Let (u, y)€ Ry be arbitrary, ‘Then, from (27), the ordered pair 34 (Uu-FY)E Re Now suppose wé Lf (not Lf,). It is desired to show that ye L$. For this purpose, note that ue L} and the L,-stability of Ry together imply that 38 Fel. Butsince Fisjusta constant nonsingular matrix, (35) implies that 3% y “Mu —ee Ly, Since the same argument can be repeated forevery w€L,itfollows that 37 u yeL%,uelp> yelp Inother words, Ry, isLy-stable. Gi) > (i): The proof is quite similar to the above. Suppose Ruy is Ly-stable, and suppose (4, CVE Rye. Then (ue, Gee Ruy. If we Lf, then the L,-stabitity of Ry, implies that Gee Lj. Since F is just a constant matrix, this in turn implies that FGeeL’. Finally e=«-FGee L}. Hence RyeisLp-stable, Ml Lemma (33) allows us to say simply “the feedback system is E,-stable” without spect- fying whether we mean u-e,u-y, oF u-e-y. Itislftto the reader to modify Definition (32) to define L, stability wfg and wb, and to prove the analogs of Lemma (33). In Lemma (33), the proof of the implication (ii) > (i) depends only on the fact that Fis ‘constant matrix, whereas the proof of the implication (i) > (ii) depends on the fact that Fis constant nonsingular matrix. The nonsingularity of F is a special feature of the feedback configuration, and is not true in general. For instance, consider the system shown in Figure 6.3, known as a multi-controller configuration. In this case, orl =|-100), rool which is singular, Hence, in an arbitrary interconnected system, u-y stability implies we stability, but the converse need not be tue.284 Input-Output Stability che Fig.6.3 ‘Now thatthe stability definitions are out ofthe way, we can at last discuss the reason for introducing relations, and speaking of the stability of relations (rather than operators), The reason simply is to divorce the notion of stability from that of well-posedness. Stability is defined above. Roughly speaking, well-posedness corresponds to the existence and unique- ness of solutions for e and y for each choice of u, though ane may wish to add requirements of causality, continuous dependence, etc. Thus a possible definition of well-posedness is that (/+FG)"! isa well-defined causal map from L%, into itself. This would allow one to solve (28) for eand write 39 e=(+FG)"u But well-posedness places no restrictions on stability. Conversely, u-e stability of (27) ‘means that, for each we Lt, ifany e€ Li. satisfy (28), then such e must in fact belong to L?. If, fora particular we L}, noe Lj satisfies (28), then this condition is deemed to be satisfied vvacuously. In this way, stability and well-posedness become independent concepts. This is desirable since well-posedness can be ensured under quite mild conditions, whereas stabil- ity is more difficult to analyze. Roughly speaking, the feedback system is well-posed if either Gor Gz contains some element of smoothing; see Vidyasagar (1980a) or Saeki and ‘Araki (1982). 6.3 RELATIONSHIPS BETWEEN /0 AND LY APUNOV STABILITY InChapter 5 we studied Lyapunov stability, which is defined for unforced ordinary dif- ferential equations; there is no input, and the system evolves under the influence of a nonzero initial state. In contrast, inthe present chapter, the word "state" isnot mentioned at all, and attention is focused on the influence of inputs upon outputs. It is therefore ‘worthwhile to relate the two types of stability, and to underline the point that both approaches tackle different facets of the same underlying issue. This is the objective of the present section. We begin with a discussion of open-loop stability, and then study feedback systems. Let us begin with discussion of linear systems of the formSee.6.3 VO and Lyapunov Stability 285 1 X)=Ax() + Bute, 2 Cxtt)+ Dut, where A, B, C, Dareall constant matrices. The transfer matrix ofthis system is 3 AGs)=Cist-Ay'B+D. The system (1-2) is said to be stabilizable if there exists @ matrix K such that A—BK is Hurwitz, and is detectable if there exists a matrix F such that AFC is Hurwitz. As for the twansfer matrix H,it will be seen in the next section that it represents an L.-stable (ie. BIBO stable) system if and only if all pols of H have negative real pars. Now the follow- ing result is well-known (Kailath (1980, Chen (1986)} 4 Theorem Suppose the system (1-2) is stabilizable and detectable. Under these con- ditions, the system is L..-stable ifand only if the associated unforced system 5 =AK) is (globally) asymptotically stable. For an extension of Theorem (4) to linear time-varying systems, see Silverman aad ‘Anderson (1966) for the continuous-time case and Anderson (1982) for the discrete-time case, ‘The converse of Theorem (4) iseven simpler. 6 — Theorem Suppose the system (5) is asymptotically stable. Then the system (1-2) is Ly-stable for eackpe 1, =| Now letus consider nonlinear systems ofthe form 7 (2) =Alt, XC), WC), YO) = Be HCO), WED), VE 20. Suppose f: R,xR"R" — R" and g: R,xR"xR” — R’ are unbiased in the sense that 8 £1,0,0 R(t, 0,0)=0, Vs 20. This ensures that O isan equilibrium of the unforced system 9 xm t, x(9, 0), Vr20. To make the theorem statements more compact, some notation and definitions are now introduced. The first one is a "local" version of Definition (6.2.2). 10 Definition The system (7) is small signal L,-stable wb ifthere exist constants r, >0 and", <0 such that286 Input-Output Stability ch 11 x(Q)=0, lull yeL and ll, <7 lll, [Note that, even if llu(t){ sr, ¥r20, (which is equivalent to saying that lIul,.<7,), the nomm iluil, can be arbitrarily large if p 0 such that, for each xe B,, there exist a time 1* and an input u* such that lNu* tS BCUNII) ands(t%, 0,0, u*) =x. tis globally reachable ifthe preceding statement holds for allxe R ‘The conditions of the definition mean simply that every state x in B, can be reached within a finite time by applying an input whose L..-norm is bounded by Bix!) 13 Definition The system (7) is uniformly observable if there exists a function cof class K such hat, with u(1)=0, we have 14 ligh-,s¢-, 0, x, 0)]!2 2a xl). ‘The inequality (14) can be stated more concisely (but less precisely) as lly ll, 2a xi). If y(J€L4, then (14) is deemed to be satisfied since, loosely speaking, llyllz=e., For a linear time-invariant system, uniform observability is equivalent to the standard notion of observability. This is because, with zero taput the output depends only ‘on the initial state. Ifthe system is unobservable, then there exists a nonzero initial state ‘which (with zero input) will produce an identically zero output, so that 14) is violated. On the other hand, ifthe system is observable, then itis not difficult to show that (14) is satisfied, Now the main theorems relating input-output stability and Lyapunov stability of the system (7) are presented. Theorem (15) isa nonlinear analog of Theorem 6). 1S Theorem Suppose 0 is an exponentially stable equilibrium of (9), that f is C', and that f, gare locally Lipschitz continuous at (0, ), ie., suppose there exist finite constants key igs r5uch that 16 fe, x,u)—flt 2, v)Il SkpLx=ail + Hu vil], ¥i20, WOx, WC, WEB, 17 gd, x,u)— gi, z, vy Sk, [lIx—2il + lu—vil], Wr20, Vex, U).(2, EB, Then the system (7) is small signal Ly-stable wb for each pel l,l. If 0 is a globally ‘exponentially stable equilibrium, and (16) and (17) hold with B, replaced by R"*™, then the system (7) is Ly-stable wb forallpel, =).See.6.3 VO and Lyapunov Stability 287 Proof The condition (16) implies that 18 Dal, x,w)lIsky, W420, V(x, we B,, where Daf denotes the partial derivative of f with respect to its second argument. Thus all the hypotheses of Corollary (5.7.77) are satisfied, and there exist a C’ function V:R, xR" — Rand constants a, B, ¥, s > Osuch that 19 oPIXI?SV(, x) SBP IIXI?, VyCt, x)<—IlxI, We20, Vee B,, 20 1D2V(, x)IISyixll, Vr20, VxeB,, where A Vylt, x) =D,VEt,)+D.VG 2M, 9) denotes the derivative of V along the trajectories of the unforced system (9). Let8=min{7, s), and suppose uis fixed input with the property that n tm ei ’ and suppose x(0)=0. Evaluate the derivative of V{Z, x(1)] along the trajectories of the forced system (7), and denote itby Vj. Then 2B Vt, x)=D, Vit, )+D2V (4, x)KC, xu) =DiV(t, x) +D2Vt, ft x, 0)+D2V(t, x) f(t, XW) MU, XO] =Vult, x) +D2Vt, x) fC, x, w)—£0, ,0)] Sothell? +ykyilxil hall if, we Bs. ‘where in the last step we use (16), (19) and (20). Now, since x(0)=0, there exists a time T>0 such that x(1)¢B, for all re [0, T). Moreover, the right side of (23) is negative when- ever lixil > yk;llull. Hence one can easily show that, 2 Vie x(OIS, max. Vie.x SBP Aj We20, vx yp From (24) and (19) it follows that288 Input-Output Stability ch. _ BTAH 25 xs C2 cmin(r, 5) ‘This last observation removes the circularity in the argument, and shows that ix)! Smin{,, 5} 'VF20. Now, from (19) and (23), it follows that hy Live xeon)? wen. 4 ean 26 GWU MON) S— Fe Vie xO} G Let W(0)=V{t, x(0)]". Then W (1) is differentiable except when x(¢)=0, and is direction- ally differentiable even there. Hence the one-sided derivative jim Warm) We a0 h exists for all 120, For notational convenience the subscript "+" is dropped hereafter. Now V = 2WW;hence it follows from (26) that ky 28 awis- we hw io, or 1 hy 29 Wit)<-— SW + fa(e il, We20. (Sapp WO+ Fe at Let (1) denote the solution of » hos age hw Then (29) implies that W ()', which in turn impliesthat W()e Ly. Since vy Hu() Ih, A(O) = W (0). a aE! 32 xD s (Vie, xiO)'2/a= Winyve, itfollows that x(-)e Lf. Finally (17) and (8) imply that 33 Hy I= Nghe, x). aC] Sky LxCe) + aCe, ‘whence y()e Li.See,6.3 VO and Lyapunov Stability 289 To complete the proof of small signal L,-stability, it only remains to demonstrate the existence of a constant 7, such that (11) holds. For this purpose, note that the inverse Laplace transform of gof G1)is hy gry exp(-12B2)EL). ‘Therefore, 38 Ug = 7kpAB oe Hence, by (6.4.31), itfollows from (30) that 36 hil, Osuch that 40 ull. yeL4 and lly Sp lulls Now, since (7) is reachable, there exist a constant > 0 and a function 8 of class K satisfying the conditions of Definition (12). Choose 5 >0 such that (8) < r2, and let x9 Bg be arbi- trary. Then, by definition, there isan input u() with lull. », which in turn shows that x(t) —> 0 as too, i 44 Corollary Suppose the system (7) is autonomous, globally reachable, and uniformly observable, Under these conditions, ifthe system isL-stable, then Visa globally attractive equilibrium of 9). ‘The proofis analogous to that of Theorem (39) andis thus omitted. ‘Thus far attention has been restricted to “open-loop” systems. Next, Theorem (15) and Corollary (44) are used to relate the extemal and intemal stability of the feedback system shown in Figure 6.4, which isthe same as Figure 515. This system is described by Liat Fig.6.4 45 X()=AX(1) + Be(s), ¥(0) = CX(0), e(1)=U(0)- Ot, YO). where x(1)e R®, ue", y(DER', and A, B, C are matrices of compatible dimensions; and@: R, xR! > R” satisfies (, 0)=0:20, 46 Theorem Consider the system (45), and suppose ® is globally Lipschitz continuous; i.e. suppose there exists a finite constant such that 47 NOE, y)—OE, y2)IE Sly, ~Yoll, ¥120, Vy), YER" Under these conditions, if the unforced system is globally exponentially stable, then theSec.6.3 VO and Lyapunov Stability 291 Jorced system is Ly-stable wb for all pé1, =]. Suppose the pair (A, B)is controllable, and the pair(C, A) is observable, Under these conditions, if the forced system is L-stable, then x=0isaglobaliy attractive equilibrium of the unforced system. Proof ‘The system (45) can be written as 48 X(0)=Ax(1) -BOLr, Cx(N] + Bun, First, suppose (47) is satisfied. Then the right side of (48) is globally Lipschitz continuous in xand u. Hence, by Theorem (15), ifx=0 is a globally exponentially stable equilibrium of the unforced system, then the forced system is L,-stable wb forall pe [1, =] Next, suppose the forced system (48) is Lp-stable. Iftcan be shown thatthe system is reachable and uniformly observable, then the global atractivity of x=0 would follow from Corollary (44), To show that (48) is reachable, let XR" be arbitrary. Then, since (A, B)is controllable pair, there exits a finite time and a continuous function e(),r€ (0, ] such that the resulting solution of %= Ax+ Be starting at x(0) = satisfies x(¢*)=X9. Now apply the control signal 49 wi =e(r)- Of, Cx) to the system (48). Then itis clear that once again we will have x(7*) =X. Showing that Hl, is bounded by a function of the form B(x) is easy and is let as an exercise. To prove thatthe system (48) is uniformly observable, suppose u(t)=0. Then, since (C, A) is an observable pair, there exist atime Tand constantsa, 6 > Osuch that 50 ani? tate? ayn trae “The proof ofthis statements not dificult andis eta anexercise. Hence 51 alle -+6lly3> bx(O)I7 Now note that ifu=0, then e( Pfr, y(4)]. Since 52 NOL, yin) Spilyie)l?, We 20, it follows that $3 Mell} sp? ily 3. Combining (50) and (53) shows that $4 ix(O)l? (ap? +b) lly This shows that (48) is uniformly observable. Now Corollary (44) enables one to conclude that the equilibrium x=0 is globally attractive. i292 Input-Output Stability ch.6 Problem 6.4 Prove Theorem (15) in the special case where f and g are autonomous, andfis C?, as follows: Define x0 Using Theorem (5.8.1), show that A is a Hurwitz matrix. Define the higher order “remainder term” f(x) =f00)- Ax Then rewrite (7) in the form, HC) = ANC) +f EXO) + [AIX(2), WO] F0XC0), 09} Interpret this equation as the exponentially stable system x= Ax driven by an input which itself depends on x. Construct an implicit inequality bounding x(1) and then use Gronwall's inequality {Lemma (5.7.1)} to get an explicit bound on x(1). Problem6.5 Repeat Problem 6.4 forthe case where fand g might be time-varying by using Bellman’ inequality instead of the Gronwall inequality. Bellman's inequality (which isan extension of Gronwal's inequality) is as follows: Suppose a(), b("), c(°) re continu- ‘ous functions defined onthe interval (0, =), and that b(¢), c(t) are nonnegative forall 120, ‘Suppose that the function x() satisfies the implicit inequality w(nsat+ bof e(ul@dr, ¥e20, Then 1 [i ] uinsacs bof ecmacmerp] fbUreisrds| de, ¥120. : it Problem 6.6 Extend Theorem (6.3.46) to time-varying systems. 64 OPEN-LOOP STABILITY OF LINEAR SYSTEMS. Before attempting t0 study the stability of interconnected systems such asin Figure 6.2, it is helpful first to obtain conditions under which the operators G, and Gz represent L,- stable subsystems, In this section, we concentrate on linear systems and obtain necessary and sufficient conditions for a linear system to be stable. The term open-loop stability refers to the fact that we study the subsystems G, and G> individually, rather than the ‘overall closed-loop system, which is described by (6.2.24).Sec.6.4 Open-Loop Stability 293 64.1 Time-Invariant Systems Throughout most of this subsection, atention is restricted to single-input, single- coutput (SISO) systems. Once the SISO case is thoroughly analyzed, the results for the “MIMO (mult-input, multi-output) case follow easily. Consider a SISO time-invariant sys- ‘em, which is characterized by a scalar transfer function h(s). IC i(s)isarational function of 5 (i¢.,a ratio of two polynomials ins), then it is well-known that such a system is L..-stable (BIBO stable) if and only if 1. Avis aproper cational fonction (ie the degre ofthe mentor polynomial of fises than or equa that ofthe denominator polynomial) and 2. Allpolesof i have negative real parts. However, the situation is more complicated if f(s) is not rational. Such a situation arises whenever h() is the transfer function of a distributed system, such as an RC transmission Jine integrated circuit), an LC transmission line (power line), or if A>) represents a simple delay, etc. In what follows, precise conditions ae given under which a scalar h(s) (rational or irrational) represents an L,-stable system. These conditions illustrate one of the chief advantages ofthe input-output approach to stability, namely, that it places lumped systems {rational fs)] and distributed systems [irrational h(s)] in a unified framework; this i much harder to achieve using Lyapunov theory. To do this, the sets A and A are introduced. Basically (as shown later), A is the set of BIBO stable impulse responses, while A is the set of BIBO stable transfer functions. The precise definitions are given next. 1 Definition The symbol A denotes the set of generalized functions (distributions) f) such that f (t)=Owhent <0, and have the form 2 f=LHdu-1)+flo.if 20. % where 8) denotes the unit delta distribution, Oty are constants, f,() is a ‘measurable function, and in addition, 3 Ebfilse [ipod Thenorm If, ofadistribution in A is defined by 4 fell Evils sora & The convolution of two distributions fand g in A, denoted by f+ g, isdefined by294 Input-Output Stability ch. [ro-necmat=[seee-oae 5 feane ‘Thus A consists of all distributions that vanish for ¢< 0, and forr> Oconsist of a sum of delayed impulses and a measurable function, with the additional property that the weights of the impulses form an absolutely summable sequence and the measurable function is abso- lutely integrable. One can thinkof A as the space L ,[0, °) augmented by delayed impulses. [Note that, in computing the convolution of two distributions, one should take 6 BU-1) + 80-0) =5(r-t-0), 7 BU-DALO=L-0. In other words, the convolution of two delayed unit impulses with delays of t and 0 respec tively is another delayed unit impulse with delay +8, while the convolution of a delayed unit impulse &(r~t) and a measurable function f,() isthe delayed measurable funtion Jolt ~1). Thus, giventwo elements f,g in A ofthe form 8 FO=TLRE-P) +f, @)= Fes Ht) + 800), 3% & their convolutionisgivenby 9 Fean=¥ Ffig de") 3B +E fiett—i+ Eeytse-ap ref hee Dade % Bi 10 Example The function Su=exp(-at) belongs to and hence to A, whenever o> 0. The distribution n()= = —\ &t- iF), T> Ogiven, ELE NEA which isa sequence of evenly spaced delayed impulses, belongs to A because the sequence weights (/(i+1)*} isabsolutely summable. However, the distribution f= ¥ rh 8a~iD), T> Ogiven, does not belong to A because the sequence {1Ai+1)) is not absolutely summable, TheSec. 6.4 Open-Loop Stability 295 distribution falt)= 5) +exp-0) belongs oA and If Remarks 1, Note that L, isa subset of A; further, if fe Ls, then " 2. If f. ge A and atleast one of them is in Ly (i.e.. does not contain any impulses), then f « g does not contain any impulses. This is clear from (9). Itis shown subse- quently thatif fe L,, ge A,then fw g¢L,ie..L, isanideal in A, ‘As mentioned previously, the set A can be interpreted asthe set of BIBO stable impulse responses; in other words, a system with an impulse response h(-) is BIBO stable if and only ifh()eA. Toprove thisimportant result, we first derive some useful properties of A. These properties imply that A is a Banach algebra with identity, and that it has no zero divisors. However, a reader who does not know what these terms mean need not worry about terms say nothing more than Lemma (12)) 12 Lemma The set A, together withthe function Wil, and the convolution +, has the following properties: (@__ Il-ll,isanormon A,and A is complete under this norm. ‘The convolution operation is commutative; ie., 13 fegagaf. Vf ged. Gii) The convolution operation isbilinear;ie., M4 fe(oe IS falgth)=fegtfeh, Vf g. hea (fg), VOER, Vf. ge, (iv) Whenever f, ge A, wehave that f ge A,and infact 16 fe gta Sifllalig lla () (isthe unitelementof A;i.e., 17 Befafed=f, VfeA.296 Input-Output Stability cho (vi) Anas no divisors of the zero element; i, 18 fxg =0>f20org =0. Proof (outline) (i) Itis easy to verify that I'll, is indeed a norm on A. The complete- ness of A under thisnorm is more difficult to show, and the fact is stated here without proof. and (ii) are obvious. (Gv) Suppose f, ge A are ofthe form (8). Then Dis, 8-H -19) WG re= +E final W+E 8 LO WP +] KU-dalode. & ia 3 ‘The first term on the right side represents the distributional part of fg, while the last three terms represent the measurable part of f +g. To compute iif gll,, we take each of the terms separately. First, Ee! a 2» BE rvene[ Bun mis = Next, mi oe u [[Zfsee-W aise Nf Neate Nae E a [é uses dt 2 [iEece- Ee] {san 7 i Finally,Sec.6.4 Open-Loop Stability 297 4 lon, 2B {fuse code drs [Yaesu dra e ee Yfelt—D gall de de j Since eee at = [iso dt {untae a ‘Note thatthe order of integration was interchanged after the first step. Putting together the four bounds (20)-(23) proves (16). (¥) is obvious. ‘The proof of (vi) is beyond the scope of this book. The reader is instead referred to Hille and Phillips (1957), Theorem 4.18.4. Suppose fe A. Then, whenever Re s 20, the integral mw +fel8) = [fnew d= Efe” 3 & converges andis well-defined. Therefore, all elements of A are Laplace-transformable, and the region of convergence ofthe Laplace transform includes the closed right half-plane 2 C, s:Res20). Now the set A can be introduced. 2%6 Definition The symbol A denotes the set of all functions f:C., -+C that are Laplace transforms of elements of A. Thus, according to Definition (26), "fe A" is just another way of saying that the inverse Laplace transform of fbelongs to A. When we deal with feedback systems, the symbol A comes in handy to keep the notation from proliferating 27 Lemma Suppose f(s)isa rational function ofs. Then fe A ifand only if,298 Input-Output Stability Cho () Fisproper, and (ii) allpoles of fhave negative real parts. Proof If (i) and (ii) hold, then it is clear that /(-), the inverse Laplace transform of f, consists of two parts: a measurable function which is bounded by a decaying exponential, and possibly an impulse at time 1 =0. Hence fe A. ic, fe A. To prove the necessity of these conditions, suppose (i3 does not hold. Then f() contains higher order impulses and there- fore does not belong to A; if Gi) does not hold, then the measurable part of ft) is not abso- lutely integrable. Having defined A, we can define its extension A, in exactly the same way that one defines Ly from L, 28 Definition The set A, consists of all generalized functions fl) which have the pro- perty that allrruncations fr of fbelong to A, forall T20, andiscalled the extension of A ‘The set A, has some very useful properties. Most physical systems, even those that are “unstable,” have impulse responses that belong to A, [for example, consider h(t) =e'l Moreover, itcan be shown that if f(r) is any regular measure that vanishes for 1 <0 and has ‘no singular part, then the corresponding operator H defined by 2B UPKNn=fae-Vs~@de ‘maps Lye into itself for all pe (1, =] if and only if he A,. Thus systems whose impulse responses le in A, are the most general (linear time-invariant) systems that one needs to consider in the present context. Now the main results ofthis subsection are stated and proved. 30 Theorem Consider the operator H defined by (29), where he A,. Then the following four statements are equivalent: (i) Hisky-stable wb. (i) HisL-stable wb, Git) His L,-stable wb forallpe {1,9 (iv) he. Moreover, ifhe A, then BM Uh sfllp Salle Illy, Wpell. Remarks Theorem (30) brings out fully the importance of the set A. According to this theorem, a necessary and sufficient condition for a system of the form (29) to have any one ‘of various forms of stability is that the impulse response belong to A. This justifies the description of A asthe set of stable impulse responsesSec.6.4 Open-Loop Stability 299 Proof Itis first proven that (iv) implies each of (i), (ji) and (ii). Accordingly, suppose hea. Civ) > (Ds If fe L, then fe A, and in fact 32 Mflla= ify. Hence, by iv) of Lemma (12), h + fe A,and 33 Whe fila Shelly lf, = Wallaiifily. Next, because fcontains no impulses, neither does h « f, which means that h « fe Land 3M he fllg= lh ofl Now (33) and (34) together imply that 35 Wha fll, SAM If. which shows that His L-stable wb, (iv) > Gi): Suppose fe L..,and write he A inthe form 36 AW=Th Stn) +hgC % Then 37 he M=Lh Sen +f hfe, & 3 Bo Mh eAMOISE In fe-a)l +f ined fol de i ° Sess. sup f(D A nse = Nfl Walla Since (37) holds for (almost) all, we see that h + fe L..,and that300 Input-Output Stability cho 39 he fllans Hall llfl, This shows that His L..-stable wb. (iv) > Gil: This part of the proof is omitted as itis a special case of a more general result for linear time-varying systems; see Theorem (75). Now the reverse implications are proved. (> (ivy: Suppose (iis true, Since Hs linear i) implies that #1 i continuous. Now, since every element in A can be approximated arbitrarily closely in the sense of distributions by an element of £, this in tum implies that A maps A into itself. Now let f (¢ hese 1 Which by assumption belongs to A. This shows that (iv) is true. Gi) > (iv): Thisis a special case of Theorem (53). Gi) > Gv): Suppose (ii is true, ic. that His L,-stable forall pe 1, ]. Then. in par- ticular, His L-stable. Asshown above, this implies (iv). 40 Theorem Consider the operator H defined by (29), where he A. Then His L-stable, and 41 y(H)=sup| hijo)! Remarks The main purpose of Theorem (40) is to prove the bound (41), For an arbi- trary pe 1, 9}, we have the inequality (31), which shows that 2 ¥(H)S Uhl, Yell =}. However, ifp Proof Letg=h +f. Then gio) then the tighter bound (41) applies. ii(jeo) fie). Using Parseval’s equality gives LT iggepi2deo=—- f kgayi? yey? 43 ligil}= 57 [tage dw= 55 {rigay! 1fU@)!? da supliyorr J jue? ao=sptigey!? npn ‘Taking the square root of both sides of (43) shows that y3(H) can be no larger than the right side of (41). Proving that Y2(H) actually equals this quantity is messy but not difficult; the reader is referred to Vidyasagar (1981), Lemma (3.1.69), for the details. i ‘The results of hs subsection up to now are summarized below: 1. Thesets Aand A have been introduced,Sec.6.4 Open-Loop Stability 301 2. It has been shown that, in a very precise sense, A is the set of stable impulse responses and A is the set of stable transfer functions, 3. The useful bounds (41) and (42) have been obtained. Actually, it can be shown that the bound (42) is exact when p = 1 and, ic.,¥\(H)=7.(H)= Mh ly. How- ever, the proofs area bit involved and are not given here. The interested reader is referred to Vidyasagar (1981), Chapter3, This subsection is concluded with a brief discussion of MIMO linear time-invariant systems. The results in thiscase follow very easily from Theorem (30), Considera system with m inputs and / outputs, with the input-output relationship He~ soar, where the impulse response matrix H()¢ A. In analogy with the SISO case, such opera- tors are the most general linear time-invariant operators that one needs to consider in the presentcontext ‘The criteria for the L,-stability of systems ofthe form (44) are given next. 45 Theorem Consider an operator 1 of the form (44), where He Al". Under these conditions, the following four statements are equivalent: () His y-stable wb. Gi) HisL-stable wo, (iii) His L,-stable wh for allpe (1, = (iy) Hea ‘The proof is leftas an exercise, sine itis entirely paralleito that of Theorem (30). Basically, Theorem (45) states that the set of matrices whose elements all belong to A is precisely the set of stable MIMO impulse response matrices, whereas the set of matrices whose elements all belong to A is precisely the set of stable MIMO transfer matrices. ‘Theorem (30) leads to the useful bound (42). A similar bound can be obtained for MIMO systems and is given below without proof. In this connection, it is worthwhile to recall the definition (6.1.27) of the norm on L. 46 Lemma Consider an operator ofthe form (44), where H()e A. Then 47 (HS IM th. Where Illy, denotes the matrix norm induced by the Euclidean vector norm, and M ison Lemmatric whose i-thelementis Why ila. Ifp =2, then302 Input-Output Stability cho 48 (A= M22, where My is the bom matrix whose i-thelementis given by 49 (Mz), =supl hyo)! Iiseasy to verify that (47) reduces to (42), and (48) reduces to (41) in the case of scalar systems. 642 Time-Varying Systems In the previous subsection, the focus was on time-invariant systems. Inthe present sub- section, we study a class of operators that represents a natural generalization of those of the form (29). Specifically, we consider operators G ofthe form 0 GD= Taos ~wrsfatt ofoae Actually, since f (t) =O whenever <0, one can rewrite (50) as SL GOH & atosee-n) | alts Dfleae, ia where 21 ists) nother words, (51) is obtained from (50) by taking the sumamation only over those indices ¢ such that $1 ‘Theorem (53) gives necessary and sufficient conditions for an operator G of the form (50) tobe L..-stable. 53. Theorem Consider an operator G of the form 50), where S40 thy F lee Loe icy SS that DEL}, 20, 56 tp [lett Datel. Then G maps Lng intoitself. Further, Gis L..-stable wb ifand only ifSec. 64 Open-Loop Stability 303 to | Remarks Note that conditions (54) to(56) are quite easy to satisfy. For instance, ifthe set of indices I(t) is finite for each finite r, and if each function g,(?) is continuous, then (54) is satisfied. Further, the index set /(¢) will indeed be finite for each finite t, provided the delay 1,9 as i>, Similarly, ifthe function gq is continuous, then (55) and (56) are satisfied. Proof Its left as a problem to show that if (54) to (56) hold, then G does indeed map Lin into itself, 87 | z= IgA) +f Lge.) dt} To show that (57) is a necessary and sufficient condition for G to be L.-stable wb, we tackle first the sufficiency, since that is much easier. “IE Suppose (57) holds and hat fe L.. Then ' 88 NGHMOI =| E sin fee] sult 9A) dr} ier) 1 1 { zr als [late Ol de fll, [tin | Seu lflle Since the right side of (58) isindependent ofr, it follows that Gfe L... and that 89 NGfL Scull fll Hence Gis L._-stable wb, “Only if” We show the contrapositive, namely that if (57) does not hold, then the ratio NGftl Afi. can be made arbitrarily large. To simplify the details, itis assumed that all the 0 such that r(1)2>&. Fix this ,and define 62 flt)=sign lett, DI, Vte10. ), where the sign of Oistakenas0. Then 3 w(t, DAC)=1e(, D1, VEL, 4], and 4 GHU=[ et. vKodrrk. ‘Therefore 65 NGfll. =sup GAMO ZK ‘Since this argument can be repeated for any k, it follows that G cannot be L..-stable wb. i ‘The proof of the "only if" part leaves open the possibility that, if r() of (61) is unbounded, then Gis L.-stable though not L.-stable wb. However, itcan be shown, using the principle of uniform boundedness, tha’ for operators of the form (50), L..stability and Lastabilty wh are equivalent properties; see Desoer and Thomasian (1963) or Desoer and Vidyasagar (1975), Theorem (4.7.5). ‘The next theorem gives necessary and sufficient conditions for G tobe L stable. 66 Theorem Consider an operator ofthe form (50), where 61 thy Y bgt ti)leLon ere Bhogal, DEL), VE20, 9 thf let, DidteL i. Under these conditions, G maps L , into iself. Further, Gis L -stable wb ifand only ifSec.6.4 Open-Loop Stability 305 0 sup] F lacesipl +f leet, Ide] =e) , the index set /(#) eventually includes all i. Thus (86) is equivalent to requiring that he A. Similarly, the contition (70) for -stability becomes308 Input-Output Stability che 87 sup] Ethie] heel at] 0. ‘The proof of this important result is well beyond the scope of this book, and can be found in Hille and Phillips (1957), p. 150. But the necessity ofthe condition (2) is quite easy to see. If I/f belongs to A. then the function L/fls) is bounded over the closed right half- plane C, [defined in (6.4.25)]. But this is the same as |/(s)! being bounded away from 0 over C,, which is what (2) says. The sufficiency of (2) is, of course, considerably more difficult oestablish, but f(s) isrationa, then the sufficiency is easily seen (Problem 6.15). Anelement fe A iscalled aunitof A ifthere existsa ge A such that f+ ¢=8(0),ie.iff has a multiplicative inverse in A. In such a case we also say that fis a unit of A. Now [Lemma (1) gives a simple necessary and sufficient condition fora given fto be a unit of A, namely that (2) musthold he Lemma Suppose Fe A”. Then the function [F()1":5 + [F(s)/" also belongs to * ifand only if 4 int Ider Fs)! >0. Resz0 Remark Note that Lemma (3) allows us to determine whether the mairix-valued func- tion [F(T belongsto A”™” by examining the scalar-valued function det [()] Proof Since the determinant of F is obtained by forming sums and products of the various components of F all of which belong to A), and since A is closed under both of these operations, itfollows that A:=det Fe A. 'If Suppose (4) holds. Then, by Lemma(I), the function I/A belongs to A. Now write ~ Adj Fis), As) s (esr where Adj f(s) denotes the adjoint matrix of F, i¢., the matrix of cofactors of #. Now ‘Adj Fe A", since the components of Adj F are determinants of various submatrices of F, Hence, if Ve A, then [F(Q'e A “Only if” Suppose (F()'eA"™. Then det (F() ‘now implies that (4) holds. i AeA. Since Ae A, Lemma (1) 65.1 SISO Systems with Constant Feedback ‘Consider the system shown in Figure 6.5, where @(s) is the transfer function ofa linear time-invariant SISO system, and k #0 isa constant. There is no loss of generality in assum- ing that k #0, since if k= 0 then there is no feedback, and the overall system is stable if and only if eA. Suppose 1 + @(s) isnot identically zero (which essentially says that the feed- back system is well-posed); then one can explicitly write down the transfer matrix relating y tou. Indeed, wehaveSec.6.5 LTI Feedback Systems 31 1 [a -eg] [ “ale | 7 Let H denote the 2x2 transfer matrix in (6). Then the feedback system is (for example) BIBO stable ifand only ifthe output ye Z2, whenever the input we L2.. Since the system is linear and time-invariant, a great many stability notions are equivalent to the requirement that Fie A” [see Theorem (6.4.45)]. Hence, throughout this subsection, feedback stability istaken tomeanthat He A’ 7 Lemma He A” ifandonlyif 1 14g Proof "Only it Suppose HeA"?. Then hy» =A?e A. But since k0, this in um impliesthat 7e A. “If” Suppose 7e A. It is shown in tum that each of the four elements of HY belongs to A. First, zy = Fe A. Next, 9 Finally, 10 iy =hy hed, This completes the proof. i The quantity 1 +K@is often referred toas the retur difference. Thus Lemma (7) states that the feedback system is stable if and only if the reciprocal of the return difference is a stable transfer function, Now the challenge is to find some easily verifiable conditions for ensuring that (8) holds. Suppose A, i., thatthe system is open-loop stable. Then 1 +k also belongs to A. Hence, by Lemma (1), it follows that IAI +kg}e A (and the feedback sys- {emis stable) ifand only if AL inf 11 +86)! >0, i.e, the return difference is bounded away from zero over C,. However, the condition ge A is very restrictive, and one would like to have a criterion that also applies to systems that are open-loop unstable. Sucha result is given next. 12 Theorem Suppose gis ofthe form312 Input-Output Stability ch.6 13 g65)=45)+ 81s). where é,¢ A and @, is rational and strictly proper. Then Hof (6) belongs to A" ifand only if(11) holds. Remarks Before proving the theorem, some remarks are in order to explain the hypothesis and value of the theorem. 1. The hypothesis on g is that it consists of a stable part which could be distributed plus an unstable part whichis lumped. In particular, this implies that gis mero- morphic on the open RHP. and that it has ony a finite numberof singularities in the open RHP, each of which isa pole of inte order. 2, The theorem is useful because it shows that (11) anecessary and sufficient con- dition for the feedback stability of a broad class of systems, not just those that are ‘open-loop stable. With this as the starting point, itis possible to derive a Nyquist- like graphical stability test. Proof “Only if Suppose Hie A. Then, as shown in Lema (7), it follows that U1 +kg)e A, which in turn implies that Fis bounded over the closed RHP, i.., that 14 sup, 1s) 0. ‘Then Lemma (1) would imply that [/ge A, which in turn implies, due to (18), that ¥e A, and the stability of H would be proved, Accordingly, the proof is completed by showing that (Q0)is true. Note that MA Gada +k) By (11), the quantity 11 +4@(s)| is bounded away from zero over the closed RHP C,.. On the other hand, d could have some zeros in C,, namely the the poles of g,. Let Ay, °°. Ry denote the poles of &, in C,,, and select some open disks B , =, B,, with B, centered at, such that none of the disks B, contains azer0 offi Since fl and dhave nocommon zeros, this can be achieved by making the disks B, sufficiently small. One other technicality is that if some A, ison the jooaxis, then B; is chosen to be a half-disk, so that B,C. Now define B tobe the union ofthe sets B through B,,and note that, by assumption, 22 inf 1i(s)1 >0, since B does not contain any zeros of . Similarly itfollows that 23 nf io 0 So, if (11) holds, then (21) and (23) show that uw inf 1g(s)1>0. sec, a What ifs¢ B? Atthe zeros of d, we have, from (19), that 25 as) ils) #0. Hence, by selecting the disks B, small enough and making use of (22), one can ensure that314 Input-Output St cho 26 inf (5)1>0. Finally, combining (24) and (26) establishes (20). As shown earlier, this completes the proof. i Now let us consider the issue of how one might go about verifying whether the condi- tion (11) holds. If is rational, then the familiar Nyquist criterion of undergraduate control theoty allows one to test whether (11) holds by examining only the behavior of the function U0) as o varies over the real numbers. In attempting to extend the test to distributed sys- tems, the main difficulty one faces isthe irrationality of the function 4(s). Specifically, write Hs) [where @ =, +2,; see (13)] in the form 27 Als)= Lae" thats) where isthe Laplace transform of a function ¢€... Now, by the Riemann-Lebesgue Jemma, | &m(/@)! —0.as 1a! ee. On the other hand, the first term. 28 igi) is an “almost period are all commensurate, * function of «, In the practically important case where the delays 1, iT. T given, the function Z4p(j0) has the form 30 Bepo)= ale"! B and is @ periodic function of w with period 2n/7. In either case, it is quite possible that 4a(je) has no specific limit as 1001 -> =», This difficulty does not arise if is a proper rational function, In spite of this, however, itis nevertheless possible to state a Nyquist-like graphical stability test. The proof in the case of commensurate delays is given by Willems (1969a) and in the noncommensurate case by Callier and Desoer (1972). To avoid techni- calites, only the case of commensurate delaysis discussed here, andthe proof is omitted. Before proceeding to the graphical stability criterion, it is necessary to introduce two preliminary concepts, namely (i) the indented janis, and (ii) the argument of the retum difference function 1+kg (jo). If @ has a pole at some point j2, on the ja-axis, then 1+4@(/2,) is undefined. To circumvent this difficulty, the ja-axis is “indented” around the poke, as shown in Figure 66, by going around the pole. Let B, denote the half-disk shown in Figure6.6:ie,letSec.6.5 LT1 Feedback Systems 315, a Fig.6.6 BL By={seCy:1s—j&l 0. Hence deleting the half-disk B, from the closed RHP does not alter whether (11) holds or not. Similarindentation can be performed around all jo-axis poles off. After the joaxis is indented, itis clear that, corresponding to each we R, there is ‘exactly one point on the indented jeo-axis whose imaginary part is o, but whose real part may or may not be zero, By a slight abuse of notation, let § J) denote the value of atthe point on the indented jo-axis whose imaginary partis; this is illustrated in Figure 6.7. be316 Input-Output Stability ch Once the j@-axis is indented as indicated above and @(j«) is suitably defined, the quantity 84/0) is a uniformly continuous function of «. It can also be assumed that 1 +k Go) #0 V0, since otherwise the condition (11) isimmediately violated and no further analysis isneeded. Thus itis possible to define a continuous function 6(j¢) such that 33° 1+kgGeo)=11 +480)! expLigGo)], and 34 O(/0)=0if 1 +K8(/0) >0, nif | +4840) <0. ‘One can also think of 9) as the argument of the complex number 1 +k#(ja@; hence we could denote itby the more suggestive notation Arg [1 +k&@)}. Now the result can be stated. 35 Theorem (Graphical Stability Test) Suppose & has the form (13), and in addition, suppose that the delays inthe distributional part of are uniformly spaced, as in(29), Let H denote the number of poles of & with positive real part. Then H of (6) belongs 108° if andonisif 36 (inf UI-+kg GU)! >0, and 37° (ii) lim (2/1) - O(—j2nn/T)) = 2m, ‘As mentioned earlier, the proof can be found in Willems (1969a). Theorem (35) has an interpretation quite similar to that of the standard Nyguist cri- terion. Congition () or (36) is equivalent tothe following statement: The plot of @(ja) is bounded away fromthe “ertical point”—L/K, Note that this sa stronger statement than "The plotof 70) does not passthrough the critical point-1/A." The latter statement suffices if ¢ isa strictly proper rational function, or even ifg, does not contain any delayed impulses, since in this case § Ja) has a well-defined limitas lol >, Butin general itisnecessary to use the more precise condition (36). Condition (i or (37) is a generalization of the familar requirement that the plot of g(j) encircle the critical point ~1/k exactly j1, times in the counterclockwise direction. If has the general form (13), then the phrase “encircle” has no meaning since $(j@) need not have a specific limit as !@! +, This is why, in (37), the phase gis evaluated at specially chosen, evenly spaced, frequencies 2n1/T. OF course, if, does not contain any delayed impulses, then one can make do with the simpler statement (40) below. The discussion can be summarized as follows: 38 Corollary Suppose j has the form (13), and suppose in addition that g,(-) does not containanydelayed impulses. Let denote the number of poles of with positive eal part Then H of (6) belongs to A” ifand only ifSec.6.5 LT1 Feedback Systems 317 S (i) 1+kg@)40, Yor RU{=, and 40 Gi) Jim 16G@)—9-jo)]= 2m. 41 Example Considera system of the form shown in Figure6.5, where st 44542 Rls) =exp(-s) Inthis case represents system with he rational transfer function (s? +4s +2){s?1)fol- lowed (or preceded) by a delay of 1 second. The objective is to determine the range of values of the gain k for which the feedback system is stable. The first step isto demonstrate that is ofthe form (13). This turns out to be surpris- ingly difficult and serves as a motivation for introducing the set Biche next subsection. By paral fraction expansion, we have Hence Clearly the first term belongs to A, since it is the product of two functions, each of which belongs to A: in fact, the first term is the Laplace transform of S(¢—1) + O.Sexp(-1-1) U (1), where U(@—1) isthe delayed unit step function. Now consider the last term, and expand itas 35e! | 35(e*~e) s= sot ‘The firstermis rational and unstable, but what is the nature of the second term? Let ce! eY-Y (1) =0, VEZ 1 Hence feL,, and eA, though 7 is of course not rational. So finally it follows that 8 = 8, +8, where318 Input-Output Stability ch als) ‘Thus gis ofthe form (13), s0 that Theorem (12) is applicable. Rofo) ‘The Nyquist plot of Zw) is shown in Figure 6.8. Note that, as w=, $j) approaches the periodic function exp(—j2). Also,1,., the number of open RHP poles of g, equals 1. From Theorem (12) and Figure 6.8, we conclude that the feedback system is stable ifand only if -2<-1k <-1,76, 010.5 0 be a given number; then the set Ag consists ofall distnbutions f(*) such that ‘f()=0fort <0, andhave the form B f=LLBU-W+LUit120, a which satisfy the conditions tt Fite cm siete Note that if G=0, then (44) reduces to (6.4.3). Hence the condition (44) is more restrictive than (6.4.3), and as a result, Ag is a subset of A for all o>0. Also, if ¢>@, then Ag is a320 Fnput-Ouipe stability cs subset of Ag. Thus, iff, ge Ag, then their convolution fg can be defined as before, and it is routine to show that» g€ Ag. More'Génerally, if fe Ag and g Ag, then their convolu- tion f « g belongs to the Set Aninjo,q)- As before, let Ag denote the set of Laplace transforms of distributions in Ag. Now define the sets A_and A_as follows: 4 A=UA,.A = UA, 30 050 Inother words, A_ consists of all distributions f which belong to A. for some o>0, and A_ is the set of Laplace transforms of distributions in A_. Note that A. isa proper subset of A, butis closed under convolution, Suppose fe A.. Then, by definition, there exists a o> 0 such that (44) holds. This means that the Laplace transform f is analytic over the shifted open half-plane (s:Res >a}. Asa consequence, all zeros of fin the open half-plane {s:Res>—G} are isolated and of finite multiplicity. In particular, all zeros of fin the closed RHP (s:Res 20) are isolated and of finite multiplicity. Thus if fe A. then feannot have a sequence of zeros, clustering at some point on the jaoaxis. This is not true in general for an arbitrary function in A, and this is one of the main motivations for introducing the set A_; see Vidyasagar etal (1982) for an example of a function in A which has a sequence of zeros clustering at a point onthe joo-axis _Now we introduce on last concept needed to define B. Suppose fe A. Then we say that/is bounded away from zero at inti thee exstsaconstantr< such that, 46 int, feat vo Inetect, (46) just states that all zeros of Fin C, Wein some compact subse thereof. Since cach of these zeros is isolated, fhas only finitely many zeros in C,. It iseasy to see that i J. @€ A_are each bounded away from zero at infinity, then so is their produc fg. 47 Definition The et Bconssts ofall functions f= d/h, whered be K.andinaddiion bisbounded away rom zeroat infty 48 Lemma Suppose ff. Then (ial singularities of fin C, are poles of rite order, and hey are alisoated: (i) all pots of fin Cen some compl subset herd, s0that f ‘has only finitely many poles in C.,; (ii) all zeros of fin C, are isolated and of finite multipli- cin. ‘The proof is aready consequence of earlier observations. 49° Examples Consider again the transfer function of Example (41), namely F s244542 foyae S82 Recall that we had to work a bit to show that is of the form (13). On the other hand,Sec.6.5 LTI Feedback Systems 321 showing that ge Bis quite simple: @ = @/b, where G(oyae*S A442 fp) S21 ‘As another example, consider a uniform RC transmission line whose input is the voltage applied at one end and whose output is the current at the other end. In this case the transfer function isthe so-called C parameter among the chain parameters, and equals, This transfer function has a pole atthe origin, but otherwise all ofits poles are on the nega- tivereal axis. To show that fbelongs o B, write itas i/d, where 5, disy=— fs)= ——p, (+1) sins s+1 Then fie A_ (though thisis perhaps not obvious), while d belongs to A_ and is bounded away from zero atinfinity ‘An example of a transfer function which does not belong to Bis provided by - 1 HS)= ase ‘which isthe voltage transfer function ofa uniform LC transmission line of unit (normalized) length. Since / has infinitely many poles on the jco-axis, i follows from Lemma (48) that h does not belongto B. 50 Lemma Suppose j, ge B: then ft, féeB. Proof Write 7=4/b, § =C/d, where &, 6, é, de A_, and in addition b, d are bounded away from zero atinfinity. Then Ineagh ese, both the numerator andthe denominator belong to As ination, ince bath And dare bounded away fram zero at infinity, soishd. Ml ‘The set of ala where be A and b 0s called the el of trations of A Lemma (S0 sates hat Bisa subring his fel of actions Next, several useful properties of the set B are proved. As the proofs are somewhat technical, they may be skipped in the first reading, and the reader can jump ahead to ‘Theorem (67), which isthe main result.322 Input-Output Stability che 52 Lemma Suppose fe and that f(o)=0 for some o>0; then the furction sb /eMs~o) is the Laplace transform of a function in Ly. Suppose fe'A and that ‘flo j)=0 for some 5 >0 and some (00; then the function s + fisM(s -0)* +0") is the Laplace transform of afunction in Ly Proof Express f(t) inthe form 53 f= EH 8U-1) += LAD 4L00. where fy isthe distributional part of f and f, is the measurable part of f. Suppose /(@) 5-0 Itis shown that each of the two functions on the right side of (54) is the Laplace transform of a function in L;. For convenience, let L denote the set of Laplace transforms of functions im, and observe that LGA. Fits, using the same reasoning as in Example (41), one can show that [Note ha the inverse Laplace transform of the function in(55)isjust 56 fee VaHn=0.¥r2H, ‘where U() denotes the unit step function. Hence the function in (56) has compact support and therefore belongs o ;, which shows that (55) is true. Moreover, the norm of this Func- tionis given by aes fe Pam UL 2 Wos filo, Wh ‘Now consider the function From (57) it follows thatthe right side of (58) is absolutely convergent. ic.,Sec.6.5 LTI Feedback Systems 323 ae MSE Vivace $n x ee defined and belongs to A Now consider the second term on the right side of (54). Suppose first that f,(s) is a rational function of s. Then soiis the term in question; moreover, its strictly proper and has no pole at s =<. Hence it belongs to L. Now, every element in L can be expressed as a limit of a sum of decaying exponentials (Kammler 1976); equivalently, every functionin L can be expressed as alimit in A of a sequence of stable rational functions. This shows that eh, and the proof is complete. The case of complex zeros follows similarly, i 61 Corollary Suppose fe A and that fhas a zero of multiplicity m at areal 6>0; then the functions (> f(sMs~)" belongs to Ly. Suppose fe A and that fhgs.a zero of mutipli- city mata point + joowhere o> O and a0. Then the function s + fis(s -0)? +0)" belongstoL Proof Apply Lemma (52)repeatedly. Wt 62 Corollary Let feA, and suppose that jo) =0 for some >0. Then the function 5 F454 Df(s\Ms ~9) belongs to A. Suppose fis-+jw)=0for some o> Vand some #0. Then the functions |-+(s + 1)? flsM(s~ 0) +07 belongs 0 A, Remarks Note the contrast between Lemma (52) Corollary (62). In the former, itis shown that f(s)/s ~) belongs to Ly, whereas here itis claimed that (s+ 1) iss ~0) belongs to A, The difference arises because Is ~9) is strictly proper, as a result of which the function f(sVs ~¢) does not have an impulsive part; in contrast, (s + 1s ~9) is not strictly proper, and as a result(s +1) f(s(s—6) could contain an impulsive part. Proof Suppose f(o)=0. Then st 1 La tye #8 Aoed 63 AS ied=fise 5 fone. ‘The case of complex zeros follows similarly. i Of course there is nothing special about the term s +1, and one can replace it by any ‘other first order polynomial as +B. 64 Corollary Let fe and suppose fhasa er of muliplicity mat o>0: then the fan tions | (5 + 1)" fisMs—0)" belongs to A. Suppose fhas a zero of multiplicity mata point 6+ jowhereo > Oand@#0. Then the functions +9(s-+ 1) fis ils a)? +7]" belongs324 Input-Output Stability cho Proof Apply Corollaries (61) and (62) repeatedly. 68 Lemma Suppose fe A_and is bounded away fram zero at infinity. Then there exists a unitibofA anda proper stable rational function such that f= i. roo SineeA..thereissomeo>Osuch thal fe Ag Satz, “2: denote he + tint rerset within," my eapetely.Debae 6 e=nl Then, from Lemma (52) and Corollary (64), it follows that ii(s):=/(s¥0(s) belongs to Ag forall 8 <<. (To prove this, one uses the fact that Re z; > — even if possibly some ofthe z, have zero real parts.) Now itis bounded away from zero at infinity, since both fand # ' have this property. Moreover, f(s) has no finite zeros in the half-plane {s: Res 2-8). Hence, by alight modification of Lemma (1), disaunitof Ag and hence of A-. Ml 67 Theorem Suppose 4, be A., and suppose that b is bounded away from zero at infinity. Then the ratio g = a/b is of the form (13). Specifically. if p\. °° py are the distinct C.-zerosofb, of multiplcitiesm , °°, my respectively, then 8 s)= $2 co) Bals)+ ES), where 8 8)=E ings nd Z,€ A, ‘and the constants ry are evaluated as Les py" Maye d =O. 0 T ost Remarks Theorem (67) states two things: (i) Ife A_and is bounded away from zero 4 infinity, then forall practical purposes b is like a proper stable rational function, To see ‘why, suppose é =a/b and express has if where isa proper stable rational function and itis aunitof A... Then the function 2 =4/ belongs toA_, and moreover Where the denominator # now has a very simple form. (ii) Suppose we start with (71) and ‘carry out. partial fraction expansion asin (69). Then the part, whichis “left over" belongsSee.6.5 LT1Feedback Systems 325 to A. This was the calculation we had to do in Example (41), and Theorem (67) makes it precise. Note that the formula (70) isthe familiar expression for the coefficients ina partial fraction expansion Proof With all the preliminary work, the proof of the theorem is actually quite easy. t, write @ in the form (71), where > of and expand 1/(s) as partial fraction sum 4] $e—u SP py To prove that § =2/0 is ofthe form (13), it is enough to show that each term 2(s\(s—p,) is. of the form (13), since a finite sum of functions ofthe form (13) is again of this form. (Note that both A and the set of strictly proper rational functions are closed under addition.) Accordingly, consider the ratio 6(s)As—p,y. Since Ce A., there exists a 6>0 such that Ge Agvand of course Rep, 20>~o. Now write at at &)- Ee-p)' eet fz} py py ‘The second term on the right side is a strictly proper rational function. As for the first term, its numerator has a zero of multiplicity j at s=p,. Hence, by Corollary (61), the fist term belongs tog and hence to A... This shows that the overall function 2(s/(s) is ofthe form (3), Now the formula (70) for the constants, follows.inthe usual fashion, i ‘This subsection is concluded with an obvious result. 78 Lemma Suppose FeB"™". Then F can be decomposed as 16 F(s)=F.ls)+ FAs), where FyeA"" and F,(s) is an xm matrix whose elements are strictly proper rational functions of. In summary, the set is very useful fora leat wo reasons: First, every element of & has the form (13); as.a consequence, the stability of feedback systems whose forward-path clement belongs to B and which have a constant gain in the feedback path, can be easily determined using a comprehensive and physically meaningful graphical stability test {Theorem (35)]. Second, determining whether or not given function belongs to B is easier than determining whether or nota given function has the form (13); compare Examples (41)326 Input-Output Stability chs and (49). At this point, itis natural to ask: Isevery function of the form (13) a member ofthe set B? The answer isno, asis easily shown. Recall that in (13) the function Z, belongs othe set A, whereas, as shown in Theorem (67), the function @, belongs to A. Thus, if one ‘chooses a function g, which belongs to A but not to A. {for example, a function which has a sequence of zeros clustering on the jeo-axis; see the example in Vidyasagar etal. (1982)], then this function is ofthe form (13) but does not belong to B. 6.5.3 Coprime Factorizations In the final subsection of this section, the notion of coprime factorizations is intro- ‘duced, and itis shown how one may analyze the feedback stability of distributed systems using this notion, 77 Definition Two elements a, be A are said 10 be coprime if there exist elements 2 YEA.uch that TR xnaty 9b=80), orequivalentty 19 Xs) d(s)+H15)b(6)=1, WsEC.. Inthis case we also say thatd, be Aare coprime. 80 Example Let 26, bis)=1 46% as ‘Then, bare coprime, since (79) is satisfied with H(s)=-05e", 5(5)=1. In order to show that two elements a, be A are coprime using Definition (77), itis necessary to display explicitly a solution pair (x, y) satisfying the relation (78), which is, often referred to as the Bezout identity. The next result gives a criterion for coprimeness that ismuch easierto verify. 81 Lemma Twoelements a, be Aare coprime ifand only if 82 inf max(14(s)1, 16(9)1) >0 Proof (Partial) “Only if’ Suppose a, b are coprime, and select , ye A such that (78) is satisfied. Then, since %, ye A, the quantities £(s), j(s) are bounded as s varies over the closed right half-plane C,.. DefineSec.6.5 LT1 Feedback Systems 327 83 j= sup (176)! +1569)11 Now itiseasy to show that, foreach fixed se C,., we have 4 1=15(5) G(s) +516) 5(5)1 SU1R(9)1 + LGU] maxt 12(s)1, 1565)1). Hence (83) and (84) imply that 85 ing max 1a(s)1, 15(5)1} 2+ >0. nitbo # “If” This part ofthe proof is given in Callier and Desoer (1978); see also Vidyasagar (1985), p.342. ‘The coprimeness condition given in Lemma (81) can be restated in another equivalent form, 86 Lemma Two elements a, beA are coprime if and only if there does not exist a sequence {5)} nC, such thar d(s;) 0, B(s,) »0asi >=. Proof tis shown that the "no common zeroing sequence” condition given above i equivalent to (82). First, suppose chat (82) tre; then obviously no sequence (s;} inC, can be found such that 14(s;)1, 16(s,)! both approach zero. Conversely, suppose (82) is false, and select a sequence {s;j in C, such that 87 max{1a(s)i, 15(5,)1) 9 0asi oes, Then itisimmediate that 1G(s,)|, 16(s;)! —+0as—>e>, ‘The condition for coprimeness given in Lemma (86) says, in effect, that @ and b have ‘no common zerosinC,. However, since the region C, isunbotnded litle care isneeded. tis possible that ds) and (s) never both vanish ata common point se C, ,but nevertheless approach zero along some common (unbounded) sequence {s;} in C,. Lemma (86) says that, in order fora and b tobe coprime,thiscannot happen either. 88 Definition Suppose 5(s)=a(s)B(s), where d, be A. Then the ordered pair (4, B) is said to be a fractional representation of p. We also say that the pair (a, b) is a fractignal representation of p. If, in addition, @ and b are coprime, then the ordered pair (a,b) is called a coprime factorization of p. We also say that the ordered pair (a, b) is a coprime factorization of p. 83 Example Consider the transfer function a cosh s which was first introduced in Example (49), representing the transfer function of an LC transmission line. As was shown in Example (49), this transfer function does not belong to328 Input-Output Stability ch.6 the set 3, soasof now we have no means of analyzing the sabiity ofa system obtained by placing cvena constant feedback gin around (as in Figure 68). Now is posible t rewrite h(s) as ius)2 2 bes), ‘As shown in Example (80), @ and b are coprime; therefore the ordered pait (2e*, +e) is 8 coprime factorization of fh. As we shall sc, this will enable us to analyze feedback sys- temsinvolving h. 9% Lemma Every rational function has a coprime factorization. Proof Suppose i(s)isarational function, and express f(s) as cs VB(s) where o.and B are polynomials with no common zeros in C... Let kequal the larger ofthe degrees of cand Band define gis) = 26, gg) = BD 91 a(sy= 2 bes)= o (s+ 7 (s+) Then G,beA, and clearly A(s)= 0, As) Bos) Note thatthe condition (102) site swonger than: As) has fl column ak forall se Cy. IfM(s) has full column rank forall se C, then certainly det (s)M(s)] >O forall 5€C,. But this is not all (102) says: it says something more, namely that the quantity is bounded away from zero. where 103. Ais 104 Definition Suppose Be M(A). Ther an ordezed pair (8, B) ia vigh-coprimefa- torizaton eof PJ( PND and (i) Nand Dare righ coprime, mthivcase we also say that(N, D) isa right-coprime factorization of P. ‘The concept ofa left-coprime factorization (Icf) isdefined analogously. ‘The next result is amatrix analog of Lemma (90). 105 Lemma {fP(s) isa matrix of rational functions of s, then P hasan refand an lef. ‘The proof of this result may be found in Vidyasagar (1978b). ‘The next result is a matrix analog of Lemma (92). Its proof is left as an exercise to the reader. 106 Lemma If? has an rof respectively, an lef), and if Ge M(A), then P+ @ has an ref (respectively, an Icf). Every matrix in M(B) has both an refandan lef. Up to now we have had a barrage of definitions and lemmas. Finally we come to the piece de resistance of this section, which is a set of necessary and sufficient conditions for the feedback stability of systems of the form, shown in Figure 6.2, inthe case where Gand. Gy ate linear time-invariant systems. Let G,, i= 1,2 denote the transfer matrices of these {wo systems, and note that G, and G> have complementary dimensions. In other words, if G has dimensions bom, then G has dimensions mxl. Asa consequence, both G G2 and. GG, are square matrices, though possibly of different dimensions. A standard identity in matrix theory, can now be ysed to show that, for cach fixed s, we have that det [1 +G (5) Gx(s)] = det 17 + G2(s) G \(s)]. Ithis determinant isnot identically zero as aSee.6.5 LTI Feedback Systems 331 funetion of s, then we say that the feedback system is well-posed, In this case itis possible to solve the system equations (6.2.27) inthe form where W466." -G10+6,6.)" 108 H=| ~ Af — Gi(I+G,G,y' (1+G,G,)" We say that the feedback system is stable if He M(A). From Theorem (6.4.45), it follows that He M(A) is a necessary and stfficient condition for various forms of stability. The next theorem can therefore be thought of as providing necessary and sufficient conditions for several types of stability at once. 109 Theorem Suppose (N,, D,) is an ref of G,, and (N,, D,) is an lof of G, for i= 1,2. Then the following statements are equivalent (ited. 110 Gi),ink 1detd(s)! >0, where M1 A=N\N2+D,D2 rt ing Ide A(s)1 >0, 2 Gi), ink, 1der. (a)! >0, where aN +D2D, 113 Remark Note that conditions (110) and (112) can be verified using graphical criteria analogous to Theorem (35) and Corollary (38). More on this later. Proof Using the matrix identities 114 A(+BAY' =(4ABY A, MIS (+BAy' =/-B(+AB) “A, ene can obtain wo equivalent expressions for H rom), one of which involves only (4G ,G2)"' and the other of which involves only (+G2G )"'. They are332 Input-Output Stability cho . [I-G20+G.G."' Gy - 6.46162)" UG 616, 4G, G7" . | UGG! -0+6,60" Gs + 1-6, 4+626))" G2|" Itis now shown, using the expression (116), that He M(A) if and only if (110) holds. The proof that He M(A) if and only if(112) holds i entirely analogous, and proceeds from the expression (117) for H Substituie 18. 6, =5)'N,,G,=%,03 in(116) and clear fractions. This gives thea WW, ied 'D, 19 A=] ww SHAT BR, Bad 'B, First, suppose (110) holds. Then, by Lemma (3), it follows that Ae M(A). Since all the matrices on the right side of (119) belong to MA), it follows that He M(A). To prove the converse, suppose He M(A). Then since fe M(A),itfollows from (119) that fMaa't Meb"B)] [Xo], 120) ~~. ~ 1A Me M(A). | 5) BAN, B80, Now select matrices X , ¥, X2, 72€ MAA) such that 1 N,X,+D,¥,=1, X2N2+ 7.0) =1. Such matrices exist because NV, D, are left-coprime, and N'2, Bare right-coprime. Now from (120)it follows that peapee el ti eer m2 a =k eat em). l Since Ae M(A), Lemma (3) now shows that(110) holds, i i 123. Example As an illustration of Theorem (109), consider a feedback system of the form shown in Figure 6.5, whereSee.6.5 LT1 Feedback Systems 333, &)= ‘coshs and kis constant, As shown in Example (89), acoprime factorization of is given by the ordered pair (/i(s), d(s)) =(2e, 1+e*). Since & is just a constant, it has the obs Coprime factorization (&, 1), Since the system is SISO, we need not bother about Ief’s and refs, as both notions coincide. Hence the function Aof(111isinthis instance given by Ais=d(s) +ki(s)= 1 +e + ke. Itisnow shown that, for every ke R, there isan so C, such that A(sq) =0. This means that (110) is violated, and hence the feedback systemis unstable. Consider the equation Le + ke Thisis aquadatic equation n.x:= e™*. Let a, Bdenote the wo roots ofthe equation 14x? +k ‘Then clearly a= 1, which shows that both roots are nonzero and that at least one root has ‘magnitude less than or equal to J. Let o denote a root such that 0< lal <1, and select 59€C, such that e-"? =a. Then A(so)=0, which means that (110) does not hold. Thus we ‘can conclude that the feedback system is unstable for all real k. Fig.6.11 ‘Now suppose the feedback operator is changed from a constant k to a pure delay of 1 second and a gain of ~0.5, as shown in Figure 6.11. Inthiscase, since the feedback operator —0.5exp(—s) belongs to A, it has the coprime factorization (0.5 exp(-s),1). Now the function A becomes Ais)=d(s) +fls)(-0.5e Hence (110) is satisfied, and we conclude thatthe system of Figure 6.11 i (BIBO) stable 124 Example As another illustration of Theorem (108), consider again the sytem of Fig- we6 5th334 Input-Output Stability ch.6 ‘Now @(s)has the coprime factorization (fi, d) with f(s)=1+e™, des Itis easy to see that § =f/d. Also, and d are coprime since (79) is satisfied with = Hence Ags) =d(s)+ kis) =d(s) + A(s Hence (110) holds, and the feedback systemis stable. Fig.6.12 125. Example Inthis example, we analyze whether the stable feedback systems of Exam- ples (123) and (124) remain stable when a small delay is inserted into the loop. Consider firs the system of Figure 6.12, which is the same as that in Figure 6.11 except for an addi- tional delay in the feedback. Itis now shown that, for every € >0, there is a t Ouch that 128. Galt)= EGj Ht iT) + Galt), Gig MUL) Finally, le denote the MeMilgn degree ofthe unstable part of Gy. Then the ransfer ‘Fancion Hof (108) belongs to MA) and onl the folloning nwo conditions hol336 Input-Output Stability ch. 129 (nk Idee +G,40)G,40}1>0,a08 130 (ii) lim 9) 2nn 1) ~ O(—j2n/T)= 2m; +H), where 131 9(ja)=Arg cet +G Jo) GsGa)} Remarks 1. Suppose G is a matrix of rational functions and that p is a pole of g. Then the McMillan degree ofp as a pole of G is the highest order it has as pole of any minor of G. The McMillan degree of the rational matrix G is the sum of the MeMiillan degrees of allo its poles. Thus isthe sum of the MeMillan degrees of all the unstable poles of G. Since the juvaxis is indented to go around the purely imaginary poles, only those poles with positive real parts should be counted in computing. Similar remarks apply oa. 2. Note that the hypothesis of Theorem (126) requires that the all delays in both G 1, and G j, must be rationally related, 3. If the system is SISO and G, is just @ constant & (so that jt2 =0), then Theorem (126) reduces to Theorem (35) Problem 6.15 Prove Lemma (1) for rational functions /: Problem 6.16 Using Theorem (35), determine the range of constant feedback gains L which can be placed around the following transfer function in order to produce a stable closed-loop system: ao-[- Problem 6.17 Using Theorem (109), determine whether or not each of the following feedback systems is stable.Sec. 6.6 ‘Time-Varying and/or Nonlinear Systems 337 6.6 TIME-VARYING AND/OR NONLINEAR SYSTEMS, The previous section was addressed to the stability of linear time-invariant feedback systems. In the present section we study time-varying and/or nonlinear systems. Two gen- eral methods are presented for the analysis of such systems. The first, known as the small gain approach, can be used to study Lp-stability for all values of pe{1, e], whereas the second method. known as the passivity approach, can be used to study L2-stability. Using the relationships between input-output and Lyapunov stability derived in Section 6.3, both approaches can also be used to analyze the Lyapunov stability of nonlinear feedback sys- tems, In particular, the small gain approach leads to the circle criterion while the passivity approach leads to the Popov criterion, 6.6.1. The Small Gain Approach ‘The starting point ofthe small gain approach isthe following result: 1 Theorem Consider he system in Figure 6.2. and suppose pe [1] is specified. Sup- pose in addition that both G, and G> are causal and Ly-stable wb, and let p= (G1) Yo = YplGa). Under these conditions, the system of Figure 6.2 is Ly-stable f 2 Yip <1 Remarks 1, The inequality (2) is often called the small gain condition, and the name of the approach derives from this ‘Theorem (1) can be interpreted as a perturbational result. Suppose we begin with two subsystems G, and G, which are stable in themselves, and then we intercon- rect them in the fashion shown in Figure 6.2. Then the resulting system is also stable provided the “loop gain’ is less than one. Theorem (1) is also valid with "wh" replaced throughout by "wig." The proof of the amended version is left as an exercise (Problem 6.18). Proof To streamline the proof, let us introduce some notation. If a. be R", then the statement "ab" is equivalent to "a, - stability becomes AA sup |@Goyt-r<1, Which isthe same.as (19). i Remarks In the proof of Lemma (18), the fact that @ is a memoryless nonlinearity is not used; the only relevant property of is that y3() , as shown in Figure 6.15. Then, through block diagram manipulations, the system is redrawn as shown in Figure 6.16, where Know acts as a feedback around G and a feed-forward around G. Note that the first external input is changed from u2 to u —Kx2 as a result of these manipulations. Now the system of Figure 6.16 can be interpreted as that in Figure 6.17. If this system is L,- stable, then so is the original one in Figure 6.2 (and conversely; see Problem 6.19). These ideas are formalized in the next resultSee.6.6 ‘Time-Varying and/or Nonlinear Systems 3al Fig.6.16 22 Theorem (Loop Transformation) Consider the system shown in Figure 6.2, and suppose pe (1, ] is specified. Suppose G2 is causal and Ly-stable wh. Under these condi tions, the system is Ly-stable wb if there exists a causal linear operator K which is Ly-stable wb such that i) G (+ KG ,)"" iscausal and Ly-stable wb, and (ii) 23 YplG\l+KG 1 y(G2-K) <1 Proof The system of Figure 6.2 is described by the familiar equations We) uy ~¥2,€2 =U $Y Vi =Giery: Now define a new output 25 2a=y2—Ker=(G2~ Ker, and eliminate y, from (24). This gives 26 ey =u) yg =u) —22~Kep=uy 29K +91). Using the fact that Ks linear, (26) can be rewritten as 27 ey =(uy~Kuz)~22- Ky. “The other equations in (24) now become 28 ep =uz+y1,¥;=Gie 22 =(G2—Koen. Clearly (27) and (28) describe the system of Figure 6.16. Now this system can be rearranged. asin Figure 6.17, with342 Input-Output Stability ch.6 Fig.6.17 29 v, =u) Kur, ~Key.di =e) +Kyi.d2 62 Mae Z1=Y By applying Theorem (1) to the system of Figure 6.17, one can conclude that this system is Ly-stable wb if () G\(U+KG1)" and G2—K are causal and L,-stable wb, and (ii) (23) holds. To complete the proof. it only remains to show that the original system (24) is also L,-stable wb. For this purpose, introduce the notation 30 Y= %lGiE+KGiY"], y=%(G2-K), and note that 1,5 < I from (23). Now, from (29), we have w [ee ey fa [ey Hence, whenever (u , u2)€L%, it follows that (v, v2)€L%,and moreover, mn ee 2 Where ky isthe induced norm ofthe linear operator K viewed as a map from 1" into L' [Now the fact that 7,y,< 1 implies that the system of Figure 6.17 is L,-stable wb. Hence, Whenever (u),u2)€L%, (which in tum implies that (v,,v2}€L%], it follows that (d,d3)€L1, (2), 23)€ 3. Moreover, inanalogy with (13) and (14) one obtains the bounds Wdyty) PL yf ot} Merl) [te tow] [ity Simm [wm % || Well liza, ‘Substituting from (32) into(33) givesSec. 6.6 ‘Time-Varying and/or Nonlinear Systems 343, Md Ip 1 tthe] flatly zy tp 1_f Ye Yolto +h) | [Meer lp] Meaty] e @)DG,0) yen 0) DG, 6) whene <<. oeatb, (©)DG,2) when a O such that 42 gelt)=E8.60-iT)+ Bale Bal % Under these conditions, the feedback system is L-stable wh if one of the following condi- tions, as appropriate, holds: Case (1). ab > 0: (i) With the jeo-axis indented around the purely imaginary poles of g asin Section 6.5.1, the plot of (ja) ishounded away from the diskD (a,b), ie,See.6.6 43 ‘Time-Varying and/or Nonlinear Systems 345 inf, WeGor-z1>0. eR 26D.) (ii) Let, denote he number of poles of & with positive real part then 44 im Arg @U22n/1)—2]~ Arglg(-j2m0/7)~z)=2mp,, Wee D(a, b) Case (2).0=a< be (i) ge A, and ii) 45 ing Reguw> Case (3).a<0<6b: (i) 8 A. and (i) the plot of §Je) is contained in the disk D(a, b), ‘and is bounded away from the circumference ofthe disk. Remarks 1 2 Nove that, in (42) all the delays inthe impulsive part of g, are commensurate Iga in (42) has no delayed impulses, then (44) is equivalent to the simpler condi- tion: The plot of (jo) encircles the disk D(a, b) in the counterclockwise direc: tiom exactly 1, times as «increases from -s»to-=. Recall that Theorem (6.3.46) relates the L-stability of a system to the global attractivity ofthe unforced equilibrium; now Theorem (40) gives asufficient con- dition for L-stability. By combining the two theorems, one can recover the circle criterion of Lyapunov stability [Theorem (5.6.37)}. However, the input-output version ofthe circle criterion is more general than the Lyapunov stability version, since the former applies even to distributed systems, delay systems, etc If —a ->0anda, bboth approach a constant k #0, then Theorem (40) reduces to the sufficiency part of the graphical Nyquist criterion [Theorem (6.5.35)]. It is shown in Theorem (126) later in this section that, ina certain sense, the circle eri- terion also gives a necessary condition. As stated, Theorem (40) requires the feedback element ® to be a memoryless non- linearity. However, this fact is not used inthe proof. Itis shown in Theorem (126) that the circle criterion guarantees L>-stabilty even if © isa dynamic nonlinear ‘map elongingto the sector(a, 5] Proof Define Then b=k-+r, a=k—r, and the map 0 +9 6(, 6)—ko belongs to the sector [-r, r]. Now apply Theorem (22) (the loop transformation theorem) with K = kl, p = 2, nd combine with Lemma (18). Since the map ®~ K belongs to the sector [-r,r], its Ly-gain is at most r Hence the systemis L -stable wb provided (i)/(I +k8)e A, and Gi)346 Input-Output Stability cho \ gga)! 47 sup 20) _j. © 1 1+keGe) | "Now it is shown that the hypotheses of the theorem enable us to deduce that the above two ‘conditions hold. ‘Consider first case (I). Since the point —1/k belongs tothe disk D (a, b), (43) and (44) show thatthe hypotheses of the Nyquist criterion {Theorem (6,5.35)] are satisfied. Hence BAL-+Kg)EA. To establish (47), we again make use of (43). From (39), only elementary algebra isnceded to show that jek rah 1 481 iffz€D (a, b). 1 Since the plot of @(/) is bounded away from the disk D (a, b), (47) follows, and the system isLy-stable wb. Next, consider Case (2). In this case k=b/2, and -I/k =-2/b, The bound (45) states that the Nyquist plot of (ja) is confined to the half-plane {z: Rez > -1/b}, and of course =I <-1/o, Hence 49 Reg(ia)+ Mk >0, Vo, and as aconsequence, 50 Argla(jw) + Vk]e(- 02, 22), Vo. Inparticular, SA Arg(@(/2nn/T) + 1) -Arg [é(-j2nn/T)+ Vk Je(-n, 2), Von Consider the limit ofthe quantity in (51) as ->». This limit must be an integer multiple of 2. Now (51) implies thatthe limit must therefore be zero. Since @ A, ithas no poles with positive real parts. Hence by Theorem (6.5.35) it follows that g/(1+43)e A. Finally, itis routine to show chat \ ' z 2ittrez>-+, 1S ittRee >= 1 2 jk 2 lieoaR |b ‘Therefore (45) implies (47), and L stability wb now follows from Theorem (22). Finally consider Case (3), Ifk=0,then a =~r, b=, and the L-stability wb of the sys- tem follows from Lemma (18), so it can be assumed that k #0. The new feature is chat ‘ab <0, so that some inequalities get reversed when both sides are multiplied by ab. As a consequence we getSec. 6. 347 1 1 zled istis a, 83 eg |S ittzeDeaoy ‘Compare (53) and (48). Now (53) shows that the point -/k lies outside D(a, b). Indeed, if k>O then -1/k <-I/, or else if k <0 then ~I/e >~L/a, In other words, the disk D(a, b) lies entirely to one side of the point —/k. If k>0, then the fact that & ja)e D(a, b) Va implies that $4 Reg(ja)+ 1k >0, Va. Asin the proof for Case (2), (54) implies that (1 +kg)e A. Similarly, if k <0,then 55 Ref(ja)+1k<0, Va, and once again 1 +Kg)€A, Now the hypotheses in Case (3) show that (47) is also satisfied. Hence the L-stability ofthe system now follows from Theorem (22). 56 Example Consider the system of Figure 6.14, with ‘Then j isa product of two functions, each of which belongs to B; hence Ze B. By Theorem (6.5.67) it follows that @ is ofthe form (41), and itis evident thatthe delays in the impulsive part of g(°) are commensurate. Hence Theorem (40) applies tog Now f isunstable, and 1, = 1. So Cases (2) and (3) of Theorem (40) are inapplicable, and only Case (1) may possibly apply. The Nyquist plot of §(ja) is shown in Figure 6.19. From the figure one can see that if toa -1S< Jets > 0.667 0. 57 Example Consider the system of Figure 6.14, with oa 544] 542 | ‘Then Ze A and thus falls within the scope of Theorem (40).x8. Input-Output Stability cho Imfti0) Repiw) os fine J—> Refiu) 1s 1 05 ° 05 1 © is shown in Figure 6.20, from which one can see that (G)=sup 18)! =4. ‘Suppose we apply the small gain theorem directly without bothering with loop transforma- tions. ‘Then Theorem (1) tells us thatthe system under study is >-stable wheneverSec.6.6 - Varying and/or Nonlinear Systems 349 Infue) 6 ' os ° Re®iiwd os 1 1s 2 2 1 ° 1 2 3 + Fig. 6.20 1 n0)< 025. WO In particular, whenever ® is a memoryless nonlinear element in the sector [0.25 +¢, 0.25~el, the system is L-stable. Now let us apply Case (3) of Theorem (40). From Figure 6.20 one can sec that the plot of @(ja)is always inthe interior of the disk D (a, b) provided Nya 0.25 0. By using Theorem (40) instead of Theorem (1), we are able toextend the upper limit from0.25 ~et00.5—e. Next let us apply Case (1) of Theorem (40). Since ., = in this case, itis desired that the Nyquist plot of (1) neither intersect nor encircle the disk D (a, b). From Figure 6.20 ‘one can see that these conditions ae satisfied provided the Nyquist po ies entirely to one side of the disk D(a, ). Thishappens provided ether 1 FS ie Oca ch <6: ‘Combined with the fact that ge A, this once again shows that the feedback system is L-350 Input-Output Stability h.6 stable whenever @ belongs to the sector [-0.25-+€, 0.5~e]. Hence there is no advantage to applying Case (1) inthis instance. 6.6.2 The Passivity Approach In this subsection, an alternative approach to L-stability is presented, known as the passivity approach. The Popov criterion is among the useful stability criteria that can be obtained using this approach. In contrast with the small gain approach which can be used to analyze L,-stability for all values of pe (1, =>], the passivity approach is naturally geared to analyzing L2-stabilty; it is, however, possible to analyze L..-stability using passivity methods, bt thisis not discussed in this book. ‘The next result, hough not the most general of its kind, is adequate for the present pur- ‘poses; itis taken from Vidyasagar (1977). A still more general result, based on the so-called dissipativity approach, can be found in Moylan and Hill (1978), 58 Theorem Consider the feedback system of Figure 6.2. Suppose there exist constants 8,751, 2, such that 89 r2ejlla ll} +8, Gurl fs, VT20, Vre La i= Then the systems Ly-stable wbif 60 5 +e2>0,8; 48) >0. Proof The system under study is described by the equations G1 ec; uy -y2,€2= U2 HV, V1 =G ie V2 =G2e2 Asaconsequence it readily follows that 620 p+ 7= 7, VT 20. Now, from (59), it follows that 6 r2eilesF, +5 tila, f= 1,2 Hence 64 r4 726i ley Ea +8) ly Fa +e le IF +83 ly Fo. Now note that We, l!}2=r. and substitute for e 2 from (61). This gives, after routine computations, 65 ey lle, Na +ealleal}a=e4L Nya ll¥2—2r+ Hey Fa)Sec. 6.6 Time-Varying and/or Nonlinear Systems 381 +eg[Ilyy Py +2p + Muy IP) ‘Combining (62), (64), and (65), and rearranging gives an implicitinequality, namely 66 (3, +€:)llys Fp +B +e )MyaN Fe S Sy, (uy ~2epUg) r+ Sy, (ug + Zeya) r— ey lay Fa ~ ep Mg NF. Using Schwar2" inequality and the triangle inequality on the right side of (66) gives frre 9 D/ site) © Cyt Wallrall 9 B40, Wyatn| 1 2eg1) Play ze SUWyi Mra My2 Mra) Ziel 1 Wag Nr ley! 0) f tea tra] [lay try tustzal| 9 ra Maa tra} This vector inequality is of the form 68 x’Axsx'Br+7'Cz, Uy ire Nay Wea yaya |*?*| Marg pa and the definitions the matrices A, B, C are obvious. Now A is positive definite, and hence hhas a symmetric square root $. By “completing the square," (68) can be rewritten as where C) 70 (Sx-48"'Bay (Sx- L8"'Bal -stable wb, -stable wb. By Lemma (6.2.33), this Remark Theorem (58) is stated for SISO systems, but it is easy to see that this fact is ‘not really used in the proof. Thus the theorem is applicable to MIMO systems as well, but an essential restriction is that both G , and G2 are "square," .e., have an equal number of inputs and outputs; otherwise quantities such as r do not make sense. On the other hand, the so-called dissipativity approach does not have any such restrictions. Several useful results can now be obtained as corollaries of Theorem (58). To state them in their original historical form, two terms are introduced. 74 Definition AnoperatorG: L3_ ~>L3q is said be passive if 7S p20,VT20, VreL2., andis strictly passiveifhere exists a constant > O suck that 16 y2Eellxll}), VT 20, VxeLy,, 177 Corollary The feedback system of Figure 6.2 is Ly-stable wh if both G, and G2 are strictly passive. Proof In this case (59) holds with €, >0, € >, and 6; =8)=0. Hence (60) is satisfied and the result follows from Theorem (58). i 78 Corollary The feedback system of Figure 6.2 is L>-stable wh ifeither (i) G, is strictly passive and has finite gain, and G2 is passive, or (ii) Gz is strictly passive and has finite gain, and G is passive. Proof Suppose (i)is tre. Selectconstantse > Oand y< such that, 19 p2elisll}y, VT20, Varela, 80 NG xiypp Syllxilp2, VT 20, VxeL2,, and observe that BL 20, VT 20, VxELy,. “Thus G, satisfies (59) with e = 8; =0. Now (80) implies that 82 U2 PNG ally, VT20, Vee Lay Pick any a¢ (0, €),and note from (79) and (82) thatSec.6.6 ‘Time-Varying and/or Nonlinear Systems 353 83 72(€-opllxil}, tallxll}y 2(€-iixii}y + G,xl1}s, 720, Wee Ly. ¥ Hence G satisfies (59) with Since both e; and 6, are positive, (60) is satisfied and the result follows from Theorem (60). If (ipholds, simply interchange the indices | and throughout. 85 Corollary Consider the feedback system of Figure 6.2, and suppose there exist real constants and 8 and a positive finite constant such that 86 y2ellxll}y, V720, Vee L. 87 UG yxlly $ylixlty, VT20, Vee Les 88 p261GyrIl}2, VT 20, Vxe Lae Under these conditions the system is L-stable wbif 89 e+5>0. Proof As in the proof of Corollary (78), (86) and (87) together imply (83). Hence G satisfies (59) with e; and 8, defined in (84). Now (68) states that G> satisfies (59) with €)=0,8;=8, Hence, forsufficiently smal) a, we have £50, +e)=e+8-020. 7 Hence (60) is satisfied and the result follows from Theorem (58). 9 848) ‘The well-known Popov criterion can now be obtained as an application of Corollary (85). A preliminary result, which is of independent interest, is stated first. 91 Lemma Suppose ge A, and define G: Lz, -9 Ly, by Gx=g +x. Define OI inf Re gw). Then (79) holds (with G , replaced by G). Proof Since G is causal, it follows that (Gx), = (Gx7)r, for all T20 and for all xe Ls, Also,354 Input-Output Stat cho 93 r=[ EN Ge Hed = [spt Grdelerdt= [xp0) Grn a Now x7¢L2 whenever x6 L2-, and therefore xy has a Fourier transform; denote itby ia). ‘Also, since «A, the function Gxy belongs o Ly, and its Fourier wansform is Gj) x74). By Parseval's theorem, 94 p= Re J feria)? 8409 nV dor 3 J Reaay lipGo)!? do xe 1z;Go)l2do sellspli} celts Since (94) is tue for every T2Oand every x€ L,,theleramais proved. Ml 95 Theorem (Popov Criterion) Consider the feedback system shown in Figure 6.14, ‘and suppose the following conditions hold: (i) g() has a distributional derivative, and 8. 8A. (i) Bisa memoryless time-invariant nonlinearity ofthe form 96 (@x\= lx). where : R->R is continuous and belongs to the sector [0,} where b could be infinite. Finally, suppose there exists aconstantq > 0 such that 97 inf Re[(1 + Jorg) ge] +5 =B>0. > Under these conditions, the functions ¢ , 2.9.2 belong to Ly whenever uy, uz and itz belong 10 Ly; moreover, there exists aconstanty <= such that 9B We May hyp lly -stability can be con- 0, since in this case the inequality (97) is weaker than (45). 16, i 4 encarta o I ! t i Fig621 L =| Proof Rearrange the system of Figure 6.14 as shown in Figure 6.21 by introducing the “multipier” 1 +gs. In the process the input u is modified to v2 =u +4iz>. which belongs toZ, by assumption. Now define G, and G2 as shown in Figure 6.21 and apply Corollary (85). The assumption that g, &¢ A imply that (1 +5) @(s)e A. Hence G has finite L>-gain wb and satisfies (87). Next, from Lemma (91), G satisfies (86) with 99 = inf Re[(1 + jag) aga). Itis now claimed that G satisfies (88) with 6= 1/b. If this can be showa, then the theorem follows, because e +5 > Oby virtue of (97). ‘Thus, to complete the proof of the theorem, it only remains to show that 100 7=r2Liri,. VT20, YreLy.. ‘Now from Figure 6.21, we see that 1OL_r(e)=o[w(H). 20) = win +n), From Remark (3) above, it can be supposed that q >0, since if q =0 the theorem follows356 Input-Output Stability ch.6 from the circle criterion. If q>0, then w is the convolution of z and the function (U/q) exp (-1/4), which shows that w (0) = 0, Now r r 402 y a She entwinds Now, since belongs othe sector (0, bit follows that 10s 0s *2 <6, vor0, Hence 106 39(9)27 1))", Vo, and asa result, 107 lw ()] (2 = Lol = Liar, vre (0, 7] Combining (104) and (107) gi " 108 <2, r>r2 Frits, thus establishing (100) and completing the proof. i ‘The inequality (97) can be given a graphical interpretation, which makes it useful in practice. Suppose we plot Re g J) versus 1m (je) as @ varies from Oto=. This graph is, called the Popov plot, in contrast to the Nyquist plot in which one plots Re @(jo) versus Im je»). Since both Re @(/ia) and w Im gia) are even functions of w, itis only necessary to draw the plot for «20. The inequality (97) means that one can draw a straight line through the point -1/4 +0 with a slope /q>0 such thatthe Popov plot lies to the right of the line and does not ouch it; such a line is called a Popov line.Sec.6.6 ‘Time-Varying and/or Nonlinear Systems 357 109 Example Let le siadsd a ImFtie Retin) Fig. 6.22 wo imbo) Reger Fig.6.23 ‘The Nyquist plot and the Popov plot of @ are shown in Figures 6.22 and 6.23, respectively. From Figure 6.23, one can see that no matter how small 1/bis (i., no matter how large bis), itisalways possible to draw a suitable Popov line, as indicated. Hence the system of Figure 6.14 is L>-stable forall time-invariant nonlinearities () in the sector (0, b ] forall finite B. Onthe other hand, inf ReaGo) So (45) is satisfied whenever 6 < 32. Thus, by applying Case (2) of the circle criterion, one sees thatthe system of Figure 6.14 is L-stable forall possibly time-varying nonlinearities in the sector [0,6] for 6 <32. Hence, by restricting the memoryless element to be time- varying, we are able o infer stability fora larger class of nonlinearities. Application: Aizerman’s Conjecture358 Input-Output Stability cho Aizerman’s conjecture was stated in Section 5.6. Now consider the input-output ver- sion of Aizerman's conjecture, stated next: Fig.6.24 110 Conjecture Consider the system of Figure 6.24, and suppose that g, 2€ A. Suppose in addition that §A\+k§)€A for all ke (0,]. Then the system in L-stable for all non- linearities@in the sector (0, b| Essentially Aizerman's conjecture (input-output version) states that ifthe system of Figure 6.24 is La-stable whenever 6(-) isa constant gain of value ke (0,6), then itis also Lz-stable forall (memoryless time-invariant) nonlinearities inte sector (0, b . In general Aizerman’s conjecture i false see Willems (1971) fora class of counterexamples. How- ever, Popov's criterion provides a means of identifying a large class of transfer functions 40 for which Aizerman’s conjecture is tue. ‘Suppose g, 8A. If g(-) contains any impulses, then g would contain higher order impulses and thus would not belong to A. Thus the assumptions ge A, eA imply that ¢ does not contain any impulses, ic. that geL,. Because geLy, it follows from the Riemann-Lebesgue lemma that @(jea) has a definite limit (namely 0) as @—>ee. Since 8A1+kg)€A Vke [0,5] by assumption, the graphical stability criterion of Theorem (6.5.35) implies that the Nyquist plot of #(/o) neither intersects nor encircles the half-line segment (—», ~1/b ]. Because the only difference between the Nyquist plot and the Popov. plot isin the vertical axis, the same is true of the Popov plot as well. Now, suppose the Popov plot of @ has the shape shown shown in Figure 6.25. Then the L-stabilty wb of the nonlinear feedback system is assured for all time-invariant nonlineasties in the sector {0,6}, because of the Popov criterion, By Theorem (6.3.46), the state-space version of Aizerman’s conjecture is satisfied by such systems. On the other hand, suppose the Popov plot of has the appearance shown in Figure 6.26. In this case, Popov's criterion is not satisfied. However, since the Popov criterion is only a sufficient condition for stability, it still does not follow that Aizerman’s conjecture i false for such a. Thus, in summary, the Popov criterion provides a readily verifiable sufficient condition for determining whether Aizerman’s conjecture is valid for a particular transfer function (-). 6.6.3 Necessity of the Circle Criterion This section isconcluded by showing tha, ina sense to be made precise below, the cit- cle criterion provides a necessary as well as sufficient condition for absolute stability. Hence in a sense the circle criterion is not overly conservative. To minimize the technicalSec.6.6 Time-Varying and/or Nonlinear Systems etme) 6 Retied Fig. 6.25 etmgie Reged Fig. 6.26 359 details, attention is restricted to systems of the form shown in Figure 6.27, where the for- ‘ward path consists of a lumped linear time-invariant system, and the feedback path consists of a causal, but not necessarily memoryless, nonlinear element ©: La, —>L2-. Leta, b be given real numbers with a 0. Before presenting the ‘main result ofthis subsection, which is Theorem (126), an important special case is treated, 112 Lemma Consider the system of Figure 6.27, and let r > be a given real number. Then the following two statements are equivalent (D The system is L-stable wb for every ® belonging to the sector (-r, 1). angeA and : 143 sup 1G! < 4. Proof “(ll)=>()" Suppose (Il) is true. Then the operator G:x 19g «xis Ly-stable, and by Theorem (6.4.40), 114 y(G)=suplayot s+ Now suppose ©: L, ~> L2, belongs to the sector (~r, r). Thea there exists an € > 0 such that ‘belongs o the sector[-r+e, re]. Hence @is L.y-stable wb, and US (®)sr-e (11) Suppose is tue. Then, in particular, the system is Ly-stable with ®=0, since the 2er0 operator belongs to the sector (rr). Since j is rational, L-stability is equivalent to .stability, and g¢ A. To prove (113), itis shown that if 113) is false then (1) isfalse. Accordingly, suppose (113) is false, i, that 117 suplgya)t >t, and select an oy such that 118 12a)! ot Tobe precise, suppose , 1 19 &oy)= exp UO. Where €lies in the interval (0, r) and 8¢ [0, 2m). Now letSec.6.6 ‘Time-Varying and/or Nonlinear Systems 361 120 t= 8/0, and define ®:L3, > L2e bY 121 (xy) =~(r-e)x(t—-0). Inother words, @ isa gain of —(r—2) cascaded witha defay of t; note that ® is not memory- less. Now «Dis also linear and time-invariant adits transfer function is 122 Ms (r~e)exp (13). From the construction itis clear that 123 1 +84) 67a))=0. Hence the function s +» 1/11 +8(s) 6(s)] is unbounded over C,; as a result, the cransfer function 141-+26) cannot represent an Ly-stable system. Thus the system under study is Ly-unstable for the particular choice of ® in (121). Since this @ belongs to the sector [-r +e, re] and hence othe sector (—r, r)itfollows that (is false. i ‘Suppose we try to modify Lemma (112) by restricting to the “closed” sector [-r, r] and changing (113)t0 124 supl@g@)! +, then there need not exist any finite ay such that 1b)! 2 Ir. Ths the eason for stating Lemma (112) in that particular form, with ® restricted to the "open" sector (=r, r) and stat- ing (113) witha non-striet inequality Lemma (112) is an example of a result for so-called absolute stability. This term refers to the study ofthe stability of an entre famity of systems, instead of just specific sys- tems. The main idea in absolute stability theory isto deduce the stability of an entire family of systems by studying only some of its members. By examining the proof of Lemma (112), the reader can easily prove the following result: If the feedback system of Figure 6.28 is ZL-stable wb for all real constants ke (~r, r) and all delays T20, then the system of Figure 6.27 is L-stable wb forall nonlinearities din the sector (-r,r). Now forthe main result, 126 Theorem Consider the system of Figure 6.27, and suppose a, b are two given real numbers witha 0, then the Nyquist plot of &(jc) does not intersect the interior of the disk D(a, b) defined in(39), and encircles the interior ofthe disk D(a, b) exactly times in the counter- clockwise direction, where i. isthe number ofpoles of é with positive ealpart. (2) Ifa then § has no poles with positive real part, and 127 Re gjw)> (3)Ifab <0, then ge A, and the Nyquist plot of &(j@) ties inside the disk D(a, 6) forall ea Remarks There are some slight differences berween Statement (II) above and the conditions in Theorem (40). These differences arise because here the nonlinear element is assumed to ie in the “open” sector (a, b). while in Theorem (40) the nonlinearity is res trited to lie in the “closed” sector [a,b]. As aconsequence, the region in the complex plane in which the Nyquist plot of @ (ja) is required to lie is a closed set here, while itis an open set in Theorem (40). The mostaotable difference arises in Case (2), ie.,4°=0, [fis permitted tolie inthe sector [0, then obviously itself must be stable, because @=0 isa permissi- blechoice. Buti can only belong o the open sector (0, b), then =O is nora permissible choice, and j itself need not be stable. The differences in the other two cases ate quite minor. Note that itis quite routine to modify Theorem (40) to provide forthe ease where ® belongs tothe open sector(a, 6). Thus Theorem (126) shows thatthe circle criterion [suit- ably modified for the fact that the sector (a, b) is open) isin fact a necessary as well as sufficient condition for absolute stability, provided the forward path element is lumped. Of course, the power of Theorem (40) lies in thatitis applicable to distributed systems as wel, Proof "(1I) > (I)" This follows from Theorem (40), after adjusting for the fact that © belongs to the sector (a, b) rather than the sector {a, b J. The details are left as an exercise. (See Problem6.22.) "> ()" Each ofthe three cases ishandled separately, Case (3). ab <0: In this case © =0 belongs to the sector (a, b). Since the system is L2- stable wb when ©=0, it follows that itself is Ly-stable wb, and since is rational, chat eA. Now define, as before,Sec, 6.6 ‘Time-Varying and/or Nonlinear Systems 363 Again, © =k is a permissible choice, and the L stability wb of the system with this particu- lar choice of © shows that 129 j,=—2—eA 1+kg "Now redraw the system of Figure 6.27 as in Figure 6.29, where , is defined above, and 130 ©, =@—kle sector (-r, r). Of course the inputs in Figure 6.29 are not the same as those in Figure 6.27, butas shown in the proof of Theorem (22), the system of Figure 6.27 is L-stable wb if the sysem of Figure 6.29 is Lz-stable wb, and the converse follows easily (see Problem 6.19). Hence the hypothesis that (1) is true implies thatthe system of Figure 6.29 is L -stable wb for every ® inthe sector (~r, r). Now apply Lemma (112). This shows that 131. sup 1a)! < or, equivalently, 1 eee 132 sup | £0) _ © | L+lb+ay2gga) 1 {As shown in the proof of Theorem (40), this implies [afer allowing for the non-strictine- quality in(132)) that {)e D(a, 6) Vo. Case 2). and (131) becomes ‘The reasoning is much the same as in Case (3). In this case & 1g ' 133 supi—2U@__j<2, @ 1 1+@2QeG@)1 > which can be easily shown to be equivalent to (127). Also, since ~2/b <—1/b, (127) implies364 Input-Output Stability Che that 134 Arg [2@)+(2b)]€ (- 172, 172), Yo. ‘Asaconsequence, 135 im Arg LG) + (2/»)]— Arg [&(-jo) + (2/6)] =0, since this quantity must bean integer multiple of 2x. Now, by assumption, the feedback sys- tem is Lz-stable if @=k/ =(6/2)1. Comparing (135) with the argument condition in ‘Theorem (6,5.35), one sees that cannot have any poles with positive real part ‘Case (1). ab > 0: Define and r as in (128), and redraw the system as in Figure 6.29. ‘Then (129) and (131) follow as before, and (131) implies, asin the proof of Theorem (40), that (J) does not intersect the interior of the disk D(a, b). Since the feedback system is, stable with = I, the Nyquist plot must encircle the point -1/k exactly , times in the coun- terclockwise direction. Since the plot of @(je) does not intersect the interior of the disk ‘D(a, b), the same encirclement condition applies to every point inthe interior of D(a, 6). . Problem 6.18 Modify the statement of Theorem (1) tothe case where the operators G and G, are f,-stable wig but not necessarily L,-stable wb, and prove the resulting state- ‘ment. Obtain estimates corresponding to (13) and (4). Problem 6.19 Consider the feedback system of Figure 6.2 and its transformed version in Figure 6.16. Suppose Kis a linear operator and that K is Lp-stable wb. Show that the sys- temof Figure 6.2 is L,-stable wh if and only if the system of Figure 6.16 is L,-stable wb. Problem 6.20 Prove (48). Problem 6.21 Obtain explicit boundson the norms We lly» and ily Ip from (73). Problem 6.22 In Theorem (126). show that (II) implies (I). Problem 6.23 Show that, in Case (2) of Theorem (126), not only is § not permitted to have any poles inthe open right hall-plane, bu itis also not permited to have any repeated pole onthe jo-axs Problem 6.24 Using Theorem (40), find some possible choices of the sector (a, b} such the feedback system of Figure 6. 14s Lz-stable when f(s asin Problem6.16. Problem 6.25 Using Theorem (95), find the possible values ofthe constant 6 > 0 such that the feedback system of Figure 614s table with GHDLSA2TSVNSH1O) * and ba time-invariant and memoryless nonlinearity belonging tothe sector {0, 6].Sec.6.7 Discrete-Time Systems 365 6.7 DISCRETE-TIME CONTROL SYSTEMS In thistrief section, the discrete-time analogs of the contents of the first six sections are presented, mostly without proof. Since in most cases the details of the discrete-time results are virtually identical to those in the continuous-time case, only the differences are highlighted. 6.7.1 Stability Definitions Let S denote the linear space ofall sequences {;),29. For p¢[1, ©) define IxeS: 5 Ix? <=), & and define 2 1,,=(xeS:xisabounded sequence } ‘Then isa subspace ofS for each pe [1, =). If we define the rorms 3 Wall =LE ba], Vrel,, a 4 Wetly=sup Lx)1, Veet. then the pair (Jp, l-11,)isa Banach space for each pe [1, =]. Note that, |, ifp <4. A sequence xe'S is said to have finite support if there is an integer NV such that )Vi>N. Clearly, if x has fnite support then x¢1, Vpe [1 s]. Hence the set S acts as the "extension" of, foreach pe [1, |. and there isno need for asymbol such as J. ‘A inary relation on Sis defined justas in Section 6.2, namely asa subset of $? S Definition Suppose Risa binary elation on S. Then Ris|,-stable if 6 Wyekxeh ByElp Ris l-stable with finite gain (wfg) if is |, stable, and in addition there exist finite con stants Yp and b, such that 7 (ERX ly > Hy hy Sy lly bp Ris [,-stable with finite gain and zero bias (wb) if itis p-stable, and in addition there exists afinite constant'y, such that B (GER xe ly > ly ly Sp aly366 Input-Output Stability che The definitions of feedback stability are now entiely analogous to Definition (6.2.32), ‘with 2, replaced by Itis left tothe reader to state and prove discrete-time analogs of the contents of Section 63. 6.7.2 Linear Time- wariant Systems ‘Suppose x, ye. Then their convolution + ye Sis defined by 9 Gays Bay Diy ashen The set 1, plays the same role in discrete-time systems that the set A does in continuous-time systems. 10 Lemma Suppose x, yel,. Thenx » yel),and Mix ey il) Sixt tyty Every linear time-invariant operator A : > Shas the representation Ax =a + x, where the sequence ae S iscalled the unit pulse response of A. 12 Theorem Suppose A: 5-5 is linear and time-invariant, and let a denote the unit pulse response ofA. Then the following four statements are equivalent: () Aisty-stablewb. Gi) Aisl_-stable wb. fi table wh forallpe|l, =. Aisly Gv) aehy. Moreover, ifa€l,, then 13° lla exllpS tally xlly, Vel, Wpel Every sequence fe !, has a-transform defined by which converges whenever Iz I$ 1. Note that z is raised to positive powers in (14), not negative powers. The symbol! denotes the set of z-transforms of sequences in. If fe, then fis analytic on the open unit disk, continuous on the closed unit disk; moreover, any zeros of fin the open unit disk are isolated.See.6.7 Discrete-Time Systems 367 ‘Theorem (12) is the discrete-time analog of Theorem (6.4.30). The next result is the analog of Theorem (6.4.40), 15 Theorem Suppose A:S—5 is linear and time-invariant, and that its unit pulse response a belongs ol. Then max 1a(e/*)| ec(.28) 16 pia. [Note that (e!) a continuous function of @, and thatthe interval [0,2] is compact; herce in(16) wecan say "max" instead of "sup" ‘The stability results for linear time-varying systems can be obtained from Theorems (6.4.53), (6.4.66), and (6.4.75) simply by replacing all integrals by summations. In Fact, itis possible to place all of these results into a unified framework by replacing R (which plays the role of the "time set” in continuous-time systems) by an arbitrary locally compact, ‘Abelian group; see Vidyasagar and Bose (1975) for details. A causal linear discrete-time system G has an input-output representation of the form 18 Theorem The operator G:5 > $ defined by (17) is |.-stable wb ifand only if eee Gisly-stable wb fandonlyif 20 supE taylmies 0. Geitng a graphical criterion to test (23) is also easy. Suppose fe, and suppose one plots fe!) as 8 inereases from 0 10 2x. If fle/*) =0 for some 8, then (23) fails at once, and, ‘no further testing isnecessary. If f(e/®) #0 V8 (0, 2n), then it follows from the fact that all eros of fi) inthe open unit disk are isolated that fhas only a finite number of zeros in the closed (and open) unit disk. By the principle of the argument, the number of zeros of f(} in the closed unit disk is precisely the number of times thatthe plot of fie/®) encircles the ori- gin in the counterclockwise direction as @ increases from 0 to 2x. The discussion can be summarized as follows: 28 Lemma Suppose fe). Then \/fei, ifand only ifthe following two conditions hold: 26 (ijie)#0, Ve 0, 2m], and 27 Gi) Argfle”™)- Arg fie) =0. Consider now the linear time-variant feedback system of Figure 6 [with §(s) changed to 4(2), of course]. Assume that (2) hasthe form 2B g(2)=8(2)+8,(2). where 2,€/), and 2, is rational but anaiytic at z =0. (The analyticity assumption is to ensure that g corresponds to a causal system.) If g, has some poles on the unit circle, then the unit circle should be “indented” so that these poles do not le inside the region enclosed by the indented circle (see Figure 6.30), By aslight abuse of notation, let g(e/*) denote the value of (z) at the unique point z on the indented ynit circle whose argument is 8. Then we have the following result, which isan analog of Theorem (6.5.35). Im: Fig. 6.30Sec.6.7 Discrete-Time Systems 369 29 Theorem Consider the system of Figure 6.5, with @(s) replaced by @(z), and suppose §(@) has the form (28). Then the feedback system is stable ifand only iti) 30 Gle!)#- 1, v0e(0,2n), and (ii) the plot of @(e"*) encircles the point -\/k exactly j1, times in the clockwise direction as 0 increases from 0 to 2n, where \1, is the number of poles of &, in the open unit disk. Notice an important difference between Theorem (29) and Theorem (6.5.35) [or Corollary (6.5.38). In the latter case, the encirclements are required to be in the counter- clockwise direction, whereas in the present case they are required to be in the clockwise direction. The difference arises because othe orientation of the contour in the two cases. In the discrete-time case the unstable” region, i. the region where the poles are forbidden to liein order to have a stable system, lis tothe leftof the unit circle. Hence, as shown in Fig- ure 6:30, the unit circle is a positively oriented curve. In contrast inthe continuous-time case, the unstable region lies to the right of the ja-axis as increases from —=t0—, ‘The discrete-time analog of the set Bis straight-forward, Givenap > 1, define 31 lip =lae8:E nip! <™ ‘ and define 32 Ly = Ulige pt ‘Then 1y.c11. Let, denote the set of z-transforms of sequences in! .. 33. Definition The set By consists of all ratios fi(zVd(z) where fi, de By, and in addition d(0) #0. Itis let to the reader to state and prove properties of the set By analogous to those of B as stated in Section 6.5.2. In particular, every fe B, has the form (28). ‘The extension of the contents of Section 6.5.3 to discrete-time systems is totally straight-forward, For a unified treatment of the coprime factorization approach, see ‘Vidyasagar (1985). 6.7.3 Nonlinear Feedback Systems ‘Now let us tum our attention to nonlinear feedback systems. The small gain theorem [Theorem (6.6.1) the passivity theorem [Theorem (6.6.58)] and its various corollaries all apply. with the obvious modifications, to discrete-time systems370 Input-Output Stability ch. Consider now the feedback system of Figure 6.14, where the forward path is linear and time-invariant with transfer function (2) [instead of @(s)), and the feedback element ® is, memoryless and of the form 3 (x), =004) for some continuous function@: Z, xR IR, where BS Z,= (0.1200) denotes the set of nonnegative integers. In analogy with (6.6.17), we say that " or ® belongs othe sector a, b Jif 36 a9? 0: The plot of ge") as ® increases from 0 to 2n does not intersect the disk D(a, b), and encircles it exactly times in the clockwise direction, where i, i the umber ofpoles of in the open unit disk fel), and Case(2).a 38 Regle!")>—1, vee {0, 2m}, Case (3). ab <0: gel), and §e!®) lies in the interior of the disk D(a, b) for all Ge (0, 2n]}. ‘The "Popov criterion” for discrete-time systems, however, noticeably different from its continuous-time counterpart, and is given next. In proving this theorem, itis helpful to notice that if ¥(z) is the z-transform of the sequence (xo, ..X2.*" Js then z.3(z) is the z- transform of the delayed sequence {0, x0, 1, ***} 39 Theorem Consider the feedback system of Figure 6.14 with @(s) replaced by (2) Suppose g€1 and that ® is a memoryless nonlinearity belonging to the sector 0, }. Then the feedback system is|y-stable fthere existsaq 2.0:such thatSec.6.7 Discrete-Time Systems 371 40 Re([1-g(e® ~ 1) ge) > We [0, 2m}, where 41 ymax, 14le%I Remarks 1 , then (40) reduces to (38). So Theorem (39) gives a less restrictive stabil- ity condition than the circle condition, However, itis not clear how one would test systematically whether there exists aq satisfying the hypotheses of the theorem. 2, Comparing Theorem (39) with the Popov criterion [Theorem (6.6.95)}, one can see several significant differences. First and foremost, the continuous-time cri- terion is only applicable to time-invariant systems, whereas no such restriction applies to the present case, Next, in the continuous-time case, both f(s) and s @(s) ate assumed tobelong to A. Such an assumption is not needed in the discrete-time case, since if (2) belongs to /,, then so does the function 2 (2). [Note that the inverse transform of the function z g(c) isthe delayed sequence {0, go. 81. °°") which belongs to / if fe /.] Similarly, in the continuous-time case, itis assumed that wy and its derivative it both belong to L2. Again, such an assumption is not needed in the discrete-time case, since if w2¢/2, then so does the delayed sequence (0, w20, }, which is the inverse transform of ziéa(z). Hence in the discrete-time case one has “true” /,-stability if (40) holds. In view of these differences, perhaps one should not even refer to Theorem (39) as a discrete-time version ofthe Popov criterion, Proof Given sequences x, ye S, define their truncated inner product y as Now the discrete-time analog of Theorem (6.6.58) isas follows: The system of Figure 6.2 is {stable if there exist constants), 8, €, 8; such that $B y Pelle} +3G.x lh, VNEZ,, VreS, for! where ” 44 xllyy=y° [38 and such that372 Input-Output Stability che 45 5, +e) >0,84e,>0. tege-) Loe Fig. 6.31 Redraw the system under studyas shown in Figure 6.31, and note that v>¢12 whenever uz6 1, Also the transfer function Ii ~g(2—W)}e! forall 20, and 46 max | _ oet02e1 | 1=q(e®=1) | Letus firstanalyze the operator G. Define it ~gle®- 1) xe" 47 Bi=, min Re ({1-g(e- D1 Re"), ~ gle = 1) e* 48 wm max IU -g(e’®— 18) ‘Then (40) implies that a 49 po-2 4 @as2—r Pr 4 Next, observe that -ge®-DI- ite = 50 WS max I1—gle*— Ii max 18(e!)1 =(1 +29). ‘Thus (49) and (50) together imply that ; si pot at Now, by analogy with Lemma (6.6.91), itfollows that,See.6.7 Discrete-Time Systems 33 82 y2Bllrll}a, VNEZ,, Vee. Also, itfollows from Theorem (15) that 53° Gyrlly> Spl llya, WNEZ,. VreS. ‘Combining (52) and (53) shows [ef. (6.6.83)] that, for all a0, WNEZ,, VxES. S40 y2B—a)lIxilh +S WG rlhh H ‘Now letus study the operator Gs. With the various quantities defined as in Figure 6.31, wehave 58 yw), Vi, and 86 H(2)=[1-9@- WG), ic., s7 4 (1-4 ~™,), Wi, ‘where w_. is taken as zero. Thus $B <1, Gat >y= Trt, = LOW) wi—4 Dow) (Hy 4 —w)). Let us examine the second summation on the right side, keeping in mind that 0S 0,0w,)4w; Sb, Vi. Now (wi) 591601) 4-1 w= Orsi —WF) Hence ll an Z baa Lowi wis 5 Dwi is & 61-4 F0,00,) 00, -w)2—2 iets 3 Applying Theorem (15) and Equation (46) to314 Input-Output Stability ch. shows that 63 lWwilypSiltllya, VNEZ.. ‘Substituting (63) into (61) gives ¥ 64-4 Low dow1—wy2-B tide iD 4 Let us retum to the first summation on the right side of (58). Since ® belongs to the sector (0, b},itfollows that 65 Sow )w.2L Foo P = ue ,e0% Lain wi 25 EOP = 5 Wat Na Finally, substituting from (64) and (65) into(58) gives 4 2 2 en? 24 2 66 <1,G yt >y2 > WGatllive ‘4 Welle. Now we are in a position to complete the proof. Equations (54) and (66) show that (43) holds with 67 ey =B ‘Note that the constant ois not yet specified. Thus the system under study is /,-stable if possible to choose a.20 such that ee opeat p70 ‘The fist inequality in (68) leads to ‘while the second inequality in (68) leads to 79 Boa- + Hence, if (51) holds, then one can choose «> O such that both (69) and (70) hold, thus estab- lishing the /,-stability of the system. iSec.6.7 Discrete-Time Systems 315 Note that actually (51) isa sufficient condition for l-stability, and (40) is just a more conversative version of his condition with replaced by the bound (50). Notes and References The foundations of feedback stability theory were lad in the early 1960's by Sandberg and Zames. Three papers by Sandberg (1964a, 1964b, 1965) and two papers by Zames (19662, 19666) are cited here, but many more could have been cited. A more complete list of the early references can be found in the book by Desoer and Vidyasagar (1975). The ‘material in Section 6.3 on the relationship between input-output and Lyapunov stability is, based on Vidyasagar and Vannelli (1982); see Hill and Moylan (1980) for related results. The analysis in Section 6.4 on open-loop stability is due to several authors, but Theorem. (6.4.75) isdue © Willems (1969). The graphical stability test of Theorem (6.5.35) isdue to Willems (1969a); see Callier and Desoer (1972) for a more general version. The algebra B of Section 6.5.2 was introduced by Callier and Desoer (1978). The idea of coprime factori- ations over A is due to Vidyasagar (1972, 1975, 1978). See the book by Vidyasagar (1985) for adiscussion of the stability of linear distributed feedback systems in an abstract setting. ‘The material in Section 6.6 owes its genesis to the early work of Sandberg and Zames referred to earlier. The form of the Popov criterion given in Theorem (6.6.95) is due to Desoer (1965). The form of the passivity theorem given in Theorem (6.6.58) is due to ‘Vidyasagar (1977); see Moylan and Hill (1978) fora more general version.7. DIFFERENTIAL GEOMETRIC METHODS In this chapter, we give a brief introduction to some results in nonlinear system theory ‘obtained using methods from differential geometry. In the preceding chapters, the attention hhas been on the stability issue, which can broadly be described as a concern as to whether or not a given system is well-behaved in some sense. The emphasis was on avoiding detailed calculations of solutions to system equations, and on obtaining broad and nonspecific con- clusions. In contrast, in the present chapter the emphasis is much more on the detailed behavior of a system: Is it reachable in te sense that, starting from a given initial state, one is able to steer the system to all nearby states? Is it observable, in the sense that, for each state, there exists atleast one corresponding input which permits us to discriminate between this state and all nearby states, by measuring only the corresponding output? Is the system feedback linearizable, in he sense that, by making an appropriate change of coordinates and. applying a nonlinear state feedback, the equations describing the system can be made to look linear? Differential geometric methods provide a powerful means to address these and other questions. Chronologically, differential geometric methods are of even more recent vintage than input-output methods, and most of the results presented in this chapter date ‘back no more than ten years. As uch, these methods are still the subject of current research, “Most of the results presented in this chapter pertain to nonlinear systems described by a set of equations ofthe form A) =MxCOL+ Zu(OBLALO. where x(1}eR". and £8, °*".fq ae vector fields on R”. (This term is defined in Section 7.1 below.) Two features ofthis system can be observed at once. First, the system is time- invariant, in that there is no explicit dependence on time. The class of systems studied in ‘Chapter 6 can either be time-invariant or time-varying, and the relative efficacy of the analysis methods presented therein remains pretty much the same. In contrast, differential ‘geometric methods are much more efficient when applied to time-invariant systems than to time-varying systems. Second, the system aboveis linear in the control, in contrast oa gen- eral nonlinear (ime-invariant) system with states and m inputs, which is described by =A, uO) Again, the analysis methods of Chapter 6 apply quite well o the more general nonlinear sys- tem, whereas the methods presented in this chapter apply only tothe less general class of systems. This is not a fundamental restriction of differential geometric snethods. In fac, it 316See.7.1 Basics of Differential Geometry 377 is possible to use differential geometric methods to siudy the more general class of nonlinear systems as well. However, the increase in complexity is enormous. Hence the attention in this chepter is focused on the less general, “linear in the control” class of systems, $0 as 10 keep the technicalities toa minimum. While the class of systems studied in this chapter is more restrictive than that in Chapter 6, the analysis is more thorough. The questions of reachability, observability, and feedback linearizabilty are all answered. Asa prelude to this, some background material in differential geometry is presented It should be emphasized that the contents ofthe present chapter represent only a small part of the results that can be found in differential geometric control theory. Also, to keep the treatment at an elementary level, a great many simplifications are made. First, it is assumed that the differential equations describing the system under study are defined on some open subset of R®, rather than on an abstract n-dimensional manifold, Second, all definitions and computations are varied out using so-called "local coordinates.” There is a price to be paid for doing things in his way. The assumption that all the action takes place ‘on some open subset of R rules out many interesting situations, such as systems defined on the circle or a torus. Using local coordinates to compute everything obscures one of the major motivations of modern differential geometry, which is to obtain a "coordinate-tree” description of various entities. A student desirous of a more general treatment of the dif- ferential geometric approach to nonlinear conirol systems should consult an advanced text, such as Isidori (1989) or Nijmeijer and van der Schaft (1990). 7.1 BASICS OF DIFFERENTIAL GEOMETRY In this section, we begin by recalling a well-known theorem from advanced calculus, namely the inverse function theorem. Then the notions ofa vector field, a form, and various types of Lie derivatives are introduced. Suppose f: R" —>IR™, and suppose each component of fs continuously differentiable with respect to each ofits arguments. (In other words, suppose thet fis C'.) Then the mxn matrix whose ij-thentry isdf/ax;iscalled the Jacoban matrix off and is denoted by fx. ‘We say that Fis smooth if every component of fhas continuous derivatives ofall orders with respect to all combinations ofits arguments. Suppose U, Vare open subsets of R" and that £.U—>Vis C', Then we say that fis a diffeomorphism of U onto Vif G)f(U)=V, Gi fis ‘one-to-one, and (ii) the inverse function f": V—>U is also C'. fis called a smooth dif- feomorphism ifboth and are smooth functions. ‘Theorem (1) can be found in most standard texts on advanced calculus, e.g., Roy- dden(1963) 1 Theorem (Inverse Function Theorem) Suppose f:R" > R" is C! at xy¢R", and let yq=flxo). Suppose [OLOX]y~n, is nonsingular. Then there exist open sets U CIR" con: taining xy and VCR" containing yo such that is a diffeomorphism of U onto V. If, in addi tion,Cis smooth, then" isalsosmooth, ie, fis asmooth diffeomorphism.378 Differential Geometric Methods ch.7 Since smooth diffeomorphisms are the only kind of diffeomorphisms used in this book, the adjective “smooth” is often omitted. ‘Throughout the remainder of the chapter, X denotes an open subset of R", where n is a fixed integer (known as the order of the system under study). 2 Definition A vector field on X isa smooth junction mapping X into R". The set of all vector fields on X is denoted by V(X), The set ofall smooth real-valued functions mapping X into Risdenoted by S(X). Note that fa, be (X),then the functions a +b and the product function ab defined by ab: x +4a(X)6(X) are also smooth functions and belong to S(X). Hence $(X) is a ring under the usval definitions of addition and multiplication. As for the set V(X), obviously it isa linear vector space over the real field. But much more is true. Suppose ac S(X)and that fe V(X) (Le. suppose a is @ smooth real-valued function and that fis a vector field on X). ‘Then the function mapping x ito a (x) f() is also a vector feld on X. Moreover, one can verify the following properties: For each a, be S(X),f, ge V(X), we have 3 alf+g)=af+ag, (a+byt=atebt, (@b)f=a(bn. ‘These show that V(X) is a module over the ring $(X), but this terminology is not used further inthis book. 4 Definition A form on Xis a smooth function mapping X into (R")*, which isthe set of Ln row vectors. The set ofall forms on X is denoted by F(X). Note thatitis customary to write vector fields as column vectors, and forms as row vee tors. S Examples Let X= R?. Then the following are vector fields on X: a}-2e2] | coshe, yea} || expt, +29) ‘The following is nora (smooth) vector field, since it does not have continuous derivatives of all orders. (Note that the frst component has continuous derivatives of order two or less, but the third derivative with respect tox is not continuous atx [aan SI eee ‘The following are examples of forms: Note that a form is written as a row vector whereas a vector fields writen as.a column vector.See.7.1 of Differential Geometry 379 [sin (x, ~3x}) x} +x], [exp(e,-2x]) x) -tanh x3]. Next we define various operations involving vector fields, forms, and smooth real- valued functions. Suppose xo€X is given. A curve in X passing through Xy is a smooth function ¢ map- ping some open interval (— \is smooth; i is called the Lie derivative of she function a with respect to the vector field f, and is denoted by L¢a. Note that Lyae S(X). Now suppose he F (X), ie., his a form on X. ‘Then the map 16 xpohootiyy issmooth and real-valued. Itis denoted by and belongs 0 (X). ‘The Lie derivative L¢a can be interpreted as the derivative of a along integral cunes of the vector field. Notice that 17 @eaye= fim + (als 0%0)]-a0%0)) We have already encountered this concept in Chapter Sin connection with taking the deriva- tive of a Lyapunov function candidate along the solution trajectories ofa particular differen- tial equation. Note that, given the differential equation (6) and a Lyapunov function candi- date a, the derivative d defined according to Definition (5.2.23) is precisely Lea. Hence there is no need to give examples ofthe computation of the Lie derivative of a smooth func- tion. As forte quantity , this scaled the inner product ofthe form h and the vee- torfieldf Its nothing more than the scala product of the ow vector h andthe column vee torf.382 Differential Geometric Methods ch.7 ‘A form his called exact if there exists a smooth function a€ $(X) such that h= da. In an abstract setting, it is not always easy to determine whether a given form is exact or not Thisis true evenif Xis restricted tobe an open subset of R", as is done here. However, if Xis an open ballin R", i¢.,a set ofthe form (x: 1x~%o l <€) for some Xo and €, then itis easy to determine whether a given form he F (X) is exact: Form the mn matrix Jy whose ij-th elementis dk Ax; then his exact ifand only if isa symmetric matrix forall xe X. Next we define avery importn: concept 18 Definition Suppose f, ge V(X). Then the Lie Bracket off and g is denoted by [f. dndisthe vector field defined by ae 19 (g)-284- Wal at ge 20 Example LetX = RR? as before, and supposef, ge V(X) are given by pnd sim (ry +22) ] f(x) [x12 |7#9 [cose —22)| ‘Then routine computations show that sin(xy x2) sin(xy 2) |* [os +42) c0s(t1 +42)] and = hg al= Bie (or + x29) 08 (x) +2) —Sin (x) +2) + 2x7 C08 (81 —¥2) © | ay xia +42) sin (x, <2) Sim (x +2)— 2, 608 (8 — 2) ‘The next several lemmas bring out several useful geometric interpretations of the Lie bracket 21 Lemma Supposef, g.are vector fields on X, and letXo€X. Then 1} a 22 {f.ll%)= lim 1 2 tt -e00)} Remarks Unfortunately Lemma (21) makes less sense in the "local coordinates" that ‘we are using than it does in a global or coordinate-free setting. In essence, Lemma (21) states thatthe Lie bracket [f, g] can be thought of as a directional derivative of the vector field g along integral curves off, in much the same way that Las the directional derivative of the real-valued function a along the integral curves of f (see (17)]. However, in the case‘See.7.1 Basics of Differential Geometry 383 ‘of an abstract manifold, one cannot just subtract the vectors gs, (Xp)] and g(%o), since they "live" in distinct tangent spaces. The extra factor [S¢-(eyay) "Pulls back" the vector ‘8ls4,(%p)] into the tangent space of X at xy; the need for this factor can only be appreciated in ‘a more abstract setting. Proof Since we are taking limits as 1 ->0, itis only necessary to compute the various ‘quantities inside the braces up toa first order term in . For small, we have 23 ,s(%q) =Xo +f) +000). Hence 24 plsy(xo)]=BIX + ef(x9)] +000 =acoe| 3] Woo) +00. Again from(23), 25 s¢_,(x)=x—Hf(x) +0(0), sothat O8-(X) at 6 ox 21-15 +000. Hence A.A) [ ae at wir aan « since 5, (Xo) =X to zeroth order in. Substituting al this in (22) and gathering terms shows that the right side of (22) equals 2 tin fu-srieen- a} HBr Ltt where the argument xo has been suppressed for clarity. Ml 29 Lemma Supposef, ge V(X), and let xq€X. Then 1 30 {f,8]0%0)= him 7 {8.884010 — Xo) Remarks Lemma (29) gives yet another interpretation of the Le bracket. Suppose we start ata point xp, and follow the integral curve ofthe vector field f for avery short time f: then, from that point, follow the integral curve of g for a duration; then, from that point,384 Differential Geometric Methods ch.7 follow the integral curve off backwards in time for a duration f; and finally, from that point, follow the integral curve of g backwards in time for a duration 1. Where do we end up? Wel, tofirst order in, we get back tothe point x». However, to second order int, we end up atx +1°UF,gh(%9). Thisis ust what Lemma (29) states. Note that the map s,. isthe inverse of the map s,,and similarly Sis the inverse ofthe maps. Hence, ifthe maps , and S,, commute (.e., 5, =5g st, forall sufficiently small), then the limit in (30) is zero. Thus cone can think of the Lie bracket {f, g] as a measure ofthe extent to which the “solution” maps 51, ands fail tocommate. This point is developed further in Lemma (40) below. Fig.7.1 Proof ‘The proof is quite routine, hough tedious. To follow it, refer to Figure 7.1. To compute the limit in (30) itis necessary to compute the various quantities up to second order int. By definition, 31 Ls (xg)] =f) $s] = M5] Hence 5 = 0%), 4 32 Flo, Bada Next, 1) implies that e at 3 Tylor Godlao= Se AO) Now, by Taylorseries, 2 ! 34 ax 4 tlno)+ FE f(x) + (07). Next, in analogy with (34),See. 7.1 Basics of Differential Geometry 385 2 38 =x, + 180K) + > 8 aX.) +007). Let us now substitute for x, from (34) and simplify by neglecting all terms which are of order higher than 2, Thus it is only necessary to estimate g(x,) to first order in ¢ since itis, ‘multiplied by 4, and itis safe to replace (2/@xK(X,)'R(%,) by [95/0X](Xy)-B(%o), since this termis multiplied by 7. Since 36 B(%,)= 8040) +1 8 my) fix) +000, (35) and (36) imply that 37 xy =xy +1 [flx) + 81%] +0107) al ot, + 2B, hee, FOS Fy f0) + FE) x9) + 5 SE 0) BUH) ‘The process is now repeated, and the results are shown below; the argument is x) unless indicated otherwise. re ole meee: 3B XC =x, Hf(K,) + z ax wr ep) +.0(07) 12g 1 Eur? 5 Ee F ai+00, exten Pte 38g tate + F Bea) em 4007) sure| ra] 0, ‘This completes the proof. Inthe remarks prior tothe proof of Lemma (29), it was stated that [f,g] is a measure of the extent to which the solution or “flow” maps sy, and s, fail to commute. The next result sheds more lighten the relationship. 40° Lemma Supposef, ge V(X). Then AL (6 g]=Oilhse552=5p Se. We-tsulficintly smal Lemma (40) states that the Lie bracket of two vector fields fand g is identically zero if and only ifthe solution maps sand s, « commute forall sufficiently small, ¢. Actually, it is easy to see that if the commutativity relationship (41) holds forall sufficiently small ¢, +, then in fact the same relationship holds for all, «for which the solution maps are defined.386 Differential Geometric Methods Ch Proof “If” Follows directly from Lemma (29). “only if” LetxoeX be arbitrary, and define ] 42 ett)=: jets e00| ste ‘Theneisalsoa vector field. In fact, comparing with (9), we see that itis the transformed ver- sion of g under the diffeomorphism s,,. Let us compute the time derivative of e(1). By definition, Sine ~ 8 e)= lim [ele +)-e(0] a -[ 2-000} | a = lim) | aS #] { (os) Now letus observe that Screen Tnother words, sc isthe composition of s< followed by ,.. Hence, by the chainrule, a 158118) a a y 9] co Fete, Now define 16 2=54(%), and apply (45) in 43). Using the fact that AT S¢o42(%a) = 842) in(45) gives r a 48 [dx] Similarly 49 Bloerax(¥0)] =Bl5<(2)]- ‘Substituting from (48) and (49) into (43) giSee.7.1 Basics of Differential Geometry 387 a Poa -«] afs,(2)]-g(@). sett) Butby Lemma (21), m8 ott ghey 532 lim, fa =[f, gz) =0. Hence ¢0, which means that e(r) = €(0) forall (when itis defined). Tocomplete the proof, note thats; 9(x) =x. Hence vx 3 dso) ‘As aconsequence, from (42), €(0)= g(x). Coupled with the fact that ¢=0, this implies that ¢(1)= g(%) V1. From the definition (42) of e(t) this means that fa ] su [da.one] ao Now we make use of the vector field transformation formula (9). Fix re R, and note that s-_, is a local diffeomorphism around xp. Apply the formula (9), with f replaced by g, and T replaced by sy, [and note that (s_..)"' =s,,.]. Then (9) shows that 58 gy) [s 109) 200| 200). set ‘where the last step follows from (54). (Recall that X is arbitrary.) This means that the vec~ (or field g remains invariant under the diffeomorphism s,.. To put it another way, if {f, g]=0, then the vector field g remains invariant under the flow of the vector field f. Now apply (11) with f and f; replaced by g, T replaced by s,.,and the time variable rreplaced by 1 Thisgives 36 SiorSe2 See St Pre- and post-multiplying both sides by , gives 57 Spc 8t1 StS p ‘This completes the proof ofthe "i" part. i388 Differential Geometric Methods ch. One last question before we move on to other topics. What happens to Lie brackets ‘when we change coordinates? Suppose f(x), g(x) are two given vector fields, and we make a coordinate change y=T(x). Then, as discussed above, f(x) and g(x) get transformed into fr(y) and gr{y) respectively. Now one can compute the Lie bracket of the vector fields either before or after the coordinate change. Do both procedures give the same answer? In other words, isittrue that ery) AY) Sy TH) er)? 58 (JOO, B10) ory) The reader should not be surprised to lear that the answer is yes. One can of course verify (58) directly by substituting for the various quantities. But a more "modern" reason for believing (58) is to note that the Lie bracket of two vector fields is defined in terms of the behavior of certain integral curves [see Lemma (29)], and the transformations of the vector fields, from fand g tof; and gr respectively, are intended precisely to ensure that the integral ccurves match in the two coordinate systems. Recall that if fs a vector field on X and a is a smooth real-valued function, then Lea is also a smooth real-valued function defined by (15). The next lemma relates repeated Lie derivatives tothe Lie bracket. 59 Lemma Suppose ae S(X)andf, ge V(X). Then 0 Lyggia=LelL ga)—LylLra) Proof The result is established via routine though lengthy computations. By definition, G1 Cpl gay =ViL a)00) fe, Hence, tis useful to compute V(L ga). Now © Lyaa=Vatn) = F 2s, . Therefore 2 -$/ 2a, , da 28) $B Mbaahi= g, ba) = Bl Beas, 8* Dey Or, ‘This can be concisely expressed. Define V2a to be the nxn matrix whose j-th element is a Ard. This matrix Va iscalled the Hessian matrix ofa, Note that V2ais symmetic. ‘Now (63) can be expressed asSec.7.1 Basics of Differential Geometry 389 OH Vilga=g Via +Va 2, ox" Therefore 65 Lplga=V(L_a) f= ViateVa Bf, and {Vat Vag+ Var 2, fo 8: ant ax® However, since V2ais asymmetric matrix, we have 7 gVat=l Vag, andso os of 68 Lplga—Lghya=Va ie FP Hiceit Thiscompletes the proof. Itis possible to prove a more general result than Lemma (59), using the concept of the Lie derivative of a form with respect 19a vector field. 69 » Definition Suppose fe V(X) and that he F(X). Then the Lie derivative of h with respectiotisalsoa orm, andis defined by ) nat Jaw Note that h isa column vector, and that hx is just the usual Jacobian matrix of h’ So far we have defined three types of Lie derivatives: Suppose f, ge V(X), a€.5(X), and he F(X). Then the Lie derivative ofthe vector field g with respect to fis just the Lie bracket {f, g). The Lie derivative of the smooth function 2 with respect o fis defined in (15) as Vat. The Lie derivative of the form h with respect to fs given by (70). Note that the Lie deriva- tives ofa vector field, a real-valued function and a form are again respectively a vector field, a real-valued function and a form. These derivatives can be related via a Leibniz type of product formula, 71 Lemma Supposef, ge V(X) and he F(X). Then 2 Li=+. Proof As usual the proof follows just by substituting forthe various expressions and clearing terms. Note that (9) isjust (x) g(x). Thus390 Differential Geometric Methods ch7 ag, | a af ay 73 (Vb, g>1,= 5 (hx) g001= ahh ae i Fal 21 i, or. in other words, [ef. Definition (69)], 4 V ++ Hence 76 + = + at gsr Ae ay] ox® Fl ox ox8| 2, axe The equality ofthe two quantities in (75) and (76) follows upon noting that f (Ah /2x)’g is justa scalar and therefore equalsits transpose. i ‘Some other properties ofthe Lie bracket are ready consequences of the definition. Lemma Suppose, g, he V(X), ae S(X), ando, BER. Thent it (ef, " 78 (f.og+Ph}=alf, g]+BIE hb), » 80 {f-fg. hi) +(e, h. f)+h,(f gl)=0. 81 fag) (hel+(Leae. Remarks Equation (79) displays the anti-symmetry of the Lie bracket, Together (78) and (79) show the the bilinearity ofthe Lie bracket. Equation (80) is known as the Jacobi identity, Equation (81) is type of product rue. In fact, if we replace the Lie bracket [6 §} by the Lie derivative symbol, then (81) can be rewriten asSee.7.1 Basics of Differential Geometry 391 82° Le(ag)=aLyg+(Lea)g, which looks just like a product rule. Proof Both (78) and (79) are ready consequences of Definition (18). The formulas (80) and (81) can be established through rostine computations; the details are left as an exer- cise. Suppose f), °:+, f,¢ V(X), and xeX. Then we say that the vector fields fy, °*", fj are linearly independent at x ifthe column vectors f)(x), °--, f(x) are linearly independent (over the field of real numbers). Linear independence of forms is defined analogously. Itis. clear, by virtue of continuity, that iff, **-, fare linearly independent at x, then they are in fact linearly independent at all points in some neighborhood of x, ie. in open set containing. x This section is concluded with one last bit of notation. To denote repeated Lie brackets of vector fields, itis convenient to introduce the "ad" symbol. Given, ge V(X), we define 83 ad?g=s, ade"'g=[f, adzel. Thus 84 adig=lf. g). ade=l6. (fg), and soon. Problem 7.1 Compute the Lie brackets of the various vector fields defined in Examples (S)and (20). Problem 7.2 Prove the following altemate version of Lemma (29): Iff, ge V(X), then [h8l)= tin Hie 8 ws stt-x} Problem 7.3 Supposef, ge V(X) are constant vector fields. Show that iC, Problem 7.4 Let Aff denote the set of affine vector fields on R", ie., the set of vector fields of the form f(x) =Ax+b, AcR™, DER", Show thatthe set Affs closed under the Lie bracket, ic. that [f gle Aff whenever f, ge Aff. Problem 7.$ Suppose fe V(X), and a, be S(X); Le., suppose fis a vector field and a, b are smooth functions. Prove the Leibniz-type product formula392 Differential Geometric Methods Ch.7 Lelab)=a Lb +b Lea. Problem 7.6 Suppose a¢ 5(X), he F(X), and fe V(X). Using Definition (47), prove the product-type formula Ly(ah) =(Layh+ al ch, Problem 7.7 Using the Jacobi identity (60), prove that iff, g, he V(X), then. Lefe.h)=(Lye.h)+ fe.Leh 7.2 DISTRIBUTIONS, FROBENIUS THEOREM In this section, we present a useful too! in differential geometry, namely the Frobenius, theorem. Along the way we introduce important concepts such as submanifolds, distribu- tions, and involutivity. 1 Definition A subser MX is a k-dimensional submanifold ( , n}. Under these conditions, the following statement is true for each xe UW: The (n~k)- dimensional subspace of (R')* spanned by the row vectors (AQ, .4(X), °° ,Aq(%)) is the same as the (n-K)-dimensional subspace of (R")* spanned by the row vectors [dyj.1(). °°, dy, (0). ‘The proofs leftas an exercise, 10 Definition Suppose M is a k-dimensional submanifold of X, and choose smooth func tions $4.5, *' +4 such that the conditions of Definition (1) are satisfied. Then the tangent space of M at xe M isthe k-dimensional subspace of R" defined by394 Differential Geometric Methods ch.7 TW TM,=(veR": \ for i= k+l, sn) A vector field fe V(X) is said o be tangent to M at x i/fix)e TM. It follows readily from Lemma (8) thatthe above definition of TM, is intrinsic, i., does not depend on the particular choice of functions used to represent Min the vicinity of x. In other words TM , is just the k-dimensional subspace of column vectors that are annihi- lated by each ofthe n —Krow vectors d6, (8), °°, iy(x) (or, to be more precise, the sub ‘space spanned by these n— k row vectors). ‘There is another way of looking at submanifolds with which its relatively easy tocom- pute. Suppose M is @ k-dimensional submanifold of X, and that xy M. Then, according to Definition (1), there exist an open neighborhood UX of xp and smooth functions cor 774 Gq Such that dd; 46x), ***,€6,(0 are linearly independent at all xe U, and such that (2) holds. Now pick smcoth functions 6,, ***, , such that (xq) =0 for #=1, +++, k and such that (46,(xo), i= 1, “+, ) isa (row) basis for R®. This is actually quite easy to do. One could even just choose 0;(x) = v;(X — Xp), Where ¥,, ***, ¥; isaset of (constant) row vectors chosen such that {Vj. ° , Vi iy +1 (0), ***, (Xo) is @ row basis for R". Now define amap 7: R" — R" by R (Tx), =0(x), sis By construction, the Jacobian 37/0x evaluated at xy is nonsingular. Thus, by the inverse function theorem (Theorem (7.1.1)], Tis locally a diffeomorphism, say on Up G U. One can think of y), +". a8 a new set of coordinates on Up. What does the submanifold M look likeinterms ofthese new coordinates? Comparing (2), we ee that B wowe(9] newer'}, ‘where is some neighborhood of @in R‘, In other words, a k-dimensional submanifold of X isa set which, after a suitable smooth change of coordinates, looks like a k-dimensional "slice" of X. ‘The tangent space of M also has a simple form in the new coordinates. If we perform the coordinate transformation 12), then in terms of the y coordinates, 6; is just the i-th coor- dinate of y. Hence do, is just the -th elementary row vector, with | in the i-th position and zeros elsewhere. To compute the tangent space at yo we can apply the formula (11) and “observe that a column vector vis annihilated by each of the elementary row vectors with a | in positions k-+1, «+, n respectively if and only if the last n —& components of v are zero. Thus, ify =7(%), then 14 TMy,=(veR":v=[v, OY,¥.R'),See.7.2 Distributions, Frobenius Theorem 395, = [veR" j2Ofori=k+1, .n). ‘A vector field fs tangent to M at a point xp if and only if the coordinate-transformed vector field fy {defined in(7.1.9)) has the form t | 5 to=l0,,) ‘The next definition introduces a very important concept. 16 Definition A k-dimensional distribution A on X is a map which assigns, taeachxe X, @ k-dimensional subspace of R" such thatthe following smoothness condition is satisfied: For each xyeX there exist an open set UX containing Xq and k vector fields fy, °"~. fk such tha (i) (f,(). = (8) ia linearly independent setfor each xe U, and (i) 17 A(x)=span (£00, ++, f4@)), Vxe U. Itis often convenient ro describe a distribution in terms ofthe vector fields that generate it, Thus, given k vector fields f), -. fy, we often define A by the formula (17). But itis {important to remember that A(x) isa subspace, and that (f(x), "+", f,(x)) isa, not the, basis, for it, To amplify this point further, suppose f, and f; are two vector fields on X with the pro- perty that f,(x) and f(x) are linearly independent at all xeX. Then span (f), 2} and spah (fy +f, f, ~fz} describe exactly the same distribution. Indeed, in order to define a k- dimensional distribution, itis often convenient to define it as the span of k or more vector fields, of which some kate linearly independent at each point. In attempting to describe a distribution by means of a set of vector fields that generate it, one can get into the following difficulty: Suppose f, --". fe V(X) are given, and we define A by (17). Then i can happen that the rank of the matrix (F(x) -°f,(x)] is not con- stant as x varies, To get around this difficulty, itis possible to define a k-dimensional distr- bution as a map wich assigns to each xe X a subspace of dimension no more than k, and then require that each open set U contain at least one point y such that the dim A(y) exactly equals k. If this definition is accepted, a great deal of verbal awkwardness is avoided because, in subsequent sections, distributions are invariably defined in terms of some gen- erating set of vector fields. In such case, if A(x) is actually ak-dimensional subspace of R", then we say that xis a regular point of A. Altematively, x isa regular point of the disribu- tion A if there is a neighborhood U of x such that the dimension of A(y) is the same for all ‘ye U. Finally. if is a given distribution and fe V(X), then we say that f belongs to A if Tox)e A(x) Vxe X, and denote itby fe A. Suppose A is a k-dimensional distribution, that UGX is an open set, and that fy." fe V(X), m2, are vector fields that span Aon U, Now suppose fe 4. Then, using Definition (16), one can show that there exist smooth functions a, «°°, dye $(X)stich that396 Differential Geometric Methods Ch.7 18 fix) Law f(x), Wxe U. is essential to note that the "coefficients" in (18) are not constants, but are smooth fune- tions ‘An interesting question in differential geometry isthe following: (Unfortunately the brevity ofthe present treatment does not permit us to explore fully just why tis question is interesting.) Suppose is a given k-dimensional, everywhere regular, distribution on X: for each xeX, does there exist a k-dimensional submanifold M, of X containing x such that every vector field fe is tangent to My at x [ie., TM, =A(x)]? If such 9 submanifold My exists for each xe X, then the distribution A is said to be completely integrable, and Mis said to be the integral manifold of 4 passing though x. But the question is: When is adistr- butioncompletely integrable? Anclegant answer to this question is provided by the next result, commonly known as the Frobenius theorem. 19 Definition A distribution Ais involutive f/f. gle A wheneverf, ge. nother words, a distribution is involutive iit is closed under the Lie bracket, 20 Theorem (Frobenius) A distribution is completely integrable ifand only ifit is invo- lutive. Ieturns out thatthe “only if" partis quite easy to prove, and the "if” part isreally the sub- stantive par ofthe theorem. Proof "Only if” Suppose A is a completely integrable k-dimensional distribution. ‘Then, for each x€X, there corresponds an integral manifold M, of dimension k. By Definition (1), this means that, for each Xo€ X, there exist a neighborhood U of xp and smooth functions $,,1, “*,@,€5(X) such that the differentials of the @, are linearly independent row vectors at xp, and such that each function 6, has constant value at all points, in M,,. Now, if we select smooth functions ),°°*,€5(X) such that {d0,(x9).i=1, ++.) is @ row basis, then the map 7: U—+R® defined by (12) is a dif- feomorphism over some neighborhood Up U of xy. Moreover, since each fe Ais tangent 10M ,ateach x€ Up, it follows that in the new coordinate system each fy€ Ar has the form fy) 21 triyy=|q, , |. Y¥ETWo) ‘The key point is to note here is that the tangency relationship holds ar all points in some neighborhood. Moreover, since the tangent space TM , is precisely A(x) forall xe Uy, itfol- lows that Ay is precisely the set of vector fields of the form (21). Now suppose fr. g7€ dr. ive. thatSec.12 Distributions, Frobenius Theorem 397 toy ] 8.9) 22 Hadela, | erty 0,4 ‘Then it readily follows from the formula (7.1.19) that the bottom n—k elements of [fr, Br are also identically zero. Thus {fy, grle Ar and hence Ay (i.e. A) is an involutive distribu- tion. If” See Appendix C. ‘The “it” part ofthe proof is by induction en the integer & (the dimension of the distribu- tion A). If k= 1, then A is just span {f} where f is a nonzero vector field. Now a one- dimensional distribution is automatically involutive, since {af, bfle spanf=A whenever 4a, be $(X), In this case, the integral manifold M , of A passing through xis just the integral curve of the vector field f passing through x. For #22, the proof is more complicated, and is, sgivenin Appendix C. Every distribution contains an infinite number of vector fields, for example, if fe A, then afe A forall ae S(X). However, in order to check whether a distribution is involutive fF not, it is only necessary to compute a finite number of Lie brackets. Suppose f,, .fqe V(X) span the k(Sm)-dimensional distribution A over some open set U CX. ‘Then every fe A has an expansion of the form (18). Hence, using the bilinearity ofthe Lie bracket and the various product formulas in Lernma (7.1.77), one can easily establish the following statement: Ais involutive if and only iff, f,]€ 4 for each i, j;or, in other words, there exist smooth functions aye $(U),1 Si, j. Sm, Such that 2B [htla)= Lay(o ho, vxe v, This leads to an alternate form of the Frobenius theorem, which is the form actually used in later sections. 24 Theorem (Alternate Frobenius) Suppose f,. “°°. fye V(X), NCX is an open set, XEN, and that the set {f,(x), °*.fq()} contains k linearly independent vectors at each XEN. Then there exist functions Oxsy. "** Qy€5(X) such that (i) A (Xp), *** + 4q(%0) are linearly independent, and ii) there exists a neighborhood VC Nof Xo such that 25 0 We V, fork+ISiSn, 1S/Sm, if ind only if the distribution spanned by {,, **- fy is involutive, ie., there exist smooth {functions &,y anda neighborhood U GNofx such that (23) holds. Remark Suppose the vectorfieldsf,,---.fi, are given by398, Differential Geometric Methods ch.7 0 Inother words, fis aconstant vector field with a"I" inthe i-th position and zeros elsewhere. Define &= span (fy, °:+, f,}. Then A isa k-dimensional distribution consisting of all vector fields ofthe form £09) 27 tw=l9 Itis clear that A is completely integrable. Indeed, given xe X. the corresponding integral ‘manifold M,, isjustthe set 2B [XER" 4) =o, fork+1 SiSn}, and functions 9.15‘, 0,€:5(X) satisfying (25) are given by 29 OX) =x, FKL ov ‘The point of the Frobenius theorem is that all completely integrable distributions can be made to look like this after a suitable change of coordinates. Of course, the theorem is non- ‘constructive in the sense that it ells us that a suitable coordinate transformation exists — it does not tell us how to find it. Nevertheless the Frobenius theorem isa very useful result, 30 Example Suppose X is an open neighborhood of the point x9 =[1 1 11’ in R?, and ‘consider the vector fields 2G 0 fox) =) 23 +215], 800=| 3 [2a o ‘Then itis routine to verify thatSee.7.3 Reachability and Observability 399 o | Uf B10) =| x3 ~Gra tx9)x5 | =O) R00, ° ] where (y= Gn +25) 5 isa smooth function of xin any neighborhood of x that doesnot intersect he surface x5 =0. Hence the two-dimensional distribution A spanned by f and gis involutive if we define X to ‘be the open ball of radius | in R° centered at xy. Now the Frobenius theorem assures us that there is a smooth function he $(X) such that (x)=0, Vex, ‘These are two simultaneous partial differential equations that the function h needs to satisfy. In general, itis not easy to solve these equations and find such a function, though the Fro- benius theorem guarantees that such a function exists. In the present instance, A(x) #247! 4? isa possiblesolution, This means that, if we look atthe surface in X defined by 2aj! 423? =constamt, then at each point x on the surface, the tangent plan¢ tothe surface atx isthe span of the two vectors f(x) and g(x). 7.3 REACHABILITY AND OBSERVABILITY ‘This section has two parts In the frst part of the section, we study the reachability of ‘nonlinear control systems ofthe form. x) +m: BX), a Bn are given vector fields in V(X) and X CR" is a given open set. The questions studied are the following: Are there simple necessary and sufficient conditions for the system (1) to be reachable? If the system (1) is nor reachable, can one carry out a ‘change of state variables in such a way that the "maximally unreachable” partis explicitly displayed? Finally, do these conditions reduce to the familiar conditions for the linear sys- tem400 Differential Geometric Methods h.7 20 xAx+Ru to be reachable, so that the nonlinear results are true generalizations of the corresponding results forlinear systems? {nthe second part ofthe section, we add an output or measurement equation of the form 3 y=hx) to the system description (1), and ask when such a system is observable. Equation (3) is the ‘nonlinear analog of the output equation 4 ystx for linear systems. The questions studied are the following: Are there simple necessary and sufficient conditions for the system (1) ~ (3) to be locally observable? If the system is not locally observable, can one carry out a change of state variables such that the system is decomposed into an observable part and an unobservable part? Finally, is it possible to decompose a nonlinear system of the form (1) ~ (3) into four subsystems, which are respec tively reachable and observable, reachable and unobservable, unreachable and observable, and last, unreachable and unobservable? 73.1 Reachal We begin with a definition of reachability 5 Definition. The system (1) is said to be (locally) reachable around a state xe X if there exists a neighborhood U of x such that, for each x€ U, there exista time T > Oanda set of control inputs (u(t), t€[0, 7}, 1SiSm) such that, ifthe system starts inthe state Xp at time 0, then it reaches the state xy at time T. . Note that the above definition is purely local: The system (1) is reachable around xy if every state sufficiently close to x can be reached from Xp. It is very difficult to analyze the “global” reachability of (1) since nonlinear systems do not satisfy superposition in general. ‘Actually, one can argue that local reachability is all that one can prove, and that the ‘equivalence of local and global reachability is a property peculiar to linear systems alone, ‘Also, for linear systems, there is complete equivalence between reachability (the ability to reach ary desired final tate from a given initial state) and controllability (the ability to reach a fixed final state from any given inital state). This is no longer the case for nonlinear sys- tems. Accordingly, the discussion here is restricted to reachability alone. For a thorough discussion of various nuances of this topic, see Isidori (1989) or Nijmeijer and van der ‘Schatt (1990) The discussion on reachability encompasses several topics. First, the notion of invari- ant distributions is introduced, and itis shown that the notion isa natural generalization, to nonlinear systems, of the notion of (A, B)-nvariant subspaces fr linear systems [see ¢&. Wonham (1979). Then itis shown that every system can be transformed locally to aSee.7.3 Reachability and Observability 401 reachable part plus a “totally unreachable” part. This extends the canonical decomposition of linear time-invariant systems into a controllable part plus a completely uncontrollable Par, introduced by Kalman (1963), to nonlinear systems. Finally, the familiar rank test for the reachability of linear systems is extended to nonlinear systems. All inall, the message is that, with appropriate tools, most ofthe familiar results for linear time-invariant systems can be extended ina natural way to nonlinear systems. To motivate the various ideas, we begin with a brief discussion of linear (time- invariant systems. Consider the linear differential equation 6 x)= Ax() + Buco). where x(2)e R",u(r)e R", Ae R™ and Be R™", Then a subspace M of Ris said to be A~ invariantif 7 AMcM. If, inaddition, 8 BIR")cM, then Missaid tobe (A, B)-invariant, Such subspaces help us to analyze the system (6) inan abstract way. For instance, suppose xy¢M; then it is clear from (6) that x(r)eM 120. Hence, if we express R” as a direct sum M +N, then the system equation (6) must have the form 9 iy =Aym +Aiam + Bin, i=Anee. Whore x; +2 is the decomposition of the vector x. Hence, itis worthwhile to make M as small (i¢., as low-dimensional as possible. Also, a necessary and sufficient condition for the system (6) tobe reachable is that there is no nontrivial (A, B)-invariant subspace of R", ice. the only (A, B)-invariant subspace of R" is" itself. To extend these ideas to nonlinear systems of the form (1), a few preliminary definitions and results are needed. 10 Definition Suppose A isa distribution on X, andfe V(X), Then Ais said to be invart- antunder , orf invariant, /{f, he A'Vhe A. This definition generalizes the notion of an A-invariant subspace for linear vector fields. Let X=IR", M a k-dimensional subspace of R", and let Ac R™*. Let f(x) be the linear vector field Ax, and let A be the distribution generated by the constant vector fields Vj, 7s Mes where (Vj, “"*, %q] isa basis for M. It is now shown that the distribution Ais f- invariant if and only if AM =0, ). Vx U. ‘Thus h satisfies the two partial differential equationsSec.7.3 Reachability and Observability 407 Ay(xy + 2x2 44x 3)+2hgxe + 3h3x5 I haes where h;=0h x, and in the second equation the extraneous factor ~ 48 has been omited. solution of the above pair of panial differential equations is givenby hod=aps. To put the system in the form (26), itis necessary to find a new set of variables y such that y3=/(X), and such that the transformation from x 10 y is a local diffeomorphism. An ‘easy way to do thisisthe following: Compute dh(m)=[0 -3 2} Choose two other row vectors @ and b such thatthe set (a, b, di (xq)} is a row basis for R’, and define X93 AW). ‘Then, since the Jacobian matrix of this transformation is nonsingular at xp, it follows from the inverse function theorem [Theorem (7.1.1)] that the above map is a local diffeomor- phisth in some neighborhood of x. In the present case, a simple choice is a=[1 0 0/,b=(0 1 Oy, and v1 u TO=| x 3 aad ‘Then the inverse transformation is =T'y)=|y2 [2] we ‘The system description in terms of the new variables can be computed directly, or by using the definition (7.1.9) for transforming vector fields. Let us adopt the laner procedure. The Jacobian matrix ofthe transformation Tis given by 1 0 oO J@-|0 1 0 0-376 2g So the vector field A (ce) gets transformed into408 Differential Geometric Methods Ch.7 fy) =JODAXI gy) -l4xy | [-14y¥yy2 =| 0 ° 2x) 1985] -38y wri L ‘Similarly the vector field Bx =: g(x) gets transformed into [ xy 42x) 44ry [vp #2y2-44y32y}? ay - 2y2 [anther tegsr85| gy ° ary) = Hence, in terms ofthe new variables, the system equations are Fy aay PN? 4 +292 44937 V1), Jae 2uy2 Ja =-38ys, ‘So the variable y, is uncoupled from the other variables y and y3 as well as from the input ue Next, we explore the relationship between the distribution 4, produced by Procedure (28) and the familiar matrix 36 W=(BAB---A’"'B) associated with the linear control system (6). In particular, itis shown tha, if one views each column vector of W as defining a constant vector field on R', then the distribution 4, pro- ‘duced by Procedure (28) is precisely the span of the column vectors of W. Now the system (6)is of the form (1) wit 37 x)= Ax, g)=b,i=1, 5m where b isthe /-th column of the matrix B. Applying Procedure (28) gives 38 Ap=span (by, “°°, ba} =span B. Since each of the vector fields generating Ay is constant, Ay has the same rank atall xg R", .e,,every xy¢ R" isa regular point of dy. Now itisclaimed thatSec.7.3 Reachability and Observability 409 39 A,=span{A'bj, 1S/Sm,0 0, the set of states reachable from Xo at time T contains a neighborhood of x. Accordingly, suppose T > O is specified. Let e,, «+, €, denote the elementary unit vectors in R®; that is, , contains a1” in the i-th row and zeros elsewhere. Now by the assumption that the linear- ized system (48) is reachable, there exists a control input ¥() (which is an m-vector valued function) that steers the system (48) from the state z= Oa time O tothe state Indeed there are n such input functions v'(-), +--, ¥"(). Now, for any r=[ry*-*rl’€R", define the control input SI w= Env Let x(, 1) denote the solution of (1) starting atthe intial condition x(0)= xp, and with the input vp. Then x(7, 1) is well-defined whenever tril is sufficiently small, say rll <2. Let B, denote the ball in R* of radius eand centered at, and define h: B, > R" by 82) b(r)=x(7. 0). ‘Note that (0) = x; ths follows from the fact that f(x) =0. Itis now shown that [JhvAr](x) is nonsingular. It then follows from the inverse function theorem [Theorem (7.1.1)} that locally his a diffeomorphism around x). Thus, given any x, sufficiently close to Xp, there existsan re R® such that 53 x/=hi (T, 0). Hence the input uy defined by (1) sters the system from xy to. This shows that the proof iscomplete once itis established that (2h/Or)(x is nonsingular. Forthis purpose, note that x satisfies the differential equation XG, DI =AG, HIE, Er (W)C) wx, 54 A, flat, H}+ Eu Interchanging the order of summation and differentiating with respect tor, givesSee.7.3 Reachability and Observability 413 ax(t, Fr) 5s “e xn +E) gets 1+ E00), (| Xen, am a mun Define $6 mio=| A 2] 0 ‘A differential equation governing m,(°) can be obtained from (55) by substituting r=0. In thiscase x(¢, 0) =p, ot 57 0.0) 58 gilt, O))=8)0%0)=boo. Also, if =0, then the las term onthe right side of 55) drops out, since (u,)pa»=0. Hence 59 y(t)=Aq my(t)+ Fv!) (0) jo = Ap my (0) + By WC). By the manner in which the functions v,(-) were chosen, we know that 60 mi7)= Hence abe) _ ax(T. 2) Oe ar which is obviously nonsingular, i 62 Example Consider a mass constrained by a nonlinear spring and a nonlinear viscous ‘damper, and driven by an external force. Such a system is described by med) +k(r) =u, \here the various quantities are defined as follows: ‘m — Massof the object414 Differential Geometric Methods ch7 1 Positionof the object, Frictional damping force k Restoring force of the spring Using the natural state vector ‘one can represent the system dynamics in the form (1), with ” ° = dean) hey) 8 1 Now let xo = 0, and note that (0) =. The matrices defined in (48) become 0 1 0 = 40) -€'0) iM where k’,d denote the derivatives of k and d respectively. Itis easy to sce thatthe linear- ized system is reachable, whatever be the constants k’(0) and d’(0). Hence, by Theorem (46), the original nonlinear system is also locally reachable around 0. fw 13.2 Observability Now let us study the observability of systems described by the state equation (1) and the output equation (3). First, a few concepts are introduced. 63 Definition Consider a system described by (1) and (3). Two states x and x, are said 1o be distinguishable ifthere exists an input function u() such that 4 yC.%0, WEG. XW where y(-,%;,W), i= 1,2 és the output function of the system (1) ~ (3) corresponding to the input function w() and the initial condition x(0)=X;. The system is said to be (locally) observable at xe X if there exists a neighborhood N of Xp such that every XEN other than Xp is distinguishable from X, Finally, the system is said be (locally) observable if it is locally observable at each Xo¢ X. As isthe case with reachability, there are several subtleties inthe observability of non- linear systems that have no analog in the case of linear systems. These are illustrated through several examples,See.7.3 Reachability and Observability als 65 Example According to Definition (63), two states Xp and x are distinguishable if (64) holds for some choice of input function u(:); there is no requirement that (64) hold for alt inputs w(). Now for a linear system described by (2) and (4), its easy to show that the following three statements are all equivalent: (i) (64) holds for some input u(); (ii) (64) holds forall inputs u();(it) (64) holds with u=0. This is because, in the case of alinear sys- tem described by (2) and (4), the output y is the sum ofthe zero-input response and the zero- state response, Since such a decomposition is not possible in general for anonlinear system, the above three statements are not equivalent inthe case of a general nonlinear system. To illustrate this point, consider thebilinear system (010). Let xp =0. Suppose we set u=0. Then a routine calculation shows that the pair (c, A) is not observable. In particular, ifwe letx, =(1 0 Oj’. then cA | x =0, cA? By familise arguments in linear systema theory, this implies thatthe states O and x; cannot be distinguished with zero input, i, However, suppose we choase a constant input u (1 become ‘Wt. Then the system equations K=(A+B)x yoex, ‘Now, as can be easily verified, the pair(c, A+B) is not observable either, but © (A+B) | x1 #0. (A +BY. ‘This means that416 Differential Geometric Methods Ch.7 9660, D#yC. x, D. Hence the states 0 and x, are distinguishable [by the constant input w(t) = 1 V4. Wt, Then the Levuscarry on the argument. Suppose we apply a constant input u (# state equation becomes, K=(A+KB)x. Now © ] foro (A+B) | =| 0.1] =:M, say. casi] [oxo Clearly M is singular. Thus the pair (¢, A-+KB) is unobservable for each fixed k. This ‘means that there is no constant input that would permit us to distinguish all nonzero states, from 0. However, let us fix k, and ask: What ae the states that cannot be distinguished from 0? These are precisely the (nonzero) states that produce an identically zero output, ic., the states x such that Mx=0, An easy calculation shows that the states that cannot be dis- Linguished from 0 with u(t) =kare {ol 0 -kI’, 2#0). ‘Now comes the important point. Given any x#0, there existsa choice ofksuch that x#a[l 0 ky VaeR. Hence, with this particular choice of input, the states 0 and xan be distinguished. So we see thatthe system under study is (locally) observable. 66 Example Itis possible ro define a system to be globally observable if every pair of states (x, X1) With x #%; is distinguishable. However, this concept is much stronger than local abservability. Consider the system [¥1] [eosx] {v2} 7[sinx]° Given y, and y2, one can uniquely determine x (o within a multiple of 2x. Hence each xis distinguishable from all other nearby states, and the system is (locally) observable. How ‘ever, since xand.x +2x cannot be distinguished, the system ismor globally observable. Ml Now let us derive necessary and sufficient conditions for the system (1) ~ (3) to be locally observable. For this purpose, itis useful t0 recall how the standard observability rank test for linear systems is derived. Consider a linear system described by (2) and (4). Suppose we know u and can measure y; assume for the sake of convenience that w() is a smooth function of i. has derivatives of all order. Then successive differentiation of theSee.7.3 Reachability and Observability 417 output equation (4) gives 67 yn=Cx, FO=CH=CAX()+CB UID, HO =CA? x(0) + CAB UE) +CB AK), ++ Hence, by successively differentiating y, we can infer the quantities 68 Cx(t), CAx(1), CA?X(),* after subtracting the known quantities CBu(t), CAB u(?), CBui(), etc. Now (68) shows that if the matrix hhas rank n, then itis possible to determine x(1) uniquely. (Of course, there is no need to go ‘beyond A’~" because of the Cayley-Hamilton theorem.) For nonlinear systems te idea is pretty much the same. Let denote the numberof out- puts, and let, h(x) denote respectively the -th components of y andi). Then DW »)=hiw, TL 3) =dhy X=dhy fx) + Du, dh, B(x) =(Leh Xx) + Day gh OO, where the Lie derivatives Lyhy and Lh, are defined in accordance with (7.1.15), and the explicit dependence on is not displayed inthe interests of clarity. Differentiating one more time gives 72 5y= (LPO Lu Lg Leh 00+ Dit (Lg ho + Eu lel g hor E Sein Le bg hyn Expressions for higher derivatives of y, get progressively nastier, but the pattern is clear418 Differential Geometric Methods Ch.7 enough. The quantity y’ is a “linear” combination of terms of the form (La. *Layhy)00), where 1 Ss Sk, andeach of the vector fields 2), °°, 2, isfromthe set a In view of the foregoing observation, Theorem (73) below seems quite plausible. However ait fon is needed to prove the theorem. 73. Theorem (Sufficient Condition for Local Observabllity) Consider the system described by (1) and (3), and suppose xe X is given. Consider the forms 1B 81. Bub TA (Aba.L,., Le, 0) 82 ‘evaluated at Xo. Suppose there are n linearly independent row vectors inthis set. Then the system is locally observable around xo Remarks The proof of Theorem (73) is based on several preliminary lemmas and is given by and by. But firstitis shown that, when specialized to the linear system described by (2) and (4), the condition of Theorem (73) reduces to the familiar condition that the matrix 'W, of (69) have rank n, Now (2) ~ (4)is of the form (1) —(3) with 75 f(x) =Ax, B(0)=b,, h(XI= 6X, where b; denotes the i-th column of the matrix B and ¢, denotes the j-th row of the matrix C. Hence, withs =0in(74), wehave 16 (dh,Xx)= Next, 77 Cth) =GAx, (Lg) =6)b, Therefore, the only nonconstant functions are L:h,. Since da = 0 ifa is aconstant function, the only nonzero vectors of the form (74) with. 1B (AL phyx)=C)A,f=1, VL When we take repeated Lie derivatives as in (74), the constant functions do not contribute anything. In fact, the only nonzero vectors ofthe form (74) are 79 GA‘ k20, 1 Theorem (73) states that if ths set contains n linearly independent row vectors, then the sys- {emis (locally) observable. Finally, the Cayley-Hamilion theorem permits one to conclude that the span of the vectors in (79) is exactly the same asthe span of the setSee.7.3 Reachability and Observability 419 80 GAL K=O. nL jel sl So the sufficient condition for observability becomes, c | | &| st rank! | |: | [ea ee To prove Theorem (73), preliminary concept isintroduced. 82 Definition Given the system described by (1) and (3), the observation space O of the system isthe linear space of functions overthe field R. spanned by all functions ofthe form 83 Ly Ly hj. 820,21. °°° 6 (EBs Bgl LSI SL Its important 1 note that the observation space consists of all linear combinations of functions ofthe form (83) wth constant rea coefficients —not functions of x, Also, ifs=0 in (82), then the "zeroth-order" Lie derivative of h, is tobe interpreted ash, itself ‘The next lemma gives an alternative and useful interpretation of the observation space. 84 Lemma For the system described by (I) and (3), let J denote the linear space of func tions over the field spanned by all functions ofthe form BS Ly by hy 520, 15/5, where), ~°>, ¥,are vector fields of the form 6 vatt Sue {for some choice ofrealnumbers u,, °°, ge. Then J=0. Proof Note thatifv, ware vector fields, we have 87 Leswh=Lyhj thehy. Now note that (i) each vector field of the form (86) isa linear combination over R of the vec- tor fields {f, 81. °°". Bn )-and (i) conversely, each vector field inthe set {f, ). °~s Bn) i8@ linear combination over R of vector fields ofthe form (86). It is obvious that (i) is true. To see (ii), observe first that is ofthe form (86) —just set u;=0Vi. Next, we have420 Differential Geometric Methods ch 88 gaiftg)-fi Hence g, is also a linear combination of vector fields of te form (86). It follows that the span ofthe vector fields (f,@.. °°". 2m) With coefficients in Ris the same as the span of all ‘vector fields ofthe form (86) with coefficients in R. That O= J now follows from repeated applications of (87). @ Lemma (89) presents another technical result needed in the proof of Theorem (73). 89 Lemma Letu,, -->,u,eR", and consider the piecewise-constant input my, Oster, nersnen 90 wns: Let y/(X)=Yj(Xo. 115° fy) denote the j-th component of the system (1) ~(3) correspond {ing to the control input w(-) and the initial state Xo, Then aa 7 [2 3| newt where v, isthe vector field Ly hao), 92 y=f+ Saye, Remarks The lemma says that if we apply a piecewise-constant control of the form (90}and then let the duration ofthe "pulses" shrink to 0, then the quantity in (91) is equal toa particule repeated Lie derivative. Lemma (89) can be proved quite easily by induction on &; the proof is left as an exer- cise. Proof of Theorem (73) Let O denote the observation space of the system, and con- sider the et of row vectors doL(xy)2s &.varies over O. Thisisa subspace of (R")*, the set of xn row vectors. Moreover, this subspace is precisely the span ofthe various row vectors in (74), and hence has dimension n by hypothesis. Now let J be as in Lemma (84). Since J=O bby Lemma (84), the hypothesis implies that there exist linearly independent row vectors of. the form dB, (Xo), ***. dB,(X9), where each B, is a function of the form (85). Hence, by the inverse function theorem {Theorem (7.1.1)} it follows that the mapSec.7.3 Reachability and Observability, 21 Bix) 93 TO)= But) is locally a diffeomorphism around xy. Choose a neighborhood N of xp such that T:.N > T(N) is a diffeomorphism. Suppose x:€ Nis indistinguishable from xp; itis shown that x; =X9. In tum, this implies that every xe Nother than x) is distinguishable from Xp, i-., thatthe systemis observable at xp, which isthe desired conclusion. Accordingly, suppose xj€N is indistinguishable from xy. This means that YC. Xo.) =¥C, x1, w) forall inputs uC). In particular, let ube a piecewise-constant input of the form (90). Then y (x0) = (Xx). Letting all, -» 0 and applying (91)shows that L 94 (Ly, dL, hy%)= (AL, + ALy, hy). j This relationship holds forall vector fields v ofthe form (86), and for all integers s 20. Now ‘comes the main point, Each function 8; in (93) is of the form (85) for a suitable choice of vector fields v,, **+,¥,. Hence 9S Bilx%o)=Bilx1), Fe 25 ie., T(x) =T(x,)- But since Tis locally adiffeomorphism, this implies that xp =x). 96 Example Consider once again the system of Example (65). This system is ef the form (1)—(3) with {(9) = Ax, £(%)= Br, h(x) =ex, Hence, to apply Theorem (73),itis necessary to examine the row vectors dh (o), (AL) 049) (AL gh)(H9). (ALL gh) OR), (AL ghee) (x0) Routine calculations show that these vectors are independent of xo, and are respectively ‘equalto cA, eB, BA, cAB, ‘However,422 Differential Geometric Methods ch? e] [foro Al =lo01}, B} [100 whichhas rank 3, Hence the systemis observable. ‘Theorem (73) gives only a sufficient condition for observability. 1s it also necessary? ‘Theorem (97) below gives a decomposition result which shows that the condition of ‘Theorem (73) is “almost” necessary; see Corollary (112). 97 Theorem (Decomposition of Unobservable Systems) Consider the system described by (1) and (3), and let x9€X be given. Let © be as in Definition (82). For each XEX, let dO(X) denote the subspace of (R")* consisting of all row vectors @a(x), A€ O. Suppose there exists aneighborhood N of x such that 98 dimdO(x)=k ). VareO}. Define ker 40 to be the distribution which assigns the subspace ker dO(X) to each xe X, ‘Thus ker €O consists of all vector fields that are annihilated by each of the forms da. as ‘0. varies over the observation space O. Suppose xy¢X is given, and that dim dO(x) is con- stant for all x in some neighborhood N of X. To be specific, suppose dimdO(x)=k is used as a shorthand for an arbitrary vector belong. ing to the appropriate subspace of (R")*. Thus (122) means: (A,)r consists ofall vectors having that particular form; (123) should be interpreted similarly Equation (122) stats that the state components 2, and 2, are controllable, while (123) states that the state components 2 and 24 are unobservable. Problem 7.8 Consider the problem of controlling a satellite in space using gas jet actuators. The angular velocity of such a satellite is govemed by Io=ax(la) +t, where lis the 3x3 inertia matrix ofthe satellite ina body-attached coordinate frame, bis the angular velocity vector, and tis the vector of externally applied torque. Suppose the coordi- nate frame is chosen to correspond to the principal axes of the satellite. Then (cf. Example (53.19)1 ag (ay fy feds and the equations goveming the motion becomeSec.7.4 Feedback Linearization: Single-Input Case 427 een tus, where and x, y, zare respectively short-forms of ,, @,, ©, (a) Show that the system is not reachable if only one actuator is used. Show that, for ‘example, ifonly u, is used, then the quantity 6, =cy? +z? is constant, whatever be 1 () Derive similarresults forthe case where only u> isused, and where only wis used, ” (b) Suppose 1, >1, >4,>0. Show that the system is reachable if any two out of the three actuators ae used. (©) Suppose J, =/ > 1.,So that =0, Show that the system is reachable if, and are used, and if wp and ware used. Show that the system is not reachable if only wand w are sed. (@) Suppose the satellite is prefectly spherical, sothat ,=, = /,. Show that the system isnot reachable unless all hree actuators are used. ‘Note: In the above simplified version of the problem, we are considering the reachabil ity of the angular velocity vector alone. For a more thorough treatment which considers the reachability of both the angular position as well as the angular vetocity, see Crouch (1984). Problem 7.9 A simplified model for steering an automobile in a plane is given by [see Nijmeijer and van der Schaft (1990), p. 52} x [sine 0 5] =] ¢080| wi +]0| u2, 6} Lo J where (x, y) is he position of the centroid of the automobile, and @ is its orientation, Show that the systemis reachable, 7.4 FEEDBACK LINEARIZATION: SINGLE-INPUT CASE In this section and the next, we study an important application of differential-geometric methods, namely the possibility of transforming a given nonlinear system to afinear system, via feedback control and a transformation of the state vector. In this section the case of, single-input systems is considered, while multi-input systems are studied inthe next section,428 Differential Geometric Methods Ch.7 ‘The problem studied in this section can be described as follows: Consider a system described by 1 x=f) +4 900), where Fand g are smooth vector fields on some open set X CR" containing 0, and f(0)=0. It is desired to know when there exist smooth functions g, se 5(X) with s(x)#0 for all x in and a local diffeomorphism T on R® with T(0)=0, such 2 veg(xy+sQx)u, 3 a=TO, then the resulting variables z and v satisfy a linear different equation of the form 4 i=Az+by, where the pair(A, b)is controllable. If this isthe case, then the system (1) is said tobe feed- back linearizable. Note that since s(x) #0 in some neighborhood of @, (2) can be rewritten ao, 1 : s(x) * 50)" where —q(x)/s(x) and 1/s(x) are also smooth functions. Hence, if we chink of vas the exter- nal input applied tothe system, then (2) lor equivalently (5)] represents nonlinear state feed- back, and a nonlinear state-dependent pre-filter, applied to the system (1). Similarly, (3) represents a nonlinear state-variable transformation. Hence the overall effect of (2) and (3) canbe depicted as shown in Figure 7.2. [is PE eran {=e Fig. 7.2 ‘The problem statement can be simplified further. Suppose suitable functions q. s, T ccan be found such that the resulting state vector z and input v satisfy (4), and the pair (A. b) is, controllable. Then itis possible to apply a further state transformation 6 zeM"s such that the resulting system isin controllable canonical form [see Chen (1984)]. ThusSee.7.4 Feedback Linearization: Single-Input Case 429 M'AMZz+M™'by, o 10 0 0 oon 0 0 8 MTAM= : o 00 0 way ~a; ~a 1 and the a,'s re the coefficients ofthe characteristic polynomial 9 IsI-a ‘Nowa further state feedback of the form 10 Flag ay oo aylt results inthe closed-loop system u Aas bi, where 010-0) fo 001-0) o 2 000-1 0 000-0 1 Iniseasy tosee that 13 T(x), ab = q(x)—a MT (x) +518), where430 Differential Geometric Methods ch.7 WM a’=lag ay a) ‘Note thatthe transformation (13) is of the same form as (2) ~ (3). Hence if the system (1) is, feedback linearizable at al, then it can be transformed to the simpler system (11). With these observations, the problem under study in this section can be precisely stated as follows: Feedback Linearization Problem (Single-Input Case) Given the system (1), do there exist(i) a smooth function qe S(X), (li) a smooth function se S(X) such that s(x) #0 {forall xin some neighborhood of O, and (ii) local diffeomorphism T:R® —R° such that T(0)=0, satisfying the following conditions: Ifnew variables v and z are defined in accor dance with (2) and 3) respectively then 1s Note that (15)isjust (11) written differently ‘Now the main result of tis section is presented. 16 ‘Theorem The feedback linearization problem for the single-input case has a solution ifand only ifthe following two conditions are satisfied in some neighborhood ofthe origin. ()_Thesetof vector fields {adyg, 0iSn-I is linearly independent. Gi) The setofvector fields (adyg, 05iSn-2) isinvolutve. Proof "Only if" Suppose the problem has a solution; ie., suppose there exist a neigh- bothood U of the origin, a diffeomorphism 7: U—+R", and suitable smooth functions 4. 5€5(X). Itis tobe shown that (i) and Gi) above are satisfied. By assumption, the transformed variables z and v satisfy (15). Let 7; denote the i-th ‘component of the vector-valued function T(x). Thus 7 g=T(x) Now 18 2,=VI, k= +u by (1) 720%) by (15). Hence (18) implies that 19 T= =L4T),and0= AT), g>=L,71. Repeating the reasoning for z2, °°", 2,-1 shows thatSec.7.4 Feedback Linearization: Single-InputCase 431 20 LpT)=Tuay,LgT=0, fori = 1, +n Finally, fori =n we have 2 i, =LyT, +u LT, by (1) q + us by (15) and (2). Hence 22 LiTy=G.LgTn Allof these equations hold for all xe U. Next, itis claimed that 2B’ Lyyg?,=O0for0Si = .Osisn-2, 33 Hits First itis established that Condition (iis true, i, thatthe set (adg, OS/ ),OSi.s=, 40 +50, AL 2=[T,(0) =~ T,00Y- Letus see what equations these new variables satisfy. For this purpose, let us first show that M2 SAT), ad}g> =0, for0Si =-)'s, whichis the same as (29), ‘Using (42) and (43), itis a simple matter to find a set of equations for the new variables zand v. First, for 1 4u =.= Finally, 45 i,=VI,i= SAT, (> +u SAT, g> = +us =1. ‘These are precisely the equations (15). Hence, to complete the proof, it only remains to show that s(x) #0 for all xe U, and that isa diffeomorphism on U. To prove the first statement, suppose there exists a point xy¢ U such that s(9) =0. Then, combining (37) with (43) evaluated with j= | gives 4 (o) for0(0)=a, 5(0) But since s(0)#0, this implies that a, =0. Let us therefore drop the term a,d7, from the ‘summation in (47), and take the inner product of the resulting sum with the column vector ‘adyg( 0). Using (42) with i= 1 and (43) with j=n—I gives 49 O= 44.4 (0)=— ap 5(0), which in turn implies thata,_; =0. The process car be continued to show that a, =0 forall i. Hence the differentials d7;(0), i= 1, ~*~, are linearly independent. By the inverse func~ tion, this shows that Tis a local diffeomorpism on some neighborhood of. iSee. 7.4 Feedback Linearization: Single-Input Case 435 Application: Robot with Flexible Joint Consider the problem of positioning a link using a motor, where the coupling between the motor and the link has some flexibility. The development below is taken from Marino ‘and Spong (1986). The system under study can be modeled as shown in Figure 7.3, where the motor shaft is coupled to the link by a linear spring. This system can be modeled by the two equations a Fig.7.3 Wj +mal sing +k (qi —42 Hija Kg, ~42)=% ‘where the various physical constants are defined as follows: J Momento inertia ofthe motor about ts axis of rotation. 1 Momento inertia of the link about the axis of rotation ofthe motor shaft {Distance from the motor shaftto the center of mass of the link. m— Mass ofthe link. 8 Acceleration due to gravity Torsional spring constant. 4) Angle ofthe link 42 Angle ofthe motor shaft. 4 Torque applied tothe motor shaft If we choose the natural setof state variables436 Differential Geometric Methods ch7 mel ne, — Kee, -. a sin, — F129) 6 x 0 w Fea) | From Theorem (16). this system can be transformed tothe form (15) if and only ifthe following two conditions hold over some neighborhood of 0: (i) The set (8, adyg, ad?g, ad?g) is linearly independent, and (ii) the set (g, adpg, ad) is involutive. Inthe present case all ofthese vector fields are constant, and foo 0 wm 00 KM 0 fe adye adje adel=| 5 yyy 9 42 The determinant of 2/72J* which is obviously nonzero. Hence Condition (i) holds. As for Condition (i), any set of constant vector fields is involutive, since the Lie bracket of two constant vector fields is zero. Hence, by Theorem (16), this system is feed- back linearizable. To construct a linearizing transformation, one can proceed, as in the proof of Theorem (16),to find aonconstant function 7; such that T (0)=Oand =0, =0, =0. | the present instance, since each of the three vector fields is constant, the conditions reduce simply to am, aM aT Ox," Oxy” Oxy Soa logical (and simple) choice isSec.7.4 Feedback Linearization: Single-Input Case 437 Tio)=xy ‘Of course this choice is not unique. Now, from (38), 222 242 SATs > =— EE cos, — Erg —49), This gives the nonlinear state transformation. To obtain the feedback control law (2), we uuse(39). This gives = 77 Inthe new variables, the system equations are ‘Two comments are worth making at this point. First, since z) =x, and z»,23.Z4 are its higher derivatives, the new state variables are physically meaningful: They are respectively the link angle, angular velocity, angular acceleration, and jerk. Second, though Theorem (16) only considers focal feedback linearization, inthe present example the linearization is actually global. This can be seen by noting that the transformation 7 mapping x into 2 is actually globally inverible, withthe inverse mapping xetn metry 1a, mal apne teat sinzy), 1, mgl xareat [eat Ezz c082))438 Differential Geometric Methods h.7 7.8 FEEDBACK LINEARIZATION: MULTI-INPUT CASE In this section, the results of the preceding section, and in particular Theorem (7.4.16), are extended to multi-input systems of the form (x) + Eu, BO). 1 where, gj, "", 5 ate smooth vector fields on some neighborhood X of, and f(0)=0. Itis reasonable to assume that the vector fields (By, °°". fe} are linearly independent in some neighborhood of the origin. Otherwise, some of the inputs are redundant, and by redefining the inputs and reducing their number, the linear independence assumption can be satisfied. The objective is to determine whether itis possible to transform the system (1) to a linear system via nonlinear feedback and a state transformation, ‘The major difference between the single-input case and the mult-input case is that, while there isa single canonical form to which all controllable single-input linear systems can be transformed, thee is more than one canonical form for multi-input systems. This necessitates the introduction of some extra concepts, namely controllability indices and the Brunovsky canonical form. Consider the linear time-invariant system 2 x=Ax+Bu, where Ac R™, Be R™", and Bhas rank m. If the system is controllable, then 3° rank(B AB A°B---A""'B]=n 1-and for i> 1 define Define rp =rank B= rank |B AB--+A'B]~rank [B AB-- A’'B} ‘Then clearly 07, $m for all i Itcan also be shown that r;2;.1. The proof is quite easy, and in any case the statement follows from the more general version for nonlinear systems given below; see Lemma (16). Thus So marg2ry 2-2 EN a Now define the Kronecker indices x, *--. x, of the system (2)as follows: kis the number of the integers r thatare greater than or equal to, Clearly 6 Ky 2KpBee Bk ZO, and P= For convenience, introduce the integersSec. 7.5 Feedback Linearization: Multi-Input Case 439 7 a= xnieh~ Now the Brunoysky canonical form of the system (2) is another linear time-invariant sys- temof the form 8 2+ iv, where A and Bhave the following special structures: 9 Block Diag(A,, "+. Aw), where A, has dimensions x;xx, and is the companion matrix corresponding to the polyno- mial s™., Inother words, O10--0 601-0 10 A= eR. 0001 000-0 ‘The j-th column of the matrix B has all zeros except for @ “I” in the G,-th row. It can be shown that the system (2) can be put into its Branovsky canonical form by state feedback and a state-space transformation. It is easy to sce that if the number of inputs m equals 1, then 7; = 1 for all 4 x, =m, and the Brunovsky canonical form js just (7.4.12). For multi-input systems, however, several canonical forms are possible. But if two multi-input systems have the same number of states and inputs, then they are "feedback equivalent,” in the sense that one system can be ‘transformed into the other via state feedback anda state-space transformation, if and only if they have the same Brunovsky canonical form. To define a Brunovsky canonical form for nonlinear systems of the form (1), we proceed in a manner entirely analogous tothe above, except for afew additional regularity assumptions which are not needed in the linear case. Given the system (1), define 1 C.=(adtgy, 1S) Sm, 0SKSi), 0Si 1 is routine to verify that the above definition reduces to (4) for linear systems of the form (2). But there isa potential complication in the case of nonlinear systems. In the linear case, each distribution A, is generated by a set of constant vector fields, and thus has the same dimension at all x. But iz the nonlinear case, dim ,(x) can vary as x varies. One way to forestall this difficulty is to assume that 0 is a regular point of each of the distributions 4,,1%0, «"*,n=1. If such an assumption is made, then dim A,(x) is constant for all x in some neighborhood of O, and as a result the integer r, is well-defined by (15) forall i. Once the integers r, are defined, the Kronecker indices x, «--, Ky and the Brunovsky canonical formare defined exactly asin the linear case. 16 Lemma We have AT OS rig Sry Vi, Proof Clearly 7.4; 20 since A, isa subset of 4,,. To show that r,.; $7), note that, by definition, 18 dima,=dima,_. +1, or, inother words, 19 dim span (C;_,tadjg,, «+, adig,,) }=dimspan C;_, +r, ‘Thus there are exactly m-~; vector fields among {adfg,, °°. ad} that are linear combi- nations of those in C;-. and the remaining r, vector fields among (adjg,, °-', ad/q)- For notational convenience only, suppose these ae the lastm vector elds, suppose 20 adjgyespanC,, adie), ~~ adie, J=n +1, ° Now the linear dependence is maintained when we take the Lie bracket of the two sides with fethus 21 adf"'gespan Gl adf'g,, =. ad", Hence 22 dimspanC,,y Sdimspan C, +r, which isthe sameas saying thatr,., <7).See.1.5 Feedback Linearization: Multi-InputCase 441 ‘To state the necessary and sufficient conditions for feedback linearizability inthe rmulti-input case, we introduce one last set of integers. Let 8 denote the largest value of i such that, #0, Thus r > Obut,=0 Vi > 8. Now define anne acca By Lemma (16), each of the m,'sis nonnegative. Note that, for a mult-input system, both 7; and m, could be zero for sufficiently large i. However, fora single input system, =n ~ 1, 1m, = L,andm,=0 for 0isn-2. Now it follows readily from (23) that a w= Lm, Since, from (15), 28 dimay=Srp, wwe get from (24) and (25) that a 5 Ly-Ln= =z ot jo" ste 26 dims dima, ate teeatt | =E Ym= Dk-dm ia +2m, += +(B- ims, after interchanging the order of summation inthe last step. In order to follow the proof below, one should be able to interpret and understand the integers m, and x, ina varity of ways. The Kronecker indices x, ~~, Kare justshe sizes ofthe various blocks in (9); there is no loss of generality in assuming thatthe blocks are arranged in nonincreasing size, which is implied by the fist part of (6). The integer 8 is ‘equal to the size of the largest block minus |, i.e., 8= x, —1, Now the integer m; is just the numberof blocks of sizei+ I. Hence ‘ & 27 Sam. E+ 1m, ans Now we can state precisely the problem under study. Feedback Linearization Problem (Multi-Input Case) Given she system (1) together with a set of integers x, “**. Xm Satisfying (6), do there exist (i)a neighborhood U of, (ii)a smooth function q: U—> R", (it) a smooth function S: U > R™" such that det S(0) #0, and442 Differential Geometric Methods ch (iv) a local diffeomorphism T:U—R° such that T(0)=0, satisfying the following condi tions: fnew variables and v are defined according to 2% 2=71), 29 v=q(x)+S¢xu, where 30 ual ual then the new variables 2 and v satisfy the linear differential equation 31 t= Az4By, with A and B in Brunovsky canonical form corresponding to the indices X\, °-*, Ky? Now we state the main result of this section, 32 Theorem Consider the system (1), and assume that the following are true: (a) The vector fields), °°. fq are linearly independent at@, so that dim dp =ro=mas in(15); and (b)0 isa regular point ofthe distribution 4, for each i20. Under these conditions, the feed: back linearization problem in the multxinput case has a soluion if and only ifthe following ‘wo conditions are satisfied: (@) dimas=n, and Gi) The distribution A, is involutive whenever m, #0. Remarks Comparing Theorem (32) with Theorem (7.4.16) for che single-input case, fone can spot an extra hypothesis inthe present instance, namely the assumption that 0 is a regular point of each of the distributions ,. But in the single-input case, this regularity assumption is a consequence of Condition (). To see this, note that in the single-input case, the distribution 4, isthe span of exactly n— 1 vector fields, namely g, adgg, “-", adf"'g, Hence, ifdim ,_ =, which is Condition (i) in the single-input case, then this set of vector fields is linearly independent. This implies that A, is the span of the i+ vector fields 1, adyg, °*:. adig, which in tur implies that @is automatically a regular point of each distri- bution 4, if Condition i) holds. Thus, in the singlesinput case, Theorem (32) reduces pre- cisely to Theorem (7.4.16). It should be noted however that Isidori (1989), Theorem 2.2, p. 250, states that even the regularity of 0 can be incorporated as part of the necessary and sufficient conditions for feedback linearizability Proof “If Suppose Conditions (i) and (ii) hold. Let 5 be, as before, the largest integer ‘such that m, #0. By (23) this implies that r= O forall j > 8, and hence that dim As =n. Now 33. dimAg, =dimAg—rs =m that is, the codimension of Ag. is mg. Now Condition (ii) states that Ag. is involutive,Sec.1.5 Feedback Linearization: Multi-InputCase 443 Hence, by the Frobenius theorem, there exist a neighborhood U of ® and smooth functions (is), LSismg), such that their differentials ae linearly independent atx =0, and 34 (x)=0, Vxe Up, 05/S8- I, 1) has sank mie. full row rank, atall xe Up, To see this, assume that there isa point x9€ Ug and constants c, “°°, Cn, such that 36 | Se; | (X9)=0, for j=1, +. ‘This means, together with (34), that the row vector 37 r= Devdhg(%o) annihilates each of the column vectors (adj, )(xq) for 0. By (26), 38 dims. =dimA, —2m5—ms., 2mg—ms Now there are two cases to consider, namely mg; =0 and mg. #0. Suppose first that mgi=0. Then it is claimed that the differentials of the 2mg functions (hg, Leha.. 1Si'Sms) are linearly independent and annihilate each of the vector fields in ‘gaa. Note that if mg. =0, then by (38) the codimension of dg.» is exactly 2mg, so what is being claimed is that if m=0, then the distribution Ag.2 is automatically involutive. Since As-2 =L¢ ~ if0<1<8-2, since both terms on the right side of (39) are zero by (34). Now itis shown that the 2mg dif- ferentials (dita. ALpho,, 1 +c, SAL thay, ad}B) > (xq) =O for0(%9)=0, forl $jsm ‘Again by Lemma(7.1.59), 43 =Lr- =~ , since the fist termi zero by (34). Hence (42) reduces to M4 ~Len 7m] Mg = 0, where Mg is defined in (35). Since ithas already been established that Mg has full row rank, (44)impliesthatc,,= Ofor 1 $i:(x) tall xe U ,. Without loss of generality, itean be assumed that U, G Uo, which is the neighborhood over which the distribution Ag. is invo- lutive. Now, of the 2ms+ms.1 exact differential forms which annihilate As.2, we have already found 2ms of them, namely the set (dig. dLhg,}- So let us select another ms. smooth functions (4g.1i,1SiSms.;) such that the set {dhg,. ALehgi. 1(x), 1SiSmg., 1SjSm, and consider the (mg +s; xm matrix Ms } # |a.,| Itisclaimed that this matrix has full row rank at all xe U; . To show this, suppose there exist ‘point xye U and a Ix(mg +mg-:) row vector that annihilates the above matrix at Xp. In other words, suppose there exist constants ¢q 1SiSmgy¢ i IS7Smg.1 Such that 48 Sco Sdhg, ad?g)>(%o)+ Dey (x9) = fork $j. ‘Again by Lemma (7.1.59), itfollows that 49 =L¢— (%9)=0, forj= 1... Fo a Letus define the vector i Lea dlehs,+ Dev dhss,;| Mo). ‘We know that this vector v already annihilates 4s.» at Xp. Now (50) implies that v also annihilates ad" g,(xo) for 0. (This cannot happen in the single-input case.) Then there are just n ~ mg functions in the set (55). In this case we must choose mg other functions, call them {fo,. 1 Simo], such that the differentials of then functions in the set SB (Lbs, 1SiSm, OS1Sk, OSK $3) are linearly independent over some neighborhood ofthe origin. ‘The selection of these addi- tional mo functions is not difficult. Even if mg >0, the differentials of the n —mo functions in (55) are linearly independent. So one can just choose some constant row vectors i, ***, Gq, to complete a row basis, and just let fo, (x) =x. In any case, one now has the ‘functions of (58). ‘The linear independence property means that the map taking the vector X into the n-vector whose components are the » functions in (58) isa local diffeomorphism. ‘Next, inanalogy with (47), form the matrixSee. 7.5 Feedback Linearization: Multi-InputCase “7 [Me |Ms.1 le 39 M= M, Mo where 60 (My),(x)= 00, 1SiSm4, 1S jm. ‘The numberof rowsof ths mati is 3 Emam In other words, the matrix M is square, At each stage, it has been ensured that the matrix continues to have full row rank even as one more block is added atthe bottom. So the final matrix M is nonsingular forall xin some neighborhood of 0. We are now in the home stretch, Let us first make a couple of observations. 1) Amongst the integers m, no more than m can be positive; this is clear from (54). 2) Suppose we define «=i +1 for those values of i for which m,>0; then these are precisely the Kronecker indices. Also, m, > 0 only when k =; for some iso that only the corresponding functions hy, appear inthe list (58). Now define 62 =h51. 7m hme Hess >0.¢nue Img >0, 6,01 0.16 2p = HOme Note that, in computing the subscripts of the @'s, we have made use ofthe relationship (24). There are exactly m of these functions 6,. For each index i, note that x; i the number of times the function 4, is Lie-differemtiated in thelist (58). Define amap2,:U—>R™ by448 Differential Geometric Methods ch.7 4:00) Lei) 6B 200= ‘ LE 600) and define a map T:U —> R" by 2400) 64 T= Zp). Then Tis a local diffeomorphism, as discussed earlier. If each @, is chosen such that $:(0)=0, which is easy to achieve since a suitable constant can be added to each hy, then T(O)=0. We shall see very shortly that this is the desired state transformation. ‘What are the differential equations governing these new variables T;(x)? By definition, 65 Ae, = = <9,00, (09 + Fu; 01> since =O forall, j. Similarly 4 6 ue All this should look familiar, because this isthe same set of manipulations as inthe single- inputease. Now ests forl $18, 1See.7.5 Feedback Linearization: Multi-Input Case 449 x =H G+ Buy by lh", Althis stage, observe that, by using Lemma (7.1.59) repeatedly, we can write 68 Ly LEG, = =(-1)" My where the matrix M is defined in (59). Thus, collecting the equations (67) as varies from | tomgives fen] «| o q(x) +Su, de where gi(X)= L705); =(-1)""'M,), and Mis the matrix of (59). Letus define 70 v=q@)+Seu. ‘Then (69) shows that i= % Finally, (66) and (71) show that, in terms of the new variables 2 and v, the system is in Brunovsky canonical form. "Only if” Suppose the feedback linearization problem has asolution. It must be shown that Conditions (i) and (i) of the theorem are satisfied. Itisnotationally convenient o renumber the m inputs as follows: To begin with, the m vector fields gi, °°", 8m are linearly independent. Now consider the set spanning A;, namely Th Cy= bie Bw Me Bude 5 EB Suppose mo20. This implies [see (23)] that exactly mo vector fields among (Uf. gj, °°. [f el) are linearly dependent on the rest and on gj. °°", Bq. Renumber the inputs so that these linearly dependent vector fields are the last m. vector fields; therefore, 73 UE, wilespan (Bh. °°. Be LE Bil °° [Ey mls ford = smote sm Note thatthe above linear dependence is preserved when we take higher order Lie brackets aswell. Thus450 Differential Geometric Methods ch7 14 adgcespan (8) (fg). 1Si =0, for0=(-1)" sy, Si fm, where 5 is the i-th element ofthe matrix S of (29).See. 7.5 Feedback Linearization: Multi-Input Case 451 By assumption, the mxm matrix $ of (29) is nonsingular for all x sufficiently near 0. Foreachk= 1, --*, 8,define 8 81 m= dm, Now itis shown that, for each k=8, °--, 1, the nj>xry submatrix of S consisting of the first ‘ny rows and the first n, columns is nonsingular. The significance of this claim is as follows: Fix k, and consider the 1xm submatrix S; of S consisting of the first 14 rows of S. Since S itself is nonsingular and hence of full row rank, this submatrix S, has rank 7), and thus con- Tk linearly independent columns. The point of the claim is that actually the first M, columns of, are linearly independent. To prove the claim, suppose first that k =8, so that 1 ="1g =m. Note that 3+ 1 is the ize ofthe largest block in the Brunovsky canonical form, and that m isthe number of such blocks. Thus Kj =x, 5+ 1. Now consider the m5xm matrix Ss consisting of the first ms rows ofS, Letting in (80) range over 1, -*", ms gives 82 But (76) states that, if j > mg, the vector field ad, isa linear combinat from S(X)] of adfgr. --, adfg,,, and some other vector fields adig, where /< 8. Now using (79) shows that, if j > mg, the function is a linear combination of Sii."~"»Sing. Repeating this argument as varies over 1, "mg shows that the lat —mg columns ofS are linear combinations of the frst ms columns. Since Ss has rank mg, it must be thatthe first ms columns ofS are linearly independent. Next, letk=8- 1, so that ns.1 =mg+mg.y. Ifms.; =0, then ng. =mg andno further argument is needed, s0 suppose ms. #0. Then kygsi = °°" =Kiy emp, =: In this ease, in addition to (82), we have the relationships [cf. (7.4.29)] et 83 sin LSES m5, 1SjSm, 84 =(-D*"s, isms + SES mg +51, 1S) Sm Now consider she (ms-+ms-,)xm submatrix $3. In this case (76) shows that if j>mg+ms.,, then ad? 'g, isa linear combination of adf'g, for k= 1, «++, mg+mg.1,as well as of adjg, for 1<5-1. Now by (79) and (7.4.23), each of the forms dL.16, for 1m-Lm=n,, ‘where tis defined in (81) and pis taken as 0. Thus (85) can be rewritten as aa 87 wO)=E Ley Cady, y(0)=0. Firstapply the forms d9), -*-.d6,, tow. Using (79) and (82), this gives 88 0= (0)= $54 (0) cap for Si.ms ‘Asa result, the expression for w canbe simplified to by 1 wa Z Lejadiay: fia Now apply the ms forms dZq),j/=1,"*,mg. and if ms.,#0 the ms, forms dy, j=mg +1, --*,mg+ms_, tog. This gives the equation,See.7.5 Feedback Linearization: Multi-Input Case 453 0 su Shamsemes can 9 0} O} | Smgsmasnt 77 Sma tmesens times CBtms +m Again, since the coefficient matrix in (92) is known to be nonsingular, it follows that 5-1 =0 forall j. The process can be repeated and shows that all cj are zero. This shows that the set of vector fields (adig,, 0 =0, for0<1<8~2, 1+u . Suppose (0)#0. Then continuity implies that (x)#0 for all x sulfciently close 100, Soif we define 7 qx)=,s00)=, 8 veg@tscyu, then (6) reduces to 9 Suppose (0)=0. Then the above scheme does not work, But suppose458 Differential Geometric Methods 7 <4h, g>(2)=0 for ll x in some neighborhood of 0. [This is not always true, even if (0)=01 Then 6) becomes 10 j= =L¢ Hence ML $= tu =Lih-+uLglgh, Where L? isa shorthand for Ly. Again, if(-glh)(0) #0, then we an define 12 q=Lihs=Lglth, and vby 8). Then (11) implies that 13 In general, suppose there is aneighborhood U of @ and an integer r 22 suchthat 14 (Ly LihNG xe U, fork =0, =. r-2, 1S (Lglf1h)(0)0. ‘Then we can define and v by (8). This results in cu In such a case the system (5) is said to have relative degree r. This terminology is quite consistent with the linear case. Consider a single-input, single-output system of the form 18 k=Ax+bu,y =x, Then the relative degree ofthis systems the smallest integer r such that 19 cA*b=0fork=0, “+, 7-2, andeA/'b 20. Fora linear system ofthe form (18), we know thatthe relative degree ris ess than or ‘equal tothe system order n. This Follows readily from (19). If 20 cAtb=Ofork=0, --,n—1, then the Cayley-Hamilton theorem implies thatSee.7.6 Input-Output Linearization 459 2 cAYy=0VK20, or in other words, the transfer matrix is identically zero, which contradicts the fact that ¢€A’"'b#0. But for nonlinear systems there is no analog of the Cayley-Hamilton theorem. Soone is forced to adopt a different strategy. 22 Lemma Consider the system (5), and suppose there exists an integer r satisfying (14) = (15). Thenr sn. This follows immediately from another result, which isaseful in its own right, 23° Lemma Consider the system (5), and suppose (14) ~ (15) hold for some integer Then the row vectors (dh, AL¢h, *-- ,ALf"h} are linearly independent in some neighbor- hood of . Proof Choose real constants, “+, ch, such that 24 Ya, ain=0. 3 Itis shown that a, =0 for alli. For this purpose, it isclaimed that 1 _[% ifitksr-l, 28 basalt) ay tgp 'h, itiek ‘The proof of (25) is by induction on ifor fixed k. If pose (25) is true for 0, “++, i= 1. Note that adi wehave ), then (25) reduces 10 (14). Now sup- adj-"g]. Hence, by Lemma (7.1.59), 26 Lysig hth = LL sap) lth Lage igh lth ight A, 0 of course is since the first term is zero by the inductive assumption. Now, if /+k(0)=(L 4210) However, BM Lqa= Soy Lylth= 0, syle 'h, a ‘where the last step follows from (14). Combining (30) and (31) shows that 32 O=0,-4 (Lyle *AV0) Now (15) implies that 0, =0, ie. that BB a= YoiLih Next, we have BM (0)= (Lag 10)=0, since da (0)=1 by assumption. Substituting forafrom (33) and using (25) leads to 3s 2 (Ug LF AYO), In tum, this combined with (15) shows that cy.» =0. The argument can be repeated all the way downto show that a, =0Vi. i Proof of Lemma (22) Since the set of row vectors (AL{h, 0SiSr-1) is linearly independent, itis obvious that Sn, Ml Inthe case of multi-input, multi-output systems, the idea is essentially the same, except that the relative degree is now a vector and not a scalar. The system (1) ~ (2) is said to have the relative degree vector r=[r, *** ral’ if (i) there exists a neighborhood U of 0 such that 36 (Lg LthXx) 2, ‘and (i) the mxm matrix S defined by 37 sy=Ly Lith, isnonsinguiaratx=0. Inthiscase, if we define the smooth m-vector valued function q by 38 Gi=Lth, define the smooth mxm matrix $ as in (36), and the new control vector v by (3), then the resulting input-output relationship is ofthe form (4)Sec.7.6 Input-Output Linearization 461 ‘The next result is the MIMO analog of Lemma (22). 39 Lemma Suppose the system (1) ~ (2) has the relative degree vector F=[r""- ral Then 4 Syn As in the scalar ease, the proof of this lemma is based on another result which is of independent interest. 41 Lemma Consider the system (1) ~ (2), and suppose it has the relative degree vector ra[rys rl’. Then the r= Sr, row vectors at 2 [AdL}h,, OSkSr,~1, 1Sism) are linearly independent insome neighborhood of, Proof Selectreal constants he, 0Sk Sr;~ 1,1 Sim, such that 43D Lay (althyXe)=0. aa Define 4 a=¥ Laulthesix. ma ‘Then, using (36), it follows that 45 Lyon Zeus tbe where 5, isdefined in (37). Now (43) shows that 46 0= (0)=(L4,a)0), for 1 R"™ are smooth functions, and in addition, q(0) =0 and S(0) is nonsingular. Two distinct cases are considered, namely: (i) the system (1) ~ (2) is input- state linearizable inthe sense of Section 7.5, and (i) the system (1) ~ (2) is input-output linearizable inthe sense of Section 7.6 Accordingly, suppose first thatthe system (1) ~ (2) satisfies the conditions of Theorem (7.5.32). Choose the feedback functions y and$ in (4) as well asa local diffeomorphism T around 0, such that the transformed state vector2=7'(x)isrelated to vby‘Stabilization of Linearizable Systems 465, S z=Az+By, where (A,B) isin Brunovsky canonical form. Next note ftom te line below (7.5.69) that 4;=L4 6 for some integer x; and smooth function 6, foreach i, Since 0) =0, it follows that q(0)=0. Also, it can be assumed without loss of generality that T(0)=0. Now the eigenvalues ofthe matrix A ae all tthe origin; hence the system (5) sno asymptotically stable. However, since the pui(A, B) is controllable, its possible to choose an mXn matrix K such that the matrix A—BK is Hurwitz, i.e., such that all eigenvalues of A-BK have negative real pars, Hence, if we apply the futher feedback control 6 Kz, then 20 (or equivalently, x =0) isan exponentially stable equilibrium of the esulting sys- tem. The scheme can be depicted as in Figure 7.5. For obvious reasons, the control law (4) {sneferred toas the inner loop control, while (6)is referred toas the outer-loop control vero Fig.75 7 Example A well-studied applicationof the above strategy is from the world of robot- ies. Consider a rigid m-link robot operating ma gravity-free environment, Thus the robot, ccan be operating in a "horizontal" plane, perpendicular to gravity, or else it can be operating inouter space. In either case, itis known [See Spong and Vidyasagar (1989), Chapters 6and 7) that the dynamics of such arobot are described by the Euler-Lagrange equations MQ y+hy.y) ‘where ys the m-vector of link angles, and wis the m-vector of the torques applied atthe vari- ‘ous joints, The matrix M, known as the inertia matrix, is symmetric and positive definite foreach y. The vector hincorporates all the Coriolisand centripetal terms. Inthe absence of ‘gravity, hsatisties (0, 0)=0.