This action might not be possible to undo. Are you sure you want to continue?
ROBUST AND OPTIMAL CONTROL
KEMIN ZHOU with JOHN C. DOYLE and KEITH GLOVER
PRENTICE HALL, Englewood Cli s, New Jersey 07632
TO OUR PARENTS
Contents
Preface Notation and Symbols List of Acronyms 1 Introduction
1.1 1.2 1.3 Historical Perspective : : : : : : : : : : : : : : : : : : : : : : : : : : How to Use This Book : : : : : : : : : : : : : : : : : : : : : : : : : Highlights of The Book : : : : : : : : : : : : : : : : : : : : : : : : :
xiii xvi xx 1
1 4 6
2 Linear Algebra
2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 2.10 2.11 2.12 3.1
Linear Subspaces : : : : : : : : : : : : : Eigenvalues and Eigenvectors : : : : : : Matrix Inversion Formulas : : : : : : : : Matrix Calculus : : : : : : : : : : : : : : Kronecker Product and Kronecker Sum : Invariant Subspaces : : : : : : : : : : : : Vector Norms and Matrix Norms : : : : Singular Value Decomposition : : : : : : Generalized Inverses : : : : : : : : : : : Semide nite Matrices : : : : : : : : : : : Matrix Dilation Problems* : : : : : : : : Notes and References : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
17
17 20 22 24 25 26 28 32 36 36 38 44 45
3 Linear Dynamical Systems
Descriptions of Linear Dynamical Systems : : : : : : : : : : : : : : vii
45
viii 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 3.10 3.11 3.12 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 5.1 5.2 5.3 5.4 5.5 5.6 5.7 5.8 6.1 6.2 6.3 6.4 6.5 Controllability and Observability : : : : : : : : Kalman Canonical Decomposition : : : : : : : : Pole Placement and Canonical Forms : : : : : : Observers and ObserverBased Controllers : : : Operations on Systems : : : : : : : : : : : : : : State Space Realizations for Transfer Matrices : Lyapunov Equations : : : : : : : : : : : : : : : Balanced Realizations : : : : : : : : : : : : : : Hidden Modes and PoleZero Cancelation : : : Multivariable System Poles and Zeros : : : : : : Notes and References : : : : : : : : : : : : : : : Normed Spaces : : : : : : : : : Hilbert Spaces : : : : : : : : : : Hardy Spaces H2 and H1 : : : Power and Spectral Signals : : Induced System Gains : : : : : Computing L2 and H2 Norms : Computing L1 and H1 Norms Notes and References : : : : : :
CONTENTS
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
47 53 58 63 65 68 71 72 78 80 89
4 Performance Speci cations
: : : : : : : :
: : : : : : : :
: : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : :
91
91 93 97 102 104 112 114 116
5 Stability and Performance of Feedback Systems
Feedback Structure : : : : : : : : : WellPosedness of Feedback Loop : Internal Stability : : : : : : : : : : Coprime Factorization over RH1 : Feedback Properties : : : : : : : : The Concept of Loop Shaping : : : Weighted H2 and H1 Performance Notes and References : : : : : : : :
117
117 119 121 126 130 134 137 141
6 Performance Limitations
Introduction : : : : : : : : : : : : : : : : : : Integral Relations : : : : : : : : : : : : : : : Design Limitations and Sensitivity Bounds : Bode's Gain and Phase Relation : : : : : : Notes and References : : : : : : : : : : : : :
: : : : :
: : : : :
: : : : :
: : : : :
: : : : :
: : : : :
143
143 145 149 151 152
CONTENTS
ix
7 Model Reduction by Balanced Truncation
7.1 7.2 7.3 7.4 Model Reduction by Balanced Truncation : : : FrequencyWeighted Balanced Model Reduction Relative and Multiplicative Model Reductions : Notes and References : : : : : : : : : : : : : : : Hankel Operator : : : : : : : : : : : : : : : : Allpass Dilations : : : : : : : : : : : : : : : : Optimal Hankel Norm Approximation : : : : L1 Bounds for Hankel Norm Approximation Bounds for Balanced Truncation : : : : : : : Toeplitz Operators : : : : : : : : : : : : : : : Hankel and Toeplitz Operators on the Disk* : Nehari's Theorem* : : : : : : : : : : : : : : : Notes and References : : : : : : : : : : : : : :
: : : : : : : : : : : : :
: : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
153
154 160 161 167 170 175 184 188 192 194 195 200 206 207 211 214 222 229 232 237 239 246 256 257 260 262 266 276 287 290
8 Hankel Norm Approximation
8.1 8.2 8.3 8.4 8.5 8.6 8.7 8.8 8.9
: : : : : : : : :
169
9 Model Uncertainty and Robustness
9.1 9.2 9.3 9.4 9.5 9.6 9.7
Model Uncertainty : : : : : : : : : : : : : : : : : : Small Gain Theorem : : : : : : : : : : : : : : : : : Stability under Stable Unstructured Uncertainties : Unstructured Robust Performance : : : : : : : : : Gain Margin and Phase Margin : : : : : : : : : : : De ciency of Classical Control for MIMO Systems Notes and References : : : : : : : : : : : : : : : : : Linear Fractional Transformations : Examples of LFTs : : : : : : : : : Basic Principle : : : : : : : : : : : Redhe er StarProducts : : : : : : Notes and References : : : : : : : :
207
10 Linear Fractional Transformation
10.1 10.2 10.3 10.4 10.5
: : : : :
: : : : :
: : : : :
: : : : :
: : : : :
: : : : :
: : : : : : : : : :
: : : : : : : : : :
: : : : : : : : : :
239
11 Structured Singular Value
11.1 11.2 11.3 11.4 11.5
General Framework for System Robustness : : Structured Singular Value : : : : : : : : : : : Structured Robust Stability and Performance Overview on Synthesis : : : : : : : : : : : : Notes and References : : : : : : : : : : : : : :
261
x
CONTENTS
12.1 12.2 12.3 12.4 12.5 12.6 12.7 Existence of Stabilizing Controllers : : : : : : : : : : Duality and Special Problems : : : : : : : : : : : : : Parameterization of All Stabilizing Controllers : : : : Structure of Controller Parameterization : : : : : : : ClosedLoop Transfer Matrix : : : : : : : : : : : : : Youla Parameterization via Coprime Factorization* : Notes and References : : : : : : : : : : : : : : : : : : All Solutions of A Riccati Equation : : : : Stabilizing Solution and Riccati Operator Extreme Solutions and Matrix Inequalities Spectral Factorizations : : : : : : : : : : : Positive Real Functions : : : : : : : : : : : Inner Functions : : : : : : : : : : : : : : : InnerOuter Factorizations : : : : : : : : : Normalized Coprime Factorizations : : : : Notes and References : : : : : : : : : : : :
12 Parameterization of Stabilizing Controllers
: : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
291
292 295 301 312 314 315 318 320 325 333 342 353 357 358 362 364 365 367 372 373 375 378 379 381 388 390 392 393 395 397 399 401 404
13 Algebraic Riccati Equations
13.1 13.2 13.3 13.4 13.5 13.6 13.7 13.8 13.9
: : : : : : : : :
: : : : : : : : :
: : : : : : : : :
: : : : : : : : :
: : : : : : : : :
: : : : : : : : : : : : : : : : : : : :
319
14 H2 Optimal Control
14.1 14.2 14.3 14.4 14.5 14.6 14.7 14.8 14.9 14.10 14.11 15.1 15.2 15.3 15.4 15.5 15.6
Introduction to Regulator Problem : : : : : : : : : Standard LQR Problem : : : : : : : : : : : : : : : Extended LQR Problem : : : : : : : : : : : : : : : Guaranteed Stability Margins of LQR : : : : : : : Standard H2 Problem : : : : : : : : : : : : : : : : Optimal Controlled System : : : : : : : : : : : : : H2 Control with Direct Disturbance Feedforward* Special Problems : : : : : : : : : : : : : : : : : : : Separation Theory : : : : : : : : : : : : : : : : : : Stability Margins of H2 Controllers : : : : : : : : : Notes and References : : : : : : : : : : : : : : : : :
365
15 Linear Quadratic Optimization
Hankel Operators : : : : : : : : : : : : : : : : : : : : : Toeplitz Operators : : : : : : : : : : : : : : : : : : : : Mixed HankelToeplitz Operators : : : : : : : : : : : : Mixed HankelToeplitz Operators: The General Case* Linear Quadratic MaxMin Problem : : : : : : : : : : Notes and References : : : : : : : : : : : : : : : : : : :
393
CONTENTS
xi
16 H1 Control: Simple Case
16.1 16.2 16.3 16.4 16.5 16.6 16.7 16.8 16.9 16.10 16.11 16.12
Problem Formulation : : : : : : : : Output Feedback H1 Control : : : Motivation for Special Problems : : Full Information Control : : : : : : Full Control : : : : : : : : : : : : : Disturbance Feedforward : : : : : Output Estimation : : : : : : : : : Separation Theory : : : : : : : : : Optimality and Limiting Behavior Controller Interpretations : : : : : An Optimal Controller : : : : : : : Notes and References : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
405
405 406 412 416 423 423 425 426 431 436 438 439
17 H1 Control: General Case
17.1 17.2 17.3 17.4 17.5 17.6 17.7 17.8 17.9 17.10 18.1 18.2 18.3 18.4
General H1 Solutions : : : : : : : : : : : : : : : : : Loop Shifting : : : : : : : : : : : : : : : : : : : : : : Relaxing Assumptions : : : : : : : : : : : : : : : : : H2 and H1 Integral Control : : : : : : : : : : : : : : H1 Filtering : : : : : : : : : : : : : : : : : : : : : : Youla Parameterization Approach* : : : : : : : : : : Connections : : : : : : : : : : : : : : : : : : : : : : : State Feedback and Di erential Game : : : : : : : : Parameterization of State Feedback H1 Controllers : Notes and References : : : : : : : : : : : : : : : : : :
441
442 445 449 450 454 456 460 462 465 468
18 H1 Loop Shaping
Robust Stabilization of Coprime Factors : : : : : : : : Loop Shaping Using Normalized Coprime Stabilization Theoretical Justi cation for H1 Loop Shaping : : : : Notes and References : : : : : : : : : : : : : : : : : : : Controller Reduction with Stability Criteria : : H1 Controller Reductions : : : : : : : : : : : : FrequencyWeighted L1 Norm Approximations An Example : : : : : : : : : : : : : : : : : : : : Notes and References : : : : : : : : : : : : : : :
469
469 478 481 487
19 Controller Order Reduction
19.1 19.2 19.3 19.4 19.5
: : : : :
: : : : :
: : : : :
: : : : :
489
490 497 504 509 513
xii
CONTENTS
20.1 20.2 20.3 Lagrange Multiplier Method : : : : : : : : : : : : : : : : : : : : : : 515 Fixed Order Controllers : : : : : : : : : : : : : : : : : : : : : : : : 520 Notes and References : : : : : : : : : : : : : : : : : : : : : : : : : : 525 Discrete Lyapunov Equations : : : : : : Discrete Riccati Equations : : : : : : : : Bounded Real Functions : : : : : : : : : Matrix Factorizations : : : : : : : : : : : Discrete Time H2 Control : : : : : : : : Discrete Balanced Model Reduction : : : Model Reduction Using Coprime Factors Notes and References : : : : : : : : : : :
20 Structure Fixed Controllers
515
21 Discrete Time Control
21.1 21.2 21.3 21.4 21.5 21.6 21.7 21.8
: : : : : : : :
: : : : : : : :
: : : : : : : :
: : : : : : : :
: : : : : : : :
: : : : : : : :
: : : : : : : :
: : : : : : : :
: : : : : : : :
: : : : : : : :
: : : : : : : :
: : : : : : : :
: : : : : : : :
: : : : : : : :
: : : : : : : :
527
527 528 537 542 548 553 558 561
Bibliography Index
563 581
Preface
This book is inspired by the recent development in the robust and H1 control theory, particularly the statespace H1 control theory developed in the paper by Doyle, Glover, Khargonekar, and Francis 1989] (known as DGKF paper). We give a fairly comprehensive and stepbystep treatment of the statespace H1 control theory in the style of DGKF. We also treat the robust control problems with unstructured and structured uncertainties. The linear fractional transformation (LFT) and the structured singular value (known as ) are introduced as the uni ed tools for robust stability and performance analysis and synthesis. Chapter 1 contains a more detailed chapterbychapter review of the topics and results presented in this book. We would like to thank Professor Bruce A. Francis at University of Toronto for his helpful comments and suggestions on early versions of the manuscript. As a matter of fact, this manuscript was inspired by his lectures given at Caltech in 1987 and his masterpiece { A Course in H1 Control Theory. We are grateful to Professor Andre Tits at University of Maryland who has made numerous helpful comments and suggestions that have greatly improved the quality of the manuscript. Professor Jakob Stoustrup, Professor Hans Henrik Niemann, and their students at The Technical University of Denmark have read various versions of this manuscript and have made many helpful comments and suggestions. We are grateful to their help. Special thanks go to Professor Andrew Packard at University of CaliforniaBerkeley for his help during the preparation of the early versions of this manuscript. We are also grateful to Professor Jie Chen at University of CaliforniaRiverside for providing material used in Chapter 6. We would also like to thank Professor KangZhi Liu at Chiba University (Japan) and Professor Tongwen Chen at University of Calgary for their valuable comments and suggestions. In addition, we would like to thank G. Balas, C. Beck, D. S. Bernstein, G. Gu, W. Lu, J. Morris, M. Newlin, L. Qiu, H. P. Rotstein and many other people for their comments and suggestions. The rst author is especially grateful to Professor Pramod P. Khargonekar at The University of Michigan for introducing him to robust and H1 control and to Professor Tryphon Georgiou at University of Minnesota for encouraging him to complete this work. Kemin Zhou xiii
xiv
PREFACE
John C. Doyle Keith Glover
xv
xvi NOTATION AND SYMBOLS Notation and Symbols R C F R+ C . ii = 1 and ij = 0 if i 6= j unit step function 2 \ 2 3 ~ := ' / j j ij Re ( ) (t) 1+(t) . and C . either R or C nonnegative real numbers open and closed lefthalf plane open and closed righthalf plane imaginary axis unit disk closed unit disk unit circle belong to subset union intersection end of proof end of example end of remark de ned as asymptotically greater than asymptotically less than much greater than much less than complex conjugate of 2 C absolute value of 2 C real part of 2 C unit impulse Kronecker delta function. jR D D @D eld of real numbers eld of complex numbers eld. C + and C + C 0 .
shorthand for (A. hx yi = 0 orthogonal complement of D.e. (A) stable invariant subspace of A X+ (A) antistable invariant subspace of A Ric(H ) g f the stabilizing solution of an ARE convolution of g and f Kronecker product direct sum or Kronecker sum angle inner product orthogonal.1 inverse of A A+ pseudo inverse of A A. e. R(A) image (or range) space of A Ker(A).1 0] L2+ L2.g.1 0] shorthand for l2 0 1) shorthand for l2 (. H2 time domain Lebesgue space subspace of L2 (.. N(A) kernel (or null) space of A X.1 1) subspace of L2 (.1 ) det(A) determinant of A Trace(A) trace of A (A) eigenvalue of A (A) spectral radius of A (A) the set of spectrum of A (A) largest singular value of A (A) smallest singular value of A ith singular value of A i (A) (A) condition number of A kAk spectral norm of A: kAk = (A) Im(A). l2+ l2.1 1) L2 0 1) L2 (. i..1 0) square integrable functions on C 0 including at 1 square integrable functions on @ D subspace of L2 (j R) with analytic extension to the rhp \ hi x?y D? S? L2 (.NOTATION AND SYMBOLS In aij ] xvii n n identity matrix a matrix with aij as its ith row and j th column element diag(a1 : : : an ) an n n diagonal matrix with ai as its ith diagonal element AT transpose A adjoint operator of A or complex conjugate transpose of A A.1 1) shorthand for L2 0 1) shorthand for L2 (. L2 (j R) L2 (@ D ) H2 (j R) . D D? D or D is unitary ? ? orthogonal complement of subspace S .
A).1 B + D lower LFT upper LFT star product F` (M Q) F u ( M Q) S (M N ) . e. BH1 and B pre x Bo open unit ball pre x R real rational. RH1 and RH2 . etc R s] Rp (s) G (s) G (z ) A B C D polynomial ring rational proper transfer matrices shorthand for GT (. (@ D ) H1 subspace of L2 (@ D ) with analytic extension to the inside of @ D subspace of L2 (j R) with analytic extension to the lhp subspace of L2 (@ D ) with analytic extension to the outside of @ D functions bounded on Re(s) = 0 including at 1 functions bounded on @ D the set of L1 (j R) functions analytic in Re(s) > 0 the set of L1 (@ D ) functions analytic in jz j < 1 the set of L1 (j R) functions analytic in Re(s) < 0 the set of L1 (@ D ) functions analytic in jz j > 1 pre x B or B closed unit ball.1 ) (discrete time) shorthand for state space realization C (sI .s) (continuous time) shorthand for GT (z . H1 (j R) .xviii NOTATION AND SYMBOLS H2 (@ D ) ? H2 (j R) ? H2 (@ D ) L 1 ( j R) L1 (@ D ) H1 (j R) H1 (@ D ) . e. A).g.1 B + D or C (zI .g..
xix .
xx LIST OF ACRONYMS List of Acronyms ARE BR CIF DF FC FDLTI FI HF i lcf LF LFT lhp or LHP LQG LQR LTI LTR MIMO nlcf NP nrcf NS OE OF OI rcf rhp or RHP RP RS SF SISO SSV SVD algebraic Riccati equation bounded real complementary inner factor disturbance feedforward full control nite dimensional linear time invariant full information high frequency if and only if left coprime factorization low frequency linear fractional transformation lefthalf plane Re(s) < 0 linear quadratic Gaussian linear quadratic regulator linear time invariant loop transfer recovery multiinput multioutput normalized left coprime factorization nominal performance normalized right coprime factorization nominal stability output estimation output feedback output injection right coprime factorization righthalf plane Re(s) > 0 robust performance robust stability state feedback singleinput singleoutput structured singular value ( ) singular value decomposition .
1983].Introduction 1. Here we see how the practical but classical methods yielded to the more sophisticated modern theory. Unfortunately. Robust control sought to blend the best of both worlds.1 Historical Perspective This book gives a comprehensive treatment of optimal H2 and H1 control theory and an introduction to the more general subject of robust control. The H1 optimal control theory was originally formulated by Zames 1981] in an inputoutput setting. 1 1 .. it may be helpful to provide some historical perspective of the statespace H1 control theory to be presented in this book. both mathematically and computationally. 1978 Ball and Helton. the standard frequencydomain approaches to H1 started running into signi cant obstacles in dealing with multiinput multioutput (MIMO) systems.1. This section is not intended as a review of the literature in H1 theory or robust control. but rather only an attempt to outline some of the work that most closely touches on our approach to statespace H1 . H1 theory seemed to many to signal the beginning of the end for the statespace methods which had dominated control for the previous 20 years. Most solution techniques available at that time involved analytic functions (NevanlinnaPick interpolation) or operatortheoretic methods Sarason. much as the H2 (or LQG) theory of the 1950's had. Hopefully our lack of written historical material will be somewhat made up for by the pictorial history of control shown in Figure 1. Indeed. in contrast to the approach adopted in the famous book by Francis 1987]: A Course in H1 Control Theory. Since the central subject of this book is statespace H1 optimal control. The strange creature that resulted is the main topic of this book. 1967 Adamjan et al.
. Unfortunately. involving several Riccati equations of increasing dimension. the 1984 procedure \solved" the general rational H1 optimal control problem and much of the subsequent work in H1 control theory focused on the 2 2block problems. 1987] give expositions of this approach.1: A picture history of control Not surprisingly. The steps in this solution were as follows: parameterize all internallystabilizing controllers via Youla et al. Encouragement came . Both Francis. presented in Doyle 1984]. 1976] obtain realizations of the closedloop transfer matrix convert the resulting modelmatching problem into the equivalent 2 2block general distance or best approximation problem involving mixed HankelToeplitz operators reduce to the Nehari problem (Hankel only) solve the Nehari problem by the procedure of Glover 1984]. either in the modelmatching or general distance forms.2 INTRODUCTION Figure 1. the associated complexity of computation was substantial. relied heavily on statespace methods. In a mathematical sense. 1987] and Francis and Doyle. although more as a computational tool than in any essential way. and formulae for the resulting controllers tended to be very complicated and have high state dimension. which will be referred to as the \1984" approach. the rst solution to a general rational MIMO H1 optimal control problem.
They showed that for the statefeedback H1 problem one can choose a constant gain as a (sub)optimal controller. In addition to providing controller formulae that are simple and expressed in terms of plant data as in Glover and Doyle 1988]. such as completing the square. and a separation argument. but using a new 2 2block solution. Petersen. This suggested the likely existence of similarly low dimension optimal controllers in the general 2 2 case. the methods in that paper are a fundamental departure from the 1984 approach. not surprising. In addition. a pair of 2 1 block problems. Foias and Tannenbaum 1988] developed an interesting class of operators called skew Toeplitz to study the 2 2block problem. Kwakernaak 1986] using a polynomial approach. in the style of Willems 1971]. Also. The very simplicity of the new formulae and their similarity with the H2 ones suggested a more direct approach. Additional progress on the 2 2block problems came from Ball and Cohen 1987]. to these cases. that a subsequent minimal realization of the controller has state dimension no greater than that of the generalized plant G.g kGk1 < 1). In particular. and nite dimensional to in nite dimensional. and Khargonekar. which are beyond the scope of this book. the Youla parameterization and the resulting 2 2block modelmatching problem of the 1984 solution are avoided entirely replaced by a more purely statespace approach involving observerbased compensators. who gave a statespace solution involving 3 Riccati equations. This book in some sense can be regarded as an expansion of the DGKF paper. Zhou and Khargonekar 1988]. Independent encouragement for a simpler approach to the H1 problem came from papers by Petersen 1987]. a formula for the statefeedback gain matrix was given in terms of an algebraic Riccati equation. Khargonekar. 1960] on linear fractional transformations). by generalizing timeinvariant to timevarying. in nite horizon to nite horizon.1. Glover. and linearquadratic di erential games. and Francis 1989]). Other approaches have been derived by Hung 1989] using an interpolation theory approach. together with a cumbersome back substitution. these papers established connections between H1 optimal control. quadratic stabilization. Jonckheere and Juang 1987] showed a connection between the 2 1block problem and previous work by Jonckheere and Silverman 1978] on linearquadratic control. and Zhou 1990]. for problems transformable to 2 1block problems. and spectral factorizations. . A ourish of activity has begun on these problems since the publication of the DGKF paper and numerous results have been published in the literature. but its use is more straightforward. Petersen. many results in DGKF paper generalize mutatis mutandis. Riccati equations. and Kimura 1988] using a method based on conjugation. Khargonekar. and the connection between frequency domain inequalities (e. The simple state space H1 controller formulae to be presented in this book were rst derived in Glover and Doyle 1988] with the 1984 approach. and Rotea 1988].1. The statespace theory of H1 can be carried much further. Historical Perspective 3 from Limebeer and Hung 1987] and Limebeer and Halikias 1988] who showed. The key to this was a return to simple and familiar statespace tools. The landmark breakthrough came in the DGKF paper (Doyle. The operator theory still plays a central role (as does Redhe er's work Redhe er.
therefore. In view of the above classi cation. The speci c choice of the topics depends on the time allotted for the course and the preference of the instructor. these two chapters alone may be used as a short course on model reductions. In particular. it may be advisable to skip those topics like Chapter 2 on the regular lectures and leave them for students to read. for example. Discrete Time Systems: Chapter 21. Model Reduction: Chapters 7 8.9 of Chapter 3 and. The diagram in Figure 1. Some other topics may only require some basic linear system theory. the diagram shows that the only prerequisite for Chapters 7 and 8 is Section 3. we have tried to collect some most frequently used formulas and results in one place for the convenience of reference although they may not have any direct connection with the main results presented in the book. we have tried to balance the broadness and the depth of the material covered in the book. Stability and Performance: Chapters 4 6. in principle. For example. parts of Chapter 5 (5:1 5:3 5:5 5:7). Robustness: Chapters 9 11. many readers may nd that it is not di cult to go directly to Chapters 9 11.2 How to Use This Book This book is intended to be used either as a graduate textbook or as a reference for control engineers. In some cases. Hence one may only cover those related sections of Chapter 13 if time is the factor. It is obvious that only some selected topics in this book can be covered in an one or two semester course. Chapter 13 on algebraic Riccati equations. parts of Chapter 13 (13:2 13:4 13:6).2 shows roughly the relations among the chapters and should give the users some idea for the selection of the topics. On the other hand. Chapter 10. For example. one only needs the knowledge of Sections 13. Although Chapters 7 8 are very much independent of other topics and can. if the book is used as a textbook.6). H2 and H1 Control: Chapters 12 19. be studied at any .4 INTRODUCTION 1. Chapter 12 (except Section 12. one possible choice for an onesemester course on robust control would cover Chapters 4 5 9 11 or 4 11 and an onesemester advanced course on H2 and H1 control would cover (parts of) Chapters 12 19. some chapters have been written su ciently selfcontained so that one may jump to those special topics without going through all the preceding chapters. readers may nd that those matrix formulas collected in Chapter 2 on linear algebra convenient in their research.2 and 13. Chapter 15 and Chapter 16.6 of Chapter 13 to understand Chapter 14. With the second objective in mind. for instance. The book is separated roughly into the following subgroups: Basic Linear System Theory: Chapters 2 3. Lagrange Method: Chapter 20. Similarly. Another possible choice for an onesemester course on H1 control may include Chapter 4.
.1. Figure 1.2: Relations among the chapters . ?.2. How to Use This Book 2 3 5 4 ? 3:9  ? A U A 5 6 7  8 ? 9 ? 10  11 ? 12 15 16 17 19 @ @ R @ I @ @ @ 13 14 13:2 13:6 13:4 ? ? ? ? ?  18 20 21 .
pole placement.1. In particular. Particularly.13. and state space realizations. system poles and zeros.3.16. 1.1: Possible choices for an onesemester course (* chapters may be omitted) Table 1.6 16 10.4.5. A course on model and controller reductions may only include the concept of H1 control and the H1 controller formulas with the detailed proofs omitted as suggested in the above table. observability.4.13.5.13.13. Chapter 3 reviews some system theoretical concepts: controllability.6 7 10 14 8 12 15 5.1 5. detectability. observer theory.7 13. Note that some of the statements in this section are not precise.5. they may serve as an introduction to sources of model uncertainties and hence to robustness problems. Robust Control 4 5 6* 7* 8* 9 10 11 Model & Controller Reductions 4 12 3.2.7 13.1 19* 19 H1 Control H1 Control Advanced Table 1.6 INTRODUCTION stage with the background of Section 3.1 15 17* 16.1 lists several possible choices of topics for an onesemester course. they are true under certain assumptions that are not explicitly stated.4. We show that for a given stable transfer function G(s) there A B is a state space realization G(s) = C D such that the controllability Gramian P and the observability Gramian Q de ned below are equal and diagonal: P = Q = = diag( 1 2 : : : n ) where AP + PA + BB = 0 .9.2.5. stabilizability. Chapter 2 reviews some basic linear algebra facts and treats a special class of matrix dilation problems.3 Highlights of The Book The key results in each chapter are highlighted below. Readers should consult the corresponding chapters for the exact statements and conditions. we show min X X B C A = max C A B A and characterize all optimal (suboptimal) X .9 5.2 16 18* 17. the balanced state space realizations are studied in some detail.
Chapter 5 introduces the feedback structure and discusses its stability and performance properties.K .1 P ^ ).1 .1 P K (I . let A G(s) = C B 2 RH1 0 g(t) = CeAt B Z kg(t)k dt 2 i . while the Poisson integral formula is used to study the design constraints . For example.C C BBA= : .1. P K ).1 = I + K (I . kg uk 1 n X w1 + +6 e e1  P ^ K + e2 ?+ w2 e We show that the above closedloop system is internally stable if and only if ^ ^ ^ ^ ^ I .P I (I . The input/output gains of a stable linear system under various input signals are characterized. The sensitivity integral relations are used to study the design limitations imposed by bandwidth constraint and the openloop unstable poles.1 2 RH1 : ^ ). P K (I . Chapter 6 introduces some multivariable versions of the Bode's sensitivity integral relations and Poisson integral formula.3. P K ). Some state Then kGk1 = sup kuk 2 and 1 kGk1 0 2 i=1 space methods of computing real rational H2 and H1 transfer matrix norms are also presented: kGk2 = trace(B QB ) = trace(CPC ) 2 and kGk1 = maxf : H has an eigenvalue on the imaginary axisg where 2 A H = . Highlights of The Book A Q + QA + C C = 0: 7 Chapter 4 de nes several norms for signals and introduces the H2 spaces and the H1 spaces. We show that H2 and H1 norms come out naturally as measures of the worst possible performance for many classes of input signals. P K Alternative characterizations of internal stability using coprime factorizations are also presented.
1: q .1 (G . Then we show that Z 1 This equality shows that the design limitations in multivariable systems are dependent on the directionality properties of the sensitivity function as well as those of the poles (and zeros).1 ) D. if G. 1 1 + 2 i ( 1 + i2 + i ) . and P and Q satisfy PA + AP + BB = 0 Q(A .1 C = 0 such that P = Q = diag( 1 2 ) with G partitioned compatibly as before. BD. in addition to the dependence upon pole (and zero) locations which is known in singleinput singleoutput systems. BD.1 2 RH1 . and let pi be the right half plane poles of the openloop system and i be the corresponding pole directions. then is stable and minimum phase.8 INTRODUCTION imposed by the nonminimum phase zeros.1 C ) + (A . Gr ) 1 r i=r+1 N Y i=r+1 1 + 2 i ( 1 + i2 + i ) . Gr ) 1 G. Chapter 7 considers the problem of reducing the order of a linear multivariable dynamical system using the balanced truncation method. and satis es respectively the following relative and multiplicative error bounds: Gr (s) = A11 B1 C1 D N Y q G. For example.1 (G . Gr (s)k1 2 i: i=r+1 A11 A12 A21 A22 C1 C2 B1 B2 D 3 5 2 RH1 On the other hand. Suppose 2 0 ln (S (j!))d! = max X ! i (Repi ) i i + H1 H1 0: G(s) = 4 is a balanced realization with controllability and observability Gramians P = Q = = diag( 1 2 ) 1 = diag( 1 Is1 2 Is2 : : : r Isr ) 2 = diag( r+1 Isr+1 r+2 Isr+2 : : : N IsN ): Then the truncated system Gr (s) = A11 B1 is stable and satis es an additive C1 D error bound: N X kG(s) . let S (s) be a sensitivity function.1 C ) Q + C (D.
3. G(s) . 1 N X inf kG . We show that for a given G(s) of McMillan degree n ^ there is a G(s) of McMillan degree r < n such that ^ ^ G(s) . D0 1 i: i=r+1 The wellknown Nehari's theorem is also shown: Q2RH. z ∆ w nominal system Chapter 10 introduces the linear fractional transformation (LFT). In particular.1. In particular. we show that an uncertain system described below with an unstructured uncertainty 2 RH1 with k k1 < 1 is robustly stable if and only if the transfer function from w to z has H1 norm no greater than 1. z M w z y G w u K . G(s) H = r+1 : Moreover. Qk1 = kGkH = 1 : Chapter 9 derives robust stability tests for systems under various modeling assumptions through the use of a small gain theorem. we show that every analysis problem can be put in an LFT form with some structured (s) and some interconnection matrix M (s) and every synthesis problem can be put in an LFT form with a generalized plant G(s) and a controller K (s) to be designed. there is a constant matrix D0 such that ^ G(s) . We show that many control problems can be formulated and treated in the LFT framework. G(s) H = inf G(s) . Highlights of The Book 9 Chapter 8 deals with the optimal Hankel norm approximation and its applications in L1 norm model reduction.
. full control (FC) problems and output estimation (OE). The construction of the controller parameterization is done via separation theory and a sequence of special problems. which are socalled full information (FI) problems. We show that an uncertain system is robustly stable for all i 2 RH1 with k i k1 < 1 if and only if the structured singular value ( ) of the corresponding interconnection model is no greater than 1. The relations among these special problems are established.10 INTRODUCTION Chapter 11 considers robust stability and performance for systems with multiple sources of uncertainties. FI dual 6 ?  FC 6 equivalent equivalent dual DF 2 OE ? 3 5 For a given generalized plant A B1 B2 G(s) = G11 (s) G12 (s) = 4 C1 D11 D12 G21 (s) G22 (s) C2 D21 D22 we show that all stabilizing controllers can be parameterized as the transfer matrix from y to u below where F and L are such that A + LC2 and A + B2 F are stable. disturbance feedforward (DF) problems. ∆1 ∆4 nominal system ∆2 ∆3 Chapter 12 characterizes all controllers that stabilize a given dynamical system G(s) using the state space approach.
A). A solution to the following ARE A X + XA + XRX + Q = 0 is said to be a stabilizing solution if A + RX is stable.Q . the methods to obtain the solutions.. c A 66 c c 6 .1. P 0 or P 0) and (P A) has no unobservable modes on the j!axis. A ). Now let A H := .L u1 F y1  Q Chapter 13 studies the Algebraic Riccati Equation and the related problems: the properties of its solutions. we study in detail the socalled stabilizing solution and its applications in matrix factorizations. Highlights of The Book z y G D22 C2 R c 11 w u B2 c ?? .1 I P S S R (sI . where X1 X2 2 C n n . Suppose (A B ) is stabilizable and suppose either A has no eigenvalues on j!axis or P is sign de nite (i.R A and let X. denoted by X = Ric(H ). In particular. If X1 is nonsingular.e. A key result of this chapter is the relationship between the spectral factorization of a transfer function and the solution of a corresponding ARE. De ne (s) = B (. then X := X2 X1 1 is uniquely determined by H . (H ) = Im X1 X2 . and some applications. (H ) be the stable H invariant subspace and X.1 B : I .sI ..3.
1(S + B X ): Chapter 14 treats the optimal control of linear timeinvariant systems with quadratic performance criteria.1 B X ) C .A .12 Then (j!) > 0 for all 0 ! 1 () 9 a stabilizing solution X to INTRODUCTION (A .1S = 0 such that (A . BR.1B X + P . LQR and H2 problems. XBR.1 S ) . BR.BR. BR.1S = 0 () the Hamiltonian matrix has no j!axis eigenvalues. SR () 9 a solution X to (j!) 0 for all 0 ! 1 (A .1 S . BR. .F B I F = .R. BR. BR . BR.1 S ) (P . BR.(A .e. XBR.1B X + P . SR.1S ) ..1 S ) . Similarly.1B H = . there exists a M 2 Rp such that = M RM: with A M = .1 S . Furthermore. .1 S ) X + X (A . SR.1 S ) X + X (A . i. We consider a dynamical system described by an LFT with A B1 B2 G(s) = 4 C1 0 D12 5 : C2 D21 0 z y G w u 2 3  K .
G(s) as given in Chapter 14.0 A 1 1 .CA C 1 1 . . Consider a dynamical system Chapter 15 solves a maxmin problem.2 C1 C1 .e.0A . the controller that minimizes kTzw k2 . is given by L Kopt (s) := A + B2 F2 + L2C2 .B2 X1 is an optimal control.(Y2 C2 + B1 D21 ): Then the H2 optimal controller.A Furthermore.B B .1. We show that there exists an admissible controller such that kTzw k1 < i the following three conditions hold: (i) H1 2 dom(Ric) and X1 := Ric(H1 ) 0 (ii) J1 2 dom(Ric) and Y1 := Ric(J1 ) 0 where Chapter 16 considers a simpli ed H1 control problem with the generalized plant A J1 := .CB2 12 1D D12 C1 B2 A C D21 B1 C2 J2 := .A : . Highlights of The Book De ne 13 H2 := . i. a full information (or state feedback) x = Ax + B1 w + B2 u _ z = C1 x + D12 u D12 C1 D12 = 0 I : Then we show that sup umin kz k2 < if and only if H1 2 dom(Ric) and X1 = w2BL2+ 2L2+ Ric(H1 ) 0 where H1 := .. .(B2 X2 + D12 C1 ) L2 := . C2 C2 .3.0 2 : F 2 H1 control problem. which is the key to the H1 theory considered in the next chapter.e. i.B B 1 1 (iii) (X1 Y1 ) < 2 .B 2 1 1 1 D21 X2 := Ric(H2) 0 Y2 := Ric(J2 ) 0 F2 := . B2 B2 . u = F1 x with F1 := .CA C . .2B1 B1 .
~ ~ Chapter 18 rst solves a gap metric minimization problem. kQk1 < and ^ A1 := A + . Finally.Z1 L1 Z1 B2 4 F1 M1 (s) = 0 I I 0 .1 = ~ I 1 1. Then we show that K stabilizing inf K (I + PK ).1M .Y1 C2 Z1 := (I . .1 N be a normalized left coprime factorization.1 : Chapter 17 considers again the standard H1 control problem but with some assumptions in the last chapter relaxed.1 (N + ~ N ) with if and only if ~N ~M q K (I + PK ). ~ ~ N M 2 . The conventional Youla parameterization approach to the H2 and H1 problems is also outlined.2B1 B1 X1 + B2 F1 + Z1 L1 C2 F1 := .C2 2 3 5 where Q 2 RH1 . The proposed technique uses only the basic concept of loop shaping methods and then a robust .14 INTRODUCTION Moreover. N M 2 : Using this stabilization result. a loop shaping design technique is proposed. Let P = M .1 I P I q 1 = K stabilizing inf This implies that there is a robustly stabilizing controller for ~ ~ P = (M + ~ M ).2Y1 X1 ). the general state feedback H1 control problem and its relations with full information control and di erential game problems are discussed. We indicate how the assumptions can be relaxed to accommodate other more complicated problems such as singular control problems.B2 X1 L1 := . the set of all admissible controllers such that kTzw k1 < equals the set of all transfer matrices from y to u in u M1 y  Q ^ A1 .1 H : 1< ~ ~ H 1. We also consider the integral control in the H2 and H1 theory and show how the general H1 solution can be used to solve the H1 ltering problem.
U .1 < 1= 2: ^ 0 I V 22 1 ^ ^^ ^ Then K = U V . b) Let K0 = 12 . Highlights of The Book 15 stabilization controller for the normalized coprime factor perturbed system is used to construct the nal controller. Chapter 21 discusses discrete time Riccati equations and some of their applications in discrete time control. two H1 performance preserving reduction methods are proposed: ^ a) Let K0 be a stabilizing controller such that kF` (G K0 )k1 < . 1 <1 . Thus the controller reduction problem is converted to weighted model reduction problems for which some numerical methods are suggested.1 2 1 where W1 and W2 are some stable.1. Special attention is paid to the controller reduction methods that preserve the closedloop stability and performance. minimum phase and invertible transfer matrices.1 I 0 p ^ 12 . Chapter 19 considers the design of reduced order controllers by means of controller reduction. In particular. Then K is also a ^ stabilizing controller such that F`(G K ) < if 1 ^ W .1 is also a stabilizing controller such that kF`(G K )k1 < .3. Chapter 20 brie y introduces the Lagrange multiplier method for the design of xed order controllers.1 be a central H1 controller such that kF`(G K0 )k1 < and 22 ^ ^ let U V 2 RH1 be such that . Finally. K0)W .1 (K . the discrete time balanced model reduction is considered.
16 INTRODUCTION .
Then an element of the form 1 x1 + : : : + k xk with i 2 F is a linear combination over F of x1 : : : xk . Let S be a subspace of Fn . Hence we shall omit most proofs and provide proofs only for those results that either cannot be easily found in the standard linear algebra textbooks or are insightful to the understanding of some related problems. 2 2. denoted by spanfx1 x2 : : : xk g := fx = 1 x1 + : : : + k xk : i 2 Fg: A set of vectors x1 x2 : : : xk 2 Fn are said to be linearly dependent over F if there exists 1 : : : k 2 F not all zero such that 1 x2 + : : : + k xk = 0 otherwise they are said to be linearly independent.1 Linear Subspaces Let R denote the real scalar eld and C the complex scalar eld. 17 . Now let x1 x2 : : : xk 2 Fn . Fn is either Rn or C n . The detailed treatment of this topic can be found in the references listed at the end of the chapter.Linear Algebra Some basic linear algebra facts will be reviewed in this chapter. most of the results presented in this book can be understood without the knowledge of the matrix dilation theory. then a set of vectors fx1 x2 : : : xk g 2 S is called a basis for S if x1 x2 : : : xk are linearly independent and S = spanfx1 x2 : : : xk g.. However. We then treat a special class of matrix dilation problems which will be used in Chapters 8 and 17 however. The set of all linear combinations of x1 x2 : : : xk 2 Fn is a subspace called the span of x1 x2 : : : xk .e. let F be either R or C and let Fn be the vector space over F. i. For the interest of this chapter.
and thus the rank of a matrix equals the maximal number of independent rows or columns. Let A 2 Fm n be a linear transformation from Fn to Fm . denoted by dim(S ). Dually.e. A set of vectors fx1 x2 : : : xk g in Fn are mutually orthogonal if xi xj = 0 for all i 6= j and orthonormal if xi xj = ij . This number is called the dimension of S . It is always possible to extend such a basis to a full orthonormal basis fu1 u2 : : : un g for Fn . The orthogonal complement of a subspace S Fn is de ned by S ? := fy 2 Fn : y x = 0 for all x 2 S g: We call a set of vectors fu1 u2 : : : uk g an orthonormal basis for a subspace S 2 Fn if they form a basis of S and are orthonormal. A full rank square matrix is called a nonsingular matrix. then ImA = spanfa1 a2 : : : an g: The rank of a matrix A is de ned by rank(A) = dim(ImA): It is a fact that rank(A) = rank(A ). A matrix A 2 Fm n is said to have full row rank if m n and rank(A) = m.! Fm : (Note that a vector x 2 Fm can also be viewed as a linear transformation from F to Fm . i. Note that (KerA)? is a subspace of Fn . it can be easily seen that dim(KerA) + dim(ImA) = n and dim(ImA) = dim(KerA)? . A : Fn 7. it is said to have full column rank if n m and rank(A) = n. It is easy .) Then the kernel or null space of the linear transformation A is de ned by KerA = N (A) := fx 2 Fn : Ax = 0g and the image or range of A is ImA = R(A) := fy 2 Fm : y = Ax x 2 Fn g: It is clear that KerA is a subspace of Fn and ImA is a subspace of Fm . More generally. Note that in this case S ? = spanfuk+1 : : : un g and fuk+1 : : : un g is called an orthonormal completion of fu1 u2 : : : uk g. where the superscript denotes complex conjugate transpose and ij is the Kronecker delta function with ij = 1 for i = j and ij = 0 for i 6= j . Let ai i = 1 2 : : : n denote the columns of a matrix A 2 Fm n . hence anything said for the general matrix case is also true for the vector case.18 LINEAR ALGEBRA such a basis for a subspace S is not unique but all bases for S have the same number of elements. Moreover.. a collection of subspaces S1 S2 : : : Sk of Fn are mutually orthogonal if x y = 0 whenever x 2 Si and y 2 Sj for i 6= j .
the columns of D?. (ii) the columns of B 2 ImA. (iii) rank A B =rank(A). Lemma 2.1 Let D = d1 : : : dk 2 Fn k (n > k) be such that D D = I . Now let A = aij ] 2 C n n . Then rank (A) + rank(B ) . a matrix M with mij as its ith row and j th column's element will sometimes be denoted as M = mij ] in this book. if it exists.2 Consider the linear equation where A 2 Fn l and equivalent: (i) there exists a solution X 2 Fl m .k) such that D D? is a unitary matrix. di i = k + 1 : : : n. and it satis es U U = I = UU . Furthermore. then the trace of A is de ned as B 2 Fn m are given matrices. so di i = 1 2 : : : k are orthonormal. Lemma 2.1. n rank(AB ) minfrank (A) rank(B )g: For simplicity. Furthermore. the solution. we will use In to emphasis that it is an n n identity matrix. Then the following statements are AX = B Trace(A) := n X i=1 aii : Trace has the following properties: Trace( A) = Trace(A) 8 2 C A 2 C n n Trace(A + B ) = Trace(A) + Trace(B ) 8A B 2 C n n Trace(AB ) = Trace(BA) 8A 2 C n m B 2 C m n : . The following lemma concerns the rank of the product of two matrices. form an orthonormal completion of fd1 d2 : : : dk g. The following results are standard: Lemma 2.3 (Sylvester's inequality) Let A 2 Fm n and B 2 Fn k . We will mostly use I as above to denote an identity matrix with compatible dimensions. is unique if and only if A has full column rank. Then there exists a matrix D? 2 Fn (n. (iv) Ker(A ) Ker(B ). Linear Subspaces 19 to see that rank(A) = rank(AT ) = rank(PA) if T and P are nonsingular matrices with appropriate dimensions. but from time to time.2. The following lemma is useful. A square matrix U 2 F n n whose columns form an orthonormal basis for Fn is called an unitary matrix (or orthogonal matrix if F = R).
Dually. The maximal modulus of the eigenvalues is called the spectral radius..1 J = diagfJ1 J2 : : : Jl g Ji = diagfJi1 Ji2 : : : Jimi g i Jij = with Pl 6 6 6 6 6 4 1 3 7 7 7 7 7 5 i 1 .4 For any square complex matrix A 2 C n n . Atij1 = .. If 2 (A) then any nonzero vector x 2 C n that satis es 2. as usual. .. A).. denoted by (A) := 1maxn j i j i where.2 Eigenvalues and Eigenvectors Ax = x is referred to as a right eigenvector of A.20 LINEAR ALGEBRA Let A 2 C n n then the eigenvalues of A are the n roots of its characteristic polynomial p( ) = det( I . : : : Tl : : : Timi : : : tijnij i tij 1 i T = T1 T2 Ti = Ti1 Ti2 Tij = tij1 tij2 where tij1 are the eigenvectors of A. i 1 2 C nij nij The transformation T has the following form: i=1 Pmi j =1 nij = n. j j denotes the magnitude. (A) := f 1 2 : : : n g if i is a root of p( ). there exists a nonsingular matrix T such that where 2 A = TJT . and with f i : i = 1 : : : lg as the distinct eigenvalues of A. This set of roots is called the spectrum of A and is denoted by (A) (not to be confused with singular values de ned later). a nonzero vector y is called a left eigenvector of A if y A= y : It is a well known (but nontrivial) fact in linear algebra that any complex matrix admits a Jordan Canonical representation: Theorem 2. That is.
: n In this case.1 + + an : Then An + a1 An. we will denote max (A) for the largest eigenvalue of A and min (A) for the smallest eigenvalue. . if A is a Hermitian matrix. . In particular. A has the following spectral decomposition: A= where yi 2 C n is given by 2 6 6 6 4 n X i=1 i xi yi y1 y2 yn . a square matrix A with all distinct eigenvalues is cyclic and can be diagonalized: 2 6 6 6 4 1 3 7 7 7 5 A x1 x2 xn = x1 x2 xn 2 .1 : In general. However.1) are called the generalized eigenvectors of A. A = A . eigenvalues need not be real. A) = n + a1 n..2. a matrix A is cyclic if its Jordan form has mi = 1 i = 1 : : : l. then there exist a unitary matrix U and a real diagonal matrix such that A = U U .. 3 7 7 7 5 = x1 x2 xn . The following theorem is useful in linear system theory. the generalized eigenvectors tijl 8l < q. i I )tijk = tij(k. if A is real and is a real eigenvalue of A.e..1 A square matrix A 2 Rn n is called cyclic if the Jordan canonical form of A has one and only one Jordan block associated with each distinct eigenvalue. then there is a real eigenvector corresponding to .. Eigenvalues and Eigenvectors 21 and tijk 6= 0 de ned by the following linear equations for k 2 (A . where the diagonal elements of are the eigenvalues of A and the columns of U are the eigenvectors of A.1 + + an I = 0: 1 For example. Theorem 2. are called the lower rank generalized eigenvectors of tijq . this is the case if A is Hermitian.5 (CayleyHamilton) Let A 2 C n n and denote det( I . and neither do their corresponding eigenvectors. In the case that all eigenvalues of a matrix A are real1 . i. De nition 2. More speci cally.2. Clearly. For a given integer q nij .
if A22 is nonsingular. then A has the following decomposition: I 0 A11 A12 = A21 A22 A21 A. .1 A12 11 0 I is nonsingular. .1 Rn = Rn. A21 A.1 . A). n.1 A12 . 2 Trace(R2 A) R1 = I R2 = R1 A + a1 I 0 = Rn A + an I = An + a1 An.1 A21 I 22 . 1 Trace(Rn. then A11 A12 = I A12 A.1 A + an.1 22 A21 A22 0 I 0 0 A22 I 0 A. Note that the CayleyHamilton Theorem follows from the fact that a1 a2 = = . A) (R1 n. = = . n Trace(Rn A) The proof is left to the reader as an exercise.1 I 1 0 = Rn A + an I: an .1 + + an I: 2.1 diag f: : : n + a1 in. The proof for the general case follows from the following lemma. Now suppose A11 is nonsingular.3 Matrix Inversion Formulas Let A be a square matrix partitioned as follows A := A11 A12 A21 A22 where A11 and A22 are also square matrices. .1 = det( I .1 A) an. with := A22 . .6 Let A 2 C n n .1 I 11 A11 0 0 I A. A) = n + a1 n. Then 1 ( I .1 + + an I = T . Lemma 2.1 + + an where ai and Ri can be computed from the following recursive formulas: . Trace A 1 .22 This is obvious if A has distinct eigenvalues.1 + i + an : : :g T = 0 and i is an eigenvalue of A. and A is nonsingular i 11 Dually. Since LINEAR ALGEBRA An + a1 An.1 + R2 n.2 + + Rn ) and det( I .
1 A21 .1 A21 A.1 A21 .1 : 11 11 .1 . if A22 is nonsingular.1 0 11 .1 + A. we have I B det . then 23 A11 A12 .1 A21 A22 .1 A12 .1 : 22 11 11 11 11 As a consequence of the matrix decomposition formulas mentioned above. A21 A.1 A12 ): 11 On the other hand. A21 A.1 A21 A22 .m I = det(In + CB ) = det(Im + BC ) C n and for x y 2 C n det(In + xy ) = 1 + y x: . A12 A.1 A12 A.A.2.1 .1 A21 A.1 A12 A. Moreover.1 A12 ).1 11 and # " . A12 A.A22 11 22 A11 A12 . A12 A.1 A11 A12 . Suppose A11 and A22 are both nonsingular matrices.1 A21 .1 11 11 11 11 . if A is nonsingular. Suppose A11 is nonsingular. and A is nonsingular i is nonsingular.1 .1 A21 A. then det A = det A22 det(A11 .3. then (A11 .1 A21 A.1 = 22 : A21 A22 .1 + A. for any B 2 C m n and C 2 C n m . Matrix Inversion Formulas with := A11 . .1 A21 ).1 A.1 = A.A.1 22 0 A22 0 A22 The following identity is also very useful. The matrix 22 ( ) is called the Schur complement of A11 (A22 ) in A. then det A = det A11 det(A22 .1 = A.1 A12 A.1 = A.1 A. we can calculate the determinant of a matrix by using its submatrices.1 A21 ): 22 In particular.1 = A.1 22 22 22 22 The above matrix inversion formulas are particularly simple if A is block triangular: A11 0 .1 A12 .A.1 + A.1 A12 (A22 . .
1 : d d .(X .1 @X TracefAX B g @ @X log det X @ T @X det X @ k @X detfX g = AT B T = BA = AT X T B T + B T X T AT = AT XB T + AXB = k(X k.4 Matrix Calculus @ @ @X F (X ) := @xij F (X ) : Let A and B be constant complex matrices with compatible dimensions. the derivative of a matrix A( ) 2 C m n with respect to a scalar 2 C is de ned as dA := daij d d d d d dA.1 T = .A.i. Then the following is a list of formulas for the derivatives2: @ @X TracefAXB g @ T @X TracefAX B g @ @X TracefAXBX g @ T @X TracefAXBX g @ k @X TracefX g @ k @X TracefAX g @ .1 dA A.1)T = (X T ).1 BAX .1 @ = @X det X = (detX )(X T ).1 = .1 )T = Pk. And nally.24 LINEAR ALGEBRA Let X = xij ] 2 C m n be a real or complex matrix and F (X ) 2 C be a scalar real or complex function of X then the derivative of F (X ) with respect to X is de ned as 2.1 = k(detX k )(X T ). In particular.1 : so that all the rules applicable to a scalar function also apply here. we have d(AB ) = dA B + A dB 2 Note that transpose rather than complex conjugate transpose should be used in the list even if the involved matrices are complex matrices.1 i=0 X i AX k.
. .5. Then we have the following properties: The eigenvalues of A B are the mn numbers i j .. and X 2 C m n . .5 Kronecker Product and Kronecker Sum 2 Let A 2 C m n and B 2 C p q . . if k = m and l = n. . we have vec(AXB ) = (B T A)vec(X ): Consequently. x1n x2n xmn Then for any matrices A 2 C k m . Kronecker Product and Kronecker Sum 25 2. then the Kronecker product of A and B is de ned as A B := 6 6 4 6 a11 B a12 B a21 B a22 B am1 B am2 B .. The eigenvalues of A B = (A Im ) + (In B ) are the mn numbers i + j .. then vec(AX + XB ) = (B T A)vec(X ): Let A 2 C n n and B 2 C m m . .2. .. j = 1 2 : : : m. if the matrices A and B are square and A 2 C n n and B 2 C m m then the Kronecker sum of A and B is de ned as A B := (A Im ) + (In B ) 2 C nm nm : Let X 2 C m n and let vec(X ) denote the vector formed by stacking the columns of X into one long vector: 2 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4 x11 x21 .. i = 1 2 : : : n. . j = 1 2 : : : m. and let f i i = 1 : : : ng be the eigenvalues of A and f j j = 1 : : : mg be the eigenvalues of B . 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 vec(X ) := xm1 x12 x22 : . 3 7 7 7 5 2 C mp nq : Furthermore. .. B 2 C n l . i = 1 2 : : : n. a1n B a2n B amn B .
and C 2 Fn m are given matrices. Then Ax = x and A( x) = ( x) for any 2 C . In other words. let 1 = 2 = = l be eigenvalues of A.7 Consider the Sylvester equation AX + XB = C (2:1) where A 2 Fn n . that S is invariant for A means that the image of S under A is contained in S : AS S . Clearly. respectively. More speci cally. For example. In general.26 LINEAR ALGEBRA Let fxi i = 1 : : : ng be the eigenvectors of A and let fyj j = 1 : : : mg be the eigenvectors of B . let 1 : : : k be eigenvalues of A (not necessarily distinct). Since the eigenvalues of B T A have the form of i (A) + j (B T ) = i (A) + j (B ). Then the eigenvectors of A B and A B correspond to the eigenvalues i j and i + j are xi yj . Lemma 2. KerA. the eigenvector x de nes an onedimensional subspace that is invariant with respect to premultiplication by A since Ak x = k x 8k. 2 The properties of the Lyapunov equations will be studied in more detail in the next chapter. Using these properties. and ImA are all Ainvariant subspaces.6 Invariant Subspaces Let A : C n 7. In particular. f0g.1) is called the \Lyapunov Equation" and the necessary and su cient condition for the existence of a unique solution is that i (A) + j (A) 6= 0 8i j = 1 2 : : : n Proof. There exists a unique solution X 2 Fn m if and only if i (A)+ j (B ) 6= 0 8i = 1 2 : : : n and j = 1 2 : : : m.! C n be a linear transformation. or Ainvariant. if B = A . Then S = spanfx1 : : : xk g is an Ainvariant subspace provided that all the lower rank generalized eigenvectors are included. B 2 Fm m . (B T 2. C n . and x be a corresponding eigenvector. and . we can show the following Lemma.1) can be written as a linear matrix equation by using the Kronecker product: A)vec(X ) = vec(C ): Now this equation has a unique solution i B T A is nonsingular. a subspace S C n is called invariant for the transformation A. (2. Equation (2. As a generalization of the one dimensional invariant subspace induced by an eigenvector. be an eigenvalue of A. and let xi be the corresponding eigenvectors and the generalized eigenvectors. the conclusion follows. if Ax 2 S for every x 2 S .
Moreover. and S234 = spanfx2 x3 x4 g are not Ainvariant subspaces since the lower rank generalized eigenvector x1 of x2 is not in these subspaces. To illustrate. consider the subspace S23 . Invariant Subspaces 27 let x1 x2 : : : xl be the corresponding eigenvector and the generalized eigenvectors obtained through the following equations: (A . S24 = spanfx2 x4 g.. Then by de nition.6. (A .e. xi 2 S 81 i t. Then it is easy to verify that S123 = spanfx1 x2 x3 g S124 = spanfx1 x2 x4 g S34 = spanfx3 x4 g S1 = spanfx1 g S3 = spanfx3 g S4 = spanfx4 g S12 = spanfx1 x2 g S13 = spanfx1 x3 g S14 = spanfx1 x4 g are all Ainvariant subspaces. but this is impossible since x1 is independent of x2 and x3 .. the subspaces S2 = spanfx2 g. and 4 6 6 4 1 1 3 1 3 7 7 5 4 > 0. . Example 2. On the other hand. S1 S3 S12 S13 . 3 3 We will say subspace S is trivial if S = f0g. Since Ax2 = x2 + x1 Ax2 2 S23 would require that x1 be a linear combination of x2 and x3 .2. S23 = spanfx2 x3 g. However. . 1 I )x2 = x1 . i. Stable invariant subspaces will play an important role in computing the stabilizing solutions to the algebraic Riccati equations in Chapter 13. Ax2 2 S23 if it is an Ainvariant subspace. 1 I )x1 = 0 (A . 1 I )xl = xl. if S is a nontrivial subspace3 and is Ainvariant. An Ainvariant subspace S C n is called a stable invariant subspace if all the eigenvalues of A constrained to S have negative real parts. and S123 are stable Ainvariant subspaces. 3 < 0.1 : Then a subspace S with xt 2 S for some t l is an Ainvariant subspace only if all lower rank eigenvectors and generalized eigenvectors of xt are in S .1.1 Suppose a matrix A has the following Jordan canonical form 2 A x1 x2 x3 x4 = x1 x2 x3 x4 with Re 1 < 0. This will be further illustrated in Example 2. then there is x 2 S and such that Ax = x.
Similarly. (iii). for p = 1 2 1. we can introduce some kind of measure for a matrix. and (iv) but not necessarily (ii). a realvalued function k k de ned on X is said to be a norm on X if it satis es the following properties: (i) kxk 0 (positivity) (ii) kxk = 0 if and only if x = 0 (positive de niteness) (iii) k xk = j j kxk. when p = 1 2 1 we have kxk1 := kxk2 := n X i=1 jxi j jxi j2 v u n uX t In particular. for example kxk2 is the Euclidean distance of the vector x from the origin. Let X be a vector space. Let x 2 C n . norm is an abstraction and extension of our usual concept of length in 3dimensional Euclidean space.7 Vector Norms and Matrix Norms In this section. then the matrix norm induced by a vector pnorm is de ned as kAxk kAkp := sup kxk p : x6=0 p kxk1 := 1maxn jxi j: i i=1 . for any scalar (homogeneity) (iv) kx + yk kxk + kyk (triangle inequality) for any x 2 X and y 2 X . we will de ne vector and matrix norms. Then we de ne the vector pnorm of x as kxkp := n X i=1 jxi jp !1=p for 1 p 1: In particular. So a norm of a vector is a measure of the vector \length". the corresponding induced matrix norm can be computed as m X kAk1 = 1maxn jaij j (column sum) j i=1 Clearly. Let A = aij ] 2 C m n .28 LINEAR ALGEBRA 2. A function is said to be a seminorm if it satis es (i).
A can be viewed as a mapping from a vector space C n equipped with a vector norm k kp to another vector space C m equipped with a vector norm k kp .7. kxk < kyk i k k < 1. It is de ned as kAkF := Trace(A A) = p v u m n uX X t i=1 j =1 jaij j2 : However. kxk kyk i there is a matrix 2 Fn m with k k Furthermore. the induced norms have the interpretation of input/output ampli cation gains. Then kxk = kyk i there is a matrix U 2 Fn and U U = I . Suppose n = m. Then jx yj for some 2 F or y = 0.9 Let A and B be any matrices with appropriate dimensions. kUxk = kxk for any appropriately dimensioned unitary matrices U .8 Let x 2 Fn and y 2 Fm . So from a system theoretical point of view. the equality holds i x = y 1 such that x = y. In fact. This is because kAkp is de ned by or induced from a vector pnorm. 3. m such that x = Uy kxk kyk. the Frobenius norm is not an induced norm. then we shall denote the Euclidean 2norm of x simply by kxk := kxk2 and the induced 2norm of A by kAk := kAk2 : The Euclidean 2norm has some very nice properties: Lemma 2.2. Then . Suppose n m. 2. 1. Another often used matrix norm is the so called Frobenius norm. 4. We shall adopt the following convention throughout the book for the vector and matrix norms unless speci ed otherwise: let x 2 C n and A 2 C m n . Vector Norms and Matrix Norms kAk2 = p 29 kAk1 = 1max i m n X j =1 max (A A) jaij j (row sum) : The matrix norms induced by vector pnorms are sometimes called induced pnorms. Moreover. The following properties of matrix norms are easy to show: Lemma 2.
particularly. . 2 (UA) = 0. the inequality becomes an equality if the F norm is used. This property is useful in some matrix perturbation problems. p 6 .1 if A is invertible. . Then for any induced matrix pnorm 2 3 kAkp 6 kA21 kp kA22 kp p 6 .10 Let A be a block partitioned matrix with 2A A 6 A11 A12 A=6 . kAm1 kp kAm2 kp kAmq kp 7 : 5 p (2:2) . 6 . (A) kAk (This is also true for F norm and any induced matrix norm). it does change its eigenvalues.30 LINEAR ALGEBRA 2. (This is also true for any induced matrix norm. kAB k kAk kB k. . p2 1 p2 1 p2 # UA = p with 1 (UA) = 2.22 4 . . and kUAV kF = kAkF . Note that although premultiplication or postmultiplication of a unitary matrix on a matrix does not change its induced 2norm and F norm.1 kAk. kUAV k = kAk. kAB kF kAk kB kF and kAB kF kB k kAkF .. for any appropriately dimensioned unitary matrices U and V . .21 . Now let U= then U is a unitary matrix and " 1 p2 1 . Am1 Am2 3 7 7 =: Aij ] 7 5 kA1q kp kA2q kp 7 7 . . kA11 k A1q A2q Amq . p 2 0 0 0 Lemma 2. in the computation of bounds for structured singular values which will be studied in Chapter 10.) 3. . kA12 k Further. 1. 4. For example. and let each Aij be an appropriately dimensioned matrix. In particular. this gives A. . let A= 1 0 : 1 0 Then 1 (A) = 1 2 (A) = 0. 4 .
1 p 1.p 6 . 7 q kA k kx k 4 q 5 j =1 mj p j p j =1 Amj xj p p 2 kA11 k kA12k kA1q kp 3 2 kx1 kp 3 p p 6 kA21 kp kA22kp kA2q kp 7 6 kx2 kp 7 76 7 sup 6 . Vector Norms and Matrix Norms will be shown... 4 . . 7 kxkp =1 6 kp =1 4 P . 6 . kxq kp 3 7 7 : 7 5 p k Aij ]kp := = = 2 Pq A1j xj 3 6 Pjq=1 A2j xj 7 7 sup k Aij ] xkp = sup 6 j=1. 2x 6 x1 x = 6 . xq 3 7 7 7 5 and note that kxkp = Then 2 kx1k 6 kx2kp 6 .. kxq kp p kAm1 kp kAm2 kp kA1q kp h i kxkp =1 3 7 7 7 5 p = h sup kAij kp kAij kp i p kxkp p : 2 . It is obvious that if the F norm is used. 6 2 4 . 5 kxkp =1 . 6 7 4P .7.2) equals the left hand side.. Let a vector x be partitioned consistently with A as 31 Proof. 54 . 5 kxkp =1 kxkp=1 q A x mj j p 2 Pq 3 j=1 2 Pq kA1j k kxj k 6 Pjq=1 A1j xj p 7 6 6 j=1 A2j xj p 7 7 6 Pjq=1 kA2j kp kxj kp 7 sup 6 6 7 kxsup 6 j=1 . then the right hand side of inequality (2.. Hence only the induced pnorm cases.2. 7 6 .... 6 P . p p 6 . . 7 4 .
11 Let A 2 Fm n . . . . . Then from the de nition of kAk. It will be seen that singular values of a matrix are good measures of the \size" of the matrix and that the corresponding singular vectors are good indications of strong/weak input or output directions. There exist unitary matrices U = u1 u2 : : : um] 2 Fm m V = v1 v2 : : : vn ] 2 Fn n such that A=U V = 01 0 0 where 2 6 6 1=6 4 and 0 . 0 1 0 . .32 LINEAR ALGEBRA 2. . there is a matrix U 2 F m n such that U U = I and ~ Az = Uz: Now let We have Ax = y.8. there exists a z 2 Fn such that kAz k = kz k : ~ ~ ~ By Lemma 2. Let = kAk and without loss of generality assume m n.. 0 p 2 0 0 . p 3 7 7 7 5 1 2 0 p = minfm ng: Proof. .4 Consequently. .. U AV has the following structure: A1 := U AV = 0 w B 4 Recall that it is always possible to extend an orthonormal set of vectors to an orthonormal basis for the whole space. Let and z x = kz k 2 Fn ~ y = Uz 2 Fm : ~ Uz V = x V1 2 Fn n U = y U1 2 Fm m be unitary.8 Singular Value Decomposition A very useful tool in matrix analysis is Singular Value Decomposition (SVD). Theorem 2.
it follows that kA1 k2 An obvious induction argument gives U AV = : This completes the proof. respectively. the singular values of a matrix A are precisely the lengths of the semiaxes of the hyperellipsoid E de ned by E = fy : y = Ax x 2 C n kxk = 1g: Thus v1 is the direction in which kyk is largest for all kxk = 1 while vn is the direction in which kyk is smallest for all kxk = 1. sin 1 sin 1 cos 1 sin 2 cos 2 2 . (A) = max (A) = and 1 is an eigenvector of AA .1) .2. ui eigenvector of A A.8. and the vectors ui and vj are. From the input/output point of view. This can be illustrated by the following 2 2 matrix: cos 2 .1) (n. 7 = ( 2 + w w) 6.. while u1 (um ) is the highest (lowest) gain observing direction. 2 The i is the ith singular value of A.1 and B 2 F(m. the ith left singular vector and the j th right singular vector. sin 2 : 1 A = cos 1 .7 4 5 0 2 2 + w w. Since 2132 607 A1 6 . v1 (vn ) is the highest (lowest) gain input direction. But since = kAk = kA1 k. It is easy to verify that Avi = A ui = The above equations can also be written as i ui i vi : i vi 2 i ui : 2 A Avi = AA ui = Hence i2 is an eigenvalue of AA or A A. Singular Value Decomposition 33 where w 2 Fn. and vi is an The following notations for singular values are often adopted: = the largest singular value of A (A) = min (A) = p = the smallest singular value of A : Geometrically. we must have w = 0.
1 U . (A). ( ): kxk=1 = Hence . in the above proof. (A)j ( ) (A) ( ) (iii) (A. The other inequality (A + ) .1 ) = 1 if A is invertible. Then (i) j (A + ) . then A.1 ) = 1= ( ) = 1= (A).12 Suppose A and are square matrices. (A) (ii) (A ) Proof. (A) follows by replacing A by A + and by . k xkg kxk=1 kxk=1 min kAxk . (i) By de nition (A + ) := kmin k(A + )xk xk=1 min fkAxk .1 = V . Hence (A. . (ii) This follows by noting that (A ) := kmin kA xk x rk=1 min x A A x = kxk=1 = (A) kmin k xk xk=1 (A) ( ): ( ) (iii) Let the singular value decomposition of A be A = U V . Hence it is often convenient to introduce the following alternative de nitions for the largest singular value : (A) := kmax kAxk xk=1 and for the smallest singular value of a tall matrix: (A) := kmin kAxk : xk=1 Lemma 2.1 ) = ( .34 LINEAR ALGEBRA 1 It is easy to see that A maps a unit disk to an ellipsoid with semiaxes of and 2 . ( ) (A + ) . max k xk (A) .
B +1 ^ ^ where B = Ik+1 0 U BV Ik0 2 F(k+1) (k+1) and rank(B ) k. Singular Value Decomposition Some useful properties of SVD are collected in the following lemma. U BV ) Ik0 rank(B) k = ^ k+1 . U BV k +1 Ik+1 0 ( . we only need to show that min kA . 8. Lemma 2. Then kA . A 2 Fm n has a dyadic expansion: 35 2 A= 5. Ak k = k+1 . Let x 2 Fk+1 ^ be such that Bx = 0 and kxk = 1. B k ( k+1 . r X i=1 i ui vi = Ur r Vr where Ur = u1 : : : ur ]. Vr = v1 : : : vr ]. B k = k . 7.2. rank(A) = r 2. 6. the conclusion follows. Ak k = k+1 : rank(B) k Proof. ImA = spanfu1 : : : ur g and (ImA)? = spanfur+1 : : : um g 4. Let B be any matrix such that rank(B ) k. Hence.8. and r = diag ( 1 : : : r ) 2 2 2 kAk2 = 1 + 2 + + r F kAk = 1 i (U0 AV0 ) = i (A) i = 1 : : : p for any appropriately dimensioned unitary matrices U0 and V0 P Let k < r = rank(A) and Ak := k=1 i ui vi . 2 . B Since B is arbitrary. It is easy to see that rank(Ak ) k and kA . B k = kU V . Then ^ ^ kA . B )x = k k+1 xk k+1 : k+1 . B k k+1 .13 Let A 2 Fm n and 1 2 r > r+1 = = 0 r minfm ng: Then 1. B k = kA . then i min kA . We shall only give a proof for part 8. KerA = spanfvr+1 : : : vn g and (KerA)? = spanfv1 : : : vr g 3.
furthermore.1 . A has a right inverse i A has full row rank. ? More generally. It can be shown that pseudoinverse is unique. Obviously. and. if Y A = I then Y is called a left inverse of A. one of the right inverses is given by X = A (AA ). Suppose m k and B B = C C . Suppose A has a singular value decomposition A=U V with 0 = 0r 0 r > 0: Then A+ = V +U with += . By duality. if a matrix A has neither full row rank nor full column rank.1 A . and. denoted by A > 0 ( 0). Lemma 2. if x Ax > 0 ( 0) for all x 6= 0. in that case. For example. Suppose A 2 Fn n and A = A 0. known also as the MoorePenrose inverse. the so called pseudoinverse. A matrix X 2 C n m is said to be a right inverse of A if AX = I . This pseudoinverse is denoted by A+ . Similarly.9 Generalized Inverses Let A 2 C m n .1 0 r 0 0 : 2. one of the left inverses is Y = (A A). which satis es the following conditions: (i) AA+ A = A (ii) A+ AA+ = A+ (iii) (AA+ ) = AA+ (iv) (A+ A) = A+ A. Then A+ = C (CC ). then there exists a B 2 Fn r with r rank(A) such that A = BB .36 LINEAR ALGEBRA 2. any matrix in the form I is a right inverse of I 0 .1 (B B ).1 B : Another way to compute A+ is by using SVD.14 Let B 2 Fm n and C 2 Fk n . Note that right (or left) inverses are not necessarily unique. Then there exists a matrix U 2 Fm k such that U U = I and B = UC . then all the ordinary matrix inverses do not exist however. .10 Semide nite Matrices A square hermitian matrix A = A is said to be positive de nite (semide nite) . One way of computing A+ is by writing A = BC so that B has full column rank and C has full row rank. is useful. A has a left inverse i A has full column rank.
1=2 (BA. then A1=2 = U 1=2 U where p p = diagf 1 : : : n g 1=2 = diagf 1 : : : n g: Lemma 2. Let V1 and V2 be unitary matrices such that B1 = V1 B1 0 C1 = V2 C1 0 37 where B1 and C1 are full row rank. Since A > 0. 2 by We can de ne square root for a positive semide nite matrix A. Then B1 and C1 have the same number of rows and V3 := B1 C1 (C1 C1 ). the conclusion follows by the fact that 0 < I . A1=2 = (A1=2 ) 0.1=2 ). Consequently.1=2 BA.1=2 ) < 1 i (BA. if X22 is the pseudoinverse of X22 .1 ) < 1.15 Suppose A = A > 0 and B = B 0.1=2 BA.1 )A1=2 : Lemma 2. X12 X22 X12 0 be partitioned as + X11 X12 = I X12 X22 X12 X22 0 I 0 X22 0 I 0 + X22 X12 I : .1=2 = I .1 ) = i (A.1 are similar. A.2. Therefore. 2 Proof. we have A > B i 0 < I . Semide nite Matrices Proof.1=2 i (A.16 Let X = X Then KerX22 + X12 X22 solves and X = X11 X12 : X12 X22 + KerX12 . Hence V3 is a unitary matrix and V3 B1 = C1 .1=2 BA.1 satis es V3 V3 = I since B B = C C . A. Clearly.1 ) < 1. and hence i (BA.1=2 BA. Then A > B i (BA. then Y = Y X22 = X12 + X11 . Finally let 0 U = V1 V3 V V2 0 4 for any suitably dimensioned V4 such that V4 V4 = I . A1=2 can be computed by using spectral decomposition or SVD: let A = U U .1=2 and BA.1=2 BA.10. A. A. A = A1=2 A1=2 : However.
C . Ker X22 Ker X12 and now + X12 X22 X22 = X12 U1 U1 = X12 U1 U1 + X12 U2 U2 = X12 : The factorization follows easily. assume U1 U2 with and Moreover 1= 1 > 0 and U = U1 U2 unitary. Then Ker X22 = spanfcolumns of U2 g + X22 = U1 U2 0 . Without loss of generality. The matrix X B is a dilation of its submatrices as indicated in the following C A diagram: X B C A d d c c d c d B A 6 ? 6 A ? c C A  . we consider the following induced 2norm optimization problem: min X X B C A (2:3) where X . B .1 0 1 0 U1 : U2 I 0 X11 X12 I 0 0 0 U X12 X22 0 U gives X12 U2 = 0.38 LINEAR ALGEBRA X22 = U1 U2 1 0 0 0 Proof. 2 2.11 Matrix Dilation Problems* In this section. Hence. and A are constant matrices of compatible dimensions.
A A): h i+ I . Theorem 2. The simplest matrix dilation problem occurs when solving min X X A : (2:4) Although (2. A A)1=2 : Proof.3). we will see that it contains all the essential features of the general problem.4). in general. Compression is always norm nonincreasing and dilation is always norm nondecreasing. \c" stands for the operation of compression and \d" stands for dilation. Norm preserving dilations are the focus of this section. Sometimes dilation can be made to be norm preserving. X A i i X X +A A 2I X X ( 2 I . A A) and let Y := X ( 2 I . A A): Now suppose X X ( 2 I .17 8 0.11.4). i there is a Y with kY k 1 such that X A X = Y ( 2 I . Similarly if X = Y ( 2 I .4) is a much simpli ed version of (2. A A)1=2 and Y Y Y Y I then X X ( 2 I .2. A A)1=2 and 2 This theorem implies that. it is immediate that 0 = kAk : The following theorem characterizes all solutions to (2. Letting 0 denote the minimum norm in (2. (2. which is in contrast to the minimization in the Frobenius norm in which X = 0 is the unique solution. The solution X = 0 is the central solution but others are possible unless 2 A A = 0I. A A)1=2 then X = Y ( 2 I .4) has more than one solution. Matrix Dilation Problems* 39 In this diagram. .
Theorem 2.1=2 X (< ) 1 (< 1): Now. returning to the problem in (2.18 8 > 0 .1=2 (< ) 1(< 1): The corresponding dual results are Theorem 2. The proof is the straightforward application of Theorem 2.1 The theorem still holds if ( 2I .3). A A). kY k 1.19. i X A X ( 2 I .19 8 0 X A X = ( 2 I .21 (Parrott's Theorem) The minimum in (2. A A)1=2 is replaced by any matrix R such that 2 I .40 LINEAR ALGEBRA Remark 2. A A = R R. such that Corollary 2.20 8 > i 0 X A ( 2 I . AA )1=2 Y: i there is a Y . Corollary 2.17 and its dual. ~ A more restricted version of the above theorem is shown in the following corollary. Theorem 2. let 0 := min X X B C A : (2:5) The following so called Parrott's theorem will play an important role in many control related optimization problems. AA ).5) is given by 0 = max C A B A : (2:6) .
Y A Z Y (^2 I . AA )1=2 Z C A (^ A Y . A A)1=2 for some Y such that kY k 1. 2 )V m 2 2 ^2 I . AA )1=2 . AA )1=2 := U (^ I . A A)1=2 A 2 = ^0I ^0I : 2 Now we are ready to show that 0 ^. 2 )1=2 V m 2 2 1=2 0 (^2 I . 41 0 ^ since compressions are norm nonincreasing. Then ^ .11.A (^2 I . we can show that (^2 I .17 we have that B = Y (^2 I .A (^2 I .m (^2 I .A (^ (^2 I . From Theorem 2.0 m ) ^In.2. A A = V (^2 I . Denote by ^ the right hand side of the equation (2. A A)1=2 X B = 2 I . Similarly.19. AA )1=2 : Using this equality. Clearly. A A)1=2 A = A (^2 I . Matrix Dilation Problems* Proof. A A)1=2 A (^2 I .Y A Z . AA = U ^ I . Then A has the following singular value decomposition: A=U 0 m V n. Now let X = . AA )1=2 I I (^ A 2 I . Theorem 2. m ^2 I0 U: 0 n.A (^2 I .6). AA )1=2 Z for some Z with ^ kZ k 1.m m Hence and Now let and U 2 C n n V 2 C m m: ^2 I . Suppose A 2 C n m and n m (the case for m > n can be shown in the same fashion). A A)1=2 Z = 2 I . A A)1=2 . AA )1=2 .19 yields C = (^2 I . A A)1=2 := V (^2 I .m U : Then it is easy to verify that (^2 I . and that 0 ^ will be shown by using Theorem 2.17 and Theorem 2. AA )1=2 A = ^: .
3).9) (2.22 considers the problem of parameterizing all solutions. . kW k 1.A Z ( 2 I . A A)1=2 : Then kY k 1 and the theorem follows by noting that W1 Y 1 i there is a 1=2 W . C A C A Z Z 1=2 0 (I .3) and an expression for 0 . A A)1=2 .Y A Z is the \central" solution analogous to X = 0 in (2. A A)1=2 : Now apply Theorem 2. As in (2.10) 1 such Proof.Y A Z + W1 (I . Z Z )1=2 . and Y with kY k kZ k 1 solve the linear equations B = Y ( 2 I .22 when > 0 .A Z ( 2I . Z Z )1=2 where W is an arbitrary contraction (kW k 1). W 1. Y Y ) 2 The following corollary gives an alternative version of Theorem 2.22 Suppose 0 . again from Theorem 2. LINEAR ALGEBRA so ^ = 0 .4). W . Theorem 2.42 Thus ^ 0. ( . Partition W as W = W1 Y to obtain the expression for X and B : X = . A A)1=2 C = ( 2 I .21 exhibits only one. AA )1=2 Z: Note that using the above expression for C we have 2I . Z Z )1=2 B = Y ( 2 I . Since that 0.7) with respect to the partitioned matrix X B to get C A ^ X B =W (I .17 (Remark 2. A 0 A)1=2 ^ ^ ^ ^ for some contraction W . The solutions X such that are exactly those of the form X = .4).1) to inequality (2. there may be more than one solution to (2. Theorem 2. A Z ) 2 I . The ^ solution X = .19 there exists a Z with kZ k C = ( 2 I . Y Y )1=2 W (I . such that W1 = (I .Y A Z + (I . Z Z )1=2 0 = (I . AA )1=2 Z: X B C A (2:7) (2:8) 1 and Z with (2. although the proof of theorem 2. 2 This theorem gives one solution to (2.
and C = U?RV .14) Note that in the case of > 0 . although the formulas given above seem to be the simplest. Then 0 = min kR + UQV k Q = min Q = min Q V U U? (R + UQV ) V ? U RV + Q U RV? : U?RV U? RV? The result follows after applying Theorem 2.1=2 (2. Let U? and V? be such that U U? and V? are unitary matrices. Z Z ). 2 . I . Matrix Dilation Problems* Corollary 2.11. A similar problem arises in H1 control theory.23 For > 0.13) 2 I . and V are constant matrices such that U U = I and V V = I . Z Z are invertible since Corollary 2.2. Y Y and I . U .20 clearly show that kY k < 1 and kZ k < 1. consider the following matrix approximation problem: 0 = min kR + UQV k Q (2:15) where R. There are many alternative characterizations of solutions to (2.1=2 Y = B ( 2 I .22 with X = Q + U RV . A = U? RV? .21 and 2.1=2 C: Z = ( (2. Corollary 2. Y Y ).22.1=2 (X + Y A Z )(I .24 The minimum achievable norm is given by 0 = max fkU? Rk kRV? kg and the parameterization for all optimal solutions Q can be obtained from Theorem 2.18 and 2. V Proof. B = U RV? . As a straightforward application of the dilation results obtained above. X B C A i where 43 (< ) (< ) (2:11) (2:12) (I . A A). AA ).11).
. Horn and Johnson 1990.1991] and Lancaster and Tismenetsky 1985]. Golub and Van Loan's book 1983] contains many numerical algorithms for solving most of the problems in this chapter. Kahan.12 Notes and References A very extensive treatment of most topics in this chapter can be found in Brogan 1991]. and Weinberger 1982].44 LINEAR ALGEBRA 2. The matrix dilation theory can be found in Davis.
pole placement. stabilizability. are studied in some detail. System interconnections and realizations. 3 3. The notions of controllability.2) where x(t) 2 Rn is called the system state. and y(t) 2 Rp is the system output. Kalman canonical decomposition. observability. in particular the balanced realization. otherwise it is called MIMO (multiple input and multiple 45 . A dynamical system with single input (m = 1) and single output (p = 1) is called a SISO (single input and single output) system. The A B C and D are appropriately dimensioned real constant matrices. The solutions of Lyapunov equations and their connections with system stability. the concepts of system poles and zeros are introduced. controllability. are discussed.1) (3. and detectability are de ned and various algebraic and geometric characterizations of these notions are summarized. Finally.Linear Dynamical Systems This chapter reviews some basic system theoretical concepts. and observer theory are then introduced. x(t0 ) is called the initial condition of the system. u(t) 2 Rm is called the system input.1 Descriptions of Linear Dynamical Systems Let a nite dimensional linear time invariant (FDLTI) dynamical system be described by the following linear constant coe cient di erential equations: x = Ax + Bu x(t0 ) = x0 _ y = Cx + Du (3. and so on.
Other reasons for using this notation will be discussed in Chapter 10.1 C D := C (sI . The impulse matrix of the dynamical system is de ned as g(t) = L. we can assume without loss of generality that t0 = 0. Note that is a real block matrix.1 B + D: Note that the system equations (3. we have G(s) = C (sI . and thus (t t1 ) is usually called the state transition matrix. The corresponding transfer matrix from u to y is de ned as Y (s) = G(s)U (s) where U (s) and Y (s) are the Laplace transform of u(t) and y(t) with zero initial condition (x(0) = 0).4) In the case of u(t) = 0 8t t0 . the dynamical system response x(t) and y(t) for t t0 can be determined from the following formulas: A B C D x(t) = eA(t. it is easy to see from the solution that for any t1 t0 and t t0 . Now given the initial condition x(t0 ) and the input u(t). the matrix function (t t1 ) = eA(t.46 LINEAR DYNAMICAL SYSTEMS output) system. not a transfer function. This will be assumed in the sequel. )Bu( )d (3.t0 ) x(t0 ) + eA(t. the notation A B .t1 ) x(t1 ): Therefore.t1 ) acts as a transformation from one state to another.1 fG(s)g = CeAt B 1+ (t) + D (t) where (t) is the unit impulse and 1+ (t) is the unit step de ned as 1+ (t) := 1 t < 0: 0 t 0 . A) B + D will be used. we have x(t) = eA(t.3) t0 y(t) = Cx(t) + Du(t): (3. A).1) and (3. Hence. Since Zt the state of a linear system at one time can be obtained from the state at another through the transition matrix.2) can be written in a more compact matrix form: x = A B _ x : y C D u To expedite calculations involving transfer matrices.
1 g(t .e.1 The dynamical system described by the equation (3.1 g(t . C = B AB A2 B : : : An. with zero initial state: x0 = 0) can be described by the convolution equation y(t) = (g u)(t) := Z1 . i (iv) The matrix A .. (ii) The matrix Wc (t) := is positive de nite for any t > 0. .2 Controllability and Observability We now turn to some very important concepts in linear system theory. hA jImB i := n=1 Im(Ai. t1 > 0 and nal state x1 . I B ] has full row rank for all in C .3.e. i. Controllability and Observability 47 The input/output relationship (i.1 The following are equivalent: (i) (A B ) is controllable. the system or the pair (A B ) is said to be uncontrollable.2. x A = x .1) or the pair The controllability (or observability introduced next) of a system can be veri ed through some algebraic or geometric criteria. De nition 3. (iii) The controllability matrix Zt 0 eA BB eA d P has full row rank or.. then x B 6= 0. Otherwise. (A B ) is said to be controllable if. for any initial state x(0) = x0 . (v) Let and x be any eigenvalue and any corresponding left eigenvector of A. there exists a (piecewise continuous) input u( ) such that the solution of (3.1 B (vi) The eigenvalues of A + BF can be freely assigned (with the restriction that complex eigenvalues are in conjugate pairs) by a suitable choice of F . )u( )d : 3.1) satis es x(t1 ) = x1 .1 B ) = Rn . in other words. )u( )d = Zt . Theorem 3.
Therefore. then v = 0. Wc (t1 ) > 0) but the controllability matrix C does not have full row rank. )Wc (t1 ).B eA (t1 . Then there exists a 0 6= v 2 Rn such that v eAt B = 0 for all t 2 0 t1 ]. v eAt B = 0 for all t or. This is again a contradiction. x1 ): Then it is easy to verify using the formula in (3. for some t1 . we have v B AB A2 B : : : An. (iii): Suppose Wc (t) > 0 for all t > 0 (in fact. v Wc (t) = 0 for all t this is a contradiction. equivalently. . )Bu( )d : Premultiply the above equation by v to get 0 = v eAt1 x(0): If we chose the initial state x(0) = e.1 (eAt1 x0 . and let the input be de ned as u( ) = .At1 v. and this is a contradiction.3) that x(t1 ) = x1 . To show that the controllability of (A B ) implies that Wc (t) > 0 for any t > 0. (i) . Hence. there exists a real vector 0 6= v 2 Rn such that v eAtB = 0 t 2 0 t1 ]: Now let x(t1 ) = x1 = 0. we have 0 = eAt1 x(0) + Zt 0 1 eA(t1 . suppose C has full row rank but Wc (t) is singular for some t1 . Wc (t) can not be singular for any t > 0.3). Since eAtBB eA t 0 for all t. and then from the solution (3. Then there exists a v 2 Rn such that v Ai B = 0 for all 0 i n . (ii) . 1.48 LINEAR DYNAMICAL SYSTEMS Proof. Hence. the controllability matrix C must be full row rank. assume that (A B ) is controllable but Wc (t1 ) is singular for some t1 > 0. Since x1 is arbitrary. and hence. set t = 0. the pair (A B ) is controllable. it can be shown that Wc (t) > 0 for all t > 0 i . and we have v B = 0: Next. evaluate the ith derivative of v eAt B = 0 at t = 0 to get v Ai B = 0 i > 0: Hence. (ii): Suppose Wc (t1 ) > 0 for some t1 > 0. In fact. Conversely. the controllability matrix C does not have full row rank. this equality holds for all i 0 by the CayleyHamilton Theorem.1 B = 0 or. in other words.
x A = x and x B = 0. we will explicitly construct a matrix F so that the eigenvalues of A + BF are in the desired locations. i.. (vi) ) (i): This follows the same arguments as in the proof of (v) ) (iii): assume that (vi) holds but (A B ) is uncontrollable. with Ac 2 R(n.2 An unforced dynamical system x = Ax is said to be stable if all the _ eigenvalues of A are in the open left half plane. . xc Ac = 1 xc . c 2 De nition 3.2. Assume that (v) holds but rank C = k < n. Then in section 3.. I B does not have full row rank for some 2 C .e.k) .e. (v) ) (iii): We will again prove this by contradiction. Re (A) < 0.4.3..e. Then.1 x B =0 i.1 TB ) is not controllable. this will result in x B AB : : : An. I B =0 i. Then x (TB ) = 0 and 0 x= x c is an eigenvector of TAT .e. the proof is completed. but this contradicts the condition (vi). (iv) ) (v): This is obvious from the proof of (iii) ) (iv). that the matrix 49 A.. Controllability and Observability (iii) ) (iv): Suppose. the controllability matrix C does not have full row rank.1 = Ac A12 TB = Bc 0 A 0 Let 1 and xc be any eigenvalue and any corresponding left eigenvector of Ac . and this is a contradiction. on the contrary. This is a contradiction since similarity transformation does not change controllability. Then there exists a vector x 2 C n such that x A.1 corresponding to the eigenvalue 1 . i.k) (n.3. Hence. there is a decomposition so that some subsystems are not a ected by the control. In that section. (i) ) (vi): This will be clear in section 3. A matrix A with such a property is said to be stable or Hurwitz. which implies that (TAT . However.1 B = x B x B ::: n. we will show that there is a transformation T such that TAT .
De nition 3. i.. the initial state x(0) = x0 can be determined from the time history of the input u(t) and the output y(t) in the interval of 0 t1].1 has full column rank for all in C . or (C A). (ii) The matrix Wo (t) := is positive de nite for any t > 0. Therefore. 7 4 5 CAn. it is more appropriate to call this stabilizability the state feedback stabilizability to di erentiate it from the output feedback stabilizability de ned later. T has full column rank or n=1 Ker(CAi. Theorem 3.1).4 The dynamical system described by the equations (3. We now consider the dual notions of observability and detectability of the system described by equations (3. Otherwise. 0.2).1 ) = 0. x B 6= 0. i (iv) The matrix A. the system.e. The following theorem is a consequence of Theorem 3.1) and (3..1.3 The dynamical system (3. or the pair (A B). I C . is said to be stabilizable if there exists a state feedback u = Fx such that the system is stable. 7 6 .2) or by the pair (C A) is said to be observable if. (ii) The matrix A . Theorem 3. 0. (iii) The observability matrix Zt 0 eA C CeA d 2 C 3 6 CA 7 6 7 O = 6 CA2 7 6 . I B ] has full row rank for all Re (iii) For all and x such that x A = x and Re (iv) There exists a matrix F such that A + BF is Hurwitz. A + BF is stable.2 The following are equivalent: (i) (A B ) is stabilizable. is said to be unobservable.50 LINEAR DYNAMICAL SYSTEMS De nition 3. for any t1 > 0.3 The following are equivalent: (i) (C A) is observable.1) and (3.
e. (i) ( (iii): Note that given the input u(t) and the initial condition x0 . 7 4 5 (n.. the output in the time interval 0 t1] is given by y(t) = CeAt x(0) + Zt 0 CeA(t. ) Bu( )d + Du(t): Since y(t) and u(t) are known. 2 De nition 3. there is a vector x0 such that Ox0 = 0 or equivalently CAi x0 = 0 8i 0 by the CayleyHamilton Theorem. then Cy 6= 0. i. Proof.1 where y(i) stands for the ith derivative of y. or the pair (C A). Now suppose the initial state x(0) = x0 . (vii) (A C ) is controllable. Since the observability matrix O has full column rank. then y(t) = CeAt x(0) = 0.e. This implies that the system is not observable since x(0) cannot be determined from y(t) 0. we will show the equivalence between conditions (i) and (iii).... i. Controllability and Observability 51 (v) Let and y be any eigenvalue and any corresponding right eigenvector of A. 6 . 4 3 2 C 3 CA 7 7 6 CA2 7 7=6 6 7 6 . Hence. y(t) = CeAt x(0) t 2 0 t1 ]: From this equation. (vi) The eigenvalues of A + LC can be freely assigned (with the restriction that complex eigenvalues are in conjugate pairs) by a suitable choice of L. First. there is no loss of generality in assuming u(t) = 0 8t. Assume that (C A) is observable but that the observability matrix does not have full column rank. there is a unique solution x(0) in the above equation. Ay = y.5 The system.1) (0) y CAn. .3. (i) ) (iii): This will be proven by contradiction. Once this is done. is detectable if A + LC is stable for some L. the rest will follow by the duality or condition (vii). This completes the proof. we have 2 y(0) 6 y_ (0) 6 . 7 x(0) 7 5 6 .2.
Similarly. It follows that a system is controllable (observable) if and only if every mode is controllable (observable). the following de nitions of modal controllability and observability are often useful. i. if the B matrix is changed into a 4 2 matrix . (iv) There exists a matrix L such that A + LC is Hurwitz. a mode of the system. For example. However. Otherwise.1 and Theorem 3. A . In particular. and as a result. Note that if 1 = 2 . Then the mode is said to be controllable (observable) if x B 6= 0 (Cx 6= 0) for all left (right) eigenvectors of A associated with . consider the following 4th order system: 2 6 A B =6 6 6 C D 4 0 0 0 1 1 1 0 0 0 1 0 1 0 0 1 0 0 0 1 0 2 1 0 3 7 7 7 7 5 with 1 6= 2 . Cx 6= 0. x2 . (ii) The matrix The conditions (iv) and (v) of Theorem 3. and 2 is not observable if = 0.52 LINEAR DYNAMICAL SYSTEMS Theorem 3. Hence any linear combination of x1 and x2 is still an eigenvector of A corresponding to 1 . In particular. a system is stabilizable (detectable) if and only if every unstable mode is controllable (observable). Then. the system is uncontrollable and unobservable for any and since in that case. equivalently. Similar arguments can be applied to check observability.2 and Theorem 3. let x = x1 . then x B = 0.4 The following are equivalent: (i) (C A) is detectable.4 are often called PopovBelevitchHautus (PBH) tests.. the mode is said to be uncontrollable (unobservable). (v) (A C ) is stabilizable. De nition 3.6 Let be an eigenvalue of A or.e.3 and the conditions (ii) and (iii) of Theorem 3. the system is not controllable. x A = x (Ax = x) and 0 6= x 2 C n . I has full column rank for all Re 0. the mode 1 is not controllable if = 0. C (iii) For all and x such that Ax = x and Re 0. both 203 6 7 x1 = 6 0 7 415 0 and 203 6 7 x2 = 6 0 7 405 1 are the left eigenvectors of A corresponding to 1 .
1 TB + D: In this section. we will consider the system structure decomposition using coordinate transformation if the system is not completely controllable and/or is not completely observable. for a system given in the Jordan canonical form. then the system is controllable even if 1 = 2 . A).3.e. the controllability and observability can be concluded by inspection.3. G(s) = C (sI .2) become _ x = TAT .1 (sI . consider a simple pendulum.1 B + D = CT . On the other hand. For example. the reader may easily verify that the system with 20 03 6 7 B=6 1 0 7 4 15 1 0 is controllable for any . the motion of the pendulum can be uniquely determined either in terms of the angle of the string attached to the pendulum or in terms of the vertical displacement of the pendulum. The interested reader may easily derive some explicit conditions for the controllability and observability by using Jordan canonical form and the tests (iv) and (v) of Theorem 3.1x + TBu y = CT . TAT .. we can regard these representations as equivalent.1 x + Du: These equations represent the same dynamical system for any nonsingular matrix T . i. To begin with. although some coordinates may not be natural descriptions of a physical dynamical system. For example.1 and Theorem 3. It is easy to see that the input/output transfer matrix is not changed under the coordinate transformation. However.1) and (3. the angular displacement is a more natural description than the vertical displacement in spite of the fact that they both describe the same dynamical system. This is true for most physical dynamical systems.1 TB : C D CT . and hence. In general. In general.1 D C D " # . in most cases. let T 2 Rn n be a nonsingular matrix and de ne x = Tx: Then the original dynamical system equations (3.! A B = TAT . 3. Kalman Canonical Decomposition 53 with the last two rows independent of each other. they may make the system analysis and synthesis much easier.3.1). let us consider further the dynamical systems related by a similarity transformation: A B 7.3 Kalman Canonical Decomposition There are usually many di erent coordinate systems to describe a dynamical system.
1 = = Aq1 q1 = T . we can now show the following theorem.6 If the controllability matrix C has rank k1 < n. G(s) = C (sI . the following theorem follows easily from the PBH tests or from the above relations. A).1 B + D = Cc (sI .1 : The controllability and observability matrices are related by Moreover. Now let q1 q2 : : : qk1 be any linearly independent columns of C . Moreover.1 Bc + D: Proof. Ac ).1 Aqk1 Aqk1 +1 Aqn qk1 qk1 +1 qn Ac A12 0 Ac Ac A12 0 Ac . Therefore. Aqi can be written as a linear combination of qi i = 1 2 : : : k1 since Aqi is a linear combination of the columns of C by the CayleyHamilton Theorem.5 The controllability (or stabilizability) and observability (or detectability) are invariant under similarity transformations.1 : qn Then the transformation x = Tx will give the desired decomposition. note that for each i = 1 2 : : : k1 . the pair (A B) is not controllable. Theorem 3. we have AT .54 LINEAR DYNAMICAL SYSTEMS C = T C O = O T . Theorem 3. Let qi i = k1 +1 : : : n be any n . k1 linearly independent vectors such that the matrix Q := q1 is nonsingular. then there exists a similarity transformation x = xc = Tx x such that c _ xc _ xc = Ac A12 0 Ac Cc Cc y = xc + Du xc xc + Bc u xc 0 with Ac 2 C k1 k1 and (Ac Bc) controllable. To see that. Since rank C = k1 < n. Using this fact. De ne qk1 qk1 +1 T := Q.
e. Kalman Canonical Decomposition 55 for some k1 k1 matrix Ac . the state space x is partitioned into two orthogonal subspaces xc 0 and 0 xc TAT . By duality. each column of the matrix B is a linear combination of qi i = 1 2 : : : k1 . note that rank C = k1 and 0 Since. hence B = Q Bc = T .e.1 Bc c for some Bc 2 C k1 m . we have rank Bc Ac Bc Ak1 .3.. for each j k1 . (Ac Bc ) controllable and with Ac stable. we have x = T . On the other hand. equivalently. To show that (Ac Bc) is controllable. if the controllability matrix C has the QR factorization QR = C . we have the following decomposition if the system is not completely observable. xc : xc .3. Im C . Write these subspaces in terms of the original coordinates x. the state xc are not a ected by the control u). the uncontrollable subspace is given by the complement of the controllable subspace. Similarly.7 If the system is stabilizable and the controllability matrix C has rank k1 < n. (Ac Bc ) is controllable.1 Bc : c 0 Cc Cc D with Ac 2 C k1 k1 . then there exists a similarity transformation T such that C = T . then T = Q.1 .1 Bc = k1 c i. For example. Corollary 3.1 3 2 Ac A12 Bc TB = 6 0 Ac 0 7 5 4 D with the rst subspace controllable from the input and second completely uncontrollable from the input (i. Hence..1 Bc 0 0 Ak1 . 2 A numerically reliable way to nd a such transformation T is to use QR factorization. Aj is a linear combination of Aic i = 0 1 : : : (k1 .1 x = q1 qk1 qk1 +1 qn So the controllable subspace is the span of qi i = 1 : : : k1 or.1 CT .1 Bc Ac0Bc 0 An. 1) by Cayleyc Hamilton theorem.
1 2 Aco 0 6 A21 A TB = 6 0 0co 6 6 0 0 D 4 Cco 0 A13 0 Bco A23 A24 Bco Aco 0 0 A43 Aco 0 Cco 0 D 3 7 7 7 7 5 . A).1 D 2 3 Ao 0 Bo = 6 A21 Ao Bo 7 4 5 Co 0 D with Ao 2 C k2 k2 .1 Bo + D: Carefully combining the above two theorems. Theorem 3.1 CT .1 3 2 Ao 0 Bo TB = 6 A21 Ao Bo 7 5 4 D Co 0 D with Ao 2 C k2 k2 and (Co Ao ) observable. Then there exists a nonsingular coordinate transformation x = Tx such that 2x 3 2 A 0 A 0 32 x 3 2 B 3 _ co co 6 xco 7 = 6 A21 Aco A13 A24 7 6 xco 7 + 6 Bco 7 u _ 7 6x 5 6 0 0 A23 0 7 6 xco 7 6 0co 7 4 _ co 4 5 4 co 5 4 5 co _ co x 0 0 A43 Aco xco 0 y = or equivalently Cco 0 Cco 0 2x 6 xco 6 xco 4 xco co 3 7 + Du 7 5 TAT .1 B + D = Co (sI . (Co Ao ) observable and with Ao stable. then there exists a similarity transformation T such that TAT . Ao ). then there exists a simiTAT . The proof is left to the reader as an exercise.1 TB CT .10 Let an LTI dynamical system be described by the equations (3.9 If the system is detectable and the observability matrix C has rank k2 < n. Similarly.1) and (3.2). we get the following Kalman Canonical Decomposition.8 If the observability matrix O has rank k2 < n. we have G(s) = C (sI . Corollary 3.56 larity transformation T such that LINEAR DYNAMICAL SYSTEMS Theorem 3.1 CT .
x(0) = 0. then the output Zt y(t) = Cco eAco(t. In other words.1 Bco + D: One important issue is that although the transfer matrix of a dynamical system A B C D is equal to its controllable and observable part " Aco Bco Cco D # their internal behaviors are very di erent.e. Moreover. xco controllable but unobservable. the transfer matrix from u to y is given by G(s) = Cco (sI . The solution to this system is given by 2 x (t) 3 6 xco(t) 7 6 xco(t) 7 4 5 co xco (t) = eAco t xco (0) y(t) = Cco xco (t) + Cco xco 3 2 Acot R e xco (0) + R0t eAco (t. This can be illustrated by the state space response for the simple system 2x 3 2 A 0 0 0 32 x 3 2 B 3 _ co 6 xco 7 = 6 0co Aco 0 0 7 6 xco 7 + 6 Bco 7 u _ 6x 7 6 0 0 A 0 7 6 xco 7 6 0co 7 4 _ co 5 4 5 4 co 5 4 5 co _ co x 0 0 0 Aco xco 0 y = Cco 0 Cco 0 2x 6 xco 6 xco 4 xco co 3 7 7 5 with xco controllable and observable.. while xco (t) and xco (t) do not show up in the output y. and xco uncontrollable and unobservable. )Bco u( )d 6 eAcotxco(0) + t eAco(t. xco observable but uncontrollable. if the initial condition is zero. Kalman Canonical Decomposition 57 where the vector xco is controllable and observable.3. xco is observable but uncontrollable. xco is controllable but unobservable. and xco is uncontrollable and unobservable. Moreover. their state space response with nonzero initial conditions are very di erent. Aco ). ) Bcou( )d : 0 .3. i. )Bcou( )d 7 7 6 0 7 6 4 5 eAco t xco (0) note that xco (t) and xco (t) are not a ected by the input u. while their input/output behaviors are the same.
since it is impossible to make the initial states exactly zero. .7. 3. then the output y(t) will grow without bound.1: State Feedback Then we have the following lemma in which the proof is simple and is left as an exercise to the reader. and the closedloop system equations are given by x = (A + BF )x + Bv _ y = (C + DF )x + Dv: D y ? e C x R  6 e B u e v 6 A F Figure 3. if Aco is not stable.1. On the other hand. we will consider how to place the system poles to achieve desired closedloop behavior if the system is controllable. In the next section. then the response xco (t) will show up in the output. any unbounded internal signal will eventually destroy the system. Since the states are the internal signals of the dynamical system. any uncontrollable and/or unobservable unstable mode will result in unacceptable system behavior. This issue will be exploited further in section 3. The problems with uncontrollable and/or unobservable unstable modes are even more profound than what we have mentioned.4 Pole Placement and Canonical Forms Consider a MIMO dynamical system described by x = Ax + Bu _ y = Cx + Du and let u be a state feedback control law given by u = Fx + v: This closedloop system is as shown in Figure 3. In particular. if the initial state is not zero.58 LINEAR DYNAMICAL SYSTEMS However.
the reader will nd that this representation is much more visually appealing than the traditional representation. the observability of the system may change under state feedback. signals owing from left to right.! C D By duality.3. and is consistent with the matrix manipulations. the following system 2 3 A B =4 1 1 1 5 1 0 0 C D 1 0 0 is controllable and observable.4.1 . i. But with the state feedback Lemma 3.11 Let F be a constant matrix with appropriate dimension then (A B) is u = Fx = .1 x the system becomes and is not completely observable. Pole Placement and Canonical Forms 59 controllable (stabilizable) if and only if (A + BF B ) is controllable (stabilizable).! x = Ax + Bu + Ly _ _ is called output injection which can be written as A + LC B + LD : A B C D 7.1 We would like to call attention to the diagram and the signals ow con vention used in this book. will also be used in this book. however. the following diagram represents the multiplication of three matrices and will be very helpful in dealing with complicated systems: z = M1 M2 M3 w: z M1 M2 M3 w The conventional signal block diagrams. It may not be conventional to let signals ow from the right to the left. For example. ~ . However.. The dual operation of the dynamical system by 20 A + BF B 4 C + DF D = 1 1 0 1 0 0 0 0 3 5 x = Ax + Bu 7. the output injection does not change the system observability (detectability) but may change the system controllability (stabilizability).e. For example. Remark 3.
Let det( I . The quadruple (A b C d) is called a state space realization of the transfer matrix G(s). 7 4 5 0 and C = C1 = b Ab b1 A1 b1 Then it is easy to verify that both C and C1 are nonsingular. .1 + + 2 sn.1 + + + a n.1 s a n + d s a1 n. .1 Tcb = A1 b1 CTc. CTc.1 b1 : 1 TcATc. a state space representation of the transfer matrix can be obtained as above.an.a 6 11 02 6 A1 := 6 0 6 .1 b + d = 1 s n + + sn.1 = 1 2 n.. given a column vector of transfer matrix G(s). First.1 d CTc.2 G(s) = C (sI .60 LINEAR DYNAMICAL SYSTEMS We now consider some special state space representations of the dynamical system described by equations (3.1 . Assume that a single input and multiple output dynamical system is given by A b G(s) = C d b 2 Rn C 2 Rp n d 2 Rp and assume that (A b) is controllable. Moreover.2). 4 0 0 .an 3 0 0 .1 + + an and de ne 2 . .. It is also easy to show that the transfer matrix is given by n.1 n .1) and (3. we will consider the systems with single inputs. A). 1 0 0 . This state space representation is usually called controllable canonical form or controller canonical form.1 s + n which also shows that.1 d where for some i 2 Rp ..7 6 .1 b An. 1 6 . 0 7 7 7 7 7 5 213 607 6 7 b1 := 6 0 7 6.1 will give the equivalent system representation An. A) = n + a1 n...a . the transformation Tc = C1 C ..
4. (A1 + b1 F )) = n +(a1 . f1) n. It is clear that the zeros of det( I . A). sn. the single input pair (A Bv) is controllable. fn ).1 B + d = 1 sn + a sn.3.1 d 6 6 . Dually. . thus showing that the eigenvalues of A + bF can be freely assigned if (A b) is controllable. 0 0 0 0 1 .1 + + + a n. .1 + +(an . consider a multiple input and single output system G(s) = A B c d B 2 Rn m c 2 Rn d 2 Rm 1 0 .a2 6 ToB = 6 . 1 0 0 and assume that (c A) is observable then there is a transformation To such that ToATo...1 n .. .. .1+ + n + d: 1 n. 0 0 0 0 0 .1 and 2 .12 If an m input system pair (A B) is controllable and if A is cyclic.1 fn x: Then the closedloop system is given by x = (A1 + b1 F )x _ and det( I . we need some preliminary results. The pole placement problem for a multiple input system (A B ) can be converted into a simple single input pole placement problem.2 + s G(s) = c(sI . 1 2 d 3 7 7 7 7 7 7 7 5 i 2 Rm This representation is called observable canonical form or observer canonical form.an. To describe the procedure.a1 6 .. . Similarly. (A1 + b1 F ) can be made arbitrary for an appropriate choice of F provided that the complex zeros appear in conjugate pairs.1 + 2 sn. Pole Placement and Canonical Forms Now consider a dynamical system equation given by 61 x = A1 x + b1 u _ and a state feedback control law u = Fx = f1 f2 : : : fn. we can show that the eigenvalues of A + Lc can be freely assigned if (c A) is observable.1 cTo. 6 6 . then for almost all v 2 Rm .1 s an Lemma 3.an 4 1 n..
.62 that the matrix B is partitioned accordingly: LINEAR DYNAMICAL SYSTEMS Proof. Now it is evident that given a multiple input controllable pair (A B ). for each i = 1 : : : k. Moreover.. by de nition we have the following lemma. then for almost any K 2 Rm n . . the pair (A B ) is controllable if and only if. .2 7 6 .13 If (A B) is controllable... By PBH test. Davison 1968].. bi v 6= 0 for each i = 1 : : : k which is clear since for each i. the set v 2 Rm such that bi v = 0 has measure zero in Rm since bi 6= 0. Let bi 2 Rm be the last row of Bi .. Lemma 3. consequently.. Without loss of generality. all the eigenvalues of A + BK are distinct and. there is a matrix K 2 Rm n and a vector v 2 Rm such that A + BK is cyclic and (A + BK Bv) is controllable.. 5 4 Bk 3 7 7 7 5 i . for almost all v 2 Rm . from the pole placement results for the single input system. i 1 and i 6= j if i 6= j . the eigenvalues of A + BF can be arbitrarily assigned by choosing a state feedback in the form of u = Fx = (K + vf )x: i 3 7 7 7 7 7 7 5 . assume that A is in the Jordan canonical form and 2J 6 1 J2 A=6 6 4 2 6 6 6 Ji = 6 6 6 4 . A proof can be found in Brasch and Pearson 1970]. Hence. the theorem does not hold. 2 The cyclicity assumption in this theorem is essential. the last row of Bi is not zero. and Heymann 1968]. 1 i where Ji is in the form of Jk 3 2B 7 6 B1 7 B = 6 . For example. the pair A= 1 0 B= 1 0 0 1 0 1 is controllable but there is no v 2 R2 such that (A Bv) is controllable since A is not cyclic. Since a matrix A with distinct eigenvalues is cyclic. and then we only need to show that. A + BK is cyclic. Without this assumption.. there is a matrix f 2 Rn so that the eigenvalues of (A + BK )+(Bv)f can be arbitrarily assigned.
^ Theorem 3. Equivalent equations are q_ = (A + LC )q + Bu + LDu . if the system states are available for feedback. More precisely. We rst show that the detectability of (C A) is su cient for the existence of an observer. the estimation of the system states from the given output information y and input u is often necessary to realize some speci c design objectives.6) Proof. (3. The interested reader may consult Kailath 1980] or Chen 1984]. then the Kalman controllable decomposition can be applied rst and the above procedure can be used to assign the system poles corresponding to the controllable subspace. q(0) and for every input u( ). with an additional forcing term L(Cq + Du .5) (3. we only need to show that the socalled Luenberger observer de ned in the theorem is indeed an observer. However.5) for q is a simulation of the equation for x. Hence. Ly x = q: ^ . furthermore.5. y). a (linear) observer is a system such as q_ = Mq + Nu + Hy x = Qq + Ru + Sy ^ so that x(t) .3. we consider such an estimation problem and the application of this state estimation in feedback control.1) and (3. The canonical form for single input or single output system can also be generalized to multiple input and multiple output systems at the expense of notation. Consider a plant modeled by equations (3. Note that equation (3.5 Observers and ObserverBased Controllers We have shown in the last section that if a system is controllable and. in most practical applications. x(t) ! 0 as t ! 1 for all initial states x(0). Further. say x. If a system is not completely controllable.2). Observers and ObserverBased Controllers 63 A dual procedure can be applied for the output injection A + LC . y) x = q ^ where L is any matrix such that A + LC is stable. To that end. which asymptotically estimates the ^ state x. which is a gain times the output error. the system states are not completely accessible and all the designer knows are the output y and input u. In this section. 3. if (C A) is detectable. then the system closed loop poles can be assigned arbitrarily through a constant feedback.14 An observer exists i (C A) is detectable. then a full order Luenberger observer is given by q_ = Aq + Bu + L(Cq + Du . An observer is a dynamical system with input of (u y) and output of.
x(t) does not converge to zero. However. To show the converse. and then simple algebra gives ^ e = (A + LC )e _ therefore e(t) ! 0 as t ! 1 for every x(0). Recall that. if (C A) is observable. Du = Cx. then the system observer poles can be arbitrarily placed so that the state estimator x can be made to approach x arbitrarily fast. assume that (C A) is not detectable. so we only need to generate the part of x in Ker C . and. Take the initial state x(0) in the undetectable subspace and consider a candidate observer: q_ = Mq + Nu + Hy x = Qq + Ru + Sy: ^ Take q(0) = 0 and u(t) 0.2). the controller has the following dynamics: _ x = (A + LC )^ + Bu + LDu . Cx(t) = 0 for all t 0.1) and (3. q(0). If C has full row rank and p := dim y. it follows that x(t) is in the unobservable subspace for all t 0. Hence. Similarly. then there is a state feedback u = Fx such that the closedloop poles of the system can be arbitrarily assigned. This is true. The idea is this: since we can measure y . It's possible to get an observer of lower dimension. De ne the error. which is the dimension of the state x. Hence. then the dimension of Ker C equals n .64 LINEAR DYNAMICAL SYSTEMS These equations have the form allowed in the de nition of an observer. The interested reader may consult Chen 1984]. x. and u(:). Thus the candidate observer does not have the required property. p. We will not pursue this issue further here. Then the equations for x and the candidate observer are x = Ax _ q_ = Mq + HCx x = Qq + SCx: ^ Since an unobservable subspace is an Ainvariant subspace containing x(0). consequently. no observer exists. if (A B ) is controllable and state x is available for feedback. we already know x modulo Ker C . Ly ^ x u = F x: ^ . for a dynamical system described by the equations (3. Now ^ let us consider what will happen if the system states are not available for feedback so that the estimated state has to be used. 2 The above Luenberger observer has dimension n. and therefore. p. e := x . Such an observer is called a \minimal order observer". so we might suspect that we can get an observer of order n . for some x(0) in the undetectable ^ subspace. q(t) 0 and x(t) 0.
then the system equation becomes ^ x : x ^ e x ^ e _ A + LC 0 _ x = . Since these proofs are straightforward. then it is said to be output feedback stabilizable.L : 0 A B G= C D then the closedloop feedback system is as shown below: y G u K In general. The controller given above is called an observerbased controller and is denoted as u = K (s)y and Now denote the open loop plant by LC K (s) = A + BF +F + LDF . 3.6. Now if (A B ) is controllable and (C A) is observable.LC A + BF + LC x ^ Let e := x . x. they can be made to be stable. then there exist F and L such that the eigenvalues of A + BF and A + LC can be arbitrarily assigned. In particular. if a system is stabilizable through feeding back the output y. Operations on Systems Then the total system state equations are given by 65 x = A _ BF _ . we present some facts about system interconnection.LC A + BF ^ and the closedloop poles consist of two parts: the poles resulting from state feedback (A + BF ) and the poles resulting from the state estimation (A + LC ). . The converse is also true and will be shown in Chapter 12. we will leave the details to the reader.3. Note that a slightly weaker result can also result even if (A B ) and (C A) are only stabilizable and detectable.6 Operations on Systems In this section. It is clear from the above construction that a system is output feedback stabilizable if (A B ) is stabilizable and (C A) is detectable.
R12 D1 C2 D1 R21 where R12 = I + D1 D2 and R21 = I + D2 D1 . a representation for the cascaded system can be obtained as G1 G2 = D1 C2 C1 Similarly. the parallel connection or the addition of G1 and G2 can be obtained as 2 3 0 B1 A1 B1 + A2 B2 = 4 A1 A 0 B2 5 : G1 + G2 = C D 2 C2 D2 1 1 C1 C2 D1 + D2 Next we consider a feedback connection of G1 and G2 as shown below: y r f G1 A2 B2 A1 B1 C1 D1 2 2 A B C C2 BDD 1 1 2 1 2 B2 = 4 0 A2 C1 D1 C2 D1 D2 3 2 A 0 2 5 = 4 B1C2 A1 B2 B1 D2 D1 D2 3 5: 6 .G2 Then the closedloop transfer matrix from r to y is given by R12 C1 . B2 R121 C1 A2 . . we shall also introduce the following de nitions.B1 R211 C2 B1 R211 1 1 2 12 1 . B2 D1 R211 C2 B2 D1 R211 5 T =4 . Note that these state space representations may not be necessarily controllable and/or observable even if the original subsystems G1 and G2 are. Hence. . .1 .66 LINEAR DYNAMICAL SYSTEMS Suppose that G1 and G2 are two subsystems with state space representations: B G1 = A1 D1 C1 1 B G2 = A2 D2 : C2 2 Then the series or cascade connection of these two subsystems is a system with the output of the second subsystem as the input of the rst subsystem as shown in the following diagram: G1 G2 This operation in terms of the transfer matrices of the two subsystems is essentially the product of two transfer matrices.1 . B D R.1 . .1C . 3 . For future reference. 2 A .
The right inverse case will be proven and the left inverse case follows by duality. Lemma 3.! GT (s) = B (sI .6.7 The transpose of a transfer matrix G(s) or the dual system is de ned as G 7. BDy C .BDy 5 DDy 3 BDy . BDy C yC 2 C DDy C A BD 4 0 A . Then GGy = = C Performing similarity transformation T = I I on the above system yields 0 I 3 2A 0 0 GGy = 4 0 A . Then y y Gy = A .! B D In particular.BDy 5 : I 3 2 . Operations on Systems 67 De nition 3.15 Let Dy denote a right (left) inverse of D if D has full row (column) rank.3. ^ ^ ^ matrix G(s) if G(s)G(s) = I ( G(s)G(s) = I ).! G (s) := GT (.1 C + D or equivalently A B . then it is simply called the inverse of G(s).1 C + D or equivalently A C A B C D 7.! B D : De nition 3. we have G (j!) := G(j!)] = G (j!).9 A real rational matrix G(s) is called a right (left) inverse of a transfer is a right (left) inverse of G. A ). Moreover. BD C . A ).sI .BDy 5 C 0 I = I: 2 A BDy C 4 0 A .8 The conjugate system of G(s) is de ned as G 7.A .C C D 7.s) = B (. Suppose DDy = I . : ^ De nition 3. if G(s) is both a right inverse and a left inverse of G(s). BDy C C BDy .BD yC y D D Proof.
1 Bm + D CAi B = Cm Aim Bm 8i 0: we have This implies that OC = Om Cm (3:7) . the natural or convenient description for a dynamical system is in terms of transfer matrices. there is a smaller dimensioned controllable and observable state space realization that has the same transfer matrix this contradicts the minimality assumption. A). input and output frequency responses are obtained from experiments so that some transfer matrix approximating the system dynamics is available. We note that if the transfer matrix is either single input or single output. De nition 3. We shall rst show that if (A B C D) is a minimal realization of G(s). some appropriate state space representation for the resulting transfer matrix is necessary.1 B + D = Cm (sI . then it is minimal. Am ).7 State Space Realizations for Transfer Matrices In some cases.68 LINEAR DYNAMICAL SYSTEMS 3. In general. let (Am Bm Cm D) be a minimal realization of G(s) with order k < n. Since the state space computation is most convenient to implement on the computer. on the contrary. then the formulas in Section 3. Next we show that if an nth order realization (A B C D) is controllable and observable. Hence. then (A B ) must be controllable and (C A) must be observable. Then from Kalman decomposition. for example.10 A state space realization (A B C D) of G(s) is said to be a minimal realization of G(s) if A has the smallest possible dimension.16 A state space realization (A B C D) of G(s) is minimal if and only if (A B ) is controllable and (C A) is observable. Then we call a statespace model (A B C D) such that A B G(s) = C D a realization of G(s). assume that G(s) is a realrational transfer matrix which is proper. Proof. certain engineering approximation or identi cation has to be carried out for example.4 can be used to obtain a controllable or observable realization. But supposing it is not minimal. Theorem 3. that (A B ) is not controllable and/or (C A) is not observable. Suppose. The realization for a general MIMO transfer matrix is more complicated and is the focus of this section. in some highly complex systems for which the analytic di erential equations are too hard or too complex to write down. Since G(s) = C (sI . Hence (A B ) must be controllable and (C A) must be observable. This occurs.
and let C1 . and Cm := Om := By Sylvester's inequality. 7 4 5 An.. therefore. We now describe several ways to obtain a state space realization for a given multiple input and multiple output transfer matrix G(s).1 Bm m Cm An. respectively. and O2 be the corresponding controllability and observability matrices.7). let us consider a 2 2 (block) transfer matrix such as G(s) = G1 (s) G2 (s) G3 (s) G4 (s) and assume that Gi (s) has a state space realization of B Gi (s) = Ai Di Ci i i = 1 : : : 4: . To illustrate. Theorem 3.3. (Am Bm ) is controllable and (Cm Am ) is observable. 2 The following property of minimal realizations can also be veri ed. Moreover.1 = C1 C2 (C2 C2 ). n rank (OC ) minfrank C rank Og and. C2 .1: Furthermore. respectively. State Space Realizations for Transfer Matrices 69 where C and O are the controllability and observability matrices of (A B ) and (C A).17 Let (A1 B1 C1 D) and (A2 B2 C2 D) be two minimal realizations of a real rational transfer matrix G(s). O1 .1 B2 = TB1 C2 = C1 T .A 7 : 6 . Then there exists a unique nonsingular T such that A2 = TA1T . rank Om Cm = k < n which is a contradiction since rank OC = rank Om Cm by equality (3. since (Am Bm Cm D) is minimal. T can be speci ed as T = (O2 O2 ). and this is left to the reader.1 . Similarly.1 m rank C + rank O . m 2 BmC AmB3 6 Cmmm 7 6 . The simplest and most straightforward way to obtain a realization is by realizing each element of the matrix G(s) and then combining all these individual realizations to form a realization for G(s).7. we have rank (OC ) = n since rank C = rank O = n by the controllability and observability assumptions.1 O2 O1 or T .
i: i=1 rank Wi = ki and let Bi 2 Rki m and Ci 2 Rp ki be two constant matrices such that Wi = Ci Bi : . if Gi (s) is a column or row vector of transfer functions. Let G(s) be a p m transfer matrix and write it in the following form: G(s) = N((ss)) d with d(s) a scalar polynomial. if the transfer matrix G(s) can be factored into the product and/or the sum of several simply realized transfer matrices. (An alternative numerically reliable method to eliminate uncontrollable and/or unobservable states is the balanced realization method which will be discussed later. In particular.) We will now describe one factorization procedure that does result in a minimal realization by using partial fractional expansion (The resulting realization is sometimes called Gilbert's realization due to Gilbert). a Kalman controllability and observability decomposition has to be performed to eliminate the uncontrollable and/or unobservable states. we shall assume that d(s) has only real and distinct roots i 6= j if i 6= j and d(s) = (s . r ): Then G(s) has the following partial fractional expansion: r X Wi s. 2 ) G(s) = D + Suppose (s . For simplicity.4 can be used to obtain a controllable or observable realization for Gi (s). then a realization for G can be obtained by using the cascade or addition formulas in the last section. A problem inherited with these kinds of realization procedures is that a realization thus obtained will generally not be minimal. 1 )(s .70 LINEAR DYNAMICAL SYSTEMS Note that Gi (s) may itself be a multiple input and multiple output transfer matrix. To obtain a minimal realization. Then a realization for G(s) can be given by 2A 0 0 0 B 0 3 6 01 A2 0 0 01 B2 7 6 0 0 A 0 B 0 7 6 7 G(s) = 6 0 0 03 A4 03 B4 7 : 6 6C C 0 0 D D 7 7 4 1 2 1 2 5 0 0 C3 C4 D3 D4 Alternatively. then the formulas in Section 3.
Hence. these tests often have to be done indirectly. An immediate consequence of part (iii) is that. then (Q A) is observable i X > 0. the Lyapunov theory is sometimes useful. Consider the following Lyapunov equation A X + XA + Q = 0 (3:8) with given real matrices A and Q. 7 .18 Assume that A is stable. a pair (A B ) is controllable if and only if the solution to ALc + LcA + BB = 0 . 3. However. This approach can. and observability of a system is very important in linear system analysis and synthesis. it is minimal. given a stable matrix A. In this section.. then the following statements hold: R (i) X = 01 eA t QeAtdt. Lemma 3. It has been shown in Chapter 2 that this equation has a unique solution i i (A) + j (A) 6= 0 8i j .. An immediate consequence of this minimal realization is that a transfer matrix with an rth order polynomial denominator does not necessarily have an rth order state space realization unless Wi for each i is a rank one matrix. controllability. a pair (C A) is observable if and only if the solution to the following Lyapunov equation is positive de nite: A Lo + Lo A + C C = 0: The solution Lo is called the observability Gramian. The following results are standard. in fact. Readers may convince themselves by trying some simple examples.3. (iii) if Q 0. B1 r Ikr Cr 3 .. we will study the relationships between the stability of A and the solution of X . Similarly. 7: 7 Br 5 D It follows immediately from PBH tests that this realization is controllable and observable. be generalized to more complicated cases where d(s) may have complex and/or repeated roots. (ii) X > 0 if Q > 0 and X 0 if Q 0. In that respect.8 Lyapunov Equations Testing stability. Lyapunov Equations Then a realization for G(s) is given by 71 2 I 6 1k G(s) = 6 6 4 C1 1 .8.
To motivate the class of realizations.20 Let C D be a state space realization of a (not necessarily stable) transfer matrix G(s). Premultiply equation (3.. Qv = 0. In many applications. 2 3. Suppose that there exists a symmetric matrix P = P = P1 0 0 0 with P1 nonsingular such that AP + PA + BB = 0: Now partition the realization (A B C D) compatibly with P as 2A A B 3 12 4 A11 A22 B1 5 : 21 2 C1 C2 D . This implies that is an unstable and unobservable mode.9 Balanced Realizations Although there are in nitely many di erent state space realizations for a given transfer matrix. we are given the solution of the Lyapunov equation and need to conclude the stability of the matrix A. some particular realizations have proven to be very useful in control engineering and signal processing. (iii) A is stable if X 0. Then we must have v Qv = 0. which contradicts the assumption that (Q A) is detectable.8) by v to get 2Re (v Xv) + v Qv = 0: Now if X > 0 then v Xv > 0. then Av = v. To see (iii). then (i) Re i (A) 0 if X > 0 and Q 0. i. Here we will only introduce one class of realizations for stable transfer matrices that are most useful in control applications.e.8).19 Suppose X is the solution of the Lyapunov equation (3. Proof.8) by v and postmultiply (3. Q 0 and (Q A) is detectable.72 LINEAR DYNAMICAL SYSTEMS is positive de nite and Lc is called the controllability Gramian. A B Lemma 3. we rst consider some simple facts. Hence (i) and (ii) hold. we assume Re 0. Let be an eigenvalue of A and v 6= 0 be a corresponding eigenvector. Lemma 3. and it is clear that Re 0 if Q 0 and Re < 0 if Q > 0. (ii) A is stable if X > 0 and Q > 0.
18 that (A11 B1 ) is controllable if A11 is stable. The above two lemmas suggest that to obtain a minimal realization from a stable nonminimal realization. Balanced Realizations Then 73 A11 B1 C1 D is also a realization of G. Proof. Suppose that there exists a symmetric matrix Q = Q = Q1 0 0 0 with Q1 nonsingular such that QA + A Q + C C = 0: Now partition the realization (A B C D) compatibly with Q as 2A A B 3 12 4 A11 A22 B1 5 : 21 2 C1 C2 D Then A11 B1 C1 D is also a realization of G. We also have 2A A B 3 2A A B 3 11 12 4 A21 A22 B1 5 = 4 011 A12 01 5 = A11 B1 : 2 22 C1 D C1 C2 D C1 2 A B Lemma 3. Moreover. it follows from Lemma 3. Moreover. (A11 B1 ) is controllable if A11 is stable. one only needs to eliminate all states corresponding to the zero block diagonal term of the controllability Gramian P and the observability Gramian Q. Hence. (C1 A11 ) is observable if A11 is stable.9.3.21 Let C D be a state space realization of a (not necessarily stable) transfer matrix G(s). the following procedure can be used to eliminate noncontrollable subsystems: . Use the partitioned P and (A B C ) to get A 0 = AP + PA + BB = A11 P1 +PP1+ 11 + B1 B1 P1 A21 + B1 B2 A21 1 B2 B1 B2 B2 which gives B2 = 0 and A21 = 0 since P1 is nonsingular. In the case where P is not block diagonal. part of the realization is not controllable: C2 D Finally.
1 2= 1 2 0 3 5 where is any nonzero number. Compute the controllability Gramian P 0 from AP + PA + BB = 0: 3. Diagonalize P to get P = U1 U2 U1 U2 unitary. Now assume that 1 > 0 is diagonal and 1 = diag( 11 12 ) such that max ( 12 ) min ( 11 ). 1 1 AU 4. If the state corresponding to the last diagonal term of P is removed. Consider a stable transfer function + G(s) = s2 3s 3s 18 18 : + + Then G(s) has a state space realization given by 2 . then it is tempting to conclude that one can also discard those states corresponding to 12 without causing much error. It is easy to check that the controllability Gramian of the realization is given by P = 0:5 2 : Since the last diagonal term of P can be made arbitrarily small by making small. this is not necessarily true as shown in the following example.4= . A dual procedure can also be applied to eliminate nonobservable subsystems. we get a transfer function . which is Q = 0:5 1= 2 : .74 LINEAR DYNAMICAL SYSTEMS A B 1.1 .1 ^ G = . However.2 G(s) = 4 4 . 2. the controllability of the corresponding state can be made arbitrarily weak. The problem may be easily detected if one checks the observability Gramian Q. Then G(s) = UCU 1 UDB 1 1 0 0 0 U1 U2 with 1 > 0 and is a controllable realization.11 0 + which is not close to the original transfer function in any sense. Let G(s) = C D be a stable realization.1 1 = s.
1.18.1) QT .1. and (C A) is observable i Q > 0. Suppose the state is transformed by a nonsingular T to x = Tx to yield the realiza^ tion " ^ ^# . . This example shows that controllability (or observability) Gramian alone can not give an accurate indication of the dominance of the system states in the input/output behavior. it will be shown that PQ has a real diagonal Jordan form and that 0.1 = where = diag( 1 2 : : : n ) and 2 = . This motivates the introduction of a balanced realization which gives balanced Gramians for controllability and observability. if a realization of a stable system is not minimal.e. Consider the similarity transformation T which gives the eigenvector decomposition PQ = T . This is a consequence of the following matrix fact.3. in the case of a minimal realization they can always be chosen such that ^ P = TPT = ^ Q = (T . then there is a transformation such that the controllability and observability Gramians for the transformed realization are diagonal and the controllable and observable subsystem is balanced. the pair (A B ) is controllable i P > 0.1 ) QT . A B Suppose G = C D is stable. Let P and Q denote the controllability Gramian and observability Gramian.1 are eigenvectors of PQ corresponding to the eigenvalues f i g.1 TB : CT D C D ^ ^ Then the Gramians are transformed to P = TPT and Q = (T . and therefore the eigenvalues of the product of the Gramians are ^ P invariant under state transformation.9. Although the eigenvectors are not unique. More generally. are called the Hankel singular values of the system. which are consequences of P 0 and Q 0.1 A B G = ^ ^ = TAT.. A is stable. Q 0. This new realization with controllability ^ ^ and observability Gramians P = Q = will be referred to as a balanced realization (also called internally balanced realization). this shows that the state corresponding to the last diagonal term is strongly observable. P and Q satisfy the following Lyapunov equations AP + PA + BB = 0 (3:9) A Q + QA + C C = 0 (3:10) and P 0. 1 2 ::: n 0. Then by Lemma 3.1 T = diag( 1 : : : n ): Then the columns of T . Balanced Realizations 75 Since 1= 2 is very large if is small. Note that ^ Q = TPQT . respectively. i. Later. The decreasingly order numbers. Furthermore.
1 = 4 I ^ But Q 0 implies Q122 = 0.1 QT1.1) QT .1(T2 ).Q 1 3 5 3 5: giving 2 (T3 ).1 = 4 2 (T3 ).22 Let P and Q be two positive semide nite matrices.1 (T2 ).1 = U1 I 0 and then > 0: 2 (T2 ).1 (T1 ).1 = 6 4 1 0 3 0 3 7 7 5 Proof. 0 3 7 7 5 2 6 (T .1 QT1.1 (T3 ).1 (T1 ).1(T2 ). So now let ^ ^ Q121 Q122 0 0 I 0 0 I 2 1 0 2 1 0 0 ^ Q121 ^ Q122 5 : Q22 3 0 ^ 121 .76 a nonsingular matrix T such that LINEAR DYNAMICAL SYSTEMS Theorem 3. Q121 .2 Q121 3 0 0 0 3 > 0: Next nd a unitary matrix U2 such that . there exists a transformation T1 such that Now let T1PT1 = I 0 : 0 0 (T1 ).1 QT1. Q121 . Since P is a positive semide nite matrix. with 1 2 0 3 diagonal and positive de nite. Then there exists 2 6 TPT = 6 4 1 2 respectively.1 = Q11 Q12 Q Q 12 22 1 and there exists a unitary matrix U1 such that 2 U1 Q11 U1 = 01 0 0 Let 0 (T2 ).1 = 4 ^ 1 ^ U2 (Q22 .2 .2 Q121 )U2 = 0 0 0 0 0 ^ 1 ^ 0 0 Q22 .
A B In the special case where C D is a minimal realization.1 = 4 and let Then 0 0 I 0 U2 0 0 3 5 T = T4 T3 T2 T1: 2 6 TPT = 6 4 2 = I. with 1 2 3 diagonal and positive de nite. 1 2 0 0 3 7 7 5 2 6 (T ).1 T such that G = TAT. a balanced realization can be obtained through the following simpli ed procedure: .1 TB CT D Gramian Q given by A B Corollary 3.23 The product of two positive semide nite matrices is similar to a positive semide nite matrix. Then it is easy to see that with the transformation given above TPQT . Proof. Let P and Q be any positive semide nite matrices.1 = 0 0 0 : 2 1 2 .9.1 QT .24 For any stable system G = C D . there exists a nonsingular has controllability Gramian P and observability 2 6 P =6 4 1 2 0 0 3 2 7 Q=6 7 6 5 4 1 0 3 0 3 7 7 5 respectively.1 = 6 4 1 0 3 0 3 7 7 5 2 with Corollary 3.1=2 0 1 2 (T4 ).3. Balanced Realizations De ne 77 .
1 s+1 u The transfer function for this system. 2.10 Hidden Modes and PoleZero Cancelation Another important issue associated with the realization theory is the problem of uncontrollable and/or unobservable unstable modes in the dynamical system.1 x2 2 1 1 g(s) = s .1 = R U . truncating those less controllable and observable states will not lose much information about the system. Find a matrix R such that P = R R. 1 = s + 1 s+1 y = 1 0 x1 x2 . Therefore.1 = and TAT. 3.78 LINEAR DYNAMICAL SYSTEMS 1. . Compute the controllability and observability Gramians P > 0 Q > 0.1 TB CT D is balanced. Diagonalize RQR to get RQR = U 2 U . : Then a state space description for this dynamical system is given by x1 = 1 .1 3. and output normal realization with P = 2 and Q = I . 1 s .1=2 . These statements will be made more concrete in Chapter 7 where error bounds will be derived for the truncation error. Assume that the Hankel singular values of the system is decreasingly ordered so that = diag( 1 2 : : : n ) and 1 2 : : : n and suppose r r+1 for some r then the balanced realization implies that those states corresponding to the singular values of r+1 : : : n are less controllable and observable than those states corresponding to 1 : : : r . let x1 = y x2 = u . 4. This problem is illustrated in the following example: Consider a series connection of two subsystems as shown in the following diagram y s. Two other closely related realizations are called input normal realization with P = I and Q = 2 .1 QT .1 _ x1 + 1 u x2 _ 0 . is stable and has a rst order minimal realization. On the other hand. Then TPT = (T ).1 1 s. Both realizations can be obtained easily from the balanced realization by a suitable scaling on the states. Let T .
if the state is available for feedback control.10. However. it is not quite so for an MIMO transfer matrix. In fact. there are fundamental di erences between these two types of interconnections as far as control design is concerned. For example. 2 0 1 which is unstable. Moreover. This implies that G(s) must have an unstable zero at 2 that cancels the unstable pole of K . the notion of \system zero" cannot be generalized naturally from the scalar transfer function zeros. The unstable mode can still result in the internal signal unbounded if the initial state (0) is not zero. Let 2 s+2 1 3 p . Hence. 3 2 p s+ 2 6 .1 s+1 s. consider the following transfer matrix 2 1 1 3 6 s+1 s+2 7 7 G(s) = 6 4 2 1 5 s+2 s+1 which is stable and each element of G(s) has no nite zeros. . the output can be unbounded if the initial state x1 (0) is not zero. it is easy to show that the unstable mode 1 is uncontrollable but observable. 2 s. This example shows that we must be very careful in canceling unstable modes in the procedure of forming a transfer function in control designs otherwise the results obtained may be misleading and those unstable modes become hidden modes waiting to blow. (s + 1)(s + 2) 0 7 7 KG = 6 7 6 4 2 1 5 s+2 s+1 p is stable. the unstable mode 1 becomes controllable but unobservable. This leads us to the next topic: multivariable system poles and zeros. For instance. One observation from this example is that the problem is really caused by the unstable zero of the subsystem s. Hidden Modes and PoleZero Cancelation 79 and is a second order system. We should also note that the above problem does not go away by changing the interconnection order: y s.1 1 u In the later case.3. Although the zeros of an SISO transfer function s+1 are easy to see. s +p 5 K =4 s. Of course. then the latter interconnection can be stabilized while the former cannot be.1 .
7 . . Then the normal rank of Q(s). However. 6 . It is a fact in linear algebra that any polynomial matrix can be reduced to a socalled Smith form through some pre. Kailath.unimodular operations. .25 (Smith form) Let P (s) 2 R s] be any polynomial matrix. i.80 LINEAR DYNAMICAL SYSTEMS Let R s] denote the polynomial ring with real coe cients. is the rank in R s] or. pp. is the maximum dimension of a square submatrix of Q(s) with nonzero determinant in R s]. 1984. In short. 7 U (s)P (s)V (s) = S (s) := 6 . To show the di erence between the normal rank of a polynomial matrix and the rank of the polynomial matrix evaluated at certain point. A matrix is called a polynomial matrix if every element of the matrix is in R s].391]. 7 6 0 7 4 0 (s) 0 5 r 0 0 0 0 and i (s) divides i+1 (s). . consider 1 Q(s) = ss2 1 : Then Q(s) has normal rank 2 since det Q(s) = s .. Let 2 s + 1 (s + 1)(2s + 1) s(s + 1) 3 P (s) = 4 s + 2 (s + 2)(s2 + 5s + 3) s(s + 2) 5 : 1 2s + 1 s The polynomial matrix P (s) has normal rank 2 since + + det(P (s)) 0 det s + 1 (s (s 2)(1)(2s 5s 1) 3) = (s + 1)2 (s + 2)2 6 0: s + 2 + s2 + + 3. De nition 3. s2 6 0. It is a fact that a square polynomial matrix is unimodular if and only if it is invertible in R s]. . . . Lemma 3. denoted normalrank (Q(s)).11 Multivariable System Poles and Zeros . . .. Q(0) has rank 1. A square polynomial matrix is called a unimodular matrix if its determinant is a nonzero constant (not a polynomial of s).e. its inverse is also a polynomial matrix. cf. sometimes we say that a polynomial matrix Q(s) has rank(Q(s)) in R s] when we refer to the normal rank of Q(s). S (s) is called the Smith form of P (s). It is also clear that r is the normal rank of P (s). . We shall illustrate the procedure of obtaining a Smith form by an example. equivalently.11 Let Q(s) 2 R s] be a (p m) polynomial matrix.and post. . then there exist unimodular matrices U (s) V (s) 2 R s] such that 2 (s) 0 3 0 0 1 6 0 2 (s) 0 07 6 .
N s G s r N s r G s p s G s N s =d s d s . . Multivariable System Poles and Zeros 81 First interchange the rst row and the third row and use row elementary operation to zero out the s+1 and s+2 elements of P (s). . 1 It is obvious that a similar notion of normal rank can be extended to a matrix in R ( ). 2. ..1 Then any real rational transfer matrix can be reduced to a socalled McMillan form through some pre. . let Rp (s) denote the set of rational proper transfer matrices. In particular. . Lemma 3. U (s)G(s)V (s) = M (s) := 6 .unimodular operations. then there exist unimodular matrices U (s) V (s) 2 R s] such that 3 2 1 (s) 0 0 0 1 (s) 6 0 2((ss)) 0 07 7 6 7 6 . 7 .and post.(s + 1) Then 21 3 2s + 1 s P1 (s) := U (s)P (s) = 4 0 (s + 1)(s + 2)2 0 5 : 0 0 0 Next use column operation to zero out the 2s + 1 and s terms in P1 . This process can be done by postmultiplying a unimodular matrix V to P1 (s): 2 1 . .(s + 2) 5 : 1 0 .s 3 1 0 5 V (s) = 4 0 0 0 1 and 21 3 0 0 P1 (s)V (s) = 4 0 (s + 1)(s + 2)2 0 5 : 0 0 0 Then we have 21 3 0 0 S (s) = U (s)P (s)V (s) = 4 0 (s + 1)(s + 2)2 0 5 : 0 0 0 Similarly.3.11. then ( ) is said to have normal rank if ( ) has normal rank .26 (McMillan form) Let G(s) 2 Rp(s) be any proper real rational transfer matrix. This process can be done by premultiplying a unimodular matrix U to P (s): 20 0 1 3 U = 4 0 1 . 7 6 . . let ( ) 2 R ( ) be a rational matrix and write ( ) = ( ) ( ) such that ( ) is a p s polynomial and ( ) is a polynomial matrix. 7 6 r (s) 5 4 0 0 r (s) 0 0 0 0 0 and i (s) divides i+1 (s) and i+1 (s) divides i (s). . .(2s + 1) .
it can be shown that the dimension of a minimal realization of G(s) is exactly the McMillan degree of G(s). Consider a 3 3 transfer matrix: 3 2 2s + 1 s 1 6 (s + 1)(s + 2) (s + 1)(s + 2) (s + 1)(s + 2) 7 7 6 7 6 7: 6 1 s2 + 5s + 3 s 7 6 (s + 1)2 G(s) = 6 7 (s + 1)2 (s + 1)2 7 6 7 6 5 4 1 2s + 1 s (s + 1)2 (s + 2) (s + 1)2 (s + 2) (s + 1)2 (s + 2) Then G(s) can be written as s + 1 (s + 1)(2s + 1) s(s + 1) G(s) = (s + 1)1(s + 2) 4 s + 2 (s + 2)(s2 + 5s + 3) s(s + 2) 2 1 2s + 1 s 2 3 5 := N (s) : d(s) . We now illustrate the above concepts through an example. If we write the transfer matrix G(s) as G(s) = N (s)=d(s) such that d(s) is a De nition 3. Then it is fairly easy to show that a complex number is a pole of G(s) if and only if it is an eigenvalue of A. Moreover. It is clear that a blocking zero is a transmission zero. 2 where deg( i (s)) denotes the degree of the polynomial i (s). the blocking zeros and the transmission zeros are the same.. the conclusion follows by letting M (s) = S (s)=d(s). De nition 3. The McMillan degree of a transfer matrix is closely related to the dimension of a minimal realization of G(s). for a scalar transfer function. In fact. De nition 3.14 The roots of all the polynomials i(s) in the McMillan form for G(s) are called the transmission zeros of G(s). A complex number z0 2 C is called a blocking zero of G(s) if G(z0 ) = 0.82 LINEAR DYNAMICAL SYSTEMS Proof.e.12 The number Pi deg( i (s)) is called the McMillan degree of G(s) scalar polynomial and N (s) is a p m polynomial matrix and if let the Smith form of N (s) be S (s) = U (s)N (s)V (s). the highest power of s in i (s). Let (A B C D) be a minimal realization of G(s).13 The roots of all the polynomials i(s) in the McMillan form for G(s) are called the poles of G. i.
11.1 .1 3 2 7 6 0 0 .1 7 6 7 6 0 0 0 . it is clear that the G(s) has the McMillan form 3 2 1 0 07 6 (s + 1)2(s + 2) 7 6 7 6 6 s+2 0 7 M (s) = U (s)G(s)V (s) = 6 7 0 6 s+1 7 7 6 4 0 0 0 5 and G(s) has McMillan degree of 4. The poles of the transfer matrix are f.2 0 1 +s + 1 4 0 1 1 . To get a minimal state space realization for G(s).2g and the transmission zero is f.1 .3 .2 1 . Note that the transfer matrix has pole and zero at the same location f. Then z0 2 C is a transmission zero of G(s) if and only if there exists a 0 6= u0 2 C m such that G(z0 )u0 = 0.1 .1 1 0 .2g this is the unique feature of multivariable systems.27 Let G(s) be a p m proper transfer matrix with full column normal rank.1 3 1 + s + 2 4 0 5 1 . Nevertheless. the Gilbert's realization procedure described in the last section cannot be used directly.1 0 0 0 7 6 7 6 1 0 0 0 0 1 0 7 5 4 0 1 0 1 0 0 0 We remind readers that there are many di erent de nitions of system zeros.1 .1 0 1 . 2 3 2 3 1 405 5 .1 0 1 0 0 3 1 3 6 0 .1 3 2 + (s + 1)2 1 1 2 . The de nitions introduced here are the most common ones and are useful in this book.1 . note that G(s) has the following partial fractional expansion: 2 3 2 3 20 0 03 1 4 1 5 1 .2 7 : G(s) = 6 7 6 0 0 1 .1 Lemma 3. .1. a careful inspection of the fractional expansion results in a 4th order minimal state space realization: 2 .2g.1 + 1 4 0 5 0 3 1 G(s) = 4 0 1 0 5 + s + 1 0 s+1 1 0 0 0 0 0 1 Since there are repeated poles at .3 .3. Multivariable System Poles and Zeros 83 Since N (s) is exactly the same as the P (s) in the previous example.1 .
1 (z0 )e1 6= 0 where e1 = 1 0 0 : : :] 2 Rm .84 LINEAR DYNAMICAL SYSTEMS Proof... 1 (z0 ) = 0.. i. .. This can be seen from the following example... assume 2 6 6 6 G(s) = U1 (s) 6 6 6 4 1 1 0 .1 : . Then 2 6 6 6 U1 (z0 ) 6 6 6 4 1 1 (z0 ) (z0 ) 0 . .. .. 0 0 2 2 (z0 ) (z0 ) 0 . On the other hand.e. Without loss of generality... m (z0 ) m (z0 ) 0 0 .2 7 = V1 (z0)u0 6= 0: 6 . . 0 0 . 0 3 7 7 7 V (z )u = 0: 7 1 0 0 7 7 5 De ne 2u 3 6 u1 7 6 . We shall rst show that there is a vector u0 2 C m such that G(z0 )u0 = 0 if z0 2 C is a transmission zero. 0 3 7 7 7 V (s) 7 1 7 7 5 for some unimodular matrices U1 (s) and V1 (s) and suppose z0 is a zero of 1 (s). . . m (z0 )um ... Consider 1 G(s) = s + 1 1 1 1 1 1 u0 = . 3 7 7 = 0: 7 5 2 Note that the lemma may not be true if G(s) does not have full column normal rank. 0 0 (s) (s) . Let u0 = V1. suppose there is a u0 2 C m such that G(z0 )u0 = 0. 0 0 (s) (s) 0 2 2 .. 7 4 5 2 6 6 6 4 um 1 (z0 )u1 2 (z0 )u2 Then This implies that z0 must be a root of one of polynomials i (s) i = 1 : : : m. Then it is easy to verify that G(z0 )u0 = 0. . m (s) m (s) 0 0 . We shall outline a proof of this lemma.
0 G(s) = s+1 s+2 0 s. s.29 Suppose z0 2 C is not a pole of G(s).1 s. In the case where the transmission zero is not a pole of G(s). Suppose zo 2 C is not a pole of G(s).1 has a zero at 1 which is not a zero of det G(s). For example. Similarly we have the following lemma: Lemma 3. Then z0 2 C is a transmission zero of G(s) if Using thep above corollary. Then z0 2 C is a transmission zero of G(s) if and only if there exists a 0 6= 0 2 C p such that 0 G(z0 ) = 0. we can con rm that the example in the last section does have a zero at 2 since 2 1 1 3 6 s + 1 s + 2 7 = 2 . For example. Lemma 3. G(s) is not required to be full column (or row) rank in this case. 0 G(s) = s+1 s+2 0 s. and only if det G(z0 ) = 0. The poles and zeros of a transfer matrix can also be characterized in terms of its state space realizations. The reason is that G(z0 )u0 may be well de ned. The following lemma is easy to show from the de nition of zeros.30 Let G(s) be a square m m proper transfer matrix and det G(s) 6 0.1 A B C D . Corollary 3.28 Let G(s) be a p m proper transfer matrix with full row normal rank.11. Then z0 is a transmission zero if and only if rank(G(z0 )) < normalrank(G(s)). It should also be noted that the above lemma applies even if z0 is a pole of G(s) although G(z0 ) is not de ned. s2 : 7 det 6 4 2 1 5 (s + 1)2 (s + 2)2 s+2 s+1 Note that the above corollary may not be true if z0 is a pole of G. we can give a useful alternative characterization of the transfer matrix transmission zeros. Let be a state space realization of G(s). Multivariable System Poles and Zeros 85 It is easy to see that G has no transmission zero but G(s)u0 = 0 for all s. Furthermore. 1 is a transmission zero.1 u0 = 1 : 0 Then G(1)u0 = 0.3. Therefore.
86
LINEAR DYNAMICAL SYSTEMS
To de ne zeros, let us consider the following system matrix
De nition 3.15 The eigenvalues of A are called the poles of the realization of G(s).
B Q(s) = A ; sI D : C De nition 3.16 A complex number z0 2 C is called an invariant zero of the system
B B rank A ; z0 I D < normalrank A ; sI D : C C The invariant zeros are not changed by constant state feedback since ; B B I 0 A ; z0 I B : rank A + BFDFz0 I D = rank A ; z0 I D C+ C F I = rank C D It is also clear that invariant zeros are not changed under similarity transformation. The following lemma is obvious. B Lemma 3.31 Suppose A ; sI D has full column normal rank. Then z0 2 C is C an invariant zero of a realization (A B C D) if and only if there exist 0 6= x 2 C n and u 2 C m such that A ; z0 I B x = 0:
Moreover, if u = 0, then z0 is also a nonobservable mode.
realization if it satis es
C
D
u
Proof. By de nition, z0 is an invariant zero if there is a vector x 6= 0 such that u
A ; z0 I B C D x =0 u B since A ; sI D has full column normal rank. C On the other hand, suppose z0 is an invariant zero, then there is a vector x 6= 0 u such that A ; z0 I B x = 0: C D u
We claim that x 6= 0. Otherwise, column normal rank, i.e., x = 0 which is a contradiction. u
B A ; sI B has full D u = 0 or u = 0 since C D
3.11. Multivariable System Poles and Zeros
Finally, note that if u = 0, then
87
A ; z0 I x = 0 C
and z0 is a nonobservable mode by PBH test.
2
B Lemma 3.32 Suppose A ; sI D has full row normal rank. Then z0 2 C is an C invariant zero of a realization (A B C D) if and only if there exist 0 6= y 2 C n and v 2 C p such that A ; z0 I B = 0: y v
Moreover, if v = 0, then z0 is also a noncontrollable mode.
C
D
B Lemma 3.33 G(s) has full column (row) normal rank if and only if A ; sI D C
has full column (row) normal rank.
Proof. This follows by noting that A ; sI B =
C D
and
I 0 C (A ; sI );1 I
A ; sI
0
B G(s)
B normalrank A ; sI D = n + normalrank(G(s)): C
2
Theorem 3.34 Let G(s) be a real rational proper transfer matrix and let (A B C D)
be a corresponding minimal realization. Then a complex number z0 is a transmission zero of G(s) if and only if it is an invariant zero of the minimal realization.
Proof. We will give a proof only for the case that the transmission zero is not a pole
of G(s). Then, of course, z0 is not an eigenvalue of A since the realization is minimal. Note that A ; sI B = I 0 A ; sI B : C D C (A ; sI );1 I 0 G(s) Since we have assumed that z0 is not a pole of G(s), we have
B rank A ; z0 I D = n + rank G(z0 ): C
88 Hence
LINEAR DYNAMICAL SYSTEMS
B B rank A ; z0 I D < normalrank A ; sI D C C if and only if rank G(z0 ) < normalrank G(s). Then the conclusion follows from Lemma 3.29. 2
Note that the minimality assumption is essential for the converse statement. For example, consider a transfer matrix G(s) = D (constant) and a realization of G(s) = A 0 where A is any square matrix with any dimension and C is any matrix with compatible dimension. Then G(s) has no poles or zeros but every eigenvalue of A is an A 0 invariant zero of the realization C D .
C D
Nevertheless, we have the following corollary if a realization of the transfer matrix is not minimal. Corollary 3.35 Every transmission zero of a transfer matrix G(s) is an invariant zero of all its realizations, and every pole of a transfer matrix G(s) is a pole of all its realizations.
Lemma 3.36 Let G(s) 2 Rp(s) be a p m transfer matrix and let (A B C D) be a minimal realization. If the system input is of the form u(t) = u0 e t , where 2 C is not a pole of G(s) and u0 2 C m is an arbitrary constant vector, then the output due to the input u(t) and the initial state x(0) = ( I ; A);1 Bu0 is y(t) = G( )u0 e t 8t 0. Proof. The system response with respect to the input u(t) = u0e t and the initial condition x(0) = ( I ; A);1 Bu0 is (in terms of Laplace transform): Y (s) = C (sI ; A);1 x(0) + C (sI ; A);1 BU (s) + DU (s) = C (sI ; A);1 x(0) + C (sI ; A);1 Bu0 (s ; );1 + Du0 (s ; );1 = C (sI ; A);1 (x(0) ; ( I ; A);1 Bu0 ) + G( )u0 (s ; );1 = G( )u0 (s ; );1 :
Hence y(t) = G( )u0 e t .
2
Combining the above two lemmas, we have the following results that give a dynamical interpretation of a system's transmission zero.
Corollary 3.37 Let G(s) 2 Rp(s) be a p m transfer matrix and let (A B C D) be a minimal realization. Suppose that z0 2 C is a transmission zero of G(s) and is not a pole of G(s). Then for any nonzero vector u0 2 C m the output of the system due to the initial state x(0) = (z0 I ; A);1 Bu0 and the input u = u0 ez0 t is identically zero:
y(t) = G(z0 )u0 ez0 t = 0.
3.12. Notes and References
89
The following lemma characterizes the relationship between zeros of a transfer function and poles of its inverse.
A B Lemma 3.38 Suppose that G = C D is a square transfer matrix with D nonsingular, and suppose z0 is not an eigenvalue of A (note that the realization is not necessarily minimal). Then there exists x0 such that
(A ; BD;1 C )x0 = z0 x0 Cx0 6= 0
i there exists u0 6= 0 such that
G(z0 )u0 = 0:
Proof. (() G(z0)u0 = 0 implies that ;1 ;1 G;1 (s) = A ; BD C ;BD
D;1 C D;1
has a pole at z0 which is observable. Then, by de nition, there exists x0 such that (A ; BD;1 C )x0 = z0 x0 and ()) Set u0 = ;D;1 Cx0 6= 0. Then Using this equality, one gets
Cx0 6= 0:
(z0 I ; A)x0 = ;BD;1Cx0 = Bu0 :
G(z0 )u0 = C (z0 I ; A);1 Bu0 + Du0 = Cx0 ; Cx0 = 0:
2
The above lemma implies that z0 is a zero of an invertible G(s) if and only if it is a pole of G;1 (s).
3.12 Notes and References
Readers are referred to Brogan 1991], Chen 1984], Kailath 1980], and Wonham 1985] for the extensive treatment of the standard linear system theory. The balanced realization was rst introduced by Mullis and Roberts 1976] to study the roundo noise in digital lters. Moore 1981] proposed the balanced truncation method for model reduction which will be considered in Chapter 7.
90
LINEAR DYNAMICAL SYSTEMS
Performance Speci cations
The most important objective of a control system is to achieve certain performance speci cations in addition to providing the internal stability. One way to describe the performance speci cations of a control system is in terms of the size of certain signals of interest. For example, the performance of a tracking system could be measured by the size of the tracking error signal. In this chapter, we look at several ways of de ning a signal's size, i.e., at several norms for signals. Of course, which norm is most appropriate depends on the situation at hand. For that purpose, we shall rst introduce some normed spaces and some basic notions of linear operator theory, in particular, the Hardy spaces H2 and H1 are introduced. We then consider the performance of a system under various input signals and derive the worst possible outputs with the class of input signals under consideration. We show that H2 and H1 norms come out naturally as measures of the worst possible performance for many classes of input signals. Some state space methods of computing real rational H2 and H1 transfer matrix norms are also presented.
4
4.1 Normed Spaces
Let V be a vector space over C (or R) and let k k be a norm de ned on V . Then V is a normed space. For example, the vector space C n with any vector pnorm, k kp , for 1 p 1, is a normed space. As another example, consider the linear vector space C a b] of all bounded continuous functions on the real interval a b]. Then C a b] 91
92
PERFORMANCE SPECIFICATIONS
becomes a normed space if a supremum norm is de ned on the space: kf k1 := sup jf (t)j: A sequence fxn g in a normed space V is called a Cauchy sequence if kxn ; xm k ! 0 as n m ! 1. A sequence fxn g is said to converge to x 2 V , written xn ! x, if kxn ; xk ! 0. A normed space V is said to be complete if every Cauchy sequence in V converges in V . A complete normed space is called a Banach space. For example, Rn and C n with the usual spatial pnorm, k kp for 1 p 1, are Banach spaces. (One should not be confused with the notation k kp used here and the same notation used below for function spaces because usually the context will make the meaning clear.) The following are some more examples of Banach spaces: lp 0 1) spaces for 1 p < 1: For each 1 p < 1, lp 0 1) consists of all sequences x = (x0 x1 : : :) such that 1 X p jxi j < 1. The associated norm is de ned as
i=0 t2 a b]
kxkp :=
1 X
i=0
jxi jp
!1=p
:
l1 0 1) space: l1 0 1) consists of all bounded sequences x = (x0 x1 : : :), and the l1 norm is
de ned as
kxk1 := sup jxi j:
i
R
Lp (I ) spaces for 1 p 1: For each 1 p < 1, Lp (I ) consists of all Lebesgue measurable functions x(t)
de ned on an interval I and
kxkp :=
Z
such that
I
jx(t)jp dt
1=p
< 1 for 1 p < 1
discussed in more detail later on. C a b] space: C a b] consists of all continuous functions on the real interval a b] with the norm de ned as kxk1 := sup jx(t)j:
t2 a b]
kx(t)k1 := ess sup jx(t)j: t2I Some of these spaces, for example, L2 (;1 0], L2 0 1) and L2 (;1 1), will be
4.2. Hilbert Spaces
93
Note that if each component or function is itself a vector or matrix, then the corresponding Banach space can also be formed by replacing the absolute value j j of each component or function with its spatially normed component or function. For example, consider a vector space with all sequences in the form of
x = (x0 x1 : : :)
where each component xi is a k m matrix and each element of xi is bounded. Then xi is bounded in any matrix norm, and the vector space becomes a Banach space if the following norm is de ned kxk1 := sup i (xi ) where i (xi ) := kxi k is any matrix norm. This space will also be denoted by l1 . Let V1 and V2 be two vector spaces and let T be an operator from S V1 into V2 . An operator T is said to be linear if for any x1 x2 2 S and scalars 2 C , the following holds: T ( x1 + x2 ) = (Tx1 ) + (Tx2 ): Moreover, let V0 be a linear subspace in V1 . Then the operator T0 : V0 7;! V2 de ned by T0 x = Tx for every x 2 V0 is called the restriction of T to V0 and is denoted as T jV0 = T0 . On the other hand, a linear operator T : V1 7;! V2 coinciding with T0 on V0 V1 is x1 : x 2 C n called an extension of T0. For example, let V0 := V1 = C n+m , 1 0 and let T = A1 A2 2 C (n+m) (n+m) T0 = A1 0 2 C (n+m) (n+m) : 0 A3 0 0 Then T jV0 = T0 .
i
De nition 4.1 Two normed spaces V1 and V2 are said to be linearly isometric, denoted
by V1 = V2 , if there exists a linear operator T of V1 onto V2 such that
kTxk = kxk
for all x in V1 . In this case, the mapping T is said to be an isometric isomorphism.
Recall the inner product of vectors de ned on a Euclidean space C n :
4.2 Hilbert Spaces
hx yi := x y =
n X i=1
2x 3 2y 3 1 1 xi yi 8x = 6 ... 7 y = 6 ... 7 2 C n : 4 5 4 5
xn yn
94
PERFORMANCE SPECIFICATIONS
Note that many important metric notions and geometrical properties such as length, distance, angle, and the energy of physical systems can be deduced from this inner product. For instance, the length of a vector x 2 C n is de ned as
kxk := hx xi and the angle between two vectors x y 2 C n can be computed from yi cos \(x y) = khx kyk \(x y) 2 0 ]: xk
The two vectors are said to be orthogonal if \(x y) = 2 . We now consider a natural generalization of the inner product on C n to more general (possibly in nite dimensional) vector spaces. De nition 4.2 Let V be a vector space over C . An inner product on V is a complex valued function, h i : V V 7;! C such that for any x y z 2 V and 2C (i) 1 hx y + z i = hx yi + hx z i (ii) hx yi = hy xi (iii) hx xi > 0 if x 6= 0. A vector space V with an inner product is called an inner product space. p It is clear that the inner product de ned above induces a norm kxk := hx xi, so that the norm conditions in Chapter 2 are satis ed. In particular, the distance between vectors x and y is d(x y) = kx ; yk. Two vectors x and y in an inner product space V are said to be orthogonal if hx yi = 0, denoted x ? y. More generally, a vector x is said to be orthogonal to a set S V , denoted by x ? S , if x ? y for all y 2 S . The inner product and the innerproduct induced norm have the following familiar properties. Theorem 4.1 Let V be an inner product space and let x y 2 V . Then (i) jhx yij kxk kyk (CauchySchwarz inequality). Moreover, the equality holds if and only if x = y for some constant or y = 0. (ii) kx + yk2 + kx ; yk2 = 2 kxk2 + 2 kyk2 (Parallelogram law) .
(iii) kx + yk2 = kxk2 + kyk2 if x ? y.
1 This is the other way round to the usual mathematical convention since we want to have h rather than for 2 C n.
y x x y x y
p
i=
x y
4.2. Hilbert Spaces
95
A Hilbert space is a complete inner product space with the norm induced by its inner product. Obviously, a Hilbert space is also a Banach space. For example, C n with the usual inner product is a ( nite dimensional) Hilbert space. More generally, it is straightforward to verify that C n m with the inner product de ned as
hA B i := Trace A B =
n m XX i=1 j =1
aij bij 8A B 2 C n m
is also a ( nite dimensional) Hilbert space. Here are some examples of in nite dimensional Hilbert spaces: l2 (;1 1): l2 (;1 1) consists of the set of all real or complex square summable sequences x = (: : : x;2 x;1 x0 x1 x2 : : :) i.e,
1 X
with the inner product de ned as
i=;1
jxi j2 < 1
1 X
i=;1
hx yi :=
xi yi
for x y 2 l2 (;1 1). The subspaces l2 (;1 0) and l2 0 1) of l2 (;1 1) are de ned similarly and consist of sequences of the form x = (: : : x;2 x;1 ) and x = (x0 x1 x2 : : :), respectively. Note that we can also de ne a corresponding Hilbert space even if each component xi is a vector or a matrix in fact, the following inner product will su ce:
hx yi :=
1 X
i=;1
Trace(xi yi ):
L2 (I ) for I R: L2 (I ) consists of all square integrable and Lebesgue measurable functions de ned
on an interval I
R
with the inner product de ned as
hf gi := f (t) g(t)dt
for f g 2 L2 (I ). Similarly, if the function is vector or matrix valued, the inner product is de ned correspondingly as
I
Z
hf gi := Trace f (t) g(t)] dt:
I
Z
96
PERFORMANCE SPECIFICATIONS
Some very often used spaces in this book are L2 0 1) L2 (;1 0] L2 (;1 1). More precisely, they are de ned as L2 = L2 (;1 1): Hilbert space of matrixvalued functions on R, with inner product
hf gi :=
Z1
;1
Trace f (t) g(t)] dt:
L2+ = L2 0 1): subspace of L2 (;1 1) with functions zero for t < 0. L2; = L2 (;1 0]: subspace of L2 (;1 1) with functions zero for t > 0.
Let H be a Hilbert space and M H a subset. Then the orthogonal complement of M , denoted by H M or M ?, is de ned as
M ? = fx : hx yi = 0 8y 2 M x 2 Hg :
It can be shown that M ? is closed (hence M ? is also a Hilbert space). For example, let M = L2+ L2 , then M ? = L2; is a Hilbert space. Let M and N be subspaces of a vector space V . V is said to be the direct sum of M and N , written V = M N , if M \ N = f0g, and every element v 2 V can be expressed as v = x + y with x 2 M and y 2 N . If V is an inner product space and M and N are orthogonal, then V is said to be the orthogonal direct sum of M and N . As an example, it is easy to see that L2 is the orthogonal direct sum of L2; and L2+ . Similarly, l2 (;1 1) is the orthogonal direct sum of l2 (;1 0) and l2 0 1). The following is a version of the socalled orthogonal projection theorem:
Theorem 4.2 Let H be a Hilbert space, and let M be a closed subspace of H. Then for each vector v 2 H, there exist unique vectors x 2 M and y 2 M ? such that v = x + y, i.e., H = M M ?. Moreover, x 2 M is the unique vector such that d(v M ) = kv ; xk. Let H1 and H2 be two Hilbert spaces, and let A be a bounded linear operator from H1 into H2 . Then there exists a unique linear operator A : H2 7;! H1 such that for all x 2 H1 and y 2 H2 hAx yi = hx A yi:
A is called the adjoint of A. Furthermore, A is called selfadjoint if A = A . Let H be a Hilbert space and M H be a closed subspace. A bounded operator P mapping from H into itself is called the orthogonal projection onto M if P (x + y) = x 8 x 2 M and y 2 M ? :
4.3. Hardy Spaces H2 and H1
97
4.3 Hardy Spaces H2 and H1
Let S
C
be an open set, and let f (s) be a complex valued function de ned on S :
A well known property of the analytic functions is the socalled Maximum Modulus Theorem. Theorem 4.3 If f (s) is de ned and continuous on a closedbounded set S and analytic on the interior of S , then the maximum of jf (s)j on S is attained on the boundary of S , i.e., max jf (s)j = s2@S jf (s)j max s2S
where @S denotes the boundary of S . Next we consider some frequently used complex (matrix) function spaces.
f (s) : S 7;! C : Then f (s) is said to be analytic at a point z0 in S if it is di erentiable at z0 and also at each point in some neighborhood of z0 . It is a fact that if f (s) is analytic at z0 then f has continuous derivatives of all orders at z0 . Hence, a function analytic at z0 has a power series representation at z0 . The converse is also true, i.e., if a function has a power series at z0 , then it is analytic at z0 . A function f (s) is said to be analytic in S if it has a derivative or is analytic at each point of S . A matrix valued function is analytic in S if every element of the matrix is analytic in S . For example, all real rational stable transfer matrices are analytic in the righthalf plane and e;s is analytic everywhere.
L2 (j R) Space L2 (j R) or simply L2 is a Hilbert space of matrixvalued (or scalarvalued) func
tions on j R and consists of all complex matrix functions F such that the integral below is bounded, i.e.,
Z1
;1
Trace F (j!)F (j!)] d! < 1:
The inner product for this Hilbert space is de ned as 1 Z 1 Trace F (j!)G(j!)] d! hF Gi := 2 ;1 for F G 2 L2 , and the inner product induced norm is given by
kF k2 := hF F i:
For example, all real rational strictly proper transfer matrices with no poles on the imaginary axis form a subspace (not closed) of L2 (j R) which is denoted by RL2 (j R) or simply RL2 .
p
1 Trace F (j!)F (j!)] d!: ? H2 Space ? H2 is the orthogonal complement of H2 in L2 . Recall the fact that a function in L2 space in the time domain admits a bilateral Laplace (or Fourier) transform.e. if g(t) 2 L2 (.1 1) and if its Fourier (or bilateral Laplace) transform is G(j!) 2 L2 (j R). H. is denoted by RH2 . are usually called Hardy spaces named after the mathematician G. Hardy (hence the notation of H). 3 See Francis 1987]. The corresponding norm is de ned as >0 Z1 kF k2 := sup 21 Trace F ( + j!)F ( + j!)] d! : 2 . case. then kGk2 = kgk2 : 2 The H space and H space de ned below together with the H spaces.1 1) = L2 (j R) L2 0 1) = H2 ? L2 (. The real rational subspace of H2 . then G 2 H2 and ? G 2 H2 . The L2 spaces de ned above in the frequency domain can be related to the L2 spaces de ned in the time domain. The real rational subspace ? of H2 . we can compute the norm for H2 just as we do for L2 .98 PERFORMANCE SPECIFICATIONS H2 Space2 H2 is a (closed) subspace of L2 (j R) with matrix functions F (s) analytic in Re(s) > 0 (open righthalf plane). and real rational transfer matrix. stable. will be denoted by RH? . which will not be 1 p 2 introduced in this book. which consists of all strictly proper rational transfer matrices with all poles in the open right half plane. Most of our study in this book will be focused on the real rational Hence. which consists of all strictly proper and real rational stable transfer matrices. p . i.1 0] = H2 : As a result. it can be shown that this bilateral Laplace (or Fourier) transform yields an isometric isomorphism between the L2 spaces in the time domain and the L2 spaces in the frequency domain (this is what is called Parseval's relations or Plancherel Theorem in complex analysis): L2 (. In fact. It is easy to see that if G 2 is a strictly proper. the (closed) subspace of functions in L2 that are analytic in the open lefthalf plane.. 1.1 It can be shown3 that kF k2 = 21 2 Z1 .
? L2 (. De ne an orthogonal projection P+ : L2 (.1 0] ? (or L2 (j R) onto H2 ).1 1) onto L2 (. Similarly.! L2 0 1) such that. Laplace Transform L2 0 1) H2 Inverse Transform P+ 6 6 P+ L2 (.1. P+ will also be used to denote the projection from L2 (j R) onto H2 . L1 (j R) Space L1 (j R) or simply L1 is a Banach space of matrixvalued (or scalarvalued) functions that are (essentially) bounded on j R. Laplace Transform Inverse Transform  L2 (j R) P. with norm kF k1 := ess sup F (j!)] : ! 2R . whenever there is no confusion.1 1) 7.1 1).1 1) P.4.1: Relationships among function spaces Other classes of important complex matrix functions used in this book are those bounded on the imaginary axis. we have g(t) = P+ f (t) with g(t) := f (t) for t < 0: 0 for t 0 In this book.1 0] Laplace Transform Inverse Transform  ? ? H2 Figure 4.3. de ne P. Hardy Spaces H2 and H1 99 Hence. the notation of functions in the time domain and in the frequency domain will be used interchangeably. Then the relationships between L2 spaces and H2 spaces can be shown as in Figure 4. for any function f (t) 2 L2 (. as another orthogonal projection from L2 (.
The real rational subspace of H1 is denoted by RH. denoted by RL1 (j R) or simply RL1 . The H1 norm is de ned as kF k1 := sup F (s)] = ess sup F (j!)] : Re(s)<0 The second equality can be regarded as a generalization of the maximum modulus theorem for matrix functions. consists of all rational proper transfer matrices with no poles on the imaginary axis. . then G(s)H2 := fG(s)f (s) : f (s) 2 H2 g H2 . . then G(s)H2 := fG(s)f (s) : f (s) 2 H2 g H2 . The L1 space in the time domain is usually used to denote signals. then G(s)L2 := fG(s)f (s) : f (s) 2 L2 g L2 .1 The notation for L1 is somewhat unfortunate it should be clear to the reader that the L1 space in the time domain and in the frequency domain denote completely di erent spaces. (ii) if G(s) 2 H1 . while the L1 space in the frequency domain is usually used to denote transfer functions and operators. H1 Space H1 is a (closed) subspace in L1 with functions that are analytic in the open righthalf plane and bounded on the imaginary axis. The real rational subspace of H1 is denoted by RH1 which consists of all proper and real rational stable transfer matrices. H1 Space .100 PERFORMANCE SPECIFICATIONS The rational subspace of L1 .3 A transfer matrix G(s) 2 H1 is usually said to be antistable or anti causal. lefthalf plane and bounded on the imaginary axis. ! 2R . which consists of all proper 1 rational transfer matrices with all poles in the open right half plane. Then a multiplication operator is de ned as MG : L2 7. Remark 4. Some facts about L1 and H1 functions are worth mentioning: (i) if G(s) 2 L1 .! L2 MG f := Gf: . H1 is a (closed) subspace in L1 with functions that are analytic in the open . The H1 norm is de ned as kF k1 := sup F (s)] = ess sup F (j!)] : Re(s)>0 ! 2R . De nition 4. ? ? ? (ii) if G(s) 2 H1 . ~ Let G(s) 2 L1 be a p q transfer matrix.
write v1 (j!0 ) as 2 ej 3 6 1 ej 7 v1 (j!0 ) = 6 . G(j!0 )] = kGk1 and denote the singular value decomposition of G(j!0 ) by G(j!0 ) = u1 (j!0 )v1 (j!0 ) + r X i=2 i ui (j!0 )vi (j!0 ) where r is the rank of G(j!0 ) and ui vi have unit length. Then kMGk = kGk1 .! Lp 2 2 i. f is a qdimensional vector function with each component in L2 .4. we have kMGk = sup fkGf k2 : f 2 L2 kf k2 1g : First we see that kGk1 is an upper bound for the operator norm: Z1 f (j!)G (j!)G(j!)f (j!) d! kGf k2 = 21 2 kGk2 21 kf (j!)k2 d! 1 . A useful fact about the multiplication operator is that the norm of a matrix G in L1 equals the norm of the corresponding multiplication operator. ~ Proof. 7 6 2.. Remark 4..2 It is also true that this operator norm equals the norm of the operator ? restricted to H2 (or H2 ). 7 4 5 1 2 q ej q . i.3. By de nition. However. i. If !0 < 1. Theorem 4.1 = kGk2 kf k2: 1 2 . rst choose a frequency !o where G(j!)] is maximum. A more accurate description of the above operator should be MG : Lq 7. we shall suppress all dimensions in this book and assume all objects have compatible dimensions. we have assumed that f has a compatible dimension.1 Z1 To show that kGk1 is the least upper bound..4 Let G 2 L1 be a p q transfer matrix. Hardy Spaces H2 and H1 101 In writing the above mapping.e.e. kMGk = kMGjH2 k := sup fkGf k2 : f 2 H2 kf k2 1g : This will be clear in the proof where an f 2 H2 is constructed. It is easy to verify that the adjoint operator MG = MG .e..
Let u(t) be a function of time. This Fourier transform is called the spectral density of u. . !o j < or j! + !oj < 0 otherwise where is a small positive number and c is chosen so that f has unit 2norm. then G(j!0 ) is real and v1 is real. It is easy to see from the de nition that Ruu ( ) = Ruu (..4 Power and Spectral Signals In this section. 2 4.T u(t + )u (t)dt if the limit exists and is nite for all . j!0 i=\ + j! and let f be given by i 0 2 6 6 f (s) = 6 6 4 1 1 +s 1 .s q q +s q . Then^ c= kGf k2 2 = 1 2 G(.1 Ruu ( )e. Assume further that the Fourier transform of the signal's autocorrelation matrix function exists (and may contain impulses). Now let i 0 be such that i . De ne its autocorrelation matrix as Ruu ( ) := Tlim 21 !1 T ZT . i. p =2 . denoted Suu (j!): Suu (j!) := Z1 .s 2 2 +s 2 . we introduce two additional classes of signals that have been widely used in engineering.j! d : .s where a scalar function f^ is chosen so that .. In this case.e. 0) and q is the column dimension of G. These classes of signals have some nice statistical and frequency domain representations.j!o )]2 + G(j!o )]2 G(j!o )]2 = kGk2 : 1 h i If !0 = 1. This in turn implies that f has unit 2norm. the conclusion follows by letting f (s) = v1 f^(s).102 PERFORMANCE SPECIFICATIONS where i 2 R is such that i 2 . ). 3 7 7 f^(s) 7 7 5 jf^(j!)j = c if j! .
if the power spectral density function Suu (j!) exists (note that Suu (j!) need not be bounded and may include impulses). as does an L2 signal. then kukP kuk1. a timelimited signal has zero power. Similarly.1 We will call signal u(t) a power signal if the autocorrelation matrix Ruu ( ) exists and is nite for all . A signal u(t) is said to have bounded power spectral density if kSuu (j!)k1 < 1. R . The set of all signals having nite power will be denoted by P . These properties are useful in establishing some input and output relationships in the next section.4. ) and Szz (j!) = G(j!)Suu (j!)G (j!).1 This expression implies that. if u 2 P . Then Rzz ( ) = g( ) Ruu ( ) g (. and output z (t). We note that if u 2 P and ku(t)k1 := supt ku(t)k < 1. The power seminorm of a signal can also be computed from its spectral density function Z1 2 = 1 Trace Suu (j!)]d!: kukP 2 .4. P could be used to model signals whose spectrum is not known but which are nite in power. For example. and moreover. input u(t). not every signal having nite 1norm is a power signal since the limit in the de nition may not exist. The set of signals having bounded spectral density is denoted as S := fu(t) 2 Rm : kSuu (j!)k1 < 1g: p The quantity kuks := kSuu (j!)k is called the spectral density norm of u(t). However. The set S can be used to model signals with xed spectral characteristics by passing white noise signals through a weighting lter. Power and Spectral Signals 103 Thus Ruu ( ) can be obtained from Suu (j!) by performing an inverse Fourier transform: Z1 Ruu ( ) := 21 Suu (j!)ej! d!: . let 2k 2k+1 for k = 0 1 2 : : : u(t) = 1 2 < t <:2 0 otherwise T Then limT !1 21 . Note also that power signals are persistent T signals in time such as sines or cosines clearly. Suu is strictly proper in the sense that Suu (1) = 0. Now let G be a linear system transfer matrix with convolution kernel g(t). The power of the signal is de ned as 1 Z T ku(t)k2 dt = pTrace R (0)] kukP = Tlim 2T uu !1 . not a norm. Thus k kP is only a seminorm.T u2 dt does not exist.T s where k k is the usual Euclidean norm and the capital script P is used to di erentiate this power seminorm from the usual Lebesgue Lp norm.
for example. In this section we are interested in answering the following question: if we know how \big" the input (disturbance) is. G(s) = 4 Gp1 (s) G1q (s) Gpq (s) . 7 = 6 .. Note that the (t) in this table denotes the unit impulse and u0 2 Rq is a constant vector indicating the direction of the input signal. how \big" is the output going to be for a given stable dynamical system? Consider a qinput and poutput linear nite dimensional system as shown in the following diagram with input u. 7 g(t) = 4 . measures the size of the output for a given input u. 5 4 . ... These norms can determine the achievable performance of the system for di erent classes of input signals The various input/output relationships. 3 2 G (s) 3 7 = 6 1. 5 gp1 (t) gpq (t) gp (t) 0 2 G (s) 6 11. and transfer matrix G 2 RH1 : z G u We will further assume that G(s) is strictly proper.. under a set of possible disturbances and system parameter variations.104 PERFORMANCE SPECIFICATIONS 4.. are summarized in two tables below. Zt z (t) = g(t . . loosely speaking.. In the timedomain an inputoutput model for such a system has the form of a convolution equation. Let the convolution kernel and the corresponding transfer matrix be partitioned as 2 g (t ) 3 2 g (t) 3 g1q (t) 11 1 6 .1. which...e. i. 7 5 4 5 Gp (s) where gi (t) is a qdimensional row vector of the convolution kernel and Gi (s) is a row vector of the transfer matrix. output z . Table 4. given di erent classes of input signals. G(1) = 0 although most of the results derived here hold for the nonstrictly proper case. Now if G is considered as an operator from the input space to the output space.1 summarizes the results for xed input signals. ..e. z=g u i.5 Induced System Gains Many control engineering problems involve keeping some system signals \small" under various conditions. We now prove Table 4. )u( ) d where the p q real matrix g(t) is the convolution kernel.. then a norm is induced on G.
5. On the other hand. u = u0 (t): If u(t) = u0 (t). Induced System Gains Input u(t) Output z (t) 105 kz k2 = kGu0 k2 = kgu0k2 u(t) = u0 (t) u0 2 Rq kz k1 = kgu0k1 kz kP = 0 kz k2 = 1 u(t) = u0 sin !0 t u0 2 Rq lim sup max jzi (t)j = kG(j!0 )u0 k1 = max jGi (j!0 )u0 j i i t!1 lim sup kz (t)k = kG(j!0 )u0 k t!1 1 kz kP = p kG(j!0 )u0 k 2 Table 4. then z (t) = g(t)u0.4.1 = Tlim 21 kGk2 ku0k2 = 0: 2 !1 T kz k2 u = u0 sin !0 t: With the input u(t) = u0 sin(!0 t).1: Output norms with xed inputs Proof.T 0 !1 Z1 Tracefg (t)g(t)g dt ku0 k2 lim 1 T !1 2T . the ith output as t ! 1 is zi (t) = jGi (j!0 )u0 j sin !0 t + argfGi (j!0 )u0 g]: (4:1) . But by Parseval's relation. kgu0k2 = kGu0k2 . so kz k2 = kgu0k2 and kz k1 = kgu0k1 . 1 Z T u g (t)g(t)u dt 0 P = Tlim 2T .
2.! L2 ) has been shown in Section 4.T 2 i=1 p 1X = 2 jGi (j!0 )u0 j2 = 1 kG(j!0 )u0 k2 : 2 i=1 2 It is interesting to see what this table signi es from the control point of view. the system's transfer function does not have a zero in every channel in the input direction at the frequency of excitation. Note that lim sup kz (t)k = kG(j!0 )u0 k t!1 for any given !0 and u0 2 Rq . that is if u0 can be chosen to be any direction. for instance.T i=1 Z T 1 . Table 4. Furthermore. Now we give a proof for Table 4. This interpretation enables us to consider the control system in the frequency domain even though the speci cations are given in the time domain.106 PERFORMANCE SPECIFICATIONS The 2norm of this signal is in nite as long as Gi (j!0 )u0 6= 0.e. Then ZTX p kz k2 = Tlim 21 jG (j! )u j2 sin2 (!0 t + i ) dt P !1 T ..1) equals jGi (j!0 )u0 j. To focus our discussion.T i=1 i 0 0 ZT p X 2 lim 1 = jGi (j!0 )u0 j T !1 2T sin2 (!0 t + i ) dt . Note that the rst row (L2 7.2 lists the maximum possible system gains when the input signal u is not required to be a xed signal instead it can be any signal in a unit ball in some function space. Hence lim sup max jzi (t)j = max jGi (j!0 )u0 j = kG(j!0 )u0 k1 i i and t!1 vp uX u lim sup kz (t)k = t jGi (j!0 )u0 j2 = kG(j!0 )u0 k : t!1 i=1 Finally. and z (t) is the tracking error. then we would require that (G(j!0 )) be small. Then we say that the system has good tracking behavior if z (t) is small in some sense. we then would require that (G(j!0 )) be small at all those frequencies. in addition. cos 2(!0t + i ) p X dt = jGi (j!0 )u0 j2 Tlim 21 !1 T . if. i. . let us assume that u(t) = u0 sin !0 t is a disturbance or a command signal on a feedback system. Now if we want to track signals from various channels. lim supt!1 kz (t)k is small. let i := arg Gi (j!)u0 . we want to track signals of many di erent frequencies.3. The amplitude of the sinusoid (4.
1 L1 L1 kuk1 = sup max jui (t)j i t kuk1 = sup ku(t)k t Z1 0 kg(t)k dt Table 4.1 .! P : By de nition.1 TracefSuu (j!)gd! TracefSuu (j!)gd! Z1 . S 7. then Szz (j!) = G(j!)Suu (j!)G (j!): Now suppose G(j!0 )] = kGk1 and take a signal u such that Suu (j!0 ) = I . Induced System Gains Input u(t) Output z (t) Signal Norms 107 Induced Norms L2 S S P L2 S P P kuk2 = 2 Z1 0 kuk2 dt kGk1 kGk1 kGk2 kGk1 max kgi k1 i kuk2 = kSuu k1 S kuk2 = 21 P kuk2 = 21 P Z1 . Then kSzz (j!)k1 = kGk2 : 1 S 7.2: Induced System Gains Proof.! S : If u 2 S .4. we have kz k2 1 Z 1 TracefG(j!)S (j!)G (j!)g d!: uu P = 2 .5.
s 2 2 +s . Now let i 0 be such that i . then Z1 2 = 1 kz kP 2 TracefG(j!)Suu (j!)G (j!)g d! . Then Z1 2 = 1 kz kP 2 TracefG(j!)G (j!)g d! = kGk2 : 2 .e.! P : If u is a power signal. . we get that kz kP kGk1 kukP : To achieve the equality.108 PERFORMANCE SPECIFICATIONS Now let u be white with unit spectral density. assume that !0 is such that G(j!0 )] = kGk1 and denote the singular value decomposition of G(j!0 ) by G(j!0 ) = u1 (j!0 )v1 (j!0 ) + If !0 < 1. i.s q q +s q .. write v1 (j!0 ) as r X i=2 i ui (j!0 )vi (j!0 ) where r is the rank of G(j!0 ) and ui vi have unit length. j!0 i=\ + j! and let the input u be generated from passing u through a lter: ^ i 0 2 6 6 u(t) = 6 6 4 1 1 1 +s 2 .. Suu = I .s . 0) and q is the column dimension of G. 2 ej 3 6 1ej 7 v1 (j!) = 6 . 7 6 2. 3 7 7 u(t) 7^ 7 5 . 7 4 5 1 2 q ej q where i 2 R is such that i 2 .1 P 7.1 and immediately..
3 7 7 7 7 5 kukP = 1: 1 kz k2 = 2 G(j!o )]2 + 1 G(.j! 2 2 +j! 2 . if !0 = 1. q .. The conclusion follows by letting u = v1 u(t) and !o ! 1. First of all. (! .j! 2 2 +j! 2 and it is easy to show Finally. Induced System Gains where Then Ruu ( ) = cos(!o ). .j! 1 1 1 +j! .j!o)] = kGk2 : 1 Similarly.! L1 : 1. )j d jgij ( )j d kuk1 kz k1 := max sup jzi (t)j i t max kgi k1 kuk1 : i ..4.5.j!o )]2 P 2 2 = G(. !o ) + (! + !o)] : . .j! .j! q q +j! 3 2 7 7 S (j!) 6 7 uu 6 7 ^^ 6 6 5 4 1 1 +j! 1 . then G(j!) is real. ) d Zt 0 0 j =1 kgi k1 kuk1 jgij ( )j d kuk1 Z 1X q 0 j =1 jgi ( )u(t . so ^^ Also. Then 109 u(t) = 2 sin(!o t): ^ p kukP = Ruu (0) = 1: ^ ^^ 2 6 6 Suu (j!) = 6 6 4 Suu (j!) = ^^ . ^ L1 7. maxi kgik1 is an upper bound on the 1norm/1norm system gain: jzi (t)j = = and Zt 0 Z tX q gi ( )u(t .j! q q +j! q .
Then kuk1 = 1 and z1 (t0 ) = = Zt 0 0 Zt X q 0 g1 ( )u(t0 . If ku(t)k1 := supt ku(t)k. there exists = diag( with 1 2 1 2 : : : n) 0 A + A + C C = 0: ::: n 0. )d kg( )k ku(t . i.110 PERFORMANCE SPECIFICATIONS That maxi kgi k1 is the least upper bound can be seen as follows: Assume that the maximum maxi kgi k1 is achieved for i = 1. Let t0 be given and set u(t0 . ) := sign(g1 ( )) 8 : Note that since g1( ) is a vector function. let t0 ! 1. and we have kz k1 = kg1 k1 . such that A + A + BB = 0 Then we have the following theorem.. ) d 0 j =1 jg1j ( )j d = kg1 k1 . Suppose A G(s) = C B 2 RH1 0 is a balanced realization. 2. the sign function sign(g1 ( )) is a componentwise operation. kz k1 R 1 kg( )k d kuk .e. then kz (t)k = Zt 0 0 Z t0 Zt g( )u(t . Z 1X q t0 j =1 jg1j ( )j d : Hence. )k d kg( )k d kuk1 : 2 And. therefore. . 1 0 Next we shall derive some simple and useful bounds for the H1 norm and the L1 norm of a stable system.
we have kGk1 = Z1 Z1 0 Re(s)>0 g(t)e. Remark 4. The inequality kGk1 follows from the Nehari Theorem of Chapter 8.5. use of the balanced realization does make the proof simple. It is easy to verify that ui uj = ij = k k 2 i k2 = 2 i k2 Z1 0 0 = Z1 ui eAt=2 BB eA t=2 ui dt = 2ui ui = 2 i ui eA t=2 C CeAt=2 ui dt = 2ui ui = 2 i : Using these two equalities.5 1 111 kGk1 Z1 0 kg(t)k dt 2 n X i=1 i where g(t) = CeAt B . ~ Proof. 1 We will now show the other inequalities. Induced System Gains Theorem 4. Then n 1 if i = j and X u u = I: i i 0 if i 6= j i=1 De ne i (t) = ui eAt=2 B and i (t) = CeAt=2 ui .st dt Re(s)>0 0 Z1 0 kg(t)k dt: To prove the last inequality. let ui be the ith unit vector. However. we have Z1 0 kg(t)k dt = Z1 X n n X i=1 0 i=1 i i dt n XZ 1 i=1 0 n X 2 i: i=1 k i i k dt k i k2 k i k2 . Since G(s) := Z1 0 g(t)e.3 It should be clear that the inequalities stated in the theorem do not depend on a particular state space realization of G(s).4.st dt Re(s) > 0 sup sup by the de nition of H1 norm.st dt g(t)e.
112 PERFORMANCE SPECIFICATIONS 2 It should be clear from the above two tables that many system performance criteria can be stipulated as requiring a certain closed loop transfer matrix have small H2 norm or H1 norm or L1 norm.1 TracefG (j!)G(j!)g d! sZ 1 . Suppose G is strictly proper. The computation of a RH2 transfer matrix norm is particularly simple.6 Computing L2 and H2 Norms kGk2 := = Let G(s) 2 L2 and recall that the L2 norm of G is de ned as s Z1 1 2 = kgk2 .6 Consider a transfer matrix A G(s) = C B 0 .e. G(1) = 0. By the residue theorem. in principle. kGk2 equals the sum 2 of the residues of TracefG (s)G(s)g at its poles in the left halfplane.. It is easy to see that the L2 norm de ned above is nite i the transfer matrix G is strictly proper.1 I = 21 j TracefG (s)G(s)g ds: The last integral is a contour integral along the imaginary axis. then the H1 norm performance is also satis ed. be computed from its de nition or from the method suggested above. this also applies to H2 functions). if L1 performance is satis ed.1 Tracefg (t)g(t)g dt where g(t) denotes the convolution kernel of G. we will generally assume that the transfer matrix is strictly proper whenever we refer to the L2 norm of G (of course. Although kGk2 can. Hence. it is useful in many applications to have alternative characterizations and to take advantage of the state space representations of G. and then we have Z1 TracefG (j!)G(j!)g d! kGk2 = 21 2 . i. We will be most interested in H2 and H1 performance in this book. Moreover. Lemma 4. and around an in nite semicircle in the left halfplane the contribution to the integral from this semicircle equals zero because G is strictly proper. 4. One straightforward way of computing the L2 norm is to use contour integral.
with G+ 2 RH2 and G. Since G is stable. using state space approach. see Chen and Francis 1992] for an application of this norm in sampleddata control. i=1 . 2 RH? .1 (G) = Ce B t < 0 0 t 0 and kGk2 2 = = Z1 Z01 0 Tracefg (t)g(t)g dt = TracefB Z1 0 Tracefg(t)g(t) g dt eA t C CeAt B g dt = Z1 0 TracefCeAt BB eA t C g dt: The lemma follows from the fact that the controllability Gramian of (A B ) and the observability Gramian of (C A) can be represented as Lo = Z1 0 eA t C CeAt dt Lc = Z1 0 eAt BB eA t dt which can also be obtained from ALc + LcA + BB = 0 A Lo + LoA + C C = 0: 2 To compute the L2 norm of a rational transfer function. Assume D = 0. G(s) 2 L2 . Then 2 2 = k G(s)] k2 + k G(s)] k2 kGk2 + 2 . we have At g(t) = L. k2 = k G(. Then we have 113 (4:2) where Lo and Lc are observability and controllability Gramians which can be obtained from the following Lyapunov equations ALc + LcA + BB = 0 A Lo + LoA + C C = 0: kGk2 = trace(B Lo B ) = trace(CLc C ) 2 Proof. 2 where k G(s)]+ k2 and k G(s)].6. Apply the impulsive input (t)ei ( (t) is the unit impulse) and denote the output by zi (t)(= g(t)ei ).s)]+ k2 can be computed using the above lemma.4. and then zi 2 L2+ and m X kGk2 = kzi k2 : 2 2 Note that this characterization of the H2 norm can be appropriately generalized for nonlinear time varying systems. Computing L2 and H2 Norms with A stable. Let ei denote the ith standard basis vector of Rm where m is the input dimension of the system. Still another useful characterization of the H2 norm of G is in terms of hypothetical inputoutput experiments. Let G(s) = G(s)]+ + G(s)].
as in the L2 case.1 H := . A control engineering interpretation of the in nity norm of a scalar transfer function G is the distance in the complex plane from the origin to the farthest point on the Nyquist plot of G. Equivalently. 2 . (j!) > 0 for all ! 2 R if and only if (j!) is nonsingular for all ! 2 R f1g.CA(+ + DR. i. Lemma 4.. f!1 Then an estimate for kGk1 is 1 k N !N g: max fG(j!k )g: This value is usually read directly from a Bode singular value plot.1 R .1 D C BR. Then it is clear that kGk1 < if and only if (j!) > 0 for all ! 2 R.C DR.1 (s) = 4 R. It is easy to compute by some simple algebra that Proof. (s) has no imaginary axis zero. Hence the 1 of a transfer function can in principle be obtained graphically.114 PERFORMANCE SPECIFICATIONS 4. The L1 norm can also be computed in state space if G is rational.1B H BR.1 (s) has no imaginary axis pole. To get an estimate.1 . set up a ne grid of frequency points. how to compute the 1 norm of an L1 transfer matrix. G (s)G(s).7 Let > 0 and A B G(s) = C D 2 RL1 : (4:3) Then kGk1 < if and only if (D) < and H has no eigenvalues on the imaginary axis where BR. Let G(s) 2 L1 and recall that the L1 norm of a transfer function G is de ned as kGk1 := ess sup fG(j!)g: The computation of the L1 norm of G is complicated and requires a search.1 D C R. Let (s) = 2I .1 3 5: . D D.7 Computing L1 and H1 Norms ! We shall rst consider. . and it also appears as the peak value on the Bode magnitude plot of jG(j!)j.4) 1 I and R = 2I . Since (1) = R > 0 and since (j!) is a continuous function of !.(A + BR.e.BD C ) (4.1D )C .
we must use some type of iterative algorithm. Of course.1 B H ).1 lable mode of (H . 2 I . Thus L1 norm computation requires a search. which does not. a contradiction will also be arrived if j!0 is assumed to be an uncontrolBR. as can the test for whether (M ) < (e. but the value of (M ) cannot.7 suggests the following bisection algorithm to compute RL1 norm: (a) select an upper bound u and a lower bound l such that l kGk1 u (b) if ( u . Similarly. Then there exists an x0 = x1 6= 0 such that x 2 Hx0 = j!0 x0 R.C Cx1 D Cx1 + B x2 = 0: Since A has no imaginary axis eigenvalues. and hence the realization has no unobservable modes on the imaginary axis.1 ).4. Then j!0 must be either an unobservable mode of BR.1 ). M M > 0).1 ( R. In principle. in contrast to L2 (H2 ) norm computation. l )= l speci ed level. kM k2 2 2 can be computed exactly with a nite number of operations. Otherwise go to next step (c) set = ( l + u )=2 (d) test if kGk1 < by calculating the eigenvalues of H for the given (e) if H has an eigenvalue on j R set l = otherwise set u = go back to step (b). the above algorithm applies to H1 norm computation as well. This contradicts our assumption. A somewhat analogous situation occurs for constant matrices with the norms kM k2 = trace(M M ) and kM k1 = M ].1 B x0 = 0: These equations can be simpli ed to (j!0 I .C DR.1 (s).1 D C R. over either or !. Computing L1 and H1 Norms 115 Thus the conclusion follows if the above realization has neither uncontrollable modes nor unobservable modes on the imaginary axis. Now suppose j!0 is an unobservable mode of ( R.C DR.1 D C R.1 D C R. Assume that j!0 is an eigenvalue of H but not a pole of .1 B H ) or an uncontrollable mode of (H .g.7. we have x1 = 0 and x2 = 0. To compute (M ). stop kGk ( u + l )=2. Bisection Algorithm . 2 Lemma 4. A)x1 = 0 (j!0 I + A )x2 = .
This assumption will often be made in the remainder of this book.1 G 1 < 1. The bisection H1 norm computational algorithm is rst developed in Boyd. for instance. Hence. ~ Remark 4. Naylor and Sell 1982] and Gohberg and Goldberg 1981].4 It is clear that kGk1 < i The H1 norm of a stable transfer function can also be estimated experimentally using the fact that the H1 norm of a stable transfer function is the maximum magnitude of the steadystate response to all possible unit amplitude sinusoidal input signals. It is also noted that there are other fast algorithms to carry out the above norm computation nevertheless. 4. there is no loss of generality to assume = 1. Balakrishnan. A more e cient L1 norm computational algorithm is presented in Bruinsma and Steinbuch 1990]. The system theoretical interpretations of the norms and function spaces can be found in Desoer and Vidyasagar 1975].8 Notes and References The basic concept of function spaces presented in this chapter can be found in any standard functional analysis textbook. and Kabamba 1989]. .116 PERFORMANCE SPECIFICATIONS . this bisection algorithm is the simplest.
6.7.2. the notion of internal stability is introduced and the relationship is established between the state space characterization of internal stability and the transfer matrix characterization of internal stability in section 5. we consider the mathematical formulations of optimal H2 and H1 control problems in section 5. Next. In section 5. The stable coprime factorizations of rational matrices are also introduced in section 5. system variations.4.1 discusses the necessity for introducing feedback structure and describes the general feedback conguration.5 considers feedback properties and discusses how to achieve desired performance using feedback control. The arrangement of this chapter is as follows: Section 5. the wellposedness of the feedback loop is de ned. and 117 .3. Central to these issues is the requirement to provide satisfactory performance in the face of modeling errors. these issues are generic in their relationship to control design objectives and procedures. These discussions lead to a loop shaping control design technique which is introduced in section 5. 5 5. Section 5. Although they may di er for each application and may have di erent levels of importance.1 Feedback Structure In designing control systems. Finally.Stability and Performance of Feedback Systems This chapter introduces the feedback structure and discusses its stability and performance properties. there are several fundamental issues that transcend the boundaries of speci c applications.
the next step in the controller design process is to determine what structure is necessary to achieve the desired performance.6  K u . The mere assumption of a feedback structure. we will consider the standard feedback con guration shown in Figure 5. and there are many obstacles to achieving the uncertaintyreducing bene ts of feedback. we cannot use feedback (or any other control structure) to cause our closedloop model set to be a proper subset of the openloop model set. To be of practical value. the feedback design problem centers around the tradeo involved in reducing the overall impact of uncertainty. For the moment.5 and in Chapter 6. however.1. when using feedback to reduce command/disturbance error while minimizing response degradation due to measurement noise. does not guarantee a reduction of uncertainty. Often.? ye ? en d Figure 5. then a feedback structure is required. Even worse is that the feedback system actually increases uncertainty and sensitivity in the frequency ranges where uncertainty is signi cantly large. It consists of the interconnected plant P and controller K forced by re . the loop gain must be small at those frequencies to avoid destabilizing the high frequency system dynamics. assuming we are given a model including a representation of uncertainty which we believe adequately captures the essential features of the plant.? upe di P . To focus our discussion. In particular. because of the type of sets required to reasonably model physical systems and because of the restriction that our controllers be causal. Feedback is only required when system performance cannot be achieved because of uncertainty in system characteristics. In other words. If the uncertainty is too great to achieve the desired accuracy of response. This tradeo also occurs. for example.1: Standard Feedback Con guration . Thus. The more detailed treatment of model uncertainties and their representations will be discussed in Chapter 9. Pre ltering input signals (or open loop control) can change the dynamic response of the model set but cannot reduce the e ect of uncertainty. a design technique must provide means for performing these tradeo s. Indeed. since for any reasonable model set representing a physical system uncertainty becomes large and the phase is completely unknown at su ciently high frequencies.118 STABILITY AND PERFORMANCE OF FEEDBACK SYSTEMS uncertainty. We will discuss these tradeo s in more detail later in section 5. what can be achieved with intelligent use of feedback is a signi cant reduction of uncertainty for certain signals of importance with a small increase spread over other signals. this requirement was the original motivation for the development of feedback systems.
e. plant input disturbance di . Then. y and all other signals are also wellde ned and the related transfer matrices are proper. Then the rst question one would ask is whether the feedback interconnection makes sense or is physically realizable. and di are speci ed and that the closedloop transfer matrices from them to u are respectively wellde ned and proper.K and regroup the external input signals into the feedback loop as w1 and w2 and regroup the input signals of the plant and the controller as e1 and e2 . s .2 is wellposed if and only if ^ I . d and di to u are not proper. Lemma 5. WellPosedness of Feedback Loop 119 command r. In general. n .2. all signals are assumed to be multivariable.2 and the system is wellposed if and only if the transfer matrix from w1 to e1 exists and is w2 proper. 1 di 3 3 i. n .1 P = . K (1)P (1) (5:1) is invertible.1 The feedback system in Figure 5. . since the transfer matrices from d and n to u are the same and di er from the transfer matrix from r to u by only a sign. d In order to be consistent with the notation used in the rest of the book..1 are xed real rational proper transfer matrices. However. and plant output disturbance d.s + 2 K = 1 s are both proper transfer functions. sensor noise n. the system is wellposed if and only if the transfer matrix from di to u exists and is proper. Then the feedback loop with the plant and the controller can be simply represented as in Figure 5. consider a simple example where . u = (s + 2) (r . d) . the transfer functions from the external signals r . and all transfer matrices are assumed to have appropriate dimensions. Now suppose that all the external signals r n d. Hence. we shall denote ^ K := .2 WellPosedness of Feedback Loop Assume that the plant P and the controller K in Figure 5. Furthermore. the feedback system is not physically realizable! De nition 5. To be more speci c.1 A feedback system is said to be wellposed if all closedloop transfer matrices are wellde ned and proper. 5.5.
^ Introduce realizations of P and K : ^ I . and hence wellposedness for most practical control systems is guaranteed. But this is equivalent to the condition that the constant term of the transfer function ^ I . KP )e1 = w1 + Kw2 : ^ Thus wellposedness is equivalent to the condition that (I .D is invertible: (5:5) .120 STABILITY AND PERFORMANCE OF FEEDBACK SYSTEMS w1 + . KP is invertible. 2 It is straightforward to show that (5.e e1 +6  P ^ K + e2 ?+ w2 e Figure 5. P (1)K (1) is invertible: A B P= C D ^ A ^ K= ^ C (5:3) " ^ B ^ : D # (5:4) ^ ^ Then P (1) = D and K (1) = D.K (1) is invertible (5:2) . The system in the above diagram can be represented in equation form as ^ e1 = w1 + Ke2 e2 = w2 + Pe1 : Then an expression for e1 can be obtained as ^ ^ (I . KP ).1 exists and is proper. in most practical cases we will have D = 0. .P (1) I The wellposedness condition is simple to state in terms of statespace realizations.1) is equivalent to either one of the following two conditions: ^ I . For example.2: Internal Stability Analysis Diagram Proof. wellposedness in (5.D I Fortunately.2) is equivalent to the condition that ^ I .
.1 0 C : .8) to get x =A x _ ~ _ x ^ x ^ where ^ ^ I .3 Internal Stability Consider a system described by the standard block diagram in Figure 5. i.2 The system of Figure 5. ^ ^ e1 = I . and write the state ^ equations in Figure 5.2 with given stabilizable and detectable realizations ^ ^ P and K .2 The system of Figure 5. the states (x x) go to zero from all initial ^ ^ states when w1 = 0 and w2 = 0.5. Internal Stability 121 5. assume that the realizations for P (s) and K (s) given in equations (5. In fact. ^ ~ for P and K is internally stable if and only if A is a Hurwitz matrix. respectively. It is also important and useful to characterize internal stability from the . ^ Let x and x denote the state vectors for P and K .9) De nition 5.3) and (5.e.2 and assume ^ the system is wellposed. no extra unstable modes are introduced by the realizations. not on speci c realizations of them as long as the realizations of P and K are It is routine to verify that the above de nition of internal stability depends only on both stabilizable and detectable.2 with w1 and w2 set to zero: x _ e2 _ x ^ e1 = = = = Ax + Be1 Cx + De1 ^^ ^ Ax + Be2 ^^ ^ C x + De2 : (5. To get a concrete characterization of internal stability.D .1 0 C e2 . Lemma 5.e.7) (5.D I C 0 Note that internal stability is a state space notion.4) are stabilizable and detectable.2 is said to be internally stable if the origin (x x) = (0 0) is asymptotically stable.8) (5.6) and (5.7) and (5.9) for e1 and e2 : x : x ^ Note that the existence of the inverse is guaranteed by the wellposedness condition. Furthermore.D I C 0 ~ Thus internal stability is equivalent to the condition that A has all its eigenvalues in 0 0 ~ A= A A + B B ^ 0 0 ^ the open lefthalf plane.D .3. i.6) (5. The above notion of internal stability is de ned in terms of statespace realizations ^ of P and K . Now substitute this into (5.. solve equations (5. this can be taken as a de nition of internal stability.
2 is internally stable. Then the wellposedness condition ^ ^ ^ implies that (I . As above let C D and A B be stabilizable and detectable realiza^ ^ C D ^ ^ tions of P and K . P K ).1 P K (I . P K )(1) is invertible. P K ) is invertible.P I (I .122 STABILITY AND PERFORMANCE OF FEEDBACK SYSTEMS transfer matrix point of view.D I C 0 x + w1 : x ^ w2 Now suppose that this system is internally stable. P K ). since the eigenvalues of 0 0 ~ A= A A + B B 0 ^ 0 ^ ^ ^ I . respectively. Lemma 5. DD) = (I . P K from (w1 w2 ) to (e1 e2 ) belongs to RH1 .P I 2 2 Now it is intuitively clear that if the system in Figure 5.2 are " # x _ _ x ^ y1 y2 e1 e2 = = = A 0 x + ^ x ^ 0 A C 0 x + ^ x ^ 0 C w1 + 0 I w2 I 0 B 0 ^ 0 B D 0 ^ 0 D y1 : y2 e1 e2 e1 e2 The last two equations can be rewritten as ^ ^ e1 = 0 C I .2 is described.1 ^ ).1 = I + K (I . the outputs (e1 e2) are also bounded. in term of transfer matrices.1 . .D e2 . Let y1 denote the output of P and y2 the output of K .1 0 C . by ^ e1 = w1 : I .D I C 0 are in the open lefthalf plane. Then the statespace equations for the system in Figure 5. Note that the feedback system in Figure 5. it follows that the transfer matrix from (w1 w2 ) to (e1 e2 ) given in (5.D .3 The system in Figure 5.11) is in RH1 .1 P ^ ).K . matrix (5:11) ^ ^ A B Proof. Furthermore. The following lemma shows that this idea leads to a transfer matrix characterization of internal stability. P K (I . Hence.K (5:10) e w . then for all bounded inputs (w1 w2 ).2 is internally stable if and only if the transfer ^ ^ ^ ^ ^ I . (I .
D .11) is in RH1 . Therefore.1 s+1 Then it is easy to compute 3 2 s+1 + .5.1 B 0 ~ ^ .D I ^ ^ I .D I is nonsingular. DD ^ I . P K )(1) = ^ ) is invertible. Finally.D . A). For example. Internal Stability 123 ^ Conversely. P K ) is invertible and the transfer matrix in (5.3.1 . 2 Note that to check internal stability. Now routine calculations give the transfer matrix from w1 to e1 w2 e2 in terms of the state space realizations: ^ I . (I . (s . Then. A). .D . 1: P = s.1 : . let an interconnected system transfer function be given by 1 ^ K = .D I Since the above transfer matrix belongs to RH1 . Stability cannot be concluded even if three of the four transfer matrices in (5. suppose that (I . in particular. since (A B C ) and (A B C ) are stabilizable and detectable.11) is ^ ^ in RH1 . It then follows that the eigenvalues of A are in the open lefthalf plane.s1)(s1+ 2) 7 e1 = 6 s + 2 7 w1 6 5 w2 4 s. it follows that ^ 0 C (sI .1 B 0 ~ ^ C 0 0 B ^ ^ ^ as a transfer matrix belongs to RH1 . (I .D + 0 C (sI . P K ).11) are in RH1 .D I C 0 0 B ^ I . This can also be seen by calculating the closedloop Amatrix with ^ any stabilizable and detectable realizations of P and K .s . it is necessary (and su cient) to test whether each of the four transfer matrices in (5.1 is proper which implies that (I .1 e2 s+1 s+2 s+2 which shows that the system is not internally stable although three of the four transfer functions are stable. ^ 0 0 C ~ A B B ^ 0 C 0 ~ is stabilizable and detectable.
1 2 RH1 . Then the system in Figure 5.7 The system is internally stable if and only if ^ (i) the number of open rhp poles of P (s)K (s) = nc + np ^ (ii) (s) := det(I . This is because all interconnected systems may be unavoidably subject to some nonzero initial conditions and some (possibly small) errors. ^ Corollary 5. 2 This corollary is in fact the basis for the classical control theory where the stability is checked only for one closedloop transfer function with the implicit assumption that the controller itself is stable.6 Suppose P 2 RH1 and K 2 RH1 . But this follows from ^ ^ ^ (I .2 is ^ internally stable i (I . ^ Corollary 5. we have Corollary 5. P (s)K (s)) has all its zeros in the open lefthalf plane (i. Also. i (I . To study the more general case. P K ). de ne ^ nc := number of open rhp poles of K (s) np := number of open rhp poles of P (s): Theorem 5.1 = I + (I .1 2 RH1 .1 P K ^ ^ and (I . Then the system in Figure 5.2 is internally stable ^ ). P K Proof. ^ P (s)K (s)). The necessity is obvious. it is su cient to show that ^ (I . To prove the su ciency..1 2 RH1 . there are some special cases under which determining system stability is simple.1 P 2 RH1 .e. P K ). P K ). Then the system in Figure 5.1 P K 2 RH1 . .1 It should be noted that internal stability is a basic requirement for a practical feedback system. ~ However. Internal stability guarantees that all signals in a system are bounded provided that the injected signals (at any locations) are bounded. and it cannot be tolerated in practice that such errors at some locations will lead to unbounded signals at some other locations in the closedloop system.5 Suppose P 2 RH1. (I . P K ).1 is stable). P K ).124 STABILITY AND PERFORMANCE OF FEEDBACK SYSTEMS Remark 5.2 is internally stable ^ ^ i K (I . P K ).4 Suppose K 2 RH1.
condition (i) is necessary and su cient for the internal stability. the MIMO version of the Nyquist stability theorem is obvious.1 ! 1 encircles the origin.8 (Nyquist Stability Theorem) The system is internally stable if and only if condition (i) in Theorem 5. Note that by SISO Nyquist stability theorem. P K ).5.1 = C D # where ^ ^ A B C + B D (I . (s) has all zeros in the open lefthalf plane if and only if the Nyquist plot of (j!) for .13) are equivalent to condition (i). 2 With this observation. ^ Now suppose that the system is internally stable. (0 0). Internal Stability 125 ^ ^ Proof.1 C DC ^ ^ ^ ^ 0 A B ^ D ^ B = B^ (I .1 B ^ ^ C = (I . (0 0). DD). Hence.1 : ~ It is also easy to see that A = A. This ^ (s)) must be in the lefthalf plane.1 have the following realizations: 2 ^ 3 ^ A BC BD ^ ^ 5 ^ 4 PK = 6 0 A B 7 ^ ^ C DC DD " A B ^ (I . P K has no unstable pole/zero cancelations. nk + np times in the counterclockwise direction.e. P K ).1 ! 1 encircles the origin. 2 . It is easy to show that P K and (I . DD). Theorem 5.12) and (5. P K ). This follows by noting that (A B ) is stabilizable i ^ ^ A BC BD (5:12) ^ ^ 0 A B is stabilizable and (C A) is detectable i ^ A BC ^ (5:13) C DC ^ 0 A ^ is detectable. the system is internally stable i A is stable. But conditions (5.1 C DC ^ D = (I . DD).7 is satis ed and the Nyquist plot of (j!) for . DD). So we only implies that all zeros of det(I .1 2 RH1 ..3. i. then (I . Proof. P (s)K A = need to show that given condition (ii). nk + np times in the counterclockwise direction.
i. for example.N M See.1 2 RH1 : More generally. Xr Yr M .1 2 RH1 =) h.. It follows from Euclid's algorithm1 that two polynomials m and n are coprime i there exist polynomials x(s) and y(s) such that xm + yn = 1 such an equation is called a Bezout identity. a leftcoprime factorization (lcf) has the form P = M . but equivalent. two transfer functions m(s) and n(s) in RH1 are said to be coprime over RH1 if there exists x y 2 RH1 such that xm + yn = 1: The more primitive. Similarly. a left coprime ~ ~ factorization P = M . 1980. h mh. de nition is that m and n are coprime if every common divisor of m and n is invertible in RH1 . two matrices M and N in RH1 are left coprime over RH1 if they have the same number of rows and if there exist matrices Xl and Yl in RH1 such that ~ ~ M N Xl = MX + NY = I: ~ l ~ l Yl Note that these de nitions are equivalent to saying that the matrix M is leftN ~ ~ invertible in RH1 and the matrix M N is rightinvertible in RH1 . e.Yl N Xl = I: (5:14) ~ Of course implicit in these de nitions is the requirement that both M and M be square and nonsingular.3 Two matrices M and N in RH1 are right coprime over RH1 if they have the same number of columns and if there exist matrices Xr and Yr in RH1 such that Xr Yr M = Xr M + Yr N = I: N ~ ~ Similarly.1 where N and M are rightcoprime over RH1 .4 Coprime Factorization over RH1 Recall that two polynomials m(s) and n(s). These two equations are often called Bezout identities. are said to be coprime if their greatest common divisor is 1 (equivalent.126 STABILITY AND PERFORMANCE OF FEEDBACK SYSTEMS 5. ~ ~ ~ ~ Similarly. Kailath. real coe cients. with. we have De nition 5. 140141]. they have no common zeros).1 . and Xr Yr Xl Yl 2 RH1 such that ~ ~ . pp.1N . A matrix P (s) 2 Rp (s) is said to have double coprime factorization if there exist a right coprime factorization P = NM .g. 1 .1N where N and M are leftcoprime over RH1 .1 nh.. Now let P be a proper realrational matrix.e. A rightcoprime factorization (rcf) of P is a factorization P = NM .
and. Yr = Yl = 0.14). 2 Remark 5. then we can take Xr = Xl = I . introduce a state feedback and change the variable v := u . Fx where F is such that A + BF is stable. (5. furthermore. ~ ~ N = N = P .D I Proof.L 3 M . right coprime factorization comes out naturally from changing the control variable by a state feedback. For example.2 Note that if P is stable.(B + LD) L 5 : F I 0 ~ ~ . and de ne 2 A + BF B . the transfer matrix from v to u is M (s) = A + BF B F I .1 = M .3 The coprime factorization of a transfer matrix can be given a feedback control interpretation. Consider the state space equations for a plant P : x = Ax + Bu _ y = Cx + Du: Next.1N are rcf and lcf.14) is satis ed.9 Suppose P (s) is a proper realrational matrix and 127 A B P= C D is a stabilizable and detectable realization. Let F and L be such that A + BF and A + LC are both stable. respectively. The theorem follows by verifying the equation (5. Then we get x = (A + BF )x + Bv _ u = Fx + v y = (C + DF )x + Dv: Evidently from these equations.4. ~ Remark 5. M = M = I.Yl = 4 F I 0 5 N Xl C + DF D I 2 3 Xr Yr = 4 A + LC . .N M C (5:15) (5:16) ~ ~ Then P = NM .5. Coprime Factorization over RH1 Theorem 5.
i..1 = M U I K = .1 2 RH1 : P I (5:19) ^ I UV . MV .e. . P = NM .128 STABILITY AND PERFORMANCE OF FEEDBACK SYSTEMS B A N (s) = C + BF D : + DF and that from v to y is Therefore u = Mv y = Nv so that y = NM .~ is invertible in RH1 .2.1 2 RH1 .N .1 N (5:17) ^ ~ ~ K = UV . ~ ~ 5.1 U: lent: 1. internal stability is equivalent to ^ I .1 0 M U . Now so that ^ I K . 2.1 = M . Consider again the standard stability analysis ^ diagram in Figure 5. The feedback system is internally stable. V M . The following conditions are equivaM U is invertible in RH . equivalently. NU is invertible in RH1 .P I or. ~ M ~ ~ 4. 1 N V ~ ~ V U 3.1 = V . We begin with any rcf's and lcf's of P and K : ~ ~ P = NM .10 Consider the system in Figure 5.2. ~ We shall now see how coprime factorizations can be used to obtain alternative characterizations of internal stability conditions. UN is invertible in RH1 . Proof.3.K . (5:18) Lemma 5.1 : N V .1 .1 NM I N V P I ^ I K .1 = M 0 0 V P I M .1 u.1 0 V . As we saw in Lemma 5.
so is the right hand side.U M U = V M . ~ ~ Corollary 5.N Furthermore. Then a particular set of state space realizations for these matrices can be given by 2 A + BF B . let F and L be such that A + BF and A + LC are stable. Then there exists a controller ^ ~ ~ K0 = U0 V0. (5.1 = V0.1N be corresponding rcf and lcf over RH1 .1 2 RH : 1 N V M 0 0 I ~ ~ ~ V M U . Coprime Factorization over RH1 Since the matrices 129 are rightcoprime (this fact is left as an exercise for the reader).19) holds i This proves the equivalence of conditions 1 and 2.U0 M U0 = I 0 : (5:20) ~ M ~ N V0 0 I . Hence. I M 0 0 V M U N V M U . Let us show condition 5 ! 1 this is obvious since ~ ~ ^ I V .1 U .1 = . UN 0 ~ ~ ~ ~ N V . The conditions 4 and 5 are implied by 2 and 3 from the following equation: ~ ~ ~ ~ V .11 Let P be a proper realrational matrix and P = NM .4.1 U0 ~ ~ with U0 V0 U0 .1 I K .1 if VN U 2 RH1 or if condition 5 is satis ed.N M = C .D I . NU : Since the left hand side of the above equation is invertible in RH1 .10 and Theorem 5.1 = M . We only need to show that either condition 4 or condition 5 implies condition 1. The equivalence of 1 and 3 is proved similarly. conditions 4 and 5 are satis ed.9.(B + LD) L 3 ~0 .N M 0 MV .1 V 0 2 RH1 N I 0 I 2 Combining Lemma 5.5. we have the following corollary.1 NM I P I = ~ M ~ . and V0 in RH1 such that ~ ~ V0 .U0 ~ V 4 F (5:22) I 0 5: ~ ~ .L 3 M U0 = 4 F (5:21) I 0 5 N V0 C + DF D I 2 A + LC .
? upe P . Using the observer ^ theory.1 = V0.20) must be invertible in RH1 .3. In fact. how to achieve the bene ts of feedback in the face of uncertainties.? ye ? en d Figure 5. Lo . Then Lemma 5.6  K u di .10 implies that each of the two lefthand side block matrices of (5. we discuss the properties of a feedback system.1 y = So d: .3: Standard Feedback Con guration Consider again the feedback system shown in Figure 5. The idea behind the choice of these matrices is as follows. (5. In particular. as Li = KP Lo = PK respectively. it is convenient to de ne the input loop transfer matrix. For convenience.1 up = Si di : And the output sensitivity matrix is de ned as the transfer matrix from d to y: So = (I + Lo ). For further discussion. re .5 Feedback Properties In this section.20) is satis ed by comparing it with the equation (5. The input sensitivity matrix is de ned as the transfer matrix from di to up : Si = (I + Li ). and output loop transfer matrix.L : (5:23) 0 Perform factorizations ^ ~ ~ K0 = U0 V0. we consider the bene t of the feedback structure and the concept of design tradeo s for con icting objectives { namely.1. nd a controller K0 achieving internal stability for example LC ^ K0 := A + BF +F + LDF .130 STABILITY AND PERFORMANCE OF FEEDBACK SYSTEMS Proof.14). 2 5.1 U0 which are analogous to the ones performed on P . Li . where Li is obtained from breaking the loop at the input (u) of the plant while Lo is obtained from breaking the loop at the output (y) of the plant. the system diagram is shown again in Figure 5.
1 = 1 (Si ) = (I + KP ) (for disturbance at plant input. S . So = Lo(I + Lo).(I + KP ).5.1P = (PS ) (for disturbance at plant input. KSod . n) . Hence. 1 (KP ) . satis es the following equations: Ti = I . di ) . n) + So Pdi + So d So (r . So Pdi KSo (r . KSod + Si di : (5. equation (5.5. equation (5. It is easy to see that the closedloop system.(I + PK ).27) These four equations show the fundamental bene ts and design objectives inherent in feedback loops. 1 (I + PK ) (I + KP ) (PK ) + 1 (KP ) + 1 .1 = (KS ) (for disturbance at plant output.25) (5.1 y r.27) shows that the e ects of disturbance di on the plant input can be made small by making the input sensitivity function Si small. T = I . Ti di KSo (r . n) . for example.24) (5. Si = Li (I + Li ). d) (Si K ) = o be made small. The notion of smallness for a transfer matrix in a certain range of frequencies can be made explicit using frequency dependent singular values.) The matrix I + Li is called input return di erence matrix and I + Lo is called output return di erence matrix. Similarly. d ) (So P ) = i i be made small and good disturbance rejection at the plant input (up ) would require that . d) .(I + PK ). particularly in the low frequency range where d and di are usually signi cant. (So ) < 1 over a frequency range would mean that the e ects of disturbance d at the plant output are e ectively desensitized over that frequency range.26) (5. For example.1 To = I .y u up = = = = To (r . if it is internally stable. Note that (PK ) . Feedback Properties The input and output complementary sensitivity matrices are de ned as 131 respectively. d) + Ton .24) shows that the e ects of disturbance d on the plant output can be made \small" by making the output sensitivity function So small. good disturbance rejection at the plant output (y) would require that . (The word complementary is used to signify the fact that T is the complement of S .1 = 1 (So ) = (I + PK ) (for disturbance at plant output.K (I + PK ).
and assume that the system is nominally stable. concerns commands and disturbance error reduction versus stability under the model uncertainty. good multivariable feedback loop design boils down to achieving high loop (and possibly controller) gains in the necessary frequency range. good disturbance rejection at the output does not necessarily mean good disturbance rejection at the plant input. they must satisfy certain performance tradeo and design limitations. Similarly.) Hence. (It should be noted that in general So 6= Si unless K and P are square and diagonal which is true if P is a scalar system.1P .e. Now the perturbed closedloop system is stable if det (I + (I + )PK ) = det(I + PK ) det(I + To) has no righthalf plane zero. then (PK ) (PK ) 1 or (KP ) 1 or (KP ) 1 () 1 () .1 ) = (P ) 1 (PK ) + 1 1 (KP ) + 1 These equations imply that Now suppose P and K are invertible.1 Hence good performance at plant output (y) requires in general large output loop gain (Lo ) = (PK ) 1 in the frequency range where d is signi cant for desensitizing d and large enough controller gain (K ) 1 in the frequency range where di is signi cant for desensitizing di . This would in general amount to requiring that k Tok be small or that (To ) be small at those frequencies where is signi cant.1 ) = (1 ) K 1 : (P . which can not changed by controller design. the closedloop system with = 0 is stable. feedback design is by no means trivial.132 then STABILITY AND PERFORMANCE OF FEEDBACK SYSTEMS 1 (So ) (Si ) (So ) (Si ) (PK ) . A major performance tradeo . 1 if (KP ) > 1: 1 () 1 () (So P ) = (KSo ) = (PK ) (KP ) 1 1: (K . typically at .. Assume that the plant model is perturbed to (I + )P with stable. Hence. Rather.K (I + PK ). i. 1 if (PK ) > 1 1 (KP ) . for example. This is true because loop gains cannot be made arbitrarily high over arbitrarily large frequency ranges. Despite the simplicity of this statement. for desensitizing d.(I + PK ). small (So ) does not necessarily imply small (Si ) in other words. good performance at plant input (up ) requires in general large input loop gain (Li ) = (KP ) 1 in the frequency range where di is signi cant for desensitizing di and large enough plant gain (P ) 1 in the frequency range where d is signi cant.
n . The resulting equation shows that disturbances and sensor noise are actually ampli ed at u whenever the frequency range signi cantly exceeds the bandwidth of P . typically some low frequency range (0 !l ): (PK ) 1 (KP ) 1 (K ) 1 and good robustness and good sensor noise rejection require in some frequency range. i. Still another tradeo is with the sensor noise error reduction. typically some high frequency range (!h 1) (PK ) 1 (KP ) 1 (K ) M where M is not too large. d) .1 (j!)] = P (j!)] 1: Similarly. n): Note that n is typically signi cant in the high frequency range. d) . To summarize the above discussion. Worst still. The speci c frequencies !l and !h depend on the speci c applications and the knowledge one has on the disturbance characteristics. This is because for small loop gain (Lo (j!)) 1 or (Li (j!)) 1 u = KSo (r . n . can make the control activity (u) quite unacceptable. Ti di P . (Lo ). the modeling uncertainties.1 (r . (Lo (j!)) 1 or (Li (j!)) 1 while (P (j!)) 1. large loop gains outside of the bandwidth of P . Ti di = Si K (r .. which in turn implies that the loop gain. However. Ti di K (r . and the sensor noise levels. This follows from u = KSo(r . d) . we have assumed P to be square and invertible for convenience. it is desirable to keep (K ) not too large when the loop gain is small. n . i.24). Feedback Properties 133 high frequency range. we have 1 P . y = To (r . Large (Lo(j!)) values over a large frequency range make errors due to d small. they also make errors due to n large because this noise is \passed through" over the same frequency range. the controller gain.e.5. should be small at those frequencies.5. should also be kept not too large in the frequency range where the loop gain is small in order to not saturate the actuators. we note that good performance requires in some frequency range. (K ). d) . since for ! such that (P (j!)) 1.4. . n . These design requirements are shown graphically in Figure 5. di : Here. d): Therefore. n .e.. The con ict between the disturbance rejection and the sensor noise reduction is evident in equation (5. n) + So Pdi + So d (r . which may cause the saturation of actuators.
.. . .e. . HH S @@ S Figure 5. The loop shaping for MIMO system can be done similarly if the singular values of the loop transfer functions are used for the loop gains. In the SISO case...6 The Concept of Loop Shaping The analysis in the last section motivates a conceptually simple controller design technique: loop shaping. ZS @Z .4: Desired Loop Gain 5. . . .. H...ZZ @ @ .134 STABILITY AND PERFORMANCE OF FEEDBACK SYSTEMS HH H 6 @ HH @ H @ XXXX @@ @ @ (L) .. the loop shaping design technique is particularly e ective and simple since (L) = (L) = jLj...Z S @ . clear the boundaries speci ed by the performance requirements at low frequencies and by the robustness requirements at high frequencies as shown in Figure 5.4... Z . which can usually be guaranteed by making L wellbehaved in the crossover region. . . .. .... L must also be chosen so that 1+ L has all zeros in the open left half plane...... (L) and (L). l HH Z . ... .. i.. . . ....... (2) The controller is given by K = L=P ..H SS @@ H.... .. S ..4. ... ! HHH ZZZ !h .. The design procedure can be completed in two steps: SISO Loop Shaping (1) Find a rational strictly proper transfer function L which contains all the right half plane poles and zeros of P such that jLj clears the boundaries speci ed by the performance requirements at low frequencies and by the robustness requirements at high frequencies as shown in Figure 5.. . log ! ... .. L should not be decreasing too fast in the frequency range of jL(j!)j 1......H . Z @@ @ . Loop shaping controller design involves essentially nding a controller K that shapes the loop transfer function L so that the loop gains. ... (L)SS @@HH..
Limitations of MIMO Loop Shaping Although the above loop shaping design can be e ective in some of applications. The loop shaping design technique can be quite useful especially for SISO control system design.1 L (or LP . Assume that can be written in the following form = ~ Wt ( ~ ) < 1: Then. To illustrate this di culty. However. More speci cally. there are severe intrinsic limitations. for robust stability. Some of these limitations are listed below: The loop shaping technique described above can only e ectively deal with problems with uniformly speci ed performance and robustness speci cations.) (2) The controller is given by K = P . there are severe limitations when it is used for MIMO system design. we would require ( To ) = ( ~ Wt To) < 1 or (Wt To) 1. The Concept of Loop Shaping MIMO Loop Shaping 135 (1) Find a rational strictly proper transfer function L which contains all the right half plane poles and zeros of P so that the product of P and P . the method can not e ectively deal with problems with di erent speci cations in di erent channels and/or problems with di erent uncertainty characteristics in di erent channels without introducing signi cant conservatism.1 L if L is the output loop transfer function (or K = LP .1 ) has no unstable poles and/or zeros cancelations. and (L) clears the boundary speci ed by the performance requirements at low frequencies and (L) clears the boundary speci ed by the robustness requirements at high frequencies as shown in Figure 5. consider an uncertain dynamical system P = (I + )P where P is the nominal plant and is the multiplicative modeling error. (This is not easy for MIMO systems. If a uniform bound is required on the loop gain to apply the loop shaping technique. L must also be chosen so that det(I + L) has all zeros in the open left half plane. we would need to overbound (Wt To ): ) (Wt To ) (Wt ) (To ) (Wt ) 1 .(L(oL ) if (Lo ) < 1 o and the robust stability requirement is implied by 1 (Lo ) (W1) + 1 (Wt ) if (Lo ) < 1: t .6.1 if L is the input loop transfer function).5.4.
. . . . . @ @ @ @@ log ! .@ @ .Z . ...@. . are not uniformly speci ed in all channels but by a weighting matrix Ws such that (Ws So ) 1.. . . . . .. 1 (Wt ) Figure 5.@ @ @ @ @ @ . .. . . this does not imply that there is no controller that will satisfy both nominal performance and robust stability except for SISO systems.. 7 @ PP @ @ @ . P@.... . P@ P @ @ @ @ @ @@ . . . This contradiction happens because the bounds 6 (Ws ) ....Z Z .@ . .5. . ..136 STABILITY AND PERFORMANCE OF FEEDBACK SYSTEMS Similarly.. as shown in the gure 5. . P @ P@ P .. ... . . if the performance requirements. .. .5: Con ict Requirements do not utilize the structure of weights. . .. .P.. . . This possibility can be further illustrated by the following example: Assume that a twoinput and twooutput system transfer matrix is given by 1 P (s) = s + 1 1 1 0 and suppose the weighting matrices are given by 2 Ws = 4 s+1 1 (s+1)(s+2) 0 s+1 1 3 2 5 Wt = 4 s+2 s+10 0 (s+1) s+10 s+2 s+10 3 5: .. ZZ . .. . .@. However.ZZ . Z . . . Ws and Wt and the bounds are only su cient conditions for robust stability and nominal performance. . . . . then it is also necessary to overbound (Ws So ) in order to apply the loop shaping techniques: ( (Ws So ) (Ws ) (So ) (L Ws ) 1 if (Lo) > 1 o) .. . say output disturbance rejection. . ... . . .@. and the performance requirement is implied by (Lo ) (Ws ) + 1 (Ws ) if (Lo) > 1: It is possible that the bounds for the loop shape may contradict each other at some frequency range.
This becomes much harder if P is nonminimum phase and/or unstable. Weighted H2 and H1 Performance 137 It is easy to show that for large . the performance objectives of a feedback system can usually be speci ed in terms of requirements on the sensitivity functions and/or complementary sensitivity functions or in terms of some other closedloop transfer functions. As shown in section 5.5.5. This motivates us to consider the closedloop performance directly in terms of the closedloop transfer functions instead of open loop transfer functions. it is also easy to show that the system with controller K = I2 gives and. However. we consider how to formulate some performance objectives into mathematically tractable problems.7. but it is essentially limited to nominally minimum phase and output multiplicative uncertain systems. it may still be di cult to nd a matrix function Lo so that K = P . The following section considers some simple closedloop performance problem formulations.1 Lo is stabilizing. However. For example. it is much more convenient to re ect the system performance objectives by choosing appropriate weighting functions. For instance. Even if all of the above problems can be avoided. the above performance objective can be written as jws (j!)s(j!)j 1 8! with jws (j!)j = .7 Weighted H2 and H1 Performance In this section. therefore. Hence some new methodologies have to be introduced to solve complicated problems. Ws S = s+2 0 1 0 s+2 1 Wt T = s+10 0 1 0 s+10 1 5. For these reasons. it will not be introduced here.1 . The socalled LQG/LTR (Linear Quadratic Gaussian/Loop Transfer Recovery) procedure developed rst by Doyle and Stein 1981] and extended later by various authors can solve some of the problems. the weighting functions are as shown in Figure 5. the performance criteria for a scalar system may be speci ed as requiring js(j!)j < 1 8! !0 js(j!)j > 1 8! > !0 1 where s(j!) = 1+p(j!)k(j!) . the robust performance criterion is satis ed.5.1 8! !0 8! > !0 : . and thus the above loop shaping technique cannot be applied.
We e Wd d ? ~ d ? en Wn n ~ Figure 5. H2 and H1 theory.138 STABILITY AND PERFORMANCE OF FEEDBACK SYSTEMS u ~ 6 Wu . hence some frequency dependent weights must be chosen. It is. We Wd Wu . Therefore. For example. The weighting matrix Wn is used to model the frequency contents of the sensor noise while We may be used to re ect the requirements on the shape of certain closedloop transfer functions.e. the weights are chosen purely as a design parameter without any physical bases. and possibly Wn Wi . Wu may be used to re ect some restrictions on the control or actuator signals.Wr re . as in the scalar case. Also. In many occasions.6. in fact. the shape of the output sensitivity function.6  6 K u e ?  Wi di ? P ~ di e . Secondly. So a very important step in controller design process is to choose the appropriate weights. First of all.. i. Wd and Wi may be chosen to re ect the frequency contents of the disturbances d and di or they may be used to model the disturbance power spectrum depending on the nature of signals involved in the practical systems.? y.6: Standard Feedback Con guration with Weights In order to use ws in control design. Similarly. and the others may be measured in terms of voltage. so . weighting functions are essential to make these components comparable. some components of the output error signal may be measured in terms of length. In general. The appropriate choice of weights for a particular practical problem is not trivial.3 into Figure 5. a rational transfer function ws is usually used to approximate the above frequency response. we might be primarily interested in rejecting errors in a certain frequency range (for example low frequencies). some components of a vector signal are usually more important than others.6 are chosen to re ect the design objectives and knowledge on the disturbances and sensor noise. for example. The weighting functions in Figure 5. The advantage of using weighted performance speci cations is obvious in multivariable system design. we shall modify the standard feedback diagram in Figure 5. and the dashed precompensator Wr is an optional element used to achieve deliberate command shaping or to represent a nonunity feedback system in equivalent unity feedback form. each component of the signal may not be measured in the same metric for example. essential that some appropriate weighting matrices be used in order to utilize the optimal control theory discussed in this book. and Wr .
. Hence. ~ d(t) = (t) and E( ) = I where E denotes the expectation. For the simplicity of presentation. i. H2 Performance ~ Assume. We may choose to minimize the expected energy of ~ the error e due to the disturbance d: E kek2 = E 2 n o Z1 0 kek2 dt = kWe So Wd k2 : 2 ~ Alternatively. that the disturbance d can be approximately modeled as an impulse with random input direction. then ~~ See = (We So Wd )Sd~d~(We So Wd ) and we may chose to minimize the power of e: Z1 kek2 = 21 Trace See (j!) d! = kWe So Wd k2 : P 2 . and some are much harder than others. we shall assume di = 0 and n = 0.7. Thus.1 In general. it is necessary to include the control signal u in the penalty function. control design may be regarded as a process of choosing a controller K such that certain weighted signals are made small in some sense. for example. Below. a controller minimizing only the above criterion can lead to a very large control signal u that could cause saturation of the actuators as well as many other undesirable problems. Weighted H2 and H1 Performance 139 these weights may be treated as tuning parameters which are chosen by the designer to achieve the best compromise between the con icting objectives. The selection of the weighting matrices should be guided by the expected system inputs and the relative importance of the outputs. as we have discussed in the last chapter. our design criterion would usually be something like this E kek2 + 2 kuk2 = ~2 2 n o We So Wd Wu KSoWd 2 2 . Di erent de nitions lead to di erent control synthesis methods. if we suppose that the disturbance d can be approximately modeled as white noise with Sdd = I .e. for a realistic controller design. A control engineer should make a judgment of the mathematical complexity versus engineering requirements. we introduce two classes of performance formulations: H2 and H1 criteria.5. Hence. There are many di erent ways to de ne the smallness of a signal or transfer matrix.
the requirement for tolerance to uncertainties. The problem is that these performance tradeo s are often overshadowed by a second limitation on high loop gains { namely. put strict limitations on the frequency range over which the loop gains may be large.140 STABILITY AND PERFORMANCE OF FEEDBACK SYSTEMS with some appropriate choice of weighting matrix Wu and scalar . the H2 norm su ers a major de ciency. This de ciency is due to the fact that the tradeo between disturbance error reduction and sensor noise error reduction is not the only constraint on feedback design. The parameter clearly de nes the tradeo we discussed earlier between good disturbance rejection at the output and control e ort (or disturbance and sensor noise rejection at the actuators). H1 Performance Although the H2 norm (or L2 norm) may be a meaningful performance measure and although LQG theory can give e cient design compromises under certain disturbance and plant assumptions. Though a controller may be designed using FDLTI models. Plant uncertainty is particularly critical in feedback systems.) The development of this paradigm stimulated extensive research e orts and is responsible for important technological innovation. in particular the ways in which they deviate from nitedimensional linear models. particularly in the area of estimation. The theoretical contributions include a deeper understanding of linear systems and improved computational methods for complex systems through statespace techniques. By allowing only additive noise for uncertainty. Note that can be set to = 1 by an appropriate choice of Wu . one may introduce a parameter and a mixed criterion kd~k2 1 kd~k2 1 kd~k2 sup n 1 kek2 + 2 kuk2 = ~2 2 o We So Wd Wu KSo Wd 2 1 : . This type of problem was the dominant paradigm in the 1960's and 1970's and is usually referred to as Linear Quadratic Gaussian Control or simply as LQG. A solution to this problem would be to put explicit constraints on the loop gain in the penalty function. For instance. the stochastic theory ignored this important practical issue. (They will also be referred to as H2 mixed sensitivity problems for the consistency with the H1 problems discussed next. the design must be implemented and operated with a real physical plant. The properties of physical systems. The major limitation of this theory is the lack of formal treatment of uncertainty in the plant itself. This problem can be viewed as minimizing the energy consumed by the system in order to reject the disturbance d. one may chose to minimize sup kek2 = kWe So Wd k1 subject to some restrictions on the control energy or control bandwidth: sup kuk2 = kWu KSo Wd k1 : ~ Or more frequently.
. if the system robust stability margin is the major concern.8. These design problems are usually called H1 mixed sensitivity problems. Chapter 18 presents another loop shaping method using H1 control theory which has the potential to overcome the limitations of the LQG/LTR method.5. Notes and References 141 This problem can also be regarded as minimizing the maximum power of the error subject to all bounded power disturbances: let e e := u ^ ~ then and W W See = W e So Wd Sd~d~ W e So Wd ^^ KSoWd u u KSo Wd kdkP sup kek2 = sup ^P ~ ~ 1 kdkP 1 1 Z 1 Trace S (j!) d! = ee ~~ 2 . For a scalar system. The discussion of internal stability and coprime factorization can also be found in Francis 1987] and Vidyasagar 1985].1 We SoWd Wu KSo Wd 2 1 : Alternatively. The idea was extended to MIMO systems by Doyle and Stein 1981] using LQG design technique. the weighted complementary sensitivity has to be limited. an H1 norm minimization problem can also be viewed as minimizing the maximum magnitude of the system's steadystate response with respect to the worst case sinusoidal inputs. Thus the whole cost function may be We So Wd W1 ToW2 where W1 and W2 are the frequency dependent uncertainty scaling matrices. The limitations of the loop shaping design are discussed in detail in Stein and Doyle 1991]. The loop shaping design is well known for SISO systems in the classical control theory. 1 5.8 Notes and References The presentation of this chapter is based primarily on Doyle 1984].
142 STABILITY AND PERFORMANCE OF FEEDBACK SYSTEMS .
1 Introduction One important problem that arises frequently is concerned with the level of performance that can be achieved in feedback design. The sensitivity integral relations are used to study the design limitations imposed by bandwidth constraints and the openloop unstable poles. and a tradeo must be performed among di erent design objectives. and the design limitations become more severe in the presence of righthalf plane zeros and poles in the openloop transfer function. These results display that the design limitations in multivariable systems are dependent on the directionality properties of the sensitivity function as well as those of the poles and zeros. It is also known that the fundamental requirements such as stability and robustness impose inherent limitations upon the feedback properties irrespective of design methods. 6 6. An important tool that can be used to quantify feedback design constraints is furnished by the Bode's sensitivity integral relation and the Poisson integral formula. These integral formulae express design constraints directly in terms of the system's sensitivity 143 . It has been shown in the previous chapters that the feedback design goals are inherently con icting. These integral relations are also used to derive lower bounds on the singular values of the sensitivity function which display the design tradeo s. while the Poisson integral formula is used to study the design constraints imposed by the nonminimum phase zeros. in addition to the dependence upon pole and zero locations which is known in singleinput singleoutput systems.Performance Limitations This chapter introduces some multivariable versions of the Bode's sensitivity integral relations and Poisson integral formula.
be the nonminimum phase zeros of L(s) and let i i i = 1 be the input directions generated from the following iterative procedure . Let zi 2 C + . i = 1 k. Let S (s) and T (s) be the sensitivity function and the complementary sensitivity function. The results presented here are some multivariable extensions of the above mentioned integral relations. Then. zI B C D =0 holds. or w L. or w L(z ) = 0 for some vector w. We shall now consider a linear timeinvariant feedback system with an n n loop transfer matrix L.1 T (s) = L(s)(I + L(s)). The vector w is called the output zero direction associated with z .1(p) = 0 for some vector .1 (s) the input and output pole directions of L(s).1(p) = 0 for some vector w. if there exist vectors and such that the relation A . It is wellknown that a nonminimum phase transfer function admits a factorization that consists of a minimum phase part and an allpass factor. by Lemma 3. In the same reference. By a slight abuse of terminology.144 PERFORMANCE LIMITATIONS and complementary sensitivity functions. the limitations imposed by the open loop nonminimum phase zeros upon the feedback properties were also quanti ed using the Poisson integral.27 and 3. respectively S (s) = (I + L(s)). respectively. This integral relation therefore suggests that in the presence of bandwidth constraints. Similarly. the integral of the logarithmic magnitude of the sensitivity function over all frequencies must equal to zero. Note also that p 2 C is a pole of L(s) if and only if it is a zero of L. which has full normal rank and a minimal state space realization (A B C D).28. recall that a point z 2 C is a transmission zero of L(s). = 1. we shall call the input and output zero directions of L. w w = 1. z is a zero of L(s) if and only if L(z ) = 0 for some vector . A result by Freudenberg and Looze 1985] further extends Bode's theorem to openloop unstable systems. which shows that the presence of openloop unstable poles makes the sensitivity tradeo a more di cult task.1(s). = 1. where = 1. Analogously. A wellknown theorem due to Bode states that for singleinput singleoutput openloop stable systems with more than one polezero excess. In the sequel we shall preclude the possibility that z is both a zero and pole of L(s).1 : (6:1) Before presenting the multivariable integral relations. zI B = 0 C D where x and w are some vectors with w satisfying the condition w w = 1. the desirable property of sensitivity reduction in one frequency range must be traded o against the undesirable property of sensitivity increase at other frequencies. and is called the input zero direction associated with z . p is a pole of L(s) if and only if L. w w = 1. a transmission zero z of L(s) satis es the relation x w A .
Integral Relations Let (A B C D) be a minimal realization of L(s) and B (0) := B Repeat for i = 1 to k 145 A .6. 6.2 R!1 lim s2C+ jsj R sup R (L(s)) = 0 .1) C D i i =0 B (i) := B (i. the input factorization of L(s) is given by L(s) = Lm (s)Bk (s) B1 (s) where Lm (s) denotes the minimum phase factor of L(s). suppose z 2 C is a zero of L(s). 2(Rez ) I . The following assumptions are needed: Assumption 6. 2(Rezi ) i i : Then. s + zi i i : i In fact. s+z : D z Note that a nonminimum phase transfer function admits an output factorization analogous to the input factorization. Consider a unit feedback system with a loop transfer function L. and the subsequent results can be applied to both types of factorizations.1 The closedloop system is stable. Assumption 6. Lm(s) can be written as A (k) Lm(s) := C BD : For example. and Bi (s) corresponds to the allpass factor associated with zi : 2Rez (6:2) Bi (s) = I . zi I B (i.2 Integral Relations In this section we provide extensions of the Bode's sensitivity integral relations and Poisson integral relations to multivariable systems.2. Then it is easy to verify using state space calculation that L(s) can be factorized as 2Re A L(s) = C B .1) .
2 L. then ln (S (j!))d! Repi : (6:6) . Then there exists .1).1(p) = 0. then Z1 0 ln (S (j!))d! max Repi : k X i=1 (6:5) (ii) If i = for all i = 1 Z1 0 k.a vector such that . Lemma 6.2 hold.146 PERFORMANCE LIMITATIONS Each of these assumptions has important implications.1 (s) Then the sensitivity function S (s) can be factorized as S (p) = 0.1 implies that the sensitivity function S (s) is analytic in C + . The proof for necessity follows by reversing the above procedure. However.16. i = 1 k. =1 S (s) = Sm (s)B1 (s)B2 (s) Bk (s) (6:3) where Sm (s) has no zeros in C + .2 states that the openloop transfer function has a rollo rate of more than one polezero excess. S (s) = (I + L(s)). Suppose that the openloop transfer function L(s) has poles pi in the open righthalf plane with input pole directions i . Hence.1 = I + L.1 and L. Assumption 6. Then Z1 0 ln (S (j!))d! 0k 1 X max @ (Repj ) j j A : j =1 (6:4) It is also instructive to examine the following two extreme cases. One instance for this assumption to hold is that each element of L has a rollo rate of more than one polezero excess. Let p be a pole of L(s). i 6= j . 2Repi i i : s + pi Theorem 6. and Bi (s) is given by Bi (s) = I . Note that most of practical systems require a rollo rate of more than one polezero excess in order to maintain a su cient stability margin. (i) If i j = 0 for all i j = 1 k. Proof. This establishes the su ciency part. Then p 2 C is a zero of S (s) with zero direction if and only if it is a pole of L(s) with pole direction .1 Let L(s) and S (s) be de ned by (6.3) and suppose that Assumptions 6. Assumption 6.2 Let S (s) be factorized in (6.1 (s). which are obtained through a similar iterative procedure as in the last section.
for which the unstable poles a ect only the integral corresponding to the largest singular value.16.2 hold. but also to the directions of poles and their relative interaction. Integral Relations Note also that the following equality holds 147 Z1 0 ln jS (j!)jd! = k X i=1 Repi if S (s) is a scalar function. Note that max ((Rep1 ) 1 1 + (Rep2 ) 2 2 ) = = = = max max amx max 1 2 A straightforward calculation then gives 1 max ((Rep1 ) 1 1 + (Rep2 ) 2 2 ) = 2 Re(p1 + p2 ) Rep1 0 1 0 Rep2 2 Rep1 0 1 0 Rep2 2 1 Rep1 (Rep1 ) 1 (Rep2 ) 2 1 Rep2 Rep1 0 0 Rep2 1 1 2 1 2 2 1 2 : . Then Z1 0 ln (S (j!))d! 2 2 2 Re(p1 + p2 ) + (Re(p1 . Clearly. The case (i) of this result corresponds to the situation where the pole directions are mutually orthogonal. Furthermore. as if the channel corresponding to the largest singular value contains all unstable poles. this result suggests that unlike in singleinput singleoutput systems. It shows that there will exist a frequency range over which the largest singular value of the sensitivity function exceeds one if it is to be kept below one at other frequencies. the limitations imposed by openloop unstable poles in multivariable systems are related not only to the locations.3 Let k = 2 and let the Assumptions 6. these phenomena are unique to multivariable systems.6. Corollary 6. The following result further strengthens these observations and shows that the interaction between openloop unstable poles plays an important role toward sensitivity properties. In the presence of bandwidth constraint. for which the integral corresponding to each singular value is related solely to one unstable pole with a corresponding distance to the imaginary axis.2. The case (ii) corresponds to the situation where all the pole directions are parallel. this imposes a sensitivity tradeo in di erent frequency ranges. p2 )) + 4Rep1 Rep2 cos \( p 1 2) (6:7) : Proof. as if each channel of the system is decoupled from the others in e ects of sensitivity properties. Theorem 6.2 has an important implication toward the limitations imposed by the openloop unstable poles on sensitivity properties.
pi k Y z + pi x ln jS (j!)j x2 + (!0. w and k Y i=1 Bi.8) is satis ed if L(s) is a proper rational transfer matrix. =2 =2] ln (S (Rej )) R = 0: (6:8) Then. y )2 d! 0 .1 0 ln (Sm (z )) . y )2 d! 0 . Theorem 6.3) and assume that lim max R!1 2 . 2 The utility of this corollary is clear. Furthermore. Next. B1 1 ( z ) : (6:9) (6:10) x ln (S (j!)) x2 + (!0. y )2 d! = ln : 0 . This result fully characterizes the limitation imposed by a pair of openloop unstable poles on the sensitivity reduction properties. ln w Bk 1 (z ) ln (S (z )): . Z1 Z1 x ln (S (j!)) x2 + (!0. however. for any nonminimum phase zero z = x0 + jy0 2 C + of L(s) with output direction w. This limitation depends not only on the relative distances of the poles to the imaginary axis. especially when these poles and zeros are close to each other and the angles between their directions are small. is the inequality (6.10). w w = 1.1 0 i=1 z . This result implies that the sensitivity reduction ability of the system may be severely limited by the openloop unstable poles and nonminimum phase zeros. This result again suggests that the multivariable sensitivity properties are closely related to the pole and zero directions.4 Let S (s) 2 H1 be factorized as in (6. The results below may be considered to be a matrix extension of the Poisson integral relation. p2 ))2 + 4Rep1 Rep2 cos2 \( 2 p 1 2 ): The proof is now completed.148 PERFORMANCE LIMITATIONS + 1 (Re(p1 . we investigate the design constraints imposed by openloop nonminimum phase zeros upon sensitivity properties. for singleinput singleoutput systems (n = 1). but also on the principal angle between the two pole directions.1 (z ) = k Y z + pi i=1 z .1 0 Note that the condition (6. pi The more interesting result. Z1 .
suppose that the condition (6. its .5 is similar to that in singleinput singleoutput systems.4 holds. for each open righthalf plane zero z 2 C + of L(s) with output direction w.3.5 Let the assumption in Theorem 6. Let z = x0 + jy0 2 C + be an open righthalf plane zero of L(s) with output direction w. hence exhibiting a tradeo between the reduction of the sensitivity over one frequency range against its increase over another frequency range. Suppose that the feedback system is designed such that the level of sensitivity reduction is given by (S (j!)) ML < 1 8! 2 .10) and inequality (6.!L x0 The following lower bound on the maximum sensitivity displays a limitation due to the open righthalf plane zeros upon the sensitivity reduction properties. (z) (z ) . (z )) ln kS (j!)k1 + (z ) ln(ML): 0 . Corollary 6. y )2 d! ( . Then. In addition. Similar to their counterparts for singleinput singleoutput systems. this result shows that for a nonminimum phase system.13) follows by applying inequality (6.!L !L ] (6:11) where ML > 0 is a given constant.3 Design Limitations and Sensitivity Bounds The integral relations derived in the preceding section are now used to analyze the design tradeo s and the limitations imposed by the bandwidth constraint and righthalf plane poles and zeros upon sensitivity reduction properties. Roughly stated. B1 1 (z ) . De ne also (z ) := Z !L x0 2 + (! .12) follows by applying inequality (6. these integral relations show that there will necessarily exist frequencies at which the sensitivity function exceeds one if it is to be kept below one over other frequencies. (z) (6:12) (6:13) kS (s)k1 ML 1 ( (S (z ))) : Proof. w Bk 1 (z ) . kS (s)k1 and ML 1 .9) 2 The interpretation of Corollary 6. Note that Z1 x ln (S (j!)) x2 + (!0.1 0 Then the inequality (6. Design Limitations and Sensitivity Bounds 149 6.6. y0 )2 d!: . (z) . (z ) (z ) .11) is satis ed.
and that it roll o at a rate of more than one polezero excess above that frequency. !1+k 1 MH d! = X Z 1 1 MH i d! .6 Suppose the Assumptions 6. bandwidth constraints in feedback design typically require that the openloop transfer function be small above a speci ed frequency.!L (1 .2 hold.14) are satis ed for some !H and !L such that !H > !L . due to the fact that (S (z )) 1. these integral relations by themselves do not mandate a meaningful tradeo between the sensitivity reduction and the sensitivity increase. suppose that the !2 !L !H ] conditions (6. (6:15) max (S (j!)) e where ML = max Pk i=1 (Repi ) i i !H . With the bandwidth (L(j!)) constraint given as such.11) can be achieved only at the expense of increasing the sensitivity at certain frequencies. k > 0 are some given constants. Of particular importance here is that the sensitivity function will in general exhibit a larger peak in multivariable systems than in singleinput singleoutput systems. ) k(!HH !L) . ln 1 . In addition. (L(j!)) MH 1 .16. The design tradeo s and limitations on the sensitivity reduction which arise from bandwidth constraints as well as openloop unstable poles can be studied using the extended Bode integral relations. Note rst that for ! !H . One way of quantifying such bandwidth constraints is by requiring the openloop transfer function to satisfy MH < 1 8! 2 !H 1) (6:14) !1+k where !H > !L . Corollary 6. These constraints are commonly needed to ensure stability robustness despite the presence of modeling uncertainty in the plant model.11) and (6. Then !L ! 1 !H . (S (j!)) = (I + 1 (j!)) L and 1 1 . since the sensitivity function can be allowed to exceed one by an arbitrarily small amount over an arbitrarily large frequency range so as not to violate the Bode integral relations. the following result again shows that the sensitivity reduction speci ed by (6. !1+k 1+k !H i=1 !H i ! Z1 . However.150 PERFORMANCE LIMITATIONS sensitivity must increase beyond one at certain frequencies if the sensitivity reduction is to be achieved at other frequencies. However. !L : 1 Proof. particularly at high frequencies. and MH > 0.
Note L that d ln jd (j!)j is the slope of the Bode plot which is almost always negative. ) !L ln ML + (!H . Bode's Gain and Phase Relation !H MH = 1+ i i(1 + k) . 1 !H k i=1 !i ! 1 !H X 1 MH = . Francis. !kH ln(1 . and further the return di erence is given by L j1 + L(j!c )j = j1 + L. MH 1+ 1+ k i=1 i !H k k !H k . !L ln ML + (!H . !1+k d! L H !H !H ln(1 . 6. It follows that \L(j!0 ) will be large if the gain L attenuates slowly and small if it attenuates rapidly.4 Bode's Gain and Phase Relation In the classical feedback theory. ): ln (S (j!))d! = 1 X1 151 !i Then Z1 0 Z !L 0 ln (S (j!))d! + Z !H !L ln (S (j!))d! + Z1 !H ln (S (j!))d! MH ln 1 . !H ln 1 . zi where zi 's are assumed to be the righthalf plane zeros of L(s) and := ln(!=!0 ).4). !L ) !2max ] ln (S (j!)) . k !L !H and the result follows from applying (6. !L ) !2max ] ln (S (j!)) . see Freudenberg and Looze 1988] and Doyle. The behavior of \L(j!) is particularly important near the crossover frequency !c where jL(j!c )j = 1 since + \L(j!c) is the phase margin of the feedback system. ! ! Z1 2 The above lower bound shows that the sensitivity can be very signi cant in the transition band.1(j!c )j = 2 sin + \2 (j!c) .6. and Tannenbaum 1992]: k 1 Z 1 d ln jLj ln coth j j d + \ Y j!0 + zi \L(j!0 ) = 2 .4.1 d i=1 j!0 . The following is an extended version of the Bode's gain and phase relationship for an openloop stable scalar system with possible righthalf plane zeros. the Bode's gainphase integral relation (see Bode 1945]) has been used as an important tool to express design constraints in scalar systems.
Doyle.5 Notes and References The results presented in this chapter are based on Chen 1992a. and Freudenberg and Looze 1988]. Since \ j!0 + z i 0 for each i. and Tannenbaum 1992].152 PERFORMANCE LIMITATIONS which must not be too small for good stability robustness. 6. 1995]. The same conclusion applies to multivariable system where the singular value plots should be wellbehaviored between the transition band. The study of analytic functions. Francis. In the classical feedback theory. it has been common to express design goals in terms of the \shape" of the openloop transfer function. zi a nonminimum phase zero contributes an additional phase lag and imposes limitations upon the rollo rate of the openloop gain. A typical design requires that the openloop transfer function have a high gain at low frequencies and a low gain at high frequencies while the transition should be wellbehaviored. If + \L(j!c ) is forced to be very small by rapid gain attenuation. 1992b. j!0 . See also Freudenberg and Looze 1988] for some multivariable generalizations. 1 A function : C 7. Some related results can be found in Boyd and Desoer 1985] and Freudenberg and Looze 1988]. and Freudenberg and Looze 1988]. Horowitz 1963]. and various integral relations in the scalar case can be found in Garnett 1981] and Ho man 1962].! R is said to be a harmonic function (subharmonic function) if 52 ( ) = 0 (52 ( ) 0) where the symbol 52 ( ) with = + denotes the Laplacian of ( ) and is de ned by 2 2 52 ( ) := (2 ) + (2 ) f f s f s f s s x jy f s f s @ f s @x @ f s @y : . the feedback system will amplify disturbances and exhibit little uncertainty tolerance at and near !c . A thorough discussion of the limitations these relations impose upon feedback control design is given by Bode 1945]. The related results for scalar systems can be found in Bode 1945]. harmonic functions1 . Horowitz 1963]. The con ict between attenuation rate and loop quality near crossover is thus clearly evident.
On the other hand. nd a lower order model. we shall study only two of them: balanced truncation method and Hankel norm approximation method.Model Reduction by Balanced Truncation In this chapter we consider the problem of reducing the order of a linear multivariable dynamical system. such that G and Gr are close in some sense. one may also desire that the approximation be in relative form Gr = G(I + rel ) so that rel is small in some norm. an rth order model Gr . the additive model reduction problem can be formulated as deg(Gr ) r inf kG . Gr k1 153 . say. This is called a relative model reduction problem. There are many ways to reduce the order of a dynamical system. However. This model reduction is usually called an additive model reduction problem. We shall be only interested in L1 norm approximation in this book. For example. A model order reduction problem can in general be stated as follows: Given a full order model G(s). Of course. Once the norm is chosen. This chapter focuses on the balanced truncation method while the next chapter studies the Hankel norm approximation method. one may desire that the reduced model be such that 7 G = Gr + a and a is small in some norm. there are many ways to de ne the closeness of an approximation.
e.1 (G . i. the requirement on the approximation accuracy at one frequency range can be drastically di erent from the requirement at another frequency range.4) 1 A21 1 + 2 A12 + B2 B1 = 0 (7. Denote the balanced Gramians by .7) A22 + A22 2 + C2 C2 = 0: (7.154 MODEL REDUCTION BY BALANCED TRUNCATION and the relative model reduction problem can be formulated as inf G.1 Model Reduction by Balanced Truncation A B Consider a stable system G 2 RH1 and suppose G = C D is a balanced realization. Gr )Wi k1 deg(G ) r o with appropriate choice of Wi and Wo .3) A11 + A11 1 + C1 C1 = 0 (7. These problems can in general be formulated as frequency weighted model reduction problems inf kW (G .. A11 A12 G = 4 A21 A22 C1 C2 B1 B2 D 5: .5) (7. Theorem 7.2) can be written in terms of their partitioned matrices as A11 1 + 1 A11 + B1 B1 = 0 (7. Gr ) 1 if G is invertible. r deg(Gr ) r 7.6) 2 A21 + A12 1 + C2 C1 = 0 A22 2 + 2 A22 + B2 B2 = 0 (7.. Then both subsystems (Aii Bi Ci ) i = 1 2 are asymptotically stable. its controllability and observability Gramians are equal and diagonal.1 Assume that 1 and 2 have no diagonal entries in common. i. a practical model reduction problem is inherently frequency weighted.8) 2 The following theorem characterizes the properties of these subsystems. then A + A + BB = 0 (7:1) A + A + C C = 0: (7:2) Now partition the balanced Gramian as = 01 0 and partition the system 2 accordingly as 2 3 Then (7.e.1) and (7. We shall see in this chapter how the balanced realization can give an e ective approach to the above model reduction problems. In general.
10) from the right by obtain B1 1 V = 0: Then multiply (7.4) can be rewritten as (7:9) V (A11 + j!I ) = 0: (7.6) by V to get A21 2 V + 2 A12 1 V = 0 1 2 A21 V + A12 1 V = 0 which gives (A21 V ) 2 = 2 (A V ) : 1 2 21 . j!I ) 2 V = 0: 1 are in Ker(A11 .19. 1 can be assumed to be positive de nite without loss of generality. which is equivalent to C1 V = 0: Multiplication of (7. j!I ). j!I ).10) from the right by 1 V to get It follows that the columns of matrix 1 such that 2V 1 1V and from the left by V 1 to (A11 .1.10) (7. j!I ) 1 + 1 (A11 + j!I ) + B1 B1 = 0 1 (A11 . Model Reduction by Balanced Truncation 155 Proof. there exists a 2V 1 = V 2: 1 2 is the restriction of 2 to the space spanned by V . It is clearly su cient to show that A11 is asymptotically stable. rst multiply (7. It is then also possible to choose 1 diagonal and 1 such that the diagonal entries of 1 are a subset of the diagonal entries of 1 .3) and (7. The proof for the stability of A22 is similar. it follows that it is possible Since 1 1 to choose V such that 2 is diagonal.20 or Lemma 3. then there exists an eigenvalue at j! for some !. j!I )V = 0 which gives Equations (7. Therefore. Multiply (7.5) from the right by 1 V and (7. Then we have (A11 .11) (A11 .11) from the right by V now gives (A11 + j!I ) 1 V = 0: Analogously.11) from the right by V and from the left by V gives V C1 C1 V = 0. Let V be a basis matrix for Ker(A11 .7. j!I ) + (A11 + j!I ) 1 + C1 C1 = 0: Multiplication of (7. By Lemma 3. Then it is obvious that i (A11 ) 0 by Lemma 3.21. Assume that A11 is not asymptotically.
1 . s2 .2 3 (s . 1)(s . s + 2 = 4 . 2) = 4 0 . 2 . 2 stable.156 MODEL REDUCTION BY BALANCED TRUNCATION 2 1 This is a Sylvester equation in (A21 V ).12) and (7. This contradicts the fact that the original system is asymptotically stable.9) implies that A11 A12 A21 A22 V 0 = j! V 0 which means that the Amatrix of the original system has an eigenvalue at j!. Because common it follows from Lemma 2. Now (7.1:4142 0 1 is also a balanced realization with = I but one of the subsystems is not stable.1:4142 5 (s + 1)(s + 2) 2 1:4142 1 is a balanced realization with On the other hand. = I and every subsystem of the realization is stable. then every subsystem is asymptotically The stability condition in Theorem 7.1 1:4142 1:4142 5 :4142 0 0 s2 + s + 2 .2 If has distinct singular values. For example. Therefore A11 must be asymptotically stable. Corollary 7.1. 2 3 Theorem 7.2 .7 that and 2 2 have no diagonal entries in (7:12) A21 V = 0 is the unique solution.1 is only su cient.2:8284 .3 Suppose G(s) 2 RH1 and 2 A A 11 12 G(s) = 4 A21 A22 C1 C2 = diag( 1 B1 B2 D 2) 3 5 is a balanced realization with Gramian 1 2 = diag( 1 Is1 2 Is2 : : : r Isr ) = diag( r+1 Isr+1 r+2 Isr+2 : : : N IsN ) and 1 > 2 > > r > r+1 > r+2 > > N .
substituting for B1 B1 .1. Let (s) := (sI . Gr (s)k1 2( r+1 + r+2 + + N ) and the bound is achieved if r = N . Then the truncated system Gr (s) = A11 B1 C1 D is balanced and asymptotically stable.A12 . The case where the multiplicity of i is not equal to one is more complicated and an alternative proof is given in the next chapter. (j!)C (j!)C (j!) : (7:13) max ~ ~ ~ ~ Expressions for B (j!)B (j!) and C (j!)C (j!) are obtained by using the partitioned form of the internally balanced Gramian equations (7. C (s)B 1 1 .1. We shall now prove the error bound for the case si = 1 for all i. A21 (s)A12 ~ B (s) := A21 (s)B1 + B2 ~ C (s) := C1 (s)A12 + C2 then using the partitioned matrix results of section 2.5). G(s) .3){(7. we assume si = 1 and N = n.1 B1 . Hence. Furthermore kG(s) . ~ ~ An expression for B (j!)B (j!) is obtained by using the de nition of B (s). Gr (j!)] = 1=2 . 1. A). C1 (s)B1 sI . Gr (s) = C (sI . B1 B2 and B2 B2 from the partitioned form of the Gramian equations (7.1 (s)B (s) = C1 C2 computing this quantity on the imaginary axis to get h i ~ ~ ~ ~ G(j!) .. i. A22 B2 ~ ~ = C (s) .1 (s) := sI . The stability of Gr follows from Theorem 7. A11 ). we get ~ ~ B (j!)B (j!) = (j!) 2 + 2 (j!): ~ ~ The expression for C (j!)C (j!) is obtained analogously and is given by ~ ~ C (j!)C (j!) = 2 (j!) + (j!) 2 : . Model Reduction by Balanced Truncation 157 where i has multiplicity si i = 1 2 : : : N and s1 + s2 + + sN = n. GN .1 (j!)B (j!)B (j!) .1 B .7. A11 .e.8). kG(s) .3.A21 sI .3){(7. A22 .1 (s)k1 = 2 N : Proof.
Gr (j!)] Ek (j!)] 2 k+1 : This is the desired upper bound. Gk (s) for k = 1 2 : : : N . Gr (j!)] = n 1=2 1 + .1 X k=r Ek (s) N . Using triangle inequality we get G(j!) .14) is 2 N at ! = 0.158 MODEL REDUCTION BY BALANCED TRUNCATION ~ ~ ~ ~ These expressions for B (j!)B (j!) and C (j!)C (j!) are then substituted into (7. is an \all pass" scalar function. we note that (0) = I . Gr (j!)] = 1=2 max 2+ . 1. Then the right hand side of (7. (j!) (j!) = . (j!) 2 (j!) : Now consider onestep order reduction. . (7. Gr (s) = by the de nition of Ek (s)..1 (j!) 2 (j!) 2+ . Gr (j!)] n 1 + j (j! )j] = 2 n : (7:15) This completes the bound for r = n .15) holds.1 X k=r G(j!) . then 2= n and G(j!) . 1.e. Noting that G(s) . i. 1 and let GN (s) = G(s). r = n . we have N . 2 The singular values i are called the Hankel singular values. (This is the only place we need the assumption of si = 1) Hence j (j!)j = 1.13) to obtain G(j!) . The remainder of the proof is achieved by using the order reduction by one step results and by noting that Gk (s) = A11 B1 obtained by the \kth" order partiC1 D tioning is internally balanced with balanced Gramian given by 1 = diag( 1 Is1 2 Is2 : : : k Isk ): Let Ek (s) = Gk+1 (s) . A useful consequence of the above theorem is the following corollary. To see that the bound is actually achieved when r = N .1 X k =r N . 1.1(j!) 1 + (j!)] (7:14) max where := . Then Ek (j!)] 2 k+1 since Gk (s) is a reduced order model obtained from the internally balanced realization of Gk+1 (s) and the bound for onestep order reduction.
4 Let Then i 159 i = 1 : : : N be the Hankel singular values of G(s) 2 RH1 .9:6954 7 6 0 0 . r) r+1 . it is not hard to construct an nth order system so that the rth order balanced model reduction error is approximately 2 r+1 but the error bound is arbitrarily close to 2(n .0:0067 7 G(s) = 6 4 0:9160 . Then kG(s)k1 = G(0) = n and G(s) has the following state space realization 2 2 3 j =1 0 and the controllability and observability Gramians of the realization are given by 6 6 G=6 6 6 4 . 2 n 7 7 7 7 7 5 P =Q= i+j 2i + 2j 1 and P = Q ! 2 In as ! 1. Next make .6:3180 .5:4682 9:6954 0:9160 .. Model Reduction by Balanced Truncation Corollary 7. For example. consider the balanced realization of a fourth order system 2 . .e. The table shows that the actual error for an rth order approximation is almost the same as 2 r+1 which would be the estimated bound if we regard r+1 = r+2 = = 4 . So the Hankel singular values j ! 1 and 2( 1 + 2 + 2 + n ) ! n = kG(s)k1 as ! 1. 2 . G(1)k1 2( 1 + : : : + N ): The above bound can be tight for some systems. 4 .0:0420 0:2893 7 5 ..4:0051 . In general.1. consider an nth order transfer function n X 2j G(s) = s + 2j with > 0. . n 2n . kG(s) ..0:0020 0:0067 0:2893 0 with Hankel singular values given by 1 =1 2 = 0:9977 3 = 0:9957 4 = 0:9952: The approximation errors and the estimated bounds are listed in the following table. For example.6:3180 3 6 5:4682 0 0 0:2378 0:0020 7 6 .19:9579 .7.. i. The model reduction bound can also be loose for systems with Hankel singular values close to each other. One method to construct such system is as follows: Let G(s) be a stable allpass function. then there is a balanced realization for G so that the controllability and observability Gramians are P = Q = I . G(s) G(s) = I .0:2378 4:0051 .
9904 7. Assume that G Wi . Bounds: 2 4=r+1 i 2 r+1 kG .9772 3. Given the original full order model G 2 RH1 . This perturbed system will have the desired properties and this is exactly how the above example is constructed. and Wo have the following state space realizations A G= C B 0 B Wi = Ai Di Ci i B Wo = Ao Do Co o with A 2 Rn n .9818 1.9772 5.9954 1. the input weighting matrix Wi 2 RH1 and the output weighting matrix Wo 2 RH1 . Now the state space realization for the weighted transfer matrix is given by 2 A 0 BC BD 3 6 BoC Ao 0 i 0 i 7 =: " A B # : Wo GWi = 6 0 0 Ai Bi 7 4 5 C 0 Do C Co 0 0 Let P and Q be the solutions to the following Lyapunov equations AP + P A + B B = 0 QA + A Q + C C = 0: (7.9904 2 1. Gr )Wi k1 is made as small as possible.17) Then the input weighted Gramian P and the output weighted Gramian Q are de ned by n n P := In 0 P I0 Q := In 0 Q I0 : .160 MODEL REDUCTION BY BALANCED TRUNCATION a very small perturbation to the balanced realization then the perturbed system has a balanced realization with distinct singular values and P = Q I . our objective is to nd a lower order model Gr such that kWo (G .991 1.996 1.9904 i 7.9914 1. Gr k1 P r 0 1 2 3 2 1.2 FrequencyWeighted Balanced Model Reduction This section considers the extension of the balanced truncation method to frequency weighted case.16) (7. Note that there is no loss of generality in assuming D = G(1) = 0 since otherwise it can be eliminated by replacing Gr with D + Gr .
e.7.. 7.1 ) QT .1 TB CT . balanced) with 1 = diag( 1 Is1 : : : r Isr ) and and partition the system accordingly as 1 2 2 = diag( r+1 Isr+1 : : : n Isn ) TAT .1 = (i.3 Relative and Multiplicative Model Reductions A very special frequency weighted model reduction problem is the relative error model reduction problem where the objective is to nd a reduced order model Gr so that Gr = G(I + rel ) . Q can be obtained from Now let T be a nonsingular matrix such that TPT = (T . Relative and Multiplicative Model Reductions It can be shown easily that P and Q satisfy the following lower order equations 161 A BC 0 A i i P P12 P12 P22 A 0 BC A o + P P12 P12 P22 A 0 BC A o A BC 0 A i i + BD B i i BD B i i =0 (7 18) : Q Q12 Q12 Q22 o + o Q Q12 Q12 Q22 + CD C o o CD C o o =0 (7 19) : : P can be obtained from The computation can be further reduced if Wi = I or Wo = I .1 0 2A A B 3 1 11 12 = 4 A21 A22 B2 5 : C1 C2 0 Then a reduced order model Gr is obtained as Gr = A11 B1 : C1 0 Unfortunately. In the case of Wi = I .3. PA + AP + BB = 0 QA + A Q + C C = 0: (7:20) (7:21) while in the case of Wo = I . there is generally no known a priori error bound for the approximation error and the reduced order model Gr is not guaranteed to be stable either.
1 C D.1 C = 0: (7:23) (b) Suppose the realization for G is weighted balanced. 1: q q . 1 1 + 2 i ( 1 + i2 + i ) . In the case where G is square and invertible. D.1 C ) Q + C (D.5 Let G G. rel is usually called the relative error.. Furthermore k k rel k1 mul k1 N Y i=r+1 N Y i=r+1 1 + 2 i ( 1 + i2 + i ) . Theorem 7. the approximation Gr obtained below also serves as a multiplicative approximation: G = Gr (I + mul ) where mul is usually called the multiplicative error. this problem can be simply formulated as min G. as a bonus.162 MODEL REDUCTION BY BALANCED TRUNCATION Of course the dual approximation problem Gr = (I + rel )G can be obtained by taking the transpose of G.1 (s) = A .e.1 (G . i. BD.1 ) D. BD C . (a) Then the weighted Gramians P and Q for the frequency weighted balanced realization of G can be obtained as PA + AP + BB = 0 (7:22) Q(A . BD.1 2 RH1 be an nth order square transfer matrix with a state space realization A B G(s) = C D : and k rel k1 is made as small as possible.BD1 .1 . P = Q = diag( 1 Is1 : : : r Isr r+1 Isr+1 : : : N IsN ) = diag( 1 2 ) with 1 > 2 > : : : > N 0 and let the realization of G be partitioned compatibly with 1 and 2 as 2 3 Then A11 A12 B1 G(s) = 4 A21 A22 B2 5 : C1 C2 D Gr (s) = A11 B1 C1 D is stable and minimum phase.1 Let Wi = I and Wo = G.1 C ) + (A . We will show below that. Gr ) 1 : degG r r .
? Lemma 7. Now the output weighted transfer matrix is given by 163 Proof. 2 In the above theorem.3.BD C 0 . Without loss of generality. z is a transmission zero of G(s) if and only if A . it is obvious that the input weighted G. BD C BD? ). zI B C D I 0 0 D D? ] 2 I 0 4 . Relative and Multiplicative Model Reductions Gramian is given by P .6 A complex number z 2 C is a zero of G(s) if and only if z is an uncontrollable mode of (A .1 . we have assumed that the system is square.C I 0 0 I 0 0 3 5 . We refer readers to the references at the end of the chapter for details. Since the input weighting matrix Wi = I .1 (G .1 C 0 5 =: A B : 4 . we shall also assume that D is normalized such that DD = I . The proof for part (b) is more involved and needs much more work.1 D C D C 0 satis es the following Lyapunov equation Q := Q Q Q Q QA + A Q + C C = 0: Hence Q is the output weighted Gramian. D) = It is easy to verify that 2 A 3 " # 0 B . zI B C D does not have full row rank. Since D has full row rank and G(s) = C D is a minimal realization.1 C A . we shall now extend the results to include nonsquare case.7. A B Proof. Now note that A . Let D? be a matrix with full row rank such D that D is square and unitary. BD. Let G(s) be a minimum phase transfer matrix with a minimal realization A B G(s) = C D and assume that D has full row rank.
1 C = A . . BDa 1 Ca ) + (A . .D? B Q) (C . To motivate our choice. we have " ^ # " ^ # G(s) = G(s) (I + ) G(s) = G(s) (I + rel mul ) ^ ^ ~ ~ ~ ~ G(s) G(s) G(s) G(s) and ^ ^ G(s) = G(s)(I + rel ) G(s) = G(s)(I + mul ): ~ However. zI BD 0 I Then it is clear that BD? : 0 does not have full row rank if and only if A . zI B C D does not have full row rank. the corresponding P and Q equations are given by PA + AP + BB = 0 .QBD?D? B Q+C C +(C . zI BD? A B A B Ga := C D = 4 C D a a ~ C D? 2 3 5 = G(s) ~ G(s) is minimum phase. D .164 MODEL REDUCTION BY BALANCED TRUNCATION = A . BD C .BD C ) Q. BD? C is stable. BD C .7 There exists a matrix C such that the augmented system A . Proof. 2 ~ Corollary 7. BD C: A. Hence an appropriate choice of C is important.BD C )+(A. there are in general in nitely many choices of C and the model reduction ~ results will in general depend on the speci c choice.B D ?~ ~ C ? ~ ~ Hence C can always be chosen so that A . BD C .D?B Q) = 0 . Q(A . note that the equation for Q can be rewritten as ~ ~ Q(A. BDa 1 Ca ) Q + Ca (Da 1 ) Da 1 Ca = 0: Moreover. By PBH test. Note that the zeros of Ga are given by the eigenvalues of 2 If the previous model reduction algorithms are applied to the augmented system Ga . BD C . . BD C BD? ). this implies that z is a zero of G(s) if and only if it is an uncontrollable mode of (A .
i. BD? C = A . W (s)W (s) = G(s)G (s): Then W (s) can be obtained as W (s) = CA BW W D . An alternative relative/multiplicative model reduction approach. BDa 1 Ca = A . In the case where the model is not stable and /or is not minimum phase. i. Relative and Multiplicative Model Reductions A natural choice might be ~ C = D?B Q: 165 The existence of a solution Q to the following socalled algebraic Riccati equation Q(A . ~ A . BD C ) Q . which does not require that the transfer matrix be minimum phase but does require solving an algebraic Riccati equation. Let W (s) 2 RH1 be a minimum phase left spectral factor of G(s)G (s). Gap Gap = I . is the socalled Balanced Stochastic Truncation (BST) method. Let Gmp be a relative/multiplicative reduced model of Gmp such that ^ Gmp = Gmp (I + and ^ Gmp = Gmp (I + rel ) mul ): ^ ^ Then G := Gap Gmp has exactly the same right half plane poles and zeros as that of G and ^ G = G(I + rel ) ^ G = G(I + mul): Unfortunately.e. BD C .7. and Gmp is stable and minimum phase. the following procedure can be used: Let G(s) be factorized as G(s) = Gap (s)Gmp (s) such that Gap ^ is an allpass. BD?D? B Q is stable will be shown in Chapter 13.e. Let G(s) 2 RH1 be a square transfer matrix with a state space realization A B G(s) = C D and det(D) 6= 0.. this approach may be conservative if the transfer matrix has many nonminimum phase zeros or unstable poles.3. BD C . QBD? D? B Q + C C = 0 such that . BD C ) + (A .
1 (G . BW (DD ).1 (C . 1: i=r+1 1 ...1 BW X + C (DD ). n Isn 3 7 7 7 5 with 1 > 2 > : : : > n 0.24) (7:25) XAW + AW X + XBW (DD ).1 G(s).1 C = 0 with AW := A .1 i=r+1 1 .8 Let G(s) 2 RH1 have the following balanced stochastic realization A11 A12 B1 A B G(s) = C D = 4 A21 A22 B2 C1 C2 D with det(D) 6= 0 and P = X = diag(M1 M2 ) where 2 3 5 M1 = diag( 1 Is1 : : : r Isr ) M2 = diag( r+1 Isr+1 : : : n Isn ): Then is stable and ^ G = A11 B1 C1 D n Y 1+ i ^ G.1 C such that AW + BW (DD ). G) 1 .1 BW X is stable. BW X ) where P is the controllability Gramian given by AP + PA + BB = 0 and X is the solution of a Riccati equation (7. i is in fact the ith Hankel singular value of the socalled \phase matrix" (W (s)). G) 1 . Theorem 7.1 (G . The realization G is said to be a balanced stochastic realization if 2 I 6 1s P =X =6 6 4 1 2 Is2 . i n Y 1+ i ^ ^ G.166 with MODEL REDUCTION BY BALANCED TRUNCATION BW = PC + BD CW = D. i .
5 are the same and i = i = 1 + i2 if G(s) is minimum phase. The error bounds for the relative and multiplicative approximations using the selfweighted balanced realization were shown by Zhou 1993]. i 3 Y 1+ i 7. The error bound for the balanced model reduction was shown by Enns 1984] and Glover 1984] subsequently gave an independent proof.1 (G .1992] where it was claimed that the following error bounds hold i ^ G. The Balanced Stochastic Truncation (BST) method was proposed by Desai and Pal 1984] and generalized by Green 1988a.1988b] and many other people. 1 = 2:51522024046226: i=2 1 . G) 1 2 i 1.4 Notes and References The balanced model reduction method was rst introduced by Moore 1981].4. i: i=r+1 n X .7. Notes and References 167 It can be shown that the balanced stochastic realization and the selfweighted balanced p realization in Theorem 7. consider a third order stable and minimum phase transfer function s3 + 2s2 + 3s + 4 G(s) = s3 + 3s2 + 4s + 4 : It is easy to show that the Hankel singular values of the phase function is given by 1 = 0:55705372196966 2 = 0:53088390857035 3 = 0:03715882438832 and a rst order BST approximation is given by ^ s + 0 00375717470515 G = s + 0::00106883052786: The relative approximation error is 2:51522024045904 and the error bound is . To illustrate. G) 1 2 1. i i=r+1 n X ^ ^ G.1 (G . It was also shown in Zhou 1993] that the frequency selfweighted method and the BST method are the same if the transfer matrix is minimum phase. The stability properties of the reduced order model were shown by Pernebo and Silverman 1982]. Improved error bounds for the BST reduction were reported in Wang and Safonov 1990. The frequency weighted balanced model reduction method was also introduced by Enns 1984] from a somewhat di erent perspective. The relative error bound for the Balanced Stochastic Truncation method was obtained by Green 1988a] and the multiplicative error bound for the BST was obtained by Wang and Safonov 1992].
Glover 1986. Discrete time balance model reduction results can be found in AlSaggaf and Franklin 1987]. Hung and Glover 1986] and references therein. i = 2:34052280324021 i=2 which is smaller than the actual error.1989]. the example in the last section gives 2 i 1 . . Limebeer and Hung 1992]. Glover.168 MODEL REDUCTION BY BALANCED TRUNCATION 3 X However. Other weighted model reduction methods can be found in AlSaggaf and Franklin 1988]. Hinrichsen and Pritchard 1990]. and references therein.
G(s) = k+1 where 1 > 2 : : : > k+1 : : : > n are the Hankel singular values of G(s). H The solution to this approximation problem relies on the allpass dilation result of a square transfer function which will be given for a general class of transfer functions. nd G(s) of McMillan degree k < n such that G(s) .Hankel Norm Approximation This chapter is devoted to the study of optimal Hankel norm approximation and its applications in L1 norm model reduction. This representation is obtained by reducing the order one dimension at a time via optimal Hankel norm approximations. G(s) is minimized. which gives ^ inf G(s) . we show that a square stable transfer function G(s) can be represented as G(s) = D0 + 1 E1 (s) + 2 E2 (s) + : : : + n En (s) where Ek (s) are allpass functions and the partial sum D0 + 1 E1 (s) + 2 E2 (s) + : : : + k Ek (s) have McMillan degrees k . This representation also gives that kGk1 2( 1 + : : : + n ) and further that there exists a constant D0 such that kG(s) . The Hankel operator is introduced rst together with some time domain and frequency domain characterizations. D0 k1 ( 1 + : : : + n ): 169 H 8 . The optimal Hankel norm approximation problem can be stated as follows: Given G(s) of McMillan ^ ^ degree n. Moreover. The allpass dilation results are then specialized to obtain the optimal Hankel norm approximations.
Then the time domain Hankel operator is .g f )(t) = . D0 1 ( k+1 + : : : + n ): We shall also provide an alternative proof for the error bounds derived in the last chapter for the truncated balanced realizations using the results obtained in this chapter.1 0 t < 0: Because of the isometric isomorphism property between the L2 spaces in the time domain and in the frequency domain. .G = P+ MGjH? .G k : . The Hankel operator associated with G will be denoted by . )f ( )d t 0 (.G f := (P+ MG)f = P+ (Gf ) for f 2 H2 i. Let g(t) denote the inverse (bilateral) Laplace transform of G(s).1 0] 7.G : H2 7.g f := P+ (g f ) for f 2 L2 (.170 HANKEL NORM APPROXIMATION The above bounds are then used to show that the kth order optimal Hankel norm ^ approximation.G and is de ned as ? . 8.! L2 0 1) .e.! H2 MG ? H2 >  L2 P+ ..g : L2 (. G(s) . G(s).1 Hankel Operator Let G(s) 2 L1 be a matrix function. together with some constant matrix D0 satis es ^ G(s) . we have k.g k = k. This is shown in the following diagram: 2 ? .1 0]: Thus R 0 g(t .G H2 ? There is a corresponding Hankel operator in the time domain. Finally we consider the Hankel operator in discrete time and o er some alternative proof of the wellknown Nehari's theorem.
1) = 0.1 CeA(t. MG jH2 .G can be computed easily from the de nition as below: let ? f 2 H2 g 2 H2 .g u(t) = Z0 . there are Gs (s) 2 RH2 and Gu (s) 2 RH? such that 2 G(s) = Gs (s) + G(1) + Gu (s): ? Now for any f 2 H2 .g can be written as . In particular. we will now further restrict G to be rational. there is no loss of generality in assuming G 2 RH1 and strictly proper.8.. G gi (since f and P+ G g are orthogonal ): ? Hence . G (s) 2 RH1 . the Hankel operator associated with G 2 RL1 depends only on the strictly causal part of G. The adjoint operator of . Then the Hankel operator .Gg = P. G(s) 2 RL1 . Hankel Operator 171 Hence. Gf and g are orthogonal ) = hGf gi = hf G gi = hf P+ G gi + hf P.g u(t) t 0 based on the past input u(t) t 0. the Hankel operator can be more speci cally decomposed as the composition of maps from the past input to the initial state and then .1.e. (G g) : H2 7. i.e.e. then . Then G can be decomposed into strictly causal part and anticausal part.G = P.Gf = P+ (Gf ) = P+ (Gs f ): Hence.! H2 or . i.. Therefore. G gi = hf P.. )Bu( )d for t 0 and has the interpretation of the system future output y(t) = . it is easy to see that . then h. if G is antistable. Now suppose G 2 RH1 has a state space realization as given below: x = Ax + Bu _ y = Cx with A stable and x(. In the state space representation.G = 0.G f gi := hP+ Gf gi = hP+ Gf gi + hP. For the interest of this book. in this book we will use the time domain and the frequency domain descriptions for Hankel operators interchangeably. Gf gi (since P. i.
) Clearly.1 0] 7.1 e.! L2 0 1).1 0] 7.L2 0 1 ) o The adjoint operators of c and o can also be obtained easily from their de nitions as follows: let u(t) 2 L2 (.g = o c : o : C n 7.A x0 iL2 (. It is easy to verify that . x0 = c u(t) for u(t) 2 L2 (.1 0] = hu c x0 iL2 (. x0 2 C n .! Rn and o : Rn 7. then h cu x0 iC n = Z0 u ( )B e.1 0] is the system state at t = 0 due to the past input and y(t) = o x0 t 0. and y(t) 2 L2 0 1).! L2 0 1) o x0 := CeAt x0 t 0: c > Cn ZZ Z L2 (.! C n c u := Z0 and .A Bu( )d (If all the data are real. o. c .1 0] .1: System theoretical interpretation of Hankel operators from the initial state to the future output.1 0].A x0 d = hu B e. respectively. then the two operators become c : L2 (.1 0] . These two operators will be called the controllability operator. and are de ned as c : L2 (. and the observability operator.1 .172 u(t) HANKEL NORM APPROXIMATION Γ g y(t) 0 t 0 t Figure 8. is the future output due to the initial state x0 with the input set to zero.g ZZ ~ Z .
Lc = Lo = Then we have Z1 Z0 0 1 eAt BB eA t dt eA t C CeAt dt: Lcx0 Lox0 for every x0 2 C n . Hankel Operator and 173 h o x0 yiL2 0 1) = Z1 0 x (t)eA t C y(t)dt = hx0 Z1 0 eA t C y(t)dtiC n = hx0 o yiC n where h iX denotes the inner product in the Hilbert space X .1 0] be a corresponding eigenvector.g : .g . Thus Lc and Lo are the matrix representations of the operators c c and o o .g . Theorem 8.g (or .e.1) by c and de ne x = cu 2 C n to get LcLo x = 2 x: (8:2) 2 .A x0 0 c x0 = B e and n o : L2 0 1) 7. In particular k.1 0] .g k = (Lc Lo).g u = c o o cu = 2 u: (8:1) Premultiply (8. )C y(t)dt 0: Let Lc and Lo be the controllability and observability Gramians of the system..8.! C o y (t) = Z1 0 eA t C y(t)dt: This also gives the adjoint of . we have n c : C 7.G.G) and the matrix LcLo have the same p nonzero eigenvalues. Let 6= 0 be an eigenvalue of . i.g y = Z1 0 B eA (t. c c x0 = o o x0 = Proof.! L2 (.g .g = ( o c) = c o : L2 0 1) 7.1 0] . and let 0 6= u 2 L2 (.! L2 (.1.1 The operator .g . Then by de nition . Therefore.
g k2 = .1 0] be a corresponding eigenvector.1 0]..1 0] v = x 2 L 0 1): i c i o i 2 i o i 2 ~ Remark 8. the Hankel operator can be written as . The proof given above suggests a way to construct this pair: nd the eigenvalues and eigenvectors of LcLo .g .2 As seen in various literature.g . suppose 2 6= 0 and x 6= 0 are an eigenvalue and a corresponding eigenvector of LcLo . let us examine some of the alternatives below: (i) Let v(t) = u(.g : L2 0 1) 7.174 HANKEL NORM APPROXIMATION Note that x = cu 6= 0 since otherwise 2 u = 0 from (8.1). and then v(t) 2 L2 0 1).e.g and 0 6= u 2 L2 (. i2 and xi such that Lc Loxi = i2 xi : Then the pairs (ui vi ) given below are the corresponding Schmidt pairs: u = ( 1 L x ) 2 L (. since G(s) is rational.! L2 0 1) or .! H2 (.1) which is impossible. 2 Remark 8. Hence k.g v)(t) = = Z 0 t 0 t<0 1 CeA(t+ )Bv( )d for t 0: 0 R 1 g(t + )v( )d 0 . Premultiply (8.g .g u = .1 Let 6= 0 be an eigenvalue of .g v = v u: This pair of vectors (u v) are called a Schmidt pair of .g . Hence.g = (Lc Lo). De ne v := 1 . there are some alternative ways to write a Hankel operator. So 2 is an eigenvalue of LcLo . i.G : H2 7. It is easy to see that u 6= 0 since c u = c c Lox = LcLo x = 2 x 6= 0. Therefore 2 is an eigenvalue of . On the other hand.t) for u(t) 2 L2 (.g .2) by c Lo and de ne u = c Lox to get (8. . u 2 L 0 1): 2 g 2 Then (u v) satisfy .g . Finally. For comparison.g is compact and selfadjoint and has only discrete spectrum.
.8. A Let G(s) = C B be an antistable transfer matrix. MGjH2 : H2 7. It will be clear in later chapters that this operator is closely related to the problem of approximating an anticausal function by a causal function. (Gv) ui = hGv ui = hv G ui = hv P+ (G u)i: ~ 8. i.g v)(t) = = Z 0 t 0 t>0 1 CeA(t.e. This transfer function dilation is the key to the optimal Hankel norm approximation in the next section.. it is more convenient to work with an anticausal operator G and view the Hankel operator associated with G as the mapping from the future input to the past output.2 Allpass Dilations This section considers the dilation of a given transfer function to an allpass transfer function.! L2 (. )v( )d 0 ^ .1 The inertia of a general complex.2.G . De nition 8.1 0] ^ (.Gv = P.G = P. . But rst we need some preliminary results and some state space characterizations of allpass functions.g : L2 0 1) 7. Then the Hankel operator associated with G(s) can be written as ^ .! H2 hP. which is the problem at the heart of the H1 Control Theory. Allpass Dilations 175 (ii) In some applications. all the eigenvalues of 0 A have positive real parts. (A) = number of eigenvalues of A on the imaginary axis. we have ^ Hence. or in the frequency domain ? ^ .G = . (A) = number of eigenvalues of A in the open left halfplane. (Gv) for v 2 H2 : ? Now for any v 2 H2 and u 2 H2 . )Bv( )d for t 0 0 R 1 g(t . square matrix A denoted In(A) is the triple ( (A) (A) (A)) where (A) = number of eigenvalues of A in the open right halfplane.
Hence (A) (P ). and 0 0 de ne ~ ~ A = U AU = A11 A12 B = U B = B1 : A A B Then U (8:3)U gives 21 22 2 AP + PA ~ ~ ~~ A P1 0 + P1 0 A + B B = 0 0 0 0 (8:4) ) B2 B2 = 0 ) B2 = (8:4) (8:5) ) A21 P1 = 0 ) A21 = (8:4) ) A11 P1 + P1 A11 + B1 B1 = ) (by part (1)) (A11 ) (A11 ) but since (A11 ) = (P1 ) = 0 (P1 ) = (A11 ) (A) (P1 ) = (A11 ) (A): 0 0 0 0 (8. C 2 C m n . Similarly. and a corresponding eigenvector. such that x A = x and x Px > 0. (8:3) Theorem 8. we can show that (A) (P ). (2) Assume (A) = 0 and that P = U P1 0 U with (P1 ) = 0 U U = I .6) (8. and hermitian matrix P = P satisfying then (1) If (P ) = 0 then (A) (2) If (A) = 0 then (P ) Proof. (1) If (P ) = 0 then observe that (8. x 2 C n . B 2 C n m . 2 C with Re( ) > 0.4) (8. then ..5) (8.176 HANKEL NORM APPROXIMATION AP + PA + BB = 0 (P ) (A) (A) (P ) (P ).e. a contradiction. (A).2 Given complex n n and n m matrices A and B. Re( ) 0.3 Given a realization (A B C ) (not necessarily stable) with A 2 C n n .7) (P1 ) (P1 ) 2 Theorem 8.3) implies 0: Now suppose (A) > (P ) then there is an eigenvalue of A. which implies x (AP + PA )x = ( + )x Px 0 i.
1 C D. G(1) = D ) DD = I .C : B D These two transfer functions are identical and both minimal (since (A B C ) is assumed to be minimal). observability and controllability are not assumed). Q = Q= .12) (8. BD C )(T ).1 (s) = A .1 T (A .8) (8. Also GG = I ) G = G.2.e. (1a) ) (1b) This is proved by constructing P and Q to satisfy (1b) as follows.1 C .1 : (8.11) (8. A).1 DC D = G = . C = C= . Hence.8. (b) there exist P Q 2 C n n such that (i) P = P .1 .1 T BD . G.BD D.e.1 C DB T )(T ).1 = A . . without loss of generality the proof will be given for the case = 1 only. P = P= .A . i. and hence there exists a similarity transformation T relating the statespace descriptions. Proof. BD C . Given (1a).1 B .9) (8. i.A = T (A .10) (8.. Q = Q (ii) AP + PA + BB = 0 (iii) A Q + QA + C C = 0 (iv) PQ = 2 I (2) Given that part (1b) is satis ed then there exists a D satisfying D D = 2I D C +B Q = 0 DB + CP = 0 and any such D will satisfy part (1a) (note.1 : .1 C B = TBD = D CT . Any systems satisfying part (1a) or (1b) can be transformed to the case = 1 p ^ p ^ ^ ^ ^ by B = B= . Allpass Dilations 177 (1) If (A B C ) is completely controllable and completely observable the following two statements are equivalent: (a) there exists a D such that GG = 2 I where G(s) = D + C (sI . BD.A D C (T ).C DB + (T .BD. (T ).. D = D= . BD C )T .1A T ) T (A .13) Further = = = = (8:9) and (8:10) ) = (8:9) ) B (8:10) ) C (8:8) ) .
Q ((1b) equation (ii)) Q gives QA + A Q + QBB Q = 0 (8:18) which together with part (1b). A).1 = .15) (8:16) (8:17) TA + A T .1 C + C (sI .CQ. 2 . and minimality implies these have a unique solution and hence T = T .sI . (1b) ) (1a) This is proved by rst constructing D according to part (2) and then verifying part (1a) by calculation.T P = .1 C .sI .1A . Firstly note that since Q = P . (8.C DB = .T .1 C CT .10) and (8. T and T satisfy identical equations.11) to (8.16) implies AT .20) (8.20) and (8. (8.CP: Equations (8. A ). Also (8.sI . C C = 0 which veri es (1b). Now setting Q = .DB (.1 = 0 which together with (8.8) and (8.21) DB Q = . A ).14 there exists a D such that D D = I and (8:19) (8.1 BD : BB Hence. equation (ii).10) implies part (1b).14) (8. A)P + P (. Further.13). T .1 C = (sI .178 HANKEL NORM APPROXIMATION Hence. equations (i) and (iv).1 BB (. A).21) imply that the conditions of part (2) are satis ed. equation (iii) implies that QBB Q = C C and hence by Lemma 2. equation (iii).9) imply (8.sI . A). A ). The following theorem dilates a given transfer function to an allpass function. Now note that ) C (sI .8) to (8.1 clearly satis es part (1b). on expanding G(s)G (s) we get G(s)G = I: Part (2) follows immediately from the proof of (1b) ) (1a) above. C (sI . A ) = CP (.1 PC (8:21) ) = .1 .1 + T .
Partition (A B C ) conformably with P . as A = A11 A12 A21 A22 ^ ^ and de ne W (s) := A B ^ D C ^ ^ A ^ B ^ C ^ D = = = = with B = B1 B2 C = C1 C2 (8:26) .8.30) (8.1 ( 2 A11 + 2 A11 . = Also de ne the error system 1 2. 2 I: B E (s) = G(s) ..1 ( 2 B1 + C1 U ) C1 1 + UB1 1.28) (8.27) (8.23) (8. Allpass Dilations A B Theorem 8.4 Let G(s) = C D with A 2 C n Cm 179 n m satisfy B 2 Cn m C 2 Cm n D 2 (8.24) (8.32) where U is a unitary matrix satisfying B2 = .C ^ De = D .2.22) (8. 6= 0 and ( 1 2 .31) (8.C2 U and .25) AP + PA + BB = 0 A Q + QA + C C = 0 for P = P = diag( Q = Q = diag( 1 2 Ir ) Ir ) with 1 and 2 diagonal. C1 UB1 ) D. D: (8:33) . W (s) = Ae De Ce e with 0 Ae = A A 0 ^ Then Be = B ^ B ^ Ce = C ..29) (8. 2 I ) = 0. U (8.
..34) (8.41) ^ A21 + B2 B = 0 (8.43) .37) (8. < 0 and 2 . The (2.1 + 2 .180 (1) (Ae Be Ce ) satisfy HANKEL NORM APPROXIMATION Ae Pe + Pe Ae + Be Be = 0 Ae Qe + Qe Ae + Ce Ce = 0 with 2 (8. > 0 and 2 . > 0 or (ii) 1 .22) which is assumed.22) and (8. < 0 then ^ ^ ^ (A B C ) is a minimal realization. 0 1 .14 there exists a unitary U satisfying (8.31).23) give (8. together with ^ ^ A11 + A + B1 B = 0 (8.31).) ^ ^ ^ (c) If P > 0 Q > 0 then the McMillan degree of the stable part of (A B C ) equals ( 1 . (d) If either (i) 1 . To verify (8.) = ( 2 . 2 ..34) to (8..39) (8.35) 3 Pe = Qe = Pe Qe = (2) E (s)E (s) = 2 I .2) blocks of (8. 0 1 0 Ir I (8. It is rst necessary to verify that there exists a unitary matrix U satisfying (8.42) ^ 2 .40) and hence B2 B2 = C2 C2 and by Lemma 2. For notational convenience it will be assumed that = 1 and this can be done A22 + A22 + B2 B2 = 0 A22 + A22 + C2 C2 = 0 without loss of generality since B C and can be simply rescaled to give = 1. as follows. (1) The proof of equations (8. 1 . Proof. (3) If (A) = 0 then ^ (a) (A) = 0 (b) If ( 1 2 ) = 0 then 4 2 4 2I 0 5 I 0 2 .1 3 0 .1 A + B B = 0 ^ ^^ A (8.38) ^ In(A) = In(.).38) is by a straightforward calculation.34) and (8.36) (8.) = In(.36) we need (8. 2 0 Ir 0 5 ..
1 A11 2 + B1 U C1 ).53) Equations (8.48) where B1 B1 was substituted using the (1 1) block of (8. A12 ) 2 + A12 2 + A21 )..A11 1 2 .50) ^^ = .51) to (8.) + (. the (1 1) blocks of (8..A + C C = 0: (8..2.53) are now veri ed in an analogous manner to equations (8.48).23).1 .35) and (8.. ^ B2 B = (8.41) ^ B1 B = (B1 B1 2 + B1 U C1 )..1 ((.1) blocks of (8.50) is a rearrangement of (8.( 2 B1 + C1 U )(B1 2 + U C1 ) + (A11 2 + 2 A11 + C1 C1 ) + 2 (A11 1 + 1 A11 + B1 B1 ) 2 (8.A12 2 ..8.22) and (8..1 ^ = . ) (8:43) 1 . B1 U C1 ) where (8.41) to (8. Equations (8.) .34) and (8. Finally to verify (8. C1 C = 0 (8. C2 C (8. Allpass Dilations which will now be veri ed.. = (A11 + 2 A11 (8.45) using the (2.23) are used.31) ) = = = 181 B2 (B1 2 + U C1 ). respectively.)A .22) and (8.1 (B2 B1 2 .A12 .43) ^ C2 C = C2 (C1 1 + UB1 ) = (.47) (8.51) ^ = 0 A12 (.45) (8.27) substituting in (8. ) (8:52) ^ C1 C = C1 C1 1 + C1 UB1 = .43) consider ^ ^ . B2 B1 = .BB . 2 A11 1 + C1 UB1 .27)!(8.37) we need (8.49) and (8.52) ^ ^ ^ ^ A 1 .22) and (8. To verify (8. A21 ) 1 .A11 .36) are hence veri ed. C1 UB1 ) 2 + 2 (A11 + 1 A11 2 .A21 ) (8:42) (8.A21 1 .23) which is assumed together with ^ ^ A11 (.44) (8.49) and nally. C2 C1 ). + 1 . A ) (8:41) (8.A 2 + 2 A .1 = (. Similarly in order to verify (8.49) = .A11 2 1 ..46) where B2 B1 and C2 C1 were substituted in (8.
Hence ^ In(A) = In(. (2) Equations (8. (8.A11 . + = 0 and .2 can be applied since equations (8. .3. ^ (3) (a) To show that (A) = 0 if (A) = 0 we will assume that there exists x 2 C n.x = 0: (8:60) Equations (8.3.60) imply that (x .) We hence need to verify that De De = I (8. ^ ^ ( + )x 1 .43) and (8. 1 . part (1b) are satis ed and Theorem 8. 0) but since it is assumed ^ that (A) = 0.x .57) ^ ) Cx = 0: (8.A ) (8:51) ^ ^ A 1 . 1 .59) Also (8:52)x and (8.. + 1 . x + C1 Cx = 0 ) x . From and 2 C such that Ax x (8:53)x.x .) = In(.1 ) = ( 1 .59) and (8. (x .182 HANKEL NORM APPROXIMATION ^ = .r ^ = x and + = 0.( 1 C1 + B1 U )(C1 1 + UB1 ) + 1 (A11 2 + 2 A11 + C1 C1 ) 1 + (A11 1 + 1 A11 + B1 B1 ) ^ C ) (8:53) = .33) makes E (s) allpass.) = 0) Theorem 8.1 ) = In(.53) have been veri ed.r and 2 C such that Ax = x and ^ ^ ^ Cx = 0 (i.55) follows by substituting the de nitions of B C De and Q.35) and (8.34).A12 . 0)A = .x = 0 .A = (A11 + 1 A11 2 . x .) ^ (c) Assume that there exists x 6= 0 2 C n.C ^ Therefore. Then (8:51)x and (8:52)x give . and the proof of part (1) is complete. part (2) can be used to show that the De given in (8.. ^ (b) Since (A) = 0 has been proved and ( 1 2 ) = 0 is assumed () ( 2 . x = 0 .A11 .54) is immediate.56) follows from De (8:55) Pe . . C1 UB1 ) = .: (8.56) ^ ^ Equation (8. (8. and show that this implies x = 0.54) De Ce + Be Qe = 0 (8.58) Now (8:51)x gives ^ .A11 ..x + x C Cx = 0 (8. (C A) is not completely observable).37) have been veri ed.A11 = .e.1 exists this implies that x = 0 and (A) = 0 is proven.58) give A12 . (8. . (8. .35) and (8. B1 U C1 ) 1 + 1 (A11 + 2 A11 1 .38) ensure the conditions of Theorem 8. 2 .55) De Be + Ce Pe = 0: (8. and (8.. (Note it is still assumed that = 1.38) is immediate.
= 1.x = 0 ^ ^ but ( + ) 6= 0 by (3a) and 1 . 13ss+ 30 G(s) . r 0) and by (3c) the ^ B C ) is n .ss+ 2)(s. 1 = diag(2 1 ).4. 3 Example 8. < 0 is assumed so that x = 0. ) is an eigenvalue of A since .42) will give the analogous conclusion and therefore all the modes in the left halfplane are controllable and observable.43) gives (A 2 Example 8. ^ ^ McMillan degree of (A ^ ^ (ii) Assume there exists x such that Ax = x and Cx = 0. Hence (C A) is ^ B ) completely controllable.4 4 . 1 ( 2 .2 .1 Take This has a balanced realization given by 2 3 2 3 . and the condition in (3b) gives their number. 2 I ) is inde nite.4 when ( 2 . ^ completely observable. part (2b) implies that ^ In(A) = In(.2 C = . Allpass Dilations 183 hence (.4 5 B = 4 1 5 C = 2 . = 2 gives 1 0 2 . W (s) = 2(.41) and (8.8.3=2 D = 1: 8 3 Theorem 8.6 0 = (1 1 0) 3 1 0 8 = 1 0 .2 Let us also illustrate Theorem 8.s 5)25) s2 s 7s 7s 30) .4 . However.9 6 Now using the above construction with 1 2 + 105 G(s) = 39(s + 2)(s s + 250 : s + 5)2 2 =4 1 0 0 0 1 0 2 0 0 2 3 5 2 ^ ^ ^ ^ 1 2 B = 1 . 2 I )) = In . + W (s) is an optimal anticausal approximation to G(s) with L1 error of 2.4 now gives 2 ^ ^ ^ ^ A = .5=2 D = 2 A = 3 .4 2 A = 4 . since P > 0 and (A) = 0 are assumed then In(A) = (0 n 0) and all the unobservable modes must be in the open ^ ^ right half plane. > 0 then by (3b) In(A) = (0 n .x 6= 0.2 4 .1 . ^ (d) (i) If 1 .3 2 B = .8 10 5 3 2 + 90 + 2)( + 2 (3 2 + + 30) W (s) = 63ss2.1 6 .2 C = 6 . The construction of 2 Theorem 8. Similarly. Similarly (8.2. > 0 or 2 . if it is assumed that (A B ) is not completely controllable then (8.7 ( + (3 . Then x (8:53)x gives ( + )x 1 . Take 1 G(s) as in the above example and permute the rst and third states of the balanced 1 realization so that = diag(2 2 1). r and the result is proven.
4. 3 Example 8.3 Optimal Hankel Norm Approximation We are now ready to give a solution to the optimal Hankel norm approximation problem based on Theorem 8.184 HANKEL NORM APPROXIMATION ^ which is veri ed by noting that A has eigenvalues of 5 and .s + 2)(s ++ 5) : ( 5)2 2 In general the error E (j!) will have modulus equal to but E (s) will contain unstable poles. ^ ^ ^ jE (j!)j = 1 by Theorem 8.3 Let us nally complete the analysis of this G(s) by permuting the second and third states in the balanced realization of the last example to obtain 1 = 2 .++ 5) s(s+.4. 1 (. part (3c). .4.6 shows that the construction of Theorem 8.s E (s) = (.3=4) so that A is stable by Theorem 8.15 . this example has been constructed to show that (A B C ) itself may not be minimal when the conditions of part (3d) are not satis ed. G(s) k+1 (G) ^ G H and then Theorem 8. 3 8.6 " 1 = diag(2 1).4 . 2 W (s) is in fact an optimal secondorder Hankel norm approximation to G(s). The following Lemma gives a lower bound on the achievable Hankel norm of the error ^ inf G(s) .5. part (2). = diag(. However.4. " # ^ ^ A B W (s) = ^ ^ = ss+ 20 +5 C D and we note that the stable part of W (s) has McMillan degree 1 as predicted by Theo^ ^ ^ rem 8.4.(s 2 + 123s+ 10) s + 21s C D s s 2 2 s E (s) = G(s) .4 can be used to achieve this lower bound. 1 . 21++ 10) : 2 ( + 2)(s 5)2 ( 2 21s 10) ^ Note that . ss+20 is in fact +5 an optimal Hankel norm approximation to G(s) of degree 1 and s + 2)(.15=2 . (A B C ) is minimal by Theorem 8. W (s) = . part (3d). We will nd ^ A = .s + 2)(. part (3b). and in this case the unstable pole at +5 is both uncontrollable and unobservable.1 2 ^ ^ A B 1 2 + 110) W (s) = ^ ^ = 2 (.20 . ^ B= 4 4 # ^ C = 15 3 ^ D = .
G(s)) = i (PQ) = i (RR Q) = i (R QR) n In 0 R QR I0 i R 0 Q R11 = i 11 = = = = i (R11 Q11 R11 ) = i (Q11 R11 R11 ) i (Q11 (P11 .8.37) and write P = P11 P12 P12 P22 Since P where with 0 it can be factorized as Q = Q11 Q12 Q12 Q22 P = RR 11 R = R0 R12 R22 P11 Q11 2 Rn n : 1 .34) and (8.61) by setting ^ ^ G(s) = (G(s) .62) follows from (8. G satisfy (8. XX ) where X = Q11 R12 1=2 1=2 i+k (Q11 P11 Q11 ) 2 i+k (P11 Q11 ) = i+k (G) 0 (8. (.36) and (8. G(s) H Proof. p m. G(s)) ^ i+k (G(s) .3.33) will be a state space realization of G(s) . then for all G(s) stable Lemma 8. G(s)) In particular.5 Given a stable. then (Ae Be Ce ) given by ^ (s). transfer function matrix G(s) with Hankel ^ i (G(s) .G(s)): ^ ^ ^ ^ Let (A B C ) be a minimal state space realization of G(s). R12 R12 ^ ^ ^ (P22 > 0 since (A B C ) is a minimal realization. rational.62) (8:63) ^ G(s) . We shall prove (8. k i (G(s)) i = 1 : : : n: k+1 (G(s)): (8. i+k (G(s)) i = 1 : : : n . Now let P = P and Q = Q (8.35) respectively (but not necessary (8. R22 = P22=2 R12 = P12 P221=2 R11 R11 = P11 . Optimal Hankel Norm Approximation singular values 1 2::: and of McMillan degree k k k+1 k+2 : : : n 185 ^ > 0. G(s)) .) ^ i (G(s) .61) (8.64) .61) only and the inequality (8. R12 R12 )) 1=2 1=2 1=2 i (Q11 P11 Q11 .
G(s) . Theorem 8. ^ Proof. approximation to G(s) if and only if there exists F (s) 2 H1 (whose McMillan ^ degree can be chosen n + k .186 HANKEL NORM APPROXIMATION where (8.64) follows from the fact that X is an n k matrix () rank(XX ) k). G . F )f ^ P+ (G . can be constructed as follows.G H k+1 (G(s)) (8. k . Then # " ^ ^ ^ (s) + F (s) = A B (8:67) G ^ ^ C D . and G(s) of McMillan degree H ^ G2H1 F 2H. By the de nition of L1 norm. r. 1 1 k ^ G(s) .6 Given a stable. .26) to (8. G(s) . for all F (s) 2 H1. F (s) satis es 2 E (s)E (s) = k+1 I: (8:65) In which case ^ G(s) . G(s) = k+1 : (8:66) (3) Let G(s) be as in (2) above. ^ ^ where G(s) 2 H1 and F (s) 2 H1 with the McMillan degree of G(s) = k and the McMillan degree of F (s) = n .31).68) . 1) such that E (s) := G(s) . then an optimal Hankel norm approximation of McMil^ lan degree k. G(s) . McMillan degree (G) k. G(s). F (s) . F (s) 1 = = f 2H? kf k2 1 2 f 2H? kf k2 1 2 f 2H? kf k2 1 2 sup sup sup ^ (G(s) . F (s))f ^ P+ (G . Let (A B C ) be a balanced realization of G(s) with corresponding = diag( 1 2 : : : k k+r+1 : : : n k+1 : : : k+r ) ^ ^ ^ ^ and de ne (A B C D) from equations (8. G(s) . transfer function G(s) then ^ ^ (1) k+1 (G(s)) = inf G(s) . rational. m m. G)f 2 2 2 ^ G. 2 We can now give a solution to the optimal Hankel norm approximation problem for square transfer functions. (2) If G(s) has Hankel singular values 1 2 : : : k > k+1 = k+2 : : : = k+r > ^ k+r+1 : : : n > 0 then G(s) of McMillan degree k is an optimal Hankel norm .
part (2). equation (8. G(s) to produce an optimal anticausal approximation F (s). the McMillan degree of (G(s) . Ga H = k+1 (Ga ) = k+1 (G) k+1 (G) ^ i.4. as in part (3). G is an optimal Hankel norm approximation of G(s).68) becomes equalities. (2) i. consider a stable. G(s)) minus ^ ).3. rational. Proof.4.5.65) holds and hence kE (s)k1 = k+1 : (8:69) Also from Theorem 8. Clearly the su ciency of part (2) can be similarly veri ed by noting that (8. k . 1.65) implies that (8. transfer function of McMillan degree n such that 1 (G) has multiplicity r1 .8. Then (1) (G(s) .r1 (F (. part (1). G H Ga . n + k . F (s))= 1 is allpass. Ga H = k+1 (Ga ): Then ^ ^ G . ^ Now de ne G(s) and F (s) via equation (8. Also let F (s) be an optimal anticausal approximation of degree n . F (. . (1) is proved in Theorem 8. and therefore (8. F (s))= k+1 (G) is allpass since k+1 (G) = 1 (G .66) holds. (2) is obtained from the forms of Pe and Qe in Theorem 8.68) is satis ed with equality. and part (1) is proven. 2 The optimal Hankel norm approximation for nonsquare case can be obtained by rst augmenting the function to form a square function.. G has McMillan degree k and it in the correct class. G ^ the McMillan degree of this F (s) will be. r k 0): (8:70) 1 ^ Hence. r1 given by the construction of Theorem 8. part (3b) 2 ^ In(A) = In(. p m (p < m). then Theorem 8.e. part (2) implies that (8.7 Let G(s) be a stable. ^ such that (G(s) .4. For example. the multiplicity of 1 (G . G). rational.68) follows from Lemma 8. Corollary 8.69) implies that the inequalities in (8. Optimal Hankel Norm Approximation 187 where (8. G 2 The following corollary gives the solution to the wellknown Nehari's problem.4. k+1 I )) = (n . Let Ga = G be an augmented square 0 ^ G be the optimal Hankel norm approximation of ^ transfer function and let Ga = ^ G2 Ga such that ^ Ga . i.26). transfer function G(s).4 ^ can be applied to G(s) . ^ (s) . Further. Now Theorem 8. 1 ( 2 .s)) = i (G(s)) i = r1 + 1 : : : n.e.s) is used since it will be stable and have wellde ned Hankel singular values. m m.4. ^ To show the necessity of part (2) suppose that G(s) is an optimal Hankel norm approximation to G(s) of McMillan degree k.
G(s) .n2 det( I .9 Let E (s) = G(s) . Firstly let the realization be transformed to. F (s) be as de ned in part (3) of Theorem 8. P12 Q21 )) = det(( . Q22 P22 ): The result now follows on observing that i (G(s)) = i (P11 Q11 ) and i2 (F (. 2 )I + Q21 P12 ) = ( .s)) = i+2k+r (G . G) i+k+r (G) Proof. det( I . The equations of Theorem 8. 1 i = 1 2 : : : n1 .3 C1 D C2 0 (i)(iv) are then satis ed by a transformed P and Q. 2 )n1 . r. F = A2 B2 .6 ^ with G(s) G(s) 2 RH1 and F (s) 2 RH.4 L1 Bounds for Hankel Norm Approximation The natural question that arise now is. ( 2 I . . P11 Q11 ) = det( I . then. n2 + 1 : : : n1 i. P = P11 P12 Q = Q11 Q21 P12 P22 Q21 Q22 PQ = 2 I implies that.s)) A1 0 B1 A B E = 4 0 A2 B2 5 = C D Re i (A1 ) < 0 Re i (A2 ) > 0 C1 C2 D in which case G = A1 B1 .8 Let an m m transfer matrix E = C D satisfy E (s)E (s) = 2 I and all equations of Theorem 8. 2 )n1 . ^ i+3k+r (G) i (F (.n1 +n2 (F (. 1 ^ i (G . If E = Gs + F with Gs 2 RH1 and F 2 RH.s)) = 2 i (Q22 P22 ): ^ Corollary 8. k . G) = k+1 (G) and for i = 1 2 ::: n . 2 )I + P12 Q21 ) = ( .3 and let A have dimension n1 + n2 with n1 eigenvalues strictly in the left half plane and n2 < n1 eigenvalues strictly in the right half plane. partitioned as.n2 det(( . does the Hankel norm being small imply that any other more familiar norms are also small? We shall have a de nite answer in this section. 2 3 . A B Lemma 8. n2 i (Gs ) = i = n1 . Then for i = 1 2 ::: 2k + r.188 HANKEL NORM APPROXIMATION 8.
part (3) (with F (s) 0) then ^ (1) (G(s) .4. The i (G(s)) are obtained from Lemma 8.8.6. transfer function with Hankel G(s) = D0 + 1 E1 (s) + 2 E2 (s) + : : : + N EN (s) where (1) Ek (s) are allpass and stable for all k. r in the open right half plane.11 Let G(s) be a stable.8. transfer function of McMillan ^ degree n and such that n (G) has multiplicity r. 2 The following lemma gives the properties of certain optimal Hankel norm approximations when the degree is reduced by the multiplicity of n .6 gives that A 2 R(n. G(s))= n (G(s)) is allpass. (2) For k = 1 2 : : : N singular values 1 > 2 : : : > N where i has multiplicity ri and r1 + r2 + : : : + rN = n. The construction of E (s) ensures that the allpass equations are satis ed and an inertia argument easily establishes that the Amatrix has precisely n1 = n + k eigenvalues in the open left half plane and n2 = n . L1 Bounds for Hankel Norm Approximation 189 Proof.r) is stable and hence F (s) can be ^ ^ chosen to be zero and therefore (G(s) . Also let G(s) be an optimal Hankel norm approximation of degree n . Hence Lemma 8. r. Lemma 8. G(s))= n (G) is allpass.5. rational m m. ^ (2) i (G(s)) = i (G(s)) i = 1 : : : n . ^ Proof.8 can be applied to give the equalities. r given by Theorem 8. In this case some precise statements on the error and the approximating system can be made.r) (n. Then there exists a representation of G(s) as (8:71) ^ Gk (s) := D0 + has McMillan degree r1 + r2 + : : : + rk . The inequalities follow from Lemma 8. Theorem 8. k .10 Let G(s) be a stable. rational m m. k X i=1 i Ei (s) . 2 ^ ^ Applying the above reduction procedure again on G(s) and repeating until G(s) has zero McMillan degree gives the following new representation of stable systems. Theorem 8.
2 Theorem 8. ^ 0 (s) which will be a constant and combining the steps gives the Finally taking D0 = G result.1 (s) = k Ek (s) for some stable. but would involve less computation ^ since the decomposition into G(s) = F (s) need not be done. Corollary 8. rational p m. m m. G(s) k+1 + k+2 + : : : + N .13 Given a stable. For p = m consider the representation of G(s) given by Theorem 8.10 (1) gives that Gk (s) . part G ^ (3d). since G(1) = 0.11. gives kD0 k ) kG(s)k1 1 + 2 +:::+ N 2( 1 + 2 + : : : + N ): For the case p 6= m just augment G(s) by zero rows or columns to make it square. This will not be 1 an optimal Hankel norm approximation in general. and such that G(1) = 0. allpass Ek (s). 2 Note that the construction of Theorem 8.10 (2) applied ^ at each step then gives that the Hankel singular values of Gk (s) will be 1 2 : : : k ^ with multiplicities r1 r2 : : : rk . transfer function G(s) with Hankel singular values 1 2 ::: k > k+1 = k+2 : : : = k+r > k+r+1 ::: n >0 . then the above argument gives upper bounds on the L1 norm of this augmented system. but will have the same L1 norm. where each i has multiplicity ri . Note also that Theorem 8.10) of degree r1 + r2 + : : : + rk . Further setting s = 1. ^ k. with GN (s) := G(s).190 HANKEL NORM APPROXIMATION ^ ^ Proof. Let Gk (s) be the optimal Hankel norm approximation to Gk+1 (s) (given by ^ Lemma 8.4. D0 k1 1 + 2 + : : : + N: Proof.12 Let G(s) be a stable. respectively. rational.11 then kG(s) . and at each step a balanced ^ ^ ^ ^ realization of Gk (s) is given by (Ak Bk Ck ) with a diagonal scaling. D0 k1 = k 1 E1 (s) + 2 E2 (s) + : : : + N EN (s)k1 1 + 2 +::: + N since Ek (s) are allpass.11 immediately gives an approximation ^ algorithm that will satisfy G(s) . transfer function with Hankel singular values 1 > 2 > : : : > N . Then (1) kG(s)k1 2( 1 + 2 + : : : + N ) (2) there exists a constant D0 such that kG(s) . relation (i) also ensures that each Gk (s) will have McMillan degree r1 + r2 + : : : + rk . Lemma 8. An upperbound on the L1 norm of G(s) is now obtained as an immediate consequence of Theorem 8. Hence Lemma 8.
6. Let Ga = G 0 ^ ^ be an augmented square transfer function and let Ga = G be the optimal Hankel ^ G2 norm approximation of Ga such that k+1 (Ga ) ^ Ga . k .12 can be applied to F (. n. D0 k1 Now Corollary 8. part (3) such that F (s) 2 RH.9 and Corollary 8. Hence G ^ the constant term of G can not be determined from Hankel norm. For example.r X i=1 i (F (. consider a stable.k. Da 1 k+1 (Ga ) + n.6. Ga . D0 can of course be obtained using any standard convex optimization algorithm: ^ D0 = arg min G .s)) i+k+r (G) 1 Note that Theorem 8.13 can be obtained in many ways.13 can also be applied to nonsquare systems. D0 k1 = k+1 (G) + kF .s) to obtain a D0 such that 1 kF . D0 k1 since by Corollary 8. We shall mention three of them: ^ Apply Corollary 8.r X i=1 i+k+r (G): for i = 1 2 ::: n .9.k.12. This is usually complicated since (G(s) . 2 ^ It should be noted that if G is an optimal Hankel norm approximation of G then ^ + D for any constant matrix D is also an optimal Hankel norm approximation. r then 1 ^ ^ kG . G . D0 : D0 i (F (.r X i=1 i+k+r (Ga ) .r X i=1 i+k+r (G): Proof. F + kF . G . has McMillan degree n . D0 k1 k+1 (G) + n. G . then there exists a D0 such that ^ k+1 (G) kG . ^ G(s)) in general has McMillan degree of n + k. An alternative is to use the unstable part of the optimal Hankel norm approxi^ mation in Theorem 8. Let G + F be obtained from Theorem 8. r.6. k . An appropriate constant term D0 in Theorem 8. L1 Bounds for Hankel Norm Approximation 191 ^ and let G(s) 2 RH1 of McMillan degree k be an optimal Hankel norm approximation to G(s) obtained in Theorem 8.s)) n.8.12 to G(s) . p m (p < m). transfer function G(s).4. G(s). D0 k1 G . G .k. The theorem follows from Corollary 8. rational.k.
G . rational. rational. D0 1 k+1 (G) + n.r X i=1 i+k+r (G) ^ ^ since i (G) = i (Ga ) and G . then k+1 (G) ^ kG .27) to (8. m m transfer function G(s).11 can be used to bound the error obtained by truncating a balanced realization. Lemma 8. Da 1 . Gk1 2( k+1 (G) + n.192 with Da = D0 . The result follows from Theorem 8. rN ) and an optimal Hankel norm approximation of the same degree.13. Corollary 8. Ga .5 Bounds for Balanced Truncation Very similar techniques to those of Theorem 8.k. 1 A less tighter error bound can be obtained without computing the appropriate D0 . strictly proper transfer function G(s). We will rst need a lemma that gives a perhaps surprising relationship between a truncated balanced realization of degree (n .15 Let (A B C ) be a balanced realization of the stable. 2 8. with Hankel singular values 1 2 ::: k > k+1 = k+2 : : : = k+r > k+r+1 ^ : : : n > 0 and let G(s) 2 RH1 of McMillan degree k be a strictly proper optimal Hankel norm approximation to G(s) obtained in Theorem 8. and let A = A11 A12 A21 A22 B = B1 B2 = C = C1 C2 1 0 0 I ^ ^ ^ ^ Let (A B C D) be de ned by equations (8. D0 Ga .r X i=1 i+k+r (G)) Proof.6. Then D 2 HANKEL NORM APPROXIMATION k+1 (G) ^ G . G .32) (where 2 = 1) and de ne Gb := Gh := then A11 B1 C1 0 ^ ^ A B ^ ^ C D .14 Given a stable. m m.k.
Gh (s)kH = 2 (Note that Gh (s) is stable if 1 > I .5. (1) In order to prove that (Gb (s) .16 Let G(s) be a stable. Hence kG(s) . Gb (s)kH kG(s) . p m. Gk (s) 1 H 2( k+1 + k+2 + : : : + N ).C ~ ^ D = . Gh (s))= but the rst term on the right hand side is allpass by Theorem 8. Gh (s))= is allpass. Then ~ (1) G(s) . ..D: (8..54) to (8. ~ (2) G(s) .3. Gb (s)k1 2 .3 ~ such that (Gb (s) . Gb (s)kH 193 2 . (3) Similarly using the fact that allpass functions have unity Hankel norms gives that kG(s) . Gb (s))= = (G(s) .72) and (8. (3) If 1 > I then kG(s) .56) and Theorem 8.73) (8:74) (This is veri ed by noting that the blocks of equations (8.4. Gh (s) = ~ C ~ B = B1 ^ B ~ B ~ D # where Now Theorem 8. transfer function with Hankel ~ singular values 1 > 2 : : : > N . part (2) and the second term is allpass by part (1) above. " Proof.. (2) kG(s) . Gh (s))= + (Gb (s) .34) and (8. Gb (s)k1 2 . . Gh (s))= is allpass.) Hence P Q = 2 I and by Theorem 8.4. Gh (s)kH + kGb (s) . where each i has multiplicity ri and let Gk (s) be obtained by truncating the balanced realization of G(s) to the rst (r1 + r2 + : : : + rk ) states.8.D is an appropriate ~ ^ there exists D choice is veri ed from equations (8. 1 1 11 0 ~ A = A0 A ^ ~ ^ C = C1 .73) are also blocks ~~ of equations (8. Gk (s) 2( k+1 + k+2 + : : : + N ). (2) (G(s) .35) for Pe and Qe .) 2 Given the results of Lemma 8. rational.1 . That D = . Bounds for Balanced Truncation (1) (Gb (s) . part (1) gives that the solutions to Lyapunov equations ~~ ~ ~ ~ ~ AP + P A + B B = 0 ~ ~ ~~ ~ ~ A Q + QA + C C = 0 are I 1 ~ ~ P = I1 Q = .72) (8. Gh(s))= is allpass we note that ~ A Gb (s) .15 bounds on the error in a truncated balanced realization are easily proved as follows. Theorem 8. part (2). .
and the proof is complete. Gi (s)) 2( k+1 + k+2 + : : : + N ) for both norms. TG = P+ MG jH2 .H2 Analogous to the Hankel operator. there is also a corresponding time domain description for the Toeplitz operator: Tg : L2 0 1) 7. Hence assume p = m.. )f ( )d t 0 . respectively. 2 In this section. Hence (GN (s) := G(s)) ~ G(s) . we consider another operator.1 X i=k ~ ~ (Gi+1 (s) . Also G ~ ~ ~ the balanced realization of Gi+1 (s) and hence Gi+1 (s) . Then a Toeplitz operator associated with G is denoted by TG and is de ned as 8. respectively. Notice that since truncation of balanced realization are also balanced.e.! L2 0 1) Tg f := P+ (g f ) = for f (t) 2 L2 0 1). Z 1 0 g(t . Gk (s) = N . In particular if G(s) 2 H1 then TG = MGjH2 .194 HANKEL NORM APPROXIMATION Proof. If p 6= m then augmenting B or C by zero columns or rows. Again let G(s) 2 L1 . will still give a balanced realization and the same argument is valid. Gi (s) 2 i+1 for both L1 and Hankel norms. the Hankel singular values of Gi (s) will be 1 2 : : : i ~ i (s) can be obtained by truncating with multiplicities r1 r2 : : : ri .6 Toeplitz Operators TG : H2 7. satisfying the ~ truncated Lyapunov equations.! H2 TGf := (P+ MG)f = P+ (Gf ) for f 2 H2 i. MG > L2 P+ H2 TG ? .
This is yet another characterization of causality. 8.G = 0 and MG is a lower block triangular operator. a multiplication operator can be decomposed as several ? Hankel and Toeplitz operators: let L2 = H2 H2 and G 2 L1 . In fact.e. we will not discuss it 2 further here. to give the readers some intuitive ideas about these operators since they are very important in the H1 control theory. MGjH? .G : .e.. MG jH? is actually a Toeplitz operator however.7 Hankel and Toeplitz Operators on the Disk* It is sometimes more convenient and insightful to study operators on the disk. some operators are much easier to interpret and compute in discrete time. To start with. MGjH2 = P. From the control system point of view. i.G TG + G H2 + G H2 Note that P. hopefully. respectively: D Let L2 (@ D ) denote the Hilbert space of matrix valued functions de ned on the unit circle @ D as Z 2 L2 (@ D ) = F ( ) : 21 Trace F (ej )F (ej ) d < 1 0 with inner product @ D := f 2 C : j j = 1g : hF Gi := 21 Z 2 0 Trace F (ej )G(ej ) d : . the future input does not a ect the past output. A fact worth mentioning is that if G is causal. it can be shown that G is causal if and only if . The objective of this section is to examine the Hankel and Toeplitz operators in discrete time (i. it is necessary to introduce some function spaces in respect to a unit disk. Hankel and Toeplitz Operators on the Disk* 195 It is seen that the multiplication operator (in frequency domain) or the convolution operator (in time domain) plays an important role in the development of the Hankel and Toeplitz operators.7. MGjH? P. In fact.e.. on the disk) and. Then the multiplication operator associated with G can be written as ? ? MG : H2 H2 7.! H2 H2 P M j ? P M j 2 2 MG = P. Let D denote the unit disk : D := f 2 C : j j < 1g := f 2 C : j j 1g and let D and @ D denote its closure and boundary. G 2 H1 . i.8.. then .G = 0.
let H2 (@ D ) be the (closed) subspace of L2 (@ D ) with matrix functions F ( ) analytic in D .4 Let F ( ) 2 L2 (@ D ).3 It should be kept in mind that.1 .e.) Example 8. ~ Analogous to the space in the half plane.1 Fi i : . (However. It can be shown that the Fourier transform (or bilateral Z transform) gives the following isometric isomorphism: L2 (@ D ) = l2 (.196 HANKEL NORM APPROXIMATION Furthermore.1 1) H2 (@ D ) = l2 0 1) ? H2 (@ D ) = l2 (. L1 (@ D ) is used to denote the Banach space of essentially bounded matrix functions with norm 202 ] kF k1 = ess sup F (ej ) : The Hardy space H1 (@ D ) is the closed subspace of L1 (@ D ) with functions analytic in D and is de ned as H1 (@ D ) = F ( ) 2 L1 (@ D ) : The H1 norm is de ned as 2D Z 2 0 F (ej )ejn d = 0 for all n > 0 : kF k1 := sup F ( )] = ess sup F (ej ) : 202 ] It is easy to see that L1 (@ D ) L2 (@ D ) and H1 (@ D ) H2 (@ D ). H2 (@ D ) = F ( ) 2 L2 (@ D ) : 21 Z 2 0 F (ej )ejn d = 0 for all n > 0 ? and let H2 (@ D ) be the (closed) subspace of L2 (@ D ) with matrix functions F ( ) analytic in C nD .1 0): Remark 8. here denotes = z .. and let F ( ) have a power series representation as follows: F( ) = 1 X ? Then F ( ) 2 H2 (@ D ) if and only if Fi = 0 for all i < 0 and F ( ) 2 H2 (@ D ) if and only if Fi = 0 for all i 0. 3 i=. it should be pointed out that these inclusions are not true for functions in the half planes or for continuous time functions. i. in contrast to the variable z in the standard Z transform.
1 7 7 u. u( ) = P1 i i=.5 G0 G1 G2 G.1 j =.1 G.2 G.1 G0 G1 G.4 G. Then the system output in the frequency domain is given by y( ) = Gd ( )u( ) = 2 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 2 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4 1 X 1 X where yi is the system time response at time t = i.1 G2 G1 G0 G1 G2 G3 G2 G3 G3 G4 G4 G5 32 76 76 76 76 76 76 76 76 76 76 76 76 76 76 54 u0 u1 u2 u.75) can be rewritten as 2 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4 y0 y1 y2 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 2 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4 G0 G.2 G1 G0 G. Then the Hankel operator associated with Gd ( ) is de ned as ? .2 G.! H2 (@ D ) ? P.3 G.3 7 7 5 Equation (8.4 G.1 G.1 G. : l2 (.2 G.1 Gi uj i+j =: 1 X i=. we have 32 76 76 76 76 76 76 76 76 76 76 76 76 76 76 54 i=. and P+ denote the orthogonal projections: P+ : l2 (. and consequently.1 ui 2 L2 (@ D ) be the system input in the frequency domain.2 y.2 y. a Toeplitz operator associated with Gd( ) is de ned as TGd : l2 0 1) 7.1 1) 7.3 G3 G2 G1 G0 G. and ui be the input at time t = i.8.4 G.3 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 y.! H2 (@ D ): Suppose Gd ( ) 2 L1 (@ D ).1 0) 7.1 Gi i 2 L1 (@ D ) be a linear system transfer matrix.1 0) or L2 (@ D ) 7.Gd = P+ MGd jH? (@ D) : 2 Similarly.2 G.! l2 0 1) or L2 (@ D ) 7.3 G.! l2 (.2 G4 G3 G2 G1 G0 G.1 G0 .7.! H2 (@ D ) .! H2 (@ D ) TGd = P+ MGd jH2 (@ D) : P1 Now let Gd ( ) = i=.1 1) 7.Gd : l2 (.5 G1 G0 G.1 G.! l2 0 1) or H2 (@ D ) 7.1 yi i 3 7 7 y2 y1 y0 y.4 G2 G1 G0 G.2 u.3 G.1 y.2 G.2 7 7 u.1 G.3 G. Hankel and Toeplitz Operators on the Disk* 197 Now let P.1 G.3 = G.2 G.3 G.1 y.! l2 0 1) or H2 (@ D ) 7.1 G5 G4 G3 G2 G1 G0 u2 7 7 u1 7 7 u0 7 : (8:75) 7 u.1 u.3 = G0 G.
we have x0 = 1 Ai.Gd and T1 is the matrix representation of the Toeplitz operator TGd . A).3 Matrices like T1 and T2 are called (block) Toeplitz matrices and matrices like H1 and H2 are called (block) Hankel matrices. x 2 R x0 = cu = B AB A2 B 6 6 6 4 u.1 0).3 7 7 7 5 2 Rn : y0 6 y1 6 o x0 = 6 y2 4 . In fact.5 Let Gd( ) 2 RH1(@ D ) and Gd( ) = C ( Gd ( ) = D + space realization of Gd ( ) and let A have all the eigenvalues in D .1 B 8i 1. The state space equations are given by P Then for u 2 l2 (. Thus these operator norms can be computed from the matrix norms of their corresponding matrix representations. .i . n .Gd k = kH1 k. H1 is the matrix representation of the Hankel operator . the output can be computed as On the other hand. 2 3 C 7 6 CA 7 7 6 7 7 = 6 CA2 7 x0 2 l2 0 1) 5 4 5 .1 B + D be a state Example 8. Lemma 8..1 : u. . 2 3 .2 u.1 Bu.1 u. Then i+1 CAi B = X Gi i i=0 i=0 1 with G0 = D and Gi = CAi.17 kGd( )k1 = T1 H1 H2 T2 1 X . .2 u. k.198 2 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4 HANKEL NORM APPROXIMATION u0 u1 u2 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 =: T1 H1 H2 T2 u. which de nes the controllability i=1 operator 2 3 xk+1 = Axk + Buk x. and so on.1 = 0 yk = Cxk + Duk : ... given x0 and uk = 0 i 0.1 I . and kTGd k = kT1 k. .
Hence. De ne the function The relation between a point j! on the imaginary axis and the corresponding point ej on the unit circle is. Suppose G(s) 2 H1 (j R) and de ne (8:77) Gd ( ) := G(s)js= 1. . . . Hankel and Toeplitz Operators on the Disk* 2 3 199 which de nes the observability operator... and oy = 2 . : 3 It is interesting to establish some connections between a Hankel operator de ned on the unit disk and the one de ned on the half plane. Of course.1 0) 6 7 c 4 5 .1 s= 1+ s+ 1. 1 : j! + 1 (s) = s +21 : p . 2 3 7 7 7 5 B AB A2 B G1 G2 G3 6 G2 G3 G4 = 6 G3 G4 G5 6 4 .Gd( ) .G(s) = . D . First de ne the map as 1 (8:76) = s. i.7. from (8. which maps the closed righthalf plane Re(s) 0 onto the closed unit disk. 2 . : 1+ Since G(s) is analytic in Re(s) > 0. . ej = j! . . . ..18 . Gd ( ) is analytic in D . .8. Lemma 8.. the Hankel operator has the following matrix representation: C AC (A )2 C 6 6 6 4 y0 y1 y2 3 7 7 7 5 2 Rn : 3 7 7 7 5 C 6 CA H = 6 CA2 6 4 .. the adjoint operators of c and o can be computed easily as B 6 B A 7 x0 = 6 B (A )2 7 x0 2 l2 (. Gd ( ) 2 H1 (@ D ).76). Proof.. including the point at 1.e. ...
? ? fd ( ) 7. In this section.! H2 (j R) is an isomorphism. a Hankel operator is closely related to an analytic approximation problem.H2 (@ D ) H2 (@ D ) ? ? H2 (j R) .8 Nehari's Theorem* As we have mentioned at the beginning of this chapter. let H1 (@ D ) denote the subspace of L1 (@ D ) with functions analytic in j j > 1.Gk.! (s)f (s) : H2 (@ D ) 7. so that f (. 8.j (j!)j2 d! + which implies that the mapping where f (s) = fd( )j = s. The lemma now follows from the commutative diagram . !2 2 1 d! = . It should be emphasized that the bilinear transformation is not an isomorphism between H2 (@ D ) and H2 (j R). Some applications .1 ) 2 H1 (@ D ).! H2 (j R) ? is an isomorphism note that if fd( ) 2 H2 (@ D ). G( ) 2 H1 (@ D ) if and only if G( .! (s)f (s) : H2 (@ D ) 7.H2 (j R ) 2 The above isomorphism between H2 (@ D ) and H2 (j R) has also the implication that every discrete time H2 control problem can be converted to an equivalent continuous time H2 control problem. where the approximation is with respect to the L1 norm.1 s+1 fd( ) 7. Minimum Distance Problem In the following. matrix in H1 equals k. For convenience. So . we shall be concerned with approximating G by an anticausal transfer matrix.1) = 0.Gd ( ) ? . and hence (s)f (s) is analytic in Re(s) < 0. Similarly. i.200 This yields HANKEL NORM APPROXIMATION d = . we shall establish that the distance in L1 from G to the nearest . . then fd (1) = 0. This is the classical Nehari Theorem.G(s) ? . one analytic in Re(s) < 0 (or j j > 1)..e.
we have . Qk1 = k. Q)f k2 sup kf k2 ? (@ D ) f 2H2 kP+ Gf k2 = sup ? (@ D ) kf k2 f 2H2 = k.19 Suppose G 2 L1.Gk and the in mum is achieved. we will only give the proof on the unit disk.18. ~ Proof. it is easy to establish that the . Hankel norm is the lower bound: for xed Q 2 H1 (@ D ).8. First. We shall give the proof on the unit disk.76) and Lemma 8.G k : . Qk1 = k. Next we shall construct a Q( ) 2 H1 (@ D ) such that kG . Q( ))f ( ) : H2 (@ D ) 7.1 0 X i Gi i Qi : i=. Hence.8. Qk1 = sup k(G kf Q)f k2 k2 ? (@ D ) f 2H2 kP+ (G .1 The lefthand side of (8. Nehari's Theorem* 201 and explicit construction of the approximation for the matrix case will be considered in the later chapters. Theorem 8.4 Note that from the mapping (8. kG .G k : Let the power series expansion of G( ) and Q( ) be (8:78) G( ) = Q( ) = 1 X i=. then Q2H.78) equals the norm of the operator ? f 7. 1 inf kG .! (G( ) . the above theorem is the same whether the problem is on the unit disk or on the half plane.! L2 (@ D ): . Remark 8.
Q. First choose Q0 to minimize 2 6 6 6 6 6 6 4 3 7 7 7 7 7 7 5 G0 .2 G. Continuing in this way gives a suitable Q. Q0 G. Q. in turn to minimize the matrix norm of (8.1 ..e.1 G3 G2 G1 G5 G4 G3 G2 G0 .Gk.1 to minimize 2 6 6 6 6 6 6 6 6 4 3 7 7 7 7 7 7 7 7 5 Again. Q0 G. Q. Q. there is a matrix representation of this operator: G0 . : : : . .78) is to select Q0 . the minimum equals the Hankel norm k.1 .G k.202 2 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4 HANKEL NORM APPROXIMATION 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 From the discussion of the last section. matrix dilation theory presented in Chapter 2). Then we have the following corollary. Q0 2 6 6 6 6 4 G3 G2 G1 G4 G3 G2 G1 G5 G4 G3 G2 3 7 7 7 7 5 : By Parrott's Theorem (i.1 .1 G0 .79). Q0 G.2 . G0 . Q. Next. Q0 G1 G4 G3 G2 G1 : 2 As we have mentioned earlier we might be interested in approximating an L1 function by an H1 function. choose Q. the minimum equals the norm of the following matrix G3 G4 G5 G2 G3 G4 G1 G2 G3 which is the rotated Hankel matrix H1 .1 G0 .1 . Q0 G3 G2 G1 G4 G3 G2 G1 G5 G4 G3 G2 G1 : (8:79) The idea in the construction of a Q to satisfy (8. Hence. the minimum equals k.
Qk2 kU k2 . 1 kU k2 1 2 2 . hH .g u = 1 v . Qk1 = .GU 2 RH2 and P+H = P+(GU ) = . Q)U = 1 V: Moreover. hU . Proof. 1 hU U i 2 2 kH k2 . Then inf. Q is allpass.8.20 Suppose G 2 L1. Q)U and note that .21 Let G 2 RH1 and 1 be the largest Hankel singular value of G. if G is a scalar function. U . Qk1 = inf .g v = 1 u: Now denote the Laplace transform (or Z transform) of u and v as U 2 RH? and 2 V 2 RH2 . This follows from the fact that ^ inf kG . Let H := (G .GU k2 2 2 + k. U i . Lemma 8. Q k1 = k. h. then ^ inf kG .G . .GU . hP H . h. U k2 . Nehari's Theorem* Corollary 8.1 0] (or l2 (. U P H i kH k2 G 2 + G G + 2 . U k2 . kG . 1 kU k2 2 2 2 kG .8. Then 0 = = = = = = = 0: kH .G k = . 1 V=U is the unique solution and G .G k = 1 .G Q2H1 and the in mum is achieved. U i . h.. kG .e.G : Q2H1 Q 2H1 2 Let G 2 RH1 and g be the impulse response of G and let 1 be the largest Hankel singular value of G.GU i 2 2 kH k2 . 203 Proof. then Q = G .1 0)) and v 2 L2 0 1) (or l2 0 1)) are the corresponding Schmidt pairs: . Suppose u 2 L2 (. U i kH k2 G G kH k2 . k. U H i kH k2 G 2 G G 2 + k. i. Qk1 = 1 and Q2H1 (G .
1 V=U . Now if G is a scalar function. i.1 I ..A 2 L2 (. (G . V=U is an allpass. H = . be the largest eigenvalue and the corresponding eigenvector of Lc Lo: ALc + LcA + BB = 0 A Lo + Lo A + C C = 0: Lc Lo = 1 2 1 : := 1 Lo : u = v = are the Schmidt pair and .B 0 2 RH2 A C 0 2 RH2 : It is easy to show that if G is a scalar. We shall prove through explicit construction below that V=U is an allpass. let G be a discrete time transfer matrix and let A G( ) = C ( . Q)U = .GU = 1 V . A). Hence.G U . then Q is uniquely determined and Q = G . The details are left as an exercise for the reader. 2 Formulas for Continuous Time Let G(s) 2 RH1 and let Let Lc and Lo be the corresponding controllability and observability Gramians: A G(s) = C B : 0 And let De ne Then 2 1. then U V are scalar transfer functions and U U = V V .204 HANKEL NORM APPROXIMATION Hence.e.1 0] c =B e At 2 L2 0 1) o = Ce U (s) = V (s) = . Formulas for Discrete Time Similarly.1 B = C B : 0 .A ? .
1 B if a state space realization of G is available) and let G1 G2 G3 6 G2 G3 G4 H = 6 G3 G4 G5 6 4 . 2 l2 (. : .8. Nehari's Theorem* Let Lc and Lo be the corresponding controllability and observability Gramians: ALcA .1 0) 3 7 7 7 5 v = are the Schmidt pair and 2 l2 0 1) where W ( ) := W T ( . . . Gi = CAi.. . 1 X 2 2 3 7 7 7 5 . . Alternatively..8. since the Hankel operator has a matrix representation H . . . Lc + BB = 0 A Lo A .i = A B 2 RH? 2 0 i=1 1 X A V( ) = CAi i = CA C 2 RH2 i=0 U( ) = G( ) = 2 1 X i=1 Gi i Gi 2 Rp q 3 7 7 7 5 (in fact.. Lo + C C = 0: 2 .. be the largest eigenvalue and a corresponding eigenvector of Lc Lo : And let 1 2 Lc Lo = 1 : De ne := 1 Lo : Then 1 205 u = B 6 B A 6 c = 6 B (A )2 4 C 6 CA 6 o = 6 CA2 4 . .1 . u and v can be obtained from a \singular value decomposition" (possibly in nite size): let B (A )i. . ..1 )..
206 HANKEL NORM APPROXIMATION H = U V Suppose H has a \singular value decomposition" = diagf 1 U = u1 u2 V = v1 v2 2 3 7 5 2 g with U U = UU = I and V V = V V = I .. Gn 0 0 0 Hence. 2 3 7 5 where u and v are partitioned such that u1i 2 Rp and v1i 2 Rq . and Krein 1978]..1 2 RH2 : In particular. v11 6 v12 v = v1 = 4 . U ( ) and V ( ) can be obtained as U( ) = V( ) = 1 X 1 X i=1 i=1 u1i .1989]. Kung and Lin 1981]. The presentation of this chapter is based on Francis 1987] and Glover 1984. u and v can be obtained from a standard SVD of Hn . Finally. then matrix H has only a nite number of nonzero elements and H = Hn 0 0 0 with 2 6 6 =6 6 6 4 0 0 : . . Power 1982]. and Safonov 1980].. Then u11 6 u12 u = u1 = 4 . Hn G1 G2 G2 G3 G3 G4 . . .. if G( ) is an nth order matrix polynomial. . Arov. Silverman. . Bettayeb. . . . .i 2 RH? 2 v1i i. Gn.9 Notes and References The study of the Hankel operators and of the optimal Hankel norm approximation theory can be found in Adamjan... Francis 1987]. Kavranoglu and Bettayeb 1994] and references therein.1 Gn Gn 0 3 7 7 7 7 7 5 8.
we need. we show that applying the SISO analysis/design method to an MIMO system may lead to erroneous results. We show by examples that the classical gain margin and phase margin are insu cient indicators for the system robustness. The relationship between models and the reality they represent is subtle and complex. the 207 . We obtain robust stability tests for systems under various model uncertainty assumptions through the use of the small gain theorem. A simple example is also used to indicate the fundamental di erence between the robustness of an SISO system and that of an MIMO system. And we also obtain some su cient conditions for robust performance under unstructured uncertainties. a set of maps. Since no single xed model can respond exactly like the true plant. However.1 Model Uncertainty Most control designs are based on the use of a design model. The quality of a model depends on how closely its responses match those of the true plant. at the very least. In particular. A mathematical model provides a map from inputs to responses.Model Uncertainty and Robustness In this chapter we brie y describe various types of uncertainties which can arise in physical systems. and we single out \unstructured uncertainties" as generic errors which are associated with all design models. The di culty associated with MIMO robust performance design and the role of plant condition numbers for systems with skewed performance and uncertainty speci cations are revealed. 9 9.
For example.g. but some aspects of physical behavior are approximated quite well using this model. a set membership statement for the parameters of an otherwise known FDLTI model is a highlystructured representation of uncertainty. and whatever mechanism is used to express these errors will be called a representation of uncertainty. Therefore. high frequency). an additive noise model is entirely inappropriate for capturing uncertainty arising from variations in the material properties of physical plants. the deviation's spectrum of the true output from the nominal will depend signi cantly on the input. Generally. a design technique must help make this leap small by accounting for the inevitable inadequacy of models. A good model should be simple enough to facilitate design. e. we are forced to use not just a single parameterized model but model sets that allow for plant dynamics which are not explicitly represented in the model structure. Wiener and Kalman ltering. However. In each case. In general. It typically arises from the use of linear incremental models at various operating points. coal composition. for certain classes of signals (e. To be practical. white noise propagation in linear systems.g. and equation coe cients in power plant control vary with aging. Representations of uncertainty vary primarily in terms of the amount of structure they contain. aerodynamic coe cients in ight control vary with ight environment and aircraft con gurations. This is equivalent to assuming that the uncertainty is generated by an additive noise signal with a bounded spectrum the uncertainty is represented as additive noise. For example. Of course. no physical system is linear with additive noise. etc.208 MODEL UNCERTAINTY AND ROBUSTNESS modeling problem is much deeper { the universe of mathematical models from which a model set is chosen is distinct from the universe of physical systems. For example. In the simplest case. the amounts of variation and any known relationships between parameters can be expressed by con ning the parameters to appropriately de ned subsets of parameter space. This re ects both our knowledge of the physical mechanisms which cause di erences between the model and the plant and our ability to represent these mechanisms in a way that facilitates convenient manipulation.. yet complex enough to give the engineer con dence that designs based on the model will work on the true plant. The term uncertainty refers to the di erences or errors between models and reality. consider the problem of bounding the magnitude of the e ect of some uncertainty on the output of a nominally xed linear system. LQG) than to the great practical signi cance o ered. the parameterized FDLTI model fails to describe the plant because the plant will always have dynamics which are not represented in the xed order model. a model set which includes the true physical plant can never be constructed. The actual construction of model sets for more general uncertainty can be quite di cult. this spectrum is assumed to be independent of the input. slag buildup.. A useful measure of uncertainty in this context is to provide a bound on the spectrum of the output's deviation from its nominal response.g. This type of uncertainty received a great deal of attention in the literature during the 1960's and 1970's.. and the attention is probably due more to the elegant theoretical solutions that are yielded (e. It is necessary for the engineer to make a leap of faith regarding the applicability of a particular design based on a mathematical model. .
The uncertainty may be caused by parameter changes.. In particular.1: Nyquist Diagram of an Uncertain Model The best choice of uncertainty representation for a speci c FDLTI model depends. P(j ω ) w(jω ) Figure 9. The statement does not imply a mechanism or structure which gives rise to . For instance. then P describes a disk centered at P with radius w(j!) at each frequency as shown in Figure 9.1) is the socalled multiplicative form: P (s) = (I + W1 (s) (s)W2 (s))P (s): (9:2) This statement con nes P to a normalized neighborhood of the nominal model P .2) compensated transfer functions have the same uncertainty representation as the raw model (i. con nes the matrix P to a neighborhood of the nominal model P . phase is completely unknown. This gives a less structured representation of uncertainty. To use such sets to describe physical systems. An alternative statement to (9. the bounds must roughly grow with frequency.9. Model Uncertainty 209 A simple example of this involves using frequencydomain bounds on transfer functions to describe a model set.1. if W1 = I and W2 = w(s)I where w(s) is a scalar function.e.. In practice. In particular.e.1) is that in (9.2) over (9. An advantage of (9. as mentioned above or by neglected dynamics or by a host of other unspeci ed e ects. at su ciently high frequencies. on the errors the model makes. i. Some other alternative set membership statements will be discussed later.1. it is generally possible to represent . This is a consequence of dynamic properties which inevitably occur in physical systems. of course. Examples of lessstructured representations of uncertainty are direct set membership statements for the transfer function matrix of the model. the statement (j!)] < 1 8! 0 (9:1) P (s) = P (s) + W1 (s) (s)W2 (s) where W1 and W2 are stable transfer matrices that characterize the spatial and frequency structure of the uncertainty. the weighting functions apply to PK as well as P ). 180o uncertainties.
This connection is developed in Safonov. These are caused by such e ects as in nitedimensional electromechanical resonance.2) have become widely used \generic" uncertainty representations for FDLTI models.2. A nonlinear element. may be quite accurately . These are usually the low frequency error components. time delays. di usion processes.1) or (9. 1966] and will not be pursued here. the weighting functions in (9. Fortunately.2) remains a strictly proper FDLTI system for all . nominal model log ω actual model Figure 9.g. We will assume that P in (9. e. More general perturbations (e. Motivated from these observations.210 MODEL UNCERTAINTY AND ROBUSTNESS some of these errors in a highlystructured parameterized form. are well suited to represent this latter class of errors. 1980] and Zames. we will focus for the moment on the multiplicative description of uncertainty.2: Typical Behavior of Multiplicative Uncertainty: p (s) = 1 + w(s) (s)]p(s) Also note that the representation of uncertainty in (9.g.1) and (9.2). in nite dimensional. When used to represent the various high frequency mechanisms mentioned above. etc. Consequently. An important point is that the construction of the weighting matrices W1 and W2 for multivariable systems is not trivial. There are always remaining higher frequency errors. The growth with frequency inevitably occurs because phase uncertainties eventually exceed 180 degrees and magnitude deviations eventually exceed the nominal transfer function magnitudes. Readers who are skeptical about this reality are encouraged to try a few experiments with physical devices. They are small ( 1) at low frequencies and increase to unity and above at higher frequencies.. time varying. the lessstructured representations.2) can be used to include perturbation e ects that are in fact certain..2) commonly have the properties illustrated in Figure 9. (9. (9. which cannot be covered this way. however. nonlinear) can also be covered by this set provided they are given appropriate \conic sector" interpretations via Parseval's theorem.
Robust Stability (RS): if K internally stabilizes every plant belong to . Nominal Performance (NP): if the performance objectives are satis ed for the nominal plant P . Robust Performance (RP): if the performance objectives are satis ed for every plant belong to . we may deliberately choose to ignore various known dynamic characteristics in order to achieve a simple nominal design model. suppose P 2 is the nominal design model and K is the resulting controller. it is treated as a conic linearity1 . . Then the closedloop feedback system is said to have Nominal Stability (NS): if K internally stabilizes the nominal model P . This is the model reduction process discussed in the previous chapters. Small Gain Theorem 211 modeled. but because our design techniques cannot e ectively deal with the nonlinearity. The conditions for which the robust stability and robust performance are satis ed under various assumptions on the uncertainty set will be considered in the following sections. Consider the interconnected system shown in Figure 9.3 with M (s) a stable p q transfer matrix. The basis for the robust stability criteria derived in the sequel is the socalled small gain theorem. 9. De nition 9. Safonov 1980] and Zames 1966]. for example. The nominal stability and performance can be easily checked using various standard techniques.9.2.2 Small Gain Theorem This section and the next section consider the stability test of a nominally stable system under unstructured perturbations. The following terminologies are used in this book.e e1 + 6 + M + e2 e + w2 ? Figure 9.3: Small Gain Theorem 1 See. w1 . As another example.1 Given the description of an uncertainty model set and a set of performance objectives.
Suppose !0 2 R+ f1g is such that (M (j!0 )) 1.6) the closedloop system is stable if det(I . We shall only prove part (a).7 (or Corollary 5. Suppose kM k1 1.kM (s)k1 > 0: (Necessity) This will be shown by contradiction. Then the interconnected system shown in Figure 9. u1 U V v1 = 1 = 0 and thus the closedloop system is either not wellposed (if !0 = 1) or unstable (if ! 2 R). Let M (j!0 ) = U (j!) (j!0 )V (j!0 ) be a singular value decomposition with up U (j!0 ) = u1 u2 vq V (j!0 ) = v1 v2 2 (j!0 ) = 6 4 1 3 7 5 2 and kM k1 = (M (j!0)) = 1 . . the closedloop system is stable if inf (I . M (j!0 ) (j!0 )) = det(I . so the system is unstable.1 (Small Gain Theorem) Suppose M 2 RH1 . M (s) (s)) 6= 0 s2C + Proof. Without loss for all 2 RH1 and k k1 1. M (s) (s)) has a zero on the imaginary axis. det(I .3 is wellposed and internally stable for all (s) 2 RH1 with (a) k k1 1= if and only if kM (s)k1 < (b) k k1 < 1= if and only if kM (s)k1 .212 MODEL UNCERTAINTY AND ROBUSTNESS Theorem 9. M ) has no zero in the closed righthalf plane for all 2 RH1 and k k1 1. M (s) (s)) 1. Indeed. We will show that there exists a 2 RH1 with k k1 1 such that det(I . There are two di erent cases: (1) !0 = 0 or 1: then U and V are real matrices. But this follows from inf (I . Thus by Theorem 5. To obtain a contradiction. Equivalently. In this case. of generality.kM (s) (s)k1 s2C + s2C + 1.. U V v1 u1 = 1 ) = 1 . (Su ciency) It is clear that M (s) (s) is stable since both M (s) and (s) are stable. The proof for part (b) is similar. for such (s). it now su ces to construct a 2 RH1 such that (j!0 ) = 11 v1 u1 and k k1 1. sup (M (s) (s)) = 1.. (s) can be chosen as = 1 v u 2 Rq p : 1 1 1 . assume = 1.
(This is stronger than what is given in the proof of Theorem 9.1 It can be shown that the small gain condition is su cient to guarantee internal stability even if is a nonlinear and time varying \stable" operator with an appropriately de ned stability notion. Remark 9.3 Suppose M 2 RH1 and kM k1 > . h 1 1 u11 1 . Choose i 0 and j 0 so that u1p ej p v1 = 6 6 4 v11 ej v12 ej v1q ej . 3 7 7 7 5 2 6 where u1i 2 R and v1j 2 R are chosen so that i j 2 .s 2 RH1 : p+ 1 and (j!0 ) = 11 v1 u1 .s 1 +s v1q q .. . Small Gain Theorem (2) 0 < !0 < 1: write u1 and v1 in the following form: 2 6 213 u1 = 6 6 4 u11 ej 1 u12 ej 2 . 0) for all i j . ~ The following lemma shows that if kM k1 > . ..1.2.) Lemma 9.9. 2 The theorem still holds even if and M are in nite dimensional.s q +s .2 The following statements are equivalent: (i) The system is wellposed and internally stable for all 2 H1 with k k1 < 1= (ii) The system is wellposed and internally stable for all 2 RH1 with k k1 < 1= (iii) The system is wellposed and internally stable for all 2 C q p with k k < 1= (iv) kM k1 . Let (s) = 1 Then k k1 = 1= i . q 3 1 2 7 7 7 5 \ 2 6 4 for i = 1 2 : : : p and j = 1 2 : : : q. Corollary 9.. This is summarized as the following corollary. . M (s) (s)) has a zero on the axis Re(s) = . there exists a destabilizing with k k1 < 1= such that the closedloop system has poles in the open right half plane. . j!0 = j j + j!0 v11 1 . j!0 = i i + j!0 3 7 5 \ j .s 1 +s s i u1p p . see Desoer and Vidyasagar 1975]. Then there exists a 0 > 0 such that for any given 2 0 0 ] there exists a 2 RH1 with k k1 < 1= such that det(I .
3 7 7 7 5 2 u1p ej p v1 = 6 6 4 6 v11 ej v12 ej v1q ej .s q+ . assume = 1.s v11 1 +s 3 7 5 h \ j . . j!0 = j j + + j!0 (s) = 1 1 6 4 Then k k1 = 1= 1 < 1 and det (I .. Since M 2 RH1 and kM k1 > 1. .. The modeling error will again . 0) for all i j . M (s) (s)). j!0 = i i + + j!0 1 .3 Stability under Stable Unstructured Uncertainties The small gain theorem in the last section will be used here to derive robust stability tests under various assumptions of model uncertainties. Re(s)>0 Now let 2 0 0 ]. . Then there exists a 0 < !0 < 1 such that kM ( + j!0)k > 1..214 MODEL UNCERTAINTY AND ROBUSTNESS Proof. Choose i 0 and j 0 so that \ for i = 1 2 : : : p and j = 1 2 : : : q. q where u1i 2 R and v1j 2 R are chosen so that i j 2 . there is a su ciently small 0 > 0 such that kM ( 0 + s)k1 := sup kM ( 0 + s)k > 1. s v1q q . . 2 The above lemma plays a key role in the necessity proofs of many robust stability tests in the sequel.s u11 1 +s 1 i u1p p. M ( + j!0 ) ( + j!0 )) = 0. Let M ( + j!0 ) = U V be a singular value decomposition with up U = u1 u2 V = v1 v2 2 vq 2 =6 4 Write u1 and v1 in the following form: 2 6 1 3 7 5 . Hence s = + j!0 is a zero for the transfer function det (I . .. Let 2 i .s 2 RH1 : p +s 9. Without loss of generality. . : 3 1 2 7 7 7 5 u1 = 6 6 4 u11 ej 1 u12 ej 2 ..
3.4: Unstructured Robust Stability Analysis We assume that the model uncertainty can be represented by an additive perturbation: = P + W1 W2 : Theorem 9. (Most of the robust stability tests discussed in the sequel can be easily generalized to unstable case with some mild assumptions on the number of unstable poles of the uncertain model.1 Additive Uncertainty Proof.4 Let = fP + W1 W2 : 2 RH1g and let K be a stabilizing controller for the nominal plant P . So and Si = (I + KP ).1 2 RH : 1 . We shall consider the standard setup shown in Figure 9.1 f 6 .1 = (I + KSo W1 W2 ). Then I K .9.1 2 RH1 since det(I + KSo W1 W2 ) = det(I + W1 W2 KSo) = (I + SoW1 W2 K ) = det(I + W2 KSoW1 ): .1 .1 Si . i. where is the set of uncertain plants with P 2 as the nominal plant and with K as the internally stabilizing controller for P .1 To = I .1 (I + K ).K (I + K ).K   Figure 9. The sensitivity and complementary sensitivity matrix functions are de ned as usual as So = (I + PK ). I (I + So W1 W2 K ) is wellposed and internally stable if (I + W2 KSo W1 ). 9. the uncertainty can be represented as W1 W2 .1 So (P + W1 W2 ) So (I + W1 W2 KSo).) In addition.4.1 .3.e. we assume that the modeling error is suitably scaled with weighting functions W1 and W2 .KSo(I + W1 W2 KSo ). Si : Recall that the closedloop system is wellposed and internally stable if and only if I K .1 Ti = I . Then the closedloop system is wellposed and internally stable for all k k1 < 1 if and only if kW2 KSoW1 k1 1.1 = (I + K ). Let = P + W1 W2. I (I + K ). Stability under Stable Unstructured Uncertainties 215 be assumed to be stable.. we encourage readers to ll in the details. .1 .
216 MODEL UNCERTAINTY AND ROBUSTNESS But (I + W2 KSo W1 ). To show the necessity.2 Multiplicative Uncertainty In this section. for all admissible .1 W1 = I . note that robust stability implies that K (I + K ).5 Let = f(I + W1 W2 )P : 2 RH1g and let K be a stabilizing controller for the nominal plant P . k k1 < 1 only if kW2 KSoW1 k1 1.e y We ? e e  Figure 9. By small gain theorem. This in turn implies that W2 K (I + K ).1 2 RH1 is guaranteed if k W2 KSoW1 k1 < 1.6 dm we .W2 r .1 2 RH1 2 RH1 with 2 Similarly. Then (i) the closedloop system is wellposed and internally stable for all 2 RH1 with k k1 < 1 if and only if kW2 ToW1 k1 1. Hence kW2 KSoW1 k1 1 is su cient for robust stability. (I + W2 KSoW1 ).K . Consider the feedback system shown in the Figure 9. .1 = KSo(I + W1 W2 KSo ).e.3.? W1  . we assume that the system model is described by the following set of multiplicative perturbation P = (I + W1 W2 )P with W1 W2 2 RH1 . (ii) the closedloop system is wellposed and internally stable for all 2 RH1 with k k1 1 if kW2 ToW1 k1 < 1.P ~ d ? Wd d ?.5: Output Multiplicative Perturbed Systems Theorem 9.1 2 RH1 for all admissible .5. z . this is true for all 9. it is easy to show that the closedloop system is stable for all 2 RH1 with k k1 1 if and only if kW2 KSoW1 k1 < 1.
This implies (I + K ). Proof. (ii). In fact.1 = s + 1 1 +1 : 1 + s+1 . We shall rst prove that the condition in (i) is necessary for robust stability. This implies (I + K ).1 are not cancelled by the zeros of So . for any given su ciently small > 0. we have also proven part (iv).9.1 has poles on the axis Re(s) = .3. It is easy to check that K s 1 To = s + 1 kTo k1 = 1 1 1 1 + s+1 s (I + K ). Then the closedloop system is wellposed and internally stable for all 2 RH1 with k k1 1 if and only if kW2 ToW1 k1 < 1. there is a 2 RH1 with k k1 1 such that (I + W2 To W1 ). 217 1 2 RH1 with k k1 (iv) the closedloop system is wellposed and internally stable for all k k1 1 only if kW2 To W1 k1 1.1 = (I + K ).1 = So (I + W1 W2 To).1 has poles on the imaginary axis. and W1 = W2 = 1. The su ciency parts of (i). We have and Example 9.1 has poles on the imaginary axis since the imaginary axis poles of (I + W1 W2 To). K (s) = 1. The proof of part (iii) is given below by exhibiting an example with kW2 ToW1 k1 = 1 but there is no destabilizing 2 RH1 with k k1 1. To show the necessity part of (v). which are the poles of P and K . Suppose kW2 ToW1 k1 > 1. assume that neither P nor K has poles on the imaginary axis. Hence kW2 To W1 k1 1 is necessary for robust stability.3.1 = s + 1 1 (I + K ).1 = K (I + K ).1 Let P (s) = 1 .1 has poles on the axis Re(s) = since can always be chosen so that the unstable poles are not cancelled by the zeros of So . Then by Lemma 9. 2 RH1 with (v) In addition. and (v) follow from the small gain theorem. suppose kW2 ToW1 k1 = 1. Hence kW2 ToW1 k1 < 1 is necessary for robust stability. From the proof of the small gain theorem. there is a 2 RH1 with k k1 < 1 such that (I + W2 ToW1 ).1 = So (I + W1 W2 To). 2 stabilizes P . Stability under Stable Unstructured Uncertainties (iii) the robust stability of the closedloop system for all does not necessarily imply kW2 To W1 k1 < 1.
1 + a1 s + s2 g(s) for some g(s) and a1 0 and s (I + K ).1 N ).1 (N + ~ N ) ~ ~ ~ ~ with M N ~ M . consider a left coprime factor perturbed plant described in Figure 9. ~ N 2 RH1 . Then (0) = . we can write (s) = . Re(s) 0. is analytic in a neighborhood of the origin since 2 RH1 . the internal stability of the closedloop system does not necessarily imply the boundedness of the arti cial signals z or w with respect to the arti cial disturbance dm . The only point where 1 + s+1 = 0 in the closed right half plane is s = 0. 2 RH1 and k k 1. Hence there is no destabilizing 2 RH1 with k k1 1. Indeed. We now claim that a1 0. The reason for the existence of a such gap may be attributed to the fact that in the multiplicative perturbed case. w. = (I + W1 W2 )P is a single system.) Thus the closedloop system is robustly stable although the transfer function from dm to z is unstable.6 Let As another example.6. along with (0) = .1 + 1 X i=1 an sn ai 2 R: ( = a1 < 0. there is really no cancelation in the physical system. and dm in Figure 9.218 MODEL UNCERTAINTY AND ROBUSTNESS 1 1 Since jTo (s)j = s+1 < 1 for all 0 6= s 2 C and Re(s) 0.3. The transfer matrices (M N ) are assumed to be a ~ . This cancelation is arti cial and is caused by the particular model representation (i. ~ ~ = (M + ~ M ). P = M .1 = K (I + K ).1. Hence. 1 + s+1 6= 0 for all 1 0 6= s 2 C .. Otherwise.1.1 = (I + K )..e.1 = s + 1 1+ 1 s+1 1 = 1 + a1 + sg 1 a sg 1 (I + K ). 9. This is the case for the above example where = (1 + )P = (a1 s + s2 g(s))=s = a1 + sg(s) and the pole s = 0 is cancelled. and K internally stabilizes ~ stable left coprime factorization of P (i. Hence (s) = .e. the signals z .5 are arti cial and they are not physical signals. d dss) s=0 implies k k1 > 1.3 Coprime Factor Uncertainty Theorem 9. 3 The gap between the necessity and the su ciency of the above robust stability conditions for the closed ball of uncertainty is only technique and will not be considered in the sequel. By assumption.1 = s + 1 1 +1 = 1 +1a+ + sg 1 + s+1 1 both of which are bounded in the neighborhood of the origin.
Hence (9. ~M w . Let K = UV .4.1 V (NU + MV ) 1 = K (I + PK ).? M .6: Left Coprime Factor Perturbed Systems the nominal system P . Stability under Stable Unstructured Uncertainties z1 ~N re K .1 ~ 1: I 1 the closedloop system is internally stable if and only if ~ ~ (N + ~ N )U + (M + ~ M )V .1 ~ 1: I 1 2 Similarly.1 M . one can show that the closedloop system is wellposed and internally stable for all k k1 1 if and only if K (I + PK ). the above is true for all k k1 < 1 if and only if U ~ ~ .1 be a right coprime factorization over RH1. All of the tests pertain to the standard setup shown in Figure 9. De ne := wellposed and internally stable for all k K (I + PK ). By Lemma 5.1 2 RH1 : By the small gain theorem.1 M .1 ~ N ~ M .e.3.1 summaries robust stability tests on the plant uncertainties under various assumptions. where is the set of uncertain plants with P 2 as the nominal plant and with K as the internally stabilizing controller of P . Then the closedloop system is k1 < 1 if and only if Proof.9.N ~ .3.1 (9:3) ~ ~ Since K stabilizes P . (NU + MV ).1 M . .1 < 1: ~ I 1 9.3) holds if and only if ~ ~ I + ( ~ N U + ~ M V )(NU + MV ).6 219 .1 2 RH1 .1 e ~ z2 y  Figure 9.10.4 Unstructured Robust Stability Tests Table 9. 2 RH1 : .
Example 9.1 : 2 RH1 k k1 1 3 Example 9. The representation P = P (I + W1 W2 ) in the second row covers similar types of errors at the inputs. s := P (1 + P ). Discussion of still other cases is left to the table. including di usion processes. the set of P can be covered by a set of feedback uncertain plants: poles is described by P with P = 1 . covering such things as unmodeled high frequency dynamics of sensors or plant. transport lags. etc. consider the following set of plants: P = (s s + 1 ++ 2) j j 2: + 1)(s . These representations are more useful for variations in modeled dynamics.1 2 P has one righthalf plane pole and P2 = s+1 2 P has no righthalf plane pole. the representation P = (I + W1 W2 )P in the rst row is useful for output errors at high frequencies (HF).2 Suppose an uncertain system with changing numbers of righthalf plane 1 P = s. : 2 R j j 1 : 1 1 Then P1 = s.220 MODEL UNCERTAINTY AND ROBUSTNESS Table 9.1 and (I + W1 W2 ). with aging. Nevertheless. Both cases should be contrasted with the third and the fourth rows which treat P (I + W1 W2 ). Note from the table that the stability requirements on do not limit our ability to represent variations in either the number or locations of rhp singularities as can be seen from some simple examples. electromechanical resonances. or across production copies of the same plant.1 should be interpreted as UNSTRUCTURED ANALYSIS THEOREM Given NS & Perturbed Model Sets Then ClosedLoop Robust Stability if and only if Robust Stability Tests The table also indicates representative types of physical uncertainties which can be usefully represented by cone bounded perturbations inserted at appropriate locations.3 As another example.1 P . For example. such as low frequency (LF) errors produced by parameter variations with operating conditions.
1 P + W1 W2 P (I + W1 W2 P ).1 P = NM . Stability under Stable Unstructured Uncertainties 221 W1 2 RH1 W2 2 RH1 Perturbed Model Sets (I + W1 W2 )P P (I + W1 W2 ) (I + W1 W2 ).1 (N + ~ N ) ~ ~ P = M .9.) .1 N = ~N ~M (N + N )(M + M ).1 ~ ~ (M + ~ M ).3.1 1 1 1 M .1 Si K I ] 1 Table 9.1 N = M LF parameter errors neglected HF dynamics uncertain rhp poles & zeros LF parameter errors neglected HF dynamics uncertain rhp poles & zeros ~ So M .1: Unstructured Robust Stability Tests (# stands for `the number'.1 P Representative Types of Uncertainty Characterized output (sensor) errors neglected HF dynamics changing # of rhp zeros input (actuators) errors neglected HF dynamics changing # of rhp zeros LF parameter errors changing # of rhp poles LF parameter errors changing # of rhp poles additive plant errors neglected HF dynamics uncertain rhp zeros LF parameter errors uncertain rhp poles 2 RH1 k k1 < 1 Robust Stability Tests kW2 ToW1 k1 1 kW2 Ti W1 k1 1 kW2 So W1 k1 1 kW2 Si W1 k1 1 kW2 KSoW1 k1 1 kW2 So PW1 k1 1 K I P (I + W1 W2 ).
4 Unstructured Robust Performance Consider the perturbed system shown in Figure 9. It is clear that the system with output multiplicative uncertainty as shown in Figure 9.P e . ~ ~ . Tzw . The performance speci cations are usually speci ed in terms of the magnitude of each component e in time domain. 9. The uncertain plant can be covered by a set of multiplicative perturbed plants: P 1 := s + 2 (1 + s 2 1 ) + 2 RH1 k k1 1 : 3 It should be noted that this covering can be quite conservative.K . For example.8 with the set of perturbed models described by a set .7: Equivalence Between Robust Stability With Output Multiplicative Uncertainty and Nominal Performance With Sensor Noise Rejection There is.222 MODEL UNCERTAINTY AND ROBUSTNESS This set of plants have changing numbers of righthalf plane zeros since the plant has no righthalf plane zero when = 0 and has one righthalf plane zero when = . Since Tzw = Ten . the set can be either parameterized set or unstructured set such as those described in Table 9. or alternatively and more conveniently some requirements on the closedloop frequency response of the transfer matrix between ~ d and e.6 e .e. hence kTzw k1 1 i supknk2 1 kek2 = kW2 ToW1 k1 1. the output multiplicative perturbed robust stability problem can be treated as a sensor noise rejection problem shown in Figure 9.W2 ? W1 e n ~ Figure 9.2. 9. i. integral of square error or the magnitude of the steady state error with . In general.7 and vise versa. Suppose the weighting matrices Wd We 2 RH1 and the performance criterion is to keep the error e as small as possible in some sense for all possible models belong to the set .1.. Nominal Performance A robust stability problem can be viewed as another nominal performance problem. with respect to bounded disturbances. This will be considered in Chapter 11. a much more general result along this line of equivalence: any robust stability problem (with open ball of uncertainty) can be regarded as an equivalent performance problem. say.5 Equivalence: Robust Stability vs.5 is robustly stable for k k1 < 1 i the H1 norm of the transfer function from w to z . L1 norm.3. is no greater than 1. in fact.
K .9. the H2 performance analysis with H1 norm bounded model uncertainty is much harder and little studied. ~ Let Ted~ denote the transfer matrix between d and e. js "w j t 2 2 .1 Wd P 2 : (9:4) And the robust H2 performance analysis problem can be stated as nding sup Ted~ 2 : To simplify our analysis. Nevertheless. this problem is sensible and important for the reason that performance is sometimes more appropriately speci ed in terms of the H2 norm than the H1 norm. then Ted~ = We (I + P K ). and assume that the model is given by a multiplicative uncertainty set = f(1 + wt )p : 2 RH1 k k1 < 1g : Assume further that the system is robustly stabilized by a controller k. consider a scalar system with Wd = 1.1 Robust H2 Performance Although nominal H2 performance analysis is straightforward and involves only the computation of H2 norms. The former design criterion leads to the socalled 9. Then P 2 w sup Ted 2 = sup 1 + s " ~ wt k k1 <1 P 2 w = 1 . Unstructured Robust Performance ~ d ? Wd d ? y We e  223 re . On the other hand.6 P 2 e  Figure 9. we will focus primarily on the H2 and H1 performance objectives with unstructured model uncertainty descriptions. respectively. P = p.8: Diagram for Robust Performance Analysis L1 optimal control framework and the latter leads to H2 and H1 design frameworks. We = ws . model uncertainties are more conveniently described using the H1 norm bounded sets and arise naturally in identi cation processes. respect to sinusoidal disturbances.4. In this section. The performance under structured uncertainty will be considered in Chapter 11.4.
1 and = 1 .. 9.224 MODEL UNCERTAINTY AND ROBUSTNESS where " = (1 + pk). the upper bounds do not seem to be insightful in the H2 setting.. i. It will be assumed throughout that the controller K internally stabilizes the nominal plant P . and. The exact analysis for the matrix case is harder to determine although some upper bounds can be derived as we shall do for the H1 case below. therefore. we have Ted~ = We So (I + W1 W2 To). Ted~ 1 1 8 2 RH1 k k1 < 1: . By scaling the error e (i. an output multiplicatively perturbed system will be analyzed rst. It should be pointed out that synthesis for robust H2 performance is much harder even in the scalar case although synthesis for nominal performance is relatively easy and will be considered in Chapter 14. The perturbed model can be described as := f(I + W1 W2 )P : 2 RH1 k k1 < 1g (9:6) with W1 W2 2 RH1 .e. sup kek2 for some small . The analysis for other classes of models can be done analogously. Then the robust performance criterion in this case can be described as requiring that the closedloop system be robustly stable and that Ted~ 1 1 8P 2 : (9:5) More speci cally.4. ". and they may shed some light on the nature of these problems.1 Wd and the robust performance is satis ed i kW2 ToW1 k1 1 and The exact analysis for this robust performance problem is not trivial and will be given in Chapter 11.e. However. The explicit system diagram is as shown in Figure 9. some su cient conditions are relatively easy to obtain by bounding these two inequalities.5. are omitted.2 Robust H1 Performance with Output Multiplicative Uncertainty kd~k2 1 Suppose the performance criterion is to keep the energy of error e as small as possible. by properly selecting We ) we can assume without loss of generality that = 1. However. For this class of models.
9.1 ] (Wd ) S (W = (I(WeWo ) W d ) ) + 1 2 To (We So) (Wd ) 1 . it is easy to see that (Ted~) (We So ) (I + W1 W2 To).7) and condition (9.7 Suppose P 2 f(I + W1 W2)P : (i) for each frequency ! 225 2 RH1 k k1 < 1g and K internally stabilizes P . Then the system robust performance is guaranteed if either one of the following conditions is satis ed (Wd ) (We So) + (W1 ) (W2 To ) 1 (ii) for each frequency ! where W1 and Wd are assumed to be invertible and (W1. Now for any frequency !. (9:7) (9:8) (W1. Unstructured Robust Performance Theorem 9.1 Wd ) : 1 . 2 Remark 9. (W1 ) (W2 To ) ( ) : Hence condition (9. (W2 ToW1 ) ( ) 1 is guaranteed for all 2 RH1 with k k1 < 1 Hence by condition (9.1 Wd ) is the condition number.1 Wd ) (We So Wd ) + (W2 To W1 ) 1 Proof.1 Wd ) and then (Ted~) (We So Wd ) (W1.1 Wd ). So it is su cient to show that Ted~ 1 1 8 2 RH1 k k1 < 1.2 It is not hard to show that either one of the conditions in the theorem is also necessary for scalar valued systems. (Ted ) ~ at all frequencies. (W1 W2 To ) (We So) (Wd ) 1 . .1 (I + W2 ToW1 ).8).4. Similarly.8) guarantee that kW2 To W1 k1 1. suppose W1 and Wd are invertible write Ted~ = We So Wd (W1. It is obvious that both condition (9.7) guarantees (Ted ) 1 for all 2 RH1 with k k1 < 1 at all ~ frequencies.1 (W1.
1Wd) Remark 9.K 6 . NP + RS RP. ~ Remark 9. and the exact stability analysis for such systems is not trivial and will be studied in Chapter 11. This conjecture is indeed true however. i.1 : (We So Wd ) + (W1.4 with the uncertainty model set given by := (I + Wd e We ). Hence it is clear that in this case nominal performance plus robust stability almost guarantees robust performance.e . the equivalent model uncertainty is structured.1 Wd ).e . k k1 < 1. as shown in Figure 9. ~ . ~ So another alternative su cient condition for robust performance can be obtained as . This is particularly the case if W1 = w1 (s)I and Wd = wd (s)I .4 Note that in light of the equivalence relation between the robust stability and nominal performance.1 Wd ) (W2 ToW1 ) 1: A similar situation also occurs in the skewed case below.5 Note that if W1 and Wd are invertible.W1 ? Wd e .9: Robust Performance with Unstructured Uncertainty vs. then Ted~ can also be written as Ted~ = We So Wd I + (W1. e .1 (I + W1 W2 )P and k e k1 < 1.3 Suppose (W1.W2 . We will not repeat all these variations. it is reasonable to conjecture that the above robust performance problem is equivalent to the robust stability problem in Figure 9. However.P ? ? .226 MODEL UNCERTAINTY AND ROBUSTNESS 1 (weighting matrices satisfying this condition are usually called round weights). Robust Stability with Structured Uncertainty Remark 9.9. this is not necessarily true for other types of uncertainties as will be shown later.We Figure 9.1 W2 ToW1 (W1. Recall that (We So Wd ) 1 is the necessary and su cient condition for nominal performance and that (W2 To W1 ) 1 is the necessary and su cient condition for robust stability.e..1 Wd ) .
if W2 is invertible.1 Wd = We So Wd I + (Wd. ~ Remark 9.4.1 I + (W1. but the uncertainty model is in input multiplicative form: := fP (I + W1 W2 ) : 2 RH1 k k1 < 1g : For systems described by this class of models.1 . Wd . see Figure 9.9..1 PW1 ) (W2 Ti W1 )(Wd.1 PW1 ) (W2 P . the robust stability condition becomes kW2 Ti W1 k1 1 and the nominal performance condition becomes kWe So Wd k1 1: ~ ~ To consider the robust performance. there are many di erent variations of su cient conditions although (9. and P are invertible and. Unstructured Robust Performance 227 We now consider the system with skewed speci cations.1 W2. Assume that P W1 W2 .10) may be the most useful one. the system performance is still measured in terms of output sensitivity.1 Wd ): Then the following results follow easily. the uncertainty and performance are not measured at the same location.6 If the appropriate invertibility conditions are not satis ed.4.1 Wd ) (We So Wd ) + (W1.8 Suppose P 2 = fP (I + W1 W2 ) : 2 RH1 k k1 < 1g and K internally stabilizes P .e. Theorem 9.1 ) (W2 ToW1 ) 1: (9:10) 9.1 : The last equality follows if W1 . let Ted~ denote the transfer matrix from d to e. can also be written as ~ Ted~ = We So Wd (W1. and Wd are square and invertible.1 PW1 ) (W2 P . For instance.1 W2. i.1 (W1.1 PW1 ) (W2 Ti W1 ) 1 (9:9) (ii) for each frequency ! (W1.8. Then ~ Ted~ = We So (I + PW1 W2 KSo). then an alter .1 )(W2 ToW1 ) .1 Wd ). Then the system robust performance is guaranteed if either one of the following conditions is satis ed (i) for each frequency ! (We So Wd ) + (Wd.1 PW1 ).3 Skewed Speci cations and Plant Condition Number native su cient condition for robust performance can be given by (Wd ) (We So ) + (PW1 ) (W2 KSo ) 1: Similar to the previous case.
Further. These classes of problems are called skewed problems or problems with skewed speci cations.1 ).2 Since. Wd = I and W2 = wt I where wt 2 RH1 is a scalar function. Alternative condition can be derived so that the condition related to nominal performance is scaled by the condition number. Since (P ) 1. ~ ~ Remark 9.10) can be written as (We So ) + (P ) (wt To ) 1 8!: Comparing these conditions with those obtained for nonskewed problems shows that the condition related to robust stability is scaled by the condition number of the plant3 . PK 6= KP .9 Suppose K is invertible. it is clear that the skewed speci cations are much harder to satisfy if the plant is not well conditioned. ~ These skewed speci cations also create problems for MIMO loop shaping design which has been discussed brie y in Chapter 5.8 It is also noted that the robust performance condition is related to the condition number of the weighted nominal model.7 It is important to note that in this case.1 Wd ) 1. P is assumed to be invertible. The idea of loop shaping design is based on the fact that robust performance is guaranteed by designing a controller with a su cient nominal performance margin and a su cient robust stability margin. the output multiplicative perturbed robust performance is guaranteed by designing a controller with twice the required nominal performance and robust stability margins.228 MODEL UNCERTAINTY AND ROBUSTNESS Remark 9.3. if (W1. See Stein and Doyle 1991].1 (I + Ti W1 W2 ). Then a su cient condition for robust performance is given by (K ) (Si ws ) + (Ti W1 ) 1 8! with (K ) := (K ) (K . So in general if the weighted nominal model is illconditioned at the range of critical frequencies. then the robust performance condition may be far more restrictive than the robust stability condition and the nominal performance condition together. This is equivalent to treating the input multiplicative plant uncertainty as the output multiplicative controller uncertainty. in general. This problem will be discussed in more detail in section 11.3 of Chapter 11. the robust stability margin or tolerances for uncertainties at the plant input and output are generally not the same. the robust stability condition is ~ given in terms of Li = KP while the nominal performance condition is given in terms of Lo = PK . Then robust performance condition (9. Remark 9. then T ~ can be written as ed Assume further that We = I Wd = ws I W2 = I where ws 2 RH1 is a scalar function.1 Si KWd: . For example. assume W1 = I . For simplicity. 2 3 ~ Ted~ = We K .
taken from the textbook Franklin. we show that the gain margin and phase margin de ned in classical control theory may not be su cient indicators of a system's robustness.5 Gain Margin and Phase Margin In this section.P ? e  Figure 9. Gain Margin and Phase Margin 229 The fact that the condition number appeared in the robust performance test for skewed problems can be given another interpretation by considering two sets of plants 1 and 2 as shown below. 2 9.wt  ? e . 1 2 := fP (I + wt ) : := f(I + wt )P : ~ 2 RH1 k k1 < 1g 2 RH1 k k1 < 1g : .0:05 0 0 0 0:7 7 6 7 s + 1)(s + 0 07) 6 0 . Powell.0s:05 0:(7(s + 1)(s + :0:13) P (s) = 6 0 7 a(s) 4 1 0 0 0 0 5 0 1 0 0 0 where a(s) = (s + 1)(s + 0:1707)(s + 0:02929). The condition number of a transfer matrix can be very high at high frequency which may signi cantly limit the achievable performance. then 2 1 if jwt j jwt j (P ) 8! ~ since P (I + wt ) = (I + wt P P . The example below.1 )P .0:2 0:1 1 0 1 3 6 .5.9.P  . 1990]. and Workman.11 may increase with the frequency: .1 1 0 7 = 1 . Let L(s) be a scalar transfer function and consider a unit feedback system such as the one shown in the following diagram: .wt ~ . shows that the condition number shown in Figure 9.10: Converting Input Uncertainty to Output Uncertainty Assume that P is invertible.
.j L0 (s) with min < < max but unstable for L(s) = e.j min L0(s) where .6 .230 3 MODEL UNCERTAINTY AND ROBUSTNESS 10 10 condition number 2 10 1 10 3 10 0 10 2 10 1 10 frequency 0 10 1 10 2 Figure 9.12 where kmax and kmin represent how much the loop gain can be increased .j max L0 (s) and for L(s) = e.L(s)  Suppose that the above closedloop feedback system with L(s) = L0 (s) is stable. min 0 and 0 max These margins can be easily read from the openloop system Nyquist diagram as shown in Figure 9.11: Condition Number (!) = (P (j!))= (P (j!)) e . Then the system is said to have Gain Margin kmin and kmax if the closedloop system is stable for all L(s) = kL0(s) with kmin < k < kmax but unstable for L(s) = kmax L0 (s) and for L(s) = kmin L0 (s) where 0 kmin 1 and kmax 1. Phase Margin min and max if the closedloop system is stable for all L(s) = e.
K = 1 and the stability margins can be easily shown to be 1 .1 a2 .. Similarly max and min represent how much loop phase lag and lead can be tolerated without causing instability. ab a2 max min max . Now let L = PK and consider a controller b K = bs+ s1 b > 0: + It is easy to show that the closedloop system is stable for any 1 < b < a: To compute the stability margins. gain margin or phase margin alone may not be a su cient indicator of a system's robustness. !0 k = 1 ! 1 k = ab ! a2 min a P = as. 1/K min 1/K max 1 φ min 1 φ max L(jω) L(jω) Figure 9. without causing instability. s1 a > 1 .5. 1 =: : kmin = a kmax = a min = .12: Gain Margin and Phase Margin of A Scalar System However. very large gain margin but arbitrarily small phase margin. respectively. 1 (ii) a < b < a and b ! a: in this case =. To be more speci c. consider three cases: (i) b = 1: in this case.9. consider a simple dynamical system with a stabilizing controller K . max = sin a2 + 1 It is easy to see that both gain margin and phase margin are very large for large a. a i. Gain Margin and Phase Margin 231 and decreased.e.
along two orthogonal transverse axes.14.1 0).e. we show through an example that the classical control theory may not be reliable when it is applied to MIMO system design. !2 (I1 .6 De ciency of Classical Control for MIMO Systems In this section. respectively. 1 1 (iii) a < b < a and b ! a : in this case k = 1 ! 1 k = ab ! 1 min i. x and y. It is easy to see that the system has at least the same gain margin and phase margin as the system with controller K = 1. Denote by !1 and !2 the angular velocities of the body with respect to the x and y axes. The problem is that the gain margin and phase margin do not give the correct indication of the system's robustness when the gain and phase are varying simultaneously.13 for a = 2 and b = 1 b = 1:9 and b = 0:55. T1 and T2 . Such a controller is complicated to construct however. Sometimes. Assume that the angular velocity of the spinning body with respect to the z axis is constant. gain margin and phase margin together may still not be enough to indicate the true robustness of a system. as shown in Figure 9. and I3 .232 MODEL UNCERTAINTY AND ROBUSTNESS min = . and z axes are I1 .13 by the dotted line. very large phase margin but arbitrarily small gain margin. respectively. Then the Euler's equation of the spinning body is given by I1 !1 . Assume further that the inertia of the spinning body with respect to the x y. respectively. I3 ) = T1 _ I1 !2 . For example. the following controller should give the reader a good idea of its construction: + + 0 55 7 2 + + Kbad = 3s:3s 3:31 0s:55s :+ 1 1s:2 s 1:51s:5s 1:71 : + + + a max min max The openloop frequency response of the system with this controller is also shown in Figure 9. I2 = I1 .. I1 ) = T2 : _ . Consider a symmetric spinning body with torque inputs. !1 (I3 . 9.1 0). . Therefore this system is less robust with respect to the simultaneous gain and phase perturbations. > k < 1 k >a min ab max max !2 but the Nyquist plot is arbitrarily close to (. it is possible to construct a (complicated) controller such that =. but the Nyquist plot is closer to (. The openloop frequency response of the system is shown in Figure 9.
b = 0:55(dashdot) and with Kbad (dotted) De ne Then the system dynamical equations can be written as !1 = 0 a _ !1 u1 !2 _ . + 1) ass.+a1) : 2 s u1 := T1 =I1 u2 T2 =I1 a := (1 . sin cos where tan := a.9.a 0 !2 + u2 : Now suppose that the angular rates !1 and !2 are measured in scaled and rotated coordinates: cos sin !1 = 1 a !1 y1 = 1 !2 . (There is no speci c physical meaning for the measurements of y1 and y1 but they are assumed here only for the convenience of discussion.5 0 0.a 1 !2 y2 cos . I3 =I1 ) : . De ciency of Classical Control for MIMO Systems 1 233 0. b = 1:9(dashed).5 Figure 9.5 1 1.a(.13: Nyquist Plots of L with a = 2 and b = 1(solid).) Then the transfer matrix for the spinning body can be computed as Y (s) = P (s)U (s) with 1 s a2 ( P (s) = s2 + a2 .5 1 0.5 0 0.5 2 2.6.5 2 1.
15: Closedloop with a \Bizarre" Controller Then the closedloop transfer function is given by 1 Y (s) = s + 1 1 0 R(s) 0 1 .14: Spinning Body Suppose the control law is chosen as u = K1 r . K1 r Figure 9.234 MODEL UNCERTAINTY AND ROBUSTNESS z x y Figure 9. y where 1 1 K1 = 1 + a2 a .a : 1 y P u f r ~ 6 .
min = 90o and gain margin kmin = 0 kmax = 1.17.e. each single loop has the openloop transfer function as 1 so each loop has phase margin max = . Figure 9.9. s w y1 6 z ? e y2 u1 P u2 e 6 . Furthermore. De ciency of Classical Control for MIMO Systems 235 and the sensitivity function and the complementary sensitivity function are given by 1 1 s 1 S = (I + P ). as shown below. i. e 6 .T11 = . s + 1 : 1 Then the maximum allowable perturbation is given by k k < 1 =1 which is independent of a.16: OneLoopAtATime Analysis Denote 1 z (s) w(s) = .a T = P (I + P ). Consider a multivariable perturbation. P = (I + )P . with = 11 12 2 RH1 21 22 kT11 k1 . then the maximum allowable perturbation is much smaller.16.1 = s + 1 .a a : s 1 It is noted that this controller design has the property of decoupling the loops. Similarly the maximum allowable perturbation on the other loop is also 1 by symmetry..1 = s + 1 a . if both loops are perturbed at the same time. as shown in Figure 9. However.6. Suppose one loop transfer function is perturbed. as shown in Figure 9.
consider = d = 11 2 R2 2 : 22 Then the closedloop system is stable for every such i 1 . det(I + T d) = (s + 1)2 s2 + (2 + 11 + 22 )s + 1 + 11 + 22 + (1 + a2 ) 11 22 has no zero in the closed righthalf plane. Then by the small gain theorem.17: Simultaneous Perturbations a 2 2 transfer matrix such that k k1 < . Hence the stability region is given by 2 + 11 + 22 > 0 1 + 11 + 22 + (1 + a2 ) 11 22 > 0: It is easy to see that the system is unstable with 1 : 11 = . the system is robustly stable for every such i 1 1 kT k1 = p1 + a2 ( 1 if a 1): In particular. ~ ? r1 f y2 ? / f 22 ? f g22 6 .236 MODEL UNCERTAINTY AND ROBUSTNESS y1 f 6 o S S f 11 12 21 6 g11 g12 g21 . 22 = p 1 + a2 . f r2 ~ Figure 9.
Some further extensions are also presented in Tits and Fan 1994]. In the same reference. some extensions of stability and performance criteria under various structured/unstructured uncertainties are given. The book by Desoer and Vidyasagar 1975] contains a quite extensive treatment and applications of this theorem in various forms.18: Stability Region for a = 1 The stability region is drawn in the Figure 9. Robust stability conditions under various uncertainty assumptions are discussed in Doyle.707 δ 11 Figure 9. Hence an MIMO method has to be used. .18.7 Notes and References The small gain theorem was rst presented by Zames 1966]. Notes and References δ 22 237 0. It was rst shown in Kishore and Pearson 1992] that the small gain condition may not be necessary for robust stability with closedball perturbed uncertainties. This clearly shows that the analysis of an MIMO system using SISO methods can be misleading and can even give erroneous results.707 0. and Stein 1982]. 9.7.9. Wall.
238 MODEL UNCERTAINTY AND ROBUSTNESS .
if c 6= 0 then F (s) can also be written as F (s) = + s(1 .1 for some and 2 C .1 Let M be a complex matrix partitioned as M = M11 M12 2 C (p1 +p2 ) (q1 +q2 ) M21 M22 239 . s). In particular.1 Linear Fractional Transformations This section introduces the matrix linear fractional transformations.Linear Fractional Transformation This chapter introduces a new matrix function: linear fractional transformation (LFT). We show that many interesting control problems can be formulated in an LFT framework and thus can be treated using the same technique. The linear fractional transformation described above for scalars can be generalized to the matrix case. De nition 10. 10 10. It is well known from the one complex variable function theory that a mapping F : C 7! C of the form a + bs F (s) = c + ds with a b c and d 2 C is called a linear fractional transformation.
we will use the general term LFT in referring to both upper and lower LFTs and assume that the contents will distinguish the situations.1 exists. We can also de ne an upper LFT with respect to u as Fu (M ) : C q1 p1 7! C p2 q2 F` (M with u ) = M22 + M21 u (I . F`(M ` ) is a transformation obtained from closing the lower loop on the left diagram similarly Fu (M u ) is a transformation obtained from closing the upper loop on the right diagram.1 exists. The reason for this is that one can use either of these notations to express . Fu (M The matrix M in the above LFTs is called the coe cient matrix. M22 ` ). In most cases. So from the above diagrams. M11 u ).1 M12 provided that the inverse (I .1 M21 provided that the inverse (I . M11 u ). The motivation for the terminologies of lower and upper LFTs should be clear from the following diagram representations of F` (M ` ) and Fu (M u ): z1 y1 M w1 u1 z2 y2  u u2 z1 y1 ` M w2 The diagram on the left represents the following set of equations: u1 = M w1 = M11 M12 u1 M21 M22 = ` y1 w1 u1 while the diagram on the right represents y2 z2 u = M w2 = M11 M12 M21 M22 2 u = u y2 : u2 w2 It is easy to verify that the mapping de ned on the left diagram is equal to F`(M ` ) and the mapping de ned on the right diagram is equal to Fu (M u ).240 LINEAR FRACTIONAL TRANSFORMATION and let ` 2 C q2 p2 and u 2 C q1 p1 be two other complex matrices. M22 ` ). Then we can formally de ne a lower LFT with respect to ` as the map F`(M ) : with C q2 p2 7! C p1 q1 ` ) = M11 + M12 ` (I .
1 N = B . It is also clear from the de nition that. It should also be clear to the reader that in writing F` (M ) (or Fu (M )) it is implied that has compatible dimensions. M22 ) is invertible. DCA 1 A . AC .1 M = AC.10.1 . it is often the case that one expression is more convenient than the other for a given problem. Tzw1 = F` (M ` ) where M may be the controlled plant and tainties or the controllers.. while M12 M21 .C . The following is true. F` (M ). respectively.1 or (C + QD). is that F` (M ) has a nominal mapping. M11 .g. is said to be well de ned (or wellposed) if (I .. and M22 re ect a prior knowledge as to how the perturbation a ects the nominal map.1 (A + QB ) = F`(N Q) with . The physical meaning of an LFT in control science is obvious if we take M as a proper transfer matrix. whenever an LFT is invoked. it will be assumed implicitly to be well de ned. Note that this de nition is consistent with the wellposedness de nition of the feedback system. A useful interpretation of an LFT. Tzw2 = Fu (M u ) may be either the system model uncer De nition 10.e. A similar interpretation can be applied to Fu (M ). F` (M 0) is well de ned hence any function that is not well de ned at the origin cannot be expressed as an LFT in terms of its variables. which requires the corresponding transfer matrix be invertible in Rp (s).DC . In that case. LFT is used to refer to the following matrix functions: (A + BQ)(C + DQ). M11 . the LFTs de ned above are simply the closedloop transfer matrices from w1 7! z1 and w2 7! z2 . Linear Fractional Transformations 241 a given object. . which is the focus of the next chapter. For example.1 C . e. It is clear that the study of an LFT that is not wellde ned is meaningless. hence throughout the book.1 B . f ( ) = 1= is not an LFT of . it is clear that Fu (N ) = F` (M ) with N = M22 M21 .2 An LFT. Then (A + BQ)(C + DQ). In some literature.1 (A + QB ) where C is usually assumed to be invertible due to practical consideration. Lemma 10. i. F`(M ).1 : .1 D D C C . This is why LFT is particularly useful in the study of perturbations. for any M . M12 M11 It is usually not crucial which expression is used however. and is perturbed by . Indeed.1.1 = F`(M Q) (C + QD).1 Suppose C is invertible.
the alternative fractional formula is more restrictive. . A = M121 M11 B = M21 . A I A N = 4 . F` (M Q) = (A + BQ)(C + DQ). .B 0 .1 cannot always be written in the form of F` (M Q) for some M however.242 LINEAR FRACTIONAL TRANSFORMATION The converse also holds if M satis es certain conditions.1 with . F` (M Q) = (C + QD). for an arbitrary LFT F`(M Q).2 Let F`(M Q) be a given LFT with M = M11 M12 .M22 M121 : However. . It should be pointed out that some seemingly similar functions do not have simple LFT representations.3 Let M Q. M11 M211 M22 C = M211 D = .M211 M22 : (b) if M12 is invertible. and G be suitably partitioned matrices: M = M11 M12 M21 M22 Q = Q11 Q12 Q21 Q22 2 A B1 B2 G = 4 C1 D11 D12 5 : C2 D21 D22 3 . (A + QB )(I + QD). then M21 M22 (a) if M21 is invertible.1 = F`(N ) with 3 2 Note that the dimension of is twice as many as Q. Lemma 10.B D 0 D 5 = Q Q : Lemma 10. For example.1 (A + QB ) with . . it can be written as (A + QB )(I + QD). The following lemma summarizes some of the algebraic properties of LFT s. M22 M121 M11 C = M121 D = . . A = M11 M211 B = M12 . neither M21 nor M12 is necessarily square and invertible therefore. .
10. Linear Fractional Transformations (i) Fu (M ) = Fl (N ) with 243 0 0 N = I I M I I = M22 M21 0 0 M12 M11 where the dimensions of identity matrices are compatible with the partitions of M and N . M12M 22 M21 .1 M .1 N = M11 .Fu (Q Then the mapping from w to z is given by 1) 2) Fl (Fu (G 1) Fu (Q 2 )) = Fu (Fl (G Fu (Q 2 )) 1 ) = Fu (N ) . Then the inverse of Fu (M ) exists and is also an LFT with respect to : (Fu (M )).1 21 M22 M22 (iii) Fu (M 1 ) + Fu (Q 2 ) = Fu (N ) with 2 M12 M11 0 Q12 N = 4 0 Q11 M21 Q21 M22 + Q22 (iv) Fu (M 1 )Fu (Q 2 ) = Fu (N ) with 3 5 = 0 1 0 2 : M11 M12Q21 M12 Q22 M11 Q12 N =4 0 M21 M22Q21 M22 Q22 2 3 5 = 0 1 0 2 : (v) Consider the following interconnection structure where the dimensions of 1 are compatible with A: z w Fu (G .M12.1 22 : .1 = Fu (N ) with N given by M . (ii) Suppose Fu (M ) is square and wellde ned and M22 is nonsingular.1.
D22 Q22 ). det P (1) 6= 0. s The following proposition is concerned with the algebraic properties of LFT s in the general control setup. Q22 D22 ). Q11 G). P22 K )(1) = 0. so the proofs are omitted. This property is particularly useful in perturbation analysis and in building the interconnection structure. then the closedloop system parameters are also LFTs of the same variables. respectively. all the above properties can be reduced to the standard transfer matrix operations if = 1 I . P22 K1 ). 6 (b) F`(P K1 ) = F`(P K2 ) implies that K1 = K2 if P12 and P21 have normal full column and row rank in Rp (s). It is usually convenient to interpret an LFT Fu (M ) as a state space realization of a generalized system with frequency structure . Then (a) G is proper if P and K are proper with det(I .1 P21 : K = Fu (Q G) = Q22 + Q21 G(I . F`(P K2 ) = P12 (I .1 G): Proof. (a) is immediate from the de nition of F`(P K ) (or wellposedness condition). K2 P22 ). Proof. In fact. Lemma 10. These properties can be straightforwardly veri ed by the de nition of LFT . K2)(I . G 0 (1) 6= 0 and P12 and 0 0 P21 are square and invertible for almost all s. (c) If P and G are proper.1 .1 . Similar results can be stated for lower LFT. and de ne . 2 Property (v) shows that if the openloop system parameters are LFTs of some variables.1 (K1 . which will be proper since det P (1) 6= 0. (b) follows from the identity F`(P K1 ) . L2 := (I .1 Q12 : (c) it is su cient to show that the formula for K is wellposed and K thus obtained is proper. det P . Let Q = P .244 2 LINEAR FRACTIONAL TRANSFORMATION A + B2 Q22 L1 C2 B2 L2 Q21 B1 + B2 Q22 L1 D21 Q12 L1 C2 Q11 + Q12 L1 D22 Q21 Q12 L1 D21 N =4 C1 + D12 L2 Q22 C2 D12 L2Q21 D11 + D12 Q22 L1 D21 where L1 := (I . and 3 5 = 0 1 0 2 . then K is proper and K = Fu (P .4 Let P = P11 P12 and let K be rational transfer function matrices P21 P22 and let G = F` (P K ).1 .
On the other hand.10.1 y z = Gw z y = P w u ) u = Fu (P . When given proper transfer matrices P and G with compatible dimensions which satisfy conditions in (c).2 A simple interpretation of the result (c) is given by considering the signals in the feedback systems.1 = P21 . Linear Fractional Transformations This expression is wellposed and proper since at s = 1 245 det(I .K u assuming this structure is wellposed. P22 K ). as shown above.1 I G 0 = det P . I 0 P .P22 P121P11 .1.1 Q12 = P21 (I . Q11 G). P22 K = (I . P22 Q21 G(I .P22 K ) 6= 0 .1 This lemma shows that under certain conditions. Then det(I .1 G): ~ . And we have hence u = Ky ) z = F` (P K )w = Gw w u z = P . P22 Q22 ) .1 Q12 : The above form is obtained by using the fact that PQ = I .1 P . 2 Remark 10. Q11 G).1 Q12 = P21 Q12 + P21 Q11 G(I .1 G)y or K = Fu (P . Hence the LFTs are both wellposed 12 and we immediately obtain that F`(P K ) = G as required upon substituting for K and (I . z P w y . since P211 exists and Q. Q11 G). . ~ Remark 10. G 0 6= 0: 0 0 We also need to ensure that F` (P K ) is wellposed: I . Q11 G) = det I . an LFT of transfer matrices is a bijective map between two sets of proper and real rational matrices. there exists a unique proper K such that G = Fl (P K ). the conditions of part (c) show that the feedback systems are wellposed.
More detailed discussion of this issue is beyond the scope of this book. the minimal dimension of depends also on the eld (real. any multivariate (matrix) polynomials are also LFTs in their indeterminates for example. Polynomials A very commonly encountered object in control and mathematics is a polynomial function. . the interested readers should consult the references in 2d or nd systems or lter theory. Here a minimal realization refers to a realization with the smallest possible dimension of . In fact.) of the realization. . In this section and the sections to follow.. It is easy to verify that p( ) can be written in the following LFT form: p( ) = F`(M In ) with 3 2 1 0 0 0 7 7 0 0 7: M= 0 1 7 .246 LINEAR FRACTIONAL TRANSFORMATION 10.. p( ) = a0 + a1 + + an n with indeterminate . For example. . i.e. .. etc. a6 1 0 1 0 a4 0 a1 0 a3 . 5 1 0 0 0 Hence every polynomial is a linear fraction of their indeterminates. . 6 6 6 6 6 4 a0 a1 an. complex..1 an p( Then with 2 6 6 6 6 4 2 2 1 2 ) = a1 1 + a2 2 + a3 1 2 + a4 1 + a5 2 + a6 : p( 1 2 ) = F` (N 3 7 7 ) N= 1 0 0 0 0 7 = 1 I2 I : 7 2 2 a5 0 0 0 a2 5 1 0 0 0 0 It should be noted that these representations or realizations of polynomials are neither unique nor necessarily minimal.2 Examples of LFTs LFT is a very convenient tool to formulate many mathematical objects. if not impossible. in multidimensional systems and lter theory. they will be examined from the LFT point of view. to nd a minimal realization for even a two variable polynomial. some commonly encountered control or mathematical objects are given new perspectives. 7 0 . it is usually very hard. . More generally. As commonly known. .
(Of course. . the inverse of dI is also an LFT as shown in Lemma 10. A system with a state space realization as x = Ax + Bu _ y = Cx + Du has a transfer matrix of G(s) = D + C (sI . F ( 1 2 0) is m ). even if it is possible.1 B = F ( A B 1 I ): u C D s . write N( 1 2 m) = N ( F( 1 2 1 2 m ) (d( 1 2 m )I ). it is not easy to see how to carry out such reduction for a complicated problem. To see that. Now the conclusion follows by the fact that the product of LFTs is also an LFT.) However. we would end up with f ( ) = F` (N I2 ) and 3 2 1 0 . Both N and dI can be represented as LFTs. since d(0 0 0) 6= 0. .2. and. Examples of LFTs 247 As another example of LFT representation. : N =4 1 0 . Rational Functions M= 1 . Then F ( 1 2 m ) can be written as an LFT in ( 1 2 m ) (some i may be repeated). Although the latter can be reduced to the former. the above LFT representation problem is exactly the problem of state space realization for a multidimensional discrete transfer matrix.3. For example. furthermore. A).10. with a nite value at the origin: F (0 0 nite. we consider a rational matrix function (not necessarily proper).1 m ) = d( 1 2 m) where N ( 1 2 m ) is a multivariate matrix polynomial and d( 1 2 m ) is a scalar multivariate polynomial with d(0 0 0) 6= 0. this is usually not a nice way to get an LFT representation for a rational matrix since this approach usually results in a much higher dimensioned than required. + f ( ) = 1 + = F`(M ) with By using the above approach. 5 : State Space Realizations We can use the LFT formulae to establish the relationship between transfer matrices and their state space realizations.
A) B ): Both formulations can correspond to the following diagram:  A B C D The following notation for a transfer matrix has already been adopted in the previous chapters: A B A B ): C D := Fu ( C D .1I 2 =: 1I 0 0 2I where zi denotes the forward shift operator.1 . A) B = D + C (I .1 z2 I .248 Now take LINEAR FRACTIONAL TRANSFORMATION = 1 I .1 0 = z10 I z . take . and let A = A11 A12 A21 A22 then its transfer matrix is B = B1 B2 z1 I B D 0 C = C1 C2 G(z1 z2 ) = D + C ( A =: Fu ( C 0 . the transfer matrix can be written as s A B G(s) = Fu ( C D ): More generally. consider a discrete time 2D (or MD) system realized by the rstorder state space equation x1 (k1 + 1 k2 ) = A11 x1 (k1 k2 ) + A12 x2 (k1 k2 ) + B1 u(k1 k2 ) x2 (k1 k2 + 1) = A21 x1 (k1 k2 ) + A22 x2 (k1 k2 ) + B2 u(k1 k2 ) y(k1 k2 ) = C1 x1 (k1 k2 ) + C2 x2 (k1 k2 ) + Du(k1 k2 ): In the same way.
we say the parameter matrix is the frequency structure of the system state space realization. The ambiguity is benign and convenient and can always be resolved from the context.10. We may rewrite it in our standard form as 1 I = I . This notation means that the transfer matrix can be expressed as an LFT of with the coe cient A B matrix C D . This notation is deliberately somewhat ambiguous and can be viewed as both a transfer matrix and one of its realizations.(p2CAI) . Examples of LFTs 249 It is easy to see that this notation can be adopted for general dynamical systems. .1 ~ ~ G(z ) = Fu (M z + 1 I ) = Fu (M Fu (N z .1 p2I = F (N z .1I ) u s where p I p N = .2II : A B G(s) = C D = Fu (M 1 I ) s Now consider a continuous system where A B M= C D z. AB +B : . multidimensional systems. I .. 2I . 2( ).g. .1 I . In this special case.1 I ) then the corresponding discrete time system realization is given by I . e. Frequency Transformation The bilinear transformation between the z domain and sdomain z+1 s= z.1 ~ M = .1 I )) = Fu (M z . with p .1 transforms the unit disk to the lefthalf plane and is the simplest example of an LFT.1I ). as far as the structure is speci ed.1I (I + z . (A)+1A) C (I .1 ( A) D The transformation from the z domain to the sdomain can be obtained similarly.2. p2I z .
FP : Constrained Structure Synthesis A B1 B2 G = 4 C1 D11 D12 C2 D21 D22 Then it is easy to show that 2 3 5 Using the properties of LFTs.P i .W2 ? n v Simple Block Diagrams A feedback system with the following block diagram can be rearranged as an LFT: z G w y with and . we can show that constrained structure control synthesis problems can be converted to constrained structure constant output feedback problems. Consider the synthesis structure in the last example and assume B K = AK DK : CK K F`(G K ) = F` (M (s) F ) .F .FP .6 i y K .K u v uf 3 5 w= d n 2 z= W2 P 0 0 W2 P W1 G=4 0 .250 LINEAR FRACTIONAL TRANSFORMATION uf W1 u d ? i F .
In particular.1. To motivate this type of uncertainty description. shown below in Figure 10. which are assumed to be unknown but lie in the interval . c.2.1 1]. X X k XX X X X c Figure 10. and k . then the corresponding control problem becomes a constant (constrained) output feedback problem. the actual mass m is within 10% of a nominal mass. and the spring sti ness is within 30% of its nominal value of k. and k are not known exactly. m. but are believed to lie in known intervals.1: Mass/Spring/Damper System The dynamical equation of the system motion can be described by F c_ k x + mx + mx = m: Suppose that the 3 physical parameters m c.10. Examples of LFTs where 251 2 3 7 7 7 7 5 A 0 B1 0 B2 6 0 0 0 I 0 6 M (s) = 6 C1 0 D11 0 D12 6 4 0 I 0 0 0 C2 0 D21 0 D22 B F = AK DK : CK K and Note that F is a constant matrix. Parametric Uncertainty: A Mass/Spring/Damper System 6 F m One natural type of uncertainty is unknown coe cients in a state space model. the block diagram for the dynamical system can be shown in Figure 10.2. not a system matrix. . Now introducing perturbations m . the actual damping value c is within 20% of a nominal value of c. we will begin with a familiar mass/spring/damper system. Hence if the controller structure is xed (or constrained).
 c(1 + 0:2 c ) d 6 + k(1 + 0:3 k ) Figure 10. 0m1 ..1 0 0 0 0 0 . To _ _ _ represent the system model as an LFT of the natural uncertainty parameters m c and k . Suppose that the input signals of the dynamical system are 1 . we shall rst isolate the uncertainty parameters and denote the inputs and outputs of k c and m as yk yc ym and uk uc um. 0:1 7 6 7 0 3 2 6 m 76 76 76 76 56 6 4 x1 x2 F uk uc um 3 7 5 3 7 7 7 7 7 7 7 7 7 5 2 6 4 uk uc um 3 7 5 2 = 6 4 yk yc ym 3 7 5 i.1 0 0 0 . as shown in Figure 10.1 m = m(1 + 0:1 m) = m . " x1 _ x2 _ # x1 6 = F` (M ) 4 x2 F 2 .1 m .c 1 k . respectively. m m (1 + 0:1 m) = F` (M1 m ) : with M1 = selected as x1 = x x2 = x F .k .0:1 1 # x1 _ x2 _ yk yc ym 3 7 7 7 7 7 7 5 2 = 6 6 6 6 6 6 4 0:3k 0 0 0 0:2c 0 .3.m 0 c .1 .m 1 0 m 1 .2: Block Diagram of Mass/Spring/Damper Equation 1 It is easy to check that m can be represented as an LFT in m : 1 1 1 0:1 . Then 2 6 6 6 6 6 6 4 " m : .252 LINEAR FRACTIONAL TRANSFORMATION x 1 s x _ 1 s x 1 m(1+0:1 m ) d F 6 .0:1 m 6 .e. and the output signals are selected as x1 and x2 .
k .c 1 1 1 1 : .0:2 yc . m .0:3 c e e 6 6 uc e 6 uk yk k Figure 10.0:1 =6 0 4 0 k 0 0 c 0 0 m 3 7 5 : x1 s x2 1 s M1 ym e  m um 6 .c . This type of modeling will form the basic building blocks for components with parametric uncertainty. and has the realization 2 G (s) 6 6 =6 6 4 A+ C+ k X i=1 k X i=1 ^ i Ai B + ^ i Ci D + k X i=1 k X i=1 ^ i Bi ^ i Di 3 7 7 7 7 5 : . m . 1 : : : k .m . Consider a linear system G (s) that is parameterized by k uncertain parameters.3: Mass/Spring/Damper System General A ne StateSpace Uncertainty We will consider a special class of state space models with unknown coe cients and show how this type of uncertainty can be represented via the LFT formulae with respect to an uncertain parameter matrix so that the perturbations enter the system in a feedback form.10. 0m1 0 0 0 3 7 7 7 7 7 7 5 2 0 0 . Examples of LFTs where 2 6 6 6 6 6 6 4 253 0 m 1 M = 0:3k k c .2.m 0 1 0 0 0 0:2c 0 .1 0 0 0 0 . F .1 .k .
. a ects the state space model. The various terms in these state equations are interpreted as follows: the nominal system description G(s). ... The structural ^ ^ ^ ^ knowledge about the uncertainty is contained in the matrices Ai Bi Ci . and D Di 2 Rny nu .1 1]. i . .4. Hence. R1 Z1 Rk Zk . given by known matrices A B C and D. . Since G (s) = Fu (M 1 I ) where s 2 M 6 6 =6 6 4 A+ C+ k X i=1 k X i=1 ^ i Ai B + ^ i Ci k X i=1 k X i=1 ^ i Bi ^ i Di 3 7 7 7 7 5 D+ we need to nd an LFT representation for the matrix M with respect to p = diag f 1 I 2 I : : : kIg : To achieved this with the smallest possibly size of repeated blocks.. They re ect how the i'th uncertainty. let qi denote the rank of the matrix " # ^ ^ Ai Bi 2 R( n+ny ) (n+nu ) Pi = ^ ^ for each i. This is shown in Figure 10. Then Pi can be written as Ci Di " Pi = Li Wi Li i Pi = Wi and M can be written as z" " # #" Ri Zi " # where Li 2 Rn qi . we consider the problem of describing the perturbed system via the LFT formulae so that all the uncertainty can be represented as a nominal system with the unknown parameters entering it as the feedback gains. is (A B C D) and the parametric uncertainty in the nominal system is re ected by the k scalar uncertain parameters 1 : : : k . Now. we have Ri i Iqi ] Zi z #{ 2 6 6 4 # M11 M = A B C D } #{ z" M12 + L1 W1 } Lk Wk 1 Iq1 } p {2 3z 76 76 54 M21 . Wi 2 Rny qi .254 LINEAR FRACTIONAL TRANSFORMATION ^ ^ ^ ^ Here A Ai 2 Rn n B Bi 2 Rn nu C Ci 2 Rny n . Ri 2 Rn qi and Zi 2 Rnu qi . and Di . and we can specify them. } { 3 7 7 5 k Iqk . say by i 2 .
Examples of LFTs i.4: LFT Representation of State Space Uncertainty Therefore. and D22 in the diagram are B2 = D12 = C2 = h h h h L1 L2 W1 W2 R1 R2 Lk Rk Zk # i i i Wk and D21 = Z1 Z2 D22 = 0 G ( ) = Fu (F`( M11 M12 M21 0 " i p) 1 I ): s . the matrices B2 C2 D12 D21 . 255 " # M = F`( M11 M12 M21 0 p ):  1I s x _ y y2 A B B2 C D D12 C2 D21 D22 1I x u u2 .e. kI Figure 10..2..10.
. . ... .. ... Mark the inputs and outputs of the 's as y's and u's... (This step is equivalent to computing the transformation in the shadowed box in Figure 10...... .... ........ ...... .. .. 2 .. ... Now the matrix \M " of the LFT can be obtained by computing the corresponding transfer matrix in the shadowed box. ...... ... ... ...7. . K .. . ...... . .6... .... . ...... ..... .... ..... ...256 LINEAR FRACTIONAL TRANSFORMATION 10.. ....... .. ........ in particular...... .......... ... ...... . ... .. ............. ..... . ...... ... .. ..... .... ........ . ........... ... ...... We shall do this in three steps: 1.....) 2 6 6 6 6 6 6 4 1 2 1 y1 y2 y3 y4 z 3 u1 7 6 7 6 u2 7 6 7 = M 6 u3 7 6 7 6 5 4 u4 w 2 3 7 7 7 7 7 7 5 .5.......... ....... ...... . . ..... ..... ... their role in modeling uncertainty........ .... Write z and y's in terms of w and u's with all 's taken out.............. . ..... . ...... ... . ... . . ...... .. . .. .. ...................... ....... .. ......... The basic principle at work here in writing a matrix LFT is often referred to as \pulling out the s".. . ... ... ............ ........ . ..... .... . We shall illustrate the above principle with an example. respectively. . . . . 3..3 Basic Principle We have studied several simple examples of the use of LFTs and... . . .... .. . . .... ... ... ... .. ..... ...... .... .. ..... .. ... .. .. . .......... ... . Consider an input/output relation a+b c 2 z = 1 + d 2 + + 1 2 w =: Gw e2 where a b c d and e are given constants or transfer functions... .. 1 ..... ............... ............ 2. .5: Multiple Source of Uncertain Structure This diagram can be redrawn as a standard one via \pulling out the s" in Figure 10. . .. . .. We would like to write G as an LFT in terms of 1 and 2 . ....6.... Figure 10. .. . We will try to illustrate this with another picture (inspired by Boyd and Barratt 1991]). 3 .. ..... . (This is essentially pulling out the s)...... Consider a structure with four substructures interconnected in some known way as shown in Figure 10....... .... .... . . .. ..... . .. .... . .. Draw a block diagram for the input/output relation with each separated as shown in Figure 10.. .... .
...... . .... ...10.. .. .. ... . .... ......... .. ..... .. ...4 Redhe er StarProducts The most important property of LFTs is that any interconnection of LFTs is again an LFT. 10... ... ..... ....... . .. .. .... ... .. Indeed.. ...... ..... .. .. ...K Figure 10..... .. . ...... ........ ........ ..... ............. .. ...... . ........ . .. .... ..... ........... .. ... ....... Redhe er StarProducts 1 257  2 .. e... can be viewed as special cases of the socalled star product.. ... ? ..... ......... . .. ..... . . ... .. ... .... .. .... ... ........ ........... ....... ..... .... .  3 ........... .... .... .. ... ..... .... .. .... .... . ....... ... .. This property is by far the most often used and is the heart of LFT machinery. ...... . .... . .... ........ ...... . ............................. ....... ..6: Pulling out the s where 2 M= Then 6 6 6 6 6 6 4 0 1 1 0 0 be ae 0 0 e 0 0 0 0 0 bd + c 0 ad 1 " d b a 1 07 7 7 0 7: 7 7 5 3 z = Fu (M )w = 1 I2 0 2 I2 0 # : All LFT examples in the last section can be obtained following the above steps... ... . . ....4. ....... .............. .. ...... ... .... . ............. . . ..g. . ..... it is not hard to see that most of the interconnection structures discussed early....... .. .... ................. . ........ .... . . ....... .. . . .. . .. ..... .... ...... ... feedback and cascade.. .... . ......... ... . ...
P22 K11 is invertible.e e ? .1 K12 : S (P K ) := K21 (I . and assume further that I .d 6 w e 6 Figure 10. In fact.8.7: Block diagram for G Suppose that P and K are compatibly partitioned matrices " # " P = P11 P12 K = K11 K12 P21 P22 K21 K22 such that the matrix product P22 K11 is well de ned and square. Then the star product of P and K with respect to this partition is de ned as # Fl (P K11 ) P12 (I . as shown in Figure 10. this star product may be well de ned for one partition and not well de ned for another however. In a block diagram. Now suppose that P and K are transfer matrices with state space representations: A B1 B2 P = 6 C1 D11 D12 4 C2 D21 D22 Then the transfer matrix 2 3 7 5 AK BK 1 BK 2 K = 6 CK 1 DK 11 DK 12 7 : 4 5 CK 2 DK 21 DK 22 " # " # 2 3 S (P K ) : w 7! z w ^ z ^ . we will not explicitly show this dependence because it is always clear from the context. this dependence appears. P22 K11 ).258 LINEAR FRACTIONAL TRANSFORMATION a z ?u4 e 2 y4 e 6 c 6 u3 b 2 y3 u1 y2 1 1 y1 u2 . K11 P22 ).1 P21 Fu (K P22 ) " # Note that this de nition is dependent on the partitioning of the matrices P and K above.
D22 DK 11 In fact.1 DK 12 DK 21 R.1 D21 BK 2 + BK 1 R.1 DK 11 C2 B2 R.1 C2 BK 1 R AK + BK 1 R.1 D21 DK 22 + DK 21 R.1 DK 12 BK 1 R.1 D22 DK 12 # # R = I . DK 11 D22 : # " # A = S B = S " " A C2 B1 D21 B2 D22 B2 D22 DK 11 BK 1 " DK 11 BK 1 CK 1 AK DK 12 BK 2 #! #! .1 D22 DK 12 # ~ C1 + D12 DK 11 R. Redhe er StarProducts z P w z z ^ 259 z ^ aa !!! a ! !! aaa K S (P K ) w w ^ w ^ Figure 10.10.4.1D21 D12 R.1 DK 11 D21 B2 R.1 D22 CK 1 ~ D11 + D12 DK 11 R.1C2 CK 2 + DK 21 R.1 CK 1 .1 C2 D12 R. it is easy to show that ~ R = I .1 CK 1 DK 21 R.8: Interconnection of LFTs has a representation 2 3 7 7 5 A B1 B2 6 6 C S (P K ) = 4 1 D11 D12 C2 D21 D22 where A B = C D " # A = B = C = D = " " " " ~ ~ A + B2 R.1 D22 CK 1 # ~ ~ B1 + B2 R.
5 Notes and References This chapter is based on the lecture notes by Packard 1991] and the paper by Doyle. and Zhou 1991]. . Packard.260 LINEAR FRACTIONAL TRANSFORMATION C = S D = S " " C1 C2 D11 D21 D12 D22 D12 D22 # " # DK 11 DK 21 " DK 11 DK 21 CK 1 CK 2 DK 12 DK 22 #! #! : 10.
in order to use the results in Chapter 9 for unstructured cases. it is necessary to re ect all sources of uncertainties from their known point of occurrence to a single reference location in the loop. a uni ed approach may relieve the mathematical burden of dealing with speci c problems repeatedly. Readers might have already had some idea about the conservativeness in such re ection from the skewed speci cation problem. arbitrarily more conservative perturbation in order to maintain a simple cone bounded representation at the reference location. in the same fashion as single unstructured uncertainty. which is a formidable problem in the standing point of Chapter 9. of course. does not mean that those special problems and their corresponding results are not important on the contrary. Furthermore. Such re ected uncertainties invariably have a great deal of structure which must then be \covered up" with a large.Structured Singular Value It is noted that the robust stability and robust performance criteria derived in Chapter 9 vary with the assumptions on the uncertainty descriptions and performance requirements. they are sometimes very enlightening to our understanding of complex problems such as those in which complex problems are formed from simple problems. We will show in this chapter that they can all be treated in a uni ed framework using the LFT machinery introduced in the last chapter and the structured singular value to be introduced in this chapter. This. where an input multiplicative uncertainty of the plant is re ected at the output and the size of the re ected uncertainty is proportional to the condition number of the plant. if a system is subject to multiple sources of uncertainties. In general. the uni ed framework introduced here will enable us to treat exactly the robust stability and robust performance problems for systems with multiple sources of uncertainties. Indeed. On the other hand. the re ected uncertainty may be proportional to the condition 261 11 .
any interconnected system may be rearranged to t the general framework in Figure 11. The assumptions which characterize the uncertainty. Given that the nominal model is an FDLTI system. The models are assumed to be FDLTI systems. performance. Performance is measured as weighted output variances. or as power. The structured singular value is de ned exactly for that purpose. and nominal models determine the analysis techniques which must be used. Thus it is highly desirable to treat the uncertainties as they are and where they are. or as weighted output Lp norms. Various modeling assumptions will be considered and the impact of these assumptions on analysis and synthesis methods will be explored in this general framework. TOOLS is a trademark of MUSYN Inc. the interconnection system has the form 2 3 P11 (s) P12 (s) P13 (s) P (s) = 6 P21 (s) P22 (s) P23 (s) 7 4 5 P31 (s) P32 (s) P33 (s) and the closedloop system is an LFT on the perturbation and the controller given by z = Fu (F` (P K ) ) w = F` (Fu (P ) K ) w: We will focus our discussion in this section on analysis methods therefore. The uncertain inputs are assumed to be either ltered white noise or weighted power or weighted Lp signals. Denote M (s) = F` (P (s) K (s)) = M11(s) M12 (s) M21(s) M22 (s) 1 2 " # SIMULINK is a trademark of The MathWorks. Various combinations of these assumptions form the basis for all the standard linear system analysis tools. Inc. Note that uncertainty may be modeled in two ways. The performance of a system is measured in terms of the behavior of the outputs or errors. such as SIMULINK1 and TOOLS2.262 STRUCTURED SINGULAR VALUE number of the transfer matrix between its original location and the re ected location. The perturbations are assumed to be themselves FDLTI systems which are normbounded as inputoutput operators. the controller may be viewed as just another system component and absorbed into the interconnection structure. many software packages. . Although the interconnection structure can become quite complicated for complex systems.1 General Framework for System Robustness As we have illustrated in the last chapter. 11.1. either as external inputs or as perturbations to the nominal model. are available which could be used to generate the interconnection structure from system components.
3. including the robust performance problem for systems with unstructured . However.1. For example. the analysis is not so simple for systems with multiple sources of model uncertainties.1: General Framework and then the general framework reduces to Figure 11.1 M12 w: z M w Figure 11.) The analysis results presented in the previous chapters together with the associated synthesis tools are summarized in Table 11. Now let w := d1 d2 " # z := e1 e2 " # then the system is robustly stable for all (s) 2 RH1 with k k1 < 1 if and only if Fu (M ) 2 RH1 for all admissible . they can be made equivalent with suitably chosen w and z .2: Analysis Framework Suppose K (s) is a stabilizing controller for the nominal plant P .1 with various uncertainty modeling assumptions.2. which is guaranteed by kM11 k1 1. Then M (s) 2 RH1 . General Framework for System Robustness z P w 263 . (Note that this is not necessarily equivalent to (I . However.K Figure 11.5. the stability of Fu (M ) does not necessarily imply the internal stability of the closedloop feedback system. In general. M11 ). where z = Fu (M )w = M22 + M21 (I .11.1 2 RH1 if belongs to a closed ball as shown in Theorem 9. consider again the multiplicatively perturbed system shown in Figure 11. M11 ).
we shall simply cover the real parametric uncertainty with norm bounded dynamical uncertainty.  F (s) . if a system is built from components which are themselves uncertain. .W1 ? e Figure 11.6 e . the uncertainty in the system level is structured involving typically a large number of real parameters. k k1 < 1.3: Multiplicatively Perturbed Systems uncertainty.K e 2 P . by absorbing any weights. The stability analysis involving real parameters is much more involved and is beyond the scope of this book and that. Thus we shall assume that (s) takes the form of : : : s IrS 1 : : : F ] : i (s) 2 RH1 j 2 RH1 g with k i k1 < 1 and k j k1 < 1.. then. in general. . instead. Figure 11. As we have shown in the last chapter. . and. Moreover.W2 3 d3 .264 STRUCTURED SINGULAR VALUE d2 d1 e1 ?ee . .1 (s)I . M11 (s) .4 is stable. Then the system is robustly stable i the intercon1 Ir1 (s) = fdiag nected system in Figure 11.1 can be applied to the analyses of the system's robust stability in two ways: . the interconnection model M can always be chosen so that (s) is block diagonal..4: Robust Stability Analysis Framework The results of Table 11..
because this test ignores the known block diagonal structure of the uncertainties and is equivalent to regarding as unstructured. This can be arbitrarily conservative3 in that stable systems can have arbitrarily large kM11k1 . This optimism is also clearly shown in the spinning body example. General Framework for System Robustness Input Assumptions Performance Speci cations Perturbation Assumptions Analysis Tests 265 Synthesis Methods LQG E (w(t)w(t) ) = I E (z (t) z (t)) 1 w = U0 (t) E (U0 U0 ) = I E (kz k2 ) 1 2 =0 kM22 k2 1 WienerHopf H2 =0 kwk2 1 kwk2 1 kz k2 1 Internal Stability kM22 k1 1 Singular Value Loop Shaping k k1 < 1 kM11 k1 1 H1 Table 11. The di erence between the stability margins(or bounds on ) obtained in (1) and (2) can be arbitrarily far apart.1. . Of course. 3 By \arbitrarily conservative.11. Only when the margins are close can conclusions be made about the general case with structured uncertainty.1: General Analysis for Single Source of Uncertainty (1) kM11 k1 1 implies stability. The exact stability and performance analysis for systems with structured uncertainty requires a new matrix function called the structured singular value (SSV) which is denoted by ." we mean that examples can be constructed where the degree of conservatism is arbitrarily large. but not conversely. (2) Test for each i ( j ) individually (assuming no uncertainty in other channels). This test can be arbitrarily optimistic because it ignores interaction between the i ( j ). see for example the spinning body example. other examples exist where it is quite reasonable.
In view of the characterization of the largest singular value of a matrix M (s) given by (11. To quantify the smallest destabilizing structured . (M (s)) can be written as (M (s)) = min f ( ) : det (I . By small gain theorem. 1 = kM k := sup (M (s)) = sup (M (j!)) 1 max is unstructured. we shall de ne (11:2) (M (s)) = min f ( ) : det (I . To be more speci c. Next increase until max so that the closedloop system becomes unstable. Now let > 0 be a su ciently small number such that the closedloop system is stable for all stable k k1 < . Then it is obvious that the robust stability margin of the feedback system with structured uncertainty is 1 = sup (M (s)) = sup (M (j!)): if max s2C + ! The last equality follows from the following lemma.1). the concept of singular values needs to be generalized.2.2 Structured Singular Value 11. M (s1 ) = 0 is structuredg ) as the largest structured singular value of M (s) with respect to the structured . Note that for any xed s 2 C + . So max is the robust stability margin. s2C + ! . the structured singular value is nothing but a straightforward generalization of the singular values for constant matrices. the feedback system becomes unstable if det(I .1 Basic Concept Conceptually. it is instructive at this point to consider again the robust stability problem of the following standard feedback interconnection with stable M (s) and (s). Since the closedloop poles are given by det(I .266 STRUCTURED SINGULAR VALUE 11. M ) = 0. w1 + 6 + . M (s) (s)) = 0 for some s 2 C + . M (s) 1 ) = 0 is unstructuredg : (11:1) In other words. the reciprocal of the largest singular value of M is a measure of the smallest unstructured that causes instability of the feedback system.e e1 M + e2 e + w2 ? One important question one might ask is how large (in the sense of k k1 ) can be without destabilizing the feedback system.
With those integers given.2. We consider matrices M 2 C n n . and their dimensions.1 Let be a structured set and M (s) 2 RH1. while the j 'th full block is mj mj . then by the de nition of . we de ne C n n as = diag 1 Ir1 : : : s IrS 1 : : : F ] : i 2 C j 2 C mj mj : (11:3) For consistency among all the dimensions. and we introduce the following notation: B = f 2 : ( ) 1g (11:4) Bo = f 2 : ( ) < 1g (11:5) . This in turn implies that (M (j !)) > 1= since ( ) < . 2 The formal de nition and characterization of the structured singular value of a constant matrix will be given below. There are two types of blocks: repeated scalar and full blocks. this structure is. It is clear that s2C + C+ sup (M (s)) = sup s2C + (M (s)) sup ! (M (j!)): Now suppose sups2C + (M (s)) > 1= . there is an underlying structure (a prescribed set of block diagonal matrices) on which everything in the sequel depends. 11.2 De nitions of This section is devoted to de ning the structured singular value. respectively. Structured Singular Value Lemma 11. we will need norm bounded subsets of . M (so ) ) = 0. the total number of blocks. di erent it depends on the uncertainty and performance objectives of the problem. The proof is complete. ^ sups2C + (M (s)) sup! (M (j!)). Two nonnegative integers. there is an so 2 f1g and a complex structured such that ( ) < and det(I . M (j !) ) = ^ 0. For each problem.2. represent the number of repeated scalar blocks and the number of full blocks. This implies that there is a 0 ! 1 and 0 < ^ 1 such that det(I . in general. De ning the structure involves specifying three things: the type of each block. To bookkeep their dimensions. we must have S X i=1 ri + F X j =1 mj = n: Often. The i'th repeated scalar block is ri ri . S and F . we introduce positive integers r1 : : : rS m1 : : : mF . Then s2C + 267 (M (s)) = sup ! sup (M (s)) = sup s2C + (M (j!)): Proof. a matrix function denoted by ( ). In other words.11. In the de nition of (M ).
all blocks are full blocks. i.e. 3 7 7 7 7 5 . De nition 11. To see this. 6 .268 STRUCTURED SINGULAR VALUE where the superscript \o" symbolizes the open ball.1 For M 2 C n unless no n (M ) is de ned as (M ) := min f ( ) : 2 1 det (I . they can come in any order. .. the full blocks in the minimal norm can 3 7 7 7 7 5 x = 6 .. ~ 6 and 2 6 6 4 x1 ~ x2 ~ 3 7 7 7 7 5 2 ..3). 5 yF xF ~ It follows that kxi k = 1 if xi 6= 0.. To keep the notation as simple as possible in (11.1 Without a loss in generality. M singular. M ) = 0g (11:6) each be chosen to be dyads (rank = 1). but restricting them as such saves a great deal in terms of notation. xF ~ := Dx = 6 6 4 6 6 d1 x1 d2 x2 dm xF . and y 6= 0. and de ne x1 x2 xi 2 C mi i = 1 : : : F xF D = diag(d1 Im1 : : : dF ImF ) where di = 1 if xi = 0 and di = 1= kxi k if xi 6= 0. Then there is an x 2 C n such that M x = x. assume S = 0.. kxi k = 0 if xi = 0 . 6 y1 y2 3 7 7 7 7 5 2 := x = 6 ~ 6 6 6 ~ 1 x1 ~ 7 2 x2 7 7 3 . Also. in which case (M ) := 0. 7 : . Remark 11. M is singular for some 2 . de ne a new ~ ~ perturbation ~ 2 C n n as ~ := diag y1 x1 : : : yF xF ] : ~ ~ . we place all of the repeated scalar blocks rst in actuality. Next de ne 2 6 6 4 x = 6 . 4 y = 6 .. Suppose that I . the full blocks do not have to be square. Now partition x conformably with : 2 6 6 4 2 makes I . Hence.
Note that D = D and D ~ = ~ D. M satis es the singularity condition with a perturbation ~ that is no larger (in the ( ) sense) and has rank 1 for each blocks but still satis es the singularity condition. we conclude that (M ) (M ) (M ) : (11:8) These bounds alone are not su cient for our purposes"because the gap between and # 1 0 can be arbitrarily large. then 1 (M ) . ( M ) = j j (M ).1 ~ x = x =) M ~ x = x =) MD ~ ~ is also singular. If = f I : 2 C g (S = 1 F = 0 r1 = n). Hence. continuity of the function : C n n ! R is apparent.2 (M ) = max (M ) In view of this lemma. The smallest one of these is associated with the largest (magnitude) eigenvalue. the function : C n n ! R is not a norm. 2 = C n n (S =0 F =1 m1 = n).3) we must have f In : 2 C g C n n: (11:7) Hence directly from the de nition of and from the two special cases above.e.1 y = x ~ . we have ~ M x = x =) MD. 2B Proof. then (M ) = (M ). then (M ) = (M ). suppose = and consider 0 2 Obviously.. so. for any 2 C . For example.2. (M ) < 1. ( ~ ) 269 ( ) and y = ~ x. for a general Proof.1 x = x =) MD.11. and de ne := (1 ) vu . the spectral radius of M . (M ) = (M ). I . If ( ) < . On the other hand. it is related to how \big" the matrix is. let u and v be unit equation (11. Then ( ) = (1 ) M M 2 and I . M ) = 0 constraint are reciprocals of nonzero eigenvalues of M . as in (11. so in some sense. Lemma 11. The only 's in If which satisfy the det (I . so I . Applying (M ). M is obviously singular. Structured Singular Value Obviously. though. Hence we have replaced a general perturbation which i. (M ) (M ). M is nonsingular. In general. We can relate (M ) to familiar linear algebra quantities when is one of two extreme sets. ~ An alternative expression for (M ) follows from the de nition.6) implies (M ) vectors satisfying Mv = (M ) u. since it doesn't satisfy the triangle inequality however.
2 .270 " # STRUCTURED SINGULAR VALUE 0 (1) M = for any > 0. The only time they do provide reliable bounds is when . 2 = 0 = 1.1 ImF : : Di 2 C ri ri Di = Di > 0 dj 2 R dj > 0 Note that for any 2 U 2 U .M ) = 0 0 1 so (M ) = 0.3 For all U 2 U and D 2 D . MUU ) = 0. we get (MU ) = (M ) as desired.1 D = det I . det (I . Thus neither nor provide useful bounds even in simple cases. DMD. M ) = det I . M ) = (2) M = . .1 .1=2 1=2 . Since det(I .2. To do this.1 : .1 since D commutes with . . 2 Therefore. DMD.8) can be tightened to max (UM ) max ( M ) = (M ) D2D inf U 2U 2B . Note that the last element in the D matrix is normalized to 1 since for any nonzero scalar . and D 2 D. But det(I .1 (11:14) where the equality comes from Lemma 11. DMD. . For all D 2 D and 2 . Also. for each U 2 U . . so (M ) = 1. Theorem 11. de ne the following two subsets of C n n : diag D1 : : : DS d1 Im1 : : : dF . det (I . Then (M ) = 0 and (M ) = . the bounds can be re ned by considering transformations on M that do not a ect (M ). it is easy to see that min maxi j i j 1 + 1 . M ) = 0 if and only if det (I . Therefore (M ) = DMD. Then (M ) = 0 and (M ) = 1. MD. However. 1+ 2 2 " # D = ( U = fU 2 : UU = In g ) (11:9) (11:10) (11:11) (11:12) (11:13) Proof. the bounds in (11. (MU ) = (UM ) = (M ) = DMD.1 ImF .1 = ( D) M ( D).1=2 1=2 1 . Since U 2 and (U ) = ( ). The argument for UM is the same.1 . but do a ect and . (U ) = ( U ) = ( ) U 2U U 2 U2 D = D: Consequently.
Structured Singular Value 271 Remark 11. there exist matrices for which is less than the in mum. This can be summarized in the following diagram. Thus local search cannot be guaranteed to obtain .1) if 2S + F 3 inf S= 0 1 2 F= 0 yes no 1 2 3 4 yes yes yes no yes no no no no no no no .3 Bounds In this section we will concentrate on the bounds max (UM ) U 2U (M ) D2D inf . be found. so the global minimum can.5 (M ) = D2D (DMD. the upper bound is always equal to (M ). This can be shown as follows: Let D be any nonsingular matrix such that D = D. and for block structures with 2S + F > 3. Theorem 11.1 : inf . Fan. DMD. they can be replaced by any set of nonsingular matrices that satisfy (11. and Doyle. Then there exist a Hermitian matrix 0 < R = R 2 D and a unitary matrix U such that D = UR and inf D .2. in principle.1U = R2D RMR. However.14) does not necessarily be restricted to Hermitian.2 Note that the scaling set D in Theorem 11. Theorem 11. DMD. which shows for which cases the upper bound is guaranteed to be equal to . the upper bound is not always equal to . For block structures satisfying 2S + F 3. The upper bound can be reformulated as a convex optimization problem. but can only yield a lower bound.1 : The lower bound is always an equality Doyle. U 2U Unfortunately. Therefore.14). For computation purposes one can derive a slightly di erent formulation of the lower bound as a power algorithm which is reminiscent of power algorithms for eigenvalues and singular values Packard. In fact. 1982]. the quantity (UM ) can have multiple local maxima which are not global.1 = inf D . 1988]). enlarging the set of scaling matrices does not improve the upper bound in inequality (11. URMR.3 and in the inequality (11. While there are open questions about convergence.4 max (MU ) = (M ). Unfortunately. ~ 11.11.2. the algorithm usually works quite well and has proven to be an e ective method to compute .12). there is no loss of generality in assuming D to be Hermitian.
This is because (M ) < 1 (exponential stability of a discrete time system matrix M ) implies for some nonsingular D 2 C n n (DMD. the scaling matrix does not need to be Hermitian. 1 7 7 7 7 7 7 7 5 1 3 D1 k .2 kn1 . F = 2 : This case was studied by Redhe er 1959]. Now let 1 3 2 6 6 6 =6 6 6 6 4 J1 = 1 . . S = 2. exponential stability is not equivalent to Lyapunov stability. (M ) = (M ). F = 1 : S = 1. It is also equivalent to the fact that Lyapunov asymptotic stability and exponential stability are equivalent for discrete time systems. F = 0 : 2 6 6 6 6 6 6 6 4 Several of the boxes have connections with standard results. kn1 . F 4 : For this case. .1 .1 7 7 7 7 7 7 7 5 2 C n1 n1 : . see Chapter 21. the matrix M can be assumed in Jordan Canonical form.. . in fact). S = 0. This is important...1 .. S = 0.2.1 ) M D DMD. . F = 1 : This is equivalent to a state space characterization of the H1 norm of a discrete time transfer function. S = 1. the upper bound is not always equal to . (M ) = (M ) = D2D DMD.) The conclusion follows by applying this result to each Jordan block. P < 0 (Lyapunov asymptotic stability). (Note that by Re!1 D1 2C n mark 11. without a loss in generality.. In fact. the bound seems to be close to .272 STRUCTURED SINGULAR VALUE S = 0. F = 0 : This is equivalent to the fact that for multidimensional systems (2d. and most systems are close to 1.. Fortunately. Then inf1 n1 (D1 J1 D1 1 ) = klim (D1 J1 D1 1 ) = j j.1 ) < 1 or (D. I < 0 which in turn is equivalent to the existence of a P = D D > 0 such that M PM . as these are the cases that arise most frequently in applications. This is a standard result in inf linear algebra. The worst known example has a ratio of over the bound of about :85.
1 (M11 ). 1 (M ) does not.4 Well Posedness and Performance for Constant LFTs Let M be a complex matrix partitioned as " M = M11 M12 M21 M22 (" # # (11:15) and suppose there are two de ned block structures. 1 and 2 . for instance.6 The linear fractional transformation Fl (M 2) is well posed . For reliable use of the theory. 2 ( ) is with respect to 2 . it is essential to have upper and lower bounds. This is exploited in the D . Of course one of the most important uses of the upper bound is as a computational scheme when combined with the lower bound. two of which were hinted at above. K iteration procedure to perform approximate synthesis (Doyle 1982]). Note that the upper bound when applied to transfer functions is simply a scaled H1 norm. The lower bound can be viewed as a natural generalization of the maximum modulus theorem. M22 2 is invertible. 11. The notations we use to keep track of these computations are as follows: 1 ( ) is with respect to 1 . Structured Singular Value 273 The above bounds are much more than just computational schemes. and. then determine how \large" Fl (M 2 ) can get for this normbounded set of perturbations. Let 2 2 2 . Another important feature of the upper bound is that it can be combined with H1 controller synthesis methods to yield an adhoc synthesis method.2. 2 (M22 ) and This section is interested in following constant matrix problems: determine whether the LFT Fl (M 2 ) is well posed for all 2 2 2 with ( 2 ) (< ). They are also theoretically rich and can unify a number of apparently quite di erent results in linear systems theory.2.11. De ne a third structure as = 0 1 0 2 : 1 2 1 22 ) 2 : (11:16) Now.4. though. Theorem 11. many major theorems in linear systems theory follow from the upper bounds and from some results of Linear Fractional Transformations. In view of these notations. (M ) all make sense. There are several connections with Lyapunov asymptotic stability. we may compute with respect to three structures. but there are further connections between the upper bound scalings and solutions to Lyapunov and Riccati equations. respectively. which will be brie y introduced in section 11. Recall that Fl (M 2 ) is well posed if I . The rst theorem is nothing more than a restatement of the de nition of . and ( ) is with respect to . Indeed. if so. which are compatible in size with M11 and M22 .
M ) becomes 1 .18) is valid. so I . 1 (Fl (M 2 )) < 1 and 1 2 B 1 . M22 2 ) det I . and de ne = diag 1 2]. det (I .M By hypothesis I . Then 2 B and. as shown in the following theorem: Theorem 11. which gives that I . M22 2 ) det (I . M ) = det (11:17) . det (I . M11 1 . ) Basically. M22 2 ) det (I .M I . M12 2 (I . with ( i ) 1. by de nition. The range of values that 1 (Fl (M 2 )) takes on is intimately related to (M ). and de ne = diag 1 2 ]. (b) for all 2 2 Bo 2 if and only if 2 (M22 ) 1. the matrix Fl (M 2 ) deviates from M11 . giving det (I . M ) = det (I .M22 2 is also nonsingular. by hypothesis. M11 . Again let 1 2 B 1 and 2 2 B 2 be given. It is easy to verify from the de nition of that (always) (M ) max f 1 (M11 ) 2 (M22 )g : We can see that 2 (M22 ) < 1. Now " # I .274 STRUCTURED SINGULAR VALUE (a) for all 2 2 B 2 if and only if 2 (M22) < 1. Fl (M 2 ) 1 must be nonsingular.M12 2 : det (I . M ) 6= 0: . We shall only prove the rst part of the equivalence. M22 2 ) 21 1 22 2 det (I . I . M ) 6= 0. and hence. (M ) < 1. The proof for the second part is similar. As the \perturbation" 2 deviates from zero.7 (MAIN LOOP THEOREM) The following are equivalent: (M ) < 1 () 8 > > < > > : 2 (M22 ) < 1 22 and 2 )) < 1: max (M ) 1 () 8 > > < > > : B 2 1 (Fl (M 2 (M22 ) 22 1 and 2 1 ( Fl ( M 2 )) max o B 1: Proof. Therefore. Fl (M 2 ) 1 ) : (11:18) But. Fl (M 2 ) 1 ) = det (I . ( Let i 2 i be given.1 M 21 1 : terms leaves det (I . Obviously 2 . Therefore. the argument above is reversed. the expression in (11. M22 Collecting the 1 2 is invertible. M is nonsingular and.
Likewise. frequency domain. and D 2 C m m be given.it is the structure that the perturbations come from however.1 C < 1: (11:20) The structured singular value of M satis es (M ) < 1: (11:21) . let be the diagonal augmentation of these two sets. which indicates that 2 Remark 11. Note that 1 ( ) appears on the right hand side of the theorem. Fl (M the claim is true. namely (" := 0m n 1 In 0n m 2 # : 1 2C m m 2 2C ) C (n+m) (n+m) : Let A 2 C n n B 2 C n m C 2 C m n .1 B < 1: (11:19) The maximum singular value of D satis es (D) < 1. D 2 ) . whether from a statespace. and j " # max 1 2C 1j 1 D + C 1 (I . Hence. the role of the perturbation structure 1 is often misunderstood. or Lyapunov approach. de ne 2 = C m m then for D 2 C m m 2 (D) = (D). 2) 1 275 is nonsingular for all i 2 B i . ~ The role of the block structure 2 in the MAIN LOOP theorem is clear .3 This theorem forms the basis for all uses of in linear system robustness analysis. Structured Singular Value Obviously.11. As an example. Now. De ne 1 := f 1 In : 1 2 C g . for A 2 C n n 1 (A) = (A). consider the theorem applied with the two simple block structures considered right after Lemma 11. so that the set 1 de nes what particular property of Fl (M 2 ) is considered. A 1 ).2. and interpret them as the state space model of a discrete time system xk+1 = Axk + Buk yk = Cxk + Duk : And let M 2 C (n+m) (n+m) be the block state space matrix of the system M= A B : C D Applying the theorem with this data gives that the following are equivalent: The spectral radius of A satis es (A) < 1. I . and 2 2C ( 2) 1 max m m A+B 2 (I .2.
1 Robust Stability The most wellknown use of as a robustness analysis tool is in the frequency domain. as in equation (11. D 2 ).1 B : 1 and The second condition implies that (I . A 1 ). and the jj jj1 norm on the transfer function from u to y is less than 1 (by replacing 1 with 1 ) z kGk1 := max D + C (zI .1 B = max 2C z2C jzj 1 j 1 1j 1 D + C 1 (I .3). Let be a block structure. but otherwise unknown. namely robust stability and inputoutput gain. Suppose G(s) is a stable. time invariant systems with one of the components normbounded. multioutput transfer function of a linear system. are also related to a particular case of the structured singular value. and the stability robustness of the uncertain di erence equation is well known.1 is well de ned for all ( 2 ) that a robust stability result holds for the uncertain di erence equation xk+1 = A + B 2 (I .1 C x k where 2 is any element in C m m with ( 2 ) 1. but otherwise unknown. We want to consider feedback perturbations to G which are themselves dynamical systems with the blockdiagonal structure of the set . Let M ( ) denote the set of all block diagonal and stable rational transfer functions that have block structures such as . assume G has q1 inputs and p1 outputs. in its necessary and su cient form for linear. M ( ) := ( ) 2 RH1 : (so ) 2 for all so 2 C + Theorem 11. What is important to note is that both of these conditions are equivalent to a condition involving the structured singular value of the state space matrix.8 Let > 0. and assume that the dimensions are such that C q1 p1 . Already we have seen that special cases of are the spectral radius and the maximum singular value.3. Here we see that other important linear system properties. This equivalence between the small gain condition.276 STRUCTURED SINGULAR VALUE The rst condition is recognized by two things: the system is stable. kGk1 < 1. For clarity. multiinput.3 Structured Robust Stability and Performance 11. D 2 ) . 11. This is the small gain theorem. A). The loop shown below is wellposed and internally stable for all ( ) 2 M ( ) with k k1 < 1 if and only if !2R sup (G(j!)) .
G(0). the peak value on the plot of the frequency response determines the size of perturbations that the loop is robustly stable against. one can nd a rational (s) such that k (s)k1 = ( c ) < 1= . This shows that sup!2R (G(j!)) < 1 is not necessary for (I . G(s) (s)).) = 0 are the complex matrices jI2 . 3 . there is a complex c 2 that each full block has rank 1 and ( c ) < 1= such that I . (G(s)) < 1 for all s 6= 0 s 2 C + .1 2 RH1 for all real rational (s) = (s)I2 with k k1 1 since (0) must be real. For example. (I .1. By Remark 11. Remark 11. using the same construction used in the proof of the small gain theorem (Theorem 9. = I2 with j j 1 for which det(I . Then sup (G(j!)) = sup jj! 1 1j = (G(j 0)) = 1: + ! 2R ! 2R On the other hand.. Similar examples with no repeated blocks are 1 generated by setting G(s) = s+1 M where M is any real matrix with (M ) = 1 for which there is no real 2 with ( ) = 1 such that det(I . G(j!o ) c is singular.e e1 G(s) + e2 ?+ w2 e Proof.1 2 RH1 with the closed ball of structured uncertainty k k1 1. : .1). Thus. and the only matrices in the form of . 2 Hence. G(s) (s)) 6= 0 for all s 2 C + f1g whenever k k1 < 1= . i. the system is robustly stable. Then there is a 0 < !o < 1 such that (G(j!o )) > . Hence det(I . (j!o ) = c . let 82 9 2 3 3 " # 0 > > 1 < = 6 7 . The following example is shown in Tits and Fan 1994]. Structured Robust Stability and Performance w1 + 6 + 277 .1. ((=) By Lemma 11. and (s) destabilizes the system.1 1 G(s) = s + 1 1 0 and = (s)I2 . sups2C + (G(s)) = sup!2R (G(j!)) . M ) = 0.e. clearly. Next. (=)) Suppose sup!2R (G(j!)) > .4 The internal stability with closed ball of uncertainties is more complicated.11. Consider " # 0 .3. G(s) (s)). 6 7 M =4 = >4 5 5 i 2 C> 2 0 .
1 2 RH1 is true only for k k1 < 0:1. G11 ). Assume Gp is a stable.278 STRUCTURED SINGULAR VALUE with 2 = 1 and 2 + 2 2 = 1.5 Let 2 RH1 be a structured uncertainty and G(s) = G11 (s) G12 (s) 2 RH1 G21 (s) G22 (s) then Fu (G ) 2 RH1 does not necessarily imply (I . G11 ).3). Partition Gp in the obvious manner Gp (s) = G11 G12 G21 G22 so that G11 has q1 inputs and p1 outputs. stability is not the only property of a closedloop system that must be robust to perturbations. Such a test will indicate the worstcase level of performance degradation associated with a given level of perturbations. the closedloop performance will degrade to the point of unacceptability. hence the need for a \robust performance" test. Let (" " # structure. the e ect that these disturbances have on error signals can greatly increase. as in equation (11. Typically. Then it is shown in Packard and Doyle 1993] that 2 (M ) = 1 and all 2 with ( ) = 1 that satisfy det(I .2 Robust Performance Often. proper transfer function with q1 + q2 inputs and p1 + p2 outputs. there are exogenous disturbances acting on the system (wind gusts. In most cases.3. Under perturbation. For example. 1 s+1 2 (k k1 < 1) but (I . sensor noise) which result in tracking and regulation errors. De ne an augmented block structure P := C q1 p1 be a block : 2 : 0 f The setup is to theoretically address the robust performance questions about the loop shown below 0 # f 2 C q2 p2 ) .1 2 RH1 whether in an open ball or is in a closed ball. ~ 1 # 11. ~ Remark 11. and so on. consider 2 " # is G(s) = 6 0 4 1 " s+1 1 s+1 10 0 0 1 0 0 3 7 5 and = admissible 1 with k k1 < 1. realrational. Then Fu (G ) = 1 2 RH1 for all 1 . long before the onset of instability. M ) = 0 must be complex.
11.3. Structured Robust Stability and Performance
 (s)
z Gp (s) w
279
The transfer function from w to z is denoted by Fu (Gp ).
Theorem 11.9 Let > 0. For all (s) 2 M ( ) with k k1 < 1 , the loop shown above is wellposed, internally stable, and kFu (Gp )k1 if and only if
!2R
sup
P
(Gp (j!))
:
Note that by internal stability, sup!2R (G11 (j!)) , then the proof of this theorem is exactly along the lines of the earlier proof for Theorem 11.8, but also appeals to Theorem 11.7. This is a remarkably useful theorem. It says that a robust performance problem is equivalent to a robust stability problem with augmented uncertainty as shown in Figure 11.5.

f
Gp (s)
Figure 11.5: Robust Performance vs Robust Stability
11.3.3 Two Block : Robust Performance Revisited
Suppose that the uncertainty block is given by
"
=
1 2
#
2 RH1
280
"
STRUCTURED SINGULAR VALUE
G(s) = G11 (s) G12 (s) 2 RH1 : G21 (s) G22 (s)
#
with k k1 < 1 and that the interconnection model G is given by
Then the closedloop system is wellposed and internally stable i sup! Let " # d! I D! = d! 2 R+ then
(G(j!)) 1.
I
; D! G(j!)D! 1 = 1G11 (j!) d! G12 (j!) : d! G21 (j! ) G22 (j! ) Hence by Theorem 11.5, at each frequency !
"
#
(G(j!)) = d inf 2R
!
"
+
Since the minimization is convex in log d! see, Doyle, 1982], the optimal d! can be found by a search however, two approximations to d! can be obtained easily by approximating the right hand side of (11.22): (1) From Lemma 2.10, we have (G(j!)) d inf ! 2R+
s q "
G11 (j!) d! G12 (j!) 1 G21 (j! ) G22 (j! ) d!
#!
:
(11:22)
kG11 (j!)k d! kG12 (j!)k 1 kG21 (j! )k kG22 (j! )k d!
!
#!
d! 2R+
inf
kG11 (j!)k2 + d2 kG12 (j!)k2 + d12 kG21 (j!)k2 + kG22 (j!)k2 !
=
kG11 (j!)k2 + kG22 (j!)k2 + 2 kG12 (j!)k kG21 (j!)k
^ d! =
8 q kG21 (j!)k > > < kG12 (j!)k > > :
with the minimizing d! given by 0
1
"
if G12 6= 0 & G21 6= 0 if G21 = 0 if G12 = 0:
(11:23)
(2) Alternative approximation can be obtained by using the Frobenius norm (G(j!)) d inf ! 2R+
G11 (j!) d! G12 (j!) 1 G21 (j! ) G22 (j! ) d!
#
F
11.3. Structured Robust Stability and Performance
s
281
= =
q
d! 2R+
inf
kG11 (j!)k2 + d2 kG12 (j!)k2 + d12 kG21 (j!)k2 + kG22 (j!)k2 ! F F F F
! F + kG22
kG11
(j!)k2
(j!)k2
F + 2 kG12 (j! )kF kG21 (j! )kF
with the minimizing d! given by ~ d! = > 0
> :
8 q kG21 (j!)kF > > < kG12 (j!)k
F
1
if G12 6= 0 & G21 6= 0 if G21 = 0 if G12 = 0:
(11:24)
It can be shown that the approximations for the scalar d! obtained above are exact for a 2 2 matrix G. For higher dimensional G, the approximations for d! are still reasonably good. Hence an approximation of can be obtained as (G(j!)) or, alternatively, as (G(j!))
" "
^ G11 (j!) d! G12 (j!) 1 G21 (j! ) G22 (j! ) ^ d! ~ G11 (j!) d! G12 (j!) 1 G21 (j! ) G22 (j! ) ~ d!
#!
(11:25)
#!
:
(11:26)
We can now see how these approximated conditions obtained in Chapter 9.
tests are compared with the su cient
Example 11.1 Consider again the robust performance problem of a system with output multiplicative uncertainty in Chapter 9 (see Figure 9.5):
P = (I + W1 W2 )P k k1 < 1:
Then it is easy to show that the problem can be put in the general framework by selecting " #
G(s) = ;W2 To W1 ;W2 ToWd We So W1 We So Wd
and that the robust performance condition is satis ed if and only if
kW2 ToW1 k1 1
and
(11:27) (11:28)
kFu (G )k1 1
282
STRUCTURED SINGULAR VALUE
for all 2 RH1 with k k1 < 1. But (11.27) and (11.28) are satis ed i for each frequency ! (G(j!)) = d inf 2R
!
"
+
d! We So W1
;W2 ToW1 ;d! W2 To Wd
1
#!
We So Wd
1:
Note that, in contrast to the su cient condition obtained in Chapter 9, this condition is an exact test for robust performance. To compare the test with the criteria obtained in Chapter 9, some upper bounds for can be derived. Let
W d! = kWe SoW1 k : kW2 To d k
Then, using the rst approximation for , we get (G(j!))
s
kW2 To W1 k2 + kWe So Wd k2 + 2 kW2 ToWd k kWe So W1 k kW2 To W1 k2 + kWe So Wd k2 + 2 (W1;1 Wd ) kW2 ToW1 k kWe SoWd k kW2 To W1 k + (W1;1 Wd ) kWe So Wd k
q
q
where W1 is assumed to be invertible in the last two inequalities. The last term is exactly the su cient robust performance criteria obtained in Chapter 9. It is clear that any term preceding the last forms a tighter test since (W1;1 Wd ) 1. Yet another alternative su cient test can be obtained from the above sequence of inequalities: (G(j!))
q
(W1;1 Wd )(kW2 ToW1 k + kWe SoWd k):
Note that this su cient condition is not easy to get from the approach taken in Chapter 9 and is potentially less conservative than the bounds derived there. 3 Next we consider the skewed speci cation problem, but rst the following lemma is needed in the sequel.
Lemma 11.10 Suppose =
1
2
:::
m=
> 0, then
+
inf max (d i )2 + (d 1 )2 = d2R+ i i
:
Proof. Consider a function y = x + 1=x, then y is a convex function and the maximization over a closed interval is achieved at the boundary of the interval. Hence for any xed d max (d i )2 + (d 1 )2 = max (d )2 + (d1 )2 (d )2 + (d1 )2 : i i
11.3. Structured Robust Stability and Performance
Then the minimization over d is obtained i (d )2 + (d1 )2 = (d )2 + (d1 )2 which gives d2 =
1
283
. The result then follows from substituting d.
2
Example 11.2 As another example, consider again the skewed speci cation problem
from Chapter 9. Then the corresponding G matrix is given by
G = ;W2 Ti W1 ;W2 KSoWd : We So PW1
"
"
#
We So Wd
So the robust performance speci cation is satis ed i (G(j!)) = d inf 2R
!
+
d! We So PW1
s
1
;W2 Ti W1
;d! W2 KSoWd
We So Wd
#!
1
for all ! 0. As in the last example, an upper bound can be obtained by taking
kW So PW d! = kW eKS W1 k : 2 o dk
Then (Wd;1 PW1 )(kW2 Ti W1 k + kWe So Wd k): (G(j!)) In particular, this suggests that the robust performance margin is inversely proportional to the square root of the plant condition number if Wd = I and W1 = I . This can be further illustrated by considering a plantinverting control system. To simplify the exposition, we shall make the following assumptions:
q
We = ws I Wd = I W1 = I W2 = wt I and P is stable and has a stable inverse (i.e., minimum phase) (P can be strictly proper).
Furthermore, we shall assume that the controller has the form
K (s) = P ;1 (s)l(s)
where l(s) is a scalar loop transfer function which makes K (s) proper and stabilizes the closedloop. This compensator produces diagonal sensitivity and complementary sensitivity functions with identical diagonal elements, namely l( So = Si = 1 +1l(s) I To = Ti = 1 + s)s) I: l(
284 Denote
STRUCTURED SINGULAR VALUE
"(s) = 1 +1l(s)
"
and substitute these expressions into G we get
l( (s) = 1 + s)s) l(
#
;1 G = ;wt I ;wt P :
ws "P
"
ws "I
The structured singular value for G at frequency ! can be computed by (G(j!)) = dinf 2R
;wt I ;wt (dP );1
ws "dP ws "I
#!
+
:
Let the singular value decomposition of P (j!) at frequency ! be
P (j!) = U V = diag( 1 2 : : : m ) with 1 = and m = where m is the dimension of P . Then
(G(j!)) = dinf 2R
" "
+
;wt I ;wt (d );1
ws "d ws "I
#!
since unitary operations do not change the singular values of a matrix. Note that
;wt I ;wt (d );1 = P diag(M M : : : M )P 1 1 2 m 2
ws "d ws "I
"
#
where P1 and P2 are permutation matrices and where
Mi =
Hence
ws "d i
" "
;wt
;wt (d i );1 :
ws "
#
(G(j!)) = dinf max 2R+ i
2 = dinf max jws "j2 + jwt j2 + jws "d i j2 + wt : 2R+ i di
ws "d i ws " # ! ;wt h 1 (d );1 i = dinf max i 2R+ i ws "d i p = dinf max (1 + jd i j;2 )(jws "d i j2 + jwt j2 ) 2R i
+
s
;wt
;wt (d i );1
#!
11.3. Structured Robust Stability and Performance
285
Using Lemma 11.10, it is easy to show that the maximum is achieved at either or and that optimal d is given by j t d2 = jww"j j so the structured singular value is
s
s
(G(j!)) = jws "j2 + jwt j2 + jws "jjwt j (P ) + (1 ) ]: P
(11:29)
Note that if jws "j and jwt j are not too large, which are guaranteed if the nominal performance and robust stability conditions are satis ed, then the structured singular value is proportional to the square root of the plant condition number: (G(j!))
p
jws "jjwt j (P ) :
(11:30)
3
This con rms our intuition that an illconditioned plant with skewed speci cations is hard to control.
11.3.4 Approximation of Multiple Full Block
= diag( so that
2 6 6 6 6 4
The approximations given in the last subsection can be generalized to the multiple block problem by assuming that M is partitioned consistently with the structure of
1 2
::: F) M1F M2F MFF
.. .
3 7 7 7 7 5
M=
and Now And hence (M )
M11 M12 M21 M22 MF 1 MF 2
.. . .. .
D = diag(d1 I : : : dF ;1 I I ): d DMD;1 = Mij di j
D2D
dF := 1: d Mij d i
j
inf (DMD;1 ) = D2D inf
286 inf D2D inf D2D
STRUCTURED SINGULAR VALUE
d kMij k di j
inf D2D
2
v u F F uX X t
v u F F uX X t
i=1 j =1
d kMij k2 di2
2
j
i=1 j =1
kMij k2 di2 : Fd
j
An approximate D can be found by solving the following minimization problem: inf D2D
F F XX i=1 j =1 F F XX i=1 j =1
2 2
di kMij k2 d2
j
2
or, more conveniently, by minimizing inf D2D
P P
kMij k2 di2 Fd
j
2
with dF = 1. The optimal di minimizing of the above two problems satisfy, respectively,
d4
and
k=
i6=k kMik k di 2 2 j 6=k kMkj k =dj
k = 1 2 ::: F ;1
(11:31) (11:32)
d4
i6=k k=P
P
Using these relations, dk can be obtained by iterations.
kMik k2 d2 F i k = 1 2 : : : F ; 1: kMkj k2 =d2 j 6=k F j
3 7 5
Example 11.3 Consider a 3 3 complex matrix 2 1 + j 10 ; 2j ;20j 6 M = 4 5j 3 + j ;1 + 3j ;2 j 4;j
with structured = diag( 1 2 3 ). The largest singular value of M is (M ) = 22:9094 and the structured singular value of M computed using the TOOLS is equal to its upper bound: (M ) = D2D (DMD;1 ) = 11:9636 inf with the optimal scaling Dopt = diag(0:3955 0:6847 1). The optimal D minimizing inf D2D
F F XX i=1 j =1
di kMij k2 d2
j
2
11.4. Overview on Synthesis
287
is Dsubopt = diag(0:3212 0:4643 1) which is solved from equation (11.31). Using this Dsubopt , we obtain another upper bound for the structured singular value: (M )
;1 (Dsubopt MDsubopt ) = 12:2538:
One may also use this Dsubopt as an initial guess for the exact optimization.
3
11.4 Overview on Synthesis
This section brie y outlines various synthesis methods. The details are somewhat complicated and are treated in the other parts of the book. At this point, we simply want to point out how the analysis theory discussed in the previous sections leads naturally to synthesis questions. From the analysis results, we see that each case eventually leads to the evaluation of kM k = 2 1 or (11:33) for some transfer matrix M . Thus when the controller is put back into the problem, it involves only a simple linear fractional transformation as shown in Figure 11.6 with
G11 G12 where G = G21 G22
"
M = F` (G K ) = G11 + G12 K (I ; G22 K );1 G21 :
#
is chosen, respectively, as
" #
P P nominal performance only ( = 0): G = 22 23 P32 P33 P P robust stability only: G = 11 13 P31 P33
2 " #
P11 P12 P13 robust performance: G = P = 6 P21 P22 P23 7. 5 4 P31 P32 P33
Each case then leads to the synthesis problem min kF` (G K )k K for = 2 1 or (11:34) which is subject to the internal stability of the nominal. The solutions of these problems for = 2 and 1 are the focus of the rest of this book. The solutions presented in this book unify the two approaches in a common synthesis framework. The = 2 case was already known in the 1960's, and the results
3
288
STRUCTURED SINGULAR VALUE
z G w
K
Figure 11.6: Synthesis Framework are simply a new interpretation. The two Riccati solutions for the = 1 case were new products of the late 1980's. The synthesis for the = case is not yet fully solved. Recall that may be obtained by scaling and applying k k1 (for F 3 and S = 0), a reasonable approach is to \solve" min ;1 inf DF`(G K )D;1 1 (11:35) K by iteratively solving for K and D. This is the socalled DK Iteration. The stable and minimum phase scaling matrix D(s) is chosen such that D(s) (s) = (s)D(s) (Note that D(s) is not necessarily belong to D since D(s) is not necessarily Hermitian, see Remark 11.2). For a xed scaling transfer matrix D, minK DF` (G K )D;1 1 is a standard H1 optimization problem which will be solved in the later part of the book. For a given stabilizing controller K , inf D D;1 2H1 DF`(G K )D;1 1 is a standard convex optimization problem and it can be solved pointwise in the frequency domain:
; sup Dinf D! F` (G K )(j!)D! 1 : 2D
!
!
D D 2H1
Indeed,
D D;1 2H1
inf
; DF` (G K )D;1 1 = sup Dinf D! F`(G K )(j!)D! 1 : ! ! 2D
This follows intuitively from the following arguments: the left hand side is always no smaller than the right hand side, and, on the other hand, given the minimizing D! from the right hand side across the frequency, there is always a rational function D(s) uniformly approximating the magnitude frequency response D! . Note that when S = 0, (no scalar blocks)
D! = diag(d! I : : : d! ;1 I I ) 2 D 1 F
which is a blockdiagonal scaling matrix applied pointwise across frequency to the frequency response F` (G K )(j!). DK Iterations proceed by performing this twoparameter minimization in sequential fashion: rst minimizing over K with D! xed, then minimizing pointwise over D! with K xed, then again over K , and again over D! , etc. Details of this process are summarized in the following steps:
11.4. Overview on Synthesis
D G D;1
289
 K
Figure 11.7: Synthesis via Scaling (i) Fix an initial estimate of the scaling matrix D! 2 D pointwise across frequency. (ii) Find scalar transfer functions di (s) d;1 (s) 2 RH1 for i = 1 2 : : : (F ; 1) such i that jdi (j!)j d! . This step can be done using the interpolation theory Youla i and Saito, 1967] however, this will usually result in very high order transfer functions, which explains why this process is currently done mostly by graphical matching using lower order transfer functions. (iii) Let D(s) = diag(d1 (s)I : : : dF ;1 (s)I I ): Construct a state space model for system ^ G(s) = D(s)
" #
I
;1 G(s) D (s)
"
#
I
:
(iv) Solve an H1 optimization problem to minimize ^ F` (G K ) over all stabilizing K 's. Note that this optimization problem uses the scaled ^ version of G. Let its minimizing controller be denoted by K . ^ ; (v) Minimize D! F` (G K )D! 1 ] over D! , pointwise across frequency.4 Note that ^ this evaluation uses the minimizing K from the last step, but that G is unscaled. The minimization itself produces a new scaling function. Let this new function ^ be denoted by D! . ^ (vi) Compare D! with the previous estimate D! . Stop if they are close, but, otherwise, ^ replace D! with D! and return to step (ii). With either K or D xed, the global optimum in the other variable may be found using the and H1 solutions. Although the joint optimization of D and K is not convex and
4
1
The approximate solutions given in the last section may be used.
290
STRUCTURED SINGULAR VALUE
the global convergence is not guaranteed, many designs have shown that this approach works very well see e.g. Balas, 1990]. In fact, this is probably the most e ective design methodology available today for dealing with such complicated problems. The detailed treatment of analysis is given in Packard and Doyle 1991]. The rest of this book will focus on the H1 optimization which is a fundamental tool for synthesis.
11.5 Notes and References
This chapter is partially based on the lecture notes given by Doyle 1984] in Honeywell and partially based on the lecture notes by Packard 1991] and the paper by Doyle, Packard, and Zhou 1991]. Parts of section 11.3.3 come from the paper by Stein and Doyle 1991]. The small theorem for systems with nonrational plants and uncertainties is proven in Tits 1994]. Other results on can be found in Fan and Tits 1986], Fan, Tits, and Doyle 1991], Packard and Doyle 1993], Packard and Pandey 1993], Young 1993], and references therein.
full control (FC) problems and output estimation 291 12 . the coprime factorization technique is used. which are socalled full information (FI) problems. K is the controller to be designed. where G is the generalized plant with two sets of inputs: the exogenous inputs w.. of the closedloop or to design the feedback control K such that the closedloop system is stable in some appropriate sense and the error signal z is speci ed.g. All of the existing results are mainly in the frequency domain although they can also be transformed to statespace descriptions. A control problem in this setup is either to analyze some speci c properties. The construction of the controller parameterization is done via considering a sequence of special problems. and control inputs u. The plant G also has two sets of outputs: the measured (or sensor) outputs y and the regulated outputs z . In this chapter.Parameterization of Stabilizing Controllers The basic con guration of the feedback systems considered in this chapter is an LFT as shown in Figure 12. Suppose that a given feedback system is feedback stabilizable. stability or performance. which include disturbances and commands. The parameterization of all internally stabilizing controllers was rst introduced by Youla et al 1976] in their parameterization. disturbance feedforward (DF) problems. the problem we are mostly interested in is parameterizing all controllers that stabilize the system..e. we consider this issue in the general setting and directly in state space without adopting coprime factorization technique. In this chapter. some performance condition is satis ed. e.1. i. We will see again that this setup is very convenient for other general control synthesis problems in the coming chapters. In this chapter we are only concerned with the basic internal stabilization problems.
K Figure 12. the construction of controller parameterization for the general output feedback problem will be considered via the special problems FI . which is essential in some control syntheses (the H2 and H1 control in Chapter 14 and Chapter 16). . This section can be either studied independently of all the preceding sections or skipped over without loss of continuity.292 PARAMETERIZATION OF STABILIZING CONTROLLERS z y G w u .6 considers an alternative approach to the controller parameterization using coprime factorizations and establishes the connections with the state space approach. In section 12. The structure of this chapter is as follows: in section 12. Section 12.1. the structure of the controller parameterization is displayed. this chapter also aims at introducing the synthesis machinery. DF .2. This general synthesis question will be called the output feedback (OF) problem. a proper controller K (s) is said to be admissible if it internally stabilizes G. A B1 B2 G11 (s) G12 (s) = 6 7 G(s) = (12:1) 4 C1 D11 D12 5 : G21 (s) G22 (s) C2 D21 D22 The stabilization problem is to nd feedback mapping K such that the closedloop " # 2 3 De nition 12. we shall examine the stabilization of di erent special problems and establish the relations among them. 12.5 shows the closedloop transfer matrix in terms of the parameterization.1: General System Interconnection (OE) problems. On the other hand. the conditions for the existence of a stabilizing controller are examined. Assume that G(s) has a stabilizable and detectable realization of the form system is internally stable the wellposedness is required for this interconnection. In section 12. these special problems are also of importance in their own right.1. FC and OE . In addition to presenting the controller parameterization.1. and at seeing how it works in the controller parameterization problem.1 Existence of Stabilizing Controllers Consider a system described by the standard block diagram in Figure 12.1 A proper system G is said to be stabilizable through output feedback if there exists a proper controller K internally stabilizing G in Figure 12. Moreover.3. In section 12. Section 12.4.
The details are left as an exercise. 293 Lemma 12. .LC2 A + B2 F + LC2 A + LC2 0 : . no matter what the compensator is. A ()) If (A B2 ) is not stabilizable or if (C2 A) is not detectable. then there are some ~ eigenvalues of A which are xed in the right halfplane. then an observerbased stabilizing controller is given by K (s) = A + B2 F + LC2 + LD22 F .2: Equivalent Stabilization Diagram 1 It should be clear that the stabilizability and detectability of a realization for G do not guarantee the stabilizability and/or detectability of the corresponding realization for G22 . let F and L be such that A + B2 F and A + LC2 are stable. then the closedloop Amatrix is given by ~ A= " " # Proof. In ~ is stable. (() By the stabilizability and detectability assumptions.1.L : F 0 L such that A + B2 F and A + LC2 are stable. there exist F and A B2 F : . particular. It follows that the realization for G22 is stabilizable and detectable. Existence of Stabilizing Controllers The following result is standard and follows from Chapter 3. 2 The stabilizability and detectability conditions of (A B2 C2 ) are assumed throughout the remainder of this chapter1. Further.K Figure 12. and these assumptions are enough to yield the following result: y G22 u .1 There exists a proper K achieving internal stability i (A B2) is stabilizable and (C2 A) is detectable. Now let K (s) be the observerbased controller given in the lemma.LC2 A + B2 F # # It is easy to check that this matrix is similar to the matrix " ~ Thus the spectrum of A equals the union of the spectra of A + LC2 and A + B2 F .12.
matrix G22 and the inherited realization from G where G22 is a submatrix.1 0 0 1 1 0 0 0 1 = s+1 3 7 7 7 7 5 is 2 G22 = 6 0 1 0 4 1 0 0 which is neither stabilizable nor detectable. A given realization for G22 may be stabilizable and detectable while the inherited realization may be not. G22 .2 is internally stable. 1 G22 = s + 1 = .1 1 1 0 2 " # " # Remark 12. For instance.1 There should be no confusion between a given realization for a transfer is a minimal realization but the inherited realization of G22 from 6 G11 G12 = 6 0 1 1 0 6 6 G21 G22 1 0 0 4 0 .1 is internally stable i the one in Figure 12. it is su cient to show that the system in Figure 12. To show the su ciency.1 and that in Figure 12.2 Suppose (A B2 C2 ) is stabilizable and detectable and G22 = A B2 . which is obvious.1 0 1 3 7 5 ~ . it will be shown that the general setup is very convenient and much more useful since any closedloop system information can also be considered in the same framework. K (s) internally stabilizes G(s) if and only if it internally stabilizes Proof.2. which is how it is usually done in the conventional Youla parameterization. Hence all stabilizing controllers for G can be obtained by using only G22 . In other words. However. .2 share the same Amatrix. we see that the stabilizing controller for G depends only on G22 . 2 From Lemma 12.294 PARAMETERIZATION OF STABILIZING CONTROLLERS " # Lemma 12. C2 D22 Then the system in Figure 12. The necessity follows from the de nition.
12. but to di erent structures than G. These special problems are fundamental to the approach taken for synthesis in this book. we now introduce the concept of algebraic duality which will play an important role. and. It is not di cult to see that K internally ~~ stabilizes G i K T internally stabilizes GT . Now. Consider a standard system block diagram z G w y . The problems are labeled as . GT and K T which are dual objects of G and K . Duality and Special Problems 295 12.2. to deal with the issues related to a system's controllability and/or observability. The special problems considered here all pertain to the standard block diagram.KT u ~ whose plant and controller are obtained by transposing G and K . It is well known that the concepts of controllability (stabilizability) and observability (detectability) of a system (C A B ) are dual because of the duality between (C A B ) and (B T AT C T ). Now consider another system shown below z ~ w ~ GT y ~ .1 Algebraic Duality and Special Problems Before we get into the study of the algebraic structure of control systems. we will discuss four problems from which the output feedback solutions are constructed via a separation argument. The notion of duality can be generalized to a general setting. 12. we consider some special problems which are related to the general OF problems stated in the last section and which are important in constructing the results for OF problems. we only need to examine the issues related to the observability and/or controllability of its dual system. We can check easily T that Tzw = F`(G K )]T = F`(GT K T ) = Tzw . So as far as stabilization or other synthesis problems are concerned. we can obtain the results for GT from the results for its dual object G if they are available. especially. So.2 Duality and Special Problems In this section. respectively. respectively. they are also of importance in their own right.2.K u where the plant G and controller K are assumed to be linear time invariant. as we shall see. And we say that these two control structures are algebraically dual.
There are also two additional structures which are standard and which will not be considered in this chapter they are SF. Disturbance feedforward. Full information. with the corresponding plant 3 A B1 B2 6 GOE = 4 C1 D11 I 7 : 5 C2 D21 0 The motivations for these special problems will be given later when they are considered. Full control. with the corresponding plant A B1 6 GFC = 6 C1 D11 6 4 C2 D21 2 2 h h h I 0 0 I 0 0 i 3 i 7 7 7 i 5 : DF. Output estimation. State feedback 2 3 A B1 B2 6 GSF = 4 C1 D11 D12 7 : 5 I 0 0 A B1 I 6 GOI = 4 C1 D11 0 7 : 5 C2 D21 0 2 3 2 3 OI. with the corresponding plant B1 D " 11# I 0 B2 7 D 7 " 12# 7 : 7 0 0 5 3 FC.296 PARAMETERIZATION OF STABILIZING CONTROLLERS A 6 6 C1 GFI = 6 " # 6 I 4 0 2 FI. Output injection . with the corresponding plant A B1 B2 GDF = 6 C1 D11 D12 7 : 5 4 C2 I 0 OE.
Duality and Special Problems 297 Here we assume that all physical variables have compatible dimensions.KFI u yDF . Then there is no feedback and the measurement is exactly w. consider the special case with C2 = 0.FC 6 equivalent . Only the disturbance. Now we examine the connection between the DF problem and the FI problem and show the meaning of their equivalence. As we shall see. In fact. The structure of transfer matrices shows clearly that FC.2. but it does a ect internal stability. provided FI is of the same structure with OF except for the speci ed parts. These relationships are shown in the following diagram: FI dual 6 equivalent . the conditions for the solvability of the DF problem are that (A B2 ) is stabilizable and (C2 A) is detectable. DF (and SF). It is clear that if any output feedback control problem is to be solvable then the corresponding FI problem has to be solvable. the controller is provided with Full Information since y = w For the FI problem. the feedback caused by C2 6= 0 does not a ect the transfer function from w to the output z . We say that these special problems are special cases of OF problems in the sense that their structures are speci ed in comparison to OF problems. Suppose that we have controllers KFI and KDF and let TFI and TDF denote the closedloop Tzw s in z GFI w z GDF w yFI .KDF u . where w is generally regarded as disturbance to the system.2 Full Information and Disturbance Feedforward x . In the FI problem.OE DF The precise meaning of \equivalent" in this diagram will be explained below. is fed through directly to the output. w. ? dual ? 12.2.12. we only need to assume that (A B2 ) is stabilizable to guarantee the " # solvability. OE (and OI) are duals of FI. To motivate the name Disturbance Feedforward. respectively.
2 There is an alternative way to see part (ii). The fact is that in the DF A . C2 0 B1 " # I 0 0 # B2 7 I # 7: 7 " 7 0 0 5 3 Proof.C2 I 0 Now connect V up with GDF as shown below " # 2 3 7 5 " problems. as claimed. Part (i) is obvious. we will give a direct proof to show the system structure. x and x as the states of ^ ^ the resulting interconnected system then its realization is 2 6 6 6 6 6 6 4 A . (i) GDF (s) = FI 0 C2 I (ii) GFI = S (GDF PDF ) (where S ( ) denotes the starproduct) GDF aa !! ! a !!aa ! a PDF 2 PDF (s) = 6 6 4 6 6 A . the disturbance w and the system states x can be solved in terms of y and u: y =: V u # " y : u # . B1 C2 " I . B1 C2 B1 B2 x =6 I 0 0 4 w . Nevertheless. Then " # I 0 0 G (s).298 PARAMETERIZATION OF STABILIZING CONTROLLERS The question of interest is as follows: given either the KFI or the KDF controller. Let x and x ^ denote the state of GDF and PDF .3 Let GFI and GDF be given as above. Part (ii) follows from applying the star product formula. B1 C2 B1 C2 C " 1 # C2 0 I 0 which is exactly GFI . Take e := x . Actually. we have the following: Lemma 12. can we construct the other in such a way that TFI = TDF ? The answer is positive. A C " 1# I 0 B1 D " 11# 0 0 7 2 7 7 12# 7 7 " 0 7 5 B D 0 3 A 6 6 C1 =6 " # 6 I 4 0 2 B1 D " 11# I 0 B2 D " 12# 0 0 3 7 7 7 7 5 0 2 Remark 12. respectively.
then problems FI and G21 (s) = A B1 : C2 I " # DF are equivalent in the above sense. and PDF be given as above. ~ The following theorem follows immediately: Theorem 12. that the internal stability is not changed if A .4 Let GFI . Duality and Special Problems z y x w GDF V w u 299 Then it is easy to show that the transfer function from (w u) to (z x w) is GFI .KFI internally stabilizes GDF if KFI internally stabilizes GFI . B1 C2 is stable.12. h i internally stabilizes GFI if KDF internally stabilizes h i F`(GFI KDF C2 I ) = F` (GDF KDF ): (ii) Suppose A . ~ . Note that the transfer function from w to yDF is Hence this stability condition implies that G21 (s) has neither right half plane invariant zeros nor hidden unstable modes. (i) KFI := KDF C2 I GDF . and.3 This theorem shows that if A . B1C2 is stable. GDF . furthermore. B1 C2 is stable. Furthermore. F`(GDF F` (PDF KFI )) = F` (GFI KFI ): Remark 12. Then KDF := F` (PDF KFI ) as shown below u y ^ PDF yDF u .2. Furthermore.
The following discussion will explain the meaning of equivalence between FC and OE problems. To put it in other words.KFC . In general. The solutions to this kind of control problem can be obtained by rst transposing GFC .300 PARAMETERIZATION OF STABILIZING CONTROLLERS 12. To examine the physical meaning of output estimation.2. the term Full Control is used because the controller has full access to both the state through output injection and the output z . rst note that z = C1 x + D11 w + u where z is to be controlled by an appropriately designed control u. which is also guaranteed by the assumptions on OF problems. So this kind of control problem can be regarded as an estimation problem. This special case motivates the term output estimation and can be obtained immediately from the results obtained for the general case. when appropriately dualized. our control objective will be to specify z in some wellde ned mathematical sense. Then I 0 6 (i) GOE (s) = GFC (s) 4 0 B2 0 I 2 3 7 5 . problem OE is dual to DF . and then transposing back. The only restriction on the controller is that it must work with the measurement y. A more conventional estimation problem would be the special case where no internal stability condition is imposed and B2 = 0. We are focusing on this particular estimation problem because it is the one that arises in solving the output feedback problem. On the other hand.5 Let GFC and GOE be given as above. and solving the corresponding FI problem. Consider the following FC and OE diagrams: z yFC GFC w u z yOE GOE w u . This problem is dual to the FI case and has the dual solvability condition to the FI problem.3 Full Control and Output Estimation For the FC problem. it is desired to nd a u that will estimate C1 x + D11 w in such de ned mathematical sense. Then the problem would be that of estimating the output z given the measurement y.KOE We have similar results to the ones in the last subsection: Lemma 12. And the solvability conditions for the OE problem are that (A B2 ) is stabilizable and (C2 A) is detectable. Thus the discussion of the DF problem is relevant here.
1 with " # A B1 B2 G11 (s) G12 (s) : 7 6 G(s) = 4 C1 D11 D12 5 = G21 (s) G22 (s) C2 D21 D22 Suppose (A B2 ) is stabilizable and (C2 A) is detectable.6 Let GFC GOE . In this section we discuss the following problem: . B2 C1 is stable. F` (GOE F`(POE KFC )) = F`(GFC KFC ): Remark 12.4 It is seen that if A . where POE is given by 301 i 3 7 7 7 5 A . Then KOE := F` (POE KFC ). Furthermore. (i) KFC := OE FC OE OE I Furthermore. which indicates that it has a stable inverse. " # F` (GFC B2 KOE ) = F`(GOE KOE ): I (ii) Suppose A .3.B2i h 0 I i h 0 0 : Theorem 12. then FC and OE problems are equivalent in the above sense.KFC u ^ internally stabilizes GOE if KFC internally stabilizes GFC .1 Problem Statement and Solution 2 3 Consider again the standard system block diagram in Figure 12. B2 C1 0 6 6 POE (s) = 6 C1 0 4 C2 I " # 2 h I . This condition implies that the transfer matrix G12 (s) from u to z has neither right halfplane invariant zeros nor hidden unstable modes. and POE be given as above. ~ 12. B2C1 is stable. B2 K internally stabilizes G if K internally stabilizes G . as shown below y u POE y ^ .12.3 Parameterization of All Stabilizing Controllers 12. Parameterization of All Stabilizing Controllers (ii) GFC = S (GOE POE ).3.
A nonconstructive proof of the theorem can be given by using the same argument as in the proof of Theorem 12.302 PARAMETERIZATION OF STABILIZING CONTROLLERS Given a plant G.1 is stable.Q with any Q 2 RH1 and I + D22 Q(1) nonsingular.(C2 + D22 F ) I . which says that a controller K stabilizes a stable plant G22 i K (I . On the other hand.5 This result is very natural considering Corollary 5.. The results can be more conveniently stated using state space representations. G22 K ). and then show that we can construct a stable Q for any . i. As we have mentioned early. suppose K0 is a stabilizing controller then Q0 := K0(I .D22 2 controllers that internally stabilize G can be parameterized as the transfer matrix from y to u below J 3 7 5 .1 for any Q 2 RH1 and I + D22 Q(1) nonsingular.5.e.7.L B2 + LD22 6 J =4 F 0 I . parameterize all controllers K that internally stabilize G. ~ Proof. However. The parameterization of all stabilizing controllers is easy when the plant itself is stable. D22 K0 (1)). 2 However.8 Let F and L be such that A + LC2 and A + B2F are stable. Note that G22 (s) is stable by the assumptions on G. Theorem 12. This parameterization for all stabilizing controllers is usually called Youla parameterization.1 . Theorem 12. the parameterization is much more complicated.1 . the stabilizing controllers for G will depend only on G22 . it is more convenient to consider the problem in the general framework as will be shown. Now suppose Q = K (I .1 2 RH1 . G22 K ). rst verify that any controller given by the formula indeed stabilizes the system G. (12:2) Remark 12.7 Suppose G 2 RH1 then the set of all stabilizing controllers can be described as K = Q(I + G22 Q). G22 K0 ). Note that the invertibility in the last equation is guaranteed by the well posedness of the interconnected system with controller K0 since I + D22 Q0 (1) = (I . Now use straightforward algebra to verify that the controllers given above stabilize G22 . if G is not stable.1 is a stable transfer matrix. then K can be solved from this equation which gives exactly the controller parameterization in the above theorem. so K0 can be expressed as K0 = Q0 (I + G22 Q0 ). and then all u y A + B2 F + LC2 + LD22F .
12. Instead. Readers should bear in mind that our objective here is to nd all admissible controllers for the OF problem. In the next two subsections. the controller equivalence is an equivalence relation. The advantages of this approach are that it is simple and that many other synthesis problems. ^ De nition 12. We will see that it is not necessary to parameterize all stabilizing controllers for these special problems to get the required tools.12. The purpose of this subsection is to characterize the equivalent classes of stabilizing controllers KFI that stabilize . which will be adopted in the later part of this chapter as an alternative approach.2 Stabilizing Controllers for FI and FC Problems In this subsection. we mainly consider the stabilizing controller characterizations for special problems. We will also see that the characterizations of equivalent classes of stabilizing controllers for special problems are good enough to construct the parameterization of all stabilizing controllers for the OF problem. The conventional Youla approach to this problem is via coprime factorization Youla et al. and then to solve the output feedback problem by the separation argument. does not give much insight into the controller construction and thus can not be generalized to other synthesis problems. Also. 1982]. we only parameterize some equivalent classes of controllers which generate the same control action.3. 1976. We will see that for di erent special problems we have di erent re ned versions of this relation. such as H2 and H1 optimal control problems in Chapters 14 and 16. we will present a novel approach to this problem without adopting coprime factorizations. we will try to build up enough tools for this objective by considering the special problems. we use the solutions to these special problems and the approach provided in the last section to characterize all stabilizing controllers of OF problems. Parameterization of All Stabilizing Controllers 303 given stabilizing controller K . i. In the following sections. The idea of this approach is to reduce the output feedback problem into some simpler problems.e. This approach. ^.KFI where the transfer matrix GFI is given in section 12. Fl (G K ) = Fl (G K ). So at rst. Desoer et al.3. Vidyasagar.2. 1985. however. written as K = K Algebraically. we rst examine the FI structure z yFI GFI w u . such as FI and OE or FC and DF which admit simple solutions.2 Two controllers K and K are of equivalent control actions if their ^ corresponding closed loop transfer matrices are identical. can be solved by using the same machinery.
Note that the stability of K2 is required to guarantee the internal stability of the whole system since w is fed directly through K2. To show this. Note that for the parameter matrix Q 2 RH1 .e. It is easy to see that the controller given in the above formula stabilizes the system GFI . u. for FI can be parameterized as KFI = F Q h i with any Q 2 RH1 .304 PARAMETERIZATION OF STABILIZING CONTROLLERS internally GFI and to build up enough tools to be used later. Obviously. where x denotes the state of the system GFI then the system with the controller KFI will be as shown in the following diagram: z ~ GFI w yFI with . in the sense of generating all admissible closedloop transfer functions. make a change of control variable as v = u . we ^ say two controllers KFI and KFI are equivalent if they produce the same closedloop transfer function from w to u. A). i. there is a choice of Q 2 RH1 such that the transfer h i functions from w to u for any stabilizing controller KFI = K1 (s) K2 (s) and for h i 0 KFI = F Q are the same. Since we have full information for feedback.9 Let F be a constant matrix such that A + B2F is stable. Hence we only need to show that the given set of controllers parameterizes all stabilizing control action. this also guarantees that Fl (GFI KFI ) = ^ Fl (GFI KFI ). Proof. Then all stabilizing controllers. our controller will have the following general form: h i KFI = K1 (s) K2(s) with K1 (s) stabilizing internally (sI . F 0]: . For this problem. Fx. Lemma 12. Note that the same situation will occur in FC problems by duality..1 B2 and arbitrary K2 (s) 2 RH1 .KFI ~ B2 D12# " 0 0 v 3 7 7 7 7 5 A + B2 F 6 6 C1 + D12 F ~ GFI = 6 " # 6 I 4 0 2 B1 D11# " I 0 ~ KFI = KFI . it is reasonable to assume that the realization of Q(s) is stabilizable and detectable. The characterization of equivalent classes of controllers can be suitably called the control parameterization in contrast with controller parameterization.
1 Q respectively. This also guarantees the identity of their resulting closedloop transfer matrices from w to z .KFC ^ Dually.3) and (12.1 B1 + I .1 B2 .6 The equivalence of KFI = KFI in the above equation can actually be shown by directly computing Q from equating the transfer matrices from w to u for the 0 0 cases of KFI and KFI . A). K1 (sI .10 Let L be a constant matrix such that A + LC2 is stable.1 B2 .3. Then the KFC = L Q set of equivalent classes of all stabilizing controllers for FC in the above sense can be parameterized as " # with any Q 2 RH1 . A). B2 K1 ).12. F )(sI . the transfer matrices from w to u with KFI and KFI are given by I . A).1 F (sI . In fact.1 K1(sI . A). F (sI .4) equal.1 B2 .1 (B2 K2 + B1 ) is stable and makes the formulas in (12. 2 0 Remark 12. A). A). ~ Now we consider the dual FC problem the system diagram pertinent to this case is z yFC GFC w u . A .1 K2 and (12:3) (12:4) I . We can verify that Q = K2 + (K1 . i Then u = Fx + v = Fx + Qw.1 B2 . we say controllers KFC and KFC are equivalent in the sense that the same injection inputs yFC 's produce the same outputs z 's. And we also have Lemma 12. . so KFI = F Q . Parameterization of All Stabilizing Controllers 305 Let Q be the transfer matrix from w to hv Q belongs to RH1 by internal stability. F (sI .1 B1 + I . K1(sI .
it does simplify the solution. Hence by Theorem 12. then Fl (PDF KFI ) = ^ FI ).3.3 Stabilizing Controllers for DF and OE Problems In the DF case we have the following system diagram: z yDF GDF w u .2. B1 C2 B1 B2 JDF = 6 F 0 I 4 I 0 . it is h to show easy h i i ^ DF in the DF structure. KDF := F` (JDF Q) ^ FI . We will further assume that A . then KDF C2 I = KDF C2 I in ^ that if KDF = K ^ the corresponding FI structure. . Remark 12. we have the following lemma which shows that the above construction of parameterization characterizes all stabilizing controllers for the DF problem.4.306 PARAMETERIZATION OF STABILIZING CONTROLLERS 12.2. Assume KFI = KFI = F Q for some Q 2 ^ ^ RH1 then KFI stabilizes GFI and F` (JDF Q) = F` (PDF KFI ) where A + B2 F . It should be pointed out that the existence of a stabilizing controller for this system is guaranteed by the stabilizability of (A B2 ) and detectability of (C2 A). B1 C2 is stable in this subsection.2. This equation characterizes an equivalent class of all controllers for the DF problem by the equivalence of FI and DF . Since KFI = K ^ KDF = KDF = F`(JDF Q).7 we have stabilizes GDF for any Q 2 RH1 . We say that the ^ controllers KDF and KDF are equivalent for the DF problem if the two transfer matrices from w to u in the above diagram are the same.7 By the equivalence between FI and DF problems. by Remarks 12. the resulting two closedloop transfer matrices from w to z are identical. We will now parameterize stabilizing controllers for GDF by invoking the relationship between the FI problem and DF problem established in section 12.C2 2 3 7 5 ^ with F such that A + B2 F is stable. Of course. We also have that if KFI = KFI . Hence this assumption is not necessary for our problem to be solvable however.1. In fact. Fl (PDF K ~ Now we construct the parameterization of equivalent classes of stabilizing controllers in DF structure via the tool we have developed in section 12.KDF The transfer matrix is given as in section 12. h i Let KDF (s) be a stabilizing control for DF then KFI (s) = KDF (s) C2 I h i ^ stabilizes the corresponding GFI .
B1 C2 .B2F B1 B2 7 6 6 0 A + B2 F 0 0 7 7 = 6 6 0 . We have already shown that the controller KDF = F`(JDF Q) for any given Q 2 RH1 does internally stabilize GDF . Parameterization of All Stabilizing Controllers 307 Lemma 12.12.F 0 I 7 5 4 I 0 . it is not necessary for the stabilization problem to be .) ^ ^ Let Q0 := F` (JDF KDF ) 2 RH1 then F`(JDF Q0 ) = F`(JDF F`(JDF KDF )) =: F` (Jtmp KDF ). Again this assumption is made only for the simplicity of the solution. Proof. 2 Now we turn to the dual OE case.F 0 I 7 5 4 I 0 0 C2 " # 0 I = : I 0 Hence F` (JDF Q0 ) = F`(Jtmp KDF ) = KDF .B2F B1 B2 7 6 6 .KOE u We will assume that A .3. The corresponding system diagram is shown as below: 2 3 z GOE w yOE .F 0 I 7 : J 4 5 C2 I 0 ^ (JDF is stabilized by KDF since it has the same `G22 ' matrix as GDF . B1 C2 + B2 F . Now let KDF be any stabilizing controller for ^ GDF then F` (JDF KDF ) 2 RH1 where A B1 B2 ^DF = 6 .11 All stabilizing controllers for the DF problem can be characterized by KDF = F` (JDF Q) with Q 2 RH1 . B2 C1 is stable. where JDF is given as above. This shows that any stabilizing controller can be expressed in the form of F` (JDF Q0 ) for some Q0 2 RH1 . where Jtmp can be obtained by using the state space star product formula given in Chapter 10: 2 3 A .B1 C2 A B1 B2 7 7 Jtmp = 6 6 F .C2 C2 3 2 A .
This also guarantees that the resulting closedloop transfer matrices are identical. " # B2 K Assume that KOE is an admissible controller for OE then KFC = OE = " L =: K for some Q 2 RH . Proof. And we can also observe the separation argument as the . Here we say controllers KOE and KOE are equivalent if the transfer matrices from yOE to z are the same.8. Lemma 12. by using again the state space star product formula given in Chapter 10.12 All admissible controllers for the OE problem can be characterized as F` (JOE Q0 ) with any Q0 2 RH1 .3. KOE = F` (JOE Q).308 PARAMETERIZATION OF STABILIZING CONTROLLERS solvable. Since KOE = KOE . This shows that any stabilizing controller can be expressed in the form of F` (JOE Q0 ) for some Q0 2 RH1 . where JOE is de ned as above.6 and the equivalent classes of controllers for ^ FC . 2 ^ JOE = 6 C1 4 C2 I 0 I 0 7 5 : 12.L B2 ^ ^ Let Q0 := F` (JOE KOE ) 2 RH1 . B2 C1 + LC2 L . We will assume the results of the special problems and show how to construct all admissible controllers for the OF problem from them. The controllers in the form as stated in the theorem are admissible since the corresponding FC controllers internally stabilize resulting GFC .B2 6 JOE = 4 C1 0 I C2 I 0 2 3 7 5 # I ^ with L such that A + LC2 is stable. KOE = F`(JOE Q) ^ stabilizes GOE for any Q 2 RH1 . Now we construct the parameterization for the OE structure as a dual case to DF . In fact. ^ Now assume KOE is any stabilizing controller for GOE then F`(JOE KOE ) 2 RH1 where 2 3 A . The parameterization of an equivalent class of stabilizing controllers for OE can be obtained by invoking Theorem 12. we have the following lemma. where A . Hence by Theorem 12. and K stabilizes G and F (J Q) = ^ FC ^ FC 1 FC ` OE Q ^ F` (POE KFC ).6. Then F`(JOE Q0 ) = F` (JOE F` (JOE KOE )) = KOE .4 Output Feedback and Separation We are now ready to give a complete proof for Theorem 12.
12. D22 6= 0.e. the mapping ^ K (s) = K (s)(I . i. For more general cases. we shall assume D22 = 0. Parameterization of All Stabilizing Controllers 309 byproduct this essentially involves reducing the OF problem to the combination of the simpler FI and FC problems. Denote x the state of system G then the openloop system can be written as x _ A B1 B2 7 6 z 5 = 4 C1 D11 D12 y C2 D21 0 x w 7: 5 u . we can see from the construction why the stability conditions of A . Moreover. Without loss of generality. Again we assume that the system has the following standard system block diagram: z G w y with . Proof of Theorem 12. Then the system ^ in terms of K has the structure z ^ G w y where 2 .K ^ u 3 A B1 B2 ^ (s) = 6 C1 D11 D12 7 : G 4 5 C2 D21 0 2 6 4 3 2 32 76 54 3 Now we construct the controllers for the OF problem with D22 = 0.K 2 u 3 A B1 B2 G(s) = 6 C1 D11 D12 7 5 4 C2 D21 D22 and that (A B2 ) is stabilizable and (C2 A) is detectable.1 is well de ned if the closedloop system is assumed to be well posed. B2 C1 in DF and OE problems were reasonably assumed and are automatically guaranteed in this case.3. B1 C2 and A .8. D22 K (s)).
Gtmp is of OE structure. Fx: Then the system can be broken into two subsystems: " # " # 2 6 4 x = A + B2 F B1 B2 _ z C1 + D12 F D11 D12 2 6 4 x w v 3 7 5 and A B1 B2 x _ 7 6 v 5 = 4 . h i Note that F 0 is actually a special FI stabilizing controller. Now let v = u . Then by Lemma 12. K stabilizes G if and only if K stabilizes Gtmp however.12 all controllers stabilizing Gtmp are given by K = F`(J Q) . Now let L be such that A + LC2 is stable. there is a constant matrix F such that A + B2 F is stable.310 PARAMETERIZATION OF STABILIZING CONTROLLERS Since (A B2 ) is stabilizable.F 0 I C2 D21 0 y 3 2 32 76 54 x w 7: 5 u 3 This can be shown pictorially below: z y G1 v Gtmp w u .K with and G1 = A + B2 F B1 B2 2 RH1 C1 + D12 F D11 D12 2 3 " # A B1 B2 6 Gtmp = 4 .F 0 I 7 : 5 C2 D21 0 Obviously.
in its general setting.C2 I 0 2 2 above construction.e. 2 3 A + B2 F + LC2 .F 0 I 4 C2 I 0 This concludes the proof.. K (s) = F` (J Q). we can get the following result from the OF problem. i.e. ~ Now we can reconsider the characterizations of all stabilizing controllers for the special problems with some reasonable assumptions.. because of its special structure. By assuming this. the stability conditions of A . i.8 shows that any stabilizing controller K (s) can be characterized as an LFT of a parameter matrix Q 2 RH1 . Parameterization of All Stabilizing Controllers where 311 3 A + B2 F + LC2 L . B1 C2 and A . rst use an output injection to reduce the OF problem to a DF problem.e. ~ Remark 12. Moreover. i. The separation argument is obvious since the synthesis of the OF problem can be reduced to FI and FC problems.3. We know that the solvability conditions for the FI problem are reduced. If we specify " # " # " # I D = 0 D = 0 C2 = 21 22 0 I 0 the OF problem. to (A B2 ) as stabilizable. becomes the FI problem.L B2 7 6 =4 F 0 I 7: 5 5 .L B2 0 I C2 I D22 7 5 and where K (s) has the stabilizable and detectable realization. .9 Theorem 12.11.F A . B2 C1 for DF and OE problems which were assumed in the last section can be dropped.e. i.13 Let L1 and F be such that A + L1 and A + B2 F are stable then all controllers that stabilize GFI can be characterized as F`(JFI Q) with any Q 2 RH1 . the latter two problems can be designed independently. Corollary 12.8 We can also get the same result by applying the dual procedure to the Remark 12. a realization of Q(s) in terms of K can be obtained as ^ Q := F` (J K ) where 3 2 ^ 4 J = 6 .12. using the same argument as in the proof of Lemma 12.B2 J =6 .
the transfer function from y to u is shown as below u JDF y A + B2 F + LC2 .8. which is required in Lemma 12. It can be seen that Lemma 12.312 where PARAMETERIZATION OF STABILIZING CONTROLLERS A + B2 F + L1 L . ~ 12. we can consider the FC problem as the special OF problem by specifying h i h i h i B2 = I 0 D12 = 0 I D22 = 0 0 : The DF (OE ) problem can also be considered as the special case of OF by simply setting D21 = I (D12 = I ) and D22 = 0. Then the proper K 's achieving internal stability are precisely those representable in Figure 12.14 Consider the DF problem. Let us recap what we have done.L B2 6 JDF = 4 F 0 I 7: 5 . In the same way. De ne J by the formula in Theorem 12.3 and K = F` (J Q) where Q 2 RH1 and I + D22 Q(1) is invertible. and assume that (C2 A B2 ) is stabilizable and detectable. i.10 It would be interesting to compare this result with Lemma 12.e. provides the natural injection matrix L = .11.B1 which satis es a partial condition in this corollary.F # 0 I " JFI = 6 6 I 4 I 0 2 3 7 7 7 7 5 0 and where L = (L1 L2) is the injection matrix for any L2 with compatible dimensions. Let F and L be such that A + LC2 and A + B2 F are stable. The condition that A . Corollary 12.11. and then all controllers that internally stabilize G can be parameterized as Fl (JDF Q) for some Q 2 RH1 .11 is a special case of this corollary.4 Structure of Controller Parameterization C2 D22 We choose F and L so that A + B2 F and A + LC2 are stable. B1 C2 is stable.C2 I 0 2 3 . We begin with a stabilizable and detectable realization of G22 # " A B2 : G22 = .Q Remark 12.B2 6 6 .
Furthermore.4. Theorem 12. it is easy to show that the transfer function between (y y1 ) and (u u1) is J . Structure of Controller Parameterization z y G D22 c 313 w u B2 ? c ? . i.e.Q Figure 12.1 is internally stable. " # " # This diagram of all stabilizing controller parameterization also suggests an interesting interpretation: every internal stabilization amounts to adding stable dynamics to the plant and then stabilizing the extended plant by means of an observer.12.L 66 c c 6 F y1 u1 . u =J y : u1 y1 It is also easy to show that the transfer matrix from y1 to u1 is zero..3: Structure of Stabilizing Controllers It is interesting to note that the system in the dashed box is an observerbased stabilizing controller for G (or G22 ).15 Assume that G22 and K are strictly proper and the system is Figure 12.. C2 R c A . The precise statement is as follows: for simplicity of the formulas. Then G22 can be embedded in a system " Ae Be Ce 0 " # where Ae = A 0 0 Aa " # Be = B2 0 # Ce = C2 0 h i (12:5) . only the cases of strictly proper G22 and K are treated.
Ly + B2 y1 Fx + y1 . Ba " # 2 12.3 leads to Figure 12. z = F` (G K )w = F`(G F` (J Q))w = F`(T Q)w: It remains to give a realization of T . Aa is stable. such that K has the form K = Ae + Be Fe + Le Ce . Elimination of the signals u and y in Figure 12.3 for some Q in RH1 . Take a minimal realization of Q: Q = Aa Ba : Ca 0 Since Q 2 RH1 . Let x and xa denote state vectors for J and Q. Thus all closedloop transfer matrices are representable as in Figure 12. Q must be strictly proper.314 PARAMETERIZATION OF STABILIZING CONTROLLERS " # and where Aa is stable.C2 x + y Aa xa + Ba u1 Ca xa xe = (Ae + Be Fe + Le Ce )xe . " # i Le := L . For K to be strictly proper. Le y _ u = Fe xe where h xe := x Fe := F Ca xa and where Ae Be Ce are as in (12. K is representable as in Figure 12.5 ClosedLoop Transfer Matrix Recall that the closedloop transfer matrix from w to z is a linear fractional transformation F`(G K ) and that K stabilizes G if and only if K stabilizes G22 .4 for a suitable transfer matrix T .Le Fe 0 where Ae + Be Fe and Ae + Le Ce are stable. " # (12:6) Proof.4.3: x _ u u1 xa _ y1 These equations yield = = = = = (A + B2 F + LC2 )x .5). (12:7) . respectively. and write the equations for the system in Figure 12.
6 Youla Parameterization via Coprime Factorization* In this section. all stabilizing controller parameterization will be derived using the conventional coprime factorization approach.16 Let F and L be such that A + BF and A + LC are stable. This is straightforward by using the state space star product formula and follows from some tedious algebra. 12. Hence it is left for the readers to verify.Q Figure 12.1N be the rcf and lcf of G22 over RH1 .B2 F 3 Proof. Youla Parameterization via Coprime Factorization* z T w 315 .1 = M . ~ ~ Theorem 12.D12F D11 D12 7 4 C1 + D12 F 5 0 C2 D21 0 A + B2 F 2 .4 of Chapter 5 before proceeding further. 2 An important point to note is that the closedloop transfer matrix is simply an a ne function of the controller parameter matrix Q since T22 = 0. Readers should be familiar with the results presented in Section 5.4: Closed loop system Theorem 12. respectively.1 )(1) 6= 0 (12:9) .1 (U0 + Ql M ) det(I + Ql N V0.12.6.17 Let G22 = NM . Then the set of all proper controllers achieving internal stability is parameterized either by K = (U0 + MQr )(V0 + NQr ).1 det(I + V0.1 NQr )(1) 6= 0 (12:8) for Qr 2 RH1 or by ~ ~ ~ ~ ~~ K = (V0 + Ql N ). Then the set of all closedloop transfer matrices from w to z achievable by an internally stabilizing proper controller is equal to F` (T Q) = fT11 + T12 QT21 : Q 2 RH1 I + D22 Q(1) invertibleg where T is given by T = T11 T21 " B1 B2 # 6 7 6 0 A + LC2 B1 + LD21 0 7 : T12 = 6 7 6 T22 .
1(MV0 .1 U0 .1) .1(I + NUZ . suppose K is proper and achieves internal stability. U0 ) ~ ~ = V0 + M . De ne Qr by the equation U0 + MQr = UZ .1 V0. NM )Qr = I: Thus K achieves internal stability by Lemma 5. N (U0 + MQr ) = MV0 .V0 N Proof. if U0 V0 U0 . NU0 + (MN .1 NQy )(1) is invertible (12. Z := MV .10.1 . we have V0 + NQr = V0 + NM .10.316 PARAMETERIZATION OF STABILIZING CONTROLLERS ~ ~ for Ql 2 RH1 where U0 V0 U0 V0 2 RH1 satisfy the Bezout identities: ~ ~ ~ ~ V0 M .1 .N M N V0 0 I Then K = (U0 + MQy )(V0 + NQy ).1 (12:12) so Qr = M . NU = M (V0 + NQr ) .1 (U0 + Qy M ) = F` (Jy Qy ) where " # ~ U0 V0.1 .1 Jy := . We shall prove the parameterization given in (12.1 ~ ~ = M .1 and where Qy ranges over RH1 such that (I + V0.1 .1 . U0 ): Then using the Bezout identity. Assume that K has the form indicated.1: (12.1(Z + NU )Z .1 N (UZ ..13) . Conversely. U0 ) ~ .8) rst.1(UZ .1 = V0. " #" # " # ~ ~ V0 .1 ) ~ ~ = M ~ ~ ~ = M .10) (12:11) V0 . NU is invertible in RH1 . and V0 are chosen such that U0V0.U0 M U0 = I 0 : ~ ~ .1 ~ ~ ~ ~ = (V0 + Qy N ).1MV Z .e. Then by Lemma 5. and de ne U := U0 + MQr V := V0 + NQr : Then ~ ~ ~ ~ ~ ~ ~ ~ MV .1 (UZ .1 ~ = M ~ = V Z . U0 N = I MV0 . i. NU0 + NUZ . NU0 = I: ~ ~ ~ ~ Moreover. Introduce an rcf of ~ ~ K over RH1 as K = UV .
Similarly.13).13) that both MQr and NQr belong to RH1 .1 ~ K = U0 V0. NU: Next. V0.10). U0 (NQr ) 2 RH1 : Finally. since V and Z evaluated at s = 1 are both invertible.1 = U0 V0.1 NQy ). U0 N ) = V0. hence so is I + V0. Theorem 12. note that (U0 + MQy )(V0 + NQy ).1 .1 N )Qy (I + V0.1 + (M .1 : To see that Qr belongs to RH1 . Youla Parameterization via Coprime Factorization* Thus.1 = (U0 + MQr )(V0 + NQr ).1 + V0. the free parameter Qy 2 RH1 in Youla parameterization is given by Qy = M .1 ~ ~ and that U0 V0. U0 ) where ~ ~ Z := MV .1 = V0. the parameterization given in (12.1 U0N ) = V0.12) and then from (12. To show that the controller can be written in the form of equation (12. we shall establish the precise relationship between the above all stabilizing controller parameterization and the parameterization obtained in the previous sections via LFT framework. U0 N )Qr = V0 (MQr ) .1 U0 .1 Qy (I + V0.1 V0. observe rst from (12. U0 V0.1 N ) = (M .1 (V0 M .12. Then ~ ~ ~ ~ Qr = (V0 M .9) can be obtained.6. We have and ~ ~ ~ ~ ~ ~ (M .19 Let the doubly coprime factorizations of G22 be chosen as 3 2 " # A + B2 F B2 .L M U N V0 0 =6 4 F I C2 + D22 F D22 I 0 7 5 .1 NQy ).1 V0.1 = V .1 U . 317 K = UV .1 : (12:14) 2 Corollary 12. so is V0 + NQr from (12. U0 V0.18 Given an admissible controller K with coprime factorizations K = ~ ~ UV .1 NQr .1 (UZ .
Glover. The conventional Youla parameterization can be found in Youla et al 1976].D22 I where F and L are chosen such that A + B2 F and A + LC2 are both stable. This follows from some tedious algebra.L B2 + LD22 6 7 Jy = 4 F 0 I 5: . Then Jy can be computed as ~ ~ V0 . and Francis 1989] for solving the H1 problem. The parameterization of all twodegreeoffreedom stabilizing controllers is given in Youla and Bongiorno 1985] and Vidyasagar 1985]. Doyle 1984].7 Notes and References The special problems FI. FC.(C2 + D22 F ) I . Zhou.D22 2 3 Proof. Desoer et al 1980].L : F 0 # ~ 12. and Francis 1987].8 and that . and they have been since used in many other papers for di erent problems. .N M # 2 3 A + B2 F + LC2 + LD22F . and OE were rst introduced in Doyle. Vidyasagar 1985].318 PARAMETERIZATION OF STABILIZING CONTROLLERS " A + LC2 .U0 ~ ~ .11 Note that Jy is exactly the same as the J in Theorem 12. The new derivation of all stabilizing controllers was reported in Lu. The paper by Moore et al 1990] contains some other related interesting results. DF. K0 := U0V0 is an observerbased stabilizing controller with " 2 Remark 12. and Doyle 1991].1 K0 := A + B2 F + LC2 + LD22 F .(B2 + LD22) L 6 =4 F I 0 7 5 C2 . Khargonekar.
1 HJ = . Then J . To see that.JHJ = .2) will be used to obtain the solutions to the equation (13.Algebraic Riccati Equations We have studied the Lyapunov equation in Chapter 3 and have seen the roles it played in some applications. Then an algebraic Riccati equation is the following matrix equation: A X + XA + XRX + Q = 0: (13:1) Associated with this Riccati equation is a 2n 2n matrix: 13 H := A R . they play the central roles in H2 and H1 optimal control.1). Roughly speaking. and R be real n n matrices with Q and R symmetric. A more general equation than the Lyapunov equation in control theory is the socalled Algebraic Riccati Equation or ARE for short. introduce the 2n 2n matrix: " # 0 .I J := I 0 having the property J 2 = . The matrix H in (13. Q.H 319 . Let A.I . Lyapunov equations are most useful in system analysis while AREs are most useful in control system synthesis particularly. It is useful to note that (H ) (the spectrum of H) is symmetric about the imaginary axis.Q .A " # : (13:2) A matrix of this form is called a Hamiltonian matrix.
real. innerouter factorizations and normalized coprime factorization.4. not necessarily hermitian. Some other special rational matrix factorizations.g. are also given in this section.1). The word \all" means that any X . there is a matrix 2 C n n such that " A R . In Section 13. the solution X is independent of a speci c choice of bases of V .5 and 13.A #" X1 = X1 X2 X2 # " # # " # : .1. is.8.1: X X 1 1 (13:3) . This chapter is devoted to the study of this algebraic Riccati equation and related problems: the properties of its solutions.2 which focuses on the stabilizing solutions.Q . the methods to obtain the solutions. The most important part of this chapter is Section 13..H are similar. 13.1 All Solutions of A Riccati Equation The following theorem gives a way of constructing solutions to (13. which is not necessarily real. Thus is an eigenvalue i . Section 13.. Positive real functions and inner functions are introduced in Section 13.A #" I = I X X . If X1 is invertible. This section is designed to be essentially selfcontained so that readers who are only interested in the stabilizing solution may go to Section 13. and some applications. then X := X2 X1 1 is a solution to the Riccati equation (13. " # Proof.713. The relationship between the stabilizing solution of a Riccati equation and the spectral factorization of some frequency domain function is established in Section 13. Since V is an H invariant subspace. satis es equation (13.1).1) and (A + RX ) = (H jV ). Postmultiply the above equation by X1 1 to get " A R . Furthermore.e. The conditions for a solution to be hermitian.1) in terms of invariant subspaces of H .Q .1 Let V C 2n be an ndimensional invariant subspace of H . maximal or minimal) solutions of a Riccati equation and their properties.3 presents the extreme (i. and so on.320 ALGEBRAIC RICCATI EQUATIONS so H and . are given in Sections 13.2 directly without any di culty. e. and let X1 X2 2 C n n be two complex matrices such that V = Im X1 : X2 . and not necessarily stabilizing. we will study all solutions to (13. not necessarily nonnegative. Theorem 13.6.
the columns of span an ndimensional invariant subspace of H .1). A X .3) also gives therefore.1.XA . such that X = X2 X1 1 and the " columns of X1 X2 form a basis of an ndimensional invariant subspace of H . 2 As we would expect.then there exist matrices X1#X2 2 C n n . Q . 2 " I X .. Equation (13.1). the converse of the theorem also holds. X2 X1 1. XRX . so (A + RX ) = (H jV ).Q . (A + RX ) = ( ). Finally note that any other basis spanning V can be represented as " # " # X1 P = X1 P X2 X2 P for some nonsingular matrix P . But. Write these two relations as " A R . All Solutions of A Riccati Equation Now premultiply (13.1).2 If X 2 C n n is a solution to the Riccati equation (13. with X1 invertible. A + RX = X1 X1 1 I X # which shows that X is indeed a solution of (13.X I . is a matrix representation of the map H jV . Proof. by de nition. The conclusion follows from the fact (X2 P )(X1 P ). and de ning X1 := I . and X2 := X completes the proof.X I h h i 321 to get i " #" A R 0 = . Multiplying this by X gives X = XA + XRX = .3) by .Q .Q . A X with the second equality coming from the fact that X is a solution to (13. Theorem 13. De ne := A + RX .A # #" I = I X X # " # : Hence.13.1 = .A = .
1 Let A = .1 7 6 7 6 6 2 7 6 . One way of constructing those invariant subspaces is to use eigenvectors and generalized eigenvectors of H .1 0 0 The eigenvalues of H are 1 1 .2 # # ..1 . 1).1 = vi+k.1 be the corresponding generalized eigenvectors associated with vi and i . i I )v i = 0 (H .1.1 are 2 6 6 6 6 4 2 3 2 3 " # " # " # v3 = 1 1 0 0 " 3 7 7 7 7 5 2 v4 6 6 =6 6 4 1 7 3=2 7 7: 0 7 5 0 " 3 All solutions of the Riccati equation under various combinations are given below: X1 = v v ]. 1g is an invariant subspace of H . . Suppose i is an eigenvalue of H with multiplicity k (then i+j = i for all j = 1 : : : k . and the corresponding eigenvector and generalized eigenvector of 1 are 1 .2 0 The corresponding eigenvector and generalized eigenvector of .3=2 7 7 6 7: v1 = 6 6 2 7 v2 = 6 17 4 5 4 5 . and let vi be a corresponding eigenvector and vi+1 : : : vi+k.4 . (H . ~ Example 13.10 6 6 .1 = spanfv1 v2 g is H invariant: let 1 2 2 1 X2 is a solution and (A + RX ) = f1 1g " # " X1 = v v ].2 2 2 .322 ALGEBRAIC RICCATI EQUATIONS Remark 13.3 2 R = 0 0 Q= 0 0 : . then X = X X .1 = spanfv1 v3 g is H invariant: let 1 3 2 1 X2 is also a solution and (A + RX ) = f1 .2 1 0 . i I )vi+k. then X = X X .1g # . it is necessary to be able to construct bases for those invariant subspaces of H . Then vj are related by ( H .1 It is now clear that to obtain solutions to the Riccati equation.2 and the spanfvj j = i : : : i + k . i I )vi+1 = vi .
X1 X2 + X2 X1 ) = . X1 = v v ].1 and 13.e. Since V is an invariant subspace of H . X1 X2 = (X1 X2 ) . i. there is a matrix representation for H jV such that ( ) = (H jV ) and H X1 = X1 X2 X2 X1 Premultiply this equation by X2 " " # " # " # : J to get # " # X1 X2 # # JH X1 = X1 X2 X2 " # " " J X1 X2 " # : (13:4) Note that JH is hermitian (actually symmetric since H is real) therefore.1. Read 3 Up to this point. Furthermore. if X1 is nonsingular.3 Let V be an ndimensional H invariant subspace and let X1 X2 2 Cn V = Im X1 : X2 Then i + j 6= 0 for all i j = 1 : : : n.X1 X2 + X2 X1 ): . then X = 0 is a solution and spanfv3 v4 g is H invariant: let 3 4 X2 (A + RX ) = f. Theorem 13..1).. (.1 . the lefthand side of (13. we have said nothing about the structure of the solutions given by Theorem 13.13.1g spanfv1 v4 g spanfv2 v3 g. i j 2 (H jV ) implies that X1 X2 is hermi.4) is hermitian as well as the righthand side: " X1 X2 # J X1 X2 " = X1 X2 J X1 = .2. (. tian.e. X2 # " X1 X2 # J X1 X2 " # i. The following theorem gives a su cient condition for a Riccati solution to be hermitian (not necessarily real symmetric). n be such that " # Proof. All Solutions of A Riccati Equation " # 323 ers can verify that the matrices constructed from those vectors are not solutions to the Riccati equation (13. and spanfv2 v4 g are not H invariant subspaces. then X = X2 X1 1 is hermitian.
" # 2 Example 13.X1 X2 + X2 X1 = 0: This implies that X1 X2 is hermitian. (() Since V is conjugate symmetric. C X2 .4 3 7 5 2 Q=6 4 0 0 0 0 0 0 7: 5 0 0 0 3 .2 This example is intended to show that there are nonreal. Theorem 13. there is a nonsingular matrix P 2 C n such that " n . . Proof. which excludes the possibility of having imaginary axis eigenvalues since if l = j! then l + l = 0. By assumption. X 2 Rn n and X1 = X1 P X2 X2 # " # I =V Im X therefore. Since i + j 6= 0.2 It is clear from Example 13.1). Therefore. i + j 6= 0 is not ~ The following theorem gives necessary and su cient conditions for a solution to be real. . That X is hermitian is easy to see by noting that X = (X1 1 ) (X1 X2 )X1 1 . v 2 V implies Then X := X2 X1 that v 2 V . V is conjugate symmetric.. is not necessary for the existence of a hermitian solution.1 is real if and only if V is conjugate symmetric.324 ALGEBRAIC RICCATI EQUATIONS This is a Lyapunov equation. and let X1 X2 2 " # n n be such that X1 is nonsingular and the columns of X1 form a basis of V .1 0 . nonhermitian solutions to equation (13. i.2 . where the over bar X denotes the complex conjugate.1 = X2 X .1 = X is real as desired.1 0 1 0 3 7 5 2 R=6 0 0 0 4 . It is also designed to show that the condition i + j 6= 0 8 i j . the equation has a unique solution: .1 that the condition necessary for the existence of a hermitian solution. ()) De ne X := X2 X1 1.2 0 . X = X2 X1 1 = . Let 2 A=6 4 1 0 0 0 0 . X2 P (X1 P ) 1 . 2 Remark 13.e.4 Let V be an ndimensional H invariant subspace.
6 . (A + RX ) C . We will take the domain of Ric.1) and that X is neither real nor hermitian. which will be denoted Ric. 4 = . 2 2 3 = 5 = j = . denoted dom(Ric). (H ) = Im X1 X2 where X1 X2 2 C n n . By nding a basis for X. stacking the basis vectors up to form a matrix. Consider the two ndimensional spectral subspaces. Hence some of the material appearing in this section may be similar to that seen in the previous section. Assume H has no eigenvalues on the imaginary axis. (H ) and X+ (H ): the former is the invariant subspace corresponding to eigenvalues in Re s < 0 and the latter corresponds to eigenvalues in Re s > 0.! X is a function. 3 13.e. For ease of reference. 0:5j 6 7 X=4 0 0 0 5 0 0 0 satis es equation (13. (H ). equivalently. to consist of Hamiltonian matrices H with two properties: H has no eigenvalues on the imaginary axis and the two subspaces in (13. and the properties of such solutions. we can set X := X2 X1 1.. (H ) Im " I 0 # (13:5) .2 Stabilizing Solution and Riccati Operator In this section. and partitioning the matrix.. This section is the central part of this chapter and is designed to be selfcontained.2. are complementary. (X1 and X2 can be chosen to be real matrices. we get " # X. Then X is uniquely determined by H . i. Then it must have n eigenvalues in Re s < 0 and n in Re s > 0. j . H 7.e.j . 2 X=6 4 2 0 0 0 0 0 0 0 0 3 7 5 is a real symmetric nonnegative de nite solution corresponding to the eigenvalues .13. if the two subspaces X. On the other hand. It is easy to show 0 0:5 + 0:5j 0:5 . i. Stabilizing Solution and Riccati Operator Then H has eigenvalues that 1 325 3 =1=.) If X1 is nonsingular or. we discuss when a solution is stabilizing.1. these will be called the stability property and the complementarity .5) are complementary. . X.
the unique solution is (. Since X1 is nonsingular and X = (X1 1 ) (X1 X2 )X1 1 . is a matrix representation of H jX.X1 X2 + X2 X1 = 0: This proves (13. . is stable. (. respectively. X is symmetric. = H. (H ) . so is the lefthand side of (13. Proof. Theorem 13.X1 X2 + X2 X1 )H. : X2 " # (13:7) . . This solution will be called the stabilizing solution. X = Ric(H ) and Ric : dom(Ric) R2n 2n 7.7) and so is the righthand side: = .5 Suppose H 2 dom(Ric) and X = Ric(H ).X1 X2 + X2 X1 ): This is a Lyapunov equation. Thus. Then (i) X is real symmetric (ii) X satis es the algebraic Riccati equation A X + XA + XRX + Q = 0 (iii) A + RX is stable .H. : X2 X2 (H. note that there exists a stable matrix H. .326 ALGEBRAIC RICCATI EQUATIONS property. It is claimed that X1 X2 is symmetric: To prove this. in Rn n such that H X1 = X1 H.X1 X2 + X2 X1 ) X1 X2 # JH X1 = X1 X2 X2 " # J X1 H.6).(.) Premultiply this equation by " " # " # (13:6) X1 X2 " # J # to get " Since JH is symmetric. (i) Let X1 X2 be as above. Since H.! Rn n : The following wellknown results give some properties of X as well as veri able conditions under which H belongs to dom(Ric).
2 Now. (iii) Premultiply (13.5 de ne X1 X2 H. : X2 X2 " X2 # (13:9) . Theorem 13. and postmultiply by X1 1 to get " " # " # . Stabilizing Solution and Riccati Operator (ii) Start with the equation 327 H X1 = X1 H.8) by I 0] to get . X2 X2 . we are going to state one of the main theorems of this section which gives the necessary and su cient conditions for the existence of a unique stabilizing solution of (13. is. I ]H I = 0: X This is precisely the Riccati equation. we must show that X.13. (H ) = Im X1 # " H X1 = X1 H. As in the proof of Theorem 13. (H ) Im " 0 # I are complementary.6 Suppose H has no imaginary eigenvalues and R is either positive semide nite or negative semide nite. Proof.2. X1 1: X X # " # (13:8) Now premultiply by X . I ]: X . so that " # X. X1 1: " # Thus A + RX is stable because H. (() To prove that H 2 dom(Ric). Then H 2 dom(Ric) if and only if (A R) is stabilizable. H I = I X1 H. This requires a preliminary step. A + RX = X1 H.1) under certain restrictions on the matrix R.
Premultiply (13.. x = x (13:11) Re < 0 0 6= x 2 Ker X1 : Premultiply (13. Now to prove that X1 is nonsingular.10) by x to get X1 H. it is claimed that Ker X1 is H. First. so that the equation XM = C always has a solution for M . x = 0. x 2 Ker X1 . we have x X2 A + I R] = 0: Then stabilizability of (A R) implies X2 x = 0.9) by I 0] to get AX1 + RX2 = X1 H. . It also implies that (A R) must be stabilizable.e. Now postmultiply (13. and a corresponding eigenvector. which is a contradiction. and Ker(X ) = 0 if and only if (C A) has no stable unobservable modes. and a minimum F norm solution is given by X yC . then " # x = 0 since X1 has full column rank.6)) to get x X2 RX2 x = 0: Since R is semide nite. this implies that RX2 x = 0.7 Suppose H has the form " A .9) by 0 I ]: .e. i. invariant. QX1 .11): (A + I )X2 x = 0: Recall that RX2 x = 0. i. let x 2 Ker X1 . and use the fact that X2 X1 is symmetric (see (13. X2 ()) This is obvious since H 2 dom(Ric) implies that X is a stabilizing solution and that A + RX is asymptotically stable. Note that Ker(X ) Ker(C ). : (13:10) Premultiply by x X2 . A X2 = X2 H. Then H. : (13:12) Postmultiply the above equation by x and use (13.328 ALGEBRAIC RICCATI EQUATIONS We want to show that X1 is nonsingular. that Ker X1 6= 0.A Then H 2 dom(Ric) i (A B ) is stabilizable and (C A) has no unobservable modes on the imaginary axis. suppose.C C . x: H. postmultiply by x. 2 Theorem 13.BB : H= . jKer X1 has an eigenvalue. H. But if both X1 x = 0 and X2 x = 0. Furthermore. X = Ric(H ) 0 if H 2 dom(Ric). This proves the claim. on the contrary. To prove this. # . Ker X1 = 0.
j!I )xi = kB z k2: Therefore B z = 0 and Cx = 0. H has no imaginary eigenvalues i (C A) has no unobservable modes on the imaginary axis. and it is also su cient if H has no eigenvalues on the imaginary axis. We'll show that X 0. j!I )x = BB z . It is clear from Theorem 13. assuming (A B ) is stabilizable. j!I )x z i = hz (A . Suppose that j! is an eigenvalue " # x is a corresponding eigenvector. j!I )x = 0 (A . j!I ) z = C Cx: (13:13) (13:14) so hx (A . j!I x = 0: C # The stabilizability of (A B ) gives z = 0.kCxk2 = h(A .C Cx . XBB X + C C = 0 or equivalently (A . j!I ) z = 0: Combine the last four equations to get z A . set X := Ric(H ).hx (A . j!I ) z i is real and hz (A . j!I ) z i = hx C Cxi = kCxk2 .2. (A . Stabilizing Solution and Riccati Operator 329 Proof.13) and (13. BB z = j!x . A z = j!z: Rearrange: Thus (A .14) (A . The Riccati equation is A X + XA . BB X ) + XBB X + C C = 0: (13:15) .6 that the stabilizability of (A B) is necessary. Then and 0 6= z Ax . Now it is clear that j! is an eigenvalue of H i j! is an unobservable mode of (C A).13. So we only need to show that. j!I B ] = 0 " A . So from (13. BB X ) X + X (A . Next. j!I )xi = hz BB z i = kB z k2 .
.e. Riccati equation 3 A X + XA .15) by x and postmultiply by x to get Cx = 0: Now postmultiply (13.. XBB X + C C = 0 Corollary 13.5).24 36 is the stabilizing solution. Re < 0 thus is a stable unobservable mode.24 > 0 . i. i. X = 18 . Let x 2 KerX .e.BB X )t dt: (13:16) Since XBB X + C C is positive semide nite. Now if Ker(X ) 6= 0. then Xx = 0. we'll show that KerX is nontrivial if and only if (C A) has stable unobservable modes. there is an x such that Ax = x Cx = 0. X is singular. BB X )x and Cx = 0. we have X= e(A. suppose (C A) has an unobservable stable mode .330 Z ALGEBRAIC RICCATI EQUATIONS 1 0 Noting that A . we get 2Re x Xx . the solution is stabilizing.15) again by x to get XAx = 0: We conclude that Ker(X ) is an Ainvariant subspace. 2 Example 13. Then the has a unique positive semide nite solution.3 This example shows that the observability of (C A) is not necessary for the existence of a positive de nite stabilizing solution. but (C A) is not detectable. Moreover. By premultiplying the Riccati equation by x and postmultiplying by x. then there is a 0 6= x 2 Ker(X ) and a such that x = Ax = (A . Conversely. BB X ) is stable.BB X ) t (XBB X + C C )e(A. Let A= 1 0 0 2 " # B= 1 1 " " # C= 0 0 : # h i Then (A B ) is stabilizable. However. Premultiply (13.8 Suppose that (A B) is stabilizable and (C A) is detectable. Finally.. BB X is stable (Theorem 13. Since (A . x XBB Xx = 0: Hence x Xx = 0. so is X .
BR.. First rewrite the Riccati equation as (A . let us assume that X 0 satis es the Riccati equation but that it is not stabilizing. i. respectively.13. DR..1 D C )x = j!x (I . BB X ) X + X (A . Hence we only need to show that any positive semide nite solution X 0 must also be stabilizing..1 D )C A . Finally.1D C I I 0 #" x = 0: 0 # . j!I B C D # has full column rank for all !. (ii) (I . BR. Stabilizing Solution and Riccati Operator 331 Proof. DR.. (C A) is not detectable.1 D )Cx = 0 i. and we have ( + )x Xx + x (XBB X + CC )x = 0: This implies B Xx = 0 Cx = 0 since Re( ) 0 and X 0.9 Suppose D has full column rank and let R = D D > 0 then the following (i) .e. X the stabilizing solution.e. To achieve that goal.R. Proof.1D C then there is an x = 0 such that 6 (A . A . statements are equivalent: " 0 is 2 Lemma 13. j!I B C D #" . It is obvious from the above theorem that the Riccati equation has a unique stabilizing solution and that the solution is positive semide nite. respectively. BB X ) + XBB X + C C = 0 (13:17) and let and x be an unstable eigenvalue and the corresponding eigenvector of A . Hence Re( ) < 0.1D )C A . BB X )x = x: Now premultiply and postmultiply equation (13. i. (A . we arrive at Ax = x Cx = 0 i. Then by the uniqueness of the stabilizing solution. " . BB X .1 D C has no unobservable modes on j!axis. which is a contradiction. BR. Suppose j! is an unobservable mode of (I .e.2. DR. A .17) by x and x. we can conclude that there is only one positive semide nite solution.e.
then there is a D? such that D? DR. i.1 D C with no imaginary eigenvalues. h 2 i Remark 13. DR.C C . Hence.1 D C ) having no imaginary unobservable modes. BR.1D C I # " # x : y # # " #" u 6= 0 v (13:19) (13:20) (A . Let H have the form H = = " " h A 0 .1 D C .1 D C ) # " # i . suppose (13.. Then we have (A . j!I B C D # (13:18) does not have full column rank.A .1 D .C D # A .e. BR. Note also that if D C = 0. B R.1 D )C . j!I )x + By = 0 (I . in some cases we will write the condition (ii) in the above lemma as (D? C A .1 D C )x = j!x (I . Conversely.R.1 D )Cx = 0 .1 D )C A . j!I B C D " u = 0: v #" Now let Then and " u = I 0 v . condition (ii) becomes (C A) with no imaginary unobservable modes. BR. BR.18) does not have full column " # rank for some ! then there exists " u 6= 0 such that v #" # A .1 D C . ~ Corollary 13.1 D C .BR. if D is square. DR.1D )Cx + Dy = 0: Premultiply (13.C (I .3 If D is not square. Of course. BR.20) by D to get y = 0.10 Suppose D has full column rank and denote R = D D > 0.1=2 is unitary and that D? D? = I .1 B : .(A . BR.1 D C B . DR. BR. j! is an unobservable mode of (I .DR. DR. the condition is simpli ed to A .1D C I x = I 0 y R.332 But this implies that " ALGEBRAIC RICCATI EQUATIONS A .
BD C ) has eigenvalues on the imaginary axis although (A B ) is controllable and (C A) is observable. even if a stabilizing solution exists. Among all solutions. let us look at an example. To illustrate the idea. BR. On the other hand.1 D C ) has no unobservable modes on the imaginary axis. Furthermore.3.7. consider " # " # " # " # 0 1 0 1 0 1 A= B= C= D= : 0 0 . . the solution is not positive de nite since (D? C A .9 and Theorem 13. symmetric. X = Ric(H ) 0 if H 2 dom(Ric). BD C = D? C = 0 0 : 1 0 A Riccati equation with the above data has a nonnegative de nite stabilizing solution since (D? C A . Hence the existence of a stabilizing solution to the Riccati equation in the above corollary is not guaranteed by the stabilizability of (A B ) and detectability of (C A). BR. (A .3 Extreme Solutions and Matrix Inequalities We have shown in the previous sections that given a Riccati equation.1 0 0 0 Then (C A) is observable. This is the consequence of the Lemma 13. in particular. Proof.1 D C ) implies the detectability (observability) of (C A) however.4 It is easy to see that the detectability (observability) of (D?C A . Furthermore. As an example. we are most interested in those which are real.1 D C ) has no stable unobservable modes. and Ker(X ) = 0 if and only if (D? C A . and. j!I B has full column rank Then H 2 dom(Ric) i (A B ) is stabilizable and C D for all !. the positive de niteness of the solution is not guaranteed by the observability of (C A) unless D C = 0.13. The connections between the extreme solutions and the stabilizing solutions will also be established in this section. in this case. if the B matrix is changed to " # 0 B= 1 then the corresponding Riccati equation has no stabilizing solution since. 2 Remark 13. (A B ) is controllable. the converse is in general not true.1 D C ) has a stable unobservable mode. ~ 13. There is another class of solutions which are interesting they are called extreme (maximal or minimal) solutions. the stabilizing solutions. However. and " # h i 0 1 A . BR. Some properties of the extreme solutions will be studied in this section. BR. Extreme Solutions and Matrix Inequalities " # 333 A . there are generally many possible solutions.
C C . i.1 3 = 2 4 = .1 = 6 36 0 7.3 6 = 3 and their corresponding eigenvectors are 3 2 3 2 3 2 3 2 3 2 3 2 .6 7 6 0 7 6 3 7 6 1 7 6 2 7 6 0 7 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7 6 6 0 7 6 .3 7 6 0 7 6 .4 Let A=6 4 1 0 0 0 2 0 0 0 .24 5 2 4 5 4 2 1 X2 0 0 0 These solutions can be ordered as Y4 Yi Y1 i = 2 3.24 0 X1 = v v v ] Y = X X . 3 " . respectively. this is only a partial ordering since Y2 and Y3 are not comparable.12 7 7 6 7 6 7 6 v1 = 6 7 v2 = 6 7: 7 v6 = 6 7 v5 = 6 7 v4 = 6 7 v3 = 6 6 0 7 6 0 7 6 0 7 6 0 7 6 6 7 6 0 7 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7 6 7 6 6 0 7 6 0 7 4 0 5 4 0 5 4 12 5 4 0 5 5 4 5 4 6 0 0 0 0 0 There are four distinct nonnegative de nite symmetric solutions depending on the chosen invariant subspaces: 2 3 " # 0 0 0 X1 = v v v ] Y = X X .3 7 0 4 0 3 1 6 7 6 7 6 7 6 7 6 7 6 6 .1 7 6 1 7 6 . BB Y4 is stable.A 1 = 1 2 = .1 = 6 4 2 3 5 2 2 1 X2 X1 = v v v ] Y = X X . Note also that only Y4 is a stabilizing solution. Y4 and Y1 are the \maximal" and \minimal" solutions.1 = 6 4 1 4 5 3 2 1 X2 # 2 # 2 # 2 2 0 0 0 0 0 0 0 0 0 0 0 0 4 0 0 0 0 3 7 5 3 7 5 3 " 18 .334 ALGEBRAIC RICCATI EQUATIONS 2 Example 13. Of course.e.3 " 3 7 5 2 B=6 4 1 1 1 3 7 5 # C= 0 0 0 : are h i A . 4 . A . Furthermore.2 5 = ..1 = 6 7 4 0 0 0 5 1 3 5 1 2 1 X2 0 0 0 " X1 = v v v ] Y = X X .BB The eigenvalues of matrix H = .
Clearly. (i): Let X be such that Q(X ) 0.1).1). Proof. maximal and minimal solutions are unique if they exist.1). ) t dt . Furthermore.4 can be stated in a much more general setting.A R) is stabilizable. we shall introduce the following quadratic matrix: Q(X ) := A X + XA + XRX + Q: (13:21) Theorem 13. ) C + . Similarly.1 . Extreme Solutions and Matrix Inequalities 335 The partial ordering concept shown in Example 13. X+ X 8 X such that Q(X ) 0 and (A + RX+ ) C . then there exists a unique maximal solution X+ to the Riccati equation (13. X. . i (A + RX+ ) C . and X. Furthermore. ) C + . Let R = . X+ . there is an F0 such that A0 := A + BF0 . we shall call a hermitian solution X. then both X+ and X. of (13. X. To study the properties of the maximal and minimal solutions. Furthermore. X 8 X such that Q(X ) 0 and (A + RX. = = Z Z 1 0 0 e(A+RX+ )t Re(A+RX+ ) t dt e(A+RX.1). the results in (i) and (ii) can be respectively strengthened to X+ > X . again consider the Riccati equation (13.) C + . < X .13. X for all hermitian solutions X of (13. i (A + RX. to the Riccati equation (13.1 : (iv) If Q(X ) > 0.1). exist.1 .1) a maximal solution if X+ X for all hermitian solutions X of (13. (A + RX. (A + RX+ ) C .3. To do that. (iii) If (A R) is controllable. X+ > X. (i) If (A R) is stabilizable. .11 Assume R 0 and assume there is a hermitian matrix X = X such that Q(X ) 0. then there exists a unique minimal solution X. In this case. Note the fact that (A R) is stabilizable (controllable) i (A B ) is. )t Re(A+RX. Since (A B ) is stabilizable.1) a minimal solution if X. We shall call a hermitian solution X+ of (13. (ii) If (.BB for some B.
1 .1 ) (Fn . Fn. with i = n. X )A0 + A0 (X0 .22). we get Xn. Then. Now let X0 be the unique solution to the Lyapunov equation Then X0 is hermitian.1 X Next..1 + Q + Fn Fn + (Fn . and we have ^ ^ 2Re x (Xn.1 )gx: . i.1 . Fn.B Xn.1 ) (Fn . Fn Fn . 1 Xi Ai + Ai Xi = .Q(X ) . Then premultiply (13. 1. (Fn . Fn.F0 F0 .22). 1 Fi = . Now using (13. Associated with fXi g.1 ): (13:24) Now assume that An is not stable. Q i = 0 1 : : : n . using (13. and fFi g for i up to n . Q(X ) 0: The stability of A0 implies that Starting with X0 . we de ne a hermitian matrix Xn with Xn. Fn. X )An + An (Xn.e. Fn. fAi g. there exists an with Re 0 and x 6= 0 such that An x = x.B Xi.Fi Fi . 1 such that Xi is hermitian and X0 X: X0 X1 Xn.336 ALGEBRAIC RICCATI EQUATIONS X0 A0 + A0 X0 + F0 F0 + Q = 0: is stable.1 An = A + BFn : (13:22) First we show that An is stable. De ne ^ F0 := F0 + B X and we have the following equation: ^ ^ (X0 .24) by x and postmultiply by x.x fQ(X ) + Fn Fn + (Fn . Assume inductively that we have already de ned matrices fXi g. X ) = .1 Xn X .1 ) = 0: Let then ^ Fn := Fn + B X (13:23) ^ ^ (Xn.1 ) (Fn . with i = n . introduce Ai = A + BFi is stable i = 0 : : : n . 1: Fn = .1 An + An Xn.1 i = 1 : : : n . Fn.1 . X )x = . X ) = . we shall also de ne a sequence of stable matrices fAi g and a sequence of matrices fFi g. we shall de ne a nonincreasing sequence of hermitian matrices fXi g.
1 Xn X: We have a nonincreasing sequence fXi g.1 ) (Fn . Xf = X+ .Fn Fn . : .1 . To establish the stability property of the maximal solution.1).1 . Q: (13:25) Then Xn is hermitian. Fn.1 . note that An is stable for any n. X )An + An (Xn . we have Xf X 8 X such that Q(X ) 0: In particular. Extreme Solutions and Matrix Inequalities 337 Since it is assumed Xn.A X . Next.1 )x = 0: An. Xn ) = . Now we introduce Xn as the unique solution of the Lyapunov equation Xn An + A Xn = . Fn..1 ) 0: Since An is stable. we have Hence the limit Xn. So we have x (Fn .1 x = (A + BFn.1 )x = (A + BFn )x = An x = x which is a contradiction with the stability of An. each term on the righthand side of the above equation has to be zero.Q(X ) . i. and we have Xf X . we have ^ ^ (Xn . Fn. X ) = . and the sequence is bounded below by Xi X . Xf := nlim Xn !1 exists and is hermitian. Fn.1 ) (Fn . we get Q(Xf ) = 0. So Xf is a solution of (13.1 )x = 0: This implies But now (Fn . Xf is the maximal solution of the Riccati equation (13. Passing the limit n ! 1 in (13. Hence.23).13.3. the eigenvalues of A . by using (13.1). BB Xf will have nonpositive real parts.1 X .X. The uniqueness follows from the fact that the maximal solution is unique. (ii): The results follow by the following substitutions in the proof of part (i): A . Hence An is stable as well. in the limit. Fn Fn 0 and. Since X is an arbitrary element satisfying Q(X ) 0 and Xf is independent of the choice of X .e. Xn )An + An (Xn.(Fn . Fn. (Xn.25).X X+ .
Now suppose X+ . Therefore. 2 Remark 13. (iv): We shall only show the case for X+ the case for X. X )(A + RX ) + (X+ . X )R(X+ . > 0 i (A + RX. X ) = . a contradiction. X )A+ + A+ (X+ . ) C + follows by analogy. X+ > X.1 k =speci ed accuracy. the following procedures can be used: (i) nd F0 such that A0 = A + BF0 is stable (ii) solve Xi : Xi Ai + Ai Xi + Fi Fi + Q = 0 (iii) if kXi . . Xi. X. .26) by x and postmultiplying by x. Note that from (i) we have (X+ . we conclude that A+ is stable. )) is controllable. X.1). X )x = 0. X. follows by analogy. This procedure will converge to the stabilizing solution if the solution exists. we get x Q(X )x = 0. The stability of A + RX+ then follows from the Lyapunov theorem. X ) = 0: Since X+ . X = . Z 1 0 e(A+RX+ )t Re(A+RX+ ) t dt . )R(X+ . For example. X > 0. X. ) = 0: ()): Suppose X+ . . Otherwise go to (iv) (iv) let Fi+1 = . X. X. X. It is easy to verify that A+ (X+ . Therefore. from Lyapunov theorem.1 > 0: A simple rearrangement of terms gives (A + RX ) (X+ . That X+ . > 0 then (A+ R(X+ . X. to nd the maximal solution. ) + (X+ . X ) + (X+ .5 The proof given above also gives an iterative procedure to compute the maximal and minimal solutions. X ) + (X+ . A+ (X+ . By premultiplying (13. > 0 i (A+ ) C . X 0. we conclude (A + RX ) C + . (X+ . X ) = 0: This equation has an invertible solution X+ . stop. X )R(X+ . X is singular and there is an x 6= 0 such that (X+ . ~ . X ) < 0 (13:26) and X+ .338 ALGEBRAIC RICCATI EQUATIONS (iii): The existence of X+ and X. Let A+ := A + RX+ we now show that X+ . ((): Assuming now that A+ is stable and that X is any other solution to the Riccati equation (13. X )A+ . )A+ . (X+ .Q(X ) + (X+ . This in turn implies X = X. follows from (i) and (ii). X )R(X+ .B Xi and Ai+1 = A + BFi+1 go to (ii).
The following example illustrates that the stabilizability of (A R) is not su cient to guarantee the existence of a minimal hermitian solution of equation (13.A R) is stabilizable. X. Example 13. 1 j j2 2 # 2 C: The maximal solution is clearly X+ = 1 0 0 0 however. there is no minimal solution. Furthermore. X+ X 8 X such that Q(X ) 0 and (A + RX+ ) C +. which will be considered in the later part of this book. X2 are two solutions to the Riccati equation (13.1). ) C .13. However. 3 The Riccati equation appeared in H1 control.1 " " # R = . Then X1 > X2 implies that X+ = X1 . to the Riccati equation (13.1 " . then there exists a unique maximal solution X+ to the Riccati equation (13. (A + RX1) C . often has R 0. . Corollary 13.13 Assume that R 0 and that there is a hermitian matrix X = X such that Q(X ) 0.1).1) are given by 1 0 0 0 # . .3. (ii) If (. then there exists a unique minimal solution X. = X2 . these conditions are only a dual of the above theorem. X 8 X such that Q(X ) 0 and (A + RX.1). (i) If (A R) is stabilizable.12 Let R 339 0 and suppose (A R) is controllable and X1 . Furthermore.1). and that (A + RX2 ) C + .1 0 0 0 " " # Q= 1 0 : 0 0 # " # Then it can be shown that all the hermitian solutions of (13.5 Let A= 0 0 0 . X. Extreme Solutions and Matrix Inequalities Corollary 13.
B Rs 1 S ) .1S~ ~ ~ ~. (A + RX.(A .11.B A . ~ . ~ ~ ~ ~ ~. and Rs are real symmetric and Rs > 0.(A . The proof is similar to the proof for Theorem 13. = = Z Z 1 0 0 e(A+RX. Rs > 0.~ ~Rs1 ~ ) .1 K := JHs = P . S Rs 1 S ) = 0: (13:28) ~ Recall that JHs and J Hs are hermitian where J is de ned as " # 0 . X B Rs 1B X + (P . We shall also make use of the following matrices: ~ ~ P S ~ T := P S T := ~ ~ : S Rs S Rs ~ We denote by X+ and X+ the maximal solution to the Riccati equation associated with ~ s . ~ ~ ~ ~ ~ ~. (A + RX+ ) C + . i (A + RX. ) t dt e(A+RX+ )t Re(A+RX+ ) t dt . respectively: Hs and H . XBRs 1B X + (P .1 Hs := A . B Rs 1S ) ~ ~ ~ where P P Rs . BRs 1 S ) X + X (A . BRs . . ) C . B Rs 1 S ) X + X (A . BRs. ~ ~.1S ~ ~ ~ ~ K := J Hs = P . ~ A . ) C .1 ~ ~ B ~ . the results in (i) and (ii) can be respectively strengthened to X+ > X . then both X+ and X.BRs 1 B . ~ .BRs " ~ ~ ~ . B~ ~ .1 (iv) If Q(X ) < 0. . In this case. BRs1 S ) .1 : Proof. . . B Rs S ~ ~. BRs 1 S ) .P + S Rs S . BRs 1 S ) and ~ ~ Rs 1 ~ ~. and X.P + SRs S " " 2 # Theorem 13.I J= : I 0 Let .B Rs 1 B ~ Hs := A~. SRs 1S ) = 0 (13:27) ~ ~ ~. exist.S . Furthermore. More speci cally.1 (iii) If (A R) is controllable.340 ALGEBRAIC RICCATI EQUATIONS . SRs 1S (A . i (A + RX+) C + .1 . (A . S Rs 1S (A . < X . X+ .11 can be used to derive some comparative results for some Riccati equations.BRs B " # # " # " # # : . X.1S . ~ . ~ ~ ~ ~. ~ (A . . X+ > X. . . )t Re(A+RX. . let .
28). . 341 ~ ~ (i) Assume that (13. and X+ 0. so we have Qs (0) = P .28) has a hermitian solution and that K K (K > K ) then ~ ~ ~ X+ and X+ exist. SRs 0. B = B and let X be a hermitian solution of (13.11 to get the existence of X+ . Denote .Rs Qs (X ) = I X while (13. and X+ X+ (X+ > X+ ). and X+ 0 (X+ > 0). and X+ X . K ) 0: X Now we use Theorem 13. L Rs L + P = 0: ~ I Qs (X ) = (K . T ) I + (L . ~ ~ solution. Assume that (13. " Proof.1 S (iii): The condition T 0 implies P . ~ X+ X+ : ~ ~ (ii): Let A = A. P . L) R (L . Then ~ ~ ~ L = . Since X is ~ arbitrary.11. ~ ~ ~ ~ (ii) Let A = A. and X+ X+ (X+ > X+ ). 2 I X # " # . Moreover. (i): Let X be a hermitian solution of (13:28) then we have Then.Rs .1 (S + B X ) and L = . and T T (T > T ). we get X+ X for all hermitian solutions X of (13.28) becomes It is easy to show that " " # K I = A X + XA .28). B = B . and X+ X+ .13. K ) X ~ ~ = P . L) 0: ~ ~ ~s ~ = L L ~ ~ Now as in part (i).1 (S + B X ). we have " # " # " I K I = I Qs (X ) := X X X I X # ~ K I = 0: X # " # " # ~ I (K . since K ~ K . there exist X+ and X+ . Apply Theorem 13.14 Suppose (A B) and (A B) are stabilizable. . SRs 1 S 0.3. Then X+ and X (iii) If T 0 (T > 0).28) has a hermitian ~+ exist.11 to obtain the existence of X+ . The existence of X+ follows immediately by applying Theorem 13. then X+ exists. L Rs L " L # s L +~ R ~ " # I (T . Extreme Solutions and Matrix Inequalities ~ ~ Theorem 13. L Rs L + P X " # ~ ~ ~ ~ A X + XA .
1B H = A . i.1 B " i .1 R.15 Suppose R is nonsingular and either one of the following assumptions is satis ed: (A1) A has no eigenvalues on j!axis (A2) P is sign de nite.1 (s).1 (s) = 6 6 H R. A).S R 3 7 5 : . (i) ) (ii): Let ^ A (s) = ^ C ^ ^^ ^ Then H = A . BR.1 C and 2 4 h " # A 0 ^ B := 6 .P .1 P (sI . P 0 or P 0.. has an eigenvalue at j!0. A ). BR .BR.(P . .e.1S R.29) S R I Lemma 13.(A .1 B + B (. then j!0 is a zero of (s).sI . Then the following statements are equivalent: (i) (j!0 ) is singular for some 0 !0 (ii) The Hamiltonian matrix . R = R . A ). (A B ) has no uncontrollable modes on j!axis.sI .1 # 3 7 7 5 : If (j!0 ) is singular.BR.1 I : (13. A ).1 . B D. A).342 ALGEBRAIC RICCATI EQUATIONS 13.1 B " #" # h i P S (sI .4 Spectral Factorizations Let A B P S R be real matrices of compatible dimensions such that P = P . A).1 SR.sI .A 4 ^ D 2 S B B .1 S ) " # 1. and j!0 is an eigenvalue of H .1 S + B (. and (P A) has no unobservable modes on the j!axis. Proof.1 B = B (.1S . Hence j!0 is a pole of . and de ne a parahermitian rational matrix function (s) = R + S (sI . SR S ) .
(A . . Then the following statements are equivalent: (i) (j!) > 0 for all 0 ! (ii) The Hamiltonian matrix .1 (s). Then i j!0 must be either an unobservable mode of ( R.15 is true. in turn.16 Suppose R > 0 and either one of the assumptions (A1) or (A2) de ned in Lemma 13.30) that j!0 is an unobservable mode of (P A) if x1 6= 0.1 SR Corollary 13.1 S h R. BR .1 S ) " # 1. This gives Px1 = 0 since P is sign de nite (P 0 or P 0).Px1 S x1 + B x2 = 0: We now consider two cases under the di erent assumptions: (a) If assumption (A1) is true.13. and this.1S . along with (13.BR. SR S ) . Similarly.31) we have x1 Px1 = 0. (b) If assumption (A2) is true. Spectral Factorizations " # 343 (ii) ) (i): Suppose j!0 is an eigenvalue ofhH but is not a pole of . a contradiction will also be derived if j!0 is assumed to be an uncontrol" # . implies x2 = 0 from (13. if x1 = 0.(P .30) implies x1 (j!0 I + A ) = 0. from (13. has no eigenvalue on j!axis. Then there exists an x0 = x1 6= 0 such that x2 Hx0 = j!0 x0 h SR " # R .1 B i x0 = 0: (13:30) (13:31) (13:32) These equations can be simpli ed to (j!0 I . A)x1 = 0 (j!0 I + A )x2 = . since (13.BR.1 ). Now suppose j! is an unobservable mode of trollable mode of (H 0 . again a contradiction.1 ).1 B i H ).1B H = A . lable mode of (H 2 .31).30). then x1 = 0 from (13.1 S R.32) imply that (A B ) has an uncontrollable mode at j!0 .1S R.1 .4.1 B H ) or an uncon .31) and (13. On the other hand. then (13.BR. BR. which is a contradiction.1 ( R. This implies.
15.I 0: # I we have Ric A . 2 G (s)G(s): .BB .C C and G(s) := C (sI .P .A " # .6 that (13.19 below.16. satis es Proof.I = # " #" I A BB P . A ). A).17 Suppose A is stable and P 0.A A BB = .BB . (i) ) (ii) follows easily from Lemma 13. (i): ()) From Corollary 13.P . Then the continuity of (j!) gives (j!) > 0.A has no eigenvalues on the imaginary axis. Then B (.15. (ii) 0 (j!) 0 for all 0 ! 1 if and only if there exists a unique real X = X 0 such that A X + XA . BB X ) C . This in turn implies from Theorem 13.1 B # I for all 0 ! 1 if and only if there exists a unique real X = X 0 such that A X + XA . Then 0 (s) = I .344 ALGEBRAIC RICCATI EQUATIONS 2 Proof.P .33) has a stabilizing solution.1 B . (i) the matrix 0 (s) = h Lemma 13. A). XBB X + P = 0 and (A . . since " A . note that (j 1) = R > 0 and det (j!) 6= 0 for all ! from Lemma 13. 0(j!) > 0 implies that a Hamiltonian matrix " # A . Let 0 < < 1 and de ne (s) := I . BB X ) C . (ii): ()) Let P = .sI .1 I 0 (j! ) > 0 i " P 0 0 I #" (sI . Furthermore. To prove (ii) ) (i). XBB X + P = 0 (13:33) and (A .BB .A # (() The su ciency proof is omitted here and is a special case of (b) ) (a) of Theorem 13. .A " #" .Ric P . G (s)G(s).
Then it is easy to verify that (A . and then assume without loss of generality that " # " # h i A11 A12 B1 A= B= C = C1 C2 0 A22 0 so that (A11 B1 ) is controllable and A11 and A22 are stable. BB X is stable and A X + X A . Y ) + 2 C C ]e. it will be assumed that (A B ) is controllable. Y )BB (X .14 that X is monotonedecreasing with .Y = Z 1 Thus X is bounded below with Y as the lower bound. there is an X = X 0 such that A .(A. Y )(A . and lim !1 X exists.(A. Spectral Factorizations 345 Then (j!) > 0 8!. (X12 ) B1 B1 X12 . Y ) + (X . BB Y ) (X .13. X BB X . BB Y ) C + (note that the existence of such a solution is guaranteed by the controllability assumption). X satis es the Riccati equation 0 e.BB Y )t dt 0: A X + XA . To show that lim !1 X exists. BB X ) C . Y )BB (X . Let X := lim !1 X then from continuity argument. Thus from part (i). The controllability assumption will be removed later. we need to show that X is bounded below for all 0 < < 1. XBB X + P = 0 and (A . X 1 X 2 if 1 2 . 2 C1 C2 = 0 A22 X22 + X22 A22 + (X12 ) A12 + A12 X12 .4. Y BB Y = 0 with (A . B1 B1 X11 ) X12 + X12 A22 + X11 A12 . BB Y ) . (X . Now suppose (A B ) is not controllable.e. 2 C1 C1 = 0 (A11 .BB Y ) t (X . Then the Riccati equation for # " X11 X12 X = (X12 ) X22 can be written as three equations A11 X11 + X11 A11 . Y ) . . 2 C2 C2 = 0 . i.. 2 C C = 0: It is easy to see from Theorem 13. 2C C = 0: This implies that X . In the following. X11 B1 B1 X11 . Let Y be the destabilizing solution to the following Riccati equation: A Y + Y A .
C1 C2 = 0 has a unique solution X12 since i (A11 . (() same as in part (i). Consequently.sI . C2 C2 = 0: We have proven that there exists a unique X such that A X + XA . Let Y11 be the antistabilizing solution to A11 Y11 + Y11 A11 .1=2 P := P . (s) = B (. the following Lyapunov equation has a unique nonnegative de nite solution X22 : A22 X22 + X22 A22 + X12 A12 + A12 X12 . A). A ). SR. A ). C C = 0 and (A . BB X = A11 . Moreover.346 and " ALGEBRAIC RICCATI EQUATIONS A .1 I and where h i " P S S R ^ P 0 0 I #" (sI .sI . Y11 B1 B1 Y11 = 0 with (A11 . BB X ) C . . the following Sylvester equation (A11 . 2 Lemma 13. X12 B1 B1 X12 . C1 C1 = 0 with (A11 . Y11 lim !1 X11 exists and satis es the following Riccati equation: 0. Furthermore.1 S B := BR.1 B # I ^ ^ ^ A := A .1 B # I ^ (s) = B (. X11 B1 B1 X11 . B1 B1 X11 ) X12 + X12 A22 + X11 A12 .1 S : Then (j!) 0 i ^ (j!) 0. A).18 Let R > 0. BR.. Then it is clear that X11 . B1 B1 Y11 ) C + . B1 B1 X11 A12 .1 I ^ ^ h i " #" ^ ^ (sI . B1 B1 X11 ) C . B1 B1 X11 ) + j (A22 ) 6= 0 8i j . XBB X . X11 = A11 X11 + X11 A11 . . B1 B1 X12 0 A22 # is stable.
Theorem 13. Then (I) The following statements are equivalent: (a) The parahermitian rational matrix P .1 S ) . Note that " 347 #" P S = I SR.1 B '(s) with '(s) = R1=2 + R. BR. A).13. 2 Now we are ready to state and prove one of the main results of this section.1 S .1 (s)] (s)'.1=2 S R 0 R1=2 h # " ^ P 0 0 I i " #" R. It is easy to verify that ^ (s) = '.4. SR.1=2 S (sI .1 (s): Hence the conclusion follows. A ).1 B X is stable. Spectral Factorizations Proof.e. Suppose either one of the following assump (s) = B (. and decomposition of Hamiltonians. A ). BR.1=2 S #" I R 1= 2 0 # : # Hence the function (s) can be written as (s) = B (.1 ' (s) ^ P 0 0 I (sI .1 B # I (j!) > 0 for all 0 ! 1: (b) There exists a unique real X = X such that (A . The following theorem and corollary characterize the relations among spectral factorizations. A).1S = 0 and that A . .. BR. R = R > 0. Riccati equations. with (A B ) stabilizable.sI .1 S ) X + X (A .1 I satis es h i " P S S R #" (sI . BR.sI . A). XBR.19 Let A B P S R be matrices of compatible dimensions such that P = tions is satis ed: (A1) A has no eigenvalues on j!axis (A2) P is sign de nite. P 0 or P 0 and (P A) has no unobservable modes on the j!axis.1 B .1B X + P . i.
1(S + B X ) and M= A B : . SR S ) . XBR. (b) ) (a) Suppose 9X = X such that A . BR. Corollary 13. respectively. BR.1 B X ) C + ) and the weakly destabilizing solution ( (A . BR. then the theorem still holds except that the solutions X in (b) and (e) are changed into the destabilizing solution ( (A . BR.1(S + B X ). (e) There exists a unique real X = X such that (A .1 (S + B X ) is stable. in Theorem 13.1S = 0 and that (A . ~ Proof.6 If the stabilizability of (A B) is changed into the stabilizability of (. BR.1 B X ) C .1 S ) .1 B X ) C + ). Let F = . BR .1 S ) X + X (A . then there exists an M 2 Rp such that = M RM: A particular realization of one such M is M= A B .19 is satis ed. BR. (c) ) (b) follows from Theorem 13.F I " # where F = . if X is the stabilizing solution.BR.5. BR.20 If either one of the conditions.1 .1 B X = A .R.6 and Theorem 13.R.(A .F I " # .1 S .(P . then M . (a){(e). (a) ) (c) follows from Corollary 13. BR. (II) The following statements are also equivalent: (d) (j!) 0 for all 0 ! 1. BR.1 S ) # has no j!axis eigenvalues.16.1S .348 (c) The Hamiltonian matrix " ALGEBRAIC RICCATI EQUATIONS .1B H = A .1 S .1 S .1 S . BR. Furthermore. BR..1B X + P .A B). SR. BR.1 2 RH1 . Remark 13. The following corollary is usually referred to as the spectral factorization theory.
e. X1 BB X1 + C1 C1 = 0: By Theorem 13.1 B From the de nition of X1 and Corollary 13.17 and note that A1 is stable then it is su cient to show that I .1 : I + B (. A ). (e) ) (d) follows the same procedure as the proof of (b) ) (a). B (. Since " # A + BF B M .j!I .1 = M . A1 := A .1 B It follows that . by assumption I + B X1 (sI . A1 ).j!I . Let = X . BB X1 is stable. Let P = C1 C1 .20.13.1 B . BB . A ).1 B (.1 B Now.j!I . A). C2 C2 = 0: To show that this equation has a solution.1 B is positive semide nite. i.1 B I + B X1 (sI . A).4.1 C2 C2 (j!I .j!I . we also have that = I + B X1 (sI . Note that this decomposition always exists since P = I . Then the equation in becomes A1 + A1 . Let X1 be the positive de nite solution to F I A X1 + X1 A .1 B = C2 (sI .1 C2 n o C2 (j!I .18 rst to get a new function with such properties. Thus M (s) has no zeros on the imaginary axis and (j!) > 0. A). (d) ) (e): Assume S = 0 and R = I otherwise use Lemma 13.1 C2 C2 (j!I . A ). A).1 B : (j!) = I + B (.7.1 X1 B . P ).1 B I + B X1 (j!I .1 0: .. A).1 B 0: n o I . A + BB X1 ). C2 C2 with C1 and C2 square nonsingular.1 C1 C1 (j!I .j!I . X1 .1 C1 C1 (sI . A ). A).1 B . I + B (. with > 0 su ciently large. recall Lemma 13. A1 ). A). Spectral Factorizations 349 It is easily veri ed by use of the Riccati equation for X and by routine algebra that = M RM . ( I .B (.sI .1 2 RH1 . de nes one such possibility. A). A ). A). X1 indeed exists and is stabilizing. Notice rst that C2 (sI .
. Corollary 13.1 D )C . Note in particular that the function can have poles on the imaginary axis. G G and M .(A + BR. Then (s) = 1 .21 Assume that G(s) := 2 RL1 is a stabilizable and detectable realization and > kG(s)k1 .1 D C ) and X # 0 if A is stable. BB X ) C . Consequently. This shows the existence and uniqueness of X . ALGEBRAIC RICCATI EQUATIONS I .R where " # " A B C D # R = 2I .j!I . .1"B X + D C ) ( A + BR.1 2 RH1 .350 Consequently.1 D C BR. BB ) = (A .1B X = Ric .1 . Then.1 B 0: We may now apply Lemma 13. D D F = R. there exists a unique solution such that (A1 .C (I + DR. A1 ). s12 and (j!) > 0. A1 ). 2 We shall now illustrate the proceeding results through a simple example.17 to the equation. B (. A particular realization of M is B A M (s) = 1=2 F R1=2 . It is easy to check that the Hamiltonian matrix H = 0 . Example 13.1 C2 C2 (j!I .1 0 " # " # does not have eigenvalues on the imaginary axis and X = 1 is the stabilizing solution to the corresponding Riccati equation and the spectral factor is given by M (s) = 0 1 = s + 1 : s 1 1 3 Some frequently used special spectral factorizations are now considered.6 Let A = 0 B = 1 R = 1 S = 0 and P = 1. there exists a transfer matrix M 2 RL1 such that M M = 2 I .
1 D C BR.1 B . In fact. DD L = (Y C" + BD )R. j! B has full column rank for (a) Suppose G (j!)G(j!) > 0 for all ! or C D all !. " # C R. This is a special case of Theorem 13.1 .1 2 RH1 .23 Let G(s) = A B be a stabilizable and detectable realization.4. DR.(A + BR.1 B X = Ric . Spectral Factorizations 351 Proof. Let " # A .1 D )C . BR.D C 1 D )C . D D in Theorem 13.C D R = 2I .19 and by using the fact that A + BR.1D )C .1 .BR.1 D C . the theorem follows by letting P = .22 Assume G(s) := A B 2 RL1 and > kG(s)k1.C (I + DR " " # = # .19.1 D C ) " # .13. GG and M .1 D C ) . we also include the following spectral factorization results which are again special cases of Theorem 13.LR 2 1= C R " # " # where and Y R = 2 I .B (I + D R D)B 0 if A is stable. Corollary 13.1 C ) # For convenience. C D A .BR.1 Y = Ric (A + BD R.(A + BR.Ric A + BR . A particular realization of M is 1=2 M (s) = A .C C S = .(A + BD R. Then. Corollary 13.C (I .1 C ) .1 B Ric . there C D exists a transfer matrix M 2 RL1 such that MM = 2 I . BR.1 C .1 D C ) C (I + DR : 2 The spectral factorization for the dual case is also often used and follows by taking the transpose of the corresponding transfer matrices.(A .19.
24 Let > 0. j! B has full row rank for all (b) Suppose G(j!)G (j!) > 0 for all ! or C D !. BD R. Then we have the following spectral factorization W W =G G where W .1 C ) . G(s) = A B 2 RH1 and C D A + BR. " # Corollary 13.C (I + DR. (iii) (D) < and H 2 dom(Ric) .1=2 A R (D C + B X ) R1=2 A .B (I . BD R. Then we have the following spectral factorization ~ ~ W W = GG ~ where W . Then the following conditions are equivalent: (i) kGk1 < . " # " # 0 (Ric(H ) > 0 if (C A) is observ .1 C ) ~ with R := DD > 0.1 C Y = Ric ~ ~ .352 ALGEBRAIC RICCATI EQUATIONS with R := D D > 0. (ii) (D) < and H has no eigenvalues on the imaginary axis.1 D)B .1 D C ) where R = 2 I .(A .19 also gives some additional characterizations of a transfer matrix H1 norm.(A + BR. D R. (iv) (D) < and H 2 dom(Ric) and Ric(H ) able). Let " # ~ ~ (A . D D.1 2 RH1 and B : W = .1D )C .1 2 RH1 and " # " # Y ~ .1=2 : ~ W = A (BD + ~ 1C2 )R R= C Theorem 13.1 D C BR.1 B H := .C R.
I 0 = 0 " I 0 I A + BR. " # .1 D C ) : 2 The equivalence between (i) and (iv) in the above corollary is usually referred as Bounded Real Lemma. ! 2 R. Positive Real Functions function is positive de nite for all !: " 353 Proof. if M (j!) has full column rank for all ! 2 R.1 B C (I + DR.13.C C . This follows from the fact that kGk1 < is equivalent to that the following (j!) := 2 I .5 Positive Real Functions A square (m m) matrix function G(s) 2 RH1 is said to be positive real (PR) if G(j!) + G (j!) 0 for all nite !.j!)G(j!) = B (. and W such that XA + A X = .I 0 H .j!I .C D 2I .25 Let A B be a state space realization of G(s) with A stable (not C D necessarily a minimal realization).Q Q B X +W Q = C D+D = W W Then G(s) is positive real and " # (13.34) (13..1 I and the fact that " # " # h i . GT (. 13. with M (s) = Q W then G(s) is strictly positive real.(A + BR. and G(s) is said to be strictly positive real (SPR) if G(j!) + G (j!) > 0 for all ! 2 R. A ).1 B # I >0 # . Theorem 13. i.36) G(s) + G (s) = M (s)M (s) A B . A).D C .35) (13.BR. Suppose there exist an X 0.1 D )C .e.5. Furthermore. Q.1D C . D D #" (j!I .
Q = B W Q W A 0 This implies that 2 3 B A 0 7 6 . Since G(s) is assumed to be positive real.A W Q B W W 2 3 7 5 G(j!) + G (j!) = M (j!)M (j!) 0 i. then M (j!)x 6= 0 for all x 2 C m and ! 2 R. then M (j!) Q W given above has full column rank for all ! 2 R..A B X +W Q B A 0 .A .26 Suppose (A B C D) is a minimal realization of G(s) with A stable and G(s) is positive real. Thus G(s) is strictly positive real.Q W 5 = 4 . 2 Theorem 13.Q W W Q B W W #" # " A B : .(XB + Q W ) 7 : 5 W W B 3 I 0 to the last realization to get Apply a similarity transformation X I B 6 G(s) + G (s) = 4 XA + A X . Furthermore.C C B D+D 0 " A 3 7 5 2 =6 4 # 0 . and W such that XA + A X = . G(s) is positive real.354 ALGEBRAIC RICCATI EQUATIONS 2 Proof.A . Proof.Q Q B X +W Q = C D + D = W W: and with M (s) = " # G(s) + G (s) = M (s)M (s) A B . if G(s) is strictly positive real. there exists a transfer matrix M (s) = # " A1 B1 C1 D1 with A1 stable such that G(s) + G (s) = M (s)M (s) . B 6 G(s)+G (s) = 4 0 . Finally note that if M (j!) has full column rank for all ! 2 R. Q.e.A .Q Q . Then there exist an X 0.
1 D + D = D1 D1 : Now the conclusion follows by de ning X = (T . M (s) = 1 .1) X1 T .(B1 X + D1 C1 ) : = D1 D1 + B1 X + D1 C1 0 B1 0 But the realization for G(s) + G (s) is given by # " # " A B + . Now let X1 following Lyapunov equation: X1 A1 + A1 X1 = .A .1 B = TB1 C = (B1 X + D1 C1 )T .27 Let A B be a state space realization of G(s) 2 RH1 with A C D stable and R := D + D > 0. BR. B X ) R 2 is minimal phase and 2 R G(s) + G (s) = M (s)M (s): " A # .13. Positive Real Functions where A and A1 have the same dimensions.1 B X + C R.A1 .(X1 B1 + C1 D1 ) 7 4 5 B1 X + D1 C1 B1 D1 D1 " # " # A1 B1 + .C1 D1 7 5 4 D1 C1 B1 D1 D1 2 3 A1 0 B1 = 6 0 .1 " # 2 B1 A1 0 6 = 4 .1 C ) X + XBR.C1 C1 : Then 355 0 be the solution of the M (s)M (s) = " .A1 .1: 2 Corollary 13.1 C ) + (A .C1 C1 .A1 . there exists a nonsingular matrix T such that A = TA1T .C1 B1 D1 A1 #" A1 B1 C1 D1 0 # B1 = 6 X1 A1 + A1 X1 .A1 .C1 D1 D1 C1 B1 D1 D1 3 2 3 7 5 W = D1 Q = C1 T .A1 .C : G(s) + G (s) = D + D + C 0 B 0 Since the realization for G(s) is minimal.1 C = 0: B Moreover.5. Then G(s) is strictly positive real if and only if there exists a stabilizing solution to the following Riccati equation: X (A . 1 (C . BR.
This follows from Theorem 13. there is an M (s) 2 RH1 such that M . . " # 2 Corollary 13.27 and the fact that G(j!)G (j!) > 0 and G(s)G (s) = A BW C 0 . This corollary follows from Corollary 13. BW X ): " # Proof. The above corollary also leads to the following special spectral factorization. BW X ) (DD ).19 and from the fact that for all ! including 1. Let P be the controllability grammian of (A B ): PA + AP + BB = 0: De ne Then there exists an X BW = PC + BD : 0 such that XA + A X + (C .1 (C .28 Let G(s) = A B 2 RH1 with D full row rank and G(j!)G (j!) > C D 0 for all !. BW X ) = 0: Furthermore.C + BW 0 " # + DD : 2 A dual spectral factorization can also be obtained easily.A .1 (s) 2 RH1 and G(s)G (s) = M (s)M (s) A BW where M (s) = CW DW and " # with a square matrix DW such that DW DW = DD CW = DW (DD ).1 (C .356 ALGEBRAIC RICCATI EQUATIONS G(j!) + G (j!) > 0 Proof.
C C . and N N = D D implies D C + B X = 0.1 " 2 3 7 7 5 I 0 on the right yields = X I 3 7 7 5 # N N = = 6 6 4 2 6 6 4 A 0 B .(XB + C D) 7 : 5 B X +D C B DD Then (a) and (b) follow easily. v 2 L2 . inner matrices will play an important role in the control synthesis theory in this book.13. Inner and coinner are dual notions.A .(A X + XA + C C ) . Conjugating the states of B A 0 6 6 . Note that N inner implies that N has at least as many rows as columns.6. Note that N need not be square. Because of these norm preserving properties. (b) (A B ) controllable. For N inner and any q 2 C m .A . kN (j!)qk = kqk.C D N N =4 DC B DD I 0 on the left and I 0 by . N is an inner i N T is a coinner. A matrix function N 2 RL1 is called allpass if N is square and N N = I clearly a square inner function is allpass. i.29 Suppose N = A B 2 RH1 and X = X C D A X + XA + C C = 0: " # 0 satis es (13:37) Then (a) D C + B X = 0 implies N N = D D. Inner Functions 357 A transfer function N is called inner if N 2 RH1 and N N = I and coinner if N 2 RH1 and NN = I . Proof. 13. 8! and kNvk2 = kvk2 since N (j!) N (j!) = I for all !.e.(XB + C D) B X +D C B DD 3 A 0 B 7 0 . we present a statespace characterization of inner transfer functions.X I 2 " # " #. We will focus on the characterizations of inner functions here and the properties of coinner functions follow by duality.6 Inner Functions Lemma 13. In this section.X I . 2 .A ..
29 to get N N = I . We simply need to use the formulas in Theorem 5. 13. Simply add the condition that D D = I to Lemma 13.358 ALGEBRAIC RICCATI EQUATIONS This lemma immediately leads to one characterization of inner matrices in terms of their state space representations.e. The former factorization in the case of G 2 RH1 will give an innerouter factorization1. we note that if NM .1 is an rcf. i. The proof of this lemma follows from straightforward calculation and from the fact that CX y X = C since Im(I . 1 A p m (p m) transfer matrix Go 2 RH1 is called an outer if Go (s) has full row rank in the open right half plane. To obtain an rcf of DR G with N inner. respectively. X + X ) Ker(X ) Ker(C ). .X C D? C D? " # " # where D? is an orthogonal complement of D such that D D? ] is square and orthogonal..9 to solve for F and Zr . The results will be proven for the right coprime factorizations. Lemma 13. The dual notion of the complementary coinner factor is de ned in the obvious way.1 with inner numerator N and inner denominator M . Given an inner N . C D Then a CIF N? is given by y N? = A . Go (s) has full row normal rank and has no open right half plane zeros. Then N is inner if and only if (a) D C + B X = 0 (b) D D = I . and X is the ob A transfer matrix N? is called a complementary inner factor (CIF) of N if N N? ] is square and inner. while the results for left coprime factorizations follow by duality. In particular. the following lemma gives a construction of its CIF. let D? denote for any orthogonal complement of D so that h . some special form of coprime factorizations will be developed.1=2 D? (with R = D D > 0) is square and orthogonal. Let G 2 Rp be a p m transfer matrix and denote R1=2 R1=2 = R. explicit realizations are given for coprime factorizations G = NM . A B C D # is stable and minimal.1 is also an rcf for any nonsingular real matrix Zr . then (NZr )(MZr ). Corollary 13.31 Let N = A B be an inner and X be the observability grammian.7 InnerOuter Factorizations In this section. For a given full column rank i matrix D.30 Suppose N = " servability grammian.
(() This can be shown by using Corollary 13. It will be proven by showing that the de nition of the inner and coprime factorization formula given in Theorem 5.(A .1=2 3 7 5 2 RH1 where and R=D D>0 F = .1 D C ) y N? = A + BF .1 > 0 on the j!axis since M 2 RH1 .R.1 such that N is inner if and only if G G > 0 on the j!axis. That G = NM .1B X = Ric . # assume that the realization of " " M := 6 4 N # 2 A + BF F C + DF BR.9 lead directly to the above realization of the rcf of G with an inner numerator. Then there exists an rcf G = NM . Furthermore.1 ) (M .1(B X + D C ) A . BR. ()) Suppose G = NM . including at 1. This factorization is G = A B is stabilizable and that A .1 D )C .7. Now suppose Proof. j!I B has full column rank for all C D C D ! 2 R. (M .1=2 R. InnerOuter Factorizations 359 Theorem 13.1) M .BR. Then a particular realization of the desired coprime factorization is " unique up to a #constant unitary multiple.20 rst to get a factorization G G = N = A + BF BZr C + DF DZr and let F and Zr be such that (DZr ) (DZr ) = I " # (13:38) (13:39) (BZr ) X + (DZr ) (C + DF ) = 0 .1 is an rcf and N N = I .1 is an rcf follows immediately once it is established that F is a stabilizing state feedback. DR. The following proof is more direct.1=2 DR.1 ) and then to compute N = GM .32 Assume p m. Then G G = (M . a complementary inner factor can be obtained as if p > m. BR.C (I .1 D C .13.X C D? C + DF D? " # " # 0: Moreover.
33 Suppose G 2 RH1 then the denominator matrix M in Theorem 13. Suppose . that C D The uniqueness of the factorization follows from coprimeness and N inner. these two factorizations are unique up to a right multiple which is a unit2 in RH1 . the formula for N? follows from Lemma 13. 2 Note that the important innerouter factorization formula can be obtained from this inner numerator coprime factorization if G 2 RH1 . Take U = I and solve (13.31.1 of Chapter 6.1 2 RH1 .1 . we need to show that H 2 dom(Ric). BR. Remark 13.R.1=2 U where R = D D > 0 and where U is any orthogonal Clearly. BR. The innerouter factorization for the general transfer matrices can be done using the method adopted in Section 6. To show that such choices indeed make sense. ~ 2 A function .360 ALGEBRAIC RICCATI EQUATIONS (A + BF ) X + X (A + BF ) + (C + DF ) (C + DF ) = 0: (13:40) .(A . N1 N2 The only inner units in RH1 are constant matrices.32 is an outer. Corollary 13. XBR.1) given in Theorem 13.1B X + C D?D? C where D? D? = I .1 D C ) X + X (A . H 2 dom(Ric) is guaranteed by the fact = # A . Note that the nonuniqueness is contained entirely in the choice of a particular square root of R.32 is an innerouter factorization. For example.19.40) to get 0 = (A + BF ) X + X (A + BF ) + (C + DF ) (C + DF ) = (A .1 1 G(s) = s+1 s+2 has innerouter factorizations but the above procedure cannot be used.1 D C ) so X " Ric(H ). is called a unit in RH1 if . by Theorem 13. BR. and thus the desired uniqueness property is established.1 D . that G = N1M1 1 = N2 M2 1 are two right coprime factorizations and that both nu" # merators are inner.7 It is noted that the above innerouter factorization procedure does not apply to the strictly proper transfer matrix even if the factorization exists. where . Hence. there exists a unit 2 RH1 such that " # " # M1 M2 . Finally. BR D C . is inner since = = N1 N1 = N2 N2 = I .BR. j! B has full column rank (or G (j!)G(j!) > 0. That is. However. s.1B H = A .1 D C ) . Clearly.1(B X + D C ): Then substitute F into (13. .C D?D? C . we have that Zr = R matrix. DR. the factorization G = N (M . By coprimeness.39) for F to get F = .
D? B Y y D? ~ ~ ~ ~ if p < m.1 such that M 2 RH1 is inner if and only if G has no poles on j!axis.1 C Y = Ric (A .1=2 ~ .1=2 ~ . Furthermore. output injection using the dual Riccati solution replaces state feedback to obtain the corresponding left factorizations.(BD + Y C )R.1 C ) Moreover. j!I B has full row rank for all ! 2 R. InnerOuter Factorizations 361 Theorem 13. where D? is a full row rank matrix such that D?D? = I . G= C D C D Then a particular realization of the desired coprime factorization is " # h i A + LC L B + LD 2 RH ~ ~ M N := ~ .1 ~ ~ .1=2 1 R C R R D where and " F = .(A . Then there exists an lcf G = M . ~ ~ ~ Theorem 13.BB 0: 0 .1 # ~ .1D. a complementary coinner factor can be obtained as # " ~? = A + LC B + LD N ~ ~ .B (I .13.B X # " X = Ric A .C R. A particular realization is 3 2 " # B A + BF M := 6 F I 7 2 RH1 5 4 N C + DF D where Suppose that the system G is not stable then a coprime factorization with an inner denominator can also be obtained by solving a special Riccati equation. .A Dual results can be obtained when p m by taking the transpose of the transfer function ~ R = DD > 0 ~ L = . The proof of this result is similar to the inner numerator case and is omitted. assume that the realization of " # " A B is detectable and that A . D R D)B .1N such that N is a coinner if and only if GG > 0 on the j!axis. including at 1.35 Assume p m. In these factorizations.7. BD R. " # matrix. BD R .1C ) 0: ~ ~ .34 Assume that G = A B 2 Rp and (A B) is stabilizable. D R. This factorization is unique up to a#constant unitary multiple. Then C D there exists a right coprime factorization G = NM .
1=2 C + DF DR .8 Normalized Coprime Factorizations M N # i. Then C D ~ ~ ~ there exists a left coprime factorization G = M .362 " ALGEBRAIC RICCATI EQUATIONS # Theorem 13. Then there is a normalized right coprime factorization G = NM . The following theorem can be proven using the same procedure as in the proof for the coprime factorization with inner numerator. In this case. the proof involves choosing F " # M is an inner.1=2 7 2 RH1 5 4 N .Y C Y = Ric A ..1=2 A + BF M := 6 F R. an lcf G = M .37 Let a realization of G be given by G= A B C D and de ne " # N ~ R = I + D D > 0 R = I + DD > 0: (a) Suppose (A B ) is stabilizable and (C A) has no unobservable modes on the imaginary axis.1N is called a normalized left h i ~ ~ coprime factorization if M N is a coinner. A particular realization is " # h i A + LC L B + LD 2 RH ~ ~ M N := 1 C I D where L = .1N such that M 2 RH1 is inner if and only if G has no poles on j!axis. if ~ ~ is an inner.C C 0 .1 with N M 2 RH1 is called a normalized right coprime factorization if M M +N N =I " 13. The normalized coprime factorization is easy to obtain from the de nition.36 Assume that G = A B 2 Rp and (C A) is detectable. Similarly. and Zr such that Theorem 13.e.1 3 2 " # BR.A " # 0: A right coprime factorization of G = NM .
(A .1 .1 B = 0 or equivalently P = Ric " (A + BF ) . Then there is a normalized left coprime factorization G = M .R.C R.1D C .C R C . It is obvious that Q = X since the Riccati equation for X can be written as F X (A + BF ) + (A + BF ) X + C + DF " # " F C + DF # =0 while the controllability grammian solves the Lyapunov equation (A + BF )P + P (A + BF ) + BR.BR.BR.1 X: Proof.1 C ) # 0: # M are given (c) The controllability grammian P and observability grammian Q of N by " P = (I + Y X ).BR. BR~ .(A . Normalized Coprime Factorizations where and " 363 F = .1B ~ .1 (B X + D C ) . BD R.8.1Y Q = X h ~ ~ ~ ~ while the controllability grammian P and observability grammian Q of M N are given by i ~ P =Y ~ Q = (I + XY ).1 N h A + LC L B + LD ~ ~ M N := ~ .1=2 ~ .1B X = Ric A .(A + BF ) # . BR.1 D C ) # 0: (b) Suppose (C A) is detectable and (A B ) has no uncontrollable modes on the imag~ ~ inary axis. We shall only prove the rst part of (c).1C ) .1=2 ~ .1 Y = Ric " ~ (A .13.1B 0 : .1=2 R C R R D i " # where and ~ L = . BD R.1 C ~ .(BD + Y C )R.
1: This shows that the stable invariant subspaces for these two Hamiltonian matrices are related by X.1 C . The iterative procedure for solving ARE was rst introduced by Kleinman 1968] for a special case and was further developed by Wonham 1968]. It was used by Coppel 1974].BR.BR. and many others for the proof of the existence of maximal and minimal solutions.1 B ~ . . Some matrix factorization results are given in Doyle 1984]. .9 Notes and References The general solutions of a Riccati equation are given by Martensson 1971].1 C .364 ALGEBRAIC RICCATI EQUATIONS " # I X Now let T = 0 I " then # " (A + BF ) .C R. The paper by Wimmer 1985] also contains comparative results for some special cases. Numerical methods for solving ARE can be found in Arnold and Laub 1984].BR .1 C ) " # " # # I = Im I + XY = T Im Y Y Hence we have P = Y (I + XY ).1B I Im P " ~ ~ 0 A .BR .1 . BD R.(A + BF ) # ~ ~ A . Dooren 1981].1 B ~ . The comparative results were obtained in Ran and Vreugdenhil 1988]. Ran and Vreugdenhil 1988].1 B 0 =T . and references therein. BD R.(A + BF ) .C R.(A .1 C . # I : = Im Y (I + XY ). BD R.1 C ) " # T . BD R.1 C = T X. or " (A + BF ) .(A .1 " # 2 13. The state space spectral factorization for functions singular at 1 or on imaginary axis is considered in Clements and Glover 1989] and Clements 1993]. The paper by Willems 1971] contains a comprehensive treatment of ARE and the related optimization problems.
This is the socalled Regulator Problem. The material in this chapter is standard.. 14 14. 365 . However. this is not practical since any physical system has the energy limitation. i. the actuator will eventually saturate. Furthermore. This is indeed the case if the controller can provide arbitrarily large amounts of energy since. Hence certain limitations have to be imposed on the control in practical engineering implementation.e.H2 Optimal Control In this chapter we treat the optimal control of linear timeinvariant systems with a quadratic performance criterion. by the de nition of controllability. but the treatment is somewhat novel and lays the foundation for the subsequent chapters on H1 optimal control. one can immediately construct a control function that will drive the state to zero in an arbitrarily short time. One might suggest that this regulator problem can be trivially solved for any T > t0 if the system is controllable.1 Introduction to Regulator Problem Consider the following dynamical system: x = Ax + B2 u x(t0 ) = x0 _ (14:1) where x0 is given but arbitrary. large control action can easily drive the system out of the region where the given linear model is valid. Our objective is to nd a control function u(t) de ned on t0 T ] which can be a function of the state x(t) such that the state x(t) is driven to a (small) neighborhood of origin at time T .
we shall be concerned exclusively with the L2 performance problem or quadratic performance problem. our problem is as follows: nd a control u(t) de ned on 0 1) such that the state x(t) is driven to the origin at t ! 1 and the following performance index is minimized: min u Z 1 " x(t) 0 # " u(t) Q S S R #" x(t) dt u(t) # (14:2) for some Q = Q . t0 kuk dt T t0 kuk2 dt t2 t0 T ] sup kuk i. and L1 norm.366 Z H2 OPTIMAL CONTROL T Z The constraints on control u may be measured in many di erent ways for example. By the substitution R1=2 u we may as well assume at the start that R = I . which is also positivede nite. without loss of generality. this . we will focus on the in nite time regulator problem. u(t) 2 L2 0 1). and. it is also generally assumed that " Q S S R u # 0: (14:3) Since R is positive de nite.e.. Hence the regulator problem can be posed as an optimal control problem with certain combined performance index on u and x. we can even assume S = 0 by using a prestate feedback u = .S x + v provided some care is exercised however. Similarly. In fact. and R = R > 0. L2 norm. Here we have assumed R > 0 to emphasis that the control energy has to be nite. it has a square root. i. and L1 norm Z T t0 kWu uk dt Z T t0 kWu uk2 dt t2 t0 T ] sup kWu uk for some constant weighting matrix Wu . T ! 1. one might also want to impose some constraints on the transient response x(t) in a similar fashion Z T t0 kWx xk dt Z T t0 kWx xk2 dt t2 t0 T ] sup kWx xk for some weighting matrix Wx .. R1=2 . Moreover. In this chapter. This problem is traditionally called a Linear Quadratic Regulator problem or simply an LQR problem. as given above.. i. S . Moreover.e. L2 norm. weighted L1 norm. So this is the space over which the integral is minimized. In this case. in terms of L1 norm. or more generally. we shall assume t0 = 0.e.
5) has a sensible solution. However. Since the matrix in (14. This will be called a standard LQR problem.e.4) and (14. This problem will be referred to as an Extended LQR problem. This is true in general. and R (or equivalently C1 and D12 ) to ensure that the optimal control law u has this property. then it can be shown that minu2L2 0 1) 01 u2 dt = 2x2 and u(t) = . 14. and R = 1: u2L2 0 1) 0 x = Ax + Bu x(0) = x0 _ min Z 1 u2 dt x = x + u x(0) = x0 : _ It is clear that u = 0 is the optimal solution. we must assume that all unstable states can be \seen" from the performance index. and the proof of this fact is left as an exercise to the reader. This can also be generalized to a more general case where (C1 A) is not necessarily detectable. the LQR problem is posed traditionally as the minimization problem u2L2 0 1) (14:5) without explicitly mentioning the condition that the control should drive the state to the origin. we shall consider the LQR problem as traditionally formulated. if the closedloopRstability is imposed on the above minimization.3) is positive semide nite with R = I .. Hence in order to ensure that the minimization problem in (14. Instead some assumptions are imposed on Q S .2.2x(t) is the optimal 0 control. On the other hand.2 Standard LQR Problem In this section. Standard LQR Problem 367 will not be assumed in the sequel. it can be factored as " Q S = C1 S I D12 u2L2 0 1) # " # h C1 D12 : i And (14.14. Standard LQR Problem .5) is sensible. The problem with this example is that the performance index does not \see" the unstable state x. x(t) = et x0 . i. the system with u = 0 is unstable and x(t) diverges exponentially to in nity.2) can be rewritten as min kC1 x + D12 uk2 : 2 min kC1 x + D12 uk2 2 (14:4) In fact.4) and (14. let us consider a simple example with A = 1 B = 1 Q = 0 S = 0. (C1 A) must be detectable. To see what assumption one needs to make in order to ensure that the minimization problem formulated in (14.
while the detectability of (C1 A) is implied by the detectability of (D? C1 A . B2 D12 C1 ) # " # i (14. Note also that D? drops out when D12 is square.1.2 0 . together with the stabilizability of (A B2 ).B2B2 : . for example. " # " # h i 2 0 1 A= B2 = C1 = 1 0 D12 = 1: 0 . In fact. B2 D12 C1 ) has no unobservable modes on the imaginary axis and is weaker than the popular assumption of detectability of (D? C1 A . B2 D12 C1 ). Finally note that (A4) is equivalent to the condition that (D? C1 A . The assumption (A3) enforces that the unconditional optimization problem will result in a stabilizing control law. which will be shown in Lemma 14.6) z = C1 x + D12 u (14.C1 D12 D12 C1 B2 # A .1 Assumption (A1) is clearly necessary for the existence of a stabilizing control function u. u 2 L2 and z 2 L2 imply x 2 L2 . guarantees by Corollary 13.e.(A .7) and suppose that the system parameter matrices satisfy the following assumptions: (A1) (A B2 ) is stabilizable h i (A2) D12 has full column rank with D12 D? unitary (A3) (C1 A) is detectable " # (A4) A . B2 D12 C1 . 2 Remark 14.A .8) Note also that if D12 C1 = 0.C1 C1 . i.10 that the following Hamiltonian matrix belongs to dom(Ric) and that X = Ric(H ) 0: H = = " " h A 0 B2 . the assumption (A3) together with (A1) guarantees that the input/output stability implies the internal stability.. j!I B2 has full column rank for all !. (A4). The above implication is not true if D12 C1 6= 0. The assumption (A2) is made for simplicity of notation and is actually a restatement that R = D12 D12 = I . .368 H2 OPTIMAL CONTROL Let a dynamical system be described by x = Ax + B2 u x(0) = x0 given but arbitrary _ (14. C1 D12 Find an optimal control law u 2 L2 0 1) such that the performance criterion kz k2 is minimized. It is interesting to point out that (A3) is not needed in the Extended LQR problem. B2 D12 C1 ). then (A4) is implied by the detectability of (C1 A).C1 D? D? C1 .
7). if p < 1. . Lemma 14.1 If u z 2 Lp 0 1) for p 1 and (C1 A) is detectable in the system described by equations (14.2 # has no eigenvalue on the ~ (14:9) (14:10) Note also that the Riccati equation corresponding to (14. Consider applying the control law u = Fx to the system (14.6) and (14.6) and (14. B2 D12 C1 = imaginary axis but is not stable. then x 2 Lp 0 1). Standard LQR Problem Then (C1 A) is detectable and A . B2 D12 C1 ) . " 369 1 0 0 .9) to get CF := C1 + D12 F AF X + XAF + CF CF = 0: (14:11) Thus X is the observability Gramian of (CF AF ).8) is (A . XB2B2 X + C1 D? D?C1 = 0: Now let X be the corresponding stabilizing solution and de ne F := .(B2 X + D12 C1 ): Then A + B2 F is stable. B2 D12 C1 ) X + X (A .2. ) = 0 _ z = CF x: The associated transfer matrix is Gc (s) = AF I CF 0 and " # kGc x0 k2 = x0 Xx0 : 2 The proof of the following theorem requires a preliminary result about internal stability given inputoutput stability. The controlled system is x = AF x x(0) = x0 _ z = CF x or equivalently x = AF x + x0 (t) x(0.14. Furthermore. then x(t) ! 0 as t ! 1. Denote AF := A + B2 F and rearrange equation (14.7).
then from Lemma 14. the system can be written as " x = AF B2 _ z CF D12 # " #" x v # x(0) = x0 : (14:12) Now if v 2 L2 0 1). Fx between v 2 L2 0 1) and those u 2 L2 0 1) that make z 2 L2 0 1) is onetoone and onto.9) and the fact that CF D12 = . So v 2 L2 0 1). Finally.(CF x + D12 v) (CF x + D12 v) + 2x CF D12 v + v v + 2x XB2 v = . u2L2 0 1) min kz k2 = v2L 0 1) kz k2 : min 2 By di erentiating x(t) Xx(t) with respect to t along a solution of the di erential equation (14. Therefore. Now let e = x . Conversely. x(t) ! 0 follows from the fact that if p < 1 then x ! 0.x CF CF x + 2x XB2 v = . x = e + x 2 Lp 0 1). u2L2 0 1) min kz k2 = kGc x0 k2 : Note that the optimal control strategy is constant gain state feedback. Lz: ^ x Then x 2 Lp 0 1) since z and u are in Lp 0 1).1 x 2 L2 0 1). Therefore. Proof. Fx. Let x ^ be the state estimate of x given by _ x = (A + LC1 )^ + (LD12 + B2 )u . if u z 2 L2 0 1). there exists L such that A + LC1 is stable.XB2. With the change of variable v = u . x then ^ ^ e = (A + LC1 )e _ and e 2 Lp 0 1).2 There exists a unique optimal control for the LQR problem. then x z 2 L2 0 1) and x(1) = 0 since AF is stable. Hence u = Fx + v 2 L2 0 1).370 H2 OPTIMAL CONTROL Proof. Moreover. It is easy to see that e(t) ! 0 ^ as t ! 1 for any initial condition e(0).13) . Since (C1 A) is detectable. kz k2 + kvk2 : (14. and this gain is independent of the initial condition x0 . we see that d x Xx = x Xx + x X x _ _ dt = x (AF X + XAF )x + 2x XB2 v = . ^ 2 Theorem 14. Thus the mapping v = u .12) and by using (14. namely u = Fx.
2 " # get Proof.CF : B2 3 7 5 " # Then it is easy to compute 2 D12 U U =6 0 4 B2 .14.X I 0 AF B2 0 I 3 7 5 " 0 3 7 5 =I # U Gc = 6 4 . i.13) from 0 to 1 to get Clearly.11) to get 2 U U =6 0 4 B2 2 .AF . let us rst note the following fact: Lemma 14. An alternative proof can be given in frequency domain. u = Fx. involving change of variables and the completion of the square.X 2 RH? : = 2 B2 0 .AF .CF D12 AF D12 CF B2 I " 2 U Gc = 6 0 4 B2 # .CF CF 0 3 I 7: AF 5 D12 CF 0 Now do the similarity transformation I .AF ..X 0 I 0 on the states of the transfer matrices and use (14.AF . The proof uses standard manipulations of state space realizations. Standard LQR Problem Now integrate (14.e.CF CF .AF B2 0 AF 0 0 .AF . To do that. is a standard technique and variants of it will be used throughout this book. From U we U (s) = .2.3 Let a transfer matrix be de ned as U := AF B2 2 RH1 : CF D12 Then U is inner and U Gc 2 RH? . the unique optimal control is v = 0. 371 kz k2 = x0 Xx0 + kvk2 : 2 2 2 This method of proof.
Hence kz k2 = kGc x0 k2 + kUvk2 : 2 2 2 Since U is inner. we get kz k2 = kGc x0 k2 + kvk2 : 2 2 2 This equation immediately gives the desired conclusion. 2 Remark 14.3 Extended LQR Problem This section considers the extended LQR problem where no detectability assumption is made for (C1 A).3. By Lemma 14.372 H2 OPTIMAL CONTROL 2 An alternative proof of Theorem 14.2 It is clear that the LQR problem considered above is essentially equivaz A I B2 C1 0 D12 I 0 0 w lent to minimizing the 2norm of z with the input w = x0 (t) in the following diagram:  K But this problem is a special H2 norm minimization problem considered in a later section.2 We have in the frequency domain z = Gc x0 + Uv: Let v 2 H2 . Gc x0 and Uv are orthogonal. ~ 14. Extended LQR Problem Let a dynamical system be given by x = Ax + B2 u x(0) = x0 given but arbitrary _ z = C1 x + D12 u with the following assumptions: (A1) (A B2 ) is stabilizable .
.1 F I . F (sI .4. x 2 L2 0 1) and the performance criterion kz k2 is minimized.1B2 . i. The closedloop block diagram is as shown in Figure 14.1 B2 = G12 (s)G12 (s): . u = Fx.e. Guaranteed Stability Margins of LQR (A2) D12 has full column rank with D12 D? unitary " # h i 373 A .sI . Lemma 14. the input/output stability may not necessarily imply the internal stability. A).. x 2 L2 0 1). Moreover. namely u = Fx. i. the internal stability is guaranteed by the way of choosing control law. min kz k = kGc x0 k2 : u2L 0 1) 2 2 Proof.14.4 Guaranteed Stability Margins of LQR Now we will consider the system described by equation (14. Then v = u . (A3) C1 D12 Find an optimal control law u 2 L2 0 1) such that the system is internally stable. min kz k2 = min kz k2 : Using the same technique as in the proof of the standard LQR problem. in this case.5 Let F = . Fx 2 L2 0 1). A).. I .(B2 X + D12C1 ) and de ne G12 = D12 + C1(sI . The proof of this theorem is very similar to the proof of the standard LQR problem except that. u2L2 0 1) v2L2 0 1) z CF D12 v 2 14. Again the mapping v = u . On the other hand. Instead. Then . let v 2 L2 0 1) and consider " # " #" # x = AF B2 _ x x(0) = x0 : Then x z 2 L2 0 1) and x(1) = 0 since AF is stable. The following result is the key to stability margins of an LQR control law. B2 (. i. we have kz k2 = x0 Xx0 + kvk2 : 2 2 And the unique optimal control is v = 0.4 There exists a unique optimal control for the extended LQR problem. 2 Assume the same notation as above. Suppose that u 2 L2 0 1) is such a control law that the system is stable.6) with the LQR control law u = Fx. and we have Theorem 14. Hence u = Fx + v 2 L2 0 1).1. Therefore. j!I B2 has full column rank for all !. A ).e.e.. Fx between v 2 L2 0 1) and those u 2 L2 0 1) that make z 2 L2 0 1) and x 2 L2 0 1) is one to one and onto.
x_ = Ax + B2u x Figure 14. A )X . A). we have B2 (.1 and from the right by (sI .j!I .sI . B2 (.1 C1 C1 (sI . F (j!I .1 B2 = 0: Then the result follows from completing the square and from the fact that D12 D12 = I . A).1 B2 +B2 (.1 B2 .1 F I + F (j!I .1 C1 C1 (sI . A) . A ).1 F F (sI .374 H2 OPTIMAL CONTROL  F u . F (sI .sI .1 F .X (sI . F F + C1 C1 = 0: Now multiply the above equation from the left by B2 (.sI . . I .1 C1 D12 + D12 C1 (sI .1 B2 I I: (14:14) (14:15) and .j!I . . A).B2 (. B2 X (sI . F B2 ). Then .1 B2 .1: LQR closedloop system Proof.1 XB2 .sI . I + B2 (.1 B2 .9) can be written as XA + A X . A). B2 (. A . A ). B2 F ). F F + C1 C1 = 0: Add and subtract sX to the equation to get . A ).sI .A ).1 B2 +B2 (.sI . A ). (.1 B2 +B2 (. A).A). A).sI . A ). A). A ). Note that the Riccati equation (14.1 F + F (sI . A . A).1 C1 C1 (sI . I . A ). A ). A).sI .1 F I . A ). B2 (.1 B2 : In particular.1 B2 = I +B2 (.B2 X = F + D12 C1 in the above equation.sI . 2 Corollary 14. B2 (. A). I .1 F F (sI . .6 Suppose D12C1 = 0. A ).1 B2 = 0: Using .sI .sI .1 B2 to get .
A .14) shows that the openloop Nyquist diagram of the system G(s) in Figure 14.1 is stable even if the openloop system G(s) is perturbed additively by a 2 RH1 as long as k k1 < 1.15) can also be given some robustness interpretation.60o max 60o 2 i.5 Standard H2 Problem The system considered in this section is described by the following standard block diagram: z y G w u  K The realization of the transfer matrix G is taken to be of the form A B1 B2 6 G(s) = 4 C1 0 D12 7 : 5 C2 D21 0 2 3 .14).x_ = Ax + B2u . De ne G(s) = .? e 14.5.14. Hence the system has at least the following stability margins: kmin 1 kmax = 1 min . the system has at least a 6dB (= 20 log 2) gain margin and a 60o phase margin in both directions. z  w  F . Standard H2 Problem 375 Note that the inequality (14.1 B2 . A).1 0) of the complex plane. it is noted that the inequality (14..F (sI .e. Next.1 B2 and assume for the moment that the system is single input. it implies that the closedloop system in Figure 14. In fact. Then the inequality (14. A similar interpretation may be generalized to multiple input systems.15) follows from taking the inverse of inequality (14. B2 F ). This can be seen from the following block diagram and small gain theorem where the transfer matrix from w to z is exactly I + F (j!I .1 never enters the unit disk centered at (.
C1 C1 . 2 .376 H2 OPTIMAL CONTROL Notice the special o diagonal structure of D: D22 is assumed to be zero so that G22 is strictly proper1 also.4 of Chapter 12 there is no loss of generality in making this assumption since the controller for D22 nonzero case can be recovered from the zero case. The following additional assumptions are made for the output feedback H2 problem in this chapter: (i) (A B2 ) is stabilizable and (C2 A) is detectable (ii) D12 has full column rank with D12 D? unitary. D11 is assumed to be zero in order to guarantee that the H2 problem properly posed.B1 B1 .7. We now state the solution of the problem and then take up its derivation in the next several sections. 2 Recall that a rational proper stable transfer function is an RH function i it is strictly proper.2 The case for D11 6= 0 will be discussed in Section 14. H2 Problem The H2 control problem is to nd a proper. .A " " # B2 .g. The rank assumptions (ii) are necessary to guarantee that the H2 optimal controller is a nite dimensional linear time invariant one.B1D21 # # h i 1 As we have discussed in Section 12. j!I B2 C1 D12 A .10 the two Hamiltonian matrices A 0 H2 := .3. realrational controller K which stabilizes G internally and minimizes the H2 norm of the transfer matrix Tzw from w to z . j!I B1 C2 D21 # has full column rank for all ! has full row rank for all !. By Corollary 13. Recall that a controller is said to be admissible if it is internally stabilizing and proper..C1 D12 " " # h D12 C1 B2 D21 B1 C2 i A 0 C2 J2 := . and D21 has full row rank h i D with ~21 unitary D? " " # (iii) " A . while the unitary assumptions are made for the simplicity of the nal solution they are not restrictions (see e.A . Chapter 17). . In the following discussions we shall assume that we have state models of G and K . and the third and the fourth assumptions together with the rst guarantee that the two Hamiltonian matrices associated with the H2 problem below belong to dom(Ric). # (iv) The rst assumption is for the stabilizability of G by output feedback.
9.14. Theorem 14. moreover. min kTzw k2 = kGc B1 k2 + kF2 Gf k2 = kGc L2k2 + kC1 Gf k2 .L2 B2 6 M2 (s) = 4 F2 0 I .(B2 X2 + D12 C1 ) L2 := . the suboptimal controllers are parameterized by a xed (independent of ) linearfractional transformation with a free parameter Q. De ne AF2 := A + B2 F2 C1F2 := C1 + D12 F2 AL2 := A + L2C2 B1L2 := B1 + L2 D21 ^ A2 := A + B2 F2 + L2C2 # " # " AL2 B1L2 : AF2 I Gf (s) := Gc (s) := C1F2 0 I 0 Theorem 14.5. kQk2 < 2 . which will be discussed in more detail in Section 14. X2 := Ric(H2 ) 0 and Y2 := Ric(J2 ) 0. . Standard H2 Problem F2 := .C2 I 0 2 3 7 5 where Q 2 RH2 .(Y2 C2 + B1 D21 ) and 377 belong to dom(Ric). (kGc B1 k2 + kF2 Gf k2 ). and.L2 : K (s) := opt F2 0 Moreover.8 makes Tzw a ne in Q and yields the Youla parameterization of all stabilizing controllers when the conditions on Q are replaced by Q 2 RH1 .7 There exists a unique optimal controller # " ^ A2 .8 The family of all admissible controllers such that kTzw k2 < equals the set of all transfer matrices from y to u in u M2 y  Q ^ A2 . For comparison with the H1 results. we recover Kopt . It is worth noting that the parameterization in Theorem 14. With Q = 0. 2 2 2 Thus. 2 2 2 2 2 The controller Kopt has the wellknown separation structure. it is useful to describe all suboptimal controllers.
D12 F2 D21 0 0 C2 De ne We have AF2 . The reason is that this approach does not generalize nicely to other control problems and is often very involved.8. In particular.6 Optimal Controlled System In this section. we look at the controller.3 that Gc B1 and U are orthogonal. We shall now analyze the system structure under the control of a such controller. We will give a brief analysis of the closedloop system.8 and the separation argument. QV k2 : 2 2 2 2 2 . The proof given here is not our emphasis.B2 F2 3 U = AF2 B2 C1F2 D12 " # V = AL2 B1L2 : C2 D21 " # Tzw = Gc B1 . we will compute explicitly kTzw k2 . (Keep in mind that all admissible controllers are parameterized by the above formula). UF2 Gf + UQV: It follows from Lemma 14. An alternative proof will be given in the later part of this chapter by using the FI and OE results discussed in section 14. UQV k2 = kGc B1 k2 + kF2 Gf .7 and 14. It will be seen that a direct consequence from this analysis is the results of Theorem 14. The idea of separation is the main theme for synthesis.378 H2 OPTIMAL CONTROL 14. K (s) = F`(M2 Q) Q 2 RH1 connected to G. Thus kTzw k2 = kGc B1 k2 + kUF2 Gf . 2 Consider the following system diagram with controller K (s) = F`(M2 Q): z y XXXX Q G y1 M2 XXX w u u1 Then Tzw = F` (N Q) with 2 B1 B2 7 6 6 0 AL2 B1L2 0 7 : 7 N =6 6 0 D12 7 5 4 C1F2 .
. so we have kTzw k2 = kGc B1 k2 + kF2 Gf . Gf V 2 RH? . Gf can be represented as A B1 I I 0 0 C2 D21 0  L2 and L2 minimizes the H2 norm of Gf and solves a special ltering problem.e.7 H2 Control with Direct Disturbance Feedforward* A B1 B2 6 G(s) = 4 C1 D11 D12 C2 D21 0 2 3 7 5 . Let us consider the generalized system structure again with D11 not necessarily zero: 14. H2 Control with Direct Disturbance Feedforward* 379 It can also be shown easily by duality that Gf and V are orthogonal. in fact. Note also that kTzw k2 is nite if and only if Q 2 RH2 . QV k2 = kGc B1 k2 + kF2 Gf k2 + kQk2 : 2 2 2 2 2 2 This shows clearly that Q = 0 gives the unique optimal control. so K = F`(M2 0) is the unique optimal controller. Similarly.8 follow easily.7. 2 and V is a coinner. Hence Theorem 14.14. First of all the transfer matrix Gc can be represented as a xed system with the feedback matrix F2 wrapped around it: A I B2 C1 0 D12 I 0 0  F2 F2 is. It is interesting to examine the structure of Gc and Gf . an optimal LQR controller and minimizes the H2 norm of Gc .7 and 14. i.
380 H2 OPTIMAL CONTROL We shall consider the following question: what will happen and under what condition will the H2 optimal control problem make sense if D11 6= 0? Recall that F`(M2 Q) with Q 2 RH1 parameterizes all stabilizing controllers for G regardless of D11 = 0 or not. suppose I 0 # D21 = 0 I # h i and D11 is partitioned accordingly D11 = D1111 D1112 : D1121 D1122 Then equation (14. i.16) holds and denotes DK := . To nd the optimal control law for the system G with D11 6= 0. For example..e. having nite H2 norm.16) is a very restrictive condition. So only D1122 can be nonzero for a sensible H2 problem.D12 D11 D21 . let us consider the following system con guration: .16) implies that " " D1111 D1112 = 0 0 D1121 0 0 0 # " # and that Q(1) = .D1122 . UF2 Gf + UQV + D11 and Tzw (1) = D12 Q(1)D21 + D11 : Hence the H2 optimal control problem will make sense. if and only if there is a constant Q(1) such that D12 Q(1)D21 + D11 = 0: This requires that and that Q(1) = . Hence from now on in this section we shall assume that (14. D12 D12 D11 D21 D21 + D11 = 0: D12 = " (14:16) Note that the equation (14.D12 D11 D21 . Now again consider the closed loop transfer matrix with the controller K = F`(M2 Q) then Tzw = Gc B1 .
but with di erent structures for G. B2 DK ) ^ K= F2 .8.8 Special Problems . Hence the controller K for the original system G will be given by " # ^ A2 . Special Problems z G w 381 .DK y f 6 f 6 ^ G u1 . A little bit of algebra will show that # " ^ A2 .5 can be used. As 14.DK Then 2 K ^ K and A + B2 DK C2 B1 + B2 DK D21 B2 ^ = 6 C1 + D12 DK C2 G 4 0 D12 C2 D21 0 3 7 5 ^ K = DK + K: ^ It is easy to check that the system G satis es all assumptions in Section 14.14. DK C2 0 ^ is the H2 optimal controller for G. B2 DK ) = F (M D ): K= ` 2 K F2 .(L2 . B2 DK C2 .. DK C2 DK In this section we look at various H2 optimization problems from which the output feedback solutions of the previous sections will be constructed via a separation argument. B2 DK C2 . All the special problems in this section are to nd K stabilizing G and minimizing the H2 norm from w to z in the standard setup.(L2 .5 hence the controller formula in Section 14.
. respectively. The special problems SF. The same comments apply to the \all" situation. OI. i. special cases of the output feedback problem since they do not satisfy all of the assumptions for output feedback (while DF and OE do). Each special problem inherits some of the assumptions (i)(iv) from the output feedback as appropriate. it will be seen that this equivalence class problem will not occur in the general output feedback case including DF and OE problems. In other words. The assumptions will be discussed in the subsections for each problem. as an equivalence class.382 H2 OPTIMAL CONTROL in Chapter 12. the minimum of kTzw k2 2. as can be seen in section 14. Hence the \unique controller" here means one of the controllers from the unique equivalence class. state feedback (SF).8. and OE are natural duals of SF. OI. we shall call these special problems.1 State Feedback Consider an openloop system transfer matrix A B1 B2 6 GSF (s) = 4 C1 0 D12 I 0 0 with the following assumptions: 2 3 7 5 . respectively. OI. and not very useful.e. 14. In particular. full information (FI). disturbance feedforward (DF). FC. FI.1 when we consider the state feedback problem. Thus the unique controller is really not unique. Once again we regard all controllers that give the same control signal u. In each case. and FC are not. and output estimation (OE). denoted as K1 = K2. the unique controller minimizing kTzw k2 3. We chose not to give the parameterization of all the elements in an equivalence class in this book since it is very messy. we will regard them as the same. where is greater than the minimum norm. FI and FC problems are neither unique nor allinclusive.8. if K1 and K2 generate the same control signal u. and DF. In that case we actually give a parameterization of all the elements in the equivalence class of the \unique" optimal controllers. Warning: we will be more speci c below about what we mean about the uniqueness and all controllers in the second and third item. full control (FC). having the same transfer function from w to u. the family of all controllers such that kTzw k2 < . The output feedback solutions will be constructed out of the FI and OE results. strictly speaking. FI.8.1. This will be much clearer in section 14. output injection (OI). the controllers characterized here for SF. the results are summarized as a list of three items (in all cases. However. K must be admissible) 1.
3 The class of all suboptimal controllers for state feedback are messy and are not very useful in this book. but it turns out that dynamics are not necessary. as are the OI problems. Let K be a stabilizing controller. ~ Proof. Notice that Tvw 2 RH2 because K stabilizes G.14. F2 ) 0 0 z AF2 B1 B2 C1F2 0 D12 I 0 0 x w 7: 5 v w The block diagram is x v . u = K (s)x.8. Special Problems (i) (A B2 ) is stabilizable " 383 h i (ii) D12 has full column rank with D12 D? unitary (iii) A . Change control variables by de ning v := u . Then Tzw = Gc B1 + UTvw . F2 Let Tvw denote the transfer matrix from w to v. State Feedback: 1. This is very much like the LQR problem except that we require from the start that u be generated by state feedback and that the detectability of (C1 A) is not imposed since the controllers are restricted to providing internal stability. so they are omitted. j!I B2 C1 D12 # has full column rank for all !. F2 x and then write the system equations as 2 6 4 3 2 32 76 54 3 x _ AF2 B1 B2 7 6 z 5 = 4 C1F2 0 D12 v (K . K (s) = F2 Remark 14. The controller is allowed to be dynamic.K . min kTzw k2 = kGc B1 k2 = (trace(B1 X2 B1 ))1=2 2.
and any other controllers achieving minimum are in the equivalence class of F2 . We where U = c 2 C1F2 D12 get Thus min kTzw k2 = kGc B1 k2 and the minimum is achieved i Tvw = 0.384 " # H2 OPTIMAL CONTROL AF2 B2 . let " # Pc (s) := AF2 B2 : I 0 . AF2 B2 K . and by Lemma 14. this is the only solution if B1 is square and nonsingular or if K is restricted to be constant and (AF2 B1 ) is controllable. We will now try to nd all of those controllers that stabilize the system and give Tvw = 0. We now give some further explanation about the uniqueness of the optimal controller that we commented on before. Then all H2 optimal state feedback controllers can be parameterized as Kopt = F` (Msf ) with 2 RH2 and F2 I Msf = : Vc (sI . Clearly. 2 2 K = F2 is a controller achieving this minimum. 2 Note that the above proof actually yields a much stronger result than what is needed. F2 )(sI . is only one of them. kTzw k2 = kGc B1 k2 + kTvw k2 : 2 2 2 Proposition 14.Vc B2 " # Proof.e. AF2 ). This means that the optimal state feedback is also optimal for the full information problem as well.1 " AF2 B1 = 0 K .. AF2 ) . The proof that the optimal Tvw is Tvw = 0 does not depend on the restriction that the controller measures just the state.1 B1 = 0 (14:17) which is achieved if K = F2 . The important observation for this issue is that the controllers making Tvw = 0 are not unique.9 Let Vc be a matrix whose columns form a basis for KerB1 (Vc B1 = 0). We only require that the controller produce v as a causal stable function Tvw of w. i.3 U is inner and U G is in RH? . To parameterize all elements in the equivalence class (Tvw = 0) to which K = F2 belongs. F2 0 # (K . Since Tvw = I . The controller given above. F2 0 we get " #!. Furthermore. all K (s) = F2 . F2 .
AF2 ) (s) 2 RH2 : (14:19) Therefore. this extra information is not used by the optimal controller. as was pointed out in the discussion of the state feedback problem. AF2 ) = F2 + (I + (s)Vc B2 ). AF2 )Pc ).1 B1 = 0: Hence we have (s) 2 RH2 : Thus all Q(s) 2 RH1 satisfying (14. AF2 ). 2 14. Special Problems Then all state feedback controllers stabilizing G can be parameterized as 385 K (s) = F2 + (I + QPc ). K (s) = F2 Q(s) . K (s) = F2 0 h h i 3. Full Information: 1. AF2 ) (s) 2 RH2 parameterizes the equivalence class of K = F2 . 2 A 6 6 C1 GFI (s) = 6 " # 6 I 4 0 B1 " 0 0 # B2 D12# " 0 0 3 7 7 7 7 5 I The assumptions relevant to the FI problem are the same as the state feedback problem.8. where Q 2 RH2 . and we get Q(s)(sI . min kTzw k2 = kGc B1 k2 = (trace(B1 X2 B1 ))1=2 2.17).18) can be written as (14:18) Q(s)(sI .14.8.1 = (s)Vc Q(s) = (s)Vc (sI . kQk2 < 2 . AF2 ).1 (s)Vc (sI .1 (s)Vc (sI . kGcB1 k2 2 2 i .2 Full Information and Other Special Problems We shall consider the FI problem rst. This is similar to the state feedback problem except that the controller now has more information (w). However.1 Q Q(s) 2 RH1 where Q is free. Substitute K in equation (14. Kopt(s) = F2 + (I + (s)Vc (sI .
one can show that every controller of the form given in item no. FC. OI. F2 x then the closedloop system is as shown below Proof. B1 C2 B1 2.386 H2 OPTIMAL CONTROL because the argument that Tvw = 0 did not depend on the restriction to state feedback only. de ne a new control variable v = u . Let K be an admissible controller such that kTzw k2 < . F2 0]: kQk2 = kTzw k2 . Then u = F2 x + v = F2 x + Qw = F2 Q y so K = F2 Q . K (s) = F2 0 " # . kGcB1 k2 < 2 . Disturbance Feedforward: A B1 B2 6 GDF (s) = 4 C1 0 D12 C2 I 0 2 3 7 5 This problem inherits the same assumptions (i)(iii) as in the state feedback problem. 2 The results for DF. and OE follow from the parallel development of Chapter 12. min kTzw k2 = kGc B1 k2 A + B2 F2 . B1 C2 . 2 2 2 2 AF2 6 6 C1F2 ~ GFI = 6 " # 6 I 4 2 B1 " 0 0 B2 D12# " 0 0 ~ K = K . Thus we only need to prove item 3. in addition to the stability condition of A .3 is admissible and suboptimal. and kTzw k2 = kGc B1 + UQk2 = kGc B1 k2 + kQk2 hence. Items 1 and 2 follow immediately from the proof of the state feedback results z ~ GFI w y with # ~ K v 3 7 7 7 7 5 I 0 Denote by Q the transfer matrix from w to v it belongs to RH2 by internal stability and the fact that D12 has full column rank and Tzw with zi= C1F2 x + D12 Qw hasi nite h h H2 norm. kGc B1 k2 : 2 2 2 2 Likewise. 1. As in the SF proof.
kGcB1 k2 2 2 Output Injection: A B1 I 6 GOI (s) = 4 C1 0 0 C2 D21 0 " # 2 3 7 5 with the following assumptions: (i) (C2 A) is detectable D (ii) D21 has full row rank with ~21 D? " unitary (iii) A . kQk2 < 2 . j!I B1 C2 D21 " # has full row rank for all !. K (s) = L2 0 # Full Control: 0 0 with the same assumptions as an output injection problem. B1 C2 B1 B2 6 M2D (s) = 4 F2 0 I . min kTzw k2 = kC1 Gf k2 = (trace(C1 Y2 C1 ))1=2 2.14. 1. min kTzw k2 = kC1 Gf k2 = (trace(C1 Y2 C1 ))1=2 2. the set of all transfer matrices from y to u in 387 u M2D y  Q A + B2 F2 . Special Problems 3.C2 I 0 2 3 7 5 where Q 2 RH2 .8. K (s) = " A B1 6 6 GFC (s) = 6 C1 0 4 C2 D21 2 h h h I 0 0 I i 3 i 7 7 7 i 5 L2 0 # . 1.
9. B2 C1 is stable. kC1 Gf k2 2 Output Estimation: The assumptions are taken to be those in the output injection problem plus an additional assumption that A . Note that F2 is the optimal state feedback in the Full Information problem and L2 is the . kC1 Gf k2 2 2 14.388 H2 OPTIMAL CONTROL " # L2 . 1. kQk2 < 3. K (s) = 2 2 Q(s) 2 . min kTzw k2 = kC1 Gf k2 A B1 B2 GOE (s) = 6 C1 0 I 4 C2 D21 0 2 3 7 5 A + L2 C2 . kQk2 < 2 .9 Separation Theory Given the results for the special problems. This essentially involves reducing the output feedback problem to a combination of the Full Information and the Output Estimation problems. where Q 2 RH . B2 C1 L2 2.L2 = A + B2 F2 + L2C2 Y2 C2 K2 (s) := F2 0 .7 using separation arguments. 14. the set of all transfer matrices from y to u in M2O  Q A + L2 C2 . we can now prove Theorem 14.B2 M2O (s) = 6 C1 0 I 4 C2 I 0 2 3 7 5 where Q 2 RH2 . B2 C1 L2 . K (s) = C1 0 u y " # 3.1 H2 Controller Structure Recall that the unique H2 optimal controller is " # " ^ A2 .B2 X2 0 2 = kGc B1 k2 + kF2 Gf k2 and min kTzw k2 2 2 # where X2 := Ric(H2) and Y2 := Ric(J2 ) and the min is over all stabilizing controllers.
Furthermore. U is inner where Gc (s) := C1F2 0 C1F2 D12 (i. and the transfer function to z becomes #" # " " AF2 B1 B2 C1F2 0 D12 # w = G B w + Uv c 1 v # (14:20) AF2 I and U (s) := AF2 B2 .3.e. ^ 14.1) and the OE cost for estimating F2 x (OE.14.2 Proof of Theorem 14. Also.2.2. y) ^ ^ ^ u = F2 x ^ where x is the optimal estimate of x. kTvw k2 is minimized by the controller " " (14:21) A + B2 F2 + L2C2 . From (14.7 z= As before we de ne a new control variable.20) and the properties of U we have that min kTzw k2 = kGcB1 k2 + min kTvw k2 : 2 2 2 But from item OE.. . Separation Theory 389 optimal output injection in the Full Control case. F2 x. U U = I ) and U Gc belongs to RH? from Lemma 14. The controller equations can be written in standard observer form as _ x = Ax + B2 u + L2(C2 x .9.1). 2 Let K be any admissible controller and notice how v is generated: v w 2 3 A B1 B2 Gv Gv = 6 . the minimum cost is the sum of the FI cost (FI.1 min kTvw k2 = kF2 Gf k2 .L2 F2 0 # and then from OE. v := u .F2 0 I 7 4 5 y u C2 D21 0 K Note that K stabilizes G i K stabilizes Gv (the two closedloop systems have identical Amatrices) and that Gv has the form of the Output Estimation problem. The wellknown separation property of the H2 solution is re ected in the fact that K2 is exactly the optimal output estimate of F2 x and can be obtained by setting C1 = F2 in OE.9.
L2 B2 6 M2 (s) = 4 F2 0 I . Consider a single input and single output two state generalized dynamical system: 2 " 6 6 6 6 " 6 6 6 6 0 6 4 h G(s) = pq pq 0 1 0 i 1 1 0 1 # " p p 0 0 0 # " 0 1 0 1 0 # # 3 7 7 7 # 7 7 7 7 7 7 5 # " : h 0 1 " i It can be shown analytically that X2 = and " 2 h # Y2 = i 2 " F2 = . kGc B1 k2 .8 u y Continuing with the development in the previous proof. it is not clear whether these stability margins will be preserved if the states are not available and the output feedback H2 (or LQG) controller has to be used.9.C2 I 0 2 3 7 5 2 . However. we see that the set of all suboptimal controllers equals the set of all K 's such that kTvw k2 < 2 . where 1 1 L2 = . Apply 2 2 item OE. 1 1 # =2+ 4+q p =2+ 4+ : p . The answer is provided here through a counterexample: there are no guaranteed stability margins for a H2 controller. kQk2 < 2 Q ^ A2 .10 Stability Margins of H2 Controllers We have shown that the system with LQR controller has at least 60o phase margin and 6dB gain margin. kF2 Gf k2 .390 H2 OPTIMAL CONTROL 14.3 Proof of Theorem 14.3 to get that such K 's are parameterized by M2 with Q 2 RH2 . 2 2 14. kGc B1 k2 .
10. Thus. k) : Note that for closedloop stability it is necessary to have a0 > 0 and a1 > 0. This will not be explored further here interested reader may consult related references. : . It is interesting to note that the margins deteriorate as control weight (1=q) gets small (large q) and/or system deriving noise gets large (large ). . modern LQG designers are obliged to test their margins for each speci c design. It can be shown that the characteristic polynomial has the form ~ det(sI . 1 7 4 5 0 . . 1. openloop stable counterexamples for which the margins are arbitrarily small.14.k 6 0 1. 1) for su ciently large and if k 6= 1. the system is unstable for arbitrarily small perturbations in k in either direction. by choice of q and . . It is also important to recognize that vanishing margins are not only associated with openloop unstable systems.( + ) 1 . however. the gain margins may be made arbitrarily small. A successful approach in this direction is the so called LQG loop transfer recovery (LQG/LTR) design technique. 4 + 2(k .k 7 : ~ 6 7 A = 6 0 1 . provide no global systemindependent guaranteed robustness properties. It is easy to see that for su ciently large and (or q and ). It may. unlike LQR solutions. In modern control folklore. 1) a0 = 1 + (1 . Note also that a0 (1 . It is easy to construct minimum phase. The point of these examples is that H2 (LQG) solutions. Then the controller implemented in the system is actually K = kKopt and the closedloop system Amatrix becomes 2 3 1 1 0 0 6 7 . The idea is to design a ltering gain (or FC control law) in such way so that the LQG (or H2 ) control law will approximate the loop properties of the regular LQR control. these have often been considered ad hoc means of improving sensitivity. 1 Kopt = 6 . Stability Margins of H2 Controllers Then the optimal output H2 controller is given by 2 1. k) and a1 2(k . A) = a4 s4 + a3 s3 + a2 s2 + a1 s + a0 with a1 = + . 4 3 7 5 391 0 Suppose that the resulting closedloop controller (or plant G22 ) has a scalar gain k with a nominal value k = 1. Like their more classical colleagues. be possible to improve the robustness of a given design by relaxing the optimality of the lter (or FC controller) with respect to error properties.
etc. Kalman ltering.. LQ optimal control.392 H2 OPTIMAL CONTROL 14. can be found in Anderson and Moore 1990] or Kwakernaak and Sivan 1972].11 Notes and References The detailed treatment of H2 related theory. .
Linear Quadratic Optimization This chapter considers time domain characterizations of Hankel operators and Toeplitz operators by means of some related quadratic optimizations..e.1) Consider rst the problem of using an input w 2 L2. i. This is 2 exactly the standard problem of computing the Hankel norm of G. the induced norm of the Hankel operator ? P+ MG : H2 ! H2 and the norm can be expressed in terms of the controllability Gramian Lc and observability Gramian Lo: ALc + LcA + BB = 0 A Lo + LoA + C C = 0: 393 .1 Hankel Operators Let G(s) be a stable real rational transfer matrix with a state space realization (15. These characterizations will be used to prove a maxmin problem which is the key to the H1 theory considered in the next chapter. to maximize kP+ z k2. 15 x = Ax + Bw _ z = Cx + Dw: 15.
kP+z k2 = supw2BH? kP+ MG wk2 = (LoLc): 2 2 2 Z Proof.1 A and completion of the squares gives c c d . we can di erentiate x(t) L.1 x) = x L.1x0 = B L.1 + L.1) = 0.1 x0 : (15:2) 2 c w2L2.A t L. kw . B Lc xk : Integration from t = .1x0 = kwk2 . kwk2 x(0) = x0 = x0 y0 where y0 solves Lcy0 = x0.2 from supw2BL2.1 2 dt (x Lc x) = kwk .1 2 . Then Lc is invertible and y0 = L. .1 xi: _ c c c _ c c c dt Using Lc equation to substitute for A L.1e(A+BB L. k 2 c 0 x0 eA t C CeAt x0 dt = x0 Lo x0 : 2 . inf 2 Proof.1 x = x (A L. for t 0 the norm of z(t) = CeAtx0 can be found kP+z k2 = 2 1 Combine this result with Lemma 15. kw .1) for any given c input w as follows: d (x L. Assume (A B) is controllable otherwise.1 A)x + 2hw B L.1)t x0 on (.1 x and equality is achieved.1 to give + z k2 sup kP+ z k2 = sup kPwk2 2 = max x0 Lo1x0 = (LoLc ): 2 x0 6=0 x0 L.394 LINEAR QUADRATIC OPTIMIZATION Although this result is wellknown. cthus proving (15.1 0]. x0 w2BL2.1 x(t) along the solution of (15.1 + L. B Lc 2 Lemma 15. then w 2 L2. Given x(0) = x0 and w = 0. We need to show inf kwk2 x(0) = x0 = x0 L. factor out the uncontrollable subspace. w 2 L2.1 xk2 kwk2 : c 2 c 2 2 c If w(t) = B e.1 to t = 0 with x(.1) = 0 and x(0) = x0 gives x0 L. Lemma 15. B L. Moreover. w = c .1x0 . can be used to c produce any x(0) = x0 given x(. 06=w2L2.1 w2L2.2). To show this. we will include a timedomain proof similar in technique to the proofs of the optimal H2 and H1 control results.1 x + x L.
3 Suppose kGk1 < 1 and x(0) = x0 .1 (B X + D C ) is stable and A X + XA + (XB + C D)R.1) = 0: sup kP+ z k2 . 2 kwk2 2 2 = = sup (x0 Lo x0 . 2 kwk2 : 2 2 w2L2. It is easy to see from the de nition of the Hankel norm that +z sup kPwkk2 k 2 06=w2L2.24. . A + BR.(A + BR. then by Corollary 13.1 Another useful way to characterize the Hankel norm is to examine the following quadratic optimization with initial condition x(. Toeplitz Operators 395 Remark 15.1 (B X + D C ) + C C = 0: The following lemma o ers yet another consequence of kGk1 < 1.1 D C BR. kwk2 ) = x0 Xx0 2 2 w2L2+ and the sup is achieved.1 D C ) . i So the Hankel norm is equal to the smallest Now w2L2. X = Ric(H ) 0. ~ 15.2 Toeplitz Operators " If a transfer matrix G 2 RH1 and kGk1 < 1.) Lemma 15. the Hamiltonian matrix A + BR.C (I + DR # R = 2I . 2 x0 L. Then sup (kz k2 . sup kP+ z k2 . 06=w2L2.1 B H= .1 x0 ) c 0 +1 (Lo Lc) 2 (Lo Lc) > 2 : Hence the Hankel norm is equal to the square root of (Lo Lc).15. (Recall that the H1 norm of a stable matrix is the Toeplitz operator norm. x0 2Rn ( sup kP+ z k2 .1D )C . 2 kwk2 2 2 0: such that the above inequality holds. D D (15:3) with = 1 is in dom(Ric).2.
(B X + D C )x]k2 x0 Xx0 : (15:4) 2 2 2 . On the other hand. To do that. ~ A dual version of Lemma 15. kR. 2 X0 ) < 0. 2 X0 ) = 0 max (X . then x 2 L2+ . So if (X0 ) is small. . use the Riccati equation (15. Remark 15. sup 2 n L2 0 1) kw k2 + x0 X0 x0 06=(x0 w)2R and max (X . H 2 dom(Ric). Thus the inequality in (15. we have the following Corollary 15.1 (B X + D C ) is stable. and complete the squares to get d (x Xx) = .1=2 Rw . AT = A B T = B .4) can be made an equality and the proof is complete.1 (B X + D C )e A+BR x0 .1=2 Rw . we rst note that1 (ii) GT (s) = " A B C D # 1 Note that since the system matrices are real. (B X + D C )x]k2 : dt If w 2 L2+ . a large (X0 ) implies that the initial condition is known very well and that min is determined essentially by the condition H 2 dom(Ric). kwk2 = x0 Xx0 . then the initial condition is probably not known very well. (i) if kGk1 < and X = Ric(H ).kz k2 + kwk2 . 2 X0 ) 0. etc. 2 As a direct consequence of this lemma. We can di erentiate x(t) Xx(t) as in the last section. Then sup kz k2 .2 The matrix X0 has the interpretation of the con dence on the initial condition x0 .1 (B X +D C )]t If we let w = R.396 LINEAR QUADRATIC OPTIMIZATION Proof.3) to substitute for A X + XA.4 Let x(0) = x0 and X0 = X0 > 0. kR. Note that the sup is achieved for a w which is a linear function of the state. then w 2 L2+ because A + BR. In that case min will be large where min denotes the smallest such that max (X .3 can also be obtained and is useful in characterizing the socalled 2 2 block mixed HankelToeplitz operator. 2 X0 ) > 0: =0 = +1 kP+z k2 < 2 if and only if (D) < . 2 x0 X0 x0 max (X . 2 (kwk2 + x0 X0 x0 ) 2 2 n 06=(x0 w)2R L2 0 1) 8 > < > : = sup <0 2 06=x0 2Rn x0 Xx0 .1 (B X + D C )x = R. so integrating from t = 0 to t = 1 gives kz k2 . The conjugate transpose is used here for the transpose for the sake of consistency in notation. 2 X0 ) < 0 max (X .
1 (B Y .1) = 0 so integrating from t = . The case in which Y is singular can be obtained by restricting x0 2 Im(Y ) and replacing Y . 2 15. ~ then w 2 L2. and complete the squares to get d .3.1 DB h C i ~ ~ where R := I . we shall assume that (A B ) is controllable hence.z k2 . . Analogous to the proof of Lemma 15. D D > 0.1(B Y .1 C is stable and ~ AY + Y A + (Y C + BD )R. where R = I .x0 Y .1 +D C )]tx0 . (B Y + D C )x]k Proof.1=2 Rw . Then sup kP. and the sup is achieved. For simplicity.BD 0 # " ~ R. If w 2 L2.1 2 2 . use kz k2 . DD .A .1 x0 : (15:6) 2 2 2 If we let w = R. kwk2 = . kwk2 x(0) = x0 = .3.1 C g Y .15. i h i G(s) = G1 (s) G2 (s) 2 RH1 h i .5 Suppose kGk1 < 1 and (A B) is controllable. and x(. Then J 2 dom(Ric).1 + D C )x = R.1 2 dt (x Y x) = . Let J denote the following Hamiltonian matrix 397 # J= A .1 + D C ) = .1 + D C )e A+BR. kR Rw . Y > 0.1 (CY + DB ) + BB = 0 (15:5) if kGk1 < 1.1x(t). D = D1 D2 .3 Mixed HankelToeplitz Operators h Now suppose that the input is partitioned so that B = B1 B2 . then x 2 L2.kz k + kwk .1 C is stable.x0 Y .1 x0 2 2 w2L2.5) to substitute for AY + Y A . (B Y . because A + BR. we can di erentiate x(t) Y . kR.1 + D C )x]k2 .x0 Y .Y fA +(Y C + BD )R.1 (B Y .BB " C + . Mixed HankelToeplitz Operators and kGk1 < 1 i kGT k1 < 1.1 (B Y . Lemma 15. Thus the inequality can be made an equality and A + (Y C + BD )R and the proof is complete.1 x0 .1=2 .1 to t = 0 gives the Riccati equation (15. A + (Y C + BD )R.1 ~ .1 by Y + . Y = Ric(J ) 0.
Since (WLc ) 1 i 9 x0 6= 0 such that the righthand side of .3 yield w2W sup kP+ z k2 . Note that the last term only contributes to kP+ z k2 then Lemma 15. R2 1 D2 C B2 . equivalently. is the sum of the Hankel operator P+ MG : H2 H2 ! H2 and the Toeplitz operator P+ MG2 : H2 ! H2 . wk2 : 2 2 2 2 2 2 through x(0). = P+ MG1 MG2 ] : W ! H2 is a mixed HankelToeplitz operator: w .2 (B2 = 0 D2 = 0). so we will prove that given condition (i). (15. equivalently. w + P+ G2 P+ w2 : ? ? Thus . Let (" # ) w1 w 2 H ? w 2 L : ? L2 = w 2 W := H2 (15:7) 1 2 2 2 w2 We are interested in a test for supw2BW kP+ z k2 < 1. Thus if Lc is invertible.1 x0 2 2 c (15:9) and the supremum is achieved for some w 2 W that can be constructed from the previous lemmas. Then kG2 k1 < 1 i (D2 ) < 1.wk2 < 1 i " # (15:8) where . and i HW := + is in dom(Ric) where R2 := I .C D2 # and w is partitioned conformably. By de nition of W . kP.24 (B1 = 0 D1 = 0) and Lemma 15. or. D2 D2 > 0. kP+ w2 k2 . de ne W = Ric(HW ). 1 w2 " # = P+ = P+ h h w1 ? w1 2 H2 w2 2 L2 G1 G2 w2 i G1 G2 P. kwk2 x(0) = x0 = x0 Wx0 .wk2 1. if w 2 W then kP+ z k2 . kwk2 = kP+ z k2 . condition (i) is necessary for (15. Proof. For HW 2 dom(Ric). Lemma 15. The following lemma generalizes Corollary 13.6 supw2BW k.1 and 15. x0 L.24. w2BW sup k.A # " h B2 .8).wk2 < 1 i the following two conditions hold: (i) (D2 ) < 1 and HW 2 dom(Ric) (ii) (WLc) < 1. We will do this by showing.398 " LINEAR QUADRATIC OPTIMIZATION A 0 . By Corollary 13. that (WLc) 1 i supw2BW k.C C .8) holds i condition (ii) holds.
4. G1 .7 below is completely straightforward and fairly short. That the expression in (15. In fact. Gz := P. (G1 z ) w1 > + < G2 z w2 > = <. given the other results in the previous sections. Suppose that A B1 B2 G(s) = G11 (s) G12 (s) = 6 C1 D11 D12 4 G21 (s) G22 (s) C2 D21 D22 " # 2 3 7 5 A B 2 RH : =: 1 C D i " # Denote h D 2 := D12 D2 := D21 D22 D22 Rx := I . is easily veri ed from the de nition of the inner product on vectorvalued L2 space. z w>: The mixed HankelToeplitz operator just studied is the socalled 2 1block mixed HankelToeplitz operator. By de nition.6. See Section 17. : W ! H2 is the operator .4 Mixed HankelToeplitz Operators: The General Case* Historically.7 in Chapter 17. : H2 ! W such that < z . the whole H1 control theory can be developed using these tools.9) is 0. Mixed HankelToeplitz Operators: The General Case* kP+ z k2 kwk2 . There is a 2 2block generalization.9) to Im(Lc ). D2 D2 " # . 1 G2 z G2 " # " # z (15:10) where P.w > = < z P+ (G1 w1 + G2 w2 ) >=< z G1 w1 > + < z G2w2 > = < P. and then the above argument generalizes in a straightforward way. so it is interesting to consider the 2 2block generalization of Lemma 15. w 6= 0 such that If Lc is not invertible. (Gz ) = (P. 2 2 (15. that (WLc) 399 1 i 9 w 2 W . But this is true i supw2BW k. The adjoint of . we need only restrict x0 in (15. : H2 ! W .9). the mixed HankelToeplitz operators have played important roles in H1 theory.10) is actually the adjoint of . 2 In the di erential game problem considered later and in the H1 optimal control problem. D 2 D 2 Ry := I . z w > for all w 2 W . we have < z .15. z 2 H2 . The proof of Lemma 15.w >=< .wk2 1. we have. 15. which is given by P (G z ) = P. MG)z . z= . by (15. we will make use of the adjoint .
7 supw2BW k.400 " LINEAR QUADRATIC OPTIMIZATION h i A 0 B2 .wk2 = kP+ Gwk2 + kP. G21 G22 wk < 1: w2BW G12 G22 # 1 = sup P+ G w2 2H2 h " w2 i 0 # 2 2 w2BW sup kP+ Gwk < 1 h i G21 G22 i = sup P. G21 G22 wk2 : 2 2 2 w2BW h i sup kP+ Gwk < 1 i sup kP.C D 2 " # " # h i A 0 C2 . HY := + Ry 1 D2 B C2 : . G21 G22 w2 w2 h # " # 2 0 3 i " w1 w2 # 7 5 : So supw2BW k. Hence w1 = P G w1 + 6 h 4 + P.wk2 < 1 . supw2BW k. : W ! Z as w P 0 . and . The mixed HankelToeplitz operator can be written as " . Proof. G21 G22 wk < 1: These two inequalities then imply that (i) and (ii) are necessary.6.BD2 # " # ? De ne W = H2 L2 .C C .wk < 1 implies and But and h " k. G21 G22 w 1 ? w2H2 w2BW sup kP.BB .A . we will show that given conditions (i) and (ii). Z = H2 L2 .wk2 < 1 holds i the following three conditions hold: (i) (D 2 ) < 1 and HX 2 dom(Ric) (ii) (D2 ) < 1 and HY 2 dom(Ric) (iii) (XY ) < 1 for X = Ric(HX ) and Y = Ric(HY ).A . Analogous to the proof of Lemma 15. 1 = + w2 0 I " # " #" G11 G12 G21 G22 #" w1 : w2 # Lemma 15. HX := + Rx 1 D 2 C B2 .
6 give the desired conclusion. kwk2 x(0) = x0 = x0 Xx0 . if w 2 W then k.5 Linear Quadratic MaxMin Problem Consider the dynamical system x = Ax + B1 w + B2 u _ z = C1 x + D12 u with the following assumptions: (i) (C1 A) is observable (ii) (A B2 ) is stabilizable h i h (15. We will do this by showing.. By saying that.3 This maxmin problem is a game problem in the sense that u is chosen to minimize the quadratic norm of z and that w is chosen to maximize the norm. kwk2 : 2 2 2L2+ w2L2+ A conventional game problem setup would be to consider the minmax problem.3 and 15. it will be seen that they are equivalent and that a saddle point exists. if the assumptions (i) and (iii) on the dynamical system are relaxed to: . kP+ w2 k2 + kP.1 x0 : 2 2 Now the same arguments as in the proof of Lemma 15. subject to some minor modi cations.wk2 . inputs u and w act as \opposing players". By de nition of W . However. It should also be pointed out that the results presented here still hold. switching the order of sup and min. h i 15. This will be elaborated in Chapter 17. In other words.11) (15.wk2 1.15. In this section. kwk2 dt = sup umin kz k2 .12) (iii) D12 C1 D12 = 0 I .5 yield sup k. x0 Y . we would like to warn readers that this may not be true in the general case where z = C1 x + D11 w + D12 u and D11 6= 0. This linear quadratic maxmin problem can be reformulated in the traditional fashion as in the previous sections: w2L2+ Z w2BL2+ 2+ sup umin 2L 1 2+ 0 kz k2 . 2 w2W . then Lemma 15.e. that (XY ) 1 i supw2BW k. we are interested in answering the following question: when sup umin kz k2 < 1? 2L i Remark 15. it is possible that supw inf u < inf u supw . G21 G22 wk2 .wk2 : 2 2 2 2 2 2 Thus if Y is invertible.wk2 .5. i. kP. kwk2 = kP+ z k2 . Linear Quadratic MaxMin Problem 401 holds i condition (iii) holds. In that case. equivalently.
B2 B2 H1 := . To answer our earlier question.9 sup umin kzk2 < 1 if and only if H1 2 dom(Ric) and X1 = 2L2+ w2BL2+ . Lemma 15.C1 C1 .B2 X2 C1 D? 2 RH2 : X2 0 0 h i " # h i (15:14) This implies that U and U? are each inner and that both U? Gc and U Gc are in RH? . where A .402 " LINEAR QUADRATIC OPTIMIZATION A . Theorem 15.B2 B2 : H2 = .A Let F2 = .1 (15:13) U? = AF2 . de ne a Hamiltonian matrix H1 and the 2 associated Riccati equation as A B1 B1 .A " # A X1 + X1 A + X1 B1 B1 X1 .C1 C1 . if the condition is satis ed. Furthermore. ~ " # It is clear from the assumptions that H2 2 dom(Ric) and X2 = Ric(H2 ) > 0. De ne Gc (s) := AF2 I C1F2 0 and " " # U (s) := AF2 B2 C1F2 D12 # " # . The following is easily proven using Lemma 13.8 U U?] is square and inner and a realization for Gc U U? is .X2 C1 D? C1F2 D? where AF2 = A + B2 F2 and C1F2 = C1 + D12 F2 .B2 X1 is an optimal control. and (iii)' D12 has full column rank. j!I B2 C1 D12 # (i)' has full column rank for all ! 2 R.1 Gc U U? = AF2 .B2 X2 and D? be such that D12 D? ] is an orthogonal matrix.29 by obtaining a statespace realization and by eliminating uncontrollable states using a little algebra involving the Riccati equation for X2 . X1 B2 B2 X1 + C1 C1 = 0: Ric(H1 ) > 0. then u = F1 x with F1 := .
de ne := u .15.14). wk2 < 1 w2BH2 " # 2 where . F2x to get z = Gc B1 w + U : Then the hypothesis implies that w2BH2 2H2 h i 403 sup min kz k2 < 1: # " (15:15) # Since by Lemma 15. and U U? z = U G c B1 w + U? Gc B1 w " P (U Gc B1 w) + P+ (U Gc B1 w + ) : = . Linear Quadratic MaxMin Problem Proof. U? Gc B1 w P. : L2+ ! W is de ned as " P (U Gc B1 w) = P.qk2 < 1: . U? Gc B1 w U? with # " # Gc B1 w W := " # (" Note that from equation (15. ()) As in Chapter 14. 1 = P+ (B1 Gc (Uq1 + U? q2 )) = P+ B1 Gc U U? q2 i " q1 q2 # ? q1 2 H2 q2 2 L2 : ) q1 q2 # where Gc U U? ] 2 RH2 has the realization in (15. (U Gc B1 w) + P+ (U Gc B1 w + ) U? Gc B1 w # Thus w2BH2 2H2 sup min kz k2 = sup min w2BH2 2H2 " and the right hand of the above equation is minimized by = . kz k2 = k U U?] z k2. : W ! H2 is given by h q . w= . So we have q2BW sup k.10) the adjoint operator . U . (U Gc B1 w) sup min kz k2 = sup U? Gc B1 w w2BH2 2H2 w2BH2 =: sup k.5.P+ (U Gc B1 w) we have 2 P.8 U U? ] is square and inner.
Petersen. the dynamical system becomes x = AF1 x + B1 w _ z = C1F1 x: So by Corollary 13. (HW ) = T Im I = Im I . W ). (H1 ) = T X. and references therein. Lemma 15. Then # " . H1 = THW T . Bryson and Ho 1975]. (() If H1 2 dom(Ric) and X1 = Ric(H1 ) > 0. B2 B2 X1 is stable.. i.1. A + B1 B1 X1 .14) is X2 1 since . Khargonekar. AF2 X2 1 + X2 1 AF2 + B2 B2 + X2 1 C1 C1 X2 1 = 0: . ..e. supw2L2+ kz k2 < 1 for the given control law.X2 0 # . Limebeer.24. kTzw k1 < 1.6 Notes and References Di erential game is a wellstudied topic. .1 X2 > 0. Anderson.AF2 transformation between H1 and HW . Glover.6 and equation (15.. . the controllability Gramian for the system (15. see e. The current setup and proof are taken from Doyle. equivalently. AF1 := A + B2 F1 C1F1 := C1 + D12 F1 : 2 Theorem 15.404 LINEAR QUADRATIC OPTIMIZATION This is just the condition (15. Note that the observability of (C1 A) implies X2 > 0. The paper by Mageirou and Ho 1977] is one of the early papers that are relevant to the topics covered in this chapter.6 also implies (WX2 1) < 1 or. # AF2 X2 1C1 C1 X2 1 2 dom(Ric) HW := and W = Ric(HW ) 0.X2B1 B1 X2 .e. Khargonekar. Using the Ric> " cati equation for X2 . i.14) we have that " . one can verify that T := " .I X2. .8). and Green 1992].1 X. .g. . and Zhou 1990]. The application of game theoretic results to H1 problems can be found in Basar and Bernhard 1991]. Now with the given control law u = F1 x. Khargonekar. Furthermore. so from Lemma 15. X2 W W X2 so H1 2 dom(Ric) and X1 = X2 (X2 . X2 # W . 15. and Francis 1989].9 will be used in the next chapter to solve the FI H1 control problem. De ne Then the Riccati equation can be written as AF1 X1 + X1 AF1 + C1F1 C1F1 + X1 B1 B1 X1 = 0: We conclude from Lyapunov theory that AF1 is stable since (AF1 B1 X1 ) is detectable and X1 > 0.1 provides a similarity .
8 discusses the H1 separation theory and presents the proof of the output feedback results. we formulate the optimal and suboptimal H1 control problems in section 16. 16 z y w u 16. which are the topics of sections 16. In section 16.2 all suboptimal controllers are characterized for a class of simpli ed problems while leaving the more general problems to the next chapter. Some other interpretations of the H1 controllers are given in section 16.1. Some preliminary analysis is given in section 16. The optimal controllers are also brie y considered in this section. Section 16. Finally.9.H1 Control: Simple Case In this chapter we consider H1 control theory. The analysis suggested the need for solving the Full Information and Output Estimation problems.3 for the output feedback results.7. section 16.10. we will focus on the suboptimal case in this book and discuss why we do so.11 presents the formulas for an optimal H1 controller. The behavior of the H1 controller as a function of performance level is considered in section 16. However.1 Problem Formulation Consider the system described by the block diagram G  K 405 .416. Speci cally.
To state the lemma. Clearly. Suboptimal H1 Control: Given > 0.406 H1 CONTROL: SIMPLE CASE where the plant G and controller K are assumed to be realrational and proper. It will be assumed that state space models of G and K are available and that their realizations are assumed to be stabilizable and detectable. nd all admissible controllers K (s) if there is any such that kTzw k1 < . However. in which the optimal controller is unique and can be obtained by solving two Riccati equations without iterations. Recall again that a controller is said to be admissible if it internally stabilizes the system.. nding an optimal H1 controller is often both numerically and theoretically complicated. Then the internal stability guarantees Tzw = F` (G K ) 2 RH1 . as shown in Glover and Doyle 1989]. This is certainly in contrast with the standard H2 theory. we brie y discuss what will happen when approaches the optimal value.1 Internal Stability and Input/output Stability Now suppose K is a stabilizing controller for the system G. we shall assume that G and K have the following stabilizable and detectable realizations. lower bandwidth. Hence any sensible controller has to be admissible. is minimized. and it is usually much cheaper to obtain controllers that are very close in the norm sense to the optimal ones.g. The following lemma provides the additional (mild) conditions to the equivalence of Tzw = F`(G K ) 2 RH1 and internal stability of the closedloop system.2 Output Feedback H1 Control 16. Optimal H1 Control: nd all admissible controllers K (s) such that kTzw k1 It should be noted that the optimal H1 controllers as de ned above are generally not unique for MIMO systems. which will be called suboptimal controllers.2. Furthermore. 16. A suboptimal controller may also have other nice properties over optimal ones. but the latter does not necessarily imply the internal stability. stability is the most basic requirement for a practical system to work. When appropriate. Knowing the achievable optimal (minimum) H1 norm may be useful theoretically since it sets a limit on what we can achieve. we focus our attention in this book on suboptimal control. e. in practice it is often not necessary and sometimes even undesirable to design an optimal controller. respectively: A B1 B2 6 G(s) = 4 C1 D11 D12 C2 D21 D22 2 3 7 5 ^ ^ K (s) = A B : ^ D C ^ " # . For the reasons mentioned above.
. we get y = 0.1 .1 Suppose that the realizations for G and K are both stabilizable and detectable. and part (b) is a dual result. Since (C A) is detectable.2. Suppose F`(G K ) has undetectable state (x0 y0)0 and mode Re 0 then the PBH test gives " 3 This can be simpli ed as " Ac . I B2 (a) detectable if C1 D12 A . 2 These relations will be used extensively below to simplify our development and to enable us to focus on input/output stability only. D22 D). Hence part (a) is proven. which is a contradiction. This implies Ay = y. Proof. Output Feedback H1 Control is 407 Lemma 16. then K is an internally stabilizing controller i Tzw 2 RH1 . then x = 0 and Cy = 0. I Cc #" #" x = 0: y # # A . I B2 C1 D12 " x ^ 1 C2 x + L2 Cy = 0 ^ DL # and Now if ^ ^ ^ BL1(C2 x + D22 Cy) + Ay . if (a) and (b) hold. I B2 C1 D12 ^ ^ ^ ^ has full column rank. Then the feedback connection Tzw = F` (G K ) of the realizations for G and K A . L2 := (I . y = 0: A .16. I B1 (b) stabilizable if C2 D21 " " # has full column rank for all Re # 0 0: has full row rank for all Re Moreover. DD22 ). The statespace equations for the closedloop are: 2 ^ ^ ^ B1 + B2 DL1 D21 7 B2 L2 C A + B2 DL1 C2 6 ^ ^ ^ ^ ^ 7 BL1 C2 A + BL1 D22 C BL1 D21 F` ( G K ) = 6 5 4 ^ ^ ^ C1 + D12 L2 DC2 D12 L2 C D11 + D12 DL1 D21 " # Ac Bc =: Cc Dc ^ ^ where L1 := (I .1 .
1 = N (M . To show Q 2 RH1 . suppose there exist a (real or in nite) frequency ! and a constant nonzero vector r such that at s = j! . (a) ) (b). Suppose that P P = I .1 2 RH1 : . let Q = NM . (1 . w = P211 (I . P211 2 RH1 and that Q is a proper rational matrix. internal stability and wellposedness follow easily from kP22 k1 1. and P inner implies that kz k2 + krk2 = kwk2 + kvk2. and a small gain argument. consider the closedloop system at any frequency s = j! with the signals xed as complex constant vectors. (b) Q 2 RH1 and kQk1 < 1. v = Qr gives v = Tvw w. (Note that kP22 k1 1 follows from the fact that P22 is a compression of P . kQk1 < 1. that kz k2 kwk2 . Twr = P211 (I . Therefore.2. P22 Q). wellposed.1 = M (M . P22 Q). Therefore. It follows that max (Q(j!)) < 1 for all !. P22 Q). kQrk krk. Proof. 1)krk2 1 . We shall show that M .2 ]kwk2 which implies kTzw k1 < 1 . By internal stability we have Q(I .1 2 RH1 . which is impossible since kz k kTzw k1 < 1.. kQk1 < 1 since Q is rational. Then setting . i. P22 Q) 2 RH1 . P22 N ).408 H1 CONTROL: SIMPLE CASE 16. Theorem 16. the system internal stability is guaranteed if (I . Then the following are equivalent: (a) The system is internally stable. 2 ) .e. Then kwk krk.) To show that kTzw k1 < 1. (b) ) (a). But as above. P22 Q)r. Note that since P Q 2 RH1.1 with N M 2 RH1 be a right coprime factorization.1 2 RH1 . i. hence.e. P inner implies that 2 + krk2 = kwk2 + kv k2 and. Also let := kTwr k1 .2 Contraction and Stability z w One of the keys to the entire development of H1 theory is the fact that the contraction and internal stability is preserved under an inner linear fractional transformation.2 Consider the following feedback system: P r  Q v P = P11 P12 2 RH1 P21 P22 " # . there exist X Y 2 RH1 such that XN + Y M = I .1 2 RH1 and (I .. and kTzw k1 < 1. P22 N ). kz k2 kwk2 + ( 2 . Let kQk1 =: < 1 and note that . To show that kQk1 < 1.
as s traverses the Nyquist contour. Of course. but is not needed . det(I .1] (this uses the fact that kP22 k1 1 and kQk1 < 1).1 2 RH1 : This implies that the winding number of det(M .3 below. the theorems and proofs can be modi ed so that the given formulae are still correct.2. Q 2 RH1 . the problem considered below contains the essential features of the H1 theory. P22 Q).1 + Y (I . equals zero. The main reason for doing so is that the general case has its unique features but is much more involved algebraically. Output Feedback H1 Control Thus 409 XQ(I . We conclude that the winding number of det M also equals zero. and the proof is complete. as will be seen in the next chapter. This equivalence enables us to focus only on input/output stability and is captured in Corollary 16. The realization of the transfer matrix G is taken to be of the form A B1 B2 6 G(s) = 4 C1 0 D12 7 : 5 C2 D21 0 The following assumptions are made: (i) (A B1 ) is stabilizable and (C1 A) is detectable (ii) (A B2 ) is stabilizable and (C2 A) is detectable (iii) D12 C1 D12 = 0 I " h i h i 2 3 (iv) B1 D = 0 . Also. P22 Q). P22 N ) det M . Nevertheless.1] and all s = j!. This assumption simpli es the theorem statements and proofs. P22 N ). An important simpli cation that is a consequence of the assumption (i) is that internal stability is essentially equivalent to inputoutput stability (Tzw 2 RH1 ). and we have det(M . P22 N ). P22 Q) = det(M . assumption (ii) is necessary and su cient for G to be internally stabilizable.1 6= 0. 2 16. we discuss a simpli ed version of H1 theory.1. det(I . for all s = j!.16. P22 Q) 6= 0 for all in 0. P22 N ) 6= 0 for all in 0.1 = (M .3 Simplifying Assumptions In this chapter. but if it is relaxed. The general case will be considered in the next chapter. Involved algebra may distract attention from the essential ideas of the theory and therefore lose insight into the problem. det M . D21 21 I # " # Assumption (i) is made for a technical reason: together with (ii) it guarantees that the two Hamiltonian matrices (H2 and J2 in Chapter 14) in the H2 problem belong to dom(Ric). Now note the fact that. Therefore.2.
However.1 are satis ed.) Corollary 16. (iii). Then a controller K is admissible i Tzw 2 RH1 . 16. (Readers should be clear that this does not mean that the realization for G need not be stabilizable and detectable. I B2 rank C1 D12 " # 2 = rank 6 4 2 I = rank 6 4 So has full column rank for all Re " D12 D? A. Proof. D22 6= 0 does not pose any problem since it is easy to form an equivalent problem with D22 = 0 by a linear fractional transformation on the controller K (s).410 H1 CONTROL: SIMPLE CASE to prove the equivalence of internal stability and Tzw 2 RH1 . I B2 C1 D12 3 # D? C1 # 0 " B2 # I 7: 5 0 0i A . 2 Two additional assumptions that are implicit in the assumed realization for G(s) are that D11 = 0 and D22 = 0. We only need to verify that the rank conditions of the two matrices in Lemma 16. D12 D12 )C1 = C1 . In point of fact. and (iv) hold. I B2 C1 D12 " A. I D? C1 # has full column rank.3 Suppose that assumptions (i). The realization for plant G is stabilizable and detectable by assumption (i). Since D? (D? C1 ) = (I .4 Suboptimal H1 Controllers In this subsection. furthermore. relaxing the assumption D11 = 0 complicates the formulae substantially. However. Then A .2. we present the necessary and su cient conditions for the existence of an admissible controller K (s) such that kTzw k1 < for a given . (D? C1 A) is detectable i (C1 A) is detectable. as will be seen in the next chapter. . The rank condition for the other matrix follows by duality. I # " 0 " 0 3 " # 7 5 A . Now suppose assumptions (i) and (iii) are satis ed and let D? be such that i h D12 D? is a unitary matrix. the last matrix has full rank for all Re 0 i (D? C1 A) is detectable. the internal stability and inputoutput stability can never be equivalent if either G or K has unstabilizable or undetectable modes. As we have mentioned many times. and.
one such controller is ^ Ksub (s) := A1 . Indeed. . . B2 B2 # . we characterize all admissible controllers that satisfy the norm condition. Note that the (1.2)blocks are a suggestive combination of expressions from the H1 norm characterization in Chapter 4 (or bounded real ARE in Chapter 13) and from the H2 synthesis of Chapter 14.e. Let opt := min fkTzw k1 : K (s) admissibleg. the optimal level. which is often called the central controller or minimum entropy controller. Recall that similar di erences arose in the norm computation problem as well. Moreover. Although not as apparent as in the H2 case. and this is one major di erence between the H1 and H2 results. Furthermore.A A J1 := .B1 B1 " . Output Feedback H1 Control 411 if the necessary and su cient conditions are satis ed.2)blocks are not sign de nite.B2 X1 L1 := . Optimal H1 controllers are more di cult to characterize than suboptimal ones. Theorem 16.7 in Chapter 13 to guarantee that H1 2 dom(Ric) or Ric(H1 ) 0. In Section 16.9 we will brie y discuss how to nd an admissible K to minimize kTzw k1 . It is also clear that if approaches in nity.2.2 C1 C1 . The proofs of these results will be given in the later sections. then these two Hamiltonian matrices become the corresponding H2 control Hamiltonian matrices. these conditions are intimately related to the existence of H1 suboptimal controllers.Z1 L1 F1 0 where " # ^ A1 := A + .16.Y1 C2 Z1 := (I . The reasons for the form of these expressions should become clear through the discussions and proofs for the following theorem.2 B1 B1 . The H1 solution involves the following two Hamiltonian matrices: A H1 := .A : The important di erence here from the H2 problem is that the (1.1 : The H1 controller displayed in Theorem 16.4 There exists an admissible controller such that kTzw k1 < i the following three conditions hold: (i) H1 2 dom(Ric) and X1 := Ric(H1) 0 (ii) J1 2 dom(Ric) and Y1 := Ric(J1 ) 0 (iii) (X1 Y1 ) < 2 .2 Y1 X1 ). i. clearly. the H1 controller also has an interesting separation structure.4. so we cannot use Theorem 13.C1 C1 " . C2 C2 # . when these conditions hold. Then. must be greater than opt for the existence of suboptimal H1 controllers. has certain obvious similarities to the H2 controller as well as some important di erences..2B1 B1 X1 + B2 F1 + Z1 L1 C2 F1 := .
the suboptimal controllers are parameterized by a xed linearfractional transformation with a free parameter Q. and then we can di erentiate x(t) X1 x(t): d _ _ dt (x X1 x) = x X1 x + x X1 x .5 If conditions (i) to (iii) in Theorem 16. we recover the central controller Ksub (s). require the ltering and full information (i.4.8. kQk1 < . The term central controller will be obvious from the parameterization of all suboptimal controllers given below. while the meaning of minimum entropy will be discussed in Section 16. Assuming that X1 exists.2.4 is constructed out of these results as well. Theorem 16.7 and 16.412 H1 CONTROL: SIMPLE CASE each of the conditions in the theorem can be given a systemtheoretic interpretation in terms of this separation.C2 2 3 7 5  Q where Q 2 RH1 . state feedback) results in sections 16.1. given in Section 16. With Q = 0 (at the \center" of the set kQk1 < ). we try to motivate the study of the OE (and hence other special problems) and show how this problem arises naturally in proving the output feedback results. Then X1 satis es the following Riccati equation: A X1 + X1 A + C1 C1 + . X1 B2 B2 X1 = 0: (16:1) Let x denote the state of G with respect to a given input w. As in the H2 case. the set of all admissible controllers such that kTzw k1 < equals the set of all transfer matrices from y to u in u M1 y ^ A1 . Suppose X1 := Ric(H1 ) exists.2 X1 B1 B1 X1 . we shall now try to give some ideas for approaching the problem.e. which is Theorem 16.4 are satis ed. The proof of Theorem 16. The following theorem parameterizes the controllers that achieve a suboptimal H1 norm less than . 16.Z1 L1 Z1 B2 6 M1 (s) = 4 F1 0 I I 0 . Speci cally.10. These interpretations. we will show that the general output feedback problem boils down to an output estimation problem which can be solved easily if a state feedback or full information control problem can be solved. The key is to use the fact that contraction and internal stability are preserved under an inner linear fractional transformation.3 Motivation for Special Problems Although the proof for output feedback results will be given later..
2B1 X1 xk2 = kvk2 .kz k + kwk . Motivation for Special Problems = x (A X1 + X1 A)x + 2hw B1 X1 xi + 2hu B2 X1 xi: Using the Riccati equation for X1 to substitute in for A X1 + X1 A gives 413 d (x X x) = . .1B1 X1 2 6 4 3 2 2 I 0 B2 D12 0 32 76 54 32 76 54 AF1 := A + B2 F1 C1F1 := C1 + D12 F1 and x _ Atmp B1 B2 7 6 = 4 .4.16.3. w 2 L2+ . . This is shown in the following diagram: XXXXX P w w XX v r u 1 Gtmp y  K . 2 kwk2 = ku + B2 X1 xk2 .kC xk2 . the closedloop system can be expressed as two interconnected subsystems below: 2 6 4 where v := u + B2 X1 x . 2 kw .F1 0 I v5 y C2 D21 0 z r x r u 3 7 5 Atmp := A + . .2 B1 X1 x: x w v 3 7 5 x _ z r 3 A F1 7 6 C1F1 5=4 .2 2 2 dt (x X1 x) = .2 B1 B1 X1 where F1 is de ned as in Section 16. and integrate (16. 2 krk2 (16:3) 2 2 2 2 2 2 (16:4) With these new de ned variables. kw . .2.2) from t = 0 to t = 1: kz k2 . B1 X1 xk + ku + B2 X1 xk : (16:2) Assume x(0) = x(1) = 0.1 B1 r := w . completion of the squares along with orthogonality of C1 x and D12 u gives the key equation d 2 2 2 2 .2 kB X xk2 + kB X xk2 +2hw B X xi +2hu B X xi: 1 1 1 1 2 1 1 1 2 1 dt Finally.
Then from standard results on Lyapunov equations (see Lemma 3. (AF1 + .3) also suggests that u = . Equation (16. Lemma 16.2 B1 X1 x is the worst disturbance input in the sense that it maximizes the quantity kz k2 . and AF1 6 = 4 C1F1 . .F1 0 I 7 : (16:6) 5 C2 D21 0 The equality (16. 2 kwk2 in (16.414 where H1 CONTROL: SIMPLE CASE P := P11 P12 P21 P22 " # While this is the basic idea behind the proof of Lemma 16. the output being estimated is the optimal FI control input F1 x and the new disturbance r is o set by the \worst case" FI disturbance input wworst . and these variables provide the connection between Tzw and Tvr .3) can be written as kz k2 + k rk2 = k wk2 + kvk2 : 2 2 2 2 " # " # .3) for the minimizing value of u = 2 2 .) It is also interesting to note that wworst is the optimal strategy for w in the corresponding LQ game problem (see the di erential game problem in the last chapter).8 below.1 Tzw and Tv( r) = . Note that wworst := . In terms of the OE problem for Gtmp . It is immediate from equality (16. . the u making v = 0 and the w making r = 0 are values satisfying a saddle point condition. We will show below that the output feedback can indeed be transformed into the OE problem.3) that kTzw k1 i kTvr k1 . To show this we rst need to prove some preliminary facts. (See Section 17.4).1Tvr . Rearrange the Riccati equation for X1 and use the de nition of F1 and C1F1 to get C1F1 C1F1 AF1 X1 + X1 AF1 + = 0: (16:7) . . Atmp B1 B2 6 Gtmp = 4 .3) motivates the change of variables to r and v as in (16.1B1 X1 Since H1 2 dom(Ric).8 for more interpretations. 2 Equation (16. Then AF1 = A + B2F1 is stable i X1 0. Notice the structure of Gtmp : it is an OE problem.1 B1 0 I 3 B2 D12 0 3 7 5 (16:5) Proof.B2 X1 x that is. the details needed for strict inequality and internal stability require a bit more work. This will be shown later.6 Suppose H1 2 dom(Ric) and X1 = Ric(H1).2 B1 B1 X1 ) is stable and hence (B1 X1 AF1 ) is detectable.B2 X1 x is a suboptimal control for a full information (FI) problem if the state x is available.19).1 B1 X1 . AF1 is stable i X1 0. Note that Tz( w) = .1B1 X1 2 2 .
1 B1 C1F1 0 I + X1 = 0: D12 0 . By Lemma 16. Proof.8. I B1 I 0 C2 " D21 C2 D21 .2 B1 B1 X1 ).16. Finally. 415 0..5) is in Lemma 16.8 Assume H1 2 dom(Ric) and X1 = Ric(H1) 0. Similarly since " # " #" # Atmp . is also equivalent to Tvr 2 RH1 .7 If H1 2.3 that internal stability for Tzw is equivalent to Tzw 2 RH1 . i. which is veri ed by the following 0. I B2 = det(A + B F . . P21 2 The following lemma connects these two systems Tzw and Tvr .1 the internal stability of Tvr . we have that " # Atmp . AF1 is stable.dom(Ric) and X1 = Ric(H1) RH1 and inner. Recall that internal and inputoutput stability are equivalent for admissibility of K in the output feedback problem by Corollary 16. which is stable by de nition. Proof. the state matrix for P211 is (AF1 + . Thus. then P in (16.e. Recall from Corollary 16. I B2 . I ) 6= 0 det tmp tmp 2 1 . Lemma 16.F1 I # has full column rank. and P211 2 RH1 . .29 upon noting that the observability Gramian of P is X1 (see (16.2 B1 X1 I has full row rank for all Re( ) 0 and since A . Hence by Lemma 16. which is the central part of the separation argument in Section 16.3.3 . I B1 = A . We may assume without loss of generality that the realization of K is stabilizable and detectable. Motivation for Special Problems This suggests that P might be inner when X1 lemma.6. Then K is admissible for G and kTzw k1 < i K is admissible for Gtmp and kTvr k1 < . That P is inner (P P = I ) follows from Lemma 13.1B1 X1 B2 . Thus internal stability is equivalent to inputoutput stability for both .7)) and " #" # " # .F1 I for all Re( ) 0 by the stability of Atmp + B2 F1 .2. So P 2 RH1 .1 2 RH1 . the internal stability of the subsystem consisting of Gtmp and controller K .
a special case of the output feedback problem because it does not satisfy all of the assumptions.2 and Lemma 16.7 along with the above block diagram that Tz( w) 1 < 1 i Tv( r) 1 < 1 or. In particular. To show (a). 16. This shows that K is an admissible controller for G if and only if it is admissible for Gtmp . So in the sections to follow. we will focus on these special problems.4 Full Information Control Our system diagram in this section is standard as before z y with 2 G w u 3  K B1 " 0 5 0 I 0 The H1 problem corresponding to this setup again is not. it should be noted that for the FI (and FC in the next section) problem. kTzw k1 < i kTvr k1 < . we need to solve a FI problem. strictly speaking. 2 From the previous analysis. # I " I 0 B " 1# 0 3 7 5 I . it is clear that to solve the output feedback problem we need to show (a) H1 2 dom(Ric) and X1 = Ric(H1 ) 0 (b) kTvr k1 < . internal stability is not equivalent to Tzw 2 RH1 since 2 6 4 A 6 6 G(s) = 6 " C1 # 6 I 4 0 0 # B2 7 D 7 " 12# 7 : 7 A. Now it follows from Theorem 16. The problem (b) is an OE problem which can be solved by using the relationship between FC and OE problems in Chapter 12. equivalently. while the FC problem can be solved by using the FI solution through duality.416 H1 CONTROL: SIMPLE CASE G and Gtmp .
h i h i Theorem 16. Assumptions (iv) and the second part of (ii) for the general output feedback case have been e ectively strengthened because of the assumed structure for C2 and D21 . a suboptimal controller exists which also uses just x. although this presents no di culties in solving this problem.4. then the equivalence class of all admissible controllers satisfying kTzw k1 < can be parameterized as K (s) = F1 . we want the methods to extend to the optimal case in a natural and straightforward way.2 Q(s)B1 X1 Q(s) where Q 2 RH1 . We have seen that in the H2 FI case.4. We will show below that. The assumptions relevant to the FI problem which are inherited from the output feedback problem are (i) (C1 A) is detectable (ii) (A B2 ) is stabilizable (iii) D12 C1 D12 = 0 I . which is more traditional. K admissible means internally stabilizing. once one gets outside the pure H2 setting.9 may each be given their own interpretations. the full information problem is more fundamental and more natural than the state feedback problem. in the H1 case. The condition that H1 2 dom(Ric) implies that X1 := Ric(H1 ) exists . This case could have been restricted to state feedback.9. One setting in which the full information case is more natural occurs when the parameterization of all suboptimal controllers is considered. The two individual conditions in Theorem 16. We simply must remember that in the FI case. Even though the optimal problem is not studied in detail in this book. . not just Tzw 2 RH1 . This di erence is essentially captured by the necessary and su cient conditions for the existence of a controller given in Theorem 16. but we believe that.16. the optimal controller uses just the state x even though the controller is provided with full information. if these conditions are satis ed. Full Information Control 417 can never have full row rank. h i (16:8) It is easy to see by comparing the H1 solution with the corresponding H2 solution that a fundamental di erence between H2 and H1 controllers is that the H1 controller depends on the disturbance through B1 whereas the H2 controller does not. Note that these conditions are the same as condition (i) in Theorem 16. It is also appropriate when studying the general case when D11 6= 0 in the next chapter or when H1 optimal (not just suboptimal) controllers are desired.9 There exists an admissible controller K (s) for the FI problem such that kTzw k1 < if and only if H1 2 dom(Ric) and X1 = Ric(H1 ) 0. Furthermore. kQk1 < .
Proof.8). Further. Now all v satisfying this inequality can be written as v = Qr 2 2 for Q 2 RH1 and kQk1 < or as u + B2 X1 x = Q(w . 2 kwk2 = ku + B2 X1 xk2 . we have kP11 k1 < 1 and kTzw k1 = kP11 k1 < .8). there is an intuitive way to see why all the equivalence classes of controllers can be parameterized in the form of (16.3): .2 B X1 xk2 = kvk2 . strengthen the assumption on (C1 "A) from detectable to observable. and A22 stable. If (C1 A) is ists an admissible controller K ^ ^ ^ C D1 D2 " # x detectable but not observable. giving the following closedloop state equations: 2 6 6 6 6 6 6 4 x1 _ x2 _ _ x ^ z u 3 A11 7 6 7 6 A21 7 6 7 = 6 B11 ^ 7 6 7 6 5 4 C11 ^ D11 2 0 ^ 12 D A22 ^ B12 0 0 0 ^ A 0 ^ C B11 B21 B12 B22 ^ B2 0 0 D12 ^2 0 D 32 76 76 76 76 76 76 54 x1 x2 7 7 7 x 7: ^ 7 w 7 5 u 3 .6. 2 kwk2 < 0 for w 6= 0. without loss of generality. This gives equation (16. 2 krk2 < 0. Remark 16. 2 kw . since Tzw = P11 and P11 P11 = I . we will show that we can. This implies that r 6= 0 and 2 2 kvk2 . to this K stabilizing Tzw . ()) For simplicity. Recall the following equality from equation (16. 2 2 2 0.418 and K (s) = F1 0 gives Tzw as h i H1 CONTROL: SIMPLE CASE Tzw = AF1 B1 C1F1 0 " # AF1 = A + B2 F1 C1F1 = C1 + D12 F1 : (16:9) Furthermore. 2 krk2 : 1 2 2 2 kTzw k1 < means that kz k2 . then change coordinates for the state of G to 1 with x2 x2 unobservable.1 Given that H1 2 dom(Ric) and X1 = Ric(H1) kz k2 . . ~ We shall now prove the theorem. (C11 A11 ) observable.2 B1 X1 x). P21 P21 by P P = I . The further condition that X1 0 is equivalent. in the proof to follow we assume that the system is normalized such that = 1. by Lemma 16. Suppose there ex# ^ ^ ^ ^ = A B1 B2 such that kTzw k1 < 1.
Full Information Control 419 If we take a new plant Gobs with state x1 and output z and group the rest of the equations as a new controller Kobs with the state made up of x2 and x. (() Suppose H1 2 dom(Ric) and X1 = Ric(H1 ) 0 and suppose K (s) is an admissible controller such that kTzw k1 < 1.9. But note that the existence of an admissible controller such that kTzw k1 < 1 is equivalent to that the admissible controller makes sup kz k2 < 1 hence.C11 C11 . But from Theorem 15. i. Hence the necessity is proven. F1 x . the latter is true if and only if H1 2 dom(Ric) and X1 = Ric(H1 ) > 0.4. Again change variables to v := u .16.A11 " # # ^ Ric(H1 ) = X1 = X1 0 0 0 0 exists for G. B21 B21 . then ^ A11 6 6 Gobs (s) = 6 "C11# 6 I 4 0 2 B11 " 0 0 # B21 D " 12# 0 0 3 7 7 7 7 5 I still satis es the assumptions of the FI problem and is stabilized by Kobs with the closedloop H1 norm kTzw k1 < 1 where ^ ^ ^ ^ A22 + B22 D12 B22 C A21 + B22 D11 B12 + B22 D2 6 7 ^ ^ ^ ^ Kobs = 4 B12 A B11 B2 5: ^ 12 ^ ^ 11 ^2 D C D D ^ If we now show that there exists X1 > 0 solving the H1 Riccati equation for Gobs . it is necessary w2BL2+ that sup umin kz k2 < 1 2L w2BL2+ 2+ since the latter is always no greater than the former by the fact that the set of signals u generated from admissible controllers is a subset of L2+ . We shall suppose that there exists an admissible controller such that kTzw k1 < 1. We can therefore assume without loss of generality that (C1 A) is observable. ^ X1 = Ric then " 2 3 A11 B11 B11 .e..
B1 X1 x. Theorem 16. 2 Remark 16. ~ The following theorem gives all full information controllers. F1 0 = Q . F1 0 w " # h i x r = w . then v = Qr and x v = u . this problem will not occur in output feedback case. and that is why \equivalence class" is used.2 It should be emphasized that the set of controllers given above does not parameterize all controllers although it is su cient for the purpose of deriving the output feedback results. P is inner and P211 2 RH1 . Nevertheless.10 Suppose the condition in Theorem 16. Q(s)B1 X1 Q(s) i Q 2 RH1 kQk1 < 1: Hence all suboptimal controllers are in the stated equivalence class. By Theorem 16.420 H1 CONTROL: SIMPLE CASE z w and r := w .B1 X1 I 0 2 3 7 5 . B1 X1 x = . By Lemma 16.Tvr v P = P11 P12 P21 P22 " # AF1 B1 B2 = 6 C1F1 0 D12 4 .9 is satis ed then all admissible controllers satisfying kTzw k1 < can be parameterized as K = F` (MFI Q): .B1 X1 I : w h i " # Hence we have K .B1 X1 I : K (s) = F1 . there is no choice of Q such that K1 belongs to the set. F1 x = K . It is clear that h i there is a suboptimal controller K1 = F1 0 with F1 6= F1 however.B1 X1 I w i h i # i " x w # K .2 the system is internally stable and kTzw k1 < 1 i Tvr 2 RH1 and kTvr k1 < 1.7. so that the closedloop system is as shown below: P r . F1 0 or Thus h h h i " x = Qh . Now denote Q := Tvr .
I 0 v and 2 3 2 3 Atmp B1 B2 2 3 x _ 6 7 .16. The parameterization of all suboptimal FC controllers follows by duality and. we have h i K = F1 .3 It is simple to verify that for Q1 = 0. are omitted. therefore. B1 X1 I This is shown pictorially in the following diagram: z r XXXXX X ^ GFI P w v r u 1 w y  K .1B1 X1 r .I 0 0 # where Q = Q1 Q2 2 RH1 and kQ2k1 < . We only need to show that F`(MFI Q) with kQ2k1 < parameterizes all FI H1 suboptimal controllers. 0 B1 F1 0 i i # B2 I 0 3 7 7 7 7 7 5 Proof. . To show that.2 B1 X1 x: Then the system equations can be written as follows: 2 3 2 . Remark 16.9. . we shall make a change of variables as before: v = u + B2 X1 x r = w .4.2 where Atmp := A + .2Q2 B1 X1 Q2 which is the parameterization given in Theorem 16.F1 # " 0 # I 7 6 x 7 6 7 6 " 6 74 r 5 4 v 5=6 7 I 0 4 0 5 u y . ~ h I 0 .2 B1 B1 X1 .1 B1 B2 3 2 x 3 x _ AF1 6 7 6 C1F1 0 D12 7 6 w 7 4 z 5=4 54 5 . Full Information Control u MFI y MFI (s) = i 2 6 6 6 6 6 4 421 A + B2 F1 " h h "  Q .2B1 X1 I .
^ Also note that Atmp + B2 F1 is stable.2 B1 X1 # . Then all icontrollers that stabilize GFI can be h parameterized as K = F` (M ) = 1 2 2 RH1 where . the transfer matrix from r to v is Tvr = N =6 4 " 0 0 # I 0 3 7 5 I and Tvr = F` (N ) = 2 .2B1 X1 # 2 6 6 6 6 6 6 4 h i then Atmp + L I Atmp + B2 F1 + L " M= . It is easy to show that With this parameterization of all controllers.5) and 2 6 4 " 6 ^ GFI = 6 6 H1 CONTROL: SIMPLE CASE Atmp . 2 . we conclude that K is an admissible controller ^ for G and kTzw k1 < i K is an admissible" controller for GFI and kTvr k1 < .2B1 X1 2 I . Hence kTvr k1 < if and only if k 2 k1 < .2 I B1 X1 Now let 1 = F1 . . " F1 I .B1 # h i 0 I B2 7 I 7: 7 7 0 5 3 Then it is easy to show that F` (M ) = F`(MFI Q).422 where P is as given in equation (16. This implies that all FI H1 controllers can be parameterized as K = F`(M ) with = h i 1 2 2 RH1 k 2 k1 < and A + 2B2 F1 6 6 M = 6 " F1 6 4 .8.2 B1 X1 B1 # " I 0 0 # B2 7 I 7: 7 7 0 5 3 So from Theorem 16.2 Q2 B X1 + Q1 1 2 = Q2 : 2 .F1 I . # Now let L = B2 F1 . .B1 = A + B2 F1 is stable. B2 F1 .2 and Lemma 16.L B2 0 3 7 7 7 7 7 7 5 I 0 : I ^ F` (GFI F` (M )) =: F` (N ).
5.11 is the same as that in (ii) of Theorem 16.16.6 Disturbance Feedforward A B1 B2 G(s) = 6 C1 0 D12 4 C2 I 0 2 3 7 5 This problem inherits the same assumptions (i)(iii) as in the FI problem. With this assumption. Y1 C1 Q(s) Q(s) " # where Q 2 RH1 . .11 There exists an admissible controller K (s) for the FC problem such that kTzw k1 < if and only if J1 2 dom(Ric) and Y1 = Ric(J1 ) 0. The assumptions that the FC problem inherits from the output feedback problem are just the dual of those in the FI problem: (i) (A B1 ) is stabilizable (ii) (C2 A) is detectable " # " (iv) B1 D = 0 . it is easy to check that the condition in Lemma 16. the condition in Theorem 16. B1 C2 is stable.2 K (s) = L1 . but for internal stability we shall add that A . kQk1 < . As expected. as in the output feedback case. Moreover. Full Control 423 16.5 Full Control A B1 6 6 G(s) = 6 C1 0 4 C2 D21 2 h h h I 0 0 I 0 0 i 3 i 7 7 7 i 5 This problem is dual to the Full Information problem. I D21 21 # Theorem 16.4.1 is satis ed so that the internal stability is again equivalent to Tzw 2 RH1 . 16. if these conditions are satis ed then an equivalence class of all admissible controllers satisfying kTzw k1 < can be parameterized as .
Suppose there is a controller KDF solving the habove problem.e. 3 2 2 I 0 B2 D12 0 32 76 54 x w v 3 3 7 5 and 2 6 4 x _ Atmp B1 B2 v 7=6 .4. if these conditions are satis ed then all admissible controllers satisfying kTzw k1 < can be parameterized as the set of all transfer matrices from y to u in u M1 y  Q A + B2 F1 . the controller KFI = KDF C2 I solves the corresponding H1 FI problem.2 Q(s)B1 X1 Q(s) where PDF is as de ned in section 12. . with kTzw k1 < i .2 B1 X1 I 0 y C2 + 32 76 54 x r 7: 5 u This is shown pictorially in the following diagram: . So again by Theorem 12. i. It is easy to verify that F` (M1 Q) = F` (PDF KFI ) with KFI = i F1 .2 of Chapter 12.424 H1 CONTROL: SIMPLE CASE Theorem 16. we shall make a change of variables as in equation (16. Hence the conditions H1 2 dom(Ric) and X1 = Ric(H1) 0 are necessarily satis ed. To show that F` (M1 Q) with kQk1 < parameterizes all DF H1 suboptimal controllers. Moreover. . the controller F`(M1 Q) solves the DF problem.4. if these conditions are satis ed then FI h is solvable. On the other hand.2 B1 X1 I 0 . kQk1 < Proof.2 B1 X1 x: .F1 0 I 5 4 . B1 C2 B1 B2 6 M1 (s) = 4 F1 0 I . Then by Theorem 12.12 There exists an admissible controller for the DF problem such that kTzw k1 < if and only if H1 2 dom(Ric) and X1 = Ric(H1 ) 0..4): v = u + B2 X1 x 2 6 4 r = w .C2 . 2 3 7 5 with Q 2 RH1 .1 B1 Then the system equations can be written as follows: x _ z r 3 AF1 7 6 C1F1 5=4 .1B1 X1 .
2 and AF1 6 P = 4 C1F1 .1 for system GDF are satis ed. the rank ^ conditions of Lemma 16. Output Estimation z r P w v r u 1 425 w XXXX XXX y ^ GDF where 2 K .8 that F` (M1 Q) ^ with Q 2 RH1 (no norm constraint) parameterizes all stabilizing controllers for GDF ^ DF F`(M1 Q)) = Q. Now it is easy to see by comparing this formula with the controller parameterization in Theorem 12. So from Theorem 16.F1 0 I 7: 4 5 C2 + .2 B1 X1 I 0 Since Atmp .7.2 and Lemma 16. B1 C2 and Atmp + B2 F1 are stable.16. B1 (C2 + .7.7 Output Estimation A B1 B2 6 G(s) = 4 C1 0 I C2 D21 0 2 3 7 5 .1B1 X1 2 0 I B2 D12 0 3 7 5 3 16. we conclude that K is an admissible controller for G and kTzw k1 < i K ^ is an admissible controller for GDF and kTvr k1 < .1B1 B1 B2 Atmp ^ DF = 6 G . and F` (M1 Q) with Q 2 RH1 and kQk1 < parameterizes all suboptimal controllers for G. So kTvr k1 < however.2 B1 X1 ) = A . . simple algebra shows that Tvr = F` (G i kQk1 < .
by duality. the essence of these remarks was made before.4 and 16. Both optimal estimators are observers with the observer gain determined by Ric(J1 ) and Ric(J2 ). .5 using separation arguments. B2 C1 is stable (ii) (C2 A) is detectable " (iv) B1 D = 0 . Moreover. The notation and assumptions for this section are as in Section 16. In contrast. It is interesting to compare H1 and H2 in the context of the OE problem.B2 . . kQk1 < . which are proven in the previous sections. 16. This essentially involves reducing the output feedback problem to a combination of the Full Information and the Output Estimation problems. Optimal H2 output estimation consists of multiplying the optimal state estimate by the output map C1 . The separation properties of the H1 controller are more complicated than the H2 controller.13 There exists an admissible controller for the OE problem such that kTzw k1 < if and only if J1 2 dom(Ric) and Ric(J1 ) 0. just as FC was to FI. we can now prove Theorems 16. Thus the discussion of the DF problem is relevant here. although they are no less interesting. D21 21 I # " # Assumption (i).426 H1 CONTROL: SIMPLE CASE This problem is dual to DF.8 Separation Theory If we assume the results of the special problems. The OE assumptions are (i) (A B1 ) is stabilizable and A . Theorem 16.2. when appropriately dualized. if these conditions are satis ed then all admissible controllers satisfying kTzw k1 < can be parameterized as the set of all transfer matrices from y to u in u M1 y  Q A + L1 C2 . Thus optimal H2 estimation depends only trivially on the output z that is being estimated. together with (iv). the H1 estimation problem depends very explicitly and importantly on the output being estimated. B2 C1 L1 . and state estimation is the fundamental problem. even though. imply that internal stability is again equivalent to Tzw 2 RH1 . This will have implications for the separation properties of the H1 output feedback controller. as in the output feedback case.2Y1 C1 6 M1 (s) = 4 C1 0 I C2 I 0 2 3 7 5 with Q 2 RH1 .
the state equations above .B2 X1 L1 := . and Y1 . The necessary and su cient conditions for the existence of an admissible controller such that kTzw k1 < are (i) H1 2 dom(Ric) and X1 := Ric(H1 ) 0 (ii) J1 2 dom(Ric) and Y1 := Ric(J1 0 (iii) (X1 Y1 ) < 2 .2 B1 B1 X1 + . given the FI and FC results.2Y1 A X1 + .Y1 C2 Z1 := (I . there is a dual form that can be obtained by inspection from the above formula. Proof. Note that all the conditions of Theorem 16. formula: _ (I . H1 2 dom(Ric) and X1 := Ric(H1 ) 0. but the formula for the controller is not. Separation Theory 427 16. Let K be an admissible controller for which kTzw k1 < . The controller K C2 D21 ] solves the FI problem hence from Theorem 16. and that is provided by (iii). the H1 controller can be written as _ x = Ax + B1 wworst + B2 u + Z1 L1 (C2 x .16. Then conditions (i) and (ii) hold. Needless to say.2Y1 C1 C1 .14 Suppose there exists an admissible controller making kTzw k1 < .9. which is an elegant combination of elements from FI and FC. . Condition (ii) follows by the dual argument. Lemma 16.2 Y1 X1 )x = As x . X1 .4 are " # ^ A1 . For a symmetric formula.Z1 L1 Ksub (s) := F1 0 ^ A1 := A + .1 H1 Controller Structure The H1 controller formulae from Theorem 16. though more complicated.4 are symmetric in H1 .2B1 B1 X1 + B2 F1 + Z1 L1 C2 F1 := . y) ^ ^ ^ ^ . L1 y ^ ^ (16. 2 We would also expect some condition beyond these two.1 where X1 := Ric(H1 ) and Y1 := Ric(J1 ).8. We have seen that condition (i) corresponds to the Full Information condition and that (ii) corresponds to the Full Control condition.11) where As := A + B2 F1 + L1 C2 + . these conditions are necessary for the output feedback case as well. It is easily shown that.8.10) u = F1 x ^ (16. To emphasize its relationship to the H2 controller formulae.2 Y1 X1 ). can be multiplied through by Z11 and put in descriptor form. A simple substitution from the Riccati equation for X1 will then yield a symmetric. J1 . .
in some sense.F1 y C2 D21 0 z y G 2 3 7 5 " r =G r tmp u u # Atmp := A + . Hence we only need to show that if conditions (i) and (ii) are satis ed. we de ne new disturbance and control variables r := w . to actually use the OE results.8 that K is admissible for G and kTzw k1 < i K is admissible for Gtmp and kTvr k1 < .14 that conditions (i) and (ii) are necessary for kTzw k1 < . The term wworst may be thought ^ of loosely as an estimate for wworst .2 B1 X1 x Then " # v := u + B2 X1 x: " # B B A v = 6 tmp 1 2 0 I 4 . It is therefore not surprising that Z1 L1 should enter in the controller equations instead of L1 .2 B1 X1 x is. which is the optimal Full Information control input.4 It has been shown from Lemma 16. condition (iii) is necessary and su cient for kTzw k1 < . 16.2 Proof of Theorem 16. The obvious questions that arise when these formulae are compared with the H2 formulae are 1) Where does the term B1 wworst come from? ^ 2) Why Z1 L1 instead of L1 ? 3) Is there a separation interpretation of these formulae analogous to that for H2 ? The proof of Theorem 16. As in section 16. we will need to verify that Gtmp satis es the following assumptions for the OE problem: .2 B1 B1 X1 : r u w u v y Gtmp  K  K Recall from Lemma 16. in the presence of this worstcase input. Furthermore. Z1 L1 is actually the optimal lter gain for estimating F1 x.2 B1 X1 x: ^ ^ These equations have the structure of an observerbased compensator. .428 H1 CONTROL: SIMPLE CASE u = F1 x ^ wworst = .8.3.4 reveals that there is a very wellde ned separation interpretation of these formulae and that wworst := . While Gtmp has the form required for the OE problem. a worstcase input for the Full Information problem.
e.6.2 X1 Y1 ). .16. T := tmp 1 0 I T . .4 (Su ciency) Assume the conditions (i) through (iii) in the and (X1 Y1 ) < 2 implies that Jtmp 2 dom(Ric) and Ytmp := Ric(Jtmp ) = Y1 (I ..2 B1 B1 X1 .Atmp : 0. The lemma follows from the dual to Lemma 16. the existence of an admissible controller for Gtmp immediately implies that assumption (ii) holds.8. The following lemma gives conditions for assumption (ii) to hold. Separation Theory (i) (Atmp B1 ) is stabilizable and Atmp + B2 F1 is stable (ii) (C2 Atmp ) is detectable 429 B1 D = 0 . i.13 with Q = 0. . C2 C2 # . (Jtmp ) = T X.13 the OE problem is solvable. Using the Riccati equation for X1 . (iv) D21 21 I Assumption (iv) and that (Atmp B1 ) is stabilizable follow immediately from the corresponding assumptions for Theorem 16. " " # " # . 2 theorem statement hold. and by Theorem 16. Ytmp C2 C2 + B2 F1 Ytmp C2 F1 0 # but this is precisely Ksub de ned in Theorem 16.2 X1 Y1 X.1 = Z1 Y1 0. (J1 ) = T Im Y1 Y1 " # Proof of Theorem 16. Then by Lemma 16. We conclude that Ksub stabilizes Gtmp and that kTvr k1 < .4. The stability of Atmp + B2 F1 follows from the de nition of H1 2 dom(Ric). .15 the OE assumptions hold for Gtmp . which gives that (Atmp . So # " I = Im I .8. From Theorem 16. Thus by Lemma 16.2 F1 F1 . one solution is " A + . one can easily verify that " # I .B1 B1 detectable. Note that the OE Hamiltonian matrix for Gtmp is Jtmp := Atmp .4.2X1 provides a similarity transformation between J and J . Ksub stabilizes G and that kTzw k1 < . Of course. then (C2 Atmp ) is Lemma 16. Proof.15 If Jtmp 2 dom(Ric) and Ytmp := Ric(Jtmp) Ytmp C2 C2 ) is stable.1Jtmp T = J1 .
.14.430 H1 CONTROL: SIMPLE CASE (Necessity) Let K be an admissible controller for which kTzw k1 < . but similar interpretations could be made for the parameterization of all suboptimal controllers (see the proofs of Theorems 16. from Theorem 16. is equivalent to state estimation. This implies that the OE assumptions hold for Gtmp and that the OE problem is solvable.13 applied to Gtmp . From Lemma 16. and Y1 := Ric(J1) 0. This is because there is no \worstcase" disturbance for the H2 Full Information problem and because the problem of estimating any output. the resulting equations are much simpler. Both terms are consequences of estimating the optimal Full Information (i. By Lemma 16. . Therefore.4 and 16. " # (I . Y1 = U " Y11 0 U # UX1 U = X11 X12 : X21 X22 Then by the same argument as in the Y1 nonsingular case.8. " # 2 We now see exactly why the term involving wworst appears and why the \observer" ^ gain is Z1 L1.2 Y1 X1 ). While an analogous output estimation problem arises in the H2 output feedback problem. X1 := Ric(H1 ) 0. We shall now show that Ytmp 0 implies that (X1 Y1 ) < 2 . So (Y1 1 1 1 1 Y1 is singular: there is a unitary matrix U such that 0 0 with Y11 > 0.e.2Y11 X11 ).2 Y1=2 X1 Y1=2 > 0.1 Y1 0. Note that the same statement holds for the H2 optimal controller. 1=2 X Y 1=2 ) < 2 or (X Y ) < 2 . We shall consider two cases: 1 1 Y1 is nonsingular: in this case Ytmp 0 implies that I . It will be stated in terms of the central controller. We now present a separation interpretation for H1 suboptimal controllers. Let UX1 U be partitioned accordingly. K is admissible for Gtmp and kTvr k1 < . we get that Ytmp = (I . state feedback) control gain. The H1 output feedback controller is the output estimator of the full information control law in the presence of the \worstcase" disturbance wworst .5). .. including the optimal state feedback. we have that Jtmp 2 dom(Ric) and Ytmp = Ric(Jtmp) 0.1 Y11 0 Ytmp = U U 0 0 I implies that 2 > (X11 Y11 ) (= (X1 Y1 )). Using the same similarity transformation formula as in the su ciency part. except that wworst = 0. J1 2 dom(Ric). . H1 2 dom(Ric).
we will indicate why the descriptor version of the controller formulae from Section 16.12) u = F1 x ^ (16. assumption (i) to controllable and observable and considering the Hamiltonians for X11 . L1 y ^ ^ (16. especially as approaches the in mal achievable norm. Since Theorem 16.. X1 ! X2 . without proof. Theorem 16.e. Most of these results can be obtained relatively easily using the machinery that is developed in the previous sections. the behavior of the H1 suboptimal solution as varies. It is possible but far less likely that conditions (i) or (ii) would fail before (iii).8. In fact.1 . Apply Theorem 16. this connection with H2 actually hints at deeper interpretations. The formulae in Theorem 16. then it is relatively straightforward to show that the descriptor formulae for = opt are optimal. While it could be argued that Ksub is a natural choice. opt X1 Y1 ) is not invertible. 2 2 If 2 1 > opt .16. Although we have not focused on the problem of H1 optimal controllers. To see this. opt is the in mum over all such that conditions (i)(iii) are satis ed. it can be computed as closely as desired by a search technique. anyone of the three conditions in Theorem 16. and Y1 As ! 1. as is (X1 Y1 ). but. consider (i) and let 1 be the largest . At = opt .3 Proof of Theorem 16. Optimality and Limiting Behavior 431 16. opt Y1 X1 )x = As x . the optimal controller is given by . etc.9 Optimality and Limiting Behavior In this section. In addition to describing the qualitative behavior of suboptimal solutions as varies. the assumptions in this book make them relatively easy to obtain in most cases. H1 ! H2 .8. 2 16. just as for the H1 norm calculation. 2kwk2 .8.2 . This fact is the result of the particular choice for the central controller (Q = 0) that was made here.9.13) . See the example below.2 term (I . then X1 ( 1 ) X1 ( 2 ) and Y1 ( 1 ) Y1 ( 2 ). denoted by opt . If only condition (iii) fails.4 does not give an explicit formula for opt .2 where As := A + B2 F1 + L1 C2 + opt Y1 A X1 + opt B1 B1 X1 + opt Y1 C1 C1 .2 _ (I .4 can fail.4 are not wellde ned in the optimal case because the . the set of all admissible controllers for G such that kTzw k1 < equals the set of all admissible controllers for Gtmp such that kTvr k1 < . i..2 . Ksub is the minimum entropy solution (see next section) as well as the minimax controller for kz k2 .1 can usually provide formulae for the optimal controller when = opt . we will discuss.5 From Lemma 16.13. Thus X1 and Y1 are decreasing functions of . The reader interested in lling in the details is encouraged to begin by strengthening . and Ksub ! K2 .4 gives necessary and su cient conditions for the existence of an admissible controller such that kTzw k1 < .
stabilizing or otherwise. Complementarity failing at = 1 means (X1 ) ! 1 as ! 1 .c2 H1 = J1 = : . If complementarity fails at = 1 . 1 (a2 + c2 ) 2 . the formulae in Theorem 16. so condition (iii) would fail at even larger values of .a The eigenvalues of H1 and J1 are given. if stability breaks down rst.4 as well as their descriptor versions are optimal at = opt . unless the eigenvectors associated with (X1 ) as ! 1 are in the null space of Y1 . In other words. Example 16. (H1 ) = Im " p(a2 +b2 ) 2 2 . and no controller whatsoever exists which makes kTzw k1 < 1 . by ( p ) ( p ) (a2 + b2 ) 2 .1.1 . then obtaining formulae for the optimal controllers is a more subtle problem. Thus condition (iii) is the most likely of all to fail rst. See part 1) of the example below.a # 1 . the controller u = . This is illustrated in the example below for the output feedback. If the complementarity condition fails rst. (but (iii) does not fail).a . respectively.1 . but X1 will be inde nite.432 H1 CONTROL: SIMPLE CASE for which H1 fails to be in dom(Ric) because the H1 matrix fails to have either the stability property or the complementarity property. 1 : 2 2 (H ) = (J ) = 1 1 If 2> a2 + b2 2 1 1 a2 + c2 .B2 X1 x is stabilizing and makes kTzw k1 = 1 . Then all assumptions for the output feedback problem are satis ed and " " 2 2 # 2 2 # 2 2 a 1. The same remarks will apply to (ii) by duality. (H1 ) and X. H1 may again be in dom(Ric). then H1 62 dom(Ric) but Ric can be extended to obtain X1 and the controller u = . If the stability property fails at = 1 . For such . The stability property will also not hold for any 1 .1 Let an interconnected dynamical system realization be given by 2 G(s) = 6 " 6 6 6 6 4 a c2 1 0 h # h 1 0 0 0 1 i b2 # " 0 1 0 3 7 7 7 7 7 5 i with jc2 j jb2 j > 0. then (X1 ) ! 1 as ! 1 . then the in mum over stabilizing controllers equals the in mum over all controllers. In view of this.B2 X1 x would make kTzw k1 < but would not be stabilizing. If condition (i) or (ii) fails rst because the stability property fails. we would typically expect that complementarity would fail rst. then X. For < 1 . (J1 ) exist and 2 X. See part 2) of the example below.b2 a 1.
Optimality and Limiting Behavior " p 2 2 2 (a +c2 ) .a X. Fur2 2 2 1 thermore. if 1 a2 + b2 and 2 2 6= b2 . (J1 ) = Im 1 We shall consider two cases: 433 # : 1) a > 0: In this case. it can be shown that kTzw k < for all 2 > a2 + b2 and 2 6= b12 . 2 2 and Tzw is stable for all It is clear that the optimal H1 norm is b12 but is not achievable. 1 = 2 ( < 0 (stable) if 2 > b12 2 1 > 0 (unstable) if a2 +b2 < 2 2< 1: 2 b2 { Suppose full information (or states) are available for feedback and let u = F1 x: Then the closedloop transfer matrix is given by 2 Tzw = " A + B2 F1 B1 = 6 6 6 4 C1 + D12 F1 0 2> # .12 b2 3 5 1 0 0 i 3 7 7 7 5 . the complementary property of dom(Ric) will fail before the stability property fails since q (a2 + b2 ) 2 . a = 0 2 when 2= b2 2 1 1 > a2 + b2 . 12 ) . 1 .9. then H1 2 dom(Ric) and 2 ( 1 X1 = p (a2 + b2 ) 2 .1. a+b2 2 4 2 p(a2 +b2 ) 1 2 2 .a 1 1 1 2 b2 and is not stable for a2 + b2 < < b2 .B2 X1 then 2 2 + b2 2 A+B2 F1 = . 1 . 2 2> Nevertheless. a + b2 b2(a2 .1 h 2 2 . 2 . a 2 p = > 0 if 2 > b12 2 1 < 0 if a2 +b2 < 2 2 < 12 : b2 Let F1 = .1.16. p(a2 +b2b)2 2 .
in this case. 1 . Since ( (a2 + b2 ) 2 . complementary property is always satis ed. H1 2 dom(Ric). Never2 1 2= theless. furthermore. X1 B2 B2 X1 = 0 . B2 B2 X1 = 0. 1 . J1 2 dom(Ric). for matrices H1 with imaginary axis eigenvalues as " 1 X. and 2 X1 = p Y1 = p (a2 + b2 ) 2 . (H1 ) = Im . 2 2 1 (a2 + c2 ) 2 . a 2 >0 > 0: (a2 + c2 ) 2 . then H1 2 dom(Ric). 1 . . a 2 >0 >0 a2 + b2 . and need to check condition (iii). H1 62 dom(Ric) since stability property fails.4 are satis ed. Note that if 2 > b12 . a is a solution to the Riccati equation . if 0 a2 + b2 . a 2 a2 + b2 . 1 . 1 . a )( (a2 + c2 ) 2 . a 2 Hence conditions (i) and (ii) in Theorem 16. we can extend the dom(Ric) to include those 2 However. A X1 + X1 A + C1 C1 + 0 2X1 B1 B1 X1 .434 H1 CONTROL: SIMPLE CASE { Suppose the states are not available then output feedback must be considered. a ) 2 2 it is clear that (X1 Y1 ) ! 1 when 2 ! b12 . J1 2 dom(Ric). and A + 0 2B1 B1 X1 . and (X1 Y1 ) = p 2 p X1 = p Y1 = p for 2> (a2 + b2) 2 . 2) a < 0: In this case.a 1 # 1 such that X1 = . 1 . and. So condition (iii) will fail 2 before condition (i) or (ii) fails.
and jb2j = jc2j. In fact.9.16. 1 . uopt = . 1 = 0:7321 and uopt = . So if states are available for feedback and u = F1 x. then A + B2 F1 = a + ba < 0. then it can be shown that (X1 Y1 ) = is satis ed if and only if 2 22 > a +b2 b2 + a > p 21 2 : a + b2 2 p ! ( p 2 (a2 + b2 ) 2 . Optimality and Limiting Behavior { For = 0 2 435 2 and F1 = . Further. a )2 2 < 2 So condition (iii) of Theorem 16. Then opt = 3 .12) and (16. In both a > 0 and a < 0 cases.13) is a constant. let a = .B2 X1 . a opt 2 opt y: For instance.1 and b2 = 1 = c2 .1:7321 1 .4 will fail before either (i) or (ii) fails. p 3 .0:7321 Tzw = 6 1 0 0 7: 5 4 . Hence the optimum is achieved. we have 2 Tzw = 6 4 6 " a + ba2 b2 a 2 # h 1 1 0 0 i 3 7 7 5 2 RH1 = . q 2 (a2 + b2 ) opt .0:7321 y. 3 2 .0:7321 It is easy to check that kTzw k1 = 0:7321. 1 .0:7321 0 . and kTzw k1 = p 21 a + b2 0 { If states are not available. the optimal for the output feedback is given by opt = p a2 + 2b2 + a 2 b2 2 if jb2 j = jc2 j and the optimal controller given by the descriptor formula in equations (16. the output feedback is considered.
.436 H1 CONTROL: SIMPLE CASE 16.1 . who also derives a separation result and a certainty equivalence principle (see also Whittle. Section 16. 1986]). his separation result has a clear interpretation for the present in nite horizon.2 2Z 1 . where i (T (j!)) is the ith singular value of T (j!).1 Therefore.10. Then the entropy of T (s) is de ned by .2 i2 (T (j!)) d! 0.4 is actually the controller that satis es the norm condition kTzw k1 < and minimizes the following entropy: . discrete time. ln det I . .2 Relations with Separation in Risk Sensitive Control Although Whittle. 16. The connection with the Qparameterization approach will be considered in the next chapter for the general case. stochastic control problem.2 Tzw (j!)Tzw (j!) d!: . 2 It is easy to see that 2Z 1 Let T be a transfer matrix with kT k1 < .10.10 Controller Interpretations This section considers some additional connections with the minimum entropy solution and the work of Whittle and will be of interest primarily to readers already familiar with them. 1981].4 in the spirit of Whittle (1981). 16.10.1 Minimum Entropy Controller I (T ) = .1 i Thus the entropy I (T ) is in fact a performance index measuring the tradeo between the H1 optimality ( ! kT k1 ) and the H2 optimality ( ! 1). . It has been shown in Glover and Mustafa 1989] that the central controller given in Theorem 16.2T (j!)T (j!) d!: I (T ) = . 2 and I (T ) show that 2Z 1X . It has been shown in Glover and Doyle 1988] that the central controller corresponds exactly to the steady state version of the optimal risk sensitive controller derived by Whittle. It is also easy to Z 1 X 2 2 lim I (T ) = 21 i (T (j! )) d! = kT k2 : !1 .I (T )). 1981] treats a nite horizon. the central controller is also called the minimum entropy controller (maxi~ mum entropy controller if the entropy is de ned as I (T ) = . ln det I .2 gives another separation interpretation of the central H1 controller of Theorem 16. .1 i ln 1 .
(C1 x)k2 + kP. and is produced by the worst case disturbance.16. and of the present state. x. .10. S. D21 w = D21 B1 Y1 1 (x(t) . with e := x . let x be given by the stable di erential equation _ x = Ax + B2 u + L1 (C2 x . (x) := sup(kP. y) + Y1 C1 C1 x with x(. is a function of the past inputs and observations. z k2 . that is consistent with the given data: S. S. which is a ltering problem with known control input. S+ . wk2 ): 2 2 In order to evaluate S. where u is not constrained to be a function of the measurements (FI case). (x). w . S. and this is easily shown to be achieved at the current state of x = Z1 x(0): ^ The de nitions of X1 and Y1 can be used to show that the state equations for the central controller can be rewritten with the state x := Z1 x and with x as de ned ^ above. and the control law acts as if this were the perfectly observed ^ state. x(t)) for t < 0 and that this gives. uk2: 2 2 2 The worst case disturbance will now reach the value of x to maximize the total stress. and future stress. 2 kP. we obtain S+ (x) = x X1 x: The past stress. In particular. 2kP. The past stress is then calculated by a completion of the square and in terms of a lter output. x. (y . which is a control problem with an unconstrained input. u(t) y(t) for . and the past stress.1) = 0: Then it can be shown that the worst case w is given by ? ? . as functions of the current state. 2e(0) Y1 1 e(0) . The central controller then combines these evaluations to give a worst case estimate. D21 w ? ? and D21 w . and evaluate the past stress. First de ne the future stress as S+ (x) := sup inf (kP+ z k2 .3. 2kP+ wk2 ) 2 2 u then by the completion of the squares and by the saddle point argument of Section 16. we see that w can be divided into two components. x. Controller Interpretations 437 continuous time. C2 x)k2 + kP. with x dependent only on D21 w (since B1 D21 = 0) and D21 w = y . We will consider the system at time. x. S. t = 0. C2 x. w. as given below and it is an interesting exercise to compare the two separation statements. (x) = . . (x) + S+ (x) . The control signal is then u = F1 x = F1 Z1 x: ^ The separation is between the evaluation of future stress. deterministic control problem. This discussion will be entirely in the timedomain.1 < t < 0.
16 There exists an admissible controller such that kTzw k1 i the following three conditions hold: (i) there exists a full column rank matrix " X11 2 R2n n X12 # # such that and H1 X11 = X11 TX Re i (TX ) 0 8i X12 X12 X11 X12 = X12 X11 " " # " (ii) there exists a full column rank matrix Y11 2 R2n n Y12 # # such that and (iii) J1 Y11 = Y11 TY Re i (TY ) 0 8i Y12 Y12 Y11 Y12 = Y12 Y11 " " # " X12 X11 . when these conditions hold.B2 X12 AK := EK TX . BK C2 X11 = TY EK + Y11 B2 CK : . we shall give in the following without proof the conditions under which an optimal controller exists and an explicit formula for an optimal controller.2Y12 X12 BK := Y12 C2 CK := . AK )+ BK where EK := Y11 X11 .11 An Optimal Controller To o er a general idea about the appearance of an optimal controller.1X12 Y12 # Y12 Y11 0: Moreover. . one such controller is Kopt (s) := CK (sEK .438 H1 CONTROL: SIMPLE CASE 16. Theorem 16.1 Y12 X12 .
13). The complete characterization of optimal controllers for the general setup can be found in Glover. Doyle. Notes and References 439 Remark 16. Limebeer.16. and Safonov 1991].1 1 X12 X11 and Y1 = Y12 Y11 . Francis 1989]. The minimum entropy controller is studied in detail in Glover and Mustafa 1989] and Mustafa and Glover 1990].4 It is simple.12 Notes and References This chapter is based on Doyle. then condition (iii) in the above theorem is equivalent to X1 0. the conditions for the existence of an optimal controller can be obtained from \taking the limit" of the corresponding conditions in Theorem 16.4. and (Y1 X1 ) 2 . . The risk sensitivity problem is treated in Whittle 1981. the controller given above is reduced to the descriptor form given in equations (16.12. ~ 16. 1986]. Glover. Y1 0.12) and (16. So in this case. and Zhou 1992].to show that if X11 and Y11 are nonsingular and if X1 = . Kasenally. Moreover. The connections between the risk sensitivity controller and the central H1 controller are explored in Doyle and Glover 1988]. Khargonekar.
440 H1 CONTROL: SIMPLE CASE .
We will also indicate how the assumptions in the general case can be relaxed further to accommodate other more complicated problems.6 considers an alternative approach to the standard H2 and H1 problems using Youla controller parameterizations. More speci cally.5 considers how the general H1 solution can be used to solve the H1 ltering problem.9.8 and 17.4 considers the integral control in the H2 and H1 theory and Section 17. Section 17. Section 17.3. and Section 17.H1 Control: General Case In this chapter we will consider again the standard H1 control problem but with some assumptions in the last chapter relaxed. Section 17. However.1 presents the solutions to the general H1 problem. the detailed proof for the general case will not be given only the formulas are presented.7 gives an 2 2 HankelToeplitz operator interpretations of the H1 solutions presented here and in the last chapter. Since the proof techniques in the last chapter can be applied to this general case except with some more involved algebra. The problems associated with relaxing the assumptions for the general standard problems and techniques for dealing with them will be considered in Section 17.2 discusses the techniques to transform a general problem to a standard problem which satis es the assumptions in the last chapter. some procedures to carry out the proof will be outlined together with some alternative approaches to solve the standard H1 problem and some interpretations of the solutions. the general state feedback H1 control problem and its relations with full information control and di erential game problems are discussed in section 17. 441 17 . Finally. Section 17.
and that the sensor noise weighting is normalized and nonsingular. that the exogenous signal w includes both plant disturbance and sensor noise. j!I B1 C2 D21 has full column rank for all ! has full row rank for all !.1 General H1 Solutions Consider the system described by the block diagram z G w y  K u where. w(t) 2 Rm1 . Relaxation of (A2) leads to singular control problems see Stroorvogel 1990]. y(t) 2 Rp2 . a normalizing procedure is given in the next section so that an equivalent new system will satisfy this assumption. and stabilizing. normalized penalty on the control u. the results presented here are still true for the complex case with some obvious modi cations. # (A4) Assumption (A1) is necessary for the existence of stabilizing controllers. as usual. For those problems that have D12 full column rank and D21 full row rank but do not satisfy assumption (A2). . u(t) 2 Rm2 . j!I B2 C1 D12 A . We will again only be interested in characterizing all suboptimal H1 controllers. The assumptions in (A2) mean that the penalty on z = C1 x + D12 u includes a nonsingular. and the state x(t) 2 Rn . Although we are taking everything to be real. The realization of the transfer matrix G is taken to be of the form A B1 B2 G(s) = 6 C1 D11 D12 4 C2 D21 0 2 3 7 5 A B = C D " # which is compatible with the dimensions z (t) 2 Rp1 . proper. G and K are assumed to be real rational and proper with K constrained to provide internal stability. The controller is said to be admissible if it is realrational.442 H1 CONTROL: GENERAL CASE 17. The following assumptions are made: (A1) (A B2 ) is stabilizable and (C2 A) is detectable (A2) D12 = " " I 0 # and D21 = 0 I # h i (A3) " A .
3.1. It is tempting to suggest that (A3) and (A4) can be dropped. General H1 Solutions 443 Assumptions (A3) and (A4) are made for a technical reason: together with (A1) they guarantee that the two Hamiltonian matrices in the corresponding H2 problem belong to dom(Ric). This issue will be discussed further later.1 In the above matrix partitioning. ~ Theorem 17.R. For the time being.1 2 3 Partition D. A further discussion of the assumptions and their possible relaxation will be discussed in Section 17. kTzw k1 < ) if and only if F111 F121 F21 0 # 6 L 7 F = 6 111 D1111 D1112 0 7 : 7 6 6 0 D L I 7 5 4 L121 D1121 D1122 0 I 0 L21 . " " 2 Im 2 Ip 0 1 D11 ~ where D 1 := R := D 1 D 1 . R. in some sense.A . but they are.1 D C + B X ] 1 1 1 F21 i ~ L11 L21 := .B1D 1 1 0 0 # 0 0 # where D1 := D11 D12 ] " # X1 := Ric(H1) F := L := " h # Y1 := Ric(J1 ) F11 := . B1 D 1 + Y1 C ]R. some matrices may not exist depending on whether D12 or D21 is square. necessary for the methods presented in the last chapter to be applicable.A .C1 D1 " # " # h i A 0 C ~ J1 := . The main result is now stated in terms of the solutions of the X1 and Y1 Riccati equations together with the \state feedback" and \output injection" matrices F and L. R := D1 D1 . R.17. we shall assume all matrices in the partition exist.C1 C1 .1 D1 C1 B .1 Suppose G satis es the assumptions (A1){(A4).e. as we have seen in Chapter 14.1 D 1 B1 C . and L11 are as follows: " Remark 17. D21 0 " # " # h i A 0 B H1 := . (a) There exists an admissible controller K (s) such that jjF` (G K )jj1 < (i.B1 B1 . F11 .
^ C1 = F21 + D11 D211 C2 ^ ^ ^. .) Some Special Cases: Case 1: D12 = I In this case 1.1 : (Note that if D11 = 0 then the formulae are considerably simpli ed. 2.g. (i) becomes > (D1121 ).Z1L21 + B2 D121 D11 ^ ^ ^. then all rational internally stabilizing controllers K (s) satisfying jjF` (G K )jj1 < are given by ^ ^ ^ A B1 B2 7 6 ^ ^ ^ 5 M1 = 6 C1 D11 D12 7 4 ^ ^ C2 D21 0 ^ D11 = . D1111 D1111 ). ^ A = A + BF + B1 D211 C2 Z1 = (I .D1121 D1111 ( 2 I . (b) Given that the conditions of part (a) are satis ed.2 D1121 D1121 ^ ^ D21 D21 = I: . . in part (a). D1111 D1111 ).2 Y1 X1 ). D1111 D1111 ).D1122 ^ ^ D12 D12 = I .D21 (C2 + F121 ) ^ ^ ^. D1121 ( 2 I . ^ B1 = . Cholesky factors) satisfying ^ D ^ ^ D12 D12 = I . D1122 ^ 12 2 Rm2 m2 and D21 2 Rp2 p2 are any matrices (e.1 D1112 . in part (b) ^ D11 = .444 H1 CONTROL: GENERAL CASE (i) > max( D1111 D1112 ] D1111 D1121 ]) (ii) H1 2 dom(Ric) with X1 = Ric(H1 ) 0 (iii) J1 2 dom(Ric) with Y1 = Ric(J1 ) 0 (iv) (X1 Y1 ) < 2 . D1112 ( 2 I .1 D1121 ^ ^ D21 D21 = I .1 D1112 and where K = F`(M1 Q) for arbitrary Q 2 RH1 2 such that kQk1 < 3 where ^ ^ B2 = Z1 (B2 + L121 )D12 ^ ^ C2 = .
(i) drops out. .Kp up The plant P has the following state space realization with Dp12 full column rank and Dp21 full row rank: 2 3 jjF` (P Kp )jj1 < . 2. yp (t) 2 Rp2 . in part (a).D1122 ^ ^ D12 D12 = I ^ ^ D21 D21 = I .D1122 ^ ^ D12 D12 = I ^ ^ D21 D21 = I: 17. in part (b) ^ D11 = . wp (t) 2 Rm1 . Loop Shifting 445 Case 2: D21 = I In this case 1. To solve this problem we rst transfer it to the standard one The objective is to nd all rational proper controllers Kp(s) which stabilize P and Ap Bp1 Bp2 P (s) = Cp1 Dp11 Dp12 Cp2 Dp21 Dp22 6 4 7 5 treated in the last section.2 Loop Shifting Let a given problem have the following diagram where zp (t) 2 Rp1 .2.2 D1112 D1112 : Case 3: D12 = I & D21 = I In this case 1. Note that the following procedure can also be applied to the H2 problem (except the procedure for the case D11 6= 0). and up (t) 2 Rm2 : zp P wp yp . in part (b) ^ D11 = . .17. in part (a). (i) becomes > (D1112 ). 2.
Then the controller for D22 6= 0 is K (I + D22 K ). ~. 0 Rp 1 0 Rp . .1 . Hence there is no loss of generality in assuming D22 = 0. ~ ~. ~. ~ Up Cp1 Up Dp11 Up Up Dp12 Rp 1 . ~. Rp 1 Cp2 Rp 1Dp21 Up Rp 1Dp22 Rp 1 # " # 3 7 7 5 A B1 B2 C1 D11 D12 C2 D21 D22 " 3 7 5 A B : = C D h i " # D12 = I z 0 # D21 = 0 I G w u and the new system is shown below: y  K ~ Furthermore. Now let ~ ~ zp = Up z wp = Up w yp = Rp y up = Rp u and ~ K (s) = Rp Kp (s)Rp G(s) = = =: Then " 2 6 6 4 2 6 4 Up ~ 0 P (s) Up 0 1 . = 2 or 1 Remove the Assumption D22 = 0 Suppose K (s) is a controller for G with D22 set to zero. jjF` (P Kp )jj = jjUp F` (G K )Up jj = kF`(G K )k for ~ since Up and Up are unitary. ~ Bp1 Up Bp2 Rp 1 Ap .446 H1 CONTROL: GENERAL CASE Normalize D12 and D21 Perform singular value decompositions to obtain the matrix factorizations Dp12 = Up " I 0 # Rp ~ ~ Dp21 = Rp 0 I Up h i ~ such that Up and Up are square and unitary.
17. In the following. and the matrix N is called a unitary dilation of D (of course. In fact. To begin with.1 D1111 D1112 and de ne " # D1112 ~ 11 := D1111 D D1121 D1122 + D1 . furthermore. N N = I .2. (Suppose we have normalized to 1). let D1 = . This transformation is called loopshifting. there are many other ways to dilate a matrix to a unitary matrix). Before we go into the detailed description of the transformation. let us rst consider a simple unitary matrix dilation problem.1 can be shown by rst transforming the general problem to the standard problem considered in the last chapter using Theorem 16.2 Suppose D is a constant matrix such that kDk < 1 then (I . Theorem 17. we will show how to construct a new system ~ transfer matrix M (s) and a new controller K such that the D11 matrix for M (s) is zero. Lemma 17. ~ and. Loop Shifting 447 Remove the Assumption D11 = 0 We can even assume that D11 = 0.D (I . we shall assume that the system G has the following realization: 2 6 6 6 6 4 N= " . i. This result can be easily veri ed.D1122 . Consider again the feedback system shown at the beginning of this chapter without loss of generality. D1111 D1111 ). kF` (G K )k1 < 1 if and only if F`(M K ) < 1. DD )1=2 # A C2 G = C1 with " B1 D1111 D1112 D1121 D1122 h i 0 I I 0 # h # " B2 # 0 0 i 3 7 7 7 7 5 I D12 = " D21 = 0 I : Suppose there is a stabilizing controller K such that kF` (G K )k1 < 1..e. 1 note that kF` (G K )(1)k < 1 by the assumption kF`(G K )(1)k = D1111 D1121 Let ( D1112 : D1122 + K (1) ) D1 2 X : D1111 D1112 <1 : D1121 D1122 + X For example.2. D D)1=2 D is a unitary matrix. D1121 (I .
Furthermore. # " A + B (I . ~ ~ ~ ~ . D D). Corollary 17. by Theorem 16. ~ M (s) = 6 ~. D11 D11 )1=2 D11 and then N N = I . Let to get where ~ G = 2 6 6 6 6 4 2 H1 CONTROL: GENERAL CASE ~ K (s) = K (s) + D1 ~ ~ ~ F`(G K ) = F` (G K + D1 ) = F` (G K ) A + B2 D1 C2 C1 + D12 D1 C2 ~ A ~ C1 ~ C2 " =: Let 6 6 4 C2 ~ B1 ~ D11 ~ D21 " ~ B2 7 ~ 5 D12 7 : 0 ~ . D11 D11 and R1 = I .448 ~ then D11 < 1.4 Let G(s) = A B 2 RH1 .D11 # 3 B1 + B2 D1 D21 D1111 D1112 D1121h D1122i+ D1 0 I # " B2 # I 0 0 3 7 7 7 7 5 ~ ~ (I . D11 D11 . Then kG(s)k1 < 1 if and only if C D kDk < 1. In summary. ~ ~ . ~ ~ ~ .1=2 C 0 " # 1 . ~ ~ A + B1 R1 1D11 C1 B1 R1 1=2 B2 + B1 R1 1D11 D12 7 6 7 ~.1 D C B (I . we have that K stabilizes G and that kF` (G K )k1 < 1 if and only if F`(G K ) internally stabilizes N and kF`(N F` (G K ))k1 < 1: Note that ~ F`(N F` (G K )) = F`(M K ) with 3 2 ~ . Lemma 17.3 There is a controller K that internally stabilizes G and kF`(G K )k1 < 1 ~ ~ if and only if there is a K that stabilizes M and F`(M K ) < 1.2. D D).1=2 2 RH M (s) = 1 (I . ~ ~ ~ ~ . DD ). D11 D11 )1=2 N= ~ ~ ~ (I . ~ R1 1=2 C1 0 R1 1=2 D12 5 4 ~ . we have the following lemma. C2 + D21 R1 1 D11 C1 D21 R1 1=2 D21 R1 1 D11 D12 ~ ~ ~ ~ ~ where R1 = I .
449 17. see Lemma 16. then Tzw = s.17. Relaxing A1 If assumption A1 is violated. although the suboptimal case is not obviously so. H1 optimal control is a silly problem. but = 0 is not admissible since it is not stabilizing.3. we indicate how the results of section 17. If the controller u = . then it is obvious that no admissible controllers exist.3 Relaxing Assumptions In this section.1. in particular. Relaxing A3 and A4 Suppose that 2 G=6 4 0 0 1 0 0 1 1 1 0 3 7 5 which violates both A3 and A4 and corresponds to the robust stabilization of an integrator. This illustrates that assumptions A3 and A4 are necessary for the techniques used in the last chapter to be directly applicable. . then the formulae presented above will yield u = 0 for both the optimal controller and the suboptimal controller with = 0. If one simply drops the requirement that controllers be admissible and removes assumptions A3 and A4. Relaxing Assumptions and kM (s)k1 < 1. x where > 0 is used. Of course. An alternative is to develop a theory which maintains the same notion of admissibility but relaxes A3 and A4. that the unstabilizable and/or undetectable modes on the j! axis must show up as poles in the closedloop system. Suppose A1 is relaxed to allow unstabilizable and/or undetectable modes on the j! axis and internal stability is also relaxed to also allow closedloop j! axis poles. but A2A4 is still satis ed. This may be thought of as a case where the H1 optimum is not achieved on the set of admissible controllers. for this system. The easiest way to do this would be to pursue the suboptimal case introducing perturbations so that A3 and A4 are satis ed. s with kTzw k1 = : + Hence the norm can be made arbitrarily small as ! 0. It can be easily shown that under these conditions the closedloop H1 norm cannot be made nite and.1 can be extended to more general cases.
The formulae given in this chapter will often yield controllers compatible with these assumptions. but the behavior as ! 0 could be problematic. In particular. It is interesting to note that the H2 and H1 design frameworks do not in general produce integral control. Roughly speaking. such problems can be reformulated so that the integrator occurs inside the loop (essentially using the internal model principle) and is hence detectable and stabilizable. cases when A has modes at s = 0 which are unstabilizable through B2 and/or undetectable through C2 arise when an integrator is included in a weight on a disturbance input or an error term. and Silverman 1986].g. Kitapci.450 H1 CONTROL: GENERAL CASE Violating A1 and either or both of A3 and A4 Sensible control problems can be posed which violate A1 and either or both of A3 and A4. we require K contain an integral. For example. Clearly. as in the above example. In many applications. We shall assume that the frequency contents of the disturbance w are e ectively modeled by the weighting Wd 2 RH1 and the constraints on control signal are limited by an appropriate choice of Wu 2 RH1 . We consider a feedback system shown in Figure 17. 17. A3. the structure algorithm of Silverman 1969] could be used to make the terms D12 and D21 full rank by the introduction of suitable di erentiators in the controller. or the closedloop H1 norm cannot be made nite. In order to let the output y track the reference signal r. and Hautus and Silverman 1983]). this would produce sensible answers for sensible problems. In these cases. Such singular ltering and control problems have been wellstudied in H2 theory and many of the same techniques go over to the H1 case (e. Relaxing A2 In the cases that either D12 is not full column rank or that D21 is not full row rank.1.4 H2 and H1 Integral Control . 1990]. We will show this process in the next section. Willems. to have nite H1 norms. A third alternative is to again introduce perturbations so that A1. improper controllers can give a bounded H1 norm for Tzw . these closedloop modes could not appear as poles in Tzw . either A3 or A4 are also violated. Willems 1981]. although the controllers will not be admissible by our de nition. In this section we show how to introduce integral control into the H2 and H1 design framework through a simple disturbance rejection problem. and A4 are satis ed. A complete solution to the singular problem can be found in Stroorvogel. An alternative approach to such problems which could potentially avoid the problem reformulation would be to pursue the techniques in the last chapter but to relax internal stability to the requirement that all closedloop modes be in the closed left half plane. The user would then have to decide whether closedloop poles on the imaginary axis were due to weights and hence acceptable or due to the problem being poorly posed.
1 Wd w: Now if the resulting controller K stabilizes the plant P and makes the norm (2norm or 1norm) between w and z1 nite. ~ Let We Wu Wd M . Figure 17. Now the problem can be reformulated as in Figure 17. We (s) is stable and minimum phase.2.) The problem with this approach is that the H2 (or H1 ) control theory presented in this chapter and in the previous chapters can not be applied to this problem formulation directly because the pole s = 0 of We becomes an uncontrollable pole of the feedback system and the very rst assumption for the H2 (or H1 ) theory is violated.K (s)M (s) ^ where there is no unstable pole/zero cancelation in forming the product K (s)M (s).1: A Simple Disturbance Rejection Problem There are several ways to achieve the integral design. Then the transfer function between w and z1 is given by z1 = We (I + PK ).6 K 6 Wd ? w u  P ? . Then without loss of generality. Now suppose there exists a controller K (s) which contains the same imaginary axis poles that achieves the performance. containing all the imaginary axis poles of We . (This follows from the wellknown internal model principle. (In general.3. and P have the following stabilizable and detectable state space . Suppose We can be factorized as follows ~ We = We (s)M (s) where M (s) is proper.We z1 Figure 17.2 can be put in the general LFT framework as in Figure 17. However..e. One approach is to introduce an integral in the performance weight We . K can be factorized as ^ K (s) = .17. H2 and H1 Integral Control 451 i.1 (s) 2 ~ RH1 .) z 62 Wu r g .g y. K (s) has a pole at s = 0. then K must have a pole at s = 0 which is the zero of the sensitivity function (Assuming Wd has no zeros at s = 0).4. K is required to have poles on the imaginary axis. and M . the obstacles can be overcome by appropriately reformulating the problem.
452 H1 CONTROL: GENERAL CASE z 62 Wu Wd ? w y1.?g M Figure 17.~ We z 1  ^ K 6 u  P .2: Disturbance Rejection with Imaginary Axis Poles z1 z2 y1 ~ We Wd w Wu M ? g ^ K P u realizations: ~ We = Ae Be " # Figure 17.3: LFT Framework for the Disturbance Rejection Problem " # " # Ce De " Wu = Au Bu Cu Du # " Wd = Ad Bd Cd Dd # M = Am Bm Cm Dm 2 " P = Ap Bp : Cp Dp # " Then a realization for the generalized system is given by G(s) = 6 4 ~ We MWd MWd 0 ~ We MP Wu MP # 3 7 5 .
4.90 1 1 " # Then we can choose without loss of generality that 1 ~ > 0: M = s + We = s + s This gives the following generalized system 2 3 . Hence the controller K1 = . H2 and H1 Integral Control Ae 0 Be Cm Be DmCd Be Dm Cp Be Dm Dd Be Dm Dp 6 7 6 0 Au 0 Bu 7 0 0 0 7 6 6 0 0 Am Bm Cd Bm Cp Bm Dd Bm Dp 7 7 6 7 6 7 6 0 Bd 0 0 0 Ad 0 7: =6 6 0 0 0 0 Ap 0 Bp 7 7 6 7 6 6 Ce 0 De Cm De Dm Cd De DmCp De DmDd De Dm Dp 7 7 6 6 0 0 0 0 Du 7 5 4 0 Cu 0 0 Cm Dm Cd Dm Cp Dm Dd Dm Dp 2 3 453 We shall illustrate the above design through a simple numerical example. 191557) 7:85(ss(++ 32:+ 100)(s19:89) s(s + 32:17)( + :89) s 17)(s .90 7 6 7 6 0 0 0 .2 1 0 + 10 Wu = ss+ 100 = .2 0 7 6 7 6 7 6 0 0 0 0 1 0 0 7 7 G(s) = 6 6 0 0 0 3 2 0 1 7 6 7 6 7 6 1 0 0 0 0 0 0 7 6 7 6 1 0 0 0 0 1 7 4 0 5 0 0 1 . 0 ^ K1 = . 0:1557) + 100)(s . We = 1 : s .100 .100 0 0 0 0 .2 1 1 0 6 6 0 7 . P = (s +s1)(s2. Let 2 3 0 1 0 . 3) = 6 3 2 1 7 Wd = 1 4 5 . 0 1 .17. 19::1557) 2 s + 32:17)( (s + ) 89) ^ which gives the closedloop 1 norm 7:854.2 1 1 0 ^ and suboptimal H1 controller K1 for each can be computed easily as )(s + 100)( . 0: K1 (s) = 2060381:4(s +s1)(s262343)(s .2060381:(4(s + 1)(s +s + 262343)(ss.K1(s)M (s) is given by 1)(s .
3) The ltering problem is to nd an estimate z of z in some sense using the measurement ^ of y. 0 069) ^ K2 = . we can solve an optimal H2 controller + 1)(s + + . z has to be generated by a causal system acting on the ^ measurements. Now we can state our H1 ltering ^ problem. a controller for = 0:001 is given by s+ 7:85(s + 1)(s + 100)(s . 7:9718) : :001)(s :7)( In this section we show how the ltering problem can be solved using the H1 theory developed early.. given T > 0 the estimate z(t) = 0 8t 2 0 T ] if y(t) = 0 8t 2 0 T ]. The restriction on the ltering problem is that the lter has to be causal so that it can be realized. We will further restrict our lter to be unbiased. an approximate H2 integral controller is obtained as + 1)(s + 100)(s 0679) K2 = (s + 043:47(s 2 + 30:93s + 411. 0:1545) K1 = (s316880(s + 1)(32)(s100)(s . 7 964) An approximate integral control can also be achieved without going through the above process by letting 1 ~ We = We = s + M (s) = 1 for a su ciently small > 0. nd a causal lter F (s) 2 RH1 if it exists such that ^2 J := sup kz .e. 20) s(s + 32)(s . 7::964) ( + )2 s + :81)( and + 100)( ^ K2 (s) = . For example. Similarly. z2k2 < 2 with z = F (s)y.. i.e. 0:1545) + 0:001)(s + + 40370)(s . ^ w2L2 0 1) 17. 0:s . i. 20) which gives the closedloop H1 norm of 7:85. Similarly.1) y = C2 x + D21 w(t) (17.454 H1 CONTROL: GENERAL CASE which is independent of as expected.2) z = C1 x + D11 w(t) (17. 0::069) : (s + 30 94s + . Suppose a dynamic system is described by the following equations x = Ax + B1 w(t) x(0) = 0 _ (17. H1 Filtering: Given a > 0.K2(s)M (s) = s43:2487(s :+ 1)(s 411:81)(ss.5 H1 Filtering kwk2 .s43:487((s 2 + 30:94s )(s411100)(ss.
" 0 2I 0 0 # . However. comparing with control problems there is no internal stability requirement in the ltering problem.F (s) z ^ A B1 0 G(s) = 6 C1 D11 . j!I B1 C2 D21 # " # has full row rank for all !. Theorem 17. Let D21 be normalized and D11 partitioned conformably as D11 = D111 D112 : D21 0 I Then there exists a causal lter F (s) 2 RH1 such that J < and J1 2 dom(Ric) with Y1 = Ric(J1) 0 where ~ R := " " # 2 if and only if (D111 ) < D11 D21 #" D11 D21 # . The above ltering problem can also be formulated in an LFT framework: given a system shown below z G(s) w y . H1 Filtering z z 2 455 ? j .5 Suppose (C2 A) is detectable and " w2L2 2 sup kz k22 < 2 : kwk2 (17:4) A .4.I 4 C2 D21 0 2 3 7 5 nd a lter F (s) 2 RH1 such that Hence the ltering problem can be regarded as a special H1 problem. z ^ F (s) y 6 4 A B1 C1 D11 C2 D21 3 7 5 w Figure 17.17. Hence the solution to the above ltering problem can be obtained from the H1 solution in the previous sections by setting B2 = 0 and dropping the internal stability requirement.4: Filtering Problem Formulation A diagram for the ltering problem is shown in Figure 17.5.
then a rational causal lter F (s) satisfying J < 2 is given by A + L21 C2 + L11 D112 C2 . Y1 C2 C2 Y1 C2 y ^ C1 0 where Y1 is the stabilizing solution to " # Y1 A + AY1 + Y1 ( .B1D21 D21 B1 C2 # " # " # Moreover. B1 D11 + Y1 C1 B1 D21 + Y1 C2 R. But for convenience.L21 .456 H1 CONTROL: GENERAL CASE J1 := " A 0 C1 C2 ~ . This approach was the main focus in the early 1980's and is still very useful in solving many interesting problems.6 Youla Parameterization Approach* In this section. we have built enough tools to give a fairly complete solution using this Q parameterization approach.B1D11 .A . While the H1 problem is much more di cult. we will allow the assumption (A2) to be relaxed to the following .1 D11 B1 C1 : . Assume again that the G has the realization A B1 B2 G = 6 C1 D11 D12 4 C2 D21 D22 2 3 7 5 with the same assumptions as before. it bears some similarities to the constant matrix dilation problem but with the restriction of internal stability. if the above conditions are satis ed.2 C1 C1 .B1 B1 .1 : i h i In the case where D11 = 0 and B1 D21 = 0 the lter becomes much simpler: z = A . C2 C2 )Y1 + B1 B1 = 0: 17. Nevertheless. We will see that this approach may suggest additional interpretations for the results presented in the last chapter and applies to both H2 and H1 problems. we shall brie y outline an alternative approach to solving the standard H2 and H1 control problems using the Q parameterization (Youla parameterization) approach. L11 D112 z = F (s)y = ^ y C1 . R. D112 C2 D112 where h " # ~ L11 L21 := . The H2 problem is very simple and involves a projection.
B1 D21 C2 ) .C1 D? D?C1 .6.17. Youla Parameterization Approach* h i 457 (A20 ) D12 is full column rank with D12 D? unitary.L2 B2 + L2 D22 M2 = 6 F2 0 I 4 .(A . we will outline the steps required to solve the H2 and H1 control problems. and D21 is full row rank with " # Next.B2 F2 3 . Because of the similarity between the H2 and H1 problems. and B1 To = 6 0 AL2 B1L2 7 2 RH1 4 5 C1F2 . B1 D21 C2 ) Then the closedloop transfer matrix from w to z can be written as F`(G K ) = To + UQV I + D22 Q(1) invertible where AF2 := A + B2 F2 AL2 := A + L2 C2 C1F2 = C1 + D12 F2 B1L2 = B1 + L2 D21 : It is easy to show that U is an inner.B2 B2 X2 = Ric 0 .(A . U U = I .(B2 X2 + D12 C1 ) L2 = .D22 where 2 3 7 5 F2 = .(Y"C2 + B1 D21 ) 2 # A . Parameterization: Recall that all controllers stabilizing the plant G can be expressed as K = F` (M2 Q) I + D22 Q(1) invertible with Q 2 RH1 and D21 ~ D? unitary.(C2 + D22 F2 ) I .C2 C2 Y2 = Ric 0: ~ ~ . and V is a coinner.B1 D? D? B1 . B2 D12 C1 . they are developed in parallel below. A + B2 F2 + L2 C2 + L2 D22 F2 . B2 D12 C1 ) " # (A .D12F2 D11 # " # " AL2 B1L2 AF2 B2 V= U= C1F2 D12 C2 D21 2 AF2 . V V = I .
AL2 6 7 R=6 ~ 5 B2 F2 Y2 .X2 C1 D? C1F2 D? # V? = " Since the multiplication of square inner matrices do not change H2 and H1 norms. ~ H2 case: In this case.D?C1 X2+ .ED21 . Projection/Dilation: At this point the = 2 and = 1 cases have to be treated separately. D12 D11 B1L2 D12 D11 D21 D12 D11 D? 7 4 ~ .E D? 6 ~ 7 6 C2 . It can be shown through some long and tedious algebra that R is antistable and has the following representation: 2 3 ~ .458 H1 CONTROL: GENERAL CASE h i " # V Unitary Invariant: There exist U? and V? such that U U? and V ? square and inner: " are + U? = AF2 . the problem can be rewritten as with the obvious partitioning.D?D11 B1L2 D? D11 D21 D? D11 D? h R = U To V U? " V? i kF`(G K )k2 2 " = R11 + Q R12 R21 R22 # 2 2 .AF2 EB1L2 .Y2+ B1 D? 7 0 . D?B1 Y2+ D? # V U U? (To + UQV ) V? " # R11 + Q R12 R21 R22 h i # " # where E := X2 B1 + C1F2 D11 . we have for = 2 or 1 kF` (P K )k = kTo + UQV k = = where AL2 B1L2 : ~ ~ . we will assume D?D11 = 0 and D11 D? = 0 otherwise. the 2norm of the closed loop transfer matrix will be unbounded since " # ~ D12 D11 D21 D12 D11 D? : R(1) = ~ D? D11 D21 D? D11 D? Now from the de nition of 2norm.
However. Therefore. Youla Parameterization Approach* = Furthermore.1=2 1 (< ) . " Q2H1 min R11 + Q R12 R21 R22 R21 R22 # i # where 0 := max " 1 " > R12 R22 0 # ) ( h 1 It is convenient to use the characterization in Corollary 2. and.1=2 (Q + R11 + WR22 Z )(I .D12 D11 D21 : The optimal controller is given by Kopt = F` (M2 Qopt): In particular. if D11 = 0 then Kopt = F` (M2 0). it should be noted that since Q is restricted to be an RH1 matrix. 459 " kR11 + Qk2 + 2 R21 0 R12 R22 # 2 2 : kR11 + Qk2 = kR11 . WW ). Hence the unique optimal solution is given by a projection to H1 : Qopt = . The H1 optimal control follows this idea. (Q + D12 D11 D21 ) has to be in H2 to guarantee that the 2norm be nite.6.17. D12 D11 D21 ) 2 H2 and (Q + D12 D11 D21 ) 2 H1 . Recall that 1 : R11 + Q R12 R21 R22 i 1 (< ) for > 0 (I .23. H1 case: Recall the de nition of the 1norm: " R11 + Q R12 (j!): = sup R21 R22 ! 1 Consider the minimization problem with respect to Q frequency by frequency R11 + Q R12 R21 R22 # " #! it becomes a constant matrix dilation problem if no causality restrictions are imposed on Q. the term R11 + Q cannot be made into an arbitrary matrix. Z Z ). D12 D11 D21 k2 + kQ + D12 D11 D21 k2 2 2 2 ? since (R11 .R11 ]+ = . generically. the problem will be signi cantly di erent from the constant matrix case. In fact.
1=2 Z = ( 2 I . Then from the last section. we begin with G having state dimension n. WW )1=2 and N = (I . Now de ne N11 N12 := R : N21 N22 . R22 R22 ). then M M . The key is Lemma 15.1 (R11 + WR22 Z )N . R22 R22 ).4 and 16. we assume further that D11 = 0. Z Z )1=2 . The optimal control law will be given by Kopt = F` (J Qopt ): 17. This is a standard Nehari problem and is solved in Chapter 8. Now let ^ Q := M .1 G := M .460 where H1 CONTROL: GENERAL CASE W = R12 ( 2 I . Z Z )1=2 with stable inverses such that if M = (I .1. Z Z ).5 or Theorem 17.1=2 R21 : The key here is to nd the spectral factors (I . we have kTzw k1 = = = " R+ Q 0 " " # 0 0 1 # R11 + Q R12 R21 " R22 0 0 R + Q # # 1 1 0 where R has state dimension 2n.5) for Q 2 H1 . WW )1=2 and (I .7 might be used in the last section to prove Theorems 16. WW .1 N N . (17:5) ^ The nal step in the H1 problem involves solving (17. and N N = (I . For simplicity.7 and its connection with the formulae given in the last section.1 2 H1 MM = I .1 ^ then the problem is reduced to nding Q 2 H1 such that ^ G+Q 1 (< ): ^ ^ Note that Q = M QN 2 H1 i Q 2 H1 .7 Connections This section considers the connections between the Youla parameterization approach and the current state space approach discussed in the last chapter.1QN . To see how Lemma 15.
7 may be applied to derive the solvability conditions and some additional arguments to construct a Q 2 RH1 from X and Y such that kTzw (Q)k1 < . the apparent state dimension of K is 3n.X2 C1 D? 6 6 .17.C2 0 0 4 D21 B1 X2 ~ ? B1 X2 D? B1 Y2+ 0 ~ 0 D " 2 RH1 (17:6) and the H1 problem becomes to nd an antistable Q such that N11 + Q N21 N12 N22 # w Let w = 1 w2 " " # 1 < : and note that N11 + Q N21 N12 N22 # 2 " 1 := = w2BL2 sup N11 + Q N21 " fw2BL2 g\fw1 2H? g 2 8 " < #" sup N12 w1 N22 w2 2 #" N11 + Q N12 w1 N21 N22 w2 #" # #" # 2 # 2 2 : The right hand side of the above equation can be written as fw2BL2 g\fw1 2H? g : 2 sup P+ 0 0 I N11 N12 N21 N22 w1 w2 2 2 + kQ w1 + P.7.B BX AL2 Y2 F2 0 = 6 1L2 1 2 6 . it turns out that X in Lemma 15. (N11 w1 + N12 w2 )k2 : 2 It is clear that the last term can be made zero for an appropriate choice of an antistable Q . Lemma 15. but some tedious state space manipulations produce cancelations resulting in the n dimensional controller .7. In fact. Connections Then " 461 2 3 7 7 7 7 5 N11 N12 N21 N22 # + AF2 0 B2 . The nal step is to obtain the controller from M2 and Q. Hence the H1 problem has a solution if and only if " o fw2BL2 g\fw1 2H? g 2 sup P+ 0 0 I #" N11 N12 N21 N22 #" w1 w2 # 2 < : But this is exactly the same operator as the one in Lemma 15. Since M2 has state dimension n and Q has 2n.7 for N is exactly W in the FI proof or in the di erential game problem.
j!I B2 C1 D12 # has full column rank for all !. w(t) 2 Rm1 . it seems unnecessarily cumbersome and indirect.5. .4 and 16.7) (17. check if sup min kz k2 < and w2B L2 0 1) u2L2 0 1) u2L2 0 1) w2B L2 0 1) min sup kz k2 < : The rst problem can be regarded as a full information control problem since the control signal u can be a function of the disturbance w and the system state x. This approach is exactly the procedure used in Doyle 1984] and Francis 1987] with Lemma 15. Although this approach is conceptually straightforward and was. this is not true in general if D11 6= 0. used to obtain the rst proof of the current state space results in this chapter. y(t) 2 Rp2 . Consider a dynamical system x = Ax + B1 w + B2 u _ z = C1 x + D11 w + D12 u (17. it will be shown that the control signal in the latter problem depends only on the system state hence. as will be shown below.7 used to solve the general distance problem.462 H1 CONTROL: GENERAL CASE formulae in Theorems 16.8) where z (t) 2 Rp1 . the state feedback H1 control is itself a very interesting problem and deserves special treatment. the optimal control signal in the second problem cannot depend on the disturbance w (the worst disturbance w can be a function of u and x). However. The following assumptions are made: (AS1) (A B2 ) is stabilizable (AS2) There is a matrix D? such that D12 D? is unitary " h i (AS3) A . On the other hand. 17. This section and the section to follow are devoted to the study of this state feedback H1 control problem and its connections with full information control and di erential game. it is equivalent to a state feedback control. In fact. and x(t) 2 Rn . u(t) 2 Rm2 . in fact.8 State Feedback and Di erential Game It has been shown in Chapters 15 and 16 that a (central) suboptimal full information H1 control law is actually a pure state feedback if D11 = 0. Nevertheless. We are interested in the following two related quadratic minmax problems: given > 0.
6 Let > 0 be given and de ne R := D1 D1 .1 D1 C1 + B X1 ] : F21 sup u2L2 0 1) w2B L2 0 1) min kz k2 < if and only if H1 2 dom(Ric) X1 = Ric(H1 0: worst + h (D11 ) < Moreover.D12D11 w and the worst wsfworst is given by D12 D11 I Fx i wsfworst = ( 2 I .R.C1 D1 R.1 f(D11 C1 + B1 X1 )x + D11 D12 ug : Proof. It is not hard to see that (D11 ) < is necessary since control u cannot feed back w directly.C1 C1 .A where B := B1 B2 . the solution to the FI problem for the general case. We now prove the condition for part (b). D11 D11 ). The condition for part (a) can be shown in the same way as in Chapter 15 and is. so the D11 term cannot be (partially) eliminated as in . an optimal u is given by and a worst w worst is given by u = . an optimal u is given by u = F21 x = . (a) w2B L2 0 1) u2L2 0 1) h i B . in fact. . State Feedback and Di erential Game Theorem 17. " " 463 2 Im 0 1 # 0 0 " # where D1 := D11 D12 ] # A 0 H1 := .1 D1 C1 B h i sup min kz k2 < if and only if (D? D11 ) < H1 2 dom(Ric) X1 = Ric(H1 ) 0: h i Moreover.D12D11 w + D12 D11 I Fx w where (b) " # worst = F11 x F := F11 := .8.17.
wsfworst ) 2 2 if conditions in (b) are satis ed. 2 kwk2 + dt (x Xx) = R0=2 (u .D12 D11 w + D12 D11 I Fx w = F11 x and that an optimal control law and a worst disturbance for problem (b) are h i u = F21 x w = wsfworst : . Integrating both equations from t = 0 to 1 with x(0) = x(1) = 0 gives d ~1 kz k2 . The conditions H1 2 dom(Ric) and X1 = Ric(H1 0 can be easily seen as necessary since w2B L2 0 1) u2L2 0 1) sup min kz k2 u2L 0 1) sup kz k2: min 2 w2B L2 0 1) It is easy to verify directly that the optimal control and the worst disturbance can be chosen in the given form. D12 D11 I Fx h i 2 . 2 kwk2 = R0=2 (u . D11 D11 = 2: Then it is easy to show that d kz k2 . F21 x) 2 . F11 x) 2 2 ^1 R0=2 (w . These relations suggest that an optimal control law and a worst disturbance for problem (a) are u = . h i 2 ^1 R0=2 (w . D11 D? D? D11 = 2 ~ R0 = I + D12 D11 ( 2 I . On the other hand. D11 D11 ).1 D11 D12 ^ R0 = I . we have if conditions in (b) are satis ed. 2 2 if conditions in (a) are satis ed. D12 D11 I Fx 2 . 2 kwk2 = u + D12 D11 w . 2 2 2 2 2 1 R0=2 (w . 2 De ne R0 = I . F11 x) 2 2 if conditions in (a) are satis ed. 2 1 R0=2 (w .464 H1 CONTROL: GENERAL CASE the FI problem. wsfworst ) 2 kz k2 . and 2 ~1 kz k2 . 2 kwk2 + dt (x Xx) = u + D12 D11 w . F21 x) .
A B1 B2 G = 6 C1 0 D12 7 : 4 5 I 0 0 2 3 . we shall consider the parameterization of all state feedback control laws.1 .7 Suppose that the assumptions (AS 1) . 2 kwk2 min 2 2 2 2 2 w2L2 0 1) u2L2 0 1) w2L2 0 1) since the solvability of problem (b) implies the solvability of problem (a). i. Parameterization of State Feedback H1 Controllers 465 Moreover. w2L2 0 1) 17. In fact. then the corresponding di erential game problem min sup kz k2 . it is easy to construct an example so that the problem (a) has a solution for a given and problem (b) does not.r h i ^ where F1 = .17. sup min kz k2 . r). the converse may not be true. Furthermore. Hence the parameterization cannot be obtained from the output feedback. if problem (b) has a solution for a given . 2 kwk2 = u2L 0 1) sup kz k2 . and Q = Q1 Q2 2 RH2 with kQ1 k1 < .1 0 0 . On the other hand. We shall assume Note that the state feedback H1 problem is not a special case of the output feedback problem since D21 = 0.9. the problems (a) and (b) are equivalent if D11 = 0.r 0 Im1 . 2 kwk2 2 2 u2L 0 1) 2 has a saddle point. if these conditions are satis ed.1 (sI . then all admissible controllers satisfying kTzw k1 < can be parameterized as ( " # ).2B1 B1 X1 +B2 F1 . However. A) 0 Im1 . The dimensions of Q1 and Q2 are m2 r and m2 (m1 . respectively.9 Parameterization of State Feedback H1 Controllers In this section.1 " # .e. Theorem 17. We shall rst assume for simplicity that D11 = 0 and show later how to reduce the general D11 6= 0 case to an equivalent problem with D11 = 0.1 B2 ^ K = F1 + Im2 + Q U Q U .(D12 C1 +B2 X1 ). A = A+ . (AS 3) are satis ed and that B1 has the following SVD: " # 0 B1 = U V UU = In V V = Im1 0 < 2 Rr r : 0 0 There exists an admissible controller K (s) for the SF problem such that kTzw k1 < if and only if H1 2 dom(Ric) and X1 = Ric(H1) 0.
I I 0 ^ Then Tvr = F` (GSF F` (Mtmp )) =: F` (Ntmp ). we make ^ ^ the same change of variables to get GSF instead of GFI : z r XXXXX X ^ GSF P w v r u 1 w y  K 2 3 Atmp B1 B2 ^ SF = 6 .466 H1 CONTROL: GENERAL CASE Proof. All controllers ^ that stabilize GSF can be parameterized as K = F`(Mtmp ) 2 RH1 where Atmp + B2 F1 + L . The conditions for the existence of the state feedback control law have been shown in the last section and in Chapter 15.2 and Lemma 16. and then Atmp + L = Atmp + B2 F1 is stable.L B2 Mtmp = 6 F1 0 I 7: 4 5 . As in the proof of the FI problem in Chapter 16.F1 0 I 7: 5 0 0 I Now let = F1 + ^ .F1 0 I 7 : G 4 5 I 0 0 So again from Theorem 16. we conclude that K is an admissible ^ controller for G and kTzw k1 < i K is an admissible controller for GSF and kTvr k1 < . It is easy to show that 2 3 Atmp + B2 F1 B1 0 6 Ntmp = 4 .8. We only need to show that the above parameterization indeed satis es the H1 norm condition and gives all possible state feedback H1 control laws. and we have 2 3 F`(Ntmp ) = ^ Atmp + B2 F1 B1 : I 0 " # . Now let L = B2 F1 .
we get the desired controller parameterization. D11 D11 )1=2 # kF` (N F` (Gg K ))k1 < 1: Note that with 2 ~ F`(N F` (Gg K )) = F` (M K ) 3 7 7 5 . we have the following lemma.2. Lemma 17. . which in turn implies that kTvr k1 = kF`(Ntmp )k1 < if and only if kQ1 k1 < .9. substituting = F1 + ^ into K = F` (Mtmp ). D11 D11 )1=2 D11 and then N N = I . D11 D11 .17. 2 The controller parameterization for the general case can also be obtained: i 0 . Parameterization of State Feedback H1 Controllers Let ^ = Q1 Q2 h h i " 467 Then the mapping from Q = Q1 Q2 2 RH2 to ^ 2 RH1 is onetoone.7 can be applied to the new system M (s) to obtain the controller parameterization for the general problem with D11 6= 0. . by Theorem 16. In summary. Hence we have h i F`(Ntmp ) = Q1 0 V and kF`(Ntmp )k1 = kQ1 k1 .1 0 # I ^ U . we have that K stabilizes Gg and kF`(Gg K )k1 < 1 if and only if F`(Gg K ) internally stabilizes N and 2 3 N= " .D11 (I .1 (sI . A): A B1 B2 6 Gg (s) = 4 C1 D11 D12 7 : 5 I 0 0 Let (I .1=2 B2 + B1 D11 R1 1D12 6 . M (s) = 6 R1 1=2 C1 0 R1 1=2 D12 4 I 0 0 where R1 := I . A + B1 D11 R1 1 C1 B1 (I . Finally. D11 D11 ).8 There is a K that internally stabilizes Gg and kF`(Gg K )k1 < 1 if and ~ ~ only if kD11 k < 1 and there is a K that stabilizes M and F` (M K ) < 1. 1 Now Theorem 17. . Furthermore.
1989]. The loopshifting procedures are given in Safonov.10 Notes and References The detailed derivation of the H1 solution for the general case is treated in Glover and Doyle 1988. The idea is also used in Zhou and Khargonekar 1988] for state feedback problems.468 H1 CONTROL: GENERAL CASE 17. The presentation of the state feedback H1 control in this chapter is based on Zhou 1992]. A fairly complete solution to the singular H1 problem is obtained in Stoorvogel 1992]. Limebeer. . and Chiang 1989]. The Youla parameterization approach is treated very extensively in Doyle 1984] and Francis 1987] and in the references therein. The H1 ltering and smoothing problems are considered in detail in Nagpal and Khargonekar 1991].
and K internally stabilizes the nominal system. The loop shaping design procedure is described in Section 18.1 (N + ~ N ) ~ ~ ~ ~ with M N ~ M ~ N 2 RH1 and ~ N ~ M < . The theoretical justi cation for the loop shaping design procedure is given in Section 18. and a particular H1 optimization problem to guarantee closedloop stability and a level of robust stability at all frequencies.1.2.3. 18 18.e. P = M . The proposed technique uses only the basic concept of loop shaping methods and then a robust stabilization controller for the normalized coprime factor perturbed system is used to construct the nal controller.1N ). we use the H1 control theory developed in the previous chapters to solve the robust stabilization of a left coprime factor perturbed plant given by ~ ~ P = (M + ~ M ). This chapter is arranged as follows: The H1 theory is applied to solve the stabilization problem of a normalized coprime factor perturbed system in Section 18. The transfer matrices (M N ) 1 ~ ~ are assumed to be a left coprime factorization of P (i. 469 h i .1 Robust Stabilization of Coprime Factors In this section.H1 Loop Shaping This chapter introduces a design technique which incorporates loop shaping methods to obtain performance/robust stability tradeo s..
470 H1 LOOP SHAPING z1 r  ~N ~ N .1: Left Coprime Factor Perturbed Systems It has been shown in Chapter 9 that the system is robustly stable i " Finding a controller such that the above norm condition holds is an H1 norm minimization problem which can be solved using H1 theory developed in the previous chapters.L # " 0 " =: 6 4 A B1 C1 D11 C2 D21 I I 3 B2 D12 7 : 5 D22 B I D D 3 # 7 7 7 7 5 .1N is given by # " h i A + LC B + LD L : ~ ~ N M = C D I Denote then the system diagram can be put in an LFT form as in Figure 18.K G(s) = 6 4 0 ~ .1 ~ 1= : I 1 # P= A B C D and let L be a matrix such that A + LC is stable then a left coprime factorization of ~ ~ P = M .6 K .M .1 f ~  Figure 18.1 M . w ~M z2 y f .? .2 with the generalized plant 2 " " # ^ K = .1 # " I P P # 3 7 5 2 = 6 " 6 6 6 4 2 A C C 0 # . Suppose P has a stabilizable and detectable state space realization given by K (I + PK ).f.1 M ~ M .
18.2: LFT Diagram for Coprime Factor Stabilization To apply the H1 control formulae in Chapter 17.(I + DD ). 1 D A 2 . we need to normalize the \D12" matrix rst. 1 I + D D).1 471 w ? i ~ N y u  ^ K Figure 18.1 D C (I + D D) 2 C " 2 .(I + DD ). 1 K + # 2~ # z1 = U z1 : ^ z2 z2 ^ Then kTzw k1 = kU Tzw k1 = kTzw k1 and the problem becomes of nding a controller ^ ~ K so that kTzw k1 < with the following generalized plant ^ ^ G= U 2 6 " 6 6 6 4 " 0 I # 0 # G I " 1 0 (I + D D). 2 0 # : I 2 I D(I + D D).(I + DD ). 1 ^ ~ Now the formulae in Chapter 17 can be applied to1 G to obtain a controller K and then .L # B# " 0 3 7 7 7 7 5 .(I + D D) = 2 (I + D D). 1 and U is a unitary matrix.1. 1 C . We shall leave the detail to the ~ the K can be obtained from K = . 2 K . Note that " " # " # 1 # 2 I = U 0 (I + D D) 2 where U = D (I + DD ). Let " ^ K = (I " D D). 2 1 2 2 D I . 1 D(I + D D). 1 . Robust Stabilization of Coprime Factors z1 z2 ~ M .
472
H1 LOOP SHAPING
2 LC LC X1 (A ; 2 ; 1 ) + (A ; 2 ; 1 ) X1 ; X1 (BB ; LL 1 )X1 + 2C C = 0 2; ;1
reader. In the sequel, we shall consider the case D = 0. In this case, we have > 1 and
It is clear that Y1 = 0 is the stabilizing solution. Hence by the formulae in Chapter 17 we have h i h i L11 L21 = 0 L and ^ ^ ^ Z1 = I D11 = 0 D12 = I D21 =
p
Y1 (A + LC ) + (A + LC )Y1 ; Y1 C CY1 = 0:
2;1
I:
The results are summarized in the following theorem.
Theorem 18.1 Let D = 0 and let L be such that A + LC is stable then there exists a
controller K such that
K (I + PK );1 M ;1 < ~ I 1 i > 1 and there exists a stabilizing solution X1 0 solving 2 LC LC X1 (A ; 2 ; 1 ) + (A ; 2 ; 1 ) X1 ; X1 (BB ; LL 1 )X1 + 2C C = 0: 2; ;1
Moreover, if the above conditions hold a central controller is given by
"
#
A ; BB X1 + LC L K= : ; B X1 0
It is clear that the existence of a robust stabilizing controller depends upon the choice of the stabilizing matrix L, i.e., the choice of the coprime factorization. Now let Y 0 be the stabilizing solution to
"
#
AY + Y A ; Y C CY + BB = 0
~ ~ and let L = ;Y C . Then the left coprime factorization (M N ) given by
h
A ; Y C C B ;Y C ~ ~ N M = C 0 I
i
"
#
is a normalized left coprime factorization (see Chapter 13).
18.1. Robust Stabilization of Coprime Factors
473
~ ~ Then P = M ;1 N is a normalized left coprime factorization and
"
Corollary 18.2 Let D = 0 and L = ;Y C where Y 0 is the stabilizing solution to AY + Y A ; Y C CY + BB = 0:
K stabilizing
inf
K (I + PK );1 M ;1 ~ = I 1
=
#
p
1 ; max (Y Q)
h
1
1;
~ ~ N M
i
2 ;1=2
where Q is the solution to the following Lyapunov equation Q(A ; Y C C ) + (A ; Y C C ) Q + C C = 0: Moreover, if the above conditions hold then for any > min a controller achieving
"
H
=: min
is given by where
K (I + PK );1 M ;1 < ~ I 1
#
#
A ; BB X1 ; Y C C ;Y C K (s) = ; B X1 0 X1 =
2 ; 1Q I ; 2 2 ; 1Y Q 2
"
;1
:
Proof. Note that the Hamiltonian matrix associated with X1 is given by " # A + 21;1 Y C C ;BB + 21;1 Y C CY H1 = : 2 ;(A + 21;1 Y C C ) ; 2 ;1 C C
Straightforward calculation shows that
"
where
I ; 2 ;1 Y H1 = 2 0 2 ;1 I
"
2
#
I ; 2;1 Y Hq 2 0 2 ;1 I
#
"
2
#;1
0 Hq = A ; Y C C : ;C C ;(A ; Y C C ) It is clear that the stable invariant subspace of Hq is given by " # I X; (Hq ) = Im Q
474 and the stable invariant subspace of H1 is given by
"
H1 LOOP SHAPING X; (H1 ) = I ; 22;1 Y 0 2 ;1 I
2
#
X; (Hq ) = Im I ; 22;1 Y Q : 2 ;1 Q
"
2
#
Hence there is a nonnegative de nite stabilizing solution to the algebraic Riccati equation of X1 if and only if 2 I ; 2 ; 1Y Q > 0 or 1 1 ; max (Y Q) and the solution, if it exists, is given by
>p
2
X1 =
2 ; 1Q I ;
2 ; 1Y Q
2
;1
:
Note that iY and Q are the controllability Gramian and the observability Gramian of h h i ~ ~ ~ ~ N M respectively. Therefore, we also have that the Hankel norm of N M p is max (Y Q). 2 ~ ~ Corollary 18.3 Let P = M ;1N be a normalized left coprime factorization and ~ ~ P = (M + ~ M );1 (N + ~ N )
with
h i
~N ~M < : 1 Then there is a robustly stabilizing controller for P if and only if
p
1 ; max (Y Q) = 1 ;
r
h
~ ~ N M
i
2
H
:
The solutions to the normalized left coprime factorization stabilization problem are also solutions to a related H1 problem which is shown in the following lemma. ~ ~ Lemma 18.4 Let P = M ;1N be a normalized left coprime factorization. Then
"
K (I + PK );1M ;1 = ~ I 1
#
"
K (I + PK );1 h I P I
#
i
1
:
18.1. Robust Stabilization of Coprime Factors
~ ~ Proof. Since (M N ) is a normalized left coprime factorization of P , we have
h
475 =I
~ ~ M N
i
ih
~ ~ M N
h
i
and
h
Using these equations, we have
~ ~ M N
"
1
#
=
~ ~ M N
i
1
= 1:
"
=
" "
=
" "
= This implies
" #
K I K I K I K I K I
# # # # #
K (I + PK );1M ;1 ~ I 1
h ih i
~ ~ ~ (I + PK );1 M ;1 M N ~ ~ ~ (I + PK );1 M ;1 M N (I + PK );1 I P ~ (I + PK );1 M ;1 ~ (I + PK );1 M ;1
1 1
" h i h h
~ ~ M N
h
i
1
i
1
~ ~ M N
1
1
~ ~ M N
i
1
:
# i
K (I + PK );1M ;1 = ~ I 1
K (I + PK );1 h I P I
1
:
2
Corollary 18.5 A controller solves the normalized left coprime factor robust stabilization problem if and only if it solves the following H1 control problem
"
I (I + PK );1 h I P K
i
#
i
and
1
p
<
1
"
inf K stabilizing
I (I + PK );1 h I P K
#
1
= =
1 ; max (Y Q)
h
1;
~ ~ N M
i
2 ;1=2
H
:
476
H1 LOOP SHAPING
0 be the stabilizing
The solution Q can also be obtained in other ways. Let X solution to XA + A X ; XBB X + C C = 0 then it is easy to verify that Q = (I + XY );1 X: Hence
min =
p
H 1 ; max (Y Q) Similar results can be obtained if one starts with a normalized right coprime factorization. In fact, a rather strong relation between the normalized left and right coprime factorization problems can be established using the following matrix fact.
1
= 1;
h
~ ~ N M
i
2 ;1=2
= 1 + max (XY ):
p
Lemma 18.6 Let M and N be any compatibly dimensioned complex matrices such that MM = M , NN = N , and M + N = I . Then i (M ) = i (N ) for all i such that 0 < i (M ) = 1. 6 Proof. We rst show that the eigenvalues of M and N are either 0 or 1 and M and N are diagonalizable. In fact, assume that is an eigenvalue of M and x is a corresponding eigenvector, then x = Mx = MMx = M (Mx) = Mx = 2 x, i.e., (1 ; )x = 0.
This implies that either = 0 or = 1. To show that M is diagonalizable, assume M = TJT ;1 where J is a Jordan canonical form, it follows immediately that J must be diagonal by the condition M = MM . The proof for N is similar. Next, assume that M is diagonalized by a nonsingular matrix T such that
M = T I 0 T ;1:
0 0
"
"
#
Then De ne
N = I ; M = T 0 0 T ;1: 0 I
"
#
and assume 0 < 6= 1. Then A > 0 and
A B := T T B D
#
,
det(M M ;# I ) =" 0 " # I 0 T T I 0 ; T T) = 0 det( 0 0 0 0
18.1. Robust Stabilization of Coprime Factors
, , , , ,
"
477
det
(1 ; )A ; B
#
;B
det(; D ; 1 ; B A;1 B ) = 0 det( 1 ; D + B A;1 B ) = 0
2
;D
=0
; A ; B =0 det ; B (1 ; )D det(N N ; I ) = 0:
"
#
This implies that all nonzero eigenvalues of M M and N N that are not equal to 1 are equal, i.e., i (M ) = i (N ) for all i such that 0 < i (M ) 6= 1. 2 Using this matrix fact, we have the following corollary.
Corollary 18.7 Let K and P be any compatibly dimensioned complex matrices. Then
"
I (I + PK );1 h I P K
#
#
i
"
=
I (I + KP );1 h I K P
" #
#
i
:
Proof. De ne
"
h M = I (I + PK );1 I P K
i
h i N = ;P (I + KP );1 ;K I : I
Then it is easy to verify that MM = M , NN = N and M + N = I . By Lemma 18.6, we have kM k = kN k. The corollary follows by noting that
" i I (I + KP );1 h 0 I N 0 ;I : I K = P ;I 0 I 0 # " # " #
2
~ ~ Corollary 18.8 Let P = M ;1N = NM ;1 be respectively the normalized left and right
coprime factorizations. Then
" #
K (I + PK );1M ;1 = M ;1 (I + KP );1 h ~ I K I 1
i
1
:
478
H1 LOOP SHAPING
h i "
Proof. This follows from Corollary 18.7 and the fact that
M ;1 (I + KP );1 I K
1
=
I (I + KP );1 h I K P
#
i
1
:
2
This corollary says that any H1 controller for the normalized left coprime factorization is also an H1 controller for the normalized right coprime factorization. Hence one can work with either factorization.
18.2 Loop Shaping Using Normalized Coprime Stabilization
;
This section considers the H1 loop shaping design. The objective of this approach is to incorporate the simple performance/robustness tradeo obtained in the loop shaping, with the guaranteed stability properties of H1 design methods. Recall from Section 5.5 of Chapter 5 that good performance controller design requires that (I + PK );1
;
(I + PK );1 P
;
(I + KP );1
;
K (I + PK );1 (18:1)
be made small, particularly in some low frequency range. And good robustness requires that ; ; PK (I + PK );1 KP (I + KP );1 (18:2) be made small, particularly in some high frequency range. These requirements in turn imply that good controller design boils down to achieving the desired loop (and controller) gains in the appropriate frequency range: (PK ) in some low frequency range and (PK ) 1 (KP ) 1 (K ) M in some high frequency range where M is not too large. The design procedure is stated below. 1 (KP ) 1 (K ) 1
Loop Shaping Design Procedure
(1) Loop Shaping: Using a precompensator W1 and/or a postcompensator W2 , the singular values of the nominal plant are shaped to give a desired openloop shape. The nominal plant G and the shaping functions W1 W2 are combined to form the shaped plant, Gs where Gs = W2 GW1 . We assume that W1 and W2 are such that Gs contains no hidden modes.
18.2. Loop Shaping Using Normalized Coprime Stabilization
re
479
;6

K
? e
di
u

P
 ? ye ? en
d
Figure 18.3: Standard Feedback Con guration (2) Robust Stabilization: a) Calculate max , where
" # !;1
I (I + G K );1 M ;1 ~s inf max = s K stabilizing K 1 r h i 2 ~ ~ = 1 ; Ns Ms <1 H
~ ~ ~ ~ and Ms Ns de ne the normalized coprime factors of Gs such that Gs = Ms;1 Ns and ~ ~ ~ ~ Ms Ms + Ns Ns = I: If max 1 return to (1) and adjust W1 and W2 . b) Select max , then synthesize a stabilizing controller K1 , which satis es
I (I + G K );1 M ;1 ;1 : ~s s 1 K1 1 (3) The nal feedback controller K is then constructed by combining the H1 controller K1 with the shaping functions W1 and W2 such that K = W1 K1 W2 :
A typical design works as follows: the designer inspects the openloop singular values of the nominal plant, and shapes these by pre and/or postcompensation until nominal performance (and possibly robust stability) speci cations are met. (Recall that the openloop shape is related to closedloop objectives.) A feedback controller K1 with associated stability margin (for the shaped plant) max , is then synthesized. If max is small, then the speci ed loop shape is incompatible with robust stability requirements, and should be adjusted accordingly, then K1 is reevaluated. In the above design procedure we have speci ed the desired loop shape by W2 GW1 . But, after Stage (2) of the design procedure, the actual loop shape achieved is in fact
"
#
480
H1 LOOP SHAPING
 W1 ;6
e
G
 W2  W2
 K1
 W1 
G Gs
;6
e
 W2  K1  W1
K

G
Figure 18.4: The Loop Shaping Design Procedure given by W1 K1W2 G at plant input and GW1 K1 W2 at plant output. It is therefore possible that the inclusion of K1 in the openloop transfer function will cause deterioration in the openloop shape speci ed by Gs . In the next section, we will show that the degradation in the loop shape caused by the H1 controller K1 is limited at frequencies where the speci ed loop shape is su ciently large or su ciently small. In particular, we show in the next section that can be interpreted as an indicator of the success of the loop shaping in addition to providing a robust stability guarantee for the closedloop systems. A small value of max ( max 1) in Stage (2) always indicates incompatibility between the speci ed loop shape, the nominal plant phase, and robust closedloop stability.
Remark 18.1 Note that, in contrast to the classical loop shaping approach, the loop
shaping here is done without explicit regard for the nominal plant phase information. That is, closedloop stability requirements are disregarded at this stage. Also, in contrast with conventional H1 design, the robust stabilization is done without frequency weighting. The design procedure described here is both simple and systematic, and only assumes knowledge of elementary loop shaping principles on the part of the designer.
~
Remark 18.2 The above robust stabilization objective can also be interpreted as the more standard H1 problem formulation of minimizing the H1 norm of the frequency
weighted gain from disturbances on the plant input and output to the controller input
18.3. Theoretical Justi cation for H1 Loop Shaping
and output as follows.
481
"
I (I + G K );1 M ;1 ~s = s 1 K1 1
= = =
#
" " " "
This shows how all the closedloop objectives in (18.1) and (18.2) are incorporated. ~
i I (I + G K );1 h I Gs s 1 K1 1 # i W2 (I + GK );1 h ;1 W2 GW1 W1;1 K 1 # h i I (I + K G );1 I K1 1 s Gs 1 # i W1;1 (I + KG);1 h W1 GW2;1 W2 G 1
#
18.3 Theoretical Justi cation for H1 Loop Shaping
The objective of this section is to provide justi cation for the use of parameter as a design indicator. We will show that is a measure of both closedloop robust stability and the success of the design in meeting the loop shaping speci cations. We rst examine the possibility of loop shape deterioration at frequencies of high loop gain (typically low frequency). At low frequency (in particular, ! 2 (0 !l )), the deterioration in loop shape at plant output can be obtained by comparing (GW1 K1 W2 ) to (Gs ) = (W2 GW1 ). Note that (GK ) = (GW1 K1 W2 ) = (W2;1 W2 GW1 K1 W2 ) (W2 GW1 ) (K1 )= (W2 ) (18:3) where ( ) denotes condition number. Similarly, for loop shape deterioration at plant input, we have (KG) = (W1 K1W2 G) = (W1 K1 W2 GW1 W1;1 ) (W2 GW1 ) (K1 )= (W1 ): (18:4) In each case, (K1 ) is required to obtain a bound on the deterioration in the loop shape at low frequency. Note that the condition numbers (W1 ) and (W2 ) are selected by the designer. Next, recalling that Gs denotes the shaped plant, and that K1 robustly stabilizes the normalized coprime factorization of Gs with stability margin , then we have
"
~ ~ where (Ns Ms ) is a normalized left coprime factorization of Gs , and the parameter is de ned to simplify the notation to follow. The following result shows that (K1 )
I (I + G K );1 M ;1 ~s s 1 K1 1
#
;1 :=
(18:5)
482
H1 LOOP SHAPING
is explicitly bounded by functions of and (Gs ), the minimum singular value of the shaped plant, and hence by (18.3) and (18.4) K1 will only have a limited e ect on the speci ed loop shape at low frequency.
Theorem 18.9 Any controller K1 satisfying (18.5), where Gs is assumed square, also
satis es
(K1 (j!))
for all ! such that
p (Gs (j!)) ; 2 ; 1 p 2 ; 1 (G (j!)) + 1 s p
(G (j!)) > 2 ; 1: p 2 ; 1, sthen (K (j!)) ' 1=p 2 ; 1, where ' denotes Furthermore, if (Gs ) 1 asymptotically greater than or equal to as (Gs ) ! 1.
Proof. First note that (Gs ) >
p
~ ~ Further since (Ns Ms ) is a normalized left coprime factorization of Gs , we have ~ ~ ~ ~ ~ ~ Ms Ms = I ; Ns Ns = I ; Ms Gs Gs Ms : Then Now can be rewritten as
I + Gs Gs > 2 I:
2 ; 1 implies that
"
~ ~ Ms Ms = (I + Gs Gs );1 < ;2 I:
I (I + G K );1 M ;1 ~s s 1 K1 1
2(I ; K G )(M Ms )(I ; Gs K1 ): 1 s ~s ~
#
(I + K1 K1 )
(18:6)
We will next show that K1 is invertible. Suppose that there exists an x such that K1 x = 0, then x (18:6) x gives ;2 x x x M Ms x ~s ~
~ ~ which implies that x = 0 since Ms Ms < ;2I , and hence K1 is invertible. Equation (18.6) can now be written as ; ; ; ~ ~ ; (K1 K11 + I ) 2 (K1 ; Gs )Ms Ms (K11 ; Gs ): (18:7) De ne W such that ~ ~ (WW );1 = I ; 2 Ms Ms = I ; 2 (I + Gs Gs );1
18.3. Theoretical Justi cation for H1 Loop Shaping
; and completing the square in (18.7) with respect to K11 yields ; ; (K1 + N )(WW );1 (K11 + N ) ( 2 ; 1)R R
483
where
N = 2Gs ((1 ; 2 )I + Gs Gs );1 R R = (I + Gs Gs )((1 ; 2 )I + Gs Gs );1 :
Hence we have
; ; R; (K1 + N )(WW );1 (K11 + N )R;1 ( 2 ; 1)I
and Use
1
;W ;1(K ;1 + N )R;1
;W ;1(K ;1 + N )R;1
; (K11 + N )
2 ; 1: 1 ; (W ;1 ) (K11 + N ) (R;1 ) to get
p
p
2 ; 1 (W ) (R)
; and use (K11 + N )
(K1 ) ; (N ) to get (K1 )
n
o;1 ( 2 ; 1)1=2 (W ) (R) + (N ) :
(18:8)
Next, note that the eigenvalues of WW N N , and R R can be computed as follows 1 ( sG ) (WW ) = 1 ; + + G(G sG ) 2 s s 4 ( (N N ) = (1 ; 2 +Gs Gs ) ))2 (Gs Gs therefore (W ) = 1 ( sG ) (R R) = 1 ; + + G(G sG ) 2 s s
p
2p min (Gs Gs ) 2 ( (N ) = max (N N ) = 1 ; 2 + = 1 ; 2 +Gs ) G ) 2( s min (Gs Gs ) p 1 + 2 (Gs ) 1=2 : 1 min s G ) 1=2 = (R) = max (R R) = 1 ; + + (G(G sG ) 2 min s s 1 ; 2 + 2 (Gs )
1 min s G ) 1=2 = 1 + 2 (Gs ) 1=2 (WW ) = 1 ; + + (G(G sG ) max 2 min s s 1 ; 2 + 2 (Gs )
p
Further.1) directly determines the frequency range over which this result is valid{a small (large ) is desirable. if (Gs ) 1= 1 asymptotically less than or equal to as (Gs ) ! 0.9. The value of (= .9. 1)1=2 (1 + 2 (Gs )) + 2 (Gs ) . p 2 .10 Any controller K1 satisfying (18. ( 2 . and is only ~ ~ sketched here: As in the proof of Theorem 18.8). 2 . the corresponding deterioration in plant input loop shape is obtained by comparing (W1 K1 W2 G) to (W2 GW1 ) where Hence. where / denotes Proof.1 = p Gs )) . we have p ( 2 . Gs K1 ): (18:9) . we now examine the possibility of deterioration in the loop shape at high frequency due to the inclusion of K1 . analogously to (18. Note that at high frequency (in particular. The proof of Theorem 18.9. (K1 ) is required to obtain a bound on the deterioration in the loop shape at high frequency. 1 (Gs (j!)) for all ! such that 1 (Gs (j!)) < p 2 : . the maximum singular value of the shaped plant. 1 : ( (K1 ) 2 (Gs ) . and the stability margin of the shaped plant. 1) 2 .1) is su ciently small.1 and ~ ~ (I + K1 K1 ) 2(I . Theorem 18. then so also will Gs K1 provided (= . 1.5) also satis es p 2 . in each case.1 p 2 .3) and (18. In an analogous manner.4) we have (GK ) = (GW1 K1W2 ) (KG) = (W1 K1 W2 G) (W2 GW1 ) (K1 ) (W2 ): (W2 GW1 ) (K1 ) (W1 ): Similarly. we have Ms Ms = (I + Gs Gs ).484 H1 LOOP SHAPING Substituting these formulas into (18. 1. K1 Gs )(Ms Ms )(I . 2 . and (Gs ).9 is that the bound on (K1 ) depends only on the selected loop shape.10 is similar to that of Theorem 18. Gs has a su ciently large loop gain. 1 (Gs ) + 1 2 The main implication of Theorem 18. In an identical manner to Theorem 18. we now show that (K1 ) is explicitly bounded by functions of . ! 2 (!h 1)) the deterioration in plant output loop shape can be obtained by comparing (GW1 K1W2 ) to (Gs ) = (W2 GW1 ). 1 + (G (j!)) s p (K1 (j!)) 1 . as we would expect. then (K (j!)) / Furthermore. Note that.
10 con rm that (alternatively ) indicates the compatibility between the speci ed loop shape and closedloop stability requirements.1 ~s s 1 K1 1 # .9 and 18. Theoretical Justi cation for H1 Loop Shaping Since (Gs ) < p 12 .1 (18:10) 2 (Gs ) : 1 .1 p 2. ( . 2 . 1 (Gs ) (M ) = 2 The results in Theorem 18. 1)Y Y where I .1 . Theorem 18.9) with respect to K1 yields (K1 + M )V V (K1 + M ) ( 2 . 2 )Gs Gs ). ( 2 .1 Y Y = ( 2 . 1 + (G ) ( 2 .1 ) + (M ): As in the proof of Theorem 18. 2)Gs Gs ).1 (V ) (Y . 1)1=2 (1 + 2 (Gs )) + 2 (Gs ) = s : p (K1 ) 1 .1 M . 485 and there exists a spectral factorization V V = I . 1) 2 (Gs ) Substituting these formulas into (18.10). 1) (Gs ) (V ) = (Y 1 + 2 (Gs ) p 2. 2 Gs (I + Gs Gs ).11 Let G be the nominal plant and let K = W1K1W2 be the associated controller obtained from loop shaping design procedure in the last section.18. 2 Gs (I + Gs Gs ).1 Gs > 0 M = 2 Gs (I + (1 . Then if " I (I + G K ). 1)(I + Gs Gs )(I + (1 .V (K (K1 ) 1 + M )Y . 1) 2 (Gs ) 1 . we have p 2 .9. ( 2 .1 Gs : Now completing the square in (18.1 : Hence we have which implies .3.1 ) = 1 . it is easy to show that 1=2 2 2 .
1 M .13) (18.15) (18. 2 (Ms ) = 2 (Gs ) : ~ 1 + 2 (Gs ) The proof for the normalized right coprime factorization is similar.1K where .(I + KG).1G .486 we have H1 LOOP SHAPING .1 ~ ~ ~ ~ Ms Ms = I .1 .16) o o 2 (W2 GW1 ) 1=2 (18:17) 1 + 2 (W2 GW1 ) 1=2 1 ~ (18:18) (Ms ) = (Ms ) = 1 + 2 (W2 GW1 ) ~ ~ and (Ns Ms ).1 M .11) (18. is a normalized left coprime factorization. ~ (Ns ) = (Ns ) = Proof. (Ns Ms ).1 ~s s 1 and K1 " I (I + G K ).1 # h i W1 (I + KG).1 h .12) (18.K (I + GK ).1 ~s = s 1 K1 1 = # " " i W2 (I + GK ). Ns Ns : 1 2 (Gs ) All other inequalities 1 2 (Ms ) = max (Ms Ms ) = ~ ~ ~ 1 + max (Gs Gs ) = 1 + 2 (Ns ) = 1 . respectively. right coprime factorization. follow from noting " # I (I + G K ).1 W2 GW1 W1.K (I + GK ).1 ~ (Ms ) (W1 ) (W2 ) n ~ min (Ms ) (W2 ) 1 + (Ns ) n ~ min (Ns ) (W1 ) 1 + (Ms ) ~ (Ns ) (W1 ) (W2 ) n ~ min 1 + (Ns ) (W1 ) (Ms ) n ~ min 1 + (Ms ) (W2 ) (Ns ) o (W2 ) o (W1 ) (W1 ) (W2 ) (18.G(I + KG).1 W2 G 1 # 1 .1 K 1 .(I + GK ).1 . Note that and Then ~ ~ Ms Ms = (I + Gs Gs ).14) (18.(I + GK ).1G .1 W1 GW2. respectively. of Gs = W2 GW1 .
and Vidyasagar 1990]. Zhu 1989]. some design examples were also shown. The H1 loop shaping using normalized coprime factorization was developed by McFarlane and Glover 1990. 1992b] Vinnicombe 1993]. and references therein. Vidyasagar 1984. The method has been applied to the design of scheduled controllers for a VSTOL aircraft in Hyde and Glover 1993]. 1985]. Glover and McFarlane 1989]. In the same references. and G. Notes and References 487 2 This theorem shows that all closedloop objectives are guaranteed to have bounded magnitude and the bounds depend only on . Glover. see ElSakkary 1985]. McFarlane.4 Notes and References .18. Georgiou and Smith 1990]. 18. The robust stabilization of normalized coprime factors is closely related to the robustness in the gap metric and graph topology. Qiu and Davison 1992a. 1992].4. W1 W2 .
488 H1 LOOP SHAPING .
the controllers designed with these methodologies typically have orders comparable to those of the plants. However. However. and then based on the reduced plant model a lower order controller may be designed accordingly. The Lagrange multiplier method developed in the next chapter is potentially useful for some problems.Controller Order Reduction We have shown in the previous chapters that the H1 control theory and synthesis can be used to design robust performance controllers for highly complex uncertain systems. Simple linear controllers are normally preferred over complex linear controllers in control system designs for some obvious reasons: they are easier to understand and computationally less demanding they are also easier to implement and have higher reliability since there are fewer things to go wrong in the hardware or bugs to x in the software. this is still largely an open research problem. On the other hand. high performance controller and subsequently proceed with a reduction of the designed controller. Another approach is to rst reduce the order of a high order plant. A potential problem associated with this approach is that such a lower order controller may not even stabilize the full order plant since the error information between the full order model and the reduced order model is not considered in the design of the controller. Therefore. This approach is usually referred to as controller reduction. A crucial consideration in controller order reduction is to take into account the closedloop so that the closedloop stability is guaranteed and the 489 19 . There are usually three ways in arriving at a lower order controller. A seemingly obvious approach is to design lower order controllers directly based on the high order models. one may seek to design rst a high order. since a great many physical plants are modeled as high order dynamical systems. a lower order controller should be sought whenever the resulting performance degradation is kept within an acceptable magnitude.
490 CONTROLLER ORDER REDUCTION performance degradation is minimized with the reduced order controllers. the approximation error only has to be made small over those critical frequency ranges that a ect the closedloop stability and performance. Unfortunately. this approach has only limited applications. Suppose K is an mth order and G22 = controller which stabilizes the closedloop system.1: Closedloop System Diagram ^ ^ Let K be a reduced order controller and assume for the sake of argument that K has the same number of right half plane poles. 19.1 where the nth order generalized plant G is given by G = G11 G21 " # 2 A B1 B2 3 G12 = 6 4 C1 D11 D12 7 5 G22 C2 D21 D22 is a p q transfer matrix. " A B2 C2 D22 # z G(s) w y . Then it is obvious that the closedloop stabil^ ity is guaranteed and the closedloop performance degradation is limited if K . Since stability is the most basic requirement for a feedback system. One simple reason is that a reduced order controller that stabilizes the closedloop system and gives satisfactory performance does not necessarily ^ make the error (K . The purpose of this chapter is to introduce several controller reduction methods that can guarantee the closedloop stability and possibly the closedloop performance as well.K (s) u Figure 19. Therefore. We are interested in investigating controller reduction methods that can preserve the closedloop stability and minimize the performance degradation of the closedloop systems with reduced order controllers. K 1 is su ciently small. .1 Controller Reduction with Stability Criteria We consider a closedloop system shown in Figure 19. we shall rst derive controller reduction methods that guarantee this property. Hence a trivial controller reduction approach is to apply the model reduction procedure to the full order controller K .K )(j!) su ciently small uniformly over all frequencies.
In view of the above lemma. KG22 = I .1C ~ 5 ~ 6 Ak + Bk D22 R. ^ I . G22 K ). ^ and K have the same number of right half plane poles and de ne Suppose K ^ := K .19. G22 K . KG22 ) so the system is stable if k Wa k1 < 1. K Wa := (I . (I .1.1 G22 : ^ Then the closedloop system with K is stable if either 19.1 Ck B2 R.1 G22 )(I . G22 K ). the controller K should be reduced in such a way so ^ ^ ^ that the weighted error Wa (K .1 ~ where R := I . G22 = (I . Now suppose K has the following state space realization 2 K = Ak Bk : Ck Dk Then " # Wa = = 3" 7 A B2 S 0 Ip G22 5 C2 D22 Iq K Ck Iq Dk 2 3 ~ ~ A + B2 Dk R. G22 K )(I .1 D22 Ck ~ D22 R. ^ Lemma 19.1 7 6 B R. the system is stable if kWa k1 < 1. Since ^ I .1 Ck Bk D22 R.1 Let K be a stabilizing controller and K be a reduced order controller. Dk D22 .1 C2 B2 R.1 C2 R. Hence in general the order of Wa is equal to the sum of the orders of G and K . K ) or (K . Controller Reduction with Stability Criteria 491 The following lemma follows from small gain theorem. On the other hand.1 Additive Reduction kWa k1 < 1 or (19:1) (19:2) k Wa k1 < 1: Proof. G22 = (I .1 G22 ) by small gain theorem. D22 Dk and R := I . K )Wa is small and K and K have 1 1 . G22 K ). G22 K = I .1 7 k 2 4 " # ! 0 Bk 0 Ip 02 A k = S B6 0 @4 #1 C A R.1. KG22 . (I .
^ where K+ is stable and a reduction is done on K+ to obtain a reduced K+ . Since K is unstable.1 where both weighted and unweighted errors are listed. K+ ) 1 we have and ^ K+ = . we need to separate K into stable part and antistable part K = K+ + K. = s . :788) ^ ^ K := K+ + K. ^ reduced order controller is given by K We shall illustrate the above procedure through a simple example. .492 CONTROLLER ORDER REDUCTION the same number of unstable poles. (s + 31:74)(ss+ 31)(s + .3 20 7 B1 = C1 = 6 1 5 4 D11 = 0 D12 = D21 = 0 A controller minimizing kTzw k2 is given by 1 " # 0 0 0 0 3 213 7 B2 = C2 = 6 1 7 5 4 5 1 D22 = 0: 148:79( + 3) K = .1 Consider a system with 2 . K is usually separated into stable and unstable parts as K = K+ + K. 9:19) :85)(s with kTzw k2 = 55:09. if K+ is reduced directly by balanced truncation without stability weighting Wa . The results are summarized in Table 19. (s 151:72(s +s0. Suppose K is unstable. 9:19 : ^ Wa (K+ .34:64 K+ = . the nal ^ = K+ + K. then the reduced order controller does not stabilize the closedloop system. On the other hand. with 3 . 9:19) : + 34:526)( . Example 19. then in order to make sure ^ that K and K have the same number of right half plane poles.1 0 4 3 A=6 0 2 0 4 0 0 . s 117:085 + 34:526 Next apply the frequency weighted balanced model reduction in Chapter 7 to ^ The kTzw k2 with the reduced order controller K is 61:69. = . (s 114:15(s + +:61) + 31:74)(s 3:85) K.
Hence having Wa (K .1622 0. K ) 1 2. Let G22 and K have the following left and right coprime factorizations.1.19.1 N = NM .1461 ^ Wa (K . 3 19. K ) < 1 (or 1 ^ ^ (K . KG ). Then h ~ ~ . G22 K ). WBT: Weighted Balance and Truncate ^ It may seem somewhat strange that the unweighted error K . Controller Reduction with Stability Criteria Methods BT WBT ^ K.e.1 stabilizes the system if ^ ^ Lemma 19.1 G22 I i h i h i and " Nn := N (V M . UN ). We should also point out that the stability criterion Wa (K .2 Coprime Factor Reduction It is clear that the additive controller reduction method in the last section is somewhat restrictive.Nn Mn i " U # " U #! ^ . K has all poles in the right half plane. respectively ~ ~ ~ ~ G22 = M .5295 0. K )Wa < 1) is only su cient. ^ < 1: V V 1 (19:3) . K ) 1 does not 1 1 ^ necessarily imply that K is not a stabilizing controller. ^ ^^ K := U V .1 and de ne h ~ ~ ~ ~ ~ ~ Nn Mn := (MV .1: BT: Balance and Truncate.1 (I .1 = G22 (I . this method can not be used to reduce the controller order if the controller is totally unstable. in particular. i.K 1 0. NU ). This happens because the balanced model reduction is not optimal in L1 ^ norm. This motivates the following coprime factorization reduction approach.69 493 kTzw k2 Table 19.1. K resulted from 1 weighted balanced reduction is actually smaller than that from unweighted balanced reduction.1 : ~ ~ 22 ~ Mn M I # " # " # ~ ~ Note that Mn Nn Mn Nn do not depend upon the speci c coprime factorizations of G22 .1 N M = V .2 Let U V 2 RH1 be the reduced right coprime factors of U and V .471 unstable 61.1 V .1 U = UV ..1 K = V .
4) are small. ^ D2 C D2 D1 (19:5) (19:6) (19:7) (19:8) It is clear from this lemma that the coprime factors of the controller should be reduced so that the weighted errors in (19. B D2 . suppose U and V have the following state space realizations h ^ ^ i " A B1 B2 # ^ ^ ^ ~ ~ U V = C D D ^ ^1 ^2 ^ and suppose D2 is nonsingular. U V i h i " .U # ^ ~ ~ < 1: . Since " h i " U . V ^ 1 2 ^ ^ Now suppose U and V have the following state space realizations ^ ^ "^# 2 A B 3 U =6 ^ ^ 7 4 C1 D1 5 ^ V ^ ^ C2 D2 ^ and suppose D2 is nonsingular.1 ^ ^ U stabilizes the system if ^ ~ K := V ~ h ^ ^ ~ ~ ~ ~ U V . . N U ).1 # C ^ = A . Then . Then the reduced order controller is given by " ^ ^ ^ . The ^ ^^ case for the left coprime factorization is analogous.3) and (19.Nn # < 1: M n 1 (19:4) Proof. ^ .1 ^ # B2 ^ K = A . D1 D2 C2 D1 D2 ^ ^ ^ ^ ^ C ^ ^ ~ ~ Similarly. N U = (MV . It is well known that K := U V .Nn Mn V .^B21D2 D1 : .1 ~^ ~^ stabilizes the system if and only if (M V . We shall only show the results for the right coprime controller reduction.Nn Mn ^ ~^ ~ ~ ~ ~ M ^ V .1 ^ ^ ^ ^ . Then the reduced order controller is given by " ^ ^ ^ .1 2 B D2.1D^2 C B1 .1 : K ^1 .1 ^ ^ ^ . let U V 2 RH1 be the reduced left coprime factors of U and V .1 2 RH1 .V the stability of the closedloop system is guaranteed if h i" U .494 CONTROLLER ORDER REDUCTION ^ ^ ~ ~ ~ ~ Similarly. U ## ^ ~ V . NU ) I . Note that there is no restriction .
. Dk D22 .R.1 Ck . 3 2 ~ ~ B2 R. It is also not hard to see that the additive reduction method in the last subsection may be regarded as a special case of the coprime factor ~ reduction if the controller K is stable by taking V = I or V = I .1 D22 Ck .1 . Note that the orders of the weighting h i " .L F 0 letting Fk = . respectively. However.B2R. De ne 3 2 i " A + LC2 B2 + LD22 L # " N # 6 A + B2 F B2 7 ~ ~ = 4 C2 + D22 F D22 5 N M = M C2 D22 I F I " # 2 Ak + Bk Fk Bk 3 h " # U =6 7 U V i = Ak + Lk Ck Bk + Lk Dk Lk : 4 Ck + Dk Fk Dk 5 ~ ~ V Ck Dk I h Fk I ~ ~ ~ ~ Then K = UV .1 D22 R.L .B2 R.Bk R.R.1C2 ~ ~ Ak + Bk R. NU = I V M .(C2 + D22 F ) and Lk = .1 7 . if K is an observerbased controller.1 N are right and left coprime factorizations over RH1 .R. K = A + B2 F + LC2 + LD22 F .Lk + Bk D22 R. UN = I: .Nn = 6 .L 3 h U =6 7 U V i = A + LC2 .1 C2 .1 = V . we get " # " " # # 2 A + B2 F . D22R.1 D22 Ck .(B2 + LD22).Bk R.1 Ck 6 7 6 . Controller Reduction with Stability Criteria 495 ^ on the right half plane poles of the controller. Let L F Lk and Fk be any matrices such that A + LC2 A + B2 F Ak + Lk Ck and Ak + Bk Fk are stable.1Ck 7 ~ ~ Mn R.1 C2 5 . K and K may have di erent number of right half plane poles.1 " # 6 7 ~ ~ ~ 7 . D22 Dk and R := I .1 A + B2 R.1 R.1 C2 R. Fk .1 = M .(B2 + LD22) ~ ~ 0 5 4 F V F 0 I C2 + D22 F I ~ ~ ~ ~ MV .B2 R.1 Ck .1 4 5 ~ ~ ~ R .Nn Mn = 4 . i.1 Dk C2 .D22 R.1 U and G22 = NM .19.B2 Dk R.1 2 ~ ~ 3 Ak + Bk D22 R.Nn Mn and are in general equal to the sum of the orders of Mn G and K .1.Nn # ~ ~ matrices .1 D22 Bk R.Bk R.1 Dk C2 ~ where R := I . Moreover.1 Ck h i 6 A + B2 Dk R.1 .e. In particular.1C2 .
1 0 4 .10:198 . Example 19.2 Consider a system with 2 . 5:3414) 6 0 .1 .8:198 .2 0 7 B1 = C1 = 6 1 0 7 B2 = C2 = 6 1 7 4 5 4 5 4 5 0 0 .8:198 .18:396 .10:198 0 8:198 7 6 7 U =6 6 0 .8:198 4 0 3 " # 6 0 .1 0 7 5 4 1 1 1 0 0 Applying the frequency weighted balanced model reduction in Chapter 7 to the weighted coprime factors.1 .8:198 0 0 7 7 4 5 1 1 1 1 Furthermore. we can chose .0:1814 1:0202 3 U =6 4 1:2244 0 7 5 ^ V 1 6:504 3 2 . we have 3 2 .8:198 .3 . Since the controller is an observerbased controller.3 0 5 4 0 .496 CONTROLLER ORDER REDUCTION i h i .Nn = .N 6 0 0 .3 0 0 1 D11 = 0 D12 = D21 = 0 A controller minimizing kTzw k2 is given by 1 h " # " # " # D22 = 0: 0 0 with kTzw k2 = 37:02.8:198 8:198 7 7=.8:198 4 0 6 .1 .N M and Mn M which have the same orders as that of the plant G.1 0 4 3 20 03 213 A = 6 0 .3 0 7 : 7 6 V 6 0 .N ~ ~ ~ ~ Therefore.1 0 7 h i 6 7 6 ~ M = 6 .8:198 7 : ~ . we obtain a rst order approximation " ^ # 2 . We shall also illustrate the above procedure through a simple example. a natural coprime factorization of K is given by 2 .8:198 . 6 67:2078(s + 1)(s + 3) 7 K =6 7 (s + 23:969)(s + 3:7685)(s .Nn Mn = .8:198 .8:198 .
the su cient stability condition is not satis ed nevertheless the controller is a stabilizing controller.4:1434 .Nn Mn V V and a second order approximation 1 ^ K = s +:2491 kTzw k2 = 52:29 6:8165 1 2 3 . j!I B2 C1 D12 # has full column rank for all ! . It is also interesting to note that the unstable pole of the second order reduced controller is not at the same location as that of the full order controller. Again we consider the feedback system shown in Figure 19.1 with a generalized plant realization given by 2 A B B 3 1 2 G(s) = 6 C1 D11 D12 7 : 4 5 C2 D21 D22 The following assumptions are made: (A1) (A B2 ) is stabilizable and (C2 A) is detectable (A2) D12 has full column rank and D21 has full row rank (A3) " A .1:2642 7 6 6 1:2244 29:5182 0 7 7 ^ V 4 5 6:504 .1:3084 1 which gives h 36:069(s + :1102) ^ K = .2 H1 Controller Reductions In this section. ^ . H1 Controller Reductions which gives 497 = 2:0928 > 1 h i " U # " U #! ^ ~ ~ .0:5288 . Moreover. ^ . 3 i " U # " U #! ^ ~ ~ . 4:7601) kTzw k2 = 39:14: Note that the rst order reduced controller does not have the same number of right half plane poles as that of the full order controller. (s + 17:3741)(s1.Nn Mn V V 1 = 0:024 < 1 19.2. we consider H1 performance preserving controller order reduction problem.19.0:1814 14:5511 1:0202 7 "^# 6 U = 6 .
. ^ ^ ^ ^. N22 Q). i. ^ such that D12 and D21 are invertible and A .1 N21 : Assume that the feedback loop is wellposed.e. j!I B1 C2 D21 # CONTROLLER ORDER REDUCTION has full row rank for all !. i. M12 21 ^ The problem to be considered here is to nd a controller K with a minimal possible ^ order such that the H1 performance requirement F`(G K ) < is satis ed. the closedloop transfer matrix from w to z is given by # Tzw = F` (N Q) = N11 + N12 Q(I . B2 D121 C1 and A . . we shall approach this problem from a di erent perspective. The following lemma is useful in the subsequent development and can be regarded as a special case of Theorem 11. B1 D211 C2 are both . and either N21 (j!) has full row rank for all ! 2 R 1 or N12 (j!) has full column rank for all ! 2 R 1 and kN k1 1 then kF` (N Q)k1 < 1 if kQk1 < 1. It is shown in Chapter 17 that all stabilizing controllers satisfying kTzw k1 < can be parameterized as K = F`(M1 Q) Q 2 RH1 kQk1 < (19:9) where M1 is of the form M1 = " ^ ^ ^ # 2 A B1 B2 3 M11 (s) M12 (s) = 6 ^ ^ ^ 7 4 C1 D11 D12 5 M21 (s) M22 (s) ^ ^ ^ C D D 2 21 22 ^ ^ ^ ^ ^. stable. Instead of choosing Q directly. Lemma 19. This 1 is clearly equivalent to nding a Q so that it satis es the above constraint and the order ^ of K is minimized.498 (A4) " A . N22 (1)Q(1)) 6= 0.3 Consider a feedback system shown below z N w y " Q u where N is a suitably partitioned transfer matrix N (s) = N11 N12 : N21 N22 Then.1 and M .1 are both stable..7 (main loop theorem).e. det(I .
H1 Controller Reductions 499 Proof. It follows from simple algebra that such that K . it follows from straightforward manipulations that .2. We shall assume N21 has full row rank. consider the closedloop system at any frequency s = j! with the signals xed as complex constant vectors. M11 0 I : a I 0 I 0 Furthermore. i. 2 ) . Also let := kTwy k1 . with stable. and kN k1 1 implies that kz k2 + kyk2 kwk2 + kuk2. Then kwk2 kyk2. Suppose that kF` (G K0 )k1 < .2 ]kwk2 kz k2 2 2 2 which implies kTzw k1 < 1 . thus to a certain degree limiting the set of attainable reduced order controllers. 1)kyk2 1 . and invertible weighting functions W1 and W2 .1 Additive Reduction Consider the class of (reduced order) controllers that can be represented in the form ^ K = K0 + W2 W1 where K0 may be interpreted as a nominal. minimum phase. N22 Q) where N21 is a right inverse of N21 . there is a Q 2 RH1 with kQk1 < ^ = F` (M1 Q).1 := " # " # where ~ R= " .19. Note that this is somewhat a special case of the above valid when K ^ ^ general problem the speci c form of K restricts that K and K0 must possess the same right half plane poles. (1 . A natural question is whether it is possible to obtain ^ a reduced order controller K in this class such that the H1 performance bound remains ^ is in place of K0 . ^ Q = F`(Ka 1 K ) where 0 I . is a stable perturbation. To show that kTzw k1 < 1. ^ kQk1 < () F` (Ka 1 K ) 1 < . ^ Suppose K is a suboptimal H1 controller.e.. Therefore.1=2 I 0 0 W1 #" 1 R11 R12 R21 R22 #" . higher order controller. The case when N12 has full column rank can be shown in the same way.1=2 I 0 0 W2 # . 2 2 2 2 2 kwk2 + ( 2 . 2 19.2. Let kQk1 =: < 1 and note that + + Twy = N21 (I . () F` (Ka 1 K0 + W2 W1 ) 1 < ~ () F` (R ) < 1 K .
1 (K . It is obvious that ^ ^ to make W2. This guarantees the existence of a Q such that kQk1 < . can be used to generate a required reduced order controller which will preserve the ^ closedloop H1 performance bound F`(G K ) < .1 (K . 1 . Then K is also a stabilizing controller such that F` (G K ) 1 < if ^ k k1 = W2.1 a R21 R22 # " Ko I ): I 0 # ~ ~ It is easy to see that R12 and R21 are both minimum phase and invertible.R22 0 3 7 7 7 7 5 .R .4 Suppose W1 and W2 are stable. This observation leads 1 to the following theorem. or ^ ^ equivalently.1 < 1 for some K . minimum phase and invertible transfer ~ matrices such that R is a contraction. ~ by invoking Lemma 19. K0 )W1.5 Assume kR22 k1 < and de ne 2 0 " # 6 . Theorem 19. The procedure below. the existence of a K such that F` (G K ) < . and hence have full column and full row rank. Consequently. Let K0 be a stabilizing controller such that ^ ^ kF`(G K0 )k1 < . See Goddard and Glover 1993].1 1 < 1: ~ Since R can always be made contractive for su ciently small W1 and W2 . we conclude that if R is a contraction and k k1 < 1 then ~ F` (R ) 1 < 1. Lemma 19.1 I ): ~ Then R is a contraction if W1 and W2 satisfy " (W1 W1 ). there are in nite many W1 and W2 that satisfy the conditions in the theorem. K0 )W1. 2 An algorithm that maximizes det(W1 W1 ) det(W2 W2 ) has been developed by Goddard and Glover 1993].R11 R0 R012 6 L = L1 L2 = F`(6 11 0 6 0 R21 021 .3.1 # " L1 L2 : L2 L 3 # Proof.R22 L2 L3 4 R12 0 . devised directly from the above theorem.1 0 0 (W2 W2 ). one would like to select the \largest" 1 W1 and W2 .500 and R is given by the star product CONTROLLER ORDER REDUCTION " R11 R12 = S (K . respectively for all ! 2 R 1.
^ ^ ^.2.6 The family of all admissible controllers such that kTzw k1 < can also K (s) = F` (M1 Q) = ( 11 Q + 12 )( 21 Q + 22 ). and UV .1 := UV .1 7 = 4 ^1 ^ 11 ^ 21 ^2 ^ 12 ^ 11 ^ 21 ^ 22 ^ 11 ^ 21 5 21 22 ^. D11 D211 D22 Proof. ^ ^.1 and V . we need the following alternative representation of all admissible H1 controllers. ^ ^. K0 )W1. # 2 A . B D.19.1 501 1 < 1. Let K0 be a full order controller such that kF`(G K0 )k1 < ~ 2.D211C2 .D22D121 7 4 ^ ^ ^. ^ ^. ^ ^ ^. H1 Controller Reductions 1. B D. . D121 2 A . B1D211D22 B1D211 3 11 12 = 6 C . D D.1 = 6 .1 (Q ~ 11 + ~ 21) := V .1 U ~ ~ where Q 2 RH1 .1 C D . ^ ^ ^. Compute W1 and W2 so that R is a contraction ^ ^ 3.1 = 6 ^ ^ ^. Note that all admissible H1 controllers are given by K (s) = F`(M1 Q) . B2 D121C1 B1 .2 Coprime Factor Reduction The H1 controller reduction problem can also be considered in the coprime factor framework. ^ D21 D211 D22 5 4 D21 ^ ^ ^ ^.D12 ^ 5 ^ ^ ^. ^ " # 2 A .1 ~ .D11 D211 D12 .1 B . Using model reduction method to nd a K so that W2. B D. ^ ^ ^. D22 D121 C1 D21 . B2 D121D11 . ^ C1 . ^ D121 C1 ^.1D 3 ^1 ^ 21 ^2 ^1 ^ 21 ^ 22 ^ ^1 ^ 21 ^2 7: . be written as Lemma 19. ^ C2 .D121 D11 4 .1 C2 . ^ D12 .B D.1C B D. B1D211C2 B2 . ^ ^ ^. For that purpose. 5 = 6 C2 . D11 D211 C2 .1D 3 ^ ^2 ^ 12 ^1 ^2 ^ 12 ^1 ^2 ^ 12 ^ 11 7 . ^ D121 D11 ^.1 ~ ~ = (Q ~ 12 + ~ 22 ).1C . ^ ^ ^. D D.1 C1 .B2D121 3 ^ ^.1 D D D. and ^ ^ ^. ^ ^ ^ ^.2. D22 D121 C1 D22 D121 D21 . kQk1 < .1 (K . B D.1 B . D22 D121 D11 2 A . ^ ^ ^ ^.1 ^ ^ ^.D211 D22 D211 ~= " ~ 11 ~ 12 ~ 21 ~ 22 ^ ^ ^. D22 D121 D11 . ^ ^ ^.1 U are respectively right and left coprime factorizations over RH1 . 19.
M12 . M12 . M11 M211 M22 " 11 21 # " # " . .1 # . to show that K = U V . we need to show that there is ^ another coprime factorization for K = UV .10) is also ^ a stabilizing controller such that kF`(G K )k1 < . . M22 M121 M21 .M121 M11 .1 .1 12 . . M11 M211 M22 M11 M211 = .1 I 0 := 0 I # . . . Note that by Lemma 19.502 where M1 = Lemma 10. we have = " M11 M12 M21 M22 # 12 22 CONTROLLER ORDER REDUCTION is given in the last subsection.6.M22 M121 ~= = .1 with U and V satisfying equation (19.M21 M21 # # ~ 11 ~ 12 ~ 21 ~ 22 2 Theorem 19.1 and a Q 2 RH1 with kQk1 < such that equation (19. Next note that from " Then the results follow from some algebra. K is an admissible controller such that kTzw k1 < if and only if there exists a Q 2 RH1 with kQk1 < such that " U V # " := 11 Q + 12 21 Q + 22 # = " Q I # (19:11) and K = UV .1 = .7 Let K0 = 12 . . U (19:10) < 1= 2: ^ 0 I V 22 ^ ^^ ^ Then K = U V . De ne " . 1 Proof. M21 . .1 = : . . M22 M121 M11 " # . M211 M211M22 ~ .1 be a central H1 controller such that kF`(G K0)k1 < 22 ^ ^ and let U V 2 RH1 be such that " .2.1 : ^ ^^ ^ ^ Hence.1 " 12 22 # " ^ . M22 M121M11 .1 # " # " ^ #! p I 0 .1M22 . U ^ V #! .M11M211 M12 .11) is satis ed.1 is also a stabilizing controller such that kF`(G K )k1 < . M121 M11 M121 " .
The above two theorems show that the H1 controller reduction problem is equivalent to a frequency weighted H1 model reduction problem. or equivalently U (I .1 : ^ V (I . 0 I h i 0 I 0 p h p I= 2 0 I= 2 i h i 3 1 p A i5 2 . Now U (I .1 ^ ^ ^^ is another coprime factorization for K .1 and Q := . U ~ . V ). V := V (I . V ). To show that K = UV .1 is a stabilizing ^ )k1 < .1 < 1= 2: ^ ~ 22 ~ 0 I 1 V ^ ^ .1 U (I . 2 Similarly.1 1 < 1 since 2 1 < 1. we need to show that controller such that kF`(G K U (I .1 1 < 1.2. Then UV . V ).1 .1 1 < .1 ^ ~ ~ ^ Then K = V U is also a stabilizing controller such that kF`(G K )k1 < . V ). U (I .8 Let K0 = ~ . V ).1 ~ 21 be a central H1 controller such that kF`(G K0)k1 < 22 ^ V 2 RH1 be such that ~ ^ ~ and let U " ~ # " ^ #! " . Theorem 19. is a contraction and p h i 3 2 0 h I 0 i5 4 p p I= 2 0 I= 2 H1 Controller Reduction Procedures . I. " U V # : = # I 0 0 I # " " . H1 Controller Reductions and partition Then and as := 503 " " ^ U ^ V # " = ^ U (I . V ).3 02 F` @4 I 0 I. # U V # h = and by Lemma 19.1 # ~ p I 0 21 .1 I ^ ^ De ne U := U (I . V ). we have the following theorem.1 = " 12 22 # .1 .1 = .1 = U V . V ). V ). U (I .19. V ).
with minimal state space real1 1 izations W1 (s) = A1w B1w C1w D1w " # W2 (s) = A2w B2w C2w D2w " # ^ ^ A B ^ and let G(s) 2 RH1 . i. Q H inf and assume " # " A1w .e. I B1w C1w D1w # " A2w .504 CONTROLLER ORDER REDUCTION ^ .1 # " # " ^ #! p I 0 . U ^ ~ 22 ~ V #! ~ .3 FrequencyWeighted L1 Norm Approximations We have shown in the previous sections that controller reduction problems are equivalent to frequency weighted model reduction problems.1 ~ 21) be a suboptimal H1 central controller (Q = 0) such 22 22 that kTzw k1 < .1 I 0 # 1 0 I 1 < 1= 2: p ^ (iii) The closedloop system with the reduced order controller K is stable and the performance is maintained with the reduced order controller. Suppose that G1 (s) = ^1 ^ 1 2 RH1 is an rth order C1 D1 optimal Hankel norm approximation of W1 GW2 ]+ . or " # "^ ~ 21 ~ . I B2w C2w D2w # . U < 1= 2 ^ 0 I V 22 (i) Let K0 = 12 .. ^ kTzw k1 = F`(G K ) 1 < : 19.9 Let W1(s) 2 RH. In this section. To that end. the frequency weighted balanced model reduction approach in Chapter 7 can be applied. Theorem 19. ^ G1 = arg degQ r W1 GW2 ]+ .1 12 .1 " . we propose another method based on the frequency weighted Hankel norm approximation method..1 (or V U ) such that the following frequency weighted H1 error " .e. i.1(= ~ . and W2(s) 2 RH.1 ^ ^ ^^ ~ ~ (ii) Find a reduced order controller K = U V .
13) can be rewritten as " A1w . G)W2 = kW1 (G . Gr := Y 0 imation. i I B1w C1w D1w #" Xi = 0 ^ Yi C1i # " # i = 1 2 : : : r: By the assumption. Assume without loss of generality that A1 has a diagonal form ^ A1 = diag 1 2 : : : r ]: ^ (The proof can be easily modi ed if A1 has a general Jordan canonical form). Partition ^ X . Y .13) ^1 Q + ZC2w = 0 QA2w .12) ^1 C1w X + D1w Y = C (19. X A + B Y .e. ^ inf W1 (G .13). Gr )W2 kH = r+1 ( W1 GW2 ]+ ) : ^ H degG r Proof. Let " # ^ A1w .3.12) and (19.15) # " ^ A1 Z is the frequency weighted optimal Hankel norm approxFurthermore..19. A (19.X B ^ ^1 1 1 1 6 0w Aw1 B1 7 = 6 0w 1w A1 1 1w ^ ^ 5 4 ^ ^1 W1 Gr = 4 B C1w D1w Y 0 C1w C1w X + D1w Y 0 3 7 5 . and Z such that ^ A1w X . X A1 + B1w Y = 0 (19.12) and (19. i.D1 Then using equations (19. i I B1w C1w D1w # has full row rank for all i and thus the existence of X and Y is guaranteed. Then there exist matrices X Y Q.14) ^ QB2w + ZD2w = B1 : (19. FrequencyWeighted L1 Norm Approximations 505 ^ have respectively full row rank and full column rank for all = i (A1 ) i = 1 : : : r. and C1 as ^ ^ ^ ^ X = X1 X2 : : : Xr ] Y = Y1 Y2 : : : Yr ] C1 = C11 C12 : : : C1r ]: Then the equations (19.X B1 2 RH. we get 2 A B Y 0 3 2 A A X . : ^ W1 = 1 ^ C1w . The general case can be proven ^ similarly. We shall assume W2 = I for simplicity. the matrix " A1w .
Gr )kH : r+1 ( W1 G]+ ) ^ H degG r Using this expression. G) kW1 (G . we have kW1 (G . G . Then 1 1 ^ ^ inf W1 (G . These conditions will of course be satis ed automatically if W1 (s) and W2 (s) have all zeros in the right half plane. To make the 1norm approximation error as small as possible. then choosing W1 (s) and W2 (s) to have all poles and zeros in the right half plane may reduce the L1 norm approximation error signi cantly. This Dr can 1 usually be obtained using any standard convex optimization algorithm. Dr )W2 is made as small as possible. G1 H = W1 G]+ . G)W2 inf W (G . (i) Let # " ^ A1 Z ^ G1 = Y 0 be a weighted optimal Hankel norm approximation of G. To further reduce the approximation error. G1 H ^ inf W1 (G . Gr )kH = = 2 ^ ^ Note that Y = C1 if W1 = I . Numerical experience shows that if the weighted Hankel approximation is used to obtain a L1 norm approximation. G)W2 = r+1 ( W1 GW2 ]+ ) : ^ ^ 1 degG r 1 H degG r The lower bound in the above corollary is not necessarily achievable. Let W1 and W2 be any antistable transfer matrices with all zeros in the right half plane. Z = B1 if W2 = I . W2 (s) 2 RH. and G(s) 2 RH1. W1 .10 Let W1 (s) 2 RH.X B 3 ^ 1 6 0w A1 B1 1 7 = W1 + G1: ^ ^ 5 ^ ^ 4 ^ C1w C1 0 ^ ^ ^ W1 G]+ + W1 G]. and the rank conditions in the ^ above theorem are actually equivalent to the statements that the poles of G1 are not zeros of W1 and W2 . ^ (ii) Let the reduced order model G be parameterized as " ^ # " ^ # ^ ( ) = A1 B ^ ( ) = A1 Z : G or G Weighted L1 Model Reduction Procedures Y D C D . . a suitable constant matrix Dr should ^ be chosen so that W1 (G . the following optimization is suggested.506 = CONTROLLER ORDER REDUCTION 2 A 0 . Corollary 19.
Then there is a M (s) = A CW (DD )1=2 and where and Then we have W2 (s)W2 (s) = M (s)M (s) BW = PC + BD CW = (DD ). . G)W2 1 = W1 (G . BW X ) (DD ). G)W2 1 = W1 (G . For example.s). G)W2 1 : In the case where W1 and W2 are not necessarily stable.3. the following procedures can be applied to accomplish this task.19. (i) Let W1n := W1 (. BW X ) PA + AP + BB = 0 XA + A X + (C .1 (C .28 if W1 and W2 are stable and W1 (1) and W"(1) have respectively full column rank and full row rank.1 (s) 2 RH1 !. take W2 (s) = M (s). factorizations must be performed rst to obtain the equivalent W1 (s) and W2 (s) so that W1 (s) and W2 (s) have all poles and zeros in the right half plane and W1 (s)W1 (s) = W1 (s)W1 (s) W2 (s)W2 (s) = W2 (s)W2 (s) ^ ^ W1 (G . Then W2 (s) has all the poles and zeros in the right half plane and ^ ^ W1 (G . FrequencyWeighted L1 Norm Approximations (iii) Find C (or B ) and D from the following convex optimization: ^ min W1 (G .1=2 (C .s) and W2n := W2n(. If the given W1 and W2 do not have all poles and zeros in the right half plane. 2 # A B 2 RH with D full row rank and W (j!)W (j!) > 0 for all assume W2 = 1 2 2 C D " # BW 2 RH1 such that M . BW X ) = 0: Finally. Spectral Factorization Procedures Let W1 2 L1 and W2 2 L1 . G)W2 1 : These factorizations can be easily done using Corollary 13. G( ))W2 : 2Rm 1 507 It is noted that the weighted Hankel singular values can be used to predict the approximation error and hence to determine the order of the reduced model as in the unweighted Hankel approximation problem although we do not have an explicit L1 norm error bound in the weighted case.
1 .1 2 RH. and 1 B+YC D : W = . BR.1B X = Ric . j! B has full row rank for all (b) Suppose W (j!)W (j!) > 0 for all ! or C D !. DR 1 D )C (A . (This step can be done using Theorem 13.1 R (D C . BR . j! B has full column rank for (a) Suppose W (j!)W (j!) > 0 for all ! or C D all !.C (I . " # A .508 CONTROLLER ORDER REDUCTION factorizations such that M1 and M2 are inners.1 may be used to factor out the undesirable poles and zeros in the weights. Let " # Y = Ric .A .11 Let W (s) = A B 2 L1 be a controllable and observable realizaC D tion. Then we have the following spectral factorization .BB X = Ric 0 0 A " # " # .1N1 and W2n = N2M2. Otherwise.A .) (iii) Perform the following spectral factorizations (ii) Let W1n = M1.C C 0 0 A with R := D D > 0.) (iv) Let W1 (s) = V1 (. B X )(I + Y X ) R1=2 A .(A .s) and W2 (s) = V2 (. (Corollary 13.23 may be used here if N1 (1) has full column rank and N2 (1) has full row rank.BR. Let " # . the factorization in Section 6.1 D C ) " # " # 0 W W =W W where W W .1=2 A + Y C C .34.s).D C ) . Lemma 19.1 be respectively the left and right coprime N1 N1 = V1 V1 N2 N2 = V2 V2 so that V1 and V2 have all zeros in the left half plane. We shall summarize the state space formulas for the above factorizations as a lemma.
BD R.6 as Ko = 12 . and 1 .19. D)B (A .1 C ) " # 0 WW = W W where W W .C R.3 2:3196 10.(A .1 2 RH.1 ~.2 where the following abbreviations are made: h 0 0 0 0 0 h i h B1 = pq2 B2 0 H = 0 0 0 0 0:55 11 1:32 18 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 3 7 7 7 7 7 7 7 7 7 7 7 7 5 213 607 6 7 607 6 7 6 7 6 7 B2 = 6 0 7 607 6 7 607 6 7 607 4 5 i 0 i .3:982 6 1 0 0 0 0 0 6 6 0 6 1 0 0 0 0 6 6 0 1 0 0 0 A=6 0 6 0 0 0 1 0 0 6 6 0 0 0 0 1 0 6 6 0 4 0 0 0 0 1 0 C2 = 0 0 6:4432 10.6 q2 = 1 and 2 .1 C Y = Ric .0:161 . An Example 509 ~ with R := DD > 0.1 ~ . We shall set up the dynamical system in the standard linear fractional transformation form x = Ax + B1 w#+ B2" # _ u "p q1 H x + 0 u z = 0 I h i y = C2 x + 0 I w where q1 = 1 10. Y C )R ~ R1 C + DB X " # 19. D R.9:9835 .1 = ~ .4.0:40727 .2 1:0002 0:10455 0:99551 : The optimal H1 norm for Tzw is opt = 1:1272. BD ~R 1 C ) ~ . The controller is reduced using several 22 22 methods and the results are listed in Table 19.1=2 : W = A + BB X (I + Y X ) (BD=2.4 An Example We consider a fourdisk control system studied by Enns 1984].0:58215 . Then we have the following spectral factorization ~ .B (I . We choose = 1:2 to compute an 8th order suboptimal controller Ko . The coprime factorizations of Ko are obtained using Lemma 19.6:004 .1 ~ 21 .3 7:1252 10.
K 1 UWRCF Unweighted right coprime factor reduction: " 12 22 # " ^ . K ) 1 . U ^ V # 1 UWLCF Unweighted left coprime factor reduction: h ^ ^ ~ 21 ~ 22 . U V ~ ~ i h i 1 SWA Stability weighted additive reduction: ^ Wa (Ko .510 UWA CONTROLLER ORDER REDUCTION Unweighted additive reduction: ^ Ko .
1 I 0 # 0 I . U V ~ ~ i h i " . K )W1. Table 19.1 " .1 1 PWRCF Performance weighted right coprime factor reduction: " .Nn Mn i " 12 22 # " ^ . U ^ V #! 1 Stability weighted left coprime factor reduction: h PWA ^ ^ ~ 12 ~ 22 .10. is no greater than r. U ^ V #! 1 PWLCF Performance weighted left coprime factor reduction: h B H/O ^ ^ ~ 12 ~ 22 .1 (Ko .1I 0 # 0 I 1 Balance reduction with or without weighting Hankel/convex optimization reduction with or without weighting Table 19. U : := ^ 0 I V 22 1 By Corollary 19.3 lists the performance weighted right coprime factor reduction errors and their lower bounds obtained using Corollary 19.4. Similarly.10. U V ~ ~ i h i ~ . An Example SWRCF Stability weighted right coprime factor reduction: 511 h SWLCF ~ ~ .Nn # Mn 1 Performance weighted additive reduction: ^ W2.1 " 12 22 # " ^ . The in Table 19.19.1 # " # " ^ #! I 0 .3 is de ned as " .1 12 .4 lists the stability weighted right coprime factor reduction errors ^ U r+1 if the McMillan degree of ^ V " # .
197 1.98 1.612 48.199 1.981 U U U U U U 1.036 1.196 1.985 5.196 1.184 27.975 3. their performances are very hard to predict.197 1.341 1.198 U U 1.197 1.27 2. What is the worst is that the small approximation error for the .91 1.199 1. Although some unweighted reduction methods and some stability weighted reduction methods do give reasonable results in some cases. Table 19.47 4.218 5.196 1. U u := ^ V 1 22 ^ ^ where u is obtained by taking the same U and V as in s .196 1.197 1.586 2.196 1.207 1. The s and u in Table 19.2 shows that performance weighted controller reduction methods work very well.195 1.273 1.196 1.674 U U 2.259 4 1.201 1. not from the unweighted model reduction.94 U U 2.031 1.236 2.195 1.622 3.375 1.196 1.527 U U U U U U 4.197 1.893 1.5 U U U 1.197 1.321 U 1.199 1.251 U ^ Table 19.467 U 13.503 1.217 1.965 3.2 1.196 1.417 1. In particular.04 2. ^ .197 1.488 15.282 27.198 1.388 3.849 3 2 1 0 U 4.401 1. the PWRCF and PWLCF are easy to use and e ective and there is in general no preference of using either the right coprime method or left coprime method.Nn Mn s := V 22 1 " # "^# 12 .197 U 23.059 U 1.501 U U U U U U U U U U U U U U U U U U U U U U U U U 23.196 U 1.99 U U 2.2: F`(G K ) with reduced order controller: U{closedloop system is unstable and their lower bounds obtained using Corollary 19.12 2.802 1.251 1.327 1.258 U 1.3 are de ned as h i " 12 # " U #! ^ ~ ~ .73 1.197 1.198 1.10.267 6 1.199 2.04 U 2.415 U U 2.198 U U 1.15 1.512 ^ Order of K PWA PWRCF PWLCF UWA UWRCF UWLCF SWA SWRCF SWLCF B H/O B H/O B H/O B H/O B H/O B H/O B H/O B H/O B H/O 7 1.485 CONTROLLER ORDER REDUCTION 5 1.199 1.
The L1 model reduction procedures discussed in this chapter are due to Zhou 1993]. however. Table 19.635 0.702 B 1.5. For example. Khargonekar and Doyle 1987] have also proposed another H1 controller reduction method with guaranteed performance for a class of H1 problems.385 0. The main results presented in this chapter are based on the work of Goddard and Glover 1993. the H1 performance is very far away from the desired level.687 0. . Although the unweighted right coprime factor reduction method gives very good results for this example. Notes and References 513 reduction criterion may have little relevance to the closedloop H1 performance. Hung and Glover 1986]. The weighted and unweighted coprime factor controller reduction methods are proposed by Liu and Anderson 1986.59 127.405 0.960 1 1 1 Table 19. i.. We should note that a satisfactory solution to the general frequency weighted L1 norm model reduction problem remains unavailable and this problem has a crucial implication toward controller reduction with preserving closedloop H1 performance as its objective.295 0. Lenz. and Ly 1990]. Normalized coprime factors have been used in McFarlane and Glover 1990] for controller order reductions. ^ Order of K Lower bounds 7 6 5 4 3 2 1 0 r+1 0. Of course.5 Notes and References The stability oriented controller reduction criterion is rst proposed by Enns 1984]. Anderson. and Zhou 1993].029 H/O 0.645 0.323 0. For normalized H1 controller. one should not be led to conclude that the unweighted right coprime factor method will work well in general.4 shows that the 7th order weighted approximation error s using the Hankel/Optimization method is very small.1994].009 0. If this is true.3: PWRCF: and lower bounds r 19.668 0. The only reason for the good performance of the right coprime factor method for this example is the special data structure of this example. 1990]. Mustafa and Glover 1991] have proposed a controller reduction method with a prior performance bounds. and Anderson 1993].626 4. Anderson and Liu 1989].750 71. then one can easily conclude that the unweighted left coprime factor method will do equally well by considering a dual problem.2 shows that this is not true because the unweighted left coprime factor method does not give good results.303 0.8 6.658 0.e.389 0. Table 19.295 0.2 2. Liu. the relationship between the B1 matrix and B2 matrix: h i B1 = pq2 B2 0 : The interested reader may want to explore this special structure further.19. The frequency weighted Hankel norm approximation is considered in Latham and Anderson 1986].
791 1.3 384.444 185.2092 0.3755 254.6 1:1 10 1:9 10.0001 0.45 3000.5 1:66 10.8421 0.6 1:9 10.01 21.4 2:145 10.5048 2.5439 0.277 397.9 288.28 9.7642 1 1 2600.3 15.5341 0.4: SWRCF: s and the corresponding u .1461 1.3182 0.3 Table 19.4 11.2 365.1 384.9 30.6 0.85 305.6 6:23 10.7018 910.5403 0.514 CONTROLLER ORDER REDUCTION ^ Order of K Lower bounds of s s u ^ Order of K Lower bounds of s s u B H/O B H/O B H/O B H/O r+1 r+1 r r 7 6 5 4 .6 1:2 10.6 383.3164 9.5473 0.38 3 2 1 0 9 10.
the controller may be limited to be a state feedback or a constant output feedback or a xed order dynamic controller. 20 20. we will try to be concise but clear. The results quoted below are standard and can be found in any references given at the end of the chapter. The fundamental idea is to formulate our optimal control problems as some constrained minimization problems. Readers should keep in mind that our purpose here is to introduce the method. we will review some Lagrange multiplier optimization methods. A point x0 2 Rn in S is said to be a (global) minimum point of f on S if f (x) f (x0 ) for all points x 2 S . We shall be interested in deriving some explicit necessary conditions that an optimal xed structure controller ought to satisfy. Then the rstorder Lagrange multiplier necessary conditions for optimality are applied to derive our optimal controller formulae. Then these tools will be used in section 20.Structure Fixed Controllers In this chapter we focus on the problem of designing optimal controllers with controller structures restricted for instance. A point x0 2 S is said to be a local minimum point of f on S if there is a neighborhood N of x0 such that f (x) f (x0 ) for all points x 2 N . not to try to solve as many problems as possible. 515 . In section 20. Hence.1.2 to solve a xed order H2 optimal controller problem. Let f (x) := f (x1 x2 : : : xn ) 2 R be a real valued function de ned on a set S Rn . we consider the constrained minimization problem.1 Lagrange Multiplier Method In this section.
1 A point x0 2 Rn satisfying the constraints H (x0 ) = 0 is said to be a regular point of the constraints if rH (x0 ) has full column rank (m) equivalently. . let (x) := H (x)z for z 2 Rm . rF (x0 ) = rf (x0 ) + rH (x0 ) = 0: In the case where the regular point conditions are either not satis ed or hard to verify. 3 7 7: 7 7 5 De nition 20.. Then there exists a unique multiplier 2 6 6 =6 6 4 1 2 . if we set F (x) = f (x) + H (x) . .e. @hm @x1 @hm @x2 @hm @xn . Theorem 20. . hi (x) = 0 i = 1 2 : : : m and m < n or equivalently H (x) := h1 (x) h2 (x) : : : hm (x) = 0: Hence we will focus on the local necessary (and su cient) conditions for the following problem: ( minimize f (x) (20:1) subject to H (x) = 0: We shall assume that f (x) and hi (x) are di erentiable and denote h i 2 6 6 rf (x) := 6 6 4 @f @x1 @f @x2 @f @xn . 3 7 7 7 7 5 rH (x) := rh1 (x) rh2 (x) h 2 i 6 6 rhm (x) = 6 6 4 @h1 @x1 @h1 @x2 @h1 @xn . . .516 STRUCTURE FIXED CONTROLLERS We will be particularly interested in the case where the set S is described by a set of functions.. m 3 7 7 2 Rm 7 7 5 such that. @h2 @x1 @h2 @x2 @h2 @xn . we have the following alternative... then rF (x0 ) = 0. and then r (x0 ) = rH (x0 )z = 0 has the unique solution z = 0..1 Suppose that x0 2 Rn is a local minimum of the f (x) subject to the constraints H (x) = 0 and suppose further that x0 is a regular point of the constraints. i. . .
Then nd a solution (or solutions) from these necessary conditions and check if the solution(s) satis es our objectives regardless of the solution(s) being a global minima or not. we have the following lemma.1 Although some second order necessary and su cient conditions for local minimality can be given. tij = tji and the e ective constraints for T (x) = 0 are given by the l(l + 1)=2 equations.3 Let T (x) = T (x) 2 Rl l be a symmetric matrix function and let x0 2 Rn be such that T (x0 ) = 0. It is a common practice in many applications that the rstorder necessary conditions are used to derive some necessary conditions for the existence of a minima.1. Since T (x) = tij (x)] is a symmetric matrix. Lemma 20. Then x0 is a regular point of the constraints T (x) = T (x) = 0 if. By de nition. Lagrange Multiplier Method straints H (x) = 0.20.2 Suppose that x0 2 Rn is a local minimum of f (x) subject to the con " 0 # 2 Rm+1 such that 0 rf (x0 ) + rH (x0 ) = 0. for P = P 2 Rl l . even if some su cient conditions can be veri ed. it is still not clear whether the minima is global. 2 Corollary 20. r Trace(T (x0 )P ) = 0 has the unique solution P = 0. and in this case. Proof. constraints are given by a symmetric matrix function. they are usually not very useful in our applications for the reason that they are very hard to verify except in some very special cases. ~ In most of the control applications. Remark 20. Furthermore. x0 is a regular point for the e ective constraints (tij (x) = 0 for i = 1 2 : : : l and i j l) if the following equation has the unique solution pij = 0 for i = 1 2 : : : l and i j l: (x0 ) := l X i=1 rtii pii + l l XX i=1 j =i+1 2rtij pji = 0: Now the result follows by de ning pij := pji and by noting that (x0 ) can be written as 0 1 (x0 ) = r @ l l XX i=1 j =1 tij pji A = r Trace(T (x0 )P ) with P = P . Then there exists 517 Theorem 20.4 Suppose that x0 2 Rn is a local minimum of f (x) subject to the constraints T (x) = 0 where T (x) = T (x) 2 Rl l and suppose further that x0 is a regular . tij (x) = 0 for i = 1 2 : : : l and i j l.
rF (x0 ) = rf (x0 ) + r Trace(T (x0 )P ) = 0: In general. . we have 2 6 x11 6 .. .e.518 STRUCTURE FIXED CONTROLLERS point of the constraints.. in the case where a local minimal point x0 is not necessarily a regular point... . then rF (x0 ) = 0.5 Suppose that x0 2 Rn is a local minimum of f (x) subject to the constraints T (x) = 0 where T (x) = T (x) 2 Rl l . . . @F (x) @F (x) @xk2 @F (x) @x12 @F (x) @x22 . Then there exist 0 6= ( 0 P ) 2 R Rl l with P = P such that 0 rf (x0 ) + r Trace(T (x0 )P ) = 0: Remark 20. Then there exists a unique multiplier P = P 2 Rl l such that if we set F (x) = f (x) + Trace(T (x)P )... 6 6 6 xk2 6 6 .e. 6 . 4 xkq @F (x) @x11 @xkq 3 7 7 7 7 7 7 7 7: 7 7 7 7 7 7 7 7 5 Then 2 6 rF (x) := 6 4 2 6 @F (x) := 6 6 6 6 @X 4 @F (x) @x11 @F (x) @x21 @xk1 . 6 . let us consider the following H2 norm minimization with constant state feedback: the dynamic system is given by x = Ax + B1 w + B2 u _ z = C1 x + D12 u .. i. we have the following corollary. @F (x) 3 7=0 7 5 @F (x) @x1q @F (x) @x2q is equivalent to .. 6 6x 6 k1 6 6 x12 6 x = VecX := 6 .. ~ As an example. Corollary 20. @F (x) @xkq 3 7 7 7 = 0: 7 7 5 This later expression will be used throughout in the sequel. i.2 We shall also note that the variable x 2 Rn may be more conveniently given in terms of a matrix X 2 Rk q ..
3) that B2 X + F = 0. Now the necessary condition for local optimum can be applied: @J (X F ) = B B + P (A + B F ) + (A + B F )P = 0 (20. the controller thus obtained is a global optimal control law. " # VecX the variable x takes the form x = ). equivalently. Hence regularity conditions are satis ed. we cannot conclude from equation (20. XB2B2 X + C1 C1 = 0: This Riccati equation has a stabilizing solution if (A B2 ) is stabilizable and if (C1 A) has no unobservable modes on the imaginary axis. For example.B 2 X and we substitute this relation into equation (20. Indeed. in this case. Hence we have F = . # " @ Trace(T (X F )P ) # " P (A + B2 F ) + (A + B2 F )P = 0 @X @ Trace(T (X F )P ) = 2(B2 X + F )P @F has a unique solution P = 0 since A + B2 F is assumed to be stable at the minimum point. if we assume that B1 is square and nonsingular. then P > 0. For simplicity. then we know that if F is such that A + B2 F is stable. It is routine to verify that J0 = Trace(B1 B1 X ) where X = X 0 satis es T (X F ) := X (A + B2 F ) + (A + B2 F ) X + (C1 + D12 F ) (C1 + D12 F ) = 0: Hence the optimal control problem becomes a constrained minimization problem. We rst verify the regularity conditions: the equation r Trace(T (X F )P ) = 0 or. (Note that in this case.2) 1 1 2 2 @X @J (X F ) = 2(B X + F )P = 0 (20. Care must be exercised to arrive at such a conclusion. and we can use the Lagrange multipliers method outlined above. Let VecF J (X F ) := J0 + Trace(T (X F )P ) with P = P . .1. Lagrange Multiplier Method 519 and the feedback u = Fx is chosen so that A + B2 F is stable and J0 = kTzw k2 2 is minimized.4) and get the familiar Riccati equation: XA + A X .3) 2 @F @J (X F ) = X (A + B F ) + (A + B F ) X + (C + D F ) (C + D F ) = 0: (20.20. in general.4) 2 2 1 12 1 12 @P It should be pointed out that. we shall assume that D12 D12 = I and D12 C1 = 0.
we shall use the Lagrange multiplier method to derive some necessary conditions for a xedorder controller that minimizes an H2 performance. we will further assume that the realization of the controller is minimal. and then the closed loop transfer matrix Tzw can be written as 2 A BC B 3 " 2 1 ~ ~ # 6 BcC2 Ac c BcD21 7 =: A B Tzw = 4 5 ~ C 0 C1 D12 Cc 0 .e. nd an ncth order controller B1 D = 0 . For technical reasons.520 STRUCTURE FIXED CONTROLLERS 20. Suppose a such controller exists.. i. We shall again consider a standard system setup z G w y  K u where the system G is nth order with a realization given by 2 A B B 3 2 6 C1 01 D12 7 : G(s) = 4 5 C2 D21 0 For simplicity. we shall assume that (i) (A B1 ) is stabilizable and (C1 A) is detectable (ii) (A B2 ) is stabilizable and (C2 A) is detectable (iii) D12 C1 D12 = 0 I (iv) h i h i " We shall be interested in the following xed order H2 optimal controller problem: given an integer nc n. D21 21 I # " # K = Ac Bc Cc 0 " # that internally stabilizes the system G and minimizes the H2 norm of the transfer matrix Tzw . (Ac Bc) is controllable and (Cc Ac ) is observable.2 Fixed Order Controllers In this section.
13) Proof. Y C2 C2 ) + XB2B2 X . .(A . Y . ? XB2B2 X ? ^ ^ (A . Moreover. Bc.7) (20. ~~ ~ kTzw k2 = Trace(B B X ) 2 521 (20:5) internally stabilizes the system G and minimizes the norm Tzw .B2 X for some factorization (20. Y C2 C2 Y + ? Y C2 C2 Y ? + B1 B1 ^ ^ (A . and ? := In . Y C2 C2 ) X + X (A . ~ ~ Let Y = Y 2 R(n+nc ) (n+nc ) and denote ~~ ~ ~ ~ ~ ~ J1 = Trace (X A + A X + C C )Y : n o @J1 = Y A + AY = 0 ~ ~ ~~ ~ @X ~ ~ has the unique solution Y = 0 since A is assumed to be stable. Hence the regularity conditions are satis ed and the Lagrange multiplier method can be applied.6). B2 B2 X . B2 B2 X ) + Y C2 C2 Y . Form the Lagrange function J as ~~ ~ J := Trace(B B X ) + J1 1 Then A matrix M is called positive semisimple if it is similar to a positive de nite matrix. Y C2 C2 ) Bc = .12) (20.6 Suppose (Ac Bc Cc) is a controllable and observable triple and K = # " Ac Bc Cc 0 Ac = .11) (20. and Y such that Ac .2.Y C2 Cc = . = Inc 1 and such that with := . . the following with M positivesemisimple conditions are satis ed: 0 0 0 0 = = = = ^^ YX = A X + XA . X . and Cc are given by ~ where X is the observability Gramian of Tzw : ~~ ~ ~ ~ ~ X A + A X + C C = 0: (20:6) Theorem 20.8) (20. ? Y C2 C2 Y ? ^ ^ ^^ rank X = rank Y = rank Y X = nc : (20.20. Fixed Order Controllers ~ with A stable. The problem can be viewed as a constrained minimization problem with the objective function given by equation (20. Then ^ ^ there exist n n nonnegative de nite matrices X .9) M .5) and with constraints given by equation (20.10) (20. B2 B2 X )Y + Y (A . XB2 B2 X + ? XB2 B2 X ? + C1 C1 AY + Y A .
18) ~ @X ~ ~ ~ It is clear that X 0 and Y 0 since A is stable. formula for Ac . .20) (20. Cc . 2.14) @Ac = 2(X12 Y12 + X22 Y22 ) = 0 @J = 2(X B + X Y C + X Y C ) = 0 (20.25) .19) (20. Now it follows from the Lyapunov theorem (Lemma 3.18) can be written as 0 0 0 0 0 0 = = = = = = " # " # X11 A + A X11 + X12 Bc C2 + C2 Bc X12 + C1 C1 X12 Ac + A X12 + X11 B2 Cc + C2 Bc X22 X22 Ac + Ac X22 + X12 B2 Cc + Cc B2 X12 + Cc Cc AY11 + Y11 A + B2 Cc Y12 + Y12 Cc B2 + B1 B1 AY12 + Y12 Ac + B2 Cc Y22 + Y11 C2 Bc Ac Y22 + Y22 Ac + Bc C2 Y12 + Y12 C2 Bc + Bc Bc : (20. (Cc Ac + X22 X12 B2 Cc ) is also observable. (20. X22 > 0 and Y22 > 0: We show rst that X22 > 0.21) (20.26) := . and : given X22 > 0 and Y22 > 0. := .18) that X22 > 0.15) 22 c 22 12 2 12 11 2 @Bc @J (20.522 ~ ~ and partition X and Y as STRUCTURE FIXED CONTROLLERS ~ ~ X = X11 X12 Y = Y11 Y12 : X12 X22 Y12 Y22 The necessary conditions for (Ac Bc Cc ) to be a local minima are @J (20.24) For clarity. we de ne . Since X 0.21) can be written as and X22 X22 12 12 + + 0 = X22 (Ac + X22 X12 B2 Cc ) + (Ac + X22 X12 B2 Cc ) X22 + Cc Cc : + Since (Cc Ac ) is observable by assumption. we shall present the proof in three steps: ~ 1.17) ~ @Y @J = Y A + AY + B B = 0: ~ ~ ~~ ~ ~ (20. we have X22 0 + X = X by Lemma 2. := Y22 1 Y12 (20.X221X12 (20. Equations (20. .17) and (20.16. Y22 > 0 follows by a similar argument..16) @Cc = 2(B2 X11 Y12 + B2 X12 Y22 + Cc Y22 ) = 0 @J = X A + A X + C C = 0 ~~ ~ ~ ~ ~ (20. Bc . Hence equation (20.22) (20.27) .23) (20.
Moreover.1 Y : Y11 .15) and (20.14) that . )Y12 )C2 = .13) follow by substituting Ac .23) and by using the fact ^ ^ ^ ^ that X = X and Y = Y . and is a positive semisimple matrix. 1 1 1 .24). Cc = .22) with some tedious algebra. (20. = . Y12 Y22 12 523 (20.Y12 X12 ).(X221X12 Y11 + Y12 )C2 = (. = .16). Bc. from equation (20. Y12 Y22 1 Y12 0 . equations (20. =I 0 that X 0 ~ ~ and 2 = .B2 (X11 + X12 .23) and some algebra.10) and (20. Bc = .29) (20. Finally. we have nc = rank X22 rank X12 nc : This implies that rank X12 = nc = rank Y12 and ^ ^ ^^ rank X = rank X12 = nc = rank Y = rank Y X: ^^ The product of Y X can be factored as ^^ Y X = (. . X12 X221 X12 . Using these formulae in equations (20.) = . X11 .12) and (20. X12 X221X12 0 .31) Then it follows from equation (20. and Cc into (20.11) follow from equations (20.(Y11 . and Y : Equations (20. X .B2 X : The formula for Ac follows from (20. and (20.B2 (X11 Y12 Y22 1 + X12 ) = . ^ ^ 3. Fixed Order Controllers ^ X ^ Y X Y := := := := . M := Y22 X22 = Y22=2 (Y22=2 X22 Y22=2 )Y22 1=2 2 .28) (20.Y11 . Equations for X . Y .30) (20.20. (.Y C2 and . It also follows from X 0 and Y and Y 0.2.14).19) and (20.20) . Y12 )C2 = . we get . = Y22 X22 .
B2 B2 X Y C2 ) is controllable.12) and (20. (Note that Ko is known to be the optimal H2 controller ^ ^ from Chapter 14). ~ Remark 20.B2 X . then Theorem 20. we get that equations (20.e. ~ . B2 B2 X . is much more complicated. but that. Ko is not necessarily minimal hence Theorem 20. equivalently. or equivalently i Ko := A .10) and (20. we can indeed nd . XB2 B2 X + C1 C1 0 = AY + Y A . In fact.6 generates the optimal controller. and are both square and .1 C K = Ac Bc = . then lower nc and try again. Using these facts. In that case. in this case. = In .11) become standard H2 Riccati equations: 0 = A X + XA . .Y C2 C2 ) t XB2B2 Xe(A. if optimal controller Ko is controllable " # Hence..B2 B2 X )t Y C2 C2 Y e(A.3 It is interesting to note that if the full order controller nc = n is considered. Y C2 C2 Y + B1 B1 ^ ^ It is a fact that X is nonsingular i (B2 X A . An alternative solution to this dilemma would be simply by direct testing: if a given nc does not generate a controllable and observable controller.B2 X 0 is controllable and observable.1 . Y 1 2 C2 ). B2 B2 X . if X and Y are nonsingular or. B2 B2 X . Furthermore. However.B2 X 0 ^ ^ i. Y C2 C2 Y C2 . it is generally hard to solve these coupled equations although some ad hoc homotopy algorithm might be used to nd a local minima. equations (20.. if X and Y are nonsingular. in general. .(A . then we have = I and ? = 0. Y C2 C2 are stable. = ( ).13) have unique solutions: Z1 ^ X = e(A.Y C2 C 0 .6 without assuming the minimality of the optimal controller by using pseudoinverse in the derivations. Y C2 C2 ) is observable and that Y is nonsingular i (A .Y C2 C2 )t dt 0 0 Z1 ^ Y = e(A.6 will not be applicable.4 We should also note that although we have the necessary conditions for a reduced order optimal controller. . It is possible to derive some similar results to Theorem 20.524 STRUCTURE FIXED CONTROLLERS Moreover. Y C2 C2 Y C2 = K = o .B2 B2 X ) t dt 0: 0 Remark 20. 0 " c # A . B2 B2 X and A . and such that . there exist unique stabilizing solutions X 0 and Y 0 to these two Riccati equations such that A . we have " # " # and observable as we assumed. in general.
. Theorem 20.3 Notes and References Optimization using the Lagrange multiplier can be found in any standard optimization textbook.5 This method can also be used to derive the H1 results presented in the previous chapters. The Lagrange multiplier method has been used extensively by Hyland and Bernstein 1984].20. The interested reader should consult the references for details.6 was originally shown in Hyland and Bernstein 1984]. Bernstein and Haddard 1989]. It should be pointed out that this method su ers a severe de ciency: global results are hard to nd. This is due to (a) only rst order necessary conditions can be relatively easily derived (b) the controller order must be xed hence even if a xedorder optimal controller can be found. the book by Hestenes 1975] contains the nite dimensional case. it may not be optimal over all stabilizing controllers. ~ 20. and the one by Luenberger 1969] contains both nite and in nite dimensional cases.3. Notes and References 525 Remark 20. and Skelton 1988] in control applications. In particular.
526 STRUCTURE FIXED CONTROLLERS .
The discrete time H2 optimal control will be considered brie y in Section 21. Section 21. Various di erent characterizations of a bounded real transfer matrix are presented in section 21.1 Discrete Lyapunov Equations Let A B . Hence we will start from the solutions of a discrete Lyapunov equation which are given in section 21.7.6 and 21. 527 .4 collects some useful matrix function factorizations and characterizations. The key is the relationship between the existence of a solution to a Riccati equation and the norm bound of a stable transfer matrix. state space criteria are stated (mostly without proof) for a transfer matrix to be inner. innerouter factorizations.1. for the existence of coprime factorizations.1) has a unique solution if and only if i (A) j (B) 6= 1 for all i j . A simpler form of a Riccati equation is the socalled Lyapunov equation.1 The equation (21. the discrete time balanced model reduction is considered in section 21. In particular.5.3. Finally. X + Q = 0: (21:1) Lemma 21. normalized coprime factorizations. and Q be real matrices with appropriate dimensions. and consider the following linear equation: AXB .Discrete Time Control In this chapter we discuss discrete time Riccati equations and some of their applications in discrete time control.2 presents the basic property of a Riccati equation solution as well as the necessary and su cient conditions for the existence of a solution to the LQR problem related Riccati equation. and spectral factorizations. 21 21. Section 21.
2 Let Q be a symmetric matrix and consider the following Lyapunov equation: AXA . Q. and.1 : S := .G(A ). then the equation (21.(A ). Let a matrix S 2 R2n 2n be partitioned into four n n blocks as 0 .1) has either no solution or more than one solution depending on the speci c data given.1 Q . Suppose that A is stable. it is easy to see that if is an eigenvalue of a simplectic matrix S . 2.528 DISCRETE TIME CONTROL 2 Proof. furthermore.I and let J = 2 R2n 2n then S is called simplectic if J . Just as the continuous time Riccati equations play the essential roles in continuous H2 and H1 theories. Lemma 21. and then the following statements hold: P (a) X = 1 Ai Q(A )i and X 0 if Q 0. and G be real n n matrices with Q and G symmetric and A nonsingular. and 1= are also eigenvalues of S . A I 0 simplectic matrix has no eigenvalues at the origin. then 1= . then the equation is called the discrete Lyapunov equation. If B = A and Q = Q .2 Discrete Riccati Equations This section collects some basic results on the discrete Riccati equations. 21. Q 0 and (Q A) is detectable. ~ The following results are analogous to the corresponding continuous time cases. Suppose that X is the solution of the Lyapunov equation then (a) j i (A)j 1 if X > 0 and Q 0. Let A. X + Q = 0 1. (b) A is stable if X 0. so they will be stated without proof.1 If i (A) j (B) = 1 for some i j . i=0 (b) if Q 0.1 . the discrete time Riccati equations play the essential roles in discrete time H2 and H1 theories. De ne a 2n 2n matrix: " # A + G(A ).1 " # S := S11 S12 S21 S22 " # . Remark 21.1 Q (A ). So the presentation of this section and the sections to follow will be very much like the corresponding sections in Chapter 13. Analogous to the continuous case. then (Q A) is observable i X > 0.1 S J = S .
1 The domain of Ric. (S ) is taken to be the subspace spanned by the generalized principal vectors corresponding to those generalized eigenvalues in jz j < 1. Then X is uniquely determined by S . De nition 21. stacking the basis vectors up to form a matrix. Here the generalized principal . denoted by dom(Ric). we make the following de nition. Then (a) X is unique and symmetric (b) I + XG is invertible and X satis es the algebraic Riccati equation A XA .1 . Then it must have n eigenvalues in jz j < 1 and n in jz j > 1. After nding a basis for X. Assume that S has no eigenvalues on the unit circle. (S ). S 7! X is a function which will be denoted Ric thus.2 In the case that A is singular.1 XA + Q = 0 (c) A .21. if the two subspaces " # are complementary. (21:2) Note that the discrete Riccati equation in (21. I Theorem 21. X + Q = 0: Remark 21.2. we get X.3 Suppose S 2 dom(Ric) and X = Ric(S ). A XG(I + XG). Discrete Riccati Equations 529 Then S is a simplectic matrix. (S ) = Im T1 T2 X. A 0 0 A .e.1 A is stable. we can set X := T2T1. As in the continuous time case.2) can also be written as A (I + XG). If T1 is nonsingular or. (S ) and Im are complementary. G(I + XG). X . all results presented in this chapter will still be true if the eigenvalue problem of S is replaced by the following generalized eigenvalue problem: " # " # I G .1 XA .1 XA = (I + GX ). X = Ric(S ).. i. equivalently. and the latter corresponds to eigenvalues in jz j > 1. and partitioning the matrix. (S ) Im 0 I " # where T1 T2 2 Rn n . consists of all (2n 2n) simplectic matrices S such that S has no eigenvalues on the unit circle and the two " # 0 subspaces X. (S ) and X+ (S ): the former is the invariant subspace corresponding to eigenvalues in jz j < 1.Q I and X. Consider the two ndimensional spectral subspaces X.
(S ) = Im T1 = Im I T1 = Im I : T2 X X Obviously. Since (21:3) We only need to show that T1 T2 is symmetric. Proof. 9 T = T1 T2 " ~ such that X. = (S. Let X := T2 T1. Since X. (S ) = Im T1 T2 and T1 is invertible. there is a stable n n matrix S.Q I # " A 0 x = .: (21:4) Premultiply by S.Q I 1 #! I G xi = 0 A # " " # I G x 1 0 A # I G x i.530 vectors corresponding to a generalized eigenvalue fx1 : : : xk g satisfying DISCRETE TIME CONTROL are referred to the set of vectors " " A 0 . T )JST = T S JST = T JT T1 XT1 = T1 T2: S. T J )TS. X is unique since " # " # I = Im I Im X1 X2 XT1 = T2 premultiply by T1 to get " # " # i X1 = X2 . .1 i = 2 : : : k: 0 A # See Dooren 1981] and Arnold and Laub 1984] for details. such that ST = TS. . (T JT )S. Now let us show that X is symmetric. (S ) is a stable invariant subspace. (a): Since S 2 dom(Ric). then " # " # X. T J )ST = (S.1. T JT = 0: . T J and note that S JS = J we get (S.
it follows that v X = 0.4) can be written as T2 T1 = T1 T2 : S " Premultiply (21. (b): To show that X is the solution of the Riccati equation.5) by . and we get v XA = 0. G(A ).1 (X .e.. Since A is assumed to be invertible.1 which is stable. T1. Hence there is a unique solution 531 T JT = 0 i. XA + (I + XG)(A ). in turn. Hence I + XG is invertible and the equation (21.6) by v . is stable. Q) = 0: We now show that I + XG is invertible.1: (21:5) i h Equivalently. h i (c): To show that X is the stabilizing solution.21. we get i "I # .1 Q . we note that equation (21.7).5) by I 0 to get A + G(A ). T1.1 X = T1 S. Thus we have X = X since T1 is nonsingular. X + Q = 0: This is the Riccati equation (21. 2 .X I h I = I X X # " # T1S.2. This. Discrete Riccati Equations This is a Lyapunov equation and S. implies v = 0 by (21.1 A = T1 S. T1.2). premultiply (21.6) can be written as A (I + XG). Suppose I + XG is not invertible. then 9 v 2 Rn such that v (I + XG) = 0: (21:7) Premultiply (21.1: The above equation can be simpli ed by using Riccati equation to get (I + GX ).1 XA .X I S X = 0: (21:6) to get .
has all eigenvalues inside the unit disc.1 X i=0 k. k = 0 1 : : :: (21:8) Then Uk+1 = SUk with U0 = T .Uk V Uk + U0 V U0 (Ui+1 V Ui+1 .Q 0 I 0 S V S .1 0 I 0 . T 2 Proof. Therefore T1 T2 = limk!1 Yk 0.1 such that (I + GX ). Ui V Ui ) Ui (S V S . = . Then S 2 dom(Ric) i (A G) is stabilizable and S has no eigenvalues on the unit circle.1 A is antistable. V = I . So Yk 0 for all k 0.532 ~ ~ Remark 21. 0 0 # Yk := . Note that Uk ! 0 as k ! 1 since S. (S ) = Im 1 then T1 T2 0.Q 0 . . Let T = T1 T2 " # and S. Let k Uk = TS. 2 " #" #" # Lemma 21.4 Suppose that G and Q are positive semide nite and that S has no eigen" # T values on unit circle. Then the ~ T2 ~ ~~ ~ Riccati equation has an antistabilizing solution X = T2 T1.A Q I since G and Q are assumed to be positive semide nite. De ning V = Further de ne = . V )Ui : .3 Let X+(S ) = Im T1 and suppose that T1 is nonsingular.5 Suppose that G and Q are positive semide nite. we get T1 T2 = U0 V U0 . Let X. and S.1 G(A ).1 X i=0 " 0 I . be such that ST = TS. " # DISCRETE TIME CONTROL ~ Lemma 21. Now k. has all eigenvalues inside of the unit disc.
1 QT1 + (A ). First. we have GT2 S. and a corresponding eigenvector.2. T2 and postmultiply x to get (21:11) . su cient.1 T2 x = T2 S.4 T2 T1 0. We now show that the stabilizability of (A G) and S having no eigenvalues on the unit circle are. 2 T2 Lemma 21. But this would imply 1 x = 0. (21. x = 0 from (21. . invariant.9) to get AT1 . x: According to Lemma 21.11). . furthermore. then 1= is an eigenvalue of A and.G(A ). in fact.1 Q)T1 . jKer T1 has an eigenvalue.1 : S= .6 Let G and Q be positive semide nite matrices and " # A + G(A ).(A ) (21.e. To prove this. let x 2 Ker T1 .1 Q . it is claimed that KerT1 is S.1 Q (A ).10) into (21.9) . i. x = T2 x: Now if T2 x 6= 0. To show this. 0 = GT2 S. = T1 S. Now to prove that T1 is nonsingular. T2 GT2 S.1 T2 = T2 S. which is impossible.10) by x to get (21:12) j j < 1 0 6= x 2 Ker T1: (A ).8) as (A + G(A ). we only need to show that T1 is nonsingular. Hence " # T we must have T2 x = 0. T2 T1S. suppose on the contrary that Ker T1 6= 0. x = GT2 x. . Rewrite (21.x = x Postmultiply (21. x = x S. Discrete Riccati Equations 533 Proof.21. : . The necessary part is obvious.10) Substitute (21. so GT2 x = 0.(A ). Ker T1 = 0. x: S. Hence Ker T1 is invariant under S. G(A ).1 T2 = T1 S. which is contradictory to the stabilizability assumption of (A G).1 Then S has no eigenvalues on the unit circle i (A G) has no uncontrollable modes and (Q A) has no unobservable modes on the unit circle. GT2 S.x S. Then S. x = 0: This in turn implies that T1 S. and premultiply the above equation by x S.
The case for (A G) having uncontrollable mode on the unit circle can be proven similarly.13) j A y: .j y and equation (21. x Qx = 0: yG = 0 Qx = 0: Substitute these relationships into (21.14) Premultiplying equation (21.j y Ax = y Gy + y x .y Gy . again a contradiction. on the contrary. ej Gy = ej x (21.1 y = ej y: Multiplying the second equation by G and adding it to the rst one give Ax .(A ). (() Suppose.1 Q)x . Then " # " # x y for some " # x 6= 0. that S has an eigenvalue ej .Qx + y = e (21.14) to get Ax = ej x ej A y = y: Since x and y cannot be zero simultaneously.1 y = ej x ). a contradiction. y (A + G(A ). i. ()): Suppose that S has no eigenvalue on the unit circle but ej is an unobservable mode of (Q A) and x is a corresponding eigenvector. 2 .13) by e.534 DISCRETE TIME CONTROL S x = ej y Proof.x Qx + x y = ej x A y: Thus It follows that . Then it is easy to verify that S x = ej 0 " # " # x 0 so ej is an eigenvalue of S . G(A .13) and (21.14) by x yield e.1 Qx + (A ). ej is either an unobservable mode of (Q A) or an uncontrollable mode of (A G).e..
(I . so j j < 1 since (I + GX ). Discrete Riccati Equations 535 Theorem 21.15) by x . (I + GX ).1 Q .1 . i. and postmultiply by x to get XAx = 0 Cx = 0: (21:16) This implies that KerX is an Ainvariant subspace. Thus is an unobservable stable mode of (Q A). (ii) or. Proof. then there is an 0 6= x 2 KerX . Then there exists a x 2 C n such that Ax = x and Cx = 0 do the same pre.1 .1 x = x. BR. DR. On the other hand.1D )C A . such that Ax = x.1 X (I + GX ).1 D C ) has no unobservable modes on the unit circle. suppose x 2 KerX .1 Q (A ).1 Ax = (I + GX ). (i) C D . X = Ric(S ) 0 if S 2 dom(Ric) and X > 0 if and only if (Q A) has no unobservable stable modes. If KerX 6= 0. I < 0. Thus X 0 by Lyapunov theorem. BR.1 A . ej I B has full column rank for all 2 0 2 ].1A is stable. But for x 2 KerX .8 Suppose that D has full column rank and let R = D D > 0 then the following statements are equivalent: A . X is singular. premultiply (21.7 Let G and Q be positive semide nite matrices and " # A + G(A ).1 : S= ..2 GXA + C C 0. I ]X 1=2 x = 0: Now X 0 G 0.15) as before to get j j2 x (XG + I ).e. Hence Xx = 0. " # .1 Xx . 2 Lemma 21.2. Furthermore. To show that the kernel of X has the refereed property.1 Then S 2 dom(Ric) i (A G) is stabilizable and (Q A) has no unobservable modes on the unit circle.2 GXA + C C = 0 (21:15) and note that by de nition (I + GX ). so Cx = 0. equivalently. x Xx = 0: This can be rewritten as x X 1=2 j j2 (I + X 1=2 GX 1=2 ).G(A ).(A ).6. (D? C A . Let Q = C C for some matrix C . X + A X (I + GX ). Now rewrite the discrete Riccati equation as A (I + XG).1D C has no unobservable modes on the unit circle.and postmultiplications on (21.5 and 21. suppose that j j < 1 is a stable unobservable mode of (Q A). and j j < 1 imply that j j2 (I + X 1=2GX 1=2 ). The rst half of the theorem follows from Lemmas 21.21.1 A is stable and A X (I + GX ).
9 Suppose that D has full column rank and denote R = D D > 0. " But this implies that A .) Q ..e. ej I B has to be invertible. BR. ej I B C D #" # u = 0: v u = I 0 .1D C I A . BR. BR.1D )Cx + Dy = 0: Premultiply (21. X = Ric(S ) 0. Then S 2 dom(Ric) i (A B ) is stabilizable and C D full column rank for all 2 0 2 ].(E ) (E ) where E = A .19) by D to get y = 0. .1D )C . ej is an unobservable mode of (I . suppose that (21. Conversely.536 DISCRETE TIME CONTROL Proof. Let S have the form . DR. Suppose that ej is an unobservable mode of (I . Q = C (I . ej I B C D #" " .17) does not have full " # column rank for some then there exists " u 6= 0 such that v Now let Then and " # " " # " A . G = BR.1 D C . ej I )x + By = 0 (I .1D C I v .1 D )Cx = 0 i. and E is assumed " # A . DR.R #" # x = I 0 .1 D )C A .1 .1D )C A . Then we have (A .1 D )Cx = 0 .1 S = E + G(E. BR. DR. Furthermore.1 D C . DR.1 D C )x = ej x (I . " # . 2 Corollary 21.1D C I y R #" # x : y u 6= 0 v (21:18) (21:19) (A . ej I B C D I 0 #" # 0 x = 0: (21:17) # does not have full column rank. BR.1 B .1D C then there is an x = 0 such that 6 (A .e. DR.1 D C )x = ej x (I .R. BR. i.G(E . DR..)1 1Q .1 D C .
1 : .21.1C C . M (z )M (z )].1 ).1 ) C C BB (A.1 ). B (A ).1 B 3 I .3.(A ) C C (A ) " # Then the following statements are equivalent: (i) kM (z )k1 < 1 (ii) S has no eigenvalues on the unit circle and kC (I . BB (.1 0 7: 4 5 . De ne M (z ) := M T (z .BB (A ).1 (B XA + D C ) + C C = 0: 21. X . .1 I .1 = 6 (A ).9 is E XE .1 B k < 1.1 B (A ) D .3 Bounded Real Functions Let a real rational transfer matrix be given by M (z ) = A B C D " # where again A is assumed to be nonsingular and the realization is assumed to have no uncontrollable and no unobservable modes on the unit circle.B (A ).1 .1 C C .1 C M (z ) = (A ) .1 S := A .1 C " # : Lemma 21. Proof.10 Let M (z) = A B 2 RL1 and let S be a simplectic matrix de ned C 0 by " # A . BB (A ). It is easy to compute that 2 A . Note again that all results hold for A singular case with the same modi cation as in the last section.(A ). A).1 C C (A ).B (A ) It is claimed that I .1 has no uncontrollable and/or unobservable modes on the unit circle. Bounded Real Functions 537 Note that the Riccati equation corresponding to the simplectic matrix in Corollary 21. X . M (z )M (z )].1 XE + Q = 0: This equation can also be written as A XA . (B XA + D C ) (D D + B XB ). Then . E XG(I + XG).
M M ).1 C C . q2 = 0. This implies e. suppose kM (z )k1 1. M M ). 2. We shall now see how to handle nonstrictly proper case.1 = ej q1 q2 : . Similar proof can be done for observability. So it is su cient to show that kM (z )k1 < 1 . Then q1 A = ej q1 . suppose that = ej is an uncontrollable mode of I . This is a contradiction since M (z ) 2 RL1 . It is clear that we have already proven that S has no eigenvalues on the unit circle i (I .I S= : I (A ). and we shall give an example below to show why this restriction is sometimes necessary for the technique to work.538 DISCRETE TIME CONTROL There are two possibilities: 1.1 C C (A ).j is an eigenvalue of A.1 2 RL1 andkM (1)k < 1: It is obvious that the right hand side is necessary.1 . q Then 9 q = 1 2 C 2n such that q2 " # " # A .. D D > 0.1 I Hence S does not have eigenvalues on the unit circle.j )M (ej ) is singular.1 C C . which again implies that M (z ) has a mode on the unit circle if q1 6= 0. Thus in particular if kM (z )k1 < 1 then I .1 0 Hence q1 B = 0 and h i h i q1 A + q2 (A ).1 = ej q2 . For that purpose we shall focus our attention on the stable system. To show " # that. the above condition may not be true if the system is not stable.1 C C q2 (A ). since max (M (1)) < 1 and M (ej ) is continuous in . we have assumed that the transfer matrix is strictly proper. BB (A ). Aq2 = e. q2 6= 0.I A .1 2 RL1 .BB (A ). Now note that " #" #" # . i.BB (A ). M (z )M (z )]. hence the claim is true. M (e.j q2 . again a contradiction.1 2 RL1 . This implies that 1 is an eigenvalue of M (e. 2 In the above Lemma. M M ). if a function is only known to be in RL1 .e. Then we have q2 (A ). BB (A ).1 = ej q q B = 0: q (A ). This contradicts to (I . To show that it is also su cient. then max (M (ej )) = 1 for some 2 0 2 ]. (I . We rst note that H1 norm of a stable system is de ned as kM (z )k1 = sup (M (z )) : jzj 1 Then it is clear that kM (z )k1 (M (1)) = kDk. On the other hand.j )M (ej ).1 . so I .1 C C (A ).
1 Q . X > 0 if (C A) is observable.1 E is stable. D D).1=2 C 0 Proof.1 B + Dk < 1 (c) 9 X 0 such that I . D D . D D can be inde nite. D D). 1= = 1= 1 539 " # 1 2 RL1 : Then kM1(z )k1 = 1.1 B Q := C (I . B XB > 0 and E XE . DD ). Then kM (z )k1 < 1 if and only if N (z ) 2 RH1 and kN (z )k1 < 1 where " # " ^# A + B (I . E XG(I + XG).1 D C G := .1 XE + Q = 0 " A B C D # 2 " # and (I + GX ). D D). E XG(I + XG).4.B (I . This is exactly the analogy of Corollary 17.1 Q (E ).(E ).1 : S := .12 Let M (z) = A B 2 RH1 and de ne C D E := A + B (I . A).11 Let M (z) = 2 RH1 .1 C: Suppose that E is nonsingular and de ne a simplectic matrix as " # E + G(E ). D D . D D = 0. but 1 . 3 Lemma 21. In general. Theorem 21.1 Then the following statements are equivalent: (a) kM (z )k1 < 1 (b) S has no eigenvalues on the unit circle and kC (I . if M 2 RL1 and kM k1 < 1.1 D C B (I .21. then I . < 1.G(E ).1 XE + Q < 0 . B XB > 0 and E XE . DD ). Moreover. (d) 9 X > 0 such that I . Bounded Real Functions Example 21. D D). X .3.1 Let 0 < < 1=2 and let z M1 (z ) = z .1=2 =: E B : N (z ) = ^ C 0 (I . X .
B XB ).1 I .1 D #! = " T 0 0 I and #" A B C D #" T 0 0 I #.540 (e) 9 X > 0 such that DISCRETE TIME CONTROL " A B C D #" X 0 0 I #" A B . (a). (g))(f) follows from Theorem 11. (c))(a) We shall only give the proof for D = 0 case.(b) follows by Lemma 21. A) = X .BB . the case D = 0 can be 6 transformed to the zero case by Lemma 21. (d))(c) can be shown in the same way as in the proof of Theorem 13. z := ej Assuming (c) is satis ed by some X 0 and considering the obvious relation with (z . G = .1 <1 (g) " A B C D #! < 1 with 1 In 0 2 0 2 := (" # : 12C 2 2 Cm p ) C (n+m) (n+p) (assuming that M (s) is a p m matrix and A is an n n matrix).10 (a))(g) follows from Theorem 11. (e))(d) follows by Schur complementary formula.1 I .5. Hence in the following we have E = A. D D .1 (B XA + D C ) + C C = 0: Proof.11.11. and Q = C C . A )X (zI . (f))(e) follows by letting X = T T . X 0 <0 C D 0 I # " # (f) 9 T nonsingular such that " TAT . Note that the Riccati equation in (c) can also be written as A XA .1 TB CT .7.1 B XA: . A )XA + A X (zI . B XB ). A XA = C C + A XB (I . A) + (z . X + (B XA + D C ) (I .
1 )(I . A). B XB )1=2 : (I . kM k1 < 1. However. Bounded Real Functions 541 The last equality is obtained from substituting in Riccati equation.13 Let M (z) = A B 2 RL1 and suppose 9 X = X such that C 0 A XA .(I . B XB ). Corollary 21. Hence ej cannot be a zero of W ..1)M (z ) = W (z .3.1 )W (z ) where A B W (z ) = .1B XA + C C = 0: Then where " # I .12 can be used to compute the H1 norm of a discrete time transfer matrix. B XB ) " # Moreover.1 Agj 6= 1.1 B XA . B XB ).1 B to get I . Remark 21. 2 The following more general results can be proven easily following the same procedure as in the proof of (c))(a). B XB )W (z ) A B W (z ) = .21. B XB ) " # Suppose W (ej )v = 0 for some and v then ej is a zero of W (z ) if v 6= 0. ~ . A).1 )M (z ) = W (z . Now premultiply the above equation by B (z .1 I . i. X + A XB (I .1 B XA = (I . BB X ).1 A that are all inside of the unit circle.4 As in the continuous time case. B XB > 0 (< 0) and j i f(I . B XB > 0 (< 0). M (z . the equivalence between (a) and (b) in Theorem 21. BB X ). all the zeros of W (z ) are given by the eigenvalues of A + B (I .1=2 B XA . the following statements hold: (1) if I .e. then kM (z )k1 1 ( 1) (2) if I . M (z . then kM (z )k1 < 1 (> 1). we get I .1 and postmultiply by (zI .j )M (ej ) > 0 for all 2 0 2 ]. M (e. Therefore.I : (I .
1 B ) D 2 The following corollary is a special case of this lemma which gives the necessary and su cient conditions of a discrete inner transfer matrix with the statespace representation. X + C C = 0: Then (a) D C + B XA = 0 implies N N = (D . inside of the unit disc).1 B ) D (b) X 0. Corollary 21. # Proof.4 Matrix Factorizations 21.1 4 5 DT C + B XA .. The results follow by noting the following equality: N (z )N (z ) = = 3 2 A 0 B 6 (A ). and N N = (D .1 C D 7 : 0 (A ).B (A ).1 (D .1 B ) D 2 3 A 0 B 6 6 7 XB + (A ). Lemma 21. (A B ) controllable.e.1 B ) D implies D C + B XA = 0.1 B ) D = D D + B XB = I .1 (D . X + C C = 0 (b) D C + B XA = 0 (c) (D . A transfer matrix is called outer if all its transmission zeros are stable (i. CA. CA.1 (A ).4.B (A ). CA.1 B ) C . A B C D # 2 RH1 is a controllable realiza0 such that . then N (ej ) = N (ej ).1 C D 7 5 4 (D . Note that if N has no poles at the origin. CA.15 Suppose that N (z) = " tion then N (z ) is inner if and only if there exists a matrix X = X (a) A XA .542 DISCRETE TIME CONTROL 21.14 Let N = A B and suppose that X = X satis es C D A XA . CA.1 C C (A ). CA.1 Inner Functions " A transfer matrix N (z ) is called inner if N (z ) is stable and N (z )N (z ) = I for all z = ej .
21.4. Matrix Factorizations
543
The following alternative characterization of the inner transfer matrix is often useful and insightful.
Corollary 21.16 Let N (s) =
nonsingular such that
"
A B C D
#
2 RH1 and assume that there exists a T
;1 P = TAT;1 TB CT D
"
#
and P P = I:
Then N (z ) is an inner. Furthermore, if the realization of N is minimal, then such T exists.
Proof. Rewrite P P = I as
"
A B
C D
#"
T T
#"
I
A B = T T : C D I
# "
#
Let X = T T , and then
"
A XA ; X + C C A XB + C D = 0: B XA + D C B XB + D D ; I
#
This is the desired equation, so N is inner. On the other hand, if the realization is minimal, then X > 0. This implies that T exists.
2
In a similar manner, Corollary 21.15 can be used to derive the statespace representation of the complementary inner factor (CIF).
Lemma 21.17 Suppose that a p m (p > m) transfer matrix N (z) (minimal) is inner h i then there exists a p (p ; m) CIF N? 2 RH1 such that the matrix N N? is
square and inner. A particular realization is
N? (z ) = A Y C Z
where Y and Z satisfy
"
#
(21:20) (21:21) (21:22)
A XY + C Z = 0 B XY + D Z = 0 Z Z + Y XY = I:
544
DISCRETE TIME CONTROL
Proof. Note that
h
the results by using the formula in Corollary 21.15.
i "A B Y # is inner. N N? = C D Z
Now it is easy to prove
2
21.4.2 Coprime Factorizations
M N
Recall " #that two transfer matrices M (z) N (z) 2 RH1 are said to be right coprime if is left invertible in RH1 i.e., 9 U V 2 RH1 such that
U (z )N (z ) + V (z )M (z ) = I: The left coprime is de ned analogously. A plant G(z ) 2 RL1 is said to have double coprime factorization if 9 a right coprime factorization G = NM ;1 , a left coprime ~ ~ ~ ~ factorization G = M ;1 N , and U V U V 2 RH1 such that " #" # ~ V U M ;U = I: (21:23) ~ ~ ~ ;N M N V
The state space formulae for discrete time transfer matrix coprime factorization are the same as for the continuous time. They are given by the following theorem.
Theorem 21.18 Let G(z) =
"
alization. Choose F and L such that A + BF and A + LC are both stable. Let ~ ~ ~ ~ U V U V N M N , and M be given as follows
A B C D
#
2 RL1 be a stabilizable and detectable reBZr Zr DZr
"
# 2 A + BF M := 6 4 F N
M := A + LC L B + LD ~ N Zl C Zl Zl D " # " # U := A + LC L ;(B + LD) V Zl;1 F 0 Zl;1 ~ ~ where Zr and Zl are any nonsingular matrices. Then G = NM ;1 = M ;1N are rcf and
lcf, respectively, and (21.23) is satis ed.
" ~ # 2 A + BF LZl;1 U := 6 F 0 4 ~ V ;(C + DF ) Zl;1 " ~# "
C + DF
3 7 5
3 7 5 #
21.4. Matrix Factorizations
545
Some coprime factorizations are particularly interesting, for example, the coprime factorization with inner numerator. This factorization in the case of G(z ) 2 RH1 yields an innerouter factorization.
j B Theorem 21.19 Assume that (A B) is stabilizable, A ;Ce I D has full column rank for all 2 0 2 ], and D has full column rank. Then there exists a right coprime factorization G = NM ;1 such that N is inner. Furthermore, a particular realization is
"
#
given by
"
# 2 A + BF M := 6 4 F N
C + DF
BR;1=2 R;1=2 DR;1=2
3 7 5
where
R = D D + B XB F = ;R;1 (B XA + D C )
and X = X
0 is the unique stabilizing solution AD X (I + B (D D);1 B X );1AD ; X + C D? D? C = 0 where AD := A ; B (D D);1 D C . Using Lemma 21.17, the complementary inner factor of N in Theorem 21.19 can be obtained as follows " # A + BF Y N? = C + DF Z where Y and Z satisfy A XY + C Z = 0
B XY + D Z = 0 Z Z + Y XY = I: Note that Y and Z are only related to F implicitly through X .
Remark 21.5 If G(z) 2 RH1 , then the denominator matrix M in Theorem 21.19 is an outer. Hence, the factorization G = N (M ;1) is an innerouter factorization. ~
Suppose that the system G(z ) is not stable then a coprime factorization with inner denominator can also be obtained by solving a special Riccati equation.
546
Theorem 21.20 Assume that G(z) := A B 2 RL1 and that (A B) is stabilizC D able. Then there exists a right coprime factorization G = NM ;1 such that M is inner if and only if G has no poles on the unit circle. A particular realization is " # 2 A + BF BR;1=2 3 M := 6 R;1=2 7 4 F 5 N ;1=2 C + DF DR
where and X = X
"
#
DISCRETE TIME CONTROL
0 is the unique stabilizing solution to A X (I + BB X );1 A ; X = 0:
R = I + B XB F = ;R;1B XA
Another special coprime factorization is called normalized coprime factorization which has found applications in many control problems such as model reduction controller design and gap metric characterization. Recall that a right coprime factorization of G = NM ;1 with N M 2 RH1 is called a normalized right coprime factorization if M M +N N =I
M i.e., if N
"
#
is an inner.
h
~ ~ Similarly, an lcf G = M ;1 N is called a normalized left coprime factorization if i ~ N is a coinner. Then the following results follow in the same way as for the ~ M continuous time case. Theorem 21.21 Let a realization of G be given by
G= A B C D
and de ne
"
#
~ R = I + D D > 0 R = I + DD > 0: (a) Suppose that (A B ) is stabilizable and that (C A) has no unobservable modes
on the imaginary axis. Then there is a normalized right coprime factorization G = NM ;1 " # 2 A + BF BZ ;1=2 3 M := 6 Z ;1=2 7 2 RH1 5 4 F N ;1=2 C + DF DZ
21.4. Matrix Factorizations
where and X = X
547
0 is the unique stabilizing solution ~ An X (I + BR;1 B X );1 An ; X + C R;1 C = 0 where An := A ; BR;1 D C . (b) Suppose that (C A) is detectable and that (A B ) has no uncontrollable modes on the imaginary axis. Then there is a normalized left coprime factorization G = ~ ~ M ;1N h i " A + LC L B + LD # ~ ~ M N := ~ ;1=2 ~ ;1=2 ~ ;1=2 Z C Z Z D where ~ ~ Z = R + CY C ~ L = ;(BD + AY C )Z ;1 and Y = Y 0 is the unique stabilizing solution ~ ~ ~ An Y (I + C R;1CY );1 An ; Y + BR;1 B = 0 ~ ~ where An := A ; BD R;1 C .
Z = R + B XB F = ;Z ;1(B XA + D C )
M (c) The controllability Gramian P and observability Gramian Q of N
by
"
#
are given
P = (I + Y X );1Y Q = X h ~ ~ ~ ~ while the controllability Gramian P and observability Gramian Q of M N
are given by
i
~ P =Y
~ Q = (I + XY );1 X:
21.4.3 Spectral Factorizations
"
The following theorem gives a solution to a special class of spectral factorization problems.
Theorem 21.22 Assume G(z) := A B 2 RH1 and > kG(z)k1. Then, there C D exists a transfer matrix M 2 RH1 such that M M = 2I ; G G and M ;1 2 RH1 . A particular realization of M is " # A B M (z ) = ;R1=2 F R1=2
#
548
where
DISCRETE TIME CONTROL
RD = 2 I ; D D R = RD ; B XB F = (RD ; B XB );1 (B XA + D C )
and X = X
0 is the stabilizing solution of ; ; As X (I ; BRD1 B X );1 As ; X + C (I + DRD1 D )C = 0 where As := A + BR;1 D C .
D
Similar to the continuous time, we have the following theorem.
Theorem 21.23 Let G(z) := A B 2 RH1 with D full row rank and G(ej )G (ej ) > C D 0 for all . Then, there exists a transfer matrix M 2 RH1 such that M M = GG . A particular realization of M is
M (z ) = A BW CW DW
where
"
#
"
#
BW = APC + BD DW DW = DD CW = DW (DD );1 (C ; BW XA)
and
APA ; P + BB = 0 A XA ; X + (C ; BW XA) (DD );1 (C ; BW XA) = 0:
21.5 Discrete Time H2 Control
Consider the system described by the block diagram
z
G
w
y

K
u
21.5. Discrete Time H2 Control
The realization of the transfer matrix G is taken to be of the form
549
2 A B B 3 " # 1 2 6 C1 D11 D12 7 =: A B : G(z ) = 4 5 C D
C2 D21
0
The following assumptions are made: (A1) (A B2 ) is stabilizable and (C2 A) is detectable (A2) D12 is full column rank with D12 D? unitary and D21 is full row rank with " # D21 unitary ~ D? (A3) (A4)
h
i
"
"
A ; ej I B2 C1 D12 A ; ej I B1 C2 D21
#
has full column rank for all 2 0 2 ] has full row rank for all 2 0 2 ].
#
kTzw k2 .
The problem in this section is to nd an admissible controller K which minimizes
Denote
Let X2 0 and Y2 and
0 be the stabilizing solutions to the following Riccati equations: Ax (I + X2 B2 B2 );1 X2 Ax ; X2 + C1 D? D? C1 = 0 ~ ~ Ay (I + Y2 C2 C2 );1 Y2 Ay ; Y2 + B1 D? D? B1 = 0:
Ax := A ; B2 D12 C1
Ay := A ; B1 D21 C2 :
Note that the stabilizing solutions exist by the assumptions (A3) and (A4). Note also that if Ax and Ay are nonsingular, the solutions can be obtained through the following two simplectic matrices:
;1 ( ;1 H2 := Ax + B2 B2 ;Ax ) C1 D?D? C1 ;B2 B2 (A1x ) ;(Ax ) 1 C1 D? D? C1 (Ax ); ;1 C ;1 ~ ~ J2 := Ay + C2;12 Ay ~ B1~D? D?B1 ;C2 C21Ay : ;Ay B1 D? D? B1 A; y
"
#
"
#
550 De ne
DISCRETE TIME CONTROL
Rb F2 F0 L2 L0
:= := := := :=
I + B2 X2 B2 ;(I + B2 X2 B2 );1 (B2 X2 A + D12 C1 ) ;(I + B2 X2 B2 );1 (B2 X2 B1 + D12 D11 ) ;(AY2 C2 + B1 D21 )(I + C2 Y2 C2 );1 (F2 Y2 C2 + F0 D21 )(I + C2 Y2 C2 );1
and
AF2 := A + B2 F2 C1F2 := C1 + D12 F2 AL2 := A + L2C2 B1L2 := B1 + L2 D21 ^ A2 := A + B2 F2 + L2C2 " # AF2 B1 + B2 F0 Gc (z ) := C1F2 D11 + D12 F0
B1L2 Gf (z ) := 1=2 AL2 1=2 (L0 D21 ; F0 ) : Rb (L0 C2 ; F2 ) Rb
"
#
Theorem 21.24 The unique optimal controller is " ^ # A2 ; B2 L0 C2 ;(L2 ; B2 L0) : Kopt(z ) := F2 ; L0 C2 L0 Moreover, min kTzw k2 = kGc k2 + kGf k2 . 2 2 2
Remark 21.6 Note that for a discrete time transfer matrix G(z) = A B 2 RH2, C D
its H2 norm can be computed as kG(z )k2 = TracefD D + B LoB g = TracefDD + CLc C g 2 where Lc and Lo are the controllability and observability Gramians
"
#
ALcA ; Lc + BB = 0 A Lo A ; Lo + C C = 0: Using the above formula, we can compute min kTzw k2 by noting that X2 and Y2 satisfy 2
the equations
AF2 X2 AF2 ; X2 + C1F2 C1F2 = 0 AL2 Y2 AL2 ; Y2 + B1L2 B1L2 = 0:
21.5. Discrete Time H2 Control
551
For example, kGc k2 = Trace f(D11 + D12 F0 ) (D11 + D12 F0 ) + (B1 + D2 F0 )X2 (B1 + D2 F0 )g 2 and kGf k2 = Trace Rb f(L0 D21 ; F0 )(L0 D21 ; F0 ) + (L0 C2 ; F2 )Y2 (L0 C2 ; F2 ) g : 2
~
Proof. Let x denote the states of the system G. Then the system can be written as
x = Ax + B1 w + B2 u _ z = C1 x + D11 w + D12 u y = C2 x + D21 w:
De ne := u ; F2 x ; F0 w then the transfer function from w (21.24) (21.25) (21.26) to z becomes
z = AF2 B1 + B2 F0 B2 C1F2 D11 + D12 F0 D12
where
"
#" #
w = G w + UR1=2 c b
; AF2 B2 Rb 1=2 : U (s) := ; C1F2 D12 Rb 1=2 It is easy to shown that U is an inner and that U Gc 2 RH? . Now denote the transfer 2 function from w to by T w . Then
"
#
1 Tzw = Gc + URb =2 T w
and
for any given stabilizing controller K . Hence if the states (x) and the disturbance (w) are both available for feedback (i.e., full information control) and u = F2 x + F0 w, then T w = 0 and kTzw k2 = kGc k2 . Therefore, u = F2 x + F0 w is an optimal full information control law. Note that = T w w T w = F` (G K )
2 1 kTzw k2 = kGc k2 + Rb =2 T w 2 kGc k2 2 2 2
G y
w

K
u
2 A B B 3 1 2 G = 6 ;F2 ;F0 I 7 4 5
C2 D21
0
552
DISCRETE TIME CONTROL
Since A ; B2 (;F2 ) = A + B2 F2 is stable, from the relationship between an output estimation (OE) problem and a full control (FC) problem, all admissible controllers for G (hence for the output feedback problem) can be written as
K = F` (Mt KFC )
2 6 A + B2F2 Mt = 6 ;F2 6 4
C2
0 0
h
I
hI ;B2i h0 Ii
0 0
i3 7 7 7 5
where KFC is an internally stabilizing controller for the following system:
h i3 2 A B1 I 0i7 6 h ^ 6 G = 6 ;F2 ;F0 h 0 I i 7 : 7 4 5
C2 D21
0 0
^ Furthermore, T w = F`(G F` (Mt KFC )) = F` (G KFC ). Hence 1 ^ min kTzw k2 = kGc B1 k2 + K kRb =2F` (G KFC )k2 : 2 2 min 2 K FC Next de ne ~ G :=
" 2 6 6 6 4
# 1 Rb =2 0 G ^ 0 I
= and
0 0 7 5 ;1=2 0 0 Rb h i3 A B1 I 0i7 h 1 1 ;Rb =2 F2 ;Rb =2F0 h 0 I i 7 7 5 D21 C2 0 0
2I 0 60 I 4
3
~ KFC := Then it is easy to see that
"
I
0 KFC : 1 0 Rb =2
#
1 ^ ~ ~ Rb =2 F`(G KFC ) = F` (G KFC )
and
KFC
1 ^ ~ ~ min kRb =2F` (G KFC )k2 = min kF`(G KFC )k2 : ~ K
FC
21.6. Discrete Balanced Model Reduction
553
L2 ~ ~ ~ A controller minimizing kF`(G KFC )k2 is KFC = 1=2 L since the transpose (or Rb 0 ~ KFC ) is a full information feedback problem considered at the beginning ~ dual) of F` (G of the proof. Hence we get
~ F` (G and
"
#
"
L2 1=2 L ) = Gf Rb 0
#
Finally, the optimal output feedback controller is given by
min kTzw k2 = kGck2 + kGf k2 : 2 2 2 K
K = F` (Mt
"
L2 ) = A + B2 F2 + L2 C2 ; B2 L0 C2 L2 ; B2 L0 : L0 ;F2 + L0 C2 L0
# "
#
The proof of uniqueness is similar to the continuous time case, and hence omitted. 2 It should be noted that in contrast with the continuous time the full information optimal control problem in the discrete time is not a state feedback even when D11 = 0. The discrete time H1 control problem is much more involved and it is probably more e ective to obtain the discrete solution by using a bilinear transformation.
21.6 Discrete Balanced Model Reduction
In this section, we will show another application of the LFT machinery in discrete balanced model reduction. We will show an elegant proof of the balanced truncation error bound. Consider a stable discrete time system G(z ) and assume that the transfer matrix has the following realization:
G(s) = A B 2 RH1 : C D
Let P and Q be two positive semide nite symmetric matrices such that
"
#
Without loss of generality, we shall assume that
APA ; P + BB 0 A QA ; Q + C C 0: P =Q=
(21:27) (21:28)
"
1 0
0 2
#
554 with
DISCRETE TIME CONTROL
1 = diag( 1 Is1 2 Is2 : : : r Isr ) 0 2 = diag( r+1 Isr+1 r+2 Isr+2 : : : n Isn )
0
where si denotes the multiplicity of i . (Note that the singular values are not necessarily ordered.) Moreover, the realization for G(z ) is partitioned conformably with P and Q:
2A A B 3 11 12 6 A21 A22 B1 7 : G(s) = 4 2 5
C1 C2 D
Theorem 21.25 Suppose 1 > 0. Then A11 is stable. Proof. We shall rst assume 2 > 0. From equation (21.28), we have A11 1 A11 ; 1 + A21 2 A21 + C1 C1 0: Assume that is an eigenvalue of A11 then there is an x = 0 such that 6
A11 x = x:
Now premultiply x and postmultiply x to equation (21.29) to get (j j2 ; 1)x 1 x + x A21 2 A21 x + x C1 C1 x 0: It is clear that j j 1. However, if j j = 1, say = ej for some , we have (21:29) (21:30)
A21 x = 0
C1 x = 0: x = ej
0
These equations together with equation (21.30) imply that
"
A11 A12 A21 A22
#" #
" #
x
0
i.e., ej is an eigenvalue of A. This contradicts the stability assumption of A, so A11 is stable. Now assume that 2 is singular, and we will show that we can remove all those states corresponding to the zero singular values without changing the system stability. For that purpose, we assume 2 = 0. Then the inequality (21.27) can be written as
"
A11 1 A11 ; 1 + B1 B1 A11 1 A21 + B1 B2 A21 1 A11 + B2 B1 A21 1 A21 + B2 B2
#
0:
2 A = A11 A12 0 A22 " # A11 1 A11 . kGk1 kDk + 2 n X C1 D i=r+1 .28) are satis ed. This proves that we can assume without loss of generality that 2 > 0. 1 + C1 C1 0: Remark 21.26 Suppose Gr = A11 B1 . Substitute this A matrix into the inequality (21.6. i.28) with 1 > 0.32) implies that Hence we have and the stability of A11 and A22 are ensured. A11 1 A11 . i=1 1 Balanced in the sense that the same inequalities as (21. 1 + A21 2 A21 + C1 C1 0: But these equations imply that hold.28).21. 1 + A12 2 A12 + B1 B1 0 A11 1 A11 . Then kG . 1 + B1 B1 0 A11 1 A11 . 1 + B1 B1 0 A21 1 A21 + B2 B2 0: A21 = 0 B2 = 0: But inequality (21. 1 + C1 C1 0 n X Theorem 21. and we obtain The subsystem with A11 still satis es inequalities (21. " # ~ In particular.27) and (21. Gr k1 2 i.7 It is important to note that the realization for the truncated subsystem Gr = A11 B1 C1 D " # is still balanced in some sense1 since the system parameters satisfy the following equations: A11 1 A11 . Discrete Balanced Model Reduction This implies that and 555 (21:31) (21:32) A11 1 A11 .27) and (21.
1: Without loss of generality. h i. The proof of the theorem 1 will follow from the following two lemmas and the bounded real lemma which establishes the relationship between the H1 norm of a transfer matrix and its realizations. a constant matrix X is said to be contractive or a contraction if kX k 1 and strictly contractive if kX k < 1).1 B1 B2 #2 0 I 6I 0 4 0 0 I 0 0 7: 5 3 " A22 . The lemma below shows that for any stable system there is a realization such that " # h i A A B is a contraction. there is another realization such that C is a contraction.1 # . and. P . we shall assume n = 1. similarly.e.1=2 AP 1=2 P .1=2 B = i " . Lemma 21. Proof.1A11 A22 A21 B2 and 2 A .1 B1 # B2 .1 A11 A21 . we have DISCRETE TIME CONTROL kG . But h A B i" P i 0 I 0 #" A B # P h Hence P .1 C1 C2 are contractive.27). Gr k1 2 r = n .1 B1 A12 .1 A 3 6 A21 .1=2 B is a contraction.1 A12 A22 .27 Suppose that a realization of the transfer matrix G satis es P = Q = diagf 2 I g then " .1 A22 7 4 11 12 5 .556 n I = I .1 A12 . It will be seen that it is more convenient to set = 1=2 . We will prove that for 2 = Then the theorem follows immediately by scaling and recursively applying this result since the reduced system Gr is still balanced.1=2 AP 1=2 P .1A11 A21 . Since P = " 2 0 0 I # satis es the inequality (21. (Note that in the following.
Md is a contraction. Note that " # 2 A11 0 B1 3 Gr = A11 B1 = 6 0 0 0 7 : 5 4 C1 D C1 0 D Hence 3 2 1 p2 B1 A11 0 0 0 7 6 0 0 0 6 0 0 7 7 6 1 (G . 2 We can now prove the theorem. i. p2 X22 0 3 7: 0 7 7 1 p2 Y22 7 .Z Considering X and Y are contractive. Dilate M to the following matrix: 0 . The other part follows by a similar argument. Proof of Theorem 21.6. Discrete Balanced Model Reduction is also a contraction. Then " # 1 p2 X11 X12 0 1 1 p2 Y11 p2 X22 Z 1 0 Y21 p2 Y22 is also contractive (strictly contractive).26.28 Suppose that X = X11 X12 Z X22 tive (strictly contractive). we can easily verify that Md Md I . 07 4 5 0 I 0 I 2 0 6 p1 Y11 6 Md = 6 2 6 Y21 4 1 p2 Y11 1 p2 X11 1 p2 Y22 Z X12 1 p2 X22 1 .I 0 1 I 7 6 . 557 Lemma 21. G ) = 6 0 1 0 A11 A12 p2 B1 7 : 7 6 r 6 2 7 6 0 0 A21 A22 p12 B2 7 5 4 1 1 1 . 2 6 M =6 4 Y Z and Y = 11 Y21 Y22 " # 2 are contrac 3 7 7 5 Proof. 5 1 p2 X11 . 0 . p2 C1 0 p2 C1 p2 C2 0 Now apply the similarity transformation 2 3 .21.e. 0 0 6 7 6 7 T = 6 0 1 .
However.1 2 2 and " # " i " A11 . We shall only consider the special case where the normalized right coprime factors are used.1 1 A 1 A12 0 0 11 2 12 2 6 1 A21 . Gr ) as a constant matrix is a contraction. the theorem follows if we can show that 1 the realization for 2 (G .1 A11 6 1 1 1 .7 Model Reduction Using Coprime Factors In this section. Gr ) to get 2 3 2 A . .1 1 A22 1 A21 1 A22 1 B2 7 7 62 2 2 2 2 6 0 1 (G .1 A12 .1 C1 0 0 2 2 It is easy to verify that as a constant matrix the right hand side of the above realization 1 for 2 (G .1 According to Theorem 21. G ) = 6 1 .1 2 21.1 p2 A21 p2 B2 7 7: A 7 ^ 6 pA (21:34) Md = 6 2 21 .1 B1 3 p2 .27 and 21.1 p12 A12 # . we consider lower order controller design using coprime factor reduction.558 DISCRETE TIME CONTROL to the realization of 1 (G .1A12 .12 (a) and (g).1 B1 7 : 7 r 7 6 1 . the contractiveness of Md follows from Lemmas 21.33) guaranteed if M are contractive.1 2 1 2 2 6 A21 1 A21 1 A22 1 B2 7 A22 5 42 2 2 2 2 1 C2 1 C2 . Gr ) can be written as 20 0 I 03 2 3 6 0 p1 I 0 0 7 6 I 0 0 0 0 7 6 7 1 1 2 6 6 I 0 0 0 7 Md 6 0 p2 I 0 p2 I 0 7 7^ 6 7 (21:33) 6 1 60 pI 0 07 60 0 I 0 07 7 4 5 4 5 2 0 0 0 0 I 0 0 0 I where 2 0 1 .1 A12 .28 by identifying 1 1 Z = A22 X11 = p .1 A11 1 . ^ Finally. this is ^ d is a contraction since both the right and left hand matrices in (21.1B1 # h X12 = .1 p 22 6 A11 1 A12 0 0 7 5 4 2 1 p2 C2 0 0 C1 .1 A11 : Y21 Y22 = 1 1 1 p2 A21 p2 B2 p2 C2 X22 C1 .1 A12 Y11 = p A21 .
i (Y X ) 2 i = i (PQ) = 1 + i (Y X ) 2A A B 3 6 A11 A12 B1 7 F = h F1 F2 i : G = 4 21 22 2 5 C1 C2 D Then the reduced coprime factors " # 2 A11 + B1F1 B1Z .21 that there exists a normalized right coprime factorization G = NM .7.30 M . " # " ^ # Lemma 21.1=2 7 2 RH1 4 F1 5 ^ N C1 + DF1 2 DZ . Model Reduction Using Coprime Factors Suppose that a dynamic system is given by 559 G= A B C D " # and assume that the realization is stabilizable and detectable.1=2 3 ^ M := 6 Z . that N Lemma 21.1 such " # M is inner. M ^ N N X i r+1 X i r+1 1 i p1 + 2 : i .29 Let given by p (Y X ). Now partitioning the system G and matrix F accordingly. i i = i= Then the Hankel singular values of i p1 + 2 < 1: i " M N # are Proof.21. Recall from Theorem 21.1=2 i =2 satisfy the following error bound. This is obvious since and i (Y X ) 0. 2 It is known that there exists a transformation such that X and Y are balanced: " # 1 0 X=Y = = 0 2 with 1 = diag 1 Is1 : : : r Isr ] > 0.
e. we would like to point out that unlike the continuous time case. we can prove a more general result. X + C C A XB + C D = 0: B XA + D C B XB + D D .e.560 DISCRETE TIME CONTROL 2 Proof. Suppose that there exists an X = X " # " 0 such that A XA . Nr Nr I. Lemma 21. Analogous to the continuous time case. + BB = 0 Nr = A11 B1 C1 D " # is stable and contractive. Remark 21.8 It should be understood that the reduced model can be obtained by directly computing X and P and by obtaining a balanced model without solving the Riccati equation for Y . N N = I . . I # (21:35) Then N is an inner. However. the reduced coprime factors in discrete time may not be normalized.. Moreover. ~ This reduced coprime factors combined with the robust or H1 controller design methods can be used to design lower order controllers so that the system is robustly stable and some speci ed performance criteria are satis ed. i. In fact. then the truncated system A A . if the realization for N= is also balanced with " # 2 A11 A12 B1 3 A B =6 5 4 A21 A22 B2 7 C D C1 C2 D X= = " 1 0 0 2 # and with 1 > 0..31 Let a realization for N (z) 2 RH1 be given by N (z ) = A B C D with A stable. i. We will leave the readers to explore the utility of this model reduction method.
8 Notes and References The results for the discrete Riccati equation are based mostly on the work of Kucera 1972] and Molinari 1975]. Notes and References Proof.35) by U to get " or A11 1 A11 .21. The normalized coprime factorizations are obtained by Meyer 1990] and Walker 1990]. Dooren 1981]. Premultiply equation (21.. Numerical algorithms for solving discrete ARE with singular A matrix can be found in Arnold and Laub 1984]. 1 + C1 C1 A11 1 B1 + C1 D = . I B2 " # " # " # A 2 B21 2 0: # : This gives " A11 B1 C1 D #" 1 0 0 I #" A11 B1 C1 D # " . I + B2 2 B2 A11 1 A11 . . The matrix factorizations are obtained by Chu 1988]. The detailed treatment of discrete time H1 control can be found in Stoorvogel 1990].1 D # # 1 2 = 0 (although the converse A21 i. kNr k1 1. A21 B1 1 A11 + D C1 B1 1 B1 + D D .8. and Walker 1989]. and then 1 " TA11 T . Limebeer.e. 1 + C1 C1 + A21 2 A21 A11 1 B1 + C1 D + A21 2 B2 = 0 B1 1 A11 + D C1 + B2 2 A21 B1 1 B1 + D D .35) by U= 561 "h I 0 0 i 0 # # I and postmultiply equation (21.1 TB1 C1 T . 1 0 0 I Now let T = 1=2 . and Iglesias and Glover 1991]. " 2 21. and references therein. It is clear that Nr is inner if B2 may not be true). Green.
562 DISCRETE TIME CONTROL .
B. Franklin (1987). D. Automat. pp. pp..Bibliography 1] Adamjan.. \In nite block Hankel matrices and related extension problems. D." AMS Transl. Liu (1989). 13.Agathoklis. 2. 266. B. Moore (1989). and Y. Krein (1971). New Jersey. and J. and M.G.Mansour (1986). PrenticeHall. Moore (1979).261 .. Krein (1978). 4] AlSaggaf.. No. pp.. 10] Anderson. 171192. 802812. Contr. 2] Adamjan. D. No. F. 6] Anderson. Vol. Control." IEEE Trans. pp.M. M. Optimiz. Vol. Auto. Automat. P." SIAM J." IEEE Trans. B. \Analytic properties of Schmidt pairs for a Hankel operator and the generalized SchurTakagi problem. 563 . 9] Anderson. vol.O." IEEE Trans. 8] Anderson. Automat. B. pp. 410414. U. Contr. 687692. Arov. 6.CAS33. New Jersey. 5] Anderson. " IEEE Trans.. D. AC34. Vol. and M.Z.O. (1967). pp. Vol. O.M. and J.Jury and M. pp. B. 1625." IEEE Trans. B. Contr. 15. 3173. Contr. E.AC32. O. O. and G. No. Englewood Cli s. Franklin (1988).. (1993). Arov. 7] Anderson. B. Vol. O. 7. \Model reduction via balanced realizations: an extension and frequency weighting techniques. 815819. No. \An error bound for a discrete reduced order model of a linear multivariable system. \An algebraic solution to the spectral factorization problem. 111. Vol. 11] Anderson. V.Z. D. D. 8.G.D. \A system theory criterion for positive real matrices. V. (1967).D. pp. and G.. ." Math. \Stability and the Matrix Lyapunov Equation for Discrete 2Dimensional Systems. O. D.I. \Bode Prize Lecture (Control design: moving from theory to practice). Optimal Filtering. PrenticeHall.4. pp. M. \Controller reduction: concepts and approaches. B." IEEE Control Systems. Vol. F. Englewood Cli s. B. Optimal Control: Linear Quadratic Methods. 3] AlSaggaf. U. AC12.AC33. 133156. Vol. USSR Sbornik.
n) which contains most classical interpolation theory. and A. No. Analysis and Synthesis Toolbox. Balakrishnan. 17] Ball. J. and C. (1990). 22] Bode. Silverman. T. 283. Dec. Doyle. pp. 13] Arnold.W. 24] Boyd. S. 2. V. Cohen (1987). Op. M. 14] Balas. 107142. C. X. vol. Norman (1988). Robust Control of Flexible Structures: Theory and Experiments. pp. J. Theory. \A bisection method for computing the H1 norm of a transfer matrix and related problems. D. H. Boston." Proceedings of the IEEE. \Optimal approximation of continuoustime systems.S. \Generalized Eigenproblem algorithms and software for algebraic Riccati equations. Academic Press. Safonov (1980). AC34. Haddard (1989). and G. Auto. Systems and Control: Foundations and Applications. Van Nostrand. T. Network Analysis and Synthesis: A Modern System Theory Approach. 785816. No. G. M." Int. 19] Basar. \A BeurlingLax theorem for the Lie group U(m. Meyer. Control. ." IEEE Trans. V. Glover.. Li. MUSYN Inc. Contr. J.M. K.AC33." Math. Vongpanitlerd (1973). and P.P. pp. 17461754..A. New Jersey." IEEE Trans. H. J. PrenticeHall. 195198. W.. Barratt.268 . and N. IEEE Conf. EnglewoodCli s. pp. 72. 12. vol. Barratt (1991). Linear Controller Design { Limits of Performance. 20] Bernstein. and Systems.. Auto. N. A. Thesis. Control. Smith (1991). B. and W. pp. Khraishi.. \A New CAD Method and Associated Architectures for Linear Controllers. S. \LQG control with an H1 performance bound: A Riccati equation approach. Caltech. G. Vol. and R. H1 Optimal Control and Related Minimax Design Problems: A Dynamic Game Approach. 18] Basar. and P. vol." Proc. C. 15] Balas. Packard. Vol. and S. 3.564 BIBLIOGRAPHY 12] Anderson.. Control. Dynamic Noncooperative Game Theory. 9. Prentice Hall. Contr. Bernhard (1991). \Sensitivity minimization in an H1 norm: parameterization of all suboptimal solutions." J. A. Kabamba (1989). 207220.A. Inc. S. 23] Boyd. W. PhD. and The MathWorks. 25] Boyd. 293305. Network Analysis and Feedback Ampli er Design. vol. 21] Bettayeb. Birkhauser. Signals. M. Albuquerque. G. Helton (1983). F. J. L. 16] Ball. pp. S. and J. and M. G. 46. D. D. Olsder (1982). J. Princeton. Laub (1984). O. (1945). Inc. D. Balakrishnan. pp.
Holt. and YC.. 35. and K.. Auto. J. Contr.BIBLIOGRAPHY 565 26] Boyd. J. Contr. T. 2. Auto. (1992b). and M. 32] Chen." Proc.. C. 36... pp. Auto. no.A. Linear System Theory and Design. and C. W. Ho (1975). 39] Chu. No. No. \Necessary and su cient condition for robust stability of linear distributed feedback systems. Vol. Amer. Auto. F. pp. 28] Brasch. \Pole placement using dynamical compensators.. pp. and J. \Multivariable gainphase and sensitivity integral relations and design tradeo s. \Structured and simultaneous Lyapunov functions for system stability problems. 22152240. \ H2 optimal sampleddata control. \On discrete innerouter and spectral factorization. 387397. Yang (1989). vol. Vol. (1992a). Modern Control Theory. and B. Vol. Vol. 3rd Ed. 335343." Int. M. Contr. pp. H1 optimization and robust multivariable control. 153170. C." Int. Steinbuch (1990).. J. and Info. 38] Chu. 15. (1991). J. 29] Brogan. 255267. 6. C. Glover (1989). J. E. S. 30] Bruinsma.A. pp." IMA J. 33] Chen. N. Control. 3443. 31] Bryson Jr. J. 797846. 2. Englewood Cli s. Desoer (1985). J. C. 34] Chen. \A fast algorithm to compute the H1 norm of a transfer function matrix"." Systems and Control Letters. Conf." submitted to IEEE Trans. Math. L. \Sensitivity integral relations and design tradeo s in linear multivariable feedback systems. (1985). 14. Francis (1992). PhD Thesis. vol. D. A. 36] Chen. B. M. pp. Prentice Hall. Rinehart and Winston." IEEE Trans." IEEE Trans. \Spectral factorization by Hermitian pencils. 287293. \Subharmonic functions and performance bounds in linear timeinvariant feedback systems. D. Contr. A. Systems and Control letters. University of Minnesota." Linear Algebra and its Applications. and C. \Rational spectral factorization using statespace methods. T. Control. 49. . J. Applied Optimal Control. Contr. (1984). 122124. 20. 37] Chen. 35] Chen. Vol.. 41] Clements. Contr. Desoer (1982).. Personal communication. (1995). 27] Boyd. A. pp." submitted to IEEE Trans. (1993). S. Hemisphere Publishing Corporation. Pearson (1970). vol. 40] Clements. 4. C. New Jersey. pp. and Q. (1988).
Y." IEEE Trans. 45] Davis. 56] Doyle. J. AC32. C. pp. Genin. pp. \Guaranteed Margins for LQG Regulators. (1982). \A transformation approach to stochastic model reduction. A. U. Auto. M. \`1 optimal feedback controllers for MIMO discrete time systems.133.. W. . C. No. A. E. pp. pp.. Academic Press. Contr. Vol. Pearson (1987). Liu." IEEE Trans. J." ONR/Honeywell Workshop.445 469. P. 121135. Vol. pp. \Feedback system design: the fractional representation approach to analysis and synthesis. and D. vol. 314 322. 43] Dahleh. Comput. W." IEEE Trans. B. pp. and R. Pal (1984). Contr. pp. \A generalized eigenvalue approach for solving Riccati equations.." IEEE Trans. Kahan. Vidyasagar (1975). and J. Automat. \Analysis of feedback systems with structured uncertainties. \Matrix quadratic equations." IEE Proceedings. Stat. W. C. Lauderdale. Kouvaritakis. Vol. Vol. Auto. (1981). Francis. pp. New York." IEE Proceedings.. Anal. Ft. \Lecture notes in advances in multivariable control.45 . Applied Math. 44] Daniel. 49] Desoer. B. AC25. W. A. (1968)." SIAM J.. 4761. Tannenbaum (1992). V. C. 10971100. J. Auto. C. (1985). 377 401. Soc. Part D. 2. 46] Davison. AC23. 756 757. pp. \Principal direction alignment: a geometric framework for the complete solution to the problem.. 12. 133. Austral. 53] Doyle. R. and H." SIAM J. and A. and M.. Weinberger (1982). (1978). and H.. 6. J. 2. J. 6. No. Feedback Control Theory. 2.. 54] Doyle. Part D. Vol. R. No.. Feedback Systems: InputOutput Properties. IEEE CDC. 36. 51] Dooren. \Structured uncertainty in control system design". P. (1984). No. Numer. 52] Doyle. 4. No. B. 48] Desai." SIAM J.C.566 BIBLIOGRAPHY 42] Copple.. Contr." IEEE Trans. Vol. and Y. Murray. NO.. A. Math. Vol. Macmillan Publishing Company. No. Saeks (1980). Latchman (1986). 56. J. B. M. Contr. Contr. 399 412. 55] Doyle. 747748." Bull. 4. pp. F. pp.. 47] Delsarte. Vol 29. vol 19. 10. J. Vol. Auto. \The NavanlinnaPick problem for matrix valued functions. 4556. C. 50] Desoer. C. Minneapolis. \On pole assignment in multivariable linear systems.. 13. \Normpreserving dilations and their applications to optimal error bounds. (1974). Sci. Kamp (1979).
J. 8. \The gap metric: Robustness of stabilization of feedback systems. \ Statespace solutions to standard H2 and H1 control problems." IEEE Trans. Auto.A. 1988. 64] Doyle. K.C. Decision Contr. Pennsylvania. J. \Control of Plants with Input Saturation Nonlinearities." Proc. Francis (1989)." Proc. 22182223. H. 30. C. on Automat. NV.BIBLIOGRAPHY 567 57] Doyle. 68] Fan." IEEE Trans. 62] Doyle. 12271232 60] Doyle. H. (1984a). K. C. J. 63] Doyle. Bodenheimer (1989). \Multivariable feedback design: Concepts for a classical/modern synthesis. pp. vol. Dec. vol. 8. California. D.. Dissertation. Bodenheimer (1994). Auto. (1985).D. of American Control Conference. 69] Fan. K. Zhou." IEEE Trans. 21st IEEE Conf. Feb. 59] Doyle. Control. LMIs and . December 1986. Control. AC31. no. pp. Model Reduction for Control System Design. . 23rd Conf. no. J. \Mixed H2 and H1 performance objectives II: optimal control.. Auto. A. and D. (1984b) \Model reduction with balanced realizations: An error bound and a frequency weighted generalization. Dec.C. J. 39. \Optimal Control with mixed H2 and H1 performance objective. J. 61] Doyle. Doyle (1991). pp 10341039. Automat. and A. Stanford. M. vol.. and J.P. C. No. \Design Examples using synthesis: Space Shuttle Lateral Axis FCS during reentry. J." in Proc. B. \Characterization and e cient computation of the structured singular value. K." IEEE Trans. 2538. Wall and G. Stein (1981). vol. vol. Packard (1986).. Enns (1987). \Performance and robustness analysis for structured uncertainty. and B. pp. 66] Enns. 67] Enns. pp. 629636. Contr. Atlanta. Pittsburgh. 8. pp. 30th IEEE Conf. \Robustness in the presence of mixed parametric uncertainty and unmodeled dynamics. June. P. Tits (1986). Stein (1982).C. and A.F.. M. AC34. C.. 1981. Las Vegas. . 20652070. Glover. Stanford University. \Review of LFTs.. 58] Doyle.Packard and K. pp.L. pp. Ph." IEEE CDC. R. pp. K. Contr. K. pp. 416. 65] ElSakkary. J." IEEE Trans.L. 240247. 831847.. 734743. Control. K. and B.. England.. Control.S. Auto." American Control Conf. Contr.. and G. Contr. J. 1." IEEE Trans. AC26. A.. Glover. Zhou (1991). Also see 1988 American Control Conference. Khargonekar. Zhou. Vol. Tits. pp. 15751587. D. Department of Aeronautics and Astronautics. Lenz. Smith. and A. C. AC36. no." Proc.
pp." SIAM J. \All optimal Hankelnorm approximations of linear multivariable systems and their L1 error bounds. 88." Int." IEEE Trans. 104. vol.. \Multiplicative approximation of linear multivariable systems with L1 error bounds. B. Science. K. 79] Georgiou. J.. Control. pp. pp. 82] Glover. 6. 95116. D. and J. I. and D. 673686. pp. B. Berlin: SpringerVerlag. Tannenbaum (1988). 726767. pp. Smith (1990).." Int.B. and M. vol. A course in H1 control theory.. T. vol. Contr. WA. 78] Georgiou. \Optimal robustness in the gap metric. (1988). no. Conf. 3. Contr. II: the singular system. B. C.P. \On the computation of the gap metric. 1988. 35. \On the fourblock problem. Lett. Auto. C." Syst. (1989). 81] Glover.P. \Linear control theory with an H1 optimality criterion. 77] Garnett. 11. K. 49. 25. Frequency Domain Properties of Scalar and Multivariable Feedback Systems. 253257. 83] Glover. No. J. pp. Doyle (1987). 741766. part 2: directionally uniform weightings and an example. Contr. 1. C." Integral Equations and Operator Theory. J. Amer. (1986). pp. AC30. J. T. No.. pp. 17051709. Lecture Notes in Contr. \Robustness analysis using singular value sensitivities". J." Operator Theory: Advances and Applications. (1989). 73] Freudenberg. vol. T. Control. pp. and Info." in Data to Model. vol.S. Automat. \Robust stabilization of linear multivariable systems: Relations to approximation. 93112 and "On the four block problem. S. 74] Freudenberg. Seattle.AC35. Vol. . Vol. Int. Willems. Vol. 72] Francis. Control Opt. 75] Freudenberg. 71] Francis. 1984. J. \Right half plane zeros and poles and design tradeo s in feedback systems. vol. (1981). vol. and D. pp. 80] Glover. vol. T. 11151193. J. Contr.S. J.A." Int. 43." IEEE Trans. J. Academic Press. \A tutorial on Hankelnorm Approximation. and A.568 BIBLIOGRAPHY 70] Foias. Contr. 76] Freudenberg. 873903. Bounded Analytic Functions. 555565. K. (1986). no.. 39. Looze (1988). \Analysis and design for illconditioned plants. Lecture Notes in Control and Information Sciences. pp. (1984). 6. Ed. " Proc.P. New York: SpringerVerlag.. (1987). 815844. Cruz (1982). vol. A. Looze and J.S. J. Looze (1985). 11 (1988). no. 32 (1988)..C. 3. K. Contr.
Lancaster. N. Automat.. M. No." SIAM J. Glover (1993). E. Automat. Optimization and Control.G. J. \Statespace formulae for all stabilizing controllers that satisfy an H1 norm bound and relations to risk sensitivity." Int. 98." Journal of Linear Algebra and its Applications. pp. \Derivation of the maximum entropy H1 controller and a state space formula for its entropy. K. 899. Basic Operator Theory. K. J. P. M (1988b). vol. vol. and D. 93] Gohberg. and J. \Performance preserving frequency weighted controller approximation: a coprime factorization approach. \A structured approximation problem with applications to frequency weighted model reduction. 11. Birkhauser. Vol. 167172." Systems and Control Letters. " IEEE Trans. Vol. K. December 1994. Glover (1994). P. 94] Golub. vol. No. 87] Glover. G. J.. Nijmeijer and J. Kasenally. Limebeer. K. and J. Vol. Goldberg (1981). H. Matrix Computations. pp. S. Hung (1992). 961965. Doyle (1989). and K. Lecture Notes in Control and Information Sciences. 96] Green. Md. Doyle (1988). pp.. 29032908. Press. pp. Boston. I. pp. vol. H. Schumacher (Eds. pp.AC33. D. no. 95] Green. (1988a). \A state space approach to H1 optimal control. Limebeer. 135. 4. Control and Optimization. and M. 88] Glover. Doyle. and S.." Proceedings of the 33nd Conference on Decision and Control. 50... Baltimore. 13231334. Vol. 10.C. Van Loan (1983).: Johns Hopkind Univ. and D." in Three Decades of Mathematical Systems Theory: A Collection of Surveys at the Occasion of the 50th Birthday of Jan C.. K. and Y. 29.M. December 1993.AC37." SIAM J. I. SpringerVerlag. \A relative error bound for balanced stochastic truncation. Texas.J. Automat. No. 211247. D. F. 283 324. San Antonio. M. \On hermitian solutions of the symmetric algebraic Riccati equation." IEEE Trans. 2. \A characterization of all solutions to the four block general distance problem. Contr. 447465. 89] Glover. Control. November 1986. 1989. 92] Gohberg. C. pp. and C. 86] Glover." IEEE Trans. and L.N. 91] Goddard. J.BIBLIOGRAPHY 569 84] Glover. P. K. 8. 34. .. \Controller reduction: weights for stability and performance preservation. Mustafa (1989). McFarlane (1989). No. 6. pp. 90] Goddard. \Balanced stochastic realizations. vol. Florida. 24. pp. Rodman (1986). and K." Proceedings of the 32nd Conference on Decision and Control. Contr. Contr. Safonov (1991).). Willems. Orlando.821830. J. \Robust stabilization of normalized coprime factor plant descriptions with H1 bounded uncertainty. 85] Glover.
\System structure and singular control. Automatic Control. Vol. CAS28. 104] Horn. S. R. (1989). J. 101] Heymann. \An improved error estimate for reducedorder models of discrete time system. 111] Hyland. Y. No. A. A. London. Bernstein (1984). Auto.J. W. 108] Hung. R." IEEE Trans. 748749. 13. 369402. S. K. D. 165172. Doyle (1988). Topics in Matrix Analysis. 28. R. Banach Spaces of Analytic Functions. I. M. \Broadband gain equalization directly from data. Contr. K.. . Y. 109] Hung. Prenticehall. (1975).. (1962). pp. pp.\H1 optimal control. Vol. SpringerVerlag. 110] Hyde. and A." Syst. Silverman (1983). 49. N0.. John Wiley & Sons. Vol." SIAM J. 7 pp. Multivariable feedback: A quasiclassical approach. and K. M. pp. 1034 1037. Vol. 38. \Comments \On pole assignment in multiinput controllable linear systems. A. 106] Horowitz. Pritchard (1990). 11. J.. Contr. Lett. Contr. Vol. \The Application of Scheduled H1 Controllers to a VSTOL Aircraft. Englewood Cli s.C. (1981).570 BIBLIOGRAPHY 97] Green. 100] Hestenes. Control. and C. Inc. (1968). Control and Optimization. Y. 98] Hautus.Part II: solution for controllers. pp. J. 105] Horn." Linear Algebra and its Applications. 102] Hinrichsen. 107] Hung..Part I: model matching. and C. Auto. C. S." IEEE Transactions on Automatic Control. Matrix Analysis. D. 6. 103] Ho man. MacFarlane (1982). Johnson (1991). L. 13501371. First Published in 1985. pp. R. and D.N. \Optimal Hankelnorm approximation of stable systems with rst order stable weighting functions. (1963). R. 317320." Int. S. pp. Glover. J.J. . Vol. No. pp. Glover (1993). M. M. and L.. \A Jspectral factorization approach to H1 control. Limebeer and J. J. AC35. Vol. Cambridge University Press." IEEE Trans. N. Glover (1986). 7.. Circuits Syst. Vol. AC29. Johnson (1990). 11251137. Optimization Theory: the nite dimensional case. 12911359. Cambridge University Press. D. 50." IEEE Trans. G. Academic Press. and A. Synthesis of Feedback Systems. M. 99] Helton. pp." " IEEE Trans. and K. 10211039. \The optimal projection equations for xedorder dynamic compensation. M.
Circuits. vol.. pp. P. 3. Rotea (1991). (1980). (1968). 18991904. and L. 13. P. Vol. no. pp. (1988). L." Proc. I. Vol. M. Glover (1990). \New results in linear ltering and prediction theory. Juang (1987). Contr. (1964). P. A. AC32.. Englewood Cli s. pp. \Discrete time H1 controllers satisfying minimum entropy criterion. 126] Kleinman. Rotea (1988)." IEEE Transactions on Automatic Control. 115] Jonckheere. and M. and M. \Fast computation of achievable performance in mixed sensitivity H1 design. Contr. D. and J.. J. H. 116] Kailath. 627638.. B. 36." IEEE Trans. 95108. Vol. pp." Proc. P. No. vol. T. 07632.BIBLIOGRAPHY 571 112] Iglesias. Vol. P. D. Contr." IEEE Trans. 123] Khargonekar. \Spectral theory of the linear quadratic optimal control problem: discretetime singleinput case. \Synthesis of H1 Controllers based on Conjugation. S. N. 118] Kalman. Vol. AC33. \H1 ." IEEE Trans." IEEE Trans. pp. 810825. pp.. Control. Auto.optimal control with statefeedback. A. P. 896906. PrenticeHall. Series D: J. 786788. Sontag (1982). 121] Khargonekar. 113] Iglesias. Glover (1991). 83.A. 9. Control. on Auto. and E. on Auto." Int. \Robust stabilization and H1 optimal control.R. Vol. 19911996. 356 361.P. vol. 14. 114115..P. Bucy (1960). \On the relation between stable matrix fraction factorizations and regulable realizations of linear systems over rings. 5. 120] Khargonekar. 125] Kishore. 275286. R. E. Tucson. Mustafa. No. and M. No. A. pp. No. Basic Engineering. Contr. pp. Linear Systems. IEEE Conf. 10311073.. Vol. A. 7. Decision and Control . Austin." Syst. P. Dec. E. E. Zhou (1990). R. Bettayeb (1994)." IEEE Trans. Automat. Control. 114] Jonckheere. pp. \Statespace approach to discretetime H1 control. Inc. Vol. pp. and K. 119] Kavranoglu. 54. A. 27. 122] Khargonekar. 35. \When is a control system optimal?" ASME Trans.J.. \Uniform stability and performance in H1 . Contr. E. 824737. and K. \Mixed H2 H1 control: a convex optimization approach. Series D: J. Basic Engineering. and J.C.A. Syst. CAS25. pp. D. . Contr.. Pearson (1992). 86. P. P. Automat." ASME Trans. pp. pp. 39. 117] Kalman. \Characterization and computation of the solution to the optimal L1 approximation problem. and R. Texas. Silverman (1978). I.. 124] Kimura.. 110. Petersen. \On an iterative technique for Riccati equation computation." IEEE Trans. Auto. Lett. Arizona." IEEE Trans. IEEE Conf. and K. 9.R. Vol. Petersen.
Session FA8E. 141] Liu. (1986). K. no. pp. 3.J." SIAM J. 344347. pp. 134] Latham. 229236. . vol. H. Tampa. Green (1992).25. pp. Linear Optimal Control Systems. and M. 3. and R. Doyle (1987). pp." SIAM J. 646677. 1980 JACC. \Optimal Hankel norm model reduction: Multivariable systems.. \A game theoretic approach to H1 control for time varying systems. P. 262283. \Controller order reduction with guaranteed stability and performance. M. \Discrete time H1 control. Vol. Tismenetsky (1985). Morf and T. and R. IEEE Conf. 130] Kwakernaak. T. Auto. pp. Levy. No... Contr. WileyInterscience. 135] Laub. K. \Frequency weighted optimal Hankelnorm approximation of stable transfer function. J." Syst. 128] Kung. 136] Lenz. D. Hung (1987). D. P. and M. Kawtani (1990). 832852. Dec. pp.N. 945 961.. N. A. Vol. 392396. V." IEEE Trans. No." Proc. Contr. Mita." IEEE Trans. Walker (1989). 132] Kucera. \A polynomial approach to minimax frequency domain optimization of multivariable feedback systems. P. New York. Control and Optimization. V. B. \Computation of 'balancing' transformations. (1972). 133] Lancaster. Discrete Linear Control. and Y. \Parameterization of state feedback H1 controllers. Contr. Academic Press. Control Opt. and G. Control Opt. 117156.N. 2D StateSpace Models { Realization and the Notions of Controllability. Control.51. \An analysis of polezero cancelations in H1 optimal control problems of the rst kind. and John C. vol. and D. M. E. P. Florida." SIAM J." Int. Khargonekar. Lin (1981). D. B. 2nd Ed. New York. 3. Anderson (1986). (1980).. 138] Limebeer. AC26. W. Vol. Kailath (1977). 30. pp. Z. pp.. P. 131] Kucera.. 139] Limebeer. John Wiley and Sons. D. AC17. K." Proc.. J. 535551. 14571493. Observability and Minimality. O.S. 26. No.. " American Control Conference. D. 129] Kwakernaak. D. J.. S. Contr.572 BIBLIOGRAPHY 127] Kung." Proc. Halikias (1988). Lett. Sivan (1972). \New results in 2D Systems Theory. J. pp. Vol. \A controller degree bound for H1 optimal control problems of the second kind. G. 16971698. K. 2.. \A contribution to matrix quadratic equations. and B. J. The Theory of Matrices: with applications. O. and D. Anderson. (1979). Khargonekar. S. 137] Limebeer. Control. H. Green. 5 pp. Automat. N.D. IEEE. S." Int. 140] Limebeer..J.
6. pp. Anderson (1990). O. 233249. Anderson. \Stabilization of LFT systems. England. and B. Vol. Contr. \Generalized H1 control theory. Vol. 2. Vol." Automatica. W. C. Contr. 50. pp. pp. 153] Mariton. M. Y. B. Automat. Contr. Y. A. 221234. Ho (1977). pp. \On the matrix Riccati equation. Mita (1992). and B. Glover (1990).. pp. pp. D. Anderson (1989). Appl. and B.. Vol. 149] Luenberger. K."Optimal. F. 4. 1749. 2. Meth. Optimization by Vector Space Methods. and K. 154] Martensson. No. Y. vol. pp. Inc. . 155] McFarlane.BIBLIOGRAPHY 573 142] Liu. 30th IEEE Conf. Z. \A homotopy algorithm for solving coupled Riccati equations.J. (1971). SpringerVerlag." IEEE Trans." Int. (1989). Vol. and J. Robust Controller Design Using Normalized Coprime Factor Plant Descriptions. O. Automat. 3. Vol. Vidyasagar (1990)." Automatica. No. John Wiley & Sons. 150] MacFarlane. C. Vol. 487497. 151] Maciejowski. vol. 147] Lu. 351357. 35. Lecture Notes in Control and Information Sciences... D. M. J. D. K." Int. pp. 369373. pp. 16.S. Dec. (1971). K. C. \Singular perturbation approximation of balanced systems.. 37.. O. 152] Mageirou. D. 3. D. 26. \Analytic Properties of the singular values of a rational matrix. and J. Doyle (1991). 26. 12391244. Contr. Multivariable Feedback Design. no. G. Y. 148] Luenberger.G. and T. American Control Conference. O. No. Contr. J. Anderson (1986). and Y. 145] Liu. Brighton. 44. pp. Ly (1990). and R. K. No." Proc." IEEE Trans. D. C. and M. Control. Bertrand (1985). (1969). D. 156] McFarlane.. M. pp. 596602. and U." Information Sciences. 507531. vol. \Controller reduction via stable factorization and balancing." Proc. \Coprime factorization controller reduction with Bezout identity induced frequency weighting. 1990. Zhou. 6. D." Int. E. 146] Liu. \Frequency weighted controller reduction methods and loop transfer recovery. Wokingham: AddisonWesley. No. \Decentralized stabilization via game theoretic methods. Vol. 393399. G.. Control. 138. 144] Liu. \An introduction tp observers. J. Hung (1983)." Automatica. \Reducedorder controller design using coprime factor model reduction. 3. J. 143] Liu. Glover. 13. 13791405.
C. 158] Meyer. Minimum Entropy H1 Control. D. Vol. C. 227228. vol. 152166.. 166] Mustafa. Telford (1990). 37. and model reduction. Nemirovski (1994).. 203 208." Annals of Mathematics. SpringerVerlag. vol. \Fractional balanced reduction: model reduction via fractional representation. 396 399. Y. 26." IEEE Trans. Vol. G. \A connection between normalized coprime factorizations and linear quadratic regulator theory. pp. pp. 165] Mustafa. and G. . P. A. 6. no. 161] Molinari. \The construction of special coprime factorizations in discrete time.Balanced Truncation". 669682. 170] Nehari. D.. 162] Moore.. pp. D.. 2. D. SpringerVerlag." IEEE Trans." IEEE Trans. \Controller Reduction by H1 . No. 15.. and R. 32. and K. Auto. M. J. Automat. K.AC36. Glover (1992). T. 168] Naylor." IEEE Trans. No. K. and A. Sell (1982). Automat. D. Vol. p. 5. W. \Synthesis of minimum roundo noise xed point digital lters. Automat. D. Roberts (1976). Linear Operator Theory in Engineering and Science. \The stabilizing solution of the discrete algebraic Riccati equation. Glover. \Principal component analysis in linear systems: controllability. 164] Mullis. 6. No. Automat. and K. Glover (1990). pp. No. Z.. 160] Meyer. (1975).AC36. Automat. (1990). 35. and P. \On bounded bilinear forms. Circuits and Systems." IEEE Trans. Vol. IEEE Trans. Interiorpoint polynomial algorithms in convex programming. Contr." IEEE Trans. 759769. Lecture notes in control and information sciences. Contr. D. Contr.. 551562. pp. and K. (1957). C. 171] Nesterov. observability. 1. Vol. Contr. June 1975. AC35. Auto. (1989). \A loop shaping design procedure using H1 synthesis." IEEE Trans. Automat. 23. B.2. G. pp. Franklin (1987). pp." IEEE Trans. Auto. and G. R. Contr. Vol. No. 163] Moore. Vol. 167] Mustafa." IEEE Trans. P. pp. 153162. Vol. SIAM. No. 588590. and A. B. (1981). pp. 12. 3. \Relations between maximum entropy / H1 control and combined H1 /LQG control. Contr. Glover (1991). 193. Contr. B." Systems and Control Letters.. A. \Filtering and smoothing in an H1 setting. G. 35. Vol... 159] Meyer.574 BIBLIOGRAPHY 157] McFarlane. (1990). No. Contr. pp. Contr. 13411345. Khargonekar (1991). 12. 169] Nagpal. \All stabilizing controllers as frequency shaped state estimate feedback.
Pitman Advanced Publishing Program. M. Auto.. and A. pp. and J. Contr.Jacobson and N. (1978). Contr.. Contr. Automat. \Principal gains and Principal phases in the analysis of linear multivariable feedback systems. J. \On a quotient norm and the SzNagy Foias lifting theorem. C." Automatica. A. \The complex structured singular value. Constant I/O similarity scaling for full information and state feedback control problems. 185] Power. and L. 19. M. A. 184] Postlethwaite. C.831 . Doyle (1988)." 1988 ACC. 382387. A.AC29. J." Syst. C.." IEEE Trans. Mach. Pandey (1993). vol. 29.BIBLIOGRAPHY 575 172] Nett. 177] Packard... I. No. Automat. 182] Petersen. Continuity properties of the real/complex structured singular value.. A. (1960). C. AC26. Vol. \A Connection between StateSpace and Doubly Coprime Fractional Representations. pp. (1982). G. 30. Hankel operators on Hilbert space. A. No. 4. (1987). P. 178] Packard. Balas (1992). Comp. 7. 176] Packard. pp. I. L. Notes on . Contr. vol. E. 174] Packard. and J.N. pp. 415428. 3.J. pp. \Optimal. MacFarlane (1981). and J. J. S. Contr. J. pp. 179] Packard.. 175] Packard. "Model Reduction via Balanced State Space Representation. 271280. 180] Parrott. K." Journal of Functional Analysis. \On Preconditioning of Matrices. Assoc. A complex variable approach to the analysis of linear multivariable feedback systems. pp. Berlin: Springer.R. G. 183] Postlethwaite. Unpublished lecture notes. pp. Silverman (1982). No. pp.. vol. Vol." IEEE Trans. (1991). Leonhardson. 181] Pernebo. S. \Disturbance attenuation and H1 optimization: a design method based on the algebraic Riccati equation. Robust Control of Multivariable and Large Scale Systems. No. 71109. 832. . Atlanta. Lett. MacFarlane (1979)." J. Vol. Zhou.. Auto. 2. C. Vol. I.Balas (1984).. Vol. 3246. C. \Structured singular value with repeated scalar blocks. Doyle (1988). AC32. Final Technical Report for Air Force O ce of Scienti c Research. 1. and A. Edmunds.AC27." IEEE Trans. and P.E." IEEE Trans. 338345. A. 38.A." IEEE Trans. Pandey. 427429. Doyle (1993). G. 311 328. 173] Osborne.
convergence and global optimality. Princeton University Press. 190] Rockafellar. in Multivariable Control." Proc. 47 65.576 BIBLIOGRAPHY 186] Ran. Vreugdenhil (1988). Vol." J. (1970). and E. 922. 2. \Inequalities for a matrix Riccati equation. 307316. Real and Complex Analysis. 8. 189] Richter. \H2 optimal control with an H1 constraint: the state feedback case. Vol. Hartmann (1981). and P. no. M. Contr." American Control Conference. Convex Analysis . T.. \Stability of interconnected systems having slopebounded nonlinearities. 99. A. Automat. 194] Qiu. AC26. R. 2. \Feedback properties of multivariable systems: the role and use of the return di erence matrix. Vol. S. (1975). 200] Safonov. \Minimizing conservativeness of robustness singular values". and R. G. (1980). (1966). (1959).1 10. pp. J. Contr. Automat. 6383. IEE. 37.M. Vol. No. No.. M.. pp. and E.G. Davison (1992a). P. New York. 1. pp." IEEE Trans. Doyle (1984). 188] Redhe er. France. 27. 770780. No." Syst. 6. Conf. Lett. McGraw Hill Book Company. R. 251256. L. Part D. Vol. 197] Safonov. \Pointwise gap metrics on transfer matrices. W. M. Laub. J. Stability and Robustness of Multivariable Feedback Systems. MIT Press. S. and G. vol. Contr. Auto.. \A Discrete State Space Model for Linear Image Processing. New Jersey. 187] Redhe er. Contr. (1987). 196] Safonov. and Physics. Ed. A. M. \On a certain linear fractional transformation.C.AC20. \Stability margins of diagonally perturbed multivariable feedback systems. M. Khargonekar (1991)." Linear Algebra and its Applications.G.time systems. R. 191] Roesser. (1960). 269286.. J. C. Nice. and J. R. pp. pp. 3. 193] Rudin. 198] Safonov. New York: Reidel. (1984). G. 6. No. M. M. 199] Safonov. L.and discrete. pp. (1982). vol. 18. 195] Qiu. A. Princeton. Minneapolis." Automatica. pp." 6th Int. \Feedback stability under simultaneous gap metric uncertainties in plant and controller." Journal of Mathematics and Mechanics.L. \Existence and comparison theorems for algebraic Riccati equations for continuous. 129. on Analysis and Optimization of Systems.G..G. M. pp. 39. Tzafestas.P. L." IEEE Trans. Vol. Vol. . Math. No." IEEE Trans. \A homotopy algorithm for solving the optimal projection equations for xedorder dynamic compensation: existence. Davison (1992b). 192] Rotea.
E. \Worstcase design in the time domain: the maximum principle and the standard H1 problem.J. 8594. (1967). 215] Vidyadagar. 207] Stein. The H1 Control Problem: A State Space Approach. Kimura (1986). M." AIAA Journal of Guidance and Control. AC14. D. Vol. PrenticeHall. 208] Stein. Hendrick (1977). MA. C." preprint." IEEE Trans. 30. Control System Synthesis: A Factorization Approach. Englewood Cli s. \Robust controllers for uncertain linear multivariable systems. K. A. No. L." preprint. No." IEEE Trans. Optimiz. 216] Vidyadagar. 211] Tits. Contr. Doyle (1991). pp. Vol. 301324. R. Contr." Automatica. C. (1990). \Adaptive control laws for F8c ight test.. \Inversion of multivariable linear systems. 202] Sarason.. vol. Automatic Control. M. 127. Vol 50. pp. (1992). 3. and R. 300301. (1988).BIBLIOGRAPHY 577 201] Safonov. 1. \Simplifying the H1 Theory via Loop Shifting. Systems and Signal Processing. H. M. Hartmann. Vol. 9. MIT Press. 38. and M. 205] Silverman. Automatic Control. (1992). C. No. G. (1984).. A. 179203. no. pp. 123142. Vol." IEEE Trans. Contr. Y. 22. (1988). Optimiz. G. matrix pencil and descriptor concepts. 204] Scherer. pp. L. \The graph metric for unstable plants and robustness estimates for feedback stability. 1. vol. Automat. pp. (1994). 29. (1985). pp. M. Cambridge." IEEE Trans." SIAM J. 13711383. 203] Scherer. (1993). 33. Contr. Auto. John Willey & Sons. and D. 217] Vinnicombe. 2467 2488." Trans. 210] Tadmor. . M. L.. C. 206] Skelton. \On the small theorem. and R.. G. and J.G. 209] Stoorvogel." Int. 212] Tits. L. Dynamic System Control. \H1 control for plants with zeros on the imaginary axis. Vol. pp. pp. pp. 270276. A. (1969). Limebeer.403417. pp. J." IEEE Trans. Vol. and H. \Beyond singular values and loop shapes.N." Mathematics of Control. M. \The small theorem with nonrational uncertainty. \H1 optimization without assumptions on nite or in nite zeros. \Frequency domain uncertainty and the graph topology." SIAM J. Chiang (1990). Fan (1994). 30. 14. A. AMS. Vol. No. \Normalized coprime factorizations for nonstrictly proper systems. G. \Generalized interpolation in H1 . Automatic Control.1. G. 214] Vidyasagar. 213] Vidyasagar.. (1992). 143166. contr.
224] Whittle. Prob. W. 7. SpringerVerlag.441455. Control. and M.Glover (1991). 221] Wang. Bongiorno (1976)." in Essays in Time Series and Allied Processes (Applied Probability Trust. Doyle. AC21. 222] Wang. P. C. Control. and J. Control. No. J. Auto.A. \Risksensitive LQG control.. AC30. pp. Vol..." Proc. Jr." Syst. no. 227] Willems." IEEE Trans. (1986).. 681697.J. H.J. Auto. Control. April 1985. Vol. No.C. and C. No. (1968). Vol. J. 229] Wonham. England.. 230] Wonham. J. 219] Wall. vol. (1985)." SIAM J. 231] Youla. W." Systems and Control Letters. pp. \On a matrix Riccati equation of stochastic control. 621634. 225] Willems. pp. D.. AC16. 313. 307317. 652665..578 BIBLIOGRAPHY 218] Walker.. pp. J. Vol 14. vol. 8. \Least Squares Stationary Optimal Control and the Algebraic Riccati Equation". 18th Allerton Conf. London). Linear Multivariable Control: A Geometric Approach.A. J. (1981). Auto. W. vol. W. Harvey (1980). 764777. 715725. Vol. Bongiorno.Doyle and K. \Tradeo s in the design of multivariable feedback systems. third edition. \Singular optimal control: a geometric approach. New York. vol.M. Kitapci and L. \Monotonicity of Maximal solutions of algebraic Riccati equations. Safonov (1990).. 13. 6. pp." Adv. J. Appl. Control.C. 317319. AC26. D." Int. Silverman (1986). Control. No. IEEE Trans. A. (1971). Auto. K. \A risksensitive certainty equivalence principle. Control Optimiz. and J. C." IEEE Trans. vol. pp. D. \Multiplicativeerror bound for balanced stochastic truncation model reduction. P. 235252." 1991 IEEE CDC.C.C. pp. Safonov (1992). H. \Modern WienerHopf design of optimal controllers: part I.C. M. pp. 232] Youla. 12651267. pp 383388. G. 6. \A tighter relativeerror bound for balanced stochastic truncation." SIAM J.E. 223] Whittle. J. 5. and M. \Stability and Model Reduction of Linear Fractional Transformation System. Contr. W. \A feedback theory of twodegreeoffreedom optimal WienerHopf design." IEEE Trans. M. Lett.. 24. 226] Willems. 323337. \Robust stabilizability of discretetime systems with normalized stable factor perturbation.AC37. . 228] Wimmer. (1990). G. J. pp. vol. 220] Wang. Jr. (1981). pp. Automat. Jabr. Contr.2. 4. pp. (1985). 52. (1985). \Almost invariant subspaces: an approach to high gain feedback design { Part I: almost controlled invariant subspaces." IEEE Trans. pp.
\An algebraic Riccati equation approach to H1 optimization. Skelton (1984). 8. pp. 3. \Modern WienerHopf design of optimal controllers: part II. Allerton Conf. Vol. 235] Youla. J.. 77108.." Automatica. G. vol. Journal of The Franklin Institute. Auto." IEEE Trans. no. (1993).AC29. . \Unstable systems and feedback: The gap metric. K. AC26. (1993). \Singleloop feedback stabilization of linear multivariable dynamical plants. Bodenheimer. 301320. 39. vol. on Automatic Control . pp. K. 10. pp. \Comparison between H2 and H1 controllers. vol. Glover. No. on Automatic Control. P. B.. N.J." IEEE Trans. 37. pp. P. (1992)." Proc. 21. K. on Automatic Control. Bongiorno (1976). \On the inputoutput stability of nonlinear timevarying feedback systems. AC11. \Frequency weighted L1 norm and optimal Hankel norm model reduction. Doyle. \Frequency weighted model reduction with L1 error bounds. Glover. A. Contr. pp. vol. (1981). pp. California Institute of Technology..A. 11. and J. K. \A note on balanced controller reduction." IEEE Trans. 244] Zhou. multiplicative seminorms. C. Jabr. Automat. 8592. K. Contr.. Vol. pp. G. Doyle (1994). Lu (1974). 239] Zames. 238] Zames. 237] Yousu .. and M. vol. AC21. Saito (1967). H. Vol. and A. \On the parameterization of H1 controllers.. Vol. and R. 8.C. D. Lett. 247] Zhou. 236] Young. pp. 2. 242] Zhou. K. 246] Zhou. and B." IEEE Trans. 319338. H. Vol. and C. pp. 15641574. K." Systems and Control Letters. (1993)." IEEE Trans. C. Khargonekar (1988). K." IEEE Trans. 240] Zames. Auto. Jabr. PhD Thesis.C. 254257. No. Auto. 241] Zhou. Control. \Mixed H2 and H1 performance objectives I: robust performance analysis. Control. Contr. \Mixed H2 and H1 control. vol." submitted to IEEE Trans. 1993. 284. of the 1990 American Control Conference.. Control. and approximate inverses. K. (1992).A. 115125. on Automat. No. 245] Zhou. parts I and II. (1966). Bodenheimer (1990). 234] Youla. 37." Syst. G. 9. 243] Zhou. California. 380385. D. and J. D. 159173. Robustness with Parametric and Dynamic Uncertainty. \Feedback and optimal sensitivity: model reference transformations. 12611265. ElSakkary (1980). No. 14421445. and P. 228 and 465. 25022507. pp. San Diego." IEEE Trans. 1988. pp. \Interpolation with positivereal functions". E." Proc. pp. K.BIBLIOGRAPHY 579 233] Youla. M.
8. 848855. Vol." IEEE Trans. S.580 BIBLIOGRAPHY 248] Zhu. pp. NO. Automat.. (1989). Q. Contr. \Graph topology and gap topology for unstable systems. . 34.
294 algebraic duality. 141 581 Bode's gain and phase relation. 489 H1 performance. 69. 179 analytic function. 548 error bounds. 71. 97 antistable. 47 operator. 128. 357 compact operator. 175 CayleyHamilton theorem. 467 coprime factorization. 48. 161. 90 CauchySchwartz inequality. 11. 158 balanced realization. 358 complementary sensitivity. 94 admissible controller. 90 basis. 168. 164 multiplicative. 297 algebraic Riccati equation (ARE). 45 Gramian. 179. 485 additive. 94 anticausal. 142 conjugate system. 195 frequencyweighted. 60 controller parameterization. 213. 155 bounded real function. 66 contraction. 41 classical control. 92 causal function. 353 discrete time. 486 coprime factors. 293 controller reduction. 408 controllability. 321 all solutions. 71 matrix. 157 stability. 532 bounded real lemma. 157 discrete time. 358. 97 balanced model reduction additive. 157 Banach space. 467 . 327 allpass. 353 discrete time. 21. 54 central solution H1 control. 335 minimal solution. 175 controllable canonical form. 523 maximal solution. 222.Index additive approximation. 27 stability property. 532 Cauchy sequence. 322 complementarity property. 335 solutions. 17 Bezout identity. 411 matrix dilation problem. 493 coprime factor uncertainty. 316. 157 additive uncertainty. 219 adjoint operator. 157 relative. 327 discrete time. 123 Bode integral. 327 stabilizing solution. 232 coinner function. 123. 51. 175 causality. 178 complementary inner function. 357 dilation.
18 column. 74 graph metric. 45. 232 gap metric. 324 generalized inverse. 548 H2 control. 142 Hermitian matrix. 546 full rank. 387. 452 xed order controller. 21. 155 gain margin. 69 Gramian controllability. 289 dom(Ric). 96 frequency weighting. 539 normalized. 173 Hardy spaces. 116 ltering H1 performance. 18 row. 75. 18 gain.582 discrete time. 305 equivalence relation. 188 Hankel operator. 297 eigenvalue. 541 disturbance feedforward (DF). 21 H1 control. 20 eigenvector. 173. 305 equivalent control. 10. 294. 268 detectability. 422 H2 control. 95 harmonic function. 484 generalized eigenvector. 141 destabilizing perturbation. 539 H1 control. 385. 211 design tradeo . 74 observability. 321 Hankel norm. 173 Hankel norm approximation. 386 D . 516 Fourier transform. 164 Frobenius norm. 21 entropy. 327 double coprime factorization. 297 H1 control. K iteration. 543 innerouter factorization. 397 Hankel singular value. 523 coprime factorization. 242 Fu . 324 generalized. 294. 298 H1 control. 525 Gilbert's realization. 71. 20. 484 graph topology. 540 Lyapunov equation. 298 H1 control. 71. 10. 153 design model. 27. 484 Hamiltonian matrix. 162. 35 generalized principal vector. 135 frequencyweighted balanced realization. 305 INDEX feedback. 523 discrete time. 29 full control (FC). 294. 123 dual system. 173. 51 di erential game. 553 normalized coprime factorization. 423 H2 control. 393 mixed HankelToeplitz operator. 21 lower rank generalized eigenvector. 242 . 10. 94 discrete algebraic Riccati equation. 523 model reduction. 405 F` . 470 cyclic matrix. 546 full information (FI). 435 equivalence class. 26. 416 H2 control. 66 duality. 459 direct sum. 61 design limitation.
537 Hermitian. 373 l2 (. 523 main loop theorem. 28. 199 Lyapunov equation. 475 loop transfer matrix (function). 406 integral control. 343. 128 LQG stability margin. 448 state feedback. 84. 448 internal model principle. 143. 94 maxmin problem. 537 inner product. 467 singular problem. 135. 89. 95 L2 (. 448 H1 control. 95 H2 stability margin. 452 H1 optimal controller. 132 H1 approach. 26. 92 innerouter factorization. 543 H2 performance. 52 kernel. 36 maximum modulus theorem. 70 discrete time.INDEX discrete time. 97 hidden modes. 357 discrete time. 448 H2 control. 467 normalized coprime factorization. 138 H1 space. H1 space. 130 loop shaping. 80 . 119. 17 linear fractional transformation (LFT). 389 ? H2 space. 448 internal stability. 459 H1 ltering. 66 isometric isomorphism. 400 McMillan degree. 77 Hilbert space. 520 H2 optimal control 365 discrete time. 511 left coprime factorization. 548 loop shaping. 322 invariant zero. 123 linear combination. 128 inputoutput stability. 91 homotopy algorithm. 97 . 91. 199 ? H2 (@ D ) space. 335 inertia. 38 dilation. 96 H2 (@ D ) space. 26. 38. 199 Hurwitz. 25 Lagrange multiplier. 93 L2 (@ D ) space. 101 inertia. 89. 430. 406 invariant subspace. 135 H2 space. 81 McMillan form. 21 inequality.1 1) space.1 1) space. 25 Kronecker sum. 179 inner function. 18 Kronecker product. 445 factorization. 20 583 Kalman canonical decomposition. 96 loop gain. 18 induced norm. 333 inverse of a transfer function. 93 L2 space. 179 inversion formulas. 241 linear operator. 275 matrix compression. 27 square root of a. 174 Jordan canonical form. 437 H1 performance. 49 image. 22 norm. 389 LQR problem. 358 input sensitivity. 367 LQR stability margin. 91 L1 space.
71 operator. 128 Parrott's theorem. 400 minimum entropy controller. 51 observer. 386 output sensitivity.584 minimal realization. 116. 397 modal controllability. 365 quadratic performance. 365. 96 plant condition number. 67. 18 Nyquist stability theorem. 175 observable canonical form. 157 relative. 100 pseudoinverse. 157. 230 Poisson integral. 79 normalized coprime factorization. 27 Frobenius. 211 . 62 observerbased controller. 387 output injection. 10. 249 multiplication operator. 28 normal rank. 94 orthogonal matrix. 94 orthogonal direct sum. 67 balanced. 203 nominal performance (NP). 77 positive real. 263 lower bound. 213. 57 polezero cancelation. 273 multidimensional system. 215 nominal stability (NS). 393 Rp (s) . 362 loop shaping. 188 optimality of H1 controller. 40 Parseval's relations. 425 H2 control. 91 optimal Hankel norm approximation. 77. 75 . 51 performance limitation. 511 orthogonal complement. 61 observable mode. 298 H1 control. 18 orthogonal projection theorem. 298 H2 control. 155 phase margin. 215 nonminimum phase zero. 435 mixed HankelToeplitz operator. 475 null space. 51 model reduction additive. 143 pole placement. 35 quadratic control. 28 seminorm. 157 multiplicative. 18 realization. 141 pole. 45 Gramian. 18. 123 observability. 165 multiplicative uncertainty. 430 optimization method. 51 modal observability. 157 model uncertainty. 73 minimax problem. 294. 141 norm. 98 multiplicative approximation. 273 synthesis. 288 upper bound. 141 phase. 80 range. 29 induced. 79 pole direction. 91 restriction. 96 PBH (PopovBelevitchHautus) tests. 220 Nehari's Theorem. 232 Plancherel theorem. 94 outer function. 36 power signal. 354 positive (semi)de nite matrix. 358 output estimation (OE). 62 operator INDEX extension.
97 RH. 142 bounds. 141 transition matrix. 23 selfadjoint operator. 395 trace. 293 stabilizing controller. 123 risk sensitive. 388 simplectic matrix. 467 rollo rate. 19 tradeo . 19. 252 unimodular matrix. 135 wellposedness. 45. 79 span. 327 right coprime factorization. 77 Redhe er starproduct. 226 structured uncertainty. 32 skewed performance speci cation. 128. 95 2 RH2 space. H2 . 178 sensitivity function. 279 robust stability (RS). 157. 500 weighting. 45. 215 Smith form. 243 winding number. 263 subharmonic function. 32 singular value decomposition (SVD). 197. 100 stability. 459 H2 control. 215 structured uncertainty. 321 Riccati operator. 31 singular vector. 128 RH1 space. 49 internal. 263 lower bound. 426 H2 control. 354 structured singular value. 17 585 spectral factorization. 278 robust stabilization. 96 Riccati equation. 273 upper bound. 68 Toeplitz operator. 67 output normal. 226. 211 state space. 97 1 RH? space. 165 return di erence. 293 stable invariant subspace. 178 Schur complement.INDEX input normal. 94. 20 spectral signal. 273 structured uncertainty. 275 H2 performance. 436 robust performance (RP). 512 regulator problem. 89 Sylvester equation. 153 integral. 448 singular value. 164. 49 stabilization. 525 singular H1 problem. 46 transmission zero. 18 unstructured uncertainty. 142 supremum norm. 327. 77 minimal. 144 separation theory. 26 Sylvester's inequality. 144 Schmidt pair. 215 H1 performance. 79 unitary matrix. 211 weighted model reduction. 230 small gain theorem. 298 H1 control. 81 uncertainty. 60 strictly positive real. 27. 343 spectral radius. 259 state feedback. 119 stability margin LQG. 365 relative approximation. 116. 525 starproduct. 409 . 310 H1 control. 259 regular point. 117. 389 stabilizability. space. 382 state space realization.
454 zero. 81 INDEX . 79 blocking.586 Youla parameterization. 84 transmission. 142 invariant. 81 direction. 316. 303. 77.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.