UFDC Home  myUFDC Home  Help 



Full Text  
NEEDBASED FEEDBACK: AN OPTIMIZATION APPROACH By DEBRAJ CHAKRABORTY A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2007 O 2007 Debraj Chakraborty To my family. ACKNOWLEDGMENTS I would like to thank my advisor, Dr. Jacob Hammer for his kind and patient guidance during my stay at University of Florida. Our long discussions, whether technical or not were intellectually stimulating and extremely enj oyable. I thank all the faculty members of my department who helped me learn the basics of Electrical Engineering. Especially I would like to thank Dr. Pramod Khargonekar, Dr. Haniph Latchman and Dr. John Schueller for serving on my PhD committee and for improving my thesis through some insightful suggestions and corrections. Lastly I would like to thank my friends Yoav Levinbook, Saravanan Vij ayakumaran, Niranj an Venkatraman and Jun Peng for comments, suggestions and especially for listening to my halfbaked ideas and theories at various stages of their development. TABLE OF CONTENTS page ACKNOWLEDGMENTS .............. ...............4..... LIST OF FIGURES .............. ...............7..... AB S TRAC T ......_ ................. ............_........8 CHAPTER 1 INTRODUCTION .............. ...............10.... 2 LITERATURE REVIEW .............. ...............15.... Related Work in MinMax Optimal Control Problems .............. ...............15.... Other Approaches for Reducing Feedback Requirements. ..........._..._ ................. .......17 Residence Time Control .............. ...............17.... 3 MATHEMATICAL PRELIMINARIES ................. ...............19........... .... Normed Linear Spaces............... ...............19. W eak Convergence............... ..............1 Alaoglu' s theorem .............. ...............20.... Weak upper semicontinuity ................. ........... ...............20...... Definition ofBorel, Lebesgue and Radon Measures................ ...............2 Riesz Representation Theorem ............_ ..... ..__ ...............22... Fubini's Theorem .............. ...............22.... Separation of Convex Sets............... ...............23.. Directional (Gateaux) Derivative .............. .... .. ...............23. Existence and Uniqueness of Solution to a LTI System ................. ................. ..........24 Jordan Canonical Representation .............. ...............24.... 4 NOTATION AND PROBLEM FORMULATION ................ ............. ......... .......26 Notation ..................... ...............26. Statement of the Problem ................. ...............28................ 5 EXISTENCE OF A SOLUTION................ ...............3 Weak Compactness of U .............. ...............30.... Alternative Problem ............... ... ............... ...............31....... Weak Upper Semicontinuity of T(u,D) .............. ...............37.... Existence of u* ................ ...............39........... .... 6 GENERALIZED FIRST ORDER CONDITIONS ................. ...............41........... ... Simplified Theorem from Warga (1970) ................. ...............41........... .. The Scaled System ................. ...............45........... .... Necessary Conditions .............. ...............47... BangBang Approximation to v* ............. ...............53..... 7 CONDITIONS FOR A PURELY BANGBANG SOLUTION ................. ............. .......57 8 CONCLUSION AND FUTURE WORK .............. ...............67.... Conclusion ............ _. ..... ...............67... Future Work............... ...............68.. Applications............... ..............6 Theoretical Research............... ...............69 LIST OF REFERENCES ............_...... ...............71... BIOGRAPHICAL SKETCH .............. ...............75.... LIST OF FIGURES Figure page 11 Schematic of intermittent feedback control ................. ...............12............... 12 Loss of feedback due to communication channel disruption ................. ......................13 61 Optimal input has one switch: M= 1.96, tf= 3.7............... ...............5...1 62 Trajectories for ten different uncertainty values of 'a'; M = 1.96, tf = 3.7 .................. .......52 63 Approximate bangbang input: 16 switches .............. ...............56.... 64 Trajectories for ten different uncertainty values of 'a'; M = 1.96, tf = 3.69 .................. .....56 71 Optimal input has one switch: M = 25, t, = 5.08 .............. ...............66.... 72 Trajectories for ten different uncertainty values of 'a'; M = 25, tf = 5.08 ................... .......66 Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy NEED BASED FEEDBACK: AN OPTIMIZATION APPROACH By Debraj Chakraborty August 2007 Chair: Jacob Hammer Major: Electrical and Computer Engineering Feedback is often used to overcome the adverse effects of perturbations and uncertainties on the performance of engineering systems. However, failures of the feedback channel cannot be completely avoided. This dissertation addresses the questions of how and for how long can desirable performance of a perturbed system be maintained after a failure of the feedback channel . Let Es be a system that is subject to a perturbation E in its parameters. The exact value of the perturbation E is not known; it is only known that s is bounded by a given constant 6. Now, let u(t) be an input function of C, and let Esu be the response of the perturbed system to the signal u(t). The nominal system is Co, and the nominal response to the signal u is Cou. Therefore, the deviation in the response caused by the perturbation is L~u Cou. To reduce the perturbation, add a "correction signal" v(t) to the input signal, so that the perturbed response becomes Es(u+v). Then, the new deviation between the perturbed and nominal cases becomes Es(u+v) Cou. The correction signal v(t) must be independent of perturbation value s, as the latter is not known. Let M be the maximal deviation allowed for the response, and let tf be the time for which Es(u+v) Cou < M. Then, the objective is to find a correction signal v(t) that maximizes tf, given only that the perturbation E is bounded by 6. EulerLagrange type first order conditions for calculating the optimal correction signal v(t) is presented. It is shown that, under rather broad conditions, the optimal correction signal v(t) is either a bangbang signal or can be arbitrarily closely approximated by a bangbang signal. CHAPTER 1 INTTRODUCTION In this work we reduce the need to communicate between the controller and the sensor measuring the system output by maximizing the time during which the feedback loop can remain open. This is motivated from the fact that in certain applications, feedback is not continually available or cannot be continually measured. Sometimes it is advantageous to temporarily stop the transmission of the feedback signal from the sensors at the system output to the controller. In other situations an unpredictable failure of the feedback channel can occur. For example, in controlling space vehicles, obstacles may accidentally disrupt the line of vision between the spacecraft and earth for varying time periods. In telemetry, the need to conserve battery life may motivate a planned reduction of the time of transmission of feedback signal. Moreover stealth applications, that hide systems from detection, prefer to minimize transmission between the controlled obj ect and the remote controller. In agriculture, measurements about soil parameters like moisture, etc., must be done manually and are consequently expensive. Usually such measurements are carried out intermittently after long intervals. In medicine, mathematical modeling and control of certain diseases have become quite common (see Panetta 2003 and references therein). However the feedback available is inherently of an intermittent nature, since measurements on the patients can only be made after long intervals. Lastly network based control systems use a common unreliable network on time shared basis with other users and hence the feedback signal may be available to the controller only intermittently. In such applications, it is relevant to ask: * Research Question 1: How long can the feedback loop be kept open, while maintaining desirable performance specifications? * Research Question 2: What is the best way to control the system when we do not have the feedback signal? Feedback is necessary due to the uncertainty inherent in any system. The main reason for this need for feedback is the lack of perfect knowledge about the system. This includes modeling inaccuracies, unmodeled nonlinearities, parametric uncertainties, spurious disturbances, input noise and measurement noise. Robustness and immunity to all this features are usually achieved through feedback strategies. Hence, it is evident that no uncertain system can be controlled indefinitely if the feedback signal is not available. However, assuming some amount of tolerance for error, there will be an interval of time in which the system will perform acceptably even without any feedback. Let us assume that the system is allowed to operate within some specified level of error and the loss of feedback occurs when the error is within this specified limit. Then the system traj ectory will usually require at least a Einite amount of time to exceed the specified error level. This is true for most real systems and in particular, for linear time invariant systems. As we illustrated in the above applications, it is interesting to know, how and for how long this period of acceptable operation without feedback may be extended. The question we ask is: what is the maximal time until which the feedback loop can be kept open so that the error remains within the specified limit. We will show that under suitable assumptions, this period is always Einite. Hence the disruption in the feedback signal, whether intended or accidental, will have to end or must be ended after this period of time. Otherwise, effective control is not possible, and the system traj ectory may exceed the tolerated error. If the disruption was accidental, this maximal time describes the upper bound of guaranteed safe operation. If the disruption carries on beyond this point the system can potentially fail. The a priori knowledge of this critical maximal time interval can help in decisionmaking. If the suspension of feedback was intentional, at that point of time the feedback is reconnected and the knowledge of the current outputs/states are utilized to bring down the system error. Once the error is reduced to near zero, the feedback may be disconnected again and the previous cycle may resume. Controller I 1 Plant Switch Figure 11: Schematic of intermittent feedback control In this study we consider linear time invariant systems with bounded inputs and the states are assumed to be available as outputs. As in most common situations, we assume that the system parameters are uncertain but are known to lie within some bounds. The Euclidean norm of the state is taken as a measure of the system error at any time instant, which is required to be always less than a prespecified upper bound. The norm of the initial condition is assumed to satisfy this bound. Under these assumptions the obj ectives for the open loop operation may be outlined briefly as follows. Find, for any permissible uncertainty, the maximal time interval for which the system error does not exceed the allowed limit. The system error is not monitored during open loop operation. Hence, such a worstcase optimization is considered to guarantee acceptable performance. A special input function is calculated to achieve maximal duration of the open loop period. Note that this is legitimate even when the loop is open, as long as no knowledge of the feedback signal is used for its computation. Existence of the time optimal controller is proved and EulerLagrange type conditions are derived for calculating the time optimal open loop input. It is shown that this time optimal input signal is either purely bangbang or can be uniformly approximated by a bangbang signal. Finally, it should be noted that, corresponding to the optimal open loop input, an infinite number of system traj ectories might be realized depending on the particular value the uncertainty takes within the allowed set The main emphasis of this formulation is to guarantee that none of the trajectories, from the infinite number possible, can exceed the allowed error bound for the maximal time interval of operation. Loss of Feedback  Signal ,/ Acceptable Flight // Envelope \ ~ Earth \\ Launch arget Site Figure 12: Loss of feedback due to communication channel disruption We conclude this introductory chapter with a hypothetical example to illustrate our problem formulation: Consider the simulated firing of a groundtoground missile as illustrated in the figure. The flight path of the missile is predetermined and the missile is regulated carefully so that it follows the prescribed flight path. The feedback about the current coordinates, velocity, angles, etc., of the missile is transmitted to the ground controller via a satellite. The controller uses this information to determine the angle and amount of thrust needed to keep the missile on track. Under this scenario even a temporary loss of signal from the satellite may lead to the catastrophic failure of the control and a subsequent loss of the missile. We are proposing a solution to the problem by defining the best way to design the control of the missile for the duration of no feedback. Our method also guarantees that the missile can stay within the acceptable flight envelope for a maximal time after loss of signal from the satellite occurs. We prove that the best control under these circumstances is either bangbang or can be approximated by a bangbang signal. CHAPTER 2 LITERATURE REVIEW Related Work in MinMax Optimal Control Problems The obj ective of time optimality, for every possible uncertainty, in open or closed loop operation embeds our problem within the framework of minmax optimal control. This area of differential game theory has been extensively researched and encompasses a wide variety of interesting results. We review a few papers directly related to our work. Isaacs (1954) was the first investigation to formulate a control problem in which two players with conflicting interests interact with each other in a game theoretical setup. These results were later published as a book in Isaacs (1965). The socalled "main" equation was derived, which can be viewed as a game theoretical extension of the HamiltonJacobiBellman equation and the dynamic programming approach (Bellman 1957). Independently, Kelendzheridze (1961) solved a pursuitevasion problem, which in turn spawned a body of research of which Pontryagin et al. (1962), Pontryagin (1966, 1967a and b), Mishchenko and Pontryagin (1967) and Mischenko (1971) were early contributors. These papers generally derived EulerLagrange type necessary conditions for the equilibrium solutions to the pursuit evasion problem. Berkovitz (1964 and 1967) formalized Isaacs' results in a classical calculus of variations setting and also derived necessary conditions for general systems under weak assumptions. Other early contributors were Fleming (1961 and 1964), Ho et al. (1965), Friedman (1971), Elliott and Kalton (1972) and Elliott et al. (1973). Numerous authors developed different definitions of the equilibrium solution and various payoff functions of which Linear Quadratic Games have been extensively developed. A complete set of references can be found in Basar (1982). However the maj ority of interesting results have been concentrated for differential games with a saddle point solution. The problem of worstcase optimization without the saddle point assumption was posed by Feldbaum (1961 and 1965). This problem was addressed by, among others, Koivuniemi (1966), Howard and Rekasius (1964) and Bellanger (1964). Witsenhausen (1968) solved the same problem with a convex cost functional for sampled data systems. As discussed next, the problem was mathematically solved by Warga (1965a). The concept of relaxed solutions in optimal control theory was introduced by Young (1937) and later in the book Young (1969). In a series of papers Warga (1965a and b, 1970 and 1971a and b) solved the minmax control problem using relaxed solutions with quite general assumptions. The existence of relaxed solutions (in the sense of Young 1937) to the minmax problems were guaranteed and the solutions were found to satisfy variants of EulerLagrange type necessary conditions. The techniques used were similar to those of Neustadt (1966 and 1967) and Gamkrelidze (1965). These results were reworked and accumulated into a book (Warga 1972), from which we use a result on conflicting controls (Theorem IX. 1.2) to derive necessary conditions for our problem. We quote a simplified version of the actual theorem in Chapter 4. It should be noted that with the exception of papers related to the Isaacs "main" equation, almost all the work reviewed above looked for open loop solutions. Apparently little has been done in search of closed loop necessary conditions characterizing solutions to the minmax problem. Conditions for existence of a stable solution were formalized with the help of concepts like stable bridges in Krasovskii and Subbotin (1974). This book contains a detailed exposition of "positional", i.e., feedback solution to differential games. In a series of papers, Ledyaev and Mishchenko (1986, 1987 and 1988) and Ledyaev (1989 and 1994) derived necessary conditions for minmax control problems of a fixed duration. This has been an area of intensive research since but most of the advances have been related to the dynamic programming approach. See Vinter (2000) and the references therein. However efforts to derive EulerLagrange type necessary conditions have been limited. Throughout this study, results and ideas from standard optimal control and mathematics references have been used. Some of them are Kolmogorov and Fomin (1957 and 1961), Liusternik and Soboley (1961), Rudin (1966), Bryson and Ho (1969), Luenberger (1969), Balakrishnan (1971), Hirsch and Smale (1974), Halmos (1982) and Zeidler (1985). Other Approaches for Reducing Feedback Requirements As outlined in the introduction, the main obj ective of this work is to reduce the duration during which the feedback loop has to be closed for controlling an uncertain plant. As far as we are aware of, such a robust time optimal formulation for reducing feedback requirements has not been dealt with in literature. However, in the context of Network Control Systems (see Nair et al. 2007), model based approaches have been developed for controlling systems with intermittent feedback. Design methodologies, in which ideal models of the system were used to guess the output of the system during the open loop, were proposed by Zhivogyladov and Middleton (2003) and Montestruque and Antsaklis (2004). Estimates of the maximum time the feedback signal may be delayed were calculated in Walsh et al. (1999). However in these works, the problem of finding the best open loop input was not dealt with and consequently, the maximum time for acceptable open loop operation has not been calculated. In conclusion, our work may have some implications for the problem of bandwidth reduction or control under communication constraints usually addressed in Network Control Systems (see Nair et al. 2007). Residence Time Control The problems of pointing and the related concept of residence time control are similar to our formulation of the problem. (See Meerkov and Runolfsson 1988 and the references therein). The pointing problem was investigated among others by Skelton (1973), Eng (1979), Cannon and Schmitz (1984) and Halyo (1983). In Meerkov and Runolfsson (1988), the residence time control problem was formulated as that of choosing a feedback control law, so as to force the system states to remain, at least on the average, within prespecified limits during some minimal period of time in spite of the disturbances that are acting on the system. The objective is very similar to that treated in this work, but the use of feedback creates a fundamental difference. Recall that in our hypothesis the feedback signal is completely absent over the period of interest. Moreover the maximum residence time was also not calculated in these papers. CHAPTER 3 MATHEMATICAL PRELIMINARIES In this chapter we will cover some of the wellknown results from functional analysis, measure theory and linear algebra that have been used repeatedly in this dissertation. This chapter is meant as an easy reference for the results derived in the later chapters, and is an attempt to make this thesis tolerably selfcontained. We start by a selection of relevant results from the theory of normed linear spaces. Most of these standard results are taken from Liusternik and Soboley (1961), Halmos (1982) and Zeidler (1984). Normed Linear Spaces Weak Convergence First we recall the definition of weak convergence in a normed linear space. Definition 3.1: Let E be a normed linear space, {xn} a seque~nce of elementsofEl andT xoEE. If for all functionals feE* (where E* is the space conjugate to E), the sequence f(xn) 4 f(xo) as namo, then we say that {xn} converges weakly to xo and we write xn 4i xo. Here xo is the weak limit of the sequence {xn). Since Hilbert spaces are selfconjugate, the above definition may be specialized to the following form for Hilbert spaces: Definition 3.2: A sequence {xn i;n a Hilbert space H converges weakly to xoEH if of the sequence {xn). A continuous functional on a compact set in a normed linear space is bounded and achieves is maximum and minimum. However, in this dissertation, most of the sets we would be interested in are not compact in the sense of strong convergence. Weak convergence and weak compactness are much less severe requirements and will shown to hold for the sets of interest in this work. Weak compactness is defined next. Definition 3.3: A set XCE is weakly compact (i.e., sequentially weakly compact) if ever in~ite equece {n} E has a subsequence {xnk} such that xnk Xo and xoEX. The existence of solutions to optimal control problems is very much dependent on the following very important theorem. Alaoglu's theorem Theorem 3.4: Every bounded sequence of elements in a Hilbert space contains a weakly convergent subsequence. This theorem guarantees the weak compactness of bounded sets in the Hilbert space. However for compactness in itself, the set of interest must be closed in the weak topology. The following result is useful for checking the sequential weak closure of convex sets: Theorem 3.5: A bounded strongly closed convex set in a Hilbert space is also weakly closed. In the analysis of existence of optimal solutions to dynamic optimization problems, we are actually interested in the functionals defined on compact or weakly compact sets. We define the concepts of weakly continuous and weakly uppersemicontinuous functionals. Weak upper semicontinuity Definition 3.6: A functional f (possibly nonlinear) on a Hilbert space H is defined to be weakly continuous at xoeH, if for any sequence xnEH and xn 42 xo, we have f(xn) 4 f(xo). Definition 3.7: A functional f (possibly nonlinear) defined on a Hilbert space H is defined to be weakly upper semicontinuous at xoeH, if for any sequence xnEH and xn 42 xo, implies lim sup f(xn) < f(xo). A generalization of the classical theorem of Weierstrass to semicontinuous functionals is often used to prove existence of solutions in optimization problems. A simplified statement is listed next: Theorem 3.8: Let f be a weakly upper semicontinuous real valued functional on a weakly compact subset S of a Hilbert space H. The functional f is bounded above on S and achieves its maximum on S. In differential game theory it is often not enough to study a single upper/lower semicontinuous functional but a family of such functionals. Next we state a very useful property of a family of upper semicontinuous functionals: Theorem 3.9: (Willard 1970) Let X and A be two topological spaces. If f, is a upper semicontinuous real valued function on X for each oceA, and if inf f,(x) exists at each xeX, then the function f(x) = inf f,(x) is upper semicontinuous on X. Definition of Borel, Lebesgue and Radon Measures The interplay between measure theory and functional analysis is extremely common in dynamic optimization and especially differential game theory. We briefly define a few standard and wellknown concepts for easy reference. For detailed exposition on these concepts see Kolmogorov and Fomin (1957) and Rudin (1966). If S be a topological space, then the smallest 0field containing every open set in S is called the Borel field of sets, and denoted C(S), and the elements of C(S) are called Borel sets. A measure defined on C(S) is called a Borel measure. Let C1 be a Borel measure. A Borel set EE C(S) is regular if both of the following properties hold: * C1(E) = inf {p1(V): EcV, V open} p(E)= sup {T(K): KcE, K( cmompact whe~never E is open or p(E) If every Borel set in S is regular then C1 is called regular. A Einite regular Borel measure is called a Radon measure. Radon measures are important from the point of this work, because of their appearance in the Riesz representation theorem for characterizing linear functionals. Finally, the Lebesgue measure on the real line R can be defined simply as follows: Let C(R) be the Borel Hield defined on the real line and consisting of intervals. Then the set function that assigns the interval [a,b], the measure (ba) is called the Borel measure on C(R). The completion of C(R) relative to the Borel measure is called the class of Lebesgue measurable sets. The extension of the Borel measure to the completion of C(R) is called the Lebesgue measure . Riesz Representation Theorem We present here a special form of the wellknown Riesz Representation theorem, which will be used in the later chapters. This specific form is a slight simplification of that presented in Evans and Gariepy (1992). Theorem 3.10: Let P be a compact subset of Rm. Denote the space of continuous functions mapping P to R by C(P,R) and each element in this space by f.Let L be a bounded linear functional on the space C(P,R). Specifically L: C(P,R)4R. Then there exists a positive Radon measure C1 on P and a C1measurable function h: PaR such that: * 1(x) =1 for C1a.e. xeP. * L~f = Sh f dC for all fe C(P,R). Fubini's Theorem Theorem 3.11: Let (X,Cx,C1) and (Y,Cy,h) be ofinite measure spaces, and let f be an (Exx~y)measurable function on XxY. If 0 < fl <0, and if $(x)= Sfxdh and V(y) = .ffy dy (xeX, y eY) then m is EXmeasurable, 'F is Ey measurable, and Sm d,= cfd(lxh) = SI d1. X XxY Y Separation of Convex Sets Let H be a topological vector space. A set A cH is convex if {ua + (1u)b) 0 < a < 1) cA whenever a,b eA. We say that a linear functional QEH* (where H* denotes the space conjugate to H), separates subset A and B of H if 40 and either f(x) <; a < Q(y) or f(x) > a 2 f(y) (xeA, yeB) for some aeR. Theorem 3.12: Let K and M be nonempty disjoint convex subsets of a topological vector space H and the interior of K is nonempty. Then there exists QEH* that separates K and M. Directional (Gateaux) Derivative Definition 3.13: Let X be a vector space, Y a normed space, and T a (possibly nonlinear) transformation defined on a domain DcX and having range RcY. Let xoeDcX and let h be arbitrary in X. If the limit: T(xo+ah) T(xo) 'DT(x;h) = lim exists, it is called the Gateaux differential of T at xo with increment h. If the above limit exists for each heX, the transformation T is said to be Gateaux differentiable at xo. We note that this definition makes sense only when xo+ah E D for all a sufficiently small and that the limit is taken in the usual sense of norm convergence in Y. If Y is the real line and the transformation T is a real valued functional on X, then the following definition may be used for the Gateaux differential: 'DT(x;h) duc T(xo + ath)a=o for each fixed xoeX. Existence and Uniqueness of Solution to a LTI System In this section we quote a widely known theorem about the existence and uniqueness of the solution to a differential equation with measurable right hand sides. In this dissertation, we consider linear systems of the form k(t) = Ax(t) + Bu(t) where u(t) is Lebesgue measurable. This theorem is used throughout to guarantee existence and uniqueness of solution for the systems under consideration. This particular version of the theorem is taken from Young (1969). Theorem 3.14: Let f(t,x) be a vector valued function with values in xspace, and suppose that in some neighborhood of (to,xo), f(t,x) is continuous in x for each t, measurable in t for each x, and uniformly bounded in (t,x). Then there exists an absolutely continuous function x(t), defined in some neighborhood of to, such that x(to) = xo and that, almost everywhere in that neighborhood: k(t) = f(t,x(t)) Suppose in addition that for some constant L, the function f(t,x) satisfies, whenever (t,xl) and (t,x2) lie in Some neighborhood N of (to,xo), the Lipschitz condition f(t,x1)f(t,x2) < L~x1x2, then in some neighborhood of to there exists one and only one absolutely continuous function x(t) such that: x(t) =xoi + /f(T,x(r)) dr. Jordan Canonical Representation Theorem 3.15 (Hirsch and Smale 1974): Let A be a nxn matrix with real elements. Then there exists a nxn real matrix P such that Ad= PAP is composed of diagonal blocks of the following two forms: h 1 F I, St=or S2 ~ 1 1 I The diagonal elements denoted by h is a real eigenvalue of A and appears as many times as the multiplicity of the minimal polynomial. In the expression for S2, F=a b _1 r 10 F =b andu I2 o 11 where (a + ib) are the complex eigenvalues of A. Each block F is repeated as many times as the multiplicity of the eigenvalue a + ib in the minimal polynomial. Ad is called the Jordan real canonical form of the matrix A. CHAPTER 4 NOTATION AND PROBLEM FORMULATION Notation The systems we consider are linear time invariant continuoustime systems given by a realization of the form (4.1) x(t) = A'x(t) + B'u(t) x(0) = xo. Where A' and B' are constant real matrices of dimensions nxn and nxm respectively. We assume that the state x of the system is available as output. The initial condition of the system is denoted by xo, and is given. In addition, we assume that there is an uncertainty about the values of the entries of the matrix A' and B'. To represent these uncertainties, we introduce two matrices DA and DB. We denote (4.2) A' = A + DA and B' = B + DB, where A, B, DA and DB are giVen COnstant matrices of appropriate dimensions. While A and B are known, the matrices DA and DB can take any values within the sets AA and AB respectively. We define AA and AB to be bounded subsets of the spaces of real (nxn) and (nxm) matrices such that AA= {DA: DA < d} and AB= {DB: DB < d}. Here the norm D = s~up Dgl where D, denotes the ij"th element of the matrix D and d is a fixed positive real number. Note that the sets AA and AB are ClOsed, bounded and hence compact. The particular values DA and DB takes in the sets AA and AB are HOt known a priori and hence represents the uncertainty in the system matrices A' and B'. Such an assumption is realistic because though the exact values of the system parameters are never known, usually the sets in which they belong are known a priori. For compactness of expression we group the uncertainties together and define D = (DA,DB) belonging to the set A = AAX B. It is further assumed that at least one eigenvalue of the nominal system matrix A has a nonnegative real part and that the pair (A',B') is stabilizable for every value of the uncertainty set A. We will show in Chapter 5 that the assumption that the nominal system is unstable guarantees the existence of a finite solution to the open loop problem. We assume that the system is performing acceptably if the Euclidean norm of the states, representing the system error, stays below a prespecified positive bound (say M). In other words, the inequality xT(t)x(t) < M, must hold as long as the system is operating. However once the state norm exceeds the bound, the system must either be stopped or some action must be taken to prevent this from happening. Here we assume that at the point in time when the feedback signal stops, the system norm denoted as the initial condition in our problem, is less than M. (4.3) x0TXO < M Then the open loop is allowed to run as long as xT(t)x(t) <; M, and the loop must be closed when this inequality can be no longer made to hold. As outlined in the introduction, we assume that there is a special input during the time when the feedback signal is absent. This input is used to keep the system error within the tolerated level for a maximal amount of time. We denote this input by u(t). For u(t) we will use weighted norm and inner product as described below. Here u(t) is a vector valued function (ul(t), ..., um(t))T, where each component is real valued and Lebesgue measurable. Each u(t) is assumed to lie in the Hilbert space L~r with the following inner product: let x(t) and y(t) be two elements of L Then the inner product some Eixed constant oc > 0. The norm and the metric are defined correspondingly. Let U be a set of bounded functions in L~ defined as Us { u E L~ : maxi ui(t) < K, t E [0,00]} where K is a fixed positive number. This set U of bounded measurable functions define the prospective inputs we shall consider for system (4.1). As will be shown in Chapter 4, the optimal solution is bangbang with possibly an infinite number of switches. Hence a smaller set, e.g., the set of piecewise continuous bounded functions, would not suffice. The set U thus defined will be shown to be compact in the weak topology and thus will facilitate the proof of existence of the optimal solution. For engineering purposes, any element in U can be approximated to any arbitrary accuracy by a piecewise continuous implementable function. Statement of the Problem We next state the mathematical problem formulation. During the open loop period the obj ective is to maximize the time during which the system error is guaranteed to stay below certain bound in the face of uncertainty. This obj ective is achieved with the help of the special input u(t) used to correct the system during the time the feedback is absent. In the introduction we described that how in various applications it is important to know both the correcting open loop input u(t), and the maximal amount of time for which the system can be left running in open loop safely. Keeping this motivation in mind we pose the following problem: Problem Statement: Using the notation of the last section the open loop problem may be formulated as follows. Problem 4.4: Find max tr subject to the following constraints: ueU x(t) = A'x(t) + B'u(t) 0 < t < tr x(0)= xo x0TXO < M xT(t~x(t) < M for 0 < t <; tr and for all (A',B') E (A+AA)x(B+AA) where (A+AA) t A+ DA : DAE A} a~ndl (B+AB) = {B+DB: DBEr B Among the main obj ectives of this work is to show that such a maximum tr exists, is finite and that it can be achieved by an input u(t)EU. Assuming they exist, the interval [0,tf] and the optimal input solution u(t) to the above problem has the following feature. If u(t) is applied over the interval [0,tf], irrespective of the value the system matrices (A',B') takes in the sets (A+AA) and (B+AB), the system trajectory stays within the allowed error level, i.e., xTOt)x(t) < M for the entire interval [0,tf]. The second objective is to find the nature of the optimal input u(t) that achieves the maximum tf. This will be achieved through the derivation of first order necessary conditions characterizing the optimal solution. CHAPTER 5 EXISTENCE OF A SOLUTION In this chapter we prove the existence of a solution to Problem 4.4. This will be achieved through the application of the generalized Weierstrass theorem stated in Chapter 3 (Theorem 3.8). Before going into the mathematical details we describe a brief outline of this chapter. The main idea is to repose Problem 4.4 in turns of a suitably defined upper semicontinuous functional. This functional is nothing but the minimum possible time (corresponding to the worst uncertainty) of open loop operation for any given ueU. We prove that the set U defined in chapter 4 is compact in the weak topology and then we show that the functional mentioned above is weakly upper semicontinuous. Thus we are ready to use a generalized Weierstrass theorem to prove that a maximal time for the open loop operation exists, is finite, and can be achieved with an input u from the set U we defined in Chapter 4. Weak Compactness of U We start with a few properties of the input set U defined in Chapter 4. Most of the standard results of this section are taken from Liusternik and Soboley (1961), Halmos (1982) and Zeidler (1984) and has been listed for reference in Chapter 3. We prove the weak compactness property of the set U of inputs, in the form of a lemma . This property, which is a consequence of the Alaoglu' s theorem (Theorem 3.4)), along with a few wellknown properties of semicontinuous functions on compact sets, is essential for proving the existence of a solution to Problem 4.4. Lemma 5.1: The set U is weakly compact in the Hilbert space L ~. Proof: The set U is obviously bounded and hence by Theorem 3.4 every infinite sequence in U has a weakly convergent subsequence. However we need to show that the set U is weakly closed. Since U is convex by definition, by Theorem 3.5 we just need to show that U is strongly closed. Consider a sequence of functions un(t)4uo(t) where uo(t)0U. Hence for some set St'c[0,c) of nonzero measure, uoi(t)>K, where te6t' and uoi denotes the i.th element of uo(t). Hence the following inequality must hold: featun(t)uot)2dt = Seatun(t) uo(t)2dt + Cn where Cn, is positive real. O 6t' > fe'* (K uoi(t))2 dt + Cn, As namo, while Cn can tend to zero, the first term remains constant and finite. So un(t) does not converge to uo(t). Alternative Problem We now define an alternative formulation for Problem 4.4 using the functional defined next. It is easier to show the existence of the optimal solution in terms of this functional. Definition 5.2: Let J(t) = x(t)Tx(t) where to [0,co). Define Sinf t:J(t) > M if J(t) > M for some t Here x(t) is related to u(t) and D through Equation 4. 1. This functional may be interpreted as the time when the system error exits the allowed envelope of operation for the first time. This functional is identical to the quantity tr defined in Problem 4.4, for trajectories obeying the constraints. In other words, for a fixed (u,D) pair, the relation x(t)Tx(t) < M is satisfied for the interval [0,T(u,D)]. Moreover for a fixed u(t), the relation x(t)Tx(t) < M is satisfied for every value of the uncertainty DEA within the interval [0, inf T(u,D) Hence it can be easily seen that Problem 4.4 is equivalent to finding sup inf T(u,D), if it is finite. In addition we need to show that the supremum over U can be achieved. We summarize in the following restatement of Problem 4.4: Problem 5.3: Show that sup inf T(u,D) < co. Find u EU such that inf T(u*,D) = sup ueU DeA DeA ueU inf T(u,D). DeA Well posed: For this problem to be well posed, the first step is to show that inf T(u,D) < DeA co. This observation is crucial to us since we are interested in running the system on open loop as long as possible. We realize that no matter what input we create for the system, nature can choose a value for DE A that will make the system traj ectory leave the set of acceptable operation in finite time. First we prove a preliminary lemma: Lemma 5.4: Let A and D be a real nxn matrices such that A has repeated eigenvalues and DEA. Also let xo be a fixed nonzero vector in Rn. Then for any d > 0, there exists a D such that D < d and (A+D) has distinct eigenvalues. Moreover let T be a nxn real matrix such that Q= T(A+D)T" results in a Q which has the following block diagonal form: (5.5) Q=whr Qi= ; biai i for i=1,....,r and Qr+1=irllr+ . Here (ai + ibi), i=1,...,r represents the complex eigenvalues and ai (i=2r+1,....,n) represents the real eigenvalues of A. Then D can be chosen such that E Tlj xoj > where Tij is the ij"th element of T and xoj is the jth element of xo. Proof: Let Ad = PAP where Ad is the real Jordan canonical form of A. Recall the Ad is COmposed of diagonal blocks of the following two forms: (possibly repeated) of A. In the expression for S2, F = _b a wee(ti)ar h ope roots of A (possibly repeated). (Theorem 3.15) Let E be a diagonal matrix that we add onto Ad such that blocks S1 and S2 are changed as follows: h+s1 1 S1' = h+8q1 h+E Where let q be the multiplicity of the Ft = r I t~ FI2 ,I eigenvalue h. Where let s be the multiplicity of the eigenvalue (af ib) and Fi= ba + yayip~ i=1,2,...s Now the eigenvalues corresponding to block S1' and S2' are distinct and of the form (h + si) and (a + yj f ib) where i=1,2,...q and j=1,2,...s. (Note: det(hI S2') = det(hI Fl) det(hI Fz)..... det(hI Fs) ) Thus (Ad + E) has distinct eigenvalues. From the construction above it can be seen that it is possible to choose such an E such that E < r for any r > 0. Now consider the reverse transformation: P (~Ad + E)P = A + P 1EP h 1 z F F~twhere h is a real eigenvalue  P1EP < k P 1EP2 < k P 12 E2 P2 = kE2  The first inequality follows from the equivalence of norms where k is a finite constant and 2 is the spectral norm of the matrix. Clearly by making E2 k~ we get that  P EP < d. Hence by defining D = P 1EP we achieve our objective. Now let C Tlj xoj = 0 with a certain choice of A with distinct eigenvalues and with a T that transforms A according to Equation 5.5. Clearly, the vector T1 [T11 T12 .... Tin] is an eigenvector of the matrix A and hence must satisfy hlT1 = AT1 where hi is the eigenvalue corresponding to T1. Now from the continuity of the eigenvalues of A w.r.t. the entries of A, we can change A to (A+D) arbitrarily keeping the eigenvalues distinct. Then the following equation must hold: hi' T1' = (A+D)T1'. Since D is arbitrary, T1' can be made noncollinear to T1 keeping T1 T1' arbitrarily small. It follows that T1' and T1 cannot be both orthogonal to xo. Theorem 5.6: Consider the system (4.1) with a fixed initial condition xo in the set {xo: 0 < xoxo <; M}. Then for any fixed u(t)EU and M < co there exists a DEA such that T(u,D) < Proof: Assume that for some fixed u(t)EU, T(u,D) = co for all De A. This implies that x(t)Tx(t) < M VDEA and Vte[0,co] >  eA't [XO + A'rB'u(T)dZ]  < M. Now fix A' at some arbitrary value in (A + AA) Such that A' has distinct eigenvalues. By Lemma 5.4 this can be done without loss of generality. Then we do a similarity transform on A' such that Q = T 1A'T is in the block diagonal form shown in Equation 5.5. Note that in this form Q and QT are commutative. Now we consider the system transformed by T and name the new state vector z(t). It is related to x(t) by the following relation: z(t) = Tx(t). We further assume that T is such that zol = E To xoy r 0. By Lemma 5.4 this also can be done without loss of generality. We seek to prove that x(t) can be made divergent with a proper choice of DEA. We claim that it is enough to prove z(t) diverges. z(t) = Tx(t)i < Tx(t) . Hence lim z(t) = O > lim x(t) = CO In terms of z(t) our hypothesis implies: z(t)Tz(t) < T M = M' VD'EA and Vt [0,co] >  eat [zo + Se Bsru(Z)dr]  M' where M'= T M and B" = TB' Let F(t) =[zo + fe B"u(r)dr]. Therefore, [ eat F(t) ]T[ Qt F(t) ] < M' >F(t)T ,(QQT F(t) <; M' > F(t)T Lt F(t) <; M' Where E = QT Q is a diagonal matrix with at least one of the elements positive, by hypothesis. For simplicity assume that the 1st diagonal element denoted by al is positive. Let us denote fi(t) each entry of F(t) by fi(t), i.e., F(t)= ... nt~ Thus for the last equation to hold at least the following must be true. (5.7) lim fi(t)= 0 Now fi(t) may be written in a general form in the following way: m m fit=, zoI + at[cos(blz) E BIjuj(z) + sin(blz) BzjUj(T)] dz where (al + ibl) is the first eigenvalue of Q and Bij is the ij"th element of the matrix B". Then Equation 5.7 implies that m m lim zol + ea"[cos(blz) E Bijuj(Z) + sin(blz) E BzjUj(T)]dz = tom j=1 j=1 > zot + e ae cobi) Bljuj(z) + sin(blz) C BzjUj(T)]dz = 0 j=1 j=1 But this is a linear equation in Bij and hence cannot hold over all values of B" in the set T(B + AB) { T(B + DB) : DB E BE ) UnlCSS Uj(t) 0 for every j=1,...,m and for all to [0,00]. However if this is true then zol = 0. Hence we have a contradiction. In the above theorem we showed that, under the assumptions made in Chapter 4, the system norm x(t) can diverge for any given input u(t) for a proper choice of the uncertainty D. This is true for arbitrarily small uncertainty sets. Now define T'(u) a inf T(u,D), i.e., T'(u) DeA is the minimum time corresponding to u(t) in which the system norm x(t) can be made to escape the set [0,M]. According to Theorem 5.6 for any Eixed u(t)EU the quantity T'(u) is Einite . Corollary 5.8: If the conditions of Theorem 5.6 are satisfied then for any fixed u(t)EU, the inequality inf T(u,D) < co holds. DeA Weak Upper Semicontinuity of T(u,D) Recall that according to Problem 5.3 our obj ective is to show that sup in~f T(u,D) < x. Define T'(u) inf T(u,D). Thus it would be enough to show that T'(u) is weakly upper semi DeA continuous in u(t). Along with the fact that the set U is weakly compact (Lemma 5.1)), this would effectively prove that the functional T'(u) is bounded over the set U and that the supremum is achieved. (Theorem 3.8). The upper semicontinuity of the functional T(u,D) in u(t), for each fixed DEA is demonstrated next. Denote the set of solutions for Equation 4.1 for all ueU and DEA by X(U,A). By Theorem 3.14, each element of X(U,A) is an absolutely continuous function in t and unique corresponding to each (u,D) pair. Lemma 5.9: For a fixed DEA, T(u,D) is weakly upper semicontinuous in u(t), i.e., as a sequence of functions un(t) 4j uo(t), the functional T(u,D) obeys the following relation: for any E > 0, and T(uo,D) < co we can choose an integer N such that T(un,D) T(uo,D) < s for n > Proof: For a fixed DEA and t is the solution to Equation 4.1. As un a uo (i.e., weakly converges), then for each to [0,co), x(t;un) 4 x(t;uo), i.e., x(un) converges pointwise to x(uo). This is because for each to [0,co), x(t;u) is a functional linear in u. 1I if i xt) = eAt [XO+ j 1)OeA'B'u(T)dr]. Now define a functional T,: X(U,A) 4 [0,co] as follows: Sinf t: xTx > M if xTx > M for some t < co Tp(x(t))= = Next consider a sequence of function xo(t) 4 xo(t) pointwise. We show that for any E > 0 there exists an integer N such that for n > N T,(x,z) T,(xo) < E. If T,(xx,) < T,(xo) then the claim is true. So we assume that T,(xx,) > T,(xo). Let T,(x,z) = to and T,(xo) = to. Now we assumed that x,z 4 xo pointwise. By the definition of to, the following is true for every E > 0: there is a tl E (to, to+s) such that xTO(tl)xo(tl) > M. Consequently, there is N such that xTn(tl>xTn(tl) M > [xTO(tl)xo(tl) M]/2 > 0 for all n > N. Therefore, to < to + s which implies that ti, to < s Thus we have shown that T,(xx,) T,(xo) < E. Now, for a fixed DEA, consider the composition map T(u,D) :U 4 X 4 [0,co]. From the above arguments it is clear that as uzz uo, x,z 4 xo pointwise and for any E >0 there exists N such that for n > N, T,(xx,) T,(xo) < E. Hence T(uzz) T(uo) < s. Thus for a fixed D, T(u,D) is weakly upper semi continuous in u(t). Hence proved. Now we need to prove that the functional T'(u) is also weakly upper semicontinuous. This is true due to Lemma 5.9 and Theorem 3.9. Corollary 5.10: The functional T'(u) inf T(u,D) is weakly upper semicontinuous in DeA u(t). Proof: The proof amounts to checking the conditions of Theorem 3.9. By Lemma 5.9), for each DEA, the functional T(u,D) is weakly upper semicontinuous in u(t). Also for each ueU, the functional T'(u) = inf T(u,D) > 0. Hence the infimum exists always, and by the DeA property above T'(u) is weakly upper semicontinuous in u(t). Existence of u" The weak upper semicontinuity of T'(u) along with the weak compactness of the set of inputs leave us posed for applying the general Weierstrass theorem (Theorem 3.8). Hence we can now conclude the following: Theorem 5.11: Let U and A be as defined above and T'(u) inf T(u,D) and let DeA 6 = suR T'(u) suC inf T(u,D). Then 6 < 00. Moreover there exists some u (t)EU such that us u DeA T'(u ) = 6. Proof: This follows directly by noting that the set U is weakly compact and the functional T'(u) is weakly upper semicontinuous in u(t) over U. (Theorem 3.8). Hence T'(u) is bounded above and it achieves it upper bound over U. In conclusion, we showed in this chapter that a solution to Problem 4.4 exists within the set of inputs U. The set U was shown to be compact in the weak topology. Then an alternative formulation of the Problem 4.4 was stated using a functional T(u,D). A proof of the wellposed ness of this new statement was presented next. In the process it was shown that under the assumptions made in Chapter 4, the system traj ectory can leave the envelope of acceptable operation in Einite time no matter what input we apply to the system. This inbuilt finiteness of the functional T(u,D), along with particular features of the formulation was utilized to prove weak upper semicontinuity of the functional T(u,D) for any Eixed De A. Another functional, namely T'(u), denoting the worst possible time for the input u was introduced and it too was shown to be weakly upper semicontinuous in u. Lastly Weierstrass theorem was applied to prove the existence of the best input u* and the finiteness of 6 = supI T'(u). The application of u* guarantees that the system error does not leave the allowed envelope for any value of the uncertainty for a maximal duration of time. However, it should be noted that even for this optimal input, the system error can leave the allowed envelope in finite time for certain values of the uncertainty. CHAPTER 6 GENERALIZED FIRST ORDER CONDITIONS In Chapter 5 we proved that the optimal solution to Problem 4.4 existed among the set of bounded and measurable input functions. Evidently it would be useful to know more about the optimal solution, with a view to facilitate it's calculation for specific problems. The conventional method for calculating solution to optimal control problems has been through characterizing first order necessary conditions. Here however the situation is complicated by the gametheoretic formulation, in that the maximization of open loop time has to be done over every possible value of the uncertainty. We draw on the considerable amount of previous research in the area of differential game theory and minmax optimal control. The relevant literature was reviewed in Chapter 2. In particular, we note that Problem 4.4 is an example of minmax or conflicting control problems studied previously among others, by Warga (1970). We use Theorem IX. 1.2 (Warga 1970), for finding the first order necessary conditions characterizing the optimal solution. However we present a modified version of the theorem, which has been simplified to suit the problem we are solving in this dissertation. Simplified Theorem from Warga (1970) In this section we state and prove the theorem along with some required modifications. We require some new notation in this chapter: Notation 6.1: Let S be a compact set on Rn and C be the Borel field of subsets of S. Then by rpm(S) we denote the set of all Radon probability measures defined on E. Moreover we denote the Banach space of continuous functions from SaR with the supremum (L1) norm by the notation C(S,R). C denotes the closure of C and int(C) denotes the interior of C. First we require the following lemma. Let H be a topological vector space and W be a convex subset of RxH .We define Wo { woeR: (wo,wH)EW} and W'= {wH: 0w,wH)EW, wo <0}). Let H* denote that space conjugate to H. Lemma 6.2: Let C' be an open convex subset of H, Osint(Wo) and OEC'. Then either there exists QH EH* such that QH f 0, QH H) > 0 > QH(C) for all wHEW' and ceC' or there exists weW such that wo <0 and wHE . Proof: For every SEW let 5o <0O and SHEC' be outside W. Recall that W'= {wH: (wo,wH) E W, wo <0}. Then W' is a nonempty convex subset of H and W'n, C' = $. Since C' is an open convex set, by the separation property(Theorem 3.12), there exists QHEH* and aeR such that QH > such that QH (H) ~ H(c) where wHEW' and ceC'. But OEC'n, W' and hence a=0. Thus we have (6.3) QH H) > 0 >H(C (HE W' and ce C') This is the first alternative of the lemma. The remaining possibility is when there exists weW such that wo <0O and wHEC'. Hence we have the second alternative of the lemma. The above lemma will be used in the proof of the following theorem. Next we state and prove the simplified version of Theorem IX. 1.2 from Warga (1970). Theorem 6.4: Let Q be a convex set, F be a compact set in R and P be any compact finite dimensional metric space. Consider two functions T1:QxF 4 R and T2:QxFxP 4 R. Now assume the following: 1. Let N = ((q,f)E QxF : T2(q,f~p)E [s,M] for every pEP ) where s,M>0. Let there be (q*,f )EN such that Tl(q*,f ) = min Tl(q,f) . 2. The functions T1 and T2 have convex Gateaux derivatives. 3. For each pair (q,f)EQxF the function T2(q,f,p):Ps R bounded and continuous in p. Then there exists a arpm(P) and an cointegrable o : P 4 R such that i. mo (p)=1 for all peP and co(P) > 0 ii. S$(p) DT2((q',f ),p;(ql~f)(q*,f ))m,(dp)t > iii. co(p)T2(q',f'p) = max co(p)a for coa.a. peP ae[s,M] where DTj((q*,f );(q,f) (q'~f )) denotes the directional derivative of T, at (q*,f*). Proof: Define the set ?v~q,f) = ( D)T((q*,f*);(q,f) (q*,f*)), D)T2((q ,f ),p;(q~f) (q ,f )) :(q,f)EQxF ) This set is convex by hypothesis. Let C2 = ( c(p)E C(P,R) : E < c(p) < M ) where E > 0. Note that C2 is a closed convex subset of C(P,R) and let C'= int(C2) T2 ,*> Then we can apply Lemma 6.2 and the first alternative of the lemma yields the following: there exists Q E CdC(P,R) such that Q f 0, and for all (q,f)EQxF (6.5) e(D)T2((q ,f ),p;(q~f) (q ,f )))~ > (6.6) (c)l < (ce C') In the above expressions note that fE C*(P,R) and hence by the Riesz representation theorem (Theorem 3.10) we can specify an integral form of 8. In particular, there exists a positive Radon measure ao on P and an cointegrable o : P 4 R such that mo (p) = 1 for all peP and co(P) > 0 C(c)~ (p)c(p)co(dp) c E C(P,R) Next we investigate the inequality (6.6). In view of the definition of C' this implies (c T2(q,f ))<0 (ceC') > $(p) [c (p)c(p)]m(dp)r > (ce C2) We claim that this equation implies claim (iii) of the theorem. To see this assume that there exists some coEC2 Such that co(p)c'(p) < c(p)co(p) for pe6P c P where co(GP) > 0. Then we can form c' EC2 Such that c'= co when pe6Potevs But, the integral inequality is not satisfied with c'(p) since: S (p)[c (p)c'(p)]m,(dp)= ) Q(p)[c (p)'(p)p)] (dp) < 0. Now note that C(P,R) is separable and hence C2 COntains a dense denumerable subset (cic2....) Moreover the set (c/p)ceC2) ;S depnse in [sM]n for all peP) and hence it followsr that the set (cl(p),c2(p),...) is dense in [s,M] for all peP. So we can write the following: $(p)c*(p) > sup mo(p)ci(p) jeN = sup co(p)a (peP) = max mo(p)a (p EP) (Since c (p)E [s,M]) Thus we have proved that the statement of the theorem is a result of the first alternative of the lemma. We now prove that the other alternative of the lemma implies a contradiction with the assumption in that (q*,f*) is found to be not minimizing. To prove this consider the set ?VV(q,f) and C' as defined above. From the lemma there exists we TV(q,f) such that wl < 0 and w2 E  For each choice of qeQ and feF, define the function hl:[0,1]>R and h2: [0,1] > C(P,R) as follows: (6.7) h1(6) = Tl(q* + 6(qq*), f + 6(ff )) h2(6) = T2(q' + 6(qq*), f*+ 6(ff*),p) where 8E [0,1]. It can be easily verified that the derivatives h: and h2 at (0,0) can be expressed in terms of the directional derivatives of T1 and T2 at (q ,f ). (6.8) hi(0)86 = 'DTi((q ,f );(ql~f) (q ,f ))68 (i=1,2) Now for sufficiently small SO > 0, h1(68) = h1(0) + h (0)68 + o(86) = h1(0) + wi SO + o(68) < hi(0) = Tl(q',f) Also we know that lim SO [h2(86) h2(0) + h2'(0)s6 ] = 0 >lim SO1 [h2(86) h2(0)] = w2 E C Hence for sufficiently small SO we can conclude that [h2(68) h2(0)]e66 C'cC' (Since OEC') h2(S6) T2(q',f ) E int(C2) T2 q ) Hence (q ,f ) is not a minimizing solution as hypothesized in assumption (1) of the theorem. The Scaled System The theorem proved above is the main tool that we will use to derive first order conditions for the solution of Problem 4.4. The functions T1 and T2 will be appropriately defined for our problem and it will be shown that all the assumptions of Theorem 6.4 are satisfied. Hence a solution to Problem 4.4 must satisfy the conclusions of Theorem 6.4. It will be shown that some useful features of the optimal solution of Problem 4.4 may be derived from careful consideration of conclusion (ii) of Theorem 6.4. With these objectives in mind we redefine Problem 4.4 for a scaled system described next. We express system (4. 1) in terms of scaled variables y(s) and w(s) as follows: y(s) x(ps) sE [0,1] v(s) au(ps) se [0,1] where 8 < 6 and 6 = sup inf T(u,D). In terms of these variables Problem 4.4 can be written as: Problem 6.9: Find min (P) subject to the following constraints: 9(s) = P(A'y(s) + B'v(s)) 0 < s < 1 y(0)= xo x0TXO < M yT(S~y(s) <; M for 0 < s < 1 and for all (A',B') E (A+AA)x(B+AB) where (A+AA) a( A+DA:DA E A) and (B+AB) a (B+DB:DB EB  It can be easily checked that Problem 4.4 and 6.9 are identical. Using standard methods, y(s) can be expressed as: y(s) = e"A' [XOi +pA'T BB'v(T)dr] Definition 6.10: Now we identify the relevant sets in Problem 6.9 with those defined in Theorem 6.4 as follows: Q Uo,,13 where Uto,l] {v(t)EU: v(t) = 0 for t > 1 ) F s[0,6+1] P ((A',B',s)E (A+AA)x (B+AB)x[0,1]} And we define the functions T1 and T2 aS follows: Tl(v(s),P) = p T2(V(S),P, (A',B',s)) = yTy In the following Theorem P (A+AA)x (B+AB)x[0, 1]. Hence if moerpm(P) then conditional probabilities co(A',B's) as well as the marginal probability co(s) can be defined in the usual way so that for any (A',B',s)eE, the joint probability co(A',B',s) = co(A',B's) co(s). Let us denote the Lebesgue measure on [0, 1] by C1. Also note that in the following theorem and thereafter the variables z and s are both used to denote the scaled time and lies in [0,1]. They are used interchangeably as appropriate. Necessary Conditions We are now ready to apply Theorem 6.4 to Problem 6.9 and thus derive necessary conditions for the optimal solution (v*,P*). From this theorem we expect to gain some insight about the characteristics of the optimal input v*. Indeed, from the theorem stated below it will become apparent that the optimal input v* may have a bangbang control feature. The result introduces the quantity z(s) over the scaled interval time [0, 1], which is similar to the classical switching function for the bangbang control input v*.(Pontryagin et al. 1962). However the exact characteristics of z(s) will have to be clarified through some further investigations. The solution to Problem 6.9 and hence the solution to Problem 4.4 is summarized in the following theorem : Theorem 6.11: Let (v (s),P*) be the solution to Problem 6.9. Then there exists ao Erpm((A+AA)x (B+AB)x[0,1]) and a C1measurable function z(s):[0,1] 4 Rm such that the following are satisfied: ii. z(s) v(s) 2 z(s) v (s) for pa a. st [0,1] and for any veUto,11 Where the support of ao is given by the set: 0Z = ( (A',B'r,s~ (A+AAvx (B+ABvx[0,1]: yT'y = M ) Proof: We can apply Theorem 6.4 since all the assumptions are clearly met We first calculate the directional derivatives of T1 and T2. It is more convenient to use the variable z as the time variable and s as the running variable in the following expressions. The advantage will be apparent from the subsequent derivation. DT1((GB)(v ;(,) (v ,P )) = (PP ) By Theorem 6.4 there exists me~rpm(P) and an cointegrable integrable o : P 4 R such that mo (p) = 1 for coa.a peP and co(P) > 0 Si5(p) D~T2 If we set p = P* in the above inequality and refer to the expression for the directional derivatives of T1 and T2 We Obtain (6.12) Si(p) DT2 >~~~ 2(p.y A (s) P*B'(v(s)v (s))pL(ds) m`p)> (p~yT(' pA (Ts *BYorjs(o)'s(ds)2 mOp (W~here Ylno, (s)=( = ~ ). >S riyr(p)yT p" A'Ts) BYor(s)o(dp)' (v(s)v (s))Cl(ds)2 > (By Fubini's Theorem (3.11)) > (s)(v(s)v (s))r(ds) > 0 where z(s) o(p)yT p A8*T'("s B'Y~o~,r(s)om(dp) 0 P Now by hypothesis, Ps ((+A (B+AB\ /nx[0,c1]} and ao is a Radon probability measure on P. Hence we can write co(dA',dB',dz) in terms of the appropriately defined conditional and marginal probabilities. The expression for z(s) can be rewritten in terms of the conditionals as follows: z(s) j 2U(A',B',Z)yT Z~P A'(ts) B'm)(dA',dB'/Z) 'lo~,](s) m3(dr) = $(A',B',Z)yT Z~P A'(s) B'm(dA',dB '/z) (dz) We claim that: Sz(s)(v(s)v.(s))p(ds) >o 0 z(s)v(s) >z(sivi(s) for pa~a. se [0,1]. To see this assume that there exists some voEU such that z(s)vo(s) < z(s)v*(s) for se6T C [0,1] where C1(6T) > 0. Then we can form v'EU such that Ivo when se6T vr~V otherwise But, the integral inequality is then not satisfied with v'(s) since: / z(s)(v'(s).(s)s))(ds) = J z(s)(v'(s)v.(s));(ds) < 0 O ST Now the third conclusion of Theorem 6.4 directly determines the form of the support set of the measure mo. From Theorem 6.4): c(p)yT~pv,*y(p )p;v ,P* )= max co(p)a for coa.a. peP. as [s,M] Recall that co(p) = +1. Hence: yT~p;v ,P )y(p;v*,P*) = M when co = 1. If c 1 the equality above cannot be satisfied. Hence ao = 1 for coa.a. peP and the measure ao has the support set defined as follows: 0Z = (peP: yT~p;v*,P )y(p;v*,P*) = M). The expression for z(s) simplifies to: zs =fy j (fA'(ts) B'mrdA',dB'/T) w(dl). + The above theorem derives necessary conditions for the optimal solution (v*,P*) to Problem 6.9. Now consider conclusion (ii) of the above theorem. z(s) v(s) > z(s) v (s) for C1a.a. s [0, 1] and for any veUtool] It follows that when each component of z(s) + 0 the optimal input solution v*(s) is bangbang and alternates between the maximum allowed input bounds (namely +K). In other words when zl(s) + 0 for some je ( 1,2,...m), this implies that for a.a. se [0,1]: I.K when z)(s)>0 YY~)'I K when zl(s)<0 This partially solves our questions about the characteristics of the optimal solution v'. However we cannot conclude anything about the corresponding components of the solution over intervals when some components of the function z(s) are zero. A completely bangbang solution would be extremely favorable from an engineering point of view, for its ease of implementation and numerical computation. However as we show in the following example, this may not always be true in this case. We will show that some components of the switching function z(s) can turn out to be zero over nonzero intervals of time. I I I I I ~_________L________~_________L________~_ ~L________~_________L_______ 0.5 1 1 .5 2 2.5 3 3.5 We would like to know more about the function z(s), in particular, whether some components of z(s) could be zero over contiguous subintervals of [0, 1]. Recall that for the solution to be purely bangbang each component of the switching function z(s) needs to be non zero almost everywhere on [0,1]. We note that our problem is linear time optimal and in most such problems studied in literature the solution turned out to be bangbang. (See Pontryagin et al. 1962). However it is interesting to note that in this case, it does not always hold. In general, it is not true that the optimal solution is bangbang and hence some components of z(s) are zero over subintervals of nonzero measure in [0, 1]. We present the following example: Example 6.13: Find max tr subject to the following constraints: ueU 0 for 0 < t <; tr and for all ae[1.2,1.4] to the problem for M = 1.96. 2 1 1 2 Time one switch: M = 1.96, tf = 3.7 Figure 61: Optimal input has x(t) = ax(t) + u(t) x(0) 1 xT(t~x(t) < M We provide a solution The solution is bangbang only over the first part [0, 1.27] of the interval. Over the rest of the interval until tf the optimal input though constant (=1.67) is not +2. For a clear understanding of the behavior of the system for different values of the uncertainty we have plotted the corresponding traj ectories for ten different uncertainty values of the pole 'a'. The solution to this problem was calculated using brute force techniques, where both the time axis (0 to 4 seconds) and the amplitude of the input [2,2] were discretized thereby forming a grid. Every input was checked on this grid to find the best solution. The continuous time solution was found by interpolation of the discrete optimal solution. O 0.5 1 1 .5 2 2.5 3 3.5 Time Figure 62: Traj ectories for ten different uncertainty values of 'a'; M = 1.96, tf = 3.7 While this feature of the optimal solution v* is interesting because of its anomaly with other known solutions to linear time optimal problems, studied in literature, it is inconvenient from an engineering point of view. It is much easier to compute a bangbang solution than otherwise, since for a bangbang solution, only the switching instants need to be computed to effectively know the solution. Hence for a bangbang solution the dynamic optimization with high complexity reduces to a simple static optimization over the space of switching instants. In the next section we show, that even over intervals where some components of z(s) are zero and consequently the optimal solution v* is possibly not bangbang, there exists a bangbang input v' which approximates v* in an appropriate sense. BangBang Approximation to v* As we discussed in the last section, it would be interesting to know if the optimal solution, which may not be bangbang, could be approximated by a bangbang input in any sense. Recall that the objective of the optimal input is to maintain the inequality xTx < M for every trajectory of the error, corresponding to any uncertainty value, during the maximal time interval tf. We show that the optimal input u* can be approximated by a bangbang input u', in the sense that xT~u',D)x(u',D) < M + s for every value of the uncertainty DEA, for at least tr seconds. Here a can be made arbitrarily small by increasing the number of switches in u'. Lemma 6.14: There exists a bangbang function u'EUto y] with a finite number of switches, such that xT(u';D)x(u';D) <; M + s for every DEA and for at least tr seconds. Here, a can be made arbitrarily small by increasing the number of switches in u'. Proof: Let x*(t,D) and x'(t,D) denote the state vectors, corresponding to uncertainty value D(DA,DB)EA at instance t when the optimal control u* and u' are applied respectively between [0,t]. Since they start from the same initial condition, x*(t,D) x'(t,D) = eAl'teA'zB'[u'u']dt We claim that for small enough t it is possible to switch u' appropriately so that eA't eA'"B'[u'u']dt = 0 V'(A',B')E((A+AA)x(B+AB)) Since eAt is invertible always, it is enough to prove that eA'TB'[u'u']dr =0. Let t be small enough so that eAt can be considered constant over [0,t]. We now propose to switch each component of u'tul',u2' ,....,Um': Once betweenn +~ duringr [0,t]. The i~ntrval before the switch we denote by til and the interval after the switch by ti2 COrresponding to the ith input ui. The switching times have to add up to the total time t, i.e., til + ti2 = t. Hence: BUlltii + Aulztl2 eA'B'[u'u']dr=eA'tB umitmi + Aumztm2 Where Auij = ui ui' (i = 1,2,...m) and j = 1 denoted the time interval before the switch and j=2 implies the time after the switch. Clearly, we can choose the switching times so as to make Auiltil + Auizti2 = 0 for every is {1,2,....m}. This proves that at time instant t we can make x*(t,D) x'(t,D) = 0 irrespective of the value of the uncertainty D(DA,DB), by switching each input once. The total number of switches required for all the inputs is exactly m. This does not guarantee that x (z,D) = x'(z,D) for all ze [0,t] VDEA. However because of the absolute continuity of x(t,D), by choosing t small enough we can guarantee that x*(t,D) x'(t,D) <: E for any given s. Recall that the optimal open loop interval is always finite. Hence the optimal interval [0,tf] can be split into arbitrarily small subintervals to form a partition P{0,tl,t2, ....,t,,tf}. We can use the above method for each subintervals [ti,ti+l], (i=1,...,p) so that x*(ti,D) = x'(ti,D) VD EA for each tieP. Also by making the partition P fine enough, we can make x (z,D) x'(z,D) < s (ze(ti,ti+l)) for any given s. It follows easily that x'(t,D) <; M + s for to[0,tf] and VDEA. Over intervals where some components of z(s) are 0, the above lemma provides an approximation of the optimal solution with a bangbang input. The approximation holds irrespective of the value of the uncertainty and hence guarantees that every state traj ectory remains within the allowed error bound for the maximal period. Moreover, since the approximation is being done by a bangbang function, all the computational advantages corresponding to a bangbang solution are inherited by u'. In general, for an accurate approximation, the number of switches required for u' may be high. However for practical purposes, the required number of switches may be computed by repeatedly calculating the maximal time for increasing number of switches. The iteration should stop when no appreciable improvement occurs with the increase in the number of switches. We consider Example 6.13 again to show that the nonbangbang optimal input can be effectively approximated by a bangbang signal, as is predicted by Lemma 6.14. The approximating bangbang input has a total of sixteen switches. Corresponding to each uncertainty value of the pole a, the state traj ectory with the actual optimal input is approximated by the traj ectory resulting from the bangbang input. To see this the reader need to compare the following figures with Figures 61 and 62. A slight relaxation of the error bound beyond M=1.96 allows for an identical maximal time of 3.7 seconds.       1.5 0.5 O: 0.5 1 1.5 O 0.5 1 1 5 2 2.5 3 3 5 Time bangbang input: Figure 63: Approximate 16 switches 0.5 O 0.5 1 1 .5 2 2.5 3 3.5 Time Figure 64: Trajectories for ten different uncertainty values of 'a'; M = 1.96, tf = 3.69 CHAPTER 7 CONDITIONS FOR A PURELY BANGBANG SOLUTION Though the necessary condition derived in Theorem 6. 11 hints that the optimal solution may be bangbang, we could not rule out the possibility that some components of the switching function z(s) could be zero over subintervals of [0, 1]. In fact we discussed a case where this is the case and the optimal solution is clearly not bangbang. Furthermore, we proposed a method in which a nonbangbang optimal input may be approximated by a high frequency bangbang mnput. However it would be interesting to know whether there are conditions under which the optimal solution itself is purely bangbang. For this we need to investigate two important questions about the switching function z(s). Firstly, we need to find conditions under which each component of the switching function z(s) is nonzero almost everywhere. Moreover we need to estimate number of zero crossings of the components of the function z(s). The following assumptions were found sufficient to guarantee a purely bangbang solution to Problem 4.4. Assumptions 7.1: * The uncertainty bound d is small. * The input bound K is sufficiently small with respect to the error bound M. * The system is controllable from each input for some value of the uncertain pair (A',B'). We first prove that all the components of the function z(s) cannot be identically zero over the entire interval [0, 1] for small uncertainties on the A and B matrices. Lemma 7.2: If the maximum disturbance d is small, the function zi(s):[0,1]4R cannot be zero C1a.e. on [0,1] for all je {1,...,m}. Proof: According to Theorem 6. 11, for P = P , S DT2 (V *), V V))Q(dp) = J z(s)(v(s)v (s))Cl(ds). Assume that z(s) = 0 a.e. on [0,1]. Then S iDT2(V *), V V*))cm(dp) = 0 for any ymeasurable v(s). Now, for p = P* the expression for DT2 (V ;V V )) may be simplified as follows : DT2 (V ),(V V )) = yT(r,A',B';P ,v.~ ) felA'(s) *B'(v(s)v (s))dsl = yT(T,A',B';P ,v ) IfeB.A'(s) B*B'v(s)ds je8A'(Ts) P*B'v (s)ds = yT(T,A',B';P ,v ) { y(Z,A',B';P ,v) y(Z,A',B';P ,v )} Now the support of ao was determined to be the set 0Z = {(z,A',B')EP: yT(T,A',B';v ,P )y(z,A',B';v ,P ) = M}. Hence the integral expression simplifies to: S DT2 (V *), V V))Q (dp) = SyT(T,A',B';P ,v*)y(t,A',B';P ,v)uadzdrd'B) M Sfa(dt,dA',dB') = SyT(T,A',B';P ,v*)y(z,A',B';P ,v)u,(dt,dA',dB') M According to our assumption: (7.3) SyT(t,A',B';B* ,v )y(t,A',B';B* ,v)co(dz,dA',dB') = M for every CLmeasurable v(t). Hence for v(t) = 0 for to [0,1] we have: (7.4) SyT(T,A',B';B ,v )eAt' XO m(d~r,dA',dB') = M. Next we use some Dirac delta and Dirac delta derivatives for v(t) to try to bring every traj ectory to near zero. While these functions are not strictly C1measurable, they can be thought of as the limits of very high amplitude measurable functions and the proof is considerably simplified as a result. Moreover, note that under the assumption of xoxo < M, and by the definition of 02, z(s) can be expressed as the following power series in s in a small enough neighborhood about s = z(s) = SyT(T,A',B')eB A' B'm3(dA',dB',dz) + s SyT(T,A',B')eBp" A*A'B'm(dA',dB',dz) + + 4 yT(TA',B')eP A' A'2B'm(dA',dB',dr)+ + .... It follows that z(s) is analytic in a small enough neighborhood about s = 0 and hence the assumption that z(s) = 0 C1a.e. implies that z(0) = 0. Choose v(t) = [6(0) 6'(0) .... Sn(0)] C xo Since z(0) = 0, (7.3) is valid for such a v(t). Then: y(0+) = xo [ (B+DB) (A+DA)(B+DB) ..... (A+DA n(B+DB) C1 XO = [DB DB +DAB ......]Cxo Hence with this input (7.3) yields: (7.5) y (0 ",v )eA"t [DB AB +DAB .........]C 'xo a = M Assuming that DA and DB are Small the left hand side of the above equation can be approximated by: y T(0 ,v)eA/t' o [D'B BD' +D'AB .....]C x where DA' and DB' are average values with respect to the measure mo. If we divide the above expression by the maximum allowed scalar disturbance, d (recall IDAl d and DB [D'B DB +D'AB ......]C'xo = d [hi Ah2 + h3B.....] Cl'xo where each element of the matrices hi lies between [1,1]. Hence for any measure ao and independent of the maximum disturbance d, the following upper bound holds: [ hi Ah2 + h3B .....] C'xo < ki xo where kl is a large enough scalar. > y ((3 ,v")eA[/to~ [D'B B '+D'AB .........]C' x <: dkl SyT *l,V )eA/t XO m < dkl M (by Equation 7.4) Now as da 0 Equation 7.5 cannot be satisfied. Hence we have a contradiction. Lemma 7.2 shows that at least one component zl(s) of the function z(s) is nonzero over some part of [0,1]. The optimal solution is bangbang over those subintervals. In the following lemma we prove that if the input bound K is small enough compared to the allowed error bound M (by Assumption 7.1), then the optimal trajectories hit the boundary only at the end of the optimal interval [0,tf]. We will show that this condition is not only enough to make optimal solution always bangbang, but also provides more information about the zeros of the function z(s). .2 denotes the Euclidean norm below. Lemma 7.6: If for some constants k2, d2 and dr defined below, (M)1/2 (A2 di) > k2 (B2 +d2) K holds, then XT(t)x(t)< Mfor all DEA te[0,6) and xT(6)x(6)= M for some DEA. Proof: Recall that J(t) = xT(t)x(t). When some input u(t) is applied to the system, the open loop operation must stop when for some disturbance D=(AA, B), the inequality J(t) > M is satisfied. Now J(t) is a continuous and differentiable function of t on any interval in [0,00). Let J(to) = M for some to corresponding to some x(t) and D .If for every permissible u(to+), J(t)to+ > 0, then for every E > 0, J(to + 8) > M and hence 6 = to. We derive conditions so that J (t)to > 0 for any permissible input and any trajectory satisfying J(to) = M. J (to>,+ 0 x,.c > 0 a xT(A'x + B'u) > 0. We assumed A' to be unstable and hence, by the Lyapunov theorem, XTA'x > 0. If xTB'u > 0, then the inequality above is trivially satisfied. Hence we assume that xTB'u < 0 and the last inequality becomes: xT(A'x B'u) > 0 e  xTA'x 2 > IIXTB'u 2  e  xTA'x 2 > IX 2.IB '2. IU 2 Now there exists 0 < kl < 1 such that  xTA'x 2 = klx1.A'2. Let k2 be the lower bound of the values of kl for any value of x(to) such that xT(to)x(to) = M and A'. Since by assumption xTA'x2 > 0, we claim that k2 > 0. This can be seen by noting that xTA'x2 < M (A2 + d2). Hence k2 = 18 M(A+d Since the right hand side is continuous in x and A', the infimum is achieved and hence k2 > 0. Hence the following implies the last inequality: k2x1 A'2 > IX 2 IB'2 IU 2z e > k3 K (where k3 is a constant resulting from the equivalence of norms: B'2 recall that u2 = kzk3 u < kzk3 K) (M)1/2(A2 di) I B2+ d2 3 The last implication follows from: A'2 =  A + DA 12 > A2 DA 2z > A2 di B'2 =  B + DB 12 < B2 + DBll 2 < B2 + d2. where dl,d2 are pOsitive constants resulting from the equivalence of norms and the bound d on the uncertainties: DA, DB < d. The significance of the above lemma is that it provides us with a simple condition when the optimal trajectories will hit the boundary only at one point in time and that time is necessarily at the end of the optimal time interval [0,tf]. It is however evident that more than one traj ectory corresponding to different values of the uncertainty (DA,DB), can hit the boundary at the same time. Lemma 7.7: Under Assumption 7. 1, each zi(s) + 0 almost everywhere on (0, 1) and the set of zeros of z(s) do not have a limit point in (0, 1). Proof: Consider the expression for z(s): z(s) = SfyT(r,A',B')ep A'(7s) B'm(dA', dB'/Z) CO(dZ). Recall that the support for aO was defined as 0Z = ((A',B',z)E(A+AA)x (B+AB)x[0,1]: yTy = M). If Lemma 7.6 holds then 0Z = ( (A',B',z)E(A+AA)x (B+AB)X I f Ty = M, ) and z(s) can be simplified as follows: z(s) = SyT(1 ,A',B')eB A'(1s) B'm(dA',dB'). Hence for each se [0,1) the expression for z(s) is given by the following power series expansion: z(s) = SyT(1,A',B')eB'A' B'm,(dA',dB') + s fyT(1,A',B')eB'A' A'B'm3(dA',dB') + 2! f.yT(1,A',B')eB'A' A'2B'm(dA',dB')+ +......... SyT(1,A',B')eP'A1' d'd' B' + s yT(1,A',B')eP'A1' mdA',dB') A'B' + fyT(!1",A',')SY'i m. lr (ladA',dB' A'2B' The above equality is true since the diameter of AA and A, are small compared to the variation in eP'A' as A' varies over As. Let us denote the coefficients by (Xo, X1, X2,  ) Hence: Xk FT~r(1~A,A',B)e'A co(dA',dB') A'kB' = [ f f2.. n]A'kB' (Where we denote the row matrix created by the integral as [ fi f2  ] and n is the order of the system. We stack that coefficients together and expand each Xk. (7.8) [Xo X1 ....Xk ....] 1 [f2  n][B' A'B' ..... A'kB'.....] or [Xol ....Xom  X11 ... ... Xkl....] 1 [f2  n[B' A'B' ..... A'mB'. ....] Now recall Lemma 7.2, which guaranteed that at least one component of zi(s) (j=1,2,...,m) is nonzero over at least a subinterval of nonzero measure in [0, 1]. Without loss of generality that this is the case with the z (s). Let us write z (s) in terms of the coefficients defined above. z (s) = Xoi + Xiis + X21S2 . Hence z (s) is analytic and in particular a power series in s over [0,1]. It is well known that a power series is either identically zero or is nonzero almost everywhere in its domain of definition [0,1] (Conway 1978). Since by Lemma 7.2, z (s) is nonzero at least over some interval of nonzero measure in [0,1], it follows that it is nonzero almost everywhere over [0,1]. We now write the coefficients of z (s) in terms of Equation 7.8. [Xol X11 ...... Xnt ......] = [fi f2 ~[C n 1 .. Where C1 is the controllability matrix for the first input. Since z (s) is nonzero, at least one of the coefficients Xil (i=0,1,....) is nonzero. Hence [fi ...fn] + 0, i.e., at least one among fi (i=1,...,n) is nonzero. Recall that by assumption (7. 1), the controllability matrices Ci corresponding to each of the m inputs are full rank. Hence at least one among the first n coefficients of each z(s) (j=1,2...,m) is nonzero. This in turn implies that each z(s) is non zero almost everywhere on (0,1). This proves that z(s) is a nonzero power series on (0,1). It is well known that zeros of analytic functions cannot include a limit point in its domain of definition. It follows that the only possible limit points in the zeros of z(s) is at s=0,1. Hence proved. From Theorem 7.2 and the above lemma, we conclude that z(s) satisfies the following: z(s~v(s) 2 z(s)v (s) for a.a. s [0, 1]. Moreover each component of z(s)+0 a.e. on (0,1). Together this implies that for a.a. se [0, 1]: I.K when z(s)>0 yi,=+K when z(s)<0 Hence, the optimal solution is purely bangbang almost everywhere on (0, 1). The value of v (s) is not known over sets of measure zero. These sets are exactly the values of s where some component of the function z(s) = 0. Recall that each zero crossing of any component of the switching function zl(s) (j=1,2...m) can potentially imply a switch between the extreme allowed values of the input (+K). The number of such points is not Einite in general and hence the solution may contain an infinite number of "switches" or sign changes between the extreme values. However, by Lemma 7.7, the zeroes of z(s), or in other words the switching instants of the optimal input v', do not have a limit point in (0, 1). The only possible limit points are at s = 0 or 1. This implies that the optimal switching sequence may have become increasingly close together at the very beginning or the end of the optimal open loop interval. However, the open loop interval is always Einite. From an engineering viewpoint, the optimal input function may be truncated near the two ends. This will result in a Einite number of switches in the optimal input solution with an arbitrarily small reduction in the maximal open loop time. This will make the computation of the switching times as well as the hardware implementation feasible. Hence the optimal v* may be approximated for every practical purpose, by an implementable function with a finite number of switches. The engineering importance of the numerical simplification in the calculation of the optimal solution cannot be over emphasized. The optimal solution, being bangbang, is effectively known by the computation of the switching instants. Moreover, a bangbang solution may be implemented with the most basic in hardware and simplest of algorithms. By a simple example, we demonstrate some of the interesting features of the problem solved above. We choose the same single dimensional system with a single input and with some uncertainty on the pole as was used in Problem 6. 13. The state is the output of the system. Let us transcribe Problem 4.4 in terms of this system: Problem 7.9: Find max tr subject to the following constraints: ueU x(t) = ax(t) + u(t) 0 < t < tr and u(t) < 2 x(0)= 1 xT(t~x(t) < M for 0 < t <; tr and for all as [1.2,1.4] We provide a solution to the problem for M = 25. The optimal input in this case has only one switch with the optimal switching instant at 1.3467 seconds. The maximum tf was found to be 5.08 seconds, and in the time interval [0, 5.08 sec], every state trajectory is guaranteed to be within [5,5]. For a clear understanding of the behavior of the system for different values of the uncertainty we have plotted the corresponding state traj ectories for ten different uncertainty values of the pole 'a'. l i l i l l i l l l l l~~L~ _l_ I_ l_ __ __ l l l l l l__~_ O 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 Time Figure 71: Optimal input has one switch: M = 25, t, = 5.08 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 Time Figure 72: Traj ectories for ten different uncertainty values of 'a'; M = 25, tf = 5.08 CHAPTER 8 CONCLUSION AND FUTURE WORK Conclusion The problem of maximizing the open loop period of operation of a linear time invariant system with bounded inputs has been considered. The parameters of the controlled system are subj ect to bounded uncertainties. An optimal controller has been derived that maximizes the time during which the control loop can be left open, while keeping the system error within pre specified bounds. The existence of such a control signal is first proved among the set of measurable functions. EulerLagrange type first order necessary conditions are then derived for calculating the optimal open loop input. It is shown that the time interval, during which the control loop can remain open, is maximized by an input, which may not be purely bangbang over the entire maximal open loop time. We further showed that in cases where the optimal input was not bangbang over certain intervals, a purely bangbang input existed which approximated the optimal input. This is of engineering importance, since the bangbang nature of the optimal solution makes the computation of switching instants computationally feasible, as opposed to solving the entire dynamic optimization problem numerically. In the general case, the possibility of a high number of switches in the approximate input solution cannot be excluded. However, under the assumption that the input bound is small compared with the allowed error bound on the system, we have shown that the optimal input itself is purely bangbang. Moreover, the sequence of optimal switching instants does not have a limit point in the interior of the maximal open loop interval. This facilitates the computation of the optimal open loop input. Because of the finiteness of the open loop interval of operation, the optimal input can then be approximated by a piecewise constant input with a Einite number of switches. Future Work The results we have obtained have potential impacts on a number of application areas, including the following. Each of these applications is a potential candidate for future research that is based on the theory developed in this dissertation. Applications Control of Space Vehicles: Space vehicles are frequently faced with intermittent loss of signal due to obstacles in the line of vision, radiation interference, limitations in power, etc. In such situations, it is critical to know how long the system can perform within its specifications without communication with its supervisory center. This question can be directly addressed within the framework of needbased feedback we have developed. Telemetry: Applications in telemetry frequently face the problem of limited power. The method we propose effectively minimizes the communication needed to control a system to a minimum and hence maximizes the longevity of the associated power sources. Stealth Systems: Stealth and spy systems, like unmanned aerial vehicles and similar devices, prefer to reduce communications with their control center so as to reduce the chances of detection. The needbased feedback framework can derive the necessary control feeds to achieve this goal and reduce the probability of detection. Agricultural Systems: Modern agricultural applications, such as soil moisture content control and fertilization control, require complex and expensive feedback processes, as most feedback data has to be collected by human experts testing in the Hield. Using needbased feedback reduces the frequency and the duration of these Hield tests, thus reducing costs and improving efficiency. Biotechnology: Biological research often involves culture and preservation of cells and other organic substances under very carefully controlled environments. Regular human surveillance is necessary and extremely expensive. Our method of reducing feedback can effectively minimize requirements of human surveillance and thus reduce involved costs. Networked Control Systems: It is increasingly common to use shared networks to control geographically separated devices. The inherent unreliability of networks may cause frequent loss of feedback signals for uncertain periods. To guarantee control obj ectives, it is crucial to have an estimate of how long the system can perform in open loop. Again, this can be answered in the framework of needba~sed feedback control. Medicine: Applications are also possible in medicine and optimization of drug dosage. Typically patients are treated with drugs at regular intervals while the feedback in terms of its effectiveness is collected after large intervals. The method proposed can be potentially used to guarantee the effectiveness of the drug when measurements are not being made. Moreover the intervals after which the patient needs to be checked may be maximized. Numerical Optimization: The actual computation of the optimal solution presents an interesting problem in numerical optimization. We conj ecture that appropriate use of combinatorics may lead to a highly efficient algorithm for calculating the solution. Theoretical Research This problem is closely related to the viscosity solutions of the HamiltonJacobiBellman equation. It seems that the socalled "exit time problem" studied widely in relation to minmax dynamic programming is similar to the maximal open loop time studied above. But the inherent nonsmoothness of the current solution makes the use of nonstandard solutions necessary and hence may present an interesting future direction of research. The theory proposed extends the results on optimal residence time of perturbed control systems. While previous results have been derived in a feedback control setting, we have shown that investigation in the framework of robust openloop optimization framework is interesting and relevant to physical situations. A complete theory requires further research in this direction. Questions about reachability of uncertain dynamical systems are closely intertwined with the problem investigated above. The concept of reachability of uncertain systems clearly depends on whether the system is in open or closed loop. However the correct characterization still needs to be investigated. The problem of stabilizing an unstable system with prespecified bounds on the inputs is related. Though this is extremely common in most controller design, a satisfactory theory is not yet available. It is thus of practical interest to calculate the minimum input bound required to stabilize an unstable uncertain system. Further theoretical applications of the proposed theory may be found in the areas of stochastic dynamical systems, Lyapunov stability analysis of switched dynamical systems, and convex and nonconvex minmax dynamical optimization. LIST OF REFERENCES A.V. Balakrishnan, Introduction to Optimization Theory in Hilbert Space, SpringerVerlag, Heidelberg, 1971. T. Basar and G.J. Olsder, Dynamic Noncooperative Game Theory, London/New York: Academic Press, 1982. P.R. Bellanger, "On the sensitivity of solutions to some optimal control problems," Ph.D. dissertation, Massachusetts Institute of Technology, Cambridge, 1964. R. Bellman, Dynamic Programming, Princeton University Press, Princeton, 1957. L.D. Berkovitz, "A differential game with no pure strategy solution", Advances in Game Theory, pp. 175194, 1964, Princeton Univ. Press, Princeton, N.J. L.D. Berkovitz, "Necessary conditions for optimal strategies in a class of differential games and control problems", SIAM~J. Control, 5, 1967, pp. 124. L.D. Berkovitz, "A variational approach to differential games", Advances in Game Theory, Annals of2ath. Study 52, Princeton University Press, Princeton, 1964, pp. 127174. A.E. Bryson, Jr. and Y.C. Ho, Applied Optimal Control, Blaisdell, 1969. R.M. Cannon, Jr. and E. Schmitz, "Initial experiments on the endpoint control of a flexible one link robot", Int. J. Robotics Res., Vol. 3, pp. 6275, Fall 1984. J.B. Conway, Functions of One Complex Variable, Springer, 1978. R.J. Elliot and N. Kalton, "The existence of value in differential games ", Mem. Amer. Math. Soc. 126, 1972. R.J. Elliot, N. Kalton and L. Markus, "Saddle points for linear differential games ", SIAMIJ Control Optimization, Vol. 11, pp. 100112, 1973. B.P. Eng, "Microprocessor based servo control for gun fire control", in Proc. IEEE IECIAnnual Conf Philadelphia, PA, Mar. 1979, pp. 329333. A.A. Feldbaum, "Dual control theory, parts I and II", Automation and Remote Control, Vol. 21, No. 9, April 1961, pp. 874880 and Vol. 21, No. 11, May 1961, pp. 10331039. (Russian originals dated September 1960, 12401249 and November 1960, pp. 14531464) A.A. Feldbaum, Optimal Control Systems, New York: Academic Press, 1965. A.F. Filippov, "Differential equations with discontinuous righthand side", Amer. Math. Soc. Transl., Ser 2, 42, pp. 199231, Providence, R.I., 1964. W.H. Fleming, "The convergence problem for differential games", Journal of Mathematical Analysis and Applications, 3, pp. 102116, 1961. W.H. Fleming, "The convergence problem for differential games", Advances in Game Theory, pp. 195210, 1964, Princeton Univ. Press, Princeton, N.J. A. Friedman, Differential Gamnes, John Wiley, 1971. R.V. Gamkrelidze, "On some extremal problems in the theory of differential equations with applications to the theory of optimal control", J. SIAM, Ser. A Control 3, 1965, pp. 106 128. P.R. Halmos, A Hilbert Space Problem Book, SpringerVerlag, New York, 1982. N. Halyo, "Terminal area automatic navigation, guidance, and control research using the microwave landing system. Part 5Design and development of digital automated landing system for steep final approach using modern control techniques," NASA Contract Rep. 3681, Apr. 1983. M.W. Hirsch and S. Smale, Differential Equations, Dynamical Systems, and Linear Algebra, Academic Press, 1974. Y.C. Ho, A.E. Bryson, Jr., and S. Baron, "Differential games and optimal pursuitevasion strategies," IEEE Trans. Automatic Control, vol. AC10, pp. 385389, October 1965. D.R. Howard and Z.V. Rekasius, "Error analysis with the maximum principle," IEEE Trans. Automatic Control, vol. AC9, pp. 223229, July 1964. R. Isaacs, Differential Gamnes, New York: Wiley, 1965. R. Isaacs, "Differential games III", Research Memorandum RM1411, The RAND Corporation, Santa Monica, California, 1954. D.L. Kelendzheridze, "On the theory of optimal pursuit", Soviet Math. Dokl., 2, 1961, pp. 654 656. A.J. Koivuniemi, "Parameter optimization in systems subj ect to worst (bounded) disturbance", IEEE Trans. Automatic Control, vol. AC11, pp. 427433, July 1966. A.N. Kolmogorov and S.V. Fomin, Elements of the Theory of Functions and Functional Analysis, Vol 1 and 2, Graylock Press, Rochester, New York, 1957 and 1961. N.N. Krasovskii and A.I. Subbotin, Game Theoretic ControlProblems, SpringerVerlag, Berlin, 1974. Y.S. Ledyaev and E.F. Mishchenko, "Extremal problems in the theory of differential games", Proceedings Steklov Inst. Math., 185, pp. 165191, 1990. Y.S. Ledyaev and E.F. Mishchenko, "Optimization conditions for strategies in differential games of fixed duration", DifferentialEquations, 23, 1988, pp. 175184 Y.S. Ledyaev, "Optimality conditions for feedback controls in problems of control under disturbance", Proceedings of the 33rd Conference on Decision and Control, Lake Buena Vista, FL, December 1994. Y.S. Ledyaev, "The maximum principle for optimal strategies in differential games", Soviet Math. Dokl., 40, pp. 545549, 1990. Y. S. Ledyaev, "On optimal strategies in differential games of fixed duration", Soviet Math. Dokl., 33, pp. 7880, 1986. L.A. Liusternik and V.J. Sobolev, Elements ofFunctionalAnalysis, Frederick Ungar, New York, 1961. D.G. Luenberger, Optimization by Vector Space M~ethods, Wiley, New York, 1969. S.M. Meerkov and T. Runolfsson, "Residence Time Control", IEEE Transactions On Automatic Control, Vol. 33, No. 4, April 1988. E.F. Mishchenko and L. S. Pontryagin, "Linear differential games," Soviet Math. Dokl., 174,. No. 1, 1967, pp. 654656. E.F. Mishchenko, "Pursuit and evasion problems in differential games", Eng. Cybernetics, 9 (1971), no. 5, 787791 L.A. Montestruque and P.J. Antsaklis, "Stability of modelbased networked control systems with timevarying transmission times," IEEE Transactions on Automatic Control, Special Issue on Networked Control Systems, Vol. 49, No. 9, pp. 15621572, September 2004. G.N. Nair, F. Fagnani, S. Zampieri and R. J. Evans, "Feedback Control under Data Rate Constraints: an Overview", to appear in Proceedings of the IEEE (special issue on Jhe Emerging Technology of Networked Control Systems"), 2007. L.W. Neustadt, "An abstract variational theory with applications to a broad class of optimization problems I, general theory", SIAM. J. Control, 4, 1966, pp. 505527. L.W. Neustadt, "An abstract variational theory with applications to a broad class of optimization problems II, applications", SIAM. 1. Control, 5, 1967, pp. 90137. J.C. Panetta and K.R. Fister, "Optimal control applied to competing chemotherapeutic cellkill strategies", SIAM Journal on Applied Ma'~thematics, Vol. 63, No. 6, 2003. L.S. Pontryagin, V.G. Boltyansky, R.V. Gamkrelidze and E.F. Mishchenko, The Mathematical Theory of Optimal Processes, Interscience Publishers John Wiley & Sons Inc., New York London, 1962 L.S. Pontryagin, "On the theory of differential games", Russ. Math. Sury., 1966, 21 (4),pp. 193 246. L. Pontryagin, "Linear differential games. I, II", Soviet Math. Dokl., 175, 1967 pp. 764766 W. Rudin, Real and Complex Analysis, McGrawHill, New York, 1966. R.E. Skelton, "Precise pointing control of space vehicles with momentum exchange controllers", Sperry RanR ep.RRRR~~~~~~~RRRRRR SP250754, 1973. R.B. Vinter, "Minimax optimal control", SIAM Journal of Control and Optimization, Vol. 44, No 3, pp 939968, 2005. R.B. Vinter, Optimal Control, Birkhauser, Boston, 2000. G. Walsh, H. Ye, and L. Bushnell, "Stability analysis of networked control systems", Proceedings ofAmerican Control Conference, June 1999. J. Warga, "Minimax problems and unilateral curves in the calculus of variations", J. SIAM, Ser. A, Control 3, pp. 91105, 1965. J. Warga, Optimal Control of Differential and Functional Equations, Academic Press, New York, 1972. J. Warga, "On a class of minimax control problems in the calculus of variations", M~ichigan2 Math. Journal, 12, 1965, pp. 289311. J. Warga, "Unilateral and minimax control problems defined by integral equations", SIAMIJ Control, 8, 1970, pp. 372382. J. Warga, "On a class of pursuit and evasion problems", J. Differential Equations, 9, 1971, pp. 155167. J. Warga, "Conflicting and minimax controls", J. Math. Anal. Appl., 33, 1971, pp. 655673. S. Willard, General Topology, AddisonWesley, Reading, MA, 1970. H.S. Witsenhausen, "A minimax control linear problem for sampled systems", IEEE Transactions On Automatic Control, Vol. AC13, No. 1, February 1968. L.C. Young, "Generalized curves and the existence of an attained absolute minimum in the calculus of variations", C. R. Sci. Lettres Varsovie, C III 30 (1937), 212234. L.C. Y oung, Lectures on the Calculus of Variations and Optimal Control Theory, W. B. Saunders, Philadelphia, 1969. E. Zeidler, Nonlinear Functional Analysis and its Applications III SpringerVerlag, New York, 1985. P.V. Zhivogyladov and R.H. Middleton, "Networked control design for linear systems", Automatica, 39, pp. 743750, 2003. BIOGRAPHICAL SKETCH Debraj Chakraborty was born in Calcutta, India. He earned his Bachelor of Engineering degree from Jadavpur University, Calcutta, in 2001. After completing his master's degree in electrical engineering from the Department of Electrical Engineering, Indian Institute of Technology, Kanpur, in 2003, he j oined the Department of Electrical and Computer Engineering of the University of Florida, Gainesville, USA. He completed his Ph.D. in electrical and computer engineering in 2007, under the guidance of Dr. Jacob Hammer at the University of Florida. From 2003 to 2007, he was appointed an Alumni Fellow at the University of Florida. PAGE 1 NEEDBASED FEEDBACK: AN OPTIMIZATION APPROACH By DEBRAJ CHAKRABORTY A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2007 1 PAGE 2 2007 Debraj Chakraborty 2 PAGE 3 To my family. 3 PAGE 4 ACKNOWLEDGMENTS I would like to thank my advisor, Dr. Jacob Hammer for his kind and patient guidance during my stay at University of Florida. Our long discussions, whether technical or not were intellectually stimulating and extremely enjoyable. I thank all the faculty members of my department who helped me learn the basics of Electrical Engineering. Especially I would like to thank Dr. Pramod Khargonekar, Dr. Haniph Latchman and Dr. John Schueller for serving on my PhD committee and for improving my thesis through some insightful suggestions and corrections. Lastly I would like to thank my friends Yoav Levinbook, Saravanan Vijayakumaran, Niranjan Venkatraman and Jun Peng for comments, suggestions and especially for listening to my halfbaked ideas and theories at various stages of their development. 4 PAGE 5 TABLE OF CONTENTS page ACKNOWLEDGMENTS...............................................................................................................4 LIST OF FIGURES.........................................................................................................................7 ABSTRACT.....................................................................................................................................8 CHAPTER 1 INTRODUCTION..................................................................................................................10 2 LITERATURE REVIEW.......................................................................................................15 Related Work in MinMax Optimal Control Problems..........................................................15 Other Approaches for Reducing Feedback Requirements......................................................17 Residence Time Control.........................................................................................................17 3 MATHEMATICAL PRELIMINARIES.................................................................................19 Normed Linear Spaces............................................................................................................19 Weak Convergence..........................................................................................................19 Alaoglus theorem....................................................................................................20 Weak upper semicontinuity......................................................................................20 Definition of Borel, Lebesgue and Radon Measures.......................................................21 Riesz Representation Theorem........................................................................................22 Fubinis Theorem............................................................................................................22 Separation of Convex Sets...............................................................................................23 Directional (Gateaux) Derivative....................................................................................23 Existence and Uniqueness of Solution to a LTI System.........................................................24 Jordan Canonical Representation...........................................................................................24 4 NOTATION AND PROBLEM FORMULATION................................................................26 Notation..................................................................................................................................26 Statement of the Problem........................................................................................................28 5 EXISTENCE OF A SOLUTION............................................................................................30 Weak Compactness of U.......................................................................................................30 Alternative Problem................................................................................................................31 Weak Upper Semicontinuity of T(u,D).................................................................................37 Existence of u*.......................................................................................................................39 6 GENERALIZED FIRST ORDER CONDITIONS.................................................................41 Simplified Theorem from Warga (1970)................................................................................41 5 PAGE 6 The Scaled System..................................................................................................................45 Necessary Conditions.............................................................................................................47 BangBang Approximation to v*...........................................................................................53 7 CONDITIONS FOR A PURELY BANGBANG SOLUTION.............................................57 8 CONCLUSION AND FUTURE WORK...............................................................................67 Conclusion..............................................................................................................................67 Future Work............................................................................................................................68 Applications.....................................................................................................................68 Theoretical Research........................................................................................................69 LIST OF REFERENCES...............................................................................................................71 BIOGRAPHICAL SKETCH.........................................................................................................75 6 PAGE 7 LIST OF FIGURES Figure page 11 Schematic of intermittent feedback control.........................................................................12 12 Loss of feedback due to communication channel disruption..............................................13 61 Optimal input has one switch: M = 1.96, t f = 3.7.................................................................51 62 Trajectories for ten different uncertainty values of a; M = 1.96, t f = 3.7.........................52 63 Approximate bangbang input: 16 switches........................................................................56 64 Trajectories for ten different uncertainty values of a; M = 1.96, t f = 3.69.......................56 71 Optimal input has one switch: M = 25, tf = 5.08.................................................................66 72 Trajectories for ten different uncertainty values of a; M = 25, t f = 5.08..........................66 7 PAGE 8 Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy NEED BASED FEEDBACK: AN OPTIMIZATION APPROACH By Debraj Chakraborty August 2007 Chair: Jacob Hammer Major: Electrical and Computer Engineering Feedback is often used to overcome the adverse effects of perturbations and uncertainties on the performance of engineering systems. However, failures of the feedback channel cannot be completely avoided. This dissertation addresses the questions of how and for how long can desirable performance of a perturbed system be maintained after a failure of the feedback channel. Let be a system that is subject to a perturbation in its parameters. The exact value of the perturbation is not known; it is only known that is bounded by a given constant Now, let u(t) be an input function of and let u be the response of the perturbed system to the signal u(t). The nominal system is 0 and the nominal response to the signal u is 0 u. Therefore, the deviation in the response caused by the perturbation is  u 0 u. To reduce the perturbation, add a "correction signal" v(t) to the input signal, so that the perturbed response becomes (u+v). Then, the new deviation between the perturbed and nominal cases becomes  (u+v) 0 u. The correction signal v(t) must be independent of perturbation value as the latter is not known. Let M be the maximal deviation allowed for the response, and let tf be the time for which  (u+v) 0 u M. Then, the objective is to find a correction signal v(t) that 8 PAGE 9 maximizes tf, given only that the perturbation is bounded by EulerLagrange type firstorder conditions for calculating the optimal correction signal v(t) is presented. It is shown that, under rather broad conditions, the optimal correction signal v(t) is either a bangbang signal or can be arbitrarily closely approximated by a bangbang signal. 9 PAGE 10 CHAPTER 1 INTRODUCTION In this work we reduce the need to communicate between the controller and the sensor measuring the system output by maximizing the time during which the feedback loop can remain open. This is motivated from the fact that in certain applications, feedback is not continually available or cannot be continually measured. Sometimes it is advantageous to temporarily stop the transmission of the feedback signal from the sensors at the system output to the controller. In other situations an unpredictable failure of the feedback channel can occur. For example, in controlling space vehicles, obstacles may accidentally disrupt the line of vision between the spacecraft and earth for varying time periods. In telemetry, the need to conserve battery life may motivate a planned reduction of the time of transmission of feedback signal. Moreover stealth applications, that hide systems from detection, prefer to minimize transmission between the controlled object and the remote controller. In agriculture, measurements about soil parameters like moisture, etc., must be done manually and are consequently expensive. Usually such measurements are carried out intermittently after long intervals. In medicine, mathematical modeling and control of certain diseases have become quite common (see Panetta 2003 and references therein). However the feedback available is inherently of an intermittent nature, since measurements on the patients can only be made after long intervals. Lastly network based control systems use a common unreliable network on timeshared basis with other users and hence the feedback signal may be available to the controller only intermittently. In such applications, it is relevant to ask: Research Question 1: How long can the feedback loop be kept open, while maintaining desirable performance specifications? Research Question 2: What is the best way to control the system when we do not have the feedback signal? 10 PAGE 11 Feedback is necessary due to the uncertainty inherent in any system. The main reason for this need for feedback is the lack of perfect knowledge about the system. This includes modeling inaccuracies, unmodeled nonlinearities, parametric uncertainties, spurious disturbances, input noise and measurement noise. Robustness and immunity to all this features are usually achieved through feedback strategies. Hence, it is evident that no uncertain system can be controlled indefinitely if the feedback signal is not available. However, assuming some amount of tolerance for error, there will be an interval of time in which the system will perform acceptably even without any feedback. Let us assume that the system is allowed to operate within some specified level of error and the loss of feedback occurs when the error is within this specified limit. Then the system trajectory will usually require at least a finite amount of time to exceed the specified error level. This is true for most real systems and in particular, for linear time invariant systems. As we illustrated in the above applications, it is interesting to know, how and for how long this period of acceptable operation without feedback may be extended. The question we ask is: what is the maximal time until which the feedback loop can be kept open so that the error remains within the specified limit. We will show that under suitable assumptions, this period is always finite. Hence the disruption in the feedback signal, whether intended or accidental, will have to end or must be ended after this period of time. Otherwise, effective control is not possible, and the system trajectory may exceed the tolerated error. If the disruption was accidental, this maximal time describes the upper bound of guaranteed safe operation. If the disruption carries on beyond this point the system can potentially fail. The a priori knowledge of this critical maximal time interval can help in decisionmaking. If the suspension of feedback was intentional, at that point of time the feedback 11 PAGE 12 is reconnected and the knowledge of the current outputs/states are utilized to bring down the system error. Once the error is reduced to near zero, the feedback may be disconnected again and the previous cycle may resume. Switch Controller Plant Figure 11: Schematic of intermittent feedback control In this study we consider linear time invariant systems with bounded inputs and the states are assumed to be available as outputs. As in most common situations, we assume that the system parameters are uncertain but are known to lie within some bounds. The Euclidean norm of the state is taken as a measure of the system error at any time instant, which is required to be always less than a prespecified upper bound. The norm of the initial condition is assumed to satisfy this bound. Under these assumptions the objectives for the open loop operation may be outlined briefly as follows. Find, for any permissible uncertainty, the maximal time interval for which the system error does not exceed the allowed limit. The system error is not monitored during open loop operation. Hence, such a worstcase optimization is considered to guarantee acceptable performance. A special input function is calculated to achieve maximal duration of the open loop period. Note that this is legitimate even when the loop is open, as long as no knowledge of the feedback signal is used for its computation. Existence of the time optimal controller is proved and EulerLagrange type conditions are derived for calculating the time optimal open loop input. It is shown that this time optimal input signal is either purely bangbang or can be uniformly approximated by a bangbang signal. 12 PAGE 13 Finally, it should be noted that, corresponding to the optimal open loop input, an infinite number of system trajectories might be realized depending on the particular value the uncertainty takes within the allowed set The main emphasis of this formulation is to guarantee that none of the trajectories, from the infinite number possible, can exceed the allowed error bound for the maximal time interval of operation. Earth Target Loss of Feedback Signal Obstacle Launch Site Acceptable Flight Envelope Figure 12: Loss of feedback due to communication channel disruption We conclude this introductory chapter with a hypothetical example to illustrate our problem formulation: Consider the simulated firing of a groundtoground missile as illustrated in the figure. The flight path of the missile is predetermined and the missile is regulated carefully so that it follows the prescribed flight path. The feedback about the current coordinates, velocity, angles, etc., of the missile is transmitted to the ground controller via a satellite. The controller uses this information to determine the angle and amount of thrust needed to keep the missile on track. Under this scenario even a temporary loss of signal from the satellite may lead to the catastrophic failure of the control and a subsequent loss of the missile. We are proposing a 13 PAGE 14 solution to the problem by defining the best way to design the control of the missile for the duration of no feedback. Our method also guarantees that the missile can stay within the acceptable flight envelope for a maximal time after loss of signal from the satellite occurs. We prove that the best control under these circumstances is either bangbang or can be approximated by a bangbang signal. 14 PAGE 15 CHAPTER 2 LITERATURE REVIEW Related Work in MinMax Optimal Control Problems The objective of time optimality, for every possible uncertainty, in open or closed loop operation embeds our problem within the framework of minmax optimal control. This area of differential game theory has been extensively researched and encompasses a wide variety of interesting results. We review a few papers directly related to our work. Isaacs (1954) was the first investigation to formulate a control problem in which two players with conflicting interests interact with each other in a game theoretical setup. These results were later published as a book in Isaacs (1965). The socalled main equation was derived, which can be viewed as a game theoretical extension of the HamiltonJacobiBellman equation and the dynamic programming approach (Bellman 1957). Independently, Kelendzheridze (1961) solved a pursuitevasion problem, which in turn spawned a body of research of which Pontryagin et al. (1962), Pontryagin (1966, 1967a and b), Mishchenko and Pontryagin (1967) and Mischenko (1971) were early contributors. These papers generally derived EulerLagrange type necessary conditions for the equilibrium solutions to the pursuitevasion problem. Berkovitz (1964 and 1967) formalized Isaacs results in a classical calculus of variations setting and also derived necessary conditions for general systems under weak assumptions. Other early contributors were Fleming (1961 and 1964), Ho et al. (1965), Friedman (1971), Elliott and Kalton (1972) and Elliott et al. (1973). Numerous authors developed different definitions of the equilibrium solution and various payoff functions of which Linear Quadratic Games have been extensively developed. A complete set of references can be found in Basar (1982). However the majority of interesting results have been concentrated for differential games with a saddle point solution. The problem 15 PAGE 16 of worstcase optimization without the saddle point assumption was posed by Feldbaum (1961 and 1965). This problem was addressed by, among others, Koivuniemi (1966), Howard and Rekasius (1964) and Bellanger (1964). Witsenhausen (1968) solved the same problem with a convex cost functional for sampled data systems. As discussed next, the problem was mathematically solved by Warga (1965a). The concept of relaxed solutions in optimal control theory was introduced by Young (1937) and later in the book Young (1969). In a series of papers Warga (1965a and b, 1970 and 1971a and b) solved the minmax control problem using relaxed solutions with quite general assumptions. The existence of relaxed solutions (in the sense of Young 1937) to the minmax problems were guaranteed and the solutions were found to satisfy variants of EulerLagrange type necessary conditions. The techniques used were similar to those of Neustadt (1966 and 1967) and Gamkrelidze (1965). These results were reworked and accumulated into a book (Warga 1972), from which we use a result on conflicting controls (Theorem IX.1.2) to derive necessary conditions for our problem. We quote a simplified version of the actual theorem in Chapter 4. It should be noted that with the exception of papers related to the Isaacs main equation, almost all the work reviewed above looked for open loop solutions. Apparently little has been done in search of closed loop necessary conditions characterizing solutions to the minmax problem. Conditions for existence of a stable solution were formalized with the help of concepts like stable bridges in Krasovskii and Subbotin (1974). This book contains a detailed exposition of positional, i.e., feedback solution to differential games. In a series of papers, Ledyaev and Mishchenko (1986, 1987 and 1988) and Ledyaev (1989 and 1994) derived necessary conditions for minmax control problems of a fixed duration. This has been an area of intensive research 16 PAGE 17 since but most of the advances have been related to the dynamic programming approach. See Vinter (2000) and the references therein. However efforts to derive EulerLagrange type necessary conditions have been limited. Throughout this study, results and ideas from standard optimal control and mathematics references have been used. Some of them are Kolmogorov and Fomin (1957 and 1961), Liusternik and Sobolev (1961), Rudin (1966), Bryson and Ho (1969), Luenberger (1969), Balakrishnan (1971), Hirsch and Smale (1974), Halmos (1982) and Zeidler (1985). Other Approaches for Reducing Feedback Requirements As outlined in the introduction, the main objective of this work is to reduce the duration during which the feedback loop has to be closed for controlling an uncertain plant. As far as we are aware of, such a robust time optimal formulation for reducing feedback requirements has not been dealt with in literature. However, in the context of Network Control Systems (see Nair et al. 2007), model based approaches have been developed for controlling systems with intermittent feedback. Design methodologies, in which ideal models of the system were used to guess the output of the system during the open loop, were proposed by Zhivogyladov and Middleton (2003) and Montestruque and Antsaklis (2004). Estimates of the maximum time the feedback signal may be delayed were calculated in Walsh et al. (1999). However in these works, the problem of finding the best open loop input was not dealt with and consequently, the maximum time for acceptable open loop operation has not been calculated. In conclusion, our work may have some implications for the problem of bandwidth reduction or control under communication constraints usually addressed in Network Control Systems (see Nair et al. 2007). Residence Time Control The problems of pointing and the related concept of residence time control are similar to our formulation of the problem. (See Meerkov and Runolfsson 1988 and the references therein). 17 PAGE 18 The pointing problem was investigated among others by Skelton (1973), Eng (1979), Cannon and Schmitz (1984) and Halyo (1983). In Meerkov and Runolfsson (1988), the residence time control problem was formulated as that of choosing a feedback control law, so as to force the system states to remain, at least on the average, within prespecified limits during some minimal period of time in spite of the disturbances that are acting on the system. The objective is very similar to that treated in this work, but the use of feedback creates a fundamental difference. Recall that in our hypothesis the feedback signal is completely absent over the period of interest. Moreover the maximum residence time was also not calculated in these papers. 18 PAGE 19 CHAPTER 3 MATHEMATICAL PRELIMINARIES In this chapter we will cover some of the wellknown results from functional analysis, measure theory and linear algebra that have been used repeatedly in this dissertation. This chapter is meant as an easy reference for the results derived in the later chapters, and is an attempt to make this thesis tolerably selfcontained. We start by a selection of relevant results from the theory of normed linear spaces. Most of these standard results are taken from Liusternik and Sobolev (1961), Halmos (1982) and Zeidler (1984). Normed Linear Spaces Weak Convergence First we recall the definition of weak convergence in a normed linear space. Definition 3.1: Let E be a normed linear space, {x n } a sequence of elements of E and x o E. If for all functionals fE (where E is the space conjugate to E), the sequence f(x n ) f(x o ) as n, then we say that {x n } converges weakly to x o and we write x n w x o Here x o is the weak limit of the sequence {x n }. Since Hilbert spaces are selfconjugate, the above definition may be specialized to the following form for Hilbert spaces: Definition 3.2: A sequence {x n } in a Hilbert space H converges weakly to x o H if x n ,y x o ,y as n for all yH. Then we write x n w x o where x o is the weak limit of the sequence {x n }. A continuous functional on a compact set in a normed linear space is bounded and achieves is maximum and minimum. However, in this dissertation, most of the sets we would be interested in are not compact in the sense of strong convergence. Weak convergence and weak 19 PAGE 20 compactness are much less severe requirements and will shown to hold for the sets of interest in this work. Weak compactness is defined next. Definition 3.3: A set XE is weakly compact (i.e., sequentially weakly compact) if every infinite sequence {x n }X has a subsequence {x nk } such that x nk w x o and x o X. The existence of solutions to optimal control problems is very much dependent on the following very important theorem. Alaoglus theorem Theorem 3.4: Every bounded sequence of elements in a Hilbert space contains a weakly convergent subsequence. This theorem guarantees the weak compactness of bounded sets in the Hilbert space. However for compactness in itself, the set of interest must be closed in the weak topology. The following result is useful for checking the sequential weak closure of convex sets: Theorem 3.5: A bounded strongly closed convex set in a Hilbert space is also weakly closed. In the analysis of existence of optimal solutions to dynamic optimization problems, we are actually interested in the functionals defined on compact or weakly compact sets. We define the concepts of weakly continuous and weakly uppersemicontinuous functionals. Weak upper semicontinuity Definition 3.6: A functional f (possibly nonlinear) on a Hilbert space H is defined to be weakly continuous at x0H, if for any sequence xnH and x n w x o we have f(x n ) f(x o ). Definition 3.7: A functional f (possibly nonlinear) defined on a Hilbert space H is defined to be weakly upper semicontinuous at x0H, if for any sequence xnH and x n w x o implies nlim sup f(x n ) f(x o ). 20 PAGE 21 A generalization of the classical theorem of Weierstrass to semicontinuous functionals is often used to prove existence of solutions in optimization problems. A simplified statement is listed next: Theorem 3.8: Let f be a weakly upper semicontinuous real valued functional on a weakly compact subset S of a Hilbert space H. The functional f is bounded above on S and achieves its maximum on S. In differential game theory it is often not enough to study a single upper/lower semicontinuous functional but a family of such functionals. Next we state a very useful property of a family of upper semicontinuous functionals: Theorem 3.9: (Willard 1970) Let X and A be two topological spaces. If f is a upper semicontinuous real valued function on X for each A, and if inf f (x) exists at each xX, then the function f(x) = inf f (x) is upper semicontinuous on X. Definition of Borel, Lebesgue and Radon Measures The interplay between measure theory and functional analysis is extremely common in dynamic optimization and especially differential game theory. We briefly define a few standard and wellknown concepts for easy reference. For detailed exposition on these concepts see Kolmogorov and Fomin (1957) and Rudin (1966). If S be a topological space, then the smallest field containing every open set in S is called the Borel field of sets, and denoted (S), and the elements of (S) are called Borel sets. A measure defined on (S) is called a Borel measure. Let be a Borel measure. A Borel set E(S) is regular if both of the following properties hold: (E) = inf {(V): EV, V open} (E) = sup{(K): KE, K compact} whenever E is open or (E). 21 PAGE 22 If every Borel set in S is regular then is called regular. A finite regular Borel measure is called a Radon measure. Radon measures are important from the point of this work, because of their appearance in the Riesz representation theorem for characterizing linear functionals. Finally, the Lebesgue measure on the real line R can be defined simply as follows: Let (R) be the Borel field defined on the real line and consisting of intervals. Then the set function that assigns the interval a,b], the measure (ba) is called the Borel measure on (R). The completion of (R) relative to the Borel measure is called the class of Lebesgue measurable sets. The extension of the Borel measure to the completion of (R) is called the Lebesgue measure. Riesz Representation Theorem We present here a special form of the wellknown Riesz Representation theorem, which will be used in the later chapters. This specific form is a slight simplification of that presented in Evans and Gariepy (1992). Theorem 3.10: Let P be a compact subset of Rm. Denote the space of continuous functions mapping P to R by C(P,R) and each element in this space by f. Let L be a bounded linear functional on the space C(P,R). Specifically L: C(P,R)R. Then there exists a positive Radon measure on P and a measurable function : PR such that: (x) = 1 for a.e. xP. L(f) = P f d for all fC(P,R). Fubinis Theorem Theorem 3.11: Let (X,X,) and (Y,Y,) be finite measure spaces, and let f be an (XY)measurable function on XY. If 0 f and if 22 PAGE 23 (x) = Yfx d and (y) = Xfy d (xX, yY) then is Xmeasurable, is Y measurable, and X d = XYf d() = Y d Separation of Convex Sets Let H be a topological vector space. A set A H is convex if {a + (1)b} 0 1}A whenever a,bA. We say that a linear functional H* (where H* denotes the space conjugate to H), separates subset A and B of H if 0 and either (x) (y) or (x) (y) (xA, yB) for some R. Theorem 3.12: Let K and M be nonempty disjoint convex subsets of a topological vector space H and the interior of K is nonempty. Then there exists H* that separates K and M. Directional (Gateaux) Derivative Definition 3.13: Let X be a vector space, Y a normed space, and T a (possibly nonlinear) transformation defined on a domain DX and having range RY. Let x0DX and let h be arbitrary in X. If the limit: DT()x;h = 0lim T(x0+h) T(x0) exists, it is called the Gateaux differential of T at x0 with increment h. If the above limit exists for each hX, the transformation T is said to be Gateaux differentiable at x0. We note that this definition makes sense only when x0+h D for all sufficiently small and that the limit is taken in the usual sense of norm convergence in Y. If Y is the real line and the transformation T is a real valued functional on X, then the following definition may be used for the Gateaux differential: 23 PAGE 24 DT()x;h = d d T(x0 + h)=0 for each fixed x0X. Existence and Uniqueness of Solution to a LTI System In this section we quote a widely known theorem about the existence and uniqueness of the solution to a differential equation with measurable right hand sides. In this dissertation, we consider linear systems of the form .x(t) = Ax(t) + Bu(t) where u(t) is Lebesgue measurable. This theorem is used throughout to guarantee existence and uniqueness of solution for the systems under consideration. This particular version of the theorem is taken from Young (1969). Theorem 3.14: Let f(t,x) be a vector valued function with values in xspace, and suppose that in some neighborhood of (t 0 ,x 0 ), f(t,x) is continuous in x for each t, measurable in t for each x, and uniformly bounded in (t,x). Then there exists an absolutely continuous function x(t), defined in some neighborhood of t 0 such that x(t 0 ) = x 0 and that, almost everywhere in that neighborhood: .x(t) = f(t,x(t)) Suppose in addition that for some constant L, the function f(t,x) satisfies, whenever (t,x 1 ) and (t,x 2 ) lie in some neighborhood N of (t 0 ,x 0 ), the Lipschitz condition  f(t,x1)f(t,x2) L x1x2, then in some neighborhood of t 0 there exists one and only one absolutely continuous function x(t) such that: x(t) = x0 + t0t f(,x()) d Jordan Canonical Representation Theorem 3.15 (Hirsch and Smale 1974): Let A be a nn matrix with real elements. Then there exists a nn real matrix P such that A d = PAP 1 is composed of diagonal blocks of the following two forms: 24 PAGE 25 S 1 = 1 1 .. 1 or S 2 = F I2 F I2 .. F I2 F The diagonal elements denoted by is a real eigenvalue of A and appears as many times as the multiplicity of the minimal polynomial. In the expression for S 2 F = a bb a and I2 = 1 00 1 where (a ib) are the complex eigenvalues of A. Each block F is repeated as many times as the multiplicity of the eigenvalue a + ib in the minimal polynomial. Ad is called the Jordan real canonical form of the matrix A. 25 PAGE 26 CHAPTER 4 NOTATION AND PROBLEM FORMULATION Notation The systems we consider are linear time invariant continuoustime systems given by a realization of the form (4.1) .x(t) = Ax(t) + Bu(t) x(0) = x 0 Where A and B are constant real matrices of dimensions nn and nm respectively. We assume that the state x of the system is available as output. The initial condition of the system is denoted by x 0 and is given. In addition, we assume that there is an uncertainty about the values of the entries of the matrix A and B. To represent these uncertainties, we introduce two matrices D A and D B We denote (4.2) A = A + D A and B = B + D B where A, B, D A and D B are given constant matrices of appropriate dimensions. While A and B are known, the matrices D A and D B can take any values within the sets A and B respectively. We define A and B to be bounded subsets of the spaces of real (nn) and (nm) matrices such that A = {D A : D A  d} and B = {D B : D B  d}. Here the norm D = Dijsup Dij where Dij denotes the ij th element of the matrix D and d is a fixed positive real number. Note that the sets A and B are closed, bounded and hence compact. The particular values D A and D B takes in the sets A and B are not known a priori and hence represents the uncertainty in the system matrices A and B. Such an assumption is realistic because though the exact values of the system parameters are never known, usually the sets in which they belong are known a priori. For compactness of expression we group the uncertainties together and define D = (D A ,D B ) belonging to the set = A B It is further assumed that at least one 26 PAGE 27 eigenvalue of the nominal system matrix A has a nonnegative real part and that the pair (A,B) is stabilizable for every value of the uncertainty set We will show in Chapter 5 that the assumption that the nominal system is unstable guarantees the existence of a finite solution to the open loop problem. We assume that the system is performing acceptably if the Euclidean norm of the states, representing the system error, stays below a prespecified positive bound (say M). In other words, the inequality xT(t)x(t) M, must hold as long as the system is operating. However once the state norm exceeds the bound, the system must either be stopped or some action must be taken to prevent this from happening. Here we assume that at the point in time when the feedback signal stops, the system norm denoted as the initial condition in our problem, is less than M. (4.3) x 0 T x 0 M Then the open loop is allowed to run as long as xT(t)x(t) M, and the loop must be closed when this inequality can be no longer made to hold. As outlined in the introduction, we assume that there is a special input during the time when the feedback signal is absent. This input is used to keep the system error within the tolerated level for a maximal amount of time. We denote this input by u(t). For u(t) we will use weighted norm and inner product as described below. Here u(t) is a vector valued function (u 1 (t), ..., u m (t)) T where each component is real valued and Lebesgue measurable. Each u(t) is assumed to lie in the Hilbert space L(t)2m with the following inner product: let x(t) and y(t) be two elements of L(t)2m. Then the inner product x,y = 0(t) x(t)Ty(t)dt, where (t) = e t for some fixed constant 0. The norm and the metric are defined correspondingly. Let U be a 27 PAGE 28 set of bounded functions in L(t)2m defined as U {u L(t)2m : maxi ui(t) K, t [0,]} where K is a fixed positive number. This set U of bounded measurable functions define the prospective inputs we shall consider for system (4.1). As will be shown in Chapter 4, the optimal solution is bangbang with possibly an infinite number of switches. Hence a smaller set, e.g., the set of piecewise continuous bounded functions, would not suffice. The set U thus defined will be shown to be compact in the weak topology and thus will facilitate the proof of existence of the optimal solution. For engineering purposes, any element in U can be approximated to any arbitrary accuracy by a piecewise continuous implementable function. Statement of the Problem We next state the mathematical problem formulation. During the open loop period the objective is to maximize the time during which the system error is guaranteed to stay below certain bound in the face of uncertainty. This objective is achieved with the help of the special input u(t) used to correct the system during the time the feedback is absent. In the introduction we described that how in various applications it is important to know both the correcting openloop input u(t), and the maximal amount of time for which the system can be left running in open loop safely. Keeping this motivation in mind we pose the following problem: Problem Statement: Using the notation of the last section the open loop problem may be formulated as follows. Problem 4.4: Find uUmax t f subject to the following constraints: .x(t) = Ax(t) + Bu(t) 0 t tf x(0) = x0 x0Tx0 M xT(t)x(t) M for 0 t tf and for all (A,B) (A+A)(B+A) where (A+ A ) {A+ D A : D A A } and (B+ B ) {B+D B : D B B }. 28 PAGE 29 Among the main objectives of this work is to show that such a maximum t f exists, is finite and that it can be achieved by an input u(t)U. Assuming they exist, the interval [0,t f ] and the optimal input solution u(t) to the above problem has the following feature. If u(t) is applied over the interval [0,t f ], irrespective of the value the system matrices (A,B) takes in the sets (A+ A ) and (B+ B ), the system trajectory stays within the allowed error level, i.e., x T (t)x(t) M for the entire interval [0,t f ]. The second objective is to find the nature of the optimal input u(t) that achieves the maximum tf. This will be achieved through the derivation of first order necessary conditions characterizing the optimal solution. 29 PAGE 30 CHAPTER 5 EXISTENCE OF A SOLUTION In this chapter we prove the existence of a solution to Problem 4.4. This will be achieved through the application of the generalized Weierstrass theorem stated in Chapter 3 (Theorem 3.8). Before going into the mathematical details we describe a brief outline of this chapter. The main idea is to repose Problem 4.4 in turns of a suitably defined upper semicontinuous functional. This functional is nothing but the minimum possible time (corresponding to the worst uncertainty) of open loop operation for any given uU. We prove that the set U defined in chapter 4 is compact in the weak topology and then we show that the functional mentioned above is weakly upper semicontinuous. Thus we are ready to use a generalized Weierstrass theorem to prove that a maximal time for the open loop operation exists, is finite, and can be achieved with an input u from the set U we defined in Chapter 4. Weak Compactness of U We start with a few properties of the input set U defined in Chapter 4. Most of the standard results of this section are taken from Liusternik and Sobolev (1961), Halmos (1982) and Zeidler (1984) and has been listed for reference in Chapter 3. We prove the weak compactness property of the set U of inputs, in the form of a lemma This property, which is a consequence of the Alaoglus theorem (Theorem 3.4)), along with a few wellknown properties of semicontinuous functions on compact sets, is essential for proving the existence of a solution to Problem 4.4. Lemma 5.1: The set U is weakly compact in the Hilbert space L(t)2m. Proof: The set U is obviously bounded and hence by Theorem 3.4 every infinite sequence in U has a weakly convergent subsequence. However we need to show that the set U is weakly closed. Since U is convex by definition, by Theorem 3.5 we just need to show that U is 30 PAGE 31 strongly closed. Consider a sequence of functions u n (t)u o (t) where u o (t)U. Hence for some set t[0,) of nonzero measure, u oi (t)K, where tt and u oi denotes the i th element of u o (t). Hence the following inequality must hold: 0etun(t)uo(t)2dt = tetun(t) uo(t)2dt + C n where C n is positive real. tet(K uoi(t))2 dt + C n As n, while C n can tend to zero, the first term remains constant and finite. So u n (t) does not converge to u o (t). Alternative Problem We now define an alternative formulation for Problem 4.4 using the functional defined next. It is easier to show the existence of the optimal solution in terms of this functional. Definition 5.2: Let J(t) = x(t) T x(t) where t[0,). Define T(u,D) = inf t:J(t) > M if J(t) > M for some t< otherwise Here x(t) is related to u(t) and D through Equation 4.1. This functional may be interpreted as the time when the system error exits the allowed envelope of operation for the first time. This functional is identical to the quantity t f defined in Problem 4.4, for trajectories obeying the constraints. In other words, for a fixed (u,D) pair, the relation x(t) T x(t) M is satisfied for the interval [0,T(u,D)]. Moreover for a fixed u(t), the relation x(t) T x(t) M is satisfied for every value of the uncertainty D within the interval []0 Dinf T(uD). Hence it can be easily seen that Problem 4.4 is equivalent to finding uUsup Dinf T(u,D), if it is finite. In addition we need to show that the supremum over U can be achieved. We summarize in the following restatement of Problem 4.4: 31 PAGE 32 Problem 5.3: Show that uUsup Dinf T(u,D) Find u U such that Dinf T(u ,D) = uUsup Dinf T(u,D). Well posed: For this problem to be well posed, the first step is to show that Dinf T(u,D) This observation is crucial to us since we are interested in running the system on open loop as long as possible. We realize that no matter what input we create for the system, nature can choose a value for D that will make the system trajectory leave the set of acceptable operation in finite time. First we prove a preliminary lemma: Lemma 5.4: Let A and D be a real nn matrices such that A has repeated eigenvalues and D. Also let x 0 be a fixed nonzero vector in R n Then for any d > 0, there exists a D such that D < d and (A+D) has distinct eigenvalues. Moreover let T be a nn real matrix such that Q=T(A+D)T 1 results in a Q which has the following block diagonal form: (5.5) Q = Q1 Q2 Qr Qr+1 where Q i = ai bibi ai for i=1,.,r and Q r+1 = a2r+1 a2r+2 .. an. Here (a i ib i ), i=1,,r represents the complex eigenvalues and a i (i=2r+1,.,n) represents the real eigenvalues of A. Then D can be chosen such that j=1n T 1j x 0j 0 where T ij is the ij th element of T and x 0j is the j th element of x 0 Proof: Let A d = PAP 1 where A d is the real Jordan canonical form of A. Recall the A d is composed of diagonal blocks of the following two forms: 32 PAGE 33 S 1 = 1 1 .. 1 or S 2 = F I2 F I2 .. F I2 F where is a real eigenvalue (possibly repeated) of A. In the expression for S 2 F = a bb a where (a ib) are the complex roots of A (possibly repeated). (Theorem 3.15) Let E be a diagonal matrix that we add onto A d such that blocks S 1 and S 2 are changed as follows: S 1 = 1 1 2 1 .. q1 1 q Where let q be the multiplicity of the eigenvalue S 2 = F1 I2 F2 I2 .. Fs1 I2 Fs Where let s be the multiplicity of the eigenvalue (a ib) and F i = a + i bb a + i i=1,2,s. Now the eigenvalues corresponding to block S 1 and S 2 are distinct and of the form ( + i ) and (a + j ib) where i=1,2,q and j=1,2,s. (Note: det(I S2) = det(I F 1 ) det(I F 2 ).. det(I F s ) ) Thus (A d + E) has distinct eigenvalues. From the construction above it can be seen that it is possible to choose such an E such that E < r for any r > 0. Now consider the reverse transformation: P 1 (A d + E)P = A + P 1 EP 33 PAGE 34  P 1 EP k P 1 EP 2 k P 1  2 E 2 P 2 = kE 2 The first inequality follows from the equivalence of norms where k is a finite constant and  2 is the spectral norm of the matrix. Clearly by making E 2 < dk we get that  P 1 EP d. Hence by defining D = P 1 EP we achieve our objective. Now let j=1n T 1j x 0j = 0 with a certain choice of A with distinct eigenvalues and with a T that transforms A according to Equation 5.5. Clearly, the vector T 1 [T 11 T 12 T 1n ] is an eigenvector of the matrix A and hence must satisfy 1 T 1 = AT 1 where 1 is the eigenvalue corresponding to T 1 Now from the continuity of the eigenvalues of A w.r.t. the entries of A, we can change A to (A+D) arbitrarily keeping the eigenvalues distinct. Then the following equation must hold: 1 T 1 = (A+D)T 1 Since D is arbitrary, T 1 can be made noncollinear to T 1 keeping T 1 T 1  arbitrarily small. It follows that T 1 and T 1 cannot be both orthogonal to x 0 Theorem 5.6: Consider the system (4.1) with a fixed initial condition x 0 in the set {x 0 : 0 < xT0x 0 M}. Then for any fixed u(t)U and M < there exists a D such that T(u,D) < Proof: Assume that for some fixed u(t)U, T(u,D) = for all D. This implies that x(t) T x(t) M D and t[0,]  e At [x 0 + 0teABu()d]  M. Now fix A at some arbitrary value in (A + A ) such that A has distinct eigenvalues. By Lemma 5.4 this can be done without loss of generality. Then we do a similarity transform on A 34 PAGE 35 such that Q = T 1 AT is in the block diagonal form shown in Equation 5.5. Note that in this form Q and Q T are commutative. Now we consider the system transformed by T and name the new state vector z(t). It is related to x(t) by the following relation: z(t) = Tx(t). We further assume that T is such that z 01 = j=1n T 1j x 0j 0. By Lemma 5.4 this also can be done without loss of generality. We seek to prove that x(t) can be made divergent with a proper choice of D. We claim that it is enough to prove z(t) diverges. z(t) = Tx(t) Tx(t) Hence tlim z(t) = tlim x(t) = In terms of z(t) our hypothesis implies: z(t) T z(t) T M = M D and t[0,]  e Qt [z 0 + 0teQBu()d]  M where M= T M and B = TB Let F(t) = [z 0 +0teQBu()d]. Therefore, [ e Qt F(t) ] T [ e Qt F(t) ] M F(t) T e(QT+Q)t F(t) M F(t) T et F(t) M Where = Q T + Q is a diagonal matrix with at least one of the elements positive, by hypothesis. For simplicity assume that the 1 st diagonal element denoted by a 1 is positive. Let us denote 35 PAGE 36 each entry of F(t) by f i (t), i.e., F(t)= f1(t)f2(t)fn(t). Thus for the last equation to hold at least the following must be true. (5.7) tlim f 1 (t) = 0 Now f 1 (t) may be written in a general form in the following way: f 1 (t) = z 01 + 0tea1t[cos(b1) j=1mB1juj() + sin(b1) j=1mB2juj()]d where (a 1 + ib 1 ) is the first eigenvalue of Q and B ij is the ij th element of the matrix B. Then Equation 5.7 implies that tlim z 01 + 0tea1[cos(b1) j=1mB1juj() + sin(b1) j=1mB2juj()]d = 0 z 01 + 0ea1 [cos(b1) j=1mB1juj() + sin(b1) j=1mB2juj()]d = 0 But this is a linear equation in B ij and hence cannot hold over all values of B in the set T(B + B ) { T(B + D B ) : D B B } unless u j (t) 0 for every j=1,,m and for all t[0,]. However if this is true then z 01 = 0. Hence we have a contradiction. In the above theorem we showed that, under the assumptions made in Chapter 4, the system norm x(t) can diverge for any given input u(t) for a proper choice of the uncertainty D. This is true for arbitrarily small uncertainty sets. Now define T(u) Dinf T(u,D), i.e., T(u) is the minimum time corresponding to u(t) in which the system norm x(t) can be made to 36 PAGE 37 escape the set [0,M]. According to Theorem 5.6 for any fixed u(t)U the quantity T(u) is finite. Corollary 5.8: If the conditions of Theorem 5.6 are satisfied then for any fixed u(t)U, the inequality Dinf T(u,D) holds. Weak Upper Semicontinuity of T(u,D) Recall that according to Problem 5.3 our objective is to show that uUsup Dinf T(u,D) Define T(u) Dinf T(u,D). Thus it would be enough to show that T(u) is weakly upper semicontinuous in u(t). Along with the fact that the set U is weakly compact (Lemma 5.1)), this would effectively prove that the functional T(u) is bounded over the set U and that the supremum is achieved. (Theorem 3.8). The upper semicontinuity of the functional T(u,D) in u(t), for each fixed D is demonstrated next. Denote the set of solutions for Equation 4.1 for all uU and D by X(U,). By Theorem 3.14, each element of X(U,) is an absolutely continuous function in t and unique corresponding to each (u,D) pair. Lemma 5.9: For a fixed D, T(u,D) is weakly upper semicontinuous in u(t), i.e., as a sequence of functions u n (t) w u o (t), the functional T(u,D) obeys the following relation: for any > 0, and T(u o ,D) we can choose an integer N such that T(u n ,D) T(u o ,D) for n N. Proof: For a fixed D and t consider the function x(u):UX(U,), where x(t;u) is the solution to Equation 4.1. As u n w u o (i.e., weakly converges), then for each t[0,), x(t;u n ) x(t;u o ), i.e., x(u n ) converges pointwise to x(u o ). This is because for each t[0,), x(t;u) is a functional linear in u. 37 PAGE 38 This can be seen by defining () = 1 if t 0 otherwise and noting that x(t) = e At [x 0 + 0()eABu()d]. Now define a functional T p : X(U,) [0,] as follows: T p (x(t)) = inf t: xTx > M if xTx > M for some t otherwise Next consider a sequence of function x n (t) x o (t) pointwise. We show that for any 0 there exists an integer N such that for n N T p (x n ) T p (x o ) If T p (x n ) T p (x o ) then the claim is true. So we assume that T p (x n ) T p (x o ). Let T p (x n ) = t n and T p (x o ) = t o Now we assumed that x n x o pointwise. By the definition of t 0 the following is true for every > 0: there is a t 1 (t 0 t 0 +) such that x T 0 (t 1 )x 0 (t 1 ) > M. Consequently, there is N such that x T n (t 1 )x T n (t 1 ) M [x T 0 (t 1 )x 0 (t 1 ) M]/2 0 for all n N. Therefore, t n t 0 + which implies that t n t o Thus we have shown that T p (x n ) T p (x o ) Now, for a fixed D, consider the composition map T(u,D) : U X Tp [0,]. From the above arguments it is clear that as u n w u o x n x o pointwise and for any 0 there exists N such that for n N, T p (x n ) T p (x o ) Hence T(u n ) T(u o ) Thus for a fixed D, T(u,D) is weakly upper semi continuous in u(t). Hence proved. Now we need to prove that the functional T(u) is also weakly upper semicontinuous. This is true due to Lemma 5.9 and Theorem 3.9. Corollary 5.10: The functional T(u) Dinf T(u,D) is weakly upper semicontinuous in u(t). 38 PAGE 39 Proof: The proof amounts to checking the conditions of Theorem 3.9. By Lemma 5.9), for each D, the functional T(u,D) is weakly upper semicontinuous in u(t). Also for each uU, the functional T(u) = Dinf T(u,D) 0. Hence the infimum exists always, and by the property above T(u) is weakly upper semicontinuous in u(t). Existence of u* The weak upper semicontinuity of T(u) along with the weak compactness of the set of inputs leave us posed for applying the general Weierstrass theorem (Theorem 3.8). Hence we can now conclude the following: Theorem 5.11: Let U and be as defined above and T(u) Dinf T(u,D) and let uUsup T(u) = uUsup Dinf T(u,D). Then Moreover there exists some u (t)U such that T(u ) = Proof: This follows directly by noting that the set U is weakly compact and the functional T(u) is weakly upper semicontinuous in u(t) over U. (Theorem 3.8). Hence T(u) is bounded above and it achieves it upper bound over U. In conclusion, we showed in this chapter that a solution to Problem 4.4 exists within the set of inputs U. The set U was shown to be compact in the weak topology. Then an alternative formulation of the Problem 4.4 was stated using a functional T(u,D). A proof of the wellposed ness of this new statement was presented next. In the process it was shown that under the assumptions made in Chapter 4, the system trajectory can leave the envelope of acceptable operation in finite time no matter what input we apply to the system. This inbuilt finiteness of the functional T(u,D), along with particular features of the formulation was utilized to prove weak upper semicontinuity of the functional T(u,D) for any fixed D. Another functional, namely 39 PAGE 40 T(u), denoting the worst possible time for the input u was introduced and it too was shown to be weakly upper semicontinuous in u. Lastly Weierstrass theorem was applied to prove the existence of the best input u* and the finiteness of uUsup T(u). The application of u guarantees that the system error does not leave the allowed envelope for any value of the uncertainty for a maximal duration of time. However, it should be noted that even for this optimal input, the system error can leave the allowed envelope in finite time for certain values of the uncertainty. 40 PAGE 41 CHAPTER 6 GENERALIZED FIRST ORDER CONDITIONS In Chapter 5 we proved that the optimal solution to Problem 4.4 existed among the set of bounded and measurable input functions. Evidently it would be useful to know more about the optimal solution, with a view to facilitate its calculation for specific problems. The conventional method for calculating solution to optimal control problems has been through characterizing first order necessary conditions. Here however the situation is complicated by the gametheoretic formulation, in that the maximization of open loop time has to be done over every possible value of the uncertainty. We draw on the considerable amount of previous research in the area of differential game theory and minmax optimal control. The relevant literature was reviewed in Chapter 2. In particular, we note that Problem 4.4 is an example of minmax or conflicting control problems studied previously among others, by Warga (1970). We use Theorem IX.1.2 (Warga 1970), for finding the first order necessary conditions characterizing the optimal solution. However we present a modified version of the theorem, which has been simplified to suit the problem we are solving in this dissertation. Simplified Theorem from Warga (1970) In this section we state and prove the theorem along with some required modifications. We require some new notation in this chapter: Notation 6.1: Let S be a compact set on Rn and be the Borel field of subsets of S. Then by rpm(S) we denote the set of all Radon probability measures defined on Moreover we denote the Banach space of continuous functions from SR with the supremum (L1) norm by the notation C (S,R). C denotes the closure of C and int(C) denotes the interior of C. 41 PAGE 42 First we require the following lemma. Let H be a topological vector space and W be a convex subset of RH We define W 0 { w 0 R: (w 0 ,w H )W} and W={w H : (w 0 ,w H )W, w 0 0}. Let H denote that space conjugate to H. Lemma 6.2: Let C be an open convex subset of H, 0int(W 0 ) and 0C. Then either there exists H H such that H 0, H (w H ) 0 H (c) for all w H W and cC or there exists wW such that w 0 0 and w H C. Proof: For every W let 0 0 and H C be outside W. Recall that W={w H : (w 0 ,w H ) W, w 0 0}. Then W is a nonempty convex subset of H and W C = Since C is an open convex set, by the separation property(Theorem 3.12), there exists H H and R such that H 0 such that H (w H ) H (c) where w H W and cC. But 0C W and hence =0. Thus we have (6.3) H (w H ) 0 H (c) (w H W and cC) This is the first alternative of the lemma. The remaining possibility is when there exists wW such that w 0 0 and w H C. Hence we have the second alternative of the lemma. The above lemma will be used in the proof of the following theorem. Next we state and prove the simplified version of Theorem IX.1.2 from Warga (1970). Theorem 6.4: Let Q be a convex set, F be a compact set in R and P be any compact finite dimensional metric space. Consider two functions T 1 :QF R and T 2 :QFP R. Now assume the following: 42 PAGE 43 1. Let H = {(q,f)QF : T 2 (q,f,p)[,M] for every pP } where ,M0. Let there be (q ,f )H such that T 1 (q ,f ) = (q,f)HminT 1 (q,f) 2. The functions T 1 and T 2 have convex Gateaux derivatives. 3. For each pair (q,f)QF the function T 2 (q,f,p):PR bounded and continuous in p. Then there exists rpm(P) and an integrable ~ : P R such that i. ~ (p)=1 for all pP and (P) 0 ii. ~(p) DT2((q*,f*),p;(q,f)(q*,f*))(dp) 0 iii. ~(p)T 2 (q ,f ,p) = a[M] max ~(p)a for a.a. pP where DT j ()(q*,f*);(q,f) (q*,f*) denotes the directional derivative of T j at (q ,f ). Proof: Define the set W(q,f) = { DT 1 ()(q*,f*);(q,f) (q*,f*), DT 2 ()(q*,f*)p;(q,f) (q*,f*) :(q,f)QF } This set is convex by hypothesis. Let C 2 = { c(p)C(P,R) : c(p) M } where 0. Note that C 2 is a closed convex subset of C(P,R) and let C= int(C 2 ) T 2 (q ,f ). Then we can apply Lemma 6.2 and the first alternative of the lemma yields the following: there exists C (P,R) such that 0, and for all (q,f)QF (6.5)(DT 2 ()(q*,f*)p;(q,f) (q*,f*)) 0 (6.6)(c) 0 (cC) In the above expressions note that C (P,R) and hence by the Riesz representation theorem (Theorem 3.10) we can specify an integral form of In particular, there exists a positive Radon measure on P and an integrable ~ : P R such that ~ (p) = 1 for all pP and (P) 0 43 PAGE 44 (c) P ~(p)c(p)(dp) c C(P,R) Next we investigate the inequality (6.6). In view of the definition of C this implies ( c T 2 (q ,f ) ) 0 (cC) P ~(p) [c*(p)c(p)](dp) 0 (cC 2 ) We claim that this equation implies claim (iii) of the theorem. To see this assume that there exists some c o C 2 such that ~(p)c*(p) ~(p)co(p) for pP P where (P) 0. Then we can form cC 2 such that c co when pPc* otherwise But, the integral inequality is not satisfied with c(p) since: P ~(p)[c*(p)c(p)](dp) = P ~(p)[c*(p)c(p)](dp) 0. Now note that C(P,R) is separable and hence C 2 contains a dense denumerable subset {c 1 ,c 2 ,.}. Moreover the set {c(p)cC 2 } is dense in [,M] for all pP and hence it follows that the set {c 1 (p),c 2 (p),} is dense in [,M] for all pP. So we can write the following: ~(p)c*(p) jNsup ~(p)cj(p) = a[,M] sup ~(p)a(pP) = a[,M] max ~(p)a(pP) (Since c (p)[,M]) Thus we have proved that the statement of the theorem is a result of the first alternative of the lemma. We now prove that the other alternative of the lemma implies a contradiction with the assumption in that (q ,f ) is found to be not minimizing. To prove this consider the set W(q,f) 44 PAGE 45 and C as defined above. From the lemma there exists wW(q,f) such that w 1 0 and w 2 C. For each choice of qQ and fF, define the function h 1 :R and h 2 : C(P,R) as follows: (6.7) h 1 () = T 1 (q + (qq ), f + (ff )) h 2 () = T 2 (q + (qq ), f + (ff ),p) where [0,1]. It can be easily verified that the derivatives h1 and h2 at (0,0) can be expressed in terms of the directional derivatives of T 1 and T 2 at (q ,f ). (6.8) hi(0) = DT i ()(q*,f*);(q,f) (q*,f*) (i=1,2) Now for sufficiently small h 1 () = h 1 (0) + h1(0)) = h 1 (0) + w 1 ) h 1 (0) = T 1 (q ,f ) Also we know that 0lim  1 [h 2 () h 2 (0) + h 2 (0) ] = 0 0lim  1 [h 2 () h 2 (0)] = w 2 C Hence for sufficiently small  we can conclude that [h 2 () h 2 (0)] CC (Since 0C) h 2 () T 2 (q ,f ) int(C 2 ) T 2 (q ,f ). Hence (q ,f ) is not a minimizing solution as hypothesized in assumption (1) of the theorem. The Scaled System The theorem proved above is the main tool that we will use to derive first order conditions for the solution of Problem 4.4. The functions T1 and T2 will be appropriately defined for our problem and it will be shown that all the assumptions of Theorem 6.4 are satisfied. Hence a 45 PAGE 46 solution to Problem 4.4 must satisfy the conclusions of Theorem 6.4. It will be shown that some useful features of the optimal solution of Problem 4.4 may be derived from careful consideration of conclusion (ii) of Theorem 6.4. With these objectives in mind we redefine Problem 4.4 for a scaled system described next. We express system (4.1) in terms of scaled variables y(s) and w(s) as follows: y(s) x(s) s[0,1] v(s) u(s) s[0,1] where and uUsup Dinf T(u,D). In terms of these variables Problem 4.4 can be written as: Problem 6.9: Find min () subject to the following constraints: .y(s) = (Ay(s) + Bv(s)) 0 s 1 y(0) = x0 x0Tx0 M yT(s)y(s) M for 0 s 1 and for all (A,B) (A+A)(B+B) where (A+ A ) {A+D A :D A A } and (B+ B ) {B+D B :D B B }. It can be easily checked that Problem 4.4 and 6.9 are identical. Using standard methods, y(s) can be expressed as: y(s) = e As [x 0 + 0seA Bv()d] Definition 6.10: Now we identify the relevant sets in Problem 6.9 with those defined in Theorem 6.4 as follows: Q U [0,1] where U [0,1] {v(t)U: v(t) = 0 for t 1 } F [0,] P { (A,B,s) (A+ A ) (B+ B )[0,1]} And we define the functions T 1 and T 2 as follows: T 1 (v(s),) = T 2 (v(s),, (A,B,s)) = y T y 46 PAGE 47 In the following Theorem P (A+ A ) (B+ B )[0,1]. Hence if rpm(P) then conditional probabilities (A,Bs) as well as the marginal probability (s) can be defined in the usual way so that for any (A,B,s), the joint probability (A,B,s) = (A,Bs)(s). Let us denote the Lebesgue measure on [0,1] by Also note that in the following theorem and thereafter the variables and s are both used to denote the scaled time and lies in [0,1]. They are used interchangeably as appropriate. Necessary Conditions We are now ready to apply Theorem 6.4 to Problem 6.9 and thus derive necessary conditions for the optimal solution (v*,*). From this theorem we expect to gain some insight about the characteristics of the optimal input v*. Indeed, from the theorem stated below it will become apparent that the optimal input v* may have a bangbang control feature. The result introduces the quantity z(s) over the scaled interval time [0,1], which is similar to the classical switching function for the bangbang control input v*.(Pontryagin et al. 1962). However the exact characteristics of z(s) will have to be clarified through some further investigations. The solution to Problem 6.9 and hence the solution to Problem 4.4 is summarized in the following theorem: Theorem 6.11: Let (v (s), ) be the solution to Problem 6.9. Then there exists rpm((A+ A ) (B+ B )[0,1]) and a measurable function z(s):[0,1] R m such that the following are satisfied: i. z(s) = s1 P yT()e*A(s)B(dAdB)(d) ii. z(s) v(s) z(s) v (s) for a.a.s[0,1] and for any vU [0,1] Where the support of is given by the set: = { (A,B,s)(A+ A ) (B+ B )[0,1]: y T y = M } 47 PAGE 48 Proof: We can apply Theorem 6.4 since all the assumptions are clearly met We first calculate the directional derivatives of T 1 and T 2 It is more convenient to use the variable as the time variable and s as the running variable in the following expressions. The advantage will be apparent from the subsequent derivation. DT 1 ()(v*,*);(v,) (v*,*) = ( ) DT 2 ()(v*,*);(v v*)=* = y T () 0e*A(s) *B(v(s)v*(s))ds By Theorem 6.4 there exists rpm(P) and an integrable integrable ~ : P R such that ~ (p) = 1 for a.a pP and (P) 0 ~(p) DT2((v*,*);(v,) (v*,*))(dp) 0 If we set = in the above inequality and refer to the expression for the directional derivatives of T 1 and T 2 we obtain (6.12) ~(p) DT2((v*,*);(v v*))(dp) 0. P~(p)yT() 0e*A(s) *B(v(s)v*(s))(ds) (dp) 0 P~(p)yT() 0e*A(s) *B[0](s)()v(s)v*(s)(ds) (dp) 0 (Where [0,] (s) = 1 when 0 s 0 otherwise ). 0 P ~(p)yT()e*A(s) B[0](s)(dp) (v(s)v*(s))(ds) 0 (By Fubinis Theorem (3.11)) 0 z(s)(v(s)v*(s))(ds) 0 where z(s) P ~(p)yT()e*A(s) B[0](s)(dp) 48 PAGE 49 Now by hypothesis, P {(A+ A ) (B+ B )[0,1]} and is a Radon probability measure on P. Hence we can write (dA,dB,d) in terms of the appropriately defined conditional and marginal probabilities. The expression for z(s) can be rewritten in terms of the conditionals as follows: z(s) 01 ~(A,B,)yT()e*A(s) B(dA,dB/) [0](s) (d) =s1 ~(A,B,)yT()e*A(s) B(dA,dB/) (d) We claim that: 0 z(s)(v(s)v*(s))(ds) 0 z(s)v(s) z(s)v (s) for a.a. s[0,1]. To see this assume that there exists some v o U such that z(s)v o (s) z(s)v (s) for sT [0,1] where (T) 0. Then we can form vU such that v vo when sTv* otherwise But, the integral inequality is then not satisfied with v(s) since: 0 z(s)(v(s)v*(s))(ds) = T z(s)(v(s)v*(s))(ds) 0. Now the third conclusion of Theorem 6.4 directly determines the form of the support set of the measure From Theorem 6.4): ~(p)y T (p;v )y(p;v ) = a[M] max ~(p)a for a.a. pP. 49 PAGE 50 Recall that ~(p) = 1. Hence: y T (p;v )y(p;v ) = M when ~ = 1. If ~ = 1 the equality above cannot be satisfied. Hence ~ = 1 for a.a. pP and the measure has the support set defined as follows: = {pP: y T (p;v )y(p;v ) = M}. The expression for z(s) simplifies to: z(s) = s1 yT()e*A(s) B(dA,dB/) (d). The above theorem derives necessary conditions for the optimal solution (v*,*) to Problem 6.9. Now consider conclusion (ii) of the above theorem. z(s) v(s) z(s) v (s) for a.a.s[0,1] and for any vU [0,1] It follows that when each component of z(s) 0 the optimal input solution v (s) is bangbang and alternates between the maximum allowed input bounds (namely K). In other words when zj(s) 0 for some j{1,2,m}, this implies that for a.a. s[0,1]: v *j (s) = K when zj(s)0 K when zj(s)0 This partially solves our questions about the characteristics of the optimal solution v*. However we cannot conclude anything about the corresponding components of the solution over intervals when some components of the function z(s) are zero. A completely bangbang solution would be extremely favorable from an engineering point of view, for its ease of implementation and numerical computation. However as we show in the following example, this may not always be true in this case. We will show that some components of the switching function z(s) can turn out to be zero over nonzero intervals of time. 50 PAGE 51 We would like to know more about the function z(s), in particular, whether some components of z(s) could be zero over contiguous subintervals of [0,1]. Recall that for the solution to be purely bangbang each component of the switching function z(s) needs to be nonzero almost everywhere on [0,1]. We note that our problem is linear time optimal and in most such problems studied in literature the solution turned out to be bangbang. (See Pontryagin et al. 1962). However it is interesting to note that in this case, it does not always hold. In general, it is not true that the optimal solution is bangbang and hence some components of z(s) are zero over subintervals of nonzero measure in [0,1]. We present the following example: Example 6.13: Find uUmax t f subject to the following constraints: .x(t) = ax(t) + u(t) 0 t tf and u(t) 2 x(0) = 1 xT(t)x(t) M for 0 t tf and for all a[1.21.4] We provide a solution to the problem for M = 1.96. Figure 61: Optimal input has one switch: M = 1.96, t f = 3.7 51 PAGE 52 The solution is bangbang only over the first part [0,1.27] of the interval. Over the rest of the interval until tff the optimal input though constant (=1.67) is not 2. For a clear understanding of the behavior of the system for different values of the uncertainty we have plotted the corresponding trajectories for ten different uncertainty values of the pole a. The solution to this problem was calculated using brute force techniques, where both the time axis (0 to 4 seconds) and the amplitude of the input [2,2] were discretized thereby forming a grid. Every input was checked on this grid to find the best solution. The continuous time solution was found by interpolation of the discrete optimal solution. Figure 62: Trajectories for ten different uncertainty values of a; M = 1.96, t f = 3.7 While this feature of the optimal solution v* is interesting because of its anomaly with other known solutions to linear time optimal problems, studied in literature, it is inconvenient from an engineering point of view. It is much easier to compute a bangbang solution than otherwise, since for a bangbang solution, only the switching instants need to be computed to 52 PAGE 53 effectively know the solution. Hence for a bangbang solution the dynamic optimization with high complexity reduces to a simple static optimization over the space of switching instants. In the next section we show, that even over intervals where some components of z(s) are zero and consequently the optimal solution v* is possibly not bangbang, there exists a bangbang input v which approximates v* in an appropriate sense. BangBang Approximation to v* As we discussed in the last section, it would be interesting to know if the optimal solution, which may not be bangbang, could be approximated by a bangbang input in any sense. Recall that the objective of the optimal input is to maintain the inequality x T x M for every trajectory of the error, corresponding to any uncertainty value, during the maximal time interval t f We show that the optimal input u can be approximated by a bangbang input u, in the sense that x T (u,D)x(u,D) M + for every value of the uncertainty D, for at least t f seconds. Here can be made arbitrarily small by increasing the number of switches in u. Lemma 6.14: There exists a bangbang function uU [0,] with a finite number of switches, such that x T (u;D)x(u;D) M + for every D and for at least t f seconds. Here, can be made arbitrarily small by increasing the number of switches in u. Proof: Let x (t,D) and x(t,D) denote the state vectors, corresponding to uncertainty value D(DA,DB) at instance t when the optimal control u and u are applied respectively between [0,t]. Since they start from the same initial condition, x (t,D) x(t,D) = eAt0teAB[u*u]d We claim that for small enough t it is possible to switch u appropriately so that 53 PAGE 54 eAt0teAB[u*u]d = 0 (A,B)((A+A)(B+B)) Since e At is invertible always, it is enough to prove that 0teAB[u*u]d = 0. Let t be small enough so that e At can be considered constant over [0,t]. We now propose to switch each component of u{u 1 ,u 2 ,.,u m } once between K during [0,t]. The interval before the switch we denote by t i1 and the interval after the switch by t i2 corresponding to the i th input u i The switching times have to add up to the total time t, i.e., t i1 + t i2 = t. Hence: 0teAB[u*u]d = e At Bu11t11 + u12t12 u21t21 + u22t22... um1tm1 + um2tm2 Where uij = ui ui (i = 1,2,m) and j = 1 denoted the time interval before the switch and j=2 implies the time after the switch. Clearly, we can choose the switching times so as to make u i1 t i1 + u i2 t i2 = 0 for every i{1,2,.m}. This proves that at time instant t we can make x (t,D) x(t,D) = 0 irrespective of the value of the uncertainty D(DA,DB), by switching each input once. The total number of switches required for all the inputs is exactly m. This does not guarantee that x (,D) = x(,D) for all [0,t] D. However because of the absolute continuity of x(t,D), by choosing t small enough we can guarantee that x (t,D) x(t,D) for any given 54 PAGE 55 Recall that the optimal open loop interval is always finite. Hence the optimal interval [0,t f ] can be split into arbitrarily small subintervals to form a partition P{0,t 1 ,t 2 ,.,t p ,t f }. We can use the above method for each subintervals [t i ,t i+1 ], (i=1,,p) so that x (t i ,D) = x(t i ,D) D for each t i P. Also by making the partition P fine enough, we can make x (,D) x(,D) ((t i ,t i+1 )) for any given It follows easily that x(t,D) M + for t[0,t f ] and D. Over intervals where some components of z(s) are 0, the above lemma provides an approximation of the optimal solution with a bangbang input. The approximation holds irrespective of the value of the uncertainty and hence guarantees that every state trajectory remains within the allowed error bound for the maximal period. Moreover, since the approximation is being done by a bangbang function, all the computational advantages corresponding to a bangbang solution are inherited by u. In general, for an accurate approximation, the number of switches required for u may be high. However for practical purposes, the required number of switches may be computed by repeatedly calculating the maximal time for increasing number of switches. The iteration should stop when no appreciable improvement occurs with the increase in the number of switches. We consider Example 6.13 again to show that the nonbangbang optimal input can be effectively approximated by a bangbang signal, as is predicted by Lemma 6.14. The approximating bangbang input has a total of sixteen switches. Corresponding to each uncertainty value of the pole a, the state trajectory with the actual optimal input is approximated by the trajectory resulting from the bangbang input. To see this the reader need to compare the following figures with Figures 61 and 62. A slight relaxation of the error bound beyond M=1.96 allows for an identical maximal time of 3.7 seconds. 55 PAGE 56 Figure 63: Approximate bangbang input: 16 switches Figure 64: Trajectories for ten different uncertainty values of a; M = 1.96, t f = 3.69 56 PAGE 57 CHAPTER 7 CONDITIONS FOR A PURELY BANGBANG SOLUTION Though the necessary condition derived in Theorem 6.11 hints that the optimal solution may be bangbang, we could not rule out the possibility that some components of the switching function z(s) could be zero over subintervals of [0,1]. In fact we discussed a case where this is the case and the optimal solution is clearly not bangbang. Furthermore, we proposed a method in which a nonbangbang optimal input may be approximated by a high frequency bangbang input. However it would be interesting to know whether there are conditions under which the optimal solution itself is purely bangbang. For this we need to investigate two important questions about the switching function z(s). Firstly, we need to find conditions under which each component of the switching function z(s) is nonzero almost everywhere. Moreover we need to estimate number of zero crossings of the components of the function z(s). The following assumptions were found sufficient to guarantee a purely bangbang solution to Problem 4.4. Assumptions 7.1: The uncertainty bound d is small. The input bound K is sufficiently small with respect to the error bound M. The system is controllable from each input for some value of the uncertain pair (A,B). We first prove that all the components of the function z(s) cannot be identically zero over the entire interval [0,1] for small uncertainties on the A and B matrices. Lemma 7.2: If the maximum disturbance d is small, the function z j (s):[0,1]R cannot be zero a.e. on [0,1] for all j{1,,m}. Proof: According to Theorem 6.11, for = 57 PAGE 58 DT2((v*,*);(v v*))(dp) = 0 z(s)(v(s)v*(s))(ds). Assume that z(s) = 0 a.e. on [0,1]. Then DT2((v*,*);(v v*))(dp) = 0 for any measurable v(s). Now, for = the expression for DT 2 ((v );(v v )) may be simplified as follows : DT 2 ((v );(v v )) = yT(*,v*) 0e*A(s) *B(v(s)v*(s))ds = yT(*,v*) 0e*A(s) *Bv(s)ds 0e*A(s) *Bv*(s)ds = yT(*,v*)e*A x0 + 0e*A(s) *Bv(s)ds y(*,v*) = yT(*,v*) {} y(*,v) y(*,v*) Now the support of was determined to be the set = {(,A,B)P: y T (,A,B;v )y(,A,B;v ) = M}. Hence the integral expression simplifies to: DT2((v*,*);(v v*))(dp) = yT(*v*)y(*v)(ddAdB) M (ddAdB) = yT(*v*)y(*v)(ddAdB) M According to our assumption: (7.3) yT(*v*)y(*v)(ddAdB) = M for every measurable v(t). Hence for v(t) = 0 for t[0,1] we have: (7.4) yT(*v*)eA*t x0 (ddAdB) = M. Next we use some Dirac delta and Dirac delta derivatives for v(t) to try to bring every trajectory to near zero. While these functions are not strictly measurable, they can be thought of as the 58 PAGE 59 limits of very high amplitude measurable functions and the proof is considerably simplified as a result. Moreover, note that under the assumption of xT0x0 M, and by the definition of z(s) can be expressed as the following power series in s in a small enough neighborhood about s = 0. z(s) = yT(,A,B')e*A B(dA,dB,d) + s yT(,A,B')e*A *AB(dA,dB,d) + + s22! yT(,A,B')e*A A2B(dA,dB,d) + .. It follows that z(s) is analytic in a small enough neighborhood about s = 0 and hence the assumption that z(s) = 0 a.e. implies that z(0) = 0. Choose v(t) = [(0) (0) n (0)] C 1 x 0 Since z(0) = 0, (7.3) is valid for such a v(t). Then: y(0+) = x 0 [ (B+D B ) (A+D A )(B+D B ) .. (A+D A )n(B+D B ) ] C 1 x 0 = [D B AD B +D A B ]C 1 x0 Hence with this input (7.3) yields: (7.5) yT(*v*)eA*t [DB ADB +DAB ]C1 x0 = M Assuming that D A and D B are small the left hand side of the above equation can be approximated by: yT(*v*)eA*t [DB ADB +DAB ]C1 x0 where D A and D B are average values with respect to the measure If we divide the above expression by the maximum allowed scalar disturbance, d (recall D A  d and D B  d (D A ,D B )), then [D B AD B +D A B ]C 1 x 0 = d [ h 1 Ah 2 + h 3 B ..] C 1 x 0 59 PAGE 60 where each element of the matrices h i lies between [1,1]. Hence for any measure and independent of the maximum disturbance d, the following upper bound holds: [ h 1 Ah 2 + h 3 B ..] C 1 x 0 k 1 x 0 where k 1 is a large enough scalar. yT(*v*)eA*t [DB ADB +DAB ]C1 x0 dk 1 yT(*v*)eA*t x0 dk 1 M (by Equation 7.4) Now as d 0 Equation 7.5 cannot be satisfied. Hence we have a contradiction. Lemma 7.2 shows that at least one component zj(s) of the function z(s) is nonzero over some part of [0,1]. The optimal solution is bangbang over those subintervals. In the following lemma we prove that if the input bound K is small enough compared to the allowed error bound M (by Assumption 7.1), then the optimal trajectories hit the boundary only at the end of the optimal interval [0,tf. We will show that this condition is not only enough to make optimal solution always bangbang, but also provides more information about the zeros of the function z(s). .2 denotes the Euclidean norm below. Lemma 7.6: If for some constants k 2 d 2 and d 1 defined below, (M) 1/2 (A 2 d 1 ) > k 2 (B 2 +d 2 ) K holds, then x T (t)x(t) < M for all D t[0,) and x T ()x() = M for some D. Proof: Recall that J(t) = x T (t)x(t). When some input u(t) is applied to the system, the open loop operation must stop when for some disturbance D=( A B ), the inequality J(t) M is satisfied. Now J(t) is a continuous and differentiable function of t on any interval in [0,). Let J(t 0 ) = M for some t 0 corresponding to some x(t) and D If for every permissible u(t 0 +), .J(t) t0+ 0, then for every > 0, J(t 0 + ) M and hence = t 0 We derive conditions so that .J (t) t0+ > 0 for any permissible input and any trajectory satisfying J(t 0 ) = M. 60 PAGE 61 .J (t) t0+ > 0 x T .x > 0 x T (Ax + Bu) 0. We assumed A to be unstable and hence, by the Lyapunov theorem, x T Ax 0. If x T Bu > 0, then the inequality above is trivially satisfied. Hence we assume that x T Bu 0 and the last inequality becomes: x T (Ax Bu) 0  xTAx  2  x T Bu  2  xTAx  2 x 2 .B 2 .u2 Now there exists 0 k 1 1 such that  xTAx  2 = k 1 x22.A 2 Let k 2 be the lower bound of the values of k 1 for any value of x(t 0 ) such that x T (t 0 )x(t 0 ) = M and A. Since by assumption x T Ax 2 0, we claim that k 2 0. This can be seen by noting that x T Ax 2 M (A 2 + d 2 ). Hence k 2 = xTx=MAinf xTAxM (A2+d2) Since the right hand side is continuous in x and A, the infimum is achieved and hence k 2 0. Hence the following implies the last inequality: k 2 x22 A 2 x 2 B 2 u 2 (M)1/2A2B2 k 3 K (where k 3 is a constant resulting from the equivalence of norms: recall that u 2 = k 2 k 3 u k 2 k 3 K) (M)1/2(A2 d1)B2 + d2 k 3 K The last implication follows from: A 2 =  A + D A  2 A 2 D A  2 A 2 d 1 B 2 =  B + D B  2 B 2 + D B  2 B 2 + d 2 where d 1 ,d 2 are positive constants resulting from the equivalence of norms and the bound d on the uncertainties: D A , D B  d. 61 PAGE 62 The significance of the above lemma is that it provides us with a simple condition when the optimal trajectories will hit the boundary only at one point in time and that time is necessarily at the end of the optimal time interval [0,tf]. It is however evident that more than one trajectory corresponding to different values of the uncertainty (D A ,D B ), can hit the boundary at the same time. Lemma 7.7: Under Assumption 7.1, each z j (s) 0 almost everywhere on (0,1) and the set of zeros of z(s) do not have a limit point in (0,1). Proof: Consider the expression for z(s): z(s) = s1 yT(,A,B')e*A(s) B(dA,dB/) (d). Recall that the support for was defined as = {(A,B,)(A+ A ) (B+ B )[0,1]: y T y = M}. If Lemma 7.6 holds then = { (A,B,)(A+ A ) (B+ B ){1}: y T y = M c } and z(s) can be simplified as follows: z(s) = yT(,A,B')e*A(s) B(dA,dB). Hence for each s the expression for z(s) is given by the following power series expansion: z(s) = yT(1,A,B')e*A B(dA,dB) + s yT(1,A,B')e*A AB(dA,dB) + s22! yT(1,A,B')e*A A2B(dA,dB) + yT(1,A,B')e*A (dA,dB) B + s yT(1,A,B')e*A (dA,dB) AB + s22! yT(1,A,B')e*A (dA,dB) A2B + 62 PAGE 63 The above equality is true since the diameter of A and B are small compared to the variation in e*A as A varies over A Let us denote the coefficients by (X 0 X 1 X 2 .). Hence: X k = yT(1,A,B')e*A (dA,dB) AkB = [ f 1 f 2 f n ]A k B (Where we denote the row matrix created by the integral as [ f 1 f 2 f n ] and n is the order of the system. We stack that coefficients together and expand each X k (7.8) [X 0 X 1 .X k .] = [f 1 f 2 .f n ][B AB .. A k B..] or [X 01 .X 0m  X 11   X k1 .] = [f 1 f 2 .f n ][B AB .. A m B..] Now recall Lemma 7.2, which guaranteed that at least one component of z j (s) (j=1,2,,m) is nonzero over at least a subinterval of nonzero measure in [0,1]. Without loss of generality that this is the case with the z 1 (s). Let us write z 1 (s) in terms of the coefficients defined above. z 1 (s) = X 01 + X 11 s + X 21 s 2 + Hence z 1 (s) is analytic and in particular a power series in s over [0,1]. It is well known that a power series is either identically zero or is nonzero almost everywhere in its domain of definition [0,1] (Conway 1978). Since by Lemma 7.2, z 1 (s) is nonzero at least over some interval of nonzero measure in [0,1], it follows that it is nonzero almost everywhere over [0,1]. We now write the coefficients of z 1 (s) in terms of Equation 7.8. [X 01 X 11 X n1 ] = [f 1 f 2 .f n ][C 1 ] Where C 1 is the controllability matrix for the first input. Since z 1 (s) is nonzero, at least one of the coefficients X i1 (i=0,1,.) is nonzero. Hence [f 1 f n ] 0, i.e., at least one among f i (i=1,,n) is nonzero. Recall that by assumption (7.1), the controllability matrices C j 63 PAGE 64 corresponding to each of the m inputs are full rank. Hence at least one among the first n coefficients of each z j (s) (j=1,2,m) is nonzero. This in turn implies that each z j (s) is nonzero almost everywhere on (0,1). This proves that z(s) is a nonzero power series on (0,1). It is well known that zeros of analytic functions cannot include a limit point in its domain of definition. It follows that the only possible limit points in the zeros of z(s) is at s=0,1. Hence proved. From Theorem 7.2 and the above lemma, we conclude that z(s) satisfies the following: z(s)v(s) z(s)v (s) for a.a. s[0,1]. Moreover each component of z(s)0 a.e. on (0,1). Together this implies that for a.a. s[0,1]: v *j (s) = K when zj(s)0 +K when zj(s)0 Hence, the optimal solution is purely bangbang almost everywhere on (0,1). The value of v (s) is not known over sets of measure zero. These sets are exactly the values of s where some component of the function z(s) = 0. Recall that each zero crossing of any component of the switching function zj(s) (j=1,2m) can potentially imply a switch between the extreme allowed values of the input (K). The number of such points is not finite in general and hence the solution may contain an infinite number of switches or sign changes between the extreme values. However, by Lemma 7.7, the zeroes of z(s), or in other words the switching instants of the optimal input v do not have a limit point in (0,1). The only possible limit points are at s = 0 or 1. This implies that the optimal switching sequence may have become increasingly close together at the very beginning or the end of the optimal open loop interval. However, the open loop interval is always finite. From an engineering viewpoint, the optimal input function may be truncated near the two ends. This will result in a finite number of switches in the optimal input 64 PAGE 65 solution with an arbitrarily small reduction in the maximal open loop time. This will make the computation of the switching times as well as the hardware implementation feasible. Hence the optimal v* may be approximated for every practical purpose, by an implementable function with a finite number of switches. The engineering importance of the numerical simplification in the calculation of the optimal solution cannot be over emphasized. The optimal solution, being bangbang, is effectively known by the computation of the switching instants. Moreover, a bangbang solution may be implemented with the most basic in hardware and simplest of algorithms. By a simple example, we demonstrate some of the interesting features of the problem solved above. We choose the same single dimensional system with a single input and with some uncertainty on the pole as was used in Problem 6.13. The state is the output of the system. Let us transcribe Problem 4.4 in terms of this system: Problem 7.9: Find uUmax t f subject to the following constraints: .x(t) = ax(t) + u(t) 0 t tf and u(t) 2 x(0) = 1 xT(t)x(t) M for 0 t tf and for all a[1.21.4] We provide a solution to the problem for M = 25. The optimal input in this case has only one switch with the optimal switching instant at 1.3467 seconds. The maximum tf was found to be 5.08 seconds, and in the time interval [0, 5.08 sec], every state trajectory is guaranteed to be within [5,5]. For a clear understanding of the behavior of the system for different values of the uncertainty we have plotted the corresponding state trajectories for ten different uncertainty values of the pole a. 65 PAGE 66 Figure 71: Optimal input has one switch: M = 25, tf = 5.08 Figure 72: Trajectories for ten different uncertainty values of a; M = 25, t f = 5.08 66 PAGE 67 CHAPTER 8 CONCLUSION AND FUTURE WORK Conclusion The problem of maximizing the open loop period of operation of a linear time invariant system with bounded inputs has been considered. The parameters of the controlled system are subject to bounded uncertainties. An optimal controller has been derived that maximizes the time during which the control loop can be left open, while keeping the system error within prespecified bounds. The existence of such a control signal is first proved among the set of measurable functions. EulerLagrange type first order necessary conditions are then derived for calculating the optimal open loop input. It is shown that the time interval, during which the control loop can remain open, is maximized by an input, which may not be purely bangbang over the entire maximal open loop time. We further showed that in cases where the optimal input was not bangbang over certain intervals, a purely bangbang input existed which approximated the optimal input. This is of engineering importance, since the bangbang nature of the optimal solution makes the computation of switching instants computationally feasible, as opposed to solving the entire dynamic optimization problem numerically. In the general case, the possibility of a high number of switches in the approximate input solution cannot be excluded. However, under the assumption that the input bound is small compared with the allowed error bound on the system, we have shown that the optimal input itself is purely bangbang. Moreover, the sequence of optimal switching instants does not have a limit point in the interior of the maximal open loop interval. This facilitates the computation of the optimal open loop input. Because of the finiteness of the open loop interval of operation, the optimal input can then be approximated by a piecewise constant input with a finite number of switches. 67 PAGE 68 Future Work The results we have obtained have potential impacts on a number of application areas, including the following. Each of these applications is a potential candidate for future research that is based on the theory developed in this dissertation. Applications Control of Space Vehicles: Space vehicles are frequently faced with intermittent loss of signal due to obstacles in the line of vision, radiation interference, limitations in power, etc. In such situations, it is critical to know how long the system can perform within its specifications without communication with its supervisory center. This question can be directly addressed within the framework of needbased feedback we have developed. Telemetry: Applications in telemetry frequently face the problem of limited power. The method we propose effectively minimizes the communication needed to control a system to a minimum and hence maximizes the longevity of the associated power sources. Stealth Systems: Stealth and spy systems, like unmanned aerial vehicles and similar devices, prefer to reduce communications with their control center so as to reduce the chances of detection. The needbased feedback framework can derive the necessary control feeds to achieve this goal and reduce the probability of detection. Agricultural Systems: Modern agricultural applications, such as soil moisture content control and fertilization control, require complex and expensive feedback processes, as most feedback data has to be collected by human experts testing in the field. Using needbased feedback reduces the frequency and the duration of these field tests, thus reducing costs and improving efficiency. Biotechnology: Biological research often involves culture and preservation of cells and other organic substances under very carefully controlled environments. Regular human 68 PAGE 69 surveillance is necessary and extremely expensive. Our method of reducing feedback can effectively minimize requirements of human surveillance and thus reduce involved costs. Networked Control Systems: It is increasingly common to use shared networks to control geographically separated devices. The inherent unreliability of networks may cause frequent loss of feedback signals for uncertain periods. To guarantee control objectives, it is crucial to have an estimate of how long the system can perform in open loop. Again, this can be answered in the framework of needbased feedback control. Medicine: Applications are also possible in medicine and optimization of drug dosage. Typically patients are treated with drugs at regular intervals while the feedback in terms of its effectiveness is collected after large intervals. The method proposed can be potentially used to guarantee the effectiveness of the drug when measurements are not being made. Moreover the intervals after which the patient needs to be checked may be maximized. Numerical Optimization: The actual computation of the optimal solution presents an interesting problem in numerical optimization. We conjecture that appropriate use of combinatorics may lead to a highly efficient algorithm for calculating the solution. Theoretical Research This problem is closely related to the viscosity solutions of the HamiltonJacobiBellman equation. It seems that the socalled exit time problem studied widely in relation to minmax dynamic programming is similar to the maximal open loop time studied above. But the inherent nonsmoothness of the current solution makes the use of nonstandard solutions necessary and hence may present an interesting future direction of research. The theory proposed extends the results on optimal residence time of perturbed control systems. While previous results have been derived in a feedback control setting, we have shown 69 PAGE 70 that investigation in the framework of robust openloop optimization framework is interesting and relevant to physical situations. A complete theory requires further research in this direction. Questions about reachability of uncertain dynamical systems are closely intertwined with the problem investigated above. The concept of reachability of uncertain systems clearly depends on whether the system is in open or closed loop. However the correct characterization still needs to be investigated. The problem of stabilizing an unstable system with prespecified bounds on the inputs is related. Though this is extremely common in most controller design, a satisfactory theory is not yet available. It is thus of practical interest to calculate the minimum input bound required to stabilize an unstable uncertain system. Further theoretical applications of the proposed theory may be found in the areas of stochastic dynamical systems, Lyapunov stability analysis of switched dynamical systems, and convex and nonconvex minmax dynamical optimization. 70 PAGE 71 LIST OF REFERENCES A.V. Balakrishnan, Introduction to Optimization Theory in Hilbert Space, SpringerVerlag, Heidelberg, 1971. T. Basar and G.J. Olsder, Dynamic Noncooperative Game Theory, London/New York: Academic Press, 1982. P.R. Bellanger, On the sensitivity of solutions to some optimal control problems, Ph.D. dissertation, Massachusetts Institute of Technology, Cambridge, 1964. R. Bellman, Dynamic Programming, Princeton University Press, Princeton, 1957. L.D. Berkovitz, A differential game with no pure strategy solution, Advances in Game Theory, pp. 175194, 1964, Princeton Univ. Press, Princeton, N.J. L.D. Berkovitz, Necessary conditions for optimal strategies in a class of differential games and control problems, SIAM J. Control, 5, 1967, pp. 124. L.D. Berkovitz, A variational approach to differential games, Advances in Game Theory, Annals of Math. Study 52, Princeton University Press, Princeton, 1964, pp. 127174. A.E. Bryson, Jr. and Y.C. Ho, Applied Optimal Control, Blaisdell, 1969. R.M. Cannon, Jr. and E. Schmitz, Initial experiments on the endpoint control of a flexible onelink robot, Int. J. Robotics Res., Vol. 3, pp. 6275, Fall 1984. J.B. Conway, Functions of One Complex Variable, Springer, 1978. R.J. Elliot and N. Kalton, The existence of value in differential games Mem. Amer. Math. Soc. 126, 1972. R.J. Elliot, N. Kalton and L. Markus, Saddle points for linear differential games SIAM J. Control Optimization, Vol. 11, pp. 100112, 1973. B.P. Eng, Microprocessor based servo control for gun fire control, in Proc. IEEE IECI Annual Conf., Philadelphia, PA, Mar. 1979, pp. 329333. A.A. Feldbaum, "Dual control theory, parts I and II", Automation and Remote Control, Vol. 21, No. 9, April 1961, pp. 874880 and Vol. 21, No. 11, May 1961, pp. 10331039. (Russian originals dated September 1960, 12401249 and November 1960, pp. 14531464) A.A. Feldbaum, Optimal Control Systems, New York: Academic Press, 1965. A.F. Filippov, Differential equations with discontinuous righthand side, Amer. Math. Soc. Transl., Ser 2, 42, pp. 199231, Providence, R.I., 1964. W.H. Fleming, The convergence problem for differential games, Journal of Mathematical Analysis and Applications, 3, pp. 102116, 1961. 71 PAGE 72 W.H. Fleming, The convergence problem for differential games, Advances in Game Theory, pp. 195210, 1964, Princeton Univ. Press, Princeton, N.J. A. Friedman, Differential Games, John Wiley, 1971. R.V. Gamkrelidze, On some extremal problems in the theory of differential equations with applications to the theory of optimal control, J. SIAM, Ser. A Control 3, 1965, pp. 106128. P.R. Halmos, A Hilbert Space Problem Book, SpringerVerlag, New York, 1982. N. Halyo, Terminal area automatic navigation, guidance, and control research using the microwave landing system. Part 5Design and development of digital automated landing system for steep final approach using modern control techniques, NASA Contract Rep. 3681, Apr. 1983. M.W. Hirsch and S. Smale, Differential Equations, Dynamical Systems, and Linear Algebra, Academic Press, 1974. Y.C. Ho, A.E. Bryson, Jr., and S. Baron, Differential games and optimal pursuitevasion strategies," IEEE Trans. Automatic Control, vol. AC10, pp. 385389, October 1965. D.R. Howard and Z.V. Rekasius, Error analysis with the maximum principle, IEEE Trans. Automatic Control, vol. AC9, pp. 223229, July 1964. R. Isaacs, Differential Games, New York: Wiley, 1965. R. Isaacs, Differential games III, Research Memorandum RM1411, The RAND Corporation, Santa Monica, California, 1954. D.L. Kelendzheridze, On the theory of optimal pursuit, Soviet Math. Dokl., 2, 1961, pp. 654656. A.J. Koivuniemi, Parameter optimization in systems subject to worst (bounded) disturbance, IEEE Trans. Automatic Control, vol. AC11, pp. 427433, July 1966. A.N. Kolmogorov and S.V. Fomin, Elements of the Theory of Functions and Functional Analysis, Vol 1 and 2, Graylock Press, Rochester, New York, 1957 and 1961. N.N. Krasovskii and A.I. Subbotin, Game Theoretic Control Problems, SpringerVerlag, Berlin, 1974. Y.S. Ledyaev and E.F. Mishchenko, Extrema1 problems in the theory of differential games, Proceedings Steklov Inst. Math., 185, pp. 165191, 1990. Y.S. Ledyaev and E.F. Mishchenko, Optimization conditions for strategies in differential games of fixed duration, Differential Equations, 23, 1988, pp. 175184 72 PAGE 73 Y.S. Ledyaev, Optimality conditions for feedback controls in problems of control under disturbance, Proceedings of the 33 rd Conference on Decision and Control, Lake Buena Vista, FL, December 1994. Y.S. Ledyaev, The maximum principle for optimal strategies in differential games, Soviet Math. Dokl., 40, pp. 545549, 1990. Y.S. Ledyaev, On optimal strategies in differential games of fixed duration, Soviet Math. Dokl., 33, pp. 7880, 1986. L.A. Liusternik and V.J. Sobolev, Elements of Functional Analysis, Frederick Ungar, New York, 1961. D.G. Luenberger, Optimization by Vector Space Methods, Wiley, New York, 1969. S.M. Meerkov and T. Runolfsson, Residence Time Control, IEEE Transactions On Automatic Control, Vol. 33, No. 4, April 1988. E.F. Mishchenko and L. S. Pontryagin, "Linear differential games," Soviet Math. Dokl., 174,. No. 1, 1967, pp. 654656. E.F. Mishchenko, Pursuit and evasion problems in differential games, Eng. Cybernetics, 9 (1971), no. 5, 787791 L.A. Montestruque and P.J. Antsaklis, Stability of modelbased networked control systems with timevarying transmission times, IEEE Transactions on Automatic Control, Special Issue on Networked Control Systems, Vol. 49, No. 9, pp.15621572, September 2004. G.N. Nair, F. Fagnani, S. Zampieri and R. J. Evans, Feedback Control under Data Rate Constraints: an Overview, to appear in Proceedings of the IEEE (special issue on The Emerging Technology of Networked Control Systems), 2007. L.W. Neustadt, An abstract variational theory with applications to a broad class of optimization problems I, general theory, SIAM. J. Control, 4, 1966, pp. 505527. L.W. Neustadt, An abstract variational theory with applications to a broad class of optimization problems II, applications, SIAM. J. Control, 5, 1967, pp. 90137. J.C. Panetta and K.R. Fister, Optimal control applied to competing chemotherapeutic cellkill strategies, SIAM Journal on Applied Mathematics, Vol. 63, No. 6, 2003. L.S. Pontryagin, V.G. Boltyansky, R.V. Gamkrelidze and E.F. Mishchenko, The Mathematical Theory of Optimal Processes, Interscience Publishers John Wiley & Sons Inc., New YorkLondon, 1962 L.S. Pontryagin, "On the theory of differential games", Russ. Math. Surv., 1966, 21 (4),pp. 193246. 73 PAGE 74 L. Pontryagin, Linear differential games. I, II, Soviet Math. Dokl., 175, 1967 pp. 764766 W. Rudin, Real and Complex Analysis, McGrawHill, New York, 1966. R.E. Skelton, Precise pointing control of space vehicles with momentum exchange controllers, Sperry Rand Rep. SP250754, 1973. R.B. Vinter, Minimax optimal control, SIAM Journal of Control and Optimization, Vol. 44, No 3, pp 939968, 2005. R.B. Vinter, Optimal Control, Birkhauser, Boston, 2000. G. Walsh, H. Ye, and L. Bushnell, Stability analysis of networked control systems, Proceedings of American Control Conference, June 1999. J. Warga, Minimax problems and unilateral curves in the calculus of variations, J. SIAM, Ser. A, Control 3, pp. 91105, 1965. J. Warga, Optimal Control of Differential and Functional Equations, Academic Press, New York, 1972. J. Warga, On a class of minimax control problems in the calculus of variations, Michigan Math. Journal, 12, 1965, pp. 289311. J. Warga, Unilateral and minimax control problems defined by integral equations, SIAM J. Control, 8, 1970, pp. 372382. J. Warga, On a class of pursuit and evasion problems, J. Differential Equations, 9, 1971, pp. 155167. J. Warga, Conflicting and minimax controls, J. Math. Anal. Appl., 33, 1971, pp. 655673. S. Willard, General Topology, AddisonWesley, Reading, MA, 1970. H.S. Witsenhausen, A minimax control linear problem for sampled systems, IEEE Transactions On Automatic Control, Vol. AC13, No. 1, February 1968. L.C. Young, Generalized curves and the existence of an attained absolute minimum in the calculus of variations, C. R. Sci. Lettres Varsovie, C III 30 (1937), 212234. L.C. Young, Lectures on the Calculus of Variations and Optimal Control Theory, W. B. Saunders, Philadelphia, 1969. E. Zeidler, Nonlinear Functional Analysis and its Applications III SpringerVerlag, New York, 1985. P.V. Zhivogyladov and R.H. Middleton, Networked control design for linear systems, Automatica, 39, pp. 743750, 2003. 74 PAGE 75 BIOGRAPHICAL SKETCH Debraj Chakraborty was born in Calcutta, India. He earned his Bachelor of Engineering degree from Jadavpur University, Calcutta, in 2001. After completing his master's degree in electrical e ngineering from the Department of Electrical Engineering, Indian Institute of Technology, Kanpur, in 2003, he joined the Department of Electrical and Computer Engineering of the University of Florida, Gainesville, USA. He completed his Ph.D. in electrical and computer engineering in 2007, under the guidance of Dr. Jacob Hamme r at the University of Florida. From 2003 to 2007, he was appointed an Alumni Fellow at the University of Florida. 75 xml version 1.0 encoding UTF8 REPORT xmlns http:www.fcla.edudlsmddaitss xmlns:xsi http:www.w3.org2001XMLSchemainstance xsi:schemaLocation http:www.fcla.edudlsmddaitssdaitssReport.xsd INGEST IEID E20101113_AAAAHQ INGEST_TIME 20101113T09:38:46Z PACKAGE UFE0021236_00001 AGREEMENT_INFO ACCOUNT UF PROJECT UFDC FILES FILE SIZE 1053954 DFID F20101113_AABCMF ORIGIN DEPOSITOR PATH chakraborty_d_Page_33.tif GLOBAL false PRESERVATION BIT MESSAGE_DIGEST ALGORITHM MD5 73a327d40d6bbd6b11e78ddae05b685e SHA1 913c36779d5b2a89879c6f859d562631fc40f6a4 48988 F20101113_AABCHI chakraborty_d_Page_45.jpg 593fc5554197ee88654c7252d75ec8ed 6dbc27122dd7121637a4a95b0417310961ab3e4f 1798 F20101113_AABCRC chakraborty_d_Page_21.txt 0cac342b2a3092df547dfe6bfc4c52f7 6d6fd66742230e20cf5834c59e80000c3c2506d3 1303 F20101113_AABCWA chakraborty_d_Page_03thm.jpg 6ca12ed3ba5b1f323729776d709b90ad 19dd99e9a4f10624b996e05eb55c75b7935e6103 F20101113_AABCMG chakraborty_d_Page_34.tif 152e2ef95e511541102e98da588f956a b6d422d03b41a8a31efb477b4e6fd7f334f8d047 50650 F20101113_AABCHJ chakraborty_d_Page_46.jpg ddd5a94264b75981d4999a8dd2f4c079 577f36152d310bc66160b4ae136c687206693ce0 1654 F20101113_AABCRD chakraborty_d_Page_22.txt 37d0e0bb30e4ff7d3e5b56f7792dc0ec 0207bb378480f4e0c14456ce3fe9df87d01f96ec 12300 F20101113_AABCWB chakraborty_d_Page_04.QC.jpg 8d2a83f9779fa64220f9d8d127c2b1d6 62eca3bead90464674322ec7b21b461a4433598e F20101113_AABCMH chakraborty_d_Page_35.tif 06bf4036abd1605906f26ae262870131 e128223acf5e9d872bbae9fae558f9cca2b08e86 48710 F20101113_AABCHK chakraborty_d_Page_48.jpg a1289feff4345dce822de9fc0a8e0f27 2f9e190ebca5776ca3691b6a6854bca3c2972851 1732 F20101113_AABCRE chakraborty_d_Page_24.txt 735351c0d16053a7eb7e52e53c892ef0 ff98068331bc0a28bd86cc3f22e610877fe6a1a1 2896 F20101113_AABCWC chakraborty_d_Page_06thm.jpg ce1710584f4d02bf68d0e877bb92dea8 b6735479023183ab80de5abacfc0c20e49b1e09c F20101113_AABCMI chakraborty_d_Page_36.tif c619941458befb199ec9d9c62b326e72 51c3e7e2f6e059d630f127e11ad23c0a5f6ff17f 43112 F20101113_AABCHL chakraborty_d_Page_49.jpg f4b54db72ca1e84d10fb2683a543b684 3ae13211c1a2438706315dce0e014abad601ce56 687 F20101113_AABCRF chakraborty_d_Page_25.txt f28b3e592c33200e2e228aefc9c66771 044d7ecf5a1857a3e22b607f351607b033a87968 10752 F20101113_AABCWD chakraborty_d_Page_07.QC.jpg 254f6923aaf0ffa38e9e21e48ff3e87c 457952ee39d3b5df45007870c8e0d4a08634d633 F20101113_AABCMJ chakraborty_d_Page_37.tif 2b34f0a0f5339b7c0a2449b6e3ebc175 459942c90747da2fe1dab2f85177ee805c39d67b 1863 F20101113_AABCRG chakraborty_d_Page_27.txt e914c9bdb7570f2f6a4827730efe5425 de2d12980b49cee23b711f477e69229f93e3c039 3072 F20101113_AABCWE chakraborty_d_Page_07thm.jpg f65aa363913b10ba76c9652db3d171c1 6ca0afa36b5a8226683067cb7de0cb6f251e4479 F20101113_AABCMK chakraborty_d_Page_38.tif 3add10a31a6b68534e08b9901d45b887 bf62677778c4a088f499cbe5745cefaf081e88a6 56396 F20101113_AABCHM chakraborty_d_Page_50.jpg 1947c5789c576262ed80bb64328caf0d 381e98123e2388fb24c3f490db7b95f5e888f568 1912 F20101113_AABCRH chakraborty_d_Page_28.txt 74851b117df2af3bf7595e893abbef49 045da60c3de3504144913564dea88018737b068d 24171 F20101113_AABCWF chakraborty_d_Page_10.QC.jpg 44c965babfe2d08fa12ad2a7e08fceeb 0beea708b3447369d5076a2bbd99a8155453c90e F20101113_AABCML chakraborty_d_Page_39.tif dc60104071e2e7f9e5ea3238029d575b 272d5c67c7e2cfab4cbc0f6852f171a4080c3694 57697 F20101113_AABCHN chakraborty_d_Page_51.jpg b56bf265ad9fe5281fbea944f872f61a c1661f016287ec39da42c5bf52dbc085f5119250 773 F20101113_AABCRI chakraborty_d_Page_29.txt f063d3260d36bd49d2f3a60f1f0b7af3 ffe88dd6b5a383ad944bc9409ef74682d23617d8 22137 F20101113_AABCWG chakraborty_d_Page_12.QC.jpg d6da916941da93d47880af92cecdfafc 03785e4cf24562b9ca7d18483bca9d00f4810f69 F20101113_AABCMM chakraborty_d_Page_40.tif a40e0b9a3fa3989fe9fb9a9f791dbba4 a16c23ce6385eefa2c3aa36f2b7529476ca77ec7 73616 F20101113_AABCHO chakraborty_d_Page_52.jpg 5c1cb443a3cb1c91659e5b60f85dccd6 d4ead899c52cd5cca58b15ca402475bb67f23c79 2058 F20101113_AABCRJ chakraborty_d_Page_30.txt ae95822520719518eff23e10d192415e 37729cef8964d6492e06f44b4f5477d0371ddeb4 6260 F20101113_AABCWH chakraborty_d_Page_12thm.jpg 31baffe33ddfa71181fb3f0ca3728806 900bb90054c857cd08994c78e8af4f53f75d8723 F20101113_AABCMN chakraborty_d_Page_41.tif 8be6109414f1348193643af1c3550142 f93ce121cb6fec835f63375ef59bd11a95497da9 65943 F20101113_AABCHP chakraborty_d_Page_53.jpg 4e827657e7725cb2bf0b7dcc3326b735 9afa499a1dbe62ecd9c42853c6e6001f0984d999 1825 F20101113_AABCRK chakraborty_d_Page_31.txt d9e4a7891d72db0e268f85b6a9d4bd69 f4859a39070632594cf509c1b7800f33aedeb02c 2388 F20101113_AABCWI chakraborty_d_Page_14thm.jpg d869a6e77efbd5408b61c08b30897df7 bfe6aa221ce9afb5a700e3d71d2d4ace68c34b0f F20101113_AABCMO chakraborty_d_Page_42.tif 1ab58596aca7553cbf2b891ad69b78e2 d99c2210546d419252a94fa5b6b1627bc8e78b44 52603 F20101113_AABCHQ chakraborty_d_Page_54.jpg 2825d7fc22f571c9b26fb1c7f2d8f050 7971c51467f02ff582ae04f2f9d5feb73b69074d 1622 F20101113_AABCRL chakraborty_d_Page_32.txt 00eb8b1814c9826a594f90214473bd2d 070db3a18081e9d1172adb7f2d5abaaaeb64664d 25930 F20101113_AABCWJ chakraborty_d_Page_17.QC.jpg 6aeb787978d21cbb8ad69a36364eeac8 0915be696c3fc1e68b9d75d0471fd7f641e97bf5 F20101113_AABCMP chakraborty_d_Page_43.tif b908dfdceca33f0272ff5c68ced5869f 9fd1103782940463a70bf936d4f1165e4de2a388 73551 F20101113_AABCHR chakraborty_d_Page_55.jpg aa8e035d86c460682b439310e8ec27dd d1a31831e6fba47ba98a17a1b3cc3f02bc1f2bb2 1053 F20101113_AABCRM chakraborty_d_Page_33.txt 9df2242a24102540565f82be03ff3494 2b5b9f1fd6f1f3be9293a8163af74494a979fed0 11003 F20101113_AABCWK chakraborty_d_Page_18.QC.jpg 1facb0658b424f79b16310c5d3801785 dc41212181b2a23db4461eae336575bd9162242a F20101113_AABCMQ chakraborty_d_Page_44.tif 9a775aa4fef10e23304bd20ae789c681 c29dd042c0f2f84e85843586ab193c06c047b514 48911 F20101113_AABCHS chakraborty_d_Page_56.jpg 313238f4ff8aabb2018dd4ed48c82e26 53612b6185e70cfca9dd42e27970ad529cbfb553 1191 F20101113_AABCRN chakraborty_d_Page_35.txt 8687c929438786f67ab5b380980e0ad3 c2b55adda1a87424f07373b3aa3c1ed941af600c 19441 F20101113_AABCWL chakraborty_d_Page_19.QC.jpg 5570508004e33e90248146f3f2e5fb16 d988ecaaae823deeb9505a1c20b29810d2831833 64145 F20101113_AABCHT chakraborty_d_Page_57.jpg 1283a9589b95bae009e0e1e29fee6b53 1d617f833de336cf36f7cb046cfa11572e0328ce 1513 F20101113_AABCRO chakraborty_d_Page_36.txt 8e0f12277df32691201ba65da068ca86 17a79d8e719fc4323fe5344706616211637c6fd3 5965 F20101113_AABCWM chakraborty_d_Page_19thm.jpg e7700003b207a3d4546ecc7e32e57198 a3514b53f7204dde07049433ea51a6492dfcf2da 52047 F20101113_AABCHU chakraborty_d_Page_58.jpg b53e9f8588909e79e68011ef2bcbb8c7 a2e924c93eb2761d7401b1138b14f1a8df512178 1722 F20101113_AABCRP chakraborty_d_Page_37.txt e73f6047f2c581bc804a93df6083a0c4 aa77e20e3607610b9adc7750cd39726e6dfa24da F20101113_AABCMR chakraborty_d_Page_45.tif 5e715eb02d9f7b67b951899e5793ac4c c6c70446441c5319af706d0cce54b6608c83d3da 22170 F20101113_AABCWN chakraborty_d_Page_21.QC.jpg f89ed6e2df1d5d075ad5c44d88480c25 108d74907c8cdd9178d1e617e91f0197cd653f4f 49886 F20101113_AABCHV chakraborty_d_Page_59.jpg 7a607d25330628f05cfeda45b73aba55 dd47016023f9f475fda73d90fede6c4005e0ff88 1607 F20101113_AABCRQ chakraborty_d_Page_38.txt 7d50f5f2005a58bc28e50c9837dae61c 66c2ad9311db2aadcda06f28fec3dab7d128d397 F20101113_AABCMS chakraborty_d_Page_46.tif b02680f824912d59cb7f855b4f315a9e 1ef6901418d3c8eb7c402175b98689694b908ef0 17422 F20101113_AABCWO chakraborty_d_Page_23.QC.jpg 01431ab1d7a5796bb42871dbe8cc12d2 af60a70f3fba4b9dde9869b1c8907c68c68529a9 69026 F20101113_AABCHW chakraborty_d_Page_60.jpg 50f45cb1d28e7d87079f36fa2083a416 210bc24e33ffa80e0a453c6b500dbd9568dbdf41 1980 F20101113_AABCRR chakraborty_d_Page_39.txt d3beb0488f02e5b2ede48e916b0a96d6 0e9deaa589c58d22321568eee1726067384cc786 F20101113_AABCMT chakraborty_d_Page_47.tif 3258fddbefb282df1991a405a6ea80af 07b52e8c90b8ee3d1cbf0441d0c5423777f31434 19540 F20101113_AABCWP chakraborty_d_Page_24.QC.jpg b924cc6cc67f1371a97510b65c1c7265 2d3507bd350dec5a007282562c51a2dbd9a62591 49521 F20101113_AABCHX chakraborty_d_Page_61.jpg f5657735574da34f14ab15f3cff26ab2 e2a312a96b69bdf279bc42eb44e4be1eff38f8fa 607 F20101113_AABCRS chakraborty_d_Page_40.txt b4f5afa584079909055f06cab4e4a58e b845ed2e2a4d1e33515db0c21604f190c2180548 F20101113_AABCMU chakraborty_d_Page_48.tif a5468c9d2ea2d347040e9368a91e41b1 7124f7db7306c4a029d617b3114e0829993f0caa 22210 F20101113_AABCWQ chakraborty_d_Page_26.QC.jpg 57eb6ded59ea8930f131b1cf21ee3442 6f857d58e231b042436eebe956758db623066ad3 7102 F20101113_AABCFA chakraborty_d_Page_01.pro 4705ba8f23dd3c93d839fb6ac536525b 563466b0052c845ccdf23c36b2f40d1eb278d319 50136 F20101113_AABCHY chakraborty_d_Page_62.jpg 68eaf8df317468dc06141720bc166294 0a8b50e4f42dcd400f443ae22ebed8fbf5c38071 1922 F20101113_AABCRT chakraborty_d_Page_41.txt 100700daa1aebd907d2f7a99f77591dd 34ae2e60c8369e1949c4be00d5acf5f40aefc1e2 F20101113_AABCMV chakraborty_d_Page_49.tif 3fbe5d592697f81578700fb301eeb893 d4c43ee48daec06ffe21a37a0dd7ae9738f08801 6040 F20101113_AABCWR chakraborty_d_Page_26thm.jpg ca19e97fc8d2870ce165157beebddd49 11ec5d4a543a898c73b31084778a5016957c0510 75370 F20101113_AABCFB chakraborty_d_Page_15.jpg 8a6d0edeb497aa1b6ceffa2545434f02 5e764165d39c48a51c2d93347f24ace70c45a632 61101 F20101113_AABCHZ chakraborty_d_Page_63.jpg 99b897da63bc133d0e2dc3c609d551b9 2a4dc36442a5683d643f230e884b15707ea529aa 1477 F20101113_AABCRU chakraborty_d_Page_42.txt bd1d08c98cb7a6fc6740c3667ae8e3fd b1b378a80105aee4a8ab046e8a1f75c8761e65dd F20101113_AABCMW chakraborty_d_Page_50.tif 14ff2c2939664f28ef79ed67d5123b16 f141b1a1c25d5227f9e96e3e2c02e530472bbd4d 6551 F20101113_AABCWS chakraborty_d_Page_27thm.jpg 6e3efce22cfd553c27c8c5295f1b0989 9e2ba47932d27ed594a528af324d08f0a3a2df98 19642 F20101113_AABCFC chakraborty_d_Page_43.QC.jpg 930daf4e51d3ae29554baa750025fc37 60db1b986fd0e7881408926d1d1ec19b0a4f1ed6 1517 F20101113_AABCRV chakraborty_d_Page_43.txt e2e01ca8a63cc827174c6bd47f4ab973 7440f50056cff888b8648b475d935ca6b0189a6f 25271604 F20101113_AABCMX chakraborty_d_Page_51.tif e13d2e1d54f95d5af9c144da1c96dfb6 a54812f120bac0acb078ec8487bcb234a6f96b62 22500 F20101113_AABCWT chakraborty_d_Page_28.QC.jpg 1a632fdf4943a8fec7f37038f74f5963 f4bd15af0f6aa322ee0f76bd0efdc8f69f56c8d6 114488 F20101113_AABCFD chakraborty_d_Page_11.jp2 af785ff03f2ca377835fcff7ae0d0f75 c47de21818e59a910f46318255a178375474d197 73199 F20101113_AABCKA chakraborty_d_Page_45.jp2 e935a1a3033552c3e9c9a9a9f4999c4e 03477d3082471a56523a49921e032ba12cfd7582 F20101113_AABCMY chakraborty_d_Page_53.tif 73ca420ce0b62cd4030a2650af88cdd7 46b9d4e71a63511c0b0056396e4af77366c3e04c 10840 F20101113_AABCWU chakraborty_d_Page_29.QC.jpg 3770cc0a9ccd7136efe43c7f887ceb03 92c9428c88302460ee0bebdac0da8d19f9dcf6b5 1368 F20101113_AABCRW chakraborty_d_Page_44.txt 2475f1f5548932170b111762e60ba806 64c98cf6c420d76aa08d7d710096896f52b82354 73168 F20101113_AABCKB chakraborty_d_Page_46.jp2 6a0b5e7b878625591c9513a1a4139a55 43e0d9d99a51534bff8ceec793be8bae67ebd2c3 F20101113_AABCMZ chakraborty_d_Page_54.tif cdc6cc92c8f4c1cb440de3ba7168eb90 a0b093a5929a91039430a0cde08b11cb8463802b 21579 F20101113_AABCFE chakraborty_d_Page_37.QC.jpg 1937f369fd1daf36b93fa89cb8f54e47 6f5d4a86b8e600520a5f4603ee4e49f838420828 3227 F20101113_AABCWV chakraborty_d_Page_29thm.jpg f1750ed1e8a665ba217b4672e7e8979a ac0ce8dee363826395107ac63fa08ea6c10cf77d 1420 F20101113_AABCRX chakraborty_d_Page_45.txt fc08c71805e487daa0ce34004e277afb 1e77647e5223af26b145436587faeb2506a6dd20 95701 F20101113_AABCKC chakraborty_d_Page_47.jp2 61b479a4903ca9c891d75eeaaedc4bba 2df093f29c2903f3b25acf5f5ba09a317f233c11 1051980 F20101113_AABCFF chakraborty_d_Page_05.jp2 0b1216b0da20331639b90ee91237b766 c78547313c1fbb21af09a914d29a47b220b95855 6729 F20101113_AABCWW chakraborty_d_Page_30thm.jpg 8813536eee3934c65d612e60c5087795 36a3e5f02859e5a754e47b08b18bae96b5d14e38 39578 F20101113_AABCPA chakraborty_d_Page_34.pro fbd7ce6fcadf636744832d47cce66498 c35772df12e2173c4431088c5f84ad0febea5885 1387 F20101113_AABCRY chakraborty_d_Page_46.txt ac9d10b6d37ac6f0a6bd14525878f828 aa33833587fb37ffd1481898e3fb6c207d7a111d 71207 F20101113_AABCKD chakraborty_d_Page_48.jp2 21f4cc42a4ebcf28b4b62b5f812b9d7a 2b66f3096bb2fc4ef2a058a7b0615e644a5605a1 36122 F20101113_AABCFG chakraborty_d_Page_42.pro f45ae5994b3bdd2abd04f28b881fd459 8eec5da778297d66da145f1725417baf335117fd 19952 F20101113_AABCWX chakraborty_d_Page_32.QC.jpg 334a623953d81c4ee805668bb2b2fadf 936a71d3e5df638ee07099ba825982f5ad0949d8 28715 F20101113_AABCPB chakraborty_d_Page_35.pro 57aec5ce4e6432a8a50e5180b570f370 56c77a1373f038436bb8dbf159eeb63eb8644067 1781 F20101113_AABCRZ chakraborty_d_Page_47.txt 276fe588910615a4285108e3faed16c7 3774597a964213399c38755ce12836ef1fa45523 61884 F20101113_AABCKE chakraborty_d_Page_49.jp2 d5ce74701e11c90393d20f8388598201 dd00cfe3703e39f8909d455f5ec2732e6f292e20 23695 F20101113_AABCFH chakraborty_d_Page_64.QC.jpg 24003afea07d30c096f024b5d6dfb9df f7f811e0183cf3e785c7df8cbb48cdc6dd4a0102 5935 F20101113_AABCWY chakraborty_d_Page_34thm.jpg 875468abe48040f26e4ca56db410989d 8f227d21f4b463fe182ba36f00858232138fdbc6 31904 F20101113_AABCPC chakraborty_d_Page_36.pro 676faf7a0f503064116c0156a0b05756 d181402b8e0d90f0d8f05ad5bf1e1e00ca50b105 84971 F20101113_AABCKF chakraborty_d_Page_50.jp2 289cb056368eb5f84941a5ccc1a1b73f ae4c7323a52c8d6fa91c18397a7611b405e97b4c 62775 F20101113_AABCFI chakraborty_d_Page_70.jp2 220488debf3a9a1cb576f194e7cfad77 57ae78f58630980a6c9d91748458b40ee25b1fee 16012 F20101113_AABCWZ chakraborty_d_Page_35.QC.jpg ade4583d07e5ba25cfac156c409fae47 5fa96906bd8e92bc6736e0b0837b7b2450164291 10265 F20101113_AABCUA chakraborty_d_Page_75.QC.jpg 8bb0b1ade05119e123073bb6f6238499 25374c77930905b9224be4f063fb8eabfd2f99a9 41545 F20101113_AABCPD chakraborty_d_Page_37.pro 6739dee70a17593b3e3b4c0ba7e0e3cb 482b85e9082f3ef0e5235702cd832c1ba7e63c1b 733058 F20101113_AABCKG chakraborty_d_Page_51.jp2 425af960885f4138f04f972fa47a0540 2540eb8768aaf6e91aeee88eb140459faeed967d 1609 F20101113_AABCFJ chakraborty_d_Page_23.txt 23b7b5ed71ed9a8b68a7d7ff2f66ae3c 0f579dce0fc06e154a0541ca7119a95ad7b1c395 5862 F20101113_AABCUB chakraborty_d_Page_08thm.jpg 928724dc03e062761c5fb40af53471eb 4d9e09c92785102785add5dc4aeb05d01657874c 37558 F20101113_AABCPE chakraborty_d_Page_38.pro 2532a5fbfed415a867fc04c8c4db7b21 08e426006647664c8fcaf9f5401a03dc3be8604e 1051986 F20101113_AABCKH chakraborty_d_Page_52.jp2 a8f1abaac834782f776a9ae8b10553f0 b98f6de49b3b03c9ef6a547371758a631dac62a8 3616 F20101113_AABCUC chakraborty_d_Page_04thm.jpg f5c76c37c0b3f90bd0df63ac14f7d495 1761ae937a56da8c794170227b992719738afbe1 15248 F20101113_AABCPF chakraborty_d_Page_40.pro bfa647e7c34bc260e15368d605f41079 9674f410dff0d7c4b2cc949ff9d405a2c204deb0 98735 F20101113_AABCKI chakraborty_d_Page_53.jp2 f954b67e250063db1f9c7491966c2672 e69dac73aeb0689d0d1cab1cbe8ad942df154239 46146 F20101113_AABCFK chakraborty_d_Page_41.pro 523348d34749db01adf16352285698a3 76289b222670b65c56b831e96d4413b2cf442fcb 6820 F20101113_AABCUD chakraborty_d_Page_16thm.jpg 3e7d05b2459dd3c309ed2a1e52215eb5 9a7fd8afc48f6023e85cf84a674ecc9a03894c51 36524 F20101113_AABCPG chakraborty_d_Page_43.pro 23d2d29dc66c8a769268852b7a47835b 575a02e511bb8193278609d34e6e9cfd6daacf99 71439 F20101113_AABCKJ chakraborty_d_Page_54.jp2 2c02aa89272be2239153b1bc6d83cfe1 6d25dfe74994d1e03ecbc781a42e522a6c869c07 F20101113_AABCFL chakraborty_d_Page_28.tif d2fd2a72e2d5a378325e29e76d4bb2d5 05b7bc665aecbab379c0a9f6b8c73a440c388194 6090 F20101113_AABCUE chakraborty_d_Page_52thm.jpg f2add0cbb5448cbf07473ba03a21a76b 8c8cc3f455c273022428f40f1b4787d15f6f8039 32024 F20101113_AABCPH chakraborty_d_Page_44.pro 707770de67cfc3655b3aa7ba72fcd4ad 9d56ffd10314b99c85a2017b22a996adc3f3ac94 109068 F20101113_AABCKK chakraborty_d_Page_55.jp2 53275b2d89e79d640e29840157a6da09 17b28cfe23db1141641de105fe5f1b1a31013131 9266 F20101113_AABCFM chakraborty_d_Page_25.QC.jpg 36fd5bb39551909749a5f158b03c2df8 59c18e1357b0eb907f752939c8c6be93a9bf0431 2271 F20101113_AABCUF chakraborty_d_Page_09thm.jpg b0964e42ccfa15a25c1cf158ac529a59 d51cba4a08557154a3ab1bcf20bd9e3f309c892f 33761 F20101113_AABCPI chakraborty_d_Page_45.pro 211d4315d8f8d1aa11091b86c6dcb283 5b1da59f64d5ebf86f640bb48450af871df2ebce 901138 F20101113_AABCKL chakraborty_d_Page_56.jp2 25e40a7445eaefa857026c8bb284f300 7ef0dbf9da7caeca40b9ce2654ff18ebcc79c576 98089 F20101113_AABCFN chakraborty_d_Page_39.jp2 6668ad2bb8fb24dce7ae48001346cb65 5ca6fb08a3c98f98af8f7ed9fb0f941dcfdf5beb 20993 F20101113_AABCUG chakraborty_d_Page_22.QC.jpg c54ded323780858d16b903a22ad259b3 c902a4fbc718a2a2c5f9dd7fdab2a7d2fa428348 43458 F20101113_AABCPJ chakraborty_d_Page_47.pro 2d766abaf6746aceda8b7555104fe91a 5d98a1382ad6a2869950e186052f8f55e3df25de 93513 F20101113_AABCKM chakraborty_d_Page_57.jp2 57fc5e11fabd94cf3ee378aff8134750 9c2ed113bf6a03e9819cceaef6ae0ec4dbeff39c 62184 F20101113_AABCFO chakraborty_d_Page_31.jpg 51c43dc3dbb72f6c441869e505e1e1c8 05b509a73c5c8fd0a8f373f7e566bd55ebe26f33 22917 F20101113_AABCUH chakraborty_d_Page_41.QC.jpg 6b81d1c079e1d1a19d850c5536e7c185 4fd25f3c738fbb7f4f193748357258b3d9797306 26820 F20101113_AABCPK chakraborty_d_Page_48.pro feb881676dec6d2bfee067660f16e469 171a3c3f4f92295fa4023c669777776c1a8bccec 75761 F20101113_AABCKN chakraborty_d_Page_58.jp2 073cdd181ccc59a792bf8b084026c5b0 5a4d4b50ba94660ee3b985c52923f7cf4fa80013 1618 F20101113_AABCFP chakraborty_d_Page_34.txt d75d9fdfc4331ab8fa5e056dfb7445c7 550b7d63bc13d47f470a8a581491388c72f8b5d8 7136 F20101113_AABCUI chakraborty_d_Page_14.QC.jpg 3854eeb6d63bf5dddd477b157d38feaa 75137d4db77b4651317935691b3d408b12b4ddf3 26353 F20101113_AABCPL chakraborty_d_Page_49.pro 7e25bc3f0cdf2addc4b6b0f41eaef8a8 62e372f39bead4765915e15845d13797ee5a1b78 73256 F20101113_AABCKO chakraborty_d_Page_59.jp2 b17f8e9711b9396f5a69559f62df80c6 ba7fcb3b25d05d313768a7bd9ada5d602b7c5788 2092 F20101113_AABCFQ chakraborty_d_Page_67.txt b376eefabdeb782ffe51787948c54c17 368eb0783c46383542413db3f2048b5745034229 6009 F20101113_AABCUJ chakraborty_d_Page_13thm.jpg 34d16ddcc58cdacd3997fb7854946848 cca429f3610d800745806aa635a5d58ea7ce1259 39351 F20101113_AABCPM chakraborty_d_Page_50.pro 9ae51df8c4d9f72fcdfcc2275eced50d bbdcaed0cf84b2a89022ecdf5880424c5a73e0e5 90025 F20101113_AABCFR UFE0021236_00001.mets FULL ba2b835b8b360fa5ed3d7e850e9fa689 0aa0e00a8094a5d95b4372971caa113d2d7720c7 24730 F20101113_AABCUK chakraborty_d_Page_72.QC.jpg 91382c33752d86ebb600c2878a373708 c3f660dd84d49f71ca500496f1f735ede7642871 69609 F20101113_AABCKP chakraborty_d_Page_61.jp2 5e43a015c513534576e99550f985188b a2176a61fc4241155c83f14c0b355796ea22f881 31158 F20101113_AABCPN chakraborty_d_Page_51.pro 11ef5fba6c267467b237ef41013e3ad8 4e19a3e02adb5651e0feeea575ad52a9e637a796 6018 F20101113_AABCUL chakraborty_d_Page_20thm.jpg aed50b54e8099aeff9190fb91722b18a 01a83518086f34701fe8c075176fed7840bb30ba 73647 F20101113_AABCKQ chakraborty_d_Page_62.jp2 943d33217d9c2e80d800ade46fa18222 7c08199aee3c3d730735351da648d03f304a209a 32592 F20101113_AABCPO chakraborty_d_Page_52.pro d0492c86f26bd9dd17f2c0357740f66c db47d6d06d020993853859204602ff80f633c342 6830 F20101113_AABCUM chakraborty_d_Page_15thm.jpg 205e8f322f502ca55e7233e3e430d408 0aebc9bd09da76303790aa29f0cc82c281ddb080 87915 F20101113_AABCKR chakraborty_d_Page_63.jp2 c9aa4751b8a21a983df700afec557423 76f3e803d971187a7ca45fb330302afd6a297e8a 21435 F20101113_AABCFU chakraborty_d_Page_01.jpg 744329f594c3d4c1948992fec0a78b10 749e7cacc194228e53fadf9e361d6caa5542a323 33891 F20101113_AABCPP chakraborty_d_Page_54.pro 6d741067a83c74856717611adf4b70d5 c9a173b0e374b3b80e6477de23e62c3706cead8e 27180 F20101113_AABCUN chakraborty_d_Page_73.QC.jpg 0fb41bde70494bf0050ba31fd6f07e98 685ead3470e79ab98d9dc4445f4d3c4096241894 104531 F20101113_AABCKS chakraborty_d_Page_64.jp2 8f8af1f41500406ee934dda87b4e072c 510d872c483906cd1c48ca44cdb14f0664959955 10032 F20101113_AABCFV chakraborty_d_Page_02.jpg fbdcf306f466405b8f03e4b8eae1d3f1 beadb5bfd1dd4a4dc579e2d80d0338fc69380d39 51587 F20101113_AABCPQ chakraborty_d_Page_55.pro 89fda23344ec9e4a3f8e76ded4ab8275 7c222877f8975ca65755a5aacb95ed730e867b7d 3967 F20101113_AABCUO chakraborty_d_Page_56thm.jpg 6055ed1357ea6315f6391d1810f7ba4c ab9b321f685cd31d9b2b905789ed83cfeeeaae34 94020 F20101113_AABCKT chakraborty_d_Page_65.jp2 710eb39bbbd9f7a235b3bc1b11fec827 6314271585b3c819c1648107e0d75257a3ba8f69 9432 F20101113_AABCFW chakraborty_d_Page_03.jpg 888081f6321eae27645aa18d1fa88c25 8f81da298de1df8ddd9c33c94aee3a294280efc5 7635 F20101113_AABCPR chakraborty_d_Page_56.pro eeb29df5865319d8e5d1e9bf14af33d9 f745f25f100be939d0419ea047e430875be45b18 13660 F20101113_AABCUP chakraborty_d_Page_33.QC.jpg 96051cf8f027b70857d0061fd27d9348 af39f6892e2592de0c85a913f93e8a7da567a789 673004 F20101113_AABCKU chakraborty_d_Page_66.jp2 b92195dc085dd268633253033b9d88c5 876e81f86acc74983895bcb67ba9c1af0edefe29 37628 F20101113_AABCFX chakraborty_d_Page_04.jpg 361400e5da92434814065aee0dc1a87d 6c8a861ecf08f9783a5126be5e11b69ebf49c3d9 43508 F20101113_AABCPS chakraborty_d_Page_57.pro 8997645345a83d903b8ae8a01c615d9a c4a61ca318318a7e9fbe6a987def4e101b6ff7e0 5263 F20101113_AABCUQ chakraborty_d_Page_46thm.jpg f33860a652f4c5fc0206f55ea58d7ba9 583f01a864cfa519af9e551f970e66c08b50788a 107738 F20101113_AABCKV chakraborty_d_Page_67.jp2 a8b49fe83545d931a42574209d5b5f14 e84d2d78bf65409268f8b0f7eea54078d4002a8c 79484 F20101113_AABCFY chakraborty_d_Page_05.jpg b453cf0a02452fd86d3c048c9df12514 116b055fdab9c396a3a0fe2dadafe04c452e5f92 30960 F20101113_AABCPT chakraborty_d_Page_58.pro 2c2aabb5ecff4b802485586d0249014c 3d088980bb086f8dba39c37482197b5d4498b0ec 6631 F20101113_AABCUR chakraborty_d_Page_69thm.jpg edc0e42114f41bbcb6964cd4137c7add 4a0e84fa594ce910ad0cf0b5fa670b7050dbe0f5 108182 F20101113_AABCKW chakraborty_d_Page_68.jp2 9aab577f3fe088cf3158e1321a70ff72 f53c0c4891e6e26eccbe2be431799e7662e06452 34191 F20101113_AABCFZ chakraborty_d_Page_06.jpg 9fe1cc2a659dfe95ca8343376506c61a 9ce45e576c83be634f45dca8f676000996c4c899 24707 F20101113_AABCUS chakraborty_d_Page_11.QC.jpg c0b7ec5a998670dca8a1354e9a8a8ff8 d87d68a92b8b7d615542ab4ca52302c2a3dbb145 111369 F20101113_AABCKX chakraborty_d_Page_69.jp2 e71a504a62d0c403bb7510d64ea3f27b d8ced191d6c5907032d727238f8354b1da1569db 33168 F20101113_AABCPU chakraborty_d_Page_59.pro c9ec638f6db224a563ebcf12bbd60529 9b8189139b5320574557fe5de7ba451c3e867c82 20874 F20101113_AABCUT chakraborty_d_Page_13.QC.jpg 48f2af475b78e33e9f28d05a351c60da d8f138cd3cf78b5e77097c6b4a7bb9226db17cdf 127583 F20101113_AABCKY chakraborty_d_Page_71.jp2 b6b8e299eacd5cdc062a9575a0d1c17b 6c0bc4f59e390f9c2fdf13849b29bff150c79f2f 46483 F20101113_AABCPV chakraborty_d_Page_60.pro a2ec6ec87b321a5ae23eef9114020780 104e388592e90d6ee15da29d7795799d1a5cc2c2 71225 F20101113_AABCIA chakraborty_d_Page_64.jpg f2d99870815acc796f8937e301dba9ac 4d0c7481cb7b71242737aed24c66b63169c71b39 5578 F20101113_AABCUU chakraborty_d_Page_05thm.jpg bd38fffdbbf8c63922cfec02f02ea50e 723e3d78dd4ccf25c57f352d687c6fd3679d1053 125179 F20101113_AABCKZ chakraborty_d_Page_72.jp2 881c762b304bf974f4bd11ec1913fe10 d93a4a3b51dbf7fab085997dd53741c7c325ee1c 34512 F20101113_AABCPW chakraborty_d_Page_61.pro 753fbb45f6a7e8305aec4b302e1de4e9 82843f328b0e0f5f4ce76f49450a37ddd1beffc1 64173 F20101113_AABCIB chakraborty_d_Page_65.jpg 41575586450f53f7e17c636f72c651b5 abc7cef932ec53ec1c206ac0881f4e12c54861a4 7051 F20101113_AABCUV chakraborty_d_Page_09.QC.jpg 86db8919f44bd0755f5fb0938b1b5311 d8268922f00f42ad0f23a842b519f4402696cdf2 32374 F20101113_AABCPX chakraborty_d_Page_62.pro f3a9cf4127d931c3dfae9d85101fb94d cdc399032a37ff0f5007bdf1a68a7e295382e410 49284 F20101113_AABCIC chakraborty_d_Page_66.jpg 3bb2432c882b5fddd2a56bf20c1f8a35 053309145e75536bd8ac9929a01bc3f1705156fd 9356 F20101113_AABCUW chakraborty_d_Page_06.QC.jpg 8813564f8cc1c1557e9fec99b70e0cc4 99349604f93a40916f46240f5abbe6d6ed70ceb7 F20101113_AABCNA chakraborty_d_Page_55.tif fbfc591a620ba85363358d39ec045a26 f31c4c55f051b16799b9678c5501b0de5b89d7ed 44915 F20101113_AABCPY chakraborty_d_Page_63.pro be61b02e0376026756334e42dbc6ca4c d3ecdaafe273672a8ea38537bf162a6af702eafb 73175 F20101113_AABCID chakraborty_d_Page_67.jpg f06be6e4d3b997830b3e59e2783b4163 bc4f09c19998b7f20026045d0cac2fbd3274573d 21104 F20101113_AABCUX chakraborty_d_Page_34.QC.jpg 4d27bdecb14658b408aaaf8017c817d4 742b2574ef234ce24023a80a6da0eed8231cf96d F20101113_AABCNB chakraborty_d_Page_56.tif 3975d5988ce5faf45b7ba1a67b2a951a 9f846e0019519a3d3e37ef7db7e939f622429084 49922 F20101113_AABCPZ chakraborty_d_Page_64.pro a63d54f129c190eec94b5d09cea7a17e d61eab14be84a3f6e2bb43c952a48d6342e08279 71539 F20101113_AABCIE chakraborty_d_Page_68.jpg a3c1852df4dfab642e4fbbdb6fd4e3b2 45b99b30c4d7e4c4e1e68b5b2749e216ef9ed902 17688 F20101113_AABCUY chakraborty_d_Page_42.QC.jpg a69e0317f100ddf8ce148a448c510321 05f692fe67b5e8c7032aeee2e68ee9837c18fb33 F20101113_AABCNC chakraborty_d_Page_57.tif 1a642c16b8ba6541d0f4d81647ac43bd dbce5a4ecc027b9974665090eacc57e2e4d92d45 73416 F20101113_AABCIF chakraborty_d_Page_69.jpg b1fa75b59e0086dbda206281fa104669 c1acbc1fd9e8201493c462ce2d25c9cd7a85fe2b 1162 F20101113_AABCSA chakraborty_d_Page_48.txt 68907ba7cf16a8ec7f12700406eabee4 01576b2acf57566f591f0be485e055975d28dd42 F20101113_AABCND chakraborty_d_Page_59.tif a89aa2f2b66f7aac1ba1e5b97d3de64a 727f9209d8571d3132ed55bcc60cd2776d0fc061 43379 F20101113_AABCIG chakraborty_d_Page_70.jpg b607419773e70047777c0af30d597e02 42a6df1ab55c6f790f2ba46d158808259a2a7d68 7144 F20101113_AABCUZ chakraborty_d_Page_71thm.jpg f8a62d49d226a952771af0e6a1e1c8e7 1b38f15e386b64cf203762ddefe09ab667a20e85 1161 F20101113_AABCSB chakraborty_d_Page_49.txt 4e957ae255ca2e71316b35b62c2ea681 e1123fe0e66b6015c3c2aa11165f0b80e6491ce0 F20101113_AABCNE chakraborty_d_Page_60.tif a236e40a9764a88b49ef90d97ca3cec4 051cafa11f8d5ef07bbb612f674c11e5701f69e9 85413 F20101113_AABCIH chakraborty_d_Page_71.jpg 3f6b894be9ea7c0b1317477ebe8d1a76 c798e7cf48eeecd9349a515dfa7410246d66b5fb 1623 F20101113_AABCSC chakraborty_d_Page_50.txt 5271831f75123939ab44d6ce873fb22f 53d467845659faf9ecda965565b14e101f91d9f1 F20101113_AABCNF chakraborty_d_Page_61.tif 80245fe048c03eb5ac87798752532700 80b034ab32343efcc50009dc59fb29d8f467de20 78146 F20101113_AABCII chakraborty_d_Page_72.jpg 02e82045c903677236958a7a81767fb4 e5c9612e9728b3591114c30b4777129384dd2cc5 4501 F20101113_AABCXA chakraborty_d_Page_35thm.jpg 330133afd59fbebb78d757ee877ad2d9 d7d045543427f70a0aee2654401f9b9fa7e0a0e2 1435 F20101113_AABCSD chakraborty_d_Page_51.txt 2531f64f07226a8d446a539a0cdff4ad dcfc9486c6aef1055982c2108352cfa195832b99 F20101113_AABCNG chakraborty_d_Page_62.tif fb78fbb27a313958de88f3fc275bf044 c47c15fb31f1a3786cc8877a678e15a0bc8ebd16 96742 F20101113_AABCIJ chakraborty_d_Page_73.jpg 9c9dc2a7697da0d939a4ee4a9617ba1a da8603cd313ba952e05b089253fc6d57943832f6 5145 F20101113_AABCXB chakraborty_d_Page_36thm.jpg 9833b309b899a045f6fef1ce9b56a4f6 2dabda8be451aad3a30f24729d20fe463ccaae22 1382 F20101113_AABCSE chakraborty_d_Page_52.txt 753bef7aed619d9d81b6e49c7a286f21 0627d70e9f4b3b3d99495cabfe394861ac72c1d0 F20101113_AABCNH chakraborty_d_Page_63.tif d9b681cc1fb958e7620c09867f32327b f07dabbd8c073d7ed038ebda5621fd501a5e461b 75349 F20101113_AABCIK chakraborty_d_Page_74.jpg 2f7e38ea71e5f09dbf7a4d7568b4d56c 306feeb3155026c7441afd1220d269e8afb70a66 5515 F20101113_AABCXC chakraborty_d_Page_38thm.jpg 7451f9c024532a4ba2cfd2f25058e6ce fda9121b775731e3bc753aa73a89081e91e0f45e 1853 F20101113_AABCSF chakraborty_d_Page_53.txt c661519fbca3704e45db6dcb2ccfc8c9 ebb7ac82938a3191e2a1025ee63c637d203a0a2c F20101113_AABCNI chakraborty_d_Page_64.tif 8226fe50a1458aa87d468662cb6c89c3 d508b565ed255825e2d725810d0f849ebb790d02 32503 F20101113_AABCIL chakraborty_d_Page_75.jpg 332b329b4987f68eec1c9747d0014274 3453808f5cd313dcc2f55a199106d88567dd40ab 2953 F20101113_AABCXD chakraborty_d_Page_40thm.jpg 6f8c47f4937204d990ea5c8caff6fdf6 0f53c39dd24eade518d261c643b5ced55215d40f 1430 F20101113_AABCSG chakraborty_d_Page_54.txt 7ca2ad9b8efadf0866e8feeef837d996 b03ced50e1a9d7be724474806e0eef0fa115621c F20101113_AABCNJ chakraborty_d_Page_65.tif cd41ce0f182a10f1585a21ef6dffe0f5 c5671eed06e4fa9c5b82cef96742dcc6819b1524 23282 F20101113_AABCIM chakraborty_d_Page_01.jp2 5d31c4a8dd33c1ebd639b8d27eb2d453 41536af06e5703663433281262592d94ce6aa54b 6460 F20101113_AABCXE chakraborty_d_Page_41thm.jpg 8580d465d55470adeb5ec0f7ae651722 ac9ad6c1d78756cdff0eb2cc691b9b9daf0077a1 2035 F20101113_AABCSH chakraborty_d_Page_55.txt 842d47f6d105d0554b4fc352f56240b6 dc40681665c14692a867e47017f497306058473d F20101113_AABCNK chakraborty_d_Page_66.tif 03ed950bc1d840ee0a5c23d6b259b023 a9fc5afa36007eae874534933b6833a12c8f589e 5831 F20101113_AABCXF chakraborty_d_Page_42thm.jpg 68c5409455a912cdd4576ea8fe1a85d6 7e60eda6274465dee5eda7a422eae8755ebe28cb 516 F20101113_AABCSI chakraborty_d_Page_56.txt 0a8483cb5338ae15018fe29e26e1b531 3a69f55ca8fed981c5740a4b4b7e034271d81795 F20101113_AABCNL chakraborty_d_Page_67.tif 121003f1178d7090f4cd923889f14902 6ac76320e08abc2b0cf2d63347bf326cbfc90512 5665 F20101113_AABCIN chakraborty_d_Page_02.jp2 34a858e41d86e3dc8c0b30a5ae297b7c 1371656f3973843664fc2c3fd272a6fa361eb4d4 5459 F20101113_AABCXG chakraborty_d_Page_43thm.jpg 8ab5a0c1bc61a659420d4f7918f3a322 5eac8b24699d6b15a9dd490beab80b94ec7b4d8c F20101113_AABCNM chakraborty_d_Page_68.tif 28a91159b71b9da7aa631ed28fe5545b ebb3fe8ae03f4cc92c648d8055b4eb99d3042e77 4374 F20101113_AABCIO chakraborty_d_Page_03.jp2 cb2e4b2d4198be92e51d63f166a77afb 54eb0c728699e327db54f2f33a9a1a5d8de3a51e 1805 F20101113_AABCSJ chakraborty_d_Page_57.txt 706faf0d21a19467a98a8d7d92343156 1ce314e6b93b9cd3e3170291fe94b8a733b2eedc 5168 F20101113_AABCXH chakraborty_d_Page_44thm.jpg 54cb7c6748b8ed672c556333a7e4dd8e 0e4b2fd4ae8308f947636dbded8c78e7e13d0541 F20101113_AABCNN chakraborty_d_Page_69.tif f277743313f4ee58900db426862119bb 49b48576e549f93e3d760dc461bbedae0a0ed737 50289 F20101113_AABCIP chakraborty_d_Page_04.jp2 def72be4865745a10ae72c22607dea35 737489827ab3e9a62eb5ccf2fbd0bfd1886916ec 1288 F20101113_AABCSK chakraborty_d_Page_58.txt 4ca84fa063b19f0e01e730928189a814 89b385beeb7ecd4183557f8f1adc9f0b154ba913 16677 F20101113_AABCXI chakraborty_d_Page_45.QC.jpg 9dd44613bba8430f3fb7fd5ac46ef39d e27fa057481b7af10d38eef8333b6b3c8a7186fc F20101113_AABCNO chakraborty_d_Page_70.tif fe93954c2457b0c1491edaee2f61652f bc06e7516debcbf9ac5a1fd3cf2a32db47698f2e 782500 F20101113_AABCIQ chakraborty_d_Page_06.jp2 9edd12b95f4a049a71f71453d41c8716 c94f43f25dd582ab7e6681c3d80a01429718d59a 1359 F20101113_AABCSL chakraborty_d_Page_59.txt 2914906ee7d9e12f24a21a33801db171 ded56660fa0c9b5efcb9fbe25617094df9e8cfcf 18114 F20101113_AABCXJ chakraborty_d_Page_50.QC.jpg f5ec54d51e1d6f76de05550bbcf9b7fe 5944b74de7fd7c5cf4e99782ca86ed1fe1370b9e F20101113_AABCNP chakraborty_d_Page_71.tif 5f2e860032333e03ab959b90c9d50408 860fe0bea8de06e2113fca5ee6bdf79037be6ffa 817868 F20101113_AABCIR chakraborty_d_Page_07.jp2 901e79d64d61bf9e93fa115223596d68 2ad3e6c6199e3eb4f02b704a2d8a4b4fce5e2fb7 1875 F20101113_AABCSM chakraborty_d_Page_60.txt ee415c5ac67e131ce1a001955bd596c6 dec7b525d1380556a9aff27b5493ca9a5ee91dbf F20101113_AABCXK chakraborty_d_Page_51.QC.jpg 5f672d3c94940eb14e93a75e3edba8a3 d6a11d8b3ac940fe25d21f863d51ee468647350e F20101113_AABCNQ chakraborty_d_Page_72.tif b706c2c9e71637515f315fe0c1a7e782 008505251521bbc4c0e882309338b16188a0f336 94921 F20101113_AABCIS chakraborty_d_Page_08.jp2 108701d6423a8d7c8cd1f45d8e223785 1d6a103d57711fd6dbdc2c57f40c13ccae04b485 1453 F20101113_AABCSN chakraborty_d_Page_61.txt 77011e012d94c862a15d2d6233d324c9 69a80cc7af974d4d931b88e57cc9f78633fbdf60 5360 F20101113_AABCXL chakraborty_d_Page_51thm.jpg 8f8cda6fc2d022c497f7b6f59798671b 98cff12f2c15c93e6972511099257a922802bb02 F20101113_AABCDW chakraborty_d_Page_58.tif e9ca02b79d8eb294c092e4a28897173b 3a115187c47ba039ee9631b50d3ab4228b1453f9 F20101113_AABCNR chakraborty_d_Page_73.tif 48ebb7da6ba90d6060eee8555e6a4d9c 140c9c48bab39a3204466c7fb59c95bfaec54c1c 23667 F20101113_AABCIT chakraborty_d_Page_09.jp2 f982c3c622fdd6d3d0e7faffd42a31f0 b80f015b0fcdbdfa248f546fa6a2d1e917239b7b 1796 F20101113_AABCSO chakraborty_d_Page_63.txt acab775e97b583fd737b0ffd7cef3d4e 348658755d186177cb200889982a5fc6e4866af3 23307 F20101113_AABCXM chakraborty_d_Page_52.QC.jpg c23cb49c0cad9c5de1f5a3547bdc00bf 04d27fb02cb0f0bc6eb3740639dc70f1ba45838c 65948 F20101113_AABCDX chakraborty_d_Page_47.jpg 330a22981bb364dd760883932b966b59 2dacae5ce48cfb62c40721804cad8059cc1e6d50 111884 F20101113_AABCIU chakraborty_d_Page_10.jp2 b267049e4015ecc4fd2d3dae06542bcd fe8a93783dc7b6ddb89b1cf098e2c2089cf95abc 1992 F20101113_AABCSP chakraborty_d_Page_64.txt 70eba206575ad7a09ca661abef18e633 d503d4147ee26d44007e0e1981dadbc7673a73d8 6170 F20101113_AABCXN chakraborty_d_Page_53thm.jpg bbe539f4c0dc06c326dcc97ccc5ae49d c337bfb123a2bc691add53a3ae4e1a858824e306 23217 F20101113_AABCDY chakraborty_d_Page_30.QC.jpg f7d2db0e66f5a8dafae9c15a5fd8e40b 3c94ab196b30eaf6ea833843600f00276b41aa21 F20101113_AABCNS chakraborty_d_Page_74.tif 5668b400a12bdfae5bf1ddaf58a1af74 544a50579d5c45ac51e33f7ceeb45284b6da47f8 97914 F20101113_AABCIV chakraborty_d_Page_12.jp2 e966495d439b531b230f2520b546a81e 0946f2a6f2ff8a0395e962c2483a77f7e50f8da9 1810 F20101113_AABCSQ chakraborty_d_Page_65.txt 25aac4b8033737b1b7d4233a53ae2a5f 6825e6cc6132a3e9f24478b8504e491ec6415c02 5309 F20101113_AABCXO chakraborty_d_Page_54thm.jpg b24745cba730347056ae6c74ce21ae99 2c9a517f5335c8f175b09516cd4e90b18d93de30 21024 F20101113_AABCDZ chakraborty_d_Page_31.QC.jpg b95d935474da00be46c8aad7b541c57c ee0dace60ac39a29d26227593a1c5ac4668bc37b F20101113_AABCNT chakraborty_d_Page_75.tif 58244bf6c16e6a2bab679de8e1632e6d d04b45264c42178c32df591d5955d555c663b085 886289 F20101113_AABCIW chakraborty_d_Page_13.jp2 1cda4b4d4c60c5e0a0dc58e715ca5f30 7be396c54e75cb9d3f17a7aad50fe22998aa7b36 869 F20101113_AABCSR chakraborty_d_Page_66.txt ca2b23f7a4bbfbb582c48c6e97b3ca6b 0b56f02b5874bd9f8ec0106e2e07b0b747920197 24951 F20101113_AABCXP chakraborty_d_Page_55.QC.jpg 45f36db2a3067496e80255e9819b29a7 b9704b6b88a993a1f371c7d64e0d88e8645b7717 938 F20101113_AABCNU chakraborty_d_Page_02.pro 8ad40482f132473fe6e2d4b0a989ebc7 236e9c161c9609c47b3170a6b7ab7bc2d21f830f 26281 F20101113_AABCIX chakraborty_d_Page_14.jp2 b525728417bf14d0f72f4d788f3479b5 4c658c0fc44c2d1a0a27252b941e9472d14f0134 2018 F20101113_AABCSS chakraborty_d_Page_68.txt 9d48fd6e2efb8abf547c831de5ba50e2 45a93ff02d15b11a9e10c0704838f0fed2cebc13 6852 F20101113_AABCXQ chakraborty_d_Page_55thm.jpg 9c4942cdba7d8347b906a99a951f3562 a232ffa3791b9702558c3551edcdfc6208137f84 626 F20101113_AABCNV chakraborty_d_Page_03.pro 972ec5ca5c66f304f91cb5be8861cac2 5ccc493322d6f2b244c659568544ee8f2f49b123 35205 F20101113_AABCGA chakraborty_d_Page_07.jpg 104461ae2096ffd3d95cec4c3d8cba2a c21e0e5efc9ab0a5c7160f5864e1d657dae8b594 112225 F20101113_AABCIY chakraborty_d_Page_15.jp2 d4eeb33cb7d093e18169760ef855dcf0 dac78b50924363c39825c1270e8b0101006b40bf 2045 F20101113_AABCST chakraborty_d_Page_69.txt 4377cbe9c069fd38fed090ffb14de080 6eba7cb71a9320267a5d04e163d1b13326260ef6 13515 F20101113_AABCXR chakraborty_d_Page_56.QC.jpg 7998eee8843a8fd3f5cea73240830e4a 96881f1b0e04b84d97bf17ed94c9a4218d438bc5 21587 F20101113_AABCNW chakraborty_d_Page_04.pro 3dc86e9a30df1ffe5996c2ff84b14ee4 db462c62838d47d172ff5f0dcf27be2ef394fc83 65498 F20101113_AABCGB chakraborty_d_Page_08.jpg 09325a005ee0b83e784eb993e9bfac97 d80a22f583b67548ae3f1e5ec14486618e7f89d5 115485 F20101113_AABCIZ chakraborty_d_Page_16.jp2 88851af1684860f681a118dc02655fd6 e26cc900b0a6bdbf4990485514dc7adc99cc7027 1123 F20101113_AABCSU chakraborty_d_Page_70.txt d92c300dd6dbc1d4ca56521567a53fbb 327f3589d60bbf848e786919d4fd7b6ae9709e04 17306 F20101113_AABCXS chakraborty_d_Page_58.QC.jpg e423327e50dc2c1cb8e5320578bd7cb9 ae9b4aaee12ba03dd2e1f6feedb59fdd58b25d1c 65470 F20101113_AABCNX chakraborty_d_Page_05.pro e0ef1c87871adcf5655c5e5b9a8490d0 3775e6963a8f0e2a50991e03326e97fbbcd9c006 75686 F20101113_AABCGC chakraborty_d_Page_10.jpg 249beeacb6c0957bc9b3518a941ebcb6 519efedc68ef73f7ac393f74cca54a1a5b897d18 2438 F20101113_AABCSV chakraborty_d_Page_71.txt 8d067f887d6bb3b133b85e58d1458653 436bf71dcbe5a1f266470bbe129ef917ca34baae 6204 F20101113_AABCXT chakraborty_d_Page_60thm.jpg e3770be44b76c35c26bb5ee6897415a9 a3242030ec4326ecfa870954ec867e18b33488aa 22363 F20101113_AABCNY chakraborty_d_Page_06.pro 33c48a076683ec99d1031c27d924b73a 3b632b36da74deffb0488dbf935539b254fea087 75316 F20101113_AABCGD chakraborty_d_Page_11.jpg b3f5e7a63558debbfaf6155e1dc65b59 bd270d84b948f66ad58aec16dc5be6690087b2cf 2321 F20101113_AABCSW chakraborty_d_Page_72.txt b30e9b65813ce1a2a22e7ef3d2a1cb6d b860ab2ce222eae4b45ed7bb3b87c6a184983c74 1051963 F20101113_AABCLA chakraborty_d_Page_73.jp2 0e486b3aca17d43ac28d5f317c252fc4 ed27152e8c48ac7ea057c7dfe4cb6e35a2659088 4967 F20101113_AABCXU chakraborty_d_Page_61thm.jpg e80cf31d112e3a4a8497a048fc8f9a2d cca382a3202f9bd04ffc4bf97fa1190f2a74a07c 22486 F20101113_AABCNZ chakraborty_d_Page_07.pro d9cf9499397e1c314f4cc8924b7bb98e c0835360f0f43d39707abe69bfeb0a099dcbdcc6 66307 F20101113_AABCGE chakraborty_d_Page_12.jpg 3fc28b1720d6baf01c75c441d4bd01dc 50028f7242f88d1656d9dbb1b80a14d47815c94d 117781 F20101113_AABCLB chakraborty_d_Page_74.jp2 4d0249f0525fb0cfad9e9645d0b145f8 aea326afd506d77dac69c0d244ff9baf0eb5c6fe 16274 F20101113_AABCXV chakraborty_d_Page_62.QC.jpg 07a69e2665f8fa0727d378c64fb70560 9e0e2ca1b3a1b993d1859c0392baa132c01a8a4f 65597 F20101113_AABCGF chakraborty_d_Page_13.jpg e98e515dd5e321bec673e36ea76a7a9b 2919d74606ab875cd354f37a7502549eac0804d1 2439 F20101113_AABCSX chakraborty_d_Page_73.txt 03fc506797ed2e234b5622d6fb884d9a c88021fe93877bfb727adb9d08781f8395439394 41612 F20101113_AABCLC chakraborty_d_Page_75.jp2 85882148d34c7e4d68358b213aa62395 4131f4e5897e00905ce69abab2d62bc3120025c7 5201 F20101113_AABCXW chakraborty_d_Page_62thm.jpg 6d398ac9ab2930a0746f7633a0d086ca 28bce92f095511308984cc5f1c7a015a4727a518 22081 F20101113_AABCGG chakraborty_d_Page_14.jpg 9c075f507e7ba300174aa0b25e384b18 ea5c808090857e71f09b1125272a693ae6a7929d 15688 F20101113_AABCQA chakraborty_d_Page_66.pro 3e9b6ef769e338d9d4c6b358b03748b3 ae4b1dd6034eb37bcc1efc1ac73e9d7f9ba5b56d 2166 F20101113_AABCSY chakraborty_d_Page_74.txt dbe5c8391207dab454cc75f2e7cd0154 ad25d939ba9d8149b82bd066a4575eac400a9e6c F20101113_AABCLD chakraborty_d_Page_02.tif 5ab6c8f8785aea923c8b0432f4c4f559 e0fb5cdea67e1923c95f8932bf59217a9f884f3f 21073 F20101113_AABCXX chakraborty_d_Page_63.QC.jpg 2c68a2301bf92e9c1118555e7b0d7666 2142af8e05dd4683a1c3bf43a0c341b5db2fb0d6 75742 F20101113_AABCGH chakraborty_d_Page_16.jpg 7350d8a10184aae9a82070dd2024b9f0 85c5383799528c9d4d3ec2b0eaac357f520df5fd 50530 F20101113_AABCQB chakraborty_d_Page_67.pro 0096da0d6576499717d546bdd35f5089 bec5cc4c5535478b57fe56179ece7d39d47e318c 728 F20101113_AABCSZ chakraborty_d_Page_75.txt ecadacc11dd1fa32e51d9e659ebca134 5663b10a6f19c8db550585ac19690d6094c95800 F20101113_AABCLE chakraborty_d_Page_03.tif f28a6be535048a59872371e83e10fc2a 3b80eba925dd383e31851ce2fc30c71368acc046 5892 F20101113_AABCXY chakraborty_d_Page_63thm.jpg 76bc1d3acf7038ff1b590fdc535dec33 f0af9d7935afb4cb21728a5127b631ee4949ef17 77141 F20101113_AABCGI chakraborty_d_Page_17.jpg a22373bd96fb7555df1ff3ccdc03c496 61c34583d042df82bca10d4f2202a8abf4b7f61d 49327 F20101113_AABCQC chakraborty_d_Page_68.pro 7d5dd7c26e80b8cb0b187f8cfea504f6 539d4a5f82c7eb63d98700d722517eff53859bfe F20101113_AABCLF chakraborty_d_Page_04.tif 2232ddbc3c1b417199767a4c57f6ab75 1079bfb3e8ff34f0304f3f6c51b5eb00fa6588d3 6572 F20101113_AABCXZ chakraborty_d_Page_64thm.jpg faf9e49bb6425737eb85d511a8c7948d fac98a789d8ee1d88d794c592bd30d814b1680c5 23485 F20101113_AABCVA chakraborty_d_Page_47.QC.jpg c9d4b2ee0dc935a0f1948899738d19f0 0b51be16cbd71f9ac0f45fca162685c42accc829 33205 F20101113_AABCGJ chakraborty_d_Page_18.jpg 8d6ff897d09ec376211f5a8f90458c47 c4f4a8ba6d73040a8259911642e98788626d7e4d 51436 F20101113_AABCQD chakraborty_d_Page_69.pro 91192d9d60a48587c0a8fa01fbdcbd7c a7e18447ed8f3f2dbb6326da6f6c98c8e5472480 F20101113_AABCLG chakraborty_d_Page_05.tif 40f86c3a858a9fd9291a7c00656b885f 8523c49e8e6d42e7392b66dadbfe26aafb338e39 3162 F20101113_AABCVB chakraborty_d_Page_75thm.jpg 3e161af44ab172ebd367b7f081e069f3 4d3468ccd3526a22599c97ba7f4a132d01528e49 61059 F20101113_AABCGK chakraborty_d_Page_19.jpg 3c6b25d95ef61f0e37fe1d1dd5fd3824 23596bffc65756f0139dbf4995fc4dd5ed24eed1 27817 F20101113_AABCQE chakraborty_d_Page_70.pro bf30ee98b0c9b6912e8c147095c329a2 1c31fcd405b1aa849ef6b47c26d3f194ae99349b F20101113_AABCLH chakraborty_d_Page_06.tif 307c0bd26477dcb9fd8b4b2d8a9c089e 54b0c539f474b88b60c8e7e34cb5bb9c7b20a193 6155 F20101113_AABCVC chakraborty_d_Page_47thm.jpg a2a6697a36689e1df5cd9a19a9b67f68 3a0b4f3c49a39b41c2acc649396f12eb87ceb7c1 60133 F20101113_AABCQF chakraborty_d_Page_71.pro e9f17093aa0ae955bb9e44c67ba697f4 5619a7cbf130c47db00b4f0f3d175529e3a06998 F20101113_AABCLI chakraborty_d_Page_07.tif fb15808b91d4120ec23bdc303443dd0f fc844c9f5a6612399f346a4b4e1d714b97fb4a8a 15407 F20101113_AABCVD chakraborty_d_Page_36.QC.jpg 95f3446191e56c40e74b21744266f3b3 2b10f7f81c975ce0bd2b27e20f8ccab44834d91e 62227 F20101113_AABCGL chakraborty_d_Page_20.jpg 7a887eca3e719e7fbfee4cc8f6d1561a ddd3f0013a06501f0be195672b23fa412b31e3db 57552 F20101113_AABCQG chakraborty_d_Page_72.pro ee16b582e07fc8c7a338a705ae714018 37c45f5aef843dcb992ce4b9ed5fe8e5ce6ffffa F20101113_AABCLJ chakraborty_d_Page_08.tif 3ef86988cbbe27d1217c97fb257d1b7d 14f87cb607eac195d414e217d9d160bf83b3237e 5068 F20101113_AABCVE chakraborty_d_Page_48thm.jpg 70eb8ce97b820eab3a781a8fce77b968 d52791ce620d71be9755f9bf440a3d1ea9d27077 64335 F20101113_AABCGM chakraborty_d_Page_21.jpg a4cd1cfb541cc8e45eea1ad270465f28 310cd2abe1b02ab931403a1e74eba716f8af98fe 60355 F20101113_AABCQH chakraborty_d_Page_73.pro 70857f9c06a316d5e1a8458f9ed177d4 b68f46265341a43579d73da6c6f1283b22554d08 F20101113_AABCLK chakraborty_d_Page_09.tif f76ba9918534d1a7919300e4cd0ab984 978b16e48afa7e4f8561db4ec3eb386d720ca8cb 5060 F20101113_AABCVF chakraborty_d_Page_58thm.jpg c9aa6b6a42f500dfd78d231bdd912946 6e5be5c3e9535d242de01e266eb7c6f80adcf00e 60673 F20101113_AABCGN chakraborty_d_Page_22.jpg 4c1273aedb37925d789e6d52fca6fb9c 2c34ae304e5893d5ad1c667d682b4f43643f2bc7 17345 F20101113_AABCQI chakraborty_d_Page_75.pro 2f12f8a92e4d402f446abbe8885388ae b66e96876e6b0d0b921dcb80f301c089f53b8c3f F20101113_AABCLL chakraborty_d_Page_10.tif a35dbd9bcc86aa8dce29ec7c940a4eed eae50601ede91a1c7c29021bbf9fad402b2eee35 22989 F20101113_AABCVG chakraborty_d_Page_60.QC.jpg 1b269ef50d34bfdccfcfa3d65939c68f 5648d312abb34ae3310e187af6988535ac834aa5 56636 F20101113_AABCGO chakraborty_d_Page_23.jpg 56a33a709ffa50386d63b0550c7abb4c c7215e877578cced45ac877a9b2367ccae652fd3 404 F20101113_AABCQJ chakraborty_d_Page_01.txt 748d74dc788aa635857b9a1533cbeafc 35af6e7890639c41d905bf1c15707b7aabf43a81 F20101113_AABCLM chakraborty_d_Page_11.tif cd679e64e125db0de5bff8ae976930b8 c8a379d6181a73f4ff62ff158be90a2d75ca11bf 16708 F20101113_AABCVH chakraborty_d_Page_44.QC.jpg 85f01b25e4e90f037dfe0561099c2922 b4cdbc47e572b0d42483af7e2ecf70792342c068 60921 F20101113_AABCGP chakraborty_d_Page_24.jpg e68be68d143f8d45f873a4f92e58bbd0 89c1dadf3b6c81968593ccf3c810ab8cafd62f86 88 F20101113_AABCQK chakraborty_d_Page_02.txt a3c53f58be416ceba947b10994ca760f 558a55ddaaa57677f71335171b5cee4970234309 F20101113_AABCLN chakraborty_d_Page_12.tif a97d2d928cf1ceb14b930fe75f080994 128a969182dddc4fd0b0a94d370e6d2d7c1d5b83 5593 F20101113_AABCVI chakraborty_d_Page_23thm.jpg 56c1ac101976e601fcd556a0a65bdd14 edd2b95937ca7b525446386642f428634d5a9a2e 25675 F20101113_AABCGQ chakraborty_d_Page_25.jpg c72afb0e1da1dd00d6bbf98a9910e07a ad8441aade670b7a131a174e8f21fa988ca803b9 80 F20101113_AABCQL chakraborty_d_Page_03.txt 107f54ab7a71f456d7d0f68b1eb9c8df 2e6e34cd504cbd9e0f35bfb406bc526ceb058f9b F20101113_AABCLO chakraborty_d_Page_13.tif 17edb547d445e9ff9008c912671773d2 6ec1a32ab3ee040f76afa768be54d2f375a9bcac 9804 F20101113_AABCVJ chakraborty_d_Page_40.QC.jpg 56aec91cc8afaf63bde739bfc70ed644 96a4367593923b22d66410916486a299749a0ed3 65294 F20101113_AABCGR chakraborty_d_Page_26.jpg 06d72fca39418b4144ebba51f8de3fa5 0693d070b46e7bbb82fa724eeb7e187590f3ec73 894 F20101113_AABCQM chakraborty_d_Page_04.txt 87ba08d9a6011fdeba2871e74e2531a0 8d0d6a38a07076e1d84457b72638e042b675a40e F20101113_AABCLP chakraborty_d_Page_15.tif 8f5293ab56d8d2a4507729be5d7ce84f a1cc8ff0bec961c9251f0767ecaf65fca04ea486 5880 F20101113_AABCVK chakraborty_d_Page_24thm.jpg c66d66a834f201104bb8788284b1848d 01cf100760534032796431d1e9bc21960ac9e90e 67906 F20101113_AABCGS chakraborty_d_Page_27.jpg 6d19cc6494bb5fcc64bf165c89df5e81 07ab3ef77a7bfeb2a9b77a1256a882a7165d4fb7 2850 F20101113_AABCQN chakraborty_d_Page_05.txt 23b6505bb3174df64f4e52921468c086 8e5a76cf60ac915150ad9d47828c32afe8e4710d 2786 F20101113_AABCVL chakraborty_d_Page_25thm.jpg 40d2a06c1f4a74e095affe6b8be48e6b 9a643a79d3389ec4e3aef4e043b1276e67391dbd 68303 F20101113_AABCGT chakraborty_d_Page_28.jpg 043ecdcce86f9a68140253a2290ac97a 68a5d84dd349e424843fa490297bc24af0e2dd3f 914 F20101113_AABCQO chakraborty_d_Page_07.txt ad70346d250c965f4a5e4557093d35ec e23251bcfc393242aaacbf7fc818db73b1aefe70 F20101113_AABCLQ chakraborty_d_Page_16.tif 1f490dcd8fe4fc6f4592e52a2118e0b7 8c2e1b31b47210ff36abe3bc95731339451ed0b3 15048 F20101113_AABCVM chakraborty_d_Page_49.QC.jpg 79b908b0748c8131b788097e18940097 e64ba4b3b097c3f32988a25293850e505eb32b1a 32462 F20101113_AABCGU chakraborty_d_Page_29.jpg 317525f06ff3b3879cd6af5beea79ab8 c933707282699b4711cfba16cd2ae180ccc5c42e 1925 F20101113_AABCQP chakraborty_d_Page_08.txt 23b544c1ca34ac6887cf7ab70c768242 d045633b12a5b0499a0bbbfb0644563d170af91e F20101113_AABCLR chakraborty_d_Page_17.tif 911f09c71207ab715e4c41a0327d8218 e52f6b80978a7f5b76fe5d3b1c146d393b086122 20405 F20101113_AABCVN chakraborty_d_Page_08.QC.jpg 95c31b3d7f2ce1534d2f2e15b2bf5e51 98ad93ab2770a15b05118641d950e96984e3335d 71300 F20101113_AABCGV chakraborty_d_Page_30.jpg ac937f052ef8fef4cb6897f33ef924b9 0a063167836eeace37c0fac1fab8bc68bb918b9a 379 F20101113_AABCQQ chakraborty_d_Page_09.txt 5cb812d8bfafb00f8019498356c840f5 70f632384a6c5a351762380ff2316915e4a90fb0 F20101113_AABCLS chakraborty_d_Page_18.tif cdda48fad50606bfe0e9d5a6719c721f b5ba3aacb00a58ea240b1f904234817c80f3030b 22519 F20101113_AABCVO chakraborty_d_Page_68.QC.jpg c39ddb0c40194da0274bac33f0ffb71d 48f707c9a536e79cd95b1bcf71956e2f760884b8 54123 F20101113_AABCGW chakraborty_d_Page_32.jpg 7850218754fef21223bdca22456d0275 e377295b5b7993d50d12747a3243e10279de061d 2146 F20101113_AABCQR chakraborty_d_Page_10.txt c2ab31733a7d393f6a0e25b70cbbe23c 406e95f8721e795d10bcbe427d63e18d895a50b2 F20101113_AABCLT chakraborty_d_Page_19.tif be12ced3e7ec31e77e59652e72068779 2c690e7161245b3d8c813e30e0930bbf44b696c0 16756 F20101113_AABCVP chakraborty_d_Page_59.QC.jpg 74a648b904d88e40d012a8bb594903a8 d8d614b11c420a1b99e718a263818351e46aa58f 43169 F20101113_AABCGX chakraborty_d_Page_33.jpg 1a9e119cfc5e92415994751bb8c6b95f 9eeb6b14580e83325c0ed18fb03d6fa37238a1d0 2151 F20101113_AABCQS chakraborty_d_Page_11.txt baf8d57d5a2a3c7e57d6137513efd786 fb1e5e3baaf5c0e66f67d3ca47f387c25c79a9b7 F20101113_AABCLU chakraborty_d_Page_20.tif a88900fc944cec038961324d3ac79804 de4f921925cfb38dd8abd911dbce1407b50bae98 22208 F20101113_AABCVQ chakraborty_d_Page_39.QC.jpg 85ef95f25ae044b2d74e8336dd918c50 56a8e63c7e8fc02c040fa949c90251863c70af73 1865 F20101113_AABCQT chakraborty_d_Page_12.txt ad2ceb9e17f3906996b8b9e84c550897 8249d45fd82966bd16aafbdccf48d8e7ce3c438a F20101113_AABCLV chakraborty_d_Page_21.tif cd59f99f4d309c3a29cbc4755eb92120 7b57aef25b37363d3e9fa197dfca6cddf89018c2 24585 F20101113_AABCEA chakraborty_d_Page_16.QC.jpg f3f7f6f5a84bb516d79d958fdd10587f d3a3e1c97948f6ccc425e76d94b585235728ac75 59982 F20101113_AABCGY chakraborty_d_Page_34.jpg 1bebfb5aeced59262d02456e1cad8802 0b650bfd431df1c260fc93d1673583b1db8cfe5c 18172 F20101113_AABCVR chakraborty_d_Page_54.QC.jpg fadbd980baa3daa7503417bd7df2df0b 1f5d6bddc666d2ef7ff1fdb43620e5d65a955696 1617 F20101113_AABCQU chakraborty_d_Page_13.txt 6e3299efbb6ad5c5278744b76f440de7 6d9e6e7049f37b2b269356c3eadd8359c8f07599 F20101113_AABCLW chakraborty_d_Page_22.tif e6354ead0639c0215c517737f0113c9c 145cb4f407e40daa592ee65d1d6d3cade0b1c773 6939 F20101113_AABCEB chakraborty_d_Page_74thm.jpg fc309c31e22437991ab6835b1036e7ed f05322f5009de0bedaa024b2918cdb144263a83e 43715 F20101113_AABCGZ chakraborty_d_Page_35.jpg 4817dbfba70905c2b9cfc827433ec1d9 23c87d2e4ac75e9aa87b8923de57afea766e0ccc 6322 F20101113_AABCVS chakraborty_d_Page_28thm.jpg 5db524d48d25091a5e59a01bf5289dbf fe18ba8fac05a30162ba9d614c5c417f66207627 F20101113_AABCLX chakraborty_d_Page_23.tif 31f7ec577e4d707067c0f4b8395f998a e5eceb1c15600da05db6335f3feecdebdbaf98a9 97592 F20101113_AABCEC chakraborty_d_Page_60.jp2 f1d838b7f5978462a16fec31626091ec b2be6a0a5b3769e0ed63616b47125915910de0a3 22096 F20101113_AABCVT chakraborty_d_Page_05.QC.jpg 48e2a226d013327feb1d8d47125e03be af09e7e993049f9df355b4833c279347c2586d8a 428 F20101113_AABCQV chakraborty_d_Page_14.txt 20f9d7fb61325e201f64748bb271decf cec037175d361fbb133c6a863059cf25338949cb 117429 F20101113_AABCJA chakraborty_d_Page_17.jp2 60c01206588b0315daedc82adfd0f99d 1d74b672b4fad2c4321e2dbbaf47b481c9842394 F20101113_AABCLY chakraborty_d_Page_24.tif 45e38cbdc8776c7ac1ee9d66b23f4fbf 62dd35f5f1cfeee7870dfe122bc220d4318f14d3 20966 F20101113_AABCED chakraborty_d_Page_57.QC.jpg 429993d89ea9975d760076ce78654a7e 287a9085ffafded50098c88bf557d5ec77d4d6b6 5472 F20101113_AABCVU chakraborty_d_Page_32thm.jpg 87c84f8cd0de13143a9410e943b391c2 0214c9c378e54ff5e8f5a406cf7a8f6dd4c12b14 2118 F20101113_AABCQW chakraborty_d_Page_15.txt e12049ebce99a8c21a6269704f2ab489 ed14f8cd0b5f8a0b070bb37a0c8847c80fb3c845 45744 F20101113_AABCJB chakraborty_d_Page_18.jp2 9c5c46198a2d8b7ec9dab0a08c2e78c8 8fee45193c54576c38c414c43da44ce0df8c201c F20101113_AABCLZ chakraborty_d_Page_25.tif 149d8fc9f6be6746886eaf2b63dca3d5 38f0605fc84585a8406f958fbdb0b76f776d021b 14054 F20101113_AABCEE chakraborty_d_Page_70.QC.jpg c198e75b891ec951dc1ec16c46624439 15aceb765cbc40dd2c35c721a18c368e9b10feb1 3898 F20101113_AABCVV chakraborty_d_Page_66thm.jpg dc5bc41885ca3a5392c0c7cc3d5f55d2 07910c1e549ba1abe550aa4abf9c48fe59d2cb03 2063 F20101113_AABCQX chakraborty_d_Page_16.txt 742a87f21c05101442a971e88e13aa3d aec02611f8359ebbf4cb63e0261f511abcd65d9d 91509 F20101113_AABCJC chakraborty_d_Page_19.jp2 fe113f26d62f2ce2bd7aef6815ad5786 14ce20fd6613894f70dd460fbfb51678d9a3ce5f 106215 F20101113_AABCEF chakraborty_d_Page_30.jp2 c15d61cd383fcc960086be9374a64e41 446221c5ed4379a6b6986e4000a0325a20c35e14 24653 F20101113_AABCVW chakraborty_d_Page_15.QC.jpg 0736a183897c32a372f5ccb39119ff2f e512f84ad9463dfa907fb7288a945f12aa563eec 2153 F20101113_AABCQY chakraborty_d_Page_17.txt ee2033f685f5aa39d6b0245c04b7e62c 43848385853dd76551c29bcc2330b81a7f52e38a 93553 F20101113_AABCJD chakraborty_d_Page_20.jp2 7fdba81105007e549f07e34594fedeaa 24e94935301c1fdbfefc79aab9d99ac55e430d14 F20101113_AABCEG chakraborty_d_Page_30.tif dedfb03fdeccf58288b417d095012616 37292a460f3104b4db2094152d503292a2fd7036 43796 F20101113_AABCOA chakraborty_d_Page_08.pro d0c9b108bb3a30add7306ca93d57e355 e72cd3c6304d4f9f8695466caef0e33322060f06 116410 F20101113_AABCVX UFE0021236_00001.xml 984b381cb2dfad94d1133ba1267276e3 0ae5d4ffcc78207ce088e73849408b53c27687a2 780 F20101113_AABCQZ chakraborty_d_Page_18.txt 50f71aaf406cc26a74b1395bc7f6be6a 61c15b0f6918eea9412313dffd30e4c877c5221f 97714 F20101113_AABCJE chakraborty_d_Page_21.jp2 5625a9664917f8432a528fc22a9757aa 82563f3831c560e1e08119fd0963f5b6425fe4e1 33028 F20101113_AABCEH chakraborty_d_Page_46.pro 33f3c2cf956a4c80d675ee8565979837 fee2c215be28c641f0148a6d2fd357210fdb9e1a 9509 F20101113_AABCOB chakraborty_d_Page_09.pro 548ef41c45adb992536acb4c03c7c85c 9d0e22a535c51ef1d01a15b1a096bb3232ead0ae 6838 F20101113_AABCVY chakraborty_d_Page_01.QC.jpg fa97f14fe0d3e11c9f50e1944c23e464 c404656350b96e8b528239de398174f420f72107 89905 F20101113_AABCJF chakraborty_d_Page_22.jp2 34c5409007894b6fd5415702a69c4139 c3049fc1e5989a577c753bddd4a088f277b82706 F20101113_AABCEI chakraborty_d_Page_14.tif b98208927ae01f503bc6265389f18c43 f422c64fd55e11bbd3dca2ae125d9e4fee2919b6 52333 F20101113_AABCOC chakraborty_d_Page_10.pro b790754706ada483e8b14c1df245e406 68a80831d4785eebe67966d73d8fa5254a89d996 1337 F20101113_AABCVZ chakraborty_d_Page_02thm.jpg ad2263d97d51dc6690d9f75739dbc5fb 8529674c04011868c0f1686fe9f3303ec6ded47c 82541 F20101113_AABCJG chakraborty_d_Page_23.jp2 1d61c90e05e86ceb1d257ee6575dcb19 cdd80c7a44503d9a17544a8e3f4b0bdd9e5ff75e 814542 F20101113_AABCTA chakraborty_d.pdf 313d6126510eb914ef64196c6ee6ced5 95b904e89ba94d84ac2944c91127aae3b7c2ea18 54606 F20101113_AABCOD chakraborty_d_Page_11.pro 53c006ae2e96ecfd8f4c41df64075f52 b8ebb2eba1cefe13f630851bcb65f2e96c80f591 88333 F20101113_AABCJH chakraborty_d_Page_24.jp2 9c9f582a8ffcc2a20ac3c3c6189a152b e2f47adbe310fde3f2f9604da30775133f1ce1da 2212 F20101113_AABCTB chakraborty_d_Page_01thm.jpg b539a9bc024fc10ce0c195056df890bd ae10bef82c934dc171621b4b888291d145622568 48995 F20101113_AABCEJ chakraborty_d_Page_44.jpg 0876c024a8530f1828f72e32148b441b 7c798ad2afa9ec0c873590600c334f3715d9e1f8 44843 F20101113_AABCOE chakraborty_d_Page_12.pro 7dadf50845ce3329689ca7d0544c7c2d f391d1b6d5fc6224b036614cb814a470eecd1f91 29483 F20101113_AABCJI chakraborty_d_Page_25.jp2 f0350f40ae0b3eb685c7ff4df374c571 2150652227742e8bdd8dc4c10e7e6dec6dbd99e3 23841 F20101113_AABCTC chakraborty_d_Page_74.QC.jpg 6c2103b6eb8c12605927749e02e9c76c 629bae551c280d5abcde392d11c14769563590f7 5956 F20101113_AABCEK chakraborty_d_Page_22thm.jpg 9f8d5375109602728ec0ccb1c5cbc3b5 0b51f4c744770856578d7a7741721d384bf2385d 36320 F20101113_AABCOF chakraborty_d_Page_13.pro c71b249ae52bba15d5da4073cf28b39a ec0d85a7b80c5838d9c18178f2d806f95dcbf0fe 21810 F20101113_AABCYA chakraborty_d_Page_65.QC.jpg 0efb5a19bca91081c35dc22666a864ca c952f5b1520076ae496c90949b0caee5bae2fd78 94708 F20101113_AABCJJ chakraborty_d_Page_26.jp2 e6cb6349e2ce3f8d7470c58abfada6a7 ab79e77efd818dc1425c2c4480b7bd2552495210 6622 F20101113_AABCTD chakraborty_d_Page_10thm.jpg 480f14d657c26b771dcb16a9bc01ec7f 7e119d9f18226a43f374f6b576e3fa9a2f80af53 6510 F20101113_AABCEL chakraborty_d_Page_39thm.jpg 11853c8b2b2a6b23af13f63440194f90 617bcd25a11c92f3359fbc79f59ba9ff5a1ba97d 10720 F20101113_AABCOG chakraborty_d_Page_14.pro 485d66643faa397eba5eb2d39ba366a3 5e2c97e54420e2f0506df847866e4f2c39871e69 24281 F20101113_AABCYB chakraborty_d_Page_67.QC.jpg 5afc8772b02d572eb231fe180deb2f33 61f2ed2b611aa494c561f5b3ceed77bc1e39a23b 100553 F20101113_AABCJK chakraborty_d_Page_27.jp2 ae69f1b4fa67cdcd3c22d4ca33a96d10 e0a8d1141bc796a4f2092084ca92ad0679cd1354 4719 F20101113_AABCTE chakraborty_d_Page_49thm.jpg a42c20eae617fa7acc1438ef8392911f 1f6cd5a343808fc83fdfb9289ccd90b4bb703766 1831 F20101113_AABCEM chakraborty_d_Page_26.txt 9f82947f74a48d2c21f800621f17c9a7 1c202d7970b6f758992f484371e499775ff53e49 51484 F20101113_AABCOH chakraborty_d_Page_15.pro 0597273d514cbb9700ecf4876db2bd0c b31f4fa31f43318c8cad15422850ff99ab1801be 24888 F20101113_AABCYC chakraborty_d_Page_69.QC.jpg 52a8cd1e13f188f872662204768b9aaf 47f4dab4b25087e1fbde1edbad233c2718afd202 99445 F20101113_AABCJL chakraborty_d_Page_28.jp2 70ca17c98d1abc7aa68a97a6a3577d61 f0f37bb27ecbcd5a691a2c5c10972eaddc0a02d9 6531 F20101113_AABCTF chakraborty_d_Page_68thm.jpg b4abeac3f4887e6cd9fe2568652d1fc6 ae4ece46d99aaa6edfbb12dda33af34e215e91b5 F20101113_AABCEN chakraborty_d_Page_52.tif cd2c8f2d53dbc1db766ace15532c5633 3343ae96a31f5e31df0826b7d44a7d50187a8d8d 52487 F20101113_AABCOI chakraborty_d_Page_16.pro e127b2d33c92cadaccb7cf0b3179400f 2d80612cc9ff035cb2fc056f729f816405c54c12 4229 F20101113_AABCYD chakraborty_d_Page_70thm.jpg 934e1c60b9c7d9443710bc6349924ab1 ef366c1009b29c728b1d58009ef8f5113e33beba 42963 F20101113_AABCJM chakraborty_d_Page_29.jp2 498d7feeb7a20e1d54dbca67430ed78a 2647bc86290d070511b36ded4f5be1695fd997b6 5269 F20101113_AABCTG chakraborty_d_Page_45thm.jpg 2273de04add84c9fe698a0dce5928166 3b8a1468bffe159001b86d10577fa60306eee2a7 45485 F20101113_AABCEO chakraborty_d_Page_53.pro 923b88bc49bfebcd57ea9111faebd5b4 43ce220aa81e22ea92f92b940b4a0e25fe84e1b9 53414 F20101113_AABCOJ chakraborty_d_Page_17.pro c1502b7bceb24e2f473c6cf4f3349556 1d0010df258bc51472e4ffe0b84df4e5cf7a322d 6617 F20101113_AABCYE chakraborty_d_Page_72thm.jpg 3100dc5ffa0b8418cd6999fff7ca6159 00a1c5fc954bd59626873c297484cb6e7f48f17e 88920 F20101113_AABCJN chakraborty_d_Page_31.jp2 6587ade60a32386f7d79cb082bb30824 90206929dfe6ca241196ec45b40f2286344f5e4a 25984 F20101113_AABCTH chakraborty_d_Page_71.QC.jpg e2f0ced4d7bea16aa7a90269224e9704 560cc5ad8a7301ec8fee8f98c7075c8428d9fca0 5990 F20101113_AABCEP chakraborty_d_Page_31thm.jpg 7ed0b0c8ecbe1c6f50c6303f16bacb34 47713a8d03ed42d4fa86a64f038fc586007a1891 19632 F20101113_AABCOK chakraborty_d_Page_18.pro a6dde839bdb4ee68c33007db79098775 f4e1387c194e29ff24c34d93e5cf48976c23ee9c 7504 F20101113_AABCYF chakraborty_d_Page_73thm.jpg b5307b852ccbaa9343c531293987a157 34e39d574600f808165fb144c690da4ea5aea5e2 16540 F20101113_AABCTI chakraborty_d_Page_48.QC.jpg cd18141963998ab7bda19750c32ffd0f 60566eb9da5b868d0fd983a0d8fa73c479e7d853 53707 F20101113_AABCEQ chakraborty_d_Page_74.pro 469b80c7005f98d35ac5da6e78a596a5 7451820a53ceef96e3d29be96c1dfcbb47d62014 41938 F20101113_AABCOL chakraborty_d_Page_19.pro 9ed9fd197857d022e3aa498e3847cfc4 eff8c90c81440f0a9149d71020b969af8702ac2e 3088 F20101113_AABCTJ chakraborty_d_Page_02.QC.jpg 1b16f1ae9531c65cc41f7b58abe5fdaa 584e1206b27c60d8ff36ef3b161cf81f53efb53f 78350 F20101113_AABCJO chakraborty_d_Page_32.jp2 25e38f81160b109f0a81f0f441756c0e 87665a64e7691546fe52b40ae22c5113f8efe7a2 6271 F20101113_AABCER chakraborty_d_Page_21thm.jpg 609d94e6b6bef6bff14932141a906b5c a4e5e7d1b979b207700e2b906b71f3c473de363d 42092 F20101113_AABCOM chakraborty_d_Page_20.pro 2ada1bb76cd2d0de89c8da9aca0c2822 99ca89bb47d2f1ab6814ef7270553dd94b4ce1fc 6700 F20101113_AABCTK chakraborty_d_Page_67thm.jpg 95fc3c1c27da036d12ed1ef5cbaabfc1 6f463a5c9eb4dacc42e7050bcabfddf8c1400e34 56162 F20101113_AABCJP chakraborty_d_Page_33.jp2 77cc29559f121f8fbb96fdd1f2f08ef9 50590f50a285e5d1d3f20499e52cc9fd6f5ade78 930 F20101113_AABCES chakraborty_d_Page_06.txt 8525699912efe3a143086f3a99b078c9 dc66f9bc3772ed9735ec219b3f3a424f51b75d98 44510 F20101113_AABCON chakraborty_d_Page_21.pro 3190fbe6893b0ab1778ebdbd4c3c3e4c c61abef3a042a382aab19bb5d211a6e7a1e92fd6 7123 F20101113_AABCTL chakraborty_d_Page_17thm.jpg 02e13b9fb2a08b671c6be27931434502 836d3fce84443ad06c044fc93287e172470c897e 81911 F20101113_AABCJQ chakraborty_d_Page_34.jp2 73840acc1c1b2a22893cb3643c368ea2 567db4ad52f0c56e3f918014e59b9055ab33bfa7 F20101113_AABCET chakraborty_d_Page_01.tif 0d4079c9fcb150b66edecd584ca36082 edf8c8f529c8a9ed469d4126db5476accdceb89f 41160 F20101113_AABCOO chakraborty_d_Page_22.pro 0c3d75cb62576fc43ad01f3fcde0baa0 20e8ff1bcaaa4938abe06b07a54d4b8ef6c870b2 17290 F20101113_AABCTM chakraborty_d_Page_61.QC.jpg 8422d2940ef61cc3b159c52d1b4f92fe 8a96105d003d221fe0f630159586b88f1b8b77ce 60936 F20101113_AABCJR chakraborty_d_Page_35.jp2 0a219ef758cad815079c424cc4494d6d 2e18e3f5e92c348898fa204a5d06f8423fdc0957 1365 F20101113_AABCEU chakraborty_d_Page_62.txt a84a83b9ecc9753281667e7dadf264c7 72740d92db38ce170d898f611d042163628f7546 37451 F20101113_AABCOP chakraborty_d_Page_23.pro 50896763e43935183c857ba2d74eeebd e074096d649590e771f4cbc843fe914b126f0c8e 5907 F20101113_AABCTN chakraborty_d_Page_37thm.jpg 62b8d613c68b2c66fdf27acd8651e70f 6289b1b451a5972781df89ae55641fd7ae53a00b 67482 F20101113_AABCJS chakraborty_d_Page_36.jp2 aec5e05bdac400d835f1f8e6148866b9 4656052a695688fc1769eceb663157fc3669fe0a 6099 F20101113_AABCEV chakraborty_d_Page_57thm.jpg 411766e5cd9c255141fbf2f1fb51a4a7 b64be0ea19027f54111f1379f0273daea1f85d7c 40907 F20101113_AABCOQ chakraborty_d_Page_24.pro 5de51a5395ac45db79c85461c0fdce49 696553b5e5aadb3d3a7bf10818465ad8d03c6443 2963 F20101113_AABCTO chakraborty_d_Page_03.QC.jpg 77ec1faf47fd1e672bf5bd74d1f4f295 d3340ef07ac8cc85d76f9183edd4d38c91115b74 88430 F20101113_AABCJT chakraborty_d_Page_37.jp2 a71e5ff2eaaa4b21ece354bebb1f79e6 9bc6ec67cda1c0a353785a940bd41a8e63b36c1c 44032 F20101113_AABCEW chakraborty_d_Page_65.pro 3f2882c22c801e945910ba66630dec2f 492ca83df553bb2702f718556fbdbe4d67222218 12968 F20101113_AABCOR chakraborty_d_Page_25.pro 199989359322a64736c8482c333ed295 c18b4448d665ac02dacea98e2ec439423900d26d 14441 F20101113_AABCTP chakraborty_d_Page_66.QC.jpg 8d28bc724c180151f684558a4978d429 9ad5bb115f415508fc2b80a197ac6ed663ab32eb 21748 F20101113_AABCEX chakraborty_d_Page_53.QC.jpg b19302b9ac73278e2c87a343ab411700 1afa01abc274a3975799603e30a8f94cf56453e8 44002 F20101113_AABCOS chakraborty_d_Page_26.pro 4772819c63541dc4293bfd02695d2dbf d011b14338b1f79faa3d052498d7259697db6f53 79447 F20101113_AABCJU chakraborty_d_Page_38.jp2 f10996ce66513a32a496afc6caefe984 459b2c66ca5e705c1b06eb56e7ce0162580478e1 21991 F20101113_AABCTQ chakraborty_d_Page_27.QC.jpg 3e3dfcb46358cc3a0663a625f732273e 12775368641fab587295760dcb54d68189bac1dd 46147 F20101113_AABCEY chakraborty_d_Page_39.pro 192ae20e5deb76a1f545e2e53711ed6d a56418daf0a22b4b56877bcffb79ef62cb26312c 35084 F20101113_AABCJV chakraborty_d_Page_40.jp2 0dded4c7b2bbebf5e4d3907fefb4b776 2a0c2b56a9f7b16cce077aec4693001df79cc27e 5921 F20101113_AABCTR chakraborty_d_Page_65thm.jpg 1a09e0ae26607d5cdd747b3e74822706 4c830ecbcd308268fd94f1a1d58a1e1d6b5b3537 20323 F20101113_AABCEZ chakraborty_d_Page_09.jpg f57eb30208feadb5aabffa7ce0da7421 0cb39637fdad489e18be3eb4f079b45e64d832a6 47050 F20101113_AABCOT chakraborty_d_Page_27.pro f75965533cd7c27a5b99921f1214c874 d13852cc926317616a8c7432764dbf35e66afe28 100483 F20101113_AABCJW chakraborty_d_Page_41.jp2 0e7dd1d3abcd3db9eec4b8fdcdce7bb2 75f63c842b20062ca80c3a8c9d4940ca894d3a77 5739 F20101113_AABCTS chakraborty_d_Page_50thm.jpg 9e3ce2bac459e2b6c775822b7bf8c902 13f9b8633e9b18c9c118292b90f1feaac8d5bac1 46384 F20101113_AABCOU chakraborty_d_Page_28.pro a28ca6e378a0f54ca4ab943791ce1ca7 4c5019ef324f4b9fa8ab40033575de7cb2ab89ef 81855 F20101113_AABCJX chakraborty_d_Page_42.jp2 ee21d258e2e20395c4cd45c8ab23ecdd 25ea9358b093551da8b0b339ff2416749c5fc321 17619 F20101113_AABCTT chakraborty_d_Page_46.QC.jpg 841132a7938676349754dd296148d7a3 ce3a41f4bdbe2125dee944dad6e8598e98d32d53 19338 F20101113_AABCOV chakraborty_d_Page_29.pro 0bb8ce430be133ae03f4532f8b21794c 5ee493c782ff4bd9218434786cbb866aea65e3dc 48849 F20101113_AABCHA chakraborty_d_Page_36.jpg 184d1d74e6e31ebd4bded9a1020dbde5 43b1ad081d2b2b33fb62ba4a2a8991d40e8f6dd6 79964 F20101113_AABCJY chakraborty_d_Page_43.jp2 bf4a066ce9fb4763c1872a99db075ec0 fd3bdd1d0cd2caea31ce63487f7ab9a64f6111dc 3301 F20101113_AABCTU chakraborty_d_Page_18thm.jpg 19fb1ab1528c854a43309422e152c808 38b5e79c41ccdb6adbb2a47309ba048e0c7219b2 49233 F20101113_AABCOW chakraborty_d_Page_30.pro a8e03baa2e767f56c72c7db554e47e3f 8045d07cc780e120453126bb5179fb849f57f722 62338 F20101113_AABCHB chakraborty_d_Page_37.jpg 1ef9150b6fc4b1804fa26c70aca52bad 30bf91247cf8f9d3eb5c04134fa53c046111bf6b 73294 F20101113_AABCJZ chakraborty_d_Page_44.jp2 c7d153509d7310e3e08aaf847f59e106 ace5b870238b65fda879c2d5ce84d6a3021118b6 21483 F20101113_AABCTV chakraborty_d_Page_20.QC.jpg 8646917d123743bbf546ae69032d4fe7 6ff55c2160f12d03c646c4f8c17ae8b0babd262b 42305 F20101113_AABCOX chakraborty_d_Page_31.pro 0ca298b8274be1c8b1df1c28fb3b7222 191810a3eb122c3d732bf61a7d49f676ef801894 55708 F20101113_AABCHC chakraborty_d_Page_38.jpg f238c5b0f4c6be9533ab472970af64e3 ccd5b96433bf7b166d7736826ad0707da0841bc2 4646 F20101113_AABCTW chakraborty_d_Page_33thm.jpg 50f142cc01151a5897e6924963ace7dd ab643fda9f6b7056ea3fc559401e2f4abc99d7d5 F20101113_AABCMA chakraborty_d_Page_26.tif a3d25d3c2005abdee16ce79594a632e6 084a0ac3c4934fa3ec45c8183b4506dd63de438b 36212 F20101113_AABCOY chakraborty_d_Page_32.pro ab65d7c644d011795bd84865d3c77c86 066de013b3a1e2ca7ea425af597961beb5c99232 66346 F20101113_AABCHD chakraborty_d_Page_39.jpg 23612c28144a875743d5863971aca767 b1990d51b49c5d0efe90c99c9f2c20c8039b5f14 17533 F20101113_AABCTX chakraborty_d_Page_38.QC.jpg 6ae6282868631a4887bd9857063b6c41 5849808a0c9970da4aecf1f8fdf39a33f6b1dffd F20101113_AABCMB chakraborty_d_Page_27.tif b93c09acc332a4f6d4fa41f4a7e3cca8 df1cf70620b72789b3660b2d2df44f481c15a78e 24462 F20101113_AABCOZ chakraborty_d_Page_33.pro 969d5905dbaec024d4023e811ccbc8f0 3aed4cf4def08fe56eb91a5475c08be630cd1f5a 27003 F20101113_AABCHE chakraborty_d_Page_40.jpg 4998136087f73e4c1368e24a882e5ec2 5acd0c19ef2ca1712d24fd9ea119e95a0544c0cc F20101113_AABCMC chakraborty_d_Page_29.tif 57a27a9f51aa149b1cef8c667799e2bd 134039d89226570c6ba0de9f3064aad0d03ea49b 68888 F20101113_AABCHF chakraborty_d_Page_41.jpg 0a1f3f64066ddc57632126047dd7d72f 4714001aa15af2fa2e93220fcde8b7e195d0a516 5270 F20101113_AABCTY chakraborty_d_Page_59thm.jpg 235be334a9c6f0c53f1e6d823efbf089 69bd2f197bd5e34c4616327776ad48364a483ac6 F20101113_AABCMD chakraborty_d_Page_31.tif 7245e1e87cce3909b6291df98d5a39eb b3acf75c67ef91b85afb846e4b91c41d5aee3bb7 55692 F20101113_AABCHG chakraborty_d_Page_42.jpg 8f3266970182ad7160d76771a5bbeb3d 82ca6706e6eb8479341c91c7a20da9bf73f023c6 1783 F20101113_AABCRA chakraborty_d_Page_19.txt c46bfdc11acb33b2781650b7dfe7b07f 5ff5f638a095664de5fd5b7121038d209b76b6ee 6687 F20101113_AABCTZ chakraborty_d_Page_11thm.jpg f00e962d7244db05d86fdb9d02c287af 358b31f51c7fa9a94ba139b629a8b9cf0cbe9fec F20101113_AABCME chakraborty_d_Page_32.tif a5cdef598f444b3a45b88b4c0ec57c9e 0e0afa712f979488c2171739b0f607fbb322c1e1 56759 F20101113_AABCHH chakraborty_d_Page_43.jpg 30ba29d0a027b0029e565f53fb91528f 6d95652e9678a34f733e6adafff90d6a9e561ddc 1716 F20101113_AABCRB chakraborty_d_Page_20.txt 588c00422074dc8a19902c46d21dffa8 c09d42a71dcbbd713d35e0c341699ed8ca48a095 