Lecture Notes in Control and Information Sciences
Editor: M. Thoma
258
Springer
London Berlin Heidelberg New York Barcelona Hong Kong Milan Paris Singapore Tokyo
Alberto Isidori, Franqoise LamnabhiLagarrigue,
andWitoldRespondek (Eds)
Nonlinear Control in the Year 2000
Volume 1
With 86 Figures
~ Springer
Series Advisory
Board • A.B. K u r z h a n s k i •
A. B e n s o u s s a n • M.J. Grimble • P. Kokotovic H. Kwakernaak • J.L. M a s s e y • M. Morari Editors Alberto Isidori, Professor
Dipartimento di Informatica e Sistemistica, Universitt~ di Roma "La Sapienza",
00184 R o m e I t a l y Franqoise LamnabhiLagarrigue, Docteur D'6tat L a b o r a t o i r e d e s S i g n a u x et S y s t e m s , CNRS SUPELEC, 91192 G i f  s u r  Y v e t t e , F r a n c e Witold Respondek, Professor L a b o r a t o i r e d e M a t h 6 m a t i q u e et I n f o r m a t i q u e , INSA d e R o u e n , 76131 M o n t S a i n t A i g n a n , F r a n c e
ISBN 1852333634 SpringerVerlag London Berlin Heidelberg British Library Cataloguing in Publication Data A catalog record for this book is available from the British Library Library of Congress CataloginginPublication Data Nonlinear control in the year 2000 / Alberto Isidori, Franqoise LarrmabhiLagarrigue, and Witold Respondek. p. cm.  (Lecture notes in control and information sciences, ISSN 01708643 ; 258259) Includes bibliographical references. ISBN 1852333634 (v. 1 : acidfree paper)  ISBN 1852333642 (v.2 : acidfree paper) 1. Nonlinear control theory. I. Isidori, Alberto. II. LamnabhiLagarrigue, F. (Franqoise), 1953 III. Respondek, W. IV. Series. QA402.35.3N66 2001 003.5dc21 00045600 Apart from an), fair dealing for the purposes of research or private study, or criticism or review, as permitted under the Copyright, Designs and Patents Act 1988, this publication may only be reproduced, stored or transmitted, in any form or by any means, with the prior permission in writing of the publishers, or in the case of reprographic reproduction in accordance with the terms of licences issued by the Copyright Licensing Agency. Enquiries concerning reproduction outside those terms should be sent to the publishers. © SpringerVerlag London Limited 2001 Printed in Great Britain The use of registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant laws and regulations and therefore free for general use. The publisher makes no representation, express or implied, with regard to the accuracy of the information contained in this book and cannot accept an}, legal responsibility or liability for any errors or omissions that may be made. Typesetting: Camera ready by editors Printed and bound at the Athenaeum Press Ltd., Gateshead, Tyne & Wear 69/3830543210 Printed on acidfree paper SPIN 10768731
Preface
These two volumes contain papers based on talks delivered at the 2nd Workshop of the Nonlinear Control Network (http://www.supelec.fr/lss/~lCN), held in Paris, June 59, 2000. The Authors of the presented papers, as well as the Editors of these two volumes, hope that "Nonlinear Control in the Year 2000" is not only one more book containing proceedings of a workshop. Two main reasons justifying our hope are that, firstly, the end of the century is a natural moment to think about past developments in nonlinear control and about its perspectives in the twentyfirst century; and, secondly, we believe that nonlinear control has reached an age of maturity which enables the community to sketch a stateoftheart of the field. We hope that papers contained in the two volumes serve to fulfill these goals: many of them have their roots and their origins in nonlinear control theories which have been developed in past three decades and which, by now, form a basis of canonical results in the field. Such papers form a bridge between the actual theory and its future developments. Many other papers contained in the book present completely new ideas and suggest new directions and, in this sense, they are directed towards the future of the nonlinear control. We would like to emphasize one peculiarity of our field: nonlinear control is an example of a theory situated at a crossroad between mathematics and engineering science. Due to this position, nonlinear control has its roots in both fields and, as we deeply believe, can bring new ideas and new results for both domains. The book reflects very well this "double character" of nonlinear control theory: the reader will find in it results which cover a wide variety of problems: starting from pure mathematics, through its applications to nonlinear feedback design, all way to recent industrial advances. Eight papers contained in the book are based on invited talks delivered at the Workshop by:
Alessandro Astolfi, John Baras, Christopher Byrnes, Bronistaw Jakubczyk, Anders Rantzer, Kurt Schlacher, Eduardo Sontag, Hector Sussmann.
Altogether the book contains 80 papers and therefore it is impossible to mention all discussed topics: to give a flavour of the presented material let us mention a few of them. For many theoretical papers a common factor is optimal control, for example subRiemannian geometry (in particular fundamental results on (non)subanalyticity of small balls and of the distance function), the use of generalized differentials in generalized Maximum Principle, singular and constrained optimal control problems. Another subdomain of nonlinear control attracting many contributions to the book is stability and asymptotic behavior. In this area, stemming from traditional Lyapunov
VI techniques, the methods based on the concept of inputtostate stability have established a leading role in analysis and design, and new ideas have emerged, such as those aiming at the analysis of nonequilibrium steadystate behaviors, or at the evaluation of asymptotic convergence for almost all initial conditions via dual Lyapunov analysis. Applications of these ideas in nonlinear control are widespread: stabilization of nonlinear systems, trajectory tracking, adaptive control. Other subdomains of nonlinear control attracting a lot of attention, and represented in the book, are that of observability and observers for nonlinear systems, sliding mode control, theory of nonlinear feedback (invariants, classification, normal forms, flatness and dynamic feedback), recursive design (backstepping and feedforwarding). The papers present in the two volumes cover various aspects of all just mentioned topics. Moreover, the book contains also papers discussing main results of a plenary Session devoted to industrial applications. We wish to thank all invited speakers and all contributors to the 2nd Nonlinear Control Network Workshop for making this conference an outstanding intellectual celebration of the area of Nonlinear Control at the turn of the century. The Editors are grateful to all the chairpersons:
Dirk Aeyels, Alfonso BaSos, Fritz Colonius, Halina Frankowska, Bronistaw Jakubczyk, Jean L~vine, Frdddric Mazenc, Philippe Mullhaupt, Elena Panteley, Pierre Rouchon, Arjan van der Schaft, Fatima Silva Leite, Alan Zinober.
Andrei Agrachev, Georges Bastin, Emmanuel Delaleau, JeanPaul Gauthier, Philippe Jouan, Antonio Loria, Gdrard Montseny, Henk Nijmeijer, Laurent Praly, Joachim Rudolph, Jacquelien Scherpen, Hebertt SiraRamirez,
Andrea Bacciotti, Antonio Bicchi, Michel Fliess, Henri Huijberts, loan Landau, Riccardo Marino, Claude Moog, Romeo Ortega, Anders Rantzer, Andrey Sarychev, Rodolphe Sepulchre, Fabian Wirth,
They have excellently played their role during the presentations and also have helped us in reviewing the contributed papers accepted for publication in this book. We would like to thank the T M R Program for the financial support which, in particular, helped numerous young researchers in attending the Workshop. We express our thanks to Christiane Bernard (Mathematics and Information Sciences Project Officer at the European Community) and Radhakisan Baheti (Program Director for Control at NSF) for their participation at the Session on Founding to Academic Research on Control. We also thank the CNRS staff as well as PhD students and Postdocs at L2S in GifsurYvette for their help in the organization of the Workshop.
Frangoise LamnabhiLagarrigue
GifsurYvette
Alberto Isidori
Rome
Witold Respondek
Rouen
Contents
Volume 1
Subanalyticity of distance and spheres in SR geometry ...... 1
Andrei Agrachev, JeanPaul Gauthier
Principal invariants of Jacobi curves .......................... 9
Andrei Agrachev, Igor Zelenko
The De Casteljau algorithm on SE(3) ......................... 23
Claudio Altafini
Trajectory tracking by cascaded backstepping control for a secondorder nonholonomic mechanical system ................
35
Nnaedozie P.I. Aneke, Henk Nijmeijer, Abraham G. de Jager
Static output feedback stabilization: from linear to nonlinear and back .....................................................
49
Alessandro Astolfi, Patrizio Colaner~
Semilinear diffusive representations for nonlinear fractional differential systems ...........................................
73
Jacques Audounet, Denis Matignon, Gdrard Montseny
C o n t r o l l a b i l i t y p r o p e r t i e s o f a class o f c o n t r o l s y s t e m s o n L i e groups ........................................................
83
Victor Ayala, Luiz A. B. San Martin
Stability analysis to parametric uncertainty extension to the m u l t i v a r i a b l e case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
93
Miguel Ayala Botto, Ton van den Boom, Josg Sd da Costa
External stability and continuous Liapunov functions ......... 103
Andrea Bacciotti
Optimal control with harmonic rejection of induction machine 113
Iyad Balloul, Maze, Alamir
VIII
Nonlinear QFT synthesis based on harmonic balance and multiplier theory .............................................
123
Alfonso BaSos, Antonio Barreiro, Francisco Gordillo, Javier Aracil
G r o u p i n v a r i a n c e a n d s y m m e t r i e s in n o n l i n e a r c o n t r o l a n d estimation ....................................................
137
John S. Baras
A g l o b a l i z a t i o n p r o c e d u r e for l o c a l l y s t a b i l i z i n g c o n t r o l l e r s . . . 171
Jochen Behrens, Fabian Wirth
Optimal control and implicit Hamiltonian systems ............ I85
Guido Blankenstein, Arian van der Schaft
Robust absolute stability of delay systems .................... 207
PierreAlexandre Bliman
S t r a t i f i c a t i o n d u s e c t e u r a n o r m a l d a n s la s p h e r e d e M a r t i n e t de p e t i t r a y o n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
239
Bernard Bonnard, Emmanuel Trdlat
N o n e q u i l i b r i u m t h e o r y for n o n l i n e a r s y s t e m s . . . . . . . . . . . . . . . . . 253
Christopher I. Byrnes
A regularization of Zubov's equation for robust domains of attraction .....................................................
277
Fabio Camilli, Lars Griine, Fabian Wirth
A remark on Ryan's generalization of Brockett's condition to discontinuous stabilizability ...................................
291
Francesca Ceragioli
A p p l i c a t i o n s o f Lie a l g e b r o i d s in m e c h a n i c s a n d c o n t r o l t h e o r y 299
Jesds ClementeGallardo
O b s e r v e r d e s i g n for locally o b s e r v a b l e a n a l y t i c s y s t e m s : convergence and separation property .........................
315
Fabio Conticelli, Antonio Bicchi
A n H c c  s u b o p t l m a l f a u l t d e t e c t i o n f i l t e r f o r b i l i n e a r s y s t e m s .. 331
Claudio De Persis, Alberto Isidori
IX A d a p t i v e control of feedback linearizable systems by orthogonal approximation functions ........................... 341
Domitilla Del Vecchio, Riccardo Marino, Patrizio Tomei
S a m p l e d  d a t a lowgain integral control of linear systems with actuator and sensor nonlinearities ............................ 355
Thomas Fliegner, Hartmut Logemann, Eugene P. Ryan
State feedbacks without asymptotic observers and generalized PID regulators ............................................... 367
Michel Fliess, Richard Marquez', Emmanuel Delaleau
Eigenstructure of nonlinear Hankel operators ................. 385
Kenji Fujimoto, Jacquelien M.A. Scherpen
D i s t r i b u t e d a r c h i t e c t u r e for t e l e o p e r a t l o n o v e r t h e i n t e r n e t ... 399
Denis Gillet, Christophe Salzmann, Pierre Huguenin
Singular LQ problems and the D i r a c  B e r g m a n n t h e o r y of constraints ....................................................
409
Manuel Guerra
Robust tracking of multlvarlable linear systems under parametric uncertainty ....................................... 423
Veit Hagenmeyer
F l a t n e s s  b a s e d c o n t r o l o f t h e s e p a r a t e l y e x c i t e d D C d r i v e . . . . . 439
Veit Hagenmeyer, Philipp Kohlrausch, Emmanuel Delaleau
S t a t e d e t e c t i o n a n d s t a b i l i t y for u n c e r t a i n d y n a m i c a l s y s t e m s 453
Mohamed Ali Hammami
C o n t r o l l a b i l i t y p r o p e r t i e s o f n u m e r i c a l e i g e n v a l u e a l g o r i t h m s . 467
Uwe Helmke, Fabian Wirth
On the diseretizatlon of slidingmodelike controllers . . . . . . . . . . 481
Guido Herrmann, Sarah K. Spurgeon, Christopher Edwards
N o n l i n e a r a d a p t i v e s t a t e s p a c e c o n t r o l f o r a class o f n o n l i n e a r systems with unknown parameters ............................ 497
Christian Hintz, Martin Rau, Dierk Schr6der
×
An observer view on synchronization ......................... 509
Henri J.C. Huijberts, Henk Nijmeijer
Regularity of the subRiemannian distance and cut locus ..... 521
Sdbastien Jacquet
Industrial sensorless control of induction motors .............. 535
Fabrice Jadot, Philippe Martin, Pierre Rouchon
Feedback invariants and critical trajectories; H a m i l t o n i a n f o r m a l i s m for f e e d b a c k e q u i v a l e n c e . . . . . . . . . . . . . .
545
Bronistaw Jakubczyk
Paths in subRiemannian geometry ........................... 569
Frgddric Jean
Observability of C°°systems for Loosingleinputs ............. 575
Philippe Jouan
Robust control of a synchronous power generator ............. 583
Matei Kelemen, Aired Francis Okou, Ouassima Akhrif, LouisA. Dessaint
................ Mazen Alamir Stabilisation of nonlinear systems by discontinuous state feedback ............................................................................. Santosuosso Hamiltonian representation of distributed parameter with boundary energy flow ... Landau......×l Volume 2 Control of a reduced size model of US navy crane using only motor position sensors .. Joachim Rudolph Dynamic feedback transformations of controllable linear timevarying systems ............... Lynch.. Arjan van der Schaft ........ dynamic 81 Nicolas Marchand......... Gilney Datum............. F...... Maschke............... Philippe Martin........ Franfoise LamnabhiLagarrigue Nonlinear observers of time derivatives from noisy measurements of periodic signals ..... 95 Lorenzo Marconi..................... time 63 Nicolas Marchand.... 45 Alan F................... 107 Riccardo Marino.......... Pierre Rouchon Asymptotic controllability implies continuousdiscrete feedback stabilizability ....... 13 Ioan D......... 1 Bdlint Kiss.. systems 137 Bernhard M.. Jean Ldvme................... O..... Iyad Balloul On the stabilization of a class of uncertain systems by bounded control ............. Mazen Alamir........................................... D...... Anderson........ Giovanni L. De Bruyne Flatnessbased boundary control of a nonlinear parabolic equation modelling a tubular reactor ...... Philippe Mullhaupt Algorithms for identification of continuous time nonlinear systems: a passivity approach ... Alberto Isidori Adaptive nonlinear control of synchronous generators .......... 55 Franfois Malrait....... B.......... 123 Riccardo Marino...
. . . . . Jonathan Pitts Diffusive r e p r e s e n t a t i o n for p s e u d o .. 237 Nicolas Petit.s l m i l a r traffic a n d s h a p i n g t e c h n i q u e s . ... . . .. . . .. ... .... Moog. . .... . .. .. . ... .. .. . . . . ... . . . . Zbigniew Bartosiewicz S i n g u l a r s y s t e m s in d i m e n s i o n 3: c u s p i d a l case a n d t a n g e n t elliptic fiat case . . .. . . . 265 Laurent Praly.. ... .. . . . .. .×11 Differentiable Lyapunov function and center manifold theory . . .. . . .. .. ..... Philippe Pinvidic C o n t r o l l a b i l i t y o f n o n l i n e a r m u l t i d i m e n s i o n a l c o n t r o l s y s t e m s . . .. Xia S t a b i l i z a t i o n o f n o n l i n e a r s y s t e m s via F o r w a r d i n g rnod{LaV } . . . . .... ... . 245 JeanFranfois Pommaret S t a b i l i z a t i o n o f a series D C m o t o r b y d y n a m i c o u t p u t f e e d b a c k 2 5 7 Richard Pothin. . . . . Denis Matignon Euler's discretization and dynamic equivalence of Nonlinear Control Systems . .. X..... . . . . .. . .. . . . . .. ..... Romeo Ortega. .. . . . . . . . .. JeanMichel Boueilh.. . . ... .. . ... . Claude H. 205 Paulo S~rgio Pereira da Silva M o t i o n p l a n n i n g for h e a v y c h a i n s y s t e m s .. .. 279 Christophe Prieur ..... ... Georgia I(aliora A robust globally asymptotically stabilizing Feedback: T h e e x a m p l e o f t h e A r t s t e i n ' s circles . . . 229 Nicolas Petit.. . . . . . .... . .. ... . .. . ... David K..... . . . . . ... . Pierre Rouchon C o n t r o l o f a n i n d u s t r i a l p o l y m e r i z a t i o n r e a c t o r u s i n g f l a t n e s s . Jacques Audounet. . .. 163 Gdrard Montseny.. ... Mondrag6n C. . Pierre Rouchon. . . 149 Radl J... . 193 Mich~le Pelletier Flatness of nonlinear control systems and exterior differential systems . .d i f f e r e n t l a l l y d a m p e d nonlinear systems . . 143 Frdddric Mazenc C o n t r o l l i n g s e l f .... . Frgddric Gugrin. 183 Ewa Pawtuszewicz... .... .. Arrowsmith.
..... William PasillasLgpine T r a j e c t o r y t r a c k i n g f o r 7rflat n o n l i n e a r d e l a y s y s t e m s w i t h a motor example ......... Lorenzo Marconi Sliding mode control of the PPR mobile robot with a flexible joint ..................Xlll Robust (TORA) s t a b i l i z a t i o n for t h e n o n l i n e a r b e n c h m a r k p r o b l e m using neural nets and evolution strategies . Andreas If ugi Vibrational control of singularly perturbed systems .. ... Santosuosso Stability criteria for timeperiodic systems via highorder averaging techniques ............ 323 Witold Respondek........ .... ....... .................. 469 Michael P............. 353 Giovanni L.. 397 Klaus R..................... 409 Andrea Serrani...... Byrnes....... Dimosthenis Fragopoulos ..... ....... .. 301 Cesdreo Raim(mdez O n c o n v e x i t y in s t a b i l i z a t i o n o f n o n l i n e a r s y s t e m s ..................... a computer algebra based approach .... Hugues Mounier Neurogenetic robust regulation design for nonlinear parameter dependent systems ....................... . ..... ... ...... 421 Hebertt SiraRam(rez T h e I S S p h i l o s o p h y as a u n i f y i n g f r a m e w o r k f o r s t a b i l i t y ............... ...... .. 315 Anders Rantzer E x t e n d e d G o u r s a t n o r m a l f o r m : a g e o m e t r i c c h a r a c t e r i z a t i o n .. .............. . Sontag C o n t r o l design of a crane for offshore lifting o p e r a t i o n s .. Alberto Isidori.. 339 Joachim Rudolph.l i k e behavior ....... 379 Kurt Schlachev . 365 Andrey Saryehev Control of nonlinear descriptor systems..................... Schneider R e c e n t a d v a n c e s in o u t p u t r e g u l a t i o n o f n o n l i n e a r s y s t e m s ....... ........ ... ... ...... Spathopoulos........... 443 Eduardo D..... .. Cristopher I........
....... Zhirabok N e w d e v e l o p m e n t s in d y n a m i c a l a d a p t i v e backstepping control ... Russell E.. Witold Respondek E x t e n d e d a c t i v e .I.. 555 Alexey N......... Sussmann Transforming a singleinput nonlinear system to a strict feedforward form via feedback .. . Scarratt. All Jafari Koshkouei List of participants at the 2nd NCN Workshop .................. 543 Ranjan Vepa O n c a n o n i c a l d e c o m p o s i t i o n o f n o n l i n e a r d y n a m i c s y s t e m s .................... 527 Issa Amadou Tall.×IV Setvalued differentials and the maximum principle of optimal control ............ Julie C.. . ... . ............................. 565 Alan S.....p a s s i v e d e c o m p o s i t i o n o f c h a o t i c s y s t e m s ........ 487 Hdctor J........ Zinober.. 623 ... Mills.......... .............
B. We denote by p(q) the s u b . T h e n u m b e r k = dim Aq is the rank of the distribution. i. connecting q0 to ql. p is H61dercontinuous with the HSlder e x p o n e n t ~ . meeting the initial condition q(0) = q0.R i e m a n n i a n distance between q0 and q. i = 1 ~ 9 " "~ rn}. .S u b a n a l y t i c i t y o f D i s t a n c e a n d S p h e r e s in SubRiemannian Geometry Andrei Agrachev 1 and J e a n . Russia agrachev@sissa. Iterated Lie brackets of the fields in /~ define a flag Aqo C A2 Q "'" C A m "'" C T q M qo qo in the following way: A m : span{[X1. The restriction of the R i e m a n n i a n structure to A is a subRiemannian structure.S. i t Laboratoire d'analyse appliqu4e et optimisation Universit de Bourgogne D4partement de Math4matiques.P. Sections of the linear bundle A are smooth vector fields. d i m M = n. ul. f r 1 Introduction Let M be a C ~ R i e m a n n i a n manifold. France gauthierQubourgogne.. Gubkina 8 117966 Moscow. A d i s t r i b u t i o n on M is a smooth linear s u b b u n d l e A of the t a n g e n t bundle T M . Moreover. is the subRiemannian distance or CarnotCaratheodory distance between q0 and qlIn the remainder of the paper we assume that A is bracket g e n e r a t i n g at a given initial point q0. q e M } . A d i s t r i b u t i o n A is bracket generating at q0 if A m : Tqo M for some m > 0 qo If A is bracket generating. p is a continuous function defined on a n e i g h b o r h o o d of q0.I. we set A={XEVecM:X(q) 6 Aq. Italy and Steklov Mathematical Institute. [X2. qo 9 "1 Xm] 9 9 .P a u l G a u t h i e r 2 1 S.e. and if q0 and ql are close. the R i e m a n n i a n length of the shortest path. We fix a point q0 6 M and study only admissible paths s t a r t i n g from this point. Via Beirut 24 34013 Trieste.](q0) : Xi 9 z~.S. 47870 21078 Dijon. the space of sections of A. [. where 1 aq' = TqoM.A.
3. Moreover. 2 Nilpotentization. i = l. 1].13.. . see [1. Main results of the paper concern subanalyticity properties of p in the case of a generic realanalytic A.15].~k. (i) where X : Oqo + lRn is an "adapted coordinate map". r]) of small enough radius are subanalytic if n > k > 3 (Th. lr.. Admissible paths are thus solutions of the Cauchy problem.7.10. .6. In adapted coordinates : x(qo)=O.) .5. 9 9 k. We prove that. u k ( ' ) ) E Lk[0. r]) are subanalytic but p is not! The results are just stated herein. i=1 q E Oqo. and transversality techniques. . .l ) ( ~ 5 ~ + ~ + 1). in a neighborhood without the pole q0). In particular. both in the normal and abnormal cases.1 ) ( ~ + ~ + l ) (Th. The nilpotentization of (1) has the following equation : k = i=1 x = 0. The main tools are the nilpotent approximation. l<i<l. (1) where u = ( u l ( . and the complete proofs can be found in [16]. We are working in a small neighborhood Oqo of q0 E M. the germ of p at q0 is subanalytic if n < ( k .1([0. endpoint mapping and geodesics Nilpotentization or nilpotent approximation is a fundamental operation in the geometric control theory and subRiemannian geometry. . 7). The bails p1 ([0.1 (Th. . . if k > 3. generically. Here we refer specially to [7]. Morsetype indices of geodesics. This paper is a new step in a rather long research line.1 ) k q . The g e r m at q0 of a Riemannian distance is the square root of an analytic germ. k (t = Z u i ( t ) X i ( q ) . . 5 ) a n d is not subanalytic i f n _> ( k .e. but up to a set of distributions of codimension co. this is a real nonholonomic analog of the usual linearization (see [2. .2 Andrei Agrachev and JeanPaul Gauthier We study mainly the case of realanalytic M and A. This is not true for a subRiemannian distance function p. q(O) = qo.12. x. X k E Vec Oqo of the subRiemannian structure under consideration.9. 8). where we fix an orthonormal frame X 1 .8]). p is never smooth in a punctured neighborhood of q0 (i. then (generically!)the balls p .It may happen that p is not even subanalytic. This statement about the balls is valid not only generically. and )(i(x) is the "nilpotentization" of Xi. n > ( k .0 ( qo)=~kl@.
t E JR. .) ~ L~[0. Moreover. Then the following identities hold for any u(. It is a welldefined smooth mapping of a neighborhood of the origin of Lk2[0. . Then f~ is a smooth mapping from a neighborhood of 0 in L~[O. 1] into M..r}.. ~ ) = ( t ~ l . where ~ is the dilation (2). u c L~[O. Let X = ( X 1 .. . . 1]..(t~)2. .~x)) ~(x) = min{[]u[[ : ](u) = x. .) ) . .. 1] is contained in the domain of f~ for e small enough. 0 u~(t) dt is the norm in Lk[0. Below (2) Ilull = (i.. t ~ 2 .. . . Clearly. 1]} = ~p(xl(.~2. We introduce a dilation dt : ll~n + 1W~.<t~<_lij=l o. l . by the formula: ~t(~1. . 1]. i = 1 . Let Ur = {u C L~[0. f~ ~ ] as e 4 0 in the C ~ topology of the uniform convergence of the mappings and all their derivatives on the balls in Lk2[0. Thus/~ is the subRiemannian distance for the nilpotentization of the original system. . 1]: Ilull .f E ~. We set: p~(x) = and: 1 min{llull : f~(u) = x. 1]. p(q) = min{llull : u a L~[0.~. 1]} is the subRiemannian distance function. . any bounded subset of L~ [0. f(u) = q}. u E L~[0. . . . We define also the mapping ] : L~[0. 1] + ~ n by the rule ] : u(. P r o p o s i t i o n 1..(t~). . u E L~[O. f. j = 1 . . u) is the solution of (i).. c ~ ]~: ](u(. .) ~~ x(1). . 1]}.... ](eu(. t % ) .)) = ~ ] ( u ( . 1]. Theorem 1.Subanalyticity of Distance and Spheres in SR Geometry 3 dim(A~0/A~ol ) = k.. 1] to 1R'~. O<_tl<_. be the sphere of radius r in L~ [0. . where x(.dt~ . Recall that p(q) = min{[]u[[ : f ( u ) = q.)) = \0 i=1 ui(t)Xix1 (qo) d t .o2. We consider the endpoint mapping f : u ~ q(1). X t ) . 1]. . Xj : Oqo + ~ k j . We set f~(u)) = ~ X ( f ( e u ) ) ....x~(qo)dt~.. I..) = x(..
. k. where hi is the Hamiltonian lift of Xi.~~uz(t)h.1].. h(r = 89E h~(r i=1 Vq E M. P r o p o s i t i o n 2. Normal geodesics are projections r e ( t ) on M of trajectories on T * M of the Hamiltonian vector field "~ associated with the Hamiltonian k h. It is easy to see that both conditions are actually intrinsic: Identity (4) doesn't depend on the choice of the orthonormal frame X 1 . which is a consequence of theorem 2 and of general facts about subanalytic sets: T h e o r e m 3. for almost all t ~ Hamiltonians a..4 Andrei Agrachev and JeanPaul Gauthier L e m m a 1.}}vivj i. a u is minimal for (1)}. T h e o r e m 2. The family of functions p~ Ig is equicontinuous for any compact K C ~ ~.. . (GohLegendre condition) If u(. then ~ is subanalytic.. 1] s.vk) E]R k. [] where {a. . + ~ uniformly on compact subsets of]~ n as ~ ~ O. Subriemannian minimizers (geodesics) can be of two types: normal or abnormal.{hj. b.. . k } ..j=l s=l <0 V(vx. . k k (4) (5) ~'~{hi. and ] ( M r \ . Identity (4) is called the Goh condition while Inequality (5) is the generalized Legendre condition.hj}(C(t)) = O Vi. . in particular..) is a strictly abnormal min imizer. vanish anyway. . r E TqM. p. any element of M r \ M r is a minimizing control for system (i). it We come to the first subanalyticity result. (3) There is the following fundamental result: P r o p o s i t i o n 3.) is a Goh control if (4) is satisfied for an appropriate r is a GohLegendre control if both (4) and (5) are satisfied. it does exist a r satisfying (3) with: {hi. Xk since hi(r i = 1. . Then M r is a compact subset of the Hilbert sphere U. Abnormal geodesics are solutions of: k = ~~ui(t)hi(r i=1 hi(r = 0.b} = ab is the Poisson bracket of the R e m a r k . If the germ of p at qo is subanalytic. We say that u(. : 3c~ E (0. Let ~t~ = {u E U. The following proposition is a modification of a result by Jacquet [13].Mr) C ~1 (r). . Inequality (5) doesn't depend on the choice of the orthonormal frame provided that (4) is satisfied.j E { 1 . .t. [0.
. We come to the following: C o r o l l a r y 1. We will also use a realanalytic version of the definition. Medium fat systems do not admit nontrivial Goh controls. then the germ of p at qo is subanalytic. X(qo) r 0 (see [5]).4 C H k T M be a set of distributions. 4 Exclusivity of Goh Distributions Controls for Rank > 2 First we'll make precise the term exclusivity. just given. If the system (2) is medium fat. then the germ of p at qo is subanalytic. The only difference with the smooth case is that the manifolds and the sections are assumed to be realanalytic. It follows directly from the definitions that a system is medium fat if and only if its nilpotentization is. Let . The space of sections is endowed with the C ~ Whitney topology and is denoted by H k T M . [] The system (2) is said to be medium fat if: TqoM = A~o + span{[X.1)k + 1. This gives the following general result. . Then the germ of the subRiemannian distance function associated with a generic germ of a rank k distribution on an ndimensional realanalytic Riemannian manifold is subanalytic. Rank k distributions on M are smooth sections of the "Grassmannization" H k T M of the tangent bundle T M . We say that . Smooth families of distributions parametrized by the manifold N are sections of the bundle p N H k T M over N x M induced by the standard projection p g : N x M ~ M . It is proved in [5] that generic germs of distributions are medium fat for n < (k . we assume that everything is analytic. [Xi. . j = 1. .4 has codimension c~ in H k T M if the subset: {D E pNHkTM : D x• ~ A.Subanalyticity of Distance and Spheres in SR Geometry 3 Subanalyticity and nilpotentization 5 In this section. VX e N } . is everywhere dense in p N H k T M . T h e o r e m 5.1)k + 1. Xj]](qo) : i. k } for any X E z~. A s s u m e that n < (k . . while the topology remains the same Whitney topology. T h e o r e m 4. If the nilpotent system (2] does not admit nonzero GohLegendre abnormal controls. VN.
. (6) is called nilpotent if it coincides with its own nilpotentization expressed in adapted coordinates. . for the subRiemannian distance between q and the fixed point q0. . We will use the special notation p m : ~ n r JR+ for the subRiemannian distance in this case. polynomial and quasihomogeneous.) .. We keep the notation ] : L~ [0. There exists a subset . is called "the flat Martinet system". Recall that all over the paper we keep the notation p(q). . q E M. More precisely. . .6 Andrei Agrachev and JeanPaul Gauthier T h e o r e m 6..1 with respect to the dilation: 6t:(xl.t ([0. . . which plays an i m p o r t a n t role below...ttx~). J r . small enough. and the notation ~ : ~ n _+ ]R+ for the subRiemannian distance. .xt)~(txt. u). are are homogeneous t E ~. . 1 . Cot. . x i k . r])). P r o p o s i t i o n 4. Combining this fact with Theorem 6 we obtain the folowing result. . they of weight . x t ) . r]) for all r. i = 1 . xi = i = 1 .4] that the small subRiemannian balls are subanalytic for any realanalytic subRiemannian structure without nontrivial Goh controls.. The vector fields Y/.. Then there exists a polynomial submersion ~ : ]~n ~ IR3 such that ( p m ) ) . In other words. Y / = tY/. T h e o r e m 7. . r]) = ~ (fil([O. For any k > 3. the distributions admitting nonzero Goh controls form a subset of codimension oo in the space of all smooth rank k distributions on M. The subRiemannian distance is defined by a given distribution .. 5 Nilpotent Systems The system: k = i=1 E = 0. t5(x) = min{llul] : ](u) = x}. ..t~x~.5 ~ . where x(.A implies the subanalyticity of the subRiemannian balls p .. . so that any vector x E ~ n takes the form x = ( x i l . Assume that k = 2.. kl = 2.4 of codimension oo in the space of rank k realanalytic distributions on M such that the relation . ) E ]Rk~. JR" is presented as a direct sum ~ " = kt = k. A special case of the system (6) with n = l = 3.)). 1] ~ ]Rn for the endpoint m a p p i n g u ~+ x(1. u) is the solution of (6).. Vr _> O. I~ k~ 0 " .9 ]Rk'. k3 = 1. Suppose that M is realanalytic and k >_ 3. . k3 ~ O. uk(. ( x t . . ks = 0.t ([0.x2. i = 1.. It was proved in [1.4 on the Riemannian manifold M . . k . . . . .k. u = ( u t ( .
We restrict ourselves to the case of maximal possible ks. Under the conditions of proposition ~ the subRiemannian balls ~([0. r > 0. i. Finally. on a ndimensional manifold. It means ks = min{n . R e m a r k .Subanalyticity of Distance and Spheres in SR Geometry 7 C o r o l l a r y 2. C o r o l l a r y 3.1 ) ( ~ + ~ + 1) and k~. Then the balls pl([0. C o r o l l a r y 4. the maximal possible ks. combining Theorem 8 with Theorem 7 we come to the following surprising result. r > 0. Assume that either k = 2 and A qo 3 dimM> (kl) s ~ A qo Or +~+1 ) and the segment (k. r]). in particular. 3. and let p be the germ of the associated subRiemannian distance function. k3. Then there exists a polynomial submersion 9 : ]Rn ~ ~3 such that (pm))i (r) = .k. are not subanalytic. and Theorem 3 we obtain the following: T h e o r e m 8. Then z " 3 p is not subanalytic. Generic germs of distributions and their nilpotentizations have the maximal possible growth vector and. generic germs are such that p is not subanalytic. In particular. k3.e. r]). Under the conditions of Proposition 5. Let now A be an arbitrary (not necessarily nilpotent) germ of a bracket generating distribution at q0 E M.~) (/~l(r)). k(k 2 1) } k3 = min{n k(k2+ 1).l ) ( ~ + ~ + 1).dim Aq0 . P r o p o s i t i o n 5. are not subanalytic. Vr > 0. &m Aq0) of the growth vector is maximal. r]) are subanalytic for all small enough r. Combining corollaries 2. Assume thatn > ( k .k3 are maximal possible. Now consider nilpotent distributions of rank greater than 2. but the function p is not subanalytic/ . (k + 1)k(k3 1)}. the subRiemannian balls ~([0. k = kl > 2. n > ( k . Let p be a germ of subRiemannian distance function associated with a generic germ of realanalytic distribution of rank k >_ 3.
Zhong Ge. Filtrations of a Lie algebra of vector fields and nilpotent approximation of control systems. Local invariants of smooth control systems. v.28. Agrachev. 5. 377448. Agrachev. B. J. 8394. Kupka. Bianchini. Optimisation and Calculus of Variations. 11. A. H. Tr~lat. SubRiemannian metrics: minimality of abnormal geodesics versus subanalyticity. Chyba. to appear. In: A. M. J. 1999. J. Agrachev. A. Longman Publishers. G. Zeitshrift. Chyba. Pitman Research Notes in Mathematics..'98. J. v.36. Agrachev. Optimal control and piecewise analyticity of the distance function.2. K. Graded approximations and controllability along a trajectory. A. Agrachev. ESAIM: Control. Armales de l'Institut Henri Poincar~Analyse non lin~aire. M~thodes ggomgtriques et analytique pour gtudier l'application exponentiele. 13. 10. 191237. 178. 1938. Subanalyticity of the subRiemannian distance. E. A. 298310. Mathem. Compactness for subRiemannian lengthminimizers and subanalyticity. Zapiski Ped. Bormard.255286. English transl. Proceed. 1988. Inst. Sept. 377403. 1998. v. Akad. S. Submited to Annales de l'Institut Henri PoincarC . Sarychev. Sussmann. Dedicated to Pontryagin. Mathem. Agrachev. R.161. A.P. March 2000. A. V. Control Optim. ESAIM: Control. v. Libknechta. Gamkrelidze. Ioffe. Observability for systems with more outputs than inputs and asymptotic observers. ESAIM: Control. I. Gauthier. Bormard. R. A. M. 903924. v. Optimisation and Calculus of Variations. B. v. Rend. M. Dokl. A. B. P. Horizontal path space and CarnotCaratheodory metric. 1987. S. 8. v. 104108. Uch.2. Agrachev. Eds. A. The tangent space in subRiemannian geometry.295..8 Andrei Agrachev and JeanPaul Gauthier References 1. Torino. v. 9. A. G.4. 1996. 4. JP. Optimisation and Calculus of Variations. Jacquet. Semin. Birkhs 1996. About connecting two points of a completely nonholonomic space by admissible curve. Confer.14. In the book: "SubRiemannian geometry". 777781. A. A. 635690. V. Bellai'che. Abnormal subRiemannian geodesics: Morse index and rigidity. V.. 1999. Moscow. in Soviet Math. A. A.223. The transcendence we need to compute the sphere and the wave front in Martinet SRgeometry. n. Sarychev. 1997. 6. 1990. Sarychev. A. I. J. 15. SubRiemannian sphere in Martinet flat case. Launay. Sarychev. Rashevsky. Acta Applicandae Mathematicae. Pacific J. Stefani.13. 1996. Reich. la sphere et le front d'onde en ggometrie SR dans le cas Martinet. Kupka. Dokl. v. 16. On Subanaliticity of CarnotCaratheodory Distances. 3. Mat. V. 12. SIAM J.5 14. v. 7. A. A. Int. v. 1989. 1993. V. Gauthier. 1992. Nauk SSSR.56 2. 4778. 3. A.. Bonnard. Dynamical and Control Systems. submitted.
Suppose 7 is a s eg m e n t of a characteristic curve and O N is a n e i g h b o r h o o d of 3' such t h a t N = O T / ( C I o . although it can be read independently. ~ = A o lr. then r O O r ) .d~ be the s t a n d a r d symplectic s t r u c t u r e on T ' M . then Hq = H O TqM is a s m o o t h hypersurface in TqM. Actually. Let . where Jacobi curves were defined. Russia agrachev@sissa. Jacobi curves are curves in the Lagrange Grassmannians. Let r : O N + N be the canonical factorization. T h e kernels of (qlH)x. these kernels form a line d i s t r i b u t i o n in H and define a characteristic 1foliation C of H . 2].I. Via Beirut 24 34013 Trieste. Let ~ be the canonical Liouville f o r m on T~M. Israel z igor%t echunix. q E M . are L a g r a n g i a n s u b m a n i f o l d s in N .S. i l A b s t r a c t . Leaves of this foliation are characteristic curves of alH. and a fundamental form. and ~r .. Differential geometry of these curves provides basic feedback or gauge invariants for a wide class of smooth control systems and geometric structures. then O'IH is a corank 1 closed 2form. which is an invariant of the parametrized curve in the Lagrange Grassmannian providing the curve with a natural projective structure.S.A. Gubkina 8 117966 Moscow.Principal Invariants of Jacobi Curves Andrei Agrachev 1 and Igor Z e l e n k J 1 S. Vq E M. ) is a welldefined s m o o t h manifold. ac. This paper is a continuation of the works [1. which is a 4order differential on the curve. A E H are transversal to 7q'M. 1 Introduction Suppose M is a s m o o t h m a n i f o l d and 7r : T * M + M is the c o t a n g e n t bundle to M. Italy and Steldov Mathematical Institute ul. t echnion. T h e quotient manifold N is in fact a symplectic manifold endowed with a s y m p l e c t i c structure ~ induced by alH. i t 2 Department of Mathematics. E T ' M . In the present paper we mainly discuss two principal invariants: the generalized Ricci curvature. TechnJonlsrael Institute of Technology Haifa 32000. Let H be a c o d im e n s i o n 1 s u b m a n if o l d in T * M such t h a t H is transversal to TqM. q E M . Jacobi curves are far going generalizations of the spaces of "Jacobi fields" along Riemannian geodesics.
where the restriction of a to the submanifold is degenerate. A similar construction can be done for a submanifold of codimension 2 in A 9 7. ( A . Agrachev. which we do not discuss here.10 A. Let Vq = V NTqM. Jacobi curve is the mapping A ~~ r from 7 to L(T~N). L(T~N) = {A C T~N : A L = A}. T*M.e. We are mainly interested in submanifolds that are dual objects to smooth control systems. Using the fact that the spaces A(t) are Lagrangian. in general. take a curve l(. it is easy to see that the form qA(. are equivalent (under linear substitutions of variables in the correspondent quadratic forms) to the "second fundamental form" of the hypersurface H~(x) C T~(x)M at the point A. where D z = {e E T~N : ~(e.)(l) depends only on d A ( 0 ) . one can obtain a lot of information on the original system just studying smooth parts of H 1. take a curve A(t) 9 L(W) with A(0) = A.D) = 0}. In particular. They are invariants of the hypersurface H in the cotangent bundle. Anyway. they are concentrated in the characteristic variety consisting of the points. Tangent vectors to the Jacobi curve "17 at a point JT(A). we have the map from TAL(W) to the space of quadratic forms on A. The "dual" normal variety H 1 and abnormal variety H ~ are defined as follows: HX= [J { A E T q M : 3 v E qEM Vq. transversal to fibers.)(l) = &(dl(0). the velocity of J'r at A is a signdefinite quadratic form if and only if the hypersurface nr(x) is strongly convex at A. A simple counting of dimension shows that this mapping is a bijection. P r o p o s i t i o n 1. H ~ . So. v ) = 1. they may have singularities. I.T~Vq)=O}. smooth manifolds. i. A(t) z = A(t). Set W = T~N and note that the tangent space TAL(W) to the Lagrange Grassmannian at the point A can be naturally identified with the space of quadratic forms on the linear space A C W. Namely. Jacobi curves are curves in the Lagrange Grassmannians. H~= U qEM e T. In particular.) in W such that l(t) 9 A(t) for all t and l(0) = I.M \ O: These varieties are not. In the codimension 2 case characteristic curves do not fill the whole submanifold. Define the quadratic form qA(. = = 0}.e.~N. Given some vector l 9 A.l). (A. Zelenko L(TTN ) be the Lagrange Grassmannian of the symplectic space TTN . VD C T. Here we call a smooth control system any submanifold V C T M . A 9 7. any differential invariant of the curves in the Lagrange Grassmannian by the action of the linear Symplectic Group produces a welldefined function on H. i.
then.e. the intersection A A ( w + A ) of the .O and ~ vanishes on the tangent lines to the characteristic curves of aIHo. 2. while H ~ has codimension > 2 in all nontrivial cases. If Vq are compact.~Vq + ~v) • is contained in Hq ~ Suppose that H 1 has codimension 1 in T*M and 7 is a characteristic curve of aiH~. It happens that the rank of velocity of any Jacobi curve generated by the system never exceeds the number of inputs. Indeed. Dimension of Vq is the number of inputs or control parameters in the control system. then H ~ = I~. which is contained in He. i.1. then H 1 has codimension 1 in T ' M . induces a canonical isomorphism A ~. Then the afline subspace A + (T. If subsets Vq C TqM are conical. where Wq is a convex set and 0 E intWq. for some v Vq. w E A. H ~ can be empty. see [1. Va > 0. in this case the Liouville form never vanishes on the tangent lines to the characteristic curves of CrIH~. let A E H~.Vq. Then the velocity of the Jacobi curve A ~~ Jr(A).Principal lnvariants of Jacobi Curves 11 Characteristic curves of ~1~ (~1~) are associated with normal Cabnormal) extremals of the control system V. A ~ possesses a canonical structure of an affine space over the linear space SyruP(A) = Sym2((W/A)*). T h e same is true for the Jacobi curves associated with characteristic curves of a]H0. T h e nondegeneracy of ~ implies that the relation w ~~ or(. T h e corresponding Jacobi curves admit a purely variational construction in terms of the original control system and in a very general setting (singularities included).. For A E H ~ 3v E Vq such that A E (TvVq) • CA. 2 Derivative Curve Let A be a Lagrangian subspace of W. For any w E A. has rank no greater than dim V~(x) (see proposition 1). if H ~ has codimension 2. The characteristic curves are actually unparametrized. A E L(W). Less inputs means more "nonholonomic constraints" on the system.(W/A)* and. H i . and any characteristic curve 7 has a canonical parametrization by the rule C~. then (T Vq) • = 1. T h e rank of the "second fundamental form" of the submanifolds H~ and Hq ~ of TqM at any point is no greater than dimVq. w) vanishes on A and thus defines a linear form on W/A.O. A E 7. We set A r = { F E L(W) : F n A = 0}.dim Vq . 3]. for any A E A r and coset (w+A) E W/A. the linear form #(.'Y) = 1. by the conjugation. w). We have + +Rv) C So A belongs to an affine subspace of dimension n . if Vq = OWq. v) . Indeed. aVq . In particular. an open everywhere dense subset of L(W). A* ~ W/A. in contrast to the previous case. this notation reflects the fact that Sym 2 C is the space of quadratic forms A) on A* that is the symmetric square of A. Let Sym2(A) be the space of selfadjoint linear mappings from A* to A. Moreover. One of the varieties H 1.
The coordinates in A induce the identification of S y m ~ A with the space of symmetric m • m matrices. A subspace F0 = {(x. (1" . Agrachev. A) = ( A .4 is represented by the matrix S. If A N F = 0. ker(A.12 A. . ~'((xl.A) a selfacljoint m a p p i n g from A* to A. this mapping has a much more straightforward description. y 2 ) . Zelenko linear subspace A and the affine subspace w + A in W consists of exactly one point. S0x) : x E ]Rn} is transversal to F if and only if det(S . F. It is easy to check that the identification W / A ".) is ample at r.(x~. In particular. F. We say that the curve A(.4 as the origin in this affine subspace and thus obtain a coordinatization of A ~ by symmetric m • m matrices. St• : x E ~ n } is ample at v if and only if the function t ~ det(St .A) is a onetoone m a p p i n g of A m onto S y m 2 (A) and the axioms of the affine space are obviously satisfied. A E A m there corresponds a m a p p i n g ( F . the correspondence F ~ ( F . S x ) : x E ]Rn} in A ~" is represented by the matrix S 1.St) has a root of finite order at v. A = IR" (~O. A) is represented by the matrix S 1 . we fix . We obtain the curve t ~+ A(t) E A(r) ~ in the affine space A(r) ~" with the pole at r. F. F n A .A ) ( w + A) d~_f F A ( w + A ) . In particular. F. Let us pick coordinates {x} in F0 and fix . y ~ ~m}. then (m.A) E S y m ~ (A) turns into the m a p p i n g from A to A. / ' . A). The operator (A.A) : W / A + A.yl). In coordinates it takes the following form: the curve t ~~ {(x. A) = A n F.So) r O. Such a vector . Let t ~~ A(t) be a s m o o t h curve in L ( W ) .A n (w + A ) .) of the kjet of A(. given A E A "~. Moreover. In particular. y ) : x . F.{(x.4"~ takes the form F .4) : A ~ . Namely. For the last linear m a p p i n g we will use the notation (A. To a pair F. while the operator (A. I.) at r. A) : A + A. Let us give coordinate representations of the introduced objects. Then A(t) e A ( r ) ~ for all t from a punctured neighborhood of v. where ( F . St such that Akr(t)n A(r) = 0.0.A* makes (F .yl). Actually. (x~. A) 1. Then any F E . F.4 as the origin in the affine space F0~. In the induced coordinatization of F0 the "point" F is ~ represented by the matrix (S . the "point" F = {(x. Sx) : x E ]~n}. the relations W = A $ A. A r is an affine subspace over S y m ~ A . The curve A(. We m a y assume that W=~"e~t" = { ( x . Fixing A E A ~" one obtains a canonical identification A _~ W / A = A*. where S is a symmetric m • m matrix. imply that T' is the graph of a linear m a p p i n g from A to A.4 = 0 ~ ]~ '~. . it is the graph of the m a p p i n g (A.) is ample at v if Sk > 0 such that for any representative A~ (. Assume that A(. We have given an intrinsic definition of an ample curve. Fixing an "origin" in A(r) ~" we make A() a vector function with values in Syrn 2 (A) and with the pole at v.) is called ample if it is ample at any point.So) 1.y2)) = ( x l . In fact.
Principal Invariants of Jacobi Curves 13 function a d m i t s the expansion in the Laurent series at v. In other words.C i v i t a connection ~7). A ( r ) ) . Note t h a t A(t) is linear o p e r a t o r from A(t) to A(t)* and Ji0(t) is linear o p e r a t o r from Ao(t) to Ao(t)*. Since the form ~ defines the canonical i s o m o r p h i s m between Ao(t) and A(t)*. A n o t h e r c h a r a c t e r i z a t i o n of Ao(r) can be done in t e r m s of the curves t ~~ (A. A E A(r):". T h e n Ao(r) E A(r) ~ is defined for all r.v a l u e d function t ~+ (St . the following o p e r a t o r R(t) : A(t) + A(t) can be defined: R(t) = . 3 Curvature operator and regular curves. A(t). only free t e r m in the Laurent expansion d e p e n d s on the choice of the "origin" we d i d to identify the affine space with t h e linear one. t a k e velocities A(t) a n d A0(t) of A(t) and its derivative curve Ao(t). for the L a u r e n t e x p a n s i o n of a curve in an affine space. see [1].) is ample. where A(t) = {(x. Using derivative curve one can construct an o p e r a t o r invariant of the curve A(t) at any its point. T h e curve A(t) in L a g r a n g e G r a s s m a n n i a n is called regular. A(t). T h e Laurent expansion at t = r of the vector function t . which a p p e a r s in the classical J a c o b i equation XT~(t)XT~(t)V+ RV(~/(t). V)Zy(t) " 0 for Jacobi vector fields V along the geodesic 7(t) (here R v is c u r v a t u r e tensor of L e v i . These curves have poles at r . if the ." in (2). Namely.+ (A. A(r)) in the linear space Horn(A. T h e curvature o p e r a t o r can be effectively used in the case of socalled regular curves.ST) 1 at t = r .1) is the Laurent e x p a n s i o n of the m a t r i x . the a d d i t i o n o f a vector to the "origin" results in the a d d i t i o n of the s a m e vector to the free t e r m in the Laurent expansion. Obviously. T h e coordinate version of the series (2. In p a r t i c u l a r . A(t) ~ Ao(r) + ~ i#o (t . Note t h a t in the case of R i e m a n n i a n g e o m e t r y the o p e r a t o r R(t) is s i m i l a r to the socalled Ricci o p e r a t o r v + RV(~/(t).A o ( t ) o A(t) (2) This o p e r a t o r is called curvature operator of A at t.Stx) : x E I~n}. More precisely. the free t e r m of the expansion is a p o i n t o f this afiine space while all other t e r m s are elements of the c o r r e s p o n d i n g linear space.). Q i ( r ) e Sym2A(r). This is the reason for the sign " . (1) where Ao(r) E A(r) ~. Assume t h a t the curve A(. A(r)) has zero free t e r m if and only if A = Ao(r).r ) i Q i ( r ) ./(t). T h e curve r .+ A0(v) is called the derivative curve of A(. v).
)..A1.' . 2 . the function t . In particular. Six) : x E ~n}. and A3 are Lagrangian subspaces of W and Ao and A1 N A2 = A3 CIAo = 0.r. 1 ~(3) 1 "WIt" 2 acting on scalar function ~ is called Schwarzian.v and one can get the following formula for the curvature operator (see [1]): R(t) = ((2s. I. the crossratio takes the form: [Ao. as we will see later in the nonregular case the curvature operator does not determine the curve up to a symplectic transformation. )2 (3) Note that the righthand side of (3) is a matrix analog of socalled Schwarz derivative or Schwarzian .A3]= (Ao. A2.A1.t ~ is a MSbius transformation (with constant coefficients) of some particular solution of this equation..+ (St .$2) L J (6) .A2. AI. Indeed. Zelenko quadratic form Ji(t) is nondegenerated for all t.i s. Stx) : x 9 ~n}.$3)1 ( $ 3 . .A3]= ($2 . . in coordinates {x}. A3. using the notion of crossratio it is possible to construct numerical invariants for a very nonrestrictive class of curves.A2. and A3 in the Lagrange Grassmannian is. then.So)(So .i ) . A2) : Ao + A2. For the nonregular curve A(t) = {(x. The crossratio IAo. Aa. (A2. 4 Expansion of the crossratio and Ricci curvature.((2s. let Ai = {(x. Agrachev. Then the function t ~+ S~1 has a simple pole at t =. Ao) : A2 + Ao. (5) This notion is a "matrix" analog of the classical crossratio of four points in the projective line.A3J of four "points" Ao.14 A.I ( S 1 . . ST = 0.ST) 1 has a pole of order greater than 1 at r and it is much more difficult to compute its Laurent expansion. It implies that in the regular case the curvature operator R(t) determines the curve completely up to a symplectic transformation. A1. Stx) : x E I~n}. A1.S 1 ) . A1.A2. A2. The operator ~ is characterized by the following remarkable property: General solution of the equation S~ = p w.A2)(A2. the following linear operator in A2: L J [Ao. At. concerning "matrix MSbius transformation" of the type (AS + B)(CS + D)1. We have (Ao. by definition. However. Suppose that the curve A() is regular and has a coordinate representation A(t) = {(x. The matrix analog of this operator has similar property. Let us recall that the differential operator: 1 " d ~o" " ~o" . Ao).). Suppose that Ao.
t) in the following way: f(to. Aa I a r e invariants of four subspaces A0. t2.tl)[ (9) This function is also symmetric. tl. A(t3)] is symplectic invariant of the curve A(t). (7) where Lt~ = (tl_t~)(ta_to) to. we have: det(S.tl. by Assumption 1. A(t2).. t . as before. It follows that X(to.t2) + f(t2. For this it is very convenient to introduce the following function det[A(to)'A(tl)'A(t2)'A(t3)] . t) # 0 )./(t) = aj. t2. tl) is a symmetric function (changing the order in (8) we obtain that X can be symmetric or antisymmetric.tl) = lnlX(to. 15 Now we are going to show how to use the crossratio in order to construct invariants of the curve A(t) in the Lagrange Grassmannian.tx.i(t) (10) One can easily obtain the following lemma. tl.t~. By the above the function (to.f(tl.Stl) have at t = tl zero of the same finite order k. so it can be expanded in the formal Taylor series at the point (t. X ( t .~ {(x.t2. tl. It can be easily expanded in formal Taylor series at any "diagonal" point (t. t ) ~s 0 (8) for any t.tz) = In [to. G(to.ta) = f(to.t l ) k x ( t 0 . t) and the coefficients of this expansion are invariants of the germ of A(.t) j.t. A1. all coefficients of the characteristic polynomial of A0. Indeed. A1.tl. a ( t l ) . o .S t 1 ) = ( t 0 . t3) is also a symplectic invariant of A(t). but the last case is impossible by the fact that X(t.).and tz.t3) . S t x ) : x E ~n} be the coordinate representation of a germ of the curve A(. t. t l ) . t2. using the fact that G(to.tl) ~ Z i. and Aa.j=O ai.tz. t3) r det [A(t0). let us try to find symplectic invariants of A(t) that are functions of t. ai.Principal Invaxiants of Jacobi Curves By construction.t)i(tl . A2.f(ta.) at t. Define also the following function f(to.j(t)(to .tl) . The function G(t0. Using this fact.to) .t3]k . As. Let. A s s u m p t i o n 1 For all parameters tl the functions t ~ det(St .
1. t 3 ) has the following formal Taylor expansion at the point (t . 1 . ~o(rl). In the sequel the Ricci curvature is denoted by p(t). . For any t the function 6 ( t o . t .7"2. (14) .t .2. one can define a canonical projective structure on the unparametrized curve A(. v3) ~~ In ([~(r0). T h e n from (7) it follows that (70. t ) : O0 G(to. The first appearing coefficient a1. i . In some cases we are interested in symplectic invariants of u n p a r a m e t r i z e d curves in Lagrange G r a s s m a n n i a n (i. 3. t 2 . t . using the Ricci curvature.2. I = 0 . 7"1.e.~(Tz). socalled a b n o r m a l extremals of vector distributions and consequently their 3acobi curves a priori have no special parametrizations. by a direct computation. i = 0.~ ) .1 (t) is called Ricci curvature of A(t).~l ] expansion up to the order two at the point (v. / where S ~ is Schwarzian defined by (4) and T/i = T / . Zelenko 1.tl. j > 1. ~(73)]'~ = ~ ( 7 ) ( r / o .y3) + .16 Lemma A.~(r. i = 0. (12).T. Denote by the function playing for A(7) the same role as the function ~ for A(t). v): In ([qo(70). 73) = ~(to.~(71).. D e f i n i t i o n 1. Let t = ~ ( r ) be a reparametrization and let A(r) = A ( ~ ( r ) ) . ~o(r~).). 3 . (11) From (11) it follows that all coefficients h i s ( t ) . i.~ ) ( ~ . this relation justifies the n a m e Ricci curvature for the invariant p(t). tl. 2 .~(r~). are symplectic invariants of the curve A(t). I.t2. v2. T2. By direct c o m p u t a t i o n it can be shown t h a t the function (T0. Note that.t3) ~ ~ a i h ( t ) ( ~ . and (13) it follows easily that: fi(v) = k ~ ( r ) . of onedimensional submanifolds in Lagrange Grassmannian).(13) \ [70.~.~(r~)]~ has the following Taylor k [~o. 3. Agrachev. A c t u a l l y .t3) + k In ([~o(ro). vl.(12) where ti = ~(7i). For example. 7. Suppose for simplicity that in the original p a r a m e t e r t the Ricci curvature p(t) = 0 and denote by/5(v) the Ricci curvature of the curve A(7). 7. Now we want to show how.).t2. t l . For this let us check how the Ricci curvature is transformed by a r e p a r a m e t r i z a t i o n of the curve A(t). T h e n from (11).r/2)(r/t .~. 1. ~O(T3)]) .j=t whereat = tt . one can get the following relation between p(r) and curvature operator R(r) for the regular curve: p(r) = 89 R ( r ) .
2. if the Ricci curvature p(t) of the curve A(t) is not identically zero we can find at least locally (i. Now let t be a projective parameter on A(.(~3 _ ~3)(~1 _ ~3) + 3  . t2.e. t. The analysis of the next terms of this expansion gives the way to find other invariants of the curve A() that do not depend on p(t).(i + 1)ai+l. 2.2(t) = 1. ~2.).O.) with Ricci curvature identically equal to zero defines a projective structure on A(. In this section we show how to find candidates for the "second" invariant of A() and then we construct a special form on unparametrized curve A(.t3) has the following Taylor expansion up to the order four at the point (t. 5 Fundamental form of the unparametrized curve. t. ~3) . t.1(t) = crl. ~1.3(t) implies that the function g(to. 3.Principal Invariants of Jacobi Curves 17 Conversely.. ~3) is a homogeneous polynomial of degree four (more precisely.1 (t) is a candidate for the second invariant (as well as the function a2. ~2. i : O. Then by definition p(t) ..2(t) = 30~3.).) ).) (any two parametrization from this set are transformed one to another by MSbius transformation). t2. t): G(to. and by (16) o/2j(t) ~ 0 and ot2. ~3) + . which we call the fundamental form of the curve A(. (t). the differential of order four on A(.t~. (17) where ~i : ti .t(t). t). The parameters of the canonical projective structure will be called projective parameters.1 (t)p4 (~0.. P4(~0. The Ricci curvature p(. (1.) (namely.~(t)).t. 1.. tl. Let oti. This together with (11) and the fact that a3.. t. ~ .(~2) The set of all parametrization of A(. (t)  3 (16) These relations imply that the function a3. ~l. . from (15) it follows easily that = (15) 1.. Then the following relation holds a~j(t) =. First note that analyzing the expansion (10) one can easily obtain the following lemma L e m m a 2.) is the first coefficient in the Taylor expansion of the function G at the point (t.j(t) + (j + 1)(~ij+l(t) In particular. We call it the canonical projective structure of the curve A(.j(t) be as in expansion (10).). and/94 (~0. t3) = cr3. in a neighbourhood of given point) a reparametrization t = ~o(r) such that fi(r) = 0 (from (14) it follows that in this case ~ ( r ) has to satisfy the equation S(~o1)(t) = P_.
using (11). In this section we consider nonregular curves having two functional invariants and prove that the above defined invariants p and A constitute a complete system of symplectic invariants for these curves (completeness of the system of invariants means that the system defines the curve uniquely up to a symplectic transformation) Assume that dimension of the symplectic space W is four and consider ample curves A(t) in L(W) such that for any t the velocity Ji(t) is a quadratic form of rank 1. then the fundamental form in this parameter has to be of the form A(t)(dt) 4. and the fact that the crossratio is preserved by Mhbius transformations it follows that 4 = as. then the canonical length element IA(t)188 is defined on A(). Since the curve A(t) is ample. then by (14) it follows that pn(7) = k S~(7).(i. by our assumptions..1(t)(dt) 4 does not depend on the choice of the projective parameter t. If t is an arbitrary (not necessarily projective) parameter on the curve A().18 A. Note that all constructions of this section can be done for the curve in the Grassmannian G(m. For projective parameter A(t) = az. length with respect to this length element) is called a normal parameter (in particular. (12). We will call this form a fundamental form of the curve A(. I. Then from (12). For arbitrary parameter it can be shown. we obtain a functional invariant of the unparametrized curve. it implies that abnormal extremals of vector distribution may have canonical (normal) parametrization). Let us fix some parameter r and let {(x.). Calculating the mcci curvature Pn (7) of A(. The velocity ~S. If t = ~(v) is the transition function between a projective parameter t and the normal parameter v.1 p(t) 2 If A(t) does not change sign.. Stx): x e A n} be a coordinate representation of A(t) such that Sr = 0.l(t) . 6 The method of moving frame. t = ~(v) = =r+d and denote by ds.l(r) the function that plays the same role for the curve A(~(v)) as the function a3. Without loss of generality we can assume that Ji(t) is nonnegative definite for any t. Agrachev. (17).be another projective parameter on A(. where A(t) is a smooth function (the "density" of the fundamental form). We will call it projective curvature of the unparametrized curve A(. The corresponding parameter 7. ~ S t 1 is .a1 : A(r)* + A(r) is d a well defined selfadjoint operator. Zelenko at+b) Let 7.e.l(t)(dt)' (18) It means that the form ~3. that A(t) = az.1(t) for A(t). 2m) ( the set of all mdimensional subspaces in the 2mdimensional linear space) instead of Lagrange Grassmannian by the action of the group GL(2m) instead of Symplectic Group. the curve S t I has a pole at r.) (i.) in the normal parameter. l(t). Moreover.e.).
).ej(r)) = 8i. one can choose two vectors fl(r) and f2(r) on the plane Ao(r) such that four vectors (el(r).one can take (el(r). Note that the curve .) defines a moving frame (el(r). It is not hard to show that the order of pole of t + w(t. So. e2(r)) or ( .t4 7)]t=r.). the vector e l ( r ) is also defined up to the sign.v)tw(t. The natural question is whether (p(. f2(r)) constitute ] symplectic basis (or D a r b o u x basis) of W (it means that cr(fi(r). vector w(t. Since vector w(t. e l ( r ) and e~(r) are linear independent if and only if I = 2). So. in addition to the vectors el(r). e2(t).) can be described by two functions. e2(t) are linear independent. e2(r) on the plane A(r). i. By the above it is easy to see that if A(. So.j). then it satisfies Assumption 1 with k = 4.fl(r). Suppose that the order of the pole is equal to I. for any 7. the invariants p() and A() are defined for A(. So for t ~ 7.) satisfies the previous assumption. f l ( r ) . r) E A(r) such that for any v E m(r)" d _. the curve . Recall that by constructions of the section 2 for the curve A() the derivative curve A0(') is defined and for any 7the subspaces A(r) and A0(v) are transversal. . . r) at t = r is equal to I if and only if I is the minimal positive integer such that the vectors el (7) and et(r) are linear independent (in particular. 7) = ( t . r) of the vectors on the plane. It implies easily that the set of points 7.4(.) can be described by the curve t ~ w(t.e 2 ( r ) ) as the canonical bases on the plane A(r).) = (19) It is clear that the curve t 4 w(t.e. up to the sign. r) is defined up to the sign. A(.f2(r)) and one can derive the structural equation for this frame: . A s s u m p t i o n 2 A(t) is a curve in L(W) with dim W = 4 such that for anyt the velocity A(t) is a quadratic form of rank 1 andet(t). Denote by u(t.e l ( r ) . where the vectors el(r) and e2(r) are linear dependent.there exists unique.4(. A(.)) is a complete system of symplectic invariants of A(). 7) and define the following vectors in A(v): ej(r) 0ja U~% Otj'Tj~'. 7) also has the pole at 7.Principal Invariants of Jacobi Curves 19 a nonpositive selfadjoint operator of rank 1.e2(r). v) the normalized curve t + u(t. is a set of isolated points in IR. So.
The curve A(t). Then the sign of A(t) and the projective curvature Pn(') determine A(.3fi Note that the coefficients in the equation (20) depend only on p and A and any symplectic basis can be taken as an initial condition of (20). It implies the following: T h e o r e m 1.. satisfying Assumption 2. I.) satisfying Assumption 2 is determined by its invariants (p(). e2(r).1_2_ A 2 (20) p")el . Assume that the curve A(. A(. The frame (el (r). It can be shown by a direct calculation that the curvature operator R(r) : A(r) + A(r) of the curve A() satisfying Assumption 2 has the f~176 the basis (ei(r)'e2(r)): R(r) = ( ~ 88 .i. 7 Flat curves. Zelenko Proposition 2.) satisfies Assumption 2 for some parametrization and its fundamental form A(t)(dt) 4 does not vanish. the curvature operator does not determine the curve A() uniquely up to a symplectic transformation.20 A. The following definition is natural. f2(r)) satisfies the following structural equation: l ea = 3e2 e 2 = ~pel + 4f2 1 l i.) uniquely up to a symplectic transformation of W and a reparametrization. Agrachev. A(t) .88 7 9 i~P! el .R depends only on p. fi (r).( n35. The curve A(. D e f i n i t i o n 2. expansion (11). As a consequence of Theorem 1. Theorem 1 implies the following result on unparametrized curves: T h e o r e m 2.O. = . and structural equation (20) one can obtain the following characterization of the fiat curve: .e. This means that in contrast to the regular case.5p'e . Remark 1.~pe2 . is called fiat if p(t) =_ O.)) uniquely up to the symplectic transformation of W.
Dynamical and Control Systems. and 3) are also equivalent for regular curves in L ( W ) with symplectic space W of arbitrary dimension (we only need to replace the power 4 in (21) by dim W).1997. No. If the previous conjecture is true. second variation and rigidity.rVIaslov index of this curve equals 2. Jacobi Curves for Singular Extremals. are equivalent. Feedbackinvariant optimal control theory . 2. Agrachev. and 3). D e f i n i t i o n 3. 4(1998). No. A. A(t3)]) = [to. R. is called flat.. 2).A. 4 . Nonregular abnormal extremals of 2distribution: existence. t3 det ([A(t0). An unparametrized curve A().II. Then the conditions 1). Agrachev.V. All other flat curves are images under symplectic transformations of L ( W ) of the segments of this unique one. 3. with the power 4 replaced by k in relation (21). 583604. J. t3] 4. Dynamical and Control Systems. 2). It happens that. t2. (21) The conditions 1). 2). i. Suppose that A(t) satisfies Assumption 1.e. A(tl). is flat if and only if one of the following condition holds: I) all coe~cients Qi(t) with i > 0 in the Laurent expansion (1) are equal to zero. satisfying Assumption 2. if its f u n d a m e n t a l form is identically zero. A(t2). Dynamical and Control Systems. No. 3. Now let us discuss the notion of flatness for unparametrized curves. The curve A(t). 343389. Regular extremals. There is an embedding of the real projective line ~IP~t into L(W) as a flat closed curve endowed with the canonical projective structure. Zelenko. J. 347383. 3. C o n j e c t u r e . In the last case these conditions are also equivalent to the fact t h a t curvature operator R(t) is identically equal to zero. t2. there exists exactly one maximal flat curve. J. . then one of the conditions 1). A. tl. 2) the derivative curve An(t) is constant. Gamkrelidze. or 3) can be taken as a definition of the flat curve. up to symplectic transformations and reparametrizations. Theorem 4. tl. I. 3) for any to. References 1.A. 3. Feedbackinvariant optimal control theory I. . 5(1999). satisfying Assumption 2 for some parametrization. Ji0(t) = 0.Principal Invariants of Jacobi Curves Theorem 21 3.
Working in ~n. We mention a m o n g others [5. In [12.16]. 1 Introduction The group of rigid body transformation S E ( 3 ) arises naturally as the configuration space of m a n y robotic systems like aerial [8] and underwater vehicles [9] or robotic manipulators [10. the generalization of closed form m e t h o d s like the De Casteljau algorithm is investigated for c o m p a c t Lie groups for which a "natural" (i.12. the search for m e t h o d s for the generation of s m o o t h trajectories on SE(3) has given rise to a rich field of literature.7.14. completely frameindependent) Riemannian structure exists. A known negative result reported in the above mentioned literature is that S E ( 3 ) cannot .The De Casteljau Algorithm on SE(3) Claudio Altafini Optimization and Systems Theory Royal Institute of Technology SE10044. In particular.17]. the natural way to generate trajectories is to solve some optimization problem. When one tries to extend such m e t h o d s from ~ n to a noncommutative Lie group.13. geometric techniques seem to appear naturally when one wants to construct trajectories in an invariant and coordinatefree way. for example minimize energy or acceleration.11.e. the resulting curve depends on the choice of the metric tensor. one would like to have a notion of distance that does not change with the way it is measured. the resulting optimal curves are normally polynomial splines. The two m o s t c o m m o n cases are analyzed. The same trajectories can also be generated in an analytic way by considering the polygonal support of the curve. Abstract. Sweden altaf ini@math. It is straightforward then to describe the problem using tools from Riemannian geometry and for SE(3) this is done is a series of papers [1.4]. T h e use of variational techniques for trajectory generation on matrix Lie groups is investigated for example in [3. Due to the lack of a biinvariant metric on SE(3). kth. Motivated by motion planning purposes for such systems. For example.15]. se Smooth closedform curves on the Lie group of rigid body motions are constructed via the De Casteljau algorithm. but rather that it represents as much as possible an intrinsic property of a system. Given a certain number of boundary conditions. as the different constructions no longer coincide.16]. using constructions like the Bdzier polynomials or schemes like the De Casteljau algorithm [6]. Stockholm. the situation becomes more complicated.
d i m e n s i o n a l s p a c e . We will call them Adinvariant pseudoRiemannian and doublegeodesic. ~ T _ _ _ ~ } where the operator ? is called the wedge product and constitutes the linear representation of the vector product in/~a i. RR T=/3 and d e t R = + l } SO(3) is a matrix Lie group with respect to the matrix product.~ ~ X O" VoE~ 3 .). This complicates things further as the choice of the metric tensor is task (or designer) biased.: 2 b. function only of the boundary data (and of the metric structure) so that it can be useful in applications in which a (noncausal) trajectory exactly matching the data is required. called ~0(3) the Lie algebra of SO(3). 2 Basic properties of SE(3) R o t a t i o n s in 3 .e.~1 0 corresponding to .t.t. the Special Orthogonal Group SO(3) defined as: S O ( 3 ) = {REGI3(I~) s.e.24 Claudio Altafini be endowed with a biinvariant Riemannian metric i. In the above mentioned works (see also [2]) a couple of preferred metric structures emerges. instead.e. Both choices have advantages and disadvantages according to the task in mind. The first consists in choosing an inner product which is nondegenerate but that can assume both negative and positive values. univocal concept of distance. : I~3 . In the second case. Consider the set of rotational matrices i. For sake of simplicity. it does not admit a natural. is the vector space of skewsymmetric matrices: s0(3) = {~ e M3(~) s. The advantage of such an algorithm with respect to the variational approach is that it gives a curve in closed form. The scope of the present paper is to extend to these two structures of SE(3) the De Casteljau method. 0(3) ~3 b. "7. On the other hand. This corresponds to have curves with both negative and positive energy and gives as geodesics the socalled screw motions. the group structure of SE(3) is discarded in favor of a cartesian product of two distinct groups (rotations and translations). we treat here only the symmetric case as it avoids us the computation of covariant derivatives. The algorithm used here is taken from [4] and produces a C 2 curve connecting two poses in SE(3) with given boundary velocities. the obtained trajectories do not seem to be the o p t i m u m of any variational problem [4]. The tangent space at the identity I.
V b e ar (1) called respectively right and left invariant representations. rotations and translations form the Special Euclidean Group SE(3) whose matrix representation utilizes the homogeneous coordinates: f SE(3) = ~g E GI4(~). the kinematic equations can assume two useful forms: = V' g and [1 = g V b V ' .The De Casteljau Algorithm on SE(3) 25 Rigid b o d y m o t i o n in 3dimensional space. The relation between spatial and body velocity at g E SE(3) is expressed by the adjoint map Ada(y) = gyg1 V Y E ~e(3) . s. In the right invariant representation. The first order kinematic equations are then: Rv b (2) A d j o i n t m a p . the left invariant representation expresses invariance to change of the inertial frame and V b E ~e(3) is called body velocity. V = [ 0 ~ ] with Q E a o ( 3 ) a n d v E ] R 3} 0 1 s. the infinitesimal generator V" is called the spatial velocity 0] because it represents the velocity ofg translated to the identity and expressed in an inertial frame. deriving g E SE(3).t. g = The Lie algebra of SE(3) is ae(3)={V 6M4(R). Considering the rotation and translation components of the kinematic equations. In particular. Invariance here is with respect to a matrix multiplication from the right and means invariance to the choice of the body fixed frame. R E S O ( 3 ) and p E ~ 3 } The elements of ar have the physical interpretation of velocities with respect to a choice of frame.t. Similarly. the right invariant representation looks like: ~" p + v ~ Similarly.
Y]:VYYV i. the bilinear form defining the Lie algebra.): Ad :7o ] LY. For SO(3) and SE(3) the following formulae are used: * exp map in SO(3): Rodriguez' formula exPso(3) : ~0(3) + SO(3) sin IJ~fl1 .26 Claudio Altafini that gives the change of basis on the Lie algebra: V s = Adg(V b) = gVbg 1 V b = Adgi ( V ' ) The derivation of the adjoint m a p with respect to g gives the Lie bracket adv(Y)=[V. then we obtain a linear representation of the operators hdg(.) and adv(.e.cos Jf~fl~ + l[~vl[ . the adjoint m a p looks like bilinear. Its solution is explicitly described in terms of the exponential map. In the matrix representation of Lie groups and algebras. the kinematic equations (1) "looks like" a linear system of 1st order differential equations. For constant velocities V s and V b.e. If instead we extract a minimal representation of the elements of the Lie algebra (i.sin []~IIQ2 i[~4P . J E x p o n e n t i a l a n d l o g a r i t h m i c m a p s in S E ( 3 ) .coslf~Jl~2 9 exp map in SE(3) eXPsE(3) : ~e(3) + SE(3) where A(~) = I + 1 . we consider a vector of length equal to the dimension of the Lie algebra).
13. Its (normalized) constant infinitesimal generator is called t w i s t and corresponds to the axis of the rigid body rototranslation. The osculating plane is related to the notion of curvature of the manifold which is also chosen by selecting a metric tensor. ( R  RT ) where r s. In order to generate a trajectory in SE(3).2. Both exponential maps are onto but not onetoone. In fact. the exponential map being onto means that every two elements can be connected by a oneparameter curve called s c r e w (Chasles Theorem).~2 ~ ~(3) V= [ ~ A 03xl 211~ll2sin I1~11 E x p a n d log in S E ( 3 ) : t w i s t s a n d screws.IIwH(1 + cos HwH). which also have closed form expressions. 9 logarithmic map in SO(3) logso(a ) : SO(3) + s0(3) R ~+ ~2 s O . The following properties are known from the literature and are collected here from the original sources [1. c o s t = tr(R)I .e. to the orbits of (complete) constant vector fields and their left/right translations. we need to introduce some corresponding tools. 2.1 ~ 9 logarithmic map in SE(3) lOgsE(3) : SE(3) 03x I where A . have to be restricted to their principal values.I ( ~ ) I 1~ + 2sin Hull . ICl < ~ defined for tr(R) ~: . by the minimum energy curve associated to some nondegenerate quadratic form on the Lie algebra. In SE(3). In ~n.The De Casteljau Algorithm on SE(3) 27 The exponential map gives the oneparameter curves corresponding to constant generators in ~r i. the corresponding logarithmic maps.t. . For a Lie group. the equivalent of straight line is given by a geodesic arc i.e.17] for sake of convenience. defining a cubic polynomial curve requires the notion of straight line and of osculating plane which are fixed implicitly by the Euclidean structure.1 Metric s t r u c t u r e for S E ( 3 ) The construction of a polynomial from boundary data in the Euclidean case is essentially a geometric construction.
Y) = K (AdgX. negative definite and Adinvariant (or biinvariant): K(X. It is known that in SE(3) there is no Adinvariant Riemannian metric which implies that there is no natural way of transporting vector fields between points of SE(3) and therefore no natural concept of distance on SE(3). .. inducing a natural concept of distance on the group through the logarithmic function.i n v a r i a n t p s e u d o .. Y) ~+ K(X. double geodesic Approach 1: A d .28 Claudio Altafini On SO(3). Adinvariant pseudoRiemannian structure 2. by Cartan 2nd criterion. a metric tensor that is nondegenerate but not necessarily positive definite.) = a K(. AdgY) V g E SO(3) Therefore. the Metric properties of SO(3). . SO(3) is endowed with the structure of a Riemannian manifold. M e t r i c p r o p e r t i e s o f S E ( 3 ) . The Killing form in SE(3) is only semidefinite negative: if we rewrite r "1 r "1 X.Y E so(3)as X _ /x~/E [ Xv J ~ 6 and Y ~_ /Y~[ E ~6. Y) = t r ( a d x 9 ady) is symmetric.) on ~0(3) and propagating it to the whole tangent bundle TSO(3) by left or right invariant translation.e. The two most common approaches to tackle this obstruction are: 1. Killing form K : ~o(3) • ~o(3) + ]~ (X. For a biinvariant metric. Semidirect product means that SE(3) as a manifold can be considered the direct product SO(3) x ~3 but its group structure includes the action of SO(3) on ~3 by isometries.R i e m a n n i a n s t r u c t u r e . The Levi decomposition for SE(3) gives the following semidirect product: SE(3) = $O(3)  ~3 with SO(3) semisimple and ~3 abelian. then a s y m m e t LYv J ricbilinearf~ [ x ~ ] T W [ Y~] based only on the Killing form corresponds Lx~ j y~ to the semidefinite quadratic matrix . the geodesics correspond to oneparameter subgroups and to their left/right translations. It consists in insisting on the notion of oneparameter subgroups and on using the logarithmic map for a pseudoRiemannian metric structure i. by choosing as inner product (.
Y)Adinv = a tr(adz~ 9adv.)~3 is the standard inner product in Euclidean spaces. Consider for example the leftinvariant representation of the system. is neither left nor right invariant with respect to coordinate change.~ 9adv~ ) + fl(x~. Such geodesics are the screw motions mentioned above. Two different representations are obtained considering or discarding the group structure in SE(3). the right (or left) invariance of the metric is preserved. x.e.p) are: vb (3) . y~)~a + fl(y~. The second method relies on the socalled double geodesic approach i. we have that the leftinvariant equation for g = (R. P r o p o s i t i o n 1. . As the action of S 0 ( 3 ) on ~3 is by isometries. which says that the geodesic curves can give either positive or negative energy.) + Z(xv. Discarding the group structure. Proof. the two structures give the same leftinvariant metric. The expression for the inner product is (X. Y E ~e(3) is then: (X..~)~a where (. in forgetting about the group structure of SE(3) and considering separately the biinvariant metric of SO(3) and the Euclidean metric of/i~ a. A p p r o a c h 2: d o u b l e g e o d e s i c .. The former is leftinvariant while the latter. The inner product of two vectors X. corresponding to the double geodesic structure. y~)~3 Although such a representation is neither right nor left invariant. Y)dgeo = a tr(ad.The De Casteljau Algorithm on SE(3) Combining this with a Klein form 29 [o I] we get the most general Adinvariant quadratic form in SE(3) w=[~i 0j whose eigenvalues are nondegenerate but can be either positive or negative according to the values of a and/3. The corresponding quadratic form was first proposed by Brockett [1].
Applying the same change of inertial frame in the group structure.) : [0. it was shown in [4] how to generate an interpolating curve in an analytic way through a combination of exponentials. Likewise.(RoP~b. R o p + p0) .30 Claudio Altafini Changing inertial frame from g to gog.(R0. a couple of iterated combinations of the geodesics gives a C ~ curve which shares the same boundary conditions with the original patching of geodesic arcs. The basic idea in ~ n is to transform the boundary conditions on the velocity into intermediate points (called "control points"). ~ t=0 Here. The generalization consists in substituting the line segments used for the construction in the Euclidean version with geodesic arcs.1 as final time is meant to simplify the calculations. d (RoR. A sketch of the (leftinvariant version of the) algorithm is as follows: * transform the 1st order boundary values in infinitesimal generators Vo1 and V~ (see [4] for the details) d7 I tO = 39oVo1 and ~. The method generalizes to Lie groups a classical construction for Euclidean spaces called the De Casteljau algorithm [6]. [] 3 Algorithm for a closedform C 2 trajectory in SE(3) The idea is to find a closed form curve g(. left invariance gives d (gog) gogVb or eq. RoVb) we reobtain (3). 1] ~+ SE(3) satisfying the following boundary conditions: "d0) = go. For this case. Deriving the two components. the time axis can be rescaled appropriately afterwards. choosing tf . The only difference between (3) and (2) is an arbitrary rotation in the ~3 part which does not modify lengths.p0) we get gog = (RoR.t = 1 : 3g! Vg d7 * get the "control points" gl : goe v~ g~ : g:eV~ . If tf r 1. where go . Rop + Po). The combination of the straight line segments connecting the extreme points to the control points gives the desired polynomial. (2) in components.
1] eV~(t) = e(1t)V~etV~ 9 construct and Vo3(t) s. In the pseudoRiemannian case. e v~ = e(at)V~etVl' t 9 [0. find the velocity V~ s. g~ = gleVl ' The three velocities obtained so far V01. V~(t). p~) E SO(3) • ~3. eV~(t) = eOt)v2(t)etV~(t) The velocities Vo2(t). VoS(t) are not constant but correspond to "polynomial" generators 9 the interpolating C 2 curve is 7(t) = goetV~(t)etVo~(t)etVo~ As in 5"0(3) and SE(3) the exponential and logarithmic maps have closed form expressions. provided that the difference of the . V11 and V~ correspond to twists in he(3) and are obtained through the maps eXPsE(3) and logsg(3 ).t.The De Casteljau Algorithm on SE(3) 31 i. the velocities V01. and as long as the data are given in a symmetric fashion (i. V~ 9 he(3) are constant. we need to split all the data: SE(3) 3 9~ ~ (R~. In the doublegeodesic case.e. the procedure above requires only linear algebra tools plus exp and log maps. 4 Simulation results The two cases above for SE(3) lead to different curves because the geodesics are different.e. use the metrics in SO(3) and ~3 to construct the various ~j e n0(3) and v j e ~3 and then recombine them in homogeneous representations The curves one obtain in the two cases are different. 9 construct Vg(t) and V~(t) s. we do not have to compute covariant derivatives). the points reached by the timeone oneparameter arcs from the extremes. Their combination (through the exponential) gives rise to curves which not anymore correspond to oneparameter subgroups. In particular. it is possible to maintain the idea of straight like independent of the amount of the rotation. instead. V11.t. in the double geodesic case. 9 using tim logarithmic m a p . but keep the same boundary values as the C o patch of geodesic arcs.t.
see Fig.~ .. 2. 5 . the more consistent is the r o t a t i o n c o m p o n e n t of the desired m o t i o n (with respect to the t r a n s l a t i o n a l p a r t ) the more the two curves will look different (compare Fig. 1).. 0 3i ~ 2. V I 1 I/ o. 2 and Fig. T h e price to p a y is t h a t the curve is not leftinvariant with respect to a c o o r d i n a t e t r a n s f o r m a tion since we "forget" a b o u t a r o t a t i o n in the Euclidean p a r t (Prop.32 Claudio Altafmi two end positions P0 and p/ is aligned with the direction of b o t h b o u n d a r y t a n g e n t vectors of the t r a n s l a t i o n a l p a r t .5 2 I.51 0. . 1 (a). The two curve generated by the De Casteljau algorithm in the special situation in which the translational components of the boundary data lie on the same direction of the translational boundary velocities. 2.V (a) Double geodesic curve (b) Adinvariant curve F i g . The two curves generated by the De Casteljau algorithm in a generic case. 1. 3). 0 0 2 ~ 2..5 0. 0 0.5 1 2.5 2 1 2 (a) Double geodesic curve (b) Adinvariant curve Fig. s. 1 1 0. In general. 0.5 I/ o...
J/ 2 (a) Double geodesic curve (b) Adinvariant curve Fig. In J. Silva Leite and P. Proc.S. . vol. FL. 3.J.The De Casteljau Algorithm on SE(3) 33 1 25 2 1. 35th Conf. Beijing. Murray (1995) Proportional Derivative (PD) Control on the Euclidean Group. Robotics Proceedings of Symposia in Applied Mathematics. Crouch (1996) Second order optimality conditions for an higher order variational problem on a Riemannian manifold. Providence. 177202.W. Koo and S. Ge and B.E.E.. 533538... T. Krishnaprasad (1995) Motion control of driftfree leftinvariant systems on Lie groups. American Mathematical Society. Tampa. IEEE Trans. F. F. The two curves generated by the De Casteljau algorithm in another generic ease. Proc. Journal of Dynamical and Control Systems 1. on Automatic Control. Silva Leite and G. 5. China. Q. G.E. Lie groups and symmetric spaces. Leonard and P. Camarinha. Crouch and F. 10911097. ASME Journal of Mechanical Design. Kun (1999) Geometric Splines.J. Silva Leite (1995) The dynamic interpolation problem on Riemannian manifolds. Bullo and R. 7. Proc. N. Proc. 40. P. R. on Decision and Control. 3.5 ~'. 14th IFAC World Congress. Brockett (1990) Some mathematical aspects of Robotics. Crouch. M. of the 3rd European Control conference. 2. Kobe. F. 4th ed. 6. Academic Press. 4.. References 1. 9. P. Balllieul ed. Sastry (1998) Output tracking control design of an helicopter based on approximate linearization. 16361641. of the 37th Conference on Decision and Control. 8. Ravani (1994) Geometric construction of B4zier motions. 116. Japan. 15391554. Italy. Rome.E. Farin (1997) Curves and surfaces for computeraided geometric design: a practical guide. 749755.D.
Park and R. L. IMA Journal of Mathematics.C. Z. on Robotics and Automation. Brockett (1994) Kinematic desterity of robotic mechanisms. V. Croke (1998) On the generation of smooth threedimensional rigid body motions. NY. Int. Proc. Zefran and V. Kumar (1998) Two methods for interpolating rigid body motions.C. Sastry (1994) A Mathematical Introduction to Robotic Manipulation. Selig (1996) Geometrical methods in Robotics. Belgium.34 Claudio Altafini 10. Li and S. Leuven.C. of the 1998 IEEE Int. 16. Conf. 576589. 13. M. Control and Information. Park and B. 12. Park (1995) Distance metrics on the rigid body motions with applications to mechanisms design. G. 3640. CRC Press. Springer. Paden (1989) Cubic splines on curved spaces.M. J. Kumar and C. F. Murray. 15. ASME Journal of Mechanical Design 117. 6. . on Robotics and Automation 12. of Robotics Research. F. 12. F. ASME Journal of Mechanical Design 117. 14. Ravani (1995) B~zier curves on Riemannian manifolds and Lie groups with kinematic applications. W. IEEE Trans. 4854. Zefran. New York. Noakes. Heinzinger and B. 465473. M. 115. J. M. 11. R. 17.B.
nl A b s t r a c t . n o n . A design methodology is presented for tracking control of a class of secondorder nonholonomic systems. nl. Box 513. The control design methodology is applied to an underactuated planar manipulator with two translational and one rotational joint (PPR). j ager@wf w. rue. In the second step. this only holds locally a n d it r e m a i n s a challenging p r o b l e m to develop control laws t h a t achieve s e m i .s m o o t h t i m e . the second subsystem is exponentially stabilized by applying a backstepping procedure. h. In the third and final step it is shown that the closedloop tracking dynamics of the extended chainedform system are globally exponentially stable under a persistence of excitation condition on the reference trajectory. de.Trajectory Tracking by Cascaded Backstepping Control for a Secondorder Nonholonomic Mechanical System Nnaedozie P. Aneke. ni j me ij er@tue.e r r o r d y n a m i c s are locally s m o o t h stabilizable. N o n h o l o n o m i c systems have a s t r u c t u r a l o b s t r u c t i o n to the existence of s m o o t h t i m e . The simulation results show a good performance of the tracking controller. and A b r a h a m G. This extended chainedform system is in cascade form and we apply a linear feedback to the first subsystem. wtb. Henk N i j m e i j e r . It is known t h a t the t i m e .g. T h e studies were p r i m a r i l y l i m i t e d to the s t a b i l i z a t i o n of firstorder n o n h o l o n o m i c s y s t e m s satisfying n o n .i n v a r i a n t stabilizing feedback laws a n d the feedback s t a b i l i z a t i o n p r o b l e m can only be solved by s m o o t h t i m e . .v a r y i n g feedback.i n t e g r a b l e acceleration constraints [3]. In the first step we transform the system into an extended chainedform system. However.I.O.v a r y i n g J a c o b i a n linearization a r o u n d a t r a j e c t o r y can be c o n t r o l l a b l e and therefore the t r a c k i n g . The method consists of three steps. 5600 MB Eindhoven. rue. Systems and Control Group P. T h e t r a c k i n g control problem for secondorder n o n h o l o n o m i c systems has h a r d l y received any attention.g l o b a l or even global t r a j e c t o r y tracking.v a r y i n g homogeneous feedback or discontinuous feedback [2]. 1 Introduction In the p a s t few years there has been an increasing interest in the control o f n o n h o l o n o m i c control systems [1]. Fax : +31 40 246 1418 edo@wfw. The Netherlands Phone: +31 40 247 2611 / 2784. de J a g e r Eindhoven University of Technology Faculty of Mechanical Engineering. nl. wtb. a. This class of s y s t e m s arises from control p r o b l e m s r e l a t e d to mechanical systems with n o n .i n t e g r a b l e k i n e m a t i c or velocity constraints. In this c o n t r i b u t i o n we consider t r a c k i n g control of secondorder nonholon o m i c systems.
In the simulations. 2 Problem Formulation Consider a 31ink underactuated planar manipulator where the first two translational joints are actuated and the third revolute joint is not actuated. 5 we draw conclusions and give recommendations for further research. 1 shows the P P R planar manipulator. In Sect. . The tracking control problem for such underactuated manipulators has received little attention. In Sect. In the third and final step. where the first subsystem is perturbed by the state of the second subsystem [4]. namely underactuated robot manipulators. we make conclusions on the exponential stability of the closedloop tracking dynamics of the extended chainedform system by using a result for cascade systems from [7]. we transform the dynamic equations of motion of the mechanical system into an extended chainedform system with secondorder derivatives (see [6]).P. the second subsystem is exponentially stabilized by applying a backstepping procedure. Oriolo and Nakamura [10] showed that planar underactuated manipulators have a structural obstruction to the existence of smooth timeinvariant stabilizing feedback laws. This system is a special case of an underactuated manipulator with secondorder nonholonomic constraints. Aneke et al. 2 we formulate the trajectory tracking problem for secondorder nonholonomic mechanical systems. In Sect. However. 3 we present the control design methodology. we focus on a subclass of the secondorder nonholonomic control systems. This extended chainedform system is in cascade form and we apply a linear stabilizing feedback to the first subsystem. finally. 4 we perform a simulation study with a 31ink planar underactuated manipulator and. We try to solve the trajectory tracking problem by using a combined cascade and backstepping approach. they do not meet Brockett's wellknown necessary condition for smooth feedback stabilization [2]. The combined cascade and backstepping approach consists of three steps. In the first step. In the second step. The inclusion of the dynamics is mandatory and results in the presence of a drift term in the statespace model which makes the control more difficult [11]. in Sect. underactuated manipulators are a challenging example of underactuated systems. Backstepping is a wellknown control design technique for nonlinear systems and the approach followed in this contribution is similar to that in [5]. We investigate the performance of the derived tracking controllers by simulation of a 31ink underactuated planar manipulator where the first two translational joints are actuated and the third revolute joint is not actuated Is].I. Cascaded nonlinear systems are nonlinear systems that can be written as two subsystems. These systems are characterized by the fact that there are more degrees of freedom than actuators. Fig.36 N.
2. Nevertheless. As we mentioned earlier. The dynamic equations of motion can be written as r~. In the case of underactuated systems in the vertical plane. 0]. The . The mass and inertia of the links are denoted by rni and Ii respectively and l denotes the distance between the center of mass of the third link and the third joint. ry. see [8.1 Equations of Motion Define q = [r~.real sin(0)/zx + rn3/cos(0)/: u : 0.y + rn31 cos(0)0 .Tracking by Cascaded Backstepping Control 37 Y l i ! ry rx" I I X Fig. ry) denotes the displacement of the third joint from the origin (x. an inverted pendulum and the Acrobot [12].g.m31 sin(0)02 . A planar PPR manipulator It should be noted that the linear approximation of this system is not controllable since the dynamics are not influenced by gravity.i. We also introduced the parameters m~ : (ml + rn2 + m3). . Also a particular feedback controller along this trajectory was presented [15].31 cos(0)0 ~ = ~1 m. r~ denote the inputs to the prismatic actuated joints.~.m3t sin(~)0 . the linear approximation around an equilibrium state is controllable due to gravity [13]. in [14] the 3link underactuated planar manipulator was shown to be controllable by constructing an input to generate a trajectory from any given initial state to any given final state in finite time. this system was shown to be controllable by constructing an input to generate a trajectory from any given initial state to any given final state in finite time [14]. e. Let rl. i.. where the configuration variables lie in the to be defined configurationspace. y) = 0 and 0 denotes the orientation of the third link with respect to the positive xaxis.16]. The trajectories consist of simple translations and rotations together with combined translational and rotational movements.r2 (1) I0.e q E C. 1. see Fig. where (r~. 1. m u : rn2 + rn3 and I : / 3 + real 2.
P.I.~2~'1.1)] tan(g) ry + A sin 0 9 J (3) r2 I m31cosOO2 + ( m . + (IIssinOcosO) vy ] I m31sinO'2 + (~sinOcosO) vr + (rnu . We therefore restrict the configurationspace of the configuration coordinates q to the subspace C given by C = {q = (r~. and v~ are given by vu Lsin0 . .~sin2 O) v.38 N. see [6]. k E N. the state q(t) corresponding to a solution ~(t) of (5) satisfies the nonholonomic constraint. .~cos~ O) vu 1 where v. this coordinate transformation is not welldefined. In the next section we will show that the system can be transformed into a more simple form.c o s 0 [A(u2cos20_202tanO) ] " ul (~/2. Aneke et al. T h e local coordinate transformation and input transformation are then given by [.~/2)} (4) When 0 = ~/2+k~.r~. 2. Therefore. (5) The nonholonomic constraint (2) is transformed into the constraint 43 .i:~ (t)sin O(t) + ~(t)cos O(t) = O.1] ~2 ~3 "rr + A ( c o s O .0) e R ~ I 0 ~ T h e dynamics of the underactuated m a n i p u l a t o r are transformed into the extended chainedform system ~'i = Ul ~'~= ~= u~ ~2ul. (2) where A = I/(rnzl) equals the effective p e n d u l u m length when the third link is treated as a rigidbody p e n d u l u m suspended under the passive joint. nonholonomic constraint of the system (1) can be written as AO(t) . Let the transformed state ~ = 0 correspond to the equilibrium q = 0.2 Transformation into Extended ChainedForm In [8] a coordinate and feedback transformation was proposed to transform the system (1) into an extended chained form.
~ d . ( S t a t e .. We therefore consider a reference trajectory qd E C.8) is globally uniformly asymptotically stable. We assume that the reference trajectory qd(t) E C Vt. . In the following sections we will investigate whether the error dynamics can be asymptotically or even exponentially stabilized.X. ~ld). Suppose that we want the third link to describe a predefined motion in time. i. The vector Ud contains Ud = [Uld. i.x.f e e d b a c k t r a c k i n g c o n t r o l p r o b l e m ) The tracking control problem is said to be solvable if we can design appropriate Lipschitz continuous timevarying statefeedback controllers of the form Ul = ul(~.Tracking by Cascaded Backstepping Control 2. i. the tracking dynamics of the original mechanical system (1) are not globally asymptotically stable.~31 ~ X32 (7) x3~ = x 2 1 u l d + ~ 2 1 ( u l . but are only asymptotically stable when the coordinatetransform (3) is welldefined.for the system (1). ~d). After transformation of this reference trajectory qd we obtain a reference trajectory ~d for the extended chainedform system (5) satisfying a system of equations given by = = = = (8) A s s u m p t i o n 1.. U(d3).~d in statespace form are given by Xll = X12 X 1 2 = U l . On the other hand. see (4).. .e.U2d . q) of the system to follow a prescribed path (qd. . .~d) (8) such that the closedloop system (7.Uld). is continuously differentiable. .3 The StateFeedback Tracking Control Problem 39 Consider the extended chainedform system (5). The dynamics of the trackingerror x = ~ . Problem 1. To that end we apply some of the results presented in [17. U2d] and its higher order derivatives.Uld X21 3g22 X22 = U 2 . Remark 1.e qd E C ~. i. U2d) in (6) are continuous.18] together with some theory for linear timevarying systems from [19]. Therefore the reference trajectory qd is continuously differentiable and the corresponding inputs (Uld. we want the state (q. Vt > O. u~ = u2(t.e when qd(t) E C. We say that the tracking control problem has been solved when the tracking dynamics of the extended chainedform system (5) are GUAS.)..e ~d = (Ud.
18] for cascaded systems. z l. (12) (iii) IIVll ~ Ilzxll ~ ~ v ( t .40 N. We state the following result from [4. Xlx.zl)<_o. 3 Cascaded Backstepping Control In this section we apply a cascade design to stabilize the error dynamics (7).I. A 1 { . v Ilzl[I ~ 7. z2) respectively. The total system (10) is a system L'l with state zl that is perturbed by the state z2 of the system Sz. In order to make conclusions on the exponential stability of the complete closedloop system we will use a result from [17. zl) continuously differentiable in (t. (11) and the perturbation term is given by g(t.P. T h e o r e m 1. zl. g(t. z2 E R m.uld. z l ) . z'2 = f2(t.ule) // ~ r . ~1) + g(t. zl) : R + • R n + R and a positive definite proper function W ( z l ) such that 6) v(t.x12) and therefore ul . which was based on a result in [4]. zl) is GUAS and there exists a continuously differentiable function V(t. We start by rewriting the tracking dynamics into a more convenient form. zt) 2~2 : z ~ = f~(t.~1 = f l ( t . 1 . Aneke et al. A2) is stabilized for ul = ula. Consider the timevarying nonlinear system given by ~1 = fx(t. zl) >__ W(zl) OV (it) OV ff+ff~fl(t.X32 x3~ = x~lul~ + ~21(ul . z2)."T12 k X12 U 1 . z2)z~. z~). z2)z2. z2). v Ilzlll ___~. fl(t. Then we design the remaining input u2 such that the remaining subsystem (A1.Uld A f k21 vt~22 = x~ u2u~d (9) Suppose that the / / subsystem is stabilized by the input ul(uld. For a survey on the stability of cascaded systems see also chapter 5 of [9]. where ~ > 0 and T > 0 are constants. (lO) where zl E R n. zl) and f2(t. zl. where ~wl : . The cascaded system (10) is globally uniform asymptotically stable (GUAS) if the following three assumptions hold: (1) ~wx : subsystem: The subsystem zl = fl(t. ~ 3 1 .18]. z2) continuous in their arguments and locally Lipschitz in z2 and (zl.
U. (16) where x21 denotes a virtual input for the backstepping procedure. Consider the timevarying subsystem A 1. R e m a r k e. zl. A stabilizing function x32 : a1(x31) for the x31subsystem is 0tl ~ . feedback u~(t. z~) satisfies 41 Ilg(t. L e m m a 1. in order to satisfy condition (2) in Theorem 1 the resulting perturbation matrix g(t. 2 d i + 2 _ ~. z~. zu)ll < ~1(I]z211) + ~2(llz~ll)llz111. V t > to. We aim at designing a stabilizing (virtual) feedback x21 for the subsystem zill. z2) is G U A S and satisfies ft?llz2(to. This is the case when choosing the feedback ul as ul = u l d + k ( t . In the system (9) the perturbation term g(t. ~ : R + + R + are continuous functions. l U l d A .1 and 2Y2 are globally Kexponentially stable.]Uld t~ldX31 .)~. x l l .k~x12.1 Stabilization of the Subsystem Az Suppose that t h e / / s u b s y s t e m ul = Uld. zl. x12). z2)z~ is given by ~2~(u~ . (13) where ~1. _ L.Uld) and depends on the.t.. z~) has to be linear with respect to the variable z2 = (x11. zl. xl~). (3) L'~ : subsystem: The subsystem i2 = fu(t. Consider the first equation x31 : x32 and assume that x32 is the virtual input. x 1 2 ) . Therefore. (see [4]) If in addition to the assumptions in Theorem 1 both . (15) where the polynomial p(A) = A2 + klA + ks is Hurwitz. (14) where the function ~(.. then the cascaded system (10) is globally Eexponentially stable. when c o n s i d e r i n g / / as the unperturbed system $2. to be designed.k l x l l . has been stabilized by choosing kl > 0. C1 > 0 . x). < Vt > to. with k : R+ • R 2 > R a linear function in (x11.z2(to))lldt ff(llz2(t0)l[).31.( .a l ( x a l ) subsystem = xau + c1(2dl C1It 2 d a + 2 ld x31 and consider the xa~ 9 2da+2X32 ~~32 ~ X211Lld "11Clltld + . 3. ks > 0. Define x3~ = xa2 .Tracking by Cascaded Backstepping Control (2) interconnection: The function g(t. For ul = ula this subsystem can be written as the timevarying linear system x31 = x3~ ~32 ~ X211/id.) is a class IC function. 2d1+1.
Aneke et al. (18) We show that the resulting closedloop system (18) is globally exponentially stable (GES) under the following assumption.e.2 c u ~d2+1~ ld ) X32 where c~ > 0 and we substituted x3~ • x3z The closedloop system becomes 9 ~ .~ 3 2 32 = ~ . The system (18) can be written as (10).e the Alsubsystem for ul = Uld. x32) for the $3~subsystem is given by E~2 = .It t+6 2r+2 ula ( r ) d r < e~. 2d142= ~31 ". given by (17). x3~). A2) In Sect.1 we saw that the Alsubsystem can be exponentially stabilized by the virtual input x21 = a~(t. and by applying Theorem 1 and L e m m a 1 globally exponentially stability follows.(C 1 C2 U2dl+2d~+3 ' ~ c 1 ( 2 d l k ld . 3. The system (16). Uld(t). The proof will be given in a forthcoming paper.r > 0 such that for all integers r > 0 ~1 _< i. we can prove the existence of a suitable Lyapunov function. The equilibrium x = 0 of (18) is globally exponentially sta ble (GES) for all reference inputs Uld(t) satisfying the persistence of excitation condition (19). 2d~t2= 3 2 v2ttld ~ + Clu~dI"I'2~31. xal.I.42 N. By Theorem 3. Suppose that x~l is the virtual input. The function Uld(t) is persistently exciting. Vt > O. 3. which we state without proof.kVlUld 71_ 2 ) U l d d l Uld) 2 9 X31 (17) (_ . The virtual input x21 is a state of the system A2 and we now continue the backstepping procedure for the A2subsystem as follows. Define x31 = x31. i. in closedloop with the linear timevarying feedback (17) is GES. A s s u m p t i o n 2. Denote the state x of the A1subsystem by x = [x31. X) and consider .P. Uld(t).. P r o p o s i t i o n 1. The result is given in Proposition 1. (19) Under Assumption 2 the closedloop system (18) is globally exponentially stable. i.t'lttld x 3 1 71. x31. EI. there exist J. then a stabilizing function x21 : a~(fild.a~(t. Define x~l : x2t .2 Stabilization of the Subsystem (•1. x32]. ~d~+l 71.12 in [9].
A2) subsystem. x) for the Atsubsystemsuch that VI(t. x)]. A2) and / / subsystems separately.c3"~21. x~2.311 11   (e3 gd. P r o p o s i t i o n 2. If the signal ~ l ( t ) . x) and consider the ~:22subsystem ~22 = (u2 .4"~22 . x) and cr3(uid(t).s + l . showing that we have successfully stabilized the (A1. I / l d ( t ) . utd(t). we have exponentially stabilized the (A1.r.A2)subsystem when ul = utd and the / / subsystem.2d4 + 1 tz4a.2d4k2~1. Z')] . r. r) = ~ d [Ot2(t . r ) for the $2~subsystem is given by Ot3(t . After substitution of or3 and ~21 respectively. the reference input uaa and its derivative htd.Its timederivative satisfies r) < . In (9) we have stabilized the (Ai. then the closedloop system is globally Kexponentially stable.3 Stability of t h e C o m p l e t e Tracking Dynamics In this section we show that the complete tracking dynamics are globally exponentially stable. U l d ( t ) . This subsystem can be stabilized by choosing the input u2 such that 1t'2 . The proof will be given in a forthcoming paper._ . A~)subsystem. Suppose that the reference trajectory utd(t) satisfies Assumption 2. are uniformly bounded in t.12 in [9] there exists a LyapunovfunctionI/i (t.u2a) . where z22 is the virtual input.+t =  d[ ] + d~ TJ uld(t). This input u~ . d (21) Accordingto Theorem3.d / d t [~2(t. cf. (22) where a z ( u ~ ( t ) . Uld(').+2 _It t. k2 > O.u~d stabilizes the (At. (6). x)]. uld(t). x) + (~1 + x2~2)/2. x) = V1 (t. Now consider the Lyapunov function given by V(x~t..~ [~3(t. x) are given by (17) and (20) respectively. Consider the system (9) in closedloop with the controller u2 given by (22) and ul given by ul = ula  kxxlx   k2x12.~ 2 d = _ z. We can now use Theorem 1 to investigate the stability properties of the complete system. We then concluded by Theorem 1 and Lemma 1 that the combined system is/Cexponentially stable when the reference input uld and its derivative tila are uniformly bounded over t. .Tracking by Cascaded Backstepping Control 43 the ~21 subsystem ~ I = z u 2 .t~3(t.2d (20) We define a new state as x22 = a:22 . The result is stated in the following proposition. A stabilizing function a3(t. Summarizing. uld(t). p(s) = s 2 + k~s + kl is Hurwitz. X)]. Utd(t). 3. (23) with kt > O. we can write (21) as _ =2~. x) _< /331]xlI2.22 ~.d / d t [c~3(t.
r~) = 0 and select ~ld(0) = 0. ~3d(O).2 s t = 15. 0 .5 t = 24. Consider the extended chainedform system (6) subject to the sinusoidal inputs Uld = .A (cos(arctan(r~ cos(at))) . (~. we conclude by Proposition 2 that the trackingerror d y n a m i c s (7) are uniformly exponentially stable. The controller . ua).0 .2  24.4.I.rla~. ~2d(t) = r~ cos(at) r l r ~ sin(2at). a r l r~/4] where ~ld(O) and (ad(O) are still to be designed parameters.19. 3 (At = 0. ~3. (25) We consider a trajectory around the origin (r~.~la(O) + rl sin(at).) and U2d = r2a ~ cos(at). We evaluate the controller (22. O. W i t h initial condition (a(O) : [~x. Ga(0) = 0 and rl = r2 = a = 1 in the reference trajectory (25).4s t =9.5 .0 ==4~ Fig.0s t .29.2 s t 19.30. ~3] given by ~d(O) : [~la(O). the closedform solution ~u E X of the reference trajectory is given by ~lu(t) . r~. Therefore. t4. Motion of the third joint and link in simulation of Fig. (24) The resulting reference trajectory qa for the mechanical system (1) is given by rxd(t) : ~ld(0) q. rl a. 2. Aneke et al.415.25 s) 4 Simulation Study In this section we apply the previous control design m e t h o d o l o g y to the 3link planar underactuated manipulator. Oa(t) = arctan(r2 cos(at)).23) by tracking a smooth trajectory (~a.rl sin(at) . ~1. where the reference input ula and its derivative flla are uniformly bounded in t and the persistence of excitation condition (19) is satisfied. rlr2 r~d(t) = ~3d(0) + ifsin(2at).0.1).A sin(arctan(r2 cos(at))).44 N.sin(at.P. ~2.
0 is shown in Fig. The parameters cl. The perturbation term in (9) depends on ut . e~ = 2. states q errors q . This control design method consisted of a combination of a cascade approach and a backstepping approach... 3.. ::. k2 = 2x/~. c a = 4 ...... ..... It follows that the closedloop tracking dynamics of the extended chainedform . 0 o:t: . c2 determine the convergence of the chained state ~a and its corresponding mechanical state ry. There still remains some freedom of choice in the control parameters. 0 (dashdotted). We choose the control parameters as k 1 = 4 . state qd I . e 1 = 2 . and when considering more complex trajectories this freedom of choice can be exploited to improve the transient response of the underactuated manipulator....qd ref..u l d .. r2 (dashed) The result of tracking the trajectory with initial condition q(0) . The state variables of the mechanical system are shown in Fig. .. .. rx (solid). 5 Conclusions We have presented a control design method for exponential tracking of a secondorder nonholonomic system.. in order obtain fast convergence to the reference trajectory ul must converge to the reference value uld fast. From the simulation it is clear that it takes about 5 seconds before the trajectory is successfully tracked... input ra 5 10 11 ' 0 5 10 0 5 10 0 5 10 Fig. k~. c4 mainly determine the convergence of the chained states ~ and its corresponding mechanical state 0. 3.. 2.Tracking by Cascaded Backstepping Control 45 ul is given by (23).rd 10 0 ref. Tracking of the trajectory (25) by the planar PPR manipulator.. Therefore...... . The parameters c3. c 4 = 4 . inputs rl (solid). 2 0 input r 5 10 0 5 errors r . . coordinates of system (1). which motivates our choice for kl. We choose dl = d2 = 0 in the virtual input x2t (17) and da = d4 = 0 in the controller us (22)...... . r~ (dashed)..
P. Japan. Reyhanoglu M.. Nijmeijer H.. Kolmanovsky I. In: 30th Conference on Decision and Control. December. Oriolo G. (1996) Nonholonomic Control of a 3 Link Planar Manipulator with a Free Joint. IEEE Control Systems Magazine 15. England..H. in practice. 23982403 12. (1996) Nonlinear Systems. Prentice Hall. 3. Yoshikawa T. December. Oriolo G. (1983) Asymptotic Stability and Feedback Stabilization. 131138 8. Shiroma N. 2036 2. van der Schaft A. However.. 2nd edn. Japan..J. (1995) Developments in Nonholonomic Control Problems. there will be d i s t u r b a n c e s a n d m e c h a n i c a l friction (in t h e free r o t a t i o n a l link) t h a t can c o n s i d e r a b l y d e t e r i o r a t e the p e r f o r m a n c e . (1996) Nonlinear Control of a Class of Underactuated Systems... 16631671 4. 681695 . Intelligence for Mechanical Systems. 16821687 7. (1998) Exponential Tracking Control of a Mobile Car using a Cascade Approach. IEEE Control Systems Magazine 15. (1995) The Swing Up Control Problem for the Acrobot. T h e s i m u l a t i o n s show a g o o d p e r f o r m a n c e of the d e v e l o p e d controller. McClamroch N. Lefeber E. References 1. Khalil H.H. 14351436 9. system are globally e x p o n e n t i a l l y s t a b l e under a persistence of e x c i t a t i o n condition on the reference t r a j e c t o r y .W. Kobayashi K. van der Schaft A. F u r t h e r m o r e .. J. November..K. Pantely E. S. Brockett R. 4955 13. Arai H. 2. 3. Sussmann H..H. In: I E E E / R S J International Workshop on Intelligent Robots and Systems '91. Panteley E. Kolmanovsky I. Systems and Control Letters 33. France.J. 221226 5. IEEE Transactions on Robotics and Automation 14.. On Global Uniform Asymptotic Stability of Nonlinear TimeVarying Systems in Cascade. New York 10. In: IFAC Workshop on Motion Control. December. and Lor/a A.)) 181191 3. Japan. 265279 6. Tanie K... Reyhanoglu M. Imura J. McClamroch N. Spong M. IEEE Transactions on Automatic Control 44. Lor/a A.W. (Eds. W. (1991) FreeJoint Manipulators: Motion Control under SecondOrder Nonholonomic Constraints.. Aneke et al. In: 35th IEEE Conference on Decision and Control. the t r a c k i n g d y n a m i c s of the mechanical s y s t e m are only e x p o n e n t i a l l y s t a b l e on a s u b s p a c e o f t h e statespace where the c o o r d i n a t e t r a n s f o r m a t i o n is welldefined. September. In: 35th IEEE Conference on Decision and Control. (1998) Nonholonomic Control of a ThreeDOF Planar Underactuated Manipulator.. (1991) Control of Mechanical Systems with SecondOrder Nonholonomic Constraints: Underactuated Manipulators..P. 2.I. McClamroch N. IEEE Transactions on Automatic Control 44..46 N. Differential Geometric Control Theory (Brockett R. O u r goal for future research is to e x t e n d t h e results to e x p e r i m e n t s w i t h an und e r a c t u a t e d R R R . Milman R.r o b o t [20]. Nijmeijer H... Nakamura Y. Jiang Z. (1999) Dynamics and Control of a Class of Underactuated Mechanical Systems. 12481253 11. Nakamura Y. (1999) A Recursive Technique for Tracking Control of Nonholonomic Systems in Chained Form.
In: IEEE International Conference on Control Applications. van Beek B.668673 . August. (1999) Output Feedback Tracking of Nonholonomic Systems in Chained Form. 183199 18.J. PrenticeHall. Shiroma N. van der Schaft A. (1999) Linear Controllers for Tracking ChainedForm Systems. (Ed.. Robertsson A. (1995) Dynamic Mobility of Redundant Robots under EndEffector Commands.. (1999) An Experimental Facility for Nonlinear Robot Control...Tracking by Cascaded Backstepping Control 47 14. de Jager B. April..)). In: Nonlinear Control and Robotics Preprints. In: IEEE International Conference on Robotics and Automation 1. In: IEEE International Conference on Robotics and Automation. 4.. Stability and Stabilization of Nonlinear Systems (Aeyels D. Hawai'i. Nijmeijer H. Mattone R. Oriolo G.r o m . (1993) Linear System Theory (Kailath T.J. 246. Nijmeijer H. (1997) Feedback Control of a 3DOF Planar Underactuated Manipulator. In: 5th European Control Conference. 1st edn. Arai H. Universits di Roma "La Sapienza'. 37073713 15. Information and System Sciences Series. Lecture Notes in Control and Information Sciences. Robertsson A. USA.. Lefeber E. c d . (1996) Controllability of a 3DOF Manipulator with a Passive Joint under a Nonholonomic Constraint. September..703709 16. New Jersey 20. Italy 17. Arai H.. Lefeber E. (Eds. Germany. Paper 772 (AP4) 19. 1... LamnabhiLagarrigue F. Springer. De Luca A. Tanie K.)). Heidelberg. Rugh W.
For nonlinear systems a sufficient condition is established and a (partial) converse is also discussed. T h e interest of such controllers rests u p o n the fact t h a t . i. The nonlinear formulation is used to derive a simple characterization of stabilizing static output feedback control laws for linear systems in terms of the intersection of two convex sets and a (generically) nonconvex set. 1 Introduction T h e static o u t p u t feedback (SOF) s t a b i l i z a t i o n p r o b l e m is p r o b a b l y one o f the most known puzzle in s y s t e m s and control. s t a t i c o u t p u t feedback controllers can be employed as backup controllers. of Electrical and Electronic Engineering Imperial College Exhibition Road London SW7 2BT.15]. b u t which are used in case of faults. it is possible to show t h a t in m a n y cases the design of a d y n a m i c a l o u t p u t feedback controller boils down to the solution of a s t a t i c o u t p u t feedback control p r o b l e m [17. ast olf i@ic. T h e s i m p l e s t a t e m e n t o f the p r o b l e m is as follows: find a s t a t i c o u t p u t feedback control such t h a t the closedloop s y s t e m is a s y m p t o t i c a l l y stable. Finally.u p controllers have to be as simple as possible. In p a r t i c u l a r . since s t a t i c o u t p u t feedback controllers are less expensive to be i m p l e m e n t e d a n d m o r e reliable in practice. Italy 2 Dept.Static O u t p u t Feedback Stabilization: from Linear to N o n l i n e a r and B a c k Alessandro Astolfi 1.v a r y i n g n a t u r e of the o u t p u t gain is chosen in the narrower set o f piecewise c o n s t a n t functions or timeperiodic functions. T h i s p r o b l e m is i m p o r t a n t in its own right. A novel necessary and sufficient condition for linear systems is proposed. as controllers which are not active d u r i n g the regular o p e r a t i o n of the plant. England a. uk A b s t r a c t . For this reason b a c k . di Elettronica e Informazione Politecnico di Milano Piazza Leonardo da Vinci 32 20133 Milano. s t a t i c feedback controllers include the f a m i l y of timevarying o u t p u t gains. Strictly speaking. The static output feedback stabilization problem for linear and nonlinear (affine) systems is discussed.e. the t i m e . This characterization is used to establish a series of simple obstructions to the solvability of the problem for linear SISO systems. A few worked out examples complete the chapter.2 a n d P a t r i z i o Colaneri 1 1 Dip. In most cases. ac.
and one has just to comment on simple particular examples and applications. Astolfi and P. It has been shown that the use of such control laws enables one to deal with problems otherwise unsolvable with timeinvariant output controllers. they introduce higher frequency components centered at multiples of the sampling frequency. Many attempts have been made in the last years. and a partial converse is also proved.e. The nonlinear analysis is useful to formulate a novel necessary and sufficient condition for SOF stabilization of linear systems. Turning back to static output feedback controllers. a sufficient condition is provided based on the solution of a suitable HamiltonJacobi inequality.50 A. see [7. Among the various technical design approaches to output feedback stabilization. Indeed they are at all effect dynamical systems. [10]. the fundamental question of the existence (in the general case) of a stabilizing static output feedback control law is still open. However. which can possibly produce undesired ripples in the input and output responses. given in terms . the parity interlacing property [22] and its extended version [20] give particularly interesting insight. Unfortunately. Among them. This observation motivates the present paper. so that at this stage we can count several nontrivial contributions to the problem. All the contributions confirm that (generically) the SOF stabilization problem for linear timeinvariant systems is intrinsically nonlinear. such as pole assignment. simultaneous stabilization of a finite number of plants and gain margin improvement. where the SOF stabilization problem is consider for affine nonlinear systems. A simple way to stabilize a (dynamically output stabilizable) linear timeinvariant system is that of using a generalized sampleddata hold output feedback with a suitable sampling period. where the state of the art is presented and the existing methods are surveyed and compared. Precisely. the cone complementary algorithm of [6]. it is worth mentioning the technique related to the solution of coupled linear matrix inequalities.g. Several necessary conditions have been worked out.1. a definite theory on the benefits of timevarying static output feedback control is still missing. Moreover. i. we can say that. Colaneri they might stabilize a linear timeinvariant system which is not stabilizable by constant static output feedback control. the linearity of the system to be stabilized does not yield any special advantage in finding analytic solutions or systematic procedures. and the ILMI approach of [21]. A few algorithms have been proposed to get through the coupling condition. see [13]. despite the simplicity of its formulation. such control laws cannot be considered as static output feedback controllers. The study of benefits and inherent limitations of generalized sampleddata control systems are on their way and some aspects are still to be understood [9]. see the recent paper [19]. see e. both numerical and speculative. Simple testable sufficient conditions have also been introduced.18] for continuous time systems and [3] for discretetime systems. which also provide a parameterization of all SOF gains [12]. Among them the most interesting seem to be the m i n / m a x procedure proposed in [8].
The static output feedback stabilization problem for system (1)(2) consists in finding.I B is fat. establishing a nested series of testable necessary conditions (obstructions). Section 7 and Section 8 contain a few examples. This sets the basis for the introduction of the novel necessary and sufficient condition provided in Section 4. this is not unique. in the case of singleinput singleoutput systems. In Section 2 the SOF stabilization problem for linear systems is recalled and a preliminary result is stated. B. where the sufficient and the necessary conditions are provided in terms of the solution of a constrained HamiltonJacobi equation along with a rank condition. whereas a few interesting issues. some obstructions. and A. ( A 3 ) m < p. B} is controllable and B has full column rank. From this equality. the matrix A + B F C has all its eigenvalues with negative real parts. i.e. respectively. if possible. This chapter is organized as follows. a static control law described by = Fy (3) such that the closedloop system is asymptotically stable. 2 Preliminary results Consider the continuoustime linear system x=Ax = C~ + Bu (1) (2) where x E ~ n . ( A 1 ) The pair {A. Finally.e. (A2) The pair {A. If such an output feedback does exist. Further results are proposed in Section 5. This can be seen as a slight modification of the necessary and sufficient condition given in [14]. C are matrices with constant real coefficients and appropriate dimensions. the transfer function W ( s ) = C ( s I . it is possible to go deeply into the structure of the problem.A ) . input and output vectors. we say that the system (1)(2) is output stabilizable and that F is a solution of the problem 1. .Static Output Feedback Stabilization 51 of the coupled solution to a linear matrix inequality and a linear equality. C} is observable and C has full row rank. some comments and hints for future research. y E /Rv are the state. and a discussion on the convexity properties for SISO systems are given in Section 6. In what follows. u E ~ m . if a solution exists. i. The SOF stabilization problem for nonlinear systems is tackled in Section 3. whenever we deal with the linear system (1)(2) we make the following standing assumptions. 1 It is obvious that.
i. Observe that by Assumption (A2) the matrix V is welldefined.I B . [14]. Setting G . The system is output feedback stabilizable if and only if there exist a symmetric positive semidefinite matrix P E ~ n x n and a matrix G E IR T M such that 0 = A'P + PA . the above equation can be rewritten as 0 = A ' P + P A .e. where 2 V = I . see e. if m > p. J2 = F ' ) is a solution of the static output feedback stabilization problem for system (1)(2) (resp. multiplying both sides of the above Riccati equation by V. Moreover.P B B ' P + C C + GIG.(2 (resp. (4)(5)).52 A. it is possible to consider the system = A'A + C'v ~I=B'A (4) (5) for whichAssumption(A3) holds. Moreover. it can be found here and there in the existing literature. Then. Proof. and observe that the static output feedback stabilization problem for system (4)(5) is solvable if and only if it is solvable for system (1)(2).PBB'P + C'C + G'G (6) (7) 0 = V ( A ' P + PA)V. whose proof is reported here for the sake of completeness. even though.g. T h e o r e m 1. A simple necessary and sufficient condition for the system to be output stabilizable is stated in the following result. Colaneri The above assumptions are without loss of generality. it follows V ( A ' P + P A ) V = O.F C + B ' P . the rank assumptions can be always enforced. in slight different forms. Consider the system (1)(2) with Assumptions (A1). Finally. Astolfi and P. notice that G V = B ~ P V and then. the solvability of the problem depends only upon the controllable and observable part of the system. by elimination of redundant control inputs or measured outputs. on the transfer function W ( s ) = C ( s I .C'(CC)IC. (Only if. if . First of all.A ) . (1)(2)) then F /2' (resp.) Assume that the system (1)(2) is output stabilizable and let F be such that A + B F C is asymptotically stable. F) is a solution of the static output feedback stabilization problem for system (4)(5) (resp. . ( A 2 ) and (A3). Then the Lyapunov Lemma implies the existence of a unique positive semidefinite solution of the Lyapunov equation 0 = (A + BFC)'P+ P(A + BFC) + C'C + C'F'FC.
e. It is worth noting that. As a result VPBB'PV T B ' PV.g.B ' P ) .V) = ( T ' G . 3 A matrix T is said to be orthogonal if T ' T = T T ' .) Assume that there exist P > 0 and G satisfying equations (6) and (7). substituting this last equation in equation (6) one has 0 = (A + BFC)'P + P(A + BFC) + C'C + C'F'FC. C'C+C'F'FC}. (8) Unfortunately.Static Output Feedback Stabilization 53 (If. . it is only possible to prove that the conditions expressed by equations (8) and (7) are necessary for static output feedback stabilizability. Theorem I provides also a parameterization of all stabilizing static output feedback control laws. the conditions in Theorem 1 cannot be recast (to the best of the authors knowledge) in an LMI framework or in a simple computational scheme. e. To recover also the sufficient part of the statement.B'P)C'(CC') I where P = P ' >_ 0 and G solve (6) and (7) and T is any orthogonal matrix. i. rank(T~(P)) < m. P > O. despite their simplicity. using simple linear arguments. Note finally > that. Consider the system (1)(2).I. C}. In the same way. This implies that there exists an orthogonal 3 matrix T such that G V = it follows F C = ( T ' G . The f a m i l y of all output feedback gains F such that the matrix A + B F C is stable is given by F = (T'G . a (nonconstant) matrix T(x) is said to be orthogonal if it is orthogonal for each fixed x. Finally. because in general the static output feedback stabilization problem is nonconvex.P B B ' P + C ' C = T~(P). C} is equivalent to the observabilityof the pair {(A+BFC). whose proof is easily obtained from the proof Theorem I. it would be tempting to replace equation (6) with the matrix inequality 0 > A ' P + P A . to obtain simpler conditions.B ' P ) ( I . as expressed in the followingstatement.B ' P ) C ' ( C C ' ) ~ = VG'GV.the existenceof P _ 0 impliesthe stability of (A+BFC). one has to add a further rank condition on 7~(F). hence G ' G = G ' T T ' G = (FC + B'P)'(FC + B ' P ) . conditions involving only the matrix P. Observe now that the observability of the pair {A. hence by the Lyapunov Lemma. by observabilityof the pair {A. Defining r = ( T ' G . Moreover. C o r o l l a r y 1.
Astolfi and P.I . Consider the system (9)(I0) and assume that the pair {f. i. 3 S t a t i c o u t p u t f e e d b a c k s t a b i l i z a t i o n for n o n l i n e a r systems In this section we consider a nonlinear system described by equations of the form = f ( x ) ~. the m a t r i x P .h(x). Theorem 2.54 A. f(0) = 0 and h(0) = 0. Given a smooth m a p p i n g y . see e. D e f i n i t i o n 2. we also assume that x = 0 is an equilibrium point. h} is said to be locally detectable (observable) if there exists a neighborhood U of the point x = 0 such that. using the more general formulation and solution of the problem given in the next section. it is fairly s t a n d a r d to recognize that equation (8) can be given an equivalent LMI formulation in the unknown P . and the mappings f(x). '~ y E JRp the measured output. it is possible to prove t h a t the equations (8) and (7) provide a necessary and sufficient condition for the solvability of the static o u t p u t feedback stabilization problem.g(x)u (9) (10) y=h(x) where x E J~R denotes the state of the system. as they involve only one unknown. h} is locally detectable and that rank (dh(O) ) = p. We are now ready to present the main result of this section. i. D e f i n i t i o n 1. . 4 This is also denoted with hl(0). In order to describe the main results of the section we need the following definitions. Suppose moreover that there exist a scalar function V(x) E C 1. (11) (12) V x E Ker(h).~V~(x)g(x)g'(x)V~(x) + h'(x)h(x) + G'(x)G(x) 0 = Yxf(x). if x(t) is any integral curve of x = f ( x ) satisfying x(0) E U. positive definite in a neighborhood of the origin. Colaneri However. we denote with 4 ker(h) the set of all x such that y . u E ~ ' ~ the control input.e.g. which is a (partial) nonlinear counterpart of Theorem 1. then h(x(t)) is defined for all t > 0 and h(x(t)) = 0 for all t >__0 implies limt~oo x(t) = 0 (x(t) = 0 for all t>o). Moreover.0. and a m • 1 matrix function G(x) E C 1 such that 0 = V~(x)f(x) . g(x) and h(x) are s m o o t h mappings defined in a neighborhood of the origin of ~'~. [11].e. These conditions are obviously simpler than conditions (6) and (7). The pair {f. Moreover.
i.rf(x(t))rl(x(t)). By equations (11) and (12) it follows 1V~(x)g(x)g'(x)V~(x) = G'(x)G(x). one has v(x(0)) = which establishes the claim. V x E Ker(h). (iii) the trajectories x(t) of the system x = f ( x ) + g(x)r/(x) starting close to the origin are such that the output y(t) = h(x(t)) and the control u(t) = r/(x(t)) are square integrable signals.l(x) = T(x)G(x) . the HamiltonJacobi equation (11) can be rewritten as 0 = V~(x)(f(x) + g(x)~(x)) + h'(x)h(x) + 7]'(x)~(x) which. equation (15) and the rank theorem. h}. see [11. ~l(x) is a function of y. V x e Ker(h). then (iv) in a neighborhood of the origin.Ig'(x)V'(x) is such that (i) for all x E Ker(h) r/(x) = 0. r/(x) = r for some smooth function r Proof. Appendix 3].h ' ( x ( t ) ) h ( x ( t ) ) . Moreover.e. . (13) (14) (ii) the system x = f ( x ) + g(x)y(x) is locally asymptotically stable.Static Output Feedback Stabilization 55 Then there exists a orthogonal matrix T ( x ) E C 1 such that the function . by local asymptotic stability of the closedloop system. by detectability of the pair {f. if p > k = rank(O~O(~) ) Vx e / 2 (15) for some constant k and some neighborhood/2 of x = O. Point (i). Point (iv). implies claim (ii). Point (ii}. Hence.(x). This means that there exists a smooth orthogonal matrix T ( x ) such that 1 ~ T ( x ) G ( x ) = ~g (x)V. This claim follows from the assumptions. (16) Point (iii). Equation (16) can be rewritten as 9 = . By equation (13). F IIh(x(t))ll ~ + IMx(t))ll2dt.
positive definite in a neighborhood 12 of the origin. in the linear case.~g (x)V~:(x). positive definite in a neighborhood of the origin. h} is locally observable. a m • 1 continuous function G(x). such that O> W x ( x ) f ( x ) . and an orthogonal matrix T(x) E C 1 such that 1 (i) 0 = Vz(x)f(x) . V x E Ker(h) (iv) p> r a . as they require the solution of a HamiltonJacobi equation. Proof The claim is trivially obtained setting V(x) = W ( x ) and G(x) such that G'(x)G(x) = V~:(x)f(x) + 1V~:(x)g(x)g'(x)V~(x) . . T h e o r e m 3. They are obviously more involved. Colaneri The sufficient conditions in Theorem 2 are the nonlinear equivalent of the sufficient conditions in Theorem 1. there exist a (nonunique) m x 1 matrix function G(x) e C t and a scalar function V(x) E C 1. V x e Ker(h) 1 . However. positive definite in a neighborhood of the origin.1W~:(x)g(x)g'(x)W~(x) + h'(x)h(x) (17) o= w~f(x). such that equations (11) and (12) are satisfied. for any fixed matrix G there exists a (unique) matrix P = P ' > 0 solving the Riccati equation (6).h'(x)h(x) > O. Observe that Theorem 2 admits the following partial converse. v x ~ Ker(h). Assume that there exists a scalar function W ( x ) E C 1. (iii) O = T ( x ) G ( x ) . as detailed in the following statement. o6 vx e ~.56 A. Assume moreover that there exists a continuous function r with r = 0. Consider the system (9)(10) and assume that the pair {f. k ( ~ ) . C o r o l l a r y 2. (18) Then.~V~:(x)g(x)g(x)'Vz(x) + h'(x)h(x) + G'(x)G(x) (ii) O= Vr(x)f(x). such that (a) x = f ( x ) + g(x)O(y) is locally asymptotically stable. Astolfi and P. Moreover. the nonlinear formulation allows to replace the HamiltonJacobi equation (11) with a HamiltonJacobi inequality involving only one unknown. Then there exist a scalar function V(x) E C 1. (b) the trajectories x(t) of the system x = f ( x ) + g(x)r starting close to the origin are such that the output y(t) = h(x(t)) and the control u(t) = r ) ave square integrable signals. whereas this fact is not in general true for the HamiltonJacobi equation (11).
one can easily show that equation (6) implies the inequality (19). T h e o r e m 4. P > 0 by Assumption (A2) and observe that equations (7) and (20) are the same. Then. Moreover. h}. By assumption (b) and the observability of the pair {f. it satisfies equation (16) with ~/(x) replaced by r Set now G(x) : r I. setting X = p . It is worth noting that the proposed condition heavily stems from the nonlinear control design methods established in the previous section. Then conditions (17) and (18) are satisfied with W ( x ) = x'Px. (20) and (21) hold. Consider the system (1)(2) with Assumptions (A1). This characterization is given in term of the intersection of two convex sets subject to a (generically) nonconvex coupling condition. claim (iv) follows directly from r being a function of y only. by the necessity part of Theorem 1 there exist a symmetric positive semidefinite matrix P and a matrix G such that such that equations (6) and (7) are satisfied.AX + BB' XC'] CX I (19) 0 = V(A'P + PA)V I = PX. let x(t) be the corresponding state trajectory of the system Jc = f ( x ) + g(x)r and define v(~) = IIh(~(t))ll ~ + ID(y(t))ll2dt.Static Output Feedback Stabilization 57 Proof. The system is output feedback stabilizable if and only if there exist two symmetric positive definite matrices X and P such that 0< XA' . the resulting output feedback controller is nonlinear. Given an initial condition ~ E /2. but (under simple regularity assumptions) a linear feedback can be computed. Then. Moreover. equation (21)). in general. we conclude that V(~') is positive definite and differentiable in/2. Moreover. 4 A new characterization for linear systems In this section we exploit the results established in Section 3 to derive a new necessary and sufficient condition for the solvability of the output feedback stabilizability problem for system (1)(2).e.I (i. (Only i]) Assume that the output feedback stabilization problem for system (1)(2) is solvable. . (lj') Assume now that there exist P > 0 and X > 0 such that (19). Finally. (20) (21) Proof.• ~ jV~ and ' T(x) = I and observe that points (i)(iii) are trivially satisfied. (A2) and (A3).
and there exists a point xe arbitrarily close to x . (ii) and (iii) in T h e o r e m (2).B ' P x satisfies conditions (i).. Colaneri f ( x ) = A x . x* < _ m < p . Consider now the the (m • n)dimensional m a t r i x D~(~) = O~ and let . for any xe in /~e N ker(Cx). by Theorem 2 there exists an orthogonal matrix T ( x ) 6 C 1 such t h a t the function y(x) = T ( x ) G ( x ) . for any point x in Be f'l ker(Cx) one has D.I(X) = 0~1 y=O C s As a matter of fact. in a neighborhood/3e C Be of xe the function ~(x) is such that ~(z) = r = r (22) To conclude the proof observe that by condition (ii) of Theorem (2) and equation (16) the function q(x) is such that k = A x + B~I(x) is locally exponentially stable. . As a result the m a t r i x A + BD~](O) has all eigenvalues with negative real. Note that. Moreover.2 g ' ( x ) V ' ( x ) = T ( x ) G ( x ) . by Assumption (A3). close r zero and such that r/(x~) = 0 are singular points of r/(x). i. However.e. In this case it is still possible s draw the same conclusions but with more involved considerations. By the Rank Theorem (see [11]) and condition (i) of T h e o r e m (2).~* = maxrank(D~/(z)) z6O w h e r e / 2 is a neighborhood of the origin. and by a simple continuity argument also the matrix A + BDq(x~) has all eigenvalues with negative real. the rank of r/(x) is nonconstant in any neighborhood of such points. (b) rank(Dr](x)) .r < x* for all x in a neighborhood Be of xe. it is possible that all points x. Astolfi and P.0 such that s (a) ~(xe) = 0. and by Corollary 2 there exist a positive definite function V ( x ) 6 C 1 and a m a t r i x G(x) 6 C 1 such t h a t conditions (11) and (12) are also satisfied.58 A. g(x) = B and h(x) = C x .
The result summarized in Theorem 4 can be easily extended to the case where not only stability but also an H ~ performance bound is taken into account.1 SOF stabilization. Notice that the set of all X satisfying the condition (19) in Theorem 4 is convex. Remark 1. state feedback and output injection In the introduction it has been noted that the problem of dynamic output feedback stabilization for system (1)(2) can be recast as a SOF stabilization . consider a system described by the equations ic = A x + B u + B w (23) and such that Assumptions (A1). and it is also convex the set of all P satisfying condition (20). Then it is possible to extend the results in Theorems 1 and 4 to prove that there exists a static output feedback control law u = F y such that the closedloop system is asymptotically stable and the H ~ norm from the input w to the output z is less then a prespecified level 7 if and only if there exist two symmetric positive definite matrices X and P such that 0<_ XA'AX + BB' CX ~ XC' I (25) O=V(A'P+PA+ I = PX. For. P B~ ' P . Unfortunately. 5. Remark 2. Moreover. (A2) and (A3) hold. ) V B (26) (27) 5 Further results and special cases In this section we focus on linear systems and propose further interesting characterizations of the SOF stabilization problem. We conclude that the problem is in general nonconvex. the SOF stabilization problem for a class of systems with special structure is addressed. which proves that 59 0r F ~ ~yy y=0 is a solution of the static output feedback stabilization problem.Static Output Feedback Stabilization where r is defined in equation (22). the coupling condition (21) is not convex.
Then. Then there exists X = X ~ > 0 such that (A + B F C ) ' X + X ( A + B F C ) < O. the separation principle implies that dynamic output feedback stabilization problems can be solved if and only if stabilization through state feedback and through output injection can be achieved. K = FC. In particular. (Only i]) Assume that the system (1)(2) is static output feedback stabilizable and F is a solution to the problem. As a result. There exists a state feedback gain K and an output injections gain L such that the systems ~ = (A + B h ' ) x and ~ = (A + LC)~ are asymptotically stable and possess a common Lyapunov function. as discussed in the following statement. This is equivalent to say the SOF stabilization problem is solvable if and only if it is possible to solve a state feedback stabilization problem and an output injection stabilization problem in a way that the corresponding closed loop systems have a common Lyapunov function. the result follows trivially from Theorem 3. 5.B K ) ' P + P ( A q.B K ) < 0 (A + L C ) ' P + P ( A + LC) < O. L = B F P = X are such that equations (28) hold. namely equation (6). Colaneri problem for an extended dynamical system. Consider the system (1)(2). The SOF stabilization problem is solvable if and only if there exist matrices (of appropriate dimensions) K.2 T h e dual o f T h e o r e m 4 Theorem 4 has been derived from Theorem 1 substituting the first equality. L and P = P' > 0 such that (A q. and indeed proves to be of interest. namely the substitution of equation (7) with an inequality might be appealing in some applications. As already remarked. whereas to prove the sufficiency we had to consider the (more general) nonlinear problem discussed in Section 3.8 in [19]. also the dual approach. However. if the idea of removing the unknown G from equation (6) and substituting the equality with an inequality is very natural. The conditions in Theorem 5 can be interpreted as follows. whenever the system . Proof. (28) Remark 3. Astolfi and P. (I]) Assume equations (28) hold. T h e o r e m 5.60 A. the necessity part of Theorem 4 is a straightforward consequence of Theorem 1. with an inequality and removing the unknown G. Moreover. The connection between static and dynamic output feedback stabilization can be strengthened.
Consider the system (1)(2) with Assumptions (A1). T h e o r e m 6.3 Minimum phase systems with relative degree one Consider the class of singleinput singleoutput linear systems with relative degree one and asymptotically stable zero dynamics. (29) (3o) 5. Consider the system (1)(2) and assume that p = m = 1 and that A. in a proper set of coordinates and after a preliminary static output feedback. where V is as in Theorem 1. To begin with.e. by asymptotic stability of the zero dynamics a(F) C ~ . let . Suppose moreover that the system has an asymptotically stable zero dynamics. Then the system is (obviously) asymptotic stabilizable by static output feedback.Static Output Feedback Stabilization 61 to be stabilized is such that there exists a positive definite matrix P with the property that A ' P 4. B and C are as in (31). (31) Moreover.P A < 0. condition (7) is automatically fulfilled if the equality sign is changed into an inequality.PBB'P + C'C + G'G O> V ( A ' P + PA)V. as discussed in the next statement. as A= h 0 B= C= [0 1 ] . We are now ready to state the main result of this section. For this class of systems a simple root locus reasoning shows that asymptotic stability can be achieved by means of high gain static output feedback. To make use of this a priori knowledge on the system it is necessary to derive a new characterization for the SOF stabilization problem. The system is output feedback stabilizable if and only if there exist a symmetric positive semidefinite matrix P E IR nxn and a matrix G E ff~mxn such that 0 = A'P + PA . Goal of this section is to show how it is possible to recover this result from the theory presented in this work. (..42) and (A3). the system is marginally stable. i. Moreover. note that the system matrices of a singleinput singleoutput linear system with relative degree one can be written. P r o p o s i t i o n 1.
Colaneri with Po = P~ > 0 such that F'Po + PoF + h'h = O.1 and satisfying A s s u m p t i o n (A1) and (A2). can be given an equivalent coordinates free formulation.. Astolfi and P. (32) .P0 .~iciently small 1 and an estimate of the minimum value of s required to achieve closedloop stability is 1 2Crma• We omit the proof of the above result which is conceptually simple. which strongly rely on a particular selection of coordinates.:o [00""1 : : ". The extension to nonlinear systems is currently under investigation. a.1 [0 a..62 A..~ ] 0 a. but it must be noted that some interesting results in this direction can be found in [4. for a su.. i.= 0 o~+0 and some matrix G such that lim G . However. we point out t h a t the result expressed in Proposition 1.e. a stabilizing output feedback gain is.1 A /oo. : " al J . 6 SISO linear properties systems. Without loss of generality it is possible to write the system in the observability canonical form.. a similar result can be obtained for systems with asymptotically stable zero dynamics and relative degree two and for multivariable square systems.+O Ot 1/2] = 0. and = then conditions (6) and (7) of Theorem I hold for some positive definite matrix P such that t"1 lim P . Finally. Furthermore..0 10 .5]. with 00. obstructions and convexity In this section we restrict our interest to linear systems described by equations of the form (1)(2) with p : m ..
They have been both included since they express obstructions of diverse complexity and nature. The system is output feedback stabilizable only if there exists a positive definite matrix X satisfying the following conditions. If they are feasible. . If this procedure fails. Remark 4. (iii) X E X. condition (iii) can be added. Theorem 7 provides an obstruction to the solvability of the static output feedback stabilization problem. we define the column vector ei as the i . and finally conditions (i) to (iv) has to be checked together. Observe that condition (iii) of Theorem 7 implies condition (ii). for convenience.k n3 "k en_4AenA n .th vector of the identity matrix of dimension n and the set X = { X = X ' I Xij = 0 i f i + j i s o d d } . has distinct roots all on the imaginary axis. T h e o r e m 7..~ "4''' ' v . First condition (i) is checked. (34) 5O e~Xen = 0 . then no solution exists. no conclusion can be drawn. (iv) the polynomial elnXen. if the procedure works. The proof is technical and it is omitted for brevity.Static Output Feedback Stabilization and C= [0001]. It is interesting to note that conditions (i) to (iii) in Theorem 6 are convex. (36) i=n1. It can be used in steps of increasing complexity. n5. as formalized in the following statement. n3. Proof.. (i) 0 <_ [XA'AX + B B ' XIC' ] CX . as described in the following statement. whereas condition (iv) is in general nonconvex. Consider the system (1)(2) with A and C as in equations (32) and (33). then conditions (i) and (ii) can be simultaneously checked. However. for low order systems the obstruction expressed in Theorem 7 can be easily checked. However.k n1 + eln_~Xen. If it is feasible. (35) . where. 63 (33) This special form allows to get more insight into the structure of the problem.
e. 7 A few illustrative examples In this section we discuss a few simple examples. If n < 4 the set of all X > 0 satisfying conditions (i) to (iv) of Theorem 7 is convex. The system passes the first test. A simple application of the necessary conditions in Theorem 7 yields the following conclusions. Theorem 7 provides a very simple test that can be also performed with pen and paper. Aim of the section is to show how the general results developed so far can be easily used in deciding the solvability of SOF stabilization problems for linear and nonlinear systems. I f n = 5 or n = 6 the conditions (i) to (iv) of Theorem 7 can be recast in a convex optimization problem. both the PIP and the E P I P do not give an obstruction to the solvability of the SOF stabilization problem. i. If n <_ 3 and B = C' the set of all X > 0 satisfying conditions (i) to (iv) of Theorem 7 is convex and conditions (i) to (iv) of Theorem 7 are also su]ficient. i. the property that guarantees the existence of a stable stabilizer with a stable inverse. In particular. systems which are Lyapunov stable and/or possess conserved quantities.( 2 ) with the matrices A and C as in equations (Se) and (S3). i. which requires . the sufficient conditions expressed in Theorem 2 is well suited to study conservative systems.e. Colaneri C o r o l l a r y 3. Therefore. the property that guarantees the existence of a stable stabilizer. However. for linear low dimensional systems. Example 1. Astolfi and P. Consider a linear 2dimensional system in observability canonical form with A:[01 and observe that v__ (37) This system satisfies the Parity Interlacing Property (PIP) [19]. and satisfies also the Extended PIP [19].e. as a quadratic optimization with linear constraints. If n < 2 the set of all X > 0 satisfying conditions (i) to (iv) of Theorem 7 is convex and conditions (i) to (iv) of Theorem 7 are also suj~icient.e. For nonlinear systems. Consider the system ( I ) . as the test expressed by condition (34) is passed by any controllable and observable system.e. the system does not pass the second test. there exists a positive definite matrix X such that condition (34) holds. This is not surprising. i.64 A. i.
yet there is no stable stabilizer with stable inverse of order less then or equal to two. we obtain the nonlinear state feedback control law = a(. Example 2.e. which also shows that this set is convex. We conclude that the system is not stabilizable by static output feedback. i. i.Static Output Feedback Stabilization 65 the matrix X to be diagonal. construct the function G ( x ) . For the considered example it is easy to prove that there exists a stable stabilizer of order one. whereas condition (20) is trivially fulfilled if P is a diagonal matrix. the order of these stabilizers m a y be large.8P1P3 . (38) and that the necessary conditions of Theorem 7 require the matrix X to be diagonal. i. As a result. we select an admissible pair {P1. Consider a linear 2dimensional system in observability canonical form with B= Observe that [41].e. It is worth noting that. C=[01]. P3}. and the scalar T ( x ) .e.8 P I P ~ 0<P1 0<P3. However. The set of all feasible P .r + .) . in general. as the system satisfies both the P I P and the E P I P there exist stable stabilizers and also stable stabilizers with stable inverse. {1. To construct the stabilizing gain F we follow the procedure outlined in the proof of Theorem 2. there exists no positive definite diagonal matrix X such that condition (34) holds. the set of all P fulfilling condition (19) (with X = p . R e m a r k 5.32P• . e. Simple calculations show that the set of all positive definite o P3 = satisfying condition (19) is described by i/x3 0] 0 < 64P?P3 .l ) and condition (20). This implies that the sufficient conditions of Theorem 4 has to be tested with X and P diagonal. is illustrated in Figure 1. 1}. For.P~ .g.
oo).(1. takes value in the set 6 (1. Astolfi and P. Example 3. The set of all pair {/'1.3). without selecting a priori a value for the pair {/'1. If the pair {/'1. This general procedure yields a family of stabilizing gains.5 4 4.5 3 P3 2.! 01.5 .5 2 2.5 5 P1 Fig./ ' 3 .45. the function F(P3). depicted in Figure 2. Colaneri 5 4. which is indeed a stabilizing gain. 1) is selected the corresponding output feedback gain is F . P3 6 (1/2. .66 A.e.5 3.z ) and condition (20). Consider a rigid body in an inertial reference frame and let zl. For P3 6 (1/2. 1. In this example it is possible to perform the calculations of the stabilizing gain in a parametric form.5 3 3. oo).1. z2 and z3 denote the angular m o m e n t u m components along a body fixed reference frame having the origin at the center of gravity and consisting of 6 Note that limp34oo F(P3) = 3. as can be seen from Figure 1 and verified with very simple calculations./'3} . where. and a very simple root locus analysis (see Figure 3) reveals that these are indeed all the static output feedback gains yielding closedloop stability./'3}. To compute the output feedback gain we have to differentiate the function q(x) at some point close to the origin and such that C z = 0. 1.~ 2 0.1 + P32 . i. which is described by the equation F : F(P3) = ~/6Pa . P3} satisfying condition (19) (with X = p . yielding local asymptotic stability for the closedloop system.
8 1. cij are constant numbers which reflect the position of the sensing The problem of stabilization using d y n a m i c feedback (either observer based or not) has been addressed in [2.16]. The static output feedback gain F = (1/2.201.8 1.Static Output Feedback Stabilization 2.2 i i i i P3 Fig.C 2 3 X 3 where the devices. and a and fl are two constant numbers depending on the location of the actuators.2 2 FIP 3) 1.4 2.for 1~ E the three principal axes. We now focus on the static o u t p u t feedback stabilization problem and we show how it is possible to make use . The only variables available for feedback are Yl = C l l ~ l Y2 : "31C12Z2 + C 1 3 X 3 (40) C 2 1 X l "Jr C 2 2 Z 2 1.8 2. The Euler's equations for the rigid body with two independent controls are ~2= (][ ]3)x3xlqu2 z3= (~~2 (39) ~)xlx2+aul+~u2 where Is > 0. 2. F(P3)as a function of P3.a 67 2. 12 > 0 and /3 > 0 denote the principal m o m e n t s of inertia. u = col(ul.u~) the control torques.4 1.
The point of departure is the consideration t h a t for system (39) there exist two positive definite functions which are natural candidates to test the conditions in T h e o r e m 2. i. The next step is to verify if it is possible to find constants A and /~ such that the function V(x) is positive definite and equation (ii) in T h e o r e m . . n a m e l y the kinetic energy. . i. In fact. Observe that for F = 1 and F = 3 the eigenvMues of the closedloop system are on the imaginary axis. As a result the function v(x) = ~P(x) + ~K(x) satisfies equation (12) for any A and/~. l(x~ x~ d~ K(x)= ~ \ii +~ + I3] and the m o d u l o of the angular m o m e n t u m vector. . i F=I =5 i 4 i 3 i 2 I 1 I 0 Real Axis t 1 2 I 3 I 4 5 Fig. 3). .68 A. . Colaneri . P(x) = ~ (x~ + x~ + x~) . of the result in Theorem 2. ~ . Astolfi and P.e. simple calculations (which can be avoided by anyone with a basic knowledge of Newton's laws) show t h a t along the trajectories of the system (39) with Ul = u2 = 0 one has 1 k=P=0. 3. The root locus of the system (38) for F E (1.e.
In the space (A./~) the set E = n. Remark 7.e. where J = and [ (/~C1~ C13)I213 flCl2I~ . the system can be nonasymptotically stabilized by static output feedback.X. for some G(z). It is worth noting that. The condition expressed in Proposition 2 can be given a simple geometrical interpretation. LC=l e22 c23 J The system is asymptotically stabilizable via static output feedback if IJ {(A. the shaded area in Figure 4. whereas to prove the necessity two eases has to be considered. if d e t J ~= 0. i. Suppose that rank [ell ] = 2.~) I /t .2.e. as shown if Figure 4. the case 7 J n (0E/{0. In the former.7 = {(A. the condition expressed in Proposition 2 is necessary and sufficient.ClaI3 ] aC1211 + C23Ia J Cli ctj 1 I (aC12+C23)IlI3 Cij = det Lc~i c2j j Remark 6. r The notation OE is used to denote the boundary of the set E. the proof of which is a consequence of the above considerations. the static output feedback stabilization problem for system (39) with output (40) is solvable if and only if fiNE#0. As a result. i.X+ T~ > 0} is a convex set. 0}) # and the case J" N E = {0. if d e t J = 0. . whereas in the latter a simple application of classical instability theorems shows that for any (static) output feedback gain the closedloop system is unstable./~) 6 Kerg} ni=1. the sufficient condition in Proposition 2 could be easily tested. 0).p) C Kerd} is either a straight line. p) = (0. if d e t J = 0 and rankC = 2./J) ] A + ~//> O} :/: O. whereas the set .3{(. or the singleton (A.3 {(A. 0}. P r o p o s i t i o n 2.2.e.=l.Static Output Feedback Stabilization 69 2 is satisfied. i. Consider the system (39) with output (~0). For the sake of brevity we omit the details of the discussion and present only the main result. The sufficient part is a consequence of Proposition (2).
"Timevarying feedback laws for decentralized control" I E E E Trans. B. can be exploited to decide the unsolvability of t h e problem.k e r J intersects the shaded region then the stabilization problem is solvable.i n p u t m u l t i . Anderson and J.o u t p u t linear systems.B.O.70 A. . to provide a description of all stabilizing gains. Colaneri F i g . For linear systems a new necessary and sufficient condition has been proposed. T h e s e necessary conditions. a n d to work out further nonlinear physically m o t i v a t e d examples. A sufficient condition for nonlinear control affine systems has also been developed. Moreover.o u t p u t linear systems a set of simple to test necessary conditions have been derived. Astolfi and P. to establish to w h a t extent the necessary condition in T h e o r e m 7 is also sufficient. otherwise the sufficient condition does not yield any conclusion. Vol. 8 Concluding remarks T h e problem of s t a t i c o u t p u t feedback s t a b i l i z a t i o n for linear a n d n o n l i n e a r systems has been studied. pp. References 1. Graphical representation of the sets E and J .7" . 4. It is worth noting t h a t the p r o o f o f the sufficient condition for nonlinear systems is i n s t r u m e n t a l to develop t h e new characterization for linear systems. If . Moore. 11331139. 1981. A u t o m .D. Control. together with a p a r t i a l converse. 26. for singleinput s i n g l e . which can be recast in t e r m s of c o n v e x c o n d i t i o n s for low dimensional systems. F u t u r e work will be directed toward the s t u d y of m u l t i . T h e theoretical p a r t is c o m p l e m e n t e d with a few worked out e x a m p l e s .
Haddad. Kucera and C. Control. Vol. A.]. Vol. 39.C. Systems and Control Letters. Vol. Vol. Journal Optim. 1992.I. 925927. 559578. l: problem formulation and reduction to optimal static output feedback".L. pp. Vol. feedback equivalence. Nett. pp. J. 36. 1990. "The order of any stabilizing regulator is sufficient a priori information for adaptive stabilization".X. 159173. pp. 6. de Souza and R. "Single loop stabilization of linear multivariable dynamical systems". [EEE Trans. Iwasaki and R. and the global stabilization of minimum phase nonlinear systems". Int. Vol. 15. 32. Vol. 1994. "Minimal complexity control law synthesis. ]sidori. 4. "Control of linear systems using generalized sampleddata functions". 3. 747758. pp.I.L. "Static output feedback stabilization: an ILMI approach". 17. pp. 34. D. . sensitivity. 291307. Cao. D. Vol. 1998. J. Control. N. 4. Automatica. IEEE Trans. Skelton. Automatica. pp. 13. Bernstein and W. 20562064. Systems and Control Letters. Autom. N. "Pole assignment for linear timeinvariant systems by periodic memoryless output feedback". 43. 12. Byrnes and A. "Asymptotic stabilization of minimum phase nonlinear systems". 8791. Y. C. 11. Control. "Robust output feedback stabilization of the angular velocity of a rigid body". Vol. Vol.W.Y. Vol. De Souza. 6.E. A. and intersample difficulties". Aeyels and J. Goodwin. 1997. C. Feuer and G. 125137. Bongiorno and C. "Stabilizability conditions of multivariable uncertain systems via output feedback". 2000. Astolfi. Vol. 10. Isidori and J. 39. 1987. 1991. P. pp. Aurora. 5. A. "Passivity. "Static output feedback controllers: stability and convexity". C. Gu. IEEE Trans. 10. "L~ stability of timevarying systems . Lu. Springer Verlag. 1999. IEEE Trans. pp. Control. pp. Grigoriadis.C. Wei. 1995. A. V. 1990. 28. 11711176. "Static output feedback . "Output feedback control of the angular velocity of a rigid body". pp.C. 19. 36. Freedman. lsidori. K. V. pp. Syrmos. Control. Astolfi. Aurora. Kabamba. A. Systems and Control Letters. "Generalized sampleddata functions: frequency domain analysis and robustness. 16411645. Willems. Sun.. 120125. Automatica. Oustry and M. 11221137. 203210. Geromel. pp. G. 7. 33. Int. 181192. Lam and Y. C. 6. F. 8. Part. Vol. Skelton. 13571359. 2. AitRami. Dorato and K. (USA). 1995. "A cone complementarity linearization algorithm for static outputfeedback and related problems". pp.Static Output Feedback Stabilization 71 2. Proc. 21. C.. Automatica Vol. Martensson. 12281240. pp. "Stabilization of a linear plant via a stable compensator having no real unstable zeros". 42. Vo].A survey". IEEE Transactions on A utomatic Control. Nonlinear Control Systems. Vol. Vol. El Ghaoui. pp. pp. T. M. pp. IEEE Transactions on Automatic Control. J. J. American Control Conference. Olbrot. 45. 15. 31. Youla. 36. pp. pp. L. B. Aurora. "Parameterization of all stabilizing controllers via quadratic Lyapunov functions".T. Pittsburgh. "A necessary and sufficient condition for output feedback stabilizability". 35. 11591168. 77.C. 22. P. Automatica. 1987. Control. 1968. 772783. 18. IEEE Trans. 10421047. 20. Mazenc and A. 3. "Robust stabilization of uncertain systems by periodic feedback". 259264. 1997. 9. Vol. 1998. Control.construction of multipliers with prescribed phase characteristics". 16. F. J.A. 1991. Vol. pp. Willems. D. Systems and Control Letters. Byrnes. 14. Autom.T. AbdaUah. Theory Applic.
e n s t . f r / g t .l i n e a r D i f f u s i v e R e p r e s e n t a t i o n s for Nonlinear Fractional Differential Systems Jacques A u d o u n e t t. e n s t .http://www. x(O) = x 0 . m e a n i n g d~x = I x . Finally. route de Narbonne.h t t p : / / w w w . .o p d / ignon/ A b s t r a c t .L i o u v i l l e fractional derivative. with 0 < a < 1. 1 Statement of the problem We are interested in the following p r o b l e m involving a n o n . & CNRS. f r / ~ m a t 3 LAAS / CNRS 7. The stability of nonllnear fractional differential equations is studied. A sufficient stability condition on the nonlinearity is given for the inputoutput stability. France audounet @~ip. a fractional version of the classical HartmanGrobman theorem for hyperbolic dynamical systems of order 1 is conjectured and reformulated. URA 820 46.~ * x . The author would like to thank Raphael Krikorian. f r .. France montseny@@laas . thanks to many different reformulations of the system using difnsive representations of dissipative pseudodifferential operators. l a a s . (1) where d a is the C a p u t o regularized version o f the socalled R i e m a n n . fr 2 Ecole Normale Suprieure des Tlcommunications TSI Dept. with causal kernel Y#(t) = ~ 1 tZ1 for the fractional integral o p e r a t o r I # o f + order ~.S e m i .~ d : = Y l . f r . avenue du Colonel Roche F31077 Toulouse Cedex 4. upstlse. t s i . F31068 Toulouse Cedex 4. from Centre de Math~matiques at l~cole Polytechnique for helpful discussions and useful references. Denis M a t i g n o n 2.l i n e a r d y n a m i c s / and a s t a t e x E ~ (for simplicity sake): dax(t) = / ( x ( t ) ) + u(t). and G 6 r a r d M o n t s e n y 3 1 MIP / CNRS Universit$ Paul Sabatier 118. rue Barrault F75634 Paris Cedex 13. The problem of asymptotic internal stability is analyzed by a more involved functional analytic method. France m a t i g n o n @ @t s i . based upon known necessary and sufficient stability conditions for linear fractional differential equations.
using Y~ * Y l . (2) can also be written x(t) .I ~ [f(x(t)) + u(t)].74 J. Montseny Problem (1) can be advantageously reformulated in the equivalent A b e l Volterra equation: x(t) = x0 t. since the fractional derivative is intrinsically a nonlocal pseudodifferential operator. 2 Diffusive formulations In subsection 2. on an infinitedimensional statespace endowed with an appropriate Hilbert structure. system (1) is transformed into a diagonal infinitedimensional system with an extra variable ~ > 0. D. Audotmet. 9 in section 4. 9 in section 3.2]) proves useful. with the help of the new variable z(t) = x(t) . it requires LaSalle invariance principle. 9 finally in section 5 we will indicate some natural extensions of the resuits. in particular stability properties are more easily examined in this context. Matignon. T h e heat .x0. (2) can also be written as: z(t) = I '~ [f(z(t)) + u(t)] . the problem is examined with null initial condition z0 = 0. as: (2) Then. n a m e l y a first order in time diffusion equation. or t h a t seem to be within reach but still need to be fully developped. in so far as the problem can first be reformulated into one (or m a n y equivalent) way(s) t h a t is classical.1. using diffusive representations of pseudodifferential operators (see [8.u(t) I.a = I/1 the Heaviside unit step.6. LaSalle invariance principle) can therefore be used. trying to use geometrical or s t a n d a r d analytical techniques (such as those used in the integer case. On the contrary. trying to extend them to the fractional differential case is of little help. and G. (3) alternatively. and of the new function Y(z) = f ( z + z0). i. the problem is reformulated in equivalent ways with m a n y advantages for the analysis. and gives strong stability of the internal state. unfortunately. t). a main comparison result is established. the problem of the initial condition alone is addressed: it requires more specific analytical tools pertaining to the properties of the heat equation.xo Y t . for the main reason that quadratic forms prove hard to fractionally differentiate. and a state ~b(~. (4) As far as stability is concerned.e. other diffusive pseudodifferential operators t h a t are dissipative: p > 0).I ~ [f(z(t)) I. either straightforward (z E C or z E C ~ . see [11]).a ( t ) ] . the use of which will be sketched as closely as possible. The paper is organized as follows: 9 in section 2. Quite standard energy m e t h o d s (Lyapunov functionals.
'a ~ 7/a ~ k'a ~ with continuous and dense injections.'a = L0+~)~. . r t) d~.~~ = L(t+{)_~ ~o (]R+). problem (4) is equivalent to (see [10]): Output 0..a . we get: o.. • = L~t+()(~+ ) and V ~ = L~l+~)_t ( ~ + ) .~(~)lr (7) 2 T h e functional spaces to be used are: 7/~ = L~o(IR+). T h e y are independent of a. Balanced f o r m Let us denote ua(~) = ~ .t) in subsection 2. sin a~ T h e energy associated to this equation is: E~(t) = ~ lfo /J~(~) Ir (6) for which it is easily proved that the following equality holds: dEa . t h a t is: Pa(~) = C.2. and 1. .(~) [B4t)) +u(t)] .[oo j[ (t) = (O.r =~r z(t) = // r lla(() r t) d(. = o ~ > olsa) (Sb) // ~(~) r Ir T h e energy associated to this equation is: E(t) = ~ 'fo (9) T h e functional spaces to be used are: 7/ = L2(/t~+).. = o ~>0. which is meaningful thanks to p > 0 only. (5a) (5b) where /~a stands for the diffusive representation of the fractional integral operator I a.1 Diagonal diffusive formulations f o r m In the scalar case.r = ~r 40 = +. 2. then perform the inverse Fourier t r a n s f o r m in the space of t e m p e r e d distributions. a heat equation is then obtained with an extra space variable y and a state ~o(y. .( .Semilinear Diffusive Representations 75 equation formulation can be recovered as follows: first let ~ = 4~2r/2 with I/E ]~. then by a straightforward change on r and a slight abuse of notations. l. ~ (I~ +) 2 and k'.
Note that these equivalent reformulations are interesting results on their own. t) dy = < N a .t) dy = < Mo. T h e y are independent of a.( ] . 9 a natural energy functional E helps prove that the system is on the nonlinearity f . V . 7/. ~o(t) > .t) dy have to be understood in the sense of duality brackets < Mo. Montseny 2. ~o(t) > . and G. with Fourier transform too(7/) = 2 sin(aTr)12~~ _ 2 _ ~ for the o u t p u t form. v ~ Output f o r m System (ha)(5b) is equivalent to: Ot~o(y. . (13) T h e functional spaces to be used are: 7/ = L2(~).I ( ] ~ ) . for 89 < a < 1 both Mo(y) oc [y12(1~ and No(y) oc l y l .H I ( ~ ) and 12~ = H . with Fourier transform no(~/) = V/too(7/) for the balanced form. Jl B a l a n c e d f o r m System (8a)(8b) is (10b) equivalent to: ~ 0. Remark 1. ~o(t) > v o ' . and No(y). T h e energy associated to this equation is: E(t) = ~ I~(y.~ are regular Lloe functions. (lla) (llb) z(t) = ~ Na(y) ~(y. D. for c~ = 89 they are proportional to the Dirac measure J. in which easily obtained. and for 0 < c~ < 89they are distributions of order 1 involving only finite parts: hence. (12) for which it is easily proved that the following equality holds: ~(t) =  IOu~(y. for the following reasons: 9 the system is local in time.76 J.2 Heat equation formulations Now. tempered distributions will be used: Mo(y).tll2dy+z(t)f(z(t))+z(t)u(t). It is clear that. which dissipative under some specific conditions analytic framework is being used. ~o(t) > = < 1.t)12dy. Y') functional regularity results can be more is provided on an energy space 7/.t) z(t) = [ Mo(y)~o(y. Matignon. Audounet. integral terms such as f~ Mo (y) ~a(y. 9 a classical (Y.
112 + Ilay~ll 2 = / ( t .e. let us consider ~01.e. together with 91 . ~(t) >) < N~./ ( t . < N o . especially for comparison results.f(t. using standard schemes of numerical analysis (see [10]).11~.[ / ( t . < N~. These features can not be captured on the original system (1) nor on any of the AbelVolterra forms (2)(4). ~ >)] < N ~ . ~. ~oj >. 7"]. r Then. Multiplying this equation by ~ _ (where 4 i = ~ + .< al . and zl > z2 on [0.~ _ and ~ + ~ _ = 0) and integrating over IR leads to: ~  AII~II ~ + II0~II 2 = > . z l ) . the following quadratic a priori estimate will be useful in the sequel: ~. thanks to / strictly decreasing. ~.f ( t . then ~ox > ~o~ and zl > z~ on [0. ~j >)  No + g j (14) such that t ~+ zj(t) = < N~. < N~.O ~ = [f(t.) is strictly decreasing on IR. we get: 10dl~IP+ II0~II 2 < 0+0 Then function [[~_[[ is positive decreasing. . Proof (Sketch of).Semilinear Diffusive Representations 77 9 on the heat equation formulations.(t) > The following theorem is an extension to the case ~ ~ 1 of a result of [1]. 2. ~0 = 0. Function ~ . < N~. ~ol >) . T h e o r e m 1. Hence. the (weak or strong) m a x i m u m principle can be used. < Nc. < No. z2)] (zl . It follows that ~vl > ~v~ a. ~oj(t) > = < n~. r > [/(t.~1 .3 A comparison result On formulation ( l l a ) . with zj = < No. ~'~(t) > be o/class C 1 on [0.z ~ ) _ < 0.(t) > + < g(t).~2 is the solution of: O t ~ . ~o2 solutions of: Ot~oj . ~02 > ) ]  .. with initial value 0.0. .( l l b ) with an extra forced t e r m denoted by g(t). thus null a. T]..9 2 > 0. 9 numerical a p p r o x i m a t i o n of diagonal diffusive formulations is straightforward. T].cg~oj = f ( t . [] . Suppose / ( t . T]. I/a1 > g~ on [0. ~1 >) .as.
and G. which is nothing but a diffusive representation for the nonlinear fractional differential inclusion:  . by a simple extension to the complexvalued case. to this end. / ~ ( t ) < 0 thanks to z f ( x ) < O.. Note that the proof needs to be performed on one of the four equivalent diffusive formulations only.g. T h e o r e m 2. r ~ 0 in 7/~ strongly as t ~ ~ . one easily gets ~e(A) < 0. mostly because of the longmemory behaviour.a. 4. and the following differential inclusion: at~(y. In the case when f is linear. 5.0) = 0. Montseny Analysis of the case Zo . . system (5a)(5b) is dissipative: from ( 7 ) .. [] Remark 2.0 First. which happens to be sufficient but not necessary. 3. the trajectory {r is precompact (see [9]). the condition on f reads A < 0. The goal is to apply LaSalle invariance principle (see e. ~(t) > . we proceed in six steps: 1. the corollary of which is the stability of system (1) subject to specific conditions.a ~ ( y . Function f can be discontinuous at 0.78 3 J. (15a) (15b) z(t) = < Na. [3]). Proof (Sketch of). Matignon.. 4 A n a l y s i s o f t h e c a s e Xo :~ 0 This seems to be a more difficult problem than the previous one. we get: IIr + 0 and x(t) 4 0 as t ~ oo. and provided f is strictly decreasing with x f ( x ) < O. Audounet. x ( t ) = < 1. t ) [ f ( z ( t ) ) + u(t)]  N~ 9 0.6]).vorOastrcx~. in which case the results are still valid. ~(. for any r 9 7/a. As soon as the input u has stopped.r >va. Remark 3. that is f ( x ) = Xx. moreover Ea = 0 if and only if r = 0 p a .e. = (16) Once the wellposedness nature of problem (15a)(15b) has been established.u(t) 9 0 . since the optimal stability result for the linear case is: I arg(m)l > a ~ (see [5. r ~ 0 in 13a weakly.t) . hence. 6. we get a main theorem. 2. though care must be taken that f is a multivalued function with f(0) 9 0. D. the solution x = x0 + z of (16) is uniquely determined as an output.
T]. . the pseudoinitial c o n d i t i o n z0 in (1) can be t a k e n into account by an e x t r a input v i n s t e a d o f u in any o f the equivalent diffusive formulations of section 2 with f . l i m z < .2 Formulation by a change of function and variable F r o m reformulation (4). i the case x0 < 0 is treated similarly. n a m e l y : v(t) = xo l~~~t+ ~ + u(t). ~g is a solution of: Ot g  = O jr(z) + g(t). [] L e m m a 3. .~(.Semilinear Diffusive Representations 79 4. and zoo = . we get ~g(.x 0 Proof. ~ et z are increasing functions of t. Otherwise. which c o n t r a d i c t s ~ increasing a n d 90 = 0.~(. a n d can therefore be i n t e r p r e t e d as a forced term. L e m m a 1.t)>. by continuity. from w h a t we d e d u c e t h a t : 9 either ~ = cte. because ~v is increasing.( l l b ) . 9 or ~ ~ ete. t ) . 3t0 such t h a t z(to) = . W i t h #(t) = . hence z(t) is constant.~.T ) < ~ ( . There exists a unique equilibrium state 9oo = ere. then / ( 0 ) = f( 0) > 0. Proof. 2 Here g has been computed in such a way as to delay the start of the diffusion process by T. for it is clear t h a t the e x t r a i n p u t will never stop: the everlasting b e h a v i o u r of the e x t r a i n p u t comes from the h e r e d i t a r y aspect of the p r o b l e m .  f ( 0 ) l[0..1 Formulation through an extra forced term F r o m r e f o r m u l a t i o n (3). we cannot expect to use the s t a b i l i t y result above ( t h e o r e m 2). ~ hence. . Suppose u = 0 from t = 0 on (the extension to u = 0 from t = to will be addressed at the end of the section).Tl.x o ::~ f ( z ( t o ) ) = O.z0 a n d a change o f function f ( z ) = f ( z 0 + z). t h a t is an e q u i l i b r i u m s t a t e .t) = ~ ( . the pseudoinitial c o n d i t i o n z0 in (1) can be taken into account by a change of variable z = z . thus as N~ > 0. in which case the c o n c a v i t y is of c o n s t a n t sign a n d negative.tT)><<N~. t h a n k s to the c o m p a r i s o n result (theorem 1) with g(t) < O. t . <N. = 0. 4. i m p l y i n g ~o is c o n s t a n t Vt > to.. [3 L e m m a 2.N . Let z0 < 01. .. we then use the h e a t e q u a t i o n f o r m u l a t i o n in b a l a n c e d form ( l l a ) . Unfortunately.gO. it is identically zero on [0.
and G. in the following sense: ~o r ~oo uniformly on any compact subset. z 4 x0.~o > > contradictory. C o r o l l a r y 2. [] Lemma 4. which is c o ~ r a d i c t o r y with ~ increasing and ~0 = 0.x 0 as t + oo. f ( z ~ ) = 0 necessarily and zoo = . Montseny Proof. Then. more precisely: VK. it is the stability/unstability generated by zo which will play the m a j o r role and enable to conclude in a similar way. Y. the m a x i m u m principle forces this limit to exist without degenerating. but ~(t) diverges in the energy space L~(IR) (weak* convergence in L~176 When u = 0 for t > to. .~v(. these techniques provide not only straightforward numerical schemes for the approximation. Thus. z d 0 as t + cx~. Matignon. Proof. the state ~o is initialized by ~o0(y) = ~o(t0.x 0 . but < Na. Otto > c9~o + N~.  k. D. [] ~o(y) >_ K for y E [Y.x 0 for f is injective.~oo > is not properly defined.. If z* = l i m z < . except if ~oo is constant.~dy > z* for K large enough. Y]. because Na and ~oo do not belong to dual spaces (except in the case a = 89 To some extent. From which we can easily deduce that ~o + +oo. O~co = No. KfY_yN. :lto. Na being positive. Remark ~. hence z = < Na. heat equation formulations) can not be overcome. limz exists and < . It can be shown that the equilibrium state ~ooo is asymptotically reached. At the equilibrium. then ]'(z) > k > O. it gives valuable information thanks to the evolution of an internal state of infinite dimension. but also very sharp estimates for the analysis of the problem (and especially for asymptotic analysis). from which the longmemory behaviour stems: this aspect is rather well controlled (from a functional analytic point of view) thanks to the specific properties of the heat equation. It is noteworthy that < Na. which implies that on any compact subset [Y.t) > tends to . Audounet. y) at time to. the concavity of ~oo is of constant sign and negative.80 J. This is certainly one of the most remarkable applications of D R of PDOs. topology of L~176 The previous analysis amounts to the maximum principle for the case a r 7" x The formulation by a heat equation (namely.z o .  f(Zoo). that is in the weak. with null input but x0 = x(t0) ~ 0 in general. Y]. which is C o r o l l a r y 1. as the autonomous dynamics generated by ~o0 # 0 is stable (diffusion).
Montseny.(t) > . and J. .M. more precisely: OtW= 0v2~ + )7(< N~. Care must be taken that the equilibrium state ~ does not belong to the energy space: specific methods from functional analysis and semilinear diffusion PDEs must be investigated in order to tackle the problem properly. J. the whole set of results obtained thanks to diffusive formulations can be extended to any other diffusive pseudodifferential operator of dissipative nature. namely: 9 x" = 0 is locally asymptotically stable i/I arg(A)l > a~. 90 = 0. a more accurate result can be conjectured. Lebanese American University. Diffusive representations of fractional and pseudodifferential operators. Physica D.~)l system is asymptotically oscillating. as a fractional version of the H a r t m a n . Matignon. V. Audotmet. The local stability of the equilibrium x* = 0 of the nonlinear fractional differential system dax = f ( x ) is governed by the global stability of the linearized system near the equilibrium dax = Ax. March 2000. Lebanon. Finally. = a~. Moreover. 2. Roquejoffre. ~. Note that nothing can be said if I arg(. that is p > 0. F = L + B with L = O~+l the linear part and B a nonlinear term of lower differential order. in order to extend the sufficient stability condition.Semilinear Diffusive Representations 81 5 Further extensions The conditions can easily be extended to the complexvMued case. and the monotonicity of f must be translated in an appropriate way. namely ~e(x" f ( x ) ) < 0. (1998). and G. pp. ~o >)  x(t) = x0+ < N. and then an infinitedimensional version of the HartmanGrobman theorem.~. 10 pages. Giovangigli. Beirut. A threshold phenomenon in the propagation ol a pointsource initiated flame. the solution and stability of (17a)(17b) is known exactly when F reduces to L. Na. Audounet. namely: T h e o r e m 3 ( C o n j e c t u r e ) . (17a) (17b) is of the form Ot~ . in Research Trends in Science and Technology. and also to the vectorvalued case. References 1. as ~e(x H f(x)) < 0. The idea is to use a semilinear diffusive reformulation of the system. 295316. D. J.G r o b m a n theorem.F(~) with F linearizable in a weak sense (unbounded operators). in which case the linearized 9 x* = 0 is not locally stable/f larg(A)l < a ~ . where A = if(O) E C.
and D. LAAS. J. Palis and W. Diffusive representation o/ pseudodifferential operators. D. Londen. f r / M a t h s / P r o c / V o l . Haraux. Temam. Journal. emath. 68 of Applied Mathematical Sciences. Syst~mes dynamiques dissipatifs et applications. Matignon. Infinite dimensional dynamical systems in mechanics and physics. Montseny 3. de Melo. G.. . 8. O. July 1996. IEEESMC. Montseny. in Nonlinear Control in the Year 2000. 2000. 1991.82 J. SMAI. and G. 6. Montseny. timeURL: 9. pp. fr/Maths/Proc/Vol. 5 of ESAIM: Proceedings. Springer Verlag. 4. vol. 18 (1971). G. URL: h t t p : / / w u v . Fractional Differential Systems: models. Audounet. in Computational Engineering in Systems Applications. MONTSENY. Ft. 12. 11. 17 of Recherche en Mathdmatiques Appliqudes. . J. Masson. eds. Stability results for fractional differential equations with applications to control processing. . S. vol. vol. 159175. MATIGNON AND G. Stability properties for generalized fractional differential systems. Internal Report. 2. 5/. pp. June 2000. http://www. D. fr/Maths/Proc/Vol. 96. emath. France. ESAIM: Proceedings. URL: http ://www. ~ . 321330. Pr~compacit~ des trajectoires de syst~mes pseudodiff~rentiels dissipatifs sous formulation diffusive. methods and applications. vol. Matignon. 10. Springer Verla8. 5 (1998).3968. 5 (1998). December 1998. 145158.. Matignon. Lille. 5/. ESAIM: Proceedings. Audounet. D. pp. 1988. pp. 7. Geometric theory o/dynamical systems: an introduction. 5/. Springer Verlag. The qualitative behavior of the solution o/nonlinear Volterra equations. 1982. A. Michigan Math. 5.. Diffusive representation/or pseudodifferentially damped nonlinear systems. emath.
1 Introduction In this paper we study controllability properties of linear control systems on Lie groups. cl ]nstituto de Matem~itica Universidade Estadual de Campinas Cx. the group of a u t m o r p h i s m of G. and X stand for an infinitesimal a u t o m o r p h i s m of G.e. If G is simple connected then it is well known that the Lie algebra aut (G) o f A u t (G) is Der (g). Postal 6065 13081970 Campinas SP. Yi (x(t) ) (1) Here z(t) E G. First of all. Folio n ~ 97001 Research partially supported by CNPq grant n* 301060//940 . the semidirect product Af (G) = A u t (G) • G. and Luiz A. considered as right invariant vector fields on G. Chile vayala@socompa. aut (G) is a Partially supported by Proyecto FONDECYTCON1CYT n ~ 1990360 and Programa de Cooperacin Cientfica lnternacional CNPqCONICYT. B. Brasil A b s t r a c t . We denote by Af (G) the a]flne group of G i. we show that for a connected Lie group G. Wealso show that the forward orbit from the identity is not in general a semigroup. This class of control systems was introduced by Markus [3] and also studied by Ayala and Tirao [1]. The control vectors Yi are elements of the Lie algebra g of G. the flow (Xt)te~ is a 1parameter group of A u t ( G ) . i. On the other hand.e. Linear control systems on Lie groups were introduced by Markus [3] and also studied by Ayala and Tirao in [1]... ucn.. A linear control system on a connected Lie group G is determined by the family of differential equations on G.Controllability P r o p e r t i e s of a Class o f Control S y s t e m s on Lie Groups Victor Ayala 1. For this class of control systems we establish controllability results in the compact case and also in the semisimple noncompact case. we discuss the group of diffeomorphism generated by the system. parametrized by the class U of the piecewise constant addmisible control: ]c(t) = X (x(t) ) + ~ i=1 u.(t). San Martin s** 1 Departamento de Matems Universidad Catdlica del Norte Casilla 1280 Antofagasta.
. Then. the linear system (1) is controllable if and only if it satisfies the Lie algebra rank condition. And the (1) system itself is induced on G by the invariant system on G27.2. This amounts to suppose that G has codimension one in the systems group G~:. According to the following cases.. In a compact and connected Lie group G. we prove: 1. induced by the decomposition of the Lie algebra ~. The vector fields of the system are then right invariant vector fields in GE... they proved that the Lie algebra rank condition characterize controllability for abelian Lie groups. We consider the global controllability property for a linear control system on a connected Lie group. where 3 is the onedimensionM center of ~E and Z = exp (3)..v = ~ @ [I. this example shows that the forward orbit of (1) from the identity is not in general a semigroup. II~). 2. On the other hand. Section 3 contains the main results about controllability: the compact case and the semisimple noncompact case. In fact.v. Section 4 contains one example showing a system which is locally controllable from the identity but not controllable from the same point on SI (2. . In [1] the authors extend the well known Kalman rank condition for a linear control systems on ~n. We assume that the system satisfies the Lie algebra rank condition. By construction G identifies with a homogeneous space of Gz. and by G~: the connected subgroup of Af (G) whose Lie algebra is 9. the invariant system on G~ defines through ~r an induced invariant control system on G.. If G is a semisimple noncompact Lie group denote by rr : G. and the system (1) is controllable if and only if Sz is transitive in this homogenous space. The original system lifts to a right invariant control system on G~. We denote by gz the subalgebra of af (G) generated by the vector fields of the system.Of course G z is exactly the group of diffeomorphisms generated by the control system. that is the systems group. To get the last result we use the notion of left reversible semigroup and the following theorem proved in [5]: If L is semisimple noncompact group and T is a semigroup with nonvoid interior then T is not left reversible unless T = L. The induced invariant system on G is controllable if the linear system (1) is controllable from the identity. whose forward orbit from the identity is a semigroup which we denote by S.j_> 1) C g is the dimension of G. For a connected Lie group G they also prove local controllability from the identity element of G if the dimension of the subspace v = Span i = 1.v + G the canonical homomorphism onto G~/Z. In particular.v. This paper is organized as tbllows: In Section 2 we study the group of diffeomorphisms generated by the linear control system 1..84 Ayala and San Martin subalgebra of Der (~).
x ) . z) . it follows that any derivation D E Der (9) induces a vector field. D i X 2 . that is. In particular any D 9 aut (G) induces a vector field. on G.Controllability on Lie Groups 2 Group of the system 85 In this section we discuss the group of diffeomorphisms generated the vector fields of the system (1). and Aut (g) the automorphism group of g. on G.D~X1 + [X1. On the other hand if G is connected then G = (~/F. z) . hence the Lie algebra of Af (G). X2]) where inside the first coordinate the bracket is that of Der (g). In fact. the Lie algebra of derivations of g. ( g . suppose that G is a Lie group and denote by g its Lie algebra. it follows that this map is indeed an isomorphism of groups.where 1 is the identity automorphism embeds G into Af (G) as a normal subgroup. By the affine group Af (G) of a group G we understand the semidirect product Af (G) = Aut (G) • G. whose flow is exp ~tD) 9 Aut (G). and since the istropy at the identity is Aut (G). In case G is connected and simplyconnected the automorphism groups Aut (G) and Aut (g) are isomorphic. So that Aut (G) is the subgroup of Ant (G) leaving F invariant.Moreover. xg (y)). The multiplication in Af (G) is given by (g. where G is the simply connected covering of G and F C G is a discrete central subgroup. Of course.From this description of the Lie algebras and the action of Aut (G) on G. a f (G) is the semidirect product aut ( G) x ~ 9. also denoted by D. X~)] = ([O1. the group G becomes identified with the homogenous space Af ( G ) / A u t (G) of Af (G). s a y / ) . and since F is discrete the Lie algebra of Aut (G) is the subalgebra aut (G) of Der (9) given by aut (G) = {D 9 Der ({~) : b ( z ) = 0 for all z 9 F}. and that the mapping x ~~ (1. y ~~ zg (y). Therefore for a simply connected group G the Lie algebra of Aut (G) is Der (g) . Recall that Aut (g) is a Lie group whose Lie algebra is Der (g). (02. D2]. while in the second that of 9. where Aut (G) stands for the automorphism group of G. which we denote by a f (G). Since any g E Aut (g) extends to an automorphism of G. X1). if G is simplyconnected then aut (G) = Der (9). (h. We start with some brief comments regarding the automorphisms of a group: intod. Now. the isomorphism is given by the mapping which assigns to g E Aut (G) its differential dgl at the identity. X is a vector field on G such that its . is the semidirect product Der (g) • g.The Lie bracket in this Lie algebra is given by  [(O1. y) = (9h. Through this embbeding left multiplication of elements of Aut (G) defines the action or Af (G) in G. In first place let X be an infinitesimal automorphism of the Lie group G. This action is transitive. We turn now to the control system (1).
x).v. so that the system itself lifts to a right invariant control system in the Lie group Af (G). . where H~ is the connected subgroup of G whose Lie algebra is Ijs and exp (~D) is the oneparameter subgroup generated by D. Suppose that [B. We denote by ttE the subalgebra of a f (G) generated by the vector fields of the system. Clearly (~D) xs I~E is a subalgebra of a.X] of m. Take B E l \ m and consider the derivation D ( X ) = [B. which implies that X = b for some derivation D E aut (G). Then ( 1 / x ) A + C so that [ B . An easy inspection shows that 9E .Of course GE is exactly the group of diffeomorphisms generated by the control system.[C. The vector fields of the system are then right invariant vector fields in GE. i. P r o p o s i t i o n 21 Let [ be a Lie algebra and m C [ be an ideal of codimension one.( ~ D ) xs [~E.C) for some x r 0 and C E m. . y = x (exp (tD) (y)). X] for some C e m and all X e m . In this case g is an ideal of codimension one in gE. x l ( e x p ( t D ) x l ) ) . and let GE stand for the connected subgroup of Af (G) whose Lie algebra is gE. X] . showing the derivation is inner. Proof. let the drift vector field of (1) be given by X . In order to simplify matters we shall consider as the systems group the group G~: = ~ x ~ HE with multiplication (tl.C is the required vector complementing m. X ] = IV. if A satisfies the given conditions then A = x (B .x2) = (tl + t ~ . . (2) (3) This group acts on G by (t. .m] = 0 and l = ~ A O m is that the derivation D is inner. We conclude this section proving a simple criteria for the existence of an one dimensional ideal complementing DE. Therefore the vector fields of the system are elements of a f (G).b where D E aut (G).Xl)" (t~. . X] for all X e m. and if G is connected it identifies with a homogenous space of G. Now. i . Then X t is an oneparameter subgroup of Aut (G). In order to state it we note that the restriction of D to I}E is a derivation of this Lie algebra.f(G). and the (1) system itself is induced on G by the invariant system on G.The connected subgroup of Aut (G) with Lie algebra ilz is the semidirect product exp (~D) • HE. A necessary and suj]icient condition f o r the existence of A E I such that [A. and denote by OE the smallest subalgebra of it invariant under D and containing the control vectors Yi.86 Ayala and San Martin flow X t is an automorphism for all t. Then A = B . m. D = ad (C) f o r some C E m .1 . that is the systems group.e.~.. Reciprocally. In general exp (~D) may be diffeomorphic to ~ or to S 1. These descriptions of BE and GE imply at once that (1) satisfies the Lie algebra rank condition if and only if I~E = g.
pick (x. we observe that the group G. This implies that. Therefore. then D (X) . The original system lifts to a right invariant control system in G. let D be a derivation of [~. Since the connected subgroup of G having Lie algebra 3 is compact. Therefore the extension of exp (tD) to Z x K fixes every element o f / " so that D induces an infinitesimal automorphism of G. where Z is abelian simplyconnected (isomorphic to the additive group of 3) and K is compact with Lie algebra ~. Now. In particular the points of Z N/~ are invariant under this group.v is generated by left translations of G and inner automorphisms.Z x K.v. hence spans the vector space 3. which in turn is isomorphic to ~.Controllability on Lie Groups 3 Controllability 87 In this section we derive some results about the controllability of the linear system (1) on a Lie group G.v is dense in S. Now. because every derivation of [~ is inner. Then y belongs to the center Z (K) of K. exp(tO)(y)) for (x. The automorphisms exp (tD) can be extended to Z x Is by putting e x p ( t D ) ( x . so that F is a lattice in Z. Concerning the controllability of a linear control system in G. it follows that exp (tD) (y) . that is.Therefore D must be identically zero on 3.0 for all X E /'. Then D (3) C 3. From now on we assume that G is connected and the system satisfies the Lie algebra rank condition. automorphisms of the form y ~+ xyx 1 . whose forward orbit from the identity is a semigroup which we denote by S. Now. the Haar measure .~. 3. By contruction G identifies with a homogeneous space of G r . In order to discuss the Lie algebra out (G) we must consider the universal covering group G of G.v. y ) = (x. But /~ is a lattice in 3.v.y) E Z x K.y for all t E ]~. it follows that Z / ( Z N F) is a compact subgroup. Let F be the central subgroup such that G = G / F . We consider the controllability property according to the different types of groups. Since exp (tD) (y) E Z (K) and this center is discrete. This is the only simplyconnected group having Lie algebra g. Summarizing. and the system (1) is controllable if and only if S r is transitive in this homogenous space. if we identify Z with 3.In order that D induces an infinitesimal automorphism of G it is necessary that exp (tD) fixes every point of F. From the Lie algebra rank condition in G r we know that the interior of S. let D be a derivation of 9.1 Compact groups If G is compact then its Lie algebra decomposes as g = 3 ~ ~ where 3 is the center of 9 and ~ is a compact semisimple Lie algebra. we have that aut (G) is isomorphic to Der (~).This reduces the Lie algebra aut (G) to a subalgebra of Der ([~).y) E F. This amounts to suppose that G has codimension one in the systems group G. it is the direct product G ". On the other hand.
its closure cl (Z) is a connected Lie subgroup contained in the center of G. if G is a semisimple Lie group a derivation D of its Lie algebra i] induces in G an infinitesimal automorphism. assume that G is connected. So that when we write G. which is isomorphic to g.B ) where D = ad (B). To see this it is enough to check that G A Z = {1}. where D is the derivation defining the drift vector field. This group acts on G by the semidirect product formula.v. /. 3.t B ) ) .From this expression it follows at once that Z f3 G = {1}.v = ~ xs G. Consider the subgroup Z = exp (3)./Z.Since A is in the center of g$ it follows that exptA = (exptD) 9 (exp ( . Clearly the Lie algebra of G. . Therefore. which is constructed as follows: By Proposition 21 we have that the Lie algebra gE associated with the system is a direct sum gE=3@g where 3 is the onedimensional center of g. Hence its Lie algebra is contained in the center of g.~. so that cl (Z) = Z. proving that G ~ G~. For this intersection we recall from Proposition 21 that a generator A of 3 is given by ( D . and consider a linear system which satisfies the Lie algebra rank condition.v + G canonical homomorphism onto Gy. We claim that G•/Z is isomorphic to G. Now. Appart from this action there is another one which comes from a homomorphism of G z onto G. exp ( . so that the flows of the vector fields of the system are measure preserving. B e i~. .v/Z. We can then perform the coset space G. According to the previous discussion the group of the system is G. which is a Lie group because Z is a normal subroup. we get exp (tA) = (t. Thus we have proved the T h e o r e m 31 In a compact and connected Lie group G. and thus in 3. In fact.2 Semisimple noncompact groups The center of any semisimple group is a discrete subgroup implying that the elements of the center are fixed under a oneparameter group of automorphisms./Z as claimed. Therefore the Lie algebra of cl (Z) is 3. so that the algebra aut (G) is just Der (g). Since the Haar measure is finite it follows that the system is controllable if and only if it satisfies the Lie algebra rank condition (see Lobry [2] and San Martin [4]). a linear system is controllable if and only if it satisfies the Lie algebra rank condition. It is a closed subgroup of GE.~/Z is 9.t B ) ) .88 Ayala and San Martin in G is biinvariant. We denote by 7r : G.v = ~ xs G with the multiplication given by (3).
and T O H x is a left reversible semigroup in H r . there exists gl E T such that glx = hx. Yi (x) x E G. Then T is left reversible in L. and finally.l g l E Hx. given by = 7r (D) (x) + ~ i=1 u. On the other hand we can show that the semigroup of certain controllable systems are reversible: P r o p o s i t i o n 33 Suppose that L is a group acting transitively on a space M .D (x) + ~ i1 uiY/(x) x E Gz. (5) where /) is the infinitesimal automorphism of G induced by D. Let T be a subsemigroup o f L and assume that T x = M . Proof. Now. the left reversibility of T n H~: ensures the existence of g2 E T n H . T h e o r e m 34 The invariant system (6) is controllable if the linear system (5) is controllable from the identity. We are required to show that h E T T 1. First the invariant system on G~: /~ . Joining together these two facts about reversible semigroups we obtain at once a necessary condition for the controllability of (5).. For this we recall first the following fact: A subsemigroup T of a group L is said to be left reversible in L if T T 1 = L.lg E T concluding the proof. so that h . and denote by Hr the isotropy group at x E M . such that h . . P r o p o s i t i o n 32 I l L is semisimple noncompact and T has nonvoid interior then T is not left reversible unless T = L.Controllability on Lie Groups 89 Thus we are laid to three control systems. Take h E L. or equivalently.l g l g 2 E T n H~. that there exists g E T such that g . The following statement was proved in [5]. The linear system m = b (+) + i=l (+) + e v. (4) where D is viewed as a right invariant vector field in G~:. Since T x = M .l h E T 1. the induced invariant control system on G. Taking g = gig2 E T it follows that h . (6) Our next objective is to prove that controllability of the linear system (5) from the identity implies the controllability of the invariant system (6).
taking into account the above description of S t . In the next section we show that if G is nilpotent then the forward orbit from the identiy is a semigroup only in the trivial case of controllability. ~r (SE) = G. This implies that ~r (Sz) is left reversible in G. Hence. Then. suppose we can prove the converse of Proposition 34. we would have also that (5) is controllable to the identity. and hence that system is controllable from every point. from the identity for the invariant system in G. and X E 9. With this equality in mind we can describe S. This seems to be a very strong property to get solely from the controllability of the invariant system (6). However. The semigroup of the system contains the oneparameter semigroup {exp (tD) : t > 0}.Taking exponentials we have exp (tD) = exp (tA)exp (tX). showing that the invariant system (6) is controllable.e be the projection into the first coordinate. even from the identity. We observe that the assumption in this proposition is controllability from a specific point. by reversing time. Since this semigroup is left reversible in the isotropy group of the identity. Hence by Proposition 32. at time exactly t.At (1) stands for the accessible set. Now. namely that ~r (Sz) = G implies that the linear system (5) is controllable from the identity. a necessary condition for S~: to be transitive is that its interior meets all isotropy subgroups. The following comments may shed some light at this problem: As discussed above D = A + X where A is a generator of the center of 9~:. Through the mapping (t. (1) where . (A) it follows that = U{t} t:>O x A. We end this section with the following remark about the algebraic properties of the forward orbit from the identy of a linear system in G. Outside the identity this property may not hold. We do not know whether controllability of the invariant projected system implies the controllability of (5). Let p : ~ x G . . it follows that controllability of (5) is equivalent to have exp (tX n) E int. This point is required in order that its isotropy subgroup meets SE in a left reversible semigroup.At (1) for all h E G and some t E ~ . g) ~~ (exp (tA). because 7r (Ss)~r ( S s ) 1 = ~r ( S ~ S ~ I ) .90 Ayala and San Martin Proof. the previous proposition together with the controllability from the identiy imply that S~ is left reversible in G ~ . It is a homomorphism of groups. g) the group G~: becomes isomorphic to direct product ~ x G. Since p.v as follows. The example below shows that in semisimple Lie groups the forward orbit is not in general a semigroup. For (5) the isotropy at a point h is {exp (tO n ) : t e ~}. (D) = p. namely the identity of G. Based on the examples of the classical linear system and of an invariant system it is natural to ask if that forward orbit is a semigroup of G. where D n = A + X h and X h = Ad (h) X.
This bilinear system is not controllable because both matrices are symmetric. Consider the derivation D = ad (X). An easy computation of brackets shows that {Y. C. Its Lie algebra is the subspace ~l (2. The group G ~ is isomorphic to IR x Sl (2. viewed as an element of g~: = gl (2. becomes the matrix D=I+X= 0 " Thus $I: is the semigroup generated by { e x p t ( 2+uu uU) :t'uEi~}" The projection rr (S~:) of Ss into SI ( 2 . ~). [X. Ayala. where and form the linear system x . Y].1R). Hence we can normalize the isomorphism so that D.Controllability on Lie Groups 91 4 Example Let G be the simple group SI (2. 14 .1R) and hence to G1+ (2. SIAM J. Hence lr (3"~:) is proper so that the invariant system is not controllable. (1974) Controllability of Non Linear Systems on Compact Manifolds. J. [X. Proceedings of Symposia in Pure Mathematics (AMS) 64. References 1. 4764 2. [X. V.IR). which is the semigroup of the bilinear system x = Xx + uYx. IR). ~ ) of trace zero 2 • 2 matrices. the group of matrices with det > 0. where Y is the right invariant 1 11 . Control 12. IR). we have got a linear control system which is locally controllable from the identity but not controllable from the same point. Since S1 (2. IR) is spanned by the identity matrix. Y]]} span sl (2. (1999) Linear Control Systems on Lie Groups and Controllability. and Tirao. ~ ) is connected. Its Lie algebra is gl (2. The centre of gl (2. This implies that the linearized system at the identity is controllable. ~ ) is the semigroup generated by exp (X + uY). Lobry.D (x) vector field given by the matrix +uY (x). and hence that our linear system is locally controllable from 1 E G. it follows that the forward orbit from the identity is not a semigroup. Summarizing.
459462 5. San Martin. Lecture Notes in Mathematics 898. (1995) Semigroup actions on homogeneous spaces. B.A. Markus. A. Semigroup Forum 50.B. (1980) Controllability of multitrajectories on Lie groups. Proceedhags of Dynamical Systems and Turbulence. L. 250265 4.92 Ayala and San Martin 3. (1987) Controllability of families of measure preserving vector fields. A. L. 5988 . L. Warwick. and Tonelli. P. Systems & Control Letters 8. San Martin.
Portugal Phone: +351218419028 . research has been focused on control design synthesis . u t l . 1 Introduction T h e basic principle behind any feedback l i n e a r i z a t i o n control s t r a t e g y is to provide a s t a t e feedback control law to cancel the s y s t e m nonlinearities while s i m u l t a n e o u s l y i m p o s i n g some desired linear d y n a m i c s [1]. This paper is concerned with stability robustness of nonlinear multivariable systems under inputoutput feedback linearization.O. Over the last decade. plant uncertainty in the nonlinear (openloop) system causes the parameters of the resulting linear system to be uncertain. yielding an uncertainty description of the closedloop in polytopic form. j .Fax: +351218498097 . n l A b s t r a c t . GCAR Avenida Rovisco Pals 1049001 Lisboa. As feedback linearization aims for a linear dosedloop system. p t Delft University of Technology Faculty of Information Technology and Systems. thus allowing the construction of a polytopic uncertainty description. t u d e l : f t . which represents imprecision of p a r a m e t e r s within the model. vdboom@its. there is a d e m a n d for a feedback l i n e a r i z a t i o n control scheme t h a t can take p l a n t u n c e r t a i n t y into account. which represents u n m o d e l e d d y n a m i c s .Fax: +31152786679 .Stability Analysis to Parametric Uncertainty: Extension to the Multivariable Case Miguel A y a l a B o t t o i. P l a n t u n c e r t a i n t y can be generally classified into two different types: p a r a metric uncertainty. T h e success of this control scheme is strongly d e p e n d e n t on the exact m o d e l description o f the system under consideration. i s t . Ton van den B o o m 2. while r e t a i n i n g the closedloop p e r f o r m a n c e and stability. and Josd Ss d a C o s t a i 1 Technical University of Lisbon lnstituto Superior T~cnico Department of Mechanical Engineering. j . Due to the nonlinearity of the process under control.t . since t h e exact m o d e l can never be o b t a i n e d . and u n s t r u c t u r e d uncertainty. how. these uncertainties can be bounded within intervals. However. The Netherlands Phone: +31152784052 . Box 5031 2600 GA Delft. with a numerical procedure.taigbotto0deta. these closedloop uncertainties will turn out to be nonlinear and state dependent. It is outhned. A procedure is presented that allows plant uncertainty to be propagated through the control design. Stabihty robustness can then be verified with the aid of hnear matrix inequalities (LMIs). Control Laboratory P.
which can easily take these uncertainties into account. although in either cases an additional state feedback is required in order to guarantee stability robustness. Developing such a procedure will be the main topic of this paper. while the nominal system matrices A. In Section 2 the stability robustness analysis problem is formulated and its solution is pointed for a particular system. A particular configuration is adopted where the uncertainties of the system are captured in nonlinear and state dependent terms. to address the problem of robustness of feedback linearization under bounded parametric uncertainty. r e = p . or adaptive control [3]. In order to analyze the influence of parametric uncertainty on the stability properties of the closedloop system. In Section 5 some conclusions are drawn. it is necessary to transport the plant uncertainty through the control design. statistical properties of such an estimation will provide confidence intervals for the estimated parameters.5]. 2 Problem statement Consider the following state space description of a multivariable discretetime system: Xk+l (A + JA(xk))zk + ( B + J B ( z k ) ) V k Yk = Czk (i) where zk E X C ~n is the state vector. identical to the closedloop system which will result from the application of the feedback linearization scheme. One of those conditions is that the system is in the model set. where a state space uncertainty description of the closedloop system is derived. This description is linked with the initial system configuration previously presented in Section 2. which parameters are to be estimated according to a nonlinear optimization of a quadratic criterion. If certain conditions are met [4. and further presents a possible formulation of the feedback linearization control scheme. The system under control will be modeled with a nonlinear autoregressive with exogeneous input (NARX) model. JA(zk) and JB(zk). The result will be then used in Section 4. This paper is divided into 5 sections. In this paper it is assumed that plant uncertainty is only due to parametric uncertainty. which can be easily checked with tools described in [6. There has been some research done on the problem of robust stability to parametric uncertainties under feedback linearization.7]. and so enabling a straightforward application of the robust stability analysis tools. . vk E V C ~ m the input vector and Yk E Y C ~P the output vector of the system.94 Miguel Ayala Botto et al.e. B and C have appropriate dimensions and represent a stable linear system. Section 3 describes the multivariable NARX model assumptions by focusing in the characterization of its parametric uncertainty. It will be assumed throughout that the system is square.. i. based on sliding mode theory [2].
such that: A~"PA~ .AL) j=l (4) which means that for each xk and vk there exist parameters A1. L (7) . a sufficient condition for the assymptotic stability of the autonomous system is proven if a positive definite matrix P is found. without loss of generality.i (xa)[vk]j = Z AiA.t h column of JB*(xk). . if a m a x i m u m bound for each element of the uncertainty terms (fA(xa) and ~Bj(xk) is known in advance. expression (6) can be rewritten as a Linear Matrix Inequality (LMI) by using Q1 = p . i=1 (5) According to Lyapunov's stability theory.. L (6) Moreover. . . . As will be shown in the outcome of this paper. the original system (1) can be described as: xk +l ". In this way.. Co: P A + 5A(xk) + ~"~SBj(xa)[vk]j E C o ( A 1 . This procedure will provide the basic ingredients for a complete stability robustness analysis of system (3)..(A + tiA(xk ) + ~']~=l tf% (Xk )[Vk]j)xk + (B + ~B(O) )v~ yk . . under some mild assumptions a quantitative measure for these bounds can be found by an offline search procedure which spans xk over the relevant regions of the system operating range. since the uncertainty in matrix A can then be captured in a polytope defined according to the following convex hull. ~ L _~ 0 with ~"~L=IAi = 1. . resulting in: A T Q . . construct the [n x n] matrices J[~j(xk) for which the ith column is equal to the j . the following expansion can be given for 6B(xk): = i=l + B(0) where dB*(xk) is a [u xp] matrix. From (2). for every bounded input vk and for any arbitrary initial state x0. such that the dynamics of the system can be recasted as: p j1 L A + 8A(xk) + Z 8B. .1 A i _ Q1 < 0 Ql>0 Vi = 1 .Cxk (3) The stability of the system can then be easily checked.. .Stability Analysis to Parametric Uncertainty 95 If IlaB(0)ll < oo. . .P < 0 P>0 Vi = 1 .
O) where the regression vector is partitioned as follows: (10) ff_ (11) . from the assymptotic stability of the autonomous system.96 Miguel Ayala Botto et al.. 3. respectively.vk) (9) where Xk and x~ represent the state vector of the process and of the desired resulting linear closedloop system. consider the following NARX description of a given square MIMO plant: uk+ = y(=k. The success of feedback linearization strongly relies on the exact cancellation of the system nonlinear dynamics via feedback. together with the assumption of having a bounded input vk..L (8) which can be solved through computationally efficient algorithms [8]. taking into account the plant uncertainties in this control loop is the first step towards a robust feedback linearization control scheme.. which cancels the system nonlinearities while simultaneously imposing some desired closedloop inputoutput dynamics [1]. expression (7) is equivalent to the following LMI: AIQ >0 Vi=l.1 Model assumptions One way to account for plant uncertainty is to assume that all plant uncertainty is captured in the model parameters. according to: uk = ~P(xk. According to the Schur's complement condition. the feedback control law describes a nonlinear and state dependent dynamic mapping between the process inputs uk and the new external inputs vk. ~.uk.xtk.. Therefore. will follow that system (3) is also bounded input bounded output (BIBO) stable. Under these assumptions. In general terms. 3 Feedback linearization with parametric uncertainties Inputoutput feedback linearization control strategy aims at finding a state feedback control law. Moreover. which underlies the assumption that both the model and the plant are structurally identical.
uki. is smaller than the (1a)level of the normal distribution.y L n .2 Approximate feedback linearization Approximate feedback linearization assumes that the first order of Taylor's expansion of the nonlinear system (10). This paper assumes that bounded confidence regions J9 for each nominal model parameter. while Axk = x k .. (xk. 50)]z:~uk (17) with 5 F ( x k . 0./}.t (15) where the probability that 0 deviates from 0 more than 60. uk1. 0. 50)]Axk q{[E(xk1. JS) and 5E(xk_ i.Stability Analysis to Parametric Uncertainty 97 where Yk corresponds to the vector with the system output measurements at time instant k. 3.. uki. uki. uk1. uki. Uki represents the system inputs at time instant k .i . Following this procedure.i. J0) = Z / ~ i ( x k .i.] r while 0 E O C ~t represents the exact parameter set of the plant.u~_.i. is an accurate representation of the original nonlinear dynamics. where: Yk1 : [YkT1.1 and Auk =UkUk1...' . available in standard statistical tables..5 E(xk1. uki. uk. uk1. the following model will be used throughout as representing the nonlinear plant dynamics: Ayk+i = F(xk1. In order to have a model description which explicitly accounts for the parametric model uncertainties. resulting in the following expression: Ayk+i = [F(xki. O) and E ( x k . 0) "l. 9.x k . O. Uk_ i).i . O)Auk (16) with matrices F ( x k . JO) computed according to the following expressions: t JF(Xk1. present in J0.l. O)Axk + E ( x k . O) containing the partial derivatives of (10) with respect to Xk and uk. Uk. uk1./~)[501i i=1 (lS) . ] O = [Oi . and so providing the means to obtain a robust stability analysis.. Or]T T (12) (13) (14) fiku = [ukru.. can be found.. respectively. This model assumption.JF(xk_i. ~) q. both evaluated at the operating point. a second Taylor's expansion around the nominal model parameters 0 is now performed to the system description (16).i .. and so the following parametric uncertainty description can be adopted as a good measure for capturing the plant uncertainty: OiE[9. although possibly conservative for some parameters.50i 0i+50i] Vi= l.i .l . is shown to be easily integrated in the approximate feedback linearization scheme. ~. computed around the previous operating point. uki. as the higher order terms of this series expansion are considered to be negligible [9].
t (see Section 3.E . . Uk1. 0)]o= ~ Therefore.1 for details).k~. 0) being respectively given by: P. since plant uncertainty is not likely to be zero. . while A. B and C are appropriate choices for the state space matrices representing the desired imposed linear dynamics._~.I C B ) A v k (23) Notice that in the case where the nominal model parameters exactly map the plant parameters. ~) = ~ . O) and Ei(xk1. O. uk1).I ( .98 Miguel Ayala Botto et al. . U. O. uk1.0)[601i (19) while ~'i(x~l. . uk1.i . O. terms gF(xk_l. for i = 1 . u k . a more general state space description which takes into account for the model parametric uncertainty will be presented in the next section. respectively. [ 0 op (*k~. u k . u.i .60) will vanish according to expressions (18) and (19). The application of (22) to the model description with uncertainties found in (17). 0 = 0. i. results in the following closedloop description: Ayk+l = (CA + JEEiCA)Axtk + (6F . t 6E(xk1.. then the confidence regions will be given by [80]i = 0.JO) and JE(xk1. . Avk the newly created input signal. Uk1. provided that matrix E ( x k ._l. 0)]o= ~ (20) (21) ~:~(**1.JEE1F)Axk + +(CB + 6 E E . . 0) is invertible for all admissible pairs (Xki._~. the feedback law will be given by (letting fall arguments for a clear notation): AUk :. resuming the closedloop description given in (23) to the following expression: zlVk+i = CAAxlk + CBz~vk (24) which corresponds to a linear dynamic system having the following state space description: Axe+ 1 = AA:c~ "4BAvk Ayk = CAxtk (25) However. which means no plant uncertainty. Uk1. ~) = bE[E(**_l. Further.e.i . 60) " E i=1 Ei(xk1.(*k~. u.F A x k + CAAx~ + CBzSvk) (22) where Ax~ is the linear state vector.
4. Ai5 and At6 can be chosen to represent any linear timeinvariant desired behaviour.~p E (33) . A22.1 + +(F4 + JF4)Afi~2 + (E + 6E)Au~ (26) In the same way. considering the new extended linear state vector to be given by: resulting the state space uncertainty description given by: (29) where matrices . the feedback law (22) can be also be expanded according to: Auk = E . A44. A6s and A66 respectively given by: A21 = 6 ~(.F 2 + AI2)Z~9~I .Stability Analysis to Parametric Uncertainty 4 Towards a state space uncertainty description 99 According to the partition of the regression vector given in (11).6A~k_2 + BZlvk] (27) where matrices Aii.~F~)Z~y~+ (F2 + '~F2)A#k. At2./~ and (~ are given by the following expanded forms: J A n At2 0 0 Al~A16] A21 A22 0 0 0 i / i 0 0 0 0 0 A4a A44 0 0 0 = [B ~'=[I o o o o A = (30) 0 0 0 o I o 0 0 0 A6s A66 J 0] r o]:r (31) (32) with matrices A21. the prediction model with uncertainties described through expression (17) can be rewriten as: ~ayk+~ = (F~ + .F3Auk1 F4Zl~k_2 + AlsAvk_l + A. + (F3 + ~ F 3 ) ~ . A43. The state space uncertainty description is then obtained by combining expressions (26) and (27) into a single state space description.i [(F1 + A t l ) A y k + ( .~+t)pxp An = l.
.1 Robust feedback linearization Since the uncertainty matrix d/} in (37) is state dependent.z)p (35) while the uncertainty matrices Jfi. in order to analyze stability robustness to parametric uncertainty of the overall closedloop system.pxp A66 = I(. Therefore. respectively. Finding these quantitative measures requires an offline procedure to be applied based on a numerical search over all admissible operating points. 4. and J/} in (29) are respectively given by: E'(F1 + An) 0 0 0 01 0 EI(F2 + A12) 0 0 0 0 E1A1500 0  (36) Eo0 F3 0 "JEE1B" 0 E1B 0 0 0 E1F400 0 EiAle 1 0 (37) At this stage.100 Miguel Ayala Botto et al. The adopted strategy which assures that only relevant regions of the operating trajectory are considered.. a similar procedure as the one presented in Section 2 can be applied in order to incorporate this uncertainty term in the system dynamics (see expression (3) for details). described in (36) and (37).. (34) Ass = E ~.4 and J/~.. a polytopic uncertainty description has to be constructed based on the maximum bounds for each element of matrices 5. it should be noticed that the state space uncertainty description (29) has the same structure as the general system (1).
Acknowledgements This research is partially supported by Nonlinear Control Network. Communications and Control Engineering Series.. G. Jonas (1995) Nonlinear blackbox modeling in system identification: a unified overview. Slotine. 5. Utkin. ECTMR #ERBFMRXCT970137. Automatica 31(12).. W. References 1. USA 6. Vadim I. New Jersey.S. a basic requirement for the success of the proposed control scheme. 16911724 7. Seber. it can be automatically checked whether matrix E(xk1. van der Schaft. Springer. USA 4. PrenticeHall. Nijmeijer. as the model output travels along this trajectory.A. bounds on the uncertainty entries of matrices JA and J/~ can be collected such that the maximum values encountered along that trajectory are used to construct the polytopic description. Inc. S. (1986) Correlation based model validity test for nonlinear models. allowing the construction of a polytopic uncertainty description. C.A. Li. New York 2. 0) in (22).. 5 Conclusions In this paper a procedure was presented which allows for parametric plant uncertainty of a general nonlinear MIMO system to be propagated through an approximate feedback linearization control scheme. resulting in a linear closedloop system with nonlinear and state dependent parametric uncertainties. u~_~. John Wiley and Sons.Stability Analysis to Parametric Uncertainty 101 consists of designing a reference trajectory Avk and apply it to the feedback linearization of the process model. A. Sj6berg. Wild. One of the advantages of this approach is that uncertainty is considered only in relevant regions of the space spanned by the z~xk vector during operation. Berlin 3. and by PRAXIS XXIBDP/20183/99. Ljung.F. as the overall stability problem could be formulated with the aid of linear matrix inequalities (LMIs). Springer Verlag. while this procedure is bein~ performed. 235244 . Moreover. This particular configuration enables a straightforward application of robust stability analysis tools based on Lyapunov's stability concepts. New Jersey. (1992) Sliding modes in control and optimization. Voon. (1989) Nonlinear regression.F. H. International Journal o/Control 4(1). JeanJacques E. Then. By means of a numerical procedure these uncertainties could be bounded within intervals.J. Billings. Weiping (1991) Applied Nonlinear Control. is invertible over all encountered operating points. (1990) Nonlinear Dynamical Control Systems. PrenticeHall. J. Lennart (1987) System Identification: Theory for the User.
Balakrishnan.B. El Gahoui. H.. Boyd.. 295300 . E.. H.A. (1994) Linear matrix inequalities in system and control.J. AyMa Botto. 15 of Studies in applied mathematics (SIAM) 9. H. S. (1999) Linear predictive control based on approximate inputoutput feedback linearisation. 8.. V. L. Feron. Verbruggen.. IEE Proceedings .B. Vol. te Br~ke. Miguel) van Can.Control Theory and Applica= tions 146(4). Josd. S~ da Costa.102 Miguel Ayala Botto et al.L.
10129 Italy bacciotti~polito. BIBS stable) if for each initial state and each bounded input u(t) : [0.u) . Then we show how they generalize to systems with inputs. We show in particular that the existence of continuous Liapunov functions with suitable properties is equivalent to a type of external stability which is more restrictive than uniform BIBS stability. The situation is exactly the same as in the theory of stability for equilibrium positions of systems without inputs (see [2]. 1 Introduction A finite dimensional autonomous nonlinear system a:=f(x. In Section 3 we introduce the definition of absolute bounded input bounded state stability (our strengthened form of external stability) and state the main result. x E R n. We extend to systems with inputs the theory developed in [2].External Stability and Continuous Liapunov Functions Andrea Bacciotti Dipartimento di Matematica del Politecnico Torino. The last section contains the proof. In this note we prove that the analogy can be further pursued. it is known that continuous Liapunov functions m a y not to exist for BIBS stable systems of the form (1). u E R m (1) is said to be bounded input bounded state stable (in short. it A b s t r a c t . In fact. It is well known that external stability of nonlinear input systems can be investigated by means of a suitable extension of the Liapunov functions method. In the recent paper [3]. [4]). . We prove that a complete characterization by means of continuous Liapunov functions is actually possible. uniform BIBS stability has been characterized by means of certain upper semicontinuous Liapunov functions. +oo) + R m the corresponding solution is bounded for t _> 0 (see [1] for a formal definition and comments). In the next section we recall the basic notions (prolongations and prolongational sets associated to a dynamical system). provided that the definition of external stability is appropriately strengthened.
we denote Q(M) = U~eMQ(x). it is convenient to begin with some intuitive considerations. we recall that from a topological point of view. they can be reviewed as set valued maps..xo) for some t > O} positive limit set. let us assume that perturbations are allowed also at arbitrary positive times: under the effect of such perturbations. Now. Let Q(x) be a set valued map from R " to R". we introduce two operators. For M C_ R ' . the possibility of taking under control the growth of the prolongational sets leads to the desired strengthened notion of stability... and the We adopt the following agreements about notation. r). For M C R n. Ura and deeply studied in [2]. x0) represents the solution of (2) such that x(O. The existence of a continuous Liapunov function actually prevents the unstable behavior of the prolongational sets.. we denote IMI = s u p ~ E M Ixl. for a locally stable equilibrium of a system without inputs ~: = f(x) .. x0) = xo. the system may jump from the present trajectory to a nearby one. it may happens that an unfortunate superposition of these jumps results in an unstable behavior even if the system is stable and the amplitude of the perturbations tends to zero. More generally. In order to illustrate the idea. Roughly speaking. Powers of Q will be defined iteratively: Q~ = Q(x) and Qk(x) = Q(Qkl(x)) for k = 1. we simply write Br instead of B(O. On the other hand. First of all. We proceed now formally to precise what we means for prolongation. (3) where x(. This phenomenon is technically described by the notion of prolongation. stability is a way to describe the behavior of the system in presence of small perturbations of the initial state. These sets depend in general on the initial state.104 Andrea Bacciotti 2 Prerequisites As already mentioned. The open ball of center x0 and radius r > 0 is denoted by B(x0. Thus. r).2. f E C1 (2) not even the existence of a continuous Liapunov function can be given for sure. very useful tools for stability analysis are provided by certain sets associated to the given system. Auslander and Seibert ([2]) discovered that the existence of a continuous generalized Liapunov function is actually equivalent to a stronger form of stability. They are defined according to . denoted by :D and I . In 1964. If x0 = O. acting on set valued maps. Next. due to T. The simplest examples are the positive trajectory issuing from a point x0 F+(xo) = {y E R " : y = x(t.
.~ O k = O . not even if Q is transitive. .) )(. P r o p o s i t i o n 2 Let K be a compact subset of R n and let Q be a transitive prolongation. Starting from the m a p F + and using repeatedly the operators :D and Z. Then Q(K) = K if and only if K possesses a fundamental system of cornpaet neighborhoods {Ki} such that Q(Ki) = Ki. or Q(x) NOK #O. we can construct several prolongational sets associated to (2). . 2 . .External Stability 105 (:DQ)(x) = 06>oQ(B(x. so that for every x the set (~)Q)(x) is closed. :DQ is not transitive. ~)) (27Q)(x) . . for every set valued m a p Q . The positive trajectory is an example of a transitive map. The following characterizations are straightforward. F+(x) C O(x) (ii) (~Q)(x) = O(x) (iii) If K is a compact subset of R a and x E K . . we see that the construction (z. For instance. . it is called a transitive prolongation. . . XK such that xo = x. . (z(v(Z(VQ) ) ) ) . . . P r o p o s i t i o n 1 a) y E (VQ)(x) if and only if there exist sequences xk r x and Yk "+ Y such that Yk E Q(x~) for each k = 1. The operators :D and 27 are idempotent. not even if Q(x) is closed for each X. it is not difficult to see that Dl(x) :'. (27Q)(x) is not closed in general. In general. Q k ( x ) . If Q is a prolongation and it is transitive. y = XK and x~ E Q(xk1) for k = 1 . / ) Q has a closed graph. then either Q(x) C K .( / ) F + ) ( x ) . . D e f i n i t i o n 1 A prolongation associated to system (2) is a set valued map Q(x) which fulfils the following properties: (i) for each x E R n. . . . However. Moreover. 2 . I . In conclusion. . 2 .(. . When 27Q = Q we say that Q is transitive. K . .) (4) gives rise in general to larger and larger sets. b) y E (27Q)(x) if and only if there exist a finite sequence of points x 0 . The following proposition will be used later (see [2]).
More precisely. T h e o r e m 1 The equilibrium xo is absolutely stable for system (2) if and only if there exists a generalized Liapunov function which is continuous in a whole neighborhood of the origin. where U is a preassigned constraint set of R m. for each admissible input there are sequences {tk} and {uk} such that 0t0 <tl <t2 < . u) is of class C 1. as already mentioned. Indeed. The first prolongation characterizes stability..<tk. +oo) ~ U. D e f i n i t i o n 2 Let a be an ordinal number. is the set of all points lying on solutions corresponding to the initial state x0 and any admissible input. A solution of (1) corresponding to an admissible input u() and an initial state z0 is a continuous curve x(. 3 Systems with input The notion of prolongation applies also to systems with inputs ([5]). we set = . it is natural to give the following definition.tk).. the so called first prolongation of (2)..uk E U for t E [tkl. and u(t) . The equilibrium xo is stable of order a (or c~stable) if D~(xo) = {x0}. The procedure saturates when a = r. x0. u(. x0.)) = x0 and coinciding with an integral curve of the vector field f(. the first uncountable ordinal number. In other words. This allows us to speak about higher order prolongations. The intuitive construction (4) can be formalized by means of transfinite induction. (2) is stable at an equilibrium x0 if and only if Dl(zo) = {x0}..tk). u(. The first prolongation in general is not transitive. ... which obviously implies De (x) = D~ (z) for each a _> 3. uk) on the interval (tkl. Let us adopt the following agreement: throughout this note an admissible input is any piecewise constant function u() : [0.106 Andrea Bacclotti is a prolongation. The equilibrium xo is said to be absolutely stable when it is "/stable. Then. it is possible to prove that an equilibrium x0 of (1) is stable if and only if Dx (x0) = {x0). Indeed. The reachable set A(z0. it is possible to prove that ZD~ = D~. let c~ be an ordinal number and assume that the prolongation Da(x) of order fl has been defined for each ordinal number fl < a.. Assume that for each u E U. the vector field f(. U) relative to the system (1) and the constraint set U..)) such that x(0. The main result in the Auslander and Seibert paper [2] is as follows. Since.
provided that Ix(t)[ > p for each t E I.) : [0. and let U = BR. f ( x ) can be thought of as a function of x and u. D~(xo) = (~D(Z(7)AR)))(x0) and so on. S The following characterization is easy. ~ Vt > 0 . consider a system of the form (2) for which there exists a continuous function V(x) which is radially unbounded and nonincreasing along solutions. one has that the composite map t ~ V(x(t)) is nonincreasing on I. IO~(x0)l _%< .~R).) of (1) defined on an interval I and corresponding to u(. even with f E C ~~ Of course. As an example. there exists . let R be a positive real number. . D e f i n i t i o n 4 A (generalized) ABIBSLiapunovfunction for (1) is an everywhere continuous. D e f i n i t i o n 3 We say that the system (1) is absolutely bounded input bounded state stable (in short A B I B S stable) if for each R > O. T h e o r e m 2 System (1) is ABIBSstable if and only if there exists an ABIBSLiapunov function. can be reinterpreted as an ABIBS Liapunov function. The proof is omitted. +oc) + B a and each solution x(. constant with respect to u. Any V(x) which is radially unbounded and nonincreasing along solutions. The proof of Theorem 2 is given in the following section.9 > 0 such that Ix01 _< R.External Stability 107 Reachable sets are the most natural candidate to play the role of the positive trajectories (3) in the case of systems with inputs. More precisely. there exists p > 0 such that for each admissible input u(. We conclude by the remark that in general an ABIBS stable system does not admit ABIBSLiapunov functions of class C 1. P r o p o s i t i o n 3 System (1) is ABIBSstable if and only if for each R > 0 there exists a compact set K C R n such that BR C K and D ~ ( K ) = K . radially unbounded function V : R n r R which enjoys the following monotonicity property: ( M P ) for all R > O. and introduce the prolongations D f ( x o ) = (~DAR)(x0) . It is proved in [4] that such systems exist. We adopt the simplified notation AR(xo) = A(xo. We are now ready to state our main result. but not a C 1 function with the same properties.).
~ E WX. there exist some 0 > 0 such that Bpo+~ C Wx. T) such that x(r) E WX. Let m0 = maxtyl<po V(y). there exists ~ > 0 such that v(~)<X+~<X+2~< v(y) for all z E B(~. ~ E WX. Since Y is continuous.0. ~) and ~) e B(~.T].0 ~ wx+. P r o o f Of course. as well. x0. By construction. For each X > m o we have (DAR~ C_ Wx. Indeed. x0. . Even in this case we proceed by contradiction. This means Y(~) < X < V(~). u(. Such a r exists since the solutions are continuous. an admissible input u() with values in BRo. Let ~ > 0 be such that X + 3 e < Y(~)). In fact. Step 2.. Let 7. ~) and y E B(.108 Andrea Bacciotti 4 The proof Sufficient part Assume that there exists a function V(x) with the required properties. in the opposite case we could find X > m0. and a positive time T such x(T. V(x(r)) = X < V(. Set for simplicity x(t) = x(t. On the other hand. A contradiction. $) in such a way that ~ E A R~ (~). ~).: v ( ~ ) < x } .)) = . In fact. and 9 E (:DAR~ but ~ ~ W x. since ~ E Wx+e. On the contrary.0.y E Wx+c.E (0. According to ( M P ) we can associate to R0 a number P0.0) implies . We note that I~l ~ p0 ~ v ( x ) < m0 ~ x ~ Wx that is. we have DTR~ = Wx. without loss of generality we can take P0 > R0. In what follows. y ~ WX. we adopt the notation wx = {~ ~ R . C_ Wx. we can now take ~ E B(~. Fix R0 = 1. This is a contradiction to Step 1: indeed. 7]. For each X > mo we have AR~ C Wx. so that V(x(t)) is nonincreasing on this interval. By the definition of the operator :D. Assume that it is possible to find X :> m0. u(.)).0). Bpo C Wx. we should have . the fact that A + 2~ < V(. while x(t) ~ Wx for t E (r. and pick any X > m0. Lemma 1 For each X > m0. it is sufficient to prove that D~~ Step 1. Ix(t)l > p0 on the interval [r.
let us set E ~ ~ = U/~<a(Z(Dff~ prove that c_ w x / o r each > mo . hence.0 E Ea ($). In addition. Note that Wxo is closed (since V is continuous) and bounded (since V is radially unbounded). C Wx. Fix A0 > m0. ~ ~ Wx+t. To end the proof.Wx. and assume that the statement Dff~ = Wx for each A > m o holds for every ordinal number fl < a. Ua<. we have ~r E Wx+~.External Stability 109 Thus. We clearly have BR C BR. (f) and y E B(/). Set K . In conclusion. Wx0 C Bn. . Assume that there are A > m0. with Ri + +oo. and D ~ ' ( W x . V(x) >_ mo. We are finally able to prove the sufficient part of Theorem 2. We arrive to define a sequence of compact sets {Wx.. 9 E WX. (f). C K _ and D~(K)C_D~'(K) . and ~ E E~~ (~). Let finally R be an arbitrary positive number.l_>p. ) = Wx. we conclude that D ~ ' (Wx) = Wx for each A > m l . and let R/ be the smallest number of the sequence {Ri} such that R < Ri.. The proof of the l e m m a is complete. for sufficiently small we can find tf such that V(x) < X + e < X + 2e < V(y) for all x E B(~. Let us choose ~ and ~) satisfying this last conditions. for some R1 > P0 > R0 = 1.~ E (~DE~~ but '9 ~ WXAs before. This is possible because of the R~ definition o f / ) . by continuity. we have V(~) < A < V(~) and. A contradiction to (5). It is not difficult to infer that also = f o r each > r0 and. Fix A1 > rnl.no 9 (5) The final step is to For sake of convenience.K .(Z(D~~ = Wx for each A > . Using the properties of V. By repeating the previous arguments. Hence. we need to make use of transfinite induction.. Let a be an ordinal number. it is not restrictive to take R1 > 2. we find Pl > R1 and define mt = maxl. and . and iterate again the procedure. we have shown that D~~ = Wx for each A > m0.} such that Bn. and such that .
. This time. we arrive to assign a compact set W~ to any dyadic reciprocal 2k A = j. We proceed as above. Necessary part The idea is to construct a Liapunov function V by assigning its level sets for all numbers of the form 2k :3 j=1. Note that (beside the endpoints 1 and 2) 4/3 is the unique dyadic reciprocal with k = 2 included in the interval [1. (6) namely. Let us start by setting R0 . Using again Proposition 2 applied to W20 and W413w e define two new sets W~o C WSl7 c W413c WSl5 c W~.. . we can find a compact set denoted by W2o such that BRo C Wuo and D~~ . and D O(Ws/ ) = W / By repeating the procedure. C W~. CW2k+. and D~ ~(K) = K .. Note that they are dense in [1.. Let Rt ~ max{2.2 k.with k "d_ 1 and 2 k1 <_ j ~ 2k. Then we turn our attention to dyadic reciprocals ~. that is all dyadic reciprocals such that 2 o < 2~ < 21.. 1.W2o..j such that k > 1 and 2 k1 _< j < 2 k. by virtue of Proposition 3. we 2k obtain sets Wx such that D~ 1(Wx) . s.. This procedure can be iterated.110 Andrea Bacciotti The proof of the sufficient part is complete. 2k We have so assigned a set to any dyadic reciprocals T with j = 1. Using again Proposition 3.1. 2].W~k+~.2. in such a way that D~~ = W). Next. +oo).l . k0. C W2. we take Rk+l ~ max{It + 2. Assuming that W2k has been defined. c Wx c W~. and ukW~h = R n . consider pairs k.. the reciprocals of the so called dyadic rationals.Wx and 2k W2. and W2o C Wx C W~. there exists a compact neighborhood /t" of W2o such that K is properly contained in W2. ]W2o[}.ch that D ~ = W s . k = 0. if )~ < p. By virtue of Proposition 2.. we find a compact set W2~ such that BR~ C W~x and D~I(W2~) = W~. Call it W413.2..with k >_ 2 and 2k2 < j < 2 k . The sequence {W2~ } satisfies the conditions W ~ CBRh+. that is 2 < = < 4. According to Proposition 3.1.. ]W2h]} and the compact set W~k+l in such a way that B/tk+~ C W~h+~ and D~ k+~ (W2k+x) .
Let N > 0. This construction can be repeated for all k and j. we remark that by construction. Bacciotti A. (1997) Global Stability and External Stability of Dynamical Systems. Theory.+. Let the integer p be such t h a t 2 k+l _< 2p _< V(x) < 2p+I. Journal of Nonlinear Analysis. Claim A is proved. Let R be given and pick the integer k in such a way t h a t Rk < R < Rk+l. References 1. Claim A. then V(y) < A. We are now ready to define the Liapunov function V(x) for all x E R n as V(x) = inf{A : x E W x } . V is locally bounded. we . (7) Assume first that V(~) < l and pick a dyadic reciprocal A in such a way that V(~) < A < l and ~: ~ Int Wx. Note that Claim A implies property ( M P ) . For all sufficiently large v. we should have x~ E Wx as well. that is V(x) ~ 2 k. 11671185 . V(x) is radially unbounded. V(x) is continuous. and let k be an integer such that N < 2 k. and this is a contradiction to (7). Methods and Applications 28. For each R there exists p such that if Ixl > p and y E An(x) then v(y) < V(x). First of all. Also the proof of the necessary part is now complete.) = 1 # V(s Ixl > R2. First. V(x~. For have x ~ W2k. (x) C_ Wx . we remark that if Ix[ > p then x ~ W2k+l.. It follows that i f y E An(x). By possibly taking a subsequence. and let A be a dyadic reciprocal such that V(x) < A < 2p+I.. so t h a t V(x) > 2 k+l. Assume t h a t we can find a point s E R " and a sequence xv + ~ such t h a t V(xv) does not converge to V(~).External Stability 111 if $ < p. and Beccari G. Claim B. We prove that the choice p = Rk+2 works. we have lim V(x~. and hence An(x) C D~(x) C_ D~ k+. Claim C. But then. Since A can be taken arbitrarily close to V(x). (x) C D~. The case V(~) > l is treated in a similar way. Andriano V.) < A. Of course x E Wa. We finally obtain an increasing family of compact sets {Wx} with the property t h a t if 2 k ~ A < 2 k+l then D ~ ~(Wx) = Wx.
Systems and Control Letters.. and Mazzi L. and Bacciotti A. Annales Institut Fourier. to appear 5. and Rosier L. Tsinias J.112 Andrea Bacciotti 2. A Necessary and Su~icient Condition for Bounded Input Bounded State Stability oI Nonlinear Systems.. to appear 4. 333354 . Regularity of Liapunov Functions for Stable Systems. Bacciotti A. SIAM Journal Control and Optimization. Bacciotti A. (1964) Prolongations and Stability in Dynamical Systems. Grenoble 14. Auslander J. Kalouptsidis N. (1987) Lyapunov Functions and Stability o] Dynamical Polysystems.. 237268 3. and Seibert P. Mathematical Systems Theory 19.
a great interest is dedicated to new trends concerning robustness and harmonic rejection. optimal control. By adding some harmonic weighting factor to the regulation term in the cost function. robustness and low cost are the key merits of induction machines. harmonic attenuation.BP 46 38402 SaintMartin d'H~res. The control law results from an optimisation algorithm associated to a cost function.C. Harmonic rejection is used to be studied a posteriori when analysing the control law. Nevertheless. namely: 9 Field oriented Control (F. two classes of techniques axe widely applied to control induction machines.): where electrical vectors are considered in coordinates related to the rotating field. Balloul@inpg. The required performances depend on the nature of the application. the interested reader is invited to consult [4] as an example of minimal energy approaches. I n d e x t e r m s : Optimisation algorithm. and [5] as an example of passivity approaches. induction machine.Mazen. In this paper a new algorithm is used to regulate flux and torque in induction machine. static error.O. 1 Introduction Certainly. fr A b s t r a c t . due mainly to its nonlinear dynamics and to the variation of some parameters (such as the rotoric resistance or the load torque). the algorithm carries out a tradeoff between regulation performances and harmonic attenuation.Alamir@inpg. one of the axes is directed in the sense of the rotoric or the statoric flux. More precisely. However. However. etc. the induction machine becomes the favorable choice for a lot of applications ranging from railroad tracking to domestic appliances. its structural simplicity hides a great functional complexity. The power range extends from tens of Watts to tens of Megawatts. The analysis is often employed to compare the harmonics related to the studied control law with the harmonics related to some classical control scheme (PWM for example). so that the instantaneous torque . fr .Optimal Control with Harmonic Rejection of Induction Machine Iyad Balloul an d Mazen A l a m i r Laboratoire d'Automatique de Grenoble ENSIEG . France Iyad. Beside the classical features (rise time.). The aim is to show the possibility of attenuating the harmonics as early as the phase of the design of the control law while preserving good regulation performances. In general. This introduction is not intended to fully describe the state of the art of induction motor control.
to. u E ~ ' ~ is the control input. In the sequel. 2 Basic result Consider the following system defined on [to.T. resulting in a hybrid closed loop system.T. and z E ~ " is the controlled output. but also provides a rigorous method to treat new performance requirements: 9 Harmonic attenuation problem can be treated when the control law is in the design phase. This expression allows a dynamic regulation of the torque [6]. to. Section (2) shows the formulation of the basic optimisation problem.114 lyad Balloul and Mazen Alamir has a scalar expression. First. 9 Model uncertainty and disturbances can be considered by the algorithm yielding a robust optimal control law. tl]: E { ~ = f ( ~ ' u.C. and then to achieve both regulation and harmonic attenuation.C. tll (2) with U some compact set in ~ m . the harmonic to be attenuated).): where commutations of the power inverter are directly controlled [2]. the cost function is constructed to generate the input profile that yields a minimal regulation error. This paper shows the use of an optimal control algorithm to regulate the flux and the torque of the induction machine. But the poor theoretical foundation (caused mainly by the hybrid system framework) is still the common drawback of these approaches. Then. u) stand for the solutions of (1) starting from . In this paper only the first point (harmonic attenuation) is presented. the most relevant property of the algorithm in the context of power machines is its ability to handle the robust optimisation problem with nonconvexadmissible control sets. X(t. we investigate the application of the deterministic version of the algorithm to control induction machines. u. The control u satisfies the following constraint: u(t) e U Vt e [to. approaches. t) (1) where x E ~ " is the state. This algorithm was first proposed to solve minmax nonconvex optimisation problems [1]. Section (3) shows the application of the algorithm: first to simply achieve a regulation goal. But certainly. These approaches seem to better match the discontinuous nature of power machines which work typically on commutation (see w hereafter). 9 Direct torque Control (D. t) . ~(to) = ~o z = q(x. u) and Z(t. In this work. The algorithm constitutes a tool for handling the constrained robust optimisation of general systems (that are not necessarily afline in control). The proposed method does not only give a theoretically wellbased reference for D.e. the cost function is modified to reaiise a compromise between regulation error and the spectral power at some frequency (i.
xo. resistances.M ~ M2 q a = I . L and M indicate respectively. Let: J(to.xo) = min u(t). this problem can be written as: P(to.1 A p p l i c a t i o n on t h e i n d u c t i o n m a c h i n e System definition First.L~ L. i k rr= R~ . The index r categorises rotoric variables.xo) : or: u6u[tO..~2] (6) Using the minimum principle.t I. The associated optimisation problem is defined as follows: P(to. the used notations are defined: 9 9 9 9 9 I. 9 and F designate respectively. + zT(t)z(t)] uEty[to. and p is the number of poles and 12 is the angular The intermediate coefficients are given by: q = L. Ts=voltages. xo.L.tl] min J(to. the deterministic version of the algorithm [1] gives a solution u E U [t~ to the above problem (6).Lr . The efficiency of the solution is demonstrated via suitable simulations presented in next section. u ) and z(t) = Z(t. u) = T(x(tD) + ftl I z T ( t ) z ( t ) d t (3) be the criterion to be optimized. xo = x(to) is the initial condition. u). 3 3. The notation U [tx . s categorises statoric ones. t o . velocity. currents. t2] with value in U. self and mutual inductances. to.u) Since the initial state xo is fix.u ) (4) (5) with x ( t l ) = X ( t l .Lr = L.t~l stands for the set of all functions defined on [tl. t 1.tl. R.to. Y ( x ( t l ) ) indicates the final desired performances. the flux and the torque. and z(t) := Z(t.Optimal Control with Harmonic Rejection 115 to with the control u.tl. V.
Numerical values of the different parameters of the motor 3.0& 00" I p T a b l e 1... q 1 pl? L.~ ) 1 pf2 L. VI . q R. and u = ky. ~r and the phase between two successive vectors is .( . V2 .'~. The numerical values considered in this paper are summarised in table (1).116 Iyad Balloul and Mazen Alamir The equations describing the induction machine in the coordinates (a.. . The other vectors in the set has the same modulus (here equal to 2000V). q pn . power machines work on c o m m u t a t i o n bases./3) are given by [8. the set U defined above can be simply extended to include the corresponding extra vectors./3) as shown in Fig. .. V~ .] pn R. 88 . 89 } Note that V0 and V7 are identical. this is simply due to the definition of c o m m u t a t i o n cycle. 89 . Id~ B = \ Id~ / Id2 with Id2 denoting the 2 x 2 identity matrix. 88 .. ~ + ..2 Admissible control set By construction. the set of admissible control can be defined by: U = { Vo . 3 Note also that when the inverter is powering the machine with multilevel control. A ( n ) = \ ...R .7]): u = h(~) = A(~)x + Bu (7) 1 1 ( ~~ + g~7 ) with x = ( ~I.] h(~) = \h~(~)] = = kP(~ ~)] where ~ a n d /" stand for flux and torque to be regulated. The tension delivered to the machine by the power inverter can be represented in the plane (a. q . Therefore. (1).
As an illustration. Tension vectors in the plane (a.2s./to IIh(x) .1 et Time step is ~ = 50/~s./3) 3. u.. An interesting feature to quote is the convergence of the algorithm after only one iteration. t. the algorithm is compelled to select an input profile that attenuates the best this harmonic while preserving good regulation performances.3 Flux and torque regulation Let hr(t) = as: Fr be the output reference (flux and torque). > 0 pr Figure (2) shows the evolution of the machine for the step reference ~P~ .4 Flux and torque regulation with harmonic attenuation The structure of the optimisation scheme gives the opportunity to treat the harmonics as early as the design phase. The horizon of control is t! . Define the criterion f t/ J(te. (8) (1 Pr O)issomeweightingmatrix. consider that a current harmonic at w0 2~f0 is to be attenuated 9 Then.hr(t)ll~dt where Q := F~ = 1 0 0 .Optimal Control with Harmonic Rejection 117 V'4~ t w. w) = . 3. Indeed. by adding the following states to the previous model: ~5 = ~ cos~ot x6 = x l s i n w o t ~7 = ~ 2 c o s ~ o t ~5(o)=o x6(O)O xT(o) = o (9) (10) (11) hs =x2sin~ot xs(O)=O (12) .to = 0. 1. xo. by including some harmonicdependant penalty term in the cost function.I='Vl" " ~ (o) v6 Fig.
.118 150 lyad Balloul and Mazen Alamir r(. .hr(t)ll~dt (13) and the final performance function (suppose that to = 0 to simplify the notation): 1 : r ( x ( t l ) ) = t~ (x~(tl) + x~(tl) + x~(tl) + x~(tl)) (14) This function gives an approximation of the power of the harmonic corresponding to w0 in the time interval [0. . 2. . . . . w) = r ( x ( t l ) ) + lib(x) ..) 1.5 0 0~8 o01 0018 o02 . . consider the cost function given by: 0 0 0 0 00 0 0 J(t0. . 9 4t000 OO08 001 0018 OO2 0 0008 O01 0015 OO2 F i g . . . .$'3 t x4 2 . . tf]. xo.~ O. without harmonic rejection the new state matrix becomes: 000 000 000 000 A(F2) Aa(t) = cos wot 0 sin wot 0 0 cos wot 0 sin wo t Now. u. . . . 2 Xl t X2 o. .5 . tl. . Motor behavior in openloop: #r = 1 et Fr = 100. . . . . .
(4) shows the evolution of the motor under this input.f0 ~0 = 170Hz. ( ~ o ) X. The algorithm provides the corresponding control profile. Current power spectrum (logarithmical scale on y) without harmonic rejection ~ Now. is the current spectral power without harmonic attenuation. Figure (5) shows a comparison between the spectrum in this case and the previous one. X . Fig.~o) .Optimal Control with Harmonic Rejection From (14): 119 2 (::) X7 8 (15) Although.x . 9 the relative gain: g = x(. 10 a 10' 5OO frequency in Hz Fig. 3. consider that the unwanted harmonic is at .(~0) where X is the current spectral power with harmonic attenuation. (3). We start by showing the current power spectrum without harmonic rejection in Fig. . the results presented in this section are preliminary. they show the efficiency of the algorithm in tackling the question of harmonic attenuation. The parameters f and g in the figure are defined as follows: 9 the frequency of the harmonic to be attenuated: f = fo.
.. . ~ .. ..120 120 100 80 EtO Iyad Balloul and Mazen Alamir 1....... i .... . ... . In practical context. .. ~ .. . .. II 01 170. .. i i t 10* 0 frequency in Hz F i g .. some harmonics are more annoying than others. : : :. Consequently. . Current power spectrum (logarithmical scale on y): ( .. . .t. F~ = 100..to = 1 7 0 H z to. The figure shows that the good attenuation of the harmonic at the selected frequency is coupled with some degradation elsewhere... 1 : 9 " . : 4O 20 1 o. Motor evolution with k~r = 1.. .... ..... . 3~3 ! X4 4o0 o. : : .. ... .... 2.. .. ... i i . ..780S)8 * 101 .. .. . . .. ... . .. .. with harmonic rejection .. . .5 . . ......) with harmonic rejection Remark 1. ... . . ... .~1 7 x2 eoo : ". . (. . .. i i i i~ i i i t.. i : : i i . .. . . 4.) without harmonic rejection... a good attenuation of these annoying harmonics will be appreciated even if it increases other harmonics as a second effect. ~ : 10 ~ r .... i .~ O.. 5. ... : 9 .5 . 0... i ! e~176 o t:~5 o ol o o18 o o2 0 OOO8 00"~ O01S OO2 F i g . . . .. . . . 2 . .. .
6. Brevet (tous pays industrialisSs). (1997). .. Nicklasson. Bornard G.s. it worths mentioning that this algorithm has already been successfully applied in its minmax version on batch processes [1] and on openchannel hydraulic systems [3]. Commande adaptative non lin~alre de machine asynchrone. (1994). Roye. D. In the present case. 5. INPG. J. Balloul I. St. Alamir M.. In this respect. partie i. In: l l t h IFAC Workshop Control Applications of Optimization. Passivitybased control of a classe of blondelpark transformable electric machines. Whereas. L. Commande des machines ~ courant alternatif.. IEEE Trans. (1999). 9701346 3. Poullaln S. the original algorithm allows a robust synthesis that can consider exogenous disturbances and parametric uncertainties. (1997). 12711287 2. In: ]~cole d'~t~ d'automatique de Grenoble. Certainly. but it lacks a theory guaranteing this convergence. References 1. in this paper only the deterministic version of the algorithm is used. But in the other hand. on Automatic Control 42(5). Georges. EspinosaP~rez G. Petersburg. In: Ecole d'~t~ d'automatique de Grenoble. L. (1997) Machine asynchrone: commande s cycles limites control. especially if the number of vectors in U is increased. an efficient control scheme would run the algorithm under a suitable supervision layer. yon Raumer. a further work is still to be done to examine workable implementation issues: 9 The algorithm involves online computation making the time of execution a crucial factor in any implementation framework. The examples studied are chosen purely to illustrate the functionality of the algorithm. Russia. Fhrthermore. T.Optimal Control with Harmonic Rejection 121 4 Conclusion A key feature of the algorithm explored in this paper is its ability to realise a double purpose of regulation as well as harmonic attenuation. Ortega R. ]~tat de l'art de la commande ~ flux orient~.. Chen M.. X. on Automatic Control 44(7).. (1999) Robust constrained control algorithm for general batch processes.. D. Roboam. C. Int. Bethoux O. D. P.. Etat de l'art de la commande ~ flux orient6. pattie ii. Thomas J. Another attractive perspective of this work is to take into account robustness consideratious. 72(14). 7. Accepted 4.. Laboratoire d'Automatique de Grenoble. 9 The algorithm has some results characterising its convergence. 14301435. de Wit C. it is not too far as a preliminary result. a Pentium200 platform gives the solution in about 10 m s . 8. this is not fast enough to allow real time application of the algorithm. Actually. Thus. Improvement of execution time could be achieved using neural networks to emulate the optimisation algorithm. IEEE Trans. (1997). Georges (2000) Nonlinear robust control of an openchannel hydraulic system based on an infinitedimensional model. it might be helpful to employ more advanced operational research techniques (more intelligent than the used combinatory one). Obviously. Ramirez J. Commande des machines ~ courant alternatif. Nonlinear H2 and Hoo optimal controllers for currentfed induction motors. Jornal of Control. Phd thesis.
]nform~tica y Sistemas Universidad de Murcia Murcia. lngenier~a de Sistemas y Autom~tica Universidad de Vigo Vigo.Nonlinear QFT Synthesis Based on Harmonic Balance and Multiplier Theory Alfonso Bafios 1 . Spain A b s t r a c t . W h e n applying absolute stability techniques. Antonio Barreiro 2. The approach is to incorporate to QFT conditions given by the application of harmonic balance and multiplier techniques. for some types of memoryless nonlinearities like the saturation or the deadzone. Well known m e t h o d s are available in the literature to predict the existence of limit cycles. and Javier Aracil a 1 Dept. Part of this success is due to the fact t h a t is much less conservative t h a n other rigorous methods to test stability. Recently. 1 Introduction T h e existence of nonlinearities such us saturations or dead zones in control systems m a y give rise to the emergence of limit cycles. Absolute stability results provide a formal framework for the analysis of global stability. balancing between different possible alternatives. giving sufficient and usually very conservative frequency domain conditions. Spain abanos@dif. Spain abarre iroQuvigo. the source of conservatism arises from the confinement of certain nonlinear relation inside a cone or conic sector. providing the designer with a very transparent tool for synthesising stabilising compensators. e a 3 Dept.9]. Many approaches to shape the cone . Francisco Gordillo a. The stability conditions are conservative because they hold for the whole family of functions inside this cone. e s 2 Dept. namely the harmonic balance and the describing function m e t h o d s [11. lngenier/a de Sistemas y Autom~tica Universidad de Sevilla SeviUa . the designer must consider the different techniques balancing between conservativeness and heuristics. urn. T h e latter has been a very popular m e t h o d for decades in spite of its approximate character. The problem of deriving conditions for a stabilising linear compensator in a uncertain nonlinear control system is addressed. multiplier theory has been considered in the literature for minimising the conservativeness of previous absolute stability criteria. In a practical problem. such as absolute stability criteria (see for example [14]).
we introduce the structure of the nonlinear feedback system that will be considered. to the actual nonlinear relation and relax the conditions have been reported in the literature. and finally comparisons and conclusions are outlined. the technique of positive multipliers is introduced and applied to the example. and we address it within the robust Q F T techniques. Quantitative Feedback Theory (QFT) [8] is specially well suited for dealing with uncertainty. In Section 4. As a frequency domain design technique. regarding the control of an electric motor in the presence of backlash. In [1. under saturation or deadzone nonlinearities. In Section 3. and in general to obtain global stability. including symmetric and asymmetric characteristics. we discuss the possible use and interpretation of the words 'frequency locus' for a nonlinear block. we present an extension of the harmonic balance to the case of asymmetric nonlinearities. The main idea is to substitute the nonlinear system by an equivalent frequency locus. having a linear compensator to be designed. In a second stage this complex locus is used jointly with the uncertain linear subsystems to compute restrictions over a stabilising compensator in the frequency domain. and that may be decomposed as a feedback interconnection of a linear system and a (usually memoryless) nonlinear system. it can also be expected that Q F T can incorporate harmonic balance methods as well as stability conditions based on multiplier theory. and apply it to the example. The goal of this work is to give a first step to provide a common Q F T framework for solving the above stability problem using different frequency domain techniques.2]. In [12] a lot of possible multipliers (or integral quadratic constraints. in particular harmonic balance and multiplier theory. On the other hand.2] the problem of obtaining the lessconservative bounding cones was addressed in the context of Q F T techniques for uncertain nonlinear systems. In Section 5. An important aspect of the linear system is that it is allowed to have uncertain linear subsystems. IQC) are listed that can be chosen for a particular problem. a complex region that may depend on the frequency. A special class of memoryless nonlinear system will be considered here. In Section 2. . These techniques have achieved a particular success in the context of antiwindup control schemes [10]. We are mainly interested in robust nonlinear control problems. to give the designer a transparent tool for evaluating the balance conservativeness/heuristics. The content of the paper is as follows. A relevant example. and what is more important.124 Bafios etal. is developed. and it has been recently shown how it can be efficiently used to adapt robust versions of classical absolute stability results such as Circle and Popov Criteria [1. it will be possible to compare different techniques. The problem considered is to compute restrictions over the linear feedback compensator for avoiding limit cycles. As a result. treating the nonlinearity by the describing function method.
usually applied to a system L ( H . N) where the linear part is a linear fractional transformation of a system G(s). The problem is complicated for the existence of uncertainty in the linear subsystems Pi(s).. The simplest variant of this method is the describing function method used when the nonlinear block has odd symmetry and the linear part is a lowpass filter. i = 1.w) # 0 where (2) H(jw) = P~(jw) + P~(jw)G(jw)P4(j~) 1 + G(jw)Pa(jw) (3) .w) is the describing function of the nonlinear system N ... N ) without uncertain blocks. Here N(a. In this simple case.. Fig. 1).Nonlinear QFT Synthesis 125 2 Limit Cycles A n a l y s i s B a s e d on the H a r m o n i c Balance Consider a SISO Lure's feedback system L(H. A direct application to the system of Fig.w) # 0 (1) for every amplitude a and every frequency w. 4 are (possibly uncertain) blocks of the plant. The Lur'e type nonlinear system: G is the feedback compensator. P. suitable for the design of G(s) in order to (usually) avoid them. that is in the form of templates.9].4 (Fig. i = 1. and (possibly) uncertain linear blocks Pi(s). that will be considered in the framework of Q F T [8]. and N is the nonlinear block of the plant A wellknown solution for analysing the existence of limit cycles if given by the harmonic balance method [11. 1 results in 1 + H(jw)N(a. there are no limit cycles if (see for example [14]) 1 + G(jw)N(a.. 1. The problem is to derive conditions for the existence of limit cycles. Templates include parametric as well as unstructured uncertainty. the controller to be designed.
s(J. The system is supposed to be embedded in the feedback structure of Figure 2. after some simple calculations we obtain PI (jw)N (a.jw) + (P(jw)N(a.)' H~(s) . For any frequency. jw) + _b(jw)N (a. In this way. (6) defines a bound over the controller G(jw) for avoiding limit cycles. although probably more conservative. It represents an electric motor driving a load through a gear in the presence of backlash. 6 is considered as a given parameter. and for every amplitude a.jw) + Pa(jw))G(jw) (6) where /3(j~) = Pl(jw)P3(jw) + P2(jw)P4(jw). Thus (6) is the final condition to be satisfied by G(fiz) for every frequency w.jw). this method would be rigorous. The nonlinear system in this example is borrowed from Example 1 in [13]. Substituting (3) into (5). where 1 1 Hi(s) . jw)G(j~z) < ~.i = 1 .. for every value of Pi(jw).~) in such a way that it includes the errors caused by the approximation of the method (see [9]).s(Jts + Bt) . 1 + Pl(jw)N(a.s + B. 4 (note that in general Pi(jw) may be uncertain). Eq. developed in [3] and adapted in [1] to robust absolute stability problems. a finite bound 6 have to be used. being the describing function incorporated as an additional transfer function with uncertain parameter a. Thus. . that in addition guarantees some finite distance to the bound given by (2). Although no details about computation will be given here. Example Condition (6) can be treated in the framework of quadratic inequalities. a realistic example is developed in the following.. another approach that can be easily integrated would be to obtain the value of ~f(. (2)(3) can be integrated in Q F T as a robust stability boundaries.a. Let L(jw) be defined as L(jw) = H(jw)N(a. that is L(jw) [<~f 1 + L(jw) (5) In this paper.126 Bafios et al. Nevertheless. condition (2) is equivalent to L(jw) [ 1 + L(j~z) < oo (4) In practice. .
39e . meaning that not only (2) is satisfied for every possible combination of parameters.8[0.0032211. the control system is transformed to a Lure's type system (Fig. 1. For the deadzone nonlinearity N(a) is a number in the interval [0. (a) f Co) Fig.20]. in m a n y practical cases this requisite is not fulfilled.0) is "protected" by a circle centred at 1. After some manipulations. and independent of w. the 6~ center is ~ and the radius 1:'~6 Results are given in Fig. like the deadzone nonlinearity considered in the previous Example. andJt = 0. in order to satisfy condition (6) and thus avoiding limit cycles. but also that the critical point (1.1 (in general. 2. 3.o(jw)G(jw) must lie above the solidline boundaries and under the dashedline boundaries. 3 Harmonic Balance with asymmetric nonlinearities The above analysis works for the case in which the nonlinearity has odd symmetry.04111.01 and with radius 0. P3.85. where P1 = kt(H1 + H2). For each frequency. 1]. ) Boundaries have been computed for Pz.6. P2 = ktH1. being the describing function N(a) given as a function of the amplitude a.00275[0. P3 = kmH1.4.Nonlinear QFT Synthesis 127 and the parameters are known to be in some intervals.o(jw) stands for a nominal value of P3(jw). Here ~ = 10 is used.1] for any value of a. (b) Transformation of the control system to a Lure's system Equation (6) can be used to compute bounds over G(jw) for avoiding limit cycles.0015. Bm E 0. (a) Modelling of the motor.b). Bt E 0. andP4 = ktHz. where P3. kt E 4. However. 2. Jm = 6. 1]. and thus it is .o(jw)G(jw).2]. km E 0.
w) = 0 Therefore.128 Bafios et al. the bias gain and the first order harmonic gain from the input of the nonlinear part to its output.. respectively.. and consider its steady state periodic output signal z(t). where No and Nt are. O) ~ 0 (7) . This signal can be expanded in series form as z(t) = Noao + Re(Nlate j~ + . where y(t) represents the output of the linear part. and for every frequency 1 + H(O)No(ao. Let the signal y(t) enter in the nonlinear block.. Stability boundaries given by the describing function method for the frequencies 10. al. The dual describing function method [4] can be useful in this task. 3.al. (dO) 20 100 ~" % I I t S t F U 10 ~ 100 50 0 Fig. ai. ~".al.al. In this method the selfsustained oscillations are assumed to be of the form y(t) = ao + Re(aieJ~t).O) = 0 1 + H(j~)Nl(ao./~z(t)d(wt) 1 F~r Nl(ao. the condition for the noexistence of limit cycles is that one of the following inequalities is fulfilled for every value of a0 and al.. J 40 ~'~ ~. 100 and 1000 rad/sec. SO 12. They can be computed with the usual expressions of the Fourier coefficients as No(ao. ) = __l ~al ~_ r [" z(t)eJWt d(wt) In order to fulfil the firstorder harmonic balance the zero and first order terms of y(t) must be equal to the corresponding terms of the output of the linear part to z(t) (neglecting the higher harmonics) 1 + n(O)No(ao.w) = ~ __. convenient to extend the above analysis to more general situations including asymmetric nonlinearities.
where b_ < 0 and b+ > 0. a similar condition to the obtained in Section 2 .Nonlinear QFT Synthesis 129 1 + H(jtv)Nl(ao. al). Example Consider again the electric motor example of Section 2. where the function N(a. the scope of application of the method is extended since the emergence of multiple equilibria and limit cycles are the most common causes of loosing global stability [5. but also for any asymmetric deadzone. Following a reasoning similar to the Section 2. since the nonlinearity is memoryless. The easiest way to use Equation (7) to avoiding the existence of multiple equilibrium points and limit cycles is to derive conditions over G(jw) to satisfy the second inequality. boundaries given in Fig. substituting N(a. the function N1 (a0. .cV) ~ 0 Notice that the dual describing function method allows to deal with a more general problem: not only limit cycles can be detected but also equilibrium points (for at = 0 or ca = 0 and a0 # 0). which is a sufficient condition. al. In this way. the describing function is independent of the frequency.w). 1]. al) take values in the interval [0. 4.4. given by g0(a0. where the deadzone nonlinearity is now considered asymmetric as in Fig. and independently of the parameters b_ and b+ . a condition identical to (6) must be satisfied. In this case. > " > Fig.. As a result. 3 are not only valid for a symmetric deadzone. In addition. el) and Nl(ao. w) by N1 (a0.w) is substituted by Yl(ao.1] for any real value of a0 and al. The difference of having an asymmetric deadzone nonlinearity is that now we need a dual describing function.al. Asymmetric deadzone nonlinearity This second condition can be embedded in the Q F T framework using (6). al) = [0.6].
centered at .N .1 ] . In the example of Section 1. For example.1 + eU)/(1 .0 (8) Thus the linear part H(jw) must avoid the critical locus . is shared by m a n y other different nonlinear blocks. or: H(jw) n . An additional way to increase security in the guaranteed stability is to define a robustness gap as in (5).N . The closed loop L(H.N . in the form of boundaries.[0. . a bounded intervalar uncertainty. N) will not have limit cycles when (2) holds. the unit saturation also has the 'frequency locus' N . including the asymmetric deadzone.1 + eU)/(1 . N) is based on H(jw) avoiding the same locus. 1] and 'critical locus' .e2) (9) where U stands for the unit circle in the complex plane. so that we can identify it with N = [0.N 1 . for N being the (symmetric or asymmetric ) deadzone as well as the saturation.o o . in the sense that H(s) with a deadzone is stable if and only if H(s) with a saturation is stable.' is changed by ( .1]. the deadzone can be represented by an arbitrary number in [0. omitting arguments for H(jw) and N(a.Y 1.HN.( . The stability (lack of limit cycles) of the loop L(H. each point z in the critical locus is changed by a forbidden circle. Putting L . for which the first harmonic is real between 0 and 1.N .1 ] . derived from the describing function.w).' = H(jw) n ( .( . which seems a rather unlikely property.N ~ . Thus.I ) ( . Otherwise.1/e in (5) on the forbidden critical locus: .1 ] .130 Bafios et al.o o . the set of all stable linear H(s) will be unable to discriminate deadzone and saturation. or equivalently when H(jw) =/s . but useful for interpretations. it will be an approximate object. 4 F r e q u e n c y Locus for t h e N o n l i n e a r B l o c k Although the main objective of this work is to obtain frequency locus for the controller G(jw). So there is some kind of approximation in this approach. however. it m a y be interesting to consider the 'frequency locus' filled by the nonlinear block N .i ) ( . then (5) amounts to II+(HN) l]>e (HN) 1 q~ 1 + eU H r Nl(1 + eU) 1 . The validity of this 'frequency locus' analysis for such nonlinear blocks is strongly confirmed by the wellknow validity of the harmonic balance principle for most of common practical cases (in which H(jw) introduces a suitable lowpass effect). in order to analyse the degree of robustness/conservativeness achieved.o % . As the nonlinear frequency response is not a so clearly defined object as the linear one. and putting e = 1/(f. 1] C R. Equation (9) shows the effect of introducing a robustness gap (f .( . It should be noticed that this property. .eu).
for different values of 6 = 1/e. Robust absolute stability in the framework of Q F T has been developed in [1.. .. In this Section.i . : .. in [7] the closed loop region for L/(1 +L) is a rectangle..1].. where d = 1 + 1/r is the rectangle halfwidth. In this case. we only consider memoryless nonlinearities like the deadzone example of Section 2 or Section 3.. which have the property of lying incrementally in [0. i. . Critical locus for the unit saturation and dead zone It is worthwhile to analyse the relationship with other robust frequency bounds for L = H N that have been proposed in the literature. . The union of all these circles forms the 'robust' critical locus.....I . where C is the given clover. In fact..... . instead of a circle like in (5).S 3 2. .1 ) C . The former gives approximated solutions to the global stability problem..2]..S ~ .. this locus is introduced to simplify the performance monitoring. . A memoryless nonlinear system N given by a relation y = f ( x ) is said to lie incrementally inside the sector [kl.eZ)...1 + r / 2 and . which is not our approach.S 0 0. while the latter gives rigorous solutions at the cost of being much more conservative... 5.. . Fig.N . 5 shows this locus for the unit saturation and dead zone. i ... For example.Nonlinear QFT Synthesis 131 z/(1 .1 =l=jr/2 and with radius r/2. and robust absolute stability criteria. formed by the union of four circles centred at . k2] if . S 2 3.e2).IS I Fg. 5. so we will considered (5) as an adequate robustness bound and frequency loci like in Fig.... Then.. .and with radius Izle/(1 ... i :. it can be seen that the corresponding open loop condition is L = H N ~ C where C is a clover.5 1 R~ ~ ~xm '~! .. . 5 Comparison with A b s o l u t e Stability Criteria An interesting question is about the comparison between the harmonic balance principle as developed in Sections 2 and 3. 2f O... the locus for H takes the form H ~ ( . i 0!6 ..
For this type of nonlinearities.1].N . and lies incrementally (possibly noncausai) W(s). More formally.W(jw)(H(jw) + 1)) = 0 defines a boundary for H(jw). or a critical locus given by .W(jw )IH(J~ . with impulse response w(t) by IIw(t)lh < 1. Closedloop stability is defined as I / O stability from any input entering additively to the feedback system to the rest of signals. the output y = Hx is in L2 and. where condition iii was based on the inequality Re((1 . in addition.( . This boundary can be obtained solving the equation 1 W(jw)(H(jw) + 1) = ja for any a E R . where N(a. x ~ E R . For each frequency w. kl < f(ar <~k2 ' for every x.I . from this last condition it is straightforward to prove that if (6) holds for this N(a.W(jw) (12) that in general are frequency dependent.0 or 1 + N(a. Ilyl12 < k 11~112 for some constant k. .w)H(jw) = 0. is stable if i)H(s) in [0. as found in [10] or [12].( .1 + 1 . iii) for some with L1 norm bounded where (10) and iv)if N is not odd then additionally w(t) > 0 for any t E R. w ) = . The proof is a minor modification of Th. the Circle and Popov Criteria used in [1] are generalised using a minor modification of multiplier stability conditions.N . Finally.~ ~ R.1 + 1 .w) = .132 Bafios etal. while the nonlinear frequency locus are circles.1 + 1 .l ( a . After some simple manipulations the result is ja 1 . a system H : L2. the critical locus as given by (12) can be represented in the complex plane as straight lines passing through the 1 point. Proof. that is ja .e is finite gain L2 stable if for every input x E L2.e + L2. Our sense of I / 0 stability is finite gain L2 stability. Following Section 4. iON is static. ja N(a. The main result is: Theorem: The system L(H. condition (6) holds for any a. 11 in [10].W(jw)(H(jw) + 1)) > e (11) The limit condition Re((1 .w) is given by (10). N). odd.W(jw)"~1 (3).o~) then the original condition (11) is satisfied (note that there is an extra degree of robustness given by the parameter ~ in (6)). where H is given by is stable. this theorem can be interpreted in terms of a stabilising nonlinear frequency locus N given by (10).
W(jw)).6 shows the frequency nonlinear locus for a nonlinearity satisfying the conditions of the above theorem. and for W(s) ."'" + s . the graph of Hl(jw) must avoid the circle Cw (w). An interesting question is how much it can be relaxed. In this case.0 (no multiplier). In general. the only influence of W(s) in the complex plane over the nonlinear frequency locus is to rotate the bound Re(z) = .~ ~.3(. It suffices to apply the inequality (6) with N given by the family of circles. introducing a particular.0. LFrom (10) it is easy to obtain that the centre c(w) and the radius r(w) of Cw(w) are given by c(w) = 0.Nonlinear QFT Synthesis Comparison of frequencial and critical locus 133 Note that for W(s) .' +. the excess of conservatism given by the Circle Criterion can be minimised using appropriate multipliers. The main difficulty is that the search space of possible multipliers W(s) is infinitedimensional. But when gl(jw)) lies on the lower (upper) halfplane.W(fiv) introduces a phase shift N becomes an offaxis circle. it is straightforward to obtain boundaries for G(jw).1 . On the other hand Fig.1 with an angle given by the phase of (1 .s1) 2 + (s . By finding appropriate multipliers W(jw).~i + s_~71 It f 1). parameterized family Of multipliers in the form a0 al b0 bl W(s) .1] (Circle Criterion). we recover the classical (and very conservative) bound from the circle criterion. . that is r = 0. However. that will be referred to as Cw(w). the conservative bound given by the Circle Criterion can be remarkable relaxed. This was arranged in [1Olin the context of antiwindup design.1) (13) For values of W(jw) close to zero. 1] C Cw(w). is shown the angle introduced by the multiplier as well as the boundaries for H(flv) in the complex plane. and how close to the describing function locus (oo. Fig. For stability based on multipliers.1]. Results in this line will clarify the limits of conservative's for both the describing function and the multiplier techniques.1 + ~ (.H 1 (jw) must avoid the real segment [0.s +1 .W ( j ~ ) ) . and we recover the circle of the Circle Criterion. N as given in (10) approaches the complex circle around [0. . the complex halfspace Re(z) < . when 1 . Note that at frequencies where (14) W(jw) is real there is no phase shift.) can be moved up (down) enough so that the forbidden region on the current halfplane is very close to [0. and for stability based on harmonic balance . Once the multiplier W(jo~) is fixed and N is identified with the complex locus defined by the circles c(w) + r(w)U given by (14).1 ] can be.5(1 + jtan(d(w))) r(~) = le(~)l where r = Angte(1 . 7 shows circles Cw(w). it is expected that the circle Cw(u.
134
Bafios et al.
B ~ q ~ m t e e fro' H ( ~ ) ~ am, 0.3. a l . 0. ~ 0. I~.R 0.,3. b l . 0. b2~ 0
[
,%
~ / . ~
ttL a.t I i
i
i
i
i
[,,.0.~.,{/
i
,
0 + i ~ all. N . i, l l , , e
.... i ...... ~ ,;oi.~:lti/i .....
9
i~ i
i
:
:
:
:
:
i
9: /
/
i
;:
i
? !
...... ~ .......
. . . . . . . . .
i
'i
.... iiii77!7 ii7i7!! 777r!i
1,1 I,I I.4 1.2 I
i
....
oA
i  o . s  o . 4 4.2
o
Fg. 6. Left: Phase shift for the multiplier W(s) = 0.3(1/(s + 1) + 1/(s  1)), right: associated critical locus
o.
FtlqUlllmy lucul IW N Imdor I n  CL~LI l l . O. Ua O, bO O.& h i . O, t ~  0
,
,
o . . . .
.o.4
.
.
.
.
.
.
.
.
.
.
.
.
~i
O2
0
'
02
*
0.4
i
0.1
i
Oil
f
i
12
Fg. 7. Frequency locus using the multipfier W(s) = 0.3(I/(s + 1) f 1/(s  1))
Example For the electric m o t o r e x a m p l e of Section 2, a p p l i c a t i o n of the s t a b i l i t y theo r e m based on multipliers is valid for b o t h the s y m m e t r i c and a s y m m e t r i c deadzone nonlinearity, since t h e y are static, and lies i n c r e m e n t a l l y in [0,1]. T h e multiplier W(s) = $ +9990 has been chosen, because it provides a phase i00
shift of almost 90 between I0 and I00 rad/s. In fact, 1  W(s) = s +s+l0 is a 100 lead network. Resuts are given in Fig. 8. Boundaries have been computed for P3,o(jw)G(jw), where P3,0(jw) stands for a nominal value of P3(j0p). Note t h a t for every frequency, P3,o(jw)G(jw) m u s t lie above the solidline b o u n d aries and under the dashedline boundaries.
Nonlinear QFT Synthesis
135

I
i i
10
t,~. (4) so i i
"
k,
280
2~
160
100
80
Fig. 8. Stability boundaries computed using multiplier W ( s ) = 999/(s + 1000), for the frequencies 10,100 and 1000 rad/sec. Boundaries of Fig. 3, given by the describing function method, can be used as a guide to measure how good is a multiplier, if the resulting boundaries are close to them. In this sense, one may conclude that the chosen multiplier give reasonable results, in the sense the corresponding boundaries, shown in Fig 8, are no overly conservative. Of course, a more rigorous study is needed in order to formalise and systematise this procedure. Conclusions The problem of designing stabilising linear compensators for uncertain nonlinear feedback systems has been addressed. A solution for a special type of nonlinear systems has been developed in the frequency domain, using the framework of QFT. The main features of Q F T are frequency discretization and an algorithmic approach for deriving boundaries under arbitrary plant uncertainty. Q F T boundaries are given in the Nichols chart, thus it is easy for the designer to decide between different solutions and compare alternative formulations. As a result, it has been found a condition (Equation (6)) for the computation of restrictions over a linear compensator using standard Q F T algorithms. This condition treats the nonlinearity as a region of the complex plane referred to as the nonlinear frequency locus. In this way, it has been possible to adapt and compare two different approaches to nonlinear stability: harmonic balance and multiplier techniques. As it is wellknown, harmonic balance is an approximate technique giving very relaxed conditions for global stability (in a qualitative sense, avoidance of multiple equilibrium points and limit cycles), while multiplier techniques gives global I/Ostability conditions but with some degree of conservatism. Since multipliers can be interpreted
136
Bafios et al.
as phase shifters of the complex circle given by the Circle Criterion, the proposed QFT approach gives a very transparent tool for electing appropriate multipliers. In this sense, we formulate and explore in a examplebased way the 'scope question' of: can we find a multiplier for for proving stability of every stable system?. This exploration suggests that the answer will be 'yes' in many practical cases. We believe that this is an interesting open question, which has links with other timeversusfrequency problems in nonlinear stability.
References
1. Bafios, A., and Barreiro, A., 2000, "Stability of nonlinear QFT designs based on robust absolute stability criteria", Int. J. Control, 73, 1,7488. 2. Barreiro, A. and Bafios, A., 2000, "Nonlinear robust stabilization by conicity and QFT techniques", Automatica, Vol. 36, No. 9. 3. Chait, Y., and Yaniv, O. 1993, "MISO computeraided control design using the Quantitative Feedback Theory". Int. J. Robust and Nonlinear Control, 3, 4754. 4. Cook, P.A., 1986, "Nonlinear Dynamical Systems", PrenticeHall. 5. Cuesta, F., Gordillo, F., Aracil, J. and OUero, A., 1999 "Global Stability Analysis of a Class of Multivariable TakagiSugeno Fuzzy Control Systems" .IEEE Trans. Fuzzy Systems, 7, 5, 508520. 6. Gordillo, F., Aracil J. and Ollero, A. (1999) "Robust Stability Analysis of MIMO Systems with Asymmetric Nonlinearities". Proceedings of the 14th World Congress of IFAC, Vol. E, 123127. 7. Gustafsson, F., and S.F. Graebe, 1998, "Closed Loop Performance Monitoring in the presence of system changes and disturbances", Automatica, 34, 13111326. 8. Horowitz, I., 1993, Quantitative Feedback Design Theory (QFT), QFT Publications, Boulder, Colorado. 9. Khalil, H.K., 1996, Nonlinear Systems, Prentice Hall, Upper Saddle River, NJ. 10. Kothare, M.V., and M. Morari, 1999, "Multiplier Theory for Stability Analysis of AntiWindup Control Systems", Automatica, 35, 917928. 11. Mees, A.1.,1981, Dynamics of Feedback Systems, Wiley, New York. 12. Megretski, A. and A. Rantzer, 1997, "System Analysis via Integral Quadratic Constraints", Automatica, 42 , 819830. 13. Oldak, S., Baril, C., and Gutman, P. O., 1994, "Quantitative design of a class of nonlinear systems with parameter uncertainty", Int. Journal of Robust and Nonlinear Control, 4, 101117. 14. Vidyasagar, M., 1978, Nonlinear Systems Analysis, Prentice Hall, Englewoods Cliffs.
Group Invariance and Symmetries in Nonlinear Control and Estimation
John S. Baras Department of Electrical and Computer Engineering, Department of Computer Science and the Institute for Systems Research, University of Maryland College Park, MD 20742, USA
b a r a s ~ isr. umd. edu
A b s t r a c t . We consider nonlinear filtering problems, nonlinear robust control problems and the partial differential equations that characterize their solutions. These include the Zakai equation, and in the robust control case two coupled Dynamic Programming equations. We then characterize equivalence between two such problems when we can compute the solution of one from the solution of the other using change of dependent, independent variables and solving an ordinary differential equation. We characterize the resulting transformation groups via their Lie Algebras. We illustrate the relationship of these results to symmetries and invariances in physics, Noether's theorem, and calculus of variations. We show how using these techniques one can solve nonlinear problems by reduction to linear ones.
1
Introduction
Symmetries have played an important role in mathematical physics as well as in systems and control. Symmetries in mathematical physics [1] are essential. Essentially all physics theories can be based in symmetries and s y m m e t r y properties. Some of the more celebrated results are: (i) Conservation laws; various physics theories. (ii) Quantum electrodynamics, elementary particles, quarks, strings. (iii) Quantum field theory, reductions, s y m m e t r y braking. Symmetries have been also fundamental in systems and control. Perhaps the most well known principle has been the unifying role that equivalences of internal and external representations and associated groups of transformations play in system theory. Some of the more celebrated results are: (a) Electrical networks and realization theory. (b) Feedback invariants. (c) Nonlinear Filtering and Estimation Algebra. (d) Parameterizations of Rational Transfer Functions. (e) Canonical forms of linear analytic systems (linear in the controls). (f) Feedback linearization. (g) Symmetries in multibody mechanical systems and continuum mechanics
138
John S. Baras
Given the rich interplay between mathematical physics and control systems, especially variational problems and optimal control, there remain many unexplored theoretical and applied aspects of symmetries for systems and control. In this paper we describe our research in this direction with focus on stochastic estimation and nonlinear control. Both mathematical physics and systems and control deal with differential equations (DE). Therefore, symmetry groups of differential equations and systems of differential equations provide a natural starting point for understanding the key methods and concepts. As a simple example consider a scalar ordinary differential equation (ODE): F x,u,~ = ~f(x,u)=0 . (1)
The left hand side of (1) can be viewed as defining a surface in IR3 (three du variables: x, u, T~)" The middle term of (1) is the ODE and its solutions are scalar valued curves. A Symmetry Group of an ODE [1] is a group of transformations on (x, u) (the independent and dependent variables) which maps any solution of the DE to another solution of the DE. Similarly for systems of DEs. Thus finding symmetry groups for (1) amounts to finding transformations (diffeomorphisms) H : IR~ + IR1 6 2 (x,u) ~ ( r ~ t (2)
which permute solution curves. Finding such groups is a celebrated old problem initiated by Lie and later extended by Ovsjannikov and many others. Continuing with this simple example, if we find such a transformation H, we can extend it to derivatives using the simple observation that if a curve passes through (x, u) with slope du/dx, its image (under H) passes through (x', u') with slope du~/dx ' where:
~' = r u) u' = r u) du' (r + r dx; = (r + r
The map:
du/dx) d~ld~)
(3)
H" :JR 3 .} ]R3
du,,
,
is an extension of H. A key theorem in the investigation of symmetry groups for ODEs is the establishment of the result that H permutes solutions of the ODE (1) iff H I
Group lnvariance and Symmetries in Nonlinear Control and Estimation
139
leaves the surface in IRa, defined by the lefthand side of (1) invariant. This equivalence gives to the problem of constructing symmetry groups of ODEs a very attractive geometric foundation. The so called LieOvsjannikov method [1] for constructing symmetry groups of ODEs is to find all H ~ that have the surface s = f ( z , u) as an invariant manifold. This idea, as worked out in this simple example, can be extended directly to n t h order ODEs and to systems of ODEs. For an n th order ODE F ( x , u , ul, u s , . . . , u a _ l ) = Un  f ( x , u l , u s , . . . , u , _ l )  0
(5)
u~ = ~
d%
one extends a transformation H on (x, u), nfold to a transformation H ~ on (x, u, u l , '  " , Un). The LieOvsjannikov method (I) extends as well. More interesting are oneparameter Lie groups which leave the solutions of (5) invariant:
x' = X(z, u;~) } ~' u(~,.;,)
The infinitesimal generator of this group
(6)
x = ~(~, ~) ~
a
+ n(~, u) ~
0
(7)
plays a fundamental role. One extends the group nfold to derivatives of all orders to get an n th order group and n th order infinitesimal generator X (n) [I]. Some of the basic results of the theory are:
1. The oneparameter Lie group leaves the ODE invariant iff its n t h order extension leaves the surface F = 0 invariant. 2. The family of all solution curves is invariant under the Lie group iff it is a symmetry group. 3. T h e o r e m (Lie}: The oneparameter Lie group is an invariance group of F  O , iff
X (") = (un  f ( x , u, ~ 1 ,  " , u. = f(.,., ul,., u.1)
u n  1 ) ) = 0,
when
The consequences of these foundations were pursued by Lie who showed how to construct the Lie group (of invariance), and that the Lie algebra of infinitesimal generators determine the local Lie group. The subject has attracted many researchers through the years. Some of the more interesting results that have been obtained are:
140
John S. Baras
9 The reduction of the intractable nonlinear conditions of group invariance to linear homogeneous equations, which determine the infinitesimal generators. 9 Invariance of an ODE under a oneparameter Lie group of point transformations leads to reduction of the order of the ODE by one. 9 Invariance of an n th order ODE under an rparameter Lie group with solvable Lie algebra is reduced to an (n  r) th order ODE plus r quadratures (integrals). 9 Invariance of a linear partial differential equation (PDE) under a Lie group leads to superposition of solutions in terms of transforms. We are more interested in symmetry and invarance groups of PDEs. These transformation groups are local Lie groups. Point symmetries are point transformations on the space of independent and dependent variables. Contact symmetries are contact transformations acting on the space of independent, dependent variables and derivatives of dependent variables. Ovsjannikov showed that if a system of PDEs is invariant under a Lie group, we can find special solutions of the system, which are called similarity solutions. A further generalization of these concepts with key significance for both systems and control and mathematical physics are the LieB~cklund symmetries (or transformations) [1]. In these transformations the infinitesimal generators of the local Lie groups depend on derivatives of the dependent variables up to any finite order:
x' = x + ~ ( ~ , u, ul, ~ 2 , . . . , ~p) + o(~ 2) } ~' ~ + ~ ( ~ , u, ul, u~, , ~p) + o ( ~ ) _
(8)
It is a basic result in the theory of such transformations that the infinitesimal generators can be computed by a simple extension of Lie~s aglorithmn. Another key result is that invariance of a PDE under a LieB/icklund symmetry usually leads to invariance under an infinite number of symmetries (connected by recursion operators) [1]. A most celebrated results in variational problems with foundational consequences in mathematical physics is E. Noether's Theorem [1]. EulerLagrange equations are the governing equations of many physical systems; they are of fundamental importance in mathematical physics. EulerLagrange equations provide the dynamics of systems from a variational formulation (typically energybased variational formulation). These ideas from mathematical physics have inspired many research efforts in systems and control: from stability theory, to dissipative systems, to communication network routing, to robot path planning (to mention just a few). In this context a physical system has independent variables ~ in IR" and dependent variables u in IRrn. The system independent variables x can take values in a domain 12 of IRn . A function is given or constructed which depends on the independent variables x, the dependent variables u, and derivatives of
Group lnvariance and Symmetries in Nonlinear Control and Estimation
141
the dependent variables up to order k, Ul, u 2 , .   , Uk. The dynamics of the system evolve so that the paths u(z) correspond to extremals of the integral
J(u)=/a
L(x,u, u l , u 2 , . . . , u k ) dx
(9)
The function L is called a Lagrangian and the integral J(u) an action integral. The path u(x) = (ul(x), u~(x), ... ,urn(x)) describes the state evolution of the system and typically has to satisfy some boundary conditions on (9/2. Such formulations are well known and used by control theorists and practitioners. Clearly, if u(x) is an extremum of (9), any infinitesimal change
u(x) ~? u(x) + ev(x)
(10)
which also satisfies the boundary conditions, should leave J(u) unchanged to order O(e). The most significant relationship of this formulation is with respect to conservation laws of a system. A conservation law of a system, is an equation in divergencefree form
div f = ~
i=1
ni f i ( x , u , u l , u ~ , . . . ,uk) = 0
(11)
Equation (11) must hold for any extremal function u(x) of (9). The vector f is called a conserved flux [1] since (11) implies that a net flow of f through any closed surface in the space x is zero. EulerLagrange equations are the (often dynamical) equations that need to be satisfied by an extremum of (9). We refer to [1, pp. 254257] for a concise and clear derivation. The EulerLagrange equations can be ODEs or PDEs dependent on the problem. As such, we m a y ask the question if they have symmetry groups (invariance groups). Noether's key idea and result was that in order to find conservation laws it is far more fruitful to investigate the invariance of the action integral (9). Noether considered LieB/icklund transformations that leave the action integral invariant:
~' = ~ + c~(~, u, ul, u ~ , . . . , up) + o ( J ) u' = u + erl(x , u, ul, u2, . . . , uP) + O(e ~)
(12)
Noether showed that the existence of such transformations lead constructively to conservation laws of the corresponding EulerLagrange equations. She established the explicit relationship between the infinitesimals ~, 7/and the conserved flux f . For a concise proof of this celebrated theorem we refer to [1, p. 257260]. This celebrated theorem induced fundamental reformulations of mathematical physics, bringing a certain degree of unification. They include:
142
John S. Baras
9 Invariance under time translation leads to energy conservation. 9 Invariance under translation or rotation in space, leads to conservation of linear or angular momentum. 9 Relativity theory formulations. The relationship with symmetry groups allows the determination of variational symmetries using Lie's algorithm. Noether's theorem resulted in many specific applications (for specific physical systems or phenomena) such as [1]: 9 Conservation of the RungeLenz vector in Kepler's problem. 9 Existence of infinity of conservation laws for the KortewegdeVries equation and other soliton equations. We close this brief review of the history of research on symmetry groups for ODEs, PDEs and dynamical systems by listing some more recent results and activities. Symmetry groups allow discovery of related DEs of simpler form. This for instance leads to transformations that map a given equation to a target equation. Comparing the Lie groups of symmetries admitted by each equation, actually allows the construction of the mapping. Such results have significant implications in facilitating the solution of new ODEs and PDEs using solutions of other ODEs and PDEs, known already. For our subject, it is important to consider transformations beyond local symmetries. These are transformations where the dependence on u and the derivatives of u is global (i.e. not just through the instantaneous values u(x)). Gauge transformations in mathematical physics and quantum field theory can be such global transformations. In the theory of symmetry groups such transformations are called potentials. Ideas, techniques and algorithms from symmetry groups have made fundamental contributions to mathematical physics. As it should be clear from this brief exposition there is great potential for similar impact and fundamental new advances by the systematic exploitation of symmetry groups in systems and control problems. Many of the advances in mathematical physics came out of application of symmetry groups in fundamental PDEs of mathematical physics. This inspires us to apply similar techniques in the fundamental PDEs of systems and control: dynamic programming, Zakai equations for nonlinear filtering, information state equations for robust control and others. In addition Noether's theorem can lead to significant advances in nonlinear optimization. The results described in the subsequent sections are a small set of what could be accomplished by such methods in systems and control. 2 Constructive Use of Symmetry Simple Example Groups of PDEs: A
An interesting, for systems and control (as we shall see) theory, application of ideas from symmetry groups is the following. Use the symmetry group of
Group Invariance and Symmetries in Nonlinear Control and Estimation
143
a PDE to compute easily solutions to new PDEs. This is a nonconventional use of s y m m e t r y groups developed by Rosencrans [14]. To explain the idea clearly, we use the simple example of the heat equation. 0u(t, x) _ 0 ~ u ( t , x) & 0x 2
(13)
It is well known [7][12] that (13) is invariant under the variable transformation
x, ) e'~
,
]
t
) e~St.
l
(14)
That is to say if u(t, x) is a solution of (13), so is u(e~St, e'x). Clearly the initial data should be changed appropriately. So if r is the initial d a t a for u, the initial data for the transformed (under 14)) solution are r This elementary invariance can be written symbolically as
etDZeSXD r _ esxD eeZ'tD~f~.
(15)
Here 0 D: = Ox 02 ax 2 (16)
Often in this paper we shall give double meaning to exponentials of partial differential operators. Thus while exp(tD 2) in (15) denotes the semigroups generated by D ~ [13], exp(sxD) is viewed as an element of the Lie group of transformations generated by xD. It is easy to verify that
r
x) = [exp(sxD)r
,
(17)
where we view exp(sxD) as such a transformation, with parameter s. Now the association
(t, s) ~=, etD'e"'D
defines a two parameter semigroup with product rule (t, s )  ( t l , s,) := (tlexp(2s) + t, s + s,)
(18)
,
(19)
because of the invariance (15). A one parameter subgroup is t = a ( e x p ( 2 e , )  1) l, S cr J (20)
144
J o h n S. B a r a s
where a, c are positive constants and r > 0 is the group parameter. To this subgroup (18) associates the one parameter semigroup of operators
H ( r ) : = expa(ezP(2er)DD~er~D
_= eer~D ea(1ea~P(2ce)) D~
(21)
It is straightforward to compute the infinitesimal generator of H M e := l i m r ~ o H ( r ) r  r _ 2acD2r _ c x D r
r
.
(22)
But in view of (21) and (22) we have the operator identity
e Mt : e  e r t O e a(1e~'p(2et))D~
(23)
To understand the meaning of (23) recall that for appropriate functions r e x p ( M t ) r is the solution to the initial value problem 02
w(0, , ) = r
Then (23) suggests the following indirect procedure for solving (24): S t e p 1: Solve the simpler initial value problem
bTu( ,
0
=
= r
02
(25)
S t e p 2: Change independent variables in u to obtain w via
w(t, z) = u(a(1  e x p (  2 c t ) ) , e x p (  c t ) z ) .
(26)
Here we have interpreted the exponential in (23) as a transformation of variables. This simple example illustrates the main point of this particular application of symmetry groups: knowing that a certain partial differential equation (such as (13)) is invariant under a group of local transformations (such as (20)) can be used to solve a more difficult equation (such as (24)) by first solving the simpler equation (such as (13)) and then changing variables. This idea has been developed by S.I. Rosencrans in [8] [14]. It is appropriate to emphasize at this point that this use of a group of invariance of a certain PDE is not traditional. The more traditional use of group invariance is discussed at length in [7] [12], and is to reduce the number of independent variables involved in the PDE. Thus the traditional use of group invariance, is
(31) (30) . denoted by exp(tL). use this group to solve a "perturbed" parabolic PDE wt = (L + P ) w (28) by a process of variable changes and the possible solution of an ordinary (not partial) differential equation. linear. Note that G induces a group G acting on the space of functions on M with values in ]Rp. for at least small t > 0.~) = r in X.~(A) . denoted by ~ ( M . The element ~ corresponding to a g in G will m a p the function A into A ~. The point of the simple example above is to illustrate a different use of group invariance which goes roughly as follows: given a parabolic PDE ut = Lu (27) and a group of local transformations that leave the solution set of (27) invariant. bi. nondegenerate elliptic partial differential operator 0~ L := i. c are smooth enough. so that /: generates an analytic group [13]. Let V be the set of solutions to Ou 07 = Lu u(0. Consider the general. i=1 (29) and assume that the coefficients alj. The operator P will be referred to as the "perturbation". 1RP). on some locally convex space X of initial functions r and appropriate domain Dora(L).e.Group lnvariance and Symmetries in Nonlinear Control and Estimation 145 just a manifestation and mathematical development of the classical similarity methods in ODE. i. A' =. One of the contributions in this paper can be viewed as an extension of the results of Rosencrans to the stochastic partial differential equations that play a fundamental role in nonlinear filtering theory. 3 The Invariance Group of a Linear Parabolic PDE. as we vary r The aim is to find a local Lie transformation group G which transforms every element of V into another element of V. that is an i n v a r i a n c e g r o u p of (30) or of L.j=l aij(x)~ n ~xi + Ebi(x) + c(x)id .
x)) + r (33) where p is a transformation acting on the variables (t. using standard Lie theory notation. We shall. )cA+r (34) where Ae. T h e o r e m 3. For that we need: D e f i n i t i o n : G is linear if there exists a Lie group of transformations 2? : M + M such that for each ~eG. a p x p matrix "multiplier" v = v(x. x). } in G. by exp(sX) where X is the infinitesimal generator of the one parameter group {Xs }. i=1 (35) Indeed this follows from an expansionof exp(sX)(x. r a fixed solution of (30).u) O  ~#~(x. Given a function Ae. The main result of Rosencrans [14]. IR) let A(s. According to standard Lie theory the infinitesimal generators of these oneparameter subgroups form the Lie algebra A(G) of the local Lie group G [7]. x) :. Clearly the elements of A[G] can be considered as first order partial differential operators in IRn+l X = 7(x..IR). there exists a ae2?. We are interested in groups G acting linearly.e. concerns the case of a single parabolic equation (30). Because of (34) one says that G acts as a m u l t i p l i e r r e p r e s e n t a t i o n of . Let X be the infinitesimal generator of {Xs }. [1)A(a(~)) + r (32) The meaning of (32) is rather obvious.~) and a solution r of (30) such that ~(A) (x) = v(x. Baras It is easy to show [14] that G and (~ are isomorphic as groups. denote X. acting on functions. r a fixed solution of the parabolic equation. u) for small s.Xs(A)(x). Clearly for linear parabolic equations G is always infinite dimensional since it always includes the infinite dimensional subgroup 7f consisting of transformations of the form A. x))u(p(t.~'(IRn.146 John S.U upon the space of solutions of (30).T(M'. T h a t is we consider subgroups of G of the form {Xs } where s "parametrizes" the elements. The way G acts on functions is basically via the "coordinate change" group Z: of M. ~v(p(t. x) . Thus XeA(G). c a scalar r 0.u) O i . We consider now oneparameter subgroups of the invariance group G of a given partial differential equation.1 [14]: Every transformation ~ in the invariance group G of a linear parabolic equation is of the form u(t. (36) . i. Now {Xs} induces a oneparameter subgroup {X.
When ~ is linear this reduces to (40) D(A) = 0 :=~ D ( f ( ( A ) ) = O.A(x) + O(s2). .' ~ A ) = I)(e'fC (A)) (41) implies (40) if we set s = 0.s E i=1 or fli(x. (3s) In view of (33) the condition for 0 to be linear is that [14] fli.(x.~O A.A(O. (39) The best way to characterize G (or G) is by computing its Lie algebra A(G) (or A((~)). x ' ) s+0 ' 8 (Z') A(x') lim u s.A(x)) + i=1 fl.u' by a specific oneparameter subgroup exp(sX) of G we can expand u' = A(x) + s T ( x . A ( z ) + O(s 2) x i = xl . By definition X e A ( G ) if 7)(A) = 0 ==~ I)(e'Y~ A) = 0 for small s.. It is not difficult to show that (40) leads to a system of partial differential equations for 7 and fli.+O S = = = lim A ( s .A(x)) OA(x) Ox.A ( x ) s . Thus e (37) A(x') = A(x) .sfli(x. A(x)) + O(s 2) fC(A)(x) = lim A(s. A direct way of doing this is the following. S = 7(x. u are transformed to x~. x) s. since d D ( e ' . x) .Group lnvariance and Symmetries in Nonlinear Control and Estimation 147 If x i .~0 S lim A ( s .. x ) .~=7~u =0 .
x) = exp(sX)a. J. the solution of the ODE ~(t. . x) = ~xp( + where ~(~(.. another solution is (42) .u) = ~. u) = uJ(x) + r for some/31. x))dr)r x)). e(t.~ i. (45) clearly displays the linearity of G near the identity.(x) 7(x.~. The latter. x)) + r and therefore x)) A(. which satisfies O A(s. entails the following. x)dt qS(t. Then in view of (39 / ~i(x. Let e(t) be the flow of the vector field )"]~i=lnfli5~. x))A(s . z))dr) as the "multiplier" u. x) = x. e(t. fin] T.x) = J(x)A(x) + ~ .148 John S. Then if A A(s. Then from (43) d A(s .x) = A(x) in view of (38) and due to the linearity assumption (42/.'/A(~(s. Baras We shall consider further the determination of A((~) in the case when G is linear. since it is the only case of importance to our interests in the present paper. ~)1 (45) qS(t. ( x ) ~ i1 +r (43) A(O. The crucial point is that (43) is a first order hyperbolic PDE and thus it can be solved by the method of characteristics.e. (46) By comparison with (33) one can view e~p(f o J(e(r. very briefly. x) = exp( s J(e(r..t.~) d = ~(~(t. x)) = J(e(t. x)) (44) e(0.t. r Let us denote by fl the vector [r is a solution of (30).))d.'.~ f l .
+ i=l ~iu. (53) For ueV the second reads o. (50) a first order hyperbolic PDE (solvable by the method of characteristics). + .1. u + o~u. Now (51) indicates how to compute (~. x)id . fli.. (51) The converse is also true: if (51) holds for some first order partial differential operator. 4t ~ 0 + 4tx~~+ x 2.Group Invariance and Symmetries in Nonlinear Control and Estimation 149 The most widely known example.Namely 0 (N implies _L)u=O (52) n 0 (N _L)(out+Z i=1 Ou ~' ~~. ~ ) ~ i=1 0 " 0 + 3`(t. 3'. 7 of t and x. for which A(G) has been computed explicitly is the heat equation (13)..tu~. as follows 0 2t 0_~+ 0 0 (47) 1. (48) for some functions a. ~) ~ + ~ Z. (54) .. Let us apply these general results to a linear parabolic equation. 2t~x + x. + 3`ut = LXu. + 3`u) = o. Clearly since o _ L is linear (40) applies and therefore ZueV if ueV. tut + ~ i=l ~. fl.(t. (49) However v is also the solution of ~ v = c~)v + ~ 0 0 ~ i=1 / 3 iOxiv + 3`v  0 v(0) = u. for small s. The infinitesimal generators of G are given in view of (38) (42) (note that xl = t here) by z = ~(t. like (40).~. Z is a generator of G. From Theorem 3. If u solves (30) so does v(s) = exp(sZ)u. then G is linear. The infinitesimal generators in this case are six. y .
so does v(s) = exp(sZ)u b u t with some new initial data. First s+0 (58) lim R ( s ) r = r (59) Furthermore from (58) etLR(r)R(s)r = erZ etLR(s)d = erZ esZ etL r = e(r+')Ze'Lr = e'LR(. a(t. (56) d In (56) [ .. Thus if u solves (30).150 or John S.ZL)u .. to generalize the ideas presented via the example of section 2. We follow Rosencrans [8][14]. or (55) d Z=[L.) has the following properties. (57) 4 U s i n g t h e I n v a r i a n c e G r o u p o f a P a r a b o l i c P D E in Solving New PDEs. Furthermore it can be shown [14] that a is independent of x. ] denotes c o m m u t a t o r and ~ Z is symbolic of O at ~ i n i=1 0 Thus the elements of A(G) in this case satisfy a Lax equation. " + s)r (60) . say R(s)r T h a t is esZ e tL = etL R(s) on X. Baras d zu = (LZ. i. Thus we consider a linear parabolic equation like (30) and we assume we know the infinitesimal generators Z of the nontrivial part of G.l Z dt~T + dtl ~ + . It is immediate from (56) that Z form a Lie algebra. Now R(.e.Z] dt on V. + doZ = 0 where ~ < dimG. In this section we use the results of the previous section. x) = a(t) and t h a t every Z satisfies an ODE dt Z dt.
s_>0 (61) From (60). x)d~. Thus M e = a(0)Lr + i=l ~. u(O) = r . (62) It is straightforward to compute M. The most important observation of Rosencrans [8] was that the limit as t ~ 0 of the transformed solution v(s) = exp(sZ)u. given Z as in (48). S t e p 1: Solve ut = Lu. This leads immediately to the following generalization of discussions in section 2: To solve the initial value problem OqW  Os  ~ Mw r w(o) where = (66) M = a(O)L + Z ~ . Let M be its generator: Me= lim R ( s ) r 1 6 2 $ 8+0 CE Dom(M).Group Invariance and Symmetries in Nonlinear Control and Estimation Or R(r)R(s)=R(r+s) 151 for r. x ) ~ + 7(0. x)id i=l (67) follow the steps given below. + 7(0. R() is a semigroup. call it w.etLM on Dom(L). (0. (61). solves the new initial value problem Ow = Mw Os w(0) = r  (64) That is e'ZetL: or e t L e "M on X (65) Ze tL . ( O . x)r (63) Note that M is uniquely determined by the Z used in (58).
Z can b e d e t e r m i n e d f r o m M. Conversely suppose M is given. Note this step requires the solution of ordinary differential equations only. To see the relation between A(G) and A(P). by repeated bracketing with L all initial derivatives of Z can be obtained. that permit solutions of ut = (L + P)u to be computed from solutions of ut : Lu. The "perturbation" which is of degree < 1't. via (48). x) Oxi + 7t(0. (68) We shall denote by A(P) the set of all perturbations like (68). x). M2]r = etLM1M~r . observe first that each generator Z in A(G) uniquely specifies an M. From the Lax equation (56) we find that dZ dT I*=o = [L. M] = at(0)L + ~2~/3.etLM~MIr : ZletLM2r . Indeed let Zi correspond to Mi.152 John S. Definition: The Lie algebra A(P) will be called the p e r t u r b a t i o n a l g e b r a of the elliptic operator L. Z] I*=o: [L. Z2]etLr (70) This establishes the claim.x).Z2etLMIr = Z l Z ~ e ' L r . . S t e p 3: Set t = 0 to v(s. Since each perturbation P is obtained from an M by omitting the component of M that involves the unperturbed operator L. which we denote by A(M). by integrating only an additional ordinary differential equation.t(0. This procedure allows easy computation of the solution to the "perturbed" problem (66)(67) if we know the solution to the "unperturbed" problem (30). Since from (57) Z satisfies a linear ordinary differential equation. v(O) J via the method of characteristics. x) ~x{ + 7(0.). Thus given an M. (63). x)id.2. So there exists a 11 correspondence between A(G) and the set of perturbed operators M.Z ~ Z l e t L r = [Zl. id. i = 1. is given by the part of M: P = i:l o /3/(0. Baras S t e p 2: Find generator Z of G corresponding to M and solve ~ = a(t)0~ v + E?~ ~ (t. It is easy to see that A(M) is a Lie algebra isomorphic to A(~. x ) ~ v : u + 7(t. 0 i=1 (69) Note that the right hand side of (69) is another perturbed operator M'. Then from (65) we have etL[M1. z)v ]. We would like to show that A(P) is a Lie algebra strongly related to the Lie algebra A(G) of the invariance group of L.t..
One significant question is: can we find the perturbation algebra A(P) without first computing A(G)." . Moreover dimA(P)) = dim(A(G)) . the invariance Lie algebra? The answer is affirmative and is given by the following result [8].of order less than or equal to one such that [L. An operator P0 of order one or less (i. Indeed consider the class of linear parabolic equations ut = Lu. .and a sequence of operators /~. within each class (orbit) {L(k)} (k indexes elements in the class) solutions to the initial value problem u(k)t = L(k)u(k) with fixed data r (independent of k) can be obtained by quadrature (i. of the Lie algebra of the invariance group of L).~x + (cx ~ + dx + e)id (72) where a. or equivalently linear parabolic equations are divided into equivalent classes (orbits). along the lines suggested here. e are arbitrary constants. x ~. It is an easy application of this result to compute the perturbation algebra O~ of the heat equation in one dimension or equivalently of L = ~:x" It turns out that A(P) is 5dimensional and spanned by (9 A(P) = Span(l. P ~ .1: The perturbation algebra A(P) of an elliptic operator L. ~0x ' X~xx)" So the general perturbation for the heat equation looks like (71) P = (ax + b).e.e. T h e o r e m 4. As.. We can define an equivalence relationship on this class by: "L: is equivalent to L2 if L2 = L: + P where P is an element of the perturbation algebra A:(P) of LI". Thus elliptic operators of the form (29). Moreover the dimension of A ( P ) i s one less than that of A(M). c.1. . where L is of the form (29). We close this section by a list of perturbation algebras for certain L. b. . from [8]. is isomorphic to a Lie subalgebra of A(G) (i. It is straightforward to rework the example of section 2. n>_O and ~ Akt~/k!.2: Assume L has analytic coefficients. x. In view of the isomorphism of A(M) and A(G) we have established [8]: T h e o r e m 4. Note that the invariance group of the heat equation is 6dimensional (47). . Pn]=AnL+Pn+I . The implications of these results are rather significant. ~ Pkta/k! converge at least for small t.Group Invariance and Symmetries in Nonlinear Control and Estimation 153 it is clear that A(P) is a Lie subalgebra of A(M). an ODE integration) from any one solution u(k0). d. of the form (69)) is in the perturbation algebra A(P) of L if there exist a sequence of scalars A:.e.
xl. Baras Elliptic operator Generators of perturbation L algebra A( P) 1. e x. Examples of perturbation algebras. x) f u(t. see [9]). (74) Here w(t). (log x)". x. so called unnormalized conditional density. We will only briefly discuss the focal points of our current understanding of the nonlinear filtering problem and we will refer the reader to [9] or the references [2][6] for details. x). D. x .xD eZD ~ I.v(t) are independent standard Wiener processes and h is such that y is a semimartingale. x z. Let us denote this conditional density by p(t.154 John S. Thus the "nonlinear filtering problem for diffusion processes" consists of a model for a "signal process" x(t) via a stochastic differential equation dx(t) = f ( x ( t ) ) d t + g ( x ( t ) ) d w ( t ) (73) which is assumed to have unique solutions in an appropriate sense (strong or weak. xD. The problem is to compute conditional statistics of functions of the signal process r at time t given the data observed up to time t. In addition we are given "noisy" observations of the process x(t) described by dy(t) = h ( x ( t ) ) d t + dv(t). o < s < t} (75) Clearly the m a x i m u m information about conditional statistics is obtained once we find ways to compute the conditional probability density of x(t) given ~t. It is more convenient to use a different function.e. z)dz (76) . the aalgebra J 7 = a{y(s). x D x 2D ~ x log xD. D T a b l e 1. x) which produces p after normalization p(t. x) _ u(t. u(t. x D D~ xD 2 1. log x. 5 Strong Equivalence of Nonlinear Filtering Problems In this section we will apply the methods described in sections 24 for parabolic equations to the fundamental PDEs governing nonlinear filtering problems. As with all symmetry group methods these techniques have a strong geometric flavor. 1 x ~D ~ 1. i.
which in It6's form is du(t.lhTh (80) (81) (82) Bj : = Mult. by hj(jth comp.g) = (• . In (78) the matrix a is given by ~(z) = a(z)a(x) T . (79) and we shall assume that a is positive definite.lh(x)Th(x))u(t. j = 1.e. This is the so called MortensenZakai stochastic PDE. the elliptic operator L: is nondegenerate. Bj have a common. dt j=l (83) We shall assume that every equation of the form (80) considered has a complete existence and uniqueness theory established on a space X. . x) + hT(x)u(t. The estimation Lie algebra introduced by Brockett [2] and analyzed in [2][6] is the Lie algebra A(E) = Lie algebra generated by A and Bj. (84) Again we shall assume that for problems considered the operators A. i..Group Invariance and Symmetries in Nonlinear Control and Estimation 155 The reason for the emphasis put on u is that it satisfies a l i n e a r stochastic PDE driven directly by the observations. Letting A : = f~. Furthermore we shall assume that continuous dependence of solutions on y() has been established. x) = s x)dt + hT(z)u(t..).. z) dy(t) at dt We shall primarily work with (80) in the present paper. dense invariant set of analytic vectors in X [10] and that the . When applying geometric ideas to (7T) it is more convenient to consider the Stratonovich version Ou(t.p. of h) we can rewrite (80) as an infinite dimensional bilinear equation P du(t) = (A + ~ Bj~/j(t))u(t). x)dy(t) (77) Here s is the adjoint of the infinitesimal generator of the diffusion process 1 n 02 which is also called the FokkerPlanck operator associated with z(.
e. and vice versa. Type 3 operations are introduced here for the first time. x) (85) & 1 . We are thus dealing with two parabolic equations. the solution of the other can be obtained by relatively simple additional computations. via the following types of operations: T y p e 1: (t.x) = r where r > 0 and r > 0. We shall then apply this result to the examples studied by Bene.x)+ hiT(x)ui(t. i = 1. To make things precise consider two nonlinear filtering problems (vector) dxi(t) = fi(xi(t)dt + gi(xi(t))dwi(t) dye(t) = h~(xi(t))dt+dv~(t) . i = 1. yj (t))ui . 2 j=t (87) where A i.~ II hi(z) 112)u~(t. created certain excitement for the possibility of a complete classification theory.(t: ~ . Our approach starts from the abstract version of (86)(i. T y p e 2: u2(t. 4 apply. Examples discovered by Beneg [11]. In this section we give an extension to the full equation (87) under certain conditions on Bj. (83)): Oui P Ot = (hi + ~ B j . i = 1. where a is a diffeomorphism. x 2) = ~(t. This is a classical problem and the methods of section 3. Bj are given by (81)(82).2 and the corresponding MortensenZakai equations in Stratonovich form aui(t. Equivalence here carries the following meaning: two nonlinear filtering problems should be equivalent when knowing the solution of one. We develop a methodology for recognizing mathematically "equivalent" problems. and will be seen to be the key in linking this problem with mathematical work on group invariance methods in ODE and PDE's.~ and recover the Riccati equations as a consequence of strong equivalence. has analyzed the effects of diffeomorphisms in zspace and he and Mitter [4] the effects of so called "gauge ~' transformations (a special case of our type 2 operations) on (80). Brockett [2]. We will first examine whether the evolutions of the time invariant parts can be computed from one another. xt). quadrature).x)yi(t).e.156 John S.2(86) D e f i n i t i o n : The two nonlinear filtering problems above are s t r o n g l y e q u i v a l e n t if u2 can be computed from ut. T y p e 3: Solving a set of ordinary (finite dimensional) differential equations (i. . Baths mathematical relationship between A(E) and the existenceuniqueness theory of (80) is well understood. We shall see how transparent Beneg' examples become from the point of view developed in this paper.
T h e o r e m 5. 5/. (where A(G) is the Lie algebra of the invariance group for the class) and PI~eA(P) (where A(P) is the perturbation algebra of the class) such that (see (65)): A 2 ____ A 1 q. we then solve the ODE associated with the characteristics of the hyperbolic PDE Ov = Z12v Os and we have (91) e 'A' = [ez' e 'a'] I.= (A' + " Bj' 'j(t))u. For piecewise constant the solution to any one of the PDEs in bilinear form is given by Ui : e(Ai+B'j~.~) 9. by the additional computation of solutions of an ODE.).n)(t~t~. i=1. t ..A 2. e' . we first compute the semigroup generated by A l. Then the two filtering problems are strongly equivalent.1: Given two nonlinear filtering problems (see (85)). p l ~ (89) eSZX~etAl : etAleSA2. We know by now what (89) means: to compute the semigroup generated by A 2. continuously dependent on y(. i = 1. are nondegenerate elliptic. e(A'+B'J~_l~.. .. p . belong to the same equivalence class there exist ZI2eA(G). P r o o f : Only a sketch will be given here.9 i=1. (92) .1.2 Since A1. One first establishes that is enough to show computability of solutions for piecewise constant ~. j = l ...). such that the corresponding MortensenZakai equations (see (86)) have unique solutions.Group Invariance and Symmetries in Nonlinear Control and Estimation 157 Our main result concerning equivalence (in a computational sense) of two nonlinear filtering problems is the following. i = 1..=0. from one another.1 . Assume that using operations of type 1 and 2 (see definition just above) these stochastic PDEs can be transformed in bilinear form au. (88) ~A'IB~ f~ )t. s >_ O. belonging to the same equivalence class (see end of section 4) (il) Bj.~_l)(tm't~~).. (90) T h a t is consider A ~ as a "perturbation" of A 1. .2.2 j=l such that: ( i ) A i. 2 belong to the perturbation algebra A(P) of (i).
. +Bj) _~ et(A'+B~)e'(A~+B~. and is an example of such more general transformations needed for systems and control problems and discussed at .B~ = A ' I.p. eSZik et(A . (93) It is now apparent that if we know (88) explicitly for i = 1. (96) We wish to show that (95)(96) are strongly equivalent only if f (the drift) is a global solution of the Riccati equation f~ + f2 = ax~ + bx + c. A 2 k Bk belong to the same class there exist 2 ZJ~eA(G). P ~ e A ( P ) such that A2I.. f and the nonlinear filtering problem (scalar x. dy(t) = x(t)dt + dr(t).)exp( f f(u)du).x2)u2(t.2(0z . . j = 1. x). That is let (defines v2) u~(t.. y) dx(t) = dw(t) "[ dy(t) x(t)dt + dr(t).).158 John S. z) . f The corresponding MortensenZakai equations in Stratonovich form are: for the linear Oul(t.x) 0t  1 2"0z(  X2)ux (t. k .~ ..BJ. y) dz(t) = f(x(t))dt + dw(t) 1. Os This completes the proof. we obtain u2 from (93) with the only additional computations being the integration of the ODEs associated with the characteristics of the hyperbolic PDEs Ov = ZJ~v. (95) for the nonlinear Ou~ & 1 .~) = ~.~ ( f u ~ ) + xl)(t)u2(t. (94) Let us apply this result to the Bene~ case.~(t. Baras More generally since A 1 I.. (98) The transformation (98) is global. x) + XiJ(t)ul (t. We consider the linear filtering problem (scalar x. (97) First let us apply to (95)(96) an operation of type 2. X).P~: .B) t.
d.1. in view of (100).e. For A 2 to belong to A ( P ) it is necessary that V be quadratic. 02 From the results of section 4. uniqueness and continuous dependence on y() for (95)(96) have been established using classical p. when (97) is satisfied is given by %~ where ) (lO5) dr(t) do(t) = (a + 1)o(t)r(t)dt . (100) Existence.r ( 0 ) 2 ) 2a(t) (103) dr(t) = o(t)(ay(t) . results. (106) 1 .1 to (95) (99). A 1. We apply T h e o r e m 5. or a potential transformation in s y m m e t r y group theory [1]. which is the same as f satisfying the Riccati equation (97). B 1._2 (O~x ~ X 2 .2o(t)bdt + o(t)dy(t) = 1 . So A1 As while B 1 = B 2 = Mult. Clearly from (72) or Table 4. using a path integral c o m p u t a t i o n showed that the solution of (99). (99) V(x) = f~ + f2. This is like a gauge transformation from m a t h e m a t i c a l physics. (102) = ~(b~ .X) at where 1 02 . Then the new function v2 satisfies OV2(t. B 2 e A ( P ) for this class.V ( x ) ) v 2 ( t .(a + 1)a2(t).r(t)at). the only possible equivalence class is t h a t of the heat equation.o2(t).x) = exp( where (x . Recall that the solution of (95) is ul(t.~) 1 02 = ~ ( ~~x~. by x. x) + z y ( t ) v 2 ( t . r(o) = do(t) = 1 .Group Invariance and Symmetries in Nonlinear Control and Estimation 159 the end of section 1. x).V) (101) 1.x 2 . o(0) = 0 (104) Beneg [11].
u) + ~ I Dp is +L(x. The dynamic game representation of the equivalent nonlinear robust control problem is as follows. A key result in establishing these equivalences was the introduction of the information state and the nonlinear PDE that it satisfies. from the point of view that strong equivalence of the linear and nonlinear filtering examples implies the Riccati equation. is when can we reduce this nonlinear problem to a linear one? Group invariance methods can be applied to this problem. Given the dynamical system ~(t) = b(x(t). which is a nonanticipating functional of y(. u) .). u(t)) + w(t).~) + ~5+ ~F(x) = 89 89 Let us next consider the information state PDE for (107)(108): Op = np.). v are LStype disturbances.89 ~TY~(x.160 John S. We also maintain that knowledge of group invariance theory makes the result immediate at the level of comparing (95) with (99). the nonlinear output robust control problem (in an H ~ sense) is equivalent to a risksensitive partially observed stochastic control problem and to a dynamic partially observed game [20][26].+ (y(t) .Cx) s tt O't (110) . u(s)) (108) _ 1 (j w(s) Is + ]v(s) I S ) ] d s + r One of the questions we want to answer. y(t) = Cx(t) + v(t) x(0) = X0 } (107) where w. As has been developed fully in [24 ]. (A(u)x + B(u)) = 1 T A(u) + 89 (109) f(x) = . Baras What we have shown here is a converse. 6 Reduction Problems of Nonlinear Output Robust Control In this section we will apply methods from s y m m e t r y groups to a problem of recent interest in nonlinear control: output robust control. In this section we apply systematic methods from symmetry groups to this fundamental PDE of nonlinear robust control. We want to find a control u(. Let us make the following structural assumptions: b(x.u) = f(x) + A(u)x + B(u) f(x) = DF(x) for some F 1 I f(x) is + f ( x ) .b(x. to minimize (p > 0) J(u) = sup sup sup {f(x0) + weL ~ wL ~ XoeL 2 [L(x(s).
(p.Group Invariance and Symmetries in Nonlinear Control and Estimation 161 The optimal control is a memoryless function of the information state since the cost (108). Under an appropriate transformation the dynamic game (107)(108) becomes linear..(C~(t)) Iz r = r A consequence of this is that the robust output control for the nonlinear system (107) is finite dimensional and easily implementable. quadratic (and therefore has a well known finite dimensional controller) with state p(t) = (5:(t). Under the structural assumptions (109). (1~ x (^t ) T ~. Y(t).Y ( t ) ~~(u(t)))~(t) + B(u(t)) .~ ( x .~)Ty1(x_ ~) + ~ + ~F(~) i (116) Similar reductions can be obtained.q) = s u p { p ( x ) . and cost J(u) = sup{r y e L :~ r (114) p(0) = p} (115) where @(P) = (P.q ( x ) } yeL ~ x ( ~ '~ . risk sensitive stochastic control problem.Y ( t ) ( c T c .(u(t))k(t) + A(u(t))k(t) _ L l y ( t ) .~):Po =/5}. under similar assumptions. it is not hard to show [21][27][28] that the PDE (110) has a finitely parameterizable solution pt(x) = ~~ (x . These .i ( x where .. can be expressed using the information state as follows [23][24]: d(u) = sup{(PT.~(t)) + r + F(x) (112) k(t) ~(o) = (A(u(t)) + IJY(t)Q(u(t)) .~ ( t ) ) T y ( t ) .l .r I pp = . The explanation for these specific results becomes clear from an invariance group perspective.C$(t) ) = ]z(t) = Y ( t ) A ( u ( t ) ) T + A ( u ( t ) ) Y ( t ) . (111) T h a t is why the information state P D E (110) is so fundamental. for the associated partially observed.#Q(u(t)) + ~ ( u ( t ) ) ) Y ( t ) + I Y(O) = Y (113) ~b(t) = 89 + 89 + 89 1 .Y (t)A(u(t) ) + Y (t)CT (y(t) .
linear. as in sections 3. 4. Requiring problem (107)(108) to be the nonlinear equivalent to a linear quadratic problem (from the perspective of equivalence of the corresponding information state PDEs.b(t) and h . Pt. understood and generalized by studying the group invariance of the information state PDE.W 2 4. A . Baras results can be obtained. ' " ..W A . (121) We now turn to the Lie transformation theory for the Information state PDE. The information state p(t. Specifically the structural assumptions (109) are completely analogous to the Beneg structural assumptions.p(t. James and Yuliar [21] point out that there is a solution of the form p(t. ~) = (~:  .. Xn) for the linear control problem satisfies the scalar P D E F ( t . in other words. and in particular using the methods of sections 3 and 4.h(t) are nvectors. and I . and constant in x we obtain the ODEs W . = rT(WAATwG)r/2rThl (119) (120) . we abbreviate to A ( t ) .(t))/2 + r (118) with W symmetric. ( A x + b) IVpl2/2 + xTGx/2 + h. x) . substituting in (117) and equating coefficients of terms quadratic. x + l = 0 . x l .Wb + W~/2) The last two equations can be rewritten as § = Ar+bWl(Gr+h). + WbATWrW2r. (117) where G .l. We shall omit the control parameter u from the notation.162 John S.G ( t ) is a symmetric matrix.(t))r w(t)(~ . x. r = wl(Wr h) . instead of writing A ( u ( t ) ) . since it plays no part in the following calculations.) implies the structural assumptions (109). We consider the invariance of (117) under an infinitesimal transformation given by a vector field of the form (note that we are not including a O/Or term) xi=l x (122) . Vp) = Pt + V p .A ( t ) is a square matrix.A T w . = ~r(w§ + W~/2 . Thus it is important to study the symmetry groups (invariance properties) of the information state PDE for the linear control problem.G .l(t) is a scalar function. Taking the gradient of (118). b .
(Vp.. pt. V p ) ~ DtV  (124) (125) ~~ ( D t ~ J ) p j . V)~) .pt.( V p .( V p .~ j=l /= 1.Ozi . D' = ~x i + P. V)~) . .=1 ap. ~t) = .1.V p ) ./~p = 0. the criterion for invariance is that X ( 1 ) F ( t . p . (123) where X (1) is the first extended infinitesimal generator.~ 4.PtVp .G x 4.tit 4ptrlp .?p) ~ = ( A z + b . z.V p ) .( V p . r/(ol) (t. " OF i=1 OF . ~Tp) ../~i) OF = Vt }. (V~l + ~lpVp . . i=l G~Pi  i=l j1 (Az + b .( A z + b . pt. ( V r I + ~lpVp .n.. where ~ + v ~ + ~ + . Theorem 4.0 whenever F(t.h) .x.. Adding up all these terms shows that (122) gives (AT (Vp) A. pt. z . Vp) .11.p. namely . (126) (127) Evaluating the entries of (122) term by term. o Pi =.=..D : I . rl~l)(t.Group Invariance and Symmetries in Nonlinear Control and Estimation 163 According to Bluman and Kumei [1]. n..~p. 0 0 0 0 Vt = ~ + Pt ~p. i=l. x. V p )  O. j=l (Di~J)pj.. ~t) .
x) = x . t.V p ) . .2: Assume p satisfies (I17).}). t.r p ) .~(~.p). Therefore we have established the following: (132) T h e o r e m 6. x.p)) = (t. if unique.O(t.t. so that (117) holds with (z.1 ( F u n d a m e n t a l T r a n s f o r m a t i o n R e l a t i o n ) : The vector fields ~ and 0 of the infinitesimal generator of a s y m m e t r y group of (117) must satisfy: 0t + (Ax + b .~(E. Hence for any choice of ~ we m a y solve for O by the method of characteristics. By definition of d~ d'~ (~) = X~(t.Vp).p) replaced by (~. where p is given by (118).parameter family of transformations of the information state variable p. x. x ) is a oneparameter family of transformations of the space variable x satisfying the system of ODEs (130. Let ~(~. t .164 John S.p)) denote the flow of the vector field Xi1 r ~ +ON (129) where ~(E. }(0. ~.t. Suppose ~ + ~ ( e . v 0 + (pt + (Ax + b . for some choice of ~ =_~(t. Baras Grouping terms. ~ + (A~ + b .V ) ~ .t.~).x. p) to the ODE (132). Then the solution E r p(~. we obtain: T h e o r e m 6.(ATXYP+ Gx + h ) . and Vp and Pt are quadratic in z.p) = p . T h a t is given ~. XTp)0p = V p . ~) is the information state for the transformed problem. We describe next how to use the fundamental transformation relation.~.t. is a one. x. ~(0. p) /~(t. x.p))  (t. (t. x.x). (1~8) Note that only l i n e a r differential operators acting on ~ and 0 are involved. ~). 6 . (131) together with the scalar ODE 0~ cg'e = q(t'~'P)' ~(O. (130) This breaks down into the system of ODEs = ~(t. solving for O involves only the solution of an ODE.p) is chosen to satisfy the Fundamental Transformation Relation (1~8) in terms of ~.~). and that O .t. ~. ( V p . t. x) are the transformed state space coordinates and i5(~.
24)implies xTI~v~ . (135) (136) where the last identity follows from (119) and (120)(121). rip . Define the following vector functions in terms of quantities determined above: ~ ( t ) .( A T + W ) W r + h Wb = ~t . V)~ .~ .x T w ~ ..( W x . .W x .t .( A T V p + G x + h) 9 +(Ax + b.x r w ~ 9 (133) This implies tit = .I / V . We shall constrain the choice of y so as to satisfy '1 = . ((Wx. F ( t ) = A + W.I/Vx  + b + W(~ . (ATW(x W(A~ ((Vp+ Wx).h + Wb . (Vp. V)~ .A T W + G . by (119). V)~ . since Wi'+Wr = WAr+WbGrh+ (WA . ~ . Now (6.W ~ . Therefore we consider a more specialized situation admitting explicit computations. .f.~Tp)) .2 is that it is too abstract to be of immediate practical use. V)~ +WE) (134) = V p . c~(t) .W ( A x + b .(Ax + b . Vr/ = . V)~ .W r .ATw . 5 = (Ax + b'Up). Rearranging terms gives ( A T ~ T p + G x + h .~))) .W2r .Group Invariance and Symmetries in Nonlinear Control and Estimation 165 7 A Case Explicitly Computable The drawback of Theorem 6.Vp). The coefficient o f x in the first bracket is ..Vp) .( A T W ) r .( W ~ ) .r) + G x + h .z T w ~ t . ~t = (Ax + b + W(x r)).W 2 + G)r .x r w ~ t . ((Wr). 7(t) = b . . (Vp+ Wx).W ( A + W ) = 0.O.
Suppose ~ is a polynomial of order N in x.(r~ + 7). .l . insert this solution in the right side of (141)./3 9~ V p .(k) (140) t + k~(k)(F(")  . . vr k=l k(k)(0( 9 ~*(~~)).( k + 1 ) ~ ( k + l ) ( 7  for k = 1 . ~ . ~.1: If we assume r / = xTW~. i. Now (138) becomes N N Z~k)x 0 ~ 1 k(k)((r~+7)O~*(k1)) = 0 Equating coefficients for each power of x forces the {~'(k)(t)} to satisfy the following system of ODEs: =(N) + Ns(N)(F(")  ") = 0 . solve (141) for s ( N .r/must satisfy the linear first order PDE (137) = rTW~ = r + ((r~ + 7).1 ) ( t ) .166 John S.. ~ . and so on. .(~ Thus we have established. = ~(1)(7). Baras Now the linear PDE which ~(t. (141) =(0) . This can be put in an even more concise form: T h e o r e m 7. t .v)~ = o. down to Z. N . (/3. x) must satisfy is: /3. N (138) r E k=0 ~(k)(t)(xC~k)' (139) where x  1 7 4 1 7 4 Then N (k factors). . then ((t. ..(N)(t). k)tensor.e. x) . a n d 3 ( k ) ( t ) i s a symmetric (0. (142) Notice the structure of this system of ODEs./ 3 . Suppose 3 ( ~ ~(N)(0) have been chosen. First we solve (140) for ~.v)r = 0. 2 . .
(k)(t)}. Pick an initial condition ~ ( 0 ) ( 0 ) .z) = x 9 (146) (derived from (131)) to determine ~'(E. let us assume that N r = rTw~ = Z k=O "~(k)(t)(x " (143) Then ~(t. ~(t. t. o(")(t)  (14s) Additional examplescases with explicit computations can be found in [29]. ~(0. Finally we determine p(e.x T w ~ = Vp 9~ . Now pick N ~(t. and solve for ((t. when n = 1 and W r is never zero.x) is completely determined by the initial conditions ~(~ ~(N)(0) and ODEs (140)(142). ~(t.~. ~:) . .Group lnvariance and Symmetries in Nonlinear Control and Estimation 167 T h e o r e m 7.~. . ~) are now explicitly computable.. .(.p). [18]. assuming ~(t.t. ~).x T w ~ = Vp" ~ . x. Or (147) which can be written out in full as N Oe = k=O (vp. x) using (140)(142) by solving for each of the {~. t. x.x) = E k1 O(k)(t)(x in other words so that (144) so that rTW~ . we solve the system of ODEs 0e = ~(t. As before. p(O.p) by solving the following first order PDE (derived from (131)(132)) by the method of characteristics (see Abraham et al.~(t.. The startingpoint is the solution p given by (118) to the linear control problem. .p) = p . o(k)(t) = (145) Now we repeat the steps described at the end of section 6.2: In the case when r / = . ~) and ~'(t. In particular. x) is uniquely determined by ~(0.~). (129)(132). x) is a polynomial of arbitrary degree in x with coefficients depending on t. E (N) (0). p. t. x). x. Thus ~(t. 287): ~5c3 __c_~ = V p . Finally we describe the procedure for computation of the transformed information state. under the assumption 7/= .
Clark.S. 18th IEEE Conf. Bluman and J. "Output feedback risksensitive control and differential games for continuoustime nonlinear systems". 317329. 11.N.:The Mathematics of Filtering and Identification. Brockett and J. 39. and their Appl. and Control. James and J. 1974.C.K.C. Elliott. 15. "Robust Hoe output feedback control for nonlinear systems". 1977. M.. M. James. 56. "Filtering Theory and Quantum Fields" presented in the C o n f . New York. Springer. G. Arch. 16. 61. L. Hazewinkel and J. "A nonlinear partially observed differential game with a finitedimensional information state". 1979.R. Elliott. A n . 6. of N A T O A S I on Stoch. Prec. M. pp. IEEE Trans. M. Anal. 1962. R. 1996. San Antonio. Ann. pp.R. 6. Tensor Analysis. June 1995. and Applications. of Math. R.M. 9. and Appl. "Analytic Vectors". M e t h o d s in S y s t e m T h e o r y ... 4.1. Darling (1994): Differential Forms and Connections.W. J. 40. "Exact Finite Dimensional Filters for Certain Diffusions with Nonlinear Drift". 1978. 7. Marsden. and Appl. p. 1957. IEEE Trans. New York. Math.W.. in Asterisque. on An. "Perturbation Algebra of an Elliptic Operator". J. SIAM. 1978. S. France. on Automatic Control.S. pp. Rosencrans. Shiryayev. Oxford. 18. 10071017.J. 20. 5. April 1994.l. LieBacklund Transformations in Applications. SpringerVerlag. Publ. Vol. Yuliar. James. 157171. Sib. Mitter. Statistics of Random Processes [. Bene~.V.. Bordeaux. 2nd ed. 8. Ratiu (1988): Manifolds. V.L. Similarity Methods for 'Differential Equations. S. 19. "Classification and Equivalence in Estimation Theory". 328368. in Richerche di Automatica. a n d G e o m . R. pp . Prec.. SpringerVerlag. Rosencrans.A. Baras. 14. Cambridge University Press. 1976. 172175. F u n e t . J. Syst. AMS Coll. Ovsjannikov. 1979. pp.S. Prec. 12. "Geometry of the Conditional Density Equation". 1980. Branch of the USSR Ac. Prec.R. 10. No.168 John S. R. Baras References 1. J. .E. Mitter. pp. 1959. of Sci.W. Baras and R. 32nd IEEE CDC. S. 1980. Systems and Control Letters. Nariboli. 1993. James and S. Liptser and A. and T. on Dec. Springer. 1977.K. S..H. Int. Sophns Lie. Vol VI. o n Alg. 2.J. Kumei (1989): Symmetries and Differential Equations. R. E. and Opt. "Risksensitive control and dynamic games for partially observed discretetime nonlinear systems". Les Arcs. No. Kristiana. "On the Analogy Between Mathematical Problems of Nonlinear Filtering and Quantum Physics". 4. 3. 780792. 13. 17. Bluman and S.R. Vol. Ibragimov. 22.W. 21. Prance.R. Syst. Abraham. 70. Nelson. Equations.S. Hille and R. 1881. G. 3. G. in Stochastics. a n d S e m l g r o u p s . E. Stoch. Phillips. on Automatic Control. for Math. 1980. of Math. No. pp. England.W. of Stoch. 23. New York.. Cole. An. Brockett. Anderson and N. M. J. WiUems (Edts).572615. 537551. 5. R. Conf. Group Properties of Diff. 1980. Baras and R.
S I A M Journal on Control and Optimization. Baras and R. pp.S. M. Baras. pp. "Finitedimensional optimal controllers for nonlinear plants". J. and Control. July 1996. 7. pp. Darling. January 1998. No. 3. and Nonlinear Hoo Control". 13421364. 1998. 223230. James and J. 6175. 29. Baras. . 28. Moore and J. 26. Bensonssan: "General FiniteDimensional RiskSensitive Problems and Small Noise Limits". Systems and Control Letters. J. Vol.R. 343348. "Robust control of setvalued discrete time dynamical systems". No. 210215. pp. James. Baras and N. IEEE Trans.B. 43. J. "FiniteDimensional Methods for Computing the Informatin State in Nonlinear Robust Control".R. infinite dimensional HJI Equations. R. Vol. of 1998 IEEE Conference on Decision and Control. Florida. 27. 41. J. No.S.Group Invariance and Symmetries in Nonlinear Control and Estimation 169 24. 1997. IEEE Transactions on Automatic Control. on Automatic Control. Vol. Vol. Tampa.S.S. 1. Proc. Baras and M. Estimation. 4. Journal of Mathematical Systems. "Partially observed differential games. 26 (1995). 25.S. Elliott and A. 34. "Robust and risksensitive output feedback control for finite state machines and hidden markov models". January 1996. pp. Patel. pp. 371374.
The hospitality of the members of the CAS and the high quality of the coffee are gratefully acknowledged.de.10. A permanent email address is fabian@math.g. de Centre Automatique et Syst~mes Ecole des Mines de Paris 77305 Fontainebleau. 28334 Bremen. Germany behrens@orthogon. fr** A b s t r a c t . ** This paper was written while Fabian Wirth was on leave from the Zentrum ffir Technomathematik. A f u n d a m e n t a l question in this area is precisely the question of the u n d e r l y i n g s o l u t i o n concept for which the constructed feedback should be i n t e r p r e t e d .15]. Often s a m pling concepts are considered.unibremen. see e. 1 Introduction It is the a i m of this p a p e r to present a p r o c e d u r e to c o m b i n e local s t a b i l i z a tion procedures with global ones to o b t a i n a g l o b a l l y defined feedback with desirable p r o p e r t i e s near the fixed p o i n t t h a t are designed using i n h e r e n t l y local arguments. Universits Bremen. It is known t h a t a s y m p t o t i c n u l l c o n t r o l l a b i l i t y is equivalent to the existence of a control L y a p u n o v function.9.5. Germany. b u t unless a d d e d s t r u c t u r e like h o m o g e n e i t y * Research partially supported by the European Nonlinear Control Network. see [16]. and in recent years n u m e r o u s p a p e r s have a p p e a r e d on t h e question on how to c o n s t r u c t s t a b i lizing feedbacks from such functions. In a neighborhood of the singular point we use linearization arguments to construct a sampled (or discrete) feedback that yields a feedback invariant neighborhood of the singular point and locally exponentially stabilizes without the need for vanishing sampling rate as the trajectory approaches the equilibrium. .6. The design procedure is twofold. [1. On the remainder of the domain of controllability we construct a piecewise constant patchy feedback that guarantees that all Carath6odory solutions of the closed loop system reach the previously constructed neighborhood.A Globalization P r o c e d u r e for Locally Stabilizing Controllers * Jochen Behrens 1 and F a b i a n W i r t h 2 ] Orthogon GmbH Hastedter Osterdeich 222 28207 Bremen. ens~p. France virth@@cas. For a nonlinear system with a singular point that is locally asymptotically nullcontrollable we present a class of feedbacks that globally asymptotically stabilizes the system on the domain of asymptotic nullcontrollability.
A question that is unfortunately neglected in the other references on this subject. [13] somewhat imprecise. This line of thought can be found in [12. the results in that paper do not provide feedback invariant sets. For our local considerations we will rely on the work of Griine. The problem of measurement noise. respectively Ancona and Bessan [1]. The ensuing Section 3 details a special class of Carath~odory feedbacks. in [8] Griine also shows that this procedure works locally for general nonlinear systems if the linearization is asymptotically nullcontrollable. Outer perturbations for the local and global procedures are discussed by Griine [8].1]. Again we will use tools from nonsmooth analysis which has the advantage that we are able to treat a more general case than the one in [1]. We do not discuss perturbation problems in this paper.3.13. persists. In Section 4 we need additional assumption on the system and on the linearization in the origin. which we will need in order to construct well defined feedbacks. Also the local construction is computable as discussed in [9]. We follow Ancona and Bessan [1] as their approach has the advantage of guaranteeing properties of all Carath6odory solutions generated by the discontinuous feedback. which renders a neighborhood of the fixed point feed . we will employ ideas from nonsmooth analysis already used in [5. In the following Section 2 we define the system class and make precise what we mean by a feedback in the Carath6odory and in the sampled sense. Here. The key in this definition is a nonsmooth "inward pointing" assumption which guarantees that there are no solutions with behavior other than the one intended in the construction of the patches. [3.9] to regularize a known control Lyapunov function. We quote some recent results showing that we can always construct a patchy feedback controlling to an open set B from its domain of attraction. this also leads to local stabilization results using homogeneous approximations in [11] which implicitly provide feedback invariant sets.172 Jochen Behrens. In fact. The local method will be combined with ideas that are inherently global in nature and that all depend in one way or another on the construction of piecewise constant feedbacks. who has shown for linear and more generally homogeneous systems how to construct discrete feedbacks that stabilize the origin where the sampling rate can be chosen to be positive [9. Chapter 12].From a practical point of view. the sampled feedbacks often require vanishing sampling intervals as the trajectory approaches the origin. We then construct a feedback in the sense of sampling for the nonlinear system. see the discussion in [15]. this appears to be undesirable./. Fabian Wirth is used [9]. however.10]. The theoretical interest in our procedure is that one obtains in this way a construction of a feedback that globally stabilizes the system without the need for increasingly faster sampling near the origin to ensure convergence. for proofs which are omitted for reasons of space we refer to [4]. Unfortunately. In a very similar vein. that are defined by patches in a piecewise constant manner. which makes some statements in [12].
u). Consider a feedback law F : ~'~ + U. = f ( x . if we consider C a r a t h ~ o d o r y solutions of (2). A n infinite sequence 7r = (ti)i_>0 with 0 . In t h e final Section 5 we show t h a t the previous constructions can be used in a c o m p l e m e n t a r y fashion. d(rr) : = s u p A i . T h e m a p F is called a Cfeedback. .f ( x . If F is not continuous (as will be the case in the feedbacks we a i m to construct) this i m m e d i a t e l y raises the question w h a t solution concepts are a p p r o p r i a t e for the solution of the discontinuous differential equation {c = f ( x . A feedback for s y s t e m (1) is a m a p F : ~ n __+ U. the i n w a r d p o i n t i n g condition is satisfied on a suitable sublevel set o f a control L y a p u n o v function. (1) where f : ~ n x U ~ ~ n is continuous a n d locally Lipschitz continuous in x uniformly in u. F(x)). T h e unique t r a j e c t o r y corresponding to an initial c o n d i t i o n x0 E ~ n and u E U : = {u : ~ ~ U [ u measurable} is d e n o t e d by x(. 2 Preliminaries We s t u d y s y s t e m s of the form .to < . < ti < t i + l and tl + ~ is called a sampling schedule or partition. Consider a feedback law F : ~ n _+ U. D e f i n i t i o n 2 ( S a m p l e d c l o s e d l o o p s y s t e m ) .A Globalization Procedure for Locally Stabilizing Controllers 173 back i n v a r i a n t a n d guarantees local e x p o n e n t i a l s t a b i l i t y for the closed loop system. u ) . F(x)) a l m o s t everywhere. A n u m b e r of concepts have been p u t forward to deal with these p r o b l e m s . We simply refer the s t a t e m e n t of necessary c o n d i t i o n s for this to possible existence results. T h e values Ai:=ti+lti. so t h a t we can a p p l y the results of Sections 3 a n d 4 to o b t a i n a global feedback strategy.r ~ n such t h a t x = f ( x . t h a t is "hybrid" in the sense t h a t we e m p l o y different notions o f feedback in different regions of the s t a t e space. Here U C ~ m is a c o m p a c t convex set with nonvoid interior. x0. Note. F(x)) (2) is an a b s o l u t e l y continuous functions x : J . D e f i n i t i o n 1 ( C a r a t h ~ o d o r y c l o s e d l o o p s y s t e m ) . T h e second notion of solution we are interested in is the following. In p a r t i c u l a r . For an interval J C ~ a Carathdodory solution 7 : J "+ ~ of J. . t h a t t h e definition of Cfeedbacks does not require or g u a r a n t e e t h a t there exist any solutions to (2) or t h a t there should be uniqueness.
[10].ti+l].F(z(ti))). Note that it is a peculiarity of sampled feedbacks. F ) with initial value x0 E R n and initial time to = 0 is defined recursively by solving ~(t) = f(x(t). (3) The m a p F is called a sampled (or discrete) feedback if we consider all sampled solutions of (3) and hsampled feedback if we consider all solutions corresponding to sampling schedules lr with d(Ir) < h. A specific point of this paper is that we allow to switch between different solution concepts in different regions of the state space. F ) . with the properties stated in Section 2. [9].r(t. on the respective interval of existence. z0. Note that the definition of sampled feedbacks guarantees existence and uniqueness of rrtrajectories in forward time. On the other hand.u). For any sampling schedule 7r the corresponding sampled or 1rtrajectory x r ( t . In particular. A set B C ]Rn is called (forward) feedbackinvariant under hsampling for system (1) with respect to the feedback F. if for any initial condition x0 E B and any sampling schedule rr with d(Tr) < h it holds that z. x(tl) = x ~ ( t i . We are interested in applying concept of patchy feedbacks that have recently been introduced in [1]. (4) the We we the . In order to obtain a well defined global feedback we will require.174 Jochen Behrens. Fabian Wirth are called intersampling times and sampling rate. nothing prevents trajectories from joining at some time instant. This is of course. z 0 . t E [tl. simply by leaving this area and returning before the next sampling instant. 3 Practical feedbacks feedback stabilization with patchy In this section we study the system = f(x. that the corresponding trajectories may for short times exist on regions of the state space where the sampled feedback is not defined. somewhat undesirable and it is one of the aims of the paper to show how it can be prevented. F) E B for a l l t > 0 . The corresponding idea has already been used in [5].x0. D e f i n i t i o n 3. replace the definition of inward pointing in terms of outer normals by appropriate concept from nonsmooth analysis. respectively. that the region where a sampled feedback is defined remains invariant under the sampled solution in the following sense. slightly extend the definition to allow for less regularity.
For Carathdodory solutions it can be shown that to each initial condition there exists at least one forward and at most one backward solution. Assume we are given a totally order index . One of the interests in this cone stems from its use in the characterization of strong invariance. r] and then follow the parabola to the right are missing.8]. i. D e f i n i t i o n 4. 457 ff. (ii) the open sets /2~ form a locally finite cover o f / 2 . 0) for some interval [1.4 the vector field ga is inward pointing on (0/2~) \ (J /2/~. i f z ~ B. which is set to be {0} if no proximal normM to y exists.y is said to be a proximal normal to B in y. The concept of patchy feedbacks is now defined as follows. In an example [1. a E . T h e arguments. along each C a r a t h d o d o r y solution x it holds that t ~4 m a x { a E A I x(t) E /2a} is nondecreasing and left continuous (with respect to the discrete topology on A). carry over to our definition. open s e t s / 2 a .] the differences to other solution concepts are explained 1. T h e o r e m 4. p. Furthermore. (iii) it holds that ifxe/2 \ p>a (iv) for every a E . p>a Patchy vector fields g and solutions to the differential equation ~ = g(x) are discussed in detail in [1] for the case that the boundaries of/2~ are C 1. g~) I a E . Let /2 C ]~n be an open d o m a i n with b o u n d a r y 0 Y / a n d let D C ~ n be open.4} such that (i) . if for every c o m p a c t set K there exists a constant e > 0 such t h a t for all ~: E K f3 0/2 \ D it holds that c11r The pair (/2.A.. see [7. contrary to what is claimed. We say that g : / 2 4 ll~'* is a patchy vector field if there exists a family of patches { ( ~ .4 is a totally ordered index set. The cone generated by taking all positive multiples of these points is called proximal normal cone to B in y. dist (x. This motivates our notion of patchy vector fields that slightly extends t h a t of [1]. .e. The ones that remain in (0. A Lipschitz continuous vector field g defined on a neighborhood of cl/2 is called inward pointing on (a/2) \ D. however. and i f a point y E B is closest to x. 13) = [1~:. L e t / 2 C ]Rn be an open domain.ylh then x . note that in this example n o t all maximal Filippov solutions are given.3.A and functions F~ : Wa 4 1 Incidentally. g) is called a patch (relative to D).A Globalization Procedure for Locally Stabilizing Controllers 175 Let B be a closed subset o f ~ n. D e f i n i t i o n 5. denoted by N~(y).
where W~ is an open neighborhood of I2~. with the necessary modifications for our case.Q : O .7(t2) 9 OB fort1 < t2 9 J and 7(t) f~ B else. which is essential for our goal. Furthermore.= f ( x . a E . (ii) for every Carath~odory solution 7 starting in x 9 O . The proof of this statement follows the ideas explained in [1]. f ( y . We now state a condition with which we can avoid solutions that touch the boundary of B several times. The following result shows that the foregoing notion of practical feedback controllability is always fulfilled. System (4) is called practically Cfeedback controllable if for every closed set Q C a n and every open set B C Q there is Cfeedback FB. it is an easy consequence of continuous dependence on the initial value that O~ (B) is open. Furthermore.y. For our discussion of practical stability we need the following definition. Let Q c ]~n be closed and B C int Q be open and bounded. s 9 [0. F can be chosen so that there exists no trajectory 7 on an interval J such that 7(tl). Note that it is obvious by definition that O . . satisfies (i) for every x 9 O(B)int Q \ cl B there exists a Carath~odory solution T with 7(0) = x. System (4) is practically Cfeedback controllable using patchy Cfeedbacks.u) 9 B and x ( s . u)) I(~ 9 N~(y)} < ellr (5) is satisfied for all y 9 a B in a neighborhood of x. f ( x ) ) is a patchy vector field on I2 :.( B ) q := {y 9 I 3t > O.U 12~ with patches (12a.( B ) i n t Q \ cl B there is a time T such that 7(T) 9 19B. T h e o r e m 1.feedback F so that for every trajectory 7 starting in x 9 O(B)int O of (2) there is a time T such that 7(T) 9 a B . u ) 9 Q . if Q is open. Fa(x))). We say that F is a patchy Cfeedback for system (4). y .Q(x)).4. D e f i n i t i o n 6. t]}.176 Jochen Behrens.(B)int Q \ cl B r U so that the system ~. f ( x .u 9 : x(t. Given a set Q c ]Rn and an open set B C Q we define the backward orbit of B relative to Q by O . Fabian Wirth U.( B ) Q C Q. Assume that for every x 9 a B there exists a u 9 U such that the inward pointing condition max{((~. C o r o l l a r y 1. FB. if f ( x . then there is a piecewise constant patchy C.
The trajectories of (7) are denoted by z(. The local design procedure we will investigate relies on linearization techniques.e. which takes the form ~ = A ( u ) s . can be characterized via the use of Lyapunov exponents. u). A(u)s) we consider for z r 0 d6(z. see also [11] for a similar result. the aim in [12] is different. and v~(z) := inf d 6 ( z . u) in x .(t) = A ( u ( t ) ) z ( t ) . i. u ) . that 0 is a singular point for (6)..+ U that stabilizes in the sense of sampling. We refer to [9] for details and revise just the facts essential to us. the property that for every z E ~ n there exists a u E U such that z(t. A(u)s)s. We thus introduce the linearization of (6) in 0 with respect to x given by ~. It is known that asymptotic nullcontrollability of system (7). A relation to this result is discussed in [4]. (7) where A(u) denotes the Jacobian of f(. see Definition 2. (6) where we assume in addition to the assumptions previously stated that f : ~ " x U + ~ n is twice differentiable in x and. For us there remains one detail to supply. We have s e e n that this is indeed not necessary. (9) . ttEh t (8) Then it can be shown that asymptotic nullcontrollability of (7) is equivalent to the existence of a t~0 > 0 small enough such that m a x v6(z) < 0 . zex~\(0} for a l l 0 < t ~ < t f 0 . Consider = f(~. u) := (s. furthermore. namely that the control Lyapunov functions constructed in [9] remain control Lyapunov functions for the nonlinear system. 4 A sufficient condition for local sampled feedback stabilization with positive sampling rate In this section we give a brief review of the result of Grfine for local stabilization at singular points. as there the feedback is defined using just extremal values of the set U thereby reducing the complexity of actually designing such a patchy feedback (in particular if U is a polyhedron).) : U + ~ n x n is Lipschitz continuous.u) := F e6T q ( s ( r ) . u ( r ) ) d r . u) .0 for all u e U.(s. z0.0.e. We assume that A(.A Globalization Procedure for Locally Stabilizing Controllers 177 In [12] the author proves the above result and imposes two further conditions: that the system be affine in the controls and locally accessible. However. z. Denoting the radial component q(s. f(0. u). The linear system (7) can be projected to the sphere. i. u) ~ 0. We are interested in constructing a locally stabilizing feedback F : ~ n _.
v. A(Fa(z))z > = ~i~<r (z). the homogeneity of V6 implies t h a t we also have Va(ax ) = a2Va(x). The following statement shows that we retain the property of being a control Lyapunov function if V6 is replaced by Va where (f > 0 and fl > 0 are small enough. A(u)z>. Recall that for a continuous function V : iRn _~ iR and v E iRn the lower directional derivative of V in x in direction v is defined by DV(x. L e m m a III. [9. 11 There exists a A > 0 such that for every p E (0. v) := l i m i n f 1 ( V ( z + tv') . z~ 0. The feedback we envisage is now given as a pointwise minimizer of (~a(z).A) there exist a$p > 0 such that for every J E (0. $p] the function V6 : ]Ra ~ iR defined by ~ ( 0 ) = 0 and V~(z) := em(~)llzli 2 . n > 0.~(z) vEcoA(U)z (11) Note that by definition we have the homogeneity property V6(az) = n ~ ( z ) . t %o.178 Jochen Behrens. (15) . v E iRn and all v >_ 0 that (13) Va(x + Tv) _< Va(x) + ~<r v> + T2llvl{2 27 (14) which can be interpreted as a Taylor inequality for V~ in direction v.xl{~ 9 inf (12) This function is Lipschitz for/3 > 0 and converges pointwise to V6 as fl + 0. L e m m a 1.1] we have for all x. We follow an approach originating in [5] and employed for our special case in [9] that obtains Lipschitz continuous Lyapunov functions from ~ . Lemma ~. (10) satisfies rain DV6(z. that is Fa(z ) is chosen so that <Ca(z). For x E iRn we denote by y~(x) a minimizing vector in (12). Fabian Wirth In [9] it is shown that v6 gives rise to a control Lyapunov function for the linearized system (7). A(u)z). Furthermore.V(x)).ya(x) By [5.~v With this notation v6 (y) is a control Lyapunov function in the following sense. which exists by continuity of V6. and introduce the vector x . We introduce the (quadratic) infconvolution of ~ given by Va(x ) := yEmz~ P~(y) + ~'z~ollY.v) <_ 2pl.
The previous result contains all the arguments necessary to prove t h a t FZ is indeed a tsampled stabilizing feedback for the linearized system (7). such that vecoA(U)z min DVz(z. Now the assertion follows from 1 2~ ~ I l y : ( z ) . For all z C ]~n and all fl > 0 it holds that v e llzll Ily~(z) . We now wish to carry the result over to a local statement for the nonlinear system (6).2] Assume that system (7) is asymptotically nullcontrollable and let p E (0.v ) < 2pV~(z) (17) for all z E It~". We quote the following result by Griine in a slightly extended m a n n e r that also states intermediate statements of the proof that we will need later on. Then there exists a/~ E (0. nor can we expect that a regular choice is possible.zll = V (z)  <  < e llzll .zll Proof. Proposition ~. Given that V/~ is a control Lyapunov function it is no real surprise t h a t it is also one for the nonlinear system (6) in a neighborhood of the origin as we show now. We denote Atzo := 7 fot A ( F z ( z o ) ) z ( v .V~(zo) < t(r ) + t~ < ~tpVz(zo) for all zo E ]Rn and all O < t < i . zo. L e m m a 2. we will however always assume that we have obtained a pointwise minimizer satisfying Fz(az) = Fz(z).Fz(zo))) . z o . To this end we need the following technical lemma. F z ( z o ) ) d r . a > 0. which is easily seen to be possible. Let 0 < 5 < 5p. there is a t > 0 such that V~(z(t. Proposition (16) 1.. Furthermore. 1].zll<%uevsup Ill(x. for z E R " . We denote a :=  inf ~e~\{o} v6(z) and ~:= sup ~e~\{o} v6(z). A).A Globalization Procedure for Locally Stabilizing Controllers 179 This choice is of course not unique. [9. By definition we have that Vz(z) < V~(z) for all z E ~n. 1 and M : = m a x { 2m:xilA(u)ll' . u ) l l } . where ~p is defined in Lemma 1.
F~(x0))) . the definition of M and (21) this implies t h a t V~(x(t.llr =   Atxo = lly:~(x~xollllf~o. For the sake of abbreviation we denote ] := f . Then by decreasing i if necessary we have t h a t [[go[] < M for all xo E c l BI(0) and all 0 < t < t. Let M be as defined in (16). xo. xo. xo. xo. (20). F~(xo)).~ <_ t(~z(xo). t > O. Consider the nonlinear system (6) and its linearization (7). and the definition of ~ we obtain for all xo E cl B1 (0) and all 0 < t < t that (6(x) . L e m m a 12. xo. F~(xo)) ..pv (xo) + . . F~(xo)). Let p E (0.z(t.At~o II II ](x(r. xo.u) (18) for all x E c l Bn(0) and Vt~(x(t. xo.A~o) < t . f ( x ( r . 1] are chosen such that the assertion of Proposition 1 hold.~). F~(x0))) .~3 p V ~ ( x o ) + t(r t s . xo. Proof.At~o) ilxoll 3 .V/3(xo) <_ tpVt~(xo) . Assume that (7) is asymptotically nullcontrollable. F~(xo))ll ___ vf~:Cllxoll~.10 (iii)] there exists a constant C > 0. < t(~o(xo). F~(xo))d~ ~t/3 < v~:llxoll IIx(t.180 Jochen Behrens. such that min DV~(x.z(t.2. 4t~o. F~(xo))ll _< Ctllxoll ~ (20) for all xo E clBl(O) and all 0 < t < L Now by L e m m a 2.V~(xo) = V~(xo + t f~o ) . Then there exists constants R > O. By [3.V~(xo) ~4_t . (14). 0 < 5 < 5p and assume furthermore that 7 > 0 and/3 E (0. F~ (xo)) . such t h a t IIx(t. Fabian Wirth T h e o r e m 2.v) <_ pV~(x) vEcol(~:.A and J'=to := ~ (19) If. for all xo E el Bn(0) and all 0 < t < t. f~o Ato) <.2M2 + t ( ~ z ( x 0 ) . f L ) + tz IIf~~ 2/3 z (21) Using Proposition 1. F~(xo)) dr. s <_ .
~.. such that XR < l p and it follows that 2 VE(z(t. r) C Q there exists a compact connected set O C B(0. we can make possible choices for r > 0 and al.(t. (ii) there exists a constant C > 0 such that for any xo E Gr and any sampling schedule ~r with d(~r) < i it holds for the ~rtrajectory defined by (3) that IIx.sup v~e~'c . r) containing the origin. The final result of this section shows that the sublevel sets of VE describe sets that are feedback invariant under tsampling for system (6) with respect to the Lsampled feedback FE. 5 Global feedback stabilization rate near the origin with positive sampling In this section the two ingredients of the final feedback design will be put together.VE(zo)<_t (3p + Xn) VE(zo)<tPVE(xo) for all x0 with I1~011< R and all 0 < t < t.xo. v'~Cll~lP five(~) . Define 181 Xr := sup H=. For r > 0 we denote the r2sublevel set by G~ := {z I Va(x) < r2}. xo. "five (x) . System (6) is called globally asymptotically feedback stabilizable. We continue to consider system (6) with the additional assumptions stated at the beginning of Section 4 and the associated linearization (7) in 0.A Globalization Procedure for Locally Stabilizing Controllers for all xo E cl B1 (0) and all 0 < t < 3. where we have used the homogeneity of Va. D e f i n i t i o n 7. If we assume in addition to the assumptions of the previous Corollary 2 that 1 fl < : ~ eo" . C o r o l l a r y 2. Now choose R > 0. Let the assumptions of Theorem 2 be satisfied.=. For any r > 0 such that G~ C B(O. if for every connected compact set Q c IRn with 0 E int Q and every open ball B(0. This implies the assertion. see [4]. E~)II _< c~"'/~ll~oll . as more concrete. R) it holds that (i) The set Gr is feedbackinvariant under isampling for system (6) with respect to the feedback FE. such that the following conditions are satisfied: . at least close to zero.FE(zo))). Choose R > 0 according to the assertions of that theorem.
As for all rl < r < r2 the set G~r is also tsampled feedback invariant we still have that the following wtrajectories remain in G~r and converge exponentially to zero.182 Jochen Behrens. Let t and/~ 6 (0. This makes the decision of switching less delicate for the price that we have concurring definitions for the feedback in a certain region of the state space.~(t. t~) (b) for every solution 7(') of (22) with 7(0) = z 6 0 . then the statement is obviously also true for D = G~. F1 (x)) with 7(0) = z on some interval [0. say at a point x 6 G~.( D ) i n t Q \ D there exists a Carath4odory solution of = f ( x . and such that t400 lim x. The linear decrease statement (19) from T h e o r e m 2 guarantees that we can satisfy the inward pointing condition (5) for OG~. Remark I. such that (a) for every x 6 O .. for every x 6 D and every rr trajectory for sampling schedules rr with d(~) < ~. (c) there is no solution 7 of (22) with 7(0) = x 60(D)intQ\D such that 7(tl) 6 0 D .( G ~ . so that we need a further variable. T h e o r e m 3. Proof. The system (6) is globally asymptotically feedback stabilizable in the sense of Definition 7 if its linearization (7) is asymptotically nullcontrollable..s a m p l i n g for system (6) with respect to the feedback F~. to demand switching between the two controllers and solution concepts if the boundary of a set D is reached. Take two values 0 < rl < r2 such that for a suitable fl the sampled feedback F~ renders G~r2 tsampled feedback invariant and exponentially stabilizes to O. 1] be such that F2 = F# is an tsampled exponentially stabilizing feedback on D ' = Gr~. ) Q . x. a "switch". such that D is feedback invariant under t . 7(t2) ~ D for some tl < t2. (ii) There exists a sampling bound t > 0 and a t . to remember which feedback strategy is applicable. with (22) there 7(T~) 9 aD. however. Fabian Wirth (i) There exists a patchy Cfeedback F1 on O . stabilizing to OGT. especially as level sets of V# do not lend themselves easily to computation. Let Q C A n be closed with 0 6 int Q. We can. F2) = O. It is of course quite unrealistic from several points of view.( D ) i n t q \ D exists a T7 > 0. Now by T h e o r e m 1 and Corollary 1 we have the existence of a ~)atchy Cfeedback F with ordered index A on O . Now we may just require that the switch between the Carathdodory and the sampled feedback is made somewhere in G ~ \ G~I.s a m p l e d feedback F2 on D. .( D ) i n t Q \ D . This concludes the proof. We choose an 0 < r < r '. relax the requirements here a bit.
This consists in the construction of feedback invariant sets for the feedback F2 that can be entered from the outside under the feedback F1. Birkhs Boston 4. 445471 2. Ancona F. 15851603 9. Clarke. In general. R.. D. L. Math. Sontag. L. F. Clarke.P. IEEE Trans. Colonius. (1984) Differential Inclusions: SetValued Maps and Viability Theory.H. A c k n o w l e d g m e n t : The authors would like to thank Lars Griine for numerous helpful comments during the preparation of this manuscript. A. Ledyaev.). J. France. L.S. Proc. J. LamnabhiLagarrigue. Aeyels.. L. Vol. Bessan A.. Yu. F.. SpringerVerlag. van der Schaft (eds.S. Yu. (1998) Nonsmooth Analysis and Control Theory. Cellina. this approach is not restricted to the feedback types we have considered here. (1999) Stabilization by sampled and discrete feedback with positive sampling rate. Contr. Contr. (1997) Asymptotic controllability implies feedback stabilization.A Globalization Procedure for Locally Stabilizing Controllers 183 6 Conclusions In this article we have presented a method to unite a local exponentially stabilizing sampled feedback with a global piecewise constant feedback interpreted in the sense of Carath6odory solutions. Ledyaev. and that have the additional property that no solutions under F1 can move away (locally) from the feedback invariant set. Clarke. 42. Contr.J. The key tools in this approach were methods from nonsmooth analysis in particular proximal normals and infconvolution. L.. & Opt. Theory Networks Syst.I. SpringerVerlag.... & Opt. (2000) A globalization procedure for locally stabilizing controllers. 13941407 6. Optim. Wolenski.J. Griine. Universitht Bremen 5. A. Stern. W. (1999) Patchy vector fields and asymptotic stabilization. 12881314 10. Gr/ine.D. New York 8. Contr.. CDRom . Kliemann.. (2000) Homogeneous state feedback stabilization of homogeneous systems. Subottin. 36. Rifford. Ledyaev.. 165182. References 1. MTNS 2000. Grllne. Berichte aus der Technomathematik. Wirth.H. (2000) Homogeneous control Lyapunov functions for homogeneous control systems. In: F. London 11. F. Yu.. but can be performed for any feedback concepts that allow for the completion of the key step in our design. Aubin. 38. (2000) Feedback stabilization and Lyapunov functions. Gr/ine. and Calculus of Variations. ESAIM: Control. (2000) The Dynamics of Control. to appear 7. pp. Stability and Stabilization of Nonlinear Systems. E.H. F. SIAM J. & Opt. R. SIAM J. Stern. Perpignan. SIAM J.S. Lecture Notes in Control and Information Sciences. 246. F. P. Springer Verlag. Berlin 3. Report 0009.. (1998) Asymptotic controllability and exponential stabilization of nonlinear control systems at singular points. 4.. A. Behrens. Automat.R.
. and Control. Differential Equations. 551598 16. Nikitin. SIAM J. (1999) Stability and stabilization: Discontinuities and the effect of disturbances.P.H. Iowa 13. Clarke and R. In: F.D. 911933 14. E.J. G. 32 15971604 15. 27992805 . Fabian Wirth 12. Contr. (1994) On Brockett's condition for smooth stabilizability and its necessity in a context of nonsmooth feedback. E. (1999) Piecewiseconstant stabilization. (1996) Practical feedback stabilization of nonlinear control systems and applications. Lai R. In Proc. S. (1995) Nonsmooth controlLyapunov functions. SIAM J.D. Ryan. E. New Orleans. Nonlinear Analysis.184 Jochen Behrens. eds. Stern. Contr. Iowa State University.J. CDC 1995. USA. & Opt. Sontag. Ames. H. Kluwer. &: Opt. Sontag. PhD thesis. Sussmann. 37. pp.
Optimal Control and Implicit Hamiltonian Systems
G u i d o Blankenstein and A r j a n van d e r Schaft Faculty of Mathematical Sciences, Department of Systems, Signals and Control University of Twente, P.O.Box 217 7500 AE Enschede, The Netherlands
{g. blankenstein, a. j. vanderschaft}@math.utwente.nl
Optimal control problems naturally lead, via the Maximum Principle, to implicit Hamiltonlan systems. It is shown that symmetries of an optimal control problem lead to symmetries of the corresponding implicit Hamiltonlan system. Using the reduction theory described in [3,2] one can reduce the system to a lower dimensional implicit Hamiltonian system. It is shown that for symmetries coming from the optimal control problem, doing reduction and applying the Maximum Principle commutes.
Abstract.
Furthermore, it is stressed that implicit Hamiltonian systems give rise to more general symmetries than only those coming from the corresponding optimal control problem. It is shown that corresponding to these symmetries, there exist conserved quantities, i.e. functions of the phase variables (that is, q and p) which are constant along solutions of the optimal control problem. See also [19,17]. Finally, the results are extended to the class of constrained optimal control problems, which are shown to also give rise to implicit Hamiltonlan systems.
1
The
Maximum
Principle
Let Q be a s m o o t h ndimensional m a n i f o l d , with local c o o r d i n a t e s d e n o t e d by q, and let U be a s m o o t h m  d i m e n s i o n a l manifold, with local c o o r d i n a t e s d e n o t e d by u. Consider a s m o o t h m a p f : Q • U + TQ, where T Q d e n o t e s the t a n g e n t bundle of Q, such t h a t the d i a g r a m
QxU
]+TQ Q
c o m m u t e s , where p : Q x U + Q is the p r o j e c t i o n o n t o the first a r g u m e n t a n d 7rq : TQ ~ Q is the canonical projection. T h e n the m a p f defines a set of s m o o t h vector fields { f ( . , u) } , e u on Q, defined by f(., u)(q) = f(q, u), q E Q. Define the nonlinear system
,~ = f(q, u)
(1)
186
G. Blankenstein and A.J. van der Schaft
The next theorem gives a summary of the Maximum Principle in optimal control theory. T h e o r e m 1. ( M a x i m u m P r i n c i p l e ) [15,4,8,18] Consider the nonlinear system defined in (1), with given initial conditions q(0) = qo. Consider two smooth functions s : Q x U + ~ (called the Lagrangian) and IC : Q + and define the cost functional J : Q• x U ~ +
T
J(q(.), u(.)) = ]0 •(q(t), u(t))dt + IC(q(T))
(2)
for some fixed and given time T E IR+. The free terminal point optimal control problem is given by minimize J over Q~ x U ~ such that (q(.), u(.)) satisfies (1), and q(0) = q0. Define the smooth Hamiltonian H : T*Q • U ~ H(q,p,u)=pTf(q,u)s (q,p,u) E T * Q • U (3)
where (q,p) ave local coordinates for T* Q. Necessary conditions for a smooth trajectory (q(.), u(.)) E QR x U 1 to solve the free terminal point optimal control problem are given by the existence of a smooth curve (q(),p()) E (T'Q) R satisfying the equations OH (l(t) = ~p (q( t ),p( t ),u(t)) OH p(t) = ~r (q(t),p(t), u(t)) OH 0 = ~ u ( q ( t ) , p ( t ) , u(t)) (4) (5) (6)
for all t E [0, T], along with the initial conditions q(O) = qo and the transversality conditions 01C p(T) = ~q (q(T)) (7)
Remark 1. Note that the Hamiltonian H in (3) can easily be defined in a coordinate free manner. Furthermore, condition (6) is usually stated as the maximization condition H(q(t),p(t), u(t)) = maxH(q(t),p(t), it)
aEU
(8)
for (almost) every time t E [0, T]. However, because we assumed U to be a smooth manifold (which is quite a strong condition and leaves out for instance constrained, or bounded, control) and H a smooth function, (6) is a necessary condition for (8) to hold.
Optimal Control and Implicit Hamiltonian Systems
187
Remark 2. Notice t h a t in the a b o v e d e s c r i p t i o n of the M a x i m u m P r i n c i p l e we only considered so called regular e x t r e m a l s ( c o n t r a r y to t h e abnormal extremals). However, since, for simplicity, we only consider the free terminal point optimal control problem, it can be proved t h a t every o p t i m a l s o l u t i o n
(q(.), u()) corresponds to a regular extremal, [18].
In the next section, the s y s t e m of equations (4)(6) will be d e s c r i b e d in a c o o r d i n a t e free m a n n e r by using the n o t i o n of an i m p l i c i t H a m i l t o n i a n system. F u r t h e r m o r e , a nice c o o r d i n a t e free i n t e r p r e t a t i o n of e q u a t i o n (7) will be given.
2
Implicit Hamiltonian Systems
It is well known t h a t the c o t a n g e n t bundle T * Q has a n a t u r a l interpret a t i o n as a s y m p l e c t i c m a n i f o l d ( T * Q , w ) , defined by the canonical twoform w : F(T(T*Q)) x F(T(T*Q)) .+ C~176 on T*Q (here F denotes the space of s m o o t h sections, i.e. F(T(T*Q)) denotes the space o f s m o o t h sections of T(T*Q), t h a t is, the space of s m o o t h vector fields on T'Q). In local c o o r d i n a t e s w is given by w = dq A dp. T h e m a p w : F(T(T*Q)) ~ F(T*(T*Q)) can be e x t e n d e d to a mapffJ : F(T(T*Q x U)) + F(T*(T*Q • U)) in the following way: Consider the trivial vector b u n d l e
T*Q x U PT~ T ' Q , and the trivial fiber bundle T*Q x U ~~ U. DeQ fine a vector field X E F(T(T*Q x U)) on T*Q x U to be horizontal if (pu).X = 0, and call it vertical if (pT.Q).X = 0. In local c o o r d i n a t e s , if X (q,p, u) = Xq(q,p, u) o~+ Xp(q,p, u) o~+ Xu(q,p, u) ~ then X is h o r i z o n t a l i f X u = 0 and X is vertical ifXq = Xp = 0. A oneform a E F(T*(T*Q • U)) on T*Q • U is called horizontal if it a n n i h i l a t e s every vertical vector field,
and called vertical if it annihilates every horizontal vector field. Now, is uniquely defined (by s k e w  s y m m e t r y ) by requiring t h a t & ( X ) = 0 for every vertical vector field X E F(T(T*Q x U)), a n d for every h o r i z o n t a l vector field X E F(T(T*Q x U)), ~(X)(q,p,u) = w(X(.,.,u))(q,p) for all (q,p, u) E T*Q • U, where X(.,, u) E F(T(T*Q)) is considered a vector field on T ' Q , for every u E U. Note t h a t ~ defines a p r e s y m p l e c t i c s t r u c t u r e on
T*QxU.
Recall t h a t a Dirac structure on a m a n i f o l d M is defined as a linear s u b s p a c e n C F ( T M ) • F(T*M) such t h a t n = D • where
D • : = { ( X ~ , a ~ ) E F ( T M ) x F(T*M) [ ( a l , X ~ ) + (a~,X1) = O,
V(Xl, al) ~ D}
where (.,.) denotes the n a t u r a l p a i r i n g between a oneform a n d a vector field, see [5,7,21,20,6,3,2]. Note t h a t D defines a s m o o t h vector s u b b u n d l e o f T M S T * M , w i t h d i m D ( z ) = d i m M , Vz E M .
188
G. Blankenstein and A.J. van der Schaft
The map & defines a Dirac structure D on T*Q x U given by
D={(X,a) EF(T(T*QxU))xF(T*(T*QxU))I
a = ~(X)}
(9)
In local coordinates, the Dirac structure given in (9) has a very simple form
D = { ( X , a ) E F(T(T*Q x V)) x F(T*(T*Q x U))]
[~ [i,!][
= o
a. 0
x,
X~
}
(10)
o where we denoted X(q,p,u) = Xq(q,p, u )~q + Xp(q,p, u).~p+X,(q,p,u)_~u and a(q, p, u) = Ctq(q,p, u)dq + ap(q, p, u)dp + au( q, p, u)du.
Consider the Hamiltonian H E C~176 • U) defined in (3). The implicit Hamiltonian system, denoted by (T* Q • U, D, H), corresponding to the Dirac structure D and the Hamiltonian H, is defined as the set of smooth solutions (q(.),p(.), u(.)) E (T*Q)ZxU R such that (4(t),[~(t), i,(t)) = X(q(t),p(t), u(t)), t E ~ , for some smooth vector field X E F(T(T*Q x U)), and
(X, dS)(q(t),p(t), u(t)) E D(q(t),p(t), u(t)),
t E~
(11)
where dH E F(T*(T*Q x U)) is the differential of the function H. For more information on implicit Hamiltonian systems, see [21,20,6,3,2]. Now, it is immediately clear that if the Dirac structure is given in local coordinates as in (10), then the equations (11), restricting t E [0,T], are precisely the equations (4)(6). Thus we have shown P r o p o s i t i o n 1. Every optimal control problem, applying the Maximum Prin
ciple, gives rise to a corresponding implicit Hamiltonian system. Remark 3. Note that equations (4)(6) are general for all optimal control problems applying the Maximum Principle, not necessarily free terminal point problems.
Now we will give a coordinate free interpretation of the transversality conditions (7). Let Y E F(TQ) be a smooth vector field on Q. Define the smooth function H y E CCr given in local coordinates (q,p) by
Hy (q, p) = p T y (q)
(12)
P r o p o s i t i o n 2. The transversality conditions (7) are equivalent to the fol
lowing: for every smooth vector field Y E F(TQ) on Q, the function Hy E C~176 at the end point (q(T),p(T)) of a solution of the free terminal point optimal control problem has the value Hy(q(T),p(T)) =  ( L y ~ ) ( q ( T ) ) where LyIC denotes the Lie derivative of the function ICE C~ to the vector field Y.
(13)
with respect
Optimal Control and Implicit Hamiltonian Systems
189
Proof. Let (q,p) be local coordinates for T ' Q around (q(T),p(T)). If one takes Y = o q ' then Hy = p and (13) is exactly (7) Now, let Y(q) = O
Yt(q)o~ + " " + Yn(q) o~, then Hy(q,p) = PlYI(q) + " " + PnYn(q), and
it follows that
Hr (q(T), p(T)) = Pl (T)Yx (q(T)) + . . . + Pn (T)Yn (q(T))
= (LyIC)(q(T))
1:3
Proposition 2 will come in need in the next section, where we show that the m o m e n t u m map, corresponding to a s y m m e t r y Lie group of the optimal control problem, will always have the value zero along solutions of the optimal control problem.
3
Symmetries
of Optimal
Control
Problems
Consider the free terminal point optimal control problem (from now on called optimal control problem) described in theorem 1. Consider a Lie group G, with smooth action r : Q • G + Q on Q, and denote the corresponding Lie algebra by {~. G is said to be a symmetry Lie group of the optimal control problem if [9,10,19]
[f(.,u),~Q]=O, L ~ Q s
VuEUandL~QIC=O,
V~E~
(14)
where ~Q E F(TQ) denotes the vector field corresponding to ~ E G and the action r (i.e. the infinitesimal generator). The action of G on Q can be lifted to an action on T*Q in the following way: let ~Q E F(TQ) and consider the function H~Q E C~176 defined in (12). Define the vector field ~Q E F(T(T*Q)) by dH~Q = i~Q~, where ~v is the canonical symplectic form on T* Q. In local coordinates, if ~Q (q) = h(q)o~ then ~Q (q, p) = h(q)O~ht ~ 0 PT O~qtq)T~p'. Lifting all the vector fields ~Q E F(TQ),~ E {~, to T*Q defines an action r : T*Q x G ~ T*Q of G on T ' Q , such that the infinitesimal
generators are given by ~q, ~ E {~. From (14) it immediately follows that the lifted action r on T*Q leaves the Hamiltonian (3) invariant. Indeed,
L~QH(q,p, u) = p ~q [q, u)h(q)  ~q (q, u)h(q)  p ~q (q)f(q, u) = pT[f(., u),~q(.)](q)  L~QC(q, u) = O, V(q,p) E T*Q,Vu E U
(15)
T 0f,
C~
T Oh
We will show that G is a symmetry Lie group of the implicit Hamiltonian system (T*Q • U, D, H) defined in section 2. Therefore we need the following
190
G. Blankenstein and A.J. van der Schaft
D e f i n i t i o n 1. [5,7,20,3,2] Consider a Dirac structure D on T*Q x U (as in (9) for example). A vector field Y E F(T(T*Q x V)) on T*Q x V is called a symmetry of D if (X, a) E D implies ([Y, X], Lya) E D, for all (X, c~) E D. A Dirac structure is called closed if it (generalizing the Jacobi identities for information. Let us just mention that closed (this follows immediately from important result is given in [5,7]. satisfies some integrability conditions a Poisson structure), see [6] for more the Dirac structure D given in (9) is proposition 4.5 in [6]). The following
Proposition
3. Let D be a closed Dirac structure on T*Q x U. Let Y E
F(T(T'Q x U)) be a vectorfleld on T*Q x U and assume that there exists a
function F E CC~ D. • U) such that (Y, dF) E D, then Y is a symmetry of
Now, consider a lifted vector field ~Q ~ F(T(T*Q)), this can be identified with a horizontal vector field ~Q E F(T(T*Q x V)) on T* Q • V. Furthermore, the function H~Q E C~176 defines a function H6Q E C~176 • U) (by H~Q(q,p,u) = H~Q(q,p), (q,p,u) E T*Q x U). Since dH~Q : i~Q~, it follows that (~Q, dH~Q) E D, given in (9), which implies by proposition 3 that ~Q is a symmetry of D. It follows that every vector field ~q,~ E G, is a s y m m e t r y of D, we call G a s y m m e t r y Lie group of D. Furthermore, since H is invariant under G, we say that G is a symmetry Lie group of the implicit Hamiltonian system (T*Q x U, D, H). Thus we have shown that every s y m m e t r y of an optimal control problem (of the form (14)) leads to a (lifted) s y m m e t r y of the corresponding implicit Hamiltonian system. In [3,2] we have presented a reduction theory for implicit Hamiltonian systems admitting a symmetry Lie group. This theory extends the classical reduction theory for symplectic and Poisson systems, as well as the reduction theory for constrained mechanical systems. We refer to [3,2] for further details. In this paper we want to apply this reduction theory to implicit Hamiltonian systems (T*Q x U, D, H) of the form (3,9) subject to the (lifted) action of the symmetry Lie group G descibed above, (14). Because of the simple form of the Dirac structure D and of the (lifted) symmetries (14), the reduction will actually come down to classical symplectic reduction. We will show that after reduction of the system (T*Q • U, D, H) to a lower dimensional implicit Hamiltonian system, one obtains the same (up to isomorphism) implicit Hamiltonian system as obtained by first reducing the optimal control problem itself (to an optimal control problem on Q/G) and then applying the Maximum Principle in theorem 1, i.e. doing reduction and applying the Maximum Principle commutes. Consider the optimal control problem defined in theorem 1, and suppose it admits a symmetry Lie group G, described in (14). Assume that the quotient manifold Q/G of Gorbits on Q is well defined (e.g. the action of G on Q is
Optimal Control and Implicit Hamiltonian Systems
191
free and proper), and denote the projection zr : Q 4 Q/G. Because of (14), every vector field f(., u), u E U, on Q projects to a well defined vector field ](., u), u E U, on Q/G, defined by rr.f(., u) = ](., u), u E U. Furthermore, the Lagrangian s u) : Q + ~ , u E U, defines a well defined function /~(., u) : Q / G ~ ~ , u E U, by s = s ~r,u E U. Finally, the function K: : Q 4 ~ defines a well defined f u n c t i o n / ~ : Q / G .4]R by/C = / ~ o lr. T h e n in [9,10] it is shown that the o p t i m a l control problem on Q projects to a well defined optimal control problem on Q / G , defined by {](., u)}uev, s and K:. T h e cost functional J : (Q/G) ~ x U ~ ~ ~ becomes
g(o(.), u(.)) =
I
2(O(t), u(t))dt + ~(O(T))
where (q('), u(.)) e (Q/G) R x U ~ denotes a time trajectory in Q / G x U. Notice that J(q(.), u(.)) = J(rr(q(.)), u(.)) for all trajectories (q(.), u(.)) E QR x U R satisfying (1). T h e reduced optimal control problem is defined by
minimize J over (Q/G) ~ x U K such that (0('), u(.)) satisfies q(t) = ](O(t), u(t) ), t E [O,T], and 0(0) = 40
where q0 = 7r(q0). As in theorem 1 we can apply the M a x i m u m Principle to obtain necessary conditions for a trajectory (~(.), u(.)) E (Q/G) ~ x U ~ to be a solution of the reduced o p t i m a l control problem. Define the reduced Hamiltonian
f t ( 4 , P , u ) = p T ] ( O , u )   ~ ( 4 , u),
(~,[~,u) E T ' ( Q / G ) • U
where (~,i5) denote local coordinates for the cotangent bundle T* (Q/G). It is easy to see that S and /2/ are related by /2/(., u) = H(., u) o lr*,u E U,, where ~r* : T*(Q/G) + T*Q is the pullback defined by ~r. Now, applying the M a x i m u m Principle to the reduced optimal control problem gives us the necessary conditions, for a s m o o t h curve (~(.), u()) E (Q/G) ~ • U ~ to be a solution of the reduced optimal control problem, of the existence of a s m o o t h curve (~(.),p(.)) E (T* (Q/G)) ~ satisfying the equations
^
~(t) = gJ~(4(t),p(t), u(t))
(16)
~(t) ~(~(t),p(t),
=
u(t))
(17)
(18)
•
0 = ~~u(~(t),p(t), u(t))
for all t E [0, T], along with the initial conditions ~(0) = q0 and the transversality conditions iS(T) =  ~  ( 4 ( T ) )
0s
192
G. Blankenstein and A.J. van der Schaft
As described in section 2, proposition l, equations (16)(18)give rise to an implicit Hamiltonian system ( T * ( Q / G ) • U, D, H), where D is a Dirac structure on T* (Q/G) x V defined analogously as in (9) (note that T* (Q/G), being a cotangent bundle, has a natural symplectic form o3). We will show that this implicit Hamiltonian system is isomorphic (to be defined precisely later on) to the implicit Hamiltonian system obtained by applying the reduction theory described in [3,2] to the system (T*Q x U, D, H). Consider the implicit Hamiltonian system (T*Q x U, D, H) with s y m m e t r y Lie group G. The action of G on T*Q is given.by r and the infinitesimal generators are given by the lifted vector fields ~q,~ E ~. Corresponding to this action, there exists a Ad*equivariant momentum map, defined in local coordinates by r : T*Q + ~* [1]
~(q,p)(~) = pT~q(q),
,dE E ~, (q,p) 9 T ' Q
(19)
Define ~(~) : T*Q + R by 4~(~)(q,p) =.q~(q,p)(~),(q,p) 9 T*Q,~ 9 ~7, then it follows that ~(~) = H~Q. Since (~Q,dH~Q) 9 D,~ 9 ~, it follows that (~q, d~(~)) 9 D, ~ 9 ~. Now, consider a solution of the optimal control problem, i.e. (q(t), i5(t), u(t)) = X(q(t), p(t), u(t)) and
(X, dH)(q(t),p(t), u(t)) 9 D(q(t),p(t), u(t)),
t 9 [0, T]
(20)
see section 2. By definition, D = D a, which implies
0 = ((X, dr
+ (~Q, dH))(q(t),p(t), u(t)) = (X, d~(r
u(t)),
t 9 [0, T], by (15), for all ~ 9 G. This means that 4~(r is constant along solutions of the optimal control problem (or rather, of the implicit Hamiltonian system (T*Q • U, D, H)), see also [19,17]. Furthermore, by the transversality conditions in proposition 2 it follows that
9 (~)(q(T),p(T)) = (L~QIC)(q(T)) = O,
~9
by (14). This means that ~(~) has actually constant value zero along solutions of the optimal control problem. Since this holds for every ~ 9 G it follows that the m o m e n t u m m a p 4~ has constant value 0 9 G* along solutions of the optimal control problem. Thus we have proved. P r o p o s i t i o n 4. The momentum map ~ : T*Q + ~* has constant value zero
along solutions of the optimal control problem (or rather, implicit Hamiltonian system (T*Q x U, D, H)).
Now we can use the reduction theory described in [3,2] to reduce the implicit Hamiltonian system to a lower dimensional implicit Hamiltonian system (T*Q x U, D, H) (note that all the constant dimensionality conditions in there are satisfied). We begin by restricting the system to a levelset of the momentum map, i.e. 4i 1(0) • U C T* Q • U. This gives again an implicit Hamiltonian
Optimal Control and Implicit Hamiltonian Systems
193
system on ~  l ( 0 ) x U. The resticted system will have some s y m m e t r y left, corresponding to the residual subgroup G~,  {g 9 G I Ad~p  p}, p 9 ~*. However, since in this case p = 0, and the coadjoint action Ad~ : G* ~ {~* is linear, Vg 9 G, it follows that Go  G. This means that the restricted system will still have G as a s y m m e t r y Lie group. Assuming that the quotient manifold ~  l ( 0 ) / G is well defined, we can project the resticted system on 4~1(0) x U to an implicit Hamiltonian system on 4)l(O)/G x U. The resulting implicit Hamiltonian system will be denoted by (4~l~0)/G x U , / ) , / ~ ) . Here g 9 C~1761 (O)/G x U) is defined by H ( . , . , u) o i = H(., u) o ~r0, u 9 U, where i : 4~1(0) + T*Q is the inclusion map, and ~r0 : O  l ( 0 ) + O  l ( 0 ) / G is the projection map. Since the Dirac structure D has the very special form (9), or (10), and the action of G on T*Q x U only acts on the T ' Q  p a r t , it can easily be seen t h a t the reduction theory described in [3,2] in this special case comes down to symplectic reduction of the symplectic manifold ( T ' Q , w), as descibed in e.g. [14,1,12], see also example 8 in [3]. It follows t h a t the reduced Dirac structure D on ~  l ( 0 ) / G x U is given by
D = { ( 2 , 5 ) 9 F(T(~I(O)/G • V)) • F(T*(4~I(O)/G x U)) I = ~0(2)}
where r : F ( T ( r x U)) ~ r(T'(r • v ) ) is the extension of the symplectic two form w0 on 4~l(0)/G, defined by ~r~w0 = /'w, as in section 2. Now we will show that the two r e d u c e d implicit Hamiltonian systems (T" (Q/G) x U, L), iI) and ( ~  l ( 0 ) / G x U, D , / ~ ) are isomorphic, defined in the following way. D e f i n i t i o n 2. [3] Let M and N be two manifolds, and let r : M + N be a diffeomorphism. Let DM be a Dirac structure on M and ON a Dirac structure on N. Then r is called a Dirac isomorphism if
(X, oL) 9 DM ~
( r . X , ( r * )  l a ) 9 ON
In this case we call DM and DN isomorphic. Consider two implicit Hamiltonian systems (M, DM, HM) and (N, ON, HN). We call the two systems isomorphic if DM and DN are isomorphic, and
HM = HN o r.
We begin by showing t h a t / ) a n d / ) are isomorphic (by some diffeomorphism r :r • V .+T*(Q/G) • U). First, consider the two symplectic m a n ifolds ( ~  l ( 0 ) / G , wo) and (T* (Q/G), Cv) (recall t h a t & denotes the canonical symplectic twoform on T*(Q/G)). It is a well known result from Satzer [16], see also [11] and [13], that these two symplectic manifolds are symplectomorphic, i.e. there exists a diffeomorphism r : ~  l ( 0 ) / G ~ T*(Q/G) such that r is a symplectic map, that is, r*& = w0. T h e m a p r induces a m a p
194
G. Blankenstein and A.J. van der Schaft
(7.,id) : q~~(o)/a x U + T ' ( Q / G ) x U, which we will also denote by 7. (it will be clear from the context which m a p is meant). Here, id : U 4 U denotes the identity m a p on U. Since /) and D are completely defined by o~0, respectively &, it follows from the fact that 7.*& = w0 that 7 is a Dirac isomorphism between /) a n d / ) . To see that the H a m i l t o n i a n s / ~ and f / a r e also related, i.e. H = jr/or, we have to go a little bit deeper into the construction described in [11,13] of the m a p 7.. In [11,13] the m a p 7 is constructed by taking a connection V : T Q ~ on the principal Gbundle Q 2~ Q/G. Precisely, V : (TQ, T r 4 (~, i d ) is a homomorphism such that V(~Q(q)) = ~,~ E G (where T r represents the tangent of the m a p r Take arbitrary a E F(T*Q) and define fl E F(T*Q) by fl(q) = a(q)  ~7*(q~(a(q))),q E Q. Then fl projects to a oneform & E F(T*(Q/G)), i.e./~ = 7r*&. Indeed,
Define the m a p roy : T*Q ~ T * ( Q / G ) by rcv(q, Otq) = (~,(~4),o~ E F(T*Q), where ~ = ~r(q), and restrict this m a p to get the m a p ZCVo : ~  1 ( 0 ) + T*(Q/G). In [11] it is shown that ~  1 ( 0 ) ~_y+oT * ( Q / G ) is a principal Gbundle. Finally, the m a p 7.: q~l(O)/G .+ T*(Q/G) is defined such t h a t the diagram
9 ~(0)
Y4 ~  ~ ( o ) / a
~ Vo " ~ , t 7.
T'(Q/G)
commutes. In [11] it is shown t h a t r : ((/'l(0)/G, w0) + (T*(Q/G),&) is a symplectomorphism (this does not depend on the connection ~r~choosen). Now we can prove that the Hamiltonians/~ a n d / : / a r e related by H = H o v. Take arbitrary z E ~I(O)/G, then there exists a point (qo,Po) E 4~1(0) C T*Q such that rro(qo,Po) = z. Let a E F ( T * Q ) be such that (q, aq) E 4  1 ( 0 ) , q E Q and (q0, aqo) = (qo,po). Then there exists an & E F ( T * ( Q / G ) ) such that ~ = rr*&, and it follows that rvo(q0,P0) = 7fro(q0, aqo) = (40, &4o)" By commutation, 7.(z) = (40, &4o) Furthermore
fl(z, u) = H(., ., u) o i(qo, aqo) = = ~T ^ ^
aqTf(qo, u) 
s
u)
4of(q0, u)  z~(~0, u) = H(q0, a~o, u) = / ~ o 7.(z, u)
and it follows that / t = / t o 7.. Thus we have shown that the two implicit Hamiltonian systems (~I(O)/G x U, D, [I) and (T*(Q/G) x U, D, [I) are isomorphic. In particular we have proved. Theorem 2. Consider an optimal control problem with symmetries described in (14}. Then reducing the optimal control problem to a lower dimensional
Optimal Control and Implicit Hamiltonian Systems
195
optimal control problem (on Q/G), called the reduced optimal control problem, and applying the Maximum Principle, is equivalent (up to isomorphism) to first applying the Maximum Principle on the original optimal control problem (on Q), and then using the reduction theory described in [3,2] to reduce the corresponding implicit Hamiltonian system.
This result seems quite natural, and it gives a nice connection between reduction of optimal control problems on the one hand and reduction of implicit Hamiltonian systems on the other hand. To appreciate the result, notice that in general ~I(I~)/G u and T*(Q/G) are not diffeomorphic (they do not even have the same dimension). However, because we proved that the momentum map has value zero along solutions of the optimal control problem (or, implicit Hamiltonian system), we could use the results in [16,11,13] to conclude that (4i 1(O)/G, wo) and (T* (Q/G), ~) are symplectomorphic (as symplectic manifolds), from which the final result in theorem 2 was obtained.
4
Generalized
Symmetries
In the previous section we showed that symmetries of optimal control problems of the form (14) lead to symmetries of the corresponding implicit Hamiltonian system, by lifting the symmetries to the cotangent bundle. However, in general the implicit Hamiltonian system has more symmetries than only those coming from the optimal control problem. These symmetries cannot necessarily be described as the lift of a vector field on Q, so they do not correspond to a symmetry of the optimal control problem. These symmetries (of the implicit Hamiltonian system) are called generalized symmetries of the optimal control problem (in [19] these symmetries are called dynamical symmetries, contrary to lifted symmetries, called geometrical symmetries). In this section it is proved that to each generalized symmetry of the optimal control problem there exists (locally) a conserved quantity, being a function of the phase variables q and p which is constant along solutions of the optimal control problem (or, implicit Hamiltonian system). (In case of a lifted symmetry, this conserved quantity is exactly the (globaly defined) momentum map, described in section 3.) We refer to [19,17] for related results. Consider the Dirac structure described in (9), corresponding to the implicit Hamiltonian system (T*Q x U,D,H) obtained by applying the Maximum Principle to the optimal control problem, see section 2. Define the following (co)distributions on T*Q [6] (denote by F vert (T(T*Q • U)) the set of all vertical vector fields on T*Q x U, and by F h~ (T*(T*Q x U)) the set of all
c. The /fpart however holds under the condition that P1 is constant dimensional and involutive.t. Define the set of admissible functions AD = {H 9 C~176 • U) [ dH 9 P1} = {H 9 C ~ ( T * Q • U) ]H(q. "}D.p) and Yp(q. where (X1.p)o~+Yp(q. LyP1 C P1. defined by Go = ker P1 being an involutive distribution.3]. if and only if 9 Y is canonical with respect to {. p. p.) 9 D} = 0 {a 9 r ( T * ( T * Q x U)) l a X 9 r ( T ( T * Q x U)) s. H~ 9 . P r o p o s i t i o n 5. Blankenstein and A. a) 9 D} Fh~ x U)) Note that all the (co)distributions are constant dimensional and that Go = ker P1 and P0 = ann GI. Furthermore. i = O.X1). (22) .40.3]. (X2.X2) = (dH2.p)o~ 9 VH1. see [6.O) 9 D} = F v*rt (T(T*Q x V))} G1 = = Oo = Px = = { X 9 F(T(T*Q • U)) I ~ 9 F(T*(T*Q • U)) s. defined in definition 1. where t'(q. dill).dH~) 9 D [6. Go = {X 9 F(T(T*Q x U)) [(X. Ly{H1. as is the case for the Dirac structure given in (9)..p.AD. * LyGi C Gi. Now.J.t. H2 9 . the vector field Y should satisfy (22). given in local coordinates by Y(q. ( X .w = 0.AD given by {H1. H~}D = (dH1.u) = Yq(q. which together with (21) and (9) implies that L~.u) = Yp(q. Finally. i. where ker denotes the kernel of a codistribution and ann the annihilator of a distribution. which implies that Yq and Yp do not depend on u.u)~~p + Yu(q.p.p.e. 0 Yq(q. which for the given vector field Y is always satisfied as direct calculation shows. LyGo C Go (so [ Y . A vector field Y 9 F(T(T*Q • U)) is a symmetry of D. LyPi C Pi. a ) 9 D} F(T(T*Q x U)) {a 9 r ( T * ( T * Q x U)) I (O. consider the Dirac structure described in (9) and let Y E F ( T ( T * Q x U)) be a symmetry of D. 1 Remark 4.p). u)= Yq(q. H~}D = {Ly H1. H1.196 G.p. (X.u)~ff. i. By proposition 5. u) o~ + Yp(q. the codistribution Px is involutive.u) = H(q. H~}D + {H1.e.p. Furthermore.p) = Yq(q. In [3] it is proved that the only /fpart always holds. van der Schaft horizontal oneforms on T*Q x U). ~ ] 9 span {~ff}). In [3] the following proposition is proved. Ly H2}D.p)} (21) along with the generalized Poisson bracket on .
Then there exists (locally) a function f1 E C~(T*Q) such that fI is constant along solutions of the implicit Hamiltonian system. Notice that (Y. aefined by dr1 = . X) + (all. together with proposition 6. where / t is defined). Then.Optimal Control and Implicit Hamiltonlan Systems 197 F(T(T*Q)) (i. From classical mechanics it is known that this implies that Y is locally Hamiltonian. H). p. Y)(q(t). . SO/1 is a conserved quantity.Y). H).p) = (PT. / t E C~(T*Q)) which is constant along solutions of the optimal control problem. In fact. Consider the implicit Hamiltonian system (T*Q • U. H). Y))(q(t). Y E rh~ ( T ( T* Q • U ) ). there exists (locally) a function of the phase variables (that is. Notice that from the above proof it follows that Y is a symmetry of (T*Q x U. that is. D. gives a Noether type of result between symmetries and conservation laws of optimal control problems.p. H). this means that to every generalized symmetry of the optimal control problem. 71 (at least locally. u(t)). H is constant along solutions of(T*Q• D. where Y(q.71. with D given in (9). D. H ) . so Y is a symmetry of H. u(t)) = (d[I. L y H = 0 only along solutions of the implicit Hamiltonian system. u) = Yq (q. p) o~ +Yp (q. D.Q). Consider the implicit Hamiltonian system (T* Q x U. p(t). since / t is a conserved quantity. Y))(q(t). that is. Proof. that is. we have the following converse of proposition 6 which. Thus we have proved.e.p(t). Then. X) + (dH. d/4) E D it follows by proposition 3 that Y is a symmetry of D. El Remark 5. D. H). there exists (locally) a function/~ E C~176 such that d H = il~w. D. p(t). along solutions (20) of (T'Q x U. D. Since (Y. d/t) E n (we call Y a Hamiltonian symmetry). assume that Y is a symmetry of the implicit Hamiltonian system (T*Q • U. That is. that is.2]. H). X)(q(t). Let Y E F(T(T*Q • U)) be a symmetry of (T*Q x U. 17" = (PT. u(t)) = (dH. Furthermore.u)o~. see also the results in [3. P) ~p+ Yu(q. H). D. D.Y E F(T(T*Q)) satisfies d / t = i~w for some (locally defined) f u n c t i o n / t E C~(T*Q). t E [0.Q). (23) t E [0. along solutions (20) of the implicit Hamiltonian system 0 = ((dH. is a symmetry of (T* Q • U. u(t)). 0 = ((d/t. Furthermore. H) only in the "weak" sense. P r o p o s i t i o n 6. with D given in (9). Y is also a symmetry of the Hamiltonian H. In particular. P r o p o s i t i o n 7. It follows that any symmetry Y of n in local coordinates has the form Y (q.p(t). Let [1 E C~(T*Q) be a conserved quantity of (T*Q • U.
such that LytH = 0. 3ul.u) O(q. H(. Consider a Lie group G with action r : Q x G + Q on Q.p) ~u ('' " 1..(. In general however. denotes a small time interval..J . as described in the reduction theory in [3. Assume that G is a s y m m e t r y Lie group of the optimal control problem in the sense that V g E G .u2(s)).0 to get (with some abuse of notation) OH(.u2E U s .198 G. In [17] this is called the control theory version of Noether's Theorem.) Notice that ifu = ul = u2 this is exactly (14)... s9 (25) where exp : • r G denotes the exponential map of the Lie group. van der Schaft Finally we remark that generalized symmetries of the optimal control problem give rise to reduction of the corresponding implicit Hamiltonian system (not of the optimal control problem itself).. It follows from proposition 6 that H~Q is . dH~o ) 9 D.J. notice that (Y~. us) or. u) o eg = H(.2]. in an infinitesimal version. V u E U ...)) = x:(. 3u2(s) E U (6'6) s. t . then in [17] it is shown that the m o m e n t u m m a p is constant along solutions of the implicit Hamiltonian system (in their terminology called biextremals of the optimal control problem). J ) . u2) = f(r u).u)or = H(. by factoring out these symmetries and using the conserved quantities to obtain a lower dimensional implicit Hamiltonian system...u~EUs. VuEU.. Define the m o m e n t u m m a p q~ : T*Q + ~* as in section 3...r. Furthermore. f(..t. The lifted action r : T*Q• T*Q defines a symmetry of the Hamiltonian H in the sense that vgEa. J > 0.(r ul) z(.u) ". In [17] the following class of symmetries is investigated. . f ( .Q).=0 which implies that L~oU 9 span{~H}. H(.Vu E U..... for every ~ 9 G. differentiate (25) to s at s .t. u2) = C(r u). 3ut. Indeed.. the reduced implicit Hamiltonian system will not like in theorem 2 correspond to a lower dimensional optimal control problem again. the same result follows from proposition 6..~O = cgH . So even in the case of generalized symmetries we can do reduction to simplify the search for solutions of the optimal control problem.. Example I. yc(Cg(. and where ( . T e g .. with (PT. so Y~ is a generalized s y m m e t r y of the optimal control problem. u ) = f(r Teg. f_..u) .ux) o r (24) + V~ E {~. so there exists a vector field Y~ 9 F(T(T*Q • U))..Y~ = ~q. In our setting. Blankenstein and A. H(.
. Also in this case the Maximum Principle gives necessary conditions to be satisfied by a solution of the constrained optimal control problem.). . . . Consider the cost functional given in (2). (a slight modification of) theorem 2 also holds in case the optimal control problem admits symmetries described in (24).Optimal Control and Implicit Hamiltonian Systems 199 constant along solutions of the implicit Hamiltonian system. by defining b(q. u)] T. bk E C~176 • U). . or. b~.( q . u) = [bl(q.qo. Necessary conditions for a smooth trajectory (q(.)) satisfies (I) and (26). . . see [9]. one could differentiate the constraints again. u()) E QR • U 9 to . u ) . it follows that the momentum map ~ is constant along solutions of the implicit Hamiltonian system.4]): the Jacobian ~ . u(. q) = 0.. Since this holds for every ~ E G. 5 Constrained Optimal Control Problems In this section we generalize the results of the previous sections to constrained optimal control problems. [15. Actually. Due to space limitations we will not persue this point any further here. . u) = 9(q. bk(q. are included in this setup.g. u) = 0 (26) We will assume that the following regularity condition is satisfied ([15. . u) E Q • U. Consider a nonlinear system (1) subject to the constraints bl (q.u) = ~ respectively b(q. Notice that by proposition 2 the momentum map has actually the value zero along solutions of the implicit Hamiltonian system. see also section 3. b(q. . b~(q. see e.) over Qa • U ~ such that (q(. . as well as "nonholonomic" constraints g(q. The constrained optimal control problem is given by minimize J(. by defining b(q. f(q. u) = 0 . Note that "holonomic" constraints 9(q) = 0.4]. and q(O) .4]. assuming that the regularity condition (27) is satisfied (if not. u) (27) has full row rank for all (q. equiv(Q alently. and we have obtained the same result as in [17] (considering the smooth case). The key observation is that the optimal control problem can be transformed by regular state feedback into an equivalent optimal control problem with statespace symmetries described in (14) (given by the action ~). Remark 6. . u)). .). and check if the regularity condition is satisfied this time) [15. u) = O.
u(t)) i (28) (29) OH cOb p(t) = '~q (q( t ). van der Schaft solve the constrained optimal control problem are given by the existence of a smooth curve (q(. a ) e F(T(T*Q x U)) x F(T*(T*Q x U))] a .u(t))~u(q(t).). u(t)) = 0 (31) for all t E [0. where H : T*Q x U + ]~ is defined in (3). This can be seen by the following.p( t ). Notice that the codistribution span {db} is constant dimensional (by the rank condition of the Jacobian (27)). along with the initial conditions q(0) = q0 and transversality conditions (7).p(.J. and . Blankenstein and A. A E ~k are Lagrange multipliers required to keep the constraints (31) to be satisfied for all time. u(0)) = 0. u(t))q(t ) + ~u (q(t). u(t))i~(t) = 0 Ob (32) and note that (31) is equivalent to (32).u(t))A(t ) along with the constraints (30) b(q(t).200 G. u(t))A(t) cOIl cOb 0 = ~ u ( q (t ). Differentiate (31) to get ~~(q(t). Then the Hamiltonian equations (11) are exactly given by (2830. a ) e F(T(T*Q • U)) • F(T*(T*Q • U)) I aq ap au 0 In 0 0 0 Xq Xp Xu o E span I obT L i)u 0[~q 0 ~u ] Xp X~ } (33) where the span is taken over C cr (T*Q x U).p( t ). X e ker (db)} (34) where db is the differential of b. The Dirac structure in (33) can be described in a coordinate free way by D = { ( X .&(X) E span {db}. Here.T]. assuming that b(q(0).)) E (T'Q) ~ satisfying the equations DH q(t) = ~ ( q ( t ) . D. Like in section 2 the equations (2831) can be described by an implicit Hamiltonian system (T*Q • U. H). p(t). Define the Dirac structure D in local coordinates by D = { ( X .32). u(t)) + ~q (q(t).
H).) over ( Q / G ) ~ x U s such that (4('). and by proposition 3 it follows that ~Q is a symmetry of D. lifting the vector fields ~q E _r'(TQ) to the vector fields ~q E F(T(T*Q)) turns G into a symmetry Lie group of the implicit Hamiltonian system (T*Q • U. u ) = 0 . the action of G on Q satisfies (14) as well as the condition L~Qb(. u E U. H). u) o ~r.. defined in (19). by b(. ~ E ker (db)} .2] to reduce the implicit Hamiltonian system (T*Q x U.. and q(0) . u(. results in the reduced implicit Hamiltonian system (~I(O)/G • U. . V~E{~ (35) Let ~r : Q + Q/G denote the projection map. Now we can use the reduction theory described in [3. .X2]) = LxI ( d b ( X ~ ) ) . has constant value zero along solutions of the constrained optimal control problem. .5 [6] it follows that D given in (34) is closed. this reduced constrained optimal control problem gives rise to an associated implicit Hamiltonian system (T ~(Q/G) • U. H) to a lower dimensional implicit Hamiltonian system. it follows that the momentum map. L g Q b j ( .e. T]. k . db)(X~) = (d(db(X1))(X2) = 0). That is. db(Xx) = db(X2) = 0 then db([Xt.. w h e r e / ) is a Dirac structure on T*(QJG) • U defined analogously as in (34) (see section 3 for the definitions of J and H). t e [0. let G be a symmetry Lie group of the constrained optimal control problem. w h e r e / t E C~176 x U) is given as in section 3. .u)=0. and since dH~Q = i~Q00. j = l . On the other hand. u(t)) = O. u(t)) and the constraints b(4(t).. Since the transversality conditions are given by (7).40Again. As in section 3. D. Indeed. The reduced Dirac structure b on 4'1(0)/G x U is given by b = {(2. Following the reduction procedure analogously to the one described in example 8 [3]. Furthermore.5) e r(T(r • U)) • F(T*(~t(O)/G x U)) I . u) : Q/G + ]~k. As in section 3 the constrained optimal control problem projects to the reduced constrained optimal control problem defined by minimize J(. 5. ~q E F(T(T*Q)) can be identified with a horizontal vector field ~q E I'(T(T*Q • U)) on T*Q • U. i.e. . D.. u) = b(. then the function b(. u E U. . Then by theorem 4. where H~Q is defined in (12) and extended to a function on T*Q • U.000(2) E span {db}.(Lx. fl).)) satisfies ~(t) = ](4(t). by (15) ~Q is a symmetry of the Hamiltonian H.Optimal Control and Implicit Hamiltonian Systems 201 that the distribution ker (db) is involutive (i.. D. see proposition 4. as in section 3. it follows that (~Q. dH~Q) E D (notice that by (35) ~Q E ker (db)).u) : Q _+ ~k projects to a well defined function b(.
This implies that r is a Dirac isomorphism between D and D.202 G. Consider the two implicit Hamiltonian systems (T*(Q/G) x U. T h e o r e m 3. u) = b(. Then it follows that b = b o 7.F vert (T(T*Q x U))] C ]vert (T(T*Q x U)). fI) and (T* (Q/G) x U. Yq and Yp do not depend on u.35). Define the m a p (r.1 ( 0 ) r T*Q the inclusion m a p and ~'0 : ~1(0) >4~1(0)/G the projection map. is equivalent (up to isomorphism) to first applying the Maximum Principle on the original constrained optimal control problem (on Q). and the corresponding (co)distributions Go = F wrt (T(T*Q x U)) N ker(db) G1 .span {db} el = Fh~ (T*(T*Q x V)) + span {db} and note that P1 is constant dimensional (by the full rank condition on the Jacobian (27)) and involutive (since G0 is involutive). Blankenstein and A.u)o 7ro. for all (z. we investigate the generalized symmetries of the implicit Hamiltonian system (T*Q x U.e. p. qo) = b(qo. u) = rq (q. p) ~ +rp (q. . Indeed. i. As in section 3 there exists a diffeomorphism 7. with i : 4 . and prove a (partial) analogue of proposition 6.. D. u) o i(qo. Consider the Dirac structure D given in (34). f'I) are isomorphic. consider a vector field Y E F(T(T*Q x U)) of the form (in local coordinates) Y (q. Proposition 5 gives necessary and sufficient conditions for a vector field Y E F(T(T*Q x U)) to be a s y m m e t r y of D.u)o i = b(.u) E ~I(O)/G x U. also denoted by v. corresponding to the constrained optimal control problem. fI) and ( ~ . Now.2] to reduce the corresponding implicit Hamiltonian system. D../5. it follows that the two reduced implicit Hamiltonian systems (~I(O)/G x U. see section 3. Consider a constrained optimal control problem with symme tries described in (1~. H). p. van der Schaft where b E CC~ x U) is defined by b(. called the reduced constrained optimal control problem.u E U.J. and then using the reduction theory described in [3. Then reducing the constrained optimal control problem to a lower dimensional constrained optimal control problem (on Q/G). This can be expressed in a coordinate free way by the condition that [Y.ker (db) P0 . Finally. u) o~. id) : #I(O)/G x V ~ T*(Q/G) x U. u) = o (qo) = o 7(z. p) ~ + r ~ (q..I ( 0 ) / G x U. and applying the Maximum Principle. H).: 9 .D.I ( 0 ) / G > T*(Q/G) such that 7*03 = w0 (where 03 is the canonical symplectic twoform on T*(Q/G) and w0 the reduced symplectic twoform on #I(O)/G). D. This result extends the result of theorem 2 to the case of constrained optimal control problems. Since also /~ = / : / o r. see section 3 for notation.
assuming that the action of the symmetry Lie group G leaves the manifold Q! invariant. q(T) e Q). In particular. where Y = (PT*Q).[I(q.il>w. Consider the implicit Hamiltonian system (T*Q x U.2] to the implicit Hamiltonian system corresponding to the original optimal control problem. it follows that (Y. "}D. Now assuming that Y E ker (db) and extending H to a function on T*Q fl U (by ft(q. In this paper we restricted ourselves for simplicity to the free terminal point (constrained) optimal control problem (i. F ~er' (T(T*QxU))] C F ~'~ (T(T*Qx U)).p. dH) E D. that is. We have the following (partial) analogue of proposition 6..e. i.. there exists (locally) a function /~r E C~176 such that dH . Remark 7.u) E T*Q x U).AD it follows that Lf~w . there exists (locally) a function of the phase variables (that is.p)} C {H E C~176 x U) I dH E Pl} . Then there exists (locally) a function Y e C~176 such that ft is constant along solutions of the implicit Hamiltonian system.Since {H EC~176 x U) ] H(q.p. where Q! is a submanifold of Q. If one furthermore assumes that Y is a symmetry of the Hamiltonian H. symmetries of the corresponding implicit Hamiltonian system.p. analogously to proposition 7.Y. D. Let Y E F(T(T*Q x U)) be a symmetry of (T*Q • U. H).0.Optimal Control and Implicit Hamfltonian Systems 203 Assume that Y is a symmetry of D. We leave it to the reader to state the corresponding converse statement. to implicit Hamiltonian systems. D.2] was used to prove that doing reduction on the optimal control problem and applying the Maximum Principle. then by (23) it follows that ~r is a conserved quantity. with D given in (34). This implies that 17"is locally Hamiltonian.V(q. The reduction theory developed for implicit Hamiltonian systems in [3.p). is equivalent (up to isomorphism) to applying the reduction theory in [3. P r o p o s i t i o n 8. by canonical lift to the cotangent bundle. H) such that Y e ker (db) and [Y. then by proposition 5 it follows that Y is canonical with respect to the bracket {.u) . /~ e C~176 which is constant along solutions of the constrained optimal control problem. It is easy to prove that all the results remain true if we add the following constraint at final time: q(T) E Q! C Q. this means that to a generalized symmetry of the constrained optimal control problem satisfying the assumptions in proposition 8. doing reduction . We showed that symmetries of the optimal control problem induce.u) = H(q. via the Maximum Principle. 6 Conclusions In this paper we showed that optimal control problems naturally lead.e.
W. John Wiley. 9. Blartkenstein and A.E. 29(11):10371040. J. The Structure and Optimization of Nonlinear Control Systems possessing Symmetries. 1975.I. 6... 5. second edition. 2. J. 1992.J. which are defined as symmetries of the corresponding implicit Hamiltonian system. Blankenstein and A. 1975. Foundations ol Mechanics.. G. van der Schaft.J. Dirac Structures and Integrability of Nonlinear Evolution Equations. Estimation and Control. Math. Soc. On representations and integrability of mathematical structures in energyconserving physical systems. Blankenstein and A. SpringerVerlag. 3. van der Schaft.H. Applied Optimal Control: Optimization. . P. T.W. 8. The key observation is that the momentum map. 1987. M. and we showed that to every generalized symmetry there corresponds a conserved quantity.C. IEEE Trans. The commutation of doing reduction and applying the Maximum Principle was also proved there. being a function of the phase variables which is constant along solutions of the optimal control problem. Dirac manifolds. Dorfman. Dalsmo and A. Trans. Dordrecht. Rishel. 1984. 11. 1978. American Math. Kummer. I. 319:631661. Reduction of implicit Hamiltonian systems with symmetry. Cambridge University Press. Marsden. 4. Karlsruhe. van der Schaft and applying the Maximum Principle commutes. Fleming and R.17]. Reidel.J. second edition. Marcus.E.J. R. Marie. Indiana Univ. Automatic Control. Symmetry and reduction in implicit generalized Hamiltonian systems. 13. A. University of Twente. 12. 1983. van der Schaft. Bryson and Y. Deterministic and Stochastic Optimal Control. 1999.E. SIAM J. University of Texas. Courant. 37(1):5491. Accepted for publication in Reports on Mathematical Physics. J. References 1. On the Constuction of the Reduced Phase Space of a Hamiltonian System with Symmetry. M. 10. June 1999. Opt. G. has the (constant) value zero along solutions of the optimal control problem.M. Cont. Grizzle. Finally. Chichester: John Wiley.W. 7. 1990. Furthermore. in the last section we generalized the results to constrained optimal control problems. Abraham and J.W. Benjamin / Cummings Publishing Company. ECC'99. Grizzle and S. Faculty of Mathematical Sciences. Marsden. Ho. Memorandum 1489. 1999. PhD thesis. This result generalizes the results obtained in [19. In Proceedings of the 5th European Control Conference. It was shown that under some conditions generalized symmetries of the corresponding implicit Hamiltonian system again give rise to conserved quantities. Libermann and C. Symplectic Geometry and Analytical Mechanics. 30(2):281291. 1993. 1981. J. Lectures on Mechanics. we described the generalized symmetries of an optimal control problem. Optimal Control of Systems Pocessing Symmetries.204 G. corresponding to a symmetry Lie group of the optimal control problem.
25(2):245259. In A.J. 1974. 5:121130.V. 1995. Canonical Reduction of Mechanical Systems Invariant under Abelian Group Actions with an Application to Celestial Mechanics. 1977. 1964. 1997. editors. Polish Academy of Sciences. A. Imperial College Press. The Mathematical Theory of Optimal Processes. 21. Melchiorri. and R. 20. Weinstein. Symmetries in optimal control. C. 16. Rep. Phys. A. Math.. van der Schaft and B. W. New York. Geometry in Nonlinear Control and Differential Inclusions. Rep. Vinter. pages 115. pages 463557. J.J. In A.S. Pontryagin. Marsden and A. Math. Modelling and Control of Mechanical Systems. In B.Optimal Control and Implicit Hamiltonian Systems 205 14. Respondek.M Maschke.J. R. lmplicit Hamiltonian Systems with Symmetry. editors. An introduction to the coordinatefree Mamimum Principle. 1998. Math. Marcel Dekker.. Phys. W. Astolfi. Jacubczyk. Gamkrelidze. Math.J. Geometry of Feedback and Optimal Control.. Control and Optimization. Warsaw. J. A. Sussmann. Indiana Univ. van der Schaft. Fryszkowski. Mischenko. D. Rzezuchowski. H.J.F. Satzer. SIAM J. pages 379393. 17. 41:203221.J. Banach Center Publications. L. Interconnected mechanical systems. 19. Tornamb6. editors. Boltyanskii.E. B.B. Sussmann. 15. A. and T.N Limebeer. H.G. Jacuhczyk and W. Reduction of symplectic manifolds with symmetry.J. van der Schaft. . Inst. V. 26(5):951976. Pergamon Press. 18. 1997. part I: geometry of interconnection and implicit Hamiltonian systems. Respondek. Symmetries and Integrals of Motion in Optimal Control. and E. 1987.
R.P. which express that circle or Popov criterion holds. fr .A. rational (that is.Robust Absolute Stability of Delay Systems PierreAlexandre Bliman* National Technical University Department of Mathematics Zografou Campus 157 80 Athens. In order to tackle this robustness issue. for their help and support.g. One of the problems one may have to deal with.I.N. delays in control loops may lead to bad performances or instabilities. Popov criterion) may be obtained equivalently by Lyapunov method or by frequency domain method. Greece P i erre .. 1 Introduction As is wellknown. but also for reasons of computational simplicity. We construct LyapunovKrasovskii functionals candidates for various such systems.17]. the usual results (circle criterion. such a property is unknown. 78153 Le Chesnay Cedex. whatever the value of the delay in a certain compact interval containing zero (delaydependent criteria). Rocquencourt B. In the classical absolute stability theory for nonlinear. The present paper is devoted to the study of absolute stability of delay systems with nonlinearities subject to sector conditions.. We are here concerned with the issue of stability of delay systems with sectorbounded nonlinearities and uncertain delays. see e. mainly for linear delay systems.a l e x a n d r e . bl iman@ inria. Phone: (33) 1 39 63 55 68.A. delayfree). whatever the normegative value of the delay (delayindependent criteria). B 1 imanO inr i a. whose decreasingness along the trajectories is expressed in terms of Linear Matrix Inequalities (LMIs).R.A1 exandre. However. one is able to provide LMI conditions under which circle or Popov criterion holds. and with uncertain delays. 105. since at least the early * On leave from I. and the coordinator of the greek team Ioannis Tsinias. Emaih p i e r r e . Fax: (33) 1 39 63 57 86. f r A b s t r a c t . Also. is the possible uncertainties on the value of these delays. France. the surveys [13. The author is indebted to the network coordinator Fran~oise LamnabhiLagarrigue. We then show that feasibility of the latter implies some frequency domain conditions. and the study of timedelay systems has attracted large interest. systems. using a class of transformations of the system.N.I. For delay systems. Permanent and corresponding address: I. with the support of European Commission's Training and Mobility of Researchers Programme ERB FMRXCT970137 "Breakthrough in the control of nonlinear systems". stability criteria independent of the values of the delays have been proposed.
16]. 3. One reduces the initial problem to the delayindependent absolute stability of an auxiliary system with two independent delays. and the machinery of the proofs. to the best of our knowledge. One is then able to state circle criterion (Theorem 4) and Popov criterion (Theorem 5) on this transformed system. A generalization of this result was then made for systems with two independent delays.15] criteria have been shown to be valid for delay systems. It was then shown that the previous results naturally imply some frequency domain conditions. whose coefficients are functions of the bound h and the other data. in a way reminiscent of some recent results on delayindependent stability of linear delay systems [9. +co). giving sufficient conditions of absolute stability for any value of the delay h E [0. In the sequel. More precisely. essentially by addition of a Lur'e term to a quadratic LyapunovKrasovskii functional. Generally speaking. In a previous paper [3]. we simply assume that there exist global solutions of the various delay systems under . after recalling some more details on the background in Sects. we attempt to establish some new sufficient conditions for absolute stability of delay systems subject to uncertainties on the delays. from the simplest to the more complex ones. considerations based on classical LyapunovKrasovskii functionals permitted to obtain some Linear Matrix Inequalities (LMIs) whose feasibility is sufficient for delayindependent absolute stability. in other words. This was.hi.23] and Popov [19.3. Section 34]. The idea is to perform a transformation of the system under study. we do not consider here the extensively studied questions of existence and uniqueness of the solutions. circle [21. Last. in [3] we provided a delayindependent version of circle criterion for systems with one delay. and Krasovskii has proposed an extension of Lyapunov method to these systems. The new results are provided in Sect. 1. we repeat the delayindependent versions of circle and Popov criteria for systems with one delay (Theorems 1 and 2 in Sect. leading in particular to some sufficient conditions for stability of linear and nonlinear delay systems independent of the value of the delay(s) [14.1 to 1. the first a t t e m p t to introduce frequency domain interpretation of LyapunovKrasovskii method for nonlinear systems (the issue of the conservativity analysis of this method is examined in [4] for linear systems and in [2] for nonlinear systems). 2). according to some transformation whose principle has been applied to linear delay systems [18]. In the present paper. similarly to the procedure used in the delayfree case.208 PierreAlexandre Bliman sixties. a version of circle criterion valid for any value of the delay h E [0. An effort is made all along the paper to present in a unified and methodical way the results. a delayindependent version of Popov criterion was finally obtained. where h is a prescribed nonnegative upper bound. obtained from circle and Popov criteria by decoupling the terms in s and z = e h8 in the transfer function of the system. and provide shortened and simpler proofs. In the present paper.
. C). Nonlinearity r : ~ + • il~P 4 ~q is assumed to fulfill a sector condition: there exists K E ~q• such that Vt e ~ + .A ) . As an example for system (9) below. . (2) where A E ~nxn.PB) R = \KC . r r > 0. 0 ] or [2h. this means by definition: Vr E r 3x E C([h.I B . B.+cx~)). . In particular. defined on [ . Last. . B E ~nxq. . . the n x n identity matrix and the n x n and n x p zero matrices.B TP 2I _ < 0 (3) is sufficient for absolute stability of system (2). . sometimes simply abbreviated I. xt denotes the mapping x(. that is H(s) " C ( s I . according to the context. for which there exist powerful numerical methods of resolution [6]. C E II~pXn. such that x0 . (1) The absolute stability problem consists in seeking conditions on H and K for stability of the interconnection. .+c~))nAC([O.p denotes the integers i .h .1.Robust Absolute Stability of Delay Systems 209 consideration. the system may be realized as: =AxBr y=Cx. The symbol  denotes Kronecker product. On. y(t)) is Notations Below. on R + clef [0. 0]. one verifies that the inequality  ~r \KCBTp 2I \r is fulfilled along the trajectories of (2). and by d i a g { M 1 . When H represents a strictly proper rational system (A. so the existence of a matrix P E ~'~ xn such that p = p T > O. On• are resp. The matrices In. 1. Searching for a Lyapunov function of the type V(x) = x T p x for a certain positive definite matrix P E ~nxn. . . 0. Problem (3) is a Linear Matrix Inequality (LMI). . i = 1 . + t). p . p "on the diagonal". V y e ~P. under sector condition (1).e. .1 Circle Criterion for Rational Systems: Equivalence Between Frequency Domain and Lyapunov Methods Consider the negative feedback interconnection of a memoryless multivariable timevarying nonlinearity r with a linear timeinvariant system characterized by its transfer function matrix H(s) of size p • q. +co). Mp} is meant the blockdiagonal matrix obtained putting the (not necessarily square) matrices Mi.r and (9) holds a. i . .1 . f ATp + PA CTK r . the m a p t ~4 r defined almost everywhere.
and fulfills sector condition: there exists K = diag{Ki} E IRpxp K _> 0. • II~ TM. by use of inputoutput techniques [21. and r is a function of yi only.2 Two A b s o l u t e S t a b i l i t y R e s u l t s for Delay S y s t e m s (8) Let us now consider the following delay system (where Ao. C1 E I~P• Bo E (9) ~=Aox+Alx(th)Bor y=Co~+Clx(th). namely 3s > O. may also be proved directly under form (4). IQ f0(c~) ' r i=l du .P B "~ K C + r I K C A .B T p 2I . d e t ( s I . decentralized (that is [13] p = q. (4) where . Vs E C with Res > 0.1PB "~ ( x(t)"~ K C + y K C A . the famous circle criterion.y K C B . such that vy e r r > 0.23]. along the trajectories of (2). the derivative of the Lyapunov function ~[Y(x(t))] is bounded from above by z(t))T( r ATp+pA CTK+ATCTK.210 PierreAlexandre Bliman KalmanYakubovichPopov lemma (see e.] \ r This leads to the LMI problem r/= diag{m} >_O.T1KCB .A) r 0 and 2 I + KH(s) + [KH(s)]* > sI . det(sI .A) r 0 and 2I + ( I + Tls)Kg(s) + [ ( I + ys)KH(s)]* > e I . A T p + PA CTK + ATCTKrl . One then verifies that. Suppose now that r is timeinvariant. for i = 1.A1 E ~ . 3r/= diag{yi} _> 0.g. [20]) permits to express equivalently the previous condition under a frequency domain form. It is obtained searching for Lyapunov function under the form V(x) = xT Px + 2 ~P y. (6) def where the nonnegative diagonal matrix r/ = diag{yi} is a free parameter. denotes complex conjugation. = (7) whose solvability is equivalent to the frequency domain condition 3e > 0.~). Co. P R = pT .BTCTKTIJ < 0.B T p 2I. This absolute stability criterion. 1. (5) A refinement of circle criterion is then Popov criterion.BTCTKy.> O. .Vs E C with Res _> 0.
22] v(r ~f CT(0)pr + /. one is led to assume solvability of a certain LMI.e . these frequency domain conditions depend explicitly upon the magnitude of the delay h. (10) For such a system. instead of (8) in the case where the nonlinearity is timeinvariant and decentralized. d e t ( s I . the linear equation x = Aox + A l x ( t .Ao .z) E j R • C with s ~ 0. ~n) _. the obtained result is a delay. .' Bo . d e t ( s I .Ao .Q are positive definite matrices. By definition. instead of (4) (with H now given by (10)). To analyze the conditions under which this property occurs. The delay does not appear in this LMI: contrary to circle criterion. are sufficient for absolute stability of system (9) under hypothesis (1).eSh A . circle criterion [21.)(sI . [9] have shown that delayindependent (asymptotic) stability of the previous linear system is equivalent to (q denotes the spectrum) Rea(A0 + A1) < 0.independent criterion.A o . see [13.e'h A1) ~s 0 and 2I + (I + ys)KH(s) + [(I + ~ls)Kg(s)]* > eI . and V(s. or 3e > 0. r162 h a. (11) where P.12] if the roots of the equations d e t ( s I . d e t ( s I .Robust Absolute Stability of Delay Systems which realizes the interconnection of r with the nonrational transfer 211 H(s) = (Co + eShC.23] and Popov criterion [19. VsE C with Res > 0. Obviously.. 37/= diag{r/i} > 0.3 DelayIndependent Stability of Linear Delay Systems Important progresses have been made recently in the study of delayindependent stability conditions for linear delay systems. whose unknowns are P and Q. Izl = 1.A o eShA1) r 0 and 2 1 + K H ( s ) + [KY(s)]* > e l . it is possible to obtain absolute stability criterion by use of a quadratic LyapunovKrasovskii functional V : C([h.15] still hold: the following condition 3e > 0. one sees that it is necessary to disconnect the term e sh from s. Hale et al. On the other hand.~ ~ of the type [14.Ao .a S A1) = 0 have negative real part for any h _> 0. Vs E C with R e s > 0. As for the rational case.zA1) ~ O. in an attempt to force the map t ~ V(xt) to decrease along the trajectories of (9). see formula (12) below.h) is asymptotically stable independently of the delay [11. 0]. 1. ) .17].
2 2.1.Ao . 9 . d e t ( s I . Then. PAl C T K T . Then (ii) holds.zA1) r O. and V(s. Furthermore.I n d e p e n d e n t Circle C r i t e r i o n In [3] is given the following result. Suppose (i) holds.Ao . They also provide method to check a slightly stronger one. Remark that this condition is equivalent [10.P B o ~ Q CTKT I < 0.e'hA1) # 0 and 2I + KH(s) + [KH(s)]* > eI . det(sI .Vh _> 0. lim x(t) = t * + oo 0 for any global solution x of (9). d e t ( s / . KCI 2I a / (12) (14) \r y(t)) / \r y(t))/ for any global solution x of (9). IzI < 1.17] have given numerical methods of verification of this condition.z) E Ca with Res > 0.Ao . Suppose (ii) holds. the functional V defined in (11) is positive definite for any solution P. namely Recr(Ao + A t ) < 0. Q E IRnxn such that P=pT >o.ef i A1 P \ KCo. z) E fir x C with Izl .Q of (12) and.zA1) r O. for any r fulfilling (1) and any h >_O. (13) where H is given in (10). [16.212 PierreAlexandre Bliman Niculescu et al. called strong delayindependent stability.1 Absolute Stability Criteria V a l i d f o r a n y h E [0) + c ~ ) D e l a y . Remarks 1 and 2] to V(s. T h e o r e m 1 Consider the two following statements.BTp The following property holds. (i) There exist P.Vs E C with Res >_ 0. [ATp+PAo+Q R d_. 3e > 0. for any r fulfilling (1) and any h > 0. Q = Q T >O. using matrix pencils.
zIn R zln = (A T + z*AT)p + P(Ao + zA1) + (1 . d e t ( s I . 0.zA1) # 0 and 2I + KH(s..P BoP O 0 [zSBo \ Iq / = B T s ' [ ( s + s ' ) P + (1 IzI~)Q]SBo. [z[ < 1. . This is the first part of (16).Ao . g(s.zA1) 1 . (4)).z) E C 2 zSBol Iq / I \ A. z) d_e__f + zCl)(sI . (13)) must be compared with (3) (resp. R < ediag{0. From (i) is deduced that.. for any z E C. z) + [KH(s. z)]* > e I .[zl2)O < O.S B o ' ~ Iq o CTZ*' /zSBo/ / / KCoKC12I / \ = . Iq} .z) aJ ( s I . Define the multivariable p x q transfer function matrix H by H(s.z)]*). and this yields (ii).Robust Absolute Stability of Delay Systems 213 Statement (12) (resp. (16) where H is defined in (15). Matrix R in (12) is of size 2n + q. Oqxn Oqxn One deduces from this that the real part of the eigenvalues of Ao + zA1 are negative when Izl < 1. Denoting for sake of simplicity S = S(s.Ao .( 2 I + KH(s. (SBo'~* ( zSBo] Iq / ~ 0 CToKT'~ [ .z) + [KH(s.Ao . We shall prove that the following property holds. from which (ii) is deduced immediatly: 3e > 0. (17) the second part of (16) is deduced from the following inequalities. valid for any (s. Sketch of the proof.z) e C 2 with Res > 0. (Co (15) to be compared with (10). and from the remark that Se > 0. Assume that (i) holds. Proposition (16) is hence proved.zA1)l Bo .
(20) + 2 E~liIs i=1 Jo The following result is enunciated in [3].h)). differentiate V along the trajectories of (the global solutions of) (9).1]. plus the Lur'e term (19). (19) in a way similar to what is done for rational systems. is a free parameter.2h) and r . However.22]. one is led to introduce also more ancient state values in V.2 DelayIndependent Popov Criterion We consider here system (9) with timeinvariant nonlinearity: = Aoz + A l z ( t  h) . In order to get a delayindependent version of Popov criterion for system (18). this addition introduces terms in x(t . defined for any r E C([2h.B0r y = Coac + C l z ( t  h) . For other details. one may try to add to the quadratic LyapunovKrasovskii functional (11) the Lur'e term 1. To prove global asymptotic stability of (9).214 Pierre. [] Remark incidentally that the fact that (i) is sufficient for asymptotic stability of (9) has been here deduced directly from (ii). Theorem 31. in the case where the input of the nonlinearity is not delayed (C1 = 0). see complete proof in [3]. and add the nonnegative term 2r y(t)) T ( K y ( t ) . it is sufficient to remark that (ii) implies that the hypotheses of circle criterion are verified for any h > 0. O]. this furnishes an alternative to the proofs k la Krasovskii [14. i=I /u~(t) dO r au. 2. One considers finally the following LyapunovKrasovskii functional. Suppose now that (ii) holds.Alexandre Bliman To get inequality (14).r y(t))). This permits to get sufficient stability results. e. (18) Throughout this section is assumed that p = q. To take this into account. and that r is decentralized and fulfills sector condition (5) for a certain nonnegative diagonal matrix K. when the input of the nonlinearity is delayed (y :~ Cox). matrix 71 = diag{~li} E ~P• 7/ > 0. . see [8. which do not permit to bound the derivative along the trajectories of (18).~n): v(r aef ( r = P kr p + h \r h)] r(Cot(0)+clr Q \r r h)J du.g. of the functional obtained as the sum of V as given by (11).
Robust Absolute Stability of Delay Systems T h e o r e m 2 Consider the two following statements. Vs E C with Res _> 0. P = pT > O. Q where = QT > O. Q of (21) and. (22) .e. for any decentralized r fulfilling (5) and any h > O. J T R J < 0.]s)KH(s) + [(I + ys)KH(s)]* > eI . on lR + . Furthermore.+2v) ~K (CoAo CIAo CoAl C1A1 CoBo C1Bo) 0p x (4n+~p) (ii) The following property holds 3e > 0. 3r/= diag{r/i} E ]Rv• > 0. d e t ( s I .e'hA1) r 0 and 21 + (I + . Q E IR~nx2n. such that ~1k O. the functional V defined in (20) is positive definite for any solution P.Ao . p.h))] t a.)] _< / =(t. P(I2 O (I~  KC~) @A1)(I2174 ) (I~  CT K) 2I~p x(4. where H is given in (10).2h) / / r / \r jTRj .he } ~ ]~(4n+2p)x(3n+2p) R d_e__f ( (I~  ATo)P + P(I~  (I~  A 1 )P (I2  KCo) . Vh _> 0.(I2  Bro)F + + ( ~K (CoAo ViA0 CoAi CiAi CoBo C1B0) 0px(4n+2p) 04n X(4n+2p) 04. (21) j d__ef diag 1 0 @ In.2h) / / r J \r . d ~[y(x.h))/ / =(t. Suppose (i) holds. 215 (i) There exist r] = diag{r/i} E ]Rpxp. Then (ii) holds.
z) + [(I + . Writing x h ' one has x(t) ) x(t . V(s. z)]* > eI.h. lim z(t) = 0 for any global solution x of (18). Suppose (i) holds. det(sI . In other words.Q E ]Rn• then (21) is fulfilled for 7/ = 0 and I2  P. where H is given in (15)./2 @ Q: as for the delayfree case.z) 6 C 2 with Res > 0. The decomposition J T R J is introduced in order to underline the proximity with the previous results.Ts)KH(s. timeinvariant nonlinearities) to a stronger stability result. for any decentralized r fulfilling (5) and any h > O. 3 r / = diag{r/i} 6 IRpXp.go .r/_> 0. the addition of a Lur'e term leads (for systems with decentralized.Ts)KH(s. compare with R in (7) and in (12). (23) .h) 9 (t .h)) When (12) is fulfilled for P. 9 t + t.zA1 ) r o and 2I + (I + . Sketch of the proof.h. We shall deduce (ii) from the following proposition: 3~ > 0. u(t .216 PierreAlexandre Bliman for any global solution x of (18). Then. the "delayindependent Popov criterion'in Theorem 2 is stronger than the "delayindependent circle criterion" in Theorem 1. U(t . Suppose (ii) holds.2h) r r .h)) =J x(t) ) x(t . [z[ < 1.e ~ Matrix j T R J in (21) is of size 3n+2p.h) r r y(t)) .
(in) In zI.~ p zIn Q zI. z)Bo w = w(s. . z)* z& () [(s + s * ) P + (1  In Izl2)q] z& S(s.z)Bo () 0 w'J T o 0 o (I2  CTK)'~ (I2  c T K )  Jw (I~  KCo) (I2  KCl) =(lz) 212p / 2(2I+ KH(s.] In "~" [(I2  (AT + z*AT))P +P(I2  (Ao + zA1)) + (1 = (AT +z'AT) +(1]zJ2) Izl2)Q] ( Iz& n) (24) ()() (. notice that. (2s) one has.Robust Absolute Stability of Delay Systems First.~ In zln In * In + (Ao + zAt) zln P zI~ ()() which permits to show. ~)Bo/ % zS(~.z)+[KH(s.o j ~T~ I z~'~ I \02pxn/ zIn = I zln ] \02pxn/ zIn [ zln ] (I2  AT)p + P(I~ V Ao) + Q P(I2  AI) (5  AT)p Q = \zI. 217 o> i z~. S(s. as in the proof of Theorem 1.o). that the eigenvalues of Ao + zA1 have negative real part whenever ]z] _< 1: this is the first part of (23). z) E C~ . for any z E C. for any (s. z) def I /_z2S(s. one verifies easily that.( I 2  BT)p 0 0 In = B[S(s. w*JT (( Is  AT)p + P(I2  Ao) + Q P(I2  A1) P(I2  Bo)\ (I2 @AT)p Q 0 ) Jw .z)]*). z)Bo I . and defining [ I S ( s . z) being defined in (17). On the other hand.
whose proof is now complete.e*h $ ) = sI(Ao+D)(A1 . say. Now.[ysKH(s.s h ) ( I . det(sIA0Ale *h) ~kO and 2I + KH(s) + [KH(s)]* > eI . [] 3 Absolute Stability C r i t e r i a Valid for any h E [0. (27) where H is given by (10). system (9). it is sufficient that 3e>0. V(s. let D be a matrix of same size than Ao. 3 r / = diag{yi} > 0. for any value of h in [0. z)]* .D e p e n d e n t Circle C r i t e r i o n 3. h].A1 (D E IRnxn). VsECwith Res_>0. The asymptotic stability is proved as in Theorem 1. One then deduces (ii).VhE[0.e 'h S . h > 0 Given D e l a y .h 1 .218 and PierreAlexandre Bliman w*J T ~IK (CoAo C1Ao CoAl C1A1 CoBo 0p x (4n+2p) T1K (CoAo C1Ao CoAl C1AI CoBo C1Bo) Opx(4n+~p) Jw + = 71sKH(s . z ) + I+l. where the upper bound h is a prescribed nonnegative quantity.1 We are now looking for sufficient conditions of absolute stability of.D ) e *h +AoD1 .}s KH(s.h]. One deduces that: 3r > 0. From the identity ( s I . z) .. For this. For other details. Izl = 1. 2I+ ( 1 ) [( ) ]* I+~71s K H ( s . hi.D 1 .e *h + Ai De_.z) >eI.z) E C2 with Res > 0.Ao  A l e . This leads finally to the second part of (23). see complete proof in [3]. (26) The previous inequality is then extended to the ball {z E C : Iz] < 1} by maximum modulus principle.
a 2 .zlz2hA1D) • 0 and 2 I + K H ( s .h X(V) dr.e .(A1 . k). where k def f t t.D ) e . a > 0.z s .(Ao + D) .' h by zl and 7. 3~ > O.c~ C1De_Sh 1 . z l . The proof follows from the fact that the latter expression is identical to the sum of three nonpositive terms: 1 .z~)]* > e I ._ e .b s = (2(1 .I ~ C o D . Let s = a + ib with (a.0)}. 3 For any s E C \ {0} such that R e s >_ O. (30) .~ .s " . then H(s) = (Co+Cle sh .by hzs.(go + D) .1) + ((1 ..z l ( A 1 .zlC1 .' h +AoD1 .D) + zsf~AoD + zlzsf~A1D) 1 Bo .~ h + A 1 D e _ S h 1_______ )_I B ~ S S (28) The previous expression leads to a "nonminimal"realization of the transfer H . using e. it is sufficient that 3 D .z l ( A 1 .a 2 .b s < 0 . Izlh Iz~l _< 1. C1. z l .a ) 2 .~ .e sh _ .e . [] In consequence.s h 219 is invert ( s I .Robust Absolute Stability of Delay Systems and the analogous one for Co. z2) d__ef (Co A.2e ~ cos b + e . the f o l l o w i n g inequality Proof. z s ) + [ K H ( s .aS). Then the desired property is verified if and only if 1 .D ~ ible. the augmented state (x.2 e ~ cos b + e .D) + z2hAoD A. = We now introduce the following auxiliary result. z 2 ) E C a with R e s _> 0. j _e oh (29) obtained by replacing in (28) e . if I . define a new transfer H ( s .a + bS(e .V(s. d e t ( s I .g. zl. one deduces that. In order t h a t (27) is fulfilled.zlzsftC1D) ( s I . z l .(Ao + D) . Lemma holds.b) E ~2 \ {(0.c o s b) bS)e ..
hi) .h2) .h~/ y ( t ..Ys) . (31) Remark.h2.hi .h i ) . Y3 def r y(t .hi .f~AoDx(t .D ) z ( t . that it does not seem possible to achieve this reduction by a change of variable: the frequency domain nature of the method seems essential. The next step towards our results consists in expressing LMI conditions sufficient for delayindependent absolute stability of (31).h C o D x ( t .r t. = C o x + C l X ( t . Such conditions for systems with two delays are given in [3].h ~ ) .4~x(t.4oX(t)+A~x(~. x(t .2h2) Y4 }'5 r r r .h~) .h2)I x ( t .2hl) / (33a) r Y1) Y(t) def= / . and in [5] for linear delay systems.hit y(t .h2) / .~o~(~.220 PierreAlexandre Bliman where H is given by (29).h~) . }'3) .2h2). (32) where x(t) x(thl) / X(t) def__ x ( t . ]I4) (33b) .2hl) + C 2 X ( t .h2) .r(t)).h l .h2. however. u(t .B0r y(Q).2h~.hi.h C i D x ( t .~h~). They are obtained by "rewriting" system (31) as {~ = .hi . = CoX(t) +C1X(t .hx .. Condition (30) may be interpreted as a condition for delayindependent absolute stability of the following auxiliary system with two independent delays i = (Ao + D ) x + (A1 .2hi)+. Y2) .h A 1 D x ( t .
(34d) O0 00 0 0 hA1D O0 O0 00 0 0 hC1D O0 . (34b) A1 def = C1 def = / ~ 0 0i/ { 00 At  o o co/ D 0 0 0 0 0 0 0 A1 .C~ : 5p x 5n. (34e) .A1. (34f) Bo d__ef Q Bo.D 0 hAo D hA I D (34c) C1 0 hCI D 0 0 . The trajectories of system (31) may be obtained as projections of trajectories of system (32): asymptotic stability of (32) implies asymptotic stability of (31). (34g) Bo : 5 n x 5 q .h A 1 D 0 hAoD 0 0 0 0 C~ d_ef .A2 : 5nx5n. Co.D hA1D 0 0 0 At .h A o D . : 5q x 5p.42 d_ef . D hAoD 0 0 A1D 0 Ao+D 0 Ao+D o) . .. 221 (34a) Co def= Co0 0 0 . 15 I5 The size of the previous matrices is Ao.D hAoD +D 0 0 Ao+D 0 0 0 0 hCoD C1 Co def Ao = hA.Robust Absolute Stability of Delay Systems and Ao + D A1 Ao 0 0 0 0 0 Co 0 . K.C1.h C 1 D 0 hCoD 0 0 {o /~ 0 0 0 0 0 C1 0 C1 hCo D hC1 D . ]C d___ef  K .h C o D .
Writing that the derivative of V should be negative definite along the trajectories of (32). [9n+5q ICCo . we are finally led to consider for system (32) LyapunovKrasovskii functionals in the class v(r ~f ~r(o)p~(o) + f 0 2h~ 0 OT(r)Ql~(v) d r + [ where q~(t) def / [ r ] J2h2 ~T(r)Q2r dr. namely V(~) ~ f r + f_~ 2hx ~I~T(T)~I~(T) dT 21 f ~)T(T)Q2~]~(T ) dT . 0]. Bo. Co. Q2 = QT > O. . C~.h~) \ r 2h~) defined for r E C([4h.I~n). ffTTzfl < 0 .42.0000)] 0 1 01 ~ 00 000 000  In. QI = QT > O. . Q~ E ~SnxSn such that 70 : 70T > O. T h e o r e m 4 Consider the two following statements. IC are defined in formulas (34a) to (34g). 2h2 (35) for any r E C([2h. r h~) h2) r .h. 2 where j def diag = {[. h2}.. Q1. (36) lr . one has the benefit of a natural class of quadratic LyapunovKrasovskii functionals. c~ T ] <~ ICCl IEC2 215q / and . In order to integrate this information. h def max{hi.BT70 (37) E ]]~(15n+Sq)x(14n+5q) .222 PierreAlexandre Bliman System (32) having independent occurrences of the two delays. ~sn). one obtains the following result. A~.70Bo\ A~70 Q1 o cT ~:r I A~p o Q.40. 0]. However we are only interested in those trajectories of system (32) fulfilling (33a). (0 There exist D E I~nxn. 79. ze def = Aro70 + 70Ao + Q1 + Q~ 70A~ 70A~ CroX:r . C1.
t r l. and hence contains only independent components.hi) xT (t . Q1 : Q1T :> 0. Suppose (i) holds. Remark that.2h2) kV(t. the 5th component of X ( t ) and the 1st component of X ( t . Q2. when (33a) is fulfilled.oo Suppose (i) holds.4hi) x T ( t .h2) x T ( t .Y(t) )T) T.. In view of the consideration preceding the statement.hi . Y(t)) ] is equal to 3"(xT (t) xT (t .' h ) ~ 0 and 2 1 + K H ( s ) + [KH(s)]" > eI .2hl) xT (t 3hl) x T ( t .Vs E C with Res > 0. Q2 : QT ~> 0. Suppose (ii) holds.Robust Absolute Stability of Delay Systems 223 50 The following property holds. 2.hi .hi.3h2) xT(t . lim x(t) = 0 for any global solution x of (9). for any r fulfilling (1) and any h E [0. Problem (37) is not convex in the unknowns D.hi . so one of the two occurrences may be omitted.A~e .h2) xT (t . The matrix 3" in (37) is related to the fact that we are concerned only by the trajectories of (32) which fulfill (33a).2hl). Q1. Then (ii) holds.Ao .2h )] Or(t. Matrix 3"TR3" in (37) is of size 14n + 5q. Then. a condition clearly stronger. in order to get (ii). Use of LyapunovKrasovskii functionals (35) instead of (36) would lead to . Se > 0.Vh E [0. More precisely. Q~.'P.. 7"~ <~ 0 instead of (37). d e t ( s I .3hl .2h2) xT(t . This is expressed in the fact that the vector x(t) x(t .I:) : . Q1.2hl . contrary to what happened in Sect. but is a LMI in 7). to prove (30). where H is given by (10).h2) x T ( t .2h2) x T ( t .RT ~> 0. once D has been chosen.2hl .h~) xT (t . it suffices. no LyapunovKrasovskii functional is available here for the initial system (9). Define (see above) v = (1 zl z~ zlz2 z~ z~ z~z2 z~z2 z 4 zg zlz~ z~ zlz g 2 2~ zlz2) T  Onehas [ Isn IZlZ2In . Proof.h].3h~) xT (t .2h )] X ( t .2hl) are equal to x(t .
In [ IzXZ~Inl = I zdn I Izl~2InJ \ z~in ] [(. one may show [3].z 2 .224 PierreAlexandre Bliman Right. ] . J I z. ]z~l ~ 1.A T ~...z I .4~)] I zJn I .]TTEff by jugate hence leads to (v) Osqxn E ]I~(14n+Sq)• and its con o > I z~io I { zgIsn J ~ Izlz2In] zi In ZlIn [ Z~Isn I.and leftmultiplying . (39) \ z~.A 2 ) P [. ] = I zJ~ I ((A0 qD) b zl(A1 (3~ I .! z~Is.I z~In ] zl In > [ z2In [ 2*~T 2* T lzlz~Inl \ z~In ] zl In +p(Ao + z~A~ + z~. \ z~in ] when Izll.qtI [. D)  z~hAoD  ziz2hAiD) .A 0 4 n! +(1 Iz~14)Q1 + (1 Iz214)Q2] ( I z2~  zlz2I. 212I~. Now..~)(. that (A0 + z12A1 + z~A2) z~In ] zlz~In] z~r.
4~ pt3o\ 7 ATP QI o Ar~ ' o o Q2 o o) ~o p o o o o o o o co~T\ o cT~T/ o c~ ]~Co ~.and leftmultiply o~T+T~../ \ zp:+.4o . Defining 8 . o > I zj+ I Iz~z~ I \ zplq ] ( /" / SSo \* /z~S~o/ lzgSBo Tr /z+S+o/ \ISq] (42) \ :~+..Cl 1~2 2Isq] .z~41 .i iz~z+~+ zpi.r.z+i.(Ao + D) .+.4~ ~.. + p~4o + Q1 + 0~ 7..Robust Absolute Stability of Delay Systems 225 Arguing as in the proof of Theorem 1.8(s. I z. To get the second part of (30). zl.zl(A1 .Jby .l Bo'~ z1Iq z2Iq Zl Z2 Iq and its conjugate.v ( s I .z~... ] \ zG ] (43) Decomposing 7r as the sum of ( and ~4~o . we then deduce the first part of (30).A2) 1 . right. ] where we have used the fact that t z:++. z2) def (st . I Bo = no I z+1<.D) + z~hAoD + Zl Z2hAi D ) . the result of this operation writes J I+~2~I+1 (40) (41) I~ z+..
2. I zlz2I. zl.z].. 3. z2) + [KH(s.42)2 1 Bo ]. and finally (ii). then z~S~o/ Isq ] SBo\" [ S .4 0 . \q] where H is given by (29). 2. More precisely (compare with (30)).226 PierreAlexandre Bliman onc shows. IzlzGI ] the expression in (42) is then proved to be larger or equal than Zl Z2 Z l Z2 (2I + KH(s.Z 2 C 2 ) ( 8 I .1. Using (39). and that r is decentralized and fulfills sector condition (5) for a certain nonnegative diagonal matrix K. (43) and (Co + z~C.2. The asymptotic stability is then proved as in Theorem 1.z2hCoD . z2)]*) .2 DelayDependent Popov Criterion As in Sect. 2.z ~ . I z~I. we shall use the fact that the following condition is sufficient for absolute stability of (18). for any h e [0. zl. Iz~l ~ 1. to system (18) with timeinvariant decentralized nonlinearity. A 1 _ z2. + z~C2) z~I.hi: . o \ iz~SBo/ z~SBo/ ~ Iz~S~oi \ Isq ] > (2• + ~(Co + =~c.q. + z]C~)(~ . see [3] for details. as in Sect.zL4~ ...) . We apply now the same ideas than those exposed in Sect. as in Theorem 1. we assume throughout this section that p .4o . i f R e s _> 0 and Iz~l. [] 3.4~)~o {[]~(C0 31. ] I (44) = IzJ~ I \zG (Co + zlC1 ... One then deduces that (30) holds.Z12Cl "~. that.ZlZ2hC1D) .
~5").Robust Absolute Stability of Delay Systems 227 3D. Jo (47) defined for any 9 E e([4h.)/ + 2E i=1 i./~(T. zl. one may use more information on ~.2h. as we are only interested .2h.2h~) . (46) with the same notations than for the system transformation conducted in Sect.zl(A1 . However.z2)]* > e I .2hl) + A 2 X ( t . as in Sect.1. The same considerations lead to study the timeinvariant analogue of (32) X = A o X ( t ) + A I X ( t .D) + z~hAoD+ ZlZ2"hA1D) ~ 0 and 2 I + ( I + y s ) K H ( s . I~1. 3~ > 0.z~) + [ ( I + ~ s ) K H ( s . zl. h del max{hl.)] / r dv o +(~) ~T sp / +(T) dv 2h.2h2). 2. ~ Iz21 _ 1.V(s.)/ e.(Ao + D) . one is led to choose for system (46) LyapunovKrasovskii functional of the form [ ~(o) ~T [ r V(V) d~f /~(_2h~) / p ~(_2h~) / \~(2h2)1 \r +f + f_ 0 / ~(T) ~ [ ~(T) /o(T.3~7 >_ O.z2) E CL with Res _> O. zl. and here !P(Y) /r /r d__ef /r /r \r / .h~}. ~/i/C. 3. Coherently with the choice made in Sect.0]. (45) where H is given by (29).2 (see formula (20)).(Co@(O)+C~@(_2hl)+C~@(_2h2))~ k~i(u) d u . d e t ( s / .2hl) + C2X(t . 3.Bo!lt(Y(t)) .2h.1. Y = CoX(t) + C1X(t .)/ e~/o(r.
A2.7 are defined after the current statement. T h e o r e m 5 Consider the two following statements.2hl) r r r r .hi .hi . o p(I3  ~o)'~ o (la  BTo )p 0 0 0 0 0 0 ( . IC are defined in formulas (34a) to (34g). Qt.13o. QI= QT>0. Q2=QT >0.IT p T ""tO "0 (/3 oAT)P o 0 0 0 0 Q2 0 Oo ) (. .228 PierreAlexandre Bliman in those trajectories of (46) constrained by x ( t .~ C[ ~o C~ where where the void matrix stands for the previous one.h2) x(t . Q~ E ]R15nxthn such that I/>0. where T~ is equal to (49) ((I3  Aro )p + p(I3  Ao) + Q~ + 0.7)=pT>0..41.h2) x(t .2hl)) X(t) = .2 "p(I3  A~) p(z3  (I3  Q. (i) There exist D E ]Rnxn.4~o T C T T Ao C2 T A tT Co ATcr + O(45n+lhp)• T A tT C2 T T A2 Co T T A2 C~ r/K: 0(4sn+lhp)• + . Co. C2. P. . 7/= diag{r/i} E ]R~px~p.7 < 0 .4o. .h2) ) .hi) x(t . and .TT jTT~J.h2)) . yields the following result. ~V(Y(t)) = (48) for certain functions x. .. y. Computing the derivative of V along the trajectories of (46) for which (48) holds.hi)) . and J.~ T l (/3  CoT/C)~ (I3 o cUc) (6 o c~rJc) 2115v + (Ia  ICCo) (I3  ICC1) (Is  lee2) . dr.
4h2) e(Y(t)) gt(Y(t .2h~)) . Then.A l e .. Then (ii) holds. h].2h~))[ ~ ( Y ( t .2h2) X ( t .2hl) X ( t .Vs E C with Res > 0. as in (21). where H is given by (10).Robust Absolute Stability of Delay Systems (ii) The following property holds.2h~))/ ~ ( Y ( t .~ 0 0 /sn j d=efdiag. in such a way that. det(sI . for any decentralized r fulfilling (5) and any h E [0. 229 3~ > 0..+ + oo Matrix J is introduced in (49) in order. X(t) def [ X ( t .2hl)) (5o) ~(Y(t . Suppose (i) holds.Vh E [0. denoting ( x(t) . Ilsp E ]~(45n.F15p)• .2hi .2h2) X (t . 9 t ..Ao . Suppose (ii) holds.~ 0 0 0 Is. lim x(t) = 0 for any global solution x of (18). to obtain a nice form for 7~.' h ) 7s 0 and 21+ ( I + q s ) g g ( s ) + [(I + qs)KH(s)]* > ~I .0 Olsn 0 0 0 0 0 Isn 0 0 0 0 0 0 Isn 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 Is.4hl) X ( t . .h].2h~)  =J ~(Y(t)) l X ( t .2h~)/ one has x(t) x(t) x ( t .2h1) ] X(t .3q = diag{qi} E ]~vxv. One has ( Isn 0 0 0 0I~0 0 0 0/5. q > 0.2 h l ) ] \ x ( t .
.4hi) X ( t . 0.. using relation (33a) and removing the repeated occurrences. then (49) is fulfilled for 7/ = 0 and /3  P .e.4h~) as a function of x(t) and its delayed values. 0. (0. X and ~ being related to the components of x and r by identities (48)..2h~) = x(t . Matrix ....) x ( t . Q1. o. One hence removes one occurrence of each of these four components.2h~) X ( t . (0.2hl) X ( t .• v = z~5.2h2). timeinvariant nonlinearities.. (51) . Q~ E ]Rs"xsn. 0.) x ( t . 0. for systems with decentralized.2hl))..) e ( v ( t ) ) = (Ip o~ o~ op op) e ( v ( t . 0. to prove (45)..4h~).2h~) = (i. 0. By definition of v.) x ( t . / /zl~. o~ o~ I. in order to get (ii).2h~ . to a stronger stability result than the one provided by Theorem 4... Proof.) x ( t .2h~) x ( t . o..1. o. as in Sect. ] T j T T ~ J J in (49) is 27n + 14p. 0. Suppose (i) holds.. o.../3  Q2: the addition of a Lur'e term leads.. i. I.2hl . which precisely expresses how the vector in the righthand side of (50) is expressed as a function of x(t) and r and their delayed values.7 is an element of ]R(3~215 Finally. 0./ 015pxhn] z4lhn ] ] z2In ] . when (37) is fulfilled for P. This permits. i.2h~) = ~(t ..2hi .) x ( o = (~. o.2h~)) = r . 0.. o .) x ( t . 0. I.2h~) = ( i . one has (compare with (38)) I~n / expressing Y o1~. 0. Similarly to what happens with the results stated in Sect. (o. o. expressed twice. 3. to construct easily the (huge) matrix J .. In view of the consideration preceding the statement..2h~). one has (0. it suffices. 2..4h~) = ~(t . the "delaydependent Popov criterion'in Theorem 5 is stronger than the "delaydependent circle criterion"in Theorem 4. Analogously to the proof of Theorem 4. 0.~. 0. one defines v E ]R~Tn• the vector x ( t . o ..230 PierreAlexandre Bliman Now. 0. / 3  Q1. the size of .
~ I \ 4~**.h A ~ D ) \ z~. / zi  ] z2I. / v Right. A~ ))p ~ \ 4~.zlz~'hA1D). Using now (39). ! 9 (52) < 4~.~)) z~i~.•) v : 4~.D) .Robust Absolute Stability of Delay Systems SO 231 sj (o~.] [(I3  (A0 + z~'. [ 41~.and leftmultiplyingjTjTT~JJ by ( 014p• )anditsconjugatehence shows that the following expression is negative (compare with (24)) 2 T  z~.//~• .~J. [ ( ( A o + D ) + z l ( A ~ .z ~ ' h d o D .4~ + z=.z2"hmoD .15o. / ..Iz214)e~] 1 4 h . \ 4:. one gets that (~  (Ao + z~A~+ d.~  z~i. / ! z~I.z2I.40+ z~A~+ z~A2)) +(1 Iz~14)el + (1 .D ) .l \ 4~./ /z. ]/4/5.I = z~  [ z~I. / +P(13  (. / /I5n~ Zl[n =  z ~ / s n  / z~In J ((A0 + D) + zl(A1 .z l z ~ . [ = (13 \2/ lz. / . I \4I~.
232 PierreAlexandre Bliman One hence deduces. i / z~o / zg~15.z~hAoD . that.2h~))] \g.z~A~ . I z~ I  \z~l iz~z~I o~ \ z~1~ l . (39) as in the previous page and (43) yields J fl O14p• ( v ) (sx(Ao + D ) .zl(A1 .z l ( A 1 . i.~o ..1 I z~. I \z~/ I~I (IZl2'~.z~.z~z~hAiD have { ~(v(t)) negative real part: this is the first part of (45).(Y(t .D) + z2~AoD + zlz2~:~D)~Bo = i / z1115.D) .~~215Iz~l I z~ I k ~.l (sI .e. then the eigenvalues of (Ao + D) + z~(A~ . \zg~. I ~ I ?\ ~' ] = I z ~ I I z~ I 9 (o~. as in the proof of Theorem 2. if Izll.D ) + z2hAoD § zlz2hAiD)l Bo'~ V/ ) " Using (52). (40)) (Ao + D ) . Ij / ~ 5 .x/~ /0~0~215 \ ~ z~i. Expressing [ ~ ( Y ( t .D) + z2hAoD + zlz2hA1D)l Bo = i z~I~. Iz~l <_ 1. I !~ ~/ ((sIAo z~ ~g~)~8o) Iz~5.~)~o) )[15pI Zl 1 ~ IPIp I~.v ( s I W ~" k (o.2h~))/ as a function of r and its delayed values. one defines now v' ~ ~14p• in the same way than v in (51).] izlz~.(Ao + D) .z~(A1 .~ } \z~15~l iz~z~1~i \ z~ I by (compare with (25).l = \z~r~l  ((s~ . I z~z~i~I I \ z~ l This implies I ~ iz~i~i \ z ~ I \ z~ l Define w E R(27n+14p)• def ( .
/ /zbr~p/ [(~ izlz~l +(1  ("/" { I~p \* zlI. i z2Iv tzlz .] by w and its conjugate proves that the following expression is negative (compare with the analogous computation in the proof of Theorem 2) / z2J.. no S 9 (Co + zl2..7~ ul (i. / ZIIPIP / \ z~i.T + z]'Cy)). / When Res < 0 and [zl[. / I~ Now. I zlZp I =2:~ I \~L ~.TJC(Co + z~C~ + ~gC2)S~o _2*~T T ~ + s ..and leftmultiplication of . fz2[ _< 1. \ =~:. / [1(/112 !2 z2/ (215v +/C(Co + z~Cl+ 4c2)s~o +~ToS'(C[ + "1 '~1 + zg'C~)~) + s.]TJT~J. \z~:~] Izll2)Q1 + (1  Iz~12)Q~](I~ St3o) + (Olovxsv) (Co C1C2) (I3  sSBo) \( I z2:~ I . / .zi. this yields the negativity of i zl~.Robust Absolute Stability of Delay Systems This shows that Jffw is equal to 233  ~ ((8'r Z~r Iz21iSp~( z~Z.j \ z~i. defining ..q(s. right.z2) as in (41).
. 2I + (I + l fls) KH(s. diagonal) matrix ^def 7= Ip zs Is 0 Is Is ts lp . Zl. Theorems 1 and 2) may be found taking D = 0. This provides the second part of (45). zl. Now. z2) 9 C : Izl I. zl. from the transformation in the beginning of Sect. Izxl = Iz21 = 1. V(s. zx.Z.3 Remarks on Theorems 4 and 5 Remark that the delayfree case (h = 0). one then deduces that /l/ll z + (2I + KH(s. [] 3. The asymptotic stability is then proved as in Theorem 1. 3. This achieves the proof of the implication (i) =~ (ii). as in the proof of Theorem 2.z2)]* > eI . (43) and (44). 1. it is straightforward. the previous inequality implies. zl.z2) + [(I + l ils) KH(s. I sKH(s.z2) 9 C 3 with Res < 0.1. Denoting ~ 9 IRP• the (nonnegative.234 PierreAlexandre Bliman Using (39). to prove that (45) indeed implies (ii). may be found taking h = 0 and D = A1.. the existence of e > 0 such that (compare with (26)). described in Sect. z2) + [KH(s.rs/ where H is given by (29). z2) k z~Is / ~ Is / //zl~Is/ L\ Z~is / /zlz~Zs/ \ z~. Iz21 < 1} by "q maximum modulus principle.1. This property is then extended to the set {(Zl. zl. Also the delayindependent case (h = +oo. z2)]*) ( Ip z2zs I 'J I z.
' h 2 + AoD~ 1 .(A1 .r as is made in Popov theory to get nonpositive values of r/ [1]. One writes here that ( s I .Ao .( A ~ .A l e 'h' = sI .e shl 8 + A1Dle sh' 1 .h + (h .s h ~ S S +A2Dle_sh ~ 1 . However.e 'h2 $ ) ( A o + D1 {. wrt the nonlinearity r verifying sector condition (1). Gantmacher F. at the cost of still more cumbersome formulas. Also. At last. 0]).D ~ ) .Robust Absolute Stability of Delay Systems 235 In order to treat the cases where the uncertainty interval on the delay is of the type [h. As an example. References 1. h .e sh~ S +AoD1 1 . (1964) Absolute stability of regttlator systems.g. and one may change r into K y . Last. hi with h > 0.R. local stability results may be easily obtained when the sector condition holds only locally. One may e. One may remark that the size of the obtained LMIs.e 8hl S . 3 to systems with two independent delays hi. one may decompose the delay as h . . 3. may be easily generalized to state analogue results for systems with multiple delays. Of course. see [7].e ah~ 8 + A ~ D 2 e _ s h 2 1 . prove that the stated convergence properties are indeed uniform wrt the initial condition r in bounded subset of C([h. and use the previous results with an uncertainty interval [0.A~e'h~)(I .n l i . remark that some generalizations or refinements of the given results may be easily carried out..h_]. one must not forget that the numerical verification of the properties we are interested in.D 1 ) e .h).e . HoldenDay Inc.1.e sh~ 8 D2 1 . which permitted to consider delayindependent stability of a 2delays systems. the techniques developed here to obtain results for systems with a unique delay. although polynomial wrt the sizes of the matrices of the problem. One may then approximate by a rational transfer the terms e sa. the case where the output of the nonlinearities is delayed too. the key manipulation is to adapt the transformation given in the beginning of Sect. h~. and wrt the initial time. even for low dimension systems. is in general NPhard. and L e m m a 3 permits to consider the delayindependent stability of a 4delays auxiliary system. and for which we have obtained here sufficient conditions.D 2 ) e .A.eSh2 + A 1 D 2 e sh~ 1 . Aizerman M.' h ~ . may be considered as well. is already large. to transpose the delaydependent results obtained in Sect.
Anal. Proc.. 1999) Stability of nonlinear delay systems: delayindependent small gain theorem and frequency domain interpretation of the LyapunovKrasovskii method 3. Dugard L. Kharitonov V. Niculescu S. IEEE/IMACS CESA'96. J. (1977) Absolute stability of automatic control systems with lag. Franklin Institute 318 no 3:143150 11.M. 228. Math. Sci. Part II 4 3 : 1 6 0 1 . (submitted.M. Karlsruhe.I.236 PierreAlexandre Bliman 2. London: 171 18.. of IFAC Workshop on Linear Time Delay Systems.J. Dion J. IFAC Syst. 14. 2 8 no 1 : 7 .W.. Popov V. (1963) Stability of motion. Stanford University Press 15.A. SIAM Studies in Applied Mathematics vol. (1985) Stability in linear delay equations. Bliman P. Automat. Lille. 2000) LyapunovKrasovskii method and strong delayindependent stability of linear delay systems. Hertz D. Bliman P.l.I.S. Bell Sys. Springer. Syst. (submitted. Differential Equations 13 no 8 (1978): 954960 9.1 0 21. (1994) Linear matrix inequalities in system and control theory. J. Chen J.. Sandberg l.. (1998) Stability and robust stability of timedelay systems: a guided tour.. Control 40 no 9:16401645 8.. (to appear.M.I. (submitted. to appear in Proc.L.P. (1962) On the stability of nonlinear automatic control systems with lagging argument. 15 7. 2000) A note on frequency domain interpretation of LyapunovKrasovskii method in control of linear delay systems 6. Applics 115:533555 10.. Halanay A.F.F.. (1963) On the absolute stability of systems with time lags.A. Niculescu S.1 6 0 8 . 28 no 2:248249 13. (1964) A frequency domain condition for stability of feedback systems containing a single timevarying nonlinear element. Chinese Math..I.S. Contr. Automat. (1983) Correction to "Linear systems with commensurate time delays: stability and stabilization independent of delay". Lecture Notes in Control and Inform. Jury E. Dion J.I. Infante E. Hale J. 27 no 2: 367375 12. Ancona..A. Dugard L..A. Italy 5.. IEEE Trans. Kamen E. Contr. (1996) On the KalmanYakubovichPopov lemma. Li H. (1982) Linear systems with commensurate time delays: stability and stabilization independent of delay. Verriest E. (1996) Asymptotic stability sets for linear systems with commensurable delays: a matrix pencil approach. Contr. of 5th European Control Conference.. Lett. 4:609626 16. Niculescu S.. Feron E. E1 Ghaoui L. Li X.A. Germany 19. J.. Balakristman V.N. Niculescu S.. IEEE Trans. (1998) Robust stability analysis of time delay systems: a survey. IEEE Trans. Kamen E. Tsen F. In: Stability and control of timedelay systems. Struct. Boyd S. (1999) On some frequency sweeping tests for delaydependent stability: a model transformation case study. Proc. Bliman P.W. Latchman H. Contr. Bliman P. Rantzer A. Automat. 2000) LyapunovKrasovskii functionals and frequency domain: delayindependent absolute stability criteria for delay systems 4. Gromova P. (1995) Frequency sweeping tests for stability independent of delay. Krasovskii N. Pelevina A. France 17. Tech. Remote Control 23:783786 20. Automat. Zeheb E.W. (1984) Stability independent and dependent of delay for delay differential systems.K.
Aggoune W. Part I 11 no 2: 228238. Zames G. (1998) Stability of nonlinear differential delay systems. Mathematics and computers in Simulation 45:257267 23. IEEE Trans. Part II 11 no 3:465477 . Contr. (1966) On the inputoutput stability of nonlinear timevarying feedback systems.Robust Absolute Stability of Delay Systems 237 22..I. Verriest E. Automat.
On consid~re le probl~me du temps minimal avec la contrainte u~ + u~ = 1 (on ~< 1). Elles sont param~tr~es par les ~quations : cgHx (I= Op' ia c3Hx cgHx cgx' Ou . Elle est dite stvicte si la singularitd est de codimension un. (p. Consid~rons le systSme : dq d't = ul Fl (q) + u2F2(q) (1) oh le contr61e u = (Ul. ulFl(q)+u~F~(q) > +A(u~ + u ~ ) .p. et conduit darts le cas Martinet h une stratification de la singularit~ en secteurs Lagrangiens. 1 Introduction Soit M une varidt~ analytique r~elle et F1. ubourgogne. us) est une application L ~176 valeurs dans un d o m a i n e b.Stratification d u S e c t e u r A n o r m a l d a n s la S p h e r e de M a r t i n e t de P e t i t R a y o n Bernard B o n n a r d 1. f r A b s t r a c t . C'est Squivalent au probl~me de minimiser la longueur d ' u n e trajectoire de (1) pour la m~trique : l(q) = (u~ + u~)89 (probl~me sous Riemannien). L'objectif de cet article est de fournir le cadre g~'om~trique pour faire une analyse de la singularit$ de l'application exponentielle le long d'une direction anormale en g~om~trie sou~Riemannienne. I1 utilise les calculs de [9]. UFR Sciences et Techniques BP 47870 21078 Dijon Cedex. On consid~re le syst~me augmentg : gl = ulFl(q) + u~F~(q). F~ deux champs de vecteurs analytiques. A = O. France { t r e l a t .~1 et les anovmales pour . [12]. Emmanuel Tt~lat 1 Universitfi de Bourgogne Dpt de MathSmatiques. Ces derni~res se projettent sur les trajectoires singuli~res du syst~me .0 oh Hx(q.8 9 On appelle extr6male une solution de ces ~quations et ggod~sique la projection d ' u n e extr~male sur M .0) et )~ est une constante normalis~e h 0 ou . q0 = u~ + u~ et on appele contr61e extrgmal une singularit~ de l'application extr~mit~ (T fix~) associ~e. bonnard}@topolog.u) : < p.A) ~ (0. On a deux types d'extr~males : les extr~males nor'males pour A = . C.
. On note 27r le secteur de U couvert par les extr~males anormales de longueur ~< r issues de {0} • V.(g~) . Sous les hypotheses ( H 1 ) . . . P 2 } = 0. P 2 } .3} Sous ces hypoth6ses le vecteur adjoint pT(0) est unique ~ un scalaire pr~s.Ur f o r m e un secteur de la boule sousR i e m a n n i e n n e homgomorphe d C t. Identifions M localement/L un voisinage U de 0 dans ~ n . r) les points ~ distance r > 0 de q(0) = 0 pour la m6trique sousRiemannienne.C ) . et B(0. n . et : u i { { P i . P2} = 0. t El .(H2) . {P1.F~l. et sont contenues dans : Pi = P2 = {P1./. od C est un c6ne positif de dimension n .F2}lT(t).'2) : (H1) K ( t ) = Vect {adkFi.3 pour n >t 4 et 1 pour n = 3.analytique relativement au riot optimal. P 2 } . Tr61at (1). La premiere ~tape de notre construction est de d6eomposer l'espace en une partie lisse et analytique et une partie non sous.F2t~ / k = 0 . r / k E iN} (premier c6ne de Pontriaguine) est de codimension un pour t E [0. On note S(0.J ( . E. . L e m m e 1 P o u r r assez petit. . T[ une trajectoire de r6fdrenee. (g2) {P~. Soit t ~~ 7(t).T. Sa trace sur la sphere est f o r m g e de deux surfaces analytiques de dimension n .(//3). pour tout t E [0. il existe r > 0 tel qu'une trajectoire anormale de longueur infgrieure d r soit stricte et globalement optimale. Bonnard. On peut supposer qu'elle est associde au contr61e u2 = 0. T ] e t est engendrd par les n . . La construction est visualisde sur la Fig. On fait alors l'hypoth~se que les conditions suivantes sont v6rifides le long de 7 pour le couple (Fi. r) la boule sousRiemannienne de rayon r.4 et rgduite d deux points pour n = 3.1 premiers ~ldments { F ~ . . N o t a t i o n .4 pour n >>. On identifie la condition initiale q(0) = q0 g 0. . On peut choisir V assez petit de sorte que toutes les extr6males anormales issues de {0} x V v~rifient les hypotheses (H1) . 7"] on suppose que : Flt~ ~ {~dkFl. P2}} r 0 le long de 3'(H3) S i n / > 3. P i } + u 2 { { P 1 . et soit V un voisinage de pT(0) dans l'espace projectif P ( T ~ U ) .1. adr'2F1. 2 La partie lisse et analytique (2) Les trajectoires anormales sont solutions des dquations : dq _ OH____a dp dt O p ' dt Oga Oq oh Ha = uxPi + u2P2.( H 3 ) . On a : P r o p o s i t i o n 1.240 B.
22r 3 Le l e m m e d e r e c o l l e m e n t P r o p o s i t i o n 2. q F3 = o . D = Ker w. C o r o l l a i r e 2 Avec les notations de la proposition prgcddente. w = dz .0 telle que : (0 ~(~) c s ( o .zva~TP1 + c2c~P2)' /5~ = . a(O) = A. Pour i = 1. z) et g = a(q)dx 2 + c(q)dy 2. y. on peut supposer le probl~me isopgrimdtrique. O n n o t e : G l = ~ + ~ . V En paramdtrant par la longueur Hn = 1(P12 + P~) = 1 et en introduisant les coordonn&s cylindriques : P1 = cos 0. g) oh 2 U est un ouvert de ~ 3 contenant 0.v/~v/~ ( u P 3 . 0. ) k s pou~ ~ # o (i 0 a(e) n L(O) r I~ od L est le lieu de coupure l'espace tangent ~ la sphere gvalug en a(e) tend vers K ( r ) lorsque e r O. r assez petit. F2 engendrent D. Fi >. c'esthdire que la mdtrique g ne ddpend pas de z. Dans les coordonn6es (q. a. de longueur r. 1. . 3 on pose : Pi = < P. On consid~re la structure sousRiemannienne (U.~2 az. P3 = A. a(e) n'est pas l'image d'une pattie compacte du cylindre P~(O) + P~(O) = 89 L 'application exponentielle n'est doric pas propre le long de l'anormale.2. on obtient les . D.Stratification du Secteur Anormal 241 py PtT~ u) Fig. q = (x. c & a n t des germes de fonctions analytiques en 0telsquea(0)=c(0)= 1. Une analyse plus fine utilise les estimations calcul&s dans le cas Martinet qui est le suivant. [28]) Soit K ( r ) le cdne de Pontriaguine dvalug d l'extrgmitd A de ia trajectoire anormale de rdfgrence. P) les extr~males normales sont solutions des dquations : P1 P2 Y2 P1 2 tSa = 0 P~ ay t 5 1 . Sans nuire h la gdndralitd de notre analyse. o~ F1. La ligne anormale L issue de 0 et paramdtrde par la longueur est L = (+t. (72 = o'ffv" Le repbre o orthonorm~ est : F1 = ~ a ' q F2 = ~ . P2 = sin 0. Les gdoddsiques anormales se projettent dans le plan y = 0 et sont des droites z = z0.~ d x . Alors K ( r ) est un sousespace vectoriel de codimension un et pour toute courbe lisse e ~~ a(e)./ 5 1 . 0). e >>. (voir [12].
Par ailleurs t o u s l e s crochets de Lie de longueur ~> 4 sont nuls. En effet dans les 6quations pr6c6dentes. en param6trant les solutions par vV~. F2] = o + Yb'~. on consid~re le syst~me de Ft 4 : f l . oh ' d6signe la d6riv6e par rapport k v.242 B. F~].IF1. Cela revient ~ 6tudier la sphere sousRiemannienne dans un C~ de la trajectoire de r6f~rence.~x + y~z + ~ O._ y2 p1 2 v~' 0 1  av V~/~ (Ay~~P' +  . Tr61at 6quations suivantes : =~' P~ Y= v~' P~ ~. On pose : .a c o s 8 + / ~ s i n O ) . bien que dans le cas plat la direction anormale ne soit pas stricte. La forme normale gradu6e est : 9 ordre 1 : g = dx ~ + dv 2 (cas plat) 9 ordre0:g=adx ~+cdy 2 aveca=(l+ay) 2etc=(l+fix+Ty) 2 et ~ l'ordre 0 les 6quations pr6c~dentes s'6crivent : cos 8 x= V= sin 0 S y~ cos 8 2 va ' 0 1 (Ay . et le syst~me 6quivaut k : O" + AsinO + a2sinOcosO . F1] = ~ ~ et [[F1. ~ 0 0 y~ 0 F~ = Oy O On a l e s r e l a t i o n s : F3 .. ~ c P~) Les variables x. F1] = 0. Bonnard. z sont gradu6es en fonction de la r~gle de [7] avec les poids suivants : le poids de x. y. Dans la section suivante on relive le probl~me sur le groupe d'Engel pour avoir une repr6sentation uniforme. 4 Le cas Martinet relev6 sur le groupe d'Engel Si q = (x./~.( A y .fl sin 0.7 de la m6trique en petits param~tres. y est un et le poids de z e s t trois. F2]. on transforme les param~tres ~. z.a/3sin ~ O + / 3 0 ' c o s O = 0 (3) La condition initiale q(0) = 0 induit la contrainte y .d r ' les 6quations se projettent d dans le plan (y.a cos 8 + 13sin 9) En introduisant le param6trage v ~ V ~ d . 0' = . E. Pour ces calculs.O) en : yt = vf~sinO.0 qui se traduit par la condition : 9' = a cos 8 . F4 = [IF1. y. le cas g6n6ral est interpr6t6 c o m m e une ddformation du cas plat. w).
Pw) on o b t i e n t : y2 PJ: + PzY + P w . 0. Les e x t r ~ m a l e s P~ = p~ = {p~. 0). ( + t . P o u r t o u t e m~trique s o u s . T o u t e m~trique s o u s . P2 sin0. L2 = 243 et on d~finit la representation : p(F1) = LI. u~{{P~. /53=P~P4./ ~ 0 = t'2t:)4. 4 et Hn = ~ l i p 21 ~_ ~ ) . 0. y est un. o n o b t i e n t : TP i~I=P~P3. Fi(q) > . Elle n'est . oh 3=0. et le poids de w e s t trois.p~} = o En p o s a n t p = (Px.L1 = (00)01 (11~176176 Stratification du Secteur Anormal 00 00 .R i e m a n n i e n n e sur Ge. Param~trage des g~od~slques dans a n o r m a l e s non triviales sont solutions de : le cas plat. Le flot a n o r m a l est donnd p a r : i:=ul. 3. on pose Pi = < p.P~}. /5~=PtP3.~ = py = pz + YPw = PwU2 = 0 Cela implique p~ ~ 0 et done u~ = 0. /5.p~. En n o t a n t P4 = A. i = 1.P~}. et p o u r 0 ~ kTr il vient : 0 = . l ' a p p r o x i m a t i o n d ' o r d r e 1 est la m~trique plate g = dx ~ + dy 2.P~}+ u2{{P~. le p o i d s de z deux. Pour calculer les extr~males normales.2.4 au syst~me invariant d gauche R = (ul L1 + u ~ L 2 ) R sur le g r o u p e d ' E n g e l G~.Py.=0 En p a r a m ~ t r a n t p a r la longueur Hn = 89 on p e u t poser : P1 = cos O. p(F2) = L~ qui p e r m e t d ' i d e n t i f i e r le syst~me sur lq. repr~sent~ ici p a r les m a t r i c e s n i l p o t e n t e s : i q2q3q~/ 1 qx 0 0 Le poids de x.P~} = o . cela ~quivaut h l'~quation du p e n d u l e : + Asin0 = 0 On d~signe p a r L la ligne a n o r m a l e issue de 0 : t ~ pas stricte et se p r o j e t t e en # = k~r. ~=uly.1 si on p a r a m b t r e p a r la longueur. y2 ~b=ul~ luxl .R i e m a n n i e n n e de M a r t i n e t s'~crit g = adx 2 + cdy ~ et se relive sur le g r o u p e Ge.
Trdlat Pour obtenir une representation uniforme des gdod6siques normales.C P I .c ) . w se calculent par quadratures. la fonction P ( u ) dtant ddveloppable en 0 selon : 1 1 V(u)=~+~6q~ ~ 1 ~ +~q3u 4 +. On peut ensuite calculer Pu et P3 en utilisant les int6grales premieres.Px . Le cas Engel contient le cas de contact plat et le cas Martinet plat restitu6s par les op6rations suivantes : 9 En posant p~ = 0. E. on obtient les gdoddsiques du cas Martinet plat. L e m m e 3 On obtient le cas Martinet plat (resp.sgroup de G e : {e p t[Ft. En effet la condition Pz = 0 (resp. y. (Ow)). qui 6quivaut pour P1 # 0 et A # 0 ~ l'~quation : 1512 = A(Pla + CPl2 . on obtient les g6oddsiques du cas Heisenberg. et x. Une autre faqon de d6duire le cas Martinet plat est d'utiliser le r6sultat gdn6ral de [7] : L e m m e 4 Le cas Martinet plat est isomdtrique ~ (Ge/H.g 3 . pro = 0) correspond alors h la condition de transversalitd.3APt2 + A. Eclatement et Martinet en droites On observe que les deux champs o et ~ c o m m u t e n t avec F1 et F2. En effet le syst~me admet trois intdgrales premi6res : P~ + P~ = 1. z. En posant g2 = 4(e2e3 + ezel + exe~) et ga = 4ete2ea on peut l'dcrire : :P'(u) = 4 P a ( u ) . on utilise la fonction elliptique de Weierstrass :P. Soit P(u) la fonction elliptique de Weierstrass (cf [20]) solution de : p'(u) =  off les complexes ei vdrifient el + e2 + e3 = 0.. dx 2 + dy ~) oft H est le so.244 B. En utilisant P1 = P2P3 on obtient : PI = . 9 En posant pw = 0. L'interpr6tation est la suivante. Bonnard. / t . On peut retrouver les solutions oscillantes et en rotation du pendule en utilisant les fonctions elliptiques de Jacobi donndes par : cnu= 5 \:P(u) e2 ' \P(u) plat e3] d6duits du cas Les cas Heisenberg Engel. Heisenberg) en minimisant la distance sousRiemannienne par rapport h la droite (Oz) (resp. la distance sousRiemannienne d'Engel hdrite donc au moins de toutes les singularitds du cas Heisenberg et du cas Martinet plat. La solution P1 peut donc s'exprimer sous la forme a P ( u ) + b . et deux Casimir : 2P1P4 + P~ = C et P4 = A.g 2 P ( u ) . I1 est int6ressant de noter que c o m m e la distance sousRiemannienne par rapport ~ une droite est plus rdgulibre que par rapport un point.
1)(E(u) . E ( K ) sont les intdgrules completes.oJ~ .z ) . Chaque extr~male normale est minimisante jusquu son premier retour en y = O.t: + 0..k s + O. L e eas g ~ n ~ r i q u e .Stratification du Secteur Anormal 245 6 Stratification dans le cas Martinet On rappelle le param~trage explicite des g~od~siques. K ( k ) ~ ~. et lorsque k '~ = 1 .y . y . z) et S~ : (x. Le premier temps con3ugud vdrifie 2 K < tlcv/A < 3 K .E ( K ) ) + k':tv/~ + 2 ~ s n ~ en u an off sn u. . . Proposition 3.u=K+tv~ ~(t) = . c'esthdire pour tVrA <~ 2 K . on a : 1. ef [4]. y(t) = . 7 dont le rble gEom~trique est le suivant : L e m m e 5 Pour la f o r m e normale graduge d'ordre 0 : g = (1 + ay)~dx 2 + (1 + fl~c + Ty)2 dy ~. A=0 x(t) = t sin r y(t) = t cos r z(t) = t_~ sin D r co~~ r o~ r ~]  ~. cn u. La trajectoire anormale est stricte si et seulement si a ~ O. Pour Etudier la situation g6n~rale on utilise la forme graduEe d'ordre 0 de la section 2. . et plus pr~cis~ment une simulation numgrique montre que t. ~] Z et les courbes dgduites des prgcgdentes en utilisant les symdtries $1 : (x.~ r In ~r. Dans ce cas dEgEn~rE.fl. Lorsque . l'application exponentielle applique tout le bord du cylindre sur la direction anormale. z) ~ ( . nEanmoins l'application exponentielle n'est pas propre.97. Elle depend de trois param~tres a. . La sphere est sousanalytique en B mais pas en A. y. les points de coupure et les points conjuguEs s'accumulent le long de l'anormale. y.E ( K ) ) . z) ~ ( x .. elle forme l'adhErence du lieu de partage. et K.x .~ 0. Le module k des fonctions de Jacobi est donne par : k s = 1P1(0) . Lorsque A + ~ .~ c n u ) 3A~ (2k ~ . La trace de la sphere avec le plan de Martinet et pour z ~> 0 est reprEsentEe Fig. 3K C o n s e q u e n c e s . K ( k ) . Les ggoddsiques normales issues de 0 poramdtrges par la longueur sont donndes par : 9 A~O. dn u et E(u) sont les fonctions elliptiques de Jacobi. Le corollaire 2 ne s'applique pas car la gEod6sique anormale n'est pas stricte.t + ~(t)= 2( (E(u) .2 .
c2. Elle est repr~sent~e Fig. 2. Tr~lat ! Fig. Les premiers p o i n t s conjugu~s localis6s au voisinage de l ' a n o r m a l e sont avant le troisi~me retour d ' u n e g$odSsique sur le p l a n de Martinet. le flot ggod4sique est intggrable par quadratures en utilisant les fonctions elliptiques de Jacobi. E oxp Pl(o) A = (. si ~ ~ 0 ( c a s s t r i c t ) . *(r. c 3 c o r r e s p o n d e n t h de grands d@lacements et il y a deux courbes car il faut consid~rer les t r a j e c t o i r e s ..o) 8 ~:(~. E. o) 2.3. ca se ramifient sur la d i r e c t i o n a n o r m a l e et sont construites ainsi : la courbe c2 est associ~e h des petits dgplacements localis~s au voisinage du col . Le pendule (3) est conservatif si et seulement si jO = O. les courbes Cl.246 B. Dans ce cas. O) e~ La section L" correspond h la p r o j e c t i o n de y = 0 d a n s le p l a n de phase d u pendule. Les trois courbes cl. a u v o l s i n a g e d e r a n o r m a l e . exp Cl f Fig. 3. T r a c e d e l a s p h e r e a v e c l e p l a n y = 0. P a r ailleurs les calculs de [9] conduisent h la conjecture suivante : Conjecture. Bonnard. C'est en fait un r~sultat qui p e r m e t de compactifier n o t r e analyse car on ne considbre que des t r a j e c t o i r e s avec un n o m b r e uniformgment borng d'oscillations (comparer avec [1]).
2. voir Fig.Stratification du Secteur Anormal 247 oscillantes du pendule (courbe cl). dans le cas de la Fig.. Alors : 9 c~ : Z = ( ~ + O ( r ) ) X 9 c2 : g = . Posons Z = rx et X = ~~f. Dans le cas strict. r assez petit. Les contacts des courbes se calculent dans la cat6gorie polynomiale.et les trajectoires en rotation (courbe ca). y ) . voir Fig. ~ X ~ + o ( x ~) ~ + o ( X 2) On en d4duit : P r o p o s i t i o n 5. Les courbes ci sont tangentes h la direction anormale car le c6ne de Pontriaguine K ( r ) coincide dans la direction anormale avec le plan (x.4. P a r m i les deux branches cl et ca. c'est l'6quivalent de la relation p = nO en optique. et seule la branche cl existe. Dans le cas plat.5. une seule appartient h la sphere . Les deux situations sont identifi6es en termes de vecteurs adjoints. elle est formde de feuilles c2. y). 27 est la droite 0 = 0. 3. e <~ r qui se recollent le long de la direction anormale. r) de petit rayon a au voisinage de la direction anormale les proprigtgs suivantes : 1. e). 4. direction anormale Gxp c3 rgM Fig. elle est l'image par l'application exponentielle d'un secteur non compact du cylindre (X + oo). pour les d6tails voir [12]. el (= cl ou ca) assocides aux spheres S(O.3 c'est la courbe ca. Les estimations sont les suivantes : 2 P r o p o s i t i o n 4. 6valu6e en 0. Dans le cas conservatif les calculs utilisent la catdgorie logexp et la thgorie de l'glimination dans cette cat6gorie. Le calcul de la sphere et donc du bord du d o m a i n e est d61icat car il y a un ph6nom~ne de c o m p e n s a t i o n . la boule B(O. L a t r a n s c e n d a n c e d u s e c t e u r . elle est homdomorphe ~ un secteur conique centrd sur L et est en particuller simptement connexe. [28]. Les calculs de [12] montrent que le positionnement des courbes d6pend de la courbure de Gauss de la m6trique Riemannienne restreinte au plan ( x . C'est un secteur non sousanalytique oh tes calculs n~cessitent l'usage des fonctions hyperboliques.
On conjecture ~videmment que la non sousanalyticit~ de la sphere reste vraie en g~n~ral mais les calculs dans le cas non conservatif sont complexes et non standards . 5. 6. Tr~lat Fig. On appelle 2secteur de Martinet de la boule SR l'intersection de la boule B(0. Bonnard. mais on peut g~n~raliser ais~ment.6. d'o~a en particulier un ph~nom~ne de stabilitY.248 B. Dans le cas plat. qui s'explique h l'aide des invariants microlocaux du secteur. R e m a r q u e . Tout secteur de Martinet est diffgrentiablement reprgsentg par la Fig. . en particulier il faut utiliser des cartes c o m m e pour les calculs des cycles limites. Dans nos calculs on a privil@gi~ la coupe de B(0. Par contre pour calculer la sphere il faut tenir compte de l'interaction entre toutes les valeurs propres du spectre pour en d~duire une m o y e n n e et c'est beaucoup plus complexe. I n v a r l a n t s m i e r o . En p e r t u r b a n t g~n~riquement on obtient un spectre en bandes =hl + O ( ~ x ) et les calculs dans le secteur utilisent tout le spectre. E. z~O *=0 Fig. et il y a un probl~me de matching. r) qui est en particulier non sousanalytique. r) avec un 2plan qui contient la direction anormale. D e f i n i t i o n 1. Les calculs montrent que les propri~t~s de la proposition 4 se g~n~ralisent tout secteur de Martinet : P r o p o s i t i o n 6. +1) et sont donc rdsonantes.l o c a u x d u p a s s a g e d u col. Dans le cas conservatif des calculs explicites permettent d'dvaluer son bord S(O.1 . les deux valeurs propres du pendule lin~aris~ au voisinage de (~r. r) avec le plan y = 0 pour des raisons g~om~triques. 0) sont ( .
La taille de la grande branche est d'ordre r et c'est donc tr~s diffdrent du cas de contact.. Dans le cas non strict il y a deux branches se ramifiant en A. u 0 A c&s non strict cas strict Fig. Dans chaque secteur les calculs sont diff~rents. . comme en section 5.7.Stratification du Secteur Anormal 249 C u t . et d'~tudier la distance sousRiemannienne par rapport ~ ces objets. etc. (1999) Compactness for subRiemannian length minimizers and subanalyticity. Report SISSA. L'analyse microlocale traduit le fait que dans la sphere sousRiemannienne toutes les directions ne sont pas identiques et qu'il faut calculer par secteurs : secteur Riemannien. secteur de Martinet. Trieste. 7. Agrachev A. Les calculs du cas conservatif conduisent ~ conjecturer que la trace du lieu de coupure L(0) sur la sphbre. au voisinage de la direction anormale. E c l a t e m e n t e n d r o i t e s . L'extr~mit~ C de la branche courte est un point conjugu~. On conqoit donc qu'en gdn~ral une param~trisation analytique de la sphere est impossible. Les rdsultats descriptifs pr~sentds ici reposent sur des calculs longs et complexes dont la difficult~ principale provient de la non validit~ du thdor~me de prdparation lisse au voisinage de la direction anormale.l o c u s . Par exemple dans le secteur de Martinet il faut utiliser les fonctions hyperboliques. en plans. en utilisant la condition de transversalitd. On conjecture n~anmoins l'existence d'une classe de Gevrey uniforme pour chaque probl~me sousRiemannien et donc la possibilitd de rdsoudre l'dquation d'HamiltondacobiBellman dans cette cat~gorie. est reprdsentde sur la Fig. On peut le faire dispara]tre en restreignant la taille du voisinage et l'observateur voit alors deux branches de m~me faille. Cette idle guide d'ailleurs la reprdsentation des vari~t~s Lagrangiennes par des families gdn~ratrices. secteur de contact. References 1. C'est une difficultd d'ordre technique. Une bonne comprdhension g~omdtrique du probl~me peut reposer simplement sur l'id~e d'dclater en droites. l'origine..
Frederick U. Naimark M. Liu W. 13. Sussmann H. 5. (1997) SubRiemannlan spheres in the Martinet flat case. Cambridge University Press. Vol. Preprint. Ouranov V. Arnold V. 18. 20. Gromov M.. (1981) Perturbation methods in applied mathematics. U. Memoirs AMS. 2. S. Hrmander L. rapport ENSTA. Bonnard B.620636. V. Birkh/iuser. NewYork. 139176. Vol. 11. Co. 17. Sarychev A. ESAIM/COCV. 16. 25. Arnold V. Cole J. Kerkovian J.. Journal of Dynamical and Control Systems. (1997) Feedback invariant control theory and differential geometry I. Chyba M. la sphbre et le front d'onde en g6om6trie sousRiemannienne de Martinet. (1946) Lectures on the calculus of variations. Eds Mir. S. Bliss G. 23. Moscou. Chicago. Vol. 343390. (1984) Symplectic techniques in physics. Singularities of differentiable mappings. 5. No. 8. (1989) Elliptic functions and applications. Bormard B. Sternberg S. No. de Morant J. (1993) Th6orie des singularit6s de l'application entr6e/sortie et optimalit6 des trajectoires singulibres dans le problbme du temps minimal. 111159. 15.. 11. Bormard. Forum Math. Kupka I. Tr$1at E. Jean F. PrePrint Dijon. SIAM Journal on Control and Optimization. (1995) Strong minimality of abnormal geodesics for 2distributions. SubRiemannian Geometry. NewYork. Pub.. 118. (1990) Lagrangian manifolds and the Maslov operator. Moyer H. Birkh/iuser. Vol. 6. 9. A. Bonnard B. SubRiemannian Geometry. Bonnard B. 12791311.250 B. SpringerVerlag. 3. 3. (1973) Sufficient conditions for a strong minimum in singular problems. 33. Mischenko A. (1999) Entropy and complexity of a path in subRiemannian geometry. Kupka I. (1995) Shortest paths for subRiemannian metrics of rank two distributions. (1983) The analysis of linear partial differential operators. Kupka I.. (1996) G$om$trie sousRiemannienne. Paris. F. 14. Lawden D. D. (1996) Tangent space in subRiemannian geometry. N564. J. V. 22. Gnillemin V. BeUai'che A. and al. Moscou. . Gamkrelidze R. E. 1. (1999) Role of abnormal minimizers in subRiemannian geometry. Eds Mir. 10.. Nikiforov A. Agrachev A. Journal of Dynamical and Control Systems. 12. 5... 4. SpringerVerlag. Vol. 3. 21. Agrachev A.. and al. and al. (1967) Linear differential operators. Vol. Regular extremals. SIAM Journal on Control and Optimization. No. Agrachev A.. 245334. Trdlat 2.. G. ESAIM/COCV. 377448. (1999) M~thodes g$om~triques et analytiques pour ~tudier l'application exponentielle. 24. (1976) M6thodes math6matiques de la m&canique classique. S6minaire Bourbaki. A. Eds Mir. 7. (1982) Fontions sp$ciales de la physique math6matique. SpringerVerlag. 2. of Chicago Press. SpringerVerlag. (1996) CarnotCarath~odory spaces seen from within. 4. (1992) Abnormal extremals. 19. (1995) Towards a geometric theory in the timeminimal control of chemical batch reactors.
Birkh/iuser. Preprint. MSmoires de la SMF. 28... 27. to appear in Journal of Dynamical and Control Systems. P. (1968) Bifurcations of planar vector fields and Hilbert's 16th problem. 29. Tr~lat E. Berlin. (2000) Some properties of the value function and its level sets for afllne control systems with quadratic cost. SSries divergentes et th$orie asymptotique. Ramis J. Symplectic geometry and analytic hypoellipticity. .Stratification du Secteur Anormal 251 26. Treves F. Roussarie R.
These conditions lead to the notion of a "control oneform" and to necessary conditions for the existence of an orbitally stable periodic motion9 As it turns out. USA Chr isByrnes@seas. This paper is concerned with the development of basic concepts and constructs for a noneqnilibrium theory of nonlinear control9 Motivated by an example of nonstabilizability of rigid spacecraft about an equilibrium (reference attitude) but stabilizability about a revolute motion. Louis. ~ t l . 1 Introduction In 1983 Brockett showed t h a t the origin of t h e nonlinear n o n h o l o n o m i c integrator x2 X3 u2 XlU2 .e. MO 63130. Byrnes Department of Systems Science and Mathematics Washington University. St. and necessary and sufficient condition for the existence of periodic orbits are discussed. Y3) r sufficiently small 9 More explicitly. even when this latter result is specialized to the equilibrium case. we review recent work on the structure of those compact attractors which are Lyapunov stable9 These results are illustrated and refined in a description of the asymptotic behavior of practically stabilizable systems taken form a recent work on bifurcations of the system zero dynamics9 These attractors can contain periodic orbits.X2Ul : Y3 is not solvable for all y = (Yl. i. * Research Supported in Part by a Grant from AFOSR. Boeing and NSF . he proved: Theorem 1 ( B r o c k e t t [3]). the criterion is new.Toward a Nonequilibrium Theory for N o n l i n e a r C o n t r o l S y s t e m s * C h r i s t o p h e r I. Y2. the equation F ( x ) = y is locally solvable. edu A b s t r a c t . A necessary condition for the origin to be a locally asymptotically stable equilibrium of the C t autonomous system x = F ( x ) is that the map F is locally onto.X 2 U I is not a s y m p t o t i c a l l y stabilizable using a C 1 s tat e or d y n a m i c feedback law because u2 = Y2 XlU2 .
As an example. [12]) have pursued the philosophy that controllability. w) takes the form jco = s(~)J~ + ~ b~ui i=1 (1) (21 [~ = S(~IR where J is the inertia matrix and S(w) is the matrix representation of the crossproduct. 19representing rotations about the rl. does not imply stabilizability of an equilibrium.e. diagonalizing a).g. we begin this paper by considering whether stabilization. S(~) = 0 . We will return to this result. r3] using Euler angles. viz. (1)(3) can be expressed in local coordinates about a reference frame R = [rl.254 Christopher I. i. One of the interpretations of this result is that controllability. Explicitly (1)(2) takes the form (4) (o3 La3WlW2J "= r cos(o) o = / sin(0) tan(p) 1 . following [7]. provided more general feedback laws. u). with continuously differentiable feedback. In contrast. respectively. r~. More specifically. the evolution of an orientation.~ 2 0 0~ 1 ] (3) Choosing principle axes (i. On the state manifold M = SO(3) x ~3.o3 OJ2 0..c o s ( 0 ) tan(P)i [sin(0)sec(~o) 0 cos(0)sec(~) J sin(O) ] [!1] wi 3 (5) . r2. Several authors (cf.e. 0. such as momentum wheels or gas jet actuators. u ( x ) ) . about a more general invariant set would be possible. that S ( x . at various points in this paper. b 4 b x w. in an appropriate sense. Byrnes As a corollary he also deduced a basic result about stabilizability of 5: = f ( x . consider the rigid body model of a satellite controlled by momentum exchange devices. angular velocity pair (R. and several of its generalizations. consider the attitude stabilization problem for a system modeled by the equations of motion of a rigid spacecraft with two actuators consisting of opposing pairs of gas jets. r3 axes. in fact does imply asymptotic stabilizability of an equilibrium.y be solvable for all sufficiently small y. [13].. 9. in the nonlinear sense. are allowed. timevarying or discontinuous. e.
are such controllable but not s m o o t h l y stabilizable systems practically stabilizable by state feedback ui = u i ( x . k ) ? We recall that a system is said to be practically stabilizable if there exists a tunable feedback law such that for all initial conditions in any fixed c o m p a c t set. it is also shown that while stabilization about the equilibrium attitude is not possible. including underactuated mechanical systems. for i = 1. the attitude of the underactuated satellite cannot be asymptotically stabilized. More explicitly. several results on the equivalence of feedback stabilizability. Since the latter m a t r i x is invertible. feedback linearization and nonunderactuation are given which foretell more recent work on underactuated mechanical systems. the nonholonomic integrator would be stabilizable about a periodic orbit or some other c o m p a c t attractor. Therefore m Z i1 biui : Yl which can only be solved for arbitrary Yl. dim span{b.dlCd3[ a3c. wi = O. so t h a t the spacecraft is underactuated.Nonequilibrium Theory for Nonlinear Systems 255 Now suppose m = 1 or m = 2. then it is shown t h a t one can design a s m o o t h feedback law which produces an orbitally stable motion about the third principal axis.)lw2J i=1 biui " Yl sin(0) tan(~o) 1 . Motivated by this example.} : 3 and hence m > 3. of course. we ask whether. A particularly i m p o r t a n t example of such a tunable feedback law is high gain o u t p u t feedback. for example. Alternatively. the feedback gain can help to steer the trajectory to an arbitrary small neighborhood of the origin. IlYtl] <:< oc. if two actuators are aligned with the first two principal axes. Let Y = Y2 . Indeed. related by question which has remained open until the recent paper [9]: Q u e s t i o n 1. The main results of [7] are actually stated for a broad class of nonlinear control systems. one can stabilize a b o u t a revolute motion.cos(O) t a n ( ~ )  sin(O)sec(~o) 0 cos(O)sec(~o) J ~2 = U2 3 with y~ = 0. W h a t are the asymptotic dynamics of a practically stabilized system inside the small ball? . In particular.2. 3. In [7]. for a suitable choice of o u t p u t function (see for example Teel and Praly [23] and Bacciotti [1]). These two questions are. Yi E and consider solving the equation [ "Jr" Gl~d2~d3] a2~.
Sundarapandian for their influence on my way of thinking about nonequilibrium nonlinear control theory. the asymptotic dynamics of practically stabilizable systems and on the geometry of periodic phenomena. Alberto Isidori.4 is contractible to xo. I s i d o r i .. Taking y = 0 in Theorem 2 leads to a remarkable consequence. If. The Brockett nonholonomic integrator is not practically stabilizable.4 contains an equilibrium x0. u) = y for all IlYll << ~ . n) f E C1 by C 1 state feedback is the solvability of the equation f ( x . . A necessary condition for global asymptotic stabilization about a compact attractor . C o r o l l a r y 1. I s i d o r l . M a r t i n ) . Victor Shubov and V. It is a pleasure to thank Roger Brockett. If `4 C Rn is a globally asymptotically stable compact attractor for a smooth dynamical system = f(x) then (i) `4 contains either an odd or an infinite number of equilibria. Clyde Martin. T h e o r e m 2 ( B y r n e s . we will review recent results on the nature of Lyapunov stable attractors in/R n. Example 1.4 C ]~n for a system = f ( x . ~Rn then . These results follow from T h e o r e m 3 ( B y r n e s . (ii} If xo E . Byrnes In this paper. Anders Lindquist. M a r t i n ) .4) << ~ . for some x with dist(x.256 Christopher I. . David Gilliam.4 C ~'~ is a globally asymptotically stable compact attractor for a smooth dynamical system ~ = f(~).4 is an equilibrium. James Ramsey. 2 Preliminaries on Compact Attractors Our first result about the existence of compact attractors extends Brockett's criteria to the case of compact attractors [10].
Ind(Fu) = x(Mc). The proof of contractibility uses Lyapunov stability and attractivity o f .4 is contractible. Moreover.4 L/V=O on ~ ' \ A Set Me = {x : Y(x) < c}. The boundary Ne =. Remark 1. smooth hypersurface. The Poincar4Hopf Theorem also holds for continuous vector fields. It is wellknown [25] that there exists a Lyapunov function V for . Fy has either an odd number of equilibria.4 and z "." As in the GaussBonnet Theorem. . For Ilyll << oo. F~ : Ne_~_l. as in [26.(~o)=O and where fJF~Jl is the Gauss map. [7]) Remark 2.OMe is a compact. Me is contractible. 4 to show that any m a p T : SP ~ Me is homotopic to a constant m a p (d'apr~s the proof of Brockett's Theorem given by ByrnesIsidori. For c > 0.f ( x ) . Therefore Ind(Fu) = 1 and.27]. Our principal interest has been in introducing the basic ideas and results. the index of the vector field can also be computed as the curvatura integra of the normalized outward normal on Are (see Milnor [19]). Wilson [25] showed that `4 ~ Me is a homotopy equivalence and so .Nonequilibrium Theory for Nonlinear Systems 257 Proof. or an infinite number. L! V < 0 on ~'~\. Remark 3. Me is a c o m p a c t nmanifold with boundary. leading to the more familiar formulations of the proofs of Brockett's Theorem for equilibria. Define Ind(Fy) = where Z i=o(fu ) F. Fu = . The degree of the Gauss m a p is also known as the "curvatura integra. in particular. By the Poincar4Hopf Theorem. Claim. so that one should certainly expect similar results in the absence of smoothness of the system.f ( x ) + y is a smooth outward normal on Ne. based on degree theory and homotopy methods.
) = q(z. compact attractor for the larger sublevel set. it is wellknown that this system can be "semiglobally practically" stabilized by "highgain" negative output feedback u : ky . y) . One can in fact replace the small and large balls in this construction by the sublevel sets of a Lyapunov function for the closedloop system. whose interior can be shown to be diffeomorphic to ~ n [7]. u E ]~.k y asymptotically stabilizes the origin for any k > O. stable equilibrium at z = 0. so that each of the sublevel sets are positively invariant. y E ~ and 1(0) = 0 p(z. [ y z+u zerostate detectable. Byrnes Compact Attractors and Practical Stabilizability Consider a system { k = / ( z ) + p(z.l . This renders the smaller sublevel set a Lyapunov stable.258 3 Christopher I. Then. = . in the odd number (counting multiplicities) of equilibria contained in the attractors. The large and small balls in practical stabifizability R e m a r k ~. but not possibly locally exponentially. for example. o) = o . 1. and therefore u = . Consider ~ ~. y) + u .z 3 + y This system is passive (positive real). in which z E l~ n . . Example 2. Suppose that the zero dynamics ~ = f ( z ) has a globally asymptotically. In particular. the results of the previous section apply to the smaller sublevel set. This will be manifested. o) = o q(o. Fig.
(k1/2.k y . In fact. 2. Moreover. as well as an early description of what is now known as "backstepping. while the other two equilibria are hyperbolic stable points. . It follows from the general results we are about to state that this attractor is Lyapunov stable for k ~ ~ 0. The Lyapunov stable compact attractor with three equilibria Example 4.Nonequilibrium Theory for Nonlinear Systems 259 On the other hand."). Consider the system I has equilibria defined via ~: Z3+y il= zky. (z. . In this case this attractor coincides with the origin. >> Fig. Consider the system { Zl = z2 Z2 Zl ezlz~ + Y ~/ Z~ .Z 3 "~ lZ. z = ky. the origin is an unstable hyperbolic point..0). Example 3. it has long been known that there are minimum phase systems with relative degree one for which output feedback renders the system unstable for every choice of the gain (see [6] for an example. which is an attractor for bounded halftrajectories (see [15]). the set of initial conditions which generate a bounded trajectory forward and backward in time is the maximal compact invariant set.]r (k1/2.k3]2) .y) = (0. We now begin a study of what can happen in such an unstable case. its 1dimensional unstable manifold and the two asymptotically stable equilibria.
the system has an unstable equilibrium at the origin and a stable limit cycle entirely contained in B6 These examples illustrate the asymptotic behavior of high gain closed loop system in two cases in which the zero dynamics are critically stable in the mildest possible way. Q ~S 1 Fig. for any 6 > 0. if k > k ' . Compact attractors: k = 500 In its linear approximation the system has two eigenvalues with positive real part for any large positive k. 3. 4.260 Christopher I. in a sense that we will now make precise. { zl z2 z2 : Zl ez~za For k >> 0 the origin is unstable but there is an asymptotically stable limit cycle. Compact attractors: k .at$ Fig. (zl. there is k* > 0 such that. As a matter of fact. Indeed. 0) is globally asymptotically stable for the zero dynamics. . Byrnes resulting from output feedback applied to a globally m i n i m u m phase system with unity high frequency gain.50 . z~) = (0.
again. in polar coordinates. a < 0 = w + br 2 t. as (see e. Moreover.. e) = ed(s) + n ( s ) = 0 and consider a branch s(e) for which Re s(O) = 0 . a C a s e B. corresponding to the linearization of the open loop system having a pair of imaginary eigenvalues + j w . < O. Case A corresponds to the situation in which the zero dynamics has a onedimensional center manifold. we set k = 1/e obtaining f ( s . corresponding to the linearization of the open loop system having a zero at 0. therefore..O(]rl 4) . In the defining equation of the rootlocus. For large k. Case B corresponds to the situation in which the zero dynamics has a twodimensional center manifold. approaches this zero from the right.Nonequilibrium Theory for Nonlinear Systems 261 C a s e A. [24. The linearization of the zero dynamics at z = 0 has a simple zero eigenvalue and all remaining spectra is in C . to determine whether for large k the real pole of the linearization of the closed loop system which approaches the zero at 0 of the linearization of the open loop system: approaches this zero from the left. the (2dimensional) restriction of the zero dynamics to its center manifold at z = 0 can be expanded. or cancels this zero. two complexconjugate eigenvalues of the linearization of the closed loop system approach the zeros at =t:jw of the linearization of the open loop system from the left or from the right and. the (1dimensional) restriction of the zero dynamics to its center manifold at z = 0 can be expanded as = a~ 3 + o(1~14). page 271]) § = ar 3 + O(Ir[ ~) . it is key to determine which case actually Occurs. Moreover. By g(s) = n ( s ) / d ( s ) we denote the transfer function of the linearization of the openloop system. The linearization of the zero dynamics at z = 0 has two purely imaginary eigenvalues and all remaining spectra is in C .g. The key issue is.
R) tend to 0 as t + oo (see Figure I). T h e o r e m 4 ( B y r n e s . ( . ~(~) 9 so(tJ . e) and setting e equal to 0 will yield an expression for the tangent to the local branch at e = 0. implicitly differentiating f(s. that s* = 0 so t h a t we are considering Case A.262 Christopher I. Consider a globally minimum phase system with positive high frequency gain Lgh. which has been practically stabilized by u = .J 0 ) = (1)ndet(Jo) sign(n'(O)) = +1 In particular. for illustration. I s i d o r i ) .j s(c) Fig._. Byrnes ~.~ ]e. 5.1 ) n det(Jo) < 0 r 08 ~ee > O. Suppose.~O Os II 9 n'(s*) = d(s*) Example 5. Let JOL denote the Jacobian of the openloop system. and all remaining spectra is in C". If ( .1 ) " d e t ( J o z ) > 0.k y . I . Suppose the linearization of the zero dynamics has a simple eigenvalue at O. The "reverse" root locus departing from imaginary openloop zeros as k decreases from +cx~ If the zero is simple. then the origin is locally exponentially stable and there exists a k* such that for k > k* all trajectories initialized in B(0. . Substitution yields d(0) = d e t ( . I.
4. The attractor . belong those systems for which an appropriately defined zero dynamics exists and is exponentially stable. which were the objects of much research during the 1980's.. these systems have relative degree one and the transmission zeros of the linearization lie in the open lefthalf plane. Fig. as depicted below. p~ o b \.1 ) n det(JoL) < O. In the late 1980's it was clear that high gain feedback together with inputoutput linearizing feedback could semiglobally stabilize those systems with critically asymptotically stable zero dynamics [6. By the 1990's it became widely appreciated that these classes of systems were feedback equivalent to a passive system and that passive systems themselves provided an underlying stability mechanism that was strong enough to allow .A is either the origin. I f ( . The root locus plot for the linearization of an exponentially minimum phase system These are the classes of systems for which the origin can be exponentially stabilized using high gain feedback and for which singular perturbations apply [5. To the first class of systems. This analysis of the Lyapunov stability and the structure of nontrivial attractors can be motivated by an analysis of the rootlocus plots of three general classes of systems.4 consists of the origin and its onedimensional unstable manifold.Nonequilibrium Theory for Nonlinear Systems 263 . 6. 9 . there exists a compact. for any R > 0 and r > 0 there exists a k" such that for k > k* there exists a compact Lyapunov stable invariant set . In particular.8].4 as t + ~ .4 is a 1manifold with boundary consistin 9 of 2 asymptotically stable equilibria.4 C Br(0) and all trajectories initialized in BR(0) tend to . It is also shown that these cases can be distinguished by the signature of a quadratic form defined on the mapping germ of the closedloop vector field. The (relative) interior of .17]. a critically asymptotically stable equilibrium of multiplicity three. Remark 5. there exists a k* such that for k > k*. In [9] it is shown that if d e t ( J ) = 0. or a 1manifold with a boundary consisting of the origin (which is a critically unstable equilibrium of multiplicity two) and a locally exponentially stable equilibrium.. Lyapunov stable attractor.
8.11] following earlier pioneering work on passivity and dissipativity in [28. for the feedback law u = .264 Christopher I. 7. The root locus plot for the linearization of a passive system This stability property is also reflected in the inequality ( . / i 0 ? i / J Fig. Consider case B.k y . T h e o r e m 5 ( B y r n e s . with k > 0 a globally minimum phase system with Lgh > O. I s i d o r l ) . The root locus plot for the linearization of a passive system undergoing a bifurcation This class of systems is precisely the class of feedback systems undergoing a nontrivial bifurcation. Fig. Of course.1 ) n det(JoL) > 0. one has a class of systems for which the closedloop poles can tend to the imaginary axis openloop zeros along an arc contained in the open righthalf plane as depicted below. Then: . for passive (or positive real) systems the closedloop poles will always tend to the imaginary axis openloop zeros along an arc contained in the open lefthalf plane as depicted in Figure 7. Our next result discuss the case of a Hopf bifurcation of the nonlinear zero dynamics. Byrnes a linearization with openloop zeros and poles in the closed lefthalf plane (see [20. When the reverse inequality holds. which in the case discussed above is a pitchfork bifurcation of the zero dynamics.21.16]).
y + eq(zl. ~ ) + ~p~(z~. In system (6). where e > 0 is a small number. z~) + p2(~.2. T h e resulting system is a singularly perturbed system that can be studied via s t a n d a r d centermanifold methods. z~) vanishes together with its first partial derivatives. z2. z~.. as in Fenichel [14]. z~) + ep~(z~. 0 [ J. ~(Zl. y) f/= q(zl. denotes derivative with respect to r) el=O z~ = eAltzx + egl(zl. y) ~/ : . If Re (d(jw)/n'(jw)) > O. ~ . e) 0.~ ( z l . z~.z2.z~) + epl(zl. e) + eq(zl. = o.4. z2. then for k > k" > 0 there is a compact. If Re (d(jw)/n'(fio)) < O.z2. e) satisfying OZ 1 [eAllz. z~.00o. ~ .0) = 0. z~) + Pl (zl. 0 0[ 1. y) ~ = Sz~ + g~(z~. y) . the graph of a m a p p i n g y = ~r(zl.k y and set k = l/e. z2. 2.4 is a smooth 2manifold with boundary consisting of a periodic orbit which bounds a smooth open surface. the matrix S is skew symmetric. rr(zl. 4 The Bifurcation of the Zero D y n a m i c s Consider the system in normal form / zl . The open surface consists of the origin and its 2dimensional unstable manifold. (9) . we introduce a "highgain" o u t p u t feedback control u = .000.A11 zl + 91(zl. y) + u. (7) The system (7) possesses an ndimensional center manifold at (zl. We can add the extra equation g = 0 and rescale time as r = tie to obtain the standard form (here " i . z~. then the origin is semiglobally stabilizable via u = ky. z~. z~) + epl(zt. + egt(zl. Lyapunov stable attractor. z2.Nonequilibrium Theory for Nonlinear Systems 265 1.e))] +~ OZ 2 [~Sz~ + ~g~(~. for i = 1. z~. gi(zl. z2. d) L (8) = . z2. and pi(Zl. y) } z~ = eSz~ + eg~(~. (6) in which the matrix A l l is Hurwitz. e)) and (000. y. ~(z~.
e) = (0. D. e) = 0. In [2]. e). z2. O.266 Christopher I. e).0) " J OeOz~ J (0. e) = (0. After using a stable manifold argument. z~) + p~(7(z2. since r ( z l . z2. For planar systems. if a vector field leaves an annulus positively invariant but has no equilibria within the annulus. z2) + p~(z~. z~. then the vector field possess a periodic orbit in the annulus. 0). 0. 0. . Birkhoff describes. augmented with t = 0. ~). z~. Moreover. the PoincardBendixson theory gives a very complete picture. ~(z~. without proof. since for any e the point (zl. d). we can conclude t h a t 7r(O. e) = Ir(7(z2. dim(z2) = 2 and S has two nonzero purely imaginary eigenvalues. which is positively invariant under a vector field f . we then have either a pitchfork bifurcation or a H o p f bifurcation. ~). 7r(zl. z~. . z2. is a regularly perturbed system and possesses a center manifold at (zl. z~. In this section we are interested in the existence and the geometry of periodic phenomena.0) The associated reduced system is ~ = Sz2 + s~(~(z~. z2. O) = 0 for i = 1. ~ ( ~ .z2. z~) + p l ( z l . G. ~)).0. e)) ~2 = Sz2 + g2(zx. this yields the zero dynamics of the original system. 0) = 0. Moreover [oo~ 1 =[ o~~ 1 L0/~z~ (0. T h e system. In particular. ~) is an equilibrium. the graph of a m a p p i n g zt = 7(z2.2. with unsealed time zl = A l l z l + g t ( z l . 0 ) = 0 and pi(zl. y. Birkhoff asserted the existence of a periodic orbit 7 C M provided there exists functions al such that (i) ~ = Oai aaj 0zi' "J=l'''''n' . z 2 . Byrnes Clearly ~r(zl. e) = O. Set tr(z2. for e = 0. a generalization of this situation. T h e Lyapunov stability of both attractors follows from stable manifold theory and an extension of the reduction theorem of center manifold theory for equilibria to the case of a c o m p a c t attractor [9] 5 Existence of Periodic Phenomena In previous sections we exhibited the existence of of periodic orbits in high gain feedback systems. Cases A and B result in the two special cases dim(z2) = 1 and S = 0. e). If there exists a solid ntorus M in IR". At (z2. We call this system the perturbed zero d y n a m i c s and note that. z2. we can focus on the associated "reduced system".
and note that dim (W" tq W " ) = 1. In particular. In [4] a generalization of this result yields a set of necessary and sufficient conditions for the existence of a periodic orbit. ~ Fig. W u are smooth invariant manifolds which intersect transversely in 7. 9. x0). Petrowski. (i) asserts t h a t do. Consider a system Jc   f(x) (10) with periodic orbit 7. 7 is hyperbolic if s + u = n + 1. we begin with a description of the behavior of solutions of a differential equation near a periodic orbit (d'aprds Lewis. These can also be constructed via the Poincar~ m a p at each P E 7dimW u=u. a result with several corollaries. 7) + 0 ( exponentially as) t + . Markus. The following result can be found in [18]. = 0 and (ii) i1 asserts t h a t the contraction (f. We introduce the notation dimW s=s.o 0 } .Nonequilibrium Theory for Nonlinear Systems fi 267 (ii) Z i=1 a. :P : D 2 +/~ T(x) = a~' (PJ . W s and W u are disc bundles over the circle. T h e o r e m 6. fl > 0 Indeed defining the oneform w = ~~aidxi. The stable and unstable manifolds of 7 are defined as w'(7) = {x0 : dist (x(t. Sternberg). W ~. and only if. 7) "+ 0 ( exponentially as ) t + oo} and w ~ ( 7 ) = {x0 : dist (x(t. 7 is e~onentially orbitally stable if. w) > 0 of f with w is positive. C o r o l l a r y 2. x0). To appreciate the necessity of this kind of geometric situation. s = n.
I f u + s . In p a r t i c u l a r = and W ~(7) = u w'(P) PE'r U W ~(P) PE'~ define c o d i m e n s i o n one f o l i a t i o n s o f W ' (7). f 2 ( v .u .268 Christopher I. WU (7) r e s p e c t i v e l y . W e see t h a t P h y p e r b o l i c i m p l i e s u + s = n + 1 w h i c h i m p l i e s 7 is h y p e r b o l i c .(M) ~_ {0}. o) = f~(v. In t h e P o r b i t a l l y s t a b l e case.s t a b l e m a n i f o l d W e ' ( 7 ) = O W e` ( P ) . F i g . such t h a t (11) 7r.2 < n .s. Byrnes R e m a r k 6. W e ' ( 7 ) is a solid torus.1. Pe'~ It follows t h a t 7 is o r b i t a l l y s t a b l e if u = 1.o).I. d i m W u ( P ) = u . t h e n t h e r e is a c e n t e r m a n i f o l d W e ( P ) for P dimWe(P) = n + 1 . w i t h or w i t h o u t b o u n d ary. a n d 7 is o r b i t a l l y s t a b l e o n W e. i > 1 . T h e r e also exists a c e n t e r . 10. O n e c a n also see t h i s in local coordinates = f~(v. o ) > o. There exists a smooth invariant manifold W e containing 7 such that W e ( 7 ) = O W e ( P ) is a codimension one foliation. T h e o r e m 7. A periodic orbit and its center manifold D e f i n i t i o n 1.1. A c i r c u l a r m a n i f o l d is a m a n i f o l d M . d i m W ' ( P ) = s .
The second example in 11. (iii) w satisfying (f. H. W ' . 1). Examples of compact circular manifolds Remark 8. 269 R e m a r k 7. F i Fig. solid tori. w) > O. W e. positively invariant circular submanifold M and a closed one f o r m w on M satisfying ( f . compact cylinders. w) > O. Proof. In general. W e. Edwards. Necessity: M = 7. (ii) a codimension one foliation on M defined by a closed oneform w. Necessary and sufficient conditions for the existence o f a periodic orbit 7 are the existence of a compact. the fourlinked unknotted torus was generated in m a t l a b using the tori4 mfile written by C. a nonlinear system ~ = f ( x ) and a periodic solution 7 give rise to (i) an invariant circular manifold M with f pointing inward on cOM. we m a y take e = 1. w = dO (although W ' . W i t h o u t loss of generality. Fix P E M . T h e o r e m 8 ( B r o c k e t t . Construct the "period map" J : M + S 1 as follows. B y r n e s ) . 11. and in this case is said to be a K ( Z . For any Q E M choose a p a t h 7~ from P to Q and consider .Nonequilibrium Theory for Nonlinear Systems Examples include circles. Any space satisfying (11) is an example of an EilenbergMacLane (classifying) space. then. W 'c. W "c provide more interesting constructions) Sufficiency: Choose a generator 71 for ~rl(M) ~ Z for which f'yl w = c > 0.
"/'2 " " ~ "/1 SO Y~.270 Christopher I. in R. 12. The Lefschetz number computes the oriented intersection of the graph of f ( z ) with the diagonal (the graph of the identity map) in MxM. If 73 is another path "~3 . Lef(T) = 1 r 0. 1.0.. and appealing to the Lefschetz Fixed Point Theorem. m o d Z.l ( O ) . Byrnes P Fig.(Q) = Y~ (Q) + t. we show that T has a fixed point by proving nonvanishing of its Lefschetz number. Remark 9. . The path of integration from P to Q on a circular manifold. is a leaf of w . and is the domain of a Poincar~ m a p Y(~0) = ~r(~0) 9 J . Finally. Therefore we may define the period m a p J : M+S 1 =~IZ via J(O) = F w. From the geometry of J it follows [4] that Jl(O) is connected. Or.
for periodic orbits. f ( x . u = a(x). If (12) has a control oneform. These conditions are satisfied by small perturbations f ( x . Then. This existence criterion for closed orbits has many corollaries and illustrations. We next turn to the appropriate analog. of Brockett's necessary condition for asymptotically stable equilibria. p ) of f(x) and therefore = f(x. p) possesses a compact. for any jointly continuous perturbation Jc = f ( x . We have actually shown that Jx(O) is contractible and that there are either an infinite number of periodic orbits on M or. if the orbits have finite multiplicity. Po) = f ( x ) . The fixed point set as the intersection of the graph and the diagonal in the product manifold. there must exist a periodic solution. which creates a periodic orbit for the closedloop system. ~ = f ( x . an odd number of orbits (counted With multiplicity). f ( x ) ) <_ 0 =~ (w. Suppose = f(x) has an asymptotically stable periodic orbit. Consider the system (12) Suppose M is a circular manifold such that f. . 2. 13. of which we shall mention two. C o r o l l a r y 3. f ) > O. Proof. nonexact one form w on M is a control oneform provided ( w . p) where for some Po. solid torus M for which f ( x . C o r o l l a r y 4. g are inward normal vector fields on a M . for each p ~ Po.MV1 = f(x) + g(x)u Nonequilibrium Theory for Nonlinear Systems 271 Fig.g(x)) ~ O. p) possesses a periodic orbit. A closed. then there exists a state feed back law. po) points inward on a M and for which there exists a closed 1form w satisfying (w.
f o r any jointly continuous perturbation where f o r some po. . ) = x(Mc) = +1. p ) is locally onto.p and substituting y = p we obtain.272 Christopher I. Consider the nonlinear system x = f ( x ) where f is locally C 1 in x and p. Proof. Moreover. Byrnes 6 Necessary Conditions For the Existence of A s y m p t o t i c a l l y Stable Equilibria. By defining f ( x .I ( . and passes Brockett's criterion but not the criterion we obtained by specializing from the periodic case to the case of equilibria. the equation f ( x . c ] .~)ll is a unit outward normal to Me. N . i. p) . Po) is that f o r all P "" Po E ]Rk. we review the recent results reported in [22] on the situation when T ( p ) + 0. T h e o r e m 9 (Byrnes. po) = f ( x ) . In the previous corollary. since the system we construct is not asymptotically stable. the map f ( . p) = f ( x . .p F. f ( x . A necessary condition f o r x = 0 to be a locally asymptotically stable equilibrium of the system x = f ( x .e. for each p . by defining F ( x . Suppose ~ = f ( x ) has an asymptotically stable equilibrium. (Sketch of Proof) Let Mc = V . f(x. we presented an analog of Brockett's necessary condition.y we obtain: C o r o l l a r y 5.~ po.o o . there must exist an equilibrium solution. In this section.F ( x ) . = OMe = V . p ) = y is locally solvable. of course. p) . IIf(~. Brockett's Theorem. Then. valid for periodic orbits of period T(p). Our final example shows that this new result is stronger... The Poincar~Hopf Theorem again gives that I n d ( F .l ( c ) . Sundarapaudian). Po) = O. and f(O.
~nd IFA C Syrup. Dover. tan1 [ f(x. These properties and claims are most easily seen using a complex variable argument.J. 2. 181191. (1968) Collected Mathematical Papers. Louis. G. Birkhoff. 3. eds.Im(z) > 0 . but has degree zero. Washington University.im(z) Therefore f ( z . Consider the system ~ = f ( z . Vol.W. R. 4. 9 1 Thus we have f(z. 2125. z # 0 O.' for x . 1) maps the closed upperhalf plane onto ~ 2 and therefore satisfies the necessary condition of the corollary.. St.S. Brocket. References 1. 2. p) for p .p)= (x~ + x~) u cos 2p t a n . z#0 ~2. W.v) = { z 2~'. (1992) Linear feedback: the local and potentially global stabilization of cascade systems. H. Proc. Brockett. D. R. Brockett. R.. p) is not locMly onto for p<l. Millmann. Byrnes. We define the complex variable z = xl + ix2. 1) is surjective. We also note that f ( x . Birkh~iuser. A.f + sin (2. Boston. < 0 [_(xl2+x~)Vsin(2ptan_l(T_~x~) ) We note that f is C ~ on ~2\{0} and C l at 0 provided P > 2" Now f ( x . Also f ( x . p ) is never surjective for p < 1. Nonlinear Control Systems Design. Differential Geometric Control Theory. W. (1983) Asymptotic stability and feedback stabilization. C. Bacciotti. I.~ P0 = 1 where X2 " 273 '(x'") : L(.. R. (2000) The geometry of periodic orbits with applications to nonlinear control systems. z=O <0. Preprint Department of Systems Science and mathematics. Sussmann.Nonequillbrium Theory for Nonlinear Systems Example 6. .
A. 10. 12281240. .. M. C. 18. Yu. Fenichel. (2000) Bifurcation analysis of the zero dynamics and the practical stabilization of nonlinear minimumphase systems.. Byrnes. 14431485. 13. Sundarpandian. E. 423424. A. H. 11. 26. 12. R. I. Inc.F. Eqs. Ortega. with Applications to Stabilization and to Adaptive Control'. Milnor. (1989) Passivity properties for stabilization of cascaded nonlinear systems. Louis. 19. 72. Diff. IFAC 27. Control Optim.... 13941407. V. C. C.. C. 14. IEEE Trans. on Decision and Control. 877888. I. A. and the global stabilization of minimum phase nonlinear systems. J. (1990) Introduction to applied nonlinear dynamical systems and chaos. pp.. (1997) Asymptotic controllability implies feedback stabilization.274 Christopher I. W. Spong. Hale. Contr. IEEE Trans. A. Byrnes. 413428. (1989) Adaptive motion control of rigid robots: A tutorial. l. Marino... Contr.. J. I. C. New York. S. M. W. Priceton University Press. Automatica. of Math. SpringerVerlag New York. Las Vegas. Martin. 8795. J. Wilson. C. A. Math. Contr. IEEE Trans. J. Aut.. Syst. 25. (1976). Control. Moylan. Control.. Math. W. Preprint Department of Systems Science and mathematics. (1985) Highgain feedback in nonlinear control systems. Aut. C. C. Syst. (1989) New results and examples in nonlinear feedback stabilization. I." submitted to Int. A. trans. 437442. J. Magalhaes. Isidori. (1991) Asymptotic stabilization of minimum phase nonlinear systems.R. M. Byrnes... Coron. Princeton. 17. Subbotin.. (1991) On the attitude stabilization of rigid spacecraft. (1984) 'A Frequency Domain Philosophy for Nonlinear Systems. (1969) Smoothing derivatives of functions and applications. R. AC36. A. Automatica J. Soc. F. 42. (2000) On the structure of globally asymptotically stable attractors. Sontag. Si9. (1992) Global asymptotic stabilization for controllable systems without drift. T.. Olvia. 31. Byrnes. no. Byrnes 5. lsidori. Submitted to Asian Journal of Control. (1967) The structure of the level surfaces of a Lyapunov function. 25.. Automatica. 16. Wilson.. S. J. 42. Isidorl. Byrues. 15. 2. Amer.Verla9. 5393. and Praiy. Proc. W... Springer. Isidori. R. 3. L. K. AC21. L. 323329. 5. (1979) Geometric singular perturbation theory for ordinary differential equations. Aut.. Wiilems. 1. of Diff. NJ. D. AC36.... J. 139.. 13691385. I.. Isidori. 295ff. l. Stability of nonlinear dissipative systems.geometric theory. Control Left. L.. J. Isidori. Byrnes. 33. (1960) The behavior of the solutions of a differential system near a periodic solution. Contr. A. Jr. (2000) Persistence of equilibria for locally asymptotically stable systems. C. (1995) Tools for semiglobal stabilization by partial state and output feedback. St. 11221137. Teel. P. Ann. J. Markus. 6. F.. Isidori. 245266. 24.. No.. of Robust and Nonlinear Control. Autom. 7. Byrnes. 27. C. 8. 12. I. 20. IEEE Trans. 9. A. J. 22. Hill.. Int. F.708711. Washington University. (1984) An introduction to infinite dimensional dynamical systems . Ledyaev. of ~3rd IEEE Con].. Jr. Ortega. feedback equivalence. D. N. 21. Clarke. l. Byrnes. (1997) Topology from the differentiable viewpoint. Wiggins. 23. SIAM J.. (1991) Passivity. Eqs.
Nonequilibrium Theory for Nonlinear Systems 275 27..F. Archive for Rat. Part I: General theory. and Anal. W. Diff. (1973) Lyaounov functions and isolatingblocks.C. 106123. Willems.. J. Dissipative dynamical systems.. 28. Eqs. Yorke.. 13. A. . J.. Mech. 45. Wilson. (1976). Jr. 321351. J.
it 2 Fachbereich Mathematik J. Germany gruene@math . France. Analysis und effiziente Simulation dynamischer Systeme" ~' This paper was written while Fabian Wirth was a guest at the Centre Automatique et Syst~mes. 60054 Frankfurt a.17].9] of this p r o b l e m .W.5.de We derive a method for the computation of robust domains of attraction based on a recent generalization of Zubov's theorem on representing robust domains of attraction for perturbed systems via the viscosity solution of a suitable partial differential equation.13. We present the resulting fully discrete scheme and show a numerical example. Several p a p e r s a n d books discuss t h e o r e t i c a l [19. the analysis of power s y s t e m s [1] a n d t u r b u l e n c e p h e n o m e n a in fluid d y n a m i c s [2. T h e knowledge of this o b j e c t is i m p o r t a n t in m a n y a p p l i c a t i o n s m o d e l e d by those s y s t e m s like e. GoetheUniversit/it. and the DFG Priority Research Program "Ergodentheorie. The hospitality of all the members of the centre is gratefully acknowledged. c aspur. di Ingegueria UniversitA de l'Aquila. * Research supported by the TMR Networks "Nonlinear Control Network" and "Viscosity Solutions and their appfications'. 1 Introduction T h e d o m a i n of a t t r a c t i o n of an a s y m p t o t i c a l l y s t a b l e fixed p o i n t has been one of the central o b j e c t s in the s t u d y of continuous d y n a m i c a l systems.1. a n d F a b i a n W i r t h 3 ~ 1 Dip. a suitable regularization enables us to apply a standard discretization technique for HamiltonJacobiBellman equations. dl Energetica.g. Postfach 11 19 32..20. Fac. While a direct discretization of the equation leads to numerical difficulties due to a singularity at the stable equilibrium. 67040 Roio Poggio (AQ). Lars Griine ~. Ecole des Mines de Paris.12] as well as c o m p u t a t i o n a l aspects [18. Italy camill iQaxcasp. Abstract.unifrankfurt.8. Fontainebleau. de 3 Center for Technomathematics University of Bremen 28334 Bremen. Germany f abian@math. .unibremen.A Regularization of Zubov's Equation for Robust Domains of Attraction* F a b i o C a m i l l i 1.M.
In this paper we use a similar optimal control technique.e. In Section 4 we introduce the regularization technique and formulate the numerical scheme. i.) we try to find the set of points which are attracted to x* for all admissible a(.4. the sublevel set characterization of the robust domain of attraction.g. but requires some knowledge about the local behavior around x* in order to avoid discontinuities in the optimal value functions causing numerical problems. but start from recent results in [4] where the classical equation of Zubov [20] is generalized to perturbed systems. The main problem in a numerical approximation is the inherent singularity of the equation at the fixed point which prevents the direct application of standard numerical schemes. in [14.e.a(t)). In particular. The techniques in these papers have in common that a numerical approximation of the optimal value function of a suitable optimal control problem is computed such that the robust domain of attraction is characterized by a suitable sublevel set of this function. This paper is organized as follows: In Section 2 we give the setup and collect some facts about robust domains of attraction. In Section 3 we summarize the needed results from [4] on the generalization of Zubov's equation for perturbed system. It might be worth noting that in particular our approach is applicable to the classical Zubov equation (i. This set has been considered e. Under very mild conditions on the problem data this equation admits a continuous or even Lipschitz viscosity solution. in [14] and [7] numerical procedures based on optimal control techniques for the computation of robust domains of attraction are presented.). . Lars Griine and Fabian Wirth Taking into account that usually mathematical models of complex systems contain model errors and that exogenous perturbations are ubiquitous it is natural to consider systems with deterministic time varying perturbations and look for domains of attraction that are robust under all these perturbations. in Section 5 we show a numerical example. Here we consider systems of the form x(t) = f(x(t).7]. Here we propose a regularization of this equation such that the classical schemes [6] and adaptive gridding techniques [11] are applicable without losing the main feature of the solution. x E 11~n where a(.278 Fabio CamiUl. Under the assumption that x* E IRn is a locally exponentially stable fixed point for all admissible perturbation functions a(. for unperturbed systems) and hence provides a way to compute domains of attraction also for unperturbed systems.15. Whereas the method in [14] requires the numerical solution of several HamiltonJacobiBellman equations (and is thus very expensive) the method in [7] needs just one such solution. and finally.) is an arbitrary measurable function with values in some compact set A C ~m.
For the system (1) satisfying (HI) we define the robust domain of attraction as i/) = {x0 6 ~ n : x(t. t h a t :Do is an open. Furthermore.D v ( z ) f ( z . We assume that the singular point 0 is uniformly locally exponentially stable for the system (1). (HI) there exist constants C. t E [0.v(x))g(x. A) and A is a compact subset of ]Rm.th.a)ll < CeOt[l~011 for any ~0 6 B(0. . i. we assume t h a t the fixed point x = 0 is singular. +cx~).(1 . 4 " For a collection of properties of (uniform) robust domains of attraction we refer to [4. and the uniform robust domain of attraction by { :Do = there exists a function/3(t) + 0 as t + oo } x0 6 ~ n : s. xo. D e f i n i t i o n 1. a) .4} . 3 Zubov's method for robust domains of attraction In this section we discuss the following partial differential equation aEA inf { . t h a t is f(0. The following sets describe domains of attraction for the equilibrium x = 0 of the system (1). In this generality.a) + 0 as t + +oo for any a 6 .xo. oo). in order to obtain a meaningful result about solutions we have to work within the framework of viscosity solutions.) E . z(0) = zo. r) and any a 6 m.4 = L ~ ( [ 0 . connected and invariant set. a)l I <_ fl(t) for all t > 0. a. which we recall for the convenience of the reader (for details about this theory we refer to [3]). Here a(. r > 0 such t h a t IIx(t. Proposition 2. Ilx(t. a)} = 0 x E ~'~ (2) whose solution will turn out to characterize the uniform robust d o m a i n of attraction Do. f is continuous and bounded in ]R" • A and Lipschitz in x uniformly in a E A.e. a 6 .~0. a(t)).A Regularization of Zubov's Equation 279 2 Robust domains of attraction We consider systems of the following form ( x ( t ) = f(x(t). There it is shown in particular. This equation is a straightforward generalization of Zubov's equation [20]. (1) with solutions denoted by x(t. a) = 0 for any a E A.xo. a). however. and t h a t the inclusion 7) C cl:D0 holds.4].
Since g is nonnegative it is immediate that v(x) 9 [0. a simple application of the chain rule shows (1 = ..a(t))dt and v(x):=suple aE.t. x. [3. (4) where the function g : ~'= x A + 1E is supposed to be continuous and satisfies (i) For any a 9 A.s. u(x).r). We now introduce the value function of a suitable optimal control problem related to (2). a) = 0 and (H2) g(x. (6) Furthermore. a) ~ go. for each t > 0 we have v(x) with = sup {(1 aEA G(x.r (3) we if for all r E Cz(Y2) and x 9 a r g m i n a ( u . a) g(y. aEa g(X. subsolution) of the equation H(r. x. g(0.a) > 0 for x r 0. (ii) There exists a constant go > 0 such that infxeB(0.e.a) := exp (/0  g(x(v.4 c~(~'~).c.) function v : ~ + ~ ) is a viscosity supersolution (resp.a)v(x(t. 1] for all x 9 ~ n .a))} (5) G(t. H ( x . (iii) For every R > 0 there exists a constant LR such that IIg(x.r have H(x. C h a p t e r III]) imply that v satisfies a d y n a m i c p r o g r a m m i n g principle. x. u. x 9 a r g m a x a (v . Lars Grline and Fabian Wirth D e f i n i t i o n 2.. Consider the functional G : ]~n x . IlY[I< R and all a 9 A.s. a). A continuous function u : Y2 + ]~ is said to be a viscosity solution of (3) if u is a viscosity supersolution and a viscosity subsolution of (3).LRIIx VII for all IIxH. v(x). Furthermore.c. standard techniques from optimal control (see e.) function u : $2 ~ ]~ (resp. D e ( x ) ) > 0 (resp. D e ( x ) ) < 0).a)) + G(x.t. a(v))dv ) .A + ]~ U {§ and the optimal value function v given by G~176 g(x(t). an upper semicontinuous (u. Du) = 0 x E ~ (resp. i.280 Fabio Camilli. we say t h a t a lower semicontinuous (l.a)II <.g. Given an open subset $2 of ll~" and a continuous function H : t9 • 1~ x R'~ + 1~.
A Regularization of Zubov's Equation implying (abbreviating G(t ) = G(x. T h e o r e m 3. Section 3] for details. Then v coincides with the restriction v}o of the function v f r o m (4). Chapter III]) shows that and v is a viscosity solution of (2). A s s u m e ( H I ) and (H2} and consider the unique viscosity solution v of (2) with v(0) = 0. Then (2) has a unique bounded and continuous viscosity solution v on ~'~ satisfying v(O) = O. Sections 4 and 5]. A s s u m e (H1).2 (i)]. continuous on ]Rn. = 0 if and only if x = O. This function coincides with v from (4). [3. In particulav the characterization Do = {x E 1Rn Iv(x) < 1} holds. Recalling t h a t v is locally bounded on ]Rn an easy application of the d y n a m i c p r o g r a m m i n g principle (5) (cp. We also obtain the following local version of this result. see [4. We end this section by stating several additional properties of v as proved in [4. For the proof see [4. and the continuity of v. Let 0 C 1Rn be an open set containing the origin. a ( r ) ) d r + G ( t ) v ( x ( t .and suboptimality principles. In particular the characterization :Do = {z E ~ " Iv(x) < 1} holds.1] Proposition (i) (ii) (iii) (iv) v(z) v(O) v is v(x) 1. which essentially follow from Soravia [16. T h e o r e m 2. T h e more difficult part is to obtain uniqueness of the solution. A s s u m e ( H I ) and (H2). T h e o r e m 1. Consider the system (1) and a function g : ]Rn x A + IR. Then the following statements hold. and let v : clO + ]R be a bounded and continuous function which is a viscosity solution of (2) on 0 and satisfies v(O) = 0 and v(x) = 1 f o r all z E 0 0 . x. Then < 1 if and only if x E :Do. a) ) v(z) = sup aEA 281 {10' G ( r ) g ( x ( r . a)) } (7) T h e next proposition shows the relation between :Do and v. since equation (2) exhibits a singularity at the origin. t. . a). x. In order to prove the following uniqueness result we use super. + 1 for x + Xo e cg:Do and for I1~11 ~ ~  We now turn to the relation between v and equation (2). Consider the system (1) and a function g : ]Rn x A ~ ~ such that ( H I ) and (He) are satisfied. T h e o r e m 3. Proposition 3. (H2).
g(y. (ii} If f(.. it is easy to see that in the situation of Figure 1 (showing one trajectory and the simplices surrounding the fixed point 0 in a twodimensional example) the piecewise linear function 9 with 1. a) = 1 = 9(xi). a)) = 1 implying (1 . c0:Do) > 0. Unfortunately. Lg > 0 uniformly in a 6 A. y E N the inequality Ig(x. In order to ensure convergence we will therefore have to use a regularization of (2).hg(xi. with constants LI. since also (8) has a singularity in 0 the fixed point argument used in [6] fails here and hence convergence is not guaranteed.v( o) < 0 for all xo E :Do \ {0}. a))v(xi + h f ( x i . then there exists a function g : ~ " r ]R satisfying (H2} such that the solution v of (2) is C ~ on a neighborhood of B. 4 Numerical solution A first approach to solve equation (2) is to directly a d o p t the first order numerical scheme from [6] to this equation.a)l < K max{llxlh Ilyll}'llx vii holds for some K > 0 and s > Ly/vr with a > 0 given by (HI}. all t > 0 and all a(. A) is convex for all x E ]~n and B C :Do satisfies dist(B. a)) + hg(xi. since for all nodes xi ~ 0 the value xi + h f ( x i . ) ) ) . of course. a) are uniformly Lipschitz continuous in R n. a)} (8) where 9 is continuous and affinely linear on each simplex in the grid and satisfies 9(0) = 0 (assuming. that 0 is a node of the grid) and 9(xi) = 1 for all xi E Ol2. hence 9(xi + h f ( x i . a) and g(. a) ) + hg(xi. and if there exists a neighborhood N of the origin such that for all x. (8).. Lars Griine and Fabian Wirth (i) The function v is a robust Lyapunov function for the system (1).h f ( x i . x i • O 9(z~)= O.4.x~=O satisfies (8).e.e. i. a) lies in a simplex with nodes xj ~ 0. T h e main idea in this is to change (2) in such a way that the "discount . then the function v is Lipschitz continuous in R n for all g with go > 0 from (H2) sufficiently large. a) )9(xi q.hg(xi..282 Fabio Camilli. Considering a bounded d o m a i n 1'2 and a simplicid grid F with edges xi covering cll2 this results in solving 9(xi) = ama~{(1 . In fact.) E . i. a ( . a ) . (iii) I l l ( x . As this situation m a y occur for arbitrarily fine grids indeed convergence is not guaranteed.
then the function v.(x) < 1} holds (iv) If f(. a) and g(. (9) The following proposition summarizes some properties of (9).a)[ > 91 > LI (11) (10) for all x r B(O. Using this gc we approximate (2) by aEA inf { .a)=max{g(x. Ilyll}' IIx . is uniformly Lipschitz on ]~n. 5) v. a) l < K max{llxll . To this end consider some parameter ~ > 0 and consider the function gc(x.D v ( x ) f ( x . a) + v(x)g.A Regularization of Zubov's Equation 283 Fig..e.. the factor g(x) in front of the zero order term v(x)) becomes strictly positive. Then for each ~ > 0 equation (9) has a unique continuous viscosity solution ve with the following properties. and thus the singularity disappears. P r o o f : Since the discount rate in (9) is strictly positive it follows by standard viscosity solution arguments [3. P r o p o s i t i o n 2. Let the assumptions of Theorem 1 hold and let v be the unique solution of (2) with v(0) = 0. a) . y E B(O. A situation of nonconvergence rate" (i. (x. a) . r/2) with C.g(y. a and r from (HI).e}. a)} = 0 x E ~n. 1. a) are uniformly Lipschitz on :Do (uniformly in A with Lipschitz constants L ! and La) and g is bounded on :Do and satisfies the inequalities Ig(x.yll for all x.(~) _ v(~) for all 9 e ~ (ii) vc "~ v uniformly in ~ n as E + 0 (iii} If s < go from (He}(ii) then the characterization :Do = {x E 1~" I v. Cr) and Ig(x.a). Chapter III] that there exists a unique . the analogous statements hold in the situation of Theorem 2.g(x. We state it in a global version on ~n.
x. t. a~). . a ) : = e x p ge(x(r.A {/0tG.)))d~a(")gO'(".a(r))dr (/0  + Gc(t)v. a~).a(v))dv[ <_ Lollx  yll ..x. a~)) = 6. Then for all x ~ :Do and all T > 0 there exists a 9 .).(. In order to see (ii) observe that the continuity of g and v implies t h a t for each 6 > 0 there exists e > 0 such that {x e ~ " l at(x.')) . (.(x(v. r. a.go +G(T)v(x(T.a). ~ (.~) we can conclude that _< <_ (a(. a) for some a 9 A} C {x 9 ~ n l v ( x ) <_ 6}. t. t. ~.'+'r Now let T _> 0 be the (unique) time with v(x(r. v. t . a) by (6) and g > 0 the stated inequality (i) follows.a(v))dv )..Hence by straightforward integration using the exponential stability and (10) we can estimate I /0 9.).y.a).~. ( x .x..(r)g(x(r.a)..(r))d. In order to see (iv) first note that (10) holds for gt for all e > 0 (with the convention go = g).)g 0. T] which immediately implies :Do = {= 9 ~ " I v. a. a~ ("))) d.a(r))dv /0' 9c(x(r.a. Lars Griine and Fabian Wirth solution vc which furthermore for all t > 0 satisfies the following d y n a m i c p r o g r a m m i n g principle vt(x) = sup with a6. ~. ~.4 such that G(x..(.a.(t) = G .x. Now fix 6 > 0 and consider the corresponding e > 0. To prove (iii) let e < go. Since 7 > 0 and x 9 ~'~ were arbitrary this shows (ii). ~..'.284 Fabio Camini.x. t..l.a).')e(~(..4 such t h a t v(x) <_ Z a(~.)a(~(. a) for all t 9 [0.a. "'.. a) > Gt (x.(x(t.(=) < 1}.a)) } (12) G.')) .at(.x.a t (. a) _> g(=.'.. (13) Since v satisfies the same principle (7) with G(z. a~) and at(r) = Gt (x. a) = Gt (x. a. x.')gO. Let x 9 ~'~ and pick some 7 > 0 and a control a7 9 . Abbreviating G(r) = G(x.(. aT)) < 6.
a)l <_ e 'g'. ( y .'.y. Y. a).t. Ig. Pick some x E Do \ B(0. a) E B(O.f is Lipschitz in x uniformly in a E A) for each m E 7)o \ B(0. a) := inf{t > 0 : x(t. a(r))dr] (14) . ~.r/2) for all (t6) For all t > 0 with x(s. t .a(v))dr r. a ) l < e'g' and using le a . Then by continuous dependence on the initial value (recall that .yll for all t > 0 and consequently sup aEA I// G.a) q~ B(O.(x.v. a). .h i < max{e=.a)lg(x(r.'.(~.b] it follows Ia.a)g(x(r. r) and some Y E N ( x ) .vll~ = ~t(L'")L~. a)g(x(T. ( y .I~V.a)l _ ~'~' _< ~tg. x.(~O'. a ) .A Regularization of Zubov's Equation 285 for all m.v.a).a) s E [0. a). (lz) _ ~'g'L~'L'Ii=V.1 ~ 1 7 Gr 6 _ sup af=A [a. I a .x.v. r) there exists a neighborhood J~f(x) such that :(t(x. t] we can estimate [a~ (x.g(x(r.r).eb}la.~).r) and x(t(y. y.x. a ) [ <_ L0llx . .z.~'.a('))g. y E B(O. r) and some L0 > 0 independent of ~ and a. implying in particular (15) for some suitable L1 > 0 and all Iv.~)lg(~(.~). t .~.. ~ B(O.Y]I. r / 2 ) ) .a(.a). t .C**'II~II + f0 ~176 ( ~ .a .t. Then for each 7 > 0 .y E B(O.t Now define T(x.(x..r) for all y E Af(x) and all a E A.O'('.. which also implies l a .r[2) and x(s.a). a(r)) .a("/)le" Lg e'L'll~. ~)  a. z.a) E B(O. a('))I a" e~ _<1 _< L~llx  vii x." ' <_L.e .a)a~(y.(V.t.a) E B(O.))d.(y)] <_ L~l]x .~). ( x .(m)  v.
(~:.)  I: z" Ce(y. Since ~ > 0 was arbitrary the Lipschitz property follows on :Do.a.r. aT) we can conclude that x(T) E B(0.(T+.~~176 r.~ (y) l < I: a.9 ( y ( r ) . y(T) E B(O.))g(y(T + r).))9(~(T+r).Here the first inequality follows by splitting up the integrals using the triangle inequality. Hence we can continue Iv=(~) .(y.286 Fabio Camilli.7". ~. ( " .(~. aT).y. r) and y(T) E B(O.)g(x(r. ~ . thus also on cl/)0 and consequently on the whole ~'~ since ve ~ 1 on lt('~ \ T)0.a~)g(x(r).. ~ ) l g ( ~ . aT) }. . ~. a ~ ( r ) ) l d r Z IG~(x.min{T(z.a. ~:. ~'. Lars GrSne and Fabian Wirth we find a7 E A such that I. r). r.(y.a~(r))dr teg'T eL'T LIlIX .r.a~(r))dr I a.T(y. y.)G.L:) + g 1 ~ since gl > LI.r.))dr < IG.. a) =: g* + ~0T < L : ' ~ ll~yll +Lll]x .r.  Lg L9 + LI ) flx .(T)) .r. a.(=) ..a~)lg(x(r).y(T)]] < e LIT.a~)Ge(y. a~). [Ix(T) . ~.Yll ~ I" G ~ ( y .~)l y < sup xE'Do.. x.a. a.aEA g(x.(x. + G.r.a~. G.a~)g(y(r). Abbreviating x(t) = x(t.(T + . Now fix some 7 > 0 and let T :..Yll Lj(gl . ~ ( r + .~. (r))d~" .~)#(~(~. (y) l  <_ I.).y) and y(t) = x(t.).(r+~))dr a~(~(T)..(x. and (15) for the second term.a~(r))drl. r). ~. ~] .~. and the third and fourth inequality follow from (16) and (17).(~(r).y]] <<g.the second followsby the triangle inequality for the first term and using x(T).a~(r))dv +a.r.
a) as long as #(z(t. and that a number of different iterative solvers for (18) are available.(x + hf(xi. a)} (18) where again vt is continuous and affinely linear on each simplex in the grid and satisfies ~. Hence although m a n y properties of v are preserved in this regularization. Note that in general the solution ve is not a robust Lyapunov function for the origin of (1) anymore. a))9.e. The grid was constructed adaptively using the techniques from [11] with a final number of about 20000 nodes. The unperturbed equation (i. Note that the adaptive gridding techniques from [11] also apply to this scheme. Here we compute it entirely in a neighborhood of 0.(0) = 0 and v.(xi.95. 5 A numerical example We illustrate our algorithm with a model adapted from [17]. some are nevertheless lost. i. a). [6. In order to explain this observation in a rigorous way a thorough analysis of the numerical error is currently under investigation. . In [17] simulations are made in order to estimate the domain of attraction of the locally stable equilibrium at the origin.hg. instead some "security factor" has to be added. Remark 2. z0. Consider & = (10/25 2~25) z + l l x l l ( ~ O 1 ) X + (azO~x~) where x = (xl. z0. We now apply the numerical scheme from [6] to (9). The numerical examples show good results also in the case where we cannot expect a globally Lipschitz continuous solution vc of (9). outside the region where the regularization is effective. along trajectories x(t. Thus we end up with 9~(xi) = ~ax{(1 .A Regularization of Zubov's Equation 287 Remark I. Note that due to numerical errors in the approximate solution it is not reasonable to take the "exact" sublevel sets ~ ( z ) < 1.11]. setting g(z. The domains shown in the figures are the sublevel sets ~ (z) < 0. Figure 2 shows the corresponding results obtained with the fully discrete scheme (18).6] yields that there exists a unique solution 9t converging to vc as h and the size of the simplices tends to 0. a)) + hg(xi. with a = 0) is introduced in order to explain the existence of turbulence in a fluid flow with Reynolds number R = 25 despite the stability of the linearization at the laminar solution. A straightforward modification of the arguments in [3. a) = [IzIIL e = 10 1~ h = 1/20.g. More precisely. we can only ensure decrease of v.(xi) = 1 for all xi E al2.x~) T E ~2. The main reason for this seems to be that in any case ve is locally Lipschitz on Do. a(t)) > ~. and in addition determine the effect of the perturbation term axlx2 for time varying perturbation with different ranges A.10. see e.e.
5. Appendix A in: M. Storey.D.'e.002 0. Bardi and I. J. Numerical solution of dynamic programming equations. c) A . Preprint 24/99. and F. Approximation of:D0 for a) A = {0}. M. Wu.JacobiBellman equations. and d) A = [3. 1997. Numerical determination of domains of attraction for electrical power systems using the method of Zubov. Birkh~iuser. M. H. J.O0 O. Physics o] Fluids 9 (1997). L. Analysis und efliziente Simulation dynamischer Systeme'. 1627. Fabio Camilli. Optimal Control and Viscosity Solutions of Hamilton. A Generalization of Zubov's method to perturbed systems. Falcone. b) A = [ . . Bardi and I. and F.0 O. Baggett and L.3] References 1. 4. Capuzzo Dolcetta. Hirsch. M. 1997.O0 [ul [ul e 0. 10431053. Griine. 2.[2. Abu Hassan and C.288 O. Boston. 2. F. Camilli. IEEE Trans. Control 34 (1981). Birkh~iuser. 3. DFGSchwerpunkt "Ergodentheorie. 1].~ d) "%%2o ~. Trefethen. ee: 191 [vl 0. Chiang. O. Lars Griine and Fabian Wirth 8. Wirth. Boston. 2].oe7 ~t o'oeT F i g . M.N. Lowdimensional Models of Subcritical Transition to Turbulence. Stability regions of nonlinear autonomous dynamical systems.~ e:oe7 o. Control 33 (1988).OO2 c) "e'~ 4. Auto. submitted. 371381. 6. 8 0 2 o.1 . Int. Optimal Control and Viscosity Solutions of HamiltonJacobiBellman equations.oo7 ~.S. Capuzzo Dolcetta.
N. Griine and F. Colonlus et al.M. 747755. 21 (1985). Differ. EQUADIFF 99. J. 323329. A. Lyapunov and their Application. On the estimation of asymptotic stability regions: State of the art and new proposals. 1964. A. 9. 20. 12. Nelepin and V. Trefethen. Driscoll.D. 17. Math.). A. (eds. 2000. P. 10.L.N. Auto. Genesio. Proceedings of the Mathematical Theory of Networks and Systems MTNS98. Berlin.B. Rev. W. Maximal Lyapunov functions and domains of attraction for autonomous nonlinear systems. Italy. Phys. R. Differ. S.A. 19. Acad. to appear. Soravia. SpringerVerlag. R.D. V. Control 3 0 (1985). 15. Zubov.B. Tartaglia. Varmelli and M. . L. Stability of Motion. and A. Berlin. L. An adaptive grid scheme for the discrete HamiltonJacobiBellman equation. Boston. Construction of the attraction region by Zubov's method. 871880. Padova.V. Hydrodynamic stability without eigenvalues. M. Sdr. 13. Vicino. Groningen. The structure of the level surfaces of a Lyapunov function.A Regularization of Zubov's Equation 289 7. Adv. 16. Sagastiz~bal. pages 353 . 6980. IEEE Trans. Wirth.356. Science 261 (1993). GroBmann. Numer. Paice and F. P. E 50 (1994). Gebhardt and S. Wilson. 4 (1999). 75 (1997). 14.W. 37053711. 275296. To appear. M. 1967. Robustness analysis of domains of attraction of nonlinear systems.E. Automatica. L.E. 18.. Wirth. A. h Equations of unbounded and degenerate control problems without uniqueness.A. Trefethen. R. Equations 3 (1967). Reddy and T. Optimality principles and representation formulas for viscosity solutions of HamiltonJacobi equations. N. Robustness of nonlinear systems subject to timevarying perturbations. GonzdJes and C. Differ. C. Falcone. Noordhoff.I. T. 1998. R. Wirth. A maximum time approach to the computation of robust domains of attraction.. Bajdaev. Birkh~iuser. Vidyasagar. 4550. 578584. 319337. Equ.C. Un algorithme pour la r~solution rapide d'~quations discr~tes de HamiltonJacobiBellman. Advances in Mathematical Systems Theory. 8. Kirin. Equations 17 (1982). Methods of A. 11. Hahn.A. F. 1 311 (1990). In F. Paris. Grfine. Chaos transition despite linear stability. Paice and F. Proe. Sci.
by means of an example. we see how the interpretation of solutions of systems with discontinuous righthand side may influence a stabilizability result for a system which does not admit a continuous stabilizing feedback law.so Duca degli Abruzzi. k : ]Rn ~ ]Rm. 24 10129 Torino. 0) = 0 and f is continuous.e.polito. f ( 0 . . called feedback law. 2 Ryan's necessary condition Let us consider a nonlinear control s y s t e m o f t h e form = f(x. We clarify in which sense Ryan's generalization of Brockett's condition to discontinuous stabilizability applies when solutions of the implemented system are intended in Filippov's sense. T h i s s y s t e m is said to be (locally) a s y m p t o t i c a l l y s t a b i l i z a b l e if there exists a function u = k(x). the origin is L y a p u n o v stable a n d a t t r a c t i v e for all C a r a t h ~ o d o r y solutions o f (2). = f(x. it A b s t r a c t . u E ]Rm. u) (1) where x E ]Rn. We want to see in which sense it is the generalization of B r o c k e t t ' s necessary condition ([4]) to discontinuous s t a b i l i z a b i l i t y when s o l u t i o n s are intended either in K r a s o v s k i i ' s or F i l i p p o v ' s sense. such t h a t the i m p l e m e n t e d s y s t e m x.A R e m a r k on Ryan's Generalization of Brockett's Condition to D i s c o n t i n u o u s Stabilizability Francesca Ceragioli Dipartimento di Matematica del Politecnico di Torino c. 1 Introduction It is a c o m m o n thought t h a t K r a s o v s k i i ' s and F i l i p p o v ' s concepts of solution (see [9. In order to s u p p o r t this i d e a u s u a l l y two p a p e r s are mentioned: [8] a n d [12]. Italy ceragiol@calvino. i. Moreover. Here we are interested in the result by Ryan.10]) are not the good ones in order to deal with discontinuous s t a b i lization p r o b l e m s of nonlinear systems. k(x)) (2) is (locally) a s y m p t o t i c a l l y s t a b l e at t h e origin.
Filippov's solutions of the differential equation = #(x) are the solutions of the differential inclusion ~ E Fg(x). On the other hand if k is a single valued discontinuous essentially bounded function it is possible to associate to it an upper semicontinuous setvalued m a p Fk by means of Filippov's operator F. Let us recall that if 9 is a locally essentially bounded vector field. compact and convex valued righthand side. Ceragioli Note that if k is discontinuous this definition m a y lose sense. First of all Ryan takes f continuous and such that A C_ l~ m convex ~ f(x. We say that system (1) is (locally) asymptotically stabilizable in Filippov's sense if there exists a locally essentially bounded function k : IRn + IRm such that the implemented system (2). by means of the Filippov operator F. the origin is Lyapunov stable and attractive for all Filippov solutions of (2).292 F.v(~)). A ) C f ( x . As admissible feedbacks. . its upper semicontinuity implies its continuity.(=):N N 6>0 ~(N)=0 where p is the usual Lebesgue measure in IRn and B(x. we can associates to 9 the setvalued m a p defined by F.6) is the sphere centered at x with radius 5. with solutions intended in Filippov's sense. i. so that the righthand side of (2) is also locally essentially bounded and its solutions can be intended in Filippov's sense.A) C_]Rn convex. (3) A condition equivalent to (3) is the following: for all A C R m c o I ( ~ . upper semicontinuous. We say that system (1) is (locally) asymptotically stabilizable in Ryan's sense if there exists a locally essentially bounded function k : IRn ~ lR~ such that 0 E Fk(O) and (5) with U(x) = Fk(x) is asymptotically stable at the origin. Ryan takes upper semicontinuous setvalued m a p s U with nonempty. In the following we suppose that k is locally essentially bounded.e. Ryan's implemented system is then the differential inclusion ~/(~. the origin is Lyapunov stable and attractive for all solutions of (5). i. (5) which has a nonempty. (4) The proof of the equivalence of these conditions easily follows from the definition of convex hull. is (locally) asymptotically stable at the origin. The notion of stabilizability adpoted by Ryan in [12] is different. coA).e. compact and convex values such that 0 E U(0). Note that in the case U is single valued.
Proof From (3) it follows that col(x. P r o p o s i t i o n 1 If f satisfies (3) and system (1) is stabilizable in Ryan's sense then it is also stabilizable in Filippov's sense. (6) This condition is satisfied by affineinput systems of the form: x=f(x)+G(x)u. Theorem 1) we get that F f ( x . A ) for all A C_ 1~m and Broekett's condition (B) does not hold then system (I) is not stabilizable in Filippov's sense (by means of a feedback law such that 0 E Fk(O)). We have the following proposition. The immediate important consequence of this result is that if a control system does not satisfy Brockett's condition (B). it stabilizes it also in Filippov's one. This means that every solution of Filippov's implemented system is also a solution of Ryan's implemented system and then if the feedback law k stabilizes system (1) in Ryan's sense. [] Ryan proves the following theorem which generalizes Brockett's one ([4]). then it is also stabilizable in Ryan's sense. Fk(x)). k(x)) C_ f ( x . coA). then (B) for each open neighbourhood Af ofO E R n. Finally we can conclude that a result analogous to Brockett's one for discontinuous stabilization in Filippov's sense holds if f satisfies the following condition for all A C_ R m cof(x. T h e o r e m 1 ([12]) Let f be continuous and satisfy condition (3). c o A ) C c o f ( x . using techniques analogous to those of Paden and Sastry ([11]. . c o A ) C c o f ( x . If system (1) is asymptotically stabilizable (in Ryan's sense). C o r o l l a r y 1 If f ( x . then a discontinuous stabilizing feedback in Ryan's sense does not exist. P r o p o s i t i o n 2 If f ( x . A ) for all A C I~m and system (1) is stabilizable in Filippov's sense (by means of a feedback law k such that 0 E Fk(O)). A) C f ( x . coA) and from this. A) = f ( x . f ( A f x IRm) contains an open neighbourhood of O.A Remark on Ryan's Condition 293 In the following proposition we show a connection between stabilizability in the sense of Ryan and in the sense of Filippov . In order to get a similar result for stabilizability in Filippov's sense we have to make a different assumption on f . The proof of this proposition is analogous to that of Proposition 1.
0). In many papers affineinput systems which do not admit continuous feedback laws are stabilized by means of discontinuous feedbacks in a sense that.0) i f ( x .(e . V. o) = e . 0) and with center on the yaxis. This system does not satisfy Brockett's condition then a continuous stabilizing feedback does not exist. We try to clarify this apparent contradiction by means of an example.y) = {2arctan/~ 0 if(x. Ceragioli The fact that for affine~input systems there is no difference in considering stabilizability in Filippov's sense or in Ryan's sense could be also easily proven directly by means of Theorem 1 in [11].294 F.w) E IR2 is the control. y) and (0. This system can be thought to represent a unicycle positioned on the plane at the point (z. y)plane passing through the points (x. ~(z. They introduce the following functions: Z= _v Z z#O. .6]. if C is the circle on the (z. 3 An Example Let us consider the system { ~ = cos Ou y.ed). y ) = (0.y) #(0. Corollary 1 is then in contradiction with these papers.o~ . sometimes seems to be that of of Filippov. v. Od(x." sin Ou 0 (T) where (z.2~. 0) E IR3 is the state and (u.y) and with orientation 0 with respect to the zaxis. We consider the feedback law proposed by Canudas de Wit and S0rdalen in [5. Moreover completely analogous results could be stated for locally bounded feedback and solutions of the implemented system intended in Krasovskii's sense. Od is the angle between the line tangent to C in P and the xaxis. though not always clear.
0) = . where 7 and k are positive constants. O) = b~(x. . The implemented system is discontinuous on the following surfaces: v= {z : We are interested in particular in the following subsets of C: ~+={zElaS: 9 ~". y ) = ~. V. 0) E IR3. y) ~(~. 0). where z = (x.'. ~r] for all (z. Note that (0.{Z ~ : ~ 3 z>0. 0) E F(u. ~x a ( x . y. 0" 0d}71" }. V. V.7 b l (x. V. Let us denote the implemented system by = a(z). #)T is the state.A Remark on Ryan's Condition 295 where n : IR + Z is such that a belongs to (~r. O)u(~. O) . Canudas de Wit and Scrdalen define the feedback law u(z. y>0. We consider Filippov solutions of this system. x#O. y. O). O. y. V. O=OdTr} : ~ < 0. w)(0. r~+V~arctan~ y a can be seen as the arclength of the circle C. A P B = ad B m X 2/~ b2 = cos 0(1 +/~2)x 2 sinS(1 + f12)x if y = 0 if y # 0 . 0)a(x.k~(~. y > 0.
0). 0.O) : (~/cosOdrOd.13] for definitions of these solutions). Acknowledgments. Z 7r g : . T h e Filippov setvalued m a p at (0.+Oa+Tr+ lim g(x. the feedback law does not stabilize the system in Filippov's sense. we consider Euler or sampling solutions (see [7. y.y. If we denote r = we have that (Irl is the radius o f t ) .0~~ g(z.y) = and O.e. Finally let us remark t h a t in m a n y other papers different strategies in order to stabilize system (7) have been proposed. Analogously if x < 0 and y > 0 are fixed. instead of Carath~odory solutions. 0 . y.y.O) = ('ycosOdrOd. ~ . g** (=. y. we have t h a t g+(x. ~) = ( .70d k~r) T.3.y.'yOddkTc) T g_+(~.y.y) = (72Y.y) : o~o~+~c lira g(z.296 F. 0) must then contain the convex hull of the following vectors: g++(y) = lim g+(x.kTr)T g(x. 0 . . y) = and O.y) = ( . 0. The same is still true if. These can be obtained by letting z + 0 + and z + 0 ..'yOd k~r) T . ~) = lim 0. discussions.~ . 0).k~) T. These strategies are not a priori in contradiction with R y a n ' s result because either the feedback is unbounded or solutions are not intended in Filippov's sense.'fsinOdrOd.TsinOdrOd.7Od d. Ceragioli ~v Let x > 0 and y > 0 be fixed.k~) r The convex hull of these vectors contains the vector (0.1.arbitrarily closed to (0.01img : ( ~ .respectively. x40+ x+0+ g++(y) = lim g_+(~. This means that all the points of the positive yaxis are equilibrium positions for the implemented system with solutions intended in Filippov's sense.y) = ( ~ 2 y .O) : (TcosOdrOd. I want to thank professor Bacciotti for m a n y helpful .t3].O) = (~/cosOdrOd. 0. 0) with any fixed y > 0.7 ~ y .+Oa~+ lim g(z. Note that there are points both on ~'+ and ~ ' .~ y . Let us now consider the point (0.7~r + kTr)T. Nevertheless it is possible to define conveniently the feedback law on the discontinuity surfaces in such a way that Carath~odory solutions exist and the implemented system is asymptotically stable. A m o n g these let us mention [2.( Y ) = .~/sinOdrOd. 77r + k~r)T.~[sinOdrOd. g+(Y) : x+O lim g+(x. i.
Cas34. Proceedings of 3rd European Control Conference. Millmarm and J. 1994. Rosier. 1998 8. 2735 2.D. 1983. 32. A. IEEE Robotics and Automation Magazine. 7381 12.W. Stability and Stabilization: Discontinuities and the Effect of Disturbances.A Remark on Ryan's Condition 297 References 1. No. IEEE Trans. 1996. M.4.l. Journal of Differential Equations.HAjek. Stern editors.. Filippov. O. Springer. Vol. NATO Science Series. On Brockett's condition for smooth stabilizability and its necessity in a context of nonsmooth feedback. Clarke. Clarke and R.R. Estimation and Control. 1979. O. France (IFAC.J. November 1992. Wolenski. Bordeaux. Kluwer Academic Publishers. Astolfi. R. Coron and L. E.H. O. 1988 10. 1992). Asymptotic stability and feedback stabilization. Vol. J.J. March 1995. Closed loop steering of unicyclelike vehicles via Lyapunov Techniques. Symp. Ledyaev.J. 171185 11. Aut. A. Scrdalen. Casalino. 30923097 3. vol. IEEE Transaction on Circuits and Systems. Exponential stabilization of a mobile robot. Canudas de Wit and O. Kluwer Academic Publishers. September 1995. in Differential Geometric Control Theory.H. in Nonlinear Analysis. Paden and S. No 11.P. A Relation between Continuous TimeVarying and Discontinuous Feedback Stabilization. 6784 9. Birkhanser. Systems and Control Letters. Rome. Bicchi and A. Brockett. Vol. on Nonlinear Control System Design. 149170. Control and Optim..S. Brockett.W. SIAM J.S. Yu. Balestrino. Proc. Differential Equations with Discontinuous Righthand Sides. 5761 6. A.J.F. Exponential Stabilization of Mobile Robots with Nonholonomic Constraints. R. 32. Journal of Mathematical Systems.H. R.. Contr. Astolfi. 17911797 7. Sastry. SCrdalen.M. 15971604 13. Nonsoomth Analysis and Control Theory. Sontag. Differential Equations and Control. 181191 5. R. Discontinuous control of nonholonomic systems. F. Stern and P. Sussmarm Eds. Ryan. 528. 1994.J. 27. Aicardi. F. January 1997. B. Series C: Mathematical and Physical Sciences. Canudas de Wit and O. Discontinuous Differential Equations I. 1. E. A. A Calculus for Computing Filippov's Differential Inclusion with Application to the Variable Structure Control of Robot Manipulators. No. II.J. Examples of Piecewise Smooth Stabilization of Driftless NL Systems with less Inputs than States. 37. Boston. G. 1999 . 3745 4.
Applications of Lie Algebroids in Mechanics and Control Theory
Jesfis ClementeGallardo 1 Faculty of Mathematical Sciences Department of Systems, Signals and Control University of Twente P.O. Box 217 7500 AE Enschede, The Netherlands
j esusOmailhost, math. utwente, nl
A b s t r a c t . We present some applications of the recently developed I:agrangian formalism for Lie algebroids, to mechanical systems. Examples treated are the rigid body with gravity, systems with symmetry and systems defined by an integrable Dirac structure. In this framework, we analyze the method of controlled Lagrangians for the asymptotic stabilization of mechanical systems with symmetry, which turns out to be naturally defined on the so called gauge algebroid.
1
Introduction
From the historical point of view, the original definition of Lie algebroids in the context of differential geometry is due to Pradines [13]. Roughly speaking, they are vector bundles with some particular properties that make them very similar to tangent bundles. Quite recently A Weinstein [16] proposed a generalization of the usual Lagrangian formalism (defined on tangent bundles) to include these objects. Later, E. Martlnez proposed a more geometrical treatment of the problem [11], which allowed him also to construct the Hamiltonian version [12]. The aim of this paper is simply to apply their results to some simple examples which are relevant in Mechanics and in Control Theory of mechanical systems. We focus our attention in two of them, namely the heavy top and mechanical systems with symmetry, which we develop in some detail. Finally, we introduce an alternative approach to the problem of Controlled Lagrangians in mechanical systems with symmetry, which might be generalized to other systems described by means of Lie algebroids. 2 The concept of Lie algebroids
First of all we need a definition of the object we are going to work on: D e f i n i t i o n 1. A L i e a l g e b r o l d o n a manifold M is a vector bundle E + M, in whose space of sections we define a Lie algebra structure (F(E), [., "]E),
300
Jesds ClementeGallardo
a n d a m a p p i n g p : E ~ T M which is a h o m o m o r p h i s m for this s t r u c t u r e in relation with the n a t u r a l Lie a l g e b r a s t r u c t u r e o f the set of vector fields (:~(M), [., "]TM). We have therefore:
p([r], ~]E)  [p(T]), P(~)]TM V r], ~ E F(E),
t h o u g h in the following we will o m i t , for simplicity, the s u b i n d e x of t h e commutator. Moreover, we can consider the action of t h e Lie bracket on the m o d u l e of sections of E :
[~,1~] = [u,5]+
(p(~)f)~ v0,~ ~ F(E),
Vf ~
C~(M)
If we t a k e c o o r d i n a t e s in the base m a n i f o l d {z i} and a basis of sections in the bundle {ea}, we can consider {(x i, A")} to be the c o o r d i n a t e s for E , w i t h respect to which we write the expression of the anchor m a p p i n g as
p(e~) = p~ O z i ,
a n d the expression of the Lie bracket in this base as
[eo,
i
a
e,]
=
T h e m a i n i d e a we have to keep in m i n d is t h a t a lie algebroid is a g e o m e t r i c a l object very s i m i l a r to a tangent bundle. T h e sections of the bundle E will play the role of vector fields (though we will need to use the anchor m a p p i n g when necessary). Can we define also analogues for o t h e r usual g e o m e t r i c a l objects as differential forms or the exterior differential as well? T h e answer is affirmative: * T h e analogue of one forms is easy to find, we take the sections o f t h e d u a l bundle E* ~ M . This definition allows us to consider an a c t i o n of sections of E on sections of E* as t h e analogue of the inner action of vector fields on differential forms. We will denote it by iv for a E F E . e T h e analogue of pforms is also easy to define: we t a k e s i m p l y the sections of the bundle (E*) ^v .+ M. 9 Finally, a basic piece of our c o n s t r u c t i o n will be the definition of t h e exterior differential d, as the o p e r a t o r which connects the a n a l o g u e of pforms with the analogue of p + 1forms. We define it as we do for the usual case, first the action on functions, a n d later the action on higher order forms:  For functions d : C~176 + AI(E) such t h a t :
(dr, a) = p(a)f
Vf e C~176
E r(E).
Applications of Lie Algebroids
301
 For higher orders forms we take the direct analogue of the usual definition d : AV(E) .+A v+l (E): p+l d0(O'l,..., trp+l) Z (   1 ) i + l p ( o ' i ) O ( o ' l , . . . , ~rl,... o'p+l) i=l {Z ( i<j
11'+j 8([r
ay], erl,... , F , , . . . , O j , . . . ap+l )
where by the symbol &i we mean that the corresponding section is omitted. With these definitions, we have extended the differential calculus on manifolds to the Lie algebroid case. We can also define an analogue for the Lie derivative by using the expression for Cartan's formula da  ia o d + doic,. One of the most interesting properties which can also be extended to the Lie algebroid framework, is that, as it happens with the cotangent bundle, the dual of any Lie algebroid is a Poisson manifold: T h e o r e m 1. Given a Lie algebvoid E, the dual bundle E* is a Poisson man
@ld.
Proof. It can be found in [16] (direct construction), or in [11] (more sophisticated definition). If we take a basis of sections of E*, as the dual of the basis {ca} of sections of E, and denote the corresponding coordinates as (x*, pa), the expression of the Poisson bracket above turns out to be:
{x;,~} = 0 {x~,,~} = p'. {,~,,,} = c~ a812~
3
The
Lagrangian
formalism
for Lie algebroids
In this section we will present the geometrical formalism of Lagrangian mechanics for Lie algebroids, as it has been presented in [16,11]. The main part of the construction in completely analogous to the well known framework for tangent bundles and Poisson manifolds, and hence, we will begin with a short s u m m a r y of this, mainly to fix notation.
3.1 G e o m e t r i c a l m e c h a n i c s in t h e u s u a l c a s e
L a g r a n g i a n f o r m a l i s m The natural framework for describing Lagrangian formalism is a tangent bundle T M . EulerLagrange equations are obtained
302
Jesds Clemente~Gallardo
from a variational principle (see [3] for a geometrical exposition of this construction and its extension to higher order Lagrangians), but they can also be formulated directly in a geometrical language by using three tools: the exterior differential d, the Liouville vector field A and the vertical endomorphism S (also known as almost tangent structure, see [8]). With these elements, we can formulate EulerLagrange formalism as follows. First we consider a Lagrangian L E C r176 M ) and a set of coordinates (x i, v i) (T in T M ; and define: 9 The e n e r g y f u n c t i o n EL = A L  L . The expression in coordinates reads
i OL EL=V~viL
9 The one form OL = S(dL). In coordinates: OL i OL = ~vi dX 9 The two form
V3 L =
dOL. The corresponding coordinate expression is
o)L 
02L d r j A d x i OxJ Ovi
02L " O~~v(dVJ A d x i
With these tools we can easily formulate geometrically EulerLagrange equations:
Proposition
1. EulerLagrange equations are the integral curves of the second order differential equations ( S O D E hereafter) vector fields F E ~ ( T M ) which verify the equation: irwL dEL (1)
3.2
G e o m e t r i c a l m e c h a n i c s o n Lie a l g e b r o i d s
The construction of geometrical mechanics in Lie algebroids has been recently carried out mainly by A. Weinstein [16] and E. Martinez [11,12]. Weinstein studied the problem in [16] for the first time and presented a Lagrangian framework for Lie algebroids and groupoids, while Martlnez developed a more geometrical construction for Lagrangian mechanics on Lie algebroids (recovering Weinstein's results) and extended it also to the Hamiltonian formalism. We will present now briefly both approaches, though for simplicity, we will use mainly Weinstein's construction in the examples we present in the following sections.
Applications of Lie Algebroids
303
Weinstein's c o n s t r u c t i o n W e i n s t e i n ' s c o n s t r u c t i o n is based in T h e o r e m 1.
T h e m a i n idea is to use this p r o p e r t y for c o n s t r u c t i n g an a n a l o g o u s p r o c e d u r e to t h a t used in the usual L a g r a n g i a n formalism. F i r s t of all we m u s t consider t h e L a g r a n g i a n of the s y s t e m . It will be a function defined on the Lie algebroid E , i.e. E E C~176 If we w a n t to use the Poisson bracket on E* we need to transfer it to E , a n d the n a t u r a l way of doing it is v i a the analogue of the Legendre t r a n s f o r m for Lie algebroids. As in the usual case, we define ~ ' L : E + E* a n d it reduces to t h e fiber derivative defined on E . In c o o r d i n a t e s we write:
OL
W i t h this we can define the energy function as we d i d in the usual case (EL = ~ L  L) and use the Poisson s t r u c t u r e on E* to define the d y n a m i c s :
dt = { x ' , E L } ;
~ ~
=
~'~,EL
(2)
T h e c o o r d i n a t e expression of these equations, whose d e d u c t i o n can be found in [16], is:
dt
(3)
x~ , cgL ~ cOL
(4)
d (cgL)
T h e first equation m u s t be considered as the algebroid a n a l o g u e o f the condition of a S O D E vector field. In this sense, Weinstein provides t h e concept of a d m i s s i b l e vectors and curves: D e f i n i t i o n 2. A t a n g e n t vector
v E T~,E is called a d m i s s i b l e if
Tr(u) = p(v).
A curve is called admissible when its t a n g e n t vectors are a d m i s s i b l e , a n d in this sense, a vector field in E is called a S O D E if its values are all a d m i s s i b l e vectors. Then, the solutions of E u l e r  L a g r a n g e equations will be curves on t h e algebroid whose t a n g e n t vectors are admissible. We will t r e a t m o r e carefully this point in the next section.
M a r t f n e z ' s a p p r o a c h T h e basic element of this f r a m e w o r k is t h e s u i t a b l e choice of the analogue of TTM, which is t h e n a t u r a l s e t t i n g for the s t u d y of geometric objects in usual l a g r a n g i a n mechanics as the d y n a m i c a l vector field, or s y m m e t r i e s of it.
304
Jestls ClementeGallardo
In this paper we will only introduce the objects, and we address the interested reader to the references above to find a stepbystep construction. T h e vector bundle we define is:
s
= {(a,b,v) E E x E x T E I r ( a ) = r(b),v E TaE, p(b) = Tar(v)}
(5)
This set is a vector bundle on E (as T T M vector space structure given by:
is a bundle on T M ) with the
A(a, bl, Vl) ~~(a, b2, v2) = (a, ~b 1 +/Jb2, )tvl t/~v2). Moreover, we can consider also a Lie algebra structure on the set of sections (see [11]). W i t h this Lie algebra structure and the projection onto the third factor pl(a,b, v) : v a,b E E v e T E we have the following result ([11]): Theorem
2. s
is a Lie algebroid.
In order to define the analogue of the EulerLagrange construction we saw above for the tangent bundle, we need three elements: the exterior differential for s (that we have already since we saw its definition for a general algebroid), the vertical endomorphism and the analogue of the Liouville vector field. Both elements require of the definition of vertical elements in I : E . Since E is a vector bundle, we can identify the vectors tangent to the fibers with elements of the fiber itself: given a E E we can define for any b E E the element b~ E Ta E as
bYaF = d F ( a + tb)]t=oVF e C~162
It is easy to understand thus t h a t we consider that the vertical elements in s are those with vanishing second factor, since because of the condition p(b) = T r ( v ) V(a, b, v) E s this means t h a t the vector v E T a E is actually a vertical vector on the vector bundle. T h e set of vertical sections will define a subbundle of s which will be denoted by Ver(E). Then we define the v e r t i c a l e n d o m o r p h i s m for Lie algebroids, as the m a p ping S : s + VerE such that S(a, b, v) = (a, O, b~). Analogously we will define the L i o u v i l l e s e c t i o n of s as A(a, b, v) = (a, 0, aa w) V(a, b, v) e L E .
In this context we can reread Weinstein's definition of admissible vectors (2). The set of admissible vectors (which provided the algebroid analogue of SODE's) will be now a subbundle of s defined by: A d m ( E ) = {(a, b, v) e f_.EIa = b}. We can also search for a basis of sections on s which allow us to write coordinate expressions of these objects. Mart~nez's choice is the following:
Applications of Lie Algebroids
305
(
a
)
v~(a) =
/
a, 0, 0  ~
(6)
according to the choice of coordinates (z i , A~, z ~, va) where (x i, Aa) identify the first and (x i, z a) the second element in the triplet (a, b, v) E s and the vector i~ 0 v~0 v = p~z ~ + OA,~. These sections play the role of {y~,} and {~6;~,} in the usual case at T T M 0 0 (with the obvious coordinate choice). With these assumptions, we can write the expression of SODE sections as F(x, A) A~..k'~+ f ~ ( x , A)V~, which under the anchor mapping pl goes into a vector field on T E pl(F)(x,A) i ~ O 4_/~(x ' O
=P~
~
"~) 0 , ~ 9
It is simple to see that in the simplest case, where the algebroid is actually a true tangent bundle, these sections are the usual vector fields related to second order differential equations. And we know from Weinstein results, that they must be also the solutions of EuleroLagrange equations in its Mgebroid version. Then we have to search for an EulerLagrange construction for Lie algebroid, whose solutions belong to Adm(E), such that they reduce to the usual EulerLagrange formalism in the case of the tangent bundle, and such that its coordinate expression coincides with Weinstein's result. The construction of such a formalism, presented in [11], is completely analogous to the usuM case. We define: 9 9 9 9 A one form OL = S(dL) The two form ~v : dOL. L The energy function EL : AL  L. And finally, we consider EulerLagrange equations as the SODE sections F which verify:
il't~L : dEL
(7)
The corresponding coordinate expression coincides with Weinstein's result (see [11]). The main advantage of this new formalism is the fact that the construction is a close analogue of the usual geometrical description of lagrangian mechanics. This will be useful for us in the following sections, where we will need to adapt some well known geometrical constructions of the usual formalism to make them work on some interesting systems defined on Lie algebroids.
306 4
Jesds ClementeGallardo Relevant examples of Lie algebroids
The aim of this section is to exhibit some interesting examples of Lie algebroids, which are relevant in the description of mechanical or even control systems. 4.1 The simplest examples
9 The first and simplest one is the case of the tangent bundle itself. In this case, the vector bundle E  T M , the base manifold is M, the Lie algebra structure is provided by the natural Lie algebra of vector fields, and the anchor mapping is the identity mapping. Of course, any mechanical system described in a tangent bundle enters trivially in the context of lie algebroids. 9 The second simplest example is the case of a Lie algebra. It can be considered as a fiber bundle over a point. Obviously, the anchor mapping must be zero, while the Lie algebra structure is the natural one. 9 The following example is an integrable subbundle of a tangent bundle D C T M . In this case, E  D, the base is M, the lie algebra structure is the restriction to D of the algebra structure of vector fields (this is why the subbundle must be integrable, in order to define a subalgebra), and the mapping p is the natural injection D ~ T M . The relevance of this example for mechanical system is obvious: any constrained system, with holonomic constraints, belongs to this class. 9 In a different level, we can find another interesting example when considering a Dirac structure. Following [6] one can define a Dirac structure as a subbundle of L C T M (~ T ' M , which must be maximally isotropic for the symmetric product: ((X,a),(Y, fl))=(X,#)+(r,a) VX, Y ETMa,gET'M
The definition of the Lie algebroid structure is thus as follows: we take L as our vector bundle on M, the natural projection onto T M as the anchor mapping, and the difficult part now is the definition of the Lie algebra structure. In [6], it is done as follows: [(X, a), (Y, g)] = ([X, Y ] , s  s  da(Y) )
Applications of Dirac structure are many, but we will mention particularly those related with control theory as are the implicit Hamiltonian systems [151.
4.2 More detailed examples of mechanical systems
We will present now a couple of more developed examples in which Lie algebroids help to describe in a suitable way mechanical systems whose usual
Applications of Lie Algebroids
307
description in t e r m s of t a n g e n t or c o t a n g e n t bundles t u r n s o u t to be m o r e involved or less elegant ( t h o u g h this m u s t be considered as a s u b j e c t i v e p o i n t of view, of course) t h a n their Lie a l g e b r o i d f o r m u l a t i o n .
T h e h e a v y t o p T h e first case t h a t we will consider is the h e a v y rigid body. It is well known [10,1] t h a t t h e m o r e direct f o r m u l a t i o n of t h e t o p d y n a m i c s is done in TSO(3) or T'SO(3) d e p e n d i n g on the f o r m a l i s m we use ( L a g r a n g i a n or H a m i l t o n i a n respectively). In those f o r m u l a t i o n s a fixed reference frame, external to the rigid b o d y is used, a n d the d y n a m i c s of t h e t o p has six degrees of freedom. T h e case w i t h o u t g r a v i t y is well known to be m o r e s u i t a b l y solved by a change in the reference f r a m e which passes to a set of axes centered in the body: the equations we o b t a i n then are the well known Euler e q u a t i o n s for the rigid body. T h e p r o b l e m with g r a v i t y is m o r e involved since we m u s t take into account the p o t e n t i a l t e r m in the lagrangian. O u r experience with the case without g r a v i t y suggests us to consider a set of axes which move with the body. T h e expression for the L a g r a n g i a n for such a case will be:
L = ~I~ Mg~/.
r~
~
E 50(3) 7 E S 2
(8)
where
r~M is the p o s i t i o n of t h e center of mass in the b o d y axes.
This L a g r a n g i a n is not a function defined on a t a n g e n t bundle, b u t on the manifold S ~ x 50(3). T h i s m a n i f o l d can be considered as an e x a m p l e of t h e Lie algebroid known as a c t i o n a l g e b r o i d , t a k i n g the following choices (see
[1@:
9 T h e bundle E is taken to be the trivial bundle S 2 x 5 0 ( 3 ) , as a bundle on the two dimensional sphere. 9 T h e Lie algebra s t r u c t u r e on the sections is n a t u r a l l y t a k e n from 50(3). A basic fact to be taken into account is the existence o f a n a t u r a l action : $ O ( 3 ) x S ~ ~ S 2 which can be used also for the e l e m e n t s o f t h e Lie a l g e b r a via the e x p o n e n t i a l m a p p i n g exp : 50(3) + S O ( 3 ) . 9 F i n a l l y the anchor m a p p i n g is o b t a i n e d from the r e p r e s e n t a t i o n o f the Lie algebra on the set o f vector fields defined by the a c t i o n above. We define the f u n d a m e n t a l vector fields p : S ~ • 50(3) + TS 2
= x7 =
d
V(expt
,s)l,=0 e T , S 2
Vs
S2
e50(3)
T h e well known r e p r e s e n t a t i o n of the algebra on the two d i m e n s i o n a l sphere gives us the desired result. T h e d y n a m i c s can be thus easily described on the Lie algebroid f r a m e w o r k . If we use Weinstein construction the corresponding E u l e r  L a g r a n g e e q u a t i o n s
308
Jesds ClementeGallardo
define the five dimensionM system: /~ = s i n r = cosr + cosr + sinr +w3
/10)1 = M g r o s i n O s i n r + (I~  Ia)w2w3 I2(o~ =  M gro sin Ocos r + (13  I1)Wl Wa ]'30~3  (I1  I2)W2Wl for (0, r coincide simplest obtained spherical coordinates on S 2 . It can be easily seen that this d y n a m i c s with other approaches to the problem as are [10,1], but it is the Lagrangian treatment known to us. In [11] an equivalent result is in the geometric approach.
Systems w i t h s y m m e t r y Let us consider now a system whose configuration space is a principal fiber bundle P ~ Q with structural group G. Let us suppose a system whose dynamics is described with a Lagrangian which is cyclic in the coordinate corresponding to the fiber of P . Such a system can be described also as a system on T P / G instead of T P , but if we want to use the usual Lagrangian formalism, we must choose the first setting in order to have a pure tangent bundle. T h e Lagrangian formalism on lie algebroids allows us to work with the system defined on the reduced tangent bundle, directly, without using objects defined on the whole bundle T P t h a t we must quotient in order to work with them. O f course, one m a y argue t h a t it is not necessary to take the Lie algebroid version, and that is true, but we consider a more natural and formally elegant framework for the description of this kind of systems this g a u g e a l g e b r o i d . A precise definition of the geometrical structure hidden behind the gauge algebroid can be found in [5]. In this short paper we will present only the direct definition, enumerating the elements which define the Lie algebroid structure:
9 The vector bundle is taken to be the aforementioned T P / G , a vector bundle over Q, whose sections are the Ginvariant vector fields. Local coordinates are given by those on Q, the corresponding velocities, and the elements of the Lie algebra ~ (see [9]). 9 The Lie algebra structure on the sections is trivially defined by using the natural Lie algebra structure of sections T P restricted to G  i n v a r i a n t vector fields. 9 And finally the anchor m a p p i n g is taken to be the differential of the projection which defines the original principal bundle which constitutes the configuration space. It is evident t h a t it goes properly to the quotient by G and it can be easily proved t h a t it is an h o m o m o r p h i s m for the Lie algebra structure we chose p = (Trr) l quotient : T P / G + T Q . We consider now a particular example which will be useful later: the well known example of the inverted p e n d u l u m on a cart. It is very well known the
Applications of Lie Algebroids usual Lagrangian description:
309
L(O,s,O,~) = I ( M + m)~ 2 + 1m1202 + mlcosO~O mglcosO,
(9)
where this Lagrangian is supposed to be defined on the full bundle T(S 1 • R) which is the usual setting for the Lagrangian description of this system. But after the discussion above, it is evident that we can also consider this cyclic Lagrangian as a function defined on the gauge algebroid T(S 1 • which is the most natural framework for the analysis of this system. It is a particularly simple example of Lie algebroid, since the corresponding Lie algebra on the set of sections is commutative and the anchor mapping is a projection. The expression for L will then be exactly the same as (9), but now it is considered as a function of the Lie algebroid, where the unnecessary dependence in the group variable (s for the cart) has been removed: L(0,0, k ) =
1 (M + m)~2 + ~ml 2'2 + mlcosO~O rnglcosO . 0
(10)
By using (2) we can write the dynamics associated to (10) as
dO dt d(ml20 + ml cos 0~) + ml sin 0~0 = rngl sin 0
a5 (M + + cos O0 = 0
5
The
controlled
Lagrangian
method
This method for the stabilization of systems with s y m m e t r y was recently proposed [2]. Given a system with symmetry, where the Lagrangian is cyclic in the group variable (as in the case we explained above), we suppose we control the group variables dynamics by fixing externally a generalized force u. For example, in the case of the inverted pendulum and the cart we have:
cos sin = d ( + mlm 0h) l+ ml 2 0~00 rngl sin 0 (M+r ) +r lcos00 ) = u
(11)
that can be also written in the form zi = 4~i(x, x, u) using the inverse of the mass matrix to transform the system above. The main idea of the method is to find a feedback law such that the dynamics above matches the dynamics given by a different Lagrangian (denoted by L~
We will consider thus the corresponding forms of maximal rank. There are two interesting geometrical interpretations of these equations [7. for a section a E I'E. preserved by the action of the dynamical section and such that the vertical subspaces o f / : E are lagrangian for the form and for iah dw (where ah is any horizontal section). with this new Lagrangian. We can divide it in two parts: the choice of the suitable Lagrangian to reproduce the dynamics (the so called m a t c h i n g t h e o r e m ) and the choice of the suitable control law for the stabilization. and that it does define a connection on the algebroid.14]). The problem arises now in the definition of the Lagrangian. in order to define. and in [11] it can be also found the definition of complete lifts of sections. The construction of the Lagrangian is based. Our aim is to present a different approach to the matching problem. which is preserved by the proposed dynamics and such that the vertical subspaces of T(TM) are Lagrangian for w and for ixh dw where Xh any horizontal vector for the connection defined by the dynamical vector field. a c E F ( / : E ) is the It is not difficult to check that this definition makes sense. its horizontal lift a h E l'(f_. a system where some interesting fixed points become (asymptotically) stable. since Poincar~ Lemma will depend on the cohomological .310 Jesfis ClementeGallardo in [2]) without explicit controls. a detailed algorithm is provided for the process in general systems with symmetry. will take place now on the b u n d l e / : E . roughly speaking. This is all we need to adapt Crampin's formalism to this setting.E) as: = 1 + r]) where U is the dynamical section in Adm(E) C s complete lift of a and a v E F ( E E ) is its vertical lift. We have seen above how we can define vertical sections on this bundle. Now we can extend the conditions for the existence of a lagrangian to the algebroid framework by borrowing them from the case above. does it exist a Lagrangian description of this dynamics? The set of conditions that must verify the system of differential equations are usually known as Helmholtz conditions (see [14] for a detailed exposition). on Poincar~ Lemma (see [7. In the rest of the paper we will focus our attention in the first part (the second part will be studied in a future paper). How can we generalize these approaches to the case of Lie algebroids? The definition of the connection given by the dynamics. In the reference above. The Inverse Problem can be formulated as follows: given a system of differential equations on some manifold M (we will consider now only the second order case). Crampin obtains Helmholtz equations as the expressions of a two form ~ of maximal rank. We follow now [7] and we consider that the dynamical vector field F defines a nonlinear connection (in general) on the bundle T T M . based in the Inverse Problem of Lagrangian dynamics. and then we can define.4].
we will have no fiber d e p e n d e n c e in the m a s s m a t r i x a n d hence the s y s t e m above b e c o m e s m u c h simpler. As an application. since in this case. y ~ } is the d u a l basis t o t h e c o r r e s p o n d i n g one on s defined by horizontal a n d vertical sections). they can be considered as a s y s t e m of differential equations defined on E = T(S 1 • ~ ) / R . If we consider the m a t r i x g as the mass m a t r i x of the new L a g r a n g i a n . b u t m i g h t be useful in t h e s t a b i l i z a t i o n p a r t of the process. Poincar$ L e m m a is valid.Applications of Lie Algebroids 311 properties of the exterior differential o f the algebroid.e. If we a p p l y the m e t h o d s above. from which we o b t a i n two free c o n s t a n t s t h a t we have chosen t r i v i a l l y in this case. k} a n d i = 0. These equations m u s t be also considered as e q u a t i o n s for t h e control law. T h e conditions on the two form a n d its p r o p e r t i e s . such t h a t : d r g ~ + ~g~g~ + ~g~gp~ = 0 1 o ~ "y 0ga# _ 0ga3. we o b t a i n a kind of Helmholtz e q u a t i o n s s u i t a b l y quotiented. we will consider t h e e q u a t i o n s (11). if we consider ~ = g ~ # X ~ A y # (where { X ~. Then. If we want an equivalent expression for the m a s s m a t r i x (the s a m e f o r m a l dependence with respect to the u n k n o w n connection) the s o l u t i o n is as follows . are t h a t it is a s y m m e t r i c non d e g e n e r a t e tensor. the analysis m u s t be carried o u t case by case. Moreover. If we consider a feedback law for t h e controls. OA'~ 8A# 1 o~ ~ 1 0r )6 where the Greek index runs on the set {0. we can simplify the s y s t e m by searching for solutions whose m a s s m a t r i x coincides with t h e m a s s m a t r i x proposed in [2]. we consider the c o n n e c t i o n associated to this d y n a m i c s . which is hidden in the section c o m p o n e n t s a n d in ~ . T h e resulting s y s t e m t u r n s o u t to be a first order differential e q u a t i o n for the control law a n d t h e connection of the m a s s m a t r i x . and i m p o s e the c o n d i t i o n s on the h o r i z o n t a l subspaces. i. which are not known in general. search for a two form defined on s such t h a t it is preserved by it. where only a connection one form k = k(O)dO on t h e a l g e b r o i d in unknown (the connection t h e y choose is a c t u a l l y a connection on t h e g a u g e algebroid).
(m + 2M)er 2 + rn(12mM + o"~')cos(20) \ 12rn2M.mcos(20) ) / I exp 1+ 212rn~M cos(20)) cos aa(m2M+m20 sin 0 12m2M . and J. Carinena and C. the Lagrangian we obtain is again a function of the gauge algebroid. Control.(m + 2 M ) a 2 + m(12mM + 0. . Rev. in Math. F. Carifiena for their support and m a n y illuminating discussions. Leonard. E. but. and that the use of this method on Lie algebroids may be useful. van der Schaft.. Phys. 3(4):379401.E.(rn + 2M)a 2 + m(12mM + a 2) cos(20) It is simple to verify that the corresponding mass matrix is non degenerate.J. Cambridge University Press. A better solution for this example. Marsden. Spinning tops. Controlled lagrangians and the stabilization of mechanical systems I. The objective of this small example was to show how the procedure works. and hence cyclic in the group variable (s in our example). IEEE Trans. To appear.312 Jestis ClementeGallardo (see [2] for the notation): (0) = 12mv/m + 2M . Evidently. Acknowledgements: The author wishes to thank Laboratoire des Signaux et Syst~mes where this work was begun. As an advantage of working on the Lie algebroid directly.M.rn cos(20) \t2m M  u )cos(20)] ) exp ( X/12m2M . A.2)cos(20) ] v~lm(m + M)gar + 2M . 2. M. 1991.m cos(20) v/'2M(M + m ) a ( exp ( x/2MlrncosO ~x (0) = arctanh ~ a x / m ~ 2M~ c~s (20) ] x/m + 2 M . as well as the extension to other interesting cases is our following objective. 3. He is also deeply indebted to F. it must be considered just as a preliminary result. Bloch. Auto. References 1. LamnabhiLagarrigue. A. 1996. this is a control law much more complicated than that obtained in [2].E. L6pez. Jose F. Geometric study of hamilton's variational principle. Audin. N. Martfnez and J.
V. A:Math. Soc. Lagrangian mechanics on fie algebroids. M. Control and Optimization. London Math. Preprint. Dodonov and V. U. Dirac manifolds.J. 263:907910. E.. 5965. Universidad de Zaragoza. 7. Weinstein.Applications of Lie Algebroids 313 4.J. Acad. 15. M. Integral systems. On the inverse problem of Lagrangian dynamics. Lie groupoids and Lie aigebroids in differential geometry. June 1990. Marsden and T. Courant. Carifiena and E. van der Schaft . and A. K. J. Mackenzie.. 14. Ann. Lagrangian mechanics and groupoids. AMS. Cannas da Silva. T. Grifone. Part II: Symmetries and Algebraic structures in Physics. second edition. Sci. Nova Science. 1992 5. XXI I(1):287334. Ratiu. Cambridge University Press. K. 8. 1981. E. R. 10. 1999. Universidad de Zaragoza. Weinstein. 11. A. Martinez. Structure presque tangente et connexions I. Trans. 1966. Introduction to Mechanics and Symmetry. J. Paris Sgr. Hamiltonian mechanics on fie algebroids. 37(1). Man'ko. 1996. . Fields Institut Com. Lecture Notes Series 124. R. 1972. Phys. 13. J.M. Relations entre propri~t6s locales et globales. Solid State physics and theory of Phase transitions.C. 5491. J. Th6orie de Lie pour les groupoi'des diff6rentiables. 6. Marthaez Generalized Jacobi equation and Inverse Problem in Classical Mechanics In V. Pradines.E. 1998. F.14 25672575. Berkeley. editors. 319(2):631661. S I A M J.. On representations and integrability of mathematical structures in energyconserving physical systems. 1978. Lectures on Geometric Models for Noncommutative algebras. Santilli. 1994. J. Hartshorn. Crampin. Fourier. Preprint. 1999. 1987. A. Dalsmo and A. 1999. Foundations of Theoretical Mechanics L SpringerVerlag. I Math. Martinez. C. 9. pages 207231. of the AMS. 16. 12. p. and Gen. Inst.
Italy bicchi@ing.e. which exhibits a local separation property. (1) where x E X . 1 1. Italy cont ice@sssup. an open subset of ~'~ c o n t a i n i n g the origin x = 0.1 Introduction Main Results In this p a p e r . 9 a local s e p a r a t i o n p r o p e r t y is achieved for the considered class o f nonlinear a n a l y t i c system. it Robotics Interdepartmental Research Center "E.30]. u E Nm is the control i n p u t vector. y E ~P is the o u t p u t vector. . i. This paper presents a novel nonlinear observer.Observer Design for Locally Observable Analytic Systems: Convergence and Separation Property F a b i o Conticelli 1 a n d A n t o n i o Bicchi ~ 1 Scuola Superiore Sant'Anna 56126 Pisa. is t h e s t a t e vector. T h e m a i n c o n t r i b u t i o n of this p a p e r is twofold: 9 a novel local nonlinear observer is presented. Piaggio" University of Pisa Via Diotisalvi. it A b s t r a c t . t h e designed observer permits to achieve local practical stability of the closedloop system. if there exists a stabilizing static state feedback. 0) = 0. locally observable s y s t e m s in the sense o f [27. The observer requires only that the nonlinear system must be locally observable for the considered real analytic input function. u) y(x) = h(x) . In fact. which ensures p r a c t i c a l stab i l i t y of t h e trivial e q u i l i b r i u m in t h e observation error d y n a m i c s . if the real state has been substituted with the current estimated one. 2 56126 Pisa. T h e vector field f ( x . we consider general n o n l i n e a r s y s t e m s of the form: = f ( x . and f(0. u) and the o u t p u t m a p h ( x ) are a s s u m e d real a n a l y t i c in the following. unipi.
Even if the proposed analysis is local.19] assuming that particular differentialgeometric conditions on the system vector fields are verified. We stress that the techniques proposed in [22. The first step is to write the input afline nonlinear system in a socalled normal observation form. Transformations. Our approach ensures local practical output stabilization of the trivial equilibrium. has been proposed in [30]. by using the time derivatives of the input. Practical stability is guaranteed since the introduced persistent perturbations can be made arbitrarily small. the observables of order higher than one are estimated in the observer design by using highpass filters. However. the simulation results show that the region of attraction has a noticeable extension. Gauthier et al. Early results on the observer design of bilinear systems without bad inputs are reported in [29]. and not uniformly locally observable. Rank conditions under which the dynamics of the observation error is linear. The paper is organized as follows. The considered system is driftless. 3 presents a local separation property. In Sect. are also investigated in [27]. which derives naturally from the application of the proposed nonlinear observer. In Sect.10]. the original system can be transformed into the observer canonical form.316 Fabio Conticelli and Antonio Bicchi The main idea in the observer design is the use of higherorder output time derivatives (henceforth called "observables") taken from the observability space associated to (1) in the assumption that the rank condition on the observability matrix is satisfied for the considered real analytic input function. Results on bilinear observers are presented in [5. this form requires . are the system immersion [8] which permits to obtain a bilinear system if the observation space is finite dimensional. Sect. Extension of the Luenberger filter in a nonlinear setting. Finally an example is reported.e.3] generalized the results in the case of inputaffine nonlinear systems without bad inputs and applied the approach to biological reactors. 1. Connections with strong observability under piecewise constant inputs are also highlight. In particular if there exists a stabilizing state feedback. In Sect. [9. Then.17. 5 the major contribution of the paper is summarized and future investigations are outlined. but meaningful. which have been proposed in the literature. 4 the proposed framework has been applied to a simple. example. 2. the equilibrium with the estimated state feedback remains locally practically stable. i. local observability of analytic nonlinear systems is briefly reminded. Moreover.1] can not be applied to the presented example. and the linearization by means output injection [16. a local separation property for the considered class of nonlinear system is proven.2 Related Work A first approach to design an observer is to transform the original nonlinear system into another one for which the design is known.
the approach presented in [1] cannot be applied to the presented example. v ~ _ l ] T = [ u T . . and global HSdel conditions on appropriate functions.t h output time derivative. A high gain observer which estimates the output derivatives combined with a globally bounded state feedback control law permits to obtain semiglobal stabilization by output feedback [7. The observable of order i = 0 . . Highgain techniques have been applied in the field of nonlinear observers. v0) = 0O0(x) f ( x . stronger conditions on the admissible inputs have to be assumed.14. Even if the results obtained concern stability analysis in the large. . and in the case of relative degree equal to the dimension of the state space n. . fullrankness of the observability matrix. In the recent reference [1] the authors employ a separation principle of a certain class of nonlinear systems. . defined as Co(x) 9 (x. v i .1 ) . and also that a diffeomorphism can be constructed using the Lie derivatives of the output along the drift nonlinear term..26]. These functions are defined recursively as O0(x) = h(x) Ol (x.20. 2 The Class of Systems Let us associate to the nonlinear system (1) the following extended output map45 : ~ " • hup "~ 4 ~(t~+l)p.. l p is r v 0 . the asymptotic stability of the equilibrium. Results on the normal observation form have been provided also by Tsinias [25.v) = r r v0) vo. . .vt~_~) ) (2) where v = [ v T .23] in case of uniformly observable inputaffine nonlinear systems [22]. .e. the input vector and its time derivatives up to order lp 1). in case of relative degree less than n.1)T] T 6 ~ r "~ is the extended input vector (i. . . Early results are due to Tornamb~ [24]. . the region of attraction. As we will show in the following..e. . and lp is an integer such that (lp + 1)p > n. . . and trajectories. it is shown the global asymptotic convergence of the estimated state..22. . the i . In [6] the authors consider single i n p u t single output inputaffine nonlinear systems. v0) 0x (3) . . u(lp . which proposed an approach based on highgain approximate cancellation of the nonlinearity. showing that with the estimated state feedback it is possible the performance recovery of the real state feedback. i.Observer Design for Locally Observable Analytic Systems 317 that the trivial input is an universal input [2] for the system. the class of considered systems is quite restricted.
~. given a point (z ~ x~ z ~ = [G OT vOT] T. v E ~m is the control input vector.(x. . . and ~ E Nn. v) E X x ~ m is the extended state vector.r zt. v) in an open neighborhood X ~ x V ~ of X x ~ . = ~ ( x . v 0 . k : lp [. Defined J ( x . (4) V(x.. m. ( ~ . v0) V0 ~ .318 Fabio Conticelli and Antonio Bicchi . defined as z0 . _ l )  0 r0 x f ( x . . We now give the following motivation of the introduced assumption. . this matrix is referred to as the Extended Out put Jacobian (EOJ) associated to the nonlinear system (1) in the following. ". ~ = [~T. l p . v) . k = 0 . Ct. z ~ + l . . . If the assumption 1 is satisfied. .+tp m] x ~n _. .[~T vT]T E ~p(tp+l)p. v0) t. ) / ] / .2 0r + ~ j=0 0vj vj+l A s s u m p t i o n 1 There exists an integer lp such that (lp + 1)p >_ n and the map 4~(x.Ct. (5) where zk E ~P.x) = 0 ..+ ~(tp+l)p. (Z) where z . Consider the extended nonlinear system associated to (1): = f(x.1 . . v ) = [ ~ } . . v t . based on the Implicit Function Theorem. . 2 l p . . . v ) ) 0x = n . . such that (x ~ v ~ E X ~ x V ~ there exist n observables taken from the (lp + 1) p . ~ 2 t .o Vl (6) ~'l~i b. . zk E ~ m . v) satisfies the rank condition: rank \ fOe(x. and ~ E ~(tp+l)p is the extended output vector. At any time instant t > t o > 0 the flow of the above nonlinear system satisfies the equation F(~. .~T] e ~(t~+l)p. Let us introduce the m a p F : ~[(tp+x)~.~) = ~1 . where (x.r F(~. .. . . and x E X.
and v = [u u . T]. being x C ~n the state vector.~T (gWT Of OW + w + 05 where w ( x .f(x.. t E [to. t) E ~n• defined as: dhl(x) A/'dhl (x) Q(x. . Remark 1. u(t)) = f(x. The nonlinear system can be viewed as a timevarying analytic system without input . 71 if and only if there exist a neighborhood ~ o of xO ' and an ptuple of integers ( k l . . t _> to > 0 of the considered nonlinear system (1) is real analytic.. T]. (10) .1 . g(z)) = 0. v) E ~n x ~ . y(n1)]T = r v) E ~". namely A ~ = I ~ x V ~ ofz ~ X ~ o f x ~ being I ~ an open neighborhood of~ ~ an unique map g : A ~ + X ~ with g(z) = x.t) are .30]. . . such that F(z. w (x.18]. al analytic timevarying functions. . and N W ~. t ) ... the observability matrix Q(x. w p ( x . then x ~ can be uniquely distinguished in a small neighborhood. such that: 9 kl >_ks>_kp > 0 and ~i=1 ki = n.1 the extended input vector. there exist two open neighborhoods. The class of nonlinear systems which are locally observable at x ~ in the interval [to. . . P (8) (9) 9 defined the differential operator: N Ow = w . T] is determined by the following proposition [27. uniform observability is a sufficient condition for our observer design. T > to _> 0. given the map [y y .t) = A/(k''lidhl (x) dhr(x) / . if given the output function y(t). The system (8) is locally observable at x ~ in the interval [to. t) y(x) = h(x) We recall the following definition [13. The system (8) is locally observable at x ~ in the interval [to. Thus. .u(n2)] T E ~ n . T h e o r e m 1. t ) ) . .27]. called the observability indices. In the case of an uniformly observable SISO (Single Input Single Output) nonlinear system [22. Assume that the input function u(t). x ~ E X. D e f i n i t i o n 1. t ) = ( w l ( x .kp).Observer Design for Locally Observable Analytic Systems 319 ones in r v) which are linear independent in X ~ • V ~ By the Implicit Function Theorem.30. the Assumption 1 is satisfied for each (x.
g(z)) = 0. This is in fact the unique implicit function g(z) = x. then the system is local weakly observable at x ~ E X ~ if and only if the observability rank condition is satisfied [11]. given the o u t p u t y(t). Vice versa. If Assumption 1 holds. by using the Implicit Function T h e o r e m . Hence. ~ ) y ( x ) = h(x) (11) (12) with ~ constant. T]. The concept of local observability at x ~ in the time interval [to. T]. . for every real analytic input function. Vff of interest. for some interval T . T > to with the assumption of real analytic input can be related to the observability of nonlinear systems under piecewise constant input. t _> to _> 0. see also [13. t _> to _> 0 which undergoes the equation F(z. We are now in the position to determine the class of nonlinear analytic systems which satisfies Assumption 1. p the T h e proof of the above proposition uses results of linear timevarying systems [21]. strong observability implies that dim (dO)(x ~ = n. . . T h e o r e m 2. T].to > 0 sufficiently small. = 1 . _ Proof. 7"]. for some T > to > _O. then lp = kt . T]. being dhi = ~ exact differentials associated to the output. v).320 Fabio Conticelli and Antonio Bicchi is nonsingularVx E X . and of perturbed differential equations [12]. if the nonlinear system is weakly controllable. T]. the system (8) is locally observable at every x ~ E X ~ in the interval [to. and t E [t0. where kl is the higher observability index. If a system is strongly observable at x ~ E X ~ then every constant input function distinguishes between nearby states. which imply that every C ~~ function (and in particular analytic) also distinguishes (see [28]). If this is the case.27]. for some interval T . the interval by choosing Assumption if the system (8) is locally observable at every x ~ E X ~ in [to. Notice that since the rows of the observability m a t r i x q ( x .to > 0 sufficiently small. t) are differentials which appear as rows in the E O J m a t r i x J ( x . Assumption 1 holds if and only if the system (8) is locally observable at every x ~ E X ~ in the interval [to. x) = 0. there exists an unique function x(t). In the analytic case. Hence if the nonlinear system(1) is strongly observable at every x ~ E X ~ then there exists T > to such t h a t it is locally . is locally weakly observable [11]. Assume that the input function u(t). . such t h a t F(z. t > to >_ 0 of the analytic nonlinear system (1) is real analytic. there exist the observability indices which satisfy the conditions of T h e o r e m 1 in [to. for all ~ of interest. then 1 holds by construction.1. after a possible reordering of the indices of the output variables. We remind that [28] the nonlinear system (1) is strongly observable at x ~ if the a u t o n o m o u s system = f~(x) = f ( x .
~ f(x. in virtue of Assumption 1. Consider the observer ~ = f ( i . T] under real analytic input.th output derivative. is asymptotically stable. . T h e o r e m 4.. This is a straightforward consequence of the above discussion and Theorem 2. Denote with s the Laplace variable.. C 1. then there exists a function "ff : X x ~t~ m r ~m of class.v) . with J+ = (jT j ) . if Assumption 2 holds.v) r i=l. and T is a small positive constant. 3 A Local Separation Property Assume that there exists a static state feedback which ensures the local asymptotic stability of the trivial equilibrium.lp the estimation of the i . for the proof. Assume that the system (1) is weakly controllable and analytic. . such that V(x) is ov positive definite. Y(s) the output Laplace transform. dim (dO)(x ~ = n. and . at tp least.Observer Design for Locally Observable Analytic Systems 321 observable at x ~ in the interval [to. then the system (1). V~ of interest. such that the equilibrium (x. under real analytic input functions. If the system is strongly observable at every x ~ E X ~ (under piecewise constant inputs}.v) = (q+ /of ~x( 8i x. (6). there exist two functions a : X + ~m.4)(5. and V : X + ~ both. and. v0) V0 = Vl (13) ~%1 = ~(x..lp r (l+Ts). v) = (0. We remind the following result. T h e o r e m 3. i. Consider the extended nonlinear system in Eq. of class C 1. satisfies Assumption 1..eta] is the estimated extended output vector. We can state the following. A s s u m p t i o n 2 Consider the nonlinear system (1).18. a(x)) < 0 in an open neighborhood of the origin x~O. v) (~ . 0) of the closedloop system /c = f(x. at least.Y(s)' . Proof. r T.vo)))j+(~. Indicate with Q a positive definite matrix. TT [yT. see for example [4. .e. v)) (14) P(~.i j T the left pseudoinverse of the EOJ. . vo) + P(x.15]. being r E i = 1 ..
lp is locally practically stable.~(~.. the equilibrium(x. t > o. T h e o r e m S .Ip is the T introduced persistent perturbation due to the estimated observables. .x . ' ) ] T I I < e. such that iflle(0)ll < 61.t~ < c.~T(0) xT(o)]TII < 8x. v)) P(~.( l + T s ) i Y(s). v)e* . vo) + P(~.x = 0 of the observation error dynamics deriving from the observer (14} is locally practically stable. . and T which depends on K. and the solutions of (15) satisfy the condition II[xr(t. eT . and f(x). t > 0 We begin with the following Lemmas. 0 . v) Si r . the equilibrium e = ic . v) (( .o.. t > 0. v)) .~. O. . and 2 hol~ and T ~ chosen sufficiently small.. g(x) are smooth vector fields. and T which depends on K. u) + P(i. Then Ve > O.e.x T ( t . . K > O. for example. vo))) J+(~. (15) y(x) = h(x) x = f(~. Assume that the origin of x = f(x) is a locally asymptotically stable equilibrium. If Assumptions 1. where e* = [0T. .x ) = ( 0 . 0 ) o f t h e closed loop system x = f ( x . where x E ~'~. L e m m a 1. For the proof see. t > t o ___0. t > o . ei = y(i) . . t ~ u) o f x = f(x) + g(x)u(t) satisfies the condition: IIx(t..v) = (Q + ( ~ x ( ~ . v) {~ . there exist gl > 0 and K > 0 such that if llx(O)l I < 81 and Ilu(t)ll < K. Consider the nonlinear system (1).~(i. the solution x(t.c). ~ T ( t .r i " i. L e m m a 2. u(t) E W~. C)ll < . Vee > 0 there exist 81 > O. i.. e*) . Consider the nonlinear system x = f(x) + g(x)u(t). then IIr < K. i = 1. . and the observation error satisfies the condition lie(t. K > O. O . The main result of this paper is the following. v0) Vo : vl 9 . .. t > t o >_ O. [18].e. r = O. i. Ve > 0 there exist 61 > O.etTp] E ~(G+I)p. then IIc(t)ll < K. if Assumption 1 holds. and T is chosen sufficiently small. such that if II[XT(0).P(i.322 Fabio Conticelli and Antonio Bicclfi The above equation can be rewritten as = f(i.
Since the highpass filter ~ is Bounded Input Bounded Output (BIBO). choose T < min{~. e) and p~(v. the Laplace transform of el results: ) . e ) .~k+l (8) s y(k)(0+) S = T Ts~ "u (I + (1 + T"~ (I + s"" + Ts~ ' ~k(.) (19) . which means that Ve* > 0. and (with an abuse of notation) Y(s) s Y ( s ) . Vt >_ 0. implies ~ Ily(0+)ll exp(7~) < e*. In fact. i.pf(v0.t'(~ e l ( s ) = ~'(s) .e. it follows that: r (8) = s (1 + Ts) Ck (8) . ~T}. where Y(k+l)(s) denotes the Laplace transform of the (k + 1 ) . then llel(t)]l < ~e" ~ TMy + e* < 7.x results + (pf(v0. . e)) = . lim pc(v. there exists My > 0 such that H)'!(t)H < My. and the functions pf(v0. ( v .Observer Design for Locally Observable Analytic Systems 323 Proof. Hence limT~ol]el(t)[[ = 0. Hence k/~ > 0. Y(s) = ~ f ~ r and dices set ( 1 .0 with their first order partial derivatives.P ( i . fixed e* < 7.y(0 +) indicates the output derivative Laplace transform. such that T < 5. Since Ck+l(S) s Ck(s) = s Y (k) ( s ) . satisfy the conditions lim p f ( v 0 . where k is chosen in the open in.8~k(8)) and ~k'll(S) : y(kl1)(s) .+0 Ilell _.0. Vt > 0.~ y ( 0 +) exp(7~).O[[ek(t)[[ : 0. Vt > 0. .e) e+0 Ilell _ 0 (17) We now prove by induction on the order of the output derivatives that the error e* can be reduced to an arbitrarily small perturbation if the constant T is sufficiently small. (14). there exists (~ > 0. v)e* + P ( i . . Vt > 0.~ " Y( s ) = T y . (l + Ts) . e) vanish at e .s ek (s) = Y(k+l)(s)+y(~)(0+)s e~(s).q e .th output derivative. e l ( t ) = T > . v) p . Assume that limT.0.  1 (y(k+1)(s) (18) + y<k)(o+) . e) where e* is the introduced perturbation due to the observables estimation.~ T Y ( s 1IT. . l p ) . Vt > 0. e) + P(~. f ( t ) . v) pr e) . the dynamics of the observation error e = i .k+. 1 t Let us consider the term y y ( 0 + ) exp(~) in the output derivative error e l ( t ) . where ~. By using Eq. it holds: limT~0 ~ Hy(0+)]l e x p ( ~ ) .' where Y(s) denotes the output Laplace transform./(t) is the output derivative filtered by 1+~'7. Vt > 0.
Let us consider the nonlinear system in Eq.__/Of . v) y(x) = h(x) . k : 1.P ( ~ .P(~. there exists an open neighborhood of the trivial equilibrium in which V is negative definite. Let us denote.. v 0 ) ) . e ) (21) As first step . w = [xT vT] T E X x ~. for convenience. v ) ) ]e ~ + e y (pf(vo. the claim follows. v ) ( ~ x ( ~ . v) po(v.vo) vo = v l . Proof. if e*(t) = 0. immediately yields lirrrr~0 flr = 0.e) = ( x .p f (VO.. consider the quadratic Lyapunov function candidate V = 89 T e. t > 0. = v) and )~ t w ' (22) . (17). we prove that if e* = 0 the equilibrium of the unperturbed system (21) is asymptotically stable. the origin e = 0 (of the unperturbed system) is locally asymptotically stable. 0.. ~ . e) + P(~. It easy to show that. v)C + P(x. . t > 0. e) .324 Fabio Conticelli and Antonio Bicchi the above expression. (16). Proof of Theorem 5. .q e . v) p#(v.e [[e(0)l I < 61 and choosing T sufficiently small such that [[r < ~ .. e)) . e) + P(~. e*)[[ < e. ~%I = ~(e + x. t > 0.. if y(k+l)(t) is bounded and by using the induction assumption. Vt > 0. By applying the Lyapunov's direct method. e)) Then.x) the system (15) results (20) f(x. i. v) pc(v. In fact. it follows [[e*[[ < ~tk~ 1 [[~k(t)[[ < K.e T Qe + e T (pf(v0. and the observation error satisfies the condition He(t.Ip." m @ = [~r v r ] r . due to Eq. In the coordinates (x. its time derivative results: ? = eT 6 = e T [(~xx(. Lemma 1 indicates that Vee > 0 there exist J1 > 0 and K > 0 such that if the initial observation error is sufficiently small...
~ ( w ) ) _ 0.e) where (25) (26) .N(w) .Q e 4 .) . ~w()) (w .e.(0. e) has an asymptotically stable equilibrium at e = 0.~(w)) and ~(. ~(w)) 415(w.~~(K(w)) (o. 0.~(*))~(w. e) .)) f Of ~(w \ = /~ ( ))) (23) (~(+) . (15).~ ( w ) = ?~(w) (~. then since from Lemma 2 there exists T > 0 sufficiently small such that Ile*(0)]l < K. e ).P(~. ~(~. from Lemma 1. satisfy the conditions: lim p~(w. e) (28) Ve > 0 there exist $1 > 0 and K > 0 such that if the initial state is sufficiently small. v)e* 4. ~(w)) 4. (23). 0.Q e + P ( ~ . ~(w))415(w. Consider the perturbed system ~" = f(w. e) . has an asymptotically stable equilibrium at (w. e*)]Tll < e.0) .q e . then. Since ~(.e.w ) + p ~ ( w .~(w)) + p d w . .) . ~(w) .e) . i.e II[x(0) T. 15(w. 0). i.N(~i. the cascade system (25). v ) p~(v. i. as a consequence of a known property deriving from the center manifold theory [18].0. v) p~(v.p f ( v 0 . the system w = f(w.P(~.15(w. v) p . e) (24) where the functions p~(w. oT] T is only a function of the observation error e. )i.pf(vo. e) vanish at e = 0 with their first order partial derivatives. (24) the closedloop system (21) is written as = f'(w. e)] (27) + p d w .~(.0. ( v .)) is such that ~(w.~(w)) and p~(w. e) = . e(0)T]T n < 51.o e+0 Using Eqs.p~(w.w) 4. t > 0.~(~. e) .w = [eT.P ( ~ . the solutions of (28) satisfy the condition I][xT(t. Vw E X x ~l~m From Theorem 4. e) = . since also e = ..e) _ 0 Ilell ~. 0) has an asymptotically stable equilibrium at w .Observer Design for Locally Observable Analytic Systems 325 notice also that @ . I1~(+) .~) . eT(t. the unperturbed system (e* = 0) in Eq.O w. . t > 0.pf(v0.~(w)ll lim p~(w. e ) .
u) = 1 + = I ul + "2 z2(x. Moreover. We apply the observer design and the local separation property presented in Theorem 5. t _> to _> 0.x 2 ( t ) u 1 ( t ) = 0. (32) . since the model is analytic and weakly controllable. In fact. vo) = Ul U2 the singularity manifold (i. We consider the problem of finding an o u t p u t feedback control law which locally stabilizes the origin x = 0. the point (X. This analysis indicates that the simple holonomic vehicle (29) is not uniformly observable [22. u) = z (x.e. First notice t h a t the first approximation of the above model around the trivial equilibrium is n o n observable. such that. u ) E X x ~2 : x l u 2 .e. u2] T E ~2 is the control input vector. a stabilizing state feedback of the extended system: X=V 0 it0 = v .18]. an open subset of ~2 containing the origin. (31) does not define a global diffeomorphism in the extended state space X x ~2. causes a loss of rank in the E O J matrix. u) = v ( x ) = 2(x. the m a p 9 l ( x . in fact an input Ub(t). in the case (lp + 1) p = n = 2. which is able only to measure the distance from the origin of a priori fixed reference frame: xu 1 y(x) = ~ ( x ~ + x ~ ) . It follows that. the system can not globally transformed in an element of the class of nonlinear systems described in the recent reference [1]. i. A stabilizing state feedback is simply u = a0(x) = [ .e. i. Fixed lp = 1.k x l . T h e assumption 1 is verified only in the region (X • ~2) \ S. the E O J matrix results: J ( x . By using Theorem 4 with V(x) = 89x T x. lp = 1. but the observability rank condition is satisfied Vx E X. This is is referred to as the full observer design. u = [ul. v0)) . satisfies the condition xl(~)u2(t) . (29) where x = [xl. is the state vector. y E ~R is the o u t p u t vector. k > 0.0) is given by the set S { ( x . at a certain time instant ~ _> to. it is also weakly locally observable Vx E X.326 4 Fabio Conticelli and Antonio Bicchi Example Consider the model of a holonomic vehicle. the set of inputs ub(t) which satisfy the above property are the bad inputs.k x2] T. det (J(x. Ub) lies on the manifold S at time t. x2] T E X.x 2 u l = 0).u) = u2 .
v0). vo) (~ . e) singularity index of the EOJ matrix: log s(x. v o ) .x .A ( v o . Example: output feedback stabilization.k Vo.Observer Design for Locally Observable Analytic Systems 327 is v = ~(x. ~. b) real and estimated x2 component. vo)) P(i. d) control input ~(x.'. 1. 6From Theorem 5. c) added state variables v0. Vo) = do(x.a o ( x ) ) . a) real and estimated xs component. r = 0 . with &o(x.(1 s +Ts) Y(s). Vo) = . Vo) : qJ+(i. vo) 1~1(s) . = vo + P(i. the output feedback given by the equations ~ *o . r\ t ' \ . A > 0. i a) ' 2 ' m b) c m ~ i io Ti i *o ~* c/ d) :i * lo t* m e) Fig. v0). 7_.~(i.lv ~ dteav~m .
5 Conclusions A local separation property for locally observable analytic nonlinear systems has been shown.328 Fabio Conticelli and Antonio Bicchi ~~(~. T .N. the control parameters are: Q . The results are shown in Fig. Hammouri: "A high gain observer for a class of uniformly observable systems". Atassi and H.e). 1999.[ . no. 3] T. 0).1. 1995.c) and 1.1. CelleCouenne and G.001. v 0 ) is reported in (30). Figs. 1.a) and 1. and ~ .. Extensive simulations have shown a meaningful extension of the region of attraction. As future investigation. First. We have proven that the trivial equilibrium of the closed loop system remains locally practically stable. Automatic Control.5 Is. 3. Fossard and D. We now introduce. G. Edited by A.K.5. 9. Bonard. Chapman &: Hall. The second step is to substitute the real state with the estimated one obtained from the proposed nonlinear observer in the control law. 1]T. Fig.0. F. it it necessary to find a stabilizing state feedback of the extended system. IEEE Trans. we are considering an extension of the presented framework to the output feedback stabilization in the large. Vol.1. 1. indicates that the determinant of the EOJ matrix becomes small in the neighborhood of the equilibrium (x. as quality measure of the current state estimation.b) indicate the observer convergence and real state stabilization by the proposed o u t p u t feedback. The initial conditions of the real and estimated state are respectively x = [1. A. 1.v)) The singularity function. NormandCyrot.v)TJ(x. 1991.v0) (33) where J ( ~ . Bonard and H.v) =~/det(J(x. v0) .~ _+ N+. Gilles: "Observability and observers". v0). 44.(0. where the number of the added state variables depends on the number of the o u t p u t derivatives considered in the observer. vol. and A . In the reported trial.d) report the plots of the added state variables v0 and of the control input v = ~(~. if assumption 1 is satisfied along the current system flow. k . 30th IEEE Conference on Decision and Control. 1.J. Nonlinear Systems. ensures practical stability of the trivial equilibrium of the nonlinear system (32).2 . pp. the singularity function s : X x Ntv . 16721687. 2. defined as s(x. Khalih "A separation principle for the stabilization of a class of nonlinear systems". G. depicted in Fig.0. . References 1.
Hestenes and T. Isidori: "New results and examples in nonlinear feedback stabilization". n. Praly: "Global stabilizability and observability imply semiglobal stabilizability by output feedback". n. Gauthier: "Bilinearization up to output injection". 11. Systems ~: Control Letters. 437442. Saberi: "Robust semiglobal stabilization of minlmumphase inputoutpnt linearizable systems via partial state and output feedback". Lin and A. 1995. Hermann and A. Kokotovi~: "Nonlinear and Adaptive Control Design". 23. Z. Silverman and H. 313325. n. 17. 365372. Esfandiari: "Semiglobal stabilization of a class of nonlinear systems using output feedback". 537556. Journal of Control. 17911794. Isidori: "Nonlinear Control Systems". pp. 7. M. Othman: "A simple observer for nonlinear systems. 3. 21. 56. Krener and W. Systems & Control Letters. Esfandiari and H. 1988. 33. 10. IEEE Trans. vol. n. lsidori: "Linearization by output injection and nonlinear observers".J. vol. and J. pp. International Journal of Control. 11. 1985. 2. Gauthier. Khaiil: "Output feedback stabilization of fully linearizable systems". no.M. Systems & Control Letters. Journal of Optimization Theory and Applications. 1986.K. 1. Automatic Control. A. 6. pp.R. n. no. 14121415. H. observers and finite dimensional filters". 19%216. vol. 6%77. M.R. Inc.. Int.J. Hwang and J. 1994. 1995. Praly: "Tools for semiglobal stabilization by partial state and output feedback". 5. 36. Respondek: "Nonlinear observers with linear error dynamics". A. pp. 12. H. 10291041. A.H. 1989. 57. I. no. vol. 9. vol. A. Ciccarella. Seinfeld: "Observability of Nonlinear Systems".E. Application to bioreactors'. John Wiley & Sons. 22.P. pp. vol. Krener: "Nonlinear controllability and observability ".Observer Design for Locally Observable Analytic Systems 329 4. 20. SIAM Journal on Control. J. Khalil and F. 16.P. 1995. H. 28th IEEE Conference on Decision and Control. 1983. pp. Automatic Control. 14. vol. n. SIAM Journal Control and Optimization. 3. vol.J. 4%52. A. Meadows: "Controllability and Observability in TimeVariable Linear Systems". n. 24. 1993. 7. pp. 1967. 721729. 13. pp. 7. F. 5. 19. IEEE Trans. 21. 22. 18. Hammouri and S. Byrnes. Levine. M. 40. 100%1037.R. A. vol. Kupka: "A finiteness criterion for nonlinear inputoutput differential systems". Florida. M. IEEE Trans. . 15. Krsti~. Tornamb~: "Use of asymptotic observers having highgains in the state and parameter estimation". vol. F. J. Marino: "Nonlinear system immersion. Dalla Mora. 2. SIAM Journal of Control and Optimization.K. pp. 12. Teel and L. Systems & Control Letters. Fliess and I. R. 137142. 1989. 6. n.. 5.P. pp. G. L. 1983. Journal of Optimization Theory and Applications. Kazakos: "Orthogonal representations of nonlinear systems and inputoutpnt maps". n. pp. 1968. Systems & Control Letters. IEEE Trans.I. 1986. and A. n. Guinn: "An Embedding Theorem for Differential Equations". and D. C. Tampa. Springer Verlag. 139149. pp. pp. Gauthier. 2. R. 1993. Germani: "A Luenbergerlike observer for nonlinear systems". 10. 23. 38. A. Hammouri. Krener and A. Kanellakopoulos and P. pp. 8. SIAM Journal Control and Optimization. 22. Teel and L. Ceile. Automatic Control. pp. M. 3rd ed. 728740. 1995. 1977. Automatic Control. 1972. no. Systems & Control Letters. 1992. J. 1992.
SIAM Journal of Control and Optimization.330 Fabio Conticelli and Antonio Bicchi 25. Systems & Control Letters. n. 28. 1982. 3. 357362. pp. 1990. n. 1990. 26. A. D. J. 29. Williamson: "Observability of bilinear systems. 20. 17. vol. Systems & Control Letters. XiaoHua and Gao. 1987. Automatica. pp. 9.J. SIAM Journal on Control and Optimization. 13. 1977. Systems and Control letters. nO. M. pp. WeiBin: "Nonlinear observer design by observer error linearization'. n. pp. 1989. vol. Tsinias: "A theorem on global stabilization of nonlinear systems by linear feedback". 27. with applications to biological control". Tsinias: "Further results on the observer design problem". 149156. 30. 27. 243254. 411418. n. . 199216. 338354. 14. pp. pp. J. Zeitz: "The extended Luenberger observer for nonlinear systems ~ . Xia. van der Shaft: "Observability and Controllability for Smooth Nonlinear Systems".
caspur. We address the problem of fault detection and isolation in presence of noisy observations for bilinear systems. A(x) :D (in the natural basis { d x l . f2(x) = }IV). edu Dipartimento di Informatica e Sistemistica Universith di Roma "La Sapienza" 00184 Rome.A n Hoosuboptimal Fault D e t e c t i o n Filter for Bilinear Systems* Claudio De Persis 1 and Alberto Isidori 1 2 I Department of Systems Science and Mathematics Washington University St. . and P = s p a n { p l . AFMC. 1 Introduction We consider in this paper the problem of fault detection and isolation for systems modeled by equations of the form rn i=1 d k = Ax + Z uiNix + g(x)m + Z p i ( x ) w i y=Cx. Note t h a t we say t h a t a distribution A (codistribution Y2) on jRn is independent of x if there is a subspace :D of JR'* (a subspace }N of the dual space of JR'*) such that. by DARPA. Pa} is "independent of x". by AFOSR under grant F496209510232. . MO 63130. y E JRP. . it A b s t r a c t .. in the natural basis {O/Oxl. In these cases. AFRL. and by MURST. under grant F306029920551. m E JR. * Research supported in part by ONR under grant N000149910697.. Italy isidor i@giannutri. dx'*} of the cotangent space to JR'* at x. O/cgx'*} of the tangent space to R'* at z.. . USA depers is@zach. Our solution rests on results derived in the geometric theory of fault detection. . Louis. . . it turns out quite useful to identify A with a m a t r i x D whose columns span 2) (and Y2 with a matrix W whose rows span }IV).. . . i=1 (1) in which x E JR'*. eust i.
i.e is affected by) the input m(). Here. Also.2. Following Theorem 3 of [5]. 2 Observability codistributions The concept of observability codistributions for an inputaffine nonlinear system and the related construction algorithms have been introduced in [3] (cf. i = 1. Bilinear systems (1) can be indifferently seen as a special subclass of the systems dealt with in either [2] or [3]. and A(u) is a matrix of analytic functions) was initiated. even though in a different setting. the residual r(.) is identically zero (see [4]). formulas (8). it is known that the problem of fault detection and isolation is solvable provided that a constant matrix L and a matrix D(u) of analytic functions exist such that the subspace 12. = y) + y) y)u (2) such that.332 Claudio De Persis and Alberto Isidori Solving the problem of fault detection and isolation for (1) means designing a filter (a residual generator) of the form = . is decoupled from) the inputs u and w and asymptotically converges to zero whenever m(. (12) and Proposition 2). First of all note that. for notational convenience.e. defined as the largest subspace of Ker{LC} which is invariant under A(u) + D(u)C for all u 9 U. we introduced the notion of observability codistribution for systems of the form (3) and we proved that the existence of the matrices L. in the cascaded system (1)(2). satisfies: (a) el (x) ~ ~2 and (b) e2(z) 9 Y for all x. in which the study of the problem of fault detection and isolation for systems of the form y = c~ (z) (v~ E/R. formulas (4) and (5)). y = Cz+v. For these systems. Related work on the problem of noise attenuation in fault detection.e. depends "trivially" on (i. In the case of stateaffine systems. the corresponding concept and algorithms are those presented in [2] (cf. for a system of the form (1). D(u) above can easily be characterized in terms of this notion. . set No = A. we examine some details of this approach for systems of the form (1) and we use the results to address the problem of fault detection in the case in which the measurements are affected by noise v. the algorithms in question assume the special form described in what follows. In [2].) depends "nontrivially" on (i. can be found in [1]. proposing a filter in which the influence of the noise on the residual is attenuated. conditioned invariant distributions and observability codistributions are independent of x.
n 1. In the terminology of [3].• 4 .m. . Consider the system (1). In the terminology of [2]. L e m m a 2. (5) j=o with i = 0 . consider the nondecreasing sequence of subspaces (of the dual space of ]Rn) Qo = Q rl span{C} m Qi+.. the distribution A : x ~ S. and set o.1.s. .((S. . ." stands for "observability subspace algorithm". . Finally. = o. . Remark 1... = Q N ( Z QiNj + span{C}). m (7) Remark 2. .a.s. where the acronym "o.a. set Q. .. . (6) i = 0. a codistribution Q : x ~ Q fulfilling the last two properties of (7) is an observability codistribution for (1). n .~?) .. [3].~ is the minimal ele ment (with respect to subspace inclusion) of the family of all subspaces of 1~n which satisfy S D 7" Ni(S O K e r C ) C S.a. . i = 0.n1.s. consider the nondecreasing sequence of subspaces (of/R") 80=7' Si+l = Si + Z Yj(Si f'l Ker{C}) j=o with i = 0 .An Hoesuboptimal Fault Detection Filter for Bilinear Systems 333 Given a fixed subspace P of ]R=. . and set S ~ = Sn1.a. Q = o. The subspace Q~.~ is the minimal conditioned invariant distribution containing the distribution P : x ~ 7". The following results describe the main properties of the subspaces thus defined. The subspace S. and hence the codistribution Q* : x ~~ Q~ is the maximal observability codistribution contained in the codistribution P • : x ~~ 7. (4) Given a fixed subspace Q of the dual space of ]Rn . is the maximal element (with respect to sub space inclusion) of the family of all subspaces of the dual space of IRn which satisfy Q C T"• QNi C Q + span{C}.s.(Q).(Q) = Q. L e m m a 1. . 1 .
~ + s p a n { C } . we first observe that the following holds.kxk + N 2 .)• and from the fact that the chosen isomorphisms preserve the structure of the system. A span{C}. c a. Q~ C P • g(x) ~ (Q~. . set A(u) := A + ~]~. Y2) = H y on the output space which transform (1) into a system of the form: kx = AllXl + Al~X~ + E ( N ~ l X l "4.. x k ) u i + t2(~)m + p2(~)w k=l ~=1 k=l k3 = E Azkxk + k=l /:1 k=l (8) Nakxk) ui + g3(f:)m + p3(k)w Yl = Clxi y2 X2 . (iii} there exists a change of coordinates ~: = col(xl. . for all u e IRm. the form (8) can be obtained in the following way: define xl = Tlx with 7'1 a matrix whose rows span Q~ and define Yl = Hly with H1 a matrix such that the rows of the matrix H1C span Q~. The proof of the various implications can be obtained by specializing the proofs of similar statements given in [2] and [3]. fii) e(x) r (Q.rl = A i l Z l q/=1 uiN~lXl (9) Yl = ClXl . and p := s p a n { p l . . i = 0. x~. .fi = col(y1.~N. is observable. since Q~ is . 1 . Then. the form (8) derives from the properties Q.)'.334 3 Claudio De Persis and Alberto Isidori A natural candidate for residual generation Specializing the results of [2] or [3] to system (1). and the subsystem . Consider system (1). P r o p o s i t i o n 1.Nil2x2)ui + el (~')m i=1 ~ = E A. in the implication (ii)=~(iii). p e } . m . x3) = T x on the state space and a change of coordinates .m=lN. . Proof. The following conditions are equivalent: (i) there exist a constant matrix L and a matrix D(u) of analytic functions such that the largest subspace V contained in Ker{LC} and invariant under A(u) + D(u)C. Note in particular that. . It is worth stressing that. in which gl is nonzero. . . satisfies: (a) e(x) • 12 and (b) PCP.u.
. and a straightforward procedure . is the maximal element of the family of all subspaces of the dual space o f / ~ " which satisfy (7). In this respect. leads itself to an immediate interpretation of the construction proposed in [5]. in which G is a suitable matrix of functions of time.to actually identify this subspace and to construct such matrices.. Indeed. if D(u).condition (ii) . Theorem 3. the inclusion P C 1. As a matter of fact. the vector field g(x) cannot belong to V for all x. the latter is nothing else than an observer for z l . L in (i) ever exist. by Proposition 2 in [2]. the subspace ~' := Ker{T1} is the largest subspace of Ker{LC} which is invariant under A(u)+ D(u)C for all u E / R " . Special form (8). Since the disturbance term does not appear in the xlsubsystem. must hold. uiY~l .to determine whether or not the subspace 1] and the pair of matrices D(u). 4 The condition (i) is precisely the condition assumed in [5]. the residual generator of [5] reduces to a system of the form = Aa. On the other hand. (11) In fact. necessarily l~ (z) # 0.GcTC1)e + gl (k)rn (13) ./ : obeying k = (All + ~ i=1 r=C1 e .C l k . L are chosen as indicated in the proof of the Proposition above. since gx ~ 0.An Hoosuboptimal Fault Detection Filter for Bilinear Systems 335 spanned by the rows of 7"1.i: + i=1 u. (lO) L=Hx .)• The observability of the subsystem (9) derives from the fact that Q~.GCT(cli: .m. and in particular subsystem (9). yielding an "observation error" e = z l .Yl) (12) r = Yl .1.N~lk + AI2y2 + Z iI uiN~y~ . with is easily obtained by setting D(u) := Do + [o Di = T 1 ~ and 0 ~2) H. i=O.based on the change of coordinates described in (iii) .. The implication (iii)=r ]~']~m=lDiul. to establish the existence of a solution of the fundamental problem of residual generation for system (1). the equivalences established in the Proposition above provide: a conclusive test . because otherwise g(a:) E (Q~.. it is easy to check that.
can be cast in a gametheoretic setting. then (12) solves the problem of residual generation. among the other things. Thus. tl) is a fixed time interval (with tl < or Proposition 1 shows. both xu and v are going to influence the residual when m = 0.336 Claudio De Persis and Albert6 Isidori I f G is chosen. that in the special new coordinates used to obtain (8) we can write m xl = A H X l + A12x~ + Z ( N ~ l x l i=l + N~lx2)ui + e l ( k ) m (15) Yl = C l x l + vt Y2 = x2 + v2 . where [O. the output of the residual residual generator must be identically zero. we address in what follows the problem of finding a filter with state variable ~:. in spite of a possible nonzero disturbance w (whose role in (15) is taken by the input x2) and of the observation noise. Consider the case in which the observed output of system (1) is corrupted by a measurement noise v. tl). m d k = Ax + Z i=1 u i N i x + g(x)m + Z p i ( x ) w i i=1 (14) y=Cx+v. as it was done in the residual generator (12). the problem of noise attenuation in the design of residual generators can be cast as a problem of state estimation in a gametheoretic.via output injection . in view of the way the noise affects the observations. driven by the observed outputs (yl. Y2). Ideally.e. To this end.as in [1] . and let v(. or Hoo. If this is not the case.the effect of x2 on the dynamics of the observation error. it is no longer possible to compensate . the problem becomes a problem of attenuating the effect of both x2 and v on the residual signal. Following the fundamental results of [1]. in this case. which for instance . as in any realistic setup. Motivated by the approach of [1]. 4 Fault detection with noisy observations In the approach to residual generation summarized in the previous section the observed output y that feeds the residual generator is supposed to be noisefree. in the absence of the fault signal (the input m to (15)). if the second component y~ of the observed output is corrupted by noise. i. . In practice. as suggested in [5]. As a matter of fact. setting.) be a function in L2[O. then the important problem of attenuating the effect that this noise m a y have on the residual r must be addressed. the special features of form (8) prove to be particularly helpful. in such a way t h a t the latter is exponentially stable for m = 0.
~+A2y~+7~II'CTV'(y. Standard methods show that. + Ilvxll~.I c T v 1 . has a solution II(t) defined and nonsingular for all t E [0. + A~x2 y. and As.(.(0) .)] dt (20) IIx. x .(0) .C. for convenience. for some fixed positive number 7 and a given choice of weighting positive definite matrices Q. + IlYx .(0).:exll~. P0. + 72FIA2MATII + CT(Q  "/2 V1)CI = 0. (17) (18) (19) m i in which. + Vl Y2 = x2 + v2 . = Alx. . then the attenuation requirement (16) is fulfilled by an estimate ~ of x. Indeed. this particular choice guarantees the attenuation properties expressed by (16). = C . M.~0Jl~o (where/70 := 7~P0) and the differential game min max m a x m a x J _< 0 (Yl. if the differential Riccati equation 1:I + ATII + HA. tl].i~)ll~ .)ll~dt <_~ .i. fo' lie. (21) with initial condition H(0) = / 7 0 .Y~) x2 xl(O) subject to the differential constraint (17).CI~).z211~. we consider the cost function J = fo' [IIC~(x~ .i'oll~. we have set Ai := Aii + Y~i=i Nllui and A~ := Alu + Y]m=l N~2ui. V. . ~(0) = ~0 9 (22) Comparing the dynamics of the estimator (22) with that of the residual generator (12) and bearing in mind the expressions of A.o tl 2 (16) for all the signals vx and v2 and for all the initial conditions z. provided by the estimator ~=A.A n Hoosuboptimal Fault Detection Filter for Bilinear Systems 337 such that. To solve this problem. fo [llv~llM._.72(11Y2 .]dt + I1". we see that (22) corresponds to choosing in (12) a "gain matrix" G of the form G = 7 ~ I I . subject to the constraints ~. where/7 is the solution of the differential Riccati equation (21).
oo) and hence e(. ~ ) . the "input" t. In fact.1 C T V . and such that the system = (A + P [ 7 . system (26) coincides with the homogeneous linear system (24) and. Lemma 3. r system = (A1 .7 2 H . ~ ) .338 Claudio De Persis and Alberto lsidori It is interesting to examine the special case in which tl ".l ( t ) is bounded on [0. (23) In order to (16) make sense for tl = r we require e() to be i n / ~ [ 0 .c T c ] p (25) in which A := A1 . If also n . C := V1/2CI .1 v l .~ .1 C . B := A2M 112 . (26) In fact.l ( t ) satisfies the differential Riccati equation ~' = A P + P A T + B B T + P [ 7 . Let P(t) be a symmetric solution of (25) satisfying o < Z~z < P(t) < ~2I for all t E [0.~ obeys = (A1 . the hypotheses of the previous lemma guarantee that the estimator (22) renders (16) fulfilled over the infinite horizon. hence.~ L T L . for some fll.) is in/:~[0.p c T C ) e is exponentially stable. L :~~ Q I / 2 c 1 .c T C ] ) p is exponentially stable. C.1 C T V . ) e if the homogeneous linear (24) is exponentially stable.7 2 I I .1 v 1 to (23) is in/~u[0. ~ ) . ~ ) and m(.1 C T V . to this end. oo) the matrix P(t) := 7 ~ H . ~ ) .~ L T L . the system = (A . Note. recall that if/Z(t) is defined and nonsingular for all t E [0.132. Then. Conditions implying the exponential stability of the latter can be derived from the works [6] and [7].) is in/:2[0.1 C I ) e + ~l(~)m . P. The following result of [7] provides the requested condition. by definition of A.l C T V .A2v2 . .7 ~ H . (~)m  A2v2  72 I I . Assume that u(t) and x(t) are bounded on [0. that the estimation error e = x l .
. Ravi R.P. (1999) A differentialgeometric approach to nonlinear fault detection and isolation. Speyer J. Khargonekar P.An Hoesuboptimal Fault Detection Filter for Bilinear Systems 339 5 Conclusion We have proposed. De Persis C. IEEE Transactions on Automatic Control 36. El Yaagoubi E.. Isidori A. Chung W. 455465 . European Journal of Control 4. De Persis C. De Persis C. 216 6.. Hammouri H. Isidori A.M.M.C. (1992) Normalized coprime factorizations for linear timevarying systems. Kinnaert M... IEEE Transactions on Automatic Control 43. Pascoal A. 152166 7. (2000) An addendum to the discussion on the paper "Fault Detection and Isolation for State Affine Systems".. Nagpal K. Khargonekar P.. European Journal of Control 3. Systems & Control Letters 4. (1991) Filtering and Smoothing in an H ~176 Setting. Submitted 5..P. 143161 2. lsidori A.H. References 1.. Systems & Control Letters 18. within a g e o m e t r i c framework. (1998) Fault detection and isolation for state affine systems. a solution to t h e p r o b l e m of fault detection and isolation for s y s t e m s of the form (1) in the presence of noisy observations. L. (1998) A game theoretic fault detection filter. (2000) On the observability codistributions of a nonlinear system.
The proposed control algorithm is repeatable since the set of initial conditions for which given performances are guaranteed is explicitly determined. it A b s t r a c t . W i t h our a s s u m p t i o n s none o f the previously p r o p o s e d controllers works for the class of s y s t e m s considered. Italia ddomit illa@hotmail. a b o u n d on the n o r m of the o p t i m a l weights is o b t a i n e d which allows us to use p r o j e c t i o n a l g o r i t h m s in the p a r a m e t e r u p d a t i n g law. By virtue of Bessel inequality. In fact [11] requires the knowledge of an u p p e r b o u n d for the nonlinear function o u t s i d e a c o m p a c t set and of an u p p e r b o u n d for the a p p r o x i m a t i o n . a n d P a t r i z i o T o m e i Universit~ di Roma "Tor Vergata" Dipartimento di lngegneria Elettronica Via di Tor vergata 110 00133 Roma.Adaptive Control of Feedback Linearizable Systems by Orthogonal Approximation Functions D o m i t i l l a Del Vecchio. The problem addressed in this paper is the control of a class of SISO feedback linearizable systems containing unknown nonlinearities with known bound on a given compact set. Projection algorithms can then be used to assure that the parameters estimates remain inside the set in which the optimal weights lie. Since the control a l g o r i t h m guarantees t h a t t h e s t a t e vector r e m a i n s in t h e c o m p a c t set where the a p p r o x i m a t i o n holds. no a p r i o r i knowledge on the bound of the norm of the optimal weights is required. uniroma2. The unknown nonlinearities axe locally approximated by a finite sum of Orthogonal basis functions so that by virtue of Bessel inequality a bound on the norm of the optimal weights derives directly from the bound of the nonlinearity. Riccardo Marino. its i n t e r n a l a d a p t a t i o n a l g o r i t h m can reconstruct the a p p r o x i m a t i o n of the u n k n o w n nonlinearities when t h e r e is sufficiently persistency of e x c i t a t i o n so t h a t s m a l l e r t r a c k i n g errors are obt a i n e d w i t h o u t increasing the control gains. By assuming t h a t an u p p e r b o u n d on the m o d u l u s of t h e nonlinear function on a given c o m p a c t set is known. 1 Introduction In this p a p e r a r e p e a t a b l e control design m e t h o d o l o g y for a class of feedback linearizable p l a n t s with unknown nonlinearities satisfying the m a t c h i n g condition is proposed. con. o r t h o g o n a l functions are used to a p p r o x i m a t e the unknown nonlinearities on the given c o m p a c t set. Marino@ing. We d e t e r m i n e the set of the initial c o n d i t i o n s and the values of the control p a r a m e t e r s t h a t assure the s t a t e vector to r e m a i n inside the c o m p a c t set in which the a p p r o x i m a t i o n holds w i t h b o u n d e d control signals.
u is the scalar control input.... Since the bound on the approximation error holds only locally. Marino and Tomei error. . = f ( x l . that obstructs the convergence of the parameter estimates to their true values. ~ ) + r y=Xl in which z = (xl . y(n1))T E R n . this latter is also required in [5]. A1. We make the following assumptions...: If(z)] < B y . . = (y~. i. X n ) T belonging to R n is the state vector. [1] requires the measure of the state derivative and a known b o u n d on the norm of the optimal weights.. In this paper we guarantee this t h r o u g h a constructive control algorithm without any dependence on the initialization of the approximation network and without introducing any forgetting factor in the adaptation dynamics..~ < l. r is a known function. V z E A C R n 9 A3. This bound is not required in [8] but it necessary to determine the size of the neighbourhood around zero in which the tracking error converges. . (1) ~ . . < U M .... A bound for the modulus of f(a~) is known when z belongs to a known compact set A = In[r~. with the exception of [8] where a bounding function for this error is assumed to be global. V I E R .nV 1...y(/~) such that ly(/)(t)l <_ er. We consider smooth bounded reference trajectories yr(t) with bounded time derivatives y(U... it is a key point to assure the state vector to remain in the desired bounded set. . u is a positive constant parameter with known bounds: O< l. with Moreover we define the vector y.342 Del Vecchio. as in [5] and in [8]. x . u is an unknown constant parameter. The bound for the disturbance is known: Id(t)l < D . . y is the scalar output to be controlled..e. for 0 ___ / ___ . A2. . d(t) is a scalar bounded exogenous disturbance. 2 Problem Statement Consider singleinput singleoutput uncertain feedback linearizable systems in the matching structure xi = xi+l i = 1. while in [10] it is not required but it is needed to determine the set in which the state vector will converge. f ( z ) : R" + R is an unknown nonlinear function belonging to C(A).] C R n.. If f ( z ) and u are known and d(t) is zero then the feedback linearizing control input is: /J (2) . x ~ .r. A4.
i. Let ok := (f. Now let us consider a general approximation with N terms SN(Z) = ~ k = 0 a k ~ k ( z ) . rnis(A) (4) be the Fourier coefficient. nonlinear adaptive techniques ([7].e. [t4]) cannot be used. ~ k ) := fA f ( ~ ) ~ k ( z ) d ~ . it can be proved ([4]) that the m i n i m u m distance between SN and f ( z ) is reached when ak = Ok and the following equality holds: NI N1 [If  SN[I2 = Ilfl] 2 . The set r of orthogonal functions is said to space of the bounded functions f in C(A) if the Parseval where Ilfll 2 = oo k=O be closed inside the equality holds: ilfll~ (7) E Of = mis(A)" . Since f ( ~ ) is not linearly parameterized. A set of bounded continuous functions {~Ok(X)}:= r for ~ belonging to the set A is said to be a set of orthogonal functions if Definition fA ~i(z)~oj(x)dx with = {mis(A) 0 for i = j for i r j (3) mis(A) is the Definition 2. where the operator ( .~k) measure of the set A .f) = fA f2(~) dx" Definition 3.Adaptive Control by Approximation Functions 343 so that the closed loop system becomes linear and by a proper choice of K it becomes asymptotically stable. We will use orthogonal functions to generate functional approximations for the unknown f ( ~ ) . 1. f ( x ) oT~(x) with 0 E R v and ~ ( x ) a known regressor vector.Y~ 0~.1 Approximation network structure Let us recall some preliminary definitions from [4]. ) is the scalar product between elements of the linear euclidean space of the functions belonging to C(A) and it is denned as ( f . k0 mis(A) (5) from which the Bessel inequality follows N1 E O~<_mis(A) lifll~ (6) k=O (f. Since both f ( ~ ) and u are not known (2) cannot be used. 2.
..S ) = 0 only i f f . where g1 Ok is defined in (4).. It can be easily shown t h a t the set r = {exp(j(mlxl + ... T h e resulting a p p r o x i m a t i o n with N t e r m s is given by lTl n . chosen to a p p r o x i m a t e the function f ( z ) ..xn)exp(j(ml xl + .1 SN :O0"l Z i=1 Oi~v~cos[( M T ~ )27r/T] + Oi. + mnXn)2rr/T) (8) m 1.m~. Marino and Tomei Since in the space of orthogonal functions t h e p r o p e r t i e s of closure and completeness are equivalent..344 Del Vecchio. then N = 2rn n . If by hypothesis A is the h y p e r c u b e of side ra. where we call S = limN. we recall from [4] the following result.S is zero. which is equivalent to fA ~ k ( ~ ) g ( z ) d z = 0 for every k. A n e x a m p l e of a c o m p l e t e set of o r t h o g o n a l functions belonging to C(A) is o b t a i n e d by generalizing the concept of Fourier series ([4]): f ( x i . So by considering (7) and (5) and the fact t h a t ( f .. + rnnx.xn) = Cm. T h e n u m b e r of a p p r o x i m a t i n g functions N when x E R'* is det e r m i n e d as follows. . SN can be also r e w r i t t e n as S g ( z ) = Y'~k=o Ok~ok(X). .)27r/T)} is c o m p l e t e and the functions are orthogonal.rnn : ~ JT/2 T/2 "'" JT/2 f ( x l . in the above relations T = 2ra. ..S .. x2.. Be m the n u m b e r of h a r m o n i c s .dxn. (9) where A = In[T/~.0 in A. By defining ~o = 1. exp(j(mlxl + . f ... Vx E A t h a t can be written in the following property.rn~ =co 1 IT~2 emlrn~. .~)2~]. T h e set of orthogonal functions r x E A C R n. ~ i ( z ) = v~cos[(M.m~ ..)2rr/T)dxl.. is c o m p l e t e in the space C(A) if and only if there is no element in C(A) different from zero t h a t is orthogonal to all the functions of the set r In other words (~k. .. x ~ . + rn. rn .x. we o b t a i n t h a t f ( ~ ) = ~"~k=0~176 Ok~k(z) .. including the zero frequency.. g) = 0 for every k is equivalent to g ( x ) .T T ~o~. .1] so t h a t M i ~ M j for i r j . v~sin[( M ~ z)2~r /T] where M i E R n is the harmonics vector whose c o o r d i n a t e s are n a t u r a l n u m bers and each coordinate can assume values belonging to [0..TI2].ooSN..1... ._l(x) = v~sin[(MTz)27[](lO) T t h a t satisfy (3).
.Adaptive Control by Approximation Functions 345 Property I. considering the set of functions f in C(A) as specified in A2 and taking (4) into account. the constant parameters Oi. can approximate functions f in C(A) with increased precision by sufficiently increasing the number of nodes. (11) k=O This result shows that networks with one layer and functions ~oi in r complete set of orthogonal functions. 0 < k < N . are given in (4) and 9 = (~0. 2. Vz E A .. .rnis(A) Z k=0 0~ _1loll _< B~. such that [r _< r for every x in A .r) : Z k0 Ok~k(X)"'... then for every ~ ~ 0 3 N(r > 0 such that N1 for N _> N(r ]f(z) E 0k~k(z)l ~ e . Therefore in our control problem we will consider a network that assures the approximation of f(x) in the given compact set A with approximation error e(x).2 The tracking problem From (12) we consider an estimation of f(x) given by N1 /(. 0N_1) T and 0k. in fact by (6) and A2 it follows that E O~ <_ mis(A) k~O so that N1 . .1) T..1. as defined in (4).~N1) 9 At this point (14) is the estimate of the functional approximation sTY(x) of f ( x ) in A with the . 0N1) T is the estimate of 0 = (00. we can establish a bound on the norm of the parameter vector 0 : (00. ON .. are also unknown. i..e. k=O Vx E A (12) in which since f(x) is unknown.. From the bound assumed for Ifl when ~ is in A ..: N1 f ( x ) = E Ok~ok(z) + e(x) .0 T ~ ( z ) ' VZ E A (14) where 0 = (00.. If we denote by r the complete set of orthogonal functions. 2 (13) Notice that the bound does not depend on N..
~e[o. 3 Main Result Let A1A4 hold. ... (ii) all the signal of the closedloop system remain bounded. the approximation error e(x) m a y be viewed as a disturbance and the robust adaptive techniques proposed in [6] are applicable provided that we can guarantee that x remains in A tbr all t . n . ~ . . Uu. . with 0 < r < r~.bM 9 (t6) T h e o r e m 3.(r)l) v1r .(e + D + v~sup.. .dllO(r)[[ + supTe[o. ~.346 Del Vecchio. urn). By virtue of (12) and by virtue of A3 we rewrite system (1) a. lb(t)] _< bM + a Vt _> O.2_ 0 the L ~ inequality is verified ly(t)  y~(t)l <_cq(O)exp(A~t)+ + l_~.. The tracking problem is formulated as in [6].t]l[. tbr any initial condition z(0) E B(r) C A..l / m 2 .s xi+t ~}n i=0 y = Xl i = I. . (ii) tbr all t . (v) when e(x) and d(t) are zero the asymptotic tracking error is zero.1 (15) in which R(x. b(O) E B(bM) C R and tbr every reference signal with Yr E B(r) the tbllowing properties hold: (i) x(t) E A. b) parameterized by the positive real k such that tbr every k > [r D. 51(0) E B(Bs) C R ~'. there exists a state feedback d y n a m i c control u ( ~ . u~~). . ur. [[O(t)i I < Bj + ~.. ~ .1 Given system (15) under assumptions A1A4 and (16) for any arbitrary real e > 0. that is to design a state feedback control and to determine the set of initial conditions so that: (i) x(t) E A Vt _> 0. (iv) arbitrarily improvable transient pertbrmances of the tracking error are given in terms of both the approximation error and the parameter estimation error.. we0. Marino and Tomei parameter vector bound given by (13). (iii) the output y is tbreed to enter asymptotically into an arbitrarily small neighbourhood of the smooth bounded reference signal satisfying A4.t) = d(t)+~(x) with Is(x)[ < e V~ E A and u0 := (UM+Um)/2 so that u = u0 + b with u0 known and b the unknown parameter satisfying Ib] < t t M  .
.Vn "~. .n~. y~.. ~ = 8 ..~nen ~.8 with 8 = (O0. . y~n . xn. . ~(x~... yr. . .yr.b .Y r ei = xi .2(~1.. We then show that the above assumption is verified for a specific choice of the ball B ( r ) and of the control parameter k..y!i1) (17) i = 2...1 Given system (15). . t ) 1 ]. 6 ) . A1 is a positive control parameter....~!~~).. *~. y~.. . .v~).1 we give the following results. v~).. ~ = ~.x* . b = b .. Before proving T h e o r e m 3.. . and n + 1 s m o o t h functions 9 ( 4 ( x ~ .R ( ~ ...y(~n~)) transforming system (15) into: di = .y(~").n.l (_~V q... First we proceed by assuming x(t) E A for every time so that A2 and (12) hold. v(~. yr..y!. It (19) . L e m m a 3.Adaptive Control by Approximation Functions (iii) for all t > 0 the L2 inequality is verified 347 fo t (y(r) .y~(t)) = 0 We will provide a constructive proof for T h e o r e m 3.y~(r))=dr < as(O) + ~ f0 t (e s + O = + 1 NII&T)II s + ~2)d~.2 ) . .... xn~....A i e i + e i + l + v i e'n : .~n.....1 in which both the dynamic control and the radius r of the ball B ( r ) are explicitly given.. ~ ) .~.1 ) T and 8i given in ( 4 ) .~. a is an arbitrarily small positive constant. x~(x~....y~. in which a~(0) and as(0) are nonnegative constants depending on the initial condition a~(0). (18) an adaptation dynamics = ~(xl. a global filtered transformation el ~ Xl . . V n . .. (iv) if d(t) = 0 and e ( z ) = 0 then l i m t + o ~ ( y ( t ) ... with wl (') and us(') Lipschits continuous functions...1)..~ T ~ ( x ) i = 1..~)) ... ~ N .. there exists a state feedback control 1 u = ~~((~'0 + b)v + ~) .
2 ) ) (ni). . if vl . By considering (24) and (23) we conclude t h a t e = ~ .y. L e m m a 3 . where x = ( x l .e i . P r o o f (sketch).1 we d e t e r m i n e a global filtered t r a n s f o r m a t i o n (18) a n d a s t a t e feedback control u (17)..y .. where v is given by (21).. a n d Ai are positive control p a r a m e t e r s .. Y ( n ... . .en) T.. 8) ..1 (sketch).. Vr...A) we o b t a i n (22).) T we can write e = z Yrz* (23) (21) (22) now by considering (20) and (19). .1 for 2 < i < n .A ) ( z ...) where 0 a21  (24) 0 .2i(z ..1 (20) and v ...)....... . P r o o f o f T h e o r e m 3. . Marino and Tomei vi(xl.xn. b) (25) . Y r = ( Y r .1 then there exists an inferior t r i a n g u l a r m a t r i x A such that e = A(x ... By a p p l y i n g L e m m a 3. yr. . b = w2 = 7 ~ P r o j ( v e n .. . . for 1 < i < n  1 and Vn(Xl. 0 .Y r ) = ( I ..&~ . we derive t h a t : :~* = A ( z .y..V~ ] are n s m o o t h functions. P r o o f (sketch). . x n ) T. an(nI) and its coefficients aij are i m p l i c i t l y defined by the A~ t h a t are arb i t r a r i l y chosen. Consider the t i m e derivatives of (18)....~T@(z) _ Anen + vn t..A i e i + x' * + vi i i = 1. 2 Given the t r a n s f o r m a t i o n s (18) a n d the relations (20). .Y r . p u t X* i+1 = ... t r a n s f o r m i n g s y s t e m (15) into s y s t e m (19)9 In p a r t i c u l a r we choose: 0 = wl = 7 1 P r o j ( @ e n .xil.. .D ) T. ) a n d defining A : = ( I . 0 .. For the c o m p l e t e p r o o f see [6]. .. n . Defining z* = ( x T .. 0\ 0 a31 a 3 2 0 an1 . .. x~.y(n) so t h a t we o b t a i n (19).348 where Del Vecchio. e = ( e l ..y(r i .
e..~ . considering (21) and (29) we finally obtain + 1 . O) >_ oTy. since by virtue of property 1) 0 and b remain bounded respectively in 121 and 122 for every time.1 ~ = _88 (29) in which VM := max~eA(V) and it is c o m p u t a b l e by considering from (24) that ~.0 Yn "''7 = (28) 3ken  en1 Vm j vi = .t v .Adaptive Control by Approximation Functions where P r o j ( y . vt > 0.~ken(~mm ) ) . @)11 ___ UylI. O) = where (26) IIt~ll 2 p ( ~ ) := . When @ = 8 r a = r a l and by (13) ~. i. 4) O P r o j ( y . 2 (30) and computing the time derivatives by taking (19) into account. O~ is Lipschitz continuous. = anl(x2 ... n. 3) I[_Proj(y.1 i = 2.. + an(n_U(x n _ y(.~)) and t h a t A4 holds. In order to prove property (i). y) <_ 0 yp if p(tO) > 0 and (gradp(tO). (27) while when O = b r a = rn~ and by (16) rD~ = bM In this way if 6 ( 0 ) E 12 then the following properties hold: 1) II&(t)ll < rn + ~. The signals vi for i < 1 < n are chosen as 131 ~. O) is the Lipschitz continuous projection defined as y ifp(O) < 0 y if p ( O ) _> 0 and (gradp(O). yp = y [ I  p(O)gradp(O)gradp(o)T] iigTadp(~)ll2 with a an arbitrary positive constant and r a is the radius of the region 12 in which @ is supposed to be. T h e resulting control input is linear with respect to the error e. /2 = B ( r a ) ..~ + ~ .i=. (The proof of the afore mentioned properties are in [9]). we proceed by considering a Lyapunov function that does not include the p a r a m e t e r estimates: w = ~~ 1 n i=l e. 2) P r o j ( y . = BI.y(1)) + . y) > 0 349 P r o j ( y . (31) .e i . .
B~ t. but our interest is concerned with the bound for ae since the f u n c t i o n a l a p p r o x i m a t i o n t h a t led us to t h e l a t t e r result holds only when z is inside A. where we have chosen A1 _< Ai for every 2 < i < n.'r='yam VArn /XM . Result (33) gives a b o u n d for e.IlY~[I) . To do this we c o m p l e t e the squares in (31) t a k i n g p r o p e r t y 1) of the Proj o p e r a t o r into account: W ___m~ I1~112+ B~ + N(Bs + a) 2 + 4(bM + a) 2 + 0 2 k so t h a t (32) W(t) < W(O)exp(2Alt) + [B~t "Jr N ( B s + c~)~ + 4(bM + 00 2 'k D2]/(2Alk).Yr) > a m l l ~ . supposing without loss of generality A1 > 1 we have from (33) (34) ~/B2t + N ( B ! + c~)2 + 4(bM + a ) 2 + 0 2 Hell < Ile(0)[I + v~ + IIU~(0)II) where AM = and since by (22) Ile(0)ll ___ ~v/iff(ll~(0)ll AMa~(ATA).u r l l 2 where Am = A. } (33) for all t _> 0.n (ATA) is the m i n i m u m eigenvalue of ATA.Yr)T AT A(~ .. by using (34) it follows t h a t II~ll _< ~/B~ + N ( B ! + c~)2 + 4(bM + a) 2 + D 2 v/iZmk 0 + Ily~ll + IIz(0)llvr. Recalling t h a t 2W(t) = ~~n=l e~(t) = Ile(t)ll 2 for every t _> 0 we o b t a i n Ile(t)ll 2 < max{2W(O). Therefore we are going to prove t h a t for a proper choice of the reference signal and of the initial c o n d i t i o n s of the system there exists a value for k t h a t constrains z ( t ) to r e m a i n inside A for every time. so t h a t Ilell > V~mll x .Yrll V~m(IIz[I . Therefore if the control can constrain z inside A for every t i m e when the endogenous a n d exogenous disturbances are m a x i m u m then z will surely r e m a i n in A in every o t h e r circumstance. Marino and Tomei Now we are going to consider the worst case in which l / ( z ) l reaches t h e m a x i m u m value BI.+ Ily~ ( )ll~/. To do this we shall c o m b i n e results (33) and (22): e r e = (~ . a n d the n e t w o r k causes an a d d i t i o n a l error i n s t e a d o f a t t e n u a t i n g the effect of f(a~) on t h e error d y n a m i c s ..N ( B ! + a) ~ + 4(bM + a) 2 + O ~ Alk : = p e2 .350 Del Vecchlo.
2 A 1 W + c2 + D ~ + NIIOll 2 + ~z k (37) from which by integrating up to time t and since ~2 _< 2W(t) for every t _> O.) V. for every t > 0. in order to assure II~(t)ll _< ~.. To show property (iv) consider the time derivative of V = W + 1 ( 1 0 r 0 + 1 b~) 2 71 7~v . we derive that ll~(t)ll < 6 + r(1 + 2.tlllOII 2 + sup~e[o. we use (35) by requiring that 6qr(l+ 2. so that we find W _< . By integrating with respect to time (37) in which we have considered again that ~)2 < 2W(t) we find fo t ~ ( r ) d r < W(0) + fo t ~2 + D~ + k NI[~[[~ + ~ dr that is (iii) with c~(0) = W(0). b and e. we derive that ~P <_2W(O)e:cp(2Alt) + ~ + D2 + Nsup~. a / ~ . + + Since we assumed that A = B(r.Adaptive Control by Approximation Functions 351 Since by hypothesis both yv and a~(0) are contained in B(r).~. A/'XM) < ra V Am  from which it is sufficient that r _< r a  < ra. (36) 1 + 2 V / ~~ From the latter we conclude that once 6 is fixed. then property (i) is assured. by taking k >_ [B~ + N ( B ! + a) ~ + 4(bM + a) ~ + D2l/(6~Am) := /% if both Yr and z(0) are inside B(r).[o..).t]llblr 2k that is (ii) with ax(0) = V ~ . To show properties (ii) and (iii) we complete the squares in (31) in order to account separately for the effects of 0. the ball centered in the origin of radius r. ' Vt > 0  (35) where + NWl + + 4(b.~ .
~ _ i=1 Aie~.V(oo) <_ oo. pp. 39. Arslan and T. 4)it guarantees properties (ii) and (iii) for the tracking error which account for the potential improvement due to a smaller a p p r o x i m a t i o n error and a small p a r a m e t e r estimation error 0. IEEE Conf. 3) it guarantees that the state remains within the approximation region in which the bound on the unknown nonlinearity is given. (ii) and (iii) of T h e o r e m 3. "Universal Approximation Bounds for Superpositions of a Sigmoidal Function". 2) it is repeatable since the initial conditions and the class of reference signals such that properties (i). 1 < i < n. Information Theory . 29872992 Phoenix.1993 3. that implies in particular that limt+oo(y(t) . "Robust Output Tracking for StrictFeedback Systems Using NeuralNet Based Approximators for Nonlinearities". 1989 . "Approximation by Superpositions of a Sigmoidal Function". vol. i=1 Since from (25) and property 1) ei. we can apply B a r b a l a t ' s L e m m a [7] [12] [13] from which it follows that lirnt+ooei(t) = O. References 1. T h e innovative features of the control algorithm are: 1) the number of approximating functions is a free p a r a m e t e r and no requirement is imposed on the optimal value of the corresponding weights and on the approximation error. Marino and Tomei that is ~r _< _ ~']~in___l Aie~ + oT'~en + en R . 3. 303314.( 1 / 7 1 ) ~ T ~ _ (1/72V)~ .yr(t)) = O. A. no. Decision and Control. IEEE Trans. 1999 2. pp. (21). and by (25) and property 4). 1 < i < n. Cybenko. Basar. G. T h e control makes use of orthogonal a p p r o x i m a t i n g functions (10) and of the online updating algorithm for the estimates of the corresponding weights (4). 930945. 4 Conclusions For a class of feedback linearizable systems (1) with A1A4. by assuming e ( z ) = 0 and d(t) = 0 we ~ind ~/ < . G. SpringerVerlag New York lnc. (25). Mathematics of Control. pp. R. a robust adaptive state feedback control (17). are bounded.(1/v)bven .1 hold are explicitly given. Barron.352 Del Vecchio. Proc. Signals and Systems.V(O) . (29) is proposed to solve a tracking problem. we obtain limt~oo E /o'~ Aiei(r)2dr <. (38) Integrating (38) with respect to time.
Automatica. vol. Mar. V. "Robustifying Nonlinear Systems Using HighOrder Neural Network Controllers". Popov. J. 20. Berlin: SpringerVerlag. pp. 1992 10. G. 729740. Kanellakopoulos. 1998 7. and P. V. New York 15. Englewood Cliffs. Nonlinear Control DesignGeometric. L. L. R. 6. pp. IEEE Trans. 1973 13. 1998 9. 44. Ann. Automatic Control . Stable Adaptive Systems. 31. Polycarpou and M. Yesildirek and F. I. vol. 1999 11. K. M. pp. vol. vol. pp. "A simple lemma on greedy approximation in Hilbert space and convergence rates for projection pursuit regression and neural network training". London: Prentice Hall. Mears. Ft. E. Sarmer and J. no. Automatic Control. M. Kolmogorov. Int. Statist. no. Marino and P. 608613. Neural Networks. 1980 5. 1992 12. Slotine. Jones. 1. 3. Kristic. IEEE Trans. Wiley. pp.. J. 43.J. "Gaussian Networks for Direct Adaptive Control". 102108. 1992 . 837863. no. Hyperstability of Control Systems. J. 3. Tomei.. Tomei. pp. A. Armaswamy. IEEE Trans. 70. "Feedback Lineaxization Using Neural Networks". M. vol. Pomet and L. A. Kokotovic (1995). N. S. 8489. K. vol. vol. Adaptive and Robust. 1989 14. 1995 8. N J: Prentice Hall. Lewis. no. IEEE Trans. J. "Stable adaptive tracking of uncertain systems using nonlinearly parametrized online approximators". 1. Automatic Control . A. Elementy teorii funktsij i funktsionai'nogo analiza. "Adaptive nonlinear regulation: Estimation from the Lyapunov equation". "Robust Adaptive StateFeedback Tracking for Nonlinear Systems". M. 11. Maxino and P. S. 37. M. Praly. 16591664. Mosca. Nonlinear and Adaptive Control design.Adaptive Control by Approximation Functions 353 4.Controi. Rovithakis. pp. R. Narendra and A. V. Fomin. 363384. no. 1995 6.
(b) the positive controller gain sequence is ultimately sufficiently small and not of class I l and (c) the reference value is feasible in a natural sense. linear systems subject to a globally Lipschitz. Nonadaptive (but possibly timevarying) and adaptive sampleddata lowgain integral control strategies are derived for asymptotic tracking of constant reference signals for exponentially stable. if a plant is known to be exponentially stable and if the sign of Go(O) is known (this information can be obtained from plant step response data). bath. uk A b s t r a c t . Such a controller design approach ('tuning regulator theory' [4]) has been successfully applied in process control. F1 iegnor~aaths. Ryan Department of Mathematical Sciences University of Bath Claverton Down. All these results are related to the wellknown principle that closing the loop around an exponentially stable. affinely sectorbounded sensor nonlinearity (the conditions on the sensor nonlinearities may be relaxed if the actuator nonlinearity is bounded).13. by lowgain integral control. and Eugene P. In particular. singleinput. 1 Introduction The paper complements a sequence [5. it is shown that applying error feedback using a sampleddata lowgain controller ensures asymptotic tracking of constant reference signals. compensated by an integral controller (see Fig. the problem of tuning the integrator gain adaptively has been addressed in . provided that (a) the steadystate gain of the linear part of the continuoustime plant is positive. for linear (uncertain) systems subject to actuator and/or sensor nonlinearities. nondecreasing. with transfer function Go. finitedimensional. Furthermore. singleoutput. see. Hartmut Logemann. United Kingdom T. singleinput.12].79] of recent results pertaining to the problem of tracking constant reference signals.Sampleddata Lowgain Integral Control of Linear Systems with Actuator and Sensor Nonlinearities Thomas Fliegner.16]). continuoustime. [3. provided that the modulus Ik I of the integrator gain k is sufficiently small and kGe(O) > 0 (see [4. ac. finitedimensional. Bath BA2 7AY. 1). nondecreasing actuator nonlinearity and a locally Lipschitz. then the problem of tracking by lowgain integral control reduces to that of tuning the gain parameter k. will lead to a stable closedloop system which achieves asymptotic tracking of arbitrary constant reference signals. singleoutput linear plant 2Sc. for example. Therefore.
is achieved by (1) provided that the positive gain sequence (kn) is ultimately sufficiently small and is not of class 11. respectively.14] and actuator and sensor nonlinearities in [5]. uo E ]~ .10]. The essence of our approach is to invoke results on discretetime lowgain control which will be summarized without proofs in Sect.14. proofs . in the input and output channel. The classes of nonlinearities under consideration contain standard nonlinear Fig. The purpose of this paper is to derive sampleddata lowgain integral control strategies for asymptotic tracking of constant reference signals for exponentially stable. 2.8. The main concern of the paper is sampleddata control of continuoustime systems. singleinput. 2. 2. our aim is to establish the efficacy of the control structure in Fig. In the adaptive case it is shown that a simple adaptation law produces a gain sequence (kn) such that (1) ensures tracking of feasible reference values. singleoutput. ~ and ~b. with actuator constraints treated in [7. In particular. 1. Sampleddata lowgain control with actuator and sensor nonlinearities ities important in control engineering such as saturation and deadzone.8. (1) where the gain sequence (kn) C ~ is either prescribed or updated adaptively.15] and infinitedimensional systems [7. and L G represents the (discretetime) lowgain integral control law U n + l = Un + k n e n . The main result of the paper in the nonadaptive case shows that tracking of reference values r.5. wherein S and H denote standard sampling and hold operations. which are feasible in an entirely natural sense.356 Thomas Fhegner et al. linear systems with transfer function G~ satisfying Gr > 0 and with monotone nonlinearities. Lowgain control system a number of papers for finitedimensional [2. Fig.
therefore. 2 for achieving the tracking objective for the considered class of nonlinear continuoustime systems.A ) . see [11. (2a) (25) (2c) w. Furthermore. is assumed to satisfy G(1) > 0.I B + D .1 = r + D~(un)) . 2 LowGain Control for Linear D i s c r e t e . has been solved for a class of continuoustime linear systems subject to actuator and sensor nonlinearities. singleinput. be recorded in this section for later use. B. We denote the l ~gain of 57 = (A. ) . then it is readily shown that l+~ReG(Z) z1 >0 Vz6C with[z[> 1 (3) for all sufficiently small ~ > 0. = Czn + D~(u. + ~ = Ax.T i m e S y s t e m s with A c t u a t o r and Sensor N o n l i n e a r i t i e s In [5].U = (A. . C. The performance of the sampleddata controller is illustrated by means of an example. D) I A powerstable. xo e R g .~. Define ~* := sup{~ > 0 [ (3) holds} . B. D) is S := {. (4) . T h e Class S o f D i s c r e t e . i. = r 2. given by G ( z ) : C(zl.T i m e L i n e a r S y s t e m s In (2a).~ + B ~ ( u . G(1) . by lowgain integral control.e. B..) .5]. These results have discretetime counterparts (see [6]) which are crucial in solving the sampleddata tracking problem and which will. the problem of tracking constant reference values. Consider a discretetime.I B + D > 0}. B. j=o If G is the transfer function of a system L" = (A. Thus. D) 6 8.C ( I . C. C. These discretetime results will subsequently be applied to demonstrate the usefulness of the control structure in Fig. D) 6 S by F. A is assumed to be powerstable. the transfer function G.Timevarying and Adaptive Sampleddata Lowgain Integral Control 357 will appear elsewhere (see [6]). Theorem 2. and so OO 0 < G(1) < Fz = ~ ICAJBI + IDl < oo. each eigenvalue of A has modulus strictly less than one.A ) . singleoutput system of the following form: x . y. the underlying class of linear discretetime systems 57 = (A. C.
by r the set of critical values of f .3 The Tracking Objective and Feasibility tracking objective is to ~ such that..f~). ) := _ f 0 ( . but not in A4. Let f : IR + ~ be locally Lipschitz. then r is necessarily in will impose a stronger condition. where ~r := {v E R [ r = r}. By upper semicontinuity of f0. 2VIL(A) := { f 9 A4 I f is globally Lipschitz with Lipschitz constant A}.f(o))~ _< . . For ~.A4 and at least one of the following holds: (i) ~ is bounded. the latter concept of a critical point coincides with that given above. if the closure of im r We Given 27 = (A. in the case of f E A4. r E A/'. by feedback. Clearly. B.is lower semicontinuous.( .r is in A r if 9 A4L(Ai) for some A1 > 0. 2.M" o f A c t u a t o r / S e n s o r Nonlinearities The following sets of monotone. the Clarke [1] directional derivative f0 (~. for functions f E A4.(~) = 0. 0 < min{cgf(~)} = f(~). the output Yn of (2) has the property Yn ~ this objective is achievable.~) := {f 9 ~ I o _< (f(~) . We are now in a position to make precise the class Ar of actuator/sensor nonlinearities: a pair (~o.358 2. v) := limsup o~ hg0 f(O + hv) . 9 := i m p . ~r n 9 5~ 0.1 ) (if f is continuously differentiable with derivative f ' . . Let f E A4: a point ~ E ]R is said to be a critical point (and f(~) is said to be a critical value) of f if f . T h e C l a s s . f . M(. .f(O) h Define f .1 We denote. an input sequence (un) C r E IR. v E ~ . Therefore. A c t u a t o r / S e n s o r N o n l i n e a r i t i e s . nondecreasing nonlinearities are first introduced: A4 := { f : R >R I f locally Lipschitz and nondecreasing} . C. namely. the determine. r 9 .2 Thomas Fliegner et al. A4L(A) C A4(A) C A4. (5) C r i t i c a l P o i n t s a n d C r i t i c a l V a l u e s . D) E S and (~. for given r as n + exp. Clearly. ~ := clos(~) 1 If f is merely locally Lipschitz. v) of f at ~ in direction v is given by fo(~. or (ii) r e A/I(A2) for some A2 > 0.~2 v~ 9 ~}. We remark that. then it would be natural to deem E IRa critical point of f if 0 belongs to the (Clarke [1]) generalized gradient cgf(~) of f at ~. then f .
It may be shown that r 9 7~ is close to being a necessary condition for tracking insofar as. + B~o(un) . r 9 Af. To achieve the objective of tracking feasible reference values r 9 7d. independent of (zo. ~ >p Ilpf=rt~of:=f. ~ + A4. P r e s c r i b e d G a i n . oo]. + D~(un))) . with each p 9 [0. we associate an operator I I p : ./3. = un + kn(r . uo) and (kn) E ~ with the properties lira supn~o o kn < k* as n + oo. solution n ~~ (zn.un + k n ( r r xo E ~ N . Define p* :/~z sup I~1 9 (0. u. We are now in a position to state the main discretetime tracking result in the nonadaptive situation. ~ < p Is < p . (~.U = (A. defined as follows: if p < o o . we will investigate integral control action of the form un+. An application of the control law (6) leads to the following system of nonlinear difference equations Xn+l = A x . then I l v f : (~r f(P) . (gn) is b o u n d e d } . and for every (xo. where N0 := 1~1 U {0}. then r ~ 7~. we describe a convenient family of projection operators. B. C . un+l . There exists k* > O. then if p = o o . (Ta) (7b) Before presenting the main results. T h e o r e m 1.Yn) ." n ~r ~} as the set of feasible reference values.q. D ) 9 S and ( ~ . C. Let . Henceforth. uo) 9 ~ N • I~. D) 9 . we assume that the gain sequence (kn) satisfies (kn) E G := {g I g : N0 + (0. Specifically. ~].4 DiscreteTime LowGaln Control Let ~' = (A. such that f o r all n and Kn := ~"~d=0kj + c~ the following hold for the (7): .~) of the initialvalue problem (kn) 9 ~. r 9 A/" and r 9 Td. f(~) . (6) where (kn) C ~ is a sequence of gain parameters (possibly constant) which is either prescribed or determined adaptively. if tracking of r is achievable whilst maintaining boundedness of ~(un). f(p) . 2.Timevarying and Adaptive Sampleddata Lowgain Integral Control and refer to the set 359 ~:= {r 9 I~. oo).r + O ~ ( u n ) ) ) = Un + k n ( r . with the property IIp o IIp = I I p (whence the terminology projection operator). uo 9 1~ ..
l f ( ~ ) > 0 } . then (un) is bounded. / f ~ r tq ~ = ~ r N 4t. an appropriate gain sequence. A d a p t i v e G a i n . Statement (iv) implies t h a t (un) converges as n + co if the set ~~(~o r) is a singleton. The hypothesis t h a t ( K n ) is unbounded ensures convergence of ~o(un) to an element of ~ r N ~ (and hence the tracking objective is achieved).r + D~o(u. R e m a r k 2.. where to* is given by (4). l0 ~ ( 0 .))l). I f / / ( p . We now consider the possibility of exploiting this output information to generate.I). r ~ C(r and (kn) is ultimately nonincreasing.+x(IrY. lirn~.oo xn = (I .~ol(~or)) = 0 ..+l=l.ooyn = r . limn. then limn~oo dist (un.A)IB~o r . xn+ 1 : 1 t. 10>0. (S) Let X E s and let the gain sequence (kn) be generated by the following adaptation law: k. R e m a r k 1. + ~ ) r E . / f ~ r N 4~ = ~ r N int(4~). by feedback.) . Cr N C(~) = 0. in turn. M linan~co ~(un) =: ~r E ~ r tq ~/i. l i r ~ i n f ~ . Let /~ denote the class of locally Lipschitz functions ~ + + ~ + with value zero only at zero and which satisfy a particular growth condition near zero. where Yn = r + D~(un)). (lOa) u0 E ~ . f locally Lipschitz. Whilst T h e o r e m 1 identifies conditions under which the tracking objective is achieved through the use of a prescribed gain sequence. It is easy to construct examples for which (kn) E ~ satisfies lim SUPn_~oo kn < k* with (Kn) bounded and ~(un) converges to ~o* ~ ~ r N45 (and so the tracking objective is not achieved). c o ) . the resulting control strategy is s o m e w h a t unsatisfactory insofar as the gain sequence is selected a priori: no use is made of the output information from the plant to update the gain. xo E ~ N .l ( O ) = {0}. f . . which. then k* may be chosen as k* = tc*/(A1A~). (10c) We now arrive at the main adaptive discretetime tracking result. Let A~ > 0 be a Lipschitz constant for ~.r ln+~ = In + x(I r .=ld.360 (i) (ii) (iii) (iv) (v) (vi) Thomas Fliegner et al. (9) This leads to the feedback system A x n + B~o(u. then the convergence in (i) to (iii) is of order pK~ f o r some p > 1.ML(A2) f o r some > 0 and some A2 > O. if ~ r tq ~__= ~r N ~. (10b) un+l = u n + l n X [ r . specifically: /2:={fif:]~++lt~+. will be true i f r r fq C(~o) = ~J.
Ce. it will not be possible to track arbitrary r E JR. the solution n ~> ( x n . r E Jff . The underlying class of real linear systems 22e = (Ae.Be. l n ) of the initialvalue problem (10) is such that statements (i) to (v) of Theorem I hold. (~o.C e A ~ I B . i. X E f. M o r e o v e r . 2. is assumed to satisfy Go(0) > 0. xo := x(0) ~ ~ N .) + r as t + co. uo. oo). Given 22e = (Ac. each eigenvalue of Ac has negative real part. the transfer function Ge. B~. B. singleoutput (y(t) E IR).Timevarying and Adaptive Sampleddata Lowgain Integral Control 361 T h e o r e m 2.lo) E 1~N x 1~ x (O. Furthermore. u . (12) has the property y(. ) = (lff 1) converges to a positive value. n E N0.T i m e Systems Tracking results will be derived for a class of finitedimensional (state space /I~N) singleinput (u(t) E I~). the output y(. we require r E 7~e where . Let 22 = (A. De) having a nonlinearity in the input and output channel: J: = A e x + B c ~ ( u ) . for given r E ~ . For each (Xo. Hurwitz. real linear systems 22e = (Ae. Be. > 0}. Ac is assumed to be Hurwitz. 3. a sequence (un) C ~ such that. D) E S. De) E Se and (~o. As in Sect. Ce. c~)).) of (11).r E A/'. 3 SampledData Control of Linear S y s t e m s with Actuator and Sensor Nonlinearities In this section we shall apply the results of Sect. for the class of systems introduced below. Be.) given by u(t) = un for t E [nr.A c ) . / f ~ " n ~ = t r n4~. by sampleddata integral control. Ce. Ge(0) = De .(n + 1)r). r > 0. y= r = r + De~(u)) . given by Go(s) = Ce(sI . . and r E 7~. D~) is denoted Se := {Se = (Ae. C. w = C~x + Dc~(u) . (lla) (llb) (11c) In (lla). g'r nC(~o) = 0 and r q~ C(r then the monotone gain ( k . resulting from applying the input u(. De) [ A . continuoustime (time domain ~ + := [0.I Be + De . 2 to solve the continuoustime lowgain tracking problem.e.. the tracking objective is to determine. Instead. Ce.1 T h e Class o f C o n t i n u o u s . by sampleddata control.
U = (A. (14a) (145) (14c) yn = r where A = e A~ . ~r~ : = { v E ~ i r 4~:=im~. (13) the discretization of (11) is given by xn+l = A x n + B~o(un) . then r 9 7~e.. it m a y be shown that r 9 7~c is close to being necessary for tracking insofar as. defining x .e A ~ It follows from (16) that ~r = ~ r and T~=7~e. + D~o(un) . for t E [nT. i. C.3 Prescribed Gain We first treat the case of a prescribed gain sequence (kn) E G. uo 9 ~ . magnitude strictly less than one. D) of Z:c = (Ae. Xo := x(O) 9 ~ g .(n + 1)r).~ = Cx. 3. w. (16) where ~r = {v 9 ~ I r = r} and 7~ denotes the set of feasible reference values for the discretization .> 0. if tracking of r is achievable whilst maintaining boundedness of ~(u). .2 Discretization of the ContinuousTime Plant Given a sequence (un) C ~ .. consider the following sampleddata lowgain controller for (11) u(t) = un . wn := w(nT) and yn := y ( n r ) Vn 9 N0.362 and Thomas Fliegner et al. D) satisfies G(1) = C c ( I . Furthermore. C. Be. B.I)A'[1Be + De = ae(0) . we define a continuoustime signal u(. the transfer function G of L' = (A. un+l = Un + k n ( r  (17b) (17c) . then. B. B = = r (e A ' r . C=Ce and D=De. Again..e. (17a) Yn = y ( n r ) .) by the standard hold operation (12) where 7. If this signal is applied to the continuoustime system (11). Ce. ~:clos(4~). (15) Since Ae is Hurwitz. + D~p(un)) .> 0 denotes the sampling period. For this purpose. n E No. n E N0. the matrix A is powerstable. for arbitrary 7. De).I ) A [ I B e . the eigenvalues of A have. yn) . 3. := x(nT) .
D) given by (15). for t E [nr. B.) is bouoaed. then limt~oo dist ( u ( t ) . such that for all (k.~o(u(t))) . the convergence in (i) to (iii) is of orderexp ( . Be. It follows from the previous subsection that xn. (i) Statement (iv) implies that u(t) converges as t 4 cx~ if ~z~ n s = 0. (iii) If H(p. we m a y infer that the tracking objective is achievable by a constant gain sequence with kn = k E (0.Timevarying and Adaptive Sampleddata Lowgain Integral Control T h e o r e m 3. as n 4 cx~. n 4" O O Since the continuoustime signal u is given by (12).~) = 9 r E ~ r f3 ~ . where K : I~+ + ~+ is defined by K(t) = K . lim 9(u. Ce. Hence S = (A. independent of (Xo.(A2) for some > 0 and some A2 > O. De) E S. wn. if~P~ O # = #cr n int(#). by part (i) of Theorem 1.a K ( t ) ) for some a > O. limt~oo x(t) = . Uo) E ~ v x l~ the following hold for the solution of the closed loop system given by (11) and (17): (i) (ii) (iii) (iv) (v) (vi) limtooo ~o(u(t)) =: ~v~ E ff'ff t3 # . if#~ n ~ = qJ~ (1 ~. Assertion (iii) follows from (i). Assertion (ii) is a consequence of (i) and the Hurwitz property of A. uo) and (kn) E ~. 9 . Moreover. where S = (A. where y(t) = ~b(C. limt~or y(t) = r . then k* may be chosen as k* = ~*/(AtA~).. (ii) and continuity of r . assertions (i). kn tends to zero sufficiently slowly in the sense that (kn) ~ l ~(~+). Thus. Yn (given by (13)) and un satisfy (14) with (A. C. (n 4.+~)r E . Let Sc = (Ac. ~e n C(9) = O and r ~ C(r the. C~. where A~ > 0 is a Lipschitz constant of 9Proof of Theorem 3. There exists k* > O. (ii) An immediate consequence of T h e o r e m 3 is the following: if (k~) E 6 is chosen such that.~) E G with the properties lim supn~c o kn < k* and Kn := ~]~'=0 kj 4 as n 4 c~. C. C.1)7). B.. r Define 363 E Af.A/It. ~ff N ~ = #ff n 45. Be.~*/(AzA2)) for all n E No. D) denotes the discretization of 2Yc = (Ae. ~].~ ( ~ ) ) = 0 . A is powerstable and G(1) = Go(0) > 0. and for every (xo.A [ 1 B ~ 9 r . (9. Let At > 0 be a Lipschitz constant for 9. where to* is given by (4).x(t) + D. then the tracking objective is achieved. then u(.+6)~b E AdL(A~) for some ~ > 0 and some A~ > 0. Remark 3. D) E S. If H(p. De) E S~ with sampling period v. (iv) and (v) follow. B. v > 0 and r E Tr p* := F~ sup I~1 ~ (0. if (kn) is ultimately nonincreasing.
We m a y now conclude that IIr _< m a x IleA~ 0<s<r m a x IleA~ O<s<r ' t 9 [nr.~'rll < M P . Define r := x(t) . C. + l n l ( r ln+l = l n + x ( l r . Yn) .IV'.r < M P . consider the following adaptive s a m p l e d . [1 . 7" > 0 and r E T~c. From the convergence of order e x p ( . ( . De) E S~ is of class S.K " for a l l .(n + 1)7"). T h e discretization . n E N 0 . invoking the hypothesis r ~ s162 and the local Lipschitz property of r that the convergence in (iii) is also of order exp ( . rl 3. ~ ) . r the monotone gain sequence (kn) = (In 1) converges to a positive value. (~. ~ o ~. If moreover k~ O ~ = k~ N 4~.U = (A. D) of . ( l n ) ) of the closedloop system given by (11) and (18) is such that statements (i) to (v) of Theorem3 hold. Be.a K . Co.y . ~o(u(t)) = ~o(u. !b) E . (n+l)r) which proves that convergence in (ii) is of order exp ( . note that by T h e o r e m l(vi). ) .4 Adaptive Gain In this subsection. (18a) (lSb) (18c) (18d) Un+l = u . where ~ = .. De) E So. I ) .oo)..U~ = (Ac. Then ~(. For all (xo.364 Thomas Fliegner et al.lo) E ]RN x ]R • (0. l0 E ( 0 .a K ( t ) ) .uo.x r. T h e o r e m 4. ~ O C(qa) = 0 and r r C ( f ) . our goal is to establish the efficacy of the a d a p t a t i o n law (9) to generate an appropriate gain sequence in the s a m p l e d .) satisfies ~(t) = Ace(t) + B c ( ~ ( u ( t ) ) . where a = l n p > 0.~r) and thus r = eA~162 + I' ~A~ 7" .K " = M exp ( . Hence. n E No . Let ~c = (Ae.). P~ooS. the theorem is an i m m e d i a t e consequence of T h e o r e m 2.r t e [nT.a K ( t ) ) in (i) and (ii). uo 9 ~ .+ 1)~). the convergence in (i) is of order exp ( . e N0.a K ( t ) ) . it is straightforward to show.d a t a setting. Thus. Be. B. u ( ' ) .) and so ko(u(t))~o~l _< M p . Therefore. the solution ( x ( ' ) .A Z l B . To show (vi). Co. there exist M > 0 and p > 1 such that I~o(un) . ( . + 1)7. for t E [nT".K " and II~'n . Yn = Y(nr) .d a t a lowgain controller for (ii) u(t) = u n . For all t e [7.a K ( t ) ) .
b 3.1] u>l. For purposes of illustration.a < r < b 3. x~(0). Let X E s By Theorem 4. 7. ~=ax~bxl+~(u).10) : (0. 1]).Timevarying and Adaptive Sampleddata Lowgain Integral Control 3. . The convergence of the gain to a positive limiting value is evident. let a : 2. w=xx. we have ~P~" A ~ :fi 0 if and only if .1 .. O. 3. Moreover. "["'. 1) 0. b > 0. the sensor nonlinearity r E A4 is the cubic r : w ~ w z.d a t a control. O. bZ].: 1 and let X : ~[~+ (the restriction of the saturation function to ]~+). 1].A4L (1) is of saturation type. if r is both nonzero (so that r r s 1 6 2 = {0}) and in the interior of 7d~ (so that ~ f3 C(~) " r then the a d a p t i n g gain (kn) converges to a positive value. u0. / 10 20 30 TKne 40 50 80 O.b z.. The transfer function Gc of the associated linear system L'~ is given by G~(s)  s~ + as + b 1 . with Gr 1 = ~ > 0 Since ~ r = {v E ~ I r = r} = {brl/3}. 0.5 Example 365 Consider the secondorder system xl = x ~ . Fig. 1 . b : 1 (in which case 7~ = [ .b . u < I u~~(u):= u. the set T~c of feasible reference values is given by Tdr : [ . 0. and the actuator nonlinearity ~ E . ue[1.d a t a controller (18) achieves the tracking objective for each feasible reference value r E 7~ = [ . Performance under adaptive sampleddata control and the feasible reference value r = 0. it follows that the adaptive s a m p l e d . For initial d a t a (xl (0).1 . bZ].4 Fig. y=r where a. 3 (generated using S I M U L I N K Simulation Software under MATLAB) depicts the system performance under adaptive s a m p l e d . Thus. and so ~ = [ . defined as follows 1 .75.
Computation. u k / p u b / p r e p r i n t a . J. Control 30. Coppus G. of Int. Control & Optim. T h i s work was s u p p o r t e d by t h e UK Engineering & Physical Sciences Research Council ( G r a n t Ref: G R / L 7 8 0 8 6 ) . Logemann H. (1997) Discretetime lowgain control of uncertain infinitedimensional systems. J.. Logemarm H. Information. 19401961 10. SIAM J. and Davison E. L. IEEE Trans. b a t h . 35.. Morari M. References 1. and Systems) 9. 574577 I Preprints in this series are available on the World Wide Web at the URL http://www.. (1983) Optimization and Nonsmooth Analysis. D 130. Signals. 26 13. (1992) Controllers with universal tracking properties. Fliegner T. SIAM J. 11201144 8. and Townley S. Lunze J. Proc. 36.uk/MATHEMATI CS/preprints.ac.. and Townley S. A. Ryan E. Pt. (1989) The selftuning robust servomechanism problem. Clarke F. (1983) Robust multivariable control of a binary distillation column. M.. Miller D. Logemarm H. 78116 11. Logemann H.366 Thomas Fllegner et al. (2000) Timevarying and adaptive integral control of infinitedimensional regular systems with input nonlinearities. (1998) Integral control of infmitedimensional linear systems subject to input saturation. Mathematics Preprint 99/241 . IEEE Trans. (submitted for publication) 6.maths. Manchester 3. Miller D. (1976) Multivariable tuning regulators: the feedforward and robust control of a general servomechanism problem. Lunze J. (to appear in Mathematics of Control. P. Cook P. P. Prentice Hall. Mathematics Preprint 98/201 . University of Bath. Control & Optim. (1985) Robust stability of systems with integral control. Fliegner T. 3547 5. Control 34. (1997) Lowgain control of uncertain regular linear systems. Control 21. K. a c . and Ryan E.html and by anonymous ftp from f t p . IMA Conf. lEE Proceedings. SIAM J. Logemann H. (1988) Robust Multivariable Feedback Control. (1999) Discretetime and sampleddata integral control of linear systems with input and output nonlinearities. . (1987) Experimentelle Erprobung einer Einstellregel fiir PIMehrgr6flenregler bei der Herstellung von AmmoniumnltratHarnstoffLSsung. London 14. (1999) Lowgain integral control of continuoustime linear systems subject to input and output nonlinearities. W. and Townley S.. J. Auto. P. Automatica 29. m a t h a . and Ryan E. Control & Optim.bath. Mathematics Preprint 99/261 . H. (submitted for publication) 7. (1998) Timevarying and adaptive discretetime lowgain control of infinitedimensional linear systems with input nonlinearities. Sha S. New York 2. P. and Ryan E. 2237 12. 201208 4. Auto. IEEE Trans. 511523 16. E. 38. IEEE Trans. P. and Wood R. University of Bath. Auto. University of Bath. 877887 15. and Ryan E. Messen Steuern Regeln 30. Acknowledgement. Wiley... Logemann H. Logemann H. Control 42. on Control: Modelling. E. Auto. Davison E. and Davison E. (1993) An adaptive tracking problem with a control input constraint.
n o n l i n e a r observers certainly r e m a i n one of the most challenging control problems. When utilized in conjunction with static state feedbacks they permit to bypass classical asymptotic observers. 2. '~ On temporary leave from Departamento de Sistemas de Control. France {~arquez. a classical PID controller where we replace the derivative term by appropriate integral ones. and E m m a n u e l Delaleau 2 1 Centre de Math~matiques et Leurs Applications Ecole Normale Sulc~rieure de Cachan 61. fr Laboratoire des Signaux et Syst~mes CNRSSup~lecUniversit~ ParisSud Plateau de Moulon 3. enscachan. Three illustrative examples are examined: 1. 3. delaleau} @iss. Richard Marquez 2. fr A b s t r a c t . We are here launching a new p r o g r a m m e for trying to elucidate this topic. France f liess@cmla. which is mainly of algebraic flavour. Venezuela. We will of course start with c o n t i n u o u s . For constant linear systems we introduce the class of exact integral observers which yield generalized PID regulators with good robustness properties. One author (RM) was also partially supported by the Consejo Nacional de Investigaciones Cient{ficas y Tecnol6gicas (CONICIT).State Feedbacks Without A s y m p t o t i c Observers and Generalized P I D Regulators* Michel Fliess 1.t i m e constant linear systems not only for obvious simplicity's Work partially supported by the European Commission's Training and Mobility of Researchers (TMR) under contract ERBFMRXTCT970137.**. 1 Introduction In spite of m a n y f u n d a m e n t a l works. a generalized PID for a second order system with a nontrivial zero dynamics. Our approach. M~rida 5101. especially nonlinear ones. Universidad de Los Andes. a real DC motor without mechanical sensors. rue JoliotCurie 91192 GifsurYvette. . We conclude by discussing possible extensions. is based on the moduletheoretic framework for linear systems and on operational calculus in Mikusifiski's setting. Facultad de Ingenien'a. Venezuela. supelec. avenue du President Wilson 94235 Cachan.
which rests on a novel exploitation of observability. (3). In this interpretation. which is based on the operational calculus via the twosided Laplace transform (see. which we call an exact integral observer.0 % +cxz) is not admissible.~ . . . [30]). e. . where f is some time function. by expressions where we only deal with measured quantities and their iterated integrals.. we know that any c o m p o n e n t of the state z = ( x l . Consider the constant linear i n p u t . which is equivalent to observability (cf. . and E.. From the constructibility property. xn) m a y be expressed as a finite linear c o m b i n a t i o n of the input and output variables. [23]). Fliess. Marquez. U m ) and y = (Yl.o u t p u t system s~ = A z + B u y = Cz where s stands for 3 " Rewrite (la) as (la) (lb) ~=A~+B s u$ By induction we obtain for any p > 1 /J z = AUsUx + ZAiIBsiu i=1 (2) Assume that (1) is observable.. R.e.. to E ( . i. 1 The very same convention was also employed by the main father of operational calculus. u = ( u l . e. . Those integrals will be given by equating to f*o~ f ( v ) d r . Delaleau sake but also for explaining our viewpoint where a s y m p t o t i c observers are replaced by exact integral ones. for p large enough. Formula (3) means t h a t the c o m p o n e n t s of the state m a y be recovered via a finite linear combination of iterated integrals of the input and output variables. . [30]).g. . Oliver Heaviside (see. Thus (2) yields. and of their derivatives. a nonzero constant function on ( . .Yp).g. 9 Xn = P + Q U (3) where P and Q are respectively n • p and n • m matrices whose entries are polynomials in the variable s 1. This endeavour. m a y be regarded as a feasible alternative to classical asymptotic observers. This leads to the following convention I which is one of the roots of our approach: A n y c o n s t a n t l i n e a r s y s t e m is a t r e s t f o r t _< to. +r162 Then.368 M. . is completing almost ten years of fiatnessbased control synthesis where the core was a revisited notion of controllability (see [14] and the references therein).
. where the ground ring R is no more a principal ideal domain (cf. at least. It satisfies r .o o but some finite tl. Consider as a matter of fact a torsion element r. Consider. with one more ingredient.g. and sliding mode [9. Our first m a t h e m a t i c a l tool is the moduletheoretic setting for finitedimensional linear systems. namely localization. The only solution of this equation is 0 and r would be pointless.33]). for (4). which is classic in c o m m u t a t i v e algebra (see. [22]) and was already extensively employed for delay systems [18]. 3 There exists a vast literature on robust asymptotic observers.. 4 For more general systems like delay ones.19]). This remark applies to the operational calculus based on the onesided Laplace transform where one should replace (la) by sz = Az + Bu + x(0) (5) Note that in this case (3) is no more an exact observer. a perturbation variable. The second tool is operational calculus in Mikusifiski's setting [28. disturbance estimation [21]. Its robustness 3 with respect to perturbations is obtained by adding a finite linear combination of iterated integrals of the o u t p u t variables. Those generalized PID regulators are defined for multivariable systems (compare with [1. freeness should be replaced by torsion freeness. for some reasons. s . This communication is organized as follows.29] (see also [10. but it also provides the features we need of operational calculus via the twosided Laplace transform. the lower bounds of integration in the iterated integrals is not . what we call a generalized PID regulator 2.32] approaches. In particular robustness is satisfied when. We define the abstract notion of regulator. e.1 ] . r ~ 0. for example.Generalized PID Regulators Consider the static state feedback 369 = U (4) where the gain/C is an rn x n matrix. without analytical difficulties. 5 Besides approaching stability and robustness under the very same perspective. Operational calculus leads us to the second convention: The module corresponding t o a c o n t r o l s y s t e m is f r e e 4.35]). r ~ [ s . observer and controller designs are now included in a unified controller synthesis. [18. It implies that any system variable is influenced by. Stability analysis becomes in this setting a consequence of robustness with respect to some peculiar perturbations 5.0. It not only permits to deal with module theory in a simple and natural manner. as developed since [11]. In the next two sections we review the moduletheoretic setting. Replacing x by its expression in (3) yields.
the next one is devoted to stability and robustness. Generalized P I D regulators are introduced in Sect. of rank m. Sects. T h e i n p u t u is said to be independent if. is a finitely generated R .33] we are replacing the derivative term sy by a suitable linear c o m b i n a t i o n of integral ones of the form s . 4. . We will of course not use it in order to avoid any confusion with PIDs in engineering! 7 An element ~ of an Rmodule M.free if.1. If R is a principal ideal domain and if M is finitely generated. A n inputoutput Rsystem $ is an R . called the torsion submodule. is said to be torsion (see. A n Rlinear system A. there exists e E R.19] and adapted if necessary to the case of modules over principal ideal domains. The set of all torsion elements is a submodule Mtor.11. such that e~ = 0. 7. . freeness and torsion freeness are equivalent (see. Yp) C S. In the conclusion we sketch some possible extensions of our work to delay systems and especially to nonlinear ones.m o d u l e . and E. 5.. which is equipped with an input u = ( u l .d y n a m i c s equipped with an output y = ( Y l .18. In the classical PID regulator [1. Fliess. e. where R is a commutative domain. although we are r e t a i n i n g the advantages of PIDs with respect to the transient response. 6 The usual mathematical abbreviation for a principal ideal domain is P1D. The module M is said to be torsion if. ao+als+s~ 3. and only if. It is torsion . or an Rsystem for short. 2 S y s t e m s over P r i n c i p a l Ideal D o m a i n s W h a t follows is.g.m o d u l e s p a n R ( u ) is free.s y s t e m .2. with the exception of the regulator. . 2.. . and only if. [22]) if. . We are i n t r o d u c i n g a generalized PID controller for a second order system given by bolbls i. 2.370 M. and 8 are e x a m i n i n g three illustrative examples: 1. A real DC m o t o r without mechanical sensors. Delaleau and the statevariable realization. . A n Rlinear dynamics I) is an R . u m ) C 7) such that the quotient module 1 ) / s p a n R ( u ) is torsion r. .e. 7. and only if. After a short overview of Mikusirlski's operational calculus in Sect. borrowed from [4. This condition is assumed to be satisfied in the sequel.13.1 Generalities Let R be a principal ideal d o m a i n 6. for example.t r i v i a l zero dynamics. . Marquez. R. . with a n o n . M = Mtor. We thus avoid taking the derivative of the measured q u a n t i t y like in the PI regulator. e ~ 0.l y and slu. the R .16. and only if. [22]). Mtor = {0}.
* N is free.1 R  from the category of finitely generated Rmodules to that of finitely generated S 1 Rmodules is called a localization functor (see.g. [22]) of Rmodules O~ N~ A~ T~ O where 9 T is torsion. [22]). y) coincide.4 Regulator An Rlinear regulator of the Rsystem A is a short exact sequence (cf. 2. For any A E A. the Rmodules S and spanR(u . and only if. 2.. The Slocalized system of a controllable R . The next result is clear.t R  A is the Slocalized system.s y s t e m is also controllable. Take an inputoutput Rsystem S.m o d u l e A is free. . sA .1 R . The functor S . This is equivalent saying that a torsion element is not altered by a regulator. output) is the Slocalized input (resp. e. The Slocalized system of an observable inputoutput R . the S .E T. The converse does not hold in general. Any basis of the free module A is called a flat output.2 Controllability and Flat Outputs 371 The Rsystem A is said to be controllable if. for any 7. a c = 0. the R . and only if. output) of 8.s y s t e m is also observable. For any Rsystem A. The Slocalized system s S is also inputoutput: its input (resp. The freeness of N is equivalent to the following fact: the restriction of the mapping A + T to Ator is onetoone.i R .m o d u l e T such that there exists ~ E S verifying. Consider the localized ring S .3 Observability The inputoutput Rsystem 8 is said to be observable if. P r o p o s i t i o n 1.1 Q A E s A is the Slocalized element of A.s y s t e m s A = S . 2.5 Localization Let S be a multiplicative subset of R. This functor is not faithful as it maps to {0} any torsion R .Generalized PID Regulators 2.
Delaleau Example I.~rq). Then.. This functor is not faithful. S . If the multiplicative set S above is R\{0}. Marquez.6 Perturbed Systems An Rsystem A pert is said to be perturbed if. where ~ denotes the canonical epimorphism :Dpert _. . urn) and of the perturbation variables vr such that 9 the control and perturbation variables are assumed not to interact.1 Classic StateVariable Representation Unperturbed StateVariable Representation Here R is the principal ideal domain IR[s] of polynomials over the field ~ of real numbers in the indeterminate s (s will of course in the sequel correspond . For a perturbed Rdynamics ~)pert. .4) is said to be not affecting the perturbation if. A matrix H E KpXm such that is called a transfer matrix of S. we will therefore not distinguish in the sequel between ~v(u) and u.372 M. The unperturbed Rsystem corresponds to the quotient module A = Apert/spann(n'). and only if.e. H is so. {0}. From spann(u ) f3 spann(rr ) = {0} it is immediate that the restriction of to spann(u ) defines an isomorphism s p a n n ( u ) ~ s p a n n ( ~ ( u ) ) . 9 the quotient module T~/spann(~(u)) is torsion. A ~r 1  A = A is the (formal) Laplace transform. tim is a basis of A. The Kvector space K  A = A is called the transfer vector space of the Rsystem A. the intersection of the image of N in A with spann(Tr ) is trivial. .m o d u l e to {0}. 3 3. the input is the union of the control variables u = (ul. and only if. and only if.e. i.. is called the Laplace functor [13] and corresponds to the simplest type of localization (see.. 2. the inputoutput R . spanR(u ) fq spanR(Tr ) = {0}. [22]). and H is uniquely defined.. R. right invertible) if.. f i t . ~rl. .v e c t o r spaces. it maps any torsion R .. . The functor K from the category of finitely generated Rmodules to that of finitedimensional K . If u is independent. has been distinguished.. By a slight abuse of notation. e.e. i.g. and ~ is the (formal) Laplace transform of A [13]. and E.s y s t e m $ is said to be left invertible (resp. A regulator (2. a perturbation 7r = (lrl.. .+/). .... The mapping A 4 A..~rq E A.1 R is the quotient field K of R. . Fliess. i.
is not affected by (4). and u~ ~~ . 3. It is free.e. since the uncontrollable K a l m a n subspace. Proposition 2.o u t p u t ~ [ s ] .. of a well known fact in control engineering.y. where K = (kz~).4} is defined by s = (A + B ~ ) (7) T h e module N C ~q is the kernel of the m a p p i n g S + T defined by (4) and (6a).s y s t e m S m a y be defined by the Kalman state variable representation [11] s = A " ~n + B " ~tm (6a) = c " + D~.~.s ~ " 1~m (6b) ~rt r where A E ~n• B E ]~n• C E ]~p• Da E ~P• We know that the dimension n of the state vector is nothing else than the dimension of the torsion module ~q/span~[sl(U). the torsion submodule. i.y kZ3x. An i n p u t . when viewed as a ~ .Generalized PID Regulators 373 to the t i m e derivation d/dt).e. . 3. T h e torsion ]~[s]module T in (2. Proof. by x~ ~+ x . in our language.v e c t o r space.2 T h e S t a t i c S t a t e F e e d b a c k as a R e g u l a t o r T h e next result is a reformulation.3 Perturbed StateVarlable Representation Pulling back (6) from S to S pert yields [17] (ert) " " s A " +B + (8a) C " ~ D~s ~ + (Sb) ~pert k xnpert / a=O U . i.. The static state feedback (4) is a regulator.
The order of e E ~[s. is the localized ring of ~[s] with respect to the multiplicative subset {s ~' I v > 0}. .l ] . The next result will be useful for the robustness analysis of Theorem 2. . wl. Proposition 3. is injective. Remark I. R.1 Generalized sLocalization PID Regulators The principal ideal domain of Laurent polynomials ~ [ s . . From now on and for the sake of simplicity we will consider a controllable and observable I~[s]system S with input u = ( u x .v e c t o r space spanned by the residues of the components of y in ~q/spanR[s](u ) is pdimensional. . From rk(C) = p. W i t h a slight abuse of notation. i. by definition. a~.s y s t e m A and sA E ~A the slocalized element of A E A . . ~n E span~[.10r). Fliess. ) the corresponding o u t p u t matrix C = (cij) verifies _ fl cij= ~ 0 i=j. s~c E ~[s]\sI~[s]. such that Proof. were already utilized in [17] for obtaining the perturbed Brunovsk~j canonical form s. + ~ . t~ E 2 . ~ . and E. Then there exists a state variable representation (8) such that the w~s in (Sb) are identically O. E ~ . we m a y choose a state ~? = ( $ 1 . 4 4. The order of 0 is. The canonical ]~[s]linear morphism st : 3 r sS. e 5~ 0. The o u t p u t y of S is said to be good if. .. . w . Yp). .. we will therefore consider S as a subset of sS. Consider a perturbed inputoutput ]R[s]system 8 pert such that the output of the corresponding unperturbed system S is good. Marquez. iCj l<i<p Then an obvious perturbation dependent state transformation yields the answer. Delaleau where w l . Perturbation dependent state transformations. urn) and o u t p u t Y : (Yl. > 0}. which are finite sums of the form )~a~s ~. . Call slocalization the localization with respect to {s v ] t. . s1]. .374 M. A more intrinsic definition reads as follows: T h e b . . ~ ~+ sA. . 8 This perturbed Brunovsk3~ form was the key ingredient for understanding the robustness of our flatnessbased predictive control strategy (see [17] and the references therein). s . . is a E 2~ such that s~e belongs to ~[s] and possesses a nonzero constant term.. the m a t r i x C in (8b) is of rank p. which are quite natural in our moduletheoretic setting. . We assume u to be independent.e. . and only if. . Write ~A the slocalized system of the ~ [ s ] . A sgeneralized PID regulator of S is a regulator of sA.
The static state feedback (4) may be written as an s generalized PID regulator Remark 2.A4 of C.A4 is an operator. The next corollary follows from Proposition 2. . Call Mikusifiski's field the quotient field . The neutral element 1 of A4 is the analogue of the Dirac measure in L.v)dl" = J0t g ( v ) f ( t . s l] is the smallest order of its entries. . Corollary 1 m a y be of course extended to other types of state feedbacks. The order 9 of P may be chosen to be any integer > O. from (2). V G ~[s. .s1] nxp. 9 The order of a matrix with entries in ]~[s. The Kalman state in (6a) may be expressed by (3) where P E ]~[s. Example 2.v)dv It follows from a famous result due to T i t c h m a r s h (cf.g(t) and the convolution product * (f . .s1] nxm.Generalized PID Regulators 4. The order of Q may be chosen > 1. we obtain the following crucial theorem.29. T h e o r e m 1. Sehwartz's distribution theory. xn are R[s]linear combination of the components of u and y.1 Some Algebraic Analysis Mikusifiski's Operational Calculus Consider the commutative ring C of continuous functions [0. From the observability assumption we know that zl.2 375 Static State Feedbacks as PID Regulators For any p > 1. 5 5. C o r o l l a r y 1. Thus.35]) t h a t there are no zero divisors in C. [28. Any element of . equation (6a) yields (2). .g)(t) = (g* f)(t) = J0t f ( v ) g ( t . + ~ ) + C with respect to the addition + ( f + g)(t) = f(t) t.
b) = 1.s y s t e m A is an R . T h e subfield C(s) C . a trajectory with left bounded support.f u n c t i o n f : ~ + C with left bounded support. sl]). T h e n s f is the derivative ~ . with left bounded support. according to the introduction. Take a mapping of the perturbation variables in . R. Write j ~ l e f t C J ~ the set of regular operators with left bounded supports.4left is also an R . The exponential e ~ . or a j ~ l ereg. .m o d u l e (this is verified .~pert such that the corresponding unperturbed system S verifies the following two properties: 9 it is controllable and observable. 11 See [20] for connections with controllability.hd such that * their images are finite sums of operators of the form ~ e a. Marquez.~ = f : ~ g(~)d~ has also a left bounded support. of the R . We refer to [29] for the notion of regular operators and their supports.376 M.2 Trajectories of Mikusifiski's Systems A Mikusi6ski's Rsystem is an R . T h e n . ~ t~ E * the real parts of the roots of b are < O. Take for instance a locally Lebes~ue integrable function g : ]~ ~ C.m o d u l e morphism A + AJ.s y s t e m A where R is a subring 1~ of. and E. The inverse in Ad of the Heaviside function is the operator s which satisfies the usual rules of operational calculus. This is the case for ~[s] or ~[s. b E ]K[s].h4trajectory 31 of A any R . Take a C l .s of rational functions in the indeterminate s with complex coefficients possesses the usual meaning. If "h. Example $. Fliess. Theorem 2. b # 0." reg with R = ~[s] or R = ~[s. Delaleau Example 3. 9 the output is good. (a.A4._. Take a perturbed inputoutput ~[s]system ~. Note that a locally Lebesgue integrable function with left bounded support belongs _utJ"tregA ~left. 5. Following [12] call . represents the usual delay operator.m o d u l e morphism A + /~/left" reg 6 Robustness and Stability The next result is an extension of familiar properties of PIDs.Mleftregis reg a ring with respect to the addition and the convolution.tra" " ft Jec~ory. 10 The Rmodule A is. . assumed to be free. sl]. Its inverse e ~ is the forward shift operator. v~ ~ 0.
~ 0. where ~(0) should be regarded as a perturbation which.a y j for a > 2. we want to transfer an initial equilibrium point to another one by following a desired trajectory. such that the canonical images of the components of the output are equal to finite sums of operators of the form a~e ~.~ w E ~ . j = 1 . In order to counteract it. The integer A is chosen such that the resulting expression for yj does not exhibit any pole at S~0. 7 Implementation of a PID Without Derivator Figure 1 is representing a P I D controller acting on an i n p u t . may easily be counteracted. control and perturbation variables. . Remark 3. .l y j ) = yj where we replace yj by its expression with respect to the state. C o r o l l a r y 2. where the real parts of the roots of d are < O. according to the previous result. 12 This perturbation is in fact a step function. The corresponding equation reads s ( s . for any ~ > O. there exists an sgeneralized PID regulator. d r e~. apply Proposition 3 and extend the state variables with a finite number of terms of the form s . which is not affecting the perturbation. p . . which is not affecting the perturbation. Then. then there exists an sgeneralized PID regulator. stability analysis is being included as a particular case in the study of perturbation attenuation. such that the canonical images of the components of y may be written as finite sums as above and of a C ~ . such that the canonical images of the components of the input and the output are as above. If the real parts of the roots of b are < O. Take a mapping of the perturbation variables in . Assume that the measurements start at t > 0.~ y j . 1 < a < A. which is not affecting the perturbation.A4 such that their images are finite sums as above and of a C ~ . c. As already stated in the introduction. . d) = 1. By utilizing high gain techniques we obtain the following corollary. Following the flatnessbased philosophy of predictive control [17].f u n c t i o n f ( t ) such that limt~+oo [ f ( t ) 1< ~. For a = 1 we obtain s ( s .~ y j ) = s l .o u t p u t system subject to a constant perturbation 1~ 7r = w e . ~ > 0.Generalized PID Regulators 377 Then there exists an sgeneralized PID regulator. Applying Corollary 1 to S pert yields an error due to the perturbations. to 0 for t < ~. It is equal to m for t > O and. d E ~ [ s ] . Proof. Note t h a t if derivatives of the control variables appear they might be eliminated by the s t a n d a r d procedure of [11].f u n c t i o n with left bounded support and which is bounded as well as all its derivatives. S pert is as above. Then one is left with (5). (c. . . Place the poles of the extended system by a static state feedback where the state variables z are dealt with via Corollary 1. according to our fundamental assumption.
 s (9) Utilize a classical P I D controller [1] for pole placement Ue = kDSye .Ye + ue S S a0 1 (11) Substituting this expression in (10) yields Ue = kD ( . thus.y ~ . Fliess. Marquez.v a .~~2 ye (12) . From (9). and ue = u . 1.s Note that y is a flat output of the unperturbed system. Nevertheless. We obtain e. and E. a generalized PID controller is to be designed. where the disturbance is taken to be 0. Delaleau 7r Fig. Define the errors Ye = Y . several approximation methods can be used (see [1]). we have sy~ = . a double integral term in the P I D controller has to be added.Y*. sensitivity to high frequency perturbations is their m a j o r drawback.kpye .gs s~y~ + a l s y e + aoYe = ue 4. ke and kl are the coefficients related to the derivative. y*) satisfy the nominal system.. T h e quantities (u*. A P I D controller of the form u~ = .a l Y e .s Y e ke)ye a0 + 1 Ue ) .kl +   8 Ye  which does not include any derivator.a l s y + a o y = u Jr w . The block diagram of a PID regulator 7.~ .a l y e .Yes kD Ue 8 kl = (alkD aoko . In the i m p l e m e n t a t i o n of the derivative.kr.keye .u*..k D s y e .378 M. ye .Y e S kl (10) where kv.1 A Simple Linear S y s t e m W i t h o u t Zero D y n a m i c s Consider the second order model edS s ~ y 4. To guarantee zero steadystate error.. proportional and integral action. with ~r = 0. R.
41 0.. .a l y e .kD)s F (ao + kp) + k.s! o. bl # 0 where y is no more a fiat o u t p u t of the u n p e r t u r b e d s y s t e m .2 2 perturbationsignal 1. 1 .Ue Thus.21 0~ 0 K 1 0.4r 1.5 o.(al t. the closedloop system dynamics results (1 + blkD)S 2 + (al + boko + blkp)s + (ao + bokp + blk..5 _ ~ 5 ~met 10 0 0 5 time 0 5 lime t tO t 10 F i g . a p p l y a controller given by (12). T h e u n p e r t u r b e d e s t i m a t i o n of the derivative t e r m sye sye = . + ~0) y = (bls + b0) u + : ~ .+oo y e ( t ) = O. (12) and (11). W i t h Ye a n d Ue given as before.2 A Linear System With Zero Dynamics Consider a second order m o d e l with a n o n . from (9). Input Systemoutput Realization of /// //" 1..t r i v i a l zero d y n a m i c s e0s (s ~ + .1) + (boktlqblki2) $ + bokt2 ~ Ye = 82 ] ( l+blkD+ b D w 8 T h e s t e a d y s t a t e error is 0 for a s u i t a b l e choice of the coefficients..$ .1 7..6~ 0.5 s 2 q. 2..5 I 1. 8 ) Ye bzkD Ye  bokD Ue 8 8" ~.u e 8 8 a0 b0 yields the generalized P I D controller Ue = (alkD . a closedloop dynamics given by ( 0.5 0.Generalized PID Regulators 379 yields.kl..kp)ye + ( aokD . 8 + 8 2 ff Ye = 1+ w 8 which implies for a suitable choice of the coefficients limt. Numerical simulation for Example 7.Y e + blue + .
and E. 14 This is a typical problem of sensorless control (see.sl 01 5 time t 10 Fig. We want to control to without measuring it 14.k. Fliess. R and L the armature resistance and inductance. Marquez. to the motor shaft angular velocity. I(t) U(t) 0 I ~ > 9 ) 3 Cr Secondmotor used for loadtorque Fig. 13 See [26] for more details. stabilizes the motor speed to around the desired speed w* if the gain kl is chosen such that J L s 3 + ( L f + R J ) s 2 + (keke + R f ) s + kckl is Hurwitz. 4 are (see.g. . [25]) Jsto = .R I + U (13a) (13b) where U is the command signal.. Delaleau Input System output Realization of perturbation 21 I/I'. ke the backemf constant. Cr the load torque. Numerical simulation for Example 7.Cr L sI = keto ./ 5 lime t 10 '!f/" 1 signal 1. DC motor diagram The control law U = U* . [5]).g.f t o + kcI . I the armature current..380 10 M. R.2 8 A R e a l D C D r i v e la The equations of the real DC motor depicted in fig.s] ! K 11 2~ 01 0 5 time t 10 o. d the motor inertia. 4. (w to*) 7 (14) where U* = ( J L s 2 + ( L f + R J) s + (kekc + R f ) ) to*)/kc. 3. e. f the viscous friction coefficient. e. kc the motor torque constant.
to the armature feedback.. [34]) should then become tractable (see [15] for a prefiminary study on flatness. Santa Ana. A speed control with a good transient behaviour is thus obtained in the absence of any mechanical sensor. The nominal behaviour is represented by dashed lines ( . for example. . They will of course rely on an interpretation of nonlinear observability which is analogous to the one used here. [15]) and therefore play a role in the future of PID control (see [2]). for s = 0. c o m ). or IRcompensation. [27]).kx + k~s (16) where the low pass filter ki/(kx + kes) attenuates high frequency disturbances 15. this extension is quite immediate 1~ thanks to the methods from [18]. In order to facilitate the practical implementation of our PIDs. w/s can be obtained from the nonperturbed equation (13b) w s ke ( LIRI+U)s i Thus. 17 Many problems related to PIDs with delays and Smith predictors (see. also. available at http ://www. computeraided methods (see [1.3. (16) corresponds. the manual CN0055 ~ CNOO55B:DC to DC Ne9ative Resistance speed control.8] (see.Generalized PID Regulators 381 Following the same lines as before. and only if. The experiment reported in fig. 5 shows the performance of the generalized PID under a load torque Cr. Preliminary results on nonlinear systems will be presented soon. The generalization to other systems is now being investigated. on a real DC motor. 9 Conclusion The conjunction between flatnessbased linear predictive control [17] and these new PID controllers 16 should be most useful in m a n y practical instances (see. at starting and braking phases. also. 1997.33]) have to be designed. also. which is often arising in practice. e. U = RI + k~w~ often used for speed control of DC motors without a tachometer (see. [6]) was given in the context of differential algebra and reads as follows: an inputoutput system is observable if. From the casestudies examined here it is clear that computer algebra will be an important ingredient. the control law (14)(15) results (15) kl ( s L l + RI + kew* + ~ s U * ) U . i~ Classical PID regulators were already employed in [17] (see.) .based Smith predictors)..g. Centent Company. c e n t e n t . For a class of delay systems. This interpretation [7. CA. any system variable may be expressed as a function is In steadystate.
and E. Valla M. R. 5 . Fliess M. (1995) Disturbance Torque Estimation in a Sensorless DC Drive.382 M. In: Digital Control: Past.J. I 0 20 0 2 4 2 3 2 II I 4 time (s) 0 5 0 2 time (s) 4 6 time Is) F i g .S. IFAC Workshop at Terrasa.. R i t t ' s characteristic sets (see [31] and [24]) will p l a y an i m p o r t a n t role in the calculations. Buja G. Marquez. (1993) Closedness of Morphisms of Differential Algebraic Sets: Applications to System Theory. AstrSm K. g i i 1 0 .. References 1. (1997) Finite Poles and Zeros of Linear Systems: an Intrinsic Approach.J. DC motor speed control: real (straight line) and nominal (dashed line) behaviour at (a) startup and (b) braking phases of the i n p u t a n d o u t p u t variables and of their derivatives up to some finite order. and Tuning. AstrSm K. Instrument Society of America. (1988) PID Controllers: Theory. NC 2.I. 152157 . Prentice Hall.. Forum Math 5:3347 7.. Design. Delaleau Speed 8O Current 5 20 Control (a) ~3 "3 3 I . Present and Future of PID Control. B. (2000) The Future of PID Control. NJ 4... Wittenmark. Hermrs. Hs T. Menis R..: 2o 0 20 2 time (s) 4 ':2 I O: E 2 time (s) . 1930 3.. (1984) Computer Controlled Systems. (1991) On Nonlinear Observability.. Diop S. Bourlrs H. Englewood Cliffs. Hs T. Internat J Control 68:897922 5. ~ 15 5 0 Ume (s) Control 2 0 ~ 15 4 speed 8O Current (b) '3. Fliess. In: Proc. 1st European Control Conference at Grenoble. AstrSm K. IEEE Trans Ind Electron 42:351357 6. ~ 10 I~ tl it .. Fliess M.J. Diop S. Research Triangle Park.. Paris.
Vol. (Eds.. (1980) Linear Systems. Taylor & Francis. D 3611:124.S. Fliess M. Control. 3rd edn. Diop S. PrenticeHall. Marquez R. Jacobson N. Lavabre M. Englewood Cliffs. In: Actes Conf6rence Internationale Francophone d'Automatique at Lille 16.K. Mounier H. (1990) Some Basic Structural Properties of Generalized Linear Systems. 2nd edn. Techniques de l'Ing6nieur. PWN.. (1995) Flatness and Defect of Nonlinear Systems: Introductory Theory and Examples. Fliess M.. Erd61yi A. (1999) Tracking Control and ~rFreeness of Infinite Dimensional Linear Systems. D 3610:114. 2. Powell J. (1994) Une Interpr6tation Alg6brique de la Transformation de Laplace et des Matrices de Transfert. (1987) Operational Calculus. Mounier H. (to appear) A Trajectorian Interpretation of TorsionFree Controllability 21. L~vine J. (1991) Nonlinear Observability. Delaleau E. MA 22. Fliess M. Boehme T. Gilliam D. & Oxford University Press. Linear Alg Appl 203204:429442 14. (1998) Controllability and Observability of Linear Delay Systems: an Algebraic Approach.D. Addison Wesley. Mounier H. Present and Fhture of PID Control. Edwards C. (1983) Operational Calculus. (2000) ContinuousTime Linear Predictive Control and Flatness: A ModnleTheoretic Setting with Examples. Fliess M. Academic Press. Franklin G.K. Oxford . ESAIM COCV 3:301314 19. 4786 15. Fliess. D 3612:113 26.. Fliess M. Mikusifiski J. Birkhs Basel. 534539 28. and Persistent Trajectories.. Reading. (1988) Commande de Machines h Courant Continu h Vitesse Variable. Marquez R. Warsaw. E.. Fliess M. Kolchin. (2000) Variations sur la Notion de Contr61abilit~.Generalized PID Regulators 383 8. (1974 & 1980) Basic Algebra I &: 1I. (2000) Sur des Pensers Nouveaux Faisons des Vers Anciens... London 10. Fliess M. Journ~e Annuelle Soc Math France. 2nd edn. Coding.) Dynamical Systems. (1962) Operational Calculus and Generalized Functions. 4568 20. In: Digital Control: Past. Kailath T.. Marquez R. Internat J Control 73:606623 18. Oxford 29. Vol... Naeini A.... Martin P. Computer Vision.F.. Flless M.. Warsaw. Holt Rinehard Winston. Fliess M. San Francisco 23.R. Louis J. 1. (2000) Commande par PID G6n6ralis6 d'un Moteur Electrique sans Capteur M6canique.. 714719 9. Identifiability. NJ 24. Multon B. IFAC Workshop at Terrasa. New York 11. E. (2000) From PID to Model Predictive Control: A Flatness Based Approach. Mikusifiski J.. Paris. (1973) Differential Algebra. Fliess M. In: Quelques Aspects de la Th6orie du ContrSle. Systems Control Lett 19:4345 13. (1998) Sliding Mode Control: Theory and Applications. In: Picci G. (1992) A Remark on Willems' Trajectory Characterization of Linear Controllability.P. Systems Control Lett 15:391396 12. New York 25. & Oxford University Press. Internat J Control 61:13271361 17. Fliess M. In: IEEE Conference on Decision and Control at Brighton. Fliess M. Rouchon P. (1994) Feedback Control of Dynamic Systems. Freeman. PWN. Spurgeon S. In: Confdrence Internationale Francophone d'Automatique at Lille 27. M.
. 2nd edn. van der Pol B. (1991) Applied Nonlinear Control. Marquez. Cambridge University Press.. New York .K. Springer. American Mathematical Society. Lee T. Springer. (1950) Differential Algebra. Ritt J. (1999) Advances in PID Control.K.G.E. Fliess...G. NJ 33. Li W. and E.. New York 32. Tan K.F.H. Delaleau 30. Tan K. Englewood Cliffs..J. Wang Q. R.J. Cambridge 31. Lee T. Yosida K.(1999) Finite Spectrum Assigment for TimeDelay Systems.384 M.H.. London 35. Bremmer H. London 34. Springer. H/igglund T. (1955) Operational Calculus Based on the TwoSided Laplace Integral. Slotine J. Wang Q. (1984) Operational Calculus. PrenticeHall.
Scherpen 2 1 Kyoto University Uji. [5.Eigenstructure of Nonlinear Hankel Operators Kenji Fujimoto 1 and Jacquelien M. The Netherlands j. e. Japan Delft University of Technology P.O.g. and provide a formal justification for the use of Hamiltonian extensions via G~. Kyoto 6110011. see e. and a lot of Hankel operator results followed.10] as a special class of nonlinear adjoints. Secondly we investigate both the eigenstructure of the self adjoint of Hankel operators and the eigenstructure of the G~teaux . Recently it was shown that singular value functions are closely related to Hankel operators [5. 1 Introduction The Hankel operator is one of the most important tools in linear control theory. firstly we consider these adjoint operators from a variational point of view.11]. scherpen@its. singular value functions which are a nonlinear extension of the Hankel singular values in the linear case play an important role in the nonlinear version of Hankel theory. Box 5031 2600 GA Delft. The nonlinear extension of balancing theory were firstly introduced in [8]. nl Abstract. which is closely related to the Hankel norm of the original system. tudelft. The results in the paper thus provide n e w insights to the realization and balancing theory for nonlinear systems.g.10. a.6. Indeed they are extremely useful for balancing and model reduction of linear systems. Nonlinear adjoint operators can be found in the mathematics literature. [1]. In particular. So called nonlinear Hilbert adjoints were introduced in [5.10]. and to give a new characterization of Hankel singular value functions for nonlinear systems. It is proved that the variational system and Hamiltonian extension can be interpreted as the G~teaux differentiation of dynamical inputoutput systems and their adjoints respectively. m. W e utilize this differentiation in order to clarify the eigenstructure of the Hankel operator. Here. and they are expected to play a similar role in the nonlinear systems theory. This paper investigates the eigenstructure of Hankel operators for nonlinear systems. The existence of such operators in inputoutput sense was shown in [6] and their statespace realizations are preliminary available in [4] where the main interest is the Hilbert adjoint extension with an emphasis on the use of portcontrolled Hamiltonian systems.A.teaux differentiation. The purpose of this paper is to clarify the eigenstructure of the Hankel operator.
It is given by { ~ = f(x. Here we assume the origin is an equilibrium. u. Some of these results are reviewed from [4. i. u j : x~ Yv x(t ~ = x ~ 1 ~ ( t 1) = x ~ . = S a ( u . Also. which has a relationship with the Hankel norm for nonlinear systems. Scherpen derivative of the square norm of Hankel operators. G~teaux differentiation . u) x(t~ ~ (1) with x(t) E ]Rn. The Hamiltonian extension Z'a of .=.u) y~ = E~(u.U: L~ [t ~ t 1] >L~[t ~ t 1] defined on a (possibly infinite) time interval [t ~ t 1] which has a statespace realization y = E(u) : {. 2 Variational systems dynamical systems as Gfiteaux derivative of This section is devoted to the statespace characterization of GRteaux differentiation of dynamical systems and their adjoints via variational systems and Hamiltonian extensions. xv. state and output respectively and they represent variation along the trajectory (u.U is now given by a Hamiltonian control system [2] which has an adjoint form of the variational system.9]. O) .0 hold. Ghteaux derivative is important for understanding the meaning of the variational systems and the Hamiltonian extensions. u). u) + u T h ( x .(x. x. y) of the original system 2Y. It is shown that the eigenstructure of these operators is closely related.e. ~ Oh + ~u~ Oh (2) ~TXv + "ff~uv The inputstateoutput set (Uv. Yv) are so called variational input. we introduce the variational system of L'.A. u(t) E IR m and y(t) E ]R r.j_ o h T ~ .386 Kenji bSajimoto and Jacquelien M. uo) : W = f ( x ' u) Tp_}_ ~ x T u a x(t ~ = x ~ p(t l) = OH T Px. other than the singular value functions given in [8] have. Furthermore this eigenstructure derives an alternative definition of the singular value functions. = pl (3) Ya  __ "~u OH T __ ~_LT~. ua) := pT f ( x . O) = 0 and h(O.u) h(x. p . Before giving the Hamiltonian extension of S. Let us consider an inputoutput system .Ou ~ ' " Ou a with the Uamiltonian H ( x . It is given by ~c y . f(O..
Then T has a G~teaux derivative at x E X if. D e f i n i t i o n 1.Eigenstructure of Nonlinear Hankel Operators 387 of Hankel operators plays an i m p o r t a n t role in the analysis of the properties of Hankel operators. see [10. To this end. u) = go(x) + g(x)u and h(x. namely there exists a neighborhood Uv C__L~[t~ 1] of 0 such that u E L~[t~ Then there holds u. whereas G~. U C_ X is open. then we can give the differential version of Proposition 2 in [4]. = (6) with the variational system Ev given in (2). Suppose that the assumptions in Theorem I hold and that uEL'~[t~176 t] ~ Ix(tl)l<oo. Then there holds u). (xv(t~ uv)) ~r (xv(tl). For the reason of space we omit the proof. (pl.[p(t~ (7) Let 2a denote the mapping ((x(t~ u).=o. i. Fr~chet differentiation is a special class of Ghteaux differentiation. f ( x . which is especially useful for analysis of nonlinear static functions. since that is the most suitable for our framework.4]. g and h. Suppose X and Y are Banach spaces. we concentrate on Ghteaux differentiation.T(x) = d T(x + e~)l. See [9] for the details. (p(tX). T h e Hamiltonian extension ~wa also has a relation with Gateaux derivative.). we state the definition of Gs differentiation. and that E is Gfiteaux differentiable. Perhaps more wellknown than the G h t e a u x derivative is the Fr~chet derivative. (s) .teaux derivative is not necessarily linear. 9~0 ~r dg (4) We write dT(x)(~) for the G~teaux derivative of T at x in the "direction" ~. uo)) = (x ~ u))'(p 1 . for all ~ E U the following limit exits: dT(x)(~) = lim T(x + e~) . u. which is the topic of Section 4. E llv C n~[t~ 11 ~ y~ E n~[t~ (5) E (u.e. Suppose that • : u ~ y in (1) is inputaffine and has no direct feedthrough. It is also noted t h a t Fr~chet derivative is linear from its definition and it does not depend on the "direction" to be differentiated. Theorem 2. Yv) of the system Sv. Let E* represent the nonlinear Hilbert adjoint of E . and T : U + Y. Ua)) ~'~ (p(t~ Ya) of the system Sa and let ~ denote the mapping ((x(t~ u). u) =_ h(x) for some analytic functions go. In the sequel. and provides a justification for the fact that it is called the adjoint form of the variational system in [2]. Theorem 1.
We only consider time invariant inputaffine nonlinear systems without direct feedthrough of the form E: { . The Hankel operator and its derivative 3 This section gives the statespace realizations of Ghteaux derivative of some operators. p0 := p(t o) and pl := p(t ~) for simplicity. u~). ya))r~• (9) holds with the inner product on 11~" x L~.• ~ = ((xO.e. Ip(t~ < ~ . Ua))~. Yv). Ua) = (dE(u))" (Ua) (11) with the Hamiltonian extension . Ya) = ~a(X ~ u)(P 1. in contrast with the portcontrolled adjoint systems.. Let the Hamiltonian function be given by H~ = pTz~. The proof follows similar arguments as Proposition 2 in [4].A. (pl.388 Kenji Fujimoto and Jacquelien M. uo))~'• r'l This implies (8) and completes the proof. i. (pl. This reduces to tI t1 = n~lt=. Scherpen P r o o f . Uv). This adjoint system has the same structure as portcontrolled adjoint systems [4]. However its statespace realization is unique and coordinate independent.y~ uo.pX)~  (x~176 Therefore ((Xlv. we obtain the following corollary. C o r o l l a r y 1. u~))~. Setting x~(t ~ = 0 and p(t 1) = 0 in Theorem 2. ~:. y~) = ~'~ (x ~ u) (X~ uv) and (p0. = f(x) + g(x)u h(x) (12) . (10) Then there holds 22a(u.• = ((~o. Ua) yields ( ~ ( x ~ u)(x ~ u~).Ua given in (3). Then we have dHv pT&v "4xTp=pT \~x +9(X)Uv) _ x vT ~gx (Ofxv IOfTp'4~xOhT Ua) T OhT T T = pTg(~)u~ .o = (x~.(~o. it uses the portcontrolled Hamiltonian systems structure. and denote x ~ := x(t~ x 1 := z(tl). E L~[t~ 1] ~ Ix(t~)l < ~ . Suppose that the assumptions in Theorem I hold and that u E L~[t~ u.~ 07 ~~ = y~ ~ .  n~l. (pO.=. ~)(pl. Substituting (x~..
Suppose that the assumptions L~[0.o o .f(x) + g(x). p(oo) = 0 vo = y+(gr(~) p) This theorem is differential version of Proposition 3 in [4].6]. oo) ' 7+(u)(t) 0 t9 := u(t) t 9 [o. o0) ~ L~ [0. ~(oo) = o ~  OhT u. along with a deeper and more detailed analysis of the Hankel operator. oe))+ L~[0. pO = (dO~:(xO)).v = O~ o C. T h e o r e m 3. It is readily checked that for linear .0] and . one can employ statespace systems to describe them which are operators of/~ n r L~[0.cr in Theorems 1 and 2 hold.. Consider the operator 27 in (12). oo) denote the flipping operators defined by Y_(u)(t) := u(t) 0 t 9 (~o.Eigenstructure of Nonlinear Hankel Operators 389 defined on the time interval (or or Here S is L~stable in the sense that u 9 L~'(oo. oo) and (dT~z(u))* : L~[0. cxD) are given by = f(~:) ~(o) = ~o oh . cr + L2[0.T_ and 7/. oo) and L~n[0. := 27 o . (dCz(u))* : ~n ( x L ~ [ 0 .~) Furthermore the Hankel operator 7/s : L~ [0.T_(u) x(~) = 0 x(O) (14) .T+ : L 2 ( .. ~ ) of E' is given by 7"/~.0] implies that 27(u) restricted to [0.~(u))T(x ) 0~: p p(O) = pl va = y + ( g r ( ~ ) p) = f(~) + g(x) y_ (u) vo = (dU~(~))'(uo): { ~ _O(]+flY(u))Tp o. cr + L~[0.o o . oc)(xL~'[0.O~T l x~ Ua p(o0) p0 = p(O) . Although the original definitions of the observability and controllability Ol>erators [5] are given by ChenFliess functional expansion [3].it'_ : L 2 ( .T_(u) x(oo) = 0 0 v. Then statespace realizations of (dOs176 * : ( x ~ n) + N". and using techniques from [4]. oo) is in L~[0. specifically: y = O~(z ~ : { ~ = f ( z ) z(0) = x ~ v 9 '= h(~) (13) ~1 = C~(u) : & = f(x) + g(x).(ua) : { [9 = . oo). . The proof of this theorem is obtained by applying the adjoint Hamiltonian extensions of Section 2. oo) r IRn. This has been proven in [5.v holds.~ T (x) p. The adjoints of their derivatives can be obtained using Theorem 2. = (dC~(u))*(p 1) : { _O(fTg. 0] t 9 [0. o o ) + L2(oo.
suppose that the assumptions of Theorems 1 and 2 hold and that there exist welldefined observability and controllability functions L0 and Lc. Further we will give a new definition of singular value functions which are closely related to the eigenstructure and the Hankel norm of nonlinear operators.. We first review what we mean by inputnormal/outputdiagonal form. we can state the result from [5. . the above characterizations yield the wellknown state space characterizations of these operators. ~'. .=(o)~= l In the sequel. in a similar way as the observability Gramian and the inverse of the controllability Gramian are solutions of a Lyapunov/Riccati equation. we need to define the following energy functions: Definition 2. 4 Eigenstructure of Hankel operators This section clarifies the eigenstructure related to the Hankel operator utilizing the Ggteaux derivative. r. The observability function Lo(x) and the controllability function Lc(x) of S in (12) are defined by Lo(x ~ := ~ Ily(t)llUdt.(z))z with vl(z) > .. . Scherpen systems. > rn(z) being the so called smooth singular value functions r Now. . we concentrate on the system L' in (12) and. In order to proceed. h) that fulfills certain technical conditions. Then there exists on a neighborhood U C V of O. =(~)=o. In [8] these functions have been used for the definition of balanced realizations and singular value functions of nonlinear systems.[o.6] that relates the singular value functions to the Hankel operator: . [8] Consider a system (f. Also they fulfill corresponding HamiltonJacobi equations. g. These functions are closely related to observability and controllability operators and Gramians in the linear case.. a coordinate transformation x = r r = O. where Lc(z) := Le(r 1 T = ~z z. Lo(z) on W := := Lo(r = 2 z T d i a g ( v l ( z ) . ~ z. .~) x(O) = x ~ u(t) = 0 (15) (16) Lr t) := min 2 Ilu(t)ll~dt. see [8]: T h e o r e m 4.A. which converts the system into an inputnormal/outputdiagonal form.390 KenjiFujimoto and Jacquelien M.
. oo) the corresponding eigenvector. . namely II~tr(v)lb/llvll~ = IISIIH holds.. .ll~/ll'~.i)L~. .~u) k 11.II~(u)ll~/llulb(u. Then the functions {&j}~=~ are singular value functions of the Hankel operator 7l ~ in the following sense: (. ?i* o ~t~(v~) = a~ vj (lS) holds for each eigenvector vj. .o o ) = 0 to z(0) = ~j and define f~j = 9r+(vj).Eigenstructure of Nonlinear Hankel Operators Theorem 391 5.~))~.d i a g o n a l form. such a relation does not hold.~u) = Ilull~d(lln~(u)ll2)(.~(u))* o'~. .~. 0. .~u) II~.~ (u))* o ~..) = 11'.2.'~ (. .(ll'... on the other hand. L rut#0 . 3u) = Ilull~ ((d~.^ =a](zj)(vj.diagonal realization of a causal L~stable inputoutput mapping 57 on a neighborhood W of O.. This eigenstructure has a close relationship with the Hankel n o r m of X' defined by IlSll~ := ma~ .n o r m a l / o u t p u t .~(u)ll~d(llull~)(.~u) .~(u).~(u)l12/ll.. in the linear case. P r o o f . 0) for j = 1 .. Moreover. 2 .u}(u~:(. [5] Let (f.~). . . ^2 .~(u)lWIlull2 (in the direction tfu) satisfies d (11~ _~)11~. . In the nonlinear case. j= l. Now we consider an eigenstructure of the operator u ~ (dT/~:(u))* o ~/~:(u).(ll'~. n. characterized by (dU(v))* o~(v) = ~ v (19) where A E R is an eigenvalue and v E L~[0. and the functions b~ ( zj ) = r(~j). 0.~. The derivative of II'~. h) be an analytic n dimensional inputnormal// output. Define on W the collection of component vectors ~ = (0. g. . .112 II~'(u)l12 .1122 II..112 ] 11. zj. Then v satisfies (19) with the eigenvalue A = IISIl~.o (21) for all variations 3u at u = v because it is a critical point..~(u) ll~<(d'~.ll~)2u.~(v)ll2/llvll=) ~ v = IISII~ v (22) . instead of (18).n. Let vj be the minimum energy input which drives the state from z ( .. . (17) The above result is quite limited in the sense t h a t it is dependent on the coordinate frame in which the system is in i n p u t . . oo) denote the input which achieves the maxi mization in the definition of Hankel norm in (20).v. Let v E L~n[0. .~ (u).) O II~t~(u)ll~ Ilull~ (20) T h e o r e m 6. This reduces to (dn(v))* o ~(v) .~.
the pair )~ and v is a pair of an eigenvalue and an eigenvector. Scherpen 1"3 and thus proves the theorem. i. Then we have v = ~y~ = gr(x)p. A pair )~ E ]~ and a nonzero v E Lr~ [0. we only prove the necessity here. The sufficiency is straightforwardly obtained.~ o C. precise properties of the eigenstructure (19) are investigated.v(u).A. Lemma 1. and therefore.392 Kenji Fujimoto and Jacquelien M. Firstly the following l e m m a gives the complete characterization of the eigenvectors.e. cxz) is a pair of an eigenvalue and an eigenvector of the mapping u ~~ (d'Hs oT/~(u) if and only if there exists x ~ E ]~n such that i =f(x)g(x)gT(x)p o(f+g~) Ir ox [u=grp p x(O)x ~ p(0) = ! ~ 1 7 6 X O~ X I  (23) gT (x)p P r o o f . Theorem 6 clarifies that the eigenstructure (19) characterizes a necessary condition for the input v to achieve the maximization in the definition of the Hankel norm (20). We can observe Ya = (dTl ~:(u) )* o 7t ~(u) = (dC~(u) )'o (dO~:(CE(u) ) )*o 0. Then (24) reduces to Ya = (dC.~ = dLo(x~ if = (o~(x~ This means = d(llO~:(x~ = ((dO~(x~ o O~(x~ dOE(x~ dOE(x~ * o O s ~ _ OL~176 T Ox o (26) We obtain from (25) that ya = (dCs176176 It follows from Theorem 3 that the the state space realization of this operator is given by { x = f ( x ) + g(x) . In the remainder of this section.(24) Let x ~ := CE(u).T_(u) OI]+g'~(u)) T = 0~ p ya = J : + ( g r ( x ) p) x(~) = 0 p(0) = ~ T ( x ~ 9 (27) Suppose u = v and Ya = )W. (28) .v(u))* o (dOE(x~ o O~:(x~ (25) Next we consider the Ghteaux derivative of Lo(x ~ in the direction OL~176 Ox.
~L~.elm cr . L e m m a 1 gives the characterization of all pairs of eigenvalues and eigenvectors.g(x)gT(x)O0~ T x(O) = x ~ y~ = C~(x~ : Ya gT (x) OL~T ax p(t) (33) If the condition (30) holds.cL(m) pmin(c) := Ilnz(u)ll2_ Ilull2 Ilnz(u)ll2 [lull2 max 117~z(u)ll~ (34) . We will now continue to investigate the necessity. then the reversetime expression of (27) reduces to 393 { Lemma ~: = .~) flulla=r .Eo.(1/A)p. ~ ) the pseudoinverse of C2 defined by C~(xt):=arg min Ilul12.. it does not deal with the existence of the solutions. Suppose there exist A E ~ and a nonzero x ~ E ~n such that (x)= OLo o A gT(x ). In order to proceed.Eigenstructure of Nonlinear Hankel Operators Let/5 :.~. 2. Next we concentrate on a special class of all pairs which are closely related to the energy functions Lo(x) and Lc(x). then it follows from L e m m a 1 that (OLc/Ox)T(x(t)) holds along the trajectory of (dT/z(u))* o 7 / s proves the lemma.  and this [] T h o u g h L e m m a 2 gives a sufficient condition of the pairs of eigenvalues and eigenvectors which are closely related to the energy functions. we define two scalar functions Pmax(C) :max II"lla=~ rain .~) Ilull2=c Ilul12 (35) . The statespace realization of C~ is given in [4] as { ~ = f(x) ..f ( x ) g(x)gT(x)fi x(O) = x ~ ( x ( ~ ) = O) = aT (~)~ [] This is identical to (23) and this completes the proof. cE(=)=~~ (31) (32) P r o o f . cOL~ o (30) Then A is the eigenvalue of the mapping u ~~ (dTt~(u))* o ~ s ing to the eigenvector correspond v = c~(~ ~ with C~ : ~ n _+ L~n[0.
dLc(71(s)) _ OLc(.7(0) = xi(c) maximizes (minimizes) the value (Lo/L~) in the level set Xt. let the levelset of Lc(x) be given by XLo=k := { x I Lc(x) = k }. min}).(. Namely Pmax (c) represents the Hankel norm under the fixed input magnitude Ilul12 = c. T h e o r e m 7./)d~l(S) _ 0 ds 011 ds holds along r/(s). Firstly we define xi(c) := Cz(ui(c)) and show the existence of Ai(c) such that ~ (xi(c)) = A. Suppose the energy functions Lo(z) and Lc(x). Take a curve r/(s) E . Since r/(s) is contained in the level set XL~=~.~ ~ . min}. min} (36) Further both the pair Amax(C) and CZ(Umax(C)).)) d ~'~('~(~ s=0 L* 2 dLo(~(s)) s=0 2 aLo(o) do(s) s=0 = O.{max.I'L~ parametrized by a scalar variable s such that 7/(0) = xi(c) holds. and the related functions Pm~x(C) and Pmi. (37) Tothis effect. Scherpen Pmax(e) is closely related to the Hankel norm because IISIIH = maxr Pmax(C) holds.(C) are the eigenvectors of u (dT/~(u))* o "Rx(u) with respect to the following eigenvalues Am~• and Amin(c) respectively. Let i denote the index such that i E {max. Let Umax(C) and Umin(C) denote the inputs which achieve the maximization and minimization in the definition of Pmax(C) and Pmin(C) in (3~) and (35) respectively. (39) . Next we can observe the following relations (38) max IIn~(u)ll2 ~ Bin II~t~(u)ll2 . :~. Therefore these functions are an alternative nonlinear extension of Hankel singular values other than ri(x)'s in Theorem 4.(c) are sufficiently smooth.394 Kenji Fujimoto and Jacquelien M. and the pair Amin(C) and CZ(Umin(C)) satisfy the condition (30). This implies . namely they satisfy Ilu~ll~ = c and 117~r(ua)ll~/lluaH2 : pa(c) (i . Then Um~x(C) and Umi. Therefore we obtain L:(.(c) = d(c) c dg+ ~ dp~(c) i = {max. Then x~(c) E XLo=~ follows from the fact that u~(c) is the input which minimizes the input energy under the constraint Ilul12 = c.A. P r o o f .(c)~~(xi(c)). Furthermore it should be noticed that Pmax and Pmin exactly coincide with the m a x i m u m and m i n i m u m Hankel singular values in the linear case.
Then we can obtain OLo(~) d~(a) [ e 0r ds OLe(~)d{(s) 0(~ ds s= ~i(e )  dLr dL~ ds ds I s=e d(p~(e) e:~/2) c dp~(c) de __ V2(e) _~_ . Namely both (OLo/Ox) and (OLc/Ox) are orthogonal to the tangent space of XLo=~.1)dimensional.d(ea[2) 2 de de This completes the proof. min{p + (c). . Secondly we prove the equation (36).. (max{p + (c). Further we can extend this result for general singular value functions.=~. From the above discussion. 2 . j G { + . [] T h e o r e m 8.(c)} > max{p++l (c). i E { 1 .} such that min{p+(c). Suppose the statespace realization is in inputnormal form. Pi+i (c)}. (C)) arid that there exists x~(c)'s satisfying Lo(g(e))OLo 2 (41) 0x (g(e)) = A (c) (e) := d (e) + (dd (42) ith (c)ldc)cl2. .at x = xi(c). The assumption that the Jacobian linearization of the system has n distinct Hankel singular values implies that the hyper plane Lo(x) = k > 0 has the form . for any vector ~ E ~ " which is not orthogonal to the tangent space of X~o=~ at x = xi(c). Because the tangent space is ( n . P l (c)} = Pmax(C).Eigenstructure of Nonlinear Hankel Operators 395 The equations (38) and (39) have to hold for all curves ~/(s) e XL. p~ (c)} = prni. Remember that ui(c) can be described by ui(c) = s It follows directly from L e m m a 2 that ui(c) = C~(xi(c))is the eigenvector o f u ~ (dTls with respect to Ai(c). Ai(c) can be expressed as Ai(c)ox Ox (40) Let ~ be the directional derivative (d~(s)/ds) of another curve ((s) = xi(s) passing through the maximizing (minimizing) state. we can conclude that (OLo/Ox) and (OLr are linearly dependent at x = xi(c). . . Suppose the energy functions Lo(x) and Le(x) are sufficiently smooth and the Jacobian linearization of the system S has n distinct Hankel singular values. P r o o f . . n } . Therefore there exists a scalar constant Ai (c) such that (37) holds. Namely it goes across the level set XLo=~ through x = xi(c)(= ((e)).p.. Then there exists locally 2n smooth singular value functions p~(c)'s.
u~ (c)))La = P~2(c) (u j (C). This equation implies the linear case result given in Theorem 8. El The following fact similar to Theorem 5 follows immediately. and even the pairs given in Theorem 5 are not unique [7]. 7t~ ( n x (u~ (c)).q of the Jacobian linearization of the system. However. It was proved that the variational system and Hamiltonian extension can be interpreted as the Ghteaux differentiation of dynamical inputoutput systems and their adjoints respectively. There are other pairs p's and u's such that the relation (43) holds. p~.1 in [12]. Then ~ (c) 's and q(c) := satisfy (43) (uj (c). We utilized this differentiation in order to clarify the eigenstructure of the self adjoint of the Hankel operator. 7 and 8 is particularly important because it is closely related to the critical points of the Hankel operator.(c) } > max{P++l (c). minimums and saddle points) are same at lease in a neighborhood of the origin. Unfortunately. P. u) itself. This means that there exist 2n critical points such that (38) and (39) hold for any curve q(s) E XLc=k. Furthermore ~ (c)'s play the role of the coordinate of the balanced realization. the eigenstrueture given in Theorems 6. In future research the application to model reduction of nonlinear systems of this new concept will be investigated. The number of the critical points (which include local maximums. U~(C))L2.396 Kenji Fujimoto and Jacquelien M. 5 Conclusions We studied the eigenstructure of Hankel operators for nonlinear systems via Hamiltonian extensions. Further the order min{p + (e). This proves the existence of A~(c)'s and ~ (c)'s satisfying (42). This fact may be especially important for applications in model reduction for nonlinear dynamical systems.+I (c) } follows from the fact that p+ (0) = p~(0) = cri holds with the Hankel singular value . C o r o l l a r y 2.4].A. Scherpen of a hyper ellipsoid whose radiuses are distinct when k is small. The property (41) can be proved in a similar way as in the proof of Theorem 7. which is closely related to the Hankel norm of the original system. the pair ~2(c) and uJi(c) is not the pair of an eigenvalue and an eigenvector of the operator u ~ 7/~ (~/~(u). In the linear ease A~(c)'s coincide with the square of singular value functions ~ ( c ) ' s respectively whereas ~ ( c ) ' s themselves are identical to the Hankel singular values. which is due to several nonuniqueness issues [7. Indeed the equation (30) reduces to xTQp = A X T (44) with observability and controllability Gramians Q and P respectively. Suppose the assumptions in Theorem 8 hold. This completes the proof. .
A. A CC. Submitted. and van der Schaft. on Decision and Control. N. and Gray.. S. M. and Gray. J. 5. Upper Saddle River. M. 9. volume 101 of Lecture Notes on Control and Information Science. K. Proc. C. S. A. (1970). 7. 8. 11.. W. W. J. pages 197222. Gray. 3. J. J. J. S... W. Fujimoto. singular value functions and Gramian generalizations for nonlinear systems. K. W. Fliess. Math. W. S. Aurora. Proc.. 4. pages 33493353. A. M. Proc. FP8. Batt. (1999a). A. Ann. Minimality and similarity invariants of a nonlinear state space realization. 189:525. J. W. 2000 Conf. Gray. M. and Gray. 10. W. (1999b).. Information Sciences and Systems. Scherpen. Crouch. M. Doyle. and Scherpen. M. (2000). A. A. pages 3944. Proc. (2000).. (1998).. Scherpen. 6.Eigenstructure of Nonlinear Hankel Operators 397 References 1. Scherpen. Matrices de Hankel. J. (1974).J. On the nonuniqueness of balanced nonlinear realizations. M. and Scherpen. J. Robust and Optimal Control. 21:143153. IFA C Workshop on Lagrangian and Hamiltonian Methods for Nonlinear Control. pages 712. A. M. Appl. Scherpen. On singular value functions and Hankel operators for nonlinear systems. (1999). S. (1993). J. Math. M. 37th IEEE Conf. E. 12. Hankel operators. (1996). Accepted for IEEE Trans. Hankel operators and Gramians for nonlinear systems. and Glover.. (2000). Zhou. and Scherpen. J. and Gray. S. . PrenticeHall. Fujimoto. Balancing for nonlinear systems. pages 23692364. K. Hamiltonian realizations of nonlinear adjoint operators. Inc. (1987). Gray. SpringerVerlag. K. Contr. Scherpen. Variational and Hamiltonian Control Systems. S. J. A. Proc. 38th IEEE Conf. Pures. on Decision and Control. Systems ~ Control Letters. J. Berlin. A. Nonlinear compact mappings and their adjoints. 2. P. On adjoints and singular value functions for nonlinear systems.
G illet @epfl.D i s t r i b u t e d A r c h i t e c t u r e for T e l e o p e r a t i o n over t h e I n t e r n e t Denis Gillet. These applications are among the most critical ones in industrial environments. and Pierre Huguenin Swiss Federal Institute of Technology C H . This paper discusses challenges in enabling teleoperation over the Inter net for the specific case of real systems which exhibit fast dynamics. The industrial equipment previously mentioned is just one example of physical systems that can be teleoperated. since they handle the functions related with the dynamic behavior of real equipment. As a consequence. an architecture is proposed to enable the remote supervision and remote tuning of distributed controlled systems. Christophe Salzmann. . Among these services. including their embedded or external control devices. To contribute to the fulfilment of the market expectations related with emerging realtime services. Didactic setups available in instructional laboratories can also be teleoperated in a flexible learning context to provide students with the necessary resources for practical experimentation. The architecture implemented for the teleoperation of both industrial equipment and didactic setups relies on a computerbased clientserver scheme with Internet as a communication channel. Switzerland Denis. A distributed clientserver architecture is proposed to provide the necessary level of interactivity for supervision and tuning.1015 Lausanne. new industrial network implementations are increasingly based on the Internet protocol to ensure a global and continuous access. [2]. However. these new capabilities are not yet exploited extensively due to the lack of predictability of the transmission delay inherent to the Internet and the lack of standardized deployment methodologies. as well as for sustainable support and immediate service from their suppliers. ch A b s t r a c t . The proposed approach is illustrated through the remote control of an inverted pendulum. without compromising the essential control tasks. and [3]. Enabling features include realtime simulation and augmentedreality visualization to enhance the user perception of the distant ongoing operations. the process and manufacturing industries increasingly request global monitoring. 1 Introduction To operate critical facilities efficiently. telemaintenance is the most requested. as described in [1].
concluding remarks and perspectives are expressed in Sect. additional devices have to be integrated. Clientserver Control Loops The remote operators are provided with client software to observe and to pilot the physical system via the Internet. The onsite server is the computer that stands for the communication and operation interface with the real system and its instrumentation. Just think about the effect of the loss of a single or numerous measurement samples for stability and safety. . 6.. 3. Then. 2.. Remote client Onsite server Physical system " J. details about the clientserver architecture and the necessary features needed at the operator side to provide a sufficient quality of service are given in Sects. Tuning loop Content adaptation loop Realtime control loop 1/. A common and dangerous mistake is to close this loop across the Internet. because wrong actions can be performed by operators if the reactions of the system are not noticed fast enough for a valid interpretation. Various ways to handle the transmission delay that occurs when switching from local to remote control are described in Sect. 4 and 5. respectively. The first one. 1. With such an architecture for teleoperation over the Internet. Even if such an approach is challenging from a robust control pointofview. The last loop is closed by the user for setting the conditions of operation or for tuning the distant realtime control loop.400 Denis Gillet et al. The second one is the control loop designed to adapt the transmitted content to the available bandwidth in order to guarantee a sufficient quality of service. This is merely to ensure the reception of the information essential to carry out the desired operations according to their respective priority. there are three imbricated control loops to implement (Fig. In Sect. This is the loop the most sensitive to the transmission delay.t Fig. such as cameras to broadcast video views of the system and of its environment. the inverted pendulum and its local control are introduced.. To overcome the drawback of the operators not being present at the equipment location. dedicated to the realtime control of the real system. it is unacceptable from an industrial pointofview. 1). has to be implemented at the server side. Finally..
No actuator exists for this axle. in two dimensions. It is worth developing a good model because it leads to a much more accurate control than empirical methods. as well as the respective velocity estimates. The rotational axis of the pendulum stick is on this cart. 2 is made up of a 4meter horizontal rail mounted on top of a case that contains both the electronics and the power supply. It emulates. The inverted pendulum This laboratoryscale system can be controlled by any type of computer equipped with a National Instruments PCI1200 family DAQ board.A Distributed Architecture for Teleoperation over the Internet 401 2 T h e I n v e r t e d P e n d u l u m a n d its Local C o n t r o l The pendulum described in this section is an inverted pendulum with two degrees of freedom. 2. using a single actuator. Hardware security switches that turn the power off can be found at each end of the rail. as well as a good knowledge of its physical parameters. The inverted pendulum shown in Fig. A representative physical model of the system is required for the controller design. This portability feature turns the setup into an ideal support for educational demonstrations or trade shows. A cart. . The only way to change the stick position is to accelerate the cart. a juggler trying to keep a broomstick in equilibrium on his fingertip. The control objective is to simultaneously keep the pendulum stick in the upright position and the supporting cart at the center of the rail. the model can also be used for realtime simulation as needed for advanced teleoperation. can slide along this rail. moved by an electric motor by means of a metallic belt. The implementation of the corresponding linear quadratic regulator (LQR) relies on both angular and longitudinal position measurements. Moreover. Motor 22// 4~ Rail Fig. The underlying statespace multivariable control principle aims at minimizing the energy spent by the controller to keep the stick raised and the cart at the center of the rail.
. 7 r ~  ... it is the optimal time to inverse the voltage polarity (for a more accurate solution. the dynamics of the electrical drive has to be taken into account).. .. 3).:: i/ i 2 3 4 ~! i~ i 1 Ume S 6 7 8 II Fg... The control sector size depends on the damping the controller can introduce and its robustness to the system nonlinearities.....~ .. Its sequence and its intuitive justification are given as follows.. such as the ones due to the angular position of the stick...~ . Thus. Once the stick crosses the lower point of its trajectory. The voltage level is a tradeoff between the horizontal range of the cart displacement (the boundaries of the rail have not to be reached) and the angular velocity of the stick..: : ' : 9 9 ... In this restricted control sector.... SvdnJ_V~__. ~e 6 . the cart accelerates and the stick swings up due to its inertia.. Then.. the statespace controller can be activated to stabilize the pendulum. The swingup algorithm implemented to move the stick from the lower stable vertical position to the upper unstable one is a very simple bang bang procedure: a constant voltage value is chosen to be applied to the motor and its sign is worked out adequately depending on the stick position in order to induce and amplify the stick oscillations.. This sequence has to be repeated until the stick reaches a predefined interval around the vertical position (Fig. .402 Denis Gillet et al... 3. Swingup control signal The swingup procedure chosen is quasioptimal from a time pointofview..} O " " i/ " . The swingup tuning parameters are the constant voltage value and the control sector size. : : Control :: ! : : : : : : i . it has accumulated the maximal possible amount of kinetic energy... the cart reaches its maximal velocity and the stick swings down by gravity.... First..
the clientside and the serverside ones. are the sampling period and the four state feedback gains. the video stream.00 400.1 13 0 Fig.0O 200.00 450. 500. Example of the transmission delay variation in Internet communication The main difficulty for the operator occurs if the time constants of the remote system are of the same order of magnitude or faster than the time delay. even in the case of a significant transmission delay.00 250. 4.00 I ~ 360.00 Packet 8tze Transmitted . There are two classes of additional solutions. Unfortunately. Since the transmission delay is incompressible. in addition to the swingup tuning ones.A Distributed Architecture for Teleoperation over the lnternet 403 The controller parameters that can be modified by local operators.40. Such schemes aim to use optimally the available bandwidth. The same settings have to be available when remote operations are considered. the delay can vary from a few milliseconds to hundreds (Fig.00  2 3 4 5 6 7 9 9 0 11 12 13 14 15 16 17 is it 20 21 2. the tuning stream and the coordination stream [4]. 3 Transmission Delay Handling The varying transmission delay (approximately half the Round Trip Time) in Internet communication is the most important problem to handle when implementing teleoperation solutions. The solution to partially alleviating this difficulty relies on an intelligent characterization and management of the data streams transmitted. the additional solutions only aim to reduce its annoying effect on the sensorimotor behavior of the operators. Enabling fast remote prototyping of the controller requires that the effect of any change in these parameter values can be noticed as soon as possible by the operators.1 2 0 0 S e e / 2600 Byte8 I 4 i j ~10~ e j ~ 300.00 150. The clientside solutions mainly rely on predicting the remote on . Depending on how the information packets transmitted are routed and how many touters are crossed. this condition appears to be true for most of the mechatronics systems. such as the measurement stream. 4).
In automatic control. The serverside solutions rely on intelligent context analysis carried out automatically by the server to feed back synthetic and composite results that can reduce the reaction time of the operators. 4 ClientServer Architecture The teleoperation server software is made of two distinctive parts. In mobile robotics. The main concept in turning a locallycontrolled setup into a remotelycontrolled one consists of moving the user interface of the monitoring and control software away from the physical system. serverside solutions can be more convenient. where the dynamic models are usually known and where the context of operation is always predefined. For local experimentation these two parts are located on the same computer.404 Denis Gillet et al. the user interface and the part handling the realtime operations. 5. The client application is used to . going operations to help the operator in anticipating the effect of its actions. clientside solutions are more interesting. including the interface with the outside world. once such a high level of information becomes available to them. Two distinctive parts result: the remote client and the onsite server. Software components at the client and server sides The remote client is a computer equipped with the functionalities necessary to observe and to act on the physical system. where the context of operation may evolve significantly and where more intelligence is embedded in the onsite system.I J~ "lJ ~ "~ ~ Sorvor Fig. The teleoperation of an inverted pendulum clearly belongs to the first class. Video Ctmmm l/~/ PhysicSystem al ~ ~ ~ .
and to allow full control of the operations. 5) are necessary to build the client and the server applications [4]. Different views of the distant equipment can be selected. Three components (Fig. the force applied on the cart (length of the horizontal line drawn at the base of the cart). such as the reference signal for the controller (triangle drawn on the target track position). The client application is made up of the GUI and the communication components. By isolating carefully these modules in the development process. 5 OperatorSide Features The user interface at the client side typically features an area showing a scope that displays all the relevant signals (see label 1 in Fig. or even a virtual hand used to apply a disturbance on the pendulum (see button 5 and the hand displayed in area 4 which appears when the button is pressed). Useful information can be added. The video camera and microphone can be seen as sensors. The main objective of such an interface is to provide the user with a general view of the equipment. it prevents unauthorized access and schedules login to avoid conflicts. the display is limited to the virtual representation which in turn can be animated using real data . The communication component allows the client and server applications to exchange information with other computers distributed in different geographical locations. In more advanced implementations the control algorithm may also be presented as a user option [5]. The client and the server application can be designed by adding a communication component to the two components used in local implementation: the realtime and the GUI ones. 6). For example. and also includes areas with dialog boxes that permit setting the controller parameters and specifying the access rights of the users (labels 2 and 3 in Fig. The server application is made up of the realtime and the communication components. it is easy to port a local solution to a remote one. 6).A Distributed Architecture for Teleoperation over the Internet 405 generate excitation signals and observe corresponding responses. The server application receives the client commands and transmits them to the equipment. The onsite server is the computer located near the real system and equipped with the hardware interface to the sensors and actuators. including an image. a virtual representation of key parts of the real system is superimposed on the video image (see the dotted lines in area 4 in Fig. 6). It also returns its physical and operational states to the client. This module also takes care of security issues regarding network management. The server may require a basic user interface for supervision of the ongoing operations. When the bandwidth available is not large enough to ensure a sufficiently high video throughput. To enhance the user perception. or port the remote solution to different physical systems. Such a composite view is called augmented reality.
: 9 : :> : :.~ Z Fig. In academia it allows access to and sharing of laboratory resources for collaborative research and flexible education. namely. the animation of the virtual image can be achieved from knowledge of only two variables. The inverted pendulum is a convenient introductory example that permits underlining the principal requirements related with the teleoperation of con . In the pendulum example. In addition. using samples generated by a realtime simulator.406 . The clientserver architecture implemented that features three hierarchical control loops poses the basis for efficient operations that require a high degree of interaction. . In industry it enables the monitoring and the maintenance of critical facilities. Denis Gillet et al. in the worst case. as well as quality of service constraints. In these cases.: :: :: : )m 84 I m 84 . It constitutes a framework for the development of adaptive schemes that cope with varying bandwidth and transmission delay. the angular and longitudinal positions of the pendulum. ::: : . . It constitutes an innovative realtime service that brings new opportunities for both industry and academia. it is important to post an indicator that announces that the data shown is provided by a model due to the absence of reliable network data. The implementation of a realtime simulator in the client software permits the user to carry out offline simulations for prevalidation purposes. User Interface of the remote client software   or. and provides the user with a sense of realtime behavior. 6 Concluding Remarks and Perspectives The teleoperation of controlled systems that exhibit fast dynamic behavior is a challenging application when it is carried out over the Internet. 6. the simulator can be used to provide synthetic data that can be posted on the user interface when packet losses occur during transmission. This gives continuity to the display. .
San Diego. Salzmann Ch.. (1998) Access to an Instructional Control Laboratory Experiment through the World Wide Web.H..A. Salzmarm Ch... Longchamp R. Philadelphia. San Diego. Gillet D. In: American Control Conference. 14651469 . In: American Control Conference. Janeiro. Latchman H. In: American Control Conference.W. In: International Conference on Engineering Education. The most important ones are: the necessity to ensure a cadenced stream of information to reproduce the dynamic behavior of the onsite system. USA. Piguet Y. the need for hybrid representation to enhance the operator perception of the ongoing operations.A Distributed Architecture for Teleoperation over the Interuet 407 trolled mecatronic systems.. Crisalle O. USA. Belgium 2.. (1999) lnteruetBased Client/Server Virtual Instruments Designs for RealTime RemoteAccess Control Engineering Laboratory.D.. Brussels.. Tzes A. References 1. Brazil 5.. Bhandari A. 14721476 4.Gillet D. USA 3. (1998) Requirements for RealTime Experimentation over the Internet. Shor M. Rio de. Bonvin D. Overstreet J. Gillet D. and the possibility to remotely perturb the systems for validation purposes. (1999) JavaBased Remote Experimentation for Control Algorithms Prototyping. In: European Control Conference. (1997) Telepresence: An Opportunity to Develop Practical Experimentation in Automatic Control Education.
Later. We apply Dirac's theory of constraints to singular linearquadratic optimal control problems and compare the results with our previous characterization of the "generalized optimal solutions" for this class of problems. we t u r n e d to the special case of s i n g u l a r l i n e a r . The Dirac's approach provides an alternative method to find the continuous term of the generalized optimal solution(s) by solving a set of equations involving Poisson brackets but omits the distributional term . pt A b s t r a c t . a n d it was suggested t h a t this could be related to gauge freedom. it was first developed in the early 1950's by P. Sarychev . [3].A.c o s t a t e in a way similar to t h a t devised by Dirac. In [7] the authors raise some questions concerning the m e a n i n g of gauge freedom in the context of o p t i m a l control p r o b l e m s a n d speculate if this could be related to a b n o r m a l extremals. the P o n t r y a g i n ' s M a x i m u m Principle generates c o n s t r a i n t s affecting the pair s t a t e . T h i s is an useful e x a m p l e because t h e appliThis work is part of the author PhD project at the University of Aveiro. who generalized the t h e o r y to cases where the phase space has no n a t u r a l s y m p l e c t i c structure. In it's original form. Volckaert a n d Aeyels [7] p o i n t e d out t h a t in singular o p t i m a l control problems.M. it was recast in the language of s y m p l e c t i c geometry by Gotay. do Quelhas 6 1200 Lisboa. The generalized optimal control(s) for a singular LQ problem is the sum of a realanalytic function with a distribution of order r concentrated at the extremes of the time interval and the corresponding trajectory is the sum of an analytic function with a distribution of order (r . Also.ISEG R.q u a d r a t i c (LQ) problems. Hence it should be possible to a p p l y a D i r a c . Portugal mguerra@iseg. under supervision of A. it is known t h a t some s i n g u l a r o p t i m a l control problems have discontinuous "generalized o p t i m a l t r a j e c t o r i e s " .Singular LQ Problems and the DiracBergmann Theory of Constraints Manuel Guerra* Universidad Tecrtica de Lisboa .1).B e r g m a n n theory of c o n s t r a i n t s provides a m e t h o d for finding singular e x t r e m a l s for a p r o b l e m of v a r i a t i o n s in L a g r a n g i a n or H a m i l t o n i a n mechanics. Dirac [2]. Nester and Hinds [4]. In order to address these issues. 1 Introduction T h e D i r a c .B e r g m a n n . We show how the DiracBergmann approach can be modified to allow for the distributional terms.utl.t y p e theory to this kind of problems.
+oo[ is fixed.~ = {u E L k [0. For each u E L k [0. ~ E / R '~.~ onto the topological completion of L k [0.~(T) = ~. Utilizing this notation. In [5] we showed that the generalized optimal trajectories of a singular LQ problem are sums of an analytic function with a distribution concentrated at the extremes of the time interval. T]. T]. LFrom the functional theoretical point of view. and showed how the generalized optimal solutions can be approximated by ordinary (suboptimal) solutions. ~ E /R '~ are fixed points and T E ]0. P E E/n• R E /R~• are symmetric matrices. It is the consequence of the existence of an infinitedimensional subspace of the space of controls in which the cost functional is constant. u). 7 ] : x~.. In the section 2 we give a brief account of our method to solve singular LQ problems and present its main results. Guerra cation of the DiracBergmann approach to these problems is straightforward and we can give an exhaustive characterization of the generalized optimal solutions using different methods [5]. T] with a weaker topology and extend both maps u ~~ J r and u ~~ x~. x(T) = 5. The DiracBergmann approach yields the analytical arc in the optimal generalized trajectory(ies). u) = foT x ( r ) ' P x ( r ) + 2u(r)'Qx(r) + u ( r ) ' R u ( r ) dr. x(O) = ~. let J r (u) denote the functional u ~+ J (x~. we provide L k [0. u E L k [0. A E IRn• B E /R"• and Q E /Rk• are arbitrary matrices. (1) where J(x. we explore the relationship between both methods and show how generalized optimal solutions can be introduced in Dirac's method.410 M. We show how generalized solutions can be introduced in the DiracBergmann framework. xu. T] . Also. T].~. T]. This paper is organized as follows. This gives some further insight about the structure of the generalized optimal solutions. but gives no information about the distributional terms when they are present. u E L k [0. one m a y represent the problem (1) in the form J~(u) ~ min. let x~. with domain in L k [0. u) + min = a x + Bu. Below. x (0) . In the section 3 we apply Dirac's method to the same type of problem.~ (T) = 5}.~. x E A C " [0. It turns out that gauge freedom is related to none of the phenomena indicated above. 2 Generalized Solutions for Singular LQ Problems Consider an optimal control problem of the type J (x. T].y denote the corresponding trajectory of the system & = A x + Bu. the problem is the one of finding a minimizer of a functional in the affine subspaze//~.
T] ~+ L~ [0. by H~ i [0. The problem (1) has finite infimum if and only if the extended functional has a minimum in the closure of//y...Singular LQ Problems 411 relative to this topology.T]. The singularity of the problem means that ker (R) r {0}.e. (2) R1 = R + H + ( Q B . then the functional u ~~ Iiul]! . r = r162 i = 1. .~ = z~u.~ + B / / ~ where zv. The topological completion of L~ [0. By integrating H~ and leaving II+u unchanged. let r i~ [0. 2. This generates a sequence of linear operators r L~ [0. i. one obtains: xa. The inputtrajectory map can be written as xu. T]. where 7/ is an Hilbert space with norm II'lin. T] ~~ i k [0. Let B1 = B H + + A B H ~ Q1 = H + Q + I I ~ 'PQA).~ with respect to the new topology. Cu(t) = ~0t u(r) dr. This space of generalized controls is a subspace of a Sobolev space.QAB .QB)II++ +II~ . k defined by r = Id. i.2.3.B ' Q ' ) H ~ + I I ~ ' . u C L~ [0. with r _< n. T] ~ L~ [0. T] with respect to the topology of II'l[! is denoted by H I. i E P/.If f is injective.e. u E L k[0.1 G e n e r a l i z e d C o n t r o l s o f Class //k_i [0. In particular.. Decomposing the control space into ker(R) and ker(R) • l e t / / + and//0 denote the orthogonal projections from •k onto ker(R) • and ker(R). t C [0.~(t) : e t A x + ~0t e(tT)ABu(v) dr.]lfulln is a norm in L k [0. Har [0. T] defined by ~pu = H + u + II~162 we have: xu. T].T]. T]. respectively. T] Consider a continuous linear operator f : L~ [0. i E SV. T]. t E [ 0 . T] denote the primitivation operator. we denote He. T] ~~ 7/..T].y(t) = etA~ + ~0t e (tT)A ( B I l + u ( v ) + ABII~162 dr + B I I ~ 1 6 2 If we consider the operator ~o: L k [0. T]. For these operators we keep the usual notation.'~ denotes the trajectory of the system k = A z + ( B I I + + A B H ~ z(O) = ~.B'A'Q')H ~ . T] and defines a topology that is weaker then the usual topology of L~ [0.
the following conditions must ufi~ hold II~  B'Q')II ~ = 0 (4) R1 > 0. Appendix 2]. z e A C n [0. 7]. Y. where ~ : L k [0. 7] and the closure of H~. In order to exist some linear subspace. the righthand side of (2) and (3) give the unique continuous extensions of the maps u ~+ x~. .~ + 2(~ou)'Q1 zeu. it follows that u E//_ inf Je(u) = r = x. (ii) the condition (4) holds and (5) is strictly positive. satisfying inf J0 (u) > <x>. Thus. T].z(T))' B # ' Q (x + z(T)) ~ min k=Az+Blv.~ = {u 6 H ~ : z ~ ( T ) + dl (z. <> (5) These conditions are known as Goh and generalized LegendreClebsch conditions. of finite codimension in L k [0.=.~ + 2 ~ou' Q l z~. (II~ It is clear that H 7 c Hk_l [0. respectively. Guerra By substituting (2) in the definition of Jy and performing all possible integrations by parts aiming to eliminate the terms that contain II~ one obtains J~( u ) = f o z~.B'Q')II~162dr+ +2(//~ (3) ( T ) ' Q z ~ . In the cases (ii) and (iii).r + (~ou)'R1 (~ou) dr+ +(x .r(T)). . the equality (3) reduces to J~(u) = f [ z. T]. ~ ( T ) + (II~ (II~ In order to proceed we need the following Proposition [1. P r o p o s i t i o n 1. The transformations (2) and (3) must have one of the following outcomes: (i) one of the conditions (4) or (5) fails.~ and u ~ Jy(u) onto the space HT.~P z§ . In the case (i).412 M.~ with respect to the topology of H 7 B(II~ ~}. but (5) is only semidefinite.ePze=. we may consider ~0u as a new control and this reduces the problem (1) into a new LQ problem is H~.~(T))'B#'Q(~ + z~=. .x and hence there exists no optimal solution in whatever sense.~ + ~ou'R l ~ou+ T +(II~ . z(0) = ~. z(T) E ~ + B I I ~ v E L~ [0.e. (iii) both the conditions (4) and (5) hold. 7] • ker(R) is the operator defined by ~u = (~u.. Let B # denote the left inverse of BII ~ i. B # B I I ~ = v. v) + (x . for all v E ker(BH~ • Then. 7] ~+ L k [0. for all u E H.z~=.
etc.Singular LQ Problems 413 with Jl(z.Fxu) + (Id . P0. we consider a sequence of transformations. n ~ (O.I ~ ) = O. {~i... For Jo to have finite infimum in some linear subspace of finite codimension in L~ [0. all problems of the type (1) can be reduced by a finite sequence of transformations of the more general type u .Fx. (8) For a given ~ 9 IR n.Ri= RN . (9) In [5] we showed how to chose a sequence of transformations such that each ~i can be partitioned qai = (~B. it is necessary and sufficient the existence of an integer r < n such that the conditions (6) to (8) hold for all i < r and either R~ > 0 or Br = O. B0. . For this particular sequence we have the following stronger version of the Proposition 2 : P r o p o s i t i o n 3. and F 9 /R kxn (see [5] for details). Q i = i .x v'Qi ( ~ A i ~ . Qi. then this is a regular LQ problem. and the corresponding matrices have the s t r u c t u r e B i = ( /3i0 ) . Pi.). A ~ . Ri > 0. Q. R). T]. . . obtaining a larger space of generalized controls. T]. it is necessary that all the following conditions hold for all i > O. . this is a new singular LQ problem and we may repeat the procedure. 2.2 The General Reduction Procedure There exist problems which can not be reduced to a regular problem by any finite sequence of transformations of the kind outlined above. If (ii) holds. n ~ = 0. H~2 C H _~ [0. . Thus. T]. (Bi A i B . starting with (A0..t B. B. corresponding to a sequence of LQ problems with matrices (Ai. w i t h B i having full column rank. Vv 9 ker(Bi) f] ker(Ri). k 2. the following Proposition holds..O. qag). . Ri) i = 1. v'Q.2.L)(u . i = 1. Bi. For any such sequence.e ~ also implies uEL/_ = x. (6) (7) Vv 9 ker(Bi) M ker(Ri)..> CL(u . If (iii) holds. (> . the existence of some T > 0 and some ~ 9 ~'~ such that inf Jr(u) > . . P r o p o s i t i o n 2. Q0. v) = f [ z ( r ) ' P z ( r ) + 2 v ( r ) ' Q l z ( r ) + v ( r ) ' R l v ( r ) dr.. For Jo to have finite infimum in some linear subspace of finite codimension in L~ [0. A ~ .}. P. where L is an idempotent matrix mapping ~ k into ker(R). ) = O. However. R0) = (A.
z 2. 7"] and a distribution of order (j . where J~(v) = for x~:~Prx. into the controllable space of (A~. (10) where T is a projection of/~2. ~(0) =~.. in [0.1).~+ C RrBv dr. R~Cu ) + Cr (~. Br) (this is always noncontrollable). W. If the conditions of the Proposition 3 hold and the conditions (8. either from the Maximum Principle or from the solution of an adequate Riccati differential equation. as well as the correspondent trajectories [5]. I f inf Jr(u) > .( C u ) + (~u. it is possible to characterize the problems of the type (1) whose infimum is finite.r. T] is optimal for the transformed probh lem (10) and w is any squareintegrable function such that w(t) E ker (Re) a. Approximations of the generalized optimal control by smooth functions can also be computed. the generalized optimal controls for the problem (1) are exactly the generalized controls that can be represented as = ~ .. x rv. uELt_= z. where ~ E L2 [0. r z + v'RrBv d r + min ~=A~+~rv. and Cr(') :/R 2n ~ / R is a quadratic form. T].3 Properties of the Generalized Optimal Solutions Using the reduction procedure outlined above.denotes the tra'~ jectory of the system k = A r x + Brv. Thus. then the extension of the functional onto Hffr satisfies J~(u) = J~. 0) + w). T h e o r e m 1. .r(T). Guerra We call the number r the "order of singularity" of the problem. The generalized optimal solution(s) for those problems can be computed. This reduces the problem (1) to the new problem: for z'Pr z + 2v'Q. concentrated at t = 0 and t = T.~ + 2vy. this shows that the problem (1) has some generalized optimal solution if and only if inf J~(u)>c~o.~. T] and a distribution of order j < r.. T]. z(T) =r~+(Idr)~. 9) hold for all i < r. z E A C n [0.e. ~). concentrated at t = 0 and t = T. ~ U~. Since the transformed problem (10) is either regular or trivial.t ((~. Below we present just a few results that are most important for the discussion of Dirac's approach presented in the next section. x(0) = ~.~ . 0 The presence of distributional terms in the generalized optimal trajectory is important because it implies that any minimizing sequence of continuous trajectories must be unbounded in a very precise sense [5].414 M. T]. The corresponding trajectory is the sum of a realanalytic function in [0. x"v. then there exists at least one generalized optimal control that is the sum of a realanalytic function in [0. v E L~ [0.
and + u)(t) ker (nr) a. and a unique realanalytic control. B. In the next two results we are using the sequence of transformations required to obtain the Proposition 3. then w and (u + w) are the smallest generalized optimal controls with respect to the norm [[11~ . ~ E 1~n such that ~ can be reached from ~. <> The conditions (ii) and (v) above mean that. P r o p o s i t i o n 5.x = x for all ~. and that there exists some point in X that can be reached from some point in X . B ~ _ l H ~ This coincides with the space of j u m p directions described by Jurdjevic [6].~(T) = ~ . (v) + u) = 0. In order to inf d~(u) > . For each ~ E X. uE/d_ = x.~)' M (a + ~ ) + ( ~ . in [O. x such that ~ can be reached from ~. ./_ . there exist unique ff E X . i + A . A s s u m e that the problem (1) satisfies inf Jy(u) > . T]. ~ E X.x for all ~. R).a)' M (~ + ~ .. where A is a distribution of order ( j .l ) concentrated in {0}U{T}. T]. "ff E X .e. in [0.~ = x ~ . Let X. (ii) ~oNw = O.o+~. <> . ~). and ~o~w(t) E ker (Rr) "L a.. P. v'Qi = 0 f o r all v E ker (Bi) fq ker (Ri) and all i < r. This control also satisfies (vi) x. an optimal generalized control is ~ = w + u where u is the unique distribution that satisfies all the following conditions (iii) xw+u. and either B . f o r which the following conditions hold (i) w is optimal with respect to (~. X denote two translations of span (BI1 ~ B I I I ~ Br_ 111~_ 1). Q. Then.0 Proposition 4.e. where M is a real matrix depending on (A. it is necessary and sufficient that the conditions of the Proposition 3 hold.. u E/. if the generalized optimal controls are not unique. = 0 or the transformed problem (10) is regular and has no conjugate point in the interval [0. w.~ .Singular LQ Problems 415 Any discontinuities and/or distributional terms that may be present in any generalized trajectory must lie in span ( B I I ~ ~ .~ . (iv) u is a distribution of order j <_ r concentrated in {0} tO { T } .T]. The corresponding cost is J~ (w + u) = J~(w)+(d .
71. 1 < j < I.~.e.g _. ( A .~) + [ H ~ ~ (x. For simplicity. H~ = B~ . we assume that the coordi nates of the control space (/R k) are such that R= 0 0 ' B= (B + B ~ Q= Q0 . with R + > 0. g (x. g] to be the m a t r i x whose entries are the functions of (x. Hence.~): xt (Q+! (/~+)i Q+ _ p). u= U0 . and B~ . Dirac called the condition (12) the primary constraint. u) : IR ~n+k ~+ IR h. u) : iR2n+k ~_~j~t. T]. W h e n we proceed through the sequence of constraints. T].~. ~. Vt E [0.~)u ~ = 0. in [0. H~ u~ satisfying H ~ (x. and the conditions that arise by differentiating H ~ i times he called the secondary constraints of order i.2Q~ = 0.1 M. H+] + [Id.g]i.~). one of the following must obtain: (i) the secondary constraints up to order j solve to give u ~ as a function of (x.416 3 3. in [0. a.j = [fi. . Guerra The DiracBergmann Approach The Dirac's Constraints For any pair of s m o o t h functions f (x. u E ~ k maximizes the Hamiltonian function at the point (x.e.2 Q + x ) .~) = 0. ~. (11) (12) where H+(x. this generates a sequence of equations involving higher order Poisson brackets. Using the n a m i l t o n i a n equation (11).2Q~ T h e condition (12) implies ~Hd 0 (x. 1 < i < h.B+ (R+)I ~! Q+)x + I('B+ (R+) 1 B+'~. u)defined by [f..~) E R un if 1 and only if u = ~ (R+) 1 (B+'~ . Since the higher order derivatives of H ~ must also be identically zero. we define the Poisson bracket [f.~) =[Id. the M a x i m u m Principle states that any absolutely continuous o p t i m a l trajectory for the problem (1) must be the projection into the statespace of some trajectory of the linear control system (k.~) = 0 a. (ii) the secondary constraints up to order j leave at least some coordinates of u ~ unsolved but the restriction of order (j + 1) can be expressed as a linear combination of all previous restrictions (including the p r i m a r y restriction). this reduces to [ H ~ +] (x.gj] = ~ o~ (~ .
.Q..B + B~ O. v ~ = H + (x. (13) = (2R~'++l O0) .~) = B+'~ . LQ problems never generate incompatible sets of constraints.~)..= ~ . Since all constraints can either be solved to u ~ or are linear homogeneous in (x.1. the maximized Hamittonian" is Hi (~. 3. v +. i i B... consider a sequence of transformations as described in the Section 2. with all the R + nonsingular. v~ For each one of these transformed problems.. one obtains [H ~ H ~ = 2 (B~ [[H~176 0 0 ~ ..2Q+x. i j=0 1 . If (ii) holds..~). parameterized by the set of square integrable functions that m a p [0.. Proceeding in the 0 0 same way.. starting with ~1 ~ = ~. If (i) obtains. R. with R + nonsingular. ~ 0 ) This yields R1 = R + 0 .P)x + i 1 ~'(~ .= (B + . (~0B~+ ( R + ) ___ j=O B?'~. u0) = (~+ + (R+)I (Q+ B0 _ B+. ~. then we just have to substitute u ~ in the Hamiltonian equation (11) to obtain the extremal trajectories.O0. ~). v+.Singular LQ Problems 417 Hence the procedure must terminate in no more then 2n steps.2 Relationship Between Dirac's Approach and the Generalized Control Approach In order to see the relationship between Dirac's approach and the previous one. B.E B? (R~) ~ ~ ) ~ + ~ .~) = x'(~"~ Q+' (R+) 1 Q+ . we obtain a sequence of transformations such t h a t the matrices of the corresponding transformed problems take the form 00) ' ~~ i "'. then for each pair (~.2... (a~+~ [H~ H+] = \ H~+t ] ' /=0. ~) + v~ where H ~ (x. We also consider the functions Gi(z.. the problem (1) is said to have gauge freedom.0+ Q. i = 1. . The corresponding transformed control is ~. = / 0 \0 .2Q~ g + (x. we have an infinite set of extremal trajectories. . Utilizing this notation. R~ 0 / ' 0 0/ o.u = (v +. T] into the remaining free coordinates of u ~ In this case. ).2. (~+.~) = B~ i ~ (x.) ~ 0 .
it is necessary that the following conditions hold for all i >>O. both approaches proceed by computing essentially the same sequence of matrices.. H + ] . (14) while the (2i) th secondary constraint reduces to u. g +] = 0. x(T) = a must be unique. ~u~. . P r o p o s i t i o n 6.Y+_~] . .+ = 1 ( R + ) . It is clear that the analytic trajectory described in the Proposition 4 must be the projection into the state space of some solution of the Hamiltonian system (11) that satisfies all the Dirac's constraints. . 7) conditions. Conversely.~). then there exists some T > 0 such that the generalized optimal solution exists and is unique for each (~.] . This implies that the Dirac's set of constraints completely solves u ~ as a function of (x. If Rr > 0. For Jo to have finite infimum in some linear subspace of finite codimension in L~ [0.~ It follows that the sequence of transformations required by the Proposition 3 yields Rr > 0 if and only if the sequence of transformations implicit in the DiracBergmann approach yields Rr > 0. [[H~ .H+_. The case when the optimal solution is not unique but the set of optimal solutions has finite dimension can only arise when T is the first conjugate point of the transformed problem (10). Contrary to the generalized control approach.H+_~] ... .H+_~]. T]. . . In this case the solution of the Hamiltonian system (11) that satisfies all the Dirac's constraints and x(0) = ~. Thus.. (15) (16) [. Guerra Hence we can give the following formulation for the Goh (4. [[H~ . . 6) and generalized LegendreClebsch (5. . . ~).1) th secondary constraint reduces to [. the DiracBergmann approach does not give a general characterization of the problems which satisfy inf J~( u ) > oo. H +] = 0. if the Dirac's constraints completely solve u ~ as a function of (x.. and inf Jr(u) > .[[Gi. Thus. [ H ~ 1 7 6 = 0. the problems that have an infinite dimensional set of optimal solutions are exactly the problems that have finite infimum and have gauge freedom..418 M. ~ such that ~ can be reached from ~. then the generalized optimal solution must be unique.~ . the ( 2 i . . .i [. [[H~176 Assuming that the conditions of the Proposition 6 hold and using the equalities (13). . .
~. (20) 9 r = ~ (R +) 1 1 ( Gi (x'. H+] (~.0. ss+] . 1 < i < r. ~1) does.~ r) . ~) satisfies the primary constraint (12) if and only if (zx. H 0] ~.~. Using the same argument we used to extend the map u ~> xu.~ l) . it follows that (x. H ~ (xl.~1) r1 + [sd.~.~1) = [s~. B) is controllable and the generalized optimal solution exists and is unique for each ~.i G1 (xl. ~ E ~ " .~ 1) .~l). that is.3 Generalized Trajectories in the 419 DiracBergmann Framework In the rest of this section we assume that (A. it follows that (x 1.)+ [[sso.~1)~o: o. H~] r i=O (19) where (x'. Using the Proposition 6. Using (17). Repeating the same procedure.~ 1) denotes the trajectory of the system (k.~") satisfies (x". sso] ~uo. H +] [Id. H~ r j=i o/ .~r) + B [Id.~'). This is not essential but avoids the lengthy analysis of the degeneracies that arise when the generalized optimal solution is not unique and/or exists only for some pairs (~'. (21) ..~1) must satisfy [sso. s++] (~1. one may show that (~. where (xl. and let u = (u +. one obtains (17) (18) (~. [H ~ H+]] Cu~ (x(0). The Dirac's set of constraints reduces to H~ (xr. u ~ denote the corresponding control. 1 ( R + ) .Singular LQ Problems 3.sr (~1.0 = (~1. ~ . ~ . (x"(O). Let (x. 1 < i < r. By differentiating H ~ (zl. H +] (x".~(0)) = ( ~ . ~") + Z [Gi.~"(O)) = (~..~) = [Id. one obtains (x.~) denote some trajectory of the Hamiltonian system (11) with (x(0).~") = [Id.0.~(0)) = ( ~ .0.~1) + [Sa. ~ .~) = (x".
Thus we define a generalized extremal to be an image of some generalized control. by some map u ~ (x .) . The projection into the state space of this generalized extremal is the generalized optimal trajectory for the pair <> For each 9 9 /Rn. 21).~.. u ~~ (x ~.<> The equations (19. ff = 0 spo. x(T) = ~. a) 9 gt 2" that satisfy all the Dirac's constraints (14.to the quotient space "/spa.r. :D = {(a.. 21) are well defined. Hence.~.r}. T[ ). For each ~.~(0)) + if) N V.. The map (~. (B ~ g ~ . .~ can also be used to obtain the unique extension onto H~r of any inputtophasetrajectory map. Guerra The Hamiltonian system (11) is linear. we have the following correspondence between generalized extremals and generalized optimal trajectories.. all the terms in the equations (19.r..[Zd. The following Proposition gives a geometric interpretation to the Proposition 4. For any u 9 H ~ . but "jumps" at time t = 0 to a point (x(0+). Not surprisingly. the argument we used to obtain the extension of the map u ~ xu. B~ .~. In the interval ]0. x~.. Proposition 7. fi E H ~ .. This trajectory reaches a point ( x ( T . 21) can be used to compute approximations of the generalized optimal solution by (suboptimal) ordinary solutions in a way similar to [5]. 2. i = 1..~(0+)). such that (~. ~) (this is uniquely defined in the interval [0.a) 9 h~2n: H O t (a. a ) = 0.) ) .r.~). For each fixed 9 the map is onetoone from 7/~ onto ( 7 / ~ + ff) N ~ .~. For each ( ~ . ~ 9 7/..~) 9 ( x ( T .. i = 1. 20.~) satisfies the equations (19.. starting at (x(0+).~) 9 7 / : x = ~}. H ~ HO]. A generalized extremal has the following structure: it may start at any point (x(0). It is possible to prove that ff C 7/. 16).~(0)) 9 7/.g)...420 M. 20. from which it is possible to "jump" to any point (~. cr) = 0./[~. ( a . let (xy. T[ the generalized extremal coincides with the unique absolutely continuous trajectory of the Hamiltonian equation (11).r.2.. the projection of 7/ into the state space is the whole state space and :D is the set of all (a. let 7/r = {(x.~. and lying in ~ .~ 9 IRn there exists one unique generalized extremal that satisfies x(O) = ~. / \ Proposition 8.) . ~(0)) .a) e n ~ n : H ~ G .. r}.g.~) ~ x~~(T) is onetoone f om y ) n V o. 20.~.) ) + .~. . ~ ( T . Let 7/ = {(a. H.(~.~) denote the generalized extremal such that (z(0). ~ ( T .~(0+)) 9 ((x(0)._~]).7.([Id.
pp. Dirac. Ann.13 N~ 635690. 2. (1964) Lectures on Quantum Mechanics. Arizona USA.Singular LQ Problems 421 References 1.(1996) Abnormal SubRiemanian Geodesics: Morse Index and Rigidity. Agrachev. Vol. Hinds. V. M. on Decision & Control.(1978) Presymplectic manifolds and the DiracBergmann theory of constraints. Jurdjevic.A. 19(11). G.J. 2. J. Dirac. Poincar~ .A. J.: Generalized Hamiltonian Dynamics (1950) Can. 129148. D. Guerra. Nester. K. Inst.V. (1999) The GotayNester algorithm in Singular Optimal Control. 6. 5.. Volckaert. 23882399. Gotay.M.A. J. Belfer Graduate School of Science. 3. M (2000) Highly Singular LQ Problems: Solutions in Distribution Spaces.Analyse non lin~aire. J. P. Phoenix. 4. Sarychev. (1997) Geometric Control Theory. Math Phys. Cambridge University Press. Proc 38 th Conf. 873874. Vol.. A. Math. A. P.M. . 7. N~ 265309.M.6. Aeyels. Dynamical Control Systems.
These This work was financially supported by the German Academic Exchange Service (DAAD) . After a system transformation we design PlDlike tracking controllers for the resulting subsystems taking nominal parameters into account. that we design a relatively simple tracking control for the nominal system and provide thereafter an analysis tool for determining its robustness with respect to the a priori given bounded parametric uncertainty. The article is concluded by a DC drive example. These studies were accompanied by the question of robustness of the developed controls: a series of papers by Schmitendorf and coworkers (see [19]. In this article the robustness of tracking controllers acting on multivariable linear systems under parametric uncertainty is investigated. the parametric uncertainty can enter the linear state equation nonlinearly in the respective system and input matrices. [11] and the references therein). [20]. 1 Introduction Tracking control of controllable linear systems has been studied profoundly in the last thirty years (refer to [6]. The work presented in this article differs from the aforementioned results with respect to two points: first. Robustness is studied under parametric uncertainty: it splits into the robust stability of an autonomous linear system and into the robust stability of the same autonomous system being perturbed by socalled "quasiexogenous" signals. [21] and [12]) provide a good overview of the attained results. Second. France hagenmey@Iss. fr A b s t r a c t .R o b u s t Tracking of Multivariable Linear S y s t e m s under Parametric U n c e r t a i n t y * Veit Hagenmeyer Laboratoire des Signaux et Syst~mes CNRSSup~Iec Plateau de Moulon 3. We extend in this article the results already obtained for the SISO case [8] to the MIMO case: by transforming the system into its controller form we develop control laws for the nominal value of the uncertain parameters. rue JoliotCurie 91192 GifsurYvette Cedex. [10]. supelec. In these papers constructive tools for choosing the right controls with respect to given bounded uncertainties are given. we follow a different philosophy in the sense.
. In contrast to the SISO case. x(O) = xo (I) where the time t E IP~.~il. the state x(t) e IRn.424 Veit Hagenmeyer control laws decouple the respective subsystems. q (2) where po is the nominal value of the parameters. The paper is organized as follows: after formulating the robust tracking control problem in section 2. we conclude the paper with a discussion in section 9. We establish the following definition: . it is assumed. followed by important remarks on controller design and tracking performance in section 7. we establish the tracking control laws for the nominal system in section 3. For both parts known stability criteria are stated which lead to the main result of the article. The main result is stated in section 6. The state vector is assumed to be known via full state measurements. . but not known by their exact values: p=Po+~. the decoupled nominal subsystems of the MIMO linear systems recouple when the uncertainty intervals are taken into account. inject a desired trajectory behaviour and stabilize the system around the desired trajectory by using PIDlike control techniques. The system matrix is A(p) : IR q ~~ IR n x n and the input matrix B ( p ) : IR q ~~ IR n x r . The fundamental hypothesis we impose on the system is that of system controllability for all possible p in (2). the input u(t) E IR ~ and the system parameters p E IRq. 2 Problem Formulation Given the MIMO linear system ~(t) = A(p)x(t) + B(p)u(t). In the following section 4 we investigate the according tracking error equation and analyze thereafter its stability in section 5. without loss of generality. /3i e[P~. that the r columns of B ( p ) are linearly independent. Nevertheless. which are constant in time. In addition. . After treating a DC drive example in section 8. second the stability of the same system being perturbed by so called "quasiexogenous" signals at the input. We investigate tracking stability in analyzing robustness of the tracking error equation under the uncertainty of the parameters. i = l . . the robustness problem splits into two parts: first as the stability of a linear autonomous system under endogenous (parametric) uncertainty.
. i = 1 .1 Tracking Control S y s t e m Transformation The theorem noted below holds for interval parameter systems (see [161. the system (1) can be transformed into ~(t) = 2(p)z(t) + h(p)u(t) (4) The transformed system matrix 2(p) reads as . with fixed controllabil. . all additional coupling enters each subsystem at its input. 2 . Thus. r . Furthermore. . R e m a r k 1 If a linear system (1) is subject to a tracking control law.Robust Tracking of Multivariable Linear Systems 425 D e f i n i t i o n 1 The tracking of the desired trajectory t ~ xd(t) is robust in the presence of parametric uncertainty with respect to the interval parameters in (2). for all p in (2). 3 3. Then there is a nonsingular transformation z(t) = S(p)x(t) (3) of the state vector x(t) and a nonsingular transformation of the input vector u(t) which reduce the system to a coupled set of r singleinput subsystems. . ity indices pi. the quasiexogenous perturbation term follows necessarily from the desired trajectory injection under parametric uncertainty. Each subsystem is in the singleinput controller form. its tracking error dynamics represent a nonautonomous system under parametric uncertainty: the calculations in this paper show. The control task we study in this paper is to track a given sufficiently smooth desired trajectory t ~~ xd(t) and to guarantee robustness for this tracking. that the tracking error system is perturbed by a socalled "quasiexogenous" perturbation at the input. [22] and [18] for nominal linear control systems): T h e o r e m 1 Suppose the system (1) is controllable. if for all parameters p in (2) the system (1) under the tracking control law the state remains bounded.
each subsystem can be represented as (9) zi.. r ) /3 = diag(flx. i.~(p) ~.i(p)= " 00] " .(p) i. . Furthermore it follows from [16] and [18] that/~(p) can be represented as h(p) = Br(p) (8) Thereby is F ( p ) E IR rx" an uppertriangular matrix with elements denoted by 7id.2(p) / (5) Ar.. The matrix/~ in (8) takes the following form (defining fli=[0 0..r. r.. .r(p)] A2..Pi Pi pj   1} r ~. j + l .....~ (p)J where each 1001 Ai. . j J i (7) 0 is an IRp' • matrix.. fir) Hence.2 (p) A2.0 1]TEIR p'.. + Z ~.~.~. + Z Z o. .Z o.426 Veit Hagenmeyer ~i(p) = "Al.... .~..~(p) is an ]Rp~xp~ matrix and each o Aid(p) = i ~.1 (p) A2.(. i= l...~. ...... i = l ....1 (p) Ar.j = l. .r (6) o o . ~(p)u~ k=l j~i k=l k=i (10) ...l(p) A1. .j = z i .(p)z....2(p)"'" Al.. . j E {1.~(p) ~.~ (p) At. j = 1 .2.)z.
r.1 } pl Pj =E k=l + E Zo. i = 1 . We stabilize the system around the desired trajectory using a PIDlike controller in placing the poles. (12) for i = 1 . to find the representation of the nominal system z(t) = A(po)~.d and the desired t ~+ zi. The nominal subsystems read therefore (compare with (10)) zi. .2 Control Law Design The tracking feedback control law to be developed is based on the nominal parameters Po given in (2). . .Robust Tracking of Multivariable Linear Systems 427 for i = 1 .j = Y. . j E { 1 . . First we execute the nominal system transformation 2(t) = S(po)x(t) (11) as in (3). . r.p.[4]). . The desired trajectory t ~~ ~.(t) to be injected are calculated from the desired behaviour xd(t) using the nominal system transformation (11) to get the expressions for the outputs of the respective chain of integrators 5d i. r are often the tobecontrolled variables in real applications (see [3]. Since r(po) is uppertriangular and invertible for all p under the assumption of controllability.(pol jr k=l + v. . The tracking feedback control law is constructed via the Brunovsk# form [22] for this nominal parameter set. a new but equivalent set of system inputs can be defined as v = F(po)U. r .~(t) (13) the control law with respect to the new input v can then be represented as . . .times with respect to time. Defining the nominal tracking error by ~(t) = ~d(t) . . For these the given controller form is natural and easy to implement (see for instance the DC drive example in section 8). we decouple the nominal subsystems by canceling the coupling terms at the input and second inject the desired nominal behaviour at the inputs of the decoupled chains of integrators. i = 1 . . .(t)+ [3(po)U(t) ^as in (4) with according nominal matrices A(po) in (5) and [~(po) = SI'(po) in (8). . . Thereafter these are derived pi. 2 .j+z. . .z. . P I . that is.i.t(t).. 3. The zi. . .
but messy and very tedious. .. = S.0.~) and rintegrals.k (Po) and 7i. . We denote this linear integrodifferential equation system in z as Z for the ongoing. i = 1 . 9 9+ Ai.428 Veit Hagenmeyer Pi PJ v..k(Po. we are going to establish the tracking error equation resulting from the problem posed beforehand..0 ~. . . Plug (14) and u : F ( p o ) . The problem to be solved in this article can now be stated as: does the control law (14). s p~ + .j. . stabilize the linear feedback system containing the unknown but bounded p a r a m e t e r s / ] around the desired trajectory ~. . . . which was developed for the nominal case ai.. po)z with S. i= 1. 2. .) = S(P..i. = k=l + F_. . i = 1 .p. .)S(P) 1. j = 0 . .(p. but we refrain from giving the exact formulae for every single coefficient 1 within the resulting equation for the general case (see section 8 for a specific example). (15) 1 We propose them to be calculated by symbolic computation for higher order applications. Therefore we go on in describing the algorithm of calculation and present the structure of the result thereafter. such t h a t the corresponding characteristic polynomials s p~+x + Ai.d? 4 The Tracking Error Equation In this section.(p. Use (3) and (11) to find 5. Pi are chosen.. .ldr+E)q. 5 (remember ~ = 5d(t) . 'd (ii) (14) +Ai.Po) in (10).k . r.. The algorithm to find the tracking error equation is as follows: 1.k~i.. (~) fO t Pi j ~ i k1 + z.r k1 (i) : cancellation of nominal terms (ii) : desired behaviour injection (iii) : stabilization around desired trajectory The coefficients Aid..l v in (4) to get a linear differential equation system in z which contains expressions of 5 d. The necessary calculations are straightforward.p. r are Hurwitz.
We remark that the respective subsystems of the tracking error equation are not only recoupled under parametric uncertainty.z(t) (16) 6. Plug (15) in Z to get . which contains only 5 as exogenous variables.. Hence.~.0 1]T E IRp'+I.p) : IR • ]Rq ~~ lRr• where g'(zd(t).p) is diagonal and every diagonal element is a linear function of zd(t).d(t) .. For this purpose. we set g'(zd(t). an augmented system of dimension n + r. the robustness problem has to be divided into two questions: 1. Use . 5.d 4. Is the autonomous tracking error system (~(zd(t). and q'(zd(t).0 in (17) to get ~(t) = ll(p)e(t) (18) . Does this linear system stay stable under the quasiexogenous perturbations q'(zd(t).~ and (16) to find the searched for tracking error system.1 Stability Analysis Stability of the Autonomous Part of the Tracking Error System In this section we study the stability properties of the internal part of the tracking error system. B is as defined as in (9) with fli = [0 0 . Define the real tracking error as e(~) = ~. . moreover the controllers for each subsystem are intertwined in their effects on the different subsystems.Robust Tracking of Multivariable Linear Systems 429 3.~ one time with respect to time to get ~. . r.p) ? R e m a r k 2 We call this perturbation as being "quasiexogenous" to make the difference with respect to real exogenous perturbations which will be included in this theory in a forthcoming publication. Derive . It is of the form 6(t) = n ( p ) e ( t ) + (17) The tracking error system m a t r i x / / ( p ) E ]R(n+r) • is of the structure of A(p) (5). . 5 5. .p) .p) = 0 in (17)) perturbed by endogenous parametric uncertainties stable? 2. . i = 1 . .
the stability can be deduced analytically in a necessary and sufficient way by the famous theorem by Kharitonov [13].1. 5. in which the coefficients of the characteristic polynomial are algebraically dependent with respect to the interval parameters. T h e A l g o r i t h m o f W a l t e r a n d J a u l i n A numerical algorithm was developed by Walter and Jaulin [23]. It determines in a necessary and sufficient way the stability of the interval characteristic polynomial resulting from the calculations above. We use the version presented in [1]: starting from the knowledge of one stable nominal linear timeinvariant system. Kharitonov's theorem gives a sufficient result when calculating each coefficient explicitly via interval analysis [17] and treating the interval coefficients thereafter as being algebraically independent (see for instance the DC drive example in section 8). To determine the exponential stability of the resulting characteristic polynomial of the IRn+rsystem with coefficients containing the interval expressions fi. Sometimes the determinant of this matrix may be difficult to be calculated or the coefficients may depend in a noncontinuous way of the uncertain parameters. we are able to determine the stability algebraically by the theorem of Frazer and Duncan [7]. then we propose to apply the following algorithm of Walter and Jaulin [23]. we use one of the following methods: K h a r i t o n o v ' s T h e o r e m If the coefficients of the characteristic polynomial are algebraically independent with respect to the interval parameters (any two coefficients are never functions of the same interval parameter). T h e T h e o r e m o f F r a z e r a n d D u n c a n In the case. If the coefficients are algebraically dependent with respect to the interval parameters.2 Stability of the Tracking Error System being Subject to the Quasiexogenous P e r t u r b a t i o n After having assured the stability of the endogenously perturbed system in the previous subsection 5. The theorem can be found in a nice version in [1]. which is based on interval analysis and projects the coefficients being dependent of interval parameters in the pole domain via set inversion. Their stability property has been intensively studied in the literature (see [1] and [2] for instance) and depends on the magnitude of both the additive and multiplicative perturbation of the desired coefficients AiS. we are going to study the stability of the lin . the theorem consists basically in determining the regularity of the n + rth Hurwitzmatrix for all p in (2).430 Veit Hagenmeyer This resulting equation represents a linear timeinvariant system of dimension IRn+r containing interval parameters.
Then the state equation is uniformly boundedinput boundedoutput stable if and only if it is exponentially stable. that this quasiexogenous perturbation is bounded as a sum of bounded terms. the linear tracking error system being stable with respect to the endogenous part of the perturbation is always stable in the boundedinput boundedoutput sense with respect to the quasiexogenous part of the perturbation. if and only if the autonomous linear tracking error system (18) is exponentially stable. .j. This effect can be taken into account whilst the controller design phase: eventually different nominal values for the uncertainty intervals can be chosen to assure stability without changing the desired coefficients Ai. exponential stability was already established in the previous subsection 5. We remember the following theorem: T h e o r e m 2 ([18]) Suppose a timeinvariant linear state equation is controllable. 6 Main Result In this section. R e m a r k 3 The exponential stability of (18) can for instance be analyzed by applying one of the different methods presented in subsection 5. 7 Remarks on Controller Performance Design and Tracking R e m a r k 4 The choice of the nominal values Po in (2) for the given uncertainty intervals influences the magnitude of both the additive and multiplicative perturbation of the desired coefficients Ai.1. Thus.Robust Tracking of Multivariable Linear Systems 431 ear system (17) including the quasiexogenous signals ~(zd(t).p) (see Remark 2 for the nomenclature). we state the main theorem. since we do not consider in this work neither unbounded desired trajectories nor unbounded desired trajectory derivatives. We see that the quasiexogenous perturbations r do not depend on ~a(t).1. Controllability follows from the basic assumptions. We remark. but on zd(t).j. which solves the problem formulated above and follows from the content of the preceding sections: T h e o r e m 3 Given the linear system (I) under the assumption of controllability for all parameters p in (2). The tracking imposed after system transformation by the presented PIDlike control technique (l f ) is robust with respect to the given uncertainty intervals of p in (2).
of the nominal rotor current I~o. the quasiexogenous perturbation stemming from parametric uncertainty will be of higher magnitude. The rotor tension/Jr and the stator tension Us are the two controls of the system.Uso and to. The linearized and singularly perturbed form of the DC drive reads as follows _~ z+ u (10) Thereby xl = o~ . we apply the theory presented in the previous sections to a separately excited DC drive model. of the nominal flux r o at w0 (which can be determined by minimization of energy losses [9]) and the values of the nominal inputs Uro and Us o are represented in Table 1. we are able to trade off the magnitude of the controller gains and the velocity of the desired trajectories to fulfill the desired control performance and the given physical constraints at the same time (see [5] for a profound discussion of this subject).432 Veit Hagenmeyer R e m a r k 5 If we design the desired trajectories such that Vt > t* : z d = O. R e m a r k 6 Scaling the desired trajectories with respect to time. 8 Example: a Separately Excited DC Drive In this section.1 as Vt > t* : !lr(zd(t). results in higher or lesser magnitudes of "d z .p) in (17) and therefore the resulting tracking error without changing the nominal poles of the closed loop system. x2 = ~ . that is decelerating or accelerating them. The values of the nominal velocity Wo.o. . . in real applications in which saturating control inputs and measurement noise have to be considered.Uro and u2 = Us . If rapid desired trajectories are to be designed. We first linearize its nonlinear model (see for instance [15]) around the nominal values of the drive and neglect thereafter the dynamics of the electrical part (it being much faster than the mechanical part [14]).4~. Ir r represent the angular velocity.p) = O. Hence. we are able to reduce the magnitude of the quasiexogenous perturbation !lti(za(t).wo. in slowing down the desired trajectories. the rotor current and the stator flux respectively. therefore we have to use higher controller gains to keep the tracking error in a reasonable range. ul = Ur . Thus. we are able to stabilize the system around these desired trajectories from t* onwards without persistent tracking error: the equation (17) reduces in this case to the linear autonomous system (18) in subsection 5.
Since the viscous friction coefficient B is not very well known.0084 Nm/rad [0. the resistances can double its value whilst in operation.4 A 12 Wb 26. the viscous friction coefficient by B. the motor constant by c. it can be represented by B = Bo + [~. 0.69 ~2 [0. The transformed system can thus be represented both for the controller design and the robustness analysis in zcoordinates by . where Rro is the value identified in cold state..021 Wb/rad 0..6654 V 32 V The inertia of the drive is denoted by J.7198 rad/s 0. Due to termic effects.4 x 1 0 . and L. which simplifies the ongoing. Therefore we have ~ = z in view of (3).3 kg m 2 0. (11) and (15). represent the rotor and stator inductance respectively.Robust Tracking of Multivariable Linear Systems !Nominal states/inputs 0J o 433 /~o ~so U~o U. The coefficients L. the rotor resistance and the stator resistance are denoted as Rr and R. 0.61 mH 45 H 0.0016 ] Nm/rad Table 2. Parameter Lr L~ Rr o fir R~ J C Bo /~ Value Unit 9.69] 120 ~2 11.. Nominal values of the states and the inputs of the DC motor Value Unit 104.. We choose the transformation matrix S(p) in (3) as which is in this case independent of any parametric uncertainty.0024. = Rro + t~. All parameter values can be found in Table 2. Parameters of the DC motor The derived model is valid for desired trajectories in the neighborhood of the nominal set point.. hence we denote R.o Table 1.
o 'o JR~o 'z1. that is c2~ .l + ~a + A2. At.1s + A~.tdr + Al. J [ 01] u (21) For the controller design we use the nominal model (see for the notation (10)) Zl.1 + JtLo z~.l + " ~ o U l Rso Lso z2..o .O s2 + A2.is + AI.o Vl ~ JR.1dr + A~. v2 = ~ s z~..oUl := vl U2 : = "02 C•$0 (23a) (23b) which yields for u J~o Ul : US : C~s''~'01 "02 (24a) (24b) We design the tracking feedback control as in (14).o Zl.l ~. c24~ 2 + BoRro cIroR..434 Veit Hagenmeyer ~.lel. o + I' el. = JR~ ~..1  .oWo C~..l (25b) where the AIj are chosen such that the corresponding characteristic polynomials s~ + At.1 + u2 (22) T h e new input v can be defined by jt~.1 = z2.le~.ofot e2.c2~. I z+ .l + j:d 1.o~o z~.o + BoR.1 J R..t (25a) R.o (26a) (26b) .o .1  CI~oR...c24'.
Robust Tracking of Multivariable Linear Systems
435
are Hurwitz. See Table 3 for our choice, which places the poles for the two uncoupled nominal closed loop subsystems at a double multiplicity of  1 0 ~ and a double multiplicity of  5 ~ respectively (the eigenvalues of the nominal open loop system are at 8.817 and  2 . 6 7 1 ). Coefficient Value Unit Al,o  100 Jy Al,a  20 1_
A~,o  25
A2,1
Table 3. Desired coeflldents

10
1
7
Using (25) and (22)(remember again z = error system
zde
(16)) , we get for the tracking
Rrb +
+ f ,h, d
( R , . Rrot ~ . ) ) zl ,"do+ 1
A1,0 ~0t el , l d r
Ct, ofr
zd
(1
~2,t = A~,o

~1,1e1,1
(27)
(28)
/o'
e2,1dr

A2,te~,l
To study the stability of this system, we derive one time with respect to time to get
el,1
:
el,2
el,2 = (1
e2,1 : e2,2
J(Rro + t ~ )
,zt,1  el,2)
J(Rro + fir) (~:d2'1  e2,2) (29a)
Rro  ..d (Rr~+ ~ ) )z1'1  A~,oel,~  A~,let,~
k2,~ = A2,oe2,1  A~,le~,2
(29b)
We go on in investigating the stability of the endogenously perturbed system, that is we set ~d = 0, i ~~ 1, 2 and ~d = 0 in (29) to get the autonomous 1,1 1,1 error system
436
Veit Hagenmeyer
0
1
0

0
Circler
~[
AL~ ~
RrB+t~rB~
J(Rc~ 0 0
0
e
Al,1
0
0 A2,o
J(P~o+/~) 1 A2,1
e
(30)
__

~2
0 0
(31)
0 xs x6J The characteristical_polynomial of this system, which contains the interval parameters B and Re, can be calculated as
s 4 + (  ~ 6  ~ ) s 3 + (  ~ 5  ~1 + ~ 6 ) s ~
+ ( ~
+ ~1~6)s + ~ 1 ~
(32)
To answer the question whether this characteristical polynomial is Hurwitz, we apply the algorithm of Walter and Jaulin presented in section 5.1. The answer is positive, the characteristical polynomial is robustly Hurwitz. After having established the stability of (30) with respect to the endogenous perturbation, we assured equally the stability of the quasiexogenously perturbed system (29) following the results of subsection 5.2. 9 Conclusions
In this article we investigate the robustness of tracking controllers working on controllable linear MIMO systems containing parametric uncertainty which is a priori known within given intervals. After having transformed the linear systems into their respective controller form we establish a nominal tracking control law via PIDlike control techniques. We analyze the stability of the tracking error equation and see, that the problem splits into the analysis of a linear autonomous system under endogenous parameter perturbation and the analysis of the same system being affected by so called quasiexogenous bounded signals at the input. For the robustness of the endogenously perturbed system we recall first Kharitonov's theorem, second a theorem by Frazer and Duncan and third a numeric algorithm established by Walter and Jaulin. The first theorem yields only a sufficient result in our general case, the second theorem is necessary and sufficient, but sometimes its analytic calculation may be very difficult, the algorithm finally gives necessary and sufficient results in a numerically guaranteed way.
Robust Tracking of Multivariable Linear Systems
437
After the robustness of the linear autonomous systems being affected by parametric uncertainty has been proven, the robustness of the quasiexogenously perturbed system follows directly from BIBO  stability of linear systems. Thus the main theorem which solves the problem of robust tracking of linear systems under parametric uncertainty is stated. Important remarks are thereafter deduced from the obtained results. They point at the relation between robust stability and controller design and discuss the tracking performance under parametric uncertainty. In the example of DC drive we apply our theoretical results and show the tracking robustness of its linearized and singularly perturbed model. In forthcoming publications we will study extensions of the work presented here to linear systems being perturbed by real exogenous perturbations and to linear systems being controlled by partial state feedback.
Acknowledgments
The author is very grateful to Dr. Emmanuel Delaleau for many intensive discussions and helpful suggestions. He is, moreover, thankful to Richard Marquez for important advises with respect to the presentation of the material, to Dr. Luc Jaulin for having him introduced to interval analysis and to Dr. Michel Kieffer for the software implementing the algorithm of Walter and Jaulin. The author wants to thank Prof. H. Nijmeijer for comments helping to improve the readability of this article.
References
1. Ackermann, J.(1993) Robust Control:Systems with Uncertain Physical Parameters. SpringerVerlag, London 2. Barmish, B. R.(1994) New Tools for Robustness of Linear Systems. Macmillan, New York 3. Bitauld, L., Fliess, M. and L~vine, J. (1997) Flatness based control synthesis of linear systems: an application to windshield wipers. Proceedings of ECC, Brussels 4. Delaieau, E. (1997) Suivi de trajectoires pour les syst~mes lin~aires. Acres Coll. Cetsis  Eea, Orsay, 149154 5. Fliess, M. and Marquez, R. (2000) Continuoustime linear predictive control and flatness: a moduletheoretic setting with examples. Internat. J. Control 73, 606623 6. Francis, B. A. (1977) The linear multivariable regulator problem. SIAM Contr. Opt. 15, 486505 7. Frazer, R. and Duncan, W. (1929) On the criteria for the stability of small motions. In Proceedings of the Royal Society A 124, 642654
438
Veit Hagenmeyer
8. Hagenmeyer, V. (2000) Robust tracking of linear systems under parametric uncertainty. Proceedings of MTNS 2000, Perpignan 9. Hagenmeyer, V., Kohlrausch, P. and Delaleau, E. (2000) Flatness based control of the separately excited DC drive. In "Nonlinear Control in the Year 2000" (this very book), eds. Isidori, A., LamnabhiLagarrigne, F. and Respondek, W., SpringerVerlag, London 10. Hunt, L. R., Meyer, G. and Su, R.(1996) Noncansal inverses for linear systems. IEEE Trans. Automat. Contr. 41, 608611 11. Hunt, L. R., Meyer, G. and Su, R.(1997) Driven dynamics of timevarying linear systems. IEEE Trans. Automat. Contr. 42, 13131317 12. Hopp, T. H., and Schmitendorf, W. E.(1990) Design of a linear controller for robust tracking and model following. Trans. ASME, J. Dynamic Syst., Measurement, and Contr. 112, 5 5 2  5 5 8 13. Kharitonov, V. L. (1978) Asymptotic stability of an equilibrium position of a family of systems of linear differential equations. Differentsial'nye Uravneniya 14, 20862088 14. Kokotovi~, P., Khalil, H. K. and O'Reilly, J. (1986) Singular Perturbation Methods in Control: Analysis and Design. Academic Press, London 15. Leonhard, W. (1996) Control of Electrical Drives. Springer, Berlin 16. Luenberger, D. G.(1967) Canonical Forms for Linear Multivariable Systems. IEEE Trans. Automat. Contr., 290293 17. Moore, R. E.(1979) Methods and Applications of Interval Analysis. SIAM, Philadelphia 18. Rugh, W. J. (1996) Linear System Theory (2nd edition). PrenticeHall, Upper Saddle River 19. Schmitendorf, W. E. and Barmish, B. R. (1986) Robust asymptotic tracking for linear systems with unknown parameters. Automatica 22, 335360 20. Schmitendorf, W. E. and Barmish, B. R. (1987) Guaranteed output stability for systems with constant disturbances. Trans. ASME, J. Dynamic Syst., Measurement, and Contr. 109, 186189 21. Schmitendorf, W. E. (1987) Methods for obtaining robust tracking control laws. Automatica 23, 675677 22. Sontag, E. D. (1998) Mathematical Control Theory (2nd edition). SpringerVerlag, New York 23. Walter, E. and Jaulin, L. (1994) Guaranteed Characterization of Stability Domains Via Set Inversion. IEEE Trans. Automat. Contr. 39, 886889
F l a t n e s s  b a s e d Control o f the Separately Excited D C Drive*
Veit Hagenmeyer, Philipp Kohlrausch**, and Emmanuel Delaleau
Laboratoire des Signaux et Syst~mes C.N.R.S.Sul~lecUniversit d Parissud Plateau de Moulon, 3 rue JoliotCurie 91 192 GifsurYvette cedex, France {hagenmey, delaleau}@Iss, supelec, fr
Due to the flatness of the separately excited DC drive a novel control scheme that achieves copper loss minimization can be designed. It makes use of an online replanification of desired trajectories by using the information of a fast converging load torque observer. Simulation results show the performance of the proposed control scheme.
Abstract.
1
Introduction
The work presented here focuses on the separately excitated DC drive and its differential flatness property. As the separately excitated DC drive configuration has two inputs (the stator and rotor voltage respectively) available, it offers a second degree of freedom with respect to the serially and parallely excited ones. Nowadays power electronics being far less expensive than some years ago, we study the advantages of this second degree of freedom combined with a nonlinear, flatness based control strategy. Thus we are able to use the nonlinear model without any simplifications and to design suitable desired trajectories for the rotor speed and the stator flux for first accelerating, second maintaining a constant angular velocity and third braking the system under the aspect of energy loss minimization with respect to copper losses due to the Joule's effect. Up to now energy loss minimization of the separately excited DC drive has only used linearized variants of the model and studied either the acceleration motion or the constant angular velocity motion. For the first time, we give a control scheme in which first the full nonlinearity of the model is respected and second in which the different motions of acceleration, constant angular speed and electrical braking are all treated by the same method. Furthermore, * The work of V.H. was financially supported by the Nonlinear Control Network (NCN) and by the German Academic Exchange Service (DAAD). The work of P.K. was financially supported by the Nonlinear Control Network (NCN). ~0"Student at the Institut fiir Regelungs und Steuerungstheorie Technische Universitgt Dresden, Germany.
440
Veit Hagenmeyer et al.
we can start the drive from rest without any discontinuous jerk and without premagnetizing the machine. Finally we are able to reinject the kinematic energy won whilst braking the drive electrically in the power supply system. The paper is organized as follows: After having briefly exposed the mathematical model of the motor in Sac. 2, we establish its differential flatness property in Sac. 3. We present the flatness based control of the DC drive which achieves the minimization of the copper losses in Sac. 4. We give some simulation results in Sac. 5 and conclude the article with a final discussion in Sec. 6.
2
Model of the Motor
The physical properties which underly the dynamic behavior of DC motors have been studied thoroughly and can be found in [11] for example. Fig. 1 depicts the equivalent circuit of this machine. l, R.
Lr
o, u, o,
Fig. 1. Equivalent circuit of a DC machine The following system of differential and algebraic equations reigns their dynamical domain:
dO

dt
=
~
(la)
je, o
dt = rm  Bw  n d4~ , + dT
(1 b) (lc)
U, = n,I,
g~ =
R~I~ + L~~t +e
(ld)
(le) (if) (lg)
r,. = cr
~, f(/8) e = cr
Flatnessbased Control of the Separately Excited DC Drive
441
where 0 and w respectively denote the angular position and angular velocity of the shaft; vm is the electromagnetic torque produced by the m o t o r and v/ the load torque. The rotor and stator voltages are expressed by Ur and Us, the corresponding resistances by Rr, Rs ; the rotor inductance is denoted by Lr; 4~s is the stator flux, In and Is are the currents of the rotor circuit and the stator circuit respectively; e represents the backelectromagnetic force induced in the a r m a t u r e winding. The parameter J expresses the m o m e n t of inertia of the rotor, B the viscous friction coefficient, c represents a constant dependent of the spatial architecture of the drive. The relation between the flux ~s and the stator current Is is nonlinear as there is a saturation effect: ~s = f ( I s ) . T h e function f ( I s ) is shown in Fig. 2. Remark that this function is strictly increasing, thus bijective with an inverse f  1 itself strictly monotonic increasing. For small stator currents, the flux is itself small and it is possible to use the c o m m o n approximation r ~ L , I s with L, as the constant autoinductance coefficient. We do not take into account the hysteresis effects which are quite negligible.
/
d~ dt
l.
Fig. 2. Typical magnetization curve of a DC motor
Using ( l b  l g ) , the state variable model for the separately excited DC m o t o r for speed control 1 is:
dt = U,  R r I r  c ~ , w d4', = U,  R , f  l ( + , ) dt
(2b) (2c)
1 The ongoing studies are quite similar in the case of an angular position control. They are not carried out for the sake of simplicity.
442 3
Veit Hagenmeyer et hi. Differential Flatness
Recall that a nonlinear system ~ = f ( x , u), u(t) E ~rn, is said to be (differentially) fiat [7,8] if there exists y = ( Y l , . . . , Ym) such that:
1. y = h ( x , u , . . . ,u(a)); 2. x = r and u = r 3. the components of y are differentially independent.
The separately excited DC m o t o r (2) is fiat [7] with the fiat o u t p u t 2 y = (w, ~a). To prove this fact one simply assumes the quite reasonable hypothesis that ~ is an (unknown) function of time which can possibly depend on the mechanical variables (w or 8) but not on any electrical variable. Thus, every variable of (2) can be expressed as a function of the flat o u t p u t and a finite number of its derivatives:
Is = f  l ( ~ , ) J~b + Bw + 71
Ir g~ = L~ c~, (JC~ + B ~ + . ) (J~o + B~o + n )
(3a)
(3b)
U, = ~, + R , f  l ( ~ , )
Remark
(3d)
1 Notice, that the flat output y = (w,4i,) used in our approach is not unique. Other physical meaningful possibilities are for instance (~, It) and (w, e). We choose (w,~,) in view of 4~, appearing in both nonlinear products in (2a} and (2b} and therefore being the physical link between the electrical and the mechanical part of the system.
Remark
2 Remark also that (3c) is singular at 4~, = O. This stems from the intrinsic physical singularity in which the motor cannot operate. See ~.~ for the avoidance of this singularity.
4
Minimization
of the
Copper
Losses
We go on in studying the design of an optimal performance trajectory for both components of the fiat o u t p u t of the separately excited DC m o t o r as the second degree of freedom at the input is nowadays usable power electronics being far less expensive than some years ago. We will denote the reference It would be y = (8, ~,) in case of a position control.
Flatnessbased Control of the Separately Excited DC Drive
443
trajectories of the components of the flat output w and ~)~ as wd and 4i, d respectively. Both wd and ~sd are functions of time. Energy loss minimization in DC and AC drives has been intensively studied in a founding article by Kusko and Galler [10]. They generally showed that DC motor losses can be reduced considerably by independently controlling the stator and field current. Several publications following this article can be found, which all use linear models of the respective DC drive for energy loss minimization. They can be divided into two different groups: one studies the acceleration motion via optimal control techniques [5], the other establishes control methods for minimizing the losses whilst maintaining a constant angular velocity applying optimal regulator theory (see [2], [3], [4]). A profound survey can be found in [6]. For the first time we propose an energy loss minimization control scheme in which all nonlinearities of the DC drive model are taken into account. The most important nonlinearity for this task seems us to be the saturation curve for the relation between stator flux and stator current (see Fig. 2). Furthermore we present for the first time an energy loss minimization control method which treats the different operating motions (as acceleration, constant angular velocity and electrical braking) at the same time. This is only possible in view of the flatness property.
4.1 Nominal Control
From the flatness property, (3c) and (3d) one obtains the nominal (openloop) control:
+ +
+ CCddq~)sd"~
Rr~~A(J&4
+ Bwd T ~)
(4a)
(4b)
U,a = ~ , d + R , f  l ( ~ , d )
where ~d and ~sd are the references trajectories of w and 45, that will be design below (w 4.3) in order to avoid singularities, i.e. that Urd and Vsd remain everywhere defined and bounded. The variable ~t denotes an estimation of the value of the load torque. We will present in the next section the construction of a simple observer for 7] under the assumption that ~ = 0. Consequently, we find (4a) and (45) from (3c) and (3d) by simply replacing the components of the flat output by their respective reference values and by setting 7] = ~ and ~ = 0. R e m a r k 3 Notice that the injection of the estimation in (4a) facilitates the work of the closed loop controller. feedback has only to cope with the unknown transients its instantaneous mean value ~ is estimated online by of the disturbance With this choice, the of the disturbance as an observer.
444 4.2
Veit Hagenmeyer et al. Observer for the Load Torque
In the preceding section we supposed throughout the calculations the load torque perturbation to be known. Hence we go on in constructing a rapidly converging observer for the load torque r~ using the measured state as it is done in [1]: dd~ d~ B
dr j~~ll(~~)+co,/vj
= 12(w  do)
(5a)
(5b)
Note that the product ~ , I v in (5a) is available by measuring both currents Iv and I,. Thereafter, ~s is determined using the magnetization curve (see Fig. 2) and the measurement of Is. Subtracting (5a) and (55) from (2a) and /t  0 results in the linear error dynamics d
(::)
ande~=rZ~.
(::)
.0.
withe1 = w  r
The observer gains 11 and 12 are chosen such that the linear error dynamics converge faster to zero than the dynamics of the observed drive. Using the knowledge of ~, we feedback online this information to the desired trajectory generator, and therefore to the nominal open loop control. 4.3 Reference Trajectory Generation
In view of the nominal control using the flatness approach (~ 4.1), we have to design the reference trajectories t ~> Wd and t ~+ ~sd for both components of the flat output. For the angular speed, this choice is quite obvious since we are designing a speed control. We can for example divide the working time interval [tini, tfin] in nonempty subintervals [ti, ti+l], i = 0 , . . . , N  1, with to =/ini and IN = tfin. At each point of linking one can impose a given speed 0~d(ti) = r In each subinterval, the function Wd can simply be chosen as a polynomial of time: OJd(t) • Y~j01i Wij (t . ti)i/(i!), .i 0, . . . , N  1, where wij e ]~. In order to obtain a continuous nominal stator voltage (4b), one has to choose in each interval a polynomial of degree at least 5 satisfying the 6 following connection conditions3: W~dk)(tf_l)= W~d~)(t+),j = i, i + 1, k = 0, 1, 2. These conditions 3 We use the standard notations: f(xo+) = lim
X>Xo
f(x) a n d f ( x o ) =
lim
X~Xo
f(x).
Flatnessbased Control of the Separately Excited DC Drive
445
lead to linear systems of equations which permit to obtain the coefficients wij of the polynomials defining the reference speed trajectory. This choice of the speed trajectory implies that the jerk db is everywhere continuous. For the trajectory of the stator flux, ~sd the choice is not directly related to the speed control. We will show that it is possible to choose its value as a function of the reference speed Wd and the estimated load torque 7~ in order to achieve copper loss minimization. We denote by P j = R,.I~ + RsI~ the total instantaneous power dissipated by the Joule's effect that are the copper losses in the rotor and stator circuit of the motor. As P j is a function of the variables of the system (2) which is a flat system, it is possible to express P j in terms of the flat output, using the expressions of the currents Ir and Is (from (ab) and (3a)). So, P j = Rr \ c~. + R,. ( f  1 (q~s)) 2. Remembering that 7",, = J& + Bw + rl, we finally obtain:
Pj =
()'
Tm
+ Rs
(7)
R e m a r k 4 Remembering that the torque produced by the motor is rm = cr one sees that a given torque can be produced by various choices of ~s and I,. All these different values of ~s and Is lead to different values of the copper losses P j . We will show below that one choice of ~s and Is results in a minimum of P j . From (2a) one deduces that the mechanical torque produced by the motor
t ~ " "t'm, = J w d + Ba.'d + 7t, is fixed for a given desired speed trajectory
t ~+ wa under a given load torque behavior t ~+ n . . consequently, the power dissipated by the Joule's effect (7) can be thought as a function of the stator flux only. If rm ~ O, Pa is the sum of two squares, the first of which is a strictly decreasing continuous function of the absolute value of the flux 14~,[ whilst the second one is a strictly increasing continuous function of ]~s ] (cf. Fig. 2). As a consequence, for a given rm., P j admits a unique minimum for ~s > 0 and a unique minimum for 4~s < 0 which is the solution of the following equation:
1 dPj
2 d 7 
rm~.
+ Rsf
1
dfl~s
 0 (8)
If Vm = 0, P j is a strictly increasing of 1~81 and P j is minimum for 4~s = 0. However, if one has to generate a speed trajectory which leads to rm(t,) = 0 in only one point of time t,, as it is the case for electromechanical braking, it is not a good idea to set at this point ~sd(t,) = 0 in view of the time demanding operation of demagnetizing and remagnetizing the drive. In order to change the sign of the motor torque rm it is better to change the sign of
. we always choose the positive solution of (8). The coefficients a and /3 have been well fitted from numerical solutions of (8) inside the range of operation of the motor 9 Therefore it is possible to obtain ~sd by the analytic expression ~.d: <d=VE : J Rr L. 1) from considerations of initial conditions for COd being equal to zero.=oaj' j '3:2~/~=3ajtj3 whereas the coefficients aj = 0. = f ( I s ) ~ L. Our approach is. In this case the expression (9b) must not become singular. i = 0. V rm. 2 (and therefore bi = 0. both the motor torque v.4 Study of Singularities In order to study the singularities at small flux it is convenient to linearize the flux characteristics 4 . (t) > O. (9a) (9b) 1 . ~ f . Using a polynomial of order 5 to design Wd as described above: 9 _ _ t (~=2bit ~/~.2 r. more general as we also consider transients by calculating 4i.~ and its first time derivative are zero. hence we show that limt.I~.0 and we have to study the corresponding singularities. Nevertheless.d = h(rm.(t) "" htt/2 t+0+) O.~ . one can think of differentiating (8) with respect to time and then solving numerically the obtained expression9 However. 4.~. In the neighborhood of t + 0 +.446 Veit Hagenmeyer et al. As we also need the first derivatives of the flux in (4b).)r~n. the rotor current only. In this case it is even possible to give analytical expressions for ~. we have preferred to find a good approximation of the solution of (8) of the form 4~. This is the option that we selected...av/tanh/Srm. j = 0. As a consequence. ~ Therefore there does not occur any singularity for the design of the desired trajectories when accelerating the DC drive from the resting position.it)] ) ~ = O. . (8) is not defined for 4 .d and by using the flatness property.2Vm. however. ~~'.) ~. When starting the system from rest. .d = h'(rm.d and ~. 1. consequently. Solving numerically (8) leads to the reference stator flux ~od at any point of time.~ Notice that (9a) is similar to the one given in [10]..o+ ( \qT. this leads to tedious calculations..
Flatnessbased Control of the Separately Excited DC Drive 447 As for the braking of the drive to stillstand.~2~(~o . J = 11. We first give the parameters of this DC drive.5 Tracking Control Using the flatness property of the separately excited DC motor. ) + vl to get the following linear system to be controlled: = vl & = v~ (10a) (10b) (lla) (llb) Using (1 l a) and the designed desired trajectories for Wd and ~ .*oa) + A2a Z' (12a) (~0  ~d)dr (12b) whereas . d ) d r v2 = CSd + . Introducing the new input v. = 45H. / L = 0.021s.. 4. the same analysis with respect to the singularity can be led.~ ' d ) "~.B ~ .l ( q ~ . The saturation curve has been approximated by f ( I .Io t a n h ( I . second show the desired trajectories which were discussed in w 4.A12 (~:~s .I. we are able to exactly linearize its model (2) by endogenous feedback [7]. The complete control scheme is depicted in Fig.61mH. f . / I o ) with L. ) = L. . d .69 ~2. we develop the controller for the new input v as follows: Yl : ~ ' d "Jr"A I I ( ~ ' . 5 Simulation Results In this section we present the results of the preceding sections for a specific separately excited DC motor. The parameters used for the simulations correspond to those identified on a real DC motor: Lr = 9. Rs = 120 . Io = 1 9 W b . The integral term is set in order to achieve a good tracking under disturbance and parameter uncertainties. Z r . ~ B u.~ ) ) + n~I~ + c ~ .2 and conclude finally with the simulation results of the controller developed in w 4.5. Notice that in the present case the linearizing feedback is a static one as the model is static state linearizable [9].d~d) + .~ij t o be chosen appropriately.10 . = R .~ b . we establish the following: u~ = ~Lr (Jr2 . L.3.3 k g m ~. B = 0 .4. third present the results of the observer and its effect on the online trajectory generation of w 4. 0 0 8 4 N m / r a d .city1 + 7 ( ~ 4 .~t (d~ . 3. c = 0.(2.
The estimated load torque ~ is used in a novel way both in the desired trajectory generator and in the flatness based controller (see Fig. I Trajectory generator r Load torque Harness based controller Ur.448 Veit Hagenmeyer et al.5s]. The chosen observer coefficients are l z = .1 respectively. 3).2. During the braking interval there exists a point for which vm = 0. 5. we apply a constant load torque at t = 2. {O. We see in Fig. 5). T h e flatness based controller was established in section 4. we lead at the midpoint of the braking interval the flux to its necessary m a x i m u m (see again (2a)) at the point where r. To show the effect of braking the system under the condition of Vm(t) > 0 for t E [4s. A22 = 300s ~ and A23 : 1000s 3. we see that the desired flux is reduced for the braking maneuver. which places the poles of (12a) at . its coefficients are Azz = 10s z. Us DC motor It. We see.3 1 2 5 N m / r a d .10 s .z .5s (see Fig. In this case we apply the optimization m e t h o d already used for the desired acceleration.4. that the field weakening normally used in a serially excited D C drive by actuating a switch is automatically imposed on the separately excitated configuration following our optimal desired trajectory design using the differential flatness property: when accelerating the drive the flux is led to its m a x i m u m (see (2a)) where the desired m o t o r torque rm. The resulting desired trajectories are presented in Fig. observer Fig. A21 = 30s . we depict in Fig. 3. To show the performance of the developed flatness based control technique minimizing copper losses. Therefore. 6 . The latter is estimated by the observer (5a) established in section 4. that the controller corrects quickly the errors s t e m m i n g from the unmodelled dynamics of ~. A12 = 20s ~.5 0 0 s . is extreme. a flatness based controller and a load torque observer J . which shows an excellent convergence behavior.5. For maintaining a certain angular velocity the flux is thereafter reduced to the necessary level.2 s z and . which is due to the braking effect of the load torque. Novel scheme linking a trajectory generator.n.5 s z and all three poles of (12b) at . which places both observer poles at .z .e It. co. is also maximal. l~ First we want to accelerate the DC drive without load torque (r~ = 0) from rest to a velocity of 1 0 0 r a d / s in one second and brake it to stillstand in the same time.1 0 0 0 s 1 and 12 = .
.. . / ~ .... .. : : . ... i 1 i. 2 . i 3 i .... .. . . 4 0 1 2 3 4 5 I/\ 0 30 I .. i :. 3 4 ! 5 0 1 2 3 4 5 20 1o . i .. ..... ~. . Having two inputs available. i . .....Flatnessbased Control of the Separately Excited DC Drive 449 the response of the closed loop system to severM load torque perturbations. Hereby we solve a longstanding problem for this subject in using the full nonlinear model of the drive and presenting a control strategy which is applicable to all different motions (accelerating. .. Our m e t h o d yields moreover the advantage of starting the drive from rest without any discontinuous jerk and without the necessity of premagnetizing the machine and the possibility to reinject the kinematic energy won whilst braking the drive electrically into the power supply system... i ... . ...... i .... that the novel control scheme finds for every point of operation consisting of the pair of desired velocity and given load torque an optimal balance between the rotor current and stator flux. 1 2 3 time [s] 4 5 0 1 2 3 time [s] 4 0 Fig. . ........ .. ::.. we propose a nonlinear flatness based control scheme under the aspect of energy loss minimization with respect to copper losses. .. .. maintaining a constant angular speed and electrical braking) of the drive. ~ "" o 5 i i i : .. All states and the inputs stay within their given physical bounds... . 4... . .... i 2 ? . . .... g' l " " : : 9 ... .............. i... ~ 5 1 9 ' . ! ....... .... We remark... i .. :. Desired trajectories 6 Conclusions In this article the relation of the separately excitated DC drive and its differentiM flatness property is investigated.. ~ s ~ o 0 ....
.... ~ ~\ .. i o i i 0 1 2 3 time [s] 4 5 0 1 2 3 time [s] 4 5 F i g ...... i ... ... ...... i/ 5 0 1 2 3 4 5 / ~o 0 . i :t 3 : 4 i 5 0 1 2 3 4 5 150 0 1 .. ! ... i. 2 ... ... i .. B r a k i n g t h e s y s t e m u n d e r load t o r q u e ~'~~ " : : : I g'[ 7 3 .... !....... ....... : ... ............. L o a d t o r q u e p e r t u r b a t i o n a n d b r a k i n g electrically .... ~i ~ o 0 i i i i ...t i. ... .. ....... . .. " ? ? '~I 0 ~" 2 0 ... .. ............ .450 Veit H a g e n m e y e r et al.... i .. 2 : ~ _ .. ........... i ...... ....... 3 4 5 : o r .......~ 4 ~_so/i 20 i .. . .. . lo ..... . 0 1 2 3 4 5 ..... i . i .... .. i . i\~] Vl 1 2 3 4 5 200 3o .... ... 1 ! 2 i 3 i 4 i 5 " 20 0 1 ....... 6.o .. i . . i. .. ~.. ....... i...... :: i... ........ ... ::. i. 1 .... .~ ~ 201 .. 2 ! 3 .... l "~176 /. i. 5...... i ~ 1 o 0 1 2 3 time [s] 4 5 0 1 2 3 time [s] 4 5 F i g ...
. References 1. Tsuchiya T.. Cooley W. Famouri P. Wang J. L6vine J.. Electric Machines and Power Systems 25:525537 7. Egami T. Acknowledgement The authors want to thank Dr. An implementation of the described control strategy to a real separately excited DC drive is currently undertaken and will be presented in a forthcoming publication. Tsuchiya T.. Bull Acad Pol Sci Set Sci Math 28:517522 10.. Egami T... IEEE Trans Automat Control 44:922937 9. Rouchon P. Wang J. Martin P. Jakubczyk B. Famouri P. Respondek W. traction drives like in subway trains for which the DC machine is still widely used. for instance. L6vine J.. (1997) Loss minimization control of a DC motor drive. Rouchon P.. Chiasson J. Applications of the work specified here concerns. (1994) Design of DC traction motor drives for high efficiency under accelerating conditions. (1983) Control means for minimization of losses in AC and DC motor drives.. Galler D. (1994) Nonlinear differentialgeometric techniques for control of a series DC motor. Kusko A. IEEE Trans Ind Appl 30:11341138 6. (1985) Efficiencyoptimized speed control system synthesis method based on improved optimal regulator theory .. IEEE Trans Ind Appl 19:561570 . (1995) Flatness and defect of nonlinear systems: introductory theory and examples.. Fliess M. Tsuchiya T.Flatnessbased Control of the Separately Excited DC Drive 451 After having designed the desired trajectories for copper loss minimization. Thereby the estimate of the load torque is used in a novel way for both an online redesign of the desired trajectories and for the enhancement of the flatness based controller.. lEEE Trans Ind Electron 32:372380 5. lEEE Trans Ind Electron 33:114125 3. second present the effect of the online redesign of the desired trajectories under load torque perturbations and finally depict the performance of the nonlinear flatness based controller. Fliess M. (1980) On linearization of control systems.. Egami T. Rudolph for helpful comments with respect to the organization of the paper. Martin P. IEEE Trans Control Systems Technology 2:3542 2. IEEE Trans Ind Electron 34:216226 4. (1986) Efficiencyoptimized speedcontrol system based on improved optimal regulator theory. J. (1999) A LieB~icldund approach to equivalence and flatness of nonlinear systems. We apply the proposed control method to a specific example: we first show results of the desired trajectory generation.. Internat J Control 61:13271361 8. (1987) Efficiencyoptimized speed control system with feedforward compensation. we establish an observer for the load torque and present thereafter a nonlinear flatness based controller.application to separately exited DC motor system.
Leonhard W. 11. Springer. Berlin . VEB Verlag Technik. 2nd edn. (1996) Control of Electrical Drives. Braunschweig 12.452 Veit Hagenmeyer et al. M/iller G. Elektrische Maschinen: Theorie rotierender elektrischer Maschinen.
based on the stabilizability of the nominal system.Hammami@fss. a continuous feedback control is proposed to exponentially stabilizes nonlinear dynamical systems using the Lyapunov approach. Furthemore. where necessary and sufficient conditions for the existence of . u is the control input and y is the output of the system. u. t n A b s t r a c t . Latter. . As a necessary condition for the existence of an exponential observer. [9]). x(t) is the state. [7]. we show that the system is detectable and stabilizable by a continuous controller. u) y = h(x) (1) where t is the time. Under some conditions on the nonlinear part.) (2) which is expected to produce the estimation ~(t) of the state x(t) of the system. In this paper. A nonlinear observer can be designed provided that the system is detectable. the linearized counterpart of the nonlinear system must be detectable. [5]) and observer canonical form method ([6]. that guewantees the observation error is globally exponentially stable. A variety of methods has been developed for constructing nonlinear observers.rnu. This observer design incorporates only the bound of the nonlinearities (uncertainties). and does not require exact knowledge concerning the structure of the plant nonlinearities.S t a t e D e t e c t i o n a n d S t a b i l i t y for U n c e r t a i n Dynamical Systems Mohamed Ali H a m m a m i Faculty of Sciences of Sfax Department of Mathematics Route Soukra BP 802 Sfax 3018 Tunisia Nohamed. [8]. we study the problems of observer configuration and feedback stability for a certain class of uncertain dynamical systems whose nominal part is linear. T h e observation of nonlinear systems has first been addressed in the literature by T h a u [1]. 1 Introduction Consider the nonlinear dynamical system { i~ = F(x. Kou et al [2] and Xia and Gao [3] have introduced and studied the notion of exponential observer. A state observer for the system (1) is a d y n a m i c system = c(~. The Lyapunovlike m e t h o d ([4].
. .A.A t ) . The system (2) is called a (local) exponential observer for the system (1). when the uncertainties are characterized by known bounds. if for all input signal u.~(t)ll < e e x p ( . 2 Problem formulation In this section. we recall some definitions and results concerning the observer design (state detection) and stability for nonlinear dynamical systems.) is the function describing the plant and k(. If in addition.x > O. the construction of the observer still remains a difficult task due to the need to solve a set of simultaneous partial differential equations to obtain the actual transformation function. 3J > 0 such that I1~(o) . Ilk(t) . u) = 0.. [131) and for uncertain systems by ([14].. on the state space and for all input signal u.454 M. u) where F(.~(o)11 < implies Ilk(t) . if these conditions are satisfied.. . [11]. [17]. v. for t > 0 and Vs > O.) + ~(~.. the function G should have the following form G(~. . Therefore. Then. the literature . k(x. Vt > o. Besides. [15]). i:(O) = x(O) implies i:(t) = x(t). For the stability problem. [18]). The above conditions are known as the two basic requirements of state observers. stabilization of dynamical systems with uncertainties (nonlinearities) has been widely studied ([16]. this problem is also solved by ([10]. we want to supply an upper bound of the nonlinearity which guarantees the stability of the reconstruction of the system state. then (2) is called a global exponential observer.. using Lyapunov techniques. . [12].x(t)ll tends to zero as t tends to infinity for any initial condition II&(0) ~(0)11. . It should be remarked that. v. A common approach is to treat the uncertainties in the system deterministically. This means that the observer and the plant have the same dynamics under the condition that the output function 9 = h(~3) copies the output function h(x). the equality. Even.Hammami coordinate transformation have been established.) is a nonlinear function that satisfies.) = F(~. h(x). one can use the Lyapunov direct method to design a state feedback control. For our case.
( A . u ) + k(x + e. In particular. that is. u E ~q. A linear system is observable if its observability m a t r i x has full rank. such that the following properties hold : .0).L C ) e + o ( e ~). In this paper.. the origin e = 0 is an equilibrium point of these dynamics.0). and this system is required to be (locally) exponentially stable for every input signal u.. B = ~OF(0. T h e system (3) is said detectable if there exists a m a t r i x K such t h a t the m a t r i x (A .. y E NP.. rank(C. . (n x q). D e f i n i t i o n 1 T h e system (4) is said to be weakly detectable at the origin if one can find a function g : ~ + x ~pn x ~P x ~q + ~'~ and a function 4~ : ~ + x ~n x ~n __+ ~ + .) defining the observer dynamic. CA.u) = Ax + Bu + f(t. and C = ~~h(0). Let the state observation error e be defined by e = & .y. This yields the following property : there exists a local exponential observer of the form (2) for (1). for u = 0. A sufficient condition for (3) to be detectable is that it is observable. A. (p x n) known constant matrices and f is Lipschitz continuous represents the uncertainties in the plant. . if and only if the linear approximation at the origin (3) is detectable.F ( x ..u) Note that..u) h(x) = Cx (4) where x E ~ n . B and C are respectively (n x n). i. These dynamics are exponentially stable if and only if the first order m a t r i x of the above equation is Hurwitz. CA nl) = n.K C ) is exponentially stable. Furthemore. Vidyasagar introduced the local asymptotic observer called "weak detector" specifying properties of the function g(. we consider nonlinear dynamical systems described by the following state equations : ( ~ = F(t.x. the origin x = 0 is an equilibrium point of the system to be observed and the error equation becomes : ~ :. T h e d y n a m i c s of e are given by the following error system : = F(x + e. observers whose state converges exponentially to the state of the plant. u) . [20]). the approximating system of (1) at the equilibrium point is given by (~=Ax+Bu (3) where A = ~F(0. In ([19]..e.x.x..State Detection and Stability 455 on nonlinear observers is nearly exclusively devoted to exponential observers.
r and . the following" weak detector" can be set up. the solution trajectories ~(t) of (5) and the input u are bounded by a sufficiently small e > 0.) E C ~176 g is smooth) and g(t.. by.e.) defined as. .L c ) T p4. this system is weakly detectable.. in definition 1 the "weak detectability" does not require the linear approximation of the system (1) to be detectable. x. u)+~x (t.. the function g(. w. y. I1~11 < ~. x. Q > O. C) is detectable. Ca : ~ + + ~ + continuous and strictly = r . u.) and r r .L(C~ .) satisfy the conditions of definition 1. Indeed. shows that for weakly detectable systems (4). the requirements of weak detectability are satisfied.11 < ~.. I1. One can check that. ~) < r . . ~)F(t. o~ 00 (94. there exists L such that (A .x) r  ~11) = (& .xll) = (~  Defining the functions 4i(. where we suppose that the pair (A. consider a detectable linear system (3). x. ~)F(t.x) = r . When the solution trajectories x(t) of (4).Hammami (i) g(.A.LC) is Hurwitz. i. (A. Ilxll < ~. x. Here. Since. ^ 9(~. According to this fact. (ii) There exist three functions Ct. we will examine the detectability of a class of dynamical system in the global sense. (i.x ) T Q ( ~ . (A . h(x). u) < r As an example. O. y.. It should be remarked that. such that ~. O. u) = A z + B u . C) is detectable.. .0 and 3r > 0. = g(~. x. . h(x).e.xll) ~(t. r increasing such that r = r vt > 0... Vidyasagar.y) it defines a Luenberger observer..xll) = r x)rP(i . .. u)+~xg(t. C) of the nominal system is detectable.P ( A  LC) = . it is proved that (&(t) .456 M.. x.x(t)) tends to zero when t tends to infinity. O) = O.Q .~). there exists a matrix K and symmetric and positive definite matrices P and Q such that. u) = g(~. u) (5) where g(. where the pair (A.xll) < ~(t..
By convection. V(t. let V(t. is to find a such feedback controller. z ) < a~ll~(t)ll 2. the state z(t) of (6} is globally exponentially stable [22]. L e m m a 1 Given a continuous system z(t) = F(t. ~ and/3 are positive scalar constants. ~1(~ 8) )(e0' . In order to study the detectability and stability of the system (4). In fact. in the sense that IIx(t)ll < ((~211z(0)112+ .x) G ~ x ~ " f~(t.et)eX')89 zl A1 i f ~ = A. u). ~(t)) (6) where z(t) is an n • 1 vector. the state feedback controller can be represented by a nonlinear function u = u(t. z). we shall say that the uncertain system (4) is feedback stable (in the global exponential stability sense). if there exists a continuous controller u = u(t) such that the system (4) with u(t) is globally exponentially stable.e~'))89 i/ ~ # X . Our goal. we will need the following lemma [21] which gives an exponential rate of convergence. ~) ~ ~ • ~". x. (7) (8) where A1. If the Lyapunov function satisfies (7} and (8). that can guarantee the exponential stability of nonlinear dynamic system in the presence of uncertainties f(t. A2. V(t. z) be the associated Lyapunov function with the following properties : Axllx(t)ll ~ _< V ( t . IIx(t)ll _ ((~llx(0)ll %x' + where ~ = ~ .State Detection and Stability 457 For the stability purpose. stability of uncertain system is then studied with respect to the equilibrium point of the nominal system (see [18]). an uncertain system has in general no equilibrium point. Provided that all states are available. ~) _< mall~(t)ll 2 + ~e~'. ha.
~11) . With the model given in (4). It turns out that. i. in the presence of uncertainties. h(x).u)F(t./3i is continuous. &) E ~n x ~ n . In order to obtain an . 0 .&. ~) _/~2(11~. e=~x. If/~i E )Coo. Vu ~ ~q /~1(11~ . u) = G(t.h(x). with input y(t) such that the estimates denoted by ~(t) converge to x(t) a s y m p t o t i c a l l y fast. u ) + ~xG(t. y E ~v where p < n and is a continuous function. D e f i n i t i o n 2 Consider the system (1) with the o u t p u t y = h(x).A.x. a continuously differentiable function V : ~ + x ~n x ~'~ + ~ + and real functions f~i : ~ + + ~ + . ~. we obtain an a s y m p t o t i c observer. u). 0 . h(x). with fll (0) = 0 and/~i (s) ~ + o o as s + +oo). the origin of the error equation =G(t. ~.x. where we treat here the global case. so that the conception of a continuous observer is possible. 3} of class )Coo (namely. u E ~q and (x.458 M./ ~ s ( l l ~ . For our case. ~.x l l ) (10) for every t _> 0. fli(0) = 0. such that F(t. u) is an observer for (4).u)+ ~ F ( t .e. W ~ ~n. We say t h a t (4) is detectable.u) og <_. x . we give a definition of detectability. if (4) is detectable. as in definition 1.u). Vt _> 0.Hammami 3 Detectability Tsinias in [23]. {i = 1.e+x. if there exists a continuous m a p p i n g G:~+ x~nx~Vx~q+~n with G ( t . the problem is to show the detectability of the system. then the system = G(t. 0 ) = 0 . 2.~ F't.~11) and (9) OV ( .h(x).V(t. strictly increasing. generalize the well known theorem of Vidyasagar [19] on the local stabilizability of nonlinear systems using state detection. is globally asymptotically (exponentially) stable. ~.
y. where P = p T > 0 is such that. u).II denotes the Euclidean norm on ~n. /3 > 0 (12) 5:) = (}  z)TP(~ . such t h a t f(t.Q < 0 (11) (. we shall assume the following assumptions. ~. u)l I < p(t.u) = and V(x.State Detection and Stability 459 exponential one. the system (~) is detectable with G(t.y) where ~(t. x. Therefore. u) ([14]. (. u) . in the section one. [15]) where r . y. Next.1 c T r z. we shall find an exponential estimation (i.y)llp(t. where AL = A . [15]) such that.z).LC.~eZ* with a > O.42) There exists a symetric positive definite m a t r i x Q(n x n) and a function ~b (see [14].~. x. where II.C) is detectable. a Lyapunov function for the nominal part can be chosen as x T P x .43) hold. one can state the following theorem. u) = P . y. as the argument used in proving the detectability of linear systems. (. Ai > 0. P1cT(c~ Y)P(t'u)2 IIC~ . Ilk(t) z(t)ll _< A l e x p . remark that V satisfies the condition (9).42) and (.e.L ( C } . then there exists a matrix L such that ReA(AL) < 0.A2t. Then. lie(t. P r o o f First.41) T h e pair (A. we will use the function V to show that (10) is satisfied for the system (4). (. u) = A } + Bu + ~(t. Then.43) There exits a positive scalar function p(t. u). ~. Suppose that (`41). A2 > 0).) : ~ + x ~n x ~q ~ ~P. PAL + AT p = . u) + . T h e o r e m 1. Let. where P is the unique positive definite solution to the Lyapunov equation (11). .
~) = ..v)rQ(i . .~.x) + 2(~ . ~) <_ (~ . ~.460 Then.x) +2(~ x)TP(~p(t.u) 2 P H~~ ~ _ ~i ~~ae_~ or(x.Cx) .e~'.) + o.. u) .y)IIp(t' '')'~e~'t . ..=)limp(t.x) + 211c~ .43) and (12). u)) Tacking into account.)lPp( u) IIC~ .x) + (~. .y.A.~. cOV(x. x. we obtain av(x.~ . u)) T P ( ~ .L c ) T p(~.x)llp(t.(.( ~ . ~) < ..x) +(~ . u) IIC .Am . ~) = (~ .x ) T P (A} + ~(t.L ( C ~ ..A x .~)llp(t. / \ r ~c) = [AJc + ~(t. ~. ~..L(C} .4. then by (11) one gets.( ~ .)llp(t. u)) Since (.. OV(x. ~) < (. y.Hammami +.f(t.~ .x)rQ(.x) + 2 IIC~' .)) Thus.u) f(t.x ) V Q ( ~ .~ .x ) r P (~(t. It follows that. ..)~ av(x. y.) holds.x ) T ( A . u) . cOY(x. u) +2(~. y. x.~)llp(t. ~) < (~ .x) + 211c(.f(t.. (.. u) .~) + 211c(i .C x ) .x ) T Q ( ~ ..x) T PIcT( C $ .y)p(t.) IIC(~ .f(t. . x.LC)(~.. u)) which impliesthat. . (A~). M.x ) T p ( A . x.x)TQ(~ .
a Lyapunov function for the nominal part can be chosen as x T p x . from l e m m a 1. x) which satisfies p(t. using this idea.State Detection and Stability Now using the fact that.u). then the system is globally exponentially stable. Note that. y.Q < 0 (13) Under the above assumption. ~. we give a sufficient condition on the uncertainties to ensure the stability of system (4). (Q)II~ .~112+ 2ae Bt. with 1 x~.pT > 0 is such that PAK + A T p = . u). with ~(t. w e Un. Hence. the theorem 1 includes the linear case which treated in the section two.~o~(P). becomes an exponential observer for the system (4). one can reach conclusions about the sign definiteness of ~r by imposing certain restrictions on the uncertainties f(t. if the nonlinearities is norm bounded by a known positive function p(t. As an example.) and Amax(.45) There exist some positive constant kl. u)ll < klJlxll + k~llull. ~ < b. Therefore. B) is stabilizable.~O) k(t) < ~ X. Va > 0. then there exists a matrix K such that ReA(AK) < 0. 4 Stability Suppose that the system (4) satisfies the following assumption. x. which means that (~(t) . one can obtain an exponential rate of convergence and the system ~ = G(t. Therefore. the condition (10). for all (z.44) The pair (A. x) < k(t)llxl] . y. where AK = A + B K . Vu e Uq (14) . u) given in (12). where Amin(. such that Ill(t. by considering the derivative of V(x) = x T p x along the trajectories of the system (4). 5:) E ~n x Nn and u E ~ . ~.) are respectively the m i n i m u m and the m a x i m u m eigenvalue of the matrix (. (. follows by using the fact that e at is bounded and the latter inequality holds for any a > 0 (which can be choosen arbitrary small) so it implies (10). Vb > 0. the system (10) is detectable. ~. (. Hence. k2.z(t)) tends to zero globally and asymptotically. ~) ~ Am~.). it follows that 461 0v(x. where P . Assume that.
. it follows that 9 <_ _~. u(t))ll < *TQ* + 211*IIIIPII (kll[*ll + k2llu(t)ll) Using the form of u(t).. + Bu + f(t.Hammami Let ko = kl + k2llKII > 0. the ratio ~ x. x. T Q . Thus. is x. Then. we have the following theorem.ll) 9 ~ Am.. (. ( e ) . Suppose that the assumptions (..(p) is maximized with the choice Q = I (see Vidyasagar [20]). under the latter two conditions. .(Q)+ 2Am. u) Now.(O) + 211PIl(k. globally exponentially stable.. *.As) hold and ko satisfies ko < 8 9 Then.. the system (4) with the controller u(t) = Kx(t).:Cp)" P r o o f Consider the Lyapunov function V = x T P x which satisfies (13).A.462 M. from (14) + 211*llllPIIIIf(t. Note that. we obtain 9 <_ _ . T h e o r e m 2. . x. using the controller u(t) and the Lyapunov equation (13).A4). The derivative of V along the trajectories of (4) .(Q)llxll ~ + 211PIl(kl + k211KII)llxll 2 9 <(Am. + k211KII))Ilxll~ _< (A~.(P)k0)llxll z Since. k0 < ! ~ . it follows that the system (4) is globally exponentially 2x stable. ..TQ~: + 211PII(k~ll~ll+ k~ll/<~. u)) = 2xTpAx + 2 x T p B u + 2xTpI(t. satisfies ~" = 2 x T p ( A .
in [25] and [26] the authors used this condition to show the stability by output feedback for a class of nonlinear system.44) and (17) ~ the system is stabilizable. the authors characterize the class of linear systems specified by (. such that < p(t. z) + eoeZt' e0 > 0. x. The design of the controller to stabilize the system represented by equations (4) subject to the bound given in assumption (`46).46) There exits a positive scalar function p(t. K x + r x)) . (. x) 2 (16) 2 [IBTpxIIp(t. we obtain = 2xTpAx + 2xTpB(Kx + r x)) + 2xTpf(t. z) ([16]). one can proof that. /~ > O.x). with the controller (15). u)). vt. is described by u(t) = g (t) + (15) where r is defined as follows : 1 r x) = BTpxp(t. its leading Markov parameter CB is symmetric positive definite.C (17) then. The derivative of V along the trajectories of (4) . Suppose that the assumptions (`42). x. stable invertible and moreover. the assumptions (`42). P r o o f As the proof of theorem 2. let considers the Lyapunov function V = x T p x which satisfies (13). Also. is globally exponentially stable.44) hold and the following condition. Then. (As) hold and that BT p =. and r x) is continuous nonlinear state feedback which used to produce an exponential stability of the whole system in the presence of nonlinearity (uncertainties f(t. using the controller u = Kx(t) + f ( t . z(t)).44) and the condition (17). The proof is nearly identical to the proof of the theorem presented in [15]. Thus. consider the system (4) and suppose that p = q. (`44). (. x. satisfies iY = 2 x T p A x + 2 x T p B u + 2 x T p f ( t . u). they prove that. The controller given in (15) consists of Kx(t) which is a linear state feedback which stabilizes the nominal system. there exists a matrix K satisfying (. In [24]. the system (~). for linear system.State Detection and Stability 463 Finally. T h e o r e m 3.
one can construct an infinity of dissipative inputoutput maps. using this lemma.B D C .Control. . Vt. References 1. it follows that (18) ~_ . x) 2 ? <_ . r y) = _x_ ~P(t'Y)~ at e0 > 0. fl > 0. This result is known as the KalmanYakubovich lemma [22]. D. In particular. yields theorem 3. y) [15]. ~ < b. 2. Since the Lyapunov equation (13) can be satisfied for any arbitrary symmetric positive definite matrix Q. u)[[ < p(t. Exponential observers for nonlinear dynamic systems. B. x) + goe~t Now using the fact that. In the above theorem. with Aa = A. simply by using compatible choices of inputs and outputs.x T Q x + 211Cxllp(t . For more general 2 Ilvlla(t.x T Q x + 211BTpxlIp(t. the KalmanYakubovich states that given any openloop strictly stable linear system.mn(Q). the output controller u(t) = Dy + r y). Vb > 0. Va > 0. x) . the authors in [27] give an output controller which stabilizes globally and exponentially a class of uncertain nonlinear dynamic systems.x T Qx + 2eoe #t Applying lemma 1 to (18).Kou. x) 211BTPxlI2P(t' x)u IIBT PxlIp(t. where r is defined as.Thau. x) q.Journal. and there exits a positive scalar function p(t. C) are output controllable. where AD = A . by (As). that BT pxp(t.' systems. x).464 M. S.v)+eoe. A1 = Amin(P). Vu. Inform and Control. Int. the triple (A.Tarn.2 x T p B iiBTpxllp(t. there exists a matrix D such that ReA(AD) < 0. it is assumed that both u(t) and y(t) E ~P and B T p = C. Then. stabilizes the system (4). y).goe~t 211BTPxllp(t. (A6) and (16). such that lie(t.A. 17 (1993) 471479. x. one can choose an infinity of outputs from which the linear system will look dissipative. x) + eoe~t ? < .Hammami One can verify. Observing the states of nonlinear dynamic systems. given the system's phisical inputs and the associated matrix B. (A4). Vx.xTQx + IIBT Pxllp(t. F.Eoefit ~" <. AS = Ama~(P) and e = 2e0. A simple choice is that the output controller is linear. 29 (3) (1976) 204216. Suppose that.Elliott and T.
13.S. SIAM.Cont.Kalman and R.Iggidr and G. AC25 (1980) 504509.Krener and W. An introduction to observers.J.Cont.Dec. 3 (1983) 4752.Cont.Autom. Applied nonlinear control. IEEE. IEEE Conf.G. R.Slotine and W. Comparative study of non linear state observation techniques. 23 (1985) 197216.Hammami. D.Bycy. J. (1991). 16 (1971) 596602.A.Aut. Syst. Xiahua Xia and Weibing Gao. 26.Qu and D. vol AC30 N10 (1985) 1025. 4.Cont. IEEE.J. 7. Int. IEEE Trans.Cont. Observability canonical (phasevariable) form for nonlinear timevariable systems. Syst. 32 (2) (1987) 166170.C.Walcott.Lett. 11 (1988) 319325. SIAM.Zak.Cont. Exponential stability of a class of nonlinear dynamic systems with uncertainties.Steinberg and M.Corless and H.Sallet. 6 (1993) pp 710.J. Practice Hall. Global stabilization of nonlinear systems with a class of unmatched uncertainties. App. Output feedback stabilization of uncertain dynamical systems. A.Sussmann. . On the output feedback stability for nonlinear uncertain control systems.J.C. Syst. vol 15 n 9 (1984) 949958. 11.Mizukami. 20. Syst. Z. EnglCliffs. J. 8. A. Nonlinear systems analysis.Cont.Lett 21 (1993) 307313.Hammouri. M.Lett. San Antonio Texas (1993) 19541959. Z.Lett.Lett.Khalil. submitted.Lett.Benabdallah and M. Prentice Hall.Qu and J. M. On the state observation and output feedback problems for nonlinear uncertain dynamic systems.Vidyasagar. 83 (1961) 95108.Zeitz. 5.Trans.Respondek. 12.Optim.Zeitz.Trans. Stabilization of a class of nonlinear systems using an observer design. N6 (1990) 14911503. Mac Millan (1992). Int.Cont.Aut. 22.J. 18. Nonlinear Systems. A.1027. Continuous feedback control guaranteeing exponential stability for uncertain dynamic systems. A separation principle for bilinear systems with dissipative drift.Cont.K. M.Chabour and H. Global stabilization of partially linear composite systems.Dec.L. Syst. B. vol 45 N6 (1987) 21092132.M. A generalization of Vidyasagar's theorem on stabilizability using state detection. 25. 38 (1983) 419431. Exponential stabilization of nonlinear systems by an estimated state feedback. Nonlinear observers with linearizable error dynamics.Kupka. 27.Krener and A.Isidori. 9. Proc of 32nd IEEE Corff. On the stabilization of nonlinear systems using state detection. IEEE Trans. H.Aut.Cont.Qu. Syst. AC vo137 (1992) 12 pp 19701974. (1991).J.Cont.Hammami. A.Corless.Cont.State Detection and Stability 465 3. 17.J.P. D. R. 21. Canonical form observer design for nonlinear time variable systems. Stabilization of planar bilinear systems using an observer configuration.Wu and K. ECC'93 (1993) 20152018. 24.ASME ser D. Tr. Z.Carroll.Auto. 18 (1992) 217222. A.V. J. On exponential observers for non linear systems. H. Linearization by output injection and nonlinear observers.Luenberger. M.Dawson.E.Math.Basic Eng.J. M. A.Kokotovic and J.Dawson. P.Syst.Bestle and M. 23. 6.Li. 15.Cont. 2nd edition (1993).Cont.Saberi. D. vol 28.Sci.Control.Gauthier and Y. 10. IEEE Trans. State observation of nonlinear uncertain dynamical systems.Zak.Optim. Int.Walcott and S. New results in linear filtring and prediction theory. 19.Cont. B.Lett 18 (1992) 301307.J.A. 17 (1991) 3742.Tsinias. 16. 14.Vidyasagar.
W i t h this point of view s t a n d a r d shift strategies as the well known R a y l e i g h iteration can be interpreted as feedbacks for the control system. For m a n y numerical m a t r i x eigenvalue m e t h o d s such as the Q R a l g o r i t h m or inverse iterations shift strategies have been i n t r o d u c e d in o r d e r to design a l g o r i t h m s t h a t have faster (local) convergence. For an a l g o r i t h m using m u l t i d i m e n s i o n a l shifts for t h e Q R a l g o r i t h m see the p a p e r of Absil.15] and in [17.4. where the l a t t e r references c o n c e n t r a t e on complex shifts.18]. Germany f abian@@math. de Zentrum flirTechnomathematik Universit~it Bremen 28334 Bremen. We analyze controllability properties of the inverse iteration and the QRalgorithm equipped with a shifting parameter as a control input.a l g o r i t h m ) t h a t the b e h a v i o r of the Rayleigh shifted a l g o r i t h m can be very complicated. I Mathematisches Institut Universit/it W/irzburg 97074 Wfir'zburg.H e r m i t i a n m a t r i ces A [4]. . This paper was written while Fabian Wirth was a guest at the Centre Automatique et Syst~mes. first results on the extension to inverse Rayleigh iteration on Grassmann manifolds using complex shifts is discussed. It is k n o w n (for instance in the case of the inverse i t e r a t i o n or its m u l t i d i m e n s i o n a l analogue. Ecole des Mines de Paris. Finally. unibremen. France. in p a r t i c u l a r if it is a p p l i e d to n o n . T h e shifted inverse i t e r a t i o n is studied in [3. de A b s t r a c t . Germany helmke@@mathemat ik. We discuss an interpretation of the system in terms of control systems on rational functions. In the case of the inverse iteration with real shifts the theory of universally regular controls may be used to obtain necessary and sufficient conditions for complete controllability in terms of the solvability of a matrix equation. uni~uerzburg. Fontainebleau.Controllability Properties of Numerical Eigenvalue Algorit hms Uwe Helmke 1 and F a b i a n W i r t h 2. In this p a p e r we interpret the shifts as control i n p u t s to the a l g o r i t h m . which u p to now has been h a r d l y studied. Sepulchre and van Dooren in this book. Mahony. It is therefore of interest to o b t a i n a b e t t e r u n d e r s t a n d i n g of the underlying control system. the Q R . The hospitality of all the members of the centre is gratefully acknowledged. Partial results on conditions for the solvability of this matrix equation are given.
This will be used in Section 2 to give a characterization of complete controllability of the system on projective space in terms of solvability of a matrix equation. As it turns out the results depend heavily on the question whether one uses real or complex shifts.468 Uwe Helmke. (1) where ut q~ ~r(A). Some concrete cases in which it is possible to decide based on spectral information whether a m a t r i x leads to complete controllable shifted inverse iteration are presented in Section 4.e. u). If the initial condition x0 for system (1) lies in an invariant subspace of A then the same holds true for the entire trajectory r x0. those x E Ii~n such that . regardless of the control sequence u. We show that the corresponding control system on the Grassmannian is never controllable except for few cases. . respectively the G r a s s m a n nian manifold for the QRalgorithm. Let A denote a real n x nmatrix with spectrum or(A) C C. In Section 1 we introduce the shifted inverse power iteration with real shifts and the associated system on projective space and discuss its forward accessibility properties. t h a t they steer every point into the interior of its forward orbit. This describes a nonlinear control system on the (n . . In particular. Fabian Wirth Here we focus on controllability properties of the corresponding systems on projective space for the case of inverse iteration. [4]. The reachable sets are characterized in terms of Grassmann simplices.1)sphere. The controllability of the inverse iteration with complex shifts has been studied in [13]. that is those sequences with the property. while the real case is treated in [14]. Therefore we m a y restrict our attention to those points not lying in a nontrivial invariant subspace of A. In Section 6 we turn to the analysis of the shifted Q R algorithm. u l . i. . ) is denoted by r x0.utI)lz(t) ' t ~ N. 1 T h e s h i f t e d i n v e r s e i t e r a t i o n on p r o j e c t i v e s p a c e We begin by reviewing recent results on the shifted inverse iteration which will motivate the ideas employed in the case of the shifted Q R algorithm. The trajectory corresponding to a normalized initial condition x0 and a control sequence u = (u0. u).u t I ) . An interpretation of these results in terms of control systems on rational functions is given in Section 5. Via the choice ut = x * ( t ) A z ( t ) we obtain from (1) the Rayleigh quotient iteration studied in [3]. In Section 3 we investigate the obtained characterization and interpret it in terms of the characteristic polynomial of A. In order to understand the controllability properties from x0 it would then suffice to study the system in the corresponding invariant subspace. We conclude in Section 7. there is an easy characterization of the set of universally regular control sequences. The shifted inverse iteration in its controlled form is given by x(t + 1) = II(A.l x ( t ) [ [ (A.
n . t 9 I~ (3) ~(0) = ~0 9 M . M is an open and dense subset of~lP n1. the set of points reachable exactly in time t is denoted by O + (~). where ut 9 U := ~ \ a(A) (the set of admissible control values). Sard's theorem implies then the existence of a control u 9 U t such that rk O~(t. u ) } . System (3) is called forward accessible [2]. The control sequence u E U t is called universally regular if (~. Corollaries 3. .2 & 3. Using the fact that the interesting dynamics of (1) are on the unit sphere and identifying opposite points (which give no further information) we then define our state space of interest to be M := ~IP" . Note that int O+(~) r ~ holds iff there is a t 9 H such that int O+(~) r ~. u) E M x U t is called regular if this rank condition holds. The system on M is now given by ~(t + 1) = (A  utI)l~(t). (2) where R]I~ . The solution of (3) corresponding to the initial value ~0 and a control sequence u 9 U N is denoted by T(t. Since a cyclic matrix has only a finite number of invariant subspaces. Ou A pair (~. defined by the cyclic vectors of A.1 denotes the real projective space of dimension n . Moreover. Similarly. A n ..~0.x \ V ( A ) .Controllability Properties of Numerical Eigenvalue Algorithms 469 {z. respectively.. which we will always assume in the following. . To keep notation short let us introduce the union of Ainvariant subspaces V(A) := U A V c V.O< d l m V <n V. I](A) is a closed algebraic subset of ~ n .u). if the forward orbit (. the 1dimensional linear subspaces of ~n.1 . u 9 U t such that 1/= ~ ( t . Vectors with this property are called cyclic and a matrix A is called cyclic if it has a cyclic vector. ~ .9+ (~) of every point ~ E M has nonempty interior and uniformly forward accessible (in time t) if there is a t 9 1~1 such that intO+(~) • 0 for all ~ 9 M.i x } is a basis of ~n. u) . By [16. . The natural projection from ~ n \ {0} to ]R]l ~n1 will be denoted by 11 Thus M consists of r. A x . . The forward orbit of a point ~ 9 M is then given by O+(~) := {7/ 9 M I ::It 9 H .1 . u) is a regular pair for every ~ E M.~. We denote the space of finite and infinite admissible control sequences by U t and U r~.3] forward accessibility is equivalent to the fact that the set of universally regular control sequences .
. Fabian Wirth Ut~g is open and dense in U t for all t large enough.470 Uwe Helmke. x. In order to analyze the controllability properties of (3) we introduce the following definition of what can be thought of as regions of approximate controllability in M C ~ I ~ 1. (For a precise statement we refer to [16]. An important subset of a control set D is its core defined by core(D) := {~ 9 D [ i n t O .1 pairwise different values in the sequence u o . M is a control set o f s y s t e m (3). . Consider the s y s t e m (3) on M . Let A 9 ~ n x n be cyclic. T h e o r e m 1. we will call a point ~ E M controllable to ~/ E M if 7/ E O + (~). We are now in a position to state a result characterizing controllability of (3). System (3) is said to be completely controllable on a subset N C M if for all ~ E N we have N C (9+ (~). (iii) D is a maximal set (with respect to inclusion) satisfying (i). . S y s t e m (3) is uniformly forward accessible in t i m e n . u t . L e m m a 1.( ~ ) n O r 0 and intO+(~) n D r 0}.1 . Here (9(~) denotes the points r/ 9 ~ I ~ 1 such that there exist t 9 1~1. By this assumption it is evident that on the core of a control set the system is completely controllable. S y s t e m (3) is completely controllable on M . .1. . The following statements are equivalent: (i) (ii} (iii} (iv) There exists a ~ 9 M such that 0 + (~) is dense in M . u0) is a regular pair. r]. As usual. A control sequence u 9 U t is universally regular i f and only i f there are n . u) E D. u0 9 int U t such that ~o(t. There exists a control set D C M with int D :/: 0. A control set of system (3) is a set D C M satisfying (i) D C c l O +(~) for a l l i E D . u0) = ~ and (r/.) The following result shows forward accessibility for (3) and gives an easy characterization of universally regular controls. 2 Controllability of the projected system By the results of the previous section we know that every point in M has a forward orbit with interior points and it is reasonable to wonder about controllability properties of system (3). see [14]. (ii) For every ~ E D there exists a u E U such that ~(1.
We have another simple characterization of IIcontrollability in terms of the existence of a universally regular periodic orbit through a cyclic vector v of A. In the following theorem we use the notation p A q = 1 to denote the fact that the two polynomials p.Controllability Properties of Numerical Eigenvalue Algorithms (v) There exists a universally regular control sequence u E U t such that t1 471 H(A$0 u . The following statements are equivalent: (i) the matrix A is lIcontrollable. already the fact that there is a control set of the system implies complete controllability on the whole state space M. 3 Polynomial characterizations of IIcontrollability As has already become evident in the last result of the previous section the question of IIcontrollability is closely linked to properties of real polynomials. if A satisfies any of the equivalent conditions of Theorem 1. C o r o l l a r y 1. The following statements are equivalent: (i) the matrix A is IIcontrollable. Let A E ]I~ n x n be cyclic with cyclic vector v and characteristic polynomial qa. $~0 . a E I~* such that t1 B = a I . then no forward orbit of (3) is dense in M. I ) 1 E ]~'I.{p(A) [ p E ~ [ z ] . Furthermore. u E Utreg. 5 i ) For every B E FA :. see [14]. p A q .I ( A . Consider the system (3) on M . Let A E I~ nxn be cyclic with characteristic polynomial q. that if the conditions of the above theorem are not met. u E Ut~eg such that I?v is a periodic point f o r system (3) under the control sequence u. I I . We will now further investigate this relationship. For brevity we will call a cyclic matrix A Ifcontrollable (for inverse iteration controllable). (ii} There exist t E I~.1} there exist t E N. Here we follow the ideas for the complex case in [13] and discuss comparable results for the real case. On the other hand it is worth pointing out. q E ~[z] are coprime. T h e o r e m 2. Consider the system (3) on M .u. This may come as a surprise. (4) The unusual fact about the system we are studying is thus that by the universally regular representation of one element of the system's semigroup we can immediately conclude that the system is completely controllable.
The reason for this is that.I ( z . Let A be lIcontrollable with characteristic polynomial qA. and the assumption guarantees that all polynomials in a neighborhood of f have simple real roots. As an immediate consequence of Theorem 1 we obtain a complete characterization of the reachable sets of the inverse power iteration given by ~(t + 1) = (A . . factors of qa(z) over the polynomial ring ~[z]. (5) Remark I. u s E ~ . (ii) There is a onetoone correspondence between a) b) c) d) The The The The forward orbits of system (6).n1. Ax. a E lt~* such that t1 p(z)=aH(zu. p A q = I . u E U~. A n .AVCV v\ U ~rf~V.~g. A 2 x .utI)a~(t). (6) for IIcontrollable matrices A E ~nxn.AVcV V = ~span{x. .u . a e ~* and r(z) ~ ~[z] such that f(z)=a+r(z)q(z). . .e. closures of the forward orbits of system (6).AVCV v. a E ~ * } .472 Uwe Helmke. /. then there is a neighborhood of A consisting of IIcontrollable matrices. F A = F A ~:={p(A) I P ( z ) = a 1 . ~EV. t (iii) For every p E IR[z]. .~0 . t E N.p A q = 1 there exist t E N.i x } .) 0~0 modq(z). (iv) There exists a monic polynomial f with only real roots and at least n .From (5) it is easy to deduce the following statement: If for a cyclic matrix A with characteristic polynomial q there exists a monic polynomial f with only real roots that are all pairwise distinct such that (5) is satisfied. keeping a and r(z) fixed. then (i) for each ~ = ~ x E ~ I ~ . This extends a result in [13] for real matrices.t we have = clO+(~) = ~ fl N xEV.. ) . C o r o l l a r y 2. Fabian Wirth t1 i. ~(0) = ~0 E ~a. small changes in the coefficients of q will only lead to small changes in the coefficients of f .1 pairwise different roots. Ainvariant subspaces of ~ n.
Let A E 1Rn• be cyclic with characteristic polynomial qA. . one sees that the matrix corresponding to p(z) = (z 2 . the three cubic roots of i and their respective complex conjugates. (i) A is not IIcontrollable. showing obstructions to this property in terms of the location of the eigenvalues are discussed in detail in [14]. Many more examples like this can be constructed. (iO If the spectrum ~(A) is symmetric with respect to rotation by a root o f unity. if it satisfies one of the followin 9 conditions (a) A has a nonveal eigenvalue of multiplicity p > 1. the companion matrices of the polynomial p(z) = z(z 2 + 1) and the 7th degree polynomial whose roots are 0. Proposition 1. some more examples are discussed in [14]. m is even.A2 e a ( A ) we have ReAl = Rem~. Several examples. furthermore. If. Let A E IRnxn be cyclic. IAII r Im2l then A is not IIcontrollable. (ii) A is II.1)(z 2 + 1) is not IIcontrollable. As.controllable. (i) If n = 1.e. Such are e. 2 then A is IIcontrollable. Using the last statement. (b) A has a real eigenvalue of multiplicity p > 2. then it is sufficient that f o r two eigenvalues of A m we have IX~ ul  < Ia2  ul. for all u > 0 in order that A is not IIcontrollable. i. (ii) If n = 3 then A is IIcontrollable if and only if the eigenvalues A1. a complete characterization of the set of cyclic matrices that is not IIcontrollable is not known. In general. A3 of A do not have a common real part. if it(A) C ~ and no eigenvalue has multiplicity# > 2. These are obtained via the following result. (i) If for two eiyenvalues A1. Let A E ]l~nxn be cyclic.Controllability Properties of Numerical Eigenvalue Algorithms 473 4 C o n d i t i o n s for II c o n t r o l l a b i l i t y The result of the previous section raises the question which cyclic matrices A admit a representation of the form (4) or equivalently when (5) is possible. For the case n < 3 the following complete result can be given. Using this corollary it is easy to construct examples of matrices that are not//controllable. With respect to this question we have the following preliminary results.g. Proposition 2. Proposition 3. ~r(A) = exp(27ri/m)cr(A) (taking into account multiplicities) and two eigenvalues of A m satisfy the condition of (i) then A is not IIcontrollable.e. i. do not satisfy ReAl = ReA2 = Re A3.
at(ut). Then. two rational functions gl (z).l z defines a bijective isomorphism between ~ n and Rat(q). Fabian Wirth Control system on rational functions There is an interesting reformulation of the inverse Rayleigh iteration as an equivalent control system on rational function spaces.A ) .474 5 Uwe Helmke. Let (lPxt)t~r~ denote the sequence in ~IP '~1 generated by the inverse power iteration (3).(z) . Then ~o induces a homeomorphism r ]RII ~n1 + ~(Rat(q)) r := ] P c ( z I .l x t = ~(g. The m a p ~o : ~ n _.A ) . i. for gt(z) := c(zI  A)lxt we obtain the divided difference scheme IFgt+l(z) = I P e ( z I . Z 'l~ t Conversely. Z 1L t then 9t E Rat(q) for all t @ 1~ and 9t(z) = c(zI . Let Rat(q) := lq(z)[P(Z)~ ( z ) I E d e g p < degq} denote the real vectorspace of all strictly proper real rational functions with fixed denomination polynomial q(z). as well as with divided difference schemes in interpolation theory. Let (c.+ Rat(q) x ~~ c(zI . We use it to transport the inverse Rayleigh iteration onto Rat(q). A) E ~ l x n x ~ n x n be an observable pair and let q(z) := d e t ( z I .A ) .l x .A) denote the characteristic polynomial.A ) .e. This connects up with the work by Brockett and Krishnaprasad [9] on scaling actions on rational functions. Let l~(Rat(q)) denote the associated projective space.I ( A .l z 0 E Rat(q) and (~at)ter~ is recursively defined by at+l(z) at(t) .u t I ) .l xt .g2(z) E Rat(q) define the same element II~gl = IP92 in ll~(Rat(q)) if and only ifgi and g2 differ by a nonzero constant factor. if go(z) = c(zI .t E No.)).A ) .gt(u.
For simplicity we focus on the complex case..p(u)q(z) ZU is again a polynomial of degree < n. that a partial flag in C n is an increasing sequence {0} ~ V1 C . To include such algorithms in our approach we have to extend the analysis to inverse iterations on partial flag manifolds.Controllability Properties of Numerical Eigenvalue Algorithms 475 for a sequence (IPxt)teNogenerated by the inverse Rayleigh iteration.. Since (8) we see (for q(ut) # 0) that the control systems (7) and (8) are equivalent.. The type of the flag ( V 1 . To this end let lib. A full analysis is beyond the scope of this paper and will be presented elsewhere. we can reformulate this algorithm as a control system on polynomials of degree < n. ..[z]) denote the associated projective space. . . . Note that for any polynomial p E ~n[z] and uEIR P(z) := p(z)q(u). 6 Inverse iteration on flag manifolds A well known extension of the inverse Rayleigh iteration (3) is the QRalgorithm. Thus the inverse Rayleigh iteration (3) on ]R]I m1 is equivalent to the divided difference control system ]Pgt+l(Z):~('gt(Z~ gt(ut)'~ut ] (7) on 1P(Rat(q)). Thus ]P(pt+l(Z)) = ~ (pt(z)q(Ut)z Z utPt(ut)q(z)) defines a control system on ~(I~.[z])..[z] := {p E ~[z] [ degp < n} denote the vectorspace of polynomials of degree < n and let P(~. i=l. ak) of dimensions ai = dirnc ~. Vk) is specified by the ktuple a = ( a l . . Equivalently.. .k. Recall. .C Vk C C" of Clinear subspaces ~ ~: of C n. . .
Ca ).l l ) lt 9 N. Ca). . . the reachable sets of the inverse iteration on the flag manifold Flag(a. C n ) have empty interior. Ca). . Except for k = 1.1.. . ut ~ a(A). k = n . Moreover. n). . . < ak < n. q ( u i ) r 0} is defined as in Section 3. a : = a l . C a ) ~ Flag(a. .utI) 1 : F l a g ( a . Vk) of t y p e a. .476 Thus Uwe Helmke.. . 2 . a E C * . T h e reachable sets are again easily seen to be equal to t1 Ua(12) : = { H ( A i=0 u i I ) . C ~ ) is called a (partial) flag manifold. C) we o b t a i n dimT~a(Y) < n . of complex n u m b e r s we o b t a i n the inverse iteration on flag manifolds (A . C n ) > n Thus we conclude Proposition 4. Now d i m F l a g ( a .1 or n = 2. . . V P 9 F l a g ( a .. ui ~ a ( A ) } .<ak<_n. 1. It is indeed a c o m p a c t c o m p l e x manifold. k = n 1 or n = 2. For any linear m a p A : CT + C n a n d any sequence (ut). a k ) . Ca) = FA 12 r where the s e m i g r o u p t1 Fa = {p(A)[p(z) = all(zc i=0 u i ) . Fabian Wirth l<al<. For any such sequence of integers a = ( a l . . . let F l a g ( a . n) we o b t a i n the (full) flag manifold F l a g ( c a ) := F l a g ( ( 1 .a l g o r i t h m with shifts is equivalent to the inverse iteration (A  utI) 1 : F l a g ( C n) + Flag(eL ) m we o b t a i n . . except for the cases k = 1. For k = 1. T h e set F l a g ( a . . . 1 < a l < . . 12 E F l a g ( a . This defines a nonlinear control s y s t e m on the flag m a n i f o l d . C " ) d e n o t e t h e set of all flags ( V 1 . since t h e Q R . we o b t a i n the Grassmann manifold Ga(C n) as a special case while for k = n a n d a = (1. Since d i m F a < n a n d since FA acts with a c c stabilizer of d i m e n s i o n > 1 on F l a g ( a . .
a r of X . For any increasing sequence a of integers 1 < a i < . For any subset a = { a i . [8]. D e f i n i t i o n 1.n. < a .al.Controllability Properties of Numerical Eigenvalue Algorithms 477 C o r o l l a r y 3.d i m e n s i o n a l subspace s p a n n e d by the columns of X . ..span 34 01 02 10) ' Y = span . 1 < a i < . We now describe in more d e t a i l the s t r u c t u r e of the reachable sets.. Every reachable set ~ A ( [ X ] ) in G k ( C n) is contained in a Grassmann simplex. < a~ _< n. the following two m a t r i c e s span rank equivalent subspaces. . <n andr= 1. . . ( R e m e m b e r t h a t A is d i a g o n a l !) L e m m a 2. let ea "~ eal ~ " " " Jr ea~ . . [7]. . . For example.n}. i f n >_ 3. . . . For us t h e y are of interest because of the following fact. . (b) R a n k equivalence defines an equivalence relation on G k ( C n).. T h e equivalence classes are called Grassmann simplices of Gk (C ~). The Q R . (a) T w o c o m p l e x linear subspaces [X].a l g o r i t h m with origin shifts is not locally accessible nor controllable. . . < a r _< n let X a denote the r x k s u b m a t r i x f o r m e d by the rows a i . . For simplicity we assume t h a t A E C nxn is a d i a g o n a l m a t r i x with has d i s t i n c t eigenvalues and we focus on the inverse i t e r a t i o n on G r a s s m a n n m a n i f o l d s (A .. For any full rank m a t r i c e s X E C n• [X] : = I m X e Gk (C") denote the k . To o b t a i n a more precise d e s c r i p t i o n of reachable sets and G r a s s m a n n simplices we consider a p r o j e c t i o n of the G r a s s m a n n i a n on a p o l y t o p e . let X . . [Y] in C n of d i m e n s i o n k are called rank equivalent if rkXa = rkYa for a l l l < a i < . . a t } C n : = { 1 . . T h e s t a b i l i z a t i o n of G r a s s m a n n m a n i f o l d s into G r a s s m a n n simplices has been introduced by Gelfand et.utI) . .i : G k ( C n) + G k ( C " ) .
~_<~_<.478 Uwe Helmke..n. t. . . denotes the i . = k}. + . .. It is a closed subface of Ak.. T h e o r e m 3. It is easily seen that the image of p in ~ n is a convex polytope.) e ~ It.t h standard basis vector of C n . In the complex case. More precisely we have ._<P~<.. reachable sets for the inverse iteration on projective space C]F n1 correspond bijectively to invariant subspaces .(TEA ([Y])). is contained in a Grassmann simplex. . . + t.(V~(~)) = Ak.+~n. [Y] E Gk (C n ) are rank equivalent if and only if (Tea ([X])) = .~ := ( ( t . We refer to it as the moment map o n Gk((~m).n denotes the hypersimplex A~. More precisely. . (b) p(Tea([X])) is a compact polytope in 1~n with vertices {ea ] d e t X a # 0}. two subspaces IX].i ) for any invertible matrix S E ~ x k defines a smooth m a p and thus p ( X ) p: a k ( ~ ) _. 7 Conclusions Controllability properties of inverse iteration schemes provide fundamental limitation for any numerical algorithm defined by them in terms of suitable feedback strategies. [7] describes the geometry of Grassmann simplices in terms of the moment map.n where A~. [X] E Gk(Cn).<ctk <n ~(x) :=  [ det Xa 12e~ [ det(XpX~')l " l<Pl <. (c) There is a bOective correspondence between (i) pdimensional reachable sets in Tea ([X]). 1 < i < n. p([g]):= p(X) on the Grassmann manifold. Then p ( X ) = p ( X S . Fabian Wirth where el. (a) Every reachable set TeA([X]). .al.. The following result by Gelfand et. For any full rank matrix X E C '~x~ define Z l<al<. (ii} Open pdimensional faces of p(Tea ([X]) ).
Batterson. 33. S. 31. R. Differences also occur for inverse i t e r a t i o n on G r a s s m a n n i a n s or flag m a n i folds.. in Math. 301316 9. to appear. Batterson. (1997) Computing an eigenvector with inverse iteration. Goresky. (1993) Discretetime transitivity and accessibility: analytic systems. V. to appear. IEEE Trans. 2. Nauk.2 5 . Moreover. Soc. P. 26. 297310 4. MacPherson. see [13]. Reachable sets are c o n t a i n e d in G r a s s m a n n simplices a n d their adherence relation is described by t h e c o m b i n a t i o n s of faces of a h y p e r s i m p l e x .. Syst. Fuhrmann. F.a l g o r i t h m is seen to be not controllable. Ipsen. Comput. Wirth F. Proc.M. V. W. (1993) Linear control semigroups acting on projective space... J. Trans.V.W. 15. Uspekhi Math... Cont. A. E.M.S.Controllability Properties of Numerical Eigenvalue Algorithms 479 of A. 2. 39. C. (1996) The Lyapunov spectrum of famihes of time varying matrices. Batterson. Albertini. Kliemarm. Gelfand. P. Smillie.. l. Batterson.. 348. 14. Syst. l. F. (1995) Dynamical analysis of numerical systems. Helmke.. Control A C . J. A 68. R (1998) Remarks on universal nonsingular controls for discretetime systems. School New Brunswick Rutgers. New Brunswick. Sontag. 624636 7. U. Cont. W.V. T h e a l g o r i t h m s are never controllable. Wirth. S. 2. Anal. T h e real case is considerably harder a n d o n l y p a r t i a l results for c o m p l e t e c o n t r o l l a b i l i t y in t e r m s of necessary or sufficient c o n d i t i o n s are given. Math. (1990) Rayleigh quotient iteration for nonsymmetric matrices. Serganova. Smillie.. New Jersey. Lett. 55. Linear Algebra Appl. SIAM J. Autom. (1987) Combinatorial geometries and torus strata on homogeneous compact manifolds. (2000) On controllability of matrix eigenvalue algorithms: The inverse power method.. (1989) Rayleigh quotient iteration fails for nonsymmettic matrices. (1993) Controllability of discretetime nonlinear systems and some related topics. Dynamics Diff. Syst. Lett. 8188 17. Contr. (2000) On controllability of the real shifted inverse power iteration. Numer. 68?3 . 1920 6. Japan Acad. E. J. F. Gelfand. & Opt. S. J. F. SIAM Rev.D. Numer. Kliemann. Cont. (1996) A Polynomial Approach to Linear Algebra. Springer Publ. U. Amer. D. Grad. F. Fuhrmarm. Krishnaprasad. F.. in p a r t i c u l a r the Q R . 43894408 12. 495528 11. Serganova. 15991622 3. New York 13. Colonius. 169178 5.. 254291 16.. PhD thesis. T (1992) Inverse iteration method with a complex parameter. Adv. S. 42. Math. Appl. References 1. Lett. Math. Lett. c o m p l e t e controllability holds if a n d only if A is cyclic. 197207 10. Colonius. Brockett. Equations 5. 107134 8. Sontag. (1989) The dynamics of Rayleigh quotient iteration. 63. R. A. Smillie. convex polyhedra and Schubert cells. Ser. SIAM J. R.. Albertini.. Helmke. P. I..M. (1980) A scaling theory for linear systems. (1987) Combinatorial geometries. Suzuki..
K. (1998) Dynamics and controllability of nonlinear discretetime control systems. Adv.). SIAM 2. Spectral scattering theory and applications. & Opt. Fabian Wirth 18. in Yajima. 23. Proceedings of a conference on spectral and scattering theory held at Tokyo Institute of Technology. T (1994) Inverse iteration method with a complex parameter. Wirth. The Netherlands. 36. Contr. F.275 . In 4th IFAC Nonlinear Control Systems Design Symposium (NOLCOS'98) Enschede. 269 . 1992 in honour of Shige Toshi Kuroda on the occasion of his 60th birthday. Wirth. Tokyo: Kinokuniya Company Ltd. Pure Math. F. (1998) Dynamics of timevarying discretetime linear systems: Spectral theory and the projected system. Stud. 447487 20.. (ed. 307310 19. Japan. II. June 30July 3.480 Uwe Helmke. Suzuki.
On the Discretization of Slidingmodelike Controllers Guido Herrmann*. B. Spurgeon. and Christopher Edwards Control Systems Research Group University of Leicester Leicester LE1 7RH. Numerical results are demonstrated for a particular example. Such an approach of [3. engg. It has been shown before that a strictly proper linear continuoustime system with sampleddata output has finite L~galn. This is because controllers are easily developed in continuous time but often need to be applied via sampleddata technology. B.uk Abstract. An extensive framework of theory has been developed for linear. Nonlinear systems often rely on Lyapunov function based analysis techniques [15].> 0. * G. In contrast.K. This result has also been shown for linear timevarying systems [7]. sampleddata control of linear systems [68]. Sarah K. ac . the Lpnorm of the discretized output signal approaches the Lrnorm of the continuous output signal of (A. project number: FMBICT983463). Herrmann is supported by a grant of the European Commission (TMRgrant. This gain converges to the L2gain associated with the continuoustime output when the sampling period approaches +0. le. the control of the nonlinear inverted pendulum. robust closed loop control using a discretized nonlinear control. U. This result is incorporated in the analysis of the discretized slidingmode based control applying techniques from nonlinear L2gain theory. The result is then compared to a Lyapunov function analysis based approach. An important result for the discretization of LTI systems is that any stable linear system (A. Furthermore. 1 Introduction Analysis of sampleddata implementations of continuoustime control systems has been of interest for a long period. In contrast to the Lyapunov function technique. C) as r ~ +0. C) followed by a sampleandhold process has finite Lpinduced norm for constant sampling time ~. . Stability of a discretized continuous slidingmode based state feedback control is proved using an L~galn analysis result for linear continuoustime systems with sampleddata output. the samplingtime constraint vanishes for a stable plant if no control is used.4] has been useful for determining samplingfrequencies which are sufficient for stable. ghlT@sun. linear timeinvariant (LTI) systems are particularly useful for the application of 7/~ techniques [6]. The technique has been applied successfully to a slidingmode based control.
z) (2) .is chosen small enough. is used to compare the results with those using a Lyapunov function based approach [3. For (1) being valid. an upper bound for the norm difference between the timesampled state z(ti) and the value of the continuous signal a~(t) is found for each particular finite sampling interval [ti. 2 A SlidingModeLike Continuous Control Law Linear. The continuous slidingmodelike control is introduced in Sect.4] is that the relation (1) can only be derived if an upper bound f on the samplingtime 7. L2gain analysis techniques have proved to be useful for nonlinear systems as the basis for nonlinear 7/~theory. 4. A problem of this approach [3.t~+~]. exponentially stabilizing nonlinear controls imply a finite L~gain. Furthermore. this upper bound ? remains finite for any control even if the nonlinear system is stable and the applied control is zero. Thus. It will be seen that both approaches have their advantages and disadvantages from a theoretical and practical point of view. a Lyapunov function can be derived from the HamiltonJacobi inequality. L~gain techniques are applied to show that the discretized control is stable and robust to the same class of uncertainty if the sampling time 7. In particular. uncertain systems shall be considered = Az + Bu + F(t. the control of an inverted pendulum.< ? is assumed. implying that a small gain relationship needs to be satisfied for stability of the nonlinear control. The states ultimately reach a coneshaped boundary layer around a slidingmode plane. Section 3 deals with the proof of stability of the discretized control while the example of the control of the inverted pendulum is provided in Sect.t. ~ _ 0. ) l l ) t~E[ti.~ ( t . The results of [6] for sampled data LTI systems are given in the Appendix.9]. This paper is structured as follows. This continuoustime state feedback control exhibits exponential stability properties and is robust with respect to matched parametric uncertainty. A practically valid example.+ll _ ~~ II~(t)ll. 2. (1) The upper bound decreases with decreasing sampling time v. a m i n i m u m s a m p l i n g frequency is always indirectly imposed. In the Lyapunov function based approach. The Lpgain result for LTI systems of [6] is incorporated. This minimum sampling frequency is not necessary in the L~induced norm approach. which is a fundamental tool of L~gain theory. a nonlinear continuoustime slidingmode based control derived from [9] will be investigated with respect to its sampleddata implementation. t ~ [t~. Thus. ti+l]: sup (ll~(t~) .2.482 Guido Herrmann et ai.
The matrices Ao and An are bounded so that The above represent the matched uncertainty and disturbances.) : ]R x ~ " + ]R" models parametric uncertainties and nonlinearities of the system lying in the range space of B. (7) . that the slidingmodelike control law outlined below forces the system states ~ to satisfy ~5(t) ~ 0 and the control is robust to matched parametric uncertainty.~ ' z l + Al~q~. . Thus.) are the linear and the nonlinear control components. It can be shown similar to [9].Q*)O(t)) (6) mainly enhances reachability of a cone shaped layer around the sliding mode. B) defines the nominal linear system.Ira.. where J2* is a stable design matrix and P2 satisfies P2YI* + ~*TP2 . .) is Caratheodory to ensure existence of solution (see [3. (4) and .10]).T. = T x = [ z tT ~T ]T. (3a) (3b) = (O+Ae)z where I + (gl+An)d~+B~u. which is assumed to be controllable with B of full rank.. whilst unmatched and actuator uncertainty has been neglected for simplicity in contrast to [9]. it is possible to define a linear transformation T such that the system (2) becomes: Zl .O(t)) d~=l_B.U is a stable design matrix.) and the associated nonnegative bounds K~ and Kn are assumed known and finite. to which the uncertainty F belongs.. The uncertain matrices stem from the transformation of F(. the nullspace dynamics reduce to z l ( t ) = L'zl (t). It is assumed that F(. The subsystem (3a) defines the nullspace dynamics and the subsystem (3b) represents the rangespace dynamics. . The linear controller part uC(zl(t). where Ira is the (m x rn) identity matrix. which corresponds to the system attaining an ideal sliding mode. u E IRm and the known matrix pair (A. This control has two parts: = (t).On the Discretization of Sfidingmodelike Controllers 483 where z E ~'*. In the absence of any uncertainty in the nullspace dynamics and if ~(t) = 0.. r + r (5) where uc(.) and UCNL(..1 (Ozl(t) + (12. a Lyapunov function V2(t) for the analysis of the rangespace dynamics is V2(f[) def = 1r (QP~q~(Q. Under certain assumptions on the class of functions. The unknown function F(.
(Pr x > 7. positive design dependent constant.f ((1 . (10) where 71 > 1.{P~ITT+:FP~I}) ).for [zT cT]T = 0 achieves robustness by counteracting the matched uncertainties and ensuring reachability of a cone shaped layer around the sliding mode..72)12" t.0 .. which is defined by the relation V2 = w2V1. (11) (12) (13) ~/2 de__/sup (llAell) +. VI(t) d~J lzlr(t)plzl(t) with P1L" + LerP1 = (8) for the nullspace dynamics.. The constant w de] Ama~(P~)d/Arnin(P1 ~) 1 ! has to be positive.ni. Ate T de_.~211P1A12P~ltI+~IIP[89 II lIP2~ II.484 and Guido Herrmann et al. I(nm) is appropriate The nonlinear control component ~(zl(t)'ck(t))B~aPz4P(t) def I JJP2~(t)JJ+6(JJz1(t)JJ+JJO(t)JJ) for [ZT ~T] T ~ 0 (9) UCIVL l (I~)~Cj~(I~) = (Z ) o . if V2 > r holds. ~(t)) d J ~r (1/1 IIP2r + '1211zl(t)ll). Hence.  j i' which guarantees that w remains positive. then the states are outside the cone shaped layer. r is de ~0(Z1 (t).A n ) .. . The constant 6 E ~ + is a design parameter which is implicitly bounded by the maximum size of the cone shaped boundary layer. 1 _> "/2 > 0. The function fined as: e(zl(t). which can be ensured by imposing the constraint j.. The gains rh and 02 in (10) have been defined so that they ensure robustness with respect to the matched disturbances and reachability of the cone shaped sliding mode layer ~/1 = max de't(( sup ~ \An IA. The expression a(llza(t)ll + IIr results in the cone shaped layer around ~b = 0. A. where w E ]R+ is a small.
Due to this characteristic. Note the contrast to [3] where relations similar to (1) need to hold and therefore the upper bound on the sampling time v remains finite for any control even if the control is 0 and the system is openloop stable. i" + +oo.. 1.e. it is easily possible to use a theorem of [9] for proving stability of the continuous time control. it is possible to apply techniques from L2gain analysis to prove that the discrete controlled system is asymptotically stable for a sufficiently small sampling time: T h e o r e m 1 There exists a ~" small enough such that for any 7" < ~. the discretized control (14) stabilizes the system (3a3b) asymptotically.(9).ti}l] : u(t) ~ = u~(t. The proof of Theorem 1 is sketched below. it is also possible to adjust the parameter 1 > 72 > 0 for a compromise of using linear control for reachability (7~ = 1) or robustness (72 r 0). A proof of stability [9] makes it necessary to show that the null space Lyapunov function V~(t) in (8) is decreasing as soon as the states have entered the cone shaped layer. The L2norm of [ZT1 ~T] T for the discrete controlled system is finite.On the Discretization of Slidingmodelike Controllers 485 However.zI. r of (56) and (9) is globally Lipschitz [3] with Lipschitz constant/C~.~) (zl + uCL (z 1 4. Since the control presented here is a modification of [9]. To symbolize the change in methodology from continuous to discretetime. it can be shown that V~ w2V1 will become ultimately smaller than an arbitrary e > 0 (slidingmodelike performance). 3 Discretizing the SlidingModeLike Control Using the L2Gain Approach Instead of the continuoustime control uC(zl. Further.Zl. . 3. Proof 1 S t e p I: The existence of solution of (2) has been proved by [3] for the time . i. 2.(ti)). An implicit bound on w will be given in Sect. uncertain system (3a3b): Vt E (tl. a timediscretized control may be applied to the linear. The constraints on the sampling time v vanishes. (14) u(t0) = 0.. where the sampling instants ti . 3. are defined by a constant sampling time r. q~) + UCNL(Zl.~)+ UN~rL(t.. d~)+uCL(zl. for a control with Lipschitz constant ICu + +0 and an exponentially stable system (3a3b}. d~) (6). by imposing a quadratic stability constraint which will limit the choice of w with an upper bound. the superscript ~(discrete) has been introduced. i .iv.0. Observe that the slidingmodelike control uCL(zl..
T C where ) iiP24. ( t ) ) + kv2(~(t)) f o r v2 ___~. (k~ ~ + 1)Vl(zl(t))) _ [ v l ( z . dp(t)) + const. Thus. 4~(t)) = max (Vl(zl(t)) + kV~(dp(t)). Step II: A global Lyapunov function can be provided by nonsmooth theory using a maxLyapunov function as described in [3. The rived from (3a3b) using the continuoustime control (6). This integrability feature of (/(zl(t). 3 > 0 for almost all t.2] Ac[O@O ] Af[~ ~] F) [~1] . the equality f t f/(~l(s).4]: V(zx(t). This implies exponentially fast decay of V(zl(t).a(t)ll. S t e p I I I : The differential system d~(t)) and il = S ~ I + All'b. exponential stability of the continuoustime controlled system. dp(s))ds = (V(zl(t).(9). Further.for V2 <_~2V1 (15) where k > 0 has to be determined [3].]. (k~ 2 + 1)Vl(zx(t)) . Further.11] that the nonsmooth maxLyapunov function satisfies for the continuoustime controlled system the relation f" < ~v. (16) = [(o + + + + + (17) system is deObserve that input u is of may be intro with input perturbation fi E II~m is now investigated. oo) implying that any signal occurring in the system is Lebesgue integrable and has finite L2norm on the finite horizon T < oo.) applies.486 Guido Herrmann et al. interval [0. it can be shown similar to [3.~v1 . the auxiliary output for (1617) duced: y ~[fA1.C~(t))Pz I [z T *T]T r 0 .(t)li. It has been shown in [3] that this function is differentiable for almost all t since the maxLyapunov function is globally Lipschitz in zl and ~ and the solution zl (t) and ~(t) is absolutely continuous in t [11]. the system (1617) has a Caratheodorytype solution if the Caratheodory type. dp(t)) allows the nonsmooth maxLyapunov function to be used in a similar fashion to the usual Lyapunov function.Hi0(t)lD for (18. f e(*l(t).M(ll.
_ k~+~ ~'~ o [~(t) ~P~m. the value of 7 should be chosen as small as possible . Similar arguments need to be made for k > 0.~(t)) is an upper bound of the norm of the uncertainty. ~(t)) from (15) and the respective subfunctions V1 in (7) and V1 + kV2 in (78) shall be investigated with respect to the newly introduced states z l and ~b.~.I : i ~ ~ d  . T a n d should not be chosen too large to ensure that the principal minors of ~1 are negative semidefinite for any value of C.~.j L~ + ~ Ilull ~ .I as well as (12")TP2 + P~[2* = . As a result. the maxLyapunov function V($1 (t).~. b ( t ) ) ) 71 ~<O. a constant 7 shall be derived determining the inputoutput characteristics from fi to . For that reason. bounded and it is permissible to define the stable matrix Thus. ~ __~ = /4. the uncertainty is compensated for V2 > w2V1 [9]. Since e(~l(t). the derivative I71(~1 (t)) + kV2(~(t)) satisfies for almost all t considering a perturbation and LCP1 + PlY' = .(t) / AL~._ . it will be seen that 7 is used similar to an L2gain... it can be shown for V~ > w2V1 ^r P2(B2UNL(Zl(t).O(V~(~(t) + kV~(r ~ . . (t). In this case. /~(t ) _~.~t~(t)) + c e ( ~ . 72 1 09) The scalars e > 0.~ I1~11~ . I1~11is /7* . Thus.2i.~.~ L~ J [o ~ " '~ .~=~P~ " The output matrix C is partitioned as follows: = In the following. 0 > 0 and 7 _~ 0 have been introduced to ensure t h a t the 1 symmetric matrix ~1 is negative semidefinite ff'l _< 0.On the Discretization of Slidingmodelike Controllers Note that the scalar r derived from (9) and (10) is bounded 0 _< 487 Ir _< &. Further. In similarity to [9].
the subfunctions V1 and V1 + kV2 of the maxLyapunov functions V(Y~l(t). Suppose ~ > 0. ca > 0 and e > 0 the inequality ~ < 0 holds when V2( ~ ( t ) ) .488 Guido Herrmalm et al.w 2 I/1 (~ 1(t)) < 0 is satisfied. it holds for almost all t: ?J=2 14. According to the Sprocedure in [12. ~ > 0. If ~ > 0.+ J . As mentioned ealier. the latter is satisfied if and only if there exists a r > 0 so that ~'~\[v~(~.(0 OV~ i J (kca2 + 1)e~ 11.5112 I (21) Thus. it follows for almost all t: < OVI (kw2+ 1)e2 11. the derivative I71 is investigated. the scalar 3' will be used similar to an L2gain.+.) v~(~l)) < o ~ for all ~b E ~ " and zl E IR(nm).3].911" 1 14. For II2 <_ca2V1.(0 J (20) Now it is of interest to determine for which ~ > 0. Sect. Assuming ~2 < 0.)p(t)) (15) have been investigated with respect to their time derivative (19). so that ~l ~ 0 for any value of C.LAT2p' + ~2 . 2. k.". V2(~(t)) w2Vl(f~l(t)) < O.(21) at those points where they define the maxLyapunov function (15). w.6. then it is possible using nonsmooth analysis to show that for almost all t the derivative iT(~l(t). This is equivalent to finding a r > 0 so that: T [. 3' and e have been determined so that the matrices ~1 and ~ are negative semidefinite. w > 0 and ~ are chosen small enough then there always exists a r > 0 so that ~ < 0. ~5(t)) .
ll ~ (22) As discussed in Step I with respect to nonsmooth theory and Lebesgue integration.I.I))2 II~ll 2ds +lIB211 Ilall 2 d~ +~:~/~2. I.dp(s))ds (23) Step IV: The system dynamics (1617) may be rewritten in terms of the signal ~ in (18) and an auxiliary output ~: (24) The signal u~(t. (25) .(kw2+ 1)0) ~ t : V(~l(s).V(~l(to). ~(t)) + ~ Ilall 2 .. which is now a function of [~:~(t) ~T(t)]T. ~:a.l min(O. ~(t)) < .~ for the continuous control and the relation ~. dp(t . St.y2 Ila(s)ll 2 ds ~2~ t ~ II~(s)ll~d.On the Discretization of Slidingmodellke Controllers satisfies: 489 72 1 V(zl (0. the dynamics (24) have for any finite horizon T < oo. as discussed in the Appendix. Considering Y+ [0]u B2 as the input signal. Z) 11'511 ds+llB~ll I1~11 ds +JC. Z.Z ~ < 2 /C~(Tax(F. I) to (34).7~ It.ZA_I(F. an inequality can be implied from (22) for any T: V(~I (t = T). Thus.T)) .. ~)+U~VL(t. ~(to)) < ~. it holds using the Lipschitz constant /t'. a finite L2gain applying the linear system (F.=llakl_< ~ 4 E k4 alakl 2: 4 = T 2 f~o Ila(s)ll ds </C..min(0. ~). (kw 2 + 1)~)V(~1 (0. ~) +"gL(~l. ~) results as in (14) from a sample and hold proc~s operating on the nonlinear control "~(~1.
V(ih (s). ~(t = T)) . Hence by (26). Note that for 7 ~ . The resulting inequality can now be added to (23) and equivalently rewritten so that for any T: V(~I (t = T).I)) =0. ~(s))ds ftT [[.I. I) and fl = B(F. ~(to)) < .I. T) holds: 7"+0 lim(%a_z(F.1))211B2112 < 1. I. 1. (s))as (29) 1 724]C~fl 2 4 r 1 . I.490 Guido Herrmann et al.(s)l[ ~ ds) )~to II.~_I(F.~(~)ll 2 ds (27) 1 72 ( f T 11/8(s)112ds+e2 +~ I .4 ~ ( % a _ / ( F . I. 1)) 2 lIB2112 < 1. (k~ 2 + 1 ) 0 ) / ] V(~(s).~_z(g. I.1 724K'~fl2 1 ( 4v2]C~(7~I(F'I'I))2 1 +r 1 .T)) =0. 1))2 lIB2112 The perturbation/~ may now be chosen to be u = fi. li~(g(F. if the sampling frequency has been chosen small enough so that 2 4/C.I ( F .4/C~(Tzlz(F. By (26).~_z(F.min(~. Both sides of the inequality may be squared and multiplied by the constant: 72 r _ 4JC~(~.min(~.4 ~ $ ( 7 ~ _ z ( F . 1. I.4/C~(TA_/(F.. 1)) 2 lIB2112 . 1. I. I.~z(F. S t e p V: Consider now the inequality (25).. 1)) 2 IIB21I 2 1 . 1))211B2112  1 (28) then it follows from (27): v(~l(t = T). 1))2 lIB2112 .. which makes the system (1617) and (24) equivalent to the discrete controlled system (3a3b). (k~o2 + 1)t~) 6(to)) f. 1))2 IIB2 II2) ' which has to be positive. 7 ~/1  21C. (26) as discussed in the Appendix with respect to the theory of [6]. this is satisfied if the sampling time r has been chosen small enough so that 2/g~(7. I) < 4/C~(Tz~z(F. 1.V(~l(to). _< .%.[i:T(to) .I.kT (to)]T. [~T(t0) ~T(I0)]T. ~(t = T)) .(7. I.V(~z(to).
5(sinc(0(t)) .5. where 0 is the angle of rotation. The system may be expressed in the form of (2) with a=[ 0 1 ] 0. 9 4 Examples of Discretized Controllers SlidingMode Based The L2gain technique is now demonstrated and compared to the Lyapunov function based discretization methodology of [4] using both the inverted pendulum system and a stable linear plant. 0.11. I)) 2 lIB2112 It can be easily derived from (15) that there exists a constant nonsmooth maxLyapunov function.0+ fl = 0 and limr~0+ 7z1I = O.4KY~(%a_l (F. asymptotic stability of the closed loop system follows. (kw 2 + 1)0) V(}l(s). which is formed by a light rod and a heavy mass attached to one end of the rod with the pivot of rotation at the other end: 0(t) = 0.l) . This would be satisfied if there is no control necessary to stabilize the system. sin(0(t))  bg(t) + u(t). .1 ' B= [~] . it follows min(0.On the Discretization of Slidingmodelike Controllers 491 Observe that the right hand side of (25) has the value 0 and the constraint on 7" in (28) vanishes if the Lipschitz constant )Uu is zero. 1 ] (31) where sinc(t) = ift = 0 " It is also of interest to investigate the discretization approaches on a stable model in the nominal case when the uncertainty is set to zero and 02 r O. I. consider a simple model of the inverted pendulum.b + 0 . Subsequently.50. First. ~(t0)) + e2 1 . Note that limr.09 < b < 0. dp(s))ds 1 724K~ 2 < V(~l(t0). b the damping coefficient and u(t) the control torque. so that: (30) cv for the This implies that the state vector [~T ~pT]ThasfiniteL2_norm" Thus. [ A s An] = [0.
The constant k from the maxLyapunov function (15) has been adjusted via numerical methods so that the sampling frequency bound 1/~for each controller is minimal. [Ao An]=0. the . By demanding 71 = 0 (11). Further.. is only used for reachability of the slidingmode. the system satisfies in this case: A= [ 0 0"51.0 .b + 0. second order systems. Hence. this constraint is necessary and sufficient so that for any state satisfying V2 < w2V1 the derivative I~'l holds for almost all t: V1 < . 8 5 ]zl] 2.0 .2]P1A12P~ ~ 10~lPl] = 0. (32) The nonlinear controllers for both systems are chosen to satisfy 71 = 1.6. the linear controller is used to compensate for the parametric uncertainty ( . the value of 0 (19) is chosen to satisfy 0. 1) for the L2norm based technique are more conservative than those calculated for the Lyapunov function analysis based approach for most choices of Z' = 12". the linear control matrix 12" (6) and the respective Lyapunov matrices P1 (8) and P2 (7) have been set. and d (20) holds 0.1) of [2.4]. Thus. the nonlinear control gain. for the L2based analysis. This ensures consistency with former results [4] and comparability of the control of the nominal and the uncertain systems and implies stable performance for both examples (20). Therefore assuming 7}1 = 0. 5 ] ' B = [~]. Thus.85/2 = P ~ . then the features of the slidingmode cone and the parameter w can be determined using nominal system and controller parameters only: Araa~ (PI. Provided the matrix L" (4) determining the slidingmode performance.572 > P2O. This constant has been set to 0. 3 is as small as possible since it is used in a similar manner to an L2gain. 1/1 = 0. the value of 72 (0 < 72 _< 1) must be chosen as large as possible. Using the Matlab toolbox. (18).492 Guido Herrmann et al.U + 27Tp1) .5 > P10 and 0. A similar result applies to the stable system with no uncertainties (32) where the nonlinear control is only used for reachability of the slidingmode region.85 dl For the considered single input. The reason for the more conservative results is connected to the conservative estimate of the L2gain 7 with respect to the output ~ (1617). There a constant ~. so that the linear control is also used for reachability. Note that the results (Fig. it is necessary to ensure that 7 > 0 in the analysis of Sect. 0 < ~ < 1 is used. The sampling frequency which is sufficient to stabilize robustly the unstable inverted pendulum (31) is generally for the L2norm based methodology more than ten times higher than for the Lyapunov function based technique.5. this choice determines a tradeoff between low sampling frequency and controller performance. For the Lyapunov function based analysis technique a similar tradeoff needs to be made [3.
. .... V a l u e s of s u f f i c i e n t s a m p l i n g f r e q u e n c i e s a g a i n s t c h a n g i n g c o n t r o l l e r p o l e s ......z (ti)lJ) (1) [3].t. . ... ... . ... ... ... . . . . . ! . ... ..... ... the very large value of the Lipschitz constant of the nonlinear control u~vL (9) could be compensated by the upper estimate of the sampling error suptor ~.. . ~2000 . . i. .. which is not straightforward for the L2norm technique. . . An advantage of this analysis technique in comparison to a Lyapunov function analysis based technique is that the samplingtime constraint vanishes for a stable plant if no control is used..+l] ( l l z ( t . : . .... . i .... ! .. .. i . .. Herrmann from the European Commission (TMRgrant.. . Furthermore in the Lyapunov function based approach. .U = 1~* u s i n g t h e L ~ .) .... . .. . .. an example of a robustly controlled inverted pendulum has shown that the presented methodology might have disadvantages when compared to the Lyapunov function analysis based technique: the results m a y be too conservative.... . i. ... 6 Acknowledgements The authors would like to acknowledge the support for G.. . .... . . .. 3oo . ! : 1000 [ o o~ . .. 5 Conclusion This work has applied tools known from nonlinear L2gain analysis and linear optimal sampleddata control to a nonlinear slidingmodelike control.. _. However..~j. ~ . It has been seen that these tools can be used to show asymptotic stability of the discretized nonlinear control9 A nonsmooth Lyapunov function has been used to derive a differential dissipation inequality and subsequently L~stability of the discretized control. ..5" tl'_~ (a) unstable inverted pendulum (31) t'l*=s (b) stable system (32) F i g ...... . i .... ~. ... 2S00 . project number: FMBICT983463)..On the Discretization of Slidingmodelike Controllers 493 3500 . 1.. .n o r m b a s e d a p p r o a c h (line) a n d t h e L y a p u n o v f u n c t i o n a n a l y s i s b a s e d a p p r o a c h (dashed) matrix ~ T ~ was approximated by a polytopic set. .. . i. ..
r IIBII IlCll e~llAII.3.T) <_ (.C.B.ti+l] : y~a(t. z(O). B. (Odt + . r d.CeA(ta)B . Note that f o Ile'A II d. finite dimensional. an inputoutput map of u 4 y. z(0). A similar argument also shows that lim~~0+/~ = O.0 T sup ~e(0. C.~) (]lCeAt(I .B. B.2. B. (p E [1. C) e~ : . i = 0 . y) = O.e(o.3. Appendix: LpGain of SampledData LTI Systems Suppose a linear. strictly proper system G. T) is bounded above by: 1 (34) Z(A.r) te[0. is given by the triple (A. A sampleandhold element may be introduced. The Lpgain of the inputoutput map u + (yZ~ _ y) is bounded from above by: "rA_I(A.3] that the inputoutput map u + (yZa _ y ) has finite Lpgain.. ti+l] for constant sampling time v>O: Vt E (ti. where A E IRnxn is a stable matrix. T] for the in/finite horizon (T< oo): (f0T[lye(0 y(0[['dr) +fl(A. .z(O).. T).y) = y(tl). y ~ (t = o. 1. Theorem 9. The bias term fi(A. ti=iT". It can be seen that the upper limit of the Lpgain ( f o fH(t)dt + Tr converges to 0 for r + 0 using the following upper estimates of fH and the term r /H(O < IIBII IlCll Ile"ll e'"A" 1 .C). It was proved by [6..eAa)z(O)[[ p) dt Note the dependence of/3 on the intial state a~(0) of the linear system.494 Guido Herrmann et al. This implies for any Lebesgueintegrable input signal u(t) E Lp[0.~) C e A t B . measuring the output signal y(t) at welldefined time instants ti and holding the output of the element constant over the interval (ti. C.r satisfy: (33) where the function fH and the term r fH(t) = sup .__y sup (liCe AtBII) and I1" II is the induced Euclidean norm. oo]). is finite for stable A.
New York 2. Feron E. Khalil H. (1999) Discretisation of sliding mode based control schemes. France 6. EIGhaoui L.. IEEE Transactions on Automatic Control. Boyd S. Department of Engineering 12. In Proceedings of the 38th Conference on Decision and Control.Corless. Ryan. 1:223242 11. New York . slidingmodelike control using nonsmooth analysis. K. Keller J. Communications and Control Engineering Series (CCES). In Proceedings of the Fifth International Workshop on Variable Structure Systems. Itkis U. 40(9):16471650 8. Edwards C. (1999) Contributions to discretetime nonlinear control using sliding mode control approaches.. Internal Second Year PhDResearch Report. Herrmann G.. Edwards C. (1992) Nonlinear Systems. (1998) Stabilization and regulation of a class of nonlinear singularly perturbed system with discretized composite feedback.. (1984) Ultimate boundedness and asymptotic stability of a class of uncertain dynamical systems via continuous and discontinuous feedback control.. (2000) Discretization of a nonlinear. P. Society for Industrial and Applied Mathematics. (1995) Inputoutput stability of sampleddata linear timevarying systems. (1999) A new approach to discretization applied to a continuous nonlinear. E. (1998) A new nonlinear. 5056 10.. In Proceedings of the Fourteenth International Symposium on Mathematical Theory of Networks and Systems. K.. (1994) Linear Matrix Inequalities in System and Control Theory. Phoenix. continuoustime control law using slidingmode control approaches. 29(4):419434 3. IEEE Transactions on Automatic Control.. D. Djenoune S. M. (1995) Optimal SampledData Control Systems.. John Wiley and Sons. Florida. continuoustime control law with small control delays . K. Chen T. O. K. 42574262 5. (1992) A new approach to the discretization of continuoustime controllers. Philadelphia 13. SpringerVerlag.. Spurgeon S. Macmillan Publishing Company.. Leicester University... Francis B.. (1976) Control systems of variable structure. Edwards C. Herrmann G.. Spurgeon S.a theoretical treatment. Balakrishnan V. Zerhouni N. Spurgeon S. Anderson B. Long Boat Key.. Herrmann G. lglesias P. Edwards C. EIMoudni A.On the Discretization of Slidingmodelike Controllers 495 References 1. Herrmann G. Herrmann G. Spurgeon S. Accepted in IMA Journal of Mathematical Control and Information 4. London 7. 37(214223):12411243 9. International Journal of Systems Science. 2nd edn. K. IMA Journal of Mathematical Control and Information.
m o s t of these systems contain nonlinearities. Some p a r a m e t e r s of the linear p a r t o f the system m a y also be u n k n o w n . A n y p l a n t c o n t a i n i n g these unknown linear p a r a m e t e r s a n d n o n l i n e a r i t i e s c a n n o t be controlled opt i m a l l y by a linear controller. M a r t i n Rau. spring characteristics.g. 1 Introduction Many technical p l a n t s in the field of m o t i o n control a n d electrical drive systems are m o d e l e d by linear differential e q u a t i o n s . i n t e g r a t o r s ) a n d the unknown nonlinearity. it is d e s i r a b l e to identify the n o n l i n e a r i t y and the u n k n o w n linear p a r a m e t e r s for m e a n s of i m p r o v i n g the controller design. which are u n k n o w n or p a r t i a l l y unknown. we present an identification method for mechatronic systems consisting of a linear part with unknown parameters and an unknown nonlinearity (systems with an isolated nonlinearity) . e. 21 80333 Muenchen. Based on this identification method we introduce an adaptive state space controller in order to generate an overall linear system behavior. backlash or excentricities. turn. Rau@e i . turn. friction. and Dierk Schr&der Technical University Munich Institute for Electrical Drive Systems Arcisstr.N o n l i n e a r A d a p t i v e S t a t e Space Control for a Class of Nonlinear S y s t e m s w i t h U n k n o w n Parameters C h r i s t i a n Hintz. Germany C h r i s t J a n . d e . the use of prior structural and parameter knowledge and the ability to completely compensate the nonlinearity. c o n t a i n i n g only e l e m e n t a r y o p e r a t o r s (gains. In this paper. sums. Nevertheless. d e A b s t r a c t . In this . Hence. The control concept starts from a nonlinear canonical form and takes advantage of the online identified parameters of the plant.Mart i n . the s y s t e m is m a p p e d into a s t r u c t u r e d recurrent neural network. H i n t z @ e i . The novelty of this approach is the simultaneous identification of the parameters of the linear part and the nonlinearity. 2 Identification with structured recurrent networks S t a r t i n g from the signal flow chart of a m e c h a t r o n i c s y s t e m . A structured recurrent neural network is used to identify the unknown parameters of the known signal flow chart. This should be done online to a u t o m a t e initial identification as well as to cope with p a r a m e t e r drift in the p l a n t .
gains by network weights and the nonlinearity by an additional neural network. :v [k] discrete system state vector.x[k]+b. u input signal. a Radial Basis Function Network (RBF). mapping.u[k]+k. Mapping of signal flow chart elements are the gains of the signal flow chart and the weights of the additional neural network for the nonlinearity.498 Christian Hintz et al. b input coupling vector. z . These parameters are adapted by a gradient based learning algorithm which uses only the system's output error. .AfZ~(z.~[k] (2) I Ac. kc. u + k . A system matrix.N's y[k]=cT. yr y output signal . sums are represented by perceptrons with linear activation. co. e. k coupling vector for static nonlinearity 2~.. Zc continuous time system state vector.z c c (1) It is called system with an isolated nonlinearity [5] because there is only one nonlinearity with a fixed input coupling into the plant. The system in (1) is discretized for the identification purpose by EulerForward and results in: z[k+l]= A. The unknown trainable parameters D K D l> l> i> K D Fig.g. 1. be. The basic signal flow chart elements and their corresponding network elements are depicted in figure 1.) yc = c T . + b . a General Regression Neural Network (GRNN) or a Multi Layer Perceptron Network (MLP). e output coupling vector. integrators by time delays. The system under consideration is of the following form: 1 z c = A c . .
one combination of the parameters w results in e[k] .w dy[k] [k] + 779 dw [k] 9 e [k] > 0 (7) The main task is to calculate d 9 [k]/dw [k]. and only if. In representation (3) the estimated nonlinearity . To find a gradient based learning rule for the unknown parameters .10.~(:$[k]) ~l[k]=dT. In order to minimize the square error cost function E[k] = ~e ~ [k] (6) dE it is necessary to evaluate the gradient ~~.~ [k] (5) The signal y[k] is equal to the signal Yc at time instant k 9h.u[k]+k.g.~[k]+b.Nonlinear Adaptive State Space Control 499 The structured recurrent network is implemented in analogy to the plant's structure in (2): ~[k+l]=h. Most drive systems are unique./). it is convenient to introduce a weight vector w. . This is done in analogy to real time recurrent learning [1. A system is unique in the sense of the given signal flow chart if. If the system is not unique a variety of parameter combinations can result in the same inputoutput behavior for the signal flow chart under consideration. In order to get correct identification results. by a R B F Network or a MLPNetwork with the weights O. where h is the sample time. (4) The operator "vec" combines all columns of a matrix into one vector.~.hfs is identified e. The detailed steps are shown in [3] and in section 4 for a mechatronic drive train.0 for k + cx~. dE [k] dw [k] . it is only possible to identify the real physical parameters if the system is unique.11]. With this gradient it is possible to use the following learning rule: w [k + 1] = w [k]  r]. The adaptation error is defined as: [k] = y [k] . and O.:~[k] (3) The discretization of the continuous system simplifies the development of the learning laws of the proposed structured recurrent neural network. and therefore the identified parameters are identical to the real physical parameters. the following facts have to be observed: 9 Due to the fact that only the output error is used for identification.
.. 9 In m o s t applications.u (8) Or 0 Or I y = Cl C2 .. t h e a l g o r i t h m is not able to find the correct p a r a m e t e r s .. z is necessary.. * If the s y s t e m shows globally i n t e g r a t i n g behavior.8.. . " ~" . if the c o n t r o l l a b i l i t y m a t r i x has full rank. and due to the i n t e g r a t i n g b e h a v i o r t h e error of t h e s t a t e s is n o t able to t e n d to zero. Cn1] J To t r a n s f o r m a system from an a r b i t r a r y s t a t e space description to t h e cont r o l l a b l e canonical form. .. Even if the identification a l g o r i t h m finds the correct p a r a m e t e r s .500 Christian Hintz et al. we want to e x t e n d the linear c o n t r o l l a b l e canonical form to nonlinear systems.. T h e a d v a n t a g e of this special s t a t e space d e s c r i p t i o n is t h a t linear s t a t e space controllers can be designed easily. the p a r a m e t e r s are not c o m p l e t e l y unknown. Since t h e o u t p u t error will never be zero. T h e linear form for S I S O syst e m s is described by equation (8). To overcome this p r o b l e m the a l g o r i t h m has to reset t h e e s t i m a t e d states such. 0 0 1 0 0 1 0 0 .. . all previous i n p u t signals c o n t r i b u t e to the current o u t p u t signal. T h e reset values o f t h e est i m a t e d states are calculated in a n a l o g y to the o b s e r v a b i l i t y m a t r i x for linear systems [2]. 0 0 = 0 0 . because at t h e beginning o f the identification procedure the wrong p a r a m e t e r s lead the s y s t e m ' s s t a t e s t o w a r d s a wrong region. This technique leads to a faster identification result even if the system does not show g l o b a l l y i n t e g r a t i n g behavior.. T h e t r a n s f o r m a t i o n is possible.9 + 1 0 0 1 Otn_ 1 . e.g.g the physical p a r a m e t e r s c a n n o t be outside a certain range. e. there are m a n y canonical forms with different p r o p e r ties for special applications. and eventually the o u t p u t error s t a y s zero. Since the p a r a m e t e r s t e n d towards their true values. . even w i t h the wrong p a r a m e t e r s . A n extension to nonlinear systems is the nonlinear controllable canonical . It is useful to i n t e r r u p t the identification process for a specific p a r a m e t e r if this p a r a m eter tends to leave this range. where the p a r a m e t e r s a i a n d ci are the coefficients of the corresponding transfer function. the s t a t e error t e n d s to zero.9]. Here. t h a t the o u t p u t error becomes zero. . by pole p l a c e m e n t . T h i s is done in a n a l o g y to the i n p u t s t a t e linearization of nonlinear s y s t e m s [4. 3 Linearizing NeuroControl (LNC) For linear systems. it is n o t possible to reduce t h e o u t p u t error to zero. a linear t r a n s f o r m a t i o n ~ = T .
To show the benefit of the proposed control method. u = = 1 ~(s + (s163 9w +al~2+.+anl~.z + . N2u.lx~ + crn2~. . + ~0xl = 7" w (12) The dynamics of the controlled system is linear and can be adjusted by an appropriate choice of the coefficients ai (pole placement or other linear techniques). The parameter 7 can be chosen such. for further studies see [4.. a nonlinear state space controller as defined in (11) is applied to the system in section 4. The calculation of v(m) will not be discussed in this paper. It is assumed. Figure 2 shows the step responses for three different types of controllers. We will apply the following control law with the reference value w to the system. .. N2nt was achieved by the . xn + (~. The nonlinear state space transformation v(me) is only possible if the controllability condition for nonlinear systems as described in [8] holds. it is possible to apply the proposed control concept in (11) in order to compensate the nonlinearity and to gain an overall linear system behavior. If a complete system model is available.l ( ~ ) ) .Tw) (11) This control law compensates all nonlinear effects (fi" and ~) and results in a linear differential equation with ai as coefficients. where Afs is set to zero. is the desired step response of the system in section 4.Nonlinear Adaptive State Space Control form (NCCF). that stationary accuracy is reached for ~ : 0. that the system under consideration is already given in the nonlinear controllable canonical form.8.9]. Its equations for a system of order n are = 1(s u = 501 + 0(s u (9) with = " = " (10) xn ~(~) ~(~) The transformation from an arbitrary nonlinear state space equation with state vector m can be achieved by a nonlinear state transformation which has to be a d i f f e o m o r p h i s m in t h e w h o l e r a n g e o f o p e r a t i o n of the system in order to be able to invert the transformation (~ = v(me) and a~e = v . The linear state space controller was adjusted according to the damping optimum criterion [7].
which are also provided by the proposed identification procedure 9 all state variables of the system which can be provided by an adaptive observer [6] 9 a symbolically calculated control law with specified linear dynamics (~i) The overall structure of the presented linearizing neurocontrol concept is depicted in figure 3. 4 Simulation Example Many subsystems in electromechanical drives can be modeled as m u l t i . Comparision of step responses same linear state space controller.b o d y systems. The first body is the rotor of a motor. ~%. Our sample system is a twobody system coupled by a damped elastic spring. The nonlinearity was not taken into account in the controller design and therefore. The adaptive component of the controller results from the adaptation of the identificator to the unknown nonlinearity and the parameters of the linear part.a. The control concept in equation (11) requires 9 information about the nonlinearity. the second body includes . These steps of the estimated states cause disturbances in the controller. N~). ~ . It is nearly identical to the desired step response N2tin. but the nonlinearity was set to 3. which can be reset. the control result is not optimal. arctan(2. the identified nonlinearity and the identified linear parameters are provided to the control law and the adaptive observer. N2ncc! is the control result with the proposed controller.502 Christian Hintz et al. 2. which is provided by the proposed identification procedure 9 information about the parameters of the linear part of the system. The additional adaptive observer [6] is necessary to avoid the utilization of the estimated states of the identificator. i : Fig. Both.
e. ? . e.Nonlinear Adaptive State Space Control 503 l ~ L II Identification Algorithm w[k] Adaptive Observer ]l" Fig. The objective of the controller design is the speed control of the second body. T = [N. 3. N2 the speed of the load and A~o the torsion angle between the two bodies. arctan(2 9N2) v:[001] CT (13) with the state vector . a friction torque. The input signal u is the active motor torque (torque at the first body). The system's state space description in continuous time is Jl = 0 1 . Linearizing neurocontrol all rotating parts of the load. all prior knowledge about the system's structure is used by setting the appropriate neurons to their known fixed values. we use a speed dependent torque at the second body. The lower part of figure 4 shows the mapping of the continuoustime signal flow chart to a discretetime structured neural network. A~ N~] (14) where N1 is the speed of the motor. e. Mapping of the signal flow chart to a recurrent neural network: gains become network weights and sums become perceptrons. by EulerForward 2. As a nonlinearity. Discretization of the continuoustime signal flow chart with the sample time h.z+ 9 + u ~L f~ J2 9 b [o]~ k . The signal flow description of this system is depicted in the upper part of figure 4. Two steps have to be performed to create the network.g. i. .g. 1.
The detailed structure of neurons ~2. # 1 . a22 and a23 of the system matrix A are fixed network weights. 4. (15) ini are the input signals of neuron j.. but given by the structure of the system. or = i i. oj is the output signal of neuron j. # 7 and # 8 contain linear activation functions. the elements a21. perceptrons ~0. The output of the perceptrons with linear activation functions is calculated according to equation (15). # 5 and ~ 9 is depicted in figure 5. It consists of a neuron with a linear activation function and a one step time delay. since they are not physical parameters. perceptrons #2. # 5 and ~ 9 are one step time delays. ~6. h 1 1 h Wl 1 w2 1 1 9 " Fig. The unknown parameter vector w for the system under consideration is w = [w0 Wl w2 w3 ~ 9 ] T = [1/J1 c d 1/J2 A/'F. #3. # 4 .504 Christian Hintz et al.]T (16) . d ! ' u w0. 5. Mapping of the signal flow chart of a twobody system to a structured neural network In the mapping in figure 4. Detailed structure of the time delay neurons As it can be seen in figure 4. Fig.
009 [ k .21 009 [k .h. e [k]. At the next timesteps.11 Owo  w2 9 At the beginning of the identification process.2] Owo Owo Owo + + wl 9 003 [k .1] Owo 009 [k .N e t w o r k are calculated in the same way as those for the linear parameters.1]/Owo is achieved by evaluating equation (18).1] Owo (19) The recursive calculation of 009 [k .1] Owo 003 [k . Equation (18) shows the necessary steps.1].o0 [k] + 02 [k .11 Owo 005 [k . all gradients are set to zero.1] 002 [k .1] Owo o06 [k . 0 5 05 [k] = o4 [k] (17) oi [k] = w 0 .0~ 005 [k . First. the outputs oj of each neuron have to be calculated (see equation (17)).1] Owo 006 [k .h . T h e o u t p u t signal ~ [k] of the identificator is equal to o9 [k .21 Owo 0o~[k2] Owo + Owo 0o~ [k .1] + w l 03 [k] 03 [k] = 02 [k . h . The a d a p t a t i o n law for the p a r a m e t e r w0 can therefore be written as: w0 [k + 11 = w0 [k] + y .w2 9 .Nonlinear Adaptive State Space Control 505 As an example. o9 [k] : 08 [k] o s [k] = w 3 . the calculation of the a d a p t a t i o n law for the unknown p a r a m e t e r w0 is shown.2] wl 9 Owo + wl 9 Owo Owo 006 [k . This gradient has to be calculated from the neuron's outputs oj and gradients from one time instance before ([k . . the nonlinearity is identified by a Multi Layer Perceptron Network (MLP).21 Owo (18) 003 [k .o0[k_l]+w0. The gradients at the current timestep are calculated recursively by the gradients of the past.1] 02 [k] = ot [k] o0 [k] = u [k] .1] The network weight w0 is a d a p t e d with the term 0(09 [ k .2] Cgwo 007 [k .o6 [k] o~ [k] = o6 [k] + N C ( y [k]) 06 [k] = w 2 .1] .11 [k . T h e gradients for the weights of the M L P .2] Owo 0o2[k1] _h.1] CgWo .2]).09 [k .03 [k] + 05 [k . the gradients of the timesteps before are known and the algorithm can start. 009 [k . o05 [k .2] 005 [k .h o4 [k] = h . In this example.1] .1]  Owo w3"h" 0o7 [k .1])/Owo. o r [k] + 09 [k .1] Owo + 009 [k .
1 Christian Hintz et al.~(q ..... This shoes. the control error tends to zero. . !':i rw4~ Fig... i ..... Identification results for the linear parameter w2 Fig. 9.. Identification results for the linear parameter w3 Figure 11 shows the difference (control error) between a linear reference system (the twobody system without isolated nonlinearity) and the nonlinear system controlled by the proposed control concept (LNC).. As the estimated parameters converge towards their true values..506 4.. oH : : : 9 ou r ~ m i m m ~ m ~m ~o ~ m .. ~ i i ! i .. Identification results for the linear parameter w] ... Simulation Results Figures 6 to 10 show the identification results (linear parameters and nonlinearity) for the twobody system in equation (13)..... 6. 8. f ...w Fig..Io J T... 7..:. that the LNCcontroller is able to compensate the nonlinearity even without exact prior parameter knowledge. Identification results for the linear parameter wo Fig. .
. .C4 4~ . . . .004 ~ 0. . . . In many problems in mechatronic systems. which takes advantage of the identification result. . This is an advantage compared to other nonlinear inputoutput (black box) identification methods. . . . The controlled system shows globally linear behavior. one can usually use fixed values when describing the relation between speed and position. .16 "0"~ . . In the field of motion control.~ ~. . To conclude. The control concept results in an adaptive nonlinear state space controller. e.0. spring constants.I ~ I). .15 o. This includes the possibility to specify parameter regions as well as fixed values to accelerate the identification procedure. If the inputoutput behavior can only be achieved by a single combination of the unknown parameters. 10. the result is unique and the identified parameters are identical to the physical parameters.~ 0. Control error nonlinear characteristic 5 Conclusion The main benefit of the presented identification approach is the ability to include the whole prior knowledge about the system into the structured neural network.~ 0. . '0. identification and control are online methods and can operate simultaneously. especially the gains of the signal flow chart and an approximation of the nonlinearity.. . Therefore it can be used in online applications to identify timevariant parameters. . . . A main property of the method are interpretable identification results. Both. . . we summarize the novelty of this approach: Simultaneous identification of the parameters of the linear part and the nonlinearity. The controller considers the identified parameters and the nonlinearity. .01~ . . . Identification results for the Fig. Simultaneous identification of the parameters of the linear part and the nonlinearity is another feature of the presented method. . ..01 8. .~ 0 0. 11. . .. .~ m 507 0.Nonlinear Adaptive State Space Control O~ lO* c.g. the identification result is unique and therefore this method provides an estimate of the nonmeasurable system states. . . .~ o ~.0~ Fig. momenta of inertia or friction characteristics. interpretable identification results and prior structural and parameter knowledge can be included. .
Hintz.(2000) Combined Identification of Parameters and Nonlinear Chrachteristics based on InputOutput Data. Einffihrung in die Methoden und ihre Anwendung.. Lenz. Regelungstechnik.. Slotine. Williams. SpringerVerlag 5. Interner Bericht. D.(1990) GradientBased Learning Algorithms for Recurrent Connectionist Networks. O. PrenticeHall 9. AddisonWesley . Vol.(1979) Entwurf nichtlinearer. NUCSS909. Zell. TUMilnchen 2. Dissertation. A.(1995) Elektrische Antriebe 2.(1991) Applied Nonlinear Control.. J. Weiping. Isidori. U.. Northeastern University. Boston. Brychcy.J. Tech. L. D. M. TU Miinchen 6. Japan.(1998) Lernf~illige neuronale Beobachter ffir eine Klasse nichtlinearer dynamischer Systeme und ihre Anwendung zur intelligenten Regelung yon Verbrennungsmotoren. Proceedings of 6th International Workshop on Advanced Motion Control (AMC2000) Nagoya. (1992) Regelungstechnik. References 1. T. SchrSder. FSllinger. SpringerVerlag 8. Rau. SchrSder. C. SchrSder. Hfithigbuch Verlag 3.(1994) Simulation Neuronaler Netze. pp. Rep. 393399 10. Oldenbourg Verlag. D.(1989) Nonlinear Control Systems. Sommer. R. Lehrstuhl fiir Elektrische Antriebssysteme.508 Christian Hintz et al. SpringerVerlag 7. R. 175180 4. College of Computer Science.(1997) Vorstrukturierte VeraUgemeinerte Rekurrente Neuronale Netze. A.(1999) Intelligent Observer and Control Design for Nonlinear Systems. MA 11. D. zeitvarianter Systeme durch Polvorgabe. pp. 12. Zipser.
dynamics.nl Abstract. hui jber t S@ClmW. It is shown that the notion of an observer is essential in the problem of how to achieve synchronization between two systems on the basis of partial state measurements of one of the systems. For instance.nijmeijer@tue.O. chaotic systems 1 Introduction Probably one of the earliest detailed accounts on synchronized motion was made by Christiaan Huygens. uk Department of Mechanical Engineering Eindhoven University of Technology P. j.An Observer View on Synchronization Henri J. Synchronization of complex/chaotic systems is reviewed from a dynamical control perspective. synchronization of oscillating physical systems is by t o d a y an i m p o r t a n t subject . Rayleigh describes in his famous treatise "The theory of sound" [29] in 1877 that two organ tubes m a y produce a synchronized sound provided the outlets are close to each other. Box 513 5600 M B Eindhoven. In fact.C. these clocks then exhibit synchronized motion. Huijberts 1 and Henk Nijmeijer 2 1 Department of Engineering Queen Mary and Westfield College Mile End Road London E1 4NS. c. In a short while. rotating bodies. or more generally rotating mechanical structures form a very important and special class of systems that. Early this century the Dutch scientist B. who around 1650 describes in his notebook [19] an experiment where two identical pendulum clocks are attached to the same (flexible) bar. Examples are given to demonstrate the main results. see [36]. with or without the interaction through some coupling. The explanation by Huijgens is remarkably accurate since by that time the differential calculus needed to describe the clocks' motion was still to be developed. An overview of recent results on the design of synchronizing systems using observers is given. Keywords: synchronization. The Netherlands h. exhibit synchronized motion. ac. United Kingdom h. van der Pol. even when initialized at different phases. studied synchronization of certain (electro) mechanical systems. observers. Many other examples of synchronized motion have been described after the 17th century. Actually.
In Section 4. Synchronization of (1) .was probably caused by the paper by Pecora and Carroll [27] where. and f and h are smooth.. are asked to work in synchrony. crx(k) = x(k + 1). Also [26] contains a rich class of motivating and illustrative examples of synchronizing systems. the paper initiates further interest in dynamical control methods in the study of synchronization problems.C. secure communication as a potential application has been indicated. Section 3 explores some of the methods to design nonlinear observers. On the other hand. the paper addresses the synchronization problem from the perspective of (nonlinear) observer design. cf. Typically robot coordination. z) (2) where z E ]Rn and g is smooth. Huijberts and H. the synchronization problem will be introduced and it will be shown that under certain circumstances the synchronization problem can be viewed as an observer problem. either identical or different. An illuminating survey on synchronization of a wide variety of mostly (electro) mechanical systems is given in [3]. we consider a second system that is driven by (4): = g(y. a denotes the timeshift.and the above mentioned surveys are illustrative for this . The growing interest in synchronization . More specifically. i. Although sofar it is still questionable whether this application can be fully realized. [5] and cooperation of manipulators.J. Section 5 contains some conclusions. y E ~q. see [24] form important illustrations of the same goal. the Pecora and Carroll paper has formed an impulse for much research along these lines.e.e. In the following section. static and dynamic methods to design synchronizing systems for nonlinear discretetime systems are discussed.. Besides (4). 2 Synchronization as an Observer Problem In this paper. System (1) is the socalled transmitter (or master). ax(t) = x(t).510 H. For discretetime systems. where it is desired that two or more mechanical systems. while for continuoustime systems cr denotes differentiation. The paper is organized as follows. among others. i. we consider a nonlinear discretetime or continuoustime system of the form h(x) (1) where x E ]Rn. The purpose of this paper is to address the synchronization problem from a control theory perspective. Hopefully. for mechanical systems synchronization is of utmost importance as soon as two machines have to cooperate. and system (2) is the receiver (or slave). Nijmeijer in some of the major physics journals.
b e a + ye~ which is a linear t i m e . z3)dynamics the known signal y = x l . the receiver (2) d e p e n d s on (1) v i a t h e drive signal y = h ( x ) . we consider the Lorenz s y s t e m { ~1 = ~ ( x 2 . e2 .~ e ~ ) 2 . or slave s y s t e m .zl. i.XlX~ T h e system (4) is known to exhibit complex or chaotic m o t i o n s for certain p a r a m e t e r s a .+1oo lim [[x(t) . e3) = (0. T h i s means. Example 1: Following [27]. t . z3) a n d where in the (z2. with s t a t e (zl..z(t)[[ = 0 (3) Typically. e3 = x3 . W i t h the system (4) viewed as t h e transmitter..x2 . no m a t t e r how (1) and (2) are initialized. as in [27].e. ~3) = ~(e~ . which explains the t r a n s m i t t e r / r e c e i v e r terminology. we have t h a t a s y m p t o t i c a l l y their states will m a t c h . or m a s t e r system. We illustrate this definition by m e a n s of an e x a m p l e . O) is s t r a i g h t f o r w a r d l y checked using the L y a p u n o v .z2 . e~. r .z~. to achieve a s y m p t o t i c synchronization.x l ) x2 rxl x2 xlx3 (4) a:3 .~e ~~ _ 2be~ (9) .z3 we o b t a i n the error d y n a m i c s ~1 = ~(e2 .el) ~ ~3 e2yea (7) . O. b > 0.z l ) ?~2 r y .v a r y i n g system. e3) = l d + e~ + d (8) with timederivativealong (7) 1 3 V ( e ~ . see (5). is s u b s t i t u t e d for zl.3 .f u n c t i o n V(el. t h a t we take as receiver d y n a m i c s { { ~1 = ~ ( z 2 . T h e s t a b i l i t y of (el. we introduce the drive signal y = ~1 (5) which can be used at the receiver. e.yz3 7.b x 3 }. Introducing the error variables el = x l . e~.b z 3 + yz2 (6) Notice t h a t (6) consists of a copy of (4).An Observer View on Synchronization 511 and (2) occurs if. z2.
So in those applications where one is able to design (2) freely. for any pair of initial conditions x(0) and z(0). Recall that the dynamics (10) is an observer for (1) if z(t) asymptotically converges to x(t). Nijmeijer showing that (et. Huijberts and H. However. we will explore some of the possibilities to solve the synchronization problem by making use of results from the theory of nonlinear observer design. In most of the synchronization literature (see e. . even if a hardware realization of (1) is given. One m a y view (10) as an identical copy of (1) with an 'innovations' term k(z. e3) asymptotically (and even exponentially!) converges to (0. a system (2) can be designed that synchronizes with (1). there are some important cases where a solution can be found. so that no synchronization will occur in general. given a system (1).512 H.g. The structure of the observer (10) deserves some further attention. it is easily checked that the system (6) indeed acts as an observer and can be put in the form (10): { ~1 = a ( z 2 . the receiver dynamics (6) asymptotically synchronizes with the chaotic transmitter (4) no matter how (4) and (6) are initialized.0. e2.b z 3 + z l z 2 +z~(yzl) Also. y) = 0 if h(z) = h(x) = y. In its full generality the problem of observer design for nonlinear systems is unsolved.y) (10) where k in (10) should be such that k(z.C. The natural way to approach the observer problem for (1) is to design another dynamical system driven by the measurements y = h(x) az=f(z)+k(z. In the Lorenz system (4.J. [27]) the systems (1) and (2) are systems that are given beforehand.z l ) Z: = rzl z~  +0 zlz3+(r za)(v zl) (11) ~a = .0). In other words. This raises the question whether.g.5) with receiver (6). The latter could be phrased as: "we can not do better than our measurements allow for". As was indicated in e. it may be worth noting that (10) is simply a computerized model and no hardware is required in building this system. Some of them will be reviewed next. [25] this problem may be interpreted as viewing (2) as an observer for (1) given the output signal y = h(x). y) which vanishes when the estimated output y : h(z) coincides with y : h(x). the synchronization problem may be tackled as an observer problem. In the following sections.
it follows that synchronizing systems can also be designed quite easily for systems (1) that m a y be transformed into a Lur'e form by means of a change of coordinates ~ = P(x) and an output transformation ~) = p(y).[18]): x l ( k + 1) = x l ( k ) + x z ( k ) 9 ~(k + 1) = ~ ( k ) .[21]. A nonlinear discretetime or continuoustime system in Lur'e form is a system of the form {~= A~ ~ ) C ~ O( + (12) where ~ E ]Rn. and g is the gravitational acceleration.. Suppose that only the phase xl is available for measurement. ~ is the coefficient of restitution and ~ = 2w~(1 + o~)A/g. E x e m p l e 2: Consider the following discretetime model of a bouncing ball ([35]. we have that y = h(x) = Xl (15) .An Observer View on Synchronization 513 3 Observer Based Design of Synchronizing Systems Probably the most successful approach to the observer design for nonlinear systems is to make use of so called Lur'e forms. Conditions in terms of f and h under which continuoustime systems can be transformed into Lur'e form are given in [20]. Here w is the angular frequency of the table oscillation.K C are in the open unit disc (for discretetime systems) or in the open left half plane (for continuoustime systems). xl(k) is the phase of the table at the kth impact.[6].e. ~) E JR. for this kind of systems a synchronizing system is given by f t = Az + Cz + 9) (13) where the matrix K is such that the eigenvalues of A . x~(k) is proportional to the velocity of the ball at the kth impact. A) is observable. A and C are matrices of appropriate dimensions. and 4~ is smooth. We illustrate the above with two examples. For discretetime systems. As may be well known.~cos(~l(k) + ~(k)) (14) In (14). the pair (C. From the above. A is the corresponding amplitude. i. see also [25] where these conditions are given in the context of synchronization. these conditions were given in [23] for the case without output transformations and in [13] for the case where also output transformations are allowed.
.~) ~n b~ni . It is clear t h a t the solutions of (19) with x . T h e case n = 3 corresponds to the usual RSssler system.exp(.e ~ z i ( k ) + e ~ ( k ) . . Therefore.J. see [2]. Nijmeijer Defining new c o o r d i n a t e s { ~i = ~i ~2 (~XI+ X2 + /3COSXi (16) we obtain the following Lur'e form: { t2 ~ l ( k + 1) = ~2(k) + (1 + cOy(k ) . T h e classes of c o n t i n u o u s . d./ 3 c o s y(k) ~2(k + 1) = ... Example { xl : x2 + axl xi x i . sofar a l m o s t all exploit L u r ' e forms.n2) (20) It is then clear t h a t the system (20) is in L u r ' e form.. ( 0 ) > 0 t h a t have no finite escape t i m e satisfy x.d) where a. . ._t . W h e n we also define a new o u t p u t ~ = l n y .C. 9 .. we m a y introduce new c o o r d i n a t e s (i = xi (i = 1. T h e r e are however o t h e r cases .t i m e systems for which a successful observer design is possible.g2 are chosen in such a way t h a t the eigenvalues of t h e m a t r i x are in the open unit disc. n 1) (19) e + bxn(x.~u(k) u(k) = ~l(k) (17) A synchronizing s y s t e m for (17) is then given by { zi(k + 1) = gizi(k) + z~(k) + ely(k) + .1 . b.~ i + l '~n1 ~n2 . (. n .] ) ) (i= 2.514 H.l .bd + e e x p ( . and when n = 4 the s y s t e m has so called h y p e r c h a o t i c flows and has two positive L y a p u n o v exponents.1).~u(k) (1 + a)y(k) /3cosy(k) z2(k + 1) (18) where gl. Hence a synchronizing s y s t e m for (20) (and thus also for (19)) m a y be designed along the lines set o u t above. the s y s t e m in the new c o o r d i n a t e s and with the new o u t p u t has the form 9 ~i = ~ i .~ = l n x n . Huijberts and H.. e > 0 and n _> 3 a r b i t r a r y . 3: Consider the hyperchaotic Rbssler s y s t e m .~ (t) > 0 for all t _> 0.xi+l (i = 2 .
Ljh(x). If both (i) and (ii) hold an observer exists of the form i=f(z)+K(h(x)h(z)) (21) with K a constant suitable (n. [11]). the observer (21) is an interesting alternative for the observer (6). need not to be l~ n. when this is restricted to a compact domain.. though some of the entries in K m a y become very large (highgain). Since it is known that (4) has an attractive compact box. [33]). no general method exists that works for all systems.. Consider the discretetime nonlinear system (1). To that end we return to the continuoustime system (1) and we introduce the following assumptions. and define the mapping r : IRn + IR~'* by (22) .An Observer View on Synchronization 515 where synchronization can be achieved without relying on a 'linearizability' assumption. like bilinear systems. We will first give a brief statement of this theorem. 1)vector (cf. Besides the above discussed cases for which a synchronizing system can be systematically designed we note that there exist further methods that may be applicable for other classes of systems. see [11]. and finally some systems admit a Kalman filterlike observer.g. Here Liyh(x) denotes the ith iterated Liederivative of the function h in the direction of f . (ii) The n functions h(x).5).L'}lh(x) define new coordinates (globally!). L~h(x). But. Also for certain mechanical systems 'physicsbased' observers can be developed.. Note that (21) obviously is of the form (10). 4 Static Synchronization and for Discretetime Systems Static Synchronization Extended Lur'e Forms 4.1 One of the methods that is extensively used in the synchronization literature for the design of synchronizing systems is based on the so called TakensAeyelsSauer Reconstruction Theorem (see e. (i) The vector field f in (1) satisfies a global Lipschitz condition on its domain.. An illustrative example of a system that fulfills (i) and (ii) is formed by the Lorenzsystem (4.
y ( k ) ) . especially when chaotic systems (1) are being considered. which gives that x ( k ) = f ~ " . one needs to repeatedly (numerically) differentiate the transmitted signal y. Thus.l ( x ( k . . ~l(k .1 ) ..N ) ) ~(k) = C{(k) (23) . the state x ( k ) m a y be reconstructed by first reconstructing x ( k .O(k . One possibility to do this for discretetime systems is by making use of so called extended L u r ' e f o r m s . This will be the topic of the next subsection.. the TakensAeyelsSauer Reconstruction Theorem also holds for continuoustime systems. also past output measurements are employed in the synchronizing system. . R e m a r k : When throughout "a" is interpreted as the timederivative. Nijmeijer The TakensAeyelsSauer Reconstruction Theorem ([34].2n + 1). .C. The static reconstruction approach set out above has the drawback that it is very sensitive to measurement errors. More specifically.[1].[30]. there is a natural generalization of (12) that gives more freedom to design synchronizing systems.[32]) then states that for generic f and h. one would need to combine the static reconstruction with some kind of filtering in order to suppress the influence of measurement noise. in order to perform reconstruction based on the TakensAeyelsSauer Reconstruction Theorem for continuoustime systems.[13] are quite restrictive.2n + 1)).2 Extended Lur'e Forms The conditions for the existence of Lur'e forms for discretetime systems given in [23]. In order to cope with this drawback.J.. this implies that there globally exists a mapping r : ~ n _+ Ftn such that = x In other words. one considers discretetime systems in so called extended Lur'e form with buffer N (N = 1. this forms a second drawback of static reconstruction methods for continuoustime systems.. 9 n . 4. Especially when one is dealing with chaotic systems (1).1 ) . Huijberts and H. In this generalization. Besides the sensitivity to measurement errors.516 H. In particular. which takes the observer based design of synchronizing systems in the direction of the static reconstruction based on the TakensAeyelsSauer Reconstruction Theorem. this numerical differentiation is very unreliable.2n + 1) by means of x ( k . These are systems of the form { {(k + 1) = A { ( k ) + 4~(fl(k). For discretetime systems however. the mapping r is an embedding.2n + 1) = ~b+(y(k . . the mapping r is a global leftinverse of r This means that once the mapping r is known.
As is well known. and possibly an o u t p u t transformation ~ = p(y).[15].[14]. .. However. T h e transformations allowed here are so called extended coordinate changes parametrized by y(k . [la] for a definition of strong observability) always m a y be transformed into an extended Lur'e form with buffer N = n . y(k .~(kN))+ (24) ~(k) = Cz(k) where the matrix K is such that the eigenvalues of A . ~ E 1~..1 m a y also be applied to generic discretetime systems (1).1. the d y n a m i c reconstruction approach with buffer N = n ... A and C are matrices of appropriate dimensions. one is now interested in the question whether a given discretetime system m a y be transformed into an extended Lur'e form with a buffer N E { 1 . y(k . . .[22] is t h a t a strongly observable discretetime system (see e.1. (ii) If there are no measurement errors.K C are in the open unit disc.[15]. n .1 ) .. As in the previous section. A) is observable. . . Some remarks on the connection between static synchronization based on the TakensAeyelsSauer Reconstruction T h e o r e m and d y n a m i c synchronization based on extended Lur'e forms with buffer N = n . the static synchronization approach m a y be applied to generic discretetime systems (1). R e m a r k : If one would like to follow the approach of extended Lur'e systems for the design of synchronizing systems for continuoustime systems. As stated above.1 m a y be applied to strongly observable systems (1)..g. the pair (C. for this kind of systems a synchronizing system is given by z ( k + 1) = A z ( k ) + r K ( ~ ( k ) . Conditions for the existence of an extended coordinate change and o u t p u t transformation t h a t transform (1) into an extended Lur'e form with a buffer i E { 1 .1 ) . .1 time steps are needed to initialize the receiver (24). and ~ is smooth. .fl(k)) 1).[22].An Observer View on Synchronization 517 where ~ E ]Rn.[14].N ) of the form ~(k) = P ( x ( k ) . . A remarkable result in [13]. . strong observability is a generic property for systems (1). . one can also achieve exact synchronization after 2n . In the d y n a m i c synchronization approach with buffer N = n . n . y ( k . and by an appropriate choice of the vector K in (24) one can obtain a deadbeat receiver which results in exact synchronization after another n time steps. Analogously to the previous subsection.1} are given in [13]..1}. the static synchronization approach results in exact synchronization after 2n .1 are in order here: (i) As stated in the previous subsection.1 time steps. . .N ) ) .. . it remains an interesting question whether the generic class to which the static synchronization m a y be applied is the same as the generic class to which the d y n a m i c reconstruction approach m a y be applied. and thus the d y n a m i c synchronization approach with buffer N = n .1 time steps: n .
Nijmeijer the extended Lur'e form as well as the synchronizing system would depend on timederivatives of the output y instead of past output measurements.518 H. Pogromsky A. it is felt that nonlinear control may provide some useful tools to address certain synchronization problems.J. [28]). of which one could mention..G. partial synchronization and generalized synchronization.. Synchronization has numerous potential applications running from coordination problems in robotics to mechanisms for secure communications.. However. This includes observers but also further aspects such as filtering (cf. Syst Control Lett 31:299305 . Landa P. (1995) Synchronization and Chaotization in Interacting Dynamical Systems. In the study of synchronization several elements from control theory turn out to be relevant.C. in many cases.I. 1995 4. 5 Conclusions We have tried to give a dynamical control view on synchronization. Precisely the latter area was mentioned in [27] as a potential field of application.Yu. References 1.g. robustness (cf. adaptive control (cf. Baier G. Even for transmitter/receiver dynamics described by partial differential equations one may expect some results along the lines set out in this paper. Aeyels D. see e. Appl Mech Rev 48:733752. [10]). Sahle S. [9]). (1981) Generic Observability of Differentiable Systems. or [4] where a general definition of synchronization is proposed. feedback control (cf. a thorough study of certain timevarying dynamical systems is required and it may be concluded that further research along these lines requires knowledge from both dynamical systems and nonlinear control theory. Huijberts and H. All in all.g. Blekhman l. There are numerous variants of synchronization defined in the literature.. The review as presented here gives only a partial view on synchronization. as said before.S. [16]). On the other hand. Likewise. [7] for a specific example of synchronizing PDE's. system inversion (cf. [17]). synchronization with timedelayed feedback has also been studied in [7]. (1997) On Selfsynchronization and Controlled Synchronization..I. see [26]. (1995) Design of Hyperchaotic Flows. Conditions for the existence of extended Lur'e forms for continuoustime systems have been obtained in e.[31]). SIAM J Control Optimiz 19:595603 2. [8]. Nijmeijer H. Blekhman I. Phys Rev E 51:R2712R2714 3. [12].. Fradkov A.L. although sofar much work remains to be done here. Rosenblum M. or system identification (cf. the fact that one needs to numerically differentiate y makes it questionable whether this approach would work for continuoustime systems. It should be clear that synchronization problems can be treated in other domains too. phase synchronization.
(1999) Zum Entwurf nichtlinearer zeitdiskreter Beobachter mittels Normalformen.. (1996) Communication by Chaotic Signals: The Inverse System Approach.. IEEE Trans Circ Systems I 44:882890 26. Lin W. Diissddorf 23.H. Nijmeijer H. ln: Nijmeijer H. Nijmeijer H. (1997) An Observer Looks at Synchronization. IEEE Trails Automat Control 37:87588O 12. CRC Press.M. Cao L. Moog C.. (1673) Horologium Oscilatorium. Willems R. Judd K.. (1983) Linearization by Output Injection and Nonlinear Observers. Huijberts H. New York. Hammouri H.. WileyVCH. (Eds. Paris. Feldman U. (1999) Control Perspective on Synchronization and the TakensAeyelsSauer Reconstruction Theorem. (1999) Discretetime Observers and Synchronization.. Nijmeijer H.J. Syst Control Lett 25:3140 24. Parlitz U. IEEE Trails Robotics Automat 15:258267 25. Cruz C. Huijberts H.C. submitted to Int J Bifurc Chaos 16.) (1999) Controlling Chaos and Bifurcations in Engineering Systems. Fossen T.I... Krener A. Huijberts H. (2000) System Identification in Communication with Chaotic Systems.. Hasler M. 2?2302 In: Schuster H. Nijmeijer H.D.An Observer View on Synchronization 519 5.Yu. Delft University of Technology. Byrues C.. (Ed. Boca Raton 8. Lilge T..M.C. 272302 . Nijmeijer H. (2000) Nonlinear Discretetime Synchrouization via Extended Observers. Nijmeijer H.C... (1999) On Existence of Extended Observer Forms for Nonlinear Discretetime Systems. Berlin. Krener A. (2000) Synchronization Through Filtering. France 20. H. Chen G... Kocarev L. Mees A. IEEE Trans Circ Systems I: to appear 17. 439455 19. (1997) Targeting Using Global Models Built from Nonstationary Data. Gauthier J. Springer.) New Directions in Nonlinear Observer Design. (1996) New Algebrogeometric Conditions for the Linearization by Inputoutput Injection.. In: [7].. In: [7].J.. (Ed. Pogromsky A. Isidori A.I.C. Lilge T. (1998) Coordination of Redundant Systems.. Xu Y...A. Schwarz W. (1999) Adaptive Observer Based Synchronization. Pogromsky A.Y. Preprint. 7992 14.A. pp. Application to Bioreactors. Glumineau A.J.J. (1985) Nonlinear Observers with Linearizable Error Dynamics.. IEEE Trans Automat Control 41:598603 13. Brunt M.C.J.G. Bergerman M. Mareels I.. (1995) Remarks on Linearization of Discretetime Autonomous Systems and Nonlinear Observer Design. Willems R.L.. Nijmeijer H.M.C. Huijberts H. Othman S.) Handbook of Chaos Control. lnt J Bifurc Chaos: to appear 9. (1992) A Simple Observer for Nonlinear Systems. Int J Circ Theory Appl 24:551579 10. Nijmeijer H... SIAM J Control Optimiz 23:197216 22. (1999) Cooperation Control of Multiple Manipulators with Passive Joints. (2000) Regulation and Controlled Synchronization for Complex Dynamical Systems. VDIVerlag. Thesis. Phys Rev Lett 231:36%372 7.H. Huijberts H. Fradkov A. (1999) Synchronization of Chaotic Systems.. Plestan F. Ph..Yu. 41%438 11.J. Liu Y. Syst Control Lett 3:4752 21. Phys Rev E 59:46914694 15. Int J Robust Nonl Control: to appear 18.J. Huijgens C. Huijberts H. The Netherlands 6.. Respondek W.J. Lilge T.
Nijmeijer H.Yu. J Stat Phys 65:579616 31. (1920) Theory of the Amplitude of Free and Forced Triod Vibration. Thorp J. (Eds.) Dynamical Systems and Turbulence. Sobiski D. (1992) An Experimental Approach to Nonlinear Dynamics and Chaos. Casdagli M..B. Abbott T..J.S. In: D. (1998) Observer Based Robust Synchronization of Dynamical Systems. Springer. (1999) Delay Embedding for Forced Systems... lnt J Bifurc Chaos 8:22432254 29. (1991) Embeddology. Reilly J.A. IEEE Trans Circ Syst I 46:841850 32.C. Kocarev L. Carroll T. Yorke J. Dover Publ. AddisonWesley.L.. J Nonl Sci. MA 36. 9:255332 33. Rayleigh J.. Sauer T. (1945) The Theory of Sound. New York 30. Nijmeijer 27. Deterministic forcing. Parlitz U.M. Hnijberts and H. (1990) Synchronization in Chaotic Systems. Van der Pol B. Reading. Takeus F. (1997) Exploiting Delay Reconstruction for Chaos Synchronization.. Harris R. Rand.520 H. Stojanovski T. Radio Rev 1:701710 ..A. Pogromsky A. Pecora L. Phys Rev Lett 64:821824 28.. Tufillaro N. (1998) PDMA1 Chaotic Communication via the Extended Kalman Filter. Physics Lett A 233:355360 34. I. Young L.J. 366381 35. Stark J.S. (1981) Detecting Strange Attractors in Turbulence. Berlin..
pt Distance A b s t r a c t .R i e m a n n i a n (or C a r n o t . ua. In this paper we make the link between the singular support of x ~~ d(x0. the distance may fail to be subanalytic [8]. a s m o o t h s u b . We already know that this function is not C 1 on any neighborhood of x0 (see [1]) and even if the data are analytic. ~(t) (~(t). T]. V x E M .Regularity of the SubRiemannian and Cut Locus Sdbastien Jacquet Departamento de Matemtica Universidade de Aveiro Campus Santiago 3810 Aveiro. setting d ( z .=T.C a r a t h e o d o r y ) distance d : M• )IRU{+oo}. Portugal j acquet Qmat. x) for a given subRiemannian distance d on a neighborhood of a point x0 of a manifold M.M. T] ) M is admissible if its derivative is t a n g e n t to the d i s t r i b u t i o n A almost everywhere i. In order to state the problem more precisely consider a s m o o t h connected n dimensional manifold M . SubRiemannian distances are obtained by minimizing length of curves whose velocity is constrained to be tangent to a given subbundle of the tangent bundle.e. We say that a H L c u r v e 7 : [0.b u n d l e A of the t a n g e n t b u n d l e T M and a s m o o t h metric tensor g on A. y ) = inf{L(7) : 7 admissible curve j o i n i n g x to y }. ~(t) The length of an admissible curve "7 is given by L(7) = j0 j. the cut locus and the set of points reached from x0 by singular minimizers. We study the regularity properties of the function x ~ d(xo. x). We make the H 6 r m a n d e r a s s u m p t i o n on A that is: Lie(A). ~(t)) dt and we can define the s u b . 1 Introduction. ~(t) (~ for almost every t in [0.
Since the results are local. In this section we give a s u m m a r y of known results on the subject and the definitions needed to state our main results. In the last section we begin a description of the cut locus (Theorem 2). . In this paper we give some results on the structure of the singularities of the value function and of the cut locus (see [3.'/(t))dtelR. ~ gT(t)(x/(t).T].T]) corresponding to each 7 in 7/T. such a u is called the control of 7. ~ d(zo. )7 e "]'/T where 3' is the solution starting at x0 of the control system In x/(t) = ~ u .x0 there are points x r z0 in which f is not continuously differentiable [1].. Xm }. The paper is devoted to study the valuefunction f : x . Under the above assumption d is a distance and as a consequence of " C h o w Raschevski" Theorem. The main difference between Riemannian (i. " '. there is an unique u E L~m([O.4.T]) by means of u E L~([0. A = T M ) and subRiemannian case is that in any neighborhood of/. 2 Minimizers. The possible existence of "abnormal minimizers" make the study of the cut locus more complex.14].2] for the study of the cut locus for contact distributions in lR 3 and [8] for the Martinet "flat" case). Like in the case of Riemannian distance we associate to each curve a s m o o t h "cost". x) in a neighborhood U of x0. being the Xi's independent at each point. d ) induces on M the same topology as the original one.13.522 S~bastien Jacquet  where Lie(A) is the Lie algebra of s m o o t h vector fields generated by C ~ (A). (1) Notice that. T]). g ) is defined by an orthonormal frame {X1. the metric space ( M . defined by E : "leTIT.e. In this case the space ~t/T of all admissible curves defined on [0. in [O. Results about subanalyticity of f have been established in [11.e.18. The aim of the section 3 is to characterize the larger open set where the function f is of class C 1.10. ( t ) X i ( 7 ( t ) ) i=1 a. the main results are stated in T h e o r e m 1 and Corollary 1. The paper is organized as follows : we recall some basic facts about minimizers in section 2. without loss of generality we m a y assume that M is an open connected neighborhood of x0 in IR n and the subRiemannian structure ( A . called energy. T] starting at :co can be identified with an open subset of L~m([0.
A curve that realizes the m i n i m u m of the energy is parametrized by constant speed and realizes also the m i n i m u m of the length. T]. In the case when T : 1 we simply note 7/.e. Such an ~ does not depend on T and i n all t h e p a p e r U will b e e q u a l t o s u c h a n Ue a n d w e d e n o t e a g a i n b y 7/T t h e s e t {u E L2m([0. Then f is given on V by fl~ = Eocr. we get that T = f(3'(1)) .g/(t)) is c o n stant almost everywhere. a(x) is a minimizer going from x0 to x. for all x in Ue it always exists a minimizer going from x0 to x and we have the following compactness result (see [1. We call minimizer an admissible curve that realizes the m i n i m u m of the energy a m o n g all admissible curves defined on the same interval joining two given points in M. T]) with [[u][ < 3Tc is the control of an admissible curve defined on [0.SubRiemannian Distance and Cut Locus 523 By our assumptions E(7) = I[u[] 2. every u e L~([0. 7] with constant speed. Proposition 1. R e m a r k . where [[ul[ is the L2norm of the control of 3'. According to our definition. i. Let K be a compact subset of U then the set of minimizers starting from x0 with endpoint in K is a compact subset of 7/T.14]). it will be more convenient to consider minimizers parametrized by arc length. Let us recall now first order necessary optimal condition. On the other side for every curve that realizes the m i n i m u m of the length there exists a reparametrization of this curve by constant speed so t h a t it becomes a minimizer of the energy.2 and when the structure is analytic for k = 1. and hence f is of class C k on V. minimizers are admissible curves parametrized on some [0. a C k m a p cr defined on V with values in 7/ such that for all z in V. For such an e and for all ~ _< e the subRiemannian ball U6= { x e M is an open set given by U6 : {7(T) : f ( x ) <~} : E(7) < T62}. For e > 0 small enough. Let V be an open subset of U that does not contain x0 and suppose that you have a C k optimal synthesis on V. T]) : I]u]l < To}. For E sufficiently small. In most of the proofs. in this case.V / ~ . .The energy of a curve depends on the parametrization while the length does not and the following inequality holds L(3') _< where the equality occurs if and only if the speed Jg~(t)(~/(t). The main result of this paper will be to prove t h a t this condition is in fact necessary for k . Regularity of f is then related to the problem of finding s m o o t h optimal synthesis and hence parametrization of minimizers.
E(3')) 9 2 1 5 so that 7 is an extremal for 0. Y) = g~( G x~. Y) for every Y in Ax and that the equality (3) is equivalent to Gwq = 2)t0+. e. there is )to 9 ]1% and a nonzero solution q : [0. We call (3'.)t0) is defined up to multiplication by a scalar and we can choose )to = 89 In this case the control of 3' can be reconstructed from (3) and (% q) is a solution of a Hamiltonian system. then 3'(t) = exp(tq0) for every t in [0. i. Such 3' will be called singular.)t = (q(T). Let us remark that Gx~ is the only vector of Ax such that (~. With the above notations the first order necessary condition for optimality can be reformulated as follows: every minimizer can be lifted to an extremal. q. (+.~) the flow of H starting at (x.Xi(x))Xi(x) i=1 m and H(x. .524 S6bastien Jacquet If 3' : [0. by the Pontryagin M a x i m u m Principal. Let :Pezp be the open subset of points q0 in T* M such that H(. In fact if 7 is an admissible curve and q satisfies the above conditions (2) and (3). It is not difficult to see that (% q.~). 1] and we will denote 3' by expq o. 1] and let p : T*M 4 M be the canonical projection. (3) where u is the control of 3'. let us define G~ : T ~ M + A~ and H : T * M + ]a by G~ = Z(~. If )t0 ~: 0 the couple (q. 7] + (]an) * of m q(t) = ..r.Z i=1 ui(t) tDXi (3"(t))q(t) (2) such that ( q . To be more precise. 1 G ~). q0).~) in T * M we shall denote by H(t)(x. T] 4 U is a minimizer then.~) = ~(~. 0) is called abnormal extremal and the corresponding curve 3' is then a singularity of the endpoint mapping. ( 7 ) ) = 2)tou. 89 is an extremal iff (% q) is a solution of the Hamiltonian system associated to H.. i= 1.qo) 9 M. If q0 is in O ~ p and (% q) is the corresponding integral curve of H .)(xo. x . For (x. qo) is Xo defined on [0. The Hamiltonian H is constant along (% q) and we have E(3') = 2H(3'0. )t0) an extremal and q an extremal lift of 3'If )t0 = 0 the extremal (% q... The m a p e : "~~T " + M is called the end point map. we define the exponential m a p by: exp : q o 9 ) poH(1)(xo.)to) is a Lagrange multiplier for the m a p 0=exE : 7ET/T' ) (7(T). q) = H ( % q)...
q0 is a singularity of exp and 7(1) is a singular value of both exp and e. L e r n m a 1.13] by using second order necessary condition for optimality proved in [5]. In [19] every Hamiltonian curve is proved to be locally minimizing. then x belongs to Regoo (y) (resp. It is a particular case of compactness result obtained in [4.expq0 and this q0 is unique. Let k be in IN U { + o o .SubRiemaxmian Distance and Cut Locus 525 It can be easily checked that the space of abnormal lifts is a vector space and that the difference of two Hamiltonian lifts is an abnormal lift. 9 7 is nonsingular. D e f i n i t i o n 1. it is not known if a general strictly abnormal minimizer is smooth or not. 3 Regularity of f. P r o p o s i t i o n 2. In this paper. In this section we describe the set of point in which f is smooth. We first state a sufficient condition for a point to be in Regoo (f). then for each a in/R. 9 ? . In particular if is an abnormal lift of expq0. Since q0 is not a singularity of exp there exist a neighborhood V of x in U and a neighborhood W of q0 in T~oU such that exPl W is a s m o o t h . P r o o f . Let K be a compact subset of Un then the set of q in T* o U such that expq is a minimizer with end point in K is compact. This result comes directly from Proposition 1. In particular. Reg~ (f) if the structure is analytic). w}.eXpqo where qo is not a singularity of exp. On the other hand. exp(q0 + a~(0)) = exp(q0) and hence the restriction of exp to any neighborhood of q0 is non injective. if a minimizer is non singular then there exists q0 such that 7 . We call Singk (f) its complementary. Until know. We call Un the set of point that are reached by x0 only by nonsingular minimizers. the inverse statement is false in general : there exist subRiemannian structures with minimizers that cannot be lifted to a normal extremal that are so called strictly abnormal curves [16]. first order necessary condition (maximum principle or Lagrange multiplier rule) will be enough to prove our results. It comes directly from Proposition 1 that Un is open and we have the following result. the set Regk(f) is defined by x E Regk(f) iff there exists a neighborhood of x such that the restriction of f to this neighborhood is of class C k. Let x be in U such that 9 there exists only one minimizer 7 going from Xo to x.
T h e cut locus will be studied in the next sections. D e f i n i t i o n 2. Let ~a be unique such that (7~. [] R e m a r k . In other words. x is not a "cut point" (see definition below). First remark that for y in V. The cut locus will be denoted by C and points in C are called cut points. ~a(0) belongs to W and exp[ W being injective. Tqo is in W and then eXpTq0 is a minimizer. Define the function fl on V by fl (Y) . 9 there is a singular minimizer going from x0 to x. then by extracting a subsequence we m a y assume that ('ra. 89 tends to some extremal (7. let 7a be a minimizer going from ~0 to ya then using the fact that 7 is the only minimizer going from x0 to z and Proposition 1. fl(y) is the length of the curve exPay where qy = (exPlw ) . Then. singular curves does not exist and the two other conditions are sufficient condition for a point to be in the cut locus.i (y~). expq is a minimizer going from x0 to a point of V.526 S~bastien Jacquet diffeomorphism. then we can construct a sequence {ya} that tends to z such t h a t expq~ is not minimizer where q~ = (exPlw). Suppose that we cannot find V such t h a t fly = fl.~. for a big enough. We proved in fact that we can choose W and V such that for all q in W. In Riemannian Geometry.e. (a(0) = q~ and gives us the contradiction. analytic) and we will prove that we can choose V such that flY = fl. T] strictly bigger than [0. {'ra} tends to 7.~a. nevertheless we will need the following Lemma. (exp Iw )1 (y)). ((0) = q0. * there is a minimizer expq going from x0 to x where q is a singularity of exp. Then a point x is in the cut locus only if at least one of the following condition is satisfied 9 there are more than one minimizer going from x0 to x. 89 is an extremal.l (y) and then fly (Y) ~ fl (Y) and the equality holds if and only if expay is a minimizer. for a big enough. In particular for T > 1 sufficiently close to 1.~/2 H (x0. 1] i. .In particular. 89 ( L e m m a 1) and since 7 is nonsingular. ~ . For each a. 7a is not singular since 7 is not. t ~ exp(tq0) minimizes on some interval [0. A point x in U is not in the cut locus (of x0) if there is a minimizer 7 which goes from x0 to z and is the strict restriction of a minimizer starting from x0. We keep the same notation of the proof of Proposition 2. then fl is smooth (resp.
. then since 3' is a minimizer. P r o o f . Let us suppose t h a t there exists an open subset V of R e g t ( f ) such t h a t IGxdf(x)Iza > 1 for all x in V. A n a l y t i c i t y in this case is used to prove t h a t each piece of a nonsingular analytic curve is non singular and then to prove t h a t the set of conjugate points along a non singular curve expq is discrete.l ( x ) is a s u b m a n i f o l d . Since 7 is nonsingular. Let us now s t u d y necessary conditions for a point to be in Regk ( f ) . Up to now there is no c o u n t e r e x a m p l e to L e m m a 2 in the non analytic case.df(x). We shall use the following result t h a t states t h a t minimizers are integral curve of the "horizontal gradient".~(t) = C~(. Let x be in U such that .)df(.[(T) = x Proposition3. The negative index of F.l ( { e x p ( T q o ) } ) is equal to the dimension of the kernel of Fr ~ t h a t is also equal to the d i m e n s i o n of the kernel of the differential of exp at q0. T h e n if D e x p ( x 0 ) has a non trivial kernel. P r o o f . Let q be in T~oU . eXpTqo fails to be a local minimizer. If we assume the structure to be analytic. vi R e m a r k . 7 i .~ there exists only one minimizer 7 going from xo to x.13] where it is proved the following : for T > 1 sufficiently close to 1 the negative i n d e x of the Hessian at eXPTqo of the restriction of E to e . a point exp(tq) is said to be c o n j u g a t e to x0 along expq if tq is a singularity of exp. it is a critical p o i n t of F and the Hessian F. Suppose that the structure is analytic.~~ of F at 3' (well defined since 7 is a critical p o i n t of F ) has to be positive. " 7 is nonsingular " 7 = eXpqo where qo is a singularity of exp then x is in the cut locus.7. T h e previous L e m m a s t a t e s t h a t in the a n a l y t i c case every nonsingular curve expq stop to m i n i m i z e after its first conjugate point.SubRiemarmian Distance and Cut Locus Lemma 527 2. : = e . T] of .~~ is c o m p u t e d in [6. R e m a r k . Let F be the restriction of E to 7 i . [12]The n o r m IG~df(x)l~x is equal to one f o r all x in R e g l ( f ) and each restriction contained in R e g l ( f ) of a m i n i m i z e r parametrized by arc length starting from xo is an integral curve of x G.~(t) ) 9 . Let x be in V and let T be s m a l l enough such t h a t a solution on [0. x ~ G~df(x) where it exists. we can replace the t h i r d sufficient condition of Proposition 2 by "x is not a cut point" by L e m m a 2.
y(t) (GT(t)df(7(t))./(t)) = 1 for almost t in I. moreover. the previous equality cannot hold and gives us the contradiction. but this last result need some more work. 7(T)) _< f(7(0)) + L(7) and then t" T f ( 7 ( T ) ) . Then IG~df(x)[a = 1 for all x in Regl(f). T h e n [G~df(x)[a < 1. In fact a continuous vector field m a y have more than one integral curve. some open arc length V for all t obtain Let us prove that the equality holds.528 S6bastien Jacquet exists and remains in V.d(x0.f(7(f)) . . 7(t)) = t we g.1 together with I.f(x). Since I~/(t)la is equal to one and IG~(t)df(7(t))la < 1. One could think that the above result implies that every z E Regl(f) can be reached from z0 by an unique minimizer. in principle.(t)df(7(tl)ladt and contradicts the fact that [G~df(x)la > 1 on Y.f(7(0)) < / 0 From (4) and (5) we obtain [G'r(t)df(7(t))['adt" (5) oT [V. 7(0)) + d(7(0).~(t)d. Let 7 be a minimizer parametrized by arclength starting from x0 and I and open interval such t h a t 7(I) C Regl(f). By derivating on I the equality f(7(t)) ./(t)la = 1 implies that "~(f) is equal t o G~f(t)df(7(t. if [G~df(x)[a < 1 on subset V of Regl(f) there exist a minimizer 7 parametrized by starting from x0 and an open interval I such that 7(t) belongs to in I. it could happen that two minimizers intersect themselves and continue together before reaching Regl(f). ~/(t))dt = jot [G~(t)df > (7(t))l~dt.'~(t) ) . ~/(t)) = (df(7(t)) . Two minimizers . From one hand I(7(T))I(7(O)) = [Jo T(df(7(t)). indeed.(4) On the other hand triangular inequality gives us f ( 7 ( T ) ) < d(0. the equality gw(t)(G.)) and then the restriction of 7 to I is an integral curve of x ~~ G~d. Let us now prove that a curve parametrized by arc length contained in Regl (f) is a piece of a minimizer starting from x0 if and only its is an integral of the horizontal gradient.(odf(7(t))[2adt <_ fo T [G.
Let us now prove that a minimizer going from x0 to a point of Regl (f) must be regular. 9 3" is nonsingular 9 z is not in the cut locus.SubRiemannian Distance and Cut Locus 529 m a y coincide in a subinterval if they are not analytic and this could be the case even if the structure is analyti