*375*
*91*
*4MB*

*English*
*Pages 577*
*Year 2019*

- Author / Uploaded
- Sethi S.P

*Table of contents : Preface to Third Edition......Page 6Preface to Second Edition......Page 8Preface to First Edition......Page 11Contents......Page 13List of Figures......Page 20List of Tables......Page 241 What Is Optimal Control Theory?......Page 25 1.1 Basic Concepts and Definitions......Page 26 1.2 Formulation of Simple Control Models......Page 28 1.3 History of Optimal Control Theory......Page 33 1.4 Notation and Concepts Used......Page 35 1.4.1 Differentiating Vectors and Matrices with Respect To Scalars......Page 36 1.4.2 Differentiating Scalars with Respect to Vectors......Page 37 1.4.3 Differentiating Vectors with Respect to Vectors......Page 38 1.4.5 Miscellany......Page 40 1.4.7 Concave and Convex Functions......Page 44 1.4.9 Saddle Point......Page 46 1.5 Plan of the Book......Page 47 2.1 Statement of the Problem......Page 51 2.1.2 Constraints......Page 52 2.1.4 The Optimal Control Problem......Page 53 2.2.1 The Hamilton-Jacobi-Bellman Equation......Page 56 2.2.2 Derivation of the Adjoint Equation......Page 60 2.2.3 The Maximum Principle......Page 63 2.2.4 Economic Interpretations of the Maximum Principle......Page 64 2.3 Simple Examples......Page 66 2.4 Sufficiency Conditions......Page 77 2.5 Solving a TPBVP by Using Excel......Page 813 The Maximum Principle: Mixed Inequality Constraints......Page 92 3.1 A Maximum Principle for Problems with Mixed Inequality Constraints......Page 93 3.2 Sufficiency Conditions......Page 102 3.3 Current-Value Formulation......Page 103 3.4 Transversality Conditions: Special Cases......Page 109 3.5 Free Terminal Time Problems......Page 116 3.6 Infinite Horizon and Stationarity......Page 126 3.7 Model Types......Page 1324 The Maximum Principle: Pure State and Mixed Inequality Constraints......Page 147 4.1 Jumps in Marginal Valuations......Page 149 4.2 The Optimal Control Problem with Pure and Mixed Constraints......Page 151 4.3 The Maximum Principle: Direct Method......Page 154 4.4 Sufficiency Conditions: Direct Method......Page 158 4.5 The Maximum Principle: Indirect Method......Page 159 4.6 Current-Value Maximum Principle:Indirect Method......Page 1695 Applications to Finance......Page 181 5.1.1 The Model......Page 182 5.1.2 Solution by the Maximum Principle......Page 183 5.2 Optimal Financing Model......Page 186 5.2.1 The Model......Page 187 5.2.2 Application of the Maximum Principle......Page 189 5.2.3 Synthesis of Optimal Control Paths......Page 192 5.2.4 Solution for the Infinite Horizon Problem......Page 2026 Applications to Production and Inventory......Page 212 6.1.1 The Production-Inventory Model......Page 213 6.1.2 Solution by the Maximum Principle......Page 214 6.1.3 The Infinite Horizon Solution......Page 217 6.1.4 Special Cases of Time Varying Demands......Page 218 6.1.5 Optimality of a Linear Decision Rule......Page 221 6.1.6 Analysis with a Nonnegative Production Constraint......Page 223 6.2 The Wheat Trading Model......Page 225 6.2.1 The Model......Page 226 6.2.3 Solution of a Special Case......Page 227 6.2.4 The Wheat Trading Model with No Short-Selling......Page 229 6.3 Decision Horizons and Forecast Horizons......Page 234 6.3.2 Horizons for the Wheat Trading Model with No Short-Selling and a Warehousing Constraint......Page 2357 Applications to Marketing......Page 245 7.1.1 The Model......Page 246 7.1.2 Solution by the Maximum Principle......Page 248 7.1.3 Convex Advertising Cost and Relaxed Controls......Page 252 7.2 The Vidale-Wolfe Advertising Model......Page 255 7.2.1 Optimal Control Formulation for the Vidale-Wolfe Model......Page 256 7.2.2 Solution Using Green's Theorem When Q Is Large......Page 257 7.2.3 Solution When Q Is Small......Page 265 7.2.4 Solution When T Is Infinite......Page 267 8.1 Nonlinear Programming Problems......Page 278 8.1.1 Lagrange Multipliers......Page 279 8.1.2 Equality and Inequality Constraints......Page 281 8.1.3 Constraint Qualification......Page 286 8.1.4 Theorems from Nonlinear Programming......Page 287 8.2.1 A Discrete-Time Optimal Control Problem......Page 288 8.2.2 A Discrete Maximum Principle......Page 289 8.2.3 Examples......Page 291 8.3 A General Discrete Maximum Principle......Page 2959 Maintenance and Replacement......Page 301 9.1.1 The Model......Page 302 9.1.2 Solution by the Maximum Principle......Page 303 9.1.3 A Numerical Example......Page 305 9.1.4 An Extension......Page 307 9.2 Maintenance and Replacement for a Machine Subject to Failure......Page 308 9.2.1 The Model......Page 309 9.2.2 Optimal Policy......Page 311 9.2.3 Determination of the Sale Date......Page 314 9.3.1 The Model......Page 315 9.3.2 Solution by the Discrete Maximum Principle......Page 317 9.3.4 Incorporation into the Wagner-Whitin Framework for a Complete Solution......Page 319 9.3.5 A Numerical Example......Page 32010 Applications to Natural Resources......Page 328 10.1.1 The Dynamics of Fishery Models......Page 329 10.1.2 The Sole Owner Model......Page 330 10.1.3 Solution by Green's Theorem......Page 331 10.2.1 The Forestry Model......Page 334 10.2.2 Determination of Optimal Thinning......Page 335 10.2.3 A Chain of Forests Model......Page 338 10.3.1 Formulation of the Model......Page 341 10.3.2 Solution by the Maximum Principle......Page 344 11.1 Models of Optimal Economic Growth......Page 351 11.1.2 Solution by the Maximum Principle......Page 352 11.1.3 Introduction of a Growing Labor Force......Page 354 11.1.4 Solution by the Maximum Principle......Page 355 11.2.1 Formulation of the Model......Page 359 11.2.2 Solution by Green's Theorem......Page 360 11.3 A Pollution Control Model......Page 362 11.3.1 Model Formulation......Page 363 11.3.2 Solution by the Maximum Principle......Page 364 11.3.3 Phase Diagram Analysis......Page 365 11.4.1 Model Formulation......Page 368 11.4.2 The Implementation Problem......Page 369 11.4.3 The Optimization Problem......Page 370 11.5 Miscellaneous Applications......Page 37612 Stochastic Optimal Control......Page 380 12.1 Stochastic Optimal Control......Page 381 12.2 A Stochastic Production Inventory Model......Page 385 12.2.1 Solution for the Production Planning Problem......Page 387 12.3 The Sethi Advertising Model......Page 390 12.4 An Optimal Consumption-Investment Problem......Page 392 12.5 Concluding Remarks......Page 39813 Differential Games......Page 400 13.1 Two-Person Zero-Sum Differential Games......Page 401 13.2 Nash Differential Games......Page 402 13.2.2 Feedback Nash Solution......Page 403 13.2.3 An Application to Common-Property Fishery Resources......Page 404 13.3 A Feedback Nash Stochastic Differential Game in Advertising......Page 407 13.4 A Feedback Stackelberg Stochastic Differential Game of Cooperative Advertising......Page 410 A.1 First-Order Linear Equations......Page 423 A.3 System of First-Order Linear Equations......Page 424 A.4 Solution of Linear Two-Point Boundary Value Problems......Page 427 A.5 Solutions of Finite Difference Equations......Page 428 A.5.1 Changing Polynomials in Powers of k into Factorial Powers of k......Page 429 A.5.2 Changing Factorial Powers of k into OrdinaryPowers of k......Page 430B Calculus of Variations and Optimal Control Theory......Page 432 B.1 The Simplest Variational Problem......Page 433 B.2 The Euler-Lagrange Equation......Page 434 B.4 The Brachistochrone Problem......Page 437 B.5 The Weierstrass-Erdmann CornerConditions......Page 440 B.6 Legendre's Conditions: The Second Variation......Page 441 B.7 Necessary Condition for a StrongMaximum......Page 442 B.8 Relation to Optimal Control Theory......Page 443C An Alternative Derivation of the Maximum Principle......Page 446 C.1 Needle-Shaped Variation......Page 447 C.2 Derivation of the Adjoint Equation and the Maximum Principle......Page 449 D.1 The Kalman Filter......Page 453 D.2 Wiener Process and Stochastic Calculus......Page 456 D.3 The Kalman-Bucy Filter......Page 459 D.4 Linear-Quadratic Problems......Page 460 D.4.1 Certainty Equivalence or Separation Principle......Page 463 D.5 Second-Order Variations......Page 464 D.6 Singular Control......Page 466 D.7 Global Saddle Point Theorem......Page 468 D.8 The Sethi-Skiba Points......Page 470 D.9 Distributed Parameter Systems......Page 472E Answers to Selected Exercises......Page 477Bibliography......Page 485Index......Page 559*

Suresh P. Sethi

Optimal Control Theory Applications to Management Science and Economics Third Edition

Optimal Control Theory

Suresh P. Sethi

Optimal Control Theory Applications to Management Science and Economics Third Edition

123

Suresh P. Sethi Jindal School of Management, SM30 University of Texas at Dallas Richardson, TX, USA

ISBN 978-3-319-98236-6 ISBN 978-3-319-98237-3 (eBook) https://doi.org/10.1007/978-3-319-98237-3 Library of Congress Control Number: 2018955904 2nd edition: © Springer-Verlag US 2000 © Springer Nature Switzerland AG 2019 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, express or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

This book is dedicated to the memory of my parents Manak Bai and Gulab Chand Sethi

Preface to Third Edition The third edition of this book will not see my co-author Gerald L. Thompson, who very sadly passed away on November 9, 2009. Gerry and I wrote the ﬁrst edition of the 1981 book sitting practically side by side, and I learned a great deal about book writing in the process. He was also my PhD supervisor and mentor and he is greatly missed. After having used the second edition of the book in the classroom for many years, the third edition arrives with new material and many improvements. Examples and exercises related to the interpretation of the adjoint variables and Lagrange multipliers are inserted in Chaps. 2– 4. Direct maximum principle is now discussed in detail in Chap. 4 along with the existing indirect maximum principle from the second edition. Chattering or relaxed controls leading to pulsing advertising policies are introduced in Chap. 7. An application to information systems involving chattering controls is added as an exercise. The objective function in Sect. 11.1.3 is changed to the more popular objective of maximizing the total discounted society’s utility of consumption. Further discussion leading to obtaining a saddle-point path on the phase diagram leading to the long-run stationary equilibrium is provided in Sect. 11.2. For this purpose, a global saddle-point theorem is stated in Appendix D.7. Also inserted in Appendix D.8 is a discussion of the Sethi-Skiba points which lead to nonunique stable equilibria. Finally, a new Sect. 11.4 contains an adverse selection model with continuum of the agent types in a principal-agent framework, which requires an application of the maximum principle. Chapter 12 of the second edition is removed except for the material on diﬀerential games and the distributed parameter maximum principle. The diﬀerential game material joins new topics of stochastic Nash diﬀerential games and Stackelberg diﬀerential games via their applications to marketing to form a new Chap. 13 titled Diﬀerential Games. As a result, Chap. 13 of the second edition becomes Chap. 12. The material on the distributed parameter maximum principle is now Appendix D.9. The exposition is revised in some places for better reading. New exercises are added and the list of references is updated. Needless to say, the errors in the second edition are corrected, and the notation is made consistent. vii

viii

Preface to Third Edition

Thanks are due to Huseyin Cavusoglu, Andrei Dmitruk, Gustav Feichtinger, Richard Hartl, Yonghua Ji, Subodha Kumar, Sirong Lao, Helmut Maurer, Ernst Presman, Anyan Qi, Andrea Seidl, Atle Seierstad, Xi Shan, Lingling Shi, Xiahong Yue, and the students in my Optimal Control Theory and Applications course over the years for their suggestions for improvement. Special thanks go to Qi (Annabelle) Feng for her dedication in updating and correcting the forthcoming solution manual that went with the ﬁrst edition. I cannot thank Barbara Gordon and Lindsay Wilson enough for their assistance in the preparation of the text, solution manual, and presentation materials. In addition, the meticulous copy editing of the entire book by Lindsay Wilson is much appreciated. Anshuman Chutani, Pooja Kamble, and Shivani Thakkar are also thanked for their assistance in drawing some of the ﬁgures in the book. Richardson, TX, USA June 2018

Suresh P. Sethi

Preface to Second Edition The ﬁrst edition of this book, which provided an introduction to optimal control theory and its applications to management science to many students in management, industrial engineering, operations research and economics, went out of print a number of years ago. Over the years we have received feedback concerning its contents from a number of instructors who taught it, and students who studied from it. We have also kept up with new results in the area as they were published in the literature. For this reason we felt that now was a good time to come out with a new edition. While some of the basic material remains, we have made several big changes and many small changes which we feel will make the use of the book easier. The most visible change is that the book is written in Latex and the ﬁgures are drawn in CorelDRAW, in contrast to the typewritten text and hand-drawn ﬁgures of the ﬁrst edition. We have also included some problems along with their numerical solutions obtained using Excel. The most important change is the division of the material in the old Chap. 3, into Chaps. 3 and 4 in the new edition. Chapter 3 now contains models having mixed (control and state) constraints, current value formulations, terminal conditions and model types, while Chap. 4 covers the more diﬃcult topic of pure state constraints, together with mixed constraints. Each of these chapters contain new results that were not available when the ﬁrst edition was published. The second most important change is the expansion of the material in the old Sect. 12.4 on stochastic optimal control theory and its becoming the new Chap. 13. The new Chap. 12 now contains the following advanced topics on optimal control theory: diﬀerential games, distributed parameter systems, and impulse control. The new Chap. 13 provides a brief introduction to stochastic optimal control problems. It contains formulations of simple stochastic models in production, marketing and ﬁnance, and their solutions. We deleted the old Chap. 11 of the ﬁrst edition on computational methods, since there are a number of excellent references now available on this topic. Some of these references are listed in Sect. 4.2 of Chap. 4 and Sect. 8.3 of Chap. 8. ix

x

Preface to Second Edition

The emphasis of this book is not on mathematical rigor, but rather on developing models of realistic situations faced in business and management. For that reason we have given, in Chaps. 2 and 8, proofs of the continuous and discrete maximum principles by using dynamic programming and Kuhn-Tucker theory, respectively. More general maximum principles are stated without proofs in Chaps. 3, 4 and 12. One of the fascinating features of optimal control theory is its extraordinarily wide range of possible applications. We have covered some of these as follows: Chap. 5 covers ﬁnance; Chap. 6 considers production and inventory problems; Chap. 7 covers marketing problems; Chap. 9 treats machine maintenance and replacement; Chap. 10 deals with problems of optimal consumption of natural resources (renewable or exhaustible); and Chap. 11 discusses a number of applications of control theory to economics. The contents of Chaps. 12 and 13 have been described earlier. Finally, four appendices cover either elementary material, such as the theory of diﬀerential equations, or very advanced material, whose inclusion in the main text would interrupt its continuity. At the end of the book is an extensive but not exhaustive bibliography of relevant material on optimal control theory including surveys of material devoted to speciﬁc applications. We are deeply indebted to many people for their part in making this edition possible. Onur Arugaslan, Gustav Feichtinger, Neil Geismar, Richard Hartl, Steﬀen Jørgensen, Subodha Kumar, Helmut Maurer, Gerhard Sorger, and Denny Yeh made helpful comments and suggestions about the ﬁrst edition or preliminary chapters of this revision. Many students who used the ﬁrst edition, or preliminary chapters of this revision, also made suggestions for improvements. We would like to express our gratitude to all of them for their help. In addition we express our appreciation to Eleanor Balocik, Frank (Youhua) Chen, Feng Cheng, Howard Chow, Barbara Gordon, Jiong Jiang, Kuntal Kotecha, Ming Tam, and Srinivasa Yarrakonda for their typing of the various drafts of the manuscript. They were advised by Dirk Beyer, Feng Cheng, Subodha Kumar, Young Ryu, Chelliah Sriskandarajah, Wulin Suo, Houmin Yan, Hanqin Zhang, and Qing Zhang on the technical problems of using LATEX. We also thank our wives and children—Andrea, Chantal, Anjuli, Dorothea, Allison, Emily, and Abigail—for their encouragement and understanding during the time-consuming task of preparing this revision.

Preface to Second Edition

xi

Finally, while we regret that lack of time and pressure of other duties prevented us from bringing out a second edition soon after the ﬁrst edition went out of print, we sincerely hope that the wait has been worthwhile. In spite of the numerous applications of optimal control theory which already have been made to areas of management science and economics, we continue to believe there is much more that remains to be done. We hope the present revision will rekindle interest in furthering such applications, and will enhance the continued development in the ﬁeld. Richardson, TX, USA Pittsburgh, PA, USA January 2000

Suresh P. Sethi Gerald L. Thompson

Preface to First Edition The purpose of this book is to exposit, as simply as possible, some recent results obtained by a number of researchers in the application of optimal control theory to management science. We believe that these results are very important and deserve to be widely known by management scientists, mathematicians, engineers, economists, and others. Because the mathematical background required to use this book is two or three semesters of calculus plus some diﬀerential equations and linear algebra, the book can easily be used to teach a course in the junior or senior undergraduate years or in the early years of graduate work. For this purpose, we have included numerous worked-out examples in the text, as well as a fairly large number of exercises at the end of each chapter. Answers to selected exercises are included in the back of the book. A solutions manual containing completely worked-out solutions to all of the 205 exercises is also available to instructors. The emphasis of the book is not on mathematical rigor, but on modeling realistic situations faced in business and management. For that reason, we have given in Chaps. 2 and 7 only heuristic proofs of the continuous and discrete maximum principles, respectively. In Chap. 3 we have summarized, as succinctly as we can, the most important model types and terminal conditions that have been used to model management problems. We found it convenient to put a summary of almost all the important management science models on two pages: see Tables 3.1 and 3.3. One of the fascinating features of optimal control theory is the extraordinarily wide range of its possible applications. We have tried to cover a wide variety of applications as follows: Chap. 4 covers ﬁnance; Chap. 5 considers production and inventory; Chap. 6 covers marketing; Chap. 8 treats machine maintenance and replacement; Chap. 9 deals with problems of optimal consumption of natural resources (renewable or exhaustible); and Chap. 10 discusses several economic applications. In Chap. 11 we treat some computational algorithms for solving optimal control problems. This is a very large and important area that needs more development. xiii

xiv

Preface to First Edition

Chapter 12 treats several more advanced topics of optimal control: diﬀerential games, distributed parameter systems, optimal ﬁltering, stochastic optimal control, and impulsive control. We believe that some of these models are capable of wider applications and further theoretical development. Finally, four appendixes cover either elementary material, such as diﬀerential equations, or advanced material, whose inclusion in the main text would spoil its continuity. Also at the end of the book is a bibliography of works actually cited in the text. While it is extensive, it is by no means an exhaustive bibliography of management science applications of optimal control theory. Several surveys of such applications, which contain many other important references, are cited. We have beneﬁted greatly during the writing of this book by having discussions with and obtaining suggestions from various colleagues and students. Our special thanks go to Gustav Feichtinger for his careful reading and suggestions for improvement of the entire book. Carl Norstr¨om contributed two examples to Chaps. 4 and 5 and made many suggestions for improvement. Jim Bookbinder used the manuscript for a course at the University of Toronto, and Tom Morton suggested some improvements for Chap. 5. The book has also beneﬁted greatly from various coauthors with whom we have done research over the years. Both of us also have received numerous suggestions for improvements from the students in our applied control theory courses taught during the past several years. We would like to express our gratitude to all these people for their help. The book has gone through several drafts, and we are greatly indebted to Eleanor Balocik and Rosilita Jones for their patience and careful typing. Although the applications of optimal control theory to management science are recent and many fascinating applications have already been made, we believe that much remains to be done. We hope that this book will contribute to the popularity of the area and will enhance future developments. Toronto, ON, Canada Pittsburgh, PA, USA August 1981

Suresh P. Sethi Gerald L. Thompson

Contents 1 What Is Optimal Control Theory? 1.1 Basic Concepts and Deﬁnitions . . . . . . . . . . . . . . 1.2 Formulation of Simple Control Models . . . . . . . . . . 1.3 History of Optimal Control Theory . . . . . . . . . . . 1.4 Notation and Concepts Used . . . . . . . . . . . . . . . 1.4.1 Diﬀerentiating Vectors and Matrices with Respect To Scalars . . . . . . . . . . . . . . . . . . . . . . 1.4.2 Diﬀerentiating Scalars with Respect to Vectors . 1.4.3 Diﬀerentiating Vectors with Respect to Vectors . 1.4.4 Product Rule for Diﬀerentiation . . . . . . . . . 1.4.5 Miscellany . . . . . . . . . . . . . . . . . . . . . . 1.4.6 Convex Set and Convex Hull . . . . . . . . . . . 1.4.7 Concave and Convex Functions . . . . . . . . . . 1.4.8 Aﬃne Function and Homogeneous Function of Degree k . . . . . . . . . . . . . . . . . . . . . . . 1.4.9 Saddle Point . . . . . . . . . . . . . . . . . . . . 1.4.10 Linear Independence and Rank of a Matrix . . . 1.5 Plan of the Book . . . . . . . . . . . . . . . . . . . . . . 2 The Maximum Principle: Continuous Time 2.1 Statement of the Problem . . . . . . . . . . . . . . . 2.1.1 The Mathematical Model . . . . . . . . . . . 2.1.2 Constraints . . . . . . . . . . . . . . . . . . . 2.1.3 The Objective Function . . . . . . . . . . . . 2.1.4 The Optimal Control Problem . . . . . . . . 2.2 Dynamic Programming and the Maximum Principle 2.2.1 The Hamilton-Jacobi-Bellman Equation . . . 2.2.2 Derivation of the Adjoint Equation . . . . . .

. . . . . . . .

. . . . . . . .

1 2 4 9 11 12 13 14 16 16 20 20 22 22 23 23 27 27 28 28 29 29 32 32 36 xv

xvi

CONTENTS 2.2.3 2.2.4 2.3 2.4 2.5

The Maximum Principle . . Economic Interpretations of Principle . . . . . . . . . . Simple Examples . . . . . . . . . . Suﬃciency Conditions . . . . . . . Solving a TPBVP by Using Excel .

. . the . . . . . . . .

. . . . . . . Maximum . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . .

39

. . . .

40 42 53 57

. . . .

. . . .

3 The Maximum Principle: Mixed Inequality Constraints 3.1 A Maximum Principle for Problems with Mixed Inequality Constraints . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Suﬃciency Conditions . . . . . . . . . . . . . . . . . . . 3.3 Current-Value Formulation . . . . . . . . . . . . . . . . 3.4 Transversality Conditions: Special Cases . . . . . . . . . 3.5 Free Terminal Time Problems . . . . . . . . . . . . . . . 3.6 Inﬁnite Horizon and Stationarity . . . . . . . . . . . . . 3.7 Model Types . . . . . . . . . . . . . . . . . . . . . . . . 4 The Maximum Principle: Pure State and Mixed Inequality Constraints 4.1 Jumps in Marginal Valuations . . . . . . . . . . . . . 4.2 The Optimal Control Problem with Pure and Mixed Constraints . . . . . . . . . . . . . . . . . . . . . . . 4.3 The Maximum Principle: Direct Method . . . . . . . 4.4 Suﬃciency Conditions: Direct Method . . . . . . . . 4.5 The Maximum Principle: Indirect Method . . . . . . 4.6 Current-Value Maximum Principle: Indirect Method . . . . . . . . . . . . . . . . . . . . 5 Applications to Finance 5.1 The Simple Cash Balance Problem . . . . . . . . 5.1.1 The Model . . . . . . . . . . . . . . . . . 5.1.2 Solution by the Maximum Principle . . . 5.2 Optimal Financing Model . . . . . . . . . . . . . 5.2.1 The Model . . . . . . . . . . . . . . . . . 5.2.2 Application of the Maximum Principle . . 5.2.3 Synthesis of Optimal Control Paths . . . 5.2.4 Solution for the Inﬁnite Horizon Problem

. . . . . . . .

. . . . . . . .

69 70 79 80 86 93 103 109

. .

125 127

. . . .

. . . .

129 132 136 137

. .

147

. . . . . . . .

159 160 160 161 164 165 167 170 180

. . . . . . . .

CONTENTS

xvii

6 Applications to Production and Inventory 6.1 Production-Inventory Systems . . . . . . . . . . . . . . . 6.1.1 The Production-Inventory Model . . . . . . . . . 6.1.2 Solution by the Maximum Principle . . . . . . . 6.1.3 The Inﬁnite Horizon Solution . . . . . . . . . . . 6.1.4 Special Cases of Time Varying Demands . . . . . 6.1.5 Optimality of a Linear Decision Rule . . . . . . . 6.1.6 Analysis with a Nonnegative Production Constraint . . . . . . . . . . . . . . . . . . . . . . 6.2 The Wheat Trading Model . . . . . . . . . . . . . . . . 6.2.1 The Model . . . . . . . . . . . . . . . . . . . . . 6.2.2 Solution by the Maximum Principle . . . . . . . 6.2.3 Solution of a Special Case . . . . . . . . . . . . . 6.2.4 The Wheat Trading Model with No Short-Selling 6.3 Decision Horizons and Forecast Horizons . . . . . . . . . 6.3.1 Horizons for the Wheat Trading Model with No Short-Selling . . . . . . . . . . . . . . . . . . 6.3.2 Horizons for the Wheat Trading Model with No Short-Selling and a Warehousing Constraint . . .

191 192 192 193 196 197 200

7 Applications to Marketing 7.1 The Nerlove-Arrow Advertising Model . . . . . . . . . 7.1.1 The Model . . . . . . . . . . . . . . . . . . . . 7.1.2 Solution by the Maximum Principle . . . . . . 7.1.3 Convex Advertising Cost and Relaxed Controls 7.2 The Vidale-Wolfe Advertising Model . . . . . . . . . . 7.2.1 Optimal Control Formulation for the Vidale-Wolfe Model . . . . . . . . . . . . . . . 7.2.2 Solution Using Green’s Theorem When Q Is Large . . . . . . . . . . . . . . . . . . . . 7.2.3 Solution When Q Is Small . . . . . . . . . . . . 7.2.4 Solution When T Is Inﬁnite . . . . . . . . . . .

. . . . .

225 226 226 228 232 235

.

236

. . .

237 245 247

. . . . .

259 259 260 262 267 268

8 The Maximum Principle: Discrete Time 8.1 Nonlinear Programming Problems . . . . . . . 8.1.1 Lagrange Multipliers . . . . . . . . . . . 8.1.2 Equality and Inequality Constraints . . 8.1.3 Constraint Qualiﬁcation . . . . . . . . . 8.1.4 Theorems from Nonlinear Programming

. . . . .

. . . . .

. . . . .

. . . . .

202 204 205 206 206 208 213 214 214

xviii 8.2

8.3

CONTENTS A Discrete Maximum Principle . . . . . . . . . . . 8.2.1 A Discrete-Time Optimal Control Problem 8.2.2 A Discrete Maximum Principle . . . . . . . 8.2.3 Examples . . . . . . . . . . . . . . . . . . . A General Discrete Maximum Principle . . . . . .

. . . . .

9 Maintenance and Replacement 9.1 A Simple Maintenance and Replacement Model . . . 9.1.1 The Model . . . . . . . . . . . . . . . . . . . 9.1.2 Solution by the Maximum Principle . . . . . 9.1.3 A Numerical Example . . . . . . . . . . . . . 9.1.4 An Extension . . . . . . . . . . . . . . . . . . 9.2 Maintenance and Replacement for a Machine Subject to Failure . . . . . . . . . . . . . 9.2.1 The Model . . . . . . . . . . . . . . . . . . . 9.2.2 Optimal Policy . . . . . . . . . . . . . . . . . 9.2.3 Determination of the Sale Date . . . . . . . . 9.3 Chain of Machines . . . . . . . . . . . . . . . . . . . 9.3.1 The Model . . . . . . . . . . . . . . . . . . . 9.3.2 Solution by the Discrete Maximum Principle 9.3.3 Special Case of Bang-Bang Control . . . . . . 9.3.4 Incorporation into the Wagner-Whitin Framework for a Complete Solution . . . . . 9.3.5 A Numerical Example . . . . . . . . . . . . . 10 Applications to Natural Resources 10.1 The Sole-Owner Fishery Resource Model . . 10.1.1 The Dynamics of Fishery Models . . 10.1.2 The Sole Owner Model . . . . . . . 10.1.3 Solution by Green’s Theorem . . . . 10.2 An Optimal Forest Thinning Model . . . . 10.2.1 The Forestry Model . . . . . . . . . 10.2.2 Determination of Optimal Thinning 10.2.3 A Chain of Forests Model . . . . . . 10.3 An Exhaustible Resource Model . . . . . . 10.3.1 Formulation of the Model . . . . . . 10.3.2 Solution by the Maximum Principle

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . .

. . . . .

269 269 270 272 276

. . . . .

. . . . .

283 284 284 285 287 289

. . . . . . . .

. . . . . . . .

290 291 293 296 297 297 299 301

. . . .

301 302

. . . . . . . . . . .

311 312 312 313 314 317 317 318 321 324 324 327

. . . . . . . . . . .

CONTENTS 11 Applications to Economics 11.1 Models of Optimal Economic Growth . . . . . . . 11.1.1 An Optimal Capital Accumulation Model 11.1.2 Solution by the Maximum Principle . . . 11.1.3 Introduction of a Growing Labor Force . . 11.1.4 Solution by the Maximum Principle . . . 11.2 A Model of Optimal Epidemic Control . . . . . . 11.2.1 Formulation of the Model . . . . . . . . . 11.2.2 Solution by Green’s Theorem . . . . . . . 11.3 A Pollution Control Model . . . . . . . . . . . . 11.3.1 Model Formulation . . . . . . . . . . . . . 11.3.2 Solution by the Maximum Principle . . . 11.3.3 Phase Diagram Analysis . . . . . . . . . . 11.4 An Adverse Selection Model . . . . . . . . . . . . 11.4.1 Model Formulation . . . . . . . . . . . . . 11.4.2 The Implementation Problem . . . . . . . 11.4.3 The Optimization Problem . . . . . . . . 11.5 Miscellaneous Applications . . . . . . . . . . . .

xix

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

335 335 336 336 338 339 343 343 344 346 347 348 349 352 352 353 354 360

. . . . . .

365 366 370 372 375 377 383

. . . .

385 386 387 388 388

. .

389

. . of . .

392

. . . . . . . . . . . . . . . . .

12 Stochastic Optimal Control 12.1 Stochastic Optimal Control . . . . . . . . . . . . . . . 12.2 A Stochastic Production Inventory Model . . . . . . . 12.2.1 Solution for the Production Planning Problem 12.3 The Sethi Advertising Model . . . . . . . . . . . . . . 12.4 An Optimal Consumption-Investment Problem . . . . 12.5 Concluding Remarks . . . . . . . . . . . . . . . . . . . 13 Diﬀerential Games 13.1 Two-Person Zero-Sum Diﬀerential Games . . . . . . 13.2 Nash Diﬀerential Games . . . . . . . . . . . . . . . . 13.2.1 Open-Loop Nash Solution . . . . . . . . . . . 13.2.2 Feedback Nash Solution . . . . . . . . . . . . 13.2.3 An Application to Common-Property Fishery Resources . . . . . . . . . . . . . . . . . . . . 13.3 A Feedback Nash Stochastic Diﬀerential Game in Advertising . . . . . . . . . . . . . . . . . . 13.4 A Feedback Stackelberg Stochastic Diﬀerential Game Cooperative Advertising . . . . . . . . . . . . . . . .

. . . .

395

xx

CONTENTS

A Solutions of Linear Diﬀerential Equations A.1 First-Order Linear Equations . . . . . . . . . . . . . . . A.2 Second-Order Linear Equations with Constant Coeﬃcients . . . . . . . . . . . . . . . . . . . . A.3 System of First-Order Linear Equations . . . . . . . . . A.4 Solution of Linear Two-Point Boundary Value Problems A.5 Solutions of Finite Diﬀerence Equations . . . . . . . . . A.5.1 Changing Polynomials in Powers of k into Factorial Powers of k . . . . . . . . . . . . . . . . A.5.2 Changing Factorial Powers of k into Ordinary Powers of k . . . . . . . . . . . . . . . . . . . . .

409 409 410 410 413 414 415 416

B Calculus of Variations and Optimal Control Theory 419 B.1 The Simplest Variational Problem . . . . . . . . . . . . 420 B.2 The Euler-Lagrange Equation . . . . . . . . . . . . . . . 421 B.3 The Shortest Distance Between Two Points on the Plane 424 B.4 The Brachistochrone Problem . . . . . . . . . . . . . . . 424 B.5 The Weierstrass-Erdmann Corner Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . 427 B.6 Legendre’s Conditions: The Second Variation . . . . . . 428 B.7 Necessary Condition for a Strong Maximum . . . . . . . . . . . . . . . . . . . . . . . . . . 429 B.8 Relation to Optimal Control Theory . . . . . . . . . . . 430 C An Alternative Derivation of the Maximum Principle 433 C.1 Needle-Shaped Variation . . . . . . . . . . . . . . . . . . 434 C.2 Derivation of the Adjoint Equation and the Maximum Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . 436 D Special Topics in Optimal Control D.1 The Kalman Filter . . . . . . . . . . . . . . . . . . . D.2 Wiener Process and Stochastic Calculus . . . . . . . D.3 The Kalman-Bucy Filter . . . . . . . . . . . . . . . . D.4 Linear-Quadratic Problems . . . . . . . . . . . . . . D.4.1 Certainty Equivalence or Separation Principle D.5 Second-Order Variations . . . . . . . . . . . . . . . . D.6 Singular Control . . . . . . . . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

441 441 444 447 448 451 452 454

CONTENTS D.7 Global Saddle Point Theorem . . . . . . . . . . . . . . . D.8 The Sethi-Skiba Points . . . . . . . . . . . . . . . . . . . D.9 Distributed Parameter Systems . . . . . . . . . . . . . .

xxi 456 458 460

E Answers to Selected Exercises

465

Bibliography

473

Index

547

List of Figures 1.1 1.2 1.3 1.4

The Brachistochrone problem . . . Illustration of left and right limits A concave function . . . . . . . . . An illustration of a saddle point . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

9 18 21 23

2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8

An optimal path in the state-time space . . . . . . . . Optimal state and adjoint trajectories for Example 2.2 Optimal state and adjoint trajectories for Example 2.3 Optimal trajectories for Examples 2.4 and 2.5 . . . . . Optimal control for Example 2.6 . . . . . . . . . . . . The ﬂowchart for Example 2.8 . . . . . . . . . . . . . Solution of TPBVP by excel . . . . . . . . . . . . . . . Water reservoir of Exercise 2.18 . . . . . . . . . . . . .

. . . . . . . .

34 44 46 48 53 58 60 63

3.1 3.2

State and adjoint trajectories in Example 3.4 . . . . . . Minimum time optimal response for Example 3.6 . . . .

93 101

4.1 4.2 4.3 4.4 4.5

Feasible state space and optimal state trajectory for Examples 4.1 and 4.4 . . . . . . . . . . . . . . State and adjoint trajectories in Example 4.3 . . Adjoint trajectory for Example 4.4 . . . . . . . . Two-reservoir system of Exercise 4.8 . . . . . . . Feasible space for Exercise 4.28 . . . . . . . . . .

. . . . .

. . . . .

. . . . .

. . . . .

128 143 147 151 157

5.1 5.2 5.3 5.4 5.5 5.6

Optimal policy shown in (λ1 , λ2 ) space . Optimal policy shown in (t, λ2 /λ1 ) space Case A: g ≤ r . . . . . . . . . . . . . . . Case B: g > r . . . . . . . . . . . . . . . Optimal path for case A: g ≤ r . . . . . Optimal path for case B: g > r . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

163 164 169 170 174 179

. . . . . .

. . . .

. . . . . .

. . . .

. . . . . .

. . . .

. . . . . .

. . . .

. . . . . .

. . . .

. . . .

xxiii

xxiv 5.7 5.8

LIST OF FIGURES Solution for Exercise 5.4 . . . . . . . . . . . . . . . . . . Adjoint trajectories for Exercise 5.5 . . . . . . . . . . .

186 187

Solution of Example 6.1 with I0 = 10 . . . . . . . . . . . Solution of Example 6.1 with I0 = 50 . . . . . . . . . . . Solution of Example 6.1 with I0 = 30 . . . . . . . . . . . Optimal production rate and inventory level with diﬀerent initial inventories . . . . . . . . . . . . . . . . . . . . . . 6.5 The price trajectory (6.56) . . . . . . . . . . . . . . . . . 6.6 Adjoint variable, optimal policy and inventory in the wheat trading model . . . . . . . . . . . . . . . . . . . . 6.7 Adjoint trajectory and optimal policy for the wheat trading model . . . . . . . . . . . . . . . . . . . . . . . . . . 6.8 Decision horizon and optimal policy for the wheat trading model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.9 Optimal policy and horizons for the wheat trading model with no short-selling and a warehouse constraint . . . . 6.10 Optimal policy and horizons for Example 6.3 . . . . . . 6.11 Optimal policy and horizons for Example 6.4 . . . . . .

199 199 200

6.1 6.2 6.3 6.4

7.1 7.2

204 207 209 212 215 216 218 219

7.12 7.13 7.14

Optimal policies in the Nerlove-Arrow model . . . . . . A case of a time-dependent turnpike and the nature of optimal control . . . . . . . . . . . . . . . . . . . . . . . A near-optimal control of problem (7.15) . . . . . . . . . Feasible arcs in (t, x)-space . . . . . . . . . . . . . . . . Optimal trajectory for Case 1: x0 ≤ xs and xT ≤ xs . . Optimal trajectory for Case 2: x0 < xs and xT > xs . . Optimal trajectory for Case 3: x0 > xs and xT < xs . . Optimal trajectory for Case 4: x0 > xs and xT > xs . . Optimal trajectory (solid lines) . . . . . . . . . . . . . . Optimal trajectory when T is small in Case 1: x0 < xs and xT > xs . . . . . . . . . . . . . . . . . . . . . . . . . Optimal trajectory when T is small in Case 2: x0 > xs and xT > xs . . . . . . . . . . . . . . . . . . . . . . . . . Optimal trajectory for Case 2 of Theorem 7.1 for Q = ∞ Optimal trajectories for x(0) < x ˆ . . . . . . . . . . . . . Optimal trajectory for x(0) > x ˆ . . . . . . . . . . . . . .

244 244 249 250

8.1 8.2

Shortest distance from point (2,2) to the semicircle . . . Graph of Example 8.5 . . . . . . . . . . . . . . . . . . .

266 267

7.3 7.4 7.5 7.6 7.7 7.8 7.9 7.10 7.11

230 231 233 238 240 241 241 242 243 243

LIST OF FIGURES

xxv

8.3 8.4

Discrete-time conventions . . . . . . . . . . . . . . . . . ∗ Optimal state xk and adjoint λk . . . . . . . . . . . . .

270 275

9.1 9.2

Optimal maintenance and machine resale value . . . . . Sat function optimal control . . . . . . . . . . . . . . . .

289 291

10.1 Optimal policy for the sole owner ﬁshery model . . . . . 10.2 Singular usable timber volume x ¯(t) . . . . . . . . . . . . 10.3 Optimal thinning u∗ (t) and timber volume x∗ (t) for the ¯(t0 ) . . . . . . . . . . forest thinning model when x0 < x 10.4 Optimal thinning u∗ (t) and timber volume x∗ (t) for the chain of forests model when T > tˆ . . . . . . . . . . . . 10.5 Optimal thinning and timber volume x∗ (t) for the chain of forests model when T ≤ tˆ . . . . . . . . . . . . . . . . 10.6 The demand function . . . . . . . . . . . . . . . . . . . . 10.7 The proﬁt function . . . . . . . . . . . . . . . . . . . . . 10.8 Optimal price trajectory for T ≥ T¯ . . . . . . . . . . . . 10.9 Optimal price trajectory for T < T¯ . . . . . . . . . . . .

316 320

11.1 11.2 11.3 11.4 11.5 11.6 11.7

. . . . . . .

340 346 347 348 351 358 359

12.1 A sample path of optimal production rate It∗ with I0 = x0 > 0 and B > 0 . . . . . . . . . . . . . . . . . . .

374

Phase diagram for the optimal growth model . Optimal trajectory when xT > xs . . . . . . . . Optimal trajectory when xT < xs . . . . . . . Food output function . . . . . . . . . . . . . . . Phase diagram for the pollution control model . Violation of the monotonicity constraint . . . . Bunching and ironing . . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

320 322 323 324 326 329 330

13.1 A sample path of optimal market share trajectories . . . 13.2 Optimal subsidy rate vs. (a) Retailer’s margin and (b) Manufacturer’s margin . . . . . . . . . . . . . . . . . . .

396

B.1 Examples of admissible functions for the problem . . . . B.2 Variation about the solution function . . . . . . . . . . . B.3 A broken extremal with corner at τ . . . . . . . . . . . .

420 421 428

404

xxvi

LIST OF FIGURES

C.1 Needle-shaped variation . . . . . . . . . . . . . . . . . . C.2 Trajectories x∗ (t) and x(t) in a one-dimensional case . .

434 434

D.1 Phase diagram for system (D.73) . . . . . . . . . . . . . D.2 Region D with boundaries Γ1 and Γ2 . . . . . . . . . . .

457 461

List of Tables 1.1 1.2 1.3

The production-inventory model of Example 1.1 . . . . The advertising model of Example 1.2 . . . . . . . . . . The consumption model of Example 1.3 . . . . . . . . .

3.1 3.2 3.3

Summary of the transversality conditions State trajectories and switching curves . . Objective, state, and adjoint equations for types . . . . . . . . . . . . . . . . . . . . .

4 6 8

. . . . . . . . . . . . . . . . various model . . . . . . . .

89 100

Characterization of optimal controls with c < 1 . . . . .

168

13.1 Optimal feedback Stackelberg solution . . . . . . . . . .

403

A.1 Homogeneous solution forms for Eq. (A.5) . . . . . . . . A.2 Particular solutions for Eq. (A.5) . . . . . . . . . . . . .

411 411

5.1

111

xxvii

Chapter 1

What Is Optimal Control Theory? Many management science applications involve the control of dynamic systems, i.e., systems that evolve over time. They are called continuoustime systems or discrete-time systems depending on whether time varies continuously or discretely. We will deal with both kinds of systems in this book, although the main emphasis will be on continuous-time systems. Optimal control theory is a branch of mathematics developed to ﬁnd optimal ways to control a dynamic system. The purpose of this book is to give an elementary introduction to the mathematical theory, and then apply it to a wide variety of diﬀerent situations arising in management science. We have deliberately kept the level of mathematics as simple as possible in order to make the book accessible to a large audience. The only mathematical requirements for this book are elementary calculus, including partial diﬀerentiation, some knowledge of vectors and matrices, and elementary ordinary and partial diﬀerential equations. The last topic is brieﬂy covered in Appendix A. Chapter 12 on stochastic optimal control also requires some concepts in stochastic calculus, which are introduced at the beginning of that chapter. The principle management science applications discussed in this book come from the following areas: ﬁnance, economics, production and inventory, marketing, maintenance and replacement, and the consumption of natural resources. In each major area we have formulated one or more simple models followed by a more complicated model. The reader may

© Springer Nature Switzerland AG 2019 S. P. Sethi, Optimal Control Theory, https://doi.org/10.1007/978-3-319-98237-3 1

1

2

1. What Is Optimal Control Theory?

wish at ﬁrst to cover only the simpler models in each area to get an idea of what could be accomplished with optimal control theory. Later, the reader may wish to go into more depth in one or more of the applied areas. Examples are worked out in most of the chapters to facilitate the exposition. At the end of each chapter, we have listed exercises that the reader should solve for deeper understanding of the material presented in the chapter. Hints are supplied with some of the exercises. Answers to selected exercises are given in Appendix E.

1.1

Basic Concepts and Deﬁnitions

We will use the word system as a primitive term in this book. The only property that we require of a system is that it is capable of existing in various states. Let the (real) variable x(t) be the state variable of the system at time t ∈ [0, T ], where T > 0 is a speciﬁed time horizon for the system under consideration. For example, x(t) could measure the inventory level at time t, the amount of advertising goodwill at time t, or the amount of unconsumed wealth or natural resources at time t. We assume that there is a way of controlling the state of the system. Let the (real) variable u(t) be the control variable of the system at time t. For example, u(t) could be the production rate at time t, the advertising rate at time t, etc. Given the values of the state variable x(t) and the control variable u(t) at time t, the state equation, a diﬀerential equation, x(t) ˙ = f (x(t), u(t), t),

x(0) = x0 ,

(1.1)

speciﬁes the instantaneous rate of change in the state variable, where x(t) ˙ is a commonly used notation for dx(t)/dt, f is a given function of x, u, and t, and x0 is the initial value of the state variable. If we know the initial value x0 and the control trajectory, i.e., the values of u(t) over the whole time interval 0 ≤ t ≤ T, then we can integrate (1.1) to get the state trajectory, i.e., the values of x(t) over the same time interval. We want to choose the control trajectory so that the state and control trajectories maximize the objective functional, or simply the objective function, T

F (x(t), u(t), t)dt + S[x(T ), T ].

J= 0

(1.2)

1.1. Basic Concepts and Deﬁnitions

3

In (1.2), F is a given function of x, u, and t, which could measure the beneﬁt minus the cost of advertising, the utility of consumption, the negative of the cost of inventory and production, etc. Also in (1.2), the function S gives the salvage value of the ending state x(T ) at time T. The salvage value is needed so that the solution will make “good sense” at the end of the horizon. Usually the control variable u(t) will be constrained. We indicate this as u(t) ∈ Ω(t), t ∈ [0, T ], (1.3) where Ω(t) is the set of feasible values for the control variable at time t. Optimal control problems involving (1.1), (1.2), and (1.3) will be treated in Chap. 2. In Chap. 3, we will replace (1.3) by inequality constraints involving control variables. In addition, we will allow these constraints to depend on state variables. These are called mixed inequality constraints and written as g(x(t), u(t), t) ≥ 0, t ∈ [0, T ] , (1.4) where g is a given function of u, t, and possibly x. In addition, there may be constraints involving only state variables, but not control variables. These are written as h(x(t), t) ≥ 0,

t ∈ [0, T ],

(1.5)

where h is a given function of x and t. Such constraints are the most diﬃcult to deal with, and are known as pure state inequality constraints. Problems involving (1.1), (1.2), (1.4), and (1.5) will be treated in Chap. 4. Finally, we note that all of the imposed constraints limit the values that the terminal state x(T ) may take. We denote this by saying x(T ) ∈ X,

(1.6)

where X is called the reachable set of the state variable at time T. Note that X depends on the initial value x0 . Here X is the set of possible terminal values that can be reached when x(t) and u(t) obey imposed constraints. Although the above description of the control problem may seem abstract, you will ﬁnd that in each speciﬁc application, the variables and parameters will have speciﬁc meanings that make them easy to understand and remember. The examples that follow will illustrate this point.

4

1. What Is Optimal Control Theory?

1.2

Formulation of Simple Control Models

We now formulate three simple models chosen from the areas of production, advertising, and economics. Our only objective here is to identify and interpret in these models each of the variables and functions described in the previous section. The solutions for each of these models will be given in detail in later chapters. Example 1.1 A Production-Inventory Model. The various quantities that deﬁne this model are summarized in Table 1.1 for easy comparison with the other models that follow. Table 1.1: The production-inventory model of Example 1.1 State variable

I(t) = Inventory level

Control variable

P (t) = Production rate

State equation

˙ = P (t) − S(t), I(0) = I0 I(t) T Maximize J = −[h(I(t)) + c(P (t))]dt

Objective function

0

State constraint

I(t) ≥ 0

Control constraints

0 ≤ Pmin ≤ P (t) ≤ Pmax

Terminal condition

I(T ) ≥ Imin

Exogenous functions

S(t) = Demand rate h(I) = Inventory holding cost c(P ) = Production cost

Parameters

T = Terminal time Imin = Minimum ending inventory Pmin = Minimum possible production rate Pmax = Maximum possible production rate I0 = Initial inventory level

1.2. Formulation of Simple Control Models

5

We consider the production and inventory storage of a given good, such as steel, in order to meet an exogenous demand. The state variable I(t) measures the number of tons of steel that we have on hand at time t ∈ [0, T ]. There is an exogenous demand rate S(t) tons of steel per day at time t ∈ [0, T ], and we must choose the production rate P (t) tons of steel per day at time t ∈ [0, T ]. Given the initial inventory of I0 tons of steel on hand at t = 0, the state equation ˙ = P (t) − S(t) I(t) describes how the steel inventory changes over time. Since h(I) is the cost of holding inventory I in dollars per day, and c(P ) is the cost of producing steel at rate P, also in dollars per day, the objective function is to maximize the negative of the sum of the total holding and production costs over the period of T days. Of course, maximizing the negative sum is the same as minimizing the sum of holding and production costs. The state variable constraint, I(t) ≥ 0, is imposed so that the demand is satisﬁed for all t. In other words, backlogging of demand is not permitted. (An alternative formulation is to make h(I) become very large when I becomes negative, i.e., to impose a stockout penalty cost.) The control constraints keep the production rate P (t) between a speciﬁed lower bound Pmin and a speciﬁed upper bound Pmax . Finally, the terminal constraint I(T ) ≥ Imin is imposed so that the terminal inventory is at least Imin . The statement of the problem is lengthy because of the number of variables, functions, and parameters which are involved. However, with the production and inventory interpretations as given, it is not diﬃcult to see the reasons for each condition. In Chap. 6, various versions of this model will be solved in detail. In Sect. 12.2, we will deal with a stochastic version of this model. Example 1.2 An Advertising Model. The various quantities that deﬁne this model are summarized in Table 1.2. We consider a special case of the Nerlove-Arrow advertising model which will be discussed in detail in Chap. 7. The problem is to determine the rate at which to advertise a product at each time t. Here the state variable is advertising goodwill, G(t), which measures how well the product is known at time t. We assume that there is a forgetting coeﬃcient δ, which measures the rate at which customers tend to forget the product.

6

1. What Is Optimal Control Theory?

To counteract forgetting, advertising is carried out at a rate measured by the control variable u(t). Hence, the state equation is ˙ G(t) = u(t) − δG(t), with G(0) = G0 > 0 specifying the initial goodwill for the product. Table 1.2: The advertising model of Example 1.2 State variable

G(t) = Advertising goodwill

Control variable

u(t) = Advertising rate

State equation

˙ G(t) = u(t) − δG(t), G(0) = G0 ∞ −ρt Maximize J = [π(G(t)) − u(t)]e dt

Objective function

0

···

State constraint Control constraints

0 ≤ u(t) ≤ Q ···

Terminal condition Exogenous function

π(G) = Gross proﬁt rate

Parameters

δ = Goodwill decay constant ρ = Discount rate Q = Upper bound on advertising rate G0 = Initial goodwill level

The objective function J requires special discussion. Note that the integral deﬁning J is from time t = 0 to time t = ∞; we will later call a problem having an upper time limit of ∞, an inﬁnite horizon problem. Because of this upper limit, the integrand of the objective function includes the discount factor e−ρt , where ρ > 0 is the (constant) discount rate. Without this discount factor, the integral would (in most cases) diverge to inﬁnity. Hence, we will see that such a discount factor is an essential part of inﬁnite horizon models. The rest of the integrand in the objective function consists of the gross proﬁt rate π(G(t)), which

1.2. Formulation of Simple Control Models

7

results from the goodwill level G(t) at time t less the cost of advertising assumed to be proportional to u(t) (proportionality factor = 1); thus π(G(t)) − u(t) is the net proﬁt rate at time t. Also [π(G(t)) − u(t)]e−ρt is the net proﬁt rate at time t discounted to time 0, i.e., the present value of the time t proﬁt rate. Hence, J can be interpreted as the total value of discounted future proﬁts, and is the quantity we are trying to maximize. There are control constraints 0 ≤ u(t) ≤ Q, where Q is the upper bound on the advertising rate. However, there is no state constraint. It can be seen from the state equation and the control constraints that the goodwill G(t) in fact never becomes negative. You will ﬁnd it instructive to compare this model with the previous one and note the similarities and diﬀerences between the two. Example 1.3 A Consumption Model. Rich Rentier plans to retire at age 65 with a lump sum pension of W0 dollars. Rich estimates his remaining life span to be T years. He wants to consume his wealth during these T retirement years, beginning at the age of 65, and leave a bequest to his heirs in a way that will maximize his total utility of consumption and bequest. Since he does not want to take investment risks, Rich plans to put his money into a savings account that pays interest at a continuously compounded rate of r. In order to formulate Rich’s optimization problem, let t = 0 denote the time when he turns 65 so that his retirement period can be denoted by the interval [0, T ]. If we let the state variable W (t) denote Rich’s wealth and the control variable C(t) ≥ 0 denote his rate of consumption at time t ∈ [0, T ], it is easy to see that the state equation is ˙ (t) = rW (t) − C(t), W with the initial condition W (0) = W0 > 0. It is reasonable to require that W (t) ≥ 0 and C(t) ≥ 0, t ∈ [0, T ]. Letting U (C) be the utility function of consumption C and B(W ) be the bequest function of leaving a bequest of amount W at time T, we see that the problem can be stated as an optimal control problem with the variables, equations, and constraints shown in Table 1.3. Note that the objective function has two parts: ﬁrst the integral of the discounted utility of consumption from time 0 to time T with ρ as the discount rate; and second the bequest function e−ρT B(W ), which measures Rich’s discounted utility of leaving an estate W to his heirs

8

1. What Is Optimal Control Theory?

at time T. If he has no heirs and does not care about charity, then B(W ) = 0. However, if he has heirs or a favorite charity to whom he wishes to leave money, then B(W ) measures the strength of his desire to leave an estate of amount W. The nonnegativity constraints on state and control variables are obviously natural requirements that must be imposed. You will be asked to solve this problem in Exercise 2.1 after you have learned the maximum principle in the next chapter. Moreover, a stochastic extension of the consumption problem, known as a consumption/investment problem, will be discussed in Sect. 12.4.

Table 1.3: The consumption model of Example 1.3 State variable

W (t) = Wealth

Control variable

C(t) = Consumption rate

State equation

˙ (t) = rW (t) − C(t), W (0) = W0 W T Max J = U (C(t))e−ρt dt + B(W (T ))e−ρT

Objective function

0

State constraint

W (t) ≥ 0

Control constraint

C(t) ≥ 0 ···

Terminal condition Exogenous Functions Parameters

U (C) = Utility of consumption B(W ) = Bequest function T = Terminal time W0 = Initial wealth ρ = Discount rate r = Interest rate

1.3. History of Optimal Control Theory

1.3

9

History of Optimal Control Theory

Optimal control theory is an extension of the calculus of variations (see Appendix B), so we discuss the history of the latter ﬁrst. The creation of the calculus of variations occurred almost immediately after the formalization of calculus by Newton and Leibniz in the seventeenth century. An important problem in calculus is to ﬁnd an argument of a function at which the function takes on its maximum or minimum. The extension of this problem posed in the calculus of variations is to ﬁnd a function which maximizes or minimizes the value of an integral or functional of that function. As might be expected, the extremum problem in the calculus of variations is much harder than the extremum problem in diﬀerential calculus. Euler and Lagrange are generally considered to be the founders of the calculus of variations. Newton, Legendre, and the Bernoulli brothers also contributed much to the early development of the ﬁeld.

Figure 1.1: The Brachistochrone problem A celebrated problem ﬁrst solved using the calculus of variations was the path of least time or the Brachistochrone problem. The problem is illustrated in Fig. 1.1. It involves ﬁnding the shape of a curve Γ connecting the two points A and B in the vertical plane with the property that a bead sliding along the curve under the inﬂuence of gravity will move from A to B in the shortest possible time. The problem was posed

10

1. What Is Optimal Control Theory?

by Johann Bernoulli in 1696, and it played an important part in the development of calculus of variations. It was solved by Johann Bernoulli, Jakob Bernoulli, Newton, Leibnitz, and L’Hˆ opital. In Sect. B.4, we provide a solution to the Brachistochrone problem by using what is known as the Euler-Lagrange equation, stated in Sect. B.2, and show that the shape of the solution curve is represented by a cycloid. In the nineteenth and early twentieth centuries, many mathematicians contributed to the calculus of variations; these include Hamilton, Jacobi, Bolza, Weierstrass, Carath´eodory, and Bliss. Converting calculus of variations problems into control theory problems requires one more conceptual step—the addition of control variables to the state equations. Isaacs (1965) made such an extension in twoperson pursuit-evasion games in the period 1948–1955. Bellman (1957) made a similar extension with the idea of dynamic programming. Modern control theory began with the publication (in Russian in 1961 and English in 1962) of the book, The Mathematical Theory of Optimal Processes, by Pontryagin et al. (1962). Well-known American mathematicians associated with the maximum principle include Valentine, McShane, Hestenes, Berkovitz, and Neustadt. The importance of the book by Pontryagin et al. lies not only in a rigorous formulation of a calculus of variations problem with constrained control variables, but also in the proof of the maximum principle for optimal control problems. See Pesch and Bulirsch (1994) and Pesch and Plail (2009) for historical perspectives on the topics of the calculus of variations, dynamic programming, and optimal control. The maximum principle permits the decoupling of the dynamic problem over time, using what are known as adjoint variables or shadow prices, into a series of problems, each of which holds at a single instant of time. The optimal solution of the instantaneous problems can be shown to give the optimal solution to the overall problem. In this book we will be concerned principally with the application of the maximum principle in its various forms to ﬁnd the solutions of a wide variety of applied problems in management science and economics. It is hoped that the reader, after reading some of these problems and their solutions, will appreciate, as we do, the importance of the maximum principle. Some important books and surveys of the applications of the maximum principle to management science and economics are Con-

1.4. Notation and Concepts Used

11

nors and Teichroew (1967), Arrow and Kurz (1970), Hadley and Kemp (1971), Bensoussan et al. (1974), St¨ oppler (1975), Clark (1976), Sethi (1977a, 1978a), Tapiero (1977, 1988), Wickwire (1977), Bookbinder and Sethi (1980), Lesourne and Leban (1982), Tu (1984), Feichtinger and Hartl (1986), Carlson and Haurie (1987b), Seierstad and Sydsæter (1987), Erickson (2003), L´eonard and Long (1992), Kamien and Schwartz (1992), Van Hilten et al. (1993), Feichtinger et al. (1994a), Maimon et al. (1998), Dockner et al. (2000), Caputo (2005), Grass et al. (2008), and Bensoussan (2011). Nevertheless, we have included in our bibliography many works of interest.

1.4

Notation and Concepts Used

In order to make the book readable, we will adopt the following notation which will hold throughout the book. In addition, we will deﬁne some important concepts that are required, including those of concave, convex and aﬃne functions, and saddle points. We use the symbol “=” to mean “is equal to” or “is deﬁned to be equal to” or “is identically equal to” depending on the context. The symbol “:=” means “is deﬁned to be equal to,” the symbol “≡” means “is identically equal to,” and the symbol “≈” means “is approximately equal to.” The double arrow “⇒” means “implies,” “∀” means “for all,” and “∈” means “is a member of.” The symbol 2 indicates the end of a proof. Let y be an n-component column vector and z be an m-component row vector, i.e., ⎡ ⎤ ⎢ ⎢ ⎢ ⎢ ⎢ y =⎢ ⎢ ⎢ ⎢ ⎣

y1 ⎥ ⎥ ⎥ y2 ⎥ ⎥ ⎥ = (y1 , . . . , yn )T and z = (z1 , . . . , zm ), .. ⎥ . ⎥ ⎥ ⎦ yn

where the superscript T on a vector (or, a matrix) denotes the transpose of the vector (or, the matrix). At times, when convenient and not confusing, we will use the superscript for the transpose operation. If y and

12

1. What Is Optimal Control Theory?

z are functions of time t, a scalar, then the time derivatives y˙ := dy/dt and z˙ := dz/dt are deﬁned as y˙ =

dz dy = (y˙ 1 , · · · , y˙ n )T and z˙ = = (z˙1 , . . . , z˙m ), dt dt

where y˙ i and z˙j denote the time derivatives dyi /dt and dzj /dt, respectively. When n = m, we can deﬁne the inner product zy = Σni=1 zi yi .

(1.7)

More generally, if ⎡ ⎢ ⎢ ⎢ ⎢ ⎢ A = {aij } = ⎢ ⎢ ⎢ ⎢ ⎣

⎤ a11

a12

a21

a22

.. .

.. .

am1 am2

···

a1k ⎥ ⎥ ⎥ · · · a2k ⎥ ⎥ ⎥ .. ⎥ ··· . ⎥ ⎥ ⎦ · · · amk

is an m × k matrix and B = {bij } is a k × n matrix, we deﬁne the matrix product C = {cij } = AB, which is an m × n matrix with components cij = Σkr=1 air brj .

(1.8)

Let E k denote the k-dimensional Euclidean space. Its elements are k-component vectors, which may be either row or column vectors, depending on the context. Thus in (1.7), y ∈ E n is a column vector and z ∈ E m is a row vector. Next, in Sects. 1.4.1–1.4.4, we provide the notation for multivariate diﬀerentiation. Needless to say, the functions introduced are assumed to be appropriately diﬀerentiable for their derivatives being deﬁned.

1.4.1

Diﬀerentiating Vectors and Matrices with Respect To Scalars

Let f : E 1 → E k be a k-dimensional function of a scalar variable t. If f is a row vector, then we deﬁne df = ft = (f1t , f2t , · · · , fkt ), a row vector. dt

1.4. Notation and Concepts Used

13

We will also use the notation f = (f1 , f2 , · · · , fk ) and f (t) in place of ft . If f is a column vector, then ⎡ ⎤ ⎢ ⎢ ⎢ ⎢ df ⎢ = ft = ⎢ ⎢ dt ⎢ ⎢ ⎣

f1t ⎥ ⎥ ⎥ f2t ⎥ ⎥ ⎥ = (f1t , f2t , · · · , fkt )T , a column vector. .. ⎥ . ⎥ ⎥ ⎦ fkt

Once again, f (t) may also be written as f or f (t). A similar rule applies if a matrix function is diﬀerentiated with respect to a scalar. ⎡ ⎤ 2 2t + 3 ⎥ ⎢ t Example 1.4 Let f (t) = ⎣ ⎦ . Find ft . 1/t e3t ⎡

⎤

2 ⎥ ⎢ 2t Solution ft = ⎣ ⎦. 3e3t −1/t2

1.4.2

Diﬀerentiating Scalars with Respect to Vectors

If F (y, z) is a scalar function deﬁned on E n ×E m with y an n-dimensional column vector and z an m-dimensional row vector, then the gradients Fy and Fz are deﬁned, respectively, as Fy = (Fy1 , · · · , Fyn ), a row vector,

(1.9)

Fz = (Fz1 , · · · , Fzm ), a row vector,

(1.10)

and where Fyi and Fzj denote the partial derivatives with respect to the subscripted variables. Thus, we always deﬁne the gradient with respect to a row or column vector as a row vector. Alternatively, Fy and Fz are also denoted as ∇y F and ∇z F, respectively. In this notation, if F is a function of y only or z only, then the subscript can be dropped and the gradient of F can be written simply as ∇F.

14

1. What Is Optimal Control Theory?

Example 1.5 Let F (y, z) = y1 2 y3 z2 + 3y2 ln z1 + y1 y2 , where y = (y1 , y2 , y3 )T and z = (z1 , z2 ). Obtain Fy and Fz . Solution Fy = (Fy1 , Fy2 , Fy3 ) = (2y1 y3 z2 + y2 , 3 ln z1 + y1 , y1 2 z2 ) and Fz = (Fz1 , Fz2 ) = (3y2 /z1 , y1 2 y3 ).

1.4.3

Diﬀerentiating Vectors with Respect to Vectors

If f : E n × E m → E k is a k-dimensional vector function, f either row or column, i.e., f = (f1 , · · · , fk ) or f = (f1 , · · · , fk )T , where each component fi = fi (y, z) depends on the column vector y ∈ E n and the row vector z ∈ E m , then fz will denote the k × m matrix ⎡ ⎤ ⎢ ∂f1 /∂z1 , ∂f1 /∂z2 , ⎢ ⎢ ⎢ ∂f2 /∂z1 , ∂f2 /∂z2 , ⎢ fz = ⎢ .. .. ⎢ ⎢ . . ⎢ ⎣ ∂fk /∂z1 , ∂fk /∂z2 ,

· · · ∂f1 /∂zm ⎥ ⎥ ⎥ · · · ∂f2 /∂zm ⎥ ⎥ ⎥ = {∂fi /∂zj }, .. ⎥ ⎥ ··· . ⎥ ⎦ · · · ∂fk /∂zm

and fy will denote the k × n matrix ⎡ ⎢ ∂f1 /∂y1 ∂f1 /∂y2 ⎢ ⎢ ⎢ ∂f2 /∂y1 ∂f2 /∂y2 ⎢ fy = ⎢ .. .. ⎢ ⎢ . . ⎢ ⎣ ∂fk /∂y1 ∂fk /∂y2

(1.11)

⎤

· · · ∂f1 /∂yn ⎥ ⎥ ⎥ · · · ∂f2 /∂yn ⎥ ⎥ ⎥ = {∂fi /∂yj }. .. ⎥ ⎥ ··· . ⎥ ⎦ · · · ∂fk /∂yn

(1.12)

Matrices fz and fy are known as Jacobian matrices. It should be emphasized that the rule of deﬁning a Jacobian does not depend on the row or column nature of the function or its arguments. Thus, fz = (f T )z = fz T = (f T )z T . Example 1.6 Let f : E 3 × E 2 → E 3 be deﬁned by f (y, z) = (y1 2 y3 z2 + 3y2 ln z1 , z1 z2 2 y3 , z1 y1 + z2 y2 )T with y = (y1 , y2 , y3 )T and z = (z1 , z2 ). Obtain fz and fy .

1.4. Notation and Concepts Used

15 ⎤

⎡

Solution.

2y

⎢ 3y2 /z1 y1 3 ⎢ ⎢ fz = ⎢ z2 2 y3 2z1 z2 y3 ⎢ ⎣ y1 y2

⎥ ⎥ ⎥ ⎥, ⎥ ⎦

⎡

⎤ 2z

⎢ 2y1 y3 z2 3 ln z1 y1 2 ⎢ ⎢ fy = ⎢ 0 0 z 1 z2 2 ⎢ ⎣ z1 z2 0

⎥ ⎥ ⎥ ⎥. ⎥ ⎦

Applying the rule (1.11) to Fy in (1.9), we obtain Fyz = (Fy )z to be the n × m matrix ⎡ ⎤

Fyz

⎢ Fy1 z1 ⎢ ⎢ ⎢ Fy2 z1 ⎢ =⎢ ⎢ .. ⎢ . ⎢ ⎣ F yn z 1

Fy1 z2 Fy2 z2 .. . Fyn z2

· · · Fy1 zm ⎥ ⎥ ⎥ · · · Fy2 zm ⎥ ∂2F ⎥ . ⎥= .. ⎥ ∂yi ∂zj ⎥ ··· . ⎥ ⎦ · · · Fyn zm

(1.13)

Applying the rule (1.12) to Fz in (1.10), we obtain Fzy = (Fz )y to be the m × n matrix ⎡ ⎤

Fzy

⎢ Fz1 y1 ⎢ ⎢ ⎢ Fz2 y1 ⎢ =⎢ .. ⎢ ⎢ . ⎢ ⎣ Fz m y 1

Fz1 y2 Fz2 y2 .. . Fzm y2

···

Fz1 yn ⎥ ⎥ ⎥ · · · Fz2 yn ⎥ ∂2F ⎥ . ⎥= .. ⎥ ∂zi ∂yj ⎥ ··· . ⎥ ⎦ · · · Fzm yn

(1.14)

Note that if F (y, z) is twice continuously diﬀerentiable, then we also have Fzy = (Fyz )T . Example 1.7 Obtain Fyz and Fzy for F (y, z) speciﬁed in Example 1.5. Since the given F (y, z) is twice continuously diﬀerentiable, check also that Fzy = (Fyz )T .

16

1. What Is Optimal Control Theory?

Solution. Applying rule (1.11) to Fy obtained in Example 1.5 and rule (1.12) to Fz obtained in Example 1.5, we have, respectively, ⎡ ⎤ ⎡ ⎤ 2y1 y3 ⎥ ⎢ 0 ⎢ ⎥ 3/z1 0 ⎥ ⎢ ⎢ 0 ⎥ Fyz = ⎢ 3/z1 ⎥ and Fzy = ⎣ ⎦. 0 ⎢ ⎥ ⎣ ⎦ 0 y1 2 2y1 y3 0 y1 2 Also, it is easily seen from these matrices that Fzy = (Fyz )T .

1.4.4

Product Rule for Diﬀerentiation

Let g be an n-component row vector function and f be an n-component column vector function of an n-component vector x. Then in Exercise 1.9, you are asked to show that (gf )x = gfx + f T gx = gfx + f T (g T )x .

(1.15)

In Exercise 1.10, you are asked to show further that with g = Fx , where x ∈ E n and the function F : E n → E 1 is twice continuously diﬀerentiable so that Fxx = (Fxx )T , called the Hessian, then (gf )x =(Fx f )x =Fx fx + f T Fxx = Fx fx + (Fxx f )T .

(1.16)

The latter result will be used in Chap. 2 for the derivation of (2.25). Many mathematical expressions in this book will be vector equations or inequalities involving vectors and vector functions. Since scalars are a special case of vectors, these expressions hold just as well for scalar equations or inequalities involving scalars and scalar functions. In fact, it may be a good idea to read them as scalar expressions on the ﬁrst reading. Then in the second and further readings, the extension to vector form will be easier.

1.4.5

Miscellany

The norm of an m-component row or column vector z is deﬁned to be 2 . z = z12 + · · · + zm (1.17) The norm of a vector is commonly used to deﬁne a neighborhood Nz0 of a point, e.g., Nz0 = {z| z − z0 < ε} , (1.18) where ε > 0 is a small positive real number.

1.4. Notation and Concepts Used

17

We will occasionally make use of the so-called “little-o” notation o(z). A function F (z) : E m → E 1 is said to be of the order o(z), if F (z) = 0. z→0 z lim

The most common use of this notation will be to collect higher order terms in a series expansion. In the continuous-time models discussed in this book, we generally will use x(t) to denote the state (column) vector, u(t) to denote the control (column) vector, and λ(t) to denote the adjoint (row) vector. Whenever there is no possibility of confusion, we will suppress the time indicator (t) from these vectors and write them as x, u, and λ, respectively. When talking about optimal state and control vectors, we put an asterisk “∗ ” as a superscript, i.e., as x∗ and u∗ , respectively, whereas u will refer to an admissible control with x as the corresponding state. No asterisk, however, needs to be put on the adjoint vector λ as it is only deﬁned along an optimal path. Thus, the values of the control, state and adjoint variables at time t along an optimal path will be written as u∗ (t), x∗ (t), and λ(t). When the control is expressed in terms of the state, it is called a feedback control. With an abuse of notation, we will express it as u(x), or u(x, t) if an explicit time dependence is required. Likewise, the optimal feedback control will be denoted as u∗ (x) or u∗ (x, t). We also use the simpliﬁed notation x (t) to mean (x(t)) , the transpose of x(t). Likewise, for a matrix A(t), we use A (t) to mean (A(t)) or the transpose of A(t), and A−1 (t) to mean (A(t))−1 or the inverse of A(t), when the inverse exists. The norm of an m-dimensional row or column vector function z(t), t ∈ [0, T ], is deﬁned to be z =

Σm j=1

0

T

zj2 (τ )dτ

12 .

(1.19)

In Chap. 4 and some other chapters, we will encounter functions of time with jumps. For such functions, it is useful to have the concepts of left and right limits. With ε > 0, these are deﬁned, respectively, for a function x(t) as x(T − ) = lim x(τ ) = lim x(T − ε) and x(T + ) = lim x(τ ) = lim x(T + ε). τ ↑T

ε→0

τ ↓T

ε→0

(1.20)

18

1. What Is Optimal Control Theory?

These limits are illustrated for a function x(t) graphed in Fig. 1.2. Here, x(0) = 1, x(0+ ) = 2, x(1− ) = 3, x(1+ ) = x(1) = 4, x(2− ) = 3, x(2) = 2, x(2+ ) = 1, x(3− ) = 2, x(3) = 3.

x(t )

4

[

3

)

2

)

(

)

(

1

0

1

2

3

t

Figure 1.2: Illustration of left and right limits In the discrete-time models introduced in Chap. 8 and applied in Chap. 9, we use xk , uk , and λk to denote state, control, and adjoint vectors, respectively, at time k, k = 0, 1, 2, . . . , T. We also denote the diﬀerence operator by Δxk := xk+1 − xk . As in the continuous-time case, the optimal values of the state variable xk and the control variable uk will have an asterisk as a superscript; thus, xk∗ and uk∗ denote the corresponding quantities along an optimal path. Once again, the adjoint variable λk along an optimal path will not have an asterisk.

1.4. Notation and Concepts Used

19

In order to specify the optimal control for linear control problems, we will introduce a special notation, called the bang function, as ⎧ ⎪ ⎪ ⎪ ⎪ ⎪ ⎨ bang[b1 , b2 ; W ] =

b1

if W < 0,

arbitrary if W = 0, ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ if W > 0. b2

(1.21)

In order to specify the optimal control for linear-quadratic problems, we deﬁne another special function, called the sat function, as ⎧ ⎪ ⎪ ⎪ y1 ⎪ ⎪ ⎨ sat[y1 , y2 ; W ] = W ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ y2

if W < y1 , if y1 ≤ W ≤ y2 ,

(1.22)

if W > y2 .

The word “sat” is short for the word “saturation.” The latter name comes from an electrical engineering application to saturated ampliﬁers. In several applications to be discussed, we will need the concept of impulse control, which is sometimes needed in cases when an unbounded control can be applied for a very short time. An example is the advertising model in Table 1.2 when Q = ∞. We apply unbounded control for a short time in order to cause a jump discontinuity in the state variable. For the example in Table 1.2, this might mean an intense advertising campaign (a media blitz) in order to increase advertising goodwill by a ﬁnite amount in a very short time. The impulse function deﬁned below is required to evaluate the integral in the objective function, which measures the cost of the intense advertising campaign. Suppose we want to apply an impulse control at time t to change the state variable from x(t) = x1 to the value x2 “immediately” after t, i.e., x(t+ ) = x2 . To compute its contribution to the objective function (1.2), we use the following procedure: given ε > 0 and a constant control u(ε), integrate (1.1) from t to t + ε with x(t) = x1 and choose u(ε) so that x(t + ε) = x2 ; this gives the trajectory x(τ ; ε, u(ε)) for τ ∈ [t, t + ε]. We can now compute imp(x1 , x2 ; t) = lim

ε→0 t

t+ε

F (x, u, τ )dτ .

(1.23)

20

1. What Is Optimal Control Theory?

If the impulse is applied only at time t, then we can calculate (1.2) as

t

J= 0

F (x, u, τ )dτ + imp(x1 , x2 ; t) +

T

F (x, u, τ )dτ + S[x(T ), T ]. t

(1.24)

If there are several instants at which impulses are applied, then this procedure is easily extended. Examples of the use of (1.24) occur in Chaps. 5 and 6. We frequently omit t in (1.23) when the impulse function is independent of t.

1.4.6

Convex Set and Convex Hull

A set D ⊂ E n is a convex set if for each pair of points y, z ∈ D, the entire line segment joining these two points is also in D, i.e., py + (1 − p)z ∈ D, for each p ∈ [0, 1]. Given xi ∈ E n , i = 1, 2, . . . , l, we deﬁne y ∈ E n to be a convex combination of xi ∈ E n , if there exists pi ≥ 0 such that l

pi = 1 and y =

i=1

l

pi x i .

i=1

The convex hull of a set D ⊂ E n is coD :=

l i=1

pi x i :

l

pi = 1, pi ≥ 0, xi ∈ D, i = 1, 2, . . . , l .

i=1

In other words, coD is the set of all convex combinations of points in D.

1.4.7

Concave and Convex Functions

A real-valued function ψ deﬁned on a convex set D ⊂ E n , i.e., ψ : D → E 1 , is concave, if for each pair of points y, z ∈ D and for all p ∈ [0, 1], ψ(py + (1 − p)z) ≥ pψ(y) + (1 − p)ψ(z).

1.4. Notation and Concepts Used

21

If the inequalities in the above deﬁnition are strict for all y, z ∈ D with y = z, and 0 < p < 1, then ψ is called a strictly concave function. In the single dimensional case of n = 1, there is an enlightening geometrical interpretation. Namely, ψ(x) deﬁned on an interval D = [a, b] is concave if, for each pair of points on the graph of ψ(x), the line segment joining these two points lies entirely below or on the graph of ψ(x); see Fig. 1.3. Reverting back to the n-dimensional case, if ψ is a diﬀerentiable function on a convex set D ⊂ E n , then it is concave, if for each pair of points y, z ∈ D, ψ(z) ≤ ψ(y) + ψ x (y)(z − y), where we understand y and z to be column vectors. Furthermore, if the function ψ is twice diﬀerentiable, then it is concave, if at each point in D, the n × n symmetric matrix ψ xx is negative semideﬁnite, i.e., all of its eigenvalues are non-positive. Finally, if ψ is a concave function, then the negative of the function ψ, i.e., −ψ : D → E 1 , is a convex function.

Figure 1.3: A concave function

22

1.4.8

1. What Is Optimal Control Theory?

Aﬃne Function and Homogeneous Function of Degree k

A function ψ : E n → E 1 is said to be aﬃne, if the function ψ(x)−ψ(0) is linear. Thus, ψ can be represented as ψ(x1 , x2 , . . . , xn ) = ni=1 ai xi + b, where ai , i = 1, 2, . . . , n, and b are scalar constants. A function ψ : E n → E 1 is said to be homogeneous of degree k, if ψ(bx) = bk ψ(x), where b > 0 is a scalar constant. In economics, we often assume that a ﬁrm’s production function is homogeneous of degree 1, i.e., if all inputs are multiplied by b, then output is multiplied by b. Such a production function is said to exhibitthe property of constant return to scale. A linear function ψ(x) = ax = ni=1 ai xi is a simple example of a homogeneous function of degree 1. Other examples n are ψ(x) = min{xi , i = 1, 2, . . . , n} and ψ(x) = a(Πni=1 xi αi )1/ i=1 αi with a > 0 and αi > 0, i = 1, 2, . . . , n. An important special case of the last example, known as the Cobb-Douglas production function, is ψ(K, L) = aK α1 Lα2 with α1 + α2 = 1, where K and L are factors of production called capital and labor, respectively, and a denotes the total factor productivity.

1.4.9

Saddle Point

An important concept in two-person zero-sum games is that of a saddle point. Let ψ(x, y), a real-valued function deﬁned on the space E n × E m , i.e., ψ : E n × E m → E 1 , be the payoﬀ of player 1 and −ψ(x, y) be the payoﬀ of player 2, when they make decisions x and y, respectively, in a zero-sum game. A point (ˆ x, yˆ) ∈ E n × E m is called a saddle point of ψ(x, y) or of the game, if ψ(ˆ x, y) ≥ ψ(ˆ x, yˆ) ≥ ψ(x, yˆ) for all x ∈ E n and y ∈ E m . Note that a saddle point may not exist, and even if it exists, it may not be unique. Note also that ψ(ˆ x, yˆ) = max ψ(x, yˆ) = min ψ(ˆ x, y). x

y

Intuitively, this could produce a picture like a horse saddle as shown in Fig. 1.4, hence the name saddle point for a point like (ˆ x, yˆ). This concept will be used in Sect. 13.1.

1.5. Plan of the Book

23

Figure 1.4: An illustration of a saddle point

1.4.10

Linear Independence and Rank of a Matrix

A set of vectors a1 , a2 , . . . , am in E n , m ≤ n, is said to be linearly dependent if there exist scalars pi not all zero such that m

pi ai = 0.

(1.25)

i=1

If (1.25) holds only when p1 = p2 = · · · = pm = 0, then the vectors are said to be linearly independent. In particular, if one of the vectors in the set {a1 , a2 , . . . , am } is a null vector, then the set is linearly dependent. The rank of an m × n matrix A, written rank(A), is the maximum number of linearly independent rows or, equivalently, the maximum number of linearly independent columns of A. An m × n matrix is of full rank if rank(A) = min{m, n}.

1.5

Plan of the Book

The book has thirteen chapters and ﬁve appendices: A, B, C, D, and E, covering a variety of topics which are listed in the table of contents and explained in the prefaces. In any given chapter, say Chap. 7, sections are numbered consecutively as 7.1, 7.2, 7.3, etc. Subsections are numbered consecutively within each section, i.e., 7.2.1, 7.2.2, 7.2.3, etc. Mathematical expressions are numbered consecutively by chapter as (7.1), (7.2), (7.3), etc. Theorems are also numbered consecutively by chapter as Theorem 7.1, Theorem 7.2, Theorem 7.3, etc. Similarly, deﬁnitions, examples, exercises, ﬁgures, propositions, remarks, and tables are numbered consecutively by chapter. These elements will be referenced throughout the

24

1. What Is Optimal Control Theory?

book by use of their designated numbers. The same scheme is used in the appendices, thus, sections in Appendix B, for example, are numbered as B.1, B.2, B.3, etc. Exercises for Chapter 1 E 1.1 In Example 1.1, let the functions and parameters of the production- inventory model be given by: h(I) = 10I, c(P ) = 20P, T = 10, I0 = 1, 000 Pmin = 600, Pmax = 1200, Imin = 800, S(t) = 900 + 10t. (a) Set P (t) = 1000 for 0 ≤ t ≤ 10. Determine whether this control is feasible; if it is feasible, compute the value J of the objective function. (b) If P (t) = 800, show that the terminal constraint is violated and hence the control is infeasible. (c) If P (t) = Pmin for 0 ≤ t ≤ 6 and P (t) = Pmax for 6 < t ≤ 10, show that the control is infeasible because the state constraint is violated. E 1.2 In Example 1.1, suppose there is a cost associated with changing the rate of production. One way to formulate this problem is to let the control variable u(t) denote the rate of change of the production rate P (t), having a cost cu2 associated with such changes, where c > 0. Formulate the new problem. Hint: Let P (t) be an additional state variable. √ E 1.3 For the advertising model in Example 1.2, let π(G) = 2 G, δ = 0.05, ρ = 0.2, Q = 2, and G0 = 16. Set u(t) = 0.8 for t ≥ 0, and show that G(t) is constant for all t. Compute the value J of the objective function. E 1.4 In Example 1.2, suppose G measures the number of people who know about the product. Hence, if A is the total population, then A − G is the number of people who do not know about the product. If u(t) measures the advertising rate at time t, assume that u(A − G) is the corresponding rate of increase of G due to this advertising. Formulate the new model.

Exercises for Chapter 1

25

E 1.5 Rich Rentier in Example 1.3 has initial wealth W0 = $1, 000, 000. Assume B = 0, ρ = 0.1, r = 0.15, and assume that Rich expects to live for exactly 20 years. (a) What is the maximum constant consumption level that Rich can aﬀord during his remaining life? (b) If Rich’s utility function is U (C) = ln C, what is the present value of the total utility in part (a)? (c) Suppose Rich sets aside $100,000 to start the Rentier Foundation. What is the maximum constant grant level that the foundation can support if it is to last forever? E 1.6 Suppose Rich in Exercise 1.5 takes on a part-time job, which yields an income of y(t) at time t. Assume y(t) = 10, 000e−0.05t and that he has a bequest function B(W ) = 0.5 ln W. (a) Reformulate this new optimal control problem. (b) If Rich (no longer a rentier) consumes at the constant rate found in Exercise 1.5(a), ﬁnd his terminal wealth and his new total utility. E 1.7 Consider the following educational policy question. Let S(t) denote the total number of scientists at time t, and let δ be the retirement rate of scientists. Let E(t) be the number of teaching scientists and R(t) be the number of research scientists, so that S(t) = E(t) + R(t). Assume γE(t) is the number of newly graduated scientists at time t, of which the policy allocates uγE(t) to the pool of teachers, where 0 ≤ u ≤ 1. The remaining graduates are added to the pool of researchers. The government has a target of maximizing the function αE(T ) + βR(T ) at a given future time T, where α and β are positive constants. Formulate the optimal control problem for the government. E 1.8 For F (x, y) deﬁned in Example 1.5, obtain the matrices Fxx and Fyy . E 1.9 Let x ∈ E m , g be an n-component row vector function of x, and f be an n-component column vector function of x. Use the ordinary product rule of calculus for functions of scalars to derive the formula (gf )x = gfx + f T (g T )x = gfx + f T gx .

26

1. What Is Optimal Control Theory?

E 1.10 Let F be a scalar function of x ∈ E n and f as deﬁned in Exercise 1.9. Assume F to be twice continuously diﬀerentiable. Show that (Fx f )x =Fx fx + f T Fxx = Fx fx + f T (Fxx )T = Fx fx + (Fxx f )T .

Hint: Set the gradient Fx = g, a row vector, and then use Exercise 1.9 to derive the ﬁrst equality. Note in connection with the second equality that the function F being twice continuously diﬀerentiable implies that Fxx = (Fxx )T . E 1.11 For Fy obtained in Example 1.5 and f deﬁned in Example 1.6, obtain (Fy f )y and verify the relation shown in Exercise 1.10. E 1.12 Use the bang function deﬁned in (1.21) to sketch the optimal control u∗ (t) = bang[−1, 1; W (t)] for 0 ≤ t ≤ 5, when (a) W (t) = t − 2 (b) W (t) = t2 − 4t + 3 (c) W (t) = sin πt. E 1.13 Use the sat function deﬁned in (1.22) to sketch the optimal control u∗ (t) = sat[2, 3; W (t)] for 0 ≤ t ≤ 5, when (a) W (t) = 4 − t (b) W (t) = 2 + t2 (c) W (t) = 4 − 4e−t . E 1.14 Evaluate the function imp(G1 , G2 ; t) for the advertising model of Table 1.2 when G2 > G1 , Q = ∞, and π(G) = pG, where p is a constant.

Chapter 2

The Maximum Principle: Continuous Time The main purpose of this chapter is to introduce the maximum principle as a necessary condition that must be satisﬁed by any optimal control for the basic problem speciﬁed in Sect. 2.1. Although vector notation is used, the reader can consider the problem as one with only a single state variable and a single control variable on the ﬁrst reading. In Sect. 2.2, the method of dynamic programming is used to derive the maximum principle. We use this method because of the simplicity and familiarity of the dynamic programming concept. The derivation also yields signiﬁcant economic interpretations. In Appendix C, the maximum principle is also derived by using a more general method similar to that of Pontryagin et al. (1962), but with certain simpliﬁcations. In Sect. 2.3, we apply the maximum principle to solve a number of simple, but illustrative, examples. In Sect. 2.4, the maximum principle is shown to be suﬃcient for optimal control under an appropriate concavity condition, which holds in many management science applications. Finally, Sect. 2.5 illustrates the use of Excel spreadsheet software to solve an optimal control problem.

2.1

Statement of the Problem

Optimal control theory deals with the problem of optimizing dynamic systems. The problem must be well posed before any solution can be attempted. This requires a clear mathematical description of the system © Springer Nature Switzerland AG 2019 S. P. Sethi, Optimal Control Theory, https://doi.org/10.1007/978-3-319-98237-3 2

27

28

2. The Maximum Principle: Continuous Time

to be optimized, the constraints imposed on the system, and the objective function to be maximized (or minimized).

2.1.1

The Mathematical Model

An important part of any control problem is the process of modeling the dynamic system under consideration, be it physical, business, or otherwise. The aim is to arrive at a mathematical description which is simple enough to deal with, and realistic enough to be able to predict the response of the system to any given input. Our model is restricted to systems that can be characterized by a set of ordinary diﬀerential equations (or, ordinary diﬀerence equations in the discrete-time case treated in Chap. 8). Thus, given the initial state x0 of the system and control history u(t), t ∈ [0, T ], of the process, the evolution of the system may be described by the ﬁrst-order diﬀerential equation, known also as the state equation, x(t) ˙ = f (x(t), u(t), t),

x(0) = x0 ,

(2.1)

where the vector of state variables, x(t) ∈ E n , the vector of control variables, u(t) ∈ E m , and f : E n × E m × E 1 → E n . Furthermore, the function f is assumed to be continuously diﬀerentiable. Here we assume x to be a column vector and f to be a column vector of functions. The path x(t), t ∈ [0, T ], is called a state trajectory and u(t), t ∈ [0, T ], is called a control trajectory or simply, a control. The terms vector of state variables, state vector, and state will be used interchangeably; similarly for the terms vector of control variables, control vector, and control. As mentioned earlier, when no confusion arises, we will usually suppress the time notation (t); thus, e.g., x(t) will be written simply as x. Furthermore, it should be inferred from the context whether x denotes the state at time t or the entire state trajectory. A similar statement holds for u.

2.1.2

Constraints

In this chapter, we are concerned with problems of types (1.4) and (1.5) that do not have state constraints. Such constraints are considered in Chaps. 3 and 4, as indicated in Sect. 1.1. We do impose constraints of type (1.3) on the control variables. We deﬁne an admissible control to be a control trajectory u(t), t ∈ [0, T ], which is piecewise continuous and satisﬁes, in addition, u(t) ∈ Ω(t) ⊂ E m ,

t ∈ [0, T ].

(2.2)

2.1. Statement of the Problem

29

Usually the set Ω(t) is determined by physical or economic constraints on the values of the control variables at time t.

2.1.3

The Objective Function

An objective function is a quantitative measure of the performance of the system over time. An optimal control is deﬁned to be an admissible control which maximizes the objective function. In business or economic problems, a typical objective function gives some appropriate measure of quantities such as proﬁt or sales. If the aim is to minimize cost, then the objective function to be maximized is the negative of cost. Mathematically, we let T J= F (x(t), u(t), t)dt + S(x(T ), T ) (2.3) 0

denote the objective function, where the functions F : E n × E m × E 1 → E 1 and S : E n × E 1 → E 1 are assumed for our purposes to be continuously diﬀerentiable. In a typical business application, F (x, u, t) could be the instantaneous proﬁt rate and S(x, T ) could be the salvage value of having x as the system state at the terminal time T.

2.1.4

The Optimal Control Problem

Given the preceding deﬁnitions we can state the optimal control problem, which we will be concerned with in this chapter. The problem is to ﬁnd an admissible control u∗ , which maximizes the objective function (2.3) subject to the state equation (2.1) and the control constraints (2.2). We now restate the optimal control problem as: T ⎧ ⎪ ⎪ max F (x, u, t)dt + S(x(T ), T ) J= ⎪ ⎪ ⎪ 0 ⎨ u(t)∈Ω(t) (2.4) subject to ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ x˙ = f (x, u, t), x(0) = x . 0 The control u∗ is called an optimal control and x∗ , determined by means of the state equation with u = u∗ , is called the optimal trajectory or an optimal path. The optimal value J(u∗ ) of the objective function will be

30

2. The Maximum Principle: Continuous Time

∗ when we need to emphasize its denoted as J ∗ , and occasionally as J(x 0) dependence on the initial state x0 . The optimal control problem (2.4) is said to be in Bolza form because of the form of the objective function in (2.3). It is said to be in Lagrange form when S ≡ 0. We say the problem is in Mayer form when F ≡ 0. Furthermore, it is in linear Mayer form when F ≡ 0 and S is linear, i.e., ⎧ ⎪ ⎪ max {J = cx(T )} ⎪ ⎪ ⎪ ⎨ u(t)∈Ω(t) (2.5) subject to ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ x˙ = f (x, u, t), x(0) = x , 0

where c = (c1 , c2 , · · · , cn ) is an n-dimensional row vector of given constants. In the next paragraph and in Exercise 2.5, it will be demonstrated that all of these forms can be converted into the linear Mayer form. To show that the Bolza form can be reduced to the linear Mayer form, we deﬁne a new state vector y = (y1 , y2 , . . . , yn+1 ), having n + 1 components deﬁned as follows: yi = xi for i = 1, . . . , n and yn+1 deﬁned by the solution of the equation ∂S(x, t) ∂S(x, t) f (x, u, t) + , (2.6) y˙ n+1 = F (x, u, t) + ∂x ∂t with yn+1 (0) = S(x0 , 0). By writing f (x, u, t) as f (y, u, t), with a slight abuse of notation, and by denoting the right-hand side of (2.6) as fn+1 (y, u, t), we can write the new state equation in the vector form as ⎞ ⎛ ⎛ ⎞ ⎛ ⎞ f (y, u, t) x ˙ x 0 ⎟ ⎜ ⎜ ⎟ ⎜ ⎟ y˙ = ⎝ (2.7) ⎠=⎝ ⎠ , y(0) = ⎝ ⎠. fn+1 (y, u, t) y˙ n+1 S(x0 , 0) We also put c = (0, · · · , 0, 1), where c has n + 1 components with the ﬁrst n terms all 0. If we integrate (2.6) from 0 to T, we see that T F (x, u, t)dt + S(x(T ), T ) − S(x0 , 0). yn+1 (T ) − yn+1 (0) = 0

In view of setting the initial condition as yn+1 (0) = S(x0 , 0), the problem in (2.4) can be expressed as that of maximizing T J= F (x, u, t)dt + S(x(T ), T ) = yn+1 (T ) = cy(T ) (2.8) 0

2.1. Statement of the Problem

31

over u(t) ∈ Ω(t), subject to (2.7). Of course, the price paid for going from Bolza to linear Mayer form is an additional state variable and its associated diﬀerential equation (2.6). Also, for the function fn+1 to be continuously diﬀerentiable, in keeping with the assumptions made in Sect. 2.1.1, we need to assume that the salvage value function S(x, t) is twice continuously diﬀerentiable. Exercise 2.5 presents the task of showing in a similar way that the Lagrange and Mayer forms can also be reduced to the linear Mayer form. Example 2.1 Convert the following single-state problem in Bolza form to its linear Mayer form:

T

u2 x− 2

max J = 0

1 dt + [x(T )]2 4

subject to x˙ = u, x(0) = x0 . Solution. We use (2.6) to introduce the additional state variable y2 as follows: u2 1 1 y˙ 2 = x − + xu, y2 (0) = x20 . 2 2 4 Then, u2 1 + xu dt y2 (0) + x− 2 2 0 T T 2 1 u dt + xx˙ dt + y2 (0) x− 2 2 0 0 T T 1 2 u2 d x− dt + x 2 4 0 0 T 1 u2 1 x− dt + [x(T )]2 − x20 + y2 (0) 2 4 4 0 T 2 1 u dt + x(T )2 x− 2 4 0 J.

y2 (T ) = = = = = =

T

32

2. The Maximum Principle: Continuous Time

Thus, the linear Mayer form version with the two-dimensional state y = (x, y2 ) can be stated as max {J = y2 (T )} subject to x˙ = u, x(0) = x0 , u2 1 1 y˙ 2 = x − + xu, y2 (0) = x20 . 2 2 4 In Sect. 2.2, we derive necessary conditions for optimal control in the form of the maximum principle, and in Sect. 2.4 we derive suﬃcient conditions. In these derivations, we shall assume the existence of an optimal control, while providing references where needed, as the topic of existence is beyond the scope of this book. In any particular application, however, the existence of a solution will be demonstrated by actually ﬁnding a solution that satisﬁes both the necessary and the suﬃcient conditions for optimality. We thus avoid the necessity of having to prove general existence theorems, which require advanced and diﬃcult mathematics. Nevertheless, interested readers can consult Hartl et al. (1995) and Seierstad and Sydsæter (1987) for brief discussions of existence results and references therein including Cesari (1983).

2.2

Dynamic Programming and the Maximum Principle

We will now derive the maximum principle by using a dynamic programming approach. The proof is intuitive in nature and is not intended to be mathematically rigorous. For more rigorous derivations, we refer the reader to Appendix C, Berkovitz (1961), Pontryagin et al. (1962), Halkin (1967), Boltyanskii (1971), Hartberger (1973), Bryant and Mayne (1974), Leitmann (1981), and Seierstad and Sydsæter (1987). Additional references can be found in the survey by Hartl et al. (1995). For discussions of maximum principles for more general optimal control problems, including those with nondiﬀerentiable functions, see Clarke (1983, 1989).

2.2.1

The Hamilton-Jacobi-Bellman Equation

Suppose V (x, t) : E n × E 1 → E 1 is a function whose value is the maximum value of the objective function of the control problem for the sys-

2.2. Dynamic Programming and the Maximum Principle

33

tem, given that we start at time t in state x. That is,

T F (x(s), u(s), s)ds + S(x(T ), T ) , V (x, t) = max u(s)∈Ω(s)

(2.9)

t

where for s ≥ t, dx(s) = f (x(s), u(s), s), x(t) = x. ds We initially assume that the value function V (x, t) exists for all x and t in the relevant ranges. Later we will make additional assumptions about the function V (x, t). Bellman (1957) in his book on dynamic programming states the principle of optimality as follows: An optimal policy has the property that, whatever the initial state and initial decision are, the remaining decision must constitute an optimal policy with regard to the outcome resulting from the initial decision. Intuitively this principle is obvious, for if we were to start in state x at time t and did not follow an optimal path from then on, there would then exist (by assumption) a better path from t to T, hence, we could improve the proposed solution by following this better path. We will use the principle of optimality to derive conditions on the value function V (x, t). Figure 2.1 is a schematic picture of the optimal path x∗ (t) in the state-time space, and two nearby points (x, t) and (x + δx, t + δt), where δt is a small increment of time and x + δx = x(t + δt). The value function changes from V (x, t) to V (x + δx, t + δt) between these two points. By the principle of optimality, the change in the objective function is made up of two parts: ﬁrst, the incremental change in J from t to t + δt, which is given by the integral of F (x, u, t) from t to t + δt; second, the value function V (x + δx, t + δt) at time t + δt. The control actions u(τ ) should be chosen to lie in Ω(τ ), τ ∈ [t, t + δt], and to maximize the sum of these two terms. In equation form this is t+δt V (x, t) = max F [x(τ ), u(τ ), τ ]dτ + V [x(t + δt), t + δt] , u(τ )∈Ω(τ )

τ ∈[t,t+δt]

t

(2.10)

34

2. The Maximum Principle: Continuous Time x

Optimal Path

V x, t

x

t

0

t

Figure 2.1: An optimal path in the state-time space where δt represents a small increment in t. It is instructive to compare this equation to deﬁnition (2.9). Since F is a continuous function, the integral in (2.10) is approximately F (x, u, t)δt so we can rewrite (2.10) as V (x, t) = max {F (x, u, t)δt + V [x(t + δt), t + δt]} + o(δt), u∈Ω(t)

(2.11)

where o(δt) denotes a collection of higher-order terms in δt. (By deﬁnition given in Sect. 1.4.4, o(δt) is a function such that limδt→0 o(δt) δt = 0.) We now make an assumption that we will return to again later. We assume that the value function V is a continuously diﬀerentiable function of its arguments. This allows us to use the Taylor series expansion of V with respect to δt and obtain V [x(t + δt), t + δt] = V (x, t) + [Vx (x, t)x˙ + Vt (x, t)]δt + o(δt),

(2.12)

where Vx and Vt are partial derivatives of V (x, t) with respect to x and t, respectively.

2.2. Dynamic Programming and the Maximum Principle

35

Substituting for x˙ from (2.1) in the above equation and then using it in (2.11), we obtain max {F (x, u, t)δt + V (x, t) + Vx (x, t)f (x, u, t)δt

V (x, t) =

u∈Ω(t)

+ Vt (x, t)δt} + o(δt).

(2.13)

Canceling V (x, t) on both sides and then dividing by δt we get 0 = max {F (x, u, t) + Vx (x, t)f (x, u, t) + Vt (x, t)} + u∈Ω(t)

o(δt) . δt

(2.14)

Now we let δt → 0 and obtain the following equation 0 = max {F (x, u, t) + Vx (x, t)f (x, u, t) + Vt (x, t)} , u∈Ω(t)

(2.15)

for which the boundary condition is V (x, T ) = S(x, T ).

(2.16)

This boundary condition follows from the fact that the value function at t = T is simply the salvage value function. The components of the vector Vx (x, t) can be interpreted as the marginal contributions of the state variables x to the value function or the maximized objective function (2.9). We denote the marginal return vector (along the optimal path x∗ (t)) by the adjoint (row) vector λ(t) ∈ E n , i.e., λ(t) = Vx (x∗ (t), t) := Vx (x, t) |x=x∗ (t) .

(2.17)

From the preceding remark, we can interpret λ(t) as the per unit change in the objective function value for a small change in x∗ (t) at time t. In other words, λ(t) is the highest hypothetical unit price which a rational decision maker would be willing to pay for an inﬁnitesimal addition to x∗ (t). See Sect. 2.2.4 for further discussion. Next we introduce a function H : E n × E m × E n × E 1 → E 1 called the Hamiltonian H(x, u, λ, t) = F (x, u, t) + λf (x, u, t).

(2.18)

We can then rewrite Eq. (2.15) as the equation max [H(x, u, Vx , t) + Vt ] = 0,

u∈Ω(t)

(2.19)

36

2. The Maximum Principle: Continuous Time

called the Hamilton-Jacobi-Bellman equation or, simply, the HJB equation to be satisﬁed along an optimal path. Note that it is possible to take Vt out of the maximizing operation since it does not depend on u. The Hamiltonian maximizing condition of the maximum principle can be obtained from (2.19) and (2.17) by observing that, if x∗ (t) and u∗ (t) are optimal values of the state and control variables and λ(t) is the corresponding value of the adjoint variable at time t, then the optimal control u∗ (t) must satisfy (2.19), i.e., for all u ∈ Ω(t), H[x∗ (t), u∗ (t), λ(t), t] + Vt (x∗ (t), t) ≥ H[x∗ (t), u, λ(t), t] +Vt (x∗ (t), t).

(2.20)

Canceling the term Vt on both sides, we obtain the Hamiltonian maximizing condition H[x∗ (t), u∗ (t), λ(t), t] ≥ H[x∗ (t), u, λ(t), t]

(2.21)

for all u ∈ Ω(t). In order to complete the statement of the maximum principle, we must still obtain the adjoint equation. Remark 2.1 We use u∗ and x∗ for optimal control and state to distinguish them from an admissible control u and the corresponding state x, respectively. However, since the adjoint variable λ is deﬁned only along the optimal path, there is no need for such a distinction, and therefore we do not use the superscript ∗ on λ.

2.2.2

Derivation of the Adjoint Equation

The derivation of the adjoint equation proceeds from the HJB equation (2.19), and is similar to those in Fel’dbaum (1965) and Kirk (1970). Note that, given the optimal path x∗ , the optimal control u∗ maximizes the left-hand side of (2.19), and its maximum value is zero. We now consider small perturbations of the values of the state variables in a neighborhood of the optimal path x∗ . Thus, let x(t) = x∗ (t) + δx(t), where δx(t) < ε for a small positive ε.

(2.22)

2.2. Dynamic Programming and the Maximum Principle

37

We now consider a ‘ﬁxed’ time instant t. We can then write (2.19) as 0 = H[x∗ (t), u∗ (t), Vx (x∗ (t), t), t] + Vt (x∗ (t), t) ≥ H[x(t), u∗ (t), Vx (x(t), t), t] + Vt (x(t), t).

(2.23)

To explain, we note from (2.19) that the left-hand side of ≥ in (2.23) equals zero. The right-hand side can attain the value zero only if u∗ (t) is also an optimal control for x(t). In general, for x(t) = x∗ (t), this will not be so. From this observation, it follows that the expression on the right-hand side of (2.23) attains its maximum (of zero) at x(t) = x∗ (t). Furthermore, x(t) is not explicitly constrained. In other words, x∗ (t) is an unconstrained local maximum of the right-hand side of (2.23), so that the derivative of this expression with respect to x must vanish at x∗ (t), i.e., Hx [x∗ (t), u∗ (t), Vx (x∗ (t), t), t] + Vtx (x∗ (t), t) = 0, (2.24) provided the derivative exists, and for which, we must further assume that V is a twice continuously diﬀerentiable function of its arguments. With H = F + Vx f from (2.17) and (2.18), we obtain Hx = Fx + Vx fx + f T Vxx = Fx + Vx fx + (Vxx f )T by using g = Vx in the identity (1.15). Substituting this in (2.24) and recognizing the fact that Vxx = (Vxx )T , we obtain Fx + Vx fx + f T Vxx + Vtx = Fx + Vx fx + (Vxx f )T + Vtx = 0,

(2.25)

where the superscript T denotes the transpose operation. See (1.16) or Exercise 1.10 for further explanation. The derivation of the necessary condition (2.25) is the crux of the reasoning in the derivation of the adjoint equation. It is easy to obtain the so-called adjoint equation from it. We begin by taking the time derivative of Vx (x, t). Thus, dVx1 dVx2 dVx dVxn = , ,··· , dt dt dt dt = (Vx1 x x˙ + Vx1 t , Vx2 x x˙ + Vx2 t , · · · , Vxn x x˙ + Vxn t ) n n = ( ni=1 Vx1 xi x˙i , i=1 Vx2 xi x˙i , · · · , i=1 Vxn xi x˙i ) + (Vx )t = (Vxx x) ˙ T + Vxt = (Vxx f )T + Vtx . (2.26)

38

2. The Maximum Principle: Continuous Time

Note in the above that Vxi x = (Vxi x1 , Vxi x2 , · · · , Vxi xn ) and

⎛

⎞⎛

⎜ V x1 x1 ⎜ ⎜ ⎜ Vx2 x1 ⎜ Vxx x˙ = ⎜ ⎜ .. ⎜ . ⎜ ⎝ Vxn x1

Vx1 x2 Vx2 x2 .. . Vxn x2

· · · Vx1 xn ⎟ ⎜ ⎟⎜ ⎟⎜ ⎜ · · · Vx2 xn ⎟ ⎟⎜ ⎟⎜ .. ⎟⎜ ⎟⎜ ··· . ⎟⎜ ⎠⎝ · · · Vx n x n

⎞ x˙ 1 ⎟ ⎟ ⎟ x˙ 2 ⎟ ⎟ ⎟. .. ⎟ . ⎟ ⎟ ⎠ x˙ n

(2.27)

Since the terms on the right-hand side of (2.26) are the same as the last two terms in (2.25), we see that (2.26) becomes dVx = −Fx − Vx fx . dt

(2.28)

Because λ was deﬁned in (2.17) to be Vx , we can rewrite (2.28) as λ˙ = −Fx − λfx . To see that the right-hand side of this equation can be written simply as −Hx , we need to go back to the deﬁnition of H in (2.18) and recognize that when taking the partial derivative of H with respect to x, the adjoint variables λ are considered to be independent of x. We note further that along the optimal path, λ is a function of t only. Thus, λ˙ = −Hx .

(2.29)

Also, from the deﬁnition of λ in (2.17) and the boundary condition (2.16), we have the terminal boundary condition, which is also called the transversality condition: λ(T ) =

∂S(x, T ) |x=x∗ (T ) = Sx (x∗ (T ), T ). ∂x

(2.30)

The adjoint equation (2.29) together with its boundary condition (2.30) determine the adjoint variables. This completes our derivation of the maximum principle using dynamic programming. We can now summarize the main results in the following section.

2.2. Dynamic Programming and the Maximum Principle

2.2.3

39

The Maximum Principle

The necessary conditions for u∗ (t), t ∈ [0, T ], to be an optimal control are: ⎧ ⎪ ⎪ ⎪ x˙ ∗ = f (x∗ , u∗ , t), x∗ (0) = x0 , ⎪ ⎪ ⎨ (2.31) λ˙ = −Hx [x∗ , u∗ , λ, t], λ(T ) = Sx (x∗ (T ), T ), ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ H[x∗ , u∗ , λ, t] ≥ H[x∗ , u, λ, t], ∀u ∈ Ω(t), t ∈ [0, T ]. It should be emphasized that the state and the adjoint arguments of the Hamiltonian are x∗ (t) and λ(t) on both sides of the Hamiltonian maximizing condition in (2.31), respectively. Furthermore, u∗ (t) must provide a global maximum of the Hamiltonian H[x∗ (t), u, λ(t), t] over u ∈ Ω(t). For this reason the necessary conditions in (2.31) are called the maximum principle. Note that in order to apply the maximum principle, we must simultaneously solve two sets of diﬀerential equations with u∗ obtained from the Hamiltonian maximizing condition in (2.31). With the control variable u∗ so obtained, the state equation for x∗ is given with the initial value x0 , and the adjoint equation for λ is speciﬁed with a condition on the terminal value λ(T ). Such a system of equations, where initial values of some variables and ﬁnal values of other variables are speciﬁed, is called a two-point boundary value problem (TPBVP). The general solution of such problems can be very diﬃcult; see Bryson and Ho (1975), Roberts and Shipman (1972), and Feichtinger and Hartl (1986). However, there are certain special cases which are easy. One such is the case in which the adjoint equation is independent of the state and the control variables; here we can solve the adjoint equation ﬁrst, then get the optimal control u∗ , and then solve for x∗ . Note also that if we can solve the Hamiltonian maximizing condition for an optimal control function in closed form u∗ (x, λ, t) so that u∗ (t) = u∗ [x∗ (t), λ(t), t], then we can substitute this into the state and adjoint equations to get the TPBVP just in terms of a set of diﬀerential equations, i.e., ⎧ ⎪ ⎨ x˙ ∗ = f (x∗ , u∗ (x∗ , λ, t), t), x∗ (0) = x0 , (2.32) ⎪ ⎩ λ˙ = −H (x∗ , u∗ (x∗ , λ, t), λ, t), λ(T ) = S (x∗ (T ), T ). x x

40

2. The Maximum Principle: Continuous Time

We should note that we are making a slight abuse of notation here by using u∗ (x, λ, t) to denote the optimal control function and u∗ (t) as the optimal control at time t. Thus, depending on the context, when we use u∗ without any argument, it may mean the optimal control function u∗ (x, λ, t), or the optimal control at time t, or the entire optimal control trajectory {u∗ (t), t ∈ [0, T ]}. In Sect. 2.5, we derive the TPBVP for a speciﬁc example, and solve its discrete version by using Excel. In subsequent chapters we will solve many TPBVPs of varying degrees of diﬃculty. One ﬁnal remark should be made. Because an integral is unaﬀected by values of the integrand at a ﬁnite set of points, some of the arguments made in this chapter may not hold at a ﬁnite set of points. This does not aﬀect the validity of the results. In the next section, we give economic interpretations of the maximum principle, and in Sect. 2.3, we solve ﬁve simple examples by using the maximum principle.

2.2.4

Economic Interpretations of the Maximum Principle

Recall from Sect. 2.1.3 that the objective function (2.3) is

T

F (x, u, t)dt + S(x(T ), T ),

J= 0

where F is considered to be the instantaneous proﬁt rate measured in dollars per unit of time, and S(x, T ) is the salvage value, in dollars, of the system at time T when the terminal state is x. For purposes of discussion it will be convenient to consider the system as a ﬁrm and the state x(t) as the stock of capital at time t. In (2.17), we interpreted λ(t) to be the per unit change in the value function V (x, t) for small changes in capital stock x. In other words, λ(t) is the marginal value per unit of capital at time t, and it is also referred to as the price or shadow price of a unit of capital at time t. In particular, the value of λ(0) is the marginal rate of change of the maximum value of J (the objective function) with respect to the change in the initial capital stock, x0 . Remark 2.2 As mentioned in Appendix C, where we prove a maximum principle without any smoothness assumption on the value function, there arise cases in which the value function may not be diﬀerentiable

2.2. Dynamic Programming and the Maximum Principle

41

with respect to the state variables. In such cases, when Vx (x∗ (t), t) does not exist, then (2.17) has no meaning. See Bettiol and Vinter (2010), Yong and Zhou (1999), and Cernea and Frankowska (2005) for interpretations of the adjoint variables or extensions of (2.17) in such cases. Next we interpret the Hamiltonian function in (2.18). Multiplying (2.18) formally by dt and using the state equation (2.1) gives Hdt = F dt + λf dt = F dt + λxdt ˙ = F dt + λdx. The ﬁrst term F (x, u, t)dt represents the direct contribution to J in dollars from time t to t + dt, if the ﬁrm is in state x (i.e., it has a capital stock of x), and we apply control u in the interval [t, t + dt]. The diﬀerential dx = f (x, u, t)dt represents the change in capital stock from time t to t + dt, when the ﬁrm is in state x and control u is applied. Therefore, the second term λdx represents the value in dollars of the incremental capital stock dx, and hence can be considered as the indirect contribution to J in dollars. Thus, Hdt can be interpreted as the total contribution to J from time t to t + dt when x(t) = x and u(t) = u in the interval [t, t + dt]. With this interpretation, it is easy to see why the Hamiltonian must be maximized at each instant of time t. If we were just to maximize F at each instant t, we would not be maximizing J, because we would ignore the eﬀect of the control in changing the capital stock, which gives rise to indirect contributions to J. The maximum principle derives the adjoint variable λ(t), the price of capital at time t, in such a way that λ(t)dx is the correct valuation of the indirect contribution to J from time t to t + dt. As a consequence, the Hamiltonian maximizing problem can be treated as a static problem at each instant t. In other words, the maximum principle decouples the dynamic maximization problem (2.4) in the interval [0, T ] into a set of static maximization problems associated with instants t in [0, T ]. Thus, the Hamiltonian can be interpreted as a surrogate proﬁt rate to be maximized at each instant of time t. The value of λ to be used in the maximum principle is given by (2.29) and (2.30), i.e., ∂H ∂F ∂f λ˙ = − =− − λ , λ(T ) = Sx (x(T ), T ). ∂x ∂x ∂x Rewriting the ﬁrst equation as −dλ = Hx dt = Fx dt + λfx dt,

42

2. The Maximum Principle: Continuous Time

we can observe that along the optimal path, −dλ, the negative of the increase or, in other words, the decrease in the price of capital from t to t + dt, which can be considered as the marginal cost of holding that capital, equals the marginal revenue Hx dt of investing the capital. In turn the marginal revenue Hx dt consists of the sum of the direct marginal contribution Fx dt and the indirect marginal contribution λfx dt. Thus, the adjoint equation becomes the equilibrium relation—marginal cost equals marginal revenue, which is a familiar concept in the economics literature; see, e.g., Cohen and Cyert (1965, p. 189) or Takayama (1974, p. 712). Further insight can be obtained by integrating the above adjoint equation from t to T as follows: T

Hx (x(τ ), u(τ ), λ(τ ), τ )dτ T = Sx (x(T ), T ) + t Hx dτ .

λ(t) = λ(T ) +

t

Note that the price λ(T ) of a unit of capital at time T is its marginal salvage value Sx (x(T ), T ). In the special case when S ≡ 0, we have λ(T ) = 0, as clearly no value can be derived or lost from an inﬁnitesimal increase in x(T ). The price λ(t) of a unit of capital at time t is the sum of its terminal price λ(T ) plus the integral of the marginal surrogate proﬁt rate Hx from t to T. The above interpretations show that the adjoint variables behave in much the same way as the dual variables in linear (and nonlinear) programming, with the diﬀerences being that here the adjoint variables are time dependent and satisfy derived diﬀerential equations. These connections will become clearer in Chap. 8, which addresses the discrete maximum principle.

2.3

Simple Examples

In order to absorb the maximum principle, the reader should study very carefully the examples in this section, all of which are problems having only one state and one control variable. Some or all of the exercises at the end of the chapter should also be worked. In the following examples and others in this book, we will at times omit the superscript ∗ on the optimal values of the state variables as long as no confusion arises from doing so.

2.3. Simple Examples

43

Example 2.2 Consider the problem: max J =

1

0

−xdt

(2.33)

subject to the state equation x˙ = u, x(0) = 1

(2.34)

u ∈ Ω = [−1, 1].

(2.35)

and the control constraint

Note that T = 1, F = −x, S = 0, and f = u. Because F = −x, we can interpret the problem as one of minimizing the (signed) area under the curve x(t) for 0 ≤ t ≤ 1. Solution First, we form the Hamiltonian H = −x + λu

(2.36)

and note that, because the Hamiltonian is linear in u, the form of the optimal control, i.e., the one that would maximize the Hamiltonian, is ⎧ ⎪ ⎪ ⎪ 1 if λ(t) > 0, ⎪ ⎪ ⎨ (2.37) u∗ (t) = arbitrary if λ(t) = 0, ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ −1 if λ(t) < 0, or referring to the notation in Sect. 1.4, u∗ (t) = bang[−1, 1; λ(t)].

(2.38)

To ﬁnd λ, we write the adjoint equation λ˙ = −Hx = 1, λ(1) = Sx (x(T ), T ) = 0.

(2.39)

Because this equation does not involve x and u, we can easily solve it as λ(t) = t − 1.

(2.40)

It follows that λ(t) = t−1 < 0 for t ∈ [0, 1) and so u∗ (1) = −1, t ∈ [0, 1). Since λ(1) = 0, for simplicity we can also set u∗ (1) = −1 at the single point t = 1. We can then specify the optimal control to be u∗ (t) = −1 for all t ∈ [0, 1].

44

2. The Maximum Principle: Continuous Time

Substituting this into the state equation (2.34) we have

whose solution is

x˙ = −1, x(0) = 1,

(2.41)

x∗ (t) = 1 − t for t ∈ [0, 1].

(2.42)

The graphs of the optimal state and adjoint trajectories appear in Fig. 2.2. Note that the optimal value of the objective function is J ∗ = −1/2.

Figure 2.2: Optimal state and adjoint trajectories for Example 2.2 In Sect. 2.2.4, we stated that the adjoint variable λ(t) gives the marginal value per unit increment in the state variable x(t) at time t. Let us illustrate this claim at time t = 0 with the help of Example 2.2. Note from (2.40) that λ(0) = −1. Thus, if we increase the initial value x(0) from 1, by a small amount ε, to a new value 1 + ε, where ε may be positive or negative, then we expect the optimal value of the objective function to change from J ∗ = −1/2 to ∗ J(1+ε) = −1/2 + λ(0)ε + o(ε) = −1/2 − ε + o(ε),

2.3. Simple Examples

45

where we use the subscript (1 + ε) to distinguish the new value from J ∗ as well as to emphasize its dependence on the new initial condition x(0) = 1 + ε. To verify this, we ﬁrst observe that u∗ (t) = −1, t ∈ [0, 1], remains optimal in this example for the new initial condition. Then from (2.41) with x(0) = 1 + ε, we can obtain the new optimal state trajectory, shown by the dotted line in Fig. 2.2 as x∗(1+ε) (t) = 1 + ε − t, t ∈ [0, 1], where the notation x∗(y) (t) indicates the dependence of the optimal trajectory on the initial value x(0) = y. Substituting this for x in (2.33) and integrating, we get the new objective function value to be −1/2 − ε. Since 0 is of the order o(ε), our claim has been illustrated. We should note that in general it may be necessary to perform separate calculations for positive and negative ε. It is easy to see, however, that this is not the case in this example. Example 2.3 Let us solve the same problem as in Example 2.2 over the interval [0, 2] so that the objective is: 2 max J = −xdt . (2.43) 0

The dynamics and constraints are (2.34) and (2.35), respectively, as before. Here we want to minimize the signed area between the horizontal axis and the trajectory of x(t) for 0 ≤ t ≤ 2. Solution As before, the Hamiltonian is deﬁned by (2.36) and the optimal control is as in (2.38). The adjoint equation λ˙ = 1, λ(2) = 0

(2.44)

is the same as (2.39) except that now T = 2 instead of T = 1. The solution of (2.44) is easily found to be λ(t) = t − 2,

t ∈ [0, 2].

(2.45)

The graph of λ(t) is shown in Fig. 2.3. With λ(t) as in (2.45), we can determine u∗ (t) = −1 throughout. Thus, the state equation is the same as (2.41). Its solution is given by (2.42) for t ∈ [0, 2]. The optimal value of the objective function is J ∗ = 0. The graph of x∗ (t) is also sketched in Fig. 2.3.

46

2. The Maximum Principle: Continuous Time

Figure 2.3: Optimal state and adjoint trajectories for Example 2.3 Example 2.4 The next example is: 1 1 2 max J = − x dt 2 0

(2.46)

subject to the same constraints as in Example 2.2, namely, x˙ = u, x(0) = 1, u ∈ Ω = [−1, 1].

(2.47)

Here F = −(1/2)x2 so that the interpretation of the objective function (2.46) is that we are trying to ﬁnd the trajectory x(t) in order that the area under the curve (1/2)x2 is minimized. Solution The Hamiltonian is 1 H = − x2 + λu. 2

(2.48)

2.3. Simple Examples

47

The control function u∗ (x, λ) that maximizes the Hamiltonian in this case depends only on λ, and it has the form u∗ (x, λ) = bang[−1, 1; λ].

(2.49)

Then, the optimal control at time t can be expressed as u∗ (t) = bang[−1, 1, λ(t)]. The adjoint equation is λ˙ = −Hx = x, λ(1) = 0.

(2.50)

Here the adjoint equation involves x, so we cannot solve it directly. Because the state equation (2.47) involves u, which depends on λ, we also cannot integrate it independently without knowing λ. A way out of this dilemma is to use some intuition. Since we want to minimize the area under (1/2)x2 and since x(0) = 1, it is clear that we want x to decrease as quickly as possible. Let us therefore temporarily assume that λ is nonpositive in the interval [0, 1] so that from (2.49) we have u = −1 throughout the interval. (In Exercise 2.8, you will be asked to show that this assumption is correct.) With this assumption, we can solve (2.47) as x(t) = 1 − t. (2.51) Substituting this into (2.50) gives λ˙ = 1 − t. Integrating both sides of this equation from t to 1 gives 1 1 ˙ )dτ = (1 − τ )dτ , λ(τ t

t

or

1 λ(1) − λ(t) = (τ − τ 2 ) |1t , 2 which, using λ(1) = 0, yields 1 1 (2.52) λ(t) = − t2 + t − . 2 2 The reader may now verify that λ(t) is nonpositive in the interval [0, 1], verifying our original assumption. Hence, (2.51) and (2.52) satisfy the necessary conditions. In Exercise 2.26, you will be asked to show that they satisfy suﬃcient conditions derived in Sect. 2.4 as well, so that they are indeed optimal. Thus, x∗ (t) = 1 − t, and using this in (2.46), we can get J ∗ = −1/6. Figure 2.4 shows the graphs of the optimal state and adjoint trajectories.

48

2. The Maximum Principle: Continuous Time

t

Figure 2.4: Optimal trajectories for Examples 2.4 and 2.5 Example 2.5 Let us rework Example 2.4 with T = 2, i.e., with the objective function: 2 1 2 max J = − x dt (2.53) 2 0 subject to the constraints (2.47). Solution The Hamiltonian is still as in (2.48) and the form of the optimal policy remains as in (2.49). The adjoint equation is λ˙ = x, λ(2) = 0, which is the same as (2.50) except T = 2 instead of T = 1. Let us try to extend the solution of the previous example from T = 1 to T = 2. Thus, we keep λ(t) as in (2.52) for t ∈ [0, 1] with λ(1) = 0. If we recall from the deﬁnition of the bang function that bang [−1, 1; 0] is not deﬁned, it allows us to choose u in (2.49) arbitrarily when λ = 0. This is an instance of singular control, so let us see if we can maintain the singular control by choosing u appropriately. To do this we choose u = 0 when λ = 0. Since λ(1) = 0 we set u(1) = 0 so that from (2.47), we have x(1) ˙ = 0. Now note that if we set u(t) = 0 for t > 1, then by integrating equations (2.47) and (2.50) forward from t = 1 to t = 2, we see that x(t) = 0 and λ(t) = 0 for 1 < t ≤ 2; in other words, u(t) = 0 maintains singular control in the interval. Intuitively, this is the correct answer since once we get x = 0, we should keep it at 0 in order to maximize the objective

2.3. Simple Examples

49

function J in (2.53). We will later give further discussion of singular control and state an additional necessary condition in Sect. D.6 for such cases; see also Bell and Jacobson (1975). In Fig. 2.4, we can get the singular solution by extending the graphs shown to the right (as shown by thick dotted line), making x∗ (t) = 0, λ(t) = 0, and u∗ (t) = 0 for 1 < t ≤ 2. With the trajectory x∗ (t), 0 ≤ t ≤ 2, thus obtained, we can use (2.53) to compute the optimal value of the objective function as 1 2 ∗ 2 −(1/2)(1 − t) dt + −(1/2)(0)dt = −1/6. J = 0

1

Now suppose that the initial x(0) is perturbed by a small amount ε to x(0) = 1 + ε, where ε may be positive or negative. According to the marginal value interpretation of λ(0), whose value is −1/2 in this example, we can estimate the change in the objective function to be λ(0)ε + o(ε) = −ε/2 + o(ε). Next we calculate directly the impact of the perturbation in the initial value. For this we must obtain new control and state trajectories. These are clearly ⎧ ⎪ ⎨ −1, t ∈ [0, 1 + ε], ∗ u(1+ε) (t) = ⎪ ⎩ 0, t ∈ (1 + ε, 2], and x∗(1+ε) (t) =

⎧ ⎪ ⎨ 1 + ε − t, t ∈ [0, 1 + ε], ⎪ ⎩

0,

t ∈ (1 + ε, 2],

where we have used the subscript (1 + ε) to distinguish these from the original trajectories as well as to indicate their dependence on the initial value x(0) = 1 + ε. We can then obtain the corresponding optimal value of the objective function as 1+ε ∗ −(1/2)(1 + ε − t)2 dt = −1/6 − ε/2 − ε2 /2 − ε3 /6 J(1+ε) = 0

= −1/6 + λ(0)ε + o(ε), where o(ε) = −ε2 /2 − ε3 /6.

50

2. The Maximum Principle: Continuous Time

In this example and Example 2.2, we have, by direct calculation, demonstrated the signiﬁcance of λ(0) as the marginal value of the change in the initial state. This could have also been accomplished by obtaining the value function V (x, t) for x(t) = x, t ∈ [0, 2], and then showing that λ(0) = Vx (1, 0). This, of course, is the relationship (2.17) at x(0) = x = 1 and t = 0. Keep in mind, however, that deriving V (x, t) is more than just ﬁnding the solution of the problem, which we have already found by using the maximum principle. V (x, t) also yields additional insights into the problem. In order to completely specify V (x, t) for all x ∈ E 1 and all t ∈ [0, 2], we need to deal with a number of cases. Here, we will carry out the details only in the case of any t ∈ [0, 2] and 0 ≤ x ≤ 2 − t, and leave the listing of the other cases and the required calculations as Exercise 2.13. We know from (2.9) that we need to solve the optimal control problem for any given t ∈ [0, 2] with 0 ≤ x ≤ 2 − t. However, from our earlier analysis of this example, it is clear that the optimal control ⎧ ⎪ ⎨ −1, s ∈ [t, t + x], ∗ u(x,t) (s) = ⎪ ⎩ 0, s ∈ (t + x, 2], and the corresponding x∗(x,t) (s) =

⎧ ⎪ ⎨ x − (s − t),

s ∈ [t, t + x],

⎪ ⎩

s ∈ (t + x, 2],

0,

where we use the subscript to show the dependence of the control and state trajectories of a problem beginning at time t with the state x(t) = x. Thus, t+x 1 1 t+x − [x∗(x,t) (s)]2 ds = − (x − s + t)2 ds. V (x, t) = 2 2 t t While this expression can be easily integrated to obtain an explicit solution for V (x, t), we do not need to do this for our immediate purpose at hand, which is to obtain Vx (x, t). Diﬀerentiating the right-hand side with respect to x, we obtain 1 x+t 2(x − s + t)ds. Vx (x, t) = − 2 t

2.3. Simple Examples

51

Furthermore, since

x∗ (t) =

⎧ ⎪ ⎨ 1 − t, t ∈ [0, 1], ⎪ ⎩

t ∈ (1, 2],

0,

we obtain Vx (x∗ (t), t) =

⎧ ⎪ ⎨ − 1 1 2(x − s + t)ds = − 1 t2 + t − 1 , t ∈ [0, 1], 2 t 2 2 ⎪ ⎩

t ∈ (1, 2],

0,

which equals λ(t) obtained as the adjoint variable in Example 2.5. Note that for t ∈ [0, 1], λ(t) in Example 2.5 is the same as that in Example 2.4 obtained in (2.52). Example 2.6 This example is slightly more complicated and the optimal control is not bang-bang. The problem is:

max J = 0

2

2

(2x − 3u − u )dt

(2.54)

subject to x˙ = x + u, x(0) = 5

(2.55)

and the control constraint u ∈ Ω = [0, 2].

(2.56)

Solution Here T = 2, F = 2x − 3u − u2 , S = 0, and f = x + u. The Hamiltonian is H = (2x − 3u − u2 ) + λ(x + u) = (2 + λ)x − (u2 + 3u − λu).

(2.57)

Let us ﬁnd the optimal control policy by diﬀerentiating (2.57) with respect to u. Thus, ∂H = −2u − 3 + λ = 0, ∂u

52

2. The Maximum Principle: Continuous Time

so that the form of the optimal control is u∗ (t) =

λ(t) − 3 , 2

(2.58)

provided this expression stays within the interval Ω = [0, 2]. Note that the second derivative of H with respect to u is ∂ 2 H/∂u2 = −2 < 0, so that (2.58) satisﬁes the second-order condition for the maximum of a function. We next derive the adjoint equation as ∂H = −2 − λ, λ(2) = 0. λ˙ = − ∂x

(2.59)

Referring to Appendix A.1, we can use the integrating factor et to obtain et (dλ + λdt) = d(et λ) = −2et dt. We then integrate it on both sides from t to 2 and use the terminal condition λ(2) = 0 to obtain the solution of the adjoint equation (2.59) as λ(t) = 2(e2−t − 1). If we substitute this into (2.58) and impose the control constraint (2.56), we see that the optimal control is ⎧ ⎪ ⎪ ⎪ 2 if e2−t − 2.5 > 2, ⎪ ⎪ ⎨ (2.60) u∗ (t) = e2−t − 2.5 if 0 ≤ e2−t − 2.5 ≤ 2, ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 0 if e2−t − 2.5 < 0, or referring to the notation deﬁned in (1.22), u∗ (t) = sat[0, 2; e2−t − 2.5]. The graph of u∗ (t) appears in Fig. 2.5. In the ﬁgure, t1 satisﬁes e2−t1 − 2.5 = 2, i.e., t1 = 2 − ln 4.5 ≈ 0.496, while t2 satisﬁes e2−t2 − 2.5 = 0, which gives t2 = 2 − ln 2.5 ≈ 1.08. In Exercise 2.2 you will be asked to compute the optimal state trajectory x∗ (t) corresponding to u∗ (t) shown in Fig. 2.5 by piecing together the solutions of three separate diﬀerential equations obtained from (2.55) and (2.60).

2.4. Suﬃciency Conditions

e2

53

t

2.5

Figure 2.5: Optimal control for Example 2.6

2.4

Suﬃciency Conditions

So far, we have shown the necessity of the maximum principle conditions for optimality. Next we prove a theorem that gives qualiﬁcations under which the maximum principle conditions are also suﬃcient for optimality. This theorem is important from our point of view since the models derived from many management science applications will satisfy conditions required for the suﬃciency result. As remarked earlier, our technique for proving existence will be to display for any given model, a solution that satisﬁes both necessary and suﬃcient conditions. A good reference for suﬃciency conditions is Seierstad and Sydsæter (1987). We ﬁrst deﬁne a function H 0 : E n ×E m ×E 1 → E 1 called the derived Hamiltonian as follows: H 0 (x, λ, t) = max H(x, u, λ, t). u∈Ω(t)

(2.61)

54

2. The Maximum Principle: Continuous Time

We assume that by this equation a function u∗ (x, λ, t) is implicitly and uniquely deﬁned. Given these assumptions we have by deﬁnition, H 0 (x, λ, t) = H(x, u∗ , λ, t).

(2.62)

For our proof of the suﬃciency of the maximum principle, we also need the derivative Hx0 (x, λ, t), which by use of the Envelope Theorem can be given as Hx0 (x, λ, t) = Hx (x, u∗ , λ, t) := Hx (x, u, λ, t)|u=u∗ .

(2.63)

To see this in the case when u∗ (x, λ, t) is diﬀerentiable in x, let us diﬀerentiate (2.62) with respect to x: Hx0 (x, λ, t) = Hx (x, u∗ , λ, t) + Hu (x, u∗ , λ, t)

∂u∗ . ∂x

(2.64)

To obtain (2.63) from (2.64), we need to show that the second term on the right-hand side of (2.64) vanishes, i.e., Hu (x, u∗ , λ, t)

∂u∗ =0 ∂x

(2.65)

for each x. There are two cases to consider. If u∗ is in the interior of Ω(t), then it satisﬁes the ﬁrst-order condition Hu (x, u∗ , λ, t) = 0, thereby implying (2.65). Otherwise, u∗ is on the boundary of Ω(t). Then, for each i, j, either Hui = 0 or ∂u∗i /∂xj = 0 or both. Once again, (2.65) holds. Exercise 2.25 gives a speciﬁc instance of this case. Remark 2.3 We have shown the result in (2.63) in cases when u∗ is a diﬀerentiable function of x. The result holds more generally, provided that Ω(t) is appropriately qualiﬁed; see Derzko et al. (1984). Such results are known as Envelope Theorems, and are used often in economics. Theorem 2.1 (Suﬃciency Conditions). Let u∗ (t), and the corresponding x∗ (t) and λ(t) satisfy the maximum principle necessary condition (2.31) for all t ∈ [0, T ]. Then, u∗ is an optimal control if H 0 (x, λ(t), t) is concave in x for each t and S(x, T ) is concave in x. Proof. The proof is a minor extension of the arguments in Arrow and Kurz (1970). By deﬁnition H[x(t), u(t), λ(t), t] ≤ H 0 [x(t), λ(t), t].

(2.66)

2.4. Suﬃciency Conditions

55

Since H 0 is diﬀerentiable and concave, we can use the applicable deﬁnition of concavity given in Sect. 1.4 to obtain H 0 [x(t), λ(t), t] ≤ H 0 [x∗ (t), λ(t), t] + Hx0 [x∗ (t), λ(t), t][x(t) − x∗ (t)]. (2.67) Using (2.66), (2.62), and (2.63) in (2.67), we obtain H[x(t), u(t), λ(t), t] ≤ H[x∗ (t), u∗ (t), λ(t), t] +Hx [x∗ (t), u∗ (t), λ(t), t][x(t) − x∗ (t)].

(2.68)

By deﬁnition of H in (2.18) and the adjoint equation of (2.31) F [x(t), u(t), t] + λ(t)f [x(t), u(t), t] ≤ F [x∗ (t), u∗ (t), t] ˙ − x∗ (t)]. (2.69) +λ(t)f [x∗ (t), u∗ (t), t] − λ(t)[x(t) Using the state equation in (2.31), transposing, and regrouping, ˙ − x∗ (t)] F [x∗ (t), u∗ (t), t] − F [x(t), u(t), t] ≥ λ(t)[x(t) +λ(t)[x(t) ˙ − x˙ ∗ (t)]. (2.70) Furthermore, since S(x, T ) is a diﬀerential and concave function in its ﬁrst argument, we have S(x(T ), T ) ≤ S(x∗ (T ), T ) + Sx (x∗ (T ), T )[x(T ) − x∗ (T )]

(2.71)

S(x∗ (T ), T ) − S(x(T ), T ) ≥ Sx (x∗ (T ), T )[x(T ) − x∗ (T )].

(2.72)

or,

Integrating both sides of (2.70) from 0 to T and adding (2.72), we have

T 0

∗

∗

∗

F (x (t), u (t), t)dt + S(x (T ), T )

T F (x(t), u(t), t)dt + S(x(T ), T ) − 0

≥ [λ(T ) − Sx (x∗ (T ), T )][x(T ) − x∗ (T )] − λ(0)[x(0) − x∗ (0)]

56

2. The Maximum Principle: Continuous Time

or, J(u∗ ) − J(u)

(2.73) ∗

∗

∗

≥ [λ(T ) − Sx (x (T ), T )][x(T ) − x (T )] − λ(0)[x(0) − x (0)], where J(u) is the value of the objective function associated with a control u. Since x∗ (0) = x(0) = x0 , the initial condition, and since λ(T ) = Sx (x∗ (T ), T ) from the terminal adjoint condition in (2.31), we have J(u∗ ) ≥ J(u).

(2.74)

Thus, u∗ is an optimal control. This completes the proof. 2 Because λ(t) is not known a priori, it is usual to test H 0 for a stronger assumption, i.e., to check for the concavity of the function H 0 (x, λ, t) in x for any λ and t. Sometimes the stronger condition given in Exercise 2.27 can be used. Mangasarian (1966) gives a suﬃcient condition in which the concavity of H 0 (x, λ(t), t) in Theorem 2.1 is replaced by a stronger condition requiring the Hamiltonian H(x, u, λ(t), t) to be jointly concave in (x, u). Example 2.7 Let us show that the problems in Examples 2.2 and 2.3 satisfy the suﬃcient conditions. We have from (2.36) and (2.61), H 0 = −x + λu∗ , where u∗ is given by (2.37). Since u∗ is a function of λ only, H 0 (x, λ, t) is certainly concave in x for any t and λ (and in particular for λ(t) supplied by the maximum principle). Since S(x, T ) = 0, the suﬃcient conditions hold. Finally, it is important to mention that thus far in this chapter, we have considered problems in which the terminal values of the state variables are not constrained. Such problems are called free-end-point problems. The problems at the other extreme, where the terminal values of the state variables are completely speciﬁed, are termed ﬁxed-end-point problems. Then, there are problems in between these two extremes. While a detailed discussion of terminal conditions on state variables appears in Sect. 3.4 of the next chapter, it is instructive here to brieﬂy indicate how the maximum principle needs to be modiﬁed in the case of ﬁxed-end-point problems. Suppose x(T ) is completely speciﬁed, i.e.,

2.5. Solving a TPBVP by Using Excel

57

x(T ) = k ∈ E n , where k is a vector of constants. Observe then that the ﬁrst term on the right-hand side of inequality (2.73) vanishes regardless of the value of λ(T ), since x(T ) − x∗ (T ) = k − k = 0 in this case. This means that the suﬃciency result would go through for any value of λ(T ). Not surprisingly, therefore, the transversality condition (2.30) in the ﬁxed-end-point case changes to λ(T ) = β,

(2.75)

is a vector of constants to be determined. where β ∈ Indeed, one can show that (2.75) is also the necessary transversality condition for ﬁxed point problems. With this observation, the maximum principle for ﬁxed-end-point problems can be obtained by modifying (2.31) as follows: adding x(T ) = k and removing λ(T ) = Sx (x∗ (T ), T ). Likewise, the resulting TPBVP (2.32) can be modiﬁed correspondingly; it will have initial and ﬁnal values on the state variables, whereas both initial and terminal values for the adjoint variables are unspeciﬁed, i.e., λ(0) and λ(T ) are constants to be determined. In Exercises 2.28 and 2.19, you are asked to solve the ﬁxed-end-point problems given there. En

2.5

Solving a TPBVP by Using Excel

A number of examples and exercises found throughout this book involve ﬁnding a numerical solution to a two-point boundary value problem (TPBVP). In this section we will show how the GOAL SEEK function in Excel can be used for this purpose. We will solve the following example. Example 2.8 Consider the problem: 1 1 2 2 max J = − (x + u )dt 2 0 subject to

x˙ = −x3 + u, x(0) = 5.

Solution We form the Hamiltonian 1 H = − (x2 + u2 ) + λ(−x3 + u), 2

(2.76)

58

2. The Maximum Principle: Continuous Time

where the adjoint variable λ satisﬁes the equation λ˙ = x + 3x2 λ, λ(1) = 0.

(2.77)

Since u is unconstrained, we set Hu = 0 to obtain u∗ = λ. With this, the state equation (2.76) becomes x˙ = −x3 + λ, x(0) = 5.

(2.78)

Thus, the TPBVP is given by the system of equations (2.77) and (2.78). A simple method to solve the TPBVP uses what is known as the shooting method, explained in the ﬂowchart in Fig. 2.6.

Guess

?

Yes

STOP

No

Figure 2.6: The ﬂowchart for Example 2.8 We will use Excel functions to implement the shooting method. For this we discretize (2.77) and (2.78) by replacing dx/dt and dλ/dt by x(t + t) − x(t) λ λ(t + t) − λ(t) x = and = , t t t t respectively. Substitution of x/t for x˙ in (2.78) and λ/t for λ˙ in (2.77) gives the discrete version of the TPBVP: x(t + t) = x(t) + [−x(t)3 + λ(t)] t, x(0) = 5,

(2.79)

λ(t + t) = λ(t) + [x(t) + 3x(t)2 λ(t)] t, λ(1) = 0.

(2.80)

2.5. Solving a TPBVP by Using Excel

59

In order to solve these equations, open an empty spreadsheet, choose the unit of time to be t = 0.01, make a guess for the initial value λ(0) to be, say −0.2, and make the entries in the cells of the spreadsheet as speciﬁed below: Enter Enter Enter Enter

-0.2 in cell A1. 5 in cell B1. = A1 + (B1 + 3 ∗ (B1ˆ2)∗ A1)∗ 0.01 in cell A2. = B1 + (-B1ˆ 3 + A1) ∗ 0.01 in cell B2.

Here we have entered the right-hand side of the diﬀerence equation (2.80) for t = 0 in cell A2 and the right-hand side of the diﬀerence equation (2.79) for t = 0 in cell B2. Note that λ(0) = −0.2 shown as the entry −0.2 in cell A1 is merely a guess. The correct value will be determined by the use of the GOAL SEEK function. Next highlight cells A2 and B2 and drag the combination down to row 101 of the spreadsheet. Using EDIT in the menu bar, select FILL DOWN. Thus, Excel will solve Eqs. (2.80) and (2.79) from t = 0 to t = 1 in steps of t = 0.01, and that solution will appear as entries in columns A and B of the spreadsheet, respectively. In other words, the guessed solution for λ(t) will appear in cells A1 to A101 and the corresponding solution for x(t) will appear in cells B1 to B101. To ﬁnd the correct value for λ(0), use the GOAL SEEK function under TOOLS in the menu bar and make the following entries: Set cell: A101. To value: 0. By changing cell: A1. It ﬁnds the correct initial value for the adjoint variable as λ(0) = −0.10437, which should appear in cell A1, and the correct ending value of the state variable as x(1) = 0.62395, which should appear in cell B101. You will notice that the entry in cell A101 may not be exactly zero as instructed, although it will be very close to it. In our example, it is −0.0007. By using the CHART function, the graphs of x∗ (t) and λ(t) can be printed out by Excel as shown in Fig. 2.7.

60

2. The Maximum Principle: Continuous Time

Figure 2.7: Solution of TPBVP by excel As we discuss more complex problems involving control and state inequality constraints in Chaps. 3 and 4, we will realize that the shooting method is no longer adequate to solve such problems. However, there is a large amount of literature devoted to computational methods for solving optimal control problems. While a detailed treatment of this topic is beyond the scope of this book, we suggest some references as well as a software in Chap. 4, Sect. 4.3. Exercises for Chapter 2 E 2.1 Perform the following: (a) In Example 2.2, show J ∗ = −1/2. (b) In Example 2.3, show J ∗ = 0. (c) In Example 2.4, show J ∗ = −1/6. (d) In Example 2.5, show J ∗ = −1/6. E 2.2 Complete Example 2.6 by writing the optimal x∗ (t) in the form of integrals over the three intervals (0, t1 ), (t1 , t2 ), and (t2 , 2) shown in Fig. 2.5.

Exercises for Chapter 2

61

Hint: It is not necessary to actually carry out the numerical evaluation of these integrals unless you are ambitious. E 2.3 Find the optimal solution for Example 2.1 with x0 = 0 and T = 1. E 2.4 Rework Example 2.6 with F = 2x − 3u. E 2.5 Show that both the Lagrange and Mayer forms of the optimal control problem can be reduced to the linear Mayer form (2.5). E 2.6 Show that the optimal control obtained from the application of the maximum principle satisﬁes the principle of optimality: if u∗ (t) is an optimal control and x∗ (t) is the corresponding optimal path for 0 ≤ t ≤ T with x(0) = x0 , then verify the above proposition by showing that u∗ (t) for τ ≤ t ≤ T satisﬁes the maximum principle for the problem beginning at time τ with the initial condition x(τ ) = x∗ (τ ). E 2.7 Provide an alternative derivation of the adjoint equation in Sect. 2.2.2 by starting with a restatement of the Eq. (2.19) as −Vt = H 0 and diﬀerentiating it with respect to x. E 2.8 In Example 2.4, show that in view of (2.47) any λ(t), t ∈ [0, 1], that satisﬁes (2.50) must be nonnegative. E 2.9 The system deﬁned in (2.4) is termed autonomous if F, f, S and Ω are not explicit functions of time t. In this case, show that the Hamiltonian is constant along the optimal path, i.e., show that dH/dt = 0. Furthermore, verify this result in Example 2.4 by a direct substitution for x and λ from (2.51) and (2.52), respectively, into H given in (2.48). E 2.10 In Example 2.4, verify by direct calculation that with a new initial value x(0) = 1+ε with ε small, the new optimal objective function value will be ∗ J(1+ε) = −1/6 + λ(0)ε + o(ε) = −1/6 − ε/2 − ε2 /2.

E 2.11 In Example 2.6, verify by direct calculation that with a new initial x(0) = 5 + ε with ε small, the objective function value will change by λ(0)ε + o(ε) = 2(e2 − 1)ε + o(ε).

62

2. The Maximum Principle: Continuous Time

E 2.12 Obtain the value function V (x, t) explicitly in Example 2.4 and verify the relation Vx (x∗ (t), t) = λ(t) for the example by showing that Vx (1 − t, t) = −(1/2)t2 + t − 1/2. E 2.13 Obtain the value function V (x, t) explicitly in Example 2.5 for every x ∈ E 1 and t ∈ [0, 2]. Hint: You need to deal with the following cases for t ∈ [0, 2]: (i) (ii) (iii) (iv)

0 ≤ x ≤ 2 − t, x > 2 − t, t − 2 ≤ x < 0, and x < t − 2.

E 2.14 Obtain V (x, t) in Example 2.6 for small positive and negative x for t ∈ [t2 , 2]. Then, show that Vx (x, t) = 2(e2−t − 1), t ∈ [t2 , 2], is the same as λ(t), t ∈ [t2 , 2] obtained in Example 2.6. E 2.15 Solve the problem: max J =

T 0

(x −

u2 )dt 2

subject to x˙ = u, x(0) = x0 , u ∈ [0, 1], for optimal control and optimal state trajectory. Verify that your solution is optimal by using the maximum principle suﬃciency condition. E 2.16 Solve completely the problem:

1

(x + u)dt

max 0

x˙ = 1 − u2 , x(0) = 1; that is, ﬁnd x∗ (t), u∗ (t) and λ(t), 0 ≤ t ≤ 1.

Exercises for Chapter 2

63

E 2.17 Use the maximum principle to solve the following problem given in the Mayer form: max[8x1 (18) + 4x2 (18)] subject to x˙ 1 = x1 + x2 + u, x1 (0) = 15, x˙ 2 = 2x1 − u, x2 (0) = 20, and the control constraint 0 ≤ u ≤ 1. Hint: Use the method in Appendix A to solve the simultaneous diﬀerential equations. E 2.18 In Fig. 2.8, a water reservoir being used for the purpose of ﬁreﬁghting is leaking, and its water height x(t) is governed by x˙ = −0.1x + u, x(0) = 10, where u(t) denotes the net inﬂow at time t and 0 ≤ u ≤ 3. Note that x(t) also represents the water pressure in appropriate units. Since high water pressure is useful for ﬁre-ﬁghting, the objective function in (a) below involves keeping the average pressure high, while that in (b) involves building up a high pressure at T = 100. Furthermore, we do not need to impose the state constraints 0 ≤ x(t) ≤ 50, as these will always be satisﬁed for every feasible control u(t), 0 ≤ t ≤ 100.

Figure 2.8: Water reservoir of Exercise 2.18 (a) Find the optimal control which maximizes 100 xdt. J= 0

Find the maximum level reached.

64

2. The Maximum Principle: Continuous Time

(b) Replace the objective function in (a) by J = 5x(100), and re-solve the problem. (c) Redo the problem with J =

100 0

(x − 5u)dt.

E 2.19 Consider the following ﬁxed-end-point problem: T 2 (g(x) + cu )dt max J = − u

0

subject to x˙ = f (x) + b(x)u, x(0) = x0 , x(T ) = 0, where functions g ≥ 0, f, and b are assumed to be continuously diﬀerentiable. Derive the two-point boundary value problem (TPBVP) satisﬁed by the optimal state and control trajectories. E 2.20 A Machine Maintenance Problem. Consider the machine state dynamics x˙ = −δx + u, x(0) = x0 > 0, where δ > 0 is the rate of deterioration of the machine state and u is the rate of machine maintenance. Find the optimal maintenance rate: T u2 −ρt −ρT e (πx − )dt + e Sx(T ) , max J = 2 0 where π > 0 with πx representing the proﬁt rate when the machine state is x, u2 /2 is the cost of maintaining the machine at rate u, ρ > 0 is the discount rate, T is the time horizon, and S > 0 is the salvage value of the machine for each unit of the machine state at time T. Furthermore, show that the optimal maintenance rate decreases, increases, or remains constant over time depending on whether the diﬀerence S − π/(ρ + δ) is negative, positive, or zero, respectively. E 2.21 Transform the machine maintenance problem of Exercise 2.20 into Mayer Form. Then solve it to obtain the optimal maintenance rate. E 2.22 Regional Allocation of Investment. Let Ki , i = 1, 2, denote the capital stock in Region i. Let bi be the productivity of capital and si be

Exercises for Chapter 2

65

the marginal propensity to save in Region i. Since the investment funds for the two regions come from the savings in the whole economy, we have K˙ 1 + K˙ 2 = b1 s1 K1 + b2 s2 K2 = g1 K1 + g2 K2 , where gi = bi si . Let u denote the control variable representing the fraction of investment allocated to Region 1 with the remainder going to Region 2. Clearly, 0 ≤ u ≤ 1, (2.81) and K˙ 1 = u(g1 K1 + g2 K2 ), K1 (0) = a1 > 0,

(2.82)

K˙ 2 = (1 − u)(g1 K1 + g2 K2 ), K2 (0) = a2 > 0.

(2.83)

The optimal control problem is to maximize the productivity of the whole economy at time T. Thus, the objective is: max{J = b1 K1 (T ) + b2 K2 (T )} subject to (2.81), (2.82), and (2.83). (a) Use the maximum principle to derive the form of the optimal policy. (b) Assume b2 > b1 . Show that u∗ (t) = 0 for t ∈ [tˆ, T ], where tˆ is a switching point and 0 ≤ tˆ < T. (c) If you are ambitious, ﬁnd the tˆ of part (b). E 2.23 Investment Allocation. Let K denote the capital stock and λK its output rate with λ > 0. For simplicity in notation, we set the productivity factor λ = 1. Let u denote the invested fraction of the output. Then, uK is the investment rate and (1 − u)K is the consumption rate. Let us assume an exponential utility 1 − e−C of consumption C. Solve the resulting optimal control problem: T −(1−u(t))K(t) [1 − e ]dt max J = 0

subject to ˙ K(t) = u(t)K(t), K(0) = K0 , K(T ) free, 0 ≤ u(t) ≤ 1, 0 ≤ t ≤ T. Assume T > 1 and 0 < K0 < 1 − e1−T . Obtain explicitly the optimal investment allocation u∗ (t), optimal capital K ∗ (t), and the adjoint variable λ(t), 0 ≤ t ≤ T.

66

2. The Maximum Principle: Continuous Time

E 2.24 The rate at which a new product can be sold at any time t is f (p(t))g(Q(t)) where p is the price and Q is cumulative sales. We assume f (p) < 0; sales vary inversely with price. Also g (Q) ≷ 0 for Q ≶ Q1 , respectively, where Q1 > 0 is a constant known as the saturation level. For a given price, current sales grow with past sales in the early stages as people learn about the good from past purchasers. But as cumulative sales increase, there is a decline in the number of people who have not yet purchased the good. Eventually the sales rate for any given price falls, as the market becomes saturated. The unit production cost c may be constant or may decline with cumulative sales if the ﬁrm learns how to produce less expensively with experience: c = c(Q), c (Q) ≤ 0. Formulate and solve the optimal control problem in order to characterize the price policy p(t), 0 ≤ t ≤ T, that maximizes proﬁts from this new “fad” over a ﬁxed horizon T. Speciﬁcally, show that in marketing a new product, its optimal price rises while the market expands to its saturation level and falls as the market matures beyond the saturation level. E 2.25 Suppose H(x, u, λ, t) = λux − 12 u2 and Ω(t) = [0, 1] for all t. (a) Show that the form of the optimal control is given by the function ⎧ ⎪ ⎪ ⎪ λx if 0 ≤ λx ≤ 1, ⎪ ⎪ ⎨ u∗ (x, λ) = sat[0, 1; λx] = 1 if λx > 1, ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 0 if λx < 0. (b) Verify that (2.63) holds for all values of x and λ. E 2.26 Show that the derived Hamiltonians H 0 found in Examples 2.4 and 2.6 satisfy the concavity condition required for the suﬃciency result in Sect. 2.4. E 2.27 If F and f are concave in x and u and if λ(t) ≥ 0, then show that the derived Hamiltonian H 0 is concave in x. Note that the concavity of F and f are easier to check than the concavity of H 0 as required in Theorem 2.1 on suﬃciency conditions. E 2.28 A simple controlled dynamical system is modeled by the scalar equation x˙ = x + u.

Exercises for Chapter 2

67

The ﬁxed-end-point optimal control problem consists in steering x(t) from an initial state x(0) = x0 to the target x(1) = 0, such that J(u) =

1 4

1

u4 dt

0

is minimized. Use the maximum principle to show that the optimal control is given by u∗ (t) =

4x0 −4/3 (e − 1)−1 e−t/3 . 3

E 2.29 Perform the following: (a) Solve the optimal consumption problem of Example 1.3 with U (C) = ln C and B = 0. Hint: Since C(t) ≥ 0, we can replace the state constraint W (t) ≥ 0, t ∈ [0, T ], by the terminal condition W (T ) = 0, and then use the transversality condition given in (2.75). (b) Find the rate of change of optimal consumption over time and conclude that consumption remains constant when r = ρ, increases when r > ρ, and decreases when r < ρ. E 2.30 Perform the following: (a) Formulate the TPBVP (2.32) and its discrete version for the problem in Example 2.8, but with a new initial condition x(0) = 1. (b) Solve the discrete version of the TPBVP by using Excel. E 2.31 Solve explicitly max J = −

2

x(t)dt 0

subject to x(t) ˙ = u(t), x(0) = 1, x(2) = 0, −a ≤ u(t) ≤ b, a > 1/2, b > 0. Obtain optimal x∗ (t), u∗ (t), and all required multipliers.

Chapter 3

The Maximum Principle: Mixed Inequality Constraints The problems to which the maximum principle derived in the previous chapter was applicable had constraints involving only the control variables. We will see that in many applied models it is necessary to impose constraints involving both control and state variables. Inequality constraints involving control and possibly state variables are called mixed inequality constraints. In the solution spaces of problems with mixed constraints, there may be regions in which one or more of the constraints is tight. When this happens, the system must be controlled in such a way that the tight constraints are not violated. As a result, the maximum principle of Chap. 2 must be revised so that the Hamiltonian is maximized subject to the constraints. This is done by appending the Hamiltonian with the mixed constraints and the associated Lagrange multipliers to form a Lagrangian, and then setting the derivatives of the resulting Lagrangian with respect to the control variables to zero.

© Springer Nature Switzerland AG 2019 S. P. Sethi, Optimal Control Theory, https://doi.org/10.1007/978-3-319-98237-3 3

69

70

3. The Maximum Principle: Mixed Inequality Constraints

In Sect. 3.1, a Lagrangian form of the maximum principle is discussed for models in which there are some constraints that involve only control variables, and others that involve both state and control variables simultaneously. Problems having pure state variable inequality constraints, i.e., those involving state variables but no control variables, are more diﬃcult and will be dealt with in Chap. 4. In Sect. 3.2, we state conditions under which the Lagrangian maximum principle is also suﬃcient for optimality. Economists frequently analyze optimal control problems involving a discount rate. By combining the discount factor with the adjoint variables and the Lagrange multipliers and making suitable changes in the deﬁnitions of the Hamiltonian and Lagrangian functions, it is possible to derive the current-value formulation of the maximum principle as described in Sect. 3.3. It is often the case in ﬁnite horizon problems that some restrictions are imposed on the state variables at the end of the horizon. In Sect. 3.4, we discuss the transversality conditions to be satisﬁed by the adjoint variable in special cases of interest. Section 3.5 is devoted to the study of free terminal time problems where the terminal time itself is a decision variable to be determined. Models with inﬁnite horizons and their stationary equilibrium solutions are covered in Sect. 3.6. Section 3.7 presents a classiﬁcation of a number of the most important and commonly used kinds of optimal control models, together with a brief description of the forms of their optimal solutions. The reader may wish to refer to this section from time to time while working through later chapters in the book.

3.1

A Maximum Principle for Problems with Mixed Inequality Constraints

We will state the maximum principle for optimal control problems with mixed inequality constraints without proof. For further details see Pontryagin et al. (1962), Hestenes (1966), Arrow and Kurz (1970), Hadley and Kemp (1971), Bensoussan et al. (1974), Feichtinger and Hartl (1986), Seierstad and Sydsæter (1987), and Grass et al. (2008). Let the system under consideration be described by the following vector diﬀerential equation x˙ = f (x, u, t), x(0) = x0

(3.1)

3.1. A Maximum Principle for Problems with Mixed Constraints

71

given the initial conditions x0 and a control trajectory u(t), t ∈ [0, T ], T > 0, where T can be the terminal time to be optimally determined or given as a ﬁxed positive number. Note that in the above equation, x(t) ∈ E n and u(t) ∈ E m , and the function f : E n ×E m ×E 1 → E n is assumed to be continuously diﬀerentiable. Let us consider the following objective: T F (x, u, t)dt + S[x(T ), T ] , (3.2) max J = 0

where F : E n × E m × E 1 → E 1 and S : E n × E 1 → E 1 are continuously diﬀerentiable functions and where T denotes the terminal time. Depending on the situation being modeled, the terminal time T may be given or to be determined. In the case when T is given, the function S(x(T ), T ) should be viewed as merely a function of the terminal state, and can be revised as S(x(T )). Next we impose constraints on state and control variables. Speciﬁcally, for each t ∈ [0, T ], x(t) and u(t) must satisfy g(x, u, t) ≥ 0, t ∈ [0, T ],

(3.3)

where g: E n × E m × E 1 → E q is continuously diﬀerentiable in all its arguments and must contain terms in u. An important special case is that of controls having an upper bound that depends on the current state, i.e., u(t) ≤ M (x(t)), t ∈ [0, T ], which can be written as M (x) − u ≥ 0. Inequality constraints without terms in u will be introduced later in Chap. 4. It is important to note that the mixed constraints (3.3) allow for inequality constraints of the type g(u, t) ≥ 0 as special cases. Thus, the control constraints of the form u(t) ∈ Ω(t) treated in Chap. 2 can be subsumed in (3.3), provided that they can be expressed in terms of a ﬁnite number of inequality constraints of the form g(u, t) ≥ 0. In most problems that are of interest to us, this will indeed be the case. Thus, from here on, we will formulate control constraints either directly as inequality constraints and include them as parts of (3.3), or as u(t) ∈ Ω(t), which can be easily converted into a set of inequality constraints to be included as parts of (3.3).

72

3. The Maximum Principle: Mixed Inequality Constraints

Finally, the terminal state is constrained by the following inequality and equality constraints: a(x(T ), T ) ≥ 0,

(3.4)

b(x(T ), T ) = 0,

(3.5)

where a : E n × E 1 → E la and b : E n × E 1 → E lb are continuously diﬀerentiable in all their arguments. Clearly, a and b are not functions of T, if T is a given ﬁxed number. In the speciﬁc cases when T is given, the terminal state constraints will be written as a(x(T )) ≥ 0 and b(x(T )) = 0. Important special cases of (3.4) are x(T ) ≥ k. We can now deﬁne a control u(t), t ∈ [0, T ], or simple u, to be admissible if it is piecewise continuous and if, together with its corresponding state trajectory x(t), t ∈ [0, T ], it satisﬁes the constraints (3.3), (3.4), and (3.5). At times we may ﬁnd terminal inequality constraints given as x(T ) ∈ Y (T ) ⊂ X(T ),

(3.6)

where Y (T ) is a convex set and X(T ) is the set of all feasible terminal states, also called the reachable set from the initial state x0 , i.e., X(T ) = {x(T ) | x(T ) obtained by an admissible control u and (3.1)}. Remark 3.1 The feasible set deﬁned by (3.4) and (3.5) need not be convex. Thus, if the convex set Y (T ) can be expressed by a ﬁnite number of inequalities a(x(T ), T ) ≥ 0 and equalities b(x(T ), T ) = 0, then (3.6) becomes a special case of (3.4) and (3.5). In general, (3.6) is not a special case of (3.4) and (3.5), since it may not be possible to deﬁne a given Y (T ) by a ﬁnite number of inequalities and equalities. In this book, we will only deal with problems in which the following full-rank conditions hold. That is, rank[∂g/∂u, diag(g)] = q holds for all arguments x(t), u(t), t, that could arise along an optimal solution, and ⎤ ⎡ ⎢ ∂a/∂x diag(a) ⎥ rank ⎣ ⎦ = l a + lb 0 ∂b/∂x

3.1. A Maximum Principle for Problems with Mixed Constraints

73

hold for all possible values of x(T ) and T. The ﬁrst of these conditions means that the gradients with respect to u of all active constraints in (3.3) must be linearly independent. Similarly, the second condition means that the gradients with respect to x of the equality constraints (3.5) and of the active inequality constraints in (3.4) must be linearly independent. These conditions are also referred to as the constraint qualiﬁcations. In cases when these do not hold, see Seierstad and Sydsæter (1987) for details on weaker constraint qualiﬁcations. Before proceeding further, let us recapitulate the optimal control problem under consideration in this chapter: ⎧ T ⎪ ⎪ ⎪ F (x, u, t)dt + S[x(T ), T ] , max J = ⎪ ⎪ ⎪ 0 ⎪ ⎪ ⎪ ⎪ ⎪ subject to ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎨ x˙ = f (x, u, t), x(0) = x , 0 (3.7) ⎪ ⎪ ⎪ g(x, u, t) ≥ 0, ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ a(x(T ), T ) ≥ 0, ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ b(x(T ), T ) = 0. To state the maximum principle we deﬁne the Hamiltonian function H : E n × E m × E n × E 1 → E 1 as H(x, u, λ, t) := F (x, u, t) + λf (x, u, t),

(3.8)

(a row vector). We also deﬁne the Lagrangian function where λ ∈ n m L : E × E × E n × E q × E 1 → E 1 as En

L(x, u, λ, μ, t) := H(x, u, λ, t) + μg(x, u, t),

(3.9)

where μ ∈ E q is a row vector, whose components are called Lagrange multipliers. These Lagrange multipliers satisfy the complementary slackness conditions μ ≥ 0, μg(x, u, t) = 0, which, in view of (3.3), can be expressed equivalently as μi ≥ 0, μi gi (x, u, t) = 0,

i = 1, 2, . . . , q.

The adjoint vector satisﬁes the diﬀerential equation λ˙ = −Lx (x, u, λ, μ, t)

(3.10)

74

3. The Maximum Principle: Mixed Inequality Constraints

with the terminal condition ⎧ ⎪ ⎨ la(T ) = Sx (x(T ), T ) + αax (x(T ), T ) + βbx (x(T ), T ), ⎪ ⎩

(3.11)

α ≥ 0, αa(x(T ), T ) = 0,

where α ∈ E la and β ∈ E lb are constant vectors. The maximum principle states that the necessary conditions for u∗ , with the corresponding state trajectory x∗ , to be an optimal control are that there should exist continuous and piecewise continuously diﬀerentiable functions λ, piecewise continuous functions μ, and constants α and β such that (3.12) holds, i.e., x˙ ∗ = f (x∗ , u∗ , t), x∗ (0) = x0 , satisfying the terminal constraints a(x∗ (T ), T ) ≥ 0 and b(x∗ (T ), T ) = 0, λ˙ = −Lx (x∗ , u∗ , λ, μ, t) with the terminal conditions λ(T ) = Sx (x∗ (T ), T ) + αax (x∗ (T ), T ) + βbx (x∗ (T ), T ), α ≥ 0, αa(x∗ (T ), T ) = 0, the Hamiltonian maximizing condition H[x∗ (t), u∗ (t), λ(t), t] ≥ H[x∗ (t), u, λ(t), t] at each t ∈ [0, T ] for all u satisfying g[x∗ (t), u, t] ≥ 0, and the Lagrange multipliers μ(t) are such that ∂H ∂L ∂g ∗ ∗ | +μ | = =0 ∂u u=u (t) ∂u ∂u u=u (t) and the complementary slackness conditions μ(t) ≥ 0, μ(t)g(x∗ , u∗ , t) = 0 hold.

(3.12)

3.1. A Maximum Principle for Problems with Mixed Constraints

75

In the case of the terminal constraint (3.6), note that the terminal conditions on the state and the adjoint variables in (3.12) will be replaced, respectively, by x∗ (T ) ∈ Y (T ) ⊂ X(T )

(3.13)

and [λ(T ) − Sx (x∗ (T ), T )][y − x∗ (T )] ≥ 0,

∀y ∈ Y (T ).

(3.14)

In Exercise 3.5, you are asked to derive (3.14) from (3.12) in the one ¯] for each T > 0, where x and dimensional case when Y (T ) = Y = [x, x x ¯ are two constants such that x ¯ > x. In the case when the terminal time T ≥ 0 in the problem (3.10) is also a decision variable, there is an additional necessary transversality condition for T ∗ to be optimal, namely, H[x∗ (T ∗ ), u∗ (T ∗ ), λ(T ∗ ), T ∗ ] + ST [x∗ (T ∗ ), T ∗ ] +αaT [x∗ (T ∗ ), T ∗ ] + β T [x∗ (T ∗ ), T ∗ ] = 0,

(3.15)

provided T ∗ is an interior solution, i.e., T ∗ ∈ (0, ∞). In other words, optimal T ∗ and x∗ (t), u∗ (t), t ∈ [0, T ∗ ], must satisfy (3.12) with T replaced by T ∗ and (3.15). This condition will be further discussed and illustrated with examples in Sect. 3.5. The discussion will also include the case when T is restricted to lie in the interval [T1 , T2 ], T2 > T1 ≥ 0. We will now illustrate the use of the maximum principle (3.12) by solving a simple example. Example 3.1 Consider the problem:

max J =

1

udt

0

subject to x˙ = u, x(0) = 1,

(3.16)

u ≥ 0, x − u ≥ 0.

(3.17)

Note that constraints (3.17) are of the mixed type (3.3). They can also be rewritten as 0 ≤ u ≤ x.

76

3. The Maximum Principle: Mixed Inequality Constraints

Solution The Hamiltonian is H = u + λu = (1 + λ)u, so that the optimal control has the form u∗ (x, λ) = bang[0, x; 1 + λ].

(3.18)

To get the adjoint equation and the multipliers associated with constraints (3.17), we form the Lagrangian: L = H + μ1 u + μ2 (x − u) = μ2 x + (1 + λ + μ1 − μ2 )u. From this we get the adjoint equation ∂L = −μ2 , λ(1) = 0. λ˙ = − ∂x

(3.19)

Also note that the optimal control must satisfy ∂L = 1 + λ + μ1 − μ2 = 0, ∂u

(3.20)

and μ1 and μ2 must satisfy the complementary slackness conditions μ1 ≥ 0, μ1 u = 0,

(3.21)

μ2 ≥ 0, μ2 (x − u) = 0.

(3.22)

It is reasonable in this simple problem to guess that u∗ (t) = x(t) is an optimal control for all t ∈ [0, 1]. We now show that this control satisﬁes all the conditions of the Lagrangian form of the maximum principle. Since x(0) = 1, the control u∗ = x gives x = et as the solution of (3.16). Because x = et > 0, it follows that u∗ = x > 0. Thus, μ1 = 0 from (3.21). From (3.20) we then have μ2 = 1 + λ. Substituting this into (3.19) and solving gives 1 + λ(t) = e1−t .

(3.23)

Since the right-hand side of (3.23) is always positive, u∗ = x satisﬁes (3.18). Notice that μ2 = e1−t ≥ 0 and x − u∗ = 0, so (3.22) holds.

3.1. A Maximum Principle for Problems with Mixed Constraints

77

Using u∗ = x in (3.16), we can obtain the optimal state trajectory x∗ (t) = et . Thus, the optimal value of the objective function is 1 ∗ J = et dt = (e − 1). 0

Let us now examine the consequence of changing the constraint x − u ≥ 0 on control u to x − u ≥ −ε, which gives u ≤ x + ε for a small ε. In this case, it is clear that the optimal control u∗ = x + ε, which we can use in (3.16) to obtain x∗ (t) = et (1 + ε) − ε. The optimal value of the objective function changes to 1 1 u(t)dt = et (1 + ε)dt = (e − 1)(1 + ε). 0

0

J ∗ increases by (e − 1)ε, which in this case equals 1This means that 1 1−t ε 0 μ2 (t)dt = ε 0 e dt, as stipulated in Remark 3.8. We conclude Sect. 3.1 with the following remarks. Remark 3.2 Strictly speaking, we should have H = λ0 F + λf in (3.8) with (λ0 , λ(t)) = (0, 0) for all t ∈ [0, T ]. However, when λ0 = 0, the conditions in the maximum principle do not change if we replace F by any other function. Therefore, the problems where the maximum principle holds only with λ0 = 0 are termed abnormal. Such problems may arise when there are terminal state constraints such as (3.4) and (3.5) or pure state constraints treated in Chap. 4. In this book, as is standard in the economics literature dealing with optimal control theory, we will set λ0 = 1. This is because the problems that are of interest to us will be normal. For examples of abnormal problems and further discussion on this issue, see Seierstad and Sydsæter (1987). Remark 3.3 The function deﬁned in (3.9) is not a Lagrangian function in the sense of the continuous-time counterpart of the Lagrangian function deﬁned in (8.45) in Chap. 8. However, it can be viewed, roughly speaking, as a Lagrangian function associated with the problem of maximizing the Hamiltonian (3.8) subject to the constraints (3.3) along the optimal path. As in this book, some people refer to (3.9) as a Lagrangian function, while others call it an extended Pontryagin function. Remark 3.4 It should be pointed out that if the set Y in (3.6) consists of a single point Y = {k}, making the problem a ﬁxed-end-point problem, then the transversality condition reduces to simply λ(T ) to equal

78

3. The Maximum Principle: Mixed Inequality Constraints

a constant to be determined, since x∗ (T ) = k. In this case the salvage function S becomes a constant, and can therefore be disregarded. When Y = X, the terminal condition in (3.12) reduces to (2.30). Further discussion of the terminal conditions can be found in Sect. 3.4 along with a summary in Table 3.1. Remark 3.5 As in Chap. 2, it can be shown that λi (t), i = 1, 2, ..., n, is interpreted as the marginal value of an increment in the state variable xi at time t. Speciﬁcally, the relation (2.17) holds so long as the value function V (x, t), deﬁned in (2.10), is continuously diﬀerentiable in xi ; see Seierstad and Sydsæter (1987). Remark 3.6 The Lagrange multiplier αi , i = 1, 2, . . . , n represents the shadow price associated with the terminal state constraint ai (x(T ), T ) ≥ 0. Thus, if we change this constraint to ai (x(T ), T ) ≥ ε for a small ε, then the change in the objective function will be −εαi + o(ε). A similar interpretation holds for the multiplier β; see Sect. 3.4 for further discussion. This will be illustrated in Example 3.4 and Exercise 3.17. Remark 3.7 In the case when the terminal constraint (3.4) or (3.5) is binding, the transversality condition λ(T ) in (3.12) should be viewed as the left-hand limit, limt↑T λ(t), sometimes written as λ(T − ), and then we would express λ(T ) = Sx (x∗ (T ), T ). However, the standard practice for problems treated in Chaps. 2 and 3 is to use the notation that we have used. Nevertheless, care should be exercised in distinguishing the marginal value of the state at time T given by Sx (x∗ (T ), T ) and the shadow prices for the terminal constraints (3.4) and (3.5) given by α and β, respectively. See Sect. 3.4 and Example 3.4 for further elaboration. Remark 3.8 It is also possible to provide marginal value interpretations to Lagrange multipliers μi , i = 1, 2, . . . , m. If we change the constraint gi (x, u, t) ≥ 0 to gi (x, u, t) ≥ ε for a small ε, then we expect the change T in the optimal value of the objective function to be −ε 0 μi (t)dt + o(ε); see Peterson (1973, 1974) or Malanowski (1984). If ε < 0, then the T constraint is being relaxed, and 0 μi (t)dt ≥ 0 provides the marginal value of relaxing the constraint. We will illustrate this concept with the help of Example 3.1. Remark 3.9 In the case when the problem (3.7) is changed by interchanging x(T ) and x(0) so that the initial condition x(0) = x0 is replaced by x(T ) = xT , and S(x(T ), T ), a(x(T ), T ) and b(x(T ), T ) are

3.2. Suﬃciency Conditions

79

replaced by S(x(0)), a(x(0)) and b(x(0)), respectively, then in the maximum principle (3.12), we need to replace initial condition x∗ (0) = x0 by x∗ (T ) = xT and the terminal condition on the adjoint variable λ by the initial condition λ(0) = Sx (x∗ (0)) + αax (x∗ (0)) + βbx (x∗ (0)) with α ≥ 0 and αa(x∗ (0)) = 0.

3.2

Suﬃciency Conditions

In this section we will state, without proof, a number of suﬃciency results. These results require the concepts of concave and quasiconcave functions. Recall from Sect. 1.4 that with D ⊂ E n , a convex set, a function ψ : D → E 1 is concave, if for all y, z ∈ D and for all p ∈ [0, 1], ψ(py + (1 − p)z) ≥ pψ(y) + (1 − p)ψ(z).

(3.24)

The function ψ is quasiconcave if (3.24) is relaxed to ψ(py + (1 − p)z) ≥ min{ψ(y), ψ(z)},

(3.25)

and ψ is strictly concave if y = z and p ∈ (0, 1) and (3.24) holds with a strict inequality. Furthermore, ψ is convex, quasiconvex, or strictly convex if −ψ is concave, quasiconcave, or strictly concave, respectively. Note that linearity implies both concavity and convexity, and concavity implies quasiconcavity. For further details on the properties of such functions, see Mangasarian (1969). We can now state a suﬃciency result concerning the problem with mixed constraints stated in (3.7). For this purpose, let us deﬁne the maximized Hamiltonian H 0 (x, λ, t) =

max

{u|g(x,u,t)≥0}

H(x, u, λ, t).

(3.26)

Theorem 3.1 Let (x∗ , u∗ , λ, μ, α, β) satisfy the necessary conditions in (3.12). If H 0 (x, λ(t), t) is concave in x at each t ∈ [0, T ], S in (3.2) is concave in x, g in (3.3) is quasiconcave in (x, u), a in (3.4) is quasiconcave in x, and b in (3.5) is linear in x, then (x∗ , u∗ ) is optimal. The result is a straightforward extension of Theorem 2.1. See, e.g., Seierstad and Sydsæter (1977, 1987) and Feichtinger and Hartl (1986). In Exercise 3.7 you are asked to check these suﬃciency conditions for Example 3.1.

80

3. The Maximum Principle: Mixed Inequality Constraints

3.3

Current-Value Formulation

In most management science and economics problems, the objective function is usually formulated in terms of money or utility. These quantities have time value, and therefore the future streams of money or utility are discounted. The discounted objective function can be written as a special case of (3.2) by assuming that the time dependence of the relevant functions comes only through the discount factor. Thus, F (x, u, t) = φ(x, u)e−ρt and S(x, T ) = ψ(x)e−ρT ,

(3.27)

where we assume the discount rate ρ > 0. We should also mention that if F (x, u, t) = φ(x, u, t)e−ρt and S(x, T ) = ψ(x, T )e−ρT , then there is no advantage of developing a current-value version of the maximum principle, and it is recommended that the present-value formulation be used in this case. Now, the objective in problem (3.7) can be written as: T −ρt −ρT max J = φ(x, u)e dt + ψ[x(T )]e . (3.28) 0

For this problem, the Hamiltonian, which we shall now refer to as the present-value Hamiltonian, H pv , is H pv := e−ρt φ(x, u) + λpv f (x, u, t)

(3.29)

and the present-value Lagrangian is Lpv := H pv + μpv g(x, u, t)

(3.30)

with the present-value adjoint variables λpv and present-value multipliers αpv and β pv satisfying pv λ˙ = −Lpv (3.31) x , λpv (T ) = Sx [x(T ), T ] + αpv ax (x(T ), T ) + β pv bx (x(T ), T ) = e−ρT ψ x [x(T )] + αpv ax (x(T ), T ) + β pv bx (x(T ), T ), (3.32) αpv ≥ 0, αpv a(x(T ), T ) = 0,

(3.33)

μpv ≥ 0, μpv g = 0.

(3.34)

and μpv satisfying We use superscript pv in this section to distinguish these from the current-value functions deﬁned as follows. Elsewhere, we do not need to

3.3. Current-Value Formulation

81

make the distinction explicitly since we will either be using the presentvalue deﬁnitions or the current-value deﬁnitions of these functions. The reader will always be able to tell what is meant from the context. We now deﬁne the current-value Hamiltonian H[x, u, λ, t] := φ(x, u) + λf (x, u, t)

(3.35)

and the current-value Lagrangian L[x, u, λ, μ, t] := H + μg(x, u, t).

(3.36)

To see why we can do this, we note that if we deﬁne λ := eρt λpv and μ := eρt μpv ,

(3.37)

we can rewrite (3.29) and (3.30) as H = eρt H pv and L = eρt Lpv .

(3.38)

Since eρt > 0, maximizing H pv with respect to u at time t is equivalent to maximizing the current-value Hamiltonian H with respect to u at time t. Furthermore, from (3.37), pv λ˙ = ρeρt λpv + eρt λ˙ .

(3.39)

The ﬁrst term on the right-hand side of (3.39) is simply ρλ using the deﬁnition in (3.37). To simplify the second term we use the diﬀerential equation (3.31) for λpv and the fact that Lx = eρt Lpv x from (3.38). Thus, λ˙ = ρλ − Lx , λ(T ) = ψ x [x(T )] + αax (x(T ), T ) + βbx (x(T ), T ),

(3.40)

where the terminal condition for λ(T ) follows immediately from the terminal condition for λpv (T ) in (3.32), the deﬁnition (3.38), α = eρt αpv

and

β = eρt β pv .

(3.41)

The complementary slackness conditions satisﬁed by the currentvalue Lagrange multipliers μ and α are μ ≥ 0, μg = 0, α ≥ 0, and αa = 0 on account of (3.33), (3.34), (3.37), and (3.41).

82

3. The Maximum Principle: Mixed Inequality Constraints

We will now state the maximum principle in terms of the currentvalue functions. It states that the necessary conditions for u∗ , with the corresponding state trajectory x∗ , to be an optimal control are that there exist λ and μ such that the conditions (3.42) hold, i.e., x˙ ∗ = f (x∗ , u∗ , t), a(x∗ (T ), T ) ≥ 0, b(x∗ (T ), T ) = 0, λ˙ = ρλ − Lx [x∗ , u∗ , λ, μ, t], with the terminal conditions λ(T ) = ψ x (x∗ (T )) + αax (x∗ (T ), T ) + βbx (x∗ (T ), T ), α ≥ 0, αa(x∗ (T ), T ) = 0, and the Hamiltonian maximizing condition (3.42) H[x∗ (t), u∗ (t), λ(t), t] ≥ H[x∗ (t), u, λ(t), t] at each t ∈ [0, T ] for all u satisfying g[x∗ (t), u, t] ≥ 0, and the Lagrange multipliers μ(t) are such that ∂L ∂u |u=u∗ (t)

= 0, and the complementary slackness

conditions μ(t) ≥ 0 and μ(t)g(x∗ , u∗ , t) = 0 hold. As in Sect. 3.1, when the terminal constraint is given by (3.6) instead of (3.4) and (3.5), we need to replace the terminal condition on the state and the adjoint variables, respectively, by (3.13) and [λ(T ) − ψ x (x∗ (T ))][y − x∗ (T )] ≥ 0, ∀y ∈ Y (T ).

(3.43)

See also Remark 3.4, which applies here as well. If T ≥ 0 is also a decision variable and if T ∗ is the optimal terminal time, then the optimal solution x∗ , u∗ , and T ∗ must satisfy (3.42) with T replaced by T ∗ along with H[x∗ (T ∗ ), u∗ (T ∗ ), λ(T ∗ ), T ∗ ] − ρψ[x∗ (T ∗ )] +αaT [x∗ (T ∗ ), T ∗ ] + β T [x∗ (T ∗ ), T ∗ ] = 0.

(3.44)

3.3. Current-Value Formulation

83

You are asked in Exercise 3.8 to show that (3.44) is the current-value version of (3.15) under the relation (3.27). Furthermore, show how (3.44) should be modiﬁed if S(x, T ) = ψ(x, T )e−ρT in (3.27). As for the suﬃciency conditions for the current-value formulation, one can simply use Theorem 3.1 as if it were stated for the current-value formulation. Example 3.2 We illustrate an application of the current-value maximum principle by solving the consumption problem of Example 1.3 with U (C) = ln C and W (T ) = 0. Thus, we solve max

C(t)≥0

T

e

J=

−ρt

ln C(t)dt + B(0)e

−ρT

0

subject to the wealth dynamics ˙ = rW − C, W (0) = W0 , W (T ) = 0, W where W0 > 0. As hinted in Exercise 2.29(a), we do not need to impose the pure state constraint W (t) ≥ 0, t ∈ [0, T ], in view of C(t) ≥ 0, t ∈ [0, T ], and W (T ) = 0. Also, the salvage function reduces to B(0), which is a constant; see Remark 3.4. Solution In Exercise 2.29(a) we used the standard Hamiltonian formulation to solve the problem. We now demonstrate the use of the current-value Hamiltonian formulation: H = ln C + λ(rW − C),

(3.45)

with the adjoint equation ∂H λ˙ = ρλ − = (ρ − r)λ, λ(T ) = β, ∂W

(3.46)

where β is some constant to be determined. The solution of (3.46) is λ(t) = βe(ρ−r)(t−T ) .

(3.47)

To ﬁnd the optimal control, we maximize H by diﬀerentiating (3.45) with respect to C and setting the result to zero: 1 ∂H = − λ = 0, ∂C C

84

3. The Maximum Principle: Mixed Inequality Constraints

which implies 1 1 = e(ρ−r)(T −t) . λ(t) β

C ∗ (t) =

(3.48)

Using this consumption level in the wealth dynamics gives ˙ (t) = rW (t) − 1 e(ρ−r)(T −t) , W (0) = W0 , W β which can be solved as ∗

W (t) = e

rt

W0 −

e(ρ−r)T (1 − e−ρt ) . ρβ

(3.49)

Setting W ∗ (T ) = 0 gives β = e(ρ−r)T (1 − e−ρT )/ρW0 . Therefore, the optimal consumption rate and wealth at time t are −ρt

e − e−ρT ρW0 e(r−ρ)t ∗ rt C (t) = , W (t) = e W0 . 1 − e−ρT 1 − e−ρT ∗

(3.50)

The optimal value of the objective function is

1 − e−ρT r−ρ 1 1 ρW0 −ρT −e + T+ ln + B(0)e−ρT . J = ρ 1 − e−ρT ρ ρ ρ (3.51) The interpretation of the current-value functions are that these functions reﬂect the values at time t in terms of the current (or, time-t) dollars. The standard functions, on the other hand, reﬂect the values at time t in terms of time-zero dollars. For example, the standard adjoint variable λpv (t) can be interpreted as the marginal value per unit increase in the state at time t, in the same units as that of the objective function (3.28), i.e., in terms of time-zero dollars; see Sect. 2.2.4. On the other hand, λ(t) = eρt λpv (t) is obviously the same value expressed in terms of current (or, time-t) dollars. For the consumption problem of Example 3.2, note that the currentvalue adjoint function ∗

λ(t) = e(ρ−r)t (1 − e−ρT )/ρW0 .

(3.52)

This gives the marginal value per unit increase in wealth at time t in time-t dollars. In Exercise 2.29(a), the standard adjoint variable was λpv (t) = e−rt (1 − e−ρT )/ρW0 , which can be written as λpv (t) = e−ρt λ(t).

3.3. Current-Value Formulation

85

Thus, it is clear that λpv (t) expresses the same marginal value in timezero dollars. In particular, dJ ∗ /dW0 = (1 − e−ρT )/ρW0 = λ(0) = λpv (0) gives the marginal value per unit increase in the initial wealth W0 . In Exercise 3.11, you are asked to formulate and solve a consumption problem of an economy. The problem is a linear version of the famous Ramsey model; see Ramsey (1928) and Feichtinger and Hartl (1986, p. 201). Before concluding this section on the current-value formulation, let us also provide the current-value version of the HJB equation (2.15) or (2.19) along with the terminal condition (2.16). As in (2.9), we now deﬁne the value function for the problem (3.7), with its objective function replaced by (3.28), as follows:

T −ρ(T −t) V (x, t) = max φ(x(s), u(s))ds + e ψ(x(T )) {u|g(x,u,t)≥0}

t

if x(T ) satisﬁes a(x(T ), T ) ≥ 0 and b(x(T ), T ) = 0, and V (x, t) = −∞, otherwise. (3.53) Then proceeding as in Sect. 2.1.1, we have " ! V (x, t)= max φ[x(τ ), u(τ )]dτ + e−ρδdt V [x(t + δt), t + δt] . {u(τ )|g(x(τ ),u(τ ),τ )≥0}

τ ∈[t, t+δt]

(3.54) Noting that e−ρδt = 1−ρδt+0(δt) and continuing on as in Sect. 2.1.1, we can obtain the current-value version of (2.15) and (2.19) as ρV (x, t) = =

max

{φ(x, u, t) + Vx (x, t)f (x, u, t) + Vt (x, t)}

max

{H(x, u, Vx , t) + Vt } = 0,

{u|g(x,u,t)≥0} {u|g(x,u,t)≥0}

(3.55) where H is deﬁned as in (3.35). Finally, we can write the terminal condition as ⎧ ⎪ ⎨ ψ(x), if a(x, T ) ≥ 0 and b(x, T ) = 0, V (x, T ) = ⎪ ⎩ −∞, otherwise.

(3.56)

86

3.4

3. The Maximum Principle: Mixed Inequality Constraints

Transversality Conditions: Special Cases

Terminal conditions on the adjoint variables, also known as transversality conditions, are extremely important in optimal control theory. Because the salvage value function ψ(x) is known, we know the marginal value per unit change in the state at terminal time T. Since λ(T ) must be equal to this marginal value, it provides us with the boundary conditions for the diﬀerential equations for the adjoint variables. We will now derive the terminal or transversality conditions for the current-value adjoint variables for some important special cases of the general problem treated in Sect. 3.3. We also summarize these conditions in Table 3.1. Case 1: Free-end point. In this case, we do not put any constraints on the terminal state x(T ). Thus, x(T ) ∈ X(T ). From the terminal conditions in (3.42), it is obvious that for the free-end-point problem, i.e., when Y (T ) = X(T ), λ(T ) = ψx [x∗ (T )].

(3.57)

This includes the condition λ(T ) = 0 in the special case of ψ(x) ≡ 0; see Example 3.1, speciﬁcally (3.19). These conditions are repeated in Table 3.1, Row 1. The economic interpretation of λ(T ) is that it equals the marginal value of a unit increment in the terminal state evaluated at its optimal value x∗ (T ). Case 2: Fixed-end point. In this case, which is the other extreme from the free-end-point case, the terminal constraint is b(x(T ), T ) = x(T ) − k = 0, and the terminal conditions in (3.42) do not provide any information for λ(T ). However, as mentioned in Remark 3.4 and recalled subsequently in connection with (3.42), λ(T ) will be some constant β, which will be determined by solving the boundary value problem, where the system of diﬀerential equations consists of the state equations with both initial and terminal conditions and the adjoint equations with no boundary conditions. This condition is repeated in Table 3.1, Row 2. Example 3.2 solved in the previous section illustrates this case.

3.4. Transversality Conditions: Special Cases

87

The economic interpretation of λ(T ) = β is as follows. The constant β times ε, i.e., βε, provides the value that could be lost if the ﬁxed-end point were speciﬁed to be k + ε instead of k; see Exercise 3.12. Case 3: Lower bound. Here we restrict the ending value of the state variable to be bounded from below, namely, a(x(T ), T ) = x(T ) − k ≥ 0, where k ∈ X. In this case, the terminal conditions in (3.42) reduce to

and

λ(T ) ≥ ψ x [x∗ (T )]

(3.58)

{λ(T ) − ψ x [x∗ (T )]}{x∗ (T ) − k} = 0,

(3.59)

with the recognition that the shadow price of the inequality constraint (3.4) is (3.60) α = λ(T ) − ψ x [x∗ (T )] ≥ 0. For ψ(x) ≡ 0, these terminal conditions can be written as λ(T ) ≥ 0 and λ(T )[x∗ (T ) − k] = 0.

(3.61)

These conditions are repeated in Table 3.1, Row 3. Case 4: Upper bound. Similarly, when the ending value of the state variable is bounded from above, i.e., when the terminal constraint is k − x(T ) ≥ 0, the conditions for this opposite case are λ(T ) ≤ ψ x [x∗ (T )]

(3.62)

and (3.59). These are repeated in Table 3.1, Row 4. Furthermore, (3.62) can be related to the condition on λ(T ) in (3.42) by setting α = ψ x [x∗ (T )] − λ(T ) ≥ 0. Case 5: A general case. A general ending condition is x(T ) ∈ Y (T ) ⊂ X(T ),

(3.63)

88

3. The Maximum Principle: Mixed Inequality Constraints

which is already stated in (3.6). The transversality conditions are speciﬁed in (3.43) and repeated in Table 3.1, Row 5. An important situation which gives rise to a one-sided constraint occurs when there is an isoperimetric or budget constraint of the form T l(x, u, t)dt ≤ K, (3.64) 0

× × → E 1 is assumed to be nonnegative, bounded, where l : and continuously diﬀerentiable, and K is a positive constant representing the amount of a budgeted resource. To see how this constraint can be converted into a lower bound constraint, we deﬁne an additional state variable xn+1 by the state equation En

Em

E1

x˙ n+1 = −l(x, u, t), xn+1 (0) = K, xn+1 (T ) ≥ 0.

(3.65)

We employ the index n + 1 simply because we already have n state variables x = (x1 , x2 , . . . , xn ). Also Eq. (3.65) becomes an additional equation which is added to the original system. In Exercise 3.13 you will be asked to rework the leaky reservoir problem of Exercise 2.18 with an additional isoperimetric constraint on the total amount of water available. Later in Chap. 7, you’ll be asked to solve Exercises 7.10–7.12 involving budgets for advertising expenditures. In Table 3.1, we have summarized all the terminal or transversality conditions discussed previously. In Sect. 3.7 we discuss model types. We will see that, given the initial state x0 , we can completely specify a control model by selecting a model type and a transversality condition. In what follows, we solve two examples with lower bounds on the terminal state illustrating the use of transversality conditions (3.61), also stated in Table 3.1, Row 3. Example 3.3 is a variation of the consumption problem in Example 3.2. It illustrates the use of the transversality conditions (3.61). Example 3.3 Let us modify the objective function of the consumption problem (Example 3.2) to take into account the salvage (bequest) value of terminal wealth. This is the utility to the individual of leaving an estate to his heirs upon death. Let us now assume that T denotes the time of the individual’s death and BW (T ), where B is a positive constant,

3.4. Transversality Conditions: Special Cases

89

denotes his utility of leaving wealth W (T ) to his heirs upon death. Then, the problem is: T −ρt −ρT e ln C(t)dt + e BW (T ) (3.66) max J = C(t)≥0

0

Table 3.1: Summary of the transversality conditions Constraint

Description

λ(T )

1

λ(T ) when ψ ≡ 0

on x(T ) x(T ) ∈ Y (T ) = X(T )

Free-end

λ(T ) = ψ x [x∗ (T )]

λ(T ) = 0

point 2

3

4

x(T ) = k ∈ X(T ),

Fixed-end

λ(T ) = β, a constant

λ(T ) = β, a constant

i.e., Y (T ) = {k}

point

to be determined

to be determined

x(T ) ∈ X(T ) ∩ [k, ∞),

lower

λ(T ) ≥ ψ x [x∗ (T )]

λ(T ) ≥ 0

i.e., Y (T ) = {x|x ≥ k}

bound

and

and

x(T ) ≥ k

{λ(T ) − ψ x [x∗ (T )]}{x∗ (T ) − k} = 0

λ(T )[x∗ (T ) − k] = 0

upper

λ(T ) ≤ ψ x [x∗ (T )]

λ(T ) ≤ 0

x(T ) ∈ X(T ) ∩ (−∞, k], i.e., Y (T ) = {x|x ≤ k}

5

x(T ) ∈ Y (T ) ⊂ X(T )

bound

and

and

x(T ) ≤ k

{λ(T ) − ψ x [x∗ (T )]}{k − x∗ (T )} = 0

λ(T )[k − x∗ (T )] = 0

General

{λ(T ) − ψ x [x∗ (T )]}{y − x∗ (T )} ≥ 0

λ(T )[y − x∗ (T )] ≥ 0

constraints

∀y ∈ Y (T )

∀y ∈ Y (T )

Note 1. In Table 3.1, x(T ) denotes the (column) vector of n state variables and λ(T ) denotes the (row) vector of n adjoint variables at the terminal time T ; X(T ) ⊂ E n denotes the reachable set of terminal states obtained by using all possible admissible controls; and ψ : E n → E 1 denotes the salvage value function Note 2. Table 3.1 will provide transversality conditions for the standard Hamiltonian formulation if we replace ψ with S, and reinterpret λ as being the standard adjoint variable everywhere in the table. Also (3.15) is the standard form of (3.44)

subject to the wealth equation ˙ = rW − C, W (0) = W0 , W (T ) ≥ 0. W

(3.67)

Solution The Hamiltonian for the problem is given in (3.45), and the adjoint equation is given in (3.46) except that the transversality conditions are from Table 3.1, Row 3: λ(T ) ≥ B, [λ(T ) − B]W ∗ (T ) = 0.

(3.68)

In Example 3.2, the value of β, the terminal value of the adjoint variable, was 1 − e−rT . β= rW0 We now have two cases: (i) β ≥ B and (ii) β < B.

90

3. The Maximum Principle: Mixed Inequality Constraints

In case (i), the solution of the problem is the same as that of Example 3.2, because by setting λ(T ) = β and recalling that W ∗ (T ) = 0 in that example, it follows that (3.68) holds. In case (ii), we set λ(T ) = B. Then, by using B in place of β in (3.47)–(3.49), we get λ(t) = Be(ρ−r)(t−T ) , C ∗ (t) = (1/B)e(ρ−r)(T −t) , and e(ρ−r)T (1 − e−ρt ) . (3.69) W ∗ (t) = ert W0 − ρB Since β < B, we can see from (3.49) and (3.69) that the wealth level in case (ii) is larger than that in case (i) at t ∈ (0, T ]. Furthermore, the amount of bequest is W ∗ (T ) = W0 erT −

eρT − 1 > 0. ρB

Note that (3.68) holds for case (ii). Also, if we had used (3.42) instead of Table 3.1, Row 3, we would have λ(T ) = B + α, α ≥ 0, αW ∗ (T ) = 0, equivalently, in place of (3.68). It is easy to see that α = β − B in case (i) and α = 0 in case (ii). Example 3.4 Consider the problem: 2 max J = −xdt 0

subject to x˙ = u, x(0) = 1, x(2) ≥ 0,

(3.70)

− 1 ≤ u ≤ 1.

(3.71)

Solution The Hamiltonian is H = −x + λu. Here, we do not need to introduce the Lagrange multipliers for the control constraints (3.71), since we can easily deduce that the Hamiltonian maximizing control has the form u∗ = bang[−1, 1; λ].

(3.72)

λ˙ = 1

(3.73)

The adjoint equation is

3.4. Transversality Conditions: Special Cases

91

with the transversality conditions λ(2) ≥ 0 and λ(2)x(2) = 0,

(3.74)

obtained from (3.61) or from Table 3.1, Row 3. Since λ(t) is monotonically increasing, the control (3.72) can switch at most once, and it can only switch from u∗ = −1 to u∗ = 1. Let the switching time be t∗ ≤ 2. Then the optimal control is ⎧ ⎪ ⎨ −1 for 0 ≤ t ≤ t∗ , ∗ u (t) = (3.75) ⎪ ⎩ +1 for t∗ < t ≤ 2. Since the control switches at t∗ , λ(t∗ ) must be 0. Solving (3.73) gives λ(t) = t − t∗ . There are two cases: (i) t∗ < 2 and (ii) t∗ = 2. We analyze case (i) ﬁrst. Here λ(2) = 2 − t∗ > 0; therefore from (3.74), x(2) = 0. Solving for x(t) with u∗ (t) given in (3.75), we obtain ⎧ ⎪ ⎨ 1−t for 0 ≤ t ≤ t∗ , x(t) = ⎪ ⎩ (t − t∗ ) + x(t∗ ) = t + 1 − 2t∗ for t∗ < t ≤ 2. Therefore, setting x(2) = 0 gives x(2) = 3 − 2t∗ = 0, which makes t∗ = 3/2. Since this satisﬁes t∗ < 2, we do not have to deal with case (ii), and we have ⎧ ⎪ ⎨ 1 − t for 0 ≤ t ≤ 3/2, 3 ∗ and λ(t) = t − . x (t) = ⎪ 2 ⎩ t − 2 for 3/2 < t ≤ 2 Figure 3.1 shows the optimal state and adjoint trajectories. Using the optimal state trajectory in the objective function, we can obtain its optimal value J ∗ = −1/4. In Exercise 3.15, you are asked to consider case (ii) by setting t∗ = 2, and show that the maximum principle will not be satisﬁed in this case.

92

3. The Maximum Principle: Mixed Inequality Constraints

Finally, we can verify the marginal value interpretation of the adjoint variable as indicated in Remark 3.5. For this, we ﬁrst note that the feasible region for the problem is given by x ≥ t − 2, t ∈ [0, 2]. To obtain the value function V (x, t), we can easily obtain the optimal solution in the interval [t, 2] for the problem beginning with x(t) = x. We use the notation introduced in Example 2.5 to specify the optimal solution as ⎧ ⎪ ⎨ −1, s ∈ [t, 1 (x + t) + 1), 2 ∗ u(x,t) (s) = ⎪ ⎩ 1, s ∈ [ 1 (x + t) + 1, 2], 2 and x∗(x,t) (s)

=

⎧ ⎪ ⎨ x + t − s, s ∈ [t, ⎪ ⎩

s − 2,

1 2 (x

+ t) + 1),

s ∈ [ 12 (x + t) + 1, 2].

Then for x ≥ t − 2, 2

−x∗(x,t) (s)ds 2 (1/2)(x+t)+1 (x + t − s)ds − (1/2)(x+t)+1 (s − 2)ds = − t

V (x, t) =

t

= (1/4)t2 − (1/4)x2 + (1/2)t(x − 2) − (x − 1). (3.76) For x < t − 2, there is no feasible solution, and we therefore set V (x, t) = −∞. We can now verify that for 0 ≤ t ≤ 3/2, the value function V (x, t) is continuously diﬀerentiable at x = x∗ (t) = 1 − t, and Vx (x∗ (t), t) = −(1/2)x∗ (t) + (1/2)t − 1 = −(1/2)(1 − t) + (1/2)t − 1 = t − 3/2 = λ(t). What happens when t ∈ (3/2, 2]? Clearly, for x ≥ x∗ (t) = t − 2, we may still use (3.76) to obtain the right-hand derivative Vx+ (x∗ (t), t) = −(1/2)x∗ (t) + (1/2)t − 1 = −(1/2)(t − 2) + (1/2)t − 1 = 0. However, for x < x∗ (t), we have x < t − 2 for which there is no feasible solution, and we set the left-hand derivative Vx− (x∗ (t), t) = −∞. Thus, the value

3.5. Free Terminal Time Problems

93

function V (x, t) is not diﬀerentiable at x∗ (t), and since Vx (x∗ (t), t) does not exist for t ∈ (3/2, 2], (2.17) has no meaning; see Remark 2.2. It is possible, however, to provide an economic meaning for λ(2). In Exercise 3.17, you are asked to rework Example 3.4 with the terminal condition x(2) ≥ 0 replaced by x(2) ≥ ε, where ε is small. Furthermore,

½

½

=

-½ ()

-3/2

Figure 3.1: State and adjoint trajectories in Example 3.4 the solution will illustrate that α = λ(2) − 0 = 1/2, obtained by using (3.60), represents the shadow price of the constraint as indicated in Remark 3.7.

3.5

Free Terminal Time Problems

In some cases, the terminal time is not given but needs to be determined as an additional decision. Here, a necessary condition for a terminal time to be optimal in the present-value and current-value formulations are given in (3.15) and (3.44), respectively. In this section, we elaborate further on these conditions as well as solve two free terminal time examples: Examples 3.5 and 3.6.

94

3. The Maximum Principle: Mixed Inequality Constraints

Let us begin with a special case of the condition (3.15) for the simple problem (2.4) when T ≥ 0 is a decision variable. When compared with the problem (3.7), the simple problem is without the mixed constraints and constraints at the terminal time T. Thus the transversality condition (3.15) reduces to H[x∗ (T ∗ ), u∗ (T ∗ ), λ(T ∗ ), T ∗ ] + ST [x∗ (T ∗ ), T ∗ ] = 0.

(3.77)

This condition along with the Maximum Principle (2.31) with T replaced by T ∗ give us the necessary conditions for the optimality of T ∗ and u∗ (t), t ∈ [0, T ∗ ] for the simple problem (2.4) when T ≥ 0 is also a decision variable. An intuitively appealing way to check if the optimal T ∗ ∈ (0, ∞) must satisfy (3.77) is to solve the problem (2.4) with the terminal time T ∗ with u∗ (t), t ∈ [0, T ∗ ] as the optimal control trajectory, and then show that the ﬁrst-order condition for T ∗ to maximize the objective function in a neighborhood (T ∗ − δ, T ∗ + δ) of T ∗ with δ > 0 leads to (3.77). For this, let us set u∗ (t) = u∗ (T ∗ ), t ∈ [T ∗ , T ∗ + δ), so that we have a control u∗ (t) that is feasible for (2.4) for any T ∈ (T ∗ − δ, T ∗ + δ), as well as continuous at T ∗ . Let x∗ (t), t ∈ [0, T ∗ + δ] be the corresponding state trajectory. With these we can obtain the corresponding objective function value T F (x∗ (t), u∗ (t), t)dt + S(x∗ (T ), T ), T ∈ (T ∗ − δ, T ∗ + δ), J(T ) = 0

(3.78) which, in particular, represents the optimal value of the objective function for the problem (2.4) when T = T ∗ . Furthermore, since u∗ (t) is continuous at T ∗ , x∗ (t) is continuously diﬀerentiable there, and so is J(T ). In this case, since T ∗ is optimal, it must satisfy dJ(T ) |T =T ∗ = 0. (3.79) dT Otherwise, we would have either J (T ∗ ) > 0 or J (T ∗ ) < 0. The former situation would allow us to ﬁnd a T ∈ (T ∗ , T ∗ + δ) for which J(T ) > J(T ∗ ), and T ∗ could not be optimal since the choice of an optimal control for (2.4) deﬁned on the interval [0, T ] would only improve the value of the objective function. Likewise, the later situation would allow us to ﬁnd a T ∈ (T ∗ − δ, T ∗ ) for which J(T ) > J(T ∗ ). By taking the derivative of (3.78), we can write (3.79) as J (T ∗ ) :=

F (x∗ (T ∗ ), u∗ (T ∗ ), T ∗ ) + Sx [x∗ (T ∗ ), T ∗ ]x˙ ∗ (T ∗ ) + ST [x∗ (T ∗ ), T ∗ ] = 0. (3.80)

3.5. Free Terminal Time Problems

95

Furthermore, using the deﬁnition of the Hamiltonian in (2.18) and the state equation and the transversality condition in (2.31), we can easily see that (3.80) can be written as (3.77). Remark 3.10 An intuitive way to obtain optimal T ∗ is to ﬁrst solve the problem (2.4) with a given terminal time T and obtain the optimal value of the objective function J ∗ (T ), and then maximize J ∗ (T ) over T. Hartl and Sethi (1983) show that the ﬁrst-order condition for maximizing J ∗ (T ), namely, dJ ∗ (T )/dT = 0 can also be used to derive the transversality condition (3.77). If T is restricted to lie in the interval [T1 , T2 ], where T2 > T1 ≥ 0, then (3.77) is still valid provided T ∗ ∈ (T1 , T2 ). As is standard, if T ∗ = T1 , then the = sign in (3.77) is replaced by ≤, and if T ∗ = T2 , then the = sign in (3.77) is replaced by ≥ . In other words, if we must have T ∗ ∈ [T1 , T2 ], then we can replace (3.77) by ⎧ ⎪ ⎪ ⎪ ≤ 0 if T ∗ = T1 , ⎪ ⎪ ⎨ H[x∗ (T ∗ ), u∗ (T ∗ ), λ(T ∗ ), T ∗ ] + ST [x∗ (T ∗ ), T ∗ ] = 0 if T ∗ ∈ (T1 , T2 ), ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ ≥ 0 if T ∗ = T2 . (3.81) Similarly, we can also obtain the corresponding versions of (3.15) and (3.44) for the problem (3.7) and its current value version (speciﬁed in Sect. 3.3), respectively. We shall now illustrate (3.77) and (3.81) by solving Examples 3.5 and 3.6. To illustrate the idea in Remark 3.10, you are asked in Exercise 3.6 to solve Example 3.5 by using dJ ∗ (T )/dt = 0 to obtain the optimal T ∗ . Example 3.5 Consider the problem: T (x − u)dt + x(T ) max J = u,T

(3.82)

0

subject to x˙ = −2 + 0.5u, x(0) = 17.5, u ∈ [0, 1], T ≥ 0.

(3.83)

96

3. The Maximum Principle: Mixed Inequality Constraints

Solution The Hamiltonian is H = x − u + λ(−2 + 0.5u), where λ˙ = −1, λ(T ) = 1, which gives λ(t) = 1 + (T − t). Then, the optimal control is given by u∗ (t) = bang[0, 1; 0.5(T − 1 − t)].

(3.84)

In other words, u∗ (t) = 1 for 0 ≤ t ≤ T − 1 and u∗ (t) = 0 for T − 1 < t ≤ T. Since we must also determine the optimal terminal time T ∗ , it must satisfy (3.77), which, in view of the fact that u∗ (T ∗ ) = 0 from (3.84), reduces to x∗ (T ∗ ) − 2 = 0. (3.85) By substituting u∗ (t) in (3.83) and integrating, we obtain

x∗ (t) =

⎧ ⎪ ⎨ 17.5 − 1.5t,

0 ≤ t ≤ T − 1,

⎪ ⎩ 17 + 0.5T − 2t, T − 1 < t ≤ T.

(3.86)

We can now apply (3.85) to obtain x∗ (T ∗ ) − 2 = 17 − 1.5T ∗ − 2 = 0, which gives T ∗ = 10. Thus, the optimal solution of the problem is given by T ∗ = 10 and u∗ (t) = bang[0, 1; 0.5(9 − t)]. Note that if we had restricted T to be in the interval [T1 , T2 ] = [2, 8], we would have T ∗ = 8, u∗ (t) = bang[0, 1; 0.5(7 − t)], and x∗ (8) − 2 = 5 − 2 = 3 ≥ 0, which would satisfy (3.81) at T ∗ = T2 = 8. On the other hand, if T were restricted in the interval [T1 , T2 ] = [11, 15], then T ∗ = 11, u∗ (t) = bang[0, 1; 0.5(10 − t)], and x∗ (11) − 2 = 0.5 − 2 = −1.5 ≤ 0 would satisfy (3.81) at T ∗ = T1 = 11. Next, we will apply the maximum principle to solve a well known time-optimal control problem. It is one of the problems used by Pontryagin et al. (1962) to illustrate the applications of the maximum principle.

3.5. Free Terminal Time Problems

97

The problem also elucidates a speciﬁc instance of the synthesis of optimal controls. By the synthesis of optimal controls, we mean the procedure of “patching” together various forms of the optimal controls obtained from the Hamiltonian maximizing condition. A simple example of the synthesis occurs in Example 2.5, where u∗ = 1 when λ > 0, u∗ = −1 when λ < 0, and the control is singular when λ = 0. An optimal trajectory starting at the given initial state variables is synthesized from these. In Example 2.5, this synthesized solution is u∗ = −1 for 0 ≤ t < 1 and u∗ = 0 for 1 ≤ t ≤ 2. Our next example requires a synthesis procedure which is more complex. In Chap. 5, both the cash management and equity ﬁnancing models require such synthesis procedures. Example 3.6 A Time-Optimal Control Problem. Consider a subway train of mass m moving horizontally along a smooth linear track with negligible friction. Let x(t) denote the position of the train, measured in miles from the origin called the main station, along the track at time t, measured in minutes. Then the equation of the train’s motion is governed by Newton’s Second Law of Motion, which states that force equals mass times acceleration. In mathematical terms, the equation of the motion is the second-order diﬀerential equation m

d2 x(t) = m¨ x(t) = u(t), dt2

where u(t) denotes the external force applied to the train at time t and x ¨(t) represents the acceleration in miles per minute per minute, or miles/minute2 . This equation, along with ˙ = y0 , x(0) = x0 and x(0) respectively, as the initial position of the train and its initial velocity in miles per minute, characterizes its motion completely. For convenience in further exposition, we may assume m = 1 so that the equation of motion can be written as x ¨ = u.

(3.87)

Then, the force u can be expressed simply as acceleration or deceleration (i.e., negative acceleration) depending on whether u is positive or negative, respectively.

98

3. The Maximum Principle: Mixed Inequality Constraints

In order to develop the time-optimal control problem under consideration, we transform (3.87) into a system of two ﬁrst-order diﬀerential equations (see Appendix A) ⎧ ⎪ ⎨ x˙ = y, x(0) = x0 , (3.88) ⎪ ⎩ y˙ = u, y(0) = y , 0 where y(t) denotes the velocity of the train in miles/minute at time t. Assume further that, for the comfort of the passengers, the maximum acceleration and deceleration are required to be at most 1 mile/minute2 . Thus, the control variable constraint is u ∈ Ω = [−1, 1].

(3.89)

The problem is to ﬁnd a control satisfying (3.89) such that the train stops at the main station located at x = 0 in a minimum possible time T. Of course, for the train to come to rest at x = 0 at time T, we must have x(T ) = 0 and y(T ) = 0. We have thus deﬁned the following ﬁxed-end-point optimal control problem: ⎧ T ⎪ ⎪ ⎪ max J = −1dt ⎪ ⎪ ⎪ 0 ⎪ ⎪ ⎪ ⎪ ⎪ subject to ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎨ x˙ = y, x(0) = x , x(T ) = 0, 0 (3.90) ⎪ ⎪ ⎪ y˙ = u, y(0) = y0 , y(T ) = 0, ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ and the control constraint ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ u ∈ Ω = [−1, 1]. Note that (3.90) is a ﬁxed-end-point problem with unspeciﬁed terminal time. For this problem to be nontrivial, we must not have x0 = y0 = 0, i.e., we must have either x0 = 0 or y0 = 0 or both are nonzero. Solution Here we have only control constraints of the type treated in Chap. 2, and so we can use the maximum principle (2.31). The standard Hamiltonian function is H = −1 + λ1 y + λ2 u,

3.5. Free Terminal Time Problems

99

where the adjoint variables λ1 and λ2 satisfy λ˙1 = 0, λ1 (T ) = β 1 and λ˙2 = −λ1 , λ2 (T ) = β 2 , and β 1 and β 2 are constants to be determined in the case of a ﬁxed-endpoint problem; see Table 3.1, Row 2. We can integrate these equations and write the solution in the form λ1 = β 1 and λ2 = β 2 + β 1 (T − t), where β 1 and β 2 are constants to be determined from the maximum principle (2.31), condition (3.15), and the speciﬁed initial and terminal values of the state variables. The Hamiltonian maximizing condition yields the form of the optimal control to be u∗ (t) = bang{−1, 1; β 2 + β 1 (T − t)}.

(3.91)

As for the minimum time T ∗ , it is clearly zero if the train is initially at rest at the main station, i.e., (x0 , y0 ) = 0. In this case, the problem is trivial, u∗ (0) = 0, and there is nothing further to solve. Otherwise, at least one of x0 or y0 is not zero, in which case the minimum time T ∗ > 0 and the transversality condition (3.15) applies. Since y(T ) = 0 and S ≡ 0, we have H + ST |T =T ∗ = λ2 (T ∗ )u∗ (T ∗ ) − 1 = β 2 u∗ (T ∗ ) − 1 = 0, which together with the bang-bang control policy (3.91) implies either λ2 (T ∗ ) = β 2 = −1 and u∗ (T ∗ ) = −1, or

λ2 (T ∗ ) = β 2 = +1 and u∗ (T ∗ ) = +1.

Since the switching function β 2 + β 1 (T ∗ − t) is a linear function of the time remaining, it can change sign at most once. Therefore, we have two cases: (i) u∗ (τ ) = −1 in the interval t ≤ τ ≤ T ∗ for some t ≥ 0; (ii) u∗ (τ ) = +1 in the interval t ≤ τ ≤ T ∗ for some t ≥ 0. We can integrate (3.88) in each of these cases as shown in Table 3.2. Also in the table we have the curves Γ− and Γ+ , which are obtained by eliminating t from the expressions for x and y in each case. The parabolic curves Γ− and Γ+ are called switching curves and are shown in Fig. 3.2. It should be noted parenthetically that Fig. 3.2 is diﬀerent from the ﬁgures we have seen thus far, where the abscissa represented the time

100

3. The Maximum Principle: Mixed Inequality Constraints Table 3.2: State trajectories and switching curves (i) u∗ (τ ) = −1 for (t ≤ τ ≤ T ∗ )

(ii) u∗ (τ ) = +1 for (t ≤ τ ≤ T ∗ )

y(t) = T ∗ − t

y(t) = t − T ∗

x(t) = −(T ∗ − t)2 /2

x(t) = (t − T ∗ )2 /2

Γ− : x = −y 2 /2 for y ≥ 0

Γ+ : x = y 2 /2 for y ≤ 0

dimension. In Fig. 3.2, the abscissa represents the train’s location and the ordinate represents the train’s velocity. Thus, the point (x0 , y0 ) represents the vector of the train’s initial position and initial velocity. A trajectory of the train over time can be represented by a curve in this ﬁgure. For example, the bold-faced trajectory beginning at (x0 , y0 ) represents a train that is moving in the positive direction and it is slowing down. It passes through the main station located at the origin and comes # to a momentary rest at the point that is y02 + 2x0 miles to the right of the main station. At this location, the train reverses its direction and speeds up to reach the location x∗ and attain the velocity of y∗ . At this point, it slows down gradually until it comes to rest at the main station. In the ensuing discussion we will show that this trajectory is in fact the minimal time trajectory beginning at the location x0 at a velocity of y0 . We will furthermore obtain the control representing the optimal acceleration and deceleration along the way. Finally, we will obtain the various instants of interest, which are implicit in the depiction of the trajectory in Fig. 3.2. We can put Γ+ and Γ− into a single switching curve Γ as ⎧ √ ⎪ ⎨ Γ+ (x) = − 2x, x ≥ 0, y = Γ(x) = (3.92) ⎪ ⎩ Γ− (x) = +√−2x, x < 0. If the initial state (x0 , y0 ) = 0, lies on the switching curve, then we have u∗ = +1 (resp., u∗ = −1) if x0 > 0 (resp., x0 < 0); i.e., if (x0 , y0 ) lies on Γ+ (resp., Γ− ). In the common parlance, this means that we apply the brakes to bring the train to a full stop at the main station. If the initial state (x0 , y0 ) is not on the switching curve, then we choose, between u∗ = 1 and u∗ = −1, that which moves the system toward the switching

3.5. Free Terminal Time Problems

101

Figure 3.2: Minimum time optimal response for Example 3.6 curve. By inspection, it is obvious that above the switching curve we must choose u∗ = −1 and below we must choose u∗ = +1. The other curves in Fig. 3.2 are solutions of the diﬀerential equations starting from initial points (x0 , y0 ). If (x0 , y0 ) lies above the switching curve Γ as shown in Fig. 3.2, we use u∗ = −1 to compute the curve as follows: x˙ = y, x(0) = x0 , y˙ = −1, y(0) = y0 . Integrating these equations gives y = −t + y0 , t2 + y0 t + x 0 . 2 Elimination of t between these two gives x=−

x=

y02 − y 2 + x0 . 2

(3.93)

This is the equation of the parabola in Fig. 3.2 through (x0 , y0 ). The point of intersection of the curve (3.93) with the switching curve Γ+ is obtained by solving (3.93) and the equation for Γ+ , namely 2x = y 2 , simultaneously, which gives y 2 + 2x0 , y∗ = − (y02 + 2x0 )/2, (3.94) x∗ = 0 4

102

3. The Maximum Principle: Mixed Inequality Constraints

where the minus sign in the expression for y∗ in (3.94) was chosen since the intersection occurs when y∗ is negative. The time t∗ that it takes to reach the switching curve, called the switching time, given that we start above it, is t∗ = y0 − y∗ = y0 +

(y02 + 2x0 )/2.

(3.95)

To ﬁnd the minimum total time to go from the starting point (x0 , y0 ) to the origin (0,0), we substitute t∗ into the equation for Γ+ in Column (ii) of Table 3.2; this gives ∗ (3.96) T = t∗ − y∗ = y0 + 2(y02 + 2x0 ). Here t∗ is the time to get to the switching curve and −y∗ is the time spent along the switching curve. Note # that the parabola (3.93) intersects the y-axis at the point (0, + 2x0 + y02 ) and the x-axis at the point (x0 + y02 /2, 0). This means that for the initial position (x0 , y0 ) depicted#in Fig. 3.2, the train ﬁrst passes the main station at the velocity of + 2x0 + y02 and comes to a momentary stop at the distance of (x0 + y02 /2) to the right of the main station. There it reverses its direction, comes to within the distance of x∗ from the main station, switches then to u∗ = +1, which slows it to a complete stop at the main station at time T ∗ given by (3.96). As a numerical example, start at the point (x0 , y0 ) =(1,1). Then, the equation of the parabola (3.93) is 2x = 3 − y 2 .

# The switching point given by (3.94) # is (3/4, − 3/2). Finally from (3.95), the switching time is t∗ = 1 + 3/2 min. Substituting into (3.96), we √ ∗ ﬁnd the minimum time to stop is T = 1 + 6 min. To complete the solution of this example let us evaluate β 1 and β 2 , which are needed to obtain λ1 and λ2 . Since (1,1) is above the switching curve, the approach to the main station is on the curve Γ+ , and therefore, u∗ (T ∗ ) = 1 and β 2 = 1. To compute β 1 , we observe # that λ2 (t #∗ ) = ∗ ∗ β 2 + β 1 (T − t∗ ) = 0 so that β 1 = −β 2 /(T # − t∗ ) = −1/ 3/2 = − 2/3. Finally, we obtain x∗ = 3/4 and y∗ = − 3/2 from (3.94). Let us now describe the optimal solution from (1, 1) in the common parlance. The position (1, 1) means the train is 1 mile to the right of the main station, moving away from it at the speed of 1 mile per minute. The control u∗ = −1 means that the brakes are applied to slow the train

3.6. Inﬁnite Horizon and Stationarity

103

down. √ This action brings the train to a momentary stop at a distance of 3 miles to the right of the main station. Moreover, the continuation of control u∗ = −1 means the train reverses its direction at that point and starts speeding toward the station. When it comes # to within 3/4 miles#to the right of the main # station at time t∗ = 1 + 3/2, its velocity of − 3/2 or the speed of 3/2 miles per minute toward the station is too fast to come to a rest at the main station without application of the brakes. So the control is switched to u∗ = +1 at time t∗ , which means the brakes are applied at that time. This action brings the √ train to a ∗ complete stop at the main station at the time of T = 1 + 6 min after the train left its initial position (1, 1). In Exercises 3.19–3.22, you are asked to work other examples with diﬀerent starting points above, below, and on the switching curve. Note that t∗ = 0 by deﬁnition, if the starting point is on the switching curve.

3.6

Inﬁnite Horizon and Stationarity

Thus far, we have studied problems whose horizon is ﬁnite or whose horizon length is a decision variable to be determined. In this section, we brieﬂy discuss the case of T = ∞ in the problem (3.7), called the inﬁnite horizon case. This case is especially important in many economics and management science problems. Our treatment of this case is largely heuristic, since a general theory of the necessary optimality conditions is not available. Nevertheless, we can rely upon an inﬁnite-horizon extension of the suﬃciency optimality conditions stated in Theorem 3.1. When we put T = ∞ in (3.7) along with ρ > 0, we will generally get a nonstationary inﬁnite horizon problem in the sense that the various functions involved depend explicitly on the time variable t. Such problems are extremely hard to solve. So, in this section we will devote our attention to only stationary inﬁnite horizon problems, which do not depend explicitly on time t. Furthermore, it is reasonable in most cases to assume σ(x) ≡ 0 in inﬁnite horizon problems. Moreover, in most economics and management science problems, the terminal constraints, if

104

3. The Maximum Principle: Mixed Inequality Constraints

any, require the state variables to be nonnegative. Thus, to begin with, we consider the problem: ⎧ ∞ ⎪ −ρt ⎪ φ(x, u)e dt , max J = ⎪ ⎪ ⎪ 0 ⎪ ⎪ ⎪ ⎪ ⎨ subject to (3.97) ⎪ ⎪ ⎪ x˙ = f (x, u), x(0) = x0 , ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ g(x, u) ≥ 0. This stationarity assumption means that the state equations, the current-value adjoint equations, and the current-value Hamiltonian in (3.35) are all explicitly independent of time t. Remark 3.11 The concept of stationarity introduced here is diﬀerent from the concept of autonomous systems introduced in Exercise 2.9. This is because, in the presence of discounting in (3.28), the stationarity assumption (3.97) does not give us an autonomous system as deﬁned there. See Exercise 3.42 for further comparison between the two concepts. When it comes to the transversality conditions in the inﬁnite horizon case, the situation is somewhat more complicated. Even the economic argument for the ﬁnite horizon case fails to extend here because we do not have a meaningful analogue of the salvage value function. Moreover, in the free-end-point case with no salvage value, the standard maximum principle (2.31) gives λpv (T ) = 0, which can no longer be necessary in general for T = ∞, as conﬁrmed by a simple counter-example in Exercise 3.37. As a matter of fact, we have no general results giving conditions under which the limit of the ﬁnite horizon transversality conditions are necessary. What is true is that the maximum principle (3.42) holds except for the transversality condition on λ(T ). When it comes to the suﬃciency of the limiting transversality conditions obtained by letting T → ∞ in Theorem 3.1, the situation is much better. As a matter of fact, we can see from the inequality (2.73) with S(x) ≡ 0 that all we need is lim λpv (T )[x(T ) − x∗ (T )] = lim e−ρT λ(T )[x(T ) − x∗ (T )] ≥ 0 (3.98)

T →∞

T →∞

for Theorem 2.1, and therefore Theorem 3.1, to hold. See Seierstad and Sydsæter (1987) and Feichtinger and Hartl (1986) for further details.

3.6. Inﬁnite Horizon and Stationarity

105

In the important free-end-point case (3.97), since x(T ) is arbitrary, (3.98) will imply lim λpv (T ) = lim e−ρT λ(T ) = 0.

T →∞

T →∞

(3.99)

While not a necessary condition as indicated earlier, it is interesting to note that (3.99) is the limiting version of the condition in Table 3.1, Row 1. Another important case is that of nonnegativity constraints lim x(T ) ≥ 0.

T →∞

(3.100)

Then, it is clear that the transversality conditions lim e−ρT λ(T ) ≥ 0 and lim e−ρT λ(T )x∗ (T ) = 0,

T →∞

T →∞

(3.101)

imply (3.98). Note that these are also analogous to Table 3.1, Row 3. We leave it as Exercise 3.38 for you to show that the limiting version of the condition in the rightmost column of Rows 2, 3, and 4 in Table 3.1 imply (3.98). This would mean that Theorem 3.1 provides suﬃcient optimality conditions for the problem (3.97), except in the free-end-point case, i.e., when the terminal constraints a(x(T )) ≥ 0 and b(x(T )) = 0 are not present. Moreover, in the free-end-point case, we can use (3.98), or even (3.99) with some qualiﬁcations, as discussed earlier. Example 3.7 Let us return to Example 3.3 and now assume that we have a perpetual charitable trust with initial fund W0 , which wants to maximize its total discounted utility of charities C(t) over time, subject to the terminal condition lim W (T ) ≥ 0.

T →∞

(3.102)

For convenience we restate the problem: ∞ −ρt e ln C(t)dt max J = C(t)≥0

0

subject to ˙ = rW − C, W (0) = W0 > 0, W and (3.102).

(3.103)

106

3. The Maximum Principle: Mixed Inequality Constraints

Solution We already know from Example 3.3 with B = 0 that we are in case (i), and the optimal solution is given by (3.50) in Example 3.2. It seems reasonable to explore whether or not we can obtain an optimal solution for our inﬁnite horizon problem by letting T → ∞ in (3.50). Furthermore, since the limiting version of the maximum principle (3.42) is suﬃcient for optimality in this case, all we need to do is to check if the limiting solution satisﬁes the condition

lim e−ρT λ(T ) ≥ 0 and lim e−ρT λ(T )W ∗ (T ) = 0.

T →∞

T →∞

(3.104)

With T → ∞ in (3.50) and (3.52), we have W ∗ (t) = e(r−ρ)t W0 , C ∗ (t) = ρW ∗ (t), λ(t) = 1/ρW ∗ (t).

(3.105)

Since λ(t) ≥ 0 and λ(t)W ∗ (t) = 1/ρ, it is clear that (3.104) holds. Thus, (3.105) gives the optimal solution. Using this solution in the objective function, we obtain J∗ =

1 r−ρ ln ρW0 + 2 , ρ ρ

(3.106)

which we can verify to be the same as (3.51) as T → ∞. It is interesting to observe from (3.105) that the optimal consumption is increasing, constant, or decreasing if r is greater than, equal to, or less than ρ, respectively. Moreover, if ρ = r, then W ∗ (t) = W0 , C ∗ (t) = rW0 , and λ(t) = 1/rW0 , which means that it is optimal to consume just the interest earned on the invested wealth—no more, no less—and, therefore, none of the initial wealth is ever consumed! In the case of stationary systems, considerable attention is focused on equilibrium where all motion ceases, i.e., the values of x and λ for which x˙ = 0 and λ˙ = 0. The notion is that of optimal long-run stationary equilibrium; see Arrow and Kurz (1970, Chapter 2) and Carlson and Haurie (1987a, 1996). If an equilibrium exists, then it is deﬁned by the ¯ μ quadruple {¯ x, u ¯, λ, ¯ } satisfying

3.6. Inﬁnite Horizon and Stationarity

107

f (¯ x, u ¯) = 0, ¯ = Lx [¯ ¯ μ ρλ x, u ¯, λ, ¯ ], μ ¯ ≥ 0, μ ¯ g(¯ x, u ¯) = 0, and

(3.107)

¯ ≥ H(¯ ¯ H(¯ x, u ¯, λ) x, u, λ) for all u satisfying g(¯ x, u) ≥ 0. Clearly, if the initial condition x0 = x ¯, the optimal control is u∗ (t) = u ¯ ¯, the optimal solution will have a transient phase. for all t. If x0 = x Moreover, depending on the problem, the equilibrium may be attained in a ﬁnite time or an approach to it may be asymptotic. If the nonnegativity constraint (3.100) is added to problem (3.97), ¯ ≥ 0 and λ¯ ¯ x = 0 in (3.107). then we may include the requirement λ If the constraint involving g is not imposed in (3.97), μ ¯ may be dropped from the quadruple. In this case, the long-run stationary equi¯ satisfying librium is deﬁned by the triple {¯ x, u ¯, λ} ¯ = Hx (¯ ¯ and Hu (¯ ¯ = 0. f (¯ x, u ¯) = 0, ρλ x, u ¯, λ), x, u ¯, λ)

(3.108)

Also known in this case is that the optimal value of the objective function can be expressed as J ∗ = H(x0 , u∗ (0), λ(0))/ρ.

(3.109)

You are asked to prove this relation in Exercise 3.40. That it holds in Example 3.7 is quite clear when we use (3.105) in (3.109) and see that we get (3.106). Also, we see from Example 3.7 that when we let t → ∞ in (3.105), we formally obtain ⎧ ⎪ ⎪ ⎪ (0, 0, ∞) if ρ > r, ⎪ ⎪ ⎨ ¯ = ¯ , C, ¯ λ) (W (3.110) (W0 , ρW0 , 1/ρW0 ) if ρ = r, ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ (∞, ∞, 0) if ρ < r.

108

3. The Maximum Principle: Mixed Inequality Constraints

This is precisely the long-run stationary equilibrium that we will obtain ¯ ≥ 0 and λ ¯W ¯ = 0 directly to the optimal if we apply (3.108) along with λ control problem in Example 3.7. This veriﬁcation is left as Exercise 3.41. Example 3.8 For another application of (3.108), let us return to Example 3.7 and now assume that the wealth W is invested in a productive activity resulting in an output rate ln W, and that the horizon T = ∞. Since ln W is only deﬁned for W > 0, we do not need to impose the terminal constraint (3.102) here. Thus, the problem is ∞ max J = e−ρt ln C(t)dt C(t)≥0

0

subject to ˙ = ln W − C, W (0) = W0 > 0, W

(3.111)

and one task is to ﬁnd the long-run stationary equilibrium for it. Note that since the horizon is inﬁnite, it is usual to assume no salvage value and no terminal conditions on the state. Solution By (3.108) we set ¯ = 0, ¯ − C¯ = 0, ρ = 1/W ¯ , 1/C¯ − λ ln W ¯ = {1/ρ, − ln ρ, −1/ ln ρ}. Since, ¯ , C, ¯ λ} which gives the equilibrium {W 0 < ρ < 1, we have C¯ > 0, which satisﬁes the requirement that the ¯ > 0. consumption be nonnegative. Also, the equilibrium wealth W It is important to note that the optimal long-run stationary equilibrium (which is also called the turnpike) is not the same as the optimal steady-state among the set of all possible steady-states. The latter concept is termed the Golden Rule or Golden Path in economics, and a procedure to obtain it is described below. However, the two concepts are identical if the discount rate ρ = 0; see Exercise 3.43. The Golden Path is obtained by setting x˙ = f (x, u) = 0, which provides the feedback control u(x) that would keep x(t) = x over time. Then, substitute u(x) in the integrand φ(x, u) of (3.28) to obtain φ(x, u(x)). The value of x that maximizes φ(x, u(x)) yields the Golden Path. Of course, all of the constraints imposed on the problem have to be respected when obtaining the Golden Path. In some cases, there may be more than one equilibria deﬁned by (3.107). If so, the equilibrium that is attained may depend on the initial

3.7. Model Types

109

starting point. Moreover, from some special starting points, the system may have an option to go to two or more diﬀerent equilibria. Such points are called the Sethi-Skiba points; see Appendix D.8. For multidimensional systems consisting of two or more states, optimal trajectories may exhibit more complex behaviors. Of particular importance is the concept of limit cycles. If the optimal trajectory of a dynamical system tends to spiral in toward a closed loop in the state space, then that closed loop is called a limit cycle. For more on this topic, refer to Vidyasagar (2002) and Grass et al. (2008).

3.7

Model Types

Optimal control theory has been used to solve problems occurring in engineering, economics, management science, and other ﬁelds. In each ﬁeld of application, certain general kinds of models which we will call model types are likely to occur, and each such model requires a specialized form of the maximum principle. In Chap. 2 we derived, in considerable detail, a simple form of the continuous-time maximum principle. However, to continue to provide such details for each diﬀerent version of the maximum principle needed in later chapters of this book would be both repetitive and lengthy. The purpose of this section is to avoid the latter by listing most of the diﬀerent management science model types that we will use in later chapters. For each model type, we will give a brief description of the corresponding objective function, state equations, control and state inequality constraints, terminal conditions, adjoint equations, and the form of the optimal control policy. We will also indicate where each of these model types is applied in later chapters. The reader may wish to skim this section on ﬁrst reading to get an idea of what it contains, work a few of the exercises, and go on to the various functional areas discussed in later chapters. Then, when speciﬁc model types are encountered, the reader may return to read the relevant parts of this section in more detail. We are now able to state the general forms of all the models (with one or two exceptions) that we will use to analyze the applications discussed in the rest of the book. Some other model types will be explained in later chapters. In Table 3.3 we have listed six diﬀerent combinations of φ and f functions. If we specify the initial value x0 of the state variable x and

110

3. The Maximum Principle: Mixed Inequality Constraints

the constraints on the control and state variables, we can get a completely speciﬁed optimal control model by selecting one of the model types in Table 3.3 together with one of the terminal conditions given in Table 3.1. The reader will see numerous examples of the uses of Tables 3.1 and 3.3 when we construct optimal control models of various applied situations in later chapters. To help in understanding these, we will give a brief mathematical discussion of the six model types in Table 3.3, with an indication of where each model type will be used later in the book. In Model Type (a) of Table 3.3 we see that both φ and f are linear functions of their arguments. Hence it is called the linear-linear case. The Hamiltonian is H = Cx + Du + λ(Ax + Bu + d) = Cx + λAx + λd + (D + λB)u.

(3.112)

From (3.112) it is obvious that the optimal policy is bang-bang with the switching function (D + λB). Since the adjoint equation is independent of both control and state variables, it can be solved completely without resorting to two-point boundary value methods. Examples of (a) occur in the cash balance problem of Sect. 5.1.1 and the maintenance and replacement model of Sect. 9.1.1. Model Type (b) of Table 3.3 is the same as Model Type (a) except that the function C(x) is nonlinear. Thus, the term Cx appears in the adjoint equation, and two-point boundary value methods are needed to solve the problem. Here, there is a possibility of singular control, and a speciﬁc example is the Nerlove-Arrow model in Sect. 7.1.1. Model Type (c) of Table 3.3 has linear functions in the state equation and quadratic functions in the objective function. Therefore, it is sometimes called the linear-quadratic case. In this case, the optimal control can be expressed in a form in which the state variables enter linearly. Such a form is known as the linear decision rule; see (D.36) in Appendix D. A speciﬁc example of this case occurs in the productioninventory example of Sect. 6.1.1. Model Type (d) is a more general version of Model Type (b) in which the state equation is nonlinear in x. Here again, there is a possibility of singular control. The wheat trading model of Sect. 6.2.1 illustrates this model type. The solution of a special case of the model in Sect. 6.2.3 exhibits the occurrence of a singular control.

3.7. Model Types

111

Table 3.3: Objective, state, and adjoint equations for various model types Objective

State

Current-value

Form of optimal

function

equation

adjoint equation

control policy

φ=

x˙ = f =

˙ = λ

Cx + Du

Ax + Bu + d

λ(ρ − A) − C

Bang-bang

(b)

C(x) + Du

Ax + Bu + d

λ(ρ − A) − Cx

Bang-bang+Singular

(c)

xT Cx + uT Du

Ax + Bu + d

λ(ρ − A) − 2xT C

Linear decision rule

(d)

C(x) + Du

A(x) + Bu + d

λ(ρ − Ax ) − Cx

Bang-bang+Singular

(e)

c(x) + q(u)

(ax + d)b(u) + e(x)

λ(ρ − ab(u) − ex ) − cx

Interior or boundary

(f)

c(x)q(u)

(ax + d)b(u) + e(x)

λ(ρ − ab(u) − ex ) − cx q(u)

Interior or boundary

integrand

(a)

Note. The current-value Hamiltonian is often used when ρ > 0 is the discount rate; the standard formulation is identical to the current-value formulation when ρ = 0. In Table 3.3, capital letters indicate vector functions and small letters indicate scalar functions or vectors. A function followed by an argument in parentheses indicates a nonlinear function; when it is followed by an argument without parenthesis, it indicates a linear function. Thus, A(x) and e(x) are nonlinear vector and scalar functions, while Ax and ax are linear. The function d is always to be interpreted as an exogenous function of time only

In Model Types (e) and (f), the functions are scalar functions, and there is only one state equation, so λ is also a scalar function. In these cases, the Hamiltonian function is nonlinear in u. If it is concave in u, then the optimal control is usually obtained by setting Hu = 0. If it is convex, then the optimal control is the same as in Model Type (b). Several examples of Model Type (e) occur in this book: the optimal ﬁnancing model in Sect. 5.2.1, the Vidale-Wolfe advertising model in Sect. 7.2.1, the nonlinear extension of the maintenance and replacement model in Sect. 9.1.4, the forestry model in Sect. 10.2.1, the exhaustible resource model in Sect. 10.3.1, and all of the models in Chap. 11. Model Type (f) examples are: The Kamien-Schwartz model in Sect. 9.2.1 and the sole-owner ﬁshery resource model in Sect. 10.1. Although the general forms of the model are speciﬁed in Tables 3.1 and 3.3, there are a number of additional modeling tricks that are useful, which will be employed later. We collect these as a series of remarks below. Remark 3.12 We sometimes need to use the absolute value function |u| of a control variable u in forming the functions φ or f. For example,

112

3. The Maximum Principle: Mixed Inequality Constraints

in the simple cash balance model of Sect. 5.1, u < 0 represents buying and u > 0 represents selling; in either case there is a transaction cost which can be represented as c|u|. In order to handle this, we deﬁne new control variables u1 and u2 satisfying the following relations: u := u1 − u2 , u1 ≥ 0, u2 ≥ 0,

(3.113)

u1 u2 = 0.

(3.114)

Thus, we represent u as the diﬀerence of two nonnegative variables, u1 and u2 , together with the quadratic constraint (3.114). We can then write |u| = u1 + u2 , (3.115) which expresses the nonlinear function |u| as a linear function with the constraint (3.114). We now observe that we need not impose (3.114) explicitly, provided there are costs associated with the controls u1 and u2 , since in the presence of these costs no optimal policy would ever choose to make both of them simultaneously positive. This is indeed the case in the cash balance problem of Sect. 5.1, where the associated transaction costs prevent us from simultaneously buying and selling the same security. Thus, by doubling the number of variables and adding inequality constraints, we are able to represent |u| as a linear function in the model. Remark 3.13 Tables 3.1 and 3.3 are constructed for continuous-time models. Exactly the same kinds of models can be developed in the discrete-time case; see Chap. 8. Remark 3.14 Consider Model Types (a) and (b) when the control variable constraints are deﬁned by linear inequalities of the form g(u, t) = g(t)u ≥ 0.

(3.116)

Then, the problem of maximizing the Hamiltonian function becomes: ⎧ ⎪ ⎪ ⎪ max(D + λB)u ⎪ ⎪ ⎨ (3.117) subject to ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ g(t)u ≥ 0. This is clearly a linear programming problem for each given instant of time t, since the Hamiltonian function is linear in u.

Exercises for Chapter 3

113

Further in Model Type (a), the adjoint equation does not contain terms in x and u, so we can solve it for λ(t), and hence the objective function of (3.117) varies parametrically with λ(t). In this case we can use parametric linear programming techniques to solve the problem over time. Since the optimal solution to the linear program always occurs at an extreme point of the convex set deﬁned by g(t)u ≥ 0, it follows that as λ(t) changes, the optimal solution to (3.117) will “bang” from one extreme point of the feasible set to another. This is called a generalized bang-bang optimal policy. Such a policy occurs, e.g., in the optimal ﬁnancing model treated in Sect. 5.2; see Table 5.1, Row 5. In Model Type (b), the adjoint equation contains terms in x, so we cannot solve for the trajectory of λ(t) without knowing the trajectory of x(t). It is still true that (3.117) is a linear program for any given t, but the parametric linear programming techniques will not usually work. Instead, some type of iterative procedure is needed in general; see Bryson and Ho (1975). Remark 3.15 The salvage value part S[x(T ), T ] of the objective function is relevant in the optimization context in the following two cases: Case (i) T is free and part of the problem is to determine the optimal terminal time; see, e.g., Sect. 9.1. Case (ii) T is ﬁxed and the problem is that of maximizing the objective function involving the salvage value of the ending state x(T ), which in this case can be written simply as S[x(T )]. For the ﬁxed-end-point problem and for the inﬁnite horizon problem, it does not usually make much sense to deﬁne a salvage value function. Remark 3.16 One important model type that we did not include in Table 3.3 is the impulse control model of Bensoussan and Lions (1975). In this model, an inﬁnite control is instantaneously exerted on a state variable in order to cause a ﬁnite jump in its value. This model is particularly appropriate for the instantaneous reordering of inventory as required in lot-size models; see Bensoussan et al. (1974). Further discussion of impulse control is given in Sect. D.9. Exercises for Chapter 3 E 3.1 Consider the constraint set Ω = {(u1 , u2 )|0 ≤ u1 ≤ x, −1 ≤ u2 ≤ u1 }. Write these in the form shown in (3.3).

114

3. The Maximum Principle: Mixed Inequality Constraints

E 3.2 Find the reachable set X, deﬁned in Sect. 3.1, if x and u satisfy x˙ = u − 1, x0 = 5, −1 ≤ u ≤ 1, and T = 3. E 3.3 Assume the constraint (3.3) to be of the form g(u, t) ≥ 0, i.e., g does not contain x explicitly, and assume x(T ) is free. Apply the Lagrangian form of the maximum principle and derive the Hamiltonian form (2.31) with Ω(t) = {u|g(u, t) ≥ 0}. Assume g(u, t) to be of the form α ≤ u ≤ β. E 3.4 Use the Lagrangian form of the maximum principle to obtain the optimal control for the following problem: max{J = x1 (2)} subject to x˙ 1 (t) = u1 − u2 , x1 (0) = 2, x˙ 2 (t) = u2 ,

x2 (0) = 1,

and the constraints u1 (t) ≥ u2 (t), 0 ≤ u1 (t) ≤ x2 (t), 0 ≤ u2 (t) ≤ 2, 0 ≤ t ≤ 2. An interpretation of this problem is that x1 (t) is the stock of steel at time t and x2 (t) is the total capacity of the steel mill at time t. Production of steel at rate u1 , which is bounded by the current steel mill capacity, can be split into u2 and u1 − u2 , where u2 goes into increasing the steel mill capacity and u1 − u2 adds to the stock of steel. The objective is to build as large a stockpile of steel as possible by time T = 2. With this interpretation, we clearly need to have x1 (t) ≥ 0 and x2 (t) ≥ 0. However, it is easily seen that these constraints are automatically satisﬁed for every feasible solution of the problem. You may ﬁnd it interesting to show why this is true. (It is possible to make the problem more interesting by assuming an exogenous demand d for steel so that x˙ 1 = u1 − u2 − d.) E 3.5 Specialize the terminal condition (3.13) in the one-dimensional case (i.e., n = 1) with Y (T ) = Y = [x, x ¯] for each T > 0, where x and x ¯ are two constants satisfying x ¯ > x. Use (3.12) to derive (3.14).

Exercises for Chapter 3

115

E 3.6 Obtain the optimal value J ∗ (T ) of the objective function for Example 3.5 for a given terminal time T, and then maximize it with respect to T by using the conditions dJ ∗ (T )/dT = 0. Show that you get the same optimal T ∗ as the one obtained for Example 3.5 by using (3.77). E 3.7 Check that the solution of Example 3.1 satisﬁes the suﬃciency conditions in Theorem 3.1. E 3.8 Starting from (3.15), obtain the current-value version (3.44) for the problem deﬁned by (3.27) and (3.28). Show further that if we were to require the function ψ to also depend on T, i.e. if S(x, T ) = ψ(x, T )e−ρT then the left-hand side of condition (3.44) would be modiﬁed to H[x∗ (T ∗ ), u∗ (T ∗ ), λ(T ∗ ), T ∗ ] + ψ T [x∗ (T ∗ ), T ∗ ] − ρψ[x∗ (T ∗ ), T ∗ ]. E 3.9 Develop the current-value formulation of Sect. 3.3 for a timevarying nonnegative discount rate ρ(t), by replacing the factors e−ρt and e−ρT in (3.28), respectively, by α(t) = e−

t 0

ρ(s)ds

and α(T ) = e−

T 0

ρ(s)ds

.

E 3.10 Begin with (3.54) and perform the steps leading to (3.55). E 3.11 Optimal Consumption of An Initial Investment Over a Finite Horizon. Begin with an initial investment of x0 . Assets x(t) at time t earn at the rate of r per dollar per unit time. A portion of the earnings is consumed at a rate of c(t) per unit time at time t, while the remainder is invested. Neither a negative consumption rate nor a consumption rate exceeding the earnings is allowed. Assets depreciate at the constant rate δ. Assume r > δ+ρ, where ρ is the discount rate applied on consumption. Find the optimal consumption rate over a ﬁnite horizon T such that the present value of the consumption stream over the ﬁnite horizon is maximized. Assume that T is suﬃciently large. Let us note that the optimal capital accumulation model treated in Sect. 11.1.1 represents a generalization of this problem. E 3.12 Show that if we require W (T ) = ε > 0, ε small, instead of W (T ) = 0 in Example 3.2, then the optimal value of the objective function will decrease by an amount βε = ε(1 − erT )/rW0 + o(ε). E 3.13 Recall Exercise 2.18 of the leaky reservoir in Chap. 2. In this problem there was no explicit constraint on the total amount of water

116

3. The Maximum Principle: Mixed Inequality Constraints

available. Suppose we impose the following isoperimetric constraint on that problem: 100 udt = K, 0

where K > 0 is the total amount of water which must be used. Assume also that the reservoir has inﬁnite capacity. Re-solve this problem for various values of K and the objective functions in parts (a) and (b) of Exercise 2.18. E 3.14 From the transversality conditions for the general terminal constraints in Row 5 of Table 3.1, derive the transversality conditions in Row 1 for the free-end-point case, in Row 2 for the ﬁxed-end-point case, and in Rows 3 and 4 for the one-sided constraint cases. Assume ψ(x) = 0, i.e., there is no salvage value and X = E 1 for simplicity. E 3.15 For solving Example 3.3, consider case (ii) by starting with t∗ = 2, and show that the maximum principle will not be satisﬁed in this case. E 3.16 Rework Example 3.4 with T = 4 and the following diﬀerent terminal conditions: (a) x(4) unconstrained, (b) x(4) = 1, (c) x(4) ≤ 1, (d) x(4) ≥ 1. E 3.17 Rework Example 3.4 with the terminal condition (3.70) replaced by x(2) ≥ ε, where ε is small. Verify that the change in the optimal value of the objective function is −ε/2 ≈ −αε + o(ε), as stipulated in Remark 3.6. E 3.18 Introduce a terminal value in Example 3.4 as follows: 2 max J = (−x)dt + Bx(2) subject to

0

x˙ = u, x(0) = 1, x(2) ≥ 0, i.e., Y = [0, ∞) in Table 3.1, Row 3, −1 ≤ u ≤ 1.

Note that for B = 0, the problem is the same as Example 3.4. Solve this problem for B = 1/2, 1, 3/2, 2, 3. Conclude that for B ≥ 2, the solution for the state variable does not change.

Exercises for Chapter 3

117

E 3.19 In Example 3.6, determine the optimal control and the corresponding state trajectory starting at the point (-4,6), which lies above the switching curve. E 3.20 Carry out the synthesis of the optimal control for Example 3.6 when the starting point (x0 , y0 ) lies below the switching curve. E 3.21 Use the results of Exercise 3.20 to ﬁnd the optimal control and the corresponding trajectory starting at the point (−1, −1). E 3.22 Find the optimal control, the minimum time, and the corresponding trajectory for Example 3.6 starting at the point (−2, 2), which lies on the switching curve. E 3.23 What is the shortest time in which a passenger can be transported in a ballistic missile from Los Angeles to New York? Assume that a missile with the ultimate mechanical and thermodynamical properties is available, but that the passenger imposes the restraint that the maximum acceleration or deceleration is 100 ft/s2 . The missile starts from rest in Los Angeles and stops in New York. Assume that the path is a straight line of length 2400 miles and ignore the rotation and curvature of the earth. E 3.24 In the time-optimal control problem (3.90), replace the state equations by ¯ > x0 , x˙ = ay, x(0) = x0 ≥ 0, x(T ) = x y˙ = u, y(0) = y0 ≥ 0, y(T ) = 0, and the control constraint by u ∈ Ω = [Umin , Umax ]. Assume a > 0 and Umax > 0 > Umin . Observe here that x(t) could be interpreted as the cumulative value of gold mined by a gold-producing country and y(t) could be interpreted as the total value of gold-mining machinery employed by the country at time t ≥ 0. The required machinery is to be imported. Because of some inertia in the world market for the machinery, the country cannot control y(t) directly, but is able to control its rate of change y(t). ˙ Thus u(t) represents at time t, the import rate of the machinery when positive and the export rate when negative. The terminal value x ¯ represents the required amount of gold to be produced in a minimum possible time. Obtain the optimal solution.

118

3. The Maximum Principle: Mixed Inequality Constraints

E 3.25 Solve the following minimum weighted energy and time problem: T 1 −( )(u2 + 1)dt max J = u,T 2 0 subject to x˙ = u, x(0) = 5, x(T ) = 0, and the control constraint |u| ≤ 2. Hint. Use (3.77) to determine T ∗ , the optimal value of T. E 3.26 Rework Exercise 3.25 with the new integrand F −(1/2)(u2 + 16) in the objective function.

=

Hint: Note that use of (3.77) gives an infeasible u. This means that we should look for a boundary solution for u. To obtain this, calculate J ∗ (T ) as deﬁned in Exercise 3.6, and then choose T to maximize it. In doing so, take care to see that x(T ) = 0, and the control constraint is satisﬁed. E 3.27 Exercise 3.26 becomes a minimum energy problem if we set F = −u2 /2. Show that the Hamiltonian maximizing condition of the maximum principle implies u∗ = k, where k is a constant. Note that the application of (3.77) implies that k = 0, which gives x(t) = 5 for all t ≥ 0 so that the terminal condition x(T ) = 0 cannot be satisﬁed. To see that there exists no optimal control in this situation, let k < 0 and compute J ∗ . It is now possible to see that limk→0 J ∗ = 0. This means that we can make the objective function value as close to zero as we wish, but not equal to zero. Note that in this case there are no feasible solutions satisfying the necessary conditions so we cannot check the suﬃciency conditions; see the last paragraph of Sect. 2.1.4. E 3.28 Show that every feasible control of the problem T max J = −udt T,u

0

subject to x˙ = u, x(0) = x0 , x(T ) = 0, |u| ≤ q, where q > 0, is an optimal control.

Exercises for Chapter 3

119

E 3.29 Let x0 > 0 be the initial velocity of a rocket. Let u be the amount of acceleration (or deceleration) caused by applying a force which consumes fuel at the rate |u|. We want to bring the rocket to rest using minimum total amount of fuel. Hence, we have the following optimal control problem: T max J = −|u|dt T,u

0

subject to x˙ = u, x(0) = x0 , x(T ) = 0, −1 ≤ u ≤ +1. Hint: Use (3.113)–(3.115) to deal with |u|. Show that for x0 > 0, say x0 = 5, every feasible control is optimal. E 3.30 Analyze Exercise 3.29 with the state equation x˙ = −ax + u, where a > 0. Show that no optimal control exists for the problem. E 3.31 By using the maximum principle, show that the problem ⎧ 1 ⎪ ⎪ max xdt ⎪ ⎪ ⎪ 0 ⎪ ⎪ ⎪ ⎪ ⎨ subject to ⎪ ⎪ ⎪ x˙ = x + u, x(0) = 0, ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 1 − u ≥ 0, 1 + u ≥ 0, 2 − x − u ≥ 0, has the optimal control

u∗ (t) =

⎧ ⎪ ⎨ 1,

t ∈ [0, ln 2],

⎪ ⎩ 1 + 2ln2 − 2t, t ∈ (ln 2, 1].

Also, provide the values of the state variable, the adjoint variable, and the Lagrange multipliers along the optimal path.

120

3. The Maximum Principle: Mixed Inequality Constraints

E 3.32 If, in Exercise 3.31, we perturb the constraint 2 − x − u ≥ 0 by 2 − x − u ≥ ε, where ε is small, then show that the change in value of the objective function equals

1

ε 0

μ3 dt + o(ε),

where μ3 is the Lagrange multiplier associated with the constraint 2−x− u ≥ 0 in Exercise 3.31. Moreover, if ε < 0, implying that we are relaxing the constraint, then verify that the change in the objective function is positive. E 3.33 Obtain the value function V (x, t) explicitly in Exercise 3.31 for every x ∈ E 1 and t ∈ [0, 1]. Furthermore, verify that λ(t) = Vx (x∗ (t), t), t ∈ [0, 1], where λ(t) is the adjoint variable obtained in the solution of Exercise 3.31. E 3.34 Solve the problem: max J = u,T

T 0

[−2 + (1 − u(t))x(t)]dt

subject to x˙ = u, x(0) = 0, x(T ) ≥ 1, u ∈ [0, 1], T ∈ [1, 8]. Hint: First, show that u∗ = bang[0, 1; λ − x] and that control can switch at most once from 1 to 0. Then, let t∗ (T ) denote that switching time, if any, for a given T ∈ [1, 8]. Consider three cases: (i) T = 1, (ii) 1 < T < 8, and (iii) T = 8. Note that λ(t∗ (T )) − x(t∗ (T )) = 0. Use (3.15) in case (ii). Find the optimal solution in each of the three cases. The best of these solutions will be the solution of the problem. E 3.35 Consider the problem: T [−3 − u(t) + x(t)]dt max J = u,T

0

subject to x˙ = u, x(0) = 0, x(T ) ≥ 1,

Exercises for Chapter 3

121

u ∈ [0, 1], √ T ∈ [1, 4 + 2 2]. The problem has two diﬀerent optimal solutions with diﬀerent values for optimal T ∗ . Find both of these solutions. E 3.36 Perform the following: (a) Find the optimal consumption rate C ∗ (t), t ∈ [0, T ], in the problem: T −ρt max J = e ln C(t)dt 0

subject to ˙ (t) = −C(t), W (0) = W0 , W where T is given and ρ > 0. (b) Assume that T is not given in (a), and is to be chosen optimally. Show for this free terminal time version that the optimal T ∗ decreases as the discount rate ρ increases. Hint: It is possible to obtain dT ∗ /dρ by implicit diﬀerentiation. E 3.37 An example, which illustrates that lim λ(t) = 0

t→∞

is not a necessary transversality condition in general, is: ∞ (1 − x)udt max J = 0

such that x˙ = (1 − x)u, x(0) = 0, 0 ≤ u ≤ 1. Show this by ﬁnding an optimal control. E 3.38 Show that the limiting conditions in the rightmost column of Rows 2, 3, and 4 in Table 3.1 imply (3.98) when T → ∞.

122

3. The Maximum Principle: Mixed Inequality Constraints

E 3.39 Consider the regulator problem deﬁned by the scalar equation x˙ = u, x(0) = x0 , with the objective function J =−

0

∞ 4 x

u2 + 4 2

dt.

¯ = (0, 0, 0), (a) Show that the long-term stationary equilibrium (¯ x, u ¯, λ) ∗ and conclude that in feedback form u (x) = u ¯ = 0 when x = x ¯ = 0. ∗

(b) By using the maximum principle and the relation u˙ ∗ = dudx(x) x, ˙ derive a diﬀerential equation for the optimal feedback control u∗ (x) and solve it with the boundary condition u∗ (0) = 0 to obtain ⎧ √ ⎪ ⎪ ⎪ −x2 / 2, x > 0, ⎪ ⎪ ⎨ ∗ u (x) = 0, x = 0, ⎪ ⎪ ⎪ ⎪ √ ⎪ ⎩ +x2 / 2, x < 0. (c) Solve for x∗ (t) and λ(t) and show that limt→∞ x∗ (t) = 0 and that the limiting condition (3.99), i.e., limt→∞ λ(t) = 0, holds for this problem. E 3.40 Show that for the problem (3.97) without the constraint g(x, u) ≥ 0, the optimal value of the objective function J ∗ = H(x0 , u∗ (0), λ(0))/ρ. See Grass et al. (2008). ¯ ≥ 0 and λ ¯W ¯ = 0 in E 3.41 Apply (3.108), along with the requirement λ view of the constraint (3.102), to Example 3.7 to verify that the long-run stationary equilibrium is as shown in (3.110). E 3.42 For a stationary system as deﬁned in Sect. 3.6, show that dH = ρλf (x∗ (t), u∗ (t)) dt and

dH pv = −ρe−ρt φ(x∗ (t), u∗ (t)) dt along the optimal path. Also, contrast these results with that of Exercise 2.9.

Exercises for Chapter 3

123

E 3.43 Consider the inventory problem: ∞ −ρt 2 2 −e [(I − I1 ) + (P − P1 ) ]dt max J = 0

subject to I˙ = P − S, I(0) = I0 , where I denotes inventory level, P denotes production rate, and S denotes a given constant demand rate. (a) Find the optimal long-run stationary equilibrium, i.e., the turnpike deﬁned in (3.107). (b) Find the Golden Rule by setting I˙ = 0 in the state equation, solve for P, and substitute it into the integrand of the objective function. Then, maximize the integrand with respect to I. (c) Verify that the Golden Rule inventory level obtained in (b) is the same as the turnpike inventory level found in (a) when ρ = 0.

Chapter 4

The Maximum Principle: Pure State and Mixed Inequality Constraints In Chap. 2 we addressed optimal control problems having constraints only on control variables. We extended the discussion in Chap. 3 to treat mixed constraints that may involve state variables in addition to control variables. Often in management science and economics problems there are nonnegativity constraints on state variables, such as inventory levels or wealth. These constraints do not include control variables. Also, there may be more general inequality constraints only on state variables, which include constraints that require certain state variables to remain nonnegative. Such constraints are known as pure state variable inequality constraints or, simply, pure state constraints. Pure state constraints are more diﬃcult to deal with than mixed constraints. We can intuitively appreciate this fact by keeping in mind that only control variables are under the direct inﬂuence of the decision maker. This enables the decision maker, when a mixed constraint becomes tight, to choose from the controls that would keep it tight for as long as required for optimality. Whereas with pure state constraints, the situation is diﬀerent and more complicated. That is, when a constraint becomes tight, it does not provide any direct information to the decision maker on how to choose values for the control variables so as not to © Springer Nature Switzerland AG 2019 S. P. Sethi, Optimal Control Theory, https://doi.org/10.1007/978-3-319-98237-3 4

125

126

4. The Maximum Principle: Pure State and Mixed Constraints

violate the constraint. Hence, considerable changes in the controls may be required to keep the constraint tight if needed for optimality. This chapter considers pure state constraints together with mixed constraints. In the literature there are two ways of handling pure state constraints: direct and indirect. The direct method associates a multiplier with each constraint for appending it to the Hamiltonian to form the Lagrangian, and then proceeds in much the same way as in Chap. 3 dealing with mixed constraints. In the indirect method, the choice of controls, when a pure constraint is active, must be further limited by constraining approximately the value of the derivative of the active state constraint with respect to time. This derivative will involve time derivatives of the state variables, which can be written in terms of the control and state variables through the use of the state equations. Thus, the restrictions on the time derivatives of the pure state constraints are transformed in the form of mixed constraints, and these will be appended instead to the Hamiltonian to form the Lagrangian. Because the pure state constraints are adjoined in this indirect fashion, the corresponding Lagrange multipliers must satisfy some complementary slackness conditions in addition to those mentioned in Chap. 3. With the formulation of the Lagrangian in each approach, we will write the respective maximum principle, where the choice of control will come from maximizing the Hamiltonian subject to both pure state constraints and mixed constraints. We will ﬁnd, however, in contrast to Chap. 3, that in both approaches, the adjoint functions may be required to have jumps at those times where the pure state constraints become tight. We begin with a simple example in Sect. 4.1 to motivate the necessity of possible jumps in the adjoint functions. Section 4.2 formulates the problem with pure state constraints along with the required assumptions. In Sect. 4.3, we use the direct method for stating the maximum principle necessary conditions for solving such problems. Suﬃciency conditions are stated in Sect. 4.4. Section 4.5 is devoted to developing the maximum principle for the indirect method, which involves adjoining the ﬁrst derivative of the pure state constraints to form the Lagrangian function and imposing some additional constraints on the Lagrange multipliers of the resulting formulation. Also, the adjoint variables and the Lagrange multipliers arising in this method will be related to those arising in the direct method. Finally, the current-value form of the maximum principle for the indirect method is described in Sect. 4.6.

4.1. Jumps in Marginal Valuations

4.1

127

Jumps in Marginal Valuations

In this section, we formulate an optimal control problem with a pure constraint, which can be solved merely by inspection and which exhibits a discontinuous marginal valuation of the state variable. Since the adjoint variables in Chaps. 2 and 3 provide these marginal valuations and since we would like this feature to continue, we must allow the adjoint variables to have jumps if the marginal valuations can be discontinuous. This will enable us to formulate a maximum principle in the next section, which is similar to (3.10) with the exception that the adjoint variables, and therefore also the Hamiltonian, may have possible jumps satisfying some jump conditions. Example 4.1 Consider the problem with a pure state constraint: 3 max J = −udt (4.1) 0

subject to x˙ = u, x(0) = 0,

(4.2)

0 ≤ u ≤ 3,

(4.3)

x − 1 + (t − 2)2 ≥ 0.

(4.4)

Solution From the objective function (4.1), one can see that it is good to have low values of u. If we use u = 0 to begin with, we see that x(t) = 0 as long as u(t) = 0. At t = 1, x(1) = 0 and the constraint (4.4) is satisﬁed with an equality. But continuing with u(t) = 0 beyond t = 1 is not feasible since x(t) = 0 would not satisfy the constraint (4.4) just after t = 1. In Fig. 4.1, we see that the lowest possible feasible state trajectory from t = 1 to t = 2 satisﬁes the state constraint (4.4) with an equality. In order not to violate the constraint (4.4), its ﬁrst time derivative u(t)+ 2(t − 2) must be nonnegative. This gives us u(t) = 2(2 − t) to be the lowest feasible value for the control. This value will make the state x(t) ride along the constraint boundary until t = 2, at which point u(2) = 0; see Fig. 4.1. Continuing with u(t) = 2(2 − t) beyond t = 2 will make u(t) negative, and violate the lower bound in (4.3). It is easy to see, however, that u(t) = 0, t ≥ 2, is the lowest feasible value, which can be followed all the way to the terminal time t = 3.

128

4. The Maximum Principle: Pure State and Mixed Constraints

Figure 4.1: Feasible state space and optimal state trajectory for Examples 4.1 and 4.4 It can be seen from Fig. 4.1 that the bold trajectory is the lowest possible feasible state trajectory on the entire time interval [0,3]. Moreover, it is obvious that the lowest possible feasible control is used at any given t ∈ [0, 3], and therefore, the solution we have found is optimal. We can now restate the values of the state and control variables that we have obtained: ⎧ ⎧ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ 0, t ∈ [0, 1), 0, t ∈ [0, 1), ⎪ ⎪ ⎪ ⎪ ⎨ ⎨ ∗ x∗ (t) = 1 − (t − 2)2 , t ∈ [1, 2], u (t) = ⎪ 2(2 − t), t ∈ [1, 2], ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ ⎩ 0, t ∈ (2, 3]. 1, t ∈ (2, 3], (4.5) Next we ﬁnd the value function V (x, t) for this problem. It is obvious that the feedback control u∗ (x, t) = 0 is optimal at any point (x, t) when x ≥ 1 or when (x, t) is on the right-hand side of the parabola in Fig. 4.1. Thus, V (x, t) = 0 on such points. On the other hand, when x ∈ [0, 1] and it is on the left-hand side of the parabola, the optimal trajectory is very similar to the one shown in Fig. 4.1. Speciﬁcally, the control is zero until it hits the trajectory at √ time τ = 2 − 1 − x. Then, the control switches to 2(2 − s) for s ∈ (τ , 2)

4.2. Optimal Control Problem with Pure and Mixed Constraints

129

to climb along the left-hand side of the parabola to reach its peak, and then switches back to zero on the time interval [2,3]. Thus, in this case, 2 3 τ 0ds − 2(2 − s)ds − 0ds V (x, t) = − t

=

2

$2

s − 4s

2

τ √

2− 1−x

= (x − 1).

Thus, we have the value function ⎧ ⎪ ⎪ ⎪ 0, x ≥ 1, t ∈ [0, 3], ⎪ ⎪ ⎨ V (x, t) = x − 1, x ≥ 1 − (t − 2)2 , t ∈ [0, 2), ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 0, 1 − (t − 2)2 ≤ x ≤ 1, t ∈ [2, 3]. This gives us the marginal valuation along the optimal path x∗ (t) given in (4.5) as ⎧ ⎪ ⎨ 1, t ∈ [0, 2), ∗ Vx (x (t), t) = (4.6) ⎪ ⎩ 0, t ∈ [2, 3]. We can now see that this marginal valuation is discontinuous at t = 2, and it has a downward jump of size 1 at that time. The maximum principle that we will state in Sect. 4.3 will have certain jump conditions in order to accommodate problems like Example 4.1. Indeed in Example 4.2, we will apply the maximum principle of Sect. 4.3 to the problem in Example 4.1, and see that the adjoint variable that represents the marginal valuation along the optimal path will have a jump consistent with (4.6). In the next section, we state the general optimal control problem that is the subject of this chapter.

4.2

The Optimal Control Problem with Pure and Mixed Constraints

We will append to the problem (3.7) considered in Chap. 3, the pure state variable inequality constraint of type h(x, t) ≥ 0,

(4.7)

130

4. The Maximum Principle: Pure State and Mixed Constraints

where we assume function h : E n × E 1 → E p to be continuously differentiable in all its arguments. By the deﬁnition of function h, (4.7) represents a set of p constraints hi (x, t) ≥ 0, i = 1, 2, . . . , p. It is noted that the constraint hi ≥ 0 is called a constraint of rth order if the rth time derivative of hi is the ﬁrst time a term in control u appears in the expression by putting f (x, u, t) for x˙ after each diﬀerentiation. It is through this expression that the control acts to satisfy the constraint hi ≥ 0. The value of r is referred to as the order of the constraint. Of course, if the constraint hi is of order r, then we would require hi to be r times continuously diﬀerentiable. Except for Exercise 4.12, in this book we will consider only ﬁrst-order constraints, i.e., r = 1. For such constraints, the ﬁrst-time derivative of h has terms in u. Thus, we can deﬁne h1 (x, u, t) as follows: h1 =

dh ∂h ∂h = f+ . dt ∂x ∂t

(4.8)

In the important special case of the nonnegativity constraint x(t) ≥ 0,

t ∈ [0, T ],

(4.9)

(4.8) is simply h1 = f. For an upper bound constant x(t) ≤ M, written as M − x(t) ≥ 0, t ∈ [0, T ], (4.10) (4.8) gives h1 = −f. These will be of order one because the function f (x, u, t) usually contains terms in u. As in Chap. 3, the constraints (4.7) need also to satisfy a full-rank type constraint qualiﬁcation before a maximum principle can be derived. With respect to the ith constraint hi (x, t) ≥ 0, an interval (θ1 , θ2 ) ⊂ [0, T ] with θ1 < θ2 is called an interior interval if hi (x(t), t) > 0 for all t ∈ (θ1 , θ2 ). If the optimal trajectory “hits the boundary,” i.e., satisﬁes hi (x(t), t) = 0 for τ 1 ≤ t ≤ τ 2 for some i, then [τ 1 , τ 2 ] is called a boundary interval. An instant τ 1 is called an entry time if there is an interior interval ending at t = τ 1 and a boundary interval starting at τ 1 . Correspondingly, τ 2 is called an exit time if a boundary interval ends and an interior interval starts at τ 2 . If the trajectory just touches the boundary at time τ , i.e., h(x(τ ), τ ) = 0 and if the trajectory is in the interior just before and just after τ , then τ is called a contact time. Taken together, entry, exit, and contact times are called junction times. In this book we shall not consider problems that require optimal state trajectories to have countably many junction times. In other words, we

4.2. Optimal Control Problem with Pure and Mixed Constraints

131

shall state the maximum principle necessary optimality conditions for state trajectories having only ﬁnitely many junction times. Also, all of the applications studied in this book exhibit optimal state trajectories containing ﬁnitely many junction times or no junction times. Throughout the book, we will assume that the constraint qualiﬁcation introduced in Sect. 3.1 as well as the following full-rank condition on any boundary interval [τ 1 , τ 2 ] hold: ⎤ ⎡ 1 ⎢ ∂h1 /∂u ⎥ ⎥ ⎢ ⎥ ⎢ 1 ⎢ ∂h2 /∂u ⎥ ⎥ ⎢ rank ⎢ ⎥ = pˆ, . ⎥ ⎢ .. ⎥ ⎢ ⎥ ⎢ ⎦ ⎣ ∂h1pˆ/∂u

where for t ∈ [τ 1 , τ 2 ], hi (x∗ (t), t) = 0, i = 1, 2, . . . , pˆ ≤ p and

hi (x∗ (t), t) > 0, i = pˆ + 1, . . . , p.

Note that this full-rank condition on the constraints (4.7) is written when the order of each of the constraints in (4.7) is one. For the general case of higher-order constraints, see Hartl et al. (1995). Let us recapitulate the optimal control problem for which we will state a direct maximum principle in the next section. The problem is ⎧ T ⎪ ⎪ max J = F (x, u, t)dt + S[x(T ), T ] , ⎪ ⎪ ⎪ 0 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ subject to ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ x˙ = f (x, u, t), x(0) = x0 , ⎪ ⎪ ⎨ (4.11) g(x, u, t) ≥ 0, ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ h(x, t) ≥ 0, ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ a(x(T ), T ) ≥ 0, ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ b(x(T ), T ) = 0.

132

4. The Maximum Principle: Pure State and Mixed Constraints

Important special cases of the mixed constraint g(x, u, t) ≥ 0 are ui ∈ [0, M ] for M > 0 and ui (t) ∈ [0, xi (t)], and those of the terminal constraints a(x(T ), T ) ≥ 0 and a(x(T ), T ) = 0 are xi (T ) ≥ k and xi (T ) = k, respectively, where k is a constant. Likewise, the special cases of the pure constraints h(x, t) ≥ 0 are xi ≥ 0 and xi ≤ M, for which hxi = +1 and hxi = −1, respectively, and ht = 0.

4.3

The Maximum Principle: Direct Method

For the problem (4.11), we will now state the direct maximum principle which includes the discussion above and the required jump conditions. For details, see Dubovitskii and Milyutin (1965), Feichtinger and Hartl (1986), Hartl et al. (1995), Boccia et al. (2016), and references therein. We will use superscript d on various multipliers that arise in the direct method, to distinguish them from the corresponding multipliers (which are not superscripted) that arise in the indirect method, to be discussed in Sect. 4.5. Naturally, it will not be necessary to superscript the multipliers that are known to remain the same in both methods. To formulate the maximum principle for the problem (4.11), we deﬁne the Hamiltonian function H d : E n × E m × E 1 → E 1 as H d = F (x, u, t) + λd f (x, u, t) and the Lagrangian function Ld : E n × E m × E n × E q × E p × E 1 → E 1 as Ld (x, u, λd , μ, η d , t) = H d (x, u, λd , t) + μg(x, u, t) + η d h(x, t).

(4.12)

The maximum principle states the necessary conditions for u∗ (with the corresponding state trajectory x∗ ) to be optimal. The conditions are that there exist an adjoint function λd , which may be discontinuous at a time in a boundary interval or a contact time, multiplier functions μ, α, β, γ d , η d , and a jump parameter ζ d (τ ), at each time τ , where λd is discontinuous, such that the following (4.13) holds:

4.3. The Maximum Principle: Direct Method

133

x˙ ∗ = f (x∗ , u∗ , t), x∗ (0) = x0 , satisfying constraints g(x∗ , u∗ , t) ≥ 0, h(x∗ , t) ≥ 0, and the terminal constraints a(x∗ (T ), T ) ≥ 0 and b(x∗ (T ), T ) = 0; d λ˙ = −Lx [x∗ , u∗ , λd , μ, η d , t]

with the transversality conditions λd (T − ) = Sx (x∗ (T ), T ) + αax (x∗ (T ), T ) + βbx (x∗ (T ), T ) +γ d hx (x∗ (T ), T ), and α ≥ 0, αa(x∗ (T ), T ) = 0, γ d ≥ 0, γ d h(x∗ (T ), T ) = 0; the Hamiltonian maximizing condition H d [x∗ (t), u∗ (t), λd (t), t] ≥ H d [x∗ (t), u, λd (t), t] at each t ∈ [0, T ] for all u satisfying g[x∗ (t), u, t] ≥ 0;

(4.13)

the jump conditions at any time τ , where λd is discontinuous, are λd (τ − ) = λd (τ + ) + ζ d (τ )hx (x∗ (τ ), τ ) and H d [x∗ (τ ), u∗ (τ − ), λd (τ − ), τ ] = H d [x∗ (τ ), u∗ (τ + ), λd (τ + ), τ ] −ζ d (τ )ht (x∗ (τ ), τ ); the Lagrange multipliers μ(t) are such that ∂Ld /∂u|u=u∗ (t) = 0, dH d /dt = dLd /dt = ∂Ld /∂t, and the complementary slackness conditions μ(t) ≥ 0, μ(t)g(x∗ , u∗ , t) = 0, η(t) ≥ 0, η d (t)h(x∗ (t), t) = 0, and ζ d (τ ) ≥ 0, ζ d (τ )h(x∗ (τ ), τ ) = 0 hold.

134

4. The Maximum Principle: Pure State and Mixed Constraints

As in the previous chapters, λd (t) has the marginal value interpretation. Therefore, while it is not needed for the application of the maximum principle (4.13), we can trivially set λd (T ) = Sx (x∗ (T ), T ).

(4.14)

If T is also a decision variable constrained to lie in the interval [T1 , T2 ], 0 ≤ T1 < T2 < ∞, then in addition to (4.13), if T ∗ is the optimal terminal time, it must satisfy a condition similar to (3.15) and (3.81), i.e., H d [x∗ (T ∗ ), u∗ (T ∗− ), λd (T ∗− ), T ∗ ] + ST [x∗ (T ∗ ), T ∗ ] + αaT [x∗ (T ∗ ), T ∗ ]

+βbT [x∗ (T ∗ ), T ∗ ] + γ d hT [x∗ (T ∗ ), T ∗ ]

⎧ ⎪ ⎪ ⎪ ≤0 ⎪ ⎪ ⎨

=0 ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ ≥0

if T ∗ = T1 , if T ∗ ∈ (T1 , T2 ),

(4.15)

if T ∗ = T2 .

Remark 4.1 In most practical examples, λd and H d will only jump at junction times. However, in some cases a discontinuity may occur at a time in the interior of a boundary interval, e.g., when a mixed constraint becomes active at that time. Remark 4.2 It is known that the adjoint function λd is continuous at a junction time τ , i.e., ζ d (τ ) = 0, if (i) the entry or exit at time τ is non-tangential, i.e., if h1 (x∗ (τ ), u∗ (τ ), τ ) = 0, or (ii) if the control u∗ is continuous at τ and the ⎤ ⎡ 0 ⎥ ⎢ ∂g/∂u diag(g) rank ⎣ ⎦ = m + p, 0 diag(h) ∂h1 /∂u when evaluated at x∗ (τ ) and u∗ (τ ). We will see that the jump conditions on the adjoint variables in (4.13) will give us precisely the jump in Example 4.2, where we will apply the direct maximum principle to the problem in Example 4.1. The jump condition on H d in (4.13) requires that the Hamiltonian should be continuous at τ if ht (x∗ (τ ), τ ) = 0. The continuity of the Hamiltonian (in case ht = 0) makes intuitive sense when considered in the light of its interpretation given in Sect. 2.2.4.

4.3. The Maximum Principle: Direct Method

135

This brief discussion of the jump conditions, limited here only to ﬁrst-order pure state constraints, is far from complete, and a detailed discussion is beyond the scope of this book. An interested reader should consult the comprehensive survey by Hartl et al. (1995). For an example with a second-order state constraint, see Maurer (1977). Needless to say, computational methods are required to solve problems with general inequality constraints in all but the simplest of the cases. The reader should consult the excellent book by Teo et al. (1991) and references therein for computational procedures and software; see also Polak et al. (1993), Bulirsch and Kraft (1994), Bryson (1998), and Pytlak and Vinter (1993, 1999). A MATLAB based software, used for solving ﬁnite and inﬁnite horizon optimal control problems with pure state and mixed inequality constraints, is available at http://orcos. tuwien.ac.at/research/ocmat software/. Example 4.2 Apply the direct maximum principle (4.13) to solve the problem in Example 4.1. Solution Since we already have optimal u∗ and x∗ as obtained in (4.5), we can use these in (4.13) to obtain λd , μ1 , μ2 , γ d , η d , and ζ d . Thus, H d = −u + λd u, Ld = H d + μ1 u + μ2 (3 − u) + η d [x − 1 + (t − 2)2 ],

(4.16) (4.17)

Ldu = −1 + λd + μ1 − μ2 = 0,

(4.18)

d λ˙ = −Ldx = −η d , λd (3− ) = γ d ,

(4.19)

γ d [x∗ (3) − 1 + (3 − 2)2 ] = 0,

(4.20)

μ1 ≥ 0, μ1 u∗ = 0, μ2 ≥ 0, μ2 (3 − u∗ ) = 0,

(4.21)

η d ≥ 0, η d [x∗ (t) − 1 + (t − 2)2 ] = 0,

(4.22)

and if λd is discontinuous for some τ ∈ [1, 2], the boundary interval as seen from Fig. 4.1, then λd (τ − ) = λd (τ + ) + ζ d (τ ), ζ d (τ ) ≥ 0,

(4.23)

− u∗ (τ − ) + λd (τ − )u∗ (τ − ) = −u∗ (τ + ) + λd (τ + )u∗ (τ + ) − ζ d (τ )2(τ − 2). (4.24)

136

4. The Maximum Principle: Pure State and Mixed Constraints

Since γ d = 0 from (4.20), we have λd (3−) = 0 from (4.19). Also, we set λd (3) = 0 according to (4.14). d Interval (2,3]: We have η d = 0 from (4.22), and thus λ˙ = 0 from (4.19), giving λd = 0. From (4.18) and (4.21), we have μ1 = 1 > 0 and μ2 = 0. Interval [1,2]: We get μ1 = μ2 = 0 from 0 < u∗ < 3 and (4.21). d Thus, (4.18) implies λd = 1 and (4.19) gives η d = −λ˙ = 0. Thus λd is discontinuous at the exit time τ = 2, and we use (4.23) to see that the jump parameter ζ d (2) = λd (2− ) − λd (2+ ) = 1. Furthermore, it is easy to check that (4.24) also holds at τ = 2. Interval [0,1): Clearly μ2 = 0 from (4.21). Also u∗ = 0 would still be optimal if there were no lower bound constraint on u in this interval. This means that the constraint u ≥ 0 is not binding, giving us μ1 = 0. Then d from (4.18), we have λd = 1. Finally, from (4.19), we have η d = −λ˙ = 0. We can now see that the adjoint variable ⎧ ⎪ ⎨ 1, t ∈ [0, 2), d λ (t) = (4.25) ⎪ ⎩ 0, t ∈ [2, 3], is precisely the same as the marginal valuation Vx (x∗ (t), t) obtained in (4.6). We also see that λd is continuous at time t = 1 where the entry to the constraint is non-tangential as stated in Remark 4.2.

4.4

Suﬃciency Conditions: Direct Method

When ﬁrst-order pure state constraints are present, suﬃciency results are usually stated in terms of the maximum principle using the direct method described in Hartl et al. (1995). We will now state the suﬃciency result for the problem speciﬁed in (4.11). For this purpose, let us deﬁne the maximized Hamiltonian H 0d (x, λd (t), t) =

max

{u|g(x,u,t)≥0}

H d (x, u, λd , t).

(4.26)

See Feichtinger and Hartl (1986) and Seierstad and Sydsæter (1987) for details. Theorem 4.1 Let (x∗ , u∗ , λd , μ, α, β, γ d , η d ) and the jump parameters ζ d (τ ) at each τ , where λd is discontinuous, satisfy the necessary conditions in (4.13). If H 0d (x, λd (t), t) is concave in x at each t ∈ [0, T ], S

4.5. The Maximum Principle: Indirect Method

137

in (3.2) is concave in x, g in (3.3) is quasiconcave in (x, u), h in (4.7) and a in (3.4) are quasiconcave in x, and b in (3.5) is linear in x, then (x∗ , u∗ ) is optimal. We will illustrate an application of this theorem in Example 4.4, which shows that the solution obtained in Example 4.3 is optimal. Theorem 4.1 is written for ﬁnite horizon problems. For inﬁnite horizon problems, this theorem remains valid if the transversality condition on the adjoint variable in (4.29) is modiﬁed along the lines discussed in Sect. 3.6. In concluding this section, we should note that the suﬃciency conditions stated in Theorem 4.1 rely on the presence of appropriate concavity conditions. Suﬃciency conditions can also be obtained without these concavity assumptions. These are called second-order conditions for a local maximum, which require the second variation on the linearized state equation to be negative deﬁnite. For further details on second-order sufﬁciency conditions, the reader is referred to Maurer (1981), Malanowski (1997), and references in Hartl et al. (1995).

4.5

The Maximum Principle: Indirect Method

The main idea underlying the indirect method is that when the pure state constraint (4.7), assumed to be of order one, becomes active, we must require its ﬁrst derivative h1 (x, u, t) in (4.8) to be nonnegative, i.e., h1 (x, u, t) ≥ 0, whenever h(x, t) = 0.

(4.27)

While this is a mixed constraint, it is diﬀerent from those treated in Chap. 3 in the sense that it is imposed only when the constraint (4.8) is tight. Since (4.27) is a mixed constraint, it is tempting to use the maximum principle (3.12) developed in Chap. 3. This can be done provided that we can ﬁnd a way to impose (4.27) only when h(x, t) = 0. One way to accomplish this is to append (4.27) to the Hamiltonian when forming the Lagrangian, by using a multiplier η ≥ 0, i.e., append ηh1 , and require that ηh = 0, which is equivalent to imposing η i hi = 0, i = 1, 2, . . . , p. This means that when hi > 0 for some i, we have η i = 0 and it is then not a part of the Lagrangian. Note that when we require ηh = 0, we do not need to impose ηh1 = 0 as required for mixed constraints. This is because when hi > 0 on an

138

4. The Maximum Principle: Pure State and Mixed Constraints

interval, then η i = 0 and so η i h1i = 0 on that interval. On the other hand, when hi = 0 on an interval, then it is because h1i = 0, and thus, η i h1i = 0 on that interval. In either case, η i h1i = 0. With these observations, we are ready to formulate the indirect maximum principle for the problem (4.11). We form the Lagrangian as L(x, u, λ, μ, η, t) = H(x, u, λ, t) + μg(x, u, t) + ηh1 (x, u, t),

(4.28)

where the Hamiltonian H = F (x, u, t) + λf (x, u, t) as deﬁned in (3.8). We will now state the maximum principle which includes the discussion above and the required jump conditions. The maximum principle states the necessary conditions for u∗ (with the state trajectory x∗ ) to be optimal. These conditions are that there exist an adjoint function λ, which may be discontinuous at each entry or contact time, multiplier functions μ, α, β, γ, η, and a jump parameter ζ(τ ) at each τ , where λd is discontinuous, such that (4.29) on the following page holds. Once again, as before, we can set λ(T ) = Sx (x∗ (T ), T ). If T ∈ [T1 , T2 ] is a decision variable, then (4.15) with λd and γ d replaced by λ and γ, respectively, must also hold. In (4.29), we see that there are jump conditions on the adjoint variables and also the Hamiltonian in the indirect maximum principle. The remarks on the jump condition made in connection with the direct maximum principle (4.13) apply also to the jump conditions in (4.29). In (4.29), we also see a condition η˙ ≤ 0, in addition to the complimentary conditions on η. The presence of this term will become clear after we relate this multiplier to those in the direct maximum principle, which we discuss next. In various applications that are discussed in subsequent chapters of this book, we use the indirect maximum principle. Nevertheless, it is worthwhile to provide relationships between the multipliers of the two approaches, as these will be useful when checking for the suﬃciency conditions of Theorem 4.1, developed in Sect. 4.4.

4.5. The Maximum Principle: Indirect Method

139

x˙ ∗ = f (x∗ , u∗ , t), x∗ (0) = x0 , satisfying constraints g(x∗ , u∗ , t) ≥ 0, h(x∗ , t) ≥ 0, and the terminal constraints a(x∗ (T ), T ) ≥ 0 and b(x∗ (T ), T ) = 0; λ˙ = −Lx [x∗ , u∗ , λ, μ, η, t] with the transversality conditions λ(T − ) = Sx (x∗ (T ), T ) + αax (x∗ (T ), T ) + βbx (x∗ (T ), T ) +γhx (x∗ (T ), T ), and α ≥ 0, αa(x∗ (T ), T ) = 0, γ ≥ 0, γh(x∗ (T ), T ) = 0; the Hamiltonian maximizing condition H[x∗ (t), u∗ (t), λ(t), t] ≥ H[x∗ (t), u, λ(t), t] at each t ∈ [0, T ] for all u satisfying g[x∗ (t), u, t] ≥ 0, and h1i (x∗ (t), u, t) ≥ 0 whenever hi (x∗ (t), t) = 0, i = 1, 2, · · · , p; the jump conditions at any entry/contact time τ , where λ is discontinuous, are λ(τ − ) = λ(τ + ) + ζ(τ )hx (x∗ (τ ), τ ) and H[x∗ (τ ), u∗ (τ − ), λ(τ − ), τ ] = H[x∗ (τ ), u∗ (τ + ), λ(τ + ), τ ] −ζ(τ )ht (x∗ (τ ), τ ); the Lagrange multipliers μ(t) are such that ∂L/∂u|u=u∗ (t) = 0, dH/dt = dL/dt = ∂L/∂t, and the complementary slackness conditions μ(t) ≥ 0, μ(t)g(x∗ , u∗ , t) = 0, η(t) ≥ 0, η(t)h(x∗ (t), t) = 0, and ζ(τ ) ≥ 0, ζ(τ )h(x∗ (τ ), τ ) = 0 hold.

(4.29)

140

4. The Maximum Principle: Pure State and Mixed Constraints

We now obtain the multipliers of the direct maximum principle from those in the indirect maximum principle. Since the multipliers coincide in the interior, we let [τ 1 , τ 2 ] denote a boundary interval and τ a contact time. It is shown in Hartl et al. (1995) that η d (t) = −η(t), ˙ t ∈ (τ 1 , τ 2 ),

(4.30)

λd (t) = λ(t) + η(t)hx (x∗ (t), t), t ∈ (τ 1 , τ 2 ),

(4.31)

Note that η d (t) ≥ 0 in (4.13). Thus, we have η˙ ≤ 0, which we have already included in (4.29). The jump parameter at an entry time τ 1 , an exit time τ 2 , or a contact time τ , respectively, satisﬁes − d d ζ d (τ 1 ) = ζ(τ 1 ) − η(τ + 1 ), ζ (τ 2 ) = η(τ 2 ), ζ (τ ) = ζ(τ ).

(4.32)

By comparing λd (T − ) in (4.13) and λ(T − ) in (4.29) and using (4.31), we have (4.33) γ d = γ + η(T − ). Going the other way, we have τ2 η d (s)ds + ζ d (τ 2 ), t ∈ (τ 1 , τ 2 ), η(t) = t

λ(t) = λd (t) − η(t)h(x∗ (t), t), t ∈ (τ 1 , τ 2 ), d ζ(τ 1 ) = ζ d (τ 1 ) + η(τ + 1 ), ζ(τ 2 ) = 0, ζ(τ ) = ζ (τ ),

γ = γ d − η(T − ). Finally, as we had mentioned earlier, the multipliers μ, α, and β are the same in both methods. Remark 4.3 From (4.30), (4.32), and η d (t) ≥ 0 and ζ d (τ 1 ) ≥ 0 in (4.13), we can obtain the conditions

and

η(t) ˙ ≤0

(4.34)

ζ(τ 1 ) ≥ η(τ + 1 ) at each entry time τ 1 ,

(4.35)

which are useful to know about. Hartl et al. (1995) and Feichtinger and Hartl (1986) also add these conditions to the indirect maximum principle necessary conditions (4.29).

4.5. The Maximum Principle: Indirect Method

141

Remark 4.4 In Exercise 4.12, we discuss the indirect method for higher-order constraints. For further details, see Pontryagin et al. (1962), Bryson and Ho (1975) and Hartl et al. (1995). Example 4.3 Consider the problem: 2 −xdt max J = 0

subject to x˙ = u, x(0) = 1,

(4.36)

u + 1 ≥ 0, 1 − u ≥ 0,

(4.37)

x ≥ 0.

(4.38)

Note that this problem is the same as Example 2.3, except for the nonnegativity constraint (4.38). Solution The Hamiltonian is H = −x + λu, which implies the optimal control to have the form u∗ (x, λ) = bang[−1, 1; λ], whenever x > 0.

(4.39)

When x = 0, we impose x˙ = u ≥ 0 in order to insure that (4.38) holds. Therefore, the optimal control on the state constraint boundary is u∗ (x, λ) = bang[0, 1; λ], whenever x = 0.

(4.40)

Now we form the Lagrangian L = H + μ1 (u + 1) + μ2 (1 − u) + ηu, where μ1 , μ2 , and η satisfy the complementary slackness conditions μ1 ≥ 0, μ1 (u + 1) = 0,

(4.41)

μ2 ≥ 0, μ2 (1 − u) = 0,

(4.42)

η ≥ 0,

ηx = 0.

(4.43)

Furthermore, the optimal trajectory must satisfy ∂L = λ + μ1 − μ2 + η = 0. ∂u

(4.44)

142

4. The Maximum Principle: Pure State and Mixed Constraints

From the Lagrangian we also get ∂L = 1, λ(2− ) = γ ≥ 0, γx(2) = λ(2− )x(2) = 0. λ˙ = − ∂x

(4.45)

It is reasonable to guess that the optimal control u∗ will be the one that keeps x∗ as small as possible, subject to the state constraint (4.38). Thus, ⎧ ⎪ ⎨ −1, t ∈ [0, 1), ∗ (4.46) u (t) = ⎪ ⎩ 0, t ∈ [1, 2]. This gives x∗ (t) =

⎧ ⎪ ⎨ 1 − t, t ∈ [0, 1), ⎪ ⎩ 0,

t ∈ [1, 2].

To obtain λ(t), let us ﬁrst try λ(2− ) = γ = 0. Then, since x∗ (t) enters the boundary zero at t = 1, there are no jumps in the interval (1, 2], and the solution for λ(t) is λ(t) = t − 2, t ∈ (1, 2).

(4.47)

Since λ(t) ≤ 0 and x∗ (t) = 0 on (1, 2], we have u∗ (t) = 0 by (4.40), as stipulated. Now let us see what must happen at t = 1. We know from (4.47) that λ(1+ ) = −1. To obtain λ(1− ), we see that H(1+ ) = −x∗ (1+ ) + λ(1+ )u∗ (1+ ) = 0 and H(1− ) = −x∗ (1− ) + λ(1− )u∗ (1− ) = −λ(1− ). By equating H(1− ) to H(1+ ) as required in (4.29), we obtain λ(1− ) = 0. Using now the jump condition on λ(t) in (4.29), we get the value of the jump ζ(1) = λ(1− ) − λ(1+ ) = 1 ≥ 0. With λ(1− ) = 0, we can solve (4.45) to obtain λ(t) = t − 1, t ∈ [0, 1]. Since λ(t) ≤ 0 and x∗ (t) = 1−t > 0 is positive on [0,1), we can use (4.39) to obtain u∗ (t) = −1 for 0 ≤ t < 1, which is as stipulated in (4.46). In the time interval [0,1) by (4.42), μ2 = 0 since u∗ < 1, and by (4.43), η = 0 because x > 0. Therefore, μ1 (t) = −λ(t) = 1 − t > 0 for 0 ≤ t < 1, and this with u∗ = −1 satisﬁes (4.41). To complete the solution, we calculate the Lagrange multipliers in the interval [1,2]. Since u∗ (t) = 0 on t ∈ [1, 2], we have μ1 (t) = μ2 (t) = 0. Then, from (4.44) we obtain η(t) = −λ(t) = 2 − t ≥ 0 which, with

4.5. The Maximum Principle: Indirect Method

143

x∗ (t) = 0 satisﬁes (4.43). Thus, our guess γ = 0 is correct, and we do not need to examine the possibility of γ > 0. The graphs of x∗ (t) and λ(t) are shown in Fig. 4.2. In Exercise 4.1, you are asked to redo Example 4.3 by guessing that γ > 0 and see that it leads to a contradiction with a condition of the maximum principle.

() ()

1

() 0 2

0

0

0

()

Figure 4.2: State and adjoint trajectories in Example 4.3 It should be obvious that if the terminal time were T = 1.5, the optimal control would be u∗ (t) = −1, t ∈ [0, 1) and u∗ (t) = 0, t ∈ [1, 1.5]. You are asked in Exercise 4.10 to redo the above calculations in this case and show that one now needs to have γ = 1/2. In Exercise 4.3, you are asked to solve a similar problem with F = −u. Remark 4.5 Example 4.3 is a problem instance in which the state constraint is active at the terminal time. In instances where the initial state or the ﬁnal state or both are on the constraint boundary, the maximum principle may degenerate in the sense that there is no nontrivial solution of the necessary conditions, i.e., λ(t) ≡ 0, t ∈ [0, T ], where T is the terminal time. See Arutyunov and Aseev (1997) or Ferreira and Vinter (1994) for conditions that guarantee a nontrivial solution for the multipliers.

144

4. The Maximum Principle: Pure State and Mixed Constraints

Remark 4.6 It can easily be seen that Example 4.3 is a problem instance in which multipliers λ and μ1 would not be unique if the jump condition on the Hamiltonian in (4.29) was not imposed. For references dealing with the issue of non-uniqueness of the multipliers and conditions under which the multipliers are unique, see Kurcyusz and Zowe (1979), Maurer (1977, 1979), Maurer and Wiegand (1992), and Shapiro (1997). Example 4.4 The purpose here is to show that the solution obtained in Example 4.3 satisﬁes the suﬃciency conditions of Theorem 4.1. For this we ﬁrst obtain the direct adjoint variable ⎧ ⎪ ⎨ t − 1, t ∈ [0, 1), d ∗ λ (t) = λ(t) + η(t)hx (x (t), t) = ⎪ ⎩ 0, t ∈ [1, 2). It is easy to see that

H(x, u, λd (t), t) =

⎧ ⎪ ⎨ −x + (t − 1)u, t ∈ [0, 1), ⎪ ⎩ −x,

t ∈ [1, 2],

is linear and hence concave in (x, u) at each t ∈ [0, 2]. Functions ⎛ ⎞ ⎜ u+1 ⎟ g(x, u, t) = ⎝ ⎠ 1−u and h(x) = x are linear and hence quasiconcave in (x, u) and x, respectively. Functions S ≡ 0, a ≡ 0 and b ≡ 0 satisfy the conditions of Theorem 4.1 trivially. Thus, the solution obtained for Example 4.3 satisﬁes all conditions of Theorem 4.1, and is therefore optimal. In Exercise 4.14, you are asked to use Theorem 4.1 to verify that the given solution there is optimal. Example 4.5 Consider Example 4.3 with T = 3 and the terminal state constraint x(3) = 1.

4.5. The Maximum Principle: Indirect Method

145

Solution Clearly, the optimal control u∗ will be the one that keeps x as small as possible, subject to the state constraint (4.38) and the boundary condition x(0) = x(3) = 1. Thus, ⎧ ⎧ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ −1, t ∈ [0, 1), 1 − t, t ∈ [0, 1), ⎪ ⎪ ⎪ ⎪ ⎨ ⎨ ∗ u∗ (t) = 0, t ∈ [1, 2], x (t) = ⎪ 0, t ∈ [1, 2], ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ ⎩ 1, t ∈ (2, 3], t − 2, t ∈ (2, 3]. For brevity, we will not provide the same level of detailed explanation as we did in Example 4.3. Rather, we will only compute the adjoint function and the multipliers that satisfy the optimality conditions. These are ⎧ ⎪ ⎨ t − 1, t ∈ [0, 1], λ(t) = (4.48) ⎪ ⎩ t − 2, t ∈ (1, 3), μ1 (t) = μ2 (t) = 0, η(t) = −λ(t), t ∈ [1, 2],

(4.49)

γ = 0, β = λ(2− ) = 1,

(4.50)

and the jump ζ(1) = 1 ≥ 0 so that λ(1− ) = λ(1+ ) + ζ(1) and H(1− ) = H(1+ ).

(4.51)

Example 4.6 Introduce a discount rate ρ > 0 in Example 4.1 so that the objective function becomes 3 −ρt −e udt (4.52) max J = 0

and re-solve using the indirect maximum principle (4.29). Solution It is obvious that the optimal solution will remain the same as (4.5), shown also in Fig. 4.1. With u∗ and x∗ as in (4.5), we must obtain λ, μ1 , μ2 , η, γ, and ζ so that the necessary optimality conditions (4.29) hold, i.e., H = −e−ρt u + λu,

(4.53)

L = H + μ1 u + μ2 (3 − u) + η[u + 2(t − 2)],

(4.54)

146

4. The Maximum Principle: Pure State and Mixed Constraints

Lu = −e−ρt + λ + μ1 − μ2 + η = 0,

(4.55)

λ˙ = −Lx = 0, λ(3− ) = 0,

(4.56)

γ[x∗ (3) − 1 + (1 − 2)2 ] = 0,

(4.57)

μ1 ≥ 0, μ1 u = 0, μ2 ≥ 0, μ2 (3 − u) = 0,

(4.58)

η ≥ 0, η[x∗ (t) − 1 + (t − 2)2 ] = 0,

(4.59)

and if λ is discontinuous at the entry time τ = 1, then λ(1− ) = λ(1+ ) + ζ(1), ζ(1) ≥ 0,

(4.60)

− e−ρ u∗ (1− ) + λ(1− )u∗ (1− ) = −e−ρ u∗ (1+ ) + λ(1+ ) − ζ(1)(−2). (4.61) From (4.60), we obtain λ(1− ) = e−ρ . This with (4.56) gives

λ(t) =

⎧ ⎪ ⎨ e−ρ , 0 ≤ t < 1, ⎪ ⎩ 0,

1 ≤ t ≤ 3,

as shown in Fig. 4.3, ⎧ ⎪ ⎪ ⎪ e−ρt − e−ρ , 0 ≤ t < 1, ⎪ ⎪ ⎨ μ1 (t) = 0, 1 ≤ t ≤ 2, ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ e−ρt , 2 < t ≤ 3, and

μ2 (t) = 0, 0 ≤ t ≤ 3,

⎧ ⎪ ⎪ ⎪ 0, 0 ≤ t < 1, ⎪ ⎪ ⎨ η(t) = e−ρt , 1 ≤ t ≤ 2, ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 0, 2 < t ≤ 3,

which, along with u∗ and x∗ , satisfy (4.29). Note, furthermore, that λ is continuous at the exit time t = 2. At the entry time τ 1 = 1, ζ(1) = e−ρ ≥ η(1+ ) = e−ρ , so that (4.35) also holds. Finally, γ = η(3− ) = 0.

4.6. Current-Value Maximum Principle: Indirect Method

147

Figure 4.3: Adjoint trajectory for Example 4.4

4.6

Current-Value Maximum Principle: Indirect Method

Just as the necessary condition (3.42) represents the current-value formulation corresponding to (3.12), we can, when ﬁrst-order pure state constraints are present, also state the current-value formulation of the necessary conditions (4.29). As in Sect. 3.3, with F (x, u, t) = φ(x, u)e−ρt , S(x, T ) = ψ(x)e−ρT , and ρ > 0, the objective function in the problem (4.11) is replaced by

max J =

T

φ(x, u)e

−ρt

−ρT

dt + ψ[x(T )]e

.

0

With the Hamiltonian H as deﬁned in (3.35), we can write the Lagrangian as L[x, u, λ, μ, η] := H + μg + ηh1 = φ + λf + μg + ηh1 .

We can now state the current-value form of the maximum principle, giving the necessary conditions for u∗ (with the state trajectory x∗ ) to be optimal. These conditions are that there exist an adjoint function λ, which may be discontinuous at each entry or contact time, multiplier functions μ, α, β, γ, η, and a jump parameter ζ(τ ) at each τ where λd is discontinuous, such that the following (4.62) holds:

148

4. The Maximum Principle: Pure State and Mixed Constraints x˙ ∗ = f (x∗ , u∗ , t), x∗ (0) = x0 , satisfying constraints g(x∗ , u∗ , t) ≥ 0, h(x∗ (t), t) ≥ 0,and the terminal constraints a(x∗ (T ), T ) ≥ 0 and b(x∗ (T ), T ) = 0; λ˙ = ρλ − Lx [x∗ , u∗ , λ, μ, η, t] with the transversality conditions λ(T − ) = ψ x (x∗ (T ), T ) + αax (x∗ (T ), T ) + βbx (x∗ (T ), T ) +γhx (x∗ (T ), T ), and α ≥ 0, αa(x∗ (T ), T ) = 0, γ ≥ 0, γh(x∗ (T ), T ) = 0;

the Hamiltonian maximizing condition H[x∗ (t), u∗ (t), λ(t), t] ≥ H[x∗ (t), u, λ(t), t] at each t ∈ [0, T ] for all u satisfying g[x∗ (t), u, t] ≥ 0, and h1i (x∗ (t), u, t)

(4.62) ∗

≥ 0 whenever hi (x (t), t) = 0, i = 1, 2, · · · , p;

the jump conditions at any entry/contact time τ , where λ is discontinuous, are λ(τ − ) = λ(τ + ) + ζ(τ )hx (x∗ (τ ), τ ) and H[x∗ (τ ), u∗ (τ − ), λ(τ − ), τ ] = H[x∗ (τ ), u∗ (τ + ), λ(τ + ), τ ] −ζ(τ )ht (x∗ (τ ), τ ); the Lagrange multipliers μ(t) are such that ∂L/∂u|u=u∗ (t) = 0, dH/dt = dL/dt = ∂L/∂t + ρλf, and the complementary slackness conditions μ(t) ≥ 0, μ(t)g(x∗ , u∗ , t) = 0, η(t) ≥ 0, η(t)h(x∗ (t), t) = 0, and ζ(τ ) ≥ 0, ζ(τ )h(x∗ (τ ), τ ) = 0 hold.

Exercises for Chapter 4

149

If T ∈ [T1 , T2 ], 0 ≤ T1 < T2 < ∞, is also a decision variable, then if T ∗ is the optimal terminal time, then the optimal solution x∗ , u∗ , T ∗ must satisfy (4.62) with T replaced by T ∗ and the condition H[x∗ (T ∗ ), u∗ (T ∗− ), λd (T ∗− ), T ∗ ] − ρψ[x∗ (T ∗ ), T ∗ ] + αaT [x∗ (T ∗ ), T ∗ ] ⎧ ⎪ ⎪ ⎪ ≤ 0 if T ∗ = T1 , ⎪ ⎪ ⎨ +βbT [x∗ (T ∗ ), T ∗ ] + γ d hT [x∗ (T ∗ ), T ∗ ] = 0 if T ∗ ∈ (T1 , T2 ),(4.63) ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ ≥ 0 if T ∗ = T2 . Derivation of (4.63) starting from (4.15) is similar to that of (3.44) from (3.15). Remark 4.7 The current-value version of (4.34) in Remark 4.3 is η(t) ˙ ≤ ρη(t) and (4.35). The inﬁnite horizon problem with pure and mixed constraints can be stated as (3.97) with an additional constraint (4.7). As in Sect. 3.6, the conditions in (4.62) except the transversality condition on the adjoint variable are still necessary for optimality. As for the suﬃciency conditions, an analogue of Theorem 4.1 holds, subject to the discussion on inﬁnite horizon transversality conditions in Sect. 3.6. We conclude this chapter with the following cautionary remark. Remark 4.8 While various subsets of conditions speciﬁed in the maximum principles (4.13), (4.29), or (4.62) have been proved in the literature, proofs of the entire results are still not available. For this reason, Hartl (1995) call (4.13), (4.29), or (4.62) as informal theorems. Seierstad and Sydsæter (1987) call them almost necessary conditions since, very rarely, problems arise where the optimal solution requires more complicated multipliers and adjoint variables than those speciﬁed in this chapter. Exercises for Chapter 4 E 4.1 Rework Example 4.3 by guessing that γ > 0, and show that it leads to a contradiction with a condition of the maximum principle. E 4.2 Rework Example 4.3 with terminal time T = 1/2.

150

4. The Maximum Principle: Pure State and Mixed Constraints

E 4.3 Change the objective function of Example 4.3 as follows: 2 (−u)dt . max J = 0

Re-solve and show that the solution is not unique. E 4.4 Specialize the maximum principle (4.29) for the nonnegativity state constraint of the form x(t) ≥ 0 for all t satisfying 0 ≤ t ≤ T, in place of h(x, t) ≥ 0 in (4.7). E 4.5 Consider the problem: max J =

T

(−x)dt

0

subject to x˙ = −u − 1, x(0) = 1, x(t) ≥ 0, 0 ≤ u(t) ≤ 1. Show that (a) If T = 1, there is exactly one feasible and optimal solution. (b) If T > 1, then there is no feasible solution. (c) If 0 < T < 1, then there is a unique optimal solution. (d) If the control constraint is 0 ≤ u(t) ≤ K, there is a unique optimal solution for every K ≥ 1 and T = 1/2. (e) The value of the objective in (d) increases as K increases. (f) If the control constraint in (d) is u(t) ≥ 0, then the optimal control is an impulse control deﬁned by the limit of the solution in (e).

E 4.6 Impose the constraint x ≥ 0 on Exercise 3.16(b) to obtain the problem: 4 max J = (−x)dt 0

subject to

Exercises for Chapter 4

151

x˙ = u, x(0) = 1, x(4) = 1, u + 1 ≥ 0, 1 − u ≥ 0, x ≥ 0. Find the optimal trajectories of the control variable, the state variable, and other multipliers. Also, graph these trajectories. E 4.7 Transform the problem (4.11) with the pure constraint of type (4.7) to a problem with the nonnegativity constraint of type (4.9). Hint: Deﬁne y = h(x, t) as an additional state variable. that we have assumed (4.7) to be a ﬁrst-order constraint.

Recall

E 4.8 Consider a two-reservoir system such as that shown in Fig. 4.4, where xi (t) is the volume of water in reservoir i and ui (t) is the rate of discharge from reservoir i at time t. Thus, x˙ 1 (t) = −u1 (t),

x1 (0) = 4,

x˙ 2 (t) = u1 (t) − u2 (t), x2 (0) = 4.

Figure 4.4: Two-reservoir system of Exercise 4.8 Solve the problem of maximizing 10 [(10 − t)u1 (t) + tu2 (t)]dt J= 0

subject to the above state equations and the constraints 0 ≤ ui (t) ≤ 1, xi (t) ≥ 0 for all t ∈ [0, 10].

152

4. The Maximum Principle: Pure State and Mixed Constraints

Also compute the optimal value of the objective function. Hint: Guess the optimal solution and verify it by using the Lagrangian form of the maximum principle. E 4.9 An Inventory Control Problem. Solve T P2 − hI + max dt P 2 0 subject to

S2 , I˙ = P − S, I(0) = I0 > 2h and the control and the pure state inequality constraints P ≥ 0 and I ≥ 0, respectively. Assume that S > 0 and h > 0 are constants and T is suﬃciently large. Note that I represents inventory, P represents production rate, and S represents demand. The constraints on P and I mean that production must be nonnegative and backlogs are not allowed, respectively. Hint: By T being suﬃciently large, we mean T > I0 /S + S/(2h).

E 4.10 Redo Example 4.3 with T = 1.5. E 4.11 Redo Example 4.6 using the current-value maximum principle (4.62) in Sect. 4.6. E 4.12 For this exercise only, assume that h(x, t) ≥ 0 in (4.7) is a second-order constraint, i.e., r = 2. Transform the problem to one with nonnegativity constraints. Use the result in Exercise 4.4 to derive a maximum principle for problems with second-order constraints. Hint: As in Exercise 4.7, deﬁne y = h. In addition, deﬁne yet another state variable z = y˙ = dh/dt. Note further that this procedure can be generalized to handle problems with rth-order constraints for any positive integer r. E 4.13 Re-solve Example 4.6 when ρ < 0.

Exercises for Chapter 4

153

E 4.14 Consider the following problem: 5 udt min J = 0

subject to the state equation x˙ = u − x, x(0) = 1, and the control and state constraints 0 ≤ u ≤ 1, x(t) ≥ 0.7 − 0.2t. Use the suﬃciency conditions in Theorem 4.1 to verify that the optimal control for the problem is ⎧ ⎪ ⎪ ⎪ 0, 0 ≤ t ≤ θ, ⎪ ⎪ ⎨ u∗ (t) = 0.5 − 0.2t, θ < t ≤ 2.5, ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 0, 2.5 < t ≤ 5, where θ ≈ 0.51626. Sketch the optimal state trajectory x∗ (t) for the problem. √ E 4.15 In Example 4.6, let t± (x) = 2 ± 1 − x. Show that the value function ⎧ √ √ −2ρ −ρ(2− 1−x) ⎪ ⎨ − 2e +2(ρ 1−x−1)e , for x < 1, 0 ≤ t ≤ t− (x), ρ2 V (x, t) = ⎪ ⎩ 0, for x ≥ 1 or t+ (x) ≤ t ≤ 3. Note that V (x, t) is not deﬁned for x < 1, t− (x) < t ≤ 3. Show furthermore that for the given initial condition x(0) = 0, the marginal valuation is ⎧ ⎪ ⎪ ⎪ e−ρ , for t ∈ [0, 1), ⎪ ⎪ ⎨ Vx (x∗ (t), t) = λd (t) = λ(t) + η(t) = e−ρt , for t ∈ [1, 2], ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 0, for t ∈ (2, 3]. In this case, it is interesting to note that the marginal valuation is discontinuous at the constraint exit time t = 2.

154

4. The Maximum Principle: Pure State and Mixed Constraints

E 4.16 Show in Example 4.3 that the value function ⎧ ⎪ ⎨ −x2 /2, for x ≤ 2 − t, 0 ≤ t ≤ 2, V (x, t) = ⎪ ⎩ −2x + 2 − 2t + xt + t2 /2, for x > 2 − t, 0 ≤ t ≤ 2. Then verify that for the given initial condition x(0) = 1, ⎧ ⎪ ⎨ t − 1, for t ∈ [0, 1), ∗ d Vx (x (t), t) = λ (t) = λ(t) + η(t) = ⎪ ⎩ 0, for t ∈ [1, 2]. E 4.17 Rework Example 4.5 by using the direct maximum principle (4.13). E 4.18 Solve the linear inventory control problem of minimizing T (cP (T ) + hI(t))dt 0

subject to ˙ = P (t) − S, I(t)

I(0) = 1,

P (t) ≥ 0 and I(t) ≥ 0,

t ∈ [0, T ],

where P (t) denotes the production rate and I(t) is the inventory level at time t and√where c, h and S are positive constants and the given terminal time T > 2S. E 4.19 A machine with quality x(t) ≥ 0 produces goods with ax(t) dollars per unit time at time t. The quality deteriorates at the rate δ, but the decay can be slowed by a preventive maintenance u(t) as follows: x˙ = u − δx, x(0) = x0 > 0. Obtain the optimal maintenance rate u(t), 0 ≤ t ≤ T, so as to maximize T (ax − u)dt 0

subject to u ∈ [0, u ¯] and x ≤ x ¯, where u ¯ > δx ¯, a > δ, and x ¯ > x0 . Hint: Solve ﬁrst the problem without the state constraint x ≤ x ¯. You will need to treat two cases: δT ≤ ln a − ln (a − δ) and δT > ln a − ln (a − δ).

Exercises for Chapter 4

155

E 4.20 Maximize

3

J= 0

(u − x)dt

subject to x˙ = 1 − u, x(0) = 2, 0 ≤ u ≤ 3, x + u ≤ 4, x ≥ 0. E 4.21 Maximize

2

J= 0

(1 − x)dt

subject to x˙ = u, x(0) = 1, −1 ≤ u ≤ 1, x ≥ 0. E 4.22 Maximize

3

J= 0

(4 − t)udt

subject to x˙ = u, x(0) = 0, x(3) = 3, 0 ≤ u ≤ 2, 1 + t − x ≥ 0. E 4.23 Maximize

J =−

4

0

e−t (u − 1)2 dt

subject to x˙ = u,

x(0) = 0,

x ≤ 2 + e−3 .

156

4. The Maximum Principle: Pure State and Mixed Constraints

E 4.24 Solve the following problem: 2 (2u − x)dt max J = 0

x˙ = −u, −3 ≤ u ≤ 3,

x(0) = e,

x − u ≥ 0,

x ≥ t.

E 4.25 Solve the following problem: 3 −2x1 dt max J = 0

x˙ 1 = x2 ,

x1 (0) = 2,

x˙ 2 = u,

x2 (0) = 0,

x1 ≥ 0. E 4.26 Re-solve Example 4.6 with the control constraint (4.3) replaced by 0 ≤ u ≤ 1. E 4.27 Solve explicitly the following problem: 2 x(t)dt max J = − 0

subject to x(t) ˙ = u(t), x(0) = 1, −a ≤ u(t) ≤ b, a > 1/2, b > 0, x(t) ≥ t − 2. Obtain

x∗ (t),

u∗ (t)

E 4.28 Minimize

and all the required multipliers. 0

T

1 2 (x + c2 u2 )dt 2

subject to x˙ = u, x(0) = x0 > 0, x(T ) = 0, h1 (x, t) = x − a1 + b1 t ≥ 0, h2 (x, t) = a2 − b2 t − x ≥ 0, where ai , bi > 0, a2 > x0 > a1 , and a2 /b2 > a1 /b1 ; see Fig. 4.5. The optional path must begin at x0 on the x-axis, stay in the shaded area, and end on the t-axis.

Exercises for Chapter 4

157

Figure 4.5: Feasible space for Exercise 4.28 (a) First, assume that the problem parameters are such that the optimal solution x∗ (t) satisﬁes h1 (x∗ (t), t) > 0 for t ∈ [0, T ]. Show that x∗ (t) = k1 et/c + k2 e−t/c , where k1 and k2 are the constants to be determined. Write down the two conditions that would determine the constants. Also, illustrate graphically the optimal state trajectory. (b) How would your solution change if the problem parameters do not satisfy the condition in (a)? Characterize and graphically illustrate the optimal state trajectory. E 4.29 With a > 0, b > 0, and γ(t)/γ(t) ˙ = −ρ(t) < 0,

T

max J = u,T

0

a (1 − e−bu(t) )γ(t)dt b

subject to x˙ = −u, x(0) = x0 > 0 given, and the constraint x(t) ≥ 0. Obtain the expressions satisﬁed by the optimal terminal time T ∗ , the optimal control u∗ (t), 0 ≤ t ≤ T ∗ , and the optimal state trajectory x∗ (t), 0 ≤ t ≤ T ∗ . Furthermore, obtain them explicitly in the special case when ρ(t) = ρ > 0, a constant positive discount rate.

158

4. The Maximum Principle: Pure State and Mixed Constraints

E 4.30 Set ρ = 0 in the solution of Example 4.6 and obtain λ, γ, η, ζ(1) for the undiscounted problem. Then use the transformation formulas (4.30)–(4.33) on these and the fact that ζ(2) = 0 to obtain λd , γ d , η d , and ζ d (1) and ζ d (2), and show that they are the same as those obtained in Example 4.2 along with ζ d (1) = 0, which holds trivially. E 4.31 Consider a ﬁnite-time economy in which production can be used for consumption as well as investment, but production also pollutes. The state equations for the capital stock K and stock of pollution W are K˙ = suK, K(0) = K0 , ˙ = uK − δW, W (0) = W0 , W where a fraction s of the production output uK is invested, with u denoting the capacity utilization rate. The control constraints are 0 ≤ s ≤ 1, 0 ≤ u ≤ 1, and the state constraint ¯ W ≤W ¯. implies that the pollution stock cannot exceed the upper bound W The aim of the economy is to choose s and u so as to maximize the consumption utility T

0

(1 − s)uKdt.

¯ , T > 1 and W0 − K0 /δ)e−δT + K0 /δ < W ¯ , which Assume that W0 < W ¯ even means that even with s(t) ≡ 0, the pollution stock never reaches W with u(t) ≡ 1.

Chapter 5

Applications to Finance An important area of ﬁnance involves making decisions regarding investment and dividend policies over time and ways to ﬁnance them. Among the ways of ﬁnancing such policies are: issuing equity, retaining earnings, borrowing money, etc. It is possible to model such situations as optimal control problems; see, for example, Davis and Elzinga (1971), Elton and Gruber (1975), and Sethi (1978b). Some of these models are simple to analyze and they yield useful insights. In this chapter we deal with two diﬀerent problems relating to a ﬁrm. The cash balance problem, in its simplest form, is a problem of controlling the level of a ﬁrm’s cash balances to meet its demand for cash at minimum total cost. The problem of the optimal equity ﬁnancing of a corporate ﬁrm, a central problem in ﬁnance, is that of determining the optimal dividend path along with new equity issued over time in order to maximize the value of the ﬁrm. Although we only deal with deterministic problems in this chapter, some of the more important problems in ﬁnance involve uncertainty. Thus, their optimization requires the use of stochastic optimal control theory or stochastic programming. A brief introduction to stochastic optimal control theory will be provided in Chap. 12, together with an application to a stochastic consumption-investment problem and references. In the next section, we introduce a simple cash balance problem as a tutorial. This model is based on Sethi and Thompson (1970) and Sethi (1973d, 1978c). We will be especially interested in the ﬁnancial © Springer Nature Switzerland AG 2019 S. P. Sethi, Optimal Control Theory, https://doi.org/10.1007/978-3-319-98237-3 5

159

160

5. Applications to Finance

interpretations for the various functions such as the Hamiltonian and the adjoint functions that arise in the course of the analysis.

5.1

The Simple Cash Balance Problem

Consider a ﬁrm which has a known demand for cash over time. To satisfy this cash demand, the ﬁrm must keep some cash on hand, assumed to be held in a checking account at a bank. If the ﬁrm keeps too much cash, it loses money in terms of opportunity cost, in that it can earn higher returns by buying securities such as bonds. On the other hand, if the cash balance is too small, the ﬁrm has to sell securities to meet the cash demand and thus incur a broker’s commission. The problem then is to ﬁnd the tradeoﬀ between the cash and security balances.

5.1.1

The Model

To formulate the optimal control problem we introduce the following notation: T = the time horizon, x(t) = the cash balance in dollars at time t, y(t) = the security balance in dollars at time t, d(t) = the instantaneous rate of demand for cash; d(t) can be positive or negative, u(t) = the rate of sale of securities in dollars; a negative sales rate means a rate of purchase, r1 (t) = the interest rate earned on the cash balance, r2 (t) = the interest rate earned on the security balance, α = the broker’s commission in dollars per dollar’s worth of securities bought or sold; 0 < α < 1. The state equations are x˙ = r1 x − d + u − α|u|, x(0) = x0 ,

(5.1)

y˙ = r2 y − u, y(0) = y0 ,

(5.2)

and the control constraints are − U2 ≤ u(t) ≤ U1 ,

(5.3)

where U1 and U2 are nonnegative constants. The objective function is: max{J = x(T ) + y(T )}

(5.4)

subject to (5.1)–(5.3). Note that the problem is in the linear Mayer form.

5.1. The Simple Cash Balance Problem

5.1.2

161

Solution by the Maximum Principle

Introduce the adjoint variables λ1 and λ2 and deﬁne the Hamiltonian function H = λ1 (r1 x − d + u − α|u|) + λ2 (r2 y − u). (5.5) The adjoint variables satisfy the diﬀerential equations ∂H = −λ1 r1 , λ1 (T ) = 1, λ˙1 = − ∂x

(5.6)

∂H = −λ2 r2 , λ2 (T ) = 1. λ˙2 = − ∂y

(5.7)

It is easy to solve these, respectively, as T

λ1 (t) = e and λ2 (t) = e

t

T t

r1 (τ )dτ

r2 (τ )dτ

.

(5.8)

(5.9)

The interpretations of these solutions are also clear. Namely, λ1 (t) is the future value (at time T ) of one dollar held in the cash account from time t to T and, likewise, λ2 (t) is the future value of one dollar invested in securities from time t to T. Thus, the adjoint variables have natural interpretations as the actuarial evaluations of competitive investments at each point of time. Let us now derive the optimal policy by choosing the control variable u to maximize the Hamiltonian in (5.5). In order to deal with the absolute value function we write the control variable u as the diﬀerence of two nonnegative variables, i.e., u = u1 − u2 , u1 ≥ 0, u2 ≥ 0.

(5.10)

Recall that this method was suggested in Remark 3.12 in Sect. 3.7. In order to make u = u1 when u1 is strictly positive, and u = −u2 when u2 is strictly positive, we also impose the quadratic constraint u1 u2 = 0,

(5.11)

so that at most one of u1 and u2 can be nonzero. However, the optimal properties of the solution will automatically cause this constraint to be satisﬁed. The reason is that the broker’s commission must be paid on

162

5. Applications to Finance

every transaction, which makes it not optimal to simultaneously buy and sell securities. Given (5.10) and (5.11) we can write |u| = u1 + u2 .

(5.12)

Also, since u ∈ [−U1 , U2 ] from (5.3), we must have u1 ≤ U1 and u2 ≤ U2 . In view of (5.10), the control constraints on the variables u1 and u2 are 0 ≤ u1 ≤ U1 and 0 ≤ u2 ≤ U2 .

(5.13)

We can now substitute (5.10) and (5.12) into the Hamiltonian (5.5) and reproduce the part that depends on control variables u1 and u2 , and denote it by W. Thus, W = u1 [(1 − α)λ1 − λ2 ] − u2 [(1 + α)λ1 − λ2 ].

(5.14)

Maximizing the Hamiltonian (5.5) with respect to u ∈ [−U1 , U2 ] is the same as maximizing W with respect to u1 ∈ [0, U1 ] and u2 ∈ [0, U2 ]. But W is linear in u1 and u2 so that the optimal strategy is bang-bang and is as follows: u∗ = u∗1 − u∗2 , (5.15) where

u∗1 = bang[0, U1 ; (1 − α)λ1 − λ2 ],

(5.16)

u∗2 = bang[0, U2 ; −(1 + α)λ1 + λ2 ].

(5.17)

Since u1 (t) represents the rate of sale of securities, (5.16) says that the optimal policy is: sell at the maximum allowable rate if the future value of a dollar less the broker’s commission (i.e., the future value of (1 − α) dollars) is greater than the future value of a dollar’s worth of securities; and do not sell if these future values are in reverse order. In case the future value of a dollar less the commission is exactly equal to the future value of a dollar’s worth of securities, then the optimal policy is undetermined. In fact, we are indiﬀerent as to the action taken, and this is called singular control. Similarly, u2 (t) represents the purchase of securities. Here we buy, do not buy, or are indiﬀerent, if the future value of a dollar plus the commission is less than, greater than, or equal to the future value of a dollar’s worth of securities, respectively. Note that if (1 − α)λ1 (t) ≥ λ2 (t), then (1 + α)λ1 (t) > λ2 (t),

5.1. The Simple Cash Balance Problem

163

Figure 5.1: Optimal policy shown in (λ1 , λ2 ) space so that if u1 (t) > 0, then u2 (t) = 0. Similarly, if (1 + α)λ1 (t) ≤ λ2 (t), then (1 − α)λ1 (t) < λ2 (t), so that if u2 (t) > 0, then u1 (t) = 0. Hence, with the optimal policy, the relation (5.11) is always satisﬁed. Figure 5.1 illustrates the optimal policy at time t. The ﬁrst quadrant is divided into three areas which represent diﬀerent actions (including no action) to be taken. The dotted lines represent the singular control manifolds. A possible path of the vector (λ1 (t), λ2 (t)) of the adjoint variables is shown in Fig. 5.1 also. Note that on this path, there is one period of selling, two periods of buying, and three periods of inactivity. Note also that the ﬁnal point on the path is (1, 1), since the terminal values λ1 (T ) = λ2 (T ) = 1, and therefore, the last interval is always characterized by inactivity. Another way to represent the optimal path is in the (t, λ2 /λ1 ) space. The path of (λ1 (t), λ2 (t)) shown in Fig. 5.1 corresponds to the path of λ2 (t)/λ1 (t) over time shown in Fig. 5.2.

164

5. Applications to Finance

Figure 5.2: Optimal policy shown in (t, λ2 /λ1 ) space Perhaps a more realistic version of the cash balance problem is to disallow overdraft on the bank account. This means imposing the pure state constraint x(t) ≥ 0. In addition, if short selling of securities is not permitted, then we must also have y(t) ≥ 0. These extensions give rise to pure state constraints treated in Chap. 4. In Exercise 5.2 you are asked to formulate such an extension and write the indirect maximum principle (4.29) for it. Exercises 5.3 and 5.4 present instances where it is easy to guess the optimal solutions. In Exercise 5.5, you are asked to show if the guessed solution in Exercise 5.4 satisﬁes the maximum principle (4.29). It is in Chap. 6 that we discuss in detail an application of the indirect maximum principle (4.29) for solving a problem called the wheat trading model.

5.2

Optimal Financing Model

In the present section, we discuss a model of a corporate ﬁrm which must ﬁnance its investments by an optimal combination of retained earnings and external equity. The model to be discussed is due to Krouse and Lee (1973), with corrections and extensions due to Sethi (1978b). The problem of the optimal ﬁnancing of the ﬁrm can be formulated as an optimal control problem. The formulations, such as those of Davis (1970), Krouse (1972), and Krouse and Lee (1973), permit the ﬁrm to ﬁnance its investments by retained earnings, debt, and/or external equity in various proportions which may vary over time. Note that earnings not retained are paid out as dividends to the ﬁrm’s stockholders.

5.2. Optimal Financing Model

165

For reasons of simplicity and ease of its solution, the model analyzed here does not permit debt as a source of ﬁnancing, but does permit retained earnings and external equity to be used in any proportions.

5.2.1

The Model

In order to formulate the model, we use the following notation: y(t) = the value of the ﬁrm’s assets or invested capital at time t, x(t) = the current earnings rate in dollars per unit time at time t, u(t) = the external or new equity ﬁnancing expressed as a multiple of current earnings; u ≥ 0, v(t) = the fraction of current earnings retained, i.e., 1 − v(t) represents the rate of dividend payout; 0 ≤ v(t) ≤ 1, 1 − c = the proportional ﬂoatation (i.e., transaction) cost for external equity; c a constant, 0 ≤ c < 1, ρ = the continuous discount rate (assumed constant); known commonly as the stockholder’s required rate of return, or the cost of capital, r = the actual rate of return (assumed constant) on the ﬁrm’s invested capital; r > ρ, g = the upper bound on the growth rate of the ﬁrm’s assets, T = the planning horizon; T < ∞ (T = ∞ in Sect. 5.2.4) . Given these deﬁnitions, the current earnings rate is x = ry. The rate of change in the current earnings rate is given by x˙ = ry˙ = r(cu + v)x, x(0) = x0 .

(5.18)

Furthermore, the upper bound on the rate of growth of the assets implies the following constraint on the control variables: y/y ˙ = (cu + v)x/(x/r) = r(cu + v) ≤ g.

(5.19)

Finally, the objective of the ﬁrm is to maximize its value, which is taken to be the present value of the future dividend stream accruing to the shares outstanding at time zero. To derive this expression, note that T (1 − v)xe−ρt dt 0

166

5. Applications to Finance

represents the present value of total dividends issued by the ﬁrm. A portion of these dividends go to the new equity, which under the assumption of an eﬃcient market will get a rate of return exactly equal to the discount rate ρ. This should therefore be equal to the present value

T

uxe−ρt dt

0

of the external equity raised over time. Thus, the net present value of the total future dividends that accrue to the initial shares is the diﬀerence of the previous two expressions, i.e., J= 0

T

e−ρt (1 − v − u)xdt;

(5.20)

see Miller and Modigliani (1961) and Sethi (1996) for further discussion. Note that in the case of a ﬁnite horizon, a more realistic objective function would include a salvage value or bequest term S[x(T )]. This is not very diﬃcult to incorporate. See Exercise 5.12 where the bequest function is linear. We will also solve the inﬁnite horizon problem (i.e., T = ∞) after we have solved the ﬁnite horizon problem. Remark 5.1 An intuitive interpretation of (5.20) is that the value J of the ﬁrm is the present value of the cash ﬂows (dividends) going out from the ﬁrm to the society less the present value of the cash ﬂows (new equity) coming from the society into the ﬁrm. The optimal control problem is to choose u and v over time so as to maximize J in (5.20) subject to (5.18), the constraints (5.19), u ≥ 0, and 0 ≤ v ≤ 1. For convenience, we restate this problem as ⎧ T ⎪ −ρt ⎪ ⎪ e (1 − v − u)xdt J = max ⎪ ⎪ u,v ⎪ 0 ⎪ ⎪ ⎪ ⎪ ⎪ subject to ⎪ ⎪ ⎨ (5.21) x˙ = r(cu + v)x, x(0) = x0 , ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ and the control constraints ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ cu + v ≤ g/r, u ≥ 0, 0 ≤ v ≤ 1.

5.2. Optimal Financing Model

5.2.2

167

Application of the Maximum Principle

This is a bilinear problem with two control variables which is a special case of Row (f) in Table 3.3. The current-value Hamiltonian is H = (1 − v − u)x + λr(cu + v)x = [(crλ − 1)u + (rλ − 1)v + 1]x,

(5.22)

where the current-value adjoint variable λ satisﬁes λ˙ = ρλ − (1 − v − u) − λr(cu + v)

(5.23)

with the transversality condition λ(T ) = 0.

(5.24)

The ﬁrst term in the Hamiltonian in (5.22) is the dividend payout rate to stockholders of record at time t. According to Sect. 2.2.1, λ is the marginal value (in time t dollars) of a unit change in the earnings rate at time t. Thus, λr(cu + v)x is the value at time t of the incremental earnings rate due to the investment of retained earnings vx and the net amount of external ﬁnancing cux. This explains why we should maximize H with respect to u and v at each t. To interpret (5.23) as in Sect. 2.2.4, consider an earnings rate of one dollar at time t. It is worth λ, on which the stockholders expect a return of ρλdt at time dt. In equilibrium this must be equal to the “capital gain” dλ, plus the immediate dividend (1 − v)dt less udt, the “claims” of the new stockholders, plus the value λr(cu + v)dt of the incremental earnings rate r(cu + v)dt at time t + dt. To specify the form of optimal policy, we rewrite the Hamiltonian as H = [W1 u + W2 v + 1]x,

(5.25)

W1 = crλ − 1,

(5.26)

W2 = rλ − 1.

(5.27)

where

Note ﬁrst that the state variable x factors out so that the optimal controls are independent of the state variable. Second, since the Hamiltonian is linear in the two control variables, the optimal policy is a combination of generalized bang-bang and singular controls. Of course, the characterization of these optimal controls in terms of the adjoint variable λ will require solving a parametric linear programming problem at each

168

5. Applications to Finance Table 5.1: Characterization of optimal controls with c < 1 Conditions on

Case A:

Case B:

W1 , W2

g≤r

g>r

Subcases

Subcases

Optimal controls

Characterization

Generalized

(1)

W2 < 0

A1

B1

u∗ = 0, v ∗ = 0

(2)

W2 = 0

A2

B2

u∗ = 0,

bang-bang

0 ≤ v ∗ ≤ min[1, g/r] u

∗

= 0, v

∗

(3)

W2 > 0

A3

–

= g/r

(4)

W1 < 0, W2 > 0

–

B3

u∗ = 0, v ∗ = 1

(5)

W1 = 0

–

B4

0 ≤ u∗ ≤ (g − r)/rc,

Singular Generalized bang-bang Generalized bang-bang

v (6)

W1 > 0

–

B5

∗

Singular

=1

u∗ = (g − r)/rc, v ∗ = 1

Generalized bang-bang

instant of time t. The Hamiltonian maximization problem can be stated as follows: ⎧ ⎪ ⎪ max {W1 u + W2 v} ⎪ ⎪ ⎪ ⎨ u,v (5.28) subject to ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ u ≥ 0, 0 ≤ v ≤ 1, cu + v ≤ g/r. Obviously, the constraint v ≤ 1 becomes redundant if g/r < 1. Therefore, we have two cases: Case A: g ≤ r and Case B: g > r, under each of which, we can solve the linear programming problem (5.28) graphically in a closed form. This is done in Figs. 5.3 and 5.4. There are seven subcases shown in Fig. 5.3 and nine subcases on Fig. 5.4, but some of these subcases cannot occur. To see this, we note from our assumption c < 1 that W1 = crλ − 1 < crλ − c = cW2 , which also gives W2 > 0 if W1 = 0. Thus, subcases A4–A7 and B6–B9 are ruled out. The remaining Subcases A1–A3 and B1–B5 are shown

5.2. Optimal Financing Model

169

adjacent to the darkened lines in Figs. 5.3 and 5.4, respectively. In addition to W1 < cW2 and W1 = 0 implying W2 > 0, we see that W2 ≤ 0 implies W1 < 0. In view of these, we can simply characterize Subcases A1 and B1 by W2 < 0, A2 and B2 by W2 = 0, A3 by W2 > 0, B4 by W1 = 0, and B5 by W1 > 0, and use these simpler characterizations in our subsequent discussion. Keep in mind that Subcase B3 remains characterized as W1 < 0, W2 > 0. In Table 5.1, we list the feasible cases, shown along the darkened lines in Figs. 5.3 and 5.4 and provide the form of the optimal control in each of these cases. The catalog of possible optimal control regimes shown in Table 5.1 gives the potential time-paths for the ﬁrm. What must be done to obtain the optimal path (given an initial condition) is to synthesize these subcases into an optimal sequence. This is carried out in the following section.

A3

A4

A

A2

A5 ,

A

A6

Figure 5.3: Case A: g ≤ r

170

5. Applications to Finance

B4 B3

B5

B6

B9

B2

,

B1

B8

B7

Figure 5.4: Case B: g > r

5.2.3

Synthesis of Optimal Control Paths

To obtain an optimal path, we must synthesize an optimal sequence of subcases. The usual procedure employed is that of the reverse-time construction, ﬁrst developed by Isaacs (1965). Reverse time can only be deﬁned for ﬁnite horizon problems. However, the inﬁnite horizon solution can usually be inferred from the ﬁnite horizon solution if suﬃcient care is exercised. This will be done in Sect. 5.2.4. Our analysis of the ﬁnite horizon problem (5.21) proceeds with the assumption that the terminal time T is assumed to be suﬃciently large. We will make this assumption precise during our analysis. Moreover, we will discuss the solution when T is not suﬃciently large in Remarks 5.2 and 5.4. Deﬁne the reverse-time variable τ as τ = T − t,

5.2. Optimal Financing Model so that

◦

y=

171

dy dt dy = = −y. ˙ dτ dt dτ

◦

As a consequence, y = −y, ˙ and the reverse-time versions of the state and adjoint equations (5.18) and (5.23), respectively, can be obtained by ◦ simply replacing y˙ by y and changing the signs of the right-hand sides. The transversality condition on the adjoint variable λ(t = T ) = λ(τ = 0) = 0

(5.29)

becomes the initial condition in the reverse-time sense. Furthermore, let us parameterize the terminal state by assuming that x(t = T ) = x(τ = 0) = αA ,

(5.30)

where αA is a parameter to be determined. From now on in this section, everything is expressed in the reverse◦

◦

time sense unless otherwise speciﬁed. Using the deﬁnitions of x and λ and the conditions (5.30) and (5.29), we can write reverse-time versions of (5.18) and (5.23) as follows: ◦

x= −r(cu + v)x, x(0) = αA ,

(5.31)

λ= (1 − u − v) − λ{ρ − r(cu + v)}, λ(0) = 0.

(5.32)

◦

This is the starting point for our switching point synthesis. First, we consider Case A. Case A: g ≤ r. Note that the constraint v ≤ 1 is superﬂuous in this case and the only feasible subcases are A1, A2, and A3. Since λ(0) = 0, we have W1 (0) = W2 (0) = −1 and, therefore, Subcase A1. Subcase A1: W2 = rλ − 1 < 0. From Row (1) of Table 5.1, we have u∗ = v ∗ = 0, which gives the state equation (5.31) and the adjoint equation (5.32) as ◦

◦

x= 0 and λ= 1 − ρλ.

(5.33)

172

5. Applications to Finance

With the initial conditions given in (5.29), the solutions for x and λ are x(τ ) = αA and λ(τ ) = (1/ρ)[1 − e−ρτ ].

(5.34)

It is easy to see that because of the assumption 0 ≤ c < 1, it follows that if W2 = rλ − 1 < 0, then W1 = crλ − 1 < 0. Therefore, to remain in this subcase as τ increases, W2 (τ ) must remain negative for some time as τ increases. From (5.34), however, λ(τ ) is increasing asymptotically toward the value 1/ρ and therefore, W2 (τ ) is increasing asymptotically toward the value r/ρ − 1. Since, we have assumed r > ρ, there exists a τ 1 such that W2 (τ 1 ) = (1 − e−ρτ 1 )r/ρ − 1 = 0. It is easy to compute τ 1 = (1/ρ) ln[r/(r − ρ)].

(5.35)

From this expression, it is clear that the ﬁrm leaves Subcase A1 provided τ 1 < T. Moreover, this observation also makes precise the notion of a suﬃciently large T in Case A by having T > τ 1 . Remark 5.2 When T is not suﬃciently large, i.e., when T ≤ τ 1 in Case A, the ﬁrm stays in Subcase A1. The optimal solution in this case is u∗ = 0 and v ∗ = 0, i.e., a policy of no investment. Remark 5.3 Note that if we had assumed r < ρ, the ﬁrm would never have exited from Subcase A1 regardless of the value of T. Obviously, there is no use investing if the rate of return is less than the discount rate. At reverse time τ 1 , we have W2 = 0 and W1 < 0 and the ﬁrm, therefore, is in Subcase A2. Also, λ(τ 1 ) = 1/r since W2 (τ 1 ) = 0. Subcase A2: W2 = rλ − 1 = 0. In this subcase, the optimal controls u∗ = 0, 0 ≤ v ∗ ≤ g/r

(5.36)

from Row (3) of Table 5.1 are singular with respect to v. This case is termed singular because the Hamiltonian maximizing condition does not yield a unique value for the control v. In such cases, the optimal controls are obtained by conditions required to sustain W2 = 0 for a ﬁnite time ◦

◦

interval. This means we must have W = 0, which in turn implies λ= 0. ◦

To compute λ, we substitute (5.36) into (5.32) and obtain ◦

λ= (1 − v ∗ ) − λ[ρ − rv ∗ ].

(5.37)

5.2. Optimal Financing Model

173

Substituting λ = 1/r, its value at τ 1 , in (5.37) and equating the righthand side to zero we obtain r=ρ (5.38) as a necessary condition required to maintain singularity over a ﬁnite time interval following τ 1 . Condition (5.38) is fortuitous and will not generally hold. In fact we have assumed r > ρ. Thus, the ﬁrm will not stay in Subcase A2 for a nonzero time interval. Furthermore, since r > ρ, ◦

we have λ (τ 1 ) = (1 − ρ/r) > 0. Therefore, W2 is increasing from zero and becomes positive after τ 1 . Thus, at τ + 1 the ﬁrm switches to Subcase A3. Subcase A3: W2 = rλ − 1 > 0. The optimal controls in this subcase from Row (2) of Table 5.1 are u∗ = 0, v ∗ = g/r.

(5.39)

The state and the adjoint equations are ◦

x= −gx, x(τ 1 ) = αA ,

(5.40)

λ= (1 − g/r) − λ(ρ − g), λ(τ 1 ) = 1/r,

(5.41)

◦

with values at τ = τ 1 deduced from (5.34) and (5.35). ◦

Since λ (τ 1 ) > 0, λ is increasing at τ 1 from its value of 1/r. A further examination of the behavior of λ(τ ) as τ increases will be carried out under two diﬀerent possible conditions: (i) ρ > g and (ii) ρ ≤ g. ◦

(i) ρ > g: Under this condition, as λ increases, λ decreases and becomes zero at a value obtained by equating the right-hand side of (5.41) to zero, i.e., at ¯ = 1 − g/r . (5.42) λ ρ−g ¯ is, therefore, an asymptote to the solution of (5.41) starting This value λ at λ(τ 1 ) = 1/r. Since r > ρ > g in this case, ¯ − 1 = r(1 − g/r) − 1 = r − ρ > 0, W2 = r λ ρ−g ρ−g

(5.43)

which implies that the ﬁrm continues to stay in Subcase A3. ◦

(ii) ρ ≤ g: Under this condition, as λ(τ ) increases, λ (τ ) increases. So W2 (τ ) = rλ(τ ) − 1 continues to be greater than zero and the ﬁrm continues to remain in Subcase A3.

174

5. Applications to Finance

Remark 5.4 With ρ ≤ g, note that λ(τ ) increases to inﬁnity as τ increases to inﬁnity. This has important implications later when we deal with the solution of the inﬁnite horizon problem. Since the optimal decisions for τ ≥ τ 1 have been found to be independent of αA for T suﬃciently large, we can sketch the solution for Case A in Fig. 5.5 starting with x0 . This also gives the value of αA = x0 eg(T −τ 1 ) = x0 egT [1 − ρ/r]g/ρ , as shown in Fig. 5.5.

A

A

ln[

]

Figure 5.5: Optimal path for case A: g ≤ r Mathematically, we can now express the optimal controls and the optimal state, now in forward time, as u∗ (t) = 0, v ∗ (t) = g/r, x∗ (t) = x0 egt , t ∈ [0, T − τ 1 ],

(5.44)

5.2. Optimal Financing Model u∗ (t) = 0, v ∗ (t) = 0, x∗ (t) = x0 eg(T −τ 1 ) , t ∈ (T − τ 1 , T ],

175 (5.45)

As for λ(t), from (5.34) we have λ(t) =

1 [1 − e−ρ(T −t) ], t ∈ (T − τ 1 , T ]. ρ

(5.46)

For t ∈ [0, T − τ 1 ], we have from (5.41), ˙ λ(t) = λ(ρ − g) − (1 − g/r), λ(T − τ 1 ) = 1/r.

(5.47)

Following Sect. A.1, we can solve this equation as 1 − g/r 1 [1 − e−(ρ−g)(T −τ 1 ) ], t ∈ [0, T − τ 1 ]. λ(t) = e−(ρ−g)(T −τ 1 −t) + r ρ−g (5.48) In this solution for Case A, there is only one switching point provided that T is suﬃciently large (i.e., T > τ 1 in this case). The switching time t = T −τ 1 has an interesting economic interpretation. Namely, it requires at least τ 1 units of time to retain a dollar of earnings to be worthwhile for investment. That means, it pays to invest as much of the earnings as feasible before T − τ 1 , and it does not pay to invest any earnings after T − τ 1 . Thus, T − τ 1 is the point of indiﬀerence between retaining earnings or paying dividends out of earnings. To see this directly, let us suppose the ﬁrm retains one dollar of earnings at T − τ 1 . Since this is the last time that any of the earnings invested will be worthwhile, it is obvious (because all earnings are paid out) that the dollar just invested at T − τ 1 yields dividends at the rate r from T − τ 1 to T. The value of this dividend stream in terms of (T − τ 1 )-dollars is τ1 r re−ρs ds = [1 − e−ρτ 1 ], (5.49) ρ 0 which must be equated to one dollar to ﬁnd the indiﬀerence point. Equating (5.49) to 1 yields precisely the value of τ 1 given in (5.35). With this interpretation of τ 1 , we conclude that enough earnings must be retained so as to make the ﬁrm grow exponentially at the maximum rate of g until t = T − τ 1 . After this time, all of the earnings are paid out and the ﬁrm stops growing. Since g ≤ r (assumed for Case A), the growth in the ﬁrst part of the solution can be ﬁnanced entirely from retained earnings. Thus, there is no need to resort to more expensive external equity ﬁnancing. The latter will not be true, however, in Case B when g > r, which we now discuss.

176

5. Applications to Finance

Case B: g > r. Since g/r > 1, the constraint v ≤ 1 in Case B is relevant. The feasible subcases are B1, B2, B3, B4, and B5 shown adjacent to the darkened lines in Fig. 5.4. As in Case A, it is obvious that the ﬁrm starts (in the reverse-time sense) in Subcase B1. Recall that T is assumed to be suﬃciently large here as well. This statement in Case B will be made precise in the course of our analysis. Furthermore, the solution when T is not suﬃciently large in Case B will be discussed in Remark 5.4. Subcase B1: W2 = rλ − 1 < 0. The analysis of this subcase is the same as Subcase A1. As in that subcase, the ﬁrm switches out at time τ = τ 1 to Subcase B2. Subcase B2: W2 = rλ − 1 = 0. In this subcase, the optimal controls u∗ = 0, 0 ≤ v ∗ ≤ 1

(5.50)

from Row (3) of Table 5.1 are singular with respect to v. As before in Subcase A2, the singular case cannot be sustained for a ﬁnite time because of our assumption r > ρ. As in Subcase A2, W2 is increasing at τ 1 from zero and becomes positive after τ 1 . Thus, at τ + 1 , the ﬁrm ﬁnds itself in Subcase B3. Subcase B3: W1 = crλ − 1 < 0, W2 = rλ − 1 > 0. The optimal controls in this subcase are u∗ = 0, v ∗ = 1,

(5.51)

as shown in Row (5) of Table 5.1. The state and the adjoint equations are ◦ x= −rx, x(τ 1 ) = αB (5.52) with αB , a parameter to be determined, and ◦

λ= λ(r − ρ), λ(τ 1 ) = 1/r.

(5.53)

Obviously, earnings are growing exponentially at rate r and λ(τ ) is increasing at rate (r − ρ) as τ increases from τ 1 . Since λ(τ 1 ) = 1/r,

5.2. Optimal Financing Model we have

177

λ(τ ) = (1/r)e(r−ρ)(τ −τ 1 ) for τ ≥ τ 1 .

(5.54)

As λ increases, W1 increases and becomes zero at a time τ 2 deﬁned by W1 (τ 2 ) = crλ(τ 2 ) − 1 = ce(r−ρ)(τ −τ 1 ) − 1 = 0,

(5.55)

which, in turn, gives 1 τ2 = τ1 + ln r−ρ

1 . c

(5.56)

At τ + 2 , the ﬁrm switches to Subcase B4. Before proceeding to Subcase B4, let us observe that in Case B, we can now deﬁne T to be suﬃciently large when T > τ 2 . See Remark 5.4 when T ≤ τ 2 . Subcase B4: W1 = crλ − 1 = 0. In Subcase B4, the optimal controls are 0 ≤ u∗ ≤ (g − r)/rc, v ∗ = 1.

(5.57)

From Row (6) in Table 5.1, these controls are singular with respect to u. To maintain this singular control over a ﬁnite time period, we must ◦

keep W1 = 0 in the interval. This means we must have W 1 (τ 2 ) = 0, ◦

◦

which, in turn, implies λ (τ 2 ) = 0. To compute λ, we substitute (5.57) into (5.32) and obtain ◦

λ= −u∗ − λ{ρ − r(cu∗ + 1)}.

(5.58)

At τ 2 , W1 (τ 2 ) = 0 gives λ(τ 2 ) = 1/rc. With this in (5.58), its right-hand side equals zero only when r = ρ. But we have assumed r > ρ throughout Sect. 5.2, and therefore a singular path cannot be sustained for τ 2 > 0, and the ﬁrm will not stay in Subcase B4 for a ﬁnite amount of time. Furthermore, from (5.58), we have ◦

λ (τ 2 ) =

r−ρ > 0, rc

(5.59)

which implies that λ is increasing and therefore, W1 is increasing. Thus at τ + 2 , the ﬁrm switches to Subcase B5.

178

5. Applications to Finance

Subcase B5: W1 = crλ − 1 > 0. The optimal controls in this subcase from Row (4) of Table 5.1 are u∗ =

g−r , v ∗ = 1. rc

(5.60)

Then from (5.31) and (5.32), the reverse-time state and the adjoint equations are ◦ x= −gx, (5.61) ◦ g−r λ= −( ) + λ(g − ρ). (5.62) rc ◦

Since λ (τ 2 ) > 0 from (5.59), λ(τ ) is increasing at τ 2 from its value λ(τ 2 ) = 1/rc > 0. Furthermore, we have g > r in Case B, which together with r > ρ, assumed throughout Sect. 5.2, makes g > ρ. This implies that the second term on the right-hand side of (5.62) is increasing. Moreover, the second term dominates the ﬁrst term for τ > τ 2 , since λ(τ 2 ) = 1/(rc) > 0, and r > ρ and g > r imply g − ρ > g − r > 0. Thus, ◦

λ (τ ) > 0 for τ > τ 2 , and λ(τ ) increases with τ . Therefore, the ﬁrm continues to stay in Subcase B5. Remark 5.5 Note that λ(τ ) in Case B increases without bound as τ becomes large. This will have important implications when dealing with the inﬁnite horizon problem in Sect. 5.2.4. As in Case A, we can obtain this optimal solution explicitly in forward time, and we ask you to do this in Exercise 5.9. We now can sketch the complete solution for Case B in Fig. 5.6. In this solution, there are two switching points instead of just one as in Case A. The reason for two switching points becomes quite clear when we interpret the signiﬁcance of τ 1 and τ 2 . It is obvious that τ 1 has the same meaning as before. Namely, if τ 1 is the remaining time to the horizon, the ﬁrm is indiﬀerent between retaining a dollar of earnings or paying it out as dividends. Intuitively, it seems that since external equity is more expensive than retained earnings as a source of ﬁnancing, investment ﬁnanced by external equity requires more time to be worthwhile. That is, 1 1 ln >0 (5.63) τ2 − τ1 = r−ρ c as obtained in (5.56), should be the time required to compensate for the ﬂoatation cost of external equity. Let us see why.

5.2. Optimal Financing Model

179

= B1 B3

B5

Earnings

Figure 5.6: Optimal path for case B: g > r When the ﬁrm issues a dollar’s worth of stock at time t = T − τ 2 , it incurs a future dividend obligation in the amount of one (T − τ 2 )dollar, even though the capital acquired is only c dollars because of the ﬂoatation cost (1 − c). Since we are attempting to ﬁnd the breakeven time for external equity, it is obvious that retaining all of the earnings for investment is still proﬁtable. Thus, there is no dividend from (T −τ 2 ) to (T − τ 1 ), and the ﬁrm grows at the rate r. Therefore, this investment of c dollars at time (T −τ 2 ) grows into cer(τ 2 −τ 1 ) dollars at time (T −τ 1 ). From the point of view of a buyer of the stock at time (T − τ 2 ), since no dividend is paid until time (T − τ 1 ) and since the stockholder’s required rate of return is ρ, the ﬁrm’s future dividend obligation at time (T − τ 1 ) is eρ(τ 2 −τ 1 ) in terms of (T − τ 1 )-dollars. But then we must have eρ(τ 2 −τ 1 ) = cer(τ 2 −τ 1 ) ,

(5.64)

180

5. Applications to Finance

which can be rewritten precisely as (5.63). Moreover, the ﬁrm is marginally indiﬀerent between investing any costless retained earnings at time (T − τ 1 ) or paying it all out as dividends. This also means that the ﬁrm will be indiﬀerent between having the new available capital of cer(τ 2 −τ 1 ) dollars at time (T − τ 1 ) as a result of issuing a dollar’s worth of stock at time (T − τ 2 ), or not having it. Thus, we can conclude that the ﬁrm is indiﬀerent between issuing a dollar’s worth of stock at time (T − τ 2 ) or not issuing it. This means that before time (T − τ 2 ), it pays to issue stocks at as large a rate as feasible, and after time (T − τ 2 ), it does not pay to issue any external equity at all. We have now provided an intuitive justiﬁcation of (5.63) and concluded that all earnings must be retained from time (T − τ 2 ) to (T − τ 1 ). Because r > ρ, it follows that the excess return on the proceeds c from the new stock issue is cer(τ 2 −τ 1 ) − ceρ(τ 2 −τ 1 ) at time (T − τ 1 ). When discounted this amount back to time (T − τ 2 ), we can use (5.63) or (5.64) to see that % & cer(τ 2 −τ 1 ) − ceρ(τ 2 −τ 1 ) e−ρ(τ 2 −τ 1 ) = celn(1/c) − c = 1 − c. Thus, the excess return from time (T − τ 2 ) to (T − τ 1 ) recovers precisely the ﬂoatation cost. Remark 5.6 When T is not suﬃciently large, i.e., when T < τ 2 in Case B, the optimal solution is the same as in Remark 5.1 when T ≤ τ 1 . If τ 1 < T ≤ τ 2 , then the optimal solution is u∗ = 0 and v ∗ = 1 until t = T − τ 1 . For t > T − τ 1 , the optimal solution is u∗ = 0 and v ∗ = 0. Having completely solved the ﬁnite horizon case, we now turn to the inﬁnite horizon case.

5.2.4

Solution for the Inﬁnite Horizon Problem

As indicated in Sect. 3.6 for the inﬁnite horizon case, the transversality condition must be changed to lim e−ρt λ(t) = 0.

t→∞

(5.65)

Furthermore, this condition may no longer be a necessary condition; see Sect. 3.6. It is a suﬃcient condition for optimality however, in conjunction with the other suﬃciency conditions stated in Theorem 2.1.

5.2. Optimal Financing Model

181

As demonstrated in Example 3.7, a common method of solving an inﬁnite horizon problem is to take the limit as T → ∞ of the ﬁnite horizon solution and then prove that the limiting solution obtained solves the inﬁnite horizon problem. The proof is important because the limit of the solution may or may not solve the inﬁnite horizon problem. The proof is usually based on the suﬃciency conditions of Theorem 2.1, modiﬁed slightly as indicated above for the inﬁnite horizon case. We now analyze the inﬁnite horizon case following the above procedure. We start with Case A. Case A: g ≤ r. Let us ﬁrst consider the case ρ > g and examine the solution in forward time obtained in (5.44)–(5.48) as T goes to inﬁnity. Clearly (5.45) and (5.46) disappear, and (5.44) and (5.48) can be written as u∗ (t) = 0, v ∗ (t) = g/r, x∗ (t) = x0 egt , t ≥ 0, λ(t) =

1 − g/r ¯ t ≥ 0. = λ, ρ−g

(5.66) (5.67)

Clearly λ(t) satisﬁes (5.65). Furthermore, W2 (t) = rλ − 1 =

r−ρ > 0, t ≥ 0, ρ−g

which implies that the ﬁrm is in Subcase A3 for t ≥ 0. The maximum principle holds, and (5.66) and (5.67) represent an optimal solution for the inﬁnite horizon problem. Note that the assumption ρ > g together with our overall assumption that ρ < r gives g < r so that 1 − v ∗ > 0, which means a constant fraction of earnings is being paid as dividends. ¯ in this case is a conNote that the value of the adjoint variable λ stant and its form is reminiscent of Gordon’s classic formula; see Gordon ¯ represents the (1962). In the control theory framework, the value of λ marginal worth per additional unit of earnings. Obviously, a unit increase in earnings will mean an increase of 1 − v ∗ or 1 − g/r units in dividends. This, of course, should be capitalized at a rate equal to the discount rate less the growth rate (i.e., ρ−g), which is precisely Gordon’s formula. For ρ ≤ g, it is clear from (5.48) that λ(t) does not satisfy (5.65). A moment’s reﬂection shows that for ρ ≤ g, the objective function can be made inﬁnite. For example, any control policy with earnings growing at

182

5. Applications to Finance

rate q, ρ ≤ q ≤ g, coupled with a partial dividend payout, i.e., a constant v such that 0 < v < 1, gives an inﬁnite value for the objective function. That is, with u∗ = 0, v ∗ = q/r < 1, we have ∞ ∞ J= e−ρt (1 − u∗ − v ∗ )x∗ dt = e−ρt (1 − q/r)x0 eqt = ∞. 0

0

Since there are many policies which give an inﬁnite value to the objective function, the choice among them may be decided on subjective grounds. We will brieﬂy discuss only the constant (over time) optimal policies. If g < r, then the rate of growth q may be chosen in the closed interval [ρ, g]; if g = r, then q may be chosen in the half-open interval [ρ, r). In either case, the choice of a low rate of growth (i.e., a high proportional dividend payout) would mean a higher dividend rate (in dollars per unit time) early in time, but a lower dividend rate later in time because of the slower growth rate. Similarly the choice of high growth rate means the opposite in terms of dividend payments in dollars per unit time. To conclude, we note that for ρ ≤ g in Case A, the limiting solution of the ﬁnite case is an optimal solution for the inﬁnite horizon problem in the sense that the objective function becomes inﬁnite. However, this will not be the situation in Case B; see also Remark 5.7. Case B: g > r. The limit of the ﬁnite horizon optimal solution is to grow at the maximum allowable growth rate with g−r and v = 1 rc all the way. Since τ 1 disappears in the limit, the stockholders will never collect dividends. The ﬁrm has become an inﬁnite sink for investment. In fact, the limiting solution is a pessimal solution because the value of the objective function associated with it is zero. From the point of view of optimal control theory, this can be explained as before in Case A when ρ ≤ g. In Case B, we have g > r so that (since r > ρ throughout the chapter) we have ρ < g. For this, as noted in Remark 5.5, λ(τ ) increases without bound as τ increases and, therefore, (5.64) does not have a solution. As in Case A with ρ < g, any control policy with earnings growing at rate q ∈ [ρ, g] coupled with a constant v, 0 < v < 1, has an inﬁnite value for the objective function. u=

5.2. Optimal Financing Model

183

In summary, we note that the only nondegenerate case in the inﬁnite horizon problem is when ρ > g. In this case, which occurs only in Case A, the policy of maximum allowable growth is optimal. On the other hand, when ρ ≤ g, whether in Case A or B, the inﬁnite horizon problem has nonunique policies with inﬁnite values for the objective function. Before solving a numerical example, we will make an interesting remark concerning Case B. Remark 5.7 Let (u∗T , vT∗ ) denote the optimal control for the ﬁnite ∗ ) denote any optimal conhorizon problem in Case B. Let (u∗∞ , v∞ trol for the inﬁnite horizon problem in Case B. We already know that ∗ ) = ∞. Deﬁne an inﬁnite horizon control (u , v ) by extendJ(u∗∞ , v∞ ∞ ∞ ing (u∗T , vT∗ ) as follows: (u∞ , v∞ ) = lim (u∗T , vT∗ ). T →∞

We now note that for our model in Case B, we have lim J(u∗T , vT∗ ) = ∞ and J( lim (u∗T , vT∗ )) = J(u∞ , v∞ ) = 0.

T →∞

T →∞

(5.68)

Obviously (u∞ , v∞ ) is not an optimal control for the inﬁnite horizon problem. Since the two terms in (5.68) are not equal, we can say in technical terms that J(u, v), regarded as a mapping, is not a closed mapping. However, if we introduce a salvage value Bx(T ), B > 0, for the ﬁnite horizon problem, then the new objective function, ⎧ ⎪ ⎨ T e−ρt (1 − u − v)xdt + Bx(T )e−ρT , if T < ∞, 0 JB (u, v) = ⎪ ⎩ ∞ e−ρt (1 − u − v)xdt + limT →∞ {Bx(T )e−ρT }, if T = ∞, 0

is a closed mapping in the sense that lim JB (u∗T , vT∗ ) = ∞ and JB ( lim (u∗T , vT∗ )) = JB (u∞ , v∞ ) = ∞

T →∞

T →∞

for the modiﬁed model. Example 5.1 We will now assign numbers to the various parameters in the optimal ﬁnancing problem in order to compute the optimal solution. Let x0 = 1000/month, T = 60 months, r = 0.15, ρ = 0.10, g = 0.05, c = 0.98.

184

5. Applications to Finance

Solution Since g ≤ r, the problem belongs to Case A. We compute τ1 =

1 ln[r/(r − ρ)] = 10 ln 3 ≈ 11 months. ρ

The optimal controls for the problem are u∗ = 0, v ∗ = g/r = 1/3, t ∈ [0, 49), u∗ = 0, v ∗ = 0,

t ∈ [49, 60],

and the optimal state trajectory is

x(t) =

⎧ ⎪ ⎨ 1000e0.05t , t ∈ [0, 49), ⎪ ⎩ 1000e2.45 ,

t ∈ [49, 60].

The value of the objective function is J

∗

49

=

e 0

−0.1t

(1 − 1/3)(1000)e

0.05t

60

dt + 49

1000e2.45 · e−0.1t dt

= 12, 578.75. Note that the inﬁnite horizon problem is well deﬁned in this case, since g < ρ and g < r. The optimal controls are u∗ = 0, v ∗ = g/r = 1/3, and

∞

J= 0

1 e−0.1t (2/3)(1000)e0.05t dt = 2000/0.15 = 13, 333 . 3

In Exercise 5.14, you are asked to extend the optimal ﬁnancing model to allow for debt ﬁnancing. Exercise 5.15 requires you to reformulate the optimal ﬁnancing model (5.21) with decisions expressed in dollars per unit of time rather than in terms relative to x. Exercise 5.16 extends the model to allow the rate of return on the assets to decrease as the assets grow.

Exercises for Chapter 5

185

Exercises for Chapter 5 E 5.1 Find the optimal policies for the simple cash balance model (Sects. 5.1.1 and 5.1.2) with x0 = 2, y0 = 2, U1 = U2 = 5, T = 1, α = 0.01, and the following speciﬁcations for the interest rates: (a) r1 (t) = 1/2, r2 (t) = 1/3. (b) r1 (t) = t/2, r2 (t) = 1/3. (c) Sketch the optimal policy in (b) in the (t, λ2 /λ1 ) space, like in Fig. 5.2. E 5.2 Formulate the extension of the model in Sect. 5.1.1 when overdraft and short selling are disallowed in the following two cases: (a) α = 0 and (b) α > 0. State the maximum principle (4.29) as it applies to these cases. Hint: Adjoin the control constraints to the Hamiltonian in forming the Lagrangian. For (b), write u = u1 − u2 as in (5.10). E 5.3 It is possible to guess the optimal solution for Exercise 5.2 when α = 0, T = 10, x0 = 0, y0 = 3, ⎧ ⎪ ⎨ 0 for 0 ≤ t < 5, r1 (t) = ⎪ ⎩ 0.3 for 5 ≤ t ≤ 10, r2 (t) = 0.1

for 0 ≤ t ≤ 10,

and U1 = U2 = ∞ (allowing for impulse controls). Show that the optimum policy remains the same for each α ∈ [0, 1 − 1/e]. Hint: Use an elementary compound interest argument. E 5.4 Do the following for Exercise 5.3 with U1 = U2 = 1, so that the control constraints are −1 ≤ u ≤ 1. (a) (b) (c) (d)

Give reasons why the solution shown in Fig. 5.7 is optimal. Compute f (t∗ ) in terms of t∗ . Compute J in terms of t∗ . Find t∗ that maximizes J by setting dJ/dt∗ = 0.

Hint: Because this is a long and tedious calculus problem, you may wish to use Mathematica or MAPLE to solve this problem.

186

5. Applications to Finance

Figure 5.7: Solution for Exercise 5.4

E 5.5 For the solution found in Exercise 5.4, show by using the maximum principle (4.29) that the adjoint trajectories are:

λ1 (t) =

⎧ ⎪ ⎨ λ1 (0) = e1.5 ,

0 ≤ t ≤ 5,

⎪ ⎩ λ (5)e−0.3(t−5) = e3−0.3t , 5 ≤ t ≤ 10, 1

and

λ2 (t) =

⎧ ⎪ ⎨ λ2 (0)e−0.1t∗ = e1.5+0.1(t∗ −t) , 0 ≤ t ≤ f (t∗ ) ≈ 6.52, ⎪ ⎩

2 3

+ 13 e3−0.3t ,

f (t∗ ) < t ≤ 10,

where t∗ ≈ 1.97. Sketches of these functions are shown in Fig. 5.8. E 5.6 Argue that as the lower and upper bounds on u go to −∞ and +∞ in Exercise 5.4, respectively, t∗ goes to 0 and f (t∗ ) goes to 5. Show that this solution is consistent with the guess in Exercise 5.3. Finally, ﬁnd the corresponding impulse solution and show that it satisﬁes the maximum principle as applied in Exercise 5.2. E 5.7 Discuss the optimal equity ﬁnancing model of Sect. 5.2.1 when c = 1. Show that only one control variable is needed. Then solve the problem.

Exercises for Chapter 5

187

Figure 5.8: Adjoint trajectories for Exercise 5.5 E 5.8 What happens in the optimal equity ﬁnancing model when r < ρ? Guess the optimal solution (without actually solving it). E 5.9 In Sect. 5.2.3, we obtained the optimal solution in Case B. Express the corresponding control, state, and adjoint trajectories in forward time. E 5.10 Let g = 0.12 in Example 5.1. Re-solve the ﬁnite horizon problem with this new value of g. Also, for the inﬁnite horizon problem, state a policy which yields an inﬁnite value for the objective function. E 5.11 Reformulate and solve the simple cash balance problem of Sects. 5.1.1 and 5.1.2, if the earnings on bonds are paid in cash. E 5.12 Add a salvage value function e−ρT Bx(T ), where B ≥ 0, to the objective function in the problem (5.21) and analyze the modiﬁed problem due to Sethi (1978b). Show how the solution changes as B varies from 0 to 1/rc. E 5.13 Suppose we extend the model of Exercise 5.12 to include debt. For this let z denote the total debt at time t and w ≥ 0 denote the

188

5. Applications to Finance

amount of debt issued expressed as a proportion of current earnings. Then the state equation for z is z˙ = wx, y(0) = y0 . How would you modify the objective function, the state equation for x, and the growth constraint (5.19)? Assume i to be the constant interest rate on debt, and i < r. E 5.14 Remove the assumption of an arbitrary upper bound g on the growth rate in the ﬁnancing model of Sect. 5.2.1 by introducing a convex cost associated with the growth rate. With r re-interpreted now as the gross rate of return, obtain the net increase in rate of earnings by the rate of increase in gross earnings less the cost associated with the growth rate. Also assume c = 1 as in Exercise 5.7. Formulate the resulting model and apply the maximum principle to ﬁnd the form of the optimal policy. You may assume the cost function to be quadratic in the growth rate to get an explicit form for the solution. E 5.15 Reformulate the optimal ﬁnancing model (5.21) with y(t) as the state variable, U (t) as the new equity ﬁnancing rate in dollars per unit of time, and V (t) as the retained earnings in dollars per unit of time. Hint: This formulation has mixed constraints requiring the Lagrangian formulation of the maximum principle (3.42) introduced in Chap. 3. Note further that it can be converted into the form (5.21) by setting U = ux, V = vx, and x = ry. E 5.16 In Exercise 5.15, we assume a constant rate of return r on the assets so that the total earnings rate at time t is ry(t) dollars per unit of time. Extend this formulation to allow for a decreasing marginal rate of return as the assets grow. More speciﬁcally, replace ry by an increasing, y ) = ρ for some strictly concave function R(y) > 0 with R (0) = r and R (¯ y¯ > y0 > 0. Obtain the optimal solution in the case when r > g > ρ, 0 < c < 1, T suﬃciently large, and y0 < y1 < y¯, where y1 is deﬁned by the relation R(y1 )/y1 = g. See Perrakis (1976). E 5.17 Find the form of the optimal policy for the following model due to Davis and Elzinga (1971): T −ρt −ρT e (1 − v)Erdt + P (T )e max J = u,v

0

Exercises for Chapter 5

189

subject to P˙ = k[rE(1 − v) − ρP ],

P (0) = P0 ,

E˙ = rE[v + u(c − E/P )], E(0) = E0 , and the control constraints u ≥ 0, v ≥ 0, cu + v ≤ g/r. Here P denotes the price of a stock, E denotes equity per stock and k > 0 is a constant. Also, assume r > ρ > g and 1/c < r/ρ < 1/c + (ck + 1)g/(ρck). This example requires the use of the generalized LegendreClebsch condition (D.69) in Appendix D.8.

Chapter 6

Applications to Production and Inventory Applications of optimization methods to production and inventory problems date back at least to the classical EOQ (Economic Order Quantity) model or the lot size formula of Harris (1913). The EOQ is essentially a static model in the sense that the demand is constant and only a stationary solution is sought. A dynamic version of the lot size model was analyzed by Wagner and Whitin (1958). The solution methodology used there was dynamic programming. An important dynamic production planning model was developed by Holt et al. (1960). In their model, referred to as the HMMS model, they considered both production costs and inventory holding costs over time. They used calculus of variations techniques to solve the continuous-time version of their model. In Sect. 6.1, a model of Thompson and Sethi (1980), similar to the HMMS model, is formulated and completely solved using optimal control theory. The turnpike solution is also obtained when the horizon is inﬁnite. In Sect. 6.2, we introduce the wheat trading model of Ijiri and Thompson (1970), in which a wheat speculator must buy and sell wheat in an optimal way in order to take advantage of changes in the price of wheat over time. In Sects. 6.2.1–6.2.3, we solve the model when the shortselling of wheat is allowed. In Sect. 6.2.4, we follow Norstr¨ om (1978) to solve a simple example that disallows short-selling.

© Springer Nature Switzerland AG 2019 S. P. Sethi, Optimal Control Theory, https://doi.org/10.1007/978-3-319-98237-3 6

191

192

6. Applications to Production and Inventory

In Sect. 6.3, we introduce a warehousing constraint, i.e., an upper bound on the amount of wheat that can be stored, in the wheat trading model. In addition to being realistic, the introduction of the warehousing constraint helps us to illustrate the concepts of decision and forecast horizons by means of examples. This section is expository in nature, but theoretical developments of these ideas are available in the literature.

6.1

Production-Inventory Systems

Many manufacturing enterprises use a production-inventory system to manage ﬂuctuations in consumer demand for their products. Such a system consists of a manufacturing plant and a ﬁnished goods warehouse to store products which are manufactured but not immediately sold. Once a product is made and put into inventory, it incurs inventory holding costs of two kinds: (1) costs of physically storing the product, insuring it, etc.; and (2) opportunity cost of having the ﬁrm’s money invested or tied up in the unsold inventory. The advantages of having products in inventory are: ﬁrst, that they are immediately available to meet demand; second, that excess production during low demand periods can be stored in the warehouse so it will be available for sale during high demand periods. This usually permits the use of a smaller manufacturing plant than would otherwise be necessary, and also reduces the diﬃculties of managing the system. The optimization problem is to balance the beneﬁts of production smoothing versus the costs of holding inventory. Works that apply control theory to production and inventory problems have been reviewed in Sethi (1978a, 1984).

6.1.1

The Production-Inventory Model

We consider a factory producing a single homogeneous good and having a ﬁnished goods warehouse. To state the model we deﬁne the following quantities: I(t) = the inventory level at time t (state variable), P (t) = the production rate at time t (control variable), S(t) = the exogenously given sales rate at time t; assumed to be bounded and diﬀerentiable for t ≥ 0, T = the length of the planning period, Iˆ = the inventory goal level,

6.1. Production-Inventory Systems

193

I0 = the initial inventory level, Pˆ = the production goal level, h = the inventory holding cost coeﬃcient; h > 0, c = the production cost coeﬃcient; c ≥ 0, ρ = the constant nonnegative discount rate; ρ ≥ 0. The interpretation of the inventory goal level Iˆ is that it is a safety stock that the company wants to keep on hand. For example, Iˆ could be 2 months of average sales or Iˆ could be 100 units of the ﬁnished goods. Similarly, the production goal level Pˆ can be interpreted as the most eﬃcient level at which it is desired to run the factory. With this notation, the state equation is given by the stock-ﬂow diﬀerential equation ˙ = P (t) − S(t), I(0) = I0 , I(t)

(6.1)

which says that the inventory at time t is increased by the production rate and decreased by the sales rate. The objective function of the model is: T c −ρt h 2 2 ˆ ˆ min J = e [ (I − I) + (P − P ) ]dt . (6.2) 2 2 0 The interpretation of the objective function is that we want to keep the ˆ and also to keep the inventory as close as possible to its goal level I, production rate P as close as possible to its goal level Pˆ . The quadratic ˆ 2 and (c/2)(P − Pˆ )2 impose “penalties” for having terms (h/2)(I − I) either I or P not being close to its corresponding goal level. Next we apply the maximum principle to solve the optimal control problem speciﬁed by (6.1) and (6.2). A stochastic extension of this problem will be carried out in Sect. 12.2.

6.1.2

Solution by the Maximum Principle

We now associate an adjoint function λ with Eq. (6.1) and can write the current-value Hamiltonian function as h ˆ 2 − c (P − Pˆ )2 . (6.3) H = λ(P − S) − (I − I) 2 2 In (6.3), we have used the negative of the (undiscounted) integrand in (6.2), since the minimization of J in (6.2) is equivalent to the maximization of −J. To apply the Pontryagin maximum principle, we diﬀerentiate (6.3) and set the resulting expression equal to 0, which gives

194

6. Applications to Production and Inventory ∂H = λ − c(P − Pˆ ) = 0. ∂P

(6.4)

From this we obtain the optimal production rate P ∗ (t) = Pˆ + λ(t)/c.

(6.5)

We should mention that in writing (6.5), we are allowing negative production (or disposal). Of course, the situation of a disposal will not arise if we assume a suﬃciently large Pˆ and a suﬃciently small I0 . Remark 6.1 If P is constrained to be nonnegative, then the form of the optimal control will be P ∗ (t) = max{Pˆ + λ(t)/c, 0}.

(6.6)

This case will be treated in Sect. 6.1.6. By substituting (6.5) into (6.1), we obtain I˙ = Pˆ + λ/c − S, I(0) = I0 .

(6.7)

The equation for the adjoint variable is easily found to be ∂H ˆ λ(T ) = 0. = ρλ + h(I − I), λ˙ = ρλ − ∂I

(6.8)

We see that (6.7) has the initial boundary speciﬁed and (6.8) has the terminal boundary speciﬁed, so together these give a two-point boundary value problem. We will employ a method to solve these two equations simultaneously, which works only in some special cases including the present case. The method is the well-known trick used to solve simultaneous diﬀerential equations by diﬀerentiation and substitution until one of the variables is eliminated. Speciﬁcally, we diﬀerentiate (6.7) with respect to t, which creates an equation with λ˙ in it. We then use (6.8) to eliminate λ˙ and (6.7) to eliminate λ from the resulting equation as follows: ˙ − S˙ = ρ(λ/c) + (h/c)(I − I) ˆ − S˙ I¨ = λ/c ˆ − S. ˙ = ρ(I˙ − Pˆ + S) + (h/c)(I − I) We rewrite this as I¨ − ρI˙ − α2 I = −α2 Iˆ − S˙ − ρ(Pˆ − S),

(6.9)

6.1. Production-Inventory Systems

195

where the constant α is given by α=

#

h/c.

(6.10)

We can now solve (6.9) by using the standard method described in Appendix A. The auxiliary equation for (6.9) is m2 − ρm − α2 = 0, which has the two real roots # # m1 = (ρ − ρ2 + 4α2 )/2, m2 = (ρ + ρ2 + 4α2 )/2;

(6.11)

note that m1 < 0 and m2 > 0. We can therefore write the general solution to (6.9) as (6.12) I(t) = a1 em1 t + a2 em2 t + Q(t), I(0) = I0 , where Q(t) is a particular integral of (6.9). We will say that Q(t) is a special particular integral of (6.9) if it has no additive terms involving em1 t and em2 t . From now on we will always assume that Q(t) is a special particular integral. Although (6.12) has two arbitrary constants a1 and a2 , it has only one boundary condition. To get the other boundary condition we differentiate (6.12), substitute the result into (6.7), and solve for λ. We obtain λ(t) = c(m1 a1 em1 t + m2 a2 em2 t + Q˙ + S − Pˆ ), λ(T ) = 0.

(6.13)

Note that we have imposed the boundary condition on λ so that we can determine the constants a1 and a2 . For ease of expressing a1 and a2 , let us deﬁne two constants b1 = I0 − Q(0), ˙ ) − S(T ). b2 = Pˆ − Q(T

(6.14) (6.15)

We now impose the boundary conditions in (6.12) and (6.13) and solve for a1 and a2 as follows: a1 = a2 =

b2 em1 T − m2 b1 e(m1 +m2 )T , m1 e2m1 T − m2 e(m1 +m2 )T b1 m1 e2m1 T − b2 em1 T . m1 e2m1 T − m2 e(m1 +m2 )T

(6.16) (6.17)

196

6. Applications to Production and Inventory

If we recall that m1 is negative and m2 is positive, then when T is suﬃciently large so that em1 T and e2m1 T are negligible, we can write a 1 ≈ b1 , b2 −m2 T e . a2 ≈ m2

(6.18) (6.19)

Note that for a large T, e−m2 T is close to zero and, therefore, a2 is close to zero. However, the reason for retaining the exponential term in (6.19) is that a2 is multiplied by em2 t in (6.13), which, while small when t is small, becomes large and important when t is close to T. With these values of a1 and a2 and with (6.5), (6.12), and (6.13), we now write the expressions for I ∗ , P ∗ , and λ. We will break each expression into three parts: the ﬁrst part labeled Starting Correction is important only when t is small; the second part labeled Turnpike Expression is signiﬁcant for all values of t; and the third part labeled Ending Correction is important only when t is close to T. Starting Correction I ∗ = (b1 em1 t )+ P ∗ = (m1 b1 em1 t )+ λ = c(m1 b1 em1 t )+

Turnpike Expression (Q)+ (Q˙ + S)+

' ( c Q˙ + S − Pˆ +

Ending Correction b2 m2 (t−T ) e (6.20) m2 ' ( b2 em2 (t−T ) (6.21) ( ' (6.22) c b2 em2 (t−T )

Note that if b1 = 0, which by (6.14) means I0 = Q(0), then there is no starting correction. In other words, I0 = Q(0) is a starting inventory that causes the solution to be on the turnpike initially. In the same way, if b2 = 0, then the ending correction vanishes in each of these formulas, and the solution stays on the turnpike until the end. Expressions (6.20) and (6.21) represent approximate closed-form solutions for the optimal inventory and production functions I ∗ and P ∗ as long as S is such that the special particular integral Q can be found explicitly. For such examples of S; see Sect. 6.1.4.

6.1.3

The Inﬁnite Horizon Solution

It is important to show that this solution also makes sense when T → ∞. In this case it is usual to assume that the discount rate ρ > 0 and the sales rate S does not grow too fast so that the objective function (6.2)

6.1. Production-Inventory Systems

197

remains ﬁnite. One can then show that the limit of the ﬁnite horizon solution as T → ∞ also solves the inﬁnite horizon problem. Note that as T → ∞, the ending correction terms in (6.20)–(6.22) disappear because e−m2 T goes to 0. We now have

Since we would like

λ(t) = c[m1 b1 em1 t + Q˙ + S − Pˆ ].

(6.23)

lim e−ρt λ(t) = 0,

(6.24)

t→∞

we would require that S + Q˙ grows slower asymptotically than the discount rate ρ. One can easily verify that this condition holds for the demand terms discussed in Sect. 6.1.4 that follows. Moreover, the condition is easy to check for any given speciﬁc demand S(t) for which the particular integral Q(t) is known. By the suﬃciency of the maximum principle conditions (Sect. 2.4), it can be veriﬁed that the limiting solution I ∗ (t) = b1 em1 t + Q, P ∗ (t) = m1 b1 em1 t + Q˙ + S

(6.25)

is optimal. If I(0) = Q(0), the solution is always on the turnpike. Note ¯ = {Q, Q˙ + S, c(Q˙ + S − Pˆ )} represents a non¯ P¯ , λ} that the triple {I, stationary turnpike. If I(0) = Q(0), then b1 = 0 and the expressions (6.25) imply that the paths of inventory and production only approach the turnpike but never attain it.

6.1.4

Special Cases of Time Varying Demands

In this section, we provide some important cases of time varying demands including seasonal demands. These involve polynomial or sinusoidal demand functions. We then solve some numerical examples of the model described in Sect. 6.1.1 for ρ = 0 and T < ∞. For the ﬁrst example, we assume that S(t) is a polynomial of degree 2p or 2p−1 so that S (2p+1) = 0, where S (k) denotes the kth time derivative of S with respect to t. In other words, S(t) = C0 t2p + C1 t2p−1 + . . . + C2p ,

(6.26)

where at least one of C0 and C1 is not zero. Then, from Zwillinger (2003), a particular integral of (6.9) is 1 1 1 Q(t) = Iˆ + 2 S (1) + 4 S (3) + · · · + 2p S (2p−1) . α α α

(6.27)

198

6. Applications to Production and Inventory

In Exercise 6.2 the reader is asked to verify this by direct substitution. For the second example, we assume that S(t) is a sinusoidal demand function of form S(t) = A sin(πBt + C) + D, (6.28) where A, B, C, and D are constants. In Exercise 6.3 you are asked to verify that a particular integral of (6.9) for S in (6.28) is πAB cos(πBt + C). (6.29) + π2B 2 It is well known in the theory of diﬀerential equations that demands that are sums of functions of the form (6.26) and/or (6.28) give rise to solutions that are sums of functions of form (6.27) and/or (6.29). Q(t) = Iˆ +

α2

Example 6.1 Assume Pˆ = 30, Iˆ = 15, T = 8, ρ = 0, and h = c = 1 so that α = 1, m1 = −1, and m2 = 1. Assume S(t) = t(t − 4)(t − 8) + 30 = t3 − 12t2 + 32t + 30. Solution It is then easy to show from (6.27) that ˙ = 6t − 24. Q(t) = 3t2 − 24t + 53 and Q(t) Also from (6.14), (6.15), and (6.16), we have a1 ≈ b1 = I0 − 53 and b2 = −24. Then, from (6.20) and (6.21), I ∗ (t) = (I0 − 53)e−t + Q(t) − 24et−8 , ˙ + S(t) − 24et−8 . P ∗ (t) = −(I0 − 53)e−t + Q(t) In Fig. 6.1 the graphs of sales, production, and inventory are drawn with I0 = 10 (a small starting inventory), which makes b1 = −43. In Fig. 6.2 the same graphs are drawn with I0 = 50 (a large starting inventory), which makes b1 = −3. In Fig. 6.3 the same graphs are drawn with I0 = 30, which makes b1 = −23. Note that initially during the time from 0 to 4, the three cases are quite diﬀerent, but during the time from 4 to 8, they are nearly identical. The ending inventory ends up being 29 in all three cases. Example 6.2 Assume that S(t) = A + Bt +

K

Ck sin(πDk t + Ek ),

(6.30)

k=1

where the constants A, B, Ck , Dk , and Ek are estimated from future demand data by means of one of the standard forecasting techniques such as those in Brown (1959, 1963).

6.1. Production-Inventory Systems

199

60 50 40 30 20 10 0

0

1

2

3

4

5

6

7

8

Figure 6.1: Solution of Example 6.1 with I0 = 10

60 50 40 30 20 10 0

0

1

2

3

4

5

6

7

8

Figure 6.2: Solution of Example 6.1 with I0 = 50

200

6. Applications to Production and Inventory

60 50 40 30 20 10 0

0

1

2

3

4

5

6

7

8

Figure 6.3: Solution of Example 6.1 with I0 = 30

Solution By using formulas (6.27) and (6.29), we obtain the particular integral πCk Dk 1 B+ cos(πDk t + Ek ). 2 α α2 + (πDk )2 K

Q(t) = Iˆ +

(6.31)

k=1

6.1.5

Optimality of a Linear Decision Rule

In Sect. 6.1.2, our emphasis was to explore the turnpike nature of the solution of the inventory model of Sect. 6.1.1. For this purpose, we made some asymptotic approximations when solving the state and adjoint diﬀerential equations under the assumption that the horizon is long. Here our focus is to solve the undiscounted version (i.e., ρ = 0) of the model exactly to ﬁnd its optimal feedback solution, and show that it is a linear decision rule as reported in the classical work of Holt et al. (1960).

6.1. Production-Inventory Systems

201

Since the two-point boundary value problem given by (6.7) and (6.8) is a linear system of diﬀerential equations, it is known via its fundamental solution matrix that λ can be expressed in terms of I in a linear way as follows: λ(t) = ψ(t) − s(t)I(t), (6.32) where ψ(t) and s(t) are continuously diﬀerentiable in t. Diﬀerentiating (6.32) with respect to t and substituting for I˙ and λ˙ from (6.7) and (6.8) with ρ = 0, respectively, we obtain ˙ + (Pˆ + ψ/c − S)s − hIˆ − ψ˙ = 0. I(h − s2 /c + s) Since the above relation must hold for any value of the initial inventory I0 , we must have ˆ s˙ = s2 /c − h and ψ˙ = (Pˆ + ψ/c − S)s − hI.

(6.33)

Also from λ(T ) = 0 in (6.8) and (6.32), we have 0 = ψ(T ) − s(T )I(T ), a relation that must hold regardless of the value of I(T ). Thus, we can conclude that s(T ) = 0 and ψ(T ) = 0. (6.34) Clearly, the solution of the diﬀerential equation given by (6.33) and (6.34) will give us the optimal control (6.5) in terms of S(t) and ψ(t). In particular, the diﬀerential equation s˙ = s2 /c − h, s(T ) = 0 is known as the Riccati equation, whose solution is given by ) √ h (T − t)). s(t) = hc tanh( c

(6.35)

(6.36)

Using (6.32) and (6.36) in (6.5), the optimal production rate P ∗ (t) is *) + ) h h ψ(t) ∗ tanh (T − t) I ∗ (t) + . (6.37) P (t) = Pˆ − c c c This says that the optimal production rate equals the production goal level Pˆ plus two adjustment terms. The ﬁrst term implies ceteris paribus that the higher the current inventory level, the lower the production rate is. Furthermore, this dependence is linear with the linear eﬀect decreasing as t increases, reaching zero at t = T. The second term depends on all the model parameters including the demand rate from time t to T.

202

6. Applications to Production and Inventory

Because of the linear dependence of the optimal production rate on the inventory level in (6.37), this rule is known as a linear decision rule as reported by Holt et al. (1960). More generally, this rule can be extended to linear quadratic problems as listed in Table 3.3(c). In Appendix D.4, we derive this rule for the problems given in Table 3.3(c), but without the forcing function d. Furthermore, the rule can be extended to a class of stochastic linear-quadratic problems that include the stochastic production planning problem treated in Sect. 12.2.

6.1.6

Analysis with a Nonnegative Production Constraint

Thus far in this chapter, we have ignored the production constraint P ≥ 0 and used (6.5) and (6.37) as the optimal decision rules. Here we will solve the production-inventory problem subject to P ≥ 0, and use (6.6) as the optimal production rule. For simplicity of analysis and exposition, we will assume also that S is a positive constant, T = ∞, and ρ > 0. These speciﬁcations make S˙ = 0, making the right hand side −αIˆ − ρ(Pˆ − S) a constant, a1 = b1 in (6.16), and a2 = 0 in (6.17). In view of its constant right-hand side, we can use Row (3) of Table A.2 to obtain its particular integral as Q=

ρ ˆ ˆ (P − S) + I, α2

(6.38)

which is a constant and thus Q˙ = 0. From (6.14) and (6.15), we now have b1 = I0 − Q = I0 − Iˆ − (ρ/α2 )(Pˆ − S) and b2 = Pˆ − S. ¯ = {(ρ/α2 )(Pˆ − S) + ¯ P¯ , λ} The turnpike is deﬁned by the triple {I, ˆ S, c(S − Pˆ )} formed from the turnpike expressions in (6.20), (6.21), I, and (6.22), respectively. Note that we could have obtained the turnpike levels directly by applying the conditions (3.108), which in this case are ¯˙ = 0, and P¯ = Pˆ + λ/c ¯ = S. I¯˙ = 0, λ

(6.39)

If I0 = Q, then the optimal solution stays on the turnpike. If I0 = Q, we must obtain the transient solution. It should be clear that the control in (6.25) may become negative, especially when the initial inventory is high. Let us complete the solution of the problem by considering three cases: I0 ≤ Q, Q < I0 ≤ Q − S/m1 , and I0 > Q − S/m1 .

6.1. Production-Inventory Systems

203

If I0 ≤ Q, then the control in (6.25) with b1 = I0 − Q0 is clearly positive. Thus, the optimal production rate is given by P ∗ (t) = m1 b1 em1 t + S = m1 (I0 − Q)em1 t + S ≥ 0.

(6.40)

Moreover, from the state in (6.25), we can obtain the corresponding I ∗ (t) as I ∗ (t) = (I0 − Q)em1 t + Q. (6.41) It is easy to see that I ∗ (t) increases monotonically to Q as t → ∞, as shown in Fig. 6.4. If Q < I0 ≤ Q − S/m1 , we can easily see from (6.40) that P ∗ (0) ≥ 0. Furthermore, P˙ ∗ (t) ≥ 0, and therefore the optimal production rate is once again given by (6.40). We also have I ∗ (t) as in (6.41) and conclude that I ∗ (t) → Q monotonically as t → ∞, as shown in Fig. 6.4. Finally, if I0 > Q − S/m1 , (6.40) would have a negative value for the initial production which is infeasible. By (6.6), P ∗ (0) = 0. We can now depict this situation in Fig. 6.4. The time tˆ shown in the ﬁgure is the time at which P ∗ (tˆ) = Pˆ + λ(tˆ)/c = 0. We already know from (6.40) that in the case when I0 = Q − S/m1 , P ∗ (0) = 0. This suggests that I ∗ (tˆ) = Q −

S . m1

(6.42)

For t ≤ tˆ, we have P ∗ (t) = 0 so that I˙∗ = −S, which gives I ∗ (t) = I0 − St, t ≤ tˆ.

(6.43)

As for the adjoint equation (6.7), we now need the boundary condition at tˆ. For this, we can use (6.4) to obtain λ(tˆ) = −cPˆ . Thus, the adjoint equation in the interval [0, tˆ ] is ˆ λ(tˆ) = −cPˆ . λ˙ = ρλ + h(I − I),

(6.44)

We can substitute I0 − St for I in Eq. (6.44) and solve for λ. Note that we can easily obtain tˆ as 1 S I0 − Q I0 − S tˆ = Q − + ⇒ tˆ = . m1 S m1

(6.45)

We can now specify the complete solution in the case when I0 > Q − S/m1 . With tˆ speciﬁed in (6.45), the solution is as follows.

204

6. Applications to Production and Inventory For 0 ≤ t ≤ tˆ : P ∗ (t) = 0, I ∗ (t) = I0 − St, and λ(t) is the solution of ˆ λ(tˆ) = −cPˆ . λ˙ = ρλ + h(I0 − St − I),

For t > tˆ : we replace I0 by Q − S/m1 and t by t − tˆ on the right hand side of (6.40) to obtain P ∗ (t) = −Sem1 (t−tˆ) . The same replacements in (6.41) gives us the corresponding I ∗ (t) = −(S/m1 )em1 t . Finally, λ(t) can be obtained by solving ˙λ = ρλ − h S em1 t + Iˆ , λ(tˆ) = −cPˆ . m1 We have thus solved the problem in every case of the initial condition I0 . These solutions are sketched in Fig. 6.4 for Iˆ = 8, Pˆ = 5, S = 6, h = 1, c = 4, and ρ = 0.1, for three diﬀerent values of I0 , namely, 25, 15, and 1. In Exercise 6.7, you are asked to solve the problem for these values and obtain Fig. 6.4.

Figure 6.4: Optimal production rate and inventory level with diﬀerent initial inventories

6.2

The Wheat Trading Model

Consider a ﬁrm that buys and sells wheat. The ﬁrm’s only assets are cash and wheat, and the price of wheat over time is known with certainty. The objective of this ﬁrm is to buy and sell wheat in order to maximize

6.2. The Wheat Trading Model

205

the total value of its assets at the horizon time T. The problem here is similar to the simple cash balance model of Sect. 5.1 except that there are nonlinear holding costs associated with storing wheat. An extension of this model to one having two control variables appears in Ijiri and Thompson (1972).

6.2.1

The Model

We introduce the following notation: T

= the horizon time,

x(t) = the cash balance in dollars at time t, y(t) = the wheat balance in bushels at time t, v(t) = the rate of purchase of wheat in bushels per unit time; a negative purchase means a sale, p(t) = the price of wheat in dollars per bushel at time t, r = the constant positive interest rate earned on the cash balance, h(y) = the cost of holding y bushels per unit time. In this section we permit x and y to go negative, meaning that borrowing money and short-selling wheat are both allowed. In the next section we disallow the short-selling of wheat. The state equations are: x˙ = rx − h(y) − pv, x(0) = x0 ,

(6.46)

y˙ = v, y(0) = y0 ,

(6.47)

and the control constraints are − V2 ≤ v(t) ≤ V1 ,

(6.48)

where V1 and V2 are nonnegative constants. The objective function is: max{J = x(T ) + p(T )y(T )}

(6.49)

subject to (6.46)–(6.48). Note that the problem is in the linear Mayer form.

206

6.2.2

6. Applications to Production and Inventory

Solution by the Maximum Principle

Introduce the adjoint variables λ1 and λ2 and deﬁne the Hamiltonian function H = λ1 [rx − h(y) − pv] + λ2 v. (6.50) The adjoint equations are: λ˙ 1 = −λ1 r, λ1 (T ) = 1, λ˙ 2 = h (y)λ1 , λ2 (T ) = p(T ).

(6.51) (6.52)

It is easy to solve (6.51) as λ1 (t) = er(T −t)

(6.53)

and (6.52) as λ2 (t) = p(T ) −

T

h (y(τ ))er(T −τ ) dτ .

(6.54)

t

The interpretation of λ1 (t) is that it is the future value (at time T ) of one dollar held as cash from t to T. The interpretation of λ2 (t) is the price at time T of a bushel of wheat less the total future value (at time T ) of the stream of storage costs incurred to store that bushel of wheat from t to T. From (6.50) the optimal control is v ∗ (t) = bang[−V2 , V1 ; λ2 (t) − λ1 (t)p(t)].

(6.55)

In Exercise 6.8 you are asked to provide the interpretation of this optimal policy. Equations (6.46), (6.47), (6.54), and (6.55) determine the two-point boundary value problem which usually requires a numerical solution procedure. In the next section we assume a special form for the storage function h(y) to be able to obtain a closed-form solution.

6.2.3

Solution of a Special Case

For this special case we assume h(y) = 12 |y|, r = 0, x(0) = 10, y(0) = 0, V1 = V2 = 1, T = 6, and ⎧ ⎪ ⎨ 3 for 0 ≤ t ≤ 3, p(t) = (6.56) ⎪ ⎩ 4 for 3 < t ≤ 6.

6.2. The Wheat Trading Model

207

We will apply the maximum principle (2.31) developed in Chap. 2 to this problem even though h(y) is not diﬀerentiable at y = 0. The answer can be obtained rigorously by using the maximum principle for models involving nondiﬀerentiable functions discussed, e.g., in Clarke (1989, Chapter 4) and Feichtinger and Hartl (1985b). For this case with r = 0, we have λ1 (t) = 1 for all t from (6.53) so that the TPBVP is 1 x˙ = − |y| − pv, x(0) = 10, 2 y˙ = v, y(0) = 0, 1 λ˙ 2 (t) = sgn(y), λ2 (6) = 4. 2

(6.57) (6.58) (6.59)

For this simple problem it is easy to guess a solution. From the fact that λ1 = 1, the optimal policy (6.55) reduces to v ∗ (t) = bang[−1, 1; λ2 (t) − p(t)].

(6.60)

Figure 6.5: The price trajectory (6.56) The graph of the price function is shown in Fig. 6.5. Since p(t) is increasing, short-selling is never optimal. Since the storage cost is 1/2 per unit per unit time and the wheat price jumps by 1 unit at t = 3, it never pays to store wheat for more than 2 time units. Because y(0) = 0, we have v ∗ (t) = 0 for 0 ≤ t ≤ 1. This obviously must be a singular

208

6. Applications to Production and Inventory

control. Suppose we start buying wheat at t∗ > 1. From (6.60) the rate of buying is 1; clearly buying will continue at this rate until t = 3, and not longer. In order to not lose money on the storage of wheat, it must be sold within 2 time units of its purchase. Clearly we should start selling at t = 3+ at the maximum rate of 1, and continue until a last sale time t∗∗ . In order to sell exactly all of the wheat purchased, we must have 3 − t∗ = t∗∗ − 3.

(6.61)

Thus, v ∗ (t) = 0 in the interval [t∗∗ , 6], which is also a singular control. With this policy, y(t) > 0 for all t ∈ (t∗ , t∗∗ ). From (6.59), λ˙ 2 = 1/2 in the interval (t∗ , t∗∗ ). In order to have a singular control in the interval [t∗∗ , 6], we must have λ2 (t) = 4 in that interval. Also, in order to have a singular control in [0, t∗ ], we must have λ2 (t) = 3 in that interval. Thus, λ2 (t∗∗ ) − λ2 (t∗ ) = 1, which with λ˙ 2 = 1/2 allows us to conclude that t∗∗ − t∗ = 2, and therefore t∗ = 2 and t∗∗ = 4. Thus ⎧ ⎪ ⎪ ⎪ 3, ⎪ ⎪ ⎨ λ2 (t) = 2 + t/2, ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 4,

(6.62)

from (6.59) and (6.60), 0 ≤ t ≤ 2, 2 ≤ t ≤ 4,

(6.63)

4 ≤ t ≤ 6.

We can now sketch graphs for λ2 (t), v ∗ (t), and y ∗ (t) as shown in Fig. 6.6. In Exercise 6.13 you are asked to show that these trajectories are optimal by verifying that the maximum principle necessary conditions hold and that they are also suﬃcient.

6.2.4

The Wheat Trading Model with No Short-Selling

We next consider the wheat trading problem in the case when shortselling is not permitted, i.e., we impose the state constraint y ≥ 0. Moreover, for simplicity in exposition we consider the following special case of Norstr¨om (1978). Speciﬁcally, we assume h(y) = y/2, r = 0, x(0) = 10, y(0) = 1, V1 = V2 = 1, T = 3, and ⎧ ⎪ ⎨ −2t + 7 for 0 ≤ t < 2, p(t) = (6.64) ⎪ ⎩ t + 1 for 2 ≤ t ≤ 3.

6.2. The Wheat Trading Model

209

The statement of the problem is: ⎧ ⎪ ⎪ max {J = x(3) + p(3)y(3) = x(3) + 4y(3)} ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ subject to ⎪ ⎪ ⎨ x˙ = − 12 y − pv, x(0) = 10, ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ y˙ = v, y(0) = 1, ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ v + 1 ≥ 0, 1 − v ≥ 0, y ≥ 0.

(6.65)

Buy

Sell

Figure 6.6: Adjoint variable, optimal policy and inventory in the wheat trading model

210

6. Applications to Production and Inventory

To solve this problem, we use the Lagrangian form of the indirect maximum principle given in (4.29). The Hamiltonian is H = λ1 (−y/2 − pv) + λ2 v.

(6.66)

The optimal control is v ∗ (t) = bang[−1, 1; λ2 (t) − λ1 (t)p(t)] when y > 0.

(6.67)

Whenever y = 0 we must impose y˙ = v ≥ 0 in order to insure that no short-selling occurs. Therefore, v ∗ (t) = bang[0, 1; λ2 (t) − λ1 (t)p(t)] when y = 0.

(6.68)

Next we form the Lagrangian L = H + μ1 (v + 1) + μ2 (1 − v) + ηv,

(6.69)

where μ1 , μ2 , and η satisfy the complementary slackness conditions: μ1 ≥ 0, μ1 (v + 1) = 0,

(6.70)

μ2 ≥ 0, μ2 (1 − v) = 0,

(6.71)

η ≥ 0,

ηy = 0.

(6.72)

Furthermore, the optimal trajectory must satisfy ∂L = λ2 − pλ1 + μ1 − μ2 + η = 0. ∂v

(6.73)

With r = 0, we get λ1 = 1 as before, and ∂L = 1/2, λ2 (3− ) = 4 + γ, λ˙ 2 = − ∂y

(6.74)

γ ≥ 0, γy(3) = 0.

(6.75)

with (3− )

= 4, and if we let tˆ denote the time Let us ﬁrst try γ = 0. Then λ2 of the last jump before the terminal time, then there is no jump in the interval (tˆ, 3). Then, from (6.74) we have λ2 (t) = t/2 + 5/2 for tˆ ≤ t < 3,

(6.76)

and the optimal control from (6.67) or (6.68) is v ∗ = 1, i.e., buy wheat at the maximum rate of 1, so long as λ2 (t) > p(t). Also, this will give

6.2. The Wheat Trading Model

211

y(3) > 0, so that (6.75) holds. Let us next ﬁnd the time tˆ of the last jump before the terminal time. Clearly, this value will not be larger than the time at which λ2 (t) = p(t). Thus, tˆ ≤ {t|t/2 + 5/2 = −2t + 7} = 1.8.

(6.77)

Since p(t) is decreasing at the start of the problem, it appears that selling at the maximum rate of 1, i.e., v ∗ = −1, should be optimal at the start. Since the beginning inventory is y(0) = 1, selling at the rate of 1 can continue only until t = 1, at which time the inventory y(1) becomes 0. Suppose that we do nothing, i.e., v ∗ (t) = 0 in the interval (1, 1.8]. Then, t = 1 is an entry time (see Sect. 4.2) and t = 1.8 is not an entry time, and tˆ = 1. Hence, according to the maximum principle (4.29), λ2 (t) is continuous at t = 1.8, and therefore λ2 (t) is given by (6.76) in the interval [1, 3), i.e., λ2 (t) = t/2 + 5/2 for 1 ≤ t < 3.

(6.78)

Using (6.73) with λ1 = 1 in the interval (1, 1.8] and v ∗ = 0 so that μ1 = μ2 = 0, we have λ2 − p + μ1 − μ2 + η = λ2 − p + η = 0, and consequently η(t) = p(t) − λ2 (t) = −5t/2 + 9/2, t ∈ (1, 1.8].

(6.79)

Since ht = 0, the jump condition in (4.29) for the Hamiltonian at τ = 1 reduces to H[x∗ (1), u∗ (1− ), λ(1− ), 1] = H[x∗ (1), u∗ (1+ ), λ(1+ ), 1]. From the deﬁnition of the Hamiltonian H in (6.66), we can rewrite the condition as λ1 (1− )[−y(1)/2 − p(1− )v ∗ (1− )] + λ2 (1− )v ∗ (1− ) = λ1 (1+ )[−y(1)/2 − p(1+ )v ∗ (1+ )] + λ2 (1+ )v ∗ (1+ ). Since λ1 (t) = 1 for all t, the above condition reduces to −p(1− )v ∗ (1− ) + λ2 (1− )v ∗ (1− ) = −p(1+ )v ∗ (1+ ) + λ2 (1+ )v ∗ (1+ ).

212

6. Applications to Production and Inventory

Substituting the values of p(1− ) = p(1+ ) = 5 from (6.64), λ2 (1+ ) = 3 from (6.78), and v ∗ (1+ ) = 0 and v ∗ (1− ) = −1 from the above discussion, we obtain − 5(−1) + λ2 (1− )(−1) = −5(0) + 3(0) = 0 ⇒ λ2 (1− ) = 5.

(6.80)

We can now use the jump condition in (4.29) on the adjoint variables to obtain λ2 (1− ) = λ2 (1+ ) + ζ(1) ⇒ ζ(1) = λ2 (1− ) − λ2 (1+ ) = 5 − 3 = 2 ≥ 0. It is important to note that in the interval [1, 1.8], the optimal control condition (6.68) holds, justifying our supposition that v ∗ = 0 in this interval. Furthermore, using (6.80) and (6.74), λ2 (t) = t/2 + 9/2 for t ∈ [0, 1),

(6.81)

and the optimal control condition (6.67) holds, justifying our supposition that v ∗ = −1 in this interval. Also, we can conclude that our guess γ = 0

Sell

Buy

Figure 6.7: Adjoint trajectory and optimal policy for the wheat trading model

6.3. Decision Horizons and Forecast Horizons

213

is correct. The graphs of λ2 (t), p(t), and v ∗ (t) are displayed in Fig. 6.7. To complete the solution of the problem, you are asked to determine the values of μ1 , μ2 , and η in these various intervals.

6.3

Decision Horizons and Forecast Horizons

In some dynamic problems it is possible to show that the optimal decisions during an initial positive time interval are either partially or wholly independent of the data from some future time onwards. In such cases, a forecast of the future data needs to be made only as far as that time to make optimal decisions in the initial time interval. The initial time interval is called the decision horizon and the time up to which data is required to make the optimal decisions during the decision horizon is called the forecast horizon; see Bes and Sethi (1988), Bensoussan et al. (1983), and Haurie and Sethi (1984) for details on these concepts. Whenever they exist, these horizons naturally decompose the problem into a series of smaller problems. If the optimal decisions during the decision horizon are completely independent of the data beyond the forecast horizon, then the latter is called a strong forecast horizon. If, on the other hand, some mild restrictions on the data after the forecast horizon are required in order to keep the optimal decisions during the decision horizon unaﬀected, then it is called a weak forecast horizon. In this section we demonstrate these concepts in the context of the wheat trading model of the previous section. In Sect. 6.3.1 we obtain a decision horizon for the model of Sect. 6.2.4 which is also a weak forecast horizon. In Sect. 6.3.2 we modify the wheat trading model by adding a warehousing constraint. For the new problem we obtain a decision horizon and a strong forecast horizon. See also Sethi and Thompson (1982), Rempala (1986) and Hartl (1986a, 1988a) for further research in the context of the wheat trading model. In what follows we obtain these horizons and verify them for some examples with diﬀerent forecast data. For more details and proofs in other situations including more general ones, see Modigliani and Hohn (1955), Lieber (1973), Pekelman (1974, 1975, 1979), Kleindorfer and Lieber (1979), Vanthienen (1975), Morton (1978), Lundin and Morton (1975), Rempala and Sethi (1988, 1992), Hartl (1989a), and Sethi (1990).

214

6.3.1

6. Applications to Production and Inventory

Horizons for the Wheat Trading Model with No Short-Selling

For the model of Sect. 6.2.4, we will demonstrate that t = 1 is a decision horizon as well as a weak forecast horizon. In Fig. 6.8 we have redrawn Fig. 6.7 with a new price trajectory in the time interval [1, 3]. Also in the ﬁgure, we have extended the initial λ2 trajectory and labeled it the price shield. Its signiﬁcance is that, as long as the new price trajectory in the interval [1, 3] stays below the price shield, the optimal solution in the interval [0, 1], which is the decision horizon, remains unchanged. That is, it is optimal to sell throughout the interval. The restriction that p(t) must stay below the price shield in [1, 3] is the reason that t = 1 is a weak forecast horizon. The optimality of the control shown in Fig. 6.8 can be concluded by obtaining the adjoint trajectory in the interval [1, 3] as a straight line with slope 1/2 and the terminal value λ2 (3− ) = p(3). This way of drawing the adjoint trajectory is correct as long as the corresponding policy does not violate the inventory constraint y(t) ≥ 0 in the interval [1, 3]. For example, this will be the case if the buy interval in Fig. 6.8 is shorter than the sell interval at the end. On the other hand, if the inventory constraint is violated, then the λ2 (t) trajectory may jump in the interval [1, 3), and it will be more complicated to obtain it. Nevertheless, the decision horizon and weak forecast horizon still occur at t = 1. Moreover, if we let T > 1 be any ﬁnite horizon and assume that p(t) in the interval [1, T ] is always below the price shield line of Fig. 6.8 extended to T, then the policy of selling at the maximum rate in the interval [0, 1] remains optimal.

6.3.2

Horizons for the Wheat Trading Model with No Short-Selling and a Warehousing Constraint

In order to give an example in which a strong forecast horizon occurs, we modify the example of Sect. 6.2.4 by adding the warehousing constraint y ≤ 1 or 1 − y ≥ 0, (6.82) changing the terminal time to T = 4, and deﬁning the price trajectory to be ⎧ ⎪ ⎨ −2t + 7 for t ∈ [0, 2), p(t) = (6.83) ⎪ ⎩ t + 1 for t ∈ [2, 4].

6.3. Decision Horizons and Forecast Horizons

215

ield Price Sh

Decision Horizon

Sell

Weak Forecast Horizon

Do Nothing

Buy

Sell

Figure 6.8: Decision horizon and optimal policy for the wheat trading model The Hamiltonian of the new problem is unchanged and is given in (6.66). Furthermore, λ1 = 1. The optimal control is deﬁned in three parts as: v ∗ (t) = bang[−1, 1; λ2 (t) − p(t)] when 0 < y < 1, ∗

v (t) = bang [0, 1; λ2 (t) − p(t)] when y = 0, ∗

v (t) = bang[−1, 0; λ2 (t) − p(t)] when y = 1.

(6.84) (6.85) (6.86)

Deﬁning a Lagrange multiplier η 1 for the derivative of (6.82), i.e., for −y˙ = −v ≥ 0, we form the Lagrangian L = H + μ1 (v + 1) + μ2 (1 − v) + ηv + η 1 (−v),

(6.87)

where μ1 , μ2 , and η satisfy (6.70)–(6.72) and η 1 satisﬁes η 1 ≥ 0, η 1 (1 − y) = 0, η˙ 1 ≤ 0.

(6.88)

Furthermore, the optimal trajectory must satisfy ∂L = λ2 − p + μ1 − μ2 + η − η 1 = 0. ∂v

(6.89)

216

6. Applications to Production and Inventory

As before, λ1 = 1 and λ2 satisﬁes λ˙2 = 1/2, λ2 (4− ) = p(4) + γ 1 − γ 2 = 5 + γ 1 − γ 2 ,

(6.90)

where γ 1 ≥ 0, γ 1 y(4) = 0, γ 2 ≥ 0, γ 2 (1 − y(4)) = 0.

(6.91)

Let us ﬁrst try γ 1 = γ 2 = 0. Let tˆ be the time of the last jump of the adjoint function λ2 (t) before the terminal time T = 4. Then, λ2 (t) = t/2 + 3 for tˆ ≤ t < 4.

(6.92)

The graph of (6.92) intersects the price trajectory at t = 8/5 as shown in Fig. 6.9. It also stays above the price trajectory in the interval [8/5, 4] so that, if there were no warehousing constraint (6.82), the optimal decision in this interval would be to buy at the maximum rate. However, with the constraint (6.82), this is not possible. Thus tˆ > 8/5, since λ2 will have a jump in the interval [8/5, 4].

Figure 6.9: Optimal policy and horizons for the wheat trading model with no short-selling and a warehouse constraint

6.3. Decision Horizons and Forecast Horizons

217

To ﬁnd the actual value of tˆ we must insert a line of slope 1/2 above the minimum price at t = 2 in such a way that its two intersection points with the price trajectory are exactly one time unit (the time required to ﬁll up the warehouse) apart. Thus using (6.83), tˆ must satisfy −2(tˆ − 1) + 7 + (1/2)(1) = tˆ + 1, which yields tˆ = 17/6. The rest of the analysis for determining λ2 including the jump conditions is similar to that given in Sect. 6.2.4. Thus, ⎧ ⎪ ⎪ ⎪ t/2 + 9/2 for t ∈ [0, 1), ⎪ ⎪ ⎨ (6.93) λ2 (t) = t/2 + 29/12 for t ∈ [1, 17/6), ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ t/2 + 3 for t ∈ [17/6, 4]. This makes γ 1 = γ 2 = 0 the correct guess. Given (6.93), the optimal policy is given by (6.84)–(6.86) and is shown in Fig. 6.9. To complete the maximum principle we must derive expressions for the Lagrange multipliers in the four intervals shown in Fig. 6.9. Interval [0, 1) : μ2 = η = η 1 = 0, μ1 = p − λ2 > 0; v ∗ = −1, 0 < y ∗ < 1. Interval [1, 11/6) : μ1 = μ2 = η 1 = 0, η = p − λ2 > 0, η˙ ≤ 0; v ∗ = 0, y ∗ = 0. Interval [11/16, 17/6) : μ1 = η = η 1 = 0, μ2 = λ2 − p > 0; v ∗ = 1, 0 < y ∗ < 1. Interval [17/6, 4] : μ1 = μ2 = η = 0, η 1 = λ2 − p > 0, η˙ 1 ≤ 0, γ 1 = γ 2 = 0; v ∗ = 0, y ∗ = 1. In Exercise 6.17 you are asked to solve another variant of this problem. For the example in Fig. 6.9 we have labeled t = 1 as a decision horizon and tˆ = 17/6 as a strong forecast horizon. By this we mean that the

218

6. Applications to Production and Inventory

optimal decision in [0, 1] continues to be to sell at the maximum rate regardless of the price trajectory p(t) for t > 17/6. Because tˆ = 17/6 is a strong forecast horizon, we can terminate the price shield at that time as shown in the ﬁgure. In order to illustrate the statements in the previous paragraph, we consider two examples of price changes after tˆ = 17/6. Example 6.3 Assume the price trajectory to be ⎧ ⎪ ⎪ ⎪ −2t + 7 for t ∈ [0, 2), ⎪ ⎪ ⎨ p(t) = t+1 for t ∈ [2, 17/6), ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 25t/7 − 44/7 for t ∈ [17/6, 4], which is sketched in Fig. 6.10. Note that the price trajectory up to time 17/6 is the same as before, and the price after time 17/6 goes above the extension of the price shield in Fig. 6.9.

Figure 6.10: Optimal policy and horizons for Example 6.3

6.3. Decision Horizons and Forecast Horizons

219

Solution The new λ2 trajectory is shown in Fig. 6.10, which is the same as before for t < 17/6, and after that it is λ2 (t) = t/2+6 for t ∈ [17/6, 4]. The optimal policy is as shown in Fig. 6.10, and as previously asserted, the optimal policy in [0,1) remains unchanged. In Exercise 6.17 you are asked to verify the maximum principle for the solution of Fig. 6.10. Example 6.4 Assume the price trajectory to be ⎧ ⎪ ⎪ ⎪ −2t + 7 for t ∈ [0, 2), ⎪ ⎪ ⎨ p(t) = t+1 for t ∈ [2, 17/6), ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ −t/2 + 21/4 for t ∈ [17/6, 4], which is sketched in Fig. 6.11.

Figure 6.11: Optimal policy and horizons for Example 6.4 Solution Again the price trajectory is the same up to time 17/6, but the price after time 17/6 is declining. This changes the optimal policy

220

6. Applications to Production and Inventory

in the time interval [1, 17/6), but the optimal policy will still be to sell in [0, 1). As in the beginning of the section, we solve (6.90) to obtain λ2 (t) = t/2+5/4 for tˆ1 ≤ t ≤ 4, where tˆ1 ≥ 1 is the time of the last jump which is to be determined. It is intuitively clear that some proﬁt can be made by buying and selling to take advantage of the price rise between t = 2 and t = 17/6. For this, the λ2 (t) trajectory must cross the price trajectory between times 2 and 17/6 as shown in Fig. 6.11, and the inventory y must go to 0 between times 17/6 and 4 so that λ2 can jump downward to satisfy the ending condition λ2 (4− ) = p(4) = 13/4. Since we must buy and sell equal amounts, the point of intersection of the λ2 trajectory with the rising price segment, i.e., tˆ1 − α, must be exactly in the middle of the two other intersection points, tˆ1 and tˆ1 − 2α, of λ2 with the two declining price trajectories. Thus, tˆ1 and α must satisfy: −2(tˆ1 − 2α) + 7 + α/2 = (tˆ1 − α) + 1, (tˆ1 − α) + 1 + α/2 = −tˆ1 /2 + 21/4. These can be solved to yield tˆ1 = 163/54 and α = 5/9. The times tˆ1 , tˆ1 − α, and tˆ1 − 2α are shown in Fig. 6.11. The λ2 trajectory is given by ⎧ ⎪ ⎪ ⎪ t/2 + 9/2 for t ∈ [0, 1), ⎪ ⎪ ⎨ λ2 (t) = t/2 + 241/108 for t ∈ [1, 163/54), ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ t/2 + 5/4 for t ∈ [163/54, 4]. Evaluation of the Lagrange multipliers and veriﬁcation of the maximum principle is similar to that for the case in Fig. 6.9. In Sect. 6.3 we have given several examples of decision horizons and weak and strong forecast horizons. In Sect. 6.3.1 we found a decision horizon which was also a weak forecast horizon, and it occurred exactly when y(t) = 0. We also introduced the idea of a price shield in that section. In Sect. 6.3.2 we imposed a warehousing constraint and obtained the same decision horizon and a strong forecast horizon, which occurred when y(t) = 1. Note that if we had solved the problem with T = 1, then y ∗ (1) = 0; and if we had solved the problem with T = 17/6, then y ∗ (1) = 0 and y ∗ (17/6) = 1. The latter problem has the smallest T such that both y ∗ = 0 and y ∗ = 1 occur for t > 0, given the price trajectory. This is one of the ways that time t = 17/6 can be found to be a forecast horizon

Exercises for Chapter 6

221

along with the decision horizon at time t = 1. There are other ways to ﬁnd strong forecast horizons. For a survey of the literature, see Chand et al. (2002). Exercises for Chapter 6 E 6.1 Verify the expressions for a1 and a2 given in (6.16) and (6.17). E 6.2 Verify (6.27). Note that ρ = 0 is assumed in Sect. 6.1.4. E 6.3 Verify (6.29). Again assume ρ = 0. E 6.4 Given the demand function S = t(t − 4)(t − 8)(t − 12)(t − 16) + 30, ρ = 0, Iˆ = 15, T = 16, and α = 1, obtain Q(t) from (6.27). E 6.5 Complete the solution of Example 6.2 in Sect. 6.1.4. E 6.6 For the model of Sect. 6.1.6, derive the turnpike triple by using the conditions in (6.39). E 6.7 Solve the production-inventory model of Sect. 6.1.6 for the parameter values listed on Fig. 6.4, and draw the ﬁgure using MATLAB or another suitable software. E 6.8 Give an intuitive interpretation of (6.55). E 6.9 Assume that there is a transaction cost cv 2 when v units of wheat are bought or sold in the model of Sect. 6.2.1. Derive the form of the optimal policy. E 6.10 In Exercise 6.9, assume T = 10, x(0) = 10, y(0) = 0, c = 1/18, h(y) = (1/2)y 2 , V1 = V2 = ∞, r = 0, and p(t) = 10 + t. Solve the resulting TPBVP to obtain the optimal control in closed form. E 6.11 Set up the two-point boundary value problem for Exercise 6.9 with c = 0.05, h(y) = (1/2)y 2 , and the remaining values of parameters as in the model of Sect. 6.2.3. E 6.12 Use Excel, as illustrated in Sect. 2.5, to solve the TPBVP of Exercise 6.11.

222

6. Applications to Production and Inventory

E 6.13 Show that the solution obtained for the problem in Sect. 6.2.3 satisﬁes the necessary conditions of the maximum principle. Conclude the optimality of the solution by showing that the maximum principle conditions are also suﬃcient. E 6.14 Re-solve the problem of Sect. 6.2.3 with V1 = 2 and V2 = 1. E 6.15 Compute the optimal trajectories for μ1 , μ2 , and η for the model in Sect. 6.2.4. E 6.16 Solve the model in Sect. 6.2.4 with each of the following conditions: (a) y(0) = 2. (b)

T = 10 and p(t) = 2t − 2 for 3 ≤ t ≤ 10.

E 6.17 Verify that the solutions shown in Figs. 6.10 and 6.11 satisfy the maximum principle. E 6.18 Re-solve the model of Sect. 6.3.2 with y(0) = 1/2 and with the warehousing constraint y ≤ 1/2 in place of (6.82). E 6.19 Solve and interpret the following production planning problem with linear inventory holding costs: ⎧ T c 2 ⎪ ⎪ P −[hI + ]dt max J = ⎪ ⎪ ⎪ 2 0 ⎪ ⎪ ⎪ ⎪ ⎨ subject to (6.94) ⎪ ⎪ 2 ⎪ I˙ = P, I(0) = 0, I(T ) = B; 0 < B < hT /2c, ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ P ≥ 0 and I ≥ 0. ˙ = P (t)−S(t), E 6.20 Re-solve Exercise 6.19 with the state equation I(t) where I(0) = I0 ≥ 0 and I(T ) is not ﬁxed. Assume the demand S(t) to be continuous in t and non-negative. Keep the state constraint I ≥ 0, but drop the production constraint P ≥ 0 for simplicity. For speciﬁcity, you may assume S = − sin πt + C with the constant C ≥ 1 and T = 4. (Note that negative production can and will occur when initial inventory I0 is too large. Speciﬁcally, how large is too large depends on the parameters of the problem.)

Exercises for Chapter 6

223

˙ = P (t) − S, E 6.21 Re-solve Exercise 6.19 with the state equation I(t) where S > 0 and h > 0 are constants, I(0) = I0 > cS 2 /2h, and I(T ) is not ﬁxed. Assume that T is suﬃciently large. Also, graph the optimal P ∗ (t) and I ∗ (t), t ∈ [0, T ].

Chapter 7

Applications to Marketing Over the years, a number of applications of optimal control theory have been made to the ﬁeld of marketing. Many of these applications deal with the problem of ﬁnding or characterizing the optimal advertising rate over time. Others deal with the problem of determining the optimal price and quality over time, in addition to or without advertising. The reader is referred to Sethi (1977a) and Feichtinger et al. (1994a) for comprehensive reviews on dynamic optimal control problems in advertising and related problems. In this chapter we discuss optimal advertising policies for two of the well-known models called the Nerlove-Arrow model and the Vidale-Wolfe model. To describe the speciﬁc problems under consideration, let us assume that a ﬁrm has some way of knowing or estimating the dynamics of sales and advertising. Such knowledge is expressed in terms of a diﬀerential equation with either goodwill or the rate of sales as the state variable and the rate of advertising expenditures as the control variable. We assume that the ﬁrm wishes to maximize an objective function (the criterion function) which reﬂects its proﬁt motives expressed in terms of sales and advertising rates. The optimal control problem is to ﬁnd an advertising policy which maximizes the ﬁrm’s objective function. The plan of this chapter is as follows. Section 7.1 will cover the Nerlove-Arrow model as well as a nonlinear extension of it. Section 7.2 deals with the Vidale-Wolfe advertising model and its detailed analysis using Green’s theorem in conjunction with the maximum principle. The switching-point analysis for this problem is a good example of the reverse-time construction technique used earlier in Chaps. 4 and 5. Ex© Springer Nature Switzerland AG 2019 S. P. Sethi, Optimal Control Theory, https://doi.org/10.1007/978-3-319-98237-3 7

225

226

7. Applications to Marketing

tensions of these models to multi-state problems are treated in Turner and Neuman (1976) and Srinivasan (1976).

7.1

The Nerlove-Arrow Advertising Model

The belief that advertising expenditures by a ﬁrm aﬀect its present and future sales, and hence its present and future net revenues, has led a number of economists including Nerlove and Arrow (1962) to treat advertising as an investment in building up some sort of advertising capital, usually called goodwill. Furthermore, the stock of goodwill depreciates over time. Vidale and Wolfe (1957), Palda (1964), and others present empirical evidence that the eﬀects of advertising linger but diminish over time. Goodwill may be created by adding new customers or by altering the tastes and preferences of consumers and thus changing the demand function for the ﬁrm’s product. Goodwill depreciates over time because consumers “drift” to other brands as a result of advertising by competing ﬁrms and the introduction of new products and/or new brands, etc.

7.1.1

The Model

Let G(t) ≥ 0 denote the stock of goodwill at time t. The price of (or cost of producing) one unit of goodwill is one dollar so that a dollar spent on current advertising increases goodwill by one unit. It is assumed that the stock of goodwill depreciates over time at a constant proportional rate δ, so that G˙ = u − δG, G(0) = G0 , (7.1) where u = u(t) ≥ 0 is the advertising eﬀort at time t measured in dollars per unit time. In economic terms, Eq. (7.1) states that the net investment in goodwill is the diﬀerence between gross investment u(t) and depreciation δG(t). To formulate the optimal control problem for a monopolistic ﬁrm, assume that the rate of sales S(t) depends on the stock of goodwill G(t), the price p(t), and other exogenous factors Z(t), such as consumer income, population size, etc. Thus, S = S(p, G; Z).

(7.2)

7.1. The Nerlove-Arrow Advertising Model

227

Assuming the rate of total production cost is c(S), we can write the total revenue net of production cost as R(p, G; Z) = pS(p, G; Z) − c(S(p, G; Z)).

(7.3)

The revenue net of advertising expenditure is therefore R(p, G; Z) − u. We assume that the ﬁrm wants to maximize the present value of net revenue streams discounted at a ﬁxed rate ρ, i.e., ∞ −ρt e [R(p, G; Z) − u] dt (7.4) J= max u≥0,p≥0

0

subject to (7.1). Since the only place that p occurs is in the integrand, we can maximize J by ﬁrst maximizing R with respect to price p while holding G ﬁxed, and then maximize the result with respect to u. Thus, ∂R ∂S ∂S =S+p − c (S) = 0, ∂p ∂p ∂p

(7.5)

which implicitly gives the optimal price p∗ (t) = p(G(t); Z(t)). Deﬁning η = −(p/S)(∂S/∂p) as the elasticity of demand with respect to price, we can rewrite condition (7.5) as p∗ =

ηc (S) , η−1

(7.6)

which is the usual price formula for a monopolist, known sometimes as the Amoroso-Robinson relation. You are asked to derive this relation in Exercise 7.2. In words, the relation means that the marginal revenue (η −1)p/η must equal the marginal cost c (S). See, e.g., Cohen and Cyert (1965, p. 189). Deﬁning Π(G; Z) = R(p∗ , G; Z), the objective function in (7.4) can be rewritten as ∞ −ρt e [Π(G; Z) − u] dt . max J = u≥0

0

For convenience, we assume Z to be a given constant. Thus, we can deﬁne π(G) = Π(G; Z) and restate the optimal control problem which we have just formulated: ⎧ ∞ ⎪ −ρt ⎪ max J = e [π(G) − u] dt ⎪ ⎪ ⎪ 0 ⎨ u≥0 (7.7) subject to ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ G˙ = u − δG, G(0) = G . 0

228

7. Applications to Marketing

Furthermore, it is reasonable to assume the functions introduced in (7.2) and (7.3) to satisfy conditions so that π(G) is increasing and concave in goodwill G. More speciﬁcally, we assume that π (G) ≥ 0 and π (G) < 0.

7.1.2

Solution by the Maximum Principle

While Nerlove and Arrow (1962) used calculus of variations, we use Pontryagin’s maximum principle to derive their results. We form the current-value Hamiltonian H = π(G) − u + λ[u − δG]

(7.8)

with the current-value adjoint variable λ satisfying the diﬀerential equation ∂H dπ λ˙ = ρλ − = (ρ + δ)λ − ∂G dG

(7.9)

and the condition that lim e−ρt λ(t) = 0.

t→+∞

(7.10)

Recall from Sect. 3.6 that this limit condition is only a suﬃcient condition. The adjoint variable λ(t) is the shadow price associated with the goodwill at time t. Thus, the Hamiltonian in (7.8) can be interpreted as the dynamic proﬁt rate which consists of two terms: (1) the current net proﬁt rate (π(G) − u) and (2) the value λG˙ = λ[u − δG] of the goodwill rate G˙ created by advertising at rate u. Also, Eq. (7.9) corresponds to the usual equilibrium relation for investment in capital goods; see Arrow and Kurz (1970) and Jacquemin (1973). It states that the marginal opportunity cost λ(ρ + δ)dt of investment in goodwill, by spending on advertising, should equal the sum of the marginal proﬁt π (G)dt from the ˙ increased goodwill due to that investment and the capital gain dλ := λdt on the unit price of goodwill. We use (3.108) to obtain the optimal long-run stationary equilibrium ¯ That is, we obtain λ = λ ¯ = 1 from (7.8) by using ¯ u or turnpike {G, ¯, λ}. ¯ = 1 and λ˙ = 0 in (7.9) to obtain ∂H/∂u = 0. We then set λ = λ

7.1. The Nerlove-Arrow Advertising Model

229

¯ = ρ + δ. π (G)

(7.11) ¯ we In order to obtain a strictly positive equilibrium goodwill level G, may assume π (0) > ρ + δ and π (∞) < ρ + δ. Before proceeding further to obtain the optimal advertising policy, let ¯ obtained by Jacquemin us relate (7.11) to the equilibrium condition for G (1973). For this we deﬁne β = (G/S)(∂S/∂G) as the elasticity of demand with respect to goodwill. We can now use (7.3), (7.5), (7.6), and (7.9) ¯ = 1 to derive, as you will in Exercise 7.3, with λ˙ = 0 and λ ¯ G β = . (7.12) pS η(ρ + δ) The interpretation of (7.12) is that in the equilibrium, the ratio of goodwill to sales revenue pS is directly proportional to the goodwill elasticity, inversely proportional to the price elasticity, and inversely proportional to the cost of maintaining goodwill given by the marginal opportunity cost λ(ρ + δ) of investment in goodwill. ¯ is that the optimal policy is to go to G ¯ as fast The property of G ¯ ¯ by as possible. If G0 < G, it is optimal to jump instantaneously to G ∗ ¯ ¯ = δG applying an appropriate impulse at t = 0 and then set u (t) = u ∗ ¯ the optimal control u (t) = 0 until the stock of for t > 0. If G0 > G, ¯ at which time the control switches goodwill depreciates to the level G, ∗ ¯ ¯ of goodwill. to u (t) = δ G and stays at this level to maintain the level G This optimal policy is graphed in Fig. 7.1 for these two diﬀerent initial conditions. Of course, if we had imposed an upperbound M > 0 on the control ¯ we would use u∗ (t) = M until so that 0 ≤ u ≤ M, then for G0 < G, ¯ the goodwill stock reaches G and switch to u∗ (t) = u ¯ thereafter. This is shown as the dotted curve in Fig. 7.1. Problem (7.7) is formulated with the assumption that a dollar spent on current advertising increases goodwill by one unit. Suppose, instead, that we need to spend m dollars on current advertising to increase goodwill by one unit. We can then deﬁne u as advertising eﬀort costing the ﬁrm mu dollars, and reformulate problem (7.7) by replacing [π(G) − u]

230

7. Applications to Marketing

Case :

Figure 7.1: Optimal policies in the Nerlove-Arrow model in its integrand by [π(G) − mu]. In Exercise 7.4, you are asked to solve problem (7.7) with its objective function and the control constraint replaced by ∞

max

0≤u≤M

J= 0

e−ρt [π(G) − mu]dt ,

(7.13)

and show that the equilibrium goodwill level formula (7.11) changes to ¯ = (ρ + δ)m. π (G)

(7.14)

¯ thus deﬁned, the optimal solution is as shown in Fig. 7.1 with With G ¯ the dotted curve representing the solution in Case 2: G0 < G. ¯ For a time-dependent Z, however, G(t) = G(Z(t)) will be a func¯ tion of time. To maintain this level of G(t), the required control is ¯ + G(t). ¯˙ ¯ u ¯(t) = δ G(t) If G(t) is decreasing suﬃciently fast, then u ¯(t) may become negative and thus infeasible. If u ¯(t) ≥ 0 for all t, then the optimal policy is as before. However, suppose u ¯(t) is infeasible in the interval ¯(t) [t1 , t2 ] shown in Fig. 7.2. In such a case, it is feasible to set u(t) = u for t ≤ t1 ; at t = t1 (which is point A in Fig. 7.2) we can no longer stay on the turnpike and must set u(t) = 0 until we hit the turnpike again (at point B in Fig. 7.2). However, such a policy is not necessarily optimal. For instance, suppose we leave the turnpike at point C anticipating the infeasibility at point A. The new path CDEB may be better than the old path CAB. Roughly the reason this may happen is that path CDEB is “nearer” to the turnpike than CAB. The picture in Fig. 7.2 illustrates

7.1. The Nerlove-Arrow Advertising Model

231

such a case. The optimal policy is the one that is “nearest” to the turnpike. This discussion will become clearer in Sect. 7.2.2, when a similar situation arises in connection with the Vidale-Wolfe model. For further details; see Sethi (1977b) and Breakwell (1968). The Nerlove-Arrow model is an example involving bang-bang and impulse controls followed by a singular control, which arises in a class of optimal control problems of Model Type (b) in Table 3.3 that are linear in control. Nonlinear extensions of the Nerlove-Arrow model have been oﬀered in the literature. These amount to making the objective function nonlinear in advertising. Gould (1970) extended the model by assuming a

Figure 7.2: A case of a time-dependent turnpike and the nature of optimal control

convex cost of advertising eﬀort, which implies a marginally diminishing eﬀect of advertising expenditures. Jacquemin (1973) assumed that the current demand function S in (7.2) also depends explicitly on the current advertising eﬀort u. In Exercise 11.6, you are asked to analyze Gould’s extension via the phase diagram analysis introduced in Chap. 11. The analysis of Jacquemin’s extension is similar.

232

7.1.3

7. Applications to Marketing

Convex Advertising Cost and Relaxed Controls

Another nonlinear extension of the Nerlove-Arrow model would involve a concave advertising cost resulting from quantity discounts that may be available in the purchase of advertising. Such an extension results in an optimal control problem with a proﬁt rate that is convex in advertising, and this has a possibility of rendering the problem without an optimal solution within the class of admissible controls discussed thus far. What is then required is an enlargement of the class to include what are known as relaxed controls. To introduce such controls, we formulate and solve a convex optimal control problem involving the Nerlove-Arrow model. Let c(u) be a strictly concave advertising cost function with c(0) = 0, c (u) > 0 and c (u) < 0 for 0 ≤ u ≤ M, where M > 0 denotes an upperbound on the advertising rate. Let us also consider T > 0 to be the ﬁxed terminal time. Then, our problem is the following modiﬁcation of problem (7.7): ⎧ T ⎪ −ρt ⎪ e [π(G) − c(u)]dt max J1 = ⎪ ⎪ ⎪ 0 ⎨ 0≤u≤M subject to ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ G˙ = u − δG, G(0) = G . 0

(7.15)

Note that with concave c(u), the proﬁt rate π(G) − c(u) is convex in u. Thus, its maximum over u would occur at the boundary 0 or M of the set [0, M ]. It should be clear that if we replace c(u) by the linear function mu with m = c(M )/M, then π(G) − c(u) < π(G) − mu, u ∈ (0, M ).

(7.16)

This means that if problem (7.15) with mu in place of c(u), i.e., the problem ⎧ T ⎪ −ρt ⎪ max J2 = e [π(G) − mu] dt ⎪ ⎪ ⎪ 0 ⎨ 0≤u≤M (7.17) subject to ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ G˙ = u − δG, G(0) = G 0 has only the bang-bang solution, then the solution of problem (7.17) would also be the solution of the convex problem (7.15). Given the

7.1. The Nerlove-Arrow Advertising Model

233

similarity of problem (7.17) to problem (7.7), we can see that for a sufﬁciently small value of T, the solution of (7.17) will be bang-bang only, and therefore, it will also solve (7.15). However, if T is large or inﬁnity, then the solution of (7.17) will have a singular portion, and it will not solve (7.17). In particular, let us consider problems (7.15) and (7.17) when T = ∞ ¯ Note that problem (7.17) is the same as the problem in and G0 < G. ¯ given Exercise 7.4, and its optimal solution is as shown in Fig. 7.1 with G by (7.14) and the optimal trajectory given by the dotted line followed by the solid horizontal line representing the singular part of the solution. Let u∗2 denote the optimal control of problem (7.17). Since the singular control is in the open interval (0, M ), then in view of (7.16), J1 (u∗2 ) < J2 (u∗2 ).

(7.18)

Thus, for suﬃciently small ε1 > 0 and ε2 > 0, we can “chatter” between ¯ + ε1 ) and G2 = (G ¯ − ε2 ) by using controls M and 0 alternately, G1 = (G as shown in Fig. 7.3, to obtain a near-optimal control of problem (7.15). Clearly, in the limit as ε1 and ε2 go to 0, the objective function of problem (7.15) will converge to J2 (u∗2 ).

Figure 7.3: A near-optimal control of problem (7.15)

234

7. Applications to Marketing

This is an intuitive explanation that there does not exist an optimal control of problem (7.15) in the class of controls discussed thus far. However, when the class of controls is enlarged to include relaxed or generalized controls, which are the limits of the approximating controls like the ones constructed above, we can recover the existence of an optimal solution; see Gamkrelidze (1978) and Lou (2007) for details. The manner in which the theory of relaxed controls manifests itself for our problem is to provide a probability measure on the boundary values {0, M }. Thus, let v be the probability that control M is used, so that the probability of using control 0 is (1 − v). With this, we transform problem (7.15) with T = ∞ as follows: ∞ ⎧ ⎪ −ρt ⎪ max = e [π(G) − vc(M )] dt J 3 ⎪ ⎪ ⎪ 0 ⎨ v∈[0,1] (7.19) subject to ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ G˙ = vM − δG, G(0) = G . 0 We can now use the maximum principle to solve problem (7.19). Thus, the Hamiltonian H = πG − vc(M ) + λ(vM − δG) with the adjoint equation as deﬁned by (7.9) and (7.10). The optimal control is given by v ∗ = bang[0, 1; λM − c(M )]. The singular control is given by ¯ = m, π (G) ¯ = (ρ + δ)m, v¯ = δ G/M. ¯ λ

(7.20) (7.21)

The way we interpret this control is by use of a biased coin with the probability of heads being v¯. We ﬂip this coin inﬁnitely fast, and use the maximum control M when heads comes up and the minimum control 0 when tails comes up. Because the control will chatter inﬁnitely fast according to the outcome of the coin tosses, such a control is also referred to as a chattering control. While such a chattering control cannot be implemented, it can be arbitrarily approximated by using alternately u∗ = M for τ v¯ periods and u∗ = 0 for τ (1 − v¯) periods for a small τ > 0. With reference to Fig. 7.3 and with G1 and G2 to be determined for the given τ , this approximate

7.2. The Vidale-Wolfe Advertising Model

235

policy of rapidly switching the control between M and 0 can be said to begin at time t1 , when the goodwill reaches G2 . After that goodwill goes up to G1 and then back down to G2 , and so on. The values of G1 and G2 , corresponding to the given τ , are speciﬁed in Exercise 7.8, and you are asked to derive them. In marketing parlance, advertising rates that alternate between maximum and zero are known as a pulsing policy. While there are other reasons for pulsing that are known in the advertising literature, the convex cost of advertising is one of them; see Feinberg (1992, 2001) for details. Another example of relaxed control appears in Haruvy et al. (2003) in connection with open-source software development. This is given as Exercise 7.14.

7.2

The Vidale-Wolfe Advertising Model

We now present the analysis of the Vidale-Wolfe advertising model which, in contrast to the Nerlove-Arrow model, does not make use of the idea of advertising goodwill; see Vidale and Wolfe (1957) and Sethi (1973a, 1974b). Instead the model exploits the closely related notion that the eﬀect of advertising tends to persist, but diminishes over subsequent time periods. This carryover eﬀect is modeled explicitly by means of a diﬀerential equation that gives the relationship between sales and advertising. Vidale and Wolfe argued that changes in the rate of sales of a product depend on two eﬀects: the action of advertising (via the response constant a) on the unsold portion of the market and the loss of sales (via the decay constant b) from the sold portion of the market. Let M (t), known as the saturation level or market potential, denote the maximum potential rate of sales at time t. Let S(t) be the actual rate of sales at time t. Then, the Vidale-Wolfe model for a monopolistic ﬁrm can be stated as S ) − bS. (7.22) S˙ = au(1 − M The important feature of this equation, which distinguishes it from the Nerlove-Arrow equation (7.1), is the idea of the ﬁnite saturation level M. The Vidale-Wolfe model exhibits diminishing returns to the level of advertising as a direct consequence of this saturation phenomenon. Note that when M is inﬁnitely large, the saturation phenomenon disappears, reducing (7.22) to the equation (with constant returns to advertising) similar to the Nerlove-Arrow equation (7.1). Nerlove and Arrow, on the

236

7. Applications to Marketing

other hand, include the idea of diminishing returns to advertising in their model by making the sales S in (7.2) a concave function of goodwill. Vidale and Wolfe based their model on the results of several experimental studies of advertising eﬀectiveness, which are described in Vidale and Wolfe (1957).

7.2.1

Optimal Control Formulation for the Vidale-Wolfe Model

Whereas Vidale and Wolfe oﬀered their model primarily as a description of actual market phenomena represented by cases which they had observed, we obtain the optimal advertising expenditures for the model in order to maximize a certain objective function over the horizon T, while also attaining a terminal sales target; see Sethi (1973a). For this, it is convenient to transform (7.22) by making the change of variable x=

S . M

(7.23)

Thus, x represents the market share (or more precisely, the rate of sales expressed as a fraction of the saturation level M ). Furthermore, we deﬁne M˙ a , δ =b+ . (7.24) r= M M Now we can rewrite (7.22) as x˙ = ru(1 − x) − δx, x(0) = x.

(7.25)

From now on we assume M, and hence δ and r, to be positive constants. It would not be diﬃcult to extend the analysis when M depends on t, but we do not carry it out here. In Exercise 7.35 you are asked to partially analyze the time-dependent case. To deﬁne the optimal control problem arising from the Vidale-Wolfe model, we let π denote the maximum sales revenue corresponding to x = 1, with πx denoting the revenue function for x ∈ [0, 1]. Also let Q be the maximum allowable rate of advertising expenditure and let ρ denote the continuous discount rate. With these deﬁnitions the optimal control

7.2. The Vidale-Wolfe Advertising Model problem can be stated as follows: ⎧ T ⎪ −ρt ⎪ e (πx − u)dt ⎪ max J = ⎪ ⎪ 0 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ subject to ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ x˙ = ru(1 − x) − δx, x(0) = x0 , ⎪ ⎪ ⎨ the terminal state constraint ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ x(T ) = xT , ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ and the control constraint ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 0 ≤ u ≤ Q.

237

(7.26)

Here Q can be ﬁnite or inﬁnite and the target market share xT is in [0, 1]. Note that the problem is a ﬁxed-end-point problem. It is obvious that the requirement 0 ≤ x ≤ 1 holds without being imposed, where x0 ∈ [0, 1] is the initial market share. It is possible to solve this problem by an application of the maximum principle; see Exercise 7.18. However, we will use instead a method based on Green’s theorem which does not make use of the maximum principle. This method provides a convenient procedure for solving ﬁxed-end-point problems having one state variable and one control variable, and where the control variable appears linearly in both the state equation and the objective function; see Miele (1962) and Sethi (1977b). Problem (7.26) has these properties, and therefore it is also a good example with which to illustrate the method. For the application of Green’s theorem we require that Q be large. In particular we can let Q = ∞.

7.2.2

Solution Using Green’s Theorem When Q Is Large

In this section we will solve the ﬁxed-end-point problem starting with x0 and ending with xT , under the assumption that Q is either unbounded or very large. The places where these assumptions are needed will be indicated. To make use of Green’s theorem, it is convenient to consider times τ and θ, where 0 ≤ τ < θ ≤ T, and the problem: θ −ρt max J(τ , θ) = e (πx − u)dt (7.27) τ

238

7. Applications to Marketing

subject to x˙ = ru(1 − x) − δx, x(τ ) = A, x(θ) = B,

(7.28)

0 ≤ u ≤ Q.

(7.29)

To change the objective function in (7.27) into a line integral along any feasible arc Γ1 from (τ , A) to (θ, B) in (t, x)-space as shown in Fig. 7.4, we multiply (7.28) by dt and obtain the formal relation udt =

dx + δxdt , r(1 − x)

which we substitute into the objective function (7.27). Thus,

J Γ1 =

Γ1

1 δx −ρt −ρt e dt − e dx . πx − r(1 − x) r(1 − x)

Figure 7.4: Feasible arcs in (t, x)-space Consider another feasible arc Γ2 from (τ , A) to (θ, B) lying above Γ1 as shown in Fig. 7.4. Let Γ = Γ1 − Γ2 , where Γ is a simple closed curve traversed in the counter-clockwise direction. That is, Γ goes along Γ1 in the direction of its arrow and along Γ2 in the direction opposite to its arrow. We now have JΓ = JΓ1 −Γ2 = JΓ1 − JΓ2 .

(7.30)

7.2. The Vidale-Wolfe Advertising Model

239

Since Γ is a simple closed curve, we can use Green’s theorem to express JΓ as an area integral over the region R enclosed by Γ. Thus, treating x and t as independent variables, we can write

, 1 δx −ρt −ρt e dt − e dx JΓ = πx − r(1 − x) r(1 − x) Γ

∂ −e−ρt ∂ δx −ρt )e = dtdx − (πx − ∂x r(1 − x) R ∂t r(1 − x)

−ρt δ e ρ = + − πr dtdx. (7.31) 2 (1 − x) (1 − x) r R Denote the term in brackets of the integrand of (7.31) by I(x) =

ρ δ − πr. + 2 (1 − x) (1 − x)

(7.32)

Note that the sign of the integrand is the same as the sign of I(x). Lemma 7.1 (Comparison Lemma) Let Γ1 and Γ2 be the lower and upper feasible arcs as shown in Fig. 7.4. If I(x) ≥ 0 for all (x, t) ∈ R, then the lower arc Γ1 is at least as proﬁtable as the upper arc Γ2 . Analogously, if I(x) ≤ 0 for all (x, t) ∈ R, then Γ2 is at least as proﬁtable as Γ1 . Proof If I(x) ≥ 0 for all (x, t) ∈ R, then JΓ ≥ 0 from (7.31) and (7.32). Hence from (7.30), JΓ1 ≥ JΓ2 . The proof of the other statement is similar. 2 To make use of this lemma to ﬁnd the optimal control for the problem stated in (7.26), we need to ﬁnd regions where I(x) is positive and where it is negative. For this, note ﬁrst that I(x) is an increasing function of x in [0, 1]. Solving I(x) = 0 will give that value of x, above which I(x) is positive and below which I(x) is negative. Since I(x) is quadratic in 1/(1 − x), we can use the quadratic formula (see Exercise 7.16) to get x=1−

2δ # . −ρ ± ρ2 + 4πrδ

(7.33)

To keep x in the interval [0, 1], we must choose the positive sign before the radical. The optimal x must be nonnegative so we have 2δ s # x = max 1 − ,0 , (7.34) −ρ + ρ2 + 4πrδ

240

7. Applications to Marketing

where the superscript s is used because this will turn out to be a singular trajectory. Since xs is nonnegative, the control us =

δxs r(1 − xs )

(7.35)

corresponding to (7.34) will always be nonnegative. Also since Q is assumed to be large, us will always be feasible. Moreover, in Exercise 7.17, you will be asked to show that xs = 0 and us = 0 if, and only if, πr ≤ δ + ρ. We now have enough machinery to obtain the optimal solution for (7.26) when Q is assumed to be suﬃciently large, i.e., Q ≥ us , where us is given in (7.35). We state these in the form of two theorems: Theorem 7.1 refers to the case in which T is large; Theorem 7.2 refers to the case in which T is small. To deﬁne these concepts, let t1 be the shortest time to go from x0 to xs and similarly let t2 be the shortest time to go from

Figure 7.5: Optimal trajectory for Case 1: x0 ≤ xs and xT ≤ xs xs to xT . Then, we say T is large if T ≥ t1 + t2 ; otherwise T is small. Figures 7.5, 7.6, 7.7, and 7.8 show cases for which T is large, while Figs. 7.10 and 7.11 show cases for which T is small. In Exercise 7.21 you are asked to determine whether T is large or small in speciﬁc cases. In the statements of the theorems we will assume that x0 and xT are such that xT is reachable from x0 . In Exercise 7.15 you are asked to ﬁnd the reachable set for any given initial condition x0 . In Figures 7.5, 7.6, 7.7, and 7.8, the quantities t1 and t2 are case dependent and not necessarily the same; see Exercise 7.20.

7.2. The Vidale-Wolfe Advertising Model

241

Theorem 7.1 Let T be large and let xT be reachable from x0 . For the Cases 1–4 of inequalities relating x0 and xT to xs , the optimal trajectories are given in Figures 7.5, 7.6, 7.7, and 7.8, respectively. Proof We give details for Case 1 only. The proofs for the other cases are similar. Figure 7.9 shows the optimal trajectory for Fig. 7.5 together with an arbitrarily chosen feasible trajectory, shown dotted. It should be clear that the dotted trajectory cannot cross the arc x0 to C, since u = Q on that arc. Similarly the dotted trajectory cannot cross the arc G to xT , because u = 0 on that arc. We subdivide the interval [0, T ] into subintervals over which the dotted arc is either above, below, or identical to the solid arc. In Fig. 7.9

Figure 7.6: Optimal trajectory for Case 2: x0 < xs and xT > xs

Figure 7.7: Optimal trajectory for Case 3: x0 > xs and xT < xs

242

7. Applications to Marketing

Figure 7.8: Optimal trajectory for Case 4: x0 > xs and xT > xs these subintervals are [0, d], [d, e], [e, f ], and [f, T ]. Because I(x) is positive for x > xs and I(x) is negative for x < xs , the regions enclosed by the two trajectories have been marked with a + or − sign depending on whether I(x) is positive or negative on the regions, respectively. By Lemma 7.1, the solid arc is better than the dotted arc in the subintervals [0, d], [d, e], and [f, T ]; in interval [e, f ], they have identical values. Hence the dotted trajectory is inferior to the solid trajectory. This proof can be extended to any (countable) number of crossings of the trajectories; see Sethi (1977b). 2 Figures 7.5, 7.6, 7.7, and 7.8 are drawn for the situation when T > t1 + t2 . In Exercise 7.25, you are asked to consider the case when T = t1 + t2 . The following theorem deals with the case when T < t1 + t2 . Theorem 7.2 Let T be small, i.e., T < t1 + t2 , and let xT be reach- able from x0 . For the two possible Cases 1 and 2 of inequalities relating x0 and xT to xs , the optimal trajectories are given in Figs. 7.10 and 7.11, respectively. Proof The requirement of feasibility when T is small rules out cases where x0 and xT are on opposite sides of or equal to xs . The proofs of optimality of the trajectories shown in Figs. 7.10 and 7.11 are similar to the proofs of the parts of Theorem 7.1, and are left as Exercise 7.25. In Figs. 7.10 and 7.11, it is possible to have either t1 ≥ T or t2 ≥ T. Try sketching some of these special cases. 2 All of the previous discussion has assumed that Q was ﬁnite and suﬃciently large, but we can easily extend this to the case when Q = ∞.

7.2. The Vidale-Wolfe Advertising Model

243

Figure 7.9: Optimal trajectory (solid lines)

Figure 7.10: Optimal trajectory when T is small in Case 1: x0 < xs and x T > xs

This possibility makes the arcs in Figs. 7.5, 7.6, 7.7, 7.8, 7.9, and 7.10, corresponding to u∗ = Q, become vertical line segments corresponding to impulse controls. For example, Fig. 7.6 becomes Fig. 7.12 when Q = ∞ and we apply the impulse control imp(x0 , xs ; 0) when x0 < xs . Next we compute the cost of imp(x0 , xs ; 0) by assessing its eﬀect on the objective function of (7.26). For this, we integrate the state equation in (7.26) from 0 to ε with the initial condition x0 and u treated

244

7. Applications to Marketing

x 1 xT x0 u* = 0

u* = Q

xs

0

T – t2

T

t1

t

Figure 7.11: Optimal trajectory when T is small in Case 2: x0 > xs and x T > xs

Impulse Control Impulse Control

Figure 7.12: Optimal trajectory for Case 2 of Theorem 7.1 for Q = ∞

7.2. The Vidale-Wolfe Advertising Model

245

as constant. By using (A.7), we can write the solution as ε −(δ+ru)ε + e(δ+ru)(τ −ε) rudτ x(ε) = x0 e 0 ru ru = x0 − e−(δ+ru)ε + . δ + ru δ + ru According to the procedure given in Sect. 1.4, we must, for u, choose u(ε) so that x(ε) is xs . It should be clear that u(ε) → ∞ as ε → 0. With F (x, u, τ ) = πx(τ ) − u(τ ) and t = 0 in (1.23), we have the impulse I = imp(x0 , xs ; 0) = lim [−u(ε)ε]. ε→0

It is possible to solve for I by letting ε → 0, −u(ε)ε → I, u(ε) → ∞, and x(ε) = xs in the expression for x(ε) obtained above. This gives x(0+) = erI (x0 − 1) + 1. Therefore,

1 − x0 1 . imp(x0 , x ; 0) = − ln r 1 − xs s

(7.36)

We remark that this formula holds for any time t, as well as t = 0. Hence it can also be used at t = T to compute the impulse at the end of the period; see Fig. 7.12 and Exercise 7.28.

7.2.3

Solution When Q Is Small

When Q is small, it is not possible to go along the turnpike xs , so the arguments based on Green’s theorem become diﬃcult to apply. We therefore return to the maximum principle approach to analyze the problem. By “Q is small” we mean Q < us , where us is deﬁned in (7.35). Another characterization of the phrase “Q is small” in terms of the problem parameters is given in Exercise 7.30. We now apply the current-value maximum principle (3.42) to the ﬁxed-end-point problem given in (7.26). We form the current-value Hamiltonian as H = πx − u + λ[ru(1 − x) − δx] = πx − δλx + u[−1 + rλ(1 − x)],

(7.37)

and the Lagrangian function as L = H + μ(Q − u).

(7.38)

246

7. Applications to Marketing

The adjoint variable λ satisﬁes ∂L = ρλ + λ(ru + δ) − π, λ˙ = ρλ − ∂x

(7.39)

where λ(T ) is a constant, as in Row 2 of Table 3.1, that must be determined. Furthermore, the Lagrange multiplier μ in (7.38) must satisfy μ ≥ 0, μ(Q − u) = 0.

(7.40)

From (7.37) we notice that the Hamiltonian is linear in the control. So the optimal control is u∗ (t) = bang[0, Q; W (t)],

(7.41)

W (t) = W (x(t), λ(t)) = rλ(t)(1 − x(t)) − 1.

(7.42)

where

We remark that the suﬃciency conditions of Sect. 2.4, which require concavity of the derived Hamiltonian H 0 , do not apply here; see Exercise 7.33. However, the suﬃciency of the maximum principle for this kind of problem has been established in the literature; see, for example, Lansdowne (1970). When W = rλ(1 − x) − 1 = 0, we have the possibility of a singular control, provided we can maintain this equality over a ﬁnite time interval. For the case when Q is large, we showed in the previous section that the optimal trajectory contains a segment on which x = xs and u∗ = us , where 0 ≤ us ≤ Q. (See Exercise 7.30 for the condition that Q is small.) This can obviously be a singular control. Further discussion of singular control is given in Sect. D.6. A complete solution of problem (7.26) when Q is small requires a lengthy switching point analysis. The details are too voluminous to give here, but an interested reader can ﬁnd the details in Sethi (1973a).

7.2. The Vidale-Wolfe Advertising Model

7.2.4

247

Solution When T Is Inﬁnite

In Sects. 7.2.1 and 7.2.2, we assumed that T was ﬁnite. We now formulate the inﬁnite horizon version of (7.26): ⎧ ∞ ⎪ −ρt ⎪ max J = e (πx − u)dt ⎪ ⎪ ⎪ 0 ⎪ ⎪ ⎪ ⎪ ⎨ subject to (7.43) ⎪ ⎪ ⎪ x˙ = ru(1 − x) − δx, x(0) = x0 , ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 0 ≤ u ≤ Q. We divide the analysis of this problem into the same two cases deﬁned as before, namely, “Q is large” and “Q is small”. When Q is large, the results of Theorem 7.1 suggest the solution when T is inﬁnite. Because of the discount factor, the ending parts of the solutions shown in Figs. 7.5, 7.6, 7.7, and 7.8 can be shown to be irrelevant (i.e., the discounted proﬁt accumulated during the interval (T − t2 , T ) goes to 0 as T goes to ∞). Therefore, we only have two cases: (a) x0 ≤ xs , and (b) x0 ≥ xs . The optimal control in Case (a) is to use u∗ = Q in the interval [0, t1 ) and u∗ = us for t ≥ t1 . Similarly, the optimal control in Case (b) is to use u∗ = 0 in the interval [0, t1 ) and u∗ = us for t ≥ t1 . An alternate way to see that the above solutions give u∗ = us for t ≥ t1 is to check that they satisfy the turnpike conditions (3.107). To do this we need to ﬁnd the values of the state, control, and adjoint variables and the Lagrange multiplier along the turnpike. It can be easily shown that x = xs , u = us , λs = π/(ρ + δ + rus ), and μs = 0 satisfy the turnpike conditions (3.107). When Q is small, i.e., Q < us , it is not possible to follow the turnpike x = xs , because that would require u = us , which is not a feasible control. Intuitively, it seems clear that the “nearest” stationary path to xs that we can follow is the path obtained by setting x˙ = 0 and u = Q, the largest possible control, in the state equation of (7.43). This gives rQ , rQ + δ

(7.44)

π ρ + δ + rQ

(7.45)

x ¯= and correspondingly we obtain ¯= λ

248

7. Applications to Marketing

by setting u = Q and λ˙ = 0 in (7.39) and solving for λ. To ﬁnd an optimal solution from any given initial x0 , the approach we take is to ﬁnd a feasible path that is “nearest” to xs ; See Sethi (1977b) for further discussion. As we shall see, for x0 < xs , such a path is obtained by using the maximum possible control Q all the way. For x0 > xs , the situation is more diﬃcult. Nevertheless, the following two theorems give the turnpike as well as the optimal path starting from any given initial ¯ = rλ(1 ¯ −x ˆ and μ ¯ such that W (ˆ x, λ) ˆ) − 1 = 0 and x0 . Let us deﬁne x ¯ μ ¯ −μ Lu (¯ x, u ¯, λ, ¯ ) = W (¯ x, λ) ¯ = 0. Thus, ¯ x ˆ = 1 − 1/rλ,

(7.46)

¯ −x μ ¯ = rλ(1 ¯) − 1.

(7.47)

¯ μ Theorem 7.3 When Q is small, the quadruple {¯ x, Q, λ, ¯ } forms a turnpike. Proof We show that the turnpike conditions (3.107) hold for the quadruple. The ﬁrst two conditions of (3.107) are (7.44) and (7.45). By Exercise 7.31 we know x ¯ ≤ x ˆ, which, from deﬁnitions (7.46) and (7.47), implies μ ¯ ≥ 0. Furthermore u ¯ = Q, so (7.40) holds and the third condition of (3.107) also holds. Finally because W = μ ¯ from (7.42) and (7.47), it follows that W ≥ 0, so the Hamiltonian maximizing condition of (3.107) holds with u ¯ = Q. 2 Theorem 7.4 When Q is small, the optimal control at any time τ ≥ 0 is given by: (a) If x(τ ) ≤ x ˆ, then u∗ (τ ) = Q. (b) If x(τ ) > x ˆ, then u∗ (τ ) = 0. ¯ for all t ≥ τ and note that λ satisﬁes the Proof (a) We set λ(t) = λ adjoint equation (7.39) and the transversality condition (3.99). By Exercise 7.31 and the assumption that x(τ ) ≤ x ˆ, we know that x(t) ≤ x ˆ for all t. The proof that (7.40) and (7.41) hold for all t ≥ τ relies on the fact that x(t) ≤ x ˆ and on an argument similar to the proof of the previous theorem. ˆ for two difFigure 7.13 shows the optimal trajectories when x0 < x ferent starting values of x0 , one above and the other below x ¯. Note that in this ﬁgure we are always in Case (a) since x(τ ) ≤ x ˆ for all τ ≥ 0.

7.2. The Vidale-Wolfe Advertising Model

249

Figure 7.13: Optimal trajectories for x(0) < x ˆ (b) Assume x0 > x ˆ. In this case we will show that the optimal trajectory is as shown in Fig. 7.14, which is obtained by applying u = 0 until x=x ˆ and u = Q thereafter. Using this policy we can ﬁnd the time t1 at which x(t1 ) = x ˆ, by solving the state equation in (7.43) with u = 0. This gives 1 x0 t1 = ln . (7.48) δ x ˆ Clearly for t ≥ t1 , the policy u = Q is optimal because Case (a) applies. We now consider the interval [0, t1 ], where we set u = 0. Let τ be any time in this interval as shown in Fig. 7.14, and let x(τ ) be the corresponding value of the state variable. Then x(τ ) = x0 e−δτ . With u = 0 in (7.39), the adjoint equation on [0, t1 ] becomes λ˙ = (ρ + δ)λ − π. We also know that x(t1 ) = x ˆ. Thus, Case (a) applies at time t1 , and ¯ So, we solve the adjoint equation with we would like to have λ(t1 ) = λ. ¯ λ(t1 ) = λ and obtain π ¯− π (7.49) + λ e(ρ+δ)(τ −t1 ) , τ ∈ [0, t1 ]. λ(τ ) = ρ+δ ρ+δ Now, with the values of x(τ ) and λ(τ ) in hand, we can use (7.42) to obtain the switching function value W (τ ). In Exercise 7.34, you are

250

7. Applications to Marketing x 1

x(0) u* = 0 x(τ)

u* = Q

x xs

0

τ Case(b)

t

t1 Case(a)

Figure 7.14: Optimal trajectory for x(0) > x ˆ asked to show that the switching function W (τ ) is negative for each τ in the interval [0, t1 ) and W (t1 ) = 0. Therefore by (7.41), the policy u = 0 used in deriving (7.48) and (7.49) satisﬁes the maximum principle. This ¯ policy “joins” the optimal policy after t1 because λ(t1 ) = λ. In this book the suﬃciency of the transversality condition (3.99) is stated under the hypothesis that the derived Hamiltonian is concave; see Theorem 2.1. In the present example, this hypothesis does not hold. However, as mentioned in Sect. 7.2.3, for this simple bilinear problem it can be shown that (3.99) is suﬃcient for optimality. Because of the technical nature of this issue we omit the details. 2 Exercises for Chapter 7 E 7.1 In Eqs. (7.2) and (7.3), assume S(p, G) = 1000 − 5p + 2G and c(S) = 5S. Substitute into (7.5) and solve for the optimal price p∗ in terms of G. E 7.2 Derive the optimal monopoly price formula in (7.6) from (7.5). E 7.3 Derive the equilibrium goodwill level formula (7.12). E 7.4 Re-solve problem (7.7) with its objective function and the control constraint replaced by (7.13), and show that the only possible singular

Exercises for Chapter 7

251

level of goodwill (which can be maintained over a ﬁnite time interval) is ¯ obtained in (7.14). the value G E 7.5 Show that the total cost of advertising required to go from ¯ to G ¯ instantaneously (by an impulse) is G ¯ − G0 . G0 < G Hint: Integrate G˙ = u − δG, G(0) = G0 , from 0 to ε and equate ¯ = lim G(ε), where the limit is taken as ε → 0 and u → ∞ in such a G ¯ 0). See also the derivation of (7.36). way that uε → cost = −imp(G0 , G; E 7.6 Assume the eﬀect of the exogenous variable Z(t) is seasonal so ¯ that the goodwill G(t) = 2 + sin t. Assume δ = 0.1. Sketch the graph ˙ ¯ ¯ of u ¯(t) = δ G + G, similar to Fig. 7.2, and identify intervals in which ¯ maintaining the singular level G(t) is infeasible. E 7.7 In the Nerlove-Arrow Model of Sects. 7.1.1 and 7.1.2, assume S(p, A, Z) = αp−η Gβ Z γ and c(S) = cS. Show that the optimal stationary policy gives u ¯/pS = constant, i.e., that the optimal advertising level is a constant fraction of sales regardless of the value of Z. (Such policies are followed by many industries.) E 7.8 Verify that G1 and G2 , which are shown in Fig. 7.3 for the pulsing policy derived from solving problem (7.19) as a near-optimal solution of problem (7.17) with T = ∞, are given by

M 1 − e−δτ v¯ M e−δτ (1−¯v) − e−δτ G1 = . , G2 = δ 1 − e−δτ δ 1 − e−δτ E 7.9 Extend the Nerlove-Arrow Model and its results by introducing the additional capital stock variable K˙ = v − γK, K(0) = K0 , where v is the research expenditure. Assume the new cost function to be C(S, K). Note that this model allows the ﬁrm to manipulate its cost function. See Dhrymes (1962). E 7.10 Analyze an extension of a ﬁnite horizon Nerlove-Arrow Model subject to a budget constraint. That is, introduce the following isoperimetric constraint: T

0

ue−ρt dt = B.

√ Also assume π(G) = α G where α > 0 is a constant. See Sethi (1977c).

252

7. Applications to Marketing

E 7.11 Introduce a budget constraint in a diﬀerent way into the NerloveArrow model as follows. Let B(t) be the budget at time t, and let γ > 0 be a constant. Assume B satisﬁes B˙ = e−ρt (−u + γG), B(0) = B0 and B(t) ≥ 0 for all t. Solve only the inﬁnite horizon model. See Sethi and Lee (1981). E 7.12 Maximize the present value of total sales in the Nerlove-Arrow model, i.e., ∞ −ρt max J = e pS(p, G)dt u≥0

0

subject to (7.1) and the isoperimetric proﬁt constraint ∞ e−ρt [pS(p, G) − C(S) − u]dt = π ˆ. 0

See Tsurumi and Tsurumi (1971). E 7.13 A Logarithmic Advertising Model (Sethi 1975). (a) With πr > ρ + δ, solve ⎧ T ⎪ −ρt ⎪ ⎪ e (πx − u)dt max J = ⎪ ⎪ ⎪ 0 ⎪ ⎪ ⎪ ⎪ ⎪ subject to ⎪ ⎪ ⎨ x˙ = r log u − δx, x(0) = x0 , ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ and the control constraint ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ u ≥ 1. (b) Find the value of T for which the minimum advertising is optimal throughout, i.e., u∗ (t) = 1, 0 ≤ t ≤ T. ¯ (c) Let T = ∞. Obtain the long-run stationary equilibrium (¯ x, u ¯, λ). E 7.14 Let Q(t) = the quality of the software at time t; Q(0) ≥ 0,

Exercises for Chapter 7

253

P (t) = the price of the software at time t; P (t) ≥ 0, D(P, Q) = the demand; D(P, Q) ≥ 0, DQ ≥ 0, DP ≤ 0, g(x) = a decreasing function; g(x) ≥ 0, g (x) ≤ 0, g (x) ≥ 0, and g(x) → 0 as x → ∞, ρ = the discount rate; ρ > 0, δ = the obsolescence rate for software quality; δ > 0. Assume that lim P D(P, Q) = 0, for each Q.

P →0

Furthermore, we assume that there is a price that maximizes the revenue (in the case when there is more than one global maximum, we will choose the largest of these) and denote it as P m (Q). We assume that 0 < P m (Q) < ∞ and deﬁne R(Q) = P m (Q)D(P m (Q), Q). By the envelope theorem (see Derzko et al. 1984), we have RQ (Q) = P m (Q)DQ (P m (Q), Q) ≥ 0. In an open-source approach to software development, the improvement in software quality is proportional to the number of volunteer programmers participating at any point in time. The volunteer programmers’ willingness to contribute to software quality is driven by fairness considerations. To capture the loss of motivation that results from the proﬁt making of the ﬁrm, we formulate the motivations of the programmers based on the current or projected future proﬁt of the ﬁrm. Then, let g(P D) be the quality improvement aﬀected by the volunteer programmers. The optimal dynamic price and quality paths can be obtained by solving the following problem due to Haruvy et al. (2003): ∞ e−rt P Ddt , max J = P (t)≥0

0

dQ/dt = g(P D) − δQ, Q(0) = Q0 .

s.t.

Because of the convexity of function g in this case, argue that the problem would require the inclusion of chattering controls. Then reformulate the problem as ∞ −rt max J = e vR(Q)dt , 0≤v≤1

0

254

7. Applications to Marketing s.t.

dQ/dt = (1 − v)g(0) + vg(R(Q)) − δQ, Q(0) = Q0 .

Apply the Green’s theorem approach to solve this problem. E 7.15 For problem (7.26), ﬁnd the reachable set for a given initial x0 and horizon time T. E 7.16 Solve the quadratic equation I(x) = 0, where I(x) is deﬁned in (7.32), to obtain its solution as shown in (7.33). E 7.17 Show that both xs in (7.34) and us in (7.35) are 0 if, and only if, πr ≤ δ + ρ. E 7.18 For problem (7.26) with πr > δ + ρ and Q suﬃciently large, ¯ by using the maximum principle. Check derive the turnpike {¯ x, u ¯, λ} to see that x ¯ and u ¯ correspond, respectively, to xs and us derived by Green’s theorem. Show that when ρ = 0, x ¯ reduces to the golden path rule. E 7.19 Let xs denote the solution of I(x) = 0 and let A < xs < B in Fig. 7.4. Assume that I(x) > 0 for x > xs and I(x) < 0 for x < xs . Construct a path Γ3 such that JΓ3 ≥ JΓ1 and JΓ3 ≥ JΓ2 . Hint: Use Lemma 7.1. E 7.20 For the problem in (7.26), suppose x0 and xT are given and deﬁne xs as in (7.34). Let t1 be the shortest time to go from x0 to xs , and t2 be the shortest time to go from xs to xT . (a) If x0 < xs and xs > xT , show that

s x ¯ − x0 x 1 1 , t2 = ln , t1 = ln s rQ + δ x ¯−x δ xT where x ¯ = rQ/(rQ + δ); assume x ¯ > xs . (b) Using the form of the answers in (a), ﬁnd t1 and t2 when x0 > xs and xs < xT < x ¯. E 7.21 For Exercise 7.20(a), write the condition that T is large, i.e., T ≥ t1 + t2 , in terms of all the parameters of problem (7.26). E 7.22 Perform the following:

Exercises for Chapter 7

255

(a) For problem (7.26), assume r = 0.2, δ = 0.05, ρ = 0.1, Q = 5, π = 2, x0 = 0.2 and xT = 0.3. Use Exercises 7.20(a) and 7.21 to show that T = 13 is large and T = 8 is small. Sketch the optimal trajectories for T = 13 and T = 8. (b) Redo (a) when xT = 0.7. Show that both T = 13 and T = 8 are large. E 7.23 Prove Theorem 7.1 for Case 3. E 7.24 Draw four ﬁgures for the case T = t1 + t2 corresponding to Figs. 7.5, 7.6, 7.7, and 7.8. E 7.25 Prove Theorem 7.2. E 7.26 Sketch one or two other possible curves for the case when T is small. E 7.27 An intermediate step in the derivation of (7.36) is to establish that ε e−ρt [πx(t) − u(t)]dt = lim [−u(ε)ε]. lim ε→0 0

ε→0

Show how to accomplish this by using the Mean Value Theorem. E 7.28 Obtain the impulse function, imp(xs , xT ; T ), required to take the state from xs up to xT instantaneously at time T as shown in Fig. 7.12 for the Vidale-Wolfe model in Sect. 7.2.2. E 7.29 Perform the following: (a) Re-solve Exercise 7.22(a) with Q = ∞. Show T = 10.5 is no longer small. (b) Show that T > 0 is large for Exercise 7.22(b) when Q = ∞. Find the optimal value of the objective function when T = 8. E 7.30 Show that Q is small if, and only if, πrδ > 1. (δ + ρ + rQ)(δ + rQ) E 7.31 Perform the following:

256

7. Applications to Marketing

ˆ when Q is small, where x ˆ is (a) Show that x ¯ < xs < x deﬁned in (7.46). (b) Show that x ¯ > xs when Q is large. E 7.32 Derive (7.48). E 7.33 Show the derived Hamiltonian H 0 corresponding to (7.37) and (7.41) is not concave in x for any given λ > 0. E 7.34 Show that the switching function deﬁned in (7.42) is concave in t, and then verify that the policy in Fig. 7.14 satisﬁes (7.41). E 7.35 In (7.25), assume r and δ are positive, diﬀerentiable functions of time. Derive expressions similar to (7.31)–(7.35) in order to get the new turnpike values. E 7.36 Write the equation satisﬁed by the turnpike level x ¯ for the model ⎧ ∞ ⎪ −ρt 2 ⎪ max J = e (πx − u )dt ⎪ ⎪ ⎪ 0 ⎨ u≥0 subject to ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ x˙ = ru(1 − x) − δx, x(0) = x . 0 Show that the turnpike reduces to the golden path when ρ = 0. E 7.37 Obtain the optimal long-run stationary equilibrium for the following modiﬁcation of the model (7.26), due to Sethi (1983b): ⎧ ∞ ⎪ ⎪ max e−ρt (πx − u2 )dt ⎪ ⎪ ⎪ 0 ⎪ ⎪ ⎪ ⎪ ⎨ subject to (7.50) # ⎪ ⎪ ⎪ ⎪ ⎪ x˙ = ru (1 − x) − δx, x0 ∈ [0, 1], ⎪ ⎪ ⎪ ⎪ ⎩ u ≥ 0. ¯ u In particular, show that the turnpike triple (¯ x, λ, ¯) is given by √ ¯ ¯ 1−x ¯ rλ r2 λ/2 , u ¯= x ¯ = 2¯ , 2 r λ/2 + δ

(7.51)

Exercises for Chapter 7

257

and # ¯= λ

[(ρ + δ)2 + r2 π] − (ρ + δ) . r2 /2

(7.52)

Show that the optimal value of the objective function is 2

2¯ ¯ 0+r λ . J ∗ (x0 ) = λx 4ρ

(7.53)

E 7.38 Consider (7.43) with the state equation replaced by x˙ = ru(1 − x) + μx(1 − x) − δx, x(0) = x0 , where the constant μ > 0 reﬂects word-of-mouth communication between buyers (represented by x) and non-buyers (represented by (1 − x)) of the product. Assume Q is inﬁnite for convenience. Obtain the turnpike for this problem. See Sethi (1974b). E 7.39 The Ozga Model (Ozga 1960; Gould 1970). Suppose the information spreads by word of mouth rather than by an impersonal advertising medium, i.e., individuals who are already aware of the product inform individuals who are not, at a certain rate, inﬂuenced by advertising expenditure. What we have now is the Ozga model x˙ = ux(1 − x) − δx, x(0) = x0 . The optimal control problem is to maximize ∞ e−ρt [π(x) − w(u)]dt J= 0

subject to the Ozga model. Assume that π(x) is concave and w(u) is convex. See Sethi (1979c) for a Green’s theorem application to this problem.

Chapter 8

The Maximum Principle: Discrete Time For many purposes it is convenient to assume that time is represented by a discrete variable, k = 0, 1, 2, . . . , T, rather than by a continuous variable t ∈ [0, T ]. This is particularly true when we wish to solve a large control theory problem by means of a computer. It is also desirable, even when solving small problems which have state or adjoint diﬀerential equations whose solutions cannot be expressed in closed form, to formulate them as discrete problems and let the computer solve them in a stepwise manner. We will see that the maximum principle, which is to be derived in this chapter, is not valid for the discrete-time problem in as wide a sense as for the continuous-time problem. In fact, we will reduce it to a nonlinear programming problem and state necessary conditions for its solution by using the well-known Kuhn-Tucker theorem. In order to follow this procedure, we have to make some simplifying assumptions and hence will obtain only a restricted form of the discrete maximum principle. In Sect. 8.3, we state without proof a more general form of the discrete maximum principle.

8.1

Nonlinear Programming Problems

We begin by stating a general form of a nonlinear programming problem. Let x be an n-component column vector, a an r-component column vector, and b an s-component column vector. Let the functions © Springer Nature Switzerland AG 2019 S. P. Sethi, Optimal Control Theory, https://doi.org/10.1007/978-3-319-98237-3 8

259

260

8. The Maximum Principal: Discrete Time

h : E n → E 1 , f : E n → E r , and g : E n → E s be continuously differentiable. We assume functions f and g to be column vectors with r and s components, respectively. We consider the nonlinear programming problem: max h(x) (8.1) subject to r equality constraints and s inequality constraints given, respectively, by f (x) = a,

(8.2)

g(x) ≥ b.

(8.3)

Next we develop necessary conditions, called the Kuhn-Tucker conditions, which a solution x∗ to this problem must satisfy. We start with simpler problems and work up to the statement of these conditions for the general problem in a heuristic fashion. References are given for rigorous developments of these results. In this chapter, whenever we take derivatives of functions, we assume that those derivatives exist and are continuous. It would be also helpful to recall the notation developed in Sect. 1.4.

8.1.1

Lagrange Multipliers

Suppose we want to solve (8.1) without imposing constraints (8.2) or (8.3). The problem is now the classical unconstrained maximization problem of calculus, and the ﬁrst-order necessary conditions for its solution are (8.4) hx = 0. The points satisfying (8.4) are called critical points. Critical points which are maxima, minima, or saddle points are of interest in this book. Additional higher-order conditions required to determine whether a critical point is a maximum or a minimum are stated in Exercise 8.2. In an important case when the function h is concave, condition (8.4) is also suﬃcient for a global maximum of h. Suppose we want to solve (8.1) while imposing just the equality constraints (8.2). The method of Lagrange multipliers permits us to obtain the necessary conditions that a solution to the constrained maximization problem (8.1) and (8.2) must satisfy. We deﬁne the Lagrangian function L(x, λ) = h(x) + λ[f (x) − a],

(8.5)

8.1. Nonlinear Programming Problems

261

where λ is an r-component row vector. The necessary condition for x∗ to be a (maximum) solution to (8.1) and (8.2) is that there exists an r-component row vector λ such that (x∗ , λ) satisfy the equations Lx = hx + λfx = 0,

(8.6)

Lλ = f (x) − a = 0.

(8.7)

Note that (8.7) states simply that x∗ is feasible according to (8.2). The system of n + r Eqs. (8.6) and (8.7) has n + r unknowns. Since some or all of the equations are nonlinear, the solution method will, in general, involve nonlinear programming techniques, and may be diﬃcult. In other cases, e.g., when h is linear and f is quadratic, it may only involve the solution of linear equations. Once a solution (x∗ , λ) is found satisfying the necessary conditions (8.6) and (8.7), the solution must still be checked to see whether it satisﬁes suﬃcient conditions for a global maximum. Such suﬃcient conditions will be stated in Sect. 8.1.4. Suppose (x∗ , λ) is in fact a solution to equations (8.6) and (8.7). Note that x∗ depends on a and we can show this dependence by writing x∗ = x∗ (a). Now h∗ = h∗ (a) = h(x∗ (a)) is the optimum value of the objective function. By diﬀerentiating h∗ (a) with respect to a and using (8.6), we obtain dx∗ dx∗ = −λfx . h∗a = hx da da But by diﬀerentiating (8.7) with respect to a at x = x∗ (a), we get fx and therefore we have

dx∗ = 1, da

h∗a = −λ.

(8.8)

We can see that the Lagrange multipliers have an important managerial interpretation, namely, λi is the negative of the imputed value or shadow price of having one unit more of the resource ai . In Exercise 8.4 you are asked to provide a proof of (8.8). Example 8.1 Consider the two-dimensional problem: ⎧ ⎪ ⎪ ⎪ max{h(x, y) = −x2 − y 2 } ⎪ ⎪ ⎨ subject to ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 2x + y = 10.

262

8. The Maximum Principal: Discrete Time

Solution We form the Lagrangian L(x, y, λ) = (−x2 − y 2 ) + λ(2x + y − 10). The necessary conditions for an optimal solution (x∗ , y ∗ ) are that (x∗ , y ∗ , λ) satisfy the equations Lx = −2x + 2λ = 0, Ly = −2y + λ = 0, Lλ = 2x + y − 10 = 0. From the ﬁrst two equations we get λ = x = 2y. Solving this with the last equation yields the quantities x∗ = 4, y ∗ = 2, λ = 4, h∗ = −20, which can be seen to give a maximum value to h, since h is concave and the constraint set is convex. The interpretation of the Lagrange multiplier λ = 4 can be obtained to verify (8.8) by replacing the constant 10 by 10 + and expanding the objective function in a Taylor series; see Exercise 8.5.

8.1.2

Equality and Inequality Constraints

Now suppose we want to solve the problem deﬁned by (8.1)–(8.3). As before, we deﬁne the Lagrangian L(x, λ, μ) = h(x) + λ[f (x) − a] + μ[g(x) − b].

(8.9)

The Kuhn-Tucker necessary conditions for this problem cannot be as easily derived as for the equality-constrained problem in the preceding section. We will write them ﬁrst, and then give interpretations to make them plausible. The necessary conditions for x∗ to be a solution of (8.1)– (8.3) are that there exist an r-dimensional vector λ and an s-dimensional row vector μ such that Lx = hx + λfx + μgx = 0, f

(8.10)

= a,

(8.11)

g ≥ b,

(8.12)

μ ≥ 0, μ(g − b) = 0.

(8.13)

8.1. Nonlinear Programming Problems

263

Note that g is appended in (8.10) in the same way f is appended in (8.6). Also (8.12) repeats the inequality constraint (8.3) in the same way that (8.11) repeats the equality constraint (8.2). However, the conditions in (8.13) are new and particular to the inequality-constrained problem. We will see that they include some of the boundary points of the feasible set of points as well as unconstrained maximum solution points, as candidates for the solution to the maximum problem. This is best brought out by examples. Example 8.2 Solve the problem: ⎧ ⎪ ⎪ ⎪ max{h(x) = 8x − x2 } ⎪ ⎪ ⎨ subject to ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ x ≥ 2. Solution We form the Lagrangian L(x, μ) = 8x − x2 + μ(x − 2). The necessary conditions (8.10)–(8.13) become Lx = 8 − 2x + μ = 0,

(8.14)

x − 2 ≥ 0,

(8.15)

μ ≥ 0, μ(x − 2) = 0.

(8.16)

Observe that the constraint μ(x − 2) = 0 in (8.16) can be phrased as: either μ = 0 or x = 2. We treat these two cases separately. Case 1: μ = 0. From (8.14) we get x = 4, which also satisﬁes (8.15). Hence, this solution, which makes h(4) = 16, is a possible candidate for the maximum solution. Case 2: x = 2. Here from (8.14) we get μ = −4, which does not satisfy the inequality μ ≥ 0 in (8.16). From these two cases we conclude that the optimum solution is x∗ = 4 and h∗ = h(x∗ ) = 16.

264

8. The Maximum Principal: Discrete Time

Example 8.3 Solve the problem: ⎧ ⎪ ⎪ ⎪ max{h(x) = 8x − x2 } ⎪ ⎪ ⎨ subject to ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ x ≥ 6. Solution The Lagrangian is L(x, μ) = 8x − x2 + μ(x − 6). The necessary conditions are Lx = 8 − 2x + μ = 0,

(8.17)

x − 6 ≥ 0,

(8.18)

μ ≥ 0, μ(x − 6) = 0.

(8.19)

Again, the condition μ(x − 6) = 0 is an either-or relation which gives two cases. Case 1: μ = 0. From (8.17) we obtain x = 4, which does not satisfy (8.18), so this case is infeasible. Case 2: x = 6. Obviously (8.18) holds. From (8.17) we get μ = 4, so (8.19) holds as well. The optimal solution is then x∗ = 6, h∗ = h(x∗ ) = 12, since it is the only solution satisfying the necessary conditions. The examples above involve only one variable, and are relatively obvious. The next example, which is two-dimensional, will reveal more of the power and the diﬃculties of applying the Kuhn-Tucker conditions. Example 8.4 Find the shortest distance between the point (2,2) and the upper half of the semicircle of radius one with its center at the origin, shown as the curve in Fig. 8.1. In order to simplify the calculation, we minimize h, the square of the distance. Hence, the problem can be stated

8.1. Nonlinear Programming Problems

265

as the following nonlinear programming problem: ⎧ . ⎪ ⎪ ⎪ max −h(x, y) = −(x − 2)2 − (y − 2)2 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎨ subject to ⎪ ⎪ ⎪ x2 + y 2 = 1, ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ y ≥ 0. The Lagrangian function for this problem is L = −(x − 2)2 − (y − 2)2 + λ(x2 + y 2 − 1) + μy.

(8.20)

The necessary conditions are −2(x − 2) + 2λx = 0,

(8.21)

−2(y − 2) + 2λy + μ = 0,

(8.22)

2

2

x + y − 1 = 0,

(8.23)

y ≥ 0,

(8.24)

μ ≥ 0, μy = 0.

(8.25)

First, we conclude that λ = 0, since otherwise λ = 0 would imply x = 2 from (8.21), which would contradict (8.23). Next, from (8.25) we conclude that either μ = 0 or y = 0. If μ = 0, then from (8.21) and (8.22), we get x = y. Solving the equation x = y together with x2 +y 2 = 1 gives: √ √ √ (a) (1/ 2, 1/ 2) and h = −(9 − 4 2). If y = 0, then solving with x2 + y 2 = 1 gives: (b) (1, 0) and h = −5, (c) (−1, 0) and h = −13. These three points are shown in Fig. 8.1. Of the three√points √ found that satisfy the necessary conditions, clearly the point (1/ 2, 1/ 2) found in (a) is the nearest point and solves the closest-point problem. The point (−1, 0) in (c) is in fact the farthest point; and the point (1, 0) in (b) is neither the closest nor the farthest point. The associated√multiplier values can be easily computed, and these are: (a) λ = 1 − 2 2, μ = 0; (b) λ = −1, μ = 4; and (c) λ = 3, μ = 4.

266

8. The Maximum Principal: Discrete Time

Closest Point

Farthest Point

(-1,0)

(1,0)

Figure 8.1: Shortest distance from point (2,2) to the semicircle The fact that there are three points satisfying the necessary conditions, and only one of them actually solves the problem at hand, emphasizes that the conditions are only necessary and not suﬃcient. In every case it is important to check the solutions to the necessary conditions to see which of the solutions provides the optimum. Next we work two examples that show some technical diﬃculties that can arise in the application of the Kuhn-Tucker conditions. Example 8.5 Consider the problem: max{h(x, y) = y}

(8.26)

(1 − y)3 − x2 ≥ 0,

(8.27)

x ≥ 0

(8.28)

y ≥ 0.

(8.29)

subject to

The set of points satisfying the constraints is shown shaded in Fig. 8.2. From the ﬁgure it is obvious that the solution point (0,1) maximizes the value of y. Hence, the optimum solution is (x∗ , y ∗ ) = (0, 1) and h∗ = 1. Let us see if we can ﬁnd it using the above procedure. The Lagrangian is L = y + λ[(1 − y)3 − x2 ] + μx + νy,

(8.30)

8.1. Nonlinear Programming Problems

267

Figure 8.2: Graph of Example 8.5 so that the necessary conditions are Lx = −2xλ + μ = 0,

(8.31)

2

Ly = 1 − 3λ(1 − y) + ν = 0, 3

2

(8.32)

λ ≥ 0, λ[(1 − y) − x ] = 0,

(8.33)

μ ≥ 0, μx = 0,

(8.34)

ν ≥ 0, νy = 0,

(8.35)

together with (8.27)–(8.29). Let us check if these conditions hold at the point (0,1). At y = 1, the constraint y ≥ 0 is not active, and we have ν = 0. With ν = 0 and y = 1, (8.32) cannot be satisﬁed. The reason for failure of the method in Example 8.5 is that the constraints do not satisfy what is called the constraint qualiﬁcation. A complete study of the topic is beyond the scope of this book, but we state in the next section a constraint qualiﬁcation suﬃcient for our purposes. For further information, see Mangasarian (1969).

8.1.3

Constraint Qualiﬁcation

Example 8.5 shows the need for imposing some kind of condition to rule out features such as the cusp at (0, 1) in Fig. 8.2 on the boundary of the constraint set. One way to accomplish this is to assume that the gradients of the equality constraints and of the active inequality constraints at the candidate point under consideration are linearly independent. Equivalently, we say that the constraints (8.2) and (8.3) satisfy the constraint qualiﬁcation at x if the following full-rank condition holds at x, that is, ⎤ ⎡ ⎢ ∂g/∂x diag(g) ⎥ (8.36) rank ⎣ ⎦ = min(s + r, s + n), ∂f /∂x 0

268

8. The Maximum Principal: Discrete Time

where ∂g/∂x and ∂f /∂x are s × n and r × n gradient matrices, respectfully, as deﬁned in Sect. 1.4.3, the notation diag(g) refers to the diagonal s × s matrix ⎤ ⎡ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎣

g1

0

0

g2

.. .

.. .

0

0

···

0 ⎥ ⎥ ⎥ ··· 0 ⎥ ⎥ ⎥, .. ⎥ ··· . ⎥ ⎥ ⎦ · · · gs

and therefore the matrix in (8.36) is an (s + r) × (s + n) matrix. Let us now return to Example 8.5 and examine whether the constraints (8.27)–(8.29) satisfy the constraint qualiﬁcation at point (0,1). In this example, s = 3, r = 0 and n = 2, and the matrix in (8.36) is ⎤ ⎡ ⎡ ⎤ 2 (1 − y)3 − x2 0 0 ⎥ ⎢ 0 0 0 0 0 ⎥ ⎢ −2x −3(1 − y) ⎥ ⎢ ⎢ ⎥ ⎥ ⎢ ⎢ ⎥ ⎥=⎢ 1 0 0 0 0 ⎥ ⎢ 1 0 0 x 0 ⎥ ⎢ ⎢ ⎥ ⎦ ⎣ ⎣ ⎦ 0 1 0 0 y 0 1 0 0 1 at point (x, y) = (0, 1). It has a null vector in the ﬁrst row, and therefore its rows are not linearly independent; see Sect. 1.4.10. Thus, it does not have a full rank of three, and the condition (8.36) does not hold. Alternatively, note that the inequality constraints (8.27) and (8.28) are active at point (x, y) = (0, 1), and their respective gradients (−2x, −3(1− y)2 ) = (0, 0) and (1, 0) at that point are clearly not linearly independent.

8.1.4

Theorems from Nonlinear Programming

In order to derive our version of the discrete maximum principle, we use two well-known results from nonlinear programming. These provide suﬃcient and necessary conditions for the problem given by (8.1)–(8.3). The Lagrangian function for this problem is L(x, λ, μ) = h + λ(f (x) − a) + μ(g(x) − b),

(8.37)

where λ and μ are row vectors of multipliers associated with the constraints (8.2) and (8.3), respectively. We now state two theorems whose proofs can be found in Mangasarian (1969).

8.2. A Discrete Maximum Principle

269

Theorem 8.1 (Necessary Conditions) If h, f, and g are diﬀerentiable, x∗ solves (8.1)–(8.3), and the constraint qualiﬁcation (8.36) holds at x∗ , then there exist multipliers λ and μ such that (x∗ , λ, μ) satisfy the Kuhn-Tucker conditions Lx (x∗ , λ, μ) = hx (x∗ ) + λfx (x∗ ) + μgx (x∗ ) = 0, ∗

f (x ) = a, ∗

(8.38) (8.39)

g(x ) ≥ b,

(8.40)

μ ≥ 0, μ(g(x ) − b) = 0,

(8.41)

∗

Theorem 8.2 (Suﬃcient Conditions) If h, f, and g are diﬀerentiable, f is aﬃne, g is concave, and (x∗ , λ, μ) satisfy the conditions (8.38)–(8.41), then x∗ is a solution to the maximization problem (8.1)– (8.3).

8.2

A Discrete Maximum Principle

We will now use the nonlinear programming results of the previous section to derive a special form of the discrete maximum principle. Some references in this connection are Luenberger (1972), Mangasarian and Fromovitz (1967), and Ravn (1999). A more general discrete maximum principle will be stated in Sect. 8.3.

8.2.1

A Discrete-Time Optimal Control Problem

In order to state a discrete-time optimal control problem over the periods 0, 1, 2, . . . , T, we deﬁne the following: Θ =

the set {0, 1, 2, . . . , T − 1},

x

k

=

an n-component column state vector; k = 0, 1, . . . , T,

u

k

=

an m-component column control vector; k = 0, 1, 2, . . . , T − 1,

k

=

an s-component column vector of constants; k=0, 1, . . . , T −1.

b

Here, the state xk is assumed to be measured at the beginning of period k and control uk is implemented during period k. This convention is depicted in Fig. 8.3.

270

8. The Maximum Principal: Discrete Time x0

x1

x2

xk

xk+1

xT −1

xT

u0

u1

u2

uk

uk+1

uT −1

0

1

2

k

k+1

T −1

T

Figure 8.3: Discrete-time conventions We also deﬁne continuously diﬀerentiable functions f : E n × E m × Θ → E n , F : E n × E m × Θ → E 1 , g : E m × Θ → E s , and S : E m × Θ ∪ {T } → E 1 . Then, a discrete-time optimal control problem in the Bolza form (see Sect. 2.1.4) is: max J =

T −1

k

k

T

F (x , u , k) + S(x , T )

(8.42)

k=0

subject to the diﬀerence equation xk = xk+1 − xk = f (xk , uk , k), k = 0, . . . , T − 1, x0 given,

(8.43)

and the constraints g(uk , k) ≥ bk , k = 0, . . . , T − 1.

(8.44)

In (8.43) the term xk = xk+1 − xk is known as the diﬀerence operator. This problem is clearly a special case of the nonlinear programming problem (8.1)–(8.3) with x = (x1 , x2 , . . . , xT , u0 , u1 , . . . , uT −1 ) as the (n+ m)T vector of variables, nT equality constraints (8.43), and sT inequality constraints (8.44).

8.2.2

A Discrete Maximum Principle

We now apply the nonlinear programming theory of Sect. 8.1 to ﬁnd necessary conditions for the solution to the Mayer form of the control problem of Sect. 8.2.1. We let λk+1 be an n-component row vector of Lagrange multipliers, which we rename adjoint variables and associate with Eq. (8.43). Similarly, we let μk be an s-component row vector of Lagrange multipliers associated with constraint (8.44). These multipliers are deﬁned for each time k = 0, . . . , T − 1.

8.2. A Discrete Maximum Principle

271

The Lagrangian function of the problem is L=

T −1

T −1

F (xk , uk , k) + S(xT , T ) +

k=0

k=0 T −1

+

λk+1 [f (xk , uk , k) − xk+1 + xk ] μk [g(uk , k) − bk ].

(8.45)

k=0

We now deﬁne the Hamiltonian function H k to be H k = H(xk , uk , k) = F (xk , uk , k) + λk+1 f (xk , uk , k).

(8.46)

Using (8.46) we can rewrite (8.45) as L = S(xT , T ) +

+

T −1

[H k − λk+1 (xk+1 − xk )]

k=0 T −1

μk [g(uk , k) − bk ].

(8.47)

k=0

We can now apply the Kuhn-Tucker conditions (8.38)–(8.41). Conditions (8.39) and (8.40) in this case give (8.43) and (8.44), respectively. Application of (8.38) results in (8.48)–(8.50) below and application of (8.41) gives the complimentary slackness conditions (8.51) below. By diﬀerentiating (8.47) with respect to xk for k = 1, 2, . . . , T − 1, we obtain ∂L ∂H k = − λk + λk+1 = 0, ∂xk ∂xk which upon rearranging terms becomes λk = λk+1 − λk = −

∂H k , k = 0, 1, . . . , T − 1. ∂xk

(8.48)

By diﬀerentiating (8.47) with respect to xT , we get ∂L ∂S ∂S = − λT = 0, or λT = . (8.49) T T ∂x ∂x ∂xT The diﬀerence equations (8.48) with terminal boundary conditions (8.49) are called the adjoint equations. By diﬀerentiating L with respect to uk and stating the corresponding Kuhn-Tucker conditions for the multiplier μk and constraint (8.44), we have ∂L ∂g ∂H k = + μk k = 0 k k ∂u ∂u ∂u or

272

8. The Maximum Principal: Discrete Time ∂H k ∂g = −μk k , ∂uk ∂u

(8.50)

μk ≥ 0, μk [g(uk , k) − bk ] = 0.

(8.51)

and We note that, provided H k is concave in uk , g(uk , k) is concave in uk , and the constraint qualiﬁcation holds, then conditions (8.50) and (8.51) are precisely the necessary and suﬃcient conditions for solving the following Hamiltonian maximization problem: ⎧ ⎪ ⎪ Hk ⎪ max ⎪ k ⎪ u ⎨ (8.52) subject to ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ g(uk , k) ≥ bk . We have thus derived the following restricted form of the discrete maximum principle. Theorem 8.3 If for every k, H k in (8.46) and g(uk , k) are concave in uk , and the constraint qualiﬁcation holds, then the necessary conditions for uk∗ , k = 0, 1, . . . , T − 1, to be an optimal control for the problem (8.42)–(8.44), with the corresponding state xk∗ , k = 0, 1, . . . , T, are ⎧ ⎪ ⎪ ⎪ xk∗ = f (xk∗ , uk∗ , k), x0 given, ⎪ ⎪ ⎪ ⎪ ⎪ T∗ ⎪ ⎨ λk = − ∂Hkk [xk∗ , uk∗ , λk+1 , k], λT = ∂S(x T ,T ) , ∂x ∂x ⎪ ⎪ ⎪ H k (xk∗ , uk∗ , λk+1 , k) ≥ H k (xk∗ , uk , λ(k+1) , k), ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ for all uk such that g(uk , k) ≥ bk , k = 0, 1, . . . , T − 1. (8.53) Section 8.2.3 gives examples of the application of this maximum principle (8.53). In Sect. 8.3 we state a more general discrete maximum principle.

8.2.3

Examples

Our ﬁrst example will be similar to Example 2.4 and it will be solved completely. The reader will note that the solutions of the continuous and discrete problems are very similar. The second example is a discrete version of the production-inventory problem of Sect. 6.1.

8.2. A Discrete Maximum Principle

273

Example 8.6 Consider the discrete-time optimal control problem: T −1 1 k 2 (8.54) − (x ) max J = 2 k=1

subject to xk = uk , x0 = 5,

(8.55)

u ∈ Ω = [−1, 1].

(8.56)

k

We will solve this problem for T = 6 and T ≥ 7. Solution The Hamiltonian is 1 H k = − (xk )2 + λk+1 uk , 2

(8.57)

from which it is obvious that the optimal policy is bang-bang. Its form is ⎧ ⎪ ⎪ ⎪ 1 if λk+1 > 0, ⎪ ⎪ ⎨ (8.58) uk∗ = bang[−1, 1; λk+1 ] = singular if λk+1 = 0, ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ −1 if λk+1 < 0. Let us assume, as we did in Example 2.4, that λk < 0 as long as xk is positive so that uk = −1. Given this assumption, (8.55) becomes xk = −1, whose solution is xk∗ = −k + 5 for k = 1, 2, . . . , T − 1.

(8.59)

By diﬀerentiating (8.57), we obtain the adjoint equation ∂H k // λ = − k /xk∗ = xk∗ , λT = 0. ∂x k

Let us assume T = 6. Substitute (8.59) into (8.60) to obtain λk = −k + 5, λ6 = 0. From Sect. A.5, we ﬁnd the solution to be 1 11 λk = − k 2 + k + c, 2 2

(8.60)

274

8. The Maximum Principal: Discrete Time

where c is a constant. Since λ6 = 0, we can obtain the value of c by setting k = 6 in the above equation. Thus, 11 1 λ6 = − (36) + (6) + c = 0 ⇒ c = −15, 2 2 so that

1 11 λk = − k 2 + k − 15. 2 2

(8.61)

A sketch of the values for λk and xk appears in Fig. 8.4. Note that λ = 0, so that the control u4 is singular. However, since x4 = 1 we choose u4 = −1 in order to bring x5 down to 0. The solution of the problem for T ≥ 7 is carried out in the same way that we solved Example 2.4. Namely, observe that x5∗ = 0 and λ5 = λ6 = 0, so that the control is singular. We simply make λk = 0 for k ≥ 7 so that uk∗ = 0 for all k ≥ 7. It is clear without a formal proof that this maximizes (8.54). 5

Example 8.7 Let us consider a discrete version of the productioninventory example of Sect. 6.1; see Kleindorfer et al. (1975). Let I k , P k , and S k be the inventory, production, and demand at time k, respectively. Let I 0 be the initial inventory, let Iˆ and Pˆ be the goal levels of inventory and production, and let h and c be inventory and production cost coeﬃcients. The problem is: T −1 1 k 2 k 2 ˆ + c(P − Pˆ ) ] − [h(I − I) (8.62) max J = 2 P k ≥0 k=0

subject to I k = P k − S k , k = 0, 1, . . . , T − 1, I 0 given.

(8.63)

Form the Hamiltonian 1 ˆ 2 + c(P k − Pˆ )2 ] + λk+1 (P k − S k ), H k = − [h(I k − I) 2

(8.64)

where the adjoint variable satisﬁes λk = −

∂H k ˆ λT = 0. = h(I k − I), ∂I k

(8.65)

8.2. A Discrete Maximum Principle xk

275

+

5 4 3 2 1 1

0

2

3

4

5

6

k

-1

-3

-6

-10 λk ∗

Figure 8.4: Optimal state xk and adjoint λk To maximize the Hamiltonian, let us diﬀerentiate (8.64) to obtain ∂H k = −c(P k − Pˆ ) + λk+1 = 0. ∂P k Since production must be nonnegative, we obtain the optimal production as (8.66) P k∗ = max[0, Pˆ + λk+1 /c]. Expressions (8.63), (8.65), and (8.66) determine a two-point boundary value problem. For a given set of data, it can be solved numerically by using spreadsheet software like Excel; see Sect. 2.5 and Exercise 8.21. If the constraint P k ≥ 0 is dropped it can be solved analytically by the method of Sect. 6.1, with diﬀerence equations replacing the diﬀerential equations used there.

276

8.3

8. The Maximum Principal: Discrete Time

A General Discrete Maximum Principle

For the maximum principle (8.53) we assumed that H k and g were concave in uk so that the set of admissible controls was convex. These are fairly strong assumptions which will now be relaxed and a general maximum principle stated. The proof can be found in Canon et al. (1970). Other references on discrete maximum principles are Halkin (1966) and Holtzman (1966). The problem to be solved is: max J =

T −1

F (xk , uk , k)

(8.67)

k=0

subject to xk = f (xk , uk , k), x0 given uk ∈ Ωk , k = 0, 1, . . . , (T − 1).

(8.68)

Assumptions required are: (i) F (xk , uk , k) and f (xk , uk , k) are continuously diﬀerentiable in xk for every uk and k. (ii) The sets {−F (x, Ωk , k), f (x, Ωk , k)} are b-directionally convex for every x and k, where b = (−1, 0, . . . , 0). That is, given v and w in Ωk and 0 ≤ λ ≤ 1 , there exists u(λ) ∈ Ωk such that F (x, u(λ), k) ≥ λF (x, v, k) + (1 − λ)F (x, w, k) and f (x, u(λ), k) = λf (x, v, k) + (1 − λ)f (x, w, k) for every x and k. It should be noted that convexity implies bdirectional convexity, but not the converse. (iii) Ωk satisﬁes the Kuhn-Tucker constraint qualiﬁcation. With these assumptions replacing the assumptions of Theorem 8.3, and since there is no salvage value term in (8.67) meaning that S(xT , T ) ≡ 0, the maximum principle (8.53) with λT = 0 holds with control constraint set g(uk , k) ≥ bk replaced by uk ∈ Ω. When the salvage function S(xT , T ) is not identically zero, the objective function in

Exercises for Chapter 8

277

(8.67) is replaced by the Bolza form objective function (8.42). In Exercise 8.20, you are asked to convert the problem deﬁned by (8.42) and (8.68) to its Lagrange form, and then obtain the corresponding assumptions on the salvage value function S(xT , T ) for the results of this section to apply. For a ﬁxed-end-point problem, i.e., when xT is also given in (8.68), the more general maximum principle holds with λT a constant to be determined. Exercise 8.17 is an example of a ﬁxed-end-point problem. Finally, when there are lags in the system dynamics, i.e., when the state of the system in a period depends not only on the state and the control in the previous period, but also on the values of these variables in prior periods, it is easy to adapt the discrete maximum principle to deal with such systems; see Burdet and Sethi (1976). Exercise 8.22 presents an advertising model containing lags in its sales-advertising dynamics. Some concluding remarks on the applications of discrete-time optimal control problems are appropriate. Real-life examples that can be modeled as such problems include the following: payments of principal and interest on loans; harvesting of crops; production planning for monthly demands; etc. Such problems would require eﬃcient computational procedures for their solution. Some references dealing with computational methods for discrete optimal control problems are Murray and Yakowitz (1984), Dunn and Bertsekas (1989), Pantoja and Mayne (1991), Wright (1993), and Dohrmann and Robinett (1999). Another reason that makes the discrete optimal control theory important arises from the fact that computers are being used increasingly in the control of dynamic systems. Finally, Pepyne and Cassandras (1999) have explored an optimal control approach to treat discrete event dynamic systems (DEDS). They also apply the approach to a transportation problem, modeled as a polling system. Exercises for Chapter 8 E 8.1 Determine the critical points of the following functions: (a) h(y, z) = −5y 2 − z 2 + 10y + 6z + 27, (b) h(y, z) = 5y 2 − yz + z 2 − 10y − 18z + 17. E 8.2 Let h be twice diﬀerentiable with its Hessian matrix deﬁned to ¯ be a critical point, i.e., a solution of hx = 0. Let Hj be H = hxx . Let x be the jth principal minor, i.e., the j × j submatrix found in the ﬁrst j

278

8. The Maximum Principal: Discrete Time

rows and the ﬁrst j columns of H. Let |Hj | be the determinant of Hj . Then, y 0 is a local maximum of h if H1 < 0, |H2 | > 0, |H3 | < 0, . . . , (−1)n |Hn | = (−1)n |H| > 0 evaluated at x ¯, and x ¯ is a local minimum of h if H1 > 0, |H2 | > 0, |H3 | > 0, . . . , |Hn | = |H| > 0 evaluated at x ¯. Apply these conditions to Exercise 8.1 to identify local minima and maxima of the functions in (a) and (b). E 8.3 Find the optimal speed in cases (a) and (b) below: (a) During times of an energy crisis, it is important to economize on fuel consumption. Assume that when traveling x mile/hour in high gear, a truck burns fuel at the rate of

1 2500 + x gallons/mile. 500 x If fuel costs 50 cents per gallon, ﬁnd the speed that will minimize the cost of fuel for a 1000 mile trip. Check the second-order condition. (b) When the government imposed this optimal speed in 1974, truck drivers became so angry that they staged blockades on several freeways around the country. To explain the reason for these blockades, we found that a crucial ﬁgure was the hourly wage of the truckers, estimated at $3.90 per hour at that time. Recompute a speed that will minimize the total cost of fuel and the driver’s wages for the same trip. You do not need to check for the second-order condition. E 8.4 Use (8.5)–(8.7) to derive Eq. (8.8). E 8.5 Verify Eq. (8.8) in Example 8.1 by determining h∗ (a) and expanding the function h∗ (10 + ) in a Taylor series around the value 10. E 8.6 Maximize h(x) = (1/3)x3 − 6x2 + 32x + 5 subject to each of the following constraints: (a) x ≤ 6 (b) x ≤ 20.

Exercises for Chapter 8

279

E 8.7 Rework Example 8.4 by replacing (2, 2) with each of the following points: (a) (0, −1) (b) (1/2, 1/2). E 8.8 Add the equality constraint 2x = y to the problem in Example 8.4 and solve it. E 8.9 Solve the problem: ⎧ ⎪ ⎪ ⎪ max h(x, y) ⎪ ⎪ ⎨ subject to ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ x2 ≤ (2 − y)3 , y ≥ 0, for (a) h(x, y) = x + y, (b) h(x, y) = x + 2y, and (c) h(x, y) = x + 3y. Comment on the solution in each of the cases (a), (b), and (c). E 8.10 Constraint Qualiﬁcation. Show that the feasible region in two dimensions, determined by the constraints (1 − x)3 − y ≥ 0, x ≥ 0, and y ≥ 0, does not satisfy the constraint qualiﬁcation (8.36) at the boundary point (1,0). Also sketch the feasible region to see the presence of a cusp at point (1, 0). E 8.11 Constraint Qualiﬁcation. Show that the feasible region in two dimensions, determined by the constraints x2 +y 2 ≤ 1, x ≥ 0, and y ≥ 0, satisﬁes the constraint qualiﬁcation (8.36) at the boundary point (1,0). Also sketch the feasible region to contrast it with that in Exercise 8.10. E 8.12 Solve graphically the problem of minimizing x subject to the constraints 1 − x ≥ 0, y ≥ 0, x3 − y ≥ 0. Show that the constraints do not satisfy the constraint qualiﬁcation (8.36) at the optimal point. E 8.13 Rewrite the maximum principle (8.53) for the special case of the linear Mayer form problem obtained when F ≡ 0 and S(xT , T ) = cxT , where c is an n-component row vector of constants.

280

8. The Maximum Principal: Discrete Time

E 8.14 Show that the necessary conditions for uk to be an optimal solution for (8.52) are given by (8.50) and (8.51). E 8.15 Prove Theorem 8.3. E 8.16 Formulate and solve a discrete-time version of the cash balance model of Sect. 5.1.1. E 8.17 Minimum Fuel Problem. Consider the problem: ⎧ ! T −1 k " ⎪ ⎪ ⎪ min J = ⎪ k=0 |u | ⎪ ⎪ ⎪ ⎪ ⎪ ⎨ subject to ⎪ ⎪ ⎪ xk = Axk + buk , x0 and xT given ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ uk ∈ [−1, 1], k = 0, 1, . . . , T − 1, where A is a given matrix. Obtain the expression for the adjoint variable and the form of the optimal control. E 8.18 Current-Value Formulation. Obtain the current-value formulation of the discrete maximum principle. Assume that r is the discount rate, i.e., 1/(1 + r) is the discount factor. E 8.19 Convert the Bolza form problem (8.42)–(8.44) to the equivalent linear Mayer form; see Sect. 2.1.4 for a similar conversion in the continuous-time case. E 8.20 Convert the problem deﬁned by (8.42) and (8.68) to its Lagrange form. Then, obtain the assumptions on the salvage value function S(xT , T ) so that the results of Sect. 8.3 apply. Under these assumptions, state the maximum principle for the Bolza form problem deﬁned by (8.42) and (8.68). E 8.21 Use Excel to solve the production planning problem given by (8.62) and (8.63) with I 0 = 1, Pˆ = 30, Iˆ = 15, h = c = 1, T = 8, and S k = k 3 − 12k 2 + 32k + 30, k = 0, 1, 2, . . . , (T − 1). This is a discrete time version of Example 6.1 so that you can compare your solution with Fig. 6.1.

Exercises for Chapter 8

281

E 8.22 An Advertising Model (Burdet and Sethi 1976). Let xk denote the sale and uk , k = 1, 2, . . . , T − 1, denote the amount of advertising in period k. Formulate the sales-advertising dynamics as x = −δx + r k

k

k

fkl (xl , ul ), x0 given,

l=0

where δ and r are decay and response constants, respectively, and fkl (xl , ul ) is a nonnegative function that decreases with xl and increases with ul . In the special case when fkl (xl , ul ) = γ lk ul , γ lk > 0, obtain optimal advertising amounts to maximize the total discounted proﬁt given by T −1 (πxk − uk )(1 + ρ)−k , k=1

where, as in Sect. 7.2.1, π denotes per unit sales revenue, ρ denotes the discount rate, and the inequalities 0 ≤ uk ≤ Qk represent the restrictions on the advertising amount uk . For the continuous-time version of problems with lags, see Hartl and Sethi (1984b).

Chapter 9

Maintenance and Replacement The problem of simultaneously determining the lifetime of an asset or an activity along with its management during that lifetime is an important problem in practice. The most typical example is the problem of optimal maintenance and replacement of a machine; see Rapp (1974) and Pierskalla and Voelker (1976). Other examples occur in forest management, such as in N¨aslund (1969), Clark (1976), and Heaps (1984), and in advertising copy management, such as in Pekelman and Sethi (1978). The ﬁrst major work dealing with machine replacement problems appeared in 1949 as a MAPI (Machinery and Applied Products Institute) study by Terborgh (1949). For the most part, this study was conﬁned to those problems where the optimization was carried out only with respect to the replacement lives of the machines under consideration. Boiteux (1955) and Mass´e (1962) extended the single machine replacement problem to include the optimal timing of a partial replacement of the machine before its actual retirement. N¨aslund (1966) was the ﬁrst to solve a generalized version of the Boiteux problem by using the maximum principle. He considered optimal preventive maintenance applied continuously over the entire period instead of a single optimal partial replacement before the machine is retired. Thompson (1968) presented a modiﬁcation of N¨aslund’s model which is described in the following section.

© Springer Nature Switzerland AG 2019 S. P. Sethi, Optimal Control Theory, https://doi.org/10.1007/978-3-319-98237-3 9

283

284

9. Maintenance and Replacement

9.1

A Simple Maintenance and Replacement Model

Consider a single machine whose resale value gradually declines over time. Its output is assumed to be proportional to its resale value. By applying preventive maintenance, it is possible to slow down the rate of decline of the resale value. The control problem consists of simultaneously determining the optimal rate of preventive maintenance and the sale date of the machine. Clearly this is an optimal control problem with unspeciﬁed terminal time; see Sect. 3.1 and Example 3.6.

9.1.1

The Model

In order to deﬁne Thompson’s model, we use the following notation: T

= the sale date of the machine to be determined,

ρ = the constant discount rate, x(t) = the resale value of the machine in dollars at time t; let x(0) = x0 , u(t) = the preventive maintenance rate at time t (maintenance here means money spent over and above the minimum required for necessary repairs), g(t) = the maintenance eﬀectiveness function at time t (measured in dollars added to the resale value per dollar spent on preventive maintenance), d(t) = the obsolescence function at time t (measured in terms of dollars subtracted from x at time t), π = the constant production rate in dollars per unit time per unit resale value; assume π > ρ or else it does not pay to produce. It is assumed that g(t) is a nonincreasing function of time and d(t) is a nondecreasing function of time, and that for all t u(t) ∈ Ω = [0, U ],

(9.1)

where U is a positive constant. The present value of the machine is the sum of two terms, the discounted income (production minus maintenance) stream during its life plus the discounted resale value at T : T [πx(t) − u(t)]e−ρt dt + x(T )e−ρT . (9.2) J= 0

9.1. A Simple Maintenance and Replacement Model

285

The state variable x is aﬀected by the obsolescence factor, the amount of preventive maintenance, and the maintenance eﬀectiveness function. Thus, (9.3) x(t) ˙ = −d(t) + g(t)u(t), x(0) = x0 . In the interests of realism we assume that − d(t) + g(t)U ≤ 0, t ≥ 0.

(9.4)

The assumption implies that preventive maintenance is not so eﬀective as to enhance the resale value of the machine over its previous values; rather, it can at most slow down the decline of the resale value, even when preventive maintenance is performed at the maximum rate U. A modiﬁcation of (9.3) is given in Arora and Lele (1970). See also Hartl (1983b). The optimal control problem is to maximize (9.2) subject to (9.1) and (9.3).

9.1.2

Solution by the Maximum Principle

This problem is similar to Model Type (a) of Table 3.3 with the freeend-point condition as in Row 1 of Table 3.1. Therefore, we follow the steps for solution by the maximum principle stated in Chap. 3. The standard Hamiltonian as formulated in Sect. 2.2 is H = (πx − u)e−ρt + λ(−d + gu),

(9.5)

where the adjoint variable λ satisﬁes λ˙ = −πe−ρt , λ(T ) = e−ρT .

(9.6)

Since T is unspeciﬁed, the required additional terminal condition (3.15) for this problem is (9.7) − ρe−ρT x(T ) = −H, which must hold on the optimal path at time T. The adjoint variable λ can be easily obtained by integrating (9.6), i.e., T π −ρT + πe−ρτ dτ = e−ρT + [e−ρt − e−ρT ]. (9.8) λ(t) = e ρ t The interpretation of λ(t) is as follows. It gives, in present value terms, the marginal proﬁt per dollar of gain in resale value at time t.

286

9. Maintenance and Replacement

The ﬁrst term represents the present value of one dollar of additional salvage value at T brought about by one dollar of additional resale value at the current time t. The second term represents the present value of incremental production from t to T brought about by the extra productivity of the machine due to the additional one dollar of resale value at time t. Since the Hamiltonian is linear in the control variable u, the optimal control for a problem with any ﬁxed T is bang-bang as in Model Type (a) in Table 3.3. Thus,

π u∗ (t) = bang 0, U ; {e−ρT + (e−ρt − e−ρT )}g(t) − e−ρt . ρ

(9.9)

To interpret this optimal policy, we see that the term {e−ρT +

π −ρt (e − e−ρT )}g(t) ρ

is the present value of the marginal return from increasing the preventive maintenance by one dollar at time t. The last term e−ρt in the argument of the bang function is the present value of that one dollar spent for preventive maintenance at time t. Thus, in words, the optimal policy means the following: if the marginal return of one dollar of additional preventive maintenance is more than one dollar, then perform the maximum possible preventive maintenance, otherwise do not perform any at all. To ﬁnd how the optimal control switches, we need to examine the switching function in (9.9). Rewriting it as e

−ρt

πg(t) π − ( − 1)eρ(t−T ) g(t) − 1 ρ ρ

(9.10)

and taking the derivative of the bracketed terms with respect to t, we can conclude that the expression inside the square brackets in (9.10) is monotonically decreasing with time t on account of the assumptions that π/ρ > 1 and that g(t) is nonincreasing with t (see Exercise 9.1). It follows that there will not be a singular control for any ﬁnite interval of time. Furthermore, since e−ρt > 0 for all t, we can conclude that the switching function can only go from positive to negative and not vice versa. Thus, the optimal control will be either U, or zero, or U followed by zero. The switching time ts is obtained as follows: equate (9.10) to zero and solve for t. If the solution is negative, let ts = 0, and if the solution is greater

9.1. A Simple Maintenance and Replacement Model

287

than T, let ts = T, otherwise set ts equal to the solution. It is clear that the optimal control in (9.9) can now be rewritten as ⎧ ⎪ ⎨ U t ≤ ts , ∗ u (t) = (9.11) ⎪ ⎩ 0 t > ts . Note that all of the above calculations were made on the assumption that T was ﬁxed, i.e., without imposing condition (9.7). On an optimal path, this condition, which uses (9.5), (9.7), and (9.8), can be restated as ∗

−ρe−ρT x∗ (T ∗ ) = −{πx∗ (T ∗ ) − u∗ (T ∗ )}e−ρT

∗

∗

(9.12)

−e−ρT {−d(T ∗ ) + g(T ∗ )u(T ∗ )}. This means that when u∗ (T ∗ ) = 0 (i.e., ts < T ∗ ), we have x∗ (T ∗ ) =

d(T ∗ ) , π−ρ

(9.13)

and when u∗ (T ∗ ) = U (i.e., ts = T ∗ ), we have x∗ (T ∗ ) =

d(T ∗ ) − [g(T ∗ ) − 1]U . π−ρ

(9.14)

Since d(t) is nondecreasing, g(t) is nonincreasing, and x(t) is nonincreasing, Eq. (9.13) or Eq. (9.14), whichever the case may be, has a solution for T ∗ .

9.1.3

A Numerical Example

It is instructive to work an example of this model in which speciﬁc values are assumed for the various functions. Examples that illustrate other kinds of qualitatively diﬀerent behavior are left as Exercises 9.3–9.5. Suppose U = 1, x(0) = 100, d(t) = 2, π = 0.1, ρ = 0.05, and g(t) = 2/(1 + t)1/2 . Then (9.3) specializes to 2u(t) , x(0) = 100. x(t) ˙ = −2 + √ 1+t

(9.15)

First, we write the condition on ts by equating (9.10) to 0, which gives ρ s . (9.16) π − (π − ρ)e−ρ(T −t ) = g(ts )

288

9. Maintenance and Replacement

In doing so, we have assumed that the solution of (9.16) lies in the open interval (0, T ). As we will indicate later, special care needs to be exercised if this is not the case. Substituting the data in (9.16) we have 0.1 − 0.05e−0.05(T −t ) = 0.025(1 + ts )1/2 , s

which simpliﬁes to (1 + ts )1/2 = 4 − 2e−0.05(T −t ) . s

(9.17)

Then, integrating (9.15), we ﬁnd x(t) = −2t + 4(1 + t)1/2 + 96, if t ≤ ts , and hence x(t) = −2ts + 4(1 + ts )1/2 + 96 − 2(t − ts ) = 4(1 + ts )1/2 + 96 − 2t, if t > ts . Since we have assumed 0 < ts < T, we substitute x(T ) into (9.13), and obtain 4(1 + ts )1/2 + 96 − 2T = 2/0.05 = 40, which simpliﬁes to

T = 2(1 + ts )1/2 + 28.

(9.18)

We must solve (9.17) and (9.18) simultaneously. Substituting (9.18) into (9.17), we ﬁnd that ts must be a zero of the function h(ts ) = (1 + ts )1/2 − 4 + 2e−[2(1+t

s )1/2 −ts +28]/20

.

(9.19)

A simple binary search program was written to solve this equation, which obtained the value ts = 10.6. Substitution of this into (9.18) yields T = 34.8. Since this satisﬁes our supposition that 0 < ts < T, we can conclude our computations. Thus, if we let the unit of time be 1 month, then the optimal solution is to perform preventive maintenance at the maximum rate during the ﬁrst 10.6 months, and thereafter not at all. The sale date is at 34.8 months after purchase. Figure 9.1 gives the functions x(t) and u(t) for this optimal maintenance and sale date policy. If, on the other hand, the solution of (9.17) and (9.18) did not satisfy our supposition, we would need to follow the procedure outlined earlier in the section. This would result in ts = 0 or ts = T. If ts = 0, we would obtain T from (9.18), and conclude u∗ (t) = 0, 0 ≤ t ≤ T. Alternatively, if ts = T, we would need to substitute x(T ) into (9.14) to obtain T. In this case the optimal control would be u∗ (t) = U, 0 ≤ t ≤ T.

9.1. A Simple Maintenance and Replacement Model

289

Figure 9.1: Optimal maintenance and machine resale value

9.1.4

An Extension

The pure bang-bang result in the model developed above is a result of the linearity in the problem. The result can be enriched as in Sethi (1973b) by generalizing the resale value equation (9.3) as follows: x(t) ˙ = −d(t) + g(u(t), t),

(9.20)

where g is nondecreasing and concave in u. For this section, we will assume the sale date T to be ﬁxed for simplicity and g to be strictly concave in u, i.e., gu ≥ 0 and guu < 0 for all t. Also, gt ≤ 0, gut ≤ 0, and g(0, t) = 0; see Exercise 9.7 for an example of the function g(u, t). The standard Hamiltonian is H = (πx − u)e−ρt + λ[−d + g(u, t)],

(9.21)

where λ is given in (9.8). To maximize the Hamiltonian, we diﬀerentiate it with respect to u and equate the result to zero. Thus, Hu = −e−ρt + λgu = 0.

(9.22)

If we let u0 (t) denote the solution of (9.22), then u0 (t) maximizes the Hamiltonian (9.21) because of the concavity of g in u. Thus, for a ﬁxed T, the optimal control is u∗ (t) = sat[0, U ; u0 (t)].

(9.23)

290

9. Maintenance and Replacement

To determine the direction of change in u∗ (t), we obtain u˙ 0 (t). For this, we use (9.22) and the value λ(t) from (9.8) to obtain gu =

e−ρt = λ(t)

π ρ

−

( πρ

1 . − 1)eρ(t−T )

(9.24)

Since π > ρ, the denominator on the right-hand side of (9.24) is monotonically decreasing with time. Therefore, the right-hand side of (9.24) is increasing with time. Taking the time derivative of (9.24), we have gut + guu u˙ 0 =

ρ2 (π − ρ)eρ(t−T ) > 0. [π − (π − ρ)eρ(t−T ) ]2

But gut ≤ 0 and guu < 0, it is therefore obvious that u˙ 0 (t) < 0. In order now to sketch the optimal control u∗ (t) speciﬁed in (9.23), let us deﬁne 0 ≤ t1 ≤ t2 ≤ T such that u0 (t) ≥ U for t ≤ t1 and u0 (t) ≤ 0 for t ≥ t2 . Then, we can rewrite the sat function in (9.23) as ⎧ ⎪ ⎪ ⎪ U for t ∈ [0, t1 ], ⎪ ⎪ ⎨ (9.25) u∗ (t) = u0 (t) for t ∈ (t1 , t2 ), ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 0 for t ∈ [t2 , T ]. In (9.25), it is possible to have t1 = 0 and/or t2 = T. In Fig. 9.2 we have sketched a case when t1 > 0 and t2 < T. Note that while u0 (t) in Fig. 9.2 is decreasing over time, the way it will decrease will depend on the nature of the function g. Indeed, the shape of u0 (t), while always decreasing, can be quite general. In particular, you will see in Exercise 9.7 that the shape of u0 (t) is concave and, furthermore, u0 (t) > 0, t ≥ 0, so t2 = T in that case.

9.2

Maintenance and Replacement for a Machine Subject to Failure

In Kamien and Schwartz (1971a), a related model is developed which has somewhat diﬀerent assumptions. They assume that the production rate of the machine is independent of its age, while its probability of failure increases with its age. Consistent with this assumption, the purpose of preventive maintenance in the Kamien-Schwartz model is to inﬂuence

9.2. Maintenance & Replacement for a Machine Subject to Failure 291

Figure 9.2: Sat function optimal control the failure rate of the machine rather than arrest the deterioration in the resale value as before. Furthermore, their model also allows for sale of the machine at any time, provided it is still in running condition, and for its disposal as junk if it breaks down for good. The optimal control problem is therefore to ﬁnd an optimal maintenance policy for the period of ownership and an optimal sale date at which the machine should be sold, provided that it has not yet failed. Other references to related models are Alam et al. (1976), Alam and Sarma (1974, 1977), Sarma and Alam (1975), Gaimon and Thompson (1984a, 1989), Dogramaci and Fraiman (2004), Dogramaci (2005), Bensoussan and Sethi (2007), and Bensoussan et al. (2015a).

9.2.1

The Model

In order to deﬁne the Kamien-Schwartz model, we use the following notation: T

= the sale date of a machine to be determined,

u(t) = the preventive maintenance rate at time t; 0 ≤ u(t) ≤ 1, R = the constant positive rate of revenue produced by a functioning machine independent of its age at any time, net of all costs except preventive maintenance,

292

9. Maintenance and Replacement ρ = the constant discount rate,

L = the constant positive junk value of the failed machine independent of its age at failure, B(t) = the (exogenously speciﬁed) resale value of the machine ˙ at time t, if it is still functioning; B(t) ≤ 0, h(t) = the natural failure rate (also termed the natural hazard ˙ rate in the reliability theory); h(t) ≥ 0, h(t) ≥ 0, F (t) = the cumulative probability that the machine has failed by time t, C(u; h) = the cost function depending on the preventive maintenance u when the natural failure rate is h. To make economic sense, an operable machine must be worth at least as much as an inoperable machine and its resale value should not exceed the present value of the potential revenue generated by the machine if it were to function forever. Thus, 0 ≤ L ≤ B(t) ≤ R/ρ, t ≥ 0.

(9.26)

u(t) ∈ Ω = [0, 1].

(9.27)

Also for all t > 0, Finally, when the natural failure rate is h and a controlled failure rate of h(1 − u) is sought, the action of achieving this reduction will cost C(u; h) dollars. For simplicity, we assume that C(u; h) = C(u)h with C(0) = 0, Cu > 0, Cuu > 0, for u ∈ [0, 1].

(9.28)

Thus, the cost of reducing the failure rate increases more than proportionately as the fractional reduction increases. But the cost of a given fractional reduction increases linearly with the natural failure rate. Hence, these conditions imply that a given absolute reduction becomes increasingly more costly as the machine gets older. To derive the state equation for F (t), we note that F˙ /(1−F ) denotes the conditional probability density for the failure of the machine at time t, given that it has survived to time t. This is assumed to depend on two things, namely (i) the natural failure rate that governs the machine in the absence of preventive maintenance, and (ii) the current rate of preventive maintenance. Thus, F˙ (t) = h(t)[1 − u(t)], (9.29) 1 − F (t)

9.2. Maintenance & Replacement for a Machine Subject to Failure 293 which gives the state equation F˙ = h(1 − u)(1 − F ), F (0) = 0.

(9.30)

Thus, the controlled failure rate at time t is h(t)(1 − u(t)). If u = 0, the failure rate assumes its natural value h. As u increases, the failure rate decreases and drops to zero when u = 1. The expected present value of the machine is the sum of the expected present values of (i) the total revenue it produces less the total cost of maintenance, (ii) its junk value if it should fail before it is sold, and (iii) the salvage value if it does not fail and is sold. That is, T ! " J= e−ρt [R − C(u)h](1 − F ) + LF˙ dt + e−ρT B(T )[1 − F (T )]. 0

Using (9.30), we can rewrite J as follows: T e−ρt [R − C(u)h + L(1 − u)h] (1 − F )dt + e−ρT B(T ) [1 − F (T )] . J= 0

(9.31)

The optimal control problem is to maximize J in (9.31) subject to (9.30) and (9.27). Remark T 9.1 In the absence of discounting, the expected junk value term 0 LF˙ (t)dt reduces to LF (T ), i.e., the junk value times the probability that the machine fails by time T. Remark 9.2 While the maintenance and replacement problem of Kamien and Schwartz is stochastic, they formulate and solve it as a deterministic optimal control problem. Bensoussan and Sethi (2007) formulate the underlying stochastic problem as a stochastic optimal control problem, and show how their solution relates to that of the KamienSchwartz model. They also provide a suﬃcient condition for an optimal maintenance and replacement policy.

9.2.2

Optimal Policy

The problem is similar to Model Type (f) in Table 3.3 subject to the free-end-point condition as in Row 1 of Table 3.1. Therefore, we follow the steps for solution by the maximum principle stated in Chap. 3. The standard Hamiltonian is H = e−ρt [R − C(u)h + L(1 − u)h](1 − F ) + λ(1 − u)h(1 − F ), (9.32)

294

9. Maintenance and Replacement

and the adjoint variable satisﬁes ⎧ ⎪ ⎨ λ˙ = e−ρt [R − C(u)h + L(1 − u)h] + λh(1 − u), ⎪ ⎩ λ(T ) = −e−ρT B(T ).

(9.33)

Since T ≥ 0 is also to be decided, we require the additional transversality condition (3.77) for an optimal T ∗ to satisfy. R − C[u∗ (T ∗ )]h(T ∗ ) + L[1 − u∗ (T ∗ )]h(T ∗ )

(9.34)

−[1 − u∗ (T ∗ )]h(T ∗ )B(T ∗ ) − ρB(T ∗ ) + BT (T ∗ ) = 0. In Exercise 9.8, you are asked to derive this condition by using (9.31)– (9.33) in (3.77). While we know from (3.79) that (9.34) has a standard economic interpretation of having zero marginal proﬁt of changing T ∗ , it is still illuminating to ﬂesh out a more detailed interpretation of each term in what looks like a fairly complex expression. A good way to accomplish that is totally what we get if we decide to sell the machine at time T ∗ + δ in comparison to selling it at T ∗ . We will do this only for a small δ > 0, and leave it as Exercise 9.9 for a small δ < 0. First we note that in solving Exercise 9.8 to obtain (9.34) from ∗ (3.77), a simpliﬁcation involved canceling the common factor e−ρT (1 − ∗ F (T ∗ )) > 0. Removing e−ρT brings the revenue and cost terms from present-value dollars to dollars at time T ∗ . The presence of the probability term 1 − F (T ∗ ) means that the machine will be replaced at T ∗ if it has not failed by time T ∗ with that probability. Its removal means that (9.34) can be interpreted as if we are at T ∗ and we ﬁnd the machine to be working, which is tantamount to interpreting (9.34) with F (T ∗ ) = 0. Now consider keeping the machine to T ∗ + δ. Clearly we lose its selling price B(T ∗ ) in doing so. But then we gain the following amounts discounted to time T ∗ : {R − C(u∗ (T ∗ ))h(T ∗ )}δe−ρδ = {R − C(u∗ (T ∗ ))h(T ∗ )}δ + o(δ), (9.35) L(1 − u∗ (T ∗ ))h(T ∗ )δe−ρδ = L(1 − u∗ (T ∗ ))h(T ∗ )δ + o(δ),

(9.36)

B(T ∗ + δ)(1 − F (T ∗ + δ))e−ρδ = B(T ∗ ) − B(T ∗ )(1 − u∗ (T ∗ ))h(T ∗ )δ −ρB(T ∗ )δ + BT (T ∗ )δ + o(δ). (9.37)

9.2. Maintenance & Replacement for a Machine Subject to Failure 295 The RHS of these equations can be obtained by noting that e−ρδ = 1 − ρδ + o(δ), B(T ∗ + δ) = B(T ∗ ) + BT (T ∗ )δ + o(δ) and F (T ∗ + δ) = F (T ∗ ) + F˙ (T ∗ )δ = F (T ∗ ) + (1 − u∗ (T ∗ ))h(T ∗ )(1 − F (T ∗ ))δ = (1 − u∗ (T ∗ ))h(T ∗ )δ + o(δ), since we had set F (T ∗ ) = 0 for interpreting (9.34) upon arrival at T ∗ and ﬁnding the machine to be working. The net gain is the sum of (9.35), (9.36) and (9.37) less B(T ∗ ), where (9.35) gives the net cash ﬂow (revenue—cost of preventative maintenance from T ∗ to T ∗ + δ), (9.36) represents the junk value L multiplied by the probability [1 − u(T ∗ )]h(T ∗ )δ that the machine fails during the short time δ when the machine is found to be working at T ∗ , and (9.37) less B(T ∗ ) has three terms −ρB(T ∗ )δ +BT (T ∗ )δ −B(T ∗ )(1−u∗ (T ∗ ))h(T ∗ )δ: the ﬁrst of which is the loss of interest ρB(T ∗ )δ on the resale value B(T ∗ ) not obtained when deciding to keep the machine to T ∗ + δ, the second term BT (T ∗ ) < 0 is the decrease in the resale value from T ∗ to T ∗ + δ, and the third term represents the loss of the entire resale value if the machine fails with the probability (1 − u∗ (T ∗ ))h(T ∗ )δ given that the machine was found to be working at time T ∗ . Moreover, if we divide the net gain by δ and then let δ → 0, we obtain the marginal proﬁt of keeping the machine from time T ∗ to T ∗ + δ, and setting it equal to zero gives precisely the transversality condition (9.34). If we separate the revenue and cost terms in the resulting expression of the marginal proﬁt, then (9.34) determining the optimal sale date T ∗ is the usual economic condition equating marginal revenue to marginal cost. Next, we analyze the problem to obtain the optimal maintenance policy for a ﬁxed T. If the optimal solution is in the interior, i.e., u∗ ∈ (0, 1), then the Hamiltonian maximizing condition gives Hu = −e−ρt h(1 − F )[Cu + L + eρt λ] = 0.

(9.38)

In the trivial cases in which the natural failure rate h(t) is zero or when the machine fails with certainty by time t (i.e., F (t) = 1), then u∗ (t) = 0. Assume therefore h > 0 and F < 1. Under these conditions, we can infer from (9.28) and (9.38) that ⎫ ⎪ ρt ∗ ⎪ ⎪ (i) Cu (0) + L + λe > 0 ⇒ u (t) = 0, ⎪ ⎪ ⎬ ρt ∗ (9.39) (ii) Cu (1) + L + λe < 0 ⇒ u (t) = 1. ⎪ ⎪ ⎪ ⎪ ⎪ (iii) Otherwise, Cu + L + λeρt = 0determines u∗ (t). ⎭

296

9. Maintenance and Replacement

Using the terminal condition λ(T ) = −e−ρT B(T ) from (9.33), we can derive u∗ (T ) satisfying (9.39): ⎫ ⎪ ∗ ⎪ ⎪ (i) Cu (0) > B(T ) − L and u (T ) = 0, ⎪ ⎪ ⎬ ∗ (9.40) (ii) Cu (1) < B(T ) − L and u (T ) = 1. ⎪ ⎪ ⎪ ⎪ ⎪ (iii) Otherwise, Cu = B(T ) − L ⇒ u∗ (T ). ⎭ Next we determine how u∗ (t) changes over time. Kamien and Schwartz (1971a, 1992) have shown that u∗ (t) is nonincreasing; see Exercise 9.10. Thus, there exists T ≥ t2 ≥ t1 ≥ 0 such that ⎧ ⎪ ⎪ ⎪ 1 for t ∈ [0, t1 ], ⎪ ⎪ ⎨ u∗ (t) = (9.41) u0 (t) for t ∈ (t1 , t2 ), ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 0 for t ∈ (t2 , T ], where u0 (t) is the solution of (9.39)(iii). Clearly, it must also be shown that u˙ 0 (t) ≤ 0 as part of Exercise 9.10. Of course, u∗ (T ) is immediately known from (9.40). If u∗ (T ) ∈ (0, 1), it implies t2 = T ; and if u∗ (T ) = 1, it implies t1 = t2 = T. For this model, the suﬃciency of the maximum principle follows from Theorem 2.1; see Exercise 9.11.

9.2.3

Determination of the Sale Date

For a ﬁxed T, we know that the terminal optimal control u∗ (T ) is determined by (9.40). If this u∗ (T ) also satisﬁes (9.34), we have determined an optimal trajectory as well as the optimal life of the machine. This, of course, is subject to the second-order condition since (9.34) is only a necessary condition for an optimal T ∗ to satisfy. It is clear that the determination of T ∗ , in most cases, will require numerical computations. The algorithm needs only to be a simple search method because it requires consideration of the single variable T. Before we go to the next section, we remark that a business is usually a continuing entity and does not end at the sale date of one machine. Normally, an existing machine will be replaced by another, which in turn will be replaced by another, and so on. The technology of the newer machines will generally be diﬀerent from that of the existing machine. In

9.3. Chain of Machines

297

what follows, we address these issues. We will choose the discrete-time setting and illustrate the use of the discrete-time maximum principle developed in Chap. 8.

9.3

Chain of Machines

We now extend the problem of maintenance and replacement to a chain of machines. By this we mean that given the time periods 0, 1, 2, . . . , T − 1, we begin with a machine purchase at the beginning of period zero. Then, we ﬁnd an optimal number of machines, say , and optimal times 0 < t1 < t2 , . . . , t−1 < t < T of their replacements such that the existing machine will be replaced by a new machine at time tj , j = 1, 2, . . . , . At the end of the horizon deﬁned by the beginning of period T, the last machine purchased will be salvaged. Moreover, the optimal maintenance policy for each of the machines in the chain must be found. Two approaches to this problem have been developed in the literature. The ﬁrst attempts to solve for an inﬁnite horizon (T = ∞) with a simplifying assumption of identical machine lives, i.e., tj − tj−1 = tj+1 − tj

(9.42)

for all j ≥ 1; see Sethi (1973b) as well as Exercise 9.16. In this case = ∞ as well. The second relaxes the assumption (9.42) of identical machine lives, but then, it can only solve a ﬁnite horizon problem involving a ﬁnite chain of machines, i.e., is ﬁnite; see Sethi and Morton (1972) and Tapiero (1973). For a decision horizon formulation of this problem, see Sethi and Chand (1979), Chand and Sethi (1982), and Bylka et al. (1992). In this section, we will deal with the latter problem as analyzed by Sethi and Morton (1972). The problem is solved by a mixed optimization technique. The subproblems dealing with the maintenance policy are solved by appealing to the discrete maximum principle. These subproblem solutions are then incorporated into a Wagner and Whitin (1958) model formulation for solution of the full problem. The procedure is illustrated by a numerical example.

9.3.1

The Model

Consider buying a machine at the beginning of period s and salvaging it at the beginning of period t > s. Let Jst denote the present value of all

298

9. Maintenance and Replacement

net earnings associated with the machine. To calculate Jst we need the following notation: xks = the resale value of the machine at the beginning of period k, k = s, s + 1, . . . , t, k Ps = the production quantity (in dollar value) during period k, k = s, s + 1, . . . , t − 1, k Es = the necessary expense of the ordinary maintenance (in dollars) during period k, Rsk = Psk − Esk , k = s, s + 1, . . . , t − 1, uk = the rate of preventive maintenance (in dollars) during period k, k = s, s + 1, . . . , t − 1, Cs = the cost of purchasing the machine at the beginning of period s, ρ = the periodic discount rate. It is required that 0 ≤ uk ≤ U sk , k ∈ [s, t − 1].

(9.43)

We can calculate Jst in terms of the variables and functions deﬁned above: Jst =

t−1 k=s

Rsk (1+ρ)−k −

t−1

uk (1+ρ)−k −Cs (1+ρ)−s +xts (1+ρ)−t . (9.44)

k=s

We must also have functions that will provide us with the ways in which states change due to the age of the machine and the amount of preventive maintenance. Also, assuming that at time s, the only machines available are those that are up-to-date with respect to the technology prevailing at s, we can subscript these functions by s to reﬂect the eﬀect of the machine’s technology on its state at a later time k. Let Ψs (uk , k) and Φs (uk , k) be such concave functions so that we can write the following state equations: ΔRsk = Rsk+1 − Rsk = Ψs (uk , k), Rss

(9.45)

Δxks = Φs (uk , k), xss = (1 − δ)Cs ,

(9.46)

given, where δ is the fractional depreciation immediately after the purchase of the machine at time s.

9.3. Chain of Machines

299

To convert the problem into the Mayer form, deﬁne Aks =

k−1

Rsi (1 + ρ)−i ,

(9.47)

i=s

Bsk

=

k−1

ui (1 + ρ)−i .

(9.48)

i=s

Using Eqs. (9.47) and (9.48), we can write the optimal control problem as follows: max[Jst = Ats − Bst − Cs (1 + ρ)−s + xts (1 + ρ)−t ] {uk }

(9.49)

subject to ΔAks = Rsk (1 + ρ)−k , Ass = 0, ΔBsk

k

= u (1 + ρ)

−k

,

Bss

= 0,

(9.50) (9.51)

and the constraints (9.45), (9.46), and (9.43).

9.3.2

Solution by the Discrete Maximum Principle

k+1 k+1 k+1 We associate the adjoint variables λk+1 1 , λ2 , λ3 , and λ4 , respectively with the state equations (9.50), (9.51), (9.45), and (9.46). Therefore, the Hamiltonian becomes k+1 k −k k −k H = λk+1 + λk+1 + λk+1 1 Rs (1 + ρ) 2 u (1 + ρ) 3 Ψs + λ4 Φs , (9.52)

where the adjoint variables λ1 , λ2 , λ3 , and λ4 satisfy the following diﬀerence equations and terminal boundary conditions: ∂H = 0, λt1 = 1, ∂Aks ∂H = − = 0, λt2 = −1, ∂Bsk ∂H −k t = − k = −λk+1 1 (1 + ρ) , λ3 = 0, ∂Rs ∂H = − k = 0, λt4 = (1 + ρ)−t . ∂x

Δλk1 = −

(9.53)

Δλk2

(9.54)

Δλk3 Δλk4

(9.55) (9.56)

The solutions of these equations are λk1 = 1,

(9.57)

300

9. Maintenance and Replacement λk2 = −1, t−1 k λ3 = (1 + ρ)−i ,

(9.58) (9.59)

i=k

λk4 = (1 + ρ)−t .

(9.60)

Note that λk1 , λk2 , and λk4 are constants for a ﬁxed machine salvage time t. To apply the maximum principle, we substitute (9.57)–(9.60) into the Hamiltonian (9.52), collect terms containing the control variable uk , and rearrange and decompose H as H = H1 + H2 (uk ),

(9.61)

where H1 is that part of H which is independent of uk and H2 (uk ) = −uk (1 + ρ)−k +

t−1

(1 + ρ)−i Ψs + (1 + ρ)−t Φs .

(9.62)

i=k+1

Next we apply the maximum principle to obtain the necessary condition for the optimal schedule of preventive maintenance expenditures in dollars. The condition of optimality is that H should be a maximum along the optimal path. If uk were unconstrained, this condition, given the concavity of Ψs and Φs , would be equivalent to setting the partial derivative of H with respect to u equal to zero, i.e., Huk = [H2 ]uk = −(1+ρ)−k +(Ψs )uk

t−1

(1+ρ)−i +(Φs )uk (1+ρ)−t = 0.

i=k+1

(9.63) Equation (9.63) is an equation in with the exception of the particular case when Ψs and Φs are linear in uk (which will be treated later in this section). In general, (9.63) may or may not have a unique solution. For our case we will assume Ψs and Φs to be of the form such that they give a unique solution for uk . One such case occurs when Ψs and Φs are quadratic in uk . In this case, (9.63) is linear in uk and can be solved explicitly for a unique solution for uk . Whenever a unique solution does exist, let this be uk

k . uk = Ust

(9.64)

9.3. Chain of Machines

301

The optimal control uk∗ is given ⎧ ⎪ ⎪ ⎪ 0 ⎪ ⎪ ⎨ k uk∗ = Ust ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ U sk

9.3.3

as k ≤ 0, if Ust k ≤ U sk , if 0 ≤ Ust

(9.65)

k ≥ U sk . if Ust

Special Case of Bang-Bang Control

We now treat the special case in which the problem, and therefore H, is linear in the control variable uk . In this case, H can be maximized simply by having the control at its maximum when the coeﬃcient of uk in H is positive, and minimum when it is negative, i.e., the optimal control is of bang-bang type. In our problem, we obtain the special case if Ψs and Φs assume the form Ψs (uk , k) = uk ψ ks (9.66) and Φs (uk , k) = uk φks ,

(9.67)

respectively, where ψ ks and φks are given constants. Then, the coeﬃcient of uk in H, denoted by Ws (k, t), is Ws (k, t) = −(1 + ρ)−k + ψ ks

t−1

(1 + ρ)−i + φts (1 + ρ)−t ,

(9.68)

i=k+1

and the optimal control uk∗ is given by uk∗ = bang[0, U sk ; Ws (k, t)], k = s, s + 1, . . . , t − 1.

9.3.4

(9.69)

Incorporation into the Wagner-Whitin Framework for a Complete Solution

Once uk∗ has been obtained as in (9.65) or (9.69), we can substitute it into (9.45) and (9.46) to obtain Rsk∗ and xk∗ s , which in turn can be used in (9.44) to obtain the optimal value of the objective function denoted ∗ . This can be done for each pair of machine purchase time s and by Jst sale time t > s.

302

9. Maintenance and Replacement

Let gs denote the present value of the proﬁt (discounted to period 0) of an optimal replacement and preventive maintenance policy for periods s, s + 1, . . . , T − 1. Then, gs =

max

t=s+1,...,T

∗ [Jst + gt ], 0 ≤ s ≤ T − 1

(9.70)

with the boundary condition gT = 0.

(9.71)

The value of g0 will give the required maximum. The mixed optimization technique presented here avoids many of the shortcomings of either pure dynamic programming or pure control theory formulations. Since the solution technique used to optimize a given machine represents a submodule of the overall method, the pure dynamic programming approach may be recognized as a special case. It should be advantageous, however, to be able to use a methodology for the submodule that is most eﬃcient for a given particular problem. Previous control theory formulations do not seem to be easily adaptable to the situation of an existing initial machine; see Sethi and Morton (1972) for other similar asymmetries. The mixed technique can also be adapted to the case of probabilistic technological breakthroughs (Exercise 9.17). Here the path of technological growth is assumed to be a tree with probabilities associated with its branches. The subproblems can be solved by using the maximum principle for stochastic networks given in Sethi and Thompson (1977). However, the number of subproblems that must be solved increases rapidly with the number of branches, thus putting computational limitations on the general usefulness of this extension. Another application of the mixed technique has been used by Pekelman and Sethi (1978) to obtain the optimal durations of advertising copies, and the optimal level of advertising expenditures for each copy.

9.3.5

A Numerical Example

To illustrate the procedure, a simple three-period example will be presented and solved for the case where there is no existing machine at time zero. Machines may be bought at times 0, 1, and 2. The cost of a machine bought at time s is assumed to be Cs = 1, 000 + 500s2 .

9.3. Chain of Machines

303

The discount rate, the fractional instantaneous depreciation at purchase, and the maximum preventive maintenance per period are assumed to be ρ = 0.06, δ = 0.25, and U = $100, respectively. Let Rss be the net return (net of necessary maintenance) of a machine purchased at the beginning of period s and operated during period s. We assume R00 = $600, R11 = $1, 000, and R22 = $1, 100. In a period k subsequent to the period s of machine purchase, the returns Rsk , k > s, depend on the preventive maintenance performed on the machine in the periods prior to period k. The incremental return function is given by Ψs (u, k), which we assume to be linear. Speciﬁcally, ΔRsk = Ψs (uk , k) = −ds + as uk , where d0 = 200, d1 = 50, d2 = 100, and as = 0.5 + 0.1s3 . This means that, in the absence of any preventative maintenance, the return in period k on a machine purchased in period s goes down by an amount ds every period from s to k, including s, in which there is no preventive maintenance. This decrease can be oﬀset by an amount proportional to the amount of preventive maintenance. Note that the function Ψs is assumed to be stationary over time in order to simplify the example. Let xks be the salvage value at time k of a machine purchased at s. We assume xss = (1 − δ)Cs = 0.75[1, 000 + 500s2 ]. The incremental salvage value function is given by Δxks = −s Cs + bs uk , ⎧ ⎪ ⎨ 0.1 when s = 0, 1,

where s =

⎪ ⎩ 0.2 when s = 2,

and bs = (0.5 − 0.05s).

304

9. Maintenance and Replacement

That is, the decrease in salvage value is a constant percentage of the purchase price if there is no preventive maintenance. With preventive maintenance, the salvage value can be enhanced by a proportional amount. ∗ be the optimal value of the objective function associated with Let Jst a machine purchased at s and sold at t ≥ s + 1. We will now solve for ∗ , s = 0, 1, 2, and s < t ≤ 3, where t is an integer. Jst Before we proceed, we will as in (9.68) denote by Ws (k, t), the coefﬁcient of uk in the Hamiltonian H, i.e., Ws (k, t) = −(1 + ρ)−k + as

t−1

(1 + ρ)−i + bs (1 + ρ)−t .

(9.72)

i=k+1

The optimal control is given by (9.69). It is noted in passing that Ws (k + 1, t) − Ws (k, t) = (1 + ρ)−(k+1) (ρ − as ), so that sgn[Ws (k + 1, t) − Ws (k, t)] = sgn[ρ − as ].

(9.73)

This implies that ⎧ ⎪ ⎪ ⎪ ≥ 0 if (ρ − as ) > 0, ⎪ ⎪ ⎨ u(k+1)∗ − uk∗ = 0 if (ρ − as ) = 0, ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ ≤ 0 if (ρ − as ) < 0.

(9.74)

In this example ρ − as < 0, which means that if there is a switching in the preventive maintenance trajectory of a machine, the switch must be from $100 to $0. Solution of Subproblems We now solve the subproblems for various values of s and t(s < t) by using the discrete maximum principle. Subproblem: s = 0, t = 1. W0 (0, 1) = −1 + 0.5(1.06)−1 < 0. From (9.69) we have

u0∗ = 0.

9.3. Chain of Machines

305

Now, R00 = 600,

R01 = 600 − 200 = 400,

x00 = 0.75 × 1, 000 = 750,

x10 = 750 − 0.1 × 1, 000 = 650,

∗ J01 = 600 − 1, 000 + 650 × (1.06)−1 = $213.2.

Similar calculations can be carried out for other subproblems. We will list these results. Subproblem: s = 0, t = 2. W0 (0, 2) < 0, W0 (1, 2) < 0, u0∗ = 0,

u1∗ = 0,

∗ = 466.9. J02

Subproblem: s = 0, t = 3. W0 (0, 3) > 0, W0 (1, 3) < 0, W0 (2, 3) < 0, u0∗ = 100,

u1∗ = 100,

u2∗ = 0,

∗ = 639. J03

Subproblem: s = 1, t = 2. W1 (1, 2) < 0, u1∗ = 0, ∗ J12 = 559.9.

Subproblem: s = 1, t = 3. W1 (1, 3) > 0, W1 (2, 3) < 0, u1∗ = 100, ∗ = 1024.2. J13

u2∗ = 0,

306

9. Maintenance and Replacement

Subproblem: s = 2, t = 3. W2 (2, 3) < 0, u2∗ = 0, ∗ J23 = 80.

Wagner-Whitin Solution of the Entire Problem With reference to the dynamic programming equation in (9.70) and (9.71), we have g3 = 0, ∗ g2 = J23 = $80,

∗ ∗ , J12 + g2 ] g1 = max [J13

= max [1024.2, 559.9 + 80] = $1024.2, ∗ ∗ ∗ g0 = max [J03 , J01 + g1 , J02 + g2 ]

= max [639.0, 213.2 + 1024.2, 466.9 + 80] = $1237.4. Now we can summarize the optimal solution. The optimal number of machines is 2, and their optimal purchase times, maintenance rates, and sell times are as follows: First Machine Optimal Policy: Purchase at s = 0 and sell at t = 1. The optimal preventive maintenance policy is u0∗ = 0. Second Machine Optimal Policy: Purchase at s = 1 and sell at t = 3. The optimal preventive maintenance policy is u1∗ = 100, u2∗ = 0. The associated value of the objective function is J ∗ = $1237.4. Exercises for Chapter 9 E 9.1 Show that the bracketed expression in (9.10) is monotonically decreasing in t. E 9.2 Change the values of U and d(t) in Sect. 9.1.3 to the new values U = 1/2 and d(t) = 3 and re-solve the problem. E 9.3 Show for the model in Sect. 9.1.1 that if it is optimal to have the maximum maintenance throughout the life of the machine, then its optimal life T must satisfy g(T ) − 1 ≥ 0. In particular, for the example in Sect. 9.1.3, show T ≤ 3.

Exercises for Chapter 9

307

E 9.4 Re-solve the example in Sect. 9.1.3 with x(0) = 40. E 9.5 Replace the maintenance eﬀectiveness function in Sect. 9.1.3 by g(t) = 2/(16 + t)1/2 and solve the resulting problem. E 9.6 Re-solve Exercise 2.20 when T is unspeciﬁed and it denotes the sale date of the machine to be determined. E 9.7 Let the maintenance eﬀectiveness function in the model of Sect. 9.1.4 be 2u1/2 g(u, t) = . (1 + t)1/2 Derive the formula for u0 (t) for this case. Furthermore, solve the problem with T = 34.8, U = 1, x(0) = 100, d(t) = 2, π = 0.1 and ρ = 0.05, and compare its solution to that of the numerical example in Sect. 9.1.3. Note that the sale date T is assumed to be ﬁxed in Sect. 9.1.4 for simplicity in exposition. E 9.8 Derive the formula in (9.34) by using (3.77). E 9.9 Redo the analysis providing the detailed economic interpretation of (9.34) when selling the machine at time T ∗ + δ, which is earlier than time T ∗ when the small δ < 0. Hint: The salvage value function required in (3.77) for the problem here is S(F (T ), T ) = e−ρT B(T )(1 − F (T )) as given in (9.31). Its partial derivative with respect to T is [−ρe−ρT B(T ) + e−ρT BT (T )(1 − F (T )). E 9.10 To show that the singular control in the third alternative in (9.39) can be sustained, we set dHu /dt = 0 for all t for which a singular control obtains. That is, u0 (t) satisﬁes Cuu u˙ 0 = Cu [ρ + (1 − u0 )h] + ρL − R + C(u0 )h.

(9.75)

Show that u˙ 0 (t) ≤ 0. Furthermore, show that u∗ (t) is nonincreasing over time. E 9.11 For the model of Sect. 9.2, prove that the derived Hamiltonian H is concave in F for each given λ and t, so that the Suﬃciency Theorem 2.1 holds.

308

9. Maintenance and Replacement

E 9.12 A ﬁrm wants to price its product to maximize the stream of discounted proﬁts. If it maximizes current proﬁts, the high price and proﬁts may attract the entry of rivals, which in turn will reduce future proﬁt possibilities. Let the current proﬁt rate R1 (p) be a strictly concave function of price p with R1 (p) < 0. The proﬁt rate that the ﬁrm believes will be available to it after rival entry is R2 < maxp R1 (p) (independent of current price and lower than current monopoly proﬁts). Whether, or when, a rival will enter is not known, but let F (t) denote the probability that entry will occur by time t, with F (0) = 0. The conditional probability density of entry at time t, given its nonoccurrence prior to t, is F˙ (t)/[1 − F (t)]. We assume that this conditional entry probability density is a strictly increasing, convex function h(p) of product price p. This speciﬁcation reﬂects the supposition that as price rises, the proﬁtability of potential entrants of a given size increases and so does their likelihood of entry. Thus, we assume F˙ (t)/[1 − F (t)] = h(p(t)) where h(0) = 0,

h (p) > 0,

h (p) ≥ 0.

Discounting future proﬁts at rate ρ, the ﬁrm seeks a price policy p(t) to ∞ e−ρt {R1 (p(t))[1 − F (t)] + R2 F (t)}dt max 0

subject to F˙ (t) = h(p(t))[1 − F (t)], F (0) = 0. The integrand represents the expected proﬁts at t, composed of R1 if no rival has entered by t, and otherwise R2 . (a) Show that the maximum principle necessary conditions are satisﬁed by p(t) = p∗ , where p∗ is a constant. Obtain the equation satisﬁed by p∗ and show that it has a unique solution. (b) Let pm denote the monopoly price (in the absence of any rival), i.e., R1 (pm ) = maxp R1 (p). Show that p∗ < pm and R1 (pm ) > R1 (p∗ ) > R2 . Provide an intuitive explanation of the result. (c) Verify the suﬃciency condition for optimality by showing that the maximized Hamiltonian is concave.

Exercises for Chapter 9

309

E 9.13 Let us deﬁne the state of a machine to be ‘0’ if it is working and ‘1’ if it is being repaired. Let λ be the breakdown rate and μ be the service rate as in waiting-line theory, so that we have P˙ 0 = −λP0 + μ(1 − P0 ), P0 (0) = 1, where P0 (t) is the probability that the machine is in the state 0 at time t. Let P1 (t) = 1−P0 (t), which is the probability that the machine is in state 1 at time t. This equation along with (9.3) gives us two state equations. In view of the equation for P˙0 , we modify the objective function (9.2) to

T

J= 0

[πx(t)P0 (t) − u(t) − kP1 (t)]e−ρt dt + x(T )e−ρT ,

where k characterizes the additional expenditure rate while the machine is being repaired. Solve this model to obtain the optimal control. See Alam and Sarma (1974). E 9.14 Starting from Ws (k, t) in (9.72), derive the result in (9.74). E 9.15 Extend the Thompson model in Sect. 9.1 to allow for process discontinuities. An example of this type of machine is an airplane assigned to passenger transportation which may, after some deterioration or obsolescence, be assigned to freight transportation before its eventual retirement. Formulate and analyze the problem. See Tapiero (1971). E 9.16 Extend the Thompson model in Sect. 9.1 to allow for a chain of machines with identical lines. See Sethi (1973b) for an analysis of a similar model. E 9.17 Extend the formulation of the Sethi-Morton model in Sect. 9.3 to allow for probabilistic technological breakthroughs. See Sethi and Morton (1972) and Sethi and Thompson (1977).

Chapter 10

Applications to Natural Resources The increase in world population is causing a corresponding increase in the demand for consumption of natural resources. As a consequence the optimal management and utilization of natural resources is becoming increasingly important. There are two main kinds of natural resource models: those involving renewable resources such as ﬁsh, food, timber, etc., and those involving nonrenewable or exhaustible resources such as petroleum, minerals, etc. In Sect. 10.1 we deal with a ﬁshery resource model, the sole owner of which is considered to be a regulatory agency. The management problem of the agency is to control the rate of ﬁshing over time so that an appropriate objective function is maximized over an inﬁnite horizon. A diﬀerential game extension known as the common property ﬁshery resource model is discussed in Sect. 13.2.3. For other applications of optimal control theory to renewable resource models including those involving predator-prey relationships, see Clark (1976), Goh et al. (1974), Jørgensen and Kort (1997), and Munro and Scott (1985). Section 10.2 deals with an optimal forest thinning model, where thinning is the process of removing some trees from a forest to improve its growth rate and quality. An extension to a chain of forests model is presented in Sect. 10.2.3. The ﬁnal model presented in Sect. 10.3 deals with an exhaustible resource such as petroleum, which must be utilized optimally over a

© Springer Nature Switzerland AG 2019 S. P. Sethi, Optimal Control Theory, https://doi.org/10.1007/978-3-319-98237-3 10

311

312

10. Applications to Natural Resources

given horizon under the assumption that when its price reaches a given high threshold, a substitute will be used instead. Therefore, the analysis of this section can also be viewed as a problem of optimally phasing in an expensive substitute.

10.1

The Sole-Owner Fishery Resource Model

With the establishment of 200-mile territorial zones in the ocean for most countries having coastlines, the control of ﬁshing in these zones has become highly regulated by these countries. In this sense, ﬁshing in territorial waters can be considered as a sole owner ﬁshery problem. On the other hand, if the citizens and commercial ﬁshermen of a given country are permitted to ﬁsh freely in their territorial waters, the problem becomes that of an open access ﬁshery. The solutions of these two extreme problems are quite diﬀerent, as will be shown in this section.

10.1.1

The Dynamics of Fishery Models

We introduce the following notation and terminology which is due to Clark (1976): ρ = the discount rate, x(t) = the biomass of ﬁsh population at time t, g(x) = the natural growth function, u(t) = the rate of ﬁshing eﬀort at time t; 0 ≤ u ≤ U, q = the catchability coeﬃcient, p = the unit price of landed ﬁsh, c = the unit cost of eﬀort. Assume that the growth function g is diﬀerentiable and concave, and it satisﬁes g(0) = 0, g(X) = 0, g(x) > 0 for 0 < x < X,

(10.1)

where X denotes the carrying capacity, i.e., the maximum sustainable ﬁsh biomass. The state equation due to Gordon (1954) and Schaefer (1957) is x˙ = g(x) − qux, x(0) = x0 ,

(10.2)

10.1. The Sole-Owner Fishery Resource Model

313

where qux is the catch rate assumed to be proportional to the biomass as well as the rate of ﬁshing eﬀort. The instantaneous proﬁt rate is π(x, u) = pqux − cu = (pqx − c)u.

(10.3)

From (10.1) and (10.2), it follows that x will stay in the closed interval 0 ≤ x ≤ X provided x0 is in the same interval. An open access ﬁshery is one in which exploitation is completely uncontrolled. Gordon (1954) analyzed this model, also known as the Gordon-Schaefer model, and showed that the ﬁshing eﬀort tends to reach an equilibrium, called a bionomic equilibrium, at the level where total revenue equals total cost. In other words, the so-called economic rent is completely dissipated. From (10.3) and (10.2), this level is simply xb =

c g(xb )p and ub = . pq c

(10.4)

Let U > g(c/pq)p/c so that ub is in the interior of [0, U ]. The economic basis for (10.4) is as follows: If the ﬁshing eﬀort u > ub is made, then total costs exceed total revenues so that at least some ﬁshermen will lose money, and eventually some will drop out, thus reducing the level of the ﬁshing eﬀort. On the other hand, if the ﬁshing eﬀort u < ub is made, then total revenues exceed total costs, thereby attracting additional ﬁshermen, and increasing the ﬁshing eﬀort. The Gordon-Schaefer model does not maximize the present value of the total proﬁts that can be obtained from the ﬁsh resources. This is done next.

10.1.2

The Sole Owner Model

The bionomic equilibrium solution obtained from the open access ﬁshery model usually implies severe biological overﬁshing. Suppose a ﬁshing regulatory agency is established to improve the operation of the ﬁshing industry. In determining the objective of the agency, it is convenient to think of it as a sole owner who has complete rights to exploit the ﬁshing resource. It is reasonable to assume that the agency attempts to maximize ∞

J= 0

e−ρt (pqx − c)udt

(10.5)

subject to (10.2). This is the optimal control problem to be solved.

314

10.1.3

10. Applications to Natural Resources

Solution by Green’s Theorem

The solution method presented in this section generalizes the one based on Green’s theorem used in Sect. 7.2.2. Solving (10.2) for u we obtain u=

g(x) − x˙ , qx

which we substitute into (10.3), giving ∞ g(x) − x˙ J= e−ρt (pqx − c) dt. qx 0

(10.6)

(10.7)

Rewriting, we have

∞

J=

e−ρt [M (x) + N (x)x]dt, ˙

(10.8)

0

where N (x) = −p +

c c and M (x) = (p − )g(x). qx qx

(10.9)

We note that we can write xdt ˙ = dx so that (10.8) becomes the following line integral [e−ρt M (x)dt + e−ρt N (x)dx],

JB =

(10.10)

B

where B is a state trajectory in (x, t) space, t ∈ [0, ∞). In this section we are only interested in the inﬁnite horizon solution. The Green’s theorem method achieves such a solution by ﬁrst solving a ﬁnite horizon problem as in Sect. 7.2.2, and then determining the inﬁnite horizon solution for which you are asked to verify that the maximum principle holds in Exercise 10.1. See also Sethi (1977b). In order to apply Green’s Theorem to (10.10), let Γ denote a simple closed curve in the (x, t) space surrounding a region R in the space. Then, , JΓ = [e−ρt M (x)dt + e−ρt N (x)dx] Γ ∂ −ρt ∂ −ρt [e N (x)] − [e M (x)] dtdx = ∂x R ∂t = −e−ρt [ρN (x) + M (x)]dtdx. (10.11) R

10.1. The Sole-Owner Fishery Resource Model

315

If we let I(x) = −[ρN (x) + M (x)] cg(x) c , = (ρ − g (x))(p − ) − qx qx2 we can rewrite (10.11) as JΓ =

e−ρt I(x)dtdx. R

We can now conclude, as we did in Sects. 7.2.2 and 7.2.4, that the turnpike level x ¯ is given by setting the integrand of (10.11) to zero. That is, − I(x) = [g (x) − ρ](p −

cg(x) c )+ = 0. qx qx2

(10.12)

In addition, a second-order condition must be satisﬁed for the solution x ¯ of (10.12) to be a turnpike solution; see Lemma 7.1 and the subsequent discussion there. The required second-order condition can be stated as I(x) < 0 for x < x ¯ and I(x) > 0 for x > x ¯. Let x ¯ be the unique solution to (10.12) satisfying the second-order condition. The procedure can be extended to the case of nonunique solutions as in Sethi (1977b); see Appendix D.8 on the Sethi-Skiba points. The corresponding value u ¯ of the control which would maintain the ﬁsh stock level at x ¯ is g(¯ x)/q x ¯. In Exercise 10.2 you are asked to show that x ¯ ∈ (xb , X) and also that u ¯ < U. In Fig. 10.1 optimal trajectories are shown for two diﬀerent initial values: x0 < x ¯ and x0 > x ¯. Let g(x)(pqx − c) π(x) = . (10.13) qx With π (x) obtained from (10.13), condition (10.12) can be rewritten as dπ(x) pqx − c =ρ , (10.14) dx qx which facilitates the following economic interpretations. The interpretation of π(x) is that it is the sustainable economic rent at ﬁsh stock level x. This can be seen by substituting u = g(x)/qx into (10.3), where u = g(x)/qx, obtained using (10.2), is the ﬁshing eﬀort required to maintain the ﬁsh stock at level x. Suppose we have attained

316

10. Applications to Natural Resources

Figure 10.1: Optimal policy for the sole owner ﬁshery model the equilibrium level x ¯ given by (10.12), and suppose we reduce this level to x ¯ − ε by removing ε amount of ﬁsh instantaneously from the ﬁshery, which can be accomplished by an impulse ﬁshing eﬀort of ε/q x ¯. The immediate marginal revenue MR from this action is MR = (pq¯ x − c)

ε . q¯ x

However, this causes a decrease in the sustainable economic rent which equals π (¯ x)ε. Over the inﬁnite future, the present value of this stream is ∞ x)ε π (¯ . e−ρt π (¯ x)εdt = ρ 0 Adding to this the cost cε/q¯ x of the additional ﬁshing eﬀort ε/q x ¯, we get the marginal cost MC =

x)ε cε π (¯ + . ρ qx ¯

Equating MR and MC, we obtain (10.14), which is also (10.12). When the discount rate ρ = 0, Eq. (10.14) reduces to π (x) = 0,

10.2. An Optimal Forest Thinning Model

317

so that it gives the equilibrium ﬁsh stock level x ¯ |ρ=0 . On account of this level satisfying the above ﬁrst-order condition, one can show that it maximizes the instantaneous proﬁt rate π(x). In economics, such a level is called the golden rule level. On the other hand, when ρ = ∞, we can conclude from (10.12) that pqx − c = 0. This gives x ¯ |ρ=∞ = xb = c/pq. The latter is the bionomic equilibrium attained in the open access ﬁshery solution; see (10.4). Finally, by denoting x ¯ obtained from (10.12) for any given ρ > 0 as x ¯ |ρ , you are asked in Exercise 10.3 to show that ¯ |ρ>0 > x ¯ |ρ=∞ = xb . x ¯ |ρ=0 > x

(10.15)

The sole owner solution x ¯ satisﬁes x ¯ > xb = c/pq. If we regard a government regulatory agency as the sole owner responsible for operating the ﬁshery at level x ¯, then it can impose restrictions, such as gear regulations, catch limitations, etc. that will increase the ﬁshing cost c. If c is increased to the level pq¯ x, then the ﬁshery can be turned into an open access ﬁshery subject to those regulations, and it will attain the bionomic equilibrium at level x ¯.

10.2

An Optimal Forest Thinning Model

Forests are another important kind of renewable natural resource, and their optimal management is becoming a signiﬁcant current problem. In Kilkki and Vaisanen (1969), a model is developed for forest growth and thinning in connection with Scotch Pine forests in Finland. Thinning is the process of removing some but not all of the trees prior to clearcutting the forest. Besides yielding a harvest of wood, the thinning process also improves the growth rate and quality of the forest. The solution method employed by Kilkki and Vaisanen was based on dynamic programming. We will use the maximum principle approach to solve the model. For related literature, see Clark (1976) and Bowes and Krutilla (1985).

10.2.1

The Forestry Model

We introduce the following notation: t0 = the initial age of the forest, ρ = the discount rate, x(t) = the volume of usable timber in the forest at time t,

318

10. Applications to Natural Resources u(t) = the rate of thinning at time t, p = the constant price per unit volume of timber, c = the constant cost per unit volume of thinning, f (x) = the growth function, which is positive, concave, and has a unique maximum at xm ; we assume f (0) = 0, g(t) = the growth coeﬃcient which is a positive, decreasing function of time.

The speciﬁc function form for the forest growth used in Kilkki and Vaisanen (1969) is as follows: f (x) = xe−αx , 0 ≤ x ≤

1 , α

where α is a positive constant. Note that f is increasing and concave in the relevant range, and it takes it maximum at 1/α. They use the growth coeﬃcient of the form g(t) = at−b , where a and b are positive constants. The forest growth equation is x˙ = g(t)f (x) − u(t), x(t0 ) = x0 . The objective is to maximize the discounted proﬁt ∞ e−ρt (p − c)udt J=

(10.16)

(10.17)

t0

subject to (10.16) and the state and control constraints x(t) ≥ 0 and u(t) ≥ 0.

(10.18)

The control constraint in (10.18) implies that there is no replanting in the forest. In Sect. 10.2.3 we extend this model to incorporate the successive replantings of the forest each time it is clearcut.

10.2.2

Determination of Optimal Thinning

We solve the forest thinning model by using the maximum principle. The Hamiltonian is H = (p − c)u + λ[gf (x) − u]

(10.19)

10.2. An Optimal Forest Thinning Model

319

with the adjoint equation λ˙ = λ[ρ − gf (x)].

(10.20)

u∗ = bang[0, ∞; p − c − λ].

(10.21)

The optimal control is

The appearance of ∞ as an upper bound in (10.21) simply means that impulse control is permitted. We do not use the Lagrangian form of the maximum principle to include constraints (10.18) because, as we will see, the forestry problem has a natural ending at a time T for which x(T ) = 0. ¯ u To get the singular control solution triple {¯ x, λ, ¯}, we must observe that due to the time dependence of g(t), x ¯ and u ¯ will be functions of time. From (10.21), we have ¯ = p − c, λ

(10.22)

which is a constant so that λ˙ = 0. From (10.20), x(t)) = f (¯

ρ or x ¯(t) = f −1 (ρ/g(t)). g(t)

(10.23)

Then, from (10.14), u ¯(t) = g(t)f (¯ x(t)) − x ¯˙ (t)

(10.24)

gives the singular control. The solution of (10.23) can be illustrated as in Fig. 10.2. Since g(t) is a decreasing function of time, it is clear from Fig. 10.2 that x ¯(t) is a decreasing function of time, and then by (10.24), u ¯(t) ≥ 0. It is also clear from (10.23) that x ¯(Tˆ) = 0 at time Tˆ, where Tˆ is given by ρ = f (0), ˆ g(T ) which, in view of f (0) = 1, gives Tˆ = e−(1/b) ln(ρ/a) .

(10.25)

In Fig. 10.3 we plot x ¯(t) as a function of time t. The ﬁgure also contains an optimal control trajectory for the case in which x0 < x ¯(t0 ). ˆ To determine the switching time t, we ﬁrst solve (10.14) with u = 0. Let x(t) be the solution. Then, tˆ is the time at which the x(t) trajectory intersects the x ¯(t) curve; see Fig. 10.3.

320

10. Applications to Natural Resources

Figure 10.2: Singular usable timber volume x ¯(t)

Figure 10.3: Optimal thinning u∗ (t) and timber volume x∗ (t) for the forest thinning model when x0 < x ¯(t0 )

10.2. An Optimal Forest Thinning Model

321

¯(t0 ), the optimal control at t0 will be the impulse cutting For x0 > x ¯(t0 ) instantaneously. To complete the to bring the level from x0 to x inﬁnite horizon solution, set u∗ (t) = 0 for t ≥ Tˆ. In Exercise 10.12 you are asked to obtain λ(t) for t ∈ [0, ∞).

10.2.3

A Chain of Forests Model

We now extend the model of Sect. 10.2.1 to incorporate successive replantings of the forest each time it is clearcut. This extension is similar in spirit to the chain of machines model of Sect. 9.3, but with some important diﬀerences. We will assume that successive plantings, sometimes called forest rotations, take place at equal intervals. This is similar to what was assumed in the machine replacement problem treated in Sethi (1973b). Let T be the rotation period, i.e., the time from planting to clearcutting which is to be determined. During the nth rotation, the dynamics of the forest is given by (10.17) with t ∈ [(n−1)T, nT ] and x[(n−1)T ] = 0. The discounted proﬁt to be maximized is given by

J(T ) =

∞

e

(k−1)ρT

k=1

=

1 1 − e−ρT

T

0

T 0

e−ρt (p − c)udt

e−ρt (p − c)udt.

(10.26)

From the solution of the model in the previous section, and the assumption that the forest is proﬁtable, it is obvious that 0 ≤ T ≤ Tˆ as shown in Fig. 10.4. We have two cases to consider, depending on whether T > tˆ or T ≤ tˆ. Case 1: T > tˆ. From the preceding section it is easy to conclude that the optimal trajectory is as shown in Fig. 10.4. Using the turnpike terminology of Chap. 7, the trajectory from 0 to A is the entry ramp to the turnpike, the trajectory from A to B is on the turnpike, and the trajectory from B to T is the exit ramp. Since u∗ (t) = 0 on the entry ramp, no timber is collected from time 0 to time tˆ. Timber is, however, collected by thinning from time tˆ to T − and clearcutting at time T. Note from Fig. 10.4 that x ¯(T ) is the amount of timber collected from impulse clearcutting u∗ (T ) = imp[¯ x(T ), 0; T ] at time T. Thus, we can write the

322

10. Applications to Natural Resources

Figure 10.4: Optimal thinning u∗ (t) and timber volume x∗ (t) for the chain of forests model when T > tˆ discounted proﬁt J ∗ (T ) of (10.26) for a given T as − T 1 ∗ e−ρt (p − c)¯ u(t)dt + e−ρT (p − c)¯ x(T ) . J (T ) = 1 − e−ρT ˆ t (10.27) Formally, the second term inside the brackets above represents

T

T−

e−ρt (p − c) imp[¯ x(t), 0; t]dt,

(10.28)

the value of clearcutting at time T. In Exercise 10.13, you are asked to show that this value is precisely the second term. For ﬁnding the optimal value of T in this case, we diﬀerentiate (10.27) with respect to T, equate the result to zero, and simplify to obtain (see Exercise 10.14) (1 − e−ρT )g(T )f [¯ x(T )] − ρ¯ x(T ) − ρ

tˆ

T−

e−ρt u ¯(t)dt = 0.

(10.29)

10.2. An Optimal Forest Thinning Model

323

If the solution T lies in (tˆ, Tˆ], keep it; otherwise set T = Tˆ. Note that (10.29) can also be derived by using the transversality condition (3.15); see Exercise 3.6. Case 2: T ≤ tˆ. The optimal trajectory in this case is as shown in Fig. 10.5. In the Vidale-Wolfe advertising model of Chap. 7, a similar case occurs when T is small; see Fig. 7.10 and compare it with Fig. 10.5. The solution for x(T ) is obtained by integrating (10.14) with u = 0 and x0 = 0. Let this solution be denoted as x∗ (t). Here (10.26) becomes J ∗ (T ) =

e−ρT (p − c)¯ x(T ). 1 − e−ρT

(10.30)

To ﬁnd the optimal value of T for this case, we diﬀerentiate (10.30) with respect to T and equate dJ ∗ (T )/dT to zero. We obtain (see Exercise 10.14) (1 − e−ρT )g(T )f [¯ x(T )] − ρ¯ x(T ) = 0. (10.31) If the solution lies in the interval [0, tˆ] keep it; otherwise set T = tˆ.

Figure 10.5: Optimal thinning and timber volume x∗ (t) for the chain of forests model when T ≤ tˆ The optimal value T ∗ can be obtained by computing J ∗ (T ) from both cases and selecting whichever is larger; see also N¨aslund (1969) and Sethi (1973c).

324

10.3

10. Applications to Natural Resources

An Exhaustible Resource Model

In the previous two sections we discussed two renewable resource models. However, many natural resources are nonrenewable or exhaustible. Examples are petroleum, mineral deposits, coal, etc. Given the growing energy shortage, the optimal production and use of these resources is of immense importance to the world. The earliest important work in this area is Hotelling (1931). Since then, a number of studies have been published such as Dasgupta and Heal (1974a), Solow (1974), Weinstein and Zeckhauser (1975), Pindyck (1978a,b), Derzko and Sethi (1981a,b), Amit (1986) and Heal (1993). In this section, we discuss a simple model taken from a paper by Sethi (1979a). The paper obtains the optimal depletion rate of an exhaustible resource that maximizes a social welfare function involving consumers’ surplus and producers’ surplus with various weights. Here we treat the special case when these weights are equal.

Figure 10.6: The demand function

10.3.1

Formulation of the Model

The model will be developed under the assumption that at a high enough price, say p¯, a substitute, preferably renewable, will become available. For example, if the price of fossil fuel becomes suﬃciently high, solar energy may become an economic substitute. In the North American

10.3. An Exhaustible Resource Model

325

context, the resource under consideration could be crude oil and its expensive substitute could be coal and/or tar sands; see, e.g., Fuller and Vickson (1987). We introduce the following notation: p(t) = the price of the resource at time t, q = f (p) is the demand function, i.e., the quantity demanded at price p; f ≤ 0, f (p) > 0 for p < p¯, and f (p) = 0 for p ≥ p¯, where p¯ is the price at which the substitute completely replaces the resource. A typical graph of the demand function is shown in Fig. 10.6, c = G(q) is the cost function; G(0) = 0, G(q) > 0 for q > 0, G > 0 and G ≥ 0 for q ≥ 0, and G (0) < p¯. The latter assumption makes it possible for the producers to make a positive proﬁt at a price p below p¯, Q(t) = the available stock or reserve of the resource at time t; Q(0) = Q0 > 0, ρ = the social discount rate; ρ > 0, T

= the horizon time, which is the latest time at which the substitute will become available regardless of the price of the natural resource; T > 0.

Before stating the optimal control problem, we need the following additional deﬁnitions and assumptions. Let c = G[f (p)] = g(p),

(10.32)

for which it is obvious that g(p) > 0 for p < p¯ and g(p) = 0 for p ≥ p¯. Let π(p) = pf (p) − g(p)

(10.33)

denote the proﬁt function of the producers, i.e., the producers’ surplus. Let p be the smallest price at which π(p) is nonnegative. Assume further that π(p) is a concave function in the range [p, p¯] as shown in Fig. 10.7. In the ﬁgure the point pm indicates the price which maximizes π(p).

326

10. Applications to Natural Resources

Figure 10.7: The proﬁt function

We also deﬁne

p¯

f (y)dy

ψ(p) =

(10.34)

p

as the consumers’ surplus, i.e., the area shown shaded in Fig. 10.6. This quantity represents the total excess amount that consumers would be willing to pay. In other words, consumers actually pay pf (p), while they would be willing to pay p yf (y)dy = pf (p) + ψ(p). p¯

The instantaneous rate of consumers’ surplus and producers’ surplus is the sum ψ(p)+π(p). Let pˆ denote the maximum of this sum, i.e., pˆ solves ψ (ˆ p) + π (ˆ p) = pˆf (ˆ p) − g (ˆ p) = 0.

(10.35)

In Exercise 10.16 you will be asked to show that pˆ < pm , as marked in Fig. 10.7. Later we will show that the correct second-order conditions hold at pˆ. The optimal control problem is: T −ρt max J = [ψ(p) + π(p)]e dt (10.36) 0

subject to Q˙ = −f (p), Q(0) = Q0 ,

(10.37)

Q(T ) ≥ 0,

(10.38)

and p ∈ Ω = [p, p¯]. Recall that the sum ψ(p) + π(p) is concave in p.

10.3. An Exhaustible Resource Model

10.3.2

327

Solution by the Maximum Principle

Form the current-value Hamiltonian H(Q, p, λ) = ψ(p) + π(p) + λ[−f (p)],

(10.39)

where λ satisﬁes the relation λ˙ = ρλ, λ(T ) ≥ 0, λ(T )Q(T ) = 0, which implies ⎧ ⎪ ⎨ 0 λ(t) = ⎪ ⎩ λ(T )eρ(t−T )

if Q(T ) ≥ 0 is not binding,

(10.40)

(10.41)

if Q(T ) ≥ 0 is binding.

To obtain the optimal control, the Hamiltonian maximizing condition, which is both necessary and suﬃcient in this case (see Theorem 2.1), is ∂H = ψ + π − λf = (p − λ)f − g = 0. ∂p

(10.42)

To show that the solution s(λ) for p of (10.42) actually maximizes the Hamiltonian, it is enough to show that the second derivative of the Hamiltonian is negative at s(λ). Diﬀerentiating (10.42) gives ∂2H = f − g + (p − λ)f . ∂p2 Using (10.42) we have ∂2H g = f − g + f . 2 ∂p f

(10.43)

From the deﬁnition of G in (10.32), we can obtain G =

f g − g f , f 3

which, when substituted into (10.43), gives ∂2H = f − G f 2 . ∂p2

(10.44)

328

10. Applications to Natural Resources

The right-hand side of (10.44) is strictly negative because f < 0, and G ≥ 0 by assumption. We remark that pˆ = s(0) using (10.35) and (10.42), and hence the second-order condition for pˆ of (10.35) to give the maximum of H is veriﬁed. In Exercise 10.17 you are asked to show that s(λ) increases from pˆ as λ increases from 0, and that s(λ) = p¯ when λ = p¯ − G (0). Case 1: The constraint Q(T ) ≥ 0 is not binding. From (10.41), λ(t) ≡ 0 so that from (10.42) and (10.35), p∗ = pˆ.

(10.45)

With this value, the total consumption of the resource is T f (ˆ p), which must be ≤ Q0 so that the constraint Q(T ) ≥ 0 is not binding. Hence, T f (ˆ p ) ≤ Q0

(10.46)

characterizes Case 1 and its solution is given in (10.45). Case 2: T f (ˆ p) > Q0 so that the constraint Q(T ) ≥ 0 is binding. Obtaining the solution requires ﬁnding a value of λ(T ) such that t∗ f (s[λ(T )eρ(t−T ) ])dt = Q0 , (10.47) 0

where

p¯ − G (0) 1 t = min T, T + ln . p λ(T ) ∗

The time t∗ , if it is less than T, is the time at which s[λ(T )eρ(t From Exercise 10.17, λ(T )eρ(t

∗ −T )

= p¯ − G (0)

(10.48) ∗ −T )

] = p¯.

(10.49)

which, when solved for t∗ , gives the second argument of (10.48). One method to obtain the optimal solution is to deﬁne T¯ as the longest time horizon during which the resource can be optimally used. Such a T¯ must satisfy λ(T¯) = p¯ − G (0), and therefore, 0

T¯

' % &( ¯ f s {¯ p − G (0)}eρ(t−T ) dt = Q0 ,

(10.50)

10.3. An Exhaustible Resource Model

329

which is a transcendental equation for T¯. We now have two subcases. Subcase 2a: T ≥ T¯. The optimal control is ⎧ ' ( ¯ ⎪ ⎨ s {¯ p − G (0)}eρ(t−T ) for t ≤ T¯, p∗ (t) = ⎪ ⎩ p¯ for t > T¯.

(10.51)

Clearly in this subcase, t∗ = T¯ and ¯

λ(T ) = [¯ p − G (0)]e−ρ(T −T ) . A sketch of (10.51) is shown in Fig. 10.8.

Figure 10.8: Optimal price trajectory for T ≥ T¯ Subcase 2b: T < T¯. Here the optimal price trajectory is % & p∗ (t) = s λ(T )eρ(t−T ) ,

(10.52)

where λ(T ) is to be obtained from the transcendental equation T ' % &( f s λ(T )eρ(t−T ) dt = Q0 . (10.53) 0

A sketch of (10.52) is shown in Fig. 10.9. In Exercise 10.18 you are given speciﬁc functions for the exhaustible resource model and asked to work out explicit optimal price trajectories for the model.

330

10. Applications to Natural Resources

Figure 10.9: Optimal price trajectory for T < T¯

Exercises for Chapter 10 E 10.1 As an alternate derivation for the turnpike level x ¯ of (10.12), use the maximum principle to obtain the optimal long-run stationary ¯ equilibrium triple {¯ x, u ¯, λ}. E 10.2 Prove that x ¯ ∈ (xb , X) and u ¯ < U, where x ¯ is the solution of (10.12) and xb is given in (10.4). E 10.3 Show that x ¯ obtained from (10.12) decreases as ρ increases. Furthermore, derive the relation (10.15). E 10.4 Obtain the turnpike level x ¯ of (10.12) for the special case g(x) = x(1 − x), p = 2, c = q = 1, and ρ = 0.1. E 10.5 Perform the following: (a) For the Schaefer model with g(x) = rx(1 − x/X) and q = 1, derive the formula for the turnpike level x ¯ of (10.12). (b) Allen (1973) and Clark (1976) estimated the parameters of the Schaefer model for the Antarctic ﬁn-whale population as follows: r = 0.08, X = 400, 000 whales, and xb = 40, 000. Solve for x ¯ for ρ = 0, 0.10, and ∞. E 10.6 Obtain π (x) from (10.13) and use it in (10.12) to derive (10.14).

Exercises for Chapter 10

331

E 10.7 Let π(x, u) = [p − c(x)](qux) in (10.3), where c(x) is a diﬀerentiable, decreasing, and convex function. Derive an expression for x ¯ satisfying an equation corresponding to (10.12). E 10.8 Show that extinction is optimal if ∞ > p ≥ c(0) and ρ > 2g (0) in Exercise 10.7. Hint: Use the generalized mean value theorem. E 10.9 Let the constant price p in Exercise 10.7 be replaced by a time dependent price p(t) which is diﬀerentiable with respect to t. Derive the equation x ¯ corresponding to (10.12) for this nonautonomous problem. Furthermore, ﬁnd the turnpike level x ¯(t) satisfying the derived equation. E 10.10 Let π(x, u) of Exercise 10.7 be π(x, u) = [p − c(x)](qux) + V (x), where V (x) with V (x) > 0 is the conservation value function, which measures the value to society of having a large ﬁsh stock. By deriving the analogue to (10.12), show that the new x ¯ is larger than the x ¯ in Exercise 10.7. E 10.11 When c(x) = 0 in Exercise 10.9, show that the analogue to (10.12) reduces to p˙ g (x) = ρ − . p Give an economic interpretation of this equation. E 10.12 Find λ(t), t ∈ [0, ∞), for the inﬁnite horizon model of Sect. 10.2.2. E 10.13 Derive the second term inside the brackets of (10.27) by comx(T ), 0; T ]. puting e−ρT (p − c) imp[¯ E 10.14 Derive (10.29) by using the ﬁrst-order condition for maximizing J ∗ (T ) of (10.27) with respect to T. Similarly, derive (10.31). E 10.15 Forest Fertilization Model (N¨aslund 1969). Consider a forestry model in which thinning is not allowed, and the forest is to be clearcut

332

10. Applications to Natural Resources

at a ﬁxed time T. Suppose v(t) ≥ 0 is the rate of fertilization at time t, so that the growth equation is x˙ = r(X − x) + f (v, t), x(0) = x0 , where x is the volume of timber, r and X are positive constants, and f is an increasing, diﬀerentiable, concave function of v. The objective is to maximize J = −c

T

e−ρt v(t)dt + e−ρT px(T ),

0

where p is the price of a unit of timber and c is the unit cost of fertilization. (a) Show that the optimal control v ∗ (t) is given by solving the equation ∂f c = e−(ρ+r)(t−T ) . ∂v p Check that the second order condition for a maximum holds for this v ∗ (t). (b) If f (v) = (1 + t) ln(1 + v), then ﬁnd explicitly the optimal control v ∗ (t) under the assumption that p/c > e(ρ+r)T . Show further that v ∗ (t) is increasing and convex in t ∈ [0, T ]. E 10.16 Show that pˆ deﬁned in (10.35) satisﬁes p ≤ pˆ ≤ pm . E 10.17 Show that s(λ), the solution of (10.39), increases from pˆ as λ increases from 0. Also show that s(λ) = p¯, when λ = p¯ − G (0). E 10.18 For the model of Sect. 10.3, assume

f (p) =

⎧ ⎪ ⎨ p¯ − p for p ≤ p¯, ⎪ ⎩ 0

for p > p¯,

G(q) = q 2 . p/3 if T ≤ 3Q0 /¯ p. (a) Show that p∗ = 2¯

Exercises for Chapter 10

333

¯ p. Moreover, (b) Show that T¯ satisﬁes T¯ + e−ρT /ρ = 1/ρ + 3Q0 /¯ ⎧ ' ( ⎪ ⎨ p¯ eρ(t−T¯) + 2 /3 if t ≤ T¯, p∗ (t) = ⎪ ⎩ p¯ if t > T¯,

for T ≥ T¯, and p∗ (t) = for T > T¯.

2¯ p ρ[¯ pT − 3Q0 ] + −ρt ρT 3 3e (e − 1)

Chapter 11

Applications to Economics Optimal control theory has been extensively applied to the solution of economic problems since the early papers that appeared in Shell (1967) and the works of Arrow (1968) and Shell (1969). The ﬁeld is too vast to be surveyed in detail here, however. Several books in the area are: Arrow and Kurz (1970), Hadley and Kemp (1971), Takayama (1974), Lesourne and Leban (1982), Seierstad and Sydsæter (1987), Feichtinger (1988), L´eonard and Long (1992), Van Hilten et al. (1993), Kamien and Schwartz (1992), and Dockner et al. (2000), and Weber (2011). We content ourselves with the discussion of three simple kinds of models. In Sect. 11.1, two capital accumulation or economic growth models are presented. In Sect. 11.2, we formulate and solve an epidemic control model. Finally, in Sect. 11.3 we discuss a pollution control model.

11.1

Models of Optimal Economic Growth

In this section we develop two simple models of economic growth or capital accumulation. The earliest such model was developed by Ramsey (1928) for an economy having a stationary population; see Exercise 11.7 for one of his models. The ﬁrst model treated in Sect. 11.1.1 is a ﬁnite horizon ﬁxed-endpoint model with a stationary population. The problem is to maximize the present value of the utility of consumption for the society, as well as to accumulate a speciﬁed capital stock by the end of the horizon. The second model incorporates an exogenously and exponentially © Springer Nature Switzerland AG 2019 S. P. Sethi, Optimal Control Theory, https://doi.org/10.1007/978-3-319-98237-3 11

335

336

11. Applications to Economics

growing population in the inﬁnite horizon setting. A technique known as the method of phase diagrams is used to analyze the model. For related discussion and extensions of these models, see Arrow and Kurz (1970), Burmeister and Dobell (1970), Intriligator (1971), and Arrow et al. (2007, 2010).

11.1.1

An Optimal Capital Accumulation Model

Consider a one-sector economy in which the stock of capital, denoted by K(t), is the only factor of production. Let F (K) be the output rate of the economy when K is the capital stock. Assume F (0) = 0, F (K) > 0, F (K) > 0, and F (K) < 0, for K > 0. These conditions imply the diminishing marginal productivity of capital as well as the strict concavity of F (K) in K. A part of this output is consumed and the remainder is reinvested for further accumulation of capital stock. Let C(t) be the amount of output allocated to consumption, and let I(t) = F [K(t)] − C(t) be the amount invested. Let δ be the constant rate of depreciation of capital. Then, the capital stock equation is K˙ = F (K) − C − δK, K(0) = K0 .

(11.1)

Let U (C) be the society’s utility of consumption, where we assume U (0) = ∞, U (C) > 0, and U (C) < 0, for C ≥ 0. These conditions ensure that U (C) is strictly concave in C. Let ρ denote the social discount rate and T denote the ﬁnite horizon. Then, a government which is elected for a term of T years could consider the following problem: T max J = e−ρt U [C(t)]dt (11.2) 0

subject to (11.1) and the ﬁxed-end-point condition K(T ) = KT ,

(11.3)

where KT is a given positive constant. It may be noted that replacing (11.3) by K(T ) ≥ KT would give the same solution.

11.1.2

Solution by the Maximum Principle

Form the current-value Hamiltonian as H = U (C) + λ[F (K) − C − δK].

(11.4)

11.1. Models of Optimal Economic Growth

337

The adjoint equation is ∂F ∂H = (ρ + δ)λ − λ , λ(T ) = α, λ˙ = ρλ − ∂K ∂K

(11.5)

where α is a constant to be determined. The optimal control is given by ∂H = U (C) − λ = 0. ∂C

(11.6)

Since U (0) = ∞, the solution of this condition always gives C(t) > 0. An intuitive argument for this result is that a slight increase from a zero consumption rate brings and inﬁnitesimally large marginal utility and therefore optimal consumption will remain strictly positive. Moreover, the capital stock will not be allowed to fall to zero along an optimal path in order to avoid the consumption rate from falling to zero. See Karatzas et al. (1986) for a rigorous demonstration of this result in a related context. Note that the suﬃciency of optimality is easily established here by obtaining the derived Hamiltonian H 0 (K, λ) by substituting for C from (11.6) in (11.4), and showing that H 0 (K, λ) is concave in K. This follows easily from the facts that F (K) is concave and λ > 0 from (11.6) on account of the assumption that U (C) > 0. The economic interpretation of the Hamiltonian is straightforward. It consists of two terms: the ﬁrst one gives the utility of current consumption and the second one gives the net investment evaluated by price λ, which, from (11.6), reﬂects the marginal utility of consumption. For the economic system to be run optimally, the solution must satisfy the following three conditions: (a) The static eﬃciency condition (11.6) which maximizes the value of the Hamiltonian at each instant of time myopically, provided that λ(t) is known. (b) The dynamic eﬃciency condition (11.5) which forces the price λ of capital to change over time in such a way that the capital stock always yields a net rate of return, which is equal to the social discount rate ρ. That is, dλ +

∂H dt = ρλdt. ∂K

338

11. Applications to Economics

(c) The long-run foresight condition, which establishes the terminal price λ(T ) of capital in such a way that exactly the terminal capital stock KT is obtained at T. Equations (11.1), (11.3), (11.5), and (11.6) form a two-point boundary value problem which can be solved numerically. In Exercise 11.1, you are asked to solve a simple version of the model in which the TPBVP can be solved analytically.

11.1.3

Introduction of a Growing Labor Force

In the preceding sections of this chapter we studied the simplest capital accumulation model in which the population was assumed to be ﬁxed. We now introduce labor as a new factor (treated the same as population, for simplicity), which grows exponentially at a ﬁxed rate g, 0 < g < ρ. It is now possible to recast the new model in terms of per capita variables so that it is formally similar to the previous model. The introduction of the per capita variables makes it possible to treat the inﬁnite horizon version of the new model. Let L(t) denote the amount of labor at time t. Since it is growing exponentially at rate g, we have L(t) = L(0)egt .

(11.7)

Let F (K, L) be the production function which is assumed to be strictly increasing and concave in both factors of production so that FK > 0, FL > 0, FKK < 0, and FLL < 0 for K ≥ 0, L ≥ 0. Furthermore, it is homogeneous of degree one so that F (mK, mL) = mF (K, L) for m ≥ 0. We deﬁne k = K/L and the per capita production function f (k) as f (k) =

K F (K, L) = F ( , 1) = F (k, 1). L L

(11.8)

It is clear from the assumptions of F that f (k) > 0 and f (k) < 0 for k ≥ 0. To derive the state equation for k, we note that ˙ + k L˙ = kL ˙ + kgL. K˙ = kL Substituting for K˙ from (11.1) and deﬁning per capita consumption c = C/L, we get (11.9) k˙ = f (k) − c − γk, k(0) = k0 ,

11.1. Models of Optimal Economic Growth

339

where γ = g + δ. Let u(c) be the utility of per capita consumption c, where u is assumed to satisfy u (c) > 0 and u (c) < 0 for c ≥ 0 and u (0) = ∞.

(11.10)

As in Sect. 11.1.2, the last condition in (11.10) rules out zero consumption. According to the position known ∞ as total utilitarianism, the society’s discounted total utility is 0 e−ρt L(t)u(c(t))dt, which we aim to maximize. In view of (11.7), this is equivalent to maximizing ∞ J= e−rt u(c)dt, (11.11) 0

where r = ρ − g > 0. Note also that r + γ = ρ + δ. Remark 11.1 It is interesting to note that the problem is an inﬁnite version of that in Sect. 11.1.1, if we consider r to be the adjusted discount rate and γ to be the adjusted depreciation rate. This reduction of a model with two factors of production to a one-sector model does not work if we jettison the assumption of an exponentially growing population. Then, the analysis becomes much more complicated. The reader is referred to Arrow et al. (2007, 2010) for economic growth models with non-exponentially and endogenously growing populations.

11.1.4

Solution by the Maximum Principle

The current-value Hamiltonian is H = u(c) + λ[f (k) − c − γk].

(11.12)

The adjoint equation is ∂H λ˙ = rλ − = (r + γ)λ − f (k)λ = (ρ + δ)λ − f (k)λ. ∂k

(11.13)

To obtain the optimal control, we diﬀerentiate (11.12) with respect to c, set it to zero, and solve (11.14) u (c) = λ. Let c = h(λ) = u−1 (λ) denote the solution of (11.14). In Exercise 11.3, you are asked to show that h (λ) < 0. This can be easily shown by

340

11. Applications to Economics

inverting the graph of u (c) vs. c. Alternatively you can rewrite (11.14) as u (h(λ)) = λ and then take its derivative with respect to λ. To show that the maximum principle is suﬃcient for optimality, it is enough to show that the derived Hamiltonian H 0 (k, λ) = u(h(λ)) + λ[f (k) − h(λ) − γk]

(11.15)

is concave in k for any λ satisfying (11.14). The concavity follows immediately from the facts that λ is positive from (11.10) and (11.14) and f (k) is concave because of the assumptions on F (K, L). Equations (11.9), (11.13), and (11.14) now constitute a complete autonomous system, since time does not enter explicitly in these equations. Such systems can be analyzed by the phase diagram method, which is used next. In Fig. 11.1 we have drawn a phase diagram for the two equations k˙ = f (k) − h(λ) − γk = 0, λ˙ = (r + γ)λ − f (k)λ = 0,

(11.16) (11.17)

obtained from (11.9), (11.13), and (11.14). In Exercise 11.2 you are asked to show that the graphs of k˙ = 0 and λ˙ = 0 are like the dotted curves in Fig. 11.1. Given the nature of these graphs, known as isoclines, it is ¯ λ). ¯ In clear that they have a unique point of intersection denoted as (k,

Figure 11.1: Phase diagram for the optimal growth model

11.1. Models of Optimal Economic Growth

341

¯ λ) ¯ is the unique solution of the equations other words, (k, ¯ − h(λ) ¯ − γ k¯ = 0 and(r + γ) − f (k) ¯ = 0. f (k)

(11.18)

The two isoclines divide the plane into four regions, I, II, III, and IV, as marked in Fig. 11.1. To the left of the vertical line λ˙ = 0, we have k < k¯ and therefore r + γ < f (k) in view of f (k) < 0. Thus, λ˙ < 0 from (11.13). Therefore, λ is decreasing, which is indicated by the downward pointing arrows in Regions I and IV. On the other hand, to the right of the vertical line, in Regions II and III, the arrows are pointed upward because λ is increasing. In Exercise 11.3, you are asked to show that the horizontal arrows, which indicate the direction of change in k, point to the right above the k˙ = 0 isocline, i.e., in Regions I and II, and they point to the left in Regions III and IV which are below the k˙ = 0 isocline. ¯ λ) ¯ represents the optimal long-run stationary equilibThe point (k, ¯ are obtained in Exercise 11.2. The next rium. The values of k¯ and λ important thing is to show that there is a unique path starting from any initial capital stock k0 , which satisﬁes the maximum principle and ¯ λ). ¯ Clearly such a path cannot start in converges to the steady state (k, Regions II and IV, because the directions of the arrows in these areas ¯ λ). ¯ For k0 < k, ¯ the value of λ0 (if any) must be point away from (k, ¯ on the other hand, selected so that (k0 , λ0 ) is in Region I. For k0 > k, the point (k0 , λ0 ) must be chosen to be in Region III. We analyze the case k0 < k¯ only, and show that there exists a unique λ0 associated with the given k0 , and that the optimal path, shown as the solid curve in Re¯ λ). ¯ It should gion I of Fig. 11.1, starts from (k0 , λ0 ) and converges to (k, be obvious that this path also represents the locus of such (k0 , λ0 ) for ¯ The analysis of the case k0 > k¯ is left as Exercise 11.4. k0 ∈ [0, k]. ˙ In Region I, k(t) > 0 and k(t) is an increasing function of t as indicated by the horizontal right-directed arrow in Fig. 11.1. Therefore, we can replace the independent variable t by k, and then use (11.16) and (11.17) to obtain 3 dλ dk [f (k) − (r + γ)]λ dλ = = . (11.19) λ (k) = dk dt dt h(λ) + γk − f (k) Thus, our task of showing that there exists an optimal path starting from any initial k0 < k¯ is equivalent to showing that there exists a solution ¯ beginning with of the diﬀerential equation (11.19) on the interval [0, k], ¯ ¯ the boundary condition λ(k) = λ. For this, we must obtain the value ¯ Since both the numerator and the denominator in (11.19) vanish λ (k).

342

11. Applications to Economics

¯ by a perturbation argument. To do so, ¯ we need to derive λ (k) at k = k, we use (11.19) and (11.18) to obtain λ (k) =

¯ − f (k)]λ [r + γ − f (k)]λ [f (k) = ¯ − γk + γ k¯ − h(λ) + h(λ) ¯ . f (k) − γk − h(λ) f (k) − f (k)

We use L’Hˆopital’s rule to take the limit as k → k¯ and obtain ¯ = λ (k) or −

¯ λ ¯ ¯ λ ¯ −f (k) −f (k) = ¯ − γ − h (λ) ¯ ¯ − γ − λ (k)/u ¯ (h(λ)) ¯ , f (k) f (k)

(11.20)

¯ 2 (λ (k)) ¯ ¯ ¯ ¯ ¯ + λ (k)[f (k) − γ] + λf (k) = 0. u (h(λ))

(11.21)

¯ = Note that the second equality in (11.20) uses the relation h (λ) ¯ 1/u (h(λ)) obtained by diﬀerentiating u (c) = u (h(λ)) = λ of (11.14) ¯ with respect to λ at λ = λ. It is easy to see that (11.21) has one positive solution and one negative ¯ because of the following solution. We take the negative solution for λ (k) consideration. With the negative solution, we can prove that the diﬀerential equation (11.19) has a smooth solution, such that λ (k) < 0. For this, let π(k) = f (k) − kγ − h(λ(k)). ¯ < 0, ¯ we have r+γ−f (k) < 0. Then from (11.19), since λ (k) Since k < k, ¯ ¯ ¯ ¯ we have λ(k − ε) > λ(k). Also since λ > 0 and f (k) < 0, Eq. (11.20) ¯ implies with λ (k) ¯ = f (k) ¯ −γ− π (k)

¯ λ (k) ¯ < 0, u (h(λ))

and thus, π(k¯ − ε) = f (k¯ − ε) − γ(k¯ − ε) − h(λ(k¯ − ε)) > 0. Therefore, the derivative at k¯ − ε is well deﬁned and λ (k¯ − ε) < 0. We can proceed as long as π (k) = f (k) − γ −

λ (k) < 0. u (h(λ(k)))

(11.22)

This implies that f (k) − kγ − h(λ) > 0, and also since r + γ − f (k) ¯ we have λ (k) < 0. remains negative for k < k,

11.2. A Model of Optimal Epidemic Control

343

˜ = 0. Then, since Suppose now that there is a point k˜ < k¯ with π(k) ˜ ≥ 0. But at k, ˜ π(k) ˜ = 0 in (11.19) implies π(k˜ + ε) > 0, we have π (k) ˜ ˜ = −∞, which is a λ (k) = −∞, and then from (11.22), we have π (k) ˜ ≥ 0. Thus, we can proceed on the whole interval contradiction with π (k) ¯ [0, k]. This indicates that the path λ(k) (shown as the solid line in Region I of Fig. 11.1) remains above the curve k˙ = f (k) − kγ − h(λ) = 0, ¯ Thus, we can set shown as the dotted line in Fig. 11.1 when k < k. ¯ λ0 = λ(k0 ) for 0 ≤ k0 ≤ k and have the optimal path starting from ¯ λ). ¯ (k0 , λ0 ) and converging to (k, ¯ in order Similar arguments hold when the initial capital stock k0 > k, to show that the optimal path (shown as the solid line in Region III of Fig. 11.1) exists in this case. You have already been asked to carry out this analysis in Exercise 11.4. We should mention that the conclusions derived in this subsection could have been reached by invoking the Global Saddle Point Theorem stated in Appendix D.7, but we have chosen instead to carry out a detailed analysis for illustrating the use of the phase diagram method. The next time we use the phase diagram method will be in Sect. 11.3.3, and there we shall rely on the Global Saddle Point Theorem.

11.2

A Model of Optimal Epidemic Control

Certain infectious epidemic diseases are seasonal in nature. Examples are the common cold, the ﬂu, and certain children’s diseases. When it is beneﬁcial to do so, control measures are taken to alleviate the eﬀects of these diseases. Here we discuss a simple control model due to Sethi (1974c) for analyzing an epidemic problem. Related problems have been treated by Sethi and Staats (1978), Sethi (1978c), and Francis (1997). See Wickwire (1977) for a good survey of optimal control theory applied to the control of pest infestations and epidemics, and Swan (1984) for applications to biomedicine.

11.2.1

Formulation of the Model

Let N be the total ﬁxed population. Let x(t) be the number of infectives at time t so that the remaining N − x(t) is the number of susceptibles. To keep the model simple, assume that no immunity is acquired so that

344

11. Applications to Economics

when infected people are cured, they become susceptible again. The state equation governing the dynamics of the epidemic spread in the population is x˙ = βx(N − x) − vx, x(0) = x0 , (11.23) where β is a positive constant termed infectivity of the disease, and v is a control variable reﬂecting the level of medical program eﬀort. Note that x(t) is in [0, N ] for all t > 0 if x0 is in that interval. The objective of the control problem is to minimize the present value of the cost stream up to a horizon time T, which marks the end of the season for that disease. Let h denote the unit social cost per infective, let m denote the cost of control per unit level of program eﬀort, and let Q denote the capability of the health care delivery system providing an upper bound on v. The optimal control problem is: T −ρt −(hx + mv)e dt (11.24) max J = 0

subject to (11.23), the terminal constraint that x(T ) = xT ,

(11.25)

and the control constraint 0 ≤ v ≤ Q.

11.2.2

Solution by Green’s Theorem

Rewriting (11.23) as vdt = [βx(N − x)dt − dx]/x and substituting into (11.24) yields the line integral ! " m JΓ = − [hx + mβ(N − x)]e−ρt dt − e−ρt dx , x Γ

(11.26)

where Γ is a path from x0 to xT in the (t, x)-space. Let Γ1 and Γ2 be two such paths from x0 to xT , and let R be the region enclosed by Γ1 and Γ2 . By Green’s theorem, we can write & % mρ − h + mβ e−ρt dtdx. JΓ1 −Γ2 = JΓ1 − JΓ2 = − (11.27) x R

11.2. A Model of Optimal Epidemic Control

345

To obtain the singular control we set the integrand of (11.27) equal to zero, as we did in Sect. 7.2.2. This yields x=

ρ ρ = , h/m − β θ

where θ = h/m − β. Deﬁne the singular state xs as follows: ⎧ ⎪ ⎨ ρ/θ if 0 < ρ/θ < N, s x = ⎪ ⎩ N otherwise. The corresponding singular control level ⎧ ⎪ ⎨ β(N − ρ/θ) if 0 < ρ/θ < N, s s v = β(N − x ) = ⎪ ⎩ 0 otherwise.

(11.28)

(11.29)

(11.30)

We will show that xs is the turnpike level of infectives. It is instructive to interpret (11.29) and (11.30) for the various cases. If ρ/θ > 0, then θ > 0 so that h/m > β. Here the smaller the ratio h/m, the larger the turnpike level xs , and therefore, the smaller the medical program eﬀort should be. In other words, the smaller the social cost per infective and/or the larger the treatment cost per infective, the smaller the medical program eﬀort should be. When ρ/θ < 0, you are asked to show in Exercise 11.9 that xs = N in the case h/m < β, which means the ratio of the social cost to the treatment cost is smaller than the infectivity coeﬃcient. Therefore, in this case when there is no terminal constraint, the optimal trajectory involves no treatment eﬀort. An example of this case is the common cold where the social cost is low and treatment cost is high. The optimal control for the fortuitous case when xT = xs is ⎧ ⎪ ⎪ ⎪ Q if x(t) > xs , ⎪ ⎪ ⎨ (11.31) v ∗ (x(t)) = v s if x(t) = xs , ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 0 if x(t) < xs . When xT = xs , there are two cases to consider. For simplicity of exposition we assume x0 > xs and T and Q to be large.

346

11. Applications to Economics

Case 1: xT > xs . The optimal trajectory is shown in Fig. 11.2. In Exercise 11.8 you are asked to show its optimality by using Green’s theorem. Case 2: xT < xs . The optimal trajectory is shown in Fig. 11.3. It can be shown that x goes asymptotically to N − Q/β if v = Q. The level is marked in Fig. 11.3. The optimal control shown in Figs. 11.2 and 11.3 assumes 0 < xs < N. It also assumes that T is large so that the trajectory will spend some time on the turnpike and Q is large so that xs ≥ N − Q/β. The graphs are drawn for x0 > xs and xs < N/2; for all other cases see Sethi (1974c).

Figure 11.2: Optimal trajectory when xT > xs

11.3

A Pollution Control Model

In this section we will describe a simple pollution control model due to Keeler et al. (1971). We will describe this model in terms of an economic system in which labor is the only primary factor of production, which is allocated between food production and DDT production. It is assumed that all of the food produced is used for consumption. On the other hand, all of the DDT produced is used as a secondary factor of production which, along with labor, determines the food output. However, when used, DDT causes pollution, which can only be reduced by natural decay. The objective of the society is to maximize the total present value of the utility of food less the disutility of pollution due to the use of DDT.

11.3. A Pollution Control Model

347

Figure 11.3: Optimal trajectory when xT < xs

11.3.1

Model Formulation

We introduce the following notation: L = the total labor force, assumed to be constant for simplicity, v = the amount of labor used for DDT production, L − v = the amount of labor used for food production, P

= the stock of DDT pollution at time t,

a(v) = the rate of DDT output; a(0) = 0, a > 0, a < 0, for v ≥ 0, δ = the natural exponential decay rate of DDT pollution, C(v) = f [L − v, a(v)] = the rate of food output to be consumed; C(v) is concave, C(0) > 0, C(L) = 0; C(v) attains a unique maximum at v = V > 0; see Fig. 11.4. Note that a suﬃcient condition for C(v) to be strictly concave is f12 ≥ 0 along with the usual concavity and monotonicity conditions on f (see Exercise 11.10), u(C) = the utility function of consuming the food output C ≥ 0; u (0) = ∞, u (C) > 0, u (C) < 0, h(P ) = the disutility function of pollution stock P ≥ 0; h (0) = 0, h (P ) > 0, h (P ) > 0.

348

11. Applications to Economics

Figure 11.4: Food output function The optimal control problem is: ∞ −ρt e [u(C(v)) − h(P )]dt max J =

(11.32)

0

subject to P˙ = a(v) − δP, P (0) = P0 ,

(11.33)

0 ≤ v ≤ L.

(11.34)

From Fig. 11.4, it is obvious that v is at most V, since the production of DDT beyond that level decreases food production and increases DDT pollution. Hence, (11.34) can be reduced to simply v ≥ 0.

11.3.2

(11.35)

Solution by the Maximum Principle

Form the current-value Lagrangian L(P, v, λ, μ) = u[C(v)] − h(P ) + λ[a(v) − δP ] + μv

(11.36)

using (11.32), (11.33) and (11.35), where λ˙ = (ρ + δ)λ + h (P ),

(11.37)

μ ≥ 0 and μv = 0.

(11.38)

and

11.3. A Pollution Control Model

349

The optimal solution is given by ∂L = u [C(v)]C (v) + λa (v) + μ = 0. ∂v

(11.39)

Since the derived Hamiltonian is concave, conditions (11.36)–(11.39) together with ¯ = constant (11.40) lim λ(t) = λ t→∞

are suﬃcient for optimality; see Theorem 2.1 and Sect. 2.4. The phase diagram analysis presented below gives λ(t) satisfying (11.40).

11.3.3

Phase Diagram Analysis

From the assumptions on C(v) or from Fig. 11.4, we see that C (0) > 0. This means that du/dv = u (C(v))C (v)|v=0 > 0. This along with h (0) = 0 implies that v > 0, meaning that it pays to produce some positive amount of DDT in equilibrium. Therefore, the equilibrium value of the Lagrange multiplier is zero, i.e., μ ¯ = 0. From (11.33), (11.37) and ¯ and v¯ as follows: (11.39), we get the equilibrium values P¯ , λ, a(¯ v) P¯ = , δ

(11.41)

¯ v )]C (¯ v) ¯ = − h (P ) = − u [C(¯ λ . ρ+δ a (¯ v)

(11.42)

From (11.42) and the assumptions on the derivatives of g, C and a, we ¯ < 0. From this and (11.37), we conclude that λ(t) is always know that λ negative. The economic interpretation of λ is that −λ is the imputed cost of pollution. Let v = Φ(λ) denote the solution of (11.39) with μ = 0. On account of (11.35), deﬁne v ∗ = max[0, Φ(λ)].

(11.43)

We know from the interpretation of λ that when λ increases, the imputed cost of pollution decreases, which can justify an increase in the DDT production to ensure an increased food output. Thus, it is reasonable to assume that dΦ > 0, dλ and we will make this assumption. It follows that there exists a unique λc such that Φ(λc ) = 0, Φ(λ) < 0 for λ < λc and Φ(λ) > 0 for λ > λc .

350

11. Applications to Economics

To construct the phase diagram, we must plot the isoclines P˙ = 0 and λ˙ = 0. These are, respectively, P =

a[max{0, Φ(λ)}] a(v ∗ ) = , δ δ

(11.44)

h (P ) = −(ρ + δ)λ.

(11.45)

Observe that the assumption h (0) = 0 implies that the graph of (11.45) passes through the origin. Diﬀerentiating these equations with respect to λ and using (11.43), we obtain dP a (v) dv |P˙ =0 = >0 dλ δ dλ

(11.46)

as the slope of the P˙ = 0 isocline, and dP (ρ + δ) |λ=0 0) above (below) the P˙ = 0 isocline and λ˙ > 0 (λ˙ < 0) to the right (left) of the λ˙ = 0 isocline. ¯ P¯ ) of these isoclines denotes the equilibThe intersection point (λ, rium levels for the adjoint variable and the pollution stock, respectively. That there exists an optimal path (shown as the solid line in Fig. 11.5) ¯ P¯ ) follows directly from the Global converging to the equilibrium (λ, Saddle Point Theorem stated in Appendix D.7. Given λc as the intersection of the P˙ = 0 curve and the horizontal axis, the corresponding ordinate P c on the optimal trajectory is the related pollution stock level. The signiﬁcance of P c is that if the existing pollution stock P is larger than P c , then the optimal control is v ∗ = 0, meaning no DDT is produced. Given an initial level of pollution P0 , the optimal trajectory curve in Fig. 11.5 provides the initial value λ0 of the adjoint variable. With these initial values, the optimal trajectory is determined by (11.33), (11.37), and (11.43). If P0 > P c , as shown in Fig. 11.5, then v ∗ = 0 until such

11.3. A Pollution Control Model

351

Figure 11.5: Phase diagram for the pollution control model

time that the natural decay of pollution stock has reduced it to P c . At that time, the adjoint variable has increased to the value λc . The optimal ¯ P¯ ). control is v ∗ = φ(λ) from this time on, and the path converges to (λ, ¯ > 0, which implies that it is optimal to At equilibrium, v¯ = Φ(λ) produce some DDT forever in the long run. The only time when its production is not optimal is at the beginning when the pollution stock is higher than P c . It is important to examine the eﬀects of changes in the parameters on the optimal path. In particular, you are asked in Exercise 11.11 to show that an increase in the natural rate of decay of pollution, δ, will increase P c . That is, when pollution decays at a faster rate, we can increase the threshold level of pollution stock at which to ban the production of the pollutant. For DDT in reality, δ is small so that its complete ban, which has actually occurred, may not be far from the optimal policy. Here we have presented a very simple model of pollution in which the problem was to choose an optimal production process. Models in which the control variable to determine is the optimal amount to spend in reducing the pollution output of an existing dirty process have also been formulated; see Wright (1974) and Sethi (1977d). For other related models, see Luptacik and Schubert (1982), Hartl and Luptacik (1992), and Hartl and Kort (1996a,b,c, 1997), Xepapadeas and de Zeeuw (1999), and Moser et al. (2014).

352

11.4

11. Applications to Economics

An Adverse Selection Model

In modern contract theory, the term adverse selection is used to describe principal-agent models in which an agent has private information before a contract is written. For example, a seller does not know perfectly how much a buyer is willing to pay for a good. A related concept is that of moral hazard, when there is present a hidden action not adversely observed by the principal. In such game situations, clearly the principal would like to know the agent’s private information which he cannot learn simply by asking the agent, because it is in the agent’s interest to distort the truth. Fortunately, according to the theory of mechanism design, the principal can design a game whose rules can inﬂuence the agent to act the way he would like. Thanks, particularly to the revelation principle, the principal needs only consider games in which the agent truthfully reports her private information. There is a large literature on contract theory, and we refer the reader to books by Laﬀont and Mortimort (2001), Bolton and Dewatripont (2005) and Cvitanic and Zhang (2013). For our purposes, we shall next consider a game between a seller and a buyer, where the buyer has private information about her willingness-to-pay for the seller’s goods; see Bolton and Dewatripont (2005).

11.4.1

Model Formulation

Consider a transaction between a seller (the principal) and a buyer (the agent) of type t ∈ [t1 , t2 ], 0 ≤ t1 ≤ t2 , represents her willingness-to-pay for seller’s goods. We assume in particular that buyer’s preferences are represented by the utility function U (q, φ, t) = ta(q) − φ,

(11.48)

where q is the number of units purchased and φ is the total amount paid to the seller. We assume a(0) = 0, a > 0, and a < 0. The seller knows only the distribution F (t), having the density f (t), t ∈ [t1 , t2 ]. The seller’s unit production cost is c > 0, so that his proﬁt from selling q units against a sum of money φ is given by π = φ − cq.

(11.49)

The question of interest here is to obtain a proﬁt-maximizing pair {φ, q} that the seller will be able to induce the buyer of type tˆ to choose.

11.4. An Adverse Selection Model

353

Thanks to the revelation principle, the answer is that the seller can oﬀer a menu of contracts {φ(t), q(t)} which comes from solving the following maximization problem: t2 [φ(t) − cq(t)]f (t)dt (11.50) max q(·),φ(·) t1

subject to (IR) tˆa(q(tˆ)) − φ(tˆ) ≥ 0, tˆ ∈ [t1 , t2 ]

(11.51)

(IC) tˆa(q(tˆ)) − φ(tˆ) ≥ tˆa(q(t)) − φ(t), t, tˆ ∈ [t1 , t2 ], t = tˆ. (11.52) The constraints (11.51), called individual rationality constraints (IR), say that the agent of type tˆ will participate in the contract. Clearly, given (11.52), we can replace these constraints by a single constraint t1 a(q(t1 )) − φ(t1 ) ≥ 0.

(11.53)

The left-hand side of the constraints (11.52), called incentive compatibility constraints (IC), is the utility of agent tˆ if she chooses the contract intended for her, whereas the right-hand side represents the utility of agent tˆ if she chooses the constraint intended for type t = tˆ. The IC constraints, therefore, imply that type tˆ agent is better oﬀ choosing the contract intended for her than any other contract in the menu. Clearly, the seller’s problem is mathematically diﬃcult as it involves maximizing the seller’s proﬁt over a class of functions. So, a way to deal with this problem is to decompose it into an implementation problem (which functions q(·) are incentive compatible?) and an optimization problem (which one is the best implementation function for the seller?)

11.4.2

The Implementation Problem

Given a menu {q(·), φ(·)} that satisﬁes the seller’s problem (11.50)– (11.52), it must be the case in equilibrium that the buyer tˆ will choose the contract {q(tˆ), φ(tˆ)}. In other words, his utility tˆa(q(t)) − φ(t) of choosing a contract {q(t), φ(t)} will be maximized at t = tˆ. Assuming that q(·) and φ(·) are twice diﬀerentiable functions, the ﬁrst-order and second-order conditions are ˙ ˙ tˆ) = 0, ˆ ˆ ˙ tˆ) − φ( tˆa (q(t))q(t) ˙ − φ(t)| t=tˆ = ta (q(t))q(

(11.54)

354

11. Applications to Economics 2 ¨ tˆa (q(t))(q(t)) ˙ + tˆa (q(t))¨ q (t) − φ(t)| t=tˆ ≤ 0.

(11.55)

From (11.54), it follows from replacing tˆ by t that ˙ ˙ − φ(t) = 0, t ∈ [t1 , t2 ], ta (q(t))q(t)

(11.56)

called the local incentive compatibility condition, must hold. Diﬀerentiating (11.56) gives, 2 ¨ = 0. ˙ + a (q(t))q(t) ˙ + ta (q(t))¨ q (t) − φ(t) ta (q(t))(q(t))

(11.57)

It follows from (11.55), (11.57), and a > 0, that q(t) ˙ ≥ 0.

(11.58)

This is called the monotonicity condition. In Exercise 11.12, you are asked to show that (11.56) and (11.58) are suﬃcient for (11.52) to hold. Since, these conditions are already necessary, we can say that local incentive compatibility (11.56) and monotonicity (11.58) together are equivalent to the IC condition (11.52). We can now ready to formulate the seller’s optimization problem.

11.4.3

The Optimization Problem

The seller’s problem can be written as the following optimal control problem: t2 [φ(t) − cq(t)]f (t)dt (11.59) max u(·)

t1

subject to q(t) ˙ = u(t),

(11.60)

˙ φ(t) = ta (q(t))u(t),

(11.61)

t1 a(q(t1 )) − φ(t1 ) = 0,

(11.62)

u(t) ≥ 0.

(11.63)

Here, q(t) and φ(t) are state variables and u(t) is a control variable satisfying the control constraint u(t) ≥ 0. The objective function (11.59) is the expected value of the seller’s proﬁt with respect to the density f (t). Equation (11.60) and constraint (11.63) come from the monotonicity condition (11.58). Equation (11.61) with u(t) from (11.60) gives the local incentive compatibility condition (11.56). Finally, (11.62) speciﬁes

11.4. An Adverse Selection Model

355

the IR constraint (11.53) in view of the fact it will be binding for the lowest agent type t1 at the optimum. We can now use the sense of the maximum principle (3.12) to write the necessary conditions for optimality. Note that (3.12) is written for problem (3.7) that has speciﬁed initial states and some constraints on the terminal state vector x(T ) that include the equality constraint b(x(T ), T ) = 0. Our problem, on the other hand, has this type of equality constraint, namely (11.62), on the initial states q(t1 ) and φ(t1 ) and no speciﬁed terminal states q(t2 ) and φ(t2 ). However, since initial time conditions and terminal time conditions can be treated in a symmetric fashion, we can apply the sense of (3.12), as shown in Remark 3.9, to obtain the necessary optimality conditions to problem (11.59)–(11.63). In Exercise 11.13, you are asked to obtain (11.67) and (11.68) by following Remark 3.9 to account for the presence of the equality constraint (11.62) on the initial state variables rather than on the terminal state as in problem (3.7). To specify the necessary optimality condition, we ﬁrst deﬁne the Hamiltonian. H(q, φ, λ, μ, t) = [φ(t) − cq(t)]f (t) + λ(t)u(t) + μ(t)[ta (q(t)u(t))] = [φ(t) − cq(t)]f (t) + [λ(t) + μ(t)ta (q(t))]u(t) (11.64) Then for u∗ with the corresponding state trajectories q ∗ and φ∗ to be optimal, we must have adjoints λ and μ, and a constant β, such that ∗ q˙∗ = u∗ , φ˙ = ta (q ∗ )u

(11.65)

t1 a(q ∗ (t1 )) − φ∗ (t1 ) = 0,

(11.66)

λ˙ = cf − μta (q ∗ )u∗ , λ(t1 ) = βt1 a (q ∗ (t1 )), λ(t2 ) = 0,

(11.67)

μ˙ = −f, μ(t1 ) = −β, μ(t2 ) = 0,

(11.68)

u∗ (t) = bang[0, ∞; λ(t) + μ(t)ta (q ∗ (t))].

(11.69)

Several remarks are in order at this point. First we see that we have a bang-bang control in (11.69). This means that the u∗ (t) can be 0, or greater than 0, or an impulse control. Moreover, in the region when u∗ (t) = 0, which will occur when λ(t) + μ(t)ta (q ∗ (t)) < 0, we will have a constant q ∗ (t), and we will have a singular control u∗ (t) > 0 if we can keep λ(t) + μ(t)ta (q ∗ (t)) = 0 by an appropriate choice of u∗ (t) along

356

11. Applications to Economics

the singular path. An impulse control would occur if the initial q(t1 ) were above the singular path. Since in our problem, initial states are not exactly speciﬁed, we shall not encounter an impulse control here. The third remark concerns a numerical way of solving the problem. For this, let us rewrite the boundary conditions in (11.67) and (11.68) and the condition (11.66) as below: t1 a(q ∗ (t1 )) − φ∗ (t1 ) = 0, λ(t1 ) = −μ(t1 )t1 a (q ∗ (t1 )) λ(t2 ) = μ(t2 ) = 0.

(11.70) (11.71)

With (11.71) and a guess of q(t2 ) and φ(t2 ), we can solve the diﬀerential equation (11.65), (11.67) and (11.68), with u∗ (t) in (11.69), backward in time. These will give us the values of λ(t1 ), μ(t1 ), q(t1 ) and φ(t1 ). We can check if these satisfy the two equations in (11.70). If yes, we have arrived at a solution. If not, we change our guess for q(t2 ) and φ(t2 ) and start again. As you may have noticed, the procedure is very similar to solving a two-point boundary value problem. Next we provide an alternative procedure to solve the seller’s problem, a procedure used in the theory of mechanism design. This procedure ﬁrst ignores the nonnegativity constraint (11.60) and solves the relaxed problem given by (11.59)–(11.62). In view of (11.52), let us deﬁne u0 (tˆ) = tˆa(q(tˆ)) − φ(tˆ) = max[ta(q(t)) − φ(t)]. t

(11.72)

By the envelope theorem, we have du0 (tˆ) ∂u0 (tˆ) = = a(q(tˆ)), dtˆ ∂ tˆ which we can integrate to obtain t t 0 0 a(q(x)dx) + u (t1 ) = a(q(x))dx, u (t) = t1

(11.73)

(11.74)

t1

since u∗ (t1 ) = 0 at the optimum. Also, since φ(t) = ta(q(t)) − u0 (t), we can write the seller’s proﬁt as t2 t [ta(q(t)) − a(q(x))dx − cq(t)]f (t)dt. (11.75) t1

t1

Then, integrating by parts, we have t2 t1 [{ta(q(t)) − cq(t)}f (t) − a(q(t))(1 − F (t))] dt t = t12 [ta(q(t)) − cq(t) − a(q(t))/h(t)] f (t)dt,

(11.76)

11.4. An Adverse Selection Model

357

where h(t) = f (t)/[1 − F (t)] is known as the hazard rate. Since we are interested in maximizing the seller’s proﬁt with respect to the output schedule q(·), we can maximize the expression under the integral pointwise for each t. The ﬁrst-order condition for that is

1 1 − F (t) (11.77) a (q(t)) = t − a (q(t)) = c, t− f (t) h(t) which gives us the optimal solution of the relaxed problem as 1 −1 −1 c t− . qˆ(t) = a h(t)

(11.78)

In obtaining (11.78), we had omitted the nonnegativity constraint (11.63) introduced to ensure that q(t) is increasing. Thus, it remains to check if dˆ q (t)/dt ≥ 0. It is straightforward to verify that if the hazard rate h(t) is increasing in t, then qˆ(t) is increasing in t. To show this, we diﬀerentiate (11.78) to obtain dˆ q (t) q (t)) g(t)a (ˆ = − , dt a (ˆ q (t))g(t) where g(t) = [t − 1/h(t)]. Clearly, if h(t) is increasing, then g(t) is increasing, and dˆ q (t)/dt ≥ 0. ˆ In this case, qˆ(t) and the corresponding φ(t) obtained from solving the diﬀerential equation given by (11.61) and the boundary condition ˆ (11.62) give us the optimal menu {φ(t), qˆ(t)}. What if h(t) is not increasing? In that case, there is a procedure called bunching and ironing given by the solution of an optimal control problem to be formulated next. This is because qˆ(t) in (11.78) is obtained by solving the relaxed problem that ignores the nonnegativity constraint (11.63), and so it may be that dˆ q /dt is strictly negative for some t ∈ [t, t¯] ⊂ [t1 , t2 ] as shown in Fig. 11.6. Then the seller must choose the optimal q ∗ (t) to maximize the following constrained optimal control problem:

t2 a(q(t)) ta(q(t)) − cq(t) − f (t)dt (11.79) max h(t) q(·) t1 subject to q(t) ˙ = u(t), u(t) ≥ 0.

(11.80)

358

11. Applications to Economics qˆ(t)

q * (θ)

t¯

t

t1

t2

t

Figure 11.6: Violation of the monotonicity constraint Now the necessary optimality conditions, with the Hamiltonian deﬁned as H(q, 0, λ, t) = (ta(q) − cq − a(q)/h)f + λu, (11.81) are

$ 4 λ˙ = − (t − 1/h)a (q) − c f, λ(t1 ) = λ(t2 ) = 0,

(11.82)

and u∗ = [0, ∞; λ].

(11.83)

We may also note that these conditions are also suﬃcient since H in (11.81) is concave in q. Integrating (11.82), we have λ(t) = −

t t1

1 z− h(z)

a (q(z)) − c f (z)dz.

Using the transversality conditions in the case when neither the initial nor the terminal state is speciﬁed for the state equation (11.80), we obtain

t2 1 a (q(z)) − c f (z)dz. z− 0 = λ(t1 ) = λ(t2 ) = − h(z) t1

11.4. An Adverse Selection Model

359

Then for u∗ (t) = 0 on an interval t ∈ [θ1 , θ2 ] ⊂ [t1 , t2 ], we must have λ(t) < 0, t ∈ [θ1 , θ2 ]. Moreover, when u∗ (t) > 0, it must be a singular control for which λ(t) = 0. But λ(t) = 0 is the same as the condition (11.77), which means that ∗ if q (t) is strictly increasing, then it must coincide with qˆ(t) in (11.78). It, therefore, only remains to determine the intervals over which q ∗ (t) is constant. Consider Fig. 11.7 qˆ(t); q * (t)

q * (t)

q * (t)

q * (t)

q * (t)

μ=0 λ 0 a constant, and and U (C) = (C − C) γ−δ > 0 a constant satisfying (γ−δ)(1−θ) < ρ < γ−δ. Let β = γ−δ and βT )/β > ¯ assume θ = 1/2 for simplicity. Also assume that K0 eβT + C(1−e KT for the problem to be well-posed (note that the left-hand side of this inequality is the amount of capital at T associated with the consumption ¯ Solve this problem to obtain explicit expressions for the optimal rate C). consumption rate and the associated capital and the adjoint trajectories. E 11.2 Perform the following: (a) Obtain the value of k¯ in Fig. 11.1 from Eq. (11.17). (b) Show that the graph of k˙ = 0 starts from +∞ when k = 0, deˆ at k, ˆ and then increases. Also obtain creases to a minimum of λ ¯ the expression for λ. ˆ (c) Show that k¯ < k.

362

11. Applications to Economics

E 11.3 Use (11.14) to show that h (λ) < 0. Then, conclude that the directions of the horizontal arrows above and below the k˙ = 0 curve are as drawn in Fig. 11.1. ¯ there exists a unique optimal path, E 11.4 Show that for any k0 > k, such as that shown by the solid curve in Region III of Fig. 11.1. E 11.5 In the formulation of the objective function for the economic growth model in Sect. 11.1.3, we took the position of total utilitarianism. Reformulate and solve the problem if our task is to maximize the present value of the utility of per capita consumption over time. E 11.6 Use the phase diagram method to solve the advertising model of (7.7) with its objective function replaced by ∞ −ρt max J = e [π(G) − c(u)]dt , u≥0

0

where c(u) represents an increasing convex advertising cost function with c(u) ≥ 0, c (u) ≥ 0, and c (u) > 0 for u ≥ 0. This is the model of Gould (1970). E 11.7 A variation of the optimal capital accumulation model with stationary population, known as Ramsey’s model, is: ∞ [u(c) − B]dt max J = 0

subject to k˙ = f (k) − c − γk, k(0) = k0 , where B = sup u(c) > 0 c≥0

is the so-called Bliss point, lim u[c(t)] = B

t→∞

so that the integral in the objective function converges, lim u [c(t)] = 0; see Ramsey (1928).

and

t→∞

(a) Show that the optimal capital stock trajectory satisﬁes the diﬀerential equation ˙ k˙ = B − u(f (k) − γk − k). ˙ u (f (k) − γk − k)

Exercises for Chapter 11

363

(b) From part (a), derive Ramsey’s rule d[u (c(t))] = u (c(t))[γ − f (k(t))]. dt (c) Assume u(c) = 2c − c2 /B and f (k) = αk, where α − γ := β > 0 and β < B/k0 < 2β. Show that the optimal feedback consumption rule is c∗ (k) = 2βk − B and the optimal capital trajectory k ∗ is given by k ∗ (t) =

1 [B − (B − βk0 )e−βt ]. β

E 11.8 Show that the trajectory x0 BLxT shown in Fig. 11.2 is optimal for the epidemic model under the stated assumptions. Assume 0 < xs < N. E 11.9 In (11.29), show by using Green’s theorem that xs = N if ρ/θ < 0. E 11.10 Show that C(v) deﬁned in Sect. 11.3.1 satisﬁes C (v) < 0 if f12 ≥ 0. Hint: Note that the usual concavity and monotonicity conditions on the production function f are f1 > 0, f2 > 0, f11 < 0 and f22 < 0. E 11.11 Show that the P c of Fig. 11.5 increases as δ in Eq. (11.33) increases. E 11.12 Show that (11.56) and (11.58) imply the (global) IC condition (11.52). Hint: The proof is by contradiction. First, begin by supposing that (11.52) is violated for some t > tˆ. Then do the same with t < tˆ. E 11.13 In problem (3.7), the terminal equality constraint b(x(T ), T ) = 0 results in the term βbx (x(T ), T ) in the terminal condition (3.11) on the adjoint variable. In problem (11.59)–(11.63), we have the equality constraint (11.62) on the initial states q(t1 ) and φ(t1 ) instead, which we can write as b((q(t1 ), φ(t1 )), t1 ) = t1 a(q(t1 )) − φ(t1 ) = 0. Now apply (3.11) in a symmetric fashion to obtain the initial conditions (11.67) and (11.68) on the adjoint variables.

Chapter 12

Stochastic Optimal Control In previous chapters we assumed that the state variables of the system are known with certainty. When the variables are outcomes of a random phenomenon, the state of the system is modeled as a stochastic process. Speciﬁcally, we now face a stochastic optimal control problem where the state of the system is represented by a controlled stochastic process. We shall only consider the case when the state equation is perturbed by a Wiener process, which gives rise to the state as a Markov diﬀusion process. In Appendix D.2 we have deﬁned the Wiener process, also known as Brownian motion. In Sect. 12.1, we will formulate a stochastic optimal control problem governed by stochastic diﬀerential equations involving a Wiener process, known as Itˆo equations. Our goal will be to synthesize optimal feedback controls for systems subject to Itˆ o equations in a way that maximizes the expected value of a given objective function. In this chapter, we also assume that the state is (fully) observed. On the other hand, when the system is subject to noisy measurements, we face partially observed optimal control problems. In some important special cases, it is possible to separate the problem into two problems: optimal estimation and optimal control. We discuss one such case in Appendix D.4.1. In general, these problems are very diﬃcult and are beyond the scope of this book. Interested readers can consult some references listed in Sect. 12.5. In Sect. 12.2, we will extend the production planning model of Chap. 6 to allow for some uncertain disturbances. We will obtain an optimal production policy for the stochastic production planning problem thus formulated. In Sect. 12.3, we will solve an optimal stochastic advertising © Springer Nature Switzerland AG 2019 S. P. Sethi, Optimal Control Theory, https://doi.org/10.1007/978-3-319-98237-3 12

365

366

12. Stochastic Optimal Control

problem explicitly. The problem is a modiﬁcation as well as a stochastic extension of the optimal control problem of the Vidale-Wolfe advertising model treated in Sect. 7.2.4. In Sect. 12.4, we will introduce investment decisions in the consumption model of Example 1.3. We will consider both risk-free and risky investments. Our goal will be to ﬁnd optimal consumption and investment policies in order to maximize the discounted value of the utility of consumption over time. In Sect. 12.5, we will conclude the chapter by mentioning other types of stochastic optimal control problems that arise in practice.

12.1

Stochastic Optimal Control

In Appendix D.1 on the Kalman ﬁlter, we obtain the optimal state estimation for linear systems with noise and noisy measurements. In Sect. D.4.1, we see that for stochastic linear-quadratic optimal control problems, the separation principle allows us to solve the problem in two steps: to obtain the optimal estimate of the state and to use it in the optimal feedback control formula for deterministic linear-quadratic problems. In this section we will introduce the possibility of controlling a system governed by Itˆo stochastic diﬀerential equations. In other words, we will introduce control variables into Eq. (D.20). This produces the formulation of a stochastic optimal control problem. It should be noted that for such problems, the separation principle does not hold in general. Therefore, to simplify the treatment, it is often assumed that the state variables are observable, in the sense that they can be directly measured. Furthermore, most of the literature on these problems uses dynamic programming or the Hamilton-Jacobi-Bellman framework rather than a stochastic maximum principle. In what follows, therefore, we will formulate the stochastic optimal control problem under consideration, and provide a brief, informal development of the Hamilton-Jacobi-Bellman equation for the problem. A detailed analysis of the problem is available in Fleming and Rishel (1975). For problems involving jump disturbances, see Davis (1993) for the methodology of optimal control of piecewise deterministic processes. For stochastic optimal control in discrete time, see Bertsekas and Shreve (1996). Let us consider the problem of maximizing

T F (Xt , Ut , t)dt + S(XT , T ) , (12.1) E 0

12.1. Stochastic Optimal Control

367

where Xt is the state at time t and Ut is the control at time t, and together they are required to satisfy the Itˆo stochastic diﬀerential equation dXt = f (Xt , Ut , t)dt + G(Xt , Ut , t)dZt , X0 = x0 ,

(12.2)

where Zt , t ∈ [0, T ] is a standard Wiener process. For convenience in exposition we assume the drift coeﬃcient function F : E1 × E1 × E1 → E1, S : E1 × E1 → E1, f : E1 × E1 × E1 → E1 and the diﬀusion coeﬃcient function G : E 1 × E 1 × E 1 → E 1 , so that (12.2) is a scalar equation. We also assume that the functions F and S are continuous in their arguments and the functions f and G are continuously diﬀerentiable in their arguments. For multidimensional extensions of this problem, see Fleming and Rishel (1975). Since (12.2) is a scalar equation, the subscript t here represents only time t. Thus, writing Xt , Ut , and Zt in place of writing X(t), U (t), and Z(t), respectively, will not cause any confusion and, at the same time, will eliminate the need for writing many parentheses. To solve the problem deﬁned by (12.1) and (12.2), let V (x, t), known as the value function, be the expected value of the objective function (12.1) from t to T, when an optimal policy is followed from t to T, given Xt = x. Then, by the principle of optimality, V (x, t) = max E[F (x, U, t)dt + V (x + dXt , t + dt)]. U

(12.3)

By Taylor’s expansion, we have V (x + dXt , t + dt) = V (x, t) +Vt dt + Vx dXt + 12 Vxx (dXt )2 + 12 Vtt (dt)2 + 12 Vxt dXt dt

(12.4)

+higher-order terms. From (12.2), we can formally write (dXt )2 = f 2 (dt)2 + G2 (dZt )2 + 2f GdZt dt, 2

dXt dt = f (dt) + GdZt dt.

(12.5) (12.6)

The exact meaning of these expressions comes from the theory of stochastic calculus; see Arnold (1974, Chapter 5), Durrett (1996) or Karatzas and Shreve (1997). For our purposes, it is suﬃcient to know the multiplication rules of stochastic calculus: (dZt )2 = dt, dZt dt = 0, dt2 = 0.

(12.7)

368

12. Stochastic Optimal Control

Substitute (12.4) into (12.3) and use (12.5), (12.6), (12.7), and the property that E[dZt ] = 0 to obtain

1 2 V = max F dt + V + Vt dt + Vx f dt + Vxx G dt + o(dt) . (12.8) U 2 Note that we have suppressed the arguments of the functions involved in (12.8). Canceling the term V on both sides of (12.8), dividing the remainder by dt, and letting dt → 0, we obtain the Hamilton-Jacobi-Bellman (HJB) equation 1 (12.9) 0 = max[F + Vt + Vx f + Vxx G2 ] U 2 for the value function V (t, x) with the boundary condition V (x, T ) = S(x, T ).

(12.10)

Just as we had introduced a current-value formulation of the maximum principle in Sect. 3.3, let us derive a current-value version of the HJB equation here. For this, in a way similar to (3.29), we write the objective function to be maximized as E 0

T

[φ(Xt , Ut )e−ρt + ψ(XT )e−ρT ].

(12.11)

We can relate this to (12.1) by setting F (Xt , Ut , t) = φ(Xt , Ut )e−ρt and S(XT , T ) = ψ(XT )e−ρT .

(12.12)

It is important to mention that the explicit dependence on time t in (12.11) is only via the discounting term. If it were not the case, there would be no advantage in formulating the current-value version of the HJB equation. Rather than develop the current-value HJB equation in a manner of developing (12.9), we will derive it from (12.9) itself. For this we deﬁne the current-valued value function V˜ (x, t) = V (x, t)eρt .

(12.13)

Then we have Vt = V˜t e−ρt − ρV˜ e−ρt , Vx = V˜x e−ρt and Vxx = V˜xx e−ρt .

(12.14)

12.1. Stochastic Optimal Control

369

By using these and (12.12) in (12.9), we obtain 1 0 = max[φe−ρt + V˜ e−ρt − ρV˜ e−ρt + Vx f e−ρt + Vxx G2 e−ρt ]. U 2 Multiplying by eρt and rearranging terms, we get 1 ρV˜ = max[φ + V˜t + V˜x f + V˜xx G2 ]. U 2

(12.15)

Moreover, from (12.12), (12.13), and (12.10), we can get the boundary condition V˜ (x, T ) = ψ(x). (12.16) Thus, we have obtained (12.15) and (12.16) as the current-value HJB equation. To obtain its inﬁnite-horizon version, it is generally the case that we remove the explicit dependence on t from the function f and G in (12.2), and also assume that ψ ≡ 0. With that, the dynamics (12.2) and the objective function (12.11) change, respectively, to dXt = f (Xt , Ut )dt + G(Xt , Ut )dZt , X0 = x0 ,

∞

E 0

φ(Xt , Ut )e−ρt dt.

(12.17) (12.18)

It should then be obvious that V˜t = 0, and we can obtain the inﬁnitehorizon version of (12.15) as 1 ρV˜ = max[φ + V˜x f + V˜xx G2 ]. U 2

(12.19)

As for its boundary condition, (12.16) is replaced by a growth condition that is the same, in general, as the growth of the function φ(x, U ) in x. For example, if φ(x, U ) is quadratic in x, we would look for a value function V˜ (x) to be of quadratic growth. See Beyer et al. (2010), Chapter 3, for a related discussion of a polynomial growth case in the discrete time setting. If we can ﬁnd a solution of the HJB equation with the given boundary condition (or an appropriate growth condition in the inﬁnite horizon case), then a result called a veriﬁcation theorem suggests that we can construct an optimal feedback control U ∗ (x, t) (or U ∗ (x) in the inﬁnite horizon case) by maximizing the right-hand side of the HJB equation

370

12. Stochastic Optimal Control

with respect U. For further details and extension when the value function is not smooth enough and thus not a classical solution of the HJB equation, see Fleming and Rishel (1975), Yong and Zhou (1999), and Fleming and Soner (1992). In the next three sections, we will apply this procedure to solve problems in production, marketing and ﬁnance.

12.2

A Stochastic Production Inventory Model

In Sect. 6.1.1, we formulated a deterministic production-inventory model. In this section, we extend a simpliﬁed version of that model by including a random process. Let us deﬁne the following quantities: It = the inventory level at time t (state variable), Pt = the production rate at time t (control variable), S = the constant demand rate at time t; S > 0, T

= the length of planning period,

I0 = the initial inventory level, B = the salvage value per unit of inventory at time T, Zt = the standard Wiener process, σ = the constant diﬀusion coeﬃcient. The inventory process evolves according to the stock-ﬂow equation stated as the Itˆ o stochastic diﬀerential equation dIt = (Pt − S)dt + σdZt , I0 given,

(12.20)

where I0 denotes the initial inventory level. As mentioned in Appendix Sect. D.2, the process dZt can be formally expressed as w(t)dt, where w(t) is considered to be a white noise process; see Arnold (1974). It can be interpreted as “sales returns,” “inventory spoilage,” etc., which are random in nature. The objective function is: T 2 2 (Pt + It )dt . (12.21) max E BIT − 0

It can be interpreted as maximization of the terminal salvage value less the cost of production and inventory assumed to be quadratic. In Exercise 12.1, you will be asked to solve the problem with the objective

12.2. A Stochastic Production Inventory Model

371

function given by the expected value of the undiscounted version of the integral in (6.2). As in Sect. 6.1.1 we do not restrict the production rate to be nonnegative. In other words, we permit disposal (i.e., Pt < 0). While this is done for mathematical expedience, we will state conditions under which a disposal is not required. Note further that the inventory level is allowed to be negative, i.e., we permit backlogging of demand. The solution of the above model due to Sethi and Thompson (1981a) will be carried out via the previous development of the HJB equation satisﬁed by a certain value function. Let V (x, t) denote the expected value of the objective function from time t to the horizon T with It = x and using the optimal policy from t to T. The function V (x, t) is referred to as the value function, and it satisﬁes the HJB equation 1 0 = max[−(P 2 + x2 ) + Vt + Vx (P − S) + σ 2 Vxx ] P 2 with the boundary condition V (x, T ) = Bx.

(12.22)

(12.23)

Note that these are applications of (12.9) and (12.10) to the production planning problem. It is now possible to maximize the expression inside the bracket of (12.22) with respect to P by taking its derivative with respect to P and setting it to zero. This procedure yields Vx (x, t) . 2 Substituting (12.24) into (12.22) yields the equation P ∗ (x, t) =

(12.24)

1 Vx2 (12.25) − x2 + Vt − SVx + σ 2 Vxx , 4 2 which, after the max operation has been performed, is known as the Hamilton-Jacobi equation. This is a partial diﬀerential equation which must be satisﬁed by the value function V (x, t) with the boundary condition (12.23). The solution of (12.25) is considered in the next section. 0=

Remark 12.1 It is important to remark that if the production rate were restricted to be nonnegative, then, as in Remark 6.1, (12.24) would be changed to

Vx (x, t) ∗ P (x, t) = max 0, . (12.26) 2

372

12. Stochastic Optimal Control

Substituting (12.26) into (12.23) would give us a partial diﬀerential equation which must be solved numerically. We will not consider (12.26) further in this chapter.

12.2.1

Solution for the Production Planning Problem

To solve Eq. (12.25) with the boundary condition (12.23) we let V (x, t) = Q(t)x2 + R(t)x + M (t).

(12.27)

˙ 2 + Rx ˙ + M˙ , Vt = Qx

(12.28)

Vx = 2Qx + R,

(12.29)

Then,

Vxx = 2Q,

(12.30)

where Y˙ denotes dY /dt. Substituting (12.28)–(12.30) in (12.25) and collecting terms gives 2

R − RS + σ 2 Q = 0. x2 [Q˙ + Q2 − 1] + x[R˙ + RQ − 2SQ] + M˙ + 2

(12.31)

Since (12.31) must hold for any value of x, we must have Q˙ = 1 − Q2 , Q(T ) = 0, R˙ = 2SQ − RQ, R(T ) = B, R2 M˙ = RS − − σ 2 Q, M (T ) = 0, 4

(12.32) (12.33) (12.34)

where the boundary conditions for the system of simultaneous diﬀerential equations (12.32), (12.33), and (12.34) are obtained by comparing (12.27) with the boundary condition V (x, T ) = Bx of (12.23). ˙ To solve (12.32), we expand Q/(1 − Q2 ) by partial fractions to obtain

1 1 Q˙ + = 1, 2 1−Q 1+Q which can be easily integrated. The answer is y−1 , y+1

(12.35)

y = e2(t−T ) .

(12.36)

Q= where

12.2. A Stochastic Production Inventory Model

373

Since S is assumed to be a constant, we can reduce (12.33) to R˙ 0 + R0 Q = 0, R0 (T ) = B − 2S by the change of variable deﬁned by R0 = R − 2S. Clearly the solution is given by T Q(τ )dτ , log R0 (T ) − log R0 (t) = − t

which can be simpliﬁed further to obtain

√ 2(B − 2S) y . R = 2S + y+1

(12.37)

Having obtained solutions for R and Q, we can easily express (12.34) as T [R(τ )S − (R(τ ))2 /4 − σ 2 Q(τ )]dτ . (12.38) M (t) = − t

The optimal control is deﬁned by (12.24), and the use of (12.35) and (12.37) yields √ (y − 1)x + (B − 2S) y . (12.39) P ∗ (x, t) = Vx /2 = Qx + R/2 = S + y+1 This means that the optimal production rate for t ∈ [0, T ] Pt∗ = P ∗ (It∗ , t) = S +

(e2(t−T ) − 1)It∗ + (B − 2S)e(t−T ) , e2(t−T ) + 1

(12.40)

where It∗ , t ∈ [0, T ], is the inventory level observed at time t when using the optimal production rate Pt∗ , t ∈ [0, T ], according to (12.40). Remark 12.2 The optimal production rate in (12.39) equals the demand rate plus a correction term which depends on the level of inventory and the distance from the horizon time T. Since (y − 1) < 0 for t < T, it is clear that for lower values of x, the optimal production rate is likely to be positive. However, if x is very high, the correction term will become smaller than −S, and the optimal control will be negative. In other words, if inventory level is too high, the factory can save money by disposing a part of the inventory resulting in lower holding costs. Remark 12.3 If the demand rate S were time-dependent, it would have changed the solution of (12.33). Having computed this new solution in place of (12.37), we can once again obtain the optimal control as P ∗ (x, t) = Qx + R/2.

374

12. Stochastic Optimal Control

Remark 12.4 Note that when T → ∞, we have y → 0 and P ∗ (x, t) → S − x,

(12.41)

but the undiscounted objective function value (12.21) in this case becomes −∞. Clearly, any other policy will render the objective function value to be −∞. In a sense, the optimal control problem becomes illposed. One way to get out of this diﬃculty is to impose a nonzero discount rate. You are asked to carry this out in Exercise 12.2. Remark 12.5 It would help our intuition if we could draw a picture of the path of the inventory level over time. Since the inventory level is a stochastic process, we can only draw a typical sample path. Such a sample path is shown in Fig. 12.1. If the horizon time T is long enough, the optimal control will bring the inventory level to the goal level x ¯ = 0. It will then hover around this level until t is suﬃciently close to the horizon T. During the ending phase, the optimal control will try to build up the inventory level in response to a positive valuation B for ending inventory.

Xt 5 4

Figure drawn for: x 0 = 2, T = 12, B = 20 S = 5, s = 2

3 2 1 t

0 T -1 -2

Figure 12.1: A sample path of optimal production rate It∗ with I0 = x0 > 0 and B > 0

12.3. The Sethi Advertising Model

12.3

375

The Sethi Advertising Model

In this section, we will discuss a stochastic advertising model due to Sethi (1983b). The model is: ⎧ ∞

⎪ −ρt 2 ⎪ max E e (πXt − Ut )dt ⎪ ⎪ ⎪ 0 ⎪ ⎪ ⎪ ⎪ ⎨ subject to ⎪ √ ⎪ ⎪ dXt = (rUt 1 − Xt − δXt )dt + σ(Xt )dZt , X0 = x0 , ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ Ut ≥ 0,

(12.42)

where Xt is the market share and Ut is the rate of advertising at time t, and where, as speciﬁed in Sect. 7.2.1, ρ > 0 is the discount rate, π > 0 is the proﬁt margin on sales, r > 0 is the advertising eﬀectiveness parameter, and δ > 0 is the sales decay parameter. Furthermore, Zt is the standard one-dimensional Wiener process and σ(x) is the diﬀusion coefﬁcient function having some properties to be speciﬁed shortly. The term in the integrand represents the discounted proﬁt rate at time t. Thus, the integral represents the total value of the discounted proﬁt stream on a sample path. The objective in (12.42) is, therefore, to maximize the expected value of the total discounted proﬁt. The Sethi model is a modiﬁcation as well as a stochastic extension of the optimal control formulation of the Vidale-Wolfe advertising model presented in (7.43). The Itˆ o equation in (12.42) modiﬁes the√VidaleWolfe dynamics (7.25) by replacing the term rU (1 − x) by rUt 1 − Xt and adding a diﬀusion term σ(Xt )dZt on the right-hand side. Furthermore, the linear cost of advertising U in (7.43) is replaced by a quadratic cost of advertising Ut2 in (12.42). The control constraint 0 ≤ U ≤ Q in (7.43) is replaced by simply Ut ≥ 0. The addition of the diﬀusion term yields a stochastic optimal control problem as expressed in (12.42). An important consideration in choosing the function σ(x) should be that the solution Xt to the Itˆ o equation in (12.42) remains inside the interval [0, 1]. Merely requiring that the initial condition x0 ∈ [0, 1], as in Sect. 7.2.1, is no longer suﬃcient in the stochastic case. Additional conditions need to be imposed. It is possible to specify these conditions by using the theory presented by Gihman and Skorohod (1972) for stochastic diﬀerential equations on a ﬁnite spatial interval. In our case, the conditions boil down to the following, in addition to x0 ∈ (0, 1), which

376

12. Stochastic Optimal Control

has been assumed already in (12.42): σ(x) > 0, x ∈ (0, 1) and σ(0) = σ(1) = 0.

(12.43)

It is possible to show that for any feedback control U (x) satisfying U (x) ≥ 0, x ∈ (0, 1], and U (0) > 0,

(12.44)

the Itˆ o equation in (12.42) will have a solution Xt such that 0 < Xt < 1, almost surely (i.e., with probability 1). Since our solution for the optimal advertising U ∗ (x) would turn out to satisfy (12.44), we will have the optimal market share Xt∗ lie in the interval (0, 1). Let V (x) denote the value function for the problem, i.e., V (x) is the expected value of the discounted proﬁts from time t to inﬁnity, when Xt = x and an optimal policy Ut∗ is followed from time t onwards. Note that since T = ∞, the future looks the same from any time t, and therefore the value function does not depend on t. It is for this reason that we have deﬁned the value function as V (x), rather than V (x, t) as in the previous section. Using now the principle of optimality as in Sect. 12.1, we can write the HJB equation as √ 4 $ ρV (x) = max πx − U 2 + Vx (rU 1 − x − δx) + Vxx (σ(x))2 /2 . U

(12.45) Maximization of the RHS of (12.45) can be accomplished by taking its derivative with respect to U and setting it to zero. This gives √ rVx 1 − x ∗ U (x) = . (12.46) 2 Substituting of (12.46) in (12.45) and simplifying the resulting expression yields the HJB equation ρV (x) = πx +

1 Vx2 r2 (1 − x) − Vx δx + σ 2 (x)Vxx . 4 2

(12.47)

As shown in Sethi (1983b), a solution of (12.47) is ¯2 2 ¯ +λ r , V (x) = λx 4ρ where

# 2 2 ¯ = (ρ + δ) + r π − (ρ + δ) , λ 2 r /2

(12.48)

(12.49)

12.4. An Optimal Consumption-Investment Problem

377

as derived in Exercise 7.37. In Exercise 12.3, you are asked verify that (12.48) and (12.49) solve the HJB equation (12.47). We can now obtain the explicit formula for the optimal feedback control as √ ¯ 1−x rλ ∗ . (12.50) U (x) = 2 Note that U ∗ (x) satisﬁes the conditions in (12.44). As in Exercise 7.37, it is easy to characterize (12.50) as ⎧ ⎪ ⎪ ¯ if Xt < X, ¯ ⎪ >U ⎪ ⎪ ⎨ ¯ ¯ if Xt = X, (12.51) Ut∗ = U ∗ (Xt ) = =U ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ X, where ¯= X and

¯ r2 λ/2 ¯ +δ r2 λ/2

√ ¯ 1−x ¯ rλ ¯ U= , 2

(12.52)

(12.53)

as given in (7.51). The market share trajectory for Xt is no longer monotone because of the random variations caused by the diﬀusion term σ(Xt )dZt in the Itˆo equation in (12.42). Eventually, however, the market share process hovers around the equilibrium level x ¯. It is, in this sense and as in the previous section, also a turnpike result in a stochastic environment.

12.4

An Optimal Consumption-Investment Problem

In Example 1.3 in Chap. 1, we had formulated a problem faced by Rich Rentier who wants to consume his wealth in a way that will maximize his total utility of consumption and bequest. In that example, Rich Rentier kept his money in a savings plan earning interest at a ﬁxed rate of r > 0. In this section, we will oﬀer Rich the possibility of investing a part of his wealth in a risky security or stock that earns an expected rate of return that equals α > r. Rich, now known as Rich Investor, must optimally allocate his wealth between the risk-free savings account and

378

12. Stochastic Optimal Control

the risky stock over time and consume over time so as to maximize his total utility of consumption. We will assume an inﬁnite horizon problem in lieu of the bequest, for convenience in exposition. One could, however, argue that Rich’s bequest would be optimally invested and consumed by his heir, who in turn would leave a bequest that would be optimally invested and consumed by a succeeding heir and so on. Thus, if Rich considers the utility accrued to all his heirs as his own, then he can justify solving an inﬁnite horizon problem without a bequest. In order to formulate the stochastic optimal control problem of Rich Investor, we must ﬁrst model his investments. The savings account is easy to model. If S0 is the initial deposit in the savings account earning an interest at the rate r > 0, then we can write the accumulated amount St at time t as St = S0 ert . This can be expressed as a diﬀerential equation, dSt /dt = rSt , which we will rewrite as dSt = rSt dt, S0 ≥ 0. (12.54) Modeling the stock is much more complicated. Merton (1971) and Black and Scholes (1973) have proposed that the stock price Pt can be modeled by an Itˆo equation, namely, dPt = αdt + σdZt , P0 > 0, Pt

(12.55)

dPt = αPt dt + σPt dZt , P0 > 0,

(12.56)

or simply, where P0 > 0 is the given initial stock price, α is the average rate of return on stock, σ is the standard deviation associated with the return, and Zt is a standard Wiener process. Remark 12.6 The LHS in (12.55) can be written also as dlnPt . Another name for the process Zt is Brownian Motion. Because of these, the price process Pt given by (12.55) is often referred to as a logarithmic Brownian Motion. It is important to note from (12.56) that Pt remains nonnegative at any t > 0 on account of the fact that the price process has almost surely continuous sample paths (see Sect. D.2). This property nicely captures the limited liability that is incurred in owning a share of stock. In order to complete the formulation of Rich’s stochastic optimal control problem, we need the following additional notation: Wt = the wealth at time t,

12.4. An Optimal Consumption-Investment Problem

379

Ct = the consumption rate at time t, Qt = the fraction of the wealth invested in stock at time t, 1 − Qt = the fraction of the wealth kept in the savings account at time t, U (C) = the utility of consumption when consumption is at the rate C; the function U (C) is assumed to be increasing and concave, ρ = the rate of discount applied to consumption utility, B = the bankruptcy parameter, to be explained later. Next we develop the dynamics of the wealth process. Since the investment decision Q is unconstrained, it means Rich is allowed to buy stock as well as to sell it short. Moreover, Rich can deposit in, as well as borrow money from, the savings account at the rate r. While it is possible to rigorously obtain the equation for the wealth process involving an intermediate variable, namely, the number Nt of shares of stock owned at time t, we will not do so. Instead, we will write the wealth equation informally as dWt = Qt Wt αdt + Qt Wt σdZt + (1 − Qt )Wt rdt − Ct dt = (α − r)Qt Wt dt + (rWt − Ct )dt + σQt Wt dZt , W0 given, (12.57) and provide an intuitive explanation for it. The term Qt Wt αdt represents the expected return from the risky investment of Qt Wt dollars during the period from t to t+dt. The term Qt Wt σdZt represents the risk involved in investing Qt Wt dollars in stock. The term (1 − Qt )Wt rdt is the amount of interest earned on the balance of (1 − Qt )Wt dollars in the savings account. Finally, Ct dt represents the amount of consumption during the interval from t to t + dt. In deriving (12.57), we have assumed that Rich can trade continuously in time without incurring any broker’s commission. Thus, the change in wealth dWt from time t to time t + dt is due to consumption as well as the change in share price. For a rigorous development of (12.57) from (12.54) and (12.55), see Harrison and Pliska (1981). Since Rich can borrow an unlimited amount and invest it in stock, his wealth could fall to zero at some time T. We will say that Rich goes bankrupt at time T, when his wealth falls zero at that time. It is clear that T is a random variable deﬁned as T = inf{t ≥ 0|Wt = 0}.

(12.58)

380

12. Stochastic Optimal Control

This special type of random variable is called a stopping time, since it is observed exactly at the instant of time when wealth falls to zero. We can now specify Rich’s objective function. It is:

T e−ρt U (Ct )dt + e−ρT B , (12.59) max J = E 0

where we have assumed that Rich experiences a payoﬀ of B, in the units of utility, at the time of bankruptcy. B can be positive if there is a social welfare system in place, or B can be negative if there is remorse associated with bankruptcy. See Sethi (1997a) for a detailed discussion of the bankruptcy parameter B. Let us recapitulate the optimal control problem of Rich Investor: ⎧ T

⎪ −ρt −ρT ⎪ max J = E e U (Ct )dt + e B ⎪ ⎪ ⎪ 0 ⎪ ⎪ ⎪ ⎪ ⎨ subject to ⎪ ⎪ ⎪ dWt = (α − r)Qt Wt dt + (rWt − Ct )dt + σQt Wt dZt , W0 given, ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ Ct ≥ 0. (12.60) As in the inﬁnite horizon problem of Sect. 12.2, here also the value function is stationary with respect to time t. This is because T is a stopping time of bankruptcy, and the future evolution of wealth, investment, and consumption processes from any starting time t depends only on the wealth at time t and not on time t itself. Therefore, let V (x) be the value function associated with an optimal policy beginning with wealth Wt = x at time t. Using the principle of optimality as in Sect. 12.1, the HJB equation satisﬁed by the value function V (x) for problem (12.60) can be written as ⎧ ⎪ ⎪ ρV (x) = max [(α − r)QxVx + (rx − C)Vx ⎪ ⎪ C≥0,Q ⎪ ⎨ (12.61) +(1/2)Q2 σ 2 x2 Vxx + U (C)], ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ V (0) = B. This problem and a number of its generalizations are solved explicitly in Sethi (1997a). Here we shall conﬁne ourselves in solving a simpler problem resulting from the following considerations.

12.4. An Optimal Consumption-Investment Problem

381

It is shown in Karatzas et al. (1986), reproduced as Chapter 2 in Sethi (1997a), that when B ≤ U (0)/ρ, no bankruptcy will occur. This should be intuitively obvious because if Rich goes bankrupt at any time T > 0, he receives B at that time, whereas by not going bankrupt at that time he reaps the utility of strictly more than U (0)/ρ on account of consumption from time T onward. It is shown furthermore that if U (0) = ∞, then the optimal consumption rate will be strictly positive. This is because even an inﬁnitesimally small positive consumption rate results in a proportionally large amount of utility on account of the inﬁnite marginal utility at zero consumption level. A popular utility function used in the literature is U (C) = lnC,

(12.62)

which was also used in Example 1.3. This function gives an inﬁnite marginal utility at zero consumption, i.e., U (0) = 1/C|C=0 = ∞.

(12.63)

We also assume B = U (0)/ρ = −∞. These assumptions imply a strictly positive consumption level at all times and no bankruptcy. Since Q is already unconstrained, having no bankruptcy and only positive (i.e., interior) consumption level allows us to obtain the form of the optimal consumption and investment policy simply by diﬀerentiating the RHS of (12.61) with respect to Q and C and equating the resulting expressions to zero. Thus, (α − r)xVx + Qσ 2 x2 Vxx = 0, i.e., Q∗ (x) = −

(α − r)Vx , xσ 2 Vxx

(12.64)

and

1 . (12.65) Vx Substituting (12.64) and (12.65) in (12.61) allows us to remove the max operator from (12.61), and provides us with the equation 1 γ(Vx )2 + rx − (12.66) Vx − lnVx , ρV (x) = − Vxx Vx C ∗ (x) =

where γ=

(α − r)2 . 2σ 2

(12.67)

382

12. Stochastic Optimal Control

This is a nonlinear ordinary diﬀerential equation that appears to be quite diﬃcult to solve. However, Karatzas et al. (1986) used a change of variable that transforms (12.66) into a second-order, linear, ordinary diﬀerential equation, which has a known solution. For our purposes, we will simply guess that the value function is of the form V (x) = A ln x + B,

(12.68)

where A and B are constants, and obtain the values of A and B by substitution in (12.66). Using (12.68) in (12.66), we see that ' A x( A − ln ρA ln x + ρB = γA + rx − A x x = γA + rA − 1 − ln A + ln x. By comparing the coeﬃcients of ln x and the constants on both sides, we get A = 1/ρ and B = (r − ρ + γ)/ρ2 + ln ρ/ρ. By substituting these values in (12.68), we obtain V (x) =

r−ρ+γ 1 ln(ρx) + , x ≥ 0. ρ ρ2

(12.69)

In Exercise 12.4, you are asked by a direct substitution in (12.66) to verify that (12.69) is indeed a solution of (12.66). Moreover, V (x) deﬁned in (12.69) is strictly concave, so that our concavity assumption made earlier is justiﬁed. From (12.69), it is easy to show that (12.64) and (12.65) yield the following feedback policies: α−r , σ2 C ∗ (x) = ρx.

Q∗ (x) =

(12.70) (12.71)

The investment policy (12.70) says that the optimal fraction of the wealth invested in the risky stock is (α − r)/σ 2 , i.e., α−r , t ≥ 0, (12.72) σ2 which is a constant over time. The optimal consumption policy is to consume a constant fraction ρ of the current wealth, i.e., Q∗t = Q∗ (Wt ) =

Ct∗ = C ∗ (Wt ) = ρWt , t ≥ 0.

(12.73)

This problem and its many extensions have been studied in great detail. See, e.g., Sethi (1997a).

Exercises for Chapter 12

12.5

383

Concluding Remarks

In this chapter, we have considered stochastic optimal control problems subject to Itˆo diﬀerential equations. For impulse stochastic control, see Bensoussan and Lions (1984). For stochastic control problems with jump Markov processes or, more generally, martingale problems, see Fleming and Soner (1992), Davis (1993), and Karatzas and Shreve (1998). For problems with incomplete information or partial observation, see Bensoussan (2004, 2018), Elliott et al. (1995), and Bensoussan et al. (2010). For applications of stochastic optimal control to manufacturing problems, see Sethi and Zhang (1994a), Yin and Zhang (1997), Sethi et al. (2005), Bensoussan (2011), and Bensoussan et al. (2007b,c,d, 2008a,b, 2009a,b,c). For applications to problems in ﬁnance, see Sethi (1997a), Karatzas and Shreve (1998), and Bensoussan et al. (2009d). For applications in marketing, see Tapiero (1988), Raman (1990), and Sethi and Zhang (1995b). For applications of stochastic optimal control to economics including economics of natural resources, see, e.g., Pindyck (1978a,b), Rausser and Hochman (1979), Arrow and Chang (1980), Derzko and Sethi (1981a), Bensoussan and Lesourne (1980, 1981), Malliaris and Brock (1982), and Brekke and Øksendal (1994). Exercises for Chapter 12 E 12.1 Solve the production-inventory problem with the state equation (12.20) and the objective function

min J = E 0

T

h c 2 2 ˆ ˆ [ (I − I) + (P − P ) ]dt , 2 2

where h > 0, c > 0, Iˆ ≥ 0 and Pˆ ≥ 0; see the objective function (6.2) for the interpretation of these parameters. E 12.2 Formulate and solve the discounted inﬁnite-horizon version of the stochastic production planning model of Sect. 12.2. Speciﬁcally, assume B = 0 and replace the objective function in (12.21) by ∞ −ρt 2 2 −e (Pt + It )dt . max E 0

E 12.3 Verify by direct substitution that the value function deﬁned by (12.48) and (12.49) solves the HJB equation (12.47).

384

12. Stochastic Optimal Control

E 12.4 Verify by direct substitution that the value function in (12.69) solves the HJB equation (12.66). E 12.5 Solve the consumption-investment problem (12.60) with the utility function U (C) = C β , 0 < β < 1, and B = 0. E 12.6 Solve Exercise 12.5 when U (C) = −C β with β < 0 and B = −∞. E 12.7 Solve the optimal consumption-investment problem: ∞

−ρt V (x) = max J = E e ln(Ct − s)dt 0

subject to dWt = (α − r)Qt Wt dt + (rWt − Ct )dt + σQt Wt dZt , W0 = x, Ct ≥ s. Here s > 0 denotes a minimal subsistence consumption, and we assume 0 < ρ < 1. Note that the value function V (s/r) = −∞. Guess a solution of the form V (x) = A ln(x − s/r) + B. Find the constants A, B, and the optimal feedback consumption and investment allocation policies C ∗ (x) and Q∗ (x), respectively. Characterize these policies in words. E 12.8 Solve the consumption-investment problem: ∞

V (x) = max J = E e−ρt (Ct − s)β dt 0

subject to dWt = (α − r)Qt Wt dt + (rWt − Ct )dt + σQt Wt dZt , W0 = x, Ct ≥ s. Here s > 0 denotes a minimal subsistence consumption and we assume 0 < ρ < 1 and 0 < β < 1. Note that the value function V (s/r) = 0. Therefore, guess a solution of the form V (x) = A(x − s/r)β . Find the constant A and the optimal feedback consumption and investment allocation policies C ∗ (x) and Q∗ (x), respectively. Characterize these policies in words.

Chapter 13

Diﬀerential Games In previous chapters, we were mainly concerned with the optimal control problems formulated by a single objective function (or a single decision maker). However, there are situations when there may be more than one decision maker, each having one’s own objective function that each is trying to maximize, subject to a set of diﬀerential equations. This extension of optimal control theory is referred to as the theory of diﬀerential games. The study of diﬀerential games was initiated by Isaacs (1965). After the development of Pontryagin’s maximum principle, it became clear that there was a connection between diﬀerential games and optimal control theory. In fact, diﬀerential game problems represent a generalization of optimal control problems in cases where there is more than one controller or player. However, diﬀerential games are conceptually far more complex than optimal control problems in the sense that it is no longer obvious what constitutes a solution; see Starr and Ho (1969), Ho (1970), Varaiya (1970), Friedman (1971), Leitmann (1974), Case (1979), Selten (1975), Mehlmann (1988), Berkovitz (1994), Basar and Olsder (1999), Dockner et al. (2000), and Basar et al. (2010). Indeed, there are a number of diﬀerent types of solutions such as minimax, Nash, Stackelberg, along with possibilities of cooperation and bargaining; see, e.g., Tolwinski (1982) and Haurie et al. (1983). We will discuss minimax solutions for zero-sum diﬀerential games in Sect. 13.1, Nash solutions for nonzero-sum games in Sect. 13.2, and Stackelberg diﬀerential games in Sect. 13.3.

© Springer Nature Switzerland AG 2019 S. P. Sethi, Optimal Control Theory, https://doi.org/10.1007/978-3-319-98237-3 13

385

386

13. Diﬀerential Games

13.1

Two-Person Zero-Sum Diﬀerential Games

Consider the state equation x˙ = f (x, u, v, t), x(0) = x0 ,

(13.1)

where we may assume all variables to be scalar for the time being. Extension to the vector case simply requires appropriate reinterpretations of each of the variables and the equations. In this equation, we let u and v denote the controls applied by players 1 and 2, respectively. We assume that u(t) ∈ U, v(t) ∈ V, t ∈ [0, T ], where U and V are convex sets in E 1 . Consider further the objective function T J(u, v) = S[x(T )] + F (x, u, v, t)dt, (13.2) 0

which player 1 wants to maximize and player 2 wants to minimize. Since the gain of player 1 represents a loss to player 2, such games are appropriately termed zero-sum games. Clearly, we are looking for admissible control trajectories u∗ and v ∗ such that J(u∗ , v) ≥ J(u∗ , v ∗ ) ≥ J(u, v ∗ ).

(13.3)

The solution (u∗ , v ∗ ) is known as the minimax solution. Here u∗ and v ∗ stand for u∗ (t), t ∈ [0, T ], and v ∗ (t), t ∈ [0, T ], respectively. The necessary conditions for u∗ and v ∗ to satisfy (13.3) are given by an extension of the maximum principle. To obtain these conditions, we form the Hamiltonian H = F + λf (13.4) with the adjoint variable λ satisfying the equation λ˙ = −Hx , λ(T ) = Sx [x(T )].

(13.5)

The necessary condition for trajectories u∗ and v ∗ to be a minimax solution is that for t ∈ [0, T ], H(x∗ (t), u∗ (t), v ∗ (t), λ(t), t) = min max H(x∗ (t), u, v, λ(t), t), v∈V u∈U

(13.6)

which can also be stated, with suppression of (t), as H(x∗ , u∗ , v, λ, t) ≥ H(x∗ , u∗ , v ∗ , λ, t) ≥ H(x∗ , u, v ∗ , λ, t)

(13.7)

13.2. Nash Diﬀerential Games

387

for u ∈ U and v ∈ V. Note that (u∗ , v ∗ ) is a saddle point of the Hamiltonian function H. Note also that if u and v are unconstrained, i.e., when, U = V = E 1 , condition (13.6) reduces to the ﬁrst-order necessary conditions Hu = 0 and Hv = 0,

(13.8)

and the second-order conditions are Huu ≤ 0 and Hvv ≥ 0.

(13.9)

We now turn to the treatment of nonzero-sum diﬀerential games.

13.2

Nash Diﬀerential Games

In this section, let us assume that we have N players where N ≥ 2. Let ui ∈ U i , i = 1, 2, . . . , N, represent the control variable for the ith player, where U i is the set of controls from which the ith player can choose. Let the state equation be deﬁned as x˙ = f (x, u1 , u2 , . . . , uN , t).

(13.10)

Let J i , deﬁned by i

i

T

J = S [x(T )] +

F i (x, u1 , u2 , . . . , uN , t)dt,

(13.11)

0

denote the objective function which the ith player wants to maximize. In this case, a Nash solution is deﬁned by a set of N admissible trajectories {u1∗ , u2∗ , . . . , uN ∗ },

(13.12)

which have the property that J i (u1∗ , u2∗ , . . . , uN ∗ ) = max J i (u1∗ , u2∗ , . . . , u(i−1)∗ , ui , . . . , u(i+1)∗ , . . . , uN ∗ ) (13.13)

ui ∈U i

for i = 1, 2, . . . , N. To obtain the necessary conditions for a Nash solution for nonzerosum diﬀerential games, we must make a distinction between open-loop and closed-loop controls.

388

13.2.1

13. Diﬀerential Games

Open-Loop Nash Solution

The open-loop Nash solution is deﬁned when the set of trajectories in (13.12) are given as functions of time satisfying (13.13). To obtain the maximum principle type conditions for such solutions to be a Nash solution, let us deﬁne the Hamiltonian functions H i (x, u1 , u2 , . . . , uN , λi ) = F i + λi f

(13.14)

for i = 1, 2, . . . , N, with λi satisfying i λ˙ = −Hxi , λi (T ) = Sxi [x(T )].

(13.15)

The Nash control ui∗ for the ith player is obtained by maximizing the ith Hamiltonian H i with respect to ui , i.e., ui∗ must satisfy H i (x∗ , u1∗ , . . . , u(i−1)∗ , ui∗ , u(i+1)∗ , . . . , uN ∗ , λ, t) ≥ H i (x∗ , u1∗ , . . . , u(i−1)∗ , ui , u(i+1)∗ , . . . , uN ∗ , λ, t), t ∈ [0, T ], (13.16) for all ui ∈ U i , i = 1, 2, . . . , N. Deal et al. (1979) formulated and solved an advertising game with two players and obtained the open-loop Nash solution by solving a twopoint boundary value problem. In Exercise 13.1, you are asked to obtain their boundary value problem. See also Deal (1979).

13.2.2

Feedback Nash Solution

A feedback Nash solution is obtained when (13.12) is deﬁned in terms of the current state of the system. To avoid confusion, we let ui∗ (x, t) = φi (x, t), i = 1, 2, . . . , N.

(13.17)

For these controls to represent a feedback Nash strategy, we must recognize the dependence of the other players’ actions on the state variable x. Therefore, we need to replace the adjoint equation (13.15) by i λ˙ = −Hxi −

N j=1

Hui j φjx = −Hxi −

N

Hui j φjx .

(13.18)

j=1,j=i

The presence of the summation term in (13.18) makes the necessary condition for the feedback solution virtually useless for deriving computational algorithms; see Starr and Ho (1969). It is, however, possible

13.2. Nash Diﬀerential Games

389

to use a dynamic programming approach for solving extremely simple nonzero-sum games, which require the solution of a partial diﬀerential equation. We will use this approach in Sect. 13.3. The troublesome summation term in (13.18) is absent in three important cases: (a) in optimal control problems (N = 1) since Hu ux = 0, (b) in two-person zero-sum games because H 1 = −H 2 so that Hu12 u2x = −Hu22 u2x = 0 and Hu21 u1x = −Hu11 u1x = 0, and (c) in open-loop nonzerosum games because ujx = 0. It certainly is to be expected, therefore, that the feedback and open-loop Nash solutions are going to be diﬀerent, in general. This can be shown explicitly for the linear-quadratic case. We conclude this section by providing an interpretation to the adjoint variable λi . It is the sensitivity of the ith player’s proﬁt to a perturbation in the state vector. If the other players are using closed-loop strategies, any perturbation δx in the state vector causes them to revise their controls by the amount φjx δx. If the ith Hamiltonian H i were maximized with respect to uj , j = i, this would not aﬀect the ith player’s proﬁt; but since ∂H i /∂uj = 0 for i = j, the reactions of the other players to the perturbation inﬂuence the ith player’s proﬁt, and the ith player must account for this eﬀect in considering variations of the trajectory.

13.2.3

An Application to Common-Property Fishery Resources

Consider extending the ﬁshery model of Sect. 10.1 by assuming that there are two producers having unrestricted rights to exploit the ﬁsh stock in competition with each other. This gives rise to a nonzero-sum diﬀerential game analyzed by Clark (1976). Equation (10.2) is modiﬁed by x˙ = g(x) − q 1 u1 x − q 2 u2 x, x(0) = x0 ,

(13.19)

where ui (t) represents the rate of ﬁshing eﬀort and q i ui x is the rate of catch for the ith producer, i = 1, 2. The control constraints are 0 ≤ ui (t) ≤ U i , i = 1, 2,

(13.20)

the state constraints are x(t) ≥ 0,

(13.21)

and the objective function for the ith producer is the total present value of his proﬁts, namely, ∞ Ji = (pi q i x − ci )ui e−ρt dt, i = 1, 2. (13.22) 0

390

13. Diﬀerential Games

To ﬁnd the feedback Nash solution for this model, we let x ¯i denote the turnpike (or optimal biomass) level given by (10.12) on the assumption that the ith producer is the sole-owner of the ﬁshery. Let the bionomic equilibrium xib and the corresponding control uib associated with producer i be deﬁned by (10.4), i.e., xib =

g(xib )pi ci i and u = . b pi q i ci

(13.23)

As shown in Exercise 10.2, xib < x ¯i , and we assume U i to be suﬃciently i i large so that ub ≤ U . We also assume that x1b < x2b ,

(13.24)

which means that producer 1 is more eﬃcient than producer 2, i.e., producer 1 can make a positive proﬁt at any level in the interval (x1b , x2b ], while producer 2 loses money in the same interval, except at x2b , where he breaks even. For x > x2b , both producers make positive proﬁts. Since U 1 ≥ u1b by assumption, producer 1 has the capability of driving the ﬁsh stock level down to at least x1b which, by (13.24), is less than x2b . This implies that producer 2 cannot operate at a sustained level above x2b ; and at a sustained level below x2b , he cannot make a proﬁt. Hence, his optimal feedback policy is bang-bang:

u2∗ (x) =

⎧ ⎪ ⎨ U 2 if x > x2 , b ⎪ ⎩ 0

if x ≤

(13.25)

x2b .

As far as producer 1 is concerned, he wants to attain his turnpike level ¯1 ≤ x2b . If x ¯1 > x2b and x0 ≥ x ¯1 , then from (13.25) producer 2 x ¯1 if x will ﬁsh at his maximum rate until the ﬁsh stock is driven to x2b . At this level, it is optimal for producer 1 to ﬁsh at a rate which maintains the ﬁsh stock at level x2b in order to keep producer 2 from ﬁshing. Thus, the optimal feedback policy for producer 1 can be stated as ⎧ ⎫ ⎪ ⎪ 1 1 ⎪ ⎪ ⎪ if x > x ¯ ⎪ ⎪ U ⎪ ⎪ ⎪ ⎨ ⎬ 1 1∗ g(¯ x ) 1 1 (13.26) ¯1 < x2b , u (x) = u ¯ = q1 x¯1 if x = x ¯ ⎪ , if x ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 0 if x < x ¯1 ⎭

13.2. Nash Diﬀerential Games

u1∗ (x) =

⎧ ⎪ ⎪ ⎪ U1 ⎪ ⎪ ⎨

g(x2b )

q 1 x2b ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 0

if x >

391

x2b

⎫ ⎪ ⎪ ⎪ ⎪ ⎪ ⎬

¯1 ≥ x2b . if x = x2b ⎪ , if x ⎪ ⎪ ⎪ ⎪ 2 if x < x ⎭

(13.27)

b

The formal proof that policies (13.25)–(13.27) give a Nash solution requires direct veriﬁcation using the result of Sect. 10.1.2. The Nash solution for this case means that for all feasible paths u1 and u2 ,

and

J 1 (u1∗ , u2∗ ) ≥ J 1 (u1 , u2∗ ),

(13.28)

J 2 (u1∗ , u2∗ ) ≥ J 2 (u1∗ , u2 ).

(13.29)

The direct veriﬁcation involves deﬁning a modiﬁed growth function ⎧ ⎪ ⎨ g(x) − q 2 U 2 x if x > x2 , b 1 g (x) = ⎪ ⎩ g(x) if x ≤ x2b , and using the Green’s theorem results of Sect. 10.1.2. Since U 2 ≥ u2b by assumption, we have g 1 (x) ≤ 0 for x > x2b . From (10.12) with g replaced by g 1 , it can be shown that the new turnpike level for producer 1 is min(¯ x1 , x2b ), which deﬁnes the optimal policy (13.26)–(13.27) for producer 1. The optimality of (13.25) for producer 2 follows easily. To interpret the results of the model, suppose that producer 1 originally has sole possession of the ﬁshery, but anticipates a rival entry. Producer 1 will switch from his own optimal sustained yield u ¯1 to a more intensive exploitation policy prior to the anticipated entry. We can now guess the results in situations involving N producers. The ﬁshery will see the progressive elimination of ineﬃcient producers as the stock of ﬁsh decreases. Only the most eﬃcient producers will survive. If, ultimately, two or more maximally eﬃcient producers exist, the ﬁshery will converge to a classical bionomic equilibrium, with zero sustained economic rent. We have now seen that a feedback Nash solution involving N ≥ 2 competing producers results in the long-run erosion of economic rents. This conclusion depends on the assumption that producers face an inﬁnitely elastic supply of all factors of production going into the ﬁshing

392

13. Diﬀerential Games

eﬀort, but typically the methods of licensing entrants to regulated ﬁsheries make some attempt to also control the factors of production such as permitting the licensee to operate only a single vessel of speciﬁc size. In order to develop a model for the licensing of ﬁshermen, we let the control variable v i denote the capital stock of the ith producer and let the concave function f (v i ), with f (0) = 0, denote the ﬁshing mortality function for i = 1, 2, . . . , N. This requires the replacement of q i ui in the previous model by f (v i ). The extended model becomes nonlinear in control variables. You are asked in Exercise 13.3 to formulate this new model and develop necessary conditions for a feedback Nash solution for this game involving N producers. The reader is referred to Clark (1976) for further details. For other papers on applications of diﬀerential games to ﬁshery management, see H¨ am¨ al¨ ainen et al. (1984, 1985, 1986, 1990).

13.3

A Feedback Nash Stochastic Diﬀerential Game in Advertising

In this section, we will study a competitive extension of the Sethi advertising model discussed in Sect. 12.3. This will give us a stochastic diﬀerential game, for which we aim to obtain a feedback Nash equilibrium by using a dynamic programming approach developed in Sect. 12.1. We should note that this approach can also be used to obtain feedback Nash equilibria in deterministic diﬀerential games as an alternative to the maximum principle approach developed in Sect. 13.2.2. Speciﬁcally, we consider a duopoly market in a mature product category where total sales are distributed between two ﬁrms, labeled as Firm 1 and Firm 2, which compete for market share through advertising expenditures. We let Xt denote the market share of Firm 1 at time t, so that the market share of Firm 2 is (1 − Xt ). Let U1t and U2t denote the advertising eﬀort rates of Firms 1 and 2, respectively, at time t. Using the subscript i ∈ {1, 2} to reference the two ﬁrms, let ri > 0 denote the advertising eﬀectiveness parameter, π i > 0 denote the sales margin, ρi > 0 denote the discount rate, and ci > 0 denote the cost parameter so that the cost of advertising eﬀort u by Firm i is ci u2 . Further, let δ > 0 be the churn parameter, Zt be the standard one-dimensional Wiener process, and σ(x) be the diﬀusion coeﬃcient function as deﬁned in Sect. 12.3. Then, in view of the competition between the ﬁrms, Prasad and Sethi (2004) extend the Sethi model dynamics in (12.42) as the Itˆo

13.3. A Feedback Nash Stochastic Diﬀerential Game in Advertising 393 stochastic diﬀerential equation # # dXt = [r1 U1t 1 − Xt − δXt − r2 U2t Xt + δ(1 − Xt )]dt + σ(Xt )dZt , X(0) = x0 ∈ [0, 1].

(13.30)

We formulate the optimal control problem faced by the two ﬁrms as ∞ 4 $ 1 −ρ1 t 2 π 1 Xt − c1 U1t dt , (13.31) e max V (x0 ) = E U1 ≥0

0

2

max V (x0 ) = E

U2 ≥0

∞

e 0

−ρ2 t

4

π 2 (1 − Xt ) −

2 c2 U2t

$

dt ,

(13.32)

subject to (13.30). Thus, each ﬁrm seeks to maximize its expected, discounted proﬁt stream subject to the market share dynamics. To ﬁnd the feedback Nash equilibrium solution, we form the Hamilton-Jacobi-Bellman (HJB) equations for the value functions V 1 (x) and V 2 (x) : 1 /2} ρ1 V 1 = max{H 1 (x, U1 , U2 , Vx1 ) + (σ(x))2 Vxx U1 ≥0 √ √ = max{π 1 x − c1 U12 + Vx1 [r1 U1 1 − x − r2 U2 x − δ(2x − 1)] U1 ≥0

1 +(σ(x))2 Vxx /2},

(13.33)

2 /2} ρ2 V 2 = max{H 2 (x, U1 , U2 , Vx2 ) + (σ(x))2 Vxx U2 ≥0

= max{π 2 (1 − x) − c2 U22 U2 ≥0 √ √ +Vx2 [r1 U1 1 − x − r2 U2 x − δ(2x − 1)] 2 /2}, +(σ(x))2 Vxx

(13.34)

where the Hamiltonians are as deﬁned in (13.14). We use the ﬁrst-order conditions for Hamiltonian maximization to obtain the optimal feedback advertising decisions √ √ U1∗ (x) = Vx1 (x)r1 1 − x/2c1 and U2∗ (x) = −Vx2 (x)r2 x/2c2 . (13.35) Since it is reasonable to expect that Vx1 ≥ 0 and Vx2 ≤ 0, these controls will turn out to be nonnegative as we will see later.

394

13. Diﬀerential Games

Substituting (13.35) in (13.33) and (13.34), we obtain the HamiltonJacobi equations ρ1 V 1 = π 1 x + (Vx1 )2 r12 (1 − x)/4c1 + Vx1 Vx2 r22 x/2c2 1 −Vx1 δ(2x − 1) + (σ(x))2 Vxx /2,

(13.36)

ρ2 V 2 = π 2 (1 − x) + (Vx2 )2 r22 x/4c2 + Vx1 Vx2 r12 (1 − x)/2c1 2 −Vx2 δ(2x − 1) + (σ(x))2 Vxx /2.

(13.37)

As in Sect. 12.3, we look for the following forms for the value functions V 1 = α1 + β 1 x and V 2 = α2 + β 2 (1 − x).

(13.38)

These are inserted into (13.36) and (13.37) to determine the unknown coeﬃcients α1 , β 1 , α2 , and β 2 . Equating the coeﬃcients of x and the constants on both sides of (13.36) and the coeﬃcients of (1 − x) and the constants on both sides of (13.37), the following four equations emerge, which can be solved for the unknowns α1 , β 1 , α2 , and β 2 : ρ1 α1 = β 21 r12 /4c1 + β 1 δ, ρ1 β 1 = π 1 − ρ2 α 2 = ρ2 β 2 =

β 21 r12 /4c1

−

β 22 r22 /4c2 + β 2 δ, π 2 − β 22 r22 /4c2 −

(13.39) β 1 β 2 r22 /2c2

− 2β 1 δ,

(13.40) (13.41)

β 1 β 2 r12 /2c1

− 2β 2 δ.

(13.42)

Let us ﬁrst consider the special case of symmetric ﬁrms, i.e., when π = π 1 = π 2 , c = c1 = c2 , r = r1 = r2 , and ρ = ρ1 = ρ2 , and therefore α = α1 = α2 , β = β 1 = β 2 . The four equations in ((13.39)–(13.42)) reduce to the following two: ρα = β 2 r2 /4c + βδ and ρβ = π − 3β 2 r2 /4c − 2βδ.

(13.43)

There are two solutions for β. One is negative, which clearly makes no sense. Thus, the remaining positive solution is the correct one. This also allows us to obtain the corresponding α. The solution is # (13.44) α = [(ρ − δ)(W − W 2 + 12Rπ) + 6Rπ]/18Rρ, # 2 β = ( W + 12Rπ − W )/6R, (13.45) where R = r2 /4c and W = ρ + 2δ. With this the value functions in (13.38) are deﬁned, and the controls in (13.35) for the case of symmetric ﬁrms can be written as √ √ √ √ β r1 1 − x βr 1 − x β r2 x βr x and u∗2 (x) = 2 , = = u∗1 (x) = 1 2c1 2c 2c2 2c

13.4. A Stackelberg Diﬀerential Game of Cooperative Advertising

395

which are clearly nonnegative as required. We return now to the general case of asymmetric ﬁrms. For this, we re-express equations ((13.39)–(13.42)) in terms of a single variable β 1 , which is determined by solving the quartic equation 3R12 β 41 + 2R1 (W1 + W2 )β 31 + (4R2 π 2 − 2R1 π 1 − W12 + 2W1 W2 )β 21 + 2π 1 (W1 − W2 )β 1 − π 21 = 0.

(13.46)

This equation can be solved explicitly to give four roots. We will ﬁnd that only one of these is positive, and select it as our value of β 1 . With that, other coeﬃcients can be obtained by solving for α1 and β 2 and then, in turn, α2 , as follows: α1 = β 1 (β 1 R1 + δ)/ρ1 ,

(13.47)

β 2 = (π 1 −

(13.48)

β 21 R1

− β 1 W1 )/2β 1 R2 ,

α2 = β 2 (β 2 R2 + δ)/ρ2 , r12 /4c1 ,

(13.49)

r22 /4c2 ,

R2 = W1 = ρ1 + 2δ, and W2 = ρ2 + 2δ. where R1 = It is worthwhile to mention that ﬁrm i’s advertising eﬀectiveness parameter ri and advertising cost parameter ci manifest themselves through Ri = ri 2 /4ci . This would suggest that Ri is a measure of ﬁrm i’s advertising power. This can be seen more clearly in Exercise 13.6 involving two ﬁrms that are identical in all other aspects except that R2 > R1 . Speciﬁcally in that exercise, you are asked to use Mathematica or another suitable software program to solve (13.46) to obtain β 1 and then obtain the coeﬃcients α1 , α2 , and β 2 by using (13.47)–(13.49), when ρ1 = ρ2 = 0.05,#π 1 = π 2 = 1, δ = 0.01, R1 = 1, R2 = 4, x0 = 0.5, and σ(x) = 0.5x(1 − x). Figure 13.1 represents a sample path of the market share of the two ﬁrms with this data. It is noteworthy to see that both ﬁrms are identical except in their advertising powers R1 and R2 . With R2 > R1 , ﬁrm 2 is more powerful and we see that this results in its capture of an increasing share of the market average over time beginning with exactly one half of the market at time 0.

13.4

A Feedback Stackelberg Stochastic Diﬀerential Game of Cooperative Advertising

The preceding sections in this chapter dealt with diﬀerential games in which all players make their decisions simultaneously. We now discuss

396

13. Diﬀerential Games

X Y

X Y

Figure 13.1: A sample path of optimal market share trajectories

a diﬀerential game in which two players make their decisions in a hierarchical manner. The player having the right to move ﬁrst is called the leader and the other player is called the follower. If there are two or more leaders, they play Nash, and the same goes for the followers. In terms of solutions of Stackelberg diﬀerential games, we have openloop and feedback solutions. An open-loop Stackelberg equilibrium speciﬁes, at the initial time (say, t = 0), the decisions over the entire horizon. As in Sect. 13.1, there is a maximum principle for open-loop solutions. Typically, open-loop solutions are not time consistent in the sense that at any time t > 0, the remaining decision may no longer be optimal; see Exercise 13.2. A feedback or Markovian Stackelberg equilibrium, on the other hand, consists of decisions expressed as functions of the current state and time. Such a solution is time consistent. In this section, we will not develop the general theory, for which we refer the reader to Basar and Olsder (1999), Dockner et al. (2000), and Bensoussan et al. (2014, 2015a, 2018). Instead, we will formulate a Stackelberg diﬀerential game of cooperative advertising between a manufacturer as the leader and a retailer as the follower, and obtain a feedback Stackelberg solution. This formulation is due to He et al. (2009). A veriﬁcation theorem that applies to this problem can be found in Bensoussan et al. (2018).

13.4. A Stackelberg Diﬀerential Game of Cooperative Advertising

397

The manufacturer sells a product to end users through the retailer. The product is in a mature category where sales, expressed as a fraction of the potential market, is inﬂuenced through advertising expenditures. The manufacturer as the leader decides on an advertising support scheme via a subsidy rate, i.e., he will contribute a certain percentage of the advertising expenditure by the retailer. Speciﬁcally, the manufacturer decides on a subsidy rate Wt , 0 ≤ Wt ≤ 1, and the retailer as the follower decides on the advertising eﬀort level Ut ≥ 0, t ≥ 0. As in Sect. 12.3, the cost of advertising is quadratic in the advertising eﬀort Ut . Then, with the advertising eﬀort Ut and the subsidy rate Wt , the manufacturer’s and the retailer’s advertising expenditures are Wt Ut2 and (1 − Wt )Ut2 , respectively. The market share dynamics is given by the Sethi model # dXt = (rUt 1 − Xt − δXt )dt + σ(Xt )dZt , X0 = x0 . (13.50) The corresponding expected proﬁts of the retailer and the manufacturer are, respectively, as follows: ∞

−ρt 2 e (πXt − (1 − Wt )Ut )dt , (13.51) JR = E 0

JM = E

∞

e 0

−ρt

5

π M Xt −

Wt Ut2

6

dt .

(13.52)

A solution of this Stackelberg diﬀerential game depends on the available information structure. We shall assume that at each time t, both players know the current system state and the follower knows the action of the leader. The concept of equilibrium that applies in this case is that of feedback Stackelberg equilibrium. For this and other information structures and equilibrium concepts, see Bensoussan et al. (2015a). Next we deﬁne the rules, governing the sequence of actions, by which this game will be played over time. To be speciﬁc, the sequence of plays at any time t ≥ 0 is as follows. First, the manufacturer observes the market share Xt at time t and selects the subsidy rate Wt . Then, the retailer observes this action Wt and, knowing also the market share Xt at time t, sets the advertising eﬀort rate Ut as his response to Wt . The system evolves over time as this game is played in continuous time beginning at time t = 0. One could visualize this game as being played at times 0, δt, 2δt, . . . , and then let δt → 0. Next, we will address the question of how players choose their actions at any given t. Speciﬁcally, we are interested in deriving an equilibrium menu W (x) for the leader representing his decision when the state is x

398

13. Diﬀerential Games

at time t, and a menu U (x, W ) for the follower representing his decision when he observes the leader’s decision to be W in addition to the state x at time t. For this, let us ﬁrst deﬁne a feedback Stackelberg equilibrium, and then develop a procedure to obtain it. We begin with specifying the admissible strategy spaces for the manufacturer and the retailer, respectively: W = {W |W : [0, 1] → [0, 1] and W (x) is Lipschitz continuous in x} U

= {U |U : [0, 1] × [0, 1] → [0, ∞) and U (x, W ) is Lipschitz continuous in (x, W )}.

For a pair of strategies (W, U ) ⊂ W ×U , let Ys , s ≥ t, denote the solution of the state equation # (13.53) dYs = (rU (Ys , Ws ) 1 − Ys − δYs )ds + σ(Ys )dZs , Yt = x. We should note that Ys here stands for Ys (t, x; W, U ), as the solution t,x depends on the speciﬁed arguments. Then JM (W (·), U (·, W (·))) and t,x JR (W (·), U (·, W (·))) representing the current-value proﬁts of the manufacturer and retailer at time t are, respectively,

=E

=E

∞ t

∞ t

t,x JM (W (·), U (·, W (·)))

e−ρ(s−t) [π M Ys − W (Ys ){U (Ys , W (Ys ))}2 ],

(13.54)

JRt,x (W (·), U (·, W (·)))

e−ρ(s−t) [πYs − (1 − W (Ys )){U (Ys , W (Ys ))}2 ], (13.55)

where we should stress that W (·), U (·, W (·)) evaluated at any state ζ are W (ζ), U (ζ, W (ζ)). We can now deﬁne our equilibrium concept. A pair of strategies (W ∗ , U ∗ ) ∈ W ×U is called a feedback Stackelberg equilibrium if t,x JM (W ∗ (·), U ∗ (·, W ∗ (·))) t,x ≥ JM (W (·), U ∗ (·, W (·))), W ∈ W, x ∈ [0, 1], t ≥ 0,

(13.56)

and JRt,x (W ∗ (·), U ∗ (·, W ∗ (·))) ≥ JRt,x (W ∗ (·), U (·, W ∗ (·))), U ∈ U , x ∈ [0, 1], t ≥ 0.

(13.57)

13.4. A Stackelberg Diﬀerential Game of Cooperative Advertising

399

It has been shown in Bensoussan et al. (2014) that this equilibrium is obtained by solving a pair of Hamilton-Jacobi-Bellman equations where a static Stackelberg game is played at the Hamiltonian level at each t, and where √ (13.58) H M (x, W, U, λM ) = π M x − W U 2 + λM (rU 1 − x − δx) √ H R (x, W, U, λR ) = πx − (1 − W )U 2 + λR (rU 1 − x − δx) (13.59) are the Hamiltonians for the manufacturer and the retailer, respectively. To solve this Hamiltonian level game, we ﬁrst maximize H R with respect to U in terms of x and W. The ﬁrst order condition gives √ λR r 1 − x ∗ , (13.60) U (x, W ) = 2(1 − W ) as the optimal response of the follower for any decision W by the leader. We then substitute this for U in H M to obtain W (λR r)2 (1 − x) H M (x, W, U ∗ (x, W ), λM ) = π M x − 4(1 − W )2 R 2 λ r (1−x) M −δx . (13.61) +λ 2(1−W ) The ﬁrst-order condition of maximizing H M with respect to W gives us W (x) =

2λM − λR . 2λM + λR

(13.62)

Clearly W (x) ≥ 1 makes no intuitive sense because it would induce the retailer to spend an inﬁnite amount on advertising, and that would not be optimal for the leader. Moreover, λM and λR , the marginal valuations of the market share of the leader and the follower, respectively, are expected to be positive, and therefore it follows from (13.62) that W (x) < 1. Thus, we set, 2λM − λR ∗ . (13.63) W (x) = max 0, M 2λ + λR We can now write the HJB equations as R ρV R = H R (x, W ∗ (x), U ∗ (x, W ∗ (x)), VxR ) + (σ(x))2 Vxx /2 R 2 2 R (V r) (1 − x) (σ(x)) Vxx − VxR δx + (13.64) = πx + x ∗ 4(1 − W (x)) 2

400

13. Diﬀerential Games ρV M

M = H M (x, W ∗ (x), U ∗ (x, W ∗ (x)), VxM ) + (σ(x))2 Vxx /2 R 2 ∗ M R 2 (V r) (1 − x)W (x) Vx Vx r (1 − x) = πM x − x + 4(1 − W ∗ (x))2 2(1 − W ∗ (x)) M −VxM δx + (σ(x))2 Vxx /2

(13.65)

The solution of these equations will yield the value functions V M (x) and V R (x). With these in hand, we can give the equilibrium menu of actions to the manufacturer and the retailer to guide their decisions at each t. These menus are √ 2VxM − VxR VxR r 1 − x ∗ ∗ W (x) = max 0, and U (x, W ) = . 2VxM + VxR 2(1 − W ) (13.66) To solve for the value function, we next investigate the two cases where the subsidy rate is (a) zero and (b) positive, and determine the condition required for no subsidy to be optimal. Case (a): No Co-op Advertising (W ∗ = 0). Inserting W ∗ (x) = 0 into (13.66) gives √ rVxR 1 − x ∗ U (x, 0) = . (13.67) 2 Inserting W ∗ (x) = 0 into (13.65) and (13.64), we have ρV M = π M x + ρV R = πx +

M VxM VxR r2 (1 − x) (σ(x))2 Vxx − VxM δx + , 2 2

(13.68)

R (VxR )2 r2 (1 − x) (σ(x))2 Vxx − VxR δx + . 4 2

(13.69)

Let V M (x) = αM + β M x and V R (x) = α + βx. Then, VxM = β M and VxR = β. Substituting these into (13.68) and (13.69) and equating like powers of x, we can express all of the unknowns in terms of β, which itself can be explicitly solved. That is, we obtain β =

2π M 2π # , βM = , 2 2 2(ρ + δ) + βr2 (ρ + δ) + r π + (ρ + δ) (13.70)

α =

β 2 r2 4ρ

, αM =

ββ M 2ρ

r2

.

(13.71)

13.4. A Stackelberg Diﬀerential Game of Cooperative Advertising

401

# Using (13.71) in (13.67), we can write U ∗ (x) = ρα(1 − x). Finally, we can derive the required condition from the right-hand side of W ∗ (x) in (13.66), which is 2VxM ≤ VxR , for no co-op advertising (W ∗ = 0) in the equilibrium. This is given by 2β M ≤ β, or 4π M 2(ρ + δ) + √

2πr 2 (ρ+δ)2 +r2 π+(ρ+δ)

2π . ≤# (ρ + δ)2 + r2 π + (ρ + δ)

(13.72)

After a few steps of algebra, this yields the required condition πM π θ := # −# ≤ 0. 2 2 2 (ρ + δ) + r π (ρ + δ) + r2 π + (ρ + δ)

(13.73)

Next, we obtain the solution when θ > 0. Case (b): Co-op Advertising (W ∗ > 0). Then, W ∗ (x) in (13.66) reduces to 2V M − VxR . (13.74) W ∗ (x) = xM 2Vx + VxR Inserting this for W ∗ (x) into (13.65) and (13.64), we have r2 (1 − x)[4(VxM )2 − (VxR )2 ] 16 VxM r2 (1 − x)[2VxM + VxR ] + 4 2V M (σ(x)) xx −VxM δx + , (13.75) 2

R 2 2

R (Vx ) r (1 − x) 2VxM + VxR (σ(x))2 Vxx R . = πx + δx + − V x 4 2VxR 2 (13.76) ρV M

ρV R

= πM x −

Once again, V M (x) = αM + β M x, V R = α + βx, VxM = β M , VxR = β. Substituting these into (13.75) and (13.76) and equating like powers of x, we have β(β + 2β M )r2 α= , (13.77) 8ρ β(β + 2β M )r2 , 8 (β + 2β M )2 r2 , = 16ρ

(ρ + δ)β = π − αM

(13.78) (13.79)

402

13. Diﬀerential Games

(β + 2β M )2 r2 . (13.80) 16 Using (13.66), (13.74), and (13.79), we can write U ∗ (x, W ∗ (x)), with a slight abuse of notation, as √ r(VxR + 2VxM ) 1 − x # ∗ = ραM (1 − x). (13.81) U (x) = 4 (ρ + δ)β M = π M −

The four equations (13.77)–(13.80) determine the solutions for the four unknowns, α, β, αM , and β M . From (13.78) and (13.80), we can obtain 2π M 2 8π 8π 2 β3 + = 0. (13.82) β + 2β− ρ+δ r (ρ + δ)r2 If we denote a1 =

2π M 8π −8π 2 , a2 = 2 , and a3 = , ρ+δ r (ρ + δ)r2

then a1 > 0, a2 > 0, and a3 < 0. From Descarte’s Rule of Signs, there exists a unique, positive real root. The two remaining roots may be both imaginary or both real and negative. Since this is a cubic equation, a complete solution can be obtained. Using Mathematica or following Spiegel et al. (2008), we can write down the three roots as 1 β(1) = S + T − a1 , 3 1 β(2) = − (S + T ) − 2 1 β(3) = − (S + T ) − 2 with S=

√ 3 1 a1 + i(S − T ), 3 2 √ 3 1 a1 − i(S − T ), 3 2

# # √ 3 3 R + Q3 + R2 , T = R − Q3 + R2 , i = −1,

where

3a2 − a21 9a1 a2 − 27a3 − 2a31 , R= . 9 54 Next, we identify the positive root in each of the following three cases: Q=

Case 1 (Q > 0): We have S > 0 > T and Q3 + R2 > 0. There is one positive root and two imaginary roots. The positive root is β = S + T − (1/3)a1 .

13.4. A Stackelberg Diﬀerential Game of Cooperative Advertising

403

Table 13.1: Optimal feedback Stackelberg solution (a) if θ ≤ 0

(b) if θ > 0

No co-op equilibrium

Co-op equilibrium

V R (x) = α + βx

V R (x) = α + βx

proﬁt V M

V M (x) = αM + β M x

V M (x) = αM + β M x

Coeﬃcients of

β=√

2π (ρ+δ)2 +rπ+(ρ+δ)

β=

2π M 2(ρ+δ)+βr 2

βM =

Retailer’s proﬁt V R Manufacturer’s

proﬁt functions, α, β, αM , β M obtained from:

βM =

α= αM =

β 2 r2 4ρ ββ M r 2 2ρ

−

π ρ+δ

−

πM ρ+δ

α=

β(β+2β M )r 2 8(ρ+δ) (β+2β M )2 r 2 16(ρ+δ)

β(β+2β M )r 2 8ρ

αM =

(β+2β M )2 r 2 16ρ

Subsidy rate W ∗ (x) =

2β M −β 2β M +β

0

=1−

α αM

Advertising eﬀort U ∗ (x) =

√ rβ 1−x 2

=

#

ρα(1 − x)

√ r(β+2β M ) 1−x 4

=

#

ραM (1 − x)

Case 2 (Q < 0 and Q3 + R2 > 0): There are three real roots with one positive root, which is β = S + T − (1/3)a1 . Case 3 (Q < 0 and Q3 + R2 < 0): S and T are both imaginary. We have three real roots with one positive root. While subcases can be given to identify the positive root, for our purposes, it is enough to identify it numerically. Finally, we can conclude that 2β M − β > 0 so that W ∗ > 0, since if this were not the case, then W ∗ would be zero, and we would once again be in Case (a). We can now summarize the optimal feedback Stackelberg equilibrium in Table 13.1. In Exercises 13.7–13.10, you are asked to further explore the model of this section when the parameters π = 0.25, π M = 0.5, r =

404

13. Diﬀerential Games

# 2, ρ = 0.05, δ = 1, and σ(x) = 0.25 x(1 − x). For this case, He et al. (2009) obtain the comparative statics as shown in Fig. 13.2.

0.6

0.6

0.55

0.5

0.5

0.4

0.45

0.3

0.4

0.2

0.35

0.1

0.3 0.25

0

0.2

-0.1

0.15 0.25

0.3

0.35

0.4

0.45

0.5

0.55

0.6

-0.2 0.25

0.3

0.35

0.4

0.45

0.5

0.55

0.6

Figure 13.2: Optimal subsidy rate vs. (a) Retailer’s margin and (b) Manufacturer’s margin

There have been many applications of diﬀerential games in marketing in general and optimal advertising in particular. Some references are Bensoussan et al. (1978), Deal et al. (1979), Deal (1979), Jørgensen (1982a), Rao (1984, 1990), Dockner and Jørgensen (1986, 1992), Chintagunta and Vilcassim (1992), Chintagunta and Jain (1994, 1995), Fruchter (1999a), Jarrar et al. (2004), Mart´ın-Herr´an et al. (2005), Breton et al. (2006), Jørgensen and Zaccour (2007), He and Sethi (2008), Naik et al. (2008), Zaccour (2008a), Jørgensen et al. (2009), Prasad and Sethi (2009). The literature on advertising diﬀerential games is surveyed by Jørgensen (1982a) and the literature on management applications of Stackelberg diﬀerential games is reviewed by He et al. (2007). Monographs are written by Erickson (2003) and Jørgensen and Zaccour (2004). For applications of diﬀerential games to economics and management science in general, see the book by Dockner et al. (2000).

Exercises for Chapter 13

E 13.1 A Bilinear Quadratic Advertising Model (Deal et al. 1979). Let xi be the market share of ﬁrm i and ui be its advertising rate, i = 1, 2.

Exercises for Chapter 13

405

The state equations are x˙ 1 = b1 u1 (1 − x1 − x2 ) + e1 (u1 − u2 )(x1 + x2 ) − a1 x1 x1 (0) = x10 , x˙ 2 = b2 u2 (1 − x1 − x2 ) + e2 (u2 − u1 )(x1 + x2 ) − a2 x2 x2 (0) = x20 , where bi , ei , and ai are given positive constants. Firm i wants to maximize T −ρT Ji = wi e xi (T ) + (ci xi − u2i )e−ρt dt, 0

where wi , ci , and ρ are positive constants. Derive the necessary conditions for the open-loop Nash solution, and formulate the resulting boundary value problem. In a related paper, Deal (1979) provides a numerical solution to this problem with e1 = e2 = 0. E 13.2 Let x(t) denote the stock of pollution at time t ∈ [0, T ] that aﬀects the welfare of two countries, one of which is the leader and the other the follower. The state dynamics is x˙ = u + v, x(0) = x0 , where u and v are emission rates of the leader and the follower, respectively. Let their instantaneous utility functions be u − (u2 + x2 )/2 and v − (v 2 + x2 )/2, respectively. Obtain the open-loop Stackelberg solution. By re-solving this problem at time τ , 0 < τ < T, show that the ﬁrst solution obtained is time inconsistent. Hint: Apply ﬁrst the maximum principle to the follower’s problem for any given leader’s decision u. Let λF denote the adjoint variable associated with the state x; Clearly λF (T ) = 0. Then apply the maximum principle to the leader’s problem, treating the follower’s adjoint equation as a “state” equation in addition to the state equation for x. Let the adjoint variables associated with x and λF be λL and μ, respectively. Clearly λL (T ) = 0. However, the transversality condition for μ will be μ(0) = 0 in view of Remark 3.9. See Basar and Olsder (1999) and Dockner et al. (2000) for further details.

406

13. Diﬀerential Games

E 13.3 Develop the nonlinear model for licensing of ﬁsherman described toward the end of Sect. 13.2.3 by rewriting (13.19) and (13.22) for the model. Derive the adjoint equation for λi for the ith producer, and show that the feedback Nash policy for producer i is given by f (v i∗ ) =

ci . (pi − λi )x

E 13.4 Consider an N-ﬁrm oligopoly. Let Si (t) denote the cumulative sales by time t of ﬁrm i ∈ {1, 2, ..., N } and deﬁne S(t) = N i=1 Si (t). Let Ai (t) denote ﬁrm i’s advertising rate. With positive constants a, b, and d, assume that the diﬀerential game has the diﬀusion dynamics S˙ i (t) = [a + b log Ai (t) + dS(t)][M − S(t)], Si (0) = Si0 ≥ 0, which means that a ﬁrm can stimulate its sales through advertising (but subject to decreasing returns) and that demand learning eﬀects (imitation) are industry-wide. (If these eﬀects were ﬁrm-speciﬁc we would have Si instead of S in the brackets on the right-hand side of the dynamics.) Payoﬀs are given by

T

Ji = 0

[(pi − ci )S˙ i (t) − Ai (t)]dt,

in which prices and unit costs are constant. Since S˙ i (t) in the expression for J i is stated in terms of the state variable S(t) and the control variables Ai (t), i ∈ {1, 2, . . . , N }, formulate the diﬀerential game problem with S(t) as the state variable. In the open-loop Nash equilibrium, show that the advertising rates are monotonically decreasing over time. Hint: Assume ∂ 2 H i /∂S 2 ≤ 0 so that H i is concave in S. Use this condition to prove the monotone property. E 13.5 Solve (13.43) to obtain the solution for α and β given in (13.44) and (13.45). E 13.6 Use Mathematica or another suitable software program to solve the quartic equation (13.46). Show that for ρ1 = ρ2 = 0.05, π 1 = π 2 = 1, δ = 0.01, R1 = 1, R2 = 4, the only positive solution for β 1 is 0.264545. Figure 13.1 gives a sample path of the optimal market shares of the two ﬁrms for this problem. Draw another sample path.

Exercises for Chapter 13

407

E 13.7 In the Stackelberg diﬀerential game of Sect. 13.4 let π = 0.25, π M = 0.5, r = 2, ρ = 0.05, and δ = 1. Obtain the coeﬃcients α, β, αM , β M , and show that W ∗ = 0.58. Graph the value functions V M (x) = αM + β M x, V (x) = α + βx, and their sum V M (x) + V (x), as the functions of the market share x. E 13.8 Suppose the manufacturer in Exercise 13.7 does not behave optimally and decides instead to oﬀer no cooperative advertising. Obtain the value functions of the manufacturer and the retailer. Compare the manufacturer’s value function in this case with VM (x) in Exercise 13.7. Furthermore, when x0 = 0.5, obtain the manufacturer’s loss in expected proﬁt when compared to the optimal expected proﬁt VM (x0 ) in Exercise 13.7. E 13.9 Suppose that the manufacturer and the retailer in the problem of Sect. 13.4 are integrated into a single ﬁrm. Then, formulate the stochastic optimal control problem of the integrated ﬁrm. Also, using the data in Exercise 13.7, obtain the value function V I (x) = αI + β I x of the integrated ﬁrm, and compare it to V M (x) + V (x) obtained in Exercise 13.7. # E 13.10 Let σ(x) = 0.25 x(1 − x) and the initial market share x0 = 0.1. Use the optimal feedback advertising eﬀort U ∗ (x) in (13.50) to determine the optimal market share Xt∗ over time. You may use MATLAB or another suitable software to graph a sample path of Xt∗ , t ≥ 0.

Appendix A

Solutions of Linear Diﬀerential Equations A.1

First-Order Linear Equations

Consider the equation x˙ + ax = b(t), x(0) = x0 ,

(A.1)

where a is a constant real number and b(t) is a given function of t. If we multiply both sides of this equation by the integrating factor eat , we get xe ˙ at + axeat = b(t)eat , which can be written at any time τ as d(x(τ )eaτ ) = b(τ )eaτ dτ . Integrating from 0 to t and then multiplying throughout by e−at , we get the solution of (A.1) as −at

x(t) = e

x0 +

t

e−a(t−τ ) b(τ )dτ .

(A.2)

0

If we generalize (A.1) by replacing the constant a by a function a(t), we get (A.3) x(t) ˙ + a(t)x(t) = b(t), x(0) = x0 . © Springer Nature Switzerland AG 2019 S. P. Sethi, Optimal Control Theory, https://doi.org/10.1007/978-3-319-98237-3

409

410

A. Solutions of Linear Diﬀerential Equations t

We can then use the integrating factor e 0 a(s)ds , and with that you are asked to show in Exercise A.1 by employing a procedure similar to that for the solution of (A.3) that t t − 0t a(s)ds x(t) = x0 e + b(τ )e− τ a(s)ds dτ . (A.4) 0

A.2

Second-Order Linear Equations with Constant Coeﬃcients

Consider the equation x ¨ + a1 x˙ + ax = b(t),

(A.5)

where a and a1 are constants and b(t) is a function of t. This equation requires two boundary conditions to be completely speciﬁed. These, for example, could be the values of x(t) at two points in time or the values of x(0) and x(0). ˙ A general solution of (A.5) has the form x(t) = xn (t) + xp (t),

(A.6)

where xn (t) is a homogeneous solution, deﬁned to be a solution of (A.5) with b(t) set at 0, and xp (t) is the particular solution. Clearly x ¨n + ¨p + a1 x˙ p + axp = b(t). a1 x˙ n + axn = 0 and x To obtain a homogeneous solution, let m1 and m2 be the roots of the auxiliary equation m2 + a1 m + a = 0. Then there are 3 cases shown in Table A.1. Next we provide the particular solution of Eq. (A.5). Since this solution depends on the function b(t), we will provide this in Table A.2. It is easy to extend Row 3 and Row 5 of Table A.2 for a polynomial P (t) of degree n. See Zwillinger (2003) for details. For solutions of higher order linear diﬀerential equations with constant coeﬃcients and many other diﬀerential equations, the reader is referred to Zwillinger (2003) and Polyanin and Zaitsev (2003).

A.3

System of First-Order Linear Equations

In vector form, a system of ﬁrst-order linear equations reads x˙ + Ax = b(t), x(0) = x0 ,

(A.7)

A.3. System of First-Order Linear Equations

411

Table A.1: Homogeneous solution forms for Eq. (A.5) Root

General solution form

m1 = m2 , real

x(t) = C1 em1 t + C2 em2 t

m1 = m2 = m, real

x(t) = (C1 + C2 t)emt

m1 = p + qi, m2 = p − qi

x(t) = ept (C1 sin qt + C2 cos qt)

Table A.2: Particular solutions for Eq. (A.5) b(t)

The particular solution of (A.5)

(1)

ert

ert /(r2 + a1 r + a)

(2)

sin θt

(a−θ2 ) sin θt−a1 cos θt (a−θ2 )2 +(a1 θ)2

(3)

P (t) = α + βt + γt2

1 a [P (t)

ert sin θ

Multiply row 2 by ert

(4)

−

a1 a P (t)

+

a21 −a P (t)] a2

Replace a1 by a1 + 2r Replace a by a + a1 r + r2 P (t)ert

(5)

Multiply row 3 by ert Replace a1 by a1 + 2r Replace a by a + a1 r + r2

where x is an n-column vector, A is an n × n matrix of constants, and b is a function of t. We will present two ways of solving the ﬁrst-order system (A.7). The ﬁrst method involves the matrix exponential function etA deﬁned

412

A. Solutions of Linear Diﬀerential Equations

by the power series ∞

tA

e

(tA)k t2 A2 + ··· = . = I + tA + 2! k!

(A.8)

0

It can be shown that this series converges (component by component) for all values of t. Also it is diﬀerentiable (component by component) for all values of t and satisﬁes d tA (e ) = AetA = (etA )A. dt

(A.9)

By analogy with (A.2), we can write the solution of (A.7) as −tA

x(t) = e

x0 +

t

e−(t−τ )A b(τ )dτ .

(A.10)

0

Although (A.10) represents a formal expression for the solution of (A.7), it does not provide a computationally convenient way of getting explicit solutions. For the second method we assume that the matrix A is diagonalizable, i.e., that there exists a nonsingular square matrix P such that P −1 AP = Λ.

(A.11)

Here Λ is the diagonal matrix ⎡ ⎢ ⎢ ⎢ ⎢ ⎢ Λ=⎢ ⎢ ⎢ ⎢ ⎣

⎤ ···

0

0

λ2 · · ·

0

.. .

.. .

···

.. .

0

0

···

λn

λ1 0

⎥ ⎥ ⎥ ⎥ ⎥ ⎥, ⎥ ⎥ ⎥ ⎦

(A.12)

where the diagonal elements, λ1 , . . . , λn , are eigenvalues of A. The ith column of P is the column eigenvector associated with the eigenvalue λi (to see this multiply both sides of (A.11) by P on the left). By looking at (A.8) it is easy to see that P −1 etA P = etΛ and P etΛ P −1 = etA ,

(A.13)

A.4. Solution of Linear Two-Point Boundary Value Problems ⎡

where

etΛ

⎢ ⎢ ⎢ ⎢ ⎢ =⎢ ⎢ ⎢ ⎢ ⎣

413

⎤ 0

···

0

0

etλ2

···

0

.. .

.. .

···

.. .

0

0

···

etλn

e

tλ1

⎥ ⎥ ⎥ ⎥ ⎥ ⎥. ⎥ ⎥ ⎥ ⎦

(A.14)

Using (A.13) into (A.10), we can write the solution to (A.7) as −tΛ

x(t) = (P e

P

−1

)x0 +

t

P e−(t−τ )Λ P −1 b(τ )dτ .

(A.15)

0

Since well-known algorithms are available for ﬁnding eigenvalues and eigenvectors of a matrix, the solution (A.15) can be computed in a straightforward manner.

A.4

Solution of Linear Two-Point Boundary Value Problems

In linear-quadratic control problems with linear salvage values (e.g., the production-inventory problem in Sect. 6.1) we require the solution of linear two-point boundary value problems of the form ⎤⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎢ x˙ ⎥ ⎢ A11 A12 ⎥ ⎢ x ⎥ ⎢ b1 ⎥ (A.16) ⎦⎣ ⎦ + ⎣ ⎦ ⎣ ⎦=⎣ ˙λ λ A21 A22 b2 with boundary conditions x(0) = x0

and

λ(T ) = λT .

The solution of this system will be of the form (A.15), restated as ⎤ ⎡ ⎤ ⎡ ⎡ ⎤⎡ ⎢ x(t) ⎥ ⎢ Q11 (t) Q12 (t) ⎥ ⎢ x(0) ⎥ ⎢ R1 (t) ⎦+⎣ ⎦=⎣ ⎣ ⎦⎣ λ(0) λ(t) Q21 (t) Q22 (t) R2 (t)

(A.17) which can be ⎤ ⎥ ⎦,

(A.18)

414

A. Solutions of Linear Diﬀerential Equations

where the λ(0) is a vector of unknowns. They can be determined by setting (A.19) λT = Q21 (T )x(0) + Q22 (T )λ(0) + R2 (T ), which is a system of linear equations for the variables λ(0).

A.5

Solutions of Finite Diﬀerence Equations

In this book we will have uses for ﬁnite diﬀerence equations only in Chaps. 8 and 9. For that reason we will give only a brief introduction to solution techniques for them. Readers who wish more details can consult one of several texts on diﬀerence equations; see, e.g., Goldberg (1986) or Spiegel (1971). If f (k) is a real function of time, then the diﬀerence operator applied to f is deﬁned as Δf (k) = f (k + 1) − f (k). (A.20) The factorial power of k is deﬁned as k(n) = k(k − 1)(k − 2) . . . (k − (n − 1)). It is easy to show that

Δk(n) = nk (n−1) .

(A.21)

(A.22)

Because this formula is similar to the corresponding formula for the derivative d(k n )/dk, the factorial powers of k play an analogous role for ﬁnite diﬀerences that the ordinary powers of k play for diﬀerential calculus. If f (k) is a real function of time, then the anti-diﬀerence operator −1 Δ applied to f is deﬁned as another function g = Δ−1 f (k) with the property that Δg = f (k). (A.23) One can easily show that Δ−1 k (n) = (1/(n + 1))k (n+1) + c,

(A.24)

where c is an arbitrary constant. Equation (A.24) corresponds to the integration formula for powers of k in calculus. Note that formulas (A.22) and (A.24) are similar to, respectively, diﬀerentiation and integration of the power function k n in calculus. By

A.5. Solutions of Finite Diﬀerence Equations

415

analogy with calculus, therefore, we can solve diﬀerence equations involving polynomials in ordinary powers of k by ﬁrst rewriting them as polynomials involving factorial powers of k so that (A.22) and (A.24) can be used. We show next how to do this.

A.5.1

Changing Polynomials in Powers of k into Factorial Powers of k

We ﬁrst give an abbreviated list of formulas that show how to change powers of k into factorial powers of k: k 0 = k (0) = 1 k

1

k

2

= k = k

(1) (1)

(by deﬁnition),

, + k (2) ,

k 3 = k (1) + 3k (2) + k (3) , k 4 = k (1) + 7k (2) + 6k (3) + k (4) , k 5 = k (1) + 15k (2) + 25k (3) + 10k (4) + k (5) . The coeﬃcients of the factorial powers on the right-hand sides of these equations are called Stirling numbers of the second kind, after the person who ﬁrst derived them. This list can be extended by using a more complete table of these numbers, which can be found in books on diﬀerence equations cited earlier. Example A.1 Express k 4 − 3k + 4 in terms of factorial powers. Solution Using the equations above we have k 4 = k (1) + 7k (2) + 6k (3) + k (4) , −3k = −3k (1) , 4 = 4, so that

k 4 − 3k + 4 = k (4) + 6k (3) + 7k (2) − 2k (1) + 4.

Example A.2 Solve the following diﬀerence equation in Example 8.7: Δλk = −k + 5, λ6 = 0. Solution We ﬁrst change the right-hand side into factorial powers so that it becomes Δλk = −k (1) + 5.

416

A. Solutions of Linear Diﬀerential Equations

Applying (A.24), we obtain λk = −(1/2)k (2) + 5k (1) + c, where c is a constant. Applying the condition λ6 = 0, we ﬁnd that c = −15, so that the solution is λk = −(1/2)k (2) + 5k (1) − 15.

(A.25)

However, we would like the answer to be in ordinary powers of k. The way to do that is discussed in the next section.

A.5.2

Changing Factorial Powers of k into Ordinary Powers of k

In order to change factorial powers of k into ordinary powers of k, we make use of the following formulas: k (1) = k, k (2) = −k + k 2 , k (3) = 2k − 3k 2 + k 3 , k (4) = −6k + 11k 2 − 6k 3 + k 4 , k (5) = 24k − 50k 2 + 35k 3 − 10k 4 + k 5 . The coeﬃcients of the factorial powers on the right-hand sides of these equations are called Stirling numbers of the ﬁrst kind. This list can also be extended by using a more complete table of these numbers, which can be found in books on diﬀerence equations. Solution of Example A.2 Continued By substituting the ﬁrst two of the above formulas into (A.25), we see that the desired answer is λk = −(1/2)k 2 + (11/2)k − 15, which is the solution needed for Example 8.7.

(A.26)

Exercises for Appendix A

417

Exercises for Appendix A E A.1 Show that the solution of Eq. (A.3) is given by (A.4). ⎡

⎤

⎡

⎤

⎢ 3 2 ⎥ ⎢ 5 0 ⎥ E A.2 If A = ⎣ ⎦ , show that Λ = ⎣ ⎦ and P = 2 3 0 2 ⎡ ⎤ ⎢ 1 1 ⎥ ⎣ ⎦. 1 −1 ⎡ ⎤ ⎢ 1 ⎥ Use (A.15) to solve (A.7) for this data, given that z(0) = ⎣ ⎦ . 2 ⎡

⎤

⎡

⎤

⎢ 3 3 ⎥ ⎢ 6 0 ⎥ E A.3 If A = ⎣ ⎦ , show that Λ = ⎣ ⎦ and P = 2 4 0 1 ⎡ ⎤ 1 3 ⎢ ⎥ ⎣ ⎦. 1 −2 ⎡ ⎤ ⎢ 0 ⎥ Use (A.15) to solve (A.7) for this data, given that z(0) = ⎣ ⎦ . 5 E A.4 After you have read Sect. 6.1, re-solve the production-inventory example stated in Eqs. (6.1) and (6.2), (ignoring the control constraint (P ≥ 0) by the method of Sect. A.4. The linear two-point boundary value problem is stated in Eqs. (6.6) and (6.7).

Appendix B

Calculus of Variations and Optimal Control Theory Here we introduce the subject of the calculus of variations by analogy with the classical topic of maximization and minimization in calculus; see Gelfand and Fomin (1963), Young (1969), and Leitmann (1981) for rigorous treatments of the subject. The problem of the calculus of variations is that of determining a function that maximizes a given functional, the objective function. An analogous problem in calculus is that of determining a point at which a speciﬁc function, the objective function, is maximum. This, of course, is done by taking the ﬁrst derivative of the function and equating it to zero. This is what is called the ﬁrst-order condition for a maximum. A similar procedure will be employed to derive the ﬁrst-order condition for the variational problem. The analogy with classical optimization extends also to the second-order maximization condition of calculus. Finally, we will show the relationship between the maximum principle of optimal control theory and the necessary conditions of the calculus of variations. It is noted that this relationship is similar to the one between the Kuhn-Tucker conditions in mathematical programming and the ﬁrst-order conditions in classical optimization. We start with the “simplest” variational problem in the next section.

© Springer Nature Switzerland AG 2019 S. P. Sethi, Optimal Control Theory, https://doi.org/10.1007/978-3-319-98237-3

419

420

B. Calculus of Variations and Optimal Control Theory

B.1

The Simplest Variational Problem

Assume a function x : C 1 [0, T ] → E 1 , where C 1 [0, T ] is a class of functions deﬁned over the interval [0, T ] with continuous ﬁrst derivatives. For simplicity in exposition, we are assuming x(t) to be a scalar function of t ∈ [0, T ], and the extension to a vector function is straightforward. Here t simply denotes the independent variable which need not be time. Assume further that a function in this class is termed admissible if it satisﬁes the terminal conditions x(0) = x0

and

x(T ) = xT .

(B.1)

We are thus dealing with a ﬁxed-end-point problem. Examples of admissible functions for the problem are shown in Fig. B.1; see Chapters 2 and 3 of Gelfand and Fomin (1963) for problems other than the simplest problem, i.e., the problems with other kinds of conditions for the end points.

Figure B.1: Examples of admissible functions for the problem The problem under consideration is to obtain the admissible function x∗ for which the functional T J(x) = F (x, x, ˙ t)dt (B.2) 0

has a relative maximum. We will assume that all ﬁrst and second partial derivatives of the function F : E 1 × E 1 × E 1 → E 1 are continuous.

B.2. The Euler-Lagrange Equation

B.2

421

The Euler-Lagrange Equation

The necessary ﬁrst-order conditions in classical optimization were obtained by considering small changes about the solution point. For the variational problem, we consider small variations about the solution function. Let x(t) be the solution and let y(t) = x(t) + εη(t), where η(t) : C 1 [0, T ] → E 1 is an arbitrary continuously diﬀerentiable function satisfying η(0) = η(T ) = 0, (B.3) and ε ≥ 0 is a small number. A sketch of these functions is shown in Fig. B.2.

Figure B.2: Variation about the solution function The value of the objective functional associated with y(t) can be considered a function of ε, i.e., V (ε) = J(y) =

T

F (x + εη, x˙ + εη, ˙ , t)dt. 0

However, x(t) is a solution and therefore V (ε) must have a maximum at ε = 0. This means / dV / / δJ = = 0, dε /ε=0

422

B. Calculus of Variations and Optimal Control Theory

where δJ is known as the variation δJ in J. Diﬀerentiating V (ε) with respect to ε and setting ε = 0 yields / T dV // = (Fx η + Fx˙ η)dt ˙ = 0, δJ = dε /ε=0 0 which after integrating the second term by parts provides / T T d dV // T (Fx˙ )ηdt = 0. = F ηdt + (F η)| − δJ = x x˙ 0 / dε ε=0 0 0 dt

(B.4)

Because of the end conditions on η, the expression simpliﬁes to / T d dV // = [Fx − Fx˙ ]ηdt = 0. δJ = / dε ε=0 dt 0 We now use the fundamental lemma of the calculus of variations T which states that if h is a continuous function and 0 h(t)η(t)dt = 0 for every continuous function η(t), then h(t) = 0 for all t ∈ [0, T ]. The reason that this lemma holds, without going into details of a rigorous proof which is available in Gelfand and Fomin (1963), is as follows. Suppose that h(t) = 0 for some t ∈ [0, T ]. Since h(t) is continuous, there is, therefore, an interval (t1 , t2 ) ⊂ [0, T ] over which h is nonzero and has the same sign. Now selecting η(t) such

η(t) is

⎧ ⎪ ⎨ > 0, t ∈ (t1 , t2 ) ⎪ ⎩ 0,

otherwise,

T it is possible to make the integral 0 h(t)η(t)dt = 0. Thus, by contradiction, h(t) must be identically zero over the entire interval [0, T ]. By using the fundamental lemma, we have the necessary condition Fx −

d Fx˙ = 0 dt

(B.5)

known as the Euler-Lagrange equation, or simply the Euler equation, which must be satisﬁed by a maximal solution x∗ . In other words, the solution x∗ (t) must satisfy Fx (x∗ , x˙ ∗ , t) −

d Fx˙ (x∗ , x˙ ∗ , t) = 0. dt

(B.6)

B.2. The Euler-Lagrange Equation

423

We note that the Euler equation is a second-order ordinary diﬀerential equation. This can be seen by taking the total time derivative of Fx˙ in (B.5) to obtain ˙ − (Fx˙ x˙ x ¨) = 0. Fx − Fxt ˙ − (Fxx ˙ x)

(B.7)

The boundary conditions for this equation are obviously the end-point conditions x(0) = x0 and x(T ) = xT . Special Case (i): When F does not depend explicitly on x. ˙ In this case, the Euler equation (B.5) reduces to Fx = 0, which is nothing but the ﬁrst-order condition of classical optimization. In this case, the dynamic problem is a succession of static classical optimization problems. Special Case (ii): When F does not depend explicitly on x. The Euler equation reduces to d Fx˙ = 0, dt

(B.8)

Fx˙ = C,

(B.9)

which we can integrate as where C is a constant. Special Case (iii): When F does not depend explicitly on t. In this important special case, the Euler equation (B.7) reduces to ˙ − (Fx˙ x˙ x ¨) = 0. Fx − (Fxx ˙ x)

(B.10)

On multiplying the left hand side of (B.10) by x˙ on the right, and adding and subtracting the term Fx˙ x ¨, transforms (B.10) to d (F − Fx˙ x) ˙ = 0. dt We can solve the above equation as F − Fx˙ x˙ = C, where C is a constant.

(B.11)

(B.12)

424

B.3

B. Calculus of Variations and Optimal Control Theory

The Shortest Distance Between Two Points on the Plane

The problem is to show that the straight line passing through two points on a plane is the shortest distance between the two points. The problem can be stated as follows: ⎧ T# ⎪ ⎪ min 1 + x˙ 2 dt ⎪ ⎪ ⎪ 0 ⎨ subject to ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ x(0) = x and x(T ) = x . 0 T √ Here t refers to distance rather than time. Since F = − 1 + x˙ 2 does not depend explicitly on x, we are in the second special case and the ﬁrst integral (B.9) of the Euler equation is ˙ + x˙ 2 )− 2 = C. Fx˙ = −x(1 1

This implies that x˙ is a constant, which results in the solution x∗ (t) = C1 t + C2 , where C1 and C2 are constants. These can be evaluated by imposing boundary conditions which give C1 = (xT − x0 )/T and C2 = x0 . Thus,

xT − x 0 ∗ x (t) = t + x0 , T which is the straight line passing through x0 and xT .

B.4

The Brachistochrone Problem

The problem arises from the search for the shape of a wire along which a bead will slide, without friction, in the least time from a given point A to another point B, under the inﬂuence of gravity; see Fig. 1.1. Let t denote the horizontal axis, x denote the vertical axis (measured vertically down), and let the (t, x) coordinates of A and B be (0, 0) and (T, b), respectively. Thus, x(0) = 0 and x(T ) = b. It is reasonable to assume b ≥ 0, so that point B is not higher than point A.

B.4. The Brachistochrone Problem

425

The time τ AB required for the bead to slide from point A to point B along a wire formed in the shape of a curve x(t) is given as sT ds , τ AB = v 0 where v represents velocity and sT is the ﬁnal displacement measured along the curve. Since ds2 = dx2 + dt2 , we can write # ds = 1 + x˙ 2 dt, where x˙ = dx/dt (note that t does not denote time here). From elementary physics, it is known that if v(t = 0) = 0 and a denotes the acceleration due to gravity, then # v(t) = 2ax(t), t ∈ [0, T ]. Then,

τ AB =

T

)

0

1 + x˙ 2 dt. 2ax

(B.13)

The purpose of the Brachistochrone problem is to ﬁnd x(t), t ∈ [0, T ], so as to minimize the time τ AB . This is a variational problem, which in view of a being a constant, can be stated as follows: T) T 1 + x˙ 2 F (x, x, ˙ t)dt = dt . (B.14) min J(x) = x 0 0 As we can see, the integral F in the above problem does not depend explicitly on t, and the problem (B.14) belongs to the third special case. Using the ﬁrst integral (B.11) of the Euler equation for this case, we have 7 ) 1 + x˙ 2 1 1 2 − x˙ = (a constant). x x(1 + x˙ 2 ) k We can reduce this to dx = dt which we rewrite as

)

dx

k2 −x x

k2 − x , x = dt.

(B.15)

426

B. Calculus of Variations and Optimal Control Theory

By performing a change of variable according to 1 2 2 2 1 − cos 2θ x = k sin θ = k 2 2

(B.16)

and recognizing that x(t = 0) = 0 corresponds to θ = 0, we can integrate (B.15) to obtain t θ 1 2k 2 sin2 θdθ = k 2 (θ − sin 2θ) = dt = t. (B.17) 2 0 0 By setting 2θ = φ in (B.16) and (B.17), we can write the solution parametrically as ⎫ ⎪ 2 t = k (φ − sin φ)/2 ⎬ , (B.18) ⎪ x = k 2 (1 − cos φ)/2 ⎭ which are known to be equations representing a cycloid, as depicted in Fig. 1.1 in Chap. 1. Furthermore, since the initial condition x(0) = 0 is already incorporated in performing the integration in (B.17), we must use the terminal condition x(T ) = b for determining the constant k. Clearly, if we let φ1 be deﬁned by the relation b 1 − cos φ1 = , T φ1 − sin φ1

(B.19)

then we can write k2 =

2T 2b = . (1 − cos φ1 ) (φ1 − sin φ1 )

(B.20)

The value of φ1 can be easily obtained numerically for any given values of b > 0 and T > 0. With these, the optimal solution x∗ (t) is the cycloid given parametrically as ( ⎫ ' ⎪ φ−sin φ ⎬ t = T φ −sin φ 1 1 (B.21) ' ( ⎪. 1−cos φ ⎭ ∗ x = b 1−cos φ 1

Furthermore, the minimum time τ ∗AB can be obtained as T7 √ 1 + (x˙ ∗ (t))2 dt. τ ∗AB = 2aJ(x∗ ) = x∗ (t) 0

(B.22)

In Exercise B.1, you are asked to obtain φ1 for T = b = 1 m, and then obtain the minimum time τ ∗AB .

B.5. The Weierstrass-Erdmann Corner Conditions

B.5

427

The Weierstrass-Erdmann Corner Conditions

So far we have only considered functionals deﬁned for smooth curves. This is, however, a restricted class of curves which qualify as solutions, since it is easy to give examples of variational problems which have no solution in this class. Consider, for example, the objective functional 1 min J(x) = x2 (1 − x) ˙ 2 dt , x(−1) = 0, x(1) = 1. −1

The greatest lower bound for J(x) for smooth x = x(t) satisfying the boundary conditions is obviously zero. Yet there is no x ∈ C 1 [−1, 1] with x(−1) = 0 and x(1) = 1, which achieves this value of J(x). In fact, the minimum is achieved for the curve ⎧ ⎪ ⎨ 0, −1 ≤ t ≤ 0, ∗ x (t) = ⎪ ⎩ t, 0 < t ≤ 1, which has a corner (i.e., a discontinuous ﬁrst derivative) at t = 0. Such a piecewise smooth extremal with corners is called a broken extremal. We now enlarge the class of admissible functions by relaxing the requirement that they be smooth everywhere. The larger class is the class of piecewise continuous functions which are continuously diﬀerentiable almost everywhere in [0, T ], i.e., except at some points in [0, T ]. Let x, deﬁned on the interval [0, T ], have a corner at τ ∈ [0, T ]. Let us decompose J(x) as T τ T J(x) = F (x, x, ˙ t)dt = F (x, x, ˙ t)dt + F (x, x, ˙ t)dt 0

0

τ

= J1 (x) + J2 (x). It is clear that on each of the intervals [0, τ ) and (τ , T ], the Euler equation must hold. To compute variations δJ1 and δJ2 , we must recognize that the two ‘pieces’ of x are not ﬁxed-end-point problems. We must require that the two pieces of x join continuously at t = τ ; the point t = τ can, however, move freely as shown in Fig. B.3. This will require a slightly modiﬁed version of formula (B.4) for writing out the variations; see pp. 55–56 in Gelfand and Fomin (1963).

428

B. Calculus of Variations and Optimal Control Theory

Figure B.3: A broken extremal with corner at τ Equating the sum of variations δJ = δJ1 + δJ2 = 0 for x∗ to be an extremal and using the fact that it must be continuous at t = τ implies Fx˙ |τ − = Fx˙ |τ + ,

(B.23)

[F − Fx˙ x] ˙ τ − = [F − Fx˙ x] ˙ τ+.

(B.24)

These conditions are called Weierstrass-Erdmann corner conditions, which must hold at the point τ where the extremal has a corner. In each of the interval [0, τ ) and (τ , t], the extremal x must satisfy the Euler equation (B.5). Solving these two equations will provide us with four constants of integration since the Euler equations are second-order diﬀerential equations. These constants can be found from the end-point conditions (B.1) and Weierstrass-Erdmann conditions (B.23) and (B.24).

B.6

Legendre’s Conditions: The Second Variation

The Euler equation is a necessary conditions analogous to the ﬁrst-order condition for a maximum (or minimum) in the classical optimization

B.7. Necessary Condition for a Strong Maximum

429

problems of calculus. The condition analogous to the second-order necessary condition for a maximum x∗ is the Legendre condition Fx˙ x˙ ≤ 0.

(B.25)

To obtain this condition, we use the second-order condition of classical optimization on function V (ε) to be a maximum at ε = 0, i.e., / T d2 V (ε) // = (Fxx η 2 + 2Fxx˙ η η˙ + Fx˙ x˙ η˙ 2 )dt ≤ 0. (B.26) dε2 /ε=0 0 Integrating the middle term by parts and using (B.3), we can transform (B.26) into a more convenient form T (Qη 2 + P η˙ 2 )dt ≤ 0, (B.27) 0

where Q = Q(t) = Fxx˙ −

d Fxx˙ dt

and

P = P (t) = Fx˙ x˙ .

While it is possible to rigorously obtain (B.25) from (B.27), we will only provide a qualitative argument for this. If we consider the quadratic functional (B.27) for functions η(t) satisfying η(0) = 0, then η(t) will be small in [0, T ] if η(t) ˙ is small in [0, T ]. The converse is not true, however, since it is easy to construct η(t) which is small but has a large derivative η(t) ˙ in [0, T ]. Thus, P η˙ 2 plays the dominant role in (B.27); i.e., P η˙ 2 can be much larger than Qη 2 but it cannot be much smaller (provided P = 0). Therefore, it might be expected that the sign of the functional in (B.8) is determined by the sign of the coeﬃcient P (t), i.e., (B.27) implies (B.25). For a rigorous proof, see Gelfand and Fomin (1963). We note that the strengthened Legendre condition (i.e., with a strict inequality in (B.25)), the Euler equation, and one other condition called strengthened Jacobi condition are suﬃcient for a maximum. The reader can consult Chapter 5 of Gelfand and Fomin (1963) for details.

B.7

Necessary Condition for a Strong Maximum

So far we have discussed necessary conditions for a weak maximum. By weak maximum we mean that the candidate extremals are smooth or

430

B. Calculus of Variations and Optimal Control Theory

piecewise smooth functions. The concept of a strong maximum on the other hand requires that the candidate extremal need only be continuous functions. Without going into details, which are available in Gelfand and Fomin (1963), we state a necessary condition for a strong maximum. This is called the Weierstrass necessary condition. The condition is analogous to the one in the static case that the objective function be concave. It states that if the functional (B.2) has a strong maximum for the extremal x∗ satisfying (B.1), then E(x∗ , x˙ ∗ , t, u) ≤ 0

(B.28)

for every ﬁnite u, where E is the Weierstrass Excess Function deﬁned as ˙ t)(u − x). ˙ E(x, x, ˙ t, u) = F (x, u, t) − F (x, x, ˙ t) − Fx˙ (x, x,

(B.29)

Note that this condition is always met if F (x, x, ˙ t) is concave in x. ˙ The proof of (B.28) is by contradiction. Suppose there exists a τ ∈ [0, T ] and a vector q such that E(τ , x∗ (τ ), x˙ ∗ (τ ), q) > 0. It is then possible to suitably modify x∗ to y, which is close to x∗ in C 1 [0, T ], such that J =

F (y, y, ˙ t)dt −

F (x∗ , x˙ ∗ , t)dt > 0,

contradicting the hypothesis that J(x) has a strong maximum at x∗ .

B.8

Relation to Optimal Control Theory

It is possible to derive the necessary conditions of the calculus of variations from the maximum principle. This is strongly reminiscent of the relationship between the ﬁrst-order conditions of classical optimization and the Kuhn-Tucker conditions of mathematical programming. First, we note that the calculus of variations problem can be stated

B.8. Relation to the Optimal Control Theory as an optimal control problem as follows: ⎧ T ⎪ ⎪ max J = F (x, u, t)dt ⎪ ⎪ ⎪ 0 ⎪ ⎪ ⎪ ⎪ ⎨ subject to ⎪ ⎪ ⎪ x˙ = u, x(0) = x0 , x(T ) = xT , ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ u ∈ Ω = En.

431

(B.30)

The Hamiltonian is H(x, u, λ, t) = F (x, u, t) + λu

(B.31)

with the adjoint variable λ satisfying λ˙ = −Hx = −Fx .

(B.32)

Maximizing the Hamiltonian with respect to u yields Hu = Fx˙ + λ = 0,

(B.33)

λ = −Fx˙ .

(B.34)

from which we obtain Diﬀerentiating (B.34) with respect to time gives d λ˙ = − Fx˙ . dt This equation with (B.32) implies the Euler equation Fx −

d Fx˙ = 0. dt

From (B.30) and (B.32), the second-order condition Huu ≤ 0 for the maximization of the Hamiltonian leads to Fx˙ x˙ ≤ 0, known as the Legendre condition. By the maximum principle, if u∗ is an optimal control with x∗ denoting the corresponding trajectory, then for each t ∈ [0, T ], H(x∗ , u∗ , λ, t) ≥ H(x∗ , u, λ, t),

432

B. Calculus of Variations and Optimal Control Theory

where u is any other control. By the deﬁnition of the Hamiltonian (B.31), x˙ ∗ = u∗ from (B.32), and Eq. (B.33), we have F (x∗ , x˙ ∗ , t) − Fx˙ (x∗ , x˙ ∗ , t)x˙ ∗ ≥ F (x∗ , u, t) − Fx˙ (x∗ , x˙ ∗ , t)u, which by transposition of the terms yields the Weierstrass necessary condition E(x∗ , x˙ ∗ , t, u) = F (x∗ , u, t) − F (x∗ , x˙ ∗ , t) − Fx˙ (x∗ , x˙ ∗ , t)(u − x˙ ∗ ) ≤ 0. We have just proved the equivalence of the maximum principle and the Weierstrass necessary condition in the case where Ω is open. In cases when Ω is closed and when the optimal control is on the boundary of Ω, the Weierstrass necessary condition holds no longer in general. The maximum principle still applies, however. Finally, according to the maximum principle, both λ and H are continuous functions of time. That is, λ(τ − ) = λ(τ + ), H(x∗ (τ ), u∗ (τ − ), λ(τ − ), τ ) = H(x∗ (τ ), u∗ (τ + ), λ(τ + ), τ ). However, λ = −Fx˙

and

H = F − Fx˙ x, ˙

which means that the right-hand sides must be continuous with respect to time, i.e., even across corners. These are precisely the WeierstrassErdmann corner conditions. Exercises for Appendix B E B.1 Solve (B.19) numerically to obtain φ1 for T = b = 1 m. Then, use the formulas (B.21) and (B.22) to compute the minimum time τ ∗AB . Note that the gravitational acceleration rate a = 9.81 m/s2 .

Appendix C

An Alternative Derivation of the Maximum Principle Recall that in the derivation of the maximum principle in Chap. 2, we assumed the twice diﬀerentiability of the value function V (x, t) with respect to the state variable x. Looking at (2.31), we can observe that the smoothness assumptions on the value function do not arise in the statement of the maximum principle. Also since it is not an exogenously given function, there is no a priori reason to assume the twice diﬀerentiability. Moreover, there arise cases in which the value function V (x, t) is not even diﬀerentiable in x. In what follows, we will give an alternate derivation. This proof follows the course pointed out by Pontryagin et al. (1962) but with certain simpliﬁcations. It appears in Fel’dbaum (1965) and, in our opinion, it is one of the simplest proofs for the maximum principle which is not related to dynamic programming and thus permits the elimination of assumptions about the diﬀerentiability of the return function V (t, x). We select the Mayer form of the problem (2.5) for deriving the maximum principle in this section. It will be convenient to reproduce (2.5) here as (C.1): ⎧ ⎪ ⎪ max {J = cx(T )} ⎪ ⎪ ⎪ u(t)∈Ω(t) ⎨ (C.1) subject to ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ x˙ = f (x, u, t), x(0) = x . 0 © Springer Nature Switzerland AG 2019 S. P. Sethi, Optimal Control Theory, https://doi.org/10.1007/978-3-319-98237-3

433

434

C.1

C. An Alternative Derivation of the Maximum Principle

Needle-Shaped Variation

Let u∗ (t) be an optimal control with corresponding state trajectory x∗ (t). We sketch u∗ (t) in Fig. C.1 and x∗ (t) in Fig. C.2 in a scalar case. Note that the kink in x∗ (t) at t = θ corresponds to the discontinuity in u∗ (t) at t = θ.

Figure C.1: Needle-shaped variation

Figure C.2: Trajectories x∗ (t) and x(t) in a one-dimensional case Let τ denote any time in the open interval (0, T ). We select a suﬃciently small ε to insure that τ − ε > 0 and concentrate our attention on this small interval (τ − ε, τ ]. We vary the control on this interval while keeping the control on the remaining intervals [0, τ − ε] and (τ , T ] ﬁxed. Speciﬁcally, the modiﬁed control is

u(t) =

⎧ ⎪ ⎨ v ∈ Ω, t ∈ (τ − ε, τ ], ⎪ ⎩ u∗ (t),

otherwise.

(C.2)

C.1. Needle-Shaped Variation

435

This is called a needle-shaped variation as shown in Fig. C.1. It is a jump function and is diﬀerent from variations in the calculus of variations; see Appendix B. Also the diﬀerence v − u∗ is ﬁnite and need not be small. However, since the variation is on a small time interval, its inﬂuence on the subsequent state trajectory can be proved to be ‘small’. This is done in the following. Let the subsequent motion be denoted by x(t) = x∗ (t) for t > τ − ε. In Fig. C.2, we have sketched x(t) corresponding to u(t). Let δx(t) = x(t) − x∗ (t), t ≥ τ − ε, denote the change in the state variables. Obviously δx(τ − ε) = 0. Clearly, (C.3) δx(τ ) ≈ ε[x(s) ˙ − x˙ ∗ (s)], where s denotes some intermediate time in the interval (τ − ε, τ ]. In particular, we can write (C.3) as δx(τ ) = ε[x(τ ˙ ) − x˙ ∗ (τ )] + o(ε) = ε[f (x(τ ), v, τ ) − f (x∗ (τ ), u∗ (τ ), τ ] + o(ε).

(C.4)

But δx(τ ) is small since f is assumed to be bounded. Furthermore, since f is continuous and the diﬀerence δx(τ ) = x(τ ) − x∗ (τ ) is small, we can rewrite (C.4) as δx(t) ≈ ε[f (x∗ (τ ), v, τ ) − f (x∗ (τ ), u∗ (τ ), τ )].

(C.5)

Since the initial diﬀerence δx(τ ) is small and since u∗ (τ ) does not change from t > τ on, we may conclude that δx(t) will be small for all t > τ . Being small, the law of variation of δx(t) can be found from linear equations for small changes in the state variables. These are called variational equations. From the state equation in (C.1), we have

or,

d(x∗ + δx) = f (x∗ + δx, u∗ , t) dt

(C.6)

dx∗ d(δx) + ≈ f (x∗ , u∗ , t) + fx δx dt dt

(C.7)

or using (C.1), d (δx) ≈ fx (x∗ , u∗ , t)δx, dt

for t ≥ τ ,

(C.8)

436

C. An Alternative Derivation of the Maximum Principle

with the initial condition δx(τ ) given by (C.5). The basic idea in deriving the maximum principle is that equations (C.8) are linear variational equations and result in an extraordinary simpliﬁcation. We next obtain the adjoint equations.

C.2

Derivation of the Adjoint Equation and the Maximum Principle

For this derivation, we employ two methods. The direct method, similar to that of Hartberger (1973), is the consequence of directly integrating (C.8). The indirect method avoids this integration by a trick which is instructive. Direct Method. Integrating (C.8) we get δx(T ) = δx(τ ) +

T

fx [x∗ (t), u∗ (t), t]δx(t)dt,

(C.9)

τ

where the initial condition δx(τ ) is given in (C.5). Since δx(T ) is the change in the terminal state from the optimal state ∗ x (T ), the change in the objective function δJ must be negative. Thus, δJ = cδx(T ) = cδx(τ ) +

T

cfx [x∗ (t), u∗ (t), t]δx(t)dt ≤ 0.

(C.10)

τ

Furthermore, since (C.8) is a linear homogeneous diﬀerential equation, we can write its general solution as δx(t) = Φ(t, τ )δx(τ ),

(C.11)

where the fundamental solution matrix or the transition matrix Φ(t, τ ) ∈ E n×n obeys d Φ(t, τ ) = fx [x∗ (t), u∗ (t)t]Φ(t, τ ), dt

Φ(τ , τ ) = I,

(C.12)

where I is an n × n identity matrix; see Appendix A. Substituting for δx(t) from (C.11) into (C.10), we have δJ = cδx(τ ) + τ

T

cfx [x∗ (t), u∗ (t), t]Φ(t, τ )δx(τ )dt ≤ 0.

(C.13)

C.2. Derivation of Adjoint Equation and the Maximum Principle This induces the deﬁnition T ∗ cfx [x∗ (t), u∗ (t), t]Φ(t, τ )dt + c, λ (t) =

437

(C.14)

τ

which when substituted into (C.13), yields δJ = λ∗ (τ )δx(τ ) ≤ 0.

(C.15)

But δx(τ ) is supplied in (C.5). Noting that ε > 0, we can rewrite (C.15) as (C.16) λ∗ (τ )f [x∗ (τ ), v, τ ] − λ∗ (τ )f [x∗ (τ ), u∗ (τ ), τ ] ≤ 0. Deﬁning the Hamiltonian for the Mayer form as

H[x, u, λ, t] = λf (x, u, t),

(C.17)

we can rewrite (C.16) as H[x∗ (τ ), u∗ (τ ), λ(τ ), τ ] ≥ H[x∗ (τ ), v, λ(τ ), τ ].

(C.18)

Since this can be done for almost every τ , we have the required Hamiltonian maximizing condition. The diﬀerential equation form of the adjoint equation (C.14) can be obtained by taking its derivative with respect to τ . Thus, T dΦ(t, τ ) dλ(τ ) cfx [x∗ (t), u∗ (t), t] = dt dτ dτ τ (C.19) −cfx [x∗ (τ ), u∗ (τ ), τ ]. It is also known that the transition matrix has the property: dΦ(t, τ ) = −Φ(t, τ )fx [x∗ (τ ), u∗ (τ ), τ ], dτ which can be used in (C.19) to obtain T dλ(τ ) =− cfx [x∗ (t), u∗ (t), t]Φ(t, τ )fx [x∗ (τ ), u∗ (τ ), τ ]dt dτ τ −cfx [x∗ (τ ), u∗ (τ ), τ ]. Using the deﬁnition (C.14) of λ(τ ) in (C.20), we have dλ(τ ) = −λ(τ )fx [x∗ (τ ), u∗ (τ ), τ ] dτ

(C.20)

438

C. An Alternative Derivation of the Maximum Principle

with λ(T ) = c, or using (C.17) and noting that τ is arbitrary, we have λ˙ = −λfx [x∗ , u∗ , t] = −Hx [x∗ , u∗ , λ, t),

λ(T ) = c.

(C.21)

This completes the derivation of the maximum principle along with the adjoint equation using the direct method. Indirect Method. The indirect method employs a trick which simpliﬁes considerably the derivation. Instead of integrating (C.8) explicitly, we now assume that the result of this integration yields cδx(T ) as the change in the state at the terminal time. As in (C.10), we have δJ = cδx(T ) ≤ 0. First, we deﬁne

λ(T ) = c,

(C.22)

(C.23)

which makes it possible to write (C.22) as δJ = cδx(T ) = λ(T )δx(T ) ≤ 0.

(C.24)

Note parenthetically that if the objective function J = S(x(T )), we must deﬁne λ(T ) = ∂S[x(T )]/∂x(T ) giving us δJ =

∂S[x(T )] δx(T ) = λ(T )δx(T ). ∂x(T )

Now, λ(T )δx(T ) is the change in the objective function due to a change δx(T ) at the terminal time T. That is, λ(T ) is the marginal return or the marginal change in the objective function per unit change in the state at time T. But δx(T ) cannot be known without integrating (C.8). We do know, however, the value of the change δx(τ ) at time τ which caused the terminal change δx(T ) via (C.8). We would therefore like to pose the problem of obtaining the change δJ in the objective function in terms of the known value δx(τ ); see Fel’dbaum (1965). Simply stated, we would like to obtain the marginal return λ(τ ) per unit change in state at time τ . Thus, λ(τ )δx(τ ) = δJ = λ(T )δx(T ) ≤ 0.

(C.25)

Obviously, knowing λ(τ ) will make it possible to make an inference about δJ, which is directly related to the needle-shaped variation applied in the small interval (τ − ε, τ ].

C.2. Derivation of Adjoint Equation and the Maximum Principle

439

However, since τ is arbitrary, our problem of ﬁnding λ(τ ) can be translated to one of ﬁnding λ(t), t ∈ [0, T ], such that λ(t)δx(t) = λ(T )δx(T ),

t ∈ [0, T ],

(C.26)

or in other words, λ(t)δx(t) = constant,

λ(T ) = c.

(C.27)

It turns out that the diﬀerential equation which λ(t) must satisfy can be easily found. From (C.27), δx ˙ d [λ(t)δx(t)] = λ + λδx = 0, dt dt

(C.28)

which after substituting for dδx/dt from (C.8) becomes ˙ = (λfx + λ)δx ˙ λfx δx + λδx = 0.

(C.29)

Since (C.29) is true for arbitrary δx, we have λ˙ = −λfx = −Hx

(C.30)

using the deﬁnition (C.17) for the Hamiltonian. The Hamiltonian maximizing condition can be obtained by substituting for δx(τ ) from (C.5) into (C.25). This is the same as what we did in (C.15) through (C.18). The purpose of the alternative proof was to demonstrate the validity of the maximum principle for a simple problem without knowledge of any return function. For more complex problems, one needs complicated mathematical analysis to rigorously prove the maximum principle without making use of return functions. A part of mathematical rigor is in proving the existence of an optimal solution without which necessary conditions are meaningless; see Young (1969).

Appendix D

Special Topics in Optimal Control In this appendix we will discuss a number of specialized topics in seven sections. These are the Kalman and Kalman-Bucy ﬁlters, the Weiner Process, Itˆo’s Lemma, linear-quadratic problems, second-order variations, singular control, and the Sethi-Skiba points. These topics are referred to but not discussed in the main body of the text. While we will not be able to go into great detail, we will provide an adequate description of these topics for our purposes. For further details, the reader can consult the references cited in the respective sections dealing with these topics.

D.1

The Kalman Filter

So far in this book, we have assumed that the values of the state variables can be measured with certainty. In many cases the assumption that the value of a state variable can be directly measured and exactly determined may not be realistic. There are two types of random disturbances present. The ﬁrst kind, termed measurement noise, arises because of imprecise measurement instruments, inaccurate recording systems, etc. In many cases the measurement technique involves observations of functions of state variables, from which the values of some or all of the state variables are inferred; e.g., measuring the inventory of a natural gas reservoir involves pressure © Springer Nature Switzerland AG 2019 S. P. Sethi, Optimal Control Theory, https://doi.org/10.1007/978-3-319-98237-3

441

442

D. Special Topics in Optimal Control

measurements together with physical laws relating pressure and volume. The second kind can be termed system noise, in which the system itself is subjected to random disturbances. For instance, sales may follow a stochastic process, which aﬀects the system equation (6.1) relating inventory, production, and sales. In the cash balance example, the demand for cash as well as the interest rates in (5.1) and (5.2) can be represented by stochastic processes. In analyzing systems in which one or both of these kinds of noises are present, it is important to be able to make good estimates of the values of the state variables. We discuss the Kalman and Kalman-Bucy ﬁlters devoted to optimal estimation of current values of state variables given past measurements. The Kalman ﬁlter will be described in this section, for which further details can be obtained from references such as Kalman (1960a,b), Bryson and Ho (1975), Anderson and Moore (1979), and Kumar and Varaiya (1986). The Kalman-Bucy ﬁlter for continuoustime linear systems will be described brieﬂy in Sect. D.3 and the readers can refer to Fleming and Rishel (1975) and Arnold (1974) for further details. Consider a dynamic stochastic system in discrete time described by the diﬀerence equation xt+1 − xt = At xt + Gt wt , t = 0, 1, ..., N − 1,

(D.1)

xt+1 = (At + I)xt + Gt wt , t = 0, 1, ..., N − 1,

(D.2)

or xt

wt

where is an n-component (column) state vector, is a k-component (column) system noise vector, At is an n × n matrix, and Gt is an n × k matrix. The initial state x0 is assumed to be a Gaussian (normal) random variable with mean and n × n covariance matrix given by ¯0 and E[(x0 − x ¯0 )(x0 − x ¯0 ) ] = Σ0 . E[x0 ] = x

(D.3)

Without loss of generality, we conﬁne ourselves to the case when wt is a standard Gaussian purely random sequence with E[wt ] = 0 and E[wt (wτ ) ] = Iδ tτ , where for t = 0, 1, ..., N, τ = 0, 1, ..., N, ⎧ ⎪ ⎨ 0 if t = τ , δ tτ = ⎪ ⎩ 1 if t = τ .

(D.4)

(D.5)

D.1. The Kalman Filter

443

Thus, the random vectors wt and wτ are independent standard normal variables for t = τ . We also assume that the sequence wt is independent of the initial condition x0 , i.e., the k × n matrix ¯0 ) ] = 0, t = 0, 1, ..., N. E[wt (x0 − x

(D.6)

The process of measurement of the state variables xt yields a rdimensional vector y t which is related to xt by the transformation y t = Ht xt + v t , t = 0, 1, ..., N,

(D.7)

where Ht is the state-to-measurement transformation matrix of dimension r × n, and v t is a Gaussian purely random sequence of r-dimensional measurement noise vectors having the following properties: E[v t ] = 0, E[v t (v τ ) ] = Rt δ tτ ,

(D.8)

¯0 )(¯ v t ) ] = 0. E[wt (v τ ) ] = 0, E[(x0 − x

(D.9)

In (D.8) the matrix Rt is the r × r covariance matrix for the random variable v t , and it is therefore positive semideﬁnite, symmetric, and nonsingular. The requirements in (D.9) mean that the additive measurement noise is independent of the system noise as well as the initial state. Given a sequence of observations y 0 , y 1 , y 2 , . . . , y i up to time i, we would like to obtain the maximum likelihood estimate of the state xi , or equivalently, to ﬁnd the weighted least squares estimate. In order to derive the estimate x ˆi of xi , we require the use of the Bayes theorem and an application of calculus to ﬁnd the unconstrained minimum of a quadratic form. This derivation is straightforward but lengthy. It yields the following recursive procedure for ﬁnding the estimate x ˆt , t = 0, 1, ..., i, i ≤ N : x ˆt = x ¯t + Kt (y t − Ht x ¯t ), x ¯

t+1

t

(D.10)

0

= (At + I)ˆ x, x ¯ given,

Kt = Pt = Σt+1 =

Pt Ht Rt−1 , −1 −1 (Σ−1 t + H t Rt H t ) , (At + I)Pt (I + At ) + Gt Gt ,

(D.11)

(D.12) (D.13) Σ0 given.

(D.14)

The procedure in expressions (D.10)–(D.14) is known as the Kalman ﬁlter for linear discrete-time processes.

444

D. Special Topics in Optimal Control

The interpretation of (D.10) is that the estimate x ˆt is equal to the mean value x ¯t plus a correction term which is proportional to the diﬀerence between the actual measurement y t and the predicted measurement ¯t . Also, Ht x Σt = E[(xt − x ¯t )(xt − x ¯t ) ], the error covariance before the measurement at time t, and ˆt )(xt − x ˆt ) ], Pt = E[(xt − x the error covariance matrix after the measurement at time t. In other words, Σt and Pt are measures of uncertainties in the state before and after the measurement at time t, respectively. Thus, the proportionality matrix Kt can be interpreted as the ratio between the uncertainty Pt in the state and the measurement uncertainty Rt . Because of this property of Kt , it is called the Kalman gain in the engineering literature. It is important to note that the propagation of Pt given by (D.13) and (D.14) is independent of the measurements. Thus, it can be computed oﬄine and stored. The computation of updated estimates by (D.10) and (D.11) involves only the current measurement and error covariance, and can therefore be done in real time. Finally, prediction of the state beyond the period up to which measurements are available can be done as ¯t+1 = (At + I)ˆ x t + Gt w ¯ t , t ≥ i, i ∈ N, x ˆt+1 = x

(D.15)

with x ˆi obtained from the ﬁlter (D.10)–(D.14).

D.2

Wiener Process and Stochastic Calculus

A continuous 1-dimensional process Z is a (standard) Wiener process on an interval [0, T ] if 1. Z has independent increments; 2. The increment Zt − Zτ is Gaussian with mean 0 and variance |t − τ | for any t, τ , ∈ [0, T ]; 3. Z0 is Gaussian with mean 0. This deﬁnition easily generalizes to deﬁne a k-dimensional Wiener process. A Wiener process is also called a Brownian motion, as it models the motion of a particle in a ﬂuid. It has been shown that a Wiener process is

D.2. Wiener Process and Stochastic Calculus

445

nowhere diﬀerentiable; a Brownian particle does not possess a velocity at any instant. Furthermore, it is a process with unbounded variation, i.e., its length in any ﬁnite interval is inﬁnite. The Wiener process is diﬃcult to draw, although Fig. D.1 is an attempt to sketch a continuous sample path that, at the same time, conveys the ﬂavor of its “wild” nature. Nevertheless, the formal time derivative of a Wiener process is termed white noise in the engineering literature. Thus, wt = dZ/dt can be regarded as a stationary process in which the random variables wt and wτ , t = τ , are independent with Ewt = Ewτ = 0 and the covariance E[wt ws ] = δ tτ . One can see that wt is a continuous time analogue of the discrete-time process wt deﬁned in the previous section. t Next we wish to deﬁne an integral s G(τ )dZτ for a rather wide class of processes G. Speciﬁcally, it will be the T class M0 of all real-valued, stochastic processes G on [0, T ] such that 0 |G(τ )|2 dτ < ∞ with probability 1. Given the wild nature of the Wiener process, the integral cannot be deﬁned in the sense of Reimann-Steiltjes for every function in M0 . Therefore, we resort to the concept of a stochastic integral in the Itˆ o sense. For this, let us deﬁne the subclass M ⊂ M0 such that any G ∈ M T satisﬁes E 0 |G(τ )|2 dτ < ∞. Let Gj ∈ M be a step process on [0, t] in the sense that there is a partition consisting of points τ 0 , τ 1 , . . . , τ m with 0 < τ 0 < τ 1 < . . . < τ m = t. For this step process, the integral equals the Riemann-Steiltjes sum t m Gj (τ )dZτ = Gj (τ k−1 )[Zτ k − Zτ k−1 ]. (D.16) 0

k=1

We then deﬁne the stochastic integral for any G ∈ M0 by taking a t sequence of step processes Gj , j = 1, 2, . . . , such that 0 |Gj (τ )−G(τ )|2 dτ converges to zero in probability as j → ∞. Then, the sequence of random variables deﬁned in(D.16) converges, as j → ∞, to t a limit in probability, t which is deﬁned as 0 G(τ )dZτ , written simply as 0 GdZ. It can be shown that the limit does not depend on the approximating sequence Gj with probability 1 for each t. It is important to note the following important properties of Itˆo’s t stochastic integral. The integral 0 GdZ can be deﬁned simultaneously for all t ∈ [0, T ], so that it is continuous on [0, T ]. Furthermore, for any H, G ∈ M0 , we have t t G(τ )dZτ = 0, E H(τ )dZτ = 0, E 0

and

0

446

D. Special Topics in Optimal Control

t

E 0

G(τ )dZτ

t 0

H(τ )dZτ

t

G(τ )H(τ )dτ .

=E 0

(D.17) Equation (D.17) serves as motivation for the frequently used symbolic notation (D.18) (dZt )2 = dt. Now that we have deﬁned the stochastic integral, it remains to specify the stochastic diﬀerential rule. T Let f, G, and X be one-dimensional stochastic processes such that E 0 |f |dt < ∞, G ∈ M0 , X is continuous, and t t X t − X0 = f (τ )dτ + G(τ )dZτ , 0 ≤ t ≤ T. (D.19) 0

0

This equation is a stochastic integral equation, for which it is customary to use the suggestive notation dXt = f (t)dt + G(t)dZt , X0 given, or simply dX = f dt + GdZ, X0 given.

(D.20)

Now let the one-dimensional process Yt = ψ(Xt , t), t ∈ [0, T ], where the function ψ(x, t) is continuously diﬀerentiable in t and twice continuously diﬀerentiable in x. Then, it possesses the stochastic diﬀerential 1 dYt = ψ t (Xt , t) + ψ X (Xt , t)dxt + ψ XX (Xt , t)G2 (t)dt 2 1 = [ψ t (Xt , t) + ψ X (Xt , t)f (t) + ψ XX (Xt , t)G2 (t)]dt 2 (D.21) +ψ X (Xt , t)G(t)dZt , Y0 = ψ(X0 , 0). Equation (D.21) is to be interpreted in the sense that its integral form from 0 to t holds with probability 1, i.e., Y (xt , t) = Y (x0 , 0) t 1 + [ψ s (xs , s) + ψ x (xs , s)f (s) + ψ xx (xs , s)G2 (s)]ds 2 0 t ψ(xs , s)G(s)dZs , w.p.1. (D.22) + 0

D.3. The Kalman-Bucy Filter

447

It is worth pointing out that the term 12 ψ xx G2 dt does not appear in the diﬀerential rule of elementary calculus. This is an important difference as seen in Chap. 12, where we discuss stochastic optimal control problems. Also, a multi-dimensional generalization of (D.16)–(D.22) is straightforward.

D.3

The Kalman-Bucy Filter

The continuous-time analogue of the Kalman ﬁlter is known as the Kalman-Bucy ﬁlter. Here, the diﬀerence equation (D.2) is replaced by the linear stochastic diﬀerential equation dXt = A(t)Xt dt + G(t)dZt , 0 ≤ t ≤ T,

(D.23)

which is a special case of the Itˆ o stochastic diﬀerential equation (D.20) introduced in Chap. 12. In this equation, Xt is an n-component (column) state vector, Zt is the value at time t of a standard k-component (column) Wiener process Z, and the matrices A(t) and G(t) of dimensions n × n and n × k, respectively, are continuous in t. Furthermore, ¯ 0 , and E[(X0 − X ¯ 0 )(X0 − X ¯ 0 ) ] = Σ0 . E(X0 ) = X

(D.24)

The measurement process (D.7) is replaced by dYt = H(t)Xt + σ(t)dξ t , Y0 = 0,

(D.25)

where ξ is a standard r-dimensional Weiner process and the k × r matrix σ(t) is such that the k × k matrix R(t) := σ(t)σ (t) is positive deﬁnite. Note that the term σ(t)dξ t in (D.25) represents the noise term, which corresponds to v t in (D.7). Thus, the term R(t) corresponds to the covariance matrix Rt in Sect. D.1 on the Kalman ﬁlter. The ﬁltering problem is to ﬁnd the weighted least square estimate of Xt given the measurements up to time t. It can be shown that the optimal estimate is the conditional expectation ˆ t = E[Xt |Ys , 0 ≤ s ≤ t]. X

(D.26)

Furthermore, it can be obtained recursively by the following KalmanBucy ﬁlter: ˆ t dt + K(t)[dYt − H(t)X ˆ t dt], X ˆ0 = X ¯ 0 , (D.27) ˆ t = A(t)X dX K(t) = P (t)H (t)R−1 (t),

(D.28)

448

D. Special Topics in Optimal Control P˙ (t) = (A(t)P (t) + P (t)A (t) − K(t)H(t)P (t) +G(t)G (t)), P (0) = Σ0 ,

(D.29)

where H (t) denotes the transpose (H(t)) and R−1 (t) means the inverse (R(t))−1 , as the notational convention deﬁned in Chap. 1. The interpretations of P (t) and K(t) are the same as in the previous section. The ﬁlter (D.27)–(D.29) is the Kalman-Bucy ﬁlter (Kalman and Bucy 1961) for linear systems in continuous time. Equation (D.29) is called the matrix Riccati equation. Besides engineering applications, the Kalman ﬁlter and its extensions are very useful in econometric and ﬁnancial modeling; see Buchanan and Norton (1971), Chow (1975), Aoki (1976), Naik et al. (1998), and Bhar (2010).

D.4

Linear-Quadratic Problems

An important problem in systems theory, especially engineering sciences, is to synthesize feedback controllers. These controllers provide optimal control as a function of the state of the system. A usual method of obtaining these controllers is to solve the Hamilton-Jacobi-Bellman partial diﬀerential equation (2.19). This equation is nonlinear in general, which makes it very diﬃcult to solve in closed form. Thus, it is not possible in most cases to obtain optimal feedback control schemes explicitly. It is, however, feasible in many cases to obtain perturbation feedback control, which refers to control in the vicinity of an optimal path. These perturbation schemes require the approximation of the problem by a linear-quadratic problem in the vicinity of an optimal path (see Sect. D.5), and feedback control for the approximating problem is easy to obtain. A linear-quadratic control problem is a problem with linear dynamics and a quadratic objective function. First, we treat a special case called the Regulator Problem:

min x (T )ST x(T ) + u

T

(x Cx + u Du)dt

(D.30)

x(0) = x0 .

(D.31)

0

subject to x˙ = Ax + Bu,

Here x ∈ E n , u ∈ E m , and the appropriate dimensional matrices C, D, A, and B, when time-dependent, are assumed to be continuous in

D.4. Linear-Quadratic Problems

449

time t. Furthermore, we shall assume the matrices C and ST to be positive semideﬁnite and, without loss of generality, symmetric, and matrix D to be symmetric and positive deﬁnite. To solve the regulator problem for an explicit feedback controller, we rewrite it as that of maximizing T J= −(x Cx + u Du)dt − x (T )ST x(T ) 0

subject to (D.31). Clearly, this is a special case of the optimal control problem (2.4) and we can apply (2.15) and (2.16) to obtain the HamiltonJacobi-Bellman equation . 0 = max −(x Cx + u Du) + Vx [Ax + Bu] + Vt (D.32) u

with the terminal boundary condition V (x, T ) = −x (T )ST x(T ).

(D.33)

By checking that V (γx, t) = γ 2 V (x, t) and V (x, t)+V (y, t) = 12 [V (x+ y, t) + V (x − y, t), we can establish that the value function V (x, t) is of a quadratic form. Thus, let V (x(t), t) = −x (t)S(t)x(t)

(D.34)

for some matrix S(t), symmetric without loss of generality. Then Vt = ˙ and Vx = −2(Sx) = −2x S. Using these relations in (D.32), we −x Sx get . ˙ x Sx = max −x Cx − u Du − 2x SAx − 2x SBu u . = − min x Cx + u Du + 2x SAx + 2x SBu . (D.35) u

To ﬁnd the minimum of the expression on the right-hand side of (D.35), we observe the following identity obtained by completing the square: x Cx + u Du + 2x SAx + 2x SBu = (u + D−1 B Sx) D(u + D−1 B Sx) +x (C − SBD−1 B S + SA + A S)x. Because matrix D is positive deﬁnite, it follows that the minimum is achieved in (D.35) by the control u∗ = −D−1 B Sx.

(D.36)

450

D. Special Topics in Optimal Control

Then from (D.35) and (D.36), we obtain, ˙ = −x [C − SBD−1 B S + SA + A S]x. x Sx

(D.37)

Since this equation holds for all x, we have the matrix diﬀerential equation (D.38) S˙ = −SA − A S + SBD−1 B S − C, called a matrix Riccati equation, with the terminal condition (D.39)

S(T ) = ST

obtained from (D.33), where ST is speciﬁed in (D.30). A solution procedure for Riccati equations appears in Bryson and Ho (1975) or Anderson and Moore (1990). With the solution S of (D.38) and (D.39), we have the optimal feedback control as in (D.36). To see that the optimal control u∗ in (D.36) maximizes the Hamiltonian H = −x Cx − u Du + Vx [Ax + Bu], let us use (D.32) to obtain 2(Du∗ ) = 2u∗ D = −2x SB(D )−1 D = −2x SB = Vx B, which is precisely the ﬁrst-order condition for the maximum of the righthand side of (D.32). Moreover, the ﬁrst-order condition yields a global maximum of the Hamiltonian, which is concave since the matrix D is positive deﬁnite. A generalization of (D.30) to include a cross-product term to allow for interactions between the state x and control u, which would be useful in the next section on the second variation, is to set

J = −x (T )ST x(T ) −

0

⎡ T

⎤⎡

⎤

⎢ C N ⎥⎢ x ⎥ (x , u ) ⎣ ⎦ ⎣ ⎦ dt, N D u

(D.40)

and the problem is to maximize J subject to the state equation (D.31). It is easy to see that the integrand in (D.40) can be rewritten as x Cx+u Du+2x N u. Furthermore, with the deﬁnition u ˜ = u+D −1 N x, the generalized problem deﬁned by (D.40) and (D.31) can be reduced to T u D−1 u ˜ + x (C − the standard regulator problem of maximizing 0 −[˜ ˜, proN D−1 N )x] − x (T )ST x(T ) subject to x˙ = (A − BD−1 N )x + B u vided that the matrix C − N D −1 N is positive semideﬁnite. We can then use formulas (D.36), (D.38), and (D.39), to obtain the solution of

D.4. Linear-Quadratic Problems

451

the transformed problem and then use the deﬁnition of u ˜ to write the feedback control of the generalized problem as u∗ (x) = −D−1 [N + B S]x,

(D.41)

where S˙ = −S(A − BD−1 N ) − (A − N D−1 B )S +SBD−1 B S + N D−1 N − C = −SA − A S + (SB + N )D−1 (B S + N ) − C

(D.42)

with S(T ) = ST .

D.4.1

(D.43)

Certainty Equivalence or Separation Principle

Suppose Eq. (D.31) is changed by the presence of the stochastic term G(t)dZt as deﬁned in (D.23) so that we have the Itˆo equation dXt = (A(t)Xt + B(t)Ut )dt + G(t)dZt , and X0 is a normal random variable with E[X0 ] = 0,

E[X0 X0 ] = Σ0 .

Because of the presence of uncertainty in the system equation, we modify the objective function in (D.40) as follows: ⎧ ⎡ ⎛ ⎞⎛ ⎞ ⎤⎫ ⎪ ⎪ ⎨ T Ct Nt ⎟ ⎜ Xt ⎟ ⎥⎬ ⎢ ⎜ max J = E ⎣−XT ST XT − (Xt , Ut ) ⎝ ⎠⎝ ⎠ dt⎦ . ⎪ ⎪ 0 ⎩ ⎭ Nt Dt Ut Assume further that Xt cannot be directly measured and the measurement process is given by (D.25), i.e., dYt = H(t)Xt + σ(t)dξ t , Y0 = 0. The optimal control Ut∗ for this linear-quadratic stochastic optimal control problem can be shown to be given by (D.41) with Xt replaced ˆ t ; see Arnold (1974). Thus, by its estimate X ˆt, Ut∗ = −D(t)−1 [N (t) + B (t)S(t)]X

452

D. Special Topics in Optimal Control

ˆ t is given by the Kalmanwhere S(t) is given by (D.42) and (D.43), and X Bucy ﬁlter: ˆ t + B(t)U ∗ ]dt + K(t)(dYt − H(t)X ˆ t dt), X(0) ˆ ˆ t = [A(t)X = 0, dX t K(t) = P (t)H (t)R−1 (t), P˙ (t) = A(t)P (t) + P (t)A (t) − K(t)H(t)P (t) + G(t)G (t), P (0) = Σ0 . The above procedure has received two diﬀerent names in the literature. In economics it is called the certainty equivalence principle; see Simon (1956). In engineering and mathematics literature it is called the separation principle; see Fleming and Rishel (1975). When we call it ˆt the certainty equivalence principle, we are emphasizing the fact that X can be used for the purposes of optimal feedback control as if it were the certain value of the state variable Xt . Whereas the term separation principle emphasizes the fact that the process of determining the optimal control can be broken down into two steps: ﬁrst, estimate Xt by using the optimal ﬁlter; second, use that estimate in the optimal feedback control formula for the deterministic problem.

D.5

Second-Order Variations

Second-order variations in optimal control theory are analogous to the second-order conditions in the classical optimization problem of calculus. To discuss the second-order variational condition is diﬃcult when the control variable u is constrained to be in the control set Ω. So we make the simplifying assumption that Ω = Rm , and thus the control u is unconstrained. As a result, we are now dealing with the problem:

max J = u

T

F (x, u, t)dt + Φ[x(T )]

(D.44)

0

subject to x˙ = f (x, u, t),

x(0) = x0 .

(D.45)

From Chap. 2, we know that the ﬁrst-order necessary conditions for this problem are given by λ˙ = −Hx ,

λ(T ) = 0,

Hu = 0,

(D.46) (D.47)

D.5. Second-Order Variations

453

where the Hamiltonian H is given by H = F + λf.

(D.48)

Since u is unconstrained, these conditions may be easily derived by the method of calculus of variations. To see this, we write the augmented objective functional as T [H(x, u, λ, t) − λx]dt. ˙ (D.49) J¯ = Φ[x(T )] + 0

Consider small perturbation from the extremal path given by (D.45)– (D.48) as a result of small perturbations δx(0) in the initial state. Deﬁne the resulting perturbations in state, adjoint, and control variables by δx(t), δλ(t), and δu(t), respectively. These, of course, will be obtained by linearizing ((D.45)–(D.47)) around the external path: dδx = fx δx + fu δu, dt

δx(0)speciﬁed,

(D.50)

dδλ = −(Hxx δx)T − δλf T − (Hxu δu), dt

(D.51)

δHu = (Hux δx)T + δλ(Hu λ)T + (Huu δu)T = (Huu δx)T + δλfu + (Huu δu)T = 0.

(D.52)

Alternatively, we may consider an expansion of the objective function and the state equation to second order since the ﬁrst-order terms vanish about a trajectory which satisﬁes ((D.44)–(D.47)). From Bryson and Ho (1975), this may be accomplished by expanding (D.49) to second order and all the constraints to ﬁrst order. Thus, we have

1 1 δ 2 J¯ = (δxT (T )Φxx δx(T )) + 2 2

⎡

T 0

⎤⎡ ⎤ H δx H xu ⎥ ⎢ ⎢ xx ⎥ (δx, δu) ⎣ ⎦⎣ ⎦ dt δu Hux Huu (D.53)

subject to dδx (D.54) = fx δx + fu δu, δx(0) speciﬁed. dt Since we are interested in a neighboring extremal path, we must determine δu(t) so as to maximize δ 2 J¯ subject to (D.54). This problem is

454

D. Special Topics in Optimal Control

a linear-quadratic problem discussed in the previous section. For this problem, the optimal control δu∗ (t) is given by the formula (D.42), provided Huu (t) is nonsingular for 0 ≤ t ≤ T. The case when Huu (t) is singular for a ﬁnite time interval is treated in Sect. D.6. Thus, recognizing that G = Φxx , C = Hxx , N = Hxu , D = Huu , A = fx , and B = fu , we have −1 δu∗ (t) = Huu [Hux + fuT S(t)]δx(t), (D.55) where −1 T S˙ +Sfx +fuT S −(Sfu +Hxu )Huu (fu S +Hux )+Hxx = 0,

S(T ) = Φxx . (D.56) While a number of second-order conditions can be obtained by proceeding further from this manner, we will be interested only in the concavity condition (or strengthened Legendre-Clebsch condition). It is possible to show that neighboring stationary paths exist (in a weak sense; i.e., δx and δu are small) if Huu (t) < 0

for

0 ≤ t ≤ T,

(D.57)

or in other words, Huu (t) is negative semideﬁnite. First-order conditions, conditions (D.57), and the condition that S(t) is ﬁnite for 0 ≤ t ≤ T represent suﬃcient conditions for a trajectory to be a local maximum. We are not being speciﬁc here because in this book we would be relying mostly on the suﬃciency conditions developed in Chaps. 2–4, which are based on certain concavity requirements. We are stating (D.57) because of its similarity to the second-order condition for a local maximum in the classical maximization problem. We must note that Hu = 0

and

Huu ≤ 0

(D.58)

form necessary conditions for a trajectory to be a local maximum.

D.6

Singular Control

In some optimization problems including some problems treated in this text, extremal arcs satisfying Hu = 0 occur on which the matrix Huu is singular. Such arcs are called singular arcs. Note that these arcs satisfy (D.58) but not the strengthened condition (D.57). While no general suﬃciency conditions are available for singular arcs, some additional necessary conditions known as the generalized Legendre-Clebsch conditions

D.6. Singular Control

455

have been developed. A good reference on singular control is Bell and Jacobson (1975). We will only discuss the case in which the Hamiltonian is linear in one or more of the control variables. For these systems, Hu = 0 implies that the coeﬃcient of the linear control term in the Hamiltonian vanishes identically along a singular arc. Thus, the control is not determined in terms of x and λ by the Hamiltonian maximizing condition Hu = 0. Instead, the control is determined by the requirement that the coeﬃcient of these linear terms remain zero on the singular arc. That is, the time derivatives of Hu must be zero. Having obtained the control by setting dHu /dt = 0 (or by setting higher time derivatives to equal zero) along the singular arc, we must check additional necessary conditions analogous to the second-order condition (D.57). For a maximization problem with a single control variable, these conditions turn out to be

2k d Hu k ∂ ≤ 0, k = 0, 1, 2, .... (D.59) (−1) ∂u dt2k The conditions (D.59) are called the generalized Legendre-Clebsch conditions. For applications of these conditions to problems in production and ﬁnance, see e.g., Maurer et al. (2005) and Davis and Elzinga (1971). The Davis-Elzinga model is covered in Exercise 5.17 in Chap. 5. For numerical solutions of singular control problems, see Maurer (1976). Example D.1 We present an example treated by Johnson and Gibson (1963): 1 T 2 x1 dt (D.60) max J = − 2 0 subject to x˙ 1 = x2 + u, x˙ 2 = −u,

x1 (0) = a,

(D.61)

x(0) = b,

(D.62)

x1 (T ) = x2 (T ) = 0.

(D.63)

Solution We form the Hamiltonian 1 H = − x21 + λ1 (x2 + u) + λ2 (−u), 2

(D.64)

456

D. Special Topics in Optimal Control

where the adjoint equations are λ˙ 1 = x1 ,

λ˙ 2 = −λ1 .

(D.65)

The optimal control is bang-bang plus singular. Singular arcs must satisfy (D.66) H u = λ 1 − λ2 = 0 for a ﬁnite time interval. The optimal control can, therefore, be obtained by dHu = λ˙ 1 − λ˙ 2 = x1 + λ1 = 0. (D.67) dt Diﬀerentiating once more with respect to time t, we obtain d 2 Hu = x˙ 1 + λ˙ 1 = x2 + u + x1 = 0, dt2 which implies u = −(x1 + x2 )

(D.68)

along the singular arc. We now verify for the example, the generalized Legendre-Clebsch condition (D.59) for k = 1:

∂ d 2 Hu − = −1 ≤ 0. (D.69) ∂u dt2

D.7

Global Saddle Point Theorem

In this section, we provide an important result for a class of stationary inﬁnite-horizon optimal control problems such as those treated in Chap. 11. In particular, we are concerned here with the one-dimensional state problem deﬁned in (3.97) without the mixed constraint and the terminal inequality constraints, i.e., ∞ −ρt max J = φ(x, u)e dt , (D.70) 0

x˙ = f (x, u), x(0) = x0 .

(D.71)

An application of the maximum principle results in an adjoint equation λ˙ = ρλ − φx − λfx

(D.72)

D.7. Global Saddle Point Theorem

457

and a Hamiltonian maximizing control u∗ (x, λ). Substituting this for u in (D.71) and (D.72) gives rise to a canonical system of diﬀerential equations (D.73) x˙ = f ∗ (x, λ) and λ˙ = ψ ∗ (x, λ). ¯ of the system (D.73) satisﬁes A saddle point (¯ x, λ) ¯ = 0 and ψ ∗ (¯ ¯ = 0. x, λ) x, λ) f ∗ (¯

(D.74)

The important issue for this problem is the existence and uniqueness of an optimal path that steers the system from an initial value x0 to the steady state x ¯. This is equivalent to ﬁnding a value λ0 so that the system ¯ A suﬃcient x, λ). (D.73) starting from (x0 , λ0 ) moves asymptotically to (¯ condition for this to happen is given in the following theorem.

Figure D.1: Phase diagram for system (D.73)

¯ be a Theorem D.1 (Global Saddle Point Theorem) Let (¯ x, λ) unique saddle point of the canonical system (D.73) of the diﬀerential equations and let x0 be a given initial state for which the vertical line x = x0 (see Fig. D.1) intersects both isoclines x˙ = f ∗ (x, λ) = 0 and λ˙ = ψ ∗ (x, λ) = 0. Assume further that the region bounded by the isoclines and the line x = x0 has a triangular shape as in Fig. D.1 (i.e., the isoclines themselves do not intersect in the open interval between x0 and x ¯). Then, there exists a unique saddle point path starting for x = x0 and ¯ leading to the saddle point (¯ x, λ).

458

D. Special Topics in Optimal Control

The proof of this theorem, based on Theorem 1.2 and Corollaries 1.1 and 1.2 from Hartman (1982), can be found in Feichtinger and Hartl (1986).

D.8

The Sethi-Skiba Points

In Exercise 2.9, we deﬁned autonomous optimal control problems. Here, we limit the discussion to autonomous systems that are discounted inﬁnite-horizon optimal control problems with one-dimensional state, deﬁned as follows: max

u(t)∈Ω

∞

e

J=

−ρt

φ(x(t), u(t))dt

0

subject to x(t) ˙ = f (x(t), u(t)), x(0) given, with ρ > 0 as the discount rate. In addition to assuming that the function φ and f are continuously diﬀerentiable, we assume that the integral in the objective function J converges for any admissible solution x(t), u(t), t ≥ 0. In such problems, there may arise multiple equilibria depending on the initial condition. Suppose x0 is an initial value for x(0), such that the system starting from it exhibits multiple optimal solutions or equilibria. Thus, at least in the neighborhood of x0 , the system moves to one equilibrium if x(0) > x0 and to another if x(0) < x0 . In other words, x0 is an indiﬀerence point from which the system could move to either of two equilibria. Such points were originally identiﬁed by Sethi (1977b, 1979c). Subsequently, Skiba (1978) and Dechert and Nishimura (1983) explored these indiﬀerence points for one-sector optimal economic growth models with nonconvex production functions, in contrast to concave production functions treated in Sect. 11.1. These points are also referred to as the DNSS points, where the acronym DNSS stands for Dechert, Nishimura, Sethi, and Skiba. Before it became known that Sethi (1977b) had already identiﬁed them prior to Skiba (1978), these points were also called Skiba points. Below we present a simple example that exhibits a Sethi-Skiba point at x0 = 0. For further discussion on these points, see Grass et al. (2008), Zeiler et al. (2010), Kiseleva and Wagener (2010), and Caulkins et al. (2015a).

D.8. The Sethi-Skiba Points

459

Example D.2 Solve the problem: ∞ −ρt e x(t)u(t)dt max J = 0

subject to x(t) ˙ = −x(t) + u(t), x(0) = x0 ,

(D.75)

u(t) ∈ [−1, +1], t ≥ 0. Let us ﬁrst solve this problem for x0 < 0. We form the Hamiltonian H = x(t)u(t) + λ(t)(−x(t) + u(t))

(D.76)

˙ λ(t) = (1 + ρ)λ(t) − u(t).

(D.77)

with Since H is linear in u, the optimal policy is u∗ (t) = bang[−1, 1; x(t) + λ(t)].

(D.78)

For x0 < 0, the state equation reveals that u∗ (t) = −1 will give the largest decrease of x(t) and keep x(t) < 0, t ≥ 0. Thus, it will maximize the product x(t)u(t) for each t > 0. We also note that the long-run ¯ = (−1, −1, −1/(1 + ρ)). stationary equilibrium in this case is (¯ x, u ¯, λ) It is also easy to verify that the solution u∗ (t) = −1, x∗ (t) = −1 + e−t (x0 + 1), and λ(t) = −1/(1 + ρ), t ≥ 0, satisﬁes (D.75), (D.77) along with the suﬃciency transversality condition (3.99), and maximizes the Hamiltonian in (D.76). Similarly, we can argue that for x0 > 0, the optimal solution is u∗ (t) = +1, x∗ (t) = 1+e−t (x0 −1) > 0, and λ(t) = 1/(1+ρ), t ≥ 0. The ¯ = (1, 1, 1/(1 + ρ). long-run stationary equilibrium in this case is (¯ x, u ¯, λ) Then by symmetry, we can conclude that if x0 = 0, both u∗ (t) = −1 and u∗ (t) = +1, t ≥ 0, yield the same objective function, and hence both are optimal. Thus, x0 = 0 is a Sethi-Skiba point for this example. Clearly, at this point, the choice between using u∗ (0) = −1 and u∗ (0) = +1 will determine the equilibrium the system approaches. Notice that once the system has moved away from x0 = 0, there is no more choice left in choosing the control. It is possible that at a Sethi-Skiba point, a decision maker can inﬂuence the equilibrium that the system would move to, by choosing a

460

D. Special Topics in Optimal Control

control from the set of possible optimal controls. This may have important implications. In a model of controlling illicit drugs, Grass et al. (2008) derive a Sethi-Skiba point, signifying a critical number of addicts, such that if there are fewer addicts than the critical number, it is optimal to use an eradication strategy that uses massive treatment spending that drive the number of addicts down to zero. On the other hand, if there are more than the critical number of addicts, then it is optimal to use an accommodation strategy that uses a moderate level of treatment spending that balances the social cost of drug use and the cost of treatment. This is a case of a classic Sethi-Skiba point acting as a “tipping point” between the two strikingly diﬀerent equilibria, one of which may be more socially or politically favored than the other, and the social planner can use an optimal control to move to the more favored equilibrium. We conclude this subsection by mentioning that the Sethi-Skiba points are exhibited in the production management context by Feichtinger and Steindl (2006) and Moser et al. (2014), in the open-source software context by Caulkins et al. (2013a), and in other contexts by Caulkins et al. (2011, 2013b, 2015a).

D.9

Distributed Parameter Systems

Thus far, our eﬀorts have been directed to the study of the control of systems governed by systems of ordinary diﬀerential or diﬀerence equations. Such systems are often called lumped parameter systems. It is possible to generalize these to systems in which the state and control variables are deﬁned in terms of space as well as time dimensions. These are called distributed parameter systems and are described by a set of partial diﬀerential or diﬀerence equations. For example, in the lumped parameter advertising models of the type treated in Chap. 7, we solved for the optimal rate of advertising expenditure at each instant of time. However, in the analogous distributed parameter advertising models, we must obtain the optimal advertising expenditure rate at every geographic location of interest at each instant of time; see Seidman et al. (1987) and Marinelli and Savin (2008). In other economic problems, the spatial coordinates might be income, quality, age, etc. Derzko et al. (1980), for example, discuss a cattle-ranching model in which the spatial dimension measures the age of a cow. Let y denote a one dimensional spatial coordinate, let t denote time,

D.9. Distributed Parameter Systems

461

and let x(t, y) be a one dimensional state variable. Let u(t, y) denote a control at (t, y) and let the state equation be ∂x ∂x = g(t, y, x, , u) ∂t ∂y

(D.79)

for t ∈ [0, T ] and y ∈ [0, h]. We denote the region [0, T ] × [0, h] by D, and we let its boundary ∂D be split into two parts Γ1 and Γ2 as shown in Fig. D.2. The initial conditions will be stated on the part Γ1 of the boundary ∂D as x(0, y) = x0 (y) (D.80) and x(t, 0) = v(t).

(D.81)

In Fig. D.2, (D.80) is the initial condition on the vertical portion of Γ1 , whereas (D.81) is that on the horizontal portion of Γ1 . More speciﬁcally, in (D.80) the function x0 (y) gives the starting distribution of x with respect to the spatial coordinate y. The function v(t) in (D.81) is an exogenous breeding function of x at time t when y = 0, which in the cattle ranching model mentioned above, measures the number of newly born calves at time t. To be consistent we make the obvious assumption that x(0, 0) = x0 (0) = v(0). (D.82)

Figure D.2: Region D with boundaries Γ1 and Γ2 Let F (t, y, x, u) denote the proﬁt rate when x(t, y) = x and u(t, y) = u at a point (t, y) in D. Let Q(t) be the price of one unit of x(t, h) at

462

D. Special Topics in Optimal Control

time t and let S(y) be the salvage value of one unit of x(T, y) at time T. Then the objective function is: T h F (t, y, x(t, y), u(t, y))dydt J= max u(t,y)∈Ω 0 0 T h (D.83) + Q(t)x(t, h)dt + S(y)x(T, y)dy , 0

0

where Ω is the set of allowable controls. We will formulate, without giving proofs, a procedure for solving the problem in (D.79)–(D.83) by a distributed parameter maximum principle, which is analogous to the ordinary one. A more complete treatment of this topic can be found in Sage (1968), Butkowskiy (1969), Ahmed and Teo (1981), Tzafestas (1982b), Derzko et al. (1984), Brokate (1985), and Veliov (2008). In order to obtain necessary conditions for a maximum, we introduce the Hamiltonian H = F + λf, (D.84) where the spatial adjoint function λ(t, y) satisﬁes

∂λ ∂H ∂ ∂H ∂ ∂H =− + + , ∂t ∂x ∂t ∂xt ∂y ∂xy

(D.85)

where xt = ∂x/∂t and xy = ∂x/∂y. The boundary conditions on λ are stated for the Γ2 part of the boundary of D (see Fig. D.2) as follows: λ(t, h) = Q(t)

(D.86)

λ(T, y) = S(y).

(D.87)

and Once again we need a consistency requirement similar to (D.82). It is λ(T, h) = Q(T ) = S(h),

(D.88)

which gives the consistency requirement in the sense that the price and the salvage value of a unit x(T, h) must agree. We let u∗ (t, y) denote the optimal control at (t, y). Then the distributed parameter maximum principle requires that H(t, y, x∗ , x∗t , x∗y , u∗ , λ) ≥ H(t, y, x∗ , x∗t , x∗y , u, λ) for all (t, y) ∈ D and all u ∈ Ω.

(D.89)

Exercises for Appendix D

463

We have stated only a simple form of the distributed parameter maximum principle which is suﬃcient for most applications in management science and economics, such as Derzko et al. (1980), Haurie et al. (1984), Feichtinger et al. (2006a), and Kuhn et al. (2015). More general forms of the maximum principle are available in the references cited earlier. Among other things, these general forms allow for the function F in (D.83) to contain arguments such as ∂x/∂y, ∂ 2 x/∂y 2 , etc. It is also possible to consider controls on the boundary. In this case v(t) in (D.81) will become a control variable. Exercises for Appendix D E D.1 Consider the discrete-time dynamics ⎧ ⎪ ⎨ xt+1 − xt = axt + wt , ⎪ ⎩ yt

=

hxt

+

(D.90)

vt,

where wt and v t are Gaussian purely random sequences with E[wt ] = E[v t ] = 0, E[wt wτ ] = qδ tτ , E[v t v τ ] = rδ tτ , where h, q, and r are constants. The initial condition x0 is a Gaussian random variable with mean μ and variance Σ0 . Use the Kalman ﬁlter (D.10)–(D.14) to obtain the recursive equations Pt+1 h t+1 xt + − h(a + 1)ˆ xt ), x ˆ0 = μ (y x ˆt+1 − xˆt = aˆ r and Pt+1 =

r[(a + 1)2 Pt + q] , p0 = rΣ0 /(r + Σ0 h2 ). r + h2 [(a + 1)2 Pt + q]

E D.2 Consider the continuous-time dynamics of the simplest nontrivial ﬁlter ⎧ ⎪ ⎨ dXt = √q dZt , x0 given, (D.91) ⎪ ⎩ dY = X + √r dξ , Y = 0, t t t 0

464

D. Special Topics in Optimal Control

where Z and ξ are standard Brownian motions, q and σ are positive constants, and X0 is a Gaussian random variable with mean 0 and variance Σ0 . Show that the Kalman-Bucy ﬁlter is given by ˆt = dX

P (t) ˆ t dt), X ˆ 0 = 0, (dYt − X r

and P (t) = where α=

#

√

rq

1 + be−2αt , 1 − be−2αt

q/r and b =

√ Σ0 − rq √ . Σ0 + rq

Hint: In solving the Riccati equation for P (t), you will need the formula u−a 1 du = ln . u2 − a 2 2a u+a E D.3 Let w(u) = u in Exercise 7.39. Analyze the various cases that may arise in this problem from the viewpoint of obtaining the Sethi-Skiba points. E D.4 The economic growth model of Sect. 11.1.3 exhibits a Sethi-Skiba point if we assume the production function f (k) to be convex initially and then concave, i.e., f (k) > 0 for k < k s and f (k) = 0 at k = k s and f (k) < 0 for k > k s for some k s ∈ (0, ∞). Analyze this problem with the additional mixed constraints 0 ≤ c ≤ f (k). See Skiba (1978) and Feichtinger and Hartl (1986).

Appendix E

Answers to Selected Exercises Completely worked solutions to all exercises in this book are contained in a forthcoming Teachers’ Manual, which will be made available to instructors by the publisher when it is ready. Chapter 1 1.1 (a) Feasible. J = −333,333. 1.3

J = 36.

1.5 (a) C = $157,861/year. (b) J = 103.41 utils. (c) $15,000/year. 1.6 (b) W (20) = 985,648; J = 104.34. 1.14

imp(G1 , G2 ; t) = (G1 − G2 )e−ρt . Chapter 2

2.4

The optimal ⎧ control is ⎪ ⎪ ⎪ 2 if 0 ≤ t ≤ 2 − ln 2.5, ⎪ ⎪ ⎨ u∗ (t) = undeﬁned if t = 2 − ln 2.5, ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 0 if t > 2 − ln 2.5.

© Springer Nature Switzerland AG 2019 S. P. Sethi, Optimal Control Theory, https://doi.org/10.1007/978-3-319-98237-3

465

466 2.17

E. Answers to Selected Exercises u∗ = bang(0, 1; λ1 − λ2 ), where λ(t) = (8e−2(t−18) , 4e−2(t−18) ).

2.18 (a) x(100) = 30 − 20e−10 ≈ 30. (b) u∗ = 3 for t ∈ [0, 100]. ⎧ ⎪ ⎨ 3 for t ∈ [0, 100 − 10 ln 2], ∗ (c) u (t) = ⎪ ⎩ 0 otherwise. 2.19

x˙ = f (x) + b(x)u, x(0) = x0 , x(T ) = 0. u˙ = [b(x)2 g (x) − 2cu{b(x)f (x) − b (x)f (x)}]/[2cb(x)].

2.22 (a) u∗ = bang[0, 1; (g1 K1 + g2 K2 )(λ1 − λ2 )]. (c) tˆ = T − (1/g2 ) ln[(g2 b1 − g1 b2 )/(g2 − g1 )b2 ]. 2.29 (a) C ∗ (t) = ρW0 e(r−ρ)t /(1 − e−ρT ). (b) C˙ ∗ (t) = K(r − ρ). 2.30 (a) λ˙ = x + 3λx2 , λ(1) = 0, and x˙ = −x3 + λ, x(0) = 1. Chapter 3 3.1

x − u1 ≥ 0, u1 − u2 ≥ 0, u1 ≥ 0, 1 + u2 ≥ 0.

3.2

X = [−1, 5].

3.9

L = F (x, u) + λf (x, u, t) + μg(x, u, t), λ˙ = −(α/α)λ ˙ −

∂L ∂x ,

μ ≥ 0, μg = 0.

$ 4 3.13 (a) λ(t) = 10 1 − e0.1(t−100) , ⎧ ⎪ ⎨ 0 if K = 300, μ= ⎪ ⎩ −10 41 − e0.1(K/3−100) $ if K < 300, u∗ (t) = bang[0, 3; λ + μ]. The problem is infeasible for K > 300.

E. Answers to Selected Exercises

467

(b) t∗∗ = min[0, ⎧ 100 − K/3], ⎪ ⎨ 0 for t ≤ t∗∗ , ∗ u (t) = ⎪ ⎩ 3 for t > t∗∗ . 3.17

λ(t) = t − 1.

3.23

11.87 min.

3.25

u∗ = −1, T ∗ = 5.

3.26

u∗ = −2, T ∗ = 5/2.

¯ = {I1 − ρ(S − P1 ), S, 2(S − P1 )}. ¯ P¯ , λ} 3.43 (a) {I, (b) I = I1 . Chapter 4 4.2

u∗ (t) = −1, μ1 = −λ = 1/2 − t, μ2 = η = 0.

4.3

One solution appears in Fig. 3.1. Another solution is u(t) = 1/2 for t ∈ [0, 2]. There are many others.

4.5 (a) u∗ = 0. ⎧ ⎪ ⎨ 1, 0 ≤ t ≤ 1 − T, ∗ (c) u = ⎪ ⎩ 0, 1 − T < t ≤ T. (e) J = −(1/8 + 1/8K). (f) J = −1/8. Chapter 5

5.1 (a) u∗ (t) =

⎧ ⎪ ⎨ 5, t ≤ 1 + 6 ln 0.99 ≈ 0.94, ⎪ ⎩ 0, t > 0.094.

(b) λ2 (t)/λ1 (t) = e3(t

2 −4t+1)/12

, u∗ (t) =

⎧ ⎪ ⎪ ⎪ −5, 0 ≤ t ≤ 0.28, ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎨ 0, 0.28 < t ≤ 0.4, ⎪ ⎪ ⎪ 5, ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 0,

0.4 < t ≤ 0.93, 0.93 < t ≤ 1.0.

468

E. Answers to Selected Exercises ∗

5.4 (b) f (t∗ ) = t∗ − 10 ln(1 − 0.3e0.1t ). (c) t∗ = 1.969327, J(t∗ ) = 19.037. 5.8 5.10

u∗ = v ∗ = 0 for all t. u∗ = 0, v ∗ = 4/5 for t ∈ [0, 49], u∗ = 0, v ∗ = 0 for t ∈ [49, 60], J ∗ = 34,420. Chapter 6

6.4

Q(t) = t4 − 160t3 + 1740t2 − 7360t + 9639.

6.9

v ∗ = sat[−V2 , V1 ; (λ2 − λ1 p)2βλ1 ].

6.10

v ∗ (t) ≈ 3e−3t , y ∗ (t) ≈ 1 − 3e−3t .

6.12

J ∗ = 10.56653. ⎧ ⎪ ⎪ ⎪ 0, ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎨ 2, ∗ u (t) = ⎪ ⎪ ⎪ −1, ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 0,

6.14

6.15

μ1 =

μ2 =

7/3 < t < 3, 3 ≤ t < 13/3, 13/3 ≤ t ≤ 6.

⎧ ⎪ ⎨ − 5 t + 5 , t ∈ [0, 1], 2 2 ⎪ ⎩ 0,

t ∈ (0, 3].

⎧ ⎪ ⎨ 0,

t ∈ [0, 1.8),

⎪ ⎩ − 1 t + 3 , t ∈ [1.8, 3]. 2 2

⎧ ⎪ ⎨ 0, η=

0 ≤ t ≤ 7/3,

t ∈ [0, 1)

8

(1.8, 3],

⎪ ⎩ − 5 t + 5 , t ∈ [1, 1.8). 2 2

E. Answers to Selected Exercises

6.16 (a) v ∗ (t) =

⎧ ⎪ ⎨ −1 for t ∈ [0, 1.8), ⎪ ⎩ 1

(b) v ∗ (t) = 1 for ⎧ ⎪ ⎪ ⎪ ⎪ −1 ⎪ ⎪ ⎪ ⎪ ⎪ ⎨ 0 ∗ 6.18 v (t) = ⎪ ⎪ ⎪ +1 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 0

6.19

u∗ (t) =

469

for t ∈ (1.8, 3]. t ∈ [0, 10]. for t ∈ [0, 1/2], for t ∈ (1/2, 23/12], for t ∈ (23/12, 29/12], for t ∈ (29/12, 4].

⎧ ⎪ ⎨ 0,

for 0 ≤ t ≤ t1 ,

⎪ ⎩ h(t − t )/c, for t < t ≤ T, 1 1

where t1 = T −

#

2BC/h. Chapter 7

7.1

p∗ = 102.5 + 0.2G.

7.7

(¯ u)/(pS) = (δβ)/(η(ρ + δ)).

7.15

The reachable set is [x0 e−δT , (x0 − x ¯)e−(δ+rQ)T + x ¯], where x ¯ = rQ/(W + rQ).

7.20 (b) t1 = 7.21

1 x0 ln s , rQ + δ x

t2 =

1 x ¯ − xs . ln rQ + δ x ¯ − xT

1 rQ(1 − x0 ) − δx0 1 xs + ln . ln rQ + δ rQ(1 − xs ) − δxs δ xT & % 1−A . imp(A, B; t) = − 1r ln 1−B T ≥

7.28

7.29 (b) J = 0.6325. 7.35

The equations corresponding to (6.28) and (6.29) can be obtained by replacing ρ by ρ + r/r. ˙ The form of (6.30) remains unchanged.

470

E. Answers to Selected Exercises Chapter 8

8.1 (a) y = 1, z = 3. (b) y = 2, z = 10. 8.2 (a) (1,3) is a relative maximum. (b) (2,10) is a relative maximum. 8.3

x = 50; x = 80.

8.6 (a) x = 4 is a local maximum. (b) x = 8 is a local maximum and x = 20 is a local and a global maximum. 8.7 (a) (0, 0) is the nearest point. (b) (1/2, 1/2) is the nearest point. √ √ 8.8 (1/ 5, 2/ 5) is the closest point. √ 8.9 (a) (2 2, 0). (b) (0, 2). (c) (0, 2). λTi = ∂F/∂xTi for i = 1, 2, . . . , n; λTn+1 = 1. Note that here T denotes the terminal time, and not the transpose operation. ⎧ ⎪ ⎪ ⎪ +1 if λk+1 b > 1, ⎪ ⎪ ⎨ k T −k λT , 8.17 uk∗ = −1 if λk+1 b < −1 , where λ = (I + A) ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 0 if |λk+1 b| < 1. 8.13

E. Answers to Selected Exercises

471

Chapter 9 9.2

ts = 5.25, T = 11.

9.4

T = ts = 2.47.

9.5

ts = 0, T = 30.

9.7

$2 4 u∗ (t) = sat[0, 1; u0 (t)], where u0 (t) = 2 − e0.05(t−34.8) /(1 + t), t1 ≈ 3; t2 − T = 34.8. Chapter 10

10.4

x ¯ = 0.734.

10.5 (a) ⎡ ⎤ 7 c c 2 8cρ ⎦ ρ ρ X⎣ . + 1− + 1− + + x ¯= 4 r Xp r Xp prX (b) For ρ = 0, x ¯ = 220,000. For ρ = 0.1, x ¯ = 86,000. For ρ = ∞, x ¯ = 40,000. 10.7

[g (x) − ρ][p − c(x)] − c (x)g(x) = 0.

10.9

[g (x) − ρ][p − c(x)] − c (x)g(x) + p˙ = 0. Chapter 11

11.1

λ(t) = λ0 e(ρ−β)t , where λ0 =

¯ − eβT )/β − KT ](2ρ − β) [K0 eβT + C(1 , eβT − e2(β−ρ)T

K(t) = K0 eβt +

C¯ λ0 (1 − eβt ) − (e2(β−ρ)t − eβt ). β β − 2ρ

Chapter 12 ' α−r ∗ (x) = 1 , c 12.5 q ∗ (x) = (1−β)σ 2 1−β ρ − rβ − 1−β ]1−β xβ , x ≥ 0. V (x) = [ ρ−rβ−γβ/(1−β)

γβ 1−β

( x,

472

E. Answers to Selected Exercises Chapter 13

13.2

u∗ (t) = 1 + λL (t), v ∗ (t) = 1 + λF (t), where λL and λF are the solution of the linear diﬀerential equation ⎞

⎛ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎝

x˙ F λ˙ L λ˙

μ˙

⎛

⎞⎛

0 ⎟⎜ x ⎟ ⎜ 0 1 1 ⎟ ⎜ ⎟⎜ ⎟ ⎜ ⎟⎜ F ⎟ ⎜ 1 0 0 ⎜ 0 ⎟ ⎟ ⎜ ⎟⎜ λ ⎟=⎜ ⎟⎜ ⎟ ⎜ ⎟⎜ ⎟ ⎜ 1 0 0 −1 ⎟ ⎜ λL ⎟ ⎜ ⎟⎜ ⎠ ⎝ ⎠⎝ μ 0 0 −1 0

⎞

⎛

⎞

⎟ ⎜ 2 ⎟ ⎟ ⎜ ⎟ ⎟ ⎜ ⎟ ⎟ ⎜ 0 ⎟ ⎟ ⎜ ⎟ ⎟+⎜ ⎟ ⎟ ⎜ ⎟ ⎟ ⎜ 0 ⎟ ⎟ ⎜ ⎟ ⎠ ⎝ ⎠ 0

with the boundary conditions x(0) = x0 , λF (T ) = 0, λL (T ) = 0, and μ(0) = 0.

Bibliography Abad PL (1982a) An optimal control approach to marketing-production planning. Optimal Control Appl Methods 3(1):1–13 Abad PL (1982b) Approach to decentralized marketing-production planning. Int J Syst Sci 13(3):227–235 Abad PL (1987) A hierarchical optimal control model for coordination of functional decisions in a ﬁrm. Eur J Oper Res 32:62–75 Abad PL (1989) Multi-product multi-market model for coordination of marketing-production decisions. Int J Syst Sci 20:2011–2027 Abad PL, Sweeney DJ (1982) Decentralized planning with an interdependent marketing-production system. Omega 10:353–359 Agnew CE (1976) Dynamic modeling and control of congestion-prone systems. Oper Res 24:400–419 Ahmed NU, Teo KL (1981) Optimal control of distributed parameter systems. Elsevier Science LTD, Amsterdam Alam M, Sarma VVS (1974) Optimal maintenance policy for equipment subject to deterioration and random failure. IEEE Trans Syst Man Cybern SMC4:172–175 Alam M, Sarma VVS (1977) An application of optimal control theory to repairman problem with machine interference IEEE Trans Reliab R-26:121–124 Alam M, Lynn JW, Sarma VVS (1976) Optimal maintenance policy for equipment subject to random deterioration and random failure. Int J Syst Sci 7:1071–1080 Allen KR (1973) Analysis of the stock-recruitment relation in Antarctic ﬁn whales. In: Parrish B (ed) Fish stock and recruitment, J. du Conseil International pour l’Exploration de la Mer, Rapports et Proc´es-Verbaux de R´eunions, vol 164, pp 132–137 © Springer Nature Switzerland AG 2019 S. P. Sethi, Optimal Control Theory, https://doi.org/10.1007/978-3-319-98237-3

473

474

Bibliography

Amit R (1986) Petroleum reservoir exploitation: switching from primary to secondary recovery. Oper Res 34(4):534–549 Amit R, Ilan Y (1990) The choice of manufacturing technology in the presence of dynamic demand and experience eﬀects. IIE Trans 22(2):100–111 Anderson RM, May RM (1992) Infectious diseases of humans: dynamics and control. Oxford University Press, Oxford Anderson BDO, Moore JB (1979) Optimal ﬁltering. Prentice-Hall, New York Anderson BDO, Moore JB (1990) Optimal control. Linear quadratic methods. Prentice-Hall, Englewood Cliﬀs Aoki M (1976) Dynamic economics: a system theoretic approach to theory and control. Elsevier, New York Arnold L (1974) Stochastic diﬀerential equations: theory and applications. Wiley, New York Aronson JE, Thompson GL (1984) A survey on forward methods in mathematical programming. J Large Scale Syst 7:1–16 Arora SR, Lele PT (1970) A note on optimal maintenance policy and sale date of a machine. Manag Sci 17:170–173 Arrow KJ (1968) Applications of control theory to economic growth. In: Dantzig GB, Veinott AF (eds) Mathematics of the decision sciences. Lectures in applied mathematics, vol 12. American Mathematical Society, Providence, pp 85–119 Arrow KJ, Chang S (1980) Optimal pricing, use, and exploration of uncertain natural resource stocks. In: Liu PT (ed) Dynamic optimization in mathematical economics. Plenum Press, New York Arrow KJ, Kurz M (1970) Public investment, the rate of return, and optimal ﬁscal policy. The John Hopkins Press, Baltimore Arrow KJ, Bensoussan A, Feng Q, Sethi SP (2007) Optimal savings and the value of population. Proc Natl Acad Sci 104(47):18421–18426 Arrow KJ, Bensoussan A, Feng Q, Sethi SP (2010) The genuine savings criterion and the value of population in an economy with endogenous fertility rate. In: Boucekkine R, Hritonenko N, Yatsenko Y (eds) Optimal control of age-structured population in economy, demography, and the environment. Routledge explorations in environmental economics. Routledge, New York, pp 20–44

Bibliography

475

Arthur WB, McNicoll G (1977) Optimal time paths with age dependence: a theory of population policy. Rev Econ Stud 44:111–123 Arutyunov AV, Aseev SM (1997) Investigation of the degeneracy phenomenon of the maximum principle for optimal control problems with state constraints. SIAM J Control Optim 35(3):930–952 Aubin J-P, Cellina A (1984) Diﬀerential inclusions: set-valued maps and viability theory. Springer, Berlin Axs¨ ater S (1985) Control theory concepts in production and inventory control. Int J Syst Sci 16:161–169 Bagchi A (1984) Stackelberg diﬀerential games in economic models. Lecture notes in control and information sciences, vol 64. Springer, Berlin Basar T (1986) A tutorial on dynamic and diﬀerential games. In: Basar T (ed) Dynamic games and applications in economics. Springer, Berlin, pp 1–25 Basar T, Olsder GJ (1999) Dynamic noncooperative game theory, 2nd edn. Society for Industrial and Applied Mathematics, Philadelphia Basar T, Bensoussan A, Sethi SP (2010) Diﬀerential games with mixed leadership: the open-loop solution. Appl Math Comput 217:972–979 Bass FM (1969) A new product growth model for consumer durables. Manag Sci 15(5):215–227 Bass FM, Bultez AV (1969) A note on optimal strategic pricing of technological innovations. Market Sci 1:371–378 Bass FM, Krishnamoorthy A, Prasad A, Sethi SP (2005a) Advertising competition with market expansion for ﬁnite horizon ﬁrms. J Ind Manag Optim 1(1):1–19 Bass FM, Krishnamoorthy A, Prasad A, Sethi SP (2005b) Generic and brand advertising strategies in a dynamic duopoly. Market Sci 24(4):556–568 Bean JC, Smith RL (1984) Conditions for the existence of planning horizons. Math Oper Res 9(3):391–401 Behrens DA, Caulkins JP, Tragler G, Feichtinger G (2000) Optimal control of drug epidemics: prevent and treat - but not at the same time? Manag Sci 46:333–347 Behrens DA, Caulkins JP, Tragler G, Feichtinger G (2002). Why presentoriented societies undergo cycles of drug epidemics. J Econ Dyn Control 26:919–936

476

Bibliography

Bell DJ, Jacobson DH (1975) Singular optimal control. Academic Press, New York Bellman RE (1957) Dynamic programming. Princeton University Press, Princeton Bellman RE, Kalaba RE (1965a) Quasilinearization and boundary value problems. Elsevier, New York Bellman RE, Kalaba RE (1965b) Dynamic programming and modern control theory. Academic Press, New York Benchekroun H, Mart´ın-Herr´ an G, Taboubi S (2009) Could myopic pricing be a strategic choice in marketing channels? A game theoretic analysis. J Econ Dyn Control 33:1699–1718 Benkherouf L, Sethi SP (2010) Optimality of (s,S) policies for a stochastic inventory model with proportional and lump-sum shortage costs. Oper Res Lett 38:252–255 Bensoussan A (2004) Stochastic control of partially observable systems. Cambridge University Press, Cambridge Bensoussan A (2011) Dynamic programming and inventory control. IOS Press, Amsterdam Bensoussan A (2018) Estimation and control of dynamical systems. Springer, New York Bensoussan A, Lesourne J (1980) Optimal growth of a self-ﬁnancing ﬁrm in an uncertain environment. In: Bensoussan A et al (eds) Applied stochastic control in econometrics and management science. North-Holland, Amsterdam, pp 235–269 Bensoussan A, Lesourne J (1981) Growth of ﬁrms: a stochastic control theory approach. In: Brockhoﬀ K, Krelle W (eds) Unternehmensplanung. Springer, Berlin, pp 101–116 Bensoussan A, Lions JL (1975) Nouvelles m´ethodes en controle impulsionnel. Appl Math Optim 1:289–312 Bensoussan A, Lions JL (1982) Application of variational inequalities in stochastic control. North-Holland, Amsterdam Bensoussan A, Lions JL (1984) Impulse control and quasi-variational inequalities. Bordas, Paris Bensoussan A, Sethi SP (2007) The machine maintenance and sale age model of Kamien and Schwartz revisited. Manag Sci 53(12):1964–1976

Bibliography

477

Bensoussan A, Hurst EG Jr, N¨ aslund B (1974) Management applications of modern control theory. Elsevier, New York Bensoussan A, Nissen G, Tapiero CS (1975) Optimum inventory and product quality control with deterministic and stochastic deterioration - an application of distributed parameter control systems. IEEE Trans Autom Control AC-20:407–412 Bensoussan A, Bultez AV, Naert PA (1978) Leader’s dynamic marketing behavior in oligopoly. In: Bensoussan A et al (eds) TIMS studies in the management sciences, vol 9. North-Holland, Amsterdam, pp 123–145 Bensoussan A, Crouhy M, Proth J-M (1983) Mathematical theory of production planning. North-Holland, Amsterdam Bensoussan A, Sethi SP, Vickson RG, Derzko NA (1984) Stochastic production planning with production constraints. SIAM J Control Optim 22(6):920–935 Bensoussan A, Liu RH, Sethi SP (2005a) Optimality of an (s, S) policy with compound Poisson and diﬀusion demands: a quasi-variational inequalities approach. SIAM J Control Optim 44(5):1650–1676 Bensoussan A, C ¸ akanyildirim M, Sethi SP (2005b) On the optimal control of partially observed inventory systems. C R Acad Sci Paris Ser I 341:419–426 Bensoussan A, Da Prato G, Delfour MC, Mitter SK (2007a) Representation and control of inﬁnite dimensional systems. In: Basar T (ed) Systems & control: foundations & applications, 2nd edn. Birkh¨auser, Boston Bensoussan A, C ¸ akanyildirim M, Sethi SP (2007b) Optimal ordering policies for inventory problems with dynamic information delays. Prod Oper Manag 16(2):241–256 Bensoussan A, C ¸ akanyildirim M, Sethi SP (2007c) A multiperiod newsvendor problem with partially observed demand. Math Oper Res 32(2):322–344 Bensoussan A, C ¸ akanyildirim M, Sethi SP (2007d) Partially observed inventory systems: the case of zero-balance walk. SIAM J Control Optim 46(1):176–209 Bensoussan A, C ¸ akanyildirim M, Sethi SP (2007e) Economic evaluation of systems that expedite inventory information. Prod Oper Manag 16(3):360–368 Bensoussan A, C ¸ akanyildirim M, Minjarez-Sosa JA, Royal A, Sethi SP (2008a) Inventory problems with partially observed demands and lost sales. J Optim Theory Appl 136(3):321–340 Bensoussan A, C ¸ akanyildirim M, Minjarez-Sosa JA, Sethi SP, Shi R (2008b) Partially observed inventory systems: the case of rain checks. SIAM J Control Optim 47(5):2490–2519

478

Bibliography

Bensoussan A, C ¸ akanyildirim M, Feng Q, Sethi SP (2009a) Optimal ordering policies for stochastic inventory problems with observed information delays. Prod Oper Manag 18(5):546–559 Bensoussan A, C ¸ akanyildirim M, Royal A, Sethi SP (2009b) Bayesian and adaptive controls for a newsvendor facing exponential demand. Risk Decis Anal 1(4):197–210 Bensoussan A, C ¸ akanyildirim M, Sethi SP (2009c) A note on ‘the censored newsvendor and the optimal acquisition of information’. Oper Res 57(3):791– 794 Bensoussan A, Keppo J, Sethi SP (2009d) Optimal consumption and portfolio decisions with partially observable real prices. Math Financ 19(2):215–236 Bensoussan A, Sethi SP, Chutani A (2009e) Optimal cash management under uncertainty. Oper Res Lett 37:425–429 Bensoussan A, C ¸ akanyildirim M, Sethi SP (2010) Filtering for discrete-time Markov processes and applications to inventory control with incomplete information. In: Crisan D, Rozovsky B (eds) Handbook on nonlinear ﬁltering. Oxford University Press, Oxford, pp 500–525 Bensoussan A, C ¸ akanyildirim M, Sethi SP, Shi R (2011a) Computation of approximate optimal policies in partially observed inventory model with rain checks. Automatica 47:1589–1604 Bensoussan A, C ¸ akanyildirim M, Sethi SP, Wang M, Zhang H (2011b) Average cost optimality in inventory models with dynamic information delays. IEEE Trans Autom Control 56(12):2869–2882 Bensoussan A, Long H, Perera S, Sethi SP (2012) Impulse control with random reaction periods: a central bank intervention problem. Oper Res Lett 40:425– 430 Bensoussan A, Chen S, Sethi SP (2014) Feedback Stackelberg solutions of inﬁnite-horizon stochastic diﬀerential games. In: Ouardighi FE, Kogan K (eds) Models and methods in economics and management science, essays in honor of Charles S. Tapiero, Series 6161, vol 198. Springer International Publishing, Cham, pp 3–15 Bensoussan A, Chen S, Sethi SP (2015a) The maximum principle for global solutions of stochastic Stackelberg diﬀerential games. SIAM J Control Optim 53(4):1956–1981 Bensoussan A, Feng Q, Sethi SP (2015b) Integrating equipment investment strategy with maintenance operations under uncertain failures. Ann Oper Res 8:1–34

Bibliography

479

Bensoussan A, C ¸ akanyildirim M, Li M, Sethi SP (2016) Managing inventory with cash register information: sales recorded but not demands. Prod Oper Manag 25(1):9–21 Bensoussan A, Chen S, Chutani A, Sethi SP (2018) Feedback Stackelberg-Nash equilibria in mixed leadership games with an application to cooperative advertising. Operations research, Working Paper, University of Texas at Dallas Berkovitz LD (1961) Variational methods in problems of control and programming. J Math Anal Appl 3:145–169 Berkovitz LD (1994) A theory of diﬀerential games. In: Basar T, Haurie A (eds) Advances in dynamic games and applications. Birkh¨auser, Boston, pp 3–22 Berkovitz LD, Dreyfus SE (1965) The equivalence of some necessary conditions for optimal control in problems with bounded state variables. J Math Anal Appl 10:275–283 Bertsekas DP, Shreve SE (1996) Stochastic optimal control: the discrete-time case. Athena Scientiﬁc, New York Bes C, Sethi SP (1988) Concepts of forecast and decision horizons: applications to dynamic stochastic optimization problems. Math Oper Res 13(2):295–310 Bes C, Sethi SP (1989) Solution of a class of stochastic linear-convex control problems using deterministic equivalents. J Optim Theory Appl 62(1):17–27 Bettiol P, Vinter RB (2010) Sensitivity interpretations of the costate variable for optimal control problems with state constraints. SIAM J Control Optim 48(5):3297–3317 Beyer D, Sethi SP (1998) A proof of the EOQ formula using quasi-variational inequalities. Int J Syst Sci 29(11):1295–1299 Beyer D, Cheng F, Sethi SP, Taksar MI (2010) Markovian Demand Inventory Models. Springer, New York Bhar R (2010) Stochastic ﬁltering with applications in ﬁnance. World Scientiﬁc, Singapore Bhaskaran S, Sethi SP (1981) Planning horizons for the wheat trading model. In: Proceedings of AMS 81 conference, 5: life, men, and societies, pp 197–201 Bhaskaran S, Sethi SP (1983) Planning horizon research - the dynamic programming/control theory interface. In: Proceedings of international AMSE winter symposium, Bermuda, pp 155–160 Bhaskaran S, Sethi SP (1987) Decision and forecast horizons in a stochastic environment: a survey. Optimal Control Appl Methods 8:201–217

480

Bibliography

Bhaskaran S, Sethi SP (1988) The dynamic lot size model with stochastic demands: a planning horizon study. Inf Syst Oper Res 26(3):213–224 Black F, Scholes M (1973) The pricing of options and corporate liabilities. J Polit Econ 81:637–659 Blaqui`ere A (1979) Necessary and suﬃciency conditions for optimal strategies in impulsive control. In: Liu P-T, Roxin EO (eds) Diﬀerential games and ˙ control theory III, part A. M. Dekker, New York, pp 1–28 Blaqui`ere A (1985) Impulsive optimal control with ﬁnite or inﬁnite time horizon. J Optim Theory Appl 46:431–439 Boccia A, De Pinho MDR, Vinter RB (2016) Optimal control problems with mixed and pure state constraints. SIAM J Control Optim 54(6):3061–3083 Boiteux M (1955) R´eﬂexious sur la concurrence du rail et de la route, le d´eclassement des lignes non rentables et le d´eﬁcit du chemin de fer. ´ ´ L’Economie Electrique, vol 2 Bolton P, Dewatripont M (2005) Contract theory. MIT Press, Cambridge Boltyanskii VG (1971) Mathematical methods of optimal control. Holt, Rinehard & Winston, New York Bookbinder JH, Sethi SP (1980) The dynamic transportation problem: a survey. Naval Res Logist Quart 27:65–87 Bourguignon F, Sethi SP (1981) Dynamic optimal pricing and (possibly) advertising in the face of various kinds of potential entrants. J Econ Dyn Control 3:119–140 Bowes MD, Krutilla JV (1985) Multiple use management of public forestlands. In: Kneese AV, Sweeney JL (eds) Handbook of natural resource and energy economics, vol 2, 1st edn, chap 12. Elsevier, London, pp 531–569 Breakwell JV (1968) Stochastic optimization problems in space guidance. In: Karreman HF (ed) Stochastic optimization and control. Wiley, New York, pp 91–100 Brekke KA, Øksendal BK (1994) Optimal switching in an economic activity under uncertainty. SIAM J Control Optim 32(4):1021–1036 Breton M, Jarrar R, Zaccour G (2006) A note on feedback Stackelberg equilibria in a Lanchester model with empirical application. Manag Sci 52(5):804–811 Brito DL, Oakland WH (1977) Some properties of the optimal income tax. Int Econ Rev 18:407–423

Bibliography

481

Brokate M (1985) Pontryagin’s Principle for control problems in age-dependent population dynamics. J Math Biol 23(1):75–101 Brown RG (1959) Statistical forecasting for inventory control. McGraw-Hill Book, New York Brown RG (1963) Smoothing, forecasting and prediction. Prentice-Hall, Englewood Cliﬀs Bryant GF, Mayne DQ (1974) The maximum principle. Int J Control 20:1021– 1054 Bryson AE Jr (1998) Dynamic optimization. Addison-Wesley, Reading Bryson AE Jr, Ho Y-C (1975) Applied optimal control. Optimization, estimation and control. Taylor & Francis, New York Buchanan LF, Norton FE (1971) Optimal control applications in economic systems. In: Leondes CT (ed) Advances in control systems, vol 8. Academic Press, New York, pp 141–187 Bulirsch R, Kraft D (eds) (1994) Computational optimal control. Birkh¨ auserVerlag, Boston Bulirsch R, Oettli W, Stoer J (eds) (1975) Optimization and optimal control. Lecture notes in mathematics, vol 477. Springer, Berlin Bultez AV, Naert PA (1979) Does lag structure really matter in optimizing advertising spending. Manag Sci 25(5):454–465 Bultez AV, Naert PA (1988) When does lag structure really matter...indeed? Manag Sci 34(7):909–916 Burdet CA, Sethi SP (1976) On the maximum principle for a class of discrete dynamical system with lags. J Optim Theory Appl 19:445–454 Burmeister E, Dobell AR (1970) Mathematical theories of economic growth. MacMillan, London Butkowskiy AG (1969) Distributed control systems. Elsevier, New York Bylka S, Sethi SP (1992) Existence of solution and forecast horizons in dynamic lot size model with nondecreasing holding costs. Prod Oper Manag 1(2):212– 224 Bylka S, Sethi SP, Sorger G (1992) Minimal forecast horizons in equipment replacement models with multiple technologies and general switching costs. Naval Res Logist 39:487–507

482

Bibliography

Caines P, Sethi SP, Brotherton T (1977) Impulse response identiﬁcation and casualty detection for the Lydia Pinkham data. Ann Econ Soc Meas 6(2):147– 163 Canon MD, Cullum CD, Polak E (1970) Theory of optimal control and mathematical programming. McGraw-Hill, New York Caputo MR (2005) Foundations of dynamic economic analysis: optimal control theory and applications. Cambridge University Press, Cambridge Carlson DA (1986a) On the existence of catching up optimal solutions for Lagrange problems deﬁned on unbounded intervals. J Optim Theory Appl 49(2):207–225 Carlson DA (1986b) The existence of ﬁnitely optimal solutions for inﬁnite horizon optimal control problems. J Optim Theory Appl 51(1):41–62 Carlson DA (1987a) On the existence of sporadically catching up optimal solutions for inﬁnite horizon optimal control problems. J Optim Theory Appl 53(2):219–235 Carlson DA (1987b) An elementary proof of the maximum principle for optimal control problems governed by a Volterra integral equation. J Optim Theory Appl 54(1):43–61 Carlson DA (1988) Suﬃcient conditions for optimality and supported trajectories for optimal control problems governed by Volterra integral equations. In: Advances in optimization and control. Lecture notes in economics and mathematical systems, vol 302. Springer, New York, pp 274–282 Carlson DA (1989) Some concepts of optimality for inﬁnite horizon control and their interrelationships. In: Roxin EO (ed) Modern optimal control; a conference in honor of Solomon Lefschetz and Joseph P. LaSalle. Marcel Dekker, New York, pp 13–22 Carlson DA (1990a) Uniform overtaking and weakly overtaking optimal solutions in optimal control: when optimal solutions are agreeable. J Optim Theory Appl 64(1):55–69 Carlson DA (1990b) The existence of catching-up optimal solutions for a class of inﬁnite horizon optimal control problems with time delay. SIAM J Control Optim 28(2):402–422 Carlson DA (1990c) Inﬁnite horizon optimal controls for problems governed by a Volterra integral equation with a state and control dependent discount factor. J Optim Theory Appl 66(2):311–336

Bibliography

483

Carlson DA (1991) Asymptotic stability for optimal trajectories of inﬁnite horizon optimal control models with state and control dependent discounting. In: Proceedings of the 4th annual workshop in analysis and its applications, pp 281–299 Carlson DA (1993) Nonconvex and relaxed inﬁnite horizon optimal control problems. J Optim Theory Appl 78(3):465–491 Carlson DA (1995) An existence theorem for hereditary Lagrange problems on an unbounded interval. In: Boundary value problems for functional diﬀerential equation. World Scientiﬁc Publishing, Singapore, pp 73–83 Carlson DA (1997) Overtaking optimal solutions for convex Lagrange problems with time delay. J Math Appl 208:31–48 Carlson DA, Haurie A (1987a) Inﬁnite horizon optimal control theory and applications. Lecture Notes in Economics and Mathematical Systems, vol 290. Springer, New York Carlson DA, Haurie A (1987b) Optimization with unbounded time interval for a class of non linear systems. Springer, Berlin Carlson DA, Haurie A (1992) Control theoretic models of environment-economy interactions. In: The proceedings of the 31st IEEE conference on decision and control, pp 2860–2861 Carlson DA, Haurie A (1995) A turnpike theory for inﬁnite horizon open-loop diﬀerential games with decoupled dynamics, In: New trends in dynamic games and applications, annals of the international society of dynamic games, vol 3. Birkh¨ auser, Boston, pp 353–376 Carlson DA, Haurie A (1996) A turnpike theory for inﬁnite horizon competitive processes. SIAM J Optim Control 34(4):1405–1419 Carlson DA, Haurie A, Jabrane A (1987) Existence of overtaking solutions to inﬁnite dimensional control problems on unbounded time intervals. SIAM J Control Optim 25(6):1517–1541 Carlson DA, Haurie A, Leizarowitz A (1991) Inﬁnite horizon optimal control: deterministic and stochastic systems, 2nd edn. Springer, New York Carlson DA, Haurie A, Leizarowitz A (1994) Equilibrium points for linearquadratic inﬁnite horizon diﬀerential games. In: Basar T, Haurie A (eds) Advances in dynamic games and applications, annals of the international society of dynamic games, vol 1. Birkh¨auser, Boston, pp 247–268 Carraro C, Filar J (eds) (1995) Control and game theoretic models of the environment. Birkh¨ auser, Boston

484

Bibliography

Carrillo J, Gaimon C (2000) Improving manufacturing performance through process change and knowledge creation. Manag Sci 46(2):265–288 Case JH (1979) Economics and the competitive process. New York University Press, New York Cass D, Shell K (1976) The Hamiltonian approach to dynamic economics. Academic Press, New York Caulkins JP, Feichtinger G, Haunschmied JL, Tragler G (2006) Quality cycles and the strategic manipulation of value. Oper Res 54:666–677 Caulkins JP, Feichtinger G, Grass D, Tragler G, (2008) Optimizing counterterror operations: should one ﬁght with ‘ﬁre’ or ‘water’ ? Comput Oper Res 35:1874–1885 Caulkins JP, Feichtinger G, Grass D, Tragler G (2009) Optimal control of terrorism and global reputation: a case study with novel threshold behaviour. Oper Res Lett 37(6):387–391 Caulkins JP, Feichtinger G, Grass D, Hartl RF, Kort PM, Seidl A (2011) Optimal pricing of a conspicuous product in a recession that freezes capital markets. J Econ Dyn Control 35(1):163–174 Caulkins JP, Feichtinger G, Grass D, Hartl RF, Kort PM, Seidl A (2013a) When to make proprietary software open source. J Econ Dyn Control 37(6):1182– 1194 Caulkins JP, Feichtinger G, Grass D, Hartl RF, Kort PM, Novak AJ, Seidl A (2013b) Leading bureaucracies to the tipping point: an alternative model of multiple stable equilibrium levels of corruption. Eur J Oper Res 225(3):541– 546 Caulkins JP, Feichtinger G, Hartl RF, Kort PM, Novak AJ, Seidl A (2013c) Multiple equilibria and indiﬀerence-threshold points in a rational addiction model. Central Eur J Oper Res 21(3):507–522 Caulkins JP, Feichtinger G, Grass D, Hartl RF, Kort PM, Novak AJ, Seidl A, Wirl F (2014) A dynamic analysis of Schelling’s binary corruption model: a competitive equilibrium approach. J Optim Theory Appl 161(2):608–625 Caulkins JP, Feichtinger G, Grass D, Hartl RF, Kort PM, Seidl A (2015a) Skiba points in free end-time problems. J Econ Dyn Control 51:404–419 Caulkins JP, Feichtinger G, Seidl A, Grass D, Hartl RF, Kort PM (2015b) Capital stock management during a recession that freezes credit markets. J. Econ Behav Organ 116:1–14

Bibliography

485

Caulkins JP, Feichtinger G, Grass D, Hartl RF, Kort PM, Seidl A (2017) Interaction of pricing, advertising and experience quality: a dynamic analysis. Eur J Oper Res 256(3):877–885 Cernea A, Frankowska H (2005) A connection between the maximum principle and dynamic programming for constrained control problems. SIAM J Control Optim 44:673–703 Cesari L (1983) Optimization - theory and applications: problems with ordinary diﬀerential equations. Springer, New York Chahim M, Hartl RF, Kort PM (2012) A tutorial on the deterministic impulse control maximum principle: necessary and suﬃcient optimality conditions. Eur J Oper Res 219(1):18–26 Chand S, Sethi SP (1982) Planning horizon procedures for machine replacement models with several possible replacement alternatives. Naval Res Logist Quart 29(3):483–493 Chand S, Sethi SP (1983) Finite production rate inventory models with ﬁrst and second shift setups. Naval Res Logist Quart 30:401–414 Chand S, Sethi SP (1990) A dynamic lot size model with learning in setups. Oper Res 38(4):644–655 Chand S, Sethi SP, Proth JM (1990) Existence of forecast horizons in undiscounted discrete time lot size models. Oper Res 38(5):884–892 Chand S, Sethi SP, Sorger G (1992) Forecast horizons in the discounted dynamic lot size model. Manag Sci 38(7):1034–1048 Chand S, Moskowitz H, Novak AJ, Rekhi I, Sorger G (1996) Capacity allocation for dynamic process improvement with quality and demand considerations. Oper Res 44:964–975 Chand S, Hsu VN, Sethi SP (2002) Forecast, solution and rolling horizons in operations management problems: a classiﬁed bibliography. Manufact Service Oper Manag 4(1):25–43 Chao R, Kavadias S, Gaimon C (2009) Revenue driven resource allocation and eﬀective NPD portfolio management. Manag Sci 55(9):1556–1569 Chappell D, Dury K (1994) On the optimal depletion of a nonrenewable natural resource under conditions of increasing marginal extraction costs. SIAM Rev 36(1):102–106 Charnes A, Kortanek K (1966) A note on the discrete maximum principle and distribution problems. J Math Phys 45:121–126

486

Bibliography

Chen SF, Leitmann G (1980) Labour-management bargaining modelled as a dynamic game. Optimal Control Appl Methods 1:11–25 Chiarella C, Kemp MC, Long NV, Okuguchi K (1984) On the economics of international ﬁsheries. Int Econ Rev 25:85–92 Chichilinsky G (1981) Existence and characterization of optimal growth paths including models with non-convexities in utilities and technologies. Rev Econ Stud 48:51–61 Chintagunta PK (1993) Investigating the sensitivity of equilibrium proﬁts to advertising dynamics and competitive eﬀects. Manag Sci 39(9):1146–1162 Chintagunta PK, Jain D (1992) A dynamic model of channel member strategies for marketing expenditures. Market Sci 11(2):168–188 Chintagunta PK, Jain D (1994) A study of manufacturer-retailer marketing strategies: a diﬀerential game approach. Lecture notes in control and information sciences. Springer, New York Chintagunta PK, Jain D (1995) Dynamic duopoly models of advertising competition: estimation and a speciﬁcation tests. J Econ Manag Strateg 4(1):109– 131 Chintagunta PK, Vilcassim NJ (1992) An empirical investigation of advertising strategies in a dynamic duopoly. Manag Sci 38(9):1230–1244 Chintagunta PK, Vilcassim NJ (1994) Marketing investment decisions in a dynamic duopoly: a model and empirical analysis. Int J Res Market 11(3):287– 306 Chow GC (1975) Analysis and control of dynamic economic systems. Wiley, New York Chutani A, Sethi SP (2012a) Cooperative advertising in a dynamic retail market oligopoly. Dyn Games Appl 2(4):347–375 Chutani A, Sethi SP (2012b) Optimal advertising and pricing in a dynamic durable goods supply chain. J Optim Theory Appl 154(2):615–643 Clark CW (1973) The economics of overexploitation. Science 181:630–634 Clark CW (1976) Mathematical bioeconomics: the optimal management of renewal resources. Wiley, New York Clark CW (1979) Mathematical models in the economics of renewable resources. SIAM Rev 21:81–99

Bibliography

487

Clark CW (1985) Bioeconomic modelling and ﬁsheries management. Wiley, New York Clarke FH (1983) Optimization and nonsmooth analysis. Wiley, New York Clarke FH (1989) Methods of dynamic and nonsmooth optimization. Society for Industrial and Applied Mathematics, Philadelphia Clarke FH, Darrough MN, Heineke JM (1982) Optimal pricing policy in the presence of experience eﬀects. J Bus 55:517–530 Clemhout S, Wan HY Jr (1985) Dynamic common property resources and environmental problems. J Optim Theory Appl 46:471–481 Coddington EA, Levinson NL (1955) Theory of ordinary diﬀerential equations. McGraw-Hill, New York Cohen KJ, Cyert RM (1965) Theory of the ﬁrm: resource allocation in a market economy. Prentice-Hall, Englewood Cliﬀs Connors MM, Teichroew D (1967) Optimal control of dynamic operations research models. International Textbook, Scranton Conrad K (1982) Advertising, quality and informationally consistent prices. Z Staatswiss 138:680–694 Conrad K (1985) Quality, advertising and the formation of goodwill under dynamic conditions. In: Feichtinger G (ed) Optimal control theory and economic analysis, vol 2. North-Holland, Amsterdam, pp 215–234 Constantinides GM, Richard SF (1978) Existence of optimal simple policies for discounted cost inventory and cash management in continuous time. Oper Res 26:620–636 Cvitanic J, Zhang J (2013) Contract theory in continuous-time models. Springer, New York Dantzig GB (1966) Linear control processes and mathematical programming. SIAM J Control 4:56–60 Dantzig GB, Sethi SP (1981) Linear optimal control problems and generalized linear programs. J Oper Res Soc 32:467–476 Dasgupta P, Heal GM (1974a) The optimal depletion of exhaustible resources. Rev Econ Stud 41:3–28 Dasgupta P, Heal GM (eds) (1974b) Symposium on the economics of exhaustible resources. The review of economic studies, vol 41. Oxford University Press, Oxford

488

Bibliography

D’Autume A, Michel P (1985) Future investment constraints reduce present investment. Econometrica 53:203–206 Davis BE (1970) Investment and rate of return for the regulated ﬁrm. Bell J Econ Manag Sci 1:245–270 Davis MHA (1993) Markov models and optimization. Chapman & Hall, New York Davis BE, Elzinga DJ (1971) The solution of an optimal control problem in ﬁnancial modeling. Oper Res 19:1419–1433 Dawid H, Feichtinger G (1996a) Optimal allocation of drug control eﬀorts: a diﬀerential game analysis. J Optim Theory Appl 91:279–297 Dawid H, Feichtinger G (1996b) On the persistence of corruption. J Econ 64:177–193 Dawid H, Kopel M, Feichtinger G (1997) Complex solutions of nonconcave dynamic optimization models. Econ Theory 9:427–439 Dawid H, Feichtinger, G, Goldstein JR, Veliov VM (2009) Keeping a learned society young. Demogr Res 20(22), 541–558 Deal KR (1979) Optimizing advertising expenditures in a dynamic duopoly. Oper Res 27(4):682–692 Deal KR, Sethi SP, Thompson GL (1979) A bilinear-quadratic diﬀerential game in advertising. In: Liu PT, Sutinen JG (eds) Control theory in mathematical ˙ economics. Marcel Dekker, New York, pp 91–109 Dechert DW, Nishimura K (1983) A complete characterization of optimal growth paths in an aggregated model with a non-concave production function. J Econ Theory 31:332–354 Deger S, Sen SK (1984) Optimal control and diﬀerential game models of military expenditure in less developed countries. J Econ Dyn Control 7:153–169 Deissenberg C (1980) Optimal control of linear econometric models with intermittent control. Econ Plan 16(1):49–56 Deissenberg C (1981) A simple model of closed-loop optimal exploration for oil. Policy Anal Inf Syst 5(3):167–183 Deissenberg C, St¨ oppler S (1982) Optimal control of LQG systems with costly observations. In: Feichtinger G (ed) Economic applications of optimal control theory. North-Holland, Amsterdam, pp 301–320

Bibliography

489

Deissenberg C, St¨ oppler S (1983) Optimal information gathering and planning policies of the proﬁt-maximizing ﬁrm. Int J Policy Inf 7(2):49–76 Deissenberg C, Feichtinger G, Semmler W, Wirl F (2004) Multiple equilibria, history dependence, and global dynamics in intertemporal optimisation models. In: Barnett WA, Deissenberg C, Feichtinger G (eds) Economic complexity: non-linear dynamics, multi-agents economies, and learning. Elsevier, Amsterdam, pp 91–122 Derzko NA, Sethi SP (1981a) Optimal exploration and consumption of a natural resource: stochastic case. Int J Policy Anal 5(3):185–200 Derzko NA, Sethi SP (1981b) Optimal exploration and consumption of a natural resource: deterministic case. Optim Control Appl Methods 2(1):1–21 Derzko NA, Sethi SP, Thompson GL (1980) Distributed parameter systems approach to the optimal cattle ranching problem. Optim Control Appl Methods 1:3–10 Derzko NA, Sethi SP, Thompson GL (1984) Necessary and suﬃcient conditions for optimal control of quasilinear partial diﬀerential systems. J Optim Theory Appl 43:9–101 Dhrymes PJ (1962) On optimal advertising capital and research expenditures under dynamic conditions. Economica 39:275–279 Dixit AK, Pindyck RS (1994) Investment under uncertainty. Princeton University Press, Princeton Dmitruk AV (2009) On the development of Pontryagin’s maximum principle in the works of A.Ya. Dubovitskii and A.A. Milyutin. Control Cybern 38(4):923–958 Dockner EJ (1984) Optimal pricing of a monopoly against a competitive producer. Optimal Control Appl Methods 5:345–351 Dockner EJ, Feichtinger G (1991) On the optimality of limit cycles in dynamic economic systems. J Econ 53:31–50 Dockner EJ, Feichtinger G (1993) Cyclical consumption patterns and rational addiction. Am Econ Rev 83:256–263 Dockner EJ, Jørgensen S (1984) Cooperative and non-cooperative diﬀerential game solutions to an investment and pricing problem. J Oper Res Soc 35:731– 739 Dockner EJ, Jørgensen S (1986) Dynamic advertising and pricing in an oligopoly: a Nash equilibrium approach. In: Basar T (ed) Dynamic games and applications in economics. Springer, Berlin, pp 238–251

490

Bibliography

Dockner EJ, Jørgensen S (1988) Optimal advertising policies for diﬀusion models of new product innovation in monopolistic situations. Manag Sci 34(1):119–130 Dockner EJ, Jørgensen S (1992) New product advertising in dynamic oligopolies. Z Oper Res 36(5):459–473 Dockner EJ, Sorger G (1996) Existence and properties of equilibria for a dynamic game on productive assets. J Econ Theory 71:209–227 Dockner EJ, Feichtinger G, Jørgensen S (1985) Tractable classes of nonzero-sum open-loop Nash diﬀerential games: theory and examples. J Optim Theory Appl 45:179–197 Dockner EJ, Feichtinger G, Mehlmann A (1993) Dynamic R & D competition with memory. J Evol Econ 3:145–152 Dockner EJ, Long NV, Sorger G (1996) Analysis of Nash equilibria in a class of capital accumulation games. J Econ Dyn Control 20:1209–1235 Dockner EJ, Jørgensen S, Long NV, Sorger G (2000) Diﬀerential games in economics and management science. Cambridge University Press, Cambridge Dogramaci A (2005) Hibernation durations for chain of machines with maintenance under uncertainty. In: Deissenberg C, Hartl R (eds) Optimal control and dynamic games: applications in ﬁnance, management science and economics, chap 14. Springer, New York, pp 231–238 Dogramaci A, Fraiman NM (2004) Replacement decisions with maintenance under uncertainty: an imbedded optimal control model. Oper Res 52(5):785– 794 Dogramaci A, Sethi SP (2016) Organizational nimbleness and operational policies: the case of optimal control of maintenance under uncertainty. In: Semmler W, Mittnik S (eds) Dynamic perspectives on managerial decision making, essays in honor of Richard F. Hartl. Dynamic modeling and econometrics in economics and ﬁnance, vol 22. Springer, Cham, pp 253–277 Dohrmann CR, Robinett RD (1999) Dynamic programming method for constrained discrete-time optimal control. J Optim Theory Appl 101(2):259–283 Dolan RJ, Jeuland AP (1981) Experience curves and dynamic demand models: implications of optimal pricing strategies. J Market 45:52–73 Dolan RJ, Muller E (1986) Models of new product diﬀusion: extension to competition against existing and potential ﬁrms over time. In: Mahajan V, Wind Y (eds) Innovation diﬀusion models of new product acceptance. Ballinger, Cambridge, pp 117–150

Bibliography

491

Dorfman R (1969) Economic interpretation of optimal control theory. Am Econ Rev 49:817–831 Dorfman R, Steiner PO (1954) Optimal advertising and optimal quality. Am Econ Rev 44:826–836 Drews W, Hartberger RJ, Segers R (1974) On continuous mathematical programming. In: Cottle RW, Krarup J (eds) Optimization methods for resource allocation. The English University Press, London Dubovitskii AJ, Milyutin AA (1965) Extremum problems in the presence of restrictions. USSR Comput Math Math Phys 5(3):1–80 Dunn JC, Bertsekas DP (1989) Eﬃcient dynamic programming implementations of Newton’s method for unconstrained optimal control problems. J Optim Theory Appl 63(1):23–38 Durrett R (1996) Stochastic calculus: a practical introduction, 2nd edn. CRC Press, Boca Raton El-Hodiri M, Takayama A (1981) Dynamic behavior of the ﬁrm with adjustment costs under regularity constraint. J Econ Dyn Control 3:29–41 Eliashberg J, Steinberg R (1987) Marketing-production decisions in an industrial channel of distribution. Manag Sci 33(8):981–1000 Elliott RJ, Aggoun L, Moore JB (1995) Hidden Markov models: estimation and control. Springer, New York El Ouardighi F, Pasin F (2006) Quality improvement and goodwill accumulation in a dynamic duopoly. Eur J Oper Res 175(2):1021–1032 El Ouardighi F, Tapiero CS (1998) Quality and the diﬀusion of innovations. Eur J Oper Res 106(1):31–38 El Ouardighi F, Jørgensen S, Pasin F (2008) A dynamic game of operations and marketing management in a supply chain. In: Petrosjan L, Yeung DK (eds) International game theory review (special issue dedicated to John F. Nash), vol 10. World Scientiﬁc, Singapore El Ouardighi F, Feichtinger G, Grass D, Hartl RF, Kort P (2016a) Autonomous and advertising-dependent ’word of mouth’ under costly dynamic pricing. Eur J Oper Res 251:860–872 El Ouardighi F, Feichtinger G, Grass D, Hartl RF, Kort P (2016b) Advertising and quality-dependent word-of-mouth in a contagion sales model. J Optim Theory Appl 170(1):323–342

492

Bibliography

El Ouardighi F, Feichtinger G, Fruchter G (2017) Accelerating the diﬀusion of innovations under mixed word of mouth through marketing-operations interaction. Ann Oper Res 264(1–2):435–458 Elton E, Gruber M (1975) Finance as a dynamic process. Prentice-Hall, Englewood Cliﬀs Erickson GM (1992) Empirical analysis of closed-loop duopoly advertising strategies. Manag Sci 38:1732–1749 Erickson GM (2003) Dynamic models of advertising competition. Springer, Boston Fan LT, Wang CS (1964) An application of the discrete maximum principle to a transportation problem. J Math Phys 43:255–260 Fattorini HO (1999) Inﬁnite-dimensional optimization and control theory. Encyclopedia of mathematics and its applications, vol 62. Cambridge University Press, Cambridge Feenstra TL, Kort PM, de Zeeuw AJ (2001) Environmental policy in an international duopoly: an analysis of feedback investment strategies. J Econ Dyn Control 25(10):1665–1687 Feichtinger G (1982a) Optimal pricing in a diﬀusion model with concave pricedependent market potential. O R Lett 1(6):236–240 Feichtinger G (1982b) Saddle-point analysis in a price-advertising model. J Econ Dyn Control 4:319–340 Feichtinger G (1982c) Optimal repair policy for a machine service problem. Optim. Control Appl Methods 3:15–22 Feichtinger G (1982d) The Nash solution of a maintenance-production diﬀerential game. Eur J Oper Res 10:165–172 Feichtinger G (ed) (1982e) Optimal control theory and economic analysis. In: First Viennese workshop on economic applications of control theory, Vienna, 28–30 October 1981. North-Holland, Amsterdam Feichtinger G (1982f) Anwendungen des maximumprinzips im operations research, Teil 1 und 2. OR-Spektr 4:171–190 und 195–212 Feichtinger G (1983a) The Nash solution of an advertising diﬀerential game: generalization of a model by Leitmann and Schmitendorf. IEEE Trans Autom Control AC-28:1044–1048 Feichtinger G (1983b) A diﬀerential games solution to a model of competition between a thief and the police. Manag Sci 29:686–699

Bibliography

493

Feichtinger G (1984a) Optimal employment strategies of proﬁt-maximizing and labour-managed ﬁrms. Optim Control Appl Methods 5:235–253 Feichtinger G (1984b) On the synergistic inﬂuence of two control variables on the state of nonlinear optimal control models. J Oper Res Soc 35:907–914 Feichtinger G (ed) (1985a) Optimal control theory and economic analysis 2. In: Second Viennese workshop on economic applications of control theory, Vienna, 16–18 May 1984. North-Holland, Amsterdam Feichtinger G (1985b) Optimal modiﬁcation of machine reliability by maintenance and production. OR-Spektr 7:43–50 Feichtinger G (1987) Intertemporal optimization of wine consumption at a party: an unusual optimal control model. In: Gandolfo G, Marzano F (eds) Keynesian theory planning models and quantitative economics, essays in memory of Vittorio Marrama, vol II. Giuﬀre, Milano, pp 777–797 Feichtinger G (ed) (1988) Optimal control theory and economic analysis, vol 3. North-Holland, Amsterdam Feichtinger G (1992a) Limit cycles in dynamic economic systems. Ann Oper Res 37:313–344 Feichtinger G (1992b) Optimal control of economic systems. In: Tzafestas SG (ed) Automatic control handbook. M. Dekker, New York, pp 1023–1044 Feichtinger G, Dockner EJ (1984) A note to Jørgensen’s logarithmic advertising diﬀerential game. Z Oper Res 28:B133–B153 Feichtinger G, Dockner EJ (1985) Optimal pricing in a duopoly: a noncooperative diﬀerential games solution. J Optim Theory Appl 45:199–218 Feichtinger G, Hartl RF (1985a) Optimal pricing and production in an inventory model. Eur J Oper Res 19:45–56 Feichtinger G, Hartl RF (1985b) On the use of Hamiltonian and maximized Hamiltonian in non-diﬀerentiable control theory. J Optim Theory Appl 46:493–504 ¨ Feichtinger G, Hartl RF (1986) Optimale Kontrolle Okonomischer Prozesse: Anwendungen des Maximumprinzips in den Wirtschaftswissenschaften. Walter De Gruyter, Berlin Feichtinger G, Hartl RF (eds) (1992) Nonlinear methods in economic dynamics and optimal control. Annals of operations research, vol 37. Baltzer, Basel Feichtinger G, Jørgensen S (1983) Diﬀerential game models in management science. Eur J Oper Res 14:137–155

494

Bibliography

Feichtinger G, Mehlmann A (1986) Planning the unusual: applications of control theory to non-standard problems. Acta Appl Math 7:79–102 Feichtinger G, Novak AJ (1992a) Optimal consumption, training, working time and leisure over the life cycle. J Optim Theory Appl 75:369–388 Feichtinger G, Novak AJ (1992b) A note on the optimal exploitation of migratory ﬁsh stocks. Dyn Control 2:255–263 Feichtinger G, Novak AJ (1994a) Optimal pulsing in an advertising diﬀusion model. Optim Control Appl Methods 15:267–276 Feichtinger G, Novak AJ (1994b) Diﬀerential game model of the dynastic cycle: 3-D canonical system with a stable limit cycle. J Optim Theory Appl 80(3):407–423 Feichtinger G, Novak AJ (2008) Terror and counterterror operations: diﬀerential game with cyclical Nash solution. J Optim Theory Appl 139:541–556 Feichtinger G, Sorger G (1986) Optimal oscillations in control models: how can constant demand lead to cyclical production? Oper Res Lett 5:277–281 Feichtinger G, Sorger G (1988) Periodic research and development. In: Feichtinger G (ed) Optimal control theory and economic analysis. Third Viennese workshop on optimal control theory and economic analysis, Vienna, 20–22 May 1987, vol 3. North-Holland, Amsterdam, pp 121–141 Feichtinger G, Steindl A (2006) DNS curves in a production/inventory model. J Optim Theory Appl 128:295–308 Feichtinger G, Veliov VM (2007) On a distributed control problem arising in dynamic optimization of a ﬁxed-size population. SIAM J Optim 18:980–1003 Feichtinger G, Wirl F (1993) A dynamic variant of the battle of the sexes. Int J Game Theory 22:359–380 Feichtinger G, Wirl F (1994) On the stability and potential cyclicity of corruption in governments subject to popularity constraints. Math Soc Sci 28:113– 131 Feichtinger G, Wirl F (2000) Instabilities in concave, dynamic, economic optimization. J Optim Theory Appl 107:277–288 Feichtinger G, Luhmer A, Sorger G (1988) Optimal price and advertising policy in convenience goods retailing. Market Sci 7:187–201 Feichtinger G, Kaitala VT, Novak AJ (1992) Stable resource-employment limit cycles in an optimally regulated ﬁshery. In: Feichtinger G (ed) Dynamic economic models and optimal control. North-Holland, Amsterdam, pp 163– 184

Bibliography

495

Feichtinger G, Hartl RF, Sethi SP (1994a) Dynamic optimal control models in advertising: recent developments. Manag Sci 40(2):195–226 Feichtinger G, Novak AJ, Wirl F (1994b) Limit cycles in intertemporal adjustment models - theory and applications. J Econ Dyn Control 18:353–380 Feichtinger G, Hartl RF, Haunschmied JL, Kort PM (1998) Optimal enforcement policies (crackdowns) on a drug market. Optim Control Appl Methods 19:169–184 Feichtinger G, Jørgensen S, Novak AJ (1999) Petrarch’s Canzoniere: rational addiction and amorous cycles. J Math Sociol 23(3):225–240 Feichtinger G, Hartl RF, Kort PM, Novak AJ (2001) Terrorism control in the tourism industry. J Optim Theory Appl 108:283–296 Feichtinger G, Grienauer W, Tragler G (2002) Optimal dynamic law enforcement. Eur J Oper Res 141:58–69 Feichtinger G, Tragler G, Veliov VM (2003) Optimality conditions for agestructured control systems. J Math Anal Appl 288:47–68 Feichtinger G, Prskawetz A, Veliov VM (2004a) Age-structured optimal control in population economics. Theor Popul Biol 65:373–387 Feichtinger G, Tsachev T, Veliov VM (2004b) Maximum principle for age and duration structured systems: a tool for optimal prevention and treatment of HIV. Math Popul Stud 11:3–28 Feichtinger G, Hartl RF, Kort PM, Veliov VM (2005) Environmental policy, the Porter - hypothesis and the composition of capital: eﬀects of learning and technological progress. J Environ Econo Manag 50:434–446 Feichtinger G, Hartl RF, Kort PM, Veliov VM (2006a) Capital accumulation under technological progress and learning: a vintage capital approach. Eur J Oper Res 172:293–310 Feichtinger G, Hartl RF, Kort PM, Veliov VM (2006b) Anticipation eﬀects of technological progress on capital accumulation: a vintage capital approach. J Econ Theory 126:143–164 Feichtinger G, Hartl RF, Kort PM, Veliov VM (2008) Financially constrained capital investments: the eﬀects of disembodied and embodied technological progress. J Math Econ 44:459–483 Feinberg FM (1992) Pulsing policies for aggregate advertising models. Market Sci 11(3):221–234

496

Bibliography

Feinberg FM (2001) On continuous-time optimal advertising under S-shaped response. Manag Sci 47(11):1476–1487 Fel’dbaum AA (1965) Optimal control systems. Academic Press, New York Ferreira MMA, Vinter RB (1994) When is the maximum principle for state constrained problems nondegenerate? J Math Anal Appl 187:438–467 Ferreyra G (1990) The optimal control problem for the Vidale-Wolfe advertising model revisited. Optim Control Appl Methods 11:363–368 Filipiak J (1982) Optimal control of store-and-forward networks. Optim Control Appl Methods 3:155–176 Fischer T (1985) Hierarchical optimization methods for the coordination of decentralized management planning. In: Feichtinger G (ed) Optimal control theory and economic analysis, vol 2. North-Holland, Amsterdam Fleming WH, Rishel RW (1975) Deterministic and stochastic optimal control. Springer, New York Fleming WH, Soner HM (1992) Controlled Markov processes and viscosity solutions. Springer, New York Fletcher R, Reeves CM (1964) Function minimization by conjugate gradients. Comput J 7:149–154 Fond S (1979) A dynamic programming approach to the maximum principle of distributed-parameter systems. J Optim Theory Appl 27(4):583–601 Forster BA (1973) Optimal consumption planning in a polluted environment Econ Rec 49:534–545 Forster BA (1977) On a one state variable optimal control problem: consumption-pollution trade-oﬀs. In: Pitchford JD, Turnovsky SJ (eds) Applications of control theory to economic analysis. North-Holland, Amsterdam Fourgeaud C, Lenclud B, Michel P (1982) Technological renewal of natural resource stocks. J Econ Dyn Control 4:1–36 Francis PJ (1997) Dynamic epidemiology and the market for vaccinations. J. Public Econ 63(3):383–406 Frankena JF (1975) Optimal control problems with delay, the maximum principle and necessary conditions. J Eng Math 9:53–64 Friedman A (1964) Optimal control for hereditary processes. Arch Ration Mech Anal 15:396–416

Bibliography

497

Friedman A (1971) Diﬀerential games. Wiley, New York Friedman A (1977) Oligopoly and the theory of games. North-Holland, Amsterdam Friedman A (1986) Game theory with applications to economics. Oxford University Press, New York Fruchter GE (1999a) The many-player advertising game. Manag Sci 45(11):1609–1611 Fruchter GE (1999b) Oligopoly advertising strategies with market expansion. Optim Control Appl Methods 20:199–211 Fruchter GE (2005) Two-part tariﬀ pricing in a dynamic environment. In: Deissenberg C, Hartl RF (eds) Optimal control and dynamic games. Springer, Dordrecht, pp 141–153 Fruchter GE, Kalish S (1997) Closed-loop advertising strategies in a duopoly. Manag Sci 43:54–63 Fuller D, Vickson RG (1987) The optimal construction of new plants for oil from the Alberta tar sands. Oper Res 35(5):704–715 Funke UH (1976) Mathematical models in marketing: a collection of abstracts. Lecture notes in economics and mathematical systems, vol 132. Springer, Berlin Fursikov AV (1999) Optimal control of distributed systems. Theory and applications (Translations of mathematical monographs). American Mathematical Society, Providence Gaimon C (1985a) The acquisition of automation subject to diminishing returns. IIE Trans 17:147–156 Gaimon C (1985b) The optimal acquisition of automation to enhance the productivity of labor. Manag Sci 31:1175–1190 Gaimon C (1986a) An impulsive control approach to deriving the optimal dynamic mix of manual and automatic output. Eur J Oper Res 24:360–368 Gaimon C (1986b) The optimal times and levels of impulsive acquisition of automation. Optim Control Appl Methods 7:259–270 Gaimon C (1986c) The optimal acquisition of new technology and its impact on dynamic pricing policies. In: Lev B (ed) Production management: methods and studies. Studies in management science and systems, vol 13. Elsevier, Amsterdam, pp 187–206

498

Bibliography

Gaimon C (1988) Simultaneous and dynamic price, production, inventory, and capacity decisions. Eur J Oper Res 35:426–441 Gaimon C (1989) Dynamic game results on the acquisition of new technology. Oper Res 37(3):410–425 Gaimon C (1994) Subcontracting versus capacity expansion and the impact on the pricing of services. Naval Res Logist 41(7):875–892 Gaimon C (1997) Planning information technology - knowledge worker systems. Manag Sci 43(9):1308–1328 Gaimon C, Burgess R (2003) Analysis of lead time and learning for capacity expansions. Prod Oper Manag 12(1):128–140 Gaimon C, Morton A (2005) Investment in changeover ﬂexibility for early entry in high tech markets. Prod Oper Manag (Special Issue on High Tech Production and Operations Management) 14(2):159–174 Gaimon C, Singhal V (1992) Flexibility and the choice of facilities under short product life cycles. Eur J Oper Res 60(2):211–223 Gaimon C, Thompson GL (1984a) Optimal preventive and repair maintenance of a machine subject to failure. Optim Control Appl Methods 5:57–67 Gaimon C, Thompson GL (1984b) A distributed parameter cohort personnel planning model using cross-sectional data. Manag Sci 30:750–764 Gaimon C, Thompson GL (1989) Optimal preventive and repair maintenance of a machine subject to failure and a declining resale value. Optim Control Appl Methods 10:211–228 Gamkrelidze RV (1965) On some extrema1 problems in the theory of diﬀerential equations with applications to the theory of optimal control. SIAM J Control Optim 3:106–128 Gamkrelidze RV (1978) Principle of optimal control theory. Plenum Press, New York Gandolfo G (1980) Economic dynamics: methods and models. North-Holland, Amsterdam Gaskins, DW Jr (1971) Dynamic limit pricing: optimal pricing under threat of entry. J Econ Theory 3:306–322 Gaugusch J (1984) The non-cooperative solution of a diﬀerential game: advertising versus pricing. Optim Control Appl Methods 5(4):353–360

Bibliography

499

Gavrila C, Caulkins JP, Feichtinger G, Tragler G, Hartl RF (2005) Managing the reputation of an award to motivate performance. Math Methods Oper Res 61:1–22 Gelfand IM, Fomin SV (1963) Calculus of variations. Prentice-Hall, Englewood Cliﬀs Gerchak Y, Parlar M (1985) Optimal control analysis of a simple criminal prosecution model. Optim Control Appl Methods 6:305–312 Gfrerer H (1984) Optimization of hydro energy storage plant problems by variational methods. Z Oper Res 28:B87–B101 Gihman II, Skorohod AV (1972) Stochastic diﬀerential equations. Springer, New York Girsanov IV (1972) Lectures on mathematical theory of extremum problems. Springer, Berlin Glad ST (1979) A combination of penalty function and multiplier methods for solving optimal control problems. J Optim Theory Appl 28:303–329 Goh BS (1980) Management and analysis of biological populations. Elsevier, Amsterdam Goh BS, Leitmann G, Vincent TL (1974) Optimal control of a prey-predator system. Math Biosci 19:263–286 Goldberg S (1986) Introduction to diﬀerence equations. Dover Publications, New York Goldstine HH (1980) A history of the calculus of variations from the 17th through the 19th century. Springer, New York G¨ ollmann L, Kern D, Maurer H (2008) Optimal control problems with delays in state and control variables subject to mixed control-state constraints. Optimal Control Appl Methods 30(4):341–365 Gopalsamy K (1976) Optimal control of age-dependent populations. Math Biosci 32:155–163 Gordon HS (1954) Economic theory of a common-property resource: the ﬁshery. J Polit Econ 62:124–142 Gordon MJ (1962) The investment, ﬁnancing and valuation of the corporation. Richard D. Irwin, Homewood Gould JP (1970) Diﬀusion processes and optimal advertising policy. InbPhelps ES et al (eds) Microeconomic foundation of employment and inﬂation theory. Norton, New York, pp 338–368

500

Bibliography

Grass D, Caulkins JP, Feichtinger G, Tragler G, Behrens DA (2008) Optimal control of nonlinear processes: with applications in drugs, corruption, and terror. Springer, New York Grimm W, Well KH, Oberle HJ (1986) Periodic control for Minimum-fuel aircraft trajectories. J Guid 9:169–174 Gross M, Lieber Z (1984) Competitive monopolistic and eﬃcient utilization of an exhaustible resource in the presence of habit-formation eﬀects and stock dependent costs. Econ Lett 14:383–388 Hadley G, Kemp MC (1971) Variational methods in economics. North-Holland, Amsterdam Hahn M, Hyun JS (1991) Advertising cost interpretations and the optimality of pulsing. Manag Sci 37(2):157–169 Halkin H (1966) A maximum principle of the Pontryagin type for systems described by nonlinear diﬀerence equations. SIAM J Control 4:90–111 Halkin H (1967) On the necessary condition for optimal control of non-linear systems. In: Leitmann G (ed) Topics in optimization. Academic Press, New York H¨am¨ al¨ainen RP, Haurie A, Kaitala VT (1984) Bargaining on whales: a diﬀerential game model with Pareto optimal equilibria. Oper Res Lett 3(1):5–11 H¨am¨ al¨ainen RP, Haurie A, Kaitala VT (1985) Equilibria and threats in a ﬁshery management game. Optimal Control Appl Methods 6:315–333 H¨am¨ al¨ainen RP, Ruusunen J, Kaitala VT (1986) Myopic Stackelberg equilibria and social coordination in a share contract ﬁshery. Mar Resour Econ 3(3):209–235 H¨am¨ al¨ainen RP, Ruusunen J, Kaitala VT (1990) Cartels and dynamic contracts in shareﬁshing. J Environ Econ Manag 19:175–192 Han M, Feichtinger G, Hartl RF (1994) Nonconcavity and proper optimal periodic control. J Econ Dyn Control 18:976–990 Hanssens DM, Parsons LJ, Schultz RL (1990) Market response models: econometric and time series analysis. Kluwer Academic Publishers, Boston Harris FW (1913) How many parts to make at once. Factory Mag Manag 10:135–136, 152 Harris H (1976) Optimal planning under transaction costs: the demand for money and other assets. J Econ Theory 12:298–314

Bibliography

501

Harrison JM, Pliska SR (1981) Martingales, stochastic integrals, and continuous trading. Stoch Process Appl 11:215–260 Hartberger RJ (1973) A proof of the Pontryagin maximum principle for initialvalue problem. J Optim Theory Appl 11:139–145 Hartl RF (1982a) A mixed linear/nonlinear optimization model of production and maintenance for a machine. In: Feichtinger G (ed) Optimal control theory and economic analysis. North-Holland, Amsterdam, pp 43–58 Hartl RF (1982b) Optimal control of non-linear advertising models with replenishable budget. Optimal Control Appl Methods 3(1):53–65 Hartl RF (1982c) Optimal control of concave economic models with two control instruments. In: Feichtinger G, Kall P (eds) Operations research in progress. D. Reidel Publishing Company, Dordrecht, pp 227–245 Hartl RF (1983a) Optimal allocation of resources in the production of human capital. J Oper Res Soc 34:599–606 Hartl RF (1983b) Optimal maintenance and production rates for a machine: a nonlinear economic control problem. J Econ Dyn Control 6:281–306 Hartl RF (1984) Optimal dynamic advertising policies for hereditary processes. J Optim Theory Appl 43(1):51–72 Hartl RF (1986a) A forward algorithm for a generalized wheat trading model. Z Oper Res 30:A 135–A 144 Hartl RF (1986b) Arrow-type suﬃcient optimality conditions for nondiﬀerentiable optimal control problems with state constraints. Appl Math Optim 14:229–247 Hartl RF (1987) A simple proof of the monotonicity of the state trajectories in autonomous control problems. J Econ Theory 40:211–215 Hartl RF (1988a) A wheat trading model with demand and spoilage. In: Feichtinger G (ed) Optimal control theory and economic analysis, vol 3. NorthHolland, Amsterdam, pp 235–244 Hartl RF (1988b) A dynamic activity analysis for a monopolistic ﬁrm. Optimal Control Appl Methods 9:253–272 Hartl RF (1988c) The control of environmental pollution and optimal investment and employment decisions: a comment. Optimal Control Appl Methods 9:337–339 Hartl RF (1989a) On forward algorithms for a generalized wheat trading model. In: Chikan A (ed) Progress in inventory research (proceedings of 4th international symposium on inventories). Hungarian Academy of Science, Budapest

502

Bibliography

Hartl RF (1989b) Most rapid approach paths in dynamic economic problems. In: Kleinschmidt P et al (eds) Methods of operations research 58 (Proceedings of SOR 12, 1987). Athen¨aun, pp 397–410 Hartl RF (1992) Optimal acquisition of pollution control equipment under uncertainty. Manag Sci 38:609–622 Hartl RF (1993) On the properness of one-dimensional periodic control problems. Syst Control Lett 20(3):393–395 Hartl RF (1995) Production smoothing under environmental constraints. Prod Oper Manag 4(1):46–56 Hartl RF, Feichtinger G (1987) A new suﬃcient condition for most rapid approach paths. J Optim Theory Appl 54(2):403–411 Hartl RF, Jørgensen S (1985) Optimal manpower policies in a dynamic staﬀmaximizing bureau. Optimal Control Appl Methods 6(1):57–64 Hartl RF, Jørgensen S (1988) Aspects of optimal slidesmanship. In Feichtinger G (ed) Optimal control theory and economic analysis, vol 3. North-Holland, Amsterdam, pp 335–350 Hartl RF, Jørgensen S (1990) Optimal slidesmanship in conferences with unpredictable chairmen. Optimal Control Appl Methods 11:143–155 Hartl RF, Kort PM (1996a) Marketable permits in a stochastic dynamic model of the ﬁrm. J Optim Theory Appl 89(1):129–155 Hartl RF, Kort PM (1996b) Capital accumulation of a ﬁrm facing an emissions tax. J Econ 63(1):1–24 Hartl RF, Kort PM (1996c) Capital accumulation of a ﬁrm facing environmental constraints. Optimal Control Appl Methods 17:253–266 Hartl RF, Kort PM (1997) Optimal input substitution of a ﬁrm facing an environmental constraint. Eur J Oper Res 99:336–352 Hartl RF, Kort PM (2004) Optimal investments with convex-concave revenue: a focus-node distinction. Optimal Control Appl Methods 25(3):147–163 Hartl RF, Kort PM (2005) Advertising directed towards existing and new customers. In: Deissenberg C, Hartl RF (eds) Optimal control and dynamic games. Springer, Dordrecht, pp 3–18 Hartl RF, Krauth J (1989) Optimal production mix. J Optim Theory Appl 66:255–273

Bibliography

503

Hartl RF, Luptacik M (1992) Environmental constraints and choice of technology. Czechoslov J Oper Res 1(2):107–125 Hartl RF, Mehlmann A (1982) The Transylvanian problem of renewable resources. R´evue Francaise d‘Automatique, Informatique et de Recherche Operationelle 16:379–390 Hartl RF, Mehlmann A (1983) Convex-concave utility function: optimal bloodconsumption for vampires. Appl Math Model 7:83–88 Hartl RF, Mehlmann A (1984) Optimal seducing policies for dynamic continuous lovers under risk of being killed by a rival. Cybern Syst Int J 15:119–126 Hartl RF, Mehlmann A (1986) On remuneration patterns for medical services. Optimal Control Appl Methods 7:185–193 Hartl RF, Sethi SP (1983) A note on the free terminal time transversality condition. Z Oper Res Ser Theory 27(5):203–208 Hartl RF, Sethi SP (1984a) Optimal control problems with diﬀerential inclusions: suﬃciency conditions and an application to a production-inventory model. Optimal Control Appl Methods 5(4):289–307 Hartl RF, Sethi SP (1984b) Optimal control of a class of systems with continuous lags: dynamic programming approach and economic interpretations. J Optim Theory Appl 43(1):73–88 Hartl RF, Sethi SP (1985a) Solution of generalized linear optimal control problems using a simplex-like method in continuous-time I: theory. In: Feichtinger G (ed) Optimal control theory and economic analysis, vol 2. North-Holland, Amsterdam, pp 45–62 Hartl RF, Sethi SP (1985b) Solution of generalized linear optimal control problems using a simplex-like method in continuous-time II: examples. In Feichtinger G (ed) Optimal control theory and economic analysis, vol 2. NorthHolland, Amsterdam, pp 63–87 Hartl RF, Mehlmann A, Novak AJ (1992a) Cycles of fear: optimal periodic blood-sucking rates for vampires. J Optim Theory Appl 75(3):559–568 Hartl RF, Feichtinger G, Kirakossian GT (1992b) Optimal recycling of tailings for the production of building materials. Czechoslov J Oper Res 1(3):181–192 Hartl RF, Sethi SP, Vickson RG (1995) A survey of the maximum principles for optimal control problems with state constraints. SIAM Rev 37(2):181–218 Hartl RF, Kort PM, Novak AJ (1999) Optimal investment facing possible accidents. Ann Oper Res 88:99–117

504

Bibliography

Hartl RF, Kort PM, Feichtinger G (2003a) Oﬀense control taking into account heterogeneity of age. J Optim Theory Appl 116:591–620 Hartl RF, Novak AJ, Rao AG, Sethi SP (2003b) Optimal pricing of a product diﬀusing in rich and poor populations. J Optim Theory Appl 117(2):349–375 Hartl RF, Kort PM, Feichtinger G, Wirl F (2004) Multiple equilibria and thresholds due to relative investment costs. J Optim Theory Appl 123:49–82 Hartl RF, Novak AJ, Rao AG, Sethi SP (2005) Dynamic pricing of a status symbol, invited talks from the fourth world congress of nonlinear analysts (WCNA 2004) Orlando, FL, June 30–July 07, 2004. Nonlinear Anal Theory Methods Appl 63:e2301–e2314 Hartman R (1982) Ordinary diﬀerential equations. Birkh¨auser, Boston Haruvy E, Prasad A, Sethi SP (2003) Harvesting altruism in open source software development. J Optim Theory Appl 118(2):381–416 Haruvy E, Prasad A, Sethi SP, Zhang R (2005) Optimal ﬁrm contributions to open source software. In: Deissenberg C, Hartl RF (eds) Optimal control and dynamic games, applications in ﬁnance, management science and economics. Springer, Dordrecht, pp 197–212 Haruvy E, Prasad A, Sethi SP, Zhang R (2008a) Competition with open source as a public good. J Ind Manag Optim 4(1):199–211 Haruvy E, Sethi SP, Zhou J (2008b) Open source development with a commercial complementary product or service. Prod Oper Manag (Special Issue on Management of Technology) 17(1):29–43 Harvey AC (1994) Forecasting, structural time series models and the Kalman ﬁlter. Cambridge University Press, New York Haunschmied JL, Kort PM, Hartl RF, Feichtinger G (2003) A DNS-curve in a two-state capital accumulation model: a numerical analysis. J Econ Dyn Control 27:701–716 Haurie A (1976) Optimal control on an inﬁnite time horizon: the turnpike approach. J Math Econ 3:81–102 Haurie A, Hung NM (1977) Turnpike properties for the optimal use of a natural resource. Rev Econ Stud 44:329–336 Haurie A, Leitmann G (1984) On the global asymptotic stability of equilibrium solutions for open-loop diﬀerential games. Large Scale Syst 6:107–122 Haurie A, Sethi SP (1984) Decision and forecast horizons, agreeable plans, and the maximum principle for inﬁnite horizon control problems. Oper Res Lett 3(5):261–265

Bibliography

505

Haurie A, Tolwinski B, Leitmann G (1983) Cooperative equilibria in diﬀerential games. In: Proceedings ACC, San Francisco Haurie A, Sethi SP, Hartl RF (1984) Optimal control of an age-structured population model with applications to social services planning. J Large Scale Syst 6:133–158 Haussmann UG (1981) Some examples of optimal stochastic controls or: the stochastic maximum principle at work. SIAM Rev 23:292–307 He X, Sethi SP (2008) Dynamic slotting and pricing decisions in a durable product supply chain. J Optim Theory Appl 137(2):363–379 He X, Prasad A, Sethi SP, Gutierrez GJ (2007) A survey of Stackelberg diﬀerential game models in supply and marketing channels. J Syst Sci Syst Eng 16(4):385–413. Erratum (2008) 17(2):255 He X, Prasad A, Sethi SP (2009) Cooperative advertising and pricing in a stochastic supply chain: feedback Stackelberg strategies. Prod Oper Manag 18(1):78–94 He X, Krishnamoorthy A, Prasad A, Sethi SP (2011) Retail competition and cooperative advertising. Oper Res Lett 39:11–16 He X, Krishnamoorthy A, Prasad A, Sethi SP (2012) Co-op advertising in dynamic retail oligopolies. Decis Sci 43(1):73–105 Heal GM (1976) The relationship between price and extraction cost for a resource with a backstop technology. Bell J Econ 7:371–378 Heal GM (1993) The optimal use of exhaustible resources. In: Kneese AV, Sweeney (eds) Handbook of natural resource and energy economics, vol 3, chap 18. Elsevier, London, pp 855–880 Heaps T (1984) The forestry maximum principle. J Econ Dyn Control 7:131–151 Heckman J (1976) A life cycle model of earnings, learning, and consumption. J Polit Econ 84:511–544 Hestenes MR (1966) Calculus of variations and optimal control theory. Wiley, New York Ho YC (1970) Diﬀerential games, dynamic optimization and generalized control theory. J Optim Theory Appl 6:179–209 Hofbauer J, Sorger G (1999) Perfect foresight and equilibrium selection in symmetric potential games. J Econ Theory 85:1–23

506

Bibliography

Hofbauer J, Sorger G (2002) A diﬀerential game approach to evolutionary equilibrium selection. Int Game Theory Rev 4:17–31 Hoﬀmann KH, Krabs W (1984) Optimal control of partial diﬀerential equations. Birkh¨auser, Basel Holly S, R¨ ustem B, Zarrop MB (eds) (1979) Optimal control for econometric models. An approach to economic policy formulation. Macmillan, London Holt CC, Modigliani F, Muth JF, Simon HA (1960) Planning production, inventories and workforce. Prentice-Hall, Englewood Cliﬀs Holtzman JM (1966) On the maximum principle for nonlinear discrete-time systems. IEEE Trans Autom Control AC-11:273–274 Horsky D (1977) An empirical analysis of the optimal advertising policy. Manag Sci 23(10):1037–1049 Horsky D, Mate K (1988) Dynamic advertising strategies of competing durable good producers. Market Sci 7(4):356–367 Horsky D, Simon LS (1983) Advertising and the diﬀusion of new products. Market Sci 2(1):1–17 Hotelling H (1925) A general mathematical theory of depreciation. J Am Stat Assoc 20:340–353 Hotelling H (1931) The economics of exhaustible resources. J Polit Econ 39:137– 175 Hwang CL, Fan LT, Erickson LE (1967) Optimal production planning by the maximum principle. Manag Sci 13:750–755 Ijiri Y, Thompson GL (1970) Applications of mathematical control theory to accounting and budgeting (the continuous wheat trading model). Account Rev 45:246–258 Ijiri Y, Thompson GL (1972) Mathematical control theory solution of an interactive accounting ﬂows model. Naval Res Logist Quart 19:411–422 Intriligator MD (1971) Mathematical optimization and economic theory. Prentice-Hall, Englewood Cliﬀs Intriligator MD (1980) Applications of control theory to economics. In: Bensoussan A, Lions JL (eds) Analysis and optimization of systems. Lecture notes in control and information sciences, vol 28. Springer, Berlin, pp 607– 626

Bibliography

507

Intriligator MD, Smith BLR (1966) Some aspects of the allocation of scientiﬁc eﬀort between teaching and research. Am Econ Rev 61:494–507 Ioﬀe AD, Tihomirov VM (1979) Theory of extremal problems. North-Holland, Amsterdam Isaacs R (1965) Diﬀerential games. Wiley, New York Isaacs R (1969) Diﬀerential games: their scope, nature, and future. J Optim Theory Appl 3:283–295 Jacobson DH, Lele MM, Speyer JL (1971) New necessary conditions of optimality for control problems with state-variable inequality constraints. J Math Anal Appl 35:255–284 Jacquemin AP (1973) Optimal control and advertising policy. Metro-Economica 25:200–207 Jacquemin AP, Thisse J (1972) Strategy of the ﬁrm and market structure: an application of optimal control theory. In: Cowling K (ed) Market structure and corporate behavior. Gray-Mills, London, pp 61–84 Jagpal S (1999) Marketing strategy and uncertainty. Oxford University Press, New York Jamshidi M (1983) Large-scale systems: modelling and control. North-Holland, New York Jarrar R, Mart´ın-Herr´ an G, Zaccour G (2004) Markov perfect equilibrium advertising strategies of Lanchester duopoly model: a technical note. Manag Sci 50(7):995–1000 Jazwinski AH (1970) Stochastic processes and ﬁltering theory. Academic Press, New York Jedidi K, Eliashberg J, DeSarbo W (1989) Optimal advertising and pricing for a three-stage time-lagged monopolistic diﬀusion model incorporating income. Optimal Control Appl Methods 10:313–331 Jennings LS, Teo KL (1997) Computation of manufacturing systems using an enhancing control. In: Chandra T, Leclair SR, Meech JA, Verma B, Smith M, Balachandran B (eds) IPMM’97: Australiasia-Paciﬁc forum on intelligent processing & manufacturing of materials, vol 1. Watson Ferguson, Brisbane Jennings LS, Sethi SP, Teo KL (1997) Computation of optimal production plans for manufacturing systems. Nonlinear Anal Theory Methods Appl 30(7):4329–4338

508

Bibliography

Jeuland AP, Dolan RJ (1982) An aspect of new product planning: dynamic pricing. In: Zoltners AA (ed) Marketing planning models. TIMS studies in the management science, vol 18. North-Holland, Amsterdam, pp 1–21 Ji Y, Mookerjee VS, Sethi SP (2005) Optimal software development: a control theoretic approach. Inf Syst Res 16(3):292–306 Ji Y, Kumar S, Mookerjee VS, Sethi SP, Yeh D (2011) Optimal enhancement and lifetime of software systems: a control theoretic analysis. Prod Oper Manag 20(6):889–904 Ji Y, Kumar S, Sethi SP (2017) Needle exchange for controlling HIV spread under endogenous infectivity. Inf Syst Oper Res 55(2):93–117 Jiang J, Sethi SP (1991) A state aggregation approach to manufacturing systems having machine states with weak and strong interactions. Oper Res 39(6):970–978 Johar M, Mookerjee V, Sethi SP (2015) Optimal software design reuse policies: a control theoretic approach. Inf Syst Front 17(2):439–453 Johnson CD, Gibson JE (1963) Singular solutions in problems of optimal control. IEEE Trans Autom Control AC-8:4–15 Jones P (1983) Analysis of a dynamic duopoly model of advertising. Math Oper Res 8(1):122–134 Jørgensen S (1982a) A survey of some diﬀerential games in advertising. J Econ Dyn Control 4:341–369 Jørgensen S (1982b) A diﬀerential games solution to a logarithmic advertising model. Oper Res Soc 33(5):425–432 Jørgensen S (1983) Optimal control of a diﬀusion model of new product acceptance with price-dependent total market potential. Optimal Control Appl Methods 4(3):269–276 Jørgensen S (1984) A Pareto-optimal solution of a maintenance-production differential game. Eur J Oper Res 18:76–80 Jørgensen S (1985) An exponential diﬀerential game which admits a simple Nash solution. J Optim Theory Appl 45:383–396 Jørgensen S (1986a) Optimal production, purchasing and pricing: a diﬀerential games approach. Eur J Oper Res 24:64–76 Jørgensen S (1986b) Optimal dynamic pricing in an oligopolistic market: a survey. In: Basar T (ed) Dynamic games and applications in economics. Springer, Berlin, pp 179–237

Bibliography

509

Jørgensen S (1992) The dynamics of extramarital aﬀairs. In: Feichtinger G (ed) Dynamic economic models and optimal control. North-Holland, Amsterdam pp 239–266 Jørgensen S, Dockner EJ (1985) Optimal consumption and replenishment policies for a renewable resource. In: Feichtinger G (ed) Optimal control theory and economic analysis, vol 2. North-Holland, Amsterdam, pp 647–664 Jørgensen S, Kort PM (1993) Optimal dynamic investment policies under concave-convex adjustment costs. J Econ Dyn Control 17(1–2):153–180 Jørgensen S, Kort PM (1997) Optimal investment and ﬁnance in renewable resource harvesting. J Econ Dyn Control 21:603–630 Jørgensen S, Kort PM (2002) Optimal pricing and inventory policies: centralized and decentralized decision making. Eur J Oper Res 138(3):578–600 Jørgensen S, Zaccour G (2001) Time consistent side payments in a dynamic game of downstream pollution. J Econ Dyn Control 25:1973–87 Jørgensen S, Zaccour G (2004) Diﬀerential games in marketing. International series in quantitative marketing. Kluwer Academic Publishers, Boston Jørgensen S, Zaccour G (2007) Developments in diﬀerential game theory and numerical methods: economic and management applications. Comput Manag Sci 4(2):159–182 Jørgensen S, Kort PM, Zaccour G (1999) Production, inventory, and pricing under cost and demand learning eﬀects. Eur J Oper Res 117:382–395 Jørgensen S, Kort PM, Dockner EJ (2006a) Venture capital ﬁnanced investments in intellectual capital. J Econ Dyn Control 30(11):2339–2361 Jørgensen S, Kort PM, Zaccour G (2006b) Advertising an event. Automatica 42(8):1349–1355 Jørgensen S, Kort PM, Zaccour G (2009) Optimal pricing and advertising policies for an entertainment event. J Econ Dyn Control 33(3):583–596 Kaitala VT (1986) Game theory models of ﬁsheries management - a survey. In: Basar T (ed) Dynamic games and applications in economics. Springer, Berlin, pp 252–266 Kalish S (1983) Monopolist pricing with dynamic demand and production cost. Market Sci 2(2):135–159 Kalish S (1985) A new product adoption model with price, advertising, and uncertainty. Market Sci 31(12):1569–1585

510

Bibliography

Kalish S, Lilien GL (1983) Optimal price subsidy policy for accelerating the diﬀusion of innovation. Market Sci 2(4):407–420 Kalish S, Sen SK (1986) Diﬀusion models and the marketing mix for single products. In: Mahajan V, Wind Y (eds) Series in econometrics and management science: innovation diﬀusion models of new products acceptance, vol V. Ballinger, Cambridge, pp 87–116 Kalman RE (1960a) A new approach to linear ﬁltering and prediction problems. Trans ASME Ser D J Basic Eng 82:35–45 Kalman RE (1960b) Contributions to the theory of optimal control. Bol de Soc Math Mexicana 5:102–119 Kalman RE, Bucy R (1961) New results in linear ﬁltering and prediction theory. Trans ASME Ser D J Basic Eng 83:95–108 Kamien MI, Schwartz NL (1971a) Optimal maintenance and sale age for a machine subject to failure. Manag Sci 17:427–449 Kamien MI, Schwartz NL (1971b) Limit pricing and uncertain entry. Econometrica 39:441–454 Kamien MI, Schwartz NL (1978) Optimal exhaustible resource depletion with endogenous technical change. Rev Econ Stud 45:179–196 Kamien MI, Schwartz NL (1982a) Market structure and innovation. Cambridge University Press, Cambridge Kamien MI, Schwartz NL (1982b) The role of common property resources in optimal planning models with exhaustible resources. In: Smith VK, Krutilla JV (eds) Explorations in natural resource economics. John Hopkins University Press, Baltimore, pp 47–71 Kamien MI, Schwartz NL (1992) Dynamic optimization: the calculus of variations and optimal control in economics and management, 2nd edn, NorthHolland, New York Kaplan W (1958) Ordinary diﬀerential equations. Addison-Wesley, Reading Karatzas I, Shreve SE (1997) Brownian motion and stochastic calculus, 2nd edn. Springer, New York Karatzas I, Shreve SE (1998) Methods of mathematical ﬁnance. Springer, New York Karatzas I, Lehoczky JP, Sethi SP, Shreve SE (1986) Explicit solution of a general consumption/investment problem. Math Oper Res 11(2):261–294

Bibliography

511

Karray S, Mart´ın-Herr´ an G (2009) A dynamic model for advertising and pricing competition between national and store brands. Eur J Oper Res 193:451–467 Keeler E, Spence M, Zeckhauser RJ (1971) The optimal control of pollution. J Econ Theory 4:19–34 Keller HB (1968) Numerical methods for two-point boundary value problems. Blaisdell, Waltham Kem MC, Long NV (eds) (1980) Exhaustible resources, optimality, and trade. North-Holland, Amsterdam Kemp MC, Long NV (1977) Optimal control problems with integrands discontinuous with respect to time. Econ Rec 53:405–420 Kendrick DA (1981) Stochastic Control for economic models. McGraw-Hill, New York Khmelnitsky E, Kogan K (1994) Necessary optimality conditions for a generalized problem of production scheduling. Optimal Control Appl Methods 15:215–222 Khmelnitsky E, Kogan K (1996) Optimal policies for aggregate production and capacity planning under an arbitrary demand. Int J Prod Res 34(7):1929– 1941 Khmelnitsky E, Kogan K, Maimon O (1995) A maximum principle based combined method for scheduling in a ﬂexible manufacturing system. Discrete Event Dyn Syst 5:343–355 Khmelnitsky E, Kogan K, Maimon O (1997) Maximum principle-based methods for scheduling FMS with partially sequence dependent setups. Int J Prod Res 35(10):2701–2712 Khmelnitsky E, Presman E, Sethi SP (2011) Optimal production control of a failure-prone machine. Ann Oper Res 182(1):67–86 Kilkki P, Vaisanen U (1969) Determination of optimal policy for forest stands by means of dynamic programming. Acta Forestalia Fennica 102:100–112 Kirby BJ (ed) (1974) Optimal control theory and its applications. Lecture notes in economics and mathematical systems, part I & II, vols 105–106. Springer, Berlin Kirk DE (1970) Optimal control theory: an introduction. Prentice-Hall, Englewood Cliﬀs Kiseleva T, Wagener FO (2010) Bifurcations of optimal vector ﬁelds in the shallow lake model J Econ Dyn Control 34(5):825–843

512

Bibliography

Klein CF, Gruver WA (1978) On the optimal control of a single-server queueing system: comment. J Optim Theory Appl 26:457–462 Kleindorfer PR (1978) Stochastic control models in management science: theory and computation. TIMS Stud Manag Sci 9:69–88 Kleindorfer PR, Lieber Z (1979) Algorithms and planning horizon results for production planning problems with separable costs. Oper Res 27:874–887 Kleindorfer PR, Kriebel CH, Thompson GL, Kleindorfer GB (1975) Discrete optimal control of production plans. Manag Sci 22:261–273 Knobloch HW (1981) Higher order necessary conditions in optimal control theory. Lecture notes in control and information sciences, vol 34. Springer, Berlin Knowles G (1981) An introduction to applied optimal control. Academic Press, New York Kogan K, Khmelnitsky E (1996) An optimal control model for continuous time production and setup scheduling. Int J Prod Res 34(3):715–725 Kogan K, Khmelnitsky E, Shtub A, Maimon O (1997) Optimal ﬂow control of ﬂexible manufacturing systems: setup localization by an iterative procedure. Int J Prod Econ 51:37–46 Kort PM, Caulkins JP, Hartl RF, Feichtinger G (2006) Brand image and brand dilution in the fashion industry. Automatica 42:1363–1370 Kotowitz Y, Mathewson F (1979a) Informative advertising and welfare. Am Econ Rev 69:284–294 Kotowitz Y, Mathewson F (1979b) Advertising, consumer information, and product quality. Bell J. Econ 10:566–588 Kreindler E (1982) Additional necessary conditions for optimal control with state-variable inequality constraints. J Optim Theory Appl 38:241–250 Krelle W (1984) Economic growth with exhaustible resources and environmental protection. Z Staatswiss 140:399–429 Krichagina E, Lou S, Sethi SP, Taksar MI (1993) Production control in a failureprone manufacturing system: diﬀusion approximation and asymptotic optimality. Ann Appl Probab 3(2):421–453 Krichagina E, Lou S, Sethi SP, Taksar MI (1995) Diﬀusion approximation for a controlled stochastic manufacturing system with average cost minimization. Math Oper Res 20(4):895–922

Bibliography

513

Krishnamoorthy A, Misra S, Prasad A (2005) Scheduling sales force training: theory and evidence. Int J Res Market 22:427–440 Krishnamoorthy A, Prasad A, Sethi SP (2010) Optimal pricing and advertising in a durable-good duopoly. Eur J Oper Res 200(2):486–497 Krouse CG (1972) Optimal ﬁnancing and capital structure programs for the ﬁrm. J Financ 27:1057–1071 Krouse CG, Lee WY (1973) Optimal equity ﬁnancing of the corporation. J Financ Quant Anal 8:539–563 Kugelmann B, Pesch HJ (1990) New general guidance method in constrained optimal control, part 1: numerical method. J Optim Theory Appl 67(3):421– 435 Kuhn M, Wrzaczek S, Prskawetz A, Feichtinger G (2011) Externalities in a life cycle model with endogenous survival. J Math Econ 47:4–5, 627–641 Kuhn M, Wrzaczek S, F¨ urnkranz-Prskawetz A, Feichtinger G (2015) Optimal choice of health and retirement in a life-cycle model. J Econ Theory 158:186– 212 Kumar S, Sethi SP (2009) Dynamic pricing and advertising for web content providers. Eur J Oper Res 197:924–944 Kumar PR, Varaiya P (1986) Stochastic systems. Estimation, identiﬁcation, and adaptive control. Prentice-Hall, Englewood Cliﬀs Kurawarwala AA, Matsuo H (1996) Forecasting and inventory management of short life-cycle products. Oper Res 44(1):131–150 Kurcyusz S, Zowe J (1979) Regularity and stability for the mathematical programming problem in Banach spaces. Appl Math Optim 5:49–62 Kushner HJ (1971) Introduction to stochastic control. Holt, Rinehart & Winston, New York Kushner HJ (1977) Probability methods for approximations in stochastic control and for elliptic equations. Academic Press, New York Kydland FE, Prescott EC (1977) Rules rather than discretion: the inconsistency of optimal plans. J Polit Econ 85:473–493 Laﬀont JJ, Mortimort D (2001) The theory of incentives: the principal-agent model. Princeton University Press, Princeton Lagunov VN (1985) Introduction to diﬀerential games and control theory. Hildermann, Berlin

514

Bibliography

Lansdowne ZF (1970) The theory and applications of generalized linear control processes. Technical report no. 10, Department of Operations Research, Stanford University Lasdon LS, Mitter SK, Warren AD (1967) The conjugate gradient method for optimal control problems. IEEE Trans Autom Control AC-12:132–138 Leban R, Lesourne J (1983) Adaptive strategies of the ﬁrm through a business cycle. J Econ Dyn Control 5:201–234 Lee EB, Markus L (1968) Foundations of optimal control theory. Wiley, New York Legey L, Ripper M, Varaiya PP (1973) Eﬀects of congestion on the shape of the city. J Econ Theory 6:162–179 Lehoczky JP, Sethi SP, Soner HM, Taksar MI (1991) An asymptotic analysis of hierarchical control of manufacturing systems under uncertainty. Math Oper Res 16(3):596–608 Leitmann G (1974) Cooperative and non-cooperative many players diﬀerential games. Springer, Wien Leitmann G (ed) (1976) Multicriteria decision making and diﬀerential games. Plenum Press, New York Leitmann G (1981) The calculus of variations and optimal control. In: Miele A (ed) Series mathematical concepts and methods in science and engineering. Plenum Press, New York Leitmann G, Liu PT (1974) A diﬀerential game model of labor-management negotiation during a strike. J Optim Theory Appl 13:427–444 Leitmann G, Stalford H (1971) A suﬃciency theorem for optimal control. J Optim Theory Appl 8:169–174 Leland HE (1972) The dynamics of a revenue maximizing ﬁrm. Int Econ Rev 13:376–385 Lele MM, Jacobson DH, McCabe JL (1971) Qualitative application of a result in control theory to problems of economic growth. Int Econ Rev 12:209–226 L´eonard D, Long NV(1992) Optimal control theory and static optimization in economics. Cambridge University Press, Cambridge Lesourne J (1973) Croissance optimale des entreprises. Dunod, Paris Lesourne J, Leban R (1982) Control theory and the dynamics of the ﬁrm: a survey. OR-Spektr 4:1–14

Bibliography

515

Levine J, Th´epot J (1982) Open loop and closed loop equilibria in a dynamic duopoly. In: Feichtinger G (ed) Optimal control theory and economic analysis. North-Holland, Amsterdam, pp 143–156 Lewis TR, Schmalensee R (1977) Non-convexity and optimal exhaustion of renewable resources. Int Econ Rev 18:535–552 Lewis TR, Schmalensee R (1979) Non-convexity and optimal harvesting strategies for renewable resources. Can J Econ 12:677–691 Lewis TR, Schmalensee R (1982) Optimal use of renewable resources with nonconvexities in production. In: Mirman LJ, Spulber PF (eds) Essays in the economics of renewable resources. North-Holland, Amsterdam, pp 95–111 Li G, Rajagopalan S (1997) A learning curve model with knowledge depreciation. Eur J Oper Res 105(1):143–154 Li G, Rajagopalan S (1998) Process improvement, quality and learning eﬀects. Manag Sci 44:1517–1532 Li T, Sethi SP (2017) A review of dynamic Stackelberg game models. Discrete Continuous Dyn Syst Ser B 22(1):125–159 Lieber Z (1973) An extension of Modigliani and Hohn’s planning horizon results. Manag Sci 20:319–330 Lieber Z, Barnea A (1977) Dynamic optimal pricing to deter entry under constrained supply. Oper Res 25:696–705 Lignell J, Tuominen MPT (1983) An advertising control model of two state variables. Eur J Oper Res 24(1):77–84 Lintner J (1963) The cost of capital and optimal ﬁnancing of corporate growth. J Financ 23:292–310 Lions JL (1971) Optimal control of systems governed by partial diﬀerential equations. Springer, New York Little JDC (1979) Aggregate advertising models: the state of the art. Oper Res 27(4):629–667 Little JDC (1986) Comment on “Advertising pulsing policies...” by V. Mahajan and E. Muller. Market Sci 5(2):107–108 Liu PT (1980) Dynamic optimization and mathematical economics. Plenum Press, New York Liu PT, Roxin EO (eds) (1979) Diﬀerential games and control theory III. Marcel ˙ Dekker, New York

516

Bibliography

Liu PT, Sutinen JG (eds) (1979) Control theory in mathematical economics. ˙ Marcel Dekker, New York Long NV, Sorger G (2006) Insecure property rights and growth: the role of appropriation costs, wealth eﬀects, and heterogeneity. Econ Theory 28:513– 529 Long NV, Vousden N (1977) Optimal control theorems. In: Pitchford JD, Turnovsky SJ (eds) Applications of control theory in economic analysis. North-Holland, Amsterdam, pp 11–34 Lou H (2007) Existence and non-existence results of an optimal control problem by using relaxed control. SIAM J Control Optim 46:1923–1941 Lou S, Sethi SP, Zhang Q (1994) Optimal feedback production planning in a stochastic two-machine ﬂowshop. Eur J Oper Res 73:331–345 Lucas RE Jr (1971) Optimal management of a research and development project. Manag Sci 17:679–697 Lucas RE Jr (1981) Optimal investment with rational expectations. In: Lucas RE Jr, Sargent TJ (eds) Rational expectations and economic practice. G. Allen & Unwin, London, pp 55–66 Luenberger DG (1969) Optimization by vector space methods. Wiley, New York Luenberger DG (1972) Mathematical programming and control theory: trends of interplay. In: Geoﬀrion AM (ed) Perspectives on optimization. AddisonWesley, Reading Luenberger DG (1973) Introduction to linear and nonlinear programming. Addison-Wesley, Reading Luenberger DG (1975) A nonlinear economic control problem with a linear feedback solution. IEEE Trans Autom Control AC-20:184–191 Luenberger DG (1979) Introduction to dynamic systems: theory, models, and applications. Wiley, New York Luhmer A, Steindl A, Feichtinger G, Hartl RF, Sorger G (1988) ADPULS in continuous time. Eur J Oper Res 34(2):171–177 Lundin RA, Morton TE (1975) Planning horizons for the dynamic lot size model: protective procedures and computational results. Oper Res 23:711– 734 Luptacik M (1982) Optimal price and advertising policy under atomistic competition. J Econ Dyn Control 4:57–71

Bibliography

517

Luptacik M, Schubert U (1982) Optimal investment policy in productive capacity and pollution abatement processes in a growing economy. In: Feichtinger G (ed) Optimal control theory and economic analysis. North-Holland, Amsterdam, pp 231–243 Luus R (1993) Piecewise linear continuous optimal control by iterative dynamic programming. Ind Eng Chem Res 32:859–865 Lykina V, Pickenhain S, Wagner M (2008) Diﬀerent interpretations of the improper integral objective in an inﬁnite horizon control problem. J Math Anal Appl 340:498–510 Macki J, Strauss A (1982) Introduction to optimal control theory. Springer, New York Magat WA, McCann JM, Morey RC (1986) When does lag structure really matter in optimizing advertising expenditures? Manag Sci 32(2):182–193 Magat WA, McCann JM, Morey RC (1988) Reply to when does lag structure really matter ... Indeed? Manag Sci 34(7):917–918 Magill MJP (1970) On a general economic theory of motion. Springer, New York Mahajan V, Muller E (1979) Innovation diﬀusion and new product growth models in marketing. J Market 43:55–68 Mahajan V, Muller E (1986) Advertising pulsing policies for generating awareness for new products. Market Sci 5(2):89–111 Mahajan V, Peterson RA (1978) Innovation diﬀusion in a dynamic potential adopter population. Manag Sci 24:1589–1597 Mahajan V, Peterson RA (1985) Models for innovation diﬀusion. Sage, Beverly Hills Mahajan V, Wind Y (eds) (1986) Innovation diﬀusion models of new product acceptance. Ballinger, Cambridge Maimon O, Khmelnitsky E, Kogan K (1998) Optimal ﬂow control manufacturing systems. Kluwer Academic Publishers, Boston Majumdar M, Mitra T (1983) Dynamic optimization with a non-convex technology: the case of a linear objective function. Rev Econ Stud 50:143–151 Malanowski K (1984) On diﬀerentiability with respect to parameter of solutions to convex optimal control problems subject to state space constraints. Appl Math Optim 12:231–245

518

Bibliography

Malanowski K (1997) Suﬃcient optimality conditions for optimal control subject to state constraints. SIAM J Control Optim 35(1):205–227 Malanowski K, Maurer H, Pickenhain S (2004) Second-order suﬃcient conditions for state-constrained optimal control problems. J Optim Theory Appl 123:595–617 Malliaris AG, Brock WA (1982) Stochastic methods in economics and ﬁnance. North-Holland, New York Mangasarian OL (1966) Suﬃcient conditions for the optimal control of nonlinear systems. SIAM J Control 4:139–152 Mangasarian OL (1969) Nonlinear programming. McGraw-Hill, New York Mangasarian OL, Fromovitz S (1967) A maximum principle in mathematical programming. In: Balakrishman AV, Neustadt LW (eds) Mathematical theory of control. Academic Press, New York Manh-Hung N (1974) Essays on the optimal dynamic exploitation of natural resources and the social rate of discount. Ph.D. Dissertation, University of Toronto Marinelli C, Savin S (2008) Optimal distributed dynamic advertising. J Optim Theory Appl 137:569–591 Mart´ın-Herr´ an G, Rinc´on-Zapatero JP (2002) Computation of subgame perfect Nash equilibria without Hamilton-Jacobi-Bellman equations. In: Zaccour G (ed), Optimal control and diﬀerential games. Essays in honor of Steﬀen Jørgensen. Kluwer Academic Publishers, Dordrecht, pp 135–151 Mart´ın-Herr´ an G, Taboubi S, Zaccour G (2005) A time-consistent open loop Stackelberg equilibrium of shelf-space allocation. Automatica 41:971–982 Mart´ın-Herr´ an G, Rubel O, Zaccour G (2008) Competing for consumer’s attention. Automatica 44:361–370 Martirena-Mantel AM (1971) Optimal inventory and capital policy under certainty. J Econ Theory 3:241–253 Mass´e P (1962) Optimal investment decisions. Prentice-Hall, Englewood Cliﬀs Maurer H (1976) Numerical solution of singular control problems using multiple shooting techniques. J Optim Theory Appl 18:235–257 Maurer H (1977) On optimal control problems with bounded state variables and control appearing linearly. SIAM J Control Optim 15:345–362

Bibliography

519

Maurer H (1979) Diﬀerential stability in optimal control problems. Appl Math Optim 5:283–295 Maurer H (1981) First and second order suﬃcient optimality conditions in mathematical programming and optimal control. Math. Program Stud 14:163–177 Maurer H, Gillessen W (1975) Application of multiple shooting to the numerical solution of optimal control problems with bounded state variables. Computing 15:105–126 Maurer H, Wiegand M (1992) Numerical solution of a drug displacement problem with bounded state variables. Optimal Control Appl Methods 13:43–55 Maurer H, B¨ uskens C, Feichtinger G (1998) Solution techniques for periodic control problems: a case study in production planning. Optimal Control Appl Methods 19:185–203 Maurer H, Kim JHR, Vossen G (2005) On a state-constrained control problem in optimal production and maintenance. In: Deissenberg C, Hartl RF (eds) Optimal control and dynamic games. Springer, Dordrecht, pp 289–308 Mayne DQ, Polak E (1987) An exact penalty function algorithm for control problems with state and control constraints. IEEE Trans Autom Control 32:380–387 McIntyre J, Paiewonsky B (1967) On optimal control with bounded state variables. In: Leondes CT (ed) Advances in control systems, vol 5. Academic Press, New York Mehra RK (1975) An optimal control approach to national settlement system planning, vol RM-75-58. International Institute of Applied Systems Analysis, Laxenburg Mehra RK, Davis RE (1972) Generalized gradient method for optimal control problems with inequality constraint and singular arcs. IEEE Trans Autom Control AC-17:69–79 Mehlmann A (1985) State transformations and the derivation of Nash closedloop equilibria for non-zero-sum diﬀerential games. Appl Math Model 9:353– 357 Mehlmann A (1988) Applied diﬀerential games. Plenum, New York Mehlmann A (1997) Wer Gewinnt Das Spiel? - Spieltheorie in Fabeln und Paradoxa. Vieweg, Braunschweig Mehrez A (1983) A note on the comparison of two diﬀerent formulations of a risky R & D model. O R Lett 2:249–251

520

Bibliography

Merton RC (1969) Lifetime portfolio selection under uncertainty: continuous-time case. Rev Econ Stat 51:247–257

the

Merton RC (1971) Optimum consumption and portfolio rules in a continuoustime model. J Econ Theory 3:373–413 Merton RC (1973) An intertemporal capital asset pricing model. Econometrica 5:867–888 Merton RC (1982) On the microeconomic theory of investment under uncertainty. In: Arrow KJ, Intriligator MD (eds) Handbook of mathematical economics, vol II. North-Holland, Amsterdam, pp 601–669 Mesak HI (1985) On modeling advertising pulsing decisions. Decis Sci 16:25–42 Mesak HI, Darrat AF (1992) On comparing alternative advertising policies for pulsation. Decis Sci 23:541–564 Michel P (1981) Choice of projects and their starting dates: an extension of Pontryagin’s maximum principle to a case which allows choice among diﬀerent possible evolution equations. J Econ Dyn Control 3:97–118 Michel P (1982) On the transversality condition in inﬁnite horizon optimal control problems. Econometrica 50:975–985 Michel P (1985) Application of optimal control theory to disequilibrium analysis. In: Feichtinger G (ed) Optimal control theory and economic analysis, vol 2. North-Holland, Amsterdam, pp 417–427 Miele A (1962) Extremization of linear integral equations by Green’s theorem. In: Leitmann G (ed) Optimization techniques. Academic Press, New York Miller RE (1979) Dynamic optimization and economic applications. McGrawHill, New York Miller MH, Modigliani F (1961) Dividend policy, growth, and valuation of shares. J Bus 34:411–433 Mirman LJ, Spulber PF (eds) (1982) Essays in the economics of renewable resources. North-Holland, Amsterdam Mirrlees J (1967) Optimum growth when technology is changing. Rev Econ Stud 34:95–124 Mirrlees J (1972) The optimum town. Swed J Econ 74:114–135 Modigliani F, Hohn F (1955) Production planning over time. Econometrica 23:46–66

Bibliography

521

Moiseev NN (1971) Numerical methods in optimal control theory. Nauka, Moscow Monahan GE (1984) A pure birth model of optimal advertising with word-ofmouth. Market Sci 3(2):169–178 Mond B, Hanson M (1968) Duality for control problems. SIAM J Control 6:114– 120 Morton TE (1978) Universal planning horizons for generalized convex production scheduling. Oper Res 26:1046–1057 Morton TE, Mitchell A, Zemel E (1982) A discrete maximum principle approach to a general dynamic market response model. In: TIMS studies in the management sciences, vol 18. North-Holland, Amsterdam, pp 117–140 Moser E, Seidl A, Feichtinger G (2014) History-dependence in productionpollution-trade-oﬀ models: a multi-stage approach. Ann Oper Res 222(1):457–481 Motta M, Rampazzo F (1996) Dynamic programming for nonlinear systems driven by ordinary and impulsive controls. SIAM J Control Optim 34(1):199– 225 Muller E (1983) Trial/awareness advertising decision: a control problem with phase diagrams with non-stationary boundaries. J Econ Dyn Control 6:333– 350 Munro GR, Scott AD (1985) The economics of ﬁsheries management. In: Kneese AV, Sweeney JL (eds) Handbook of natural resource and energy economics, vol 2, 1st edn, chap 14. Elsevier, London, pp 623–676 Murata Y (1982) Optimal control methods for linear discrete-time economic systems. Springer, New York Murray DM, Yakowitz SJ (1984) Diﬀerential dynamic programming and Newton’s methods for discrete optimal control problems. J Optim Theory Appl 43(3):395–414 Muzicant J (1980) Systeme mit verteilten Parametern in der Bio¨okonomie: Ein Maximumprinzip zur Kontrolle altersstrukturierter Modelle, Diss., Inst. f. Unternehmensforschung, Techn. University Wien Nahorski Z, Ravn HF, Vidal RVV (1984) The discrete-time maximum principle: a survey and some new results. Int J Control 40:533–554 Naik PA, Mantrala MK, Sawyer A (1998) Planning pulsing media schedules in the presence of dynamic advertising quality. Market Sci 17(3):214–235

522

Bibliography

Naik PA, Prasad A, Sethi SP (2008) Building brand awareness in dynamic oligopoly markets. Manag Sci 54(1):129–138 N¨aslund B (1966) Simultaneous determination of optimal repair policy and service life. Swed J Econ 68:63–73 N¨aslund B (1969) Optimal rotation and thinning. For Sci 15:446–451 N¨aslund B (1979) Consumer behavior and optimal advertising. J Oper Res Soc 20:237–243 Neck R (1982) Dynamic systems with several decision-makers. In: Feichtinger G, Kall P (eds) Operations research in progress. D. Reidel Publishing Company, Dordrecht, pp 261–284 Neck R (1984) Stochastic control theory and operational research. Eur J Oper Res 17:283–301 Nelson RT (1960) Labor assignment as a dynamic control problem. Oper Res 14:369–376 Nepomiastchy P (1970) Application of optimal control theory and penalty function techniques to solution of a particular scheduling problem (in Russian). Zhurnal Vychiﬂitel’noi Matemtiki i Matematichefkoi Fiziki 10(4) Nerlove M, Arrow KJ (1962) Optimal advertising policy under dynamic conditions. Economica 29:129–142 Neustadt LW (1976) Optimization: a theory of necessary conditions. Princeton University Press, Princeton Nguyen D (1997) Marketing decisions under uncertainty. Kluwer Academic Publishers, Boston Norstr¨ om CJ (1978) The continuous wheat trading model reconsidered. An application of mathematical control theory with a state constraint, W.P. 58– 77-78, GSIA, Carnegie Mellon University, Pittsburgh Novak AJ, Feichtinger G (1993) Optimal treatment of cancer diseases. Int J Syst Sci 24:1253–1263 Oberle HJ (1979) Numerical computation of singular control problems with application to optimal heating and cooling by solar energy. Appl Math Optim 5:297–314 Oberle HJ (1986) Numerical solution of minimax optimal control problems by multiple shooting technique. J Optim Theory Appl 50:331–358

Bibliography

523

Oberle HJ, Grimm W (1989) BNDSCO: a program for the numerical solution of optimal control problems. Report 515, Institute for ﬂight systems dynamics, German aerospace research establishment DLR, Oberpfaﬀenhofen, Germany Oberle HJ, Sothmann B (1999) Numerical computation of optimal feed rates for a fed-batch fermentation model. J Optim Theory Appl 100(1):1–13 O˘guzt¨ oreli MN, Stein RB (1983) Optimal control or antagonistic muscles. Biol Cybern 48:91–99 Øksendal BK (1998) Stochastic diﬀerential equations: an introduction with applications. Springer, New York Olsder GJ (1976) Some thoughts about simple advertising models as diﬀerential games and the structure of coalitions. In: Ho YC, Mitter SK (eds) Directions in large-scale systems, many-person optimization and decentralized control. Plenum Press, New York, pp 187–205 Oniki H (1973) comparative dynamics (sensitivity analysis) in optimal control theory. J Econ Theory 6:265–283 Oren SS, Powell SG (1985) Optimal supply of a depletable resource with a backstop technology. Oper Res 33:277–292 Osayimwese I (1974) Rural-urban migration and control theory. Geogr Anal 4:147–161 Ozga S (1960) Imperfect markets through lack of knowledge. Quart J Econ 74:29–52 Palda KS (1964) The measurement of cumulative advertising eﬀects. PrenticeHall, Englewood Cliﬀs Pantoja JF, Mayne DQ (1991) Sequential quadratic programming algorithm for discrete optimal control problems with control inequality constraints. Int J Control 53(4):823–836 Parlar M (1983) Optimal forest ﬁre control with limited reinforcements. Optimal Control Appl Methods 4:185–191 Parlar M (1984) Optimal dynamic service rate control in time dependent M/M/S/N queues. Int J Syst Sci 15:107–118 Parlar M (1986) A problem in jointly optimal production and advertising decisions. Int J Syst Sci 17(9):1373–1380 Parlar M, Vickson RG (1980) An optimal control problem with piecewise quadratic cost functional containing a ’dead zone’. Optimal Control Appl Methods 1:361–372

524

Bibliography

Parlar M, Vickson RG (1982) Optimal forest ﬁre control: an extension of Park’s model. For Sci 28:345–355 Pauwels W (1977) Optimal dynamic advertising policies in the presence of continuously distributed time lags. J Optim Theory Appl 22(1):79–89 Pekelman D (1974) Simultaneous price-production decision. Oper Res 22:788– 794 Pekelman D (1975) Production smoothing with ﬂuctuating price. Manag Sci 21:576–590 Pekelman D (1979) On optimal utilization of production processes. Oper Res 27:260–278 Pekelman D, Rausser GC (1978) Adaptive control: survey of methods and applications. In: Bensoussan A et al (eds) Applied optimal control, TIMS studies in management sciences, vol 9. North-Holland, Amsterdam, pp 89– 120 Pekelman D, Sethi SP (1978) Advertising budgeting, wearout and copy replacement. J Oper Res Soc 29:651–659 Pepyne DL, Cassandras CG (1999) Performance optimization of a class of discrete event dynamic systems using calculus of variations techniques. J Optim Theory Appl 100(3):599–622 Perrakis S (1976) A note on optimal equity ﬁnancing of the corporation. J Financ Quant Anal 11(1):157–164 Pesch HJ (1989a) Real-time computation of feedback controls for constrained optimal control problems, part 1: neighboring extremals. Optimal Control Appl Methods 10(2):129–145 Pesch HJ (1989b) Real-time computation of feedback controls for constrained optimal control problems, part 2: a correction method based on multiple shooting. Optimal Control Appl Methods 10(2):147–171 Pesch HJ (1994) A practical guide to the solution of real-life optimal control problems. Control Cybern 23(1–2):7–60 Pesch HJ, Bulirsch R (1994) The maximum principle, Bellman’s equation, and Canath´eodory’s Work. J Optim Theory Appl 80(1):203–229 Pesch HJ, Plail M (2009) The maximum principle of optimal control: a story of ingenious ideas and missed opportunities. Control Cybern 38 (4A):973–995 Peterson DW (1973) The economic signiﬁcance of auxiliary functions in optimal control. Int Econ Rev 14:234–252

Bibliography

525

Peterson DW (1974) On sensitivity in optimal control problems. J Optim Theory Appl 13(1):56–73 Peterson FM, Fisher AC (1977) The exploitation of extractive resources: a survey. Econ J 87:681–721 Peterson DW, Zalkin JH (1978) A review of direct suﬃcient conditions in optimal control theory. Int J Control 28:589–610 Petrov IP (1968) Variational methods in optimum control theory. Academic Press, New York Pickenhain S (2010) On adequate transversality conditions for inﬁnite horizon optimal control problems - a famous example of Halkin. In: Cuaresma JC, Palokangas T, Taraysev A (eds) Dynamic systems, economic growth, and the environment. Dynamic modeling and econometrics in economics and ﬁnance, vol 12. Springer, Berlin, pp 3–22 Pickenhain S, Lykina V (2006) Suﬃciency conditions for inﬁnite horizon optimal control problems. In: Seeger A (ed) Recent advances in optimization. Lecture notes in economics and mathematical systems, vol 563. Springer, Berlin, pp 217–232 Pickenhain S, Lykina V, Wagner M (2008) On the lower semicontinuity of functionals involving Lebesgue or improper Riemann integrals in inﬁnite horizon optimal control problems. Control Cybern 37:451–468 Pierskalla WP, Voelker JA (1976) Survey of maintenance models: the control and surveillance of deteriorating systems. Naval Res Logist Quart 23:353–388 Pindyck RS (ed) (1978a) Advances in the economics of energy and resources, vol II. J.A.I. Press, Greenwich Pindyck RS (1978b) The optimal exploration and production of nonrenewable resources. J Polit Econ 86:841–862 Pindyck RS (1978c) Gains to producers from the cartelization of exhaustible resources. Rev. Econ Stat 60:238–251 Pindyck RS (1982) Adjustment costs, uncertainty, and the behavior of the ﬁrm. Am Econ Rev 72(3):415–427 Pitchford JD, Turnovsky SJ (eds) (1977) Applications of control theory to economic analysis. North-Holland, Amsterdam Pohjola M (1984) Threats and bargaining in capitalism: a diﬀerential game view. J Econ Dyn Control 8:291–302

526

Bibliography

Polak E (1971) Computational methods in optimization. Academic Press, New York Polak E (1973) A historical survey of computational methods in optimal control. SIAM Rev 15:553–584 Polak E, Yang TH, Mayne DQ (1993) A method of centers based on barrier functions for solving optimal control problems with continuous state and control constraints. SIAM J Control Optim 31:159–179 Polyanin AD, Zaitsev VF (2003) Handbook of exact solutions for ordinary differential equations, 2nd edn. Chapman & Hall/CRC, Boca Raton Pontryagin LS, Boltyanskii VG, Gamkrelidze RV, Mischenko EF (1962) The mathematical theory of optimal processes. Wiley, New York Prasad A, Sethi SP (2004) Competitive advertising under uncertainty: stochastic diﬀerential game approach. J Optim Theory Appl 123(1):163–185 Prasad A, Sethi SP (2009) Integrated marketing communications in markets with uncertainty and competition. Automatica 45(3):601–610 Prasad A, Sethi SP, Naik P (2012) Understanding the impact of churn in dynamic oligopoly markets. Automatica 48:2882–2887 Presman E, Sethi SP (2006) Inventory models with continuous and Poisson demands and discounted and average costs. Prod Oper Manag 15(2):279– 293 Presman E, Sethi SP, Zhang Q (1995) Optimal feedback production planning in a stochastic N -machine ﬂowshop. Automatica 31(9):1325–1332 Presman E, Sethi SP, Suo W (1997a) Optimal feedback production planning in stochastic dynamic jobshops. In: Yin G, Zhang Q (eds) Mathematics of stochastic manufacturing systems. Lectures in applied mathematics, vol 33. American Mathematical Society, Providence, pp 235–252 Presman E, Sethi SP, Suo W (1997b) Existence of optimal feedback production plans in stochastic N -machine ﬂowshops with limited buﬀers. Automatica 33(4):1899–1903 Presman E, Sethi SP, Zhang H, Zhang Q (1998a) Analysis of average cost optimality for an unreliable two-machine ﬂowshop. In: Proceedings of the fourth international conference on optimization techniques and applications, Curtin University of Technology, Perth, Australia, pp 94–112 Presman E, Sethi SP, Zhang H, Zhang Q (1998b) Optimality of zero-inventory policies for an unreliable manufacturing system producing two part types. Dyn Contin Discrete Impuls Syst 4(4):485–496

Bibliography

527

Prskawetz A, Feichtinger G, Luptacik M (1998) The accomplishment of the Maastricht criteria with respect to initial debt. J Econ 68:93–110 Pytlak R, Vinter RB (1993) PH2SOL solver: an O(N) implementation of an optimization algorithm for a general optimal control problem. Research report C93-36, Centre for Process Systems Engineering, Imperial College, London Pytlak R, Vinter RB (1999) Feasible direction algorithm for optimal control problems with state and control constraints: implementation. J Optim Theory Appl 101(3):623–649 Raman K (1990) Stochastically optimal advertising policies under dynamic conditions: the ratio rule. Optimal Control Appl Methods 11:283–288 Raman K, Chatterjee R (1995) Optimal monopolist pricing under demand uncertainty in dynamic markets. Manag Sci 41(1):144–162 Ramsey FP (1928) A mathematical theory of savings. Econ J 38:543–559 Rao RC (1970) Quantitative theories in advertising. Wiley, New York Rao RC (1984) Advertising decisions in oligopoly: an industry equilibrium analysis. Optimal Control Appl Methods 5(4):331–344 Rao RC (1985) A note on optimal and near optimal price and advertising strategies. Manag Sci 31(3):376–377 Rao RC (1986) Estimating continuous time advertising-sales models. Market Sci 5(2):125–142 Rao RC (1990) Impact of competition on strategic marketing decisions. In: Day G, Weitz B, Wensley R (eds) Interface of marketing and strategy. JAI Press, Greenwich Rapoport A (1980) Mathematische Methoden in den Sozialwissenschaften. Physica-Verlag, W¨ urzburg Rapp B (1974) Models for optimal investment and maintenance decisions. Almqvist & Wiksell/Wiley, Stockholm/New York Rausser GC, Hochman E (1979) Dynamic agricultural systems: economic prediction and control. North-Holland, New York Raviv A (1979) The design of an optimal insurance policy. Am Econ Rev 69:84– 96 Ravn HF (1999) Discrete time optimal control. PhD thesis, Technical University of Denmark

528

Bibliography

Ray A, Blaqui`ere A (1981) Suﬃcient conditions for optimality of threat strategies in a diﬀerential game. J Optim Theory Appl 33:99–109 Reinganum JF (1982) A dynamic game of R & D: patent protection and competitive behavior. Econometrica 50:671–688 Rempala R (1980) On the multicommodity Arrow-Karlin inventory model. In: Proceedings ﬁrst international symposium on inventories, Hungarian Academy of Science, Budapest Rempala R (1982) On the multicommodity Arrow-Karlin inventory model, part II: horizon and horizontal solution. In: Proceedings second international symposium on inventories, Hungarian Academy of Science, Budapest Rempala R (1986) Horizon for the dynamic family of wheat trading problems. In: Proceedings fourth international symposium on inventories, Budapest Rempala R, Sethi SP (1988) Forecast horizons in single product inventory models. In: Feichtinger G (ed) Optimal control theory and economic analysis, vol 3. North-Holland, Amsterdam, pp 225–233 Rempala R, Sethi SP (1992) Decision and forecast horizons for one-dimensional optimal control problems: existence results and applications. Optimal Control Appl Methods 13:179–192 Rempala R, Zabczyk J (1988) On the maximum principle for deterministic impulse control problems. J Optim Theory Appl 59(2):281–288 Richard SF (1979) A generalized capital asset pricing model. TIMS Stud Manag Sci 11:215–232 Ringbeck J (1985) Mixed quality and advertising strategies under asymmetric information. In: Feichtinger G (ed) Optimal control theory and economic analysis, vol 2. North-Holland, Amsterdam, pp 197–214 Rishel RW (1965) An extended Pontryagin principle for control systems whose control laws contain measures. J Soc Ind Appl Math Control 3:191–205 Rishel RW (1985) A partially observed advertising model. In: Feichtinger G (ed) Optimal control theory and economic analysis, vol 2. North Holland, Amsterdam, pp 253–262 Roberts SM, Shipman JS (1972) Two-point boundary value problems: shooting methods. Elsevier, New York Robinson B, Lakhani C (1975) Dynamic price models for new-product planning. Manag Sci 21(10):1113–1122

Bibliography

529

Robson AJ (1981) Suﬃciency of the Pontryagin conditions for optimal control when the time horizon is free. J Econ Theory 24:438–445 Robson AJ (1985) Optimal control of systems governed by partial diﬀerential equations: economic applications. In: Feichtinger G (ed) Optimal control theory and economic analysis, vol 2. North-Holland, Amsterdam, pp 105– 118 Rockafellar RT (1970) Convex analysis. Princeton University Press, Princeton Rockafellar RT (1972) State constraints in convex control problems of Bolza. SIAM J Control 19:691–715 Russak B (1976) Relations among the multipliers for problems with bounded state constraints. SIAM J Control Optim 14:1151–1155 Russell DL (1965) Penalty functions and bounded phase coordinates. SIAM J Control 2:409–422 Sage AP (1968) Optimum systems control. Prentice-Hall, Englewood Cliﬀs Salukvadze ME (1979) Vector-valued optimization problems in control theory. Academic Press, New York Samaratunga C, Sethi SP, Zhou X (1997) Computational evaluation of hierarchical production policies for stochastic manufacturing systems. Oper Res 45(2):258–274 Samuelson PA (1972) The general saddle point property of optimal-control motions. J Econ Theory 5:102–120 Sarma VVS, Alam M (1975) Optimal maintenance policies for machines subject to deterioration and intermittent breakdowns. IEEE Trans Syst Man Cybern SMC-5:396–398 Sasieni M (1989) Optimal advertising strategies. Market Sci 8:358–370 Scalzo RC (1974) N person linear quadratic diﬀerential games with constraints. SIAM J Control 12:419–425 Schaefer MB (1957) Some considerations of population dynamics and economics in relation to the management of marine ﬁsheries. J Fish Res Board Can 14:669–681 Schilling K (1985) On optimization principles in plant ecology. In: Aubin JP et al (eds) Dynamics of macrosystems. Lecture notes in economics and mathematical systems, vol 257. Springer, Berlin, pp 63–71

530

Bibliography

Seidl A, Kaplan EH, Caulkins JP, Wrzaczek S, Feichtinger G (2016) Optimal control of a terror queue model. Eur J Oper Res 248(1):246–256 Seidman TI, Sethi SP, Derzko NA (1987) Dynamics and optimization of a sales advertising model with population migration. J Optim Theory Appl 52(3):443–462 Seierstad A (1982) Diﬀerentiability properties of the optimal value function in control theory. J Econ Dyn Control 4:303–310 Seierstad A (1985) Existence of an optimal control with sparse jumps in the state variable. J Optim Theory Appl 45:265–293 Seierstad A, Sydsæter K (1977) Suﬃcient conditions in optimal control theory. Int Econ Rev 18:367–391 Seierstad A, Sydsæter K (1983) Suﬃcient conditions applied to an optimal control problem of resource management. J Econ Theory 31:375–382 Seierstad A, Sydsæter K (1987) Optimal control theory with economic applications. North-Holland, Amsterdam Selten R (1975) Reexamination of the perfectness concept for equilibrium points in extensive games. Int J Game Theory 4:25–55 Sethi SP (1973a) Optimal control of the Vidale-Wolfe advertising model. Oper Res 21:998–1013 Sethi SP (1973b) Simultaneous optimization of preventive maintenance and replacement policy for machines: a modern control theory approach. AIIE Trans 5(2):156–163 Sethi SP (1973c) An application of optimal control theory in forest management. J Manag Sci Appl Cybern 2:9–16 Sethi SP (1973d) A note on modeling simple dynamic cash balance problems. J Financ Quant Anal 8:685–687; Errata, 13, 585–586 Sethi SP (1974a) Suﬃcient conditions for the optimal control of a class of systems with continuous lags. J Optim Theory Appl 13:542–552 Sethi SP (1974b) Some explanatory remarks on the optimal control of the Vidale-Wolfe advertising model. Oper Res 22:1119–1120 Sethi SP (1974c) Quantitative guidelines for communicable disease control program: a complete synthesis. Biometrics 30:681–691 Sethi SP (1975) Optimal control of a logarithmic advertising model. Oper Res Quart 26:317–319

Bibliography

531

Sethi SP (1977a) Dynamic optimal control models in advertising: a survey. SIAM Rev 19(4):685–725 Sethi SP (1977b) Nearest feasible paths in optimal control problems: theory, examples, and counterexamples. J Optim Theory Appl 23(4):563–579 Sethi SP (1977c) Optimal advertising for the Nerlove-Arrow model under a budget constraint. Oper Res Quart 28(3):683–693 Sethi SP (1977d) A linear bang-bang model of ﬁrm behavior and water quality. IEEE Trans Autom Control AC-22:706–714 Sethi SP (1978a) A survey of management science applications of the deterministic maximum principle. TIMS Stud Manag Sci 9:33–67 Sethi SP (1978b) Optimal equity ﬁnancing model of Krouse and Lee: corrections and extensions. J Financ Quant Anal 13(3):487–505 Sethi SP (1978c) Optimal quarantine programs for controlling an epidemic spread. J Oper Res Soc 29(3):265–268 Sethi SP (1979a) Optimal depletion of exhaustible resources. Appl Math Model 3:367–378 Sethi SP (1979b) Optimal pilfering policy for dynamic continuous thieves. Manag Sci 25(6):535–542 Sethi SP (1979c) Optimal advertising policy with the contagion model. J Optim Theory Appl 29(4):615–627 Sethi SP (1979d) A note on the Nerlove-Arrow model under uncertainty. Oper Res 27(4):839–842; Erratum (1980) 28(4):1026–1027 Sethi SP (1983a) Optimal long-run equilibrium advertising level for the Blattberg-Jeuland model. Manag Sci 29(12):1436–1443 Sethi SP (1983b) Deterministic and stochastic optimization of a dynamic advertising model. Optimal Control Appl Methods 4(2):179–184 Sethi SP (1983c) Applications of optimal control to management science problems. In: Proceedings of the 1983 World Conference on Systems, Caracas, Venezuela Sethi SP (1984) Application of the maximum principle to production and inventory problems. In: Proceedings third international symposium on inventories, Budapest, pp 753–756 Sethi SP (1990) Decision and forecast horizons in dynamic optimization. In: Singh MG (ed) Systems and control encyclopedia supplementary, vol 1. Pergamon Press, Oxford, pp 192–198

532

Bibliography

Sethi SP (1996) When does the share price equal the present value of future dividends? - a modiﬁed dividend approach. Econ Theory 8:307–319 Sethi SP (1997a) Optimal consumption and investment with bankruptcy. Kluwer Academic Publishers, Boston Sethi SP (1997b) Some insights into near-optimal plans for stochastic manufacturing systems. In: Yin G, Zhang Q (eds) Mathematics of stochastic manufacturing systems. Lectures in applied mathematics, vol 33. American Mathematical Society, Providence, pp 287–315 Sethi SP (1998) Optimal consumption - investment decisions allowing for bankruptcy: a survey. In: Ziemba WT, Mulvey JM (eds) Worldwide asset and liability modeling. Cambridge University Press, Cambridge, pp 387–426 Sethi SP (2010) i3: incomplete information inventory models. Decision line, October, 16–19. Reprinted in Proceedings of the ﬁrst international conference on best practices in supply chain management, Bhubaneswar, India, 22–23 November 2012, pp 658–662 Sethi SP, Bass FM (2003) Optimal pricing in a hazard rate model of demand. Optimal Control Appl Methods 24:183–196 Sethi SP, Chand S (1979) Planning horizon procedures in machine replacement models. Manag Sci 25:140–151; Erratum (1980) 26(3):342 Sethi SP, Chand S (1981) Multiple ﬁnite production rate dynamic lot size inventory models. Oper Res 29(5):931–944 Sethi SP, Lee SC (1981) Optimal advertising for the Nerlove-Arrow Model under a replenishable budget. Optimal Control Appl Methods 2(2):165–173 Sethi SP, Lehoczky JP (1981) A comparison of Ito and Stratonovich formulations of problems in ﬁnance. J Econ Dyn Control 3:343–356 Sethi SP, Li T (2017) A review of dynamic Stackelberg game models. Discrete Contin Dyn Syst Ser B 22(1):125–159 Sethi SP, McGuire TW (1977) Optimal skill mix: an application of the maximum principle for systems with retarded controls. J Optim Theory Appl 23:245–275 Sethi SP, Morton TE (1972) A mixed optimization technique for the generalized machine replacement problem. Naval Res Logist Quart 19:471–481 Sethi SP, Sorger G (1989) Concepts of forecast horizons in stochastic dynamic games. In: Proceedings of the 28th conference on decision and control, Tampa, Florida, pp 195–197

Bibliography

533

Sethi SP, Sorger G (1990) An exercise in modeling of consumption, import, and export of an exhaustible resource. Optimal Control Appl Methods 11:191–196 Sethi SP, Sorger G (1991) A theory of rolling horizon decision making. Ann O R 29:387–416 Sethi SP, Staats PW (1978) Optimal control of some simple deterministic epidemic models. J Oper Res Soc 29(2):129–136 Sethi SP, Taksar MI (1988) Deterministic and stochastic control problems with identical optimal cost functions. In: Bensoussan A, Lions JL (eds) Analysis and optimization of systems. Lecture notes in control and information sciences. Springer, New York, pp 641–645 Sethi SP, Taksar MI (1990) Deterministic equivalent for a continuous-time linear-convex stochastic control problem. J Optim Theory Appl 64(1):169– 181 Sethi SP, Taksar MI (2002) Optimal ﬁnancing of a corporation subject to random returns. Math Financ 12(2):155–172 Sethi SP, Thompson GL (1970) Applications of mathematical control theory to ﬁnance. J Financ Quant Anal 5(4–5):381–394 Sethi SP, Thompson GL (1977) Christmas toy manufacturer’s problem: an application of the stochastic maximum principle. Opsearch 14:161–173 Sethi SP, Thompson GL (1981a) Simple models in stochastic production planning. In: Bensoussan A, Kleindorfer PR, Tapiero CS (eds) Applied stochastic control in econometrics and management science. North-Holland, Amsterdam, pp 295–304 Sethi SP, Thompson GL (1981b) A tutorial on optimal control theory. Inf Syst Oper Res 19(4):279–291 Sethi SP, Thompson GL (1982) Planning and forecast horizons in a simple wheat trading model. In: Feichtinger G, Kall P (eds) Operations research in progress. D. Reidel Publishing Company, Dordrecht, pp 203–214 Sethi SP, Zhang Q (1994a) Hierarchical decision making in stochastic manufacturing systems. Systems and control: foundations and applications. Birkh¨auser, Boston Sethi SP, Zhang Q (1994b) Hierarchical controls in stochastic manufacturing systems. SIAG/CST Newsl 2(1):1–5 Sethi SP, Zhang Q (1994c) Asymptotic optimal controls in stochastic manufacturing systems with machine failures dependent on production rates. Stochastics Stochastics Rep 48:97–121

534

Bibliography

Sethi SP, Zhang Q (1995a) Hierarchical production and setup scheduling in stochastic manufacturing systems. IEEE Trans Autom Control 40(5):924– 930 Sethi SP, Zhang Q (1995b) Multilevel hierarchical decision making in stochastic marketing - production systems. SIAM J Control Optim 33(2):538–553 Sethi SP, Zhang Q (1998a) Asymptotic optimality of hierarchical controls in stochastic manufacturing systems: a review. In: Aronson JE, Zionts S (eds) Operations research: methods, models, and applications. Quoram Books, Westport, pp 267–294 Sethi SP, Zhang Q (1998b) Optimal control of dynamic systems by decomposition and aggregation. J Optim Theory Appl 99(1):1–22 Sethi SP, Zhang H (1999a) Average-cost optimal policies for an unreliable ﬂexible multiproduct machine. Int J Flex Manuf Syst 11:147–157 Sethi SP, Zhang H (1999b) Hierarchical production controls for a stochastic manufacturing system with long-run average cost: asymptotic optimality. In: McEneaney WM , Yin G, Zhang Q (eds) Stochastic analysis, control, optimization and applications: a volume in honor of W.H. Fleming. Birkh¨auser, Boston, pp 621–637 Sethi SP, Zhang Q (2001) Near optimization of stochastic dynamic systems by decomposition and aggregation. Optimal Control Appl Methods 22:333–350 Sethi SP, Zhang Q (2004) Problem 4.3 feedback control in ﬂowshops. In: Blondel VD, Megretski A (eds) Unsolved problems in mathematical systems and control theory. Princeton University Press, Princeton, pp 140–143 Sethi SP, Zhou X (1996a) Optimal feedback controls in deterministic dynamic two-machine ﬂowshops. O R Lett 19(5):225–235 Sethi SP, Zhou X (1996b) Asymptotic optimal feedback controls in stochastic dynamic two-machine ﬂowshops. In: Yin G, Zhang Q (eds) Recent advances in control and optimization of manufacturing systems. Lecture notes in control and information sciences, vol 214. Springer, New York, pp 147–180 Sethi SP, Thompson GL, Udayabhanu V (1985) Proﬁt maximization models for exponential decay processes. Eur J Oper Res 2:101–115 Sethi SP, Suo W, Taksar MI, Zhang Q (1997a) Optimal production planning in a stochastic manufacturing system with long-run average cost. J Optim Theory Appl 92(1):161–188 Sethi SP, Zhang H, Zhang Q (1997b) Hierarchical production control in a stochastic manufacturing system with long-run average cost. J Math Anal Appl 214:151–172

Bibliography

535

Sethi SP, Zhang Q, Zhou X (1997c) Hierarchical production controls in a stochastic two-machine ﬂowshop with a ﬁnite internal buﬀer. IEEE Trans Robot Autom 13(1):1–13 Sethi SP, Suo W, Taksar MI, Yan H (1998a) Optimal production planning in a multiproduct stochastic manufacturing systems with long-run average cost. Discrete Event Dyn Syst 8(1):37–54 Sethi SP, Zhang H, Zhang Q (1998b) Minimum average-cost production planning in stochastic manufacturing systems. Math Models Methods Appl Sci 8(7):1251–1276 Sethi SP, Zhang H, Zhang Q (2000) Optimal production rates in a deterministic two-product manufacturing system. Optimal Control Appl Methods 21:125– 135 Sethi SP, Yan H, Zhang H, Zhang Q (2002) Optimal and hierarchical controls in dynamic stochastic manufacturing systems: a survey. Manuf Serv Oper Manag 4(2):133–170 Sethi SP, Zhang H, Zhang Q (2005) Average-cost control of stochastic manufacturing systems. Stochastic modelling and applied probability. Springer, New York Sethi SP, Prasad A, He X (2008a) Optimal advertising and pricing in a newproduct adoption model. J Optim Theory Appl 139(2):351–360 Sethi SP, Yeh DHM, Zhang R, Jardine A (2008b) Optimal maintenance and replacement of extraction machinery. J Syst Sci Syst Eng 17(4):416–431 Shani U, Tsur Y, Zemel A (2005) Characterizing dynamic irrigation policies via Green’s theorem. In: Deissenberg C, Hartl RF (eds) Optimal control and dynamic games. Springer, Dordrecht, pp 105–117 Shapiro C (1982) Consumer information, product quality, and seller reputation. Bell J Econ 13:20–25 Shapiro A (1997) On uniqueness of Lagrange multipliers in optimization problems subject to cone constraints. SIAM J Optim 7(2):508–518 Sharomi O, Malik T (2017) Optimal control in epidemiology. Ann Oper Res 251:55–71 Shell K (ed) (1967) Essays on the theory of optimal economic growth. The MIT Press, Cambridge Shell K (1969) Applications of Pontryagin’s maximum principle to economics. In: Kuhn HW, Szego GP (eds) Mathematical systems theory and economics. Springer, Berlin

536

Bibliography

Siebert H (1985) Economics of the resource-exporting country: intertemporal theory of supply and trade. JAI Press, Greenwich Silva GN, Vinter RB (1997) Necessary conditions for optimal impulsive control problems. SIAM J Control Optim 35(6):1829–1846 Simaan M, Cruz JB Jr (1975) Formulation of Richardson’s model of arms race from a diﬀerential game viewpoint. Rev Econ Stud 42:67–77 Simon HA (1956) Dynamic programming under uncertainty with a quadratic criterion function. Econometrica 24:74–81 Simon HA (1982) ADPULS: an advertising model with wearout and pulsation. J Market Res 19:352–363 Singh MG (1980) Dynamical hierarchical control. North-Holland, Amsterdam Singh MG, Titli A, Malanowski K (1985) Decentralized control design: an overview. Large Scale Syst 9:215–230 Singhal K (1992) A noniterative algorithm for the multiproduct production and work force planning problem. Oper Res 40(3):620–625 Singhal K, Singhal J (1986) A solution to the Holt et al. model for aggregate production planning. OMEGA 14:502–505 Skiba AK (1978) Optimal growth with a convex-concave production function. Econometrica 46:527–539 Smith VL (1972) Dynamics of waste accumulation: disposal versus recycling. Quart J Econ 86:600–616 Snower DJ (1982) Macroeconomic policy and the optimal destruction of vampires. J Polit Econ 90:647–655 Solow RM (1974) Intergenerational equity and exhaustible resources. Rev Econ Stud 41:29–45 Solow RM, Wan FY (1976) Extraction costs in the theory of exhaustible resources. Bell J Econ 7:359–370 Sorger G (1989) Competitive dynamic advertising: a modiﬁcation of the case game. J Econ Dyn Control 13:55–80 Sorger G (1995) Discrete-time dynamic game models for advertising competition in a duopoly. Optimal Control Appl Methods 16:175–188 Sorger G (1998) Markov-perfect Nash equilibria in a class of resource games. Econ Theory 11:79–100

Bibliography

537

Sorger G (2002) Existence and characterization of time-consistent monetary policy rules. In: Zaccour G (ed) Optimal control and diﬀerential games. Kluwer Academic Publishers, Boston, pp 87–103 Sorger G (2005) A dynamic common property resource problem with amenity value and extraction costs. Int J Econ Theory 1:3–19 Southwick L, Zionts S (1974) An optimal-control-theory approach to the education-investment decision. Oper Res 22:1156–1174 Spence M (1981) The learning curve and competition. Bell J Econ 12:49–70 Spiegel MR (1971) Schaum’s outline series: theory and problems of calculus of ﬁnite diﬀerences and diﬀerence equations. McGraw-Hill, New York Spiegel MR, Lipschutz S, Liu J (2008) Schaum’s outline of mathematical handbook of formulas and tables, 3rd edn (Schaum’s Outline Series). McGraw-Hill, New York Spremann K (1985) The signaling of quality by reputation. In: Feichtinger G (ed) Optimal control theory and economic analysis, vol 2. North-Holland, Amsterdam, pp 235–252 Sprzeuzkouski AY (1967) A problem in optimal stock management. J Optim Theory Appl 1:232–241 Srinivasan V (1976) Decomposition of a multi-period media scheduling model in terms of single period equivalents. Manag Sci 23:349–360 Stalford H, Leitmann G (1973) Suﬃciency conditions for Nash equilibria in Nperson diﬀerential games. In: Blaqui`ere A (ed) Topics in diﬀerential games. North-Holland, Amsterdam, pp 345–376 Starr AW, Ho YC (1969) Nonzero-sum diﬀerential games. J Optim Theory Appl 3:184–206 Steindl A, Feichtinger G (2004) Bifurcations to periodic solutions in a production/inventory model. J Nonlinear Sci 14:469–503 Steindl A, Feichtinger G, Hartl RF, Sorger G (1986) On the optimality of cyclical employment policies: a numerical investigation. J Econ Dyn Control 10:457–466 Stepan A (1977) An application of the discrete maximum principle to designing ultradeep drills - the casing optimization. Int J Prod Res 15:315–327 Stern LE (1984) Criteria of optimality in the inﬁnite-time optimal control problem. J Optim Theory Appl 44:497–508

538

Bibliography

Stiglitz JE, Dasgupta P (1982) Market structure and the resource depletion: a contribution to the theory of intertemporal monopolistic competition. J Econ Theory 28:128–164 Stoer J, Bulirsch R (1978) Einf¨ uhrung in die Numerische Mathematik II, Heidelberger Taschenb¨ ucher, vol 114. Springer, Berlin St¨ oppler S (1975) Dynamische Produktionstheorie. Westdeutscher-Verlag, Opladen St¨ oppler S (1985) Der Einﬂußder Lagerkosten auf die Produktionsanpassung bei zyklischem Absatz - Eine kontrolltheoretische Analyse. OR-Spektr 7:129–142 Sulem A (1986) Solvable one-dimensional model of a diﬀusion inventory system. Math Oper Res 11(1):125–133 Swan GW (1984) Applications of optimal control theory in biomedicine. M. Dekker, New York Sweeney DJ, Abad PL, Dornoﬀ RJ (1974) Finding an optimal dynamic advertising policy. Int J Syst Sci 5(10):987–994 Sydsæter K (1978) Optimal control theory and economics: some critical remarks on the literature. Scand J Econ 80:113–117 Sydsæter K (1981) Topics in mathematical analysis for economists. Academic Press, London Takayama A (1974) Mathematical economics. The Dryden Press, Hinsdale Tan KC, Bennett RJ (1984) Optimal control of spatial systems. George Allen & Unwin, London Tapiero CS (1971) Optimal simultaneous replacement and maintenance of a machine with process discontinuities. Revue francaise d’informatique et recherche op´erationnelle 2:79–86 Tapiero CS (1973) Optimal maintenance and replacement of a sequence of machines and technical obsolescence. Opsearch 19:1–13 Tapiero CS (1977) Managerial planning: an optimum and stochastic control approach. Gordon Breach, New York Tapiero CS (1978) Optimum advertising and goodwill under uncertainty. Oper Res 26(3):450–463 Tapiero CS (1981) Optimum product quality and advertising. Inf Syst Oper Res 19(4):311–318

Bibliography

539

Tapiero CS (1982a) Optimum control of a stochastic model of advertising. In: Feichtinger G (ed) Optimal control theory and economic analysis. NorthHolland, Amsterdam, pp 287–300 Tapiero CS (1982b) A stochastic model of consumer behavior and optimal advertising. Manag Sci 28(9):1054–1064 Tapiero CS (1983) Stochastic diﬀusion models with advertising and word-ofmouth eﬀects. Eur J Oper Res 12(4):348–356 Tapiero CS (1988) Applied stochastic models and control in management. North-Holland, Amsterdam Tapiero CS, Farley JU (1975) Optimal control of sales force eﬀort in time. Manag Sci 21(9):976–985 Tapiero CS, Farley JU (1981) Using an uncertainty model to assess sales response to advertising. Decis Sci 12:441–455 Tapiero CS, Soliman MA (1972) Multi-commodities transportation schedules over time. Networks 2:311–327 Tapiero CS, Venezia I (1979) A mean variance approach to the optimal machine maintenance and replacement problem. J Oper Res Soci 30:457–466 Tapiero CS, Zuckermann D (1983) Optimal investment policy of an insurance ﬁrm. Insurance Math Econ 2:103–112 Tapiero CS, Eliashberg J, Wind Y (1987) Risk behaviour and optimum advertising with a stochastic dynamic sales response. Optimal Control Appl Methods 8(3):299–304 Taylor JG (1972) Comments on a multiplier condition for problems with state variable inequality constraints. IEEE Trans Autom Control AC-12:743–744 Taylor JG (1974) Lanchester-type models of warfare and optimal control. Naval Res Logist Quart 21:79–106 Teng JT, Thompson GL (1983) Oligopoly models for optimal advertising when production costs obey a learning curve. Manag Sci 29(9):1087–1101 Teng JT, Thompson GL (1985) Optimal strategies for general price-advertising models. In: Feichtinger G (ed) Optimal control theory and economic analysis, vol 2. North-Holland, Amsterdam, pp 183–195 Teng JT, Thompson GL, Sethi SP (1984) Strong decision and forecast horizons in a convex production planning problem. Optimal Control Appl Methods 5(4):319–330

540

Bibliography

Teo KL, Moore EJ (1977) Necessary conditions for optimality for control problems with time delays appearing in both state and control variables. J Optim Theory Appl 23:413–427 Teo KL, Goh CJ, Wong KH (1991) A uniﬁed computational approach to optimal control problems. Longman Scientiﬁc & Technical, Essex Terborgh G (1949) Dynamic equipment policy. McGraw-Hill, New York Th´epot J (1983) Marketing and investment policies of duopolists in a growing industry. J Econ Dyn Control 5:387–404 Thompson GL (1968) Optimal maintenance policy and sale date of a machine. Manag Sci 14:543–550 Thompson GL (1981) An optimal control model of advertising pulsation and wearout. In: Keon JW (ed) ORSA/TIMS special interest conference on market measurement and analysis. The Institute for Management Sciences, Providence, pp 34–43 Thompson GL (1982a) Continuous expanding and contracting economies. In: Deistler M, Furst E, Schwodiauer G (eds) Games, economic dynamics, time series analysis. Physica-Verlag, Berlin, pp 145–153 Thompson GL (1982b) Many country continuous expanding open economies with multiple goods. In: Feichtinger G (ed) Optimal control theory and economic analysis. North-Holland, New York, pp 157–168 Thompson GL, Sethi SP (1980) Turnpike horizons for production planning. Manag Sci 26:229–241 Thompson GL, Teng JT (1984) Optimal pricing and advertising policies for new product oligopoly models. Market Sci 3(2):148–168 Thompson GL, Sethi SP, Teng JT (1984) Strong planning and forecast horizons for a model with simultaneous price and production decisions. Eur J Oper Res 16:378–388 Tidball M, Zaccour G (2009) A diﬀerential environmental game with coupling constraints. Optimal Control Appl Methods 30:197–207 Tintner G (1937) Monopoly over time. Econometrica 5:160–170 Tintner G, Sengupta JK (1972) Stochastic economics: stochastic processes, control, and programming. Academic Press, New York Tolwinski B (1982) A concept of cooperative equilibrium for dynamic games. Automatica 18:431–447

Bibliography

541

Tousssaint S (1985) The transversality condition at inﬁnity applied to a problem of optimal resource depletion. In: Feichtinger G (ed) Optimal control theory and economic analysis, vol 2. North-Holland, Amsterdam, pp 429–440 Tracz GS (1968) A selected bibliography on the application of optimal control theory to economic and business systems, management science and operations research. Oper Res 16:174–186 Tragler G, Caulkins JP, Feichtinger G (2001) Optimal dynamic allocation of treatment and enforcement in illicit drug control. Oper Res 49(3):352–362 Treadway AB (1970) Adjustment costs and variable inputs in the theory of the competitive ﬁrm. J Econ Theory 2:329–347 Troch I (ed) (1978) Simulation of control systems with special emphasis on modelling and redundancy. Proceedings of the IMACS Symposium, NorthHolland, Amsterdam Tsurumi H, Tsurumi Y (1971) Simultaneous determination of market share and advertising expenditure under dynamic conditions: the case of a ﬁrm within the Japanese pharmaceutical industry. Econ Stud Quart 22:1–23 Tu PNV (1969) Optimal educational investment program in an economic planning model. Can J Econ 2:52–64 Tu PNV (1984) Introductory optimization dynamics. Springer, Berlin Turner RE, Neuman CP (1976) Dynamic advertising strategy: a managerial approach. J Bus Admin 7:1–21 Turnovsky SJ (1981) The optimal intertemporal choice of inﬂation and unemployment. J Econ Dyn Control 3:357–384 Tzafestas SG (1982a) Optimal and modal control of production-inventory systems. In: Optimization and control of dynamic operational research models. North-Holland, Amsterdam, pp 1–71 Tzafestas SG (1982b) Distributed parameter control systems: theory and application. In: International series on systems and control. Pergamon Press, Oxford Uhler RS (1979) The rate of petroleum exploration and extraction. In: Pindyck RS (ed) Advances in the economics of energy and resources, vol 2. JAI Press, Greenwich, pp 93–118 Valentine FA (1937) The problem of Lagrange with diﬀerential inequalities as added side conditions. In: Contributions to the theory of the calculus of variations, vol 1933–1937. University Chicago Press, Chicago

542

Bibliography

Van Hilten O, Kort PM, Van Loon PJJM (1993) Dynamic policies of the ﬁrm: an optimal control approach. Springer, New York Vanthienen L (1975) A simultaneous price-reduction decision making with production adjustment costs. In: Proceedings XX international meeting of TIMS held in Tel Aviv, Israel on 24–29 June 1973, vol 1. Jerusalem Academic Press, Jerusalem, pp 249–254 van Loon PJJM (1983) A dynamic theory of the ﬁrm: production, ﬁnance and investment. Lecture notes in economics and mathematical systems, vol 218. Springer, Berlin van Schijndel GJCT (1986) Dynamic shareholder behaviour under personal taxation: a note. In: Streitferdt L et al (eds) Operations research proceedings 1985. Springer, Berlin, pp 488–495 Varaiya PP (1970) N-person nonzero-sum diﬀerential games with linear dynamics. SIAM J Control 8:441–449 Veliov VM (2008) Optimal control of heterogeneous systems: basic theory. J Math Anal Appl 346:227–242 Verheyen PA (1985) A dynamic theory of the ﬁrm and the reaction on governmental policy. In: Feichtinger G (ed) Optimal control theory and economic analysis, vol 2. North-Holland, Amsterdam, pp 313–329 Verheyen P (1992) The jump in models with irreversible investments. In: Feichtinger G (ed) Dynamic economic models and optimal control, vol 4. Elsevier Science, Amsterdam, pp 75–89 Vickson RG (1981) Schedule control for randomly drifting production sequences. Inf Syst Oper Res 19:330–346 Vickson RG (1982) Optimal control production sequences: a continuous parameter analysis. Oper Res 30:659–679 Vickson RG (1985) Optimal conversion to a new production technology under learning. IEE Trans 17:175–181 Vickson RG (1986/1987) A single product cycling problem under Brownian motion demand. Manag Sci 32(10):1223–1370 Vidale ML, Wolfe HB (1957) An operations research study of sales response to advertising. Oper Res 5(3):370–381 Vidyasagar M (2002) Nonlinear systems analysis, 2nd edn. SIAM, Philadelphia Villas-Boas JM (1993) Predicting advertising policies in an oligopoly: a model and empirical test. Market Sci 12:88–102

Bibliography

543

Vinokurov VR (1969) Optimal control of processes described by integral equations I. SIAM J Control Optim 7:324–336 Vinter RB (1983) New global optimality conditions in optimal control theory. SIAM J Control Optim 21:235–245 von Weizs¨acker CC (1980) Barriers to entry: a theoretical treatment. Lecture notes in economics and mathematical systems, vol 185. Springer, Berlin Vousden N (1974) International trade and exhaustible resources: a theoretical model. Int Econ Rev 15:149–167 Wagener FOO (2003) Skiba points and heteroclinic bifurcations, with applications to the shallow lake system. J Econ Dyn Control 27(9):1533–1561 Wagener FOO (2005) Structural analysis of optimal investment for ﬁrms with non-concave production. J Econ Behav Org 57(4):474–489 Wagener FOO (2006) Skiba points for small discount rates. J Optim Theory Appl 128(2):261–277 Wagener FOO (2009) On the Leitmann equivalent problem approach. J Optim Theory Appl 142(1):229–242 Wagner HM, Whitin TM (1958) Dynamic version of the economic lot size model. Manag Sci 5:89–96 Wang PKC (1964) Control of distributed parameter systems. In: Leondes CT (ed) Advances in control systems, vol 1. Academic Press, New York, pp 75– 170 Wang WK, Ewald CO (2010) A stochastic diﬀerential ﬁshery game for a two species ﬁsh population with ecological interaction. J Econ Dyn Control 34(5):844–857 Warga J (1962) Relaxed variational problems. J Math Anal Appl 4:111–128 Warga J (1972) Optimal control of diﬀerential and functional equations. Academic Press, New York Warschat J (1985a) Optimal control of a production-inventory system with state constraints and a quadratic cost criterion. RAIRO-OR 19:275–292 Warschat J (1985b) Optimal production planning for cascaded production inventory systems. In: Bullinger HJ, Warnecke HJ (eds) Toward the factory of the future. Springer, Berlin, pp 669–674 Warschat J, Wunderlich HJ (1984) Time-optimal control policies for cascaded production-inventory systems with control and state constraints. Int J Syst Sci 15:513–524

544

Bibliography

Weber TA (2011) Optimal control theory with applications in economics. The MIT Press, Cambridge Weinstein MC, Zeckhauser RJ (1975) The optimum consumption of depletable natural resources. Quart J Econ 89:371–392 Welam UP (1982) Optimal and near optimal price and advertising strategies for ﬁnite and inﬁnite horizons. Manag Sci 28(11):1313–1327 Westphal LC (1974) Toward the synthesis of solutions of dynamic games. In: Leondes CT (ed) Advances in control systems, vol 11. Academic Press, New York, pp 389–489 Whittle P (1982/1983) Optimization over time: dynamic programming and stochastic control, vol I–II. Wiley, Chichester Wickwire K (1977) Mathematical models for the control of pests and infectious diseases: a survey. Theor Popul Biol 11:182–238 Wiener N (1949) The interpolation and smoothing of stationary time series. MIT Press, Cambridge Wirl F (1984) Sensitivity analysis of OPEC pricing policies. OPEC Rev 8:321– 331 Wirl F (1985) Stable and volatile prices: and explanation by dynamic demand. In: Feichtinger G (ed) Optimal control theory and economic analysis, vol 2. North-Holland, Amsterdam, pp 263–277 Wirl F, Feichtinger G (2005) History dependence in concave economies. J Econ Behav Organ 57:390–407 Wonham WM (1970) Random diﬀerential equations in control theory. In: Bharucha-Reid AT (ed) Probabilistic methods in applied mathematics, vol 2. Academic Press, New York Wright C (1974) Some political aspects of pollution control. J Environ Econ Manag 1:173–186 Wright SJ (1993) Interior-point methods for optimal control of discrete-time systems. J Optim Theory Appl 77(1):161–187 Wrzaczek S, Kaplan EH, Caulkins JP, Seidl A, Feichtinger G (2017) Diﬀerential terror queue games. In: Dynamic games and applications, vol 7. Springer, Berlin, pp 4, 578–593 Wrzaczek S, Kuhn M, F¨ urnkranz-Prskawetz A, Feichtinger G (2010) The reproductive value in distributed optimal control models. Theor Popul Biol 77(3):164–170

Bibliography

545

Xepapadeas A, de Zeeuw A (1999) Environmental policy and competitiveness: the Porter hypothesis and the composition of capital. J Environ Econ Manag 37(2):165–182 Yang J, Yan H, Sethi SP (1999) Optimal production planning in pull ﬂow lines with multiple products. Eur J Oper Res 119(3):26–48 Yin G, Zhang Q (1997) Continuous-time Markov chains and applications: a singular perturbation approach. Springer, New York Yong J, Zhou XY (1999) Stochastic Controls: Hamiltonian Systems and HJB Equations. Springer, New York Young LC (1969) Calculus of variations and optimal control theory. W.B. Saunders, Philadelphia Zaccour G (2008a) On the coordination of dynamic marketing channels and two-part tariﬀs. Automatica 44:1233–1239 Zaccour G (2008b) Time consistency in cooperative diﬀerential games: a tutorial. INFOR: Inf Syst Oper Res 46(1):81–92 Zeidan V (1984a) Extended Jacobi suﬃciency criterion for optimal control. SIAM J Control Optim 22:294–301 Zeidan V (1984b) A modiﬁed Hamilton-Jacobi approach in the generalized problem of Bolza. Appl Math Optim 11:97–109 Zeiler I, Caulkins JP, Grass D, Tragler G (2010) Keeping options open: an optimal control model with trajectories that reach a DNSS point in positive time. SIAM J Control Optim 48(6):3698–3707 Zelikin MI, Borisov VF (1994) Theory of chattering control: with applications to astronautics, robotics, economics, and engineering (Systems & control: foundations & applications). Birkh¨auser, Boston Zhang R, Ren Q (2013) Equivalence between Sethi advertising model and a scalar LQ diﬀerential game. In: 25th Chinese control and decision conference (CCDC), 25–27 May 2013 Zhang R, Haruvy E, Prasad A, Sethi SP (2005) Optimal ﬁrm contributions to open source software. In: Deissenberg C, Hartl RF (eds) Optimal control and dynamic games. Springer, Dordrecht, pp 197–212 Zhou X, Sethi SP (1994) A suﬃcient condition for near optimal stochastic controls and its application to manufacturing systems. Appl Math Optim 29:67–92

546

Bibliography

Ziemba WT, Vickson RG (eds) (1975) Stochastic optimization models in ﬁnance. Academic Press, New York Zimin IN, Ivanilov YP (1971) Solution of network planning problems by reducing them to optimal control problems (in Russian). Zhurnal Vychiﬂitel’noi Matematiki i Matematichefkoi Fiziki 11:3 Zi´olko M, Kozlowski J (1983) Evolution of body size: an optimal control model. Math Biosci 64:127–143 Zoltners AA (ed) (1982) Marketing planning models. TIMS studies in the management sciences, vol 18. North-Holland, Amsterdam Zwillinger D (2003) Standard mathematical tables and formulae. Chapman & Hall/CRC, Boca Raton

Index A Abad, P.L., 473, 538 Adjoint equation, 36–38, 42, 271 Adjoint variables, 10, 38 Adjoint vector, 35 Admissible control, 28 Advertising model, 5, 6 Aﬃne function, 22 Aggoun, L., 383, 491 Agnew, C.E., 473 Ahmed, N.U., 462, 473 Alam, M., 291, 309, 473, 529 Allen, K.R., 330, 473 Amit, R., 324, 474 Amoroso-Robinson relation, 227 Anderson, B.D.O., 442, 450, 474 Anderson, R.M., 474 Anti-diﬀerence operator, 414 Aoki, M., 448, 474 Applications to biomedicine, 343 Applications to ﬁnance, 159 Applications to marketing, 225 Arnold, L., 367, 370, 442, 451, 474 Aronson, J.E., 474, 534 Arora, S.R., 285, 474 Arrow, K.J., 11, 54, 70, 106, 226, 228, 335, 336, 339, 383, 474, 520, 522 Arthur, W.B., 360, 475 Arugaslan, O., x

Arutyunov, A.V., 143, 475 Aseev, S.M., 143, 475 Aubin, J.-P., 475, 529 Autonomous, 61 Axs¨ ater, S., 475 B Backlogging of demand, 5, 371 Bagchi, A., 475 Balachandran, B., 507 Balakrishman, A.V., 518 Bang-bang, 51, 99, 110, 111, 113, 162, 167, 168, 231, 273, 286, 289, 301, 390, 456 Bang function, 19 Bankruptcy, 380 Barnea, A., 515 Barnett, W.A., 489 Basar, T., 385, 396, 405, 475, 477, 479, 483, 489, 508, 509 Bass, F.M., 475, 532 Bayes theorem, 443 Bean, J.C., 475 Behrens, D.A., 11, 70, 109, 122, 360, 458, 460, 475, 500 Bell, D.J., 49, 455, 476 Bellman, R.E., 10, 33, 476 Benchekroun, H., 476 Benkherouf, L., 476

© Springer Nature Switzerland AG 2019 S. P. Sethi, Optimal Control Theory, https://doi.org/10.1007/978-3-319-98237-3

547

548 Bennett, R.J., 538 Bensoussan, A., 11, 70, 113, 213, 291, 293, 336, 339, 383, 385, 396, 397, 399, 404, 474–479, 506, 524, 533 Bequest function, 7, 377 Berkovitz, L.D., 10, 32, 385, 479 Bernoulli, Jacob, 9 Bernoulli, Jakob, 10 Bernoulli, Johann, 9, 10 Bertsekas, D.P., 277, 366, 479, 491 Bes, C., 213, 479 Bettiol, P., 41, 479 Beyer, D., x, 369, 479 Bhar, R., 448, 479 Bharucha-Reid, A.T., 544 Bhaskaran, S., 479, 480 Bionomic equilibrium, 313, 391 Black, F., 378, 480 Blaqui`ere, A., 480, 528, 537 Bliss,G.A., 10 Bliss point, 362 Blondel, V.D., 534 Boccia, A., 132, 480 Boiteux, M., 283, 480 Boiteux problem, 283 Bolton, P., 352, 480 Boltyanskii, V.G., 10, 27, 32, 70, 96, 141, 433, 480, 526 Bolza, 10 Bolza form, 30, 270, 277, 280 Bookbinder, J.H., xiv, 11, 360, 480 Borisov, V.F., 545 Boucekkine, R., 474 Boundary conditions, 86, 372 Boundary interval, 130 Bourguignon, F., 480

Index Bowes, M.D., 317, 480 Brachistochrone problem, 9, 424 Breakwell, J.V., 231, 480 Brekke, K.A., 383, 480 Breton, M., 404, 480 Brito, D.L., 360, 480 Brock, W.A., 383, 518 Brockhoﬀ, K., 476 Brokate, M., 462, 481 Broken extremal, 427 Brotherton, T., 482 Brown, R.G., 198, 481 Brownian Motion, 378 Bryant, G.F., 32, 481 Bryson, A.E. Jr., 113, 135, 141, 442, 450, 453, 481 Bryson, Jr., A.E., 39 Buchanan, L.F., 448, 481 Bucy, R., 448, 510 Budget constraint, 88 Bulirsch, R., 10, 135, 481, 524, 538 Bullinger, H.J., 543 Bultez, A.V., 404, 475, 477, 481 Bunching and ironing, 357, 359 Burdet, C.A., 277, 281, 481 Burgess, R., 498 Burmeister, E., 336, 481 B¨ uskens, C., 519 Butkowskiy, A.G., 462, 481 Bylka, S., 297, 481 C Caines, P., 482 C ¸ akanyildirim, M., 383, 477–479 Calculus of variations, 9, 419 Canon, M.D., 276, 482 Canonical system, 457 Capital accumulation model, 336 Caputo, M.R., 11, 482 Carath´eodory, C., 10

Index Carlson, D.A., 11, 106, 482, 483 Carraro, C., 483 Carrillo, J., 484 Case, J.H., 385, 484 Cass, D., 484 Cassandras, C.G., 277, 524 Caulkins, J.P., 11, 70, 109, 122, 360, 458, 460, 475, 484, 485, 499, 500, 512, 530, 541, 544, 545 Cellina, A., 475 Cernea, A., 41, 485 Certainty equivalence, 451, 452 Cesari, L., 32, 485 Chahim, M., 485 Chain of forests model, 321, 323 Chain of machines, 297 Chand, S., 221, 297, 485, 532 Chandra, T., 507 Chang, S., 383, 474 Chao, R., 485 Chappell, D., 485 Charnes, A., 360, 485 Chattering, 545 Chattering controls, 234, 253 Chatterjee, R., 527 Chen, S., 396, 397, 478, 479 Chen, S.F., 486 Cheng, F., x Chiarella, C., 486 Chichilinsky, G., 486 Chikan, A., 501 Chintagunta, P.K., 404, 486 Chow, G.C., 448, 486 Chutani, A., 478, 479, 486 Clark, C.W., 11, 283, 311, 312, 317, 330, 389, 392, 486, 487 Clarke, F.H., 32, 207, 487

549 Clemhout, S., 487 Coddington, E.A., 487 Cohen, K.J., 42, 227, 487 Common-property ﬁshery resources, 389 Comparison lemma, 239 Complementary slackness conditions, 73, 74, 76, 81, 82 Computational methods, ix, 135, 277 Concave function, 20, 79 Connors, M.M., 11, 487 Conrad, K., 487 Constantinides, G.M., 487 Constraint of rth order, 130 Constraint qualiﬁcations, 73, 130, 131, 267 Constraints, 28 Consumption-investment problem, 377 Consumption model, 7, 8 Contact time, 130 Continuous wheat trading model, 204 Control of pest infestations, 343 Control trajectory, 2, 28 Control variable, 2, 28 Control vector, 28 Convex combination, 20 Convex function, 20, 79 Convex hull, 20 Convex set, 20 CorelDRAW, ix Cottle, R.W., 491 Cowling, K., 507 Crisan, D., 478 Critical points, 260 Crouhy, M., 213, 477 Cruz, J.B., Jr., 536

550 Cuaresma, J.C., 525 Cullum, C.D., 276, 482 Current-value adjoint variables, 86 Current-value formulation, 70, 80, 111, 115, 280 Current-value functions, 80, 82, 84 Current-value Hamiltonian, 81, 83, 104, 111 Current-value Lagrange multipliers, 81 Current-value Lagrangian, 81 Current-value maximum principle, 83, 147 Cvitanic, J., 352, 487 Cycloid, 10 Cyert, R.M., 42, 227, 487 D Dantzig, G.B., 474, 487 Da Prato, G., 477 Darrat, A.F., 520 Darrough, M.N., 487 Dasgupta, P., 324, 487, 538 D’Autume, A., 488 Davis, B.E., 159, 164, 188, 383, 455, 488 Davis, M.H.A., 366, 488 Davis, R.E., 519 Dawid, H., 488 Day, G., 527 DDT, 346, 348, 351 Deal, K.R., 388, 404, 405, 488 Dechert, D.W., 458, 488 Decision horizon, 213, 215, 216, 219 Deger, S., 488 Deissenberg, C., 488–490, 497, 502, 504, 519, 535, 545 Deistler, M., 540 Delfour, M.C., 477

Index De Pinho, M.D.R., 132, 480 Derived Hamiltonian, 53 Derzko, N.A., 54, 253, 324, 383, 460, 462, 463, 477, 489, 530 DeSarbo, W., 361, 507 Descarte’s Rule of Signs, 402 Dewatripont, M., 352, 480 de Zeeuw, A.J., 351, 492, 545 Dhrymes, P.J., 251, 489 Diﬀerence equation, 270, 442 Diﬀerence operator, 414 Diﬀerential games, 385 Diﬀerentiation with scalars, 12 Diﬀerentiation with vectors, 13, 14 Direct contribution, 41 Discount factor, 6 Discount rate, 6 Discrete maximum principle, 259, 269, 270 Discrete-time optimal control problem, 269, 270 Distributed parameter systems, 460 Dixit, A.K., 489 Dmitruk, A.V., 489 Dobell, A.R., 336, 481 Dockner, E.J., 11, 335, 361, 385, 396, 404, 405, 489, 490, 493, 509 Dogramaci, A., 291, 490 Dohrmann, C.R., 277, 490 Dolan, R.J., 490, 508 Dorfman, R., 491 Dornoﬀ, R.J., 538 Drews, W., 491 Dreyfus, S.E., 479 Dual variables, 42 Dubovitskii, A.J., 132, 491

Index Dunn, J.C., 277, 491 Durrett, R., 367, 491 Dury, K., 485 Dynamic eﬃciency condition, 337 Dynamic programming, 32, 366, 433 E Economic applications, 335, 383 Economic interpretation, 40, 84, 175, 337 Educational policy, 25 Eigenvalues, 412, 413 Eigenvectors, 412, 413 El-Hodiri, M., 491 Eliashberg, J., 361, 491, 507, 539 Elliott, R.J., 383, 491 El Ouardighi, F., 491, 492 Elton, E., 159, 492 Elzinga, D.J., 159, 188, 455, 488 Ending correction, 196 Entry time, 130 Envelope theorem, 54, 253, 356 EOQ, 191 Epidemic control, 343 Equilibrium relation, 42 Erickson, G.M., 492 Erickson, L.E., 11, 404, 506 Euler, 9 Euler equation, 421, 422, 427, 428 Euler-Lagrange equation, 422 Ewald, C.-O., 543 Excel, ix, 57–60 Exhaustible resource model, 111, 324 Exit time, 130 F Factorial power, 414 Fan, L.T., 360, 492, 506

551 Farley, J.U., 539 Fattorini, H.O., 492 Feedback Nash equilibrium, 388, 393 Feedback Nash stochastic diﬀerential game, 392 Feedback Stackelberg equilibrium, 403 Feedback Stackelberg stochastic diﬀerential game, 395 Feenstra, T.L., 492 Feichtinger, G., x, xiv, 11, 39, 70, 79, 85, 104, 109, 122, 132, 136, 140, 207, 225, 335, 351, 360, 361, 458, 460, 463, 464, 475, 484, 485, 487–496, 499–504, 509, 512, 513, 515–517, 519–522, 527–530, 533, 537, 539–542, 544 Feinberg, F.M., 235, 495, 496 Fel’dbaum, A.A., 36, 433, 438, 496 Feng, Q., 291, 336, 339, 383, 474, 478 Ferreira, M.M.A., 143, 496 Ferreyra, G., 496 Filar, J., 483 Filipiak, J., 496 Finite diﬀerence equations, 414 First-order linear equations, 409 First-order pure state constraints, 135 Fischer, T., 496 Fisher, A.C., 525 Fishery management, 392 Fishery model, 312, 389 Fishing mortality function, 392

552 Fixed-end-point problem, 77, 86, 98, 99, 113 Fleming, W.H., 366, 367, 383, 442, 496 Fletcher, R., 496 Fomin, S.V., 419, 420, 422, 427, 429, 430, 499 Fond, S., 496 Forecast horizons, 213 Forest fertilization model, 331 Forestry model, 111 Forest thinning model, 317, 321 Forgetting coeﬃcient, 5 Forster, B.A., 496 Fourgeaud, C., 496 Fraiman, N.M., 291, 490 Francis, P.J., 343, 496 Frankena, J.F., 496 Frankowska, H., 41, 485 Free-end-point problem, 86 Free terminal time problems, 93 Friedman, A., 385, 496, 497 Fromovitz, S., 269, 518 Fruchter, G.E., 404, 492, 497 Fuller, D., 325, 497 Full-rank condition, 23, 72, 130, 131 Fundamental lemma, 422 Funke, U.H., 497 F¨ urnkranz-Prskawetz, A., 463, 513, 544 Fursikov, A.V., 497 Furst, E., 540 G Gaandolfo, G., 498 Gaimon, C., 291, 360, 361, 484, 485, 497, 498

Index Gamkrelidze, R.V., 10, 27, 32, 70, 96, 141, 234, 433, 498, 526 Gandolfo, G., 493 Gaskins, D.W. Jr., 498 Gaugusch, J., 498 Gaussian, 442, 443 Gavrila, C., 499 Geismar, N., x Gelfand, I.M., 419, 420, 422, 427, 429, 430, 499 General discrete maximum principle, 276 Generalized bang-bang, 113, 167 Generalized Legendre-Clebsch condition, 189, 454–456 Geoﬀrion, A.M., 516 Gerchak, Y., 499 Gfrerer, H., 499 Gibson, J.E., 455, 508 Gihman, I.I., 375, 499 Gillessen, W., 519 Girsanov, I.V., 499 Glad, S.T., 499 Global Saddle Point Theorem, 343, 350, 456, 457 Goal Seek, 57, 59 Goh, B.-S., 135, 311, 499 Goh, C.J., 540 Goldberg, S., 414, 499 Golden Path, 108 Golden Rule, 108, 123 Goldstein, J.R., 488 Goldstine, H.H., 499 G¨ollmann, L., 499 Goodwill, 5, 226 Goodwill elasticity of demand, 229 Gopalsamy, K., 499 Gordon, H.S., 312, 313, 499 Gordon, M.J., 181, 499

Index Gordon’s formula, 181 Gould, J.P., 231, 257, 362, 499 Grass, D., 11, 70, 109, 122, 360, 458, 460, 484, 485, 491, 500, 545 Green’s theorem, 225, 237, 239, 245, 254, 257, 314, 344, 346, 363, 391 Grienauer, W., 495 Grimm, W., 500, 523 Gross, M., 500 Gruber, M., 159, 492 Gruver, W.A., 512 Gutierrez, G.J., 404, 505 H H¨am¨ al¨ainen, R.P., 392, 500 Hadley, G., 11, 70, 335, 500 Hahn, M., 500 Halkin, H., 32, 276, 500 H¨ am¨ al¨ainen, R.P., 392 Hamilton, 10 Hamiltonian, 35, 41, 73, 271, 337, 437 Hamiltonian maximizing condition, 36, 39, 74, 97, 99, 118, 437 Hamilton-Jacobi-Bellman (HJB) equation, 32, 36, 366, 368, 371, 393 Hamilton-Jacobi equation, 371, 394 Han, M., 500 Hanson, M., 521 Hanssens, D.M., 500 Harris, F.W., 191, 500 Harris, H., 360, 500 Harrison, J.M., 379, 501 Hartberger, R.J., 32, 436, 491, 501

553 Hartl, R.F., x, 11, 32, 39, 70, 79, 85, 95, 104, 131, 132, 135–137, 140, 141, 149, 207, 213, 225, 281, 285, 351, 360, 361, 458, 460, 463, 464, 484, 485, 490, 491, 493, 495, 497, 499–505, 512, 516, 519, 535, 537, 545 Hartman, R., 458, 504 Haruvy, E., 235, 253, 504, 545 Harvey, A.C., 504 Haunschmied, J.L., 360, 484, 495, 504 Haurie, A., 11, 106, 213, 385, 392, 463, 479, 483, 500, 504, 505 Haussmann, U.G., 505 He, X., 396, 404, 505, 535 Heal, G.M., 324, 487, 505 Heaps, T., 283, 505 Heckman, J., 505 Heineke, J.M., 487 Hestenes, M.R., 10, 70, 505 HJB equation, 36, 376 HMMS model, 191 Ho, Y.-C., 39, 113, 141, 385, 388, 442, 450, 453, 481, 505, 523, 537 Hochman, E., 383, 527 Hofbauer, J., 505, 506 Hoﬀmann, K.H., 506 Hohn, F., 213, 520 Holly, S., 506 Holt, C.C., 191, 200, 202, 506 Holtzman, J.M., 276, 506

554 Homogeneous function of degree k, 22 Horsky, D., 506 Hotelling, H., 324, 506 Hritonenko, N., 474 Hsu, V.N., 221, 485 Hung, N.M., 504 Hurst, Jr., E.G., 11, 70, 113, 477 Hwang, C.L., 506 Hyun, J.S., 500

Index

J Jabrane, A., 483 Jacobi, 10 Jacobson, D.H., 49, 455, 476, 507, 514 Jacquemin, A.P., 228, 229, 231, 507 Jagpal, S., 507 Jain, D., 404, 486 Jamshidi, M., 507 Jardine, A., 535 Jarrar, R., 404, 480, 507 I Jazwinski, A.H., 507 Ijiri, Y., 191, 205, 506 Jedidi, K., 361, 507 Ilan, Y., 474 Jennings, L.S., 507 Illustration of left and right limits, Jeuland, A.P., 490, 508 18 Ji, Y., 508 Imp, 19 Jiang, J., 508 Impulse control, 19, 113, 242, 243 Johar, M., 508 Impulse control model, 113 Johnson, C.D., 455, 508 Impulse stochastic control, 383 Jones, P., 508 Imputed value, 261 Jørgensen, S., x, 11, 311, 335, 361, Incentive compatibility, 353, 354 385, 396, Indirect adjoining method, 404, 405, 489–491, 137, 147 493, 495, 502, 508, 509 Indirect contribution, 41 Joseph, P.D., 452 Individual rationality, 353 Jump conditions, 135 Inﬁnite horizon, 6, 103 Jump Markov processes, 383 Instantaneous proﬁt rate, 29 Junction times, 130 Interior interval, 130 Intriligator, M.D., 336, 506, 507, K 520 Kaitala, V.T., 392, 494, 500, 509 Inventory control problem, 152 Kalaba, R.E., 476 Investment allocation, 65 Kalish, S., 360, 497, 509, 510 Ioﬀe, A.D., 507 Kall, P., 501, 522, 533 Isaacs, R., 10, 170, 385, 507 Kalman, R.E., 442, 448, 510 Isoperimetric constraint, 88, 252 Kalman-Bucy ﬁlter, 442, 447, 448 Itˆ o stochastic diﬀerential equation, Kalman ﬁlter, 366, 441–443, 447 366, 393 Kamien, M.I., 11, 290, Ivanilov, Y.P., 360, 546 293, 296, 335, 360, 510

Index Kamien-Schwartz model, 111, 291 Kaplan, E.H., 360, 530, 544 Kaplan, W., 510 Karatzas, I., 337, 367, 381–383, 510 Karray, S., 511 Karreman, H.F., 480 Kavadias, S., 485 Keeler, E., 511 Keller, H.B., 346, 511 Kemp, M.C., 11, 70, 335, 486, 500, 511 Kendrick, D.A., 511 Keon, J.W., 540 Keppo, J., 383, 478 Kern, D., 499 Khmelnitsky, E., 11, 511, 512, 517 Kilkki, P., 317, 318, 511 Kim, J-H.R., 455, 519 Kirakossian, G.T., 503 Kirby, B.J., 511 Kirk, D.E., 36, 511 Kiseleva, T., 458, 511 Klein, C.F., 512 Kleindorfer, G.B., 274, 512 Kleindorfer, P.R., 213, 274, 512, 533 Kleinschmidt, P., 502 Kneese, A.V., 480, 505, 521 Knobloch, H.W., 512 Knowles, G., 512 Kogan, K., 11, 478, 511, 512, 517 Kopel, M., 488 Kort, P.M., 11, 311, 335, 351, 360, 361, 404, 458, 460, 463, 484, 485, 491, 492, 495, 502–504, 509, 512, 542 Kortanek, K., 360, 485

555 Kotowitz, Y., 512 Kozlowski, J., 546 Krabs, W., 506 Kraft, D., 135, 481 Krarup, J., 491 Krauth, J., 361, 502 Kreindler, E., 512 Krelle, W., 476, 512 Krichagina, E., 512 Kriebel, C.H., 274, 512 Krishnamoorthy, A., 475, 505, 513 Krouse, C.G., 164, 513 Krutilla, J.V., 317, 480, 510 Kugelmann, B., 513 Kuhn, H.W., 535 Kuhn, M., 463, 513, 544 Kuhn-Tucker conditions, 260, 262, 269, 271 Kumar, P.R., 513 Kumar, S., x, 442, 508, 513 Kurawarwala, A.A., 513 Kurcyusz, S., 144, 513 Kurz, M., 11, 54, 70, 106, 228, 335, 336, 474 Kushner, H.J., 513 Kydland, F.E., 513 L Laﬀont, J.J., 352, 513 Lagrange, 9 Lagrange form, 30 Lagrange multipliers, 69, 70, 73, 74, 82, 260 Lagrangian, 73 Lagrangian form, 70, 76, 114 Lagrangian maximum principle, 70 Lagunov, V.N., 513 Lakhani, C., 528 Lansdowne, Z.F., 246, 514 Lasdon, L.S., 514

556 Leban, R., 11, 335, 514 Leclair, S.R., 507 Lee, E.B., 164, 514 Lee, S.C., 252, 532 Lee, W.Y., 513 Left and right limits, 17 Legendre, 9 Legendre’s conditions, 428 Legey, L., 360, 514 Lehoczky, J.P., 337, 381, 382, 510, 514, 532 Leibniz, 9 Leitmann, G., 32, 311, 385, 419, 486, 499, 500, 504, 505, 514, 520, 537 Leizarowitz, A., 483 Leland, H.E., 514 Lele, M.M., 507, 514 Lele, P.T., 285, 474 Lenclud, B., 496 L´eonard, D., 11, 335, 514 Leondes, C.T., 481, 519, 543, 544 Lesourne, J., 11, 335, 383, 476, 514 Lev, B., 497 Levine, J., 515 Levinson, N.L., 487 Lewis, T.R., 515 L’Hˆopital’s rule, 342 Li, G., 515 Li, M., 479 Li, T.