# Diﬀerential Equations

Linear, Nonlinear, Ordinary, Partial

A.C. King, J. Billingham and S.R. Otto

   Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, São Paulo Cambridge University Press The Edinburgh Building, Cambridge  , United Kingdom Published in the United States of America by Cambridge University Press, New York www.cambridge.org Information on this title: www.cambridge.org/9780521816588 © Cambridge University Press 2003 This book is in copyright. Subject to statutory exception and to the provision of relevant collective licensing agreements, no reproduction of any part may take place without the written permission of Cambridge University Press. First published in print format 2003 - - - - - - ---- eBook (NetLibrary) --- eBook (NetLibrary) ---- hardback --- hardback ---- paperback --- paperback

Cambridge University Press has no responsibility for the persistence or accuracy of s for external or third-party internet websites referred to in this book, and does not guarantee that any content on such websites is, or will remain, accurate or appropriate.

Contents

Preface Part One: Linear Equations 1 Variable Coeﬃcient, Second Order, Linear, Ordinary Diﬀerential Equations 1.1 The Method of Reduction of Order 1.2 The Method of Variation of Parameters 1.3 Solution by Power Series: The Method of Frobenius Legendre Functions 2.1 Deﬁnition of the Legendre Polynomials, Pn (x) 2.2 The Generating Function for Pn (x) 2.3 Diﬀerential and Recurrence Relations Between Legendre Polynomials 2.4 Rodrigues’ Formula 2.5 Orthogonality of the Legendre Polynomials 2.6 Physical Applications of the Legendre Polynomials 2.7 The Associated Legendre Equation

page ix 1 3 5 7 11 31 31 35 38 39 41 44 52 58 58 60 64 69 71 71 77 79 80 93 96 100 107

2

3

Bessel Functions 3.1 The Gamma Function and the Pockhammer Symbol 3.2 Series Solutions of Bessel’s Equation 3.3 The Generating Function for Jn (x), n an integer 3.4 Diﬀerential and Recurrence Relations Between Bessel Functions 3.5 Modiﬁed Bessel Functions 3.6 Orthogonality of the Bessel Functions 3.7 Inhomogeneous Terms in Bessel’s Equation 3.8 Solutions Expressible as Bessel Functions 3.9 Physical Applications of the Bessel Functions Boundary Value Problems, Green’s Functions and Sturm–Liouville Theory 4.1 Inhomogeneous Linear Boundary Value Problems 4.2 The Solution of Boundary Value Problems by Eigenfunction Expansions 4.3 Sturm–Liouville Systems

4

vi

CONTENTS

5

Fourier Series and the Fourier Transform 5.1 General Fourier Series 5.2 The Fourier Transform 5.3 Green’s Functions Revisited 5.4 Solution of Laplace’s Equation Using Fourier Transforms 5.5 Generalization to Higher Dimensions Laplace Transforms 6.1 Deﬁnition and Examples 6.2 Properties of the Laplace Transform 6.3 The Solution of Ordinary Diﬀerential Equations using Laplace Transforms 6.4 The Inversion Formula for Laplace Transforms

123 127 133 141 143 145 152 152 154 157 162

6

7

Classiﬁcation, Properties and Complex Variable Methods for Second Order Partial Diﬀerential Equations 175 7.1 Classiﬁcation and Properties of Linear, Second Order Partial Diﬀerential Equations in Two Independent Variables 175 7.2 Complex Variable Methods for Solving Laplace’s Equation 186

Part Two: Nonlinear Equations and Advanced Techniques 8 Existence, Uniqueness, Continuity and Comparison of Solutions of Ordinary Diﬀerential Equations 8.1 Local Existence of Solutions 8.2 Uniqueness of Solutions 8.3 Dependence of the Solution on the Initial Conditions 8.4 Comparison Theorems Nonlinear Ordinary Diﬀerential Equations: Phase Plane Methods 9.1 Introduction: The Simple Pendulum 9.2 First Order Autonomous Nonlinear Ordinary Diﬀerential Equations 9.3 Second Order Autonomous Nonlinear Ordinary Diﬀerential Equations 9.4 Third Order Autonomous Nonlinear Ordinary Diﬀerential Equations Group Theoretical Methods 10.1 Lie Groups 10.2 Invariants Under Group Action 10.3 The Extended Group 10.4 Integration of a First Order Equation with a Known Group Invariant

201 203 204 210 211 212 217 217 222 224 249 256 257 261 262 263

9

10

CONTENTS

vii

10.5 10.6 10.7 11

Towards the Systematic Determination of Groups Under Which a First Order Equation is Invariant Invariants for Second Order Diﬀerential Equations Partial Diﬀerential Equations

265 266 270 274 275 280 303 303 306 351 372 372 381 388 417 418 418 422 426 429 433 436 447 447 452 467 472 484 495 502 505 509 511 517 526 534 536

Asymptotic Methods: Basic Ideas 11.1 Asymptotic Expansions 11.2 The Asymptotic Evaluation of Integrals Asymptotic Methods: Diﬀerential Equations 12.1 An Instructive Analogy: Algebraic Equations 12.2 Ordinary Diﬀerential Equations 12.3 Partial Diﬀerential Equations Stability, Instability and Bifurcations 13.1 Zero Eigenvalues and the Centre Manifold Theorem 13.2 Lyapunov’s Theorems 13.3 Bifurcation Theory Time-Optimal Control in the Phase Plane 14.1 Deﬁnitions 14.2 First Order Equations 14.3 Second Order Equations 14.4 Examples of Second Order Control Problems 14.5 Properties of the Controllable Set 14.6 The Controllability Matrix 14.7 The Time-Optimal Maximum Principle (TOMP) An Introduction to Chaotic Systems 15.1 Three Simple Chaotic Systems 15.2 Mappings 15.3 The Poincar´ Return Map e 15.4 Homoclinic Tangles 15.5 Quantifying Chaos: Lyapunov Exponents and the Lyapunov Spectrum

12

13

14

15

Appendix 1 Linear Algebra Appendix 2 Continuity and Diﬀerentiability Appendix 3 Power Series Appendix 4 Sequences of Functions Appendix 5 Ordinary Diﬀerential Equations Appendix 6 Complex Variables Appendix 7 A Short Introduction to MATLAB Bibliography Index

Preface

When mathematical modelling is used to describe physical, biological or chemical phenomena, one of the most common results is either a diﬀerential equation or a system of diﬀerential equations, together with appropriate boundary and initial conditions. These diﬀerential equations may be ordinary or partial, and ﬁnding and interpreting their solution is at the heart of applied mathematics. A thorough introduction to diﬀerential equations is therefore a necessary part of the education of any applied mathematician, and this book is aimed at building up skills in this area. For similar reasons, the book should also be of use to mathematically-inclined physicists and engineers. Although the importance of studying diﬀerential equations is not generally in question, exactly how the theory of diﬀerential equations should be taught, and what aspects should be emphasized, is more controversial. In our experience, textbooks on diﬀerential equations usually fall into one of two categories. Firstly, there is the type of textbook that emphasizes the importance of abstract mathematical results, proving each of its theorems with full mathematical rigour. Such textbooks are usually aimed at graduate students, and are inappropriate for the average undergraduate. Secondly, there is the type of textbook that shows the student how to construct solutions of diﬀerential equations, with particular emphasis on algorithmic methods. These textbooks often tackle only linear equations, and have no pretension to mathematical rigour. However, they are usually well-stocked with interesting examples, and often include sections on numerical solution methods. In this textbook, we steer a course between these two extremes, starting at the level of preparedness of a typical, but well-motivated, second year undergraduate at a British university. As such, the book begins in an unsophisticated style with the clear objective of obtaining quantitative results for a particular linear ordinary diﬀerential equation. The text is, however, written in a progressive manner, with the aim of developing a deeper understanding of ordinary and partial diﬀerential equations, including conditions for the existence and uniqueness of solutions, solutions by group theoretical and asymptotic methods, the basic ideas of control theory, and nonlinear systems, including bifurcation theory and chaos. The emphasis of the book is on analytical and asymptotic solution methods. However, where appropriate, we have supplemented the text by including numerical solutions and graphs produced using MATLAB†, version 6. We assume some knowledge of
† MATLAB is a registered trademark of The MathWorks, Inc.

x

PREFACE

MATLAB (summarized in Appendix 7), but explain any nontrivial aspects as they arise. Where mathematical rigour is required, we have presented the appropriate analysis, on the basis that the student has taken ﬁrst courses in analysis and linear algebra. We have, however, avoided any functional analysis. Most of the material in the book has been taught by us in courses for undergraduates at the University of Birmingham. This has given us some insight into what students ﬁnd diﬃcult, and, as a consequence, what needs to be emphasized and re-iterated. The book is divided into two parts. In the ﬁrst of these, we tackle linear diﬀerential equations. The ﬁrst three chapters are concerned with variable coeﬃcient, linear, second order ordinary diﬀerential equations, emphasizing the methods of reduction of order and variation of parameters, and series solution by the method of Frobenius. In particular, we discuss Legendre functions (Chapter 2) and Bessel functions (Chapter 3) in detail, and motivate this by giving examples of how they arise in real modelling problems. These examples lead to partial diﬀerential equations, and we use separation of variables to obtain Legendre’s and Bessel’s equations. In Chapter 4, the emphasis is on boundary value problems, and we show how these diﬀer from initial value problems. We introduce Sturm–Liouville theory in this chapter, and prove various results on eigenvalue problems. The next two chapters of the ﬁrst part of the book are concerned with Fourier series, and Fourier and Laplace transforms. We discuss in detail the convergence of Fourier series, since the analysis involved is far more straightforward than that associated with other basis functions. Our approach to Fourier transforms involves a short introduction to the theory of generalized functions. The advantage of this approach is that a discussion of what types of function possess a Fourier transform is straightforward, since all generalized functions possess a Fourier transform. We show how Fourier transforms can be used to construct the free space Green’s function for both ordinary and partial diﬀerential equations. We also use Fourier transforms to derive the solutions of the Dirichlet and Neumann problems for Laplace’s equation. Our discussion of the Laplace transform includes an outline proof of the inversion theorem, and several examples of physical problems, for example involving diﬀusion, that can be solved by this method. In Chapter 7 we discuss the classiﬁcation of linear, second order partial diﬀerential equations, emphasizing the reasons why the canonical examples of elliptic, parabolic and hyperbolic equations, namely Laplace’s equation, the diﬀusion equation and the wave equation, have the properties that they do. We also consider complex variable methods for solving Laplace’s equation, emphasizing their application to problems in ﬂuid mechanics. The second part of the book is concerned with nonlinear problems and more advanced techniques. Although we have used a lot of the material in Chapters 9 and 14 (phase plane techniques and control theory) in a course for second year undergraduates, the bulk of the material here is aimed at third year students. We begin in Chapter 8 with a brief introduction to the rigorous analysis of ordinary diﬀerential equations. Here the emphasis is on existence, uniqueness and comparison theorems. In Chapter 9 we introduce the phase plane and its associated techniques. This is the ﬁrst of three chapters (the others being Chapters 13 and 15) that form an introduction to the theory of nonlinear ordinary diﬀerential equations,

PREFACE

xi

often known as dynamical systems. In Chapter 10, we show how the ideas of group theory can be used to ﬁnd exact solutions of ordinary and partial diﬀerential equations. In Chapters 11 and 12 we discuss the theory and practice of asymptotic analysis. After discussing the basic ideas at the beginning of Chapter 11, we move on to study the three most important techniques for the asymptotic evaluation of integrals: Laplace’s method, the method of stationary phase and the method of steepest descents. Chapter 12 is devoted to the asymptotic solution of diﬀerential equations, and we introduce the method of matched asymptotic expansions, and the associated idea of asymptotic matching, the method of multiple scales, including Kuzmak’s method for analysing the slow damping of nonlinear oscillators, and the WKB expansion. We illustrate each of these methods with a wide variety of examples, for both nonlinear ordinary diﬀerential equations and partial diﬀerential equations. In Chapter 13 we cover the centre manifold theorem, Lyapunov functions and an introduction to bifurcation theory. Chapter 14 is about time-optimal control theory in the phase plane, and includes a discussion of the controllability matrix and the time-optimal maximum principle for second order linear systems of ordinary diﬀerential equations. Chapter 15 is on chaotic systems, and, after some illustrative examples, emphasizes the theory of homoclinic tangles and Mel’nikov theory. There is a set of exercises at the end of each chapter. Harder exercises are marked with a star, and many chapters include a project, which is rather longer than the average exercise, and whose solution involves searches in the library or on the Internet, and deeper study. Bona ﬁde teachers and instructors can obtain full worked solutions to many of the exercises by emailing solutions@cambridge.org. In order to follow many of the ideas and calculations that we describe in this book, and to fully appreciate the more advanced material, the reader may need to acquire (or refresh) some basic skills. These are covered in the appendices, and fall into six basic areas: linear algebra, continuity and diﬀerentiability, power series, sequences and series of functions, ordinary diﬀerential equations and complex variables. We would like to thank our friends and colleagues, Adam Burbidge (Nestl´ Ree search Centre, Lausanne), Norrie Everitt (Birmingham), Chris Good (Birmingham), Ray Jones (Birmingham), John King (Nottingham), Dave Needham (Reading), Nigel Scott (East Anglia) and Warren Smith (Birmingham), who read and commented on one or more chapters of the book before it was published. Any nonsense remaining is, of course, our fault and not theirs. ACK, JB and SRO, Birmingham 2002

Part One Linear Equations

1

CHAPTER ONE

Variable Coeﬃcient, Second Order, Linear, Ordinary Diﬀerential Equations

Many physical, chemical and biological systems can be described using mathematical models. Once the model is formulated, we usually need to solve a diﬀerential equation in order to predict and quantify the features of the system being modelled. As a precursor to this, we consider linear, second order ordinary diﬀerential equations of the form P (x) dy d 2y + R(x)y = F (x), + Q(x) 2 dx dx

with P (x), Q(x) and R(x) ﬁnite polynomials that contain no common factor. This equation is inhomogeneous and has variable coeﬃcients. The form of these polynomials varies according to the underlying physical problem that we are studying. However, we will postpone any discussion of the physical origin of such equations until we have considered some classical mathematical models in Chapters 2 and 3. After dividing through by P (x), we obtain the more convenient, equivalent form, d 2y dy + a0 (x)y = f (x). + a1 (x) dx2 dx (1.1)

This process is mathematically legitimate, provided that P (x) = 0. If P (x0 ) = 0 at some point x = x0 , it is not legitimate, and we call x0 a singular point of the equation. If P (x0 ) = 0, x0 is a regular or ordinary point of the equation. If P (x) = 0 for all points x in the interval where we want to solve the equation, we say that the equation is nonsingular, or regular, on the interval. We usually need to solve (1.1) subject to either initial conditions of the form y(a) = α, y (a) = β or boundary conditions, of which y(a) = α and y(b) = β are typical examples. It is worth reminding ourselves that, given the ordinary differential equation and initial conditions (an initial value problem), the objective is to determine the solution for other values of x, typically, x > a, as illustrated in Figure 1.1. As an example, consider a projectile. The initial conditions are the position of the projectile and the speed and angle to the horizontal at which it is ﬁred. We then want to know the path of the projectile, given these initial conditions. For initial value problems of this form, it is possible to show that: (i) If a1 (x), a0 (x) and f (x) are continuous on some open interval I that contains the initial point a, a unique solution of the initial value problem exists on the interval I, as we shall demonstrate in Chapter 8.

The solutions of the homogeneous equation are elements of the null space of L. These results can be proved rigorously. Unfortunately. Ly = f . 1. there is a lot more work to do.1. but nonconstructively. An initial value problem. by studying the operator Ly ≡ dy d 2y + a0 (x)y. if we actually want to construct the solution of a particular equation. we will outline a series of subproblems that are more tractable. SECOND ORDER DIFFERENTIAL EQUATIONS y given y(a) = α and y' (a) = β α calculate the solution for x > a a x Fig. is then given formally as y = L−1 f . (ii) The structure of the solution of the initial value problem is of the form y= A u1 (x) + B u2 (x) Complementary function + G(x) . . This subspace is completely determined once its basis is known. Particular integral where A.4 VARIABLE COEFFICIENT. + a1 (x) dx2 dx and regarding L : C 2 (I) → C 0 (I) as a linear transformation from the space of twice-diﬀerentiable functions deﬁned on the interval I to the space of continuous functions deﬁned on I. B are constants that are ﬁxed by the initial conditions and u1 (x) and u2 (x) are linearly independent solutions of the corresponding homogeneous problem y + a1 (x)y + a0 (x)y = 0. The solution of the inhomogeneous problem. Before we try to construct the general solution of the inhomogeneous initial value problem.

Dividing through by Zu1 we have 2 du1 1 dZ + + a1 = 0. The solution is therefore ˜ . This can then be integrated to give U (x) = c exp − u2 (t) 1 t a1 (s) ds dt + c. the term multiplying U is zero.2) we obtain d 2 u1 d 2U dU du1 +U u +2 + a1 (x) 2 1 dx dx dx dx2 We can now collect terms to get U d 2 u1 du1 + a0 (x)u1 + a1 (x) dx2 dx + u1 d 2U dU + dx2 dx 2 du1 + a1 u1 dx = 0. say y(x) = u1 (x). Z dx u1 dx which can be integrated directly to yield x log |Z| + 2 log |u1 | + a1 (s) ds = C. dx dx dx d 2y d 2 u1 d 2U dU du1 +U = u1 + 2 . + a1 (x) dx2 dx (1. for some constant C. and. dx2 dx2 dx dx dx2 If we substitute these expressions into (1. where s is a dummy variable. We will look for a solution of the form y(x) = U (x)u1 (x). and only need to ﬁnd the second solution. since u1 (x) is a solution of (1.1. ˜ for some constant c. Thus Z= c exp − u2 1 x x a1 (s) ds = dU dx where c = eC . dU du1 u1 + U dx dx + a0 (x)U u1 = 0.1 The Method of Reduction of Order As a ﬁrst simpliﬁcation we discuss the solution of the homogeneous diﬀerential equation d 2y dy + a0 (x)y = 0.2). by deﬁning Z = dU/dx.2) on the assumption that we know one solution. Now. We have therefore obtained a diﬀerential equation for dU/dx. Diﬀerentiating y(x) using the product rule gives dy dU du1 = u1 + U .1 THE METHOD OF REDUCTION OF ORDER 5 1. have u1 du1 dZ +Z 2 + a1 u1 dx dx = 0.

= 2+ 2) 2 −t t 1−t t 2(1 + t) 2(1 − t) This gives the second solution of (1.2) as x u2 (x) = x 1 1 = x − + log t 2 and hence the general solution is 1 1 1 + + t2 2(1 + t) 2(1 − t) 1+t 1−t x dt = x log 2 1+x 1−x − 1. and x u2 (x) = u1 (x) 1 u2 (t) 1 t exp − a1 (s)ds dt (1.3) as the second part of the complementary function. we have a1 (x) = −2x/(1 − x2 ). − 2x dx2 dx given that y = u1 (x) = x is a solution. y = Ax + B x log 2 1+x 1−x −1 . ˜ We can recognize cu1 (x) as the part of the complementary function that we knew ˜ to start with. Example Let’s try to determine the full solution of the diﬀerential equation (1 − x2 ) d 2y dy + 2y = 0. 1 − s2 x u2 (x) = x 1 exp − log(1 − t2 ) dt = x t2 dt . . dx2 1 − x2 dx 1 − x2 Comparing this with (1. t2 (1 − t2 ) We can express the integrand in terms of its partial fractions as t2 (1 1 1 1 1 1 1 = 2+ + . We ﬁrstly write the equation in standard form as 2 d 2y 2x dy + − y = 0. After noting that t t a1 (s) ds = the reduction of order formula gives x − 2s ds = log(1 − t2 ).6 VARIABLE COEFFICIENT. This result is called the reduction of order formula. SECOND ORDER DIFFERENTIAL EQUATIONS x y(x) = u1 (x) c exp − u2 (t) 1 t a1 (s) ds dt + cu1 (x).2).

when diﬀerentiated again. comprising u1 (x) and u2 (x). dx dx dx dx Since u1 and u2 are solutions of the homogeneous equation. W c1 c2 = 0 f . = c1 + + 2 2 dx dx dx dx dx2 dx dx This form can then be substituted into the original diﬀerential equation to give c1 du1 dc1 du2 dc2 d 2 u1 d 2 u2 du1 du2 + c2 + a1 c1 + c2 + + dx2 dx dx dx2 dx dx dx dx d 2 u1 du1 + a0 u1 + a1 dx2 dx d 2 u2 du2 + a0 u2 + a1 dx2 dx + a0 (c1 u1 + c2 u2 ) = f. we ﬁnd that dy du1 dc1 du2 dc2 = c1 + u1 + c2 + u2 . W . which can be written in matrix form as u1 u1 u2 u2 f u2 .2 The Method of Variation of Parameters Let’s now consider how to ﬁnd the particular integral given the complementary function. for c1 = dc1 /dx and c2 = dc2 /dx.1. the ﬁrst two terms are zero. dx dx dx dx (1. dx dx dx which. As the name of this technique suggests.2 THE METHOD OF VARIATION OF PARAMETERS 7 1.4) This can be rearranged to show that c1 + c2 + du2 dc2 du1 dc1 + = f. and assume that y = c1 (x)u1 (x) + c2 (x)u2 (x). dx dx (1. Diﬀerentiating. (1.5) We now have two simultaneous equations. which gives us du2 dc2 du1 dc1 + = f. dx dx dx dx dx We will choose to impose the condition u1 and thus have dy du1 du2 = c1 + c2 . we take the constants in the complementary function to be variable.4) and (1. yields d 2y d 2 u1 d 2 u2 du1 dc1 du2 dc2 + c2 .5). These can easily be solved to give c1 = − where W = u1 u2 − u2 u1 = u1 u1 u2 u2 c2 = f u1 . dc1 dc2 + u2 = 0.

Using integration by parts. we can evaluate this. we have made no comment so far about why the procedures we used in the derivation should work at all! It turns out that this has a close connection with the theory of continuous groups. dx2 The homogeneous form of this equation has constant coeﬃcients. These expressions can be integrated to give x c1 = − f (s)u2 (s) ds + A. W (s) x c2 = f (s)u1 (s) ds + B. 4 4 Although we have given a rational derivation of the reduction of order and variation of parameters formulae. and ﬁnd that 1 1 1 y(x) = − x2 cos x + x sin x + cos x 4 4 8 is the required particular integral. The general solution is therefore 1 1 y = c1 cos x + c2 sin x − x2 cos x + x sin x. W (s) The particular integral is therefore y(x) = f (s) u1 (s)u2 (x) − u1 (x)u2 (s) W (s) ds.8 VARIABLE COEFFICIENT. W (s) We can now write down the solution of the entire problem as x y(x) = u1 (x) − f (s)u2 (s) ds + u2 (x) W (s) x x f (s)u1 (s) ds + Au1 (x) + Bu2 (x).6) This is called the variation of parameters formula. = cos2 x + sin2 x = 1. . which we will investigate in Chapter 10. (1. The variation of parameters formula then gives the particular integral as y= since W = cos x sin x − sin x cos x x s sin s cos s sin x − cos x sin s 1 ds. Example Consider the equation d 2y + y = x sin x. with solutions u1 (x) = cos x. x u2 (x) = sin x. We can split the particular integral into two integrals as x y(x) = sin x 1 sin x 2 x s sin s cos s ds − cos x s sin 2s ds − 1 cos x 2 x s sin2 s ds = s (1 − cos 2s) ds. SECOND ORDER DIFFERENTIAL EQUATIONS is called the Wronskian.

dx . To see this. and u2 (x) = 3x2 . let’s pause to discuss some further properties of the Wronskian. with α1 and α2 not both zero. Example The functions u1 (x) = f (x) and u2 (x) = kf (x). u1 (x) = 2x. since their Wronskian is W = f f kf kf = 0. we can show by diﬀerentiating W = u1 u2 − u1 u2 directly that dW + a1 (x)W = 0. u2 ∈ C 1 (a. then two elements v1 . b). note that. Notice that. If u1 . Now let V = C 1 (a. These are homogeneous equations of the form Ax = 0. with k a constant.2). b). b). b). and hence x2 and x3 are linearly independent on (−1. Example The functions u1 (x) = x and u2 (x) = x3 are linearly independent on the interval (−1. u1 (x) u2 (x) u1 (x) u2 (x) α1 α2 = 0 0 . α2 ∈ R. α2 ∈ R such that α1 u1 (x) + α2 u2 (x) = 0 ∀x ∈ (a. v2 ∈ V are linearly dependent if ∃ α1 . The contrapositive of this result is that if W ≡ 0 on (a. 1). such that α1 v1 +α2 v2 = 0.1. the Wronskian of these two functions is W = x2 x3 2x 3x2 = 3x4 − 2x4 = x4 . b) be the set of once-diﬀerentiable functions over the interval a < x < b. since u1 (x) = x2 .2. If the functions u1 and u2 are solutions of (1. then u1 and u2 are linearly independent on (a. by direct diﬀerentiation. the Wronskian of two linearly dependent functions is identically zero on (a. ∃ α1 . are linearly dependent on any interval. b) are linearly dependent. Recall that if V is a vector space over R. In other words. this also gives α1 u1 (x) + α2 u2 (x) = 0 or.2 THE METHOD OF VARIATION OF PARAMETERS 9 1. u2 (x) = x3 .1 The Wronskian Before we carry on. 1). 2 This quantity is not identically zero. which only have nontrivial solutions if det(A) = 0. that is W = u1 (x) u2 (x) u1 (x) u2 (x) = u1 u2 − u1 u2 = 0. in matrix form.

3. and u1 (x1 ) u2 (x1 ) hence u1 (x1 ) = ku2 (x1 ) and u1 (x) = ku2 (x) for some constant k. We just need to exclude the possibility that W is ever zero. and satisﬁes the initial conditions u(x1 ) = 0. homogeneous equation. Theorem 1. which has u1 (x) = x2 and u2 (x) = x as its linearly independent solutions. u ≡ 0.2) by linearity. The vectors and are then linearly dependent. Suppose that W (x1 ) = u1 (x1 ) u2 (x1 ) 0. We end this section with a couple of useful theorems. nonsingular ordinary diﬀerential equation (1. the Wronksian is −x2 . The nonsingularity of the diﬀerential equation is crucial here. the Wronskian is identically positive. Hence u1 and u2 are linearly dependent – a contradiction. Example Consider the equation y + 1 1 y + 1− 2 x x x y = 0. The function u(x) = u1 (x) − ku2 (x) is also a solution of (1. and since the exponential function does not change sign. To ﬁnd this constant. Theorem 1. SECOND ORDER DIFFERENTIAL EQUATIONS This ﬁrst order diﬀerential equation has solution x W (x) = W (x0 ) exp − x0 a1 (t)dt . u (x1 ) = 0. is the only solution. then .10 VARIABLE COEFFICIENT.2) has a unique solution. Proof From Abel’s formula.2).2 (The Sturm separation theorem) If u1 (x) and u2 (x) are the linearly independent solutions of a nonsingular.1 If u1 and u2 are linearly independent solutions of the homogeneous.7) which is known as Abel’s formula. identically negative or identically zero. then the Wronskian is either strictly positive or strictly negative. it is usually necessary to know more about the solutions u1 (x) and u2 (x). which vanishes at x = 0.2). We will describe a technique for doing this in Section 1. the obvious solution. This is because the coeﬃcient of y also vanishes at x = 0. This gives us an easy way of ﬁnding the Wronskian of the solutions of any second order diﬀerential equation without having to construct the solutions themselves. Using Abel’s formula. If we consider the equation x2 y − 2xy + 2y = 0. this has Wronskian W (x) = W (x0 ) exp − dt x0 t = A x0 W (x0 ) = x x for some constant A. (1. (1. Since (1. This means that u1 ≡ ku2 .

8) and hence dy = an (n + c)xn+c−1 . In other words.11) . This means that u1 (xi ) and u2 (xi ) are nonzero. From this we have seen that we can easily construct the second independent solution and. in the inhomogeneous case. If we consider any two of the zeros of sin ωx. (1.1 The Roots of the Indicial Equation Diﬀer by an Integer Consider the equation x2 d 2y dy 1 + x2 − +x dx2 dx 4 y = 0.3. We also know that W (x) is of one sign on [x1 . u1 (x) must change sign. it is immediately clear that cos ωx has a zero between them.3 Solution by Power Series: The Method of Frobenius Up to this point. a0 .3 SOLUTION BY POWER SERIES: THE METHOD OF FROBENIUS 11 the zeros of u1 (x) and u2 (x) occur alternately. as we claimed.1. are as yet undetermined. Proof Suppose that x1 and x2 are successive zeros of u2 (x). so that W (xi ) = u1 (xi )u2 (xi ) for i = 1 or 2. with the simplest case ﬁrst. in which we cannot determine a solution of the homogeneous problem by inspection. . since u1 (x) and u2 (x) are linearly independent. As an example of this. we have considered ordinary diﬀerential equations for which we know at least one solution of the homogeneous problem. . since u2 (x2 ) is zero. we will give some idea of why and when this method can be used. x2 ].9) (1.10) where the constants c. and hence u1 has a zero in [x1 . a1 . We now turn our attention to the more diﬃcult case. dx n=0 d 2y = an (n + c)(n + c − 1)xn+c−2 . We proceed by example. dx2 n=0 ∞ ∞ (1. Later on. (1. x2 ]. . successive zeros of u1 (x) are separated by successive zeros of u2 (x) and vice versa. Now if u2 (x1 ) is positive then u2 (x2 ) is negative (or vice versa). 1. As we noted at the start of the chapter. This is known as the method of Frobenius. This suggests that we can look for a solution of the form ∞ ∞ y = xc n=0 an xn = n=0 an xn+c . For the moment. consider the equation y +ω 2 y = 0. we will restrict our attention to the case where the variable coeﬃcients are simple polynomials. 1. which has solution y = A sin ωx + B cos ωx. . we will just try to make it work. We must devise a method that is capable of solving variable coeﬃcient ordinary diﬀerential equations in general. the particular integral. Since the Wronskian cannot change sign between x1 and x2 .

∞ ∞ am−2 xm+c = m=2 n=2 an−2 xn+c .12 VARIABLE COEFFICIENT. SECOND ORDER DIFFERENTIAL EQUATIONS We substitute (1.11).12) begin at the same power of x. and hence. and n = ∞ implies that m = ∞). (1.10) into (1. If we let m = n + 2 in the last summation (notice that if n = 0 then m = 2. after simplifying the terms in the ﬁrst summation. We now extract the ﬁrst two terms from the ﬁrst summation to give a0 c2 − ∞ 1 4 xc + a1 (c + 1)2 − 1 4 ∞ 1 4 xc+1 + n=2 an (n + c)2 − xn+c + n=0 an xn+c+2 = 0. The two summations in (1.12) becomes a0 c2 − ∞ 1 4 xc + a1 (c + 1)2 − 1 4 ∞ 1 4 xc+1 + n=2 an (n + c)2 − xn+c + m=2 am−2 xm+c = 0. namely x2+c .12) Notice that the ﬁrst term is the only one containing xc and similarly for the second term in xc+1 .8) to (1. and hence a0 c2 − ∞ 1 4 xc + a1 (c + 1)2 − 1 4 ∞ 1 4 xc+1 + n=2 an (n + c)2 − xn+c + n=2 an−2 xn+c = 0. ∞ an (n + c)2 − n=0 1 4 ∞ xn+c + n=0 an xn+c+2 = 0. (1. which gives ∞ ∞ x2 n=0 an (n + c)(n + c − 1)xn+c−2 + x n=0 ∞ an (n + c)xn+c−1 + x2 − 1 4 an xn+c = 0. Since the variables in the summations are merely dummy variables. . n=0 We can rearrange this slightly to obtain ∞ an (n + c)(n + c − 1) + (n + c) − n=0 1 4 ∞ xn+c + n=0 an xn+c+2 = 0.

we need to take some care to avoid simple slips. (1. and implies that c = ± 1 . . It is here that the diﬀerent structures come into play.13) must hold for all values of x. The coeﬃcient of xc is therefore a0 c2 − 1 4 = 0. 2 We now progress to the individual terms in the summation. (1.3 SOLUTION BY POWER SERIES: THE METHOD OF FROBENIUS 13 Since the last two summations involve identical powers of x. if we choose c = − 1 2 the indicial equation is satisﬁed for arbitrary values of a1 . and a1 will act as the second arbitrary constant for the solution. proportional to xc+1 . the coeﬃcient of each power of x must be zero.14) we obtain a2 = − For n = 3. if we were to do this. We therefore assume that a0 = 0. Since (1. If we were to use the solution a0 = 0. we ﬁnd that a1 (c + 1)2 − 1 4 = 0. . However. 2·1 . the series (1.1.14) an = − n(n − 1) By putting n = 2 in (1.13) Although the operations above are straightforward. This is just equivalent to increasing c by 1. Now. We start by rearranging to give an−2 . We solve it by observation as follows. . we therefore let c = − 1 . . This is 4 called the indicial equation. which means that c must satisfy c2 − 1 = 0. a3 = − a1 . Choosing c = 1 gives a1 = 0. progressing to the 2 next term. The general term yields an n− 1 2 2 − 1 4 + an−2 = 0 for n = 2. and. we would ﬁnd that we had 2 constructed a solution with one arbitrary constant. In order to generate this more general solution.8) would have a1 xc+1 as its ﬁrst term. 3·2 a0 . This is called a recurrence relation. Up to this point. 3. most Frobenius analysis is very similar. we can combine them to obtain a0 c2 − ∞ 1 4 xc + a1 (c + 1)2 − 1 4 1 4 xc+1 + n=2 an (n + c)2 − + an−2 xn+c = 0.

. we must show that (1. c = α. Starting from (1. This series splits naturally into two proportional to a0 and a1 . 1/2 x x since we can recognize the Taylor series expansions for sine and cosine. (2n)! (2n + 1)! a5 = − (1. This particular diﬀerential equation is an example of the use of the method of Frobenius.15) is true for n = 0 and that. both solutions can be generated by putting c = α in the recurrence relation. Now. formalized by y(x) = a0 Frobenius General Rule I If the indicial equation has two distinct roots. using the forms of the even and odd coeﬃcients given in (1. namely y = x−1/2 a0 1 − The solution is therefore cos x sin x + a1 1/2 .15). SECOND ORDER DIFFERENTIAL EQUATIONS For n = 4. using the expression for a3 in terms of a1 .14 VARIABLE COEFFICIENT. 5·4 5·4 3·2 5! A pattern is emerging here and we propose that a0 a1 a2n = (−1)n . β (α < β). we substitute c = − 1 . and one of the coeﬃcients of xk becomes indeterminate on putting c = α. a2n+1 = (−1)n . 4·3 and substituting for a2 in terms of a0 gives a4 = − a0 a0 a0 1 − = = . although we will not dwell upon the details here. a4 = − 1 a1 a1 a3 =− − = . x4 x2 + − ··· 2! 4! + x−1/2 a1 x − x5 x3 + − ··· 3! 5! . as given by substituting n + 1 for n in (1.8). 4·3 2·1 4·3·2·1 4! Similarly for n = 5. † In the usual way. a2 . we obtain a2(n+1)+1 .15). when the value of a2n+1 is substituted into the recurrence relation. whose diﬀerence is an integer.15) This can be proved in a straightforward manner by induction. y = x−1/2 a0 + a1 x − a1 x3 a0 x4 a1 x5 a0 x2 − + + + ··· 2! 3! 4! 5! . 2 and write out the ﬁrst few terms in the summation y = x−1/2 (a0 + a1 x + a2 x2 + · · · ).† We can now deduce the full solution.

Collecting similar terms gives ∞ an {2(n + c)(n + c − 1) + (n + c)}xn+c−1 n=0 ∞ + n=0 an {3 − 2(n + c)(n + c − 1) − (n + c)}xn+c = 0. and hence ∞ ∞ an (n + c)(2n + 2c − 1)x n=0 n+c−1 + n=0 an {3 − (n + c)(2n + 2c − 1)}xn+c = 0.3 SOLUTION BY POWER SERIES: THE METHOD OF FROBENIUS 15 In the above example the indicial equation was c2 − 1 = 0. which has solutions c = 4 ± 1 . As in the previous example. The coeﬃcient of xc+1 was a1 (c + 1)2 − 1 = 2 4 0. let’s assume that the solution can be written as the power series (1. The various terms can be multiplied out. in other words a1 is indeterminate.2 The Roots of the Indicial Equation Diﬀer by a Noninteger Quantity We now consider the diﬀerential equation dy d 2y + 3y = 0.16) + (1 − x) 2 dx dx As before. (1. We now extract the ﬁrst term from the left hand summation so that both summations start with a term proportional to xc .1.3. whose diﬀerence is an integer. This gives ∞ a0 c(2c − 1)xc−1 + n=1 an (n + c)(2n + 2c − 1)xn+c−1 . this can be diﬀerentiated and substituted into the equation to yield 2x(1 − x) ∞ ∞ 2x(1 − x) n=0 an (n + c)(n + c − 1)xn+c−2 + (1 − x) n=0 ∞ an (n + c)xn+c−1 +3 n=0 an xn+c = 0. When we choose the lower of the two values (c = − 1 ) this expression does not 2 give us any information about the constant a1 .8). which gives us ∞ ∞ an (n + c)(n + c − 1)2xn+c−1 − n=0 ∞ ∞ n=0 an (n + c)(n + c − 1)2xn+c ∞ + n=0 an (n + c)xn+c−1 − n=0 an (n + c)xn+c + 3 n=0 an xn+c = 0. 1.

Case I: c = 0 In this case. this implies that c = 0 or c = 1 . Since a0 = 0. a1 = a0 − For n = 2 a2 = a1 − 1 3 = −3a0 − 1 3 = a0 . which yields 3 1 = −3a0 . . we can combine the two summations to give a0 c(2c − 1)xc−1 ∞ + n=1 {an (n + c)(2n + 2c − 1) + an−1 {3 − (n − 1 + c)(2n + 2c − 3)}}xn+c−1 = 0. . 2. As in the previous example we can now consider the coeﬃcients of successive powers of x. which gives ∞ a0 c(2c − 1)xc−1 + n=1 ∞ an (n + c)(2n + 2c − 1)xn+c−1 + n=1 an−1 {3 − (n − 1 + c)(2n + 2c − 3)} xn+c−1 = 0. a0 c(2c − 1) = 0. SECOND ORDER DIFFERENTIAL EQUATIONS ∞ + n=0 an {3 − (n + c)(2n + 2c − 1)}xn+c = 0. . considering each root of the indicial equation separately.17) We now need to solve this recurrence relation. which gives the indicial equation. Finally. m=1 We again note that m is merely a dummy variable which for ease we rewrite as n. which then becomes ∞ am−1 {3 − (m − 1 + c)(2(m − 1) + 2c − 1)}xm+c−1 . We recall that this holds for n 1. we can rewrite the recurrence relation (1. . (1. so we start with n = 1. We start with the coeﬃcient of xc−1 . We now let m = n + 1 in the second summation.17) as an = an−1 (n − 1)(2n − 3) − 3 n(2n − 1) = an−1 2n2 − 5n n(2n − 1) = an−1 2n − 5 2n − 1 . . The general term in the summation shows that an = an−1 (n + c − 1)(2n + 2c − 3) − 3 (n + c)(2n + 2c − 1) for n = 1.16 VARIABLE COEFFICIENT. Notice that these 2 roots do not diﬀer by an integer.

Now progressing to n = 3. frob = polyval(a.18) Note that there is no obvious way of writing this solution in terms of elementary functions. A simple MATLAB† function that evaluates (1. Figure 1. (2n − 1)(2n − 3) n=0 ∞ This can be tidied up by putting 3a0 = A. this would be very complicated.2 can then be produced using the command ezplot(@frob.3 SOLUTION BY POWER SERIES: THE METHOD OF FROBENIUS 17 where we have used the expression for a1 in terms of a0 . The function polyval evaluates the polynomial formed by the ﬁrst 100 terms in the sum (1./(2*n-3). 5·3 which again can be proved by induction. † See Appendix 7 for a short introduction to MATLAB. Although we could now use the method of reduction of order. a4 = a3 Finally.1]). In addition. It is easier to consider the second root of the indicial equation.[-1. (2n − 1)(2n − 3) n=0 ∞ (1. since we have constructed a solution. 9·7 3 7 = a0 3 .x). so that the solution is y=A xn . a simple application of the ratio test shows that this power series is only convergent for |x| 1. (2n − 1)(2n − 3) 5 9 = a0 3 . an = 3a0 ./(2*n-1). . 7·5 1 5 = a0 3 .18) in an eﬃcient manner.   which sums the ﬁrst 100 terms of the series. for reasons that we discuss below. for n = 5 we have a5 = a4 In general.18) is   function frob = frob(x) n = 100:-1:0. we have a3 = a2 and for n = 4.1. a = 1. We conclude that one solution of the diﬀerential equation is ∞ y=x c n=0 an x = x n 0 3a0 xn .

2. SECOND ORDER DIFFERENTIAL EQUATIONS Fig. 3. 1.18) can be written as 1 y = − A 3x − 2 + 3x1/2 (1 − x) log 6 1 + x1/2 (1 − x) 1/2 . formalized by . (1. . which explains why the radius of convergence of the power series solution (1.18) is |x| 1. since all successive ai will be written in terms of a2 .3). This diﬀerential equation is an example of the second major case of the method of Frobenius. . Case II: c = 1 2 In this case. . This expression has a logarithmic singularity in its derivative at x = 1.16) given by (1. The second solution of the equation is therefore y = Bx1/2 (1 − x). (1.17) to give n− 1 2 an = an−1 (2n − 2) − 3 n + 1 2n 2 = an−1 2n2 − 3n − 2 2n2 + n = an−1 (2n + 1)(n − 2) n(2n + 1) = an−1 n−2 n . which gives a1 = a0 (−1). to determine an analytical formula for the ﬁrst solution.18 VARIABLE COEFFICIENT.18). The solution of (1. we simplify the recurrence relation (1. . For example.18). for 0 x 1. We again recall that this relation holds for n 1 and start with n = 1. ai = 0 for i = 2. Substituting n = 2 gives a2 = 0 and. we ﬁnd that (1. We can now use this simple solution in the reduction of order formula.

. 2.3 The Roots of the Indicial Equation are Equal Let’s try to determine the two solutions of the diﬀerential equation x dy d 2y + 2y = 0. We know that there must be two solutions.8).3.3 SOLUTION BY POWER SERIES: THE METHOD OF FROBENIUS 19 Frobenius General Rule II If the indicial equation has two distinct roots. c = α. β (α < β).1. + (1 + x) 2 dx dx ∞ We substitute in the standard power series. which gives ∞ x n=0 an (n + c)(n + c − 1)xn+c−2 + (1 + x) n=0 ∞ an (n + c)xn+c−1 +2 n=0 an xn+c = 0. (1. 1. Now shifting the series using m = n + 1 (and subsequently changing dummy variables from m to n) we have ∞ a0 c2 xc−1 + n=1 {an (n + c)2 + an−1 (n + c + 1)}xn+c = 0. The recurrence relation is then an = −an−1 When n = 1 we ﬁnd that a1 = −a0 2 . For the moment let’s see how far we can get by setting c = 0.19) where we have combined the two summations. We can extract the ﬁrst term from the left hand summation to give ∞ ∞ a0 c2 xc−1 + n=1 an (n + c)2 xn+c−1 + n=0 an (n + c + 2)xn+c = 0. The indicial equation is c2 = 0 which has a double root at c = 0. 12 n+1 n2 for n = 1. This can be simpliﬁed to give ∞ ∞ an (n + c)2 xn+c−1 + n=0 n=0 an (n + c + 2)xn+c = 0. . whose diﬀerence is not an integer. . . . but it appears that there is only one available to us. (1. the general solution of the equation is found by successively substituting c = α then c = β into the general recurrence relation.

n! ∞ y = a0 n=0 (−1)n which can also be written as y = a0 x ∞ (−1)n xn−1 (−1)n xn + (n − 1)! n! n=1 n=0 = a0 (1 − x)e−x . which we write out more fully as x ∞ dy d 2y + 2y = + (1 + x) dx2 dx a0 c2 xc−1 + n=1 {an (n + c)2 + an−1 (n + c + 1)}xn+c = 0. as this gets rid of most of the terms. . and leaves us with dy d 2y + 2y = a0 c2 xc−1 . a0 = (−1)n (n!)2 n! (n + 1) n x .19). where we regard y as a function of both x and c. a2 = −a1 Using n = 3 gives a3 = −a2 and we conclude that an = (−1)n One solution is therefore ∞ 3 3·2 = a0 2 2 .20) x 2 + (1 + x) dx dx Let’s now take a partial derivative with respect to c. This gives us an as a function of c for n 1. 22 1 ·2 4 4·3·2 = −a0 2 2 2 . dx ∂x ∂c This gives x ∂2 ∂x2 ∂y ∂c + (1 + x) ∂ ∂x ∂y ∂c +2 ∂y ∂c = a0 ∂ 2 c−1 (c x ). ∞ = a0 −x (−1)m xm + e−x m! m=0 This solution is one that we could not have readily determined simply by inspection. ∂c ∂y ∂x = ∂ ∂x ∂y ∂c . SECOND ORDER DIFFERENTIAL EQUATIONS and with n = 2. Consider (1. We could now use the method of reduction of order to ﬁnd the second solution. making use of ∂ ∂ d = . but we will proceed with the method of Frobenius so that we can see how it works in this case. 2 3 1 ·2 ·3 (n + 1)! n+1 a0 .20 VARIABLE COEFFICIENT. (1. The best we can do at this stage is to set an (n + c)2 + an−1 (n + c + 1) = 0 for n 1.

the other coeﬃcients are. ∂c dc n=0 n=0 using a similar technique as before to deal with the diﬀerentiation of xc with respect to c. we diﬀerentiate the power series (1. c=0 an = − an−1 (n + c + 1) . (n + c)2 . ∂c which we can tidy up to give ∂ 2 c−1 (c x ) = c2 xc−1 log x + xc−1 2c. Diﬀerentiating using the product rule we have ∂ ∂ ∂ 2 c−1 (c x ) = c2 (xc−1 ) + xc−1 (c2 ). ∂c ∂c ∂c Diﬀerentiating the exponential gives ∂ 2 c−1 (c x ) = c2 (log x ec log x )x−1 + xc−1 2c. We need to be careful when evaluating the right hand side of this expression. Putting c = 0 gives ∂y ∂c ∞ ∞ ∞ = c=0 n=0 dan dc dan dc ∞ x + log x c=0 n=0 n an |c=0 xn . ∂y We conclude that is a second solution of our ordinary diﬀerential equation. ∂c Substituting this form back into the diﬀerential equation gives x ∂2 ∂x2 ∂y ∂c + (1 + x) ∂ ∂x ∂y ∂c +2 ∂y = a0 {c2 xc−1 log x + xc−1 2c}. so that we have ∂ ∂ ∂ 2 c−1 (c x ) = c2 (ec log x x−1 ) + xc−1 (c2 ). ∂c ∂c ∂c We rewrite xc−1 as xc x−1 = ec log x x−1 . We begin with the recurrence relation.20) has a repeated root at c = 0.3 SOLUTION BY POWER SERIES: THE METHOD OF FROBENIUS 21 Notice that we have used the fact that a0 is independent of c.8) (carefully!) to give dan n ∂y = xc x + an xn xc log x. We therefore need to determine which is . although a0 is not a function of c. ∂c c=0 To construct this solution. ∂c Now letting c → 0 gives x ∂ 2 ∂y ∂x2 ∂c + (1 + x) c=0 ∂ ∂y ∂x ∂c +2 c=0 ∂y ∂c = 0. Note that.1. c=0 Notice that this procedure only works because (1.

c+j+1 c+j j=1 n n = c=0 j=1 1 1 −2 . (c + 1)2 (c + 2)2 This process can be continued to give an = (−1)n a0 which we can write as an = (−1) a0 n n j=1 (c (c + 2)(c + 3) . (n!)2 j j j=1 j=1 In this expression. (c + 1)2 (c + 2)2 . . . . a2 = −a1 (c + 3) . We now take the logarithm of this expression. SECOND ORDER DIFFERENTIAL EQUATIONS Starting with n = 1 we ﬁnd that a1 = whilst for n = 2. changed the ﬁrst summation and removed the extra term that this incurs. recalling that the logarithm of a product is the sum of the terms. . (c + 1)2 and substituting for a1 in terms of a0 gives us a2 = a0 (c + 2)(c + 3) . . (c + n)2 + j + 1) + j) n j=1 (c 2. (c + n + 1) . j+1 j j=1  2 Since we know an when c = 0. we can write dan dc = c=0 (−1)n a0 n j=1 (j n j=1 + 1) j   n j=1 1 1 −2 . we have manipulated the products and written them as factorials. j+1 j j=1 n  =  n+1 n (n + 1)!  1 1 (−1)n a0 −1−2 . which leads to     log(an ) = log((−1)n a0 ) + log  n n (c + j + 1) − 2 log  n n (c + j) j=1 j=1 = log((−1)n a0 ) + j=1 log(c + j + 1) − 2 j=1 log(c + j). (c + 2)2 −a0 (c + 2) . Now diﬀerentiating with respect to c gives 1 dan = an dc and setting c = 0 we have 1 dan an dc n n j=1 1 1 −2 .22 VARIABLE COEFFICIENT.

ak = − (k + c)2 − 1 we ﬁnd that a2 = − a0 a0 =− . c = α.3). it is rather diﬃcult to get all of u2 (x) this way. which. we ﬁnd that a1 = 0. to ﬁnd the second linearly independent solution. and. The indicial equation has roots c = ±1. Having got these. we can assume a general series form for u2 (x). (1. we ﬁnd that n=0 an x ∞ a0 (c2 − 1)xc + a1 (1 + c)2 − 1 xc+1 + k=2 ak (k + c)2 − 1 + ak−2 xk+c = 0. we obtain dan dc = (−1)n a0 c=0 (n + 1) (φ(n + 1) − 2φ(n) − 1) . There are several other cases that can occur in the method of Frobenius. (−1)n n! n! n=0 This methodology is formalized in Frobenius General Rule III If the indicial equation has a double root. we will not go into here.1. Of course. n! 1 . and ﬁnd the coeﬃcients in the usual way.22) using the method of Frobenius. by choosing either of these. due to their complexity. (2 + c)2 − 1 (1 + c)(3 + c) . If we now look for the general solution of ak−2 . This can be used in the reduction of order formula. Example Let’s try to solve x2 y + xy + (x2 − 1)y = 0. One method of dealing with these is to notice that the method outlined in this chapter always produces a solution of ∞ the form y = u1 (x) = n=0 an xn+c . y = ∞ n+c . The second independent solution is (∂y/∂c)c=α where an = an (c) for the calculation.3 SOLUTION BY POWER SERIES: THE METHOD OF FROBENIUS 23 Simplifying. If we look for a solution in the usual form. but the ﬁrst few terms are usually easy enough to compute by expanding for small x. m m=1 ∞ n where we have introduced the notation φ(n) ≡ The second solution is therefore ∞ (1. (1.21) y = a0 n=0 (−1) n (n + 1) (n + 1) n {φ(n + 1) − 2φ(n) − 1}xn + x log x . one solution is obtained by putting c = α into the recurrence relation.

If we assume a solution structure of this form and substitute it into (1.3. we obtain one of the linearly independent solutions of (1.4 Singular Points of Diﬀerential Equations In this section. However.1 that the Wronskian of (1. by choosing c = 1. and hence A = 1.22) is W = A/x for some constant A. u1 (x) = y(x. c) = a0 xc 1 − . 2x 4 t2 exp − 1 ds s dt The second linearly independent solution of (1. variable coeﬃcient diﬀerential equation of the form P (x) d 2y dy + R(x)y = 0. we use the reduction of order formula.24 VARIABLE COEFFICIENT. 2. since we know that u1 = x + · · · and u2 = −1/2x + · · · .3). This gives us a solution of the form y(x. we need to further reﬁne our ideas about singular points. 1. it is straightforward to pick oﬀ the coeﬃcients b2n . (1. If x0 is a singular point and (x − x0 )Q(x)/P (x) and (x − x0 )2 R(x)/P (x) have convergent . note that we showed in Section 1.2. Now.22).22). We consider a second order.23) Before we proceed. . SECOND ORDER DIFFERENTIAL EQUATIONS a4 = − a0 a2 = . Now. if c = −1. In order to ﬁnd the structure of the second linearly independent solution. Substituting for u1 (x) gives u2 (x) = x 1− x2 + ··· 8 x 1 2 2 t x2 + ··· 8 x2 = x 1− + ··· 8 = x 1− t 1 − + ··· 8 x 1 1 t2 dt 1 + + ··· t2 4 t 1 1 − 2 + log x + · · · . + Q(x) 2 dx dx (1. 1) = x 1 − x4 x2 + − ··· 2 · 4 2 · 42 · 6 . 2−1 (4 + c) (1 + c)(3 + c)2 (5 + c) x4 x2 + − ··· (1 + c)(3 + c) (1 + c)(3 + c)2 (5 + c) and so on. 4 2x where v(x) = 1 + b2 x2 + b4 x4 + · · · . . we must have W = x(1/2x2 ) + 1/2x + · · · = 1/x + · · · . are singular. . the coeﬃcients a2n for n = 1. Finally.22) therefore has the structure u2 (x) = 1 1 u1 (x) log x − v(x). we give some deﬁnitions and a statement of a theorem that tells us when the method of Frobenius can be used. and for what values of x the inﬁnite series will converge.

24) There are singular points where x2 − x = 0. P x −x x−1 Again.1. otherwise. P which can be multiplied out to give xQ = −1 − 2x + · · · .3 SOLUTION BY POWER SERIES: THE METHOD OF FROBENIUS 25 Taylor series expansions about x0 . using the binomial expansion. Consider the expression x(1 + x) (1 + x) xQ = 2 = = −(1 + x)(1 − x)−1 . P {1 − (1 − x)} Expanding in powers of (1 − x) using the binomial expansion gives (x − 1) Q 1−x =2 1− P 2 {1 + (1 − x) + (1 − x)2 + · · · }. x0 is called an irregular singular point. P Since xQ/P and x2 R/P have convergent Taylor series about x = 0. . + (1 + x) 2 dx dx (1. Now consider the other singular point. so we rewrite x as 1 − (1 − x). then x = x0 is called a regular singular point. We note that (x − 1) (x − 1)(1 + x) (x − 1)(1 + x) 1+x Q = = = . Example Consider the equation x(x − 1) d 2y dy + y = 0. Let’s start by looking at the singular point x = 0. Now x2 R x2 x = 2 = = −x(1 − x)−1 . x2 R = −x(1 + x + x2 + · · · ) = −x − x2 − x3 − · · · . P This power series is convergent provided |x| < 1 (by considering the binomial expansion used above). which is convergent provided |x| < 1. x = 1. and hence (x − 1) 2 − (1 − x) Q = . P x −x (x − 1) Upon expanding (1 − x)−1 using the binomial expansion we have xQ = −(1 + x)(1 + x + x2 + · · · + xn + · · · ). at x = 0 and x = 1. this is a regular singular point. P (x2 − x) x(x − 1) x At this point we need to recall that we want information near x = 1.

Ostberg and Perkins (1966) and is due to Fuchs.5 An outline proof of Theorem 1. R(x) = R0 + xR1 + · · · . which is again convergent provided |x−1| < 1. If x0 is a regular singular point. Therefore x = 1 is a regular singular point. x4 1. SECOND ORDER DIFFERENTIAL EQUATIONS which is a power series in (x − 1) that is convergent provided |x − 1| < 1.3 holds. We also need to consider (x − 1)2 R (x − 1)2 (x − 1)2 x−1 x−1 = 2 = = = P x −x x(x − 1) x {1 − (1 − x)} = (x − 1){1 − (1 − x)}−1 = (x − 1){1 + (x − 1) + (x − 1)2 + · · · }.3. which cannot be n=0 an x expanded about x = 0.3 We will now give a sketch of why Theorem 1. then there exists a unique series solution in the neighbourhood of x0 .23). where ρ is the smaller of the radii of convergence of the series (x − x0 )Q(x)/P (x) and (x − x0 )2 R(x)/P (x). assuming that P (x) = P0 + xP1 + · · · . We are more concerned here with using the result to tell us when a series solution will converge. We give an outline of why the result should hold in Section 1.3.24). is unique. Kuller. − dx2 dx For this equation. where ρ is the smaller of the radii of convergence of the series for Q(x)/P (x) and R(x)/P (x). When x = 0 is an ordinary point. The radii of convergence of xQ(x)/P (x) and x2 R(x)/P (x) ∞ are both unity. The series solution ∞ n+c is not guaranteed to exist. We have already seen that x = 0 is a regular singular point. which converges for |x − x0 | < ρ.3 If x0 is an ordinary point of the ordinary diﬀerential equation (1. and hence the series solution n=0 an xn+c exists. Consider the equation P (x)y + Q(x)y + R(x)y = 0. Theorem 1. and will converge for |x| < 1. Example Consider the diﬀerential equation d 2y dy + y = 0. Proof This can be found in Kreider.26 VARIABLE COEFFICIENT. since xQ/P = −1/x3 . x = 0 is a singular point but it is not regular. . Q(x) = Q0 + xQ1 + · · · . then there exists a unique series solution in the neighbourhood of x0 which converges for |x − x0 | < ρ.5. Example Consider the diﬀerential equation (1.

When x = 0 is a singular point. c(c − 1) = 0. after using dy d 2 y d dy = −s2 . is most easily shown using the theory of complex variables. as claimed by the theorem. this is done in Section A6. When the terms are ordered. Q(s) = 2s3 P 1 s − s2 Q 1 s ˆ . In particular. the simplest case to consider is when P (x) = P1 x + P2 x2 + · · · . The indicial equation. 1. P (s) 2 + Q(s) ds ds s2 dy ds where ˆ P (s) = s4 P 1 s ˆ . This makes it clear that. . P (x) P1 P (x) P1 so that both of these quantities have a Taylor series expansion about x = 0. by making the assumption that R−1 + R0 + · · · . with the usual possibilities for its types of roots. Let’s now try to include the y and y terms as well.1. We can then ask what form Q(x) and R(x) must take to ensure that a series solution exists. the indicial equation was formed from the y term in the equation alone. we can choose c = 0 and ﬁnd a solution of the form y = a0 y0 (x) + a1 y1 (x). R(s) = R 1 s . following Frobenius General Rule I.6 The point at inﬁnity Our discussion of singular points can be extended to include the point at inﬁnity by deﬁning s = 1/x and considering the properties of the point s = 0. = s2 2 dx ds dx ds in (1. we ﬁnd that 2 ˆ ˆ d y ˆ dy + R(s)y = 0. x Then.23). When x = 0 is an ordinary point. . That the series converges. This is a quadratic equation in c.3. As x → 0 Q0 x2 R(x) R−1 xQ(x) → → .5. when P (x) = P1 x + · · · . after substitution of the Frobenius series into the equation. R(x) = P1 c(c − 1) + Q0 c + R−1 = 0. we ﬁnd that the ﬁrst two terms in the expansion are P0 a0 c(c − 1)xc−2 + {P0 a1 c(c + 1) + P1 a1 c(c − 1) + Q0 a1 c} xc−1 + · · · = 0.3 SOLUTION BY POWER SERIES: THE METHOD OF FROBENIUS 27 we can look for a solution using the method of Frobenius. has two distinct roots that diﬀer by an integer and. the coeﬃcient of xc−1 gives the indicial equation as Q(x) = Q0 + Q1 x + · · · . these choices of expressions for the behaviour of Q(x) and R(x) close to x = 0 are what is required to make it a regular singular point.

(c) x cos(log |x|). u2 }. SECOND ORDER DIFFERENTIAL EQUATIONS For example. which we will study in Chapter 3. which is singular at s = 0. u2 . We will study the behaviour of solutions of Bessel’s equation in the neighbourhood of the point at inﬁnity in Section 12.28 VARIABLE COEFFICIENT. and we conclude that the point at inﬁnity is an irregular singular point. −x dx2 dx 2 d y dy + xy = 0. Which of these pairs of functions are linearly independent on the interval [−1. .7. P (s) = s2 . x 2 +2 dx dx (b) u1 = x−1 sin x. Exercises 1.2 Find the Wronskian of (a) x. 1]? Find the general solution of dy d 2y + y = x3/2 ex . W (x) = y y y u1 u1 u1 u2 u2 u2 . In this case. and hence has a regular ˆ ˆ ˆ singular point at x = 0. (a) u1 = ex . −2 dx2 dx d 2y (b) + 4y = 2 sec 2x.4 is zero everywhere. dx2 1 d 2y 1 dy + 1 − 2 y = x. e−x . show that the Wronskian of {y. has P (x) = x2 . x2 . (x − 1) d 2y dy + y = 0. Bessel’s equation also has a singular point at inﬁnity. Bessel’s equation. Use the reduction of order method to ﬁnd the second independent solution. 1. Q(x) = x and R(x) = x2 − ν 2 . u1 . In addition.3 1. subject to y(0) = y (0) = 0. ˆ Since P (0) = 0. dx2 If u1 and u2 are linearly independent solutions of (d) y + p(x)y + q(x)y = 0 and y is any other solution. Q(s) = s and R(s) = 1/s2 − ν 2 .1 Show that the functions u1 are solutions of the diﬀerential equations given below. where ν is a constant. x sin(log |x|). (c) + dx2 x dx 4x (a) d 2y + y = f (x).2. Hence ﬁnd a second order diﬀerential equation which has solutions y = x and y = log x. (b) ex . 1. ˆ ˆ s2 R(s)/P (s) = 1 − s2 ν /s.

Use the method of Frobenius to determine this constant.5 1. + {γ − (α + β + 1)x} dx2 dx Show that x = 0 and x = 1 are regular singular points and obtain the roots of the indicial equation at the point x = 0.7 Find the Wronskian of the solutions of the diﬀerential equation (1−x2 )y −2xy +2y = 0 to within a constant. + (x + 1) 2 dx dx Show that one of them is a regular singular point and that the other is an irregular singular point.6 1.EXERCISES 29 1. For which values of k is there a polynomial solution? Let α.10 (where k is a real constant) in power series form. Express a1 x(1 − x) .9 1. β. +4 2 dx dx which are valid for x > 0. there are two linearly independent solutions of Frobenius form. Find the two linearly independent solutions of each of the diﬀerential equations 1 d 2y 1 dy + y = 0. (a) x2 2 + x x − dx 2 dx 2 d 2y dy (b) x2 2 + x (x + 1) − y = 0. Show that if γ is not an integer. 1. (b) Find two linearly independent Frobenius solutions of (x2 − 1)2 dy d 2y − xy = 0. Obtain one series solution of the equation in the form x(1 − x) ∞ y=A n=1 n2 xn−1 = Au1 (x). Find the general solution of the diﬀerential equation x 2x d 2y dy − ky = 0 + (1 + x) dx2 dx 1. dx dx using the method of Frobenius. Show that the indicial equation for dy d 2y − 4y = 0 + (1 − 5x) dx2 dx has a double root. What is the radius of convergence of this series? Obtain the second solution of the equation in the form u2 (x) = u1 (x) log x + u1 (x) (−4x + · · · ) .8 (a) The points x = ±1 are singular points of the diﬀerential equation d 2y dy − y = 0. γ denote real numbers and consider the hypergeometric equation d 2y dy − αβy = 0.

dx dx d 2y dy (b) x2 2 + − 2y = 0. SECOND ORDER DIFFERENTIAL EQUATIONS and a2 in terms of a0 for each of the solutions.14 1. Show that equation (a) has no solution of Frobenius type but that equation (b) does.12 1. if f (x) and g(x) are nontrivial solutions of the diﬀerential q(x). Verify directly that this function satisﬁes the diﬀerential equation. Show that x = 0 is a regular singular point of the diﬀerential equation d 2y dy − y = 0. but you will need to use the reduction of order formula to determine the structure of the second solution. ∗ Show that. if q(x) 0. 1. f (x) equations u + p(x)u = 0 and v + q(x)v = 0. Hint: It is straightforward to ﬁnd one solution.11 Show that each of the equations d 2y dy (a) x3 2 + x2 + y = 0. + x(1 − x) dx2 dx Find two linearly independent Frobenius solutions and show that one of these solutions can be expressed in terms of elementary functions.13 1. and p(x) vanishes at least once between any two zeros of g(x) unless p ≡ q and f = µg. dx dx has an irregular singular point at x = 0. Find the two linearly independent solutions of the ordinary diﬀerential equation 2x2 x(x − 1)y + 3xy + y = 0 in the form of a power series.15 . no nontrivial solution of u + q(x)u = 0 can have more than one zero. µ ∈ R (this is known as the Sturm comparison theorem). 1. ∗ Show that.30 VARIABLE COEFFICIENT. Obtain this solution and hence ﬁnd the general solution of equation (b).

The Legendre functions are the solutions of Legendre’s equation.6.CHAPTER TWO Legendre Functions Legendre’s equation occurs in many areas of applied mathematics. The equation is usually deﬁned for −1 < x < 1 for reasons that will become clear in Section 2.1) − 2x dx2 dx where n is a parameter. and takes the form d 2y dy + n(n + 1)y = 0. Substitution of this series into (2.1 Deﬁnition of the Legendre Polynomials. called the order of the equation. a series solution will be convergent for |x| < 1. a second order linear diﬀerential equation with variable coeﬃcients. (1 − x2 ) 2. We can see immediately that x = 0 is an ordinary point of the equation. and seek a power series solution of (2.3.1) in the form ∞ y= i=0 ai xi+c . physics and chemistry in physical situations with a spherical geometry. (2.1) leads to ∞ ∞ (1 − x2 ) i=0 ai (i + c)(i + c − 1)xi+c−2 − 2x i=0 ∞ ai (i + c)xi+c−1 +n(n + 1) i=0 ai xi+c = 0. We are now in a position to examine Legendre’s equation in some detail using the ideas that we developed in the previous chapter. and. Pn (x) We will use the method of Frobenius. by Theorem 1. Tidying this up gives ∞ ∞ ai (i + c)(i + c − 1)xi+c−2 + i=0 i=0 ai {n(n + 1) − (i + c)(i + c + 1)}xi+c = 0. and collecting like powers of x we get ∞ a0 c(c − 1)xc−2 + a1 c(c + 1)xc−1 + i=2 ai (i + c)(i + c − 1)xi+c−2 . This equation was introduced by Legendre in the late 18th century.

5·4 5! n(n + 1) 2 {3 · 2 − n(n + 1)}n(n + 1) 4 x − x + ··· 2! 4! {3 · 2 − n(n + 1)} −{3 · 2 − n(n + 1)}n(n + 1) a2 = a0 . so that a1 is arbitrary and acts as a second constant. convergent for |x| < 1. we choose c = 0. . If n is a positive integer. 2 (i − 1)(i − 2) − n(n + 1) ai−2 for i = 2. when n is an integer. The indicial equation is therefore c(c − 1) = 0. which gives c = 0 or 1. . . a5 = {4 · 3 − n(n + 1)} {4 · 3 − n(n + 1)}{2 · 1 − n(n + 1)} a3 = a1 .2) where a0 and a1 are arbitrary constants. Following Frobenius General Rule I. a3 = For i = 4. If n is not a positive integer. Rearranging the series gives us a0 c(c − 1)xc−2 + a1 c(c + 1)xc−1 ∞ + i=2 {ai (i + c)(i + c − 1) + ai−2 (n(n + 1) − (i + c − 2)(i + c − 1))}xi+c−2 = 0.2) as y = a0 un (x) + a1 vn (x). In general we have. we have two inﬁnite series solutions. 3! 5! (2. . one of the inﬁnite series terminates to give a simple polynomial solution. . 4·3 4! {2 · 1 − n(n + 1)} a1 . 3. i(i − 1) The solution of Legendre’s equation is therefore y = a0 1 − +a1 x + {2 · 1 − n(n + 1)} 3 {4 · 3 − n(n + 1)}(2 · 1 − n(n + 1)) 5 x + x + ··· . 3·2 −n(n + 1) a0 .32 LEGENDRE FUNCTIONS ∞ + i=0 ai {n(n + 1) − (i + c)(i + c + 1)}xi+c = 0. a4 = For i = 5. a2 = For i = 3. If we write the solution (2. ai = For i = 2.

so we have to pick oﬀ the Legendre polynomial as the ﬁrst row of the matrix p. pout = [pout.2. 2 8 4 8 Graphs of these Legendre polynomials are shown in Figure 2. It is convenient to write the solution in the form y = + BQn (x). Q1 (x) = 1 x log 2 1+x 1−x − 1.pout(2.1.1. 2. end plot(x. which we will meet later. Q3 (x) = (5x3 − 3x) log 2 4 These functions are singular at x = ±1.:).x.x.:). x.’P_2(x)’.:). 1.:). which was generated using the MATLAB script ' $x = linspace(-1.’. The ﬁrst ﬁve of these are 1 P0 (x) = 1. Simple expressions for the Qn (x) are available for n = 0.3). (1.x) generates both the Legendre polynomial of order n and the associated Legendre functions of orders 1 to n.’-. of degree n converges for |x| < 1 where we have deﬁned Pn (x) = un (x)/un (1) for n even and Pn (x) = vn (x)/vn (1) for n odd. pout = [].’P_3(x)’.. P2 (x) = (3x2 − 1).. P1 (x) = x. 2 1 35 4 15 2 3 (5x3 − 3x). p].1 DEFINITION OF THE LEGENDRE POLYNOMIALS. P4 (x) = x − x + . − r)!(n − 2r)! where m is the integer part of n/2. v1 (x). The polynomials Pn (x) are called the Legendre polynomials and can be written as m Pn (x) = r=0 (−1)r 2n r!(n (2n − 2r)!xn−2r . 2 3 Q2 (x) = 1 (3x2 − 1) log 4 1+x 1−x 1 3 − x. for k = 1:4 p = legendre(k.’:’) legend(’P_1(x)’.pout(4. APn (x) ↑ ↑ Polynomial Inﬁnite series.pout(3.:). 1+x 1−x 2 5 − x2 + .’--’.’P_4(x)’. Note that by deﬁnition Pn (1) = 1.xlabel(’x’) & % Note that the MATLAB function legendre(n. Pn (x) 33 then u0 (x). 3 using the reduction of order formula.x). u2 (x) and v3 (x) are the ﬁrst four polynomial solutions.4). p=p(1..500).pout(1. Notice that part of the inﬁnite series has . In particular P3 (x) = Q0 (x) = 1 log 2 1+x 1−x . Firstly.6).34 LEGENDRE FUNCTIONS Fig. P3 (x) and P4 (x). 1 − x2 for −1 < x < 1. Abel’s formula. 1 − s2 . shows that the Wronskian is W = P1 (s)Q1 (s) − P1 (s)Q1 (s) = W0 . This is just an inhomogeneous version of Legendre’s equation of order one.2. (1. The variation of parameters formula. 2. Example Let’s try to ﬁnd the general solution of (1 − x2 )y − 2xy + 2y = 1 .7). been summed to give us the logarithmic terms. (1. The complementary function is yh = AP1 (x) + BQ1 (x). The Legendre polynomials P1 (x). then shows that the particular integral is x yp = P1 (s)Q1 (x) − P1 (x)Q1 (s) ds. Graphs of these functions are shown in Figure 2.1. P2 (x). (1 − s2 ){P1 (s)Q1 (s) − P1 (s)Q1 (s)} We can considerably simplify this rather complicated looking result. The ﬁrst four Legendre functions. 2 = 1+x 1−x The general solution is this particular integral plus the complementary function (yp (x) + yh (x)). Q0 (x).2 THE GENERATING FUNCTION FOR Pn (x) 35 Fig. G(x.2. t) around t = 0 are Pn (x). 1/(1 − s2 ) = 1 + s2 + · · · W = 1 + s2 + · · · for s for s 1. 2. t) = (1 − 2xt + t2 )−1/2 = ∞ Pn (x)tn . Q1 (s) = 1 s log 2 1+s 1−s − 1.3) .2. From the binomial theorem. t). We start with the assertion that G(x. This function is deﬁned in such a way that the coeﬃcients of the Taylor series of G(x. Since P1 (s) = s. Q1 (x). it is helpful to introduce a generating function. This means that x x yp = Q1 (x) 1 2 x Q1 (x) − x 2 P1 (s) ds − P1 (x) 1 2 (x − 1) log 4 Q1 (s) ds 1 − x . n=0 (2. We conclude that W0 = 1. Q2 (x) and Q3 (x). We can determine the constant W0 by considering the behaviour of W as s → 0. 1. 2.2 The Generating Function for Pn (x) In order to make a more systematic study of the Legendre polynomials. {1 + (−2xt + t2 )}−1/2 = 1+ − 1 2 (−2xt + t2 ) + −1 2 2! −3 2 (−2xt + t2 )2 + · · · 1 = 1 + xt + (3x2 − 1)t2 + · · · = P0 (x) + P1 (x)t + P2 (x)t2 + · · · . using the binomial expansion. We start by working with (1 − 2xt + t2 )−1/2 = ∞ Zn (x)tn .4) with respect to t. n=0 Now we diﬀerentiate (2.4) where. which leads to (x − t)(1 − 2xt + t2 )−3/2 = 2 ∞ Zn (x)ntn−1 . Using the binomial expansion formula. n=0 (2. With a little extra work we can derive (2. We ﬁrst diﬀerentiate with respect to x. n=0 Zn (x)tn . 2 as expected. ∂t = t2 {(x − t)(1 − 2xt + t2 )−5/2 3(x − t) + (1 − 2xt + t2 )−3/2 − 1} +(x − t)(1 − 2xt + t2 )−3/2 2t. which we can simplify to get (1 − 2xt + t2 )−3/2 {3t2 (x − t)2 (1 − 2xt + t2 )−1 − 1 + 2t(x − t)} ∞ = n=0 Zn (x)n(n + 1)tn . n=0 Multiplying this last result by t and diﬀerentiating with respect to t gives ∞ Zn (x)n(n + 1)tn = n=0 ∂ 2 {t (x − t)(1 − 2xt + t2 )−3/2 }. which is convergent for t2 − 2xt < 1 (for |x| < 1 we can ensure that this holds by making |t| small enough). let’s consider the ﬁrst terms in the Taylor series expansion about t = 0. .3). we know that Zn (x) is a polynomial of degree n. which gives us t(1 − 2xt + t2 )−3/2 = and again gives 3t2 (1 − 2xt + t2 )−5/2 = ∞ ∞ Zn (x)tn .36 LEGENDRE FUNCTIONS Just to motivate this formula. which is just Legendre’s equation. . as required. For example.3) gives (1 + 2t + t2 )−1/2 = By the binomial expansion we have that (1 + 2t + t2 )−1/2 = {(1 + t)2 }−1/2 = (1 + t)−1 ∞ ∞ Pn (−1)tn . This means that Zn (x) = αPn (x) + βQn (x) where α. Therefore (1 − x2 )Zn (x) − 2xZn (x) + n(n + 1)Zn (x) = 0. 2. We will give a few examples of its application. The generating function. Special Values The generating function is useful for determining the values of the Legendre polynomials for certain special values of x. Finally. n=0 = 1 − t + t2 − · · · + (−t)n + · · · = n=0 (−1)n tn .2. to obtain (1 − 2t + t2 )−1/2 = Since (1 − 2t + t2 )−1/2 = (1 − t)2 −1/2 ∞ Zn (1)tn . substituting x = −1 in (2. as well as some recurrence formulae. We conclude that ∞ ∞ (−1) t = n=0 n=0 n n Pn (−1)tn . we have Zn (1) = 1. . and therefore Pn (−1) = (−1)n for n = 1. Since we know that Pn (1) = 1. . can be used to prove a number of interesting properties of Legendre polynomials. we conclude that Zn (x) ≡ Pn (x). . n=0 ∞ = (1 − t)−1 = tn . . we need to show that Zn (1) = 1. (2. β must be zero.3). β are constants. t) = (1 − 2xt + t2 )−1/2 . for any t. As Zn (x) is a polynomial of degree n. n=0 at least for |t| < 1.2 THE GENERATING FUNCTION FOR Pn (x) 37 Combining all of these results gives ∞ ∞ ∞ (1 − x2 ) n=0 Zn (x)tn − 2x n=0 Zn (x)tn + n=0 n(n + 1)Zn (x)tn = (1 − x2 )3t2 (1 − 2xt + t2 )−5/2 − 2xt(1 − 2xt + t2 )−3/2 +(1 − 2xt + t2 )−3/2 {3t2 (x − t)2 (1 − 2xt + t2 )−1 − 1 + 2t(x − t)} = 0. This is done by putting x = 1 in the generating function relationship. G(x. Starting with P0 (x) = 1 and P1 (x) = x.38 LEGENDRE FUNCTIONS 2. and hence (n + 1)Pn+1 (x) − (2n + 1)xPn (x) + nPn−1 (x) = 0. Pn (x)tn − n=0 Pn (x)tn+1 ∞ =n n=0 Pn (x)tn−1 − 2xn n=0 n Pn (x)tn + n n=0 Pn (x)tn+1 . (2.5). n=0 Diﬀerentiation of (2. Equating coeﬃcients of t on both sides shows that xPn (x) − Pn−1 (x) = (n + 1)Pn+1 (x) − 2xnPn (x) + (n − 1)Pn−1 (x). (2. which gives 2P2 (x) − 3x2 + 1 = 0. Firstly.3). 2 By iterating this procedure. we diﬀerentiate the generating function.3 Diﬀerential and Recurrence Relations Between Legendre Polynomials The generating function can also be used to derive recurrence relations between the various Pn (x).3) with respect to t gives (x − t)(1 − 2xt + t2 )−3/2 = ∞ ntn−1 Pn (x).5) This is a recurrence relation between Pn+1 (x). In a rather similar manner we can generate a recurrence relation that involves the derivatives of the Legendre polynomials. and ﬁnd that (x − t)(1 − 2xt + t2 )−3/2 = 2 ∞ Pn (x)ntn−1 . which can be used to compute the polynomials Pn (x). we substitute n = 1 into (2. we diﬀerentiate with respect to t.3). we can generate the Legendre polynomials Pn (x) for any n. n=0 . with respect to x to get P2 (x) = t(1 − 2xt + t2 )−3/2 = ∞ tn Pn (x). n=0 We now multiply through by (1 − 2xt + t ) to obtain (x − t)(1 − 2xt + t ) which leads to x n=0 ∞ ∞ ∞ 2 −1/2 ∞ = (1 − 2xt + t ) 2 n=0 ∞ Pn (x)ntn−1 . Pn (x) and Pn−1 (x). and hence 1 (3x2 − 1). Starting with (2. An Example From Electrostatics In electrostatics. and the most useful of these is Rodrigues’ formula. the potential due to a unit point charge at r = r0 is V = 1 . dx1 1 2n n! dn {(x2 − 1)n }. V = r 1 Pn (cos θ) a n=0 a ∞ n . As we would expect from the symmetry of the problem. for n = 1. Pn (x) = For example.4 Rodrigues’ Formula There are other methods of generating the Legendre polynomials. there is no dependence upon the azimuthal angle. (r. φ). z = r cos θ. z = a. y = r sin θ sin φ. dxn . and hence V =√ 1 1 = 2 + a2 − 2ar cos θ a r r2 r 1 − 2 cos θ + 2 a a −1/2 . x = r sin θ cos φ. 2. In terms of spherical polar coordinates. θ. We can now use the generating function to write this as a power series. |r − r0 | 1 x2 + y2 + (z − a)2 If this unit charge lies on the z-axis. this becomes V = . This means that x2 + y 2 + (z − a)2 = x2 + y 2 + z 2 − 2az + a2 = r2 + a2 − 2az.2. at x = y = 0.4 RODRIGUES’ FORMULA 39 Combining these gives ∞ ∞ nt Pn (x) = (x − t) n n=0 n=0 tn Pn (x). P1 (x) = 1 21 1! d1 {(x2 − 1)} = x. and by equating coeﬃcients of tn we obtain the recurrence relation nPn (x) = xPn (x) − Pn−1 (x). φ. a enclose the point ξ = z and be traversed in an anticlockwise sense. since Pn (1) = 1 2π 2π 1dθ = 1. centred on ξ = z with radius | z 2 − 1|. 0 .6) which is known as Schl¨ﬂi’s representation.40 LEGENDRE FUNCTIONS whilst for n = 2. it is convenient to switch from the real variable x to the complex variable z = x + iy. Rodrigues’ formula can also be used to develop an integral representation of the Legendre polynomials. which we leave as an exercise. We deﬁne the ﬁnite complex Legendre polynomials in the same way as for a real variable. since | cos φ + i sin φ cos θ| = we have |Pn (cos φ)| 1. we now choose C to√ a circle. {(x2 − 1)2 } = dx2 4 · 2 dx 2 The general proof of this result is by induction on n. In order to show how this is done. we can write z = cos φ and use Laplace’s representation to show that |Pn (cos φ)| Now. after some simple manipulation. f (n) (z) = for n n! 2πi C f (ξ) dξ. 1 2π 2π | cos φ + i sin φ cos θ|n dθ. when z is an interior point of C. Pn (z) = 1 2π 2π z+ θ=0 z 2 − 1 cos θ n dθ. The contour C must. we have Pn (z) = 1 2n+1 πi C (2. Now. 1]. Putting ξ = z + z 2 − 1eiθ gives. P2 (x) = 1 22 2! d2 1 d 1 {4x(x2 − 1)} = (3x2 − 1). 0 Laplace’s representation is useful. To simplify √ matters. (ξ − z)n+1 (ξ 2 − 1)n dξ. Pn (z) = 1 dn 2 z −1 2n n! dz n n . amongst other things. will be useful to us here. For z ∈ [−1. Recall from complex variable theory (see Appendix 6) that. In particular Rodrigues’ formula. if f (z) is analytic and single-valued inside and on a simple closed curve C. using Rodrigues’ formula. cos2 φ + sin2 φ cos2 θ cos2 φ + sin2 φ = 1. This is known as Laplace’s representation. (ξ − z)n+1 0. of course. be with z = 1. In fact it is also valid when z = 1. for providing a bound on the size of the Legendre polynomials of real argument. To show this. . for which 1 −1 Pn (x)Pn (x)dx = (−1)n 2n n! 1 −1 (−1)n 2n n! 1 −1 (n) (x2 − 1)n Pn (x)dx = Noting the fact that (x2 − 1)n 1 d 2n {(x2 − 1)n }dx. This result is interesting in its own right. but for the time being consider the case f (x) = Pm (x) with m < n.7) where the Kronecker delta is deﬁned by δmn = 1 0 for m = n.5 ORTHOGONALITY OF THE LEGENDRE POLYNOMIALS 41 2. 2n n! dx2n (x2 − 1)n = x2n + · · · + (−1)n . this must also hold for m > n. By symmetry. (2.2.5 Orthogonality of the Legendre Polynomials Legendre polynomials have the very important property of orthogonality on [−1.8) where f (n) (x) is the nth derivative of f (x). that is 1 −1 Pn (x)Pm (x)dx = 2 δmn . Let’s now consider the case n = m. 2n + 1 (2. Rodrigues’ formula shows that 1 −1 f (x)Pn (x)dx = 1 2n n! 1 1 f (x) −1 1 dn {(x2 − 1)n }dx dxn d n−1 {(x2 − 1)n }dx dxn−1 = 1 n n! 2 f (x) d n−1 {(x2 − 1)n } dxn−1 =− 1 2n n! 1 −1 − f (x) −1 f (x) −1 d n−1 {(x2 − 1)n }dx. for m = n. 1]. 1]. note that if f ∈ C[−1. so that 1 −1 Pm (x)Pn (x)dx = (−1)n 2n n! 1 −1 (n) (x2 − 1)n Pm (x)dx. dxn−1 Repeating this integration by parts (n − 1) more times gives 1 −1 f (x)Pn (x)dx = (−1)n 2n n! 1 −1 (x2 − 1)n f (n) (x)dx. Since the nth derivative of an mth order polynomial is zero for m < n. we have 1 −1 Pm (x)Pn (x)dx = 0 for m < n. † What we can present here is the procedure for obtaining the coeﬃcients of such an expansion for a given function f (x) belonging to C[−1. We have now shown that the Legendre polynomials are orthogonal on [−1. To evaluate the remaining integral. we note that. Multiplying by Pm (x) and integrating over [−1. 3 · 2 · 1 = (2n)!. Kuller. It is also possible to show that these polynomials are complete in the function space C[−1. and will be omitted here. dx2n we can see that 1 −1 Pn (x)Pn (x)dx = (−1)n 2n n! 1 1 −1 (x2 − 1)n 1 (2n)!dx 2n n! = (−1)n (2n)! 22n (n!)2 −1 (x2 − 1)n dx. forming the inner product with Pm (x)) gives 1 −1 1 ∞ f (x)Pm (x)dx = −1 Pm (x) n=0 an Pn (x)dx. which we will discuss in Chapter 4. we use a reduction formula to show that 1 −1 (x2 − 1)n dx = 22n+2 n!(n + 1)!(−1)n . according to a fundamental result due to Weierstrass (the Weierstrass polynomial approximation theorem) any function which is continuous on some interval can be approximated as closely as we wish by a polynomial. 1]. This means that a continuous function can be expanded as a linear combination of the Legendre polynomials. 2n + 1 and hence 1 −1 2 Pn (x)dx = This completes the derivation of (2. . The completeness then follows from the application of this theorem. 1].7). The proof of the completeness property is rather diﬃcult. The treatment of completeness of other solutions of Sturm–Liouville problems may be more complicated. † Just to give a ﬂavour of the completeness proof for the case of Legendre polynomials. To do this we write ∞ f (x) = a0 P0 (x) + a1 P1 (x) + · · · + an Pn (x) + · · · = n=0 an Pn (x). 1] (more precisely. However. There is an easier proof of the ﬁrst part of this. We will return to this topic in Chapter 4. . using the idea of a self-adjoint linear operator. we can deduce that any polynomial can be written as a linear combination of Legendre polynomials. for example for Bessel functions. Ostberg and Perkins (1966). because of their polynomial form. (2n + 2)! 2 . .42 LEGENDRE FUNCTIONS and hence that d 2n 2 (x − 1)n = 2n(2n − 1) . A complete proof of this can be found in Kreider. 1]. . Let’s consider a couple of examples. 1 2 P0 (x) + P2 (x). .5 ORTHOGONALITY OF THE LEGENDRE POLYNOMIALS 43 Interchanging the order of summation and integration leads to 1 −1 ∞ 1 f (x)Pm (x)dx = n=0 ∞ an −1 Pn (x)Pm (x)dx = n=0 an 2 2am δmn = . (2. . (2.8) shows that and therefore. . This is called a Fourier–Legendre series. am = so that a0 = 1 2 1 −1 2m + 1 2 1 −1 x2 Pm (x)dx. 23 3 1 a1 = (2 · 1) + 1 2 1 x2 · xdx = 3 x4 2 4 = 0.7). consider f (x) = ex . = −1 2 . −1 1 a2 = (2 · 2) + 1 2 x3 1 5 3x5 − x2 (3x2 − 1)dx = 2 2 2·5 3·2 −1 am = 0 for m = 3. when f (x) = x2 .2. (2.9) 1 −1 f (x) = n=0 f (x)Pn (x)dx Pn (x). 3 3 A ﬁnite polynomial clearly has a ﬁnite Fourier–Legendre series. ex dx = −1 1 e − e−1 . In this case x2 = am = and hence a0 = 1 2 1 2m + 1 2 1 −1 ex Pm (x) dx. Firstly. This means that am = and we write the series as ∞ 2m + 1 2 2n + 1 2 1 −1 f (x)Pm (x)dx. 3 Also. x2 · 1dx = 1 −1 1 x3 2 3 1 = −1 1 12 = . 4. 2m + 1 2m + 1 using the orthogonality property. Secondly. 2 . y. Qy . . y. 1) where f has a right and left derivative. . t) − Qx (x + δx. is related to the temperature gradient. . are easily computed. z) with sides of length δx. The thermal conductivity. we will encounter the Legendre equation in situations where we have to solve partial diﬀerential equations containing the Laplacian in spherical polar coordinates. t) − Qy (x. ∇T . of a body appears in Fourier’s law. a5 . . speciﬁc heat capacity c and thermal conductivity k. by the simple linear relationship Q = −k∇T . k. y. Q = (Qx . δy and δz. which gives an = (2n + 1) an−2 − an−1 2n − 3 for n = 2. This is best done by using Rodrigues’ formula.10) from which the values of a4 . Instead we content ourselves with a statement that the Fourier–Legendre series converges uniformly on any closed subinterval of (−1. is the amount of heat required to raise the temperature of a unit mass of a substance by one degree. We will not examine the convergence of Fourier–Legendre series here as the details are rather technical. (2. Recall that the speciﬁc heat capacity. y. z. y. density ρ. t + δt) − T (x. . the temperature change in this element over a time interval δt is determined by the diﬀerence between the amount of heat that ﬂows into the element and the amount of heat that ﬂows out. t)} δxδyδz = {Qx (x. which states that the heat ﬂux per unit area. z + δz. (2. y. z. In general. An extension of this result to the space of piecewise continuous functions is that the series converges to the value + − 1 2 f (x0 ) + f (x0 ) at each point x0 ∈ (−1. Qz ). 1) in which f is continuous and diﬀerentiable. If we now consider a small element of our solid body at (x.11) .6. z. y + δy. 2. y. 2.1 Heat Conduction Let’s derive the equation that governs the evolution of an initial distribution of heat in a solid body with temperature T . per unit time. We will prove a related theorem in Chapter 5. which gives ρc {T (x.6 Physical Applications of the Legendre Polynomials In this section we present some examples of Legendre polynomials as they arise in mathematical models of heat conduction and ﬂuid ﬂow in spherical geometries. z. To proceed with this calculation it is necessary to ﬁnd a recurrence relation between the an . . t)} δtδyδz + {Qy (x. 3. y. z. t)} δtδxδz + {Qz (x.44 LEGENDRE FUNCTIONS a1 = 3 2 1 −1 x ex dx = 3 e−1 . t) − Qz (x. t)} δtδxδy. c. z. . z. when the surface temperature is held at f (θ) = T0 sin4 θ in spherical polar coordinates. neither does the solution. θ.1 contains the values of relevant properties for three everyday materials. y. we obtain ∂Qx ∂Qy ∂Qz ∂T =− + + ρc = −∇ · Q. gives the diﬀusion equation. t + δt). ∂t ∂x ∂y ∂z Substituting in Fourier’s law. φ). Some typical physical properties of copper. (2. is the amount of heat crossing the x-orientated faces of the element. uniform sphere of unit radius. y.13). for example.1 × 10−5 3. δx.1. Example Let’s try to ﬁnd the steady state temperature distribution in a solid. Taking the limit δt. θ) = R(r)Θ(θ). z. This gives Θ d dr r2 dR dr + R d sin θ dθ sin θ dΘ dθ = 0. this equation takes the simple form ∇2 T = 0. and hence Laplace’s equation takes the form ∂T ∂ 1 ∂T 1 ∂ r2 + 2 sin θ = 0. t) − Qx (x + δx.13) which is known as Laplace’s equation. Q = −k∇T . This temperature distribution will satisfy Laplace’s equation. during the time interval (t. water (at room temperature and pressure) and glass. each with area δyδz. Note that a typical term on the right hand side of this. {Qx (x. . It must be solved in conjunction with appropriate boundary conditions.2. T (r. (2.1 × 10−4 6. φ. δy. δz → 0. which drive the temperature gradients in the body.4 × 10−7 Table 2. (r. Since the equation and boundary conditions do not depend upon the azimuthal angle. t)} δtδyδz.12) ∂t where K = k/ρc is called the thermal diﬀusivity. z.8 c J kg−1 K−1 386 4186 840 K m2 s−1 1. ∂T = K∇2 T. (2.6 PHYSICAL APPLICATIONS OF THE LEGENDRE POLYNOMIALS 45 ρ kg m−3 copper water glass 8920 1000 2800 k J m−1 s−1 K−1 385 254 0. When the temperature reaches a steady state (∂T /∂t = 0). r2 ∂r ∂r r sin θ ∂θ ∂θ Let’s look for a separable solution. Table 2. and using sin2 θ = 1 − µ2 .46 LEGENDRE FUNCTIONS and hence 1 d R dr r2 dR dr =− d 1 Θ sin θ dθ sin θ dΘ dθ . the temperature must . to within an arbitrary multiplicative constant. This is d dr r2 dR dr = n(n + 1)R. We have introduced the subscript n for Θ so that we can specify that this is the solution corresponding to a particular choice of n. and hence Θn (θ) = Pn (cos θ). R) → (λr. We must now consider the solution of the equation for R(r). is the Legendre polynomial. dθ d −1 d d = = . By inspection. with 1 d R dr r2 dR dr =− d 1 Θ sin θ dθ sin θ dΘ dθ = constant = n(n + 1). namely 1 d sin θ dθ Changing variables to µ = cos θ. sin θ dΘ dθ + n(n + 1)Θ = 0. equivalently (1 − µ2 )y − 2µy + n(n + 1)y = 0. At the centre of the sphere. Since a physically meaningful solution must in this case be ﬁnite at µ = ±1 (the north and south poles of the sphere). This choice of constant may seem rather ad hoc at this point. dµ dµ or. leads to d dy (1 − µ2 ) + n(n + 1)y = 0. the solution. dµ dµ dθ sin θ dθ y(µ) = Θ(θ). by the method of Frobenius or by noting that the equation is invariant under the transformation group (r. we can ﬁnd the solution of this equation as Rn (r) = An rn + Bn r−1−n . we have found the solution of Legendre’s equation that is bounded at µ = ±1. We now have an ordinary diﬀerential equation for Θ. Speciﬁcally. where again the subscript n denotes the dependence of the solution on n and An and Bn are arbitrary constants. y(µ) = Pn (µ). this equality can only be valid if both sides are equal to some constant. but all will become clear later. λR) (see Chapter 10). Since the left hand side is a function of r only and the right hand side is a function of θ only. Note that we have just solved our ﬁrst boundary value problem. which is Legendre’s equation. .2. It is best to work in terms of the variable µ = cos θ.9). This can be done using the speciﬁed temperature on the surface of the sphere. 6. Note that the temperature at the centre of the sphere is 8T0 /15. we expect to obtain a ﬁnite Fourier–Legendre series. [0 1 0 pi])   .6 PHYSICAL APPLICATIONS OF THE LEGENDRE POLYNOMIALS 47 be ﬁnite. A4 = · T0 . The solution is therefore T = T0 16 8 8 P0 (cos θ) − r2 P2 (cos θ) + r4 P4 (cos θ) . A straightforward calculation of the integral gives us A0 = 1 16 5 32 9 16 · T0 . . A2 = − · T0 . the general solution for the temperature will be a linear combination of these solutions. which is a polar plot in a plane of constant φ. 2 15 2 105 2 315 Am = 0 for m = 5. so we set Bn = 0. Since the function that we want to expand is a ﬁnite polynomial in µ. We substitute r = 1 into the general expression for the temperature to get ∞ f (θ) = T0 sin4 θ = n=0 An Pn (cos θ).3 using the MATLAB script   ezmesh(’r*cos(t)’.’r*sin(t)’. 15 21 35 (2. From (2. ∞ T = n=0 An rn Pn (cos θ).3. The remaining task is to evaluate the coeﬃcients An .. We then have ∞ T0 (1 − µ2 )2 = n=0 An Pn (µ).. An = 2n + 1 2 1 −1 T0 (1 − µ2 )2 Pn (µ)dµ. As this is a solution for arbitrary n. . T = T0 sin4 θ at r = 1. 8*r^4*(35*cos(t)^4/8-15*cos(t)^2/4+3/8)/35’.’8/15-8*r^2*(3*cos(t)^2-1)/21+. Our solution of Laplace’s equation therefore takes the form T = An rn Pn (cos θ). A3 = 0. We produced Figure 2. .14) This solution when T0 = 1 is shown in Figure 2. The An will therefore be the coeﬃcients in the Fourier–Legendre expansion of the function f (θ) = T0 sin4 θ. A1 = 0. 2.3.14). Mathematically. S where u is the velocity of the ﬂuid. so any ﬂux out of V at one place is balanced by a ﬂux into V at another place. is smooth. given by (2. whilst the fourth speciﬁes the ranges of r and t. Fig. u. Physically. it is possible to make the further assumption that the ﬂow is irrotational (∇ × u = 0). since the ﬂuid is incompressible.48 LEGENDRE FUNCTIONS The function ezmesh gives an easy way of plotting parametric surfaces like this. If the ﬂuid velocity ﬁeld. we can express this as u · n dS = 0. we must have ∇ · u = 0 throughout the ﬂow. V As this applies to an arbitrary volume V within the ﬂuid. and hence u·n is the normal component of the ﬂuid velocity out through S. Although ﬂuid ﬂows into and out of V . the mass of ﬂuid within V remains constant.6. y and z coordinates as parametric functions of r and t = θ. we can use the divergence theorem to rewrite this statement of conservation of mass as ∇ · u dV = 0. If we also suppose that the ﬂuid is inviscid (there is no friction as one ﬂuid element ﬂows past another).2 Fluid Flow Consider a ﬁxed volume V bounded by a surface S within an incompressible ﬂuid. this means that . The ﬁrst three arguments give the x. 2. n is the outward unit normal to S. The steady state temperature in a uniform sphere with surface temperature sin4 θ. † For simply connected ﬂow domains. assuming that the ﬂow is steady.2. axisymmetric solution of Laplace’s equation is ∞ φ= n=0 An rn + Bn r−1−n Pn (cos θ).6 PHYSICAL APPLICATIONS OF THE LEGENDRE POLYNOMIALS 49 there is no spin in any ﬂuid element as it ﬂows around. We can then use the MATLAB script x = linspace(-2.2.4. First of all. ∂r 1 3 2a On substituting (2.^(3/2)). We know from the previous example that the bounded.2) the streamlines are given by ψ=U r− a3 r2 sin θ = U y 1 − a3 (x2 + y 2 ) 3/2 = constant. known as the velocity potential. To ﬁx the constants Bn . we note that (see Section 7. there can be no ﬂux of ﬂuid into a solid body so that u · n = n · ∇φ = ∂φ/∂n = 0 where the ﬂuid is in contact with a solid body.2. with u = U i where i is the unit vector in the x-direction.^2). Z = Y. contour(X.16) The streamlines (continuous curves that are tangent to the velocity vector) are shown in Figure 2. irrotational.^2+Y. v = linspace(0. As for boundary conditions.400). notice that there can be no ﬂow through the surface of the sphere. (2.Y. incompressible ﬂows are therefore governed by the two equations ∇ · u = 0 and ∇×u = 0. y = linspace(0.15) The ﬂow at inﬁnity has potential φ = U x = U r cos θ./(X.*(1-1. [X Y] = meshgrid(x.Z.4 when a = U = 1. The solution is therefore φ=U r + a3 2 r2 cos θ. let’s consider what happens when a sphere of radius r = a is placed in a uniform stream. the velocity potential satisﬁes Laplace’s equation. In order to obtain this ﬁgure.2. we see that we must take A1 = 1 and An = 0 for n > 1.200). As an example of such a ﬂow. inviscid and irrotational.15) into this boundary condition. This gives us the boundary condition on the radial velocity as ur = ∂φ = 0 at r = a. In other words.15). we ﬁnd that B1 = Bn = 0 for n > 1. Since P1 (cos θ) = cos θ.v) † We will consider an example of viscous ﬂow in Section 6. The ﬂow at inﬁnity must be uniform. Inviscid. Substituting into ∇ · u = 0 gives ∇2 φ = 0. ∇×u = 0 if and only if u = ∇φ for some scalar potential function φ. and (2. .2. it is clear that the problem will be best treated in spherical polar coordinates.y). p. We would now like to apply Newton’s second law to the motion of the ﬂuid within V . t) · ∇) u(x. t) + · · · ∂t dt = δt ∂u (x. We can now use Newton’s second law on the ﬂuid within S to obtain −pn dS = S V ρ Du dV.50 LEGENDRE FUNCTIONS The command meshgrid creates a grid suitable for use with the plotting command contour out of the two vectors. Fig. This means that the ﬂuid acceleration is ∂u Du = + (u · ∇) u. t) + · · · . In order to complete this example. Dt After noting that pn dS = i S S (pi) · n dS + j S (pj) · n dS + k S (pk) · n dS. .4. t) + (u(x. The force exerted by the ﬂuid on a surface S by the ﬂuid outside S is purely a pressure. we need an expression for the acceleration of the ﬂuid. ∂t since u = dx/dt. or time derivative following the ﬂuid particle. In other words. we can Taylor expand to give u(x(t + δt). 2. for small δt. we must consider the pressure in our ideal ﬂuid. t + δt) − u(x(t). In order to do this. The streamlines for inviscid. t) + · ∇u(x. which. t) = δt dx ∂u (x. the volume enclosed by S. x and y. The vector v speciﬁes the values of ψ for which a contour is to be plotted. Let’s consider the change in the velocity of a ﬂuid particle between times t and t + δt. the force on a surface element with area dS and outward unit normal n is −pn dS. irrotational ﬂow about a unit sphere. which acts normally to S. Dt ∂t where D/Dt is the usual notation for the material derivative. 18) p 1 2 + |∇φ| ρ 2 = 0.6 PHYSICAL APPLICATIONS OF THE LEGENDRE POLYNOMIALS 51 where i. φ is given by (2. applied to each of these integrals. This expression simpliﬁes on the surface of the sphere to give 1 9 p|r=a = p∞ + ρU 2 1 − sin2 θ . the stream-facing poles of the sphere. Since V is arbitrary. 2 for irrotational ﬂow (u · ∇) u = ∇ 1 u · u and we can write (2. Its implication for inviscid. Dt ρ and for steady ﬂows 1 (u. φ.2. 1 2 |∇φ| 2 1 = − ∇p. with p = p∞ + 1 ρU 2 .∇) u = − ∇p.16). k are unit vectors in the coordinate directions. ρ 2 (2. the divergence theorem. so that |∇φ| = φ2 + r 2 1 2 φ = U2 r2 θ 1− a3 r3 2 cos2 θ + 1 r2 r+ a3 2r2 2 sin2 θ .17) as 2 ∇ and hence ∇ which we can integrate to give p 1 2 + |∇φ| = C. 2 4 This shows that the pressure is highest at θ = 0 and π. and Bernoulli’s equation gives 1 p = p∞ + ρU 2 1 − 2 a3 1− 3 r 2 1 cos θ + 2 r 2 a3 r+ 2 2r 2 sin2 θ . Since (u · ∇) u ≡ ∇ 1 u · u − u × (∇ × u). where we have written p∞ for the pressure at inﬁnity. j. V and hence that ρ V Du + ∇p Dt dV = 0. irrotational ﬂow is that the pressure can be calculated once we know the velocity potential. In our example. u = ∇φ. and drops below p∞ over a portion of the rest of 2 . shows that pn dS = i S V ∂p dV + j ∂x V ∂p dV + k ∂y V ∂p dV = ∂z ∇p dV.17) We earlier used the irrotational nature of the ﬂow to write the velocity ﬁeld as the gradient of a scalar potential. ρ (2. Du 1 = − ∇p. ρ which is known as Bernoulli’s equation. corresponding to separable axisymmetric solutions of Laplace’s equation.52 LEGENDRE FUNCTIONS the boundary.20) where we have written Z (m) = d m Z/dxm .20) and (2. dx and therefore that y = (1 − x2 )m/2 A dm dm Pn (x) + B m Qn (x) . p = p∞ − 5 ρU 2 . on the equator. Y is the solution of (1 − x2 )Y − 2(m + 1)xY + (n − m)(n + m + 1)Y = 0. with the lowest pressure. (1 − x2 ) m2 d 2y dy + n(n + 1) − − 2x dx2 dx 1 − x2 y = 0. 2.21) (2. we ﬁnd that Φ = eimφ with m an integer.7 The Associated Legendre Equation We can also look for a separable solution of Laplace’s equation in spherical polar coordinates (r. It trivially reduces to Legendre’s equation when m = 0.19) This equation is known as the associated Legendre equation. n where dm dm Pn (x). If we write Legendre’s equation in the form (1 − x2 )Z − 2xZ + n(n + 1)Z = 0. from what we know about the Legendre polynomials. n dxm dx m m The functions Pn (x) and Qn (x) are called the associated Legendre functions. we get (1 − x2 )[Z (m) ] − 2(m + 1)x[Z (m) ] + (n − m)(n + m + 1)Z (m) = 0. The equation for y is. Pn (x) = 0 if m > n m Pn (x) = (1 − x2 )m/2 . If we deﬁne Y = (1 − x2 )−m/2 y. and diﬀerentiate m times. A comparison of (2. φ).19) are m y = APn (x) + BQm (x). If we seek a solution of the form y(θ)Φ(φ)R(r). θ. Qm (x) = (1 − x2 )m/2 m Qn (x). m Clearly. which has solution Z = APn (x) + BQn (x).21) shows that dm Y = m [APn (x) + BQn (x)] . after using the change of variable x = cos θ. dxm dx We have now shown that the solutions of the diﬀerential equation (2. the connection is more profound than this. 8 The implications of this and the modelling assumptions made are discussed more fully by Acheson (1990). that is not necessarily axisymmetric. (2. However. (2. θ = π/2. sin2 θ ∂φ2 Separable solutions take the form Sn = (Am cos mφ + Bm sin mφ)F (θ). By homogeneous we mean of the form xi y j z k . Substituting this expression for u in Laplace’s equation. like u = rn Sn (θ. µ = cos θ. z) will look.7 THE ASSOCIATED LEGENDRE EQUATION 53 and Qm (x) is singular at x = ±1. with m an integer. The function F (θ) satisﬁes 1 d sin θ dθ sin θ dF dθ + n(n + 1) − m2 sin2 θ F = 0. Example: Spherical harmonics Let’s try to ﬁnd a representation of the solutions of Laplace’s equation in three dimensions. which can be derived in much the same way as those for the ordinary Legendre functions. There are various recurrence formulae and orthogonality relations between the associated Legendre functions. . 2 1 P3 (x) = 3 3 (5x2 − 1)(1 − x2 )1/2 = (sin θ + 5 sin 3θ).2. 2 8 where x = cos θ. we ﬁnd that 1 ∂ sin θ ∂θ sin θ ∂Sn ∂θ + 1 ∂ 2 Sn + n(n + 1)Sn = 0. It is straightforward to show from their deﬁnitions n that 1 1 P1 (x) = (1 − x2 )1/2 = sin θ. If we now make the usual transformation. sin2 θ ∂φ2 A homogeneous solution of order n = i + j + k in (x. It is simplest to work in spherical polar coordinates. in terms of which Laplace’s equation takes the form ∂ ∂r r2 ∂u ∂r + 1 ∂ sin θ ∂θ sin θ ∂u ∂θ + 1 ∂2u = 0. 2 2 P2 (x) = 3(1 − x2 ) = 3 sin2 θ = 3 (1 − cos 2θ). y and z. ∂2u ∂2u ∂2u + 2 + 2 = 0. φ). ∂x2 ∂y ∂z that are homogeneous in x. in the new coordinates. we get (1 − µ2 )y − 2µy + n(n + 1) − m2 1 − µ2 y = 0. F (θ) = y(µ). P2 (x) = 3x(1 − x2 )1/2 = 3 sin θ cos θ = 3 sin 2θ. y. 2. . (1 − x2 )Pn (x) = nPn−1 (x) − nxPn (x). 1. = 2 cos5 θ.m Pn (cos θ) + Dn. . (1 − x2 )y − 2xy + 6y = 0. by the method of Frobenius.54 LEGENDRE FUNCTIONS which is the associated Legendre equation. 5 (7x3 − 3x)(1 − x2 )1/2 . . 1. . . .2 2. n m The quantities rn cos mφ Pn (cos θ). .4 Pn+1 (x) − Pn−1 (x) = (2n + 1)Pn (x). 4n2 − 1 (b) −1 1 Pn (x)Pn+1 (x)dx = 2 for n = 0. Solve the axisymmetric boundary value problem for Laplace’s equation. xPn (x)Pn (x)dx = 2n for n = 0. . . What is the simplest solution of this equation? By using this simple solution and the reduction of order method ﬁnd a closed form expression for Q1 (x).5 What value will this series have at x = 0? Establish the results 1 (a) −1 1 xPn (x)Pn−1 (x)dx = 2n for n = 1.m Qm (cos θ)}. Our solutions can therefore be written in the form m u = rn (Am cos mφ + Bm sin mφ){Cn. . 2.6 2. Exercises 2.3. . 2n + 1 (c) −1 2.3 Find the ﬁrst four nonzero terms in the Fourier–Legendre expansion of the function f (x) = 0 for −1 < x < 0. 1.8 Show that 2 (a) P3 (x) = 1 (b) P4 (x) = 15x(1 − x2 ). . . 2 . We will see how they arise in a quantum mechanical description of the hydrogen atom in Section 4. 0 < θ < π. Prove that (a) (b) 2. which appear as typical terms in the solution.7 Determine the Wronskian of Pn and Qn for n = 0. are called spherical harmonics and appear widely in the solution of linear boundary value problems in spherical geometry. 2. . 2. 2. 1 for 0 < x < 1. Use the generating function to evaluate (a) Pn (1).6. θ) = 0 for 0 < r < a. . ∇2 T T (a.1 Solve Legendre’s equation of order two. (b) Pn (0). . incompressible. (b) During contraction. where s = cos θ with θ the usual spherical polar coordinate. . 1). and symmetric with respect to the aortal axis. Show that having (R sin α)2 = R2 (1 − a2 ) constant gives ∂φ = f (s) = ∂n Us ˙ R(1 − s/a) for a < s < 1.5. irrotational. some of the blood ﬁlling the ventricle cavity is ejected though the opening with mean speed U = U (t) into the attached cylindrical aorta.14 (a) Suppose Pn (x0 ) = 0 for some x0 ∈ (−1. prove the orthogonality property 1 −1 m Plm (x)Pn (x)dx = 0 for l = n. i is the unit vector in the aortic ﬂow direction and a = cos α(t). ∗ Find the solution of the Dirichlet problem. m Pn (x) = 0.11 2. As a result. (a) State a partial diﬀerential equation appropriate to the ﬂuid ﬂow for this situation. This occurs suﬃciently rapidly that we can assume that the ﬂow is inviscid.9 2. θ. φ) = sin2 θ cos 2φ. (b) Show that Pn with n 1 has n distinct zeros in (−1.10 2.12 ∗ Prove that |Pn (x)| < n2 and |Pn (x)| < n4 for −1 < x < 1. Project A simpliﬁed model for the left ventricle of a human heart is provided by a sphere of time-dependent radius R = R(t) with a circular aortic opening of constant area A. Evaluate 1 −1 m [Pm (x)] dx. ∗ The self-adjoint form of the associated Legendre equation is d m2 m (1 − x2 )Pn (x) + n(n + 1) − dx 1 − x2 Using this directly. where n is the outward unit normal at time t. 2 2.10). ∇2 Φ = 0 in r > 2 subject to Φ → 0 as r → ∞ and Φ(2. as shown in Figure 2. Show that x0 is a simple zero. Derive Equation (2.13 2.EXERCISES 55 2. 1). During contraction we suppose that the opening remains ﬁxed whilst the centre of the sphere moves directly toward the centre of the opening and the radius R(t) decreases accordingly. for −1 < s < a. show that a point on the surface of the ventricle has velocity ˙ ˙ Rn − (Ra)i. which relates the geometry to the mean aortal speed. 2. where S is the surface of V . for a solution to exist.56 LEGENDRE FUNCTIONS i Aorta U (t ) α (t ) R (t ) V entricle Fig. Deduce that ˙ R = [a(a − 1)/(a + 1)]U . The total momentum in the direction of i is the blooddensity times the integral I= V 1 ∇φ · idV = S (sφ)|r=R dS. A simple model for the left ventricle of the human heart. (c) Show that. (d) Let V = V (t) denote both the interior of the sphere at time t and its volume. −1 f (s)ds = 0.5. Hence show that ∞ 1 I = 2πR2 n=0 cn −1 sPn (s)ds = 4 2 3 πR c1 = V 3 2 1 sf (s) ds. −1 . .EXERCISES 57 (e) Use the answer to part (c) to show that 4I = (1−a)(a2 +4+3a)V U . (f) Explain how this model could be made more realistic in terms of the ﬂuid mechanics and the physiology. You may like to refer to Pedley (1980) for some ideas on this. 1) has a regular singular point at x = 0. who ﬁrst used them to analyze planetary orbits.CHAPTER THREE Bessel Functions In this chapter.1 The Gamma Function and the Pockhammer Symbol Before we use the method of Frobenius to construct the solutions of Bessel’s equation. These are named after the German mathematician and astronomer Friedrich Bessel. By deﬁnition. +x 2 dx dx (3. it will be useful for us to make a couple of deﬁnitions. The gamma function is deﬁned by ∞ Γ(x) = 0 e−q q x−1 dq.1) with ν real and positive. ∞ Γ(1) = 0 e−q dq = 1. P x P x2 both of which are polynomials and have Taylor expansions with inﬁnite radii of convergence. as we shall discuss later. Bessel’s equation can be written in the form x2 d 2y dy + x2 − ν 2 y = 0. x2 x2 − ν 2 xQ x2 x2 R = 2 = 1. usually in a cylindrical geometry. and we will discuss some examples of these at the end of this chapter. = = x2 − ν 2 . . Any series solution will therefore also have an inﬁnite radius of convergence. 3. Bessel functions occur in many other physical problems. Note that (3.2) Note that the integration is over the dummy variable q and x is treated as constant during the integration. We will start by considering the function evaluated at x = 1. Using the notation of Chapter 1. for x > 0. We also note that ∞ Γ(x + 1) = 0 e−q q x dq. (3. we will discuss a class of functions known as Bessel functions. 3) (3. We therefore have the useful result. We will often need to know Γ (1/2). . Q = Q. Q)-plane. and hence Γ 1 2 2 π/2 ∞ r=0 =4 θ=0 e−r r dr dθ. Suppose that x = n is a positive integer. consider the deﬁnition. ˜2 Since the limits are independent.3. we can combine the integrals as Γ 1 2 2 ∞ ∞ 0 =4 0 e−(Q 2 ˜ +Q2 ) ˜ dQ dQ. (3. 2q √ If we introduce the new variable Q = q. Firstly.4) = 0 e−q q −1/2 dq. Then Γ(n + 1) = nΓ(n) = n(n − 1)Γ(n − 1) = · · · = n(n − 1) . where ˜ Q = r cos θ and Q = r sin θ. . 2 ˜ The limits of integration give us the positive quadrant of the (Q. we have Γ 1 2 ∞ 1 2 2 1 = 2π − e−r 2 ∞ = π. 2 · 1 = n!. as required. 2 and integrating with respect to r gives Γ Finally. We therefore have the recursion formula Γ(x + 1) = xΓ(x). to obtain Γ 1 2 = 2 0 e−Q dQ 2 2 0 ˜ e−Q dQ . Γ 1 2 ∞ (3.1 THE GAMMA FUNCTION AND THE POCKHAMMER SYMBOL 59 which can be integrated by parts to give Γ(x + 1) = −q x e−q ∞ 0 ∞ + 0 e−q xq x−1 dq = x 0 ∞ e−q q x−1 dq = xΓ(x). =2 Γ 2 0 2 ∞ ∞ this integral ˜ We can also write this integral in terms of another new variable. Γ(n + 1) = n! for n a positive integer. 1 −1/2 dq.5) . so that dQ = becomes ∞ 2 1 e−Q dQ. 0 =2 0 e−Q dQ = 2 √ π. Performing the integration over θ we have Γ 1 2 2 2 ∞ = 2π 0 re−r dr. ˜ If we now change to standard polar coordinates we have dQ dQ = r dr dθ. y(x) = n=0 an xn+c . so that. Alternatively. For example. A relationship between the gamma function and the Pockhammer symbol that we will need later is Γ(x) (x)n = Γ(x + n) (3. . this yields ∞ ∞ x2 n=0 an (n + c)(n + c − 1)xn+c−2 + x n=0 an (n + c)xn+c−1 . and. which is not integrable. To derive this.60 BESSEL FUNCTIONS We can use Γ(x) = Γ(x + 1)/x to deﬁne Γ(x) for negative values of x. We can repeat this to give Γ(x) (x)n = Γ(x + n − 1) (x + n − 1) = Γ(x + n). Using the recursion relation (3. . We also choose to deﬁne (α)0 = 1. Now Γ(x) (x)n = Γ(x) {x(x + 1)(x + 2) . Note that the gamma function is available in MATLAB as the function gamma. (α)r is a product of r terms.6) for x real and n a positive integer.2). (x + n − 1)} . the integrand diverges like 1/q as q → 0. . It is deﬁned as (α)r = α(α + 1)(α + 2) . .3). The Pockhammer symbol is a simple way of writing down long products. Γ − 1 2 = Γ −1 + 1 2 = −2Γ −1 2 1 2 √ = −2 π. Γ(x) = Γ(x+1)/x shows that Γ(x) ∼ 1/x as x → 0. From the deﬁnition. . Note that (1)n = n!. We also ﬁnd that Γ(x) is singular at x = 0. Γ(x) (x)n = Γ(x + 1) {(x + 1)(x + 2) . (x + n − 1)} = {Γ(x)x} {(x + 1)(x + 2) . we start with the deﬁnition of the Pockhammer symbol. (α + r − 1). (x)n = x(x + 1)(x + 2) . in general. .1). 3.2 Series Solutions of Bessel’s Equation ∞ We can now proceed to consider a Frobenius solution. (α)1 = α and (α)2 = α(α + 1). . (x + n − 1). for example. . . When substituted into (3. (x + n − 1)} . . (3. The recurrence relation that we obtain from the general term of (3. 2(2 ± 2ν) Substituting for a2 in terms of a0 gives a0 a0 = 2 . and hence an = 0 for n odd.7). 2ν. which we studied in the previous chapter. we can consider both roots of the indicial equation at once.3. From the second term of (3. 4(4 ± 2ν) a0 . is not an integer. We can now rearrange this equation and combine corresponding terms to obtain ∞ ∞ an {(n + c)2 − ν 2 }xn+c + n=0 n=0 an xn+c+2 = 0. so that c = ±ν. . we have a1 {(1 ± ν)2 − ν 2 } = a1 (1 ± 2ν) = 0. which yields a2 = − and now with n = 4. (3.2 SERIES SOLUTIONS OF BESSEL’S EQUATION ∞ 61 +(x2 − ν 2 ) n=0 an xn+c = 0. . We can extract two terms from the ﬁrst summation and then shift the second one to obtain a0 (c2 − ν 2 )xc + a1 {(1 + c)2 − ν 2 }xc+1 ∞ + n=2 [an {(n + c)2 − ν 2 } + an−2 ]xn+c = 0. .7) is an (n ± ν)2 − ν 2 + an−2 = 0 for n = 2. Note that it is not possible for the series solution to terminate and give a polynomial solution. which implies that a1 = 0 since 2ν is not an integer. 3. This makes the Bessel functions rather more diﬃcult to study. which is what happens to give the Legendre polynomials. We can now distinguish various cases. We will now determine an expression for the value of an for general values of the parameter ν. which is proportional to xc+1 . The recurrence relation gives us an−2 . an = − n(n ± 2ν) Let’s start with n = 2. . a4 = − a2 . We start with the case for which the diﬀerence between the two roots of the indicial equation. Using Frobenius General Rule II. a4 = (4 ± 2ν)(2 ± 2ν) · 4 · 2 2 (2 ± ν)(1 ± ν)22 (2 · 1) .7) The indicial equation is therefore c2 − ν 2 = 0. 2n + 1. we can show that the other solution is ∞ Y0 (x) = J0 (x) log x − n=0 (−1)n φ(n) (n!)2 x2 4 n . which is called Weber’s Bessel function of order zero. 1 d 2y 1 dy + 1− 2 + dx2 x dx 4x . The general solution of Bessel’s equation.1). beginning with 2ν an odd integer. By setting ν = 0 in the expression (3. 22n (1 ± ν)n n! With a suitable choice of a0 we can write this as y(x) = A x2 /4 x±ν = AJ±ν (x).8) where we have used the Pockhammer symbol to simplify the expression. As an example. let’s consider the case ν = 1 2 so that Bessel’s equation is y = 0. We now consider the case for which 2ν is a nonzero integer. 2 (1 ± ν)n n! (3. Let’s now consider what happens when ν = 0. one solution is ∞ J0 (x) = n=0 (−1)n 1 (n!)2 x2 4 n . (−1)n ±ν Γ(1 ± ν) 2 (1 ± ν)n n! n=0 ∞ n These are the Bessel functions of order ±ν. From this expression for a2n . In this case. the solution takes the form y(x) = AJn+1/2 (x) + BJ−n−1/2 (x). Using Frobenius General Rule III. This expression can be derived by evaluating ∂y/∂c at c = 0. y(x) = a0 x±ν ∞ (−1)n n=0 x2n . (−1)n 2 Γ(1 ± ν) n=0 (1 ± ν)n n! ∞ n (3.62 BESSEL FUNCTIONS We can continue this process. for arbitrary constants A and B. (3.9) Remember that 2ν is not an integer. with x2 /4 x±ν J±ν (x) = ±ν .8) for an and exploiting the fact that (1)n = n!. is therefore y(x) = AJν (x) + BJ−ν (x).21). and we ﬁnd that a0 a2n = (−1)n 2n . in which case the indicial equation has a repeated root and we need to apply Frobenius General Rule III. Note that we have made use of the function φ(n) deﬁned in (1. plot(x.6 in a similar way. " plot(x.2 we show the ﬁrst two Weber’s Bessel functions of integer order. all of these Bessel functions are oscillatory. 7. Yn (x) → −∞. We note that the second solution of Bessel’s equation can also be determined using the method of reduction of order as x y(x) = AJν (x) + BJν (x) 1 dq.1 using the MATLAB script # x=0:0. plot(x.4048. where Yν is Weber’s Bessel function of order ν deﬁned as Yν (x) = Jν (x) cos νπ − J−ν (x) .besselj(1. x1/2 x The recurrence relations (3. so this case requires careful treatment.x)). Finally we consider what happens when 2ν is an even integer.x)).5201 and 8.besselj(2.2 SERIES SOLUTIONS OF BESSEL’S EQUATION 63 We considered this example in detail in Section 1. all to four decimal places. using the MATLAB functions bessely. J−1/2 (x) = πx 2 cos x. We generated Figure 3. subplot(2.1).8317. i > 1. 5.3).2. As you can see. q Jν (q)2 In Figure 3. subplot(2.02:20.2. sin νπ (3.4).x)). title(’J_0(x)’) title(’J_1(x)’) title(’J_2(x)’) title(’J_3(x)’) ! We produced Figures 3.2. plot(x. subplot(2. whilst the ﬁrst three nontrivial zeros of J1 (x) are 3. and hence ν is an integer. . A rather lengthy calculation allows us to write the solution in the form y = AJν (x) + BYν (x). and that Ji (0) = 0 for j < i. besseli and besselk.2) J1/2 (x) = 2 sin x.1735. then show that Jn+1/2 (x) and J−n−1/2 (x) are products of ﬁnite polynomials with sin x and cos x.3. which we will derive later. Note that J0 (0) = 1.1.2. πx cos x sin x + a1 1/2 .10) Notice that the denominator of this expression is obviously zero when ν is an integer.22).2. and found that y(x) = a0 This means that (see Exercise 3.besselj(0.3.besselj(3. but that Ji (0) = 0 (j) for i > 0. The ﬁrst three zeros of J0 (x) are 2.21) and (3. 3. Notice that as x → 0.0156 and 10.5 and 3.2).1 we show Ji (x) for i = 0 to 3.x)). In Figure 3. subplot(2.6537. for ν = n. n an integer Rather like the Legendre polynomials. J2 (x) and J3 (x). we have. The functions Y0 (x) and Y1 (x).1. 3. Fig.9). The functions J0 (x). J1 (x). xn Jn (x) = n 2 Γ(1 + n) ∞ i=0 −x2 /4 = i!(1 + n)i i ∞ i=0 (−1)i x2i+n . 22i+n i!(1 + n)i Γ(1 + n) .64 BESSEL FUNCTIONS Fig. From (3. there is a simple function that will generate all of the Bessel functions of integer order. 3. In order to establish this. 3.2. it is useful to manipulate the deﬁnition of the Jν (x).3 The Generating Function for Jn (x). 3.3 THE GENERATING FUNCTION FOR Jn (x), n AN INTEGER 65 Using (3.6) we note that Γ(1 + n)(1 + n)i = Γ(1 + n + i), and using (3.4) we ﬁnd that this is equal to (n + i)!. For n an integer, we can therefore write Jn (x) = i=0 (−1)i x2i+n . 22i+n i!(n + i)! (3.11) Let’s now consider the generating function g(x, t) = exp 1 1 x t− 2 t . (3.12) The series expansions of each of the constituents of this are exp (xt/2) = j=0 (xt/2)j x , exp − = j! 2t x ∞ i=0 (−x/2t) , i! i both from the Taylor series for e . These summations can be combined to produce ∞ ∞ g(x, t) = j=0 i=0 (−1)i xi+j tj−i . 2j+i i!j! ∞, this becomes (−1)i x2i+n 22i+n i!(n + i)! tn . Now, putting j = i + n so that −∞ n ∞ i=0 g(x, t) = n=−∞ Comparing the coeﬃcients in this series with the expression (3.11) we ﬁnd that g(x, t) = exp 1 1 x t− 2 t = n=−∞ Jn (x)tn . (3.13) We can now exploit this relation to study Bessel functions. Using the fact that the generating function is invariant under t → −1/t, we have ∞ ∞ Jn (x)tn = n=−∞ n=−∞ Jn (x) − 1 t n = n=−∞ Jn (x)(−1)n t−n . Now putting m = −n, this is equal to −∞ J−m (x)(−1)m tm . m=∞ Now let n = m in the series on the right hand side, which gives ∞ ∞ Jn (x)tn = n=−∞ n=−∞ J−n (x)(−1)n tn . Comparing like terms in the series, we ﬁnd that Jn (x) = (−1)n J−n (x), and hence that Jn (x) and J−n (x) are linearly dependent over R (see Section 1.2.1). This explains why the solution of Bessel’s equation proceeds rather diﬀerently when ν is an integer, since Jν (x) and J−ν (x) cannot then be independent solutions. 66 BESSEL FUNCTIONS The generating function can also be used to derive an integral representation of Jn (x). The ﬁrst step is to put t = e±iθ in (3.13), which yields e±ix sin θ = J0 (x) + or, in terms of sine and cosine, e ±ix sin θ ∞ ∞ ∞ n=1 (±1)n eniθ + (−1)n e−niθ Jn (x), = J0 (x) + 2 n=1 J2n (x) cos 2nθ ± i n=0 J2n+1 (x) sin(2n + 1)θ. Appropriate combinations of these two cases show that cos(x sin θ) = J0 (x) + 2 n=1 ∞ J2n (x) cos 2nθ, (3.14) sin(x sin θ) = 2 n=0 J2n+1 (x) sin (2n + 1) θ, (3.15) and, substituting η = π 2 − θ, we obtain cos(x cos η) = J0 (x) + 2 n=1 ∞ (−1)n J2n (x) cos 2nη, (3.16) sin(x cos η) = 2 n=0 (−1)n J2n+1 (x) cos (2n + 1) η. (3.17) Multiplying (3.16) by cos mη and integrating from zero to π, we ﬁnd that π cos mη cos(x cos η) dη = η=0 πJm (x) 0 for m even, for m odd. Similarly we ﬁnd that π sin mη sin(x cos η) dη = η=0 0 πJm (x) for m even, for m odd, from (3.17). Adding these expressions gives us the integral representation Jn (x) = 1 π π cos (nθ − x sin θ) dθ. 0 (3.18) We shall now describe a problem concerning planetary motion that makes use of (3.18). This is the context in which Bessel originally studied these functions. Example: Planetary motion We consider the motion of a planet under the action of the gravitational force exerted by the Sun. In doing this, we neglect the eﬀect of the gravitational attraction of the other bodies in the solar system, which are all much less massive than the Sun. Under this approximation, it is straightforward to prove Kepler’s ﬁrst and second laws, that the planet moves on an ellipse, with the Sun at one of its foci, 3.3 THE GENERATING FUNCTION FOR Jn (x), n AN INTEGER 67 and that the line from the planet to the Sun (PS in Figure 3.3) sweeps out equal areas of the ellipse in equal times (see, for example, Lunn, 1990). Our aim now is to use Kepler’s ﬁrst and second laws to obtain a measure of how the passage of the planet around its orbit depends upon time. We will denote the length of the a ea Fig. 3.3. The elliptical orbit of a planet, P , around the Sun at a focus, S. Here, C is the centre of the ellipse and A and A the extrema of the orbit on the major axis of the ellipse. semi-major axis, A C, by a and the eccentricity by e. Note that distance of the Sun from the centre of the ellipse, SC, is ea. We also deﬁne the mean anomaly to be µ = 2π Area of the elliptic sector ASP , Area of the ellipse which Kepler’s second law tells us is proportional to the time of passage from A to P. Let’s now consider the auxiliary circle, which has centre C and passes through A and A , as shown in Figure 3.4. We label the projection of the point P onto the auxiliary circle as Q. We will also need to introduce the eccentric anomaly of P , which is deﬁned to be the angle ACQ, and can be written as φ = 2π Area of the sector ACQ . Area of the auxiliary circle We now note that, by orthogonal projection, the ratio of the area of ASP to that of the ellipse is the same as the ratio of the area of ASQ to that of the auxiliary circle. The area of ASQ is given by the area of the sector ACQ ( 1 φa2 ) minus the 2 area of the triangle CSQ ( 1 ea2 sin φ), so that 2 µ = π and hence µ = φ − e sin φ. (3.19) 1 2 2a φ − 1 ea2 sin φ 2 , 1 πa2 2 Now, in order to determine φ as a function of µ, we note that φ − µ is a periodic 68 BESSEL FUNCTIONS Fig. 3.4. The auxiliary circle and the projection of P onto Q. function of µ, which vanishes when P and Q are coincident with A or A , that is when µ is an integer multiple of π. Hence we must be able to write φ−µ= n=1 An sin nµ. (3.20) As we shall see in Chapter 5, this is a Fourier series. In order to determine the constant coeﬃcients An , we diﬀerentiate (3.20) with respect to µ to yield dφ −1= nAn cos nµ. dµ n=1 We can now exploit the orthogonality of the functions cos nµ to determine An . We multiply through by cos mµ and integrate from zero to π to obtain π µ=0 ∞ dφ − 1 cos mµ dµ = dµ π cos mµ µ=0 πm dφ dµ = Am . dµ 2 Since φ = 0 when µ = 0 and φ = π when µ = π, we can change the independent variable to give πm Am = 2 π π cos mµ dφ. φ=0 Substituting for µ from (3.19), we have that Am = 2 mπ cos m (φ − e sin φ) dφ. φ=0 2 m Jm (me), Finally, by direct comparison with (3.18), Am = so that . 1 1 φ = µ + 2 J1 (e) sin µ + J2 (2e) sin 2µ + J3 (3e) sin 3µ + · · · 2 3 3.4 DIFFERENTIAL AND RECURRENCE RELATIONS 69 Since µ is proportional to the time the planet takes to travel from A to P , this expression gives us the variation of the angle φ with time. 3.4 Diﬀerential and Recurrence Relations Between Bessel Functions It is often useful to ﬁnd relationships between Bessel functions with diﬀerent indices. We will derive two such relationships. We start with (3.9), multiply by xν and diﬀerentiate to obtain d ν d [x Jν (x)] = dx dx (−1)n x2n+2ν 22n+ν n! Γ(1 + ν + n) n=0 = (−1)n (2n + 2ν) x2n+2ν−1 . 22n+ν n! Γ(1 + ν + n) n=0 Since Γ(1 + ν + n) = (n + ν)Γ(n + ν), this gives a factor that cancels with the term 2(n + ν) in the numerator to give d ν (−1)n x2n xν xν−1 [x Jν (x)] = . dx 22n+ν−1 n! Γ(ν + n) n=0 We can rewrite this so that we have the series expansion for Jν−1 (x), as d ν xν xν−1 [x Jν (x)] = ν−1 dx 2 so that d {xν Jν (x)} = xν Jν−1 (x). dx (3.21) ∞ n=0 ∞ 22n (−1)n x2n , n! Γ(ν)(ν)n Later, we will use this expression to develop relations between general Bessel functions. Note that by putting ν equal to zero d {J0 (x)} = J−1 (x). dx However, we recall that Jn (x) = (−1)n J−n (x) for n an integer, so that J1 (x) = −J−1 (x) and hence J0 (x) = −J1 (x), where we have used a prime to denote the derivative with respect to x. In the same vein as the derivation of (3.21), d −ν d (x Jν (x)) = dx dx d dx −x2 /4 x−ν xν 2ν Γ(1 + ν) n=0 n! (1 + ν)n (−1)n x2n−1 n . 22n+ν−1 n! Γ(1 + ν + n) n=0 n = (−1)n x2n 22n+ν n! Γ(1 + ν + n) n=0 = Notice that the ﬁrst term in this series is zero (due to the factor of n in the numerator), so we can start the series at n = 1 and cancel the factors of n. The series 70 BESSEL FUNCTIONS can then be expressed in terms of the dummy variable m = n − 1 as d −ν 1 (x Jν (x)) = ν−1 dx 2 1 2ν+1 ∞ m=0 (−1)m+1 x2m+1 22m+2 m! Γ(ν + m + 2) m=0 (−1)m x2m+1 . m! Γ(ν + m + 2) =− 22m Using the fact that Γ(ν + m + 2) = Γ(ν + 2) (ν + 2)m and the series expansion of Jν+1 (x), (3.11), we have −x2 /4 d x x−ν Jν (x) = − ν+1 , dx 2 Γ(ν + 1) m=0 m! (2 + ν)m and consequently d x−ν Jν (x) = −x−ν Jν+1 (x). dx (3.22) ∞ m Notice that (3.21) and (3.22) both hold for Yν (x) as well. We can use these relationships to derive recurrence relations between the Bessel functions. We expand the diﬀerentials in each expression to give the equations ν Jν (x) + Jν (x) = Jν−1 (x), x where we have divided through by xν , and Jν (x) − ν Jν (x) = −Jν+1 (x), x where this time we have multiplied by xν . By adding these expressions we ﬁnd that Jν (x) = and by subtracting them 2ν Jν (x) = Jν−1 (x) + Jν+1 (x), x which is a pure recurrence relationship. These results can also be used when integrating Bessel functions. For example, consider the integral I= xJ0 (x) dx. 1 {Jν−1 (x) − Jν+1 (x)} , 2 This can be integrated using (3.21) with ν = 1, since I= xJ0 (x) dx = (xJ1 (x)) dx = xJ1 (x). 3.5 MODIFIED BESSEL FUNCTIONS 71 3.5 Modiﬁed Bessel Functions We will now consider the solutions of the modiﬁed Bessel equation, d 2y dy − (x2 + ν 2 )y = 0, +x (3.23) dx2 dx which can be derived from Bessel’s equation using x → ix, so that the solutions could be written down as conventional Bessel functions with purely imaginary arguments. However, it is more transparent to introduce the modiﬁed Bessel function of ﬁrst kind of order ν, Iν (x), so that the complete solution of (3.23) is x2 y(x) = AIν (x) + BI−ν (x), provided ν is not an integer. As was the case when we introduced the function Yν (x), there is a corresponding function here Kν (x), the modiﬁed Bessel function of second kind of order ν. This is deﬁned as π {I−ν (x) − Iν (x)} . Kν (x) = 2 sin νπ Note that the slight diﬀerence between the deﬁnition of this Bessel function and that of Weber’s Bessel function of order ν, (3.10), occurs because these functions must agree with the deﬁnition of the ordinary Bessel functions when ν is an integer. Most of the results we have derived in this chapter can be modiﬁed by changing the argument from x to ix in deriving, for example, the recurrence relations. The ﬁrst few modiﬁed Bessel functions are shown in Figures 3.5 and 3.6. Note the contrast in behaviour between these and the Bessel functions Jν (x) and Yν (x). Equations (3.21) and (3.22) also hold for the modiﬁed Bessel functions and are given by d d {xν Iν (x)} = xν Iν−1 (x), x−ν Iν (x) = x−ν Iν+1 (x) dx dx and d d {xν Kν (x)} = −xν Kν−1 (x), x−ν Kν (x) = −x−ν Kν+1 (x). dx dx 3.6 Orthogonality of the Bessel Functions In this section we will show that the Bessel functions are orthogonal, and hence can be used as a basis over a certain interval. This will then allow us to develop the Fourier–Bessel series, which can be used to represent functions in much the same way as the Fourier–Legendre series. We will consider the interval [0, a] where, at this stage, a remains arbitrary. We start from the fact that the function Jν (x) satisﬁes the Bessel equation, (3.1), and make a simple transformation, replacing x by λx, where λ is a real constant, to give d2 d J (λx) + x Jν (λx) + (λ2 x2 − ν 2 )Jν (λx) = 0. (3.24) 2 ν dx dx We choose λ so that Jν (λa) is equal to zero. There is a countably inﬁnite number x2 72 BESSEL FUNCTIONS Fig. 3.5. The modiﬁed Bessel functions of ﬁrst kind of order zero to three. Note that √ they are bounded at x = 0 and monotone increasing for x > 0, with In (x) ∼ ex / 2πx as x → ∞. Fig. 3.6. The modiﬁed Bessel functions of the second kind of order zero and order one. Note that they are singular, with K0 (x) ∼ − log x and Kn (x) ∼ 2n−1 (n − 1)!/xn for n a positive integer, as x → 0. These functions are monotone decreasing for x > 0, tending to zero exponentially fast as x → ∞. of values of λ for which this is true, as we can deduce from Figure 3.1. Now choose µ = λ so that Jν (µa) = 0. Of course Jν (µx) also satisﬁes (3.24) with λ replaced by µ. We now multiply (3.24) through by Jν (µx)/x and integrate between zero and a, which yields a Jν (µx) x 0 d2 d 1 Jν (λx) + (λ2 x2 − ν 2 )Jν (λx) dx = 0. J (λx) + 2 ν dx dx x (3.25) 3.6 ORTHOGONALITY OF THE BESSEL FUNCTIONS 73 Notice that we could have also multiplied the diﬀerential equation for Jν (µx) by Jν (λx)/x and integrated to give a Jν (λx) x 0 d2 d 1 Jν (µx) + (µ2 x2 − ν 2 )Jν (µx) dx = 0. J (µx) + 2 ν dx dx x d dx d d Jν (λx) − Jν (λx) dx dx d Jν (µx) dx (3.26) We now subtract (3.25) from (3.26) to give a Jν (µx) 0 x x + x(λ2 − µ2 )Jν (λx)Jν (µx) dx = 0. (3.27) We have simpliﬁed these expressions slightly by observing that xJν + Jν = (xJν ) . Now consider the ﬁrst term of the integrand and note that it can be integrated by parts to give a Jν (µx) 0 d dx − x d d Jν (λx) dx = Jν (µx)x Jν (λx) dx dx a a 0 x 0 d d Jν (λx) Jν (µx)dx. dx dx Similarly for the second term, which is eﬀectively the same with µ and λ interchanged, a Jν (λx) 0 d dx − x d d Jν (µx) dx = Jν (λx)x Jν (µx) dx dx a a 0 x 0 d d Jν (µx) Jν (λx)dx. dx dx Using these expressions in (3.27), we ﬁnd that a (λ2 − µ2 ) 0 xJν (λx)Jν (µx)dx = Jν (λa)aµJν (µa) − Jν (µa)aλJν (λa). (3.28) Finally, since we chose λ and µ so that Jν (λa) = Jν (µa) = 0 and µ = λ, a xJν (µx)Jν (λx)dx = 0. 0 This is an orthogonality relation for the Bessel functions, with weighting function a w(x) = x. We now need to calculate the value of 0 xJν (µx)2 dx. To this end, we substitute λ = µ + into (3.28). For 1, neglecting terms in 2 and smaller, we ﬁnd that a −2µ 0 xJν (µx)Jν (µx + x) dx (3.29) = a [µJν (µa + a)Jν (µa) − (µ + )Jν (µa)Jν (µa + a)] . In order to deal with the terms evaluated at x = a(µ + ) we consider Taylor series 74 BESSEL FUNCTIONS expansions, Jν (a(µ + )) = Jν (aµ) + aJν (aµ) + · · · , Jν (a(µ + )) = Jν (aµ) + aJν (aµ) + · · · . These expansions can then be substituted into (3.29). On dividing through by and considering the limit → 0, we ﬁnd that a x[Jν (µx)]2 dx = 0 a2 1 a [J (µa)]2 − a2 Jν (µa)Jν (µa) − Jν (µa)Jν (µa). 2 ν 2 2µ We now suppose that Jν (µa) = 0, which gives a x[Jν (µx)]2 dx = 0 a2 [J (µa)]2 . 2 ν a2 [J (µa)]2 δλµ , 2 ν In general, a xJν (µx)Jν (λx) dx = 0 (3.30) where Jν (µa) = Jν (λa) = 0 and δλµ is the Kronecker delta function. We can now construct a series expansion of the form f (x) = i=1 Ci Jν (λi x). (3.31) This is known as a Fourier–Bessel series, and the λi are chosen such that Jν (λi a) = 0 for i = 1, 2, . . . , λ1 < λ2 < · · · . As we shall see later, both f (x) and f (x) must be piecewise continuous for this series to converge. After multiplying both sides of (3.31) by xJν (λj x) and integrating over the interval [0, a] we ﬁnd that a a ∞ xJν (λj x)f (x) dx = 0 0 xJν (λj x) i=1 Ci Jν (λi x) dx. Assuming that the series converges, we can interchange the integral and the summation to obtain a ∞ a xJν (λj x)f (x) dx = 0 i=1 Ci 0 xJν (λj x)Jν (λi x) dx. We can now use (3.30) to give a ∞ xJν (λj x)f (x) dx = 0 i=1 Ci a2 λj a2 2 2 [Jν (λj a)] δλj λi = Cj [Jν (λj a)] , 2 2 and hence Cj = 2 a2 [Jν (λj a)] 2 0 a xJν (λj x)f (x) dx. (3.32) Example Let’s try to expand the function f (x) = 1 on the interval 0 x 1, as a Fourier– Bessel series. Since J0 (0) = 1 but Ji (0) = 0 for i > 0, we will choose ν to be zero for our expansion. We rely on the existence of a set of values λj such that J0 (λj ) = 0 2. end & % . Jν (x) is positive for even m and negative for odd m.6 ORTHOGONALITY OF THE BESSEL FUNCTIONS 75 for j = 1.3. This phenomenon always occurs in series expansions relative to sequences of orthogonal functions. end_int = (ir-1)*pi+7*pi/8. . and is called Gibbs’ phenomenon. In Figure 3. it can be shown that the positive zeros of J0 (x) lie in the intervals nπ + 3π/4 < x < nπ + 7π/8. (see Figure 3. Notice the oscillatory nature of the solution. Chapter 15) for a full discussion of this problem. . If we introduce the variable y = λj x (so that dy = λj dx). we need to be able to calculate the zeros of Bessel functions. we have Cj = 2 λ2 j [J0 (λj )] ∞ i=1 2 1 2 λ2 [J0 (λj )] j λ λj y=0 d (yJ1 (y)) dy dy = 2 . The Bessel– Lommel theorem on the location of the zeros of the Bessel functions Jν (x) states 1 1 that when − 1 < ν 2 2 and mπ < x < (m + 2 )π. . A simple MATLAB script that uses this result is '$ global nu ze nu=0.21) with ν = 1. . This implies that Jν (x) has an odd number of zeros in the intervals (2n − 1)π/2 < x < 2nπ for n an integer. λj J1 (λj ) = and hence [yJ1 (y)]0 j = 2J1 (λj ) λj [J0 (λj )] 2 2 J0 (λi x) = 1 for 0 λi J1 (λi ) x < 1.7. where J0 (λi ) = 0 for i = 1. .1). which is more pronounced in the neighbourhood of the discontinuity at x = 1. Using (3. . the integral becomes Cj = 1 2 2 [J0 (λj )] λj 2 λj yJ0 (y) dy.7 we show the sum of the ﬁrst ﬁfteen terms of the Fourier–Bessel series. The interested reader is referred to Watson (1922. y=0 Using the expression (3. In fact. Before we can give a MATLAB script that produces Figure 3. ze(ir) = fzero(’bessel’.[start_int end_int]). Here we merely state a couple of simple results and explain how these are helpful. we have Cj = 2 [J0 (λj )] 2 2 0 1 xJ0 (λj x) dx. This allows us to use the ends of these intervals as an initial bracketing interval for the roots of J0 (x). We will need to determine these values either from tables or numerically. 2.32). . for ir=1:20 start_int = (ir-1)*pi+3*pi/4.

The series is shown for x > 1 only to demonstrate that convergence is only guaranteed for the associated interval.x). 3.76 BESSEL FUNCTIONS Fig. By deﬁning the variables nu and ze as global. A MATLAB script for this is .    The MATLAB function fzero ﬁnds a zero of functions of a single variable in a given interval. together with the function  function bessel = bessel(x). Fourier–Bessel series representation of the function f (x) = 1 on the interval [0. in the script below. Consequently the zeros of J0 (x) can be used as the bracketing intervals for the determination of the zeros of J1 (x). In particular.7. The zeros of J1 (x) interlace those of J0 (x). We can see this by noting that J1 (x) = −J0 (x) and that both functions are continuous. global nu bessel = besselj(nu. The expansion is only valid in the interval x ∈ [0. 1) (truncated after ﬁfteen terms). this allows us to use the computed positions of the zeros. 1). we make them available to other functions. ze.

The constant can then be found by considering the behaviour of the functions close to x = 0 (see Exercise 3. The solution can be written as x y(x) = s sin νπ f (s) (Jν (s)Yν (x) − Jν (x)Yν (s)) ds + AJν (x) + BYν (x).33) 2ν Here we have made use of the fact that the Wronskian associated with Jν (x) and Yν (x) is 2ν/x sin νπ. 3. (1. end & 3. end_int = ze(ir+1). for ir=1:19 start_int = ze(ir). We can now deﬁne a MATLAB function $' function fourierbessel = fourierbessel(x) global ze n=15. (3. +x 2 dx dx can be dealt with by using the technique of variation of parameters (see Section 1. a = 2. end fourierbessel = X*a’.7 Inhomogeneous Terms in Bessel’s Equation So far we have only concerned ourselves with homogeneous forms of Bessel’s equation.ze(k)*x(:)).ze(1:n)). which can be derived using Abel’s formula. for k = 1:n X(:.7 INHOMOGENEOUS TERMS IN BESSEL’S EQUATION$ 77 % The zeros of the other Bessel functions can be found by exploiting similar analytical results. +x 2 dx dx .k) = besselj(0. & % which can be plotted with ezplot(@fourierbessel. [0 2]) to produce Figure 3. Example Let’s ﬁnd the general solution of x2 d 2y dy + x2 y = x2 .7.[start_int end_int]).5). The inhomogeneous version of Bessel’s equation.7)./ze(1:n). ze1(ir) = fzero(’bessel’.' global nu ze ze1 nu=1. x2 d 2y dy + (x2 − ν 2 )y = f (x).2)./besselj(1.

At next order we ﬁnd that a1 (22 − ν 2 ) = 0.33) with f (x) = x2 and ν = 0. For the general terms in the summation we have an−2 an = − .78 BESSEL FUNCTIONS This can be determined by using (3. Example Let’s ﬁnd the particular integral of d 2y dy + (x2 − ν 2 )y = x. we obtain an expression similar to (3. Although it is clear.34) dx2 dx We will look for a series solution as used in the method of Frobenius. namely x2 ∞ y(x) = n=0 an xn+c . and hence y(x) = 1 + AJ0 (x) + BY0 (x). {(n + 1)2 − ν 2 } Note that since a1 = 0. It now remains to determine the general term in the sequence.7). 2 In order to integrate sJ0 (s) we note that this can be written as (sJ1 (s)) and similarly for sY0 (s). For n = 2 we ﬁnd that a2 = − a0 1 =− 2 32 − ν 2 (1 − ν 2 )(32 − ν 2 ) 1 . an = 0 for n odd. We will defer discussion of the case ν = 1. Substituting this into (3. Note that x needs to match with the x on the right hand side so that c = 1 and a0 = 1/(1 − ν 2 ). simple solutions are not always easy to spot a priori. so that the general solution is x y(x) = πs (J0 (s)Y0 (x) − J0 (x)Y0 (s)) ds + AJ0 (x) + BY0 (x). and consequently. unless ν = 2. that y(x) = 1 is the particular integral solution. (12 − ν 2 )(32 − ν 2 )(52 − ν 2 ) and then with n = 4 and using the form of a2 we have a4 = . a0 (c2 − ν 2 )xc + a1 {(1 + c)2 − ν 2 }xc+1 ∞ + n=2 c [an {(n + c)2 − ν 2 } + an−2 ]xn+c = x. which gives πx y(x) = (J1 (x)Y0 (x) − J0 (x)Y1 (x)) + AJ0 (x) + BY0 (x). +x (3.34). so that the expression in the brackets is merely the Wronskian. we have a1 = 0. 2 But we note that J1 (x) = −J0 (x) and Y1 (x) = −Y0 (x). with hindsight.

(2i − 1)2 − ν 2 This can be manipulated by factorizing the denominator and extracting a factor of 22 from each term in the product (of which there are n + 1). They are undeﬁned when µ ± ν is an odd negative integer. 3. n+1 Hence the particular integral of the diﬀerential equation is y(x) = x (−1)n 22n+2 n=0 x2n 1 2 + 1ν 2 n+1 1 2 − 1ν 2 . Since α and β could be fractional. solutions of the equation x2 d 2y dy + (x2 − ν 2 )y = xµ+1 +x 2 dx dx (3. n+1 (3. we will restrict our attention to x 0.35) is correct.ν and are called Lommel’s functions. The series expansion of the solution of (3. a case that is discussed in depth by Watson (1922). ˜ Example Let’s try to express the solutions of the diﬀerential equation d 2y − xy = 0 dx2 in terms of Bessel functions. This gives a2n = (−1)n 22n+2 n+1 i=1 1 1 i − 2 + 1ν 2 ∞ n+1 i=1 1 (−1)n = 2n+2 2 i − 1 − 1ν 2 2 1 1 2 + 1 2 ν n+1 1 2 − 1ν 2 .35) In fact.3. ˜ .36) are commonly referred to as sµ. We start by introducing the function y . We can use this to check that (3. We substitute this expression into the diﬀerential equation and seek values of α and β which give Bessel’s equation for y . we consider the transformation ˜ y(x) = xα y (xβ ).36) is y(x) = x µ−1 (−1)m 1 x Γ 1µ − 1ν + 1 Γ 1µ + 1ν + 2 2 2 2 2 2 Γ 1µ − 1ν + m + 3 Γ 1µ + 1ν + m + 3 2 2 2 2 2 2 m=0 ∞ 2m+2 1 2 .8 SOLUTIONS EXPRESSIBLE AS BESSEL FUNCTIONS 79 The general expression is n+1 a2n = (−1)n i=1 1 . the Airy functions. Note that we need to use (3. In order to determine some of these. This is called Airy’s equation and has solutions Ai(x) and Bi(x).6).8 Solutions Expressible as Bessel Functions There are many other diﬀerential equations whose solutions can be written in terms of Bessel functions.

but at the outset we will consider the .12.1 Vibrations of an Elastic Membrane We will now derive the equation that governs small displacements. ˜ 2 The coeﬃcient of y gives us that α = 1 . ˜ Considering the coeﬃcient of y we note that we require x3 ∝ x2β which gives β = 3 . Newton’s second law of motion in the vertical. which gives ˜ ˜ x2β y + (2α + β − 1) β 1 α(α − 1) x y + − 2 x3 + ˜ β β β2 y = 0. as shown in Figure 3. ˜ which has solutions K1/3 (2x3/2 /3) and I1/3 (2x3/2 /3). The Airy functions Ai and Bi are available in MATLAB through the function airy. The general solution of Airy’s equation in terms of Bessel’s functions is therefore y(x) = x1/2 AK1/3 2 3/2 x 3 + BI1/3 2 3/2 x 3 . In fact. We start by considering a small membrane with sides of length δSx in the x-direction and δSy in the y-direction. y. t). z-direction gives ∂2z ρδSx δSy = δSy {T sin(ψx + δψx ) − T sin(ψx )} + δSx {T sin(ψy + δψy ) − T sin(ψy )}. 3. ˜ ˜ ˜ dx2 These expressions can now be substituted into Airy’s equation to give ˜ ˜ ˜ ˜ α(α − 1)xα−2 y + (2αβ + β 2 − β)xα+β−2 y + β 2 xα+2β−2 y − xα+1 y = 0. It is now convenient to multiply the entire equation by x−α+2 /β 2 (this means that the coeﬃcient of y is x2β ). of an elastic membrane. assumed constant for small vibrations of the membrane. We will eventually consider the angles ψx and ψy to be small. ψx +δψx and ψy . z = z(x. Ai(x) = x/3K1/3 (2x3/2 /3)/π. ∂t2 where ρ is the constant density (mass per unit area) of the membrane and T is the tension. The equation is now ˜ 2 x2β y + xβ y + ˜ ˜ − 2xβ 3 2 − 1 9 y = 0. ˜ ˜ dx Diﬀerentiating again we obtain d 2y = α(α − 1)xα−2 y + (2αβ + β 2 − β)xα+β−2 y + β 2 xα+2β−2 y . ψy +δψy with the horizontal.8.9 Physical Applications of the Bessel Functions 3.80 BESSEL FUNCTIONS Diﬀerentiating with respect to x we have dy = αxα−1 y + βxα+β−1 y . which makes angles ψx . A graph of this Airy function is shown in Figure 11.9.

Accordingly we expand the trigonometric functions and ﬁnd that ρ ∂2z δψx δψy = cos ψx + cos ψy + ··· . T ∂t2 δSx δSy where we have divided through by the area of the element. to be smaller. Also. 3.9 PHYSICAL APPLICATIONS OF THE BESSEL FUNCTIONS 81 T δSy ψx + δψx T T δS x y ψx ψy + δψy ψy T x Fig. A small section of an elastic membrane.37) By diﬀerentiating these expressions with respect to x and y respectively we ﬁnd that sec2 ψx ∂2z ∂2z ∂ψx ∂ψy = = . tan ψy = . T ∂t2 ∂Sx ∂Sy We can now use the deﬁnition of the partial derivatives. changes in these angles. 2 ∂x ∂x ∂y ∂y 2 For small slopes. We now consider the limit as the size of the element shrinks to zero. cos ψx and sec2 ψx are both approximately unity (and similarly for variation in the y direction). tan ψx = ∂z ∂z . δSx δSy . . Consequently we ﬁnd that ∂ψx ∂ψy ρ ∂2z = cos ψx + cos ψy . using the formula for arc length we have ∂z ∂x 2 dSx = 1+ dx ≈ dx.8. and therefore let δSx and δSy tend to zero. sec2 ψy .3. δψx and δψy . ∂x ∂y (3.

2 ∂t ∂x (3. t). One-Dimensional Solutions of the Wave Equation We will start by considering the solution of this equation in one dimension. ∂2z ∂2z = c2 2 . ξ = x − ct and η = x + ct (see Section 7. 2 2 T ∂t ∂x ∂y (3.82 BESSEL FUNCTIONS when |∂z/∂x| 1. ∂2z ∂2z ρ ∂2z = + 2 = ∇2 z. ≈ ≈ ∂Sx ∂x ∂Sy ∂y 1. 2001). We will deﬁne appropriate boundary conditions in due course. At this stage we have not speciﬁed the domain of solution. If we look for solutions of the form z ≡ z(x. . see Billingham and King. Similarly dSy ≈ dy when |∂z/∂y| ∂ψx ∂ψx ∂ψy ∂ψy . independent of y. as illustrated in Figure 3.9. A one-dimensional solution of the two-dimensional wave equation. known as characteristic Fig. t) of an elastic membrane.1 for an explanation of where . variables. Consequently Combining this information yields the governing equation for small deﬂections z = z(x. we need to solve the one-dimensional wave equation.39) is to deﬁne new variables.9.38) the two-dimensional wave equation.39) where c = T /ρ. 3. The easiest way to solve (3. y. This equation also governs the propagation of small-amplitude waves on a stretched string (for further details.

∂ξ and with respect to ξ ξ z= and hence we have that F (s) ds + g(η) = f (ξ) + f (η). f (x − ct) is constant. and one. represented by g(x+ct).43) . represented by f (x − ct). we obtain d’Alembert’s solution of the onedimensional wave equation. (3. a g(x) = 1 1 z0 (x) + 2 2c x u0 (s) ds. −cf (x) + cg (x) = u0 (x). the solution y = f (x − ct).40) with respect to time. (3. one. z(x. z(x. we ﬁnd that f (x) = 1 1 z0 (x) − 2 2c x u0 (s) ds. a On substituting these into (3. Similarly.40) This represents the sum of two waves.39) becomes ∂2z = 0. The functions f and g can be determined from the initial displacement and velocity. This can be integrated to yield x (3. g(x + ct) is constant on the paths x = −ct + constant. from right to left at speed c.3. z(x. Solving the simultaneous equations (3. propagating from left to right without change of form at speed c.42) where a is an arbitrary constant.40) with t = 0 gives us f (x) + g(x) = z0 (x). If ∂z (x.42). ∂t and hence when t = 0. and simply note that on the paths x = ct + constant. for example. we have ∂z = −cf (x − ct) + cg (x + ct). (3. t) = f (x − ct) + g(x + ct). x−ct (3. If we now diﬀerentiate (3. 0) = u0 (x). t) = 1 1 {z0 (x − ct) + z0 (x + ct)} + 2 2c x+ct u0 (s) ds.41) and (3. ∂t then (3.41) −cf (x) + cg(x) = a u0 (s) ds. ∂ξ∂η Integrating this with respect to η gives ∂z = F (ξ). consider.40). To see this. 0) = z0 (x).9 PHYSICAL APPLICATIONS OF THE BESSEL FUNCTIONS 83 these come from). In terms of these variables.

the two-dimensional wave equation becomes 1 ∂2z 1 ∂2z ∂2z 1 ∂z + 2 2. D’Alembert’s solution for an initially stationary top hat displacement. Two-Dimensional Solutions of the Wave Equation Let’s now consider the solution of the two-dimensional wave equation. if u0 = 0. 3. for an elastic. a top hat initial displacement. Fig. Such solutions must be periodic in θ with period 2π. c2 τ rR r Θ c .44). On substituting this into (3. We seek a separable solution. each with the same shape but half the amplitude of the initial displacement. the solution consists of a left-travelling and a right-travelling wave.44) 2 ∂t2 c ∂r r ∂r r ∂θ We will look for solutions in the circular domain 0 r a and 0 θ < 2π. a string released from rest. disc-shaped membrane ﬁxed to a circular support. is shown in Figure 3. (3.10. z = R(r)τ (t)Θ(θ). The boundary condition is z = 0 at r = a.84 BESSEL FUNCTIONS In particular.10. The solution when z0 = 0 for |x| > a and z0 = 1 for |x| < a.38). In cylindrical polar coordinates. we ﬁnd that rR + R 1 Θ ω2 1 τ = + 2 =− 2. = 2 + (3.

the general form of the displacement of the membrane is ∞ ∞ z(r. c2 We can simplify this by introducing a new coordinate s = λr where λ = ω/c. which is. The other boundary condition is that the membrane is constrained not to move at the edge of the domain. θ) and the membrane is released from rest. The angular frequency. R c Θ where n2 is the next separation constant. after all. Consequently. so that ∂z/∂t = 0. Consequently the functional form of the natural modes of oscillation of a circular membrane is z = Jn (λni r) (A cos nθ + B sin nθ) eiωi t . An appropriate solution is τ = eiωt . This gives us Θ = A cos nθ + B sin nθ. using the linearity of the wave equation to add all the possible solutions of the form (3. 0) = G(r. we require B = 0. a drum. with n a positive integer for 2π-periodicity. We could however have tackled an initial value problem. Finally. (3. ω. This gives the condition Jn (λa) = 0. since Yn (s) is unbounded at s = 0. t) = i=0 n=0 Jn (λni r) (Ani cos nθ + Bni sin nθ) cos ωi t. θ. Specifying the value of λ = ω/c prescribes the frequency at which the membrane will oscillate. Here we have considered the natural modes of oscillation. Let’s consider an example where the initial displacement of the membrane is speciﬁed to be z(r. We need a solution that is bounded at the origin so.11 shows a few of these natural modes when a = 1. which we created using the MATLAB function ezmesh (see Section 2.9 PHYSICAL APPLICATIONS OF THE BESSEL FUNCTIONS 85 where −ω 2 /c2 is the separation constant. This is what we would expect for the vibrations of an elastic membrane. so we need only consider a solution proportional to cos ωi t. . when t = 0. θ. which gives d 2R dR + s2 − n2 R = 0. Consequently the solutions can be written as s2 R(s) = AJn (s) + BYn (s). The fact that the membrane is released from rest implies that the temporal variation will be even.1).3. so that R(s) = 0 at s = λa.46) where ωi = cλni and the values of λni are solutions of (3. is yet to be determined. We can now write ω2 r2 r2 R + rR Θ + 2 =− = n2 . R(r) satisﬁes r2 d 2R dR + +r 2 dr dr ω2 2 r − n2 R = 0.46). which represents a time-periodic solution.45).6.45) which has an inﬁnite number of solutions λ = λni . Figure 3. +s 2 ds ds which is Bessel’s equation with ν = n. (3.

11. Multiplying (3. . (3. and with sin mθ we have 1 π 2π ∞ sin mθ G(r. 0) = G(r. using 2π 2π cos mθ cos nθ dθ = 0 0 sin mθ sin nθ dθ = πδmn . θ) dθ = 0 i=0 Ami Jn (λni r). for m and n integers.47) through by cos mθ and integrating from 0 to 2π we have 1 π 2π ∞ cos mθ G(r. Using the condition that z(r. Six diﬀerent modes of oscillation of an elastic membrane. θ. 3. θ) we have the equations ∞ ∞ G(r. given by (3.47) In order to ﬁnd the coeﬃcients Ani and Bni we need exploit the orthogonality of the Bessel functions.30). θ) = i=0 n=0 Jn (λni r) (Ani cos nθ + Bni sin nθ) . and of the trigonometric functions. θ) dθ = 0 i=0 Bmi Jn (λni r).86 BESSEL FUNCTIONS Fig.

1977).49) into its various frequency components. We can expand (3.50) Using simple trigonometry.16) and (3.2 Frequency Modulation (FM) We will now discuss an application in which Bessel functions occur within the description of a modulated wave (for further information. using (3. (3.3. 3. about which we will have more to say in Chapter 5. we have Amj = and Bmj = 2 a2 π [Jn (λj a)] 2 a 2π 2 a2 π [Jn (λj a)] 2 a 2π rJn (λj r) r=0 0 cos mθ G(r.17) to give ∞ Fc (t) = cos(2πfc t) J0 (β) + 2 n=1 ∞ (−1)n J2n (β) cos 2n(2πfm t) + sin(2πfc t) 2 n=0 (−1)n J2n+1 (β) cos (2n + 1) (2πfm t) .9 PHYSICAL APPLICATIONS OF THE BESSEL FUNCTIONS 87 Now. The expansion in θ is an example of a Fourier series expansion. θ) is even. θ) dθ dr rJn (λj r) r=0 0 sin mθ G(r.49) An example of this signal is shown in Figure 3.12. We can do this by exploiting (3.48) where β = K2 a/fm . . (3.9. The phase shift of the carrier is related to the time integral of the modulating wave Fm (t) = cos 2πfm t. so that the actual signal is t Fc (t) = cos 2πfc t + 2πK2 0 a cos(2πfm t)dt . Bmj = 0.48) to give Fc (t) = cos(2πfc t) cos {β sin(2πfm t)} − sin(2πfc t) sin {β sin(2πfm t)} . (3.32). a constant called the modulation index. Amj = 0. θ) dθ dr. see Dunlop and Smith. We would now like to split the signal (3. 2 cos(2πfc t) cos {(2n)2πfm t} = cos {2πfc t + (2n)2πfm t} + cos {2πfc t − (2n)2πfm t} . which we can integrate to obtain Fc (t) = cos {2πfc t + β sin(2πfm t)} . Note that if the function G(r. The expression for a frequency carrier comprises a carrier frequency fc and the modulating signal with frequency fm . and similarly if it is odd.

known as sidebands. fm = 1. the ﬁrst few terms of which are given by Fc (t) = J0 (β) cos(2πfc t) − J1 (β) [cos {2π(fc − fm )t} − cos {2π(fc + fm )t}] + J2 (β) [cos {2π(fc − 2fm )t} + cos {2π(fc + 2fm )t}] − J3 (β) [cos {2π(fc − 3fm )t} − cos {2π(fc + 3fm )t}] + · · · .50). and 2 sin(2πfc t) sin {(2n + 1)2πfm t} = − cos {2πfc t + (2n + 1)2πfm t} + cos {2πfc t − (2n + 1)2πfm t} . with β = 1. have frequencies fc ± nfm for . This means that the main carrier signal has frequency fc and amplitude J0 (β). we ﬁnd that Fc (t) can be written as Fc (t) = J0 (β) cos(2πfc t) ∞ + n=1 ∞ J2n (β) [cos {2πt (fc + 2nfm )} + cos {2πt (fc − 2nfm )}] − n=0 J2n+1 (β) [cos {2πt (fc − (2n + 1)fm )} − cos {2πt (fc + (2n + 1)fm )}] .88 BESSEL FUNCTIONS Fig.2 and fc = 1. 3. Substituting these expressions back into (3. An example of frequency modulation. and that the other components.12.2.

k 3.2 and fc = 1.2 Using the deﬁnition Jν (x) = show that J1/2 (x) = 2 πx xν 2ν Γ(1 + ν) ∞ i=0 (−x2 /4)i . . The frequency spectrum of a signal with β = 0. are shown in Figure 3. fm = 1. and amplitudes Jn (β). Exercises 3.2.1 The functions J0 (x) and Y0 (x) are solutions of the equation x dy d 2y + xy = 0. where i φ(i) = k=1 1 .13. i!(1 + ν)i sin x. 3. n an integer.13.EXERCISES 89 Fig. known as the frequency spectrum of the signal. + dx2 dx ∞ Show that if Y0 (x) is the singular solution then Y0 (x) = J0 (x) log x − i=1 φ(i) (i!)2 − x2 4 i . These amplitudes.

Determine the coeﬃcients of the Fourier–Bessel series for the function f (x) = x on the interval 0 x 1 in terms of the Bessel function J1 (x) (and repeat the exercise for J2 (x)). (d) dx2 d 2y − α2 y = 0. 3. +β dx2 dx d 2y dy + n(n + 1)y = 0. Find the Wronskian of the functions (a) Jν (x) and J−ν (x). of the differential equations d 2y (a) − x2 y = 0.4 3. (e) dx2 dy d 2y (f) + γy = 0. Calculate the Fourier–Bessel expansion for the functions . (b) x 2 + dx dx d 2y (c) x 2 + (x + 1)2 y = 0. 0 3. 3.9 show that Jν (0) = 0 for ν a nonzero integer and J0 (0) = 1. When µ = 1 and ν = 0 show that this is equivalent to xJ1 (x).11 in terms of the Bessel function J0 (x). Modify the MATLAB code used to generate Figure 3.6 Determine the series expansions for the modiﬁed Bessel functions Iν (x) and I−ν (x). Determine the coeﬃcients of the Fourier–Bessel series for the function f (x) = 1 for 0 −1 for 1 x < 1.90 BESSEL FUNCTIONS 3. (g) (1 − x2 ) 2 − 2x dx dx Using the expression Jν (z) = 1 2π 2π 3. where possible in terms of Bessel functions.10 3. x 2. Give the solutions. dx2 dy d 2y + y = 0.7 to check your answers. (b) 2νJν (x) = xJν+1 (x) + xJν−1 (x).3 ∗ Show that (a) 2Jν (x) = Jν−1 (x) − Jν+1 (x). dx d 2y + α2 y = 0.8 cos (νθ − z sin θ) dθ.5 3. (b) Jν (x) and Yν (x) for ν not an integer. Determine the series expansion of xµ Jν (x) dx.7 3.

EXERCISES 91 (a) f (x) = (b) f (x) = x2 + 2 for 0 3 for 1 x < 1. In this case. Here. We will consider some problems associated with the baking of inﬁnitelylong. (c) In many baking problems. t) = f (r)eωt . ∂t The diﬀusivity. When for 0 r r1 . in terms of the Bessel function J1 (x). D1 and the heat ﬂux and the temperature are continuous at r = r1 . subject to the boundary conditions ∂T /∂r = 0 at r = 0 and ∂T /∂r = h(Ta − T ) at r = r0 . is a function of the properties of the food. 0) = T0 T1 for 0 r for r1 < r r1 . (d) Solve the initial–boundary value problem when the initial temperature is uniform in each layer. the baking of a loaf of bread). 3. r0 . is governed by the diﬀusion equation. D0 D(r) = for r1 < r r0 . of (E3. there are two distinct layers of food. determine the eﬀect of changing r1 /r0 and D0 and D1 on the possible values of ω. x for 0 2 − x for 1 x < 1.12 Construct the general solution of the diﬀerential equation d 2y dy + (x2 − ν 2 )y = sin x. . as we showed in Chapter 2. T (r. 0) = T0 (r). x 3.1) with D(r) = D0 . and thereby improve the quality of mass-produced baked foods. D. x 2. for example. cylindrical. the diﬀusion equation becomes ∂T 1 ∂ = ∂t r ∂r rD(r) ∂T ∂r . determine the temperature for t > 0. +x 2 dx dx Project This project arises from attempts to model the baking of foodstuﬀs. a constant. (E3. axisymmetric foodstuﬀs under axisymmetric conditions (a ﬁrst approximation to. (b) If the initial temperature proﬁle within the foodstuﬀ is T (r. with T (r.13 (a) Look for a separable solution. Determine the possible values of ω. A key element of such modelling is the temperature distribution in the food. x2 ∂T = ∇ · (D∇T ) .1) 3. which. h > 0 is a heat transfer coeﬃcient and Ta is the ambient temperature.

either from your library or the Internet.92 BESSEL FUNCTIONS (e) Find some realistic values of D0 and h for bread dough. How does the temperature at the centre of baking bread vary with time. if this simple model is to be believed? .

we will consider the dynamics of a string ﬁxed at x = y = 0 and x = l. . In order to be able y ω θ(x) x=l x Fig. In order to ground our discussion in terms of a real physical problem.1. as shown in Figure 4. x = a and x = b. for which the boundary conditions are speciﬁed at two diﬀerent points. to formulate a diﬀerential equation that captures the dynamics of this string. A string rotating about the x-axis. y = 0. but one whose ends are motionless. we will make several assumptions. This is rather like a skipping rope. The solutions of boundary value problems have some rather diﬀerent properties to those of solutions of initial value problems. and rotating steadily about the x-axis at constant angular velocity. Green’s Functions and Sturm–Liouville Theory We now turn our attention to boundary value problems for ordinary diﬀerential equations. 4.CHAPTER FOUR Boundary Value Problems. ﬁxed at x = 0 and x = l.1.

94 BOUNDARY VALUE PROBLEMS (i) The tension in the string is large enough that any additional forces introduced by the bending of the string are negligible in comparison with the tension force. either due to air resistance on the string or to the ﬁxing of the string at each of its ends. sin θ = and hence d 2y T 2 + ρω 2 y dx 1+ dy dx 2 2 dy dx 1+ dy dx 2 −1/2 .1) This equation of motion for the string must be solved subject to y(0) = y(l) = 0. y = 0. We denote the constant angular velocity of the string about the x-axis by ω. tan θ = dy/dx. that |dy/dx| 1. the leading order boundary value problem is d 2y + λy = 0. however. in which the string is stationary.† If. (ii) The tension force acts along the local tangent to the string. (iii) The slope of the string. This is now a linear boundary value problem. (4. with no elementary solution apart from the equilibrium solution.2. Newton’s ﬁrst law shows that the forces that act on the string. † This nonlinear boundary value problem can. . must be in balance. a mass per unit length. is small. and hence its displacement from the x-axis. = 0. but it has a constant line density. dx2 (4.2) where λ = ρω 2 /T . we invoke assumption (iii). easily be studied in the phase plane (see Chapter 9). If we divide through by δx and take the limit δx → 0 we obtain T d {sin θ (x)} + ρω 2 y dx 1+ dy dx 2 = 0. By deﬁnition. ρ. (vi) The thickness of the string is negligible. by elementary trigonometry. as shown in Figure 4. and hence that T sin θ(x + δx) − T sin θ(x) = −ρω 2 y δx2 + δy 2 . and the angle that the tangent to the string makes with the x-axis by the function θ(x). (iv) The eﬀect of gravity is negligible. including the centrifugal force. which we can solve by elementary methods. which constitutes a rather nasty nonlinear boundary value problem. so. however. Working in a frame of reference that rotates with the string. (v) There is no friction. and is of constant magnitude. subject to y(0) = y(l) = 0. T .

To ensure that y. it should be. so we must have β = 0. Note that there is an inﬁnite sequence of eigenvalues. the displacement of the string. The values of λ for which there is a nontrivial solution of this problem.BOUNDARY VALUE PROBLEMS 95 y T θ(x + δx) δy θ(x) T x x +δx x δx Fig. we obtain m2 + λ = 0. are the eigenfunctions. (4. This gives αl = nπ with n an integer. as expected. y = 0. we have either e−βl − eβl = 0 or cos αl = 0. whilst the corresponding solutions. . . . (4.5) Note that the solution that corresponds to n = 0 is the trivial solution. e−βl + eβl sin αl = 0. The latter leaves the second of 2 equations (4. Since y(0) = 0. and hence sin αl = 0.4) From the ﬁrst of these. A small element of the rotating string. and hence either β = 0 or αl = n + 1 π for n an integer. If we look for a solution of the form y = Aemx . are called the eigenvalues of the boundary value problem. . which are real and have magnitudes that tend .4) with no solution.3) At this stage. (4. which gives y = an sin nπx l for n = 1. y = sin λ1/2 x. we write An = an /2i for an real. so that the solution is y = Aeiλ x + Be−iλ x . although. namely λ = n2 π 2 /l2 .2. and y(l) = 0 gives eiλ 1/2 l − e−iλ 1/2 l = 0. 4. which means that λ1/2 is real. and hence 1/2 1/2 m = ±iλ1/2 .3). 2. we do not know whether λ is real. with λ1/2 = nπ/l and y = An einπx/l − e−inπx/l = 2iAn sin (nπx/l). If we write λ1/2 = α + iβ and equate real and imaginary parts in (4. since it represents a ratio of real quantities. we ﬁnd that e−βl − eβl cos αl = 0. B = −A. is real.

length l and tension T .1 Inhomogeneous Linear Boundary Value Problems Continuing with our example of a rotating string. (4. where there may be either no solution or many solutions depending on the form of F (x). subject to y(0) = y(l) = 0. ω= nπ l T . 0 However. The linearized boundary value problem is then d 2y + λy = F (x). for example. let’s consider what happens when there is a steady. 1937) known as the Fredholm alternative.7) When λ is not an eigenvalue. In terms of the physical problem.7) has a unique solution B=− 1 λ1/2 sin λ1/2 l l F (s) sin λ1/2 (x − s) ds. are in complete contrast to the solutions of an initial value problem. This situation also arises for any system that can be written in the form Ax = b.6). imposed external force T F (x) acting towards the x-axis. as we shall see in Chapter 8. sin λ1/2 l = 0. A famous result (see.6) dx2 We can solve this using the variation of parameters formula. (1. when λ is an eigenvalue. for which. we have sin λ1/2 l = 0. (4. These cases. so that there is a solution for arbitrary values of B. subject to a few mild conditions on the form of the ordinary diﬀerential equation. for a string of given line density ρ.96 BOUNDARY VALUE PROBLEMS to inﬁnity as n tends to inﬁnity. Courant and Hilbert. shows that either there is a unique solution of Ax = b. the eigenfrequencies. 0 (4. a nonequilibrium steady motion is only possible at a discrete set of angular velocities.9) If F (s) does not satisfy this integral constraint there is no solution. ρ 4.8) F (s) sin λ1/2 (l − s) ds = 0. . To satisfy the boundary condition y(0) = 0 we need A = 0. or Ax = 0 has nontrivial solutions. where A is a linear operator. (4. which gives y(x) = A cos λ1/2 x + B sin λ1/2 x + 1 λ1/2 0 x F (s) sin λ1/2 (x − s) ds. and (4. we can prove theorems that guarantee the existence and uniqueness of solutions. whilst y(l) = 0 gives B sin λ1/2 l + 1 λ1/2 0 l F (s) sin λ1/2 (x − s) ds = 0. namely y(x) = B sin λ1/2 x + provided that l 1 λ1/2 0 x F (s) sin λ1/2 (x − s) ds.

We say that there is a resonance. (4.9) there is no steady solution.1 Solubility As we have now seen. whilst varying spatially like sin λ1/2 x. (4. a (4. 4. Let’s examine this further for the general boundary value problem (p(x)y (x)) + q(x)y(x) = f (x) subject to y(a) = y(b) = 0.1 INHOMOGENEOUS LINEAR BOUNDARY VALUE PROBLEMS 97 In terms of the physics of the rotating string problem. and.11) gives b b y(x) (p(x)u (x)) + q(x)u(x) dx = a a u(x)f (x) dx. using u(a) = u(b) = 0. since u(x) is a solution of the homogeneous problem. that is a linear combination of the eigensolution and the response to the forcing.1. The size of B depends upon how the steady state was reached.8). In practice. b] gives b b u(x) (p(x)y (x)) + q(x)y(x) dx = a a u(x)f (x) dx.1) were no longer negligible. for arbitrary forcing of the string there is a unique steady solution.10) If u(x) is a solution of the homogeneous problem.11) Now. if ω is not an eigenfrequency. linear boundary value problem may have no solutions. an inhomogeneous. until the nonlinear terms in (4. Substituting this into (4. then multiplying (4. if the string were forced in this way.10) by u(x) and integrating over the interval [a. just as for the unforced problem. b a u(x) (p(x)y (x)) dx = [u(x)p(x)y (x)]a − b b b p(x)y (x)u (x) dx a =− a p(x)y (x)u (x) dx. This reﬂects the fact that the forcing has a component that drives a response at the eigenfrequency. there is a solution.4. using integration by parts. using y(a) = y(b) = 0. If ω is an eigenfrequency and the forcing does not satisfy (4. the amplitude of the motion would grow linearly. If ω is an eigenfrequency and the forcing satisﬁes (4. (4.12) . Integrating by parts again gives b − a p(x)y (x)u (x) dx b = − [p(x)u (x)y(x)]a + b b (p(x)u (x)) y(x) dx = a a (p(x)u (x)) y(x) dx. b u(x)f (x) dx = 0. so that (p(x)u (x)) +q(x)u(x) = 0 and u(a) = u(b) = 0.9).

which we call a solvability or solubility condition. although correct. and that . b Au1 (b) + Bu2 (b) = a f (s) {u1 (b)u2 (s) − u1 (s)u2 (b)} ds. is not the most convenient one to use. are linear combinations of solutions of the homogeneous problem.13) to obtain b y(x) = a f (s) {u1 (b)u2 (s) − u1 (s)u2 (b)} {u1 (a)u2 (x) − u1 (x)u2 (a)} ds W (s) u1 (a)u2 (b) − u1 (b)u2 (a) x + a f (s) {u1 (s)u2 (x) − u1 (x)u2 (s)} ds.6). must be orthogonal to the solution of the homogeneous problem.14). The function G(x.1. W (s) Provided that u1 (a)u2 (b) = u1 (b)u2 (a). s) for a s < x.15) where G(x. f (x).10) is therefore (4.14) This form of solution. To improve it. which appear in (4.2. W (s) (4. We say that the forcing term.98 BOUNDARY VALUE PROBLEMS A necessary condition for there to be a solution of the inhomogeneous boundary value problem (4. (4. we can solve these simultaneous equations and substitute back into (4.14) as a product. it makes sense to look for a solution of the inhomogeneous boundary value problem in the form b y(x) = a f (s)G(x. Note that variations in the form of the boundary conditions will give some variation in the form of the solvability condition.2 The Green’s Function Let’s now solve (4.10) using the variation of parameters formula. (1. and are therefore themselves solutions of the homogeneous problem. 4. s) ds. s) v1 (x)v2 (s) = G> (x. W (s) (4.13) where u1 (x) and u2 (x) are solutions of the homogeneous problem and W (x) = u1 (x)u2 (x)−u1 (x)u2 (x) is the Wronskian of the homogeneous equation.12). s) is known as the Green’s function for the boundary value problem. it is clear that G is continuous at s = x.3. for x < s b. we note that the functions v1 (x) = u1 (a)u2 (x) − u1 (x)u2 (a) and v2 (x) = u1 (b)u2 (x) − u1 (x)u2 (b). From the deﬁnition. The boundary conditions show that Au1 (a) + Bu2 (a) = 0. They also satisfy v1 (a) = v2 (b) = 0. a terminology that we will explore in more detail in Section 4. s) = v1 (s)v2 (x) = G< (x. Because of the way that they appear in (4. This gives x y(x) = Au1 (x) + Bu2 (x) + s=a f (s) {u1 (s)u2 (x) − u1 (x)u2 (s)} ds.

10). actually satisﬁes the inhomogeneous diﬀerential equation. G> (a. If we substitute this into the diﬀerential equation (4. we obtain x b y (x) = a G<. s)f (s) ds − G> (x. s)f (s) ds. since. x)f (x) + x G>.4. using Abel’s formula. x)f (x). using the deﬁnition of C.x (x. for example. For this to be useful. C is a constant. we must now show that y(x). as deﬁned by (4. to give x b y (x) = a x G<x (x. the integral terms vanish and. If we now diﬀerentiate under the integral sign. and hence x (py ) = a b v1 (s)(pv2x )x f (s) ds + p(x)v1 (x)v2x (x)f (x) + x x (pv1x )x v2 (s)f (s) ds − p(x)v1x (x)v2 (x)f (x) b = a v1 (s)(pv2x )x f (s) ds + x (pv1x )x v2 (s)f (s) ds + Cf (x).x = ∂G< /∂x. s = x+ ) = v1 (x)v2 (x) − v1 (x)v2 (x) = W = C . s)f (s) ds + x G> (x. s)f (s) ds + x b G>x (x. if we choose C = 1. Since v1 and v2 are solutions of the homogeneous problem. s)f (s) ds = 0. s)f (s) ds = a v1 (s)v2x (x)f (s) ds + x v1x (x)v2 (s)f (s) ds. If we calculate the partial derivative with respect to x. To do this. s = x− ) − Gx (x.15). s)f (s) ds + G< (x. we obtain x b v1 (s)(pv2x )x f (s) ds + a x x (pv1x )x v2 (s)f (s) ds + Cf (x) b +q(x) a v1 (s)v2 (x)f (s) ds + x v1 (x)v2 (s)f (s) ds = f (x). our representation provides us with a solution of the diﬀerential equation. .x (x. p(x) b where. s) = 0 and y(a) = a G> (a. This simpliﬁes. by virtue of the continuity of the Green’s function at x = s.1 INHOMOGENEOUS LINEAR BOUNDARY VALUE PROBLEMS 99 y(a) = y(b) = 0. we split the range of integration and write x b y(x) = a G< (x. where G<. (py ) + qy = f . we ﬁnd that Gx (x. and also determine the value of C.

   sinh 1 G(x. s = x− ) − Gx (x.100 BOUNDARY VALUE PROBLEMS As an example. Gx (x. s) = AB sinh(1 − x) sinh s for 0 s < x. we require AB = −1/ sinh 1. it is useful to be able to work with the simplest possible type of linear diﬀerential operator. Appropriate combinations of these that satisfy v1 (0) = v2 (1) = 0 are v1 (x) = A sinh x and v2 (x) = B sinh(1 − x).1 Self-Adjoint Operators We deﬁne a linear diﬀerential operator. b] → C[a. and the ﬁnal Green’s function is   − sinh(1 − x) sinh s for 0 s < x. we developed the idea of a Green’s function. which gives the Green’s function G(x. 4. b].16) . L : C 2 [a. with which we can solve inhomogeneous boundary value problems for linear. In addition. AB sinh(1 − s) sinh x for x < s 1.2. s) =  sinh(1 − s) sinh x    − for x < s 1. sinh 1 We will return to the subject of Green’s functions in the next chapter. second order ordinary diﬀerential equations. We will now develop an alternative approach that draws heavily upon the ideas of linear algebra (see Appendix 1 for a reminder). 4. (4. as being in selfadjoint form if L= d dx p(x) d dx + q(x). s = x+ ) = −AB cosh(1 − x) sinh x − AB sinh(1 − x) cosh x = −AB sinh 1. consider the boundary value problem y (x) − y(x) = f (x) subject to y(0) = y(1) = 0. Before we start. sinh 1 The solution of the inhomogeneous boundary value problem can therefore be written as x which is continuous at s = x.2 The Solution of Boundary Value Problems by Eigenfunction Expansions In the previous section. y(x) = − 0 f (s) sinh(1 − x) sinh s ds − sinh 1 1 f (s) x sinh(1 − s) sinh x ds. The solutions of the homogeneous problem are e−x and ex . Since p(x) = 1.

18) Note that p(x) = 0 for x ∈ [a. For example. d 2y dy + λy = 0. q(x) = exp a2 (t) a2 (x) x a1 (t) dt .2. . For example.21) 4. a2 (t) (4.22) β1 y(a) + β2 y(b) + β3 y (a) + β4 y (b) = 0.2 EIGENFUNCTION EXPANSIONS 101 where p(x) ∈ C 1 [a.19). b].5) x p(x) = exp a1 (t) a0 (x) dt . Since each of these is dependent on the values of y and y at each end of [a. we need to know the boundary conditions. if L1 ≡ a2 (x) d2 d + a0 (x) + a1 (x) dx2 dx (4. This deﬁnition encompasses a wide class of second order diﬀerential operators. By studying inhomogeneous boundary value problems of the form Ly = f . b]. if we follow the above procedure. b] and is strictly nonzero for all x ∈ (a. b]. b]. and kept in self-adjoint form. 2 2 This can be simpliﬁed. or d dx p(x) dy dx + q(x)y = f (x).20) for −∞ < x < ∞. The reasons for referring to such an operator as self-adjoint will become clear later in this chapter. we refer to these as mixed or coupled boundary conditions. It is unnecessarily complicated to work with the boundary conditions in this form. dx2 /2 y. In general these will be of the form α1 y(a) + α2 y(b) + α3 y (a) + α4 y (b) = 0.17) is nonsingular on [a. − 2x dx2 dx (4. the self-adjoint form of the equation is d dx e−x 2 dy dx + λe−x y = 0. consider Hermite’s equation. we can write it in self-adjoint form by deﬁning (see Exercise 4. (4. and q(x) ∈ C[a. but.19) we are therefore considering all second order. and we can start to simplify matters by deriving Lagrange’s identity. This is not in self-adjoint form. linear diﬀerential operators. by writing u = e−x obtain d 2u − (x2 − 1)u = −λu.4. nonsingular. to (4.2 Boundary Conditions To complete the deﬁnition of a boundary value problem associated with (4. (4. b).

when the functions on which D operates are zero at x = 0 and x = 1. y2 for all y1 and y2 . y2 = [ p (y1 y2 − y1 y2 )]a . on the interval [0. +γ 2 dx dx with γ. we ﬁnd that 1 φ1 . Ly2 − Ly1 . satisﬁes y1 . δ ∈ R. g = a f (x)g(x) dx. y1 (Ly2 ) − y2 (Ly1 ) = y1 (py2 ) + qy2 − y2 (py1 ) + qy1 = y1 (py2 ) − y2 (py1 ) = y1 (py2 + p y2 ) − y2 (py1 + p y1 ) = p (y1 y2 − y1 y2 ) + p (y1 y2 − y1 y2 ) = [ p (y1 y2 − y1 y2 )] .102 BOUNDARY VALUE PROBLEMS Lemma 4. 1]. written T . Dφ2 = 0 1 1 φ1 (φ2 + γφ2 + δφ2 ) dx = φ1 φ2 1 1 1 0 − 0 φ1 φ2 dx + γφ1 φ2 1 0 1 0 − 0 γφ1 φ2 dx + 0 1 δφ1 φ2 dx 1 = − φ1 φ2 where + 0 φ1 φ2 dx − 0 γφ1 φ2 dx + 0 ¯ δφ1 φ2 dx = Dφ1 . Proof From the deﬁnition of L. b (4. (4. then y1 (Ly2 ) − y2 (Ly1 ) = [ p (y1 y2 − y1 y2 )] . φ2 . let’s see if we can construct the adjoint to the operator D≡ d d2 + δ. After integrating by parts and applying these boundary conditions.23) over [a. dx dx .23) Now recall that the space C[a.1 (Lagrange’s identity) If L is the linear diﬀerential operator given by (4. d2 d ¯ D ≡ 2 −γ + δ. b].24) This result can be used to motivate the following deﬁnitions. If we now integrate (4. For example. The adjoint operator ¯ ¯ to T . T y2 = T y1 . b] then y1 . and if y1 .16) on [a. b] is a real inner product space with a standard inner product deﬁned by b f. y2 ∈ C 2 [a. b].

This is the case of periodic boundary conditions.1 Eigenfunctions belonging to distinct eigenvalues of a Hermitian linear operator are orthogonal. and hence (4. (i) p(a) = p(b) = 0. whether or not L is Hermitian depends only upon the boundary values of the functions in the space upon which it operates.3 . Eigenvalues and Eigenfunctions of Hermitian Linear Operators The eigenvalues and eigenfunctions of a Hermitian. In other words. This is the case of singular boundary conditions. The signiﬁcance of this result becomes apparent when we examine the eigenvalues and eigenfunctions of Hermitian linear operators. As an example of how such boundary conditions arise when we model physical systems. T y2 = T y1 . or self-adjoint. the periodic conditions p(0) = p(l). are called unmixed or separated.2. linear operator L are the nontrivial solutions of Ly = λy subject to appropriate boundary conditions. y1 (b)y2 (b) − y1 (b)y2 (b) = 0.4.25) b a = 0. b). y(0) = y(l) are frequently used to represent the repeating structure of the lattice. Theorem 4.2 EIGENFUNCTION EXPANSIONS 103 A linear operator is said to be Hermitian. operator if and only if p (y1 y2 − y1 y2 ) and hence p(b) {y1 (b)y2 (b) − y1 (b)y2 (b)} − p(a) {y1 (a)y2 (a) − y1 (a)y2 (a)} = 0. In the study of the motion of electrons in a crystal lattice. It is clear from (4. (ii) p(a) = p(b) = 0. Conditions (iii).24) that L is a Hermitian. This leads to boundary conditions y(0) = y(a) = 0 – separated boundary conditions. Note that this doesn’t violate our deﬁnition of p as strictly nonzero on the open interval (a. We have therefore shown that our linear diﬀerential operator is Hermitian with respect to a pair of unmixed boundary conditions.25) is satisﬁed. (iii) α1 yi (a) + α2 yi (a) = 0 and β1 yi (b) + β2 yi (b) = 0.25) can occur. These conditions then have nontrivial solutions if and only if y1 (a)y2 (a) − y1 (a)y2 (a) = 0. with at least one of the αi and one of the βi nonzero. or selfadjoint. each of which involves y and y at a single endpoint. consider a string that is rotating (as in the example at the start of this chapter) or vibrating with its ends ﬁxed. if y1 . yi (a) = yi (b) and yi (a) = yi (b). y2 for all y1 and y2 . There are three diﬀerent ways in which (4. (4. 4.

φn (x).26) By making use of the orthogonality of the eigenfunctions. . y2 = 0. (4. we begin by ﬁnding the eigenvalues and eigenfunctions of L.26) with φn . and y1 . 4. . Ly2 = y1 . y2 = y1 . a result that we will prove once we have deﬁned the notion of a complex inner product. . and we cannot use this argument. after taking the inner product of (4. y2 = λ1 y1 .19) with f ∈ C[a.27) Next. λ2 y2 = λ2 y1 . φn . Unfortunately. y2 . we expand the solution of the boundary value problem in terms of the eigenfunctions. we expand f (x) in terms of these eigenfunctions. as ∞ f (x) = n=1 cn φn (x). Then Ly1 . . φn . We denote these eigenvalues by λ1 . we ﬁnd that the expansion coeﬃcients are cn = f. . .104 BOUNDARY VALUE PROBLEMS Proof Let y1 and y2 be eigenfunctions that correspond to the distinct eigenvalues λ1 and λ2 . b] is not ﬁnite dimensional. . If the space of functions C 2 [a. λ2 . Since λ1 = λ2 . and the eigenfunctions by φ1 (x). . φn (4. and y1 and y2 are orthogonal.28) . As we shall see in the next section. extended into a basis). all of the eigenvalues of a Hermitian linear operator are real. . φ2 (x). b] and unmixed boundary conditions. . C 2 [a.2. as ∞ y(x) = n=1 dn φn (x). . y1 . so that the Hermitian property Ly1 . b] were of ﬁnite dimension. we would now argue that the orthogonal eigenfunctions generated by a Hermitian operator are linearly independent and can be used as a basis (or in the case of repeated eigenvalues. . . . y2 = λ1 y1 . We will have to content ourselves with presenting a credible method for solving inhomogeneous boundary value problems based upon the ideas we have developed. and simply state a theorem that guarantees that the method will work in certain circumstances. Ly2 gives (λ1 − λ2 ) y1 .4 Eigenfunction Expansions In order to solve the inhomogeneous boundary value problem given by (4. (4. . λn . . y2 . y2 = 0. Next.

Example Consider the boundary value problem −y = f (x) subject to y(0) = y(π) = 0. the eigenfunctions are solutions of y + λy = 0 subject to y(0) = y(π) = 0. n2 n=1 ∞ = 2(−1)n+1 . . We therefore write ∞ (4. rather artiﬁcial. in detail in Chapter 5.29) is y(x) = In the case f (x) = x. and the solution of the inhomogeneous problem (4.1 on Bessel functions.29) directly. In this case.19) to obtain ∞ ∞ L n=1 dn φn (x) = n=1 cn φn (x).29) f (x) = n=1 cn sin nx.28) into (4.27) and (4. there is no solution if cm = 0 and an inﬁnite number of solutions if cm = 0. n3 n=1 ∞ π x sin nx dx 0 π sin2 nx dx 0 cn sin nx. cn = so that y(x) = 2 (−1)n+1 sin nx. n We will discuss the convergence of this type of series. known as a Fourier series. We have therefore constructed a solution of the boundary value problem with dn = cn /λn . as we saw in Section 4. which we already know to be λn = n2 .1. many boundary value problems for which this eigenfunction expansion method is the only way to proceed analytically. of course. such as the example given in Section 3. however. φn (x) = sin nx. This process will work correctly and give a unique solution provided that none of the eigenvalues λn is zero.4.28) converges and deﬁnes a function in C 2 [a. This example is. if the series (4.9. From the linearity of L and the deﬁnition of φn this becomes ∞ ∞ dn λn φn (x) = n=1 n=1 cn φn (x). and we could have integrated (4.2 EIGENFUNCTION EXPANSIONS 105 and substitute (4. There are. When λm = 0. b].

1] (as we shall show in Theorem 4. 2 2 − m(m + 1) m=2 ∞ Bm Pm (x). If u and y are bounded on [−1. 1]. then. we ﬁrst construct the eigenfunction solutions by solving Ly = λy. Having seen that this method works.4). These Legendre polynomials are orthogonal over [−1. The choice 2 − λ = n(n + 1). Since the Legendre polynomial. is the bounded solution of the homogeneous problem. the left hand side of this equation vanishes. 1]. subject to the condition that y should be bounded on [−1. after multiplying through by u and integrating over [−1.106 BOUNDARY VALUE PROBLEMS Example Consider the inhomogeneous equation (1 − x2 )y − 2xy + 2y = f (x) on −1 < x < 1. then . we ﬁnd that u(1 − x2 )y 1 −1 − u (1 − x2 )y 1 −1 1 = −1 u(x)f (x) dx. and then expand y(x) = we ﬁnd that {2 − m(m + 1)} Bm = Am for m The required solution is therefore y(x) = Am 1 A0 + B1 P1 (x) + Pm (x).30) with f ∈ C[−1. b] and subject to unmixed boundary conditions at both endpoints. 1]. Now. linear diﬀerential operator deﬁned on a closed interval [a. we can now state a theorem that gives the method a rigorous foundation. If we now write ∞ f (x) = m=0 Am Pm (x). we have 1 −1 P1 (x)f (x) dx = 0. and form a basis for C[−1. 0. 1]. If u(x) is a solution of the homogeneous problem. 1]. which has bounded solutions yn (x) = Pn (x). gives us Legendre’s equation of integer order.2 If L is a nonsingular. to solve the boundary value problem. We begin by noting that there is a solubility condition associated with this problem. (4. u = P1 (x) = x. with B1 an arbitrary constant. so 1 that −1 u(x)f (x) dx = 0. Theorem 4. with n a positive integer. ∞ m=0 where A1 = 0 by the solubility condition. which is (1 − x2 )y − 2xy + (2 − λ)y = 0.

After examining some solution techniques that are applicable to such equations in general. λ1 . and we will devote the rest of this chapter to a study of the properties of the solutions of such systems. 4. . we can construct a set of orthogonal solutions that can be used as the basis for a series expansion of the solution of the physical problem in question. is subject to the unmixed boundary conditions α1 y(a) + α2 y (a) = 0. b] in which y is continuous. which deﬁnes the eigenfunctions and eigenvalues. and that we can deduce many properties of such equations independent of the functional form of the coeﬃcients. (ii) The eigenfunctions that correspond to these eigenvalues form a basis for C[a. b] converges uniformly to y on any subinterval of [a. (4. linear diﬀerential operator. β1 y(b) + β2 y (b) = 0. namely the Fourier–Legendre and Fourier–Bessel series. For a self-adjoint. this is d dx p(x) dy dx + q(x)y = λy. b]. (4.3.3 Sturm–Liouville Systems In the ﬁrst three chapters. We will not prove this result here.31) which.4. since they frequently arise in models of physical systems in spherical and cylindrical geometries. Ly = λy.1 The Sturm–Liouville Equation Sturm–Liouville equations are of the form (p(x)y (x)) + q(x)y(x) = −λr(x)y(x). second order. This is an example of a Sturm–Liouville system. . which can be ordered so that |λ0 | < |λ1 | < · · · < |λn | < · · · and n→∞ lim |λn | = ∞. in each case. † For a proof see Ince (1956). (4. We saw that. in its simplest form.32) 2 2 2 2 with α1 + α2 > 0 and β1 + β2 > 0 to avoid a trivial condition. . and the series expansion relative to this basis of a piecewise continuous function y with piecewise continuous derivative on [a.† Instead. .33) . we studied the particular cases of Legendre’s equation and Bessel’s equation.3 STURM–LIOUVILLE SYSTEMS 107 (i) L has an inﬁnite sequence of real eigenvalues λ0 . 4. we have studied linear second order diﬀerential equations. we return to the equation. In this chapter we will see that Legendre’s and Bessel’s equations are examples of Sturm–Liouville equations.

Some Examples of Sturm–Liouville Equations Perhaps the simplest example of a Sturm–Liouville equation is Fourier’s equation. (a. Recall that Legendre’s equation is λ 2x dy d 2y + − y = 0. This can be written as ((1 − x2 )y ) = −λy. we know that this has solutions of the form √ √ Jν (x λ) and Yν (x λ). which is usually solved for 0 < x < a. the number λ is the eigenvalue. If λ = n(n + 1). λ) = −λy(x. We will also make some assumptions about the behaviour of the coeﬃcients of (4. is x2 y + xy + (λx2 − ν 2 )φ = 0.108 BOUNDARY VALUE PROBLEMS which can be written more concisely as Sy(x. We can write Legendre’s equation and Bessel’s equation as Sturm–Liouville problems.31).33). This can be rearranged into the form ν2 y = −λxy. (iii) p(x) > 0 and r(x) > 0.35) (4. are critically dependent upon the given boundary conditions.34) This is a slightly more general equation than (4. Bessel’s equation. It is often more important to know the properties of λ than it is to construct the actual solutions of (4. λ) = −λr(x)y(x. In (4. (4. from the results of Chapter 3. dx2 1 − x2 dx 1 − x2 and we are usually interested in solving this for −1 < x < 1. We discussed a physical problem that leads naturally to Fourier’s equation at the start of this chapter. b). q(x) and r(x) are real-valued and continuous.37) √ √ which has solutions cos(x λ) and sin(x λ). which may be complex. on an open interval. namely that (i) p(x). Similarly. and we will meet another at the beginning of Chapter 5.36) . b). (4. x Again.33). λ). whose possible values. of the real line. λ). where the diﬀerential operator S is deﬁned as Sφ ≡ d dx p(x) dφ dx + q(x)φ.33) for x ∈ (a. y (x. We seek to solve the Sturm–Liouville equation. we showed in Chapter 2 that this has solutions Pn (x) and Qn (x). (ii) p(x) is diﬀerentiable. (4. (xy ) − (4.33).

of the form α0 y(a. since p(x) = 1 − x2 = 0 when x = ±1. λ) = 0. In many problems. Bessel’s equation has regular endpoints for 0 < a < b < ∞. but singular endpoints if a = −1 or b = 1. This is because of the self-adjoint nature of the diﬀerential operator. x = a. Legendre’s equation has regular endpoints if −1 < a < b < 1. with regular endpoints.3 STURM–LIOUVILLE SYSTEMS 109 Although the Sturm–Liouville forms of these equations may look more cumbersome than the original forms. The endpoint x = a is a singular endpoint if a = −∞ or if a is ﬁnite but the conditions (4. and similarly if the other endpoint. we prescribe unmixed.36). b) is a regular endpoint if a is ﬁnite and the conditions (4. . is an eigenfunction. where the coeﬃcient functions satisfy the conditions (4. β0 2 2 2 2 and β1 are real. (4. to be solved subject to a separated boundary condition at each regular endpoint and a Friedrichs boundary condition at each singular endpoint. b). is singular. Fourier’s equation has regular endpoints if a and b are ﬁnite. we will see that they are very convenient for the analysis that follows. (ii) On an interval with one or two singular endpoints.36) do not hold on the closed interval [a. with α0 + α1 > 0 and β0 + β1 > 0. We can now deﬁne the Sturm–Liouville boundary value problem to be the Sturm–Liouville equation. λ) = 0. these are equivalent to Friedrichs boundary conditions. or separated. c] for each c ∈ (a. 4. but singular endpoints if a = 0 or b = ∞. since q(x) = −ν 2 /x is unbounded at x = 0. b). A nontrivial solution. b) there exists A ∈ R+ such that |y(x. β0 y(b. [a. λ) ≡ 0. (p(x)y (x)) + q(x)y(x) = −λr(x)y(x) for x ∈ (a. of the interval (a. α1 . b]. Similar deﬁnitions hold for the other endpoint. x = b. and λ is the corresponding eigenvalue. Note that this boundary value problem is homogeneous and therefore always has the trivial solution. λ)| A for all x ∈ (a. that for some c ∈ (a. For example. (i) On a ﬁnite interval. there exists B ∈ R+ such that |y(x.38) These boundary conditions are said to be real if the constants α0 .4. c] for some c ∈ (a. λ) + β1 y (b.36) hold on the closed interval [a. c]. y(x. x = b. λ)| B for all x ∈ [c. b). The endpoint. b). λ) + α1 y (a. We can now deﬁne the types of boundary condition that can be applied to a Sturm–Liouville equation. y = 0. boundary conditions.2 Boundary Conditions We begin with a couple of deﬁnitions.3. the boundary conditions that arise in models of physical problems are usually boundedness conditions.

In other words. 2 for some A ∈ R+ . 1. with corresponding eigenvalues λ = λn = n2 π 2 . appropriate boundary conditions are B for x ∈ 1 . in the form |y(x. 1 . 2. . λ) = 0. the eigenfunctions exist for all real. ∞ . λ2 . 1 2 Finally. . subject to the boundary conditions y(0. whilst x = 1 is a regular endpoint. Legendre’s equation is (1 − x2 )y (x. The eigenfunctions of this system are sin λn x for n = 1. which are appropriate √ since both endpoints are regular. λ)| B for x ∈ [0. for example with c = 0. let’s examine Bessel’s equation with ν = 1. . . x = 0 is a singular endpoint. The solution that is bounded at x = 0 is J1 (x λ). which we write as λ = λ2 . 1). 2. over the interval (0. (xy ) − y = −λxy. where J1 (λn ) = 0. with corresponding eigenvalues λ = λn = n(n + 1). λ)| A for x ∈ 0. Pn (x) for n = 0. for some A. λ)| 2 . we have a discrete spectrum. The eigenvalues are solutions of J1 ( λn ) = 0. 1). λ) for x ∈ (−1. The set of eigenvalues for a Sturm–Liouville system is often called the spectrum. whereas for the second there is a continuous spectrum. but now the eigenvalues lie on the half-line [0. λ) = −λy(x. . . λ) = −λy(x. 1). λ) = 0. |y(x. 0]. We will focus our attention on problems that have a discrete spectrum only. . Since both endpoints are now singular. y (x. We therefore apply Friedrichs boundary conditions. positive λ. Let’s now consider Bessel’s equation with ν = 1.110 BOUNDARY VALUE PROBLEMS Some Examples of Sturm–Liouville Boundary Value Problems Consider Fourier’s equation. and we know that the only eigenfunctions bounded at both the endpoints are the Legendre polynomials. . 1 . x Because of the form of q(x). The eigenfunctions are again J1 (x λ). Note that this is singular at both endpoints. In Chapter 3 we constructed the solutions of this equation using √ the method of Frobenius. . In the ﬁrst of the Bessel function examples above. |y(x. 2 √ for some A. In Chapter 2 we used the method of Frobenius to construct the solutions of Legendre’s equation. λ) for x ∈ (0. B ∈ R+ . ∞). λ)| A for x ∈ (−1. Suitable boundary conditions are therefore |y(x. since p(±1) = 0. . B ∈ R+ . . |y(x. y(1. 1). λ) = y(1. ∞). λ)| A for x ∈ 0. . but now for x ∈ (0.

φ2 + φ1 . Note that this reduces to the deﬁnition of a real inner product if φ1 and φ2 are real.3 Properties of the Eigenvalues and Eigenfunctions In order to further study the properties of the eigenfunctions and eigenvalues. α (4. φ1 ∗ . Proof If we substitute y1 (x) = y(x. λ)(y ∗ (x. λ). (4. By integrating by parts. S.40) If a is a ﬁnite singular endpoint and the functions y1 and y2 satisfy the Friedrichs boundary condition at a. so that we can take the limits α → a+ and β → b− when the endpoints are singular. b]. λ) b a = 0. y2 (x) ∈ C 2 [a. we have y ∗ (x. b). we begin by deﬁning the inner product of two complex-valued functions over an interval I to be φ1 (x). λ) into Green’s formula over the entire interval. φ2 = 0 with φ1 ≡ 0 and φ2 ≡ 0. β] ⊂ (a. Now if x = a is a regular endpoint and the functions y1 and y2 satisfy a separated boundary condition at a. This means that the inner product has the properties (i) (ii) (iii) (iv) φ1 . φ = I |φ|2 dx 0. φ2 = φ2 . b] be twice-diﬀerentiable complex-valued functions. λ) = p(x) y(x. Sy(x.35). y(x. Theorem 4. We can now derive several results concerning the eigenvalues and eigenfunctions of a Sturm–Liouville boundary value problem.3 STURM–LIOUVILLE SYSTEMS 111 4. φ. with equality if and only if φ(x) ≡ 0 in I.3 The eigenvalues of a Sturm–Liouville boundary value problem are real. If φ1 . φ3 . then ∗ ∗ p(a) y1 (a)(y2 (a)) − y1 (a)y2 (a) = 0. λ) and y2 (x) = y ∗ (x. φ1 + φ2 .4. a2 φ2 = a∗ a2 φ1 . y1 = p(x) y1 (x)(y2 (x)) − y1 (x)y2 (x) β . . x→a+ ∗ ∗ lim p(x) y1 (x)(y2 (x)) − y1 (x)y2 (x) = 0. The inner products are deﬁned over a subinterval [α. φ3 = φ1 .41) Similar results hold at x = b.1) ∗ ∗ y2 . φ2 . Sy1 − Sy2 .39) which is known as Green’s formula. a1 φ1 . φ2 + φ3 = φ1 . it is straightforward to show that (see Lemma 4. [a.3. we say that φ1 and φ2 are orthogonal. Let y1 (x). 1 where a superscript asterisk denotes the complex conjugate. λ). (4. φ3 + φ2 . and the Sturm–Liouville operator. 1 φ1 . λ)) − y (x. φ3 . is given by (4. φ2 (x) = I φ∗ (x)φ2 (x) dx. λ)y ∗ (x. λ) − Sy ∗ (x.

41). 1. when λ = 2n for n = 0. λ)y(x. By using the method of Frobenius. λ) y (x. we ﬁnd that b ˜ r(x)y(x. Now. b) with respect to the weighting function r(x). λ) ∗ ˜ − y (x. (a.33) and its complex conjugate. Since r(x) > 0 and y(x. Example Consider Hermite’s equation. we must have λ = λ∗ . ˜ ˜ y(x. and hence λ ∈ R. λ). λ) and y ∗ (x. so that.38).40) and (4. so that b ˜ r(x)y(x. If we evaluate this formula. notice that the separated boundary condition. that are bounded at inﬁnity for λ = 2n then . a ˜ so that the eigenfunctions associated with the distinct eigenvalues λ and λ are orthogonal with respect to the weighting function r(x). λ) dx = 0. we need ∗ y1 (a) (y2 (a)) − y1 (a)y2 (a) = 0. λ) and y(x. ˜ Theorem 4. For the pair of equations (4. H1 (x) = 2x and H2 (x) = 4x2 − 2. from Green’s formula.21).39). a (4. (4. λ)|2 dx = 0. λ)(λ − λ∗ ) dx = (λ − λ∗ ) a b r(x)|y(x.4 If y(x. λ) dx = 0. λ). The solutions of (4. the self-adjoint form of the equation. using the fact that the functions y(x. λ) are eigenfunctions of the Sturm–Liouville ˜ then these eigenfunctions are orthogonal over boundary value problem.20). with λ = λ. Sy(x. we can show that there are polynomial solutions. 2. at x = a takes the form α0 y1 (a) + α1 y1 (a) = 0.112 BOUNDARY VALUE PROBLEMS making use of (4. For example. (4.43) (4. we ﬁnd that b a r(x)y(x. (4.43)2 and (4. α0 y2 (a) + α1 y2 (a) = 0. . x = b. . This clearly shows that ˜ p(x) y(x. λ) = Sy(x. λ)y ∗ (x. λ) are solutions of (4. λ) is nontrivial. . Taking the complex conjugate of the second of these gives ∗ α0 y2 (a) + α1 (y2 (a)) = 0.42) Proof Firstly.44) to have a nontrivial solution. Hn (x). λ)y(x. H0 (x) = 1. λ) . y(x. λ) y(x.44) since α0 and α1 are real. λ) ∗ →0 as x → a and x → b. . ∗ (4. ∗ A similar result holds at the other endpoint.

from Theorem 4. λi ). r(x)y(x. (4. f (x) − i=0 Ai φi (x) . We will try to establish when we can write a piecewise continuous function f (x) in the form ∞ f (x) = i=0 ai φi (x). Approximation in the Mean and Completeness We can now deﬁne a sequence of orthonormal eigenfunctions φn (x) = which satisfy φn (x). Now N N ||RN ||2 = f (x) − i=0 Ai φi (x). We now try to minimize this error by minimizing its norm b N 2 ||RN || = RN (x).3 STURM–LIOUVILLE SYSTEMS 113 take the form un = e−x condition 2 /2 Hn (x).4. and.45) r(x)y(x. λi ). for some ﬁnite N . where δnm is the Kronecker delta.4 Bessel’s Inequality. (4. λn ) r(x)y(x. which is the mean square error in the approximation.3. where the Ai are to be determined so that this provides the most accurate approximation to f (x).46) Taking the inner product of both sides of this series with φj (x) shows that aj = f (x). The error in this approximation is N RN (x) = f (x) − i=0 Ai φ(x. λn ) . 4.47) using the orthonormality condition.46). λn ). N fN (x) = i=0 Ai φ(x. satisfy the orthogonality 2 ∞ −∞ e−x Hn (x)Hm (x) dx = 0 for n = m. (4. The quantities ai are known as the expansion coeﬃcients. or generalized Fourier coeﬃcients. RN (x) = 2 a f (x) − i=0 Ai φi (x) dx. φm (x) = δnm .45). we start by approximating f (x) by a ﬁnite sum. φj (x) .4. In order to motivate the inﬁnite series expansion (4. (4.

. We can now use the orthonormality of the eigenfunctions. Approximations by the method of least squares are often referred to as approximations in the mean. for a given orthonormal system. (4. The error is therefore smallest when Ai = ai for i = 0. (4. . φi (x) i = ||f (x)||2 + i=0 {−Ai ai − A∗ a∗ + A∗ Ai } i i i N = ||f (x)||2 + i=0 |Ai − ai |2 − |ai |2 . so the most accurate approximation is formed by simply truncating the series (4. . ∞ b |ai |2 = i=0 a |f (x)|2 dx. As the right hand side of this is independent of N . it follows that ∞ b |ai |2 i=0 a |f (x)|2 dx. For complete orthonormal systems. since the norm of RN (x) is positive.45). to obtain N ||RN (x)||2 = ||f (x)||2 − i=0 N N Ai f (x). In addition. i=0 N Ai φi (x) N − i=0 Ai φi (x). . N b |ai |2 i=0 a |f (x)|2 dx. . φ2 (x). . This shows that the sum of the squares of the expansion coeﬃcients converges. RN (x) → 0 as N → ∞. because of the way the error is minimized.47).46) after N terms. 1. . (4. . then the orthonormal system is said to be complete. which determines the coeﬃcients ai . i=0 Ai φi (x) . so that Bessel’s inequality becomes an equality. φi (x) − i=0 A∗ φi (x). If. f (x) + i i=0 N A∗ Ai φi (x).49) for every function f (x). N . φ1 (x). any piecewise continuous function can be approximated in the mean to any desired degree of accuracy by choosing N large enough.114 BOUNDARY VALUE PROBLEMS N = ||f (x)||2 − N f (x). . and the expression (4.48) which is Bessel’s inequality. f (x) + i=0 Ai φi (x).

and we can write ∞ 0 f (x) = i=1 ai Jν (λi x). the eigenvalues. . on the interval [0. then the limit and the integral can be interchanged.32).50) . (4.33). p(x)y (x) = R(x) cos θ(x). and we say that i=0 ai φi (x) converges in the mean to f (x). They satisfy the orthogonality relation a √ √ xJν ( µx)Jν ( λx) dx = 0. P1 (x). so that λ = λi for i = 1. If. although we could have used an analytical framework. we will establish that the nth eigenfunction has n zeros in the open interval (a. should be seen as a necessary condition for the validity of the expansion. λ2 a. The Legendre polynomials P0 (x).4. Jν (λ2 x). however. . . the Bessel functions Jν ( λx) satisfy the Sturm–Liouville equation. but. in other words that f (x) ∞ has an expansion in terms of the φi (x). b). . the series i=0 ai φi (x) is uniformly convergent. of For example. . .33). . √ if λ and µ are distinct eigenvalues. . . . the ∞ expansion is valid.5 Further Properties of Sturm–Liouville Systems We conclude this section by investigating some of the qualitative properties of solutions of the Sturm–Liouville system (4. and the expansions of Chapters 2 and 3 are special cases √ the more general results of this chapter.3 STURM–LIOUVILLE SYSTEMS 115 The completeness of orthonormal systems. 4. Using the regular endpoint condition Jν ( λa) = 0 and the singular endpoint√ condition at x = 0. with ai = consistent with (3. φ2 (x). we introduce the Pr¨ fer u substitution. . In particular. 2 a2 {Jν (λi a)} 2 0 a xJν (λi x)f (x) dx. . . (4. a] are both examples of complete orthogonal systems (they can easily be made orthonormal). 2. . .3. . The completeness of the system φ1 (x). . for an arbitrary function f (x). We will take a geometrical point of view in order to establish this result. q(x) = −ν 2 /x and r(x) = x. that is the zeros √ of Jν (x). does not necessarily imply that f (x) = i=0 ai φi (x). on the interval [−1. To achieve this. with p(x) = x. 1] and the Bessel functions Jν (λ1 x). can be written as λa = λ1 a. the question of convergence requires a more detailed investigation. . y(x) = R(x) sin θ(x). as expressed by b N →∞ N 2 lim f (x) − a i=0 ai φi (x) ∞ dx = 0.

which shows some typical solution curves for various values of λ. is however independent of R and is. (4.3. tan δ = δ π. . as sketched in Figure 4. for any initial condition θ(a) = γ.51)1 . (4. This is the set of small line segments of slope dθ/dx in the (θ. we obtain −cosec2 θ and hence dθ 1 = (q(x) + λr(x)) sin2 θ + cos2 θ. (4. by analogy with polar coordinates. x) plane. there is a . we can see that. γ and δ. for 2 ﬁxed x. dx p(x) If we now diﬀerentiate (4. some simple manipulation gives 1 dR = dx 2 1 − q(x) − λr(x) R sin 2θ. dx y y2 p 2 The Pr¨fer substitution has therefore changed our second order linear diﬀerential u equation into a system of two ﬁrst order nonlinear diﬀerential equations in R and θ over the interval a < x < b. As λ increases from −∞.52) subject to θ(a) = γ. . 2.116 BOUNDARY VALUE PROBLEMS The new dependent variables. If we now consider the separated boundary conditions α1 y(a) + α2 y (a) = 0. dθ/dx = q(x) + λp(x) at θ = n + 1 π.52) (py ) p (y ) dθ 1 = − = −λr − q − cot2 θ.52) by drawing its direction ﬁeld. 1. . β1 y(b) + β2 y (b) = 0. θ(b) is an increasing function of λ. . θ = tan−1 y py (4. and hence give the trivial solution for all x. since R = 0 at any point x = x0 would mean that y(x0 ) = y (x0 ) = 0. In addition.53). we call R the amplitude and θ the phase angle. The equation for θ. such that tan γ = y(a) α2 =− for 0 p(a)y (a) p(a)α1 β2 y(b) =− for 0 p(b)y (b) p(b)β1 γ π. we can deﬁne two phase angles. which. Note that dθ/dx = 1/p(x) at θ = nπ. are then deﬁned by R2 = y 2 + p2 (y )2 . relatively easy to analyze.52) and (4. Nontrivial eigenfunction solutions of the Sturm–Liouville equation have R > 0. as we shall see. increases with increasing λ.52). If we now write cot θ(x) = py /y and diﬀerentiate.51) and. R and θ. p(x) (4.54) We need to add this multiple of nπ because of the periodicity of the tangent function. We can infer the qualitative form of the solution of (4. and the eigenvalue problem that we have to solve is (4. From Figure 4.53) (4. θ(b) = δ + nπ for n = 0.4. which is independent of λ.

q(x) = −1/x and r(x) = x. Because of the way we have labelled these.3 STURM–LIOUVILLE SYSTEMS 117 ﬁrst value. in the interval γ θ δ + nπ. yn (x) = Rn (x) sin θ(x. Returning now to the example y + λy = 0 subject to y(0) = y(π) = 0. . We can also see that λ1 = 22 and the ﬁrst eigenfunction. yn (x).5 A regular Sturm–Liouville system has an inﬁnite sequence of real eigenvalues. . Theorem 4. at x = π . 2. with λn → ∞ as n → ∞. we have eigenvalues λn = n2 and eigenfunctions sin nx for n = 1. λ0 = 1 and y0 (x) = sin x is the zeroth eigenfunction. Fig. λ0 < λ1 < · · · < λn < · · · . y1 (x) = sin 2x. and so on. To see why this should be so. there is a sequence of values for which θ = δ + nπ for n = 1. . We can prove another useful result if we add an additional constraint. The direction ﬁeld (lines of slope dθ/dx) for (4. for which θ(b) = δ. We can formalize 2 this analysis as a theorem. . which corresponds to Bessel’s equation of order one. . nπ. 2π. Each of these is associated with an eigenfunction. If q(x) < 0. have n zeros in the interval a < x < b. .52) when p(x) = x. λn ). 2. λ = λ0 . which has no zeros in 0 < x < π. then all the eigenvalues are positive. consider the boundary value problem d dx p(x) dy dx + (λr(x) + q(x)) y = 0 subject to y(a) = y(b) = 0. As λ increases further. This has a zero when sin θ = 0 and. . . The corresponding eigenfunctions. . has one zero in 0 < x < π. there are precisely n zeros at θ = π. 4.3. .4. . . .

1968): (i) A system is completely speciﬁed by a state function. Typical solutions of (4.118 BOUNDARY VALUE PROBLEMS Fig. 4.4. This shows that b λ= a −q(x)y 2 + p(x) (y ) 2 b dx a r(x)y 2 dx . If we multiply through by y and integrate over [a. ψ(r. which corresponds to Bessel’s equation of order one. For example.3. b]. or wave function. with ψ. which is positive when p and r are positive and q is negative. 4. q(x) = −1/x and r(x) = x. Three important postulates of quantum mechanics are (see Schiﬀ. t). ψ = 1. using integration by parts. the theory that governs the behaviour of matter on very small length scales. we obtain b b b λ a r(x)y 2 dx = a −q(x)y 2 dx − a y d dx p(x) dy dx dx b = a −q(x)y 2 + p(x) (y ) 2 dx.6 Two Examples from Quantum Mechanics One of the areas of mathematical physics where Sturm–Liouville equations arise most often is quantum mechanics. if the system consists of a particle .52) for various values of λ when p(x) = x.

as classical physics would suggest.56) when V is independent of time.3 STURM–LIOUVILLE SYSTEMS 119 moving in an external potential. For example. i ∂ψ = Eψ. (ii) For every system there exists a certain Hermitian operator. ψ = 1. t)ψ. the energy must be one of the eigenvalues of the boundary value problem associated with (4. A. (4. The energy is said to be quantized. which governs the evolution of the wave function. called the Hamiltonian operator.57). i For a single particle of mass m moving in a potential ﬁeld V (r. Hermitian operator. we will ﬁnd that the energy of the system is not arbitrary. and u and ∇u continuous throughout the domain of solution. the operators that correspond to momentum and energy are −i ∇ and i ∂/∂t. ∂t where 2π ≈ 6. t). o Let’s look for a separable solution of (4. This is just the sum of the kinetic and potential energies. we quantize the classical result (4. then |ψ(r. the classical (as opposed to quantum mechanical) total energy is E = p2 /2m + V (r. Instead. Since i T = ET . T 2mu where E is the separation constant. and hence the eigenvalues are of great interest. and hence we need R3 |ψ(r)|2 d 3 r = ψ. − 2 . and ﬁnd that i 2 T =− ∇2 u + V (r) = E.57) 2m We seek solutions of this equation subject to the conditions that u should be ﬁnite. and arrive at 2 ∂ψ =− ∇2 ψ + V (r. and any measurement of the property gives one of the eigenvalues of A. t). H.55) by substituting the appropriate operators for momentum and energy. After we impose appropriate boundary conditions. o i ∇2 u + (V (r) − E) u = 0. (iii) To each observable property of the system there corresponds a linear. t)|2 d 3 r is the probability of ﬁnding the particle in a small volume d 3 r that surrounds the point r. such that ∂ψ = Hψ. ∂t and hence E is the energy of the particle.62 × 10−34 J s−1 is Planck’s constant.56) ∂t 2m This is Schr¨dinger’s equation.55) where p is the momentum of the particle. (4. To obtain the quantum mechanical analogue of this. The equation for u is then the timeindependent Schr¨dinger equation. (4. We write ψ = u(r)T (t).4.

What energies can the particle have? Since the probability of ﬁnding the particle outside the region 0 < x < a is zero.120 BOUNDARY VALUE PROBLEMS Example: A conﬁned particle Let’s consider the one-dimensional problem of a particle of mass m conﬁned in a region of zero potential by an inﬁnite potential at x = 0 and x = a.m sin mφ) . The steady state wave function. − Example: The hydrogen atom The hydrogen atom consists of an electron and a proton. (4. θ. with which we began this chapter. so that S − q2 n(n + 1) S+2 E+ r r S = 0 for r > 0.2). φ) in the form ψ = R(r)Y (s)Θ(φ). .m cos mφ + En. By continuity. and we have a solution in the form m ψ(r. with n m. 2. We must therefore solve the eigenvalue problem ∂2u = Eu. However. 1 − s2 Θ for some separation constant λ. (4. . where Dn. this is precisely the problem. we obtain Θ = Am cos mφ + Bm sin mφ. ψ(r).m are arbitrary constants and R satisﬁes the diﬀerential equation r2 R + 2rR − n(n + 1)R + 2r2 E + We can simplify matters by deﬁning S = rR. We conclude that the allowed energies of the particle are E = En = 2 λn /2m = 2 n2 π 2 /2ma2 .58) 2m ∂x2 subject to u = 0 at x = 0 and x = a. given by (4. then satisﬁes (4. then the equation for Y is the associated Legendre equation. we therefore have ψ = 0 at x = 0 and x = a.57) which. for periodicity. If λ = n(n + 1). with λ = 2mE/ 2 . after rescaling r and E to eliminate the constants. If we take Θ /Θ = −m2 with m = 1. where −q is the charge on the electron and q that on the proton. where s = cos θ. The bounded solution of this is Y = Cn Pn (s). Substituting this into (4. .60) q2 r R = 0. . Since the mass of the proton is much larger than that of the electron. (4. which we studied in m Chapter 2. let’s assume that the proton is at rest. θ. We can now look for separable solutions in spherical polar coordinates (r.59) 2 The potential V (r) = V (r) = −q 2 /r is the Coulomb potential due to the electrical attraction between the proton and the electron.m and En. φ) = R(r)Pn (s) (Dn. we must have ψ = 0 there.59) gives us r2 (1 − s2 )Y R R + 2r + 2r2 (E − V (r)) = − R R Y − 1 Θ = λ. we write as ∇2 ψ − 2V ψ = −2Eψ for |r| > 0.

of the fourth order equation y (4) + λy = 0 subject to y(0) = y (0) = y(π) = y (π) = 0 for 0 x π. B and C? Write the generalized Legendre equation. corresponds to a possible steady state for our model of the hydrogen atom. and is in agreement with the experimentally observed values. which will satisfy (4. Each energy level.EXERCISES 121 to be solved subject to the condition that S/r → 0 as r → ∞.1 Use the eigenfunction expansion method to ﬁnd a solution of the boundary value problem y (x) = −h(x) for 0 < x < 2π.6 4. However. . . What are 2 2 the corresponding eigenfunctions? Show that the equation dy d 2y + {λB(x) − C(x)} y = 0 + A(x) dx2 dx can be written in Sturm–Liouville form by deﬁning p(x) = exp What are q(x) and r(x) in terms of A. there is both a continuous and a discrete part to the spectrum. as the electron has too much energy. 2π]. Some thought and experimentation leads one to try a solution in the form S = rn+1 e−µr . . the allowable energy levels of the electron in the hydrogen atom. y (0) = y (2π).3 Comment on the diﬃculties that you face when trying to construct the Green’s function for the boundary value problem y (x) + y(x) = f (x) subject to y(a) = y (b) = 0. . (b) y (x) + ω 2 y(x) = f (x) subject to y(0) = y(π/2) = 0. 2. 2. This is an eigenvalue problem for E. 2 A reduction of order argument then shows that there are no other solutions with E < 0 that satisfy the condition at inﬁnity. subject to the boundary conditions y(0) = y(2π). the eigenvalues when (a) y(0) = y (π) = 0. . .7 as a Sturm–Liouville equation. nontrivial solution of (4. Exercises 4. this is not the end of the matter. Determine the eigenvalues.60) provided that µ = q 2 /(n + 1) and E = − 1 µ2 . 4. (b) y (0) = y(π) = 0.5 4. 4. are 2 2 n + 1 . n + 1 and n2 respectively. . with h ∈ C[0. 1. These states with positive energy can be shown to be unstable. (1 − x2 ) m2 d 2y dy + n(n + 1) − − 2x dx2 dx 1 − µ2 y = 0. In other words. 4. EN = −q 4 /2N 2 for N = 1. (c) y (0) = y (π) = 0. λ. A(x) dx . Find the Green’s function for the boundary value problems (a) y (x) = f (x) subject to y(−1) = y(1) = 0.4 Show that when y + λy = 0 for 0 < x < π. .2 4. .60). as it is possible to show that every positive value of E corresponds to a bounded. where n = 0.

. 1. subject to ˜ α1 y(a) + β1 y(b) + α1 y (a) + β1 y (b) = γ1 ˜ ˜ α2 y(a) + β2 y(b) + α2 y (a) + β2 y (b) = γ2 .12 4. ˜ is self-adjoint provided that b ˜ ˜ β1 β2 − β1 β2 = (α1 α2 − α1 α2 ) exp ˜ ˜ a A(x) dx .13 . These are known as the Laguerre polynomials. show that every solution of Airy’s equation.122 BOUNDARY VALUE PROBLEMS 4.10 Show that −(xy (x)) = λxy(x) is self-adjoint on the interval (0. there are polynomial eigenfunctions. Show that these eigenfunctions are orthogonal on the interval (0. 4. Prove that sin mx has at least one zero between each pair of consecutive zeros of sin nx.8 Consider the singular Sturm–Liouville system xe−x y + λe−x y = 0 for x > 0. Find the eigenvalues of the system consisting of Fourier’s equation and the conditions y(0) − y (0) = 0 and y(π) = 0. π). 2. b). . 4. with x = 0 a singular endpoint and x = 1 a regular endpoint with the condition y(1) = 0. vanishes inﬁnitely often on the positive x-axis.9 with y bounded as x → 0 and e−x y → 0 as x → ∞. Show that the boundary value problem y (x) + A(x)y (x) + B(x)y(x) = C(x) for all x ∈ (a.14). 1). ∗ Using the Sturm comparison theorem (see Exercise 1. when m > n. 4. Show that when λ = 0. . and at most once on the negative x-axis.11 4. y (x) − xy(x) = 0.

we note that the temperature in the bar satisﬁes the one-dimensional diﬀusion equation ∂T ∂2T = K 2 for 0 < x < L. the temperature of the bar is T = T0 . (1. the temperature will approach T1 throughout the bar. 0) = 1. it is convenient to deﬁne dimensionless variables. but that the uninsulated end is suddenly cooled to a temperature T1 < T0 .1) and (5. t ) = 0. In order to quantify this. and that. ˆ ˆ ∂ x2 ˆ ∂t (5. ˆ T0 − T1 L L In terms of these variables.2) become ˆ ˆ ∂T ∂2T = for 0 < x < 1. 0) = T0 . ∂t ∂x (5. as time progresses. and. T (x. T (0. The initial and boundary conditions are ∂T (L.4) ∂x ˆ As you can see.1) where x measures distance along the bar and K is its thermal diﬀusivity (see Section 2. t) = T1 . x= .6. It is usual to measure lengths in terms of metres. Suppose also that. The use of dimensionless variables has additional advantages. cylindrical metal bar of length L whose curved sides and one end are insulated from its surroundings. it .CHAPTER FIVE Fourier Series and the Fourier Transform In order to motivate our discussion of Fourier series.2) ∂x Before we solve this initial–boundary value problem. t= 2. thin. T (0. at ﬁrst sight. Let’s suppose that we have a long. t) = 0. we have managed to eliminate all of the physical constants from the problem. t ) = 0. we shall consider the solution of a diﬀusion problem. initially. T − T1 x ˆ Kt ˆ T = . (5. How does the temperature in the metal bar vary for t > 0? Physical intuition suggests that heat will ﬂow out of the end of the bar. which makes it essential to use them in studying most mathematical models of real physical problems. (5.3) ˆ ∂T ˆ x ˆ ˆ ˆ T (ˆ. Consider the length of the metal bar in the diﬀusion problem that we are studying. (5. by choosing appropriate dimensionless variables.1).

Similarly. and we will defer any further discussion of them until Chapter 11. the solution when t the small time solution. We can now continue with our study of the diﬀusion of heat in a metal bar by ˆ x ˆ ˆ seeking a separable solution. we can state that a variable is small or large in a meaningful way. which we deﬁned above. The same argument applies to all of the dimensional. t ) = 0 means that X(0) = 0. ˆ . and leads to a bar that lies between x = 0 and x = 1. the separation constant. (4. In eﬀect. For any given ˆ ˆ problem. On substituting this into (5. . No dimensionless groups appear in the simple diﬀusion problem given by (5. . X F 2ˆ ˙ The equation F = −k 2 F has solution F = ae−k t . 2 There is therefore a countably-inﬁnite sequence of solutions. which can themselves be said to be large or small in a meaningful way.3) and (5.4). . whilst problems in molecular dynamics may be made dimensionless using the mean atomic separation in a hydrogen atom. This is the most sensible length with which to make the problem dimensionless. Fourier’s equation. For example. the only length that is deﬁned in the problem is the length of the bar itself. 2. . t = Kt/L2 . L. whilst a bar of length 10−2 m = 1 cm is short. Is this a reasonable thing to do? In fact. For example.124 FOURIER SERIES AND THE FOURIER TRANSFORM seems reasonable to say that a metal bar with length 100 m is long. the condition that there should be no ﬂux of heat through the bar at x = 1 leads ˆ to kA cos k = 0. using the ˆ 1. it is essential to choose a length scale that is relevant to the physics of the problem itself. which we studied in Chapter 4.37). whilst X + k 2 X = 0. The other advantage of using dimensionless variables is that any physical constants that remain explicitly in the dimensionless problem appear in dimensionless groups. x x ˆ ˆ The condition T (0. and hence that B = 0. T (ˆ. really does represent the behaviour of the temperature in the metal bar when diﬀusion has had little eﬀect on the initial state. that characterizes the full length of the bar. t ) = X(ˆ)F (t ). not some basically arbitrary length. physical quantities that are used in a mathematical model. 1 ˆ Tn = An exp −π 2 n + 2 2 1 ˆ t sin π n + 2 x . problems in celestial mechanics are often best made dimensionless using the mean distance from the Earth to the Sun. has solution X = A sin kˆ + B cos kˆ. and hence shows that k= π + nπ for n = 0. By dividing each variable by a suitable constant with the same dimensions.3) x we obtain ˙ F X = = −k 2 . These length scales are enormously diﬀerent from 1 m. 1. ˆ dimensionless time. we choose a bar of length 1 m as a reference relative to which we measure the length of our actual bar. such as 1 m. over the length scale.

which shows that ∞ An sin π n + n=0 1 2 x ˆ = 1. ˆ (5. as t → ∞. After noting that 1 sin π n + 0 1 1 2 x sin π m + ˆ 1 2 x ˆ dˆ x = 0 1 [cos {π (m − n) x} − cos {π (m + n + 1) x}] dˆ = 0 for m = n.5)*x(:)). ˆ (5. and is shown in Figure 5. 0) = 1. 2 2 (5. Note that the function ceil rounds its argument upwards to the nearest integer (the ceiling function). as expected. ˆ x The only information we have not used is the initial condition.9) This is a Fourier series solution. We produced Figure 5. t ) = 4 1 exp −π 2 n + (2n + 1)π 2 n=0 ∞ 2 1 ˆ t sin π n + 2 x .6) x and integrate over ˆ We now multiply this expression through by sin π m + the length of the bar. and hence T → T1 . (2n + 1)π we conclude that 1 An = 2 0 sin π n + x ˆ dˆ = x and hence that ˆ x ˆ T (ˆ.FOURIER SERIES AND THE FOURIER TRANSFORM 125 Since this is a linear problem.1 at various times.7) and 1 sin2 π n + 0 x ˆ dˆ = x 0 1 1 [1 − cos {π (2m + 1) x}] dˆ = ˆ x for m = n. the general solution is ∞ ˆ x ˆ T (ˆ. ˆ ˆ It is clear that T → 0.5).t) acc = 10^-8.5) We are now left with the task of determining the constants An .^2*t)/pi. end heat = X*a’. for k = 0:n X(:.1 by plotting the MATLAB function $' function heat = heat(x./(2*N+1). a = 4*exp(-pi^2*(N+0. and that the . n=ceil(sqrt(-log(acc)/t)-0.5).k+1) = sin(pi*(k+0. N = 0:n. t ) = n=0 An exp −π 2 n + 1 2 2 1 ˆ t sin π n + 2 x .8) 1 2 4 . & % This adds enough terms of the series that exp{−π 2 (n + 1 )2 t} is less than some 2 small number (acc = 10−8 ) in the ﬁnal term. T (ˆ. 1 2 (5. ˆ ˆ x 2 1 2 1 (5. and the Fourier series is just another example of a series expansion in terms of an orthogonal basis.4). a number of questions arise. ˆ which were ﬁrst solved by Fourier himself in the 1820s. should come as no surprise.9). However. since Fourier’s equation is a Sturm–Liouville equation. Under what conditions does a series of the form ∞ An sin π n + n=0 1 2 x ˆ . given by (5. similar to the Fourier–Legendre and Fourier–Bessel series. that led to the development ˆ of the theory of Fourier series. heat ﬂows out of the cool end of the bar. It is precisely problems of this sort. as given by (5.1. at various times.126 FOURIER SERIES AND THE FOURIER TRANSFORM Fig. x = 0. 3 5 7 9 11 2 2 That the basis functions are orthogonal. The solution of the initial–boundary value problem.9) with t = 0 and x = 1/2 and 1 ˆ leads to the interesting results 1− 1 π 1 1 1 1 + − + − + ··· = . 3 5 7 9 11 4 1+ 1 1 1 1 1 π − − + + − ··· = √ . (5. Note also that (5.8).3) and (5. We have developed the Fourier series solution of this problem without too much attention to mathematical rigour. 5.7) and (5. . using the method described above. 5. if f is an even function of t. Fourier series can arise as solutions of diﬀerential equations.12) with complex Fourier coeﬃcients Cn = 1 1 (An − iBn ) = 2 T 1 2T −1T 2 e−2πint/T f (t) dt. complex form as ∞ f (t) = n=−∞ Cn e2πint/T . f (t) = f (t ± 2nπ) for n = 1. but they are also useful for representing periodic functions in general. (5. (5.1 General Fourier Series The most general form for a Fourier series representation of a function. Using (5. with period T is f (t) = 1 A0 + An cos 2 n=1 ∞ 2πnt T + Bn sin 2πnt T . a Fourier sine series. Equations (5. This means that the resulting Fourier series is a sum of just the odd functions sin(2πnt/T ). An = 0. (5. 2.10) and (5. the resulting expansion is a Fourier cosine series.11). For example. which is piecewise continuous. however. Bn = T 1 2T −1T 2 sin 2πnt T f (t) dt. such as the Fourier–Legendre and Fourier–Bessel series.11) can also be written in a more compact.14) The partial sums of this series. Similarly.13) As we have seen. consider the function of period 2π. The technical details are. rather more straightforward for the Fourier series. n n=1 N .5. we conclude that f (t) = 1 2 1 (−1)n+1 sin nt = 2 sin t − sin 2t + sin 3t − · · · n 2 3 n=1 ∞ . since cos(2πnt/T ) is an even function of t.10) and (5. (5. (5. f (t). deﬁned by f (t) = t for −π < t < π. An and Bn . fN (t) = 2 (−1)n+1 sin nt.11) Note that if f is an odd function of t. . and is this convergence uniform? Does every function have a convergent Fourier series representation? These questions can also be asked of expansions in terms of other sequences of orthogonal functions.10) and. .1 GENERAL FOURIER SERIES 127 converge. are given by 2 An = T 1 2T −1T 2 cos 2πnt T 2 f (t) dt. . the Fourier coeﬃcients. (remember that f (t) has period 2π. (5.2. fN . both of which are crucial in what follows. 5.2 for various N . but the size of the region where this occurs decreases – a sure sign of nonuniform convergence.14) gives the value zero. 25 and 35. and therefore f → π as t → π − . Before we proceed. but f → −π as t → π + ) the series (5. Lemma 5. This nonuniform. In the neighbourhood of t = ±π. t = ±π. The function being approximated is shown as a dotted line.14). and f (t) appears not to become smaller as N increases. Fig. the diﬀerence between the partial sums. 15. for N = 5. oscillatory behaviour close to discontinuities is known as Gibbs’ phenomenon. the mean of the two limits as t approaches π from above and below. but that the convergence is not uniform. An inspection of this ﬁgure suggests that the series does indeed converge. we need to construct the Dirichlet kernel and prove the Riemann–Lebesgue lemma.128 FOURIER SERIES AND THE FOURIER TRANSFORM are shown in Figure 5. At the points of discontinuity.1 (The Dirichlet kernel) Let f be a bounded function of period T . The partial sums of the Fourier series for t. which we also met brieﬂy in Chapter 3. † See Appendix 2. The Dirichlet kernel. (5.15) be the nth partial sum of the Fourier series of f (t). sin( πt ) (5.3 for T = 2π. t) = m=−n Cm e2πimt/T (5.17) Fig.1 GENERAL FOURIER SERIES 129 with f ∈ P C[a. Dn for n = 10. If we deﬁne the Dirichlet kernel. 5.16) Dn (t) = T  2n + 1 when t = 0. t) = 1 T 1 2T −1T 2 f (τ )Dn (t − τ ) dτ.5. then Sn (f. . b]†. 20.13). to be a function of period T given for − 1 T < t < 1 T by 2 2   sin{(2n+1) πt } T when t = 0. which is illustrated in Figure 5. with Cm given by (5. Dn (t).3. 30 and 40 and T = 2π. and let n Sn (f. the sequence {Dn (t)} is closely related to the delta function. we obtain d−π/λ I(λ) = − c−π/λ f τ+ π sin λτ dτ. . the Dirichlet kernel becomes more and more concentrated around the origin. d] be a subinterval of [a. Lemma 5.13) and (5. As we will see in the next section. t) → f (t) as n → ∞. What we need to show is that. n Dn (t) = m=−n e2πint/T . as n → ∞.15). and hence that Sn (f. putting m = r − n gives the simple geometric progression 2n Dn (t) = e−2πint/T r=0 e2πirt/T = e−2πint/T sin (2n + 1) πt e4πi(n+1/2)t/T − 1 T = e2πit/T − 1 sin πt T and the result is proved. Lemma 5. as λ increases.18) The argument for f (t) cos λt is identical. When t = 0. b] → R. Let [c.17). Clearly Dn (0) = 2n + 1.130 FOURIER SERIES AND THE FOURIER TRANSFORM Proof From (5. as n increases. and the contributions from the positive and negative parts of the integrand cancel out. As we can see in Figure 5.1 tells us that the nth partial sum of a Fourier series can be written as (5. Deﬁne d I(λ) = c f (t) sin λt dt.19) † As we shall see. (5. b] on which f is continuous. t) = m=−n 1 2T 1 T 1 2T −1T 2 f (τ )e−2πinτ /T dτ 1 2T e2πint/T = where 1 T n −1T 2 f (τ ) m=−n e2πin(t−τ )/T dτ = 1 T −1T 2 f (τ )Dn (t − τ ) dτ.17). λ (5. the period of oscillation of the trigonometric function becomes smaller.2 (The Riemann–Lebesgue lemma) If f : [a. f ∈ P C[a. b] and f is bounded. a simple integral that involves the underlying function and the Dirichlet kernel†. the Dirichlet kernel just picks out the value f (τ ) at τ = t in (5. then b λ→∞ b lim f (t) sin λt dt = lim a λ→∞ f (t) cos λt dt = 0. a Proof This seems intuitively obvious. n Sn (f. since. By making the substitution t = τ + π/λ. this is actually a convolution integral.3. 2). Theorem 5. |I(λ)| Kπ 1 + λ 2 d−π/λ f (t) − f t + c π λ dt.18) and (5. at points where f (t) is continuous. it is also uniformly continuous there and. then the right hand side of (5. 2 2 Note that.1. for h small enough that f is continuous when t0 − h t < t0 .1 (The Fourier theorem) If f and f are bounded functions of period T with f . and hence that I(λ) → 0 as λ → ∞. b].20) Now. Similarly. we can ﬁnd a constant λ0 such that π ∀ λ > λ0 . f has well-deﬁned left and right derivatives f− (t0 ) = lim h→0 f− − f (t0 − h) f (t0 + h) − f+ . since f is continuous on the closed interval [c. since f is piecewise continuous. 2 2 f → f− as t → t− . then. with Cn given by (5. which we know exists by Theorem A2. Applying this result to all of the subintervals of [a. Let K be the maximum value of |f | on [c. f ∈ P C[a. given any constant . 0 0 since f is piecewise continuous. f+ (t0 ) = lim .12). < f (t) − f t + λ d − c − π/λ Since we can also choose λ0 so that Kπ/λ < /2 ∀ λ > λ0 .10) converges pointwise to f (t).20) shows that |I(λ)| < . f → f+ as t → t+ . b] on which f is continuous completes the proof. with f− = f+ if f is continuous at t = t0 . d]. remembering that | sin λt| 1.1 GENERAL FOURIER SERIES 131 Adding (5. (5. Proof At any point t = t0 ∈ (− 1 T. 2 2 τ →− 1 T + 2 lim f (τ ) + τ→ 1 T − 2 lim f (τ ) for t = − 1 T or 1 T . 1 T ). h→0 h h with f− (t0 ) = f+ (t0 ) if f is continuous at t = t0 . By the mean value theorem (Theorem A2. . converges pointwise to 1 2 1 2 τ →t− lim f (τ ) + lim f (τ ) + τ →t for − 1 T < t < 1 T . If we also assume that λ is large enough that π/λ d − c. (5.19) gives c 2I(λ) = − c−π/λ f t+ d−π/λ π sin λt dt + λ π λ d f (t) sin λt dt d−π/λ + c f (t) − f t + sin λt dt.5. (5. f− − f (t0 − h) = f (c)h for some t0 − h < c < t0 .13). d]. 2 After using a similar argument for t > t0 . but note that this is consistent with our earlier discussion of Gibbs’ phenomenon. t) − where −1T 2 1 {f (t + τ ) + f (t − τ )} D(τ ) dτ . τ ) is bounded and piecewise continuous.22) that Sn → 1 (f− + f+ ) as 2 n → ∞ and the result is proved. for suﬃciently small τ . and we conclude from (5. using the facts that Dn (t) = Dn (−t). 2 1 2T 1 1 (f− + f+ ) = 2 T −1T 2 g(t.21) for all h > 0 such that f is continuous when t0 − h < t0 < t0 + h. 2| sin(πτ /T )| which is bounded as τ → 0. 2 2 Theorem 5. Now. we arrive at |f− − f (t0 − h)| |f− − f (t0 − h)| + |f (t0 + h) − f+ | Mh (5. piecewise continuous.17) we make the change of variable τ = t + τ and. τ ) sin (2n + 1) πτ T dτ . t) = 1 2T 1 T 1 2T −1T 2 f (t + τ )D(τ ) dτ 1 = T and hence that Sn (f. periodic functions have a convergent Fourier series expansion. in (5.12).22) g(t. in any closed subinterval of [− 1 T. (5. By exploiting the periodicity of f . converges uniformly to f (t) if and only if f (t) is continuous there. 2 2 the right hand side of (5.132 FOURIER SERIES AND THE FOURIER TRANSFORM Since f is bounded. τ )| M |τ | .2 (Uniform convergence of Fourier series) For a function f (t) that satisﬁes the conditions of Theorem 5. However.1. We will not give a proof. 2 sin(πτ /T ) Considered as a function of τ .21) shows that |g(t. returning to the question of whether all bounded. The function g therefore satisﬁes the conditions of the Riemann–Lebesgue lemma. g(t. 1 T ]. (5. except possibly at τ = 0. τ ) = f (t + τ ) − f+ + f (t − τ ) − f− . there exists some M such that 1 M h. and f and Dn have period T . we note that this . we ﬁnd that Sn (f. the same proof works for t = − 1 T or 1 T with minor modiﬁcations.13). with Cn given by (5. Finally. 5. the Fourier transform produces a function of a continuous variable whose value indicates the frequency content of the original .1). We will prove this result rigorously later. for example in models of ﬁnancial derivatives. If we now let kn = 2πn/T and ∆kn = kn − kn−1 = 2π/T . kn becomes a continuous variable and the summation becomes an integral.2 THE FOURIER TRANSFORM 133 diﬃcult question was not ﬁnally resolved until 1964.13). f (t) = 1 2π ∞ −∞ ˜ e−ikt f (k) dk. the rational numbers).24) we immediately have an inversion formula.12) and (5. we have 1 f (t) = 2π ∞ ∆kn e n=−∞ ikn t 1 2T −1T 2 e−ikn t f (t ) dt . T .5.12) and (5. This is a minor. (5. detail. and are clearly rather peculiar (see. o and references therein.25) Note that some texts alter the deﬁnition of the Fourier transform (and its inverse) to take account of the factor of 1/2π in a diﬀerent √ for example with the transform way.23) which is known as the Fourier integral. but irritating. f (t). as we have seen in Theorem 5. If we now momentarily abandon mathematical rigour. However. For further details. consider (5. If we now deﬁne the Fourier transform of f (t) as ˜ F[f (t)] = f (k) = ∞ −∞ eikt f (t) dt. which are widely applicable to real world problems. In the same way as the Fourier series expansion of a periodic function decomposes the function into its constituent harmonic parts. so that we obtain f (t) = 1 2π ∞ −∞ eikt ∞ −∞ e−ikt f (t ) dt dk. see K¨rner (1988). Figure A2. such functions do arise in the theory of Brownian motion and stochastic processes. as T → ∞.13) gives f (t) = 1 T ∞ e2πint/T n=−∞ 1 2T −1T 2 e−2πint /T f (t ) dt . when Carleson showed that such functions exist whose Fourier series expansions diverge at a ﬁnite or countablyinﬁnite set of points (for example. (5. The Fourier transform maps a function of t to a function of k. which does not aﬀect the basic ideas.1. (5. even though they are continuous.2 The Fourier Transform In order to motivate the deﬁnition of the Fourier transform. which deﬁne the complex form of the Fourier series expansion of a periodic function. What happens as the period. Of course. and its inverse each multiplied by a factor of 1/ 2π. for example. these functions cannot possess a bounded piecewise continuous derivative. tends to inﬁnity? Combining (5. The advantage of this is that every generalized function has a Fourier transform and an inverse Fourier transform. −∞ Two regular sequences of good functions are equivalent if. We will be more concerned here with the use of the Fourier transform to solve partial diﬀerential equations. we need to give (5. g(x). If f (x) is an ordinary function such that (1 + x2 )−N f (x) is integrable from −∞ to ∞ for some N . 5. and two generalized functions are equal if their deﬁning sequences are equivalent.26) exists. the classic.134 FOURIER SERIES AND THE FOURIER TRANSFORM function. What restrictions must we place on the function f (t) for its Fourier transform to exist? This is a diﬃcult question. for which the reader is referred to Lighthill (1958). and very accessible.25) a ﬁrmer mathematical basis. is said to be regular if. for any good function F (x). e−x and sech2 x are good functions. For example fn (x) = G(x)/n is a regular sequence for any good function G(x). Before we can proceed. the regular sequences {G(x)/n} and {G(x)/n2 } are equivalent. A generalized function. ∞ n→∞ lim −∞ fn (x)F (x) dx (5. for example in electrical engineering and seismology. with ∞ ∞ f (x)F (x) dx = lim −∞ n→∞ −∞ fn (x)F (x) dx. for any good function F (x).24) and (5. For example. the limit (5. is a function in C ∞ (R) that decays rapidly enough that g(x) and all of its derivatives tend to zero faster 2 than |x|−N as x → ±∞ for all N > 0. is a regular sequence of good functions. A good function. f (x). We will not go into great detail. and our approach is to treat f (t) as a generalized function. Generalized functions are therefore only deﬁned in terms of their action on integrals of good functions. with ∞ n→∞ lim −∞ fn (x)F (x) dx = lim 1 n→∞ n ∞ G(x)F (x) dx = 0.1 Generalized Functions We begin with some deﬁnitions. ∞ n→∞ ∞ lim −∞ fn (x)F (x) dx = f (x)F (x) dx.26) exists and is the same for each sequence. and that the ordinary functions in whose Fourier transforms we are interested form a subset of the generalized functions. This has led to the widespread use of the Fourier transform to analyze the form of time-varying signals. for any good function F (x). {fn (x)}. −∞ . for any good function F (x). introduction to the subject.2. For example. then the generalized function f (x) equivalent to the ordinary function is deﬁned as any sequence of good functions {fn (x)} such that. A sequence of good functions. −∞ Then sgn(x) can be identiﬁed with the ordinary function sgn(x) = −1 for x < 0. We can see from † Since generalized functions are only deﬁned through their action on integrals of good functions.2 THE FOURIER TRANSFORM 135 For example. ∞ −∞ ∞ In fact. ∞ ∞ I(x)F (x) dx = −∞ −∞ F (x) dx. . sgn(x) = 2H(x) − I(x). is deﬁned so that for any good function F (x). ∞ ∞ 0 sgn(x)F (x) dx = −∞ 0 F (x) dx − F (x) dx. is deﬁned so that for any good function F (x). The unit function is the generalized function equivalent to the ordinary function f (x) = 1. I(x). — The Heaviside function. — The unit function. — The sign function. The zero generalized function is very simple. the generalized function equivalent to zero can be represented by either of the sequences {G(x)/n} and {G(x)/n2 }. 1 for x > 0. The generalized function H(x) is equivalent to the ordinary step function† H(x) = 0 for x < 0. 1 for x > 0. H(x). is deﬁned so that for any good function F (x). is deﬁned so that for any good function F (x). δ(x). 2 A useful sequence of good functions that deﬁnes the unit function is {e−x /4n }. −∞ using the deﬁnition of the Heaviside and unit functions. ∞ ∞ H(x)F (x) dx = −∞ 0 F (x) dx. −∞ No ordinary function can be equivalent to the delta function. sgn(x). ∞ δ(x)F (x) dx = F (0).5. since we can note that ∞ −∞ {2H(x) − I(x)} g(x) dx = 2 ∞ H(x)g(x) dx − ∞ I(x)g(x) dx −∞ 0 =2 0 g(x) dx − ∞ g(x) dx = −∞ 0 g(x) dx − g(x) dx. the value of H at x = 0 does not have any signiﬁcance in this context. Let’s consider some more useful generalized functions. — The Dirac delta function. This is just the deﬁnition of the function sgn(x). 2.4. which is equivalent to δ(x).27). However.136 FOURIER SERIES AND THE FOURIER TRANSFORM (5. satisﬁes T T 1 2T F (0) = lim n→∞ −1T 2 fn (x)F (x) dx. Since we can integrate by parts using any . is a useful way of deﬁning δ(x). Fig.1 that a sequence based on the Dirichlet kernel.2 Derivatives of Generalized Functions Derivatives of generalized functions are deﬁned by the derivative of any of the equivalent sequences of good functions. the function becomes more and more concentrated about the origin as n increases. the sequence fn (x) = n π 1/2 e−nx . As we can see in Figure 5.4 for various n. The sequence (5.3. 5. 5. 2 (5.17) and Theorem 5. this only works on a ﬁnite domain. On an inﬁnite domain. fn (x) = Dn (x)/2π = sin (2n + 1) πt /2π sin πt . eﬀectively plucking the value of F (0) out of the integral.27) which is illustrated in Figure 5. 5.28) . which is all we need to know. dx dx since xδ(x) = 0. consider the Fourier transform of the function e−|x| . This does not allow us to represent δ (x) in terms of other generalized functions. Now consider the derivative of the modulus function. formally.2 THE FOURIER TRANSFORM 137 member of the sequence. For example. The proof is straightforward. we can also. the Fourier transform of a function f (x) is deﬁned as F[f (x)] = ∞ −∞ eikx f (x) dx. −∞ and ∞ ∞ f (0)δ(x)F (x) dx = −∞ −∞ δ(x)[f (0)F (x)]dx = f (0)F (0). which is valid because it works for the equivalent sequence of good functions. since ∞ f (x)δ(x)F (x) dx = f (0)F (0). d d d d |x| = {x sgn(x)} = x {sgn(x)} + sgn(x) (x).2. using the deﬁnition. Using the product rule.5. treating it as if it were zero at inﬁnity. ∞ Another useful result is f (x)δ(x) = f (0)δ(x). We can deﬁne the modulus function in terms of the function sgn(x) through |x| = x sgn(x). but does show us how it acts in an integral.3 Fourier Transforms of Generalized Functions As we have seen. 1 + ik −1 + ik 1 + k2 (5. We can also show that H (x) = δ(x). which. dx dx dx dx We can now use the fact that sgn(x) = 2H(x) − I(x) to show that d d |x| = x {2H(x) − I(x)} + sgn(x) = 2xδ(x) + sgn(x) = sgn(x). since ∞ −∞ H (x)F (x) dx = − ∞ −∞ H(x)F (x) dx = − 0 ∞ F (x) dx = − [F (x)]0 = F (0). For example ∞ −∞ δ (x)F (x) dx = − ∞ −∞ δ(x)F (x) dx = −F (0). do so for the equivalent generalized function. is F e−|x| = 0 ex eikx dx + −∞ 0 ∞ e−x eikx dx = 1 1 2 − = . Since {e−x /4n } is a sequence equivalent to the unit function.3 The Fourier transform of a good function is a good function. Lemma 5.2. Making the change of variable y = ax gives F[f (ax)] = 1 a ∞ −∞ eiky/a f (y) dy. note that from the deﬁnition F[f (ax)] = ∞ −∞ 2 eikx f (ax) dx. ˜ which is equal to f (k/a)/a as required. so that F[αf + βg] = αF[f ] + βF[g]. c = cI(x). we conclude that π F[c] = 2πcδ(k). . and we can deal with the Fourier transform of an equivalent sequence instead.4 The Inverse Fourier Transform If we can show that (5. As an example of how this result can be used.25) holds for all good functions. so F[e−a|x| ] = 2 1 a1+ k 2 = a 2a .138 FOURIER SERIES AND THE FOURIER TRANSFORM We would also like to know the Fourier transform of a constant.10 for an alternative method).5). (5. the Fourier transformation is clearly a linear transformation. By writing z = x − 2ikn and deforming the contour of integration in the complex plane (see Exercise 5. ˜ ˜ Another useful result is that. F[f (ax)] = f (k/a)/a for a > 0. for example F[ce−x 2 /4n ∞ ]=c −∞ eikx−x 2 /4n dx = ce−nk 2 ∞ −∞ e−(x/2 √ √ 2 n−ik n) dx. c. we know that F[e−|x| ] = 2/(1 + k 2 ). it follows that it holds for all generalized functions. a2 + k 2 Finally. we ﬁnd that F[ce−x 2 /4n ] = ce−nk 2 ∞ −∞ e−z 2 /4n dz = 2πc n −nk2 e . treated as a generalized function. To prove this. We begin with a useful lemma. and 2 { n e−nk } is a sequence equivalent to the delta function. Instead we note that. However.29) This means that linear combinations of functions can be transformed separately. 5. it is not clear whether F[c] = c ∞ −∞ eikx dx is a well-deﬁned integral. if F[f (x)] = f (k). π using (3. we can exchange the order of integration and arrive at F e− ∞ x2 ˜ f (−x) = ∞ −∞ ∞ ∞ f (t) −∞ −∞ ei(k−t)x− x2 dx dt 2 = −∞ f (t)e−(k−t) 2 /4 exp − 1/2 x−i (k − t) 2 dx dt. by making the change of variable x = x + i(k − t)/2 3/2 . Theorem 5. Since f is a good function. dxN ˜ All derivatives of f therefore decay at least as fast as |k|−N as |k| → ∞ for any ˜ N > 0. we note that for F e− x2 1 2π ∞ −∞ ˜ e−ikx f (k) dk.2 THE FOURIER TRANSFORM 139 Proof If f (x) is a good function. checking that this ¯ change of contour is possible in the complex x-plane. > 0. If we diﬀerentiate p times and integrate by parts N times we ﬁnd that ˜ f (p) (k) (−1)N (ik)N 1 |k|N ∞ −∞ ∞ −∞ eikx dN {(ix)p f (x)} dx dxN dN {xp f (x)} dx. ∞ −∞ ˜ f (−x) = eikx− x2 ∞ −∞ e−ixt f (t) dt dx.5. Now. its Fourier transform clearly exists and is given by ˜ F[f ] = f (k) = ∞ −∞ eikx f (x) dx. and hence f is a good function. we ﬁnd that ∞ −∞ exp − 1/2 x−i (k − t) 2 2 ∞ dx = −∞ e− x2 ¯ d¯ = x π . This means that F e− x2 ˜ f (−x) = π ∞ −∞ e−(k−t) 2 /4 f (t) dt. . then the inverse Fourier transform is given by f (x) = Proof Firstly.3 (The Fourier inversion theorem) If f (x) is a good function with Fourier transform ˜ f (k) = ∞ −∞ eikx f (x) dx. 23). For good functions. Similarly. π so we can write 1 F e− 2π x2 ∞ −∞ ∞ −∞ e−(k−t) 2 /4 dt = π ∞ −∞ e−t 2 /4 dt = 2π. Note that f ∗ g = g ∗ f . We ﬁnd that F[f (x)] = ∞ −∞ f (x)eikx dx = −ik ∞ −∞ f (x)eikx dx = −ikF[f ]. this result also holds for generalized functions. x∈R ∞ −∞ and hence 1 F e− 2π x2 ˜ f (−x) − f (k) π 1 2π π ∞ −∞ max |f (x)| x∈R 2 e−(k−t) 2 /4 |t − k| dt = 1 2π max |f (x)| 4 x∈R e−X |X| dX → 0 as ∞ −∞ ∞ −∞ → 0. this can be done using integration by parts. the solutions of diﬀerential equations can often be written as a convolution because of the key property F[f ∗ g] = F[f ]F[g]. We conclude that f (k) = 1 1 ˜ F f (−x) = 2π 2π eikx e−ixt f (t) dt. max |f (x)| . This is precisely the Fourier integral. (5. 1 ˜ f (−x) − f (k) = 2π f (t) − f (k) t−k π e−(k−t) 2 /4 {f (t) − f (k)} dt.2. since the good function f must tend to zero as x → ±∞.140 FOURIER SERIES AND THE FOURIER TRANSFORM In addition. Since generalized functions are deﬁned in terms of sequences of good functions. we need to be able to ﬁnd the Fourier transform of a derivative. the second derivative is F[f (x)] = −k 2 F[f ]. and hence the result is proved.5 Transforms of Derivatives and Convolutions Fourier transforms are an appropriate tool for solving diﬀerential equations on the unbounded domain −∞ < x < ∞. Since f is a good function. As we shall see below. which is a function of x only. In order to proceed. We can also deﬁne the convolution of two functions f and g as f ∗g = ∞ −∞ f (y)g(x − y) dy. . 5. we write down the deﬁnition of the Fourier transform of f ∗ g.3 Green’s Functions Revisited Let’s. for a linear operator L with domain −∞ < x < ∞. Taking the Fourier transform of both sides of this equation and exploiting the linearity of the transform. forget our previous deﬁnition of a Green’s function (Section 4. since this is independent of z.3 GREEN’S FUNCTIONS REVISITED 141 To derive this. dx2 We will solve this equation subject to G → 0 as |x| → ∞. and can be extracted from the inner integral to give F[f ∗ g] = ∞ ∞ f (y) y=−∞ x=−∞ eikx g(x − y) dx dy. we have F[f ∗ g] = ∞ ∞ f (y) y=−∞ z=−∞ eik(z+y) g(z) dz dy. we can use the Fourier transform to ﬁnd G.1. and deﬁne it instead. Now f (y) is independent of x. Note that the factor eikx is independent of y and so can be taken inside the inner integral to give F[f ∗ g] = ∞ x=−∞ ∞ y=−∞ eikx f (y)g(x − y) dy dx. (5. and now extracting the factor eiky . for the moment.29). dx2 d2 − 1. Since the limits of integration are independent of one another.2).5. gives F d 2G − F[G] = F[δ(x)]. as the solution of the diﬀerential equation LG = δ(x) subject to G → 0 as |x| → ∞. 5. consider the operator L≡ so that LG = δ is d 2G − G = δ(x). dx2 . we can exchange the order of integration so that F[f ∗ g] = ∞ y=−∞ ∞ x=−∞ eikx f (y)g(x − y) dx dy. By making the transformation z = x − y (so that dz = dx) in the inner integral. F[f ∗ g] = ∞ −∞ eikx ∞ −∞ f (y)g(x − y) dy dx. For example. allows us to write F[f ∗ g] = ∞ y=−∞ f (y)eiky dy ∞ z=−∞ eikz g(z) dz = F[f ]F[g]. If we assume that G is a good function. using the fact that F[G ] = −k 2 F[G]. and is continuous there. we get p(x) dG dx + − − q(x)G(x) dx = − δ(x) dx = 1.2. → 0 this reduces to If p(x) is continuous at x = 0. p(0) which.28) means that 1 G(x) = − e−|x| .2. is the result that we obtained in Section 4. If we integrate between x = − and x = . the free space Green’s function depends only on x. as before.1. 1 . From the deﬁnition F[δ(x)] = Therefore −k 2 F[G] − F[G] = 1. On any interval that does not contain the point x = 0. Firstly. The solution is φ = G ∗ P . and the deﬁnition that we gave in Section 4. the answer is.142 FOURIER SERIES AND THE FOURIER TRANSFORM Firstly we need to determine the Fourier transform of the delta function. when we take the limit dG dx x=0 + = x=0− 1 . whilst the other deﬁnition depends upon both x and s. To see this. Suppose that we need to solve Lφ = P . G is clearly the solution of the homogeneous problem. which is sometimes referred to as the free space Green’s function. to be able to solve the inhomogeneous diﬀerential equation. apart from a sign diﬀerence. . We can see some similarities by considering the self-adjoint problem d dx p(x) dG dx + q(x)G = δ(x). k2 + 1 G(x − y)P (y) dy = ∞ ∞ −∞ LG(x − y)P (y) dy = −∞ δ(x − y)P (y) dy = P (x). note that L(G ∗ P ) = L ∞ −∞ ∞ −∞ eikx δ(x) dx = eikx x=0 = 1.1. There are both diﬀerences and similarities between this deﬁnition of the Green’s function. After rearrangement this becomes F[G] = − which we know from (5. Once we know the Green’s function we can therefore write down the solution of Lφ = P as φ = G ∗ P . 2 Why should we want to construct such a Green’s function? As before. This gives us F ˜ ∂2φ ∂2φ ∂2φ ˜ + − k 2 φ = 0. where f (k) = F[f ]. we require that φ is bounded as y → ∞. namely φ(x. Using the so that F[e formula for the inverse transform gives e−a|x| = 1 2π ∞ −∞ e−ikx 2a dk. with the inverse g transform of g (k) = e−|k|y . so that ˜ ˜ φ(k. We can invert this Fourier transform of the solution using the convolution theorem. We deﬁne ˜ φ(k. We will return to consider the Green’s function for partial diﬀerential equations in Section 5.1.4 SOLUTION OF LAPLACE’S EQUATION USING FOURIER TRANSFORMS 143 The end result from constructing the Green’s function is the same as in Section 4. and hence ˜ φ(k. To ﬁnd g(x) = F −1 [˜(k)] = F −1 [e−|k|y ]. f (x). Boundary value problems for partial diﬀerential equations where the value of the dependent variable is prescribed on the boundary are often referred to as Dirichlet problems. which gives A(k) = 0. for y > 0 subject to φ = f (x) on y = 0 and φ bounded as y → ∞.5. 0) = f (k). 0) = f (x). and hence e = F [2a/(a + k )]. It now remains to satisfy the condition at y = 0. = ∂y 2 ∂x2 ∂y 2 ˜ This has solution φ = A(k)e|k|y +B(k)e−|k|y .2. We can solve this Dirichlet problem using a Fourier transform with respect to x.28) that F[e−|x| ] = 2/(1 + k 2 ). We begin by taking the Fourier transform of Laplace’s equation.5. ∇2 φ = 0. y) dx = F[φ].4 Solution of Laplace’s Equation Using Fourier Transforms Let’s consider the problem of solving Laplace’s equation. noting that F ˜ ∂2φ ∂2φ ∂2 F [φ] = . ˜ By putting y = 0 we therefore ﬁnd that B(k) = f (k). We take ˜ ˜ ˜ the Fourier transform of this condition to give φ(k. the solution is just the convolution of the boundary condition. 5. = 2 2 2 ∂y ∂y ∂y which is easily veriﬁed from the deﬁnition of the transform. a2 + k 2 . y) = ∞ −∞ eikx φ(x. However. we note from ˜ ˜ (5. y) = B(k)e−|k|y . Since ˜˜ F −1 [f g ] = F −1 [F[f ] F[g]] = f ∗ g. We can express the solution of the inhomogeneous boundary value problem as an integral that involves the Green’s function and the inhomogeneous term. and exploit the fact that F[f (ax)] = f (k/a)/a. −a|x| 2 2 −a|x| −1 2 2 ] = 2a/(a + k ). y) = f (k)e−|k|y . + (x − ξ)2 (5. and the condition at y = 0 tells us that B(k) = − and hence ˜ ˜ e φ(k.144 FOURIER SERIES AND THE FOURIER TRANSFORM We can exploit the similarity between the Fourier transform and its inverse by making the transformation k → −x.2 for details of how to create this type of contour plot in MATLAB. y > 0. + x2 ) 1 2π ∞ −∞ eikx 2a 2a 1 dx = . which leads to e−a|k| = and hence F −1 e−|k|y = π(y 2 y .31) Figure 5. 1 −1 Using the formula we have derived. consider the two-dimensional. . irrotational ﬂow in the upper half plane (see Section 2. x → k.† Finally. The solution φ can now be written as the convolution integral ∞ φ(x. let’s consider the Neumann problem for Laplace’s equation in the upper half plane.6. φ= y π = 1 −1 dξ y 1 tan−1 = y 2 + (x − ξ)2 π y tan−1 1−x y + tan−1 1 π 1+x y (5. inviscid.6. F 2 a2 + x2 2π a + x2 In the deﬁnition of the convolution we use the variable ξ as the dummy variable rather than y to avoid confusion with the spatial coordinate in the problem. 0) = f (x). driven by the Dirichlet condition φ(x. y) = −∞ f (ξ) y y dξ = 2 + (x − ξ)2 } π{y π ∞ −∞ y2 f (ξ) dξ. ξ−x y . y) = −f (k) −|k|y ˜ f (k) . y) = B(k)e−|k|y . with ∂φ/∂y = f (x) at y = 0. y + x2 −∞ † See Section 2.2). where f (x) = 1 0 for −1 < x < 1. In order to invert this Fourier transform we recall that ∞ 2y eikx 2 dx = 2πe−|k|y . so that f ∗g = ∞ −∞ f (ξ)g(x − ξ) dξ.5 shows some contours of equal φ. |k| |k| .30) As an example. This is the same as the Dirichlet problem except that the boundary condition is in terms of a derivative. elsewhere. As ˜ before we ﬁnd that φ(k. x2 . 5. . Rn . through the integral δ(x)F (x) d n x = F (0). 5. φ. f (ξ) log{y 2 + (x − ξ)2 } dξ. 5.5. xn ) be a vector in Rn .31). we can deﬁne the delta function in Rn .5. .5 GENERALIZATION TO HIGHER DIMENSIONS 145 −10 −8 −6 −4 −2 2 4 6 8 10 Fig. This allows us to use the techniques that we have developed above to solve other partial diﬀerential equations. y) = 1 2π ∞ −∞ e−|k|y 1 = F log y 2 + x2 |k| 2π .5. given by (5. |k| In other words − and hence the solution is φ(x.1 The Delta Function in Higher Dimensions If we let x = (x1 . . and integrate both sides with respect to y to obtain ∞ x=−∞ eikx log(y 2 + x2 ) dx = −2π e−|k|y . Contours of constant potential function.5 Generalization to Higher Dimensions The theory of Fourier transforms and generalized functions can be extended to higher dimensions. . δ(x). δ(xn−1 )F (x1 . the volume of a thin. Cartesian Coordinates In Cartesian coordinates we have δ(x) = δ(x1 )δ(x2 ) .2 Fourier Transforms in Higher Dimensions If f (x) = f (x1 . πr Spherical Polar Coordinates In R . 2π θ=0 ∞ r=0 ∞ δ(x) d 2 x = 2π 0 ∞ a(r)δ(r)r dr. Again. 0)d n−1 x = · · · = F (0. . k2 . in R2 . the isotropic volume element can be written as d 3 x = 4πr2 dr. and F is a good function in each of the coordinates xi . δ(xn ). in R3 ˜ f (k1 . x2 . xn−1 . kn ).32) For example. the delta function must be a multiple of δ(r). By symmetry. x3 )ei(k1 x1 +k2 x2 +k3 x3 ) dx1 dx2 dx3 . . (r. and the same argument gives 3 δ(x) = 1 δ(r). . and a multiple of δ(r). . . k2 . δ(xn−1 ) R δ(xn )F (x1 . . . . 0). . δ(r) dr = 0 1 . 2πr2 5. . . . spherical shell. We therefore let δ(x) = a(r)δ(r) and. xn−1 . . . x2 . . . . . Plane Polar Coordinates In terms of standard plane polar coordinates. we can deﬁne the Fourier transform of f as ˜ f (k) = Rn f (x)eik·x d n x. since δ(x)F (x) d n x Rn = Rn−1 δ(x1 ) .5. the distance from the origin. xn ) dxn d n−1 x = Rn−1 δ(x1 ) . . xn ) and k = (k1 . in other words independent of θ.146 FOURIER SERIES AND THE FOURIER TRANSFORM This is a multiple integral over the whole of Rn . θ). 0. (5. . . . . k3 ) = ∞ x3 =−∞ ∞ x2 =−∞ ∞ x1 =−∞ f (x1 . where r is now |x|. . 2 so we can take a(r) = 1/πr. . and hence δ(x) = 1 δ(r). noting that d 2 x = r dr dθ. . . δ(x) must be isotropic. x3 )ei(k1 x1 +k2 x2 +k3 x3 ) dx1 dx2 dx3 . Since G vanishes at inﬁnity. This result is. k3 ) = so that F ∂G = ∂x1 ∞ −∞ ∞ −∞ ∞ ∞ −∞ ∞ −∞ ∞ −∞ ∞ −∞ (5. In this case.33) The inversion of Fourier transforms in higher dimensions is considerably more difﬁcult than in one dimension.5. (ii) The inversion formula is f (x) = 1 (2π)n ˜ f (k)e−ik·x d n k. although we will not provide the details. analogous to the one-dimensional result that we derived in Section 5.34) G(x1 . that: (i) If f (x) is a generalized function. The Fourier transform of G is ˜ G(k1 . x3 )ei(k1 x1 +k2 x2 +k3 x3 ) dx1 dx2 dx3 ∂x1 ei(k1 x1 +k2 x2 +k3 x3 ) G ∞ −∞ ∞ −∞ = −∞ ∞ −∞ − Gik1 ei(k1 x1 +k2 x2 +k3 x3 ) dx1 dx2 dx3 . We introduced the idea of a Green’s function for a linear operator in Section 5.5 GENERALIZATION TO HIGHER DIMENSIONS 147 We can proceed as we did for one-dimensional functions and show. k2 . ∂G (x1 . we conclude that ∂G = −ik1 ∂x1 ∞ −∞ ∞ −∞ ∞ −∞ ˜ Gei(k1 x1 +k2 x2 +k3 x3 ) dx1 dx2 dx3 = −ik1 G.3.2. The inversion formula. then shows that G(x) = − 1 8π 3 R3 e−ik·x 3 d k. |k|2 (5.34).35) In order to evaluate this integral. we ﬁnd that ˜ (−ik1 )2 + (−ik2 )2 + (−ik3 )2 G = δ(x)eik·x d 3 x = 1. of course. after integrating by parts.5. Rn (5. x2 . we seek a solution of ∇2 G = δ(x) subject to G → 0 as |x| → ∞.33). If we now take the Fourier transform of (5. Example: Laplace’s equation Let’s try to construct the free space Green’s function for Laplace’s equation in R3 . R3 ˜ and hence G = −1/|k|2 . then its Fourier transform exists as a generalized function. we need to introduce spherical polar coordinates . (5. as we shall see. x2 . 36) ∞ 0 in terms of a convolution integral. Now that we know the Green’s function. Example: The wave equation Let’s try to solve the three-dimensional wave equation†. c2 ∂t2 subject to the initial conditions ∇2 u = u(x.39) ∂u (x. ∂t (5. ˜ 2 ∂t (5. (5.2. ˜ † See Section 3.37) for sound waves. The Fourier transform of (5.37) is ˜ ∂2u = −c2 k 2 u. 0) = f (x. z. y.38). y. by a direct generalization of the results presented in Section 5. 1 ∂2u for t > 0 and x ∈ R3 .39) is ∂u ˜ ˜ (k. φ=− 4π R3 |x − y| This result is fundamental in the theory of electrostatics. θ. and Billingham and King (2001) for a derivation of (5. z 4πr using the standard result. 0) = 0.5.35) becomes G(x) = − 1 4π 2 1 8π 3 ∞ k=0 ∞ k=0 π θ=0 2π φ=0 π e−ikr cos θ 2 k sin θ dk dθ dφ k2 dk = − 1 2π 2 ∞ k=0 =− e−ikr cos θ ikr 1 2π 2 r ∞ θ=0 sin kr dk kr =− z=0 1 sin z dz = − . 0) = 0.148 FOURIER SERIES AND THE FOURIER TRANSFORM (k. where Q is the distribution of charge density. ˜ The general solution of (5. y. z. 0) = f (k). then become u(k. Ablowitz and Fokas. ˜ (5. for example. z). (sin z/z) dz = π/2 (see. . and φ the corresponding electrical potential. with the line θ = 0 in the x-direction.1 for a derivation of the wave equation in two dimensions. ∂t u = A(k)eickt + B(k)e−ickt . The initial conditions.37) (5. φ). Then k · x = |k| |x| cos θ = kr cos θ and (5. 1997).38) 2 2 2 where u is the Fourier transform of u and k 2 = k1 + k2 + k3 . We ﬁnd that Q(y) 3 1 d y.9. we are able to solve the inhomogeneous problem ∇2 φ = Q(x) subject to φ → 0 as |x| → ∞. 1 5. .4 5. It is somewhat easier to interpret solutions like (5. Determine the Fourier series expansion of the function f (x) = Using this series.40) in the large time limit.38) can be written as a continuous spectrum of plane wave solutions travelling in all directions. ﬁnd an expression for δ(g(x)) in terms of the sequence of delta functions {δ(x − xi )}. t) = 1 8π 3 ˜ f (k) {exp (ickt − ik · x) − exp (−ickt − ik · x)} d 3 k. What is the general solution of the equation (x − a)f (x) = b.3 Show that 1 δ(x). x for 0 < x < π. 2α Express δ(ax + b) in the form µδ(x + σ) for appropriately chosen constants µ and σ. at speed c. (5. the terms exp (±ickt − ik · x) represent plane travelling wave solutions of the three-dimensional wave equation.7 5. . 2ick The inversion formula.40) is a weighted sum of these plane waves at diﬀerent wavenumbers k. Since the integral (5. where (a) f (x) = ex . which move perpendicular to themselves.x = ±ckt. + 2 + ··· = 32 5 8 0 for −π < x 0. Exercises 5. (c) f (x) = x2 . We will do this in Section 11. if f is a generalized function? If g(x) = 0 at the points xi for i = 1. .6 5. since they remain constant on the planes k.8 . then shows that u(x. 2. with |x|/t ﬁxed. d −|x| d Show that sgn(x) = 2δ(x).2 Determine the Fourier series expansion of f (x) for −π < x < π.2 using the method of stationary phase. 5.33). we can see that the Fourier transform has revealed that the solution of (5.2.EXERCISES 149 and when we enforce the initial conditions. in the direction ±k. this becomes u= ˜ ˜ f (k) ickt e − e−ickt . What is e ? dx dx Calculate the Fourier transforms of (a) δ(αx) = 1 1 π2 . . t 1. (b) f (x) = | sin x| . show that 1+ 5. n. |α| 1 (b) δ(x2 − α2 ) = {δ(x + α) + δ(x − α)}.5 5.37) subject to (5. 2ick (5.40) R3 For any given value of k. evaluate the integral π ekx einx dx.150 FOURIER SERIES AND THE FOURIER TRANSFORM (a)   0 f (x) = −1  1 for x < −1 and x > 1. F[f (x − a)] = eika F[f (x)].9 xe−x 0 for x > 0. ∂t ∂x for −∞ < x < ∞. for 0 < x < 1. e−λt u(x. t) = √ 4πkt 5. subject to u(x. ex 0 for x < 0. t > 0. ∂2u ∂u = k 2 − λu. (b) f (x) = (c) f (x) = 5. Verify that √ 2 J(0) = π and hence evaluate F e−x . (x − y)2 4kt can be written in convolution form as exp − f (y) dy. . 0) = f (x). Show that the Fourier transform has the ‘shifting property’.11 Use Fourier transforms to show that the solution of the initial boundary value problem. for x < 0.10 If ∞ J(k) = −∞ e−(x−ik/2) dx. 5. For n ∈ N and k ∈ Z.12 ∞ −∞ u→0 as |x| → ∞. for x > 0. 2 show by diﬀerentiation under the integral sign that dJ/dk = 0. for −1 < x < 0. with λ and k real constants. k 2 + n2 Obtain the Fourier series of the function (of period 2π) deﬁned by f (x) = cosh kx for − π < x < π. −π and hence show that π ekx cos nx dx = −π (−1)n+1 2n sinh kπ . 5. where k0 = ω/c0 . + n2 )2 5. What is the physical signiﬁcance of these like e boundary conditions? (c) Show that φ = e−iωt G ∗ q = e−iωt 2ik0 x q(y)eik0 (x−y) dy + −∞ ∞ x q(y)e−ik0 (x−y) dy . 2 (b) Find the Green’s function solution of G + k0 G = δ(x) that behaves ±ik0 x as x → ±∞. (d) Show that. Use Fourier transforms to show that G(x) = − 1 8π 3 R3 e−ik·x 3 d k. φ∼ e−i(ωt−k0 x) 2ik0 ∞ −∞ q(y)e−ik0 y dy. (a) Show that if Q = q(x)e−iωt then a separable solution exists with 2 φ = e−iωt f (x) and f + k0 f = q(x).14 The free space Green’s function for the modiﬁed Helmholtz equation in three dimensions satisﬁes ∇2 − m2 G = δ(x1 . x2 . with G → 0 as |x| → ∞.13 The forced wave equation is ∂2φ 1 ∂2φ − 2 2 = Q(x. t). Verify by direct substitution that this function satisﬁes the modiﬁed Helmholtz equation. ∂x2 c0 ∂t where Q is the forcing function. as x → ∞. k 2 + m2 Use contour integration to show that this can be simpliﬁed to 1 −mr e . What is the physical interpretation of this result? 5.EXERCISES 151 Find the value of 1 k 2 + n2 n=1 as a function of k and use (4. x3 ). 4πr where r = |x|. G(x) = − . which satisﬁes Q → 0 as |x| → ∞.49) to evaluate ∞ n=1 ∞ (k 2 1 . and hence that the integral converges. we need s > k. with t representing time. ∞). This shows that 1 for Re(s) > Re(k). the kernel is e−st and I = R+ = [0. K(x.CHAPTER SIX Laplace Transforms Integral transforms of the form T [f (x)] = I K(x. which is how we will need to treat s when we use the inversion formula. in order that e−(s−k)t → 0 as t → ∞.1 Deﬁnition and Examples ∞ 0 The Laplace transform of f (t) is L[f (t)] = F (s) = e−st f (t) dt. 6. it is closely related to the Fourier transform. such as linear integral equations. (6. In this chapter we will study the Laplace transform. 0 Now. we have L[ekt ] = 0 ∞ ∞ 0 e−st ekt dt = e−(s−k)t dt = e−(s−k)t −(s − k) ∞ . This transform allows us to solve linear initial value problems. (6. As we shall see. can also be solved using integral transforms. are useful tools for solving linear diﬀerential equations. we need Re(s) > Re(k). for which the kernel is eikx and I = R. Substituting this into (6.5).1) We will consider for what values of s the integral is convergent later in the chapter. where k is a constant. for which the usual notation for the original variable is t and for the transform variable is s.1). In Chapter 5 we met the Fourier transform. L[ekt ] = s−k . where I is some interval on the real line. Example 1 Consider f (t) = ekt . The Fourier transform allows us to solve linear boundary value problems whose domain of solution is the whole real line. and begin with some examples. Other types of equation. k) is the kernel of the transform and k is the transform variable. If s is a complex-valued variable. k)f (x) dx. As we shall see. we have rather glossed over the question of when the Laplace transform of a particular function actually exists.4. we obtain L[tn ] = 0 ∞ e−x x s n 1 1 dx = n+1 s s n+1 ∞ 0 e−x xn dx = Γ(n + 1) .1 The Existence of Laplace Transforms So far.5). In the same vein.25). in Section 6. We can combine the two fractions to show that s L[cos ωt] = 2 for Re(s) > 0. and involves an integral in the complex s-plane. a result that we could have obtained In order to use the Laplace transform. 6. we will need to know how to invert it. If we let x = st so that dx = s dt. By deﬁnition. sn+1 If n is an integer. For the Fourier transform. In fact the deﬁnition of the Laplace transform is an improper integral and should be written as N L[f (t)] = lim N →∞ e−st f (t) dt. so that we can determine f (t) from a given function F (s). s + ω2 It is easy to show that L[sin ωt] = ω/(s2 + ω 2 ) using the same technique. We have.1 DEFINITION AND EXAMPLES 153 Example 2 Consider f (t) = cos ωt. however. 2 L[cos ωt] = 1 2 ∞ 0 e(iω−s)t + e−(iω+s)t ds = 1 2 1 1 + s − iω s + iω .1. where ω is a constant. Since cos ωt = 1 (eiωt + e−iωt ).6. We will return to give an outline derivation of the inversion formula. For the moment. (6. there are a lot of techniques available that allow us to invert the Laplace transform in an elementary manner. The inverse of a Laplace transform is rather similar. s − a2 s − a2 Example 3 Consider f (t) = t for n ∈ R. provided Re(s) > 0. n . s a L[cosh(at)] = 2 . this gives L[t ] = n!/s directly using integration by parts. n ∞ 0 L[tn ] = e−st tn dt. L[sinh(at)] = 2 . 0 . discussed in passing that the real part of s needs to be greater than some constant value in some cases. we will deal with the problem of inverting Laplace transforms by trying to recognize the function f (t) from the form of F (s). this inversion process involves the integral formula (5. .2 Properties of the Laplace Transform Theorem 6. ∞). The proof of this is rather technical. L et t2 2 N = lim N →∞ et e−st dt. if lims→∞ F (s) = 0. If f is of exponential order. using the triangle inequality. if f is not of exponential order and grows too quickly as t → ∞. s2 /(s2 + 1) → 1 as s → ∞. we therefore have |L[f (t)]| < and hence |L[f (t)]| → 0 as |s| → ∞. Kuller.154 LAPLACE TRANSFORMS We will now ﬁx matters by deﬁning a class of functions for which the Laplace transform does exist. We say that a function f (t) is of exponential order on 0 t < ∞ if there exist constants A and b such that |f (t)| < Aebt for t ∈ [0. and the interested reader is referred to Kreider. the integral will not converge. and is not therefore the Laplace transform of a function of exponential order. Theorem 6. Proof Consider the deﬁnition of L[f (t)] and its modulus |L[f (t)]| = 0 ∞ e−st f (t) dt 0 ∞ e−st f (t) dt = 0 ∞ e−st |f (t)| dt. ∞). In contrast. consider the Laplace transform of the 2 function et . then L[f ] → 0 as |s| → ∞. Conversely.2 If f is of exponential order on [0. 6. 2 0 It should be clear that e grows faster than est for all s. |L[f (t)]| < 0 ∞ e−st Aebt dt = 0 ∞ Ae(b−s)t dt = A (b−s)t e b−s ∞ = lim 0 Y →∞ A (b−s)Y A e − . ∞) has a Laplace transform. Ostberg and Perkins (1966). For example.3 (Linearity) The Laplace transform and inverse Laplace transform are linear operators. For example. Theorem 6. s−b Provided s > b. b−s b−s A . then F (s) is not the Laplace transform of a function of exponential order.1 (Lerch’s theorem) A piecewise continuous function of exponential order on [0. and hence that the Laplace transform of et does not exist. so that the integrand 2 diverges. consider L−1 [1/(s2 − 5s + 6)]. This is extremely useful.4 (First shifting theorem) If L[f (t)] = F (s) for Re(s) > b. so the Laplace transform is a linear operator. We can proceed by splitting this rational function into its partial fractions representation and exploit the linearity of the inverse Laplace transform. If we now deﬁne F (s) = L[f ] and G(s) = L[g]. . that is.1. we obtain αL−1 [F ] + βL−1 [G] = L−1 [αF + βG] . For example. for example L[2t2 − t + 1] and L[e3t + cos 2t] easily in terms of the Laplace transforms of their constituent parts.6. As we mentioned earlier. then L[eat f (t)] = F (s − a) for Re(s) > a + b. Often we can recognize the constituents of an expression whose inverse Laplace transform we seek. L[eat f (t)] = 0 ∞ e−st eat f (t) dt = 0 ∞ e−(s−a)t f (t) dt = F (s − a). L[αf (t) + βg(t)] = 0 ∞ ∞ e−st (αf (t) + βg(t)) dt =α 0 e−st f (t) dt + β 0 ∞ e−st g(t) dt = αL[f (t)] + βL[g(t)]. this gives αL−1 [L[f (t)]] + βL−1 [L[g(t)]] = L−1 [αL[f (t)] + βL[g(t)]] . which we would prefer to avoid when possible. This is a direct consequence of Lerch’s theorem. provided that Re(s − a) > b. Theorem 6. We have L−1 = L−1 − 1 1 = L−1 s2 − 5s + 6 (s − 2)(s − 3) 1 1 1 1 + = −L−1 + L−1 = −e2t + e3t . Taking the inverse Laplace transform of both sides gives αf (t) + βg(t) = L−1 [αL[f (t)] + βL[g(t)]] . the Laplace inversion formula involves complex integration. Proof By deﬁnition. since we can calculate.2 PROPERTIES OF THE LAPLACE TRANSFORM 155 Proof By deﬁnition. For every function f (t) the Laplace transform L[f (t)] is uniquely deﬁned and vice versa. it is one-to-one and onto. 6. so the inverse Laplace transform is also a linear operator. s−2 s−3 s−2 s+3 We pause here to note that the inversion of Laplace transforms using standard forms is only possible because the operation is a bijection. Since we can write f (t) as L−1 [L[f (t)]] and similarly for g(t). we ﬁrstly note that L[t2 ] = 2/s3 .5 (Second shifting theorem) If the Laplace transform of f (t) is F (s).1. where H is the Heaviside step function. We recall from the previous section that s for Re(s) > 0. can be written in terms of any dummy variable of integration. L[g(t)] = 0 ∞ g(t)e−st dt = a ∞ f (t − a)e−st dt. Proof By deﬁnition. s3 2 . L e3t cos 4t = s−3 for Re(s) > 3.156 LAPLACE TRANSFORMS Example 1 Consider the function f (t) = e cos 4t. since the deﬁnition of the Laplace transform. (s − 3)2 + 42 Example 2 Consider the function F (s) = s/(s + s + 1). L [cos 4t] = 2 s + 42 Using Theorem 6. Using the ﬁrst shifting theorem then gives us √ √ s 1 −t/2 3t 3t −1 −t/2 L −√ e . (6. s = L−1 s2 + s + 1 s+ s+ 1 2 2 1 +3 2 4 √ −L −1 1 √ 3 s+ 3 2 1 2 2 + 3 4 . What is its inverse Laplace transform? We begin by completing the square of the denominator. By writing τ = t − a. to determine the inverse transform of the function e−3s /s3 .4. The second shifting theorem then shows immediately that L−1 e−3s 1 = H(t − 3)(t − 3)2 . cos sin =e 2+s+1 s 2 2 3 Theorem 6. For example. this becomes L[g(t)] = 0 ∞ f (τ )e−sτ e−sa dτ = e−sa F (s).1). then the Laplace transform of the function g(t) = H(t−a)f (t−a) is e−sa F (s). using Example 3 of Section 6. which gives F (s) = s+ and hence L −1 2 3t s 1 2 2 + 3 4 = s+ s+ 1 2 1 2 +3 2 4 − s+ 1 2 1 2 2 + 3 4 . A useful corollary of (6.6.2) is that. for example. we will need to be able to calculate the Laplace transform of the derivative of a function.2) g(t) = 0 f (τ )dτ. This means that L[f ] = sL[f ] − f (0). so that. t L[g] = L 0 f (τ )dτ = 1 L[f ]. we will assume that the values of s are restricted so that e−st f (t) → 0 as t → ∞. + ω2 ) We know that L[sin ωt]/ω = 1/(s2 + ω 2 ) so that f (t) = L−1 t 1 1 1 1 sin ωt = L−1 L s s2 + ω 2 s ω 1 1 sin ωτ dτ = 2 (1 − cos ωt) .3 THE SOLUTION OF ORDINARY DIFFERENTIAL EQUATIONS 157 6. s and hence L−1 1 F (s) = s t f (τ ) dτ. except where f (t) is discontinuous. Since g(0) = 0 by deﬁnition. we have L[f ] = sL[g]−g(0). g (t) = f (t). 0 This can be useful for inverting Laplace transforms. L[f ] = 0 ∞ e−st f (t) dt. ω ω = 0 Let’s now try to solve the simple diﬀerential equation dy − 2y = 0. if t (6. it is trivial to solve this . dt subject to the initial condition y(0) = 1. Of course. we ﬁnd that L[f ] = e−st f (t) ∞ 0 − 0 ∞ −se−st f (t) dt. At this stage. By deﬁnition. F (s) = s(s2 1 . After integrating by parts.3 The Solution of Ordinary Diﬀerential Equations Using Laplace Transforms In order to be able to take the Laplace transform of a diﬀerential equation. Using the initial conditions then shows that Y (s) = 1 . Using the initial condition and manipulating the equation gives 1 . For example. we can now solve the diﬀerential equation d 2y dy + 6y = 0. dt where Y (s) = L[y(t)]. Let’s now consider the solution of the coupled system of equations dy1 dy2 + y1 = y2 . where G(s) = L[g] = L[f ] = sF (s) − f (0). Y (s) = which immediately shows that y(t) = e3t − e2t . We ﬁnd that s2 Y (s) − sy(0) − y (0) − 5(sY (s) − y(0)) + 6Y = 0. Many of the diﬀerential equations that we will try to solve are of second order. but it is a useful illustrative example. It is also straightforward to show by induction that L[f (n) ] = sn F (s) − sn−1 f (0) − sn−2 f (0) − · · · − f (n−1) (0). so we need to determine L[f ].2) shows that L[g ] = sG(s) − g(0).158 LAPLACE TRANSFORMS separable equation. We begin by taking the Laplace transform of the diﬀerential equation.3) We can obtain the same result by integrating the deﬁnition of the Laplace transform of f twice by parts. which gives L dy − 2L[y] = sY (s) − y(0) − 2Y (s) = 0. Although we could 1 1 − . Y (s) = s−2 which is easily inverted to give y(t) = e2t . If we introduce the function g(t) = f (t). We conclude that L[f ] = L[g ] = s2 F (s) − sf (0) − f (0). s2 − 5s + 6 In order to invert the Laplace transform we split the fraction into its constituent partial fractions. s−3 s−2 . dt dt subject to the initial conditions that y1 (0) = y2 (0) = 1. − y2 = y 1 . (6. −5 dt2 dt subject to the initial conditions y(0) = 0 and y (0) = 1 using Laplace transforms. (6. Note that 1 L−1 = 1. s+1 and. Note that the reduction of a system of diﬀerential equations to a system of algebraic equations is the key beneﬁt of these transform methods. using the ﬁrst shifting theorem. The transform of the equations is sY1 − 1 + Y1 = Y2 . sY2 − 1 − Y2 = Y1 . L−1 This means that L−1 1 1 1 − = L−1 = 1 − e−t . The solution of these algebraic equations is s s+2 . + dt2 dt In order to take the Laplace transform of this equation. where L[yj (t)] = Yj (s). We can calculate this directly. using the second shifting theorem. it is easier to use the standard solution techniques described in Appendix 5. The real power of the Laplace transform method lies in the solution of problems for which the inhomogeneity is not of a simple form. s(s + 1) s s+1 e−s = H(t − 1)(1 − e−(t−1) ). as L[δ(t − a)] = 0 ∞ δ(t − a)e−st dt = e−sa . for simple scalar equations. Of course. and hence Y (s) = e−s + (s + 1)y(0) + y (0) . for example d2 y dy = δ(t − 1). y2 (t) = cosh 2t + 2 sinh 2t. we need to know the Laplace transform of the delta function. Y1 (s) = 2 s −2 s −2 which we can easily invert by inspection to give √ √ √ √ y1 (t) = cosh 2t. s and. we will solve them directly using Laplace transforms. so that the diﬀerential equation becomes s2 Y (s) − sy(0) − y (0) + sY (s) − y(0) = e−s . s(s + 1) Judicious use of the two shifting theorems now allows us to invert this Laplace transform.3 THE SOLUTION OF ORDINARY DIFFERENTIAL EQUATIONS 159 combine these two ﬁrst order equations to obtain a second order equation. L−1 . s(s + 1) 1 = e−t .6. Y2 (s) = 2 . We can show this by taking the Laplace transform of the convolution integral. However. shows that y(t) = y(0) + y (0)(1 − e−t ) + H(t − 1)(1 − e−(t−1) ). and hence L[f ∗ g] = ∞ τ =0 f (τ )e−sτ dτ ∞ z=0 e−sz g(z) dz = L[f ] L[g]. as shown in Figure 6. we deﬁne the convolution of two functions as t f ∗g = 0 f (τ )g(t − τ ) dτ.1 The Convolution Theorem The Laplace transform of the convolution of two functions plays an important role in the solution of inhomogeneous diﬀerential equations.3. and using the linearity of the Laplace transform. Since e−st is independent of τ it can be moved inside the inner integral so that we have L[f ∗ g] = 0 ∞ 0 t e−st f (τ )g(t − τ ) dτ dt. if f (t) and g(t) have Laplace transforms. when using Laplace transforms.160 LAPLACE TRANSFORMS Combining all of these results. If we switch the order of integration. With this deﬁnition. this becomes τ =∞ L[f ∗ g] = τ =0 ∞ t=τ e−st f (τ )g(t − τ ) dt dτ. we can transform the inner integral into L[f ∗ g] = 0 ∞ ∞ f (τ ) 0 e−s(z+τ ) g(z) dz dτ.1. This result is most useful in the form L−1 [F (s)G(s)] = L−1 [F (s)] ∗ L−1 [G(s)]. so that dz = dt. If we now introduce the variable z = t − τ . 6. to obtain L[f ∗ g] = 0 ∞ e−st 0 t f (τ )g(t − τ ) dτ dt. which is itself a function of t. just as it did for Fourier transforms. . We now note that the domain of integration is a triangular region delimited by the lines τ = 0 and τ = t. then L[f ∗ g] = L[f (t)]L[g(t)]. Example 1 Consider the Laplace transform F (s) = 1/(s − 2)(s − 3). (s − 2)(s − 3) s−2 s−3 We now need to calculate the convolution integral. and hence L−1 1 1 1 = L−1 ∗ L−1 = e2t ∗ e3t .3 THE SOLUTION OF ORDINARY DIFFERENTIAL EQUATIONS 161 τ t=τ t Fig. of course. The domain of integration of the Laplace transform of a convolution integral. This is. namely 1/(s − 2) = L[e2t ] and 1/(s − 3) = L[e3t ]. This rational function has two easily recognizable constituents. the same result as we obtained using partial fractions. which is t t e2t ∗ e3t = τ =0 e2τ e3(t−τ ) dτ = e3t τ =0 e−τ dτ = e3t −e−τ t 0 = e3t −e−t + 1 = −e2t + e3t . + 1)(s − 2) .1. Example 2 Consider the Laplace transform F (s) = (s2 s .6. 6. We can derive this in an informal way using the Fourier integral. 6. The integral equation (6. (6. we note that F = 1/s and L[K] = 1/s2 . if we treat t as time. in particular of partial diﬀerential equations. we obtain Y = F + L[y ∗ K] = F + Y L[K]. However. and hence is the Laplace transform of t cos t ∗ e2t = 0 e2τ cos(t − τ ) dτ = t 1 2 = t 0 e2τ ei(t−τ ) + e−i(t−τ ) dτ = ei(t−τ )+2τ e−i(t−τ )+2τ + 2(2 + i) 2(2 − i) 0 2 2t 1 e − (2 cos t − sin t) . Example 3: Volterra integral equations A Volterra integral equation can be written in the standard form t y(t) = f (t) + 0 y(τ )K(t − τ ) dτ for t > 0. Y (s) = 2 s −1 This gives us the solution. for which f (t) = 1 and K(t) = t. and hence Y = F . We will need an inversion formula.162 LAPLACE TRANSFORMS This is the product of the Laplace transforms of cos t and e2t . these techniques are often inadequate to invert Laplace transforms that arise as solutions of more complicated problems. 1 − L[K] t where F (s) = L[f ] and Y = L[y]. y(t) = cosh t.4) can therefore be written as y = f + y ∗ K.4 The Inversion Formula for Laplace Transforms We have now seen that many Laplace transforms can be inverted from a knowledge of the transforms of a few common functions along with the linearity of the Laplace transform and the ﬁrst and second shifting theorems.23). For example.4) The function K is called the kernel of the equation. is an integral over the history of the solution. . to solve y(t) = 1 + 0 (t − τ )y(τ ) dτ. and hence s . 5 5 This is somewhat easier than the partial fractions method. If we now take a Laplace transform of this equation. The integral is in the form of a convolution and. (5. Since h(t) is bounded as t → ∞.5) Note that the contour of integration is a vertical line in the complex s-plane. g(t) therefore has a Laplace transform G(s). γ−i∞ (6. sometimes called the Bromwich inversion integral. As b → ∞. which exists for Re(s) > γ. from the deﬁnition of the Laplace transform.6. as we would expect. . we arrive at the inversion formula. so that ds = −i dk. Consider the integral I(t) = 1 2πi est C β ds. If we now make the change of variable s = γ − ik. Example 1 We start with the simple case G(s) = β/(s − α). since |G(s)| is algebraically small when |s| 1. the semicircular part of the contour becomes large and. Since G(s) is only guaranteed to exist for Re(s) > γ. we conclude from Jordan’s lemma (see Section A6. since we have not shown that this change in the contour of integration is possible. if the contour can be closed in this way. On the straight part of the contour C. the result will depend crucially on the residues at any poles of est G(s). We now deﬁne h(t) = e−γt g(t)H(t). As we shall see in the following examples.5). as b → ∞ we recover the inversion integral (6. Finally.2. As we have seen. we ﬁnd that† g(t) = 1 2πi γ+i∞ est γ−i∞ 0 ∞ e−sT g(T ) dT ds. the Fourier integral (5. and hence the inverse Laplace transform is g(t) = βeαt .4) that the integral along the semicircle tends to zero. I(t) is equal to the sum of the residues at any poles of est G(s) enclosed by C.5) by closing the contour of integration using a large semicircle in the left half plane.4 THE INVERSION FORMULA FOR LAPLACE TRANSFORMS 163 Let g(t) be a function of exponential order. and hence that the straight boundary of C lies to the right of the pole at s = α. The residue theorem then shows that I(t) = βeαt . † This is the point at which the derivation is informal. g(t) = 1 2πi γ+i∞ est G(s) ds. Let’s assume that the contour encloses the simple pole of G(s) at s = α. s−α (6.23) exists. so that I(t) = g(t).6) where the closed contour C is shown in Figure 6. e−ikt e−(γ−ik)T g(T ) dT dk. By the residue theorem. and shows that h(t) = and hence that e−γt g(t) = 1 2π ∞ −∞ 1 2π ∞ −∞ e−ikt ∞ −∞ ∞ 0 eikT h(T ) dT dk. this contour lies to the right of any singularities of G(s). It is often possible to simplify (6. in particular with γ the smallest real number such that e−γt g(t) is bounded as t → ∞. the point s = 0 is a branch point. as expected. Choosing the contour C as we did in Example 1.2.164 LAPLACE TRANSFORMS Im(s) s = a + ib C Re(s) s = a − ib Fig. G(s) = 1/(s−α)3 . Example 2 Let’s try to determine the inverse Laplace transform of G(s) = 1/(s2 + 1). we again ﬁnd that g(t) is the sum of the residues at the two simple poles of est G(s). since G(s) = 1 2i 1 1 − s−i s+i . In this case G(s) has simple poles at s = ±i. and the deﬁnition of 1/2 . The contour C used to evaluate the inverse Laplace transform of β/(s − α). We can check this result by noting that L[t2 ] = 2 Γ(3)/s3 = 2/s3 and using the ﬁrst shifting theorem. and hence that. Example 4 Consider the inverse Laplace transform of G(s) = e−αs /s. s1/2 . 6. g(t) = 1 it e − e−it = sin t. The simplest way to calculate the residue of est G(s) at s = α is to note that est eαt eαt = e(s−α)t = 3 3 (s − α) (s − α) (s − α)3 1 1 + (s − α)t + (s − α)2 t2 + · · · 2 . Since G(s) contains a fractional power. 2i Example 3 We now consider a Laplace transform that has a nonsimple pole. and hence that g(t) = 1 t2 eαt . as shown in Figure 6. the integral along the straight contour AF tends to g(t) as b → ∞. by the inversion formula (6. s1/2 = + |s|eiθ/2 and the real part of s1/2 is positive. the integral around CB is zero. such as C. We need to introduce a branch cut in order to make G(s) single-valued. In this way. we ensure that we use the correct value of s1/2 on either side of the branch cut. so that. It is convenient to place the branch cut along the negative real axis.3. This contour also includes a small circle around the origin of radius . Because of this. (6. G(s) → 0 exponentially fast as b → ∞. On the circular arcs AB and EF . which avoids the branch cut.3. Im(s) A s = a + ib B E C D δ Re(s) s = a − ib F Fig. since G(s) has a simple pole there. We can parameterize these lines as s = xe±iπ respectively.6. if s = |s|eiθ with −π < θ < π. . by Cauchy’s theorem (see Appendix 6). we use CB . As before.4 THE INVERSION FORMULA FOR LAPLACE TRANSFORMS 165 G(s) is incomplete. The keyhole inversion contour. and hence the contributions to the integral from these arcs tend to zero. We conclude that g(t) = − lim 1 b→∞ 2πi →0 + BC CD + DE est−αs s 1/2 ds. since s1/2 has positive real part. 6. CB .5). Since est G(s) is analytic within CB . s1/2 = x1/2 eiπ/2 = ix1/2 . We cannot now integrate est G(s) along any contour that crosses the negative real axis.7) Let’s consider the contributions from the lines BC and DE. Along BC. used for inverting Laplace transforms with a branch cut along the negative real axis. and is often referred to as a keyhole contour. we can interchange the order of summation and integration.166 LAPLACE TRANSFORMS and hence. 2t1/2 . 2 (2n − 1) (2n − 2)(2n − 4) . Γ 2 1 2 (2n − 3)(2n − 5) . . (2n − 1)! n=1 ∞ (6. . . 5 2 1 .. . (2n − 1)! ∞ ∞ 0 e−X X n−3/2 dX = n=1 α t1/2 2n−1 using the change of variable X = xt. we note that ∞ I= 0 e−xt sin αx1/2 dx = x ∞ 0 e−xt x using the Taylor series expansion of sin αx1/2 . x (αx1/2 )2n−1 dx.8) In order to calculate this integral. . Since this series is uniformly convergent for all x. 3 · 2 · 1 √ √ 1 π π = n−1 = 2(n−1) . . . and hence + BC DE est−αs s 1/2 ds = 2i 0 e−xt sin αx1/2 dx. . 3 · 1 2 = n−1 (2n − 1)! 2 (2n − 1)(2n − 2) . since Γ n− 1 2 = n− = we ﬁnd that 3 2 1 2n−1 Γ n− 3 2 = n− 3 2 n− √ π. 3 · 1 · √ Γ n− 1 π (2n − 3)(2n − 5) . 4 · 2 2 (2n − 1)(n − 1)! and hence ∞ I=2 n=1 α 2t1/2 2n−1 √ π =2 π (2n − 1)(n − 1)! α/2t1/2 0 √ α/2t1/2 ∞ 0 (s2 )n−1 ds (n − 1)! n=1 √ =2 π e−s ds = π erf 2 α . Now.. Similarly 1/2 ∞ ds = 0 DE e−xt+αix x ∞ 1/2 dx. in the limit as → 0 and b → ∞ est−αs s est−αs s 1/2 0 ds = ∞ BC e−xt−αix x 1/2 dx. so that I= α2n−1 (2n − 1)! n=1 ∞ ∞ 0 ∞ e−xt xn−3/2 dx = 1 α2n−1 (2n − 1)! tn−1/2 n=1 Γ n− 1 2 . For a one-dimensional ﬂow. This is a uni-directional ﬂow with velocity u(x. we conclude that L−1 e−αs s = 1 − erf α 2t1/2 ∞ x = erfc α . as shown in Figure 6. Balancing forces in the x-direction. ux + vy = 0. the kinematic viscosity. t) in the x-direction only and associated scalar pressure ﬁeld p(x. Example 5: Flow due to an impulsively-started ﬂat plate Let’s consider the two-dimensional ﬂow of a semi-inﬁnite expanse of viscous ﬂuid caused by the sudden motion of a ﬂat plate in its own plane. As → 0 we therefore have −π ds = π i dθ = −2πi. To be consistent with u = u(y.9) CD If we now use (6. so that est−αs s 1/2 −π ds = π e teiθ −α 1/2 iθ/2 e CD eiθ est−αs s 1/2 1/2 i eiθ dθ = −π e π teiθ −α 1/2 iθ/2 e i dθ. δt → 0.8) and (6.4. it is found experimentally that τ = µ∂u/∂y. ∂x ∂y where ρ is the density and τ is the shear stress. This reduces the x-momentum equation to ut = νuyy . where µ is the dynamic viscosity (see Acheson. u. On this curve we have s1/2 = 1/2 eiθ/2 . we ﬁnd that ρ Du =ρ Dt ∂u ∂u +u ∂t ∂x =− ∂p ∂τ + . 2 We can parameterize the small circle CD using s = eiθ . which we derived in Chapter 2. as we shall see in the following example. t). t). 1990). where we have introduced the quantity ν = µ/ρ. Here we have used the convective derivative. We will use Cartesian coordinates with the x-axis lying in the plane of the plate and the y-axis pointing into the semi-inﬁnite body of ﬂuid. We now consider the streamwise momentum within a small element of ﬂuid. is solely a function of y and t. Flows with high values of ν are extremely . shows that the streamwise velocity. (6.6. we now insist that ∂/∂x ≡ 0.4 THE INVERSION FORMULA FOR LAPLACE TRANSFORMS 167 where erf(x) is the error function. deﬁned by 2 erfc(x) = √ π e−q dq. which we derived in Chapter 2. deﬁned by 2 erf(x) = √ π x 0 e−q dq. and taking the limit δx. where θ runs from π to −π. The continuity equation. y.9) in (6. δy. 2t1/2 where erfc(x) is the complementary error function. y.7). 2 Laplace transforms of the error function and complementary error function arise very frequently when solving diﬀusion problems. t) p (x.10) This is known as Rayleigh’s problem. The x-momentum balance on a small element of ﬂuid. y. Our ﬁnal initial–boundary value problem is ∂2u ∂u = ν 2. we will take a Laplace transform with respect to time. u → 0.4. lava and mucus.13) (6. for example tar. whilst the values for air and water at room temperature and pressure. . t) p (x + δx. y. y. Equation (6.10) states that the only process involved in this ﬂow is the diﬀusion of x-momentum into the bulk of the ﬂuid.168 LAPLACE TRANSFORMS τ (x. suddenly heated up at one end. ∂t ∂y to be solved subject to u = 0 when t = 0 for y > 0. (6. u = U.12) (6. water and inert gases. t) τ (x.5 × 10−5 m2 s−1 .11) (6. t)] = U (y. so that L[u(y. s) = 0 ∞ e−st u(y. at y = 0 for t > 0. the kinematic viscosity of lava is around 10 m2 s−1 . For example. as y → ∞ for t > 0. a semi-inﬁnite bar of metal. 6. for example. y + δy. Of course. whilst those with low viscosity include air. 10−6 m2 s−1 and 1. t) dt. are very similar. respectively. In order to solve this initial–boundary value problem. insulated along its sides. t) Fig. viscous. this initial–boundary value problem can also be thought of as modelling other diﬀusive systems. which gives ∞ U (0. which therefore has the solution U (y.16) Using (6. (6.18) .2) to determine the Laplace transform of ∂u/∂t. t)e−st dt = 0 ∞ Ue−st dt = U . and hence U (y.17) We can now invert this Laplace transform using the convolution theorem.15) The Laplace transform of this condition is U (0. s) = U {L[f (t)] + f (0)} L erfc .6. We can also consider what happens when the velocity of the plate is a function of time. Since u → 0 as y → ∞. After using (6. e s From the result of the previous example. to give t u(y. Now B(s) = F (s). at y = 0 for t > 0. (6.4 THE INVERSION FORMULA FOR LAPLACE TRANSFORMS 169 Taking two derivatives of this deﬁnition with respect to y shows that L ∂2u ∂2U = . which becomes u = Uf (t). (6. (6. we must also have U → 0 as y → ∞. t) = U erfc y √ 2 νt . t) = U τ =0 {f (t − τ ) + f (0)} erfc y √ 2 ντ dτ .14) The transform variable only appears as a parameter in this equation.2). These are easy to plot. s and hence B(s) = U/s. y-derivatives are not aﬀected by a Laplace transform with respect to t. s) = A(s)es 1/2 y/ν 1/2 + B(s)e−s 1/2 y/ν 1/2 . We now need to transform the boundary condition (6. All we need to change is the boundary condition (6. we can see that sF (s) = L[f (t)] + f (0). s) = u(y. Some typical velocity proﬁles are shown in Figure 6. where F (s) is the Laplace transform of f (t). s) = UF (s). since the error and complementary error functions are available as erf and erfc in MATLAB. ∂y 2 ∂y 2 In general.5. and hence U (y.12).10) becomes sU = ν ∂2U . s) = UF (s)e−s 1/2 y/ν 1/2 = UsF (s) e−s 1/2 y/ν 1/2 s y √ 2 νt . s) = 0 u(0. so that A(s) = 0. We conclude that U −s1/2 y/ν 1/2 . we ﬁnd that U (y. (6. ∂y 2 (6.12). end plot(u. As a ﬁnal example. with f (t) = sin ωt.16). and then plots the solution at the points given in yout at the times given in tout.num2str(t))). t).y) nu = 1. (6.18) numerically.yout).0.19) 2πi γ−i∞ s2 + ω 2 where γ > 0. We proceed as we did in Example 4. started from rest. xlabel(’u’). ylabel(’y’) title(strcat(’t = ’.18) is the most convenient form to use for numerical integration. is$ ' function rayleigh(yout. let’s consider what we would do if we did not have access to a computer. since the integrand has poles at s = ±iω.tout) for t = tout u = []. integrand = df. In the example shown. The analysis . function f = f(t) df = 2*cos(2*t). we can calculate the ﬂow proﬁles for any f (t). gives √ γ+i∞ ωest−y s/ν 1 u(y.*erfc(y/2.170 LAPLACE TRANSFORMS Exact solutions of the equations that govern the ﬂow of a viscous ﬂuid. which corresponds to an oscillating plate. MATLAB is able to evaluate erfc(y/0) = erfc(Inf) = 0 for y positive. along with the inversion formula (6.3.t. f (t) = sin 2t.t. evaluating the integral on the contour CB shown in Figure 6.t. for y = yout u = [u quadl(@integrand. (6. are very useful for verifying that numerical solution methods are working correctly.10^-4. f = df + f0. Xlim([0 1]) pause(0.5) end function integrand = integrand(tau. but wanted to know what happens when the plate oscillates. t) = ds.0. (6. df = f(t-tau). f0 = 0. In fact./(eps+sqrt(nu*tau))).5). for any speciﬁed f (t). & % This uses the MATLAB function quadl to evaluate the integral correct to four decimal places. Although (6. Note that we add the inbuilt small quantity eps to the denominator of the argument of the error function to avoid MATLAB producing lots of irritating division by zero warnings. Their stability can also be studied more easily than for ﬂows where no analytical solution is available. A MATLAB function that evaluates u(y.y)].19) gives us the most helpful way of approaching this speciﬁc problem. By evaluating (6. the Navier– Stokes equations. Since F (s) = ω/(s2 + ω 2 ).

The second term. The ﬂow due to an impulsively-started ﬂat plate when U = 1 and ν = 1. t) = ω π ∞ 0 σ √ω e−σt sin νy dσ + e− 2ν y sin ωt − 2 + ω2 σ ω y . which comes from the branch cut integrals. This second term therefore gives the large time behaviour of the ﬂuid. and represents the initial transient that arises because the plate starts at rest. . tends to zero as t → ∞. except that now the integral around CB is equal to the sum of the residues of the integrand at the poles at s = ±iω. represents the oscillatory motion of the ﬂuid.5. which comes from the residues at the poles. whose phase changes with y and whose amplitude decays exponentially fast with y. 6.8).4 THE INVERSION FORMULA FOR LAPLACE TRANSFORMS 171 Fig.6. In addition. the integral around the small circle CD tends to zero as → 0. proceeds exactly as it did in Example 4. You can see what this solution looks like by running the MATLAB function that we gave above.20) The ﬁrst term. After taking all of the contributions into account (see Exercise 6. we ﬁnd that u(y. 2ν (6.

(f) 1/(s4 + 9s2 ). solve the integral and integro-diﬀerential equations 1 t (a) y(t) = t + y(τ )(t − τ )3 dτ . = x − 2y.5).1 Find the Laplace transforms of the functions. = 0.6 Determine the inverse Laplace transform of s F (s) = 2 (s + 1)(s − 2) using (a) the inversion formula (6.4 dy dx dy dx +x+ = 0. where F (s) = L[f (t)]. (b) sinh(ωt).172 LAPLACE TRANSFORMS Exercises 6. −4 dt2 dt dx d 2x dy + x + 2y = t. and (b) the residue theorem. subject to y(0) = 0. + 2t dt2 dt 6. solve the initial value problems (a) (b) (c) d 2y π + 9y = 18t. (g) (2s − 4)/(s − 1)4 . y dt2 2 = 0. (a) 1/(s2 − 3s + 2). (e) (2s+3)/(s2 −4s+20). 6. y (0) = 1. Find the inverse Laplace transforms of the functions. Specify the values of s for which each transform exists. 6 0 t dy (b) . (c) cosh(ωt). (h) (s2 + 1)/(s3 − s2 + 2s − 2).5 Using Laplace transforms. subject to x(0) = x (0) = 0. Using Laplace transforms.2 6. (d) (e) (f) 6. subject to y(0) = 1 and y (0) = 0. (b) 1/s3 (s+1). (i) s3 /(s + 3)2 (s + 2)2 . + 5x + 3 2 dt dt dt subject to x (0) = x(0) = y(0) = 0. (a) t(t+1)(t+2). (c) 1/(2s2 −s+1). d2 d − 4 + 4 y = f (t). (d) 1/s2 (s+1)2 . −x+2 = e−t . subject to y(0) = 1. subject to the initial conditions dt2 dt2 x(0) = y(0) = 1 and x (0) = y (0) = 0. 3. dt dt dt dt d 2x d 2y = −2x + y. (d) e−t sin t. with 2 dt dt f (t) = t if 0 t + 2 if t t 3. subject to y(0) = −2. Hence solve the initial value problem d 2x dx − 4x = 1. 6.3 d 2y dy + 3y = f (t). subject to x(0) = y(0) = 1. y(τ ) cos(t − τ ) dτ = dt 0 Show that L[tf (t)] = −dF/ds. .

Evaluate the integral (6. By expanding the denominator. g. subject to the initial conditions φ(y. 6. s) = s (1 + e−κL ) where κ = s/D. L]. Project The voltage.1). 0) = 0 and the boundary conditions φ(0. Consider the solution of the diﬀusion equation ∂φ ∂2φ =D 2 ∂t ∂y for t > 0 and y ∈ [0.19). so that it works when i(t) is supplied by an input routine input. and hence verify that (6. v(t).8 6. y(x.20) holds. of a uniform string with constant tension T and line density ρ subject to a uniform gravitational acceleration. ∂t2 ∂x where c = T /ρ (see Section 3. Use Laplace transforms to determine the solution when the string is ﬁxed at x = 0. Sketch the solution. (b) Write a MATLAB script that inverts the Laplace transform of this equation directly. and has y = yt = 0 for x 0 when t = 0. downwards satisﬁes ∂2y ∂2y = c2 2 − g. t) = 1. show that the solution can be written as ∞ φ(y.7 Find the inverse Laplace transform of F (s) = 1 .10 (a) Show that a small displacement.9 by (a) expressing F (s) as partial fractions and inverting the constituent parts and (b) using the convolution theorem. C (a) Solve this equation using Laplace transforms when i(t) = H(t − 1) − H(t).EXERCISES 173 6. 6. t). + dt2 R dt L dt where C is the capacitance. and explain how and why the qualitative form of the solution depends upon the sign of x − ct. R the resistance and L the inductance.m. Φ(y. t) = φ(L.11 . satisﬁes yx → 0 as x → ∞ and is allowed to fall under gravity. Show that the Laplace transform of the solution is 1 e−κy + e−κ(L−y) . (b) Such a string is semi-inﬁnite in extent. in an RLC circuit with implied current i(t) is given by the solution of the diﬀerential equation v di d 2v 1 dv + = . (s − 1)(s2 + 1) 6.9. t) = j=0 (−1)j erfc y + jL √ 2 Dt − erfc (j + 1)L − y √ 2 Dt .

174 LAPLACE TRANSFORMS (c) Compare your results for various values of the parameters R. (d) Extend this further to consider what happens when the implied current is periodic. . L and C.

since these must be of fundamental signiﬁcance. Properties and Complex Variable Methods for Second Order Partial Diﬀerential Equations In this chapter. Second Order Partial Diﬀerential Equations in Two Independent Variables Consider a second order linear partial diﬀerential equation in two independent variables. Each of these types of equation has distinctive properties. We can write this change . y)φ = f (x. for example Laplace’s equation.39). y) ∂x2 ∂x∂y ∂y (7. for example the wave equation. These mathematical properties are related to the physical properties of the system that the equation models. and not just a result of our choice of coordinate system. we will consider a class of partial diﬀerential equations.1) ∂φ ∂φ + d2 (x. the diﬀusion equation and the wave equation are the canonical examples. +d1 (x. y) 2 + 2b(x. y) + d3 (x. for example the diﬀusion equation.1) allow us to classify the equation into one of three distinct types: elliptic.13). with particular application to certain problems in ﬂuid mechanics. We will also demonstrate that complex variable methods are powerful tools for solving boundary value problems for Laplace’s equation. (2. y). (3. We will show that the ﬁrst three terms of (7. y) ∂2φ ∂2φ ∂2φ + c(x.1 Classiﬁcation We would like to know about properties of (7.1 Classiﬁcation and Properties of Linear. (2. parabolic. which we can write as a(x.CHAPTER SEVEN Classiﬁcation. 7.1) that are unchanged by an invertible change of coordinates. as we have already seen.12). 7.1. We will discuss what sort of boundary conditions are appropriate. and why solutions of these equations have their distinctive properties. or hyperbolic. y) ∂x ∂y Equations of this type arise frequently in mathematical modelling. of which Laplace’s equation.

6) This shows that the sign of ac−b2 is independent of the choice of coordinate system. The Jacobian is the factor by which the transformation changes inﬁnitesimal volume elements. a change of coordinates should not aﬀect its qualitative behaviour. . Substituting these into (7. we ﬁnd that 2 2 φx = ξx ψξ + ηx ψη . y)) .3) B = aξx ηx + b (ηx ξy + ξx ηy ) + cξy ηy . η) . η)ψ = g (ξ. with ∂ (ξ.4) We do not need to consider the other coeﬃcient functions here. We can express (7. (7. φxx = ξx ψξξ + 2ξx ηx ψξη + ηx ψηη + ξxx ψξ + ηxx ψη .4) in a concise matrix form as A B which shows that det A B B C = det a b b c ∂ (ξ. y) (7. η)ψη + b3 (ξ. which allows us to classify the equation. (7.1) gives us Aψξξ + 2Bψξη + Cψηη + b1 (ξ. (7. for example. ∂ (x.1) is a model for a physical system. if (7. Writing φ(x. where 2 2 A = aξx + 2bξx ξy + cξy . ∂x2 ∂y — A parabolic equation has ac = b2 . y) ξx ηx ξy ηy ∂ 2 φ ∂φ = 0. and similarly for the other derivatives. η) ∂ (x. y) → (ξ (x. Laplace’s equation ∂2φ ∂2φ + 2 = 0. for example.2) In particular. the diﬀusion equation K † Note that ∂ (ξ. y) . − ∂x2 ∂y is the Jacobian of the transformation. η) = 0. y) 2 B C = ξx ξy ηx ηy a b b c ξx ηx ξy ηy .176 CLASSIFICATION AND COMPLEX VARIABLE METHODS of coordinates as† (x. η)ψξ + b2 (ξ. η) and using subscripts to denote partial derivatives. (7. η) = det ∂ (x. 2 2 C = aηx + 2bηx ηy + cηy . η (x. — An elliptic equation has ac > b2 . y) ≡ ψ(ξ.5) .

and hence (p1 dy − q1 dx) (p2 dy − q2 dx) = 0. Tricomi’s equation. η)ψη + b3 (ξ. and hence A = C = 0.1 CLASSIFICATION OF PARTIAL DIFFERENTIAL EQUATIONS 177 — A hyperbolic equation has ac < b2 . We can then choose ξ and η so that p1 ξx + q1 ξy = p2 ηx + q2 ηy = 0. η)ψ = g (ξ.1) can be written in canonical form by choosing the canonical coordinate system. 2 2 C = aηx + 2bηx ηy + cηy = (p1 ηx + q1 ηy ) (p2 ηx + q2 ηy ) . in terms of which the second derivatives appear in the simplest possible way. or nonexistence.9) . ξ is constant on curves with dx p1 q2 dy = . η)ψξ + b2 (ξ. φxx = xφyy . As a less trivial example. For example. we can factorize A and C to give 2 2 A = aξx + 2bξx ξy + cξy = (p1 ξx + q1 ξy ) (p2 ξx + q2 ξy ) . is elliptic for x < 0 and hyperbolic for x > 0.7) As we shall see. of characteristic curves for the three types of equation that determines the distinctive properties of their solutions. or simply characteristics.8) The curves where ξ is constant and the curves where η is constant are called the characteristic curves.1) takes its canonical form ψξη + b1 (ξ. this is the easiest equation to use to determine (ξ. This means that q1 dy = .1. it is the existence. y)-plane. with the two factors not multiples of each other. As we shall see. equations of mixed type are possible. We call (ξ. (7. in terms of which (7. the wave equation 1 ∂2φ ∂2φ − 2 = 0.1.2 Canonical Forms Any equation of the form given by (7.7. (7. dx p2 η is constant on curves with We can therefore write p1 dy − q1 dx = p2 dy − q2 dx = 0. (7.9. 7. c2 ∂x2 ∂y Note that. although these three examples are of the given type throughout the (x. η) the characteristic coordinate system. consider the hyperbolic equation φxx − sech4 x φyy = 0. η) . which gives a dy 2 − 2b dx dy + c dx2 = 0. Hyperbolic Equations: ac < b2 In this case. We discussed the reduction of the wave equation to this canonical form in Section 3. for example. η).

η)ψξ + b2 (ξ. and we can take ξ = x + cos y as the characteristic coordinate.178 CLASSIFICATION AND COMPLEX VARIABLE METHODS Equation (7. dx The characteristic curves are therefore given by x + cos y = constant. with φ(x.10) so we can only construct one set of characteristic curves. η) .12) . A suitable choice for the other coordinate is η = y. This gives us A = 0 and. with φ(x. (η − ξ) (ψξ − ψη ) 4 − (ξ − η) 2 .9) in terms of these variables. 2 2 C = aηx + 2bηx ηy + cηy = (pηx + qηy ) . 2 2 2 dy − sech2 x dx = 0. η = y − tanh x. Parabolic Equations: ac = b2 In this case. (7. B = 0. η). For any set of curves where η is constant that is never parallel to the characteristics.11) (7. On writing (7. We can now see that the diﬀusion equation is in canonical form. we ﬁnd that its canonical form is ψηη = sin2 η cos η ψξ . since AC = B 2 . y) = ψ(ξ. C does not vanish.13) 2 (7.12) in terms of these variables. On writing (7. and the canonical form is ψηη + b1 (ξ. The characteristic curves satisfy dy 2 − 2cosec y dx dy + cosec2 y dx2 = (dy − cosec y dx) = 0. 2 2 A = aξx + 2bξx ξy + cξy = (pξx + qξy ) . and hence dy = cosec y.7) shows that the characteristics are given by dy 2 − sech4 x dx2 = dy + sech2 x dx and hence dy = ±sech2 x. in the whole (ξ. η). and the characteristic coordinates are ξ = y + tanh x. we ﬁnd that its canonical form is ψξη = in the domain (η − ξ) < 4. (7. We therefore take ξ to be constant on the curves p dy − q dx = 0. η)ψη + b3 (ξ. y) = ψ(ξ. dx The characteristics are therefore y ± tanh x = constant. η)ψ = g (ξ. consider the parabolic equation ψxx + 2cosec y φxy + cosec2 y φyy = 0. η)-plane. As a further example.

we can simplify by making A = C and B = 0. on √ b ∓ i ac − b2 dy = .15). we can make neither A nor C zero.18) (7. ∇2 ψ. (7. and the canonical form is ψξξ + ψηη + b1 (ξ. η)ψ = g (ξ. and noting that these two equations form the real and imaginary parts of aχ2 + 2bχx χy + cχ2 = 0. 1 − η2 (7. and hence.14) B = aξx ηx + b (ηx ξy + ξx ηy ) + cξy ηy = 0. η) . Laplace’s equation is in canonical form.17) in the domain |η| < 1.17) in terms of these variables. η = tanh x. In this case. x y and hence that √ −b ± i ac − b2 χx . dx and hence y ∓ i tanh x = constant. .15) Now χ is constant on curves given by χy dy + χx dx = 0. η)ψξ + b2 (ξ. so that the second derivatives form the Laplacian. from (7.16) we can deduce ξ and η. η). (7. = χy a (7. since no real characteristic curves exist. we must solve 2 2 2 2 A − C = a ξx − ηx + 2b (ξx ξy − ηx ηy ) + c ξy − ηy = 0. the reader is referred to Kevorkian (1990) and Carrier and Pearson (1988). Clearly. with φ(x. Instead.16) dx a By solving (7. On writing (7. We can do this by deﬁning χ = ξ + iη. χ = ξ + iη is constant on the curves given by dy = ±i sech2 x. We can therefore take χ = y + i tanh x. we ﬁnd that its canonical form is ψξξ + ψηη = 2η ψη . consider the elliptic equation φxx + sech4 x φyy = 0. η)ψη + b3 (ξ. y) = ψ(ξ. For more detailed information.7. and hence ξ = y.1 CLASSIFICATION OF PARTIAL DIFFERENTIAL EQUATIONS 179 Elliptic Equations: ac > b2 In this case. We can now describe some of the properties of the three diﬀerent types of equation. For example. In order to proceed.

1. for a formal statement and proof).43). y)-plane upon which we know the Cauchy data. We conclude that discontinuities in the second derivative can propagate on characteristic curves. Suppose that φξξ is not continuous on C. ∂φ/∂η(ξ. η − φξξ ξ0 . y). the use of the independent variable t.1). η ? If we evaluate the equation on either side of the curve and subtract. and hence C is a characteristic curve. and that the functions that appear as coeﬃcients in (7. Indeed. we saw in Section 3. We can therefore satisfy the equation if A(ξ0 . a local existence theorem (see Garabedian. 0) + η 1 ∂2φ ∂φ (ξ.9. which is then parameterized by ξ. discontinuities in the second derivative of φ can propagate on characteristic curves. In general. instead of y suggests that the equation has an evolutionary nature. In a Cauchy problem. To see this. which corresponds to propagation on the characteristic curves. 0). not necessarily a characteristic. we can write (7. More speciﬁcally. For the wave equation. The eﬀects of the initial data propagate into the (x. and hence construct a power series solution. but that φ satisﬁes the hyperbolic equation on either side of C. ∂η 2! ∂η where the orthogonal coordinate system (ξ. 0) from (7. We can then write φ(ξ. y) = ξ0 . In addition. we can also deduce higher derivatives from derivatives of (7. 0) + η 2 2 (ξ. Can we choose ξ + − in such a way that the equation is satisﬁed. η) φξξ ξ0 . η = 0.3). time. The initial value problem for the wave equation that we studied in Section 3. consider the Cauchy problem for a hyperbolic equation of the form (7.3 Properties of Hyperbolic Equations Hyperbolic equations are distinguished by the existence of two sets of characteristics. We know φ(ξ. (3. we ﬁnd that + − A(ξ0 . Firstly. and. we specify a curve C in the (x. ∂φ/∂n. 1964.1) can also be expanded as power series in the neighbourhood of C.3). This allows us to establish two key properties. as shown in Figure 7. even though φξξ ξ0 . η) is set up so that η = 0 is the curve C. For the wave equation. so it is inconsistent to specify initial data upon a characteristic curve.9. When does the coeﬃcient of ∂ 2 φ/∂η 2 vanish on C? Precisely when C is a characteristic curve. consider a curve C in the (x. y) is only dependent on the initial conditions that lie between the two characteristics through (x. the solution at any point (x. given by ξ(x. η = φξξ ξ0 . x ± ct = constant. but for . 0) and the normal derivative. Does this problem have a solution in general? Let’s assume that φ and ∂φ/∂n on C can be expanded as power series. 0) + · · · . As we have seen. η) = 0. we can deduce ∂ 2 φ/∂η 2 (ξ. valid in the neighbourhood of C.1 that solutions propagate at speeds ±c. y)-plane on the characteristics. φ and the derivative of φ normal to C. The formal statement of this informally presented procedure is the Cauchy– Kowalewski theorem.3).1) in terms of this new coordinate system as (7. characteristics are carriers of information. Provided that C is not a characteristic curve. Secondly.180 CLASSIFICATION AND COMPLEX VARIABLE METHODS 7. y)-plane.1. η) = φ(ξ. this is immediately obvious from d’Alembert’s solution.1 is a Cauchy problem. φ and its ﬁrst derivatives are continuous. provided that the coeﬃcient of ∂ 2 φ/∂η 2 does not vanish on C.

usually arise naturally. or initial value.10. there is no timelike variable. rather than initial value problems. In fact.7.1 CLASSIFICATION OF PARTIAL DIFFERENTIAL EQUATIONS 181 Characteristics y (x.1. the wave equation. 0) = v0 (x). For physical problems that can be modelled using elliptic equations. boundary value problems. physical problem. Laplace’s equation is written in terms of the spatial variables.4 Properties of Elliptic Equations For elliptic equations. for example the steady state diﬀusion problem discussed in Section 2.1. Consider the Cauchy. 0. 7.6. the solution of the Cauchy problem for elliptic equations does not depend continuously on the initial conditions. C. y) depends only upon the initial data on the marked part of the initial line. φt (x. y) x Initial line C Fig. 7.20) . The domain of dependence of the solution of a hyperbolic equation.1. Solutions are therefore inﬁnitely-diﬀerentiable. † We say that the problem is ill-posed. (7. The solution at (x.19) (7. −∞ < x < ∞. x and y. for example.† We can easily demonstrate this for Laplace’s equation. A and C are never zero. and is therefore not a sensible representation of any real. 0) = φ0 (x). problem φxx + φtt = 0 for t subject to φ(x. Moreover. and there are no characteristics. as shown in Figure 3. discontinuities in φ and its derivatives can also propagate on characteristics. in which only second derivatives appear.

Now. For example in two dimensions. and hence that ψ must attain its maximum value on the boundary of D. we conclude that ψ has no local maximum in the interior of D. say at x = x0 . we do need to do a little work in order to exclude this possibility in general. we can make arbitrarily small. Let’s consider some more of its properties. which also satisﬁes Laplace’s equation. the boundary of D. the closer Φt (x. and. Let ψ = φ + |x|2 . and nowhere in the interior of D. at 0 and φyy 0. and φ cannot be a solution of Laplace’s equation. However. Consider the function ˆ ˆ ˆ ˆ Φ = Φ + φ. as we have seen. we have ∇2 φ = φxx + φyy < 0. y) = (x0 . unless φ is a constant. and conclude that φ(x) max φ for all x ∈ D. An arbitrarily small change in the boundary data produces an arbitrarily large change in the solution. However. with > 0 and |x|2 = x2 + y 2 in two dimensions and 2 |x| = x2 + y 2 + z 2 in three dimensions. . Since ∇2 ψ 0 at an interior local maximum. The larger the value of λ. arises in many diﬀerent physical contexts. let Φ(x. it is possible to have φxx = φyy = 0 at an interior local maximum. attains its maximum value on ∂D.19) subject to (7. ∂D where l is the greatest distance from ∂D to the origin. But. Then φ attains its maximum and minimum values on ∂D. |Φ| increases without bound for t > 0 as λ increases. t). a solution is φ = φ(x. if the Taylor series expansion close to 4 4 (x. λ 1 1 ˆ ˆ As λ → ∞. Theorem 7.182 CLASSIFICATION AND COMPLEX VARIABLE METHODS If φ0 = 0 and v0 = 1 λ ˆ sin λx. by deﬁnition φ(x) ψ(x) ψ(x0 ) = φ(x0 ) + |x0 |2 max φ + l2 .20). φxx maximum with φxx < 0 or φyy < 0. Although this is clearly not a solution of Laplace’s equation. At a a local maximum in the interior of D. and hence φ attains its minimum value on ∂D. However. and φ a solution of Laplace’s equation in D. Laplace’s equation is the simplest possible elliptic equation. bounded. t) be the general solution of (7. φx = φy = 0. The smaller the initial condition.1 (The maximum principle for Laplace’s equation) Let D be a connected. 0) = λ sin λx → 0. Since this is true for all > 0. Proof The idea of the proof is straightforward. ∂D Similarly. We then have ∇2 ψ = ∇2 φ + ∇2 |x|2 = k > 0. 0) is to v0 (x). max |φ| = λ2 sinh λt → ∞ as λ → ∞. φt (x. where 1 ˆ φ(x. where k = 4 in two dimensions and k = 6 in three dimensions. y0 ) is φ = (x − x0 ) + (y − y0 ) + · · · . the larger the solution. for example. for t > 0. −φ. open set in two or three dimensions. t) = 2 sin λx sinh λt.

Theorem 7.1 CLASSIFICATION OF PARTIAL DIFFERENTIAL EQUATIONS 183 We can use the maximum principle to show that the Dirichlet problem for Laplace’s equation has a unique solution.22). The maximum value is clearly a2 /4. and that ˆ 0) = φ(0. so that ∇2 φ = 0.21) How big is φ(0. subject to φ = 0 on ∂X. min ψ ∂X φ(0. Since X is symmetric about both coordinate axes. 0). with ψ(x) = 0 on ∂D. 0) a2 . y) |x| |y| + <1 a b .25) (7.21) subject to (7. (7. ∂X We can therefore bound φ(0.22) (7. 4 Finally.7. the other type of boundary value problem that often arises for Laplace’s . y > 0.1 then shows that φ(0. We can also see that φ = ψ on ∂X.23) (7. and hence that there is unique solution. and hence that a2 b2 4(a2 + b2 ) φ(0.26) (x. 0) max ψ. consider the boundary value problem ∇2 φ = −1 for x ∈ X = with a > b > 0.25) into Laplace’s equation. with φ(x) = f (x) on ∂D.24) By Theorem 7. (7. we can ﬁnd the minimum value of ψ by 2 determining the value of the radius r0 for which the circle x2 + y 2 = r0 just touches the straight line x/a + y/b = 1. We deﬁne ψ = (x2 + y 2 )/4. so that 0 ψ(x) 0 for x ∈ D.2 The Dirichlet problem ∇2 φ = 0 for x ∈ D. 0) using the maximum and minimum values of ψ = (x2 + y 2 )/4 on the boundary of X. which forms the boundary of X in the quadrant 2 x > 0. has a unique solution. Proof Let φ1 and φ2 be solutions of (7. The maximum principle can also be used to study other elliptic equations. For example. 0)? We can obtain some bounds on this quantity by transforming (7. This means that φ1 = φ2 . and hence ψ ≡ 0. Theorem 7. so that ∇2 ψ = 1. A little algebra shows that r0 = a2 b2 /(a2 + b2 ). (7. ψ attains its maximum and minimum values of ∂D. Then ψ = φ1 − φ2 satisﬁes ∇2 ψ = 0 for x ∈ D. and ˆ ˆ ˆ let φ = φ + ψ.1.

4. Kφxx = φt with K > 0. Then φ attains its maximum value on x = 0. In this case. Such problems can only be solved if the boundary data satisﬁes a solubility condition. 7. or x = L. such a ﬂow can only exist if the total ﬂux into D through ∂D is zero. For example. any localized disturbance is therefore felt everywhere in −∞ < x < ∞ instantaneously. We can also prove a maximum principle for the diﬀusion equation. irrotational ﬂuid ﬂow.2. Consider the domain 0 x L.5 Properties of Parabolic Equations Parabolic equations have just one set of characteristics. ∂D (7.29). with ∂φ (x) = g(x) on ∂D. we ﬁnd that ∇2 φ d 3 x = D ∂D n. the Neumann problem speciﬁes a steady ﬂow in D with the normal velocity of the ﬂuid on ∂D given by g(x). Since the ﬂuid is incompressible. as we can see from Example 5 of Section 6. As we have seen. Proof This is very similar to the proof of the maximum principle for Laplace’s equation. incompressible. as expressed by (7. Theorem 7. solutions are inﬁnitely-diﬀerentiable with respect to x.1. where the normal derivative of φ is speciﬁed on the boundary of the solution domain. . ∂n is g(x) d 2 x = 0.6. Theorem 7.8(a). 0 t T . 0 t T .4). and we leave it as Exercise 7.27) (7.27) over D and use the divergence theorem and (7. This solubility condition has a simple interpretation in terms of inviscid.∇φ d 2 x = ∂D g(x) d 2 x = 0. 0 t T .184 CLASSIFICATION AND COMPLEX VARIABLE METHODS equation is the Neumann problem (see Section 5. t is constant on the characteristic curves.29) Proof If we integrate (7.4 (The maximum principle for the diﬀusion equation) Let φ be a solution of the diﬀusion equation. or t = 0.28). for the diﬀusion equation. 0 x L. which we introduced in Section 2.3 A necessary condition for the existence of a solution of the Neumann problem ∇2 φ = 0 for x ∈ D ⊂ R3 . In addition.28) (7.

33).33) x L. with ∂φ/∂n = 0 on ∂D.1 CLASSIFICATION OF PARTIAL DIFFERENTIAL EQUATIONS 185 We can use this maximum principle to show that the solution of the initial– boundary value problem given by ∂φ ∂2φ = K 2 for t > 0. consider the case f = φ(1 − φ). is unique. t for x ∈ D. ¯ φt ¯ ¯ K∇2 φ + f φ. If 0 φ(x. and that for a canonical example. and secondly φ = φ. t for x ∈ D. (7. we ﬁnd that 0 φ(x. If φ(x. φ = 0. We leave the details as Exercise 7. (7. φ1 and φ2 prescribed functions. so that parabolic equations share some of the features of both hyperbolic and elliptic equations. αφ − β ∂n ∂n and ¯ φ(x. If there also exist constants α and β with α2 + β 2 = 0. We will meet reaction–diﬀusion equations frequently in Part 2. 0) = φ0 (x) for 0 and φ(0. t > 0. t) for x ∈ D. for further details). 1991. x. . φ(L. t > 0. we say that φ is a subsolution of (7. t) 1 for t 0. which are parabolic. Similarly. t > 0. φ = φ. the diﬀusion equation.7. t > 0. such that ¯ ∂φ ∂φ ¯ αφ − β for x ∈ ∂D. ∂t ∂x subject to φ(x. 2001). We have now seen that parabolic equations have a single set of characteristics. This problem arises in a model for the propagation of chemical waves (see Billingham and King. We will not prove this result here (see Grindrod. t) for x ∈ D. x. t) is a bounded function that satisﬁes φt K∇2 φ + f φ. x. t > 0. there is a maximum principle.31) (7.30) ¯ with K > 0 and f a smooth function.33). To see how this theorem can be used.8(b). We end this section by stating a useful theorem that holds for reaction–diﬀusion equations. 0) φ(x. t) φ(x. then by taking ﬁrstly ¯ ¯ φ = 1. 0 < x < L.32) with φ0 . 0) 1. t) = φ1 (t). if φ(x. (7. t) = φ2 (t) for t > 0. t) is a bounded function that satisﬁes ¯ we say that φ is a supersolution of (7.5 (A comparison theorem for reaction–diﬀusion equations) Consider the reaction–diﬀusion equation φt = K∇2 φ + f (φ. Theorem 7. then ¯ φ(x. 0) for x ∈ D.

Elementary calculus shows that the stream function is constant on any streamline in the ﬂow. such as aircraft wing sections.2 Complex Variable Methods for Solving Laplace’s Equation In Section 2. ψ(x.186 CLASSIFICATION AND COMPLEX VARIABLE METHODS In other words. imply that w(z) is an analytic function of the complex variable z = x + iy in the domain occupied by the ﬂuid. which. the comparison theorem allows us to determine upper and lower bounds on the solution (see Exercise 7. deﬁned by u = ψy . Once φ is known. Let’s consider some examples. incompressible. (i) A uniform stream ﬂowing at an angle α to the horizontal has u = U cos α. If we now look at the deﬁnitions we have for the components of velocity. we can determine the pressure using Bernoulli’s equation. with our smoothness assumption. for example. and the change in ψ between any two streamlines is equal to the ﬂux of ﬂuid between them. for compatibility we need φx = ψy and φy = −ψx .6. satisﬁes the continuity equation identically. (2.6. 7.2 we described how steady. inviscid. 7. so that dw = u − iv = U e−iα . ux + vy = 0. and vice versa. φy ).1 The Complex Potential Recall from Section 2. These are the Cauchy–Riemann equations for the complex function w = φ + iψ and. a uniform stream.18). can be found easily.2 that the velocity in a two-dimensional ideal ﬂuid ﬂow. so that. 1960). with ∂φ/∂n = 0 on the boundary of the body. dz † See Appendix 6 for a reminder of some basic ideas in the theory of complex variables. . where φ is the velocity potential and u = ∇φ the velocity ﬁeld. give simple descriptions of the ﬂow past many simplyconnected body shapes. v) = (φx . irrotational ﬂuid ﬂow. except for the simplest body shapes. for example.† The quantity w(z) is called the complex potential for the ﬂow and. dz where q is the magnitude of the velocity and θ the angle that it makes with the x-axis. Milne-Thompson. ∇2 φ = 0. for suﬃciently smooth functions. since powerful complex variable methods can. in the ﬂuid. for simple ﬂows. without too much eﬀort. we can easily compute the components of the velocity. We can also see that dw = φx + iψx = u − iv = qe−iθ . y). u = (u. However. solutions of this boundary value problem are hard to ﬁnd. commonly referred to as ideal ﬂuid ﬂow. once we know w. We can therefore introduce a stream function. twodimensional ﬂow is an exception. v = −ψx .2. v = U sin α. In general. past a rigid body can be modelled as a boundary value problem for Laplace’s equation. satisﬁes the continuity equation.9 for another example). Appropriate conditions at inﬁnity also need to be prescribed. channels with junctions and ﬂows with free surfaces (see.

For example. C. In each case. close to z = z0 . as expected. The streamlines are therefore given by ψ = constant. using the methods that we will discuss in Chapter 12. w= 2π for a source at z = z0 . the points where the complex potential is not analytic must be excluded from the ﬂow domain. and we need to include it. consider a uniform stream of magnitude U ﬂowing parallel to the x-axis and a source of strength 2πm situated . where r is the polar coordinate centred at z = z0 . 7. By taking real and imaginary parts. we need to invoke some extra physics close to this point. this can be done by deﬁning an inner asymptotic region. The ﬁrst of these complex potentials is analytic over the whole z-plane. since m 1 dw = qe−iθ = .7. Its complex potential is m log(z − z0 ). if there are no sources or vortices in the ﬂow domain that we are considering. In reality. dz 2π z − z0 we have that q = m/2πr. Alternatively. we need to take this into account. For a point vortex. The streamlines are straight lines passing through z = z0 and. 2π If we now deﬁne the circulation to be C u·dr. ψ = κ log r/2π. the point vortex has circulation κ. For example. that encloses z0 is u · n dl = m. It is simple to show that the ﬂow through any closed curve. and that the ﬂow is purely radial. (ii) A point vortex is a system of concentric circular streamlines centred on z = z0 . where C is any closed contour that encloses z = z0 . where r is the polar coordinate centred on z = z0 . with complex potential w= iκ log (z − z0 ) . (iii) A point source of ﬂuid at z = z0 ejects m units of ﬂuid per unit area per unit time.2 COMPLEX VARIABLE METHODS FOR SOLVING LAPLACE’S EQUATION 187 and hence w = U e−iα z. a family of concentric circles. but we wish to use sources and vortices to model the ﬂow. close to the singularity the eﬀect of viscosity becomes important. any real source of ﬂuid will be of ﬁnite size. A word of warning is required here.2. complex potentials can be added together to produce more complicated ﬂuid ﬂow ﬁelds. and hence r = constant. we can see that φ = −κθ/2π. C as expected. The second and third fail to be analytic at z = z0 . and. We will see an example of this in the next section.2 Simple Flows Around Blunt Bodies Since Laplace’s equation is linear.

34) therefore represents the ﬂow past a blunt body of the form given by (7.35).34) with U = m = a = 1.5 −2 −2 −1. We would prefer a . whilst for x > 0 there is another streamline with ψ = 0. dw/dz = 0. and hence both components of the velocity vanish there. and can be disregarded.5 2 Fig.5 BLUNT BODY 0 y −0. There is a stagnation point at the origin. The complex potential is w = U z + m log(z − a).2. As there is no ﬂow across a streamline. At the point z = a − m/U .188 CLASSIFICATION AND COMPLEX VARIABLE METHODS on the x-axis at z = a. The streamlines are shown in Figure 7. However. we can replace it with the surface of a rigid body without aﬀecting the ﬂow pattern. it has its drawbacks. so that the velocity ﬁeld associated with it is an even function of y. since it is an inverse method (for a given ﬂow ﬁeld. given by Uy . Adding potentials is one way to construct ideal ﬂuid ﬂows around rigid bodies. The complex potential (7. (7. This is called a stagnation point of the ﬂow.5 −1 −0. 7.35) y = (a − x) tan m which is a blunt-nosed curve.34) x−a This is an odd function of y.5 1 0.5 1 1. Note that the singularity due to the point source does not lie within the ﬂow domain. The ﬂow past a blunt body given by the stream function (7. as we would expect from the symmetry of the ﬂow.2. The x-axis (y = 0) is a streamline with ψ = 0. we can introduce a rigid body bounded by any streamline). from which we can deduce that y ψ = U y + m tan−1 . (7. 2 1.5 −1 −1.5 0 x 0.

The force on this inﬁnitesimal arc is due to the pressure. since f (z) is analytic in |z| < a. Let the tangent to this element of arc make an angle θ with the x-axis. which we discussed in Section 2. z When α = 0. When α = 0. ψ = 0 on z = aeiθ . as we shall now show. By the circle theorem. is real. we need the circle theorem. By taking real and imaginary parts. this gives the potential w=U z+ a2 z + iκ log z. .3). causes a nonzero lift force on the body. f ∗ (a2 /z) is analytic in |z| > a. Finally. potential and stream function. since w ∼ f (z) + f ∗ (0) as |z| → ∞.2. as it breaks the symmetry (see Figure 7. Before we describe such a method. the ﬂow in the far ﬁeld is given by f (z). we note that the complex potential is not unique. 2π (7. ds. and still have the circle as a streamline. and. which. with dw =U dz 1− a2 z2 . The complex potential for the stream is w = U e−iα z. and the boundary of the circle is a streamline.6. since we can add a point vortex. without introducing a singularity in the ﬂow domain.36) The introduction of this nonzero circulation about the circle has important consequences. Secondly. the complex potential for the ﬂow when the circular boundary is introduced is w = U e−iα z + U eiα a2 . we can recover the velocity ﬁeld.6 (The circle theorem) If w = f (z) is a complex potential for a ﬂow with no singularities in the domain |z| a. as expected. consider an inﬁnitesimal element of arc on the surface of the body. centred upon z = 0. z = aeiθ with 0 θ 2π. Theorem 7.7. whereby we can calculate a stream function for a given rigid body. the ﬂow is symmetric about the x-axis. and given by −p sin θ ds + ip cos θ ds = ipeiθ ds. as it is the sum of complex conjugate functions.2 COMPLEX VARIABLE METHODS FOR SOLVING LAPLACE’S EQUATION 189 direct method. Proof Firstly. Therefore. Before leaving this problem. as a2 /z is just an inversion of the point z. on the boundary of the circle. which has no singularities in |z| a. and dw/dz ∼ U as |z| → ∞. so that f (z) + f ∗ (a2 /z) = f (aeiθ ) + f ∗ (ae−iθ ). In order to calculate the force exerted by an ideal ﬂuid ﬂow upon a rigid body. then w = f (z) + f ∗ (a2 /z) is a complex potential for the same ﬂow obstructed by a circle of radius a centred at the origin. We can illustrate this theorem by ﬁnding the complex potential for a uniform stream of strength U ﬂowing past a circle of radius a.

(2. Flow past a circle.3. dz dz ∗ dz ∗ = dw ∗ dw . it is convenient to manipulate this. the total force on the body is Fx + iFy = body ipeiθ ds = body ip dz. since this is equal to the tangential angle at the surface of the rigid body. Now. Fx − iFy = Now note that dw dz so that dw dz 2 2 1 ρi 2 body dw dz 2 dz ∗ . = dw dz dw dz ∗ = dw dw∗ . taking the complex conjugate.18). 7. dz . p = p0 − 1 ρq 2 and qe−iθ = dw/dz. Since dz = dx + i dy = cos θ ds + i sin θ ds = eiθ ds. with no circulation and with κ = 20.190 CLASSIFICATION AND COMPLEX VARIABLE METHODS κ=0 5 5 κ = 20 0 0 y −5 −5 y 0 x 5 −5 −5 0 x 5 Fig. dz. 1 Fx + iFy = − ρi 2 dw dz 2 1 dw i p0 − ρ 2 dz body 2 dz. where θ 2 is the angle of the ﬂow. so that Fx + iFy = As p0 is a constant. from Bernoulli’s equation. Firstly. body To get a more usable result.

we could induce an inﬁnite velocity in the z-plane. Conformal mappings are so called because they also preserve the angle between line segments except at any isolated points where f (z) is zero or inﬁnity (see Exercise 7. such as a half-plane or a circle. the complex potential in the z-plane. and a unique inverse transformation can be deﬁned. it is also the circulation around an aerofoil that provides the lift. dz dζ dz dζ If |f (z)| is bounded and nonzero. (7. we have a correspondence between ﬂow and geometry in both planes. z-plane. which is unphysical. where the shape of the body is simpler. For example. dw∗ = dw. After evaluating the residue. ζ = f (z). Any streamline in the z-plane transforms to a streamline in the ζ-plane because of this correspondence in complex potentials. and we have a transformation. If we can transform this problem into a new plane. if we seek w = w(z). by deﬁning W (ζ) = w(z).7. Specifically. we may be able to solve for the ﬂow in the ζ-plane. except perhaps for isolated points on the boundary of the domain to be transformed. For the example of ﬂow around a circle with circulation κ. which tries to lift the body in the direction of increasing y. 7. and hence Fx − iFy = 1 ρi 2 dw dz 2 dz. we must have f (z) equal to zero or inﬁnity at these corner points. we have Fx − iFy = 1 ρi 2 U body 1− a2 z2 + iκ 2πz 2 dz. This lift force arises from the pressure distribution around the circle.37) body In this form. the ζ-plane.3 Conformal Transformations Suppose we have a ﬂow in the z-plane past a body whose shape cannot immediately be seen to be the streamline of a simple ﬂow. there is a practical diﬃculty with this . which is just due to the simple pole.2. and the ﬂow outside it. if |f (z)| = ∞. and then invert the transformation to get the ﬂow in the original. The complex velocity is dw dζ dw dw = = f (z). which maps the surface of the rigid body. the force can usually be evaluated using the residue theorem. with pressures at the lower surface of the body higher than those on the upper surface. Such points can cause diﬃculties. A consequence of this is that if we want to map a domain whose boundary has corners to a domain with a smooth boundary. onto a ﬂow outside a half-plane or circle in the ζ-plane. we ﬁnd that Fx − iFy = −iρU κ. we say that the transformation between the z. As we shall see in Example 4 in the next section. so that there is a vertical force of magnitude ρU κ. Although beautifully simple in principle. so that dw∗ = (dw/dz)dz.10). then.and ζ-planes is conformal.2 COMPLEX VARIABLE METHODS FOR SOLVING LAPLACE’S EQUATION 191 Since the boundary of the body is a streamline.

then z ranges from i − ∞ to i + ∞. If we now choose iπK = 2i. Example 2: Mapping a strip onto a half-plane Consider the strip of width 2 that lies in the z-plane shown in Figure 7. Mapping a three-quarter-plane to a half-plane. to have arg ζ = π. dictionaries of common conformal transformations are available (see. MilneThompson. How can we map this onto the half-plane in the ζ-plane shown in Figure 7. If we further require that the image of z = −1 should be ζ = −1. since the angle of the boundary changes there. arg A = π/3. Note that this transformation is not conformal at z = 0.5(b)? Let’s try the transformation z = K log ζ + L. On C D . the choice L = −i makes z range from −i − ∞ to −i+∞. to have arg ζ = 0. f (z)? Fortunately. arg z = π. and we require B C .4. so z = K log |ξ|+iπK −i. z = K log |ξ| + L. arg ζ = 0. arg z = −π/2.192 CLASSIFICATION AND COMPLEX VARIABLE METHODS method: how can we construct the transformation. On BC. the image of AB. How can we map this onto the half-plane in the ζ-plane shown in Figure 7. arg ζ = π. This means that the family of transformations ζ = |A|eiπ/3 z 2/3 will map the three-quarter-plane to the half-plane. The transformation is therefore z= 2 log ζ − i. the image of BC. On AB. and ζ = eiπ/3 z 2/3 . This means that π = arg A + nπ and hence n = 2/3. as required. as required. then |A| = 1. ζ = ieπz/2 . Let’s try to make things clearer with some examples. If K is real. π which is conformal in the ﬁnite z-plane. 1952).4(b)? Let’s try the transformation ζ = Az n . so if ζ = ξ + iη.4(a). 7. Example 1: Mapping a three-quarter-plane onto a half-plane Consider the three-quarter-plane that lies in the z-plane shown in Figure 7. for example.5(a). . so that arg ζ = arg A + n arg z. which shows that arg A = nπ/2. On A B . and we require A B . η (a) z-plane y B (b) A x A B C ξ ζ-plane C Fig.

(7. To justify this. in the three-quarterplane and strip respectively. in both of these examples. we can use the conformal mapping to solve the Dirichlet problem for Laplace’s equation. which we derived using Fourier transforms. We can solve Laplace’s equation in the half-plane using the formula (5. namely ∇2 φ = 0 for −∞ < x < ∞ and −1 < y < 1. y(ξ. note that if φ(x. and hence ξ = −eπx/2 sin 1 πy .2 COMPLEX VARIABLE METHODS FOR SOLVING LAPLACE’S EQUATION 193 (a) (b) y D C η C' D' A' B' i x −i A B ξ ζ-plane z-plane Fig.39) and (7. . 1) = e−x 2 (7.7. 2 In the ζ-plane. η = eπx/2 cos 2 1 πy .30). We can illustrate this by solving a Dirichlet problem for Laplace’s equation in a strip. y) = Φ(x(ξ. a function that is harmonic in the (ξ.5. η)plane is harmonic in the (x. φ(x. 0) = 0 4 exp − π2 log2 (−ξ) for 0 < ξ < ∞.40) become ∂2Φ ∂2Φ + = 0 for −∞ < ξ < ∞ and 0 < η < ∞.38) so that. we can diﬀerentiate to show that dζ ∂2φ ∂2φ + 2 = ∂x2 ∂y dz 2 ∂2Φ ∂2Φ + ∂ξ 2 ∂η 2 . η)). η). We now pause to note that. −1) = 0. Mapping a strip to a half-plane. (7. y)-plane. ∂ξ 2 ∂η 2 subject to φ(x. and vice versa (see Exercise 7. 7. for −∞ < ξ < 0.11).40) We know that the transformation that maps this strip in the z = x + iy-plane to the upper-half-plane in the ζ = ξ + iη-plane is ζ = ieπz/2 . We can then easily write down the result in the z-plane using the inverse transformation. provided ζ = f (z) is conformal. by mapping to the half-plane.39) for −∞ < x < ∞. (7. subject to φ(x.

41) To show this.6(a) in the z-plane then map to the corresponding points labelled in Figure 7. so that the outside of the circle maps to the outside of the plate. so that z = ia cos θ. and has been rotated through π/2 radians. as shown in Figure 7. is Φ(ξ. This can be done using z= a2 1 i ζ+ 2 ζ .6(b). so that 2 dw dw ∼ −2i . η) = and hence φ(x.194 CLASSIFICATION AND COMPLEX VARIABLE METHODS and Φ → 0 as ξ 2 + η 2 → ∞.6. we must have dw/dζ ∼ U i/2. we write the surface of the circle as ζ = aeiθ . y) = 1 π 0 s=−∞ 1 π 0 s=−∞ 4 η exp − π2 log2 (−s) ds.6(a) in the z-plane. so that the stream at inﬁnity in the ζ-plane is half the strength of the one in the z-plane. as shown in Figure 7. . The solution. the ﬂuid is on the right hand side in the ζ-plane. (7. dz dζ Since we require that dw/dz ∼ U .30). (a) y •ia E A (b) η B' x D B • −ia C C' a θ A' E' ξ D' ζ-plane z-plane Fig. The points labelled in Figure 7.6(b). 7. and consequently is on the right hand side in the z-plane. (ξ − s)2 + η 2 eπx/2 cos eπx/2 sin 1 2 πy 1 2 πy 4 exp − π2 log2 (−s) +s 2 + eπx cos2 1 2 πy ds. As |ζ| → ∞. Mapping a ﬂat plate to a unit circle. Example 3: Flow past a ﬂat plate Consider a plate of length 2a positioned perpendicular to a uniform ﬂow with speed U . Following the sequence A B C D E . given by (5. We want to map the exterior of the plate to the exterior of a circle in the ζ-plane. z ∼ 1 iζ.

around a Joukowski proﬁle.7. and a sharp tail. the ends of the plate. 2 2 ζ 2z ζ + a2 = 0. which are not physical. and lead to ﬁnite velocities. but which still pass through the point ζ = 1 and enclose the point ζ = −1. From (7. we ﬁnd that w= 1 U 2 z 2 + a2 + z + √ a2 z 2 + a2 + z . Let’s now consider the ﬂow of a stream of strength U . Since dz/dζ = 0 at ζ = 1. In contrast.44) maps the exterior of a unit circle in the ζ-plane to the exterior of a ﬂat plate in the z-plane. ζ2 − and hence ζ = −i choosing this root so that |ζ| → ∞ as |z| → ∞. unless we can make the ﬂow in the transform plane have a stagnation point there. to give w(ζ) = a2 1 1 iπ/2 Ue ζ + U e−iπ/2 . viscosity will become important close to these points. This can be achieved by including an appropriate circulation around the aerofoil. and the velocity will be inﬁnite there. We must choose the strength of the circulation to make the velocity ﬁnite at the trailing edge. are mapped to shapes. i z 2 + a2 + z . Example 4: Flow past an aerofoil As we have seen. w(ζ) = V e−iβ ζ + V eiβ ¯ + ¯ ¯ 2π ζ . at z = ζ = 1. so that we can use the circle ¯ theorem.42) Next.2 COMPLEX VARIABLE METHODS FOR SOLVING LAPLACE’S EQUATION 195 We can now solve the ﬂow problem in the ζ-plane using the circle theorem. the transformation z= 1 2 ζ+ 1 ζ (7. we need to write this potential in terms of z. we must begin by making another conformal transformation. These shapes have two distinguishing characteristics: a blunt nose. If the centre of the circle in the ζ-plane is at ζ = ζc . known as Joukowski proﬁles. (7. the transformation is not conformal at the trailing edge. as shown in Figure 7. it does give rise to inﬁnite velocities at z = ±ia. (7. the exterior of circles of radius a > 1 whose centres are not at the origin. or trailing edge.41). ζ = (ζ − ζc ) /a. at an angle α to the horizontal. This is called the Kutta condition on the ﬂow.7. so that 1 iκ ¯ ¯ log ζ. that look rather like aerofoils. and arise because of the sharpness of the boundary at these points. The ﬂow in the ζ-plane is a uniform stream and a point vortex. From the potential (7.42). In reality. or leading edge.43) Although this result is correct.

5 η 0 −0. where ζ = ξ + iη and z = x + iy.46) 1 aU 2 1 ¯ e−iα ζ + eiα ¯ ζ + iκ ¯ log ζ.45) ¯ ¯ We can therefore make the trailing edge.5 −2 −1 0 1 η y 0. a stagnation point .7. we note that z= 1 2 1 ¯ aζ + ζc + ¯ aζ + ζc . 2π ζ (7. 7.5 1 ξ c ζ = 0. and hence V = aU/2 and β = α.5 −2 −1 0 1 y 0.5 −1 −1.2 0 − 0. so that we need U e−iα = 2V e−iβ /a. and the corresponding circles in the ζ-plane. The complex potential is therefore w(ζ) = ¯ ¯ and hence dw ¯ 1 ¯ = 2 aU dζ 1 e−iα − eiα ¯2 ζ + iκ ¯.5 −1 −1. ¯ ¯ and hence that z ∼ aζ/2 as |ζ| → ∞. ζ = (1 − ζc ) /a ≡ ζs .2 0 −1 −0.5 0 x 0.4 0.5 1 0.2 −1 −0. 2π (7.25( −1+i) 1.25 1. To determine V .196 CLASSIFICATION AND COMPLEX VARIABLE METHODS ζc = 0.5 1 0. Two examples of Joukowski proﬁles in the z-plane.5 ξ Fig.5 0 x 0.5 0 −0.

of the one-dimensional wave equation.47) ¯ ¯ ¯ This expression is real. Transform the equation to canonical form in (a) the hyperbolic region and (b) the elliptic region. after some algebra. where y1 . φ(x. we need to evaluate Fx − iFy = 1 ρi 2 dw dz 2 dz = body 1 2U 1 ρi 2 ¯ |ζ|=1 dw ¯ ¯ dζ 2 2 ¯ dζ ¯ dζ dz = 1 ρia 2 ¯ |ζ|=1 ¯ ¯ e−iα − eiα /ζ 2 + iκ/2πaζ 1 ¯ + ζc )2 2 1 − 1/(aζ ¯ dζ.1 Find and sketch the regions in the (x. in the hyperbolic region. We ﬁnd. (b) φxx − 2φxy + 3φyy + 24φy + 5φ = 0. t). y3 and y4 are the values of the solution at successive corners of any quadrilateral whose edges are characteristics. y2 . Exercises 7. From this. we can see that the lift force on the aerofoil is directed perpendicular to the incoming stream. The streamlines for the ﬂow around the aerofoils illustrated in Figure 7. x > 0.4 .2 is elliptic. that Fx − iFy = −iρκU e−iα .EXERCISES 197 by choosing κ = iπaU 1 ¯ e−iα ζs − eiα ¯ ζs . (c) φxx + 6yφxy + 9y 2 φyy + 4φ = 0. 7. φxx = xφyy . 7. Determine the characteristics of Tricomi’s equation.39). Determine the type of each of the equations (a) φxx − 5φxy + 5φyy = 0.8 with α = π/4. or equivalently. Since we know that the ¯ integrand is analytic for |ζ| > 1. In order to calculate the force on the Joukowski proﬁle. parabolic and hyperbolic. Show that any solution. (3. satisﬁes the diﬀerence equation y1 −y2 +y3 −y4 = 0.3 and reduce them to canonical form. This integral can be evaluated using residue calculus. because |ζs | = 1 and hence the quantities e−iα ζs and eiα /ζs are complex conjugate. we can evaluate the residue at the point at inﬁnity ¯ by making the transformation ζ = 1/s and evaluating the residue at s = 0. y)-plane where the equation (1 + x)φxx + 2xyφxy + y 2 φyy = 0 7.7 are shown in Figure 7. Fx + iFy = ρκU ei(π/2+α) . (7.

7. The streamlines for ﬂow past the two examples of Joukowski proﬁles shown in Figure 7.5 −1 −2 −1 0 x 1 2 Fig. . with α = π/4. 0) = φt (x. 7. 0) = 0 for x 0. subject to φ(x.8. x > 0.198 CLASSIFICATION AND COMPLEX VARIABLE METHODS 1 0.5 0 −0.5 Consider the initial–boundary value problem c−2 φtt + 2kφt − φxx = 0 for t > 0.5 0 −0.7.5 −1 −2 −1 0 x 1 2 y y 1 0.

7. Show that the image of a circle of radius r in the z-plane under the transformation ζ = (z + 1/z) /2 is an ellipse. ∂n 7. Consider the initial–boundary value problem φt = K∇2 φ − φ3 for x ∈ D. and ∂φ = 0 for x ∈ ∂D. y(ξ.30) to (7.13 Show that the mapping ζ = sin (πz/2k) maps a semi-inﬁnite strip in the z-plane to the upper half ζ-plane. Show that.32) is unique.25) and (7. 2(a2 + b2 ) 7. obtain an equation for the jump in a second derivative of φ across the characteristic curve x = ct.12 7. t) = 1 2 t for t 2 0. Consider the boundary value problem given by (7. η)). Show that the angle between two line segments that intersect at z = z0 in the complex z-plane is unchanged by a conformal mapping ζ = f (z).6 By diﬀerentiating the canonical form of the equation. where r and θ are polar coordinates and a and b are positive constants.11 Use Theorem 7. with a < R < b.8 7. 7. provided that f (z0 ) is bounded and nonzero. (a) Prove Theorem 7. then dζ ∂2φ ∂2φ + 2 = ∂x2 ∂y dz 2 ∂2Φ ∂2Φ + ∂ξ 2 ∂η 2 .9 where m(R) is the maximum value of φ on the circle r = R.EXERCISES 199 φ(0. By solving this equation. show that a2 b2 2(a + b)2 φ(0. uniformly in D. η). subject to φ(x. (b) Hence show that the solution of the initial– boundary value problem given by (7. . y) = Φ(x(ξ. t > 0.7 The function φ(r.5 to show that φ → 0 as t → ∞. 0) a2 b2 . with k > 0.4. θ) satisﬁes Laplace’s equation in region a < r < b. show that the jump in φtt across x = ct is e−kcx . 0) = φ0 (x) for x ∈ D. Using the maximum principle and the fact that ∇2 (log r) = 0. if φ(x. By using suitable functions ψ = Ax2 + By 2 that satisfy ∇2 ψ = 1 with A > 0 and B > 0.26). show that m(R) log b a m(b) log R a − m(a) log R b . ζ = ξ + iη and z = x + iy. 7.10 7.

Hint: Consider the eﬀect of the transformation ζ = z/ (z − 1 − i) on this region. Hint: Firstly apply a Joukowski transformation to map the section to a straight line.15 Find the function T (x. and then use an inverse Joukowski transformation to map this straight line to a circle. . joined at z = 1 to a horizontal ﬂat plate of unit length. Find the lift on an aerofoil section that consists of a unit circle centred on the origin. when a unit stream is incident at an angle of 45o .14 7.200 CLASSIFICATION AND COMPLEX VARIABLE METHODS 7. y) that is harmonic in the lens-shaped region deﬁned by the intersection of the circles |z − i| = 1 and |z − 1| = 1 and takes the value 0 on the upper circular arc and 1 on the lower circular arc.

Part Two Nonlinear Equations and Advanced Techniques 201 .

.

The initial condition gives c = 0 and the solution y = t3 . how sensitive are the solutions of the diﬀerential equation to changes in the initial conditions? If we performed two experiments with nearly identical initial conditions (exact repetition of an experiment being impossible in practice). it is fairly obvious that y = 0 is another solution that satisﬁes the initial condition. However. Uniqueness. Another question that we should consider arises from the process of mathematical modelling. If a diﬀerential equation provides a model of a real physical system. namely 0 (t − c)3 for 0 t c. does the prescription of initial conditions specify a unique solution? Not necessarily. If we have a diﬀerential equation for which solutions exist. dt If we integrate this separable equation we obtain y = (t − c)3 . y= for any c 0. A related issue is that of uniqueness of solutions. for t > c. . we are considering the question of existence of solutions. For example.CHAPTER EIGHT Existence. this equation has an inﬁnite number of solutions that satisfy the initial condition. When we ask whether an equation actually has a solution. This is not always the case. should we expect nearly identical outcomes? In this chapter we will prove some rigorous results related to the issues we have discussed above. Consider the equation dy = 3y 2/3 subject to y(0) = 0 for y > 0. the equation dy = dt y2 − 1 has no real solution for |y| < 1. In fact. Continuity and Comparison of Solutions of Ordinary Diﬀerential Equations Up to this point in the book we have implicitly assumed that each of the diﬀerential equations that we have been studying has a solution of the appropriate form.

CONTINUITY AND COMPARISON 8.2) This is an integral equation for the unknown solutions. we obtain t y(t) = y0 + t0 f (y(s). In general. For continuous f (y. t) | |y − y0 | b.3) t yk+1 (t) = y0 + t0 f (yk (s). s) ds.1) dt Here. 1 n t . f (y1 (s). As an example of how this works. t) subject to y(t0 ) = y0 in the domain |t − t0 | < α. α > 0 deﬁnes the size of the region where we will be able to show that a solution exists. . within which we will make certain assumptions about the behaviour of f . (8. UNIQUENESS. s) ds. (y0 . . (8. t). n! n=0 k+1 yk+1 (t) = 2 As k → ∞. {yk (t)} → y∞ (t). t0 y2 (t) = y0 + . y2 (t) = 2 + 0 1 2(1 + s) ds = 2 1 + t + t2 . R = {(y. We begin by deﬁning a closed rectangle. .1) with respect to t and apply the initial condition. if the sequence of functions (8. t0 ). Our strategy now is to use this to produce a set of successive approximations to the solution that we seek. s) ds. {yk (t)} → 2et .204 EXISTENCE. 2 . then y∞ (t) is a continuous solution of the integral equation.1 Local Existence of Solutions We will start with the simplest type of scalar diﬀerential equation by proving a local existence theorem for dy = f (y. t y1 (t) = y0 + t0 t f (y0 . consider the simple diﬀerential equation y = y subject to y(0) = 2. . the correct solution. centred upon the initial point. which is equivalent to the original diﬀerential equation. y = y(t). s) ds. We deﬁne y0 (t) = y0 . (8. In this case. We will do this using the initial condition as the starting point. t y1 (t) = 2 + t 0 2 ds = 2(1 + t). y0 (t) = 2.3) converges uniformly to a limit. we can then . |t − t0 | a} . If we integrate (8.

yk+1 (t) = y0 + t0 f (yk (s). Theorem A2. It is possible for a function to satisfy a Lipschitz condition in a region R without having a continuous partial derivative everywhere in R. We will shortly prove that. Proof We proceed by induction. t) = Since (c. which says that a function continuous on a bounded region is itself bounded. Our assumption about the continuity of the ﬁrst partial derivative automatically leads to functions that satisfy a Lipschitz condition. f (y.1. we must show that the elements of this sequence are well-deﬁned. we are going to use the successive approximations (8. t)(y2 − y1 ) for some c with y1 < c < y2 . it is suﬃcient that f (y. so that there exist strictly positive. Prior to using this. t) | |y − y0 | b. we must establish that the point (yk (s). t y0 (t) = y0 . the mean value theorem. we have ∂f (c. If (y1 . in order to ensure that the sequence (8.1 If α = min (a. ∂f /∂y is not continuous on the line y = 0. t) are two points in R.† As a preliminary to the main proof. b/M ). (y1 . We assume that t yn (t) = y0 + t0 f (yn−1 (s). s) remains in the rectangle R = {(y. D. t) ∈ R. so that it is Lipschitz with K = 1. s) ds are well-deﬁned in the interval I = {t | |t − t0 | < α}.3) converges to a continuous function. recall Theorem A2. t) and ∂f 0 ∂y (y. real constants M and K such that |f (y. as it is constant. if yk+1 (t) is to be deﬁned on some interval I. t)| < M and ∂f (y. to show that y is a solution of the original diﬀerential equation. t). ∂f /∂y ∈ C (R). |t − t0 | a} for all s ∈ I. t) ∈ R. and on this interval |yk (t) − y0 | < M |t − t0 | b. we need f . t)| < |y2 − y1 | in the unit square centred on the origin. t) |f (y2 . . t) = t|y| satisﬁes |f (y2 . Lemma 8. Speciﬁcally.8. t) − f (y1 .3) to establish the existence of a solution. As we stated earlier.1). However. s) ds † If these conditions hold in a larger domain.4) ∂f ∂y (c. (8. Functions that satisfy the inequality (8. and hence (8. states that f (y2 .4) are said to satisfy a Lipschitz condition in R. t)| < K |y2 − y1 | ∀(y2 . t) − f (y1 . then the successive approximations. In other words. where |f | < M . which is the key to proving the main result of this section.1 LOCAL EXISTENCE OF SOLUTIONS 205 diﬀerentiate under the integral sign. It is clear that y0 (t) is deﬁned on I.2. t) < K ∂y for all (y. t) ∈ R. For example. t) and (y2 . t) are continuous in R. we can use the local result repeatedly until the solution moves out of D. t) − f (y1 . ∂y < K.

means that the successive approximations cannot cross these lines either. deﬁned by (8. We can now proceed to establish the main result of this section. s)| ds t0 K t0 |yn (s) − yn−1 (s)| ds. t) remains in R for t ∈ I. s) ds t0 |f (yn (s). (8.1. as shown in Figure 8. .206 EXISTENCE. t |yn+1 (t) − yn (t)| = t0 t {f (yn (s).3). t)| < M implies that the solution of the diﬀerential equation.1. To see rather less formally why we chose α = min(a. For n = 0 we have t |y1 (t) − y0 (t)| = t0 f (y0 (s). y = y(t). note that the condition |f (y. CONTINUITY AND COMPARISON is well-deﬁned on I. |t − t0 | α. s) ds M |t − t0 |. s) − f (yn−1 (s). By deﬁnition. and the length of the interval I depends upon whether they meet the horizontal sides (α = b/M ) or the vertical sides (α = a). it is straightforward to show that |yn+1 (t) − yn (t)| M K n |t − t0 |n+1 for n (n + 1)! 1. b/M ). so that the point (yn (t). Using the deﬁnition (8. s) − f (yn−1 (s). then the successive approximations yk (t). so we have yn+1 (t) deﬁned in I. t) that satisﬁes the initial condition y(t0 ) = y0 .1.5) In order to use this. s)| ds M |t − t0 | Mα b.3). These lines intersect the boundary of the rectangle R. converge on I to a solution of the diﬀerential equation y = f (y.1 (Local existence) If f and ∂f /∂y are in C 0 (R). s) ds. t yn+1 (t) = y0 + t0 f (yn (s). which we established in Lemma 8. Now t t |yn+1 (t) − y0 | = t0 f (yn (s). s)} ds t |f (yn (s). cannot cross lines of slope M or −M through the initial point (y0 . we need to estimate the value of yn+1 (t) − yn (t). The relationship |yk (t) − y0 | < M |t − t0 |. Theorem 8. we have. which is what we claimed. UNIQUENESS. t0 ). as shown in Figure 8. Proof We begin with the identity yj (t) = y0 (t) + {y1 (t) − y0 (t)} + {y2 (t) − y1 (t)} + · · · + {yj (t) − yj−1 (t)} j−1 = y0 (t) + n=0 {yn+1 (t) − yn (t)} . by using the continuity bound on f . for n 1 and |t − t0 | < α. If we now repeatedly use these inequalities.

6) If we denote by y(t) the limit of the sequence {yk (t)}.7) Each term in this inﬁnite series is dominated by M (Kα)n+1 /K(n + 1)!.1 LOCAL EXISTENCE OF SOLUTIONS 207 Fig. and hence that |yn+1 (t) − yn (t)| M (Kα)n+1 . (8. 8. K(n + 1)! (8. from (8.8. ∞ y(t) = y0 (t) + n=0 {yn+1 (t) − yn (t)} . a positive . α. The two cases that determine the length of the interval.5). we now have.1.

M and K. |y(t + h) − y(t)| |y(t + h) − yk (t + h)| + |yk (t + h) − yk (t)| + |yk (t) − y(t)|. y(t). (k + 1)! lim f (yk (s). and we have now proved that the sequence of successive approximations. and M K (Kα)n+1 M Kα = e −1 . α so that t k→∞ t M (Kα)k+1 Kα e → 0 as k → ∞. s) ds. We now show that the limiting function. since the right hand side of (8.8) is independent of k.2).1. using the estimate (8. (8.9). and consequently y(t) satisﬁes (8. converges. (8.1 with the knowledge that y(t) = limk→∞ yk (t) is welldeﬁned. immediately claim that (y(s).3). UNIQUENESS. |yk (t) − y(t)| so that |y(t + h) − y(t)| 2 k By choosing k suﬃciently large. We can now return to Lemma 8. s) ∈ R. (k + 1)! For ﬁxed α. we can use some of the auxiliary results that we derived above.2). we have k. The comparison theorem applied to the series in (8.7) shows that it converges absolutely on the interval |t − t0 | < α. (n + 1)! K n=0 ∞ (8. t t {f (y(s).208 EXISTENCE. and. so that. actually satisﬁes the integral equation. (8. . To prove that y(t) is continuous on |t − t0 | < α. Now ∞ |y(t) − yk (t)| = n=k {yn+1 (t) − yn (t)} M K ∞ m=k (Kα)m+1 (m + 1)! M (Kα)k+1 K (k + 1)! (Kα)m m! m=0 ∞ M (Kα)k+1 Kα e = K (k + 1)! k for |t − t0 | < α. CONTINUITY AND COMPARISON constant. s)} ds t0 K t0 |y(s) − yk (s)| ds α M (Kα)k+1 Kα e .8) which is a ﬁnite constant.9) Using the Lipschitz condition in the interval |t − t0 | < α and Lemma 8. + |yk (t + h) − yk (t)| . using the triangle inequality. Consider y(t + h) − y(t) = y(t + h) − yk (t + h) + yk (t + h) − yk (t) + yk (t) − y(t). s) ds = t0 t0 f (y(s). s) − f (yk (s).

the successive approximations are not guaranteed to converge. For example. y2 . exists for −∞ < t < 1. α = min(a. . . than it is to establish local existence. if y = y + t. and the equivalence of the integral and diﬀerential equations can be established by diﬀerentiation under the integral sign. the successive approximations will converge. . In fact. .10) y1 y2 .1 LOCAL EXISTENCE OF SOLUTIONS 209 for suﬃciently small h. ∂f /∂y = 2y −1/3 . Remarks (i) For a particular diﬀerential equation.1 1 . the point of Theorem 8. Theorem 8. −∞ < t < ∞. f =    f1 (y1 . and therefore a solution exists in some rectangle containing the initial point. t) | |t| a. t) are vectors with n components and t is a real scalar. Since M = maxR y 2 = (1 + b)2 . However. t) f2 (y1 . yn . y2 . yn . b/(1 + b)2 ). independent of the particular choice of a. . is discontinuous at y = 0. . −∞ < y < ∞. From simple calculus. its partial derivative. where    y=         (8. If y = 3y 2/3 .8. f (y.1. since. . so the successive approximations are guaranteed to converge in this domain. t) = y + t and ∂f /∂y = 1 are continuous for −∞ < t < ∞. It is rather more diﬃcult to determine regions of global existence like this. This should be no surprise. 2 (1 + b) 4 1 so that. . . Theorem 8. both f (y. yn . t) = 3y 2/3 . . (iii) We can also consider the local existence of solutions of systems of ﬁrst order equations in the form y = f (y. . yn     . y2 . y(t) is continuous. the solution. as we noted in the introduction to this chapter. α 4 . The vector form of the Lipschitz condition can be established by requiring f and ∂f /∂yi to be . Let’s now try to determine the values of the constants a and b. . . . max b>0 b 1 = . therefore shows that a solution exists for |t| 4 y = 1/(1 − t). which is continuous in −∞ < y < ∞. |y − 1| b} . Hence.1. fn (y1 . . (ii) Consider the equation y = y 2 subject to y(0) = 1.1 is easy to use. as we claimed at the beginning of this section. t) . Consequently. the solution of this equation is not unique for some initial data. namely R = {(y. and we will not attempt it here. t). . using the continuity of yk (t). Since this equation satisﬁes the conditions of Theorem 8.

2 (Gronwall’s inequality) If f (t) and g(t) are non-negative functions on the interval α t β. see Coddington and Levinson (1955). β]. It then follows that for any points (y. 8. we have h (t) = f (t)g(t) Using the idea of an integrating factor. t) and (z. h(t) exp − α g(s) ds − L 0. . t so that h(α) = L. β].1. The proof then follows that of Theorem 8. t f (t) Proof Deﬁne L exp α g(s) ds for t ∈ [α. (iv) For a selection of results similar to. t) in D. where ||.210 EXISTENCE.|| is a suitable vector norm. UNIQUENESS. and by the fundamental theorem of h(t)g(t) for t ∈ [α. t h(t) = L + α f (s)g(s) ds. the solution of the initial value problem (8. with modulus signs replaced by vector norms. so that ∂f ∂yi K for all (y. In order to prove this. but in some ways better than. ||f (y. By hypothesis. L is a non-negative constant and t f (t) then L+ α f (s)g(s) ds for t ∈ [α. t f (t) L exp α g(s) ds . t)|| K||y − z||. since g(t) 0. under the same conditions as Theorem 8. t) ∈ D ⊂ Rn+1 . CONTINUITY AND COMPARISON continuous with respect to all of their components and t. we need a result called Gronwall’s inequality. β]. since f (t) h(t). t) − f (z.1. Theorem 8.1) is unique.2 Uniqueness of Solutions We will now show that. Lemma 8. this can be rewritten as d h(t) exp − dt Integrating from α to t gives t g(s) ds α 0. and.1 step by step and line by line. f (t) calculus. h(t).

3 Dependence of the Solution on the Initial Conditions A solution of y = f (y. t0 ) depends continuously on the three variables t0 . . β). t1 ∈ (α. We can now apply Gronwall’s inequality to the non-negative function |y1 (t) − y2 (t)| with L = 0 and g(s) = K > 0 to conclude that |y1 (t) − y2 (t)| 0. since the modulus of a function cannot be negative. If f . t) that passes through the initial point (y0 . s) ds. and hence y1 (t) ≡ y2 (t). then ∀ > 0 ∃δ > 0 such that if |t0 − t1 | < δ and |y0 − y1 | < δ then |y0 (t) − y1 (t)| < ∀t ∈ (α. s) ds.10) exists and is unique.3 DEPENDENCE OF THE SOLUTION ON THE INITIAL CONDITIONS 211 Theorem 8. s)} ds t t |{f (y1 (s). y0 and t. Just as we saw earlier when discussing the existence of solutions. this uniqueness result can be extended to systems of ordinary diﬀerential equations. s) − f (y2 (s). Earlier. at least for values of t close to the initial point. We can now prove a theorem about this. ∂f ∈ C 0 (R). y1 ). y0 and t. then the solution of the initial ∂y value problem y = f (y. y0 ) and y = y1 (t) the solution that passes through (t1 . s) lie in the region R.3 Let y = y0 (t) be the solution of y = f (y. |y1 (t) − y2 (t)| = 0. s) − f (y2 (s). Proof We proceed by contradiction. β). where K > 0 is the Lipschitz constant for the region R. y = 3y 2/3 subject 1/3 to y(t0 ) = y0 has solution y = (t − t0 + y0 )3 . which depends continuously on t0 . with y1 (t0 ) = y2 (t0 ) = y0 . ∂f /∂y ∈ C 0 (R) and both y0 (t) and y1 (t) exist on some interval α < t < β with t0 . the solution of (8. Under the conditions given in Remark (iii) in the previous section. t) that passes through the initial point (t0 . we hypothesized that solutions of identical diﬀerential equations with initial conditions that are close to each other should remain close. t = t0 . y = y1 (t) and y = y2 (t).2 (Uniqueness) If f . s)}| ds t0 t0 K|y1 (s) − y2 (s)| ds for |t − t0 | α. t) subject to y(t0 ) = y0 is unique on |t − t0 | < α.11) and hence the points (yi (s). Theorem 8. However. (8. y2 (t) = y0 + t0 f (y2 (s). Suppose that there exist two solutions. 8. These functions satisfy the integral equations t t y1 (t) = y0 + t0 f (y1 (s). Taking the modulus of the diﬀerence of these leads to t |y1 (t) − y2 (t)| = t0 {f (y1 (s).8. For example.

8. (y1 . it would seem plausible that 0 y of this traps the solution of the initial value problem in the range 1 y 1 + t. For example. we will prove all of the results in this section for one side of a point only. since 0 e−ty 1 1. s)} ds. this reduces to t |y0 (t) − y1 (t)| (M + 1)δ + t1 K|y0 (s) − y1 (s)| ds. and the properties of its solution are not immediately obvious. However. s) − f (y1 (s). we need some deﬁnitions. y = e−ty subject to y(0) = 1 is a rather nasty nonlinear diﬀerential equation.4 Comparison Theorems Since a large class of diﬀerential equations cannot be solved explicitly. Note that. t) satisﬁes a one-sided Lipschitz condition in a domain D if. we can take their diﬀerence and.11). In order to prove this result rigorously. it is useful to have a method of placing bounds on the solution of a given equation by comparing it with solutions of a related equation that is simpler to solve. by using the upper bound on f and the Lipschitz condition on R. † We have assumed that t1 > t0 . t) − F (y1 . we have |y0 (t) − y1 (t)| < . by splitting up the range of integration. CONTINUITY AND COMPARISON Proof Since y1 (t) and y2 (t) satisfy the integral equations (8. Firstly. s) ds + t0 t t1 {f (y0 (s). UNIQUENESS. and. and the proof is complete. t) ∈ D. . A function F (y. s) ds + t0 t1 {f (y0 (s). but the argument needs only a slight modiﬁcation if this is not the case. If |t1 − t0 | < δ and |y1 − y0 | < δ. we have |y0 (t) − y1 (t)| (M + 1)δ exp {K(β − α)} . t). for some constant K. we need some preliminary results and ideas. s) − f (y1 (s). write† y0 (t) − y1 (t) = y0 − y1 + Taking the modulus gives |y0 (t) − y1 (t)| |y0 − y1 | + t1 t t1 t f (y0 (s). to simplify things. t) K(y2 − y1 ) for (y2 . y2 > y1 ⇒ F (y2 . If we now choose δ < exp {−K(β − α)} /(M + 1). f (y0 (s). s)} ds |y0 − y1 | + M (t1 − t0 ) + t1 K|y0 (s) − y1 (s)| ds. We can now apply Gronwall’s inequality to obtain t |y0 (t) − y1 (t)| (M + 1)δ exp t1 K ds = (M + 1)δ exp {K(t − t1 )} .212 EXISTENCE. Integration for 0 t < ∞ and 0 y < ∞. since |t − t1 | < β − α.

and hence the result. dt 0 and multiply through by This leads to e−Kt σ(t) e−Ka σ(a).8. then σ(t) σ(a).12) It is known. so that σ(t) 0 for t0 t t1 and σ(t0 ) = 0. t) and F or G is one-sided Lipschitz on this strip.4 COMPARISON THEOREMS 213 If σ(t) is diﬀerentiable for t > a.4 If g(t) is a solution of y = F (y.4 Let f (t) and g(t) be solutions of the diﬀerential equations y = F (y. Note that if σ (t) 0 for t a.4. Deﬁne σ(t) = f (t) − g(t). σ(t) ≡ 0. with f (a) = g(a). then σ(t) σ(a)eK(t−a) for t ∈ [a. Deﬁne t0 to be the largest t in the interval a t t1 such that f (t) g(t). t) G(y. This contradicts the hypothesis that σ(t1 ) = f (t1 ) − g(t1 ) > 0. provided that F satisﬁes a one-sided Lipschitz condition for t a. If F (y.4. t) K {f (t) − g(t)} = Kσ(t). Comparison theorems.3 If σ(t) is a diﬀerentiable function that satisﬁes the diﬀerential inequality σ (t) Kσ(t) for t ∈ [a. then σ (t) K is called a diﬀerential inequality. b] and some constant K.4 and the preceding two lemmas can be made two-sided in the neighbourhood of the initial point with minor modiﬁcations. b]. t) G(y. t) respectively. t) for t a and f (t) is a function that satisﬁes the diﬀerential inequality f (t) F (f (t). t). and hence f (t0 ) = g(t0 ). (8. and hence the result is proved. f (t) g(t) for all t a.3. Kσ(t) for t > t0 . Proof We will proceed by contradiction. Theorem 8. on the strip a t b. then. Proof We write the inequality as σ (t) − Kσ(t) e−Kt > 0 to obtain d −Kt e σ(t) 0. t) for t a with f (a) = g(a). then f (t) g(t) for a t b Proof Since y = F (y. By Lemma 8. such as Theorem 8. t) − F (g(t). Lemma 8. and f satisﬁes the diﬀerential equation. Note that this is really just a special case of Gronwall’s inequality. since σ(t) is non-negative for t t0 . Consider as an example the initial value problem y = t2 + y 2 subject to y(0) = 1. that . Lemma 8. Note that Theorem 8. t) and z = G(z. Suppose that f (t1 ) > g(t1 ) for some t1 > a. either from analysis of the solution or by numerical integration. The one-sided Lipschitz condition shows that σ (t) = f (t) − g (t) F (f (t). and hence σ (t) However. are of considerable practical use. σ(t) σ(t0 )eK(t−t0 ) 0. g satisﬁes a diﬀerential inequality of the form described in Lemma 8. This gives us the result immediately.

214

EXISTENCE, UNIQUENESS, CONTINUITY AND COMPARISON

the solution of this equation ‘blows up’ to inﬁnity at some ﬁnite value of t = t∞ . We can estimate the position of this blowup point by comparing the solution of (8.12) with the solutions of similar, but simpler, diﬀerential equations. Since t2 + y 2 y 2 for 0 t < ∞, the solution of (8.12) is bounded below by the solution of y = y 2 subject to y(0) = 1, which is y = 1/(1−t). Since this blows up when t = 1, we must have t∞ 1. Also, since t2 + y 2 1 + y 2 for 0 t 1, an upper bound is provided by the solution of y = 1 + y 2 subject to y(0) = 1, namely y = tan t + π . This 4 blows up when t = π/4. By sandwiching the solution of (8.12) between two simpler solutions, we are able to conclude that the blowup time satisﬁes π t∞ 1. 4 Comparison theorems are also available for second order diﬀerential equations. These take a particularly simple form for the equation y (x) + g(x)y(x) = 0, (8.13)

from which the ﬁrst derivative of y is absent. In fact, we can transform any linear, second order, ordinary diﬀerential equation into this form, as we shall see in Section 12.2.7. We will compare the solution of (8.13) with that of z (x) + h(x)z(x) = 0, and prove the following theorem. Theorem 8.5 If g(x) < h(x) for x x0 , y(x) is the solution of (8.13) with y(x0 ) = y0 = 0, y (x0 ) = y1 , these conditions being such that y(x) > 0 for x0 x x1 , and z(x) is the solution of (8.14) with z(x0 ) = y0 and z (x0 ) = y1 , then y(x) > z(x) for x0 < x x1 , provided that z(x) > 0 on this interval. Proof From (8.13) and (8.14), y z − yz = (h − g) yz. Integrating this equation from x0 to x < x1 , we obtain
x

(8.14)

y (x)z(x) − y(x)z (x) =
x0

(h(t) − g(t)) y(t)z(t) dt.

By hypothesis, the right hand side of this is positive. Also, by direct diﬀerentiation, d dx y(x) z(x) = y (x)z(x) − y(x)z (x) > 0, z 2 (x)

so that y/z is an increasing function of x. Since y(x0 )/z(x0 ) = 1, we have y(x) > z(x) for x0 < x x1 . As an example of the use of Theorem 8.5, consider Airy’s equation, y (t) − ty(t) = 0 ¨ (see Sections 3.8 and 11.2), in −1 < t < 0 with y(0) = 1 and y(0) = 0. By ˙ making the transformation t → −x, we arrive at y (x) + xy(x) = 0 in 0 < x < 1, x < x1 . with y(0) = 1 and y (0) = 0. The solution of this is positive for 0 If we consider z (x) + z(x) = 0 with z(0) = 1 and z (0) = 0, then clearly z =

EXERCISES

215

cos x, which is positive for 0 x < π/2, and hence for 0 x 1. Since the equations and boundary conditions that govern y(x) and z(x) satisfy the conditions of Theorem 8.5, we conclude that y(x) > cos x for 0 < x < 1. In fact, we can see from the exact solution, y(x) = Bi (0)Ai(−x) − Ai (0)Bi(−x) , Bi (0)Ai(0) − Ai (0)Bi(0)

shown in Figure 8.2, that x1 ≈ 1.986.
1 y cos x

0.5

0

−0.5

0

0.2

0.4

0.6

0.8

1 x

1.2

1.4

1.6

1.8

2

Fig. 8.2. The exact solution of y + xy = 0 subject to y(0) = 1 and y (0) = 0.

Exercises 8.1 Determine the integral equations equivalent to the initial value problems (a) y = t2 + y 4 subject to y(0) = 1, (b) y = y + t subject to y(0) = 0, and determine the ﬁrst two successive approximations to the solution. 8.2 Show that the functions (a) f (y, t) = te−y for |t| (b) f (y, t) = t + y for |t|
2 2
2

1, |y| < ∞, 2, |y| 3,

are Lipschitz in the regions indicated, and ﬁnd the Lipschitz constant, K, in each case.

216

EXISTENCE, UNIQUENESS, CONTINUITY AND COMPARISON

8.3 8.4

How could successive approximations to the solution of y = 3y 2/3 fail to converge to a solution? Let f (y, t) and g(y, t) be continuous and satisfy a Lipschitz condition with respect to y in a region D. Suppose |f (y, t) − g(y, t)| < in D for some > 0. If y1 (t) is a solution of y = f (y, t) and y2 (t) is a solution of y = g(y, t), such that |y2 (t0 ) − y1 (t0 )| < δ for some t0 and δ > 0, show that, for all t for which y1 (t) and y2 (t) both exist, {exp (K|t − t0 |) − 1} , K where K is the Lipschitz constant. Hint: Use the Gronwall inequality. Find upper and lower bounds to the solutions of the diﬀerential equations |y2 (t) − y1 (t)| δ exp (K|t − t0 |) + (a) y = sin(xy) subject to y(0) = 1/2 for x (b) y = y 3 − y subject to y(0) = 1/4. 0,

8.5

8.6

If σ(t) ∈ C 1 [a, a + ] and positive satisﬁes the diﬀerential inequality σ Kσ log σ, show that σ(t) σ(a)eK(t−a) for t ∈ [a, a + ].

8.7

For each ﬁxed x, let F (x, y) be a nonincreasing function of y. Show that, if f (x) and g(x) are two solutions of y = F (x, y) and b > a, then |f (b) − g(b)| |f (a) − g(a)|. Hence deduce a result concerning the uniqueness of solutions. This is known as the Peano uniqueness theorem.

CHAPTER NINE

Nonlinear Ordinary Diﬀerential Equations: Phase Plane Methods

9.1

Introduction: The Simple Pendulum

Ordinary diﬀerential equations can be used to model many diﬀerent types of physical system. We now know a lot about second order linear ordinary diﬀerential equations. For example, simple harmonic motion,

d 2θ + ω 2 θ = 0, dt2

(9.1)

describes many physical systems that oscillate with small amplitude θ. The general solution is θ = A sin ωt + B cos ωt, where A and B are constants that can be ﬁxed from the initial values of θ and dθ/dt. The solution is an oscillatory function of t. Note that we can also write this as θ = Ceiωt +De−iωt , where C and D are complex constants. In the real world, the physics of a problem is rarely as simple as this. Let’s consider the frictionless simple pendulum, shown in Figure 9.1. A mass, m, is attached to a light, rigid rod of length l, which can rotate without friction about the point O.

Fig. 9.1. A simple pendulum.

218

PHASE PLANE METHODS

Using Newton’s second law on the force perpendicular to the rod gives −mg sin θ = ml and hence d 2θ + ω 2 sin θ = 0, dt2 (9.2) d 2θ , dt2

where ω 2 = g/l and t is time. For oscillations of small amplitude, θ 1, so that sin θ ∼ θ, and we obtain simple harmonic motion, (9.1). If θ is not small, we must study the full equation of motion, (9.2), which is nonlinear. Do we expect the solutions to be qualitatively diﬀerent to those of simple harmonic motion? If we push a pendulum hard enough, we should be able to make it swing round and round its point of support, with θ increasing continuously with t (remember there is no friction), so we would hope that (9.2) has solutions of this type. In general, nonlinear ordinary diﬀerential equations cannot be solved analytically, but for equations like (9.2), where the ﬁrst derivative, dθ/dt does not appear ˙ explicitly, an analytical solution is available. Using the notation θ = dθ/dt, the ˙ as a function of θ instead of t. Note that trick is to treat θ d d 2θ = 2 dt dt dθ dt = ˙ ˙ ˙ dθ dθ dθ ˙ dθ = d = =θ dt dt dθ dθ dθ 1 ˙2 θ . 2

This allows us to write (9.2) as d dθ 1 ˙2 θ 2 = −ω 2 sin θ,

which we can integrate once to give 1 ˙2 θ = ω 2 cos θ + constant. 2 Using ω 2 = g/l, we can write this as 1 2 ˙2 ml θ − mgl cos θ = E. 2 (9.3)

This is just a statement of conservation of energy, E, with the ﬁrst term representing kinetic energy, and the second, gravitational potential energy. Systems like (9.2), which can be integrated once to determine a conserved quantity, here energy, E, are called conservative systems. Note that if we try to account for a small amount of friction at the point of suspension of the pendulum, we need to add a term proportional to dθ/dt to the left hand side of (9.2). The system is then no longer conservative, with dramatic consequences for the motion of the pendulum (see Exercise 9.6). From (9.3) we can see that dθ =± dt 2E 2g cos θ. + 2 ml l

9.1 INTRODUCTION: THE SIMPLE PENDULUM

219

We can integrate this to arrive at the implicit solution
θ

t=±
θ0

2E ml2

, cos θ

(9.4)

+

2g l

where θ0 is the angle of the pendulum when t = 0. Note that the two constants of integration are E and θ0 , the initial energy and angle of the pendulum. Equation (9.4) is a simple representation of the solution, which, if necessary, we can write in terms of Jacobian elliptic functions (see Section 9.4), so now everything is clear . . . , except of course that it isn’t! Presumably equation (9.4) gives oscillatory solutions for small initial angles and kinetic energies, and solutions with θ increasing with t for large enough initial energies, but this doesn’t really leap oﬀ the page at you. From (9.4) we have a quantitative expression for the solution, but we are really more interested in the qualitative nature of the solution. Let’s go back to (9.3) and write 2E 2g ˙ cos θ. + θ2 = 2 ml l ˙ Graphs of θ2 as a function of θ are shown in Figure 9.2(a) for diﬀerent values of E. — For E > mgl the curves lie completely above the θ-axis. — For −mgl < E < mgl the curves intersect the θ-axis. — For E < −mgl the curves lie completely below the θ-axis (remember, −1 cos θ 1). ˙ We can now determine θ as a function of θ by taking the square root of the curves in Figure 9.2(a) to obtain the curves in Figure 9.2(b), remembering to take both the positive and negative square root. — For E > mgl, the curves lie either fully above or fully below the θ-axis. ˙ — For −mgl < E < mgl, only ﬁnite portions of the graph of θ2 lie above the θ-axis, so the square root gives ﬁnite, closed curves. — For E < −mgl, there is no real solution. This corresponds to the fact that the pendulum always has a gravitational potential energy of at least −mgl, so we must have E −mgl. As we shall see later, the solution with E = mgl is an important one. The graph ˙ of θ2 just touches the θ-axis at θ = ±(2n − 1)π, for n = 1, 2, . . . , and taking the square root gives the curves that pass through these points shown in Figure 9.2(b). ˙ ˙ How do θ and θ vary along these solution curves as t increases? If θ is positive, θ ˙ = dθ/dt is, by deﬁnition, the rate at increases with t, and vice versa (remember, θ which θ changes with t). This allows us to add arrows to Figure 9.2(b), indicating in which direction the solution changes with time. We have now constructed our ˙ ﬁrst phase portrait for a nonlinear ordinary diﬀerential equation. The (θ, θ)plane is called the phase plane. Each of the solution curves represents a possible solution of (9.2), and is known as an integral path or trajectory. If we know the 2g 2E ˙ ˙ cos θ0 , the integral path that passes initial conditions, θ = θ0 , θ = θ0 = 2 +
ml l

220

PHASE PLANE METHODS

˙ through the point (θ0 , θ0 ) when t = 0 represents the solution. Finally, note that, since θ = π is equivalent to θ = −π, we only need to consider the phase portrait for −π θ π. Alternatively, we can cut the phase plane along the lines θ = −π and θ = π and join them up, so that the integral paths lie on the surface of a cylinder.

˙ ˙ Fig. 9.2. (a) θ2 as a function of θ for diﬀerent values of E. (b) θ as a function of θ – the phase portrait for the simple pendulum. The cross at (0, 0) indicates an equilibrium solution. The arrows indicate the path followed by the solution as t increases. Note that the phase portrait for |θ| π is the periodic extension of the phase portrait for |θ| π.

Having gone to the trouble of constructing a phase portrait for (9.2), does it tell us what the pendulum does in a form more digestible than that given by (9.4)? Let’s consider the three qualitatively diﬀerent types of integral path shown in Figure 9.2(b). ˙ (i) Equilibrium solutions The points θ = 0, θ = 0 or ±π, represent the two equilibrium solutions of (9.2). The point (0, 0) is the equilibrium with the pendulum vertically downward, (π, 0) the equilibrium with the pendulum vertically upward. Points close to (0, 0) lie on small closed trajectories close to (0, 0). This indicates that (0, 0) is a stable equilibrium point, since a small change in the state of the system away from equilibrium leads to solutions that remain close to equilibrium. If you cough on a pendulum hanging downwards, you will only excite a small oscillation. In contrast, points close to (π, 0) lie on trajectories that take the solution far away from (π, 0), and we say that this is an unstable equilibrium point. If you cough on a pendulum balanced precariously above its point of support, it will fall. Of course, in practice it is impossible to balance a pendulum in

9.1 INTRODUCTION: THE SIMPLE PENDULUM

221

this way, precisely because the equilibrium is unstable. We will reﬁne our deﬁnitions of stability and instability in Chapter 13. (ii) Periodic solutions Integral paths with −mgl < E < mgl are closed, and represent periodic solutions. They are often referred to as limit cycles or periodic orbits. The pendulum swings back and forth without reaching the upward vertical. These orbits are stable, since nearby orbits remain nearby. Note that the frequency of the oscillation depends upon its amplitude, a situation that is typical for nonlinear oscillators. For the simple pendulum, the amplitude of the motion is θmax = cos−1 (E/mgl), and (9.4) shows that the period of the motion, T , is given by
θmax

T =2
−θmax 2E ml2

dθ +
2g l

. cos θ

Small closed trajectories in the neighbourhood of the equilibrium point at (0, 0) are described by simple harmonic motion, (9.1). Note that, in contrast to the full nonlinear system, the frequency of simple harmonic motion is independent of its amplitude. The idea of linearizing a nonlinear system of ordinary diﬀerential equations close to an equilibrium point in order to determine what the phase portrait looks like there, is one that we will return to later. In terms of the phase portrait on the cylindrical surface, trajectories with E > mgl are also stable periodic solutions, looping round and round the cylinder. The pendulum has enough kinetic energy to swing round and round its point of support. (iii) Heteroclinic solutions The two integral paths that connect (−π, 0) and ˙ (π, 0) have E = mgl. The path with θ 0 has θ → ±π as t → ±∞ (we will prove this later). This solution represents a motion where the pendulum swings around towards the upward vertical, and is just caught between falling back and swinging over. This is known as a heteroclinic path, since it connects diﬀerent equilibrium points.† Heteroclinic paths are important, because they represent the boundaries between qualitatively diﬀerent types of behaviour. They are also unstable, since nearby orbits behave qualitatively diﬀerently. Here, the heteroclinic orbits separate motions where the pendulum swings back and forth from motions where it swings round and round (diﬀerent types of periodic solution). In terms of the phase portrait on a cylindrical surface, we can consider these paths to be homoclinic paths, since they connect an equilibrium point to itself.‡ We have now seen that, if we can determine the qualitative nature of the phase portrait of a second order nonlinear ordinary diﬀerential equation and sketch it, we can extract a lot of information about the qualitative behaviour of its solutions. This information provides rather more insight than the analytical solution, (9.4),
† Greek hetero = diﬀerent ‡ Greek homo = same

222

PHASE PLANE METHODS

into the behaviour of the physical system that the equation models. For nonconservative equations, we cannot integrate the equation directly to get at the equation of the integral paths, and we have to be rather more cunning in order to sketch the phase portrait. We will develop methods to tackle second order, nonconservative, ordinary diﬀerential equations in Section 9.3. Before that, we will consider the simpler case of ﬁrst order nonlinear ordinary diﬀerential equations.

9.2 First Order Autonomous Nonlinear Ordinary Diﬀerential Equations An autonomous ordinary diﬀerential equation is one in which the independent ¨ variable does not appear explicitly, for example (9.2). The equations θ+t2 θ = 0 and 2 th x = t−x are nonautonomous. Note, however, that an n order nonautonomous ˙ ordinary diﬀerential equation can always be written as a (n + 1)th order system of autonomous ordinary diﬀerential equations. For example x = t − x2 is equivalent ˙ ˙ to x = y − x2 , y = 1 with y = 0 when t = 0. ˙ In this section we focus on the qualitative behaviour of solutions of ﬁrst order, autonomous, ordinary diﬀerential equations, which can always be written as dx = x = X(x), ˙ dt (9.5)

with X(x) a given function of x. Of course, such an equation is separable, with the solution subject to x(0) = x0 given by
x

t=
x0

dx . X(x )

As we found in the previous section, solving the equation analytically is not necessarily the easiest way to determine the qualitative behaviour of the system (see Exercise 9.2).

9.2.1 The Phase Line Consider the graph of the function X(x). An example is shown in Figure 9.3(a). If X(x1 ) = 0, then x = 0, and hence x = x1 is an equilibrium solution of (9.5). ˙ For the example shown in Figure 9.3 there are three equilibrium points, at x = x1 , ˙ x2 and x3 . We can also see that x = X(x) < 0, and hence x is a decreasing function of t for x < x1 and x2 < x < x3 . Similarly, x = X(x) > 0, and hence ˙ x increases as t increases, for x1 < x < x2 and x > x3 . By analogy with the phase plane, where we constructed the phase portrait for a second order system in the previous section, we can draw a phase line for this ﬁrst order equation, as shown in Figure 9.3(b). The arrows indicate whether x increases or decreases with t. Clearly, the diﬀerent types of behaviour that are possible are rather limited by the constraint that the trajectories lie in a single dimension. Both for this example and in general, trajectories either enter an equilibrium point or head oﬀ to inﬁnity.

9.2 FIRST ORDER EQUATIONS

223

In particular, periodic solutions are not possible.† Solutions that begin close to x = x1 or x = x3 move away from the equilibrium point, so that x = x1 and x = x3 are unstable. In contrast, x = x2 is a stable equilibrium point, since solutions that start close to it approach it.
(a) X(x)

x2 x1 x3

x

(b) x •1 x2 • x3 •

Fig. 9.3. (a) An example of a graph of x = X(x). (b) The equivalent phase line. ˙

From Figure 9.3(b) we can see that x → −∞ as t → ∞ x → x2 as t → ∞ x → ∞ as t → ∞ when x0 < x1 , when x1 < x0 < x3 , when x0 > x3 .

The set D−∞ = {x | x < x1 } is called the domain of attraction or basin of attraction of minus inﬁnity. Similarly, D2 = {x | x1 < x < x3 } is the domain of attraction of x2 and D∞ = {x | x > x3 } that of plus inﬁnity. All of this information has come from a qualitative analysis of the graph of x = X(x). ˙

9.2.2 Local Analysis at an Equilibrium Point If x = x1 is an equilibrium solution of (9.5), what happens quantitatively close to ¯ x1 ? Firstly, let’s move the equilibrium point to the origin by deﬁning x = x − x1 ,
† If there is some geometrical periodicity in the problem so that, for example, by analogy with x π and x = −π is equivalent to x = π, we can construct a the simple pendulum, −π phase loop, around which the solution can orbit indeﬁnitely.

x → 0 (x → x1 ) as t → −∞. x = x1 is a nonhyperbolic equilibrium point.224 PHASE PLANE METHODS so that d¯ x ¯ = X(x1 + x). This ¯ is consistent with our qualitative analysis (note the slope of X(x) at the equilibrium points in Figure 9. a piece of quantitative information well worth knowing. Now. we will assume that the right hand sides in (9. d¯ x ≈ X (x1 )¯ for x x ¯ 1. ordinary diﬀerential equation can be written as a system of two ﬁrst order equations. y = Y (x. and the equilibrium point is therefore stable. y). x ¯ 1. for all of the systems that we study.6). In the neighbourhood of the equilibrium point. If X (x1 ) = 0. and the equilibrium point is therefore unstable. in the form x = X(x. ˙ ˙ (9. ¯ ¯ ¯ (9.3(a)). solutions that do not start at a hyperbolic equilibrium point cannot reach it in a ﬁnite time.6) 2 where X = dX/dx. in order to sort out what happens close to the equilibrium point. X(x1 ) = 0. We will consider this further in Chapter 13. ¯ 1 X(x1 + x) = X(x1 ) + xX (x1 ) + x2 X (x1 ) + · · · . .2. we now know that solutions approach stable equilibrium points exponentially fast as t → ∞. and. (9. Let’s deﬁne x = θ and y = x. we say that x = x1 is a hyperbolic equilibrium point. and this analysis determines how solutions behave in its neighbourhood. † Continuously diﬀerentiable functions are continuous and have derivatives with respect to the independent variables that are continuous. y = x = θ = −(g/l) sin θ.3 Second Order Autonomous Nonlinear Ordinary Diﬀerential Equations Any second order. and hence ˙ ˙ ¨ ¨ g x = y.2). x → 0 ¯ ¯ (x → x1 ) as t → ∞. dt This has solution x = k exp {X (x1 )t} for some constant k. If X (x1 ) > 0. y = − sin x. If X (x1 ) = 0. assuming that X (x1 ) = 0. ¯ dt Since x = x1 is an equilibrium solution. ˙ ˙ l From now on. autonomous. and we need to retain more terms in the Taylor expansion of X.7) For example. If X (x1 ) < 0. dt We can expand X(x1 + x) as a Taylor series.7) are continuously diﬀerentiable†. 9. In particular. and hence d¯ x ≈ X(x1 ) + xX (x1 ). which governs the motion of a simple pendulum. Moreover. y). consider (9. and hence that a solution exists and is unique in the sense of Theorem 8.

1 The Phase Plane In Section 9. y0 ). This will become clearer in the next section. using x = x0 + x. y0 ) = 0. to obtain ¯ ¯ ¯ X(x0 + x. and ˙ ˙ this point represents an equilibrium solution of (9.9.9) . and are therefore equilibrium points. y0 ) + · · · . y0 ) = Y (x0 . and we deduce that integral paths can meet only at equilibrium points. we saw how the solutions of the equations of motion of a simple pendulum can be represented in the phase plane. y0 + y ) = x ¯ ¯ ¯ ∂X ∂X (x0 . ¯ ∂x ∂y ∂Y ∂Y (x0 . We begin by shifting the origin to an equilibrium point at (x0 . Since the solution through such a point exists and is unique. we proceed as we did for ﬁrst order systems. y0 ) that are not ordinary have X(x0 . At an equilibrium point. x ¯ d¯ x ≈ xXx (x0 . y0 ) and Y (x0 . dy/dx. Integral paths cannot cross at ordinary points! Points (x0 . y0 ) are not both zero. ¯ y = y0 + y . dt (9. y) This slope is uniquely deﬁned at all points x = x0 and y = y0 at which X(x0 . ¯ ∂x ∂y 1 and y ¯ 1.3.3 SECOND ORDER EQUATIONS 225 9. y = y0 . ¯ ¯ dt d¯ y ≈ xYx (x0 . and only as t → ±∞. This is possibly the most useful piece of information that you need to bear in mind when sketching phase portraits. ¯ ¯ (9. and hence In the neighbourhood of the equilibrium point. y0 ) + · · · .2 Equilibrium Points In order to determine what the phase portrait looks like close to an equilibrium point.8) dt where we have used the notation Xx = ∂X/∂x.7). y0 ). y0 )u. Let’s now consider the phase plane for a general second order system. we deduce that integral paths cannot cross at ordinary points.3. The most convenient way of writing this is in matrix form. These are known as ordinary points. y0 + y ) = x Y (x0 + x. y)-plane. 9. y0 ) + y (x0 . x = y = 0 when x = x0 . y) dy = = = . plotted in the (x. ¯ Now we Taylor expand the functions X and Y . y(t)) of (9.7). dx dx/dt x ˙ X(x. The slope of an integral path is dy/dt y ˙ Y (x. as du = J(x0 . y0 ). y0 ) + y Yy (x0 . remembering that X(x0 . An integral path is a solution (x(t). or phase plane. In other words. y0 ) + y Xy (x0 . y0 ) + y (x0 . the slope. of the integral paths is not well-deﬁned. y0 ) = 0.1. y0 ) = Y (x0 .

ﬁrst order ordinary diﬀerential equations with constant coeﬃcients. ˆ with k1 and k2 constants. which suggests that we can look for solutions of the form u = u0 eλt .12) shows that x. so are the eigenvectors u1 and u2 . J(x0 . in both the transformed and untransformed coordinate systems. This is the equation of the integral paths in the transformed coordinates.11) for any constants A1 and A2 . There are now three main cases to consider. real. This type of equilibrium point is called an unstable node. u = u1 eλ1 t and u = u2 eλ2 t . y0 ) the Jacobian matrix at the equilibrium point (x0 . (i) Distinct.9) represents a pair of linear. y0 ) Yx (x0 . The equation of the integral paths. . ˆ λ /λ λ1 0 0 λ2 . λ = λ1 and λ = λ2 . and corresponding unit eigenvectors u0 = u1 and u0 = u2 . where Λ= ˙ ˙ ˆ ˆ ˆ Therefore x = λ1 x. where v = (ˆ. and (9. y0 ) . and hence ˆ y = k3 xλ2 /λ1 . ˙ v = P −1 JP v = Λv. (9. y ) and P = (u1 .9) gives λu0 = Ju0 . distinct and nonzero. so that ˆ x = k1 eλ1 t . (9.11) suggests that the form of the solution is simpler if we make a linear transformation so that u1 and u2 lie along the coordinate axes. These provide us with two possible solutions of (9. The local phase portrait is sketched in Figure 9.10) This is an eigenvalue problem. Such a transformation is given by u = P v.226 PHASE PLANE METHODS where u= x ¯ y ¯ . so the equilibrium point ˆ ˆ is unstable. We call J(x0 . Substituting this into (9. y0 ) Xy (x0 . y → 0 as t → −∞.9) gives ˙ P v = JP v. y = k2 eλ2 t . Substituting this form of solution into (9. possibly equal.13). exponentially fast. y0 ) = Xx (x0 . Note that the transformed coordinate axes are integral paths. λ2 > 0) The solution (9. (9. There are two eigenvalues. possibly complex. y0 ) Yy (x0 . y = λ2 y .4. then shows that they all meet at the equilibrium point at the origin.12) (9. When the eigenvalues λ1 and λ2 are real. u2 ) is a x ˆ matrix with columns given by u1 and u2 . y0 ). positive eigenvalues (λ1 . Equation (9.13) where k3 = k1 2 1 /k2 . so that the general solution is the linear combination u = A1 u1 eλ1 t + A2 u2 eλ2 t . where u0 and λ are constants to be determined. ˆ (9.9).

and ˆ ˆ vice versa. (ii) Distinct. This type of equilibrium point is called a saddle point. An unstable node with λ2 > λ1 sketched in (a) the transformed and (b) the untransformed coordinate systems. the linearization that we have performed in order to obtain (9.9. but with the arrows reversed. this is the phase portrait close to the equilibrium point. Figure 9.5(a). This type of equilibrium point is called a stable node. λ2 < 0) In this case. and thereby causes the integral ˆ paths to bend as they do.13).9) becomes inappropriate and we must consider how this local phase portrait ﬁts into the full.4. Figure 9. On the other integral paths. global picture. negative eigenvalues (λ1 . real. (iii) Real eigenvalues of opposite sign (λ1 λ2 < 0) In this case. The situation is as shown in Figure 9.5(b). (a) y (b) − y u2 u1 − x x Fig. separatrix. and we recover the previous case.3 SECOND ORDER EQUATIONS 227 Remember. the solution grows more rapidly in the u2 -direction than the u1 -direction. x → ±∞ as y → 0. and are called the stable and unstable separatrices†.4(a) illustrates the situation when λ2 > λ1 . although the separatrices are straight trajectories † singular form. As the integral paths head away from the equilibrium point. are therefore the only ones that enter the equilibrium point. This means that y ˆ grows more quickly than x as t increases. all we have to do is consider the situation with the sense of t reversed. Remember. . The separatrices of saddle points usually represent the boundaries between diﬀerent types of behaviour in a phase portrait. x → 0 and ˆ y → ±∞ as t → ∞. When λ2 > 0 > λ1 .4(b) shows that when λ2 > λ1 . 9. shown in Figure 9.4. See if you can spot the saddle points in Figure 9. the coordinate axes are the only integral paths in the (ˆ. y )-plane that enter the equilibrium x ˆ point.2(b). as shown in Figure 9. The integral paths in the directions of the eigenvectors ˆ u1 and u2 . given by (9.

so that the equilibrium point is unstable. The remaining term in (9.8) shows that d¯/dt = y ¯ ¯ y x Xx (x0 . Let’s now consider what the phase portrait looks like when the eigenvalues λ1 and λ2 are complex. y0 ) = 0.14) means that the solution grows exponentially with t. y0 ) < 0. and hence the spiral is anticlockwise if Xx (x0 . A saddle point with λ2 > 0 > λ1 sketched in (a) the transformed and (b) the untransformed coordinate systems. There are two cases to consider. The general solution. (a) y (b) − y u2 u1 − x x Fig. (9. as shown in Figure 9. y0 )¯.5. whilst for λ1 = λ2 < 0 it is stable.4. so we can write λ = α ± iβ. but refer you to Exercise 9. and we will not consider it here. try looking on the positive y -axis.14) is oscillatory. 9. clockwise if Xx (x0 . with α and β real. When y = 0.11). they must be complex conjugate. it should be clear that when λ1 = λ2 > 0 the equilibrium point is unstable.14) . y0 ) > 0. it is easiest just to consider the sign of d¯/dt on the positive x-axis. then becomes u = eαt A1 u1 eiβt + A2 u2 e−iβt . This type of equilibrium point is called an unstable spiral or unstable focus. and we conclude that the integral paths spiral away from the equilibrium point. (9. (i) Eigenvalues with strictly positive real part (α > 0) The term eαt in (9. However. Since the eigenvalues are the solutions of a quadratic equation with real coeﬃcients.† † If Xx (x0 . To determine whether the sense of rotation is clockwise or anticlockwise.6. ¯ (9. The degenerate case where λ1 = λ2 is slightly diﬀerent from that of a stable or unstable node. they will start to bend as they head away from it and become governed by the full nonlinear equations.228 PHASE PLANE METHODS in the neighbourhood of the saddle point.

slightly diﬀerently. As before. for complex conjugate eigenvalues.3 SECOND ORDER EQUATIONS 229 Fig. and has the phase portrait shown in Figure 9. for example u1 . and write λ1 = µ + iω. Re(Ju1 )) = (Im(λ1 u1 ). Re(λ1 u1 )) = (µIm(u1 ) + ωRe(u1 ). anticlockwise spiral. we can deﬁne P = (Im(u1 ). but now JP = J (Im(u1 ). An unstable. we can transform the system into a more convenient form by deﬁning the transformation matrix. If we choose either of the eigenvectors. µRe(u1 ) − ωIm(u1 )) = P and hence P −1 JP = µ ω −ω µ . µ ω −ω µ . but with the sense of t reversed. (ii) Eigenvalues with strictly negative real part (α < 0) This is just the same as the previous case. ˙ and u = P v. but with the arrows reversed. Re(u1 )) .6. This is a stable spiral or stable focus. Note that. 9. P .9. . v = P −1 JP v.6. Re(u1 )) = (Im(Ju1 ).

with λ = ±iβ. We will not give the proof. v1 − 2µv1 + ω 2 + µ2 v1 = 0. (9. nonlinear terms. then v1 = µv1 − ωv2 . As we shall see in Chapter 13. In this case. As an example. . nonlinear. nonlinear. Theorem 9.7)? Yes. Note that all of the above analysis is rather informal. Can we guarantee that the behaviour that we have deduced persists when we study the full. v2 )T . onto and continuous and has a continuous inverse) from Rn to Rn deﬁned in ¯ a neighbourhood of x that maps trajectories of the nonlinear system to trajectories of the local linearized system. If at least one eigenvalue has zero real part. An important example where it may not be necessary to consider higher order terms in the Taylor expansions is when the eigenvectors are purely imaginary. terms in the Taylor expansions may be to make the local solutions spiral into or out of the equilibrium point. Each of the ﬁve cases we have discussed above is indeed an example of a hyperbolic equilibrium point. which holds for autonomous nonlinear systems of arbitrary order. then there is homeomorphism (a mapping that is one-toone. v2 = ωv1 + µv2 . 0) = 0 −g l 1 0 . However. the behaviour in the neighbourhood of a nonhyperbolic equilibrium point is determined by higher order. (9. with g x = y. the eﬀect of higher order. y = − sin x. We can formalize this in the following theorem. terms in the Taylor expansions of X and Y . We say that the equilibrium point is a linear centre. This type of equilibrium point is called a centre. ¨ ˙ This is the usual equation for damped simple harmonic motion.1 (Hartman–Grobman) If x is a hyperbolic equilibrium point of the ¯ ˙ nth order system x = f (x). there are many physical systems where a linear centre persists. but refer the interested reader to the original papers by Hartman (1960) and Grobman (1959). with natural frequency ω and damping given by −µ. but a nonlinear spiral. even in the presence of the higher order. On the other hand. eliminating v2 . provided that the equilibrium point is hyperbolic.14) with α = 0 shows that the solution is purely oscillatory. consider the simple pendulum. the equilibrium point is said to be nonhyperbolic. and is called a nonlinear centre. nonlinear system. ˙ ˙ l This has J= 0 1 − g cos x 0 l and J(0. ˙ ˙ and. The integral paths are closed and consist of a set of nested limit cycles around the equilibrium point. since it is based on a simple linearization about the equilibrium point.230 PHASE PLANE METHODS If v = (v1 . we can.

9. In other words. and a dot denotes d/dt. ˙ with F0 a constant force.3 SECOND ORDER EQUATIONS 231 The eigenvalues of J(0.7.3 An Example from Mechanics Consider a rigid block of mass M attached to a horizontal spring. A spring-mounted. the origin cannot be a spiral and must be a nonlinear centre. 9. the obvious answer is that there is a conserved quantity. the force exerted by the spring must exceed the frictional force for the block to move. Therefore. mass M Friction here x U Fig. x = U for xe − F0 /k x xe + F0 /k. and this occurs when k|x − xe | < F0 . as shown in Figure 9.7. We model the frictional force as F (x) = ˙ F0 −F0 for x < U . Here we know that if we map x → −x and t → −t. As we can see in Figure 9. This immediately gives us a solution. the phase portrait close to the origin is indeed a nonlinear centre. rigid block on a conveyor belt. and vice versa. xe is the equilibrium length of the spring. ˙ spring Block.2. . the block moves at the same speed as ˙ the conveyor belt. the energy. When x = U . and hence the phase portrait should not change. ˙ for x > U .3. ¨ ˙ where x is the length of the spring. k is the spring constant. Under this transformation. a stable spiral would become an unstable spiral. 0) are λ = ±i g/l. M x = F (x) − k(x − xe ). because of the reversal of time. which parameterizes the set of limit cycles. the equations are unchanged. F (x) is the frictional force exerted on the block by the conveyor ˙ belt. From Newton’s second law of motion and Hooke’s law. another way of looking at this is to ask whether the system has any symmetries. reﬂecting the phase portrait in the y-axis and reversing time. What is it about this system that allows the centre to persist in the full nonlinear analysis? Well. However. 9. The block lies ﬂat on a horizontal conveyor belt that moves at speed U and tries to carry the block away with it. since the phase portrait should not change.

17) We have left out the overbars for notational convenience. This system has a single ¯ equilibrium point at x = 1 + F0 . when it moves oﬀ on the limit cycle through the point D in Figure 9. so the equilibrium point is a linear centre. for y > U . The phase portrait is sketched in Figure 9. k (9.18) The solutions for y = U are therefore concentric circles. we obtain ¯ ¯ ¯ ¨ x = F − x + 1. with the spring stretched so far that it can immediately overcome the frictional force.15) as the system x = y. which carries it along with it until the spring is stretched far enough that the force it exerts exceeds the frictional force. xe . until it reaches C. For x < 1 − F0 and x > 1 + F0 . U= F0 = kxe xe We can now write (9. and the solution follows a diﬀerent circular trajectory. We conclude that an integral path that meets y = U with x in this range follows this trajectory until x = 1 + F0 . In fact. the slope of the integral paths is discontinuous there. Thereafter. Since y = 0 < U . The solution follows the circular trajectory until it reaches B. we have already seen that the line y = U for 1 − F0 x 1 + F0 is itself a solution. since x (x − 1 − F ) + y y = 0.17) is discontinuous at y = U . the solution remains on the periodic solution .8. M . At this point. y = F (y) − x + 1.8. At this point. If we now ¯ deﬁne dimensionless variables x = x/xe and t = t/ M/k. This occurs at the point D.15) ¯ ¯ ˙ We also have the possible solution x = U for 1 − F0 ¯ just two dimensionless parameters.232 PHASE PLANE METHODS Our model involves ﬁve physical parameters. (9. we can integrate to obtain ˙ ˙ {x − (1 + F0 )} + y 2 = constant 2 {x − (1 − F0 )} + y 2 = constant 2 for y < U . trajectories simply cross the line y = U . We now need to take some care with integral paths that meet the line y = U . ˙ ˙ M . the direction of the frictional force changes. x ¯ x ¯ (9. Since the right hand side of (9. The Jacobian matrix has eigenvalues ±i. the block is stationary relative to the conveyor belt. and we conclude that the equilibrium point remains a centre when we take into account the eﬀect of the nonlinear terms. k. y = 0. This corresponds to an initially stationary block. with x ˙ y ˙ = 0 −1 1 0 x y + 0 F0 + 1 .16) ¯ 1 + F0 . consider the trajectory that starts at the point A. ¯ ¯ ˙ > U. There are now (9. F0 and U . For example. F0 U ¯ ¯ . ¯ where ¯ ¯ ˙ F (x) = ¯ F0 ¯ −F0 for for ¯ ˙ x < U. However. the system is linear in the neighbourhood of this point.

A model for the way in which the two species interact with each other and their environment is x = x (A + a1 x + b1 y) . If P is a typical size of a population.3 SECOND ORDER EQUATIONS 233 through D. so that the more of species y that is available. b1 > 0. rigid block on a conveyor belt. b1 . and regard x and y as continuous functions of time. the faster the population of x grows. The phase portrait for a spring-mounted.4 Example: Population Dynamics Consider two species of animals that live on an island with populations X(t) and Y (t). In contrast. if A < 0. If species y competes with x for the available resources. a1 . b1 < 0. 9. a2 and b2 are constants. so the frictional force remains constant. so we can say that x does not rely on eating species y to survive. and therefore x must need to eat species y to survive. the population of species x grows when x 1 and y 1. If species x eats species y. y D B C • A x Fig. The term a1 x represents the eﬀect of overcrowding and competition for resources within species x.9. If A > 0. the population of x dies out. the speed of the block is always less than that of the conveyor belt. t. and the block undergoes simple harmonic motion.8. B. we can deﬁne x(t) = X(t)/P and y(t) = Y (t)/P as dimensionless measures of the population of each species. ˙ (9. On this periodic solution. so we require that a1 < 0.3. Let’s now consider what each of the terms that model the diﬀerence between the birth and death rates represents. ˙ y = y (B + b2 x + a2 y) . 9.19) where A. . The term b1 y represents the interaction between the two species. We can interpret each of these equations as Rate of change of population = Present population × (Birth rate − Death rate) .

0) = 3 0 −3 0 0 2 −3 −1 . and the integral paths are organized around them. The unit eigenvectors of P4 are u± = ( 1/3. For saddle points it is usually important to determine the directions of the eigenvectors. These are the lines on which either x = 0 ˙ or y = 0. At equilibrium points. and is therefore a saddle point. P2 has eigenvalues −1 and −2 and is therefore a stable node. The ﬁnal equilibrium point. ˙ (9. 0) and P4 = (1/2. P1 has eigenvalues 3 and 2 and is therefore an unstable node. The vertical nullclines. for example sheep and goats (see Exercise 9. as integral paths can only meet at equilibrium points.10 for an example of a predator– prey system). The four diﬀerent possibilities show that there are four equilibrium points. Xx Yx Xy Yy = 3 − 4x − 2y −2y −2x 2 − 2x − 2y . but they are not really important for sketching the phase por√ trait. 2). Speciﬁcally. so their eigenvalues can be read directly from the diagonal elements. The ﬁrst thing to do is to determine where any equilibrium points are. and what their types are. Equilibrium points are the main structural features of the system. 0). x = y = 0. where x = 0 (only y is varying so the integral ˙ ˙ paths are vertical on the nullcline). since these determine the directions in which the separatrices leave the equilibrium point. we can consider the nullclines. These separatrices are the only points that meet P4 . P3 has eigenvalues −3 and −1 and is therefore also a stable node. and hence ˙ ˙ (x = 0 or 2x + 2y = 3) and (y = 0 or 2x + y = 2) .234 PHASE PLANE METHODS We will consider two species that do not eat each other. but compete for resources. P1 = (0. 0) = . J(0. Next. This will tell us almost everything we need to know in order to sketch the phase portrait. ˙ y = y(2 − 2x − y). Since √ 2 > 1. P2 = (0. The Jacobian matrix is J= and hence J(0. . and we will see that determining their global behaviour is the key to sketching the full phase portrait. J(3/2.20) We will use this system to illustrate the full range of techniques that are available to obtain information about the phase portrait of second order systems. 2) = −1 −4 0 −2 . P3 = (3/2. J(1/2. Three of these matrices have at least one oﬀ-diagonal element equal to zero. 1) = −1 −1 −2 −1 . We could work out the direction of the eigenvectors for these three equilibrium points. P4 has one positive and one negative eigenvalue. are given by x = 0 or 2x + 2y = 3. P4 . ∓ 2/3)T . we will study the typical system x = x(3 − 2x − 2y). 1). has eigenvalues λ = λ± = −1 ± 2.

are examples of situations where two species competed for the same resources. in this case. either species x or species y dies out completely. for 0 < x < 3/2. We will conﬁne our attention to the positive quadrant. Apart from the stable separatrices of the saddle point P4 . We must ﬁrst deﬁne the MATLAB function . are given by ˙ y = 0 or 2x + y = 2. on the x-axis. there are some questions that we really ought to answer before we can be conﬁdent that Figure 9. and the extinction of Neanderthal man after the arrival of modern man in Europe. can we be sure that there are no limit cycle solutions? An associated question is. and the fact that integral paths can only cross at equilibrium points. Returning to our simple model. can we prove that we have sketched them correctly? More speciﬁcally. Depending upon the initial conditions. We can. Note that the equilibrium points lie at the points where the nullclines intersect.3 SECOND ORDER EQUATIONS 235 The horizontal nullclines. Firstly. the displacement of the native red squirrel by the grey squirrel in Britain.9(a). if we think that a system does possess a limit cycle. because we are not interested in negative populations.9(b). labelled S1 and S2 . what does the phase portrait look like far from the origin? We will consider some mathematical tools for answering these questions in the next four sections. where y = 0. and x < 0. also test whether our phase portrait is correct by solving equations (9. The directions of the arrows are determined by considering the signs of x and y at any point. as shown in Figure 9. if species x is absent initially. by competing for the same resources there can only be one eventual winner in the competition to stay alive. These separatrices therefore represent the boundary between two very diﬀerent types of behaviour. Moreover.9. Although neither species eats the other.9(b) is an accurate sketch of the phase portrait of (9. since the position of the stable separatrices of P4 is so important. This is clear from the biology of the problem. It is not necessary that one of the species kills the other directly. x = x(3 − 2x). can we prove it? Secondly. which means that x is decreasing ˙ with t. However. From all of this local information. and similarly for species y. all trajectories asymptote to either P2 or P3 as t → ∞. the nullclines are not integral paths. we can sketch a plausible and consistent phase portrait. since x = 0 ˙ when x = 0 and y = 0 when y = 0. there is no way for it to appear spontaneously.19) numerically in MATLAB. the fact that the coordinate axes are integral paths prevents any trajectory from leaving the positive quadrant. the coordinate axes are themselves integral ˙ paths.20). Although this is obviously a very simple model. ˙ ˙ For example. so that x > 0. which means that x is ˙ ˙ increasing with t. for x > 3/2. In general. Simply a suﬃcient advantage in numbers or enough extra eﬃciency in exploiting natural resources could have been enough for modern man to consign Neanderthal man to oblivion. however. how do we know that S1 originates at P1 and that S2 originates at inﬁnity? Finally. since. All of the information we have amassed so far is shown in Figure 9.

9. dy=dy’.y) dy(1) = y(1)*(3-2*y(1)-2*y(2)).9. dy(2) = y(2)*(2-2*y(1).20).    .236 PHASE PLANE METHODS (a) y P2 P4 P1 2x + y = 2 (b) y P3 x + y = 3/2 x P2 S2 P4 S1 P1 P3 x Fig.  function dy = population(t. and (b) the full phase portrait for the population dynamical system (9.y(2)). (a) Local information at the equilibrium points and the nullclines.

7). Figure 9.5 The Poincar´ Index e The Poincar´ index of an equilibrium point provides a very simple way of detere mining whether a given equilibrium point or collection of equilibrium points can be surrounded by one or more limit cycles. 9. However. that does not pass through any equilibrium points.9.y). Let Γ be any smooth. nonself-intersecting curve in the (x. when the initial value of y is suﬃciently large. For small enough initial values of y. Then [t y] = ode45(@population. and x → 0 as t → ∞. Fig. y = y(2) = 2. closed. autonomous ordinary diﬀerential equations given by (9. The dashed line is y. 9.19). there .10.19) for various initial conditions.3 SECOND ORDER EQUATIONS 237 which returns a column vector with elements given by the right hand sides of (9. [0 10]. y)-phase plane. y → 0 as t → ∞. with initial conditions x = y(1) = 1. as shown in Figure 9.10 shows the eﬀect of holding the initial value of x ﬁxed. the type of behaviour changes. in this case for 0 t 10. the solid line is x. not necessarily an integral path.3. The results can then be displayed with plot(t. The solution of (9. [1 2]) integrates the equations numerically. At any point P lying on Γ. Consider the second order system of nonlinear.9(b). consistent with the phase portrait sketched in Figure 9.11. and varying the initial value of y.

On the curve where Γ1 and Γ2 intersect. Proof This result is also obvious. Proof This result is obvious. we call nΓ the Poincar´ index. which must be an integer.11. (iv) If Γ is a limit cycle then nΓ = 1. the Poincar´ index cannot change e from its initial value. (ii) If Γ does not enclose any equilibrium points. e We can now deduce some properties of the Poincar´ index. anticlockwise circuit around Γ.238 PHASE PLANE METHODS is a unique integral path through P . . On this circle. 9.12. The deﬁnition of the Poincar´ index. but must also be an integer. e (i) If Γ is continuously distorted without passing through any equilibrium points. (v) If integral paths either all enter the region enclosed by Γ or all leave this region. e Γ P ψ Fig. then Γ1 and Γ2 . If ψ changes by an amount 2nΓ π as P makes one complete. Proof Using the previous result. that it makes with the horizontal (tan ψ(P ) = Y /X). ψ(P ). so these contributions cancel and nΓ = nΓ1 + nΓ2 . nΓ = 1. and therefore nΓ = 0. The angle ψ varies continuously as P moves around Γ. (iii) If Γ is the sum of two closed curves. the amount by which ψ varies on traversing Γ1 is equal and opposite to the amount by which ψ varies on traversing Γ2 . we can continuously shrink Γ down to an arbitrarily small circle without changing nΓ . nΓ = nΓ1 + nΓ2 . nΓ does not change. with a well-deﬁned direction that we can characterize using the angle. Proof Consider the curves shown in Figure 9. ψ(P ) is almost constant. nΓ = 0. Proof Since nΓ must vary continuously under the action of a continuous distortion.

20). result (iv) shows that the sum of the Poincar´ indices of the equilibrium points enclosed by a e limit cycle must be 1. A corollary of this is that a limit cycle must enclose at least one node.12.5. and this is what all of this has been leading up to. since these are integral paths. Now result (iii) e e shows that nΓ is simply the sum of the Poincar´ indices of all the equilibrium points enclosed by Γ. The sum of two curves. but we will not consider these here. (9. Proof Γ can be continuously deformed into a small circle surrounding the saddle point. and hence nΓ = −1. focus or nonlinear centre. focus or nonlinear centre in the positive quadrant of the phase plane. In particular. If a limit cycle exists. Finally. nΓ = −1. there is no node. (vi) If Γ encloses a single node. However. The linearized solution then shows that integral paths either all enter or all leave the region enclosed by Γ. Γ can be continuously deformed into one of the limit cycles that encloses the equilibrium point. a node. focus or nonlinear centre has Poincar´ index n = 1. and we conclude that no limit cycle exists. e whilst a saddle has Poincar´ index n = −1. The direction of the integral paths makes a single clockwise rotation as P traverses Γ.9. . Now let’s return to our example population dynamical system. These results show that we can deﬁne the Poincar´ index of an equilibrium e point to be the Poincar´ index of any curve that encloses the equilibrium point and e no others. Γ can be continuously deformed into a small circle surrounding the equilibrium point. (vii) If Γ encloses a single saddle point. nΓ = 1. For a nonlinear centre. 9. where the linearized solution gives the phase portrait shown in Figure 9. focus or centre. and result (iv) shows that nΓ = 1. and the previous result shows that nΓ = 1.3 SECOND ORDER EQUATIONS 239 Γ1 Γ2 Fig. Proof For a node or focus. it cannot cross the coordinate axes. Nonhyperbolic equilibrium points can e have other Poincar´ indices.

which is a contradiction. focus or nonlinear centre. there are no limit cycles in any simply-connected region of the (x. We now consider a way of doing this that works for many types of system. Note that the restriction of this result to simply-connected regions (regions without holes) is crucial. Proof Let Γ be a limit cycle. y) is continuously diﬀerentiable in some simply-connected region D in the (x. .20). y)). Y (x.2 (Bendixson’s negative criterion) For the second order system of nonlinear. Stokes’ theorem states that D ∂Y ∂X + ∂x ∂y dx dy = Γ (Xdy − Y dx) . there are no limit cycle solutions if (ρX)x + (ρY )y does not change sign. the Poincar´ e index can be used to determine which sets of equilibrium points a limit cycle could enclose if it existed. y). Theorem 9. If we now apply this theorem to the system (9. y)-phase plane. and hence the result is proved. where D is the region enclosed by Γ. then the integral is either strictly positive or strictly negative†.3. Although this tells us that any limit cycle solution must be intersected by the line 6x + 4y = 5. and hence D ∂X ∂Y + ∂x ∂y dx dy = 0. We need a more powerful version of Bendixson’s negative criterion. However.7).3 (Dulac’s extension to Bendixson’s negative criterion) If ρ(x.6 Bendixson’s Negative Criterion and Dulac’s Extension Theorem 9. 9. we can write the right hand side of this as X Γ dx dy −Y dt dt dt = Γ (XY − Y X) dt = 0. If Xx + Yy does not change sign in D.240 PHASE PLANE METHODS When a system does possess a node. For the vector ﬁeld (X(x. which is provided by the following theorem. if indeed no limit cycles exist. we ﬁnd that Xx + Yx = 5 − 6x − 4y. but it would be useful to have a way of ruling out the existence of limit cycles altogether. † ignoring the degenerate case where Xx + Yy ≡ 0 in D. y)-phase plane where Xx + Yy does not change sign. autonomous ordinary diﬀerential equations given by (9. we cannot rule out the possibility of the existence of limit cycles in this way.

so that all integral paths enter S. but. ∂x ∂y x y and hence that there can be no limit cycle solutions in the positive quadrant. of the phase plane with outward unit normal n that has (X.9. and can be omitted on ﬁrst reading.20). Y ) · n 0 on its boundary. in which integral paths that represent chaotic solutions can indeed wander around indeﬁnitely without being attracted to a periodic solution. any subset.3 SECOND ORDER EQUATIONS 241 Proof Since ∂ (ρX) ∂ (ρY ) + ∂x ∂y dx dy = Γ ρ (Xdy − Y dx) . As we shall see. Note that it is crucial that the region I be bounded.7 The Poincar´–Bendixson Theorem e We begin with a deﬁnition. it cannot be simply-connected. Returning again to (9. of the phase plane. e I.7). Now that we know how to try to rule out the existence of limit cycles. do indeed behave as sketched. is a positively invariant set. as expected. Similarly. then I contains at least one limit cycle. S. as is a limit cycle. if I contains a limit cycle. of the phase plane and all integral paths in I − when t = 0 remain there when t < 0.3.9. shown in Figure 9. As a less prosaic example. D the proof is as for Bendixson’s negative criterion. we shall see in Chapter 15 that the Poincar´–Bendixson theorem does not hold for third and higher e order systems. . and I contains no equilibrium points. a negatively invariant set is a closed subset. 9. We ﬁnd that 1 2 ∂ (ρX) ∂ (ρY ) + = − − < 0. I − . Consider the second order system (9. The details of the proof are rather technical. y) is continuously diﬀerentiable. For example. we can choose ρ = 1/xy and consider the positive quadrant where ρ(x. in contrast to Bendixson’s negative criterion. we say that I + is a positively invariant set.4 (Poincar´–Bendixson) If there exists a bounded. invariant region. I does not need to be simply-connected. The Poincar´–Bendixson e theorem says that the integral paths in I cannot wander around for ever without asymptoting to a limit cycle. an equilibrium point is both a positively and a negatively invariant set. Although this seems obvious. If I + is a closed subset of the phase plane and any integral path that lies in I + when t = 0 remains in I + for all t 0. Theorem 9. This will e also allow us to prove that the stable separatrices of P4 . how can we go about showing that limit cycles do exist for appropriate systems? We will consider this in the next section using the Poincar´–Bendixson theorem.

13(b)). x0 ) be the solution of (9. lies in D. There are two possibilities. all the points on an integral path that enters a stable node or focus at x = xe have xe as their only ω-limit point. xi−2 . {ti }. Secondly. x0 ). 9. which is denoted by ω(x0 ). the integral path through xi enters D and remains there (see Figure 9.242 PHASE PLANE METHODS Proof of the Poincar´–Bendixson Theorem e We will assume that I is a positively invariant region. then xi lies in the segment of L between xi−1 and xi+1 . We call the set of all ω-limit points of x0 the ω-limit set of x0 . all points on an integral path that asymptotes to a stable limit cycle as t → ∞ have all the points that lie on the limit cycle as ω-limit points. Lemma 9. and xi−1 therefore lies in the segment of L between xi−2 and xi . . and L ⊂ I a transversal through x0 ∈ I. x(t. We now need two further deﬁnitions. (a) xi − 1 xi xi + 1 L (b) L xi xi − 1 xi − 2 D Fig. Similarly. x0 ) meets L for the ith time. The proof when I is negatively invariant is identical. We say that L is a transversal. In other words. Firstly.13(a)). lies in D. In this case. Proof Consider the invariant region D bounded by the integral path from xi−1 to xi and the segment of L between xi−1 and xi . it is always possible to construct a transversal through x0 . that passes through x0 when t = 0 intersects L in a monotone sequence. If x0 is not an equilibrium point. since the slope of the integral paths is well-deﬁned in the neighbourhood of x0 . the integral path through xi−1 originates in D (see Figure 9. In this case.1 Let I be a positively invariant region in the phase plane. Let x(t. We say that x1 is an ω-limit point of x0 if there exists a sequence. Let L be a ﬁnite curve such that all integral paths that meet L cross it in the same direction. such that ti → ∞ and x(ti . if xi is the point where x(t. x0 ) → x1 as i → ∞.7) with x = x0 when t = 0. For example.13. and xi therefore lies in the segment of L between xi−1 and xi+1 . xi+1 . if it exists. if it exists. but with time reversed. The intersection of an integral path in an invariant region with a transversal. The integral path.

intersects L at most once. by Lemma 9. lying in L such that {ˆ i } → x and {¯ i } → x as x i → ∞. let x2 ∈ ω(x1 ). by Lemma 9. since it relies crucially on the fact that a closed curve separates an inside from an outside (the Jordan curve theorem). the intersections of the integral path through x0 with L should be monotonic. y = r sin θ.2. such that x1i → x2 as i → ∞. x1 ). e Proof Let x0 be a point in I. r = (x2 + y 2 )1/2 and x = r cos θ. However. By Lemma 9. θ). From the deﬁnitions.4.2 Consider the region I. the integral path though x1 must intersect L at a monotonic sequence of points x1i . If x1 ∈ x(t. so the integral path through x0 must pass arbitrarily close to each of the points x1i as t → ∞. and hence is a closed curve. the chain rule gives r = (xx + y y)(x2 + y 2 )−1/2 = cos θ x + sin θ y. But x1i ∈ ω(x0 ).9. ˙ (9. Proof Since x1 lies in x(t. Lemma 9. since Lemma 9. ω(x0 ).22) . the integral path through x1 . x0 ). and we conclude that x1i = x2 . ˙ y = x + y − y(x2 + y 2 ). since there are no equilibrium points in I. and hence that ω(x1 ) is the closed limit cycle through x1 .1. ˙ ˙ ˙ ˙ ˙ For (9. is a closed curve. ˆ ¯ Proof Suppose that ω(x0 ) intersects L twice. and hence x1 ∈ ω(x1 ). Let L be a transversal through x2 . x0 ) ⊂ I and ω(x0 ) ⊂ I. at x and x.3 If x1 ∈ ω(x0 ) is not an equilibrium point and lies on x(t. Since x2 ∈ ω(x1 ). x0 ).21) In order to analyze these equations. the Poincar´–Bendixson theorem. x(t. then x(t.21). the integral path through x0 . ω(x1 ) = ω(x0 ).1. Hence the result is proved by contradiction. Let L be a curve through x1 transverse to the integral paths. x0 ). Lemma 9. x0 ). the integral path through x1 is a closed curve and. this cannot occur. The ω-limit set of x0 . If x1 ∈ x(t.3 SECOND ORDER EQUATIONS 243 Note that Lemma 9.1 does not hold in systems of higher dimension. must be a limit cycle.3. (r. also passing through x0 . x1 ) can only meet L once. {ˆ i } and {¯ i }. it is convenient to write them in terms of polar coordinates. transverse curve L and point x0 ∈ I deﬁned in Lemma 9. However. We can now ﬁnally prove Theorem 9. r = cos θ r cos θ − r sin θ − 2r3 cos θ + sin θ r cos θ + r sin θ − r3 sin θ ˙ (9. Choose x1 ∈ ω(x0 ). then. and hence x(t. Example Consider the system x = x − y − 2x(x2 + y 2 ). We can therefore ﬁnd ˆ ¯ x x x sequences of points.1 states that these intersections must be monotonic.

and therefore integral paths are directed away from the origin.21). e Γ = Γ0 Γ = Γ1 Ι Fig. ˙ √ Therefore r > 0 for 0 < r < 1/ 2 and r < 0 for r > 1. θ) | r0 r r1 } √ with 0 < r0 < 1/ 2 and r1 > 1.23) . Remembering that if r > 0. ˙ 1 cos θ r cos θ + r sin θ − r3 sin θ − sin θ r cos θ − r sin θ − 2r3 sin θ θ= r 1 = 1 + r2 sin 2θ. such that all integral paths enter the region I. Since 0 cos2 θ 1. 9. we conclude that r(1 − 2r2 ) r r(1 − r2 ). ˙ θ= For (9. we can deﬁne a closed. 2 2 1 + (y/x) x r (9.14. 2 xy − y x ˙ ˙ 1 1 ˙ ˙ = (cos θ y − sin θ x) . If we can show that there are no equilibrium points in I. as shown in Figure 9. From the deﬁnition of θ = tan−1 (y/x). bounded. the Poincar´–Bendixson theorem shows that at least one limit cycle exists in I. The region I that contains a limit cycle for the system (9. r ˙ ˙ ˙ is an increasing function of t. annular region D = {(r.21).244 PHASE PLANE METHODS = r − r3 1 + cos2 θ .14.

A corollary of the Poincar´–Bendixson theorem is that. there is at least one limit cycle in I. there are no equilibrium points in I. In general. P4 .15. so that this projection is useful for studying the behaviour of a second order system for x 1. x2 + y 2 = r0 . if I is a closed. which is enclosed by any limit cycles. v). but the most useful is the Poincar´ projection. 9. Therefore. y) | x 0. and I contains no limit cycles. and hence that there √ no equilibrium points for 0 < r < 2. In particular. it cannot be simply-connected. 2x + y 2. focus or centre.20). y)-plane. In order to relate u and v to x and y. and denote the point a unit distance above the origin as O . e invariant region of the phase plane.3. with r0 > 3. We label the origin of the (x. There are several ways of doing this. as shown in Figure 9. if we can construct a region in the phase plane that traps a limit cycle in this way. and hence. The separatrix S2 lies in R2 in the neighbourhood of the saddle point.9. In this example. we conclude that θ > 0 √ provided r < 2. We can set up a Cartesian coordinate system. Consider the region R1 = {(x. We now consider a point B in the (x. y 0. bounded. P4 . Since we can make r0 arbitrarily large. C approaches the u-axis. we ﬁrstly consider the triangles OAB and . y)-plane. (u. as shown in Figure 9. We can use similar ideas to determine the positions of the stable separatrices of P4 for our example population dynamical system. u-axis in the same direction as the y-axis and v-axis vertically upwards. as x → ∞. provided that are e 1 < r1 < 2. y)-plane as O. 2x + 2y 3. Consider the plane that passes e through the line x = 1. by the Poincar´– Bendixson theorem. √ 245 shown in Figure 9. Now consider the region R2 = (x. x2 + y 2 2 r0 . (9.3 SECOND ORDER EQUATIONS ˙ Since −1 sin 2θ 1. with origin a unit distance above x = 1. and we conclude that S1 must enter R2 through its 2 curved boundary. y = 0. No integral path enters the region R2 through any of its straight boundaries. v)-plane. The separatrix S1 lies in R1 in the neighbourhood of the saddle point. 2x + 2y 3} . The Poincar´ projection maps the point B to the point C where e the line O B intersects the (u. y) | x 0.16. but which does not lie within I. it is often useful to map the phase plane to a new coordinate system where the point at inﬁnity is mapped to a ﬁnite point. No integral path enters the region R1 through its boundaries.8 The Phase Portrait at Inﬁnity In order to study the behaviour of a second order system far from the origin. y 0. in this plane. Note that we know that a limit cycle must enclose at least one node. we conclude that S2 originates at inﬁnity. P1 . 2x + y 2. there is an unstable focus at r = 0. and we conclude that S1 must originate at the equilibrium point at the origin. perpendicular to the (x.15. we can deduce that all of the integral paths that meet the boundary of I terminate at an equilibrium point contained within I.

u = x/y.25) (9. and hence ˙ u = uv x − v y. which show that (1 + v)/m = 1/ x2 + y 2 . The nature of these equilibrium points remains the same after the projection. Figure 9. but then move away. ˙ ˙ ˙ ˙ is therefore a Poincar´ projection. The regions R1 and R2 for the population dynamical system (9. v = −2/3. whilst v < 0.15. which corresponds to P3 .20).20). By eliminating m. The analogous transformation. shown in Figure 9.25) close to the u-axis for v < 0 and u > 0.20) for x 1 lead into the ﬁnite (x.9(b).20) for x 1.24) This has ﬁnite equilibrium points at u = 0. v = −2. we need to consider (9. v)-plane. so that P3 is a stable node and P4 is a saddle point. ˙ Integral paths close to the u-axis therefore start parallel to it. OA B . ˙ v (9. v = v 2 x. we ﬁnd that u = −u 1 + ˙ u . and can be used to examine the behaviour for y 1. 9. and u = 2. v = −1/y is also a Poincar´ e projection. which corresponds to P4 . We conclude that the integral paths of (9. u ˙ 1.17 is a sketch of the phase portrait in the (u. In order to determine the behaviour of (9. y)-plane as sketched in Figure 9.16. . This shows that u = y/x and x2 + y 2 − m = x2 + y 2 /x. e If we apply this to the population dynamics example (9. we consider the similar triangles BB C and BOO . v = −1/x. which are similar. v = −(2u + 3v + 2).246 PHASE PLANE METHODS y R2 P4 S1 S2 R1 2x + y = 2 2x + 2y = 3 x Fig. When −v 1. The simple change of variables u = y/x. Secondly. we ﬁnd that v = −1/x.

y) is called the Hamiltonian. y) = constant. which we will now investigate.9 A Final Example: Hamiltonian Systems Let Ω be a region of the (x. 2 Hamiltonian systems have several general properties.3. A Poincar´ projection. which we studied earlier. 9. y). y(t)). 9.5 The integral paths of a Hamiltonian system are given by H(x. One example is the simple pendulum. this has x = y = Hy .26) is called a Hamiltonian system and H(x.9. the total energy. Theorem 9. 0) m y x B = (x. y(t)) and dH = Hx x + Hy y = Hx Hy − Hy Hx = 0. y)-plane and H : Ω → R be a real-valued. As we have seen. y).3 SECOND ORDER EQUATIONS 247 v u y O′ C = (u. The two-dimensional system of ordinary diﬀerential equations. is a Hamiltonian for the system. e We have now answered all of the questions concerning the phase portrait of (9.20) that we asked at the end of Section 9. . y = −ω 2 sin x = −Hx . H = H(x(t). y) Fig. ˙ ˙ so that H = 1 y 2 − ω 2 cos x. Proof On an integral path (x(t). ˙ (9.3. 0) A = (x.16. x = Hy (x.4. ˙ y = −Hx (x. v) 1+ v B′ O 1 A' = (1. Such systems occur frequently in mechanics. continuously diﬀerentiable function deﬁned on Ω. ˙ ˙ dt so that H is constant.

the integral paths are therefore given by the curves y 2 − 2ω 2 cos x = constant. 2 2 Note that Hxy = Hxx Hyy at x = xe since the eigenvalues are nonzero.248 PHASE PLANE METHODS v u P3 P4 Fig. the eigenvalues of J satisfy 2 λ2 = Hxy (xe . ye ). If Hxy > Hxx Hyy at x = xe . the level curves of H are closed and surround x = xe . A sketch of the phase portrait of the system of equations (9. there are two complex conjugate imaginary eigenvalues. then it is either a saddle point or a centre. Since the integral paths are given by H = constant and the 2 conditions Hx = Hy = 0 and Hxx Hyy > Hxy are those for a local maximum or minimum at x = xe . If Hxy < Hxx Hyy at x = xe . ye )Hyy (xe . We conclude that the equilibrium point is a nonlinear centre. .6 If x = xe is an equilibrium point of a Hamiltonian system with nonzero eigenvalues.25). For the simple pendulum.17. there is one positive and one negative eigenvalue. Proof The Jacobian of (9. so xe is 2 a saddle point. Theorem 9. ye ) − Hxx (xe . x=xe Since Hxy = Hyx .26) at x = xe is Hyx −Hxx Hyy −Hxy J= . 9.

For example.9. z). There are also some interesting. y. e nor. |δx2 | 1 (δx1 × δx2 ) . We will examine some elementary techniques for studying chaotic solutions in Chapter 15. z). so the area of the triangle is unchanged under the action of a Hamiltonian system. 2. z)-phase space. is there an analogue of the Poincar´–Bendixson theorem. most of the useful results that hold for the phase plane do not hold in three or more dimensions. which have fractal or noninteger dimensions. where x0 + δx1 and x0 + δx2 . However.4 THIRD ORDER EQUATIONS 249 Theorem 9. A simple way to get a grasp of this is to remember that cars.27) can be analyzed in terms of integral paths in a three-dimensional. 9. e In third or higher order systems. z). which drive about on a plane. The area of this triangle is |A|. Proof Consider a small triangle with one vertex at x0 .26). and |δx1 |. . ˙ z = Z(x. but aircraft. in three or more dimensions. and useful. integral paths can be attracted to strange or chaotic attractors. y. which have an extra dimension to use. Up to O(|δxi |). There is no analogue of the Poincar´ index or Bendixson’s negative criterion. often hit each other. Since any area in the phase plane can be broken up into inﬁnitesimal triangles. a Taylor expansion shows that A= d δxi ∼ (δxi Hyx (x0 ) + δyi Hyy (x0 ). we therefore have 2 d d dA = δx1 × δx2 + δx1 × δx2 dt dt dt = δy2 {δx1 Hyx (x0 ) + δy1 Hyy (x0 )} + δx2 {δx1 Hxx (x0 ) + δy1 Hxy (x0 )} −δy1 {δx2 Hyx (x0 ) + δy2 Hyy (x0 )} − δx1 {δx2 Hxx (x0 ) + δy2 Hxy (x0 )} = 0.4 Third Order Autonomous Nonlinear Ordinary Diﬀerential Equations The solutions of the third order system x = X(x. and represent chaotic solutions. closed curves no longer divide the phase space into two distinct regions.7 (Liouville’s theorem) Hamiltonian systems are area-preserving. (x. y. dt for i = 1. 2 If all of the vertices move along integral paths of the Hamiltonian system (9. do so more rarely. but this is not the case in a three-dimensional phase space. y.4). ˙ y = Y (x. where δxi = (δxi . ˙ (9. integral paths inside a limit cycle in the phase plane are trapped there. as we noted earlier. the Hamiltonian system is area-preserving. inside and outside the curve. conservative third order systems for which a more straightforward analysis is possible (see also Exercise 13. δyi ). The crucial diﬀerence is that. −δxi Hxx (x0 ) − δyi Hxy (x0 )) . the other two vertices at 1.

28). dt dt and hence k 2 y 2 − z 2 and k 2 x2 + z 2 are constant on any integral path. These remain centres when nonlinear terms are taken into account. The phase portrait is sketched in Figure 9. using the argument that we described earlier for the simple pendulum. 0   0 ±1 0 0 . t → −t. . ˙ z = −k 2 xy. Consider the integral paths on the surface of the cylinder x2 + y 2 = 1. and hence all integral paths lie on the surface of a cylinder with its axis pointing in the z-direction. 0)T . The zero eigenvalue. J =  Yx Yy Yz  =  −z 2 2 −k y −k x Zx Zy Zz 0 This gives   0 ∓1  . We can also see that d 2 (x + y 2 ) = 2xx + 2y y = 2xyz − 2yxz = 0. 0) and (±1. This surface is two-dimensional. ˙ y = −xz. ±1. 0) each have eigenvalues λ = 0. We can conﬁrm that this phase portrait is correct by noting that this system actually has two other conserved quantities. 1. 0. the points (0. ˙ ˙ dt Therefore.or z-axis is an equilibrium point. where the equilibrium points are saddles. Finally. y. so that x= ˙ 1 − x2 1 − k 2 x2 .18. corresponds to the fact that the y-axis is completely made up of equilibrium points. x2 + y 2 is constant on any integral path.250 PHASE PLANE METHODS Example Consider the third order system x = yz. any point on the x-. 0 0 0 J(±1. consider the integral path with x = 0. The remaining two eigenvalues are real and of opposite sign. ˙ (9. From (9. the hyperboloidal cylinder k 2 y 2 − z 2 = constant. 0) =  0 ∓k 2 The points (±1. y = z = 1 when t = 0. 0) and the Jacobian matrix is     Xx X y X z 0 z y 0 −x  .18 shows.28) In this rather degenerate system. so we should be able to analyze the behaviour of integral paths on it using phase plane techniques. and the elliptical cylinder k 2 x2 + z 2 = constant. d d 2 2 (k y − z 2 ) = (k 2 x2 + z 2 ) = 0. There are equilibrium points at (0. k 2 x2 + z 2 = 1 and x2 + y 2 = 1. 0. ±k. Similarly. and control the dynamics on the cylinder x2 + y 2 = 1. 0. On this integral path. with eigenvector (0. since the system is unchanged by the transformation z → −z. Integral paths therefore lie on the intersection of the circular cylinder x2 + y 2 = constant. 0) each have eigenvalues λ = 0. ±1. and are therefore linear centres on x2 + y 2 = 1. This is precisely what the phase portrait in Figure 9. 0) =  0 0 0 0 ∓k 2 0 J(0. ±ik. ±1.

which gives† x t= 0 √ 1− s2 ds √ . In Sections 12. By integrating this equation once. 1 − k 2 s2 This is the deﬁnition of the Jacobian elliptic function sn(t . Determine the coordinates of the two equilibrium points and show that there is a homoclinic orbit associated with one of them. k).2.EXERCISES 251 Fig. What types of behaviour occur inside and outside the homoclinic orbit? † Note that the properties of the particular integral path with x = 0. k) and z = dn(t . On this integral path y and z are also Jacobian elliptic functions. 9.3 and 12.1 Consider the second order.2. autonomous ordinary diﬀerential equation x = 3x2 − 1. k). provided that k = 1. The Jacobian elliptic functions will also prove to be useful in Section 12. y = z = 1 when t = 0 ensure that the arguments of the square roots remain positive. ¨ where a dot represents d/dt. . Exercises 9. y = cn(t . and secondly when k 1. Sketch the phase portrait in the (x.2. ﬁrstly when k is close to unity. k).5. The phase portrait that we have just determined now allows us to see qualitatively that these elliptic functions are periodic with t.18. The phase portrait of the third order system (9.28) on the surface of the cylinder x2 + y 2 = 1 when k = 2. x)-phase ˙ ˙ plane. obtain a relation between x and x.4 we will develop asymptotic expansions for sn(t .

(f) 9.3 . whilst when j = 2 it is an improper node. Determine the equation of motion and write it as a pair of ﬁrst order equations.2 By sketching the curve x = X(x). 9. 9. for A = (a) (c) (e) 1 1 −4 3 4 3 −5 −1 2 −2 −2 −1 . y = x + y − 2xy 2 . ˙ ˙ 9.252 PHASE PLANE METHODS 9. the equilibrium point at the origin is known as a star.8 Consider the system of ordinary diﬀerential equations dx dy = x x2 y 2 + 1 . ˙ ˙ (c) x = y 2 − 1. Sketch a phase portrait in which (a) a separatrix connects the saddle points. and hence sketch the phase portrait in each case. y = sin x. qualitative or quantitative? Find the eigenvalues and eigenvectors of the matrix A. λ. (b) no separatrix connects the saddle points. dt dt .5 9. By using the Cayley–Hamilton theorem. What happens after a long time? Find all of the equilibrium points of each of the following systems. y = y 2 − x2 . and then sketch the ˙ phase portrait of the linear system u = Au. Sketch the phase portrait. ˙ ˙ (b) x = −3y + xy − 10.4 ˙ Consider a second order linear system u = Au. Note that in the case j = 1. . and determine their type. Sketch the phase portrait in each case. Now check that this qualitative analysis is correct by actually solving each equation with initial conditions x = x0 when t = 0. Show that the equilibrium points are either stable points or saddle points. (a) x = x − y. determine the equilibrium points and ˙ corresponding domains of attraction when (a) X(x) = x2 − x − 2. where u = (x. The weight at the end of a simple pendulum experiences a frictional force proportional to its velocity. = y (y − 2 − x) .6 9. y)T . show that there must exist a linear transformation that takes A to either Λ1 = λ 0 0 λ or Λ2 = λ 0 1 λ . (c) X(x) = sin x. when the constant matrix A has two equal eigenvalues. Which method do you think is easier to use. (b) 1 1 1 −2 1 2 2 2 2 1 −1 1 . A certain second order autonomous system has exactly two equilibrium points. both of which are saddles. . (b) X(x) = e−x − 1. (d) .7 ˙ Solve the linear system of equations v = Λj v for j = 1 and j = 2.

dt 9.10 can be used to model the population of sheep (s) and rabbits (r) in a closed ecosystem. Consider the system dy dx = x(x2 + y 2 − 1). dt ds = s(2 − r − s).EXERCISES 253 9.12 Sketch a possible phase portrait in each of the cases where a limit cycle can surround the equilibrium points. dt dt Show that the x. sketch the phase portrait for r > 0 and s > 0. y)plane. (b) the Poincar´ index. a stable node and an unstable focus. 9. y) has a local minimum at one of the equilibrium points. e Sketch the phase portrait. a saddle point. two saddle points. Assuming that there are no limit cycles. Show that there are no limit cycle solutions using (a) Dulac’s extension to Bendixson’s negative criterion with auxiliary function ρ(x. ˙ 9.9 Determine the position and type of each equilibrium point in the (x. and hence deduce that it is a nonlinear centre. (a) (b) (c) (d) an unstable node. . If x = −1 and y = −1 when t = 0. b1 and b2 positive. how does the solution behave as t → ∞? The second order system dr = r(3 − r − s). Sketch the phase portrait. predator species and its herbivorous prey in a closed ecosystem. y) = 1/xy. models the populations of a carnivorous. Find the directions of the separatrices close to any saddle points. = y(x2 + y 2 − 2). taking care to ensure that your sketch is consistent with the position of the horizontal nullcline. Determine dy/dx as a function of x and y. What happens to the populations? Use the concept of the Poincar´ index to determine which of the following e can be surrounded by a limit cycle in the phase portrait of a second order system.11 with the constants A. ˙ y = y(B − b2 x). and integrate once to obtain an equation of the form E(x. Sketch the phase portrait. B. a saddle point. Which animal becomes extinct? Explain how the system x = x(−A + b1 y). Which variable is the predator population. Show that the coordinate axes are integral paths. an unstable focus and a stable node. y) = constant. Determine the position and type of all the equilibrium points.and y-axes are integral paths. x or y? Determine the type of each of the equilibrium points. Show that E(x.

y = 1 + x + y 2 .13 Use the Poincar´ index. y) = eax+by . use Dulac’s extension to Bendixson’s negative criterion with auxiliary function ρ(x. ˙ (a) x = 2x + 2y − x(2x2 + y 2 ). dt dy = x + y + xy 2 − 2yx4 − 2y 3 x2 − y 5 dt in terms of polar coordinates (r.14 (Hint: For (f). ˙ ˙ ˙ x = 1 − x3 + y 2 . ˙ ˙ x = 2xy + x3 . = x + y − (x2 + 2y 2 )y dt dt in terms of polar coordinates. and use the Poincar´–Bendixson theorem to show that at least one limit cycle e solution exists. ˙ 9. ˙ ˙ (b) x = x − y − x(x2 + 3 y 2 ). and use the Poincar´–Bendixson e theorem to show that at least one limit cycle solution exists. (b) Write the system of ordinary diﬀerential equations dy dx = xy − x2 y + y 3 . θ). θ). = −x − y + x2 + y 2 . Show that there is a single equilibrium point at the origin and that it is nonhyperbolic. y = −2x + 2y − y(2x2 + y 2 ). (a) (b) (c) (d) (e) (f) x = y.) Write each of the following systems in terms of polar coordinates (r. dt dt (a) Write the system of ordinary diﬀerential equations dx = x − y − y 3 − 2x5 − 2x3 y 2 − xy 4 . ˙ ˙ x = −(1 − x)3 + xy 2 . y = −x − y + x2 + y 2 . = y 2 + x3 − xy 2 dt dt in terms of polar coordinates. y = y + y 3 . ˙ 2 ˙ x = y + x .15 (a) Write the system dy dx = x − y − (2x2 + y 2 )x. θ). ˙ 2 2 9. Sketch the phase portrait. y = 2xy. Show that dr/dt = 0 when θ = 0 or π. (r. y = x + y − y(x2 + 1 y 2 ). (b) Use Dulac’s extension to Bendixson’s negative criterion (with an auxiliary function of the form eax+by for some suitable constants a and b) to show that there is no limit cycle solution of the system with dy dx = y. Show that the lines θ = ±π/4 and θ = ±3π/4 are integral paths. Bendixson’s negative criterion or Dulac’s extene sion as appropriate to show that the following systems have no limit cycle solutions. ˙ x = x. y = −x2 + y − y 2 + y 3 . 9.16 . and then use the Poincar´–Bendixson e theorem to show that there is at least one limit cycle solution.254 PHASE PLANE METHODS 9. ˙ y = 1 + x2 − (1 − x)y.

Find the equilibrium point that corresponds to a small perturbation of the Newtonian case ( = 0). ds ds which arises in a thermal ignition problem (see Section 12. one of which is a saddle. and classify them. 2 dθ h (E9. Show that this separatrix asymptotes to the other equilibrium point as s → −∞. dθ2 h where h is the angular momentum of the particle.2) numerically. nonlinear system? (b) If the particle moves at relativistic speeds. and the term u2 is called Einstein’s correction.1) is modiﬁed. = p + q − 2. (Hint: First show that S must meet the p-axis with p > 2. Consider S. it can be shown that (E9.17 Consider the second order system dp dq = −pq. to γ d 2u + u = 2 + u2 .2.2) where is a small positive constant. α = 2. .) Project A particle of mass m moves under the action of an attractive central force of magnitude γm/rα . What feature of the linear approximation will carry over to the solutions of the full. du/dθ)-phase plane. θ) are polar coordinates and γ and α are positive constants. appropriate to a gravitational attraction. γ and h appropriate to each of the three planets nearest to the Sun (you’ll need to ﬁnd an appropriate astronomy book in your library). where (r. Next show that S must go on to meet the p-axis with p < 2. (E9. Show that there are just two ﬁnite equilibrium points.1) 9.EXERCISES 255 9. and relate what you obtain to part (b) above. Finally.3). the stable separatrix of the saddle that lies in p > 0. in the case of the inverse square law.18 (a) Find the equilibrium points in the (u. and show that it is a centre. show that u = 1/r satisﬁes the equation d 2u γ + u = 2 uα−2 . use the coordinate axes and S up to its second intersection with the p-axis to construct a trapping region for S. and hence draw the phase portrait for the values of . By using Newton’s second law of motion in polar form. (c) Use MATLAB to solve (E9.

and we are stuck unless we can use some other property of the equation. ∂x ∂y dx In order to stress the equal role played by the independent variables in this equation. An inspection reveals that the substitution x = λˆ. v = y/x. y) = constant. y → λˆ. In order to introduce the general idea. x y The quantity y/x → y /ˆ is also invariant under the transformation. If we use this ˆ x invariant quantity.CHAPTER TEN Group Theoretical Methods In this chapter we will develop an approach to the solution of diﬀerential equations based on ﬁnding a group invariant. since dˆ y =f dˆ x y ˆ x ˆ . the equation becomes v+x dv = f (v). y = λˆ. as a new dependent variable. Another simple class of equations. takes the form ∂φ ∂φ dy + = 0. this is neither separable nor exact. where λ is any real constant. leaves the form of the equation x y unchanged. ∂x ∂y This has the solution φ(x. dx is dy = g(y) f (x)dx + constant. The solution of the separable equation. ﬁrst order ordinary diﬀerential equations. dx . we will usually write ∂φ ∂φ dx + dy = 0. often referred to as exact equations. dy = f (x)g(y). Let’s now consider the equation y dy =f . let’s begin by considering some simple. We say that the equation is invariant under the transformation x → λˆ. dx x In general.

). − ) and y = g(x1 . y → λˆ forms a group under composition of transformations. y1 = y + . so that f (x. Some simple one-parameter groups are: (a) Horizontal translation. to show that the set of transformations M( ) forms a group. 10. x1 = x and y1 = y. this is an example of a Lie group. More x y speciﬁcally. so that if x1 = f (x. ) is a well-deﬁned transformation from D into R2 . y. + δ). ).10. y1 . x1 = x.1 Lie Groups Let D be a subset of R2 on which x → x1 = f (x. or Lie group.e x = eδ+ x. x2 = f (x1 . + δ) and y2 = g(x. Finally. V( ): (c) Magniﬁcation. y = e− y1 . 0) = y. y. y. ). ﬁrstly note that when = 0. For example. a simple rearrangement gives x = e− x1 . y → y1 = g(x. g(x.1 LIE GROUPS 257 which is separable. if x1 = f (x. (ii) the transformation with − gives the inverse transformation. ) and y1 = g(x. M( ): (d) Rotation. and then show how this can be used as a practical tool for the solution of diﬀerential equations. . y1 = e y. y. y1 . (iii) the composition of two transformations is also a member of the set of transformations. if x2 = eδ x1 and y2 = eδ y1 then x2 = eδ . x If we regard the parameter λ as a continuous variable. the set of transformations x → λˆ. with solution dv = f (v) − v dx + constant. Secondly. then x = f (x1 . then x2 = f (x. We will discuss exactly what we mean by a group and a Lie group below. y1 = x sin + y cos . 0) = x. − ). so that. ). δ). y. y1 . δ) and y2 = g(x1 . R( ): x1 = x + . y. y. This set of transformations forms a one-parameter group. if (i) the transformation with = 0 is the identity transformation. y1 = y. x1 = e x. y1 . x1 = x cos − y sin . so that the inverse transformation is given by M(− ). y. we begin by developing as much of the theory of Lie groups as we will need. and similarly with y2 . In the following sections. named after the mathematician Sophus Lie. y1 = g(x. We also assume that x1 and y1 vary continuously with the parameter . y. y. It is the invariance of the diﬀerential equation under the action of this group that allows us to ﬁnd a solution in closed form. H( ): (b) Vertical translation.

y. =0 (10. we obtain the inﬁnitesimal magniﬁcation M( ). so that the original group is isomorphic to the translation group V( ). Consider what happens when 1. d d a result that is correct at leading order by virtue of the inﬁnitesimal nature of the transformation. which we deﬁned earlier. y1 ) = C2 + for some constants C1 and C2 . we can Taylor expand to obtain x1 = x + dx1 d + ··· . y1 = g(x. η(x. ξ η Integration of this gives solutions that are. We can also write this in the form dx1 dy1 = =d . This means that. we are giving the ﬁnite form of the group. y1 = e y forms the magniﬁcation group 1. ). y1 ∼ (1 + )y. Since = 0 gives the identity transformation. We will now show that every inﬁnitesimal transformation group is similar. expressible in the form F1 (x1 . we obtain x1 = x + ξ(x. y). y1 = x sin + y cos forms the rotation group R( ). y) = =0 dy1 d . y) as new variables. Using the approximations cos = 1 + · · · . We will show later that every one-parameter group is associated with a unique inﬁnitesimal group. y1 ). y1 ∼ y + x. y) = y. = η(x1 . or isomorphic. y1 ). y) = x. ). This is called the inﬁnitesimal form of the group. Since = 0 corresponds to the identity transformation. y1 = y + η(x. we obtain the inﬁnitesimal rotation group as x1 ∼ x − y. so that ξ(x. and hence ξ(x.1. y. When = 0 this gives the identity transformation. y1 ) = F2 (x. the transformation x1 = x cos − y sin . The transformation x1 = e x. we can deduce that F1 (x1 . y1 ) = C1 and F2 (x1 . to a translation group. =0 If we now introduce the functions ξ(x. For example. This means that if we deﬁne u = F1 (x. y) and F2 (x1 .1 The Inﬁnitesimal Transformation By deﬁning a Lie group via the transformation x1 = f (x. =0 y1 = y + dy1 d + ··· . y) = x and η(x.258 GROUP THEORETICAL METHODS 10. y) = −y and η(x. then the group can be represented by u1 = u and v1 = v + . sin = + · · · for 1. y) + . y) and v = F2 (x. y) = dx1 d . in principle. Using e = 1 + + · · · for group as x1 ∼ (1 + )x. . Consider the equations that deﬁne ξ and η and write them in the form dx1 dy1 = ξ(x1 . y1 ) = F1 (x. and which we shall soon see is exact. y). we can make any inﬁnitesimal transformation group look like H( ) or V( ). by using a change of variables.1) and just retain the ﬁrst two terms in the Taylor series expansions.

0) + Since φ(x1 . φ (x1 . y) under an inﬁnitesimal transformation. ) = φ(x1 . 0) = φ(x. ). =0 . Any inﬁnitesimal transformation is completely speciﬁed by U φ. ∂x ∂y ∂φ ∂φ +η . y1 = y + x. We ﬁnd that ∂φ ∂φ δφ = φ(x1 . y1 ) = U x1 Let’s now consider a group deﬁned in ﬁnite form by x1 = f (x. y1 . ∂x ∂y φ (x1 . so that Uφ = Ux ∂φ ∂φ + Uy . y) = −y and η(x. y) = +η ξ + ··· . ) as a function of . y1 . y + η) − φ(x. 0) + φ (x1 . y). ∂x ∂y (10.2). y1 ) then U φ(x1 . δφ. ∂x ∂y If we retain just this single term. y). we can see that δφ can be written in terms of the quantity Uφ = ξ or. y1 . y1 . 0) = dφ d = U 2 φ. y) = φ(x + ξ. For example. ∂x1 ∂y1 and if a group acts on (x. ∂x ∂y ∂φ ∂φ + U y1 . y) = x. 0) + · · · .10. 2 = =0 ∂φ dy1 ∂φ dx1 + ∂x1 d ∂y1 d =ξ =0 =0 = and ξ ∂φ ∂y1 d d ∂φ ∂φ +η = U φ. If we regard φ(x1 . 0) = ∂φ ∂x1 dφ d +η =0 1 2 φ (x1 .1 LIE GROUPS 259 10. y) to produce new values (x1 . ∂x ∂y ξ(x.2) This is called the inﬁnitesimal generator of the group. which is consistent with the way we derived the inﬁnitesimal transformation. we ﬁnd that φ(x1 . y1 = g(x. with a prime denoting d/d . y1 . y. ) and a function φ = φ(x. y. U x = ξ and U y = η.1. y1 . From the deﬁnition (10. y1 ) − φ(x. y1 . so that the transformation is given by x1 = x − y. in operator notation. that occurs in a given smooth function φ(x. if U φ = −y ∂φ ∂φ +x . y1 . U ≡ξ ∂ ∂ +η .2 Inﬁnitesimal Generators and the Lie Series Consider the change.

For example. U 4 y = y. y1 = y + η + 1 2 2 ξ + ··· . U 3 x = y. ξ ∂ξ ∂ξ +η ∂x ∂y ∂η ∂η +η ∂x ∂y + ··· . ) = e In particular. ∂x ∂y U is therefore a cyclic operation with period 4 and the equations of the ﬁnite form of the group are x1 = x − y − 1 2 1 3 1 4 x+ y+ x + ··· 2! 3! 4! . y). ∂x ∂y 1 2 1 3 y+ y + · · · = ye . ∂φ ∂φ +x . if an inﬁnitesimal group is represented by Uφ = x then x1 = x + x + y1 = y + y + 1 2 1 3 x+ x + · · · = xe . and can be written more compactly in operator form as φ(x1 . if we take φ(x. U 4 x = x. if an inﬁnitesimal group is represented by U φ = −y then U x = −y. 2! 3! The ﬁnite form is therefore M( ). ) = φ(x. y.260 GROUP THEORETICAL METHODS we have 1 2 2 U φ + ··· . As a further example. the magniﬁcation group. U 3 y = −x. 0) = x. y1 . 2! 3! ∂φ ∂φ +y . 0) + U φ + φ(x1 . y. These two relations are a representation of the group in ﬁnite form. y1 . It should now be clear that we can calculate the ﬁnite form of the group from the inﬁnitesimal group (via the Lie series) and the inﬁnitesimal group from the ﬁnite form of the group (via expansions for small ). U 2 x = −x. =x+ ξ+ Similarly. U 2 y = −y. 2 This is known as a Lie series. x1 = x + U x + 1 2 2 1 U x + · · · = x + ξ + 2U ξ + · · · 2 2 1 2 2 U φ(x. U y = x.

Using the Lie series. and we have the rotation group R( ). y1 ) = Ω(x. y0 ) we then obtain a family of curves. y0 ) and apply the inﬁnitesimal transformation to it. y1 ). Ω(x1 . this gives dx1 /d = x1 and dy1 /d = y1 . y1 = y + x − =y 1− 1 2! 2 1 3 1 3 1 4 y− x+ y + ··· 2! 3! 4! +x − 1 3! 3 + 1 4! 4 − ··· + ··· = y cos + x sin . with x = x1 and y = y1 at = 0. ∂x ∂y A function Ω(x.10. subject to x1 = x and y1 = y at = 0 d d is an exact relationship according to (10.3).2 Invariants Under Group Action Let x1 = f (x. y0 + η). when x1 and y1 are derived from x and y by the operations of the group. all of which are solutions of (10. using the fact that dx1 dy1 = ξ(x1 . 2! 2! so that a necessary and suﬃcient condition for invariance is U Ω = 0. y) is said to be invariant under the action of this group if. This ﬁrst order system can readily be integrated to give x1 = xe and y1 = ye . ) be the ﬁnite form of a group and let the inﬁnitesimal transformation associated with the group have inﬁnitesimal generator Uφ = ξ ∂φ ∂φ +η . and hence that ∂Ω ∂Ω ξ +η = 0. y1 ) = Ω(x. ). whose solution is Ω(x.3) ξ η Since this equation has only one solution. y). on the curve dy dx = . it follows that a one-parameter group has only one invariant.1). y. This family of . y. = η(x1 .3). (10. If we repeat this procedure inﬁnitely often. which we denote by Ωf (x. Now let’s take a point (x0 . we can obtain a curve that is an integral of the diﬀerential system given by (10. y1 = g(x. ∂x ∂y Ω(x1 . a constant. y) = C. By varying the initial point (x0 . so that it is mapped to (x0 + ξ. y) = C. y1 ). For the ﬁrst example above.2 INVARIANTS UNDER GROUP ACTION 261 =x 1− Similarly. y) + U Ω + U (U Ω) + · · · . y) + U Ω + This is a partial diﬀerential equation for Ω. we can write 1 2 2 1 2 U Ω + · · · = Ω(x. There is a rather more concise way of doing this. 10. 1 2! 2 + 1 4! 4 − ··· −y − 1 3! 3 + ··· = x cos − y sin .

dx1 fx + pfy It can easily be veriﬁed that the triple given by (x. If we write x1 = x + ξ(x. Ωf (x1 . y). For example. This gives the intuitively obvious result that circles are invariant under rotation! 10. 1 + ξx + p ξy Expanding this using the binomial theorem for small . y) becomes (x1 . This is conveniently written as U Ωf = F (Ωf ) for some arbitrary nonzero function F. ) form a group of transformations in the usual way. y1 ) = Ωf (x. p . y1 = y + η(x.4) The inﬁnitesimal generator associated with this three-element group is U φ = ξφx + ηφy + ζφp . y1 ) under the action of the group. . y. y. y). This extended group also has an inﬁnitesimal form associated with it. y.262 GROUP THEORETICAL METHODS curves is invariant under the action of the group in the sense that each curve in the family is transformed into another curve of the same family under the action of the group. then p1 = ηx + p(1 + ηy ) .3 The Extended Group If x1 = f (x. Under the transformation this becomes p1 = dy1 gx + pgy = = h(x. y) + U Ωf + 2! The most general condition that forces the ﬁrst two terms to be constant is that U Ωf = constant should represent one family of curves. y. p) forms a group under the transformations above. the rotation group is represented in inﬁnitesimal form by U φ = −y ∂φ ∂φ +x . This is known as the extended group of the given group. This means that Ωf (x1 . because U n Ωf = U n−1 (U Ωf ) = U n−1 (constant) = 0. (10. Using the Lie series we can write 1 2 2 U Ωf + · · · . Suppose (x. ). y x which can be easily integrated to give Ω = x2 + y 2 = C. ) and y1 = g(x. y1 ) = constant must represent the same family of curves. we ﬁnd that p1 = p + ηx + (ηy − ξx ) p − ξy p2 = p + ζ. we can extend the group by regarding the diﬀerential coeﬃcient p = dy/dx as a third independent variable. ∂x ∂y The equation for the invariants of this group is − dy dx = .

4 INTEGRATION OF A FIRST ORDER EQUATION 263 It is of course possible. This is called Lie’s fundamental problem and is considerably more diﬃcult than (i). ∂x ∂y ∂φ ∂φ ∂φ +η +ζ ∂x ∂y ∂p We have seen that a suﬃcient condition for the invariance property is that U φ = 0. In this case we can give a recipe for using this invariance to solve the equation. let’s consider a general ﬁrst order ordinary diﬀerential equation. where φ is constant. To do this. (ii) No obvious group of transformations can be spotted and we need a more systematic approach to construct the group. are the two independent solutions of the simultaneous system dx dy dp = = .10. y. F (x. We are now at the stage where we can use the theory that we have developed in this chapter to solve ﬁrst order diﬀerential equations. though algebraically more messy. ξ η ζ Let u(x. it is helpful to outline two generic situations that may occur when you are confronted with a diﬀerential equation that needs to be solved.4 Integration of a First Order Equation with a Known Group Invariant To show more explicitly that this group invariance property will lead to a more tractable diﬀerential equation than the original. ξ η and v(p. y) = α be a solution of dx dy = . Prior to this. p) = 0. 10. Indeed. to form further extensions of a given group by including second and higher derivatives. ﬁnding v is simply a matter of integration. We now show that if we know U . ∂x ∂y ∂p The solution curves of this partial diﬀerential equation. y) = β be the other independent solution. so we are faced with solving ξ ∂φ ∂φ ∂φ +η +ζ = 0. that is invariant under the extended group U φ=ξ derived from Uφ = ξ ∂φ ∂φ +η . no general solution is known to Lie’s fundamental problem for ﬁrst order equations. recall the earlier result that any group with one parameter is similar to the . (i) It is straightforward to spot the group of transformations under which the equation is invariant. x.

U1 f . Then U1 f = U x1 ∂f ∂f ∂f + U y1 = .264 GROUP THEORETICAL METHODS translation group. 0 1 0 This particularly simple system has solutions x1 = constant and p1 = constant. In principle. given by 1 dy1 = . one solution of this system is u(x. as they are separable. The solution of the original equation can then be obtained by returning to the (x. y) = 2x and η(x. If we solve the system dy dx = . so that ξ(x. Solving dy1 1 1 = = dy η y . y) so that by a simple integration we can obtain y1 as a function of x and y. so that the diﬀerential equation can be put in the form p1 = F (x1 ) for some calculable function F . ξ η we ﬁnd that y/x1/2 = eC . y1 ) reduce U φ to the group of translations parallel to the y1 axis. Let the change of variables from (x. the most general diﬀerential equation invariant under U1 in the new x1 . Example The diﬀerential equation dy = dx 1 x + y2 is invariant under the transformation x = e−2 x1 . y = e− y1 . y) and the second is equivalent to the simultaneous system dx dy dy1 = = . y) variables. The inﬁnitesimal transformation associated with this is x1 = x+2 x. and call the inﬁnitesimal generator of this group U1 f . y) to (x1 . so that x1 = y/x1/2 . it is straightforward to solve equations of this form. ∂x ∂y ∂x ∂y The ﬁrst of these equations has the solution x1 = u(x. dy η(x. y1 = y + y. y1 variables will therefore be a solution of the simultaneous system dy1 dp1 dx1 = = . ξ η 1 Again. is identical to U1 f . ξ +η = 1. As the extended group of translations. y) = y. y) = α. ∂x1 ∂y1 ∂y1 from which we see that U x1 = 0 and U y1 = 1. ξ ∂x1 ∂x1 ∂y1 ∂y1 +η = 0. or more explicitly. This can be used to eliminate x from the second independent solution.

However. Using the diﬀerential equation to eliminate dy/dx.4) shows that dy dy1 + = dx1 dx ηx + (ηy − ξx ) dy − ξy dx dy dx 2 + ··· . 10. on rearranging.6) . y). given the function F . ∂x ∂y (10. x1 x2 + 1 1 . gives 1 dy1 = 1 dx1 x1 1 − 2 x1 x2 + 1 1 . the fundamental problem is to determine two functions ξ and η that satisfy this ﬁrst order partial diﬀerential equation. Of course. 2 + y 2 )2 ∂x ∂y (x + x (x + x2 + y 2 ) (10. This is the separable equation that we are promised by the theory we have developed. we ﬁnd that the equation will be invariant under the action of the group provided that ξ ∂F ∂F +η = ηx + (ηy − ξx ) F − ξy F 2 .10. y). The ﬁnal integration of the equation can be achieved by the successive substitutions z = log x1 and t2 = 1 + e−2z . Some simple calculus then shows that the original diﬀerential equation transforms to 1 3 dy1 x 2 1 dx1 = dy1 x1 −1 dx1 which. (10. y) dx and an inﬁnitesimal transformation of the form x1 ∼ x + ξ(x. Example Consider the equation dy y . − (1 + 2x) y x + x2 − y 2 ∂F ∂F = = . as the following example shows.5) So. = dx x + x2 + y 2 In this case. this is an underdetermined problem and has no unique solution.5 Towards the Systematic Determination of Groups Under Which a First Order Equation is Invariant If we consider a diﬀerential equation in the form dy = F (x. 2. y1 ∼ y + η(x. by choosing a special form for either ξ or η there are occasions when the process works.5 DETERMINING GROUPS UNDER WHICH AN EQUATION IS INVARIANT 265 gives y1 = log y.

linear. dx2 Writing x1 = x + ξ and y1 = y + η we have already shown that dy1 dy + = dx1 dx = Now d d 2 y1 = dx2 dx1 1 and d dx dy1 dx1 =1 = dp + (Πx + Πy p + Πp px ) + · · · dx dy1 dx1 = d dx dy1 dx1 dx dx1 ηx + (ηy − ξx ) dy − ξy dx dy dx 2 (10. We therefore take x1 = y/x and hence y1 = y as our new variables. In terms of these variables. we can ﬁnd the solution ξ = x/y.6 Invariants for Second Order Diﬀerential Equations First order diﬀerential equations can be invariant under an inﬁnite number of oneparameter groups. y. The inﬁnitesimal transformation in this case is x1 = x + x/y.266 GROUP THEORETICAL METHODS and (10. we now solve dx dy = x/y 1 to obtain y/x = constant. dx1 1 = + ··· . 10. let’s consider the simplest form of a variable coeﬃcient. (10.5) takes the form (x + x2 − y 2 )η − (1 + 2x)yξ = (x + x2 + y 2 )2 ηx + (ηx − ξx )y(x + x2 + y 2 ) − y 2 ξy . =− dx1 (1 + x2 ) 1 with solution y1 = − tan−1 x1 +C. following the procedure outlined in the last section.6) becomes dy1 1 . but after some trial and error. second order. this reduces to (x + x2 − y 2 ) + (1 + 2x)yξ = y(x + x2 + y 2 )ξx + y 2 ξy . It is not easy to solve even this equation in general. Second order diﬀerential equations can only be invariant under at most eight groups. dx 1 + (ξx + ξy p) . diﬀerential equation. y. y1 = y + and. The solution of our original diﬀerential equation can therefore be written in the form y + tan−1 (y/x) = C. d 2y + q(x)y = 0. p) . dx dx = p + Π (x.7) dy dy + Π x. To see where this ﬁgure comes from. If now we choose η = 1.

ρ (x) + q(x)ρ(x) = 0.11) by noting that the coeﬃcient of each power of y must be zero. (10. ˜ (10. Substitution of this into (10.8)3 . we obtain ξyy = 0. Equation (10. η At this stage notice that ρ(x) and ζ(x) satisfy the original second order equation. which can be integrated to give ˜ η = ρ (x)y 2 + η (x)y + ζ(x). ˜ η ˜ ξ (x) = 2˜ (x).10. There are also second order equations for ξ(x) and η (x). which gives us four independent equations. 2ηxy − ξxx + 3q(x)yξy = 0.8)1 can be integrated to give ξ = ρ(x)y + ξ(x). ηyy − 2ξxy = 0. the solution of which will involve four constants.10) and (10.6 INVARIANTS FOR SECOND ORDER DIFFERENTIAL EQUATIONS 267 so that d 2 y1 d 2y 2 = dx2 + dx1 Πx + Πy dy d 2y d 2y + Πp 2 − 2 dx dx dx dy dx ξx + ξy dy dx + ··· .9) gives ρ (x)y 2 + η (x)y + ζ (x) − q(x)y {˜(x) − 2ξ (x)} + {ρ(x)y + ξ(x)} q (x)y ˜ η + ρ (x)y 2 + η (x)y + ζ(x) q(x) = 0. (10. 3ρ(x)q(x)y + 3ρ (x)y + 2˜ (x) − ξ (x) = 0.8)2 gives ηyy = 2ρ (x).9) We can ﬁnd a solution of (10. ηxx + (2ξx − ηy )q(x)y + ξq (x)y + q(x)η = 0.10) (10.11) (10. so the original equation is invariant under at most . From (10. ζ (x) + q(x)ζ(x) = 0. Some simple calculation leads to {ηxx + (2ξx − ηy ) q(x)y + ξq (x)y + q(x)η} + {2ηxy − ξxx + 3q(x)yξy } + {ηyy − 2ξxy } dy dx 2 dy dx − ξyy dy dx 3 = 0. The condition for invariance is therefore Πx + Πy d 2y dy d 2y + 2 Πp − 2 dx dx dx ξx + ξy + ξq (x1 )y1 + ηq(x1 ) = 0. the solution of which gives rise to a further four ˜ constants. Each of these constants will generate a one-parameter group that leaves the original equation invariant. η (x) − q(x) (˜(x) − 2ξ (x)) + ξ(x)q (x) + η (x)q(x) = 0. If we now set the coeﬃcients of powers of dy/dx to zero.7). η Finally.8) (10.

Let’s consider the group generated by ρ(x) in more detail. This means ˜ 2 that. dx 2ρ (x) Now dx1 dx1 dx = = 2ρ2 d(x1 y1 ) dx d(x1 y1 ) which gives d dx dx1 d(x1 y1 ) =2 ρ dy dy d 2y −ρ y−ρ +ρ 2 dx dx dx dx1 = C1 . dy1 1 1 1 = = = . 1 dy ρy − 2 ρ dx ρ =2 ρ dy −ρy . dy η(x. For example. dx 2Cρ2 which gives y1 = 1 2C x x0 dx ρ(x) = 2 ρ 2y x x0 dt ρ2 (t) .7) is invariant under magniﬁcation in y. d(x1 y1 ) = 2ρ d 2y + q(x)y dx2 = 0. we have to integrate dy dx = . 2ρ2 (t) x0 so that d 1 (x1 y1 ) = 2 . following the ideas of the previous section. We can write the diﬀerential equation in terms of these new variables by noting that x dt x1 y1 = . y) ρ (x)y 2 + Cρρ y ρ (x) (y 2 + Cρy) and. since y = Cρ(x). ξ(x. y) η(x.268 GROUP THEORETICAL METHODS eight one-parameter groups. In the usual way. we should deﬁne x1 = y/ρ(x). y) which leads to the equation y ρ (x)y + 1 ξ (x) + ζ(x) dy 2 = . Note that η (x) = 1 ξ (x) by direct integration. 1 dy1 = . Some of these groups are obvious. dx ρ(x)y + ξ(x) This has a solution of the form y = Cρ(x) provided that ξ(x) = Cρ2 (x) and ζ(x) = 0 for some constant C. Now. dx Integrating this expression gives . since (10. it must be invariant under x1 = x. y1 = e y.

which in general is diﬃcult to solve. it is not particularly useful to us in ﬁnding solutions. 0 y and we can see that this suggests using the new variables x1 = x and y1 = log y. This means that dy dx = . Although the invariance property produces this via a route that is rather diﬀerent from that taken in Chapter 1. consider the equation y + In this case. The point can perhaps be reinforced by considering the nonautonomous equation Y1 + Y12 + 1 y + ey = 0. For this example. so the group invariance property has allowed us to reduce the order of the original equation. deﬁned by y = −2 log x + Y .10. which has ξ = 0 and η = y. 4x2 . The exception to this is if we can spot a solution. The counter argument to this is that the group theoretical method gives both the technique of reduction of order and an algorithm for reducing the original equation to a ﬁrst order diﬀerential equation. Let’s now consider the invariance x1 = x and y1 = e y. you could of course argue that you could have guessed the solution u1 (x) = x1/2 and then reduced the order of the equation to obtain the second solution u2 (x) = x1/2 log x. we can see that Y1 = 1/2x is a solution. we ﬁrst introduce a new dependent variable. x Using the methods derived in this chapter. As an example of this. ρ2 (t) x0 x This is just the reduction of order formula. Putting Y1 = y1 leads to Y1 + Y12 + q(x) = 0. This is a ﬁrst order equation. Since y = ey1 (y1 +y12 ) under this change of variables.6 INVARIANTS FOR SECOND ORDER DIFFERENTIAL EQUATIONS 269 so that x1 = C1 x1 y1 + C2 or. when the equation will linearize. y = C1 ρ(x) dt + C2 ρ(x). which we derived in Chapter 1. we obtain y1 +y12 +q(x) = 0. in terms of the original variables. This gives us y + 1 eY Y + 2 = 0. dz 2 1 y = 0. Writing Y1 = 1/2x + V −1 linearizes the equation and it is straightforward to show that the general solution is y = C1 x1/2 + C2 x1/2 log x. 4x2 This is a form of Ricatti’s equation. Y . x x The invariance of this under x-magniﬁcation suggests introducing z = log x and leads to Y + d 2Y + eY = 0. we obtain 1 = 0. At this stage.

0) = e− δ(x1 ). 2 T (x. so that x/t1/2 is an invariant of the transformation (can you spot which other groups it is invariant under?). 0) = H(−x). it is readily established from the ordinary diﬀerential equation that the solution is 1 T (x.7 Partial Diﬀerential Equations The ideas given in this chapter can be considerably extended. since δ(ax) = δ(x)/a. 0) = δ(x).12) For the initial condition T (x. T = eµ T1 . particularly to the area of partial diﬀerential equations. we can use the fact that (10. (10. 10. (10. t) = √ π 2 e−s √ x/ 4Dt ∞ ds = 2erfc √ x 4Dt . The analysis of this equation is of course much simpler than the original one.270 GROUP THEORETICAL METHODS which we can immediately integrate to 1 2 dY dz 2 = C − eY . t) = t−1/2 F (x/t1/2 ). namely ∂2T ∂T =D 2. t = e2 t1 .13) which is shown in Figure 10. which is also invariant under the transformation.14) . we ﬁnd that 1 1 DFηη + ηFη + F = 0. t = e2 t1 . this reduces the partial diﬀerential equation to the ordinary diﬀerential equation D 1 dT d 2T = 0. consider the equation for the diﬀusion of heat that we derived in Chapter 2. if the initial condition is T (x. If we therefore look for a solution of the form T (x. + η dη 2 2 dη (10. ∂t ∂x This equation is invariant under the group of transformations x = e x1 . 2 2 A suitable solution of this is Ae−η and hence 2 /4D (for example. and now t1/2 T is invariant as well. using the method of Frobenius). x/t1/2 is an invariant.12) is also invariant under the two-parameter group x = e x1 . The initial condition transforms to eµ T1 (x1 . As a simple example. t) = At−1/2 e−x /4Dt . Finally.1. As before. If we write η = x/t1/2 (here η is known as a similarity variable). so that the choice µ = − makes both diﬀerential equation and initial condition invariant.

or the transformation of a partial diﬀerential equation into an ordinary diﬀerential equation. If you ﬁnd one. this will lead to a reduction of order for an ordinary diﬀerential equation. (10.7 PARTIAL DIFFERENTIAL EQUATIONS 271 Fig. d dt Since ∞ ∞ ∞ T dx = 0.10. and hence T (x. Notice the similarity between this solution and the sequence (5. 0) dx = −∞ −∞ δ(x) dx = 1. we can integrate (10. then you need to make a systematic search for invariants. Given an ordinary or partial diﬀerential equation. We end this chapter with a recommendation.15) This is known as the point source solution of the diﬀusion equation. t) = √ 2 1 e−x /4Dt . magniﬁcation or rotation. .13). from (10. Then. Further details of extensions to the basic method can be found in Bluman and Cole (1974) and Hydon (2000). try to ﬁnd a simple invariant.4. If you cannot ﬁnd an invariant by inspection. The solution. of the diﬀusion equation.1. t) dx = 1. such as translation. which we can use to deﬁne the Dirac delta function. −∞ T (x.14).27). 4πDt (10. To ﬁnd the constant A. A = 1/ 4πD. 10.12) to obtain the integral conservation law. we must have ∞ −∞ √ T (x. as shown in Figure 5.

272 GROUP THEORETICAL METHODS This can involve a lot of calculation. (b) U φ = x ∂x ∂y ∂φ ∂φ + . show that it has a solution in integrating factor form. invariant under the groups ∂φ ∂φ +x .3 10.2 and hence integrate each equation. (c) dx is invariant under a group of transformations with inﬁnitesimal generator (a) x U φ = ax ∂φ ∂φ + by . Find the general diﬀerential equation of ﬁrst order. For the case p(x) = xm . P dy − Q dx = constant.4 Show that the diﬀerential equation d 2y + p(x)y 2 = 0 dx2 is invariant under at most six one-parameter groups. . dx 2xy 3 − x4 10. ∂x ∂y 2 10. reduce the order of the equation.1 Show that each of the diﬀerential equations dy dy . dx 2 dy = y + x2 . ∂x ∂y ∂φ ∂φ + ay . and is best done by a computer algebra package. + y = x4 dx dx dy (b) x − yxm = 0. y) Q(x. Hydon (2000) gives a list and critique of some widely available packages. y) is invariant under a group with inﬁnitesimal generators ξ and η. Exercises 10. P η − Qξ Hence ﬁnd the solution of dy y 4 − 2x3 y = . (c) U φ = y ∂x ∂y If the diﬀerential equation (a) U φ = dy dx = P (x.

Can you ﬁnd any solutions of the resulting diﬀerential equation? The equation ut = uxx − up .6 10. and the corresponding ordinary diﬀerential equations. Find some invariants of this equation. Can the point source problem be resolved for this sink-like equation? . ∂y 10. ut = (uux )x . with p ∈ R+ . arises in mathematical models of tumour growth. Find a similarity reduction of the porous medium equation.7 10.8 Derive the integrating factor for y + P (x)y = Q(x) by group theoretical methods.5 Find the most general ﬁrst order diﬀerential equation invariant under the group with inﬁnitesimal generator U φ = exp p(x)dx ∂φ .EXERCISES 273 10. and other areas.

Thin aerofoil theory . or when one geometrical length in the problem is much shorter than another. Acheson. — In aerodynamics. These crosssections are usually long and thin with aspect ratio = l/L 1. Moreover. dimensionless parameters usually arise when one physical process occurs much more slowly than another. where an apparently negligible physical eﬀect. Here. Some other examples are: — Waves on the surface of a body of ﬂuid or an elastic solid. it is crucial to be able to calculate the lift force on the crosssection of a wing due to the ﬂow of an inviscid ﬂuid around it. dimensionless parameters arise naturally when we use dimensionless variables. becomes important in a small region. for example. are said to be of small amplitude if = a/λ 1. numerical solution. to solve numerically. for example. a dimensionless parameter that measures the ratio of acceleration to viscous forces. if not impossible. Billingham and King. with amplitude a and wavelength λ. Examples occur in many diﬀerent areas of applied mathematics. ρ is the ﬂuid density. Small. This is an example of a regular perturbation problem. and L a typical horizontal length. pressure changes due to accelerations are much greater than those due to viscous stresses. where viscosity becomes important (see. and we will meet several in Chapter 12. Re is the Reynolds number. Often. where l is a typical vertical width. here viscosity. The solution for Re−1 except for a small neighbourhood of the plate. where the problem is simpliﬁed throughout the domain of solution. — In the high speed ﬂow of a viscous ﬂuid past a ﬂat plate of length L.CHAPTER ELEVEN Asymptotic Methods: Basic Ideas The vast majority of diﬀerential equations that arise as models for real physical systems cannot be solved directly by analytical methods. This is an example of a singular perturbation problem. U the ﬂuid velocity away from the plate and µ the ﬂuid vis1 has the ﬂuid velocity equal to U everywhere cosity. 1990). However. the presence of a small parameter often leads to a singular perturbation problem. 2001). it may be possible to use an asymptotic method to obtain an approximate solution. if one or more small. As we shall see. known as a boundary layer. which we discussed at the beginning of Chapter 5. A simpliﬁcation of the governing equations based on the fact that 1 leads to a system of linear partial diﬀerential equations (see. which can be diﬃcult. as expressed by Re = ρU L/µ 1. the only way to proceed is to use a computer to calculate an approximate. dimensionless parameters appear in the diﬀerential equation.

. These series are asymptotic expansions. by deﬁnition. g( ) → 0. we will look at series solutions of diﬀerential equations. as → 0. since it is used to gauge the size of f ( ). For example. and say that f is much less than g. 2 e −1∼ . 1960. This notation tells us nothing about the constant A. for 1. cos = o( −1 ). The crucial diﬀerence between asymptotic expansions and the power series that we met earlier in the book is that asymptotic expansions need not be convergent in the usual sense. e− = O(1). If lim →0 f( ) = 1. 11. when 1.1 Asymptotic Expansions In this chapter. For example. cos = O(1). all of which can be found from the Taylor series expansions of these functions. The order notation only tells us how functions behave as → 0. (ii) We also have a notation available that displays more information about the behaviour of the functions. We can illustrate this using an example. whilst in Chapter 12. We say that f is of order g for small . Milne-Thompson. g( ) lim →0 for some nonzero constant A. Here g( ) is a gauge function. 1964). cos − 1 = O( 2 ). which are all. and say that f ( ) is asymptotic to g( ) as example.1 ASYMPTOTIC EXPANSIONS 275 exploits the size of to greatly simplify the calculation of the lift force (see. for example. For we write f ( ) ∼ g( ).1. sin = O( ). (iii) If lim →0 1 − cos ∼ 1 2 . we write f ( ) = O (g ( )) for 1. 11. For example sin = o(1). f( ) = 0. there is some useful notation for comparing the sizes of functions that we will introduce here and use extensively later. we will study how to approximate various integrals as series expansions. 1010 = O(1).1 (i) If Gauge Functions f( ) = A. sin ∼ . and Van Dyke. but ﬁrstly. It is not meant to be used for comparing constants.11. g( ) e− = o(log ) we write f ( ) = o (g ( )). of O(1).

∞ x e−t dt tN +1 N! xN +1 ∞ x e−t dt = N !e−x . (11.1) for x > 0. as we shall see.2) provides an extremely eﬃcient method of calculating Ei(x). since the integrals will not be convergent.2 Example: Series Expansions of the Exponential Integral. |RN | = N ! and hence |RN | = O e−x x−(N +1) for x 1. In order to develop the power series representation of Ei(x) about x = 0. Ei(x) Let’s consider the exponential integral ∞ Ei(x) = x e−t dt.1. We therefore have to be a little more cunning in order to obtain our power series. we would like to use the series expansion of e−t .3) 1. xm (11. N +1 N −1 (N − 1)! x (−1) x which is unbounded as N → ∞. The ratio of terms is a ﬁxed value of N by our ﬁrst example of an asymptotic expansions. Moreover. it converges to Ei(x) as x → ∞. tN +1 (11. However. xN +1 Therefore.276 ASYMPTOTIC METHODS: BASIC IDEAS 11. the integral for Ei(x) does not converge as x → 0. note that ∞ Ei(x) = x 1 t e−t − 1 1+t ∞ dt + x 1 dt t(1 + t) . Firstly. for x even moderately large. Using the fact This result is exact. We can integrate by parts to obtain Ei(x) = − e−t t ∞ − x ∞ x e−x e−t − dt = t2 x e−t dt. However. the (N + 1)th and N th N xN (−1)N N ! =− .2) does not converge in the usual sense. by the ratio test.2) where RN = (−1)N N ! ∞ x e−t dt. let’s consider how big RN is for x that x−(N +1) > t−(N +1) for t > x.2) at neglecting RN . term by term integration is not possible as things stand. t2 Doing this repeatedly leads to N Ei(x) = e−x m=1 (−1)m−1 (m − 1)! + RN . This is asymptotic expansion. if we truncate the series expansion (11. In common with most useful (11. so the series diverges. t ∞ x (11. Now.

These diﬃculties do not arise for the asymptotic expansion. whilst for Ei(40) this rises to approximately 3. however. the result is the diﬀerence of many large terms. Note that the convergent series (11. A further rearrangement then gives ∞ Ei(x) = 0 1 t e−t − x 1 1+t x dt − 0 1 −t e − 1 dt t − 0 ∞ 1 dt − log 1+t x 1+x x = 0 1 t e−t − 1 1+t dt − log x − 0 1 −t e − 1 dt.4) to calculate the function for any given x. We conclude that asymptotic expansions can also be convergent series.2).5772 . but can be shown to be equal to minus Euler’s constant.1 shows the relative error using this expansion compared with that using the ﬁrst 25 terms of the convergent series representation. which we could evaluate numerically. we can interchange the order of summation and integration. (11. We can now use the power series representation of e−t to give x Ei(x) = −γ − log x − 0 1 (−t)n dt. For example.4) also provides an asymptotic expansion valid for x 1. . practical problems associated with this. . The philosophy that we have used earlier in the book is that we can now use (11. we can take the limit x → 0 in the ﬁrst integral. even if we take enough terms in the series to get an accurate answer. . although we will not prove this here. Consider the two-term expansion Ei(x) ∼ e−x x 1− 1 x . The advantages of the asymptotic series over the convergent series are immediately apparent.3 × 106 . the largest term in the series for Ei(20) is approximately 2.11. n · n! n=1 ∞ (11.8 × 1014 .4). t n=1 n! ∞ Since this power series converges uniformly for all t. if we take x = 5 we need 20 terms of the series in order to get three-ﬁgure accuracy. With this rearrangement. Figure 11. with the second integral giving Ei(x) ∼ − log x as x → 0. There are. the series converges very slowly. Secondly. γ = 0. Unless the computer used to sum the series stores many signiﬁcant ﬁgures. there will be a large roundoﬀ error involved in calculating the Ei(x) as the small diﬀerence of many large terms. t The ﬁrst term is just a constant. Firstly.1 ASYMPTOTIC EXPANSIONS ∞ 277 = x 1 t e−t − 1 1+t dt − log x 1+x . For example. (11. and ﬁnally arrive at Ei(x) = −γ − log x − (−x)n .4) This representation of Ei(x) is convergent for all x > 0. .

solid line. dashed line. δn = (− log )−n . 2. let N SN (x) = n=0 fn (x). δn = n . If we can write N f( ) ∼ n=0 an δn ( ) as → 0. δn = n/2 . Such a sequence is called an asymptotic sequence of gauge functions. — If SN is a convergent series representation of S(x).1. — If SN is an asymptotic series representation of S(x).3 Asymptotic Sequences of Gauge Functions Let δn ( ). For example.2).4). for some sequence of constants an . SN (x) ∼ S(x) as x → ∞ (or whatever limit is appropriate) for any ﬁxed N 0.1. For example. .278 ASYMPTOTIC METHODS: BASIC IDEAS Fig. δn = (cot )−n . Ei(x) has sion relative to the asymptotic sequence δn ( ) for . . we say that f ( ) has an asymptotic expan1. 1. SN → S(x) as N → ∞ for ﬁxed x within the radius of convergence. . 11. and the two-term asymptotic expansion (11. 11. The relative error in calculating Ei(x) using the ﬁrst 25 terms of the convergent series expansion (11. n = 0. be a sequence of functions such that δn = o(δn−1 ) for 1. To summarize.

N f (x. We say that the expansion becomes nonuniform as x → 0. so the expansion is uniformly valid. 1 3 1 2 sin x − cos x + O( 4 ) 2! 3! for all x. 2 3 4 Some sequences of gauge functions are clearly easier to use than others. if δn ( ) = (1 − e )n . For example. Now consider sin(x + ) = sin x + cos x − √ √ 1/2 as → 0. sin = n=1 (−1)n+1 2n−1 (2n − 1)! . 1 1 5 sin = δ0 + δ1 + δ2 + δ3 + · · · . e + e−1/ ∼ n=0 n n! as → 0+ . f (x. Note that each successive term is smaller by a factor of /x. Now let’s consider functions of a single variable x and a parameter .1 ASYMPTOTIC EXPANSIONS 279 an asymptotic expansion. . For example.2). ). ) ∼ n=0 fn (x)δn ( ) as → 0. If f has an asymptotic expansion. . For example. Since e−1/ is exponentially small. However. a given function can have diﬀerent asymptotic expansions with respect to diﬀerent sequences of gauge functions. for any N > 0. when x = O( ). N e ∼ n=0 n n! N as → 0+ . ∞ if δn ( ) = n . . we say that the expansion is uniformly valid in R. . If this is not the case. 2. We can think of this as a typical solution of an ordinary diﬀerential equation with independent variable x and a small parameter. To determine an asymptotic . provided x . we say that the expansion becomes nonuniform in some subdomain. and therefore the expansion fails to be asymptotic when x = O( ). The asymptotic expansion of a function relative to a given sequence of x−1 gauge functions is unique. that is valid for all x in some domain R. relative to the asymptotic sequence x−n e−x for 1. . since each an can be calculated in turn from a0 = lim f( ) . for any N > 0.11. x+ = x 1+ x ∼ √ x 1+ 2x − 2 8x2 + O( 3 ) as → 0. it will only appear in the asymptotic expansion after all of the algebraic terms. We also note that diﬀerent functions may have the same asymptotic expansion when we consider only a ﬁnite number of terms. →0 δ0 ( ) →0 m=0 n−1 for n = 1. (11. an = lim f ( ) − an δm ( ) /δn ( ).

5).6) To see that this is a solution of Airy’s equation.5) with f and g analytic functions of t. (11. has f (t) = t. Such integrals arise very commonly in this context. takes this form with f (t) = it. λ real and positive and C a contour that joins two distinct points in the complex t-plane. scaled variable.2 The Asymptotic Evaluation of Integrals We have already seen in Chapters 5 and 6 that the solution of a diﬀerential equation can often be found in closed form in terms of an integral. This is the trivial asymptotic expansion valid for X = O(1)†. For |t| 3 1 3 π 3 − 3 t . Another common example is the solution of Airy’s equation. so we need Re(t ) > 0 on the contour C. y − xy = 0. note that Ai − xAi = 1 2πi 1 3 (t2 − x)ext− 3 t dt = − C 1 3 1 2πi C 1 3 d ext− 3 t dt dt = 0.10). 2 < arg(t) < 6 or − 2 > arg(t) > − 6 . we deﬁne a new. √ √ x + = 1/2 X + 1. which we discussed in Section 6. it is often more convenient to ﬁnd an integral expression for the solution and then seek an asymptotic expansion of the integral.4. In terms of X. given by x = X. . which we wrote in terms of Bessel functions in Section 3. The solution y = Ai(x) can also be written in integral form as Ai(x) = 1 2πi ext− 3 t dt. the Fourier transform. one or both of which may be at inﬁnity. † and of course valid for all X in this simple example.8. As we shall see later.280 ASYMPTOTIC METHODS: BASIC IDEAS expansion valid when x = O( ). which we studied in Chapter 5. with Ai(x) ∼ √ x−1/4 exp − 2 x3/2 /2 π. 11. a procedure that we will use again and again later. 1. we can proceed by considering two simpler cases ﬁrst. Rather than diving straight in and trying to evaluate I(λ) in (11. Instead of trying to develop the asymptotic solution of a diﬀerential equation as some parameter or variable becomes small. xt − 1 t3 ∼ provided that ext− 3 t → 0 as |t| → ∞ on the contour C. whilst the Laplace inversion integral. the contour C in 3 3 Figure 11. in particular. This means that an appropriate contour C originates 3 with arg(t) = − 2π and terminates with arg(t) = 2π (for example. C 1 3 (11. with X = O(1) as → 0. and therefore − 6 < arg(t) < π π 5π π 5π 6 . Ai(x) is distinguished from Bi(x) in that Ai(x) → 0 as x → ∞. Let’s consider integrals of the type I(λ) = C eλf (t) g(t) dt. For example.

∞ ∞ 0 2 −x e x ∞ 0 ˆ e−t dt = ˆ2 π −x e . and we can therefore 2 use Laplace’s method to approximate Kν (x) as Kν (x) ∼ 0 ∞ 1 2 e−x(1+ 2 t ) dt = e−x ∞ 0 e− 2 xt dt. We will see how we can justify this rigorously later. . this becomes Kν (x) ∼ using (3. In fact. Can we ﬁnd the asymptotic form of the gamma function when λ 1? Since the deﬁnition is not quite in the form that we require for Laplace’s method. for x 2x 1. Γ(1 + λ) = 0 tλ e−t dt = eλ log t−t dt. of the gamma function.2.5). we need to make a transformation. of the modiﬁed Bessel function.2 THE ASYMPTOTIC EVALUATION OF INTEGRALS 281 11. which we met in Section 3. where f (t) = g(t) = 1. If we let t = λτ .7) In this case. Example 1 Consider the integral representation ∞ Kν (x) = 0 e−x cosh t cosh νt dt. For t 1. For the moment. cosh t ∼ 1 + 1 t2 .1. Kν (x). this is in the form (11.11. we ﬁnd that Γ(1 + λ) = λ1+λ 0 ∞ eλ(log τ −τ ) dτ. Example 2 Consider the deﬁnition. we can approximate the integral by simply considering the contribution from the neighbourhood of the point where f (t) takes its maximum value. 1 2 ˆ After making the substitution t = t 2/x.7) can be estimated by considering the contribution from the neighbourhood of the maximum value of f (t) alone.7) with f (t) = − cosh t and g(t) = cosh νt. This is known as Laplace’s Method Integrals of the form (11. let’s consider some examples.1 Laplace’s Method Consider the complex integral along the real line b I1 (λ) = a eλf (t) g(t) dt with f (t) real and λ 1. (11. When x 1. The maximum value of f (t) = − cosh t is at t = 0. the integrand is largest at the point where f (t) is largest.

and provides an excellent approximation to the gamma function. this is an asymptotic. 1 2 ˆ If we now let T = T /λ1/2 . λn+1 λn+1 n=0 n=0 ∞ N N since. 1+t In this case. rather than convergent.282 ASYMPTOTIC METHODS: BASIC IDEAS This is in the form of (11. Laplace’s method states that Γ(1 + λ) is dominated by the contribution to the integral from the neighbourhood of τ = 1. We use a Taylor series expansion. log τ − τ = log {1 + (τ − 1)} − 1 − (τ − 1) = −1 − (τ − 1)2 + · · · for |τ − 1| 2 and extend the range of integration. for λ > 2. since the integrand is exponentially small away from t = 0. 10]. 1 ˆ2 2πλλ+ 2 e−λ as λ → ∞. ∞ 0 tn e−λt dt = 1 λn+1 0 τ n e−τ dτ = 1 λn+1 Γ(n + 1) = n! λn+1 . 10 I(λ) ∼ 0 e−λt dt = 1 1 1 − e−10λ ∼ as λ → ∞. Since f (t) = −1 = 0 and the maximum value of f occurs at t = 0 for t ∈ [0. Since f (τ ) = 1/τ − 1 and f (τ ) = −1/τ 2 < 0. we ﬁnd that Γ(1 + λ) ∼ λλ+ 2 e−λ and hence λ! = Γ(1 + λ) ∼ √ 1 ∞ −∞ ˆ e− 2 T dT . We can also extend the range of integration to inﬁnity without aﬀecting the result. We can justify this procedure using the following lemma. . Example 3 Consider the integral 10 I(λ) = 0 e−λt dt. to give N I(λ) ∼ n=0 0 ∞ (−1)n tn e−λt dt = (−1)n Γ(n + 1) (−1)n n! = as λ → ∞. as shown in Figure 11. series representation of I(x). 1 1.2. λ λ In fact we can use the binomial expansion (1 + t)−1 = 1 − t + t2 − t3 + · · · + (−1)n tn + · · · . Note that. as we found for Ei(x).7) with f (τ ) = log τ − τ . f has a local maximum at τ = 1. 1 (11.8) This is known as Stirling’s formula. even though this is only convergent for |t| < 1. f = −t and g = 1/(1 + t). to obtain Γ(1 + λ) ∼ λ1+λ e−λ ∞ −∞ e− 2 λT dT as λ → ∞. using the substitution τ = λt.

The gamma function. Γ(1 + λ) = λ!. with g either bounded or having only integrable singularities. and −1 < α0 < α1 < · · · < αN (which ensures that e−λt g(t) has at worst an integrable singularity at t = 0). 11. . so that we can replace g with its asymptotic expansion and extend the range of integration to inﬁnity.2 THE ASYMPTOTIC EVALUATION OF INTEGRALS 283 Fig. then I(λ) ∼ 0 ∞ N N e−λt n=0 an tαn dt = n=0 an λ−αn −1 Γ(αn + 1) as λ → ∞. Lemma 11.8). and its asymptotic approximation using Stirling’s formula.11. Note that this lemma simply says that the integral is dominated by the contribution in the neighbourhood of t = 0.2. (11. N g(t) ∼ n=0 an tαn as t → 0.1 (Watson’s lemma) If A I(λ) = 0 e−λt g(t) dt. for A > 0.

284 ASYMPTOTIC METHODS: BASIC IDEAS Proof We begin by separating oﬀ the contribution to the integral from the neighbourhood of the origin by noting that N N I(λ) − n=0 δ 0 an λ−αn −1 Γ(αn + 1) = I(λ) − n=0 ∞ 0 N ∞ an 0 A tαn e−λt dt e−λt g(t) dt − e−λt n=0 an tαn dt + δ e−λt g(t) dt for any real δ with 0 < δ < A. Now. and also A δ e−λt |g(t)| dt < Ge−λδ . ∞ δ tαn e−λt dt < e−(λ−1)δ δ ∞ tαn e−t dt < e−(λ−1)δ Γ(αn + 1). Next. we have N I(λ) − n=0 δ N an λ−αn −1 Γ(αn + 1) δ ∞ 0 N < 0 e−λt n=0 an tαn dt + K 0 e−λt tαn+1 dt − e−λt n=0 an tαn dt A + δ N ∞ e−λt g(t) dt ∞ 0 ∞ δ < − n=0 an δ tαn e−λt dt + K e−λt tαn+1 dt − K e−λt tαn+1 dt A + δ N ∞ δ ∞ e−λt g(t) dt < n=0 |an | tαn e−λt dt + Kλ−αn+1 −1 Γ(αn+1 + 1) A δ +K δ e−λt tαn+1 dt + e−λt |g(t)| dt. . we make use of the asymptotic behaviour of g. Since N g(t) − n=0 an tαn < Ktαn+1 for some K > 0 when 0 < t < δ and δ is suﬃciently small. since e−λt < e−(λ−1)δ e−t for t > δ.

2.11). If we assume that the integral is dominated by the contribution from the neighbourhood of the point of stationary phase.3. again along the real line. g(t) = t. I2 (λ) ∼ t0 +δ t0 −δ 1 g(t0 ) exp iλ F (t0 ) + F (t0 )(t − t0 )2 2 t0 +δ dt = eiλF (t0 ) g(t0 ) exp t0 −δ 1 iλF (t0 )(t − t0 )2 2 dt . Let’s consider the situation when there is a single point of stationary phase at t = t0 . we can illustrate why this should be so by considering. of the form b I2 (λ) = a eiλF (t) g(t) dt.11. and the integral can be evaluated by considering the sum of the contributions from each of these points. 1 F (t) ∼ F (t0 ) + (t − t0 )2 F (t0 ) + O (t − t0 )3 2 for |t − t0 | 1. A]. Then. as can be seen in Figure 11.2 THE ASYMPTOTIC EVALUATION OF INTEGRALS 285 for t ∈ [δ. A simple modiﬁcation of this proof can be used to justify the use of Laplace’s method in general. since F (t0 ) = 0. (11. 11. Points where F (t) = 0 are called points of stationary phase. For the moment.2 The Method of Stationary Phase Let’s now consider an integral. as an example. This ﬁnally shows that I(λ) − n=0 an λ−αn −1 Γ(αn + 1) < Kλ−αn+1 −1 Γ(αn+1 + 1) + Ge−λδ N +e−(λ−1)δ n=0 |an | δ ∞ tαn e−t dt + K δ ∞ tαn+1 e−t dt = O(λ−αn+1 −1 ). The rapid oscillations of this integrand lead to almost complete cancellation. and g(t)eiλF (t) = t cos λ(1 − t)2 + it sin λ(1 − t)2 . with F (t) real and λ 1. where G = N A δ |g(t)| dt. the case F (t) = (1 − t)2 . We will consider this more carefully in the next section. except in the neighbourhood of the point of stationary phase.9) Integrals of this type arise when Fourier transforms are used to solve diﬀerential equations (see Chapter 5). provided that F (t0 ) = 0 (see Exercise 11. and hence the result is proved. which has F (t) = 0 when t = 1.

illustrated . and. 2 √1 this becomes I2 (λ) ∼ eiλF (t0 ) g(t0 ) 2 λ|F (t0 )| δ −δ √1 2 λ|F 2 λ|F (t0 )| (t0 )| ei sgn{F (t0 )}T 2 dT. The function t cos λ(1 − t)2 for λ = 5. I2 (λ) ∼ eiλF (t0 ) g(t0 ) We now just need to calculate ∞ 2 λ|F (t0 )| ∞ −∞ e±iT dT. 10 and 20.3. C2 and C3 in the complex T -plane. consider the contours C1 . at leading order as λ → ∞. for some δ 1. 2 To do this. If we now let T = (t − t0 ) 1 λ|F (t0 )|.286 ASYMPTOTIC METHODS: BASIC IDEAS Fig. 11. 2 J= 0 eiT dT = 2 1 2 ∞ −∞ eiT dT.

11.4. 2 . 2 and eiT dT = C1 0 2 R eiT dT → 0 2 ∞ eiT dT as R → ∞. 2 By Cauchy’s theorem eiT dT = 0. The contours C1 . eiT dT = C3 R 2 0 ˆ e−T eiπ/4 dT → − ˆ2 π iπ/4 e as R → ∞. C2 and C3 in the complex T -plane. On the remaining contours. Considering the contour C2 ﬁrst. 11. √ by Jordan’s lemma.2 THE ASYMPTOTIC EVALUATION OF INTEGRALS 287 in Figure 11. Im(T) C3 C2 C1 Re(T) Fig. eiT dT = C2 π/4 0 2 π/4 eiR 2 (cos 2θ+i sin 2θ) iReiθ dθ R 0 e−R 2 sin 2θ θ → 0 as R → ∞. C1 +C2 +C3 2 and hence ∞ 0 e iT 2 √ dT = π iπ/4 e .4.

we get half of the contribution from the stationary phase point. in general.5. a unique point of stationary phase and. we can see that .11) Note that the contribution from the endpoints is of O(λ−1 ). noting that ∂ 2 F+ 2 ∂kx = kx =0 c 2 ky 2 + kz .2. F (t0 ) = 0 and F (t0 ) = 2. . Example Consider (5. v. F+ (k) = c kx + ky + kz − vkz 16π 3 ick ∂F+ = ∂kx ckx 2 2 2 kx + ky + kz and k = (kx . In this case. F (t) = (1 − t)2 . and g(t) = t. there are no points of stationary phase. ˜ f (k) 2 2 2 . therefore. Since (5. the solution of an initial value problem for the wave equation that we considered in Section 5. we can.40) is written in a form independent of the orientation of the coordinate axes. of O(λ−1/2 ). Let’s analyze this solution when t 1 at a point x = vt that moves with constant velocity. 1). Starting with the kx integral. so that g(t0 ) = 1. λ|F (t0 )| In our example. so that t0 = 1.288 ASYMPTOTIC METHODS: BASIC IDEAS Similarly.10) If a = 1 or b = 1. we can assume that v = v(0. which is zero when kx = 0. If a > 1 or b < 1. integrate by parts to obtain I2 (λ) ∼ − i λ g(a) iλF (a) g(b) iλF (b) e e − F (b) F (a) +O 1 λ2 . This is. 2 We conclude that I2 (λ) ∼ eiλF (t0 ) esgn(F (t0 ))iπ/4 g(t0 ) 2π = O(λ−1/2 ) as λ → ∞. when a < 1 < b.40). (11. ∞ 0 e−iT dT = 2 √ π −iπ/4 e . kz ). Consider ∞ ∞ −∞ ∞ −∞ I+ = where g(k) = −∞ g(k)eitF+ (k) dkx dky dkz . 0. ky . whilst contributions from points of stationary phase are larger. λ (11. We conclude that b a teiλ(1−t) dt ∼ e−iπ/4 2 π as λ → ∞.

This is consistent with the fact that disturbances propagate at speed c. we can use the method of stationary phase on the ky integral. a simple change of variable.12) with e an arbitrary unit vector. except on the surface of a sphere of radius ct. 2 2 2 F− (k) = −c kx + ky + kz − vkz . 0. 0. we ﬁnd that I− ∼ 1 8π 2 c2 t ∞ −∞ ˜ f (0. We therefore assume that v = c. kz ) dkz . kz )eitF+ (ky . I+ is much larger. (5. and arrive at ∞ 1 ˜ f (0. when v = c. in the direction of e. However. of O(1/t2 ). ky . of O(1/t). F+ (ky . 1). Since the z-direction can be chosen arbitrarily in this problem. u(vt) ∼ I+ ∼ 2 c2 t 8π −∞ where v = c(0.38). where the solution is given by u(cte) ∼ 1 8π 2 c2 t ∞ −∞ ˜ f (se) ds = O 1 t . so that we are considering a point moving in the z-direction with constant speed c. kz )e−i(c+v)tkz dkz . 16π 3 ick Similarly. kz ) = ˆ ˆ When v = c. kz ) ˆ 2 2 . and also depends upon the total frequency content of the initial conditions. ˆ ˆ 1 2 c2 t 8π ∞ −∞ ˜ f (0. I− = with −∞ g(k)eitF− (k) dkx dky dkz . 0. The amplitude of the solution therefore decays like 1/t. ∞ ˜ f (se) ds. kz )ei(c−v)tkz dkz . kz ) dkz . kz = tkz . and follow the analysis through in the same way.11. −∞ . as given by the integral of their Fourier transform.2 THE ASYMPTOTIC EVALUATION OF INTEGRALS 289 the method of stationary phase shows that I+ ∼ eiπ/4 where ˜ f (0. I− = O(1/t2 ) for t and therefore ∞ 1 ˜ f (0.kz ) dky dkz . Since we have already decided to consider the case v = c. 1. we conclude that the large time solution is small. shows that I+ = O(1/t2 ). kz ) = c ky + kz − vkz . 0. I+ ∼ 2 c2 t 8π −∞ g (ky . 0. which gives us I+ ∼ If we now deﬁne ∞ ∞ −∞ ∞ −∞ 2π ct ∞ −∞ ∞ −∞ 2 2 ky + kz 1/4 g (ky . (11.

1. f and g analytic functions of t = x + iy and C a contour in the complex t-plane. because we have shown rigorously. 11. then (11. The Cauchy–Riemann equations for the analytic function f (t) are (see Appendix 6) ∂ψ ∂φ ∂ψ ∂φ = . the integral is dominated by the contributions from the endpoints of the range of integration. from which we conclude that the lines of constant φ are perpendicular to the lines of constant ψ. . In the absence of such points. that this sort of integral is dominated by the neighbourhood of the point on C1 where the real-valued function φ(t) is largest. Figure 11. we need to know what the curves φ = constant and ψ = constant look like for an analytic function f (t) = φ(x.290 ASYMPTOTIC METHODS: BASIC IDEAS In summary.2. =− . Laplace’s method and the method of stationary phase are special cases of this. Moreover. on which the imaginary part of f (t) is a constant. Our strategy for evaluating the integral (11. in other words on the lines of constant ψ.13) is therefore to deform the contour of integration into one of steepest descent. f (t) = φ(t) + iψ0 . in Lemma 11.13) becomes I(λ) = eiλψ0 C1 eλφ(t) g(t) dt.13) for λ 1. The Method of Stationary Phase Integrals of the form (11.9) can be estimated by considering the contribution from the neighbourhood of each point of stationary phase. We say that the lines where ψ is constant are contours of steepest descent and ascent for φ. Recall that ∇φ is normal to lines of constant φ. If we deform C onto a contour C1 . where F (t) = 0.3 The Method of Steepest Descents Let’s return to consider the more general case. ∂x ∂y ∂y ∂x which show that ∇φ · ∇ψ = ∂φ ∂ψ ∂φ ∂ψ ∂φ ∂φ ∂φ ∂φ + =− + = 0. and we can simply use Laplace’s method. we can deform the contour C without changing I(λ). φ changes most rapidly in the direction of ∇φ. y).5 illustrates these ideas for the function f (t) = t2 . y) + iψ(x. In order to take this further. Since f (t) and g(t) are analytic. I(λ) = C eλf (t) g(t) dt (11. This is what we really want to do. ∂x ∂x ∂y ∂y ∂x ∂y ∂y ∂x and hence that ∇φ is perpendicular to ∇ψ.

I(λ) = C g(t)eλt dt. f (t) = t2 = (x + iy)2 = x2 − y 2 + 2ixy. 2 Case 1: ψ(P ) > 0. and without loss of generality.2 THE ASYMPTOTIC EVALUATION OF INTEGRALS 291 Fig. with C1 the steepest descent contour on which ψ = ψ(P ) and C2 the contour on which φ = φ(Q). On C1 we can therefore make the change of variable f (t) = f (P ) − τ. ψ(Q) > 0 and φ(P ) = φ(Q) In this case. P and Q. The real part of f varies from φ(P ) to φ(Q) as τ varies from zero to φ(P ) − φ(Q). f (P ) λ . 11. y) = x2 − y 2 . In this case. φ(P )−φ(Q) g(t)eλf (t) dt = −eλf (P ) C1 0 g(t) −λτ e dτ f (t) ∼ −eλf (P ) g(P ) 1 as λ → ∞. φ is constant on the hyperbolas x2 − y 2 = φ and ψ is constant on the hyperbolas xy = 1 ψ. of f (t) = t2 . the ends of the contour C lie in the upper half plane. in the complex t-plane. and the lines of steepest ascent and descent for the analytic function f (t) = t2 . The real part. We can deform C into the contour C1 +C2 . and hence φ = x2 − y 2 and ψ = 2xy. 2 with C a contour that joins two points. Therefore.11. as shown in Figure 11. φ(x. Since dτ /dt = −f (t). as shown in Figure 11. whilst ψ = ψ(P ) is constant.5. with 0 τ φ(P ) − φ(Q). Example: f (t) = t2 When f (t) = t2 .6. we take φ(P ) > φ(Q).5.

with ψ constant on C1 and C3 and φ < φ(P ) constant on C3 . On C2 . The exact choice of contour C2 is not important. using Watson’s lemma. its contribution is exponentially smaller than the contributions from C1 and C3 . This is very similar to case 1. we must deform C into C1 + C2 + C3 . . where G = maxC2 |g(t)| and L = length of C2 .7. except that we now also have. since φ(Q) < φ(P ). λ Let’s consider a more interesting example. φ = φ(Q) and the imaginary part. We conclude that I(λ) ∼ −eλf (P ) g(P ) 1 as λ → ∞. We ﬁnd that I(λ) ∼ −eλf (P ) g(P ) g(Q) + eλf (Q) f (P ) f (Q) 1 as λ → ∞. We can write g(t)eλf (t) dt = eλφ(Q) C2 C2 g(t)eiλψ(t) dt GLeλφ(Q) C1 g(t)eλf (t) dt. ψ. using Laplace’s method. a contribution at leading order from C3 . ψ(Q) > 0 and φ(P ) = φ(Q) In this case. as shown in Figure 11.292 ASYMPTOTIC METHODS: BASIC IDEAS Im(t) φ = φ(P) φ = φ(Q) P • C1 C2 ψ = ψ(Q) ψ = ψ(P) Re(t) • Q Fig. varies. The contours C1 and C2 in case 1. since. 11. provided that g(P ) is bounded and nonzero.6. as in case 1. f (P ) λ Case 2: ψ(P ) > 0.

which is the portion of the real axis −1 x 1. . (3. 11. as shown in Figure 11.18). The contribution from C2 . We therefore deform the contour C. whilst the contributions from C1 and C3 are dominated by the neighbourhoods of the endpoints on the real axis. On C1 we make the change of variable t = −1 + iy. and hence φ = −y.2 THE ASYMPTOTIC EVALUATION OF INTEGRALS 293 Im(t) φ = φ(P) = φ(Q) φ < φ(P) Q • C3 • P C2 ψ = ψ(Q) ψ = ψ(P) Re(t) C1 Fig. Example: The Bessel function of order zero with large argument We saw in Chapter 3 that the Bessel function Jn (λ) has an integral representation.11. P = −1 and Q = 1. is exponentially small whatever the choice of Y (Y = 1 in Figure 11. The contours C1 . Setting n = 0 and making the change of variable t = sin θ leads to 1 π 1 −1 J0 (λ) = √ eiλt dt. ψ = x. C2 and C3 in case 2.7. so that 1 π eiλt 1 dt = π 1 − t2 Y 0 √ C1 e−iλ e−λy 2iy + y 2 i dy ∼ e−iλ i √ π 2i ∞ 0 y −1/2 e−λy as λ → ∞. In this case. y = Y > 0. 1 − t2 (11.8. into C1 + C2 + C3 . so that φ(P ) = φ(Q) = 0. The contours of steepest descent through P and Q are just the straight lines ψ = x = −1 and ψ = x = 1 respectively.8).13) with f (t) = it.14) This is of the form (11.

dt ∼ √ 1 − t2 2πλ So. and arrive at 1 π Similarly.294 ASYMPTOTIC METHODS: BASIC IDEAS Fig. we can write ˆ √ this integral in terms of a gamma function. The contours C1 . f (t) = t2 . t = 0 is neither a local maximum nor a local minimum. It is a saddle point. 1 π and hence J0 (λ) ∼ π 2 cos λ − λπ 4 as λ → ∞. 11.8. Since f (t) = 2t. By making the change of variable y = λy. using Laplace’s method. Γ(1/2) = π. In . However. is this the whole story? Let’s return to our original example. dt ∼ √ 1 − t2 2πλ eiλt eiλ−iπ/4 as λ → ∞. C2 and C3 for evaluating the asymptotic behaviour of J0 (λ). the real and imaginary parts of f are stationary at t = 0. √ C3 √ C1 eiλt e−iλ+iπ/4 as λ → ∞.

exponentially small. and.5. f (Q) λ On the steepest descent contour C3 . to arrive at g(t)eλt dt = i C3 ψ(P ) 2 ψ(Q) g(iy)e−λy dy. 11. since this is the only line with ψ constant that connects the two valleys.1. g(t)eλt dt ∼ −eλf (P ) C1 2 g(P ) 1 . P and Q lie in diﬀerent valleys of the function f (t) = t2 . As usual. C2 .11. using Laplace’s method as in cases 1 and 2. The contours C1 .2 THE ASYMPTOTIC EVALUATION OF INTEGRALS 295 fact. 2 . C3 . f (P ) λ g(t)eλt dt ∼ eλf (Q) C5 2 g(Q) 1 .9. C4 and C5 in case 3.4 and Appendix 6). as shown in Figure 11. we were able to deform the contour C within one of the valleys. The real part of f (t) = t2 has a ridge along the line y = 0 and a valley along the line x = 0.9. no analytic function can have a local maximum or minimum. as can be seen in Figure 11. In cases 1 and 2. we can make the simple change of variable t = iy. since its real and imaginary parts are harmonic (see Section 7. Case 3: ψ(P ) > 0 and ψ(Q) < 0 In this case. the integrals on C2 and C4 are Im(t) C2 C1 P C3 ψ = ψ(P) > 0 Re(t) ψ = ψ(Q) < 0 C4 C5 Q Fig. and we must deform the contour so that it runs through the line of steepest descent at the saddle point. which we studied above.

note that this contour passes through T = −i in the direction of 1 + i. In order to calculate the leading order contribution from C1 . We .18) Similarly. so that Ai(x) = (−x)1/2 2πi e(−x) C 3/2 (−T − 1 T 3 ) dT. Example: The Airy function. this is the dominant contribution. 3 there are saddle points at τ = ±1.15) If φ(P ) 0 and φ(Q) 0. We can now deform C into the steepest descent contour through τ = −1. Ai(x) = 1 2πi ext− 3 t dt. since arg(t) = arg(τ ). the integral is dominated by the contribution from the steepest descent contour through the saddle point. λ (11. In our example. The contours of constant imaginary part.17) This is now in the standard form. 3 (11.19) In this case. 3 (11. C. Ai(x). The contours of constant imaginary part. are shown in Figure 11. Laplace’s method shows that g(t)eλt dt ∼ −ig(0) C3 2 ∞ −∞ e−λy dy = −ig(0) 2 π . for |x| As we have seen. what if the contour of integration. the integral is dominated by the contribution from the neighbourhood of this saddle point. extends to inﬁnity? For such an integral to converge. The integral along C2 is exponentially smaller. y = 0. given by (11. with f (τ ) = τ − 1 τ 3 . (11. As we have seen. f (T ) = −T − 1 T 3 and f (T ) = −1 − T 2 . C 1 3 1 (11.10. and. we can make the transformation t = (−x)1/2 T . Finally. if we let P and Q head oﬀ to inﬁnity along the two diﬀerent valleys. ψ.15). we write τ = −1 + iy and use Laplace’s method to show that Ai(x) ∼ x1/2 2πi ∞ −∞ ex 3/2 1 (−2/3−y2 ) i dy = √ x−1/4 exp − 2 x3/2 3 2 π as x → ∞. if g(0) is bounded and nonzero. so there are saddle points at 3 T = ±i. are shown in Figure 11. it must extend into the valleys of f (t) as |t| → ∞.296 ASYMPTOTIC METHODS: BASIC IDEAS This integral is dominated by the behaviour close to the saddle point.11. Since the contour C is vertical at this point. By deforming C into the contour C1 + C2 + C3 . when x < 0.6).10. For x > 0 we can make the change of variable 3 t = x1/2 τ and. marked as C in Figure 11. ψ.16) with arg(t) = ± 2π as |t| → ∞. Since f (τ ) = 1 − τ 2 . deform back to the original contour to obtain Ai(x) = x1/2 2πi ex C 3/2 (τ − 1 τ 3 ) dτ. we can see that the integral is dominated by the contributions from the two saddle points. (11.

2 (−x)−1/4 √ sin π π 2 (−x)3/2 + 3 4 as x → −∞.10. 2 −1/4 (−1 + i) (−T − 1 T 3 ) dt ∼ (−x) 3 2πi π − 2 i(−x)3/2 e 3 . Contours where the imaginary part of f (τ ) = τ − 1 τ 3 is constant.20) The transition from exponential decay as x → ∞ to decaying oscillations as x → −∞ is shown in Figure 11. Also shown is the behaviour predicted by our analysis . (−x)1/2 2πi and hence Ai(x) ∼ e(−x) C3 3/2 (−x)−1/4 (1 + i) 2πi π 2 i(−x)3/2 e3 . 3 therefore make the substitution T = −i + (1 + i)s and use Laplace’s method to arrive at (−x)1/2 2πi e(−x) C1 3/2 1/2 (−T − 1 T 3 ) dt ∼ (−x) e 2 i(−x)3/2 (1 + i) 3 3 2πi ∞ −∞ e−2(−x) 3/2 2 s ds = Similarly. 11.2 THE ASYMPTOTIC EVALUATION OF INTEGRALS 297 Fig.11. (11.12.

just the neighbourhood of the single point of stationary phase. which can be seen to be in excellent agreement with Ai(x). where the real part of f is constant. even for moderate values of |x|. 2 with a < 1 < b. we give a justiﬁcation for the success of the method of stationary phase. the real axis with a x b. however.10).11. 2 . C3 . Consider the example that we looked at earlier. because the integrand is analytic.13. b I2 (λ) = a teiλ(1−t) dt. and. for large |x|. along 3 with a single contour. into C1 + C2 + C3 + C4 + C5 . as shown in Figure 11. To end this section. (11. The same arguments as those that we used above show that the largest contribution comes from the neighbourhood of the saddle point on the steepest descent contour. C2 . we obtain the same result. Using Cauchy’s theorem on the analytic function teiλ(1−t) . 11. even though the contour is diﬀerent. This is.298 ASYMPTOTIC METHODS: BASIC IDEAS Fig. we can deform the contour C. Contours where the imaginary part of f (T ) = −T − 1 T 3 is constant.

t By integrating by parts repeatedly.12. (f) log(1 + ). (a) sinh . (b) tan . show that the series is asymptotic as x → ∞.4 Show that the following are asymptotic sequences. provided that the complex variable is restricted to a domain that you should determine.3 et dt as x → ∞. (e) cot . x x2 x x By considering the error in terminating the expansion after the term in x−n . 11. Exercises 11. 11.2 11. (c) sinh(1/ ). as → 0+ . Consider the integral I(x) = e−x 1 x 11. (g) (1 − cos )/(1 + cos ). . (d) e− .1 Calculate the order of magnitude of the following functions in terms of the simplest function of possible. (h) exp {− cosh (1/ )}. Show that e−1/ = o( n ) as → 0 for all real n. develop a series expansion of the form I(x) = 1 1 2 3! + + 3 + 4 − (1 + 1 + 2 + 3!) e1−x + · · · . The Airy function Ai(t).EXERCISES 299 Fig.

.13. use Laplace’s method to show that ∞ 0 e4λ t −t dt ∼ 1 + (t/λ)2 3/2 1/2 2 π 3λ2 e as λ → ∞. . The contours C1 .5 ). . C2 . 1. . which has an integral representation Kν (z) = 1 2 ∞ −∞ eνt−z cosh t dt. 6 11. C3 . 1.300 ASYMPTOTIC METHODS: BASIC IDEAS Fig. .7 Use Laplace’s method to show that the modiﬁed Bessel function Kν (z). (b) log(1 + 11. 11. . 11. . n (a) sin 1 . as n → ∞. → 0. n = 0. C4 and C5 for the method of stationary phase.6 After making the substitution t = uλ. as Use Watson’s lemma to show that ∞ 0 (t + t2 )−1/2 e−λt dt ∼ π λ 1− 9 1 + 4λ 32λ2 as λ → ∞. . n = 0. .

10 Show that as λ → ∞ 1+5i (a) 4 1+5i eiλt dt ∼ i eiλt dt ∼ 2 2 e16iλ 8 i 1 − . λ 11. 11.8 You may assume that the integral is dominated by the contribution from the neighbourhood of the maximum value of νt − z cosh t. F (t0 ) = 0 and F (t0 ) = 0.12 Consider the integral Q I(λ) = P eλ( 2 t 1 2 +it) dt.9 Use the method of stationary phase to show that π/2 eiλ(2t−sin 2t) dt ∼ 0 1 Γ 2 4 3 6 λ 1/3 eiπ/6 as λ → ∞.) Determine the next term in the asymptotic expansion. 11. (Hint: Let z = Reiθ . 11. λ Show that Stirling’s formula for Γ(z + 1) when z 1 holds for complex z provided |z| 1 and −π < arg(z) < π. 2 I(λ) ∼ −i 11.13 2πeλ as λ → ∞.9) when F (t) has a single point of stationary phase at t = t0 with a < t0 < b. Use the method of stationary phase to show that I2 ∼ 2 g(t0 )eiλF (t0 )+iπsgn(F 3 (t0 ))/6 Γ 1 3 6 λ|F (t0 )| 1/3 for λ 1. Show that if P = − 1 and 2 2 Q = 2. where P and Q are points in the complex t plane. (2 + i)λ Show that if P = − 1 + i and Q = 1 − 3i. using Kν (z) ∼ π −ν e 2ν 2ν z ν . .EXERCISES 301 can be approximated as ν → ∞. I(λ) ∼ e(2+2i)λ as λ → ∞. Use the method of stationary phase to show that ∞ 0 1 eiλ( 3 t −t) dt ∼ 1+t 2 1 3 π −2iλ/3+iπ/4 e as λ → ∞.11 Consider the integral (11. Sketch the contours of constant real and imaginary parts of 1 t2 + it. λ 32λ2 (b) −5−i π iπ/4 e . λ 11. with z = O(1) and positive.

D−2m (x) ∼ x−2m e−x as x → ∞.14 From the integral representation. Iν (x) = 1 π π ex cos θ cos νθ dθ − 0 sin νπ π ∞ 0 e−x cosh t−νt dt. D−2m+1 (x) = 2 1 xe−x /4 (m − 1)! ∞ 0 e−s sm−1 x2 + 2s −m+1/2 ds. Show that for real x and ﬁxed m. show that ex Iν (x) ∼ √ . for m a positive integer.302 ASYMPTOTIC METHODS: BASIC IDEAS 11. D−2m+1 ∼ x−2m+1 e−x 2 /4 . of the modiﬁed Bessel function of order ν.15 as x → ∞ for real x and ﬁxed ν. The parabolic cylinder functions are deﬁned by D−2m (x) = 2 1 xe−x /4 (m − 1)! ∞ 0 e−s sm−1 x2 + 2s −m−1/2 ds. 2 /4 . . 2πx 11.

CHAPTER TWELVE Asymptotic Methods: Diﬀerential Equations In this chapter we will apply the ideas that we developed in the previous chapter to the solution of ordinary and partial diﬀerential equations. and the ideas that we will use are far more transparent. and hence 1 0 2 4 .1. we get x3 − x = 0. it is rather cumbersome to use. These solutions are obviously not exact when is small but nonzero.1 Example: A Regular Perturbation Consider the cubic equation x3 − x + = 0.2) We can now substitute this into (12. and hence x0 = −1. 0 or 1. an asymptotic solution) of the form x = x0 + x1 + 2 x2 + O( 3 ).1 An Instructive Analogy: Algebraic Equations Many of the essential ideas that we will need in order to solve diﬀerential equations using asymptotic methods can be illustrated using algebraic equations. (3x0 − 1) x1 + 1 = 2x1 + 1 = 0. At O( 2 ). We will 0 concentrate on the solution with x0 = 1. These are called the leading order solutions of the equation. 12. At O( ). 0 or 1. Let’s suppose instead that we only need to ﬁnd the solutions for 1. We will consider two illustrative examples. and hence x = −1. At leading order we obviously get x3 − x0 = 0.1) Although there is an explicit formula for the solution of cubic equations.1) and equate powers of . If we simply set = 0. (12. (12. These are much more straightforward to deal with than diﬀerential equations. which we can rearrange into a hierarchy of powers of x3 − x0 + {(3x0 − 1) x1 + 1} + 0 2 in the form 3 3x0 x2 + 3x2 − 1 x2 + O 1 0 = 0. 12. 3x0 x2 + 3x2 − 1 x2 = 3 + 2x2 = 0. so let’s try to improve the accuracy of our approximation by seeking an asymptotic expansion of the solution (or more succinctly. and hence x1 = − 1 . This gives us x0 + x1 + 2 x2 3 − x0 + x1 + 2 x2 + + O 3 = 0.

The function x3 − x + for = 0. which we neglect at leading order when x = O(1). of course.2 shows how the function x3 + x2 − 1 changes qualitatively for = 0 and 0< 1. continue to higher order if necessary. Figure 12. However. The equation changes from quadratic to cubic.2).3) At leading order for 1. 12. and = 0. since we have found asymptotic expansions for all three roots of the cubic equation using the simple expansion (12.1.1). we know that a cubic equation is meant to have three solutions. and hence x = ±1. This is a regular perturbation problem.2 Example: A Singular Perturbation Consider the cubic equation x3 + x2 − 1 = 0. This shows 8 that x=1− 3 1 − 2 8 2 + O( 3 ) for 1.304 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS x2 = − 3 . where the solution for = 0 is qualitatively diﬀerent to the solution when 0 < 1.1. We could. Similar expansions can be found for the other two solutions of (12. Fig. broken line. .1 shows that the function x3 − x + is qualitatively similar for = 0 and 0 < 1. becomes comparable to the term x2 for suﬃciently large x. Figure 12. x2 − 1 = 0. the term x3 . solid line. The key point is that we have implicitly assumed that x = O(1). and the number of solutions goes from two to three. 12. speciﬁcally when x = O( −1 ). However. (12. What’s happened to the other solution? This is an example of a singular perturbation problem.1.

4) now becomes X3 + X2 − 2 = 0. with α to be determined. So. and assume that x = O( α ).5) Since only 2 and not appears in this rescaled equation. as expected. we expand X = X0 + 2 3 2 X1 + O( 4 ). X0 = 0 will just give us the two solutions with x = O(1) that we have already considered. Equation (12. we can construct the two O(1) solutions.2. If we deﬁne a scaled variable. The function x3 + x2 − 1 for = 0.12. (12. solid line. 12. and hence α = −1. Of course. X0 + X0 = 0. Since we know that there must be three solutions. If α > 0. At O( 2 ). in the same manner as we did for the previous example.4) We must choose α in order to obtain an asymptotic balance between two of the terms in (12.1 AN INSTRUCTIVE ANALOGY: ALGEBRAIC EQUATIONS 305 Fig. the ﬁrst two terms are large. how can we proceed in a systematic way? If we expand x = x0 + x1 + O( 2 ). the ﬁrst two terms are small and cannot balance the third term. which is of O(1).1. At leading order. and = 0.4). (12. and hence X0 = −1 or 0. This requires that 3α+1 3 X = O( 2α X 2 ). x = ±1 + O( ). x = α X. This means that x = −1 X = O( −1 ).3) becomes 3α+1 X3 + 2α X 2 − 1 = 0. we conclude that the other solution cannot have x = O(1). and we can choose α so that they are of the same asymptotic order. This gives 3α + 1 = 2α. (12. so we take X0 = −1. and hence 3α+1 = O( 2α ). (−1 + 2 X1 )3 + (−1 + 2 X1 )2 − 2 + O( 4 ) = 0 . broken line. with X = O(1) for 1. If α < 0.

Ordinary Diﬀerential Equations The solution of ordinary diﬀerential equations by asymptotic methods often proceeds in a similar way to the solution of algebraic equations.8) . and equate powers of the small parameter. subject to y0 (0) = 0 and y0 (1) = 1. (12. sinh 1 Similarly. subject to the boundary conditions y(0) = 0. y0 − y0 = 0. Therefore X = −1 + 12. we will begin by considering some simple. This determines a sequence of simpler equations and boundary conditions that we can solve.1 Regular Perturbations Consider the ordinary diﬀerential equation y + 2 y − y = 0. + O( 4 ). We seek a solution of the form y(x) = y0 (x) + y1 (x) + O( 2 ).7) show that y1 − y1 = −2y0 = −2 y0 (0) + y1 (0) + O( 2 ) = 0. which we discussed in the previous section. which has solution sinh x y0 (x) = . we obtain (y0 + y1 ) + 2 (y0 + y1 ) − (y0 + y1 ) + O( 2 ) = y0 + y1 + 2 y0 − y0 − y1 + O( 2 ) = 0. we could solve this constant coeﬃcient ordinary diﬀerential equation analytically using the method described in Appendix 5. 12.2.6) and retain terms up to O( ). to be solved for 0 x 1. sinh 1 If we now substitute the expansion for y into (12.2 2 2 + O( 4 ) = 0. substitute into the equation and boundary conditions. At leading order. (12. We assume that an asymptotic expansion of the solution exists. constant coeﬃcient linear ordinary diﬀerential equations before moving on to study both nonlinear ordinary diﬀerential equations and some simple partial diﬀerential equations. y(1) = 1. and hence cosh x . In order to introduce the main ideas.6) Of course. the boundary conditions (12. and hence x = −1/ + + O( 3 ).306 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS gives −1 + 3 2 X1 + 1 − 2 2 X1 − and hence X1 = 1. y0 (1) + y1 (1) + O( 2 ) = 1. but it is instructive to try to construct the asymptotic solution when 1.7) (12.

(12. One subtle point is that. for example.3. for this simple. (12.2 ORDINARY DIFFERENTIAL EQUATIONS 307 and hence y1 (0) = 0.2. although we believe that the next term in the asymptotic expansion of the solution. cosh 1 12. and using the constants in the homogeneous solution. and by checking that the asymptotic solution makes sense in terms of the underlying physics of the system that we are modelling. our faith is justiﬁed.10) sinh 1 cosh 1 for 1. is uniformly smaller than the two retained terms for suﬃciently small. and in particular for nonlinear problems.6).8) in the form y1p = kx sinh x.11) The observant reader will notice that this is the same as the previous example. this is not possible. to be solved for 0 x 1. as can be seen in Figure 12.2 The Method of Matched Asymptotic Expansions Consider the ordinary diﬀerential equation y + 2y − y = 0. by comparing asymptotic solutions with numerical solutions. The ratio of the second to the ﬁrst term in this expansion is (1 − x) tanh 1. (12. Although. y(1) = 1. This leads us to believe that the asymptotic solution (12. where we were able to ﬁnd an upper bound for RN . given by (11. we arrive at y1 = (1 − x) and hence sinh x sinh x + (1 − x) + O( 2 ).3). We again seek an asymptotic solution of the form y(x) = y0 (x) + y1 (x) + O( 2 ). and an act of faith is involved in trusting that our asymptotic solution provides a valid representation of the exact solution.9) By seeking a particular integral solution of (12. which we write as O( 2 ) in (12.12. This faith can be bolstered in a number of ways. constant coeﬃcient ordinary diﬀerential equation.1.1. which is uniformly small for 0 x 1. The situation is analogous to the example that we studied in Section 12.10) is uniformly valid in the region of solution. to satisfy the boundary conditions (12. subject to the boundary conditions y(0) = 0.10). We do not have a rigorous estimate for the size of the neglected term in the way that we did when we considered the exponential integral. The sensible applied mathematician always has a small. y1 (0) = 0. but with the small parameter moved to the highest derivative term. y1h = A sinh x + B cosh x. For (12.9). in general. we could write down the exact solution and prove that the remainder is of O( 2 ).12) (12. we have not proved this. nagging doubt at the back of their mind about the validity of an asymptotic solution. y(x) = sinh x . .

The problem is that for = 0. and is analogous to the example we studied in Section 12. which gives y1 (x) = 1 (1 − x)e(x−1)/2 . 2 . At leading order. 2y0 − y0 = 0. At O( ) we have 1 2y1 − y1 = −y0 = − e(x−1)/2 . 2 for some constant k. How can we satisfy two boundary conditions using only one constant? Well. which has the solution y0 = Aex/2 for some constant A. 12. the boundary conditions require that y0 (0) = 0 and y0 (1) = 1. and we just use y1 (1) = 0. Again.308 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS Fig.1.2). which gives y0 (x) = e(x−1)/2 .6) when = 0. Let’s proceed by satisfying the boundary condition y0 (1) = 1. This equation can be solved using an integrating factor (see Section A5. the equation is of ﬁrst order. second order equation. of course we can’t.25.3. we cannot satisfy both boundary conditions. and therefore qualitatively diﬀerent from the full. This is a singular perturbation. However. Exact and asymptotic solutions of (12. 4 to be solved subject to y1 (0) = 0 and y1 (1) = 0.2. which gives 1 y1 (x) = − xe(x−1)/2 + ke(x−1)/2 .

which we neglected at leading order.13) for 1.2 ORDINARY DIFFERENTIAL EQUATIONS 309 Finally. for some constant A. where x = O( ). to be solved subject to Y0 (0) = 0. the term y . In order to ﬁx a second boundary condition for (12. At leading order. and hence α = 1.14). the second term in this equation is large. which clearly does not 2 satisfy the boundary condition y(0) = 0. becomes important because y varies rapidly. in some small neighbourhood of x = 0. which does satisfy y(1) = 1. and to obtain an asymptotic balance at leading order we must have 1−2α = O( −α ). x→0 (12. x = 1 does not lie in the inner region. (the inner expansion). +2 2 dX dX (12. At O( ) we obtain the equation for Y1 (X) as Y1 + 2Y1 = Y0 = A(1 − e−2X ). we now know that y ∼ e(x−1)/2 Y ∼ A(1 − e −2X for x 1 (the outer expansion). across which y adjusts to satisfy the boundary condition.11) gives Y dY − Y = 0. we must have X→∞ lim Y (X) = lim y(x). that this is correct. So x = X. (12. At leading order. We now expand Y (X) = Y0 (X) + Y1 (X) + O( 2 ). . We must therefore introduce a boundary layer at x = 0. with α > 0 (so that x → 0. This suggests that y → e−1/2 (1 + 1 ) as x → 0. with the solution in the outer region. d 2Y dY − Y = 0. 1) and X = O(1) as We rescale by deﬁning x = α X. + 2 −α dX 2 dX Since α > 0. Y0 + 2Y0 = 0. However. with outer solution y(x). The solution is Y0 = A(1 − e−2X ). which means that 1 − 2α = −α.15) which gives A = e−1/2 .12. It is usual to refer to the region where x 1 as the outer region. we will have to make sure that the solution in the inner region is consistent. We will see below. and the boundary layer region where x = O( ) as the inner region with inner solution Y (X). The idea is that. The other boundary condition is to be applied at x = 1. Substituting this into (12. in a sense that we will make clear below.14) 1−2α d 2 and Y (0) = 0. ) for X = O(1). x = O( ) For these two expansions to be consistent with each other. where we make this vague notion of “consistency” more precise. and write y(x) = Y (X) for X = O(1). this gives 1 y = e(x−1)/2 1 + (1 − x) 2 + O( 2 ).

ˆ When X = O(1). In terms of the ˆ intermediate variable. the two-term asymptotic expansions are Y1 = y ∼ e(x−1)/2 + Y ∼ A(1 − e−2X ) + 1 (1 − x)e(x−1)/2 for 2 x 1.16) and (12. and the inner and outer expansions are known as matched asymptotic expansions. (12. we need A = e−1/2 . that we made above. whereby we make the inner and outer expansions consistent. It should be . 2 (12. consistent with the crude assumption. which we need to determine. This can now be solved using an integrating factor.4. where x 1.17) Since (12. The point is that in such a region we expect both expansions to be valid. we can expand the exponentials as Taylor series. ˆ ˆ ˆ We deﬁne x = β X with 0 < β < 1. and also c2 = − 1 e−1/2 . To summarize.15). 1 AX(1 + e−2X ) − c2 (1 − e−2X ) 2 for X = O(1). 2 (12.16) ˆ X gives X = β−1 ˆ X. for some constant c1 . the outer expansion becomes ˆ Y ∼ e−1/2 exp 1 2 β ˆ X + 1 (1 − 2 β ˆ X)e−1/2 exp 1 2 β ˆ X . since exp(−2 we have ˆ X) = o( n ) for all n > 0 (it is exponentially small for β < 1). and write y = Y (X). 2 for some constant c2 . x = O( ). is known as asymptotic matching. We can determine the constants A and c2 by forcing the two expansions to be consistent in the sense that they should be equal in an intermediate region or overlap region. Now.310 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS Integrating this once gives 1 Y1 + 2Y1 = A X + e−2X 2 + c1 .17) must be identical. 1 ˆ ˆ Y = A + A β X − c2 + o( ). A comparison between the one-term inner and outer solutions and the exact solution is given in Figure 12. X. and the solution that satisﬁes Y1 (0) = 0 is 1 AX(1 + e−2X ) − c2 (1 − e−2X ). and ﬁnd that 1 ˆ Y = e−1/2 1 + 2 Since x = X = β β ˆ X 1 + e−1/2 + o( ). This 2 process. the inner expansion is β−1 ˆ Y ∼ A 1 − exp −2 + 1 A 2 β−1 ˆ X β−1 ˆ X 1 + exp −2 β−1 β−1 ˆ X − c2 1 − exp −2 ˆ X .

4. Exact and asymptotic solutions of (12. A little later. and is much easier to use. Van Dyke’s Matching Principle The use of an intermediate variable in an overlap region can get very tedious in more complicated problems. we can write y(x) in terms of X = x/f ( ). A method that works most. but let’s start with the explanation.M ) . of the time.12.11) when = 0. with x = f ( )X. 12. .2 ORDINARY DIFFERENTIAL EQUATIONS 311 clear that the inner expansion is a poor approximation in the outer region and vice versa. Fig. is Van Dyke’s matching principle. but not all. In order to analyze how the outer expansion behaves in the inner region.25. We denote this by y (N. we can write Y (X) in terms of x. we will show how to construct. This principle is much easier to use than to explain. a composite expansion that is uniformly valid for 0 x 1. using the inner and outer expansions. Similarly. the M th order inner approximation of the outer expansion. and retain M terms in the resulting asymptotic expansion. Let N y(x) ∼ n=0 φn ( )yn (x) be the outer expansion and N Y (X) ∼ n=0 ψn ( )Yn (X) be the inner expansion with respect to the asymptotic sequences φn ( ) and ψn ( ).

Van Dyke’s matching principle states that Y (2. 2 for x = O(1). In terms of the outer variable. the M th order outer approximation of the inner expansion. 2 Composite Expansions Although we now know how to construct inner and outer solutions. We denote this by Y (N. as expected. Van Dyke’s matching principle states that y (N. it would be more useful to have an asymptotic solution valid uniformly across the whole domain of solution. we can write the composite solution of any order as M (M. y (2.2) . Let’s see how this works for our example. Note that. We can construct such a uniformly valid composite expansion by using the inner and outer expansions.N ) . We simply need to add the two expansions and subtract the expression that gives their overlap.2) = e−1/2 1 + for X = O(1).2) . . For our example problem.17) or (12.1) = e(x−1)/2 + e−1/2 (1 − e−2X ) − e−1/2 = e(x−1)/2 − e−1/2−2x/ for 0 x 1 as → 0. the inner expansion is Y (X) ∼ A(1 − e−2x/ ) + 1 x A 1 + e−2x/ 2 − c2 1 − e−2x/ 1 ∼ Y (2.4. In terms of the inner variable.N ) .312 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS and retain M terms in the resulting expansion. in terms of Van Dyke’s matching principle. 1 1 y (2. and therefore gives A = e−1/2 and c2 = − 1 e−1/2 rather more painlessly than before.M ) .M ) = Y (M.2) = e−1/2 1 + x + 2 2 . the one-term composite expansion is y ∼ y0 + Y0 − y (1.N ) = y ∼ yc n=0 N yn (x) + n=0 Yn (X) − y (M. and shows good agreement with the exact solution across the whole domain. 0 x 1 in the example.2) = A + Ax − c2 .2) or Y (2. (12.2) = y (2. The overlap is just the expression that appears to the appropriate order in the intermediate region. the outer expansion is 1 1 y(x) ∼ exp − + X 2 2 + 1 1 1 (1 − X) exp − + X 2 2 2 1 1 X+ 2 2 . ∼ y (2. In terms of the outer variable. valid in the inner and outer regions. This composite expansion is shown in Figure 12.16). or equivalently the matched part.

as we have seen. we ﬁnd that we can only obtain an asymptotic balance at leading order by taking α = 1. since then this solution is only needed for X 0. as before. to be solved for −1 < x < 1.18) We will try to construct the leading order solution for 1. Interior Layers Singular perturbations of ordinary diﬀerential equations need not always result in a boundary layer. and hence y = k − x for some constant k.2 ORDINARY DIFFERENTIAL EQUATIONS 313 The Location of the Boundary Layer In our example. and. Why is this correct? Let’s see what happens if we assume that there is a boundary layer at x = x0 .19) (12. YXX + 2x0 YX = 0. For x0 > 0 this grows exponentially as X → −∞. Y0 becomes exponentially large. for a leading order balance we need In this case. subject to the boundary conditions y(−1) = 2. At leading order. we can construct an asymptotic solution. As an example. If y(−1) = 2 we need y = 1−x. when 1−2α −2x0 X YXX + 2XYX + 2 α X = 0. X = O(1). At leading order. As before. if x0 = 0 and x0 = 1 this is an interior layer. with α > 0 and Y . For a leading order asymptotic balance we need α = 1/2. This forces us to take x0 = 0. As X → −∞. (12. and hence a boundary layer with width of O( 1/2 ). whilst for x0 < 0 this grows exponentially as X → ∞. (12.18) becomes 1−2α YXX + 2(x0 + α X)( −α YX + 1) = 0. whilst if y(1) = 3 we need y = 4−x. In either case. We clearly cannot satisfy both boundary conditions with the same outer solution. X = O(1) for 1. at leading order. with Y . = O( −α If x0 = 0. In terms of these scaled variables. This suggests that we need x0 = 0. we cannot match these exponentially growing terms with the outer solution. y(1) = 3. and hence α = 1. consider y + 2xy + 2x = 0. YXX + 2XYX = 0. so let’s look for a boundary or interior layer at x = x0 by deﬁning y(x) = Y (X) and x = x0 + α X. The leading order outer solution satisﬁes 2x(y + 1) = 0. 1−2α ). . when we constructed the outer solution. Strictly speaking. we chose to satisfy the boundary condition at x = 1 and assume that there was a boundary layer at x = 0. and cannot be matched to the outer solution.12. and hence Y = A + Be . We deﬁne scaled variables y(x) = Y (X) and x = x0 + α X. so that x = x0 + X and Y + 2Y − = 0. Y0 = A0 + B0 e−2X .

k) is available in MATLAB as ellipj.5 for various values of . k) for various values of δ. the Jacobian elliptic function x = sn(t. 2 which leads to the one-term composite solution y ∼ yc = −x + 1 2 5 + 3 erf x √ for −1 x 1 and 1. the solution that corresponds to sn(t.2. k) satisﬁes the equation dx = dt 1 − x2 1 − k 2 x2 . since the interior layer is at x = 0. the outer solution must be y= 1 − x for −1 4 − x for O( 1/2 x < O( )<x 1/2 ). Note that the boundary conditions at x = ±1 are only satisﬁed by the composite expansion at leading order. we must have Y → 4 as X → ∞ and Y → 1 as X → −∞. with δ the ﬁrst quarter period of x(t). 1. dX 2 and hence X Y =B+A −∞ e−s ds. it seems reasonable to assume that the period of the solution is large but ﬁnite. 0.21) subject to x = 0 when t = 0.314 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS which.18. 0) in the phase portrait shown in Figure 9.6 shows sn(t. and has periodic solutions for k = 1. .3 Nonlinear Problems Example 1: Elliptic functions of large period As we have already seen in Section 9. and hence the period tends to inﬁnity as k → 1. When k = 1. (12. after multiplying by the integrating factor. k) reaches its ﬁrst maximum. 2 Now.4. (12. When k is close to unity. gives 2 d eX YX = 0. This allows us to ﬁx the constants A and B and ﬁnd that 3 Y (X) = 1 + √ π X −∞ e−s ds = 2 1 (5 + 3 erf (x)) . Figure 12. k) is a heteroclinic path that connects the equilibrium points (±1. We seek an asymptotic solution of the form x = x0 + δ 2 x1 + δ 4 x2 + O δ 6 .20) This is illustrated in Figure 12. and we can see that the period does increase as δ decreases and k approaches unity. Since y → 4 as x → 0+ and y → 1 as x → 0− . Can we quantify this? 1. and seek an asymptotic solution for Let’s assume that k 2 = 1 − δ 2 . with 0 x 1. 12. The function sn(t. eX . The quarter period is simply the value of t when sn(t.

subject to x0 (0) = 0. (12. and we treat this as the outer expansion.5. dx1 1 1 = −2x0 x1 + x2 = −2 tanh tx1 + tanh2 t. subject to x1 (0) = 0. 0 dt which has solution x0 = tanh t. (12. so we should expect any asymptotic expansion that we develop to become nonuniform as x → 1.12. The composite asymptotic solution. dt 2 0 2 Using the integrating factor cosh2 t. At O(δ 2 ).21) is dx = 1 − x2 dt 1 + δ2 x2 1 − x2 1/2 x4 1 1 = 1 − x2 + δ 2 x2 + δ 4 + O(δ 6 ). Using a binomial expansion. 12. 4 . of (12. 4 1 tanh t − t sech2 t . At leading order.18). dx0 = 1 − x2 .2 ORDINARY DIFFERENTIAL EQUATIONS 315 Fig.20). we can ﬁnd the solution x1 = We can now see that 1 x ∼ 1 − 2e−2t + δ 2 + O(δ 4 ) as t → ∞. 2 8 1 − x2 This binomial expansion is only valid when x is not too close unity.

128 We can now see that the fourth term in this expansion becomes comparable to the third when δ 4 e2t = O(δ 2 ). there is no nonuniformity in this expansion. dx2 1 = −x2 − 2x0 x2 + x0 x1 − 1 dt 8 x4 0 1 − x2 0 subject to x2 (0) = 0. and hence as t → log(1/δ). x ∼ 1 − 2e−2t + δ 2 − 4 128 (12. we don’t really want to know the exact expression for x2 . just its behaviour as t → ∞. dt 32 and hence from solving this simple linear equation. 12. Although x approaches unity as t → ∞. . Fortunately.22) 1 2t e as t → ∞. we ﬁnd that dx2 1 + 2x2 ∼ − e2t as t → ∞. x2 ∼ − This shows that 1 1 4 2t δ e + O(δ 6 ) as t → ∞. when x = 1 + O(δ 2 ).6. Solving this problem would be extremely tedious. The Jacobian elliptic function sn(t. so we need to go to O(δ 4 ). Using the known behaviour of the various hyperbolic functions. At this order.316 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS Fig. k) for various values of δ.

Figure 12.12. ’--’) xlabel(’log_{10}\delta’).Texact. Tasymp = 4*log(4.22). Comparing this with (12. and the solution is 1 (12. 8 4 for t = O(1). which gives 1 x = 1 − δ 2 {cosh (4 log 2 − 2T ) − 1} + O(δ 4 ) (12. The agreement is very good for all δ 1. end plot(log10(d).21)./d). at leading order.23). with the outer solution. k → 1− .k)]. Writing the inner expansion in terms of the outer variables and retaining terms up to O(δ 2 ) leads to 1 1 x = 1 − eK e−2t + δ 2 + O(δ 4 ). dX = − 2X (2X + 1).’off’. This is the quarter period of the solution.’TolX’. δ On substituting these inner variables into (12. we ﬁnd that.Tasymp. so the period τ satisﬁes 1 1 τ = log + T0 . for del = d k = sqrt(1-del^2).24) 4 when T = O(1).^(-7:0. 10^-10).options. 4 δ and hence τ = 4 log 4 δ + O(δ 2 ).T2. whose behaviour as x → 1 is given by (12. t = log + T. dT ¯ Using the substitution X = X − 1 brings this separable equation into a standard 4 form. ylabel(’T’) legend(’exact’. T2 = 2*log(4/del).23) X = {cosh (K − 2T ) − 1} . We produced this ﬁgure using the MATLAB script ' $Texact = []. 8 and hence K = 4 log 2. (12. but only logarithmically fast.’asymptotic’) & % . for δ 1.7 shows a comparison between the exact and analytical solutions.25:0). We conclude that the period of the solution does indeed tend to inﬁnity as δ → 0. x = 1 when T = T0 = 2 log 2 + O(δ 2 ).22) shows that we need 1 eK e−2t = 2e−2t .2 ORDINARY DIFFERENTIAL EQUATIONS 317 We therefore deﬁne new variables for an inner region close to x = 1 as 1 x = 1 − δ 2 X. Texact = [Texact 2*fzero(@ellipj.log10(d). 4 We now need to determine the constant K by matching the inner solution. From this leading order approximation. options = optimset(’Display’. d = 10. For a suﬃciently large volume of material.318 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS This uses the MATLAB function fzero to ﬁnd where the elliptic function is zero. This positive feedback mechanism can lead to the unwanted. it releases heat more rapidly. This decomposition is usually more rapid the higher the temperature. and the material ignites. which controls the details of its execution. As a material gets hotter. This leads to the possibility of thermal ignition. ranging from packets of powdered milk to haystacks. heat production dominates heat loss. The most common physical mechanism that can break this feedback loop is the diﬀusion of heat through the material and out through its surface. using the asymptotic expression as an initial estimate. Determining the critical temperature . and so on. The rate of heat production due to decomposition is proportional to the volume of the material and the rate of heat loss from its surface proportional to surface area. In this case. A comparison of the exact period of the elliptic function sn(t.7. we can obtain a composite expansion. Note that the function optimset allows us to create a variable options that we can pass to fzero as a parameter. and set the convergence tolerance to 10−10 . uniformly 4 valid for 0 t 1 τ = log(4/δ) + O(δ 2 ). Example 2: A thermal ignition problem Many materials decompose to produce heat. combustion of many common materials. 1 + 1 δ 2 − 2e−2t . by adding the solutions in the inner and outer regions and subtracting the matched part. which heats it more rapidly. as 4 1 x = tanh t + 2e−2t + δ 2 tanh t − t sech2 t − cosh log 4 16 δ2 − 2t + O(δ 4 ). and potentially disastrous. Fig. k) for k = √ 1 − δ2 . we turn oﬀ the output of intermediate results. 12. Finally. which governs the ﬂow of heat in a body. per unit volume. y.12. The Arrhenius law can be derived from ﬁrst principles using statistical mechanics.8 shows that the reaction rate is zero at absolute zero (T = 0). † For more background on combustion theory. for example.25) The rate of combustion of the material can be modelled using the Arrhenius law. R = Ae−Ta /T .1. see Buckmaster and Ludford (1982). δz → 0.12). although we will not attempt this here (see. We now need to include the eﬀect of a chemical reaction that produces R(x. 12.2 ORDINARY DIFFERENTIAL EQUATIONS 319 below which it is safe to store a potentially combustible material is an important and diﬃcult problem. ∂t and hence.6. per unit time. also a constant. we showed how to derive the diﬀusion equation. by adding a term R δt δx δy δz to the right hand side of (2. for a steady state solution (∂/∂t = 0). k∇2 T + R = 0. The Arrhenius reaction rate law. (2. z.8. After deﬁning u = T /Ta and rescaling distances Fig. On taking the limit δt. In Section 2. 1970). Flowers and Mendoza. we arrive at ρc ∂T = −∇ · Q + R. t) units of heat. and since Fourier’s law of heat conduction is Q = −k∇T . when it increases. with R → A as T → ∞.11). Figure 12. . δy. where A is a constant and Ta is the activation temperature. and remains small until T approaches the activation temperature. (12. Ta . δx. It is important to note that T is the absolute temperature here.† We now want to develop a mathematical model of this problem. . A small trick that makes the study of this problem easier is to replace (12. ua is the dimensionless absolute temperature of the surroundings and ra the dimensionless radius of the sphere. 1. we arrive at the nonlinear partial diﬀerential equation ∇2 u + e−1/u = 0.26) as du 1 =− 2 dr r r 0 s2 e−1/u(s) ds < 0.26) and (12. + dr2 r dr subject to du = 0 at r = 0.28) for a given value of . u = dr at r = 0. which was ﬁrst studied by Gel’fand (1963). we would expect that the larger ra .320 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS with k/A. Note that. the dimensionless temperature at the centre of the sphere. as we discussed earlier. Our task is therefore to construct an asymptotic solution of the initial value problem given by (12. Note that by using the integrating factor r2 . we can conclude that the material will ignite. and hence conclude that u is a monotone decreasing function of r. The temperature of the sphere decreases from centre to surface. the smaller ua must be to prevent the ignition of the sphere. (12. see Billingham (2000). we can write (12. u = ua at r = ra .26) Here. we deﬁne a new variable u = u/ .† A positive solution of this boundary value problem represents a possible steady state solution in which these two physical processes are in balance. u = 1 at r = 0.27) with du = 0. † For all the technical details of this problem.26) and (12.28) become d 2 u 2 dˆ 1 1 u ˆ + exp − + dr2 r dr u ˆ subject to dˆ u = 0. and then determine the corresponding value of ua = u(ra ). For a uniform sphere of combustible material with a spherically-symmetric temperature distribution. ˆ dr = 0.26) and (12. If no such steady state solution exists. Asymptotic analysis: Region I Since u = at r = 0.27) (12.28) We can then solve the initial value problem given by (12. dr (12. with u = O(1) for ˆ ˆ In terms of this variable.28) when 1. this becomes d 2 u 2 du + e−1/u = 0. (12. ˆ ˆ ˆ Now. φ2 = e−1/ φ1 = 1 3 e−2/ . + 2 dr r dr subject to dˆ1 u = u1 = 0 at r = 0. note that 1 e−1/ u ˆ ∼ ∼ 1 1 −1 exp − (1 + φ1 u1 ) ˆ e−1/ exp φ1 ∼ 1 ∼ e−1/ 1 1 exp − (1 − φ1 u1 ) ˆ φ1 1 u1 ˆ 1+ u1 . ˆ dr and hence u2 = ˆ 1 4 120 r .31) This means that u∼1− ˆ 1 −1/ 2 1 e r + e−2/ r4 6 120 3 (12. ˆ dr Using the integrating factor r2 . We’re going to need more than this to be able ˆ to proceed. ˆ (12. (12.32) . + dr2 r dr 6 subject to dˆ2 u = u2 = 0 at r = 0.26) then shows that + φ2 d 2 u2 u ˆ 2 dˆ2 + 2 dr r dr 1 2 d 2 u1 u ˆ 2 dˆ1 + 2 dr r dr 1 ∼ − e−1/ 1+ φ1 u1 . Firstly.12. 1 e−1/ . which we can check later. so let’s look for an asymptotic expansion of the form u = 1 + φ1 ( )ˆ1 + φ2 ( )ˆ2 + · · · . Equation (12. u ˆ 2 dˆ1 d 2 u1 = −1. (12.29) + 2 dr r dr which has the trivial solution u = 1. we ﬁnd that the solution is u1 = − 1 r2 . we therefore take φ1 = and hence expand 1 1 u = 1 + e−1/ u1 + 3 e−2/ u2 + · · · .2 ORDINARY DIFFERENTIAL EQUATIONS 321 At leading order. ˆ u u φ1 1 are to be determined. using (12. ˆ provided that φ1 φ1 . we need a three-term where φ2 asymptotic expansion to be able to determine the scalings for the next asymptotic region. As we shall see. Similarly. u ˆ d 2 u 2 dˆ = 0. ˆ 6 1 d 2 u2 u ˆ 2 dˆ2 = r2 .30).30) In order to obtain a balance of terms. 36) shows that we are interested (12. R → e−k/2 R. with U = O(1) and R = O(1) for 1.33) is nonlinear and nonautonomous. We can therefore make the nonlinear transformation dU . Since (12. After some thought. As r increases. 2) and P2 = (2. ds ds subject to 1 p ∼ e−2s . There are two ﬁnite equilibrium points.31). 6 (12. = p + q − 2. Note that this is why we needed a three-term expansion to work out the scalings for the next asymptotic region. Asymptotic analysis: Region II In this region we deﬁne scaled variables u = + 2 U . + 2 dR R dR to be solved subject to the matching condition 1 U ∼ − R2 as R → 0.31). the ﬁrst comparable to the second. do not depend at all on the functional ˆ form of the term e−1/ u .36) 3 The problem is still nonlinear. (12.26) becomes d 2U 2 dU + eU = 0. This is a sign that we need another term in the expansion in order to capture the eﬀect of the only nonlinearity in the problem. 0) in the (p.29) or (12. However. we ﬁnd that P1 is a saddle point and P2 is an unstable.33) is invariant under the transformation U → U + k. At leading order. which usually means that we must resort to ﬁnding a numerical solution. (12. q) phase plane. at P1 = (0.33) (12.322 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS for 1. since for suﬃciently large r the second term is comparable to the ﬁrst. but is now autonomous. Equation (12.35) . you will ﬁnd that it is impossible to match the solutions in the new region to the solutions in region I. After determining the Jacobian at each of these points and calculating the eigenvalues in the usual way.34) Equation (12. q ∼ 2 − e−2s as s → ∞.34) become dq dp = −pq. p(s) = e−2s eU .33) and (12. (12. q(s) = 2 + e−s dR after which (12. you notice that the equations at O(1) and O( 1 e−1/ ) in region I. and the third is 1/2 1/2 e . when r = O( 1/2 e1/2 ). This expansion is not uniformly valid. r = e1/2 R. R = e−s . What would have happened if we had taken only a two-term expansion and looked for a new asymptotic region with r = O( 1/2 e1/2 ) and u = O( )? If you try this. when r = O( e1/2 ) nonuniformity to occur is therefore when r = O( e1/2 ) and u = + O( 2 ). we have seen in Chapter 10 that we can often make progress using group theoretical methods.29) and (12. anticlockwise spiral. which could be replaced by e−1/ without aﬀecting (12. so we can use the phase plane methods that we studied in Chapter 9. (12. and 2 therefore √ 7 s/2 s+B as s → −∞. The behaviour of the solution of (12. Since U = 2s + log p and s = − log R. The proof that this is the case is Exercise 9. The eigenvalues of P2 are 1 (1 ± i 7). Fig. Since the solution asymptotes to P2 . for which we calculated the solution numerically using MATLAB (see Section 9.35) subject to (12.3.4). When log R = O( −1 ).12. for 1. this solution must be represented by one of the stable separatrices of P1 . 12. this shows that √ A 7 U ∼ −2 log R + log 2 − √ sin log R − B as R → ∞. p ∼ 2 + Ae sin 2 for some constants A and B. the second term is comparable to the ﬁrst term.9. . What happens to this separatrix as s → −∞? A sensible. and we have a further nonuniformity. the unique integral path that represents the solution is the stable separatrix of P1 that lies in the half plane p > 0. q)-phase plane. which comes with some hints. and correct.17. as shown in Figure 12.37) 2 2 R We conclude that u∼ + 2 (log 2 − 2 log R) as R → ∞. we can determine its behaviour as √ → −∞ s by considering the local solution there.9. since p = e−2s eU > 0. Furthermore.2 ORDINARY DIFFERENTIAL EQUATIONS 323 in a solution that asymptotes to P1 as s → ∞. guess would be that it asymptotes to the unstable spiral at P2 .36) in the (p. In terms of these variables. as shown in Figure 12. as can be seen in Figure 12.39) At leading order. (12. and the solution is complete. If ra = O( e1/2 ).39). the solution in region III shows that ua ∼ 1 2 + 2 log(ra / ) < uamax . namely uamax ∼ + 2 log pmax 2 e1/ 2 ra . so that We can now determine ua = u(ra ). 1 + 2S (12.10. the fact that p < pmax ≈ 3. which also satisﬁes (12. To ﬁnd the matching conditions.322. U1 ∼ log 2 as S → 0. which gives ¯ ¯ U0 ∼ 1 − 2S.38) + exp + − ¯ + 1 + 2S 2 dS dS U ¯ We expand U as ¯ ¯ ¯ U = U0 + U1 + · · · . . 1 . with u → 0 as S → ∞. leads to a maximum value of ua for which a steady state solution is available. r = ra lies in region I. the solution is given by the exponential term in (12. we need to consider the solution in region II. we write expansion (12.26) becomes ¯ ¯ 1 dU d 2U 1 = 0. dS and hence that ¯ U1 = 1 log (1 + 2S)2 2 2 (1 + 2S)2 . for ua suﬃciently small that ra ua e1/2ua . The oscillations in p lead to oscillations in ua as a function of .37) in region I in terms of the new variables and retain terms up to O( ). In other words.38) as ¯ U0 = which satisﬁes (12. e1/2 . and we predict that the sphere will not ignite. if ra = O( e3/2 ). we ﬁnd that ¯ dU0 2 ¯ + exp (1 + 2S) U1 = 0. In particular.39).40) Finally. and hence R → ∞. If ra 1 −1/ 2 ra ∼ . ua ∼ − 6 e a steady state solution is always available. 2 1/2 S/ e . r= e (12.9. This expansion remains uniform. for S = O(1) and 1. (12.324 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS Asymptotic analysis: Region III ¯ We deﬁne scaled variables u = U and S = log R = − 1 − log + log r. We conclude that u= 1 + 2S + (1 + 2S)2 log 2 (1 + 2S)2 + O( 3 ). with U = O(1) and S = O(1) for 1. At O( ). At O( ). where a dot denotes d/dt.4 The Method of Multiple Scales Let’s now return to solving a simple. 12. subject to y0 (0) = 1.10. Fig.41) for t 0. ua . Since this is an initial value problem. y0 (0) = 0. we expand y = y0 (t) + y1 (t) + O( 2 ) for 1. ¨ ˙ to be solved subject to y(0) = 1. we can think of y developing as a function of time. and hence at which the material will not ignite. At leading order. Consider y + 2 y + y = 0. As usual. as a function of in region II when ra = 109 . constant coeﬃcient ordinary differential equation that. t. which has ¨ ˙ solution y0 = cos t. ˙ (12. ¨ ˙ . 12. linear. y0 + y0 = 0. seems like a regular perturbation problem. The ambient temperature. y(0) = 0.12.2 ORDINARY DIFFERENTIAL EQUATIONS 325 We conclude that uamax (the critical ignition temperature or critical storage temperature) gives an asymptotic approximation to the hottest ambient temperature at which a steady state solution exists.2. y1 + y1 = −2y0 = 2 sin t.42) (12. at ﬁrst sight. let’s examine the exact solution of (12. We conclude −1 . consists of deﬁning a new slow time variable. the solution is a decaying oscillation. which slowly ˙ reduces the amplitude of the oscillation. so that when t = O(1/ ).11. the ratio of the second term to the ﬁrst term in this expansion is asymptotic to t. we need to keep the exponential rather than its Taylor series 1 and t expansion if we are to construct a solution that is valid when t = O( −1 ).41) is an undamped. in its most basic form. This means that y(t) ∼ cos t + (−t cos t + sin t) (12.41) subject to the boundary conditions (12. and is therefore no longer small when t = O( −1 ). two-term asymptotic solution. Figure 12.43). although it correctly captures the fact that e− t ∼ 1 − t for −1 . (12. Fig.44). y1 (0) = 0. with the decay occurring over a timescale of O( −1 ). The term 2 y represents the eﬀect of weak damping. T = O(1).11.326 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS subject to y1 (0) = 0. After noting that the particular integral solution ˙ of this equation takes the form y1p = kt cos t for some constant k. The problem with the asymptotic expansion (12.42).11 shows that the two-term asymptotic expansion. The exact solution.43) is only valid for t To see where the problem lies. rapidly becomes inaccurate once t = O( −1 ).50) of (12. linear oscillator. which is y = e− t cos 1− 2t +√ 1− 2 sin 1− 2t . At leading order. that the asymptotic solution given by (12. (12.1. (12. T = t. 12. and the .44) As we can see from Figure 12.43) as → 0. and oneterm multiple scales solution (12. we ﬁnd that y1 = −t cos t + sin t. As t → ∞. (12. The method of multiple scales.43).41) when = 0.43) is that. (12. ∂y0 (0. As we have seen.12. dt ∂t ∂T (12.48) + y1 = − 2 ∂t2 ∂t∂T ∂t dT dT Because of the appearance on the right hand side of the equation of the terms cos t and sin t. The functions A0 (T ) and B0 (T ) are. 0) + (0. 0) = 0. which are themselves solutions of the homogeneous version of the equation. At leading order. After noting that ∂ ∂ d = + . with each term a function of both t. B0 (0) = 0. 0) = 1.2 ORDINARY DIFFERENTIAL EQUATIONS 327 slow decay can therefore be accounted for. to capture the slow decay. The terms proportional to sin t and cos t in (12. ∂ 2 y0 + y0 = 0. ∂t ∂y ∂y (0. as yet. 0) = 0. the particular integral solution will involve the terms t sin t and t cos t. ∂ 2 y1 ∂y0 dA0 ∂ 2 y0 dB0 −2 =2 + A0 sin t − 2 + B0 cos t. we must remember that the ‘constants’ of integration can actually be functions of the other variable.47) . it is precisely terms of this type that lead to a nonuniformity in the asymptotic expansion when t = O( −1 ). the only derivatives are with respect to t. to capture the oscillation. ∂t2 subject to y0 (0. (12. undetermined.46) 2 ∂2y ∂y ∂y +2 2 + y = 0. 0) = 1. and T . +2 ∂T 2 ∂t ∂T (12. and the boundary conditions show that A0 (0) = 1.45) Although this is a partial diﬀerential equation. At O( ). to keep the asymptotic expansion (12. so we can solve as if it were an ordinary diﬀerential equation in t. T ) + y1 (t.48) are known as secular terms and. ∂t ∂T (12.41) becomes ∂2y ∂2y + +2 2 ∂t ∂t∂T to be solved subject to y(0. T . We then look for an asymptotic solution y = y0 (t. T ) + O( 2 ). This means that y0 = A0 (T ) cos t + B0 (T ) sin t. However. This suggests that we will need to expand up to O( 2 ) to construct a two-term multiple scales expansion.53) We already know that y0 = e−T cos τ. T ) + After noting that d ∂ = + 1+a dt ∂T equation (12.41) becomes ∂2y ∂2y ∂2y + + 2a 2 2 + 2 2 ∂τ ∂τ ∂τ ∂T to be solved subject to y(0. This shows that the phase of the oscillation changes by an O(1) amount when t = O( −2 ). 0) + (0. Although this looks like a bit of a cheat. b. (12. 0) = 0. (12. so the solutions are A0 = e−T and B0 = 0. and remains valid when t = O( −1 ). We conclude that y ∼ e−T cos t = e− t cos t (12. We know that 1 1 (12.50) for 1. with y = y0 (τ. we needed to consider the solution up to O( ). 3 +b + ··· ∂ . this approach works for a wide range of problems. The initial conditions for (12. to be determined. we must eliminate them by choosing A0 and B0 appropriately.52) +b 3 + ··· (12. This asymptotic solution is consistent with the exact solution.328 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS uniform. we can take the exact solution. we need dB0 dA0 + A0 = 0. + B0 = 0.47). . T ) + y1 (τ.11.49) are given by (12. 0) = 1. as can be seen in Figure 12.44). In order to develop a one-term multiple scale expansion. In order to proceed to ﬁnd more terms in the asymptotic expansion using the method of multiple scales. we will −2 see below that this solution is only valid for t . 2 2 for 1. we seek a solution that is a function of the two timescales T = t. T ) + O( 3 ). including nonlinear diﬀerential equations. since we are only doing this because we know the exact solution. ∂τ ∂T (12.44). 1+a 2 2 2 2 y2 (τ. . . with the constants a.49) dT dT and hence y1 = A1 (T ) cos t + B1 (T ) sin t. ∂τ ∂2y ∂y ∂y +2 2 + y + O( 3 ) = 0. +2 2 ∂T ∂τ ∂T ∂y ∂y (0. as a guide. In this example. In fact.51) y ∼ e− t cos 1 − 2 t + sin 1 − 2 t . τ = 1 + a 2 +b 3 + · · · t. (12. In order to capture this. . y1 = A1 (T ) cos τ + B1 (T ) sin τ. which leads to a nonzero right hand side in (12. B1 = This gives us y ∼ e−T cos τ + a+ 1 2 T e−T + e−T sin τ. ∂y0 ∂y1 (0. and study the chaotic response of the circuit. and we must therefore remove it by choosing a = −1/2. For 1 this is a linear oscillator with a weak nonlinear term. It is straightforward (at least if you’re an electrical engineer!) to build an electronic circuit from resistors and capacitors whose output is governed by (12. We could have deduced this directly from the diﬀerential equation for B1 .12. We conclude that τ= 1− 1 2 2 + ··· t. ˙ For |y| < 1 this term tends to drive the oscillations to a greater amplitude. whilst for |y| > 1.2 ORDINARY DIFFERENTIAL EQUATIONS 329 At O( 2 ). Example 1: The van der Pol Oscillator The governing equation for the van der Pol oscillator is d 2y dy + y = 0.51). since the term proportional to e−T is secular. the part of the O( ) term proportional to T will lead to a nonuniformity in the expansion when T = O( −1 ). 0) = A1 (0) = 0. However. ∂ 2 y2 ∂ 2 y0 ∂y0 ∂ 2 y0 ∂ 2 y1 ∂y1 + −2 + y2 = −2a 2 − 2 −2 2 2 ∂τ ∂τ ∂τ ∂T ∂T ∂τ ∂T =2 dA1 + A1 sin τ − 2 dt dB1 1 + B1 − a + dt 2 1 2 e−T cos τ.54). this term damps the oscillations. as expected. ∂τ ∂T a+ e−T . 0) = B1 (0) − 1 = 0. It is also straightforward to construct a forced van der Pol oscillator. a reasonable guess at the behaviour of the solution for 1 would be that there is an oscillatory solution with unit .54). Since the damping is zero when y = 1. + (y 2 − 1) dt2 dt (12. and hence A1 = 0. + B1 = dt dt At O( ) the boundary conditions are y1 (0.54) for t 0. and hence obtain (12. (y 2 −1)y. It was in this context that this system was ﬁrst studied extensively as a prototypical nonlinear oscillator. 0) + (0. a+ 1 2 T e−T + e−T . In order to remove the secular terms we need dA1 dB1 + A1 = 0. We seek an asymptotic solution of the form y = y0 (t. . as shown in Figure 12. As for our linear example. 2. . A0 (T ). † To get cosn θ in terms of cos mθ for m = 1. Therefore y ∼ 2(1 + 3e−T )−1/2 cos t for 1. . φ0 (T ).56) In order to pick out the secular terms on the right hand side. and phase.12.55) We will use the method of multiple scales. dT 4 0 dT 4 0 To suppress the secular terms we therefore need dφ0 = 0. we note that† sin θ cos2 θ = sin θ − sin3 θ = sin θ − This means that ∂ 2 y1 + y1 = ∂t2 2 1 dφ0 dA0 1 + A3 − A0 sin(t + φ0 ) + 2A0 cos(t + φ0 ) + A3 sin 3(t + φ0 ). . φ0 (0) = 0. The boundary conditions show that A0 (0) = 1. ∂ 2 y1 ∂ 2 y0 ∂y0 2 − (y0 − 1) + y1 = −2 ∂t2 ∂t∂T ∂t =2 dA0 dφ0 sin(t + φ0 ) + 2A0 cos(t + φ0 ) + A0 sin(t + φ0 ) A2 cos2 (t + φ0 ) − 1 .330 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS amplitude. but incorrect. 0 dT dT 1 1 1 3 sin θ + sin 3θ = sin θ + sin 3θ. dT dA0 1 = A0 (4 − A2 ). n. T ) + y1 (t. 4 4 4 4 (12. it is more convenient to write the solution in terms of an amplitude. T ) + O( 2 ). . dy (0) = 0.56). and deﬁne a slow time scale. At O( ). and we conclude that the amplitude of the oscillation actually tends to 2 as t → ∞. use einθ = cos nθ + i sin nθ = (eiθ )n = (cos θ + i sin θ)n and equate real and imaginary parts. we ﬁnd that y0 = A0 (T ) cos {t + φ0 (T )} . guess by considering the solution with initial conditions y(0) = 1. T = t. dt (12. Let’s investigate this plausible. For this problem. the solutions are φ0 = 0. 0 dT 8 Subject to the boundary conditions (12. A0 = 2(1 + 3e−T )−1/2 . 12. subject to y(0) = 1. when = 0.12. dt2 (12. At leading order ∂ 2 x0 + x0 = 0. which satisﬁes (12. As usual. Let’s now use the method of multiple scales to see how this small perturbation aﬀects the oscillation after a long time.57) The initial conditions.21). this function is oscillatory and. T ) + k 2 x1 (t.2 ORDINARY DIFFERENTIAL EQUATIONS 331 Fig. at leading order. We can see this more clearly by diﬀerentiating (12.58) . T ) + O(k 4 ). y(0) = 0. ∂t2 (12. of which there must be two for this second order equation.54). dx/dt = 1 when t = 0. 12. in terms of which (12. we deﬁne T = k 2 t and x = x(t.21) and eliminating dx/dt to obtain d 2x + (1 + k 2 )x = 2k 2 x3 .21). (12. k).1. performs simple harmonic motion. When k 1. ˙ Example 2: Jacobian elliptic functions with almost simple harmonic behaviour Let’s again turn our attention to the Jacobian elliptic function x = sn(t. The leading order solution of the van der Pol equation.57) becomes ∂2x ∂2x ∂2x + k 4 2 + (1 + k 2 )x = 2k 2 x3 . + 2k 2 2 ∂t ∂t∂T ∂ T We seek an asymptotic solution x = x0 (t. T ). from (12. are x = 0 and. we must set the coeﬃcients of sin(t + φ) and cos(t + φ) to zero. As we saw in Chapter 9. in contrast to the solution of the van der Pol equation that we studied earlier. dy d 2y + y − y 3 = 0. which gives 1 A = 1. This has solution x0 = A(T ) sin {t + φ(t)} . is appropriate for weakly perturbed linear oscillators. If we take the analysis to O(k 4 ). we ﬁnd that the amplitude of the oscillation is also dependent on k (see King. Let’s begin by considering the leading order problem. 4 for k 1.59). Can we make any progress if the leading order problem is nonlinear? We will concentrate on the simple example. 0) = 1. with with dy (0) = v0 > 0. φ(0) = 0. as we would expect from the analysis given in Section 9. = 0. +2 dt2 dt subject to y(0) = 0.4. However.332 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS subject to x0 (0.60) when . the leading order amplitude of the oscillation does not change with t.60) small and positive. since dy/dt does not appear in (12. and A(0) = 1. ∂ 2 x1 ∂ 2 x0 + x1 = 2x3 − x0 − 2 0 ∂t2 ∂t∂T = 2A3 sin3 (t + φ) − A sin(t + φ) − 2 3 3 dφ A − A + 2A 2 dT dφ dA cos(t + φ) + 2A sin(t + φ) dT dT 1 dA cos(t + φ) − A3 sin 3(t + φ). We can see that. 4 and hence x = sin 1 1 − k 2 t + O(k 2 ). the period of the oscillation changes by O(k 2 ) even at this lowest order. as we have described it above. dT 2 (12. 12.2. φ = − T.61) (12. 0) = 0 and ∂x0 ∂t (0. At O(k 2 ).59) = sin(t + φ) − 2 In order to remove the secular terms. 1988). This gives us two simple ordinary diﬀerential equations to be solved subject to (12.5 Slowly Damped Nonlinear Oscillations: Kuzmak’s Method The method of multiple scales. dt (12. we begin in the usual way by deﬁning a slow time scale.12. This was ﬁrst done by Kuzmak (1959). T = t.60) is negligible when v0 . . and possibly some change in its phase.63) √ 1 − 1 − 2E √ 1 + 1 − 2E .4. we can see that θ (T ) ≡ ω(T ) is the frequency of the oscillation at leading order and φ(T ) the change in phase. For small and positive.62) becomes dˆ y =± ˆ dt where k= 1 − y2 ˆ 1 − k2 y2 .2 ORDINARY DIFFERENTIAL EQUATIONS 333 = 0. (12. (12. However. ˆ 1/2 (12. we ﬁnd that y = 1− √ 1 − 2E 1/2 sn E √ 1 − 1 − 2E 1/2 t. we can integrate once to obtain 1 dy = ± E − y2 + y4 . Example 2 in the previous section 1. This is to be expected.62) 2 2 with E = v0 from the initial conditions. y varies periodically in a manner described by the Jacobian elliptic function sn. If we now assume that v0 < 1/2†. we expect that weak damping leads to a slow decrease in the amplitude of the oscillation. k . since the nonlinear shows that y ∼ v0 sin t when v0 term in (12. but nonzero. is small. so it is now convenient to deﬁne ψ= θ(T ) + φ(T ). when the amplitude and phase of the oscillation have changed signiﬁcantly. although not in full generality.61). and hence y. and scale y and t using √ 1/2 1/2 √ 1 − 1 − 2E ˆ y. In order to construct an asymptotic solution valid for t = O( −1 ). T ).65) In the absence of any damping ( = 0). y = 1 − 1 − 2E E (12. In addition. (12. the frequency of the leading order solution depends upon the amplitude of the oscillation. by analogy with what we found using the method of multiple scales in the previous section. both of which we must † The usual graphical argument shows that this is a suﬃcient condition for the solution to be periodic in t. t = ˆ t. dt ∂ψ ∂T where a prime denotes d/dT . θ(0) = 0. dt 2 (12.63) with the system that we studied in Section 9. and seek a solution y ≡ y(ψ. Since dy ∂y ∂y = (θ + φ ) + . for a nonlinear oscillator.64) If we compare (12. 72) 1 √ 1− s2 ds √ 1 − k 2 s2 (12. T ) + O( 2 ). T ) + y1 (ψ. (12.69). ∂ψ ∂T ∂y ∂y (0.66) ∂y ∂y +2 2 + y − y 3 = 0. shows that P (T ) = 4ω(T ) where K(k) = 0 1− 1 − 2E(T ) E(T ) 1/2 K(k(E)).61) become (ω + φ ) 2 ∂2y ∂2y + + 2 (ω + φ ) 2 ∂ψ ∂T ∂ψ +2 (ω + φ ) 2 ∂2y ∂y + (ω + φ ) 2 ∂T ∂ψ (12.73) is the complete elliptic integral of the ﬁrst kind. (12. show that 2 E(0) = v0 . At leading order. 0) = 0.66) and (12. k(T ) .334 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS determine as part of the solution. ∂ψ 2 (12. (12. 0) + (0. The initial conditions. can be determined by noting that the quarter period is √ 1/2 1− 1−2E(T ) ds 1 P (T ) = ω(T ) . we need to go to O( ) to determine the . (12. 0) = v0 . this has solution   1/2 E(T ) y0 = 1 − 1 − 2E(T ) sn  1 − 1 − 2E(T ) where k is given by (12. In terms of these new variables. we obtain ω 2 (T ) subject to y0 (0. (12. (ω(0) + φ (0)) We now expand y = y0 (ψ.71) The period of the oscillation. 4 0 E(T ) − s2 + 1 s4 2 which. and substitute into (12. 0) = v0 . As in the method of multiple scales. 0) = 0.67). φ(0) = 0.60) and (12.70)  ω(T ) As we have seen. ∂ψ 1/2 (12.67) ∂ 2 y0 3 + y0 − y0 = 0. ω(0) ∂y0 (0. after a simple rescaling.69)   ψ . ∂ψ ∂T subject to y(0.64).68) (12. P (T ). 75) which we write as L(y1 ) = R1odd + R1even . note that we have three unknowns to determine. whilst ∂y0 /∂ψ is an even function of ψ. Firstly. we can now treat y0 as a function of ψ and E. whilst for the method of multiple scales.77) (12. so that ∂y0 /∂ψ is a solution of the homogeneous version of (12. In addition. . so that 1 dθ = ω ≡ ω(E) = dT 4K(k(E)) E √ 1 − 1 − 2E 1/2 .76) Now. + 1 − 3y0 y1 = −2ωφ − 2ω ∂ψ 2 ∂ψ 2 ∂ψ∂E dT ∂ψ ∂ψ √ 1 − 2E 1/2 sn {4K(k(E))ψ . For convenience.2 ORDINARY DIFFERENTIAL EQUATIONS 335 behaviour of the solution on the slow time scale. we have the freedom to choose this new time scale so that the period of the oscillation is constant. and therefore orthogonal to ∂y0 /∂ψ. . by diﬀerentiating (12. equivalent to E(T ) and φ(T ).76) to be periodic. E(T ). Since ∂y0 /∂ψ is even in ψ. Both y0 and ∂ 2 y0 /∂ψ 2 are odd functions.78) ∂2 2 + 1 − 3y0 . R1even = −2ω 2 ∂ψ ∂ψ∂E dT ∂ψ ∂ψ (12. φ(T ) and ω(T ).† This is equivalent to the elimination of secular terms in the method of multiple scales. the right hand side must be orthogonal to the solution of the homogeneous equation. we take P = 1. ∂ψ which is the equivalent of the secularity condition that determines the amplitude † Strictly speaking. the theory of linear ordinary diﬀerential equations with periodic coeﬃcients. with y0 (ψ.74) We also need to note for future reference the parity with respect to ψ of y0 and its derivatives. E) = 1 − At O( ) we obtain ω2 ∂y0 ∂ 2 y1 ∂ 2 y0 ∂y0 ∂ 2 y0 dE 2 −ω − 2ω . ∂ψ 2 (12.68) with respect to ψ. k(E)} . this is a result that arises from Floquet theory. this means that 1 0 ∂y0 R1even dψ = 0. (12. (12. Since we introduced θ(T ) simply to account for the fact that the period of the oscillation changes with the amplitude. For the solution of (12. we had just two. T . where L = ω2 and R1odd = −2ωφ ∂y0 ∂ 2 y0 ∂y0 ∂ 2 y0 dE −ω − 2ω .76).12. we ﬁnd that L ∂y0 ∂ψ = 0. using the results of Section 9. we ﬁnd that 1 0 ∂y0 ∂ψ 2 dψ = 16K 2 (E) 1 − √ 1 1 − 2E 0 1 − sn2 (4Kψ . k) dψ = where 1 3k 2 1 + k 2 L(k) − 1 − k 2 K(k) . k) dψ. k) We now need a standard result on elliptic functions‡. 2ω (12. ‡ See. . 1 L(k) = 0 1 − k 2 s2 ds. the symbol E is already spoken for in this analysis. 1 − s2 is the complete elliptic integral of the second kind§. Using (12. = 16K(E) 1 − √ 1 − 2E 0 1 − sn2 (ψ .74). 2 1/2 ω 0 ∂y0 ∂ψ (1− dψ =2 0 √ 1−2E ) dy0 E− 2 y0 = + 1 4 2 y0 1 .78). k) 0 1 − k 2 sn2 (ψ .79) To proceed. k) K 1 − k 2 sn2 (4Kψ . (12.336 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS of the oscillation in the method of multiple scales.4. for example. we ﬁrstly note that 1 0 ∂y0 ∂ψ 2 (1− dψ = 4 0 √ 1−2E ) 1/2 ∂y0 dy0 ∂ψ 4 = ω and hence that ∂ ∂E 1 2 (1− 0 √ 1−2E ) 1/2 1 4 2 E − y0 + y0 dy0 . k) dψ 1 − k 2 sn2 (ψ . (12. then show that 4E dE =− 2 dT 3k K 1 + k 2 L (k) − 1 − k 2 K(k) . (12. Byrd (1971). we can write this as† d dE 1 ω 0 ∂y0 ∂ψ 2 dψ dE + 2ω dT 1 0 ∂y0 ∂ψ 2 dψ = 0. § Although the usual notation for the complete elliptic integral of the second kind is E(k).80) Secondly.81) † Note that the quantity in the curly brackets in (12. namely that K 1 − sn2 (ψ . Equation (12.79) is often referred to as the action.79) and the deﬁnition of ω. Figure 12. E = E(0)e−2T . (12.35 0. this reduces to the multiple scales result. dE/dT = −2E (see Exercise 12.12.1 0. 0. For v0 1. Unfortunately.15 0.4 0. along with the initial condition. We now need to ﬁnd an equation that determines the phase.5 T 2 2. Note that we can simultaneously calculate θ(T ) numerically by solving (12.81) when E(0) = 0.68) with respect to E.13 shows the function E(T ) calculated for E(0) = 0.25 E(T) 0. It is straightforward to integrate (12. we need to determine the solution at O( ) in order to make progress. dE ∂ψ 2 .82) ∂y0 ∂E = −2ω dω ∂ 2 y0 . since the complete elliptic integrals of the ﬁrst and second kinds can be calculated using the built-in function ellipke.81) using MATLAB. By diﬀerentiating (12.05 0 0 0.5 1 1.13. and also the corresponding result using multiple scales on the linearized problem.5 3 Fig.74) subject to θ(0) = 0.10). unlike the method of multiple scales.71). we ﬁnd that L We also note that L ψ ∂y0 ∂ψ = 2ω 2 ∂ 2 y0 . determines E(T ). 12.45 Kuzmak Linearized multiple scales 0. and the corresponding multiple scales solution of the linearized problem.2 0. The solution of (12.45.3 0. ∂ψ 2 (12.45. φ(T ). E(T ) = E(0)e−2T .2 ORDINARY DIFFERENTIAL EQUATIONS 337 This equation. 83) The general solution of the homogeneous version of (12. ∂ψ 2 where y1even is the part of the particular integral generated by R1even . and hence from (12. ω ∂ψ (12.75) that dK ∂y0 ∂y0 ∼4 ψ as ψ → ∞.76) is therefore a linear combination of the even function ∂y0 /∂ψ and the odd function y0odd .338 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS and hence that L (y0odd ) = 0. Noting from (12. this reproduces (12. To ﬁnd the particular integral solution of (12. we can show that 1 0 φ ∂y0 ∂y1 dψ = ∂ψ ∂ψ dω/dE 1 0 φ ∂ ∂y0 ∂ 2 y0 dψ = ∂ψ ∂ψ∂E 2dω/dE ∂E 1 0 ∂y0 ∂ψ 2 dψ.79). we note from (12. 4ωdK/dE + dω/dE (12.86) that d dT e2T ∂ φ dω/dE ∂E 1 ω 0 ∂y0 ∂ψ 2 dψ = 0. whilst at O( ) we ﬁnd that d dT 1 e2T 2ω 0 ∂y0 ∂y1 dψ + φ ∂ψ ∂ψ 1 0 ∂y0 ∂ψ 2 dψ = 0. (12.82) that L − We therefore have y1 = A(T ) ∂y0 dω ∂y0 ∂y0 + B(T ) ω + ψ ∂ψ ∂E dE ∂ψ − φ ∂y0 ψ + y1even . the coeﬃcient of ψ must be zero. For y1 to be bounded as ψ → ∞. and integrate over one period of the oscillation.66) by ∂y/∂ψ. and is itself even in ψ. where y0odd = ω dω ∂y0 ∂y0 + ψ .76). ∂E dE ∂ψ (12. At leading order. After taking into account the parity of the components of each integrand. using what we know about the parity of the various components of y1 .86) Now. ∂E dE ∂ψ this means that B(T ) = φ /ω . we ﬁnd that d dT 1 e2T (ω + φ ) 0 ∂y ∂ψ 2 dψ = 0. .84) φ ∂y0 ψ ω ∂ψ = −2ωφ ∂ 2 y0 = R1odd .85) The easiest way to proceed is to multiply (12. 0) = −E (0) (0. E(0)). Although we know that φ(0) = 0.6 The Eﬀect of Fine Scale Structure on Reaction–Diﬀusion Processes Consider the two-point boundary value problem d dx D x.2 ORDINARY DIFFERENTIAL EQUATIONS 339 Using (12. 0) = 0. 0) ∂ψ ∂ψ + B(0) dω φ (0) (E(0)) − dE ω(E(0)) = −E (0) ∂y0 ∂y0 (0.80). E(0)) + B(0)ω(E(0)) 2 ∂ψ ∂ψ∂E ∂y0 ∂y1even (0. x = 0 for 0 < x < 1.87) This is a second order equation for φ(T ). Although the multiple scales solution gives a good estimate of E(T ). this becomes d dT e2T φ ωdω/dE = 0. are y1 (0.87) we also need to know φ (0). the initial conditions.90) dx We can think of θ(x) as a dimensionless temperature.88) By substituting the solution (12. so that this boundary value problem models the steady distribution of heat in a conducting material.84) for y1 into the second of these. When 1. (12. the leading order multiple scales solution of the linearized √ problem. To see how the method works in general for damped nonlinear oscillators. E(0)) − φ (0) (0. and therefore vanish when ψ = 0. At O( ). (12.67). E(0)) + (0. and .87) that φ(T ) = 0. all of the terms that do not involve φ (0) are odd in ψ.12.2. x. x dθ dx + R θ. and hence from (12.89) subject to the boundary conditions dθ = 0 at x = 0 and x = 1. The numerical and Kuzmak solutions are indistinguishable. (12. this material has a ﬁne scale structure varying on a length scale of O( ). as shown in ﬁgure 12.13. E(0)). and the numerical solution of the full problem when = 0. E(0)) − φ (0) (0. ∂E ∂ψ Using (12. ∂ψ ∂E ∂ψ (12. We conclude that φ (0) = 0. we ﬁnd that ω(E(0)) A(0) ∂ 2 y0 ∂ 2 y0 (0. y = v0 e−T sin t. Figure 12. E(0)) (0. the interested reader is referred to Bourland and Haberman (1988).85). to be able to solve (12. it does not give an accurate solution of the full problem.14 shows a comparison between the leading order solution computed using Kuzmak’s method. (12.01. 12. ω(E(0)) ∂y1 ∂y0 ∂y0 (0. 2 1.6 1. R (see Section 12.4 −0. In order to use asymptotic methods to determine the solution at leading order when 1. If we integrate (12. a coarse. Physically. background structure varying on a length scale of O(1).6 0.2.3.8 1 T 1.90) states that no heat can escape from the ends of the material.91) for a solution to exist.2 −0. Example 2 for a derivation of this type of balance law). 12. (12. (Dθx )x (where subscripts denote derivatives).90).6 0.8 0 0.4 0.4 0. and its production by some chemical reaction.01 compared to asymptotic solutions calculated using the leading order Kuzmak solution and the leading order multiple scales solution of the linearized problem.61) when v0 = 0.60) subject to (12. with sources of heat in some regions of the material balanced by heat sinks in other regions.45 and = 0. x.6 −0.91) simply says that the overall rate at which heat is produced must be zero for a steady state to be possible.4 1. The steady state temperature distribution represents a balance between the diﬀusion of heat.340 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS 0. since (12.14.2 0.8 Numerical Kuzmak Linearized multiple scales 0. we begin by introducing the fast variable. In the usual way ˆ . x = x/ . The numerical solution of (12.8 2 2 Fig. 0 x dx = 0 (12.89) over 0 x 1 and apply the boundary conditions (12. we ﬁnd that the chemical reaction term must satisfy the solvability condition 1 R θ(x).2 0 y −0. s) (12. we ﬁnd that (Dθ1ˆ )x = − (Dθ0ˆ )x − (Dθ0x )x . we .92) θ2 (x. (12.2. s) ∂x 1/ 0 1/ − s ds D(x. ˆ θ0 = α(x) 0 ds + f0 (x). and we will not consider it below. x = 1/ . D ≡ D(x. which gives x ˆ 2 2 {(Dθx )x + R} = 0. s) x ˆ 0 When x is large. with α(0) = 0. there are terms in this expression that are of O(ˆ2 ). x). x). x) + O( 3 ). (12.93) with R ≡ R (θ(x. Since α(0) = 0. x ˆ We can integrate (12. It is quite awkward to apply a boundary ˆ ˆ ˆ condition at x = 1/ ˆ 1. and then once x more. x. x) + At leading order. x ˆ x ˆ which we can write as (Dθ1ˆ + Dθ0x )x = −α (x).94) to obtain Dθ1ˆ + Dθ0x = −α (x)ˆ + β(x). D(x.95) −α(x) ∂ ∂x x ˆ 0 x−s ˆ ds + β(x) D(x. but we shall see that we can determine the equation satisﬁed by θ at leading order without using this. x).96) This is a ﬁrst order ordinary diﬀerential equation for α(x). and become comparable with the leading order term in the expansion for θ when x = 1/ . x x Substituting for θ0 and integrating again shows that x ˆ (12.89) and (12. s) At O( ). D(x.90) become (Dθx )x + {(Dθx )x + (Dθx )x } + ˆ ˆ ˆ ˆ subject to θx + θx = 0 at x = x = 0 and x = 1. x ˆ subject to ˆ θ0ˆ = 0 at x = x = 0. s) = 0. x) + θ1 (x. We now expand θ as ˆ ˆ θ = θ0 (x. These are ˆ x secular.12. We must eliminate this secularity by taking ˆ 1/ lim →0 α (x) 0 ds ∂ − α(x) 2 D(x. ˆ ˆ ˆ (12. (Dθ0ˆ )x = 0. x We can integrate this once to obtain Dθ0ˆ = α(x).94) θ1 = f1 (x) − f0 (x)ˆ − α (x)ˆ x x 0 ds D(x.2 ORDINARY DIFFERENTIAL EQUATIONS 341 (see Section 12. s) ds . (12.4). 95). x). at leading order. s) 1 D(x. s) R(θ0 (x). we therefore require that x lim 1/ →0 1 d 2 θ0 − 2 dx2 2 2 ∂ ∂x 1/ 0 1/ β(x) 0 s 1/ − s ds D(x. θ varies only on the coarse. u) du ds. 1/ ds For example. we must also eliminate the ˆ secular terms that are of O(ˆ) when x is large in (12. x. s) x ˆ + γ(x) 0 ds D(x. Note that each of the terms in (12. and hence an equation satisﬁed by θ0 .98) 1 d 2 θ0 2 ∂ x − ˆ 2 dx2 ∂x x ˆ 0 β(x) 0 s x−s ˆ ds D(x. ˆ which we can integrate twice to obtain θ2 = f2 (x) − f1 (x)ˆ + x x ˆ x ˆ (12. s)   R(θ0 (x). x. as we shall see. we arrive at  1/ d s  d 2θ 2 dx 2 0 D(x.97) In order to determine β. we must consider one further order in the expansion. x) ˆ x x ˆ = −β (x) − R(θ0 (x).97) to eliminate β(x). 0 D(x. We now have simply θ0 = f0 (x). At O( 2 ).92) gives (Dθ2ˆ + Dθ1x )x = − (Dθ1ˆ )x − (Dθ0x )x − R(θ0 (x).99) In order to eliminate the secular terms of O(ˆ2 ). This does not mean that the ﬁne scale structure has no eﬀect.342 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS conclude that α ≡ 0. O(1) length scale. s) . (12. 0 (12. Since we now know that θ0 = O(1) when x = 1/ . This means that. we can see that θ0 satisﬁes the reaction–diﬀusion equation ¯ D(x)θ0x x ¯ + R(θ0 (x). We therefore require that x ˆ f0 = dθ0 = lim →0 dx 1/ β(x) 0 ds D(x.s) ds dθ 0 0 lim − 1/ ds →0  dx2 dx 0 D(x. u) du ds  = 0. (12. x. s) 1 D(x. s) −β (x) 0 s ds − D(x. s) R(θ0 (x).96) is of O(1) for 1. (12. x.s) is the mean value of 1/D over the ﬁne spatial scale.100) . x. x) = 0. s) −β (x) 2 0 s ds − D(x. 0 After multiplying through by a suitable integrating factor. u) du ds = 0.s) 1/ s +2 2 0 1 D(x. 0 If we now use (12. x)R(ˆ). x) = lim ¯ 2D(x) 2 x 0 d dX 2 1/ s D(X. The homogenized diﬀusion coeﬃcient and reaction term are simply given by the terms D0 (x) and R0 (θ. This asymptotic analysis. A1 and A2 . x. u) du ds 0 (12. whilst the corresponding solution of (12. steady state reaction diﬀusion equation.100) analytically. when ˆ x ˆ x D(ˆ) = 1/ (1 + A1 sin k1 x) and R(ˆ) = 1 + A2 sin k2 x for some positive constants ˆ ˆ k1 . and satisﬁes a standard reaction–diﬀusion equation on the coarse scale. x.  (12. but that these are at most of O( ). modiﬁed by a measure of the mean value of the ratio of the ﬁne scale variation of each. In particular. A1 = A2 = 1/2. x) = KR0 (θ. We can illustrate this for a simple case where it is straightforward to ﬁnd the exact solution of both (12.101) 1/ →0 0 1 D(x. D0 (x) = 1/(1 + x) and R0 = 2x − 1. s) s R(θ0 (x). The analytical solution of (12.89) that vanishes at x = 0 (the solution would otherwise contain an arbitrary constant in this case) is θ= 1 2 1 4 x − x + 2 4 + 2 1 (1 − 2x) 2 2 cos x 1 cos 8 + 3 sin 16 2x − sin x 1 1 − x− 4 16 3 2x −2 .100) correctly reproduces the leading order part of this. we ﬁnd. shows that the leading order temperature does not vary on the ﬁne scale.s) ds 0 1/ ds D(X. which is called homogenization. with A1 < 1 and A2 < 1. Figure 12. Homogenization has been used successfully in many more challenging applications than this linear. where    ¯ D(x) = lim exp −2 →0  ¯ R(θ0 (x). the ﬁne scale structure modiﬁes the reaction term and diﬀusion coeﬃcient through (12. x). for example. where ¯ ¯ R.102). we would ﬁnd that there are variations in the temperature on the ﬁne scale. ¯ ¯ One case where D and R take a particularly simple form is when D(x. x).12.102).102) are the ﬁne scale averages.101) and (12. x) = R0 (θ(x).89) and (12. On substituting these into (12. 1/ K = lim →0 2 0 1 ˆ D(s) 0 s ˆ R(u) du ds . . However. given by the constant K. we ﬁnd that K = 1.s) 0   dX  . ¯ ¯ with D the homogenized diﬀusion coeﬃcient and R the homogenized reaction term.101) ¯ and (12. k2 . x) = ˆ ˆ x ˆ x ˆ D0 (x)D(ˆ) and R(θ(x). after cancelling a constant common factor between D and ¯ that we can use D(x) = D0 (x) and R(θ.15 shows a comparison between the exact and asymptotic solutions for various values of when k1 = k2 = 1.2 ORDINARY DIFFERENTIAL EQUATIONS 343 on the coarse scale. If we were to seek higher order corrections. or dissipative. (12.2.344 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS 0.9 1 Fig.2 0.1 0. 12. 1989). If there were a ﬁrst derivative term in this problem.2 0.7 The WKB Approximation In all of the examples that we have seen so far.2 analytical. 2 1 2 sin x in assessing the strength of elastic media with small carbon ﬁbre reinforcements (Bakhvalov and Panasenko. ε = 0. although we would have a singular perturbation problem.05 0. φ(x)y(x). Analytical and asymptotic solutions of (12. 12.103) the procedure fails as there are no terms to balance with the leading order term.5 x 0. This procedure doesn’t always work! For example.7 0.6 0.4 0.8 0. boundary layer.25 asymptotic analytical.15 θ 0. a ﬁrst derivative term can always be removed.15. the procedure would work.05 0 0 0. if we consider the two-point boundary value problem 2 y (x) + φ(x)y(x) = 0 subject to y(0) = 0 and y(1) = 1. we have used an expansion in algebraic powers of a small parameter to develop perturbation series solutions of differential or algebraic equations. ε = 0.1 analytical.89) when D = 1/(1+x) 1 + and R = (2x − 1) 1 + 1 sin x . ε = 0. Suppose that we have a diﬀerential equation . This procedure has been at its most complex when we have needed to match a slowly varying outer solution with an exponentiallyrapidly varying.1 0. However.3 0. there is a sense in which 2 y + φy = 0 is a generic second order ordinary diﬀerential equation. we can remove the ﬁrst u 2 derivative term. + exp ψ0 + ψ1 + O( ) (ψ0 )2 2 ψ0 + ψ1 + O( ) = exp ψ0 + ψ1 + 1 (2ψ0 ψ1 + ψ0 ) + O(1) . and we need to develop a perturbation method that can deal with it. Because of this. The Basic Expansion A suitable method was proposed by Wentzel. If we substitute these into (12.104) If φ(x) > 0. 2ψ0 2 dx (12. ψ1 = − ψ0 1 d =− (log ψ0 ) .103). and hence (ψ0 )2 = −φ(x). we obtain (ψ0 )2 + (2ψ0 ψ1 + ψ0 ) + φ(x) + O( 2 ) = 0. 1 ψ1 = − log (φ(x)) + constant. say for x > 0. we can simply integrate these equations to obtain x ψ0 = ±i φ1/2 (t) dt + constant. ψ0 + ψ1 ψ0 + ψ1 + O( ) . u u x By choosing 2u + p = 0 and hence u = exp − 1 p(t) dt . Kramers and Brillioun (and perhaps others as well) in the 1920s.12.2 ORDINARY DIFFERENTIAL EQUATIONS 345 of the form w + p(x)w + q(x)w = 0. we can easily show that W + 2u +p W + u u u + p + q W = 0. The appropriate asymptotic development is of the form y = exp Diﬀerentiating this gives y = exp and y = exp ψ0 + ψ1 + O( ) ψ0 + ψ1 (x) + O( ) 2 ψ0 (x) + ψ1 (x) + O( ) . By writing w = W u. 4 . If. we will have an oscillatory solution on one side of the zero and an exponentially growing or decaying solution on the other. and. y= C( ) exp |φ(x)|1/4 x |φ(t)|1/2 dt + D( ) exp − |φ(x)|1/4 x |φ(t)|1/2 dt + ··· . or dissipative.107) Although this equation currently contains no small parameter. (12. in which we neglected terms of O(δ 2 ). in front of the exponential. If φ(x) < 0. or dispersive. then some minor modiﬁcations give an exponential. in this case. (12. and it cannot appear in the ﬁnal result when we change variables back from x to x.346 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS All of these constants of integration can be absorbed into a single constant. φ(x) changes sign. one or other of the expansions (12. (12.105) and (12. (12. and can be written as y= A( ) exp i φ1/4 (x) x φ1/2 (t) dt + B( ) exp −i φ1/4 (x) x φ1/2 (t) dt + ··· . say for x < 0. in order that it remains uniform. x 4 . however. solution of the form.106) will be appropriate. ¯ ¯ x In terms of x and Y (¯) = Y (x). we must have ψ1 ψ0 / . we obtain the generic form of the equation. ψ1 = − log 1 = 0. we can introduce one in a useful way by deﬁning x = δx. Example 1: Bessel functions for x We saw in Chapter 3 that Bessel’s equation is y + 1 ν2 y + 1− 2 x x y = 0. We have introduced this artiﬁcial small ¯ parameter as a device to help us determine how the Bessel function behaves for x 1. 1 If we make the transformation y = x−1/2 Y . We will consider how to deal with this combination of dispersive and dissipative behaviour after studying a couple of examples. If x is large and positive.105) Note that this asymptotic development has assumed that x = O(1). Y + 1+ 1 2 − ν2 x2 Y = 0.107) becomes ¯ 1 ¯ δ2 Y + 1 + δ2 2 − ν2 x2 ¯ ¯ Y = 0.106) If φ(x) is of one sign in the domain of solution.108) By direct comparison with the derivation of the WKB expansion above. (12. we can have ¯ x = O(1) by choosing δ suﬃciently small. The leading order solution is rapidly oscillating. which may depend upon . x ¯ ψ0 = ±i 1 dt = ±i¯. (12.109). which generates a solution of the form y1 = exp c1 x − 0 q(t) + ··· p(t) x ∼ C1 ( ) exp − 0 q(t) p(t) . oscillatory functions of x as x → ∞. Note that. Example 2: A boundary layer Let’s try to ﬁnd a uniformly valid approximation to the solution of the two-point boundary value problem y + p(x)y + q(x)y = 0 subject to y(0) = α.112) The second solution of (12.2 ORDINARY DIFFERENTIAL EQUATIONS 347 so that ¯ Y ∼ A(δ) exp and hence y∼ 1 x1/2 Aeix + Be−ix i¯ x δ + B(δ) exp − i¯ x δ + ··· . 2ψ0 ψ1 + ψ0 + p(x)ψ1 + q(x) = 0.14). We can show that the WKB method is not restricted to ordinary diﬀerential equations with terms in just y and y by considering a further example. (12.110) has ψ0 = −p(x) and hence x ψ0 = − Equation (12. it is more appropriate to use the integral representation (11.110) Using the solution ψ0 = 0 of (12.109) and substitute into (12. (12. y(1) = β when 1 and p(x) > 0.111) (12. we obtain at O(1/ ) ψ0 {ψ0 + p(x)} = 0. although we can clearly see that the Bessel functions are slowly decaying. As we have already seen. If we assume a WKB expansion. p(t) . x ψ1 = − log p(x) + 0 q(t) dt. is the required expansion.111) gives ψ1 = −q(x)/p(x). y = exp ψ0 (x) + ψ1 (x) + O( ) .12. and at O(1).111) then gives p(t) dt + c2 . we cannot determine the constants A and B using this technique.110) and substituting into (12. This is precisely the composite expansion that we would have obtained if we had used the method of matched asymptotic expansions instead. Using the expansions (12. p(t) x Combining (12.106).105) and (12. so that the asymptotic solution can be written as 1 y ∼ β exp x q(t) dt p(t) 1 0 x x + p(0) α − β exp p(x) 1 0 q(t) dt p(t) exp − p(t) dt + 0 q(t) dt . φ(x) > 0 for x > 0. β = C1 exp − p(0) 1 0 q(t) p(t) + 1 C2 exp − p(1) 1 p(t) dt + 0 0 q(t) dt .113) ∼ C2 1 exp − p(x) p(t) dt + 0 0 q(t) dt .112) and (12. so we can write 1 y ∼ β exp x q(t) p(0) dt + α − β exp p(t) p(x) 1 0 q(t) dt p(t) exp − p(0)x .113) then gives the general asymptotic solution as x y = C1 exp − 0 q(t) p(t) + 1 C2 exp − p(x) 1 0 x p(t) dt + 0 0 q(t) dt . p(t) The term exp − 1 p(t) dt is uniformly small and can be neglected. we will also insist that |φ(x)| x−2 for |x| 1. φ(x) < 0 for x < 0. Connection Problems Let’s now consider the boundary value problem 2 y (x) + φ(x)y(x) = 0 subject to y(0) = 1. the last exponential in this solution is negligibly small unless x = O( ) (the boundary layer). φ1 > 0.114) We can now apply the boundary conditions in (12. y → 0 as x → −∞.109) to obtain α = C1 + C2 .115) with (12. p(t) Finally. (12.116) To prevent nonuniformities as |x| → ∞.348 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS The second independent solution therefore takes the form y2 = exp c2 − x 0 p(t) dt x − log p(x) + 0 x x q(t) dt + · · · p(t) (12. p(t) 1 (12. φ(x) ∼ φ1 x for |x| 1. we can immediately write A( ) y = 1/4 exp i φ (x) x 0 φ1/2 (t) dt + . Writing (12. y dt2 † Note that equation (12.17). in ¯ terms of which (12.120) If we now rewrite (12. ¯¯ d¯2 x subject to y (0) = 1.120). The WKB solutions will therefore be valid in two outer regions 2/3 . We will need a small inner region. 3 1 4 4 We can now see that the second term becomes comparable to the ﬁrst when −x = O( 2/3 ). making use of φ ∼ 2/3 φ1 x at leading ¯ order.115) in the inner region. Equation (12.2 ORDINARY DIFFERENTIAL EQUATIONS 349 B( ) exp −i φ1/4 (x) y = C( ) exp − 1 0 x 0 φ1/2 (t) dt + · · · for x > 0. 1/3 .12. and we can expect the asymptotic matching solutions are valid. In fact the overlap domain can be reﬁned to 2/3 Exercise 12. (12.121) becomes ¯ d 2y = t¯. (12. Hence there is an overlap of the domains for which the inner and outer 2/3 |x| 1. φ ∼ φ1 x.119) 1/4 1/6 3 φ (−¯) x 1 y∼ ¯ A( ) 1/4 φ1 x1/4 1/6 ¯ 2 1/2 ¯ exp i x3/2 φ1 3 + B( ) 1/4 φ1 x1/4 1/6 ¯ 2 1/2 ¯ exp −i x3/2 φ1 3 as x → ∞.117) and (12.121)† in terms of a standard equation by deﬁning t = −φ1 x.121) is valid for |x| 1. and we would expect its solution to be valid in the same domain. and the with |x| inner solution must match with the outer solutions. so we can estimate the sizes of the terms in the WKB expansion. ¯ (12.117) |φ(t)|1/2 dt − x 1 log |φ(x)| + · · · 4 for x < 0. and the matching conditions (12.115) shows that the only rescaling possible near to the origin is in a region where x = O( 2/3 ). For |x| 1. we arrive at d 2y ¯ + φ1 xy = 0. namely 2/5 (see |x| process to be successful.118) The problem of determining how A and B depend upon C is known as a connection problem. A similar estimate of the solution for x > 0 also gives a nonuniformity when x = O( 2/3 ). ¯ (12.119) and (12.118) in terms of x leads to ¯ x = x/ 2/3 = O(1) for ¯ the matching conditions C( ) 2 1/2 x y ∼ 1/4 ¯ exp − (−¯)3/2 φ1 as x → −∞. For x < 0 1 1 0 (−φ1 t)1/2 dt − log(−φ1 x) y ∼ C( ) exp − 4 x ∼ C( ) exp − 2 1/2 1 1 φ (−x)3/2 − log φ1 − log(−x) . centred on the origin. ¯ (12. and 1.121) We can write (12. and can be solved by considering an inner solution in the neighbourhood of the origin. we must have b = 0.122) In Section 11. The same technique can be used for Bi(t). We can therefore satisfy the matching condition (12. so that only the Airy function Ai appears in the solution. the solution is real for x > 0.2. (12.350 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS This is Airy’s equation.8. (12. 2 3 Bi(t) ∼ π −1/2 t−1/4 exp 1 Ai(t) ∼ √ (−t)−1/4 sin π 1 Bi(t) ∼ √ (−t)−1/4 cos π 2 3/2 t 3 as t → ∞.124) Using this known behaviour to determine the behaviour of the inner solution. As x → ∞.125). and hence The boundary condition y (0) = 1 then gives a = 1/Ai(0) = Γ ¯ determines C( ) through (12. (12. 2 π (−t)3/2 + 3 4 as t → −∞.120) by taking A( ) = B ∗ ( ) = aφ1 1/6 √ .119).122). This determines all of the unknown constants. as of course it should be.125) 32/3 . and completes the . ¯ In order to satisfy the matching condition (12. so the solution can be written as 1/3 1/3 y = aAi(t) + bBi(t) = aAi −φ1 x + bBi −φ1 x .3 we determined the asymptotic behaviour of Ai(t) for |t| 1 using the method of steepest descents. and we ﬁnd that 1 2 Ai(t) ∼ π −1/2 t−1/4 exp − t3/2 . 2i π 1/6 Since A and B are complex conjugate. ¯ ¯ ¯ (12. which we met in Section 3. and C( ) = 1 −1/2 π 2 1/6 1/6 φ1 a. ¯ −1/4 2 1/3 3/2 π 1 1/3 ¯ sin ¯ + φ x y ∼ a √ φ1 x ¯ 3 1 4 π ∼√ a 1/12 πφ1 x1/4 ¯ 1 2i exp 2 1/2 3/2 π iφ x + ¯ 3 1 4 π 2 1/2 ¯ − exp − iφ1 x3/2 − 3 4 . 2 3 (12.123) 2 π (−t)3/2 + 3 4 . shows that 1 1/3 y ∼ a π −1/2 φ1 x ¯ ¯ 2 +bπ −1/2 φ1 x ¯ 1/3 −1/4 −1/4 exp − 2 1/3 ¯ φ x 3 1 3/2 3/2 exp 2 1/3 ¯ φ x 3 1 as x → −∞. subject to φ0 (1. Since φ = O(1) on the boundary. subject to F (1) = 0. we arrive at 1 1 r F + F − 2 F = − . . At O( 2 ). r r a Using the variation of parameters formula. 12. This arises naturally as the equation that governs time-harmonic solutions of the wave equation. at leading order.12. so the bounded solution that satisﬁes the boundary condition is f (r) = r. If we write z = eiωt φ(x). θ) = sin θ. we expand φ = φ0 + 2 φ2 + O( 4 ). ∇2 φ0 = 0. If we again seek a separable solution. we obtain (12. The Airy function Ai(t) is shown in Figure 11. (12. and hence φ0 = r sin θ. the task is usually rather more diﬃcult than we have found it to be for ordinary diﬀerential equations.3 Partial Diﬀerential Equations Many of the asymptotic methods that we have met can also be applied to partial diﬀerential equations. θ) = 0. exponential decay as t → ∞. which clearly shows the transition from dispersive oscillatory behaviour as t → −∞ to dissipative. r φ = r sin θ + 1. 8 The two-term asymptotic expansion of the solution can therefore be written as φ2 = 1 2 r − r3 sin θ + O( 4 ). and obtain.126) subject to φ(1. this has the bounded solution 1 r − r3 sin θ. This has solutions of the form f = Ar +Br−1 . 8 which is bounded and uniformly valid throughout the circle. Example 1: Asymptotic solutions of the Helmholtz equation As an example of an elliptic partial diﬀerential equation. we obtain 1 1 f + f − 2 f = 0. φ2 = F (r) sin θ.12. subject to φ2 (1. ∇2 φ + 2 φ = 0 for r 1. As you might expect. θ) = sin θ for 1. r r subject to f (1) = 1. (12. with = ω/c. c2 ∂t2 which we met in Chapter 3. 1 ∂2z = ∇2 z. We will proceed by considering four generic examples.3 PARTIAL DIFFERENTIAL EQUATIONS 351 solution. If we seek a separable solution of the form φ0 = f (r) sin θ.126). let’s consider the solution of the Helmholtz equation.126) gives ∇2 φ2 = −r sin θ. 0) = f0 (x) for x 0. (12. The inner solution. in this case at x = 0 only. f0 (0) = 0 and 0 −∞ f0 (x) dx = ftot .127) Note that φ = 0 satisﬁes both the partial diﬀerential equation and the far ﬁeld boundary condition. f0 → 0 as x → −∞. Example 2: The small and large time solutions of a diﬀusion problem Consider the initial value problem for the diﬀusion equation. and also a composite solution valid condition at r = 0 if B ¯ at leading order for all r a. We can gain a lot of insight by asking how the solution behaves just after the initial state begins to diﬀuse (the small time solution. It can be helpful to think of c as a distribution of heat or a chemical concentration. is therefore φ = exp r−1 . The result.130) We could solve this using either Laplace or Fourier transforms . This suggests that we need a boundary layer near r = 1. 0 for x > 0. ¯¯ at leading order. which does not shed much light on the structure of the solution. (12. ∂2c ∂c = D 2 for −∞ < x < ∞ and t > 0. ¯ which will match with the far ﬁeld solution if A = 0. but not the boundary condition at r = 1. where c = 0 initially. however. t 1 The general eﬀect of diﬀusion is to smooth out gradients in the function c(x. ∂t ∂x to be solved subject to the initial condition c(x. 2 ∇φ − φ = 0 subject to φ(1. θ) = 1 and φ → 0 as r → ∞. t 1). The small time solution. would be in the form of a convolution integral.352 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS Let’s also consider a boundary value problem for the modiﬁed Helmholtz equation. t).128) (12. For this reason. and satisfy the boundary ¯ = 1. Diﬀusion will also spread the initial data into x > 0. This smoothing is particularly pronounced at points where c is initially discontinuous. (12. . This has solution ¯ r ¯ ¯ φ = A(θ)er + B(θ)e−¯. If we deﬁne r = 1 + r with r = O(1) in the ¯ ¯ boundary layer for 1.129) with f0 ∈ C 2 (R). we anticipate that there will be three distinct asymptotic regions. and after a long time (t 1). we obtain φrr − φ = 0. b and d are constants of integration. where β.7. in x > 0 we would expect c to be small. However.12.131) Note that c increases in regions where f0 > 0. . where we have noted that c is small. C = − b + 4D 2 log (2βD + x) + d. x The solutions of these equations are A= 1 x2 + βx + Dβ 2 .2.132) † Note that we are using the small time. In this region we expect a gradual smoothing out of the initial data. and vice versa. — Region III: x > 0. t and substitute into (12. B = −D (Axx + 2Ax Cx ) . and the solutions in regions I and III will not match together without a boundary layer centred on the origin. and hence 1 c(x. Ax Bx = 0. so we expect an immediate change from c = 0 to c nonzero. — Region II: |x| 1. t) ∼ f0 (0). in the WKB expansion that we developed in Section 12. we obtain A = DA2 . as physical intuition would lead us to expect. The major feature in this region will be an instantaneous smoothing out of the initial discontinuity. c(x. 2 (12. The WKB solution in region III is therefore c = exp − 1 t x2 + βx + Dβ 2 4D + b log t − b + 1 2 log (2βD + x) + d + o(1) . namely region II. B = b. it is convenient to ﬁnd the solution in region III. If we try a WKB expansion of the form† c(x. t.128) gives f1 = Df0 . Substituting this into (12. 2f2 = Df1 . c(x.128). There will be a ﬂux from x < 0 into this region. Before setting up this boundary layer. Note also that as x → 0− . Let’s begin our analysis in region I by posing an asymptotic expansion valid for t 1.3 PARTIAL DIFFERENTIAL EQUATIONS 353 — Region I: x < 0. t) = f0 (x) + tDf0 (x) + t2 D 2 f0 (x) + O(t3 ). t) = exp − A(x) + B(x) log t + C(x) + o(1) . Thinking about the physics of the problem before doing any detailed calculation is usually vital to unlocking the structure of the asymptotic solution of a partial diﬀerential equation. t) = f0 (x) + tf1 (x) + t2 f2 (x) + O(t3 ). (12. 4D −∞ For this to be consistent with (12. η = x/tα . b = 1/2 and d = log (−2DG).134) b+ log 2βD + t1/2 η + d + o(1) η −∞ The solution of (12. c is exponentially small. x < 0 there is a slow smoothing of the initial data that involves algebraic powers of t. For |x| = O(t1/2 ) there is a boundary layer. √ and hence G = −f0 (0)/ 4πD. 1 − ηc0η = Dc0ηη . in terms of η.135) (12. we can balance terms in this equation to ﬁnd a distinguished limit when α = 1/2.131) and (12.132). In terms II. we have. we can use integration by parts to show that ∞ c0 = f0 (0) + G e−s 2 /4D −∞ ds − 2D −η2 /4D e +O η 1 −η2 /4D e η2 . (12. at leading order. where c = O(1)). (12. we need ∞ f0 (0) + G e−s 2 /4D ds = 0. In order that this is consistent with the matching condition (12. If we now expand as c = c0 (η) + o(1). c must match with the solution in region I as η → −∞. and hence |x| = O(tα ) of η. (12. c0 ∼ exp − 1 2 βη Dβ 2 η2 − 1/2 − + b log t− t 4D t as η → ∞.135) we need β = 0. given by (12. region II. we deﬁne a new variable. x > 0. As η → ∞. with α > 0 to be determined. given by (12. This leaves us with c0 ∼ exp − η2 + log (−2DG) − log η . For |x| t1/2 . For |x| t1/2 . ∂t t ∂η t ∂η Since c = O(1) in the boundary layer (remember. The boundary layer therefore has thickness of O(t1/2 ). In region 1. |η| = O(1). c ∼ F = f0 (0). with the . which is a typical diﬀusive length scale. The structure of the solution that we have constructed allows us to be rather more precise about how diﬀusion aﬀects the initial data.354 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS In the boundary layer. we arrive at the matching conditions c0 ∼ f0 (0) as η → −∞.128) becomes D ∂2c ∂c α ∂c − η = 2α 2 .135). As η → −∞.133) is c0 (η) = F + G e−s 2 /4D ds.133) 2 If we now write the solutions in regions I and III. driven by a diﬀusive ﬂux across the boundary layer.131). The large time solution. The superscript ± indicates whether the solution is for η > 0 or η < 0. This looks like a sensible gauge function. a new asymptotic structure emerges. introduced by the linear terms. scaled variable. which causes a nonuniformity in the expansion.137). 1 F± − ηF± = DF± . is to account for differences in the solution for x > 0 and x < 0. ˙ at leading order.3 PARTIAL DIFFERENTIAL EQUATIONS 355 solution changing by O(1) over a small length of O(t1/2 ).138) (12.12. with η = O(1) for t In order to match with the solution in the inner region. is not calculable by asymptotic methods. we will try to solve this using the expansion c = c0 (t)F± (η) + o(c0 (t)). subject to F± → 1 as η → 0± and F± → 0 as η → ±∞. t) = c0 (t) 2 ˙ ± ± x + α1 (t)x + β1 (t). As |x| → ∞. t) → c0 (t) as η → 0. If we proceed. at leading order. t) = c0 (t) + c1 (x. ∂η c0 (t) ∂η 2 (12. It is straightforward to substitute this into (12. speciﬁcally when x = O c0 (t)/|c0 (t)| . we introduce a ˙ ˙ 1 in this outer region.137) Motivated by the matching condition (12. and hence c0 = Ct−α for some constants C and α. 2 (12. (12. When t is suﬃciently large. The ﬁrst and third terms are of O(c0 (t)). with |c1 | |c0 | for t 1 to ensure that the expansion is asymptotic. (12. η = x/ c0 (t)/|c0 (t)|.138) becomes. and. and similarly for β1 . when t = O(1). Should we include the second term in the leading order balance ˙ ˙ or not? Let’s see what happens if we do decide to balance these terms to get the ˙ richest limit. and seek a solution of the form c(x. and the structure of the solution is rather diﬀerent from that which we discussed above for t 1. where x = O(1). If we substitute this into (12.138). This small time solution continues to evolve. whilst the second term is of ˙ c0 (t) d O(c0 (t) dt |c0 (t)| . diﬀusion will have spread out the initial data in a more or less uniform manner. 2D (12. t) + · · · . c1 grows quadratically. In terms of η. We must then have c0 (t)/c0 (t) = O(t). We will start our asymptotic development where x = O(1). which can be integrated to give c1 (x. In order to deal with this. t 1 After a long time. but this leads to some options.128) becomes ∂c 1 |c0 (t)| d ˙ − η ∂t 2 c0 (t) dt c0 (t) |c0 (t)| ˙ |c0 (t)| ∂ 2 c ˙ ∂c = D . we obtain c0 (t) = Dc1xx .139) . we need c(η. which we shall consider next.136) + − ± The distinction between α1 and α1 .128). using integration by parts.140) As η → 0. we quickly ﬁnd that F± = η 2 + 2D is a solution. so that we require A ±B ± ± 0 ∞ e−s /2D ds = 0. This is now a diﬀerential equation for c0 in the form ftot . we ﬁnd that F± = (η + 2D) A ± B 2 ± ± 0 ∞ e−s /2D ds + O (s2 + 2D)2 2 1 −η2 /4D e η5 . As η → ±∞. Firstly. However. if we look for a quadratic solution. (s2 + 2D)2 2 (12.133). we therefore require that A± = 1/2D. we can show that F± = (η 2 + 2D) A± + B ± η + O(η 3 ). and hence the size of c0 (t). the general solution of (12.142) This just says that mass is conserved during the diﬀusion process. (12.18 we ﬁnd that −∞ F± (η) dη = 2πD and hence that c0 (t) = c0 (t) = |c0 (t)|1/2 ˙ ∞ −∞ 3/2 . Secondly. F± (η) dη √ ∞ In Exercise 12. t) dx = −∞ −∞ f0 (x) dx = ftot . by Taylor expanding the integrand.142) to obtain ∞ −∞ c0 (t)F± (η) dx = c0 (t) c0 (t) |c0 (t)| ˙ ∞ −∞ F± (η) dη = ftot . (s2 + 2D)2 2 (12. some further work is needed.141) In order to determine α. we can integrate (12.139) is F± = (η 2 + 2D) A± + B ± 0 η e−s /2D ds . (s2 + 2D)2 2 The outer solution can therefore be written as F± = 1+ η2 2D η 1∓ 0 e−s /2D ds (s2 + 2D)2 2 ∞ 0 e−s /2D ds .356 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS This is slightly more diﬃcult to solve than (12. we can write down the composite expansion c = cinner + couter − (cinner )outer = c0 (t)F± (η). Using the method of reduction of order. 2D To match with the inner solution.128) and apply the initial condition. and use this in (12. to obtain ∞ 0 c(x. 0) = 0.43). y0 (x. This 2 solution is plotted in Figure 12. Our asymptotic solution therefore takes the form  f  √ tot for |x| = O(t1/2 ). which is the ˙ usual diﬀusive length scale. It only “remembers” the area under the initial data.17 we plot c(0. is in excellent agreement with the exact solution. 0) = y1t (x. governs the small amplitude motion of an elastic string.1.144) = c2 2 − 2 ∂t ∂x ∂t subject to the initial conditions ∂y (x. represents a weak. In Figure 12. at leading order.143) c(x.17(b) we can see that the asymptotic solution. and we can clearly see the structures that our asymptotic solutions predict emerging for both small and large times. t) = 1 ex+Dt erfc √4Dt + Dt .9. 2 (12. for t > 0 and −∞ < x < ∞.16 at various times. t) = 1 {Y0 (x − ct) + Y0 (x + ct)} . subject to y1 (x. proportional to the velocity of the string. namely c(x.9. yt . we ﬁnd (see Exercise 12. y0t (x.145) we obtain ∂ 2 y0 ∂ 2 y0 − c2 = 0. (3.  4πDt (12. (12. ± 4πDt The success of this approach justiﬁes our decision to choose c0 (t) in order to obtain the richest distinguished limit.144) and (12.17(a).147) 2 2 ∂t ∂x ∂t The initial value problem given by (12. and has d’Alembert’s solution. − c2 =− (12. t) = 1 + o(1) for t 2 √ Figure 12. (12. c(0. which we met in Section 3. We can now calculate that c0 (t)/|c0 (t)| = O(t1/2 ). The additional term. ∂t with 1. t) ∼ 1/ 4πDt as t → ∞. t) as a function of Dt. (12.ftot / 4πDt. Our asymptotic solution 1. The one-dimensional wave equation.144) with = 0. 0) = 0.1. linear damping. consistent with Figure 12. 0) = Y0 (x).146) is one that we studied in Section 3. Notice that the large time solution has “forgotten” the precise details of the initial conditions. In predicts that c(0. Example 3: The wave equation with weak damping (i) Linear damping Consider the equation ∂2y ∂y ∂2y . 0) = 0. 0) = Y0 (x).146) 2 ∂t ∂x2 ∂ 2 y1 ∂ 2 y1 ∂y0 .145) y(x.3 PARTIAL DIFFERENTIAL EQUATIONS 357 . subject to y0 (x. The form of the initial conditions suggests that we should consider an asymptotic expansion y = y0 + y1 + O( 2 ). On substituting this into (12.148) √ 12.18) that √ x an exact solution is available. If we consider the particular case f0 (t) = ex . t) ∼  √ftot F (η)  for |x| t1/2 . (12. for example due to drag on the string as it moves through the air. without change of form. In terms of the characteristic variables. The solution of the diﬀusion equation with f0 (x) = ex at various times.149) The initial conditions show that F1 (x) + G1 (x) = 0. ξ = x − ct. This is illustrated in Figure 12. and F1 (x) − G1 (x) = 1 {xY0 (x) − Y0 (x)} .16. 8c (12.358 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS Fig.18 for the initial condition Y0 (x) = 1/(1 + x2 ). η = x + ct.and right-travelling waves is clearly visible. 4c . 2 Integrating this expression twice gives the solution y1 = 1 {ξY0 (η) − ηY0 (ξ)} + F1 (ξ) + G1 (η). This solution represents the sum of two waves. (12. 12. each with speed c. The splitting of the initial proﬁle into left. and with half the initial amplitude.147) becomes −4c2 ∂ 2 y1 =c ∂ξ∂η ∂y0 ∂y0 − ∂ξ ∂η = 1 c {Y0 (ξ) − Y0 (η)} . one travelling to the left and one travelling to the right. in conjunction with (12.149).17. 2 which. x−ct (12. (12. deﬁning a slow time scale T = t.144) becomes ytt + 2 ytT + 2 yT T = c2 yxx − yt − 2 yT . We will proceed using the method of multiple scales. In terms of these new independent variables. t. The solution of the diﬀusion equation with f0 (x) = ex at x = 0. which can be integrated once to give F1 (x) − G1 (x) = Finally. and therefore that our asymptotic We can now see that y1 = O(t) for t expansion becomes nonuniform when t = O( −1 ). and looking for a solution y = y(x. (12.3 PARTIAL DIFFERENTIAL EQUATIONS 359 Fig. T ).151) . shows that 1 1 y1 = − t {Y0 (x + ct) + Y0 (x − ct)} + 4 4c x+ct Y0 (s) ds. xY0 (x) − 2 0 1 Y0 (s) ds + b. 12.150) 1.12. F1 (x) = −G1 (x) = 1 8c x 1 4c x xY0 (x) − 2 0 Y0 (s) ds + b. G0ηT = − G0η . t. T ) + y1 (x.152) As usual in the method of multiple scales. 0) = Y0 (η). t. but now the solution is y0 (x. 0) = 1 1 Y0 (ξ). we must have 1 1 F0ξT = − F0ξ . 2 2 (12.360 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS Fig.146). 12. If we now seek an asymptotic solution of the form y = y0 (x. the presence of the terms of the right hand side causes y1 to grow linearly with t. (12. T ) + G0 (η. 16. t. 20. 8. as before. at leading order we obtain (12. with F0 (ξ.155) . The solution of (12. when = 0 and 0.153) (12.144) when c = 1 and Y0 (x) = 1/(1 + x2 ) at equal time intervals. T ). t = 0. In order to eliminate them. T ) + O( 2 ).1.18. 2 2 (12. G0 (η. We ﬁnd that −4cy1ξη = F0ξ − G0η + 2 (F0ξT − G0ηT ) . 4. we need to go to O( ) to determine F0 and G0 . 12.154) On solving this equation. T ) = F0 (ξ. G0 = e−T Y0 (η). T ) + y1 (x. y = y0 (x.3 PARTIAL DIFFERENTIAL EQUATIONS 361 If we solve these equations subject to the initial conditions (12.144) leads to an exponential decay of the amplitude of the solution over the slow time scale. G0ηT = − c2 G3 . we obtain F0 = and hence 1 − t/2 e {Y0 (x − ct) + Y0 (x + ct)} . in (12. The solutions are F0ξ = and hence Y0 (ξ) 4 + c2 T {Y0 (ξ)} 2 . so let’s go straight to a multiple scales expansion. as before.  ξ y0 = −  ∞ Y0 (s) 4+ c2 T {Y0 (s)} 2 ∞  Y0 (s) 4+ c2 T {Y0 (s)} 2 ds + η ds . t. At leading order. to eliminate secular terms in (12.153).157) subject to the initial conditions ∂y (x.159). (12. 2 2 . y0 = (ii) Nonlinear damping What happens if we replace the linear damping term yt with a nonlinear damping term. t. 0) = Y0 (x). (12. (12. 0η Assuming that F0 (s) and G0 (s) are integrable as s → ±∞. we integrate the expression (F0ξ − G0η )3 twice to obtain ξ ξ 3 F0ξ (s) ds − 3G0 (η) 0 0 2 F0ξ (s) ds + 3F0 (ξ) 0 η η η G2 (s) ds − ξ 0ξ 0 G3 (s) ds.12. We conclude that.153).158) ∂t We would again expect a nonuniformity when t = O( −1 ). T ). yt . 0η 2 2 to be solved subject to (12.156) 2 This shows that the small term. (yt )3 ? We must then solve ∂2y ∂2y = c2 2 − 2 ∂t ∂x ∂y ∂t 3 1 −T 1 e Y0 (ξ). t = O( −1 ).18 shows how this slow exponential decay aﬀects the solution. we need 0η 1 1 3 F0ξT = − c2 F0ξ .159) In order to see clearly which terms are secular. Figure 12. consistent with our interpretation of this as a damping term. G0η = Y0 (η) 4 + c2 T {Y0 (η)} 2 . we have (12. we can see that the terms that become unbounded as ξ and η become large are those associated with 3 F0ξ and G3 .152) and (12. (12. At O( ). for t > 0 and −∞ < x < ∞. 0) = 0.153).160) . y(x. (12. −4cy1ξη = c2 (F0ξ − G0η )3 + 2 (F0ξT − G0ηT ) . This is a device with a tip that senses whether it is in oil or water. oil companies are interested in how much oil is coming out of their oilwells. For obvious reasons. deforms around and is penetrated by the probe. Write down the governing equations and boundary conditions. Although this measurement gives a straightforward way of distinguishing between oil and water when only one liquid is present. A measurement of the current between the core and the cladding is then made to determine the conductivity of the surrounding medium. although it is possible that we can ﬁnd an analytical solution. and an array of probes deployed to give a measurement of how the oil fraction varies across the well. we can use an asymptotic solution technique. We will pre-empt all of this ﬂuid mechanical complexity by considering what happens if. We can break the modelling process down into separate steps. Solve the governing equations using either a numerical method or an asymptotic method. As we discussed at the start of Chapter 11. when one or more of the dimensionless parameters is small. these integrals cannot be determined analytically. 1995). The geometry of the electrical probe. is earthed. which is several orders of magnitude higher in saline water than in oil. is shown in Figure 12. A voltage is applied to the core of the probe. Deﬁne dimensionless variables and identify dimensionless constants. and are of O(T −1/2 ) for T Example 4: The measurement of oil fractions using local electrical probes As we remarked at the beginning of the book. One tool that can be lowered into a producing oilwell to assist with this task is a local probe. or cladding. rather than at the surface. the diﬃculty lies in interpreting the change in conductivity as a droplet of oil approaches. this is highly unlikely when studying real world problems.19. The output from the probe can be time-averaged to give the local oil fraction at the tip. Note that. many problems that arise in engineering are susceptible to mathematical modelling. which is made from sharpening the tip of a coaxial cable like a pencil. meets. the amplitudes of the waves do decay as T increases. there is clearly a diﬃcult ﬂuid mechanical problem to be solved before we can begin to relate the conﬁguration of the oil droplet to the current through the probe (see Billingham and King. whilst the outer layer. for example when dipping the probe into a beaker containing a single liquid. but we can see that 1. How thin must this oil layer become before the current through the probe is eﬀectively equal to that of a probe in pure water? . in the course of the interaction of an oil droplet with a probe. If we want to understand and model this process. Let’s now discuss an example of this type of situation. (i) (ii) (iii) (iv) Identify the important physical processes that are involved.362 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS In general. We will consider a simple device that distinguishes between oil and water by measuring electrical conductivity. a thin layer of oil forms on the surface of the probe. and often want to make this measurement at the point where oil is entering the well as droplets. Equation (12. σ ∂φ = 0.163) represents continuity of potential and continuity of current at an interface. σ is the conductivity.164) using cylindrical polar coordinates coaxial with the probe. and ∂/∂n is the derivative in the direction normal to the interface. . Some approximations are. In order to answer this question. as shown in Figure 12. † This.19.12. so that φ= 1 0 at the surface of the core. the ratio of a typical ﬂuid speed to the speed of light. (12. The electrostatic potential. however. (12. we have the jump conditions [φ] = 0. is an axisymmetric solution of Laplace’s equation. φ. ∂n (12. at the earthed surface of the cladding. see Billingham and King (2001). To complete the problem.161) We will assume that the conducting parts of the probe are perfect conductors. so obvious that justifying them rigorously is a little too pedantic. which is diﬀerent in each medium (oil. (12. we have the far ﬁeld conditions that φ → 0 as r2 + z 2 → ∞ outside the probe. and r = 0 at the tip. 12.163) Square brackets indicate the change in the enclosed quantity across an interface. in itself. Here r0 and r1 are the inner and outer radii of the insulator. for example oil and water or oil and insulator. is an asymptotic approximation that can be made rigorous by deﬁning a small parameter.162) At interfaces between diﬀerent media. and φ ∼ φ∞ (r) as z → ∞ for r0 < r < r1 .19. water and insulator). we must solve a problem in electrostatics.3 PARTIAL DIFFERENTIAL EQUATIONS 363 Fig. ∇2 φ = 0. For a simple introduction to electromagnetism. A cross-section through an axisymmetric electrical probe. since the speed at which oil–water interfaces move is much less than the speed at which electromagnetic disturbances travel (the speed of light)†.165) (12. 166) d 2 φ∞ 1 dφ∞ = 0.165). Our task is to solve the boundary value problem given by (12. This has solution φ∞ = log(r1 /r) . except for a uniform layer of oil on its surface of thickness h r0 . so we let η measure displacement across the ﬁlm. ∂ 2 φ/∂η 2 = 0. which we will not consider for the moment. we have δj where δj = σo . ∂η ∂n (12. The quantities h and r0 are the natural length scales with which to measure displacements across and along the ﬁlm. σj ∂φ ∂φ =δ at interfaces. φ = 0 at r = r1 . This is an example of a problem where the governing equation and boundary conditions are fairly straightforward. The potential is likely to change rapidly across this ﬁlm compared with its variation elsewhere. for r0 < r < r1 . with η = 0 at the surface of the probe and η = 1 at the surface of the water. make some progress by looking for an asymptotic solution in the thin ﬁlm. The ﬁrst thing to do is to set up a local coordinate system in the oil ﬁlm. with ξ = 0 at the probe tip and ξ = 1 a distance r0 from the tip. this provides us with an orthogonal coordinate system. and hence φ varies linearly across the ﬁlm.168) . since we expect variations of φ in the water and the insulator to take place over the geometrical length scale r0 .161) to (12.163)2 . However. in this case we have one region where the aspect ratio is small – the thin oil ﬁlm. (12. We can.161) becomes ∂2φ ∂2φ + δ 2 2 = 0. Away from the tip and the edge of the probe. and let ξ measure displacement along the ﬁlm. and (12. log(r1 /r0 ) (12.167) Turning our attention now to (12. with φ = A(ξ)η + B(ξ). + dr2 r dr Finally. but the geometry is complicated. 2 ∂η ∂ξ where δ= h r0 1. however. At leading order. and a numerical method will have diﬃculty handling this. we will assume that the probe is surrounded by water.364 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS The far ﬁeld potential must satisfy ∇2 φ∞ (r) = subject to φ = 1 at r = r0 . Problems like this are usually best solved numerically. 169) becomes 1 at the surface of the core. the ﬁlm of oil is too thin to prevent a current from passing from core to cladding through the water. since oil and the insulator have similar We expect that δw conductivities. both much less than that of saline water. to show that at the surface of the probe.162). Also. ∂φ = 0 at the surface of the insulator. 1 and δi = O(1). Returning now to (12. and hence φ = 0 in the water. ∂φ  = 0 at the surface of the insulator. note that we have two small parameters.3 PARTIAL DIFFERENTIAL EQUATIONS 365 with the subscripts o. We will not show how to do this here.161) subject to (12. The key point is that this asymptotic analysis allows us to eliminate the thin ﬁlm from the geometry of the problem at leading order. and hence h r0 σo /σw . from which it is straightforward to calculate . and hence A(ξ) = 0 there.168) with j = w. (12.12. δw → 0) have to be treated with care. ∂n ∂η We can now combine all of the information that we have. from the conditions at the surface of the cladding and core.169) in the region outside the probe is geometrically simple. If K 1. as the ﬁnal result usually depends on how fast one parameter tends to zero compared with the other. Double limiting processes like this (δ → 0.169) becomes ∂φ/∂n = 0 at the surface of the probe. (12. w and i indicating oil. and instead include its eﬀect in the boundary conditions (12. The fact that the oil ﬁlm is thin allows us to apply these conditions at the surface of the probe at leading order.169). however. δ and δw . at leading order (12. at the interface between oil and insulator. B(ξ) = 0 at the cladding and B(ξ) = 1 at the core. water and insulator respectively. extract one vital piece of information from our analysis. We conclude that. We can. In this case. φ= 0 at the surface of the cladding. we obtain the richest asymptotic balance by assuming that hσw δ =K= = O(1) as δ → 0. We have proceeded on the basis that K = O(1).165) and (12. What happens if K 1 or K 1? If K 1. as it is outside the scope of this book. We conclude that if K 1. δw r0 σ o and ∂φ ∂φ =K = KA(ξ). and the oil cannot be detected by the probe. at leading order (12.170) ∂n These are precisely the boundary conditions that would apply at leading order in the absence of an oil layer. (12. at leading order ∂φ/∂η = 0. This then shows that φ = 1 − η in the oil ﬁlm over the core and φ = 0 in the rest of the ﬁlm. The solution of (12.169) ∂n  Kφ at the surface of the cladding.164). (12. the potential in the water satisﬁes  K (φ − 1) at the surface of the core. and easily achieved using a computer. 3 Construct a composite expansion.169). For h the oil ﬁlm is eﬀectively invisible. with 1. equivalents of the conductivities. 1. For a typical oil and saline water. eﬀectively insulating it from the water outside. This is such a small length that. 12. local probes are used with alternating rather than direct current driving the core.366 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS the current ﬂowing from core to cladding. for 0 x 1. due to the way that the impedances† of oil and water change with the frequency of the driving potential. and can lead to practical diﬃculties with this technique. Hence show that 2 3 e2 Γ 3 2 1/3 . valid up to O( ). and the current through the probe is determined by the thickness of the ﬁlm. Show that a boundary layer is possible only at x = 0. 1. hc ≈ 10−9 m. Exercises 12. any thin oil ﬁlm coating a probe insulates it from the external ﬂuid. subject to y(0) = 0 and y(1) = 1. In reality. † the a. In this case the oil ﬁlm dominates the response of the probe. x4 + (1 − 3 )x3 − (1 + 3 )x2 − (1 + )x + 1 = 0. or an intermediate variable. . One helpful eﬀect of this is to increase the value of hc . Use the method of matched asymptotic expansions to determine two-term inner and outer expansions. which you should match using either Van Dyke’s matching principle. 1 In each case. Hence show that y (0) ∼ 4e2 − 4e2 + 8e2 tan−1 1 2 as → 0. hc We conclude that there is a critical oil ﬁlm thickness. Determine the leading order outer and inner approximations to the solution of y + x1/2 y + y = 0 for 0 subject to y(0) = 0 and y(1) = 1. hc = r0 σo /σw . for h hc the external ﬂuid is eﬀectively invisible. in practice. sketch the left hand side of the equation for = 0 and The function y(x) satisﬁes the ordinary diﬀerential equation y + (4 + x2 )(y + 2y) = 0. x4 + (1 − 3 )x3 − (1 − 3 )x2 − (1 + )x + 1 = 0.c.1 Determine the ﬁrst two terms in the asymptotic expansion for 0 < of all the roots of each of the equations (a) (b) (c) (d) 12. whilst for h = O(hc ) both the oil and water aﬀect the current through the boundary conditions (12. x3 + x2 − 1 = 0.2 x3 + x2 − x + = 0. when y (0) ∼ −2/3 x 1. with 1. Show that it is not possible to have a boundary layer at either x = −1 or x = 2. . Determine the rescaling needed for an interior layer at x = 0. Use the ˙ method of multiple scales to determine a two-term asymptotic expansion. −3 uniformly valid for all t when 1. y(2) = 7. and hence show that x0 = 1/2 when α = −1 and β = 1.7 subject to y(0) = α. and the leading order inner solution. (b) Assuming that there is an interior layer at x = x0 . for Consider the ordinary diﬀerential equation y+ y+y+ ¨ ˙ 2 y cos2 t = 0. 12.8 subject to y = 1.5 12. subject to y(0) = 1. y(0) = 0. where a dot denotes d/dt. y(1) = β. x 2. Match these two solutions. and hence show that y(0) ∼ 2 as → 0. Sketch the leading order solution. 12.4 The function y(x) satisﬁes the ordinary diﬀerential equation y + (1 + x)y − y + 1 = 0. Show that a term of O( log ) is required in the inner expansion. and 1. Hence show that 1 + 1 as → 0. 1. y (0) ∼ 2 Consider the boundary value problem (2y + y ) + 2xy − 4x2 = 0 for −1 subject to y(−1) = 2.EXERCISES 367 12. Consider the ordinary diﬀerential equation y + yy − y = 0 for 0 x 1. dy/dt = 0 when t = 0. Find the leading order outer solution away from this interior layer. determine the leading order inner and outer solutions when α = 0 and β = 3. (a) Assuming that there is a boundary layer at x = 0. −2 . of uniformly valid for t d 2y + y = y3 dt2 dy dt 2 . and hence show that y(0) = 2 − 3 log + O( ) for 2 1. where a prime denotes d/dx. subject to the boundary conditions y(0) = y(1) = 0. with α and β constants. with 1. for 0 x 1. Determine a two-term inner expansion and a one-term outer expansion. Match the expansions using either Van Dyke’s matching principle or an intermediate region. Now determine the outer solutions up to O( ). Match the two-term inner and outer expansions. for t 0.6 12. determine the leading order inner and outer solutions. Use the method of multiple scales to determine the leading order solution. subject to y(0) = 1. (12.1) dy (0) = 0. (b) 1 1 1− k2 1 − x2 1 k 1 − k 2 + k 2 x2 dx 1 k = 1 3k 2k 2 − 1 L − 2 k2 − 1 K . Hence show that y ﬁrst becomes negative when t = t0 ∼ T0 / . with T = t.81) becomes dE/dT = −2E at leading order. where a dot denotes d/dt. T ).4). You should check that your solution is consistent with the linearized solution when 1. Use Kuzmak’s method to determine the leading order ap√ proximation to the solution when 0 < 1 and 1 < yi < 2. k ≡ k(E) = Hints: .9 Consider the ordinary diﬀerential equation y+ y+y+ ¨ ˙ 2 3 y = 0. to show that τ =t+ 2 at. 1. y ∼ e−T /2 cos τ for 12. Use the ˙ method of multiple scales. . y(0) = 0.10 12. provided √ that yi < 2.) Show that when v0 1. yi − 1 where T0 = 1 3K k √ √ 1 2 2 4 1 + 2E 1 + 2E + 1 2 yi (yi −2) 0 × (2k 2 − 1) L 1 k 1 − 2 (k 2 − 1) K √ 1 + 2E + 1 √ 2 1 + 2E 1 k 1/2 dE.2) dt Use a graphical argument to show that when = 0. 2 dt dt subject to y(0) = yi > 1. Consider the initial value problem dy d 2y −2 − y + y 3 = 0. y is positive. (You will need to consider the ﬁrst three terms in the asymptotic expansion for y.11 Show further that a = −1/8 and determine the next term in the asymptotic expansion. (E12. y ≡ y(τ.368 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS 12. (E12. for t 0. (a) The leading order solution can be written in terms of the Jacobian elliptic function cn (see Section 9. 17 subject to y(0) = 1 and y → 0 as x → −∞.13 ˆ x ˆ x Show that when D(x. and satisﬁes an equation of the form F (x) ∂ ∂θ = ∂t ∂x ∂θ ¯ D(x) ∂x . x/ ) = D0 (x)D(x/ ).2. ¯ where D(x) is given by (12.14 →0 D0 (x) ∂θ ∂x . 12. . 12. If D(x. D x. 12. when λ 1. when 1. ˆ ˆ the homogenized diﬀusion coeﬃcient and reaction term given by (12.16 that satisﬁes V (±∞) = 0 and y(±∞) = 0. when Solve the connection problem 2 y (x) + cx2 y(x) = 0. ∂x When 0 < 1. θ satisﬁes the diﬀusion equation ∂θ ∂ = ˆ ∂x ∂t where ˆ t= lim 12. x.101).121)).6. x)R(ˆ). Find the ﬁrst two terms in the WKB approximation to the solution of the fourth order equation y (x) = {1 − V (x)} y(x) 1.101) and (12. t 2 2 1/ s ds ˆ 0 D(s) . x ∂θ ∂x for 0 < x < 1. ∂θ = 0 at x = 0 and x = 1 for t > 0. Use the WKB method to ﬁnd the eigensolutions of the diﬀerential equation y (x) + (λ − x2 )y(x) = 0.15 subject to y → 0 as |x| → ∞. Give a physical example that would give rise to the initial–boundary value problem ∂ ∂θ = ∂t ∂x subject to θ(x.102) can be written in the simple form described in Section 12. By determining the next term in the WKB expansion. and F (x) is a function that you should ˆ determine.12 12. verify that the overlap domain for the inner and outer solutions of the boundary value problem 2/5 (12. 0) = θi (x) for 0 x 1. at leading order. show that. x) = D0 (x)D(ˆ) and R(θ(x). at leading order. use homogenization theory to show that.115) is 2/3 |x| (see the footnote just after (12. θ is a function of x and t only. x) = R0 (θ(x).EXERCISES 369 12. 19 Use Fourier transforms to solve (12. 0) = 0. How does the solution for c > |α| diﬀer from the solution when c < |α|? Project: The triple deck Consider the innocuous looking two-point boundary value problem y + x3 y + (x3 − )y = 0. t) = 0. L) = 0. . 0) = 0. (E12.20 What feature of the solution arises as a result of the reaction term. t > 0. 0) = f (x) and the boundary condition c(0. 0) = f (x). subject to the boundary conditions u(x. subject to the initial condition u(x. Your solution should be uniformly valid in the domain of solution.128) subject to (12. Find the small time solution of the reaction–diﬀusion equation ut = Duxx + αu.18 12. Find a uniformly valid leading order asymptotic solution of (utt − c2 uxx ) + ut + αux = 0 for −∞ < x < ∞.129) with f0 (x) = √ x ex . t > 0.370 ASYMPTOTIC METHODS: DIFFERENTIAL EQUATIONS 12. ut (x. 0) = f (x).23 12. Find a uniformly valid solution of the hyperbolic equation (ut + ux ) + (t − 1)2 u = 1 for −∞ < x < ∞. when 1. t > 0. show that −∞ F± (η) dη = 2πD. t) = 1 ex+Dt erfc √4Dt + Dt . 2 with 1.22 when 1. and hence show that c(x. y) = g(y) and u(x.3) 12. for |x| α. Your solution should remain valid for large x and t. subject to the initial conditions u(x.24 (a) Show that the outer solution is y = e1−x . when 1. 12. so you will need to resolve any boundary layers that are required. 0) = f0 (x) 0 for |x| < α. Find the leading order asymptotic solution of (uxx + uyy ) + ux + βuy = 0 for x > 0. subject to the initial condition c(x. t) ∼ 1/ 4πDt as t → ∞. αu? Find the leading order solution of the partial diﬀerential equation ct = (D + x)2 cxx for x > 0. when 1. By comparing this result with the large time asymptotic solution that we derived in Sec√ ∞ tion 12. Use this 2 √ expression to show that c(0. subject to y(0) = 1 . 12.3. y(1) = 1. 0 < y < L. subject to the initial condition u(x.21 12. u(0. with water ﬂowing past it at 10 m s−1 . µ = 1 and µ = 1 . ∇ · u = 0. estimate the maximum thickness of the boundary layer on a plate 1 m long. show ¯ that there are two possibilities. (e) The matching between the lower and middle decks requires that x→∞ ¯ lim y = lim y ∗ .EXERCISES 371 (b) Since the outer solution does not satisfy the boundary condition at x = 0. where would you expect a triple or double deck structure to appear? You can ﬁnd some more references and background material in Sobey (2000). ¯ ¯ and x∗ = x/ 1/2 . Re where Re is the Reynolds number. 2 As there is now a danger of confusion in our notation. (u · ∇u) = −∇p + .1. (c) Show that the leading order solution in the lower deck is y = 1 e−¯ . Can you see the structure of the triple deck? (g) The equations for steady.01 and 0. 2 (f) Integrate (E12. By writing x = x/ µ . If these are to be solved in −∞ < x < ∞. This nomenclature comes from the study of boundary layers in high Reynolds number ﬂuid ﬂow. We will refer to the region where x = O(1) as the upper deck. From the data in this book. and hence that the solution in the middle deck is y ∗ = exp 1 − 1 2 (x∗ ) 2 . ¯ ∗ x →0 which is satisﬁed automatically without ﬁxing any constants. there must be a boundary layer.001. Show that the composite solution takes the form y = e1−x + exp 1 − 1 2 (x∗ ) 2 1 x + e−¯ − e. (d) Apply the simplest form of matching condition to show that a = e. y > f (x) subject to u = 0 on y = f (x) and u ∼ U i as x2 + y 2 → ∞. ¯ 2 x and in the middle deck y ∗ = a exp −1/2 (x∗ ) 2 . 0. and refer to the region where x = O(1) as the lower deck. we will deﬁne x = x/ . and refer to the region where x∗ = O(1) as the middle deck. and compare the numerical solution with the asymptotic solution for = 0. high Reynolds number ﬂow of a viscous Newtonian ﬂuid with velocity u and pressure p are 1 2 ∇ u.3) numerically using MATLAB. We are going to have to work harder to determine the behaviour of the other integral paths near the origin. y = −y. 13. the qualitative nature of the solutions changes. We will study what happens when an equilibrium point has one or more eigenvalues with zero real part. These are nonhyperbolic equilibrium points. . y) | x = 0. we will look at an example of a global bifurcation.2. This shows that points on the y-axis close to the origin approach the origin as t → ∞. Finally. and rescaling time so that the negative eigenvalue has unit magnitude. 0) = 0. locally. ˙ y = −y + Q(x. Instability and Bifurcations In this chapter. This is the set of points that are. We will also consider diﬀerential equations that contain a parameter. 0) = Q(0. Next. the system will have the form x = P (x.1) where P and Q are nonlinear functions with P (0.1. of the origin is ω s (0. which we were unable to study by simple linearization in Chapter 9. y) ∈ U } . At parameter values where one or more equilibrium points is nonhyperbolic. which was concerned with phase plane methods for nonlinear. We say that the local stable manifold† in some neighbourhood. a local bifurcation point. After shifting the equilibrium point to the origin. attracted to the origin in the direction of the eigenvector that corresponds to the negative eigenvalue. y = y0 e−t ˙ ˙ that passes through the origin.1 Zero Eigenvalues and the Centre Manifold Theorem Let’s consider the structure of an equilibrium point which has one zero eigenvalue and one negative eigenvalue. U . U ) = {(x.CHAPTER THIRTEEN Stability. y). † We will carefully deﬁne what we mean by a manifold in Section 13. and examine what can happen to the qualitative form of the solutions as the parameter varies. we will build upon the ideas that we introduced in Chapter 9. ordinary diﬀerential equations. autonomous. The linear approximation to (13. (x. ˙ (13. y).1) is x = 0. we will introduce the idea of Lyapunov functions. and show how they can be used to study the stability of nonhyperbolic equilibrium points. writing the system in terms of coordinates with axes in the directions of the eigenvectors. which has the solution x = 0. 1 (The centre manifold theorem) The equilibrium point at the origin of the system (13. y) ∈ U dx .13. is stable/unstable. so that ω c (0.1. then x(t) ∈ ω c (0.1) is stable/unstable if and only if the equilibrium point at x = 0 of the ﬁrst order diﬀerential equation x = P (x. U ) = (x. The local stable and centre manifolds of the system (13. we deﬁne a local centre manifold for (13.1. so that if x(0) ∈ ω c (0.1 ZERO EIGENVALUES AND THE CENTRE MANIFOLD THEOREM 373 In order to work out what is going on. excluding the stable manifold. The main idea now is that the qualitative behaviour of the solution in y x ωs ωc Fig.1) to be a C k curve ω c (0. (x. is determined by the behaviour of the solution on the centre manifold. the neighbourhood of the origin. ˙ (13. h(0) = dh (0) = 0. U ) in a neighbourhood U of the origin such that (i) ω c (0. U ) when x(t) ∈ U . Integral paths in the . (ii) ω c (0. y) | y = h(x).1). a solution that initially lies on the centre manifold remains there when it lies in U .2) where y = h(x) is the local centre manifold. This means that the local dynamics are governed by a ﬁrst order diﬀerential equation. U ) is an invariant set within U . 13. In other words. h(x)). U ). U ) is the graph of a C k function that is tangent to the x-axis at the origin. Theorem 13. This gives us the picture of the local stable and centre manifolds shown in Figure 13. but we can proceed to determine the local form of the centre manifold by assuming that h(x) can be represented as a power series.5) into (13. INSTABILITY AND BIFURCATIONS neighbourhood of the local centre manifold are attracted onto the centre manifold exponentially fast. h(x)). which automatically satisﬁes (13. we write A + 2B → 3B. which we label A and B. rate kab2 . Equating powers of x gives us a2 = 0 at O(x2 ) and a3 = +1 at O(x3 ). Example: Travelling waves in cubic autocatalysis If two chemicals. x ≈ x0 / 1 − 2ax2 t.3) (13. so that x ∼ x0 / −2ax2 t as t → −∞. 1 we can ignore the quartic term we therefore have x = ax3 − x4 + · · · . This algebraic 0 0 behaviour on the centre manifold is in contrast to the exponential behaviour that occurs on the unstable separatrix of a linear saddle point.3). so we need to substitute (13. y = h(x)? Since y = xdh/dx. Proof This can be found in Wiggins (1990). h(x) = a2 x2 + a3 x3 + · · · . On the centre manifold.2) shows that ˙ ˙ dh P (x. For |x| ˙ ˙ ˙ and just consider x ≈ ax3 . and we conclude that the local phase portrait is of a nonlinear saddle.6) (13. (13.3) analytically.4) We will not usually be able to solve (13. (13. react through a mechanism known as cubic autocatalysis. how can we determine its equation. y = −y + y 2 + x2 y + x3 ˙ (13. dx subject to h(0) = dh (0) = 0. This is of the form (13. as shown in Figure 13.374 STABILITY. Integral ˙ paths that begin on the local centre manifold therefore asymptote to the origin as t → −∞.5) with a > 0. dx (13.1. if x = x0 on the centre manifold when t = 0.2. let’s consider the system ˙ x = ax3 + xy. As an example. so that the centre manifold is given by y = h(x) = +x3 + · · · . h(x)) = −h(x) + Q(x.4).1). This shows that (2a2 x + 3a3 x2 + · · · )(ax3 + a2 x3 + a3 x4 + · · · ) = −a2 x2 − a3 x3 + (a2 x2 + a3 x3 + · · · )2 + x2 (a2 x2 + a3 x3 + · · · ) + x3 .7) . so that x > 0 for x > 0 and x < 0 for x < 0.1 Construction of the Centre Manifold Now that we know that a local centre manifold exists. More speciﬁcally. 13. The chemical B is known as the autocatalyst. see Billingham and King (2001). ∂t ∂x ∂2b ∂b = D 2 + kab2 . 13. waves of chemical reaction will propagate away from the initiation site. in the absence of any underlying ﬂuid ﬂow. Speciﬁcally. We can study one of these by looking for a wave that propagates from left to right at constant speed v > 0. The local phase portrait for the system (13.1 ZERO EIGENVALUES AND THE CENTRE MANIFOLD THEOREM 375 y ωs ωc x Fig. since it catalyses its own production. ∂t ∂x (13. ‡ For more background and details on reaction–diﬀusion equations and travelling wave solutions.7). the faster it is produced by the reaction (13.8) where t is time and D the constant diﬀusivity of the chemicals.3. We can derive the partial differential equations that govern this situation using the arguments that we described in Section 12. (13. example 2. If these two chemicals then react in a long thin tube. where y = x − vt. If a small amount of the autocatalyst is introduced locally into a spatially uniform expanse of A with a = a0 . represented by a solution of the form a = a(y). so that their concentrations only vary in the x-direction.02 × 1023 ) of molecules of the substance.13. are chemical reaction and one-dimensional diﬀusion. since the rate of change of the amount of each chemical in a control volume is equal to the total diﬀusive ﬂux through the bounding surface plus the total rate of production of that chemical within the volume.8) become nonlinear ordinary diﬀerential † A mole is a ﬁxed number (Avogadro’s number ≈ 6. we ﬁnd that ∂2a ∂a = D 2 − kab2 . the main physical processes that act.‡ With a and b functions of y alone.2. Such a solution is called a travelling wave solution. where k is the reaction rate constant and a and b are the concentrations of the two chemicals.2.6). The greater the concentration of B. which are measured in moles m−3 †. . b = b(y). c = 0. we need 0 b a0 . and hence a + b = a0 . so that k1 = 0. which we can solve to obtain a + b = k0 + k1 e−vy/D . c → 0 as y → ∞. and 0 β 1 for a physically meaningful solution. c → 0 as y → −∞.11) are b = c = 0 and b = a0 . γ → 0 as z → −∞.14) b . (13. The dimensionless wave speed. V = D v ka2 D 0 .10) This represents the unreacted state ahead of the wave. since we require that both of the chemical concentrations should be positive for a physically meaningful solution.376 STABILITY.16) (13. so we also require that b → a0 .9). is now the only parameter.11) (13. V . INSTABILITY AND BIFURCATIONS equations. γ → 0 as z → ∞. γ= a0 ka2 c 0 z= D a0 ka2 0 y.16) in the (β. The second of these represents the fully reacted state behind the travelling wave.10) then shows that k0 = a0 . b → 0 as y → ∞.13) in a more convenient form by deﬁning the dimensionless variables β= so that dβ dγ = γ. B. we obtain a linear diﬀerential equation for a + b. The next step is to study the solutions of (13.15) and (13. (13. dy dy subject to b → 0. solutions . The boundary condition (13. (13. D d 2a da d 2b db − kab2 = 0. V .12) Note that.15) (13. β → 1. +v 2 dy dy dy dy a → a0 . The only equilibrium points of (13. = −V γ − β 2 (1 − β) . We also require that a+b is bounded as y → −∞. γ) phase plane and determine for what values of the wave speed. where all of the chemical A has been converted into the autocatalyst. If we now add equations (13.9) and arrive at a second order system db dc = c.14) subject to (13. (13. dz dz subject to β → 0.11) to (13. We can therefore eliminate a from (13. D 2 + v + kab2 = 0.13) We can write the boundary value problem given by (13. D = −vc − kb2 (a0 − b) .9) to be solved subject to the boundary condition (13. √ Fig. with z0 = 0. As a preliminary. as shown in Figure 13.3.17) √ √ we ﬁnd that we can satisfy this equation with k = −1/ 2 and V = 1/ 2. The Jacobian is J= 0 −2β + 3β 2 1 −V . and determine the nature of the two equilibrium points.18) simply shows that the solution can be given an arbitrary displacement in the z-direction and remain a solution. The presence of z0 in (13. Since dγ β 2 (1 − β) = −V − . A plausible functional form is γ = kβ(1 − β).1 ZERO EIGENVALUES AND THE CENTRE MANIFOLD THEOREM 377 exist. . This gives dβ 1 = − √ β(1 − β). What about other values of V ? Let’s go along our usual route.3. dz 2 and hence 1 e(z−z0 )/ 2 1 √ . √ So now we know that a solution exists for V = 1/ 2. as we would expect for a wave that propagates at constant speed without change of form. which satisﬁes the boundary conditions. γ = γe = − √ √ β = βe = . 13.13. (13. P1 = (1. dβ γ (13. The analytical solution of the cubic travelling wave problem.18) 2 (1 + e(z−z0 )/ 2 )2 1 + e(z−z0 )/ 2 √ This is an exact solution for V = 1/ 2 and any constant z0 . let’s see if we can guess a solution of this system. 0) and P2 = (0. 0). β = βe (z). 378 STABILITY, INSTABILITY AND BIFURCATIONS At P1 , the eigenvalues are real and of opposite sign, so that P1 is a saddle point. Boundary condition (13.16) shows that we need the integral path that represents the solution to asymptote to P1 as z → −∞. Since the unstable separatrices of P1 are the only integral paths that do this, the solution must be represented by the unstable separatrix of P1 that lies in the physically meaningful region, 0 β 1, shown in Figure 13.4 as S1 . The other boundary condition, (13.15), shows that we need S1 to asymptote to the other equilibrium point, P2 , as z → ∞, if it is to represent a solution. γ P1 P2 S2 S1 β Fig. 13.4. The local behaviour in the neighbourhood of the two equilibrium points. At P2 = (0, 0), the eigenvalues are −V and zero, with associated eigenvectors e− = (1, −V ) and e0 = (1, 0), respectively, so that this is a nonhyperbolic equilibrium point. Since V > 0, there is a local stable manifold in the direction of e− and also a local centre manifold tangent to the β-axis (the direction of e0 ). We can construct a local approximation to the centre manifold, γ = h(β), by assuming that γ ∼ Aβ 2 as β → 0 for some constant A. The governing equations, (13.14), then show that dγ dβ ∼ 2Aβ ∼ 2Aβγ ∼ 2A2 β 3 , dz dz and hence that 2A2 β 3 ∼ −V Aβ 2 − β 2 (1 − β). By equating coeﬃcients of β 2 , we ﬁnd that A = −1/V , and hence that the local centre manifold has γ ∼ −β 2 /V as β → 0. This means that γ = dβ/dz < 0 on the local centre manifold. Points on the centre manifold in β > 0 are therefore attracted to P2 as z → ∞ with, from dβ/dz ∼ −β 2 /V , β ∼ V /z as z → ∞. In contrast, points on the centre manifold in β < 0 are swept away as z increases. This type of behaviour is characteristic of a new type of equilibrium point, known as a saddle–node. To the right of the stable manifold, the point behaves like a stable node, attracting integral paths onto the centre manifold and into the origin, whilst to the left of the stable manifold the point is unstable, as shown in Figure 13.4. 13.1 ZERO EIGENVALUES AND THE CENTRE MANIFOLD THEOREM 379 Apart from the local stable manifold, on which β = O(e−V z ) as z → ∞ (recall that the negative eigenvalue of P2 is −V ), any other integral paths that asymptote to the origin as z → ∞ do so on the centre manifold, with β ∼ V /z. Finally, for any given value of V , we need to determine whether the integral path S1 lies to the right of the stable manifold of P2 , which we label S2 , and therefore enters the origin and represents a solution, or whether S1 lies to the left of S2 , is swept away from the origin into β < 0, and therefore does not represent a physically meaningful solution. We will return to this problem in Section 13.3.3. 13.1.2 The Stable, Unstable and Centre Manifolds We end this section by deﬁning more carefully what we mean by a manifold, and generalizing the deﬁnitions of the stable, unstable and centre manifolds to nth order systems. Let’s begin with some deﬁnitions. A homeomorphism is a mapping f : L → N that is one-to-one, onto and continuous and has a continuous inverse. Here, L and N are subsets of Rn . A C k diﬀeomorphism is a mapping f : L → N that is one-to-one, onto and k times diﬀerentiable with a k-times diﬀerentiable inverse. A smooth diﬀeomorphism is a C ∞ diﬀeomorphism and a homeomorphism is a C 0 diﬀeomorphism. An m-dimensional manifold is a set M ⊂ Rn for which each x ∈ M has a neighbourhood U in which there exists a homeomorphism φ : U → Rm , where m n. A manifold is said to be diﬀerentiable if there is a diﬀeomorphism rather than a homeomorphism φ : U → Rm . For example, a smooth curve in R3 is a one-dimensional diﬀerentiable manifold, and the surface of a sphere is a two-dimensional diﬀerentiable manifold. Now that we have deﬁned these ideas, let’s consider the behaviour of the solutions of dx = f (x), dt (13.19) ¯ where x, f (x) ∈ Rn . In the neighbourhood of an equilibrium point, x, of (13.19), there exist three invariant manifolds. s (i) The local stable manifold, ωloc , of dimension s, is spanned by the eigenvectors of A whose eigenvalues have real parts less than zero. u (ii) The local unstable manifold, ωloc , of dimension u, is spanned by the eigenvectors of A whose eigenvalues have real parts greater than zero. c (iii) The local centre manifold, ωloc , of dimension c, is spanned by the eigenvectors of A whose eigenvalues have zero real parts. s Note that s + c + u = n. Solutions lying in ωloc are characterized by exponential u decay and those in ωloc by exponential growth. The behaviour on the centre manifold is determined by the nonlinear terms in (13.19), as we described earlier in this section. For a linear system these manifolds exist globally, whilst for a nonlinear system they exist in some neighbourhood of the equilibrium point. 380 STABILITY, INSTABILITY AND BIFURCATIONS Example Let’s consider the simple system x = x(2 − x), ˙ y = −y + x, ˙ (13.20) and try to determine the equations of the local manifolds that pass through the equilibrium point at the origin. The Jacobian at the origin is J= 2 0 1 −1 , which has eigenvalues 2 and −1 and associated eigenvectors of (3, 1)T and (0, 1)T u s respectively, so that ωloc is the line y = x/3 and ωloc is the y-axis. We can solve the nonlinear system directly, since the equation for x is independent of y, and the equation for y is linear. The solution is x= 2A , A + 2e−t y = e−t B + 2et − 4 log(Aet + 2) , A where A = 0 and B are constants. There is also the obvious solution x = 0, s y = Be−t , the y-axis, which gives the local stable manifold, ωloc , and also the s x global stable manifold, ω (¯ ). Points that lie in the local unstable manifold, u ωloc , have y → 0 as t → −∞. Since y ∼ e−t (B − 4 log 2/A) as t → −∞, we must x have B = A + 4/A log 2, so that the global unstable manifold, ω u (¯ ), is given in parametric form by 2A 4 4 , e−t A + log 2 + 2et − log(Aet + 2) A + 2e−t A A The phase portrait is sketched in Figure 13.5. y . ωs ωu x ωs ωu Fig. 13.5. The phase portrait of the system (13.20). 13.2 LYAPUNOV’S THEOREMS 381 13.2 Lyapunov’s Theorems Although we are now familiar with the idea of the stability of an equilibrium point in an informal way, in order to develop the idea of a Lyapunov function, we need to consider some more formal deﬁnitions of stability. Note that, in this section, we will consider systems of autonomous ordinary diﬀerential equations, written in ˙ vector form as x = f (x), that have an isolated equilibrium point at the origin. An equilibrium point, x = xe , is Lyapunov stable if for all > 0 there exists a δ > 0 such that for all x(0) with |x(0) − xe | < δ, |x(t) − xe | < for all t > 0. In other words, integral paths that start close to a Lyapunov stable equilibrium point remain close for all time, as shown in Figure 13.6. ε δ xe x0 |x − xe| < ε ∀t>0 Fig. 13.6. A Lyapunov stable equilibrium point of a second order system. An equilibrium point, x = xe , is asymptotically stable if there exists a δ > 0 such that for all x(0) with |x(0) − xe | < δ, |x(t) − xe | → 0 as t → ∞. This is a stronger deﬁnition of stability than Lyapunov stability, and states that integral paths that start suﬃciently close to an asymptotically stable equilibrium point are attracted into it, as illustrated in Figure 13.7. Stable nodes and spirals are both Lyapunov and asymptotically stable. It should be clear that asymptotically stable equilibrium points are also Lyapunov stable. However, Lyapunov stable equilibrium points, for example centres, are not necessarily asymptotically stable. Now that we have formalized our notions of stability, we need one more new concept. Let V : Ω → R, where Ω ⊂ Rn and 0 ∈ Ω. We say that the function V is positive deﬁnite on Ω if and only if V (0) = 0 and V (x) > 0 for x ∈ Ω, x = 0. For example, for n = 3, V (x) = V (x1 , x2 , x3 ) = x2 + x2 + x2 is positive deﬁnite in R3 , 1 2 3 whilst V (x1 , x2 , x3 ) = x2 is not, since it is zero on the plane x2 = 0, and not just 2 at the origin. Note that if −V (x) is a positive deﬁnite function, we say that V is a negative deﬁnite function. In the following, we will assume that V is continuous 382 STABILITY, INSTABILITY AND BIFURCATIONS |x − xe| → 0 as t → ∞ x0 xe δ Fig. 13.7. An asymptotically stable equilibrium point of a second order system. and has well-deﬁned partial derivatives with respect to each of its arguments, so that V ∈ C 1 (Ω). We can now introduce the idea of the derivative of a function V (x) with respect ˙ to the system x = f (x) = (f1 , f2 , . . . , fn ), which is deﬁned to be the scalar product V ∗ (x) = ∇V · f (x) = ∂V ∂V f1 (x) + · · · + fn (x). ∂x1 ∂xn This derivative can be calculated for given V (x) and f (x), without knowing the solution of the diﬀerential equation. In particular, dV ∂V ∂V ∂V ∂V = x1 + · · · + ˙ xn = ˙ f1 (x) + · · · + fn (x) = V ∗ (x), dt ∂x1 ∂xn ∂x1 ∂xn so that the total derivative of V with respect to the solution of the equations coincides with our deﬁnition of the derivative with respect to the system. This allows us to prove three important theorems. Theorem 13.2 If, in some region Ω ⊂ Rn that contains the origin, there exists a scalar function V (x) that is positive deﬁnite and for which V ∗ (x) 0, then the origin is Lyapunov stable. The function V (x) is known as a Lyapunov function. Proof Since V is positive deﬁnite in Ω, there exists a sphere of radius r > 0† contained within Ω such that V (x) > 0 for x = 0 and |x| < r, V ∗ (x) † the set of points with |x| r in Rn . 0 for |x| r. 13.2 LYAPUNOV’S THEOREMS 383 ˙ Let x = x(t) be the solution of the diﬀerential equation x = f (x) with x(0) = x0 . By the local existence theorem, Theorem 8.1, extended to higher order systems, this solution exists for 0 t < t∗ with t∗ > 0. This solution can then be continued for t t∗ , and we denote by t1 the largest value of t for which the solution exists. There are two possibilities, either t1 = ∞ or t1 < ∞. We now show that, for |x0 | suﬃciently small, t1 = ∞. From the deﬁnition of the derivative with respect to a system, dV (x(t)) = V ∗ (x(t)) for 0 dt We can integrate this equation to give t t < t1 . V (x(t)) − V (x0 ) = 0 V ∗ (x(s)) ds 0, since V ∗ is negative deﬁnite. This means that 0 < V (x(t)) V (x0 ) for 0 t < t1 . Now let satisfy 0 < r, and let S be the closed, spherical shell with inner and outer radii and r†. By continuity of V , and since S is closed, µ = minx∈S V (x) exists and is strictly positive. Since V (x) → 0 as |x| → 0, we can choose δ with 0 < δ < µ such that for |x0 | δ, V (x0 ) < µ, so that 0 < V (x(t)) V (x0 ) < µ for 0 t < t1 . Since µ is the minimum value of V in S, this gives |x(t)| < for 0 t < t1 . If there exists t2 such that |x(t2 )| = , then, when t = t2 , we also have, from the deﬁnition of µ, µ V (x(t2 )) V (x0 ) < µ, which cannot hold. We conclude that t1 = ∞, and that, for a given > 0, there exists a δ > 0 such that when |x0 | < δ, |x(t)| < for t 0, and hence that the origin is Lyapunov stable. The proofs of the following two theorems are rather similar, and we will not give them here. Theorem 13.3 If, in some region Ω ⊂ Rn that contains the origin, there exists a scalar function V (x) that is positive deﬁnite and for which V ∗ (x) is negative deﬁnite, then the origin is asymptotically stable. Theorem 13.4 If, in some region Ω ⊂ Rn that contains the origin, there exists a scalar function V (x) such that V (0) = 0 and V ∗ (x) is either positive deﬁnite or negative deﬁnite, and if, in every neighbourhood N of the origin with N ⊂ Ω, there exists at least one point a such that V (a) has the same sign as V ∗ (a), then the origin is unstable. Theorems 13.2 to 13.4 are known as Lyapunov’s theorems, and have a geometrical interpretation that is particularly attractive for two-dimensional systems. The equation V (x) = c then represents a surface in the (x, y, V )-space. By varying c (through positive values only, since V is positive deﬁnite), we can obtain a series of contour lines, with V = 0 at the origin a local minimum on the surface, as shown ˙ in Figure 13.8. Since x = f (x), the vector ﬁeld f represents the direction taken by † the set of points with |x| r in Rn . 384 STABILITY, INSTABILITY AND BIFURCATIONS an integral path at any point. The vector normal to the surface V (x) = c is ∇V , 0, integral paths cannot point into the exterior of the so that, if V ∗ = ∇V · f region V (x) < c. We conclude that an integral path that starts suﬃciently close to the origin, for example with V (x0 ) < c1 , cannot leave the region bounded by V (x) = c1 , and hence that the origin is Lyapunov stable. Similarly, if V ∗ < 0, the integral paths must actually cross from the exterior to the interior of the region. Hence V will decrease monotonically to zero from its initial value when the integral path enters the region Ω, and we conclude that the origin is asymptotically stable. (a) V (b) y V(x) = c2 V(x) = c1 < c2 V(x) = c2 y V(x) = c1 < c2 ∇V x f(x) x Fig. 13.8. (a) The local behaviour and (b) a contour plot of a Lyapunov function near the origin. Although we can now see why a Lyapunov function is useful, it can take considerable ingenuity to actually construct one for a given system. Example 1 Consider the system x = −x − 2y 2 , y = xy − y 3 . ˙ ˙ The origin is the only equilibrium point, and the linearized system is x = −x, y = 0. ˙ ˙ The eigenvalues are therefore 0 and −1, so this is a nonhyperbolic equilibrium point. Let’s try to construct a Lyapunov function. We start by trying V = x2 + αy 2 . This is clearly positive deﬁnite for α > 0, and V (0, 0) = 0. In addition, V∗ = dV = 2x(−x − 2y 2 ) + 2αy(xy − y 3 ) = −2x2 + 2(α − 2)xy 2 − 2αy 4 . dt If we choose α = 2, then dV /dt = −2x2 − 4y 4 < 0 for all x and y excluding the origin. From Theorems 13.2 and 13.3 we conclude that the origin is both Lyapunov and asymptotically stable. As a general guideline, it is worth looking for a homogeneous, algebraic Lyapunov function when f has a simple algebraic form. 13.2 LYAPUNOV’S THEOREMS 385 Example 2 ¨ Consider the second order diﬀerential equation θ + f (θ) = 0 for −π θ π, with θf (θ) positive, f diﬀerentiable and f (0) = 0. We can write this as a second order system, ˙ x1 = x2 , x2 = −f (x1 ), ˙ where x1 = θ. The origin is an equilibrium point, but is it stable? In order to construct a Lyapunov function, it is helpful to think in terms of an equivalent physical system. By analogy with the model for a simple pendulum, which we ˙ discussed in Section 9.1, we can think of f (θ) as the restoring force and θ as the angular velocity. The total energy of the system is the sum of the kinetic and θ ˙ potential energies, which we can write as E = 1 θ2 + 0 f (s) ds. If this energy 2 were to decrease/not grow, we would expect the motionless, vertical state of the pendulum to be asymptotically/Lyapunov stable. Guided by this physical insight, we deﬁne x1 1 f (s) ds. V = x2 + 2 2 0 Clearly V (0, 0) = 0, and, since 0 1 f (s) ds is positive by the assumption that θf (θ) 0, V is positive deﬁnite for −π x1 π. Finally, dV = f (x1 )x2 + x2 · −f (x1 ) = 0. dt By Theorem 13.2, V is a Lyapunov function, and the origin is Lyapunov stable. V∗ = Example 3 Consider the diﬀerential equation x + x + x + x2 = 0. We can write this as a second ¨ ˙ order system, x1 = x2 , x2 = −x1 − x2 − x2 , ˙ ˙ (13.21) 1 where x1 = x. This has two equilibrium points, at (−1, 0) and (0, 0). It is straightforward to determine the eigenvalues of these equilibrium points and show that both are hyperbolic, with (−1, 0) a saddle point and (0, 0) a stable, clockwise spiral. We can now construct a Lyapunov function that will give us some idea of the domain of attraction of the stable equilibrium point at the origin. Consider the function 1 1 2 (x1 + x2 ) + x3 . 2 2 3 1 This function vanishes at the origin and is positive in the region V = Ω= 2 (x1 , x2 ) | x2 > −x2 − x3 , 2 1 3 1 x which is sketched, along with the phase portrait, in Figure 13.9. Let’s consider the curve V = 1 , which passes through the saddle point and the 6 point ( 1 , 0), as shown in Figure 13.10. If V = V0 < 1 , we have a curve that encloses 2 6 the origin, but not the saddle point. By taking V = V0 < 1 arbitrarily close to 6 1 6 , we can make the curve V = V0 arbitrarily close to the saddle point. As we are 386 STABILITY, INSTABILITY AND BIFURCATIONS 1.5 1 0.5 x2 0 −0.5 −1 −1.5 −2 −1.5 −1 −0.5 x1 0 0.5 1 Fig. 13.9. The region Ω and the phase portrait of the system (13.21). The region Ω lies to the right of the curved, broken line, which is V = 0. interested in the domain of attraction of the equilibrium point at the origin, we will focus on Ω0 , the subset of Ω given by V < V0 < 1 , with V0 close to 1 . 6 6 Since V∗ = dV = x2 x1 + x2 + −x1 − x2 − x2 x2 = −x2 1 1 2 dt 0, we immediately have from Theorem 13.3 that the origin is Lyapunov stable. To prove asymptotic stability requires more work, as V ∗ = 0 on x2 = 0, which could allow trajectories to escape from the region Ω0 through the two points where V = V0 meets the x1 -axis, which are labelled as A and B in Figure 13.10. There are various ways of dealing with this. The obvious one is to choose a diﬀerent Lyapunov function, which is possible, but technically diﬃcult. We will use a phase plane analysis. Consider SA , the integral path through the point A. All other integral paths through the boundary of Ω0 in the neighbourhood of A enter Ω0 . The integral path SA cannot, therefore, lie along the boundary of Ω0 . If SA does not enter Ω0 , it must intersect the integral path that enters Ω0 at x2 = 0+ , which is not possible. We conclude that the integral path through A enters Ω0 , as shown in Figure 13.11. A similar argument holds at B. Since the Lyapunov function, V , is monotone decreasing away from the x1 -axis, there cannot be any limit cycle solutions in Ω0 . Finally, since Ω0 has all integral paths entering it, and contains a single, stable equilibrium point and no limit cycles, 13.2 LYAPUNOV’S THEOREMS 387 Fig. 13.10. The regions Ω and Ω∗ and the curves V = 0, V = 1 6 and V = V0 < 1 . 6 we conclude from the Poincar´–Bendixson theorem, Theorem 9.4, that all integral e paths in Ω0 enter the origin, which is therefore asymptotically stable, and that Ω0 lies within the domain of attraction of the origin. In fact, we can see from Figure 13.9 that this domain of attraction is considerably larger than Ω0 , and is bounded by the stable separatrices of the saddle point at (−1, 0). Example 4 Consider the system ˙ x = x2 − y 2 , y = −2xy. ˙ This is a genuinely nonlinear system with an equilibrium point at the origin. The ˙ linear approximation at the origin is x = y = 0, so both eigenvalues are zero. Let’s ˙ try a Lyapunov function of the form V = αxy 2 − x3 , which has V (0, 0) = 0. Since V∗ = dV = (αy 2 − 3x2 )(x2 − y 2 ) − 4αx2 y 2 = 3(1 − α)x2 y 2 − αy 4 − 3x4 , dt we can choose α = 1, so that V ∗ = −y 4 − 3x4 , which is negative deﬁnite. We can see that V = x(y 2 − x2 ) = 0 when x = 0 or y = ±x, so that V changes sign six times on any circle that surrounds the origin. In particular, in every neighbourhood of the origin there is at least one point where V has the same sign as V ∗ , so that all of the conditions of Theorem 13.4 are satisﬁed by V , and hence the origin is unstable. 8 Fig. and the interested reader is referred to Coddington and Levinson (1955) for further information. in order to sort out what happens close to the equilibrium point.2 −1 A −0.1 First Order Ordinary Diﬀerential Equations Let’s consider the ﬁrst order ordinary diﬀerential equation. The phase portrait in the neighbourhood of the point A and the saddle point at (−1.1 −0.86 −0. x = x1 is a nonhyperbolic equilibrium point. 13.2.15 −0.98 −0.388 STABILITY.2 0.3 Bifurcation Theory 13. For example. 0).88 −0. .82 −0. Further reﬁnements exist of the Lyapunov Theorems 13. and we need to retain more terms in the Taylor expansion of X.11.4 that we have studied in this section.5). we saw that a simple linearization about x = x1 determines the local behaviour and stability.15 V = V0 0.94 −0.9 x1 −0.05 −0. whose hyperbolic equilibrium points we studied in Section 9. if X(x1 ) = X (x1 ) = 0 and X (x1 ) = 0. 1 d¯ x ≈ X (x1 )¯2 x dt 2 for x ¯ 1.96 −0.92 −0.6).1 0.2 to 13.84 −0. For a hyperbolic equilibrium point at x = x1 . If X (x1 ) = 0. 13.05 x2 0 −0.3. INSTABILITY AND BIFURCATIONS 0. (9. (9. dx dt x Fig.12. Focusing on this case. for example. The addition of a small positive constant to X(x) would shift the graph upwards by a small amount. for example the addition of a small extra term to X(x).13. This nonhyperbolic equilibrium point therefore attracts solutions from x x1 and repels them from x < x1 .12.22) . and give two equilibrium solutions. x ¯ x 2 A system that contains one or more nonhyperbolic equilibrium points is said to be structurally unstable. we can see that x → 0 (x → x1 ) as t → ∞ ¯ for t0 < 0. ˙ (13. The graph of X(x1 + x) close to x = 0 is shown in Figure 13. Consider the equation x = µ − x2 . in contrast to the faster. Let’s investigate this further. 13. a change in the number of equilibrium points or in their stability. Note that the rate at which solutions asymptote to x = x1 from x > x1 is algebraic. exponential approach associated with stable hyperbolic equilibrium points. This means that a small perturbation. whilst the addition of a small negative constant would shift the graph downwards and lead to the absence of any equilibrium solutions. Consider the function shown in Figure 13.3 BIFURCATION THEORY 389 and hence x≈− ¯ 2 . not to the solution but to the model itself. X (x1 )(t − t0 ) for some constant t0 .12 ¯ ¯ ¯ for X (x1 ) < 0. can lead to a qualitative diﬀerence in the structure of the set of solutions. whilst x → −∞ as t → t0 for t0 > 0. The graph of d¯/dt = X(x1 + x) ≈ 1 X (x1 )¯2 for X (x1 ) < 0. x = 0 is called a bifurcation point. whilst the other is unstable. This suggests that there are saddle–node bifurcations at λ = 0. y = λ − 2λy − y 2 ≈ λ − y 2 .390 STABILITY. Graphs of x = µ − x2 for (a) µ > 0. When µ = 0. which shows the position of the equilibrium solutions as a function of µ. For y 1 and λ 1.13. for reasons that will become ˙ clear in Section 13.3.14. we can see that the equilibrium point with the larger value of y is stable.2. with stable equilibria as solid lines and unstable equilibria as dashed lines. as shown in Figure 13. (b) µ < 0. The bifurcation associated with x = µ − x2 is called a saddle–node bifurcation. INSTABILITY AND BIFURCATIONS where µ is a parameter. which is precisely the normal ˙ form for the saddle–node bifurcation. dy ˙ = −2λ − 2y = ∓ λ2 + λ. can be written in the form x = µ − x2 in the ˙ neighbourhood the bifurcation point. and certainly looks as if it contains two saddle– node bifurcations. in other words one that contains a bifurcation point where two equilibrium solutions meet and then disappear. The equation x = µ−x2 is called the normal ˙ form for the saddle–node bifurcation. ˙ Example Consider the ordinary diﬀerential equation (13. The bifurcation diagram is shown in Figure 13. so there are no real equilibrium points for −1 < λ < 0 and two equilibrium points otherwise. All of the terms on the right hand side of . at the equilibrium points. y = 1. x x x Fig. Any ﬁrst order system that undergoes a saddle–node bifurcation.15. The point µ = 0. Now note that. as shown in Figure 13. because the qualitative nature of the phase line changes there. dy Using the analysis of the previous section. x (b) . y = 0 and λ = −1. We can now draw a bifurcation diagram.23) y = λ − 2λy − y 2 . ˙ √ This has equilibrium points at y = −λ ± λ2 + λ. 13. (a) . whilst for µ < 0 there are no equilibrium points.13. X(x) = −x2 and x = 0 is a nonhyperbolic equilibrium point of the type that we analyzed above. For µ > 0 there are two hyperbolic equilibrium points √ at x = ± µ. 3 BIFURCATION THEORY 391 stable unstable x µ Fig. 13. To examine the neighbourhood of the other bifurcation point. The bifurcation diagram for y = λ − 2λy − y 2 . y = 1 + Bx. we shift the origin there using λ = −1 + Aµ.13. but the only one that is sure to be smaller than at least one of the others is the second.23) are small. The saddle–node bifurcation diagram.15. ˙ (13.14. We make no assumption about how big λ is compared with y 2 . 13. Fig. since y 1 means that λy λ. where . 0) for µ > 0. the Taylor expansion of f (x. 0) > 0. INSTABILITY AND BIFURCATIONS A and B are constant scaling factors that we will ﬁx later. Provided that fµ (0. (ii) if fµ (0. and vice versa for fxx (0. 0)x. In terms of µ and x. Theorem 13. 0) for µ < 0. whilst for µ < 0. 0)µ2 .15. (i) if fµ (0.23) becomes x=− ˙ A A µ − 2Aµx − Bx2 ≈ − µ − Bx2 B B for µ 1 and x 1. 0)µ. 2 2 There are therefore equilibrium points at x= −A1 ± A2 − 4A0 A2 1 ∼± 2A2 −2fµ (0. 0) shows that x ∼ A0 (µ) + A1 (µ)x + A2 (µ)x2 for |x| ˙ where 1 1 A0 (µ) = fµ (0. By choosing A = −1 and B = 1 this becomes the normal form for the saddle–node bifurcation. 0) there are two hyperbolic equilibrium points. there exists a continuous curve of equilibrium points in the neighbourhood of (0. at the equilibrium points. µ) about (0. whilst for µ > 0. 0) < 0. indicates that the sense of the bifurcation is reversed with respect to λ. 0) = 0. 0) < 0. 0) there are two hyperbolic equilibrium points.5 (Saddle–node bifurcation) Consider the ﬁrst order diﬀerential equation x = f (x. then there are no equilibrium points in the bourhood of (0. (13. λ = −1 − µ. neighneighneighneigh- If fxx (0. µ). We can formalize the notion of a saddle–node bifurcation using the following theorem. which is tangent to the line µ = 0 there. in a suﬃciently small bourhood of (0. ˙ with f (0. 0). Since f (0. 0) > 0. A2 (µ) = fxx (0. 0) = 0. 0)fxx (0. dx ˙ ∼ A1 (µ) + 2A2 (µ)x ∼ fxx (0. 0). 0) = 0 and fxx (0. 0)fxx (0. 0) = fx (0. This shows that the location of the equilibrium points is as described in the theorem. Proof We will give an informal proof. the equilibrium point with the larger value of x is stable. 1 and |µ| 1. 0) = fx (0. in a suﬃciently small bourhood of (0. 0)µ + fµµ (0. as can be seen in Figure 13.392 STABILITY. dx . In addition. 0) = 0. Finally. A1 (µ) = fxµ (0. Note that the required change of coordinate. then there are no equilibrium points in the bourhood of (0. 0) 1. whilst the other is unstable. 0)µ for |µ| fxx (0. A continuous ﬂow. If the CSTR has constant volume V . and assume that the idea is to convert as much of the reactant A into the autocatalyst B as possible. and characterizes the period for which a ﬂuid element typically remains within the CSTR. b0 − b db = kab2 + . 0). dt tres where b0 is the inlet concentration of the autocatalyst. The rate of change of the total amount of species A in the CSTR is equal to the rate at which it is produced by chemical reaction plus the rate at which it ﬂows in minus the rate at which it ﬂows out. stirred tank reactor (CSTR).25) . CSTR Reactants in MIXER Products and unused reactants out Fig. to which fresh chemicals are continuously supplied and from which the resulting reactants are continuously withdrawn. This is simply a large container. Similarly. we assume that the concentrations of the chemicals are spatially uniform. we have d (V a) = −V kab2 + q (a0 − a) . Example: The CSTR Many industrially important chemicals are produced in bulk using a continuous ﬂow. and given by a(t) and b(t).16.24) The residence time. constant inlet and (by conservation of mass) outlet ﬂowrate q and inlet concentration of A given by a0 . The stability of the equilibrium points is therefore determined by the sign of fxx (0. β = .13.3 BIFURCATION THEORY 393 since x = O(|µ|1/2 ). τ = ka2 t. is the time it takes for a volume V of fresh reactants to ﬂow into the CSTR.7) are fed into a CSTR.16. stirred tank reactor (CSTR). as sketched in Figure 13. Let’s consider what happens when chemicals A and B that react through the cubic autocatalytic reaction step (13. We now deﬁne dimensionless variables b a α = . 0 a0 a0 (13. dt and hence da a0 − a = −kab2 + . 13. tres = V /q. dt tres (13. B. Since the CSTR is well stirred. A stirrer ensures that the chemicals in the CSTR are well mixed. R(z) has a repeated root at z = 0. R(z).26) (13. In particular. The steady states are therefore given by the points of intersection of these two curves. α + β → 1 + β0 as τ → ∞. and hence for t an initial transient. just the reactant A supplied by the inlet ﬂow.24) and (13. dz = R(z) − F (z). where dz/dτ = 0. The solution is α + β = 1 + β0 + ke−τ /τres . Case 1: No autocatalyst in the inﬂow. where k is a constant. F (z) = 2 (13. a0 (13. F (z). as shown in Figure 13. The term ke−τ /τres represents τ τres . dτ where R(z) = (1 − z) (z + β0 ) . INSTABILITY AND BIFURCATIONS in terms of which (13.28) z . and can thereby eliminate β from (13.27) to obtain 1 + β0 − (α + β) d (α + β) = . it is more convenient to work in terms of z = 1 − α. occurs when the rate of reaction.25) become dα 1−α = −αβ 2 + .27) b0 .26). The straight line F (z) always passes through z = 0. α + β ∼ 1 + β0 . Clearly.394 STABILITY. which is therefore always a steady state. dτ τres a linear equation that we can solve for α + β using an integrating factor. is balanced by the rate at which A ﬂows into the CSTR. z = 0 is the only steady state. β0 = 0 When β0 = 0. whilst F (z) is a straight line through the origin. which decays to zero exponentially fast over a time scale given by the residence time. We therefore assume that α + β = 1 + β0 . dτ τres where τres = ka2 tres . A simple calculation of the steady state solutions shows that for τres < 4. ﬁrst order ordinary diﬀerential equation for z. In the analysis that follows. whilst for . The function R(z) is a cubic polynomial.26) and (13. for tres . The state z = 0 represents a CSTR that contains no autocatalyst. This gives us a nonlinear. dτ τres dβ β0 − β = αβ 2 + . the extent to which the reactant A has been converted to B. and hence when F (z) = R(z).17. τres We can now see that a steady state. β0 = 0 We can now add (13. and the reaction stops. .13. 13. and we ﬁnd that z = 0 and z = z2 are stable states. with β0 = 0.3 BIFURCATION THEORY 395 τres > 4. This situation. since it is probably desirable to run the system with just a single species entering the CSTR. is a realistic one. This is known as washout. and z1 → 0. we can now see that. z2 will also slowly decrease. z2 → 1 as τres → ∞. Fig. z2 = 1 2 1+ 1− 4 τres .17. so that no autocatalyst remains in the CSTR. where z1 = 1 2 1− 1− 4 τres . if τres is slowly decreased. and that when τres decreases past 4. we also want to make the residence time as small as possible. The only available steady state is z = 0. are doomed to failure when there is no autocatalyst entering the CSTR. the saddle–node bifurcation point. The stability of the steady states is easily calculated. at z = z1 and z = z2 . The curves R(z) and F (z) when β0 = 0. There is a saddle–node bifurcation at τres = 4. and hence two other steady states. τres . Note that 0 < z1 < 1 and 1 < z2 < 1. However. We clearly want to run the CSTR in the state z = z2 . whilst z = z1 is unstable. The 2 2 steady states are sketched in Figure 13. to increase the rate at which B is produced.18. However. the situation changes dramatically. where more than half of the reactant A is converted to B. there are two further points of intersection between F (z) and R(z). Attempts to recover the desirable state z = z2 by increasing the residence time. 18. (z± + β0 ) (2 − β0 − 3z± ) There are now two saddle–node bifurcations. we must simultaneously solve R(z) = F (z) and R (z) = F (z).30). 13.19 and 13. For 0 < β0 < 1 . where. τ± = 1 . from (13.20. We conclude that there are no such points of tangency for β0 > 1 .29) (z + β0 ) (−3z + 2 − β0 ) = (13. β0 > 0 Let’s now see how the situation is aﬀected by including some autocatalyst in the inﬂow.30) Eliminating τres between these two equations gives 2z 2 − z + β0 = 0. The bifurcation diagram when β0 = 0. Case 2: Autocatalyst in the inﬂow.396 STABILITY. In order to determine when F (z) is tangent to R(z). . there is a unique solution for 0 τres τ+ and τres > τ− . at τres = τ± . and three 8 solutions for τ+ < τres < τ− . and hence 8 that there is a unique stable solution. the cubic polynomial R(z) has a single positive root at z = 1. and hence the points of tangency are at z = z± = 1 1± 4 1 − 8β0 . which gives (z + β0 ) (1 − z) = 2 z . When β0 > 0. as shown in Figure 13. τres 1 . τres (13. and is strictly positive for 0 z < 1. as shown in Figures 13.19. INSTABILITY AND BIFURCATIONS Fig. ˙ (13. we again 8 have the possibility. there is no possibility of washout. dx ˙ = µ − 2x = dx −µ at x = µ. This is itself a bifurcation. We will study two other types of bifurcation.31) This system has equilibrium points at x = 0 and x = µ.13. if not of a washout of the autocatalyst. 13. τres . and is known as a codimension two 27 bifurcation. Such a bifurcation can only occur in a system that has at least two parameters (here β0 and τres ). whilst when µ = 0 there ˙ are always two equilibrium points. if τres is now increased again. note that when β = 1 the two saddle–node points merge and disappear 8 at τres = 64 . . In this case. When µ = 0 we again have the nonhyperbolic equilibrium point given by x = −x2 . Finally. However. there will be a dramatic increase in the concentration of B as τres increases past τ− . The curves R(z) and F (z) when β0 > 0. We conclude that for β0 > 1 . just a unique 8 steady state solution for any given residence time. This change in the steady state from when a parameter is increased to when it is decreased is known as hysteresis.3 BIFURCATION THEORY 397 Fig. at least of a dramatic decrease in the concentration of B when τres falls below τ+ . µ at x = 0. For 0 < β < 1 . Consider the normal form x = µx − x2 .19. ˙ 2 with f (0. This is called a transcritical bifurcation. there exist two continuous curves of equilibrium points in the neighbourhood of (0. 0) = 0. For each µ = 0 there are two hyperbolic equilibrium points in the neighbourhood of x = 0. 0).20. so this sort of bifurcation is often referred to as an exchange of stabilities. stable for µ < 0.21.398 STABILITY. 13. 0). . 0)fµµ (0. unstable for µ < 0. the equilibrium point with the larger value of x is stable. x = −µ is The bifurcation diagram is shown in Figure 13. INSTABILITY AND BIFURCATIONS Fig. 0) = fx (0.6 (Transcritical bifurcation) Consider the ﬁrst order diﬀerential equation x = f (x. 0) − fxx (0. 0) > 0.9. We will leave the informal proof as Exercise 13. 0) = fµ (0. Provided that fxx (0. If fxx (0. 0) = 0 and fµx (0. At the bifurcation point. 0) > 0. 0) < 0. unstable for µ > 0. and vice versa for fxx (0. µ). These curves intersect transversely at (0. whilst the other is unstable. and hence x = µ is stable for µ > 0. The bifurcation diagram when β0 > 0. Theorem 13. the two equilibrium solutions pass through each other and exchange stabilities. then. 0) transverse to the line µ = 0. (i) if fµx (0. 0) = 0 and fxxx (0. 0) = 0. whose bifurcation diagram is shown ˙ in Figure 13. there is a single equilibrium points for µ < 0. 0) < 0. there is a single equilibrium points for µ > 0.22(a). consider the normal form x = µx − x3 . A similar bifurcation. Theorem 13. Provided that fµx (0.21. ˙ (13. µ). close point for µ > 0 and three equilibrium to (0. close point for µ < 0 and three equilibrium (ii) if fµx (0.22(b). in which the two new equilibrium points created at the bifurcation point are unstable.3 BIFURCATION THEORY 399 stable unstable x µ Fig. 0)fxxx (0. Finally. The transcritical bifurcation.7 (Pitchfork bifurcation) Consider the ﬁrst order diﬀerential equation x = f (x. and three equilibrium √ points at x = 0 and x = ± µ for µ > 0. ˙ with f (0. One curve passes through (0. 0) = fx (0. The bifurcation at x = 0. The bifurcation diagram is shown in Figure 13. In addition. 0) > 0. 0) = fxx (0. µ = 0 is called a supercritical pitchfork bifurcation. 0)fxxx (0.32) This has a single equilibrium point at x = 0 for µ < 0. whilst the other is tangent to µ = 0 at x = 0. 13.13. then. 0). 0). 0). 0) < 0. If fxxx (0. is the subcritical pitchfork bifurcation. to (0. with normal form x = µx + x3 . the single equilibrium point and the outer two of the three equilib- . 0) = fµ (0. 0) = 0. there exist two continuous curves of equilibrium points in the neighbourhood of (0. they often describe the qualitative behaviour of complicated physical systems. as shown in Figure 13.400 STABILITY. at a critical value the beam buckles. x. whilst the middle of the three equilibrium points is unstable. as shown in Figure 13. of the midpoint of the beam. INSTABILITY AND BIFURCATIONS rium points are stable.23(b). as we saw for the CSTR. and vice versa for fxxx (0. Consider a straight beam of elastic material under compression from two equal and opposite forces along its axis. Another simple example is the buckling beam.9. the beam undergoes a simple axial compression and x = 0. For suﬃciently small applied forces. As the applied forces are slowly increased. 13. The (a) supercritical and (b) subcritical pitchfork bifurcation. Although these bifurcations seem rather abstract. we obtain the bifurca- . Now consider the displacement. because the unbuckled state becomes unstable. it is equally likely to buckle left or right. We leave the informal proof as Exercise 13. 0) > 0. one applied at each end. (a) supercritical x µ (b) subcritical x stable unstable µ Fig. If the beam has perfect left–right symmetry.22. When the displacement of the midpoint of the beam is plotted as a function of magnitude of the applied forces.23(a). as in Figure 13. This suggests that the pitchfork bifurcation is itself structurally unstable. second order systems that possess an equilibrium point with an eigenvalue with zero real part are structurally unstable.13. a beam. ˙ (13. with the pendulum being brought to rest as t → ∞. with phase portrait shown in Figure 9.2) to model the eﬀect of friction. Since a centre has two eigenvalues with zero real part. You can investigate this by trying Exercise 13. for example a ruler held between your foreﬁngers.23. contains a nonlinear centre.2(b).2 Second Order Ordinary Diﬀerential Equations As for ﬁrst order equations. No matter how small µ is. is added to the left hand side of (9. Now consider the second order system x = µ − x2 . Of course. This will gradually reduce the energy and the amplitude of the motion. the nonlinear centre ˙ is transformed into a stable focus. and will buckle in a preferred direction. the simple. will not have perfect symmetry. 13. An elastic beam under compression (a) before and (b) after buckling. For example. frictionless pendulum. we conclude that this system is structurally unstable.7. (a) F0 (b) F1 > F0 F0 F1 > F0 Fig.3. consider what happens if we allow for a tiny amount of friction.6).33) . A small change in the governing equations can change the qualitative nature of the phase portrait. with µ > 0.3 BIFURCATION THEORY 401 tion diagram corresponding to a supercritical pitchfork. In terms of the physics. in practice. ˙ y = −y. This is reﬂected in what happens to the phase portrait when a term µdθ/dt. and all integral paths asymptote to θ = θ = 0 as t → ∞ (see Exercise 9. 13.22(a). which occurs when a complex conjugate pair of eigenvalues passes through the imaginary axis away from the origin. a saddle and a node collide and disappear. and all trajectories spiral into the origin from inﬁnity.8) the transcritical bifurcation. 1983). with normal form x = µx − x2 . which we studied in Section 13. nonhyperbolic equilibrium point at (0. whilst P− is a saddle point. ˙ y = −y. whilst for µ > 0 there are two equilibrium √ √ points at P+ = (+ µ. for which the limit cycle is unstable and exists for µ < 0.25. which is the centre manifold for the bifurcation. 0). In each of these examples. with normal forms x = µx ± x3 . When µ = 0 a linear analysis shows that the origin is a centre. and hence has eigenvalues with zero real part. as shown in Figure 13. ˙ (13. we will consider one other type of bifurcation.33) is its normal form. this is an example where the nonlinear terms.402 STABILITY. This shows why it was important to study bifurcations in ﬁrst order systems. We will not consider the more degenerate case where both eigenvalues pass through zero (see Guckenheimer and Holmes. one eigenvalue passes through zero at the bifurcation point. cause the integral paths to spiral into the equilibrium point. see Arrowsmith and Place (1990). since y = −y. At such a point. The normal form of the Hopf bifurcation is most easily written in terms of polar coordinates as r = µr + ar3 . and no other equilibrium points. This is called the Hopf bifurcation. we can ignore the dynamics in the y-direction.36) Let’s assume that a < 0. ˙ where the dynamics are controlled by x = µ − x2 . However. and concentrate on the dynamics on the x-axis. analogous to the corresponding bifurcation in a one-dimensional system. Trajectories spiral onto the limit cycle both from the origin and from inﬁnity. INSTABILITY AND BIFURCATIONS For µ < 0 there are no equilibrium points. 0) and P− = (− µ. In eﬀect. In general. This is just the normal form ˙ of the saddle–node bifurcation in a ﬁrst order equation. for equations that model real physical systems. ˙ y = −y.35) The behaviour on the x-axis is. any limit cycle solutions . 0). Note that ˙ this is a supercritical Hopf bifurcation. and no other equilibrium points. and (13. For more on bifurcation theory. (13. The subcritical Hopf bifurcation.3. Note that. all integral paths asymptote to the x-axis as t → ∞. ˙ (13. However. given by r = −µ/a. here r = ar3 . The point P+ is a stable node. For µ 0 there is a stable focus at the origin. θ = ω.1.24. a saddle–node. The point µ = 0.34) and the supercritical and subcritical pitchfork bifurcations. x = y = 0 is called a saddle–node bifurcation point. and has no counterpart in ﬁrst order systems. ˙ ˙ θ = ω. The various phase portraits are shown in Figure 13. When µ = 0 there is a single. but there ˙ is also a stable limit cycle. since the important dynamics of higher order systems occur on a lower order centre manifold. Two other simple types of bifurcation are (see Exercise 13. occurs when a > 0. in each case. For µ > 0 there is an unstable focus at the origin. 13. y = Y (x. µ). Theorem 13. we can appeal to the Hopf bifurcation theorem. In order to demonstrate that a Hopf bifurcation occurs when a pair of eigenvalues crosses the imaginary axis. y. (c) µ < 0. (b) µ = 0. and it is therefore crucial to determine the position of these bifurcations. that exist are formed at Hopf bifurcations. A saddle–node bifurcation in a second order system. y.3 BIFURCATION THEORY 403 (a) µ > 0 y P− P+ x (b) µ = 0 y x (c) µ < 0 y x Fig. µ).8 (Hopf ) Consider the second order system x = X(x. for (a) µ > 0.24. ˙ ˙ Suppose that .13. µ0 ]. The amplitude of the limit cycle grows like |µ|1/2 and its period tends to 2π/|ω| as µ → 0. (i) X(0. The stability of the limit cycle is the same as that of the origin in µ < 0. and given by α(µ) ± iβ(µ).25. We will not prove this theorem here. µ > 0. For µ < 0 there exists a neighbourhood of the origin that does not contain a limit cycle solution. µ) = Y (0. The idea of the proof is to go through a series of algebraic transformations of the diﬀerential equations that reduce them to the . with α and β real. µ) = 0 for each µ ∈ [−µ0 . where a= 1 (Xxxx + Yxxy + Xxyy + Yyyy ) 16 + 1 {Xxy (Xxx + Xyy ) − Yxy (Yxx + Yyy ) − Xxx Yxx + Xyy Yyy } . For µ > 0 there exists a neighbourhood of the origin that does not contain a limit cycle solution. µ 0 and (b) a < 0. (ii) If aα (0) > 0. (13.404 STABILITY.37) 16ω with all of these partial derivatives evaluated at the origin with µ = 0. a unique limit cycle solution bifurcates from the origin in µ < 0 as µ passes through zero. 13. Then (i) If aα (0) < 0. µ0 ] for some µ0 > 0. a unique limit cycle solution bifurcates from the origin in µ > 0 as µ passes through zero. (iv) a = 0. (iii) the eigenvalues of the equilibrium point at the origin are complex conjugate for µ ∈ [−µ0 . A Hopf bifurcation for (a) a < 0. 0. The stability of the limit cycle is the same as that of the origin in µ > 0. INSTABILITY AND BIFURCATIONS < y (b) µ > 0 y x x Fig. 0. (ii) the Jacobian matrix evaluated at the origin with µ = 0 is 0 ω −ω 0 for some ω = 0. 36). (ii) and (iii) of Theorem 13.39). y = Y (x.37). µ) = µx − ωy + x3 . a2 and a3 . after considerable eﬀort. we ﬁnd that a = 3/8. µ) = ωx + µy + y 2 . Example Consider the second order system ˙ x = X(z. ˙ (13.38) where µ and ω are constants. For |z| 1. Speciﬁcally. y and µ. We begin by deﬁning z = x + iy. y. 1 2 2 1 If we now take this expression. The Hopf bifurcation at µ = 0 is therefore subcritical. and hence that aα (0) = 3/8 > 0.8. using the same transformations as those used in the proof of the Hopf bifurcation theorem. This means that the origin is a stable focus for µ < 0.39). (13. (13.36) for suﬃciently small x.3.39) (13.13. we ˙ ˙ arrive. Let’s now try to write (13. For as straightforward an explanation of the details as you could hope for.38) in the normal form (13.2. see Glendinning (1994). we will concentrate on a concrete example to show how the proof of the theorem works. and replace z and z ∗ using (13. and is algebraically very unpleasant. If we now substitute this particular choice of X and Y into (13. in terms of which (13. which we discussed in Section 9.3 BIFURCATION THEORY 405 normal form. w = z − a1 w2 − a2 ww∗ − a3 w∗2 = z − a1 z − a1 w2 − a2 ww∗ − a3 w∗2 −a2 z − a1 w2 − a2 ww∗ − a3 w∗2 −a3 z ∗ − a∗ w∗2 − a∗ ww∗ − a∗ w2 1 2 3 2 2 z ∗ − a∗ w∗2 − a∗ ww∗ − a∗ w2 1 2 3 = z − a1 z 2 − a2 zz ∗ − a3 z ∗2 + 2a2 + a2 a∗ z 3 1 3 + 3a1 a2 + a2 a∗ + 2a2 z 2 z ∗ 2 3 + 2a1 a3 + a∗ a2 + a2 + 2a∗ a3 zz ∗2 + (a2 a3 + 2a∗ a3 ) z ∗3 + O(|z|4 ). The system therefore satisﬁes conditions (i). but can be used to simplify the nonlinear part by an appropriate choice of the constants a1 . ˙ 4 8 The next step is to make a quadratic near identity transformation. at 1 1 2 3 w = λz − i (z − z ∗ ) + (z − z ∗ ) ˙ 4 8 .40) The idea is that such a transformation leaves the linear part of the equation unchanged when written in terms of w. Instead.38) can be written concisely as 1 1 2 3 z = λz − i (z − z ∗ ) + (z + z ∗ ) . y. diﬀerentiate. we can eliminate quadratic terms from (13. with an unstable limit cycle emerging from the origin in µ < 0. z = w + a1 w2 + a2 ww∗ + a3 w∗2 . The linear part of the system is in the normal form for an equilibrium point at the origin with complex conjugate eigenvalues λ = µ + iω. a linear centre for µ = 0 and an unstable focus for µ > 0. 3 4 A2 = −2a2 −2a∗ 3 1 i + 2a3 λ∗ 4 A3 = −2a3 1 i + 2a1 λ + (a1 + a∗ ) 2 4 1 3 i − a2 (λ + λ∗ ) + 2 8 1 + i (3a2 − 2a3 − 4a1 ) + 3a1 a2 + a2 a∗ + 2a2 (2λ + λ∗ ) .406 STABILITY. 2 3 4 1 i + 2a1 λ + (a∗ + a2 ) 1 4 1 3 i − a2 (λ + λ∗ ) + 2 8 .40). INSTABILITY AND BIFURCATIONS 1 2 − (2a1 z + a2 z ∗ ) λz − i (z − z ∗ ) 4 1 2 − (a2 z + 2a3 z ∗ ) λ∗ z ∗ + i (z − z ∗ ) 4 +3 2a2 + a2 a∗ λz 3 + 3a1 a2 + a2 a∗ + 2a2 (2λ + λ∗ ) z 2 z ∗ 1 3 2 3 + 2a1 a3 + a∗ a2 + a2 + 2a∗ a3 (λ + 2λ∗ ) zz ∗2 + 3 (a2 a3 + 2a∗ a3 ) λ∗ z ∗3 + O(|z|4 ). of w to eliminate z from the right hand side. ˙ where A1 = −2a1 1 i + 2a1 λ + a∗ 3 4 1 1 i − a2 (λ + λ∗ ) + 2 8 i i i . retaining only cubic terms and larger. a2 = . 1 2 2 1 We can now use the deﬁnition. (13. and arrive at w = λ w + a1 w2 + a2 ww∗ + a3 w∗2 ˙ − + 1 i + 2a1 λ 4 w2 + 2a1 w3 + 2a2 w2 w∗ + 2a3 ww∗2 1 i − a2 (λ + λ∗ ) 2 − ww∗ + a∗ w3 + (a1 + a∗ ) w2 w∗ + (a∗ + a2 ) ww∗2 + a3 w∗3 3 2 1 w∗2 + 2a∗ w∗3 + 2a∗ w∗2 w + 2a∗ w∗ w2 1 2 3 1 i + 2a3 λ∗ 4 1 1 1 2 2 3 + i (2a1 w + a2 w∗ ) (w − w∗ ) − i (a2 w + 2a3 w∗ ) (w − w∗ ) + (w + w∗ ) 4 4 8 +3 2a2 + a2 a∗ λw3 + 3a1 a2 + a2 a∗ + 2a2 (2λ + λ∗ ) w2 w∗ 1 3 2 3 + 2a1 a3 + a∗ a2 + a2 + 2a∗ a3 (λ + 2λ∗ ) ww∗2 + 3 (a2 a3 + 2a∗ a3 ) λ∗ w∗3 + O(|w|4 ). a3 = . 4λ 2λ 4 (λ − 2λ∗ ) 1 + i (2a1 − a2 ) + 3 (2a1 + a2 a∗ ) λ. 1 2 2 1 We can now see that we can eliminate all of the quadratic terms by choosing a1 = − which leaves us with w = λw + A1 w3 + A2 w2 w∗ + A3 ww∗2 + A4 w∗3 . appears in the normal form in this way. we ﬁnd that v = λv + (A1 − 2λb1 ) v 3 + {A2 − (λ + λ∗ ) b2 } v 2 v ∗ + (A3 − 2λ∗ b3 ) vv ∗2 ˙ + {A4 + (λ − 3λ∗ ) b4 } v ∗3 + O(|v|4 ). . which we discuss below. we recover the normal form (13. This is an example of a homoclinic bifurcation. For |µ| suﬃciently small. with a = A2 (0). 1 2 2 4 1 i + 2a3 λ∗ 4 1 1 i − a2 (λ + λ∗ ) + 2 8 A4 = −2a∗ 1 + a3 1 + i (a2 − 2a3 ) + 3 (a2 a3 + a∗ a3 ) λ∗ .41) and try to eliminate the cubic terms as well. we can see that the limit cycle is circular. At ﬁrst sight. ˙ using v 2 v ∗ = v|v|2 . we conclude that we cannot eliminate the v 2 v ∗ term. w = v + b1 v 3 + b2 v 2 v ∗ + b3 vv ∗2 + b4 v ∗3 .37). However. the limit cycle moves away from the neighbourhood of the origin and becomes increasingly distorted as it begins to interact with a saddle point that lies close to (−1. the coeﬃcient of b2 is 1. as deﬁned in (13. However. We went through the details of this example purely to illustrate the steps involved in the proof.12176. Indeed. Since we need b2 = O(1) for |v| 1 and λ + λ∗ = 2µ. when µ = µ0 ≈ −0.13.4). If we now write v = reiθ and separate real and imaginary parts. 1 4 The next step is to make another near identity transformation. −1). which is small when µ µ 1.26 shows the unstable limit cycle solution for various µ < 0. it would appear that we can eliminate all of the cubic terms in the same way as we did the quadratic terms. The Hopf bifurcation theorem is far easier to use. and is destroyed. (13.37). and hence that the simplest normal form is v ∼ λv + av|v|2 . A little algebra then shows that a = A2 (0) = 3/8. We should reiterate that this transformation and the ensuing algebra is not what needs to be done whenever the system you are studying contains a Hopf bifurcation point. the hard part of the proof the Hopf bifurcation is to show that a.3 BIFURCATION THEORY 407 −2a∗ 2 1 i + 2a3 λ∗ 4 1 + i (4a3 − 3a2 + 2a1 ) + 2a1 a3 + a∗ a2 + a2 + 2a∗ a3 (λ + 2λ∗ ) .36). In fact. consistent with our earlier calculation of a from (13. as |µ| increases. although the algebra is now easier since we only retain cubic terms. Figure 13.3. Proceeding exactly as we did for the quadratic terms. the limit cycle collides with this saddle point. We calculated these solutions numerically using MATLAB (see Section 9. and we are done. all of the bifurcations that we have studied so far have been local bifurcations.01. which can occur when a limit cycle interacts with a saddle point. 13. −0. the crucial features are usually limit cycles and the separatrices of any saddle points. That is. −0.12176.3 Global Bifurcations With the exception of the previous example. 13. .3.1. the amplitude of the limit cycle grows until.36) for µ = −0. A limit cycle is formed in a Hopf bifurcation when the bifurcation parameter. −0.12 and −0. they have arisen because of a change in the nature of an equilibrium point as a parameter passes through a critical value. −0. µ. The unstable limit cycle solution of (13.27 illustrates an example of a homoclinic bifurcation.005. when µ = µ2 > µ1 .002. −0. INSTABILITY AND BIFURCATIONS Fig. Figure 13.408 STABILITY. is equal to µ1 .26. A global bifurcation is one that occurs because the qualitative nature of the solutions changes due to the interaction of two or more distinct features of the phase portrait. the limit cycle collides with the saddle point.05. As µ increases. For second order systems. labelled S2 in Figure 13. since a local analysis is of no help. A homoclinic bifurcation. since an arbitrarily small change to the equations can destroy the connection. but quite another to be able to quantify when the bifurcation occurs. This is the homoclinic bifurcation point.27. This region is constructed so that there are three qualitatively diﬀerent possibilities. For µ > µ2 . We must usually resort to numerical methods. enters the origin as z → ∞. where γH (β) = −β 2 (1 − β)/V is the horizontal isocline. (a) (b) (c) µ1 < µ < µ2 µ = µ2 µ > µ2 Fig. γ = − 27 V is the local minimum of 3 γH (β). Of course. γH (β) < γ < 0 . With the separatrices joined up like this. . it is one thing to describe a global bifurcation in qualitative terms. Note that the point β = 2 . there is no limit cycle.26.1. γ) | β < 1. the system is structurally unstable. labelled S1 in Figure 13. as we did for the example shown in Figure 13. 13.14).3 BIFURCATION THEORY 409 forming a connection between the stable and unstable separatrices.1 we performed a local analysis of the two equilibrium points of (13.13. The region R is shown 4 in Figure 13.28. We begin by deﬁning the region R= (β. Example: Travelling waves in cubic autocatalysis (continued) In Section 13.4. This is a global problem involving the relative positions of S1 and the stable manifold of P2 . — Case (a): S2 enters R through the β-axis.4. − V <γ<0 3 27 (β. we need to determine for what values of V the unstable separatrix of P1 . In order to determine for what values of V a physically meaningful solution exists. γ) | 0 < β < 2 3 4 2 . We will therefore conﬁne ourselves to a single example of a global bifurcation where analytical progress is possible. to show that f (V ) = O(V 2 ) for V f (V ) ∼ V for V 1. and this solution enters the origin on the stable manifold. We will leave the next stage of our argument as Exercise 13. since f (V ) < 1 for V suﬃciently small. and a unique physically meaningful solution when f (V ) 1 (cases (b) and (c)). all integral paths that meet the straight parts of the boundary of D(V0 ) also enter D(V0 ).14) when f (V ) < 1 (case(a)).29. since S2 is directed along the vector e− = (1.14). In addition. Proof When V = V0 > 0. f (V ) is continuous. In case (b). from (13. where γ0 is the value of γ where S2 crosses the line β = 1 (which it does. −V ) as it enters the origin.11. Note that S2 cannot enter R through the γ-axis. and there is no physically meaningful solution of (13. — Case (a): f (V ) is equal to the value of β where S2 crosses the β-axis leaving the region R. We now need to determine which of these three cases arises for each value of V . (b) and (c) are illustrated in Figure 13. INSTABILITY AND BIFURCATIONS — Case (b): S2 = S1 . as illustrated in Figure 13. we deﬁne the region D(V0 ) = (β.1 f (V ) is strictly monotone increasing for V > 0. ∂ ∂V dγ dβ = −1 < 0. when V = V1 > V0 . as illustrated in Figure 13. it lies outside D(V0 ) in a suﬃciently small neighbourhood of the origin when V = V1 . S1 lies above S2 and is therefore attracted into the origin on the centre manifold. We conclude that when V = V1 > V0 . Finally.410 STABILITY. but we will not do this here (see Billingham and Needham. and therefore that f (V1 ) > f (V0 ). since dβ/dz = γ < 0). Cases (a). 1991. In case (c). helped by some hints. S2 = S1 represents a physically meaningful solution. since dβ/dz = γ < 0 there. It cannot represent a physically meaningful solution in this case. γ) | 0 γ γS (β)|V =V0 . — Case (b): f (V ) = 1. — Case (c): S2 enters R through its lower boundary. In particular. — Case (c): f (V ) = 1 − γ0 . and hence asymptotes to P1 as z → −∞. in which you 1. f (V ) > 1 for V .28. Now. Lemma 13. From (13. 0 β 1 . We can do this by deﬁning a function f (V ). S1 lies below S2 and therefore does not enter the origin and is swept into the region β < 0. for more details). This means that the slope of the integral path through any ﬁxed point is strictly monotone decreasing as V increases. In case (a). and that are asked.17).29. S2 cannot pass through the boundary of D(V0 ). where γ = γS (β) is the equation of S2 within the region R. Deﬁned is this way. all integral paths that meet the curved boundary of D(V0 ) (a boundary given by S2 when V = V0 ) enter D(V0 ). These arguments can be made more rigorous. 1. The region R and the three possible types of global behaviour of S1 and S2 .28. and f (V ) is continuous.13. we conclude from the intermediate value theorem that there exists a V ∗ and f (V ) < 1 for V < V ∗ . suﬃciently large. 13.3 BIFURCATION THEORY 411 (a) γ P2 S2 P1 β S1 (b) γ P2 P1 β S2 = S1 (c) γ P2 S1 S2 P1 β Fig. unique value V = V ∗ . such that f (V ) 1 for V ∗ When V = V there is therefore a global bifurcation. since we have now shown that . f (V ) is strictly monotone increasing by Lemma 13. In order to determine the numerical value of V ∗ . case (b) occurs when V = V ∗ and case (c) occurs when V > V ∗ . 13.412 STABILITY. The deﬁnition of f (V ) and the region D(V ) in each of the three cases. This global bifurcation comes about purely because of the relative positions of the manifolds S1 and S2 .29. we would usually have to solve . case (a) occurs when V < V ∗ . INSTABILITY AND BIFURCATIONS (a) γ f(V) < 1 D S2 β β=1 (b) γ f(V) = 1 β D S2 = S1 β=1 (c) γ f(V) > 1 β D S1 β=1 Fig. show that ω1 +ω2 +ω3 is a constant of the motion. ˙ 13. y = −y − x2 . we do know that when V = ∗ V ∗ the solution asymptotes to the origin on the stable manifold with β = O(e−V z ) as z → ∞. V = 1/ 2. y = −4xy. However. y = xy − y 3 .3 ¨ ˙ Consider the second order diﬀerential equation θ + k θ + f (θ) = 0. ˙ 2 ˙ (b) x = x . The reader is referred to Billingham and Needham (1991) for further details. Use arguments based on Lyapunov’s theorems to determine the stability of the equilibrium points of ˙ (a) x = x2 − 2y 2 .8). and to King and Needham (1994) for a similar analysis when the diﬀusion coeﬃcient is not constant. θf (θ) > 0 for θ = 0 and f (0) = 0. ω3 ) is the angular velocity of the body relative to its principal axes. y = −y − x2 . where f is continuously diﬀerentiable on the interval |θ| < α. y = x2 y − y. the solution asymptotes to the origin on the centre manifold. ˙ ˙ (b) x = xy 2 − x. Euler’s equations for a rigid body spinning freely about a ﬁxed point in the absence of external forces are Aω1 − (B − C)ω2 ω3 = 0. ˙ C ω3 − (A − B)ω1 ω2 = 0. (13. By considering a suitable Lyapunov function. with β ∼ V /z as z → ∞.1 Sketch the phase portraits of the systems ˙ (a) x = −x − 2y 2 . whilst this decay is only algebraic for V > 1/ 2. whilst for V > V ∗ . This must therefore be the unique solution that corresponds to V = V ∗ . ˙ where A.4 . when V = 1/ 2.14) numerically. This has implications for the selection of the speed of the waves generated in an initial value problem for equations (13. Exercises 13. including in your sketch the unstable and centre manifolds. with α > 0 a given constant. Find all of the steady states of Euler’s equations. B and C are the principal moments of inertia. and that the solution with V = 1/ 2 asymptotes to zero expo√ nentially fast as z → ∞. In conclusion. show that the origin is an asymptotically stable equilibrium point. in particular that localized initial inputs of autocat√ alyst generate waves with the minimum speed. ω2 . Using the ideas that we 2 2 2 developed in Section 9. which has β = O(e−z/ 2 ) as z → ∞. we have shown that a unique travelling wave solution exists for √ √ each V 1/ 2. 13. and we √ conclude that the global bifurcation point is V = V ∗ = 1/ 2. and ω = (ω1 .EXERCISES 413 the governing equations (13.4. ˙ in the neighbourhood of the origin.2 13. ˙ ˙ (c) x = y + x2 . We were also able to show that there is an √ √ analytical solution.18). ˙ B ω2 − (C − A)ω3 ω1 = 0. ˙ Sketch the bifurcation diagram for each system. Deduce that the steady state with ω1 = ω0 . f (u. Suppose that W has a local minimum at r = 0.5 is a Lyapunov function for the case when A is the largest moment of inertia. express this equation of motion in ˙ ˙ ˙ terms of ﬁrst derivatives only. 13. write each system in the normal form appropriate to the bifurcation. Sketch the bifurcation diagram when = 0. ω2 = ω3 = 0 is unstable if C < A < B or B < A < C. In one case there are two saddle–node bifurcation points. in the other there are no bifurcation points. By using the Lyapunov function 1 V = W + m u2 + v 2 + w 2 . INSTABILITY AND BIFURCATIONS 2 2 2 2 Sketch the phase portrait on the surface of the sphere ω1 + ω2 + ω3 = ω0 . Close to each bifurcation point. Show that V = 2 2 2 2 2 B(A − B)ω2 + C(A − C)ω3 + Bω2 + Cω3 + A ω1 + 2ω0 ω1 2 . ˙ (c) y = −y(4y 2 + λ2 − 1). nonconservative force. and show that each system has two bifurcation points of the same type. Suggest a Lyapunov function that will establish the stability of this state when A is the smallest moment of inertia. y. Are these states asymptotically stable? Perform a simple experiment to verify your conclusions. w). What do the level curves of V represent physically? Is the origin asymptotically stable? If an additional. 13. ˙ When = 0 this is the normal form for a transcritical bifurcation. r By writing x = u. ˙ (b) y = y 2 + 4λ2 − 1.7 . which you should determine.6 13. so that the state ω1 = ω0 . so that its equation of motion is m¨ = −∇W. ω2 = ω3 = 0 is stable. also acts on the particle. 2 show that the origin is a stable point of equilibrium for the particle. (a) Consider the system x = − + µx − x2 . v. so that m¨ = −∇W + f .414 STABILITY. z). A particle of mass m lies at r = (x. Consider the three systems (a) y = y(y − 1)(y − 2λ). dealing separately with the cases > 0 and < 0. z) and moves in a potential ﬁeld W (x. y = v and z = w. r describe qualitatively how the stability of the point of equilibrium at the origin is aﬀected. y. but stable otherwise. In this case. but where.3.35). Plot how the limit cycle solution changes with µ. rate k2 b. Use a near identity transformation to write this system in normal form. and breaks down to form the ﬁnal product C through the chemical reaction B → C. Project Consider the CSTR system that we studied in Section 13. ˙ Consider the system x = µx − ωy + x3 . ¯ ¯ Hence show that f (V ). and conﬁrm that this is consistent with the Hopf bifurcation theorem. subject to γ ∼ −β as β → 0.17) subject to the boundary condition (13. 13.11 (b) Repeat part (a) when V 1.10 13. µ = 0 and µ < 0 for the normal forms of (a) the transcritical and (b) the supercritical pitchfork bifurcation. Sketch the bifurcation diagram when is small.17) and (13. ˆ = −1 − ˆ γ ˆ dβ Integrate this equation numerically using MATLAB. the autocatalyst is itself unstable. and hence show that f (V ) ∼ β ∗ V 2 as V → 0.9 13. by rescaling γ = V γ and using an ¯ asymptotic expansion ¯ γ (β) = γ0 (β) + V −2 γ1 (β) + O(V −4 ). valid when V 1.34) and (13. ˆ β2 dˆ γ ˆ ˆ .EXERCISES 415 (b) Consider the system x = µx + 2 x2 − x3 . What can you deduce from the answers to the two parts of this question? Sketch the phase portraits when µ > 0. Give an informal proof of the transcritical bifurcation theorem and the pitchfork bifurcation theorem.15). Use the Hopf ˙ bifurcation theorem to determine the nature of the Hopf bifurcation at µ = 0. 13. Use MATLAB to solve this system of equations numerically.1. ˙ with 0. you will need to seek ˆ a rescaling of the form β = φ(V )β.15). When = 0 this is the normal form for a supercritical pitchfork bifurcation. at leading order.3. (a) Seek an asymptotic solution of (13. in addition. y = ωx + µy + y 3 . γ = ψ(V )ˆ . but nonzero. as deﬁned in Section 13.12 where β ∗ is a constant that you should determine numerically. and determine φ γ and ψ by seeking an asymptotic balance in (13.3. You should ﬁnd that. given by (13. and determine for what range of values of µ it exists. . satisﬁes 1 f (V ) = V + 1 − V −1 + O(V −3 ) for V 6 1.8 13. . Show that the largest steady state loses stability through a Hopf bifurcation. Determine the range of values for which there are three steady state solutions. and secondly when there is. Without making any quantitative calculations. indicating the location of any limit cycles. (b) Using the ideas that we developed in Section 13. dτ τres β0 − β β dβ = αβ 2 + − . Draw the complete bifurcation diagram. show that the steady state solutions are again given by the points of intersection of a cubic polynomial and a straight line through the origin. ﬁrstly when there is no autocatalyst in the inﬂow. there is more than one limit cycle. dτ τres τ2 (E13. INSTABILITY AND BIFURCATIONS (a) Show that the concentrations of A and B now satisfy the equations dα 1−α = −αβ 2 + . (d) Use MATLAB to integrate (E13.3.2) numerically. (c) Now restrict your attention to the case where there is no autocatalyst in the inﬂow.1. Show that the smallest of these steady states is stable and that the middle steady state is a saddle point.2) where τ2 is a dimensionless constant that you should determine. What advice would you give to an engineer trying to maximize the output of C from the CSTR? Hint: For some parameter ranges.416 STABILITY. sketch the position of the steady states in the (τres . z)-plane.1) (E13. Use the Hopf bifurcation theorem to determine when this is supercritical and when it is subcritical.1) and (E13. and hence investigate what happens to the limit cycle that forms at the Hopf bifurcation point. whose location you should determine. and the government attempts to control the state of the system to a target state (low inﬂation.CHAPTER FOURTEEN Time-Optimal Control in the Phase Plane Many physical systems that are amenable to mathematical modelling do not exist in isolation from human intervention. We wish to drive the system to the target state in the shortest possible time. foreign exchange rates. for example. for which the Treasury has a complicated mathematical model. We will also restrict our attention to timeoptimal control. For the case of two herbivorous species. high growth) by varying several control parameters (most notably taxes and government spending). for which linear equations are a good approximation. continually introducing animals of the less successful species or some combination of these two methods of control. which we know. The crucial results that we will work towards are the properties of the controllability matrix and the application of the time-optimal maximum principle. this may well be undesirable. for which the cost function is the time taken for the system to reach the target state. as it is often necessary to control small deviations from a steady state. and we would like to know how to intervene to maintain the island close to a state of equilibrium. for which we have studied a simple mathematical model. if left uncontrolled. growth. we saw that one species or the other will eventually die out. however. high employment. for example when steering a ship. consists of two populations of diﬀerent species coexisting on an isolated island. which involves linear. and in this short. These are examples of optimal control problems. If the island is under human management. Optimal control is a huge topic. which the government tries to minimize (some function of. government borrowing and. Each of these actions has a cost associated with it. Although we will give proofs of the various results that we need. These are. introductory chapter we will study just about the simplest type of problem. which we studied in Chapter 9. is unstable. unemployment. We could choose between either continually culling the more successful species. A good example is the British economy. . extremely important. one would hope. There is also a cost associated with any particular action. The optimal control leads to the economy reaching the target state with the smallest possible cost. constant coeﬃcient ordinary diﬀerential equations. consumer spending and inﬂation). the environmental cost of any government action or inaction). The state of the system (the economy) is given by values of the dependent variables (for example. Another system. the main thing is to know how to apply them. these conditions guarantee the existence and uniqueness of the solution for a given control u. linear ordinary diﬀerential equations with constant coeﬃcients. can we choose the control vector u(t) so that x = x1 . linear. so that we can allow for discontinuous changes in its components. As we have seen. u. x2 (t). constant coeﬃcient equations. associated with the control problem. and hence J = t1 . We say that x(t) is the state vector. . Note that there are at most n independent control variables for an nth order linear system of this form. . 14. dx1 = A11 x1 + A12 x2 + B11 u1 (t) + B12 u2 (t). whose components u = (u1 (t).418 TIME-OPTIMAL CONTROL IN THE PHASE PLANE 14. um (t)) are the control variables. u(t). xn (t)) are the state variables. whose components x = (x1 (t). Note that we use the superscript 0 to indicate the initial state and the superscript 1 to indicate the ﬁnal state. t) dt. t1 J= 0 g0 (x(t). 2 × 2 matrices. and u is the control vector. dt where A and B are constant. . so that we seek to minimize the time taken to reach the state x1 . reaching x1 when t = t1 ? If there is a cost function. ﬁrst order. we have an optimal control problem. Consider the system of ordinary diﬀerential equations ˙ x = f (x. . . In particular. in other words. . such that we seek controls u for which J is a minimum. If g0 = 1. dt We can write this more concisely using matrix notation as dx = Ax + Bu. We assume that f is continuously diﬀerentiable with respect to x. linear. dx = Ax + Bu(t). . for ﬁrst order. but that u is merely integrable. the target state.2 First Order Equations A good way of introducing many of the important. basic concepts of control theory is to study one-dimensional systems. t). A control problem takes the form of a question: if x = x0 when t = 0. when t = t1 ? In other words. subject to x(0) = x0 . u and t. This is the type of problem that we will be studying. constant coeﬃcient equations. we have a timeoptimal control problem. u2 (t). the control variables. can the system be controlled from x0 to x1 . dt dx2 = A21 x1 + A22 x2 + B21 u1 (t) + B22 u2 (t). .1 Deﬁnitions Let’s begin by formalizing some of the ideas that we introduced above. dt and for second order. . 4) . u must be negative for at least some of the period 0 t t1 . (14. so that d −t (e x) = e−t u(t).2) Since we want to control the system to x = 0 when t = t1 . time-optimal control. they cannot be controlled. and what should the control be? If there is no attempt to control the deviation. how quickly can it be controlled. x = x0 et and the tightrope walker falls oﬀ.3). Moreover. Can this deviation from the vertical be controlled at all? If so. we can solve (14. The time-optimal control problem is meaningless if the control is unbounded. This is typical of many types of problem where the idea is to control small deviations from equilibrium as quickly as possible – linear.1) using the integrating factor e−t . A very simple model for the tightrope walker is dx = x + u. u = 0. We will see later how to scale a slightly more realistic problem so that u lies in this convenient range. we can see that to push x back towards equilibrium as quickly as possible. so that (14. From now on we restrict our attention to bounded controls with −1 u(t) 1. (14.14. we need to push in the appropriate direction as hard as we can by taking u(t) = −1. Let’s assume that x = x0 > 0 when t = 0.3) gives x0 − 0 t1 e−τ dτ = 0. it is easy to choose u to satisfy (14. since we can control the system in an arbitrarily short time t1 by making |u| suﬃciently large.2 FIRST ORDER EQUATIONS 419 Example: The tightrope walker A tightrope walker is inherently unstable. if her deviations from the vertical become too large. and she falls oﬀ the tightrope.3) Clearly. This is unrealistic. and hence t1 = − log(1 − x0 ). Walking a tightrope is rather like trying to balance a pencil on its tip. Although the walker has an equilibrium position. For a general control u(t). is an unstable equilibrium state. we must have x0 + 0 t1 e−τ u(τ ) dτ = 0. Now. (14. x = 0. if u(t) is not bounded below (the tightrope walker can apply an arbitrarily large restoring force). The upright position. the components of a bounded control vector satisfy −1 ui (t) 1. and she must push herself back towards the vertical to stay upright. dt and hence t x = x0 et + et 0 e−τ u(τ ) dτ (14. In general. for example with u(t) = −et x0 /t1 . since x0 > 0. Intuitively.1) dt where x represents her angular deviation from the vertical and u her attempts to control her balance. it is unstable. 5) For t > t1 we take u = 0. and she falls oﬀ. If . Some time-optimal solutions are shown in Figure 14. we say that the system is completely controllable. The controllable set at time t1 is the set of initial states that can be controlled to the origin in time t1 . if x0 < 0. (14. For the tightrope walker problem. The controllable set is the set of initial states that can be controlled to the origin in some ﬁnite time. 14. the system cannot Equation (14. this is the set C(t1 ) = x | |x| 1 − e−t1 . This is just the union of all the controllable sets at time t1 0. For the tightrope walker problem. pushing in the other direction. The tightrope walker system is not completely controllable.1. of the tightrope walker problem. x = 0. and the system remains in its equilibrium state. the controllable set is C= t1 0 C(t1 ) = {x | |x| < 1} . 1. Similarly.4) shows that t1 → ∞ as x0 → 1− . Some time-optimal solutions. This is a good point at which to deﬁne the reachable set.5). If C is the whole real line.1).420 TIME-OPTIMAL CONTROL IN THE PHASE PLANE with solution x(t) = 1 + (x0 − 1)et . since C = R (see Exercise 14. (14. the time-optimal control is u(t) = 1.1. C = R. We can now make a couple of deﬁnitions. If x0 be controlled back to equilibrium. Fig. The tightrope walker cannot push hard enough. 6) 1. From (14. since |u(t)| x1 e−t1 − x0 = 0 t1 1. whilst for |x0 | < 1 the origin is reachable for − log(1 − |x0 |). R(t1 . For |x0 | set does not contain the origin. x1 = x0 et1 + et1 0 t1 e−τ u(τ ) dτ. x0 ). t1 The reachable set is shown for two diﬀerent cases in Figure 14. x1 x1 t1 Fig.2.2. the reachable deﬁnes the points in the reachable set. consistent with what we know about the controllable set.2). 14. R(t1 . the reachable set in time t1 . The reachable set for the tightrope walker problem lies between the curved lines. . x0 ). a fact that will prove to be important later. e−τ u(τ ) dτ 0 t1 e−τ |u(τ )| dτ 0 t1 e−τ dτ = 1 − e−t1 .14. Note that the boundaries of the reachable set are given by the solutions with the controls u(t) = ±1. is the set of points x1 for which there exists a control u(t) such that x(t1 ) = x1 .2 FIRST ORDER EQUATIONS 421 x = x0 when t = 0. (14. and hence. so that (x0 − 1)et1 + 1 x1 (x0 + 1)et1 − 1. 3 Second Order Equations For second order. constant coeﬃcient systems. Some examples of convex and nonconvex sets of points in the plane.3. We will show that our intuitive notion that only the extreme values of the bounded control are used in the time-optimal control is correct. This is known as bangbang control. if S is a convex set and x. linear. S is connected and any closed loop lying in S can be shrunk continuously to a point without leaving S. in the (x1 .3 for some examples). is one for which the line segment between any two points in S lies entirely within S (see Figure 14. 14. All of the following can be generalized to systems of higher order.422 TIME-OPTIMAL CONTROL IN THE PHASE PLANE 14. S. Fig. 14. y ∈ S.3. Before we can do this. we can examine the behaviour of the state variables. . x2 )-phase plane (see Chapter 9). there are a few mathematical ideas that we must consider.1 Properties of sets of points in the plane — A convex set. Note that necessary but not suﬃcient conditions are that S must be both connected (any two points in S can be joined by a curve lying within S) and simply-connected (S has no holes†). Formally. cx + (1 − c)y ∈ S for all c such that 0 c 1. x1 and x2 . and some of it to nonlinear. nonautonomous systems. † More formally. some of the tangents to the set will meet the boundary at more than one point along a straight part of the boundary (see Figure 14. for each pair of points in S. An open set has S = Int(S).4). the set of points with |x| 1 (the closed unit disc) is closed. of a set S is the set of all the exterior points of S.5). all of which lie in S (see Figure 14. — The interior. — A closed set contains all of its boundary points.4. If S is convex.4). the line segment joining them is entirely made up of interior points (see Figure 14. and can therefore be written as ∂S = {x | (x ∈ Int(S))} ∩ x | (x ∈ Int(S C )) . — A boundary point of S is a point.14. it is said to be open. For example. — A set S is strictly convex if. . — The exterior. Ext(S). of a set S is the set of all the interior points of S. Fig. Int(S). the complement of S. For a strictly convex set.3 SECOND ORDER EQUATIONS 423 — An interior point of S is a point x ∈ S for which there exists a disc of points centred on x. Note that the boundary of the open unit disc is the unit circle. the set of points with |x| < 1 (the open unit disc) is open. 14. Note that all discs centred on a boundary point of S contain a point that is not in S. — An exterior point of S is a point in the interior of S C . |x| = 1. the tangent to any boundary point does not meet S at any other boundary point. and discs centred on them. Examples of interior and boundary points of S. but not strictly convex. that lies in neither the interior nor the exterior of S (see Figure 14. For example. not necessarily in S. — If S does not contain any of its boundary points.6). — The boundary of a set S is the set of all boundary points of S. 14. the unit matrix. 14. we deﬁne the matrix exponential of At to be ∞ exp(At) = k=0 Ak tk . dt (14. We can see immediately that d exp(At) = dt ∞ k=1 Ak tk−1 = A exp(At). k! (14.5. Note that this power series is convergent for all A and t. (k − 1)! a property that the matrix exponential shares with its scalar counterpart. subject to x(0) = x0 .7) with A an n × n matrix of constants. .2 Matrix solution of systems of constant coeﬃcient ordinary diﬀerential equations Consider the system of n diﬀerential equations dx = Ax + b(t). In order to be able to write the solution of this equation in a compact form. Fig.424 TIME-OPTIMAL CONTROL IN THE PHASE PLANE Fig. 14. but not strictly convex sets.6.8) with A0 = I. Examples of tangents to convex and strictly convex sets. eat . Examples of strictly convex and convex.3. so that (14. note that we know from the Cayley–Hamilton theorem (Theorem A1.14. A .2). exp(A) exp(B) = exp(B) exp(A) unless AB = BA (see Exercise 14. an example of which is given by (14. A. . and hence so can exp(At). using the product rule. dP = A exp(At) exp(−At) − exp(At)A exp(−At) = 0. ˙ ˙ . Example: Simple harmonic motion Simple harmonic motion with angular frequency ω is governed by d 2x + ω 2 x = 0. the solution can be written as x = exp(At)x0 . {exp(At)} at −at −1 = exp(−At). . and hence from its Taylor series expansion. d n P/dtn = 0. that. Firstly. dt using the fact that. in line with the result e e = 1 for the scalar exponential function. for an n × n matrix A and k n. In particular. dt2 We can write this as a system of two ﬁrst order equations in the usual way as x1 = x2 . (14. A exp(At) = exp(At)A. Secondly. Similarly. x2 = −ω 2 x1 . Note that. t x = exp(At) x0 + 0 exp(−Aτ )b(τ ) dτ . . I. we can use a matrix integrating factor to write d dx − exp(−At)Ax = {exp(−At)x} = exp(−At)b.8). for a 2 × 2 matrix.9) This is the generalization to an nth order system of the ﬁrst order solution. from the deﬁnition. we note that P (0) = I. . A exp(At) is a linear combination of I and A. In other words. When the equation is inhomogeneous. Ak can be written as a linear combination of 2 n−1 . again.1) which states that every matrix satisﬁes its own characteristic equation. (14. P (t) = I. exp(−At) dt dt and hence.7) is homogeneous.4). These results mean that when b(t) = 0. in general.3 SECOND ORDER EQUATIONS 425 Now consider the product P (t) = exp(At) exp(−At). Finally. 4 Examples of Second Order Control Problems Example 1: The positioning problem Consider the one-dimensional problem of positioning an object of mass m in a frictionless groove using a bounded applied force F (t) such that −Fmax F (t) Fmax . A2n+1 = ω 2n (−1)n A. ˙ ˙ dt2 F (t) Fmax Fmax Fmax x. x2 (0) = x0 . (14. with |u| dx1 dx2 = x2 . x(t1 ) = 0. −ω 2 0 0 −ω 2 = −ω 2 I. x2 = x. With the initial condition x1 (0) = x0 . and hence ∞ ∞ ∞ 0 −ω 2 1 0 . Newton’s second law gives F (t) = m d 2x .426 TIME-OPTIMAL CONTROL IN THE PHASE PLANE with x = x1 . exp(At) = k=0 ∞ Ak t k A2n t2n A2n+1 t2n+1 = + k! (2n)! (2n + 1)! n=0 n=0 ∞ =I (−1)n ω 2n t2n (−1)n ω 2n t2n+1 +A (2n)! (2n + 1)! n=0 n=0 1 sin ωtA = ω cos ωt −ω sin ωt 1 sin ωt ω cos ωt = cos ωtI + . 1 dt dt . x(t1 ) = 0. x0 = 1 m m Fmax m 1. In terms of the generic equation (14. x1 (0) = x0 . x(0) = 0. . x2 (0) = 0.10) minimizing t1 . If we let x1 = we have a problem in the standard form.7). subject to x(0) = X. = u. this means that the solution is 1 2   x0 2 0 x = x0 exp(At) =  x1 cos ωt + ω sin ωt  . A= We now note that A2 = This means that A2n = ω 2n (−1)n I. u = ˙ X. −ωx0 sin ωt + x0 cos ωt 1 2 14. but for now let’s just construct the solution. so that x2 (0) = 0. We have omitted this for clarity.14. We will return later to the problem of determining the time-optimal control if the particle is not initially stationary. Let’s now think what the optimal control might be. We will prove later that this is indeed the optimal method of control. This gives T = 1 t1 . 0 t T . and t1 .7. The full solution is 1  x0 − 1 t2 for 0 t x0 .14) 2 x0 . we can † Note that. for x1 t 2 x1 1 2 t − 2 x1   x2 =  −t t − 2 x0 1 for 0 for x0 1 t t x0 . (14. which gives us x1 = 1 2 t − 2T t + T 2 + x0 . then decelerate as strongly as we can by pushing as hard as possible in the opposite direction for T t t1 . x2 = −T when t = T 1 2 to ﬁx the constants of integration.13) 1 2 0 t + 2x0 0 0. since there is no dependence on a second control function in this problem. x2 = t − 2T for T 1 2 t t1 . If x1 represents the deviation of the ship from a straight path. the time at which we have to switch from accelerating as hard as possible to decelerating as hard as possible. but it must be at rest when we get it there. x2 )-phase plane in Figure 14. B= 0 1 . We can obtain this by using x1 = x2 = 0 when t = t1 . and tends to drift oﬀ course if it is not controlled. we take u = −1 for 0 t T . Presumably we should push as hard as we can for some period of time. Assuming that x0 > 0. We want to push the particle to the origin as quickly as possible. and various solutions plotted in the (x1 . so that the particle is at rest at the origin when t = t1 . strictly speaking.  1 1 2 x1 = (14. We can easily integrate 1 (14. Example 2: The steering problem / The positioning problem with friction The forced motion of a ship is unstable. . and t1 = 2 x0 . 1 A typical solution is plotted as a function of t. (14. and use x1 = x0 − 1 T 2 .4 EXAMPLES OF SECOND ORDER CONTROL PROBLEMS 427 In terms of our matrix notation† A= 0 0 1 0 . x2 = −t for 0 1 2 t T. the total time taken to control the particle to the origin. the matrix B should have another column of zeros. so that the periods of acceleration and deceleration are 2 equal.10) and ﬁnd that 1 x1 = x0 − t2 .12) Now we just need to ﬁnd T .11) Then we take u = 1 for T t t1 . 1 (14. can be written as dx1 dx2 = x2 .15) The drag force caused by the resistance of the water to the lateral motion of the ship is represented by the term −qx2 .16) . = −qx2 + u. A= 0 0 1 −q . (14. dt dt (14. = −x1 + u.7. u(t). This is exactly the same as the positioning problem. Example 3: Controlling a linear oscillator The equation for a linear oscillator (simple harmonic motion) of unit angular frequency. with q a positive constant. In terms of our matrix notation. A= 0 1 −1 0 . B= 0 1 . dt dt In terms of our matrix notation. B= 0 1 . but in a groove with friction. 14. write a simple model as dx1 dx2 = x2 . Some time-optimal solutions of the positioning problem when the particle is initially stationary. subject to an external force.428 TIME-OPTIMAL CONTROL IN THE PHASE PLANE Fig. A= 0 0 1 0 .8). — C is open if and only if 0 ∈ Int(C). we often want to construct C(t1 ). dt dt In terms of our matrix notation. the controllable set for any time t1 . and of C. Example 4: The positioning problem with two controls Suppose that in the positioning problem we are also able to change the velocity of the particle through an extra control. nonautonomous systems of ordinary diﬀerential equations. Any state controllable to zero in time t1 is also controllable to zero in time t2 > t1 . There are a number of statements that we can make about the geometry of these sets. so that dx2 dx1 = x2 + u1 . 1 2 (14.1 If t1 < t2 .18) With the exception of the steering problem (see Exercise 14. If we apply the control u= v(t) 0 for 0 t for t1 < t t1 . In other words. Theorem 14. the trajectory reaches x = 0 when t = t1 . the controllable set. — C(t) and C are connected sets.5 Properties of the Controllable Set Apart from determining the time-optimal control. B= 1 0 0 1 . x1 (t1 ) = x2 (t1 ) = 0.14. We can think of this as the problem of stopping a child on a swing as quickly as possible. we also have that C(t) and C are symmetric about the origin and convex. C(t1 ) ⊂ C(t2 ).17) minimizing t1 . Note that this also implies that these sets are simply-connected. Proof Let x0 be a point in C(t1 ). t2 . which means that C(t1 ) ⊂ C(t2 ) (see Figure 14.9). — If t1 < t2 . constant coeﬃcient equations. = u2 .16) subject to x1 (0) = x0 . u = 0 is an equilibrium state. These results hold for nonlinear. Therefore x0 ∈ C(t2 ). since all convex sets are simply-connected. we will solve all of these problems below after we have studied some more of the theory relevant to this type of control problem. C(t1 ) ⊂ C(t2 ). since x = 0. we must solve (14. For linear. the controllable set never gets smaller as t increases. . (14.5 PROPERTIES OF THE CONTROLLABLE SET 429 If the oscillator is not initially at rest and we wish to bring it to a halt as quickly as possible. with control u = v(t). x2 (0) = x0 . then remains there for t1 t t2 . 14. All points on controllable trajectories are controllable. C(t1 ) ⊂ C(t2 ). Therefore y ∈ C(τ2 −τ1 ) and. If t1 < t2 . . y ∈ C(t1 ). with control u(t). the system follows the same trajectory. x = X(t + τ1 ). X(τ2 ) = 0.430 TIME-OPTIMAL CONTROL IN THE PHASE PLANE Fig. X(τ1 ) = y. and when t = τ2 . Theorem 14.9. and reaches x = 0 when t = τ2 −τ1 . τ2 t1 (see When t = τ1 .8. 14. In other words. 14. all points on controllable trajectories are controllable. with τ1 Figure 14. y ∈ C(t1 ).9). If we now consider the solution with control v(t) = u(t + τ1 ) and Fig. by Theorem 14. Proof Let x = X(t) be the trajectory containing both x0 and y.1.2 If x0 ∈ C(t1 ) and y is a point on the trajectory from x0 to 0. initial condition x(0) = y. each of these trajectories lie in C(t1 ).2. as shown in Figure 14.14. and hence. we can also ﬁnd a ˆ control v(t) that steers y(t1 ) to 0 in some time t2 . Therefore y0 can be controlled to the origin in time t1 + t2 using the control u(t) = v(t) for 0 t ˆ v(t) for t1 < t t1 . 14. Therefore y0 ∈ C(t1 + t2 ) ⊂ C. t2 . Proof If C is open. that lies entirely within C. which we write as D(0. r) on a path y(t) with y(t1 ) ∈ D(0.4 C is open if and only if 0 ∈ Int(C). By continuity of the solutions of the underlying diﬀerential equations. the control v(t) steers y0 into the disc D(0. by deﬁnition. r) ∈ C. C is therefore open. trajectories that connect each point to the origin (see Figure 14. Theorem 14. and let y0 be another point within this disc. Now suppose that u = v(t) is a control that steers some point x0 to 0 in time t1 . D(x0 . . r0 ) ∈ C for all x0 ∈ C.10. and hence. r0 ) be a disc of radius r0 centred on x0 . by deﬁnition. all points on Fig.5 PROPERTIES OF THE CONTROLLABLE SET 431 Theorem 14. Therefore the union of these two trajectories is a curve made up of points in C(t1 ) that connects x0 and y0 . by deﬁnition there is a disc of radius r centred on 0. if r0 is suﬃciently small. so clearly 0 ∈ Int(C).3 C(t1 ) and C are connected sets. Since D(0. C(t1 ) and C are connected sets. Since C = t1 0 C(t1 ). r). Proof If x0 ∈ C(t1 ) and y0 ∈ C(t1 ). r) at some time t1 . It is less straightforward to prove that 0 ∈ Int(C) implies that C is open. C is also connected. By deﬁnition. for r0 suﬃciently small.11. C(t1 ) is connected. By Theorem 14. all of its points are interior points. If 0 ∈ Int(C). Let D(x0 . there are.10). 11. Theorems 14. ui (t) 1 is crucial here. nonautonomous systems of ordinary diﬀerential equations. dt with A and B constant matrices. We now focus on linear. 14. This means that x0 ∈ C(t1 ) if and only if there is a control u(t) such that x0 = − 0 t1 exp(−Aτ )Bu(τ ) dτ. † Note that the fact that the control variables are scaled so that −1 . Recall that if dx = Ax + Bu. Therefore.19) Theorem 14. (14. the solution is t x = exp(At) x0 + 0 exp(−Aτ )Bu(τ ) dτ . cx0 + (1 − c)x1 = − 0 t1 exp(−Aτ )B cu0 (τ ) + (1 − c)u1 (τ ) dτ.4 all hold for nonlinear.5 C(t1 ) is symmetric about the origin and convex.† Therefore C(t1 ) is symmetric about the origin. Proof If x0 ∈ C(t1 ) with control u(t). C is open if and only if 0 ∈ Int(C). constant coeﬃcient equations.432 TIME-OPTIMAL CONTROL IN THE PHASE PLANE Fig.3 and 14. Now note that the set of bounded controls U = {u(t) | −1 ui (t) 1} is convex. if u0 and u1 are controls that steer x0 and x1 to the origin in time t1 .19) shows that −x0 ∈ C(t1 ) with control −u(t).2. since if u0 ∈ U and u1 ∈ U . (14. cu0 + (1 − c)u1 ∈ U for 0 c 1.1. 14. 14. 12). since det(A − λI) = λ2 .3 tells us that C(t1 ) ⊂ C(t2 ) for t1 < t2 . B= 0 1 .6 C is symmetric about the origin and convex. The line segment that joins x0 to x1 therefore lies entirely within C(t1 ). Although a union of convex sets is not necessarily convex (see Figure 14. rank(M ) = 2. Although we will not do so here. Proof A union of symmetric sets is symmetric.6 The Controllability Matrix For a second order system.6 THE CONTROLLABILITY MATRIX 433 and hence the control cu0 (τ ) + (1 − c)u1 (τ ) ∈ U steers cx0 + (1 − c)x1 to the origin in time t1 . A union of convex sets is not necessarily convex. Fig. 14. .12. C is a union of nested convex sets and therefore is itself convex. since its columns are linearly independent. Theorem 14.20) We will show that the system is completely controllable (C = R2 ) if and only if (i) rank(M ) = 2. Clearly.14. (ii) all the eigenvalues of A have zero or negative real part. Example: The positioning problem For the positioning problem A= and hence M= 0 1 1 0 . (14. we deﬁne the controllability matrix to be M = [B AB]. Also. so C = t1 0 C(t1 ) is symmetric. since Theorem 14. by deﬁnition. and hence. 0 0 1 0 . it is straightforward to generalize this result to nth order systems. 14. C(t1 ) is convex. 13). Theorem 14. . Let’s now prove that these properties of the controllability matrix do determine whether or not a second order. 1. y. Now suppose that 0 ∈ Int(C). Since C(t1 ) is convex (Theorem 14.434 TIME-OPTIMAL CONTROL IN THE PHASE PLANE the matrix A has the repeated eigenvalue zero. yT Ak B = 0 for k = 0. Since C(t1 ) ⊂ C. If rank(M ) = 1. if x0 ∈ C(t1 ). . zT x0 = z · x0 0 (see Figure 14.5).6). The tangent and outward normal to the set C(t1 ) at the origin. the origin must be a boundary point of C(t1 ). If rank(M ) = 0. and hence 0 ∈ Int(C). . This means that yT B = yT AB = 0. Now. There are therefore no controls. This means that yT exp(−Aτ )B = 0. x0 lies on the straight line through the origin perpendicular to y. there is a single direction.4 are also completely controllable (see Exercise 14.19 then shows Fig. (14. . orthogonal to every column of M .19) shows that yT x0 = y · x0 = − 0 t1 yT exp(−Aτ )Bu(τ )dτ = 0. We conclude that the system is completely controllable.13. All of the other examples that we described in Section 14. 2. and for all x0 ∈ C(t1 ). and hence 0 ∈ Int(C). C = {0}. linear. Equation 14. and hence by the Cayley–Hamilton theorem. which is a closed set. Since 0 ∈ C(t1 ). .7 0 ∈ Int(C) if and only if rank(M ) = 2. there is a tangent to C(t1 ) through 0 with outward normal z. and hence so does B. 14. 0 ∈ Int(C(t1 )) at any time t1 . constant coeﬃcient system is controllable. M has only zero entries. Therefore if rank(M ) = 1. Proof Suppose that rank(M ) < 2. There is then a tangent to C.9 If rank(M ) = 2 and A has at least one eigenvalue with positive real part. Since the system can only be completely controllable if 0 ∈ Int(C). z = 0 for 0 t t1 . for t1 suﬃciently 0 large. This is a contradiction. 0 1 (|z1 | + |z2 |) dτ → ∞ as t1 → ∞. so that n · x0 = − 0 t1 t1 t1 nT exp(−Aτ )Bu(τ ) dτ =− 0 z(τ )u(τ ) dτ = 0 (|z1 | + |z2 |) dτ. . Since this must hold for all controls u. In particular. If any eigenvalue has zero part. the corresponding component of z will be either a polynomial or a periodic function of t. 0 for each eigenvalue λi of A. the system is not completely controllable. Suppose that rank(M ) = 2 and A has eigenvalues with zero or negative real parts. with n · x0 p and n · y p (see Figure 14. since −u is also an admissible control. Consider a point y ∈ C. and hence rank(M ) < 2. and we conclude that C = R2 . However. Let z = n {exp(−At)B}.8 If rank(M ) = 2 and Re(λi ) system is completely controllable.6 THE CONTROLLABILITY MATRIX 435 that t1 0 zT exp(−Aτ )Bu(τ ) dτ 0 for all controls u. − 0 t1 zT exp(−Aτ )Bu(τ ) dτ 0. but that C = R2 . n · x > p. zT exp(−At)B = 0 for 0 t t1 . Theorem 14. If we now set t = 0 we get zT B = 0.14). this theorem shows that a necessary condition for the system to be completely controllable is that rank(M ) = 2. Now choose a control with components ui (t) = −sgn(zi (t)). zT AB = 0. By choosing an appropriate coordinate system. that separates y from each x0 ∈ C. we can make each component of z a sum of terms proportional to e−λi τ .14. and hence t1 0 zT exp(−Aτ )Bu(τ ) dτ = 0 for 0 t t1 . with equation n · x = p. and if we diﬀerentiate and put t = 0. t In each case. This means that z is orthogonal to all of the columns of M . the Proof We proceed by contradiction. Theorem 14. Because rank(M ) = 2. R(t. Proof Let λ be an eigenvalue of A with positive real part.14. Proof Let u = u∗ (t) be the optimal control for x(0) = x0 . x0 ). t)-space. The controllable set therefore lies on one side of a line in the plane.7 The Time-Optimal Maximum Principle (TOMP) Consider the set reachable from x0 in time t. This integral converges as t1 → ∞. ∂RT (t. Therefore y(t1 ) ∈ Int(R(t1 .10 The time-optimal trajectory lies in the boundary. Now suppose that y(t0 ) ∈ Int(R(t0 . x0 )). x2 . If we now apply the optimal control. As time increases. so that eT A = λeT . so that e · x0 c. and let y(t) be a solution with u = u∗ . D(y(t0 ). A tangent to the set C that separates each x0 ∈ C from y ∈ C. u∗ . 14. and is bounded above by some constant c.436 TIME-OPTIMAL CONTROL IN THE PHASE PLANE Fig. x0 )). which we label RT (t. so that any trajectory that starts in . r). r). and must also lie in R(t1 . x0 ). that lies entirely within R(t0 . x0 ). and e the associated eigenvector. where t∗ is the ﬁrst time when t∗ ∈ R(t∗ .15). the shortest time in which this can be achieved is t∗ . x0 ). and eT Ak = λk eT . This means that eT exp(−Aτ ) = e−λτ eT . Theorem 14. This concludes our proof that the controllability matrix has the properties that we outlined at the start of this section. to all of the points within D(y(t0 ). x0 ). There must therefore be a disc of suﬃciently small radius r. 14. and hence eT x0 = e · x0 = − 0 t1 e−λτ eT Bu(τ ) dτ. they will lie in the neighbourhood of y(t1 ) when t = t1 > t0 . If the system can be controlled to the origin. and hence C = R2 . x0 ) (see Figure 14. this set traces out a volume in (x1 . The reachable set. ∂RT (t. If v(t) is an arbitrary control and y(t) the corresponding solution. y1 = exp(At1 ) x0 + 0 t1 exp(−Aτ )Bv(τ ) dτ .11 (The time-optimal maximum principle (TOMP)) The control u(t) is time-optimal if and only if there exists a nonzero vector. 2. Theorem 14. . v(t) The components of the time-optimal control are ui (t) = sgn hT {exp(−At)B} – bang-bang control. .5). RT (t. the time-optimal trajectory must lie entirely within this boundary. . x0 )) remains there. x0 ).x0 ) i for i = 1.15. m n · y1 . Since the origin lies in the boundary. Proof Let u(t) be a control that steers x from x0 when t = 0 to x1 ∈ ∂R(t1 . so there is a tangent at x1 with normal n such that n·x= sup y1 ∈R(t1 .14. Int(R(t. x0 ). . The reachable set is convex (this can be proved in the same way that we proved that the controllable set is convex in Theorem 14. h. . such that for 0 t t1 . x0 ) when t = t1 . 14.7 THE TIME-OPTIMAL MAXIMUM PRINCIPLE (TOMP) 437 Fig. hT {exp(−At)Bu(t)} = sup hT {exp(−At)Bv(t)} . (14. . 1 or −1. we must take vi (t) = sgn hT exp(−At)B i for 0 t t1 . We shall see that this is the main result that we need to solve time-optimal control problems in the phase plane. because exp(At) is a nonsingular matrix since it always has an inverse. If we let hT = nT exp(At1 ). each component of the time-optimal control must take one of its extreme values. 1 0 −t 1 .22). Since (14. exp(−At). and change when hT exp(−At)B i changes sign – bangbang control. we must have hT exp(−Aτ )Bu(τ ) = sup v hT exp(−Aτ )Bv(τ ) . and we have therefore proved the ﬁrst part of the theorem.21) Note that h is nonzero.21) holds for all t1 > 0. To obtain the maximum value of the right hand side of (14. In other words.22) All of the steps of this argument can be reversed. B= 0 1 . 0 1 0 0 . Example: The positioning problem For this problem A= Since A2 = 0. and hence nT exp(At1 ) 0 t1 t1 exp(−Aτ )Bu(τ ) dτ = sup v nT exp(At1 ) 0 exp(−Aτ )Bv(τ ) dτ. we ﬁnd that t1 0 hT exp(−Aτ )Bu(τ ) dτ = sup v 0 t1 hT exp(−Aτ )Bv(τ ) dτ . This means that n · exp(At1 )x0 + n · exp(At1 ) 0 t1 t1 exp(−Aτ )Bu(τ ) dτ = sup v n · exp(At1 )x0 + n · exp(At1 ) 0 exp(−Aτ )Bv(τ ) dτ . exp(−At) = I − At = and exp(−At)B = −t 1 .438 TIME-OPTIMAL CONTROL IN THE PHASE PLANE and y1 = x1 when v = u. (14. and must u(t). so the integral paths are given by dx1 = ±x2 . Rather than just solving the equations. The TOMP therefore shows that the time-optimal control is u(t) = sgn(β cos t − α sin t) = sgn {a cos(t + b)} . or even whether. hT exp(−At)B = β cos t − α sin t. Let’s now consider what happens if the particle is initially moving. cos t sin t − sin t cos t . if we write hT = (α. this change in sign occurs. as illustrated in Figure 14. β). with x1 (0) = x0 and x2 (0) = x0 . We also constructed the unique solution with bang-bang control that changes sign just once. whilst on S− we need u = −1. The TOMP does not tell us when. We know that the time-optimal control is bang-bang with u = ±1. we ﬁnd that hT exp(−At)B = β − αt. for some constants a and b.7 THE TIME-OPTIMAL MAXIMUM PRINCIPLE (TOMP) 439 Therefore. Since cos(t + b) changes sign at intervals of π. The TOMP now shows that this is indeed the time-optimal solution.14. The ﬁrst change of sign occurs when t = T0 with 0 < T0 − sin t cos t . exp(−At) = and exp(−At)B = With hT = (α. the system must be controlled onto one of the curves S± . This means that the time-optimal. which are labelled as 2 2 S± in Figure 14.16.16. Any initial conditions that lie on these curves can be controlled to the origin without changing the sign of u. dx2 and hence are parabolas. The only two time-optimal paths that reach the origin are therefore the appropriate branches of x1 = ± 1 x2 . let’s think about what 1 2 the time-optimal solution will look like in the phase plane. when the control must change sign. an initially stationary particle. On S+ we need u = 1. For any other initial conditions. . However. with 1 x1 = k ± x2 2 2 for some constant of integration k. bang-bang control changes sign at most once. with x1 (0) = x0 and x2 (0) = 0. so must π. we have 1 seen that the control must change sign at least once if the solution is to reach the origin. β). Example: Controlling a linear oscillator For a linear oscillator with unit angular frequency. we would perhaps have expected the control to change sign when x2 = 0. subsequently change when t = Tn = T0 + nπ. dt dt . The governing equations. A typical example is shown in Figure 14. In fact. Let’s consider the solution when u = ±1 in the phase plane. . . These trajectories are circles. 0) and (−1. with the control changing sign there. and a succession of semicircles of increasing radius with centres alternating between (1.17. 2 enter the origin. 2. Proceeding as we did for the positioning problem. . the time-optimal solution is out of phase with this intuitive solution in order to allow the velocity to be zero precisely when the swing is vertical. as shown in Figure 14.16). centred on (±1. Thinking in terms of stopping a child on a swing. show that (x1 ∓ 1) and hence that (x1 ∓ 1)2 + x2 = k 2 . we might have expected to push in the opposite direction to the motion.440 TIME-OPTIMAL CONTROL IN THE PHASE PLANE Fig. dx2 dx1 + x2 = 0. The time-optimal trajectories for the positioning problem. . 0).18. 14. The sign of the control must change with period π. k. for n = 0. 0). We can now construct the time-optimal solution by considering the solutions that meet S± . Intuitively. 2 for some constant of integration. so the time-optimal solution consists of part of S+ or S− . 1.16.17. we can see that only the circles marked S± in Figure 14. (14. given by (x ∓ 1)2 + x2 = 1. we obtain 2 1 x1 = x0 + x0 t − t2 + 1 2 2 t1 1 0 −t 1 . This immediately gives us x2 = x0 − t. In contrast. The optimal control time is 2 therefore t1 = x0 . If we now integrate the equation for x1 . u1 (τ ) dτ. 14. 0 .7 THE TIME-OPTIMAL MAXIMUM PRINCIPLE (TOMP) 441 Fig. β). however. provided that α = 0.17. If. 2 which is therefore only appropriate when x0 > 0. β − αt). Example: The positioning problem with two controls For this problem exp(−At)B = With hT = (α.14. as was the case for the positioning problem with just one control variable. u1 = sgn(α) does not change sign in the time-optimal control. with u2 = −1. u2 does not change sign and u1 is undetermined. The bang-bang trajectories for the problem of controlling a linear oscillator. Let’s begin by considering this case. hT exp(−At)B = (α. The TOMP therefore says that the time-optimal control u2 = sgn(β − αt) changes sign at most once. α = 0. 14. where 1 x0 + (x0 )2 2 2 2 x0 1 1 −x0 + (x0 )2 . as marked in Figure 14. 2 and hence when 1 −x0 − (x0 )2 2 2 2 x0 1 1 x0 − (x0 )2 . since |u1 | 1. where the sign of u1 changes. since t1 = |x0 | in regions A and B where the time-optimal solution 2 . this can only be achieved when 1 −x0 − (x0 )2 1 2 2 x0 . Since x1 (t1 ) = 0. 2 2 2 Outside these two regions. and then follows the boundary of region A to the origin. 1 2 2 However. Similarly in region D. when x0 < 0 and u2 = 1 we 2 have nonunique time-optimal paths in region A. each time-optimal path meets the boundary of region A. In region C.18. the time-optimal control is unique. Similarly.19.19. we have a time-optimal trajectory for any control u1 such that x0 2 0 1 u1 (τ ) dτ = −x0 − (x0 )2 . A typical time-optimal trajectory for the problem of controlling a linear oscillator.442 TIME-OPTIMAL CONTROL IN THE PHASE PLANE x2 S ¯ x x x1 Fig. with u1 = u2 = −1 in region C and u1 = u2 = 1 in region D. 2 2 2 This is marked as region B in Figure 14. Note that. 14. 2) when t1 < t2 . x = x + u(t) with |u| ˙ 1. In practice. 2) and the set controllable to x = 2 in time t1 . and some time-optimal solutions for the positioning problem with two controls. 2) ⊂ C(t2 . Show that R(t1 . 14. C(t1 . constant coeﬃcient systems. is not unique. linear. 2). ﬁnd the reachable set from x = 2 in time t1 . 2) ⊂ R(t2 . and sketch them. that the other two qualitatively diﬀerent. x = x+u. ˙ ˙ For the tightrope walker system. by solving the governing equations. the boundaries of C(t1 ). are straight lines in these two regions. the diﬃculty with applying bang-bang control lies in determining when the control needs to change by making measurements of the state of the system. the reachable set in time t1 . Although we will not discuss this further. This is known as measurement–action lag (see. This means that C(t1 ) is not strictly convex. are completely controllable. C and D.2 . Exercises 14. Show.EXERCISES 443 Fig. 1995). B.19. ﬁrst order. is not completely ˙ controllable. Marlin. for example. controllable sets that are not strictly convex are always associated with nonunique time-optimal solutions.1 We have seen that the tightrope walker system. 2) and C(t1 . R(t1 . The regions A. x = −x + u and x = u. 2 1 2 (x1 .7 Use the controllability matrix to show that Examples 2.8 Consider the positioning problem. −1 B= B= B= 1 0 1 1 1 1 0 1 . exp(At) and exp(−At). 2 1 . x2 ) | x1 0} . 1 2 14.444 TIME-OPTIMAL CONTROL IN THE PHASE PLANE 14. 1 2 2 . 14. 1 x0 + (x0 )2 . 2 1 2 (x1 . dt subject to x(0) = x0 = (x0 . x2 ) | x1 > 0} .6 14.3 Consider the system of ordinary diﬀerential equations dx = Ax + b(t).4 are completely controllable systems.4 Show that exp(At) exp(Bt) = exp((A + B)t). 14. 14. convex but not strictly convex. Classify each of the following sets of points in the plane ﬁrstly as either open. ˙ If x = Ax+Bu with u the vector of bounded controls. or not convex. (a) (b) (c) (d) (x1 . and hence write down both the general solution and the solution when b is independent of t when A = (i) 0 1 0 0 (ii) 1 0 0 2 (iii) 1 −q 0 0 . M . . x2 ) | x2 + x2 1 ∩ {(x1 . x0 )T . b2 (t)). x2 ) | x2 + x2 < 1 . 2 1 2 2 Determine C(t1 ). x2 ) | x2 + x2 1 . and secondly as either strictly convex. with b(t) = (b1 (t). Show that for initial conditions lying above the curves S ± the switching time is T = x0 + 2 and that the optimal control time is 1 t1 = x0 + 2 x0 + (x0 )2 . the controllable set in time t1 . or neither open nor closed. 0 1 . x2 ) | x2 + x2 1 ∩ {(x1 . construct the controllability matrix. 1 2 (x1 . x2 ) | (x1 − 1)2 + x2 1 ∩ (x1 .5 if and only if the matrices A and B commute. closed. 3 and 4 given in Section 14. Determine the 1 2 exponential matrices. x2 ) | (x1 − 1)2 + x2 < 1 ∪ (x1 . and hence determine whether the system is completely controllable when (a) (b) (c) A= A= A= 0 0 −1 0 −1 0 1 . . Let the populations of the pest and the beetle be denoted by x1 (t) and x2 (t) creatures per square metre respectively. Show that when x2 is initially positive. ˙ with |u| 1 representing the rate at which beetles are released into the environment. x2 = −x2 + u(t). x1 = X > 0 and x2 = 0. 2(1 + x0 ) 2 . Show that   1 qt 1 (1 − e )  . show that the optimal control 1 time is 1 2 0 2 q x1 + exp (q 2 x0 ) − 1 . a genetically engineered. q2 14. The population of a pest is increasing exponentially. ˙ ˙ subject to x1 (0) = x0 . Since the beetles are harmful to crops. If x1 (0) = x0 and x2 (0) = 0. To control the pest. sterile.11 What is the upper limit on X below which the system is controllable? Consider the system x1 = x1 + x2 . Use the time-optimal maximum principle to show that the optimal control changes sign just once. it is desirable to remove them as soon as possible. and the target is to reduce both populations to zero simultaneously and in the shortest possible time. x2 (0) = x0 . t1 = log exp 1 q 2 Show that the integral paths for u = ±1 are x1 = k − 1 {qx2 ± log (1 ∓ qx2 )} . t1 = log 2X − 1 14. ˙ x2 = −x2 + u(t). predatory beetle is introduced into the environment. x2 )-phase plane. with |u(t)| 1. Initially.9 Consider the steering problem. Sketch the time-optimal trajectories and the controllable set in time t1 in the (x1 .10 with k a constant. the time-optimal solution on which the control does not change sign has x0 = − 1 (x0 )2 2 . Use the time-optimal 1 2 maximum principle to show that the sign of the optimal control changes at most once. exp(−At) =  q 0 eqt Use the TOMP to show that the time-optimal control has at most one change of sign. Model equations for the two populations are x1 = x1 − x2 .EXERCISES 445 14. Show that the optimal control time is √ X + 1 + X 2 + 4X . Consider the second order system of ordinary diﬀerential equations dx = Ax + Bu. Use the time-optimal maximum principle to show that the sign of u(t) can change no more than once in the time-optimal control. dt where x is the vector of state variables. where t1 is the time taken to control the system to the origin.12 Determine the matrices exp(−At) and exp(At) when 1 A= √ 2 1 1 1 −1 . Show that the time-optimal solutions on which the control does not change sign have 1 1 x0 = ± sinh t1 − √ cosh t1 + √ 1 2 2 1 1 . u the vector of bounded control variables with |ui (t)| 1.446 TIME-OPTIMAL CONTROL IN THE PHASE PLANE 14. x0 = ± − √ cosh t1 + √ 2 2 2 . . and B= 1 0 . 448 AN INTRODUCTION TO CHAOTIC SYSTEMS If we now deﬁne the dimensionless variables x= we obtain ˆ x + δ x − x + x3 = F (t ).3). and arises in many other contexts in mechanics. A mechanical system whose small amplitude oscillations are governed by the forced Duﬃng equation. so that (15. mky y(t) Fig. t = 2a2 (l − a) µ(l − a) t. 15. This solution 2 5 † For more on the dynamics of the forced Duﬃng equation. (15. .2) with F = 0 is known as Duﬃng’s equation. With F = 0.2) becomes x + δ x − x + x3 = γ cos ωt. γ = 3 and y = dy/dt = 0 when t = 0. ¨ ˙ where δ=k ma ¨ ˙ ˆ . ma The reason for our rather odd choice of dimensionless parameters will become clear later. when δ = 1 .2) is the forced Duﬃng equation. ¨ ˙ (15. see Arrowsmith and Place (1990) and references therein.1. (15.2) l ˆ Y.3) φ(t) a m .† We will assume that the forcing takes the form F = γ cos ωt (dropping the hat on the time variable for convenience).3. µ2 l(l − a) (15.2 shows the solution. Figure 15. Equation (15. F (t ) = − φ + k φ µ(l − a) 2m2 a4 . calculated numerically using MATLAB (see Section 9.4). ∼a+ 2 a 2a . After e investigating homoclinic tangles in Poincar´ return maps. The rings are joined by a spring of natural length l > a that obeys Hooke’s law with elastic constant µ. we discuss the computation of the Lyapunov spectrum of a diﬀerential equation. which are important because ordinary diﬀerential equations can be related to mappings through the Poincar´ return map. Let’s assume that the relative displacement of the two rings.1. which contain chaotic e solutions. between the wires. ˙ y (15. Finally. so that a2 + (y − φ)2 = a 1 + In terms of Y .1 A Mechanical Oscillator Consider the mechanical system that consists of two rings of mass m threaded onto two horizontal wires a distance a apart. from which a quantitative measure of chaos can be obtained. We then look at some examples of mappings. a. we investigate how their existence can be established by examining the zeros of the Mel’nikov function. 15. On the assumption that a frictional force of magnitude mk y opposes the motion of the lower ring. and that of the lower ring by y(t). is much less than the distance. what happens to the lower ring? We denote the displacement of the upper ring from a ﬁxed vertical line by φ(t).CHAPTER FIFTEEN An Introduction to Chaotic Systems In order to introduce the idea of a chaotic solution. as shown in Figure 15. ma 2ma3 Y2 Y2 . we will begin by studying three simple chaotic systems that arise in diﬀerent physical contexts.1.1) becomes µ(l − a) µl ¨ ˙ ¨ ˙ Y + kY − Y + Y 3 ∼ −φ − k φ.1 Three Simple Chaotic Systems 15. Y = y − φ. If we move the upper ring.1) where a dot denotes d/dt. (15. Newton’s second law in the horizontal ˙ direction shows that −µ a2 + (y − φ)2 − l (y − φ) a2 + (y − φ)2 − mk y = m¨. Adjacent chaotic solutions diverge exponentially fast. Indeed. However. solutions with diﬀerent initial conditions. they are indistinguishable. when δ = y = dy/dt = 0 when t = 0. unpredictable in the long term. and remain in a bounded region. 2 γ= 3 5 and We can now give an informal deﬁnition of a chaotic solution as a bounded. we know from Chapter 8 that the solution of (15. a property that we will later measure using the Lyapunov spectrum. in practice.2. We say that the system exhibits sensitive dependence upon initial conditions.15. and soon diverge completely.1 THREE SIMPLE CHAOTIC SYSTEMS 449 behaves erratically. we can only know the initial state of a physical system with a ﬁnite degree of accuracy. The two solutions begin close to each other. 15. 1 . After a sufﬁcient time has elapsed. but. . aperiodic.3). that has a random aspect due to its sensitive dependence on initial conditions. The solution of the forced Duﬃng equation. where they undergo repeated folding. (15. Fig. will diverge in a chaotic system. drift further apart. as time increases. it is tempting to think of it as ‘random’ in some sense. but which are close enough together that.3) exists and is unique. dy/dt = 0 when t = 0. this time with y = 10−4 . recurrent solution. and are.3 superimposes another solution. in practice. In practice. Figure 15. which decays to produce A. C → D rate k3 c. dy/dt = 0 when t = 0 (broken line). P + C → A + C rate k1 pc.8) (15.2 A Chemical Oscillator The next example that we will consider is a well-stirred system of reacting chemicals. (15.7) (15. known as the cubic crosscatalator. 15. catalysis of precursor decay. which has at its heart the cubic autocatalytic step that we studied in Chapter 13.9) In addition to the reactant. cubic autocatalysis. A + 2B → 3B rate k1 ab2 . (15.6) (15. uncatalyzed conversion. P. precursor decay. A. when δ = 1 .1. The solution of the forced Duﬃng equation. 15.4) (15. A → B rate ku a. and autocatalyst. and is produced by the decay of the autocatalyst. which accelerates the decay of the precursor. there is a precursor.5) (15. catalyst decay.3). C. The action of C both at . and a catalyst.450 AN INTRODUCTION TO CHAOTIC SYSTEMS Fig. autocatalyst decay. The reaction scheme is P → A rate k0 p. γ = 3 and 2 5 y = dy/dt = 0 when t = 0 (solid line) and y = 10−4 . B.3. B → C rate k2 b. y = z = ±6 2.1 THREE SIMPLE CHAOTIC SYSTEMS 451 the start of the reaction cascade. we can derive the dimensionless governing equations (see Exercise 15. one at the origin and two at x = 27. and κ. β and γ are the dimensionless concentrations of A. Equations (15. Figure 15. and as a product at the end of the sequence P → A → B → C is the essential ingredient that leads to complex behaviour. B and C. The best way to get a feel for the dynamics of the Lorenz equations. associated with complex eigenvalues. and a one-dimensional stable manifold.11) were derived by Lorenz (1963) as the leading order approximation to the behaviour of an idealized model of the Earth’s atmosphere. ˙ 3 y = −10(y − z). catalyzing the decay of the precursor. and χ are dimensionless constants. Under the assumption that the precursor.5 for the solution with x = y = z = 1 when t = 0. we will focus on a single chaotic solution. ˙ (15. The two equilibrium points away from the origin each have a two-dimensional unstable manifold.1) α ˙ ˙ β γ ˙ = κ (1 + ηγ) − αβ 2 − α. = β − χγ. To claim that these simple equations model the weather is perhaps going a little too far.054. P. = 0.71. η = 0. is to type lorenz in MATLAB. is in large excess and decays slowly.25. = αβ 2 − β + α. The solution is continually attracted towards the equilibrium point close to the stable manifold. each of which is unstable. and then spiralling away close to the unstable manifold. which runs an animated simulation. η. There are √ three equilibrium points. The reactant A also decays spontaneously to produce B. The system is rather like two copies of the cubic crosscatalator system interacting with each other. as shown in Figure 15. Typical solutions bounce back and forth between the two equilibrium points away from the origin. Scott and Showalter 1992).15. (15. . but they certainly have very interesting dynamics.005 and χ = 0. decaying to the inert product D. and then spirals away close to the unstable manifold before the process begins again.10) where α. Although this system possesses a very complicated set of diﬀerent types of behavior (see Petrov. 15. and hence oscillatory behaviour. continually being attracted towards an equilibrium point along a trajectory close to the stable manifold.1.11) z ˙ = −xy + 28y − z. and C is itself unstable.3 The Lorenz Equations Our third example is the system of three autonomous diﬀerential equations. 8 x = − x + yz. The single equilibrium point has a one-dimensional stable manifold and a two-dimensional unstable manifold associated with complex eigenvalue. and an indication of their iconic status as the ﬁrst chaotic system to be discovered. known as the Lorenz equations.4 shows the solution when κ = 0. (15.10). 1 = 0. Before we give some basic deﬁnitions. which maps to zero.909090 . . (15. → 11 → . which maps to 0. . .11111 .5. χ = 0. we will see later that it is often helpful to relate the solutions of diﬀerential equations to those of mappings.4. . 15. . The map also has many periodic solutions. 1) to itself. or ﬁxed point of the map. Another equilibrium point of the map is 1 = 0. Example: A shift map The shift map is deﬁned by x → ax|1. for example. .25. . which has period 2. 15.12) which maps [0. .125 maps to 0. Let’s focus on the case a = 10. let’s consider three examples of nonlinear mappings. where b|c means the remainder when b is divided by c.005. The shift map is then equivalent to shifting the decimal point one place to the right and throwing away the integer part.2 Mappings Although this book is about diﬀerential equations. η = 0.054. an 8 equilibrium.09090909 . . For example. = 0. which illustrate how complicated the solutions of these deceptively simple systems can be. → 11 = 0.452 AN INTRODUCTION TO CHAOTIC SYSTEMS Fig. . sometimes known as diﬀerence equations. The solution of the cubic crosscatalator equations.71. 9 1 10 1 11 = 0.25 and α(0) = β(0) = γ(0) = 0. . with κ = 0. π and e. . . .449 . √ for example 2. but the corresponding √ solutions eventually diverge. For √ example. . for example 11 . .15. since they behave in an apparently random manner.4142135 . which correspond to rational numbers with no ﬁnite decimal representation. . . whilst x2 maps to 0. . and therefore is eventually mapped to zero. . 1). The chaotic solutions. Each solution initially at an irrational number therefore satisﬁes our conditions to be called a chaotic solution. (15. x1 maps to 0.5. Every rational number has either a ﬁnite decimal expansion.2 MAPPINGS 453 Fig.4213562 → 0.135 . have decimal expansions that do not repeat themselves √ and have no pattern. but after four iterations of (15. for example 2 − 1 ≈ 0. as are the periodic solutions. 15. . consider x = x1 = 2 − 1 and x = x2 = 2 − 1 + 10−5 π ≈ 0. . . This is a simple example of sensitive dependence on initial conditions. are dense in [0.2135623 → .1421356 . . with x(0) = y(0) = z(0) = 1.12). Irrational numbers. . for example 1 . In addition. which correspond to the irrational numbers. . → 0. two irrational numbers may be arbitrarily close to each other. . These diﬀer by just 10−5 π. The solution of the Lorenz equations.4142449 . and initially close solutions diverge. 8 1 in which case it is part of a periodic solution. → 0.11). or has a repeated decimal expansion. eventually leads to chaotic solutions. Similarly. x0 . xn xn − r−1 r r2 x2 − r(1 + r)xn + (1 + r) = 0.13). The state xn = 0 represents the complete absence of the species. Figure 15. we obtain the equation for xn . (15. xn+2 = xn = rxn+1 (1 − xn+1 ). as r increases. The nontrivial ﬁxed point lies in the meaningful range only if r > 1. The ﬁxed points are therefore x = 0 and x = (r −1)/r. which models the eﬀect of overcrowding and competition for resources. Let’s now determine whether there are any solutions of period 2. The interested reader is referred to Arrowsmith and Place (1990). and a stable 4-cycle emerges. since we want to concentrate on maps relevant to diﬀerential equations. or 2-cycles. but loses stability in a bifurcation† at r = 3. 1. with 0 < r 4 and n an integer. Let’s begin by trying to ﬁnd the ﬁxed points of (15. since we know that it must also be satisﬁed by the two ﬁxed points. Starting from an initial population. The interval [0. it can be shown that the nontrivial ﬁxed point is stable for r < 3. and for xn xn+1 ∼ rxn .454 AN INTRODUCTION TO CHAOTIC SYSTEMS Example: The logistic map Consider the logistic map xn+1 = rxn (1 − xn ). n This equation is easy to factorize.13) gives a measure of the population in subsequent generations. This was produced using the MATLAB script † a ﬂip bifurcation . and. the factor (1 − xn ). determines the size of the next generation.13). then 0 xn 1 for n 0. These satisfy xn+1 = xn . 1] is also the physically meaningful range for this map. nonlinear equation. By eliminating xn+1 .6 shows the period doubling process as a bifurcation diagram. The logistic map is a simple model for the growth of the population of a single species. We will not go into the details of this process. In fact.13) where r is a constant. the points of period 2 eventually lose stability. This process is known as period doubling. which is positive provided r > 3. When xn is not small. Note that if 0 x0 1. The discriminant of the quadratic factor is r2 (r2 − 2r − 3). so that the next generation grows by a factor of r. is no longer close to unity. (15. as r increases. These satisfy xn+1 = rxn (1 − xn ). where the points of period 2 emerge and are stable. and the full. This means there are points of period 2 for r > 3. (15. and hence xn = rxn (1−xn ). Fig. end xout = []. starting from randomly generated initial conditions (x = rand(1)). The bifurcation diagram for the logistic map.6.005:4 x = rand(1). and then saves the next 400 iterates of the map before plotting them. Figure 15. pause(0. for j = 1:400 x = r*x*(1-x).’. for j = 1:100 x = r*x*(1-x). 15.7 shows 100 iterates of the logistic map with r = 4 for two initial conditions separated by just 10−16 .' for r = 0:0.01) end & 15. hold on. .2 MAPPINGS$ 455 % This iterates the map 100 times. The apparently random nature of the solution can be seen.xout.’.3) axis([0 4 0 1]). xout = [xout x]. end plot(r*ones(size(xout)). as can the fact that these initially very close solutions have completely diverged after about 50 iterations of the map.’MarkerSize’.

7. at suﬃciently low ﬂow rates.8. and look for the nearby values of x and r at which f (5) (x) actually touches the straight line.† At suﬃciently high ﬂow rates. A system that behaves in an intermittent manner exhibits bursts of chaotic behaviour interspersed with simpler. . 15. the ﬂow of water down a pipe is smooth and steady. spatially localized bursts (see. We will also discuss another phenomenon that can occur in the logistic map. Consider the ﬁfth iterate of the logistic map. for example. Many other systems. f (5) (x). It is straightforward † Strictly speaking. known as intermittency. which is shown in Figure 15.7. For example. or turbulent. Mathieu and Scott. and hence part of a periodic solution of the logistic map of period 5.456 AN INTRODUCTION TO CHAOTIC SYSTEMS Fig. or laminar. which often begins to occur as a parameter is changed. 2000). and is a prelude to full chaos. at suﬃciently low Reynolds numbers. f (x) = x. the image of which is shown in Figure 15. We will focus on the point close to x = 0.8 for r = 3. Note that there are four points where the curve is close to the straight line that represents the identity mapping. A sequence of 100 iterates of the logistic map with r = 4 and x0 = 0. more regular behaviour. However. at intermediate ﬂow rates.65. but does not touch it. can exhibit intermittency. both maps and ﬂows. turbulence can appear in intermittent.1 + 10−16 .1 and x0 = 0. the ﬂow becomes unsteady and chaotic. This is then a ﬁxed point of f (5) (x).

8 0. see Guckenheimer and Holmes (1983).10.3 0. For further examples of intermittency.6 f (x) 5 0. For values of r slightly greater that rc .8 0. For r slightly less than rc . The solution therefore exhibits almost steady behaviour before moving away and behaving irregularly.1 0. This is intermittency. we can see in Figure 15. that g(x) = f (5) (x) − x and its derivative are zero when r = rc ≈ 3. Figure 15. one stable. which is also shown in Figure 15.9 0.8. Clearly.3 0.7. points initially close to xc are attracted to the ﬁxed point.1 0 0 0.2 0. one unstable. r = rc is a bifurcation point of the map f (5) .15. and also close to the other points where the curve is close to the straight line.9 shows f (5) (x) in the neighbourhood of this point when r = rc and for two values of r slightly above and below rc .4 0. analogous to the saddle–node bifurcation that we discussed in Section 13. as shown in Figure 15. The equivalent solution of the logistic map displays almost periodic behaviour interrupted by chaotic bursts.73817237526634 and x = xc ≈ 0. and is known as a tangent bifurcation.5 0.6 0. whilst for r < rc .65. 15. . and that iterates of the map can be trapped close to x = xc before moving away.9 1 x Fig.66045050397608.2 0.3.2 MAPPINGS 1 457 0. We can see that for r > rc there are two ﬁxed points. using MATLAB. Also shown is the image of the point x = 0.5 0. and therefore the solution of the logistic map is attracted to a stable periodic solution of period 5.10. to show.7 0. locally there are no ﬁxed points.4 0.9 that f (5) (x) is very close to the straight line. The ﬁfth iterate of the logistic map for r = 3.1.7 0.

15.655 0.655 0.68 x (b) r = rc − 1.68 0.5 × 10−3 .645 0.665 f5(x) 0.655 0.67 0.675 0.65 0.67 0.65 0. 0. The iterates of a point in this neighbourhood are also shown.66 0.645 0.64 0.458 AN INTRODUCTION TO CHAOTIC SYSTEMS 0.655 0.64 0.65 0.645 0.68 0.66 0.655 0.675 0.66 0.65 0.66 0.67 0.645 0.65 0.675 0.675 0.67 0.67 0.64 0.73817237526634.665 f5(x) 0.645 0. The function f 5 (x) at rc and rc ± 1.67 0.9.675 0.66 0.66 0.645 0.665 f5(x) 0.68 x (c) r = rc + 1.68 x (a) r = rc ≈ 3.5 × 10−3 Fig.64 0.5 × 10−3 0. .655 0.675 0.68 0.64 0.64 0.665 0.665 0.665 0.65 0.

10.3 0.5 0.6 x 0.2 0.6 n 0.15.3 0.9 0. n yn+1 = xn sin θ + yn cos θ − x2 cos θ.8 0.7 0.7 n 0. the deﬁnition of which we will discuss in more detail below. two-dimensional map. 15.8 0.2 0 200 400 n 600 800 1000 Fig.9 0. is ∂xn+1 ∂xn ∂yn+1 ∂xn ∂xn+1 ∂yn ∂yn+1 ∂yn cos θ + 2xn sin θ sin θ − 2xn cos θ − sin θ cos θ = = 1. n This is the most simple area-preserving. We can see that it preserves areas.5 x 1000 2000 n 3000 4000 5000 0.11 we show the results of an iteration using the MATLAB script . since the determinant of its Jacobian. Example: The H´non map e The H´non map is a nonlinear.4 0. quadratically nonlinear map with linear part a rotation through the angle θ about the origin. The action of the logistic map and its ﬁfth iterate with x1 = xc and r = rc − 1 × 10−6 .4 0. In Figure 15. which is deﬁned by e xn+1 = xn cos θ − yn sin θ + x2 sin θ.2 MAPPINGS xn+1 = f(xn) = rx(1−x) 1 1 xn+1 = f (xn) (5) 459 0.

owe their appearance to the existence of periodic solutions.24. Chaotic solutions exist near these points. po(ii)=x+i*y. y=st. In particular.11. In fact.5:0. if ((abs(x)>10)|(abs(y)>10)) break end end end plot(po. In addition. xn=cosa*x-sina*y+x^2*sina. or islands. for its=1:1000 ii=ii+1. there is a 5-cycle consisting of the ﬁve points at the centre of the chain of ﬁve islands. yn=sina*x+cosa*y-x^2*cosa. e It is clear that this apparently simple map gives rise to extremely complex behaviour. Fig.’.5 x=st.’MarkerSize’. axis([-1 1 -1 1]) & $% Note that we have used MATLAB’s ability to plot complex numbers. and that | is the logical or operator.3. for st=-0.’. If we consider .05:0. Iterates of the H´non map for cos θ = 0.460 AN INTRODUCTION TO CHAOTIC SYSTEMS ' cosa=0. 15. y=yn. for cos θ = 0. The regions where there are concentric sets of points.24 we see evidence of a period 5 structure. sina=sqrt(1-cosa^2). x=xn.24 and cos θ = 0.2) axis equal. separated by ﬁve further unstable equilibrium points. ii=0. it is necessary to conﬁne the iterates by stopping the calculation once the points have left the domain of interest. A ﬁxed point.15. Figure 15. A point of period k. of the map satisﬁes ¯ x = g(¯ ). We also note that if we consider certain regions in more detail. † Note that.12. x.14) where g : Rn → Rn is a diﬀeomorphism. other periodic structures become evident (for instance.11 using MATLAB’s ability to zoom in on a small region of an existing ﬁgure).12. only the H´non map is a diﬀeomorphism. Fig. e since the logistic.2 MAPPINGS 461 other values of cos θ. 15. we can see structures with even longer periods (see. 15.2. when cos θ = 0. x∗ . tent and horseshoe maps are not bijective and have discontinuous derivatives. and x∗ = gm (x∗ ) for x all positive integer m < k. .34 we ﬁnd a structure of period 11). of the maps that we consider in this section.1 Fixed and Periodic Points of Maps Consider the map xn+1 = g(xn ). which we obtained from Figure 15.24 in the neighbourhood of one of the e hyperbolic ﬁxed points. satisﬁes x∗ = gk (x∗ ). that is g is a bijection and is dif¯ ferentiable with a diﬀerentiable inverse†. for example. Iterates of the H´non map for cos θ = 0. (15. of dimension u. it cannot also asymptote to a diﬀerent equilibrium tote to x point as n → ∞. ωloc . We say that the equilibrium point is hyperbolic if none of the eigenvalues of its Jacobian have modulus unity. stable manifold of an equilibrium point cannot intersect itself. and that it is nonhyperbolic otherwise. Theorem 15. Proof Recall that every point x ∈ Rn has a unique image and preimage. x c (iii) The local centre manifold. (i) (ii) (iii) (iv) The The The The stable manifolds of diﬀerent equilibrium points cannot intersect. ˆ xn+1 = Dg(¯ )ˆ n . x. ωloc . there exist three invariant manifolds. unstable manifold of an equilibrium point cannot intersect itself. is spanned by the eigenvectors of Dg(¯ ) whose eigenvalues have modulus greater than unity. g(x∗ ).1 Consider the diﬀeomorphism (15. since g is a diﬀeomorphism.14) about a ﬁxed point. s (i) The local stable manifold. just as they do for systems of diﬀerential equations. the solution of xn+1 = g (0)xn is xn = (g (0)) x0 . unstable manifolds of diﬀerent equilibrium points cannot intersect. is spanned by the eigenvectors of Dg(¯ ) whose eigenvalues have modulus equal to unity. In the same way as we found for diﬀerential systems. but on the unstable manifold. x∗ . xx (15. Since g is continuous. x. . is spanned by the eigenvectors of Dg(¯ ) whose eigenvalues have modulus less than unity. this is not possible. nonlinear system. at leading order. ωloc . Similarly. of dimension c. (iv) As for case (iii). x (15. in the neighbourhood of an equilibrium point of a map.462 AN INTRODUCTION TO CHAOTIC SYSTEMS ¯ Let’s now consider the linearization of (15. x These local manifolds exist as global manifolds when we consider the full. of dimension s.15) ¯ where Dg(¯ ) is the Jacobian matrix associated with g at x = x. must asymp¯ as n → ∞. For one-dimensional x n maps. (15.15) has solution n ˆ xn = (Dg(¯ )) x0 . Points on the stable manifold in the neighbourhood of x∗ must map to points on the stable manifold in the neighbourhood of g(x∗ ). x u (ii) The local unstable manifold. that is mapped to a point of intersection of the stable manifold with itself.16) and the stability of the ﬁxed point depends upon the size of the moduli of the n eigenvalues. ¯ (i) Since a point in the stable manifold of an equilibrium point. which proves (i). (iii) Consider a point.14). (ii) As for case (i) but with n → −∞. and hence the stability of x = 0 depends upon whether |g (0)| is greater than or less than unity. If we let ¯ ˆ xn = x + xn we ﬁnd that. x → 3(1 − x) < 0. 15. 15. This will prove to be crucial later. 0). After one iteration. with f n (x) → −∞ as n → ∞ for x ∈ (1. Clearly. Note that f (x) = 1 when x = 1 or x = 2 and that the only equilibrium 3 3 Fig. when the function f is as shown in Figure 15. and case (i) applies for subsequent iterations. all points with x > 1 are therefore mapped to a point with x < 0. (ii) If x > 1. point of the map is zero. x → 3x < x. We will concentrate on the case s = 3. which will prove to be of direct relevance later. . ∞). Example: A tent map A tent map is a map x → f (x) where f : R → R and f (x) = sx for x s(1 − x) for x 1 2.15.2. f n (x) → −∞ as n → ∞ for x ∈ (−∞. (i) If x < 0.13. Let’s now consider where this tent map sends various sets of points. 1 2. either of the same equilibrium point or a diﬀerent one. we will examine two more examples of maps.2 MAPPINGS 463 Although this theorem eliminates several possibilities. The function f (x) when s = 3.13. it is possible for a stable manifold to intersect an unstable manifold.2 Tents and Horseshoes Before we return to consider systems of diﬀerential equations. 1] have images in the set 9 9 9 9 9 9 [0.d1 d2 d3 . 3 ] ∪ [ 6 . . . Similarly. 1] to leave K1 = [0.2. . 0. To see this.2x2 x3 . . (r + 1)/3n ]. each 3 x in K2 has ternary form 0. 9 9 9 9 9 9 and so on. Lemma 15. . which can be generated iteratively using R0 = {0} . . .x21 x22 x23 . K= n=0 Kn . . 1] = [0. .22x3 x4 . . . .20x3 x4 .x11 x12 x13 . and must be uncountable. with f n (x) → −∞ as n → ∞. . 2.† † This is just a variation of Cantor’s famous diagonalization proof that the real numbers are uncountable. 0.0x2 x3 . that are not expelled from [0. The index r takes the values in the set Rn . . . < x2 = 0. We can now deﬁne y = 0. = d1 /3 + d2 /32 + d3 /33 + · · · . 1] as n → ∞ can be constructed in an iterative way by deleting the middle third from K0 = [0. x → f (x) > 1. Proof Suppose that K is countable. . . Rn = {Rn−1 . . in ternary form. or 0. with di = 0 or 2 for n = 1. so that all the members of K can be ordered as x1 = 0. 1]. We can see that excluding the middle third of each successive subinterval is equivalent to excluding numbers that have di = 1 for i = 1. then deleting the middle third 3 3 from each of the remaining intervals to leave K2 = [0. 1 ] ∪ [ 2 . and case (ii) applies for subsequent iterations. 3n − 1 − Rn−1 } for n We then have ∞ 1. . 1 ] ∪ [ 2 . . 1 ] ∪ [ 2 . 7 ] ∪ [ 8 .02x3 x4 . The elements of K cannot therefore be ordered. .y1 y2 y3 . 2 . 1 ] ∪ [ 2 . 1]. if it lies in [ 2 .1 Cantor’s middle-third set.464 AN INTRODUCTION TO CHAOTIC SYSTEMS (iii) If x ∈ 1 . . 1]. so the remaining points have f n (x) → −∞ as n → ∞. After one iteration. .0222 . which tends to zero as n → ∞. 2. Then 0 < y < 1. according to which of the four subintervals of K2 it lies in. note that each point in K1 has ternary form 0. 3 3 Only points in the set [0. and so on. . A more convenient deﬁnition of K is provided by looking at the numbers in [0. 3 ] ∪ [ 6 . . . ] and 0. 0. 0.222 . . and hence the length of K is zero. if it lies in [0. has uncountably many points. so that K is a positively invariant set for the tent map with s = 3. . 3 3 If we continue this process. completely disconnected set. . K. . We can write all points in K as x = 0.00x3 x4 . 1]. 1] expressed in ternary. Note that the length of Kn is 2n /3n . 7 ] ∪ [ 8 . with Kn the union of the 2n closed subintervals [r/3n . . K. All points initially in K remain in K as n → ∞. we can see that the set of points. . 1 ] = 3 [0. . or base three form. with ym = 0 if xmm = 2 and ym = 2 if xmm = 0. which are mapped to [0. each of which has length 3−n . ]. 1 ] ∪ [ 2 . . . . 1]. (iv) Now consider points with x ∈ [0. all points with 1 < x < 3 3 3 2 3 are therefore mapped to a point with x > 1. y ∈ K and y = xn for all n – a contradiction. . which is known as Cantor’s middle-third set and was ﬁrst constructed by Cantor in 1883 as an example of an inﬁnite. . It is now helpful to show that each point in the invariant set can be uniquely labelled by a sequence of 0’s and 1’s. y) ∈ R2 | 0 x 1. however small. which can be visualized in terms of stretching and folding the unit square in the opposite way. Other Cantor sets can be constructed by successively removing diﬀerent parts of [0. to itself.15. of [0. as shown in Figure 15. 15. as we shall see. 1]. The map is only deﬁned on the unit square. typical of a chaotic system. D. where ak = 0 1 for points in the left half of D. Λ. Λ= · · · ∩ f −2 (D) ∩ f −1 (D) ∩ D ∩ f (D) ∩ f 2 (D) ∩ · · · .14. f : D → D. is the intersection of the image of D under any number of forward or backward iterations. Fig. y) such that x ∈ Kx and y ∈ Ky . and then folds the resulting strip back on itself.14. where Kx and Ky are the Cantor middle-third sets in each direction. D = (x. Example: The Smale horseshoe map Consider a two-dimensional map from the unit square. The invariant set. by construction. is shown in Figure 15. This is the Smale horseshoe map. The Smale horseshoe map. which is both positively and negatively invariant. 1] contains inﬁnitely many subintervals that do not intersect K. for points in the right half of D. The inverse map. and points that are mapped out of D are discarded. each application of the map removes the middle third of each remaining strip in each direction. 0 y 1 . As we can see. we append ak to the right of the symbol sequence. so we conclude that Λ is a two-dimensional Cantor set given by the set of points (x. After each forward iteration of the map. .15. K contains no subinterval.2 MAPPINGS 465 Nonetheless. An invariant set that is also a Cantor set is. and [0. The function f contracts the square in the horizontal direction and expands it in the vertical direction. 1]. 16. 15. . The inverse Smale horseshoe map.15. 15. Fig.466 AN INTRODUCTION TO CHAOTIC SYSTEMS Fig. The construction of the left half of the bi-inﬁnite sequence. the action of the horseshoe map is simply to shift the decimal point in the bi-inﬁnite sequence one place to the right. which maps the point x0 ∈ Rn and an interval (t0 . and hence the horseshoe map.17. The digits to the right of the decimal point reﬂect the vertical location and to the left the horizontal location.t0 . that is easier to work with than the original system.17. We introduce the ﬂow evolution operator or ﬂow operator. 15. Consider the autonomous system of ordinary diﬀerential equations dx = f (x).´ 15. 15. a2 a1 a0 .3 THE POINCARE RETURN MAP 467 as illustrated in Figure 15.3 The Poincar´ Return Map e Now that we have seen several examples of maps.17) dt where f : Rn → Rn and x ∈ Rn . . as shown in Figure 15. . The map is therefore equivalent to the shift map x → 2x|1. For the same reasons that the shift map that we studied as our ﬁrst example had chaotic solutions. e and then develop a technique for deciding whether this map is chaotic. . which is known as the Bernoulli shift map. t1 ) ∈ R . We can then represent any point in Λ as the bi-inﬁnite Fig.a−1 a−2 . φt1 . . we will demonstrate how the solutions of a system of ordinary diﬀerential equations can be related to the solutions of a map.18. for points in the top half of D. as shown in Figure 15.16. so does the Bernoulli shift map. the Poincar´ map. where ak = 0 1 for points in the bottom half of D. sequence of symbols. The construction of the right half of the bi-inﬁnite sequence. For any point in Λ. a= . (15. After each backward iteration of the map we append ak to the left of the symbol sequence. . t0 (x0 ) = φt1 −t0 (x0 ). so that φ is a group operator (see Chapter 10).t1 · φt1 . φt · φs ≡ φt+s and −1 (φt ) = φ−t .t0 ≡ φt2 . That is x(t0 ) = x0 evolves to x(t1 ) = x1 in the ndimensional phase space during the time interval (t0 .17) subject to x = x0 when t = t0 . dt This is the continuous version of the logistic map. Consequently we have that φ0 is the identity transformation. Note that φt0 . t1 ). 15.18. and we write φt1 . and we can therefore integrate it subject to the initial . t2 ). Example: The logistic equation Let’s construct the ﬂow operator for the logistic equation. The logistic equation is separable. In addition. which we studied earlier. we can express the ﬂow evolution operator in terms of the length of the time interval alone.468 AN INTRODUCTION TO CHAOTIC SYSTEMS Fig. dx = x(1 − x). since (15.t0 for all t1 ∈ (t0 .17) is an autonomous system. the invariant set. The points in Λ. where x1 = x(t1 ) and x is the solution of (15.t0 (x0 ) = x0 and φt2 . to the point x1 ∈ Rn . This is e illustrated in Figure 15. x(t) = φt (x0 ) = x0 et .3 THE POINCARE RETURN MAP 469 condition that x = x0 when t = 0 to obtain the solution. Note that an equilibrium point of the Poincar´ return map corresponds to a limit cycle that intersects Σ e once.17) is to sample the solution with period τ . P : Σ → Σ. We can now verify that φ0 (x0 ) = x0 and φt · φs (x0 ) = φt (φs (x0 )) = and. in complete contrast to the solutions of the logistic map. since φs (x0 ) = x0 es /(x0 es − x0 + 1). This shows that choosing to use a continuous rather than a discrete system can have dramatic consequences for the behaviour of the solutions. and a periodic solution of period k corresponds to a limit cycle that intersects Σ k times. This allows us to track the trajectory of particles in a stroboscopic way. which we can use to analyze the behaviour of integral paths close to a periodic solution.´ 15. Equilibrium points of (15. and can be written in terms of the ﬂow operator as φt+T (x∗ ) = φt (x∗ ) for all t ∈ R. If γ intersects Σ at x = p.19 for n = 3 where Σ is a plane. and consider the e intersections of γ with Σ ⊂ Rn such that (i) Σ has dimension n − 1.17) with n = 2 when e f (x) = 4x + 4y − x(x2 + y 2 ) 4y − 4x − y(x2 + y 2 ) . A more useful map. as expected. then the Poincar´ return map. so that φn (x0 ) = x(t0 + nτ ) for n ∈ Z. x0 es et − x0 es + (x0 es − x0 + 1) x0 es+t − x0 + 1 φs (x0 )et . and hence the evolution operator. .17) with period T satisfy x∗ (t) = x∗ (t + T ) for all t ∈ R. Let γ be a trajectory of (15. (ii) Σ is transverse (nowhere parallel) to the integral paths. and its next intersection with Σ is at x = q. is the Poincar´ return map. maps the point p to the point q. Example Let’s try to construct the Poincar´ return map for (15. Periodic solutions of (15.17). φt · φs (x0 ) = x0 es et x0 es+t = . One way of obtaining a map from the system (15.17) satisfy φt (x∗ ) = x∗ for all t ∈ R (or equivalently f (x∗ ) = 0). (iii) all solutions in the neighbourhood of γ pass through Σ. φs (x0 )et − φs (x0 ) + 1 which is equal to φs+t (x0 ). x0 et − x0 + 1 Note that solutions of the logistic equation all have x → 1 as t → ∞. which sepa√ rates the two equilibrium points at x = 27. e and Σ is the positive x axis. and. Indeed the equations are symmetric about this plane. The deﬁnition of the Poincar´ return map. θ = −4t.22) and (9. along with the Poincar´ return e map. and hence deﬁne a Poincar´ return map. we must proceed e numerically. 4 − x2 + x2 e8t 0 0 The orbit next returns to this axis when θ is an integer multiple of 2π.20. using (9.19. we need to deﬁne an event function . ˙ These expressions can be integrated with initial conditions in Σ. Example: The Lorenz equations Solutions of the Lorenz equations repeatedly cross the plane y = −z. the positive x-axis. y = z = ±6 2. so that x1 = P (x0 ) = 4x2 e4π 0 . 15. In MATLAB.23). that is when t = π/2.470 AN INTRODUCTION TO CHAOTIC SYSTEMS Fig. z → −z. In order to investigate this map. which rapidly asymptotes to x = 2. It is easier to write this system in terms of plane polar ˙ coordinates. since they are unchanged by the transformation y → −y. We can therefore use the plane y = −z as Σ. so that r(0) = x0 and θ(0) = 0. 4 − x2 + x2 e4π 0 0 The solution for x0 = 10 is shown in Figure 15. we ﬁnd that r = r(4 − r2 ) and θ = −4. to obtain r= 4x2 e8t 0 . but only as the solution approaches Σ with y + z increasing (direction = 1). e  function [value.    This function returns the value zero when the solution intersects Σ (when y + z = y(2)+y(3) = 0). and allows the integration to continue (isterminal = 0).[0 500]. which are in the function lor # function dy = lor(t.   This integrates the Lorenz equations. isterminal. We can then use the commands   options = odeset(’Events’. In this case. 15.y) value = y(2)+y(3). dy=dy’.´ 15. direction = 1.20. Note that the function odeset allows us to create a variable options that we can pass to ode45 as an argument. we tell ode45 to detect the . isterminal = 0. which controls its execution. The solution when x0 = 10 and the corresponding Poincar´ return map. dy(2) = 10*(y(3)-y(2)). starting from x = y = z = 1 when t = 0. dy(3) = -y(2)*y(1)+28*y(2)-y(3).y) dy(1) = -8*y(1)/3+y(2)*y(3).3 THE POINCARE RETURN MAP 471 Fig. " ! for 0 t 500. solution = ode45(@lor. options).@lorenzevent). [1 1 1]. direction] = lorenzevent(t. If this equilibrium point is stable.21.17). produced by ode45. in solution. since the points where the solution meets Σ lie on a simple curve. of the map. which corresponds to a limit cycle solution e † The variable solution. The Poincar´ e return map is shown in Figure 15. Fig. is a structure. the limit cycle is stable.17) is associated with an equilibrium point. e 15. A Poincar´ return map for the Lorenz equations. x. 1982. . The dynamics are. As you can see.4 Homoclinic Tangles Let’s now consider a two-dimensional Poincar´ return map. with f : Σ → Σ and Σ ⊂ e R2 . Let’s assume that a particular equilibrium ¯ point. apparently chaotic (see Sparrow.21. however. and eﬀectively contains more than one type of data in its deﬁnition. for a detailed discussion of the Lorenz equations). (15. the map is eﬀectively onedimensional. associated with a three-dimensional system. Each limit cycle solution ¯ of (15.ye. The variable† solution then contains the points where the solution crosses Σ.te. x of the Poincar´ return map. at times solution.472 AN INTRODUCTION TO CHAOTIC SYSTEMS event coded for in lorenzevent. 15. it is possible for these manifolds to intersect. ω s (¯ ). Proof The Poincar´ return map is associated with a ﬂow. the x x ¯ ¯ intersection of the stable and unstable manifolds of x. the manifolds become tangled. Such a point asymptotes to x as n → ±∞. Since the stable and unstable manifolds associated with this trajectory vary continuously. so that. . as shown in Figure 15. 1988). Under successive applications of the map and its inverse.22. ¯ A homoclinic point is a point x = x that lies in the set ω s (¯ ) ∩ ω u (¯ ). and thus have chaotic solutions. then all positive and negative iterates of x0 have the same orientation.4 HOMOCLINIC TANGLES 473 of (15. but is not actually necessary (see Wiggins. and a one-dimensional x x unstable manifold. has a one-dimensional stable manifold.17). We will now discuss why this should be so. Figure 15. and should not be read as a rigorous proof. Although we will proceed through lemmas and theorems. The manifolds then contain embedded horseshoe maps.22. if this intersection is transverse. As we saw earlier. We will now assume that the map f is area-preserving.24 shows e the trajectory through the points x0 and f (x0 ). As we shall see. as shown in Figure 15. our approach is fairly informal. intersecting inﬁnitely often. for any set D ⊂ Σ. a discrete set of points that is a subset of the stable and unstable manifolds.15. the area of f (D) is equal to that of D. A homoclinic tangle. Fig. a homoclinic point is mapped to a homoclinic orbit.23. ω u (¯ ). either US or SU.2 If x0 is a transverse homoclinic point. their orientations can at most rotate and cannot ﬂip during their passage from one side of Σ to the other. 15. This assumption greatly simpliﬁes the following discussion. Lemma 15. 24. US and SU orientation at a homoclinic point. the next transverse homoclinic point along say ω s after x0 cannot be f (x0 ).474 AN INTRODUCTION TO CHAOTIC SYSTEMS Fig. Remember that the manifolds ¯ are stable or unstable with regard to the equilibrium point x. 15. 15.3 If x0 is a transverse homoclinic point. then there must be another transverse homoclinic point between x0 and f (x0 ). Lemma 15. The stable and unstable manifolds of the limit cycle. s u ∑ s u x0 f (x0) u s Fig. In other words. not x0 . .23. Fig. If ω s (¯) and ω u (¯) intersect transversally at a point ¯ other than x. which we label a and b. the distance between the points along ω u must increase. Note that. then the map has a ¯ horseshoe map embedded within it. then the next homoclinic point along ω s .25. Theorem 15. one associated with x0 and one associated with z. z = f (x0 ). these lobes must overlap. therefore. We can assume. There are.26. Note that this means that there are at least two transverse homoclinic orbits.25. if there is a homoclinic tangle. this theorem shows that the . bounded by short sections of ω s and long sections of ω u . Consequently. Lemma 15. the distance between these images must decrease. as shown in Figure 15. two further transverse intersection points. 15. However.4 HOMOCLINIC TANGLES 475 Proof If the orientation associated with x0 is US. since horseshoe maps have chaotic orbits. We therefore have a ﬁnite area covered by an inﬁnite set of lobes of ﬁnite area equal to the area of A0 (recall that we are assuming that the map preserves areas). and conclude that there must be an inﬁnite number of homoclinic points. which is bounded by the stable and unstable manifolds through x0 and z.2 (Smale–Birkhoﬀ ) Let f : R2 → R2 be a diﬀeomorphism with a x x hyperbolic equilibrium point x. z. as shown in Figure 15. in other words. The orientation of successive homoclinic points. and by Lemma 15. This leads to long thin lobes. that they overlap between z and f (x0 ). We can repeat this argument indeﬁnitely. without loss of generality.15. Similarly if x ∈ L0 then f j (x) ∈ Lj which is bounded by the portions of ω u and ω s between f j (x0 ) and f j (z).2. has SU orientation. The image of L0 is the lobe L1 .4 The existence of a single transverse homoclinic point ensures the existence of an inﬁnite number of transverse homoclinic points. which is bounded by the portions of ω u and ω s between f (x0 ) and f (z). Proof We consider the image of the points within the lobe L0 . Since we know that ¯ ¯ f n (x0 ) → x and f n (z) → x along ω s as n → ∞. 15. with overlap between D − and D + . f k (D) is stretched along the unstable manifold and contracted along the stable manifold. there is an algebraic test. Then ¯ (i) f k (D) contains the iterates of f k (x0 ) for k ∈ Z. that can be used to determine whether a system has a transverse homoclinic point. D − . which we deﬁne below. x0 . is a horseshoe map. x of a map f : R2 → R2 . Our aim here is simply to convince you that the forward and backward maps of the region D.26. which uses the idea of Markov partitioning. A more detailed proof. (iii) For k > 0. as shown in Figure 15. The homoclinic points a and b.476 AN INTRODUCTION TO CHAOTIC SYSTEMS f j(x0) f j(z) f(z) f(x0) L1 b a ωu x0 L0 z Fig. . Consider a region D that contains a transverse homoclinic point. associated with an equilibrium point. (ii) For k < 0. (iv) For some backward iterate −q. Proof We will simply give an outline of the proof here. As we shall see later. intersect.27. f k (D) is stretched along the stable manifold and contracted along the unstable manifold. (v) The map f −(p+q) (D + ) = D − . which maps the region D + into a horseshoe shaped region. Mel’nikov’s method. f −q (D) = D − will have a horseshoe shape and intersect with f p (D) = D + for some forward iterate p. is given in Guckenheimer and Holmes (1983). existence of a homoclinic tangle implies the existence of chaotic orbits. A horseshoe map arising from the dynamics in a homoclinic tangle. by Theorem 15. 1988). We start by considering . y) and associated diﬀerential equations ˙ x = f (x) where f = (−∂H/∂y. t+T ) for some period T > 0. 15. 15.9) with Hamiltonian H = H(x.27.3.4 HOMOCLINIC TANGLES 477 ωs x0 D ωu D+ x D− Fig.2. dt ∂y dt ∂x (15. written in component form. dx ∂H dy ∂H =− . We consider a Hamiltonian system (see Section 9. We will focus on two-dimensional Hamiltonian systems perturbed by a periodic function of time only.19) 1. ˙ with g(x. t). and hence.1 Mel’nikov Theory We will now describe an algebraic test with which we can determine whether a system has a homoclinic tangle.15. (15. and the unperturbed system. t) = g(x. ∂H/∂x)T or. = .4. in the form x = f (x) + g(x. although the method can also be used for perturbed non-Hamiltonian systems (see Wiggins.18) We will consider this system under the inﬂuence of a small perturbation that is a function of space and periodic in time. chaotic solutions. The interior of the homoclinic orbit must contain concentric limit cycles surrounding a centre at x∗ .30. A homoclinic orbit of the unperturbed Hamiltonian system. 15. We will now introduce the idea of the distance between points on the stable and unstable manifolds of the saddle point. x0 . since map. This is not true for the solutions of the perturbed system. Let t0 denote the time at which we wish to consider the fate of the solutions. We would now like to ﬁnd some way of distinguishing between the two cases – either disjoint or tangled manifolds. the stable and unstable manifolds of the saddle point do not necessarily connect smoothly. this is equivalent to a ¯ Poincar´ return map with Σ the plane t = 0.9. The idea is to ﬁnd an expression for this distance and determine whether it can be . e of the unperturbed system of diﬀerential equations is also an equilibrium point of ¯ the associated map. 15. its orbits are invariant under arbitrary transformations in time. Since the unperturbed system is autonomous.28. as shown in Figure 15. with. |x −x0 | 1. so that the orbits x0 (t) and x0 (t − t0 ) are the same. we show the orbit x0 (t) as a dashed line and the stable and unstable manifolds of the saddle point of the perturbed system as solid lines.4. and that the stable and unstable manifolds of x0 are the same for the unperturbed system of diﬀerential equations and the associated map. t) = g(x. Since g(x. this system is area-preserving. t)-phase space. For the perturbed point of the map to the point x .3 Perturbed System (0 < 1) We can think of the perturbed system as autonomous in the three-dimensional (x. t + nT ). We can deﬁne an associated map by stroboscopically sampling the ﬂow at t = nT for n ∈ Z. x0 x∗ x = x0(t) Fig.3. as shown in Figure 15.28. Note that the equilibrium point.29. In Figure 15. y. and its equilibrium ¯ points are either nonlinear centres or saddles.478 AN INTRODUCTION TO CHAOTIC SYSTEMS 15.2 Unperturbed System ( = 0) As we saw in Section 9. We assume that the inﬂuence of the perturbation is to modify the equilibrium 1. We suppose that x0 is a saddle point ¯ and that there is a homoclinic orbit x0 (t) such that x0 (t) → x0 as t → ±∞.4. The Mel’nikov function. If D(t. we construct the normal to the unperturbed homoclinic orbit. D(t. The perturbed stable and unstable manifolds of the associated map. and is given by D(t. Fig. t0 ) has simple zeros. We can do this by deﬁning the Mel’nikov function. 15. The perturbed (solid lines) and unperturbed (broken lines) stable and unstable manifolds of the saddle point. zero.4 HOMOCLINIC TANGLES 479 Fig. t0 ) · d(t.30. and hence N(t. t0 ) · f (x0 (t − t0 )) = 0. t0 ).29. Firstly.15. x0 (t). In component form. then there must be transverse homoclinic intersections. which then provides an algebraic suﬃcient condition for the existence of transverse homoclinic intersections and hence chaotic dynamics. The tangent to the orbit is f . t0 ) connects corresponding points on the stable and unstable manifolds. . Here N(t. t0 ) = N(t. is proportional to the component of the distance between corresponding points on the stable and unstable manifolds in the direction normal to the unperturbed homoclinic orbit. 15. and we can conclude that the system has chaotic solutions. t0 ). t0 ) is a normal to the unperturbed orbit and d(t. diﬀerentiating the functions D s. t0 ) − D u (t. we now introduce the binary operation ∧ such that u ∧ v = u1 v2 − v1 u2 .19). We now substitute (15. ˙ s. t).u (15.u (t. ˙ dt .20) into the perturbed system. t0 ).u = f (x0 ) + Df (x0 )x1 + g(x0 . t0 ). f1 (x0 (t − t0 ))). and obtain s. t) + O( 2 ). t0 ). t0 ) = f ∧ x1 + f ∧ x1 . t0 ) ∼ (xs (t. where D s. t0 ) = f ∧ xs. t0 ) and D u (t.u x1 (t.u ˙ ˙ ˙ s. so that D ≡ f ∧ d. t0 ) = (d1 (t.u (t. t0 ) = N(t. t0 ). t0 ) − xu (t. where f = (f1 . f2 ). We can now use perturbation theory to obtain an expression for D(t. 1 and hence d(t. t0 ). D s (t. since x0 is a solution of the unperturbed equation. such that D ∼ D s (t. so the Mel’nikov function is D(t. t0 ) = xs (t.u (t. 1 (15. However. This equation is automatically satisﬁed at leading order. t0 ) = Df (x0 (t − t0 ))x1 (t. The vector displacement between corresponding points on orbits xs (t) and xu (t) in the stable and unstable manifolds of x is d(t.u (t.u D s. 1 1 It is now convenient to introduce two subsidiary Mel’nikov functions. d2 (t. t0 ) + g(x0 (t − t0 ). Note that N is not a unit normal.21) Now. whilst at O( ).u s. We also introduce the superscript s.u ˙1 x0 + xs. t) ˙ s. t0 ) − xu (t.u = f (x0 + x1 ) + g(x0 + x1 . We assume that points on the orbits associated with x remain close to points on the homoclinic orbit x0 (t − t0 ). we obtain s. t0 ) ∼ x0 (t − t0 ) + xs. where d(t.480 AN INTRODUCTION TO CHAOTIC SYSTEMS N1 f1 + N2 f2 = 0. t0 ))T . (15. s. For notational convenience.20) where Df (x0 ) is the Jacobian of the unperturbed system. t0 )) 1 1 and D ∼ (f ∧ xs − f ∧ xu ) .u (t. so we can take N = (−f2 (x0 (t − t0 )). since there is a product rule associated with the ∧ operator.u so that we can discuss the stable and unstable cases simultaneously. writing xs. t0 ).u . t0 ). we note that d ˙ f = f (x0 (t − t0 )) = Df (x0 (t − t0 ))x0 (t − t0 ). t0 ) · d(t. t0 ) = −f2 d1 + f1 d2 . so that we can write their locations as a position on the homoclinic orbit plus a small perturbation proportional to . t0 ) + D s (∞. t0 ) = f ∧ g. there are chaotic solutions. we integrate from −∞ to t0 to obtain t0 −∞ t0 −∞ ˙ D u (t. which gives ∞ t0 ˙ D s (t. t0 ) = f ∧ g dt. ∞ t0 and hence −D s (t0 . we have D(t0 ) ∼ (D s (t0 ) − D u (t0 )) = − ∞ −∞ ∞ −∞ f ∧ g dt =− Note that f (x0 (t − t0 )) ∧ g(x0 (t − t0 ). there are no transverse homoclinic intersections. which gives D u (t0 . t0 ) with (t0 ).23) We note that as t → −∞ along the unstable orbit and as t → ∞ along the stable orbit. both solutions tend to x .2.22) and (15. (iii) If we can ﬁnd one zero of a Mel’nikov function. and hence ˙ ˙1 D s. t0 ). we ﬁnd that ˙ D s. the dynamics of the system ensure that there will be inﬁnitely many zeros.u ˙ D s.u (t. so that D u (−∞. and hence by Theorem 15. (15.15.u + f ∧ xs. t) dt.u . 1 Substituting from (15.u + f ∧ g.22) Now consider the stable case integrated from t0 to +∞. 1 If we now use the identity Df f ∧ x + f ∧ Df x ≡ (∇ · f )f ∧ x. Considering the unstable manifold ﬁrst. then ω u and ω s intersect transversally.u (t.21) into this equation gives s. (i) If D(t0 ) has simple zeros for suﬃciently small .4 HOMOCLINIC TANGLES 481 ˙ which implies that f = Df f . . t0 ) = (Df f ) ∧ x1 + f ∧ Df xs.u (t. t0 ) = (Df f ) ∧ xs. t0 ) dt = f ∧ g dt. (ii) If D(t0 ) is bounded away from zero. t0 ) = D s (∞. Adding (15. and note that ∇ · f = 0 since the unperturbed system is Hamiltonian. t0 ) dt = ∞ t0 f ∧ g dt. t0 ) = t0 −∞ f ∧ g dt.23) and changing notation slightly so that we replace (t0 . (15. t0 ) − D u (−∞. We can now construct the Mel’nikov function. 0) and (x. The vectors f and g are f (x0 (t − t0 )) = and g(x0 (t − t0 )) = so that their wedge product is f ∧ g = y 0 (t − t0 )(γ cos ωt − δy 0 (t − t0 )). and it is straightforward to show that the origin is a saddle point and that the other two equilibria are centres. 0 γ cos ωt − δy 0 (t − t0 ) .3). This orbit must pass through the origin. For what values of δ. However. 2 4 2 The unperturbed system has equilibrium points at (x.31. y) = (0. H is a constant on the orbit. The unperturbed phase portrait is shown in Figure 15. so the homoclinic orbit is given by 1 y 2 = x2 − x4 .482 AN INTRODUCTION TO CHAOTIC SYSTEMS Example Consider the perturbed Hamiltonian system given by the forced Duﬃng equation. 2 Setting y equal to zero. where H = 0. 0). We will take y(0) = 0 and x(0) = 2. We can √ diﬀerential equation for the homoclinic orbit and ﬁnd that x0 (t) = 2secht. y) = (±1.24) This is a Hamiltonian system with d x y dt and Hamiltonian H= = . ¨ ˙ when 1. We now need to determine the equation of the homoclinic orbit. = y x − x3 − (δy − γ cos ωt) dt The unperturbed system is d dt x y = y x − x3 ∂H/∂y −∂H/∂x . y 0 (t − t0 ) = − 2sech(t − t0 ) tanh(t − t0 ). we ﬁnd that the homoclinic √ orbit also meets the x-axis √ then solve the where x = ± 2. (15. 1 2 1 4 1 2 y + x − x . x = x − x3 − (δ x − γ cos ωt). (15. γ and ω are there chaotic solutions? We begin by rewriting the system as d x y . y 0 (t − t0 ) x (t − t0 ) − (x0 (t − t0 ))3 0 . This gives us √ √ x0 (t − t0 ) = 2sech(t − t0 ). D(t0 ) = − ∞ −∞ √ − 2sech(t − t0 ) tanh(t − t0 ) √ × cos ωt + δ 2sech(t − t0 ) tanh(t − t0 ) dt. and we can infer that the system is chaotic. Now.15. .1.32 shows the Poincar´ e return map when = 0. Figure 15.4 HOMOCLINIC TANGLES 483 Fig. 15. 4 2 This means that there are transverse homoclinic points when this condition is satisﬁed. using the deﬁnition of the Mel’nikov function. 4 2 provided that √ πω 3 2π γωsech δ< . δ = 0.31. γ = 1 and ω = 1 . This can be integrated to give D(t0 ) = − This has simple zeros when √ πω 3 2π γωsech δ= sin ωt0 . The phase portrait of (15. 3 √ 4δ πω sin ωt0 + − 2πγωsech 2 3 .24) with = 0. we have seen that neighbouring trajectories diverge. can we characterize and quantify the chaos? In this ﬁnal section. What can we say about other nonlinear systems of diﬀerential equations? If we can solve such a system numerically.1 .5 Quantifying Chaos: Lyapunov Exponents and the Lyapunov Spectrum We have now seen how to determine analytically when chaotic solutions exist for weakly. In fact this divergence is usually exponentially fast. 15. we will introduce the ideas of the maximum Lyapunov exponent and the Lyapunov spectrum.5. and it appears to have chaotic solutions. The Poincar´ return map for the forced Duﬃng equation when e γ = 1 and ω = 1 . δ = 0. 3 = 0.484 AN INTRODUCTION TO CHAOTIC SYSTEMS Points in the Poincaré return map 10 5 dx/dt 0 −5 −10 −4 −3 −2 −1 0 x 1 2 3 4 Successive iterations of the Poincaré return map 4 2 0 x −2 −4 0 100 200 300 400 500 t 600 700 800 900 1000 Fig.32.1. 15. periodically perturbed Hamiltonian systems. If we can quantify this rate of 15. which we can use for this purpose. Lyapunov Exponents of Systems of Ordinary Diﬀerential Equations In a chaotic system. One way of doing this is to renormalize after regular time intervals τ .5 LYAPUNOV EXPONENTS AND THE LYAPUNOV SPECTRUM 485 divergence.27) will lead L L to exponentially growing solutions. Let x(t) be a reference trajectory. Consider the system dx = f (x).33. we can quantify the chaotic solutions. x dt (15.) is a linear evolution operator. This gives us the basic numerical recipe for computing λmax from two neighbouring solutions. so that the calculation can continue with a trajectory that is within a small distance of the reference orbit. and consider a neighbour¯ ing trajectory.33. we deﬁne the maximum Lyapunov exponent with respect to a reference trajectory of a ﬂow as λmax = ||∆x(0)||→0 t→∞ lim 1 ||∆x(t)|| log . as shown in Figure 15. y(t). whilst if λ > 0 the trajectories diverge. As t → ∞.27) Formally ∆x(t) = φt (∆x(0)) where φt (. Fig. and the trajectories ¯ diverge. and consequently we need to ensure that we can indeed linearize by considering neighbouring trajectories (∆x(0) → 0).26) 2 where ||x|| = i xi is the Euclidean norm (see Section A1. Recall that the idea which is central to this deﬁnition is linearization about the ¯ trajectory x(t). t ||∆x(0)|| (15. The maximum Lyapunov exponent is then given by the average of the . ∆x(t) is governed by the linearized equation d∆x = Df (¯(t))∆x. and λ gives a measure of the rate of divergence. as shown in Figure 15. If λmax is positive. In this limit. If λ < 0. so that L L φt (α∆x) = αφt (∆x). The idea is to integrate forward in time until the two trajectories become a given distance apart.2). Neighbouring trajectories diverge exponentially fast.25) ¯ where f : Rn → Rn . the trajectories actually converge. Formally.15. and then scale ∆x.34. numerical integration of (15. This can be avoided in a nonlinear system by renormalizing. that has y(0) = x(0) + ∆x(0). dt (15. we expect that ∆x(t) = y(t) − x(t) ∼ ∆x(0)eλt . 15. with δ = 1 and γ = 3 . .2. We can now construct the maximum Lyapunov exponent using the MATLAB function .486 AN INTRODUCTION TO CHAOTIC SYSTEMS Fig. . 15. 2. (15.28) provided that we deﬁne our renormalization to be ∆xn (0) = δ ∆xn−1 (τ ) for n = 1. . λmax = lim 1 n→∞ (n + 1)τ n log j=0 ||∆xj (τ )|| .34. Renormalization of diverging trajectories. ||∆xn−1 (τ )|| where δ 1 and we note that ||xn (0)|| = δ. a solution 2 5 of which is shown in Figure 15. . exponent calculated between each renormalization.3). ||∆xj (0)|| (15. Example: The forced Duﬃng equation Consider the forced Duﬃng equation. 1. y. y).0. which shows how the average maximum Lyapunov exponent converges to a value of about 0.:)-yout2(end. the number of separate evaluations of the maximum Lyapunov exponent. The built-in function norm(delxe) calculates the Euclidean norm of the vector delxe. and can depend upon the direction of the initial displacement from the reference trajectory. y+del). area. As we saw above. x dt We can write the solution of this system as ∆x(t) = M (t)∆x(0). x(t). For an n-dimensional system of diﬀerential equations. [tout1 yout1]=ode45(eqn.tspan.del0. can have an eﬀect.@duffing)) produces Figure 15. del = del0*delxe/nd. t = t+tau. Consider the system (15.1. the vector of initial data. volume and hypervolume elements.25). nd=norm(delxe). and eqn. ls(i) = log(nd/del0).35.:). The command plot(lyapunov(1000. and hence is chaotic. 15. del0 = δ. [tout2 yout2]=ode45(eqn. † This is just the maximum Lyapunov exponent. we can overcome these problems by deﬁning n quantities that characterize the growth of line†. neighbouring trajectories diverge exponentially fast – an indication that the solution has sensitive dependence upon initial conditions.5 LYAPUNOV EXPONENTS AND THE LYAPUNOV SPECTRUM$ 487 for i=1:n tspan = [t t+tau].tspan. avls=ls.y. Since this is positive. We will see how to overcome this problem in the next section.tau. y = yout1(end. ls=zeros(1.5. delxe= [yout1(end. the time evolution of small variations ∆x(t) about the reference trajectory. 15. this is rather a blunt tool. ∆x0 .n). a handle containing the name of the equation to be integrated. avls(i) = sum(ls)/i/tau. One diﬃculty associated with calculating the maximum Lyapunov exponent in this way is that the choice of the direction of the initial displacement. del=del0*[1 1]/sqrt(2). is governed by ¯ d∆x = Df (¯(t))∆x. .2 The Lyapunov Spectrum Although we now have a way of characterizing the rate of divergence of neighbouring trajectories.' function avls = lyapunov(n.eqn) t=0.:)]. tau = τ .[0 0]. end & % The arguments of the function lyapunov are n.01.

where the evolution operator M (t) is the fundamental matrix. We can construct M numerically.35. 1. the maximum exponents can be ordered as λ1 Lyapunov exponent. . As before. . The λ2 ··· λn . 0. with M (t) = (∆x(1) (t). . . this is a sign of . The fundamental matrix M (t) has n eigenvalues {mi (t)} and the Lyapunov exponents are deﬁned as λi = lim 1 log |mi (t)| for i = 1. and λ1 = λmax . t→∞ t (15. ∆x(2) (t). estimated over 1000 iterations.29) We say that the set of n Lyapunov exponents is the Lyapunov spectrum. The ﬁrst column of M is ∆x(1) (t). ∆x(n) (t)). . . 0)T . The maximum Lyapunov exponent of the forced Duﬃng equation. . The j th column is ∆x(j) subject to the initial condition ∆x(j) (0) = (0. if one of the exponents is positive. 0. . the solution subject to the initial condition ∆x(1) (0) = (1.488 AN INTRODUCTION TO CHAOTIC SYSTEMS Fig. By using this construction we ﬁnd that ∆x(j) (t) = M (t)∆x(j) (0). . 0)T where the one is in the j th position. 15. . . . . 2. . n. . . . . . .

t = t+tau. delx= [youtx(end.eqn) t=0. Since the other two elements of the Lyapunov spectrum are negative. There are several further points that we can make about the Lyapunov spectrum. y+del0*ez). for i=1:n tspan = [t t+tau]. (15.tspan. [toutx youtx]=ode45(eqn. There may be diﬀerent sets of exponents in different regions of phase space.:)]/del0. lambda = avlambda.i) = sum(lambda. ez = [0 0 1]. One of these is positive.5 LYAPUNOV EXPONENTS AND THE LYAPUNOV SPECTRUM 489 sensitive dependence upon initial conditions. . delz]). y+del0*ex). [toutz youtz]=ode45(eqn.15.tspan.:). [touty youty]=ode45(eqn.tspan.tau. avlambda(:.del0. [tout yout]=ode45(eqn.2)/i/tau. dely= [youty(end. ey = [0 1 0]. Figure 15.:)]/del0. dely.:)-yout(end. avlambda = zeros(3. We should note that there are some diﬃculties associated with calculating the Lyapunov exponents numerically. It is also notoriously hard to obtain convergence for Lyapunov exponents. lambda(:. y = yout(end.:)-yout(end. this indicates that neighbouring solutions diverge in one direction and approach each other in the two perpendicular directions.n). Example: The cubic crosscatalator We can calculate the Lyapunov spectrum of the cubic crosscatalator equations. delz= [youtz(end.y. end & $% The built-in function eig(A) calculates the eigenvalues of the matrix A.:)]/del0.10) using the MATLAB function ' function avlambda = lyapunovspectrum(n. which indicates that the solution depends sensitively upon initial conditions. ex = [1 0 0]. especially where there are regions characterized by rotation.:)-yout(end.36 shows the estimates of the three Lyapunov exponents converging over 1000 iterations.tspan. such as in the neighbourhood of centres.i) = log(abs(m)). y). y+del0*ey). m = eig([delx. and we need to use a renormalization scheme to calculate the Lyapunov spectrum. This means that the Lyapunov spectrum may depend upon the starting condition. If µi are the eigenvalues of Df (x∗ ). . This means that the fundamental matrix is M = exp (Df (x∗ )t) (see Section 14. dt with Df (x∗ ) a constant matrix. and so on.29). so that line elements grow like eΛ1 t . x = x∗ .3.2). note that. For the cubic crosscatalator. d∆x = Df (x∗ )∆x. On substituting these into the deﬁnition (15. so that line and area elements expand.490 AN INTRODUCTION TO CHAOTIC SYSTEMS Fig. but that Λ3 is negative. N ΛN = j=1 λj This allows us to consider the growth of various elements.36. so that volume elements contract. we obtain λi = Re(µi ). in the neighbourhood of an equilibrium point. we can see that Λ1 and Λ2 are positive. — We can also deﬁne the N -dimensional Lyapunov exponent. The Lyapunov spectrum of the cubic crosscatalator equations. To see this. then the eigenvalues of M are eµi t . area elements grow like eΛ2 t . — The Lyapunov exponents of an equilibrium point are just the real parts of its eigenvalues. 15. 3 Consider the map H(x) = 3x. — Lyapunov exponents characterize average rates of expansion (λi > 0) and conn traction (λi < 0) in phase space.10). which means that the product of its eigenvalues is unity. −2 + 3x. . . . f (xi ) 15. If a(x) = a1 a2 a3 .4 (a) Write down the map for the iterates in the determination of the roots.4) to (15. 1 . 15. we note that if a dynamical system is chaotic then at least one Lyapunov exponent is positive. This is given by xi+1 = xi − f (xi ) . After deﬁning suitable dimensionless variables. . For conservative systems. has i=1 λi < 0. for which volume elements contract. .2 for x ∈ 0. and hence the sum of the Lyapunov exponents is zero.1 Write down the equations that govern the concentrations of the chemicals involved in the cubic crosscatalator scheme. we can try to determine the zeros of f (x. A n dissipative system. since the determinant of the Jacobian is unity. 1 . . Exercises 15. 2}. Finally. 3n n=1 with an ∈ {0.EXERCISES 491 — Every point in the basin of attraction of an attractor has the same Lyapunov spectrum. show that a(H(x)) = σ(a(x)). ∞ 15. 2}. derive (15.9). noting carefully the conditions on the initial concentration of P and the rate at which P decays under which the equations are a good approximation. where σa = b if bn = an+1 . Determine the period 2 points of the map f : x → 4x|1. a) = xn − a using the Newton–Raphson iteration. 2 for x ∈ 1 . 2 Prove that the only points that remain in [0. (15. In order to determine the nth roots of a. i=1 λi = 0.? Show that this point is periodic of period 3 under H. with an ∈ {0. and show for n = 2 that xi+1 = 1 2 xi + a xi . 1] have the form x= an . For what value of x is a(x) = 002002002002 . 5 Determine the ﬁxed point of the map yn+1 = 1 for yn > 0. xn are the Fibonacci numbers. φ)-plane. Show that if = 0 then there is a pair of heteroclinic orbits connecting saddle points of (±π. γ and are positive constants. and 1. θ) and hence calculate the Poincar´ e return map P : R → R as the map of successive intersections of the orbit x(t) with the positive y-axis. Deduce that one of the orbits is given by θ0 (t) = 2 tan−1 (sinh t). comment on the properties of the composition f (g(x)). B= 1 0 1 1 . use MATLAB to calculate a Poincar´ e return map for (i) the forced Duﬃng equation and (ii) the cubic crosscatalator equations. and hence the corresponding maps are areapreserving. the ﬁxed points of the map are a1/n . dt2 where α.7 Express the system ˙ x = x3 + xy 2 − x − 2y. and determine the unstable manifolds. where x = (x.492 AN INTRODUCTION TO CHAOTIC SYSTEMS (b) Show that.6 and show that with yn = xn /xn+1 . φ0 (t) = 2sech(t). 1 + yn 15. The equation of motion of a forced simple pendulum is d 2θ + sin θ = (α + γ cos ωt). 0) in the (θ. (c) Consider A= 1 0 0 −1 .8 15. of the maps f . stating a condition for which these are the same set of points. Discuss the diﬀerent options for f and g given that the maps are area-preserving. B are real 2 × 2 matrices. Consider two maps f and g such that f (x) : x → Ax and g(x) : x → Bx. (a) Determine a condition for f (x) to have a ﬁxed point other than the trivial one. and A. y = yx2 + y 3 − y + 2x ˙ in terms of the polar coordinates (r. for general n. 15. Comment on the ﬁxed points of the composition of f (g(x)) and g(f (x)).9 . determine the stability of these points by considering xi + for Do you think that this is a good way of determining the nth roots of a? 15. where applicable. (b) If det(A) = det(B) = 1. where φ = dθ/dt. g and f (g). Show that orbits starting on the y-axis outside the circle r = e2π / (e2π − 1) never return to the y-axis. y)T . 15. After deﬁning a suitable plane Σ. y.11 15. Find its Lyapunov exponents. (b) Write a MATLAB script that iterates points forward under the inﬂuence of the pulsed ﬂow ∞ ψ(x.y1 ) ∂ψ ∗ ∂x .y0 ) (E15. x + 2y modulo 1). and is given by xn+1 = f (xn ). but that the map x1 − x0 = ∂ψ ∗ ∂y y1 − y0 = − (x0 . yn ) and f (x. y) = (x + y modulo 1. (a) Show that the map deﬁned in (E15. t) is governed by the equations x= ˙ ∂ψ ∂ψ . y1 − y 0 = − (x0 . Calculate the Lyapunov spectrum of the Lorenz equations. y1 ) its position after the impulse.1) 15. Show that this map is area-preserving. where xn = (xn . 2) of the unstable manifold from (−π.y1 ) (E15. We can integrate (E15. 2 Arnol’d’s cat map maps the torus T2 = R2 /Z2 to itself. y)δ(t − t0 ). 2 15. 0) can be expressed as ∞ M (t0 ) = −∞ φ0 (t − t0 )(α + γ cos ωt) dt. and deduce that M (t0 ) = 2π α + γsech πω cos(ωt0 ) . Project The two-dimensional motion of a particle in a ﬂow with stream function ψ = ψ(x. (x0 . y. ∂y ∂x (E15. y0 ) is the position of the particle before the impulse and (x1 .2) is not area-preserving.12 Consider the motion of a particle under the inﬂuence of an impulsive Stokes ﬂow.3) is area-preserving. 0) and the stable manifold to (π. for which momentum is negligibly small. where the stream function is that associated with a point force at .1) to show that x1 − x0 = ∂ψ ∗ ∂y . and ψ = ψ ∗ (x.2) where (x0 . t) = n=0 ψ ∗ (x.y0 ) ∂ψ ∗ ∂x . (x0 . y=− ˙ .EXERCISES 493 Show that the Mel’nikov function for the perturbation problem of intersection near (0.10 Hence show that there is chaos for small if γ > α cosh( 1 πω). y)δ(t − n). (x − d)2 + (y + h)2 (c) (d) (e) (f) Here α is the strength parameter (see Otto. and determine the Lyapunov exponents associated with the ﬂow. for more details). .3). 1 ) and ( 1 . y) = (d. for instance (− 1 . Now consider the ﬂow associated with Stokeslets that are not on a vertical line. 1989). Use the area-preserving map (E15. 3 ). 1 ) and ( 1 . determine the nature of any ﬁxed or periodic points. 1 ) and (0. y) = α(x − d) 1 log 2 (x − d)2 + (y + h)2 (x − d)2 + (y − h)2 − 2hy . h) above a solid plane at y = 0. 2 2 2 2 2 2 2 2 By constructing the Jacobian associated with the ﬂow. and its stream function is given by ∗ ψd. Show that this leads to chaotic dynamics. Investigate the combination of other fundamental solutions of Stokes ﬂow. This model can be extended to other ﬂows (see Ottino.494 AN INTRODUCTION TO CHAOTIC SYSTEMS (x. 1 ) or (− 1 . 2001.h (x. Yannacopolous and Blake. Consider the eﬀect of alternating Stokeslets at (0. 1984). This is similar to Aref’s blinking vortex which leads to chaotic advection (Aref. 3 ) with 2 2 the same strengths. This will give a stroboscopic plot of the trajectory of the point. This is known as a Stokeslet. y. . . y =    y1 y2 . (α + β)x = αx + βx. y.    . together with the existence of inverses and identities for the scalars and vectors. or scalars. (αβ)x = α(βx). 1x = x. associativity and distributivity for the vectors and scalars. z. . x + 0 = x. αx =  . xn     . β ∈ R. . .1 Vector Spaces Over the Real Numbers Let V be a set of objects called vectors. and let R denote the real numbers. . yn are n-dimensional vectors that can be written in terms of their coordinates.  x1 x2 . . xn + yn αxn it is straightforward to verify that Rn is a vector space over R. .  . x. x + (−x) = 0. .  . . (x + y) + z = x + (y + z). These conditions are the familiar laws of commutativity.APPENDIX 1 Linear Algebra A1. z ∈ V and α. . Examples (i) V = Rn . If we then deﬁne vector addition and scalar multiplication by     αx1 x1 + y1  αx2   x2 + y2      x+y =  . . }. for all x. α(x + y) = αx + αy. . (i) (ii) (iii) (iv) (v) (vi) (vii) (viii) x + y = y + x. of the form V = {. The set V forms a vector space over R if. . so that    x=       .   0 0 . and it has a basis with an inﬁnite number of elements.  . . . A basis for V is B = {1. x ∈ [α. All other bases also have n elements. it is again easy to verify that V is a vector space over R. . for some αi ∈ R. . bn form a basis for V . 1        = {b1 . . . . . for every x ∈ V . . and the αi are not all zero. .  There are other bases for Rn . and the zero function by 0(x) = 0. A vector space V is ﬁnite dimensional if it has a basis with a ﬁnite number of elements. cos x.. we can write x as a linear combination of the elements of B. (ii) Consider the vector space ∞ V = n=0 an cos nx an ∈ R. x2 . The subset B forms a basis for V if. . . . . . it is said to be inﬁnite dimensional. which contains an inﬁnite number of elements. If span (b1 . bn is called the span of these vectors. b2 . . for x ∈ [α. . (αf )(x) = αf (x). cos 2x. The  1 0   . A subset B = {b1 . 0 subset   0  1     . .496 LINEAR ALGEBRA (ii) V = Pn = an xn + an−1 xn−1 + · · · + a1 x + a0 | an ∈ R. . β]. bn ) = V . we say that the set of vectors x1 . β] . A typical member of V would be an object of the form 6x3 − 2x + 1. . }. . . . which consists of convergent Fourier series deﬁned on −π x π.  . If it is not ﬁnite dimensional.. xn is linearly dependent. π] . . b2 . . then b1 . . . 0     .. The set of all linear combinations of b1 . . but this is the simplest. . . β]. If α1 x1 + α2 x2 + · · · + αn xn = 0. so that x = α1 b1 + α2 b2 + · · · + αn bn . This shows that V is inﬁnite dimensional.  . since      x1 x2 . . xn    = x1 b1 + x2 b2 + · · · + xn bn . . so that V is the set of polynomials of degree n with domain x ∈ [α. Examples (i) Consider V = R       B=      n . x ∈ [−π.     forms a basis. If we deﬁne vector addition and scalar multiplication by (f + g)(x) = f (x) + g(x). . bn } . bn } of a vector space V is said to be linearly independent if α1 b1 + α2 b2 + · · · + αn bn = 0 implies that α1 = α2 = · · · = αn = 0. . . z + β y.2 Inner Product Spaces The inner or dot product of two elements. ||kx|| = |k| ||x|| for all real k. (ii) x. we can deﬁne a norm b ||f || = a f 2 (x) dx. . V over R. 1]. using the properties of the Riemann integral. ||x + y|| ||x|| + ||y||.. g = a f (x) g(x) dx is an inner product. {xi } for i 1. deﬁned on an interval I = [a. there are diﬀerent types of norm. αn ∈ R. then 1 f. if I = [−1. with x. ||x|| = 0 if and only if x = 0. A set of nonzero vectors in an inner product space. and gives the The Euclidean norm of a vector is deﬁned to be ||x|| = size or length of x. y. . x 0. 5 Two nonzero vectors. z ∈ V . in an inner product space are said to be orthogonal if x. the inner product is a mapping. It is straightforward to conﬁrm. V × V → R. of Rn is deﬁned to be x. C(I). that b . x and y.A1. x. x2 . x . α2 . For example. This means that α1 x1 . . of a vector x must have the four properties (i) (ii) (iii) (iv) ||x|| is a non-negative real number. f (x) = x and g(x) = x3 .2 INNER PRODUCT SPACES 497 A1. Using the inner 1 2 3 product discussed above. A vector space with an inner product deﬁned on it is called an inner product space. To see this. (iii) αx + βy. . . An important example is the space of all real-valued functions. g = −1 x4 dx = 2 . xj = αj xj . . and hence. . αj = 0 for j 0. y = x1 y1 + x2 y2 + · · · + xn yn . the triangle inequality. : f. b]. . . whose members are mutually orthogonal is necessarily linearly independent. In C(I). . ||x||. xj = 0. x = 0 if and only if x = 0. xn ) and y = (y1 . by property (ii) above. for x. . with the three properties (i) x. A useful relationship between the inner product and this norm is the Cauchy– Schwartz inequality. y = 0. | x. . so that ||x|| = x2 + x2 + x2 . x = (x1 . For a general vector space. This is familiar in R3 . xj + · · · + αn xn . z = α x. suppose that α1 x1 + α2 x2 + · · · + αn xn = 0 for α1 . α. z . yn ). xj + α2 x2 . A norm. . β ∈ R. y2 . y = x1 y1 + x2 y2 + x3 y3 . . y | ||x|| ||y||. the vector space of polynomials of degree n. then T (x) = λ1 T (b1 ) + λ2 T (b2 ) + · · · + λn T (bn ). Ostberg and Perkins (1966) for a discussion of completeness. (ii) T (λx) = λT (x) ∀λ ∈ R.† This can be useful when proving rigorous results about diﬀerential equations. Examples (i) If V = R and T : R → R is deﬁned by     x1 x2 T  x2  =  x3  . † See Kreider. and we take a general vector x = λ1 b1 + λ2 b2 + · · · λn bn . x ∈ V . T . x3 x1 3 3 3 then    x1 + y1 x2 + y2 T (x + y) = T  x2 + y2  =  x3 + y3  = T (x) + T (y). Kuller. The function space C(I) is complete under the sup norm.498 LINEAR ALGEBRA We can also deﬁne the sup norm through ||f || = sup {f (x) | x ∈ I} . λx3 λx1   so that T is a linear transformation. into itself. y ∈ V . Linear transformations can be represented by considering their eﬀect on the basis vectors. T is a linear transformation. It follows immediately from this deﬁnition that T (λx + µy) = λT (x) + µT (y) and T (0) = 0. from a vector space. A1. and T (an xn + · · · + a1 x + a0 ) = nan xn−1 + · · · + a1 . (ii) If V = Pn . x3 + y3 x1 + y1    λx2 λx1 T (λx) = T  λx2  =  λx3  = λT (x). . and can be identiﬁed with the operation of diﬀerentiation. denoted by T : V → V . is linear if (i) T (x + y) = T (x) + T (y) ∀x. V .3 Linear Transformations and Matrices A transformation. = λ1 (α11 b1 + α21 b2 + · · · + αn1 bn ) + · · · + λn (α1n b1 + α2n b2 + · · · + αnn bn ). If T (bj ) = α1j b1 + α2j b2 + · · · + αnj bn . for which the condition for nontrivial solutions is det(A − λI) = 0. . . .  . . . This gives λ2 − 5λ + 6 = 0 so that λ = 2 or 3.     x1 x2 T  x2  =  x3  . . αn2 . . . αnn is a representation of the transformation. .. . α2n   λ2       T . 1 0 0 The Eigenvalues and Eigenvectors of a Matrix  A1.4 THE EIGENVALUES AND EIGENVECTORS OF A MATRIX 499 which. = . This is known as the characteristic equation associated with the matrix A.. . . 4 2 −1 1 . .  . where I is the n × n identity matrix. using the standard deﬁnition of matrix multiplication. . αn1 αn2 . In example (i) above. . .   . The eigenvectors satisfy 4 −1 2 1 x y =λ x y . the eigenvalues of A are deﬁned as those values of λ for which the equation Ax = λx has a nontrivial solution.  .4 If A is an n × n matrix and x is an n × 1 column vector. the eigenvectors of A are the corresponding values of x. αnn λn λn We say that the matrix    A=       α11 α21 .. .A1. . Example Find the eigenvalues and eigenvectors of the matrix A= The eigenvalues satisfy det 4−λ 2 −1 1−λ = 0. . For each of these values of λ. x3 x1 so that  0 1 0 A =  0 0 1 . ..  . α1n α2n .. we can write as      α11 α12 . αn1 α12 α22 . . . This deﬁning equation can be rearranged into the form (A − λI)x = 0.. α1n λ1 λ1  λ2   α21 α22 . . . 0  0 λ2 0 . . 0 ˙ We can write this in matrix form as x = Ax  0 0 A= 1 0 2 1 . . xn )T ˙ and a dot denotes diﬀerentiation. 2) for any nonzero real α. the superscript T denotes the transpose of the vector.. such that   λ1 0 . . For λ = 3. −1)T corresponding to λ = −1. 2x − y = 0. . if an n × n matrix A has n distinct eigenvalues λ1 . λ2 . 0. x2 . since ˙ the diﬀerential equations are all decoupled. Example Let’s try to ﬁnd a simpliﬁcation of the system of diﬀerential equations x1 ˙ x2 ˙ x3 ˙ = = = x3 . . . Here.  . . with  1 0 .500 LINEAR ALGEBRA For λ = 2. the system changes to y = B −1 ABy. . Consider the system of diﬀerential equations x = Ax. 2 2 1 The eigenvalues satisfy −λ det  0 2   0 1 −1 − λ 0  = 0. . If we write x = By. 0. . 2x1 + x2 . λ2 .  0 0 . 2 1−λ T so that (λ − 2)(λ + 1)2 = 0. These eigenvectors span a two-dimensional subspace of R3 . 2)T corresponding to λ = 2. The associated eigenvectors are α(1. and hence λ = 2 or −1. . x − y = 0 T so that the eigenvector is β (1. . where x = (x1 . so that the eigenvector is α (1. which is considerably easier to analyze if B −1 AB is diagonal. there exists a nonsingular matrix B. The eigenvectors are 2 linearly independent and span R . . λn . whose columns are the eigenvectors of A.    B −1 AB = diag(λ1 . Example Find the eigenvalues and eigenvectors of the matrix   0 0 1 A =  0 −1 0  . λn The only nonzero entries of this matrix are on its diagonal. . . . λn ) =  . and are the eigenvalues of A. and β(1. x1 . A particularly useful result concerning eigenvalues is that. .. 1) for any nonzero real β. . x2 = −y1 − 1 (1 − 2 2 √ √ √ √ 5)y2 − 1 (1 + 5)y3 and x3 = −y1 + 1 (1 + 5)y2 + 1 (1 − 5)y3 . An−1 . y3 = (1 − 5)y3 . in terms of the 2 2 2 original variables. and hence linearly independent. λ = −1. y2 = c2 exp{ 1 (1 + 5)t}. 2 2       1 √ 1 √ 1  −1  . and the corresponding eigenvectors. . det(A − λI) = 0. can be written as a linear combination of I. To do this we ﬁrstly ﬁnd the √ eigenvalues. 1982). y2 = (1 + 5)y2 . ˙ 2 2 √ These have the simple solutions y1 = c1 exp(−t).1 (Cayley–Hamilton) Every square matrix A satisﬁes its own characteristic equation.  − 1 (1 − 5)  . A. the solution is x1 = y1 + y2 + y3 . (1 + 5). The proof of this theorem is rather involved. Theorem A1. 2 2 √ √ 1 1 −1 5) 5) 2 (1 + 2 (1 − These are orthogonal.  − 1 (1 + 5)  . and form a basis for R3 .4 THE EIGENVALUES AND EIGENVECTORS OF A MATRIX 501 We need to simplify the structure of the √ matrix A. 2 2 1 1 5) 5) −1 2 (1 + 2 (1 − with inverse   B −1 =  we ﬁnd that  1 1 √ 5 1 √ 5  1 √ −1 √ 1 1 √ (3 − 5) − 2√5 (1 − 5)  . The transformed system of diﬀerential equations therefore takes the form √ √ 1 1 ˙ ˙ y1 = −y1 . we will often make use of the Cayley–Hamilton theorem. . Finally. . . . Finally. After choosing our matrix B to be   1 1 √ 1 √ B =  −1 − 1 (1 − √ 5) − 1 (1 + √ 5)  . (1 − 5)) 2 2 1 5) 2 (1 − −1 B −1 AB =  0 0 0√ 1 (1 + 5) 2 0 is the representation of A with respect to the basis of eigenvectors. with k n.A1. y3 = 2 √ c3 exp{ 1 (1 − 5)t}. A2 . 2 5 √ √  1 1 √ (1 + − 2√5 (3 + 5) 5) 2 5  0 √ 1 √ 1 0 √  = diag(−1. 1 (1 + 5) and 1 (1 − 5). and we will not consider it here (see Morris. The Cayley–Hamilton theorem shows that Ak . b] to denote the vector space of continuous functions f : [a. The set of all continuous functions f : (a. b) or C 0 (a. . b) and x→a+ lim f (x) = f (a). b). form a vector space. b). b] → R. A function f is continuous on [a. deﬁnition of continuity is that f is continuous at a point x = x0 if. and. x→c The function f is continuous on the interval (a. C 3 (a. in which the higher derivatives are deﬁned in terms of limits of derivatives of one order lower. We say that f is continuous at x = c if and only if lim f (x) = f (c). but equivalent. . 2π) → R : f (x) = sin x. b). b). ) such that |f (x) − f (x0 )| < when |x − x0 | < δ. b). for which there are a ﬁnite number of discontinuities. b) if and only if it is continuous for all x ∈ (a. b). b] if it is continuous on (a. . An alternative. We can make similar deﬁnitions of C 2 (a. Examples of functions that live in these spaces are: (i) f : (0. there is no singular behaviour at the ends of the interval. C n (a. at each of which f (x) jumps by a ﬁnite amount. . f (x) is not continuous at x = c. C ∞ (a. b). These are known as piecewise continuous functions. There is an important class of functions of this form.APPENDIX 2 Continuity and Diﬀerentiability Let f be a real-valued function deﬁned on some open interval that contains the point x = c. We will use C[a. b) → R (R denotes the set of real numbers) forms a vector space denoted by C(a. . The set of all once-diﬀerentiable functions on (a. we say that f is uniformly continuous on I. b). . In other words. x→b− lim f (x) = f (b). which we denote by P C(a. . If δ depends upon but not upon x0 in some interval I. If the limit of {f (x) − f (c)} /(x − c) exists as x → c. Since f (x) = cos x and f (x) = − sin x = . It can be shown that. for every > 0. there exists a function δ(x0 . b) forms a vector space which we denote by C 1 (a. then it is also uniformly continuous there. if a function is continuous on a closed interval. and denote its derivative by f (c) = lim x→c f (x) − f (c) x−c . If limx→c+ f (x) = limx→c− f (x). like the continuous functions. . we say that f is diﬀerentiable at x = c. 3]. 1 (iii) f : R → R : f (x) = n=1 n! sin (n!) x. nowhere diﬀerentiable function f (x) = ∞ 1 n=1 n! sin (n!)2 x. is continuous on R but nowhere diﬀerentiable. since f (x) = 0 for x < 0 and f (x) = 2 for x 0.1. 3] → R : for 0 x < 1.1. but f ∈ C n (R) for any n > 0. x2 f (x) = cos x for 1 x 2. We conclude that f ∈ C ∞ (0. 2π).1) which is plotted in Figure A2. (ii) f : (−1. this function can be diﬀerentiated as many times as we like and the result will be a continuous function. We conclude that f ∈ C 2 (−1. so that f ∈ / C 0 (R). no more than twice. This function is discontinuous at x = 1 and x = 2. . 1) → R : f (x) = x2 for x 0. (iv) f : [0.  −x for 2 < x 3. This function can be diﬀerentiated twice. 1). which is illustrated in Figure A2.CONTINUITY AND DIFFERENTIABILITY 503 −f (x). but. f (x) = 0 for x < 0. A2. e   (A2. The continuous. and therefore f ∈ P C[0.2. This function. ∞ 2 Fig. then ∃c ∈ (a. then ∃K such that |f (x)| < K ∀x ∈ [a. Theorem A2. . b) such that f (b) − f (a) = (b − a)f (c).2 (The mean value theorem) If f : (a. b]. A2. b) ⊃ C 1 (a. b) · · · ⊃ C n (a. b) · · · . it is continuous at c (but the converse of this is false. bounded interval is bounded.504 CONTINUITY AND DIFFERENTIABILITY Fig. A cross indicates the value of the function at a point of discontinuity. If f is diﬀerentiable at c ∈ (a. b) ⊃ C 2 (a. b]. Theorem A2. b). see example (iii) above) so that C 0 (a. b).1 If f : [a. a continuous function on a closed.2. b] → R and f ∈ C[a. b) → R and f ∈ C 1 (a. The piecewise continuous function given as example (iv). In words. We conclude this section with two useful theorems that you should recall from your ﬁrst course in real analysis. f (0) = −1. Since 1 f (x) = . f (0) = 1. (1 + x)n+1 . the series takes the form x2 xn (n) f (0) + · · ·. b) for some a < 0 < b. n (−1) n! (n) f (x) = . for example.1 Maclaurin Series The Maclaurin series is a power series about x = 0. 1+x 1 f (x) = − . (1 + x)4 . . . A3. f (0) = 2. (−1)n sin x = x − 3! 5! (2n + 1)! n=0 cos x = 1 − x2 x2n x4 + − ··· = . (1 + x)3 −3 · 2 f (3) (x) = . f (n) (0) = (−1)n n!. (−1)n 2! 4! (2n)! n=0 ∞ (A3.APPENDIX 3 Power Series Many commonly encountered functions can be expressed as power series. we can develop a power series representation for a suﬃciently diﬀerentiable function. fractional powers of x or logarithms. f (3) (0) = −3!. (1 + x)2 2 f (x) = . a prime denotes a derivative with respect to x and f (n) (x) is the nth derivative. f (x) = f (0) + xf (0) + f (0) + · · · + 2! n! Here. If f ∈ C ∞ (a. provided that it involves no singularities. ∞ x5 x2n+1 x3 + − ··· = . Example Let’s determine the Maclaurin series expansion for the function f (x) = 1/(1 + x).1) In general. n! 2n+1 2 n=0 2n n=0 ∞ ∞ n (A3. Its existence also requires reasonable behaviour of the function at x = x0 . . (1 + x)3 2 −3 · 2 3! f (3) (x) = . 1+x 2 1 1 f (x) = − . which is a power series about x = x0 . which says that   converges if limn→∞ |an+1 /an | < 1. . we can usually use the ratio test. 2! n! Example Let’s determine the Taylor series of the function f (x) = 1/(1 + x) at the point x0 = 1. f (3) (x0 ) = − 4 . f (x) = f (x0 ) + (x − x0 )f (x0 ) + (x − x0 )2 (x − x0 ) (n) f (x0 ) + · · · + f (x0 ) + · · · . and for a C ∞ function takes the form. ∞ the series an  n=0 diverges if limn→∞ |an+1 /an | > 1.2 Taylor Series The Maclaurin series is a special case of the Taylor series. f (n) (x0 ) = (−1)n n+1 . f (x0 ) = 3 . f (x0 ) = .2) A3. 4 (1 + x) 2 . (1 + x)n+1 2 f (x) = the Taylor series is f (x) = (x − 1)n (−1)n n! (x − 1)n (−1)n 1 = .3 Convergence of Power Series In order to determine for what values of x a power series is convergent.506 POWER SERIES the Maclaurin series is f (x) = 1 + (−1)x + 2 x2 x3 xn + (−1)3! + · · · + (−1)n n! + ··· 2! 3! n! ∞ = 1 − x + x2 − x3 + · · · + (−1)n xn + · · · = n=0 (−1)n xn . Since 1 1 . . f (x0 ) = − 2 . A3. (1 + x)2 2 2 2 f (x) = . n n! (−1) n! f (n) (x) = . y) is a scalar ﬁeld that has suﬃcient partial derivatives at the point (x0 . Taylor Series for Functions of Two Variables If f = f (x. which deﬁnes the radius of convergence of the series as two units from the point x = 1. Notice that the function f (x) is singular at x = −1. also diverges. Note that another useful test for the convergence of a series is the comparison test.2). = an 2 This means that the series converges when −1 < x < 3. 0). which says that if (i) if (ii) if A3. A useful composite result from the theory of Taylor series is that if f (x) = ∞ n n=0 an (x − x0 ) is a Taylor series with radius of convergence R (it converges for |x − x0 | < R).0) 2 +y 2 ) about the point (0.y0 ) + + (x − x0 )(y − y0 ) (x0 .0) . This deﬁnes the domain of convergence of the power series (A3. ak for k > k0 .y0 ) ∂f ∂y (x0 . (x0 . we need an = n→∞ an+1 = (x − 1)n+1 (−1)n+1 . y) = f (x0 .4 TAYLOR SERIES FOR FUNCTIONS OF TWO VARIABLES 507 In the previous example. y0 ). This is often written as |x − 1| < 2. y0 ) + (x − x0 ) (x − x0 )2 ∂ 2 f 2! ∂x2 ∂f ∂x + (y − y0 ) (x0 . (A3. n=1 nan (x − x0 ) This ‘term by term’ diﬀerentiation result allows us to develop a method for solving diﬀerential equations using power series – the method of Frobenius. 2n+2 lim an+1 x−1 < 1. the Taylor series about this point is f (x. ∞ n=0 bn ∞ n=0 bn an converges and bk an diverges and bk also converges. ∂f ∂y = 0.4 ∞ n=0 ∞ n=0 ∞ n=0 an and ∞ n=0 bn are series of positive terms then ak for k > k0 .y0 ) ∂2f ∂x∂y + (x0 .y0 ) (y − y0 )2 ∂ 2 f 2! ∂y 2 + ··· . and f (x) = ∞ n−1 has the same radius for convergence as the series for f (x).A3. 2n+1 so that. It is worth pointing out that the Taylor series of a polynomial is a terminating series and hence has an inﬁnite radius of convergence. and will not have a Taylor series about this point. which we discuss in Chapter 1.2). (0. for convergence.y0 ) Example Let’s ﬁnd the ﬁrst two terms in the Taylor series of e−(x ∂f ∂x = (0. (x − 1)n (−1)n . f is diﬀerentiable for all x such that |x − x0 | < R. x= x y . (0. Example Let’s ﬁnd the linear approximation to the vector ﬁeld f= about the point x0 = 0 1 ex − 1 (1 − y)ex . f (x) = J(f ) After calculating the Jacobian.508 POWER SERIES ∂2f ∂x2 so that e−(x 2 = (0. (0. the Taylor series takes the form 0 1 x y−1 x− 0 1 + ··· .0) ∂2f ∂y 2 = −2.0) ∂2f ∂xdy = 0. f (x) = f (x) g(x) . . We can also generalize the Taylor series from scalar. where ∂f  ∂x (x0 ) J(f )(x0 ) =   ∂g (x0 ) ∂x   ∂f (x0 )  ∂y   ∂g (x0 ) ∂y is called the Jacobian matrix. this gives f (x) = 1 0 0 −1 + ··· = x 1−y + ··· . where we study nonlinear diﬀerential equations.0) +y 2 ) = 1 − x2 − y 2 + · · · .to vector-valued functions of two variables. This result proves to be very useful in Chapter 9. Since f (x0 ) = 0. Note that the deﬁnition of the Jacobian can easily be generalized to higher dimensional vector ﬁelds. Using the Taylor series for each component of this vector we can write f (x) = f (x0 ) + J(f )(x0 )(x − x0 ) + · · · . Let’s deﬁne a vector ﬁeld. APPENDIX 4 Sequences of Functions The concepts that we brieﬂy describe in this section are used in Chapter 5, where we discuss the convergence of Fourier series. Consider a sequence of functions, {fk (x)} for k = 1, 2, . . . , deﬁned on some interval of the real line, I. We say that {fk (x)} converges pointwise to f (x) on I if limk→∞ fk (x0 ) exists for all x0 ∈ I, and fk (x0 ) → f (x0 ). For example, consider the sequence {fk (x)} = {x + 1/k} on R. This sequence converges pointwise to f (x) = x as k → ∞. The functions fk (x) and their limit, f (x) = x, are continuous. In contrast, consider the sequence {fk (x)} = {xk } on [0, 1]. Although each of these functions is continuous, as k → ∞, fk (x) converges pointwise to f (x) = 0 for 0 x < 1, 1 for x = 1, which is not continuous. As a ﬁnal example, consider the sequence fk (x), deﬁned on x 0, with fk (x) = k 2 x(1 − kx) for 0 x 1/k, 0 for x 1/k, (A4.1) 1/k, fk (x0 ) → 0 as which is illustrated in Figure A4.1. Since fk (x) = 0 for x k → ∞ for all x0 > 0. Moreover, fk (0) = 0 for all k, so we conclude that {fk (x)} converges pointwise to f (x) = 0, a continuous function. However, the individual members of the sequence don’t really provide a good approximation to the pointwise limit, ﬁrstly because the maximum value of fk (x) is k/4, which is unbounded as k → ∞, and secondly because ∞ ∞ fk (x) dx = 1/6, 0 0 f (x) dx = 0. In order to eliminate this sort of behaviour, we need to introduce the concept of uniform convergence. A sequence of functions {fk (x)} is uniformly convergent to a function f (x) on an interval of the real line, I, as k → ∞ if for every > 0 there exists a positive integer, K( ), which is not a function of x, such that |fk (x)−f (x)| < for all k K( ) and x ∈ I. This means that, for suﬃciently large k, we can make fk (x) arbitrarily close to f (x) over the entire interval I. For example, to see that the sequence deﬁned by (A4.1) does not converge uniformly to zero, note that, for a given value of , we can only guarantee that |fk (x) − f (x)| = fk (x) < for k 1/x, which is not independent of x. 510 SEQUENCES OF FUNCTIONS Fig. A4.1. The sequence of functions given by (A4.1). Theorem A4.1 If {fk (x)} is a sequence of continuous functions deﬁned on an interval of the real line, I = [a, b], that converges uniformly to f (x), then (i) f (x) is continuous (ii) for every x ∈ I x k→∞ x lim fk (s) ds = a a f (s) ds. A proof of this theorem can be found in any textbook on analysis. APPENDIX 5 Ordinary Diﬀerential Equations Chapters 1 to 4 are concerned with the solution of linear second order diﬀerential equations with variable coeﬃcients. In this appendix, we will review simple methods for solving ﬁrst order diﬀerential equations, and second order diﬀerential equations with constant coeﬃcients. A5.1 Variables Separable We will start by looking at a class of equations that can be rewritten in the form dy = f (t). dt These equations can be integrated with respect to t to give g(y) g(y) and, using the chain rule, g(y) dy = f (t) dt. dy dt = dt f (t) dt, (A5.1) Example Determine the solution of the diﬀerential equation dy + ey cos t = 0, dt subject to the condition that y(0) = 1. Firstly, we convert the equation to the form (A5.1), so that dy = − cos t. dt Integrating with respect to t we have e−y −e−y = − sin t + C. Using the boundary condition we ﬁnd that C = −e−1 , and rearranging gives the solution 1 y(t) = log . e−1 + sin t 512 ORDINARY DIFFERENTIAL EQUATIONS A5.2 Integrating Factors We now consider the linear, ﬁrst order diﬀerential equation dy + P (x)y = Q(x). dx You should recognize this as a type of equation that can be solved by ﬁnding an x integrating factor. If we multiply through by exp P (t)dt , where t is a dummy variable for the integration, we have x exp P (t)dt dy + exp dx x x P (t)dt P (x)y = Q(x) exp d dx x P (t)dt . x We can immediately see that the left hand side of this is so we have d dx exp x exp P (t)dt y , P (t)dt y = Q(x) exp P (t)dt . This can be directly integrated to give x x x s y = A exp − P (t)dt + exp − P (t)dt Q(s) exp P (t)dt ds, where A is a constant of integration. Here it is important to notice the structure of the solution as y = yh + yp where x yh = A exp − P (t)dt is the solution of the homogeneous diﬀerential equation, dyh + P (x)yh = 0, dx and x x s yp = exp − P (t)dt Q(s) exp P (t)dt ds is the particular integral solution of the inhomogeneous diﬀerential equation dyp + P (x)yp = Q(x). dx The key idea of an integrating factor used here may seem rather like pulling a rabbit out of a hat. In fact, the derivation can be performed systematically using the idea of a Lie group (see Chapter 10). Example Let’s ﬁnd the solution of the diﬀerential equation dy − y = ex , dx A5.3 SECOND ORDER EQUATIONS WITH CONSTANT COEFFICIENTS 513 subject to the condition that y(0) = 0. The integrating factor is x exp so that we have e−x −1 dx = e−x , dy d −x − e−x y = e y = 1. dx dx We can integrate this to obtain e−x y = x + c, where c is a constant. Since y(0) = 0, we must have c = 0 and hence y = xex . A5.3 Second Order Equations with Constant Coeﬃcients We will now remind you how to solve second order ordinary diﬀerential equations with constant coeﬃcients through a series of examples. Example Solve the second order ordinary diﬀerential equation d 2y dy + 2y = 0, +3 dx2 dx subject to the boundary conditions that y(0) = 0 and y (0) = 1. We can solve constant coeﬃcient equations by seeking a solution of the form y = emx , which we substitute into the equation in order to determine m. This gives d 2y dy + 2y = m2 + 3m + 2 emx = 0. +3 dx2 dx Since emx is never zero, we require m2 + 3m + 2 = (m + 2)(m + 1) = 0. This gives m = −1 or m = −2. The general solution of the equation is therefore y(x) = Ae−x + Be−2x , with A and B constants. The ﬁrst boundary condition, y(0) = 0, yields A + B = 0, whilst the second, y (0) = 1, gives −A − 2B = 1. (A5.3) (A5.2) We now need to solve (A5.2) and (A5.3), which gives us A = 1 and B = −1, so that the solution is y(x) = e−x − e−2x . It is possible to solve this equation by direct integration, but the method we have presented above is far simpler. Notice that it is not necessary that both boundary conditions are imposed at the same place. Boundary conditions at diﬀerent values of x will just lead to slightly more complicated simultaneous equations. 514 ORDINARY DIFFERENTIAL EQUATIONS Example: Equal roots Solve the diﬀerential equation dy d 2y + y = 0, +2 dx2 dx subject to the boundary conditions y(0) = 1 and y (0) = 0. Again we can look for a solution of the form y = emx . We ﬁnd that m satisﬁes the quadratic equation m2 + 2m + 1 = (m + 1)2 = 0, so that m = −1 is a repeated root, and we have only determined one solution. In fact, the full solution has the form (A5.4) y(x) = Ae−x + Bxe−x . The boundary condition y(0) = 1 gives A = 1. Since y (x) = −Ae−x + B(−x + 1)e−x , y (0) = 0 gives −A + B = 0, and hence B = 1. The solution is therefore y(x) = e−x + xe−x = (x + 1)e−x . Example: Imaginary roots Solve the diﬀerential equation d 2y + 4y = 0, dx2 subject to the boundary conditions y(0) = 1 and y (0) = 3. As usual, we seek a solution of the form y = emx , and ﬁnd that m satisﬁes 2 m + 4 = 0. This means that m = ±2i. The general solution is therefore y(x) = Ae2ix + Be−2ix . (A5.5) We could proceed with the solution in this form. However, it may be better to use eiα = cos α + i sin α. Substituting this into (A5.5) gives y(x) = A(cos 2x + i sin 2x) + B(cos 2x − i sin 2x) ˜ ˜ = (A + B) cos 2x + i(A − B) sin 2x = A cos 2x + B sin 2x. ˜ ˜ We have introduced new constants A and B, which we determine from the boundary ˜ = 1 and B = 3/2. The solution is therefore ˜ conditions to be A y(x) = cos 2x + 3 sin 2x. 2 Example: An inhomogeneous equation Solve the diﬀerential equation d 2y dy + 2y = ex , +3 dx2 dx subject to the boundary conditions y(0) = y (0) = 1. A5.3 SECOND ORDER EQUATIONS WITH CONSTANT COEFFICIENTS 515 Initially we consider just the homogeneous part of the equation and solve d 2 yh dyh + 2yh = 0. +3 2 dx dx As we saw earlier, the general solution of this is yh (x) = Ae−x + Be−2x . We now have to ﬁnd a function f (x), the particular integral, which, when substituted into the inhomogeneous diﬀerential equation, yields the right hand side. We notice that when we diﬀerentiate ex we get it back again, so we postulate that f (x) takes the form αex . Substituting y = f (x) = αex into the diﬀerential equation gives df d 2f + 2f = αex + 3αex + 2αex = ex . +3 dx2 dx From this expression we ﬁnd that we need α = 1/6. The solution of the inhomogeneous equation is therefore 1 y(x) = Ae−x + Be−2x + ex . 6 It is at this point that we impose the boundary conditions, which show that A = 5/2 and B = −5/3, and hence the solution is y(x) = 5 −x 5 −2x 1 x e − e + e . 2 3 6 Example: Right hand side a solution of the homogeneous equation Solve the diﬀerential equation equation d 2y dy + 2y = e−x , +3 2 dx dx subject to the boundary conditions y(0) = y (0) = 1. Following the previous example, we need to ﬁnd a function which, when substituted into the left hand side of the equation, yields the right hand side. We could try y = αe−x , but, since e−x is a solution of the equation, substituting it into the left hand side just gives us zero. Taking a lead from the case of repeated roots, we try a solution of the form y = f (x) = αxe−x . Since f (x) = αxe−x , f (x) = α(e−x − xe−x ) = α(1 − x)e−x , f (x) = α(−1 − (1 − x))e−x = α(x − 2)e−x , on substituting into the diﬀerential equation we obtain α(x − 2)e−x + 3α(1 − x)e−x + 2αxe−x = 3αe−x = e−x , which gives α = 1/3. The solution is therefore 1 y(x) = Ae−x + Be−2x + xe−x . 3 516 ORDINARY DIFFERENTIAL EQUATIONS We now need to satisfy the boundary conditions, which give A = 8/3 and B = −5/3, and hence 8 5 1 y(x) = e−x − e−2x + xe−x . 3 3 3 This technique is usually referred to as the trial solution method. Clearly it is not completely satisfactory, as it requires an element of guesswork. We present a more systematic method of solution in Chapter 1. APPENDIX 6 Complex Variables The aim of this brief appendix is to provide a reminder of the results that are needed in order to be able to invert Laplace transforms using the Bromwich inversion integral (6.5), and to understand the material on the asymptotic evaluation of complex integrals in Section 11.2. We have had to be selective in what we have presented, and note that this appendix is no substitute for a proper course on complex variables! A good textbook is Ablowitz and Fokas (1997). A6.1 Analyticity and the Cauchy–Riemann Equations Consider a complex-valued function of a complex variable, f (s), with s = x + iy. The natural way to deﬁne the derivative of f is df f (s + ∆s) − f (s) = lim , ds ∆s→0 ∆s provided that this limit is independent of the way that ∆s = ∆x + i∆y tends to zero. A function f (s) is said to be analytic in some region, R, of the complex s-plane if df /ds exists and is unique in R. Theorem A6.1 If the complex-valued function f (s) = u(x, y) + iv(x, y), where s = x + iy and u, v, x and y are real, is analytic in a region, R, of the complex s-plane, then ∂v ∂u = , ∂x ∂y ∂v ∂u =− . ∂x ∂y (A6.1) These are known as the Cauchy–Riemann equations. Proof By deﬁnition, df f (s + ∆s) − f (s) = lim ds ∆s→0 ∆s = lim If ∆y = 0, ∂u ∂v u(x + ∆x, y) − u(x, y) + i {v(x + ∆x, y) − v(x, y)} df = lim = +i , ds ∆x→0 ∆x ∂x ∂x u(x + ∆x, y + ∆y) + iv(x + ∆x, y + ∆y) − u(x, y) − iv(x, y) . ∆x→0 ∆x + i∆y ∆y→0 518 COMPLEX VARIABLES whilst if ∆x = 0, df ∂v ∂u u(x, y + ∆y) − u(x, y) + i {v(x, y + ∆y) − v(x, y)} = lim = −i . ds ∆y→0 i∆y ∂y ∂y These are clearly not equal unless the Cauchy–Riemann equations hold. Theorem A6.2 If the Cauchy–Riemann equations, (A6.1), hold in a region, R, of the complex s-plane for some complex-valued function f (s) = u(x, y) + iv(x, y), where s = x + iy and u, v, x and y are real, then, provided that all of the partial derivatives in (A6.1) are continuous in R, f (s) is analytic in R. We will not give a proof of this here. A6.2 Cauchy’s Theorem, Cauchy’s Integral Formula and Taylor’s Theorem The contour integral of a complex-valued function f (s) along some contour C in the complex s-plane is deﬁned to be b f (s) ds = C a f (s(t)) ds (t) dt, dt where s(t) for a t b is a parametric representation of the contour C. By convention, the integral around a closed contour is taken in the anticlockwise direction. Theorem A6.3 (Cauchy’s theorem) If f (s) is a single-valued, analytic function in a simply-connected domain D in the complex s-plane, then, along any simple closed contour, C, in D, f (s) ds = 0. C Proof If f (s) = u + iv and ds = dx + i dy, f (s) ds = C C (u dx − v dy) + i C (v dx + u dy). If df /ds is continuous in D, then u and v have continuous partial derivatives there. We can therefore use Green’s theorem in the plane to write f (s) ds = − C D ∂v ∂u + ∂x ∂y dx dy + i D ∂u ∂v − ∂x ∂y dx dy. Since f is analytic in D, the Cauchy–Riemann equations hold, and the result is proved. If df /ds is not continuous in D, the proof is rather more technical, and we will not give it here. and z0 is a point inside C. centred on the point s = z. An extension of Cauchy’s integral formula is Theorem A6. (s − z)n+1 (A6.3) ∞ . analytic function in a simply-connected domain D in the complex s-plane. (s − z) C where C is any simple closed contour in D that encloses the point s = z. we ﬁnd that ds = (s − z) 2π 0 Cδ iδeiθ dθ = 2πi. f (s) ds = (s − z) f (s) ds.5 If f (s) is a single-valued. (s − z) C Cδ which we can rewrite as C f (s) ds = f (z) (s − z) Cδ ds + (s − z) Cδ By writing the ﬁrst of these two integrals in terms of polar coordinates. and are given by f (n) (z) = n! 2πi f (s) ds.4 (Cauchy’s integral formula) If f (s) is a single-valued.A6. Theorem A6. the result is proved.6 (Taylor’s theorem) If f (z) is analytic and single-valued inside and on a simple closed contour.2 CAUCHY’S THEOREM AND TAYLOR’S THEOREM 519 Theorem A6. then f (z) = f (z0 ) + (z − z0 )f (z0 ) + 1 1 (z − z0 )2 f (z0 ) + · · · = (z − z0 )n f (n) (z0 ). δeiθ We can deal with the second integral by noting that. This means that f (s) − f (z) ds (s − z) |f (s) − f (z)| |ds| |s − z| δ |ds| = 2π . C. The proof of this is similar to that of Cauchy’s integral formula. Proof Let Cδ be a small circle of radius δ. Cδ Cδ Cδ Since → 0 as δ → 0. Using Cauchy’s theorem. then all the derivatives of f exist in D. 2! n! n=0 (A6. |f (s) − f (z)| < for suﬃciently small |s − z| = δ. analytic function in a simply-connected domain D in the complex s-plane. then f (z) = 1 2πi f (s) ds. (s − z) f (s) − f (z) ds. since f (s) is continuous. s = z +δeiθ .2) C where C is any simple closed contour in D that encloses the point s = z. for any point z inside γ. and let γ be any circle centred on z0 with radius ρ < δ. Equation (A6.520 COMPLEX VARIABLES Proof Let δ be the minimum distance from z0 to the curve C. w − z0 (w − z0 ) (w − z0 ) and this series is uniformly convergent inside γ.7 (Cauchy’s inequality) If f (z) is analytic for |z| < R with ∞ f (z) = n=0 an z n . z n+1 |f (z)| |dz| rn+1 1 Mr Mr 2πr = n . This means that f (z) = 1 2πi γ 1 f (w) dw = w−z 2πi + (z − z0 ) n γ f (w) 1 dw + (z − z0 ) w − z0 2πi f (w) γ n+1 dw f (w) γ (w − z0 ) 2 dw + ··· 1 2πi (w − z0 ) + ··· . If z0 = 0 we get the simpler form f (z) = f (0) + zf (0) + 1 2 z f (z) + · · · . Proof By Taylor’s theorem. so that. an = so that |an | 1 2π |z|=r 1 2πi |z|=r f (z) dz.2) then gives us (A6. By Cauchy’s integral formula. if Mr is the supremum of |f (z)| on the circle |z| = r < R.3). This series converges in any circle centred on z0 with radius less than δ. f (z) = We also have 1 1 1 1 = = z−z w−z w − z0 − z + z 0 w − z0 1 − w−z0 0 = 1 w − z0 1+ z − z0 (z − z0 ) + 2 + ··· w − z0 (w − z0 ) n 2 1 2πi C f (w) 1 dw = w−z 2πi γ f (w) dw. we have |an | < Mr /rn . w−z = 1 z − z0 (z − z0 ) + 2 + ··· + n+1 + · · · . 2! Theorem A6. since |z − z0 | / |w − z0 | < 1. |z − z0 | < ρ < δ. 2π rn+1 r . then. s2 . we will not give a proof here. which is the contour C deformed into small circles around each singular point and joined to the original ˆ position of C by straight contours. If f is not analytic within C. sn . c−1 = 0. Theorem A6. Since f is analytic within C.A6. C For an analytic function.1. s = s1 .8 Any function f (s) that is single-valued and analytic in an annulus r1 |s − s0 | r2 for r1 < r2 real and positive has a series expansion ∞ f (s) = n=−∞ cn (s − s0 )n . The coeﬃcient c−1 is called the residue of f (s) at s = s0 . Although we quote this important theorem. which underlies the techniques of residue calculus.3 The Laurent Series and Residue Calculus Theorem A6. and we have Cauchy’s theorem. We do note however that f (s) ds = 2πic−1 .3 THE LAURENT SERIES AND RESIDUE CALCULUS 521 A6. where aj is the residue of f (s) at s = sj . which is convergent for r1 < |s − s0 | < r2 . Then n f (s) ds = 2πi C j=1 aj . .9 (The residue theorem) Let C be a simple. f (s) ds = 0. The integrals along the straight parts of C cancel out in the limit as they approach each other. ˆ Proof Consider the closed contour C. so we conclude that n f (s) ds = ˆ C C f (s) ds − j=1 Cj f (s) ds = 0. . . ˆ C ˆ by Cauchy’s theorem. This is known as the Laurent series. this result shows that we can calculate the integral by determining the coeﬃcient of 1/(s − s0 ) in the Laurent series. . closed contour. and let f (s) be a complex-valued function that is single-valued and analytic on and within C. and its coeﬃcients are given by cn = 1 2πi f (s) ds (s − s0 )n+1 C for any simple closed contour that encloses s = s0 and lies in the annulus r1 |s − s0 | r2 . except at n isolated singular points. . shown in Figure A6. Note that. A6.1. closed contour is equal to 2πi times the sum of the residues enclosed by C.5). dsm−1 1 g(s). in particular to evaluate the integral in the Laplace inversion formula.522 COMPLEX VARIABLES and hence n f (s) ds = 2πi C j=1 aj . (6. and hence d m−1 g = (m − 1)!a−1 + m! a0 (s − s0 ) + · · · . C x x Cn x • • •• C1 C2 Fig. The residue theorem simply states that the integral around a simple. (s − s0 )m a−(m−1) a−m + + ··· . The contours C and Cj . and Laurent series expansion f (s) = we can write f (s) = where g(s) = a−m + a−(m−1) (s − s0 ) + · · · + a−1 (s − s0 )m−1 + · · · . for a function with a pole of order m at s = s0 . m (s − s0 ) (s − s0 )m−1 . This can be used in a variety of ways. 1 (Jordan) Let CJ be a semi-circular arc of radius R centred at s = s0 in the left half of the complex s-plane. CJ R→∞ for t a positive constant. for which m = 2.4 JORDAN’S LEMMA 523 which shows that the residue of f (s) at s = s0 is a−1 = d m−1 1 m {(s − s0 ) f (s)} (m − 1)! dsm−1 . for which m = 1. s=s0 In particular. . provided that |F (s)| → 0 uniformly as |s| → ∞ in the left half plane. and at a double pole.2. A6. If F (s) is a complex-valued function of s and |F (s)| → 0 uniformly on CJ as R → ∞. Lemma A6. a−1 = (s − s0 )f (s0 ). sin θ 2θ/π. s=s0 A6. at a simple pole. then lim est F (s) ds = 0.A6. a−1 = d 2 (s − s0 ) f (s) ds . Fig.4 Jordan’s Lemma Jordan’s lemma allows us to neglect an integral that often arises when evaluating inverse Laplace transforms of functions F (s). with the proviso that there is a circle of convergence rather than an interval of convergence. 2t A6.2. are convergent. singularities and other features associated with the complex plane.4) with zq(z) and z 2 r(z) analytic at z = 0. π |I| R ets0 K(R) 0 e−2tRθ /π dθ = π |ets0 | K(R) 1 − e−2tR → 0 as R → ∞. as shown in Figure A6. If we assume a series solution. Since |F (s0 + Reiθ )| → 0 as R → ∞. the convergence theorem for a real diﬀerential equation at a regular singular point. iθ and hence 3π/2 |I| R π/2 3π/2 ets0 +tRe F (s0 + Reiθ ) dθ iθ =R π/2 ets0 etR cos θ+itR sin θ F (s0 + Reiθ ) dθ 3π/2 = R ets0 π/2 π etR cos θ F (s0 + Reiθ ) dθ = R ets0 0 e−tR sin θ F s0 + Rei(θ + 2 ) π dθ . |F (s0 + Reiθ )| < K(R) for some real-valued. we will prove that power series solutions of w + q(z)w + r(z)w = 0. after making the change of variable θ = θ + π/2.3. .524 COMPLEX VARIABLES Proof On CJ . although their solutions tend to have a richer structure due to the appearence of branch cuts. ∞ w= n=0 an z n+c . positive function K(R). such that K(R) → 0 as R → ∞. In this section. (A6. In addition. so 3π/2 I= CJ est F (s) ds = π/2 ets0 +tRe F (s0 + Reiθ )iReiθ dθ. s = s0 + Reiθ . Note that these are precisely the conditions for Theorem 1. since sin θ 2θ/π.5 Linear Ordinary Diﬀerential Equations in the Complex Plane Linear ordinary diﬀerential equations in the complex variable z = x + iy can be studied in much the same way as those in a real variable. In particular. the solution of linear second order equations by power series works in the same way as the real variable case. rn rn If we take the modulus of (A6. Hence n=0 an z n converges for |z| < R as we claimed. Let R be the smaller of these two radii.5 ORDINARY DIFFERENTIAL EQUATIONS IN THE COMPLEX PLANE 525 and expansions zq(z) = q0 + q1 z + · · · .7). for n λ. we can deﬁne a sequence of coeﬃcients An with |an | An using An = |an | n−1 n(n − λ)An = K s=0 An (µ + s + 1) /rn−s The deﬁnitions of An−1 and An for n λ show that for 0 n < λ. in general. n−1 z 2 r(z) = r0 + r1 z + · · · . r r If we now divide through by n(n−λ)An−1 and let n → ∞. The indicial equation always has one root that will produce a well-deﬁned series solution. they each have Taylor expansions that converge in some disc centred on z = 0. But. . Since zq(z) and z 2 r(z) are analytic at z = 0. from our ∞ deﬁnition. we can now see that |qn | n−1 n |an | |n + c1 − c2 | K s=0 |as | |c1 | + s + 1 . .6). An−1 An−1 = K(µ + n) . . (A6. n(n − λ)An − (n − 1)(n − 1 − λ) . there exist r and K = K(r) such that K K . We can rewrite (A6. Then. we ﬁnd that An /An−1 → ∞ 1/r.A6. and we have used the fact that c1 + c2 = 1 − q0 . then the indicial equation is a0 {c(c − 1) + cq0 + r0 } = 0. by Cauchy’s inequality (Theorem A6. This shows that the radius of convergence of n=0 An z n is r. an f (c + n) + s=0 as {(c + s)qn−s + rn−s } = 0. when special care would be needed.5) in the form n−1 an n (n + c1 − c2 ) = − s=0 as {(c1 + s) qn−s + rn−s } . |an | An . We will proceed assuming that we are not dealing with a diﬃcult case in which f (c + n) vanishes.6) where c1 and c2 are the solutions of the indicial equation. |rn | for r < R and n = 0. 1. (A6. so that the radius of convergence of n=0 an z n is at least r ∞ for all r < R.5) where f (c) = c(c − 1) + cq0 + r0 . . rn−s Writing |c1 − c2 | = λ and |c1 | = µ. and. 2. This can be changed using the format command. In this appendix. >> (4+5-6)*14/5. † MATrix LABoratory ‡ We will often use the built-in subroutine ode45 to solve ordinary diﬀerential equations. our aim is to show complete newcomers to MATLAB enough for them to be able understand the material in the book that uses MATLAB. In particular. exp(0. the graphical display of data. and contains extremely eﬃcient routines. The three expressions separated by commas are evaluated successively.APPENDIX 7 A Short Introduction to MATLAB MATLAB† is a programming language and environment that is both powerful and easy to use. Γ(x). For example.1 Getting Started MATLAB can be used in two ways.4918 ans = 1.4). The default is to display ﬁve signiﬁcant digits of the answer. For a more extensive guide to the power of MATLAB. § See Trefethen and Bau (1997) for an introduction to numerical linear algebra that uses MATLAB extensively to illustrate the material. and much more. gamma(0. and scripts. and the calculated answers displayed. and obtain results immediately. as ﬁles. MATLAB has all the functionality of a scientiﬁc calculator.0000 Note that pi is a built-in approximation to π and gamma is the gamma function. and call them from MATLAB. which are just sequences of MATLAB commands. A7. Alternatively. . the numerical integration of ordinary diﬀerential equations‡.4. we can simply type commands into MATLAB at the command prompt (>>). It contains built-in functions that perform most of the tasks that we need in order to illustrate the material in this book.3.5)/sqrt(pi) ans = 8. We introduce several other MATLAB functions in the main part of the text. for example. we can save functions.4000 ans = 1. For example. An explanation of this is given in Section 9. MATLAB was originally developed to provide an easy way to access programs that act on matrices. which have arguments and return values. As we shall see. to solve systems of linear equations and to extract the eigenvalues of a matrix§. see Higham and Higham (2000). for example. as storage is allocated as it is needed.2204e-016 ans = 1.A7. D = linspace(1.0000i † The precise value of eps varies depending upon the system on which MATLAB is installed. Vectors and Matrices As well as being able to perform calculations directly.49182469764127 ans = 1. exp(0. can only perform calculations with a ﬁnite precision. each calculation is subject to an error of the order of eps.. F = [1+i.7500 2. 2+2*i] A = 1 B = 1 2 3 C = 0 0.4).45. the machine precision. C = 0:0. 4 5 6]. Most programming languages require the user to declare the type of each variable.2. A calculation of sin π = 0 leads to an error comparable in magnitude to eps.4000 D = 1.0000 1. which corresponds to double precision arithmetic. VECTORS AND MATRICES 527 >> format long >> (4+5-6)*14/5.5000 1. gamma(0.2246e-016 This tells us that the machine precision is about 10−16 †.0000 + 2.. which can be accessed with the command help.2 VARIABLES. MATLAB allows the use of variables. >> A = 1. It should be noted that MATLAB.0000 E = 1 2 3 4 5 6 F = 1.sin(pi) ans = 2. This is not necessary in MATLAB. E = [1 2 3.40000000000000 ans = 1. .0000 + 1.1:0. For example.3000 0. A7. Typing help format at the command prompt lists the many other formats that are available.5)/sqrt(pi) ans = 8.2000 0.0000i 2.00000000000000 MATLAB has comprehensive online documentation.5)’. for example a double precision scalar or a single precision complex array.2 Variables. >> eps. since numbers can only be represented digitally with a ﬁnite precision.1000 0. In particular. B = 1:3.2500 1. like any programming language. c denotes the spacing as the vector runs from a to b. — C is another row vector. For example. The semicolon is used to denote a new row. For example.D ??? Error using ==> Matrix dimensions must agree. 1. However. X and x denote distinct variables.0000 + 15. The function linspace(a./C.0000 + 12. just denotes a vector with entries running from a to b at unit intervals.3500 1. It is also possible to operate on matrices element by element. The matrices B and C must be of the same size. — B is a 1 × 3 matrix (a row vector).2000 2.0000i 12. a:b. as an example. for example >> B. as another example. we can just treat this as a scalar. >> C + D’ ans = 1.0000 1.2000 0. — E is a 2 × 3 matrix. For example.7000 >> C . real entry. The colon notation. Note that i is the square root of −1. Matrix addition and subtraction take the obvious form.n) generates a row vector running from a to b with n evenly spaced points.0500 2. We also need to remember that MATLAB variables are case-sensitive. and its adjoint if A is complex. Remember. .0000i 15. Note that A’ is the transpose of A when A is real. MATLAB has a rich variety of functions that operate upon matrices (type help elmat).0000 + 9. Of course.528 A SHORT INTRODUCTION TO MATLAB Note that . >> E’*F ans = 9. In the notation a:c:b.. 2.0000 This takes each element of B and multiplies it by the corresponding element of C. matrix multiplication takes the obvious form. which we do not really need for the purposes of this book. allows the current command to overﬂow onto a new line. B.. — D is a 5 × 1 matrix (a column vector). The use of a full stop in.b. — A is a 1 × 1 matrix with the single. an m1 × n1 matrix can only be multiplied by an m2 × n2 matrix if n1 = m2 .0000i >> E*F ??? Error using ==> * Inner matrix dimensions must agree. always denotes element by element calculation. — F is a complex column vector.4000 The one exception to the rule that matrices must have the same dimensions comes when we want to add a scalar to each element of a matrix.*C ans = 0 0.6000 1. 0998 4 25 9 36 0.1000 1.4000 This adds 1 to each element of C. For example >> sin(C) ans = 0 >> E.3 0. Most of MATLAB’s built-in functions can also take matrix arguments.3000 1. >> E(:. E(2.0000 1.2000 1. if we save the commands . For example. It is often useful to be able to extract rows or columns of a matrix.3894 User-Deﬁned Functions We can add to MATLAB’s set of built-in functions by deﬁning our own functions.2955 0.1987 0.A7.^2 ans = 1 16 A7.3). even though 1 is a scalar.3 USER-DEFINED FUNCTIONS 529 >> C + 1 ans = 1. For example.:) ans = 3 6 ans = 4 5 6 extracts the third column and second row of E. .  in a ﬁle named x2.2500 It is always prudent to write functions in such a way that they can take a matrix argument. >> x2(0:-0.0100 0. The most useful of these is the plot command. which must. Note that the semicolon suppresses the output of a. if we had written x2 using a = x^2.0900 0. For example.5) ans = 0 0. For example. we will be able to access this function.y) plots the data contained in the vector y against that contained in the vector x.4 Graphics One of the most useful features of MATLAB is its wide range of diﬀerent methods of displaying data graphically (type help graphics).1600 0. of course have the same dimensions.m. MATLAB would return an error unless x were a square matrix.^2. A7.1:-0.  function a = x2(x) a = x. when the result would be the matrix x*x. The basic idea is that plot(x. from the command prompt. For example.0400 0. which simply squares each element of x. and the command legend(’label 1’.x2. to reset the limits of the y axis. ’label 3’) adds a legend that names each line. There are many other ways of controlling the axes (type help axis).x2.y2. there is also an easy way of plotting functions to get an idea of what they look like. and individual points plotted if necessary. A7. ’label 2’. YLim([a b]). Note that MATLAB automatically picks an appropriate range for the y axis.x3. y = log(gamma(x)).y1. plot(x1. x 10. For a comprehensive list of options. for example.’o-’) plots the ﬁrst data set as a dashed line.y) produces a plot of the logarithm of the gamma function with 1 add axis labels and titles with >> xlabel(’x’). We can Fig. and the third as a solid line with circles at the discrete data points. The style of each line can be controlled. plot(x. For example. ezplot(@sin) . By default. For example. type help plot.y2.1. For example. the lines are produced in diﬀerent colours to distinguish between them.y3) produces three lines in the obvious way. plot(x1. A plot of log Γ(x).y3.’x’.1.500).10. ylabel(’log \Gamma(x)’) >> title(’The logarithm of the gamma function’) This produces the plot shown in Figure A7. produced by MATLAB. the second as discrete crosses. As we shall see in the main part of the book.’--’. We can override this using. The command plot can produce graphs of more than one function.x3.530 A SHORT INTRODUCTION TO MATLAB >> x = linspace(1.y1. which produces a mesh plot of a function of two variables (see Section 2. This includes functions supplied by the user. the k th entry of A. but makes the script more readable. ezplot(@x2) plots the function x2 that we deﬁned earlier.A7. since the last tic. if we save the script # x = linspace(0. toc elapsed_time = 0. A7. We can see that allocating k 2 to A(k).2*pi.5 PROGRAMMING IN MATLAB 531 produces a plot of sin x. by passing the function handle as a parameter. For example.sin(k*x)) pause end " ! to a ﬁle sinkx. in seconds. A = (1:50000). FOR loops should be avoided where possible. sin 3x and sin 4x for 0 x 2π. using element by element squaring of 1:50000. toc elapsed_time = 0. FOR Loops There are many examples of FOR loops in the main text. and then type sinkx at the command prompt. For example. for k = 1:4 plot(x. for example. so this script successively plots sin x. We will also use the command ezmesh. so that. A(k) = k^2. ELSEIF structure. the commands in sinkx. executes about forty times faster than performing the . and the variable toc contains the time elapsed. Note that the indentation of the lines is not necessary. if a script needs to execute quickly.1 for an explanation). As a general rule.4010 The command tic sets an internal clock to zero. >> tic. These allow functions to take other functions as arguments. known as vectorizing the function. The quantity @sin is called a function handle.500). in favour of matrix and vector operations.5 Programming in MATLAB All of the control structures that you would expect a programming language to have are available in MATLAB. The basic syntax is for c = x statements end The statements are executed with c taking as its value the successive columns of x. end.6. for k = 1:50000.m. The editor supplied with MATLAB formats scripts in this way automatically. The command pause waits for the user to hit any key before continuing. The most commonly used of these are FOR loops and the IF. sin 2x.^2.m are executed. This is a MATLAB script.0100 >> tic. expecting MATLAB to successively increase the size of A at each iteration of the loop. Another way of speeding up a MATLAB script is to preallocate the array. for k = 1:50000. The IF. end. In the above sequence. the array A already exists. A(k) = k^2. For an in-depth discussion of vectorization. else statements end For example. and is of the correct size when the FOR loop executes. we can see that the execution time becomes huge compared with that when A is already known to be a vector of length 50000. >>tic. see Van Loan (1997). consider the function ' function f = f1(x) f = zeros(size(x)).toc elapsed_time = 94.532 A SHORT INTRODUCTION TO MATLAB same task using a FOR loop.1160 If we set A to be the empty array. . ELSEIF Structure The basic syntax for an IF. for k = 1:length(x) if (x(k)<0)|(x(k)>3) disp(’x out of range’) elseif x(k)<1 f(k) = x(k)^2. and then allocate k 2 to A(k). ELSEIF statement is if expression statements elseif expression statements elseif expression statements . else f(k) = exp(-x(k)). elseif x(k)<2 f(k) = cos(x(k)). end end &$ % . Consider >>A = [].

else disp(’x out of range’) end & % Note that ==. ELSEIF statement. which uses a FOR loop and an IF. y = f1(x). this preallocation of the matrix signiﬁcantly speeds up the execution of the function. not =. it is still not as eﬃcient as it could be.1) far more eﬃciently than f1. As we have already discussed.4320 >> tic.. ..^2 + ((x>=1)&(x<2)). and displays a warning if any of the elements of x lie outside the range [0. and zeros elsewhere. The function f2 evaluates f (x) in a vectorized manner.1300 It is clear that the vectorized function f2 evaluates (A2.*x. The function length(x) calculates the length of a vector x. The vector (x>=0) is the same size as x. at the points in the vector x. In fact. and contains ones where the corresponding element of x is positive or zero.*cos(x). Now consider >> tic.*exp(-x). Note that we initialize f as a vector of zeros with the same dimensions as x using f = zeros(size(x)). although this is a transparent way of programming the function. The logical function or is denoted by | in MATLAB. Consider the function \$ ' function f = f2(x) if ((x<0)|(x>3)) == zeros(size(x)) f = ((x>=0)&(x<1)).A7.2. y = f2(x). toc elapsed_time = 1. which is plotted in Figure A2. avoiding the use of a FOR loop. toc elapsed_time = 0.1).5 PROGRAMMING IN MATLAB 533 This evaluates the function (A2. is used to test the equality of two matrices. + ((x>=2)&(x<=3)). 3]. where the function is deﬁned. The operator & acts in the obvious way as the logical function and.

and Ludford. Cambridge University Press.. Translations Series 2. N. Aref. Glendinning.C. Garabedian. 1994. 143.J. J. and Fokas.. Oxford University Press..C. Cambridge University Press. 1990.F. and weakly damped oscillators’. D. M. 65. 296.. Theory of Laminar Flames.. J. H. Wave Motion.. 1983. 1969. Grobman. Appl. A. 1955. Similarity Methods for Diﬀerential Equations. Bourland. Math. ‘Stirring by chaotic advection’. 1991. John Wiley. Billingham. An Introduction to Dynamical Systems. Springer-Verlag. Handbook of Elliptic Integrals for Engineers and Scientists.J. 880. Billingham.Bibliography Ablowitz. Cambridge University Press. 1970. J. G. ‘The modulated phase shift for strongly nonlinear. E. Chapman and Hall.. Courant. ‘Steady state solutions for strongly exothermic thermal ignition in symmetric geometries’. and King. 325–351. Dokl. Billingham. 737–748.. Amer. J. Grindrod. 1995. and Panasenko.. 1964..M. Guckenheimer. 6.H. P.S. SIAM J. McGraw-Hill. R. Dynamical Systems . 1971. John Wiley. and Place.F. D. Kluwer.J. N. 1–21. and Mendoza. C. and Noble.. E. 2001. P. Elementary Fluid Dynamics. Billingham. J. J. 1. Akad.J. Telecomunications Engineering.G. Coddington. J. P. Gel’fand.E.J. B.W. IMA J. Oxford University Press. Stability.. Arrowsmith. Patterns and Waves: The Theory and Applications of Reaction-Diﬀusion Equations. Soc. Math. J. Methods of Mathematical Physics. D.. 48. 283–313. Math. slowly varying. ‘Homeomorphisms of systems of diﬀerential equations’. A.. Homogenization: Averaging Processes in Periodic Media.. and King.M. Buckmaster. 1977. and Needham. Brauer. Benjamin. Nauk SSSR. P. I. and Holmes. Complex Variables. 2000. Interscience.. Fluid Mech. 128. Partial Diﬀerential Equations. and Pearson. D.. Byrd. F. J. and Haberman.D. 1982. Bluman. G. Springer-Verlag. 1988. Cambridge University Press. 1989. P..S. C. G. Partial Diﬀerential Equations: Theory and Technique. 29..K.. ‘A note on the properties of a family of travelling wave solutions arising in cubic autocatalysis’.D. Flowers. Qualitative Theory of Ordinary Diﬀerential Equations. of Appl. Carrier.S. J. 1990. 295–381. ‘The interaction of a ﬂuid/ﬂuid interface with a ﬂat plate’.R. G..... F.. Acheson.. Academic Press. 1937.M.A. Properties of Matter. 1991. 1997. 1959. Cambridge University Press. and Smith.A. D. A. ‘Some problems in the theory of quasilinear equations’. and Hilbert. Nonlinear Oscillations. D. 1988. Fluid Mech. R. J. 1984. Dynamics and Stability of Systems... 1963. Dunlop. 1974. Bakhvalov. and Levinson.. Instability and Chaos: An Introduction to the Theory of Nonlinear Diﬀerential Equations. Theory of Ordinary Diﬀerential Equations. 33–49. and Cole.

An Introduction to Interactive Boundary Layer Theory.. S. G. Lorenz.J..C. 1988. Chem. L.K. Theoretical Aerodynamics. MATLAB Guide. M. Springer-Verlag. Cambridge University Press.. ‘Periodic approximations to an elliptic function’. 1968.R. 1960. ‘Transport and mixing in Stokes ﬂow: the eﬀect of chaotic dynamics on the blinking stokeslet’. McGraw-Hill. 2000. F. K¨rner. Scott.. K. Oxford University Press. Ordinary Diﬀerential Equations. 97.O. Lunn. 1956. J. Prentice Hall. Milne-Thompson. 1992. 348. 6191–6198. o Kreider. S. L. Kevorkian. 1997. Wiggins. SIAM.. 1–26. 1964.. 2001.R. Petrov. The Lorentz Equations: Bifurcations. and Bau.. 11. E. Cambridge University Press. Quantum Mechanics... J. King.M. Otto..J.BIBLIOGRAPHY 535 and Bifurcations of Vector Fields. 610–620. T. 1952. Cambridge University Press.. 1982. 1982.M.. 1959. Ostberg. 1990.C.J.. Applicable Analysis.. Sci. . Milne-Thompson. 1963 ‘Deterministic non-periodic ﬂows’. 2000. A. E..T. 1997.. Ottino. T. King... and Showalter. Dover.. Addison-Wesley. 1980. Landau. J..E. A. 1995. D. Van Dyke. P. Cambridge University Press.E. Lighthill.. Fluid Mech. P. ‘A lemma in the theory of structural stability of diﬀerential equations’.D.. The Kinematics of Mixing: Stretching. Kuller. A. Proc. J. 130–141. J. Theoretical Hydrodynamics. Amer.M. Macmillan. Hartman.L. Phil. L.. and Higham. S. 1988. D. Marlin.. Chaos and Strange Attractors. C. Chaos and Transport.L.. Cambridge University Press. and Lifschitz. Yannacopoulos. M. 1960. Math. Sobey. D. Perturbation Methods in Fluid Mechanics. Ince. Soc. S.R. Pergamon. A Treatise on the Theory of Bessel Functions. Wadsworth and Brooks.. D. G. The Fluid Mechanics of Large Blood Vessels.W. Macmillan. 23.M. ‘Mixed-mode oscillations in chemical systems’.. An Introduction to Turbulent Flow. Linear Algebra: An Introduction. Trefethen. 1990. 1989. Trans. 2000. 1958. J. Introduction to Applied Nonlinear Dynamical Systems and Chaos.. Schiﬀ. Process Control: Designing Processes and Control Systems for Dynamic Performance. Springer-Verlag. Kuzmak.. E...I. 1988. ‘The eﬀects of variable diﬀusivity on the development of travelling waves in a class of reaction-diﬀusion equations’. 1959. Cambridge University Press. Wiggins. SIAM.. 271–278. Mech. D. Partial Diﬀerential Equations: Analytical Solution Techniques.G. Symmetry Methods for Diﬀerential Equations. III. M. J. Appl. and Blake. Van Loan. Global Bifurcations and Chaos. Introduction to Fourier Analysis and Generalised Functions.N. Cambridge University Press. 1994. J.J. Soc. L. Morris. Phys. Fourier Analysis. ‘Asymptotic solutions of nonlinear second order diﬀerential equations with variable coeﬃcients’. Oxford University Press. 20. V.J. and Needham. Introduction to Scientiﬁc Computing.E. 1922. Watson.N. 1966. J. Springer-Verlag.. Pedley.. Van Nostrand Reinhold.. 730–744.W. 27.. C. R.. A First Course in Mechanics. Springer-Verlag.. Numerical Linear Algebra. 229–260. An Introduction to Linear Analysis. N. Hydon.. T. Atmos. 430. Lond. I. A. McGraw-Hill. 2000. 1990. A. Academic Press.N. Higham.. L. and Scott. Theory of Elasticity. and Perkins.. Math.F. Mathieu. Sparrow. R.

497 Cayley–Hamilton theorem. 195. 10. 450 comparison theorems. 498 complex analysis branch cut. 293 Fourier–Bessel series. 271 conservative systems. 402 pitchfork. 10 ν = 1 . 518 control bang-bang. 165 poles and residues. 71 orthogonality. 408 homoclinic. 99 aerofoils. 425. 253E Bernoulli shift map. 336 completely controllable. 214. 422 variables. 191 Joukowski transformation. 212 complete elliptic integral ﬁrst kind. 312 conformal mapping. 114 Bessel–Lommel theorem. 422. 190 Bessel functions. 499 characteristic variables. 457 subcritical Hopf. 334 second kind. 75 bifurcation codimension two. 79 Bessel’s inequality. 195 connection problems. 77. 517 Arrhenius law.Index Abel’s formula. 376 inhomogeneous. 96 bounded controls. 464 cantor middle-third set. 240 Dulac’s extension. 358 chemical oscillator. 114. 437 basin of attraction. 502 contour integral. 457 transcritical. 502 uniform. 54E. 58. 402 supercritical pitchfork. 28E 2 inhomogeneity. 79. 519 Cauchy’s theorem. 305 autocatalysis. 397 diagram. 464 Cauchy problem. 390. 418 536 . 467 Bernoulli’s equation. 348 conservation law. 274 Airy’s equation. 28. 379. 186 Cauchy–Schwartz inequality. 374. 454 global. 511 chaotic solutions. 521 composite expansions. 287. 164 Cauchy–Riemann equations. 240. 518 Cauchy–Kowalewski theorem. 462 centre manifold theorem. 392. 402. 34. 77 Lommel’s functions. 372 chain rule. 64 modiﬁed Bessel functions of ﬁrst and second kind. 449 islands. 390 ﬂip. 74 generating function. 502 piecewise. 402 subcritical pitchfork. 58. 180 Cauchy–Riemann equations. 418 vector. 399 supercritical Hopf. 186 keyhole contour. 121E. 402 saddle–node. 319 asymptotic balance. 42. 12. 408 Hopf. 69 Weber’s. 183. 399 tangent. 399. 388 boundary value problem. 434. 402 bifurcation theory. 420. 223 Bendixson’s negative criterion. 218 continuity. 419 Cantor diagonalization proof. 46. 444E completeness. 108 ν = 1. 373. 501 centre manifold local. 71. 350 analyticity. 298. 93. 370. 409 bang-bang control. 62 Bessel’s equation. 180 Cauchy’s integral formula. 398. 71 recurrence relations. 42. 163. 165 branch point. 460 characteristic equation.

55. 280. 270 Duﬃng’s. 175. 269 Schr¨dinger’s. 517 characteristic (matrices). 148 equation Airy. 289. 370 turbulence. 384 nonlinear centre. 170 porous medium. 186. 468 Lorenz. 352 dimensionless groups. 128 Legendre polynomials. irrotational. 413E forced Duﬃng. 95. 400. 104 eigensolutions. 369E eigenvalues. 79. 23 Airy. 184. 143. 470. 425 diﬀusion. 393. 274. 475 smooth. 79. focus. 352 CSTR (continuous ﬂow. 220 Euclidean norm. 143. 226 unstable spiral. 90E. 448 Euler. 68. 45. 179. 186 boundary conditions. 87 Friedrichs boundary conditions. 96 frequency modulation. 290. 223 eigenfunctions. 371. 103.INDEX control problem optimal. 229 unstable. 280 537 . 126 Fredholm alternative. 49 viscosity. 193 logistic. 227 stable spiral. 95 expansions. 182. 127 generalised coeﬃcients. 280. 273 reaction–diﬀusion. 170 Reynolds number. 230 hyperbolic. 186. 123 point source solution. 357 equilibrium point centre. 451. 294. 175. 489. 486. 482. 351 Laplace. 185. 461. 175. 108. 145 inverse. 456 stream function. 296 Bessel. 492E forced wave. 48 Bernoulli’s equation. 467 ﬂuid dynamics. 71. 415 cubic autocatalysis. 497 exchange of stabilities. 194 ﬂow past an aerofoil. 195 Navier-Stokes equations. 186 triple deck. 122E. 193 domain of attraction. 485. 276 Floquet theory. 167 Fourier integral. 151. 123 Dirac delta function. 119 o Tricomi’s. 124 Helmholtz. 378 stable. 43 sine. 44 Fourier–Bessel series. 379 diﬀerentiable. 162 wave. 140 higher dimensions. 509 Bessel functions. 224. 132 Fourier theorem. 433 controllable set. 148. 177 Volterra integral. 271 diﬀusion problem. 502 diﬀusion equation heat. 398 existence. 450. 350 Bernoulli. 124 Fourier’s law. 374 cubic crosscatalator. 135. 448. 103. 123. 39. stirred tank reactor). 499 eigenvectors. 189 complex potential. 496. 492E D’Alembert’s solution. 418 controllability matrix. 74. 133 Fourier series. 499 elastic membrane. 113 Gibbs’ phenomenon. 271 Dirichlet kernel. 274 circle theorem. 475 nonhyperbolic. 80 electrostatics. 195 inviscid. 183. 204 exponential integral. 151 Fourier. 133 linearity. 274 variables. 379. 147. 51. 109 functions φ(n). 130. 190 Cauchy–Riemann. 418 local. 186 ﬂow past a ﬂat plate. 214. 224. 335 ﬂow evolution operator. 493E modiﬁed Helmholtz. 227. 127 uniform convergence. 138 shifting. 357 diﬀeomorphism. 220 stable node. 456 velocity potential. 131 Fourier transform. 288. 352 convolution integral. 375 Ricatti’s. 180. 370. 75. focus. 71. 49 boundary layer. 372. 138 inversion formula. 401 saddle. 228 equilibrium solutions. 352 Navier–Stokes. 186 Kutta condition. 160. 133. 74 cosine. 128 Dirichlet problem. 126 Fourier–Legendre series. 125. 418 time-optimal. 150 Fourier’s equation. 420 convolution. 220 unstable node. 487.

512 intermittency. 379 homoclinic path. 293 complementary error. 243 Jordan’s lemma. 112 Hermitian. 225. 258 magniﬁcation. 135 unit function. 322. 167. 135 error. 153. 287. 29 hysteresis. 210. 270. 101. P1 (x). 230 Heaviside function. 247. 266 PDEs. 134 Heaviside. 74 Laguerre polynomials. 248. 314. 120 general solution. 135 generating functions Bessel functions. 35 Gibbs’ phenomenon. 262 inﬁnitesimal generators. 164 inverse. 357 cost (control).) Bessel (large argument). 102. 54E Laplace’s representation. 135 Heaviside. 260 one-parameter. 480. 143. 212. 160 ﬁrst shifting theorem. 39 Schl¨ﬂi’s representation. 357 inner product. 135 gamma function.538 INDEX functions (cont. 75. 270 rotation. 281 gauge. 377. 45. 40 a special values. 508 Jordan curve theorem. 135 multi-dimensional delta function. 281. 381 sequences of. 52. P3 (x) and P4 (x). 121E. 250. 108 associated. 373 integrating factors. 134 Green’s formula. 331. 397 initial value problem. 163. 300E. 521 leading order solutions. 71. 167 gamma. 113 Kuzmak’s method. 38 Rodrigues’ formula. 447. 134 Dirac delta. 128 good functions. 261 two-parameter. 303 Legendre equation n = 1. 251. 352 Bromwich inversion integral. 280 inversion. 256. 61. 54E. 472 Hopf bifurcation theorem. 153. 111 integral equations. 6 Legendre polynomials. 380. 58. 162 integral path. 219. 234. 181. 106 P0 (x). 403. 162 linearity. 122 Mel’nikov. 221 solutions. 294 gauge functions. 154 parabolic cylinder. 221 homoclinic point transverse. 474 homoclinic tangles. 523 Kronecker delta. 40 orthogonality. 158 of a derivative. 216E group extended. 302E negative deﬁnite. 418 Dirac delta. 54E generating function. 145 sign function. 111 Green’s function. 259 isomorphic. 333. 155. 33 associated Legendre polynomials. 103 heteroclinic path. 31. 509 sign function. 37 Legendre’s equation. 288. 32 . 281. 381 of exponential order. 41 recurrence relations. 261 group theoretic methods. 3. 122 Laplace transform. 172E Volterra. 415E hypergeometric equation. 142 Gronwall’s inequality. 52 Fourier-Legendre series. 135 Jacobian elliptic. 257. 138 Jacobian. 497 complex functions. 302 positive deﬁnite. 141 free space. 163 convolution. 477 Hartman-Grobman theorem. 368E Kronecker delta. 193 Laplace’s method. 512 Hamiltonian system. 479 modiﬁed Bessel. 98. 64 Legendre polynomials. 456 Inverse Fourier transform. 156 Hermite’s equation. P2 (x). 157 second shifting theorem. 156 Laplace’s equation. 346 Bessel of order zero. 43. 275 good. 155. 154 ODEs. 300E Laurent series. 31. 135. 101 Laguerre polynomials. 332 Lagrange’s identity. 152. 322. 35. 221 homeomorphism. 275 generalized function. 135 unit function. 270 group invariants.

452 Bernoulli shift. 496 linear independence. 475 Lommel’s functions. 402. 60. 447 nonautonomous. 432 normal form. 498 linearity Fourier transform. 257 539 . 8 regular point. 390 pitchfork bifurcation. 94 second law. 477 method of Frobenius. 9. 11–24. 23. 24 successive approximations. 222 complementary function. 3. 359. 255. 511 autonomous. 4 coupled systems. 367E method of reduction of order. 322. 9. 218. 249 Lipschitz condition. 507 General Rule I. 381 Lyapunov stable. 33 lorenz. 125 contour. 205. 19. 350 Method of variation of parameters. 505 manifold. 332. 269 formula. 459 e horseshoe. 269. 62 indicial equation. 221 Liouville’s theorem. 310 MATLAB functions airy. 318 ellipke. 493E Mel’nikov theory. 25 nonautonomous. 496 linear operator Hermitian.INDEX generalized. 257 magniﬁcation. 71 negatively invariant set. 379 maps. 257 rotation. 307. 50. 429. 463 matched asymptotic expansions. 222 nonsingular. 225 integrating factors. 426. 154 linearization of nonlinear systems. 513 separable variables. 402. 77. 511 singular point. 32 General Rule II. 454 Poincar´ return. 234 ODEs. 33. 333. 257 Lie series. 76. 50 eig. 154 Lie group. 33 legendre. 259 limit cycles. 31 Lerch’s theorem. 3. 63 matrix exponential. 3 second order equations (constant coeﬃcients). 103 self-adjacency. 210 one-parameter group horizontal translation. 314. 439 linear transformations. 169 erfc. 225 particular integral. 103 self-adjoint. 415E transcritical bifurcation. 475 logistic. 169 ezmesh. 467 H´non. 54E. 504 Mel’nikov function. 372. 61 General Rule III. 144. 489 ellipj. 42. 50 ode45. 5. 17 subplot. 444 mean value theorem. 96. 15. 489 plot. 285–290. 414E nullclines. 121 order. 237. 33. 487 Lyapunov function. 452 Smale horseshoe. 203. 31. 241 Neumann problem. 467 e shift. 60 legend. 4. 204 uniqueness. 54E. 98 modiﬁed Bessel functions. 63 ceil. 290. 6 method of stationary phase. 8. 158 existence. 34. 318 gamma. 484 maximum. 184 Newton’s ﬁrst law. 79 Lyapunov exponents. 529 polyval. 451 meshgrid. 351 formula. 231. 337 erf. 301E method of steepest descents. 428. 203. 138 Laplace transform. 465 tent. 424. 512 irregular singular point. 100 linear oscillator controlling. 242 linear dependence. 381 Maclaurin series. 485 spectrum. 221 limit point ω. 27. 80. 325. 80 besselj. 366E method of multiple scales. 257 vertical translation. 7. 212 lobe. 48 fzero. 10 ordinary points. 33 linspace. 204 integral paths. 13 method of matched asymptotic expansions.

351 canonical form. 381 exchange of. 408 set Cantor’s middle-third. 260 rotation. 445E product rule. 208. 506 Pr¨fer substitution. 431 controllable. 418 steering problem. 378. 113 particular integral. 124 similarity variable. 425 Smale-Birkhoﬀ theorem. 30 Sturm separation theorem. 389 stable manifold global. 97 Riemann integral. 237. 423 reachable. 103. 120 Rayleigh’s problem. 146 planetary motion. 175 parabolic. 379. 429. 409 triangle inequality. 66 Pockhammer symbol. 41 orthonormality. 227 Schr¨dinger’s equation. 60 Poincar´ return map. 98. 420. e 387 population dynamics. 507 convergence. 39. 71 Legendre polynomials. 506 radius of convergence. 389 Sturm comparison theorem. 475 solubility. 258 inverse. 497 Bessel functions. 53. 327 separatrices. 433. 398 Lyapunov. 107 regular endpoint. 270 hyperbolic. 418 Taylor series. 216E period doubling. 257 magniﬁcation. 109 target state. 425 quantum mechanics. 253E. 497 . 177 Cauchy problem. 507 ratio test. 454 periodic solutions. 143 group theoretic methods. 175 Fourier transforms. 323. 429 convex. 233. 39 saddle point. 274. 177. 119 o secular terms. 467. 375. 419. 274. 301E structurally unstable. 304 phase plane. 242 travelling wave solution. 462 stable node. 423 completely disconnected. 253E e Poincar´ projection. 146 stability asymptotic. 512 PDEs asymptotic methods. 234. 438. 221 perturbation problem regular. 111. 444E with friction. 109 singular endpoint. 5. 521 resonance. 381 structurally unstable. 245 plane polar coordinates. 245 e Poincar´–Bendixson theorem. 427 Stirling’s formula. 303 singular. 97 spherical harmonics. 436 transformations identity. 318 tightrope walker. 217. 492E e Poincar´ index. 380 local. 426. 464 closed. 331. 358 classiﬁcation. 54 spherical polar coordinates. 507 thermal ignition. 418 Time-optimal maximum principle. 175. 130 Rodrigues’ formula. 443E time-optimal control problem. 418 state vector. 270 Peano uniqueness theorem. 441 positively invariant set. 227. 180 characteristic variables. 168 reachable set. 497 Riemann–Lebesgue lemma. 115 u predator–prey system. 42. 422 open. 241 power series comparison test.540 INDEX orthogonality. 282. 423 simple harmonic motion. 175 separable solution. 464 connected. 252E. 118 hydrogen atom. 175 d’Alembert’s solution. 11 Sturm–Liouville problems. 427 with two controls. 227 stable spiral. focus. 83. 506 functions of two variables. 421 residue theorem. 257 inﬁnitesimal. 233 positioning problem. 378. 180 elliptic. 372. 241. 219 phase portrait. 421 strictly convex. 261 transversal. 229 state variables.

77. 380 local. 344. 83 WKB approximation. 300E wave equation.INDEX uniqueness. 283. 418 unstable manifold global. 90E. 227 unstable spiral. 369E Wronskian. 311. 210. 54E. 82 d’Alembert’s solution. focus. 228 van der Pol Oscillator. 8. 98 541 . 9. 496 Watson’s lemma. 329 Van Dyke’s matching principle. 9. 366E Vector space. 379. 28E. 462 unstable node. 292. 495 basis.