Author: Bressan A.   Piccoli B.  

Tags: mathematics   control theory  

ISBN: 1-60133-002-2

Year: 2007

Text
                    AIMS Series on Applied Mathematics
Volume 2
Introduction to the
Mathematical Theory of Control
Alberto Bressan and Benedetto Piccoli
With 102 figures and 107 exercises
A 1 M s American Institute of Mathematical Sciences


EDITORIAL COMMITTEE Editor in Chief: Alberto Bressan (USA) Members: H. Mete Soner (Turkey), Eitan Tadmor (USA), Luigi Ambrosio (Italy), Peter S. Constantin (USA). Alberto Bressan Department of Mathematics Penn State University University Park, Pa. 16802 USA E-mail: bressan@math.psu.edu Benedetto Piccoli Istituto per le Applicazioni del Calcolo ’’Mauro Picone” Viale del Policlinico 137 00161 Roma (Italy) E-mail: b.piccoli@iac.cnr.it AMS 2000 subject classifications: 49J15, 49J30, 49N05, 49N25, 93B05, 93B52, 70Q05, 34K35, 35B37 ISBN-10: 1-60133-002-2; ISBN-13: 978-1-60133-002-4 © 2007 by the American Institute of Mathematical Sciences. All rights reserved. This work may not be translated or copied in whole or part without the written permission of the publisher (AIMS, P.O. Box 2604, Springfield, MO 65801-2604, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights. Printed in the United States of America. aimsciences.org

To Wen and Alessia

PREFACE The present book originated from a set of lecture notes developed by the first author, at S.I.S.S.A. and at Penn State University. Its primary aim is to provide an introduction to the mathematical theory of nonlinear control systems. Care has been taken to make the exposition as self-contained as possible. A preliminary chapter covers the basic theory of O.D.E’s with coefficients measurable w.r.t. time, while an extended appendix collects several results from functional analysis, geometry and measure theory. All this background material could be previously found scattered in the literature. Readers will find these results conveniently collected, and supplied with concise proofs. The theory of finite-dimensional, deterministic control has been largely developed in the years 1960-1980’s, and has now reached a “mature” stage. This classical theory makes the content of Chapters 1 to 8. After the introduction and a review of O.D.E. theory, Chapter 3 develops the main concepts and properties of nonlinear control systems. In particular, we discuss the relationship between control systems and differential inclusions, the properties of the control-to-trajectory map, the structure of reachable sets, and various local and global controllability results. The last two sections introduce the notion of chattering controls, and provide a proof of the well known “bang-bang theorem” for linear systems. Chapter 4 contains some basic results on asymptotic stabilization. We first review the method of Lyapunov functions, to analyze the stability of dynamical systems described by O.D.E’s. Then we discuss the existence of a stabilizing feedback control, in the case of a linear control system. The last section contains a result on the local stabilization of non-linear systems, obtained by a linearization method. The next chapters deal with optimal control problems. Chapter 5 intro- duces the main types of optimization problems, with terminal cost and with running cost. We prove here some fundamental results on the existence of optimal controls, under two different types of assumptions: either the system
VIII has linear dynamics, or, in the nonlinear case, we assume that the sets of admissible velocities are convex. The famous Pontryagin Maximum Principle is discussed at length in Chap- ter 6, together with other related necessary conditions for optimality. Proofs of the basic theorems are supplemented by several pictures, together with various examples where optimal controls can be explicitly computed. The following two chapters deal with sufficient conditions for optimal- ity. In Chapter 7 we derive the first order P.D.E. of dynamic programming, describing how the minimal cost changes as a function of the initial state. Assuming the regularity of this “optimal value function”, or the existence of a regular feedback synthesis, one derives conditions which guarantee the global optimality of a given trajectory. Chapter 8 provides a concise introduction to the theory of viscosity so- lutions for Hamilton-Jacobi equations. A key result states that, even with minimal regularity assumptions, the optimal value function can be character- ized as the unique viscosity solution to the appropriate H-J equation. This represents an alternative way to obtain sufficient conditions for optimality. The last two chapters present some recent topics which were never before included in a textbook. These can be used as optional material, providing a flavor of current research. Chapter 9 develops the theory of patchy feedbacks for asymptotic sta- bilization and optimal control. After some basic definitions and examples, we describe the construction of stabilizing feedbacks, prove their robustness w.r.t. inner and outer perturbations, and show the existence of nearly opti- mal patchy feedbacks. Considerable effort was made here in order to polish the exposition, resulting in much shorter and more transparent proofs, compared with the original papers [2], [3], and [4]. Finally, Chapter 10 contains an introduction to the theory of impulsive control of Lagrangian systems. This theory, initiated in the 1980’s indepen- dently by Aldo Bressan [20] and Charles Marie [66], is concerned with me- chanical systems which are controlled not by applying external forces but by implementing some frictionless constraints. In this case, the equations of mo- tion contain also the time derivative of the control function. When this control is discontinuous, the motion thus has an impulsive character. A key issue here is to understand what is the correct dynamics corre- sponding to discontinuous controls, and how to reduce these systems to a more tractable form, such as a differential inclusion. When the derivative of the control enters linearly, under a crucial commutativity assumption one can follow a construction of H. Sussmann [83] and “integrate” the equations, thus eliminating the singularities. On the other hand, without commutativity assumptions, the trajectory de- termined by a discontinuous control t i—> u(t) is not uniquely defined. Indeed, at each time т where и has a jump, one should additionally specify a contin- uous path joining the left and right limits u(t—), u(t+). This leads to the basic concept of “graph completion” introduced in [16], which has now found
IX diverse applications also outside the realm of control theory [37]. In a last section, for systems where the derivative of the control enters also quadrati- cally, we show that the dynamics can be described by a suitable differential inclusion. In the References, we made no attempt to collect the extremely vast lit- erature published on the theory of control during the past 50 years. Our list is thus restricted to some major treatises, and to a small number of seminal papers which may claim historical relevance. An exception is made for the last two chapters, where we quote several recent papers on topics of current research. The book can be used for a one- or two-semesters course on control theory, at a beginning graduate level. Having only one semester at disposal, one can cover the Introduction, Chapters 3-6 (possibly skipping the more advanced sections 3.8-3.10), and sections 7.3-7.4 in Chapter 7. Results from O.D.E. theory given in Chapter 2, as well as the background material collected in the Appendix, can be presented during the course, when needed. Chapters 9 and 10 may provide topics for individual students’ projects, at the end of the semester. The text is supplemented by a large collection of figures, which help the reader understanding the key geometric ideas and building intuition. Several homework problems are listed at the end of each chapter. For science or engineering students, this book provides a richly illustrated overview of the basic techniques and results in the theory of nonlinear control. More mathematically oriented students can use this text as a useful introduc- tion, before tackling more advanced monographs on geometric control theory [1], [56], or the theory of control for infinite-dimensional systems, described by partial differential equations [63], [64], [65].

Contents 1 Introduction....................................................... 1 2 Review of Differential Equations.................................. 13 2.1 Fundamental theory........................................... 14 2.2 Linear systems............................................... 21 2.3 Differentiability with respect to initial data............... 26 2.4 A transversality theorem..................................... 30 Pro blems ...................................................... 32 3 Control Systems................................................... 35 3.1 An equivalent differential inclusion......................... 36 3.2 Fundamental properties of trajectories ...................... 37 3.3 Closure ..................................................... 44 3.4 Density ..................................................... 47 3.5 Reachable sets............................................... 51 3.6 Linear systems............................................... 56 3.7 Local controllability of nonlinear systems................... 59 3.8 Lie brackets and controllability............................. 61 3.9 Chattering controls.......................................... 65 3.10 The Bang-Bang theorem....................................... 67 Prob lems ...................................................... 69 4 Asymptotic stabilization.......................................... 75 4.1 Lyapunov stability........................................... 75 4.2 Stabilization of linear control systems...................... 79 4.3 Stabilization of nonlinear systems........................... 83 Pro blems ...................................................... 85 5 Existence of Optimal Controls..................................... 87 5.1 Mayer problems............................................... 87 5.2 The problem of Bolza......................................... 93
XII Contents Problems....................................................... 95 6 Necessary conditions .......................................... 99 6.1 The Mayer problem with free terminal point................100 6.2 Computation of optimal controls...........................104 6.3 The Mayer problem with terminal constraints...............110 6.4 Variable terminal time....................................115 6.5 The problem of Bolza......................................119 6.6 Linear-quadratic optimal control..........................125 Problems.......................................................127 7 Sufficient Conditions........................................133 7.1 Existence T PMP...........................................134 7.2 Convexity + PMP...........................................135 7.3 Dynamic Programming.......................................137 7.4 Relations between the P.M.P. and the P.D.E. of Dynamic Programming.....................148 7.5 Linear-quadratic case.....................................150 7.6 Optimal syntheses.........................................154 Problems.......................................................161 8 Viscosity solutions for Hamilton-Jacobi equations..............165 8.1 The method of characteristics.............................166 8.2 One-sided differentials...................................170 8.3 Viscosity solutions.......................................174 8.4 Stability properties......................................176 8.5 Comparison theorems.......................................178 8.6 Dynamic programming (revisited)...........................185 8.7 The Hamilton-Jacobi-Bellman equation .....................189 8.8 Infinite horizon problems.................................192 Problems.......................................................197 9 Patchy Feedbacks...............................................199 9.1 Patchy vector fields......................................201 9.2 Asymptotic feedback stabilization.........................206 9.3 Robustness................................................211 9.4 Nearly optimal patchy feedbacks ..........................219 Problems.......................................................229 10 Impulsive Control Systems .....................................233 10.1 Mechanical systems controlled by moving constraints......235 10.2 Generalized trajectories for commuting vector fields.....241 10.3 The non-commutative case: graph completions..............247 10.4 Systems with quadratic impulses..........................252 10.5 Optimization problems for commutative impulsive systems. . . . 257
Contents XIII Problems.....................................................259 A Appendices.....................................................263 A.l Normed spaces............................................263 A.2 Banach’s contraction mapping theorem....................265 A.3 Brouwer’s fixed point theorem ..........................267 A.4 A compactness theorem...................................273 A.5 Review of Lebesgue measure theory.......................274 A.6 Differentiability of Lipschitz continuous functions.....277 A.7 Multifunctions..........................................279 A.8 Convex sets.............................................283 A.9 Convex cones............................................288 A. 10 Lie brackets and Frobenius’ theorem....................294 Problems.....................................................300 References.......................................................305 Index............................................................311

1 Introduction Since the beginnings of Calculus, differential equations have provided an effec- tive mathematical model for a wide variety of physical phenomena. Consider a system whose state can be described by a finite number of real-valued pa- rameters, say x = (xi, ••• If the rate of change x = dx/dt is entirely determined by the state x itself, then the evolution of the system can be modelled by the ordinary differential equation i = g(x'). (1.1) If the state of the system is known at some initial time to, the future behavior for t > to can then be determined by solving a Cauchy problem, consisting of (1.1) together with the initial condition nr(to) = ^o- (1-2) We are here taking a spectator’s point of view: the mathematical model al- lows us to understand a portion of the physical world and predict its future evolution, but we have no means of altering its behavior in any way. Celestial mechanics provides a typical example of this situation. We can accurately cal- culate the orbits of moons and planets and exactly predict time and locations of eclipses, but we cannot change them in the slightest amount. Control theory provides a different paradigm. We now assume the pres- ence of an external agent, i.e. a “controller”, who can actively influence the evolution of the system. This new situation is modelled by a control system, namely x = f(x,u), (1.3) where U is a family of admissible control functions. In this case, the rate of change x(t) depends not only on the state x itself, but also on some external parameters, say и = (ui, • • • , um), which can also vary in time. The control function n(-), subject to some constraints, will be chosen by a controller in order to modify the evolution of the system and achieve certain preassigned
2 1 Introduction goals — steer the system from one state to another, maximize the terminal value of one of the parameters, minimize a certain cost functional, etc... In a standard setting, we are given a set of control values U C lRm. The family of admissible control functions is defined as Z7 = < u : IR i—> IRm ; и measurable, u(t) G U for a.e. t . (1-4) The system (1.1) can then be written as a differential inclusion, namely x G Т’(ж) where the set of possible velocities is given by F(x) = {y; y = f(x, u) for some и G U (1-5) (1.6) Clearly, every admissible trajectory of the control system (1.3) is also a solu- tion of (1.5). Under some regularity assumptions on f, it turns out that the converse is also true: given any absolutely continuous trajectory t »—> x(t) of (1.5), one can select a measurable control function t >—► u(t) G U such that i(t) = /(x(f),u(f)) at almost every time t. Differential inclusions often provide a convenient ap- proach for the analysis of control systems. Fig. 1.1. A differential equation vs. a differential inclusion. Figure 1.1 illustrates the basic difference between an O.D.E and a differen- tial inclusion. In the first case, we have a deterministic model: to each initial state Xq there corresponds one single trajectory t h-> x(t). On the other hand, the evolution described by (1.5) is non-deterministic. Given an initial state #0, several different trajectories t1—► x(t) are possible. Remark 1.1 Differential inclusions are sometimes used as non-deterministic models, when the future behavior of a system cannot be predicted due to lack of information. It should be clear, however, that is not the point of view of
1 Introduction 3 control theory. Here the non-determinacy reflects the possible different strate- gies of a rational controller, who will make his choices in order to achieve a specific goal. The control law can be assigned in two basically different ways. In “open loop” form, as a function of time: t u(t), and in “closed loop” or feedback, as a function of the state: x i—► ufx). Implementing an open loop control и = u(t) is in a sense easier, since the only information needed is provided by a clock, measuring time. On the other hand, to implement a closed loop control и = u[x) one constantly needs to measure the state x of the system. Designing a feedback control, however, yields some distinct advantages. In particular, feedback controls can be more robust in the presence of random perturbations. For example, assume that we seek a control u(-) which steers the system from an initial state P to the origin. If the behavior of the system is exactly described by (1.1), this can be achieved, say, by the open loop control t i—> u(t). In many practical situations, however, the evolution is influenced by additional disturbances which cannot be predicted in advance. The actual behavior of the system will thus be governed by x = f(x,u) + (1.7) where t i—► //(£) is a perturbation term. In this case, if the open loop control и = u(t) steers the system (1.1) to the origin, this same control function may not accomplish the same task in connection with (1.7), when a perturbation is present. In Figure 1.2 (left) the solid line depicts the trajectory of the system (1.1), while the dotted line illustrates a perturbed trajectory x(•). We assumed here that the disturbance tj(-) is active during a small time interval [£i, hs presence puts the system “off course”, so that the origin is never reached. Alternatively, one can solve the problem of steering the system to the origin by means of a closed loop control. In this case, we would seek a control function и = u(x) such that all trajectories of the O.D.E. z = g(x) = f(x,u(x)) (1.8) approach the origin as t —► oo. This approach is less sensitive to the presence of external disturbances. As illustrated in Figure 1.2 (right), in the presence of an external disturbance ??(•), the trajectory of the system does change, but our eventual goal - steering the system to the origin would still be attained. Various examples of control system are described below. Example 1.1 (boat on a river). Consider a river with straight course. Using a set of planar coordinates, assume that it occupies the horizontal strip 5 = {(a? 1,2:2) : — 00 < < 00, — 1 < Х2 < 1}. Moreover, assume that speed of the water is given by the velocity vector v(ii,z2) = (1 - 0).
4 1 Introduction Fig. 1.2. The effect of a perturbation on an open loop and on a feedback control. If a boat on the river is merely dragged along by the current, its position will be determined by the differential equation (±1,±2) = (1 - ^2’ 0). On the other hand, if the boat is powered by an engine, then its motion can be modelled by the control system (a?i,±2) = v + u = (1 - X2 T , u2), (1.9) where the vector u = (1x1,112) describes the velocity of the boat relative to the water. The set U of admissible controls consists of all measurable functions u : IR IR2 taking values inside the closed disc U = < (а?], (J2) • (1-Ю) The constant M accounts for the maximum speed (in any direction) that can be produced by the engine. Given an initial condition (xi,#2)(0) = (±i,t2), solving (1.9) one finds Xi(t) = Xi T t + / ui(s)ds— / {X2+ U2(r)drl ds, Jo Jo \ Jo / #2 (*) = #2 + / 112(5) ds (-1 < X2 < 1). Jo In particular, the constant control u = (1/1,112) = (—2/3,1) takes the boat from a point (xi,— 1) on one side of the river to the point (^i,l) on the opposite side, in two units of time. It is not difficult to show that if M > 0 the boat can be steered from any point P on the river to any other point Q. Observe that for the system (1.9)-(1.10) the admissible trajectories coin- cide with the solutions to the differential inclusion
1 Introduction 5 Fig. 1.3. Velocities of the water and of the boat. (ii,dr2) € F(xi,x2) = < (yi,t/2) : - 1 + ^)2 + y% < M Example 1. 2 (fishery management). Consider a fish population living in a lake. A simple model describing how its size x(t) varies in time is provided by the O.D.E. x = x(a — x). (1-11) Here the constant a describes the maximum sustainable amount of fish which can be present in the lake. Next, assume that some fish is harvested from the lake, at rate и = u(t). For example, one may think of и as the number of fishermen active at time t. In this case, the evolution of the fish population is described by x = x(a — x) — xu. (1.12) This provides another example of a control system. In a realistic situation, one may select the harvesting rate и = u(t) in order to maximize the total amount of fish caught during a given time interval. Notice that if we adopt a constant harvesting rate u(t) = й < о, the fish population will approach the asymptotic limit x = a — й. As t oc, the choice u(t) = a/2 maximizes the average amount of fish caught in unit time. Indeed xu = (a — й\й = max (а— . cv>0 In several situations, the optimal harvesting of natural resources leads to control problems of similar type. Example 1. 3 (cart on a rail). Consider a cart which can move without friction along a straight rail (Figure 1.4). For simplicity, assume that it has unit mass. Let ?/(0) = у be its initial position and г;(0) = v be its initial velocity. If no forces are present, its future position is simply given by
6 1 Introduction y(t) = y-^-vt. Next, assume that a controller is able to push the cart, with an external force и = u(t). The evolution of the system is then determined by the second order equation y(t) = u(t). (1.13) Calling Xi (£) = y(t) and = v(t) respectively the position and the velocity of the cart at time we can rewrite (1.13) as a first order control system: (ii,i2) = (x2,u). (1-14) Given the initial condition a?i(0) = у, .r2(0) = v, solving (114) one finds Assuming that the force satisfies the constraint |u(t)| < 1, the control system (1.14) is equivalent to the differential inclusion (^1,^2) C F(x\,x2) = {(^2,tu); - 1 < w < 1} . Fig. 1.4. A cart moving along a straight, frictionless rail. We now consider the problem of steering the system to the origin. More precisely, we want the cart to be at the origin with zero speed. For example, if the initial condition is (7/, v) = (2,2), this goal is achieved by the open-loop control &(£) — 1 if 0 < t < 4, if 4 < t < 6, if t > 6. A direct computation shows that (j?i (t), a^W) = (0?0) for t > 6. Notice, how- ever, that the above control would not accomplish the same task in connection
1 Introduction 7 with any other initial data (i/, v) different from (2,2). This is a consequence of the backward uniqueness of solutions to the differential equation (1.14). A related problem is that of asymptotic stabilization. In this case, we seek a feedback control function и = u(j?i,j?2) such that, for every initial data (^, v), the corresponding solution of the Cauchy problem (±1,£2) = (^2, и(Х1,Х2У), (.rbz2)(0) = (у, v) approaches the origin as t —> oo, i.e. lim (rri,X2)(t) = (0,0). t—>oc There are several feedback controls which accomplish this task. For example, one can take u(xi,X2) = —x\ — a?2- Because of backward uniqueness, it is clear that there cannot be any Lip- schitz continuous feedback и = u(xi,X2) which steers every initial condition exactly to the origin within finite time. This goal, however, can be accom- plished by the discontinuous feedback law -1 The multifunction «(^l,^) = 1 0 if X2 > 0, jq > or if X2 < 0, Xi > #2/2, if X2 < 0, x\ < x|/2 or if X2 > 0, x\ < -#2/2, if x\ = X2 = 0. (1.15) F(xi,x2) = j (^2,^); w € [-1,1 and the trajectories of the corresponding equation (±i, ±2) = (^2, u(zi,z2)) are shown in figure 1.5. Fig. 1.5. A discontinuous feedback steering every initial point to the origin. Example 1. 4 (car steering). We consider here a mathematical model de- scribing the motion of a car in a large parking lot. At a given time, the position
8 1 Introduction of a car is determined by three scalar parameters: the coordinates (re, y) of its barycenter В € 1R2 and the angle 0 giving its orientation, as in Figure 1.6. The driver controls the motion of the car by acting on the gas pedal and on the steering wheel. The control function thus has two scalar components: speed u(t) of the car and the turning angle a(t). The motion is thus described by the control system ' ii = и cos 0, ±2 = U sin f), (1-16) в -au. It is reasonable here to impose bounds on speed of the car and on the steering angle, say u(t) E [—m, M], o(t) 6 [—a, a]. A frequently encountered problem is the following: given the initial po- sition, steer the car into a parking spot. The typical maneuver needed for parallel parking is illustrated in Figure 1.6. Fig. 1.6. Car parking. In connection with a control system of the general form (1.3), a wide range of mathematical questions can be formulated. A first set of problems is concerned with the dynamics of the system. Given an initial state ж, one would like to determine which other states x e IRn can be reached using the various admissible controls utU. More precisely, given a control function и = u(i), call 11—► a?(Z, u) t he solution to the Cauchy problem ±(t) = /(z(t), ?/(£)), ж(0) = ж, and define the reachable .set at time t as R(t) = {x(t, u); и e Z/}. For general nonlinear systems, explicit formulas describing R(t) are not avail- able. However, one can analyze several topological and geometric properties
1 Introduction 9 of this reachable set. The closure, boundedness, convexity, and the dimension of the set R(t) provide useful information on the control system. In addition, it is interesting to study whether R(t) is a neighborhood of the initial point x for all t > 0. In the positive case, the system is said to be small time locally controllable at x. Another important case is when the union of all reachable sets R(t) as t—->oc includes the entire space IR". We then say that the system is globally controllable. The dependence of the reachable set R(t) on the time t and on the set of controls U is also a subject of investigation. For example, if U is defined by (1.4), one may ask whether the same points in R(f) can be reached by using controls which are piecewise constant, or take values within the set of extreme points of U. Being able to perform the same tasks by means of a smaller set of control functions, easier to implement, is quite relevant in practical applications. Different kind of problems arise in connection with controls in feedback form. Here one basic goal is to construct a feedback control и = u(x) such that the resulting dynamics determined by the differential equation (1.8) has certain desired properties. For example, one could seek a control which steers every initial state asymptotically toward the origin, or stabilizes the system in a neighborhood of a periodic orbit, etc... The regularity of the feedback control is often a major issue of investiga- tion. Ideally, one would like the function x h-> ?i(x) to be smooth, or at least continuous. However, for some nonlinear systems it turns out that certain tasks cannot be accomplished by any continuous feedback law. This raises the question of what kind of discontinuities can be allowed in a feedback control, and how to interpret the solution to the resulting O.D.E. (1.8) when the right hand side is a discontinuous function of the state x. A further key issue related to feedback control is robustness. In general, the differential equation (1.3) provides only an approximate description of reality. External disturbances may affect the evolution of the system. Since these cannot be predicted in advance, it is important to design a control such that the system’s behavior will not be much affected by these small perturbations. Continuous feedback laws are usually robust, but the problem can become quite delicate when discontinuous feedbacks are implemented. A second, very important area of control theory is concerned with opti- mal control. In many applications, among all strategies which accomplish a certain task, one seeks an optimal one, based on a given performance criterion. In mathematical terms, a performance criterion can be defined by an integral functional of the form J= L{t,x,u)dt. (1.17) Jo The value of J will have to be optimized among all admissible trajectories of (1.3), with a number of initial or terminal constraints.
10 1 Introduction For example, among all controls which steer the system from the initial point x to some point on a target set J? at time T, we may seek the one that minimizes the cost functional (1.17). This problem is formulated as min I L(t,x,u)dt (1.18) Jo subject to x = f(t,x, u), z(0) = x, x(T) e ft. (1-19) Observe that if (1.3)-(1.4) takes the simple form x = u, u(t) eV = JRn, (1.20) and if 12 = {?;} consists of just one point, then we do not have any con- straint on the derivative x. Our problem of optimal control thus reduces to the standard problem in the Calculus of Variations: min f L(t,x,x)dt, ж(0) = x, x(T) = у. (1-21) *(•) Jo Roughly speaking, the main difference between the problem (1.18)-(1.19) and (1.21) is that in (1.21) the derivative x is unrestricted, while in (1.18)-( 1.19) it is constrained within the closed set F(x) introduced at (1.6). The basic mathematical theory of optimal control has been concerned with three main issues: (i) Existence of optimal controls. Under a suitable convexity assumption, op- timal solutions can be constructed following the direct method in the Calculus of Variations, i.e., as limits of minimizing sequences, relying on compactness and lower semi-continuity properties. When the convexity condition is not satisfied, the problem usually does not admit any optimal solution. In some special cases, however, the existence of optimal control can still be proved, using a variety of more specialized techniques. (ii) Necessary conditions for the optimality of a control. The ultimate goal of any set of necessary conditions is to isolate a hopefully unique candi- date for the minimum. The major result in this direction is the celebrated Pontryagin Maximum Principle , which extends to control systems the Euler-Lagrange and the Weierstrass necessary conditions for a strong lo- cal minimum in the Calculus of Variations. These first order conditions have been supplemented by several high order conditions, which provide additional information in a number of special cases. (iii) Sufficient conditions for optimality. For some special classes of optimal control problems, one finds a unique control u*(*) which satisfies the Pon- tryagin’s necessary conditions. In this case, u* provides the unique solution to the optimization problem. For general nonlinear systems, however, conditions which guarantee the optimality of a control u*(-) can only be obtained by a global analysis.
1 Introduction 11 Toward this goal, a standard technique is to embed (1.18)-(1.19) in a family of problems, obtained by varying the initial conditions. The value function V, defined as = min uEU subject to x = f(t, x, u), x(r) = y, x(T) e J?, can then be characterized as the solution to a first order Hamilton- Jacobi partial differential equation and computed by dynamic program- ming methods. In turn, from the knowledge of the function V and its gradient VXV, one can determine the optimal control и in feedback form. The strong nonlinearity of the Hamilton-Jacobi equation and the possible lack of regularity of the value function V account for the main difficulties toward a rigorous mathematical analysis. In this direction, a major step forward has been provided by the theory of viscosity solutions. In addition to the fundamental theory, valid for control systems of the general form (1.3), a wealth of results are available for some special systems which can be analyzed in much greater detail. In particular, consider the linear system with constant coefficients x = Ax 4- Bu, (1-22) where x 6 IRn, и 6 IRW and the matrices А, В have dimension n x n and n x m, respectively. For a given control t the corresponding solution of (1.22) admits the explicit integral representation rr(t, u) = efAa?(0) + I s^ABu(^s)ds. Jo This allows an in-depth study of all the relevant properties of the system. Another important class consists of semi-linear systems, having the form m i = fo(x) + ^fi(x)ui, 1=1 (1.23) where fa. fa, • • • , fm are smooth vector fields on IRn. In general, there exists no explicit representation for the trajectories of (1.23) in terms of integrals of the control. Nevertheless, a rich mathematical theory has been developed for these systems, applying techniques and ideas from differential geometry and the theory of Lie algebras.

2 Review of Differential Equations In the basic model of a control system, as soon as a control function и — u(f) is assigned, the evolution can be determined by solving the O.D.E. z = g(t, x) = f(x, u(0) (2.1) In this chapter we review various aspects of the theory of differential equations, with particular focus on issues that arise from applications to control systems. Let J? be an open set in IR x IRn. Given a function g : J? IRn, by a (Caratheodory) solution of the O.D.E. x = g(t,x) (2.2) we mean an absolutely continuous function 11—> defined on some interval which satisfies (2.2) almost everywhere. Equivalently, we require that x(t) = x(to) -|- / g(s,x(s))ds J to for every t € [£оД1] where the function is defined. In the classical theory of ordinary differential equations, it is assumed that the function g is continuous w.r.t. both variables. For applications to the theory of control, however, it is important to consider also the case where g is only measurable w.r.t. the variable t. This more general setting is needed for two main reasons: In many practical situations, a controller acts on the system using a fi- nite number of switches that can be turned on and off. In a mathe- matical model, the control function thus takes the form t »—► u(t) = with Uj(t) = 1 or Uj(t) = 0 if at time t the J-th switch is turned on or off, respectively. The function u(-) thus takes values in a discrete set. All non-constant controls are necessarily discontinuous. In several optimization problems, the existence of an optimal control can be established only within the class of all measurable functions u(-). Quite often, the optimal control is actually discontinuous.
14 2 Review of Differential Equations When a measurable control function и = u(t) is inserted in the equation describing a control system, we obtain an O.D.E. of the form (2.1), whose right hand side is only measurable w.r.t. the time t. 2.1 Fundamental theory We begin this chapter by proving a basic result on the local existence and uniqueness of solutions to the initial value problem for the O.D.E. (2.2). Basic assumptions. Throughout this section we assume that the function g : J? i—* IRn satisfies the following conditions: (A) For every x the function t—>g(t, x) defined on the section 41x = {t : (t, rr) € 12} is measurable. For every t, the function x^>g(t. x) defined on the section = {x : (t, x) G f?} is continuous. (B) For every compact К C (1 there exist constants C#, Lk such that |3(t,x)| < , \g(t,x)-g(t,y)\< LK\x-y\ for all (t, x), (t y) £ К. (2-3) Fig. 2.1. The horizontal and vertical sections J2«, and the approximation of w by a piecewise constant function. Theorem 2.1.1. (Existence of solutions). Given a map g : 41 IRn, consider the Cauchy problem X = g(t,x), x(to) = x0, (2-4)
2.1 Fundamental theory 15 for some (to,^o) € ft- (i) If g satisfies the assumptions (A), (B), then there exists e > 0 such that (2.4) has a local solution x(-) defined for t€ [to, to + e]. (ii) Assume, in addition, that the function g is defined on the entire space IR x IRn and there exist constants C, L such that \g(t,x)\<C, \g(t,x) — g(t,y)\ < L\x — y\ for all t,x,y. (2.5) Then, for every T > to, the initial value problem (2.4) has a global unique solution x(-) defined for all t G [to, T]. Moreover, the solution depends contin- uously on the initial data Xq . Proof. The proof will be achieved in several steps. 1. We first prove (ii). Assuming that (2.5) holds, we shall construct a forward solution of (2.4), on any given interval [to, Т]. In view of applying Theorem A.2.1 of the Appendix, define A = lRn. The initial condition xq € IRn plays here the role of a parameter. Moreover, let X be the space of all continuous functions from [to,T] into IRn with the “weighted” norm lk(-)llt = toI5^T e which is equivalent to the usual C° norm lk(-)llco = Wf)l- Finally, define the map Ф : А x X 1—► X by setting Ф(т0, w(-))(t) = xq + [ g(s,w(s))ds, t G [t0,T]. J t0 (2.6) (2.7) 2. To prove that Ф is well defined, for each function w(-) G X we need to show that the composite map s 1—► g(s, w(sf) is integrable. This is not entirely obvious, because g itself is not continuous. We argue as follows: Given the function w(-), consider the sequence of piecewise constant functions as in Fig. 2.1, with = w Л -u *0 n----I \ y / if to + - I/ to 4----- у t g By (A), the maps t—>g(t, w^tf) are all measurable. Moreover, the second assumption in (2.5) implies lim |<?(£,w(£)) — g(t, wl/(t))| < lim L|w(t) — wp(f)| = 0. V—*OO P—>OO Hence the function t—>g(t, w(tf) is measurable, being the limit of a sequence of measurable maps. Since |^(s,w(s))| < C for all s, the integral in (2.7) is
16 2 Review of Differential Equations well defined and depends continuously on t. Hence, Ф is well defined and takes values inside X. 3. The continuous dependence of Ф on Xq is obvious. To study its depen- dence on w(-), consider any two functions w, w' e X and set ||w — w'||j = 6. Recalling the definition (2.6), we have |w(s) — wz(s)| < Se2Ls for all s 6 [a, b\. Moreover, the assumption of Lipschitz continuity in (2.5) implies e 2Lt |Ф(яо, w)(t) — Ф(^о, w')(t)| = e 2Lt I g(s,w(sy) — g(s,w'(s))ds JtQ <e 2Lt 1 L\w(s) — w'(s)| ds */ to < e~2Lt I L8e2Ls ds J t() 6 < 2 for all t G [*0, Т]. Therefore, ||^(xo,w)-^(x0,w')llt | Ik-w'llf (2-8) 4. We can now apply fixed point Theorem A.2.1, obtaining the existence of a unique continuous mapping xq—»#(•) such that x = Ф(хв,х), i.e. x(t) = Xq + / g(s,x(xy)ds for all fG[t0,T]. J to By definition, #(•) is the required solution to (2.4). This achieves a proof of (ii). 5. Finally, we prove the statement (i) concerning local existence, without the additional assumption (2.5). Choose e > 0 small enough so that the cylinder K = {(t,x): \t-to\<£, |x - x0| < e} is entirely contained inside the domain Г2. Then consider a smooth cut-off function ф : IR x IRn i—> [0,1] such that ф = 1 on A\ while 0 = 0 outside some larger compact set K', with К С К' С 12. Observe that the function if if (t,x) G ii, (t,x) a, (2-9)
2.1 Fundamental theory 17 satisfies (A) and (B), together with the extra assumption (2.5), because it vanishes outside the compact set K'. By the previous steps, there exists a solution x(-) to the Cauchy problem i(t) = 9^(f.x(f)\ x(to) = x0, (2.Ю) defined on arbitrarily large interval [io, Т]. We now recall that the cylinder К is a neighborhood of the point (io,xo). Therefore, for some e > 0 sufficiently small, the point (t,x(t)) remains inside К as i G [io, to + e]. Since g and coincide on K, the function jr(-) thus provides a local solution to the original problem (2.4), restricted to the smaller time interval [io , io + e]. Remark 2.1 The construction of a solution backward in time, on an interval [to — £, to] is entirely analogous. It can be reduced to the previous case by reversing time (i.e. setting т = —t) and considering the equation dx(r) dr -s(-T,x(r)). The next lemma provides a useful tool for estimating the distance between two solutions of a differential equation. It represents the main ingredient in several uniqueness proofs, and in this respect it can replace the original version of Gronwall’s Lemma , which has more mathematical content and also a longer proof. Lemma 2.1.2. (Gronwall). Let z(-) be an absolutely continuous nonnegative function such that z(io) < 7, z(t) < a(t)z(t) + /3(i) for a.e. t € [io, T], (2.11) for some integrable functions and some constant 7 > 0. Then for every t G [t0, Г] the following holds >t z(t) < 7 exp to a(s)ds\+ / /3(s)expl / a(cP) da / J t0 \JtQ ds . (2.12) Proof Notice that the right hand side of (2.12) is precisely the solution to the linear Cauchy problem w(i0) = 7, w(t) = a(i)w(i) 4- Z?(t). In order to establish the inequality, consider the absolutely continuous func- tion V>(i) = exp / a(s)ds) z(t) — / /?(s)exp Jt0 / L JtQ ds
18 2 Review of Differential Equations Using (2.11), a direct computation shows 'ijfit) < 0 for almost every t. There- fore V>(£) < V>(*o) = < 7 for all t e [to? Т]. (2-13) Multiplying (2.13) by exp( a(s)d.s), from the definition of -0 we obtain (2-11). We can now establish the uniqueness of the solution to the Cauchy problem (2.4), whose local existence was proved in Theorem 2.1.1. Theorem 2.1.3. (Uniqueness). Let g : J? i—> IRn satisfy the assumptions (A) and (B), as in Theorem 2.1.1. Let Xi(-),X2^) be solutions of (2.4), defined on the intervals [to, tj, [to, t?] respectively. If T = min{ti,t2}? then x\(t) = ^(t) for all t € [to, Т]. Proof. Let Lk be the Lipschitz constant in the assumption (B), corresponding to the compact set К = |((,ari(t)), : t ё [io,r]|. Then the absolutely continuous function z(t) = |^i(t) — #2 (01 satisfies z(t0) = 0, i(t) < |ii(t) — 0:2 (01 < Lxzff) for a.e. t. Applying Gronwall’s lemma with a = Lk, 0 = 7 = 0, this implies z(t) < 0 for all t e [to, Т]. In general, the solution to the Cauchy problem (2.4) is defined only locally, for t in a neighborhood of the initial time tg. If a solution cannot be extended beyond a certain time T, two cases may arise (see Fig. 2.2): 1. As t —> T—, the point (t, x(t)) approaches the boundary df2 of the domain 12 where g is defined. 2. As t—, the solution blows up, i.e. \x(t)| —>oo. The next theorem shows that these are actually the only two possibilities. Theorem 2.1.4. (Maximal solutions). Let the basic assumptions (A), (B) hold. Let T > to be the supremum, of all times т such that (2.4) has a solution #(•) defined on [to,r]. Then, either T = oo, or else ( hm_ ^(t)| + (2.14)
2.1 Fundamental theory 19 Fig. 2.2. Two maximally extended solutions. Proof. Assume T < oo. If (2.14) does not hold, then there exist Л/, s > 0 and a sequence ty^T— such that, for all v > 1, |ж(^)| < Af, 2j(ip)), сИ2) >e. By possibly taking a subsequence, we can assume that x(ty) converges to a point Xoo, with (T, ^oo) e J?. Choose p > 0 so small that the cylinder К = {(t,x) : \t - T\ < p, |x - Zoo| < p} is entirely contained in J?. As in (2.9), construct a function : IRxIRn »—> IRn such that = g on A, and in addition g^ satisfies the global bounds |^(*,ж)|<С, |<z+(t,rr)-5f(t,y)| < L|x-y| for all t,x,y, (2.15) for some constants C, L > 1. Fix 5 > 0 small enough so that (2C+l)5<p, and choose и so large that XqQ I < T try < By (ii) in Theorem 2.1.1, the Cauchy problem У(«) = 9\t, y(tv) = x(t„) has a solution ?/(•) on [ty, T + <5]. We can now define an extension x of x by setting f tQ<t<ty X^ = \y{t} if tv<t<T + 8.
20 2 Review of Differential Equations Since |?/(t) | < C, for t e [t^, T 4- J] we have \y(t) ~^oo| < \y(t) - + |^(L/) -Zed < C(t-tp)4-5 < 2Cb + 8<p. Therefore, for all t 6 [tp, T4-5], the point (t, y(tfi) remains inside the compact К where g and g^ coincide. The function £(•) thus provides a solution of the original problem (2.4), defined on the strictly larger interval [to, T 4-5]. This contradicts the maximality of T, thus proving the theorem. In the case where g is defined on the entire domain [to, oo] x Rn, to establish the the global existence of the solution to (2.4) it thus suffices to prove that x(-) remains bounded on bounded intervals of time. In general, a-priori estimates on the size of |#(t)| can be obtained by a comparison with a scalar O.D.E., as we now describe. Theorem 2.1.5. (A-priori bounds). Let g : 12 i—> Rn satisfy the basic assumptions (A)-(B). Let 0 : [to, ti] xR i-> R be a scalar function, measurable in t and continuous in x, such that 4>(t,r) > max \g(t,x)\ for all t,r. (2.16) |x|=r Let r : [£o3i] ► R be an absolutely continuous function such that r(t) >zp(t,r) for a.e. t e [Wi], r(t0) > Ы • (2.17) If the set К = {(t,x) : to < t < t±, |x| < r(t)} is contained in (2, then the Cauchy problem (2.4) has a solution defined on the entire interval [to, ti], which satisfies |#(t)| < r(t), for all t e [to, *1]. (2.18) Proof Define the auxiliary function ^*(t,rr) = \g(t,j:) \-9(г’ г^й) if if kl < r(t), И > r(t). (2-19) Otherwise stated, g*(t, x) = g(t, 7гг(гг)), where 7r*(a?) denotes the perpendicu- lar projection of x € Rn on the ball centered at the origin with radius r(t). From the basic assumptions (A)-(B) it follows that #* is globally bounded and Lipschitz continuous w.r.t. x on the domain [to, tj x Rn. Therefore, the Cauchy problem y(t) = 2/(0) = x0, (2.20) has a unique global solution x* : [to, ti] IRn . We now observe that the two maps t h-> |.r*(t)| and t •—* r(t) are both Lipschitz continuous and satisfy \x*(to)| < r(to), together with 4k*(0l < |ff*(tz*(0)| < max |5(t,x)| < ^(t, r(0) < r(t) at i I |i|=r
2.2 Linear systems 21 at almost every time t € [^о,й]. Hence |a?*(£)| < r(t) for all t e (2.21) By definition g*(£,x) = g(tyx) whenever |rr| < r(t). Therefore, the function #*(•) coincides with the unique solution of the Cauchy problem (2.4). By (2.21) this proves the theorem. In many applications, useful estimates can be obtained from the above theo- rem by a judicious choice of the function r(-). Corollary 2.1.6. Assume that the function g = g(t,x) satisfies the bound |^,х)| < C(1 4-|®|) (2.22) for some constant C and all x € lRn. Then any solution of (2.4) satisfies the a priori estimate |x(t)| (1 + |x(i0)|) • (2.23) Indeed, for t > the estimate (2.23) is obtained by taking ^(t, r) = C(1 + r). Solving the scalar Cauchy problem r = C(1 T r), r(t0) = r0 = hr(t0) I we find r(t) = eC(t"fo)r0+ f Cec(-s-to} ds = eC('-to)|z(f0)| + (eC(s-to) - 1) . This yields (2.23) in the case t > to- For t < to the estimate is obtained in an entirely similar way, reversing the direction of time. 2.2 Linear systems In this section we consider differential equations of the form x = A(fi) x (2.24) p=-pA(t) (2.25) where t—>A(t) is a measurable map from an interval [a, b] into the set of n x n matrices. We regard x as a column vector and p as a row vector. The equations (2.24)-(2.25) are thus a short-hand notation for
22 2 Review of Differential Equations ( an • • • ain \ ^nl ’ ’ ’ ®nn / Throughout this book, the norm of a matrix A is defined as ||A|| = max |Arc| = max|pA|. |x| = l |p| = l In the special case where the n x n matrix A(t) = A is independent of time, the solution to the Cauchy problem x = Ax, #(0) = x (2.26) can be written in the form x(t) = etAx. (2.27) Here the exponential matrix elA is defined as the limit of the absolutely con- vergent series fk дк ‘‘л - E -jr <2-28> k=0 We recall that, by definition, A0 = I is the n x n identity matrix. Observe that, if В = R~rAR for some invertible matrix 7?, then = Re^R-1. Indeed, for every k > 1 one has (RBR-1^ = RBR-1 • RBR-1 • • • RBR~l = RB^"1. The actual computation of the exponential matrix etA can thus be carried out by reducing A to a more convenient canonical form B. and then computing etD. Example 2.1. Assume that A is a 6 x 6 matrix, with det(a-^) = (C-A)«-M)3(C-(a + i/3))(C-(a-f/3)), so that Л is a simple real eigenvalue, /i is a multiple eigenvalue and a±i(3 are a pair of complex conjugate eigenvalues. Assume that the geometric multiplicity of /1 is 1. Then there exists an invertible matrix R that reduces A to the canonical form B = R~'AR = /Л 0 0 0 0 0 \ 0/1100 0 0 0 /z 1 0 0 0 0 0 // 0 0 0 0 0 0 a -(3 \0 0 0 0 (3 a /
2.2 Linear systems 23 In this case one has (ext 0 0 0 0 ° \ 0 te*“ («2/2)6*“ 0 0 tB _ 0 0 eMt 0 0 e = 0 0 0 0 0 0 0 0 0 eat cos/3t —eat sin (3t 0 0 0 eat sin eat cos j3t ) and etA = Re^R-1. The next theorem provides the global existence and uniqueness of solut ions to linear systems, with general time-dependent coefficients. Theorem 2.2 .1. (Existence of solutions for linear systems). Assume that ||A(t)|| < L for some constant L and all t E [a, b]. Then, for any to e [a, b] and every initial condition Xo E !Rn, the Cauchy problem x = A(t)x, x(to) = Xq, (2.29) has a unique solution defined on the entire interval [a, 6]. This solution satisfies |x(t)|<eL|‘ tol|ar0|. (2.30) Proof. Indeed, the local existence and uniqueness are obtained applying The- orem 2.1.1, (i) with g(t,x) — A(t)x. For t > to, the estimate (2.30) follows from Theorem 2.1.5, with ^(t, r) = Lr, r(f) = eL^“tol|xo|. Since ||Л|| = ||— A||, reversing time we obtain (2.30) also in the case t < to- Of course, an entirely similar result holds for (2.25). The systems (2.24) and (2.25) are related by a fundamental property: Theorem 2.2 .2. (Adjoint systems). Let #(•), p(-) be any two solutions of (2.24), (2.25) respectively, defined on the same interval of time: t E [a,b]. Then their inner product p(t) • x(t) is constant. Proof. This is verified by the direct computation ^-(p- x) = p- x+p>x = —pAlf) - x + p - A(t}x = 0. dt Since the system (2.24) is linear and homogeneous, the set of solutions is a linear space. In other words, if #(•) and ?/(•) are solutions of (2.24), then the linear combination Xx + py provides yet another solution, for every choice of X,p E 1R. To obtain the general solution to a Cauchy problem, it thus suffices to construct a set of n linearly independent solutions. This motivates the following construction. Let
24 2 Review of Differential Equations be the elements of the standard basis in IRn. For a fixed time s and each j = 1,..., n, call 11—► Vj(t, s) the solution to the Cauchy problem = e,. Construct the n x n matrix Af(£,s) whose columns are given by the vectors vi, V2, •.., vn • Namely M(t,s) = vi(t,s) v2(t,s) vn(t,s) This is called the fundamental matrix solution of (2.24). For a fixed value of s, the map t M(t, s) provides a matrix-valued solution to the problem = A(t)M{t, s), M (s, s) = I, (2-31) where I denotes the n x n identity matrix. Theorem 2.2 .3. (Properties of the fundamental matrix solution). As- sume that the matrices A(t) are uniformly bounded, with coefficients depending measurably ont € [a, b\. Then for every € lRn, the function x(T) — M(T, .s)£ provides the solution to the Cauchy problem. = A(t)x(t), x(s)=£. (2.32) The fundamental matrix solution M satisfies M(t, s)M(s,r) = M(t,r) for all (2.33) = (2-34) Moreover, if h : [a, b]lR.n is integrable, then any function satisfying x(t) = M(t,r)x(r) + У M(t, s)h(s) ds (2.35) is a solution to x = A(t)x(t) + hft). (2.36)
2.2 Linear systems 25 Proof. The first statement follows immediately from (2.31). To prove (2.33) we observe that, for every £ € Rn, r e [a, 6], the functions ^1(0 = Af(t, s)Af(s, r)£, ^2(t) = are both solutions of the Cauchy problem J^rr(£) = A(t)x(t), x(s) = M(s, r)£ . By uniqueness, Xi(t) = xz(t) for every t. Since £ € IRn was arbitrary, this proves (2.33). From (2.33) it now follows M(t, s)M(s, t) = I, hence the map s h-> = [Af(s,£)]-1 is differentiable. An elementary computation yields — [M(M)M(M)] = 0 = M(s,t) + M(t,s)[A(s)M(M)]. (2.37) Multiplying (2.37) on the right by [M(s,t)]-1 we obtain (2.34). The last assertion is verified by straightforward differentiation, writing (2.35) in the equivalent form xft) = Л/(£,т) я:(т) -F [ M(r,s)h(s)ds By Theorem 2.2.1, each bounded, measurable matrix-valued function A(-) determines a fundamental matrix solution M(•, •). The next result states that M depends continuously on A, in the appropriate norms. Theorem 2.2 .4. (Continuous dependence of the fundamental matrix solution). The map A(-) i—* Af (•, •) is continuous w.r.t. the distances M(-)-OllL1= [b\\A(t)-A'(t)\\dt, J a \\M(•, •) - M'(-, -)llco = max \\M(t,s) - M\t,s)||. a<s,t<6 Proof. Fix any s G [«,5] and let v be any unit vector. For t > s, define v(t) = M(t, s)v, v'(t) = M'(t, s)v, z(t) = v(f) - ?/(£), and observe that z(s) = 0. Then z = v — i)' — A(t)v - A'(fyu', and ^|z(f)| < H(t)« - + \A(t)v' - А'(*У| <||A(t)||-|z| + ||A(t)-A'(t)||-K|.
26 2 Review of Differential Equations Since we are assuming the bound ||A'(t)|| < L, there holds |«'WI < exP IH'MII da^ < ець~а\ We now apply Gronwall’s Lemma with a(t)=||A(t)||5 Ж=еЬ(6-")||А(0-Л/(01|. 7 = 0, obtaining |z(i)| < eL^ £ ||A(<t) - A'(a)|l exp ||A(C)II da. The above estimate shows that max max \M(t, s)v — M'(t, s)v| (2.38) a<s<t<b |v|=l <e^exp( / ||A(cr)|| da j • / ||A(cr) - >l'(cr)|| da. a ! J a Clearly, the right hand side of (2.38) approaches zero as A—+A' in the L1 norm. The estimates for t < s are obtained in the same way, considering the systems dv л/ x dv' 4/z . , . r , . — = — A(—r)f, -т— — — A (—г) r, t = —t e [—6, —a]. 2.3 Differentiability with respect to initial data A common problem in the theory of optimal control is to test whether a given control function ?/*(•) is optimal. This is usually done by comparing the trajectory t i—> #(f,u*) with other nearby trajectories. The basic ingredient in this analysis is a detailed description of how the solution of the Cauchy problem i(t) = g(t, #(£)), x(t) = £ (2.39) changes, as the initial data t, £ are varied. Throughout the following, we denote by £ н-> rr(f,T, £) the solution of (2.39), while Dxg(t,x) is the n x n Jacobian matrix of first order partial derivatives dgi/dxj at the point (t,x). Theorem 2.3.1. (Directional derivatives). Let g : J? h-> ]Rn satisfy the basic assumptions (A)-(B) and be continuously differentiable w.r.t. x. Let £(•) = x(-, to^o) be the solution of (2.4), defined for t G [fo, h]- For Vq G IRn, call ?;(•) the solution of the linear Cauchy problem
2.3 Differentiability with respect to initial data 27 Then i’(t) = Dxg(t,x(t)) v(t), v(t0) = v0- lim £->0+ x(t,t0,x0 + ev0)-x(t) --------------------------V(t) = I), e (2.40) (2.41) the limit being uniform for t € [£o,*i], |^o| < 1- Fig. 2.3. The first order variation of a solution, as the initial point жо is changed to Xo + EVo. Proof. 1. As in Fig. 2.3, for e sufficiently small define x£(t) = x(t, to, x0 + ev0), y£(t) = x(f) + £v(t). (2.42) To prove the theorem, we need to show that lta _0 (2 43) £-*0+ £ Observe that x£(-) is the fixed point of the map w Ф(ж0 + w), with ф(х0 + evq, w(-)\(t) = xo -I- evq + / p(s,w(s))ds, J to We recall that Ф is contractive with respect to the equivalent norm ||-||| defined at (2.6). Thinking of y£(/) as an approximate fixed point, by (A.6) in Theorem A.2.1 of Section A.2, one has 1 2 - ||xe - ?/e||t < - ||Ф(х0 + ev0, ye) - УеIIt • It therefore suffices to show that, uniformly for |uq| < 1,
28 2 Review of Differential Equations lim sup - z0 + 6v0 + / g(s,y£(s,voy)ds - s/e(t,v0) JtQ = 0. (2.44) 2. By the equations (2.4) and (2.40), satisfied by £(•) and by v(-) respectively, we have жо + / .g(s,£(s)) ds — x(t) = 0, Ao ev0 + / Dx^(s,^(s)) J to • ev(s) ds — Ev(t) — 0 . The quantity in (2.44) can thus be estimated as 1 — T T E 1 E / g(s, x(s) -I- Ev(sf) ds — x(t) — Ev(t) J to #o + evo+ / g(s,x(sf)ds J to •'to ) + 11 lPxg(s, £(s) + (tev(s)) — Dxp(s, f(s))] • ev(s) da ds J to A < I I i(s) + crev(s)) - Dxg(s,x(s))\\ • |v(s)|d<7ds. (2.45) J to Jo Let К C 12 be a compact set containing a neighborhood of the graph of x and Lk as in assumption (B). Then |v(s)| is bounded by eLl<s|vq|, hence it is clear that the right hand side of (2.45) approaches zero, uniformly for t G [to, Zj, |Vo| < 1. In turn, this proves (2.43), hence (2.41). For each t € [Zo, й], Theorem 2.3.1 states the existence of all directional derivatives for the map £ •—► rr(Z,Zg,£). In the next theorem, we observe that these derivatives depend continuously on the point Xo where they are com- puted, and conclude that the map is differentiable. Theorem 2.3 .2. (Differentiability w.r.t. the initial point). Let g : Г2 h-* IR” satisfy the basic assumptions (A)-(B) and be continuously differentiable w.r.t. x. Let f(-) be a solution to (2.4), defined on [Zg,ti]. Then, for any t € [Zg,Zi], the map £ i—► x(t,to,£) is continuously differentiable in a neighborhood of .Tg. Its Jacobian matrix at a given point Xo is D^t,t0^)^xo = Af(Mo), (2.46) where A/(-,-) denotes the fundamental matrix solution to the linear problem v(t) = Dxg(t,x(t,t0,xQ)) (2-47)
2.3 Differentiability with respect to initial data 29 Proof. The n partial derivatives of the map £ i—► a;(t,to,C) at £ = are defined as x(t, fо, .lq “I- c e^) x(t, to, xq) . hm —------------------------------ i = 1, • • • , n, £—0 e where {ei, • • • , en} denotes the standard basis in IRn. By Theorem 2.3.1 these limits exist, being equal to Vi(t) = M(L ^o)e?? where M(•, •) is the fundamental matrix solution to (2.47). To complete the proof, it thus suffices to show that these partial derivatives depend continuously on the point xq. Let be a sequence of initial points, with lim^-^o = Xo- Then the cor- responding trajectories x(-, Zq, Cp) converge to □?(•, to, xo) uniformly on [to, ^1], and the matrix-values functions t i—► A„(t) = Dxg(t,x(t,to,£tU')>) converge in L1 to the map t i—► A(t) = Dxg(t, x(t, to, xo)). By Theorem 2.2.4, the corre- sponding fundamental matrix solutions A/p(-, •) converge uniformly to Af(-, •). In particular, lim M„(t, to)ei = M(t, to)ei, y—*oo proving the continuity of the partial derivatives. The last result in this section is concerned with the differentiability of the solution x(-,r, £) w.r.t. the initial time t, under the additional assumption that g is continuous also w.r.t. time. Theorem 2.3 .3. (Differentiability w.r.t. the initial time). In addition to the basic assumptions (A)-(B), let the function g be continuous in both variables t,x and continuously differentiable w.r.t. x. Let a:(-,Zo,^o) be a so- lution of (2.4), defined on some interval Then, for every t e [fo,^i] the map r 1—► x(t,r, xq) is continuously differentiable in a neighborhood of to- More precisely, one has DTx(t, r, ^o)(T=to = “Af (t, to)g[to, xo), (2.48) where •) is the fundamental matrix solution for (2.47). Proof. 1. Call £(t) = x(to,r,xo) the value at time t = to of the solution to x = g(t,x), ж(т) = ж0- Since x(t,r, Xq) = x(t,to,^r)), from Theorem 2.3.2 it follows DTx(t,r,x0) = D^x(t,t0,x0)-^- = M(t,t0)^L (2.49) 2. Using the assumption that g is continuous in both variables, we now com- pute C(t-) - ж0 т - to 1 rto —- -----— / g(s,x(s.T,xoy)ds = -p(t0,x0). T — to JT Inserting this value in (2.49) one obtains (2.48). л— - liD? dr r^tQ
30 2 Review of Differential Equations 2.4 A transversality theorem Let g : [to, x lRn t—> IRn be a continuously differentiable vector field, and let Л4 C IR',+1 be a n-dimensional manifold. More precisely, assume that there exists a Cl mapping ф : IRn+1 i—> 1R such that M can be represented as the zero level set of ф : M = {(t,x); ф(},х) = 0}, (2.50) and such that the gradient of ф does not vanish on any point of M : Чхф)(1,х) = (<^,0X1, • • • ,фХп)(1,х) / (0,0, • • • ,0) for all (f,x) € M. (2.51) Let ./,*(•) be a solution to the differential equation ±(t) = g(t,x(tf). (2.52) If (т, x(t)) e M. we say that #(•) intersects M transversally at the point (r, x(r)) if x(r)) + Vx(/>(r, x(r)) • ,g(r, t(t)) / 0 . (2.53) This means that the vector (1, д(т, .t(t)) is not tangent to M at the point (t, a:(r)). The next theorem states that “almost all” trajectories of (2.52) have only transversal intersections with Л4. Theorem 2.4.1. (Generic transversality of trajectories). Let g = g(t,x) be continuously differentiable w.r.t. both t and x. Assume that, for every Xo in an open ball В C IR71, the solution t i—* x(t,to,xo) of (2.52) with initial condition x(fo) = xo is defined on the whole interval [to^i]- Let Л4 C JRn-H1 be an n-dimensional embedded manifold, as in (2.50)-(2.51). Call N the set of all points Xo E В such that (t,t(t, io^o)) £ АЛ for some r 6 [to? but the intersection is not transversal. Then meas(ff) = 0. Proof. 1. The manifold Л4 admits a countable open covering {Л4}} such that, for every i, there exists a chart : Ai Mi, with Ai C IRn a bounded open set and at1 diffeomorphism. Since the countable union of negligible sets has measure zero, it suffices to prove the theorem assuming that M = <p(A), where A is a bounded open set in lRn and <p : A i—> IRn+1 is a Cl embedding, i.e.,: Rank(D^(T/)) = n for all у e A. 2. For each у € A, if 92(1/) = (t, x(t, to, #o)) for some r, xq, set Ф(у) = j;q. Otherwise stated, if </?(?/) = (t(?/), £(?;)), then tf'O/) = x(to, r(y), C(,v)) (2.54) whenever the right hand side of (2.54) is defined and lies in B. By Theo- rems 2.3.2 and 2.3.3, x(fo,T, £) is continuously differentiable w.r.t. t, £. There- fore, Ф is a C1 function defined on some open subset A' C A, taking values inside B.
2.4 A transversality theorem 31 Fig. 2.4. The only two non-transversal trajectories are the ones originating from x i and X2 • 3. Let (t/i,--- ,yn) be coordinates on A. If = (r,£), from (2.46), (2.48) it follows that the partial derivatives of Ф w.r.t. t/i, • • • ,yn are the vectors dyi \dyt dyij where Af(-, •) denotes the fundamental matrix solution for the linear system v(t) = Dxg(t,i(t,T,O) v(t). The Jacobian determinant det(D^(?/)) vanishes if and only if the n vectors (2.55) are linearly dependent. Observing that the first n vectors in (2.55) form a basis for the tangent space to M at the point y)(y), we conclude that det(£>^(?/)) = 0 if and only if the vector (l,^(r, £)), with £ = аг(т, to,^o), is tangent to M at (r, £), i.e. if and only if the crossing at (r, £) = <p(y) is not transversal. 4. By the previous analysis, Af = P| Л4 where Л4 = |^(т/); у G Af, det(D#(y)) < e} . E>0 Computing the area of J\T£ using the ^/-coordinates, one obtains meas(A4) < / |det(£>’Zz(?/))| dy < emeas(A). J{y€A', |det(D^(?/))|<£} Letting £ 0 we obtain meas(AQ = 0, proving the theorem.
32 2 Review of Differential Equations Example 2.2. Every solution of the differential equation x = l^l1/2 with ж(0) < 0 intersects the manifold M. = {(tx); x = 0} C IR2 tangentially, at some time t > 0. Of course, this does not contradict the previous theorem since the function g(t,x) = |x|1?Z2 is not differentiable on M . Problems 2.1. Consider the Cauchy problem z(0) = Xq , with a• > 1. Show that, for every xq > 0, the solution becomes unbounded in finite time. Find the time T, depending on q and Xq, such that lim x(t) = oo . 2.2. Consider the Cauchy problem: x = 2\/|ж[, z(0) = 0. Show that for every a < 0 < b there exists a unique solution xab, defined for every t e IR, such that xab(t) < 0 for t < a. xab(t) > 0 for t > b and = 0 for b < t < a. 2.3. Let : J? нч. IR7' satisfy the assumption (A). Prove that (B) is satisfied if the following holds. The function g is continuously differentiable w.r.t. x and, for every compact К C 12 there exist constants Cj<, Lk such that \g(t, x)| < CK, ||£>^(t,x)|| < LK for all (t,x) e K. 2.4. Show that in Theorem 2.1.1 the global existence of solutions still holds if the assumption (2.5) is replaced by the following weaker condition: There exists an integrable function ф : [io,T] i—► R± such that for all (t,y) e IR x IRn. Hint: use the equivalent norm l|w(-)||t = max e v’(s)ds |w(0|. £€[a,b]
2.4 A transversality theorem 33 2.5. Show that the conclusion of Theorem 2.1.5 remains valid if the definition (2.16) of is replaced by / x \ V>(£, r) = max ( — , g(f,x) > for all t,r . |x|=r \|x| / Hint: the derivative of the map 11—► |x(t)| is computed by the inner product 2.6. Given the differential equation x = x , let x(-) be the solution with initial data x(0) = 1. Moreover, for £ > 0, call xe(-) the solution with perturbed initial data xe(0) = 1 + e. For t < 1, compute the vector „(,) _ Um . 4 7 €->0 8 Check that it satisfies the corresponding linearized equation (2.40) with v(0) = 1. 2.7. Consider the linear system of differential equations Xi = x2, x2 = x3, ±з ——Xi- Compute the solution for a given initial data (xi, x2, x3)(0) = (xi, x2, x3). 2.8. Write the second order equation x + x + x = /(t) as a first order system. Write a formula for the solution, with initial data x(0) = x(0) = 0. 2.9. Consider a linear pendulum with external force: 0-0 = 0(0) = 0o, 0(0) = 0. (a) Compute the solution for u(t) = sign(0(£)) and 0q = тг/2. (b) Consider the solution 0e for the same forcing term, corresponding to initial data 0£(O) = тг/2 + е. Determine the tangent to the curve £ —> 0е(тг) at £ = 0. (c) Consider the forcing terms ue such that u£(t) = 0 for 0 < t < £ and ue(t) = sign(0(t)) otherwise. Let ye be the corresponding solutions for Oq = 7t/2. Compute the tangent to the curve £ —* y£ (тг) at £ = 0. Hint: compute the times t = t(e) where 0(t) = 0.

3 Control Systems In this chapter we begin a study of the control system: x = f(t,x,u), iz(-) e Z7, (3.1) where the set of admissible controls is defined as U — { w(-) measurable, u(t) G U for all t} (3.2) We shall assume the following basic hypothesis: (H) The set U C Hlm of control values is compact, f2 is an open subset of 1R x IRn, the function f : Г2 x U i—> IRn is continuous in all variables and continuously differentiable w.r.t. x. We say that an absolutely continuous function #(•) defined on some interval [a, b] is a solution of (3.1) if its graph {(t, x^t)); a < t < b} is entirely contained in <2, and if there exists a measurable control u, taking values inside U, such that ±(t) = f(t, u(t)) for almost every t G [а, Ь]. The main goal of this chapter is to provide information about the dynamics of the system (3.1). We will show how to construct solutions to (3.1), and study how trajectories depend on the choice of the control function u(-). The first section compares a control system with a differential inclusion, showing that the two mathematical models are essentially equivalent. In Sec- tion 3.2 we apply the basic results on O.D.E’s with measurable right hand side, and derive the existence and uniqueness of the solution, for a given control function t u(t). As the control ?/(•) varies, a whole set of possible trajectories is obtained. Some basic properties of this family of trajectories are described in the follow- ing Sections 3.3-3.4. In particular, it is important to understand which points can be reached at a given time T, by suitably choosing the control function. Preliminary results in this direction are given in Section 3.5. The case of linear systems with constant coefficients is studied in detail in Section 3.6. This is a classical topic in engineering literature, with exten-
36 3 Control Systems sive applications. Thanks to an explicit formula representing the trajectories, precise results can here be given. In turn, every nonlinear system can be locally approximated by a linear one, in a neighborhood of a given point. By a linearization method, from a global controllability result valid for linear systems, in Section 3.7 we thus obtain a local result valid for general nonlinear systems. For a special class of systems, where the control variable enters linearly in the equations, further controllability results are given in section 3.8, based on the analysis of Lie brackets. According to the analysis in Section 3.3, the set of trajectories of a control system is closed if, at each point x. the set of possible velocities F(x) C IRn is closed and convex. When this key property fails, one can always construct an auxiliary system, where the velocity sets F(x) are the convex hulls of the original ones. These “chattering systems” are introduced in Section 3.9. They are particularly useful in connection with optimization problems, because one can prove that an optimal control problem always admits a generalized solu- tion in the form of a “chattering control”. In the special case of linear systems, for any chattering control one can find a genuine control of the original sys- tem which steers the system exactly to the same terminal point. For example, points reached by controls u(t) G [—1,1] can also be reached using controls which take values only in the two end-points: u(t) = 1 or u(t) = — 1. This is the content of the famous bang-bang theorem, given in Section 3.10. 3.1 An equivalent differential inclusion In connection with (3.1), define the multifunction F(t.x) = {f(t,x,a;) : a; e U} and consider the differential inclusion x e F(t,x). (3.3) (3-4) Observe that, for each (L^), the set of admissible velocities x in (3.1) is given in parametrized form, as the image of the fixed set U c IRm. On the other hand, when we study the differential inclusion (3.4), we do not assume any parametrization of the set F(t,x) C IRn, see Figure 3.1. The next result shows that these two approaches are essentially equivalent. Theorem 3.1.1. (Filippov). An absolutely continuous function x : [a, b] Rn is a trajectory of (3.1) if and only if it satisfies (3.4) almost everywhere. Proof. 1. The fact that every solution of (3.1) is a solution of (3.4) is an immediate consequence of the definitions.
3.2 Fundamental properties of trajectories 37 Fig. 3.1. Parametrized and non parametrized set of velocities. 2. Viceversa, let x(-) be a solution of (3.4). Fix an arbitrary element w in the control set U. and define the multifunction W(t\ = f {w e U: w) = ±(i)} if x(t) € F(t,x(t)), [ {w} otherwise. Notice that the second alternative holds if either the function rr(-) is not differentiable at the time t. or else if ±(i) exists but does not lie in the set F(£, By assumption, this happens only on a set of times of measure zero. We now define the control и by choosing u(t) as the first element of the set W(Z) w.r.t. the lexicographical ordering. Such an element exists because each set W(£) is compact. 3. By construction it follows that x(t) = /(L z(f), ?/(£)) for almost every t e [a, b\. To prove that the control function u(-) is measurable, we use Lusin’s theorem and construct a sequence of disjoint compact sets J?,... with Ji C [a,b], meas [a, b] \ |^J Ji I = 0, \ / and such that the derivative £(•) is well defined and continuous, restricted to each Ji. Therefore, restricted to Л, the bounded multifunction W has closed graph. By Theorem A.7.3 in the Appendices, for t e Ji the lexicographic selec- tion t u(t) e IV(i) is measurable. Since the sets Ji cover almost the entire interval [a, b], we conclude that the selection и : [a, b] h-> U is measurable. 3.2 Fundamental properties of trajectories Let /, U satisfy the basic hypothesis (H). Let an initial value x be given. For any measurable control и : [0, T] U, the Cauchy problem
38 3 Control Systems ±(t) = /(t,a?(t), u(t)), rr(O) = ж, (3.5) is equivalent to ±(t) = g(£,a:(t)), #(0) = x, (3.6) where g(t,x) = f(t,x,u(t)) satisfies the basic assumptions (A) and (B) of Chapter 2. By the results proved in Chapter 2, if (0,x) E Г2, there exists e > 0 such that the Cauchy problem (3.6) has a unique local solution, defined on the small time interval [0,s]. To study how the solution ;r(-,u) varies with the control u. we first consider the globally bounded case, assuming (H*) The set U C IRm of control values is compact. The function f = f(t,x,u) is defined and continuous on the entire space IR x IR7' x U. continuously differentiable w.r.t. ж, and satisfies |/(Ж,«)| < C, \\Dxf(t,x,u)\\ < L. (3.7) for some constants C, L and all t,x,u. Theorem 3.2.1. (Global existence and continuous dependence). Let the assumption (H*) hold. Then, for every T > 0, и G U, the Cauchy problem (3.5) has a unique solution x(-,u) defined for all t E [0,T]. The “input output” map u(-) (—► rr(-, u) is continuous from L1 ([0, T]; IR771) into C° ([0, T]; IRn). Proof For each control u, the trajectory #(-, iz) is the fixed point of the trans- formation w i—> Ф(и, w) defined by Ф(и, w)(t) = x + I f(s,w(s),u(s)) ds. (3.8) ./o The theorem is proved showing that Ф : Л x X X satisfies the assumptions of the Contraction Mapping Theorem A.2.1 in the Appendix. Here we use the spaces X = C°([0, T]; R”) and Л = Ll([0,T]; Rm), so that the control func- tion u(-) plays the role of a parameter. On the space X we use the equivalent norm Mt = m<K e“2Lt|w(f)|. (3.9) 1. By the hypothesis (H*), the map Ф is well defined. If (zzp)p>i is a sequence of admissible controls approaching a in the L1 norm, for any subsequence и„' we can extract a further subsequence u„" such that и„"(Р) —> u(t) for almost every t. By the Lebesgue dominated convergence theorem, it follows lim / |/(s, w(s), u„(s)) — f(s, zz(s))l ds = 0. p-oo I I Since the subsequence was arbitrary, we conclude that the above limit holds for the entire sequence uu. Therefore, for each continuous function w, the image Ф(ир, ш) converges to $(zz, w) uniformly on [0, Т]. This proves the continuity of Ф w.r.t. u.
3.2 Fundamental properties of trajectories 39 2. For any fixed u. the second inequality in (3.7) implies |/(i,x, u) - f(t,y,u)\< L\x - y|. To prove that Ф is a strict contraction w.r.t. the second variable, assume ||w — w'||| = 6. Recalling the definition (3.9), one has |w(s) — w'(s)| < Se2Ls for all s € [0,T]. Therefore, for every t > 0 we obtain e 2Lt |Ф(и, - Ф(и, w')(t)| = = e~2Lt f(s, w(s), u(sf) — f(s, w'(s),u(sy) ds L\w(s) — w'(s)\ds L6e2La ds < -. 2 This implies ||<?(u,w) -<Z>(u,w')llt < — wZ^+ * (3.10) We can now apply Theorem A.2.1 and obtain the existence of a unique fixed point x = Ф(и,х), for every given control function u. By the definition of Ф, this provides a solution to the Cauchy problem (3.5). In general, if f is only continuous w.r.t. u, under the assumptions of Theo- rem 3.2.1, the map u(-) •—> a?(-,u) may not be Lipschitz continuous from L1 into C°. This can be easily seen from the example i(t) = f(t,x,u) = u1/3, #(()) = 0, u(t) e [0,1]. The input-output map, however, turns out to be always Lipschitz continuous w.r.t. the distance d(u, v) = meas {t E [0, T]; u(t) v(t)} (3-11) on the set of admissible controls Id = {и : [0, T] i—> U; и measurable} . (3-12) Proposition 3.2.2. Let T > 0 be given, and assume that f, U satisfy the basic hypothesis (H*). Then the input-output map u(-) •—> x(-,u) is Lipschitz continuous from the complete metric space 1Л = {и : [0, T] »—> U; и measurable} , with the distance d(-,->) defined at (3.11), into C°([0, T]; IRn).
40 3 Control Systems Proof. Identifying controls which differ on a set of measure zero, it is clear that U is a complete metric space. Observing that ||u — v||li < d(u,v) • max |o> —o/|, from the proof of Theorem 3.2.1 it follows that the transformation Ф defined at (3.8) maps U x C° continuously into C°, and that (3.10) holds. Therefore, all assumptions of Theorem A.2.1 hold, where the metric space (U,d) plays the role of A and where C° with the equivalent norm || • || f plays the role of X. Define the constant M = max {|/(t,#,u?)|; t G [0,T], |a?| < |^| + CT. cuGU}. Using (A.6) with у = :r(-,u), Л = v, we obtain ||.7,-(-. u)-®(-,v)||f < 2||ar(*,«)-#(v,x(-,u))||t = 2 max e-2Lt / f(s, x(s, u),u(s)) ds — / /(s,x(s, u), v(s)) ds t€[0,T] Jo Jo <2 / |/(s,x(s,u),u(s)) — f(s, x(s,u), v(s))| ds Jo < 2T-2Md(u,v}. Hence, for any u. v G U one has max |x(t,u) — j:(t,v)| < ^MTe2LT • d(u, v), te[o,r] proving the Proposition. If in place of (H*), the control system satisfies the weaker hypothesis (H), for a given control и : [0, T] U the solution x(-,u) of (3.5) may not be defined on the entire interval [0, Т]. Indeed, as in figure 2.2, the norm of the solution \x(t. u)\ may approach -hoc, or else (t,;r(t,u)) may approach the boundary of the domain C, before time T. From Theorem 2.1.5 we immediately obtain an estimate on the solutions to the Cauchy problem (3.5). Theorem 3.2.3. (A-priori bounds on trajectories). Let f : J?xU »-► IRn satisfy the basic assumption (H). Let ф : [to, ti] x IR •—> R be a scalar function, measurable in t and continuous in x, such that max |/(t, x, a?)| for all t,r. (3.13) |x|=r, iveu Let 11—> r(t) be an absolutely continuous function such that r(t)>-0(t,r) for a.e. tc[to,U], r(to) > |#| • (3-14)
3.2 Fundamental properties of trajectories 41 If the set К = {(t,x) : 0 < t < T, |x| < r(£)} is contained in L2, then for every admissible control и : [0, T] i—> U the Cauchy problem (3.5) has a solution #(•) defined on the entire interval [0, T], which satisfies |^(£)|<r(£), for all t 6 [0, Т]. (3.15) Proof. Given any control и : [0, T] i—► U, consider the function g(t,x) = f(t, x, u(tY). The result then follows from Theorem 2.1.5, applied to the Cauchy problem (3.6). As in Corollary 2.1.6, as a special case we obtain Corollary 3.2.4. In addition to the hypothesis (H), assume that f : [0, -Foo) x ]Rn x U satisfies \f(t, x, u)| < C(1 4- |ж|) for all t,x,u. Then, for every admissible control ti(-), the solution to (3.1) with x(0) = x satisfies |яг(£,и)| < ect|.t| + (eCt - 1) . (3.16) For future applications, we now prove a further related result: if the solu- tion t i—> x(t,u) corresponding to a given control w(-) is well defined on the whole interval [0, T], then the same holds for all controls v(-) sufficiently close to и in the L1 norm. Proposition 3.2.5. Let the basic hypothesis (H) hold. Let the solution x(,u) of (3.5) corresponding to the control и be defined on the entire interval [О, Т]. Then there exists p > 0 such that (i) For every vEU with ||zz - v||u < p, the trajectory #(-, r) is defined on the entire interval [0, T]. The map v(-) »—> x(,v) is continuous from Ll into C°. (ii) For every control in the set Up = {t'E U\ d(u,v) < p}, the trajectory x(-,v) is defined on the entire interval [0,Т]. The map ?;(•) f-> x(,v) is Lipschitz continuous from Up into C°, w.r.t. the distance (3.11). Proof. Construct a smooth cut-off function ф : IR x IRn н-> IR such that ф = 1 on a neighborhood Af of the graph of the trajectory {(t, x(t, u)); t e [0,T]}, while ф = 0 outside a compact set К C 12. Then the function ft (4 «)/(*, a:,u) if (*,*) € 12, (r. satisfies the global bounds (H*). Hence the previous results apply to /f. By the continuous dependence proved in Theorem 3.2.1, for all controls v(-) with ||v ~ ^IIl1 sufficiently small, the solution 1i—► x\t,v) of x = f\t, x, v), z-(0) = x (3.18)
42 3 Control Systems is defined on [0, T] and remains inside J\f. The same holds whenever the dis- tance d(u, v) defined at (3.11) is sufficiently small. Since f = onff, x\-,v) coincides with the solution x(-,v) to the original problem (3.5). From The- orem 3.2.1 we thus deduce the statement (i), while (ii) is a consequence of Proposition 3.2.2. The continuous dependence of trajectories on the control function и is a basic result. However, in the analysis of optimal control problems, stronger regularity properties are needed. In the next theorem we consider a reference control и and a one-parameter family of perturbed control functions, of the form u£ = и + eAu. If f is differentiable also with respect to the control values, we show that the corresponding family of trajectories x(•,?/ + eAu) is differentiable w.r.t. the parameter e. Theorem 3.2.6. (Differentiability w.r.t. the control). In addition to the basic hypothesis (H), assume that f is defined on V, with V open neighbor- hood ofXJ, and is continuously differentiable w.r.t. u. Let u(-) EU be a control whose corresponding solution #(-, u) of (3.5) is defined on [0, Т]. Then, for ev- ery bounded measurable Au(-) and every t E [0, T], the map e h-► x(t, u + eAu) is differentiable. Its derivative at e = 0 is d —-x(t, и T еЛ?/)|е=0 = / M(t, s)Duf (s, x(s, u), u(s)) • Au(s) ds. (3.19) Jo Here Duf denotes the n x m matrix of partial derivatives dfi/duj, and M is the matrix fundamental solution for the linearized problem v(t) = Dxf(t,x(t, u),u(t)) • v(t). (3.20) Proof. The result will first be proved under the additional assumptions (H*), then in the general case. 1. Call z(t) the right hand side of (3.19). Recalling Theorem 2.2.3, z is a solution to z(t) = A(t)z(t) T Duf(t, x(t, u), u(t)) ♦ Au(t), z(0) = 0, (3.21) with A(t) = Dxf(t,x(t,u),u(t)). Next, define x£(t) = x(t, и + eAu), ?/c(t) = x(t,u) + Ez(t). (3.22) To prove the theorem, we need to show that lim =0 (323) £—о г 2. Observe that же(-) is the fixed point of the map w i—> Ф(и+еДи, w) defined at (3.8), which is contractive with respect to the equivalent norm || • ||| defined
3.2 Fundamental properties of trajectories 43 at (3.9). Thinking of as an approximate fixed point, using the estimate (A.6) in the Appendix with к = we obtain 1 2 -|ke-3/e||t < -||Ф(ы + еД«, 3/e) — 3/ellf- To prove (3.23) it therefore suffices to show that lim I sup - £_>0 UG|O,T] £ = 0. (3-24) 3. Recalling (3.21) and the definition (3.22) of yc, we obtain 1 E 4- ez(s), u(s) 4- eAu(s)^ ds — x(t) - ez(t) 4- j Duf x(s,u), u(s)j - eAu(s) ds — x(t) ~ £z(t) 4- / / l/9r/(.s, x(s. u) + crs2(s), u(s) + aeAu(s)) Jo Jo L —Dxf(s, x(s,u), ?i(s))j • ez(s) dads + J [z)u/(s, x(s, u) + aez^sy u(s) 4- aeAu^s)) —Duf(s, x(s,u), u(s eAu(s) dads У У ||dx/(s, x(s,u) -I- aez(s), u(s) 4- aeAu(s)) —Dxf(s. x(s,u), u(s)) | • |z(s)| dads 4- У У ||£>u/(s, a;(s, u) 4- aez(sy u(s) 4- aeAu^s)) —Duf(s, x(s,u), tz(s)) • IAu(s)| dads . (3.25) By the Lebesgue Dominated Convergence Theorem, the right hand side of (3.25) converges to zero, proving (3.24) and hence (3.23). This completes the proof under the additional assumption (H*). 4. To cover the general case, it suffices to consider an auxiliary function defined as in (3.17), which satisfies (H*) and coincides with f for all (t,x) in a neighborhood of the graph {(£,a?(t, u)); t e [0,T]}. The result then holds for the system (3.18), hence for the original system (3.5) as well.
44 3 Control Systems 3.3 Closure In this section we study one of the key qualitative properties of the set of trajectories, namely its closure. Consider a sequence of admissible controls uy and assume that the corresponding solutions .?(•, of (3.5) converge to #(•) uniformly on [0,T]. Our main concern is whether this limit trajectory is a solution of the original control system. This is the case if we can find a control u(-) such that x(t) = f(t, x(t), u(t)) for a.e. time t. In general, this may not be the case. Indeed, even if the trajectories xy(-) converge, the controls Uy(/) might have a highly oscillatory behavior and not converge in L1. Example 3.1 . Consider the system on IR: x(t) = u(t), z(0)=0, u(t) e {-1,1} a.e. (3.26) For t e IR, define 1 if sin(i/t) > 0, —1 if sin(i/t) < 0. As shown in figure 3.2, the sequence of trajectories t ь-> x(t, uy) converges to zero uniformly for all t e IR. However, .t(£) = 0 is not a solution of (3.26). Fig. 3.2. A sequence of highly oscillatory controls and their trajectories. The closure of the set of trajectories is best studied within the framework of differential inclusions. Theorem 3.3.1. (Closure of the set of trajectories). Assume that the multifunction (t,x) »—> F(t,j?) is Hausdorff continuous on IR x IRn with compact convex values. Then the set of trajectories of (3.4) is closed in C°([0,T];Rn).
3.3 Closure 45 Proof. 1. Let xp(-) be a sequence of trajectories of (3.4) tending to #(•) uni- formly on [0,T]. Since the sets F(t, x) are uniformly bounded as Lx range in a compact domain, the xI/(-) are uniformly Lipschitz continuous. Therefore, the function x(-) is Lipschitz continuous as well, hence differentiable a.e. on [0,T]. To prove the theorem, we thus need to show that x(t) G F(t,x(t)). (3.28) at each time r where the time derivative x(t) exists. 2. Assume, on the contrary, that x(t) exists but the inclusion (3.28) fails. Since the set of velocities F(t, x(r)) is compact and convex, by Lemma A.8.5 in the Appendix the two sides of (3.28) can be strictly separated by a hyperplane. As shown in figure 3.3, there exists s > 0 and a unit row-vector p G lRn such that p у <p • i(r) — 3s for all у G F(t, x(r)). By continuity, there exists 6 > 0 such that, for \t — t| < 6 and |x' — x(t)| < J, one still has p • у < p • x(t) — 2s for all у G F(t, xf). (3.29) Recalling that the map t *—> x(t) is differentiable at t = t, we can choose r' G ]t, t + 5] such that X^T \— ±(r) < e, |x(t) - х(т) I < 8 for all t G [r, r']. т — т By uniform convergence, we now have On the other hand, for all v sufficiently large the bound (3.29) implies xy(r') -хДт) T' — T ! p • Xy(t) dt <p • x(t) — 2s . T— T This contradiction proves (3.28) Using Theorem 3.1.1, the previous result can be applied to the control system (3.1). Corollary 3.3.2. Let the basic assumptions (H) hold. Let xl/(«) be a sequence of solutions to (3.1) converging to x(-) uniformly on [О, Т]. If the graph {(£,x(t)); t G [0, T]} is entirely contained in L2 and all sets of velocities F(t,x) = {f(t,x,u); и G U} are convex, then x(-) is also a trajectory of the control system (3.1).
46 3 Control Systems F(T, x(t)) Fig. 3.3. Left: the speed x(t) is strictly separated from the set F(t, x(r)). Right: if the sequence of trajectories xv^) converges to rr(-) at time t = r, it cannot converge also at time r'. Remark 3.1. In the above results, the key assumption is the convexity of the sets of velocities F(t,x) C IRn, not the convexity of the set of controls U C IRm. If U is convex and the function f is affine w.r.t. the control variable u, then every set F(£, x} is convex. The most general function f of this type can be written in the form f(t,x, u) — g(t, x) + B(t, x)u where B(t,x) is an n x m matrix, for each given t,x. On the other hand, if U is convex but f is non-linear w.r.t. the variable u, then the sets F(t,x) = {f(t,x,u); и € U} may not be convex, in general. Example 3.2 . Consider the control system on Ш2 (±i,±2) = (zz, 1 - u2) и e U = [-1. 1]. Here U is convex. However, consider the sequence of rapidly oscillating con- trols u^t) as in (3.27). Starting from the origin at time t = 0, the corre- sponding trajectories t •—> x^t) = х(1,и„) converge to the null trajectory = (0,0). uniformly w.r.t. t. However, this is not a trajectory of the system, for any control u(-). Here the non-convexity of the set of velocities F{t,x) = {(s/1,2/2); 2/2 = 1 — 3/1 , У1 € [-1,1]} yields a set of trajectories which is not closed.
3.4 Density 47 3.4 Density In this section we study what happens if the set U of admissible controls is re- placed by a smaller one. For example, instead of using all measurable controls и : [0, T] •—> [—1,1], suppose we can only use piecewise constant controls, or, say, controls taking only the two values +1, —1. Do these limitations substan- tially reduce our ability to control the system? Two results in this direction are presented below. Theorem 3.4 .1. (Density of trajectories with piecewise constant con- trols). In connection with the Cauchy problem (3.5), let f, U satisfy the basic hypothesis (H). Then the family of trajectories corresponding to piecewise con- stant controls is dense in the set of all solutions (with measurable controls). Proof. Given a measurable control и 6 U, assume that the corresponding solution x(-,u) of (3.5) is defined on [0, Т]. Construct a sequence of piecewise constant controls uy G Id converging to и in L1. If f satisfies the global bounds (H*), by Theorem 3.2.1 the corresponding trajectories x(-,uy) converge to x(-,u) uniformly on [0, Т]. To prove that the uniform convergence holds also in the general case, it suffices to consider an auxiliary function f\ defined as in (3.17), which satisfies (H*) and coincides with f for (t, x) in a neighborhood of the graph {(t, x(t, w)); t 6 [0, T]}. The result then holds for the system (3.18), hence for the original system (3.5) as well. The next result describes under what conditions one can replace the set U of control values by a smaller set U' C U and still be able to approximate all the trajectories of the original system. By cd(S) we denote the closed convex hull of a set S C Rn, i.e. the intersection of all closed convex sets which contain S. Theorem 3.4 .2. (Approximation using a smaller set of controls). Let f, U satisfy the assumptions (H). Consider a subset U' C U such that cd{f(t,x,u)-, и G U'} D {f(t.x,u); и G U} forallt,x. (3.30) The every trajectory of i = f(t,x,u}, x(0) = x, u(t)GU. (3.31) can be approximated by a trajectory of x = f(t,x,u), rr(O) = x, u(t) G U', (3.32) uniformly on bounded intervals.
48 3 Control Systems Fig. 3.4. Constructing a trajectory which remains within the tube Г£ around x Proof. 1. Let t x(t) be a trajectory of (3.31), corresponding to the admissi- ble control t и-► u(t), say, defined for t e [О, Т]. We have to show that x can be approximated by trajectories of (3.32), uniformly on [0,Т]. By Theorem 3.4.1, it is not restrictive to assume that u is piecewise constant, right continuous. 2. Fix a radius p > 0 arbitrarily small, and consider the tube around the reference trajectory x By the assumption (H), there exists a Lipschitz constant L such that u) — f{t,x\u)\ < L\xfx'\ for all (t, x), (£, x') ё Г, uCU. Define the positive, increasing function V’(t) = eLt - | . О Choose e > 0 sufficiently small so that £^(T) < p. Note that this implies that the tube Ге = |(t,x); t € [0,T], |rr - x(Z)| < s^(t)} . (3.33) is contained inside Г. 3. The theorem will be proved by constructing a piecewise constant control и : [0, T] U' whose trajectory remains inside the tube Ге. namely \x(t, u) - x(t)\ < s^(t) for all te[0,T]. (3.34) The main constructive procedure is shown in figure 3.4, At time to = 0 we choose an arbitrary control value uq € U'. We define u(P) = Uq until the first
3.4 Density 49 time ti when the trajectory x(-) hits the boundary of Д, We then choose a new control value Ui € U' such that, at time t = ti, the velocity ±(ti) = f(ti, rr(Zi), щ) points strictly inside the tube Г£. We take u(t) = u\ until the next time, say t2 > £1, when the trajectory hits again the boundary of Г£, etc... In the following, we describe the inductive step more in detail, and show that in a fine number of steps one can cover the whole interval [0,T]. 4. Assume that the piecewise constant control и has already been constructed on the time interval [0, ti] and the corresponding solution satisfies \x(t) — x(t)| < for all t € [0, ti]. Define the unit vector pi = and choose Ui e U' such that щ), pi) > (f(ji, x(ti), u(ti)), pi) - . (3.35) In other words, among all possible velocities, we choose one which has almost the largest possible inner product with the unit vector pi, in the direction of x(ti) — x(ti). Notice that an element щ G U' satisfying the inequality (3.35) certainly exists, because of the key assumption (3.30). 5. We now extend our solution rr(-) beyond time ti using the constant control u(t) = Our main concern is the distance between x(t) and the reference solution x(t). Using (3.35) and the Lipschitz continuity of f w.r.t. x, at the time t = ti the derivative of the distance can be estimated as follows: ft{ l*(0 -*(01 ,+ = (/(*», «(*<)) - /(£», , pi) - eLeLti < x(ti), - f(ti, x(ti), Ui) , Pi) + i(ti), щ) — /(ij, x(ti), -u»)| - eLeLti < + Le (eLti - I) - eLeLti ____eL -___3 ’ Therefore, on some open interval ]ti, ti + the solution remains strictly inside the tube Г£. We can thus define u(t) = щ on the interval [ti, tj+ih where is the first time > ti when the solution satisfies |x(t) = e&(t). At time t = we choose a new control value , and so on. 6. To complete the proof, we have to make sure that, in a finite number of inductive steps, we can cover the whole time interval [0, Т]. By assumption, the control й is piecewise constant, say u(t) = uj € U for t G Ij = [tj, Tj+i[, j = 1,... ,N. It thus suffices to show that, in a finite number of steps, we can
50 3 Control Systems cover each of the intervals Ij. For every boundary point (t, y) G дГ£, with \y — £(т)| = £^(t), t G Ij, there exists a control w G U' such that the solution to the Cauchy problem x = f(t,x,w), x(r) - у, satisfies |x-(t) — x(t)\ < for all fe]r, t + J[, (3.36) for some 6 — \T.y) > 0 depending on r, y. By continuity, the same control can be used for all points (rf,yf) 6 дГ£ sufficiently close to (r,y). The corresponding length in (3.36) depends continuously on rf,yf. We can now cover the compact set |(т,у); т 6 [г,, t>+1] , |y - i(r)| with finitely many open sets Jt where the function 5 = \r,y) remains uni- formly positive. Therefore, in the previous construction, we can always achieve > ti T 6 for some fixed 5 > 0. The inductive procedure thus terminates after a finite number of steps. An important case where the key assumption (3.30) holds is the following. Assume that the control system is linear w.r.t. u: x = h(t, x) + A(t, x) - u, each A(t,x) being an n x m matrix. Assume U' C U and call F,F' the corresponding multifunctions, defined as in (3.3). Then, by linearity, coU'DU => cdF'(t,x) D F(t,x) for all t,x. (3.37) The implication (3.37), however, is usually false when f is nonlinear. Example 3.3 . Let f.g be smooth vector fields on lRn. Then the set of tra- jectories of x = f(x) + g(x)u(t) a:(0) = 0, u(t) G { — 1,1} a.e. is dense in the set of trajectories x = f(x) + g(x)u(t) rr(O) = 0, u(t) G [— 1,1] a.e., because co{ —1,1} = [—1, 1] and the velocity x is a linear function of u. Example 3.4 . On IR2, consider the systems
3.5 Reachable sets 51 (ii,±2) = (u, 1 - it2), (^1,ж2)(0) = (0,0), u(t) € U = [-1,1], (3.38) (±i,i2) = (tz,l -zz2), (a?i,x2)(0) = (0,0), u(t) e U' = {-1,1}. (3.39) Then coU' = U, but the set of solutions to (3.39) is not dense in the set of solution to (3.38). Indeed, taking w(f) = 0 one obtains a solution to (3.38): (xi,x2)(t) = (0,t). This solution cannot be approximated by trajectories of (3.39), because u(t) G {-1,1} implies ±2(t) = 0. In this case, the key assumption (3.30) fails. 3.5 Reachable sets In this section we consider a control system whose dynamics is independent of time: ± = /(ж, ?z), a:(0)=^, w(-)eZ7, (3.40) where the set U of admissible controls is given by (3.2). The reachable set R(r, x) at time t, starting from x, is then defined as R(r, x) = < x(t) ; rr(-) is a solution of (3.40) with a?(0) = x Fig. 3.5. The reachable set at time т > 0, starting from the point x. More generally, given a set К C IRn of initial states, we define R(t, K) as the reachable set at time t, starting from points in K: R(r,K) = |z(t) ; x(-) is a solution of (3.1) with x(0) € k\ . The next theorem establishes the closure of the reachable sets, under a suitable convexity assumption. In a later chapter, this closure property will be of great importance, providing the existence of optimal controls.
52 3 Control Systems Theorem 3.5.1. (Closure of the reachable set). Let f, U satisfy the basic assumption (H). Assume that the graphs of all solutions of (3.40) starting from any point x 6 К are contained in some compact set К' c L2, for t G [О, Т]. If all sets of velocities F(x) = {f(x,w); tv € U} are convex, then, for every т G [0,T], the reachable set R(t,K) is compact. Proof. Let be a sequence of points in R(r, K) tending to By defini- tion, for each v we have £y = xu(r) for some solution to (3.40) with хДО) G K. By assumption, the graphs of these solutions remain inside the compact set K'. Therefore the derivatives xy are uniformly bounded and the solutions xy(-) are uniformly Lipschitz continuous. According to Theorem A.4.1 we can extract a subsequence which converges to a limit function x(J uniformly on [0,т]. Clearly x(0) = lim^—oo хДО) G К because К is closed. Moreover, be- cause of the convexity of the sets F(x), by Corollary 3.3.2, there is a control function zz(-) G IL such that x(t) = /(x(t),u(f)) for a.e. t G [0, т]. We now have £ = lim £y = lim ху(т) = x(r) G R(t, K), u—►ОС V—»oo showing that the reachable set R(r,K) is closed. Example 3.5 . On IR2 consider the system (see figure 3.6) (±i, ±2) = (^, ^i), Cei,#2)(0) = (0,0), u(t) G U = {-1,1}. (3.41) On a fixed interval [0, T], consider the sequence of rapidly switching controls defined at (3.27). The corresponding trajectories satisfy xi(£,u„) = / uy(s)ds —> 0 uniformly on [0,T], Jo X2(t>uiS) = / [^i(s,Wx/)] ds —> 0 uniformly on [0, Т]. Jo However, the limit trajectory x(t) = (0,0) is not an admissible solution of the system (3.41). Indeed, if t ► (xi (t), x2(t)) is a solution, then X\(t) G { — 1,1} implies X\(t) / 0 at almost every time t. Hence x2(T) = f x2(t)dt>0. Jo We conclude that the reachable set /?(т) starting form the origin is not closed, for every т > 0. In general, it is impossible to give an explicit formula for the reachable sets. In certain cases, some information can be obtained by comparing the reachable set with the sub-level sets of a given function ф : Hn h-> IR.
3.5 Reachable sets 53 Fig. 3.6. Each set of velocities F(x) C IR.2 for the system (3.41) contains exactly two vectors and is not convex. A highly oscillatory control produces a trajectory xu that remains close to the origin, but always has a strictly positive second component. The reachable set R(t) is not closed because it does not contain its lower boundary. Theorem 3.5.2. (Outer estimates on the reachable sets). Consider the control system (3.J0), satisfying the basic assumptions (H). Let ф : IR7'i—> IR be a C1 function such that \7ф(х) • /(a?, u) < 1 whenever </>(rr) € [0, T], и G U . If the set К of initial states satisfies К C {z ; ф(х) < 0} , (3.42) then for every т G [0, T] the reachable set R(t, K) is contained in the corre- sponding sub-level set of ф, namely (see figure 3.7) R(r, К) С {x; ф(х) < г} . (3.43) Fig. 3.7. The reachable set R(r, K) is contained in the level set {ф(х) < r}. Proof Let x(-) be any solution of the control system, starting from a point xtK. Using the chain rule, we obtain ^</>(x(t)) = V0(x(i)) • i(t) = V0(x(t)) • /(x(i), u(f)) < 1. at
54 3 Control Systems Therefore </>(х(т)) < 0(ж(О)) + [ Jo ) dt<r. This proves (3.43). Theorem 3.5.3. (Inner estimates on the reachable sets). Consider the control system (З.^О), satisfying the basic assumptions (H). In addition, as- sume all sets of velocities F(x) = {f(x, ca); co G U} are convex. Let ф : IRn 1—> IR be a C1 function whose level sets are bounded and such that max Vd(a;) • f(x,cS) > 1 whenever t G [0,T], ф(х) E [0,T]. If the set К of initial states satisfies К 3{x; ф(х) < 0} , (3.44) then for every т G [0, T] the reachable set R(r,K) contains the corresponding sub-level set of ф, namely R(r, К) Э {37; 0(x) < t} . (3.45) Fig. 3.8. Construction of a trajectory reaching the point y. Proof. 1. Consider any point у G JRn such that ф(у) < т. To prove the theorem, we need to construct a trajectory #(•) of the control system such that a:(0) G К and x(r) = y. Because of the assumption (3.44), it suffices to construct a trajectory t x(t) such that z(t) = y, ф(х(1)) < t for all t G [0,t] . (3.46) 2. As an intermediate step, given any £ > 0, we will construct a solution such that я(т) = у, 0(z(t)) < t + e(r - t) for all t G [0, r] . (3.47)
3.5 Reachable sets 55 This will be achieved defining a piecewise constant control function, starting at t = т and working backwards in time. Consider the tube-like domain (see figure 3.8) Г€ = | (£, ж); ф(х) < t + е(т - t) |. Set = t. By assumption, there exists a control value uq 6 U such that (3.48) By (3.48) and the continuity of V^>, the solution to the backward Cauchy problem x = f(x,u0), я(т) = У satisfies V0(s(t)) • /(rr(t), u0) >1-6 on some open interval ]t — 5, t[. Therefore </>(a;(f)) < t + e(r — t) t € ]t - 5, t[ . (3.49) We then set u(t) = uq on an interval [ti, to], where ti is the first time where the solution hits the boundary of the tube Te, i.e. ti — inf < r; </>(x(ty) < t + 2s(t — t) for all t e [t',r] |. We then choose a control value щ such that V</>(x(ti)) • /(x(ti),ui) > 1. This guarantees that the backward solution to x = /(ж, satisfies (3.49) on some open interval ]ti — 5, t± [. The solution can thus be prolonged backwards up to the next time t2 < ti where it hits the boundary of the tube Ге, etc... By a uniform continuity argument, the entire interval [0, t] can be covered in a finite number of steps, so that 0 = t?\r < ♦ • • < £2 < < to = t, for some integer N. This yields a solution of the control system which satisfies (3.47). 3. Since £ > 0 in (3.47) was arbitrary, we can now consider a sequence £m —► 0 and construct solutions rrm(-) so that Хт(т) = у , фт(0) < t + £m(r - t) for all Ш>1, t € [0, t] . (3.50) By the boundedness of the sub-level sets of ф, all these solutions are uni- formly bounded, hence their speeds |±m| remain uniformly bounded. By the compactness Theorem A.4.1 in the Appendix, we can find a subsequence which converges to some function t •—> x(t) uniformly on [0, г]. According to Theo- rem 3.3.1, this limit trajectory x(-) is itself a solution of the control system (3.40). By (3.50) and the continuity of ф, it is clear that (3.46) holds. This completes the proof.
56 3 Control Systems 3.6 Linear systems In this section we analyze in greater detail the case of a linear system : x — A(t)x + B(t)u, x(0) — x, u(t) € U. (3.51) Here x e IRn, U C IRm, while A(t) and B(t) are respectively n x n and n x m matrices, continuously depending on t. Calling M(t,s) the matrix fundamental solution for the homogeneous problem x = A(t)x, (3.52) the solution of the Cauchy problem (3.51) can be written as x(t) = M(t, 0) x + I M(t,s) B(s)u(s) ds. (3.53) Jo We begin by examining the case where x = 0 and the set U is the entire space IRm. The reachable set at a time t > 0 is then described by f ft >| R(t) = \ M(t,s) B(s)u(s)ds; и e 1?([0,*]; IRW) к (3.54) I Jo J Lemma 3.6.1. For each t > 0, The reachable set (3.54) for the linear system (3.51), starting at x = 0, is a vector subspace of№\ Proof. Consider any two points Xi, x? € 1R(£). Assume that these can be reached at time t using the controls ui(-), U2(-)- Then, for any Ai,A2 € IR the point x = Ai^i T А2Ж2 can be reached using the control u(t) = Apui(t) + A?^^)- Indeed, by linearity M(t, s) B(s) (Aitii(s) + A2u2(s)) ds = Ai M(t, s) B(s) U1(s) ds + A2 M(t, s) B(s) u2(s) ds = A;Xi T A2#2 • Complete information on the reachable sets R(t) can be obtained for the linear system x = Ax + Bu u(t) e IR/n (3.55) where the matrices A, В are constant in time. In this case, the controllability matrix is defined as the n x nm matrix C(A,B) = (B, AB,..., An~} B). Theorem 3.6.2. (Reachable subspace for a linear system). For every t > 0, the reachable set for the linear system (3.55) starting at the origin is precisely the subspace spanned by the columns of the n x nm controllability matrix C(A, B).
3.6 Linear systems 57 Proof. 1. In the case of constant coefficients, the matrix fundamental solution takes the form M(£,s) = wjiere ifc Лк ЛА = 1 Л k\ k=0 Therefore 7?(t) = {f'e^^Bu^ds-, L^fO^R”1) (3.56) 2. We recall that и € IRm is a column vector, while В is a n x m matrix. Let bi,..., bm be the column vectors of B, so that В = I by (3.57) We need to show that 7?(t) = span/ Akbj ; к = 0,..., n — 1, j = Since we already know that R(t) is a subspace of IR71, to prove the theorem it suffices to show that, for every row-vector p e IRri, p • x = 0 for all x e R(i) (3.58) if and only if p Akbj = 0 for all к = 0,1,... ,n — 1, j = (3.59) 3. Assume that (3.59) holds. By the Cayley-Hamilton theorem, the matrix A is a root of its characteristic polynomial. Hence there exist real numbers co, Ci, ... , cn_i such that n—1 Ап = ^аА*. i=0 By induction on k, it follows that every matrix Ak can be written as a linear combination of the n matrices A0 = Z, A, A2,..., An-1. From (3.59) it thus follows p • Akbj =0 for all к > 0 , j = 1,..., m. In turn this implies p . etA bj = 0 for all t > 0, j = 1,..., m,
58 3 Control Systems and hence /»t pt m P‘x(t,u) = p l e^~s^ABu(s) ds = / • e^~s^A bj Uj(s) ds = 0 Jo Jo J=1 for every control function и = (ui,..., um). Therefore (3.58) holds. 4. Next, assume that (3.58) holds. Fix any j € {!,...,m} and choose the vector-valued control function u(t) in (3.57) so that zzj(Z) = l, = 0 for i^j. The assumption (3.58) now implies p-x(t) = 0. Differentiating several times this identity w.r.t. t we obtain 0 = = P,i(0=p-(^4a;(0 + bj)> d^ 0 = -^\p-x(t)] = p-x(t) = p Ai(t) = p A(Ax(t) + bj), dk dk ° = = p'dtkx^=p'A ^Ax^ + b^‘ Notice that here к > 0 is an arbitrary integer. At time t = 0 we have x = 0 and the above identities take the simpler form 0 = p • 5j , 0 = pAbj, ... , 0 = p-Afc-1fy, ... Therefore (3.59) holds. We say that the linear system (3.55) is completely controllable if, for every t > 0, the reachable set R(t) starting from the origin coincides with the entire space IRn. From the above theorem we immediately obtain Corollary 3.6.3. The linear system (3.55) is completely controllable if and only if Rank(B, AB, A2B, ... , Лп-1 в) = n. (3.60) Indeed, the dimension of reachable subspace R(t) equals the number of lin- early independent column vectors in the controllability matrix C(A, B). Hence Rft) = Rn if and only if the rank of this matrix is n.
3.7 Local controllability of nonlinear systems 59 Remark 3.2. If the linear system (3.55) is completely controllable, then from any initial point xq 6 IRn we can reach every other point x\ at a given time t > 0. Indeed, let u(-) be a control function that steers the system from the origin to the point x\ — etAXQ. This means #(£) = [ a)ABu(s)ds = • Jo Using this same control function u(-) in connection with the initial data z(0) = xq, at time t the system is steered to the point x±. Example 3.6 Consider the linear system x = Ax -Ь Bu. #(0) = 0, x e IR3, и € IR, where /-1 0-l\ / 0\ A = I 0 1-1 I , В = I 1. \ 0-1 1/ \-i/ The controllability matrix is given by: / ° i 1\ C(A,B) = 1 2 4 . \-l -2 -4/ (3.61) The rank of C(A, B) is two, thus the system is not controllable. Let us find a row vector p = (рьР2,Рз) which is orthogonal to B, AB and A2B\ pB = p AB = pA2B = 0. This yields the system of three linear homogeneous equations P2 - Рз = Pi + 2p2 - 2p3 = pi + 4p2 - 4p3 = 0 . A non-trivial solution is p = (0,1,1). FYom the proof of Theorem 3.6.2, it follows that the reachable set at any time t > 0 coincides with the subspace orthogonal to p, namely R(t) = {(^i,^2,rr3); лг2+^з = 0}« 3.7 Local controllability of nonlinear systems We now consider a general nonlinear system, with dynamics independent of time: x = /(x,u) u(t) e U С Г. (3.62)
60 3 Control Systems Given a point x E IR71, we say that the system is (small time) locally con- trollable at x if, for every t > 0, the set R(t,x) contains a neighborhood of x. Roughly speaking, this means that the system can be steered from x to all nearby points, within a small interval of time. From the global controllability theorem for linear systems, by a lineariza- tion argument one can deduce a result on local controllability, valid for general nonlinear systems. Theorem 3.7.1. (Small time local controllability). Consider the control system (3.40) and, assume that set U of admissible control values contains a neighborhood of the origin 0 E IRm. At a given point x E IRn, assume that (a) /(£,0) = 0, (b) Defining the matrices of partial derivatives of f w.r.t. x and и computed at the equilibrium point (a?, 0) A = Dxf(x, 0), В = Du(x, 0), (3.63) the linearized system x = A • (x — x) 4- В и (3.64) is completely controllable, i.e. A, В satisfy (3.60). Then the system (З.4О) is locally controllable at the point x. Proof. 1. Fix any т > 0. Since the system (3.64) is controllable, there exists n control functions ..., u^ : [0, t] 1—> IRm such that the corresponding solutions X{(t) = A • (xi(t) — z) + Bu^\t), ^i(0) = x, reach n points, say У\ =^i(t), ... , yn — xn(r) . By an approximation argument, as in Theorem 3.4.1, we can assume that the controls are piecewise constant, hence uniformly bounded on the time interval [0,т]. 2. Since U contains a neighborhood of the origin, for every choice of 0 = (01,..., 0n) € IRn with |0| sufficiently small, the control n i=l is admissible, taking values inside the set U. 3. Choosing 0 = 0 E IRm, the solution of the Cauchy problem
3.8 Lie brackets and controllability 61 x = f(x,u), rr(O) = x (3.65) corresponding to to the control u(t) = ua(t) = 0 is the constant function x(t) = x. Next, call t »—> x(t,ue) the solution of (3.65) corresponding to the control ug. According to Theorem 3.2.6, the partial derivatives of the map в х(т, uq) at 0 = 0 are computed by dx(r, ue) dOi Г e(T~s}ADuf(x(s),0) u®(s) ds Jo The previous construction implies \ dGi дх(т, ue) dOn = Rank уi Therefore, by the Implicit Function Theorem, as 0 varies in a neighborhood of 0 G IRm, the image of the map G > х(т, iz#) covers a whole neighborhood of the point x = x(r,uo). Example 3.7 Consider the motion of a forced pendulum, described by the equation x(t) + sina?(t) = u(t) u(t) e [—1,1]. (3.66) Taking Xi = x, X2 = x, we can rewrite (3.66) as a first order system, namely ±1 ±2 % 2 — sinxi + и When x — 0 e IR2. The matrices of partial derivatives in (3.63) are computed The controllability matrix here is (B, AB) = This matrix has full rank, hence the system is locally controllable at the origin. 3.8 Lie brackets and controllability To derive further controllability properties for non-linear systems, some basic tools from differential geometry are needed.
62 3 Control Systems Given two smooth vector fields f and g on ]Rn, their Lie bracket is defined as [/, s) = Dxg • f - Dxf g. In other words, [/, g] is the directional derivative of g in the direction of f minus the directional derivative of f in the direction of g. For various char- acterizations and properties of Lie brackets, we refer to section A. 10 in the Appendix. The set of vector fields over IRn, with the Lie Bracket operation, is a Lie Algebra , i.e. a vector space endowed with a bracket operation [•, •] such that the following holds: (LAI) The bracket operation is bilinear: for every ai, «2 € IR. and vector fields Д, /2 and g, one has [ai/i + a2f2,g] = «1 [fi,g] +«2 [/2,3], Ь> oil fi + CK2/2] = «1 [9, /1] + «2 b> /2]- (LA2) The bracket operation is antisymmetric: for every vector fields f and g, one has LAs] = -b,/l- (LA3) The Jacobi identity holds: for every vector fields /, g and h, one has [/• Ь- Л]] + [fl, [h, /]] + [/г. [f, 5]] = 0. Next, consider a nonlinear control system on IRn, having the special form i - ^Juifi(x), u(t) = (ui(t),...,um(t)) € U, (3.67) i=l assuming that the vector fields f , are C°°. We define the Lie Algebra generated by (3.67) as the smallest subspace £ C C°°, closed for the bracket operation, which contains all vector fields ft, with i — 1,..., m. For a given point x 6 IRn, we also consider the vector space £(x) = {/(я) : / e £} c IR" . We can now state the following: Theorem 3.8.1. (Local controllability for nonlinear systems). Con- sider the control system (3.67), and assume that the control set U C lRm contains a neighborhood of the origin. Fix any initial point x. If £(x) = IRn (3.68) for all x sufficiently close to x, then for every t > 0 the reachable set R(t,x) is a neighborhood of x.
3.8 Lie brackets and controllability 63 Proof. As a preliminary, notice that, by choosing the control и = 0, the state of the system remains constant. This clearly implies R(t,x) C R(t,x) whenever 0 < t < t. Moreover, by possibly choosing a smaller set of controls, we can assume that the set U is symmetric w.r.t. the origin, i.e. и G U if and only if —u e U. The proof will be given in several steps. 1. Let 8 > 0 be given. For each к = 1,..., n we construct a diffeomorphism фк from an open set Vk C IRA to a /с-dimensional manifold Mk C R(ke,x). In particular, this will show that the reachable set R(ne, x) contains an open set Mn = фп(Уп)-> and hence has non-empty interior. We proceed by induction on k. 2. Let к = 1. By (3.68) there exists an admissible control u G U such that 52^-1 Uifi(x) 7^ 0. Let t i—> 7(Z) be the trajectory corresponding to the con- stant control u(t) = u, with 7(0) = x. Choosing Ji G]0, s[ sufficiently small, the restriction of 7 to the open interval ] — <5i[ is a diffeomorphism. We then set Vi =] — #i[, 0i = 7, and Л4Х = 7(] — Ji, <5i[)- 3. By induction, assume that for some к < n we have already constructed a diffeomorphism фк from an opens set Vk to some /с-dimensional manifold Mk = Фк(Ук) C R(ke,x). Assume that, for every x G Mk and every и G IRm, the velocity vector 52^=1 U*-A(x) i4 s always tangent to Mk . By (A.62) it follows that also the brackets [fa, fj](x) are tangent to Mk and, by induction, C(x) is contained in the tangent space to Mk. Since к < n, this yields a contradiction with (3.68). By the previous argument, there exists a point x G Mk and a control u G IRm such that ^2™x Uifi(x) is not tangent to Mk. Since the control set U contains a neighborhood of the origin, we can here choose u G U. By continuity, the vector 52™ X Uifify) is still not tangent to Mk , for every у G Mk sufficiently close to x. To fix the ideas, let x = Фк(т]ъ • • •, flk)- For convenience, we shall use the exponential notation в •—► (exp0/)(x) to denote the solution of the Cauchy problem dw rz . /лЧ — = /(w), w(0) = x. With this notation, we now define (m \ (771 i=l / Here (т/i,..., ?7fc_|_i) range in an open set Vfc+X, with 77^+1 £] — suffi- ciently close to the origin and (771,..., rjk) ranging in a small neighborhood of (Чь • • • >%)• This achieves the inductive step. 4. When к = n, our inductive argument shows that the reachable set Я(тге, x) C IRn has non-empty interior, since it contains the open set Mn =
64 3 Control Systems Фп(Уп)- Choose any interior point я* E A4n and let w*(-) be a control steering the system from the origin to x* at time T = tie. Since the system (3.67) is linear homogeneous w.r.t. the control и and we are assuming that U — —U, the reversed control uf(t) = -u‘(2T-t) e и steers the system from x* back to x, during the time interval t E [T 2Т]. 5. For each у e Л4п, t G [T, 2T], call t i—> x(t,y) the solution to the Cauchy problem m ±(t) = '4 (О Л(®(0) > = у i=\ As у ranges in the open set A4n, the terminal points x(2T, y) cover a whole neighborhood of x. Observing that x(2T,y) E 7?(T, Л4П) C R(2T,x), since T = tie can be arbitrarily small, we conclude that the system is locally controllable. Under very similar assumptions, one can also establish a global result. Theorem 3.8.2. (Global controllability for nonlinear systems). Con- sider the control system (3.67), and assume that the control set U C HU" contains a neighborhood of the origin. Let (3.68) hold for every x E IRn. Then for every initial point x, one has R(x) = (J R(t,x) = !Rn. (3.69) r>0 Proof. To establish the result, it suffices to show that the set R(x) is at the same time open and closed. As before, we can assume that the control set U C JRm is symmetric, i.e. U = — U. 1. Let x* E R(x). Hence x* E R(r,x) for some r > 0. According to Theorem 3.8.1, J?(e, x*) is an open neighborhood of U for every e > 0. Hence R(x) D R(r + £,#*) contains x* in its interior. 2. Now consider any point x* in the closure R(x). Again by Theorem 3.8.1, for a fixed e > 0, the reachable set R(e, aU) intersects R(x). In other words, there exists r > 0 and a point у E R(r, x) П R(e, #*). Let й : [0, t] »—> U be a control steering the system from x to y. and let u* : [0, e] i—> U be a control steering the system from яг* to y. Define the new control и : [0, т + s] i—> U by setting
3.9 Chattering controls 65 1l(t} = I if *€ [°’ТЬ [ -и*(т + е —t) if te]r, т + е]. One now checks that this control tz(-) steers the system from x to a:*. Hence x* e R(r 4- e, x) C R(x), proving that R(x) is closed. 3. We have proved that the non-empty set R(x) C IRn is at the same time open and closed. Since lRn is connected, this implies R(x) = lRn. Remark 3.3. The two previous controllability results refer to systems of the form (3.67). In this case, if the set U of admissible control values is a symmetric neighborhood of the origin, then the sets of admissible velocities m f(ж) = {52e u| k i=l are also symmetric. Namely, у G F(x) if and only if — у G F(x). A considerably more difficult problem is the controllability of systems with drift, having the form m x = 7o(z) + , «=1 u(«) = (ui(t),...,um(t)) e u. For results in this direction we refer to [87], [52], [57], or to the monograph [56]. 3.9 Chattering controls In cases where the sets of admissible velocities F(t,x) = {f(t,x,u); и G U} are not convex, the reachable sets R(t) may not be closed. We next describe a natural construction that associates with the system (3.1) an auxiliary system ± e U# for a.e. t, (3.70) in such a way that the trajectories of (3.70) are precisely the solutions to the differential inclusion x — F^(t,x) = coFityX). Since F(t, x) C lRn, by Caratheodory’s Theorem A.8.1 in the Appendix, every point in coF(t, x} can be obtained as a convex combination of at most n + 1 elements:
66 3 Control Systems cdF(t, x) = < (#□,... ,0n) C An, Mi € U for all i > . I i=0 J (3-71) Here An = J 6» - (0O,...,0„); ^0г = 1, 0, > 0 for all г I (3.72) I i=0 J is the standard simplex in IRn+1. Motivated by (3.71), we define the compact set U* = U x ... x U x 4 c ]R(n+1)w+(n+1) (3.73) and consider the control system (3.70) with n Г(?,х,иГ) = fft.x, (uo,...,un,(0o, ...^n))) = 52^/(6,x,Ui). (3.74) i=0 Generalized controls of the form u$ = (uq, ... ,un.O) taking values in are called chattering controls. In practical applications, they can be ap- proximated by rapidly switching the control value u(t) among the values uo(t),..., un(t). Here the length of time during which и = щ should be pro- portional to Gift). The above construction provides a representation of the closure R(t) of the reachable set for (3.5) as the reachable set R^(t) for the ’’chattering” system x = f$(t,x,u?y г?(£) e ий, a?(0) = x. (3.75) Theorem 3.9.1. Let f.U satisfy the basic assumption (H). Assume that the graphs of all solutions to (3.5) on [0,T] are contained in some compact set ГУ с P. Then, for every t С [0, T], the closure R(t,x) of the reachable set for the system (3.5) coincides with the reachable set R$(t,x) for the system (3.75). Proof. 1. From definitions (3.73), (3.74), by Caratheodory’s theorem it follows {f$(t,x,afly, if € Ua} = Рй(£,гг) = cdF(t, x) = co {f(t,x,u); и € U} . (3.76) 2. By Corollary A.8.2, the sets F$(t,x) are compact, convex. Therefore, The- orem 3.3.1 yields the closure of the set of trajectories for the differential in- clusion x e F$(t, x). In particular, the reachable set R^(t, x) is closed. 3. By (3.76), we can apply Theorem 3.4.2 and deduce that the set of solutions of x 6 F(t,x) is dense on the set of solutions of x e 0(t,x). In particular, the closure R(t, x) of the reachable set for the system (3.5) contains R$(t,x). Together with the previous step, this yields the equality R(t,x) = R$(t,x).
3.10 The Bang-Bang theorem 67 3.10 The Bang-Bang theorem If the sets of velocities F(t,x) = {f(t,x,u); и G U} are not convex, then we have seen that the reachable sets may not closed. A noteworthy exception occurs in the case of systems with linear dynamics: x = A(t)x(t) + h(t,u(t)), u(t) e U, x(0)=±. (3.77) Indeed, in this case the application of Lyapunov’s theorem implies that every point reachable with a chattering control can also be reached by an admissible control t »—> u(t) € U of the original system. Th eorem 3.10.1. (Reachable sets for linear systems). Consider the sys- tem (3.77). Assume that U C lRm is compact, A(t) is an n x n matrix de- pending continuously on t, and h : [0, T] x U i—> IRn is continuous. Then for every r € [0,T], the reachable set R(r,x) is a compact, convex subset ofJRn. Proof. 1. By the continuity of A and h and the compactness of the set U, as t e [0,t], all trajectories of the system remain uniformly bounded. In particular, the closure of the reachable set R(r,x) is compact. 2. We claim that the reachable set R(r, x) for the system (3.77) coincides with the reachable set R\r, x) for the chattering system x = A(t) x(t) -T У2 0t(^)^(L w^W), a?(0)=x, (3.78) 2=1 with u^z\t) G U, 0(t) G An for every t G [0, т]. Indeed, fix any point £ G R^(r,x). Let Af(-,-) be the matrix fundamental solution for the linear homogeneous system v = A(t) v. If £ = for some solution of the chattering system (3.78), we have J1 £ = М(т, 0)£ + / У^М(т, s)h(s, u^(sf)ds. 2=1 for some control functions : [0,r] •—> U and coefficients в = (во,... ,вп) G An. We now use the result 14. in the Appendix A.5 (Lyapunov Theorem), with f^(s) = M(r, s)hfs, «W(s)). This provides the existence of nd-1 disjoint sets Jo,... ,Jn C [0, r] such that Jo U • • • U Jn = [0, t] and [ y^ei(s)M(r,s)h(s,u(z\s))ds = I M(r,s)h(s,u^\s))ds. (3.79) 2=0 2=1
68 3 Control Systems If ?/* : [0, t] i—► U is the control defined by = u^(t) for t G Ji, from (3.79) it follows £ = M(r,tyx + / Jo Af(r, s)/t(s, u*(s)) ds. Hence £ is reached at time r by the trajectory of the original system (3.77) corresponding to the admissible control u*(-). This proves that 7?#(т, £) C Л(т, x). Since the converse inclusion is obvious, the two sets coincide. 3. By the previous step, it now suffices to show that the reachable set 1$(т, x) for the chattering system (3.78) is compact and convex. Because of Theorem 3.9.1, the boundedness and the closure of Яй(т, x) are clear. To prove its convexity, observe that t x(t) is a trajectory of the chattering system (3.78) if and only if x(t) G co < A(t)x(t) + h(t, и); и G U > for a.e.t. Equivalently, x(t) — A(t) x(t) G G(t) = co{h(t, cu); w € U} . If now .Ti(-) and ^2(*) are two trajectories of the chattering system, for any Л G [0.1] their convex combination x(t) = Aa?i(f) + (1 — A)j?2(£) provides yet another trajectory. Indeed ±(t) - A(f)x(t) = A(iq(t) - A(/)rr1(t)) + (1 - A)(i2(t) - A(^2«) G G(t) because each set G(t) is convex. In particular, x(r) = Aa?i(r) T (1 — A)a?2(r) G Я#(т, ±), proving the convexity of the reachable set for the chattering system. As a special case, consider a linear system where the admissible controls take values in a convex polytope with vertices uq,... ,aqv £ Rrn x = A(t) x + B(t) и u(t) G U# = co{oq, ... , cj/v} . (3.80) In addition, consider the system x = A(t) x + B(t) и u(i) G U = {uq, ... , oqv} . (3.81) where the controls are allowed to take values only on the vertices of the poly- tope. In this case, the admissible control functions u(-) are called bang-bang controls. Indeed, they must be piecewise constant, jumping between the ex- treme points of LP. By the previous result, the reachable sets for the two systems are exactly the same.
3.10 The Bang-Bang theorem 69 Corollary 3.10.2. (The bang-bang theorem). Assume that the nxn ma- trix Aft) and the n x m matrix В ft) depend continuously on time. Then, for every initial point ж(0) = x and any т > 0, the reachable sets R^(r,x) and R(r,x) for the systems (3.80) and (3.81) are compact, convex and coincide. Indeed, in this case the function h(t, u) = Bft)u is linear w.r.t. u. Hence, for any t, x, the convex hull of the set of velocities Fft, x) = {A(t)x F В ft)aii; i = 1,..., TV} for the system (3.81) is precisely the set of velocities F^(t, x) = {л(£)я + B(t)cd ; ш G co{o?i,...,} for the system (3.80). Problems 3.1. A forced linear pendulum is described by the system x + x = и uft) e IR , (3.82) Write (3.82) as a first order system, taking Xi = x, x% = x. Show that this system is completely controllable. Next, assume that the admissible controls are only those taking values inside the set U = [—1,1]. Show that the reachable set at time т > 0, starting from the origin, satisfies 7?(t) C {(^i, ж2); x2 + x% < r2} . Hint: apply Theorem 3.5.2 with ф(х) = y/x2 + x^. 3.2. As in Chapter 1, consider the system f Xi = и cos 3 , < X2 = U sin $, 3 -au. (3.83) modelling the steering of a car in a parking lot. Assume that the two components of the control satisfy the constraints uft) € [-1,1], a(t) e [-1,1]. and let an initial position (х1(0),а?2(0), #(0)) = (#i,#2,0) be given. Show that the system is locally controllable at the point (zi,^#)- Con- struct explicitly some control functions t uft) and 11—► aft) which steer the system from (^1, 3) to the origin (0,0,0) at some later time т > 0. By a suitable translation and rotation of coordinates, show that the car can be steered from any initial position to any other terminal position.
70 3 Control Systems 3.3. Prove that in Theorem 3.2.3 the inequality (3.13) can be replaced by the weaker assumption <^(Z,r) > sup |x |=r, u?eu /(t, x, cu) 3.4. Consider the control system fii = (x? + l)u p1(0) = 0 |,.^|<1яр (±2 = (1+^1) 1 |x2(0) = 0 I ( )l - Show that the reachable set is compact for t < тг/2, but neither closed nor bounded when t > тг/2. Hint: when t > тг/2, choose (3 > 0 such that arctan/? > % — t. Compute the infimum inf{ar2; (/3,x2) € R(t)} and show that it can not be attained by any admissible trajectory. 3.5. Consider the linear control system (ii, ±2) = (^2, u), u(t) G [-1,1]. Fix any constant A G [—1,1] and a time r > 0. Call xx G IR2 the point reached at time r starting from the origin, using the constant control u(t) = A. According to the bang-bang theorem, the point xx can also be reached using some control t u(t) G { — 1,1} taking values only in the extreme points. Determine explicitly such control function. 3.6. On IR2 consider the system (±1, x2) = ? u2 - a?2), (rri, Я2ХО) = (o, 0), u(t) e U = [-1,1]. (3.84) On a fixed interval [0,7], consider the sequence of rapidly switching con- trols as in (3.27). Prove that, as v —> 00, the corresponding trajectories satisfy Xi(t,uy) 0 , x2(t, Uy) t uniformly for t G [0,7*]. However, show that the function t 1-* (0, t) is not a solution of the system (3.84). In addition, prove that, for every r > 0, the reachable set R(r) C IR2 starting from the origin is not closed. 3.7. Prove a converse to Theorem 3.3.1, namely: if the set of trajectories of the differential inclusion (3.4) is closed, then all the sets F(t,x) are convex. 3.8. Consider the control system (±1, ±2? £3) = (^1 T x2 + и, sin x% + u2 , x\ + sin x2 + cos x% — 1),
3.10 The Bang-Bang theorem 71 with x(0) = 0 e IR3 and |u| < 1. Prove that it is locally controllable at the origin. 3.9. Consider the control system on IRn x — u, t(0)=0, u(t) G U = {o>i,...,cu/v} . Describe the reachable set 7?(t) at any time r > 0. For a given point у 6 R(t), construct a control function и : [0,т] U that steers the system to the point у and is discontinuous at no more than n times, say 0 < ti < t? < ... < tn < r. 3.10. Consider the linear system on IRn x = Ax + bu, rr(O) = 0, u(t) € U = [0,1]. where b is a fixed vector in IRn. Given a continuous control function u(t) : [0, T] i—> [0,1], construct a sequence of bang-bang controls : [0, T] i—* {0,1} as follows. Divide [0, T] into equal subintervals Ik =]tk , tfc+i], tk = к/v, к = 0,1,... On each Д, define 1 if tk < t < tk + 0 if tk + < t < tk-^-i Prove that, as p —* oo, the corresponding trajectories #(-, converge to x(-,u) uniformly on [0,Т]. Hint: consider first the case where T = 1 and the function tt(-) is constant. 3.11. Let f : IRn i—► ]Rn be a bounded, smooth vector field, with n > 2. Show that the system x = /(ж) и u(t) e [—1,1] is not locally controllable in the neighborhood of any point x G IRn. 3.12. Consider the control system (±1,^2) = u(t)e[-i,l]. Show that this system does not satisfy the assumptions of Theorem 3.7.1 at the point x = (0,0). Nevertheless, by a direct computation prove that the system is small-time locally controllable at the origin. Hint: explicitly compute the trajectories corresponding to the piecewise constant controls «(*) = 1 if t e [0, A[ -1 if te [A,r] or u(t) = -1 if t € [0,A[ 1 if t 6 [A, t]
72 3 Control Systems 3.13. Consider the control system щ X2 = (a/2 — t)(l + (f — л/2)Х|)х2, with initial data (xi, #2)(0) = (0,0). Show that for every control function и : [0, \/2] —> U = {-1,4-1} the corresponding trajectory is defined and does not explode. What happens if we take co(t7) as control set? 3.14. Consider the linear control system (3.51), assuming that x — 0 and that the set U C IRW is compact. Fix a row vector p e IRn and call t i—> p(£) the solution to the linear adjoint problem p(*) =-Р(0л(*) > Р(т)=р. Show that every trajectory of (3.51) starting from the origin satisfies u) < / ( maxp(s) • B(s)lu ) ds. Jo J Viceversa, prove that there exists a measurable control function 11—> ?i*(t) such that p - x(r, u*) = / ( maxp(s) • B(s)tu ) ds. Jo \^и / Conclude that, for every row-vector p E IRn, the reachable set at time r satisfies r max p • у = I maxp(s) • B(s}w ) ds . ytR(r) Jo / Letting p range over the unit sphere in IRn, this provides a characterization of the reachable sets for linear control systems. 3.15. For the equation (3.82) of the forced linear pendulum, assume that the control function satisfies the constraint u(t) e [—1,1]. Let R(t) C IR2 be the reachable set starting from the origin. For 0 < t < t', prove that R(t) is a neighborhood of the origin and R(t) G R(t'). Moreover, show that Ut>o/?(0 = Di2. Hint: for any unit vector p 6 IR2, use the previous problem and estimate max p • у. 3.16. Consider the control system ±1 = щ X2 = X1(O) = -l X2(O) = 1 |Uj(Z)| — 1 a-e-, where f is given by
3.10 The Bang-Bang theorem 73 1 ®1 /(*1)= 0 if x\ < 0, /(Xi) = exp if X\ > 0. Describe the reachable set R(f) for t > 0. Is it possible to reach the point (—1,0)? If yes find the corresponding control. Hint: first reach the zone where Xi > 0, then it is possible to change the second coordinate by suitably choosing the control.

4 Asymptotic stabilization Consider a control system described by x = f(x, u). Assume that /(z, 0) = 0, so that x E IRn is an equilibrium point when the null control и = 0 is applied. In general, this equilibrium may not be stable: a trajectory which starts at a point xq « x may get very far from x at large times. For many engineering applications, an important problem is to design a feedback control U = U(x) such that the resulting system x = /(x, u(x)) is asymptotically stable at the point x. In the first section of this chapter we recall some basic stability results in the theory of ordinary differential equation. The second section deals with linear control systems. We prove here the main theorems on global stabiliza- tion, by means of a linear feedback control. Using a linearization technique, in Section 4.3 we prove a similar result on local feedback stabilization, valid for nonlinear systems. We conclude the chapter observing that, even if a control system is glob- ally stabilizable, because of topological obstructions the stability may not be achieved by any continuous feedback. In some cases, it is thus necessary to construct discontinuous feedback controls. A special class of discontinuous controls, in the form of patchy feedbacks, will be discussed in Chapter 9. 4.1 Lyapunov stability Consider the differential equation x = g(x) хЕПп, (4T)
76 4 Asymptotic stabilization let t i—> x(t,xo) be the solution taking initial data ж(0) = xq. We say that the system (4.1) is Lyapunov stable at the origin if the following holds (see figure 4.1). (LI) For every e > 0 there exists S > 0 such that if |tq| < 6 then for every t > 0 we have |я?(£,я?о)| < (L2) For every x^ e IRn we have lirnt_>+oo x(t, x0) = 0. The first condition means that for every assigned ball, the entire solution will remain inside this ball provided that the initial condition is sufficiently close to the origin. Such condition is also called stability. The second condition says that the origin attracts every trajectory of the system. Fig. 4.1. Lyapunov stability. For a general nonlinear system, checking its stability is not an easy task. A standard method for proving Lyapunov stability is to construct a positive function that decreases monotonically along all trajectories of the system. We review here the basic ideas of this approach. Given an open set 12 C lRn, a C1 function V : J? i—► IR , is called a Lya- punov function for (4.1) on 12 if the following conditions hold. i) V is proper, i.e. for every r > 0 the sub-level set {a; : V(z) < r} is compact; ii) V is positive definite, i.e. V(0) =0 and V(x) >0 for every x 0; iii) V is strictly decreasing along trajectories of the system: For x / 0 we have W(a?) • g(x) < 0. Theorem 4.1 .1. If the system (4-1) admits a Lyapunov function on IRn then it is Lyapunov stable. Proof. Let V be a Lyapunov function for (4.1). 1. Let us first prove that for every s > 0 there exists 5 > 0 such that {:r : V(.t) < £} C L?(0,6), where B(0,s) is the ball centered at the origin with radius e. Indeed, assume by contradiction that there exist e > 0 and a sequence x„ such that x„ £ B(0, e) and V(a:p) —> 0. Then by i) there exists a
4.1 Lyapunov stability 77 converging subsequence, that we indicate again by xv, and x B(0,s) such that xu —► x. In particular x 0, but by continuity we get V(x) = 0, thus we reach a contradiction. 2. We now establish (LI). For any given e > 0, by 1. there exists 8' such that {a; : V(x) < <5'} C B(0,s) and, by ii), there exists 8 such that B(0, <5) C {# : V(x) < J'}. Take xq E B(0,5), then У(яо) < 8' and, by iii), V\x(t,xo)) < 0, hence V(x(t,xof) < 8' for every t > 0. Finally x(t,xo) e B(0,e) for every t > 0. 3. Next, for every xq, we establish the limit V(rr(t,xo)) ~> 0 as t —* oo. By iii) the function t V(#(£, rro)) is monotone decreasing. If V(x^t, j?o)) > c > 0 for all t > 0, define A = max |w(x)-<z(ar); V(x) e [c, V(x0)]} • Observe that A is the maximum of a strictly negative function on a compact set, hence A < 0. We now have ^-V(x(t,x0)) - VV(i(t,x0)) -g(x(t,xoy) < A, at hence V(z(£,a:o)) < ^(яо) + At —> —oo as t —» oo . reaching a contradiction. 4. Finally we prove (L2). Fix x$ and e > 0. By i) there exists 6 such that {я : У(ж) < J} C B(0,c) and by 3. V(rr(t,x0)) —► 0 as t —► oo. Therefore there exists T such that for every t > T we have V(x(t,a?o)) < 8, hence x(t,Xfy) e B(0,s). This completes the proof. A complete characterization of Lyapunov stability can be given in the case of a linear system with constant coefficients: x = Ax x e IRn. (4.2) Theorem 4.1 .2. The system (4-%) is Lyapunov stable if and only if all the eigenvalues of A have strictly negative real part. Proof If v is an eigenvector of A whose corresponding eigenvalue A has non- negative real part, then the function 11—> eXtv is a solution of (4.2) which does not approach the origin as t —> oo. To prove the converse, we consider an invertible matrix R such that В = R~1AR is in canonical Jordan form, as in Example 2.1. If all eigenvalues of A have strictly negative real parts, say JfA < — 8 < 0, and their multiplicity is < A:, then
78 4 Asymptotic stabilization ||eM|| = ||ЛегВ7Г1|| < C • tk~xe~St 0 as t —> oc. (4.3) Here the constant C depends on the matrix R. By(4.3), every trajectory t i—> x(fi) = etAx(0) of the linear system approaches the origin. Using Theorem 4.1.1 we can provide a further characterization the set of linear systems which are Lyapunov stable. In the following the symbol A* will indicate the transpose of the matrix A. Similarly, the transpose of the column vector x E Rn is the row-vector ж*. Theorem 4.1 .3. For a linear system (4-2) the following conditions are equiv- alent. (i) All the eigenvalues of A have strictly negative real part; (ii) There exists a symmetric, strictly positive definite matrix P such that A*P + PA = -I; (Hi) There exists a symmetric matrix P such that V(x) = x* Px is a Lyapunov function. Proof, (i) implies (ii). By hypothesis ||eM'eM|| <Ce~£t for some constants C > 0 and s > 0, depending on the matrices R. В in (4.3). We can thus define the matrix I etA* etA dt о by an absolutely convergent integral. This matrix is positive definite, because for any x 0 x^Px = / x*etA etAxdt = / |etj4rr|2 dt > 0 . jo jo Moreover, r+°o г f+oo (ptA*tA\ A*P+PA=( ^A*etA‘etA + etA‘etAA^ dt — J ---^-dt = -I. (ii) implies (iii). Let P be as in (ii). Since the matrix P is positive definite, the quadratic form V(x) = x*Px is proper, positive definite and it satisfies: VV(a?) • Ax = x'PAx 4- x'A'Px = -x*x < 0. (iii ) implies (i). This is a direct consequence of Theorems 4.1.1 and 4.1.2.
4.2 Stabilization of linear control systems 79 Example 4.1 Consider the linear differential equation (4.2) with n = 2 and A = -1 1 0-1 The matrix A has a unique eigenvalue A = — 1 of double multiplicity. We can then apply Theorem 4.1.3. The equation A*P 4- PA = — I can be written as 2pn = ~l, 2p12 = 2p2i =Pu , 2p22 = P12 + P21, and the solution is found to be 1/2 1/4 1/4 3/4 The corresponding Lyapunov function is given by: izz x 1 2 1 3 2 V(x) = -XX + -371^2 + ^2’ 4.2 Stabilization of linear control systems In this section we study the linear control systems x = Ax F Bu. ueIRw, (4.4) where A is an n x n matrix and В an n x m matrix. If all the eigenvalues of the matrix A have negative real part, the system is already stable in connection with the null control и = 0. In the case where the uncontrolled system (4.2) is unstable, our aim is to find a linear feedback U(x) = Fx, with F an m x n matrix, such that the resulting linear system ± = (4 + BF)rr (4.5) is Lyapunov stable at the origin. In this case, the control U(x) = Fx is called a stabilizing feedback. By Theorem 4.1.2, this will be the case if and only if all eigenvalues of the matrix A 4- BF have negative negative real part. The main result given below will show that, if the system (4.4) is controllable, then one can always construct a stabilizing feedback. Toward a general result on stabilization by linear feedbacks, some preliminary lemmas are needed. The next result shows that the space IRn can be decomposed as a sum of a subspace where the system is completely controllable, and another subspace where the control has no effect. Lemma 4.2.1. Let d — rank[B, AB. A2B,..., An~1B] be the rank of the con- trollability matrix for the linear system (4-4)- Then, by a linear change of variables, the control system can be urritten in the form:
80 4 Asymptotic stabilization у = A\y + A2z + B\u z = A3Z, (4-6) where (y, z) € IR6' x JR"~d. Proof. Let Vi be the subspace of dimension d, generated by the columns of the controllability matrix [B, AB, A2B,..., An~lB], and let {^i,..., a basis of Vi. Choose n — d additional linearly independent vectors . ,vn so that {vi,...,vn} forms a basis of IRn, and call V2 the space spanned by the vectors , vn. In terms of the basis {t?i,..., ?;n}, the system (4.4) takes the desired form (4.6). Indeed, the space V\ is А-invariant, namely Ax e Vi for every x € Ц. Moreover, Bu 6 Ц for every и G IRm, hence the component of Bu along the subspace V2 is zero. Lemma 4.2.2. (Pole shifting). If the control system (4-4) ™ controllable, then for any given set of real numbers Ai,..., An 6 IR one can find an m x n matrix F such that the square matrix A + BF has Ai,..., An as eigenvalues. Proof. 1. Assume first that m = 1. In this case the matrix В reduces to a column vector b. The controllability matrix [6, Ab,.... An-16] is thus an n x n matrix which, by assumption, has rank n. We can thus use its columns as a new basis of IRn. With respect to this basis, the system takes the form x = Ax + bu = /0 ••• 0 a0 \ 1 • •• 0 cti \0 1 an_i / (4-7) Notice that here the numbers qq, ..., an_i are precisely the coefficients of the characteristic polynomial of A : n—1 det (А/ - A) = An - £2 aiXJ j=o Indeed, by the Cayley-Hamilton theorem, the matrix A is a root of its char- acteristic polynomial. The vector A(An“1b) can thus be written as a linear combination of b, Ab,..., An~lb in the form (n—1 \ n—1 ^ajAj b^^ajA^b. j=0 J j=0 2. Next, consider the auxiliary linear system x = Ax 4- bu = / 0 1 0 \ 0 0 ••• 1 \«o «1 • • • ^4-1 / (4-8)
4.2 Stabilization of linear control systems 81 Notice that the characteristic polynomial of A coincides with the one of A and A. Therefore, using the basis {b, Ab, ..., An~1b}, the control system (4.8) can be put in the same canonical form (4.7). We conclude that there exists a change of coordinates that transforms the original system (4.4) into the form (4.8). 3. Up to a change of coordinates, it thus suffices to prove the result for the special system (4.8). Given the numbers Ai,..., An, compute the coefficients /3j of the polynomial n—1 (А-Л1)---(Л-Ап) = Лп-^/ЗйЛ J=o Defining the row-vector F = (/?q — qq, /?i — <*1, • • • , /3n-i - »n-i) we obtain A + bF = / 0 1 ••• 0 \ 0 0 ••• 1 \0o 01 • • • 0n-l / The characteristic polynomial of this matrix is computed as n—1 det (А/ - (Я+ bF)) = An - 0jxj j=o By (4.9), its eigenvalues are Ai,..., An, as required. This proves the theorem in the case where m = 1, i.e. the matrix В has just one column. 4. We now deal with the general case m > 1. As a first step, we construct a basis of linearly independent vectors tq, г>2,..., vn 6 IRn such that ui = Buq for some uq e IR™ and v<+i = Avi 4- Вщ (4.10) for some controls щ, u%, ..., ttn-i 6 IR771. This can be done by induction. We start by choosing uq such that Vi = Buq 0. Next, assume that Vi,..., Vk have already been constructed and call W the subspace generated by the vectors Vi,... ,Vk. If к < n, we claim that there exists a control Uk such that Ufc+i = Avk + Buk £ W, so that the induction can continue. In the opposite case, we would have Avk + Bu E W for every и E Rm. In particular, this implies Avk € W and hence Bu G W for every u. By (4.10), this in turn implies Avi E W for every i = l,...,fc. From the above, we conclude that Aw E W for every w E W, and В и E W for every и E IRm. Therefore, all columns of the controllability matrix [B, AB,..., An-1B] lie in W. By assumption, these columns span the whole space IRn. If к < n we thus obtain a contradiction.
82 4 Asymptotic stabilization 5. Recalling (4.10), we now construct an m x n matrix F so that Fvi = Ui, i = 1,..., n — 1. The above equations can clearly be satisfied because the vectors Vi are linearly independent. With this choice, we compute the n x n matrix [V1, (Л + BF)v\, , I.A + BFp-'n] = (V1,...,vn). (4.11) Notice that this is precisely the controllability matrix for the system with scalar control x = (A + BF)x + xiu и € IR . Since the vectors Vi,... vn generate the entire space IRn, this system is com- pletely controllable. By the previous steps dealing with the case m — 1, there exists a row vector f such that the matrix A + BF + iq f has Ai, ... , Xn as eigenvalues. Since = Bizq, the conclusion of the theorem is achieved by taking F = F + uq f. Combining the two previous lemmas we can now prove the main result on feedback stabilization of linear systems. Theorem 4.2.3. (Feedback stabilization of linear systems). Consider the linear control system (4-4)> and let (4-6) be its decomposition, in terms of a fully controllable component у and a non-controllable component z. If all the eigenvalues of the matrix A3 have strictly negative real part, then there exists a linear feedback U = Fx that renders the system Lyapunov stable at the origin. In particular, the conclusion holds when the system (4-4) is completely controllable. Proof. By construction, the linear system on IR'7 у = Aiy + BiU is completely controllable. Using Lemma 4.2.2, there exists an m x d matrix Fi such that all eigenvalues of Ai + ByFi have strictly negative real parts. Implementing the feedback control U = F^y we obtain the linear system = (A1 +nB1F1 42) • (4.12) z J \ о As J \z J 7 Consider the eigenvalues of the matrix in (4.12). These are given by the eigen- values of Aj + B}F[, which by construction have negative real part, together with the eigenvalues of A3, which by assumption have negative real part. Therefore the system (4.12) is Lyapunov stable. We conclude the section with a useful lemma:
4.3 Stabilization of nonlinear systems 83 Lemma 4.2.4. (Hautus) If the system (4-4) ls completely controllable then rank(AZ — A, B) = n for every A 6 IR. In particular rank(A, B) = n. Proof If the conclusion does not hold, there would exists a row vector p 0 such that p • w = 0 for every column vector w in the (n + m) x n matrix (AI — A, B). Then pA = Xp and pB = 0, hence pAkB = XkpB = 0 for every integer к. In particular p • w = 0 for every column vector v in the controllability matrix [B, AB, , An-1B]. In view of Corollary 3.6.3, this yields a contradiction with the assumption of complete controllability. 4.3 Stabilization of nonlinear systems We now consider a general nonlinear system: x = f(x,u) u(P) eUc IRm. (4.13) The first result gives conditions for the existence of a stabilizing feedback on a small neighborhood of the origin, and is obtained by a linearization technique. Theorem 4.3 .1. Consider the control system (4-13), satisfying the standard conditions (H) in Chapter 3. Assume that f is continuously differentiable and that U G IRm is contains a neighborhood of the origin. If (4)/(o,o) = o, (b) Setting 4 = ^(0,0), В = ^(0,0), OX OU the linearized system x = Ax + Bu with is completely controllable. Then there exists a neighborhood of the origin 12 such that the system restricted to (2 admits a continuous stabilizing feedback. Proof. By Lemma 4.2.2, there exists a matrix F such that A = A + BF has all eigenvalues with negative real part. Hence by Theorem 4.1.3 there exists a symmetric matrix P such that A*P + PA = —I. In particular, the quadratic form V (x) = x* Px is a Lyapunov function for the linear system x = Ax + BFx. We claim that V is also a Lyapunov function for the nonlinear system x — f^x.Fx) (4.14) restricted to a small neighborhood of the origin. Indeed, by assumption we can write f(x, u) = Ax + Bu + B(x, u)
84 4 Asymptotic stabilization with Therefore, every trajectory of (4.14) sufficiently close to the origin satisfies V(x(<)) = VK(i) • /(x, Fx) = 2x*P {Ax + BFx + Л(х, Fx)) = —x*x + 2x*PR(x, Fx) < 0 . By Theorem 4.1.1, this shows that the system is Lyapunov stable, restricted to a neighborhood of the origin of the form f2 = {# G IRn ; V(x) < 5} . For linear systems, our earlier results showed that, if the system can be sta- bilized at the origin, then the stabilization can always be achieved by means of a linear feedback, say U = Fx. A similar statement may not true, even locally, in the case of nonlinear systems. Indeed, one can give examples of nonlinear control systems which can be stabilized, but only using a discontin- uous feedback. In a global setting, it is easy to understand how topological obstructions can prevent the existence of a continuous stabilizing feedback. Example 4.2. As in Example 1.1 in the Introduction, consider the problem of steering a boat on a river to a given point P along a shore. Let v(x) be the velocity of the water at point x, and let и be the speed of the boat relative to the water. We seek a feedback control U — ф(х) such that al trajectories of the resulting O.D.E. x = v(x) + ф(х) (4-16) converge to P as t -> oo. It is now easy to see that, if an island is present in the middle of the river, there can be no continuous feedback ф performing the desired task. Indeed, there must be a curve 7 separating points whose trajectories pass to the left or to the right of the island, as in figure 4.2. Fig. 4.2. Non-existence of a smooth feedback. In the above example, the non-existence of a continuous feedback was due to the topological properties of the set L? of admissible states for the
4.3 Stabilization of nonlinear systems 85 control system. One can also identify topological properties of the map f which prevent the existence of a smooth stabilizing feedback. Recall that, by the Hautus Lemma 4.2.4, for every controllable linear system the map (x, u) h-> Ax + В и is onto. A similar necessary condition can be formulated in the nonlinear case (for a proof see [80]) Theorem 4.3 .2. Consider the control system (4-13), assume U = IRm and f : IRn x lRm —► lRn continuously differentiable. If there exists a smooth stabilizing feedback in a neighborhood of the origin, then the map f is open at 0, i.e. for every e > 0 there exists 8 > 0 such that {y\ |y| < <5} c {f(x,u); |x| + |u|<e}. (4.17) Example 4.3 (Brockett) Consider the control system on IR3 x = fi(x)ui + f2(x)u2 (4-18) with /i(xi,x2,x3) = (1,0, -x2), f2(x1,x2,x3) = (0,1,Xj). Computing the Lie bracket, one finds = (0,0,2). Therefore, the Lie algebra £ generated by /1 and /2 satisfies £(x) D span|/i(x), f2(x), |/i,/2](^)} • for all x e IR3. According to Theorem 3.8.2, every initial point can be steered to the origin in finite time. However, a smooth stabilizing feedback to the origin for system (4.18) does not exits. Indeed, the map: (x,U) f(x,u) = (?li, U2, U2Z1 — ^1^2) • is not open at the origin. To verify this claim, consider the velocity v68 = (0,0,5), 8 > 0, and assume that v6 = f(x,u) for some (x,u). Then, equating the first two components, we get щ = u2 = 0. Hence 8 = u2X\ — щх2 = 0, reaching a contradiction. Problems 4.1. Consider the linear differential equation x = Ax on IR2, with A=(~1 ° 0 -2/’ Compute a Lyapunov function for the system on ]R2.
86 4 Asymptotic stabilization 4.2. Consider the linear differential equation x = Ax on IR3, with /-1 2 0 \ A= 0-10 . \ 0 0-3/ Is the system Lyapunov stable? If yes, find a Lyapunov function for this system, checking that it satisfies the three conditions (i) (iii). 4.3. Write the linear control system ±i=2x2 + wi, x2 = -rr3, ±3 = z3 + ui.. in the standard form x = Ax 4- Bu. Find a linear feedback и — Fx which stabilizes the system to the origin. 4.4. Consider the linear control system x = Ax + Bu, with /100\ /1\ A= 0 1 1 , B = I 0 I . \101/ \° / Prove that this system is completely controllable. Find a linear feedback matrix F such that (A+BF)3 = 0. Compute the solutions for the resulting system with feedback: x = Ax + BFx. 4.5. Consider the linear control system (4.4), with n > 2, m — 1. Assume that there exists a constant A G IR such that = A a\j for j = 1,..., n and bn = A . Is the system controllable? Hint: Use Hautus Lemma 4.2.4. 4.6. Consider the equations of a forced pendulum: x(t) + sinx(^) = u(t), |u(£)| < 1. Find a feedback control и = u(x,x) which locally stabilizes the system asymptotically at the origin. Can the system be also locally stabilized by a feedback at the (unstable) equilibrium point x = 7Г ? 4.7. Consider the control system: x = Ax 4- Bu, у = Cx, z = Dz 4- <p(x) 4- V;(71)? where x, y, z, и € IRn, А, В, C, D n x n matrices, with В and C invertible while det(D) = 0. Prove that there exists no feedback stabilizing this system to the origin. Hint: Apply Theorem 4.3.2.
5 Existence of Optimal Controls In this chapter we begin the analysis of optimal control problems, focusing- first on the existence of optimal solutions. In Section 5.1 we consider problems in Mayer form, where the cost to be minimized depends only on the terminal point of the trajectory. For gen- eral nonlinear systems, following [42] the existence of optimal controls will be proved under a key assumption. Namely, the sets of admissible velocities should all be convex. When this convexity assumption fails, one can still construct an optimal solution, but within a class of “relaxed trajectories”, corresponding to chat- tering controls. In the special case of linear control systems, one can push the analysis a bit further, and show the existence of a genuine trajectory of the original system which reaches exactly the same terminal point. As in [67], this argument yields again the existence of optimal controls. In Section 5.2 all results are extended to the more general Bolza problem, where a running cost is also present. 5.1 Mayer problems As usual, we consider the control system x = /(£,#, гл), и € U, (5-1) where L( = { u(-) measurable, u(t) e U for all t} . (5.2) Given an initial state x, a set of admissible terminal conditions S C IR x lRn, and a cost function ф : IR x lRn f—> IR, we consider the optimization problem min ф(Т, x(T,uY) (5.3) u€Z/,T>0 with initial and terminal constraints
88 5 Existence of Optimal Controls z(0) = x, (Г, ж(Т)) e S. (5.4) When, as in (5.3), the performance criterion depends only on the terminal time T and on the terminal point x(T) of the trajectory, we say that the problem is in Mayer form. Remark 5.1 The maximization problem max ^(T, x(T, u)) иеи,т>о is of course equivalent to (5.3), choosing ф = —ф. Under suitable hypotheses, we shall prove the existence of an admissible control u* whose corresponding trajectory #*(•) = j?(-,u*) satisfies the con- straints (5.4) and yields the minimum in (5.3). The key assumption will be the convexity of the sets F(t,x) = {f(t,x,w)', ai e U}, (5.5) which guarantees the closure of the set of trajectories of (5.1). To begin with the simplest case, assume that the terminal time T > 0 is fixed, and (5.3), (5.4) take the form inf ф(х(Т, u)), я(0) = x, x(T) € S', иЕЫ where ф is continuous and S is a closed subset of IRn. If the assumptions of Theorem 3.5.1 of Chapter 3 hold and the set of trajectories reaching the target set S is nonempty, then an optimal control exists. Indeed, by Theorem 3.5.1 of Chapter 3, the reachable set R(T) is compact. Hence there exists a point #min in the nonempty compact set R(T) A S where ф is minimal. Any control u* that steers the system to the point xmin, i.e. such that a;(T, u*) = a\nin3 is clearly optimal. A similar result holds for the more general problem (5.1)-(5.4). On the control system (5.3) we make the following assumptions, somewhat stronger than the hypotheses (H) at the outset of Section 3.1. (H) The set U C IRm of control values is compact, f : [0, oo) x IRn x U »—► IR™ is continuous in all variables, continuously differentiable w.r.t. x, and satisfies the bound |/(t, x, w)| < <7(1 -F |^|) for all t,x,u. (5.6) Theorem 5.1.1. (Existence of optimal controls). Let the assumptions hold. Assume that the sets of velocities F(t,x) in (5.5) are convex, the cost function ф is continuous, and the target set S is closed and contained in some strip [0, T] x IRn. If some trajectory rr(-) satisfying the constraints (5.4) exists, then the problem (5.1)-(5.4) has an optimal solution.
5.1 Mayer problems 89 Proof. 1. By assumption, there is at least one admissible trajectory reaching the target set S. Therefore, we can construct a sequence of controls uy : [О, Ty\ i—► U whose corresponding trajectories tp(-) starting at x satisfy (Tu, xp(Tp)) ES, lim <£(TM, ^(Tp)) = inf ф(7\ x(T,u)). (5.7) U—>OQ W(.) Since S C [0,T] x IR/1, we have Ty < T for every v. We can now prolong each function xy to the entire interval [0, T] by setting xy(P) = xy(Ty) for t € [Ty,T]. 2. By the assumption (H), as in Corollary 3.2.4 all trajectories satisfy the uniform bound fxy(t, u)\ < (eCt — 1) + ec<|^| for all te[0,T]. (5.8) Since f is uniformly bounded on bounded sets, the sequence xy(/) is uniformly Lipschitz continuous. Using Ascoli’s compactness theorem, by possibly taking a subsequence we can assume that Ty^>T* for some T* < T, and xy(•)—►£*(•) uniformly on [0, Г*]. 3. Because of the convexity of the sets F(t,x), Corollary 3.3.2 in Section 3.3 implies that #*(•) is an admissible trajectory of (5.1), i.e. there exists a control u* : [0,T*] i—> U such that ±*(i) = /(t, x*(t), for a.e. t e [0, T*]. Clearly, rr*(O) = x. Since S is closed, the first relation in (5.7) implies (T*, x*(T*)) = lim (T„ ^(TJ) 6 S. v—>oo 4. Finally, from (5.7) and the continuity of ф it follows ф(Т*, x‘(T*)) = lim ф(Т„, xv(Ty)) = inf ф(Т, x{T,«)). v—>oo u,T Therefore the control u*(-) is optimal. Examples of approximate solutions are shown in figure 5.1. We remark that the main difficulties faced in the above proof are: • The sequence of control functions up(-) is bounded, but may be highly oscillatory, hence it may not converge in L1. • Similarly, the sequence of derivatives xy = f(t.xy,uy) is bounded but may not converge in L1. On the positive side, taking a subsequence one has:
90 5 Existence of Optimal Controls • The sequence of times Ty converges to some T, • The sequence of trajectories яД-) converges uniformly to some function £*(•). whose derivative remains in the convex closure of the velocity sets, i.e. x*(t) € coF(Z, x*(t)) for a.e. time t. Hence, if all sets F(t, rr*(Z)) are convex, then there exists an optimal control such that i*(i) = f(t, #*(£), w*(0) f°r a-e- [0, Т]. Fig. 5.1. Two successive approximations xu to the optimal trajectory x*. Example 5.1 Consider a system of the form £(£) = /о(^) + Л(^) UiW , z(0) = x, e [-1,1], i = 1,..., m , i=l where /0, /i,..., fm are bounded, continuously differentiable vector fields on lRn. Then, given a time T > 0 and any continuous function ф : IRn IR, the problem min </>(x(T)) u() has at least one optimal solution. Remark 5.2 The above proof is a typical example of the Direct Method for proving the existence of optimal solutions. The basic steps are: (1) Construct a minimizing sequence x^-). (2) Show that some subsequence converges to a function #*(•). (3) Prove that .?*(•) is an admissible trajectory and satisfies the appropriate initial and terminal conditions. (4) Prove that ^*(-) attains the minimum value for the optimization problem.
5.1 Mayer problems 91 Several extensions of Theorem 5.1.1 are possible: (i) The assumption that S С [0, T] x lRn is used in order to ensure that the time intervals [0, Tu] are uniformly bounded. The theorem still holds if S is closed and ф(Т, x) —* oo as T—>oo. (ii) The assumption (5.6) on f is used to achieve the a-priori estimate (5.8). Using Theorem 3.2.3 of Chapter 3, one can replace (5.6) by any other hy- pothesis which guarantees that, as t e [0,T], all solutions of (5.1) starting at x remain inside a fixed compact set. (iii) The continuity assumption of the cost function ф can be replaced by lower semi-continuity: </>(£, re) < liminf ф(1',x') for all t,x. t' —>t, x' —*x The relevance of the various assumptions in Theorem 5.1.1 is illustrated by the following examples. Example 5.2. In connection with the system (±i, ±2) = (u, ^i), e { — 1, 1} , consider the optimization problem: min x2(T), z(0) = (0,0), x(T) e IR2 . Here the terminal time T is fixed but there are no constraints on the final state x(T). By analysis in Example 3.5, it is clear that for any given T > 0 this problem has no solution. Using highly oscillatory controls, one can construct a sequence of trajectories xI/(-) such that xu(T) -+ (0, 0). However, the second component хъ^Т} cannot be zero, for any admissible trajectory. Therefore this minimum cost cannot be attained. Notice that in this example the sets of velocities F(x) = {f(x,w); ш = ±1} are not convex. Example 5.3. Consider the optimal control problem in one space dimension max x(T, u) u(-) for the control system with dynamics x = их2, rr(O) = 1, where u(t) E U = [—1, 1] and T > 1 is any fixed terminal time. For any e g]0, 1] define the control
1 1\ 1 - t ’ £ ) 92 5 Existence of Optimal Controls The corresponding trajectory is found to be x£(t) = min If T > 1, there is no uniform bound on this set of solutions. In particular, x£(T) — e-1, thus there exists no optimal control. This shows the necessity of the assumption (5.6), or soma alternative assumption providing a uniform bound on the reachable sets. Example 5.4. Consider the optimal control problem given by: min ф(х(Т)), x = и, |u| < 1, z(0) = 1, iz(-) where T is fixed, ф(х) = (x — l)sign(z) and we set sign(O) = 0. For every T > 1 the infimum of the cost is —1, but is never exactly attained. Notice that here the cost function ф is not continuous. In fact, it is not even lower semicont inuous. The next result is concerned with the case where the velocity sets F(f, x) in (5.5) are not necessarily convex. In this case, one can construct the associated chattering system n x = /tt(t,x,utt) = , (uo, • • • ,un,0) - u9 e W = Un+1xAn i-0 (5-9) as in (3.72), (3.74) of Section 3.9. Applying Theorem 5.1.1, we can find an optimal solution t : [0, T*] h-> IR77 to the relaxed problem min ф(Т\ ж(Т, tt#)), (5.10) subject to the initial and terminal constraints (5.4). In general, ^opt(-) is not a trajectory of the original system (5.1). However, if the control system has the special form x(t) + A(t) • x(t) + h(t, u{t)), (5-11) then there exists some trajectory #*(•) of the original system (5.1) which reaches the same terminal point as ^pt. Clearly, this yields an optimal solution to the optimization problem (5.1)-(5.4). Theorem 5.1.2. (Existence of optimal controls for linear systems). Consider the optimization problem (5.3)-(5.4) for the linear system (5.11). Assume that the functions A. h, ф are continuous, the set U is compact and the target set S С [0, T] x lRn is closed. If there exists at least one trajectory which reaches the target set S, then the minimization problem admits an optimal solution.
5.2 The problem of Bolza 93 Proof. 1. For the linear system (5.11), the corresponding chattering system (5.9) can be written as x = f*(t,x,u“) = A(t) -x + ^O^t)- h(t,Ui(ty), u* e 1/“ = U x • • • U x An . i=0 (5-12) The optimization problem (5.3)-(5.4) for the chattering system (5.12) satisfies all the assumptions of Theorem 5.1.1. therefore it admits an optimal solution 4₽t: [0, T] ~ IR’*. 2. By Theorem 3.10.1 of Section 3.10, there exists a trajectory a:*(-) = ж(-,и*) of the original system (5.11) such that t*(T) = t(T). Since every trajectory of (5.11) is also a trajectory of (5.12), we have Ф(Т, x*(T)) = 0(T, z’pt(T)) = inf ф(Т, x(T,u*)) < inf ф(Т, x(T,u)), ueu (5.13) proving that the trajectory #*(•) is optimal. We remark that the values of the two infima in (5.13) coincide. Otherwise stated, the introduction of chattering controls, changing the system (5.11) into (5.12), does not improve the optimal performance of the system. 5.2 The problem of Bolza This section is concerned with the minimization problem of Bolza: mm L(t, x(t), u(t)) dt + ф(Т, x(T, w)) (5-14) for the control system (5.1) with initial and terminal constraints (5.4). Apparently, the performance criterion (5.14) looks more general than (5.3), because, in addition to the terminal cost ф(Т, ж(Т)), we here also have a running cost L = L(t,x, u). However, the problem (5.13) can be easily recast in Mayer form, introducing the additional state variable #o(t) = / L(s, x(s), u(sf) ds. Jo (5.15) The optimization problem (5.14), (5.1), (5.4) is equivalent to min {х0(Т,и) + ф(Т, x(T,u))}, (5.16) for the system on IR71"1"1
94 5 Existence of Optimal Controls ( ±o = L(t,x, u) [ x = f(t,x,u) u(t) G U. (5.17) with initial and terminal constraints жо(О) = 0, x(0) = x, (T, x(T)) G S. (5.18) The existence of optimal solutions for the problem (5.14), (5.1), (5.4) can thus be obtained by studying the equivalent problem (5.16)-(5.18). Theorem 5.2.1. (Existence of solutions for the Bolza problem). Let the assumptions (H) on f, U hold. Assume that S С [0, T] x IRn is closed, L is continuous and the set of trajectories reaching the target set S is nonempty. If ф is continuous and the sets F+(t,x) = {(yo,y) G IRn+1; уо > L(t,x,w)y = f(t,x,cS) for some cu G U} (5.19) are all convex, then the minimization problem (5.14) for the system (5.1) with initial and terminal constraints (5.4) has an optimal solution. Proof. 1. Because of (H), all admissible trajectories remain uniformly bounded as t G [0, Т]. Since L is continuous and the set U is compact, there exists a constant M such that |L(t, x(t), cj)| < M for all t G [0, Т]. w G U, and every admissible trajectory #(•). 2. Consider the auxiliary minimization problem on lRn+1, in Mayer form: inf |х0(Г, u0 ,u) + ф(Т, z(T,w))} (5.20) for the control system f xo = fo(t,x,uo,u) = + (1 - u0(t))T(t, x(t), u(t)), z . [ x = f(t,x,u) ' ’ ' Here the control functions satisfy (u0(t), iz(t)) € [0,1] x U C Rm+1. (5.22) In addition, we impose the initial and terminal conditions (x0,x)(0) = (0,x) e IRn+1, (Т,ж(Т))ё5. (5.23) 3. We can now apply Theorem 5.1.1 to the problem (5.20)-(5.23), observing that the sets
5.2 The problem of Bolza 95 F(t,xo,x)-{(fo,/)(t,x , Uq, u); uq E [0,1], и E U = {(yo,y) e F+(t,x); y0 < м} C IRn+1 are all compact and convex. This yields an optimal control (uj, и*) : [0, T*] i—► [0,1] x U for the auxiliary problem (5.20)-(5.23). Let #*(•) be the correspond- ing trajectory. 4. Since L(t,:r*(t), zz*(t)) < Af, we must have uq(£) = 0 for almost all t, other- wise the control (0, u*) would achieve a strictly better performance. Therefore, the control ?/*(•) is optimal also for the original problem (5.14), (5.1), (5.4). In the case of a system with linear dynamics, the convexity assumption on the sets of velocities can be relaxed, also in the case of Bolza problems. Indeed, Theorem 5.1.2 can be extended also to case of a running cost. Theorem 5.2.2. (Solutions for the Bolza problem with linear dynam- ics). Consider the control system with linear dynamics (5.11), with initial and terminal constraints (5.4)- Assume that the functions A,h are continuous, the set U is compact and the target set S С [0, T] x lRn is closed. If there exists at least one trajectory which reaches the target set S, and if the functions ф,ао,Ьо are continuous, then the minimization problem min < / [ao(^) • #(£) + ho(t, wW)] u)) 1 (5.24) U’T [Jo J admits an optimal solution. Proof. Introduce a new scalar variable Яд, satisfying ±o(t) = a0(t) • x(t) + hv(t,u(ty), ^o(O) = 0. The Bolza problem (5.24) can then be reformulated as a Mayer problem on IRn+1, namely mm {х0(Г) + ф(Т,х(ТУ)}. This problem satisfies all the assumptions in Theorem 5.1.2, hence it admits an optimal solution. Problems 5.1. Consider the optimal control problem (5.1), (5.3) with the more general constraints x(0) ек, (T, x(T)) e S, x(t) e P for all t e [о, T].
96 5 Existence of Optimal Controls Here the initial point is not fixed, but can vary within a compact set К C IRn. In addition, we require that the entire trajectory remains inside a closed set J?. Under the same assumptions as in Theorem 5.1.1, show that if there exists at least one trajectory satisfying all constraints, then the minimization problem admits an optimal solution. 5.2. For the control system (5.1), starting from the origin, consider the problem of reaching a point in a closed set К G ]Rn in minimum time. Write this problem in the standard form (5.3), (5.4), for suitable S and ф. Let the assumptions in (H) hold and assume that the sets of velocities F(f, x) in (5.5) are convex. If some trajectory reaching a point of К from the origin exists, prove that the minimum time problem admits an optimal solution. 5.3. Let t •—> y(t) G lRn be any continuous function. Consider the optimal tracking problem fT min / |?/(i) — x(t,u)\2 dt, **(•) Jo for the control system m i = /o(®) + 52/i(®)«i(t) Uj(t) e [-1,1], г=1 with initial data jr(O) = x G JR71. We assume that Jo, • • •,fm are C1 vector fields with sub-linear growth, so that \fi(x)| < C(1 + |.r|). Prove that this problem admits an optimal solution. 5.4. Consider the optimization problem min / L(t, x(f), x(t}) dt *(•) Jo where the minimum is sought among all functions x : [0, T] »—> IRn such that t(0) = 0, x(T) = £, and which are Lipschitz continuous with constant «, i.e. |ar(t) — .t(^)| < n\t — £'| for all t, t' G [0, T]. Formulate this problem as an optimal control problem with fixed terminal time. Assuming that the Lagrangean function L is continuous w.r.t. all variables and convex w.r.t. x, prove that the problem admits an optimal solution. 5.5. Consider the car parking problem, modelled by the system (1.16). Given an initial position (.T](0), x2(0). #(())) = (ah,#2,0), prove that there exists a control function t ►—> (u(£), o(£)) steering the car to the ori- gin in minimum time. Notice that in this case, the angle is defined up
5.2 The problem of Bolza 97 to multiples of 2тг. As target set we should thus take {(xi,#2>0) = (0,0,2Ъг), к integer}. 5.6. Recall the equations (1.9) describing the motion of a boat on a river. Given any two points A = (a, — 1) and В = (b, 1) on opposite shores, show that there exists a control u(-) steering the boat from A to В in minimum time. Notice that for this control system we have the state constraint o:2(t) 6 [—1,1] at every time t. 5.7. Consider the linear control system: x = Ax + Bu, x € IRn, и € IR™, and the optimal control problem: max ^(Т,х(ТУ). и Prove that if lim^^^ x) = +00 and the system is controllable then there exists no optimal control. What happens if the system is not con- trollable? 5.8. A satellite moves around the earth on a fixed plane with dynamics, in polar coordinates, given by: r(t) = u(t), r(0) = 1, 0(t) = 1, 0(0) = 0, with 0 < и < 1 the energy given to rockets. The satellite gets energy e by solar panels, thus: e(t) = /?(0(t)) — u(t), e(0) = 0, e(t) > 0, where (3(0) = 1 if 0 € [0,7r] and vanishes otherwise, reflecting the fact that for half orbit the earth sends shade over the satellite. Consider the optimal control problem: maxr(T) + e(T), r(T) > Ci, e(T) > C2. и 1) Verify that the set of admissible controls (which ensure all constraints e > 0, r(T) > Ci and e(T) > C2) for the case T = 2тг, Ci = C2 = 0, is given by л /'2tv U := < и : [0, 2тг] —► [0,1] : / u(s) ds < тг I Jo Hint: Notice that on [0, я] the increase of e is given by %, while it vanishes on [7г, 2тг]. 2) Show that if T is allowed to be arbitrarily large, then there exists
98 5 Existence of Optimal Controls no optimal control. Hint: is there any a priori bound on r 7 3) Assume T = 2А?тг. Show that there exists a solution if and only if A;7t > Ci and fc?r + 1 > (7i + C2. 5.9. Consider the control system: ±1 = их] + 1, X2 = tz, #i(0) = #2(0) = 0. with и € U and the optimal control problem: inax</)(T, 2?(T)), G S, и where ф is continuous and S is a compact set. 1) Prove that there exists no optimal control if S = {(—1,0)} and U = [-1,1]. Hint: can 2:1 became negative? 2) Prove that there exists an optimal control if ф = X2, S — {(2,2:2) • 0 < x2 < 100} and U = {0, +1}. 3) Prove that there exists an optimal control for ф = Xi — T, S = {(2:1,0) : 2 < 24 < 100} and U = {—1,4-1}. 5.10. Consider the optimal control problem: 2:1 = 2:2, %2 = 4" .(/(J>1) 4-u, #i(0) = 2:2(0) = 0, max^(ar(T')), ^2(T} = — 24 (T)2, и where |u| < 1, g and ф are continuous. Prove that if lim <7(2:1 )/|2:iI = a > 1, |ii |—>4-00 then there exists an optimal control (regardless of the growth of ф at infinity).
6 Necessary conditions In this chapter we introduce the Pontryagin Maximum Principle (PMP) and show how it can be used in order to compute optimal controls and optimal trajectories. Section 6.1 deals with optimization problems where the payoff depends only on the terminal point, which is not subject to any constraints. This is a case where the key ideas can be more clearly described. To test the opti- mality of a control w*(J, one has to consider various possible perturbations, and check that none of these produces a better payoff. In particular, to derive Pontryagin’s necessary conditions, one constructs a family of “needle varia- tions” , changing the values of the control u* only in a neighborhood of a given time t. Remarkably, in order to compute how all these different perturbations affect the terminal payoff, it suffices to transport one single “adjoint vector” backward along the trajectory. Imposing that the change in the payoff (up to first order) is non-positive for all admissible variations, one obtains the famous Pontryaging Maximum Principle [72]. In general, the PMP yields a (highly non-linear) system of O.D.E’s for the optimal trajectory and for a corresponding adjoint vector, which must then be solved with appropriate boundary conditions. In Section 6.2 we present various examples where these equations can be explicitly solved, thus determining the optimal controls and the optimal trajectories. In Section 6.3 we discuss the PMP in the case where terminal constraints are present. The proof is now considerably more difficult, ultimately relying on the use of Brouwer’s fixed point theorem. Sections 6.4 and 6.5 extend the earlier results to problems with variable terminal time, and to minimization problems containing both a terminal cost and a running cost. In particular, we show how the classical Euler-Lagrange and the Weierstrass necessary conditions in the Calculus of Variations can be derived from the PMP. The special case of a linear system with quadratic cost, most important for engineering applications, is discussed in Section 6.6.
100 6 Necessary conditions For an alternative approach to necessary conditions for optimal control problems, see [27], [28]. 6.1 The Mayer problem with free terminal point We consider here an optimal control problem in Mayer form max^(2?(T, u)) (6.1) u£U subject to ±(t) = f(t, rr(t),u(t)), ж(0)=ж. (6.2) For a given set U C IRm, the family of admissible control functions is defined as U — {и : [0, T] h-> U, и measurable} . (6.3) Notice that here the terminal time is fixed, but we do not put any constraints on the terminal state x(T). Our main goal is to derive necessary conditions in order that a trajectory £*(•) = x(-,u*) be optimal. We make the following- assumptions. (0) The set 1? C IR x IRn is open, the function f = f(t,x,u) is continuous on I? x U and continuously differentiable w.r.t. x. The payoff function ф : IR” i—► BFt is differentiable. Notice that here we are not making any assumption on the set U of admis- sible control values. In particular, we may well have U = IR™. Our main goal is to derive necessary conditions for the optimality of a control £*(•). These conditions will provide a basic tool for the actual computation of optimal controls. Theorem 6.1.1. (Pontryagin Maximum Principle, free terminal po- int). Consider the optimal control problem (6.1)-(6.3), under the assumptions (C>). Let u*(-) be a bounded admissible control whose corresponding trajectory #*(•) = x(-,?i*) is optimal. Call p : [0,T] »—► IRn the solution of the adjoint linear equation p(i) = -p(t) • Dxf(t,x*(t),u\t)), p(T) = \7г1>(х*(ТУ). (6.4) Then the maximality condition PW • f(t,x*(t),u*(ty) = max |p(t) • f(t,x*(t), w)} (6.5) holds for almost every time t € [0, Т].
6.1 The Mayer problem with free terminal point 101 In the above theorem, rr, /, v represent column vectors, Dxf is the n x n Jacobian matrix of first order partial derivatives of f w.r.t. x, while p is a row vector. In coordinates, the equations (6.4), (6.5) can be rewritten as = Pi(T) =-^(x’(T)), (6.6) U.l j J=1 n ( n 1 52Pi(^) • = max < Vpi(t) • fi(t,. (6.7) г=1 L i=l ) Proof. To understand the derivation of the Pontryagin Maximum Principle (PMP), it is best to consider first the case where the optimal control 1? : [0, T] i—> U is continuous. In the last step of the proof we shall extend the result to a general control u*, measurable but possibly discontinuous. 1. Fix any time т > 0 and any admissible control value w € U. For e > 0 small, consider the perturbed control function |u*(i) otherwise. This is called a “needle variation” of u, due to the shape of its graph, as shown in figure 6.1. Fig. 6.1. A ’’needle variation” of the optimal control u*. 2. Let t h-> x£(t) = x(tyue) be the corresponding trajectory of the control system (6.2). Of course, xe(t) = x*(i) for t < r — e. At time t = т we have = lim£_0+ JTT_e /(t, ar£(i), w)dt - | f^_£ f(t, x*(t). dt} = /(r, x*(t), w) - /(r, x*(r), . (6-9)
102 6 Necessary conditions As shown in figure 6.2, at time r the curve e i-> же(т) thus admits the tangent vector t?(r) = lim Xe^— rr*(r), cj) — /(t, x*(t), tz*(t)) . (6.10) e^o+ e On the remaining time interval [т, T], all trajectories x£ are solution to the same O.D.E., namely x(t) = /(f,z(f),u*(f)). According to Theorem 2.3.1, for every t € [т, T] the tangent vector / x . i. xeU) “ x*(t) v(t) = lim £->0+ £ is well defined and provides a solution to the linearized equation v(t) = Dx/(t,a:*(t),?i*(t)) • v(t), (6.И) with initial data given by (6.9). In particular, replacing u* with the controls u£ defined at (6.8), as e in- creases the terminal point x(T, u£) of the trajectory is shifted in the direction of the vector v(T). Fig. 6.2. The perturbed trajectories xe, corresponding to the needle variations ue . 3. By assumption, the control «*(•) is optimal, therefore ^(х*(Т))>Ж(Г)) for every e > 0. Differentiating w.r.t. s we obtain 0 > lim = ^fe(r)) e—»0-|- e (is E—04" = W(x*(T))-v(T). (6.12)
6.1 The Mayer problem with free terminal point 103 4. Summarizing the previous arguments, we have shown the following: For each т e]0,T] and every e U, let t be the solution to the linear Cauchy problem (6.11) with initial data v(t) = /(Л £*(TX ^*(r), W*(T)) • Then the terminal value v^T^(T) has non-positive inner product with the gradient of the payoff function at the terminal point ж*(Т), namely W(z*(T)) • v{r^(T) < 0. (6.13) Here we regard as a row vector and as a column vector. 5. The family of inequalities (6.13), one for each (t,oj), can be rewritten in a more convenient form. Instead of transporting each of the (infinitely many) tangent vectors forward in time along the trajectory ж*(-), it is more convenient to transport the single row-vector p(T) = V'0(^*(T)) backward in time. Let t p(i) be the solution of the backward Cauchy problem Ж = -p(f) • Dxf(t, ?(t)y(t)), p(T) = W(z*(T)). (6.14) Notice that (6.14) is the adjoint linear equation for (6.11). According to The- orem 2.2.2, the inner product p(t) • te[r,T] is constant in time. Therefore, by (6.13) we have p(r) • г/т’ш)(г) = p(T) • v(T>u,)(T) < 0. Recalling (6.10), for every т e]0, T] and cu e U we obtain p(r) • (/(^ я*(т), w) - /(r, x*(r), u*(t))) < 0. (6.15) By continuity, it is clear that the above inequality also holds for r = 0. Since w 6 U is arbitrary, (6.15) yields the maximality condition (6.5), for every t e [о, T] 6. To complete the proof of the PMP, we need to extend the previous argu- ments to the case where the optimal control tz*(-) is measurable, bounded, but possibly discontinuous. By a theorem of Lebesgue, at almost every time т e [0, T] the bounded measurable function t /(£, x*(t), u*(t)) is quasi-continuous, i.e. lim; [ f(T,x*(T),u*(T))\dt = 0. (6.16) £->0+ e jT_e | I Fix any such time т and any control value u> e U. For e > 0 small, define the control function u£(-) as in (6.8) and let же(-) = be the corresponding trajectory. At time t = r one has
104 6 Necessary conditions дг£(т) = т*(т) 4- I - /(£,я*(£),гд*(£))] dt. Letting e —> 0+, thanks to (6.16) we again recover the existence of a tangent vector v(r), as in (6.10). The rest of the proof goes through without changes, also in this more general case. Notice that now we obtain the validity of the maximality condition (6.5) not at all times t e [0,T] but only almost everywhere, i.e. at every Lebesgue point t of the map t»—► Fig. 6.3. Geometric meaning of the optimality condition. The geometric meaning of the maximality condition (6.5) is illustrated in figure 6.3. Using the adjoint linear equation (6.14) we transport the row-vector p(T) = V^(x*(T)) backward along the optimal trajectory. The maximality condition means that, at each time r, among the possible attainable speeds ±(t) e F(r, x*(t)) = |/(т, ж*(т), cd), Cd€U| we should choose the one that maximizes the inner product р(т) • ±(т). Remark 6.1 In addition to the assumptions of Theorem 6.1.1. let U be convex and assume that f is also differentiable w.r.t. u. Then the maximality condition (6.5) implies p(t) • Duf(t, x\t). u*(tY) (id - u*(t)) <0 Vid e U. (6.17) Otherwise, one would have u*(£) + e(cd — ti*(t)) 6 U and p(t) • /(t, x*(t), u*(i) + e(w - w*(t))) > P(t) /(t &*((), «*(*)) for e > 0 sufficiently small. 6.2 Computation of optimal controls The Pontryagin Maximum Principle motivates a practical method for find- ing optimal solutions to the problem (6.1) (6.3). We first define the function
6.2 Computation of optimal controls 105 и = и(1,х,р), from [0,T] x IRn x IRn into the control set U, in terms of the maximality condition p • f(t,x, u(t, x,p)) = max p • (6.18) In other words, u(t, x,p) = argmax p • f(t, x, cu). (6.19) We then solve the system of 2n differential equations in the variables x,p: (i = f(t,x,u(t,x,pY) r ^p- -p. Dxf(t,x,u(t,x,p)), with boundary conditions z(0) = x, p(T) = W(z(T)). (6.21) By Theorem 6.1.1, if a bounded optimal control и exists, it must be found among the solutions of (6.20)-(6.21). In general, this method encounters two main difficulties: (i) The map (i,x,p) h-> u(t,x,p), implicitly defined by the maximality condi- tion (6.18), may be multivalued and/or discontinuous. (ii)The equations (6.20)-(6.21) do not constitute a Cauchy problem on IR2n, but a (usually harder) two point boundary value problem. Indeed, at time t = 0 we are assigning the initial value д?(0), but the initial value p(0) is not explicitly known. Instead, an implicit equation is given, involving the terminal values p(T), x(T). In particular cases, however, the equations for p and x can be uncoupled, and a solution is more easily found. Example 6.1 (Linear pendulum with external force). Let q be the position of a linearized pendulum (see figure 6.4), controlled by an external force u, with magnitude constraint |u(t)| < 1, Vi. For simplicity, let us assume that the initial position and velocity are both zero, and that the motion is determined by the equations q(«) + q(t) = u(t), 9(0) = 9(0) =0. We wish to maximize the displacement q(T) at a fixed terminal time T. Introducing the variables x\ = q, x% = q, the optimization problem can be formulated as max^i(T, u), uea where the dynamics is described by
106 6 Necessary conditions {X\ = x2 X2 = -Xi + U £1(0) = 0, £2(0) = 0, and the set of admissible controls is U = {и : [0,T] »—> [— 1,1], и measurable } . Fig. 6.4. Trajectories in the phase plane for the linear pendulum. Notice that here we have /(^w) = a?2 —Xi + и Dxf = 0 1 -1 0 In this case, the adjoint equations (6.6) take the form Pi — P2 P2 = -Pi Р1(П = 1- p2(T) = o. These equations can be solved for p independently of x, yielding Pi(t) = cos(T — t), P2(T) = sin(T - t). By (6.18), the optimal control u* satisfies Pi£2 + p2(—£i + ti*) = max <! prx2 + p2(~£i + ^) |u/|<l I Therefore, the optimal control is u* = sign(p2(T)) = sign(sin(T - <)). (6.22)
6.2 Computation of optimal controls 107 Notice that trajectories corresponding to the constant controls и = 1 or и = — 1 are circles centered at (1,0) or (—1,0), respectively. The case where T = Зтг/2 is illustrated in figure 6.4. According to (6.22), the optimal control is «*(*)={; if 0 < t < тг/2 , if тг/2 < t < Зтг/2. Example 6.2 . To appreciate the effect of the map и = u(t,x,p) being mul- tivalued, consider the problem max а?з(Т, и), и for the system (±i, i2, i3) = (a, -xi, x2 — я?), (^1,^2,^з)(0) = (0,0,0) with the control constraint |u| < 1 for all t e [0, Т]. The equations (6.6), (6.7) here take the form (Р1,Р2,Рз) = (P2 + 2zip3,p3,0), (p1,p2,p3)(T) = (0,0,l), ' u(t) = 1 if pi > 0, < u(t) = —1 if pi < 0, k u(t) e [—1,1] if pi = o. Solving for p3,p2, we find рз(*) = 1, p2(t)=T-t Vt6[0,T]. In turn, the value of pi can be found from the equations Pi = (-1 + 2u)p3 = -1 + 2sign(p!), P1(T) = 0, pi(0) = p2(0) = T, (6.23) with the convention: sign(0) = [—1,1]. From (6.23) it follows, see figure 6.5. ' _3 (T < 2 V 3 0 u*(t) = - t)2 if t G if t G M) .w, -1 if 0 < t < T/3 1/2 if T/3<t<T. P1W = Observe that, on the interval [T/3, T], the value a* — 1/2 is derived not from the maximality condition (6.18), but from the equation р^ = (—1 + 2u) = 0. An optimal control with this property is called singular. Example 6.3 . We now study a case where is not linear, hence the terminal value p(T) = V^(x(T)) is not a priori known. Consider the problem
108 6 Necessary conditions Fig. 6.5. Graph of pi. max [xi(T, u) — x^T, u)] , for the system (±ь±2) = (#2,ti), (#i,#2)(0) = (0,0), |u| < 1 Vie[0,T]. Pontryagin’s equations (6.6), (6.7) take here the form (pi,p2) = (0,-P1), (P1,P2)(T) = (l,-2x2(T>), u(t) = sign(P2(t)). We thus find Pi = l, P2(t)=P2(T) + (T-i). Observe that p2 is strictly decreasing function. Hence the corresponding con- trol и = sign(p2) is either constantly = 1, or constantly = —1, or else it has the form ( 1 if 0 < t < r (a o .x u^~ \ — 1 if r<t<T, (6’24) for some r G [0, T] at which p2 changes sign. The optimality of the constant controls и = ±1 is easily ruled out. Indeed, for u(t) = 1 we have a;2(Z) = P?(T) = — 2T < 0. For u(t) = —1, we have xz(t) = —t, p2(T) = 2T > 0. In both cases, the maximality condition is not satisfied. Now consider a control и satisfying (6.24), for some r. The corresponding trajectory now is Г2(/) = {2т-Ип'>т ’ Ж1(<) = / xAs}dS' Since at the switching time r we have р2(т) = 0, we deduce
6.2 Computation of optimal controls 109 0 = р2(т) = p2(T) + T - т = -2ar2(T) + T - т = -2(2т - T) + T - r, hence т = 3T/5. The optimal control is zm _ f 1 if 0 < t < 3T/5, [-lif 3T/5<f<T. Example 6.4 . To see that the conditions of Theorem 6.1.1 are not sufficient for optimality, consider the problem max a:2(T, zz), и for the system described by = (“j, = СП > l«(t)l<i vtG[0,T]. y.r2 J \x\ ) \^2(0) ) (6.25) The constant control zz*(t) = 0 yields the trajectory a;*(£) = 0. The corre- sponding adjoint vector satisfying (Pi, P2) = (-2^, 0) = (0,0), (pi,p2)(T) = (0,1) is (pi, p2)(£) = (0Д), hence the maximality condition (6.7) holds for all t. However, zz* is not optimal. In fact, any control и 7^ zz* yields z2(T, u) = dt > rr2(T, zz*) = 0. Remark 6.2. The above example shows that the Pontryagin Maximum prin- ciple provides only a “first order” necessary condition for the optimality of a control zz*(-)- Roughly speaking, this means the following. Assume that there exists a family of perturbed controls u£(•) such that, calling xe(-) the corre- sponding trajectories, one has ^е(Г))-Ж(П) Um ---n-----n----- £^o+ ||zZe - 1Z*||L1 Then the control zz*(-), which is clearly not optimal, will not satisfy the PMP. However, the non-optimality of zz* may not be detected by the PMP if the increase in the payoff function ф is of higher order w.r.t. the control perturbation, say Hm ^(Г))-У>(х*(Т)) = Q £-*0+ 11^-71*11^ j = l,2,...,fc- 1,
110 6 Necessary conditions ^(Г))-У>(Ж*(Т)) e->0+ ||?Ze— > 0 for some к > 2. For a discussion of higher order necessary conditions for optimality, we refer to [48], [58], [59]. Example 6.4 (continued). For the control system (6.25), consider the per- turbed control functions _ ( e if t E [0, T/2], [ > |-£if te]T/2, Т]. Notice that — u*||Li = H^Hl1 — eT. The terminal points of the corre- sponding trajectories are (xf(T), ^(T)) = (0, s3T3/12). Recalling that the payoff function is ^(^1^2) = ^2, in this case we have d3 = T3/2 > 0. £=0+ This again shows that the constant control u*(f) = 0 is not optimal, but the increase in the value function is very small, namely of third order w.r.t. to the control perturbation. This cannot be detected by the PMP. 6.3 The Mayer problem with terminal constraints This section is concerned with the optimization problem max (#(T, ^)) (6.26) u^U for the control system described by x = f(t, x, u), #(0) = x, u(t)ev, £ e [0, T]. (6.27) Here the terminal time T is fixed and the terminal point x(T) satisfies the constraint x(T) € S = {x e IRn ; 0Дж) = 0, г = 1,..., к}. (6.28) In addition to the assumptions (<0>), we now assume that all functions 0o, Фъ • • • j Фк ' Dln ► Dt are continuously differentiable.
6.3 The Mayer problem with terminal constraints 111 Theorem 6.3.1. (Pontryagin Maximum Principle with terminal con- straints). Let u* be a bounded admissible control, whose corresponding tra- jectory x* is optimal for the maximization problem (6.26)-(6.28). Assume that the gradients V0o,... ,^фк are linearly independent at the terminal point x*(T). Then there exists a nontrivial, absolutely continuous vector function p(-) which satisfies the equations p(t) = -p{tyDxf(t,x^(tfiu^t)Y (6.29) p(f) • f(t, x*(f), u*(t)) = max |p(t) • fft, z*(£), a;)} (6.30) at almost every time t G [0, T], together with the terminal conditions к р(Т) = £л^(х*(Т)) (6.31) i=0 for some constants Ao,..., Xk, wiift Ao > 0. Remark 6.3. If u* is optimal in connection with the minimization problem min 0o (^(T, u)) for the system (6.27) with constraints (6.28), then all conclusions of PMP continue to hold, with (6.30) replaced by p(t) • f(t, x*(t), = min (p(0 • f(t, x*(t), a/)} . (6.32) The equations (6.29)-(6.31) have a nice geometrical interpretation, which motivates the subsequent proof (see figure 6.6). For every co € U and every time r where ?/*(•) and is quasi-continuous, consider the one-parameter family of control functions u£ defined at (6.8). By changing the control function from u* to u£, the terminal point of the trajectory will be shifted from x*(T) to x(T, u£). For e > 0 small, the direction of the shift is approximately given by the vector: V(T’W) = j-x(T,ue) de = v(T), £=0 where v(-) denotes the solutions to the linear Cauchy problem v(t) = Dxf(t, z*(t), u*(£)) • v(t), v(t) = /(т, ж*(т),о?) - /(t,x*(t),u*(t)) . (6.33) Otherwise stated, if Af(-, •) denotes the fundamental matrix solution for the linear system: v(t) = Dxf(t,x\t),u*(tf) • 'u(t), then
112 6 Necessary conditions ^-x(T,u£) =v(r’“) = • [/(r,x*(r),w) - /(t,x*(t),u*(t))] . d£ e=0 (6.34) Define the convex cone Г as the positive span of all such vectors v^r,u;b Г = span4- 0 < т < T, lj g U, zz*(-) is quasi-continuous at r } . (6.35) Next, let Ts+ — Ts+(x*(Ty) be the tangent cone to the set S+ = 0o(^) > 0о(я*СП)> — ° Vz = 1,..., at the point rr*(T). More precisely, Ts+ = {v; V0oCr*CT)) -v > 0, V0i(x*(T)) -v = 0 Vz = l,...,fc}. Observe that Г represents a cone of ’’feasible directions”, i.e. a set of direc- tions along which we could shift the terminal point ж(Т) by suitably changing the control zz*. On the other hand, Ts+ represents the cone of ’’profitable di- rections”, i.e. the set of directions along which we should shift the terminal point, in order to raise the value of 0o(x(7'))^ keeping x(T} inside the target set S, see Figure 6.6. Fig. 6.6. The cone Г of feasible directions and the cone Ts+ of profitable directions. The PMP amounts precisely to the assertion that the two cones Г and 7$+ are weakly separated. More precisely, the conditions (6.29)-(6.31) can be restated in the form p(T) • z; < 0 WgF, (6.36) p(T) • v > 0 Vz; G Ts+ . (6.37)
6.3 The Mayer problem with terminal constraints 113 Indeed (6.36) holds if and only if p(T) • < 0 for every admissible control value ш G U and every time т at which u* is quasi-continuous, Recalling that the product p(t) • v(t) remains constant in time, this is the case if and only if p(r) • [/(г, £*(t), u>) — /(т, и*(т))^ <0 Vt G [0,T], cj G U, hence if and only if (6.30) holds. The equivalence of two conditions (6.31) and (6.37) is proved in Lemma A.9.2 of Section A.9. The proof of Theorem 6.3.1 relies on the possibility of “combining” con- trol variations. More precisely, let To,... ,тдг € (0,T] be distinct times where u*(-) is quasi-continuous. For any given uo, • • • € U and вц,..., Ок > 0, consider the control function „ if t e [Tj -eOj, Tj] for some j € {0,..., 2V}, , ( u*(t) otherwise. Replacing the control u* by we will show that, as e increases, the terminal point of the trajectory is shifted in the direction N = 52^v(T'’Wj)- (6-39) £=0+ j=0 Therefore, by suitably choosing the values of Tj, cjj, 6j, one can shift the ter- minal point x(T) along any direction v of the cone Г. Assuming that this cone of feasible directions is large, so that it cannot be separated from 7^+, the theorem will be proved by contradiction, showing that u* is not optimal. Proof. The Pontryagin Maximum Principle with terminal constraints will be proved in several steps. 1. By Lemma A.9.3 in the Appendix A.8, the assumption that Г and T§+ are not weakly separated implies that there exists finitely many control values £ U and times tq,...,t/v where ?/*(•) is quasi-continuous, such that, setting v(Tj,Wi) = - /(^^(тДиЧт;))] (6.40) as in (6.34), the cone Г* = span+ |v(To’"1 * * o) * * * * * *,...; v(TN’“w)| (6.41) is not weakly separated from Ts+. Observe that, for any 3 > 0 and every Lebesgue point т for u*(-), there exists another Lebesgue points т' т such that d X —ar(T,u£ie)
114 6 Necessary conditions Therefore, again by Lemma A.9.3 of Section A.8, it is not restrictive to assume that the times Tj in (6.40) are all distinct, say, 0 < T0 < <rN <T. 2. Consider the parametrized family of control variations u£j defined at (6.38), where £ € [0, f] and 0 ranges over the standard TV-dimensional simplex {N 1 O = (0o,...,ON)-, 0,>О ^2^ = 11. (6.42) г=0 J One eventual goal is to apply Lemma A.9.4 of Section A.8 to the map X(s,0) = z(7>^). (6.43) 3. With u£j as in (6.38), we claim that the limit x(T,uej) - x(T,u*) lim-------—------------ E—>0 £ j=0 (6.44) holds uniformly w.r.t. 0 € A^. Indeed, for Z? = 0,..., TV define the controls if t E [Tjfor some j e {0,. e,e( (u*(£) otherwise. ' ’ ' By induction on Д we will show that the limit ,. x(T,u*e)—x(T,u*) lim --------------------= e-»0+ £ £ ^0jAf(T,Tj) [/(трХ*(тД^) - /(г^х’(т,),и*(т7))] (6.46) j=0 is valid for all t € [т^,Т], uniformly w.r.t. 0 6 An. Indeed, assume that (6.46) holds for some index L Then «eV) -a:(^+1’ue,e) fTe+i J Tf+i ~E0f+l [/(t,x(t,u^V),^+i) - /(t,x(t,i4 e),u*(t))] dt + [/(t,x’(t),^+l) - /(t,x‘(t),u*(t))J + [/(t,x*(t),u‘(i)) - f(t,x(t,uee e),u*(t))] | dt.
6.4 Variable terminal time 115 The assumption that is bounded allows us to use Theorem 2.3.2 of Section 2.3. Together with the inductive hypothesis, this yields .. -x*(t€+i) lim ------------------------- £->0+ £ Га;(т€+1,^ 0)-a?‘(r€+i) - a:(r€+i, u* 0) = Inn ---------------------------1----------’------------------- £->o+ e £ i = [/(7>>a:*('5),%) - f(rj,x*(rj'), u*(r,))] j=0 + lim I [f(t,x*(t),CLie+i) - /(t,x*(t),u*(t))]dt e^0+ Jre+1-e0e+l + lim I - f(t,x*(t),we+i)] dt + lim / [/(£, #*(£), — f(t, x(t^ Д w*(t))] dt £^0+ A£+1-£^+1 £ j=0 +&e+i [f(.t,x*(t),ue+i) - €+1 = - /(т,-,х*(т^),и*(т,))] . (6.47) j=o Observing that for every t e [r^_|_i,T], it holds .. -x*(t) aj^+i,^1)-x*(r£+i) lim -----------------= Af(t, * lim ------------------------------- e->0+ £ £->0+ £ from (6.47) it follows (6.46) with £ replaced by £ T 1. By induction, when £ = N, t = T, we recover (6.44), proving our claim. 4. We can now apply Lemma A.9.4 in the Appendix A.8 to the map X intro- duced at (6.43), with £ > 0 suitable small. Since Г* and T$+ are not weakly separated, there exists (e, 0) such that 0o(X(f, 0)) > </>o(^*(T)), &(X(e, 0)) = Фг^{Т}\ Vz = 1,..., fc. (6.48) Recalling (6.41), this shows that the control u~g steers the system to some point in S, and achieves a value фо(х(Т, иё 0)) strictly better than u*. Hence, tz* is not optimal. This contradiction proves the theorem. 6.4 Variable terminal time This section is concerned with the optimization problem
116 6 Necessary conditions max фо (T, x(T, u)) (6.49) u&A for the control system described by x = f(t, x(t), u(tf), #(0) = £, u(t) E U a.e., (6.50) where the terminal time and the terminal point are subject to the constraints фг(Т,х(Т,и)) =0, г= (6.51) An optimal solution of (6.49)-(6.51) is now a pair (T*,rz*), where u* : [0,T*] I—» U is measurable and the corresponding trajectory x*(-) yields the maximum in (6.49) among all those which satisfy (6.51). Theorem 6.4.1. (PMP, variable terminal time). Consider the opti- mal control problem (6.49)-(6.51), under the usual assumptions (ф). Let u* : [0,T*] i—> U a bounded optimal control for the problem, (6.49) (6.51), and let #*(•) be the corresponding optimal trajectory. Assume that f is continuously differentiable w.r.t. both t and x, and that the vectors = • • • > fa2-)’ i = 1,..., k, are linearly independent at the point (T*, x*(T*)). Then there exists a nontrivial absolutely continuous row-vector p(-) such that p(t) = -p(t) • Dxf(t,x*(t\u*(tf), (6.52) p(t) • f(t, z*(t), zz*(t)) = max |p(i) • f(t, z*(£), c^)| (6.53) at almost every time t E [0,T*]. Moreover, there exist constants Xo,...,Xk with Ao > 0 such that (Pl, • • • .p„)(T*) = £ Л, (T*, х*(Г)) / (0,..., 0), (6.54) max р(Г*)./(Г’,ж*(Т*),и;) = -^А^(Т*,а;’(Г*)) . (6.55) wEu *—' (Jt i=0 Finally, the function t i—► p(£)-±*(£) in (6.53) coincides a.e. with an absolutely continuous function, satisfying u*(t))} = P(0 Dtf(t, x*(t), u*(t)). (6.56) Proof. 1. We shall apply Theorem 6.3.1 to an auxiliary optimization problem in n + 1 space variables with fixed terminal time T*. Set x = (xq, x) E IRn+1, u = (и$,и) E IRm+1 and consider the problem max </>o(x(T*, u)), (6.57)
6.4 Variable terminal time 117 for the (n + 1)-dimensional system f ±o(r) = u0(r), [ i(r) = uo(t)/(xo(t), x(t), u(t)) , (ar0, z)(0) = (0,5) subject to the constraints </>i(x(T*)) = 0, i = 1,..., k. (6.58) (6.59) (6.60) (6.61) uo(t) e |, 2 , u(r) e U for a.e. r G [0,7*]. The rationale behind these definitions is the following. The additional state variable xq plays the role of the time t in (6.50). Indeed, (6.58) implies dx dx dr dx0 dr dx0 U Moreover, the new time variable г is a reparametrization of the old time t. In this way, we can allow t to range in a variable time interval [0, T], while т always ranges on [0,T*]. 2. Assume now that и* : [0, T*] U is optimal for the original problem (6.49)-(6.51). Then u* = (l,u*) is optimal for (6.57)-(6.61). Indeed, let v = (vo,v) be another admissible control, with 0o(x(T*,v)) > </>0(T*,a:(T*,w*)), &(x(T*, v))=0 (1 < i < k). (6.62) Since Xot'r') = Vq(t) e [1/2, 2], one can invert the function r i—> ^o(^) = £(t) and construct a control u“(t) = v(r(t)). Using this control in the original system (6.50), at the terminal time T = xq(T*) by (6.62) one obtains </>o(T, x(T,rz“)) ></>0(Г*, x(T*,u*)), <к(Т,х(Т,и*)) = 0 (l<i<fc), against the optimality of u*. 3. Applying Theorem 6.3.1 to the optimal control u* = (l,tz*) for the prob- lem (6.57)-(6.61) and recalling that xq = t, we obtain the existence of an absolutely continuous adjoint vector p = (po?p) with the properties Po(<) = ~p(t) Dtf(t, , (6.63)
118 6 Necessary conditions PoW • 1 + p(t) • f(t, = max (po(£)u>o + p(i) • cvof(t,x*(t'),cu)\, (6.64) j<w0<2, ueu I- J k / (po(T*),.. . ,pn(T*)) - £ A, (^, gi,..., g J (T*,x*(T*)). (6.65) i=0 4 ' for some constants Ло,..., Afc with Aq > 0. Since the maximum in (6.64) is attained when cuq = 1 and cj = a*(i), we must have PoW = ~P(C • № z*(f), a.e. (6.66) Since po is absolutely continuous, from the first equation in (6.63) and from (6.66) we deduce the identity (6.56). The adjoint linear equation (6.52) follows from (6.63). The maximality condition (6.53) is derived from (6.64). The terminal condition (6.54) is a consequence (6.65). Finally, (6.66) and (6.53) together imply -po(0 = max p(Z) • /(t, x*(t), w) Vte [0,T*]. Indeed, the two sides are continuous and, by (6.53), they coincide almost everywhere. Together with (6.65), at t = T* this yields (6.55). Remark 6.4. Defining the payoff function </>o(T, x) = — T in (6.49), one obtains a minimum time problem. In this case, дфц/dxi = 0 for all i = I,..., n. Therefore, defining the target set at time T as 5(T) = {x; фг(Т, x) = 0, i = 1,..., к} , (6.67) the conditions (6.54) states that the adjoint vector p(T*) is perpendicular to S(T*) at the terminal point x(T*). Remark 6.5. In connection with the system (6.50), define the Hamiltonian H(t, x,p, u) = p • /(t, x, u). If u*(-) is an optimal control and x*(-),p(-) denote the corresponding trajec- tory and adjoint variable, from (6.50) and (6.52) it follows that the corre- sponding system of O.D.E’s has the Hamiltonian form dx* 3H(t,x*(t),p(t), u*(Z)) dt dp dp dH(ty x* (t), p(t), u* (t)) dt dx (6.68)
6.5 The problem of Bolza 119 6.5 The problem of Bolza This section is concerned with the Bolza problem with running cost min I L(t, x(t\ u(t))dt (6.69) uezv /л subject to x = f(t, x(t), u(t)), д:(0) = ж, ?z(t)eU, (6.70) with the terminal constraints Фг(Т, x(T, it)) = 0, i = l, ...,fc. (6-71) Assuming that L is continuous in all variables and continuously differentiable w.r.t. t,x, the above problem can be recast in Mayer form, introducing the auxiliary variable xn+i(t) = / L(s, a:(s),u(s)) ds Jo This yields a maximization problem with (n + 1)-dimensional state variable: min жп+1(Т), uEu (6.72) subject to the terminal constraints (6.71), for the system with dynamics ( Xi = fi(t, x(£), u(tY) i = 1,..., n [±n+i = L(t,x(t),u(t)) e u (6.73) and initial conditions (j?i,...,xn+i)(0) = (ж1,...,жп,0). (6.74) From Theorem 6.4.1 we deduce Theorem 6.5 .1. (PMP, Bolza problem). Let f and L be continuous in all variables and continuously differentiable w.r.t. t,x. Let the bounded con- trol и* : [0, T*] U be optimal for the problem (6.69)-(6.71) and assume that vectors ..., (T*, rc*(T*)), i = 1,..., k, are linearly inde- pendent. Then there exists a nontrivial adjoint vector p = (p1?... ,pn) and constants Ло,..., Afc with Ao > 0 such that, for almost every t e [0, Г*], UXi uXi (6.75)
120 6 Necessary conditions p(0‘ /(^ u*(0) + АоД*, rr*(i), u*(t)) = min [p(t) fit, #*(<), ш) + A0L(t, x*(t), w)|, (6.76) d г . . _z + .. . T. * z 4 df(t,x*,u*) x dL(t,x*,u*) -jAp(t)-f(t,x ,u ) + XoL(t,x ,u ) !> = p(t)-——- + A0 —— , (6.77) (Pi.... ,p„)(T*) = £ Аг ^(Т*,а?*(Т)),..., ^(Г ,Z(T))^ , (6.78) min {p(T*) /(Г*, s*(T*), w) + Aoi(T*, x’(T*), u>)} (6.79) Proof. Notice that min f L = — maxj(—L), thus the control u* solves also a maximization problem for the cost with changed sign. We can then ap- ply Theorem 6.4.1 to the corresponding problem (6.71), (6.73) and (6.74) on JRn+1 with cost —jrn+i. In this case, we would have an adjoint vector p = (рь ... ,pn+1) satisfying the evolution equation (6.52). However, since neither f nor L depend on Tn+i, the evolution equation for pn-i-i is trivial: 4;Pn+i(t) = - ]Tpj(t)-^—(t,x*(t),u*(t)) Ol ‘ OX^^. i ~Рп+1(^)д( ~ (t,X*{t),U*{t)}=0. ОХп^\ Hence pn+i(t) = pn+i(T*) = Ao for all t G [0, T*]. We then set pi = — рг for г = 1,..., A: and pn+i = pn+i = Ao. The identities (6.75) (6.79) now follow from the corresponding statements in Theorem 6.4.1. For example, (6.76) follows from p(t) • /(i, x*(i), u*(t)) 4- Ao (-b)(t, = max {p(0 • u) + Ao x*(t), w) multiplying by —1 and recalling the definition of p. Remark 6.6. For the control system (6.50), consider the problem of reaching a (possibly moving) target set S in minimum time. If S(T) is described by (6.67), this can be formulated either in the Mayer form (6.49)-(6.51), taking </>o(T, ж) = T, or in the Bolza form, taking L(t,x, u) = 1. Assume that the trajectory rr* = #(-,?/*) is optimal for the problem of minimizing the time T subject to
6.5 The problem of Bolza 121 яг(Т, и) € S = {x; Фг(х) = 0, г = 1, , (6.80) ± = /(ж,и), ar(O) = x, u(t) e U. (6.81) Notice that here f does not explicitly depend on time. Using (6.75), (6.77) and (6.78), in this special case where L = 1 we obtain p(t) = -P(t) Dxf(x*(t), p(t) • /(#*(£), !£*(£)) = COriSt. к p(T*) = i=l Remark 6.7. If L(i, x, u) > 0 for every t, x, u, then the Bolza problem (6.69)- (6.71) can be reformulated as a minimum time problem. Indeed, consider first the autonomous case min / L(a?(£),w(£)) df, (6.82) u Jo subject to (6.80)-(6.81), with free terminal time T. By a simple rescaling of time, (6.82) is transformed into the minimal time problem for the auxiliary system dx = G и, = e s dr L{x,u) To handle the general case, just observe that any time-dependent problem can be rewritten as an autonomous one, introducing the auxiliary variable x„4-i = t, together with the equations £n+i = 1, жп+1(0) = 0. As an application of Theorem 6.5.1, we shall derive the usual necessary conditions for an extremum, for the standard problem in the Calculus of Variations: min / L(t, x(t), ±(t)) dt, (6.83) *(•) Jo subject to z(0)=z, x(T)=y, (6.84) with x, у € lRn. Theorem 6.5 .2. (Euler-Lagrange and Weierstrass necessary condi- tions). Assume that L is continuously differentiable w.r.t. all variables t,x,x. Let #*(•) be a Lipschitz continuous function which attains the minimum for the problem (6.83) (6.84)- Then (i) The function t i—> ^(t, x*(t), £*(£)) coincides a.e. with an absolutely con- tinuous function, such that d \dL(+ * •*< — -T— lt.X ,X ) dt dx. ox (6.85)
122 6 Necessary conditions (ii) The function t ►—> L(t,x*(t), ±*(t)) — |4(t, x*(i), £*(£)) • x*(t) coincides a.e. with an absolutely continuous function, such that d r/ * V- dL , * — L(t,X ,X )- > 7— (t,X ,X )• Xa dt dxi L i=l (6.86) (Hi) For almost every t G [0, T] and every ш G IRn, one has L(t,x*(t),w) > L(t,x*(t),x*(t)) + J*)’-(ш-х*(0)• (6.87) Proof. The problem (6.83)-(6.84) is a special case of (6.69) (6.70), for the system with simple dynamics i(t) = u(t), u(t) G HV1 with fixed terminal time and terminal point. By Theorem 6.5.1 there exists a nontrivial, absolutely continuous adjoint vector (pi,... ,pn) and a constant Aq > 0 such that dL Pi(t) = -Xo — (t,x*(t),x*(t)), (6.88) A0L(t,a:*(i),i*(Z))+^2pi(t)i’‘(<) = min < A0L(t,ar*(Z),w) + ^p^Z)^ > . i=i weIR" I «=i ) (6.89) If Aq = 0, since p is nontrivial and iv is arbitrary, the infimum in (6.89) would be identically equal to — oo, and could never be attained as a minimum. Therefore, multiplying Ao and all components of the adjoint vector p by the same positive constant, we can assume Aq = 1. In order that w = ±*(f) yield the global minimum in (6.89), we must have d dwi L(i,ar*(t),w) + 57pi(t) i=l hence Pi(t) = -f^(«, ±*(t)) (6.90) dxi at a.e. time t. Taking the derivative w.r.t. time and using (6.88) we obtain (6.85). From (6.89), with Aq = 1, we also deduce L(t,x*(t),u) > L(i, rr*(t), x*(t)) + p(f) • (£*(*) - Vcj G IRn. (6.91) By (6.90), p = —dL/dx, hence (6.91) yields (6.87). Finally, since Aq = I, (6.77) and (6.90) together imply (6.86).
6.5 The problem of Bolza 123 Fig. 6.7. In dimension n = 1, the graph of L must lie above the tangent line at In this example, the condition (6.87) holds provided that the derivative satisfies i*(t) < a or x*(t) > b. Remark 6.8. The necessary condition (6.87) has a clear geometric meaning (see figure 6.7). Namely, the graph of the function w >—> L(f, a?*(t), a;) lies entirely above the tangent plane at the point u = x*(t). Example 6.5 A landing vehicle separates from a spacecraft at time t = 0 at an altitude h from the surface of the planet, with initial velocity vq. For simplicity, consider vertical motion only and assume that the gravity acceleration g is constant. Let Xi denote denote the altitude, X2 the velocity and let u(t) be the thrust exerted by the rocket motor, subject to |tz(£) | < 1, with a suitable rescaling. The equations of motions are = (aj2, u — g), (a;1,a:2)(0) = (Ji, v0) For a soft landing at some time T we require (x1,a;2)(T) = (0,0). As performance index, we choose a linear combination of fuel consumption and total time. This leads to the problem min f (|iz(t) | + k) dt. Jo The Pontryagin’s equations (6.75)-(6.76) here take the form (P1,P2) = (0, -Pl),
124 6 Necessary conditions Pi^2 + P2(u* - g) + A0(|u*| + k) = min {Pix2 +р2(ш - g) 4- А0(|ш| + к)}. kl<i These in turn yield Pi(0 = рг, p2(t)=p2 + (T-t)pi, for some constants Pi,p2, while the control law is determined by -1 if p2(t) > Ao, 0 if - Ao < p2(t) < Ao, 1 if p2(t) < -Ao. Since p2 is a linear function of t, observing that u* = 1 for t sufficiently close to T (to avoid crashing), by the minimality condition the optimal control must have the form u*(t) = < 0 if л < t < t2, if и < t < T, for some switching times 0 < Ti < т2 < T. The terminal condition (6.78) here does not yield any additional information. Indeed, we have <^i(x) = #i, ф2(х) = x2, and (6.78) states that the vector (pi,p2)(T) is a linear combi- nation of V^i = (1,0) and V</>2 = (0,1). On the other hand, from (6.79) we further deduce Pi(T>2(T) +P2(T)(u*(T) - p) + A0(|u*(T)| + k) = p2(l - p) + A0(l + k) = 0. From the relations P2(n) = Ao, P2(t2) = -Ao, x 1 + к P2\T) = -Ao j _ = ' — 1 if t < Ti, 1 and the fact that p2 is a linear function of t it now follows ti + r2 2 1 - g _ t2 - л 1 + к ~ 2 Using this additional condition on the switching times, a unique optimal con- trol iz* is determined, satisfying the terminal conditions (xi,x2)(T) = (0,0) at some time T. Example 6.6. An enemy airplane flies at constant speed V at an altitude h above the ground. Using a?i,a;2 as coordinates in a vertical plane, its position at time T is given by (rr1,ar2)(T) = (fc + VT, h). (6.92) A rocket is launched from the origin at time t — 0, with zero initial speed. Its motor can produce an acceleration of magnitude A whose direction can be
6.6 Linear-quadratic optimal control 125 controlled. Calling x = (^1,^2) its position and v = (^1,^2) its velocity, the equations of motion are: (±i, X2, Vi, ^2)^) = (*h> ^2, A cos zi(i), Asin u(t) — g), where g is the gravity acceleration and the angle и is the (unrestricted) control function. One seeks the control law и = u*(t) that will intercept the airplane in minimum possible time. We are thus considering a minimum time problem, with terminal constraint (xi,£C2)(^) = (H VT, h). Pontryagin’s equations here take the form (Р1,Р2,91,9г) = (0,0, -pi,-p2), </i(i)cos u*(t) + q2(f)sin u*(t) = min {<?i(f)cos w + (ftWsin w}, i.e., tan u*(f) — — 92(£)/9i(i). The terminal conditions (6.92) further imply 91(7') = 92(7) = 0, hence 91W =Pi(7-t), 92(t) =p2(T-t). Since <7i, (72 are linear functions of /, the optimal control law thus has the form u*(t) = arctan — This shows that optimal controls must be constant. 6.6 Linear-quadratic optimal control In this last section we apply the Pontryagin Maximum Principle to a spe- cial class of optimal control problems, frequently encountered in engineering applications, with linear dynamics and quadratic cost. Consider the system: ±(t) = A(f) x(t) + B(t) u(t), x e IRn, и e IRm, (6.93) where A is a n x n matrix and В is a n x m matrix. The optimal control problem, with fixed terminal time T, takes the form • fT mm / Jq [г? R(t)u + xr*Q(£)^] dt, a:(0) = x, (6.94) where U is the class of all measurable controls, R is a symmetric m x m matrix, Q is a symmetric n x n matrix, and t denotes the transpose. We assume that each Q(t) is positive semi-definite and that the matrices R(t) are uniformly positive definite, so that xfQ(t)x > 0, ulR(t)u > 0|tt|2
126 6 Necessary conditions for some 0 > 0 and all x € IRn, и E Rm, t E [О, Т]. In particular, this implies that the symmetric matrices R(t) are invertible. Roughly speaking, we wish to keep the system close to the origin during the whole interval [0, Т]. The integral in(6.94) penalizes the distance of x from the origin, and the energy spent by controlling the system. The problem (6.94) is in Bolza form, hence we shall apply Theorem 6.5.1. The minimality condition (6.76) can be written as p(t)- [4(t)x*(t) + B(t)tz*(0] +Ao[(M*(i))t^)M*(Z) + (a:*(Z))t(2(^*(i)] = = min |p(t) - [x(t)x*(i) + 4- Ao|o?7?(£)w + (a;*(t))<Q(f)a;*(t)j Notice that Aq 7^ 0. Otherwise, since there is no final constraint, (6.75) and (6.78) would imply р(«) = -p(t) • лад, p(T) = 0, hence p(i) = 0, reaching a contradiction. Multiplying Aq and each component of the adjoint vector p by the same positive constant, we can assume Aq = 1. Since the dynamics is differentiable w.r.t. и and и can vary in the whole space IR™, the minimization condition implies that the following quantity is vanishing: This, in turn, implies p(t) • Z?(£) + 2(u*(t)//?(£) = 0. Hence the optimal control satisfies u‘(i) = -17?-1(^(i)pt(f). (6.95) The optimal trajectory for the linear-quadratic optimal control problem is thus found by solving the two-point boundary value problem x = A(t)x - , p = — pA(t) - 2xtQ(f), with initial and terminal conditions x(0) = x , p(T) = 0 • The corresponding optimal control is then obtained by inserting the solution p(-) in (6.95).
6.6 Linear-quadratic optimal control 127 Problems 6.1. In Theorem 6.3.1, assume that the target is defined also by means of some one-sided constraints, i.e. S = {a; G IR71; ф^х) = 0 i = l,...,fc, </>j(rr) > 0 j = Let ж*(-) = a;(-,?z*) be optimal. At the terminal point x*(T), assume that ^(x*(T)) = 0 > = 1,...,£, ф^х\Т))>0 j = and assume that the k + £ gradient vectors V<^o(^(T)),..., V0fc(^*(T)), V4>i(z*(T)),..., V^*(T)), are linearly independent. Prove the existence of a nontrivial adjoint vector p(-) such that (6.29)-(6.30) hold, together with к £ p(T) = £ Aj;V^(x*(T)) + £ A'V^(X*(T)) i=0 j=0 with Ai,..., А^ G IR, Ao, A^, ..., A^ > 0. 6.2. Show that the trajectory (xi(t), x2(£)) = (0,0) is optimal for the problem max xi(T) и for the system (±i,±2) = (tt,O), u(t) e [—1,1], with initial and terminal constraints (^i,^2)(0) = (0,0), (^i,a?2)(T) G S = {ж G IR2; ф(х) > 0} , where ф^х^х?) = x% — x*. However, there exists no adjoint vector p(-) which satisfies (6.29), (6.30) together with p(T) • V > o Vv e ts+ where Ts+ = {(3/1,372); У1 > 0} is the tangent cone at the origin to the set S+ = {(rri,x2) G S, хг > 0} . Explain why this does not contradict the maximum principle proved in the previous exercise.
128 6 Necessary conditions 6.3. As in (1.9) consider the control system (ii,i2) = (l-arl + ux, u2) modelling the motion of a boat on a river. Here the point (^1,^2) is constrained to the strip {(^1,^2)» ^2 £ [— 1, 1]} and the set U of admissible controls consists of all measurable functions и : IR 1—> IR2 taking values inside the closed disc U — { (cJl, 1л12 ) * у CJ j M } • Assume that the initial position is (jti, ж2)(0) = (—1,0). Write the Pontryagin necessary conditions, for a trajectory which reaches a given point у = (1,6) on the opposite shore in minimum time. Find the trajectory which reaches some point у — (1,^2) 011 the opposite shore, in minimum time (Notice that in this second case the coordinate X2 is free). Write the Pontryagin optimality conditions for this problem. 6.4. A population of fish in a lake evolves according to the equation x — x(a — x) — x и . (6.96) Here the control u(t) 6 [0,1] denotes the intensity of fishing activity. An initial population x(0) = x and a terminal population x(T) = у are assigned and one wants to maximize the payoff function 0(T) = I [x(t)u(t) — k?z2(£)] dt Jo i.e. the total amount of fish caught from the lake minus the cost of fishing. We assume here that о. к > 0 and 0 < x, у < a. Write the PMP necessary conditions for this problem. 6.5. As in (1.14), consider the system (±1,^2) = (^2/tt), u(t) e [-1,1] modelling the control of a cart on a straight rail. Given any point x = (^1,^2), find the control which steers the system from the origin to x in minimum time. Compute the minimum time function T = T(x). Show that it is smooth on the entire plane IR2 except along the two curves 7+ = {(^i,.t2); xi = x%/2, x2>0}, 7" = {(a?i,x2); Xi = -x^/2, x2 < 0}.
6.6 Linear-quadratic optimal control 129 6.6. Consider the system e [-1,1], discussed in Example 6.4. Fix any point x = (Si,x2) € IR2 with x% > х^/З. Write the Pontryagin equations satisfied by a control u(-) steering the system from the origin to x in minimum time T. Show that there exists a control reaching x, having one of the forms u(t) = +1 if t € [0, r], -lif te]r,T], u(t) = (-1 if t e [o,r], ( +1 if t €]т, T], (6.97) satisfying the additional condition т > T/2. Show that such control is unique if 0, while there are two such controls if x\ =0. These are the only optimal ones. Hint: if u(-) is a control of the form (6.97) but with т < T/2, show that there exists a second control u(-) of the same type, which steers the system to the same point x(T, u) = x(T,u) but at an earlier time T < T. 6.7. Let A, В be an n x n and n x m matrix, respectively. Let z : [0, T] i—> IRn be a smooth function, a > 0. Write the Pontryagin necessary conditions for the optimal tracking problem min [ \x(t) - z(t)|2 + a|u(t)|2 dt, u Jo in connection with the linear system on IRn x = Ax 4- Bu, z(0) = 0, u(t) 6 IRm. 6.8. A control ?/*(•) is called a strong [weak] local maximizer for the problem (6.1)-(6.3) if there exists 6 > 0 such that i/?(a;(T,u*)) > ^(t(T, u)) for every control и 6 U such that ||u — u*||li < d [respectively: for every control и e U such that supte[0 Tj \u(t) — u*(£)| < J]. (i) Assuming that u* is a strong local minimizer, prove that all state- ments of Theorem 6.1.1 still hold. (ii) Assume that the set U C IRm of admissible controls is convex, and let u* be a weak local minimizer. Prove that the statements in Theorem 6.1.1 still hold, with the maximality condition (6.5) replaced by (6.17). 6.9. Recalling the definitions introduced in the previous problem, consider the optimal control problem: max x(T), utu '
130 6 Necessary conditions x = и3 — и2, ж(0) = 0, with Т > 0 and Ы = {и : [0, Т] —> [—2, 2] и measurable}. (i) Prove that u*(^) = 0 is a weak local maximizer. (ii) Show that (6.5) fails. In particular, u*(t) = 0 is not a strong local maximizer, and it does not satisfy the Pontryagin necessary conditions. 6.10. Assume that the function #*(•) provides a minimum for the standard problem of the Calculus of Variations (6.83)-(6.84). Prove that, for almost every t € [0, T], n QU) = £ x*(t), x\t))^ > 0 € IRn, iJ=0 i.e. the quadratic form Q is positive semi-definite (Legendre necessary condition). 6.11. Let £*(•) afford the minimum for the standard problem of the Calculus of Variations (6.83)-(6.84). Assume that the derivative x* is piecewise continuous, with a jump at some intermediate time t. Prove that the right and left limits i(t+), x(t~) of the derivative satisfy the Erdmann corner conditions: OX ox L(t, x*(t), i*(f+)) — L(t, rr*(t), £*(£+)) i*(t+) — ±*(i_) 6.12. A particle of mass m moves on a smooth horizontal plane with rectangu- lar coordinates y. Initially the particle is at rest at the origin. During a fixed time interval [0,T], the particle can be accelerated by applying a force F(t) of constant magnitude |F| = к and arbitrary direction. The angle 0(t) made by the force with the positive a?-axis is the (unrestricted) control variable. At the terminal time T, we want the particle to be mov- ing along a given line parallel to the ж-axis (say, the line {y = c}), with maximum speed. (i) Find conditions on к, с, T which guarantee the existence of at least one control $(•) satisfying the requirements. (ii) Show that the optimal control is given by 0*(t) = arctan(a T bt) for suitable constants a, b. 6.13. Suppose that a cup is initially filled for 4/5 of its capacity with hot coffee, at temperature 0^. Cold milk is then poured inside at a rate u(t) e [0,1],
6.6 Linear-quadratic optimal control 131 until the cup is completely full. The coffee can be drank as soon as its temperature decreases to a prescribed value 0± < 0q. We wish to minimize this time. Calling x(t) the amount of liquid in the cup and 0(t) its temperature, we have the equation x = u, x(0) = 4/5, 0(0) = 0O • Find the control function rz(-) which minimizes the time at which the state (ж,0) = (l,0i) is reached. 6.14. As a variant to the previous problem, suppose that the cup is initially already full of coffee, at the temperature 0q. Milk is poured into the cup, at a controlled rate u(t) 6 [0,1], so that the liquid overflows. The total amount of milk is limited by u(t) dt — ttiq. In this case the total amount of liquid in the cup x(t) = 1 remains constant in time. The thermodynamic equation describing the temperature of the liquid in the cup is bilinear, namely 0 = — 3 — и — Ou . Here the first term on the right hand side corresponds to the heat loss to the surroundings, the second term is due to the inflow of cold milk, and the third term is overflow. Determine the admissible control u(-) which reduces the temperature of the cup to the value 0i in minimum time. 6.15. Consider the linear quadratic problem min / [jq(£) + x£(t) + cu2(t)] dt Ц) Jo for the system ±1 = X2 ( #i(0) = Xi ±2 = и [ £2(0) = ^2 Write the necessary conditions for the optimality of a control u*(-).

7 Sufficient Conditions Consider the optimization problem in Bolza form inf < I L(t, x(t, u),-u(t)) dt + (T, x(T,u)) U^u [Ло (7-1) (7.2) (7-3) for the system x = f(t,xyu), with initial and terminal constraints u(t) E U a.e., x(t) = x, (T,x(T)) e S. It is well known that the PMP provides only a necessary condition for op- timality: a control tt*(-) may satisfy Pontryagin’s conditions and yet not be optimal. For example, it may provide a local minimum to the functional (7.1), but not the global one. In this chapter we describe four techniques for proving global optimality. (I) Prove that an optimal solution exists. If u*(-) is the only admissible control that satisfies the PMP, then u* must be the optimal one. (II) Consider a Mayer problem with fixed terminal time. If the set of ad- missible controls reaching the target S is convex, and if the functional и ^(^(T, u)) is convex, then any control satisfying the PMP is in fact optimal. (Ill) Embed the problem (7.1)-(7.3) in a family of problems, by varying the initial data (f, £)• Compute trajectories satisfying Pontryagin Maximum Principle. If such trajectories cover in suitably regular way the state space, then all trajectories are optimal. (IV) Compute the optimal value function V = V(t, ж), for all initial data (7,ж), by solving a related Hamilton-Jacobi partial differential equation. Then, any control ?/*(•) which attains the optimal value V(t,x) is optimal.
134 7 Sufficient Conditions The first three techniques will be covered in the following sections. The fourth approach, based on the value function, will be discussed at the end of this chapter in a special case, i.e. for linear-quadratic problems. A more general treatment, based on the theory of viscosity solutions to Hamilton-Jacobi equa- tions, will be given in Chapter 8. As in Chapter 3, we shall always assume that the control system satisfies the basic assumptions (II) below, except for the special Linear-Quadratic problems treated in Section 7.5. (H) The set U C IRm of control values is compact, is an open subset of IR x П1п, the functions f : J? x U i-> IRn, L : L2 x U IR are continuous in all variables and continuously differentiable w.r.t. x. 7.1 Existence + PMP. Theorem 7.1.1. Let the hypotheses (H) hold. Assume that an optimal so- lution for the problem (7.1)-(7.3) exists. Assume that the control functions ,n/c(-) are the only admissible ones that satisfy the PMP. Then, among the controls ,Uk, the one which yields the lowest value of the cost (7.1) is optimal. Indeed, by Theorem 6.3.1, the optimal control u*(-) satisfies the Pontryagin Maximum Principle. By the above assumptions, we must have u* = Uj for some j e {1,..., k}, and the result is obvious. We illustrate the theorem with a simple example. Example 7.1 Consider the system (®1,Ж2) = (u,Xj), (xi,x2)(0) = (0,0), |u(t)| < I te [0,2]. We seek an optimal control for the problem max {xi(T, и) + x%(T, tz)} и with fixed terminal time T = 2. Observe that the reachable set at any time t is bounded. Indeed |a;i(«)| < t, |x2(<)| < j s2 ds = y. Moreover, for each t, x, the set of admissible velocities F(t,x) = {(u,: — 1 < и < 1}
7.2 Convexity + PMP. 135 is a compact, convex subset of IR2. Therefore, by Theorem 5.1.1 in Chap- ter 5, an optimal solution exists. The PMP yields the adjoint equation and maximization condition (P1,P2) = (-2xiP2,0), (Р1,рг)(2) = (1,1), u = sign(pj). These in turn imply p?(t) = pi = -2±ip2 = -2u = -2sign(pi), pi(0) = 0, pi(2) = 1. The graph of pi(-) therefore is the union of finitely many arcs of parabolas. By direct inspection, we find the only two solutions: Pi(i) = 5-t2, if 0 < t < 1, if 1 < t < 2. The corresponding extremal controls are ut(f) = | ! 1 = 1, if 0 < t < 1, if 1 < t < 2. The trajectory corresponding to the control u*(-) is (xi,x2)(t) = (t, t3/3) te [0,2]. The trajectory corresponding to ?J(-) is / t3\ I —t, — ] if 0 < t < 1, (xi,X2)(t) = < z q /, _ \ ( t - 2, ( } ) if 1 < t < 2. \ 3 3 J ~ ~ Computing #i(2) + #2(2) in the two cases, it is found that iz*(-) is the only optimal control. 7.2 Convexity + PMP. We consider here the Mayer problem with fixed terminal time: inf </>o(*(T,u)), (7.4) uQU for the system (7.2), with initial and terminal constraints x(to) = xq, x(T) € S = {x; = 0, i = 1, • • • , k}. (7.5) We recall that a function f : A —> IR, where A is a convex subset of a vector space, is called convex if for every x, x' G A and Л € [0,1] one has: f(Xx + (1 - A)x') < A/(x) + (1 - A)/(a/).
136 7 Sufficient Conditions Theorem 7.2.1. In addition to the basic hypotheses (H), let Duf be contin- uous and assume that the set of admissible controls which steer the system to the target set Us = {и : [0, T] U, x(T, u) € S} is convex. In addition, assume that the functional и > фо(х(Т, и)) from Us into IR is convex. Then, any trajectory &*(•) = a?(-,u*), which satisfies Pon- tryagin’s equations, with к p(T) = V^0(a:*(T')) + £ Х^ф&*(ТУ) (7.6) 2=1 for some Ai, • • • , Xk € IR, is optimal. Proof. Let the control «*(•) and its corresponding trajectory x*(*) satisfy the PMP. Assume that there exist a different control u^ E Us whose trajectory t (£) = x(t, г?) satisfies 0о(^(Т)) <0о(х‘(П)- (7-7) A contradiction is then obtained as follows. Define the 1-parameter family of controls ti«(i) = eu+(t) + (i-e)u*(t) ce [0,1]. Notice that E Us, because Us is convex. Let M(•, •) denote the fundamental matrix solution of the linear problem v(t) = Dxf(t,x*(t),u*(t)} • v(t). By Theorem 3.2.6, we then have Фф^х(Т,и^) _ dx(T,ut) £=0 = V^(a-*(T)) / • Duf(t,x*(t),u*(t)) • (г?(<)-u*(t))dt Jo = 0 = (7.8) because x(T, u^) E S for all Using (7.6) and (7.8) we now obtain <У0о(а;(Т,ы?)) dx(T>ut) -----—------ = Уф0(х (Г))--------—----- f=0 = V<Ao(x*(T)) • I Duf(t,x*(t),u*(t)) (u\t) - u*(t))dt Jo = p(T)- I Af(T,i)-D,J(i,2:*(0,u*(f))-(wt(t)-u*(t))</i Jo = [ p(t) Duf(t, x*(f),u*(t)) • (uf(t) - u*(t))dt Jo
7.3 Dynamic Programming 137 because, by the PMP (see Remarks 6.2 and 6.3 in Chapter 6), the last inte- grand is a.e. nonnegative. In particular, this implies фо(х(Т,и*У) > ф0(х(Т,и*)) - | [</>о(^(Г,и*)) - ф0(х(Т,и]У)] , (7.9) for all £ > 0 sufficiently small. On the other hand, the convexity of the map и i—► ф0(х(Т, и)) implies <£о(я(Т,?/)) < </>0(x(T,u*)) -£ [ф0(х(Т,и*У) - <£0(а:СГ,г?))] V£ € [0,1]. This yields a contradiction with (7.7) and (7.9), proving the optimality of the control ?z*. Observe that the assumptions of Theorem 7.2.1 are satisfied whenever фо : IRn и-> IR is convex, the control set U and the target set S are both convex and the system is affine, having the form x = A(t)x + B(t)u + c(t). Indeed, if u, u' e ZY, £ 6 [0,1], one has Ф0(х(т,£и+(i -e)u')) = Фо^^и) + (i - ew,tz')) < С Фо(х(Т, u)) + (1 - £)ф0(х(Т, и')). This technique applies in particular to the Examples 6.1 and 6.3 in Chapter 6. 7.3 Dynamic Programming This section is concerned with the minimization problem inf ф(Т,х(Т,иУ) (7.10) uet/ for the control system x = f(t,x,u), u(t) e U a.e., (7.11) subject to the terminal constraints (T,j:(T))€S, (7.12) where S C IR71-*-1 is a closed target set. We always assume that the basic hypotheses (H) hold and that ф : S •—► IR is continuous and bounded below. Given an initial condition ж(^о) = xo, (7-13)
138 7 Sufficient Conditions call £4o.To the family of all measurable controls и : [to? T] U, for some T > to? such that the corresponding trajectory of (7.11), (7.13) satisfies the condition (7.12) at the terminal time T. In order to keep track of the initial conditions, we write x(-,u,to,xo) for the solution of x = f(t,x,u) x(to) = Xq- (7-14) We now consider a family of optimization problems with different initial conditions z(s) = y, and study how the values of these optimization problems vary with s, y. We thus define the Value Function as V(s,?/)= inf x(T, u, s, ?/)), (7.15) adopting the convention that V(s, y) = oc if Us>y is empty. Some basic prop- erties of V are proved in the next theorems. Theorem 7.3 .1. (Properties of the Value Function). Let V = V(s,y) be the value function for the problem (7.10)-(7.12). Then (i) For every и € U and every admissible trajectory x(-,u), the map t V(t, xft, u)) is nondecreasing. (ii) If и* : [t0,T] i—> U is optimal for the problem (7.10)-(7.12), then V(t,x(t,u)) is constant for t E [fo?^1]- Proof. 1. Let и : [to? U be any admissible control. We use the notation xi = x(ti, u, to'Xo) and assume that V(t^xx) = V(t0,x0)-6 (7.16) for some e > 0. A contradiction is then derived as follows. Let v : [ti, T] i—> U be a control in for which ^(T,x(T,v,tx,x^ < V^Xi) + (7.17) Define the control й : [t0,T] •-> U by setting uft) if te [t0,ti], vft) if te^T], Then u e UtQ,xG because x(T. u. to, Xq) = x(T,v,ti,Xi). Moreover, (7.16) and (7.17) together imply г[>(Т, x(T,u,to,xo)) < V(to,xo) - |, u(t) = contrary to the definition of V. This proves (i).
7.3 Dynamic Programming 139 2. Let t h-> u*(t) be an optimal control, for the initial data z(£q) = #o- Calling t и-> x*(t) = x(t,u*the corresponding trajectory and using (i), we deduce V(t0,x0) = i/>(T,x*(T)) > V(t,x*(t)) > V(to,xo) e [to, Т]. This proves (ii). Our main goal is to derive a partial differential equation satisfied by the value function V = V(s,y). In the following, we write dsV= £-V, respectively for the partial derivative of V w.r.t. the time s. and for the gra- dient of V w.r.t. the space variables у = (t/i,..., yn). We regard 4yV as a row vector, while f(s,y,u) E IRn is a column vector. Theorem 7.3 .2. (P.D.E. of Dynamic Programming). Assume that the value function V is C1 on some open set Q C IR x IRn, not intersecting the target set S. Then, at every point (s,y) € Q, the function V satisfies the Hamilton-Jacobi equation dsVM)+ inf{VyV(S,y)-/(S,y,^)} = 0. (7.18) Proof. 1. Let (s,y) E Q and consider any constant control u(t) = w E U. Then by Theorem 7.3.1 at = dsV(s,y) + VyV(s,2/) • f(s,y,u) > 0. Since uEU was arbitrary, the left hand side of (7.18) is > 0. 2. To prove the converse inequality, consider a sequence of controls uy : [s, Tv\ h-> U such that ^(Tp,x(Tp,Up,s,z/)) < V(s,7/) -I-1/”3, (Tp, x(Ty,uu,s,y\) e S. (7.19) Since (s,y) S, we can assume Ty > s + 6 for some 6 > 0 and all у > 1. Calling x„(t) = x(t,uy,s,y), we claim that, for all v sufficiently large, there exists ty E [s, s + i/-1] such that •^(L/) --- f(fv, ^1/(^1/)» ^1/(^1/)), (7.20) a5V(tp,Xp(tp)) + VyV(tp,j:p(^)) • /(^,Xp(Zp),Up(tp)) < i. (7.21) Indeed, if our claim does not hold, one would have
140 7 Sufficient Conditions for a.e. t 6 [s, s + и !], hence > V (.s + p 1,xJ/(s + b' *)) > V(s,y) + in contradiction with (7.19). 3. By (7.21) we have inf (tu,xu{tu)) + 7?уУ(^,^(^)) • f(tu,Xy(ty),w) < As y—>oc we have ty—*s, ху(1у)—+у. Therefore, the continuity of dsV. 4yV and f, we conclude that the left hand side of (7.18) is < 0. This completes the proof. Remark 7.1 (Bolza Problem). Consider a minimization problem in the more general form (7.1). Then the value function V(y, s) = inf L(t, x, u) dt + ^(T, rr(T)) Ф) = у (7.22) satisfies the Hamilton-Jacobi equation asV(S,y)+ inf{VvV(m) -f(s,y,iv) + L(s,y,a>)} = 0. (7.23) cue и Remark 7.2 (Minimum Time Problem). If the problem is autonomous, i.e., f = f(x,u), tl) = L = L(t,ii). S = {(f,.r); t € IR, x € £}, then the value function V is independent of time. Hence (7.23) takes the form inf {VV(y)-/(y,w) + L(y,u?)} =0. cueu In particular, consider the minimum time problem: x = /(x,u), u(t) E U for a.e. ,t x(s) = y, x(T) 6 S'. Then the value function V = V(y) describes the minimum time taken by a trajectory starting at у in order to reach the target set S'. On a domain where V is C1, one has inf{VV(7/)./(t/,iv) + l} = 0. (7.24)
7.3 Dynamic Programming 141 Remark 7.3 (Optimal Feedback Controls). By theorem 7.3.1(ii), if a trajectory x*(-) = rr(-,u*) is optimal, then £V(t, x* (t)) = 0. Therefore, ^^W) + W^W)‘/(^*W^*(0) = 0 for a.e. t at which V(t,rr*(t)) is differentiable. The knowledge of V therefore allows us to derive an expression for the optimal control u* = u*(t,x) in feedback form, solving the equation dtV(t,x) + VxV(Z,x) • f(t, я, u*) = 0 in terms of u*. Equivalently, u*(£, x) = arg —max VxV(t, x) • /(£, жси). cuEU In general, however, the trajectories of ± = f(t, x, u*(t, x)) need a careful interpretation. Indeed, the function (£,x) u*(t, x) may be discontinuous, multivalued, or not everywhere defined. By definition, a control u* : [to, T] h-> U is optimal for the problem (7.10)- (7.13) if and only if (T,z(T,u*)) € S, ^(T,x(T,u*)) = V(t0,z0). The knowledge of the value function V thus provides a straightforward crite- rion for optimality. Using Theorem 7.3.2, one may hope to recover V as the (unique) solution to the Hamilton-Jacobi equation (7.18), with boundary data on the target set S: V(s,y) =t/>(s,y) (s,y) e S. (7-25) This approach runs into a major difficulty. Since (7.18) is strongly nonlinear, globally defined C1 solutions do not exist, in general. In fact, in most cases, the value function is continuous but only piecewise differentiable, with derivatives dsV, V.V having jumps along a finite number of submanifolds of positive codimension. Example 7.2 Consider the minimum time problem on IR: for the system dt, x = u, u(t) € [-1,1] x(s) = у,
142 7 Sufficient Conditions with target set x(T)eS = {xelH\ |ж| = 2}. The value function independent of time, is then equal to the distance of у from the set consisting of two points {—2, 2}: V(S/) = ||»|-2|. As in (7.24), it satisfies the Hamilton-Jacobi equation min {VyV-u>+l} = 0 (7.26) uc[—1,1] almost everywhere on IR2\S, together with the boundary conditions V(y)-0 if 12/1 = 2. (7.27) Clearly, V is not differentiable at у = 0. Observe that V cannot be charac- terized as the unique piecewise C1 solution of (7.26)-(7.27). Indeed, we can choose any piecewise constant function Л : IR i-> {—1,1} such that y* h(x) dx = 0, and define y W(?/) = [ h(x)dx, J-2 This provides another piecewise C1 solution to the same equation with the same boundary conditions. Given a continuous function W : IR x IRn IR, we shall seek sufficient conditions which guarantee that W is the value function for the minimization problem (7.10)-(7.12). If W is globally C1 and satisfies the equation (7.18), then a uniqueness theorem for classical solutions of Hamilton-Jacobi equations would guarantee the equality W = V. If W is only piecewise Cl, the problem is more subtle. A result in this direction is Theorem 7.3.3. Consider the problem (7.10)-(7.12). Let Q C IRn+1 be an open set containing the closed target set S, and letW : Q i—♦ IR be a continuous function such that (i) W > V, (ii) W = on S, (iii) At every boundary point (t,x) € dQ, one has W(t, x) = M = sup W(s, y), (7.28) (s,2/)€Q
7.3 Dynamic Programming 143 (iv) There exist finitely many manifolds Mi, • • • ,Mw C Q with dimension < n, such that W is continuously differentiable and satisfies the H-J equation Ws(s,y) + min{%(s,y) • /(s,y,u>)} = 0 cuGLJ at every point (s, y) in the open set <2\ U Л4,. Then W coincides with the value function V on the closure of the domain Q. Fig. 7.1. A nearly-optimal trajectory is approximated by one with piecewise con- stant control. In turn, this is approximated by a trajectory having transversal inter- sections (or no intersection at all) with the manifolds Mt where W is not smooth. Proof. 1. Because of (i), it suffices to prove W < V. Assume, on the contrary, that W(tQ,x$) > V(Iq,Xq) for some (to^o) € Q- Since И7(/0,т0) < M, this of course implies V(to,^o) < Af. By the continuity assumptions on W, there exist e, 6 > 0 such that V(t0,z0) + e < M (7.29) |z - x0| < 6 W(t0,x) > V(t0,x0) + e. (7.30) 2. Let iz* : [to,!1] U be a control whose corresponding trajectory £*(•) = z(-, w*, to, xq) satisfies (T,x*(T)) e S, il>(Tyx*(T)) < V(t0,x0) + e. (7.31) If uy : [to ,T] i—> U is a sequence of piecewise constant controls approaching w* in the L1 norm, by Theorem 3.2.1 in Chapter 3 the trajectories x„(t) = x(t, Uy, T, x*(T)),
144 7 Sufficient Conditions with the same terminal point as x*, converge to #*(•) uniformly on [to, Т]. In particular, there exists a piecewise constant, left continuous control : [to?T] *—► U such that the trajectory а;й(-) = x(-, u\ T, x^(T)) satisfies (ar^to) - zo| < (7.32) 3. Let t"] be an interval on which the control u$(t) = ш is constant, and such that (t,x#(t)) € Q for all t Then - IV(t',a;l,(t')) > 0. (7.33) Indeed, by the Transversality Theorem 2.4.1 in Chapter 2, there exists a se- quence of points zm^>x^(t') such that the solution xm(-) to the Cauchy prob- lem tu), Xffi^t ) — Zm has only transversal crossings with each manifold Mj. Hence (t, xm(t)) € Uj-Mj only at finitely many times t. We can thus use the assumption (iii) and obtain W",M*")) - W(t',xm(t')) (7.34) = / [Ws($,zm(s)) + Wy(s,j:m(s)) • /(МпМu)] ds > 0. (7.35) Since W is continuous, letting m—>oc in (7.34), we recover (7.33). 4. We now consider two separate cases. First, the graph {(t, x^(t)) : t G [to? ^]} is assumed to be entirely contained in Q. If is constant on the intervals with to < ti < • • • < t/v = T, from (7.33) it follows N W(T,x\T)) - W(t0,x*(t0)) = [Wi.^(ti)) - > 0. i=l (7.36) By construction, хЦТ) = x*(T), hence (7.36) and (7.31) together imply Wo,^(*o)) < W(T,x*(TY) = Ж*‘СП) < V(to,xo) + £. (7.37) Recalling (7.32), this yields a contradiction with (7.30). 5. It remains to consider the other case, where (t,x$(t)) is not always inside Q. Since Q is a neighborhood of the closed target set S, the time f where the trajectory rH(-) makes its last entrance in Q satisfies f = inf{i<T; (S,x*(s))eQ Vse[t,T]}<T.
7.3 Dynamic Programming 145 Moreover, by the assumption (ii), lim = M. t^>T+ Choose т e ]т, T] such that (7.38) W(r, ^(t)) > M - E. (7.39) Recalling (7.29), the same argument used in (7.37) now yields Ж(т,^(т)) < РИ(Т,^(Т)) = ^(T,xtt(T)) < V(t0,x0) + e< M - e, a contradiction with (7.39). Remark 7.4. In the case where Q C 1R+ x IRn, the conclusion of the theorem still holds if we assume that (7.28) is satisfied only for the boundary points (t, x) € dQ with t > 0. In practice, one can use Theorem 7.3.3 in the following way. Assume that, for every initial data (t, y) in a neighborhood Q of the target set S we can construct a trajectory 11—> xs,y(t), which reaches the target set and achieves a cost W(s,y). Even if this trajectory satisfies the Pontryagin Maximum Prin- ciple, a priori there is no guarantee that it be optimal. However, assume that the function IT is piecewise smooth, and satisfies the P.D.E. of dynamic pro- gramming (7.18), outside finitely many smooth manifolds All,..., of lower dimension. Then, using the theorem, we would like to conclude that our function W actually coincides with the value function, i.e. W(t, x) = V{t,x) for all (2, x) € Q. By construction, the assumption IT > V is automatically satisfied. The only remaining condition to achieve is (iii), requiring that the global maximum of W on Q should be attained at all boundary points of Q. For this purpose, it is often useful to replace the domain Q with the sublevel set Q' = {(t,x) 6 Q; W(t,x) = M}, for a suitable value of the constant M. Another situation where the same arguments in Theorem 7.3.3 can be applied is illustrated below. Corollary 7.3.4. Consider the problem (7.10)-(7.12), where the target set is S = {T} x lRn. Let Q =]£,T[xIRn and let W : Q i—> IR be a continuous function such that (i) W > V, (ii) W = onS={T}x IRn. (iii) There exist finitely many manifolds A4i, * • • , C Q with dimension < n, such that W is continuously differentiable and satisfies the H-J equation
146 7 Sufficient Conditions Ws(s,y) + min{Wy(s,p) • f(s,y,a>)} = 0 at every point (s, y) in the open set Q\ U Mi. Then W coincides with the value function V, on the closure of the domain Q- Proof. By (i), again it is enough to prove the inequality W < V. All steps 1. to 4. in the proof of Theorem 7.3.3 remain valid. Indeed, they are based on the continuity of W and on the assumption (iii). By the form of the domain Q, it is obvious that any trajectory (£,a^(£)) cannot leave or re-enter Q before the terminal time T. Therefore, the situation considered in step 5. never occurs, and the proof is complete. Example 7.3 For the linear system (±i,±2) = (x2,u), u(t) € [—1,1] a.e., (7-40) consider the problem of reaching the origin in minimum time min{T; (aq,a;2)(T) = (0,0)}. Since the system is autonomous, the value function does not depend on t and satisfies the Hamilton-Jacobi equation min W(yi,2/2) • (y2,w) =-I- (7-41) Pontryagin’s equations yield (РьРг) = (0, -Pi), w(t) = -sign(p2(i)). Hence pi is constant and p2 is a linear function of t. which can change sign at most once. Therefore, the only controls which may be optimal are bang-bang with at most one switching: Observe that the trajectories of (Jq, ±2) = (rr2,±l) are arcs of parabolas aq = C ±^2/2* Any initial condition (?/i,?/2) can be connected to the origin by two such arcs. More precisely, call If У1 > ^(^2), the point (?/i, У2) can be steered to the origin along the two parabolas I X if A Xi = I Pl + у
7.3 Dynamic Programming 147 The point of intersection (xi,^) of the above parabolas (see figure 7.2) is computed as / 2 z- - \ / У1 , У2 (xi,x2) = I — + — , Recalling that |^21 = Ы = 1, the total time taken to travel along the two arcs is ^(271,3/2) = I3/2 - x2\ + Ы = 3/2 + 2 -4 For 2/1 < 92(2/2) the analysis is entirely similar, showing that W(y\,y2) = W(—2/1, —y2). Observe that W is smooth except at the origin and on the two manifolds Г 2/2 Л41 = < (2/1,3/2) : 3/2 = у, 3/2 < 0 f 3/2 A42 = < (3/1,3/2) : 3/1 = ~y5 3/2 > 0 Moreover, in the region where 2/1 > ^(^2), the Hamilton-Jacobi equation (7.41) takes the form f dw dw ^€[-1,1] [ дуг dy2 = -1. + min Therefore, on Q = IR2 all of the assumptions in Theorem 7.3.3 hold. Hence all trajectories are optimal. Fig. 7.2. Optimal trajectories reaching the origin in minimum time.
148 7 Sufficient Conditions 7.4 Relations between the P.M.P. and the P.D.E. of Dynamic Programming Consider an optimization problem in Mayer form max (7-42) uElA for the control system x = f(t,x, u), w(£)cU. (7.43) Call V — V(r,y) the maximum payoff attainable with the initial condition x(t) = y. By Theorem 7.3.2, on regions where V is smooth, this value function satisfies the Hamilton-Jacobi equation dTV + H(r,y,W) = 0, (7-44) where H(r,y,p) = max p f(r,y,u>). (7-45) Aim of this section is to establish a basic connection between the P.D.E. (7.44) and the Pontryagin equations (6.4)-(6.5). V=V T Fig. 7.3. A characteristic curve. As a preliminary, we recall here the method of characteristics, for the solution of the first order P.D.E. (7.44). Consider any differentiable curve t (£,#(£)) in t-x space (see fig. 7.3). We seek a system of O.D.E’s describing how the function V varies, restricted to this particular curve. From (7.44) it follows ~V(t,x(t)) = dtV + i-VV = -H(t,i,VV) + iVV. (7.46) at Of course, (7.46) cannot be solved by itself, because the right hand side con- tains the unknown quantity W(t, x(t\). We thus introduce the gradient vector p(t) = W(£,z(£)), so that
7.4 Relations between the P.M.P.and the P.D.E. of Dynamic Programming 149 Differentiating (7.44) w.r.t. хг, i = 1,... , n, we obtain d dH dp, dH air' + дх, + Эх, Op, ~ ' and we also have ^t,x(t)) = %+± ir£. p.48) Cot- С/i J=1 J The identity VXiXj we thus obtain — VxjXi yields dpi/dxj = dpj/dxi. From (7.47) and (7.48) d / / \ \ —Pi(t,x(t)) = ЭЯ ЭЯ ул. d^_ dxi “ dxi dpj " lj dxj (7.49) Notice that, in order to compute the time evolution of p = W along an arbitrary curve, according to (7.49) one also needs to know all the quantities dpi/dxj. However, if we choose a special curve such that x = dH/dp, then the last two terms on the right hand side of (7.49) cancel out. We thus obtain the closed system of n + n O.D.E’s Xi = Pi = dH dpi ’ dH dxi > i = 1,..., n . (7.50) For each terminal point x, the hamiltonian system (7.50) can be solved with terminal conditions x(T) = x, p(T) = VV(T,x). (7-51) Next, we show that, under suitable regularity conditions, the hamiltonian system of O.D.E’s (7.50) coincides with the Pontryagin equations. Indeed, consider the hamiltonian function (7.45) and write H(t,x,p) = p • f(t, x,u(t,x,p)), (7-52) where u(t, x, p) = arg max p • f(t, x, u). (7.53) To avoid technicalities, we suppose the above maximum is assumed at an interior point of U. This is certainly the case if U = lRm. The above definitions then imply ЭИ d f d f du d I (7.54)
150 7 Sufficient Conditions Indeed, p • df /ди = 0 at an interior point where the maximum is attained. Moreover, by the same argument we obtain yy Q f du — = f(t,x,u(t,x,p)) + p- — (t,x,u(t,a:,p)) = f(t,x,u(t,x,p)). op OU op (7.55) By (7.54)-(7.55), the hamiltonian system (7.50) is equivalent to the two vector equations x = f(t,x, u(t, х,рУ) , df p = —p • — (t,x.u(t,x,pY), (7.56) dx with u(t,x,p) defined by the maximality condition (7.53). In conclusion, the above analysis has shown that the equations of char- acteristics for the Hamilton-Jacobi P.D.E. of dynamic programming coincide with the Pontryagin’s equations (7.56), (7.53). 7.5 Linear-quadratic case In this section we apply the methods of Dynamic Programming to a special class of optimal control problems with linear dynamics and quadratic costs. More precisely, we consider the system: x(t) = A(t) x(t) + B(t) u(t), x e ИГ, и e IRm, (7.57) with A n x n matrix and В n x m matrix, and the optimal control problem mm / Js [??#(t)u + xfQ(t)x] dt + х(ТУ5х(Т), z(s) = У, (7.58) where U is the class of all measurable and a.e. bounded controls, T is fixed, R is a m x m matrix, Q and S arc n x n matrices, and * denotes the transpose. We make the following assumption: (H*) The functions t A(t), t —> B(£), t R(t), t —> are all measurable and bounded on compact intervals. The matrices R(t),Q(t) and S are symmetric. Q(t) and S are positive semi-definite, i.e. xtQ(t}x > 0 for every x e ПГ and same for S', while R(t) is positive definite, i.e. z?H(t)u > 0 for every и G ПГ and equality holds only for и = 0. Using the expression for solutions to linear systems, we notice that the cost function is quadratic in и and y. With this in mind, as a candidate for the value function, we consider W(s,y) = у'Р(з)у, (7.59) where P is a symmetric n x n matrix to be determined.
7.5 Linear-quadratic case 151 Theorem 7.5 .1. The function (7.59) is continuously differentiable and is the value function if and only if matrix valued function t h-> P(t) is a solution to the system of O.D.E’s ( P(t) = - P(t)A(t) - A4t)P(t) - Q(t) ( . \P(T) = S ’ U J The matrix equation (7.60) is known as the Riccati Differential Equation. Proof. 1. Assume first that the function W in (7.59) is differentiable and coincides with the value function. The terminal condition at time t = T clearly implies P(T) = S. Recalling Remark 7.1, we have: 0 = <9.SIT + min • [A(s)z/ + B(s)cu] Tcuf/?(s)cj + ylQ(s)y^. (7.61) Since the expression in braces is a convex quadratic function of cu, the mini- mum is attained for the value of satisfying: ° = {^yW ' + + Wffl(s)w + y'<2(s)i/j = [p(S)y + P‘(s)y] + 7?(s)w + R\s^. Recalling that P(s) is symmetric and R(s) is symmetric and invertible, we obtain (7.62) Substituting this value of w in (7.61), for every у 6 IRn we have 0 = y* [p(s) - P(s)B(s)/?-1(s)Bt(s)P(s) - 2X‘(s)F(s) - Q(s)] У- Since А*(з)Р(8)у is scalar, it is equal to its transpose yl P(s)A(sfy. In con- clusion, if the value function has the form W(s, у) = у*Р($)у, then the differ- ential equation in (7.60) must hold. 2. Next, assume that P solves (7.60). Observe that the matrix P(t) is symmetric for every time t. Indeed, its transpose Pl solves the same equation and has the same terminal condition P*(T) = S = P(T). We want to apply Corollary 7.3.4 to the function W(s, y, z) = W(s, y) + z on Q =] — oo, T[xIRn x IR and with dQ = S — {ffT,y,z) : у € IRn,z 6 IR}. Define xSiV as the trajectory corresponding to the feedback control u(t,x) = -R~1(t)Bt(t)P(t)x
152 7 Sufficient Conditions as in (7.62), with initial condition xs,y(s) = y. Then the same computation as in 1. shows that W is constant along (xs,y, z) for z(s) = 0. In particular: W,?/,0) = W(T,x„,v(T),z(T)) = T I [us,y(t)R(t)ua,y(t) + x*s,y(i)Q(t)arSiJ)(O] dt + x^y(T)Sxs,y(T), J s where us,y(fi) = w(xs,y(t)Y Hence the assumption (i) holds. Since P(T) = S, the condition (ii) is also satisfied. Finally, the computations of 1. show that W satisfies the P.D.E. of dynamic programming on the entire domain Q. Therefore, W coincides with the minimum value function. To solve LQ problems we only need to show that there exists a solution to the Riccati Differential Equation: Theorem 7.5 .2. The backward Cauchy problem (7.60) has a unique solution defined for all t € ] — ос, Т]. Proof. The Riccati equation is an O.D.E. on Rn*n whose right hand side is continuous w.r.t. time and smooth (a quadratic polynomial) w.r.t. the space variable. Therefore, the Cauchy problem admits a unique local solution. This solution can be continued backwards up to a maximal interval [f, T], with lim ||P(f)|| = +oo. (7.63) t—T+ We claim that (7.63) cannot happen for a finite time t. Indeed, for every t < T, the value function W(t,y) is bounded by the cost of the trajectory corresponding to control и = 0, i.e. 0 < W(t,y) = у1Р(у)у < f y'dT + у1М*(Т,1)8М(Т,1)у. By assumption (//*) there exists C > 0 such that ||M(t, t)|| < 1 + C(r - t). This shows that ||P(t)|| is bounded on every bounded interval of time. Hence no blow-up occurs and the solution of (7.60) is well defined for all times t <T. Once the solution to (7.60) is found, an optimal control can be easily recovered from the HJB equation. In fact since W in (7.59) is differentiable the optimal control is given by (7.62) thus we have: Corollary 7.5.3. The optimal feedback control for the problems (7.57) (7.58) is given by u\s,y) = -R~1(s)Bt(s)P(s)y, where P solves (7.60).
7.5 Linear-quadratic case 153 Next, we show how the asymptotic stabilization problem can be solved also in terms of a linear-quadratic optimization. Consider now the linear au- tonomous system i(t) = + zeIRn, (7.64) where A is an n x n matrix and В is an n x m matrix. For given strictly positive, symmetric matrices P Q consider also the optimal control problem min, y* [u*Ru 4- xlQx\ dt. x(0) = x. (7.65) i.e. problem (7.57) (7.58) with A.B.R.Q constant and S = {0}. We assume that the linear system (7.64) is controllable. We shall study the limit of optimal trajectories, as T tends to oo. More precisely, we claim that, as T —> oo, the optimal trajectories tend to the solutions of a stabilizing feedback. Given T > 0 let us denote by the value function of (7.64)-(7.65) and by P^ the solution to (7.60) with S = {0}. First, since (A, B) is controllable, for every x and T we can find a control й : [0, T] —> lRm such that the corresponding trajectory satisfies x(T) = 0. For every T > T. we can prolong й on [0, T], setting u(t) = 0 for t > T, thus we get: Гт х^Р<т)(0)^ = V^T\x) < / [й*Рй 4- xlQx\ dt= 4- x*Qx] dt. Jo Jo This proves that the function T i—> xJP(T\tyx is uniformly bounded. Given Ti < T2 let u be the optimal control for x and time T2, then: гТг xJP^'^x — V^T2\x) = / 4- x*Qx\ dt Jo fT1 > / [й*Яй + xfQx] dt > V(Tl\x) = xJP^^x. Jo Hence, for each x € IRn, the map T н-> х1Р^т\0)х is nondecreasing and uni- formly bounded. Therefore, it admits a limit as T —► 00. Since x is arbitrary, we conclude that P^1 —► P^ for some symmetric, positive definite matrix Poo- We claim that the feedback control defined as u(x) = -R~1BtP(Xix. (7.66) stabilizes the system (7.64) asymptotically to the origin. The existence of a monotone limit implies that the time derivative of the map T 1—> P(T)(0) tends to zero. Since the map t 1—> P^T\t) satisfies (7.60) and p(T-e)(0) = P(t\e), we have 0 = - lim inf p(T\0) — lim -^-(P^T\s)) T—>oo dT v 7 T^oodsy
154 7 Sufficient Conditions = lim (P^XtyBR-^pC1")^) - Р(Т\О)Л - A(P(T)(0) - (?), T—»oo Therefore the matrix satisfies the algebraic Riccati equation 0 = PooBR-1BtPao-PxA-AtPoo-Q. I (7.67) Now fix an initial point x(0) = x, and let x(-) be the trajectory of the linear system (7.65), using the feedback control (7.66).Using (7.67) we find ^xXt)Pxx(t) = (xt(t)At+ut(x(t))Bt)P00x(t)+xt(t)P00(Ax(t')+Bu(x(t))) = = -x^^^BR^B^^ + Q)x(t). The definition of value function \РТ) implies rT xl о 0 oo + Q]x(t)dt Г d dt = - —хХ^РоаХХ) dt = XtP<x)X - х‘(Т)РооХ(Т) . Jo Taking the limit as T —► oc we obtain lim х1Р^х < x^P^x - lim х*(Т)Роож(Т'), T—>oo T—>(x> and hence lim xt(T)Poox(T) = 0. T -too Since Рю is positive definite, this implies we conclude lim^oo #(7") = 0. The other property of a stabilizing feedback is also easily checked. 7.6 Optimal syntheses In this section we illustrate an alternative method to ensure optimality for various initial conditions. We consider the optimal control problem: inf [ L(x(t, u), u(t)) dt, (7.68) «(•)ew Jto for the system x = f(x,u), a.e., (7.69) with initial and terminal constraints z(0) = x0, x(T) = 0. (7.70)
7.6 Optimal syntheses 155 We assume hypothesis (H) holds. A synthesis on an open set J? is a collection of trajectories Г — {^х : x € 12,7* steers x to S}, and Г is optimal if every is. The main goal of this section is to prove that a synthesis, formed by trajectories satisfying the Pontryagin Maximum Principle, is optimal provided it covers the entire space in a regular fashion. Given a single trajectory satisfying PMP there is no regularity condition which ensures optimality, as shown by next example. Example 7.4 Consider the control system (±i,±2) = (u, 1 + xl\ (xi(0),z2(0)) = (0,0), |it| < 1 and the problem of reaching the point (0,1) in minimum time. If we choose the control и = 0 then we get the trajectory 7 given by: ^i(f)=0, x2(t) = t, (7.71) that reach the final point in time 1. The equation for the adjoint variable is: (Ai, A2) = (-2a;iA2,0) and, using (7.71), one easily verifies that defining Ai(Z) = 0, A2(t) = 1, 7 satisfies Pontryagin Maximum Principle. Moreover 7 and the corresponding control are analytic, even more: polynomial. However, one easily check that any control й satisfying u 7^ 0, /* й($) ds = 0, t + I ( I й(т) dr\ ds = 1, Jo Jo \Jo / steers the origin to (0,1) in time t < 1, thus 7 is not optimal. The degeneracy of the example can not happen for a synthesis, indeed we introduce a concept of regularity ensuring optimality. Our interest is to con- sider the case in which Г is generated by a feedback which is piecewise smooth in the following sense. We can partition the space into regular manifolds of various dimensions, such that the feedback is smooth on each region. To give a precise definition we need some notation. A set P G IRn is said a curvilinear open polytope of dimension p, if there exists a polytope (i.e. bounded closed region intersection of a finite number of half-spaces) P' G IRP and a smooth map ф : 1RP —> lRn, injective with jacobian having maximal rank at every point, such that ф(Р' \ дР') = P. Let 12 be an open set containing the origin. We say that P is a Boltyanskii- Brunovsky regular synthesis, briefly BB synthesis, if the following holds. There exists a 6-tuple S = (P, Pi,Р2,П, 27, it) such that
156 7 Sufficient Conditions (BB1) P is a collection of curvilinear open polyhedra and <2 is disjoint union of elements of P. If P3 Pk e P and Рк П Pj 0 then Pk C dPj and dim(Ffc) < dim(Pj). {0} € P and the elements of P are called “cells”. (BB2) P\{{0}} is the disjoint union of P\ (the set of “type I cells”) and P^ (the set of “type II cells”), (BB3) the feedback и : {x : 3Pi E Pi,x E Pi} —> U and П : Pi —> P are maps, 27 : P2 —> Pi is a multifunction, with non empty values, such that the following properties are satisfied: i) The function и is of class C1 on each cell. ii) If P] € Pi then f(x, u(x)) E TxPi (the tangent space to Pi at x) for every x E Pi. In addition, for each x £ Pi, if we let £x be the maximally defined solution to the initial value problem £ = /«, £(0) = ®, ее Fl, (7.72) and define tx = sup Dom(£x), then the limit £x(tx —) := lim^ £ж(£) exists and belongs to 77(Pi). iii) If P2 E P2, then for each x E P2 and P E ^7(P?) there exists a unique curve : [0, [ —► J? such that the restriction of to ] 0, t? [ is a maximally defined integral curve of the vector field /(-,!/(•)) on P, and ef (o) = x. iv) On every cell Pj E P\, x —> tx is a continuously differentiable function, and (t,x) —> £x(£), (£,#) “► := ?z(&rW) are continuously differentiable maps on the set E(P) := {(t,x) : x E Pi , t E [0,^]}. If P2 E P‘2 the same holds for every , u* , with P E 27(P2). v) For every x E f?\{0}, the trajectory : [0, Tx\ —> IRn, yx e Г, is obtained by piecing together the trajectories on every single cell. Moreover, yx changes cell a finite number of times. Remark 7.5. Notice that condition (BB1) essentially means that the collec- tion of open polyhedra form a ’’triangulation” of the set 12, see Figure 7.4. In the same figure we represent the two type of cells to illustrate properties (BB2) and (BB3). The cell Pi is of type I and dimension equal to 2, then there are trajectories running on it, corresponding to the feedback u, which end on the cell 77(Pi). On the contrary, the cell P2 is of type II and dimension equal to 1, hence from every point there start some trajectories entering other cells. More precisely, from every point of P2 it starts one trajectory entering the 2-dimensional cell P3 and one entering the 2-dimensional cell P4, this means S(P2) = {P3,P4}. Theorem 7.6.1. (Boltyanskii-Brunosky sufficiency theorem) Let Г be а В В synthesis on IRn formed by trajectories satisfying PMP, then Г is opti- mal.
7.6 Optimal syntheses 157 Fig. 7.4. Example of BB synthesis. Proof. We define a candidate value function Wr in the following way: ^(1) = I L(4x(t),u(?ix(t)))dt. Jo We claim that Wr satisfies the assumptions of Theorem 7.3.3. 1. First, fix x belonging to a cell Pi of maximal dimension n (which nec- essarily is of type I). By BB3) ii), from every x, in a neighborhood of ж, it starts a trajectory £x corresponding to it(rr), defined up to time tx and run- ning on the cell Pi. Then, by BB3) iv) the functions x —> tx, (t,x) —► £x(t) and (t,x) —» ux(t) := u(^x(t)) are continuously differentiable. By BB3) ii), the trajectories £x end on the cell 77(Pi) with x £x(tx) continuously differ- entiable. Then we can use again BB3 ii), or iii), and iv) for the cell 77(Pi) and prolong the functions x —> tx, (t,x) £x(t) and (t,x) —> ux(t) in a continu- ously differentiable fashion. Using BB3) v), by recursion we prove that these functions are defined and continuously differentiable up to times Tx. Then, it easily follow that Wr is differentiable at x. 2. Let us denote by (A, Ao) the adjoint covector along (7z,7fc)- Setting x = x + ev, we want to compute: pTx />тх Xo(Wr(x+Ev)-Wr(xy) = / A0L(7x(t),?ix(t))dt— / XqL(^x (t), ux(t))dt. Jo Jo From 1. it follows: ||7x(t) -7И011 =O(e)- (7.73) Assume Tx < Tx, the other case being similar. In the following we consider integrands which are also of order O(e), hence we can compute all integrals up to Tx, possibly defining the integrands to be zero after Tx and adding a o(e) term. Hence we can write:
158 7 Sufficient Conditions Ao(Wr(z + ev) - ВД = = [ X0(L(yx(t),ux(t)) - L(7S(f),uI(t)))di Jo + I A0(i(7i(t),uI(t)) - L(^x(t),ux(t)))dt + o(e) Jo = Л + /2 + o(e). (7.74) We start estimating Д: Ii = Ao [ I DyL(yx(t) + 0(yx(t)• (yx(t) - ix(t))d0dt Jo Jo = Ao У У [£>JZ£(7x(0 + ^(7x(0-7®(0)>ux(0) -DyL^x(t) + #(7x(0 ~ 7*(0)> «*(<))] ’ (7x(0 - ^x^dOdt +Ao У У [£>y£(7S(t) + 0(^x(t)-yx(t)),ux(t)) - Иу£(7±(/),г4г(0)] •(7x(0 -'Ti^dO dt + [ X(jDyL(yx(t),ux(t)) • (7x(0-7x(t))dt, (7.75) Jo and the first two terms can be estimated as o(s). The equation (6.75) can be compactly written as A(f) = —A(t) • Dyf(^x(t),ux{x)) - X0DyL(yx(t),ux(t)). (7.76) In turn, this yields Ii= [ <-A(0 - A(t) • Dy/(7S(i),tti(0),7x(<) “ 7xW)dt + o(e) Jo JTs d = -y ^(А(0,7х(0-7гда + I (A(t),/(7x(t),«x(0) -/(7х(0,«х(0))л Jo - [ (A • £>v/(7i(t),tti(O),7x(0-7x(0M* +°(e) Jo = (A(0),7i(0) - 7г(0)) - (А(Тг),7х(Г^) - 7s(rx)> + [ (A(0,/(7s(<),Wx(0) - f(li(t),ux(t)))dt •h + I (A(t)r /(7x(0, ux(t)) - fbiit),ux(t)))dt Jo + / (A • £>уУ(7г(0,иг(0),7х(«) - 7ж(0)^ + «(£)• Jo
7.6 Optimal syntheses 159 Since 7x(Ti) = 7±(T±) = 0, the second addendum vanishes, while the sum of the last two integrals is of order o(s), indeed we can argue as for (7.75) replacing L with f. From the minimization condition of PMP (6.76) it follows (A(t) , /(7x(<),WzW)> + A0L(7x(t),«®(i)) < (A(i),/(7£(t),ux(t))) + A0L(7s(f),Ux(t)), for every x and almost every t. Therefore yT£ Л > (A(0),x - x) - Aq / A0(L(7^(t),ux(i)) - L(7^(^),Ux(f)))df + o(e). Jo Notice that the second addendum is precisely the term /2 in (7.74). Dividing (7.74) by e, using the above inequality, and passing to the limit as e goes to zero, it follows: X0(DyWr(x),v) > (A(0),v), and equality holds, since both terms are linear in v. Now Aq 0 otherwise A = 0, but this would contradict the non triviality of adjoint covector. Hence, it is possible to normalize Ao = 1 and finally obtain: DyWr(x) = A(0). Using again the the minimization condition of PMP (6.76), we have: (A(t),/(7x(0>w)> + M7x(t),w) > 0 for every ш € U and almost every t, with equality holding for ш = ux(t). From the continuity of /, L, A, and ux near 0, we get: {DyWr^x^f^x^}) 4-T(rr,cu) > 0, for every w with equality holding for cj = ux(t). Since x and Pi are arbitrary, this proves iv) of Theorem 7.3.3. 3. Conditions ii) and iii) hold trivially. Finally, since Wr(x) is the cost of the trajectory 7X, by definition of value function we have i). Therefore we can apply Theorem 7.3.3 and conclude that Wr coincides with the value function, but this exactly means that Г is optimal. Remark 7.6 Theorem 7.6.1 can be proved also for synthesis on an open set 12, assuming that Wr satisfies (iii) of Theorem 7.3.3. Various generalizations can be find in [71]. Example 7.3 (continued) Consider again the minimum time to origin prob- lem for the controlled equation (7.40). After straightforward computations, one checks that the optimal trajectories, represented in Figure 7.2, correspond to the discontinuous feedback (1.15) of Chapter 1. The collection of optimal trajectories form a BB synthesis, see Figures 7.5. There are four cells all of
160 7 Sufficient Conditions type I, on which it is defined the feedback (1.15). The first two Pi and P2 are of dimension 2 and are located, respectively, below and above the curve = —sign(x‘2)x’2/2. The cells P3 and P4 are of dimension 1 and form the same curve. The trajectories on cell Pi reach cell P3, hence Я (Pi) = P3, while trajectories on cell P2 reach P4, hence = P4. It is easy to check that assumptions (BB1)-(BB3) hold, hence this synthesis is аВ В synthesis. Example 7.5 Consider the minimum time problem to the origin for the control system: ( = и 1 • 12 5 [ X2 = —#1 - |^i where и e U = [—1,1]. The trajectories corresponding to constant controls ±1 can be described giving X2 as a function of x\. They are, respectively, cubic polynomials of the following type: X2 — T a a e IR #2 — “б" T + Q Ct G IR. (7.77) Consider the curve S = {(^1,^2) € IR2 • #1 = — 1}- It is easy to check that any trajectory running on S satisfies the PMP. In fact if 7 : [to, ^1] is such curve, then the corresponding control is constantly equal to 0 and the evolution equation for the covector is Ai — A2(l + ^i), A2 — 0. Hence the covector A is constant along 7. Moreover the maximality condition maxA • /(7(^),o;) = Ai?i(£) + ^A2, is always satisfied taking Ai = 0.
7.6 Optimal syntheses 161 Given b > 0, consider the trajectories 71 : [—6,0] »—> IR2 for which there exists to e [0,6] such that 71 corresponds to control -Fl on [—6, —to] and 71 corresponds to control —1 on [—to, 0]. Define also the trajectories 72 : [—6,0] 1—► IR2, b > 2, for which there exists ti e [b. 2] such that 72 corresponds to control — 1 on [—6, — ti] and 72 corresponds to control -hl on [—ti,0]. For every b > 2, these trajectories cross each other in the region of the plane above the cubic (7.77) with a = 0 and determine a curve К of points admitting two optimal trajectories. We use the symbols хл (b, to) and x~+(b, ti) to indicate, respectively, the initial points of 71 and 72. Explicitly we have 1 _ . . (2to — 6)3 (2to — &)2 2 ^0 x+- =2t0-b x+~ = -I- - v 0 J- +$ + -% (7.78) о 2 о + , „ , (b-2tO3 (£>-2<i)2 2 tf xi+=b-2t, z2+ = -----+ (779) О 2 о As b varies in [2, 4-oo[, the equation £+”(Mo) = z"+(Mi), (7.80) describes the set K. From (7.78), (7.79) and (7.80) it follows: to — b — ti ty — 2t2 T (2 4- 36)ti ~h (—62 — 26)^ — 0. Solving for ti we obtain three solutions: t\ = 0, t\ — 6, = 1 ~h —. The first two solutions are trivial, while the third determines a point of A, so that: f 21 К = Haq, x2) : zi = —2, x2 > -- > • I о I The optimal synthesis is a BB synthesis and is portrayed in Fig. 7.6. Notice that both S and К are cells of the BB synthesis, with S of type I and К of type II. Remark 7.7 A complete theory of two dimensional time-optimal syntheses, with many examples, can be find in [12]. Problems 7.1. Consider the optimal control problem: x = A x + В и, x(0) = ж,
162 7 Sufficient Conditions s Fig. 7.6. BB synthesis for Example 7.5. min / \xTQx 4- uTRu] dt ueu ,/0 L J where x E IRn, и E IRm, U — L1([0, T]; IR™), Q and R positive definite. Prove that Theorem 7.1.1 can be applied to this case. 7.2. Consider the optimal control problem: ±1 = 4- T2, x<2 — 2X2 + w, |u| < 1, max^j(T') + z(0) = 0, with T fixed. Determine the optimal control. 7.3. Prove that Theorem 7.3.3 can be applied to the case of unbounded value functions, replacing assumption hi) with: lim IV(ir) = 4-oo. x—^dQ 7.4. Consider a geostationary satellite and assume that the local motion around a stable orbit is given by the linear system: x = Ax 4- Bu. To ensure transmission, the satellite must track a given trajectory ?/(£), for t E [О, Т]. Assuming that the fuel consumption amounts to ulR.u for given matrix R positive definite, find the trajectory that minimizes the running cost sum of the fuel consumption and the distance from y(t). (Hint: Write the running cost and reduce to an LQ problem by adding a fictitious variable.) 7.5. Consider the minimum time to origin problem for the linearized pendulum with external force: x = x 4- u, x E IP, |u| < 1.
7.6 Optimal syntheses 163 (a) Use PMP to compute a candidate optimal trajectory for every initial point (x,i). (b) Compute a (discontinuous) feedback u(x) so that all trajectories of (a) solve x = x 4- u(x). (c) Compute the cost function W for trajectories of (a). Show that W is C1 outside two piecewise smooth curves contained in the first and third quadrant. (d) Show that we can apply Theorem 7.3.3 to W, thus all trajectories of (a) are optimal. (e) Show that trajectories of (a) form a BB synthesis and compute the corresponding cells. How many cells cover the set {(#,£) : x2 T x2 < 7r}. 7.6. Consider the minimum time to origin problem for the controlled equation (7.40). (a) Is the value function Lipschitz continuous? (b) A function W : IRn —> IR is said semi-concave if for every com- pact convex К C IRn there exists a constant Ck > 0 such that: W(скЕ-f- (1 — a)y) > aW(a?) + (1 — a)W(y) — Ck(x — y)2 for every x,y€K and a € [0,1]. Is the value function semi-concave? 7.7. Recall Example 7.5 and consider the open region 1? and the synthesis represented in Figure 7.7. The open region L? does not contain the point Fig. 7.7. BB synthesis on a region Г2. B, while the synthesis is the same as the optimal ones except for points on the left of К, for which we take the trajectories 71 instead of 72 (defined in Example 7.5.) We can define a candidate value function W computing the
164 7 Sufficient Conditions time along the trajectories of such synthesis. Prove that all assumptions of Theorem 7.3.3 are verified except iii).
8 Viscosity solutions for Hamilton-Jacobi equations Aim of this Chapter is to provide a concise introduction to the theory of vis- cosity solutions for first order nonlinear PDEs, and illustrate its applications to problems of optimal control. In the first section we review the classical method of characteristics, to construct solutions of the first order P.D.E. F(x, u, Viz) = 0 x G Q C IRn . (8.1) In general, the local smooth solutions obtained by this technique cannot be extended globally to the entire domain f2. Indeed, when two or more charac- teristic curves meet at a same point, a singularity occurs. In a typical situation, a boundary value problem for (8.1) will thus have no global smooth solutions. On the other hand, it may well have infinitely many piecewise smooth solutions, which satisfy the equation at almost every point of the domain. We then face the question of how to single out a unique “good” solution, relevant for whatever application we may have in mind. An answer is provided by the theory of viscosity solutions, introduced by Crandall and Lions in [34]. In essence, the main results show that • Letting e ОТ, the solutions ue(-) to the parabolic problems F(x, ue, Vue) = e Au£ converge to a unique limit zz(-). • This limit function и can be uniquely characterized by imposing certain inequalities on its upper and lower differentials, at each point x G L? where they exist. In Sections 8.2 and 8.3 we discuss some definitions and properties of super- and sub-differentials, and introduce the notion of upper and lower viscosity so- lution. The stability of viscosity solutions w.r.t. uniform convergence is proved
166 8 Viscosity solutions for Hamilton-Jacobi equations in Section 8.4. A basic comparison theorem, between an upper and a lower viscosity solution, is proved in Section 8.5. In turn, this yields the uniqueness of solutions in the viscosity sense. Applications to control systems are worked out in the last three sections. Namely, we characterize the value function for an optimal control problem as the unique viscosity solution to the corresponding first order P.D.E. prob- lem. This provides an alternative approach to the construction of optimal trajectories and to the study of sufficient conditions for optimality. In addition to [35], and [36], for a comprehensive monograph on the subject we refer to [8]. 8.1 The method of characteristics Consider a first order, scalar P.D.E., having the general form (8.1). It is conve- nient to introduce the variable p = Vn, so that (pi,... ,pn) = (uX1,..., uXn). Throughout the following, we assume that the F — F(x,u,p) is a continuous function, mapping IRn x IR x IR" into IR. Given the boundary data u(x) = й(х) x e 312, (8.2) a solution can be constructed (at least locally, in a neighborhood of the bound- ary) by the classical method of characteristics. The idea is to obtain the values u(x) along curves s h-> j;(s) starting from the boundary of 12, solving a suitable O.D.E. (see figure 8.1). Fig. 8.1. The method of characteristics. Fix a point у E 312 and consider an arbitrary differentiable curve s i—> x(s) with x’(0) = y. Call u(s) = u(x(s)), p(s) = p(x(sY) = Vu(x(s)). We seek an O.D.E. describing the evolution of и and p = Vu along the curve. Denoting by an upper dot the derivative w.r.t. the parameter s, we clearly have
8.1 The method of characteristics 167 Pj — 'U'XjXi • (8.3) In general, pj thus depends on the second derivatives of u, which at this stage are not available. Differentiating the basic equation (8.1) w.r.t. Xj we obtain dF dF dxj + du Uxj = 0. Hence dF dF dF dpi XjX' dxj du ^J (8.4) Instead of taking an arbitrary curve, we now choose a curve such that Xi = dF/dpi. By this specific choice, the right hand side of (8.3) is computed by (8.4), and can thus be expressed in terms of the variables x,u,p. We thus obtain a closed system of n+l+n equations where the second order derivatives uXiX. do not appear: ( +. - dF. 1 ~ дрг < й = Ё.гргт£- p 9F-. 9Fp dxj du i = 1, . . . , П j = 1,... ,n. (8-5) This leads to a family of Cauchy problems, which in vector notation take the form i=f^ dp й = p- op •n = _ dF * dx du P ' x(0) = у < u(0) = u(p) p(0) = Vu(j) yedn. (8.6) The resolution of the first order boundary value problem (8.1)-(8.2) is thus reduced to the solution of a family of O.D.E’s, depending on the initial point y. As у varies along the boundary of 12, we expect that the union of the above curves x(-) will cover a neighborhood of Э12, where our solution и will be defined. Remark 8.1. If F is linear w.r.t. p, then the derivatives dF/dpt do not depend on p. Therefore, the first two equations in (8.5) can be solved independently, without computing p from the third equation. Example 8.1. The equation |V-u|2 -1 = 0 x e n on R2 corresponds to (8.1) with F(x,u,p) = p2 + p| — 1. Assigning the boundary data
168 8 Viscosity solutions for Hamilton-Jacobi equations и = 0 x € dfi, a solution is provided by the distance function u(x) — dist (ж, dii). The corresponding equations (8.6) are x = 2p, и — p • x — 2 , p = 0. Choosing the initial data at a point у we have .r(0) = 7/, u(0) = 0, p(0) — n , where n is the interior unit normal to the set f? at the point y. In this case, the solution is constructed along the ray x(s) = у + 2sn, and along this ray one has u(.x) = |rr - y\. Assuming that the boundary dii is smooth, in general the distance function will be smooth only on a neighborhood of this boundary. If J? is bounded, there will certainly be a set 7 of interior points x where the distance function is not differentiable (fig. 8.2). These are indeed the points such that dist (.r, dQ) = |x — 7/11 = |.t — 7/21 for two distinct points 7/1,7/2 € dii. Fig. 8.2. Singularities of the distance function. The previous example shows that, in general, the boundary value problem for a first order P.D.E. does not admit a global C1 solution. This suggests that we should relax our requirements, and consider solutions in a generalized sense. We recall that, by Rademacher’s theorem, every Lipschitz continuous function и : ii 1—> IR is differentiable almost everywhere. It thus seems natural to introduce a concept of generalized solutions. A function и is a generalized solution of (8.1)-(8.2) if и is Lipschitz continuous on the closure f?, takes
8.1 The method of characteristics 169 the prescribed boundary values and satisfies the first order equation (8.1) at almost every point x G f?. Unfortunately, this concept of solution is far too weak, and does not lead to any useful uniqueness result. Recall Example 7.2. There exist infinitely many generalized solutions to the equation: x E [0,1], rr(O) = #(!) = 0, see figure 8.3 (left). Fig. 8.3. Left: Infinitely many generalized solutions. Right: a solution and a smooth approximation. Therefore, one seeks a new concept of solution for the first order equation (8.1), having the following properties: 1. For every boundary data (8.2), a unique solution exists, depending contin- uously on the boundary values and on the function F. 2. This solution и coincides with the limit of vanishing viscosity approxima- tions. Namely, и = ue, where the ue are solutions of F(x, ue, Vue) = s Au£ . 3. In the case where (8.1) is the Hamilton-Jacobi equation for the value func- tion of some optimization problem, our concept of solution should single out precisely this value function. In connection with Example 8.1, we see that the distance function if xe [0, 1/2], if xe [1/2, 1], is the only one, among those shown in figure 8.3 (left), that can be obtained as a vanishing viscosity limit. Indeed, any other generalized solution и with polygonal graph has at least one strict local minimum in the interior of the interval [0,1], say at a point x. If u£ —► и uniformly on [0,1], for some sequence of smooth solutions to |14| — 1 = euxx ,
170 8 Viscosity solutions for Hamilton-Jacobi equations then each u£ will have a local minimum at a nearby point x£, as shown in figure 8.3 (right). But this is impossible, because |«x(^e)| - 1 = “I / > °- In the following sections we shall introduce the definition of viscosity solution and see how it fulfils the above requirements. 8.2 One-sided differentials Let u : J? »—> ]R be a scalar function, defined on an open set 12 C IRn. The set of super-differentials of и at a point x is defined as D+«(x) = LelR"; limsup ~ "W ~ ~ < o) . I y->x \y ~ J In other words, a vector p e IR" is a super-differential iff the plane у i—► u(x) +p-(y — x) is tangent from above to the graph of и at the point x (fig. 8.4 (left)). Similarly, the set of sub-differentials of и at a point x is defined as D~u(x) = pE IR"; lin[in{Ufa)-«W-P-(»-») У-+Х \y — a? I so that a vector p 6 IR" is a sub-differential iff the plane у нч- u(x) + p- (y — x) is tangent from below to the graph of и at the point x (fig. 8.4 (right)). Fig. 8.4. Super and sub-differentials. Example 8.2. Consider the function (fig. 8.5) u(x) = if if if x < 0, xe [0,1], X > I. In this case we have = 0, D tt(O) = [0, oo[,
8.2 One-sided differentials 171 Fig. 8.5. Example of super and sub-differentials. D+u(x) = D u(x') = {1/2\/t} rr G]0,1[, D+u(l) = [0, 1/2], p-tz(l) =0. If tp e C1, its differential at a point x is written as V<p(x). The following characterization of super- and sub-differentials is very useful. Lemma 8.2.1. Let и e C(J?). Then (i) p e D+u(x) if and only if there exists a function tp 6 C1(f?) such that Vcp(z) = p and и — (p has a local maximum at x. (ii) p € D~u(x) if and only if there exists a function ip e such that \7(p(x) = p and и — <p has a local minimum at x. By adding a constant, it is not restrictive to assume that <p(x) = u(x). In this case, we are saying that p E D+u(x) iff there exists a smooth function ip > и with V<p(rr) = p, <p(x) = u(x). In other words, the graph of <p touches the graph of и from above at the point x (fig. 8.6 (left)). A similar property holds for subdifferentials: p G D~u(x) iff there exists a smooth function < ?/, with V<p(x) = p, whose graph touches from below the graph of и at the point x. (fig. 8.6 (right)). Fig. 8.6. Characterization of super and sub-differentials. Proof of Lemma 8.2.1. Assume that p E D+u(x). Then we can find 5 > 0 and a continuous, increasing function ст : [0, оо[ь-> IR, with cr(0) = 0, such that u(y) < u(x) + p (y - x) + a(\y - ar|)
172 8 Viscosity solutions for Hamilton-Jacobi equations for \y — ж| < 6. Define p(r) = f ff(t) dt Jo and observe that p(0) = pz(0) — 0, p(2r) > a(r)r. By the above properties, the function 4>(у) = «(or) + p • (y - x) + p(2|y - x|) is in and satisfies <^(x) = tt(rr), = p. Moreover, for \y — <t| < 6 we have w(y) - V’(y) < - ж|) |j/ - x| - p(2|-y - x|) < 0. Hence, the difference и — ip attains a local maximum at the point x. To prove the opposite implication, assume that D<p(x) — p and и — <p has a local maximum at x. Then Ito Supа(э)-„(х)-р.(,-х) s Um y(jO-y(»)-p-(!>-x) _ 0 y^x y^x |г/-ж| This completes the proof of (i). The proof of (ii) is entirely similar. □ Remark 8.2. By possibly replacing the function ip with <p(y) = <p(y)±|?/—x|2, it is clear that in the above lemma we can require that и — ip attains a strict local maximum or local minimum at the point x. This is particularly important in view of the following stability result. Lemma 8.2.2. Let и : L? > IR. be continuous. Assume that, for some ф € C1, the function и — ф has a strict local minimum (a strict local maximum) at a point x € fl. If um —> и uniformly, then there exists a sequence of points xm —► x with um(xTn) —> u(x) and such that um — ф has a local minimum (a local maximum) at xm. Proof. Assume that и — ф has a strict local minimum at x. For every p > 0 sufficiently small, there exists ep > 0 such that u(y) — ф(у) > u(x) - 0(x) + ep whenever \y - a;| = p . By the uniform convergence um —> u, for all m > Np sufficiently large one has um(y) - u(y) < ep/^ for \y - x| < p. Hence
8.2 One-sided differentials 173 «m(У) - 0(У) > Um(x) - 0(x) + y 1У - *1 =P, This shows that ит—ф has a local minimum at some point with |жт— x\ < p. Letting p,ep —> 0, we construct the desired sequence {а?т}. This situation is illustrated in fig. 8.7 (left). On the other hand, if x is a point of non-strict local minimum for и — 0, the slightly perturbed function um — ф may not have any local minimum xm close to a?, see fig. 8.7 (right). Fig. 8.7. Convergence of strict local minima. Some simple properties of super- and sub-differentials are collected in the next lemma. Lemma 8.2.3. Let и G C(J2). Then (i) If и is differentiable at x, then D+u(x) = D u(x) = {Vu(a?)} . (8-7) (ii) If the sets D+u(x) and D~u(x) are both non-empty, then и is differen- tiable at x, hence (8.7) holds. (iii) The sets of points where a one-sided differential exists: P+ = {zGJ2; P+u(x)^0}, Q- = {xeQ; D~u(x) ± 0} are both non-empty. Indeed, they are dense in 12. Proof. Concerning (i), assume и is differentiable at x. Trivially, Vu(x) G D±u(x). On the other hand, if p G C1(I2) is such that и — p has a local maximum at x, then V<p(rr) = Vu(e). Hence D+u(x) cannot contain any vector other than Vu(rc). To prove (ii), assume that the sets D+u(x) and D~u(x) are both non- empty. Then there we can find 6 > 0 and <^i. p2 C C^J?) such that (fig. 8.8 (left)) <£l(z) = u(x) = p2(x\ ^1(Z/) < tz(i/) < p2(y) \y - z| < S.
174 8 Viscosity solutions for Hamilton-Jacobi equations By a standard comparison argument, this implies that и is differentiable at x and Vu(x) = V^i(x) = V<p2(#)- To prove (iii), consider any open ball B(xo,p) C j? and define the smooth function (fig. 8.8 (right)) , . . 1 PW = ~2------1-------12 P2 - F - ZO|2 Since —> +oo as |rr—a?o | P, the the continuous function u—p attains a local maximum at some interior point у 6 B(xo.p). By Lemma 8.2.1, the super- differential of и at у is non-empty. Indeed, \7<р(у) E D+u(y). The previous argument shows that, for every Xq E ii and p > 0, the set has non-empty intersection with the ball Z?(xo,p). Therefore is dense in P. The case of sub-differentials is entirely similar. Fig. 8.8. Left: a function и having an super- and a sub-differential at a given point x is differentiable. Right: pushing down the graph of until it touches the graph of u, one finds a point у where D+u is non-empty. 8.3 Viscosity solutions In the following we consider the first order partial differential equation F(x, u(x), Уф)) = 0 (8.8) defined on an open set Г2 G IRn. Here F : £? x IR x IRn IR is a continuous (nonlinear) function. A function и E C(<2) is a viscosity subsolution of (8.8) if F(.r, u(x),p) < 0 for every x E 12, p E D+u(x). Similarly, и E C(L2) is a viscosity supersolution of (8.8) if
8.3 Viscosity solutions 175 F(j:, u(x),p) > 0 for every x e /2, p 6 D u(x). We say that и is a viscosity solution of (8.8) if it is both a supersolution and a subsolution in the viscosity sense. Similar definitions also apply to evolution equations of the form ut + H(t, x, u, Vu) = 0, (8.9) where Vu denotes the gradient of и w.r.t. x. Recalling Lemma 8.2.1, we can reformulate these definitions in an equivalent form: A function и G С(Г2) is a viscosity subsolution of (8.9) if, for every C] function <p = <p(t,x) such that и — <p has a local maximum at (£,x), there holds <Pt(t, x) + H(t, x, u, V<p) < 0. Similarly, и G C(f2) is a viscosity supersolution of (8.9) if, for every C1 function ip = <p(t,x) such that и — p has a local minimum at (t,x), there holds <pt(t,x) T H(t,x,u, V92) > 0. Remark 8.3. In the definition of subsohition, we are imposing conditions on и only at points x where the super-differential is non-empty. Even if и is merely continuous and nowhere differentiable, there are infinitely many of these points. Indeed, by Lemma 8.2.3, the set of points x where D+u(x) 0 is dense on <2. Similarly, for supersolutions we impose conditions only at points where D~u(x] 0. Remark 8.4 If и is a C1 function that satisfies (8.8) at every x 6 12, then и is also a solution in the viscosity sense. Viceversa, if и is a viscosity solution, then the equality (8.8) must hold at every point x where и is differentiable. In particular, if и is Lipschitz continuous, then by Rademacher’s theorem it is a.e. differentiable. Hence (8.8) holds a.e. in Г2. Example 8.3. Set F(x,t4,nx) = 1 — |ux|. Then the function u(x) = |#| is a viscosity solution of 1 —|nx|=0 (8.10) defined on the whole real line. Indeed, и is differentiable and satisfies the equation (8.10) at all points x 0. Moreover, we have P+w(0) = 0, F"u(0) = [-1, 1]. To show that и is a subsolution, there is nothing else to check. To show that и is a supersolution, take any p G [—1, 1]. Then 1 — |p| > 0, as required. It is interesting to observe that the same function u(x) = |ж| is NOT a viscosity solution of the equation |tzx|-l = 0. (8.11) Indeed, at x = 0, taking p = 0 e we find |0| -1 < 0. In conclusion, the function u(x) = |ж| is a viscosity subsolution of (8.11), but not a supersolution.
176 8 Viscosity solutions for Hainilton-Jacobi equations 8.4 Stability properties For nonlinear P.D.E’s, the set of solutions may not be closed w.r.t. the topol- ogy of uniform convergence. In general, if un —> и uniformly on a domain Г2, to conclude that и is itself a solution of the P.D.E. one should know, in addition, that all the derivatives Daun that appear in the equation converge to the corresponding derivatives of u. This may not be the case in general. Example 8.4. A sequence of solutions to the equation |ux| -1=0, u(0) = ?z(l) = 0 (8.12) is provided by the saw-tooth functions (fig. 8.9) Clearly um —> 0 uniformly on [0,1], but the zero function is not a solution of (8.12). In this case, the convergence of the functions un is not accompanied by the convergence of their derivatives. Fig. 8.9. A sequence of saw-tooth functions. The next lemma shows that the uniform limit of viscosity solutions is itself a viscosity solution. Quite remarkably, nothing at all is assumed here about the convergence of derivatives. Lemma 8.4.1. Consider a sequence of continuous functions um, which pro- vide viscosity sub-solutions (super-solutions) to ^Wrn) — 0 X E • Asm —> oo, assume that Fm —> F uniformly on compact subsets of J?xIRxIRn and um —► и in C(f2). Then и is a subsolution (a supersolution) of (8.8). Proof. To prove that и is a subsolution, let ф 6 Cl be such that и — ф has a strict local maximum at a point x. We need to show that F(x,</>(jr), V0(a:)) < 0. (8.14) By Lemma 8.2.2, there exists a sequence —► x such that um — ф has a local maximum at xw, and um(a?m) —♦ u(x) as m —► oo. Since um is a subsolution,
8.4 Stability properties 177 (8.15) Taking the limit in (8.15) as m —> oo, we obtain (8.14). The above result should be compared with Example 8.4. Clearly, the func- tions un in (8.13) are not viscosity solutions. The definition of viscosity solution is naturally motivated by the properties of vanishing viscosity limits. Theorem 8.4.2. Let u£ be a family of smooth solutions to the viscous equa- tion F(x, ue(x\ = e Au£ . (8.16) Assume that, as e —> 0+, we have the convergence u£ —> и uniformly on an open set ft C HU1. Then и is a viscosity solution of (8.8). Proof. Fix x e ft and assume p E D+u(x). To prove that и is a subsolution we need to show that F(x, u(x), p) < 0. 1. By Lemma 8.2.1 and Remark 8.2, there exists <p E C1 with V</?(x) = p, such that и — <p has a strict local maximum at x. For any S > 0 we can then find 0 < p < 8 and a function ф E C2 such that |V^)-V^)|<<5 if \y-x\<p, (8.17) <<5 (8.18) and such that each function u£ — ф has a local maximum inside the ball В(х; /?), for e > 0 small enough. 2. Let x£ be the location of this local maximum of u£ - ф. Since u£ is smooth, this implies \7ф(х£) = Vu(r£), Ди(х£) < Аф(х£), hence from (8.16) it follows F(x,u£(xe), Wfe)) < e Аф(х£). (8.19) 3. Extract a convergent subsequence x£ —> x. Clearly |i—a?| < p. Since ф E C2. we can pass to the limit in (8.19) and conclude F(x, u(x), V^(i)) < 0 (8.20) By (8.17)-(8.18) we have |V^(i) — p\ < |VV>(£) — Vy?(^)| 4- |V<p(ai) - V^(x)| < 8 + 8. Since 8 > 0 can be taken arbitrarily small, (8.20) and the continuity of F imply F(x, u(x),p) < 0, showing that и is a subsolution. The fact that и is a supersolution is proved in an entirely similar way.
178 8 Viscosity solutions for Hamilton-Jacobi equations 8.5 Comparison theorems A remarkable feature of the notion of viscosity solutions is that on one hand it requires a minimum amount of regularity (just continuity), and on the other hand it is stringent enough to yield general comparison and uniqueness theorems. The uniqueness proofs are based on a technique of doubling of variables, which reminds of Kruzhkov’s uniqueness theorem for conservation laws [60]. We now illustrate this basic technique in a simple setting. Theorem 8.5.1. (Comparison). Let 12 C IR71 be a bounded open set. Let U\,U2 € C(12) be, respectively, viscosity sub- and supersolutions of и + II(x, Vu) = 0 x € 12. Assume that Ui(rr) < U2(x) for all x 6 dS2. Moreover, assume that H : 12 x IRn i—> IR is uniformly continuous in the x-variable: \H(x,p) - H(y,p)\ <w(|x-j/|(H-|p|)), (8.21) for some continuous and non-decreasing function uj : [0, оо[ь-► [0, oo[ with cu(0) = 0. Then ui(x) < U2 (x) for all ж e 12. (8.22) Proof. To appreciate the main idea of the proof, consider first the case where ui,U2 are smooth. If the conclusion (8.22) fails, then the difference Ui — U2 attains a positive maximum at a point 6 12. This implies p = Vui(rro) = Vu2(^o)- By definition of sub- and supersolution, we now have ui(x0) + H(x0,p) < o, . . U2(x0) + H(x0,p) >0. 1 (ii) Subtracting the second from the first inequality in (8.23) we conclude щ (a?o) ~ ^2(^0) < 0, reaching a contradiction. Next, consider the non-smooth case. We can repeat the above argument and reach again a contradiction provided that we can find a point xq such that (fig. 8.10 (left)) (i) ui(rro) > ^2(^0), (ii) some vector p lies at the same time in the upper differential D+u1(xq) and in the lower differential D~U2(xo\ A natural candidate for tq is a point where Ui —U2 attains a global maximum. Unfortunately, at such point one of the sets D+u\(xq) or D~U2(xq) may be empty, and the argument breaks down (fig. 8.10 (right)). To proceed further, the key observation is that we don’t need to compare values of щ and U2 at exactly the same point. Indeed, to reach a contradiction, it suffices to find nearby points xe and y£ such that (see fig. 8.11)
8.5 Comparison theorems 179 Fig. 8.10. Comparison of viscosity solutions. (f) Ui(z£) > U2(t/e), (ii’) some vector p lies at the same time in the upper differential D+ui(x£) and in the lower differential D~U2(yeY Fig. 8.11. Geometric motivation for the proof of Theorem 8.5.1. Can we always find such points? It is here that the variable-doubling tech- nique comes in. The trick is to look at the function of two variables &e(x,y) = Ui(x) - U2(y) - (8.24) This clearly admits a global maximum over the compact set ft x J?. If щ > U2 at some point xo, this maximum will be strictly positive. Moreover, taking e > 0 sufficiently small, the boundary conditions imply that the maximum is attained at some interior point (xe,ye) 6 ft x Г2. Notice that the points x£, y£ must be close to each other, otherwise the penalization term in (8.24) will be very large and negative. We now observe that the function of a single variable
180 8 Viscosity solutions for Hamilton-Jacobi equations X b-> Ui(x) - ( и2(?/г) + ——— } = Ul(rr) - </?i (x) (8.25) attains its maximum at the point x£. Hence by Lemma 8.2.1 ——— = V<^i(xe) e £)+uj(xe). £ Moreover, the function of a single variable / I £ у | 2 \ У ” U2(y) - I tii(Xe) - ——-------- = U2(l/) - ^(у) (8.26) \ / attains its minimum at the point y£. Hence e D~U2(ye). We have thus discovered two points x£, yE and a vector p = (xe — у£)/e which satisfy the conditions (i’)-(ii’). We now work out the details of the proof, in several steps. 1. If the conclusion fails, then there exists Xq € 42 such that ui(x0) _ «2(^0) = max {tzi(x) - u2(x)} = 6 > 0. (8.27) For £ > 0, call (x£,y£) a point where the function Ф£ in (8.24) attains its global maximum on the compact set 42 x 42. By (8.27) one has &e(x£,yE) > 6 > 0. (8.28) 2. Call M an upper bound for all values |ui(x)|, |- as x e 42. Then la? — vl2 Ф£(х,т/)<2М-Ц-^, Z£ Фе(х,у) <0 if \x — y\2>M£. Hence (8.28) implies \x£ - y£\ < \/M£ . (8.29) 3. By the uniform continuity of the functions u2 on the compact set 42, for £f > 0 sufficiently small we have |w2(x) - li2(y)| whenever |x — y\ < VM£f. (8.30) We now show that, choosing £ < £r, the points xe, y£ cannot lie on the bound- ary of 42. For example, if x£ e dQ. then by (8.29) and (8.30)
8.5 Comparison theorems 181 Фг(хе,у£) < (ui(a:e)-u2(xe)) + |игке) -«г(?/г)| - ~ < 0 + 5/2 + 0, against (8.28). 4. Having shown that xe^ye are interior points, we consider the functions of one single variable <pi,(p2 defined at (8.25)-(8.26). Since xe provides a local maximum for щ —(pi and y€ provides a local minimum for U2 — <£2. we conclude that p£ = ~~ 6 О В~Ы2(Уе)- From the definition of viscosity sub- and supersolution we now obtain UiM + < 0. u2(&) + /%,p£) >0. 5. Observing that Ui(a:£)-U2(Ze) <Фг{х£,у£) < щ(хе)-и2(хе) + \и2(х£)-и2(Уе)\ (8.31) ke ~ tfel2 2г by (8.27) we see that |w2(^e) - U2(ye)| ke -&|2 2e > 0. Hence, by the uniform continuity of U2, 4^^° 2s (8.32) as s —> 0. 6. Recalling (8.28) and subtracting the second from the first inequality in (8.31) we obtain 6 < Фе(х£,Уе) < tZi(2?£) - U2(ye) < \H(xe,p) - H(ye,p)\ < w((|are - ye\ (1 + ke - ye|e-1)). (8.33) This yields a contradiction, Indeed, by (8.21) and (8.32) the right hand side of (8.33) becomes arbitrarily small as s —► 0. An easy consequence of the above result is the following uniqueness result for the boundary value problem и + H(x, Vu) = 0 x e P, (8.34) U = x e <ЭР. (8.35)
182 8 Viscosity solutions for Hamilton-Jacobi equations Corollary 8.5.2. (Uniqueness). Let Pl C IRn be a bounded open set. Let the Hamiltonian function H satisfy the equicontinuity assumption (8.21). Then the boundary value problem (8.34)-(8.35) admits at most one viscosity solu- tion. Proof. Let be viscosity solutions. Since iq is a subsolution and U2 is a supersolution, and щ = U2 on <ЭР, by Theorem 1 we conclude ui < U2 on 12. Reversing the roles of iq and /q* we deduce U2 < iq, completing the proof. By similar techniques, comparison and uniqueness results can be proved also for Hamilton-Jacobi equations of evolutionary type. Consider the Cauchy problem utTH(Lx,Vu) =0 (Lj:) e]0,T[xIRn, (8.36) u(0, x) = й(х) хеПп. (8.37) Here and in the sequel, it is understood that Vu = (uXi,..., uXn) always refers to the gradient of и w.r.t. the space variables. Theorem 8.5.3. (Comparison). Let the function H : [0, T] x IR” x IR” satisfy the Lipschitz continuity assumptions \H(t,x,p) - H(s,y,p)\ < C(|i - s| + |x - y|) (1 + |p|), (8.38) \H(t,x,p)-H(t,x,q)\ < C\p — q\. (8.39) Let u,v be bounded, uniformly continuous sub- and super-solutions of (8.36) respectively. Ifu(f),x) < v(0, x) for all x e lRn, then u(t,x) < v(t,x) for all (t,x) e [0, T] x IR”. (8.40) Toward this result, as a preliminary we prove Lemma 8.5.4. Let и be a continuous function on [0, T] x IR”, which provides a subsolution of (8.36) for t e]0,T[. If ф G Cl is such that и — ф attains a local maximum at a point (T, Xq), then <j>t(T,x0) + H(T,x0,V</>(T,x0)) <0. (8.41) Proof. We can assume that (T, x'o) is a point of strict local maximum for и — ф. For each e > 0 consider the function Each function и — ф£ will then have a local maximum at a point (t£,x£), with t£<T, (te,x£) -> (T,x0) as e -> 0 T . Since и is a subsolution, one has 0e.t(te,xe) + H(^,a:e.V0e(te,a:e)) < • (8-42) Letting e —> 0T, from (8.42) we obtain (8.41).
8.5 Comparison theorems 183 Proof of Theorem 8.5.3. 1. If (8.40) fails, then we can find A > 0 such that sup < u(t, x) — v(t, a?) — 2AO = a > 0. t,X t J (8.43) Assume that the supremum in (8.43) is actually attained at a point (to?^o), possibly with to = T. If both и and и are differentiable at such point, we easily obtain a contradiction, because uf(fo,^o) + Vu) < 0, vt(to^o) + Vv) >0, Vu(to,^o) = Vr(Zo^o), ut(to,xo) - vt(t0, xq) - 2A > 0. 2. To extend the above argument to the general case, we face two technical difficulties. First, the function in (8.43) may not attain its global maximum over the unbounded set [0, T] x Moreover, at this point of maximum the functions u, v may not be differentiable. These problems are overcome by inserting a penalization term, and doubling the variables. As in the proof of Theorem 8.5.1, we introduce the function Ф£(1,х,з,у) = u(t,x)-v(s,y)-A(t+s)-e(|a;|24-|j/|2)-^(|t-s|2 + |2:-y|2) . Thanks to the penalization terms, the function Ф£ clearly admits a global max- imum at a point (t£, x£, y£) € (]0, T] x IRn) . Choosing e > 0 sufficiently small, one has <P£(t£,x£,s£.y£) > тахФ£^,х^,х) > a/2 . t,x 3. We now observe that the function (t,x) u(t, x) — [«(se,tfe) + A(t + .sj +c(|.r|2 + Ы2) + Fqt _ <J£|2 + |x - = u(t, x) — ф(1,х) takes a maximum at the point (t£^x£). Since и is a subsolution and ф is smooth, this implies A + 2(*e /е) + Я (t£, xe, 2(*£ 2 + 2£aA < 0. (8.44) Notice that, in the case where t£ = T, (8.44) follows from Lemma 8.5.4. Similarly, the function
184 8 Viscosity solutions for Hamilton-Jacobi equations (s,y) v(s,y) - [u(te,xe) - A(ie + s) - e(|®e|2 + |a|2) - ^j(l*e - «I2 + l®e - 3/|2)] = v(s,y) -il>(s,y) takes a maximum at the point (t£,x£). Since v is a supersolution and ф is smooth, this implies . 2fa-se) О + н s£, y£, Уе) n A > n -----2-------2ey£ 1 > 0. £z-----------J (8.45) 4. Subtracting (8.45) from (8.44) and using (8.38)-(8.39) we obtain 2A < H (se, ye, - 2ey£') - H (te, xe, + 2sxe) < Co(|l-£| + lifel) + C(\t£ - se| + |are - ad) (1 + + c(|zd + |ad)) (8.46) To reach a contradiction we need to show that the right hand side of (8.46) approaches zero as e —> 0. 5. Since u, v are globally bounded, the penalization terms must satisfy uniform bounds, independent of e. Hence kel, lad < -^= • |ie - sd, !•'=- - ad < c'e. (8.47) for some constant C". This implies е(|х£| + |Уе|) <2CZ4/i. (8.48) To obtain a sharper estimate, we now observe that Ф£ (££, ;r£, se, ?/£) > Ф£ (Ze, же, t£, )T£), hence u{t£.xe) - v(se,y£) - X(t£ + se) - E(|a?e|2 + lad2) “^2 (I*® “ S®|2 + Iх® “ ^®|2) > u(t£,x£) — v(t£1xe) — 2Xt£ — 2e|:re|2, 72 (1^ -»d2 + l^- ad2) < v(ie,a?e)-v(se,ae) + A(te-Se)+E(|a:e|2 - ladT (8.49) By the uniform continuity of v, the right hand side of (8.49) tends to zero as e —> 0, therefore
8.6 Dynamic programming (revisited) 185 Re ~ «el2 + l^e ~ &|2 £2 as e —> 0. (8.50) By (8.47), (8.48) and (8.50), the right hand side of (8.46) also approaches zero, This yields the desired contradiction. □ Corollary 8.5.5. (Uniqueness). Let the function H satisfy the assumptions (8.38)-(8.39). Then the Cauchy problem (8.36)-(8.37) admits at most one bounded, uniformly continuous viscosity solution и: [0. T] x IR" > IR. 8.6 Dynamic programming (revisited) Consider again a control system of the form x(t) = f(x(t), e U. (8.51) We now assume that the set U C IRW of admissible control values is compact, while f : IRn x U Hl" is a continuous function such that for all x. у E IR" and и E U |f(x,u)|<C, \f(x,u)~ f(y,u)\<C\x-y\ (8.52) for some constant C. Given an initial data ar(s) = T/eIR", (8.53) under the assumptions (8.52), for every choice of the measurable control func- tion zz(-) € U the Cauchy problem (8.51)-(8.53) has a unique solution, which we denote as t i—> x(t; s, y, u) or sometimes simply as t x(t). We seek an admissible control function u* : [s, T] »—> U. which minimizes the sum of a running and a terminal cost J(s,y,u) = У h(x(t), u(t)) dt + д(х(Т)У (8.54) Here it is understood that x(t) = z(t; s, y, u), while h:IRn xUh+IR, g : IR" i—> IR are continuous functions. We shall assume that the functions h.g satisfy the bounds |/i(z,u)| <G \g(x)\ <C, (8.55) \h(x, u) - h(3/,u)| < C\x-y\, \g(x) -^(t/)| < (8.56) for all x, у E IR", и E U. As in the previous sections, we call
186 8 Viscosity solutions for Hamilton-Jacobi equations Z7 = : IR i—> IRrn measurable, u(t) G U for a.e. 1j* the family of admissible control functions. According to the method of dy- namic programming, an optimal control problem can be studied by looking at the value function: V(s^y)= inf J(s,y,u). (8.57) We consider here a whole family of optimal control problem, all with the same dynamics (8.51) and cost functional (8.54). The main interest is on how the minimum cost varies, as a function of the initial conditions (8.53). Indeed, we will show that the value function V can be characterized as the unique viscosity solution to a Hamilton-Jacobi equation. Toward this goal, a basic step is provided by Bellman’s principle of dynamic programming. Fig. 8.12. Dynamic Programming Principle. Theorem 8.6.1. (Dynamic Programming Principle). For every r e [s,T] and у G IR", one has V(s, t/) = inf < / s, y, u), u(t)) dt T V(т, <г(т; s, ?/,)) (8.58) In other words (fig. 8.12), the optimization problem on the time interval [.$•, T] can be split into two separate problems: • As a first step, we solve the optimization problem on the sub-interval [т, T], with running cost h and terminal cost g. In this way, we determine the value function V(t, •), at time r. • As a second step, we solve the optimization problem on the sub-interval [s, r], with running cost h and terminal cost V(r, •), determined by the first step. At the initial time s. by (8.58) we are saying that the value function V(s, •) obtained in step 2 is the same as the value function corresponding to the global optimization problem over the whole interval [s,Т].
8.6 Dynamic programming (revisited) 187 Proof. Call Jr the right hand side of (8.58). 1. To prove that JT < V(s, y), fix s > 0 and choose a control и : [s,T] i—► U such that J(s, у, и) < V(s, y) + e. Observing that V(t, x(t',s, y, uf) < h(x(t; s, y, u), u(ty) dt + g(x(T; s, y, u)), we conclude JT < h(x(t; s, y, u), u(t)) dt + V (т, ж(т; s, y, u)) <J(s,y,u) < V(s,?/) + s. Since £ > 0 is arbitrary, this first inequality is proved. 2. To prove that V(s,?/) < JT, fix г > 0. Then there exists a control u' : [s, t] U such that У h(x(t, s, y, u'), u(tf) dt + V(т, ж(т; s, ?/, u7)) < JT + £. (8.59) Moreover, there exists a control и" : [т, T] i—► U such that J(t, .t(t; s, ?/, u7), u")<V(t, х(т;з,у,и')) Те. (8.60) One can now define a new control и : [s,T] »—► A as the concatenation of u\u”‘. if £e[s,r], u(t) “[«"(t) if fe]r,T]. By (8.59) and (8.60) it is now easy to check that V(s, y) < J(s, y, u) < JT + 2e. Since e > 0 can be arbitrarily small, this second inequality is also proved. The next lemma establishes the Lipschitz continuity of the value function. This property will be later used, to characterize the value function as the unique viscosity solution to the corresponding Hamilton-Jacobi equation. Lemma 8.6.2. Let the functions f,g,h satisfy the assumptions (8.52), (8.55) and (8.56). Then the value function V : [0, T] x IRn i—> R in (8.57) is bounded and Lipschitz continuous. Namely, there exists a constant C' such that |V(M|<C', (8.61) | V(s,y) - V(s',у')| < C'(|s - s'| + |y - y'\\ (8.62)
188 8 Viscosity solutions for Hamilton-Jacobi equations Proof. 1. Let s > 0 and у E IR” be given. By (8.55), for every control function и : [s, T] U, the corresponding cost satisfies |J(.s.7/. w)| <C(T-.S) + C. Hence (8.61) holds with C = C(T + 1). 2. To prove (8.62), let s, у be given. Fix any s > 0 and choose a near-optimal control function и : [s, T] h-> U such that J(s, и) < V(s, у) T £. (8.63) If we use the same control function и in connection with another initial con- dition ($.?/'), by (8.52) the corresponding trajectories satisfy |x(t; .s, y, u) — x(t, s,u)| < ес^~^\у — y'\ . Hence, from (8.54) and (8.54) we deduce J(s, y', «) < ./(s. y, u) + ^r(7eC(t-s)|y' - y\ds + Cec(r~s}\y - y'\. By (8.63), this implies V(s,y') < V(s.y) + e(C + l)eCT\y — y'\. (8.64) Since e > 0 was arbitrary, interchanging the role of y' we conclude that the value function V is Lipschitz continuous w.r.t. y: \V(s,y)- V(.M/)| <C0|y-y'|. (8.65) with Co = (C + 1) eCT. 3. To prove the Lipschitz continuity of V w.r.t. the time variable s, let у € IR” and 0 < s < s' < T be given. Choose a near-optimal control function и : [s, T] и-> U for which (8.63) holds, and call t x(P) = x(t\s,y,u) the corresponding trajectory. We now have V(s', x(.s')) < J(.s, у, и) - Г h(t, x(t)) dt < V(s, y) + e - C(.s' - «). J s On the other hand, by the dynamic programming principle, V(.s, у) < Г h(t, x(t)) dt + V(s', x(.s')) < V(s',x(.s')) - C(.s' - .s). The uniform bound on f stated in (8.52) implies |r(S')-x(.s)H|T(.s')-y|<C'|.s'-S|.
8.7 The Hamilton-Jacobi-Bellman equation 189 Using the above inequalities we conclude IV(s,y) - V(S',y)| < |V(s,y) - У(з',ф'))1 + |V(.<;r(.S')) - V(e',y)| <С0СУ-з|+С0СК-^|. This proves the Lipschitz continuity of the value function w.r.t. time, com- pleting the proof. 8.7 The Hamilton-Jacobi-Bellman equation The main goal of this section is to characterize the value function as the unique solution of a first order P.D.E., in the viscosity sense. In turn, this will provide a sufficient condition for the global optimality of a control function u(-). As in the previous section, we assume here that the set U is compact and that the functions fig.h satisfy the bounds (8.52), (8.55) and (8.56). Theorem 8.7.1. In connection with the control system (8.51), consider the value function V = V(s,y) defined by (8.57) and (8.5f). ThenV is the unique viscosity solution of the Hamilton-Jacobi-Bellman equation + =0 (t,x) €]0,Т[х1Б'“, (8.66) with terminal condition V(T.x) = g(x) xE Rn, (8.67) and Hamiltonian function H(x, p) = min { f(x, w) • p + hfx, o>)}. (8.68) Proof. By Lemma 8.6.2, the value function is bounded and uniformly Lipschitz continuous on [0, T] x IR". The terminal condition (8.67) is obvious. To show that V is a viscosity solution, let € C1 (]0, T[ xlR"). Two separate statements need to be proved: (Pl) If V - ip attains a local maximum at a point (to,^o) e]0, T[ xIRn, then <a(*o,Zo) + nun {/(я70,cu) • V</?(to,^o) + h(x0,w)} > 0. (8.69) (P2) If V - tp attains a local minimum at a point (to,^o) €]0, T[ xlRn, then + niin {/(ж0,й?) • V<p(io,^o) + h(xG,w) < 0. (8.70)
190 8 Viscosity solutions for Hamilton-Jacobi equations 1. To prove (Pl), we can assume that V(to-^o) = V(£,t) < <p(t,x) for all t,x. If (8.69) does not hold, then there exists cu € U and 0 > 0 such that 9?t(fo,#o) + №o,^) * W(to,£o) + h(xQ,(jj} < -0. (8.71) We shall derive a contradiction by showing that this control value ш is “too good to be true”. Namely, by choosing a control function iz(-) with u(t) = и for t e [to- t0 + 5] and such that и is nearly optimal on the remaining interval [to + <5- T], we obtain a total cost J(to,#o,4) strictly smaller than V(to,^o)- Indeed, by continuity (8.71) implies (/?t(t, x) + • V^(L x) < - 0. (8.72) whenever |t — t0| < <5, |t- T01 <C8, (8.73) for some 8 > 0 small enough and C the constant in (8.52). Let T(t) = t(Z; to,To,cu) be the solution of i(t) =/(x(t),u>), rc(f0) = zo, i.e. the trajectory corresponding to the constant control u(t) = cu. We then have V(to + d. x(t0 + <5)) - V(to,x0) < ¥?(to + 5, z(t0 + <*)) - <p(to, x0) fto+S d = Jt dt /•to+<5 x x = I x(t)) + f(x(t),w) -Vip(t, £(<))} dt /•<o+<5 < - / /i(T(t),u)dt-^, (8.74) Jto because of (8.72). On the other hand, the Dynamic Programming Principle (8.58) yields /»to+<5 V(t0 + 5, x(t0 + <5)) - V(to,xo) > - h(t, x(t)) dt. (8.75) v/ to Toget her. (8.74) and (8.75) yield a contradiction, hence (Pl) must hold. 2. To prove (P2), we can assume that V(fo^o) = ^o^o). V(t,T) > <p(t,x) for all t,x.
8.7 The Hamilton-Jacobi-Bellman equation 191 If (P2) fails, then there exists 0 > 0 such that Ptfaxo) + /(#o?tu) • V99(^0? ^o) + h(xo,u) > 6 for all uEU. (8.76) In this case, we shall reach a contradiction by showing that no control function u(-) is good enough. Namely, whatever control function u(-) we choose on the initial interval [to? to 4- <5], even if during the remaining time [to + <S? T] our control is optimal, the total cost will still be considerably larger than V(to, Xq). Indeed, by continuity, (8.76) implies <^t(t,x) + • V<£>(t,x) > 0 — h(x.u) for all w e U, (8.77) for all t,x close to to,#o? be. such that (8.73) holds. Choose an arbitrary control function и : [to, to + 5] •—> A, and call t i—> x(t) = x(t,to, xq, u) the corresponding trajectory. We now have V(<o + <5, x(to + <5)) - V(t0, x0) > fp(t0 + 8, x(t0 + <9) - <p(to, x0) fin+6 J rt()+6 = / x(t)) + f(x(t),u{t)) • x(t))dt J to rto+6 > / 0 — h(x(t), u(t)) dt, J to because of (8.77). Therefore, for every control function u(-) we have /•to+<5 V(t0 + &, x(t0 + 6)) + h(x(t),u(t))dt>V(to,xo) + 89. (8.78) J to Taking the infimum of the left hand side of (8.78) over all control functions u, we see that this infimum is still > V(to,^o) + On the other hand, by the Dynamic Programming principle (8.58), this infimum should be exactly V(to, #o)- This contradiction shows that (P2) must hold, completing the proof. One can combine Theorems 8.5.3 and 8.7.1, and obtain sufficient conditions for the optimality of a control function. The usual setting is the following. Consider the problem of minimizing the cost functional (8.54). Assume that, for each initial condition (s, y). we can guess a “candidate” optimal control us'y : [s,T] U. We then call V(s,y) = J{s,y,us'y) (8.79) the corresponding cost. Typically, these control functions us'y are found by applying the Pontryagin Maximum Principle, which provides a necessary con- dition for optimality. On the other hand, consider the true value function
192 8 Viscosity solutions for Hamilton-Jacobi equations V, defined at (8.57) as the infimum of the cost over all admissible control functions u(-) G Ы. By Theorem 8.7.1, this function V provides a viscos- ity solution to the Hamilton-Jacobi equation (8.66) with terminal condition V(T, y) = g(y). If our function V at (8.79) also provides a viscosity solution to the same equations (8.66)-(8.67), then by the uniqueness of the viscosity solution stated in Theorem 8.5.3, we can conclude that V = V. Therefore, all controls us,y are optimal. 8.8 Infinite horizon problems This section is devoted to infinite horizon problems, where trajectories are defined for all times t > 0. Consider again a control system of the form x = f(x, и) я:(0) = у, (8.80) with U C IR"' compact and f : IR " x U IR". For a given measurable control function ?/(•) e U, we denote by t i—► x(t;y, u) the solution to the Cauchy problem for (8.80). We seek an admissible control function u* : [0, oo[«-> U, which minimizes the exponentially discounted cost J(y, u) = f e~ath(x(t), u(t)) dt. (8.81) Jo where a > 0 and t i—* x(t) = x(t; y. u) denotes the solution of the Cauchy problem (8.80) corresponding to the control u(-). To simplify the exposition, we assume that both f and h are uniformly bounded and Lipschitz continuous, namely \f(x.u)\<C, \h(x,u)\<C, (8.82) |/(x, u) - f(y, u)\<L\x- y\, |Л(.т, u) - h(.r, y)\< L\x-y\, (8.83) for some constants C, L. The above assumptions on f imply that the trajectory t и-> ;r(t; y. u) is well defined for all times t > 0. Call U = : IR IR'" measurable, u(t) G U for a.e. the family of admissible control functions. Thanks to assumptions (8.82)- (8.83). the cost J(y,u) is well defined for every admissible control function. We define the value function by: V(?;)= inf J(y,u). (8.84) u( ) ел/ As for the finite horizon case, one can now prove the uniform continuity of the value function.
8.8 Infinite horizon problems 193 Lemma 8.8.1. Let the functions f,h satisfy the assumptions (8.82)-(8.83). Then the value function V in (8.84) bounded and uniformly Holder contin- uous, namely |V(y)| <Co, 1Ш- Vy.y'eIR", (8.85) for some constants Cq.Ci and 0 < 7 < 1. If a > L, then V is actually Lipschitz continuous. Proof 1. Consider any control и : [0, oo[«—► U. The uniform bound (8.82) yields |J(t/,tt)I < I e~at\h(x(t,y,u),u)\dt < I Ce~at dt = — . Jo Jo ° This yields the first inequality in (8.85), with Co = C/a. 2. Consider any two initial values y, y', with \y-y' | < 1. In connection with the same control function и : [0, oo[«—► U, by (8.83) the corresponding trajectories satisfy \x(t,y,u) -x(t,y',u)\ < eLt\y — y'\. By the Lipschitz continuity of the cost function h. for every time T > 0 we thus have the estimate \J(y,u) — J(y',u)\ < /0T e~at\hfr(t,y,u),u) - h(x(t,y', u),u)| dt -j- e~at^\h(x(t,y.u),u)\ + \hfx(t,y',u),u)\^ dt < fy e~Qt • LeLt\y — y'\dt + 2e~aTCo/a . (8.86) If L < a, letting T —► 00 we obtain \J(y, u) - J(y',u)\ < —~—х\у~у'\- (8.87) a — L If L > a, we choose т = -7 ln|j/-y'|,- JLj When L = a, from (8.86) we deduce \J(y, u) - J(y', U)| <-Lln|y~y/| \y - y'\ + Coly - y'l < G \y - y'\V (8.88) Q for 7 < 1 and a suitable constant C\. When L > o, the above choice of T yields |J(y,u) - J(y',u)\ < z^exp(^ln|y-y'||+2exp(f <тЬ\У~ y'\(L~a)/L + 2C0|j/ - y'\“/L ^С^у-уГ- (8.89)
194 8 Viscosity solutions for Hamilton-Jacobi equations 3. Consider again two distinct points у y', with \y — y'\ < 1. Fix any e > 0 and choose a near-optimal control tz(-) such that J(y,u) < V(y) + e If we use this same control in connection with the initial value z(0) = y', we obtain a cost < J(y,u) + Ci|y-y'|7 < V(y) +e + Ci|y-y'|7. Therefore V(y') < J(y’, u) < V(y) + e + G|y - y'R. Since c > 0 can be chosen arbitrarily small, arid the points y, y' can be inter- changed, this proves the Holder continuity of the value function. In the case L < a, by (8.87) we can take 7 = 1, hence the value function V is Lipschitz continuous. Also in this case, the value function V can be characterized as the unique viscosity solution to a Hamilton-Jacobi equation. First we prove a Dynamic Programming Principle. Theorem 8.8.2. (Dynamic Programming Principle - Infinite Hori- zon). For every т > 0 and у € IR7’, one has Vfy) = inf | [ e~ath(x(t:,y,u), u(t)) dt T e~aTV(xfr, y, ?z)) 1 . (8.90) n(-) l./o J Proof. Call JT the right hand side of (8.90). 1. We first prove that JT < Vfy). For every control и € U. we have: J(y, u) — J e~ath(<x(t;yiu)^ u(t))dt T j e~ath(x(t\y,u), u(tf) dt = f e~at'h(x(t',y,u\ uft)) dt + e~ar I e~°lth(<x{T T t-,y,u), ufr + t)) dt. Jo Jo Define the control u : [0, Too) i—> U by uft) = uft T т), then Jfy,u) = f e~ath(x(t-, y,u), u(t)) dt T e~aT J(x(r, y, w), u) > JT. Jo Since и was arbitrary, this establishes our claim. 2. To prove that Vfy) < JT, fix e > 0. Then there exist a control и! e Id such that I e~ath(x(t-,y,u'), u'(t))dt T е~ат V(x(r; у, uf) < JT T г (8.91) Jo
8.8 Infinite horizon problems 195 and a control u" e U such that 7(я(т; ?/, и'), и") < V(я(т; г/, г/)) + 5. (8.92) One can now define a new control й G U as the concatenation of uf up to time т and u" time shifted: 1 u"(t — t) if te]r, +oo). Then, by (8.91) and (8.92), we get: V(y) < = f e~ath(x(t\y,u), й(1)) dt + J e~ath(<x(t]y1u)1 u(t)) dt = I e~ath(x(t;y,u,)y u'(t))dt + e~ar J(x(r; y, u'), u") Jo < JT + e(l + e"QT). Since £ > 0 can be arbitrarily small, this second inequality is proved. We are now ready to show that the value function is solution of a first order P.D.E., in the viscosity sense. Theorem 8.8.3. In connection with the control system (8.51), consider the value function V = V(y) defined by (8.84) and (8.81). Then V is viscosity solution of the Hamilton-Jacobi-Bellman equation - [ - aV + H(x, VV)] =0 x e lRn, (8.93) and Hamiltonian function given by (8.68). Proof. To show that V is a viscosity solution, let e C1 (lRn). Two separate statements need to be proved: (Pl) If V — kp attains a local maximum at a point xq G IRn, then -aip(xo) 4- min {f(xo,ui) • V^(t0) + h(xo,u)} > 0. (8.94) (P2) If V — attains a local minimum at a point tq G IRn, then —Q(/?(x0) + min {f(xo,aj) • V(£(a7o) + /г(жо,^) < 0- (8.95) 1. To prove (Pl), we can assume that V(zq) = (X#o)> V(z) < <^Gr) for all x. Fix (j€U and consider the trajectory x(t) = x(t; 0, Xq,iv\ solution of
196 8 Viscosity solutions for Hamilton-Jacobi equations ±(Z) = /(#(£), о?), я(0) = Xq. Then: 99(^0) - y>(x(t)) < V(z0) - V(x(t)) thus by the Dynamic Programming Principe (Theorem 8.8.2), we can write: ^(^0) - 9?(я(/)) < [ e~a8h(s,x(s)) ds + V(a;(^))(e“Qt - 1). Jo Letting t —> 0, we get: -Dip(x0) • < h(x0,u) - aV(xo). Letting cj vary in U we get the desired conclusion. 2. To prove (P2), we can assume that V(xo) = <^(#0), V(#) ></>(#) for all x. Fix e > 0. By the Dynamic Programming Principe (Theorem 8.8.2), for every t > 0 there exists a control t?(-) and a corresponding trajectory xl such that V(x0) > Г e“QS/i(^(s),u£(s)) ds + e-atV(x\t)) - te. Jo We can thus write ¥?(a?o) - <p(z‘(0) > v(xo) ~ ^(*‘(0) > > [ e"QS/l(j;t(s),u‘(s))(/s + V(xt(Z))(e-at-l)-t£. (8.96) Jo On the other hand: <p(x0) - ^(x‘(t)) = - [ W(s)) • /(®‘(s), u‘(s)) ds. (8.97) Jo Using (8.96) and (8.97), by continuity we can write: — J [v^C^o) ’ /(^o, w*(s)) + ^(^o, W4S))] ds + / (1 — е-а5)/г(а?Отt££(s)) c?5 +V(a/(t))(l - e~at) > -te + o(Z), where the Landau symbol o(t) denotes a quantity such that lim*—о о(0Л — 0- The second integral is also o(t), thus we can write: -t H(x0, V(f?(iro)) + ^(^£(0)(l ~ e-Qt) > — te + o(i). Dividing by — t and passing to the limit, we get the conclusion since e is arbitrary. A comprehensive study of control problems in infinite horizon can be found in [25].
8.8 Infinite horizon problems 197 Problems 8.1. Consider the Cauchy problem: ut + a(t,x)ux = 0, u(0, a?) = uq(x), (8.98) where t > 0, x G IR. a and uq are smooth functions. a) Prove that there exists a smooth solution u(t, x) to (8.98). b) Find the solution to (8.98) for a(t,x) = x and uo(x') = sin(x). 8.2. Let 12 = IR \ {0} and consider the equation: (|uj - 1) (x2 - 1) = 0 Vrr e P, u(0) = 0. (8.99) Find all solutions и to (8.99) such that there exist a sequence sn, En —* 0, and a sequence un, solution to (|uT| — 1) (x2 — 1) = en uxx, converging to и uniformly on J?. 8.3. Find £>±(0) for the following functions: a) u(x) = x sin Q); b) u(x) = x2 sin Qj; c) u(x,y) = |x| - 8.4. Consider the function u(x) = min < |ж — z\ : z e : n 6 IN and compute the sets (defined in Lemma 8.2.3.) 8.5. Write the Dynamic Programming Principle (Theorem 8.6.1) and equation (8.66) for a Linear-Quadratic optimal control problem. 8.6. Use Theorem 7.3.3 to derive sufficient conditions for a function U, solving the HJB equation (8.66) almost everywhere, to be a viscosity solution. 8.7. Consider the minimum time to origin problem for the linearized pendulum with external force: x + x = u, x G IR, |u| < 1. (a) Compute the value function V as in Problem 6.5; (b) Show that V is a viscosity solution of the corresponding HJB equation directly computing D±V at every point. 8.8. Consider Example 6.5. (a) Compute the cost function V corresponding to the constructed syn- thesis;
198 8 Viscosity solutions for Hamilton-Jacobi equations (b) Determine the curves along which V is not differentiable; (c) Show that the synthesis is optimal verifying that V is a viscosity solution to the corresponding HJB equation. 8.9. Consider the infinite horizon optimal control problem (8.51), (8.81). As- sume that x G IR. n(a?, u) = xe + x u, with U = {?/ : |u| < 1}. For a = 2 is the value function Lipschitz contin- uous?
9 Patchy Feedbacks This chapter contains an introduction to the theory of patchy feedbacks and its applications to problems of asymptotic stabilization and of optimal control. We consider a nonlinear control system on IRn x = /(z, u), u(t) € U C Brn , (9.1) assuming that the map f : IRn x U i—> IRn is smooth. We shall focus the attention on two classical problems: (I) Asymptotic Feedback Stabilization: Construct a feedback control и = U(x} such that all trajectories of the resulting O.D.E. i = g(X) = f(x, t7(x)) (9.2) approach the origin as t oo. (II) Optimal Control Problem: Let a terminal cost ф = and a running cost function L = L(x, u) be assigned. For a given initial data я(0) = г/, (9.3) find an optimal control ?/(•) and a terminal time т > 0 which minimize the total cost r Ф = ^(ж(т)) + / L(rr(t), u(tY) dt. (9.4) Jo In an ideal situation, the above problems would be solved by a C1 feedback control x •—> U(x). In this case, for every initial state у e IRn, the Cauchy problem (9.2)-(9.3) has a unique classical solution, continuously depending on the initial data. In various cases, however, smooth feedback controls cannot be used. (I) For the problem of asymptotic stabilization, several examples of control systems are known, where each initial state can be steered toward the origin
200 9 Patchy Feedbacks as t —* oo by an open-loop control t u(t). However, no smooth (or even continuous) feedback control function x »—> U(x) can accomplish the same task, not even locally in a small neighborhood of the origin. See Examples 4.2 and 4.3 of Chapter 4. In all these examples, asymptotic stabilization can be achieved only by a discontinuous feedback. (II) For nonlinear optimization problems, it is well known that the optimal feedback can be discontinuous, with a very complicated structure, wdiile op- timal controls 11—► u(t) can have infinitely many switchings. Moreover, a dis- continuous optimal feedback may not be robust w.r.t. perturbations. In other words, one may find arbitrarily small functions ei,e2 such that the solution to the perturbed equation x = f(x, U(x + «1 (t))) + 62(t) (9-5) achieves a performance far worse than the optimal one. See [71]. It is worth noting that the use of a discontinuous feedback control x i—► U(x) leads to the theoretical problem of how to interpret solutions for the O.D.E. (9.2), when the right hand side is discontinuous w.r.t. the state variable x. Indeed, the standard theory of O.D.E’s does not include any general result in this direction. Different approaches can be followed: • Consider only solutions in a strong sense. We recall that a Caratheodory solution of (9.2)-(9.3) is an absolutely continuous map 11—» x(t) which satisfies (9.3) together with (9.2) at a.e. time t. Equivalently, ®(i) = У + / s(z(s)) ds. Jo (9-6) In general, we expect that these strong solutions will have the asymptotic convergence or optimality properties required by the original control problem. However, if one does not impose any regularity assumption of the feedback C7(x), there is no guarantee that any Caratheodory solution wall exist. The simplest example is if x < 0 , if x > 0 , z(0) = 0, where no solution exists, forward in time. • Define a weaker concept of solution. For example, if g is measurable and bounded, one can consider “Krasovsky solutions” (see for example [50]) of the O.D.E. (9.2). These are, by definition, the trajectories of the differential inclusion x e G(x) = p) сб|д(у); \y - ят| < £ j . e>0
9.1 Patchy vector fields 201 It is not difficult to show that, in this relaxed sense, at least one solution always exists. In some cases, however, too many solutions are obtained. Not all of them may have the desired properties. In the present chapter we outline a recent approach, introduced in [2], based on patchy feedbacks. These are piecewise constant controls, whose discontinuities are sufficiently tame in order to guarantee the existence of Caratheodory solutions forward in time. At the same time, this class is suffi- ciently broad to solve a wide class of stabilization and optimization problems. Section 9.1 contains the basic definitions of patchy vector fields and patchy feedbacks, and a summary of their basic properties. In particular, we prove the existence of forward solutions and the stability of the solution set under small perturbations. In Section 9.2 we show how to solve the asymptotic stabilization problem by means of a patchy feedback. Section 9.3 is concerned with the robustness of patchy feedbacks, showing that they still perform well also in the presence of small external perturbations, or small measurement errors in the state of the system. Finally, in Section 9.4 we construct a patchy feedback which is nearly optimal, for a general class of optimization problems with free terminal time. 9.1 Patchy vector fields We begin with some definitions. Throughout the following, the boundary and the closure of a set 12 are denoted by dS2 and 12, respectively. Definition 9.1 (single patch). By a patch we mean a pair (12, g} where 12 C IRn is an open domain with smooth boundary <912, and g is a Lipschitz continuous vector field defined on a neighborhood of the closure 12, which points strictly inward at each boundary point x E d(2. Calling п(ж) the outer normal at the boundary point x, and denoting the inner product by a dot, we thus require sup n(rr) • g(x) < 0. (9.7) хЕдГ2 Definition 9.2 (Patchy vector field). We say that g : 12 i—» IRn is a patchy vector field on the open domain 12 if there exists a family of patches {(12a, gafi a e Л} such that - A is a totally ordered set of indices, - the open sets 12a form a locally finite covering of 12, - the vector field g can be written in the form = 9a(x) if x € \ [J fig . fl>ot (9-8)
202 9 Patchy Feedbacks We shall occasionally adopt the longer notation (12, g, (12Q, <7а)а6Л) to in- dicate a patchy vector field, specifying both the domain and the single patches. By setting o*(t) = max {a € A ; x E 12a}, (9.9) we can write (9.8) in the equivalent form g(x) = <7а.(ж)(я) for all x E £2. (9.10) Remark 9.1. It is important to observe that the patches (12Q, ga) are not uniquely determined by a patchy vector field (12, g). Indeed, whenever a < /3, by (9.8) the values of gn on the set 12Q A 12^ are irrelevant. Of course, the values of ga for x outside the domain J? don’t matter either. Therefore, if the open sets 12a form a locally finite covering of 12 and if for each a E Л the vector field ga satisfies na(x) • ga(x) < -6 < 0 for all x e n n dfia \ [J , (9.11) /3>a then the vector field g in (9.8) is still a patchy vector field. Indeed, without changing the function 9, one can suitably redefine the values of each ga on the set U/3>« or outside 12, and achieve the strict inequality nQ(x)-ga(x) < < 0 for all x E df2a . Remark 9.2. For convenience, we are always assuming that the single patches 12tt are open, while the vector fields ga are defined on the closure 12Q . In certain cases, it would be natural to choose patches of the form 12x = {# E 42; n • x < c} , 1?2 = {x e 12; n • x > c} , for some unit vector n. In this way, however, the union 12i U12'2 does not cover all of 12, because it does not contain the points where n • x = c. This situation is easily fixed, replacing 121 by a slightly larger open set which contains also these boundary points. The resulting vector field 0(z) = f 51 (*) if if n • X < c, n • X > c, can still be written in patchy form. Remark 9.3. In some situations it is convenient to adopt a more general def- inition of patchy vector field than the one formulated above. Indeed, one can consider patches (12a, g(^ where the boundary of the domain 12Q is only piece- wise smooth. For example, 12a could be the intersection of a ball and finitely
9.1 Patchy vector fields 203 many half-spaces. In this more general case, the inward-pointing condition (9.7) can be reformulated by asking that, at each boundary point x 6 d!2a, the vector ga(x) lies strictly in the the interior of the tangent cone to J?a at the point x. Namely ga(x) e int7bo(.-r), (9.12) where the tangent cone is defined by T^(x) = ^elRn lim inf t|0 d(x + tv, 12Q) 1 t ~ T All the results concerning patchy vector fields, stated in Theorem 9.1.1 below, remain valid with this more general formulation. We also observe that, by slightly modifying the domains 12a, we can always obtain a patchy vector field where all boundaries dQa are smooth. Definition 9.3 (Patchy feedback). Let (12, g, (12a, ра)а6Л) be a patchy vector field. Assume that there exist control values va G U such that, for each a € Д, there holds ^a(^) = f(x, va) for all x G 12Q \ [J I?/?. (9.13) /3>a Then the piecewise constant map U(x) = va if x G 12Q \ U^. (9.14) (3>a is called a patchy feedback control on J?. Recalling (9.9), the patchy feedback control can thus be written on the form t7(j?) — va*^x) • The next theorem collects the basic properties of Caratheodory solutions for the O.D.E. i = g(x) = go.(l)(x), (9.15) where g is a patchy vector field (see figure 9.1). Theorem 9.1.1. (Trajectories of patchy vector fields). Let (12, g, (1?Q, <7а)аел) be a patchy vector field. Then the following holds. (i) If 11—> x(t) is a Caratheodory solution to (9.15) on an open time interval J, then its derivative t x(t) is piecewise and has a finite set of jumps on any compact subset J' C J. The function 11—> a* (#(£)) defined at (9.9) is non-decreasing and left-continuous. Moreover the one-sided limit of x satisfies *lim i(t) = g(x(r)) for every r G J. (9.16)
204 9 Patchy Feedbacks Fig. 9.1. Trajectories of a patchy vector field. (ii) For each xq G 12, the Cauchy problem for (9.15) with initial condition яг(0) = jtq has at least one local forward Caratheodory solution and at most one backward Caratheodory solution. (iii) The set of Caratheodory solutions of (9.15) is closed. More precisely, assume that xy : i—> f2 is a sequence of solutions and, as v oo, there holds ay —> a, by —> b, xy(t)x{t) for all t e]a,b[. (9.17) Then f(-) is itself a Caratheodory solution of (9.15). Proof, (i) Since {f2a : a G Л} is a locally finite covering of 12, the function t h-> a*(t) can take only finitely many values on each compact interval. In particular, its lim-sup and lim-inf are well defined at each time r. To prove that a* is left continuous, fix any r G J. Since f2Q.(r) is open, x(t) G for all t sufficiently close to t. Therefore lim inf o*(t) > Q*(r). (9.18) On the other hand, we now show that a = lim sup a*(t) < а*(т) . (9.19) t-*r- Indeed, since o* takes finitely many values on compact intervals, we can choose 8 > 0 such that o*(Z) < a. t G ]r—6, т] C J, then there exists s G ]r—6, t] with a*(s) = a. The inward-pointing condition (9.7) now implies that x(t) G 12« for every t G [s, г]. Hence o*(r) > a. Together, (9.18) and (9.19) show that the map t и-> o*(i) is left-continuous, non -decreasing and thus piecewise constant on compact sets. The identity (9.16) is now clear, because x = and the map t i—> «*(#(£)) is piecewise constant. Finally, x is piecewise C1. (ii) To prove the local existence of solutions, forward in time, define a = max{a; x$ G f2Q}. Because of the inward pointing condition (9.7), the solution to the Cauchy problem
9.1 Patchy vector fields 205 x = x(0)=xo, remains inside for all t > 0. Moreover, by definition of the index ct, we have x(t) Qp for all (3 > a and t € [0, J]. Hence, this function x(-) provides also a solution of the original equation x = g(x), on the interval [0, <5]. To show backward uniqueness, let a?i(-), Я2С) be any two Caratheodory so- lutions with jq(0) = £2(0). If they do not coincide for all times t < 0, by continuity there will be a minimum time r such that jq(t) = X2 (t) for all t e [t, 0] but xi(tk) / %2(tk) for an increasing sequence of times tk r-. For i = l,2, set ct*(£) = max{a; Xi(t) e By (i), the maps t o*(i) are piecewise constant and left continuous. Hence there exists 6 > 0 and indices Qi, such that for every t E [t — 5, t], we have q*(£) = &i, i = 1, 2. But then di = oq(r) = a^C7) = ^2- Moreover, #i(-) and are solutions of the same differential equation x — g^x) with same initial data. Since is continuously differentiable, they coincide on an interval [r — 5, т]. This contradiction yields the uniqueness result. (iii) Finally, let xy : [ay, by] »—> ]Rn be a sequence of solutions of (9.15), satisfying (9.17). We need to prove that the limit x(-) is also a Caratheodory solution. First, we observe that on any compact subinterval J C ]a, b[ the functions xy are uniformly continuous and intersect a finite number of domains say with indices cq < &2 < • ’ • < Qm. For each 1/, the function o*(t) = max {a e A ; xu(t) e is non-decreasing and left continuous, hence it can be written in the form = if By taking a subsequence we can assume that, as и —► 00, Vj —> tj for all j. By a standard convergence result for smooth O.D.E’s, the function x provides a solution to x = gQ . (x) on each open subinterval Ij = ]tj, tj+i [• Since the domains are open, there holds x(t) Qfj for all /3 > aj, t E Ij. On the other hand, since gQj is inward pointing, a limit of trajectories xy = gaj(xy) taking values within 31aj must remain in the interior of I2aj. Hence a*(x(f)) = aj for all t E Ij, completing the proof of (iii). Example 9.1 Consider the covering of J? = IR2 consisting of = IR2, = {(^1,^2); xi < -x%}, = {(^1,^2); X!>X%},
206 9 Patchy Feedbacks Fig. 9.2. The patchy vector field in Example 9.1. and the family of inward-pointing vector fields 91 : f?i —► IR2, 92 ' —+ IR2, 9з : -* IR2, 31(2:1,2:2) = (0,1), 32(2:1,2:2) = (-1,0), 33(2:1,2:2) = (1,0). Then the vector field 3 on IR2 defined as 3(2:1,2:2) = (0,1) if (-1,0) if (1,0) if |2?i I < X% . 2 Xi < — a?2 X1 > x% is the patchy vector field associated with : a = 1,2,3} and {ga : a = 1,2,3}, see figure 9.2. Notice that the O.D.E. x = g(x) has three forward solutions and one backward solution starting at the origin. For every initial point of the form у = (£, у/?) with £ > 0, there exist two forward Caratheodory solutions, but no solution backward in time. Remark 9.4. As shown in the previous example, the set of all trajectories of a patchy vector field starting from a given point is closed but not connected, in general. For many applications, this is actually a desirable property. Indeed, it avoids all the topological obstructions that can prevent the existence of continuous stabilizing feedbacks. See Example 4.3 of Chapter 4. 9.2 Asymptotic feedback stabilization In this section we describe a method for constructing discontinuous feedbacks which stabilize nonlinear controllable systems. The basic idea can be explained as follows. Assume that for each initial point у we can construct an open-loop control t •—► uy(t) that steers the sytem from у into a small neighborhood the origin. By continuity, this control uy will still perform the same task for
9.2 Asymptotic feedback stabilization 207 all initial points sufficiently close to y. By patching together a locally finite family of these controls uy, we shall eventually obtain a feedback control on the entire space IRn. We begin by introducing a minimal set of assumptions, in order to have a stabilizing feedback for the system (9.1). Given a point у € lRn, and a control tz(-), we denote by ; г/, iz(-)), or briefly x(-; u), the solution to the Cauchy problem x = f(x(t),u(t)), x(0) = y. Definition 9.4 (Asymptotically controllable systems). The system (9.1) is said to be globally asymptotically controllable (to the origin) if the following holds. 1. Attractiveness: for each у e IRn there exists some admissible open-loop control t i—> uy(t) 6 U such that the trajectory x(-;uy) starting at у is defined for all t > 0 and satisfies x(t\uy) —> 0 as t —► oo. 2. Lyapunov stability: for each e > 0 there exists 6 > 0 such that for every у e IRn with I?/1 < 6 there is an admissible control uy as in 1. such that |a;(t;?iy)| < £ for all t > 0. We will show that the above conditions are sufficient for the existence of a stabilizing discontinuous feedback. The basic step in the construction of a stabilizing patchy feedback is provided by the next lemma. This provides a feedback that steers every point in the spherical shell Sr,s = {#; r < |x| < s} into the inner ball Br = {x\ |rr | < 7~}. Lemma 9.2.1. Let the system (9.1) be globally asymptotically controllable to the origin. Then, for every 0 < r < s there exist T > 0y R > 0 and a patchy feedback control U : f2 i—> U, defined on some open domain satisfying {x ; r < |a?| < s} С 12 C {x ; |x| < R} , such that the following holds. For every initial state xq € 12 with |xq| > r, the Cauchy problem for (9.13) admits at least one Caratheodory solution, forward in time. Moreover, for every such solution 11—> 7(t), there exists ty <T such that |7(M<r. (9.20) Proof. 1. For every point у in the spherical shell Sr,s defined above, there exists an open-loop control uy : [0,^] —> U that steers the system from у inside the open ball Bs. The corresponding trajectory t ^(t) thus satisfies я^(0) = у and xy(ty) e Br (see figure 9.3). By an approximation argument, we can assume that the control uy is piecewise constant and right-continuous, so that Uy(fi) = Uj te [tj, tj+ib j = 0,...,A-l,
208 9 Patchy Feedbacks Fig. 9.3. Steering the system from у into the smaller ball Bs with an open-loop control. with to = 0, £дг = ty. It is clearly not restrictive to assume that the map t i—> xy(t) is one-to-one. Otherwise we can simply delete closed loops and obtain a trajectory without self-intersections. In particular 0. Fig. 9.4. Construction of a section Zj of the flow tube. 2. We first construct a patchy feedback in a tube-like domain around the trajectory {xy(t); t € [0, fj}. For each j = 0,1,... N - 1, given Cj > 0, we construct a tube Fj around the portion of the trajectory {*”(*) 5 t e [tj, ij+i]}. Set Xj = xy(tj) and consider the (n — l)-dimensional ball B' with radius Sj, centered at Xj and perpendicular to the vector f(xj, Uj) B' = {z G IRn ; \x-Xj\<£j, (х-Xj, f(xj, Uj)) = o|.
9.2 Asymptotic feedback stabilization 209 For each b € B' and t G [tj — Ej , tj+i + call t i—► X(b, t) the solution to the Cauchy problem X = f(x, Uj) , x(tj) — b • Observe that the tube (see figure 9.4) Bj = b 6 Bj? tj — Ej < t < 4~ Ej together with the vector field f(x,Uj) does not yet constitute a patch, accord- ing to our previous definition. Indeed, its boundary is not entirely smooth, and the vector field is tangent to the lateral surface, not inward pointing. We fix this problem by defining (see figure 9.5) Qj = {%(£, 6); b e B', tj - Ej 4- Cj\b - Xj\2 < t < tj+i +Ej|, where Cj is a constant large enough so that tj — Ej 4- Cj£2 > tj+i + Ej. This guarantees that the lower boundary of , say d~Q3; = |x(£,b); beBj, t = tj — Ej 4- Cj\b - xj2, is smooth. Moreover, the vector field f(x,Uj) is strictly inward pointing at every point of d~(lj . Fig. 9.5. Replacing the flow tube by a patch Qj . We now patch together the domains starting from and proceed- ing by backward induction, as in figure 9.6. For every j = 0,1,...,7V — 1, call d+flj = |x(Z, 6); b e Bj, t>Cjlb — Xj\2, t = 4- the upper boundary of Qj. We begin by choosing En-\ so small that all points on the upper boundary d+ lie inside the ball Br. By induction, after Ej has been determined, we choose Ej-i > 0 small enough so that all points on the upper boundary d~ Qj-i are contained inside f2j. Since the vector field f(x, Uj) points strictly inward on all the lower bound- aries d~Qj, by Remark 9.1 the family |(f?j , /(•, Uj)); j = 0,1,..., N — 1 j yields a patchy vector field.
210 9 Patchy Feedbacks Fig. 9.6. Patching together the domains ,..., . 3. For every initial point у in the spherical shell 5r,s, let Г2У be the open set on which a patchy feedback is constructed, as in the previous two steps. These sets form an open covering of the compact set Sr,s. We can thus extract a finite subcovering, say Qyi,..., I2!Jm. For every i = 1,..., m, let , /(•, Uij)), j = 1,..., rrii, be the patches defining the patchy vector field on Qyi. We can then define a patchy vector field on the whole set D by taking all patches (J?^-,/(•, w^)) and ordering them according to the lexicographic order. Namely, (i, J) -< (г',/) if either i < i' or i = i' and j < j'. Defining {rrii tyi + 2 eij j=l R — max sup < |rr|; x E ftyi the proof is completed. Using the above lemma, we can now prove: Theorem 9.2.2. (Asymptotic stabilization with a patchy feedback). If the system (9.1) is globally asymptotically controllable to the origin, then it admits a stabilizing patchy feedback. Proof. 1. We apply Proposition 9.2.1 to all spherical shells Sk = ; 2-/c“1 < И < 2-fc}, where к ranges over all positive and negative integers. For every k, this yields a patchy feedback (f4,£ , /(•, Uk,ef), t = 1, 2,..., Nk , defined on the open domain Nk f4 = [J c{x; 2-fe"2 < |l| < (9.21) This feedback steers all points of the spherical shell Sk into the smaller ball В = {x; |x| < 2-fc“1}. Thanks to the property (ii) in the definition of
9.3 Robustness 211 asymptotic controllability, the construction can be performed so that the up- per bounds Rk in (9.21) satisfy Rk —> 0 as к —> oo. Therefore, the family of all patches ftk,£ constitutes a locally finite open covering of the space IRn \ {0}. 2. The index set A = {(fc, £); к integer, f can be totally ordered lexicographically, i.e. by letting (fc,£) -< (V,f) if either к < к' or else к = к' and £ < (!. The piecewise constant map I/* : IRn\{0} »—► U defined by {/‘(я) = Uk,t if x^ilk^\ [J (M)<(k',r) provides the desired patchy feedback control. 9.3 Robustness In practical applications, one has to take into account the presence of several perturbations which may degrade the performance of a feedback control. For example: (i) The model equation, described by the function f in (9.1), may not be precisely known. (ii) The evolution of the system may be affected by random external pertur- bations. (iii) While implementing the feedback, the state of the system may not be accurately measured. As a result, instead of the dynamics x = f(x, C/(x)), the controlled system will actually evolve according to x = f(x, U(x + £i(*))) +e2(t), (9.22) (9.23) for some small perturbations ei, e2. Here the ‘‘inner perturbation” ei accounts for measurement errors, while the “outer perturbation” s2 models and external disturbances. In the above framework, it is important to design a feedback control which is robust, i.e. it still accomplishes the desired task in the presence of (suffi- ciently small) perturbations. As shown in this section, this important property is actually achieved by all patchy feedbacks. We shall first prove some stabil- ity results valid for patchy vector fields, then give an application to patchy feedbacks.
212 9 Patchy Feedbacks Together with the O.D.E. (9.15) determined by a patchy vector field g, consider the perturbed equation x = g(x) 4- w , (9.24) where t »—> w(t) is a (possibly discontinuous) function with bounded variation. By definition, a solution of (9.24) is a function t i—> x(t) such that x(t) = a:(0) + / g(x(s)) ds 4- [w(t) - w(0)] . (9.25) Jo for every t. Notice that the right hand side of (9.24) contains the derivative of w. In particular, if the function w(-) has a jump at a time t, the same is true for rr(-). We recall that a function ф : [0, T] i—> IRn has bounded variation if N Tot.Var.{0} = sup l<№) “ 0(^-i)l < 00 • o=to<ti< -<tN=T i=1 The total variation norm of ф is then defined as HIlBV-Tot.Var.W + HIlL- (9-26) The next lemma extends the result in part (iii) of Theorem 9.1.1 to the case where impulsive perturbations are also present. Lemma 9.3.1. Let (12,g, (12a, да)аел) be a patchy vector field. Let x„ : [0, T] i-* IRn be a sequence of solutions to the perturbed equations xu = р(жр) 4- Wy , (9.27) where Wy : [0, T] i—> IRn are BV functions such that Tot. Var.{Wi,} —► 0 as у —> oo . (9.28) Assume that the solutions х„(-) take values inside a compact subset К C 12 and, converge pointwise to a function x : [0, T] i—> K. Then x(-) is a Caratheodory solution of the unperturbed equation (9.15). Proof. 1. To establish the lemma, it suffices to show that lim [ \g(xy(t)) - g(x(tf)\dt = 0. (9.29) Jo Indeed, (9.28) implies that |wi,(t) — w(0)| —> 0 uniformly for t G [0, Т]. If (9.29) holds, we can thus conclude x(t) = lim (яД0)4- / g(xy(t))dt + Iw^t) - wp(0)l ) \ Jo / = x(0) 4- I g(x(t))dt (9.30) Jo
9.3 Robustness 213 for every r G [0,T], proving the lemma. For each t G [0, T], consider the indices = max {a G Л; x(t) G = max {a G A; xy(t) G . By the definition of patchy vector fields, one has p(x(t)) = pQ(t)(a;(t)), = Pa„(z„(*)) • Since each vector field ga is bounded and continuous, we know that xy(f) —> x(i) uniformly for t G [0,Т]. To prove (9.29) it thus suffices to to show that lim meas({£ G [0, T]; au(t) / a(t)}) = 0. (9.31) 2. Observing that f2Q(r) is open and ху(т) —> ж(т), for all v sufficiently large one has xy(r) G . Therefore lim inf ay(r) > a(r) for all r E [0, T], i/—»oo and hence Jnn^ meas({r G [0,T]; аДт) < q(t)}) =0. (9.32) The following steps will establish the identity lim measure [0,T]; ou(r) > а(т)}) =0. (9.33) 3. Toward a proof of (9.33), we first show that, for every fixed time т G [0, T], limsup meas( {t G [0,т]; оД^) > q(t)} 1 = 0. (9.34) v—>oc ' ' Assume, on the contrary, that limsup measf{t G [0,r]; ap(Z) > o(r)}") > 0. (9.35) P—>oo ' ' A contradiction is obtained as follows. Since the patches which intersect the compact set К are finitely many, there exists a unique index /3 > such that limsup measf{7 G [0,r]; ay(t) > /3}} = 0, (9.36) l/—->OO ' ' while limsup measf{£ G [0, t] ; ay(t) =/3}} = 7]q > 0 . (9.37) v—►oo ' Consider the distance function from the complement of the set Ф^(х) = dist (x ;
214 9 Patchy Feedbacks Since the vector field gp is strictly inward pointing on the boundary сИЛз, we can find 5 > 0, p > 0 such that, for every solution of the O.D.E. y(0 = ^(уЮ), one has ^Фз(у(<)) > <5 whenever 0 < 03(y(t)) < p. (9.38) at More generally, if t x(t) is a solution of the patchy O.D.E. (9.15), then ^Фр(х(1)) > 6 for a.e. t such that Фр(х(Ь)) < p and x(f) G ftp \ (9.39) ^Фр(х(1)) = 0 for a.e. t such that x(t) ftp , (9.40) ^<^(z(Z)) > -C for a.e. t G [0, T], (9.41) with C = supx€/< |<7(#)| • We also observe that Фр, being a distance function, satisfies \Фр(х)-Ф0(у)\<\х-у\. (9.42) Next, consider the functions t Ф^(жг(^)), where x„ are the perturbed solutions in (9.27). Observing that ^(r) = rCp(O) + [wr(r) - wI/(0)] + / ^(^p W) dt Jt€[0,r],<*„(t)=p + [ g(xl/(t\)dt+ I g^x^t^dt Jte[0,r],a^(t)>P Jt€[0,r],a,,(t)<P and using the previous four estimates (9.39)-(9.42), we obtain (see Pb. 9.3 at the end of this section for more details) Фр(х»(тУ) > -Tot.Var.{wp} - C • measf{f G [0, r]; f / (9.43) + min ip , <5 • meas I {£ G [0, t] ; a„(t)=/3}]>. Letting у —► oo, by (9.28) and (9.36) the first two terms on the right hand side of (9.43) approach zero. Using (9.37) to estimate the third term, we conclude limsup Фр(х„(тУ) > min<p, tfr/of >0. This implies 0/j(x(r)) = lim фр(х^т)) > 0. P—>oo Hence x(r) G i2p and q(.t(t)) > /3, reaching a contradiction.
9.3 Robustness 215 4. Next, assume that (9.33) fails. Then we can find an index /3 such that limsup measf {t G [0,T]; ct(f) = (3 and ay(t) > (3}\ > 0. i/—►oo ' ' In particular, we can find a time т G [0, T] such that a(r) = 0, lim sup meas 0. This provides a contradiction with (9.34). Hence (9.33) must hold. Together, (9.32) and (9.33) yield (9.31), completing the proof. We now consider the O.D.E. determined by patchy vector field g in the presence of internal and external perturbations: i = g(x + ei(t))+e2(t). (9.44) Assuming that the the perturbations ei, 62 are sufficiently small, the next theorem shows that every solution of (9.44) remains close to some solution of the unperturbed equation (9.15). Theorem 9.3.2. (Robustness for patchy vector fields). Let g : J? i—> IRn be a patchy vector field. Given any compact subset К C 12, and any T,e > 0, there exists 6 > 0 such that the following holds. If у : [0, T] К is a solution of the perturbed equation (9.44) and ||ei||sv<(5, IIMl00 < <5 (9.45) then there exists a solution x : [0, T] •—> 12 of the unperturbed equation (9.15) such that |a;(t) — £ for all tG[0,T]. (9.46) Proof. 1. By contradiction, assume that the conclusion was not true. Then there exist a sequence of perturbations ei^y , 62,^ with ||ei,P||BV —* 0, 11^2,1/||l°° 0 as u—>oo, (9-47) and corresponding solutions xy : [0, T] > K. which do not approach any solution of the original equation (9.15). Namely H^-^IIl~([0,t]) > £ (9.48) for some e > 0, every v > 1, and every solution x{-} of (9.15). 2. Define the functions yy(fi} = xy(t) -h eijIZ(t). These functions satisfy yv(t) = xv(f) + ei,„(t) = + e2>1/(t)) +ei,„(f).
216 9 Patchy Feedbacks Setting = ei?p(t) + / e2^(s) ds Jo we can thus write M*) = 5(M0)+ «’„(<). Notice that Tot.Var.{w^} < Tot.Var.{ei?I/} + ► 0 as z/—>oo. By the definition of solution, for every t e [0, T] one has = 6/Д0) + [ g(y^s))ds\ + [w^t) - wp(0)]. (9.49) \ Jo / The first terms on the right hand side of (9.49) are bounded and uniformly Lipschitz continuous w.r.t. t. The last terms converge to zero uniformly on [0,T]. Therefore, by Ascoli’s compactness theorem there exists a subsequence (?/^)fc>i that converges to some function z(-), uniformly for t e [0, Т]. 3. By Lemma 9.3.1, this limit function z(-) provides a solution to the un- perturbed equation (9.15). Notice that ||ei>IZ||BV —> 0 implies ||ei,p||l=» —* 0. Therefore lim inf ||xp - x||L« < hm ( + \\Уик - z||l~ ) = 0- i/—>oc к—>oo X / This yields a contradiction with (9.48), thus proving the theorem. Remark 9.5. According to the above theorem, every solution of the perturbed equation is close to some solution of the original one. In general, however, these solutions cannot be chosen with the same initial data. For example, fig. 9.7 shows the trajectories of a patchy vector field g, and a solution ?/(•) of the perturbed equation (9.44). In this case, the solution z(-) to the initial value problem z = g(z), z(0) = xQ = ?/(0), is unique, but very different from ?/(•). In order to find a trajectory £(•) of the O.D.E. (9.15) which remains always close to ?/(•), one has to start from a different initial point. Example 9.2. Consider the patchy vector field on IR2 ( (1,1) ifx2>0, S(xi,a:2) - ifa.2<o. For each v > 1, let x„ : [0, T] i—> IR.2 be the solution to the perturbed Cauchy problem
9.3 Robustness 217 Fig. 9.7. Two trajectories z(-) and rr(-) of the patchy vector field g, and a solution y(-) of the perturbed equation (9.44). ip(t) = glx^t) + ei,i,(t)) + e2>p(t), жДО) = (0, 2 "). (9.50) Calling = [(& —1) 2-1/, k 2-I/[, the inner and outer perturbations are here taken to be ° (t\-\ (°> °) if Z odd o (t\=(\ - I (Oi _22-^) jf tel^k, feeven, e2,v(t)_0. The corresponding solutions x„ are shown in figure 9.8. Observe that, as и —* oo, one has the uniform convergence xjj) —> x(t) = (t,o) t e [о, T]. However, the limit function rr(-) is not a solution of the unperturbed equation. This does not contradict Theorem 9.3.2, because lim ||ei p||l«> = 0 but lim inf Tot.Var. {ei ^} > 0 . I/—»OO ’ V—ЮО This example shows that the size of internal perturbations should be measured in the total variation norm, not in the L°° norm. Fig. 9.8. A solution xp(-) of the patchy O.D.E. (9.50) with internal perturbations.
218 9 Patchy Feedbacks As a further application of Lemma 9.3.1, we now prove a robustness result for patchy feedbacks, in connection with a stabilization problem. To avoid technicalities, we shall assume that the patchy feedback is defined on the entire space IRn, and that the function f in (9.1) satisfies the sub-linear growth condition |/(ж,и)| < cy(l + |x|) for all и G U , x G IRn. (9.51) Theorem 9.3.3. (Robustness of stabilizing patchy feedbacks). Let the system (9.1) satisfy the growth condition (9.51). As in (9.13)-(9.14), let U(x) = be a patchy feedback defined on lRn. Given a compact set K$, assume that every solution of i = g(x) = (9.52) starting inside Kq reaches the ball Br centered at the origin with radius r, within time T. Then, given e > 0. there exists 6 > 0 such that, if the pertur- bations Ei, E2 ‘ [0,T] i—> IRn satisfy Ikillsv < 6, (9.53) then any solution of (9.23) with initial data x(0) G enters the ball Br+e within time T. Proof. 1. Assume, on the contrary, that the conclusion does not hold. Then one can find e > 0 and a sequence of perturbations E\>y ,Е2^ч with ||£i,i/||bv o,, ||£2,^||l«>-> о as (9.54) and corresponding solutions t •—» x„(t) of iy(t) = f(xv(t), U(x^(t) + Ei>lz(t))) + e2, </(i), (9.55) such that the following holds. For every v > 1, гГр(О) G Kq , |xp(£)| > r + £ for all t G [0,T]. (9.56) 2. By assumption, g(x) = f(x, U(x)) is a patchy vector field. We can write (9.55) in the equivalent form iy = g(x + + e2jP(t), (9.57) with ei,p(<) = £i,p(t)> e2,„(0 - + /(^(i), U(xv(t) + ei,p(t))) + ei,p(t), U(xv{t} + £i,p(t))) • (9.58)
9.4 Nearly optimal patchy feedbacks 219 By (9.54), we also have ll^i,p||bv o,, 11^2,1/||ь°° “* 0 as и —> oo . (9.59) 3. By the sublinear growth condition, all trajectories : [0,T] * IRn remain uniformly bounded. Repeating the steps 2 -3 in the proof of Theorem 9.3.2, we can choose a subsequence x„k which converges to a limit function jr(-) uniformly for t G [0,T]. Using Lemma 9.3.1 we again conclude that rr(-) is a solution of the unperturbed equation (9.52). However, by (9.56) it follows z(0) G , |x(t)| > r + г for all t G [0,T], (9.60) contradicting the main assumption of the theorem. This achieves the proof. 9.4 Nearly optimal patchy feedbacks Consider a general optimization problem with running cost L, terminal cost and free terminal time: min < '0(я:(Т))+ I L(x(t), u(t)) dt I , (9.61) T»w() I Jo for the nonlinear control system x = f(x, u) u(t) G U, (9.62) with fixed initial datum. The minimum is sought over all times T > 0 and all measurable control functions и : [0,T] i—► U. Aim of this section is to show that this problem can be approximately solved by a patchy feedback, with an arbitrary degree of accuracy. For convenience, we list here all the basic assumptions. (H) The set of admissible control values U C lRm is a compact, the function f : JR" x U i—> IRn is continuous w.r.t. both variables, and twice continuously differentiable w.r.t. x. In addition, f satisfies the sub-linear growth condi- tion (9.51) for some constant Cf . Both the terminal cost i—> IR and the running cost L : IRn x U IR are continuous and non-negative. More precisely V>(x) > 0, Л(Ж, Ii) > Q(] > 0 for all x G IRn, uGU. (9.63) Throughout the following, V denotes the value function for the optimiza- tion problem (9.61)-(9.62), namely
220 9 Patchy Feedbacks V(y) = inf T, x(’,u) (9.64) where the minimization is taken over all T > 0 and all solutions of t i—> x(t, u), z(0, u) = y, corresponding to a measurable control и : [0, T] i—> U. The main result can be stated as follows. Theorem 9.4.1. Existence of a nearly optimal patchy feedback). Let the functions 'ip.L.f in (9.61)-(9.62) satisfy the assumptions (H). Let e > 0 and a compact set К C IRn be given. Then there exist a closed terminal set S C HVZ and a patchy feedback и = U(я) defined on the complement IRn \ S such that the following holds. For each у e К, every Caratheodory solution of x = f(xy t/(rr)) , z(0) = у (9.65) reaches the set S within finite time. Calling т = inf {t; x(t) € S'} the first time where the trajectory reaches S, we have V>(z(t)) + I L(x(t), dt <V(y)+e. (9.66) Jo We recall that, by well known properties of patchy vector fields, for every ini- tial point у e IRZ'\S the O.D.E. (9.65) has at least one forward Caratheodory solution. According to (9.66), all of the solutions starting from the compact set К are nearly optimal, for the cost (9.61). Proof. The proof of Theorem 9.4.1 will be given in several steps. 1. Various reductions can be performed. Taking a smooth approximation, we can assume that € C°°. Moreover, approximating the cost function L by a more regular function, it is not restrictive to assume that L is twice continuously differentiable w.r.t. x. Recalling that L(x. u) > Qq > 0, we can now replace f(x,u) by g(x,u) = (9.67) L[x. u) and consider the equivalent problem inf T,u(-) (9.68) with dynamics x = g(x,u), x(0) = у. Notice that the function g in (9.67) is continuous w.r.t. both variables x,u, and twice continuously differentiable w.r.t. x. Moreover it satisfies the growth condition
9.4 Nearly optimal patchy feedbacks 221 |5(z,u)| < (1 + |x|) for all и e U . In the following, we thus assume without loss of generality that the running cost is simply L(x,tz) = 1, so that the minimization problem (9.61) reduces to (9.68). 2. Choose a constant M such that M > 1, M > max 'ф(х). хек v To fix the ideas, throughout the following we assume that 0 < s < 1/8 and that the compact set К is contained in the open ball Вp centered at the origin with radius p. Because of the sub-linear growth condition (9.51), the a priori bound (2.23) of Corollary 2.1.6 holds. In particular, during the time interval t e [О, 2M] every trajectory of (9.65) starting form a point у 6 К C Bp will remain inside the open ball Bp, with p = eCf2M (p + 1). (9.69) 3. Let V = V(y) be the value function for the optimization problem (9.68), with dynamics (9.62). We claim that V is semi-concave. More precisely, there exists a constant к such that, for any y, y' e Bp, one has v(y') < v(y) + w • (у’ - у) + к - (9.70) for some vector w e D+V(y) in the super differential of V at the point y, see Section 8.2. Indeed, from the theory of optimal control [6] it is well known that the optimization problem (9.68), (9.62) with initial data з:(0) = у has at least one solution, within the class of chattering controls. Let t i—> x(t) = x(t; y,u,0) be an optimal chattering trajectory, with x(0) =?/, x(f) = /(#(£), 14(f)) t e [0, r] , i=0 for some measurable functions (it, 0) = (tto, • • •, , 0n) satisfying щ : [0, г] U, ^:[0,t]h^[0,1], n £>(*) = 1. (9.71) i=0 For any other initial data y'. we can consider the same chattering control (й, 0), always stopping at the same terminal time t = r. This yields the cost Vu'\y') = T + У,й,0У) . (9.72)
222 9 Patchy Feedbacks The regularity assumptions on /, ф w.r.t. the variable x imply that, as y' varies in the ball Bp, the map y' yu’<9’T(y) is twice continuously differentiable. Moreover, its C2 norm remains bounded: 11^й’*,Т|1са(Вд) — K- (9.73) Since r e [0, Tmax] while both й and 0 in (9.71) range over compact sets, this bound is uniform, i.e. in (9.73) we can take a constant к > 1 which does not depend on the particular chattering control, or on the time r. Observing that V(y) = V^T(y), V(y’) < V“’^’T(y/) for all y' e Bp, the inequality (9.70) follows from (9.73), choosing w = VV^’^y). (9-74) 4. As shown in the previous step, the value function V(y) = min Уй’ё’т(у) й,0,т is Lipschitz continuous on the ball Bp. By (9.74), the constant к > 1 in (9.73) also provides a Lipschitz constant for V, namely |У(д:) — V(?/)| < к |x — y\ for all x,y£Bp. (9.75) By Rademacher’s theorem, sec Theorem A.6.1, V is differentiable almost everywhere. At each point x E Bp where the gradient W(x) exists, if V(j:) < ф(х) then one has the relation, see Theorem 7.3.2: min {W(x) • /(.t,tz)} + 1 = 0. (9.76) Consider the open set V = {# ; V(x) < 'ф(х)} . Given 6 > 0, we can choose finitely many points y\,... ym G Bp П T) such that W(?;? ) is well defined for each i = 1,..., m, and moreover m ВрПЪС \jB(yi,8). (9.77) 1=1 Define the approximate value function IT(x) = min {ф(х), fVi(x), ... , VLmGr)}, (9.78) where
9.4 Nearly optimal patchy feedbacks 223 Wi(x) = V(j/j) + VV(yi) • (x - yi) + k|x -yi\2. (9.79) We claim that, by choosing 8 > 0 sufficiently small, for all x G Bp the following relations hold. V(z) < W(x) < V(x) + e, (9.80) min {Vl4\(x) •/(x, ?/)}-h l < £ whenever кИДгг) = W(a?), (9.81) Indeed, the first inequality in (9.80) follows from (9.70). Next, since f is continuous and U is compact, we can find Ji e]0,1] such that the following conditions hold. If x 6 Bp , w = WQ/) exists and min {w • f(y, n)} -h 1 = 0, izGU |w' - w| < 2k<5i , |ж — y\ < , then | min {w' • f(x,u)} + 1| < £. (9.82) We now choose 8 > 0 such that Given any x G Bp, if j is an index such that |rr — yj | < <5, recalling the Lipschitz condition (9.75) we find W(x) < V(j/j)+|VV(%)| |ж-у,|+ф-у,|2 < У(ж)+2ф-у,|+к|ж-г/;|2, / к8^“ i Ж(ж)-У(ж) < min |г, (9.83) This already yields (9.80). Comparing (9.70) with (9.79) we notice that W) - V(x) > к |Z ~ У<|2 • Hence from (9.83) it follows |x — yi\ < <51 whenever Ж(^) = W(x). (9.84) Observing that |v^(x)-vw^(%)| < 2/< |x — yj\ < 2k(5i , from (9.82) we deduce the inequality (9.81). This establishes our claim. 5. By the definition of PK, it is clear that all level sets where Wi is constant are spheres. Indeed, for any given constant c we can write
224 9 Patchy Feedbacks {x; Wi(x)=c} = {rr; |x-#i|=r}, with Xi = yi — /2к and a suitable radius r. For each i = 1,..., m, consider the set ^ = {хеВ-р- Wi(x) = Ж(х)}. (9.85) In this step we show that there exists a minimum radius rmin > 0 and a maximum radius rmaa, such that, if x E Pi, then the level set where Wz — Wj\x) is a sphere of radius r with 'min _ max • (9.86) Indeed, since e< 1/2, by (3.17) it follows (9.87) Calling Mf = max |x|<p,uEU from (9.87) we deduce 1 2M~f ’ Therefore |VWi(x)| 2k 1 4к Mf On the other hand, by (9.79) and (9.84) we have min • |VlV,(x)| < IVWiG/Jl +2K\x-Vi\ < /с + 2к<51 < Зк. Hence |V^(x)| < 3 2k “2 T'max • (9.88) 1 6. We are now ready to construct the near-optimal patchy feedback on the open set Г2 = {z e Bp; W(x) < . (9.89) The terminal set S will then be defined as S = EV1 \P. Given r) > 0 small, for each point x G T>i consider the point (see figure 9.9) Pi 2 1 r+3 X Xi
9.4 Nearly optimal patchy feedbacks 225 and the ball Bx — B(px, |ж — xJ/3) centered at px with radius r = |x — Xf|/3. By (9.81), there exists a nearly-optimal control value и = ux e U such that (9.90) Consider the lens-shaped region ri = B(Pi, \ B(Xi ’ Iх “ “ *0- (9'91) Its upper boundary will be denoted as д+Ггх = дГх \ Bfa , |z - - V) • (9.92) Moreover, for z € д+ Гх, we write п^(г) for the outer unit normal at the point z. W.=W.(x) Fig. 9.9. Construction of a lens-shaped patch. We claim that, by choosing p > 0 sufficiently small, the following holds: VU'(z)-/(z,^)<-l + 2e for all z e Гх , (9.93) n<(z) • < -7) for all z&d+rtx. (9.94) Moreover, the constant p > 0 can be chosen uniformly valid for all i — 1,..., m and all x e Bi. For fixed z, x this is clear because, as p 0, the diameter of the set Гх approaches zero. Moreover, as z varies on the upper boundary д+Гх , all the unit normals n^z) approach the vector VWi(^)/| VWi(x)|. Therefore, both inequalities (9.93)-(9.94) follow from (9.90). We now observe that f = /(a;, u) is uniformly continuous on the compact domain Bp x U. Moreover, on each set 7?г, the gradient VWi(j:) is uniformly
226 9 Patchy Feedbacks Lipschitz continuous and bounded away from zero. Finally, the radius of each level set, where Wi is constant, by (9.86) is bounded above and below. This allows us to choose a constant rj > 0 uniformly valid for all i,x. 7. To achieve a nearly optimal feedback, we would need the inequality VW(z) • /(г, u?) < -1 + 4s for all z G Ггх . (9.95) If W(z) = Wi(z) for all z e Гх, this is a trivial consequence of (9.93). However, we must also consider the case where some of the points z G Г? lie in a region where W(z) = Hj(z) < W2(z). for some different index j. For this purpose, we observe that the set where Wi = Wj is always a hyperplane, say = Wi(x) = WjCr)} = {я; nij-x = Cij}. (9.96) for a suitable constant Cij and a unit normal vector nZJ . The orientation of riij will be chosen so that {.t ; Wi(z) < Wj(x)} = {a?; nij - x < Cij] . We claim that, by choosing 77 > 0 sufficiently small, uniformly w.r.t. г,х, one of the following two cases occurs (see figure 9.10). CASE 1: At every point z G Гх П Hij one has nij-/(z,O < -т/. (9.97) CASE 2: At every point z G Гх one has VW^z) -f(z,ux) < — l+4e. (9.98) Indeed, assume that (9.97) fails. Then there exists a point z* G Гх D'Hij such that n0-/(z*,uf) > -77. (9.99) By (9.96) and the orientation of the unit vector n4- , we can write VW/z*) = V^(z*) -(3nij (9.100) for some constant (3 > 0. Together, (9.93) and (9.99) now imply VHA(z‘)-/(^,uf)= VHA(z*)-/(z‘,7zn-/?no-/(^,«?) /qwn < -l + 2e + /3r] < — 1+Зг, 1 ' provided that we choose t] > 0 sufficiently small. Since f and VWj are uni- formly Lipschitz continuous, from (9.101) it follows that (9.98) is valid for all
9.4 Nearly optimal patchy feedbacks 227 z sufficiently close to 2*. By reducing the size of r] > 0, we can make the di- ameter of the lens-shaped domain Г? as small as we like. Hence the inequality (9.98) will hold for all z G . To prove or claim, it remains to observe that the functions f and are uniformly continuous, and that the constant /3 in (9.100) remains uniformly bounded. Hence the constant 7/ > 0 can be chosen uniformly valid for all г, We now define the smaller domain Pf = r?\ U {ze]Rn; (9.102) jeli where Ц C {1,... , m} is the set of indices j i for which CASE 1 applies. By the previous analysis, for each j such that Wj(z) = W}(z) for some z G Г®, two cases can occur. If CASE 1 applies, then the vector field is strictly inward-pointing along the portion of the boundary dQ? where Wi = Wj. On the other hand, if CASE 2 applies, then (9.98) holds on the entire domain Г? . Notice that 12? always contain a ball centered in x. Fig. 9.10. If the domain Г? intersects the half-space where Wj < W/, two cases must be considered. Left: in Case 1, the vector field points toward the set where Wt < Wj. As a patch we then take the shaded region Pf G Г*. Right: In Case 2, points toward the set where Wj < VK.We can now take Г2? = Г?, because the control u* is nearly optimal on this whole region. 8. Consider the family of all domains J??, as i G {1,..., m} and x ranges over the closure of the set Q = {.r G Bp ; W(x) < Since all these domains are defined by the same 77 > 0, they are “uniformly large”. More precisely, for each x G 12, consider the union U« >taken over all i G {1,..., m} such that x ET>V Then this set contains the ball with radius t] centered at x. It now remains to select finitely many domains 12? which cover the compact set J2. This last step, however, must be done with some care because on the lower portion of the boundary d~Q* = 012? П Bfa , |x - - 7/) (9.103) the vector field may not be inward-pointing. To cope with this prob- lem, we first observe that there exists a uniform constant h > 0 such that
228 9 Patchy Feedbacks W^z) < W^-th, (9.104) for every i, x and every z E d (7? . Wc now set M* = max { IF(x) ; x E Bp }, and split the domain 1? in sub-domains of the form = {ж E (2; NB - (£ + l)h < W(x) < M* - th} . (9.105) For each £, we cover the compact set with finitely many patches constructed as in step 7. choosing x E . After a relabelling, this yields the patches (see figure 9.11) (fytt,/(>4<>))> a (9.106) On the collection of all patches (9.106) we define the lexicographic order: (Ла) -< (Ла') iff either ( < f! or f — and a < a'. Fig. 9.11. The domain f? = U is covered by a family of patches . We claim that the above construction yields a patchy vector field: g(x) = f(x,u^Q) iff X E \ (ЬХ'.а') (9.107) Indeed, according to Remark 9.1, it suffices to check that, for each patch the vector field /('Ль) = is inward pointing at every point of the set Q n dQt,a \ IJ Рг,а/ . (€,»)<«',а') In the present case, this is clear, because the only boundary points where is not inward pointing are those on the lower boundary d~ i2f. Since x E we have IV(rr) < Af* — £/?,, and hence by (9.104) W(z) < M* -(€ + 2)Л for all z E d f2f . Therefore, given any point z E d , either W(z) = 'ф(г) and z 12, or else z is contained in a patch f2^yOt> with as required in Remark 9.1.
9.4 Nearly optimal patchy feedbacks 229 9. To complete the proof, we now check that the patchy feedback that we have constructed is nearly optimal. We recall that, by the analysis in step 7, for every г, x we have VWi(z))-/(z, u?)<-l + 4e for all zef2*. (9.108) Now take any initial point у 6 К and let t •—> x(t) be any Caratheodory solution of the Cauchy problem x = g(x) x(0) = у , with g defined at (9.107). Call г > 0 the first time at which x(t) reaches the boundary of the set C = {x* E Bp ; W(x) < 'ф(х)} . By (9.108) we have dt < (-1+4s)t. Since 0 < s < 1/8, the above implies W(y) — W(x(r)} T l-4g ~ 2W^ ~ ~2M’ By the definition of p at (9.69), it follows that the trajectory 11—> x(t) remains inside the open ball Bp for all t E [0,т]. Therefore, at time t = r we must have W(j;(t)) = VJ(^(T))- Stopping at time r, since W(a?) > t^(x’) > 0 and V(t/) < M, the total cost can be estimated as , ,7 / n . W(y) - Vy(x(T)) , т + ^(ж(т)) < --------- 1 — 46 , V(w) + e , ... . , + 1) S TT £1,i' 1—4e Since 6 > 0 was arbitrary, this completes the proof. Problems 9.1. Consider the covering of IR2 consisting of T2i = IR2, J?2 = {(#i, ^2); ^2 > Xi}, 1^3 = {(^1,^2); < 0, x\ < —^2}, and the family of inward- pointing vector fields g± = (cos(a), sin(o)), g2 = (0? 1), Рз = (—1, —1)- Let g be the associated patchy vector field, see figure 9.12. For every a E [0, 2тг] compute the set of forward and of backward solutions from the origin. 9.2. Let (f?#,^) be a single patch, and assume that the open set is bounded, with smooth boundary. Using (9.7), give a detailed proof of the inequality (9.38).
230 9 Patchy Feedbacks Fig. 9.12. The patchy vector field in Problem 9.1. Hint: since the boundary dQp is smooth, there exists p > 0 such that the following holds (see figure 9.13). If a? € I? is any point such that dist^x; < p, then there exists a unique closest point 7r(rr) E with In this case, one has / X X ~ where n7r(T) is the unit outer normal to the set at the boundary point 7t(j:). Fig. 9.13. The distance from the boundary is strictly increasing, along the trajec- tories of the inward-pointing vector field gp. 9.3. Give a detailed proof of the estimate (9.43). Hint: consider first the case where < p for all t E [0,т]. Prove
9.4 Nearly optimal patchy feedbacks 231 that (9.43) holds if wy = 0. Then argue that, since Фр is Lipschitz con- tinuous with constant one, the presence of the perturbation wy can de- crease the terminal value Фр(яг(т)) by an amount < Tot.Var{wzy}. In the case where supfe[0 Фр(хр(£)) > p, define the time r' = sup{t G [0, г]; Ф@(ху(1)) > p} and study the function t i—► Фр(жм(^)) on the interval [г', т]. 9.4. Consider the control system (4.13). Assume attractiveness, Lyapunov sta- bility, and let the assumptions of Theorem 4.3.1 hold. Prove that, on every ball B(0, R) centered at the origin with radius R > 0, there exists a patchy feedback stabilizing the system to the origin defined by a finite number of patches. Hint: Considering a level curve of a (local) Lyapunov function, one can use a single patch in a neighborhood of the origin. 9.5. Consider the control system (4.13) and assume that there exists a C1 function V such that V (0) = 0, V(x) > 0 for x 0 and V(x) oo when |#| —> oo. Moreover, assume that inf W(x) • /(.r,?z) < 0 for all x. u£U Prove that there exists a stabilizing patchy feedback.

10 Impulsive Control Systems This last chapter contains an introduction to the theory of impulsive control systems , described by equations such as x = x, u, u). (10.1) Here x 6 IRn is the state variable, the control и ranges in a set U C IRm and we assume that Ф is continuously differentiable w.r.t. all variables. When the control и is absolutely continuous, its derivative й is an integrable function, de- fined almost everywhere. A solution of (10.1) can thus be defined in the usual Caratheodory sense, i.e. as an absolutely continuous function which satisfies the differential equation at a.e. time t. On the other hand, when the control и is discontinuous, its derivative must be interpreted as a distribution. This gives to the system (10.1) an impulsive character, because the corresponding trajectory can then be discontinuous as well. In this case, the previous concept of Caratheodory solution is no longer applicable, and an alternative definition is needed. We shall focus on two main cases. First, we shall consider systems with vector-valued controls, where the derivative of the controls enters linearly in the equations: m X = f(t, X, 'll) + • (10.2) 1=1 Later, we study the case where the derivatives also enters quadratically: 771 ГП x = f(t,x,u)-\-^^gi(t,x,u)Ui-\- hij(t,x,u)uiUj . (10.3) i=l i,j = l As a motivation for the above models, in Section 1 we introduce a class of con- trolled Lagrangian systems, whose equations naturally have impulsive char- acter. The theory of control of mechanical systems by means of moving con- straints was initiated independently by Aldo Bressan and by Charles-Michel
234 10 Impulsive Control Systems Marie, around 1980. The memoir [20] was motivated by problems of optimal control for the ski or the swing, later studied in [21]. In [66] one can find a more general geometric approach, also including some mechanical applications. The connections between the two approaches were clarified in [24]. In order to define a generalized concept of solution for (10.1), which is consistent with the classical one when и is absolutely continuous, a natu- ral approach is to approximate the measurable control u(-) by a sequence of smooth control functions and study the limits of the correspond- ing trajectories a;^(*) as p —> oo. For a given initial datum t(0) = x, two possibilities may arise. CASE 1: As v —► oo, the sequence of Caratheodory solutions x^y\-) converges to a unique limit x(-) which does not depend on the choice of the approxi- mating sequence. It is thus appropriate to define this limit x = x(^u) as the generalized solution to the Cauchy problem, corresponding to the control u. CASE 2: As и —» oo, the sequence x^ may diverge, or converge to different limits depending on the choice of the approximating sequence , In this case, additional information is needed in order to determine a well defined solution. The main results presented in this chapter can be summarized as follows: • For the system (10.2), if the vector fields gi satisfy a crucial commutativity assumption, then CASE 1 occurs. Discontinuous trajectories, corresponding to controls having jumps, can be uniquely defined as limits of smooth solutions. By a suitable change of variables, the presence of the time derivative й can be entirely eliminated from the equations. This reduces the system to a standard form, without impulsive character. All the classical results of control theory can then be applied to this equivalent system. • Still for the system (10.2), if the vector fields gi do not commute, there is no canonical way to define the trajectory determined by a general discontinuous control t I—> u(t). A unique trajectory 11—* x(t, u) can be still defined under ad- ditional conditions, namely: (i) The control function tz(-) should have bounded variation, (ii) At each time r where и has a jump, a continuous path joining the left and right limits u(r—), 'u(r-h) of the control, should be specified. This leads to the concept of graph completion of the control function tz(-), first introduced in [16]. • In cases where the equation contains the square of the derivative of the control, trajectories corresponding to discontinuous controls typically blow up instantly. One thus needs to restrict the analysis to absolutely continuous controls with square integrable derivative. The set of trajectories of (10.3) can be described in terms of an auxiliary differential inclusion.
10.1 Mechanical systems controlled by moving constraints 235 The final section of this chapter is concerned with optimization problems for impulsive systems. When the commutativity assumptions hold, we show that a problem in Mayer form can be reduced to a standard optimization problem for a suitable non-impulsive control system. This auxiliary optimiza- tion problem can then be analyzed by well established techniques, such as the Pontryagin Maximum Principle or the PDE of dynamic programming. 10.1 Mechanical systems controlled by moving constraints Consider a mechanical system described by N Lagrangian variables Qi,..., q^. As usual, upper dots will denote derivatives w.r.t. time. Let the kinetic energy be given by the quadratic form 1 N T(q,q) = - ^2 (Ю.4) t,J=l and assume that the system is affected by external forces having components Qz = Qi(t,q,q). The motion of the (uncontrolled) system is thus determined by the equations 1=1.....N- d»-5» There are two distinct ways in which an external controller can influence the system, both physically meaningful, leading to substantially different sets of equations. • The controller can apply additional forces, whose components (f)i(q. u) de- pend continuously on the state q of the system. In this case, one obtains the system of equations d дТ ЭТ ла? = а? + <Ш’’’, + л('!’“) i-ь.(юс) This leads to a standard control system, where the right hand side depends continuously on the control u(-). • The controller can directly prescribe the values of some of the coordinates as functions of time. Namely, let N = n + m and assign the values qi(t) = Uj(t), j = n + l,...,n + mof the last m coordinates as functions of time. Then the evolution of the first n = N — m free coordinates will be determined by an impulsive system of equations, linear or quadratic w.r.t. the time derivatives of the control functions.
236 10 Impulsive Control Systems Example 10.1 Consider a small child riding on a swing, pushed by his mother. His motion can be described as a forced pendulum, say of length p and mass m (see fig. 10.1, left). In addition to the gravity acceleration, the child is subject to a force и = u(t) exerted by the mother who is pushing. Denote by 0 the angle formed by the swing with a vertical line. The motion is then described by the equation mp0 — — mgp sin 0 + и . (10-7) Calling w = 0, we obtain a control system in standard (non-impulsive) form: lu = — g sin 0 4—— и. mp Fig. 10.1. Left: a child pushed on a swing. Right: a boy on a swing, standing up and down. Example 10 .2 Next, consider an older boy riding on the same swing. By standing up or kneeling down, he can change at will the radius of oscillation (see fig. 10.1, right). We describe this new system in terms of two variables: the angle 0 and the radius of oscillation r. The kinetic energy is given by T(r, 0, r, 0) = — (r2 + r2#2), (10.8) while the potential energy is V(r, 0) — mgr cos 0. The control implemented by the boy amounts to assigning the radius of oscil- lation as a function of time, i.e. r = u(t), for some control function u. Calling L = T — V the associated Lagrangian function, the evolution of the remaining coordinate 0 = 0(t) is now determined by the equation
10.1 Mechanical systems controlled by moving constraints 237 dfrL _ dL dt дв dO ’ which in this case yields 2mr0r + mr20 = — mgr sin# . (10.9) Calling cu = 0 the angular velocity, we obtain an impulsive control system of the form (10.2), namely • . g sint* 2w . ,inim # = iv , v =-----------------u. (10.10) и и Observe that in the above equation, the derivative of the control enters only linearly. Example 10.3 Consider a bead of mass m that can slide freely along a rigid bar of negligible mass. The bar can rotate, with one end fixed at the origin (fig. 10.2, left). Calling 6 the angle formed by the bar with a vertical line, and with r the distance of the bead from the origin, the kinetic energy of the system is again given by (10.8). However, assume that now we assign the angle в = u(t) as a function of time, and regard the radius r as a free variable. The motion is now governed by the equation d_ dL _ dL dt dr dr r = rd2 + g cos в . Introducing the radial velocity v = r, we thus obtain the impulsive control system r = v , v = g cos v 4- ru2 . Notice that in this case, the equation contains the square of the derivative of the control u. We now describe a general framework for the impulsive control of La- grangian systems. Consider a system described by N = n + m Lagrangian coordinates, say Qi,...,^n, Qn+ь • • Qn+m- Let (10.4) represent its kinetic energy and and assume that the system is affected by external forces having components Qi = Qi(t,q,q)- The motion of the (uncontrolled) system with m + n degrees of freedom is thus determined by the equations (10.5). As- sume now that a controller prescribes the values of the last m coordinates gn+i,... ,gn+m as functions of time. This will be achieved by implementing m frictionless constraints. Here frictionless means that forces produced by the constraints make zero work in connection with any virtual displacement of the remaining free coordinates More precisely, call ФД£) the components of the additional applied forces, needed in order to achieve the equalities
238 10 Impulsive Control Systems Fig. 10.2. Left: a bead sliding on a rotating bar. Right: two masses joined by a rigid bar, with the first mass contrained on the t/-axis. qn+i(t) = Ui(t) i = (10.11) The motion is now determined by the equations = 7Ti + + #»(*) i = l,...,n + m. (10.12) dt dql dql The assumption that the constraints are frictionless means that the following identities hold: ф1(^ = ... = фп(^) = 0. (10.13) Remarkably, there is no need to compute the remaining forces Фп+1 ,..., Фп+т in order to completely determine the evolution of the system. Indeed, the coordinates gn+1,..., qn+m are already assigned by (10.11). Of course, their time derivatives 9n+1=ui(f), ... ,qn+m = um(t) are determined as well. We now show that the evolution of components g1,..., qn can be derived from the first, n equations in (10.12), taking (10.13) into account. This is done in two steps. STEP 1: In connection with the quadratic form (10.4), introduce the conjugate moments Pi = Pi(q,q) = qt- = 52Aj(9)qj- (10.14) i=l Moreover, define the Hamiltonian function - n+m H(q,p) = -j BlJ (q) pipj , (10.15)
10.1 Mechanical systems controlled by moving constraints 239 where BIJ are the components of the (n+m) x (n+m) inverse matrix В = A x. In other words, 1 0 £ BljAjk = J=1 if i = k, if i 7^ k. STEP 2: Solve the system of Hamiltonian equations for the first n variables 9* Pi ^p) - + Qi(t,q,q) i = l,...,n. (10.16) Notice that (10.16) is a system of 2n equations for q},.... qn,pi,... ,pn, where the right hand sides also depends on the remaining components pJ? j = n + 1,... , n + m. We can remove this explicit dependence by inserting the values qn+t=iii(t) z = l,...,m, < Pj = Pj(Pl- •• ,Pn, Qn+1, . • • ,Qn+m) j = n + l,...,n + m. (10.17) In (10.17), to express Pj as a linear combination of pi,... ,pn, gn+1,..., qn+my we proceed as follows. Let C = (Cij) be the inverse of the rn x m submatrix (B'^)i,j=n+l,...,n+m> SO that if j = /с, if j ± k, j,k e {n + 1,.., ,n +m} . (10.18) Recalling that p = Aq, q = Bp, we multiply by Cji both sides of the identity n n+m 6f = £^Pfc + £ в'кРк, fc=l Jc=n+1 then we sum over z = n + l,...,n + m. By (10.18), this yields n+m n+m n Pj — £ Cjiq1 - £ ^2CjiBlkpk j = n+l,...,n + m. (10.19) i=n+l i=n+lfc=l Inserting in (10.16) the values pn+i? • • • ,Pn+m given at (10.19), we obtain a closed system of 2n equations for the 2n variables (71,...,gn,pi,... ,pn. We now take a closer look at the equation of motion derived at (10.16)- (10.17). For simplicity, we shall first assume that there are no external forces, i.e. Qi(t,q,q) = 0. The extension to the general case is straightforward. Fix an index i G {l,...,n}. Inserting the values (10.19) for the last m components in (10.16) and recalling the definition of the Hamiltonian function at (10.15), we obtain
240 10 Impulsive Control Systems = E"., B'> p, + £”i", в» (£;_+”, cj( «< - Ей”, E*”., с„в'Ч) (10.20) Next, using again (10.15) and (10.19) we compute (1 । у^п-|-тп . 1 \ ()B^k 2 ' 2^j=l 2_-rfc=n4-l '2 2^j,fc=n+l ) dqi PjPk = -lY^J^Pipk - e;=1 eSi (z:in+i ckhqh - e^:+1 e?=1 ckhB^ — lvn+m dBjk (\^n+m .h y^n+m v^n TDhf^ A 2^j,bn+l dqi \2^h=n+l ^jh4 Z^h=n+1 2^=1°jh^> 14) x (ер:г+1 ckrqr - е:г+1 e;=i скгв^Ре). (10.21) Recalling that дп+г = щ, and that the matrices C(q) = (О) (7)) are invertible, a direct inspection of the above equations reveals that: • The right hand side of (10.20) is always an affine function of the derivatives fti,..., um . • The right hand side of (10.21) is an affine function of the derivatives ill ,..., um if and only if d (a} ———=0 for all i e {1,..., n} , j, к e {n + 1,..., n 4- m} . uq1 (10.22) Following [20], systems whose equations of motion are affine w.r.t. the time derivatives of the control will be called fit for jumps. If there exists a coordinate system for which the derivatives щ do not appear at all in the equations, we say that the system is strongly fit for jumps. From the above analysis it thus follows Theorem 10. 1.1. The system described by (10.11)-(10.13) is “fit for jumps” if and only if the external forces Qi are affine functions of the derivatives qJ, j = n 4- 1,..., n 4- rn, and the identities (10.22) hold. Theorem 10. 1.2. The system (10.11) (10.13) is “strongly fit for jumps” pro- vided that the external forces Qi depend only on the variables t,ql, and more- over the identities (10.22) hold, together with B'\q) = 0 i e {1,..., n} , j € {n 4- 1,..., n 4- rn} . (10.23)
10.2 Generalized trajectories for commuting vector fields 241 Example 10.4 Consider again the swing with variable radius of oscillation, described by a Lagrangian system having kinetic energy (10.8). Assigning the radial coordinate r = ufj) as a function of time, we obtain the control system (10.10), which is linear w.r.t. u. This would follow from the above theorem, observing that the gravity force Q = —gm sin в does not depend on r and that the matrices A, В = A-1 here have the form On the other hand, if we assign the angular coordinate 0 = u(t) as a function of time, the remaining radial coordinate is determined by the equation r = g cos и + r u2 . Here the right hand side contains the square of the derivative of the control function u. Of course, the assumptions of the Theorem 10.1.1 in this case are not satisfied. 10.2 Generalized trajectories for commuting vector fields The aim of this section is to provide a definition of generalized solution to an impulsive control system, where the right hand side depends linearly on the derivative of the control. As a preliminary, we observe that, by introducing additional variables Xq = t and ;rn+i = tii,..., xn+m = um with equations •ГО = 1 , ^n+1 = , • • • ? *Гп4-т = > (10.24) the system (10.2) can be put in the simpler form x = F(x) -F Gi(x)ui. (10.25) г=1 The new vector fields F, Gi on 1RA (N — 1 + n + m) do not depend explicitly on the variables t, u. For simplicity, we shall thus consider the Cauchy problem determined by (10.25), together with the initial data z(0) = x € JRn. (10.26) Our construction rely on a crucial commutativity assumption on the vector fields Gi. Precisely, we assume that all their Lie brackets vanish identically: [Gi,Gj] = 0 i,j = l,...,m. (10.27) We recall that the Lie bracket of two vector fields f, g is defined as
242 10 Impulsive Control Systems lf,g] = (Dxg)f-(Dxf)g. This is the directional derivative of g in the direction of /, minus the di- rectional derivative of f in the direction of g. It is worth noting that the assumption (10.27) is trivially satisfied if m = 1. As usual, we shall impose some smoothness and sublinear growth condi- tions on the vector fields F,Gi, which guarantee that the Cauchy problems have a unique, globally defined solution. (Л) The vector fields F, Gi are twice continuously differentiable. Moreover, there exists a constant C such that, for all x € IRA and i = 1,.... m, |F(x)|<C(l + |x|), Gi(z)| <C(1 + H). In the following, we shall use the notation т t—> х(т) = (ехрт/)(гг). to indicate the solution to the Cauchy problem ^-x(r) = /(x(r)), x(0)=x. (10.28) ат For clarity of exposition, let us first consider the case where no drift is present, i.e. F(x) = 0. Our impulsive control system thus reduces to х = ^С{(х)щ. (10.29) i=l According to Frobenius’ Theorem A. 10.3 in the Appendix, there exists a unique map и (—> Ф(и) = (exp UiGi j (a?) (10.30) \ i=i / from HV” into 1RA such that Ф(0)=х, ^-Ф(и) = СДФ(и)), for all w = (ub. ,.,ит) e IRm. (10.31) We claim that, for every continuously differentiable control function и = (ui,...,um) : [0, T] •—> lRm, the formula x(t, u) = 0(u(f) - u(0)) (10.32) provides a solution to the impulsive Cauchy problem (10.29), (10.26). Indeed, this follows at once from the properties of the function Ф at (10.31): x(0) = Ф(0) = x,
10.2 Generalized trajectories for commuting vector fields 243 I 'll' Q = 12 аГф(и(0- u(0))ui(<) = dt = £^(ф(и(0-«(о)))м*) = i=l i=l It is important to observe that, while the equation (10.29) involves the time derivative of u, the solution formula (10.32) does not. Indeed, (10.32) makes perfectly good sense for an arbitrary measurable function 1i—> u(t) G IRm. We can thus use this formula as a definition of generalized solution to the Cauchy problem (10.29), (10.26), for arbitrary measurable controls u(-). The previous analysis shows that this reduces to the usual concept of solution when и e C1. Next, we wish to extend the above construction to the more general equa- tion (10.25), where a drift is present. We shall always rely on the commuta- tivity assumptions (10.27), but we do not make any assumption on the Lie brackets [F, G$]. Consider any C1 control function t i—> u(t), and let t i-► x(t, u) be the corresponding solution of the Cauchy problem (10.25), (10.26). We seek a representation formula which does not explicitly involve the time derivative of u, and hence remains meaningful also for controls which are discontinuous. Fix an arbitrary value u* = (u*,..., e IR"1 and consider the auxiliary trajectory t > £(£, u) = I exp ^2(tz* — Ui(tY)Gi j (a?(Z, u\). \ i=l / (10.33) We shall derive a differential equation satisfied by £. Writing Z)x(exp 52 ViGi) for the differential of the map x h-> (exp 52 viGi)(x), consider the function m Dx ( exp ^(u* - Ui)G^ (a?) t=i (10.34) where the differential is computed at the point x — (exp 52 (^г — u* )GJ(£). In other words, given £ e IRA and и e IRm, the value of F*(£,n) is obtained by (i) computing the vector F at the point x = (exp 52 (u* — (ii) pulling back this vector from the point x to the point £, using the differential of the map x и-> (10.35) ПЫ = Щ ~ ui Theorem 10.2.1. Let the vector fields F,Gi be continuously differentiable, satisfy (Jit) and the commutativity assumptions (10.27). Let и : [0, T] h-> IR™
244 10 Impulsive Control Systems be a C1 control function, and let x(-,u) be the corresponding solution of the Cauchy problem (10.25), (10.26). Then the function £ defined at (10.33) pro- vides a solution to the Cauchy problem 4 = (Ю-36) C(0) = [eXpJ2« -иД0))сЛ (®). (10.37) X i=l / Proof. Since x(0) = x, the initial condition (10.37) is clear. To prove (10.36), we shall use the identities d Ovt m \ exp^UjGj I (x) = Gi j=i / (10.38) Computing the time derivative of (10.33), one obtains C = - E(=i Gi ((exPE£=l(«J - Ч?)^)^)) йг + [^(ехрЦ™,^ -u^Gj jer)] x = ЕГ1 Gi ((exp£™ j(uJ - Uj)Gj)(x)) щ + ^(exp^jl^wj - Uj)Gj)(x) F(x) + ^(exp ^^(uy-Uj)Gjj(x) • £’"i Gi(x)tii = [£>x(exp£”Lj(u* -u>(t))G>)(a:)] • F ((exp£™ -Wj)Gj)(£)) (10.39) Indeed, the two summations involving the time derivatives щ cancel out each other, according to Lemma A. 10.4 in the Appendix. Remarkably, the differential equation (10.36) does not involve any of the derivatives йъ. Indeed, as a result of the transformation (10.33), the contri- bution of the terms G/Uz has been cancelled out. A Caratheodory solution of (10.36) is thus well defined even when the control и is only measurable. Moreover, as soon as £(t,u) is known, the value of x(t,u) can be recovered by inverting (10.33): m x(t,u) = (exp^2(uj(t) -u*)Gi)(£(t,u)). (10.40) i=l Motivated by the previous analysis, a concept of generalized solution can now be introduced. Definition 10.1 (Generalized solution for a commutative, impulsive control system). Let the vector fields F, Gi satisfy the assumptions in The- orem 10.2.1. Let и : [0, T] i—> IRm be a measurable control function. Then we
10.2 Generalized trajectories for commuting vector fields 245 say that x : [0, T] i—> IRл is a generalized solution of the impulsive Cauchy problem (10.25)-(10.26) if, for some и* e IR™ the function £(•) in (10.33) is a Caratheodory solution of (10.36)-(10.37). Remark 10.1. If x(-) satisfies the above definition for one particular choice of the constant w* € IR™, then the same is true with any other choice. In other words, the above definition of generalized solution does not depend on u*. In the commutative case, the impulsive Cauchy problem (10.25)-(10.26) can thus be solved in two steps. (i) Choose any u* € IRm, for example u* = u(0), and compute the solution t h-> £(t, u) to the standard control problem (10.36)-(10.37). (ii) Recover x(t, u) from £(£,?/), using the formula (10.40). The existence and uniqueness of generalized solutions is an immediate consequence of this construction. If the vector fields Gi are C2 (i.e. twice continuously differentiable), then the same is true of the exponential map (10.35). The differential Dx(exp UiGi)(x) is thus a C1, and the function F* in (10.34) is continuously differentiable w.r.t. both £ and u. Applying the standard O.D.E. theory to the Cauchy problem (10.36)-(eq 10.315) we obtain Theorem 10.2.2. Assume that the vector fields F. Gi satisfy (J^J together with the commutativity assumption (10.27). Let и : [0, T] i—> IR™ be any bounded, measurable control function. Then the impulsive Cauchy problem (10.25), (10.26) has a unique generalized solution x(-,u), pointwise defined forte [0,T]. Remark 10.2 As a special case, assume that the control zz(-) is piecewise continuously differentiable, with jumps occurring at finitely many times 0 < Ti < T2 < • • • < Tfc < T. To fix the ideas, let и be right continuous, so that 'u(tj) = lim^Tj_|_ u(t). In this case, a piecewise continuous map t i—> x(t), continuous from the right, is a generalized solution of (10.25) provided that (i) x(-) is a classical solution of (10.25) inside each open subinterval Tj[ where и is smooth. (ii) At each time т where и has a jump, the left and right limits of x(f) as t —> т satisfy m ®(r+) = (ехр^2(иДт+) - иг(т-))Сг) (ге(т-)). i=l (10.41) Indeed, the function t £(t) remains continuous even at times r = Tj where ?i(-) has a jump. The representation formula (10.40) now yields
246 10 Impulsive Control Systems m х(т+) = (ехр^2(иДт+) -<)G,)(C(t)) 2=1 m = (ехр]Г(и,(т+) -Ui(r-))Gi^ 2=1 m •(expJ2(Ui(T-) -u*)G^(C(r)) = ( exp52(ui(r+) - Ui(r-))Gi) (x(r-)). 2=1 The next result shows that the concept of generalized solution is robust: a generalized solution x(-, u) is obtained as the unique limit of Caratheodory solutions x(-,tz^^), as the discontinuous control function и is approximated by a sequence of more regular functions u^. Theorem 10.2.3. Under the same assumptions on F Gi and и as in Theorem 10.2.2, let u^ : [0, T] »-> IRm, у > 17 be a sequence of absolutely continuous, uniformly bounded control functions such that, as v oo, u(">(0) ->u(0), и^(Т) -> u(T), ||UM - u||L1 _ o. (10.42) Call t *—> Xy(t) the Caratheodory solutions to xv = F(xI/) + Gi(xy)u^ , 2=1 x(0) = X e IR* (10.43) and let #(•) be the generalized solution of (10.25)-(10.26). Then, as у oo one has Xy(T) t(T) , [ \xy(t) -x(t)\dt -> 0. (10.44) Jo Indeed, for a fixed u* G IRm, the solutions of tW = F*(C,(t), uy^), in C(0) = (exp ^2« 2=1 (0))ф), converge to the unique solution of the Cauchy problem (10.36)-( 10.37), uni- formly on [0,Т]. The limits in (10.44) are thus an immediate consequence of (10.42) and of the representation formula (10.40). Example 10.5. Adding the variable х’з = u, the impulsive system (10.10) takes the standard form (х1,х2,хз) = 6^2, — ,0^-i-fo, _i^ = F(x) + G(x)u . (10.45) \ хз / \ хз /
10.3 The non-commutative case: graph completions 247 Since m = 1, in this case the commutativity assumptions are automatically satisfied. Solving the differential equation x — G(x) we find 2 (expuG)(xliX2Jx3) = xi, 7-------—x3 + и . \ \x3 4- uy J Choosing u* = 1 as reference value for the control, from (10.33) we obtain 2 (Ci,6,6) = fci, ^'<3— 72, X3 + 1 ~u). (10-46) At this stage, it is useful to recall the physical meaning of our variables: x\ = Ci =0 is the angle formed by the swing with the vertical direction, X2 = 6 is the angular velocity, while x3 = и = r is the radius of oscillation. In (10.46) we therefore have £3 = 1, while £2 = Or2 is the angular momentum. The differential equation satisfied by £ can of course be recovered using the general formula (10.34). However, it is more convenient to derive it directly from (10.9). Observing that ^-(0r2) = вг2 4- 20rr = (-~—П— - | r2 4- 20rr = -grsinO , (10.47) at \ r r I we obtain for £ the non-impulsive system (41,4г,4з) = (^, -pusin6, 0) = F*(6«)- (10.48) Notice how the terms involving r = й cancel each other out in (10.47). As soon as the solution £(•) is computed, the evolution of the original variables Xi can be recovered using the identity (Ж1,®2,^з)(«) = (б(0, (Ю-49) x CL I / 10.3 The non-commutative case: graph completions In this section we consider again the impulsive control system rn x = F(rr) + СУх)щ , i=l x(0) = X , (10.50) but we drop the crucial assumption (10.27) on the commutativity of the vector fields Gi. In this case, Frobenius’ Theorem A. 10.3 cannot be applied and one cannot construct any map и Ф(и) having the properties (10.31). If u(-) is a discontinuous control, we can still approximate и by a sequence of Lipschitz controls u^\ in L1 and pointwise almost everywhere. However,
248 10 Impulsive Control Systems the corresponding trajectories will now heavily depend on the approxi- mating sequence. Example 10.6 Consider the impulsive system on IR2 = (1,0)^i + (0,^i)w2 = G](x)ui + G2(a:)w2 , (10.51) with initial condition (a?i,x2)(0) = (0,0). (10.52) Observe that in this case the vector fields Gi,G2 do not commute. Indeed, their Lie bracket is [Gi,G2] = (0,1). Consider the discontinuous control func- tion (M2)(0= (Ю.53) One way to approximate the discontinuous control и by more regular con- trol functions is as follows. (u (0,0) (0, 1 + v(t - 1)) (Ki-1), 1) (1,1) if t e [o, i - i/i/], if t € [1 - 1/p, 1], if t € [1, 1 + l/И , if t € [1 + 1/p, 2]. (10.54) The corresponding Caratheodory solutions of the Cauchy problem (10.51)- (10.52) are computed as (0,0) (i/(f- 1), 0) (1,0) if t e [o, 1], if t e [1, 1 + i/H, if t e [1 + l/p, 2]. (10.55) As v —> oo, the above sequence of trajectories converges (pointwise and in L1 to the limit trajectory / f (0,0) if t e [o, 1], if t e [1, 2]. (10.56) Next, consider a second approximating sequence f (0,0) if t e [o, i - i/И, c)"1.4p))(<) = < (1 +i/(t- 1), 0) (1, Hi-l)) if if t e t e [i -Ш i], [i, i + i/H, (10.57) I (1,1) if t € [1 + 1/^, 2]. The corresponding solutions (10.51)-(10.52) are now 1 «2И)(0 = * (a?i,a;2)(f,u(‘/)) =
10.3 The non-commutative case: graph completions 249 (xi, Xz)(t, U^) = < (0,0) (l + i/(Z- 1), 0) (1, -1)) (1,1) if if if if t e [0, 1 - 1/p], t e [i - i/p, i], te [1, i + 1/И, t e [1 + i/p, 2]. (10.58) As » oc, in this second case the limit trajectory ( J(o>o) (xl,X2)(t) = j L’jj if t e [o, 1], if t € [1, 2]. (10.59) is still well defined, but different from (10.56). The above example shows that, in the non-commutative case, the limit of the approximating trajectories depends not only on the pointwise values of u, but also on the way we approximate и by more regular controls. Observe that in the first case the values of change from (0,0) to (0,1), and then to (1,1). In the second case, the values of vary from (0,0) to (1,0), and then to (1,1). This suggests that, in the noncommutative case, a unique trajectory can be determined only if, at every time r where и has a jump, we specify along which path the transition from u(r—) to u(t+) takes place. The next definition, introduced in [16], makes this more precise. Definition 10.2 (Graph completion). Consider any function и : [0,T] i—> lRm. A Lipschitz continuous path 7 = (70,71, • • •, 7m) • [0, S'] ► [0, T] x IR™ is a graph-completion of и if • 7(0) = (0,7i(0)), 7(5) = (T, ii(T)), • 7o($i) < 70(^2) for all 0 < $i < s2 < S', • for each t € [0,T] there exists some s such that 7(5) = (£, The path 7 thus provides a continuous parametrization of the graph of и in the (t, u) space. In the case where и is piecewise continuous, at a time т where и has a jump the path 7 must include an arc joining the left and right points (r,u(r—)), (t,?i(t+)). Lemma 10.3.1. A graph completion of и exists if and only if и has bounded total variation. Proof. l.Let 7 = (70,7i ? • • • ,7n) be a graph-completion of u. For any finite sequence 0 = to < t\ < • • • < tk = T we can choose parameter values 0 = So<3i<---<Sfc = 5 such that 70 (sj) = tj. We then have ^2 MM - MM “ 3 3 |7(s)|ds- Taking the supremum over all increasing sequences to < ti < • • • < tk, к > 1, we obtain
250 10 Impulsive Control Systems Fig. 10.3. Two different graph-completion of the same control function 11—> u(t). Tot.Var. |?(s)|c/s ОС . because 7 is Lipschitz continuous. 2. Viceversa, assume that the control function u(-) has bounded variation. As a consequence, и is a.e. continuous, with at most countably many points of jump. Moreover, it admits left and right limits u(r—), u(r+) at every time t. We shall construct a graph-completion of и by bridging each of its jumps with a straight segment. For each т G [0, T], consider the total variation of и restricted to the half-open subinterval [0, r] V(r) = sup MM “ ufe-i)l • 0<tO<tl<...<<N<T Set S = T + V(T) and define the path 7 : [0, S] [0, T] x HT as follows. Observing that the map 11—> t + V(t) is strictly increasing, given s G [0, S] there exists exactly one т G [0, T] such that т + V(т—) < s < т 4- V(т-h). We consider various cases: (i) If s = г + V(r), we set 7(5) = (r, w(r)). This happens, in particular, if и is continuous at r. (ii) If т + V(t—) < s < т + V(r), say s = 0[r + V(r)] + (1 — 0)[t + V(r—)] for some 0 G [0,1], we set 7(5) = (t, 0u(r) + (1 — 0)u(r—)) . (iii) If т + V(r) < s < r + V(t+), say s = 0[r 4- V(t+)] + (1 — 0)[r + V(r)] for some 0 G [0,1], we set 7(5) = (7-, ^z(r-h) 4- (1 — 0)u(r)^ .
10.3 The non-commutative case: graph completions 251 It is now easy to check that the above construction satisfies all conditions required by the definition of graph completion. In particular, the map s i—> 7(5) is absolutely continuous, being Lipschitz continuous with constant L = 1. Remark 10.3. The graph completion constructed above is in a sense “canon- ical”, assuming that the control function u(-) takes values in the Euclidean space IRm. For certain applications, however, it is more natural to consider controls taking values in an m-dimensional manifold AL In this case, per- forming the above construction w.r.t. different charts will give rise to different graph-completions. Using graph-completions, we can now uniquely determine solutions to the impulsive Cauchy problem (10.50). Let u(-) be a control function with bounded variation, and let 7 = (70,71, • • • ,7m) be a graph-completion of u. Consider the related Cauchy problem d ш —y(s) = F(2/(s))7o(s) + 52Gi(j/(s))7i(s), y(0) = x. (10.60) i=l Since 7 is Lipschitz continuous, its derivative 7 = ^7(5) is a bounded mea- surable function, defined for a.e. s e [0,5]. Therefore, by Theorems 2.1.1 and 2.1.3 the Cauchy problem (10.60) has a unique Caratheodory solution s i—► ?/(s,7). Following [16] we now introduce Definition 10.3 (Trajectory determined by a graph-completion). Let £/(-,7) be the unique Caratheodory solution of (10.60). Then the (possibly multivalued) function t = {?/(s,7); 7o(s) = t} (10.61) is called the generalized trajectory of (10.50) determined by the graph- completion 7 of u. Remark 10.4. It can be shown that the trajectory #(-,7) depends on the path 7 itself, but not on the way it is parametrized. In particular, let 7 : [0, 5] [0, T] x be another graph-completion of и such that 7(0(s)) = 7(3) s e [0,5] for some absolutely continuous, strictly increasing ф : [0,5] [0,5]. Then the generalized trajectories ^(-,7) and #(-,7) coincide. Example 10.6 (continued). For the discontinuous function и in (10.53), consider the graph-completion 7 : [0,4] i-> [0, 2] x IR2 defined as
252 10 Impulsive Control Systems (s, 0, 0) (1, 0, s - 1) (1, s-2, 1) Дз-3, 1, 1) if t € [0, 1], if <€[1,2], if t € [2,3], if t € [3,4]. (10.62) The generalized trajectory t h-» x(t, 7) is *(<,?) = f (0,0) 1(1,0) if if *e[o, i][, t e]i, 2]. (10.63) Notice that this coincides with (10.56) for all t e [0, 2], t 1, while ж(1,7) is multivalued. Observe that the curve 7 in this case is precisely the limit of the graphs of the approximating functions u^. A different graph-completion, following the construction in Lemma 10.3.1, is achieved by bridging the jump at time r = 1 with one single straight segment. This yields the path 7(s) = < (s, 0, 0) (1, s — 1, s- 1) (s-1, 1, 1) if t € [0, 1], if <€[1,2], if < e [2,3]. (10.64) The corresponding trajectory of (10.50) is given by (0,0) if ~ ((1, 1/2) if < € [0. 1], <€ [1,2]. (10.65) Using graph-completions, trajectories of the impulsive system (10.50) can be computed by solving a Cauchy problem for the O.D.E. (10.60). For detailed results on the dependence of generalized solutions on the path 7, and on approximations with smooth control functions, we refer to [16]. 10.4 Systems with quadratic impulses We consider here a control system where the right hand side contains also the square of the derivative of the control function: m m x = f(x) + '^gl(x)ul + У2 hij(x) Uitij . (10.66) i=l i,j=l Here the state of the system is described by the variable x € IRn, while u(t) G IRm is the value of control. The upper dot denotes a derivative w.r.t. time. We assume that the functions /, and hij = hji are at least twice continuously differentiable and have sub-linear growth, so that m m l/(x)| + £ |Pi(x)| + £ IM*)I < c(1 + И). i=l i»J=l
10.4 Systems with quadratic impulses 253 We remark that the case where these vector fields depend also on time and on the control и can be easily rewritten in the form (10.66). Indeed, if m m x = f(t,x, u) 4- и) щ + hij(t, x, u) iiiUj , t=i ij=i it suffices to introduce the additional state variables xq = t and xn+j = Uj, with equations (10.24). This yields a new control system of the form (10.66), on the extended state space x 6 Notice that now we can no longer use discontinuous functions as controls. Indeed, if some components of и have a jump at some time t, the right hand side of (10.66) would formally contain the square of a Dirac delta distribution, which is not bounded. Example 10.7 Consider a ring sliding without friction along a rotating bar, as in fig. 10.2. Assume that we want to assign the angle 0 = v(t) as a discon- tinuous function of time, say f в~ if 0 < f < 1, - |0+ if i < t < 2. (10.67) Consider any sequence of smooth approximations t ^(t), with \\0y —0||ц 0 as z/ —> oo. The radial coordinate satisfies r = g cosv + rv2. Hence, if r(l—) > 0, as у —> oo we have and hence r(t, Vy) —* oo, r(t,Vy) —> oo for every t > 1. As admissible controls, we shall thus consider a family of absolutely con- tinuous functions ti(-) with derivative in L2. For example L( = [и : [О, T] w IRm ; и absolutely continuous, /* |ti(£)|2 dt < k\ . Jo J (10.68) Given the initial condition х(0) = ^, (10.69) a natural problem is to describe the set of all possible trajectories. The main goal of the following analysis is to provide a characterization of the closure of this set of trajectories, in terms of an auxiliary differential inclusion.
254 10 Impulsive Control Systems It will be convenient to work in an extended state space, and use the variable x = [ 0 | e lR1+n. For a given x. consider the set \rx where co(S) denotes the closed convex hull of a set S C JR1-1"71. Notice that у i—> F(y) is a convex, compact valued multifunction on IR1+n, Lipschitz continuous w.r.t. the Hausdorff metric, see Section A.7 and [5]. For a given interval [0, S'], the set of trajectories of the differential inclusion € F(£(s)), as ^(°) = Qt) (10.71) is a non-empty, closed, bounded subset of C([0, S]; IR1+n). Consider one par- ticular solution, say s i—► x(s) = defined for s e [0, S']- Assume that T = Xq(S) > 0. Since the map s a?o(s) is non-decreasing, it admits a generalized inverse $ = $(£) iff xo(s)=t. (10.72) Indeed, for all but countably many times t e [0, T] there exists a unique value of the parameter s such that the identity on the right of (10.72) holds. We can thus define a corresponding trajectory t i—► x(t) = x(s(t)) € IRn. (10.73) This map is well defined for almost all times t 6 [0,T]. To establish a connection between the original control system (10.66) and the differential inclusion (10.71), consider first a smooth control function u(-). Define a reparametrized time variable by setting (10.74) Notice that the map t s(t) is strictly increasing. The inverse map s i-> t($) is uniformly Lipschitz continuous and satisfies dt ds
10.4 Systems with quadratic impulses 255 Let now x : [0, T] f-* IRn be a solution of (10.66) corresponding to the smooth control и : [0, T] IRm. We claim that the map s i-> f(s) = I Л \ 1 is a \ x\f\s)) / solution to the differential inclusion (10.71). Indeed, setting (10.75) / «o(s) \ \/(x(s))vg(s) + E™ i S«(a:(s))«o(s)vi(s) + £™=1 hij(x(s))vj(s)vj(s)/ (10.76) Hence (10.71) holds, because by (10.75) m ZX?<s) = i. i=0 The following theorem shows that every solution of the differential inclu- sion (10.71) can be approximated by smooth solutions of the original control system (10.66). Theorem 10.4.1. (Reduction of a quadratic impulsive system to a differential inclusion). Let the vector fields f,gi,hij in (10.66) be Lipschitz continuous. Let x : [0, S’] i—> H14~n be a solution to the multivalued Cauchy problem at (10.70)-(10.71). Let the first component satisfy xq(S) = T > 0. Then there exists a sequence of smooth control functions и” : [0, T] ь-> IRm such that the corresponding solutions s i—> ^(s) = r(S) \ of the equations (10.75)-(10.76) converge to the map s > x(s) uniformly on [0, S]. Moreover, defining the function x(if) = as in (10.73), one has lim I \x(t) - xu(t)| dt = 0. (10.77) p-^oo Jo Proof. By assumptions, the extended vector fields 9i =
256 10 Impulsive Control Systems are Lipschitz continuous. Consider the set of trajectories of the control system d rn m £(«) = /vo + 52 ’-Wi + 52 ViV3 ’ i=l M=1 (10.78) where the controls Vi satisfy the constraints ^o(s) C [0,1], m 52 = 1 s € [0, S']. (10.79) According to Theorem 3.4.2, the set of trajectories s i-> £($) = . ,^n)(s) of (10.78)-(10.79) is dense on the set of solutions to the differential inclusion defined by (10.70)-( 10.71). Hence there exists a sequence of control functions s i—> v%s) = (l?q, ..., v^) (s), v > 1, such that the corresponding solutions s tp(s) of (10.78) converge to £(•) uniformly for s e [0, S']. In particular, this implies the convergence of the first components: ^(S) = / [^«</s^r0(S) = T. Jo (10.80) We now observe that the “input-output map” v(-) •—> i(-,^) from controls to trajectories is uniformly continuous as a map from L1 ([0, S]; R1+m) into C([0,S]; IR.1 ^n). By slightly modifying the controls vy in L1, we can replace the sequence vy by a new sequence of smooth control functions wy : [0, S] h-> IR1+m with the following properties: Wq(s) > 0 for all s e [0, S], v > 1, (10.81) Cs / [wo (s)]2 cLs = T for all v > 1, (10.82) Jo lim / |wI/(s) — vl/(s)| ds = 0. (10.83) Jo This implies the uniform convergence £m ||i(-,w1')-i(-)||C((0,S];]R—) = 0- (10.84) By (10.81), for each v > 1 the map s^^(s)= / [w£(s)]2ds Jo is strictly increasing. Therefore it has a smooth inverse s = sy(t) We now define the sequence of smooth control functions uy : [0, T] >—> IR™ by setting
10.5 Optimization problems for commutative impulsive systems 257 «ИО) . ®оИО (10.85) By construction, the solutions t xy(t\ uy) of the original system (10.66) for the controls uy coincide with the trajectories t i—> , xy)(sy (t)), where xy = (л-q, ..., rr^) is the solution of (10.78) with control wy = To prove the last statement in the theorem, define the increasing functions *00 = [ [w(r)]2 Jo t^s)= [S[w^r)]2dr Jo dr, and let t и-> s(t), t >—> sy(t) be their respective inverses. Notice that each sy is smooth. Moreover we have d . . ds lim f |s(t) - s"^)! dt = lim f ' \t(s) p-oo Jo < 1, ty(s)\ds = 0. (10.86) (10.87) Using (10.86), we obtain the estimate f \x(t) — #P(£)| dt = /* |a:(s(t)) — j:Z7(.s(/))| dt + о Jo [ \xy(s(f))-xy(sy(t))\dt S J° fT \x[s)-x^s)\ds + C- / |s(t)-Z(0|dt. Jo Here the constant C denotes an upper bound for the derivative w.r.t. s, for example (10.88) where the supremum is taken over a compact set containing the graphs of all functions tp(-). By (10.84) and (10.87), the right hand side of (10.88) vanishes in the limit у oo. This completes the proof of the theorem. 10.5 Optimization problems for commutative impulsive systems This last section is concerned with optimization problems for the impulsive system (10.25), assuming that the vector fields F, Gi on 1RA satisfy the
258 10 Impulsive Control Systems sub-linear growth condition (♦) and the commutativity assumptions (10.27). Assume that the control values u(t) are constrained within a compact set U C IRm, and let the initial values x(0)=x, г/(0)=й (10.89) be assigned. The family of admissible controls will be denoted by U = {и : [0, T] U, и measurable, tt(O) = й, T > 0} . Given a continuous cost function J = J(t,x) and a closed target set S C IR x IR77, we consider an optimization problem in Mayer form, with a terminal cost, terminal constrains and variable terminal time: min J(T, x(T, u)), subject to (T, x(T))eS. (10.90) Fix an arbitrary value if = ..., G IR m. According to Theorem 10.2.1, the trajectory t x(t,u) is then provided by the representation for- mula (10.40), where £ is the solution to the Cauchy problem (10.36)-(10.37). As we now show, the minimization problem (10.90) can be reformulated as a standard Mayer problem for the auxiliary variable £ G IRA . We begin by defining the functions v ' ( +oo if (t,x) f S. = min Js (t, (. (10.91) \ i=l / Observe that the minimum in (10.91) is always attained (unless it equals 4-oc), because Js is lower semicontinuous and U is compact. Next, we consider the optimization problem min J*(T, £(!») (10.92) u£U for the control system (10.36)-(10.37). As a consequence of the representation formula (10.40), we obtain Theorem 10.5.1. Let the vector fields F,Gi be continuously differentiable and let the commutativity assumptions (10.27) hold. Assume that there ex- ists at least one admissible trajectory of (10.60) that satisfies the constraint (T, x(T\u)) G S. Then, the following are equivalent: (i) The control function uEU is optimal for the Mayer problem (10.90), in connection with the impulsive system (10.60). (ii) The control u(-) is optimal for the optimization problem (10.92), in con- nection with the control system (10.36)-(10.37). Moreover, the assignment w — u(T) yields the minimum value of J$(T. •) in (10.91).
10.5 Optimization problems for commutative impulsive systems 259 If the function J* is sufficiently regular, the new optimization problem for the variable £ can now be studied by standard techniques. For the case of a swing, described in Example 10.2, optimization problems were discussed in [61]. Problems 10.1. A girl riding a swing is standing when the angle of the swing with a ver- tical line is increasing and kneeling down in the opposite case. Hence if 0 < rmm < r < rmax then и = rmax when 9 • 9 < 0 and и = rmin when 9 • 9 > 0. Assume that she starts with zero velocity and an angle of тг/4 with respect to the vertical position. (a) Compute the position of the swing in time for 0 < t < тг. (b) Is the swing going to stop as t —► oo? Hint: Refer to Example 10.2 and compute the control r(t). For simplicity, use the normalization g = 1 in (10.10). 10.2. Consider a mechanical system consisting of two point masses on a vertical x-y plane (see figure 10.2, right). The mass at A is constrained to move vertically on the i/-axis. The mass at В is connected to A by a rigid bar of length p. The system is thus described in terms of two lagrangian variables: the height h and the angle 9. Call the two masses, and g the gravity acceleration. Assuming that h(t) = u(t) is assigned as a function of time (by frictionless constraints), write the equation determining the motion of the remaining free variable 9(t}. Is this sytem fit for jumps ? 10.3. For i = 1,2, consider functions gi(t, x,u) from IR x IRn x IRm into IRn. Inserting additional variables as in (10.24), define the corresponding func- tions C?i, G2 : IRA •-> RN, with N = 1 + n + m. Find conditions on <71, which guarantee that the Lie Bracket [Gi,G2] : IRA >—► IRA vanishes identically. 10.4. Consider the impulsive control system: ±1 = #1 + axi ui X2 = 1 + (—2^2 + ж2) щ + (я?? 4- ж2) W2, Find conditions on a for the commutativity condition (10.27) to hold. 10.5. Consider the impulsive control system: ±1 = й1 + X2 ±2 = Щ + U2)
260 10 Impulsive Control Systems and the control given by (10.53) for t 6 [0, 2]. Define the graph completion: <^(s) = < (S,(0,0)) (l,((s-l),(S («-1, (1,1)) 0 < s < 1 1) + a(s — l)(s — 2))) 1 <s<2 . 2<s<3 For every a, compute the corresponding generalized trajectory xa. 10.6. Consider a bead sliding along a bar as in Example 10.3 and consider the control function: u(t) = (l-t)°. (a) Determine the values of a for which a solution is defined; (b) Compute the trajectory for a = 3/4. 10.7. Consider the impulsive control system (10.10) with u(f) 6 U = [1,3]. Let the initial conditions be 0 = 0, ш = — 1. Consider the optimization problem max cj(T) subject to 0(T) = 0 and 0 < T < 2тг. Find the optimal control. 10.8. Give a detailed proof of Remark 10.4. In other words, show that the trajectory determined by a graph completion 7(-) is independent of the way the path 7 is parametrized. 10.9. Write a representation formula for the solution of the Cauchy problem (10.25)-(10.26) assuming that all vector fields commute, i.e. [F, Gi] = 0, [Gi, Gj] = 0 for all z, j = 1,..., m. 10.10. Consider the impulsive control system (10.25), assuming that FGZ are smooth and globally Lipschitz continuous, so that rn |F(x) - F(y)| + £ |СДх) - Gi(y)| < L |x - y\. i=l Consider a set of uniformly Lipschitz continuous control functions A = {a : [0,T] IR”1; lu«W - «i(s)| < C - s| for all t,s € [0, T]} . 1=1 Call t x(t, u) the solution of the Cauchy problem (10.25)-(10.26). Prove that the map tz(-) (—> ж(-,и) is Lipschitz continuous w.r.t. the C° norm: max |rr(^,u) — rr(t,v)| < C• maX] |u(t)—v(t)| for all u,v^A. Hint: Using an equivalent norm
10.5 Optimization problems for commutative impulsive systems 261 lk(-)llt = check that the Picard operator о [f(z(s)) + Gi(#(s)) йДз)] ds i=l satisfies the assumption of the Contraction Mapping Theorem A.2.1. given in the Appendix. Use an integration by parts to prove the Lipschitz con- tinuity of Ф w.r.t. u.

Appendices We collect here various definitions and basic results of mathematical analysis, which constitute the main background material used throughout the book. A.l Normed spaces Let X be a vector space. A norm || • || on X is a mapping X i—► IR+ with the following properties. For every vectors x,y € X and every real number A e IR one has • Non degeneracy: ||.t|| > 0 whenever x 0, • Homogeneity: ||Aj:|| = |A| ||x||, • Sub-additivity: ||rr + y\\ < ||x|| + \\y\\. Roughly speaking, the norm ||j:|| provides a measure of how big is the vector x G X. A subset А С X is convex if, given any two points x,x' € A and A 6 [0,1], the convex combination A.r+(1 — X)x' also lies in A. From the above properties it follows that, for every r > 0, the ball Br = {x e X \ ||ж|| < r} centered at the origin with radius r is convex. Indeed, if ||z|| < r, ||t/|| < r and 0 € [0,1], then the norm of the convex combination satisfies ||0ar+(l-0)y|| < ||0ж|| + Н(1-ЭД = 6*lkll + (1-0)lly|l < 0r+(l-0)r = r. We recall that a sequence of points is called a Cauchy sequence if, for every e > 0 there exists an integer N large enough so that ||xm — яп|| < 8 whenever m,n > N .
264 A Appendices Intuitively, this condition means that the elements of the sequence are getting closer and closer to each other. We say that the normed space X is complete if every Cauchy sequence converges to some limit point in X. A complete normed space is called a Banach space. Example A.l. Let X be the space of all polynomial functions defined on the interval [0,1], with norm IM - 11.(01 This is a normed space but not a Banach space. To see this, consider the sequence of polynomials Pn(o = E Ь k=0 This is a Cauchy sequence, but it has no limit in the space X. Indeed, as n —> oo, the polynomials pn converge to the function et uniformly on the interval [0,1]. However, this exponential function is not a polynomial and does not lie in the space X. Therefore, our space X is not complete. Several well known Banach spaces will be used throughout this book. We recall here the main examples. • The finite dimensional space IR” with the Euclidean norm M - y.r'f +.T^ + ••• + .7(2. • The space C°([a, 6]) of all continuous functions on the closed interval [a, 6], with norm ll/llc" = max |/(t)|. tG[a,b] • The space C^Qa, 6[) of all continuously differentiable functions on the open interval ]a, b[, with norm ll/lld = sup |/(t)| + sup 0)1. a<t<b a<t<ib • The space Ьх([а, 6]) of Lebesgue integrable functions on the interval [a, 6], with norm J a • The space L°°([a, 6]) of Lebesgue measurable, essentially bounded functions on the interval [a, b], with norm
A.2 Banach’s contraction mapping theorem 265 ||/||l~ = ess- sup |/(t)| = inf { r > 0; meas{t e [a, 6]; | f(t)| > r} = 0 t€[a,6] Given a subset J? C lRm, we say that a map f : J? i—► IRn is Lipschitz continuous if there exists a constant L such that l/(®) - /(y)l < L\x- y| (A.2) Vt, у 6 SI. The space Lip(f?; IRn) of all these Lipschitz continuous mappings is a Banach space with norm II/IIlip = sup |/(x)| + sup /(y)l x£.Q 1*^ У\ A.2 Banach’s contraction mapping theorem One of the main interests of mathematicians is to solve equations. In many cases, an explicit formula for the solution cannot be achieved. Still, one might be able to prove that a unique solution exists, depending continuously on the parameters that describe the problem. According to Banach’s theorem, this is possible if the equation can be written in the form z = <Z>(z,A) (A.3) and the map x i—> Ф(А, x) is a strict contraction for each given value of the pa- rameter A. Indeed, the fixed point of this mapping can be found by a standard iterative procedure. Theorem A.2.1. (Contraction Mapping Theorem). Let X be a Banach space, Л a metric space, and let Ф : А x X i—> X be a continuous mapping such that, for some к < 1, ||Ф(А,Ж) - Ф(А, г/)|| < к ||х - у\\ ЧХ,х,у. (А.4) Then, for each А € Л there exists a unique fixed point x(A) € X such that х(А)=Ф(А,х(А)). (A.5) The map A—>x(A) is continuous. Moreover, for any A 6 А, у € X one has ||у-х(А)||<-1- ||y-^(A,y)||. (A.6) 1 Av
266 A Appendices Fig. A.l. The approximating sequence, obtained by iteration. Proof. Fix any point у € X. For each fixed A € A, consider the sequence (see figure A.l) Уо=У, У1 = Ф(А,?/о), ", y^+i = Ф(Х,уО, "• By induction, for every 2/ > 0 one checks that b+i - Ы < lli/1 - 2/01| = ||2/ - Ф(Х, y)||. (A.7) Since к < 1, the sequence y„ is Cauchy and converges to some limit point, which we call x(A). By the continuity of Ф we now have x(X) = lim y„ = lim Ф(Х,уи-х) = Ф (A, lim y^-x) = Ф(А,х(А)), P —>OO P—>ОО \ P —>00 / hence (A.5) holds. The uniqueness of a?(A) is proved observing that, if = Ф(А,я?1), x2 = Ф(Х,х2), by (A.4) it follows ||tfi -Я2Ц = ||Ф(А,Х1) - Ф(А, ж2)|| < к||Я1 -#2||- The assumption к < 1 thus implies aq = x2. Next, observe that (A.7) yields V V 1 11^+1 —2/11 <52ll%+i — 2/jll < Пз/-ф(А’2/)11 ||y-#(A,2/)ll • j=0 j=0 (A.8) Letting i/—>00 in (A.8) we obtain (A.6). To show that the fixed point x depends continuously on A, let (An)n>i be a sequence of parameters converging to A*. Using (A.6) with A = An, у = x(A*), we obtain ||x(A*) - x(An)|| < -T_ ||x(A*) - Ф(АП, x(A*))|| = 1 — К = гЧ ИA*’ *(0 - ф(А- A*))II • (A-9) 1 rv Since Ф is continuous also w.r.t. the variable A, the right hand side of (A.9) tends to zero as n^oc. Hence rr(An)—*x(A*).
А.З Brouwer’s fixed point theorem 267 A.3 Brouwer’s fixed point theorem This section contains a short proof of the classical fixed point theorem of Brouwer: a continuous mapping from a closed ball В C IR71 into itself has at least one fixed point. A corollary of this result plays a key role in the proof of the Pontryagin Maximum Principle. Brouwer’s theorem will first be proved in case of a continuously differen- tiable map. Later, the regularity assumption will be removed by an approxi- mation argument. With this in mind, we recall here a standard mollification technique, to approximate an arbitrary continuous function by a smooth one. As usual, we say that a function f is smooth, or equivalently f G C°°, if f is к times continuously differentiable for every integer к > 1. Lemma A.3.1. Let f : IR71 > IR/71, be a continuous map and consider a C°° function ф : IR71 i—► [0,1] with compact support, such that For each e > 0 define the mollified approximation dy. Then each f£ is smooth. As e —> 0, the functions f£ converge to f uniformly on bounded sets. We are now ready to prove the main result of this section. Here and in the sequel (•, •) denotes the inner product of two vectors in IR/1. For simplicity, we prove the theorem for the unit ball: the proof for any ball is obtained in the same way. Theorem A.3.2. (Brouwer). Let f be a continuous map from the closed unit ball В C IRn into itself. Then there exists a point x* G В such that X* = f(x*) (A.10) Proof. 1. We first prove the theorem under the additional assumption that f G C1 is a continuously differentiable mapping defined on the whole space IR77, taking values strictly within the interior of the ball B. Assuming that no fixed point of f exists, we will derive a contradiction. For any x e IR/1, consider the ray originating from f(x) and containing x, as in figure A.2. By assumptions, x f(x) and |/(z)| < 1. Therefore, the ray crosses the unit sphere at exactly one point, which we call g(x). In other words, з(х) = /(x) + А(ж)(ж - /(a:)), where A = A(or) > 0 is implicitly defined by the equation
268 A Appendices Fig. A.2. A continuous map x g(x) whose values lie on the surface of the unit ball. 1 = lffk)|2 = l/(z)|2 + 2A(/(z), ;г-/(я)> + Л2|я-/(ге)|2. (A.11) The assumption |/(#)| < 1 for all x implies that (A.11) has two distinct roots, with opposite signs. Therefore, the map x X(x) is continuously dif- ferentiable, and hence g G C1 as well. Moreover, we notice that |a?| = 1 = ж. Fig. A.3. The map and the wrould-be polynomial P. 2. We now consider a family of maps, depending on an additional parameter t e [0,1]. V’(t)k) = (1 — t)x + tg(x). (A.12) Clearly, is the identity mapping, while = g. Since f takes values strictly inside the unit ball B, for each t G [0,1[ the construction of g implies И < 1 => |</5(t)Cr)| < 1, И = 1 =► <f^(x)=X, kl >1 => kt)WI>i. (A.13)
А.З Brouwer’s fixed point theorem 269 3. Let L be a Lipschitz constant for g, so that \g(x) - g(y)\ < L\x - y\ 4x,y € B, and choose т G]0,1[ such that t G [0, t] Lt 1 1-^2* (A.14) We claim that, for t G [0,r] the map is one-to-one C1 mapping from В onto itself, whose inverse is also C1. Indeed, by the first condition in (A.13) each maps В into itself. To show that is one-to-one and onto, let any point у G В be given. By (A.14), for t G [0,t] the map X » Ф^х) = p(x) (A.15) is Lipschitz continuous with constant < |. Hence, by the Contraction Mapping Theorem A.2.1, it has a unique fixed point x = x(y). The third condition in (A. 13) implies that |я?(?/)| < 1. From (A. 12) and (A. 15) it now follows <£(f)(z) = (1 - t)x + tg(x) = y, proving that is bijection. The smoothness of the inverse function ip^ is proved by checking that the Jacobian matrix V<^(t)(a?) = (1 - f)I — tVg(x') is invertible when t G [0, т]. 4. For t G [0,1], consider the function P(t) = / det(V(£>(t)(a:)) dx = / det ((1 — t)I + tS7g(x)) dx. (A.16) Jb J в ' ' For each x G B, the determinant of an n x n matrix whose entries depend linearly on a parameter t is a polynomial of degree < n. Therefore, the integral P(t) itself must be a polynomial of degree < n. This leads to a contradiction. Indeed, for t G [0, t], the transformation is a C1 one-to-one map of В C IRn onto itself. By the formula for a change of variables in a multiple integral, we compute P(t) = / det(V^(t)(x))drr J в = vol (^(t)(B)) = vol(B). For t > 0 small, the polynomial P(t) is thus constantly equal to the volume of the unit ball В C IRn. On the other hand, since |<P(i)(a?) | = |^(x)| = 1 for all x, when t = 1 we find P(l) = j det(V</(x)) dx = I Odx — 0. J в J в
270 A Appendices No polynomial having such behavior exists (see figure A.3). This contradiction proves the theorem, under the additional assumption that f is a C1 mapping defined on the whole space lRn. 5. To prove the theorem in the general case where f : В i—► В is only assumed to be continuous, we first extend f to a map f defined on the entire space IRn, by setting 7(x) = /(тг(а;)). Here 7Г : IR” >—> В is the perpendicular projection on the unit ball: if ' [ ж/|ж| if |.r| > 1. Next, we construct the approximations A(x) = (l-e)£ dy' (A17) J1R7' £ \ £ / where ф is a smooth mollifier, as in Lemma A.3.1. For each e > 0, the smooth function f£ maps the entire space IRn strictly into the interior of the unit ball В because of the additional factor 1 — e present in (A. 17). By the previous arguments, this map has at least one fixed point, say Xe = /(ж£) e В . By the compactness of the closed ball B, we can find a subsequence eu —> 0+ such that the corresponding fixed points converge: x£v x*. Since f£ f uniformly on B, this implies /(®*) = lim f£ (x£ ) = lim ix£u = x*, eu—*0 е„ —>0 proving the theorem in the general case. The above theorem has several far-reaching extensions. A few are described below. Corollary A.3.3. Let v:B» IR” be a continuous vector field defined on the n-dimensional unit ball which points outward at the boundary: (v(x),x) > 0 whenever |x'| = 1. (A.18) Then there exists a point x* e В with v(x*) — 0. From a geometric point of view (see Figure A.4), the condition (A. 18) means that the angle 0 between a unit vector x on the surface of the ball В and its image f(x) satisfies |0| < 7r/2. In particular, this implies that |x + v(x)| > 1 whenever |ar| = 1.
А.З Brouwer’s fixed point theorem 271 Fig. A.4. The outward-pointing condition. Proof. Consider the perpendicular projection 7Г : IRn—>B, and define f{x) = 7t(z) - t;(7r(a;)). Since v is bounded, we can assume |v(a:)| < M for all x € B. Hence f maps the ball Вм+i centered at the origin with radius M + 1 into itself. By Brouwer’s theorem, f has a fixed point ж*. If |ж*| > 1, then (A.18) implies (x*, x*) = (тг(а;*) — -и(тг(а;*)), x*) < (г—г, < (я*,#*). \М / This contradiction implies |ж*| < 1, hence v(x*) = x* — /(/) = 0. Next, we show that in Brouwer’s theorem the unit ball can be replaced by any bounded, closed convex set К C IRn. Corollary A.3.4. Let К be any compact convex subset of IRn. Then every continuous map f : К i—» К has a fixed point. Proof. Choose a > 0 so large that К is contained inside the closed ball aB centered at the origin with radius a. Let тг : НС h-► К be the perpendicular projection onto К. This is characterized by the properties 7г(гг) € К, k(x) — ж| = min \y — x| уек for every x E IRn. By Brouwer’s theorem, the composed map g : aB i—► aB defined by g(x) = /(тг(д:)) has a fixed point x*. Clearly, x* E K, hence In our last application, we consider a mapping f from a compact set К into lRn. Given a point wq 6 K, from information on how f behaves on the boundary Ж we deduce the existence of a solution to the equation = w0. The key assumption here requires that each boundary point x E dK remains sufficiently close to its image.
272 A Appendices Corollary A.3.5. Let К be a compact, convex neighborhood of a point w G IRn, with boundary dK. Assume that the continuous map f :K^ IRn satisfies \f(y)-y\ < |y-w| forallyedK. (A.19) Then there exists x$ & К such that /(^o) — w. Fig. A.5. A map whose range contains the origin. Proof. By possibly performing a translation of coordinates, we can assume w = 0. Choose a constant a > 0 so large that К is contained in the closed ball aB centered at the origin with radius a. Let ir : IRn i—> К be the radial projection on К. This projection (not to be confused with the perpendicular projection!, see figure A.6) is defined by setting 7r(x) = x if x G К, while 7г(х) = Xx, Л = max{A' > 0; X'x G K} in the case x^K. We claim that the continuous function v(x) = /(тг(х)) x G aBn points outward at each point x on the boundary of the ball aBn. Indeed, assume |x| = a and у = тг(х) = Xx for some A G (0,1]. Since у G dK, from (A.19) it follows X{v(x),x) = X{f(Xx),x) = (f(y),y) = (y,y) + -y,y} > Ы2 -l№) - y\ • li/l > 0, proving our claim. By Corollary A.3.3, there exists a point x* such that v(x*) = 0. The result thus holds with Xq = 7r(x*).
A.4 A compactness theorem 273 Fig. A.6. The perpendicular projection and the radial projection on the set K. A.4 A compactness theorem This section contains a simple version of Ascoli’s compactness theorem, which is used in several existence proofs. We say that a sequence /„(•) is uniformly Lipschitz continuous if all f„ are Lipschitz continuous with the same constant L, i.e. |/p(t) — /p(s)l < L\t ~ sl for s G [a, b], v > 1. Theorem A.4.1. Let (fv)v>\ be a bounded, uniformly Lipschitz continuous sequence of functions from a compact interval [a,b] into IRn. Then there ex- ists a subsequence fu> converging to some Lipschitz continuous function f, uniformly on [a, b]. Proof 1. Let be a dense sequence of points in [a, b]. From the original sequence /p, we extract a subsequence (/i,m)m>i such that /1,тп(£1) con- verges. This is possible because f„ is bounded. Assume now that a sequence (A,m)m>i has been constructed, which converges at the points ii,--- From this sequence we can extract from this further subsequence (A+i,m)m>i which converges also at the point tk+i- By induction on к, we thus obtain a double sequence of functions fk,m> k,m > 1, such that each subsequence Лд, A,2, А,з> • • • converges at the points <1,^2, • • •,tk- 2. By the previous step, the diagonal sequence of functions /1,1,/2,2, /3,3, • • • converges to some value f(tj) at each point tj of our sequence. Moreover, the uniform Lipschitz continuity of all functions f„ implies \f(ti) - < L\ti - tj\ for all ij. Therefore, the map f can be uniquely extended by continuity to a Lipschitz continuous map, still called /, defined on the entire interval [a, b]. 3. We claim that fmim—*f uniformly on [u,b]. Indeed, fix any e > 0. Choose p so large that
274 A Appendices ж (a-2°) 2=1 where B(£,p) = [t — p,t + p\ denotes the ball centered at t with radius p. Notice that (A.20) is possible because the points tj are dense on [a, b\. Choose N so large that f for all m > N, о For every m > N and every t e [a, b], since t e B(ti,e/(3L)) for some i < p, we have the estimate |/тп,т(*) - /(t)| < |/m,m(*) “ /m,m(^)| + |/m,m(^) “ /(Ml + |/(M “ /(01 This establishes the uniform convergence of the subsequence fm,m, completing the proof. A.5 Review of Lebesgue measure theory This section collects some of the main definitions and results of Lebesgue measure theory. For the proofs, see Folland [F] or Rudin [Ru 2]. As a preliminary, we recall that a function / : [a, b] i—► IRn is continuous if at each point r there holds f(r) = lim/(f) - We say that the restriction of / to a subset J C [a, b] is continuous if, for every т 6 J, one has /(t) = t /(t). A function / : [a, b] *—> IR, is lower sernicontinuous if at every point r it satifies /(r) < liminf /(f) • These concepts are illustrated in figure A.7. 1. (Measurable sets) A bounded set A C IRn is measurable iff, for every e > 0, there exists V open and К compact such that К С А С V and meas(V\K) < e. Here V \ К = {ж; x e V, x £ K} denotes the set- theoretic difference of the two sets. 2. (Measurable functions) A function / : [a, b] IRn is Lebesgue measur- able if, for every open set V C IRn, the preimage f~} (V) = {t 6 [a, b]; /(£) G
A.5 Review of Lebesgue measure theory 275 Fig. A.7. The function f is lower semicontinuous. Moreover, it is continuous re- stricted to the compact set J. У} is a Lebesgue measurable set. Every continuous function and every lower semicontinuous function is measurable. 3. (Properties valid almost everywhere) We say that a property P holds almost everywhere (a.e.) on a set A if there exists a null set N such that meas(A) = 0 and every point of A\N has the property P. 4. (Pointwise converging sequences) Consider a sequence (fy)u>i of mea- surable functions. If one has the pointwise convergence /Дх) —> f(x) for a.e. point ж, then the limit function f is measurable. 5. (Integrable functions) If f is measurable and |/(t)| < </>(t) for all t, for some integrable scalar function </>, then f itself is integrable. Functions /, f which differ only on a set of measure zero are identified. With this equivalence relation, the space of all Lebesgue integrable functions f : [a, b] i—> IRn is written L^a, b\; IRn). This is a Banach space with norm 6. (Lebesgue Dominated Convergence Theorem) Consider a sequence of integrable functions fy which converge to f at a.e. point in [a, b]. Moreover, assume that there exists an integrable function such that and |/i/(t)| < for all I/ > 1 and t G [a, b\. Then >b rb f(t)dt= lim / fy(t)dt. > J a 7. (Severini - Egoroff’s Theorem) Let fy : [a, 6] ]Rn be a sequence of measurable functions which converges a.e. to f. Then, for every e > 0, there exists a set J such that meas(J) < в and fy^f uniformly on [a, 6]\J.
276 A Appendices 8. (Characterization of measurable functions) For any function f : [a, b] i—> lRn, the following statements are equivalent: (i) /is measurable. (ii) For every e > 0, there exists a compact set J C [a, b] with meas([a,6] \ J) < £, and such that the restriction of / to J is continuous. (iii) There exists a sequence of disjoint compact subsets Jk C [a, 6] with meas ( [a, b] \ Jk I = 0 \ fc=i / (A.21) and such that the restriction of / to each set Jk is continuous. (iv) There exists a sequence of disjoint compact subsets Jk C [a, b] satisfying (A.21), such that the restriction of / to each set Jk is measurable. 9. (Absolutely continuous functions) A function / : \a,b] »—> III7' is ab- solutely continuous if, for every s > 0, one can choose <5 > 0 such that the following holds. If ]$$, i = 1,... N is any finite collection of disjoint intervals with total length 52 “ S<l S’ then 10. (Differentiation vs. integration) If / is absolutely continuous on [a, b], then its derivative /' is defined almost everywhere. Moreover, it satisfies the Fundamental Theorem of Calculus: /W = /(*o)+/ f'(s)ds for all t,toe[a,b]. Every Lipschitz continuous / defined on an interval [a, 6] is absolutely contin- uous. 12. (Lebesgue points) A point т € [a, b] is a Lebesgue point for an integrable function f if When this happens, we also say that / is quasi-continuous at r. In this case, the integral function
A.6 Differentiability of Lipschitz continuous functions 277 F(t) = f f(s) ds J a is differentiable at the point r and one has the identity F'(r) = f(r)- We say that a point £ is a Lebesgue point for the set A C IR if lim — meas(A П \t — e, t + el) = 1. 2e 13. (Lebesgue Theorem) Let f : [a, b] i—> IRn be integrable. Then almost every point t E [a, b] is a Lebesgue point for f. If A is a measurable set, then almost every t E A is a Lebesgue point of A. 14. (Lyapunov’s Theorem on convex combinations) Let •• • , € Lx([a, b];JRn) be integrable vector valued functions. Let 0or" , • [a, b] «—> [0,1] be measurable weight functions such that = f°r everY Then there exist a partition of [a, b] into disjoint measurable subsets Jo, • • •, J к such that /•b i к \ к л / b>w/(i)wU=£ (A-22) \i=0 / i=0Jji Observe that, for each t E [a, 6], the integrand on the left hand side of (A.22) is a convex combination of the vectors ., f^k\t) with coefficients 0O(*),..., Ok(t) € [0,1]. The right hand side can also be interpreted as a convex combination, but where the coefficients are allowed to take only the two values 0 or 1. A.6 Differentiability of Lipschitz continuous functions We recall that a function f : IRn i—► IR™ is differentiable at a point x if there exists a linear map Df(x) : IRn »—> IR™ such that f(x + Л) - /(x) - D/(x) • h = h™ |/i| where | • | indicates the Euclidean norm in Rn. We say that f is locally Lipschitz continuous if, for every compact set К C IRn, there exists a costant Lk > 0 such that |/(s) - /(у) I < LK |x - 7/1 for all X, у e K. It is well known that a Lipschitz continuous function of a single real variable is absolutely continuous, hence differentiable almost everywhere, see 10. in Section A.5. The following theorem shows that the same is true for functions of several variables.
278 A Appendices Theorem A.6.1. (Rademacher) Let f : IRn »—> IRm be locally Lipschitz continuous. Then f is differentiable almost everywhere. Proof. Without loss of generality we may assume that f is Lipschitz contin- uous, with a uniform constant L > 0. Moreover, considering separately each component, it suffices to prove the theorem in the case m = 1. 1. Let v € IRn with |v| = 1. For every x e IRn the map t »-> f(x+tv) of a single real variable is Lipschitz continuous, hence differentiable almost everywhere. Therefore, the directional derivative of f along v, written Dvf(x), exists for a.e. x e IRn. In particular, letting v vary among the standard basis of IRA we see that the vector grad /(ar) = is well defined for a.e. x E RA 2. For every smooth function ф with compact support in Rn, we can write x 4- tv) — f(x) dx. (A.23) Letting t —► 0+ in (A.23), by the dominated convergence theorem (see 6. in Section A.5) we obtain t = [ ^~{х)ф{х)Фх JjR." dxi / (v • grad /(х))ф(х) dx. JlRn Since in the above equality ф is arbitrary, for every v e IRn it follows that Dvf(x) = v - grad f(x) for a.e. x € RA (A.24) 3. Let {гр}^>1 be a countable dense subset of the unit sphere of IRA Define By = {x € IR'1; Dv„f(x) = • grad f(x)}, В = Q By. l/=l By the previous step, one has meas(lRn \ B) < meas(IRn \ By) = 0. (A.25) 17=1 We claim that f is differentiable at every point x € B, and its differential is Df(x) = grad /(x). Indeed, for every unitary vector v E IRn, define:
A.7 Multifunctions 279 Q(x,v,t) = № + ~ /M - V . grad Ц*). Our claim will be proved by showing that ^lim Q(x, v,t) = 0 (A.26) for every x € В uniformly w.r.t. v. Now fix x 6 В and 5 > 0. Choose N so large that, for every unit vector v, one can find z/ = L/(t>) G {1,..., N} such that - ',“l *= 2(1 + ^)L (A.27) Then by (A.27) we get: |Q(x, v,t) - Q(x,v,,, t)| < /(x + tv) - f(x + tv„) t + |(v-Vp) - grad /(x)| < L\v — v„\ + |grad /(x)| |г> - vv\ < (1 + Vn)L|u-u„| < (A.28) Notice that, by definition of В, (A.26) holds for every vu. Hence there exists 6 > 0 such that p |Q(x,v^,t)| < - Vte]0,5], v = (A.29) Consider now w G IRn \ {0}, |w| < 6. Set v = w/|w|, so that w = tv with t G]0, <5]. Using (A.28) and (A.29), we get f (ж + w) - f(x) - w • grad f (a?) |w| = |Q(x,v,t)| < |Q(x,u,t) - Q(x,vt/,i)| + |Q(x,vv,t)| < | + j = e. Since e > 0 was arbitrary the theorem is proved. f{x + tv) - f(x) —-----------------v • grad f(x) A.7 Multifunctions In the following, X is a Banach space with norm || • ||. The distance between a point x and set А С X is defined as the smallest distance between x and points in A, i.e. d{x, A) = inf H# - a||. a€A The open E-neighborhood around the set A is denoted by B(A,e) = {x € X : d(x,A) < e}.
280 A Appendices The Hausdorff distance between two (nonempty) compact sets А, А! С X is defined as df^A, A') = max {d(x, A'),d(xf, A); x G A, x' G A'} . Equivalently, c?h(A, A') can be defined as the infimum of all radii p > 0 such that A is contained in the p-neighborhood around A' while A' is contained in the p-neighborhood around A (see figure A.8). dH(A,A') = infb>0; AcB(A',p) and A'cB(A,p' Fig. A.8. The Hausdorff distance between the two sets A, A' is rnax{p, p'}. A multifunction F from X to Y is a map that associates to each point x G X a set F(x) C Y. We say that F is compact valued if F(x) is a non- empty compact subset of У, for every x G X. The multifunction F is bounded if all its values are contained inside a fixed ball В C Y. A multifunction F with compact values is said to be Hausdorff continuous if lim dH(E(y),F(x)) = 0 for every x G X. y^x We say that the multifunction F : X i—> Y has closed graph if its graph Graph(F) = ((x, у); у G F(x)} is a closed subset of X x Y. This condition means that, whenever x„ —► x, у и —> у and у у е F(xjf) for every и > 1, then we also have the inclusion У € F(x).
A. 7 Multifunctions 281 In the Introduction, we remarked that a general control system can alter- natively be written in the form of a differential inclusion. The next lemma shows that, under a natural assumption, the continuity of the function f in (1.3) implies the continuity of the corresponding multifunction F in (1.5). Lemma A.7.1. Let U be a compact subset o/IRm and let f : Rn x U i—> IRn be continuous. Then the multifunction F defined by F(x) = {f(x,u) : utU] is Hausdorff continuous. Proof. 1. Consider any closed ball В C lRn. On the compact set В xU the continuous function f is uniformly continuous. Therefore, for any e > 0, there exists S£ > 0 such that |ж — x'\<5£, \u — => |/(x, u) — /(з/,г/)| < e (A.30) whenever x, x' € B, u,u' 6 U. 2. We now show that dH(F(x), F(x'))<e (A.31) whenever x,x' e B, \x — x'| < 6£. Indeed, consider any point у 6 F(x). Then у = f(x, u) for some и e U. By (A.30), the point y' = f(x', u) G F(x') satisfies \y' — y\e. Hence F(x) is contained in the ^-neighborhood B(F(x'), s) around the set F(x'). Inverting the roles of x, x', we obtain F(x') C B(F(x),s). This proves (A.31). Since e > 0 and the ball В were arbitrary, the theorem is proved. Given a multifunction 11—» F(t) with non-empty values, a natural problem is to construct a selection, i.e. a single-valued, function f such that f(t) € F(t) for every t. If the sets F(t) are non-empty, then the existence of some selection would follow simply from the axiom of choice in abstract set theory. However, one would like to have a selection with some additional properties, such as continuity or at least measurability. The following analysis will establish the existence of a measurable selection, for a wide class of multifunctions. On IRn, the lexicographical ordering is defined as follows. For any two distinct points x = (a?i,..., xn) and у = (т/i,..., i/n), we write x -< у if one of the following alternatives holds. xi <yi, Zi = У1 and X2 < У2 5 Я1 = У2 , ^2 = У2 , and x3 < y3 , ж i — y\,..., xn—i — Уп—i, and xn < yn .
282 A Appendices Notice the analogy with the ordering of words in a dictionary. We now show that every compact set К C IRn has one first point £ = (aril,... , £n), w.r.t. the lexicographical order. Indeed, its coordinates can be determined inductively as follows: Let {ei, ©2, ... , en} be the standard basis of unit vectors in IRn. Define £1 — min (ei , x) , K\ = {# e К , (ei, x) = £1} £2 = min (©2 , x) , x^Ki K-2 = {x G Kx , (e2, x) = £>} Cn — niin , x) , — {x G Kn—i , (en , x) — Cn} • xEKn-i By induction, it is clear that all sets К D K\ D К 2 2 • • • D Kn are compact, hence the components £i,...,£n are well defined. The set Kn contains the single point £, which precedes all other points of К in the lexicographical order. If t t—> F(t) C HV? is a multifunction with compact values, we can now define the selection t •—►£(£) € F(£), where £(£) is the first point of the compact set F(t) w.r.t. the lexicographical order. Our next goal is to show that this selection is measurable. A preliminary result is needed. Lemma A.7.2. Let t 1—> F(t) C IRn be a bounded multifunction with closed graph, defined for t in a closed set J C R. Then for each vector v € IRn, the function = min (v, y) is lower semicontinuous. Proof Fix any r 6 J and consider a sequence of points tk 6 J with tk r, such that lim inf = lim <p(tk). t—►T, teJ k—^OQ Choose yk G F(tk) such that = P • У к- Since this sequence of vectors is uniformly bounded, we can extract a convergent subsequence, say —> У- Clearly у e F(t) because F has closed graph. We now have — min (v, y) < (v, y) = lim tp(tk) — liminfcp(Z), yeF(t) ’ k->oo t-^T proving the lemma. Theorem A.7.3. Let t 1—> F(t) be a bounded multifunction with closed, graph, defined for t € [a, b]. For each t, let £(t) € F(t) be the lexicographic selection. Then the map t»—* £(t) is measurable.
A.8 Convex sets 283 Fig. A.9. Three examples of lexicographic selections. Proof. Let £ = (£i,...,£n). To prove measurability, we need to check that each component 11—> £Д<) is measurable. This is clear for the first component, because £i = min (ei , x), xeF(t) and the result follows from Lemma A.7.2. By induction, assume that we already proved that the maps £i(-)> • • • ? £>() are all measurable. Then there exist countably many disjoint compact subsets Jk as in (A.21) such that the maps £i,..., are all continuous when restricted to each J*.. Consider the multifunction t i—> Fj(t) defined as Fj(t) = {x € F(t); (ee,x)=&(t), € = 1,2 ... j} . By construction, each set Fj (t) is compact and non-empty. Restricted to each set Jfc, the multifunction t Fj(t) has closed graph. Hence by Lemma A.7.2 the function Cj+i(i)= min (ee,x) is lower semicontinuous. Since the sets Jk cover the entire interval [a, b] except a set of measure zero, this proves that £j+i is also measurable. By induction on J, we conclude that all components of £(•) are measurable functions. A.8 Convex sets Let A be any set contained in a vector space. We say that A is convex if, for any two points x, xf e A, the segment that joins them is entirely contained in A. Otherwise stated, A is convex if Ox + (1 — 0)x' € A for every x, x' e A, 0 e [0,1].
284 A Appendices The convex hull of a set A is the smallest convex set which contains A. It can be represented as the set of all convex combinations of finitely many elements of A, namely {N N > 52 Ar > 1, Xi&A, Ai€[0,l], 5ZAi = 1f' (A-32) i=l i=l J An well known result of Caratheodory states that, for subsets of IRn, the convex hull can be obtained by taking convex combinations of only n + 1 points. Theorem A.8.1. (Caratheodory). If A C IRn, then {n+l n+1 У^ XiXi; Xi € A, Xi e [0,1], Xi = 1 > . (A.33) i=l i=l J Proof. Fix any element x G co A. Let N be the smallest integer such that we can write x as a convex combination of N points of A. We need to show that N < n + 1. 1. Assume, on the contrary, that N > n + 1. Consider the set N < (0 = (#i,..., On) ; У2 @ixi — x у k 1=1 0i e [o, i], Notice that (9 yields all possible ways of obtaining x as a convex combination of the Xi. Since О is a compact subset of IR1V, it contains a first point 0 = (#i,..., On) w.r.t. the lexicographical ordering. We claim that 0j = 0 for some index j. This will imply that x is the convex combination of N — 1 points, proving the theorem. 2. If 0 < Oi < 1 for all i = 1,..., TV, consider the equations N OiXi = 0 , 1=1 N 2> = °- 1=1 (A.34) We regard (A.34) as a linear homogeneous system of n + 1 equations in the N variables 0\,... ,0n- If N > n + l, this system has a nontrivial solution, say A = (Ai,..., Xn) / (0,..., 0). Choosing s > 0 sufficiently small, the two vectors 0 + eX and 0 — eX both lie in 3. Since 0 is the mid-point of the above two vectors, it cannot be the first element of 3 in w.r.t. the lexicografical order. We thus reach a contradiction, and the theorem is proved. The above theorem has some important consequences. In particular: Corollary A.8.2. The convex hull of a compact set К C IR77 is compact.
A.8 Convex sets 285 Proof. Let An be the standard n-dimensional simplex, defined as An = = (0q, • • * , вп); 0j = l, Gi > 0 for every i j C IRn+1. i=0 (A.35) Then co К is compact, being the image of the compact set К x • • • x К x An through the continuous mapping n (•ГО? ’ * ' 1 I > GiXi . i=0 A useful technique, that we already encountered in the proof of Brouwer’s theorem, is the continuous projection of points in IRn onto a convex subset. The following theorem collects the basic facts about perpendicular projections. Theorem A.8.3. Let A C IRn be a closed, convex set. Then, for every x E IRn there exists a unique point 7r(x) E A such that |тг(ж) — ж| = inf |a — ж|. (A.36) a£A The perpendicular projection x 1—> 7r(a;) has the following additional properties: (x — тг(х), a — 7г(ж)) < 0 for all x 6 IRn, a E A. (A.37) |тг(д?) — 7г(я')| < |z — a/| for all j?,j?'ElRn. (A.38) Proof. 1. To construct the projection, for any given x E lRn consider a se- quence of points ym E A such that lim \ym — x\ = inf |a — ж|. тп—-ос aCA This sequence is clearly bounded, hence by compactness we can extract a subsequence converging to a limit point y. We then set у = 7г(я). Clearly 7г(а?) E A because the set A is closed. 2. To prove the uniqueness, assume that y\,y2 E A are distinct points such that - ж| = |т/2 - ж| = inf |а - х|. а^А Then, taking the convex combination у = and observing that у— x and y2 —yi are perpendicular to each other (see figure A. 15), using Pythagoras’ theorem we obtain I |2 I 12 |У2 ~ У112 . p 1 j \y - Я = \У1 - я----------- я 4 aeA
286 A Appendices 3. To prove (A.37), consider the segment joining 7r(z) with a. Points on this segments can be parametrized as y(t) = тг(ж) + (a — тг(ж)), for t G [0,1]. We now compute -*i2 dt = 2 (а — 7г(яг), 7г(я:) — x) . t=o (A.39) By convexity, y(t) G A for all t G [0,1]. Since the distance ||?/(t) — ж|| attains a minimum at t = 0, the right hand side of (A.39) must be non-negative. Hence (A.37) holds. 4. Finally, we establish the contraction property (A.38). By the previous step we have (x — 7r(z) , 7г(я/) — 7г(я)) < 0 , (x' — тг(х') , 7г(х) — 7г(я/)) < 0 . Therefore (х — х', X — х') — (тг(х) - 7г(я/) , 7г(х) — 7г(я/)) = = — xf) + (тг(ж) — 7г(д/)) ? — х') ~ W37) “ 7Г(379)^ — 0 • A point х G A is an extreme point of a convex set A if it cannot be obtained as a strict convex combination of two distinct points of A. More precisely, x G A is an extreme point if x = Xxi T (1 — A)#2 for some , a?2 € A 0 < A < 1 implies Xi = x% = x. Theorem A.8.4. (Krein-Milman). Let E be a Banach space such that its dual E* separates points. Then every compact convex set К С E contains an extreme point. Fig. A. 10. The extreme points of a set A and the separation of two convex sets К, K' by hyperplanes.
A.8 Convex sets 287 Remark A.l (row and column vectors). Up to now, by a vector x e lRn we usually meant a column vector. In particular, the state x(t) G IRn of a system and the value u(t) G IRm of a control are always regarded as column vectors. However, in some cases it is useful to work also with row vectors. A linear mapping IRn i—► IR can be conveniently identified with a row vector p = (p1?...,pn), writing p-x = (pi,...,pn) n Pi^i • 2=1 Gradients of scalar functions ф : IRn IR will always be regarded as row vectors, having the form дф дф \ dxi ’ ’ dxnJ The next lemma is concerned with the separation of two compact convex sets by means of a hyperplane. Lemma A.8.5. (Separation of convex sets) Let K,K' C IRn be disjoint, closed, convex sets, with К compact. Then they can be strictly separated by a hyperplane. More precisely, there exists e > 0 and a unit row-vector p € IRn such that min p • v > sup p • v' + e. (A.40) Proof. 1. Choose two sequences of points x„ G К, x'y G K' such that lim \xy - x'y\ = inf {\y - y'\- у G K, y' G K'}. I/—>OC Since К is compact, the sequence xy is bounded; hence the sequence x'y is bounded as well. By taking a subsequence, we can assume xy-+x € K, xy^x' G K'. This of course implies |jr — xf| = min{|2/ - t/'|; у € К, у G К'}. (A.41) 2. We claim that the conclusion of the theorem is satisfied by Indeed, by (A.41), x' is the perpendicular projection of x on Kf, while x is the perpendicular projection of x' on K. Therefore, (A.37) implies p • x = min p • v, p’x' = max p • v'. (A.42) vGK v'EK1 Our claim is now an immediate consequence of (A.42).
288 A Appendices Lemma A.8.6. (Support hyperplanes). Let К C IRn be closed, convex, with boundary dK. Then every boundary point w G Ж admits a support hyperplane. More precisely, there exists a unit vector p such that p • w — max p • y. (A.43) Proof. Choose a sequence of points xv К with xu—+y. Let у у = 7r(xp) be the perpendicular projection of Xy on K. Observe that y^^w as i/—>oo. By possibly taking a subsequence, we can assume that there exists a unit vector p such that . - у у Ру = ;--------;—*p as V~>0C. - Уу\ Recalling (A.37), we have Pv • Уу > Ру ’ У for all у e К, и > 1. For every у € К, this yields p • w = lim py • w = lim pu • yu > lim py • у = p • y, y—>OO У—ЮО y—>OQ proving the theorem. A.9 Convex cones In this section we study a special class of convex sets, namely convex cones. We recall that a set Г C lRn is a cone if Л x e Г whenever x 6 Г, A > 0 . Let V C IRn be any set of vectors. The span and the positive span of V are defined respectively as {N У A^; г=1 {N 57 Аг^; i=l Now consider a closed set S C lRn tangent cone to S at the point £ as N > 1, Vi e V, Xi e IR N > U eV, Xi > 0 and fix any point £ 6 S. We define the Ts(O = {<> = lim e 1dU + ev, S) = 0 e—o+ v 7
A.9 Convex cones 289 Example A.2. Consider the set (see figure A.11) S = {(#1,^2) C IR2 ; sq > 0, X2 > 0, xj + — 1} Then the tangent cone to S at the point £1 = (1,0) is ^s(Ci) = {(2/ь?/2) C IR2 ; yi < 0, У2 > 0} . The tangent cone to S at the point £2 = (1/a/2 , l/\/2) is ^>(£2) — {(У1,Уз) E IR2; yi + У2 < 0} . Fig. A.11. Examples of tangent cones. We shall be mainly interested in the case where the set S is defined in terms of finitely many equations or inequalities (see figure A. 11) S = {.т; ^(ж) = 0, i = k}, (A.44) S+ = {x\ фо(х) > фо (£), ФЛХ>) = * • • , (A.45) where </>o, • • • , Фк are C1 functions, £ G S'. The next lemma is an immediate consequence of the implicit function theorem. Lemma A.9.1. (Representation of tangent cones). Let фо,’",фк • IRn и-> IR. be continuously differentiable. Assume that, at the point £ G S, the gradients V</>o(£)>”’ ^Фк(Я) are linearly independent. Then the tangent cones at £ to the sets (A.44), (A-4$) are, respectively: Ts($ = {v, V0i(£)-v = O, i = (A.46) Ts+^) = {v; V</>0(£) • v > 0, Wf(£H = 0, г = !,•••,&}, (A.47)
290 A Appendices Lemma A.9.2. Under the same assumptions of Lemma A.9.1, a vector p satisfies p • v > 0 for all v e Ts+ (£) (A.48) if and only if p can be written as a linear combination к р = ^Х^ф^) (A.49) i=O with Aq > 0. Proof. Indeed, if (A.49) holds with Aq > 0, by (A.47) we have p • v = AqV0o(£) • v > 0 for all v G (£). To prove the converse, assume that (A.48) holds and choose additional smooth functions 0^+1, • • • , фп-i • •—> IR such that the set of gradients Wo(£), V0i(e), ••• , forms a basis in IRn. Let {do, • • • , bn-i} be the dual basis, so that _ , , fl if i = j , ь, - 10 ,f . . Observe that, by (A.47), the vectors do, ibfc+ь • • • , ±dn_i all lie inside Ts+ (£). Write the vector p as a linear combination n—i i=0 where Xi = p • bi, i = 0,1, • • • , n — 1. If Aq <0 then p • do < 0, in contrast with (A.48). If Xi 0 for some i > k, then either p • bi < 0 or p • (—dj < 0. In any case, this contradicts (A.48). We say that two cones Г, Г' are weakly separated if there exists a unit vector p such that p • v < 0, p • v' > 0, for all v G Г, v' G Г'. The next two lemmas on the separation of tangent cones play a key role in the proof of Pontryagin’s Maximum Principle. Lemma A.9.3. Let Г = span+(V) for some set V C IRn, let Г' = Ts+(£), with Т$+,£ as in Lemma A.9.1. Assume that the cones Г,Г' are not weakly separated. Then there exist finitely many vectors Vo,--- ,vn € V such that span* {t>o, • • • , тдг} is not weakly separated from Г'. Moreover, there exists e > 0 such that, if \wi - vj < e i = 0, • • • ,N, then the cones span+{wo, • • • , w^} and Г' still cannot be weakly separated.
A.9 Convex cones 291 Proof. 1. Let {vj}j>o be a sequence everywhere dense in Г. Assume that, for all v > 1, there exists a unit vector py such that • Vj < 0 j = 1, • • ♦ , y, pu • v' > 0 for all v' G Г'. (A.50) By taking a subsequence, we can assume py-^p for some unit vector p. By (A.50) and the density of the sequence {vj} it follows p • v < 0 for all v G Г, p • v' > 0 for all vr G Г', a contradiction. Hence, for some A, span+{vo, • • • ,vn} is not weakly sepa- rated from Г'. 2. The proof of the second statement is similar. If no e > 0 with the desired property exists, then one can construct sequences (v(0>l/), • • • , v^yf) and unit vectors py such that for every у > 1, v’ G Г' and i = 0, • • • ,7V, one has: |^i,p - < p pv-v' > 0. (A.51) Taking a subsequence, we can assume Pv—tp. Then (A.51) implies p • Vi < 0 p • v1 > 0 for all v1 G Г', i = 0, • • • , N, showing that the cones span+{t?o? • • • ,vn} and Г' are weakly separated, a contradiction. Lemma A.9.4. Let Г = span+{vo, • • • , v^} C IRn and let Г' = T$+(£), with S+,£ as in Lemma A.9.1. Define the unit simplex An as in (A.35) and let X : [0, e] x An i—> IRn be a continuous map such that in, г—0 uniformly for 9 G An- If Г and Г' are not weakly separated, then there exists some £, 9 such that фв(Х(ё, 0)) > 0o(^), 0)) = 0, i = 1, • • • , к. (A.53) Remark A.2. In connection with the constrained optimization problem max |</>o(X(e,0)); (e,0) e [0,e] x 4N, X(e,0) 6 S’} , (A.54) the conclusion of the lemma states that the optimal value for (A.54) is strictly larger than <fo(£) (see Figure A.12).
292 A Appendices Fig. A. 12. If the cones Г and Г' are not weakly separated, then £ is not an optimal solution to the constrained optimization problem (A.54). Proof, of Lemma A.9.4- It is clearly not restrictive to assume that </>o(£) = 0. Assume that the cones Г and Г' are not weakly separated. Consider the vector-valued function ф = (0o, • • •, <^fc) ’ Its (fcfl)xn Jacobian matrix will be denoted by \?ф = (V0o, • • • , V</>&). Moreover, define the cone Г* =span+{V^ Vi; i = 0, ••• ,JV} C IRfe+1. 1. We claim that the vector w = (1,0,-•• ,0) e lRfc+1 lies in the interior of the cone Г*. Indeed, if w is not in the interior of Г*, by Lemma A.8.6 there would exist some unit vector p = (po, • • ,р&) such that pw=maxpp. (A.55) yer* Since Г* is a cone containing the origin, both quantities in (A.55) must be zero, i.e. po = 0. Now consider the linear functional on IRn к V p • V0 • V = ‘ V' i=l By (A.47) and the definition of Г*, this functional is identically zero on Г'. On the other hand, by (A.55) it is non-positive on Г. The two cones Г, Г' are thus weakly separated. This contradiction proves our claim. 2. By the previous step, there exist к + 2 vectors, say Wq, • • • , Wk+i 6 such that
A.9 Convex cones 293 w E int co{wq, • • • , Wfc+i}, (A.56) TV Wi — CijV0 • Vj г = 0, • • • ,fc-h 1, (A.57) j=0 for some nonnegative coefficients Cij. Observe that it is not restrictive to assume that > 0 for all i,j. Define the map h : (e, $0,..., $fc+i) «—> (б,0о,- • • ,#tv) by setting (\ -1 \ fc+i 57 I = - 57 ^iCij г,£ J i=0 3. Define a new map (б,$) i—> У (б,??) from [0,f] x Ak+i hito IRfc+1 by setting Y(e,,..., i?fc+1) = ^(X(A(6, tf0, • • •, tffc+i))) • To prove the lemma, it now suffices to construct a point (б, $□,..., tfk+i), with e > 0, such that Г(ё, t?o, • • •, dk+1) = (c, 0,..., 0) € IRfc+1. (A.58) Observe that, by (A.52) and (A.57), we have lim (A.59) € г=0 uniformly for 0 6 Дк+i • By (A.56), the map fc+i 0 : ($o, • ' , tffc+i) ^2 ^iWi i=0 is a linear bijection between the unit simplex Дк+i and a convex neighborhood К of w, namely К = co {wq, ..., Wk+i} • Call 0-1 : К i—> Дк+i its inverse mapping. For б > 0, define the map f(: К Rfc+1 by setting As б —> 0+, we have lim /c(w) = w uniformly for w € K. In particular, this holds for points w € dK on the boundary of К. Since w = (1,0,..., 0) € int К, we can apply Corollary A.3.5 and deduce the existence of б > 0 sufficiently small and w € К such that /?(w) = w. Hence (A.58) holds with d = 0-1(w). This completes the proof.
294 A Appendices A. 10 Lie brackets and Frobenius’ theorem In the following, we shall use the exponential notation 0 i—> (exp0/)(a:) to denote the solution of the Cauchy problem dW , ZAK — = /(w), w(0) = x. (A.60) (1U Moreover, for a fixed 0, we shall denote by (exp#/)* = Z)x(exp^/) the Jaco- bian differential of the map x (expOf)(x). (A.61) Given two smooth vector fields f and g on IRn, their Lie bracket is the vector field defined as [f,5] = Dxg f - Dxf-g, In other words, the Lie bracket [/, <?] is the directional derivative of g in the direction of /, minus the derivative of f in the direction of g. The following lemma provides various equivalent constructions of Lie brackets. Lemma A. 10.1. (Characterization of Lie brackets). The Lie bracket can be equivalently characterized as [/, 5] = lim [(exp eg)(exp e/) - (expE/)(expsp)j, (A.62) [/,g] = liml [(exp(-eff))(exp(-£/))(expE5)(exp£/)], (A.63) [f,g] = lim |[(exp(-e/))tp(exps/) - p], (A.64) Proof. In the following estimates, the Landau symbol o(s2) indicates a higher order infinitesimal w.r.t. б2, i.e. a quantity such that o(£2)/e2 —> 0 as e —> 0. 1. To prove (A.62), we observe that (exps/)(x) = x + ef(x) + — (Dxf) f(x) + o(e2), (expc5)(expE/)(x) = x + ef(x) + ^-(Dxf) f(x) +eg(x) + e2(Dxg) f(x) + у(Dxg) • g(x) + o(e2). (A.65) Interchanging f and g in (A.65) and subtracting the results we obtain (ехре<?)(ехрг/)(х) - (exp ef) (exp eg) (x) = e\(Dxg) f - (Dxf) • д') + o(e2). Hence (A.62) follows.
A. 10 Lie brackets and Frobenius’ theorem 295 2. Concerning (A.63), we have lim£_o e~2 [(exp(-Eg)) (exp(-e/))(expep)(expe/)(x)j = lime_0 (ехр(-£^))Дехр(-£/))ф- •lime_o e~2 |\ехрез)(ехр£/)(а:) - (exp e/) (exp Eg) (a:)^ = [/,<?] (я:)- 3. Finally, to prove (A.64), we observe that the value at time t = 0 of the solution of the linear Cauchy problem i>(£) = £>x/((expt/)(a;)) • v{t), v(e) = p((expe/)(a;)), (A.66) is precisely v(0) = (exp(—e/))* g(expef)(x). (A.67) From (A.66)-(A.67) it follows v(0) — v(e) - eDxf(x) v(e) + o(e) = g(x) + eDxg(x) f(x) - eDxf(x) g(x) + o(e) (A.68) Letting e —> 0 in (A.68) we obtain (A.64). One can think of (A.63) as a recepy for constructing the Lie bracket. Starting from any point ж, let us move along the vector field f for a time E, then along g for time e, then along — f for time e, and finally along — g for time e. As shown in figure A. 13, in the end we reach a point ж + e2 [/, </] (ж) + o(e2). Fig. A.13. Two constructions of the Lie bracket [/,<?]. Example 3.8. Consider again the car steering problem (1.16) discussed in the Introduction. This control system is described by the three variables P = (Ж1,Ж2,0). Here (ж1,жг) yield the position of the barycenter of the car, while the angle в determines the orientation. If the car advances with unit speed steering to the right, its motion satisfies
296 A Appendices -T7 — f(P) = (cos#, sin#, -1). at On the other hand, if the car steers to the left, its motion satisfies dP = g(P) = (cos 0, sin0, 1). In a typical parking problem, one needs to shift the car on one side, without changing its orientation, see figure 1.6 in Chapter 1. This is obtained by the following maneuver: 1. right-forward, 2. left-forward, 3. right-backward, 4. left-backward. Indeed, according to (A.62) the sequence of these four actions generates the Lie bracket [/, g] - (Dg) f - (Df) • g — (2 sin #, - 2 cos #, 0) which is precisely the desired direction. According to (A.62), the Lie bracket [/, g] vanishes if the flows generated by the vector fields /, g commute. Roughly speaking, [/, g] measures the non- commutativity of these flows, at an infinitesimal level. The next result provides a converse to Lemma A. 10.1. To avoid technicalities, we assume here that the vector fields /, g are globally defined on IRn and satisfy a sub-linear growth condition such as |Ж)|<с(1 + И). (A.69) Fig. A.14. Proof of the commutativity property.
A. 10 Lie brackets and Frobenius’ theorem 297 Lemma A. 10.2. (Commuting vector fields). Let fig be smooth vector fields on IRn, satisfying the sub-linear growth condition (A.69). Assume that their Lie bracket vanishes identically: [fig] = 0. Then they commute, namely (exp sffiexptgfix) = (exp tg) (exp sfi) (ж) (A.70) for every x E lRn and every s.t 6 IR. Proof. For each 9 E [0, t], consider the points (see figure A. 14) Q(0) = (exp(t - 9)gfiexpsfifiexp9gfix), P(9) = (exp sfifiexp9gfix). We claim that ^W) = o. (A.71) au This is clearly true, provided that we can show b(0)=ff(P(0)). (A.72) av A proof of (A.72) goes as follows. Fix 9, and call T] ~ Xе (tj) = (expTj/)(exp^)(f) the solution to the Cauchy problem = ЛаЛт?)), ®0(O) = (ехр0р)(ж). ar/ Moreover, let g i—► ve(g) be the solution of the linearized equation = Dxf(xe^) • veM, V*(O) = <7(^(0)). (A.73) dg According to Theorem 2.3.1, for every g E [0, s] one has the identities (exp tj/) (exp 0g)(x) = t/(»j) • In particular, when g = s we have -^P(0)=tr’(s). at/ To prove (A.72), it thus suffices to show that ^(tj) = f°r all 9 € [0,t] • (A.74) Using the assumption [/, g] = 0, we can write fg(xe^) = Dg(xe(7))).f(xe(Tj)) = Df(xe(ri))-g(xe(r/)). (A.75)
298 A Appendices Together, (A.73) and (A.75) imply -Р(^(т?))| < |Дг/(^(7?)) • (ve(rj) -5(^(7/))) j < C |ve(?7) -<z(x0(t/))| When T] = 0, by (A.73) we have г/0(О) — д(ж0(О)) = 0. An application of Gronwall’s Lemina 2.1.2 now yields (A.74). Hence (A.72) and (A.71) also hold. In turn, (A.71) yields 6^(0) = Q(t). By the definition of Q, this is precisely the conclusion of the lemma. We now consider a family of vector fields Gi,..., Gm on IR5. assuming that all their Lie brackets vanish identically: [G^GjCrHO for all г, j € {1,..., m} , x G IR 5 . (A.76) The following theorem shows that, in this case, the equations x = Gi(x) can be simultaneously integrated. Theorem A.10.3. (Frobenius). Let Gi,... , Gm be smooth vector fields on IR5, satisfying a sub-linear growth condition (A.69) and the commutativity assumptions (A. 76). Then, for any given x G IR 5, there exists a unique C1 map и = (щ,... , zzm) i—► Ф(и) from IRm into IR5 such that Ф(0) = x, дФ (и) = СкЩи» duk for all fcG{l,...,m}, и G IRm. (A.77) Proof. The sub-linear growth condition guarantees that, for every Cauchy problem of the form at x(0) = y. the corresponding solution t x(t) = (exp tGifiy) is well defined for all t G IR and does not blow up in finite time. 1. To prove uniqueness we observe that, by (A.77) the map t i—> x(t} = Ф(£,0,... ,0) satisfies = Gi(x(t)), z(0) = x. Therefore, for every щ G IR, when t — ui we must have Ф(гг1,0,... ,0) = х(щ) = (expuiGi)(rr). Similarly, the map t i—+ x(fi) = Ф(и\, t, 0,.... 0) satisfies
A. 10 Lie brackets and Frobenius’ theorem 299 -^-x(t) = G2(ic(t)), ж(0) = 0,..., 0) = (exp tiiGi)(я). dt Therefore, for every U2 6 IR, when t = ?i2 we must have 0(t£i, tz2,..., 0) = a:(u2) = (expu2G2)(expiAiGi)(z). Repeating the above argument, we conclude that the map Ф : lRzn > IRa must be given by Ф(и) = (expwmGm) ••• (expuiGi)(£). (A.78) In particular, Ф is unique. 2. It remains to show that the map Ф defined by (A.78) has the required properties (A.77). The condition Ф(0) = x is clearly satisfied. To compute the partial deriva- tives of Ф, we observe that, by Lemma A. 10.2, we can arbitrarily permute the order of the terms in (A.78). In particular, for every fixed index k G {1,..., m}, we can equivalently write Ф(и) = (expufcGfc)(y), with у = (expumGm) ••• (expitfc+iGfc+i)(expufe-iGfe_i) (expUiGiXx). Differentiating the above expression, we obtain д . d -—Ф(и) = — дик дик (expufcGfc)(y) = Gfc ((expufcG*:)(y)) = Gfc(*(u)). thus proving (A.77). Notice that, with a slight abuse of notation, if the commutativity assump- tions (A.76) hold, one can write Ф(и) = | exp^'UiGj j (ж). (A.79) \ i=i / We conclude with a useful lemma. In the following, Dz(exp VjGj) denotes the differential of the map x i—> (exp VjGj)(x). Lemma A. 10.4. With the same assumptions as in Theorem A. 10.3, let (m \ expy%jGj (*) j=i / Then (A.80)
300 A Appendices Proof. Consider the map m m t ( exp VjGj^ (exp tGi)(x) — (exp tGf) ( exp У^ vjGj^ (x) j=i j=i = (expiGi)U). Computing the derivative of the first and last term w.r.t. t at the time t = 0, one obtains precisely (A.80). Problems A.l. Let X be a Banach space. We say that the norm || • || is strictly convex if Д? ~h 7/ —-— < 1 whenever 11 x 1| = ||?/|| = 1, x/y. Prove that the following three statements are equivalent: • The norm || • || is strictly convex. • Given 0 < 0 < 1 and any two distinct vectors ж, у E X. x =4 y, one has ||fe + (1 - 0)i/|| < max{||x||, IIj/II} . • Every point x E dB on the boundary of the unit ball В — {ж E X ; ||ж|| < 1} is an extreme point of B. A.2. Show that the set of equations 2ж = cos у, 2y = sin ж admits a solution (ж*,?/*) inside the unit disc. Hint: apply a fixed point theorem to the map /(ж, у) = cosy, | sin ж) . A.3. Let t н-> F(Z) C IRn be a multifunction with closed graph but possibly unbounded values. Prove that F admits a measurable selection. Hint: define r(t) = min3z€F(t) |t/| and consider the multifunction F*(t) = {x€F(t); |z| = r(i)}. A.4. The contingent cone to a set S at the point £ € S' is defined as Cs(£) = : lim inf s-1d(£ + ev,S) = ()j> . Compute the tangent and contingent cones at 0 E IR2 to the spirals (writ- ten in polar coordinates ж = pcosO, у = psin0) Si = {(r, 0); r = 6, 0 < 0 < oo}, $2 = {0} U {(r, 0); r = ee, - oo < 0 < oo}, S3 = {0} U {(r,0); r = 0-1, 0 < 0 < oo}.
A. 10 Lie brackets and Frobenius’ theorem 301 A.5. Consider the function f : [0,1] i—> IR defined by setting f(t) = t if t is irrational, f(t) = 0 if t is rational. Show that f is continuous only at the point t = 0. Find the Lebesgue points of f. Construct a compact set К C [0,1] such that meas(/C) >1/2 and the restriction of f to the set К is continuous. Hint: let , <72, ... be the list of all rational points in [0,1]. Let Ik by an open interval containing qk with length 2“fe“1. Consider the union of all intervals Ik and its complement. A.6. Let f : [a, b] x IRm h-> ]Rn be continuous and let и : [a, b] > IRm be a bounded measurable map. If r is a Lebesgue point for u, show that r is also a Lebesgue point for the composed mapping t f(t, Hint: for every 8 > 0 prove that lim_ ^-meas{£ 6 [r — e,r + e] : |/(t,u(t)) - /(t,u(t))| > 5} = 0. A.7. Let A c IR3 be the set A = {(x,j/,0); x2 + y2 = 1} U {(1,0,1)} U {(1,0, -1)} . Prove that its convex hull co A is compact. However, show that the set of extreme points of co A is not closed. A.8. In Theorem A.2.1, assume that ||Ф(А,ж) — Ф(Л',x)|| < L || A — A'||, for all A, A' € A, x E X. Show that the distance between the two fixed points is bounded by ^(A) -xfA'XI < _А^||Д_ A'|| 1 n. for all A, A' € A. A.9. Prove the representation formula (A.32) for the convex hull of the set A. A. 10. Given an open set 12 Q IRn, consider the Banach space C1(J2) of all continuously differentiable mappings f : 12 IR, with norm ||/||C1 = sup |/(x)| + sup|V/(x)|. Here V/ = (/X1, ..., fXn) denotes the gradient of f. If 12 is convex, prove that every f € C^J?) is Lipschitz continuous and ||/||ыр < ll/llc1- How- ever, show that the above can fail if 12 is not convex. Hint: let 12 be the union of two disjoint open circles, tangent at one point. A. 11. On the plane IR2, consider the closed disc D = {(x,y); x2 + y2 < 4},
302 A Appendices and the circumference s = {(x,y); - I)2 + y2 = 1} Observe that S C D. Find a point £ such that the two tangent cones 7p(£) and Ts(£) are weakly separated. A. 12. Describe the tangent cones to the sets S,S+ C IR3 defined as S = | (x, y, z); x2 + y2 = 1, x + z — o|, | (ar, ?/, z); x2 + y2 = 1, x + z = 0 , z > o|, at the point £ = (0,1,0). A. 13. Show that the two sets K = {(x,y); x2+y2<l}, K' = {(x,y)- |x + 2| + |2/-2| < 1} . can be strictly separated, i.e. find explicitly a unit vector p and e > 0 so that (A.40) holds. A.14. Let К, К' C IRn be two compact convex sets such that max p • x < max p • x хек ~ хек' for every unit row-vector p 6 Rn. Prove that К С K'. In particular, show that K' contains the closed ball 22(0, r) centered at the origin with radius r > 0 if and only if max p • x > r x£K' for every unit row-vector p G IRn. A. 15. In the plane IR2. consider the two sets 2<={(ar, 2/); x > 0, у > 0, д?2/>1|, К' = {(я, у); у = 0} . Notice that К, К' are closed and disjoint. However, prove that they cannot be strictly separated, i.e. one cannot find any unit vector p and s > 0 such that (A.40) holds. A. 16. Let x i—> F(x) C IRn be a Hausdorff continuous multifunctions with con- vex, compact values. Fix a point у 6 lRn and define f(x) 6 F(x) as the perpendicular projection of у on the set F(t), i.e. (see figure A. 15) = 7Гг(х)(у). Prove that the map x f (x) provides a continuous selection of F.
A. 10 Lie brackets and Frobenius’ theorem 303 A. 17. On the space IR2, consider the multifunction x !-> F(x) = € IR2; |y| < 1, |y-a:| > l/з} . Show that F is Hausdorff continuous with compact (but not convex) values (see figure A. 15). Prove that there exists no continuous selection x i—> /(x) e F(x). Fig. A. 15. A continuous selection of a convex-valued multifunction and a continuous multifunction without any continuous selection. A.18. Let the assumptions of Theorem A.10.3 hold. For any u,u' G IRm and x G IR a , prove the identity m \ / m \ / m \ exp^2(ui +u'i)Gi j (я) = j exp^u'G; j I exp^PujGj I (ж). i=l / \ i=l / \ i=l / A. 19. Consider two linear vector fields on IRn, say f(x) = Ax, g(x) = Bx for some n x n matrices A, B. Compute the Lie bracket [/, g] and show it is also a linear vector field.

References [1] A. A. Agrachev and Yu. Sachkov, Control Theory from the Geometric View- point. Springer Ver lag, Berlin, 2004. [2] F. Ancona and A. Bressan, Patchy vector fields and asymptotic stabilization. ESAIM - Control, Optimiz. Calc. Var. 4 (1999), 445 471 [3] F. Ancona and A. Bressan, Flow stability of patchy vector fields and robust feedback stabilization SIAM J. Control Optim. 41 (2003), 1455-1476. [4] F. Ancona and A. Bressan, Nearly time optimal stabilizing patchy feedbacks. Ann. Inst. Henri Poincare, Analyse Non Lineaire, 24 (2007), 279-310. [5] J.P. Aubin and A. Cellina, Differential Inclusions. Springer-Ver lag, 1984. [6] J.P. Aubin and H. Frankowska, Set Valued Analysis. Birkhauser, Boston, 1990. [7] A. Bacciotti, Local Stabilizability of Nonlinear Control Systems. World Scien- tific, River Edge, NJ, 1992. [8] M. Bardi and I. Capuzzo Dolcetta, Optimal Control and Viscosity Solutions of Hamilton-Jacobi-Bellman Equations. Birkhauser, 1997. [9] G. Barles, Solutions de Viscosite des Equations de Hamilton-Jacobi. Springer, 1994. [10] L.D. Berkovitz, Optimal Control Theory. Springer-Verlag, New York, 1974. [11] V.G. Boltyanskii, Sufficient conditions for optimality and justification of the dynamic programming method, SIAM J. Control, 4 (1966), pp. 326-361. [12] U. Boscain and B. Piccoli, Optimal Syntheses for Control Systems on 2-D Manifolds. Springer, New York, 2004. [13] A. Bressan, Singularities of stabilizing feedbacks, Rend. Sem. Mat. Univ. Pol. Torino 56 (1998), 87-104. [14] A. Bressan, Impulsive control of Lagrangian systems and locomotion in fluids, Discr. Cont. Dynam. Syst., to appear. [15] A. Bressan and B. Piccoli, A Baire category approach to the bang-bang prop- erty, J. Diff. Equations 116 (1995), no. 2, 318-337. [16] A. Bressan and F. Rampazzo, On differential systems with vector-valued im- pulsive controls, Boll. Un. Matematica Italiana 2-B, (1988), 641-656. [17] A. Bressan and F. Rampazzo, Impulsive control systems with commutative vector fields, J. Optim. Theory & Appl. 71 (1991), 67-84. [18] A. Bressan and F. Rampazzo, On systems with quadratic impulses and their application to Lagrangean mechanics, SIAM J. Control 31 (1993), 1205-1220.
306 References [19] A. Bressan and F. Rampazzo, Moving constraints as stabilizing controls in classical mechanics, preprint (2007). [20] Aldo Bressan, Hyper-impulsive motions and controllizable coordinates for La- grangean systems Atti Accad. Naz. Lincei, Memorie, Serie VIII, Vol. XIX (1990), 197-246. [21] Aldo Bressan, On some control problems concerning the ski or swing, Atti Accad. Naz. Lincei, Memorie, Serie IX, Vol. I, (1991), 147-196. [22] R. W. Brockett, Asymptotic stability and feedback stabilization, in Differen- tial Geometric Control Theory, R. W. Brockett, R. S. Millman and H. J. Suss- mann Eds., Birkhauser, Boston, (1983), 181-191. [23] P. Brunovsky, Existence of regular synthesis for general problems, J. Diff. Equations, 38 (1980), pp. 317-343. [24] F. Cardin and M. Favretti, Hyper-impulsive motion on manifolds. Dynam. Contin. Discrete Impuls. Systems 4 (1998), 1-21. [25] D. A. Carlson, A. B. Haurie, and A. Leizarowitz, Infinite Horizon Optimal Control: Deterministic and Stochastic Systems. Springer-Verlag, 1991. [26] L. Cesari, Optimization Theory and Applications. Springer-Verlag, 1983. [27] F.H. Clarke, Optimization and Nonsmooth Analysis. Wiley Interscience, New York, 1983. [28] F. H. Clarke, Necessary conditions in dynamic optimization, Memoirs Amer. Math. Soc. 816, vol. 173 (2005). [29] F. H. Clarke, Yu. S. Ledyaev, E. D. Sontag, and A. I. Subbotin, Asymptotic controllability implies feedback stabilization. IEEE Trans. Automat. Control 42 (1997), 1394-1407. [30] F. H. Clarke, Yu. S. Ledyaev, R. J. Stern, and P. R. Wolenski, Nonsmooth Analysis and Control Theory. Graduate Texts in Mathematics 178, Springer- Verlag, New York, 1998. [31] R. Conti, Problemi di Controllo e di Controllo Ottimale, U.T.E.T., Torino, 1974 (Italian). [32] R. Conti, Processi di controllo lineari in IRn, Pitagora Editrice, Bologna, 1985 (Italian). [33] J. M. Coron, A necessary condition for feedback stabilization. Systems Control Lett. 14 (1990), 227 232. [34] M. Crandall and P. L. Lions, Viscosity solutions of Hamilton-Jacobi equations, Trans. Amer. Math. Soc. 277 (1983), 1-42. [35] M.G. Crandall, L.C. Evans and P.L. Lions, Some properties of viscosity so- lutions of Hamilton-Jacobi equations, Trans. Amer. Math. Soc.. 282 (1984), pp. 487-502. [36] M. Crandall. II. Ishii and P. L. Lions, User’s guide to viscosity solutions of second order partial differential equations, Bull. Amer. Math. Soc. 27 (1992), 1-67. [37] G. Dal Maso, P. LeFloch, and F. Murat, Definition and weak stability of nonconservative products. J. Math. Pures Appl. 74 (1995), 483-548. [38] K. Dcimling, Nonlinear Functional Analysis, Springer-Verlag, Berlin, 1985. [39] K. Deimling, Multivalued Differential Equations, W. de Gruyter, Berlin, 1992. [40] J. Dieudonne, Foundations of Modern Analysis, Academic Press, New York, 1960. [41] L.C. Evans and M.R. Janies, The Hamilton-Jacobi-Bellmann equation for time-optimal control. SIAM J. Control, 27 (1989), pp. 1477-1489.
References 307 [42] A.F. Filippov, On certain questions in the theory of optimal control, SIAM J. Control, 1 (1962), pp. 76-84. [43] A.F. Filippov, Differential equations with multivalued discontinuous right hand side, Dokl. Akad. Nauk. USSR, 151 (1963), pp. 65-69. [44] A.F. Filippov, Classical solutions of differential equations with multivalued right hand side, SIAM J. Control, 5 (1967), pp. 609-621. [45] W.H. Fleming and R.W. Rishel, Deterministic and Stochastic Optimal Con- trol. Springer-Verlag, New York, 1975. [46] G.B. Folland, Real Analysis, Wiley-Interscience, New York, 1984. [47] A. T. Fuller, Study of an optimum non linear control system, J. Electronics Control, 15 (1963), pp. 63-71. [48] R. Gabasov and F.M. Kirillova, High order necessary conditions for optimal- ity, SIAM J. Control, 10 (1972), pp. 127-168. [49] R. V. Gamkrelidze, On sliding optimal regimes, Soviet Math. Dokl., 3 (1962), pp. 390-395. [50] O. Hajek, Discontinuous differential equations I, J. Differential Equations 32 (1979), 149-170. [51] P. Hartman, Ordinary Differential Equations, S.H. Hartman, Baltimore, 1973. [52] H. Hermes, On local controllability. SIAM J. Control Optim. 20 (1982), 211- 220. [53] H. Hermes, Nilpotent and high-order approximations of vector field systems, SIAM Review, 33 (1991), pp. 238-264. [54] H. Hermes and J.P. LaSalle, Functional Analysis and Time Optimal Control, Math, in Sci. & Engr., vol 56, Academic Press, 1969. [55] A. Isidori, Nonlinear Control Systems. Third edition. Communications and Control Engineering Series. Springer-Verlag, Berlin, 1995. [56] V. Jurdjevich, Geometric Control Theory. Cambridge University Press, 1996. [57] M. Kawski, High-order small-time local controllability, in Nonlinear control- lability and optimal control, H.J. Sussmann ed., M. Dekker, New York, 1990. [58] H.W. Knobloch, High order necessary conditions in optimal control theory, Lecture Notes in Control and Information Sciences, Springer-Verlag, Berlin, 1981. [59] A.J. Krener, The high order maximal principle and its applications to singular extremals, SIAM J. Control, 15 (1977), pp. 256-293. [60] S. Kruzhkov, First order quasilinear equations with several space variables, Math. USSR Sbomik 10 (1970), 217-243. [61] J. Kulkarni and B. Piccoli, Pumping a swing by standing and squatting: do children pump time optimally?, Control Systems Magazine, IEEE 25 (2005), 48-56. [62] E.B. Lee and L. Markus, Foundations of Optimal Control Theory, John Wiley, New York, 1967. [63] J.P. Lions, Optimal Control of Systems Governed by Partial Differential Equa- tions, Springer-Verlag, Berlin, 1976. [64] I. Lasiecka and R. Triggiani, Control theory for partial differential equations: continuous and approximation theories. I. Abstract parabolic systems. Ency- clopedia of Mathematics and its Applications 74. Cambridge University Press, 2000. [65] I. Lasiecka and R. Triggiani, Control theory for partial differential equations: continuous and approximation theories. II. Abstract hyperbolic-like systems
308 References over a finite time horizon. Encyclopedia of Mathematics and its Applications 75. Cambridge University Press, 2000. pp. 645-1067. [66] C. Marie, Geometric des systeines mecaniqucs a liaisons actives, in Symplectic Geometry and Mathematical Physics, 260-287, P. Donato, C. Duval, J. El- hadad, and G. M. Tuynman Eds., Birkhauser, Boston, 1991. [67] L. W. Neustadt, The existence of optimal control in absence of convexity conditions, J. Math. Anal. Appl., 7 (1963), pp. 110-117. [68] H. Nijmejer and A.J. van der Schaft, Nonlinear Dynamical Control Systems, Springer-Verlag, New York, 1990. [69] C. Olech and A. Lasota, On the closedness of the set of trajectories of a control system, Bull. Acad. Polon. Sci., 16 (1968), pp. 711-716. [70] B. Piccoli, Classification of generic singularities for the planar time optimal synthesis, S.I.A.M. J. Control Optim. 95 (1996), 59-79. [71] B. Piccoli and H.J. Sussmann, Regular synthesis and sufficiency conditions for optimality, SIAM J. Control Optim. 39 (2000), no. 2, pp. 359-410. [72] L.S. Pontryagin and V.G. Boltyanskii, R.V. Gamkrelidze and E.F. Mishenko, The Mathematical Theory of Optimal Processes, John Wiley, New York, 1962. [73] F. Rampazzo, On the Riemannian structure of a Lagrangean system and the problem of adding tiine-dependent coordinates as controls European J. Mechanics A/Solids 10 (1991), 405-431. [74] L. Rifford, Semiconcave control-Lyapunov functions and stabilizing feedbacks. SIAM J. Control Optim. 41 (2000), 659-681. [75] R.T. Rockafellar, Convex Analysis, Princeton University Press, 1970. [76] W. Rudin, Functional Analysis, McGraw-Hill, New York, 1973. [77] W. Rudin, Real and Complex Analysis, McGraw-Hill, New York, 1974. [78] G. V. Smirnov, Introduction to the Theory of Differential Inclusions, A.M.S. Graduate Studies in Mathematics, vol. 41 (2002). [79] E. D. Sontag, A Lyapunov-like characterization of asymptotic controllability, SIAM. J. Control Optim. 21 (1983), pp. 462-471. [80] E. D. Sontag, Mathematical Control Theory. Deterministic Finite-dimensional Systems, Second edition. Springer-Verlag, New York, 1998. [81] G. Stefani, Polynomial approximations to control systems and local control- lability, Proc. 24th IEEE Conf. Decision & Control, Ft. Lauderdale (1985), pp. 33-38. [82] A. I. Subbotin, A theory of generalized solutions to first-order PDEs with an emphasis on differential games. In Advances in nonlinear dynamics and control: a report from Russia, 189 229, Birkhuser, Boston, 1993. [83] H. J. Sussmann, On the gap between deterministic and stochastic ordinary differential equations, Ann. Prob. 6 (1978), 17-41. [84] H. J. Sussmann, Lie brackets, real analyticity and geometric control theory, in Mathematical Control Theory, R.W. Brockett, R.S. Milimam and H.J. Suss- mann eds., Birkhauser, Boston Inc. (1983), pp. 1-115. [85] H. J. Sussmann, The structure of time-optimal trajectories for single-input system in the plane: The C°° nonsingular case, SIAM J. Control, 25 (1987), pp. 433-465. [86] H.J. Sussmann, A general theorem on local controllability, SIAM J. Control, 25 (1987), pp. 158-194. [87] H.J. Sussmann and V. Jurdjevic, Controllability of nonlinear system, J. Diff. Equations, 12 (1972), pp. 95-116.
References 309 [88] D. V. Tran, M. Tsuji and D. T. S. Nguyen, The Characteristic Method and its Generalizations for First Order Nonlinear Partial Differential Equation, Chapman&Hall/CRC, 1999. [89] T. Wazewski, Systemes de commande et equations au contigent, Bull. Acad. Polon. Sci., 9 (1961), pp. 151-155. [90] L.C. Young, Calculus of Variations and Optimal Control Theory, Chelsea Publ. Co., New York, 1980. [91] J. Zabczyk, Mathematical Control Theory: an Introduction, Birkhuser, Boston, 1992. [92] V. Zeidan, Sufficient conditions for the generalized problem of Bolza, Trans. Amer. Math. Soc., 275 (1983), pp. 561-586.

Index absolutely continuous functions, 276 algebraic Riccati equation, 154 asymptotic stability, 199 Banach space, 264 Banach’s contraction mapping theorem, 265 Bang-Bang theorem, 69 Bolza problem, 93 Brouwer theorem, 267 Caratheodory solution, 13 Caratheodory theorem, 284 Cauchy problem, 14, 37 Cauchy sequence, 263 chattering controls, 66 closed loop control, 3 constrained optimization, 291 control system, 1, 35 controllability matrix, 56 convex cone, 288 convex function, 135 convex hull, 284 convex set, 263, 283 differentiability w.r.t. initial data, 26 differential inclusion, 2, 36 direct method, 90 dynamic programming, 137, 186 dynamic programming principle, 186 Euler-Lagrange and Weierstrass necessary conditions, 121 feedback, 3 fit for jumps, 240 Frobenius theorem, 298 fundamental matrix solution, 24 graph completion, 249 Gronwall’s lemma, 17 Hamilton-Jacobi-Bellman equation, 189 Hausdorff distance, 280 impulsive control system, 233 input-output map, 38 Lebesgue dominated convergence theorem, 275 Lie algebra, 62 Lie bracket, 62, 294 linear O.D.E., 21 linear system, 11, 56 linear-quadratic problem, 125, 150 Lipschitz continuous, 265 lower semicontinuous, 274 Lyapunov function, 76 Lyapunov stability, 76 Lyapunov’s theorem, 277 Mayer problem, 88 minimum time problem, 121 multifunction, 36, 280 necessary condition, 10, 99 needle variation, 101 normed space, 263
312 Index O.D.E., 13 open loop control, 3 optimal control, 9 optimal synthesis, 155 ordinary differential equation, 1, 13 partial differential equation, 165 patch, 201 patchy feedback, 203 patchy vector field, 201 pole shifting, 80 Pontryagin Maximum Principle, 10, 100, 111, 116 Rademacher’s theorem, 278 reachable set, 8, 51 Riccati differential equation, 151 small time local controllability, 9, 60 stability, 76, 207 stabilizing feedback, 79 strongly fit for jumps, 240 sub-differential, 170 sufficient condition, 10, 133 super-differential, 170 trangent cone, 288 transversality, 30 value function, 186 viscosity solution, 175