Martin Hermann, Masoud Saravi - Nonlinear Ordinary Differential Equations - Analytical Approximation and Numerical Methods-Springer (2016) PDF

Martin Hermann · Masoud Saravi
Nonlinear
Ordinary
Differential
Equations
Analytical Approximation and
Numerical Methods
Nonlinear Ordinary Differential Equations
Martin Hermann Masoud Saravi
•
Nonlinear Ordinary
Differential Equations
Analytical Approximation and Numerical
Methods
123
Martin Hermann Masoud Saravi
Department of Numerical Mathematics Department of Science
Friedrich Schiller University Shomal University
Jena Amol
Germany Iran
ISBN 978-81-322-2810-3 ISBN 978-81-322-2812-7 (eBook)

DOI 10.1007/978-81-322-2812-7
Library of Congress Control Number: 2016937388
© Springer India 2016

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made.
Printed on acid-free paper
This Springer imprint is published by Springer Nature

The registered company is Springer (India) Pvt. Ltd.
To our wives,
Gudrun and Mahnaz
and our children,
Alexander, Sina and Sayeh
Preface
Nonlinear ordinary differential equations (ODEs) play a fundamental role in sci-

entific modeling, technical and economic processes as well as in inner mathematical
questions. Examples include the complicated calculation of the trajectories of space
gliders and airplanes, forecast calculations, spread of AIDS, and the automation
of the parking maneuver of road vehicles on the basis of mechatronic systems.
However, the mathematical analysis of nonlinear ODEs is much more compli-
cated than the study of linear ones. The complexity of nonlinear problems can be
illustrated by the following comparison: if the amount of linear problems is set into
relation to the size of the earth then the amount of nonlinear problems corresponds
to the size of our universe. Therefore, the special type of the nonlinearity must be
taken into consideration if a nonlinear problem is theoretically and/or numerically
analyzed.
One of the greatest difficulties of nonlinear problems is that, in general, the
combination of known solutions is not again a solution. In linear problems, for
example, a family of linearly independent solutions can be used to construct the
general solution through the superposition principle.
For second- and higher order nonlinear ODEs (more generally speaking, systems
of equations) the solutions can seldom be represented in closed form, though
implicit solutions and solutions involving non-elementary integrals are encoun-
tered. It is therefore important for the applications to have analytical and numerical
techniques at hand that can be used to compute approximate solutions.
The first part of this book presents some analytical approximation methods
which have been approved in practice. They are particularly suitable for nonlinear
lower order ODEs.
High-dimensional systems of nonlinear ODEs are very often found in the
applications. The solutions of these systems can only be approximated by appro-
priate numerical methods. Therefore, in the second part of this book, we deal with
numerical methods for the solution of systems of first-order nonlinear
(parameter-depending) ODEs, which are subjected to two-point boundary
vii
viii Preface
conditions. It is precisely this focus on analytical as well as numerical methods that

makes this book particularly attractive.
The book is intended as a primary text for courses in the theory of ODEs and the
numerical treatment of ODEs for graduate students. It is assumed that the reader has
a basic knowledge of functional analysis, in particular analytical methods for
nonlinear operator equations, and numerical analysis. Moreover, the reader should
have substantial knowledge of mathematical techniques for solving initial and
boundary value problems in linear ODEs as they are represented in our book [63], A
First Course in Ordinary Differential Equations—Analytical and Numerical
Methods (Springer India, 2014). Physicists, chemists, biologists, computer scien-
tists, and engineers whose work involves the solution of nonlinear ODEs will also
find the book useful both as a reference and as a tool for self-studying.
In Chaps. 1–4, we have avoided the traditional definition–theorem–proof format.
Instead, the mathematical background is described by means of a variety of
problems, examples, and exercises ranging from the elementary to the challenging
problems. In Chap. 5, we deviate from this principle since the corresponding
mathematical proofs are constructive and can be used as the basis for the devel-
opment of appropriate numerical methods.
The book has been prepared within the scope of a German–Iranian research
project on mathematical methods for nonlinear ODEs, which was started in 2012.
We now outline the contents of the book. In Chap. 1, we present the basic
theoretical principles of nonlinear ODEs. The existence and uniqueness statements
for the solutions of the corresponding initial value problems (IVPs) are given. To
help the reader understand the structure and the methodology of nonlinear ODEs,
we present a brief review of elementary methods for finding analytical solutions of
selected problems. Some worked examples focusing on the analytical solution of
ODEs from the applications are used to demonstrate mathematical techniques,
which may be familiar to the reader. In Chap. 2, the analytical approximation
methods are introduced. Special attention is paid to the variational iteration method
and the Adomian decomposition method. For smooth nonlinear problems both
methods generate series, which converge rapidly to an exact solution. These
approximation methods are applied to some nonlinear ODEs, which play an
important role in the pure and applied mathematics. In Chap. 3, another four
analytical approximations methods are considered, namely the perturbation method,
the energy balance method, the Hamiltonian approach, and the homotopy pertur-
bation method. These methods have proved to be powerful tools for approximating
solutions of strongly nonlinear ODEs. They are applied in many areas of applied
physics and engineering. The accuracy of the approximated solutions is demon-
strated by a series of examples which are based on strongly nonlinear ODEs.
The numerical part of the book begins with Chap. 4. Here, numerical methods
for nonlinear boundary value problems (BVPs) of systems of first-order ODEs are
studied. Starting from the shooting methods for linear BVPs [63], we present the
simple shooting method, the method of complementary functions, the multiple
shooting method, and a version of the stabilized march method for nonlinear BVPs.
A MATLAB code for the multiple shooting method is provided such that the reader
Preface ix
can perform his own numerical experiments. Chapter 5 is devoted to nonlinear

parametrized BVPs since in the applications the majority of nonlinear problems
depend on external parameters. The first step in the treatment of such BVPs is their
formulation as a nonlinear operator equation Tðy; λÞ ¼ 0 in Banach spaces, and to
apply the Riesz–Schauder theory. Under the assumption that Ty ðz0 Þ is a Fredholm
operator with index zero, where z0 ðy0 ; λ0 Þ is a solution of the operator equation,
the concept of isolated solutions and singular solutions (turning points and bifur-
cation points) is discussed. Appropriate extended systems are presented to compute
the singular points with standard techniques. In each case, the extended operator
equation is transformed back into an extended nonlinear BVP. Since the operator
equation is bad conditioned in the neighborhood of singular points, we use the
technique of transformed systems to determine solutions near to singularities. In
addition to the one-parameter problem, we study the operator equation under the
influence of initial imperfections, which can be described by a second (perturbation)
parameter. Feasible path-following methods are required for the generation of
bifurcation diagrams of parametrized BVPs. We present some methods that follow
simple solution curves. The chapter is concluded with the description of two
real-life problems.
At this point it is our pleasure to thank all those colleagues who have particularly
helped us with the preparation of this book. Our first thanks go to Dr. Dieter Kaiser,
without whose high motivation, deep MATLAB knowledge, and intensive com-
putational assistance this book could never have appeared. Moreover, our thanks
also go to Dr. Andreas Merker who read through earlier drafts or sections of the
manuscript and made important suggestions or improvements. We gratefully
acknowledge the financial support provided by the Ernst Abbe Foundation Jena for
our German–Iranian joint work.
It has been a pleasure working with the Springer India publication staff, in
particular with Shamim Ahmad and Ms. Shruti Raj.
Finally, the authors wish to emphasize that any helpful suggestion or comment
for improving this book will be warmly appreciated and can be sent by e-mail to:
[email protected] and [email protected].
Martin Hermann
Masoud Saravi
Contents
1 A Brief Review of Elementary Analytical Methods for Solving

Nonlinear ODEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Analytical Solution of First-Order Nonlinear ODEs . . . . . . . . . . . 1
1.3 High-Degree First-Order ODEs . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.4 Analytical Solution of Nonlinear ODEs by Reducing
the Order . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 16
1.5 Transformations of Nonlinear ODEs . . . . . . . . . . . . . . . . . .... 21
1.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 29
2 Analytical Approximation Methods . . . . . . . . . . . . . . . . . . . . . . . . 33
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2 The Variational Iteration Method . . . . . . . . . . . . . . . . . . . . . . . . 33
2.3 Application of the Variational Iteration Method . . . . . . . . . . . . . . 38
2.4 The Adomian Decomposition Method . . . . . . . . . . . . . . . . . . . . 44
2.5 Application of the Adomian Decomposition Method . . . . . . . . . . 48
2.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3 Further Analytical Approximation Methods and Some
Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.1 Perturbation Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.1.1 Theoretical Background . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.1.2 Application of the Perturbation Method . . . . . . . . . . . . . . 63
3.2 Energy Balance Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.2.2 Application of the Energy Balance Method . . . . . . . . . . . 74
3.3 Hamiltonian Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
3.3.2 Application of the Hamiltonian Approach. . . . . . . . . . . . . 88
xi
xii Contents
3.4 Homotopy Analysis Method . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

3.4.2 Application of the Homotopy Analysis Method. . . . . . . . . 107
3.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
4 Nonlinear Two-Point Boundary Value Problems . . . . . . . . . . . . . . . 121
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
4.2 Simple Shooting Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.3 Method of Complementary Functions. . . . . . . . . . . . . . . . . . . . . 131
4.4 Multiple Shooting Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
4.5 Nonlinear Stabilized March Method . . . . . . . . . . . . . . . . . . . . . . 146
4.6 MATLAB PROGRAMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
4.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
5 Numerical Treatment of Parametrized Two-Point Boundary
Value Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
5.2 Two-Point BVPs and Operator Equations . . . . . . . . . . . . . . . . . . 169
5.3 Analytical and Numerical Treatment of Limit Points . . . . . . . . . . 172
5.3.1 Simple Solution Curves . . . . . . . . . . . . . . . . . . . . . . . . . 172
5.3.2 Extension Techniques for Simple Turning Points . . . . . . . 179
5.3.3 An Extension Technique for Double Turning Points . . . . . 186
5.3.4 Determination of Solutions in the Neighborhood
of a Simple Turning Point . . . . . . . . . . . . . . . . . . . . . . . 190
5.4 Analytical and Numerical Treatment of Primary Simple
Bifurcation Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
5.4.1 Bifurcation Points, Primary and Secondary Bifurcation
Phenomena. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
5.4.2 Analysis of Primary Simple Bifurcation Points . . . . . . . . . 195
5.4.3 An Extension Technique for Primary Simple Bifurcation
Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
of a Primary Simple Bifurcation Point . . . . . . . . . . . . . . . 204
5.5 Analytical and Numerical Treatment of Secondary Simple
Bifurcation Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
5.5.1 Analysis of Secondary Simple Bifurcation Points . . . . . . . 232
5.5.2 Extension Techniques for Secondary Simple
Bifurcation Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
of a Secondary Simple Bifurcation Point . . . . . . . . . . . . . 244
5.6 Perturbed Bifurcation Problems . . . . . . . . . . . . . . . . . . . . . . . . . 247
5.6.1 Nondegenerate Initial Imperfections . . . . . . . . . . . . . . . . . 247
5.6.2 Nonisolated Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . 252
5.6.3 Solution Curves Through Nonisolated Solutions . . . . . . . . 267
Contents xiii
5.7 Path-Following Methods for Simple Solution Curves . . . . . . . . . . 269

5.7.1 Tangent Predictor Methods. . . . . . . . . . . . . . . . . . . . . . . 269
5.7.2 Arclength Continuation . . . . . . . . . . . . . . . . . . . . . . . . . 273
5.7.3 Local Parametrization . . . . . . . . . . . . . . . . . . . . . . . . . . 275
5.7.4 Detection of Singular Points . . . . . . . . . . . . . . . . . . . . . . 280
5.8 Parametrized Nonlinear BVPs from the Applications . . . . . . . . . . 288
5.8.1 Buckling of Thin-Walled Spherical Shells . . . . . . . . . . . . 288
5.8.2 A Bipedal Spring-Mass Model . . . . . . . . . . . . . . . . . . . . 292
5.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
About the Authors
Martin Hermann is professor of numerical mathematics at the Friedrich Schiller

University (FSU) Jena, Germany. His activities and research interests include sci-
entific computing and numerical analysis of nonlinear parameter-dependent ODEs.
He is also the founder of the Interdisciplinary Centre for Scientific Computing
(1999), where scientists of different faculties at the FSU Jena work together in the
fields of applied mathematics, computer science, and applications. Since 2003, he
has headed an international collaborative project with the Institute of Mathematics
at the National Academy of Sciences Kiev (Ukraine), studying, for example, the
sloshing of liquids in tanks. Since 2003, Dr. Hermann has been a curator at the
Collegium Europaeum Jenense of the FSU Jena (CEJ) and the first chairman of the
Friends of the CEJ. In addition to his professional activities, he volunteers in
various organizations and associations. In German-speaking countries, his books
Numerical Mathematics and Numerical Treatment of ODEs: Initial and Boundary
Value Problems count among the standard works on numerical analysis. He has
also published over 70 articles in refereed journals.
Masoud Saravi is professor of mathematics at the Shomal University, Iran. His
research interests include the numerical solution of ODEs, partial differential
equations (PDEs), integral equations, differential algebraic equations (DAE), and
spectral methods. In addition to publishing several papers with his German col-
leagues, Dr. Saravi has published more than 15 successful titles on mathematics.
The immense popularity of his books is a reflection of his more than 20 years of
educational experience, and a result of his accessible writing style, as well as a
broad coverage of well laid-out and easy-to-follow subjects. He has recently retired
from Azad University and currently cooperates with Shomal University and is
working together with the Numerical Analysis Group and the Faculty of
Mathematics and Computer Sciences of FSU Jena (Germany). He started off his
academic studies at the UK’s Dudley Technical College before receiving his first
xv
xvi About the Authors
degree in mathematics and statistics from the Polytechnic of North London, and his
advanced degree in numerical analysis from Brunel University. After obtaining his
M.Phil. in Applied Mathematics from Iran’s Amir Kabir University, he completed
his Ph.D. in Numerical Analysis on solutions of ODEs and DAEs by using spectral
methods at the UK’s Open University.
Chapter 1
A Brief Review of Elementary Analytical
Methods for Solving Nonlinear ODEs
1.1 Introduction
Nonlinear ordinary differential equations (ODEs) are encountered in various fields

of mathematics, physics, mechanics, chemistry, biology, economics, and numerous
applications. Exact solutions to those equations play an important role in the proper
understanding of qualitative features of many phenomena and processes in various
areas of natural science. However, even if there exists a solution, only for a few
ODEs it is possible to determine this exact solution in closed form. Although some
methods have been proposed in the literature to determine analytical (closed form)
solutions to special ODEs, there is not a general analytical approach to evaluate
analytical solutions. Surveys of the literature with numerous references, and useful
bibliographies, have been given in [5, 67, 89, 106], and more recently in [50].
In this chapter, we present the basic theoretical principles of nonlinear ODEs.
The existence and uniqueness statements for the solutions of the corresponding ini-
tial value problems (IVPs) are given. To enable the reader to understand the structure
and the methodology of nonlinear ODEs, we present a brief review of elementary
methods for finding analytical solutions of selected problems. Some worked exam-
ples focusing on the analytical solution of ODEs from the applications are used to
demonstrate mathematical techniques, which may be familiar to the reader.
1.2 Analytical Solution of First-Order Nonlinear ODEs
The general form of a first-order ODE is
f (x, y, y ) = 0. (1.1)
Since (1.1) is order one, we may guess that the solution (if it exists) contains a
constant c. That is, the family of solution will be of the form
© Springer India 2016 1

M. Hermann and M. Saravi, Nonlinear Ordinary Differential Equations,
DOI 10.1007/978-81-322-2812-7_1
2 1 A Brief Review of Elementary Analytical Methods …
F(x, y, c) = 0. (1.2)
Moreover, if (1.1) can be written as y = g(x, y), then we search a family of

solutions y = G(x, c). We mention, that the constant c can be specified by an
additional condition.
As we know, a linear equation has only one general solution whereas a nonlinear
equation may also have a singular solution. Consider the IVP
y = f (x, y), y(x0 ) = y0 . (1.3)
This equation may have no solution or one or more than one solution. For example,
the IVP
(y )2 + y 4 = 0, y(0) = 1,
has no real solution. The IVP
y − 2y = 0, y(0) = 1,
has one solution y(x) = e2x and the IVP

√
y = y, y(0) = 0,
1
has more than one solution, namely, y(x) ≡ 0 and y(x) = x 2 . This leads to the
4
following two fundamental questions:
1. Does there exist a solution of the IVP (1.3)?
2. If a solution exists, is this solution unique?
The answers to these questions are provided by the following existence and unique-
ness theorems.
Theorem 1.1 (Peano’s existence theorem)
Let R(a, b) be a rectangle in the x-y plane, and assume that the point (x 0 , y0 ) is
inside it such that
R(a, b) ≡ {(x, y) : |x − x 0 | < a, |y − y0 | < b}.
If f (x, y) is continuous and | f (x, y)| < M at all points (x, y) ∈ R, then the
IVP (1.3) has at least one solution y(x), which is defined for all x in the interval
|x − x0 | < c, where c = min{a, b/M}.
Proof See, e.g., the monograph [46].

1.2 Analytical Solution of First-Order Nonlinear ODEs 3
Theorem 1.2 (Uniqueness theorem of Picard and Lindelöf)

Let the assumptions of Theorem 1.1 be satisfied. If in addition ∂∂ yf (x, y) is continuous
and bounded for all points (x, y) ∈ R, then the IVP (1.3) has a unique solution y(x),
which is defined for all x in the interval |x − x0 | < c.
Moreover, we have
Theorem 1.3 (Picard iteration)
Let the assumptions of Theorem 1.2 be satisfied. Then, the IVP (1.3) is equivalent to
Picard’s iteration method
x
yn (x) = y(x0 ) + f (t, yn−1 (t))dt, n = 1, 2, 3, . . . (1.4)
x0
That is, the sequence y0 , y1 , . . . converges to y(x).
We note that in this chapter our intention is not to solve a nonlinear first-order
ODE by Picard’s iteration method, but to present Theorem 1.3 just for discussing the
existence and uniqueness theorems.
Now, as before, let c ≡ min{a, b/M} and consider the interval I : |x − x0 | ≤ c
and a smaller rectangle
T ≡ {(x, y) : |x − x0 | ≤ c, |y − y0 | ≤ b}.
By the mean value theorem of differential calculus, one can prove that if ∂ f /∂ y is
continuous in the smaller rectangle T ⊂ R(a, b), then there exists a positive constant
K such that
| f (x, y1 ) − f (x, y2 )| ≤ K |y1 − y2 |, (x, y1 ) , (x, y2 ) ∈ T. (1.5)
Definition 1.4 If the function f (x, y) satisfies the inequality (1.5) for all (x, y1 ),
(x, y2 ) ∈ T , it is said that f (x, y) satisfies a uniform Lipschitz condition w.r.t. y in
T and K is called Lipschitz constant.
Remark 1.5 We can conclude that if ∂ f /∂ y is continuous in T , then f (x, y) satisfies

a uniform Lipschitz condition. On the contrary, it may f (x, y) satisfy a uniform
Lipschitz condition, but ∂ f /∂ y is not continuous in T . For example, f (x, y) = |x|
satisfies a Lipschitz condition in a rectangle, which contains (0, 0), but ∂ f /∂ y is not
continuous.
The following result can be easily proved.

Proposition 1.6 If on a closed bounded region R(a, b) the function f (x, y) is a
continuous function of its arguments and satisfies a Lipschitz condition, then there
exists a unique solution y(x) of (1.3), which is defined on an interval I . Furthermore,
the sequence of functions {yn (x)}∞1 defined by (1.4) converges to y(x) on I .
As we pointed out it may be difficult to find closed form solutions whenever we

deal with nonlinear problems. We are going to introduce some analytical methods,
known as elementary methods, which are based on integration. But, first of all, let
us write (1.1) in the form
M(x, y)d x + N (x, y)dy = 0. (1.6)
This equation can be solved in the following cases.

(i) Separable equations
If (1.6) can be written in the form M(x)d x + N (y)dy = 0, we say that the variables
are separable, and the solution can be obtained by integrating both sides with respect
to x, i.e.

dy
M(x)d x + N (y(x)) · d x = c.
dx
Then, for simplicity one can write

M(x)d x + N (y)dy = c.
Example 1.7 Solve the IVP (y + 1)e x y + x y 2 = 0, y(0) = 1.
Solution. The given equation can be written as

1
1+ dy = −xe−x d x.
y
Integration yields
y + ln(y) = (x + 1)e−x + c.
Using y(0) = 1, we obtain c = 0. Hence y + ln(y) = (x + 1)e−x .

(ii) Homogeneous equations
We call f (x, y) homogeneous of degree k if f (t x, t y) = t k f (x, y), for some con-
stant k = 0. If in (1.6), M(x, y), and N (x, y) have the same degree, or if f (x, y) is
of degree zero, then the substitution z = y/x will reduce the given equation into a
separable form. Let y = x z, then y = z + x z . Substituting this into the equation
gives z + x z = f (z). After separating the variables, we obtain
z 1
= .
f (z) − z x
y(2x 2 − x y + y 2 )
Example 1.8 Solve y = .
x 2 (2x − y)
Solution. Since the degree of
y(2x 2 − x y + y 2 )
f (x, y) =
x 2 (2x − y)
is zero, the substitution y = x z and y = z + x z leads to

2 1 dx
− 2 dz = .
z3 z x
Thus,
1 1
− 2 = ln(|x|) + c.
z z
Since z = y/x, we have
x x2
− 2 = ln(|x|) + c.
y y

(iii) Exact equations
The total differential of the function z(x, y) is given by
dz = z x d x + z y dy. (1.7)
The next definition is based on the connection between (1.6) and (1.7). We say that
(1.6) is exact, if M(x, y)d x + N (x, y)dy is the total differential of z(x, y). Then,
the ODE can be written as dz = 0. Now, we have the following result.
Theorem 1.9 If M(x, y) and N (x, y) have continuous first partial derivatives over
some domain D, then (1.6) is exact if and only if M y = N x .
Proof See e.g., [46].
In the case of an exact equation, it holds z x = M and z y = N . Define z(x, y) by

z(x, y) ≡ M(x, y)d x + g(y).
x
From this we have

∂
zy = M(x, y)d x + g (y),
∂y x
and this is equal to N (x, y). Therefore

∂
g (y) = N (x, y) − M(x, y)d x.
∂y x
We can determine g(y) if g (y) is independent from x, and this is true whenever
the derivative of the right-hand side w.r.t. x is zero. Performing this differentiation
leads to

∂ ∂ ∂N ∂ ∂
N (x, y) − M(x, y)d x = − M(x, y)d x
∂x ∂y x ∂x ∂x ∂ y x
∂N ∂M
= − .
∂x ∂y
For an exact equation, the last expression is zero.

Example 1.10 Find the family of solutions of the differential equation

cos(2x) − 2y sin(2x) − 2x y 3 d x + cos(2x) − 3x 2 y 2 dy = 0.
Solution. The variables are not separable. We have
M(x, y) = cos(2x) − 2y sin(2x) − 2x y 3 , N (x, y) = cos(2x) − 3x 2 y 2 .
Then M y = N x = −2 sin(2x) − 6x y 2 . Hence the equation is exact. We have z x =

M = cos(2x) − 2y sin(2x) − 2x y 3 . Integration w.r.t. x gives
sin(2x)
z(x, y) = + y cos(2x) − x 2 y 3 + g(y).
2
Differentiation w.r.t. y leads to
z y = cos(2x) − 3x 2 y 2 + g (y).
Now, since z y = N (x, y) = cos(2x) − 3x 2 y 2 , it is
g (y) = 0, i.e. g(y) = c.
Therefore, the family of solutions will be
sin(2x)
+ y cos(2x) − x 2 y 3 = c.
2

(iv) Non-exact equations

Let us come back to (1.6) and assume that this equation is not exact. Thus, it holds
M y = N x . To determine a solution in that case can be very difficult or impossible,
even if there exists a solution. But, sometimes Eq. (1.6) can be transformed into
an exact equation by multiplying it with a function F(x, y) ≡ 0, which is called
the integrating factor that is usually not unique. If such a function exists then the
modified ODE
F(x, y)M(x, y) + F(x, y)N (x, y) = 0
will be exact. Therefore (F M) y = (F N )x , and we have
M Fy + F M y = N Fx + F N x .
We can rewrite this formula as

N Fx − M Fy
= M y − Nx . (1.8)
F
This is a first-order partial differential equation that may have more than one solution.
But to compute such a solution will be more difficult than to solve the original
Eq. (1.6). Fortunately, if we restrict the integrating factor to be a function of the
single variable x or y, then we can find F. Suppose F = F(x), then Fy = 0, and
(1.8) is reduced to
dF M y − Nx
= .
F N
It is not difficult to show that

M y − Nx
F = exp dx .
N
On the other hand, if we have F = F(y), then Fx = 0, and it holds

M y − Nx
F = exp − dy .
M
Thus, we can say that F = F(x), when (M y − N x )/N is a function of x, and

F = F(y), when (M y − N x )/M is a function of y.
Example 1.11 Determine a particular solution of the ODE
2 tan(y) sin(2x)d x + (sin(y) − cos(2x))dy = 0.
Solution. We have M y = 2(1 + tan2 (y)) sin(2x) and N x = 2 sin(2x). Hence, the
equation is not exact, but (M y − N x )/M = tan(y) = g(y). Thus

F(y) = exp − g(y)dy = exp − tan(y)dy
= exp (ln(cos(y)) = cos(y).
Multiplying the differential equation by cos(y), we get
2 sin(y) sin(2x)d x + cos(y)(sin(y) − cos(2x))dy = 0.
This equation is exact, and it can be shown that its solution is
2 sin(y) cos(2x) + cos2 (y) = c.
Example 1.12 Determine a particular solution of the IVP

π
cos(x) sin(x) − sin(y) d x + sin(x) cos(y)dy = 0, y = 0.

2
Solution. We have M y = − cos(x) cos(y) and N x = cos(x) cos(y). Hence, the
equation is not exact, but
M y − Nx
= −2 cot(x) = g(x).
M
Thus
F(x) = exp − g(x)d x = exp −2 cot(x)d x
1
= exp − 2 ln(sin(x)) = .
sin2 (x)
Multiplying the differential equation by 1/ sin2 (x), we get
cos(x) cos(x)
sin(x) − sin(y) d x + dy = 0.
sin (x)
2 sin(x)
This equation is exact. The corresponding family of solutions is
sin(y)
ln(sin(x)) + = c.
sin(x)
To determine the particular solution, which satisfies the given initial condition, we
substitute x = π/2 and y = 0 into this family of solutions, obtaining c = 0. Hence,
the solution of the IVP is
sin(y)
ln sin(x) + = 0.
sin(x)
(v) Reducible equations

Some nonlinear ODEs can be reduced to linear or separable equations by forming
the dependent variable z = g(y). The most common examples of such equations are
the Bernoulli and the Riccati equation. In the rest of this section, we will consider
these two equations. But first, we remember that the general form of a linear ODE is
y + p(x)y = q(x). (1.9)
When q(x) = 0, this equation is reduced to a separable equation. Assume q(x) = 0

and consider (1.9). This equation is not exact but it can easily be shown that it has
the integrating factor
F(x) = exp p(x)d x .
Multiplying the given equation by this factor, we get

exp p(x)d x (y + p(x)y) = exp p(x)d x q(x).
This equation is exact. Integrating both sides yields

exp p(x)d x y = exp p(x)d x q(x)d x + c.
Hence,

y = exp − p(x)d x exp p(x)d x q(x)d x
(1.10)
+ c exp − p(x)d x ,
which is the general solution of (1.9).

(a) Bernoulli equation
The general form of a Bernoulli equation is
y + p(x)y = q(x)y n , (1.11)
where n is a real number. Applications of this equation can be seen in the study of
population dynamics and in hydrodynamic stability. When n = 0 or n = 1, this
equation is reduced to a linear equation. Therefore, we assume n = 0 or n = 1.
If we set z ≡ y 1−n , then z = (1 − n)y y 2−n . Thus,
1−n
y = z.
y 2−n
Substituting these expressions into (1.11), we obtain the following linear ODE for
the function z
z + (1 − n) p(x)z = (1 − n)q(x). (1.12)
Example 1.13 A mass m is accelerated by a time-varying force exp(−αt)v 3 , where

v is its velocity. There is also a resistive force βv, where β is a constant, owing to its
motion through the air. The ODE modeling the motion of the mass is therefore
mv = exp(−αt)v 3 − βv.
Find an expression for the velocity v of the mass as a function of time. Given is the
initial velocity v0 .
Solution. We write the equation in the form
β 1
v + v = e−αt v 3 .
m m
This is a Bernoulli equation with n = 3. Thus, z = v 1−3 = v −2 , and the derivative

is v = −z /(2v −3 ). Substituting these expressions into the Bernoulli equation,we
obtain
2β 2
z − z = − e−αt .
m m
It is not difficult to show that the general solution of this equation is

2e−αt 2β
z= + c exp t .
2β m
m(α + )
m
Since z = v −2 , we have
⎡ ⎤−2

⎢ 2e−αt 2β ⎥
v=⎢
⎣ + c exp t ⎥ .
2β m ⎦
m α+
m
Now, the initial condition v(0) = v0 must be used to find the particular solution.
Example 1.14 For the two-dimensional projectile motion under the influence of
gravity and air resistance, using the assumption that the air resistance is proportional
to the square of the velocity, the model equation relating the speed s(t) of the projectile
to the angle of inclination θ(t) is
ds
− tan(θ)s = b sec(θ)s 3 ,
dθ
where b is a constant, which measures the amount of air resistance. Solve this
equation.
Solution. The given equation is a Bernoulli equation with n = 3. We have z =
s 1−3 = s −2 , and the derivative is z = −2s s −3 . Thus s = −z /(2s −3 ). Substituting
these expressions into the Bernoulli equation, we get
dz
+ 2 tan(θ)z = −2b sec(θ).
dθ
This is a linear ODE and one can show that its general solution is

c − b sec(θ) tan(θ) + ln(sec(θ) + tan(θ))
z= .
sec2 (θ)
Since z = s −2 , it follows
sec(θ)
s= .
c − b sec(θ) tan(θ) + ln(sec(θ) + tan(θ))

(b) Riccati equation
The general form of a Riccati equation is
y = p(x) + q(x)y + r (x)y 2 . (1.13)
This equation appears in numerous physical applications and is closely related to

linear homogeneous second-order ODEs. If we substitute
u (x)
y=−
q(x)u(x)
into (1.13), we obtain

r (x)
u (x) − + q(x) u (x) + p(x)r (x)u(x) = 0. (1.14)
r (x)
Obviously, if in (1.13) it holds p(x) ≡ 0, then we have a Bernoulli equation. There-

fore, we assume p(x) = 0. In that case, Eq. (1.13) cannot be solved by elementary
methods. But, if a particular solution y1 is known, then the solution of this equation
can be obtained by the substitution y = y1 + 1/z, where z is the general solution of
the ODE
z + (q + 2r y1 )z = −r.
Example 1.15 Show that y = cos(x) is a particular solution of the ODE

1 − sin(x) cos(x) y − y + cos(x)y 2 = − sin(x),
then solve this equation.
Solution. We have y = cos(x), and therefore y = − sin(x). Substituting these

functions into the equation leads to
(1 − sin(x) cos(x))(− sin(x)) − cos(x) + cos(x) cos2 (x)

= − sin(x) + sin2 (x) cos(x) − cos(x) + cos(x) 1 − sin2 (x) = − sin(x).
Thus, y = cos(x) is a particular solution of the ODE. We set y ≡ cos(x) + 1/z.

Then, y = − sin(x) − z /z 2 . Substituting these expressions into the ODE, we obtain

z
1 − sin(x) cos(x) − sin(x) −
z2
2
1 1
− cos(x) + + cos(x) cos(x) + = − sin(x).
z z
By simplifying this expression, we get

1 − sin(x) cos(x) z + 1 − 2 cos2 (x) z = cos(x).

Since 1 − sin(x) cos(x) = 1 − 2 cos2 (x), the integration of the above equation
yields
sin(x) + c
z= .
1 − sin(x) cos(x)
We also have y = cos(x) + 1/z, i.e. z = 1/(y − cos(x)). Therefore,
c cos(x) − 1
y= .
c − sin(x)
Remark 1.16 Another choice for y(x) is y(x) = y1 (x) + z(x), where z(x) is the
family of solutions of the equation
z + ( p − 2qy1 ) = qz 2 .
Example 1.17 Solve the ODE y + 2x y = y 2 + x 2 + 1. A corresponding particular

solution is y1 (x) = x.
Solution. Let y = x + z. Then, y = 1 + z . Substituting these expressions into the

given equation, we get z = z 2 . It is not difficult to show that the family of solutions
of this equation is z = 1/(c − x). That is, y = x + 1/(c − x).
1.3 High-Degree First-Order ODEs
So far, we have considered first-order ODEs with degree one. In this section, we
briefly discuss equations with a higher degree.
The general form of such an equation is
(y )n + p1 (x, y)(y )n−1 + · · · + pn−1 (x, y)y + pn (x, y) = 0. (1.15)
Unfortunately, solving (1.15) is extremely difficult, and in the majority of cases

impossible, even if a solution exists. However, there are some special cases of (1.15),
which can be solved analytically. Here, we will discuss two of these special cases.
(i) Equations solvable for y
Suppose (1.15) can be factorized in the form
[y − f 1 (x, y)][y − f 2 (x, y)] · · · [y − f n (x, y)] = 0.
We set each factor equal to zero, i.e.,
y − f i (x, y) = 0, i = 1, 2, . . . , n,
and determine the solutions
φi (x, y, c) = 0, i = 1, 2, . . . , n,
of these factors. Now, the family of solutions of (1.15) can be written as
φ1 (x, y, c)φ2 (x, y, c) · · · φn (x, y, c) = 0.
Example 1.18 Solve the ODE x 2 (y )2 + x y − y 2 − y = 0.

Solution. In terms of y , this equation can be compared with an algebraic equation
of degree two. Hence, we can write
(x y − y)(x y + y + 1) = 0.
From first factor we obtain y − c1 x = 0 and from the second one, we get x(y + 1) −
c2 = 0. Thus
(y − c1 x) x(y + 1) − c2 = 0
is the solution.
(ii) Equations solvable for x or y

Suppose we can write (1.15) in the form
y = f (x, y ). (1.16)
Differentiating w.r.t. x gives
y = F(x, p, p ),
where y = p. This equation has the two variables x and p. The corresponding
solution can be written in the form
φ(x, y , c) = 0. (1.17)
The elimination of y from (1.16) and (1.17) gives the desired solution. A similar
operation can be done when the independent variable is missing.
The methods presented in (i) and (ii) can be helpful for solving (1.15). However,
when n ≥ 3, we encounter difficulties to factorize (1.15) or to eliminate y . To
recognize this, let us consider the equation
y = 3x y − (y )3 .
Differentiation w.r.t. x yields
y = 3y + 3x y − 3y (y )2 .
It follows
3(x − p 2 ) p = −2 p.
If we rewrite this equation in the form
2 p d x + 3(x − p 2 ) dp = 0,
where p = y , then it is not difficult to show that the equation has the integrating
√
factor F = p. Multiplying this equation by F and integrating term by term, we
obtain
√
p p(x − p 2 ) = c, i.e. y y (x − (y )2 ) = c.
The next step is to eliminate y from this equation and the given equation. Obviously,
this elimination is nearly impossible.
We close this section by considering two special nonlinear ODEs.
1.3 High-Degree First-Order ODEs 15
(iii) Clairaut equation

The general form of a Clairaut equation is
y = x y + f (y ). (1.18)
Differentiating w.r.t. x gives
dp
x + f ( p) = 0,
dx
where y = p. We can have dp/d x = 0 or x + f ( p) = 0. If dp/d x = 0, then

p = y = c. Now, putting y = c in (1.18), we obtain
y = cx + f (c).
This is the family of solutions of the Clairaut equation. One can eliminate y from
(1.18) and x + f (y ) = 0. It results in a singular solution because it does not
contain c.
Example 1.19 Solve the ODE y = x y − 2(y )3 , and find its singular solution.
Solution. Obviously, this is a Clairaut equation. Hence, its family of solutions is

y = cx − 2c3 . We have x + f (y ) = x − 6(y )2 = 0. Now, if we eliminate y from
this equation and the given equation, we obtain
x
23
y=4 ,
6
which is a singular solution.
(iv) Lagrange equation
A more general form of a Clairaut equation is given by
y = f (y ) + xg(y ), (1.19)
and known as Lagrange equation. If we differentiate it w.r.t. x and set y ≡ p, then

this equation can be written in the form
dx g ( p) f ( p)
− x= .
dp p − g( p) p − g( p)
Obviously, this is a linear equation and can be solved by the formula given in (1.10).
Example 1.20 Solve the equation y = x(y )2 − (y )2 .

Solution. This is a Lagrange equation with f (y ) = −g(y ) = (y )2 . Therefore,
dx 2p 2p
− x =− .
dp p − p2 p − p2
That is,
dx 2 2
+ x =− .
dp p−1 p−1
The general solution of this equation can be found by Eq. (1.14). We obtain
c c
x =1+ =1+ .
( p − 1)2 (y − 1)2
By elimination of y from this equation and the given equation, we get

c
y = x +c+2 .
x −1
What makes the solution of the Lagrange equation tedious is the elimination of y .
In general, this difficulty is the same difficulty as it exists in the case (iii).
1.4 Analytical Solution of Nonlinear ODEs

by Reducing the Order
We can simplify a nonlinear ODE to a first-order equation by reducing its order, in

particular when one of its dependent or independent variables is missing. We consider
the second-order ODE in general form
f (x, y, y , y ) = 0, (1.20)
and deal with two special cases.

1. Suppose (1.20) does not contain y. Hence, we have
f (x, y , y ) = 0. (1.21)
The substitution y = p can be used to transform this equation into a first-order

ODE. Since y = p , we get
f (x, p, p ) = 0.
2. Suppose (1.20) does not contain x, i.e., it is an autonomous ODE. So we have
f (y, y , y ) = 0. (1.22)
1.4 Analytical Solution of Nonlinear ODEs by Reducing the Order 17
Let y = p, then y = p(dp/dy). Therefore, (1.22) can be written as
f (y, p, p ) = 0.
This is an equation of first-order and can be solved for p.
Example 1.21 A classical problem of mechanical engineering is the hanging cable

problem. The shape of a hanging cable of uniform density is governed by the ODE
2
d2 y d dy
2
= 1+ ,
dx T dx
where d and T are the constant density and the tension in the cable at its lowest point,
respectively. Solve this ODE.
Solution. Set y ≡ p, then y = p . Hence, the equation becomes

d
p = 1 + p2 .
T
Obviously, the variables are separable, and we can write
dp d
= d x.
1+ p2 T
Integrating both sides, we obtain

d d
sinh ( p) = x + c1 , i.e., p = y = sinh
−1
x + c1 .
T T
One more integration leads to

T d
y= cosh x + c1 + c2 .
d T
We mention that, for this example, we could also use the method presented in Case 2.

Example 1.22 The ODE
2
d2 y dy dy
= +1
dx2 dx dx
occurs in the analysis of satellite orbits. Find all solutions of this equation.
Solution. Set y ≡ p, then y = p(dp/dy). Substituting these expressions into the

ODE, we get
dp 2
p = p + 1 p.
dy
Obviously, p = 0 or dp/dy = p2 + 1. From the first possibility, we get y = 0, i.e.,

y = c. For the second possibility, we can write dp/( p 2 + 1) = dy. Obviously, the
variables are separable. Therefore, we integrate both sides, and obtain tan−1 ( p) =
y + c1 , i.e., y = tan(y + c1 ). We can write
y cot (y + c1 ) = d x.
Integrating this equation yields

ln sin (y + c1 ) = x + ln(c2 ), i.e., y = sin−1 ec2 x − c1 .
Example 1.23 In the theory of the plane jet, which is studied in the hydrodynamics
[21], we are encountered with the ODE
ay + yy + (y )2 = 0, a > 0,
which is subjected to the conditions y(0) = y (0) = 0 and y (x) = 0, when

x → +∞. Determine the solution of this problem.
Solution. Let us write

d
(ay + yy ) = 0.
dx
By integration we get
ay + yy = c1 .
Using the given initial conditions y(0) = y (0) = 0, we obtain c1 = 0. Hence, the
equation becomes
ay + yy = 0.
One more integration w.r.t. x leads to
y2
ay + = c2 .
2
This equation can be written as
2a dy
= d x,
b2 − y 2
with b2 = 2c2 . The variables are separable. By integration, we obtain
2a y
tanh−1 = x + c3 .
b b
We use the initial condition y(0) = 0, and obtain c3 = 0. Thus,

y

2a bx
tanh−1 = x, i.e., y = b tanh .
b b 2a
What makes the solution of such equations difficult is the integration. Let us
consider a next example.
Example 1.24 The mathematical model for the vibration of the simple pendulum is
given by the following ODE
g
θ + sin(θ) = 0,
l
where θ denotes the inclination of the string to the downward vertical, g is the
gravitational constant and l is the length of pendulum. Solve this equation.
Solution. In absence of an independent variable, we set θ ≡ p, then θ = p(dp/dθ).
Substituting these expressions into the given equation, we obtain
dp g
p + sin(θ) = 0.
dθ l
This is a first-order ODE and it can be shown that
p2 g
− cos(θ) = c1 .
2 l
Now, if we insert p = θ into the above equation, we get

g
θ = ±2 c1 + cos(θ).
l
The next integration is impossible. We have to use a numerical method to obtain the
solution (see e.g., the monographs [57, 63]).
Let us consider a last example, which exhibits another difficulty.
Example 1.25 Solve the ODE y = −λe y , y(0) = y(1) = 0.
Solution. This is the well-known Bratu-Gelfand equation in 1D planar coordinates.

The range of applications of this equation is including the solid fuel ignition model
in thermal combustion theory, physical applications arising from chemical reaction
theory, radiative heat transfer, nanotechnology as well as the expansion of the uni-
verse [63].
Set y ≡ p, then y = p(dp/dy). Then, the ODE becomes
dp
p = −λe y , i.e., p dp = −λe y dy.
dy
The integration of the above equation gives
p2
= −λe y + c1 .
2
This equation can be written as

p = ± a 2 − 2λe y ,
where a 2 = 2c1 . Let a 2 − 2λe y = z 2 . Then
2z z
−2λy e y = 2z z, i.e., y = .
a2− z2
The substitution of y = p into this expression leads to
2z z 2dz
= ±z, i.e. 2 = ±d x.
a2 − z2 a − z2
The integration of both sides gives
2 z
tanh−1 = c2 ± x.
a a
It follows
a
a
z = a tanh b ± x , b ≡ c2 .
2 2
Since a 2 − 2λe y = z 2 , the solution can be formulated in term of y as

a
a2
λe y = a 2 1 − tanh2 b ± x = a
.
2 cosh2 b ± x
2

⎛ ⎞
⎜ a ⎟
y(x) = 2 ln ⎝ √ a
⎠ .
2λ cosh b ± x
2
√ when we apply the boundary conditions. If we use y(0) = 0, we

Difficulties occur
come to a = 2λ cosh(b). Then, the second boundary condition y(1) = 0 implies
y(x) = 0, with λ = 0. This is the trivial solution. For λ = −π 2 , the exact solution is

√ x 1
y = −2 ln 2cos π − .
2 4
This BVP will be studied in the next chapter by approximation methods.
1.5 Transformations of Nonlinear ODEs
In many cases, a change of variables may be useful to solve a nonlinear ODE. Possible
choices are z = g(x), z = h(y), or other forms. Although the right choice of the
variables is difficult, in the majority of cases the presence of the variables in the given
ODE may help to find such a choice.
2
Example 1.26 Solve the IVP xe y y − e y = 2 , y(1) = 0.
x
Solution. Because of the presence of the function e y in the ODE, one can guess
z = e y . Hence, z = e y y . Substituting this into in the ODE and simplifying, we get
z 2
z − = 2.
x x
This is a linear ODE and it can be shown that its particular solution is
1
z = 2x − .
x
Transforming back the above equation by z = e y , we obtain

1
y = ln 2x −
x
as the solution of the given IVP.

Example 1.27 Solve the ODE
2y = (x − y) cot(x) − [(x − y) cot(x)]−1 + 2.

Solution. The presence of (x − y) encourages us to choose z = x − y. Hence,

y = 1 − z . If we substitute this into the ODE, we get
1 1
z + cot(x)z = cot(x)z −1 .
2 2
This is a Bernoulli equation and can easily be solved.
Example 1.28 Solve the ODE x 3 y + 2x 2 = (x y − y)2 .
Solution. We may write this equation as
y
2
x y + 2 = y − .
x
One can choose z = y/x. Then, y = z + x z and y = x z + 2z . Substituting these

expressions into the above equation and simplifying the result, yields
x 2 z + 2x z + 2 = x 2 (z )2 .
This equation does not contain the function z. Therefore, we set z ≡ p and get
z = p . By substituting these expressions into the equation, we obtain
x 2 p + 2x p + 2 = x 2 p 2 ,
which is a Riccati equation. It can easily be seen that p1 = −1/x is a particular

solution. Hence, we have z = − ln(x). That is, y = −x ln(x).
There are several approaches to find the family of solutions. The first one is to
set p ≡ −1/x + 1/z. With this choice, the family of solutions can be determined
as described above. This solution method brings to mind that another choice may be
y = −x ln(z). If we use this change of variable, we come to the following Cauchy-
Euler equation
x 2 z + 2x z − 2z = 0,
which can easily be solved.

A list of the exact solutions of some nonlinear ODEs and the corresponding
transformations can be found in [96].
Sometimes a group of transformations is required to obtain an ODE, which can
be solved exactly.
Example 1.29 Solve the ODE x 4 y + (x y − y)3 = 0.
Solution. If we rewrite the equation as
4
1
y + (x y − y)3 = 0,
x
1.5 Transformations of Nonlinear ODEs 23
we can set u ≡ x y − y and v ≡ 1/x. With these changes, the equation is transformed
into
du
= −vu 3 .
dv
Separating the variables, it can easily be shown that
1
u= .
v2 + c1 2
Substituting back the expressions for u and v, we obtain

x
x y − y = .
1 + c1 2 x 2
This is a first-order linear ODE. It is not difficult to show that

⎛ ⎛ ⎞⎞
x
y = x ⎝c2 + ln ⎝ ⎠⎠
c1 + c12 + x 2
is the desired solution.

Let us now consider a class of ODEs, which is given by
a 2 b c
y + (y ) + y + 2 y + d x r y s = 0, (1.23)
y x x
where a, b, c, d and r = −2, s = 1 are arbitrary constants (see [96]). A variant of

this equation is obtained by setting y ≡ eu and inserting this expression into (1.23).
It results
b c
u + (a + 1)(u )2 + u + 2 + d x r e(s−1)u = 0. (1.24)
x x
A large number of nonlinear ODEs, which play an important role in applied physics,
can be obtained from (1.23) or (1.24) by a specific choice of the constants. Typical
examples are:
1. Ivey’s equation
(y )2 2
y − + y + ky 2 = 0,
y x
Occurring in space-charge theory, it is a special case of (1.23) with a = −1,

b = 2, c = 0, d = k, r = 0 and s = 2.
2. Thomas-Fermi’s equation
y = x −1/2 y 3/2 .
This equation is a special case of (1.23) with a = b = c = 0, d = −1, r = −1/2

and s = 3/2.
3. Emden-Lane-Fowler equation
2
y + y = ym .
x
It is a special case of (1.23) with a = 0, b = 2, c = 0, d = 1, r = 0 and s = m.
4. Poisson-Boltzman’s equation
α
y + y = ey .
x
In case of plane, cylindrical, or spherical symmetry, we have α = 0, 1, 2, respec-
tively. This is a special case of (1.24) with a = −1, b = α, c = 0, d = −1,
r = 0, and s = 2.
5. Bratu’s equation
y = −λe y .
This is a special case of (1.24) with a = −1, b = c = 0, d = λ, r = 0 and s = 2.

A reliable procedure for obtaining a particular solution of such equations is to use
the special ansatz y ≡ ax n .
Let us consider the first three equations. When we set y ≡ ax n , we obtain y =
nax n−1 and y = an(n − 1)x n−2 . Substituting these expressions into the given
equations and equating both sides, gives the values of a and n. Let us first have a
look at Ivey’s equation
(y )2 2
y − + y + ky 2 = 0.
y x
We set y ≡ ax n and substitute y, y and y into the equation. It follows
n(n − 1)ax n−2 − an 2 x n−2 + 2nax n−2 + ka 2 x 2n = 0,
that is,
n(n − 1) − n 2 + 2n + kax n+2 = 0.

Since this relation must be satisfied for all x, we set x = 1. It can easily be seen
that a solution of the resulting algebraic equation is n = −2 and a = 2/k. Thus, a
particular solution of Ivey’s equation is
2
y= .
k x2
To compute a solution of the Thomas-Fermi’s equation and the Emden-Lane-
Fowler’s equation, we use the change of variables y ≡ ln(u) and transform these
equations into the form (1.23). Then, we set u = ax n and proceed as before.
Let us consider the Poisson-Boltzman’s equation with α = 2, i.e.
2
y + y = ey .
x
If we set y ≡ ln(u), we obtain
2 (u )2
u + u − = u2.
x u
Now, when we set u ≡ ax n , we have u = anx n−1 and u = an(n − 1)x n−2 . The
substitution of these expressions into the above ODE yields
n(n − 1) − n 2 + 2n = ax n+2 .
As before we can set x = 1. A solution of the resulting algebraic equation is n = a =

−2. Thus, we have u = −2x −2 , and a particular solution of the Poisson-Boltzman’s
equation is
2
y = ln − 2 . (1.25)
x
Another strategy by which one can attempt to determine a particular solution of

a second-order ODE is to reduce the order by a group of transformations. In some
cases it is then easier to solve the nonlinear first-order ODE. We will see that this is
true for the Poisson-Boltzman’s equation. Let us rewrite this equation in the form
x 2 y + 2x y = x 2 e y .
When we set u ≡ x y and v ≡ x 2 e y , the ODE is transformed into the equation
dv v(u + 2)
= .
du v−u
This equation cannot be solved analytically. But if we write
(v − u)dv − v(u + 2)du = 0,

then one may choose v = u and u = −2, i.e., x y = −2 and x 2 e y = −2. These two
relations lead to

2
y = ln − 2 . (1.26)
x
Comparing (1.25) with (1.26), we see that we have determined the same particular
solution as before.
Let us come back to the Thomas-Fermi’s equation
y = x −1/2 y 3/2 .
Setting u ≡ x 3 y and v ≡ x 4 y , we obtain
dv u 3/2 + 4v
= .
du v + 3u
We write

u 3/2 + 4v du − (v + 3u)dv = 0,
and demand
u 3/2 + 4v = 0 and v + 3u = 0.
It is not difficult to show that the intersection points are (144, −532) and (0, 0). If
we choose u = 144 or v = −532, we obtain the particular solution
144
y=
x3
of the Thomas-Fermi’s equation.
As we could see above, sometimes it is possible to solve some classes of ODEs
without imposing any additional condition. But this is not our goal. The main task
in this book is to solve initial or boundary value problems.
Remark 1.30 The first integral method, which is based on the ring theory of commu-
tative algebra, was first proposed by Feng [37] to solve the Benjamin-Bona-Mahony,
Gardner and Foam drainage equations. Here, we use a part of the first integral method
for the analytical treatment of the Foam drainage equation, without applying the divi-
sion theorem. First we transform this equation into an ODE by ξ = x + ct, which is
known as the wave variable. Then, we solve it by elementary methods.
Example 1.31 Solve the Foam drainage equation
√
∂u ∂ u ∂u
+ u −
2
= 0.
∂t ∂x 2 ∂x
Solution. First, we write the above equation in the form

2 √
∂u ∂u 1 ∂u u ∂2u
+2 u− √ − = 0.
∂t ∂x 4 u ∂x 2 ∂x 2
Using the wave variable ξ ≡ x + ct, this partial differential equation can be trans-
formed into the following ODE (see [107])
√
(u )2
u u
cu + 2uu − √ − = 0,
4 u 2
where (·) denotes the derivative w.r.t. the variable ξ. Assume u = v 2 , then u = 2vv
and u = 2((v )2 + vv ). If we substitute these expressions into the ODE, we get
2cvv + 4v 3 v − 2v(v )2 − v 2 v = 0.
Writing this equation in the form
2cvv + 4v 3 v − (v 2 v ) = 0,
and then integrating, we obtain
cv 2 + v 4 − v 2 v = c1 .
If we assume c1 = 0, we get
v 2 (c + v 2 − v ) = 0.
Thus, v 2 = 0, that is u = 0, which is an obvious solution. Moreover, it also applies

v = c + v 2 , where v = dv/dξ. By separating the variables, we can write
dv
= dξ.
c + v2
A simple integration leads to

1 v
√ tan √ = ξ + c2 .
c c
Since u = v 2 and ξ = x + ct, we get

1 u
√ tan = x + ct + c2 .
c c
It follows √ 2
u = c tan−1 c (x + ct + c2 ) .
This is a solution of the Foam drainage equation.

On the other hand, if we assume c1 = 0, then it can be shown that

−1 u 1 u
√ tan−1 + √ tan−1 = x + ct + c2 ,
2b a + b a+b 2b a − b a−b
where
c c2
a≡ and b ≡ + c2 .
2 4
This is another solution of the Foam drainage equation in implicit form.

The Benjamin-Bona-Mahony, Gardner equations can also be solved by this pro-
cedure (see [107]). We end this chapter by considering another example.
Example 1.32 Solve the following Burgers’ equation
∂u ∂u ∂2u
+u + 2 = 0, u(x, 0) = x.
∂t ∂x ∂x
Solution. Set ξ ≡ x + ct, then the PDE is transformed into
cu + uu + u = 0,
where (·) denotes the derivative w.r.t. the variable ξ. This is an autonomous ODE.
We set u ≡ p. Then, u = p(dp/du). Inserting these expressions into the ODE
yields
dp
cp + up + p = 0.
du
Obviously, p = 0, i.e., u = a is a solution. Moreover, it also applies
dp
= −(c + u).
du
Thus,
1
p = u = − (c + u)2 + c1 .
2
Let us assume that c1 = 0. It follows
du 1
− = dξ.
(c + u)2 2
By integration, we obtain
1 1 1
= ξ + c2 = (x + ct) + c2 .
c+u 2 2
Using the initial condition u(x, 0) = x, we determine the integration constant c2 as
1 1
c2 = − x.
c+x 2
This leads to following particular solution
2(c + x) 2x − c2 t (c + x)
u(x, t) = −c = .
2 + ct (c + x) 2 + ct (c + x)
1.6 Exercises
Exercise 1.1 Solve the following differential equations:

(1) tan(y)d x + (x
− sin2 (y))dy = 0,
y
(2) xd x + sin2 (yd x − xdy) = 0, y(1) = 0,
x
(3) sin(θ) − cos(θ) dr + r sin(θ) + cos(θ) dθ = 0,
(4) y (e x + y) d x + (e x + 2x y) dy = 0,
2 sin(2x)d2x + sin(y)
(5) 2 tan(y) − cos(2x) dy = 0,
(6) y 2x − x y + y d x − x 2 (2x − y)dy = 0,
(7) cos(2r
) − 2θ sin(2r
) − 2r θ 3
dr + cos(2r ) − 3r 2 θ2 dθ = 0,
(8) y ln(x) − ln(y) d x + √ (y − x ln(x) + x ln(y)) dy = 0, y(1) = 1,
(9) x − 2x sin(y)
= −x x sin(y),
(10) x 3 yd x + 3x 4 − y 3 dy = 0,
(11) y 3 sec2 (x)d x = 1 − 2y 2 tan(x) dy,
(12) (5x − 3y)d x + (4x + 2y + 1)dy = 0,
(13) 44(x − 2y + 2)d x − (x − 2y − 3)dy = 0,
(14) 2y(yd x + 2xdy) + x 3 (5yd x + 3xdy) = 0,
(15) 2x(2yd x + d x + dy) + y 3 (3yd x + 5xdy) = 0,
(16) y + x sin(2y) 2xe−x cos2 (y),
2
=
(17) 2x + 5x y y + 4x y + 3y 4 = 0,
2 3
(18) 2y 2 − ln(x) + 3x(y − 2x)y = 0,

√
(19) x y − 3y = x 5 3 y,
(20) x (x − sec(y) ln(x)) y + x tan(y) − y sec(y) = 0.
Exercise 1.2 Show that by choosing the appropriate change of variables, the
following equations lead to equations with separable variables:
(1) y = eay f (ay + ln(x)),

(2) x y + n = f (x)g(x n e y ),
(3) y + nx −1 ln(y) = y f (x)g(x 2n ln(y)),
(4) cos(y)y + cot(x) sin(y) = f (x)g(sin(x) sin(y)),
(5) sin(2x)y + sin(2y) = cos2 (y) f (x)g(tan(x) tan(y)).
Exercise 1.3 By introducing the appropriate change of variables, solve the following
equations:
y
(1) x y = y + x n+1 sec ,

x

(2) x y − y = y − x ,
2 4
3
(3) 3y 2 y = 2 y 3 − 4 − 5x y 3 − 4 ,
(4) x 2 y + x(y + 1)y + y = 0,
(5) x yy + 2x(y )2 + 3yy = 0,
(6) x(y + x)y + x(y
)2 − (y − x)y = y,

(7) x y x 2 − y 2 + x y = y − x 2 x 2 − y 2 ,
(8) x 2 y + x y + y = y −3 ,
(9) y = f (x)eay + g(x),
(10) cos(x)y + (y )2 = 1.
Exercise 1.4 Find the family of solutions of the following ODEs by reducing the
order of the equation:

(1) x 4 y = y y + x 3 ,
(2) yy + (y + 1)(y )2 = 0,
(3) yy + (y )3 = (y )2 ,

(4) e y y = x,
(5) yy = (y )2 1 − y sin(y) − yy cos(y) .
Exercise 1.5 For following nonlinear ODEs, find a particular solution:

(1) x 2 y − (y )2 + 2y = 0,

(2) x y + 3y = xe−y ,
(3) x 2 y − 2(y )3 + 6y = 0,
2
(4) y + y = y m , m = 3,
x
15
(5) y − 2 y = 3y 2 .
x
Exercise 1.6 For the following equation, discuss the particular solution for different
values of a, b, and c:
x 2 y + ax y = bx 2 ecy .
1.6 Exercises 31
Exercise 1.7 By using a group of transformations, find the family of solution of the
following ODEs:
(1) x 3 y = e y/x ,
(2) x 2 y + x y − y = x 4 y 2 ,
(3) tan(y)y = x n−1 cos(y),
(4) x y + y = x 2 eay (y )3 ,
(5) y − cot(x) tan(y) = sin(x) sin(y).
Chapter 2
Analytical Approximation Methods
2.1 Introduction
As we mentioned in the previous chapter, most of the nonlinear ODEs have no

explicit solutions, i.e., solutions, which are expressible in finite terms. Even if an
explicit solution can be determined, it is often too complicated to analyze the principal
features of this solution. Due to such difficulties, the study of nonlinear mathematical
problems is the most time-consuming and difficult task for researchers dealing with
nonlinear models in the natural sciences, engineering, and scientific computing. With
the increasing interest in the development of nonlinear models, a variety of analytical
asymptotic and approximation techniques have been developed in recent years to
determine approximate solutions of partial and ordinary differential equations. Some
of these techniques are the perturbation method, the variational iteration method, the
homotopy perturbation method, the energy balance method, the variational approach
method, the parameter-expansion method, the amplitude-frequency formulation, the
iteration perturbation method, and the Adomian decomposition method.
In this chapter, we present the variational iteration method and the Adomian
decomposition method since these techniques have good convergence characteristics
and can be used to treat strongly nonlinear ODEs.
2.2 The Variational Iteration Method
The variational iteration method (VIM) was first proposed by He (see e.g. [49, 50])
and systematically elucidated in [51, 54, 126]. The method treats partial and ordi-
nary differential equations without any need to postulate restrictive assumptions that
may change the physical structure of the solutions. It has been shown that the VIM
solves effectively, easily, and accurately a large class of nonlinear problems with
approximations converging rapidly to accurate solutions, see e.g. [127]. Examples

DOI 10.1007/978-81-322-2812-7_2
34 2 Analytical Approximation Methods
for such problems are the Fokker–Planck equation, the Lane–Emden equation, the
Klein–Gordon equation, the Cauchy reaction–diffusion equation, and biological pop-
ulation models.
To illustrate the basic idea of the VIM, we consider, the ODE
L y + N (y) = f (x), x ∈ I, (2.1)
where L and N are linear and nonlinear differential operators, respectively, and
f (x) is an given inhomogeneous term defined for all x ∈ I . In the VIM, a correction
functional of the Eq. (2.1) is defined in the following form
x
yn+1 (x) = yn (x) + λ(τ ) L yn (τ ) + N ( ỹn (τ )) − f (τ ) dτ , (2.2)
0
where λ(τ ) is a general Lagrange multiplier, which can be identified using the vari-
ational theory [38]. Furthermore, yn (x) is the nth approximation of y(x) and ỹn (x)
is considered as a restricted variation, i.e., δ ỹn (x) = 0.
By imposing the variation and by considering the restricted variation, Eq. (2.2) is
reduced to
x
δ yn+1 (x) = δ yn (x) + δ λ(τ )L yn (τ )dτ
0
τ τ =x
= δ yn (x) + λ(τ ) Lδ yn (ξ)dξ (2.3)
0 τ =0
x τ

− λ (τ ) Lδ yn (ξ)dξ dτ .
0 0
Obviously, in (2.3) we have used integration by parts, which is based on the

following formula

λ(τ )yn (τ )dτ = λ(τ )yn (τ ) − λ (τ )yn (τ )dτ . (2.4)
In the next sections, we will also use two other formulas for the integration by parts,
namely

λ(τ )yn (τ )dτ = λ(τ )yn (τ ) − λ (τ )yn (τ ) + λ (τ )yn (τ )dτ , (2.5)
and
λ(τ )yn (τ )dτ = λ(τ )yn (τ ) − λ (τ )yn (τ ) + λ (τ )yn (τ )
(2.6)
− λ (τ )yn (τ )dτ .
2.2 The Variational Iteration Method 35
Now, by applying the stationary conditions for (2.3), the optimal value of the
Lagrange multiplier λ(τ ) can be identified (see e.g. [50], formula (2.13) and the
next section). Once λ(τ ) is obtained, the solution of the Eq. (2.1) can be readily
determined by calculating the successive approximations yn (x), n = 0, 1, . . ., using
the formula (see Eq. (2.2))
x
yn+1 (x) = yn (x) + λ(τ ) L yn (τ ) + N (yn (τ )) − f (τ ) dτ , (2.7)
0
where y0 (x) is a starting function, which has to be prescribed by the user.

In the paper [50] it is shown, that the approximate solution yn (x) of the exact
solution y(x) can be achieved using any selected function y0 (x). Consequently, the
approximate solution is given as the limit y(x) = limn→∞ yn (x). In other words,
the correction functional (2.2) will give a sequence of approximations and the exact
solution is obtained at the limit of the successive approximations. In general, it is
difficult to calculate this limit. Consequently, an accurate solution can be obtained
by considering a large value for n. This value depends on the interval I where a good
approximation of the solution is desired.
Let us consider, the following IVP
y (x) + y(x)2 = 0, y(0) = 1. (2.8)
The corresponding exact solution is
y(x) = (1 + x)−1 = 1 − x + x 2 − x 3 + x 4 − x 5 + x 6 + · · · . (2.9)
Here, we have
L y = y, N (y) = y 2 , f (x) ≡ 0.
To determine the Lagrange multiplier, we insert these expressions into (2.2) and
obtain
x
dyn (τ )
yn+1 (x) = yn (x) + λ(τ ) + ỹn (τ )2 dτ . (2.10)
0 dτ
Making the above correction functional stationary w.r.t. yn , noticing that δ ỹn (x) = 0
and δ yn (0) = 0, it follows with (2.3)

dyn (τ )
x
δ yn+1 (x) = δ yn (x) + δ λ(τ ) dτ
0 dτ
τ τ =x
d δ yn (ξ)
= δ yn (x) + λ(τ ) dξ
dξ
τ =0
0
x τ
d δ yn (ξ)
− λ (τ ) dξ dτ
0 0 dξ
x
= δ yn (x) + λ(τ )δ yn (τ )|τ =x − λ (τ )δ yn (τ )dτ
0
.
= 0.
Thus, we obtain the equations
1 + λ(x) = 0 and λ (x) = 0. (2.11)
Now, we substitute the solution λ = −1 of (2.11) into (2.10). It results the successive
iteration formula
x
dyn (τ )
yn+1 (x) = yn (x) − + yn (τ )2 dτ . (2.12)
0 dτ
We have to choose a starting function y0 (x), which satisfies the given initial con-
dition y(0) = 1. Starting with y0 (x) ≡ 1, we compute the following successive
approximations
y0 (x) = 1,
y1 (x) = 1 − x,
1
y2 (x) = 1 − x + x 2 − x 3 ,
3
2 1 1 1
y3 (x) = 1 − x + x 2 − x 3 + x 4 − x 5 + x 6 − x 7 ,
3 3 9 63
13 1
y4 (x) = 1 − x + x 2 − x 3 + x 4 − x 5 + · · · − x 15 ,
15 59535
43 1
y5 (x) = 1 − x + x 2 − x 3 + x 4 − x 5 + x 6 − · · · − x 31 .
45 109876902975
In Fig. 2.1 the first iterates y0 (x), . . . , y4 (x) are plotted.
Comparing the iterates with the Taylor series of the exact solution (see (2.9)), we
see that in y5 (x) the first six terms are correct. The value of the exact solution at
x = 1 is y(1) = 1/2 = 0.5. In Table 2.1, the corresponding value is given for the
iterates yi (x), i = 0, . . . , 10.
d
In the above example, the linear operator is L = . More generally, let us
m
dx
d
assume that L = , m ≥ 1.
dxm
In [86], the corresponding optimal values of the Lagrange multipliers are given.
It holds
2.2 The Variational Iteration Method 37
Fig. 2.1 The first successive

iterates yi (x) for the IVP
(2.8). The solid line
represents the exact solution
y(x)
Table 2.1 The successive i yi (1)

iterates at the right boundary
for the IVP (2.8) 0 1
1 0.0
2 0.66666667
3 0.42857143
4 0.51954733
5 0.49529971
6 0.50094000
7 0.49983557
8 0.50002547
9 0.49999645
10 0.50000045
λ = −1, for m = 1,
λ = τ − x, for m = 2,
(2.13)
(−1)m
λ= (τ − x)m−1 , for m ≥ 1.
(m − 1)!
Substituting (2.13) into the correction functional (2.2), we get the following iteration
formula
x
(−1)m
yn+1 (x) = yn (x) + (τ − x)m−1 (L yn (τ ) + N (yn (τ )) − f (τ ))dτ ,
0 (m − 1)!
(2.14)
where y0 (x) must be given by the user.
2.3 Application of the Variational Iteration Method
In this section, we will consider some nonlinear ODEs and show how the VIM can
be used to approximate the exact solution of these problems.
Example 2.1 Solve the following IVP for the Riccati equation
y (x) + sin(x)y(x) = cos(x) + y 2 , y(0) = 0.
Solution. In (2.1), we set
dy
Ly ≡ , N (y) ≡ sin(x)y − y 2 , f (x) ≡ cos(x).
dx
Thus, the correction functional (2.2) is

x
dyn (τ )
yn+1 (x) = yn (x) + λ(τ ) + sin(τ ) ỹn (τ ) − ỹn (τ )2 − cos(τ ) dτ .
0 dτ
Since L is the first derivative, i.e., m = 1, a look at formula (2.13) shows that λ = −1
is the optimal value of the Langrange multiplier. The resulting successive iteration
formula is
x

yn+1 (x) = yn (x) − yn (τ ) + sin(τ )yn (τ ) − yn (τ )2 − cos(τ ) dτ . (2.15)
0
Let us choose y0 (x) ≡ 0 as starting function. Notice that y0 (x) satisfies the given
initial condition. Now, with (2.15) we obtain the following successive approximations
y1 (x) = sin(x),
y2 (x) = sin(x),
..
.
yn (x) = sin(x).
Obviously, it holds limn→∞ yn (x) = sin(x). The exact solution is y(x) = sin(x).
Example 2.2 Determine with the VIM a solution of the following IVP for the second
order ODE
y (x) + ω 2 y(x) = g(y(x)), y(0) = a, y (0) = 0. (2.16)

2.3 Application of the Variational Iteration Method 39
This problem is the prototype of nonlinear oscillator equations (see, e.g., [30, 99]).
The real number ω is the angular frequency of the oscillator and must be determined
in advance. Moreover, g is a known discontinuous function.
Solution. To apply the VIM, we set (see formula (2.1))
L y ≡ y + ω 2 y, N (y) ≡ −g(y), f (x) ≡ 0.
The corresponding correction functional is

x
d 2 yn (τ )
yn+1 (x) = yn (x) + λ(τ , x) + ω yn (τ ) − g( ỹn (τ )) dτ .
2
0 dτ 2
Before we identify an optimal λ, we apply the formula (2.5) for the following inte-
gration by parts
τ =x
x
∂λ(τ , x)
λ(τ , x)yn (τ )dτ = λ(τ , x)yn (τ ) |ττ =x
=0 − yn (τ )
0 ∂τ τ =0
x
∂ 2 λ(τ , x)
+ yn (τ )dτ .
0 ∂τ 2
Using this relation in the correction functional, imposing the variation, and making
the correction functional stationary, we obtain

∂λ(τ , x)
δ yn+1 (x) = δ yn (x) + − δ yn (τ )
λ(τ , x)δ yn (τ ) |τ =x
∂τ τ =x
x 2
∂ λ(τ , x)
+ + ω 2 λ(τ , x) δ yn (τ )dτ
0 ∂τ 2
.
= 0.
Thus, the stationary conditions are
∂ 2 λ(τ , x)
δ yn : + ω 2 λ(τ , x) = 0,
∂τ 2
δ yn : λ(τ , x)|τ =x = 0, (2.17)

∂λ(τ , x)
δ yn : 1− = 0.
∂τ τ =x
The solution of the Eqs. in (2.17) is
1
λ(τ , x) = sin(ω(x − τ )), (2.18)
ω
which leads to the following iteration formula
yn+1 (x) = yn (x)

2
1 x d yn (τ ) (2.19)
+ sin(ω(x − τ )) + ω yn (τ ) − g(yn (τ )) dτ .
2
ω 0 dτ 2
Example 2.3 Let us consider the following IVP of the Emden-Lane-Fowler equation
(see e.g. [48])
2
y + y + x k y μ = 0, y(0) = 1, y (0) = 0.
x
This ODE is used to model the thermal behavior of a spherical cloud of gas acting
under the mutual attraction of its molecules.
Solve this equation for k = 0 and μ = 5, which has a closed form solution.
Solution. For the given parameters, the IVP to be solved is
2
y + y + y 5 = 0, y(0) = 1, y (0) = 0.
x
Obviously, there is a singularity at x = 0. To overcome this singularity, we set
y ≡ z/x. Then, we get
z + x −4 z 5 = 0, z(0) = 0, z (0) = 1.
We set
d2z
Lz ≡ , N (z) ≡ x −4 z 5 , f (x) ≡ 0.
dx2
Thus, the correction functional (2.2) is

x
d 2 z n (τ ) −4 5
z n+1 (x) = z n (x) + λ(τ ) + τ z̃ n (τ ) dτ .
0 dx2
Looking at formula (2.13), we obtain for m = 2 the Lagrange multiplier λ = τ − x.

Therefore, the corresponding iteration formula (2.14) is

x
d 2 z n (τ ) −4 5
z n+1 (x) = z n (x) + (τ − x) + τ z n (τ ) dτ .
0 dx2
Starting with z 0 (x) = x, we obtain the following successive approximations
z 0 (x) = x,
x3
z 1 (x) = x − ,
6
x3 x5
z 2 (x) = x − + ,
6 24
x3 x5 x7
z 3 (x) = x − + − .
6 24 432
−1/2
x2 x4 5x 6 x2
lim z n (x) = z(x) = x 1 − + − + ··· = x 1 + .
n→∞ 6 24 432 3
Thus −1/2
z(x) x2
y(x) = = 1+
x 3
is the exact solution of the given IVP.
Example 2.4 One of the problems that has been studied by several authors is Bratu’s
BVP (see, e.g., [18, 71, 80, 87, 100, 108]), which is given in one-dimensional planar
coordinates by
y = −αe y , y(0) = 0, y(1) = 0, (2.20)
where α > 0 is a real parameter. This BVP plays an important role in the theory of
the electric charge around a hot wire and in certain problems of solid mechanics.
The exact solution of (2.20) is

cosh(0.5(x − 0.5)θ)
y(x) = −2 ln ,
cosh(0.25θ)
where θ satisfies √
θ= 2α cosh(0.25θ).
Bratu’s problem has zero, one or two solutions when α > αc , α = αc , and α < αc ,
respectively, where the critical value αc satisfies

1 = 0.25 2αc sinh(0.25θ).

In Chap. 5, the value αc is determined as
αc = 3.51383071912.
Use the VIM to solve the BVP (2.20).
Solution. Let us expand e y and use three terms of this expansion. We obtain
∞
yi y2
y + αe y = y + α ≈ y + α 1 + y + .
i=0
i! 2
Setting
d2 y y2
Ly ≡ , N (y) ≡ α 1 + y + , f (x) ≡ 0,
dx2 2
the corresponding correction functional (2.2) is

x
d 2 yn (τ ) ỹn (τ )2
yn+1 (x) = yn (x) + λ(τ ) + α 1 + ỹn (τ ) + dτ .
0 dτ 2 2
Looking at formula (2.13), we obtain for m = 2 the Lagrange multiplier λ = τ − x.

Therefore, the corresponding iteration formula (2.14) is

x
d 2 yn (τ ) yn (τ )2
yn+1 (x) = yn (x) + (τ − x) + α 1 + yn (τ ) + dτ .
0 dτ 2 2
(2.21)
Let us start with y0 (x) = kx, where k is a real number. The next iterate is

x
k2τ 2
y1 (x) = kx + α (τ − x) 1 + kτ + dτ .
0 2
Integrating by parts leads to
αx 2 αkx 3 λk 2 x 4
y1 (x) = kx − − − .
2! 3! 4!
Substituting y1 (x) into the right-hand side of (2.21), we obtain the next iterate
αx 2 αkx 3 αk 2 x 4
y2 (x) = kx − − −
2! 3! 4!
x
ατ 2 2αkτ 3 α αk
+ (τ − x) − − + (3α − 5k 2 )τ 4 + (2α − k 2 )τ 5
0 2 3 24 24

5α2 k 2 τ 6 α2 k 3 τ 7 α2 k 4 τ 8
+ + + dτ .
144 144 1152
Again, integration by parts yields
αx 2 αkx 3 α(k 2 − α)x 4 4α2 kx 5 α2 (5k 2 − 3α)x 6

y2 (x) = kx − − − + +
2! 3! 4! 5! 6!
5α2 k(k 2 − 2α)x 7 25α3 k 2 x 8 35α3 k 3 x 9 35α3 k 4 x 10
+ − − − .
7! 8! 9! 10!
(2.22)
The function y2 (x) must satisfy the initial conditions (see formula (2.20)). For a given
α, the equation y2 (1) = 0 is a fourth degree polynomial in k. When an appropriate
k is chosen from the corresponding four roots, the function y2 (x) can be accepted as
an approximation of the exact solution y(x) for x ∈ (0, 1).
Let us consider Bratu’s problem with α = 1. The polynomial in k is
35k 4 − 3250k 3 + 128250k 2 − 3137760k + 1678320 = 0.
Solving this algebraic equation by a numerical method, the following approximated

roots are obtained:
k1 = 0.546936690480377, k2 = 55.687874088793869,
k3,4 = 18.311166038934306 ± 35.200557613929831 · i.
When we substitute k = k1 and α = 1 into (2.22), the next iterate is determined. In

Table 2.2, y2 (x) is compared with y(x) for x = 0.1, 0.2, . . . , 0.9.
Next, let us consider Bratu’s problem with α = 2. The polynomial in k is
7k 4 − 290k 3 + 5490k 2 − 71136k + 78624 = 0.
Solving this algebraic equation by a numerical method, the following approximated

roots are obtained:
Table 2.2 Numerical results for Bratu’s problem with α = 1; θ = 1.51716459905075436852

18444212962
x y2 (x) y(x) Relative error
0.1 0.049605613312791 0.049846791245413 0.004838384309118
0.2 0.088710501172383 0.089189934628823 0.005375421099202
0.3 0.116898896821669 0.117609095767941 0.006038639627618
0.4 0.133864761192820 0.134790253884190 0.006866169212538
0.5 0.139428101877514 0.140539214400472 0.007906067553444
0.6 0.133548005691170 0.134790253884190 0.009216157379500
0.7 0.116331618650518 0.117609095767941 0.010862060532661
0.8 0.088038206883444 0.089189934628823 0.012913203156517
0.9 0.049077324553953 0.049846791245413 0.015436634379762
Table 2.3 Numerical results for Bratu’s problem with α = 2; θ = 2.357551053877402042593

9799885899
x y2 (x) y(x) Relative error
0.1 0.110752223723751 0.114410743267745 0.031977062988159
0.2 0.199192293844683 0.206419116487609 0.035010432976831
0.3 0.263300560244860 0.273879311825552 0.038625595742077
0.4 0.301549640325367 0.315089364225670 0.042971059761337
0.5 0.313080790744633 0.328952421341114 0.048249015866102
0.6 0.297850396206136 0.315089364225670 0.054711361209852
0.7 0.256726478126018 0.273879311825552 0.062629168976661
0.8 0.191509696354622 0.206419116487609 0.072228872919732
0.9 0.104847118151250 0.114410743267745 0.083590271711759
k1 = 1.211500000137995, k2 = 25.631365803713045,
k3,4 = 7.292852812360195 ± 17.564893217135829 · i.
As before, when we substitute k = k1 and α = 2 into (2.22), the next iterate is

determined. In Table 2.3, y2 (x) is compared with y(x) for x = 0.1, 0.2, . . . , 0.9.
The results in the Tables 2.2 and 2.3 show that the VIM is efficient and quite
reliable. Only two iterations have lead to acceptable results. There is no doubt, if
more terms of the expansion and/or more iterates are used, the VIM will generate
far better results.
2.4 The Adomian Decomposition Method
The Adomian decomposition method (ADM) is a semi-analytical technique for solv-

ing ODEs and PDEs. The method was developed by the Armenian-American math-
ematician George Adomian [7, 8, 9]. The ADM is based on a decomposition of the
solution of nonlinear operator equations in appropriate function spaces into a series
of functions. The method, which accurately computes the series solution, is of great
interest to the applied sciences. The method provides the solution in a rapidly con-
vergent series with components that are computed elegantly. The convergence of this
method is studied in [1, 2, 73].
Let the general form of an ODE be
F(y) = f,
where F is the nonlinear differential operator with linear and nonlinear terms. In
the ADM, the linear term is decomposed as L + R, where L is an easily invertible
operator and R is the remainder of the linear term. For convenience L is taken as the
highest-order derivative. Thus the ODE may be written as
2.4 The Adomian Decomposition Method 45
L y + Ry + N (y) = f (x), (2.23)
where N (y) corresponds to the nonlinear terms.

d
Let L be a first-order differential operator defined by L ≡ . If L is invertible,
dx
then the inverse operator L −1 is given by
x
L −1 (·) = (·)dτ .
0
Thus,
L −1 L y = y(x) − y(0). (2.24)
d2
Similarly, if L 2 ≡ , then the inverse operator L −1 is regarded as a double
dx2
integration operator given by
x τ
L −1 (·) = (·)dt dτ .
0 0
It follows
L −1 L y = y(x) − x y (0). (2.25)
We can use the same operations to find relations for higher-order differential
d3
operators. For example, if L 3 ≡ , then it is not difficult to show that
dx3
1 2
L −1 L y = y(x) − y(0) − x y (0) − x y (0). (2.26)
2!
The basic idea of the ADM is to apply the operator L −1 formally to the expression
L y(x) = f (x) − Ry(x) − N (y(x)).
This yields
y(x) = Ψ0 (x) + g(x) − L −1 Ry(x) − L −1 N (y(x)), (2.27)
where the function g(x) represents the terms, which result from the integration of
f (x), and
⎧ d
⎪
⎪ y(0), for L = ,
⎪
⎪
⎪
⎪
d x2
⎪
⎪ d
⎨ y(0) + x y (0), for L 2 = ,
Ψ0 (x) ≡ d x32
⎪
⎪ 1 d
⎪
⎪ y(0) + x y (0) + x 2 y (0), for L 3 = ,
⎪
⎪ 2! d x43
⎪
⎪ 1 1 d
⎩ y(0) + x y (0) + x 2 y (0) + x 3 y (0), for L 4 = .
2! 3! dx4
Now, we write
∞
∞

y(x) = yn (x) and N (y(x)) = An (x),
n=0 n=0
where
An (x) ≡ An (y0 (x), y1 (x), . . . , yn−1 (x))
are known as the Adomian polynomials. Substituting these two infinite series into
(2.27), we obtain
∞
∞
∞

yn (x) = Ψ0 (x) + g(x) − L −1 R yn (x) − L −1 An (x). (2.28)
n=0 n=0 n=0
Identifying the zeroth component y0 (x) by Ψ0 (x) + g(x), the remaining components
yk (x), k ≥ 1, can be determined by using the recurrence relation
y0 (x) = Ψ0 (x) + g(x),

(2.29)
yk (x) = −L −1 Ryk−1 (x) − L −1 Ak−1 (x), k = 1, 2, . . .
Obviously, when some of the components yk (x) are determined, the solution y(x)
can be approximated in form of a series. Under appropriate assumptions, it holds

n
y(x) = lim yk (x).
n→∞
k=0
The polynomials Ak (x) are generated for each nonlinearity so that A0 depends only
on y0 , A1 depends only on y0 and y1 , A2 depends on y0 , y1 , y2 , etc. [7]. An appropriate
strategy to determine the Adomian polynomials is
A0 = N (y0 ),
A1 = y1 N (y0 ),
1 2
A2 = y2 N (y0 ) + y N (y0 ),
2! 1
1
A3 = y3 N (y0 ) + y1 y2 N (y0 ) + y13 N (3) (y0 ),
3!
1 2 1
A4 = y4 N (y0 ) + y2 + y1 y3 N (y0 ) + y12 y2 N (3) (y0 )
2! 2!
1 4 (4)
+ y1 N (y0 ),
4!
..
.
dk
where N (k) (y) ≡ N (y).
dy k
2.4 The Adomian Decomposition Method 47
The general formula is

∞

1 ∂k
Ak = N yn λ n
, k = 0, 1, 2, . . . (2.30)
k! ∂λk n=0 λ=0
A Mathematica program that generates the polynomials Ak automatically can

be found in [16]. Moreover, in [20] a simple algorithm for calculating Adomian
polynomials is presented. According to this algorithm the following formulas for the
Adomian polynomials result:
A0 = N (y0 ),
A1 = y1 N (y0 ),
1 2
A2 = y2 N (y0 ) + y N (y0 ),
2 1
1
A3 = y3 N (y0 ) + y1 y2 N (y0 ) + y13 N (3) (y0 ),
6
1 2 1 1 4 (4)
A4 = y4 N (y0 ) + y1 y3 + y2 N (y0 ) + y12 y2 N (3) (y0 ) +

y N (y0 ).
2 2 24 1
(2.31)
Before we highlight a few examples and show how the ADM can be used to solve
concrete ODEs, let us list the Adomian polynomials for some classes of nonlinearity.
1. N (y) = exp(y):
A0 = exp(y0 ), A1 = y1 exp(y0 ),

1 2 1 3 (2.32)
A2 = y2 + y1 exp(y0 ), A3 = y3 + y1 y2 + y1 exp(y0 );
2! 3!
2. N (y) = ln(y), y > 0:

y1
A0 = ln(y0 ), A1 = ,
y0
(2.33)
y2 1 y12 y3 y1 y2 1 y13
A2 = − , A3 = − 2 + ;
y0 2 y02 y0 y0 3 y02
3. N (y) = y 2 :
A0 = y02 , A1 = 2y0 y1 ,
(2.34)
A2 = 2y0 y2 + y12 , A3 = 2y0 y3 + 2y1 y2 ;
4. N (y) = y 3 :
A0 = y03 , A1 = 3y02 y1 ,
(2.35)
A2 = 3y02 y2 + 3y0 y12 , A3 = 3y02 y3 + 6y0 y1 y2 + y13 ;
5. N (y) = yy :
A0 = y0 y0 , A1 = y0 y1 + y0 y1 ,

(2.36)
A2 = y0 y2 + y1 y1 + y2 y0 , A3 = y0 y3 + y1 y2 + y2 y1 + y3 y0 ;
6. N (y) = (y )2 :
A0 = (y0 )2 , A1 = 2y0 y1 ,

(2.37)
A2 = 2y0 y2 + (y1 )2 , A3 = 2y0 y3 + 2y1 y2 ;
7. N (y) = cos(y):
A0 = cos(y0 ), A1 = −y1 sin(y0 ),

1
A2 = −y2 sin(y0 ) − y12 cos(y0 ), A3 = −y3 sin(y0 ) − y1 y2 cos(y0 )
2
1 3
+ y sin(y0 );
6 1
(2.38)
8. N (y) = sin(y):
A0 = sin(y0 ), A1 = y1 cos(y0 ),
1 2
A2 = y2 cos(y0 ) − y sin(y0 ), A3 = y3 cos(y0 ) − y1 y2 sin(y0 ) (2.39)
2 1
1
− y13 cos(y0 ).
6
2.5 Application of the Adomian Decomposition Method
In this section, we consider some IVPs for first-order and second-order ODEs.
Example 2.5 Solve the IVP
y (x) = 1 − x 2 y(x) + y(x)3 , y(0) = 0.
Solution. This is Abel’s equation and its exact solution is y(x) = x. We apply the
ADM to solve it. First, let us look at formula (2.23). We have
L y ≡ y, Ry ≡ x 2 y, N (y) ≡ −y 3 , f (x) ≡ 1.

2.5 Application of the Adomian Decomposition Method 49
Using (2.27), we obtain
y(x) = Ψ0 (x) + g(x) − L −1 Ry(x) − L −1 N (y(x))

x x
= y(0) + x − τ y(τ )dτ +
2
y(τ )3 dτ .
0 0
Now, applying formula (2.28), we get
∞ ∞
∞
x x
yn (x) = x − τ 2
yn (τ ) dτ − An (τ ) dτ .
n=0 0 n=0 0 n=0
The Adomian polynomials, which belong to the nonlinearity N (y) = y 3 , are given
in (2.35). Setting
y0 (x) = Ψ0 (x) + g(x) = y(0) + x = x,
the recurrence relation (2.29) yields

x x x x
y1 (x) = − τ dτ −
3
A0 (τ )dτ = − τ dτ + τ 3 dτ = 0, 3
x
0 0
x 0 0
y2 (x) = − τ 2 y1 (τ )dτ + 3 y0 (τ )2 y1 (τ )dτ = 0,

0 0
..
.
Hence
∞

y(x) = yk (x) = x + 0 + · · · + 0 + · · · = x.
k=0
y (x) + x e y(x) = 0, y(0) = 0.
Solution. Here, we have
L y ≡ y, Ry ≡ 0, N (y) ≡ x e y , f (x) ≡ 0.
y(x) = Ψ0 (x) + g(x) − L −1 Ry(x) − L −1 N (y(x))

x x
= y(0) − τ e y(τ ) dτ = − τ e y(τ ) dτ .
0 0
∞ ∞

x
yn (x) = − τ An (τ ) dτ .
n=0 0 n=0
The Adomian polynomials, which belong to the nonlinearity exp(y), are given in
(2.32). Setting
y0 (x) = Ψ0 (x) + g(x) = y(0) = 0,

x
x2
y1 (x) = − τ · 1 dτ = − ,
0 2
x x 3
τ2 τ x4
y2 (x) = τ · dτ = dτ = ,
2 2 8
x 4 2 2
0 0
x 5
τ 1 τ τ
y3 (x) = − τ + − · 1 dτ = − dτ
0 8 2 2 0 4
x6
=− ,
24
..
.
Thus, 2 n
x
1 x
yn (x) = − τ An−1 (τ )dτ = − , n = 1, 2, . . . ,
0 n 2
and it holds
∞

1 1 1 x2
y(x) = yn (x) = − x 2 + x 4 − x 6 + · · · = − ln 1 + .
n=0
2 8 24 2
y (x) + 2y(x)y (x) = 0, y(0) = 0, y (0) = 1.
L y ≡ y , Ry ≡ 0, N (y) ≡ 2yy , f (x) ≡ 0.

y(x) = Ψ0 (x) + g(x) − L −1 Ry(x) − L −1 N (y(x))

x τ

= y(0) + x y (0) − 2 y(t)y (t)dtdτ .
0 0
∞ ∞
x τ
yn (x) = x − 2 An (t) dtdτ .
n=0 0 0 n=0
The Adomian polynomials, which belong to the nonlinearity yy , are given in (2.36).
Setting
y0 (x) = Ψ0 (x) + g(x) = y(0) + x y (0) = x,

τ x τ
x
y1 (x) = −2 y0 (t)y0 (t) dtdτ = −2 t dtdτ
0 0 0 0
x 2 x
τ x3
= −2 dτ = − τ 2 dτ = − ,
2 3
0 x τ 0

y2 (x) = −2 y0 (t)y1 (t) + y0 (t)y1 (t) dtdτ
0 0
x τ
t3 4·2 x τ 3
=2 1 · + t · t 2 dtdτ = t dtdτ
0 0 3 3 0 0

2 x 4 2 5
= τ dτ = x ,
3 0 15
x τ

y3 (x) = −2 y0 (t)y2 (t) + y1 (t)y1 (t) + y2 (t)y0 (t) dtdτ
0 0
x τ 3
2 5 2 t 2 4
= 1 · t + t · + t · t dtdτ
0 0 15 3 3
x τ x
2 · 17 17 17 7
=− t dtdτ = −
5
τ 6 dτ = − x .
15 0 0 15 · 3 0 315
Obviously, it holds
∞
x3 2 17 7
y(x) = yn (x) = x − + x5 − x + ···
n=0
3 15 315

y (x) − y (x)2 + y(x)2 = e x , y(0) = y (0) = 1.
L y ≡ y , Ry ≡ 0, N (y) ≡ −(y )2 + y 2 , f (x) ≡ e x .
We write
N (y) = N1 (y) + N2 (y), N1 (y) ≡ −(y )2 , N2 (y) ≡ y 2 .
y(x) = Ψ0 (x) + g(x) − L −1 N1 (y(x)) − L −1 N2 (y(x))

x τ x τ
= y(0) + x y (0) + et dtdτ + y (t)2 dtdτ
x τ 0 0 0 0
− 2
y(t) dtdτ
0 0
∞ ∞
x τ
yn (x) = 1 + x + e − x − 1 + x
An (t) dtdτ
n=0 0 0 n=0
∞

x τ
− Bn (t) dtdτ .
0 0 n=0
The Adomian polynomials An , which belong to the nonlinearity (y )2 , are given in

(2.37), and for the nonlinearity y 2 , the Adomian polynomials Bn are given in (2.34).
Setting
y0 (x) = Ψ0 (x) + g(x) = e x ,

x τ x τ
y1 (x) = A0 (t)dtdτ − B0 (t)dtdτ
0 x 0 τ 0x 0τ
t 2 t 2
= e dtdτ − e dtdτ = 0.
0 0 0 0
This implies yn (x) ≡ 0, n = 1, 2, . . ., and we obtain the exact solution of the given
problem:
y(x) = e x + 0 + 0 + · · · = e x .
The convergence of the ADM can be accelerated if the so-called noise terms
phenomenon occurs in the given problem (see, e.g., [9, 10]). The noise terms are
the identical terms with opposite sign that appear within the components y0 (x) and
y1 (x). They only exist in specific types of nonhomogeneous equations. If noise terms
indeed exist in the y0 (x) and y1 (x) components, then, in general, the solution can be
obtained after two successive iterations.
By canceling the noise terms in y0 (x) and y1 (x), the remaining non-canceled
terms of y0 (x) give the exact solution. It has been proved that a necessary condition
for the existence of noise terms is that the exact solution is part of y0 (x).
y (x) − y (x)2 + y(x)2 = 1, y(0) = 1, y (0) = 0.
Solution. As in the Example 2.8, we set
L y ≡ y , Ry ≡ 0, N1 (y) ≡ −(y )2 , N2 (y) ≡ y 2 , f (x) ≡ 1.

x τ x τ
x2
y(x) = 1 + + y (t) dtdτ −
2
y(t)2 dtdτ
2 0 0 0 0
∞
x τ ∞

x2
yn (x) = 1 + + An (t) dtdτ
n=0
2 0 0 n=0
x τ ∞

− Bn (t) dtdτ .
0 0 n=0
The Adomian polynomials An , which belong to the nonlinearity (y )2 , are given in

(2.37), and for the nonlinearity y 2 , the Adomian polynomials Bn are given in (2.34).
Setting
1
y0 (x) = 1 + x 2 ,
2

x τ x τ 2
t2
y1 (x) = t 2 dtdτ − 1+ dtdτ
0 0 0 0 2
x τ τ x
t3 1 5 1 3
= t + tdτ −
+ t dτ
0 0 0 3 20 3 0
x 3 x
τ 1 5 1 3
= dτ − τ + τ + τ dτ
0 3 0 20 3
1 4 1 6 1 4 1 2
= x − x − x − x
12 120 12 2
1 6 1 2
=− x − x .
120 2
1
Comparing y0 (x) with y1 (x), we see that there is the noise term x 2 . Therefore, we
2
can conclude that the solution of the given IVP is y(x) = 1.
Several authors have proposed a variety of modifications of the AMD (see, e.g.,
[12]) by which the convergence of the iteration (2.29) can be accelerated. Wazwaz
[124, 126] suggests the following reliable modification which is based on the assump-
tion that the function h(x) ≡ Ψ0 (x) + g(x) in formula (2.27) can be divided into
two parts, i.e.,
h(x) ≡ Ψ0 (x) + g(x) = h 0 (x) + h 1 (x).
The idea is that only the part h 0 (x) is assigned to the component y0 (x), whereas
the remaining part h 1 (x) is combined with other terms given in (2.29). It results the
modified recurrence relation
y0 (x) = h 0 (x),
y1 (x) = h 1 (x) − L −1 Ry0 (x) − L −1 A0 (x), (2.40)
−1 −1
yk (x) = −L Ryk−1 (x) − L Ak−1 (x), k = 2, 3, . . .
y (x) − y(x)2 = 2 − x 4 , y(0) = y (0) = 0.
L y ≡ y , Ry ≡ 0, N (y) ≡ −y 2 , f (x) ≡ 2 − x 4 .
y(x) = Ψ0 (x) + g(x) − L −1 Ry(x) − L −1 N (y(x))

x τ
1 6
=x − x +
2
y(t)2 dtdτ .
30 0 0
∞ ∞
1 x τ
yn (x) = x − x 6 +
2
An (t) dtdτ .
n=0
30 0 0 n=0
The Adomian polynomials, which belong to the nonlinearity y 2 , are given in (2.34).
Dividing
1
h(x) = x 2 − x 6
30
1
into h 0 (x) ≡ x 2 and h 1 (x) ≡ − x 6 , and starting with y0 (x) = x 2 , the recurrence
30
relation (2.29) yields
x τ
1 6 2 2
y1 (x) = − x + t dtdτ
30 0 0
x 5
1 τ
= − x6 + dτ
30 0 5
1 1
= − x 6 + x 6 = 0.
30 30
This implies
yk (x) = 0, k = 1, 2, . . .
Thus, we can conclude that the exact solution of the given IVP is y(x) = x 2 .
Let us compare the modified ADM with the standard method. The ADM is based
on the recurrence relation (2.29). Here, we have to set
1 6
y0 (x) = x 2 − x .
30
Now, the recurrence relation (2.29) yields
τ x
x
1 6 2 1 1 9 1 5
y1 (x) = t − t dtdτ =
2
τ −
13
τ − τ dτ
0 0 30 0 11700 135 5
1 6 1 10 1
= x − x + x 14 .
30 1350 163800
x τ
1 1 6 1 10 1
y2 (x) = 2 t2 − x 14 −
t x + x 6 dtdτ
0 0 16380030 1350 30
x
1 227 1 13 1 9
= − τ 21 + τ 17 − τ + τ dτ
0 51597000 62653500 3510 135
1 10 1 227 1
= x − x 14 + x 18 − x 22 .
1350 49140 1127763000 1135134000
Since
y(x) = y0 (x) + y1 (x) + y2 (x) + · · · ,
we see that the terms in x 6 and x 10 cancel each other. The cancelation of terms is
continued when further components yk , k ≥ 3, are added.
This is an impressive example of how fast the modified ADM generates the exact
solution y(x) = x 2 , compared with the standard method.
Many problems in the mathematical physics can be formulated as ODEs of Emden-

Fowler type (see, e.g., [26, 30]) defined in the form
α
y (x) +y (x) + β f (x)g(y) = 0, α ≥ 0,
x (2.41)
y(0) = a, y (0) = 0,
where f and g are given functions of x and y, respectively. The standard Emden-
Lane-Fowler ODE results when we set f (x) ≡ 1 and g(y) ≡ y n .
Obviously, a difficulty in the analysis of (2.41) is the singularity behavior that
occurs at x = 0. Before the ADM can be applied, a slight change of the problem
is necessary to overcome this difficulty. The strategy is to define the differential
operator L in terms of the two derivatives, y + (α/x)y , which are contained in the
ODE. First, the ODE (2.41) is rewritten as
L y(x) = −β f (x)g(y), (2.42)
with
−α d d α
L≡x x .
dx dx
The corresponding inverse operator L −1 is

x τ
−1 −α
L (·) = τ t α (·)dtdτ .
0 0
Applying L −1 to the first two terms of (2.41), we obtain

α
L −1 y (x) + y (x)
x τx
−α α
= τ t α y (t) + y (t) dtdτ
t
0 x 0
τ τ
α α−1 α−1
= τ y (τ ) − αt y (t)dt + αt y (t)dt dτ
0 0 0
x
= y (τ )dτ = y(x) − y(0) = y(x) − a.
0
Now, operating with L −1 on (2.41), we find
y(x) = a − β L −1 ( f (x)g(y)). (2.43)
It is interesting, that only the first initial condition is sufficient to represent the solution
y(x) in this form. The second initial condition can be used to show that the obtained
solution satisfies this condition.
Let us come back to the Adomian decomposition method. As before, the solution
y(x) is represented by an infinite series of components
∞

y(x) = yn (x). (2.44)
n=0
In addition, the given nonlinear function g(y) is represented by an infinite series of

Adomian polynomials (as we have done it for N (y))
∞

g(y(x)) = An (x), (2.45)
n=0
where
An (x) ≡ An (y0 (x), y1 (x), . . . , yn−1 (x)).
Substituting (2.44) and (2.45) into (2.43) gives
∞
∞

−1
yn (x) = a − β L f (x) An (x) . (2.46)
n=0 n=0
Now, the components yn (x) are determined recursively. The corresponding recur-
rence relation is
y0 (x) = a,
yk (x) = −β L −1 ( f (x)Ak−1 (x)), k = 1, 2, . . . ,
or equivalently
y0 (x) = a,
x τ
−α α
(2.47)
yk (x) = −β τ t ( f (t)Ak−1 (x))dtdτ , k = 1, 2, . . .
0 0
2
y (x) + y (x) + e y(x) = 0, y(0) = y (0) = 0.
x
Solution. This problem is a special case of (2.41), where α = 2, β = 1, f (x) ≡ 1,

g(y) = exp(y) and a = 0. A particular solution of the ODE is

2
y(x) = ln .
x2
Obviously, this solution does not satisfy the initial conditions. We will see that the
ADM can be used to determine a solution which satisfies the ODE as well as the
initial conditions.
For the nonlinearity g(y) = exp(y), the Adomian polynomials are given in
Eq. (2.32). Using the recurrence relation (2.47), we obtain
y0 (x) = 0,
τ
x x
τ3 x
τ
y1 (x) = − τ −2 t 2 · 1 dtdτ = − τ −2 dτ = − dτ
0 0 0 3 0 3
1
= − x 2,
6
x τ x
τ 4
−t 2 t
y2 (x) = − τ −2 dtdτ =t2 · τ −2 dtdτ
0 0 6 0 0 6
x x 3
τ5 τ
= τ −2 dτ = dτ
0 30 0 30
1 4
= x ,
120
x τ 4 x τ 6
−2 t t4 −2 t
y3 (x) = − τ t ·
2
+ dtdτ = − τ dtdτ
0 0 120 72 0 0 45
x τ
τ7 τ5
=− τ −2 dτ = − dτ
0 315 0 315
1
=− x 6,
1890
x τ
1 6 1 2 1 4 1 1 6
y4 (x) = − τ −2 t2 − t − t · t − · t dtdτ
0 0 1890 6 120 6 216
x x
61 61
= τ −2 τ 9 dτ = τ 7 dτ
0 204120 0 204120
61
= x .
8
1632960
Thus, we have
1 1 4 1 61
y(x) = − x 2 + x − x6 + x8 + · · ·
6 120 1890 1632960
In [125] further variants of the general Emden-Fowler equation are discussed, and
solved by the ADM.
2.6 Exercises
Exercise 2.1 Solve following ODEs by VIM or ADM:

(1) y (x) − y(x) = −y(x)2 , y(0) = 1,

y(x) y(x) 2
(2) y (x) = 1 + + , y(0) = 0,
x x
5 y(x)
(3)
y (x) + y (x) + exp(y(x)) + 2 exp = 0, y(0) = y (0) = 0,
x 2
(4) y (x) + 2y(x)y (x) − y(x) = sinh(2x), y(0) = 0, y (0) = 1,
π
(5) y (x) + cos(y(x)) = 0, y(0) = 0, y (0) = ,
2
8
(6) y (x) + y (x) + 2y(x) = −4y(x) ln(y(x)), y(0) = 1, y (0) = 0,

x
(7) y (x) + y (x)2 + y (x)2 = 1 − cos(x), y(0) = y (0) = 0, y (0) = 1.
Exercise 2.2 Given the following IVP for the Emden-Fowler ODE
2
y (x) + y (x) + αx m y(x)μ = 0, y(0) = 1, y (0) = 0.
x
Approximate the solution of this IVP by the VIM and the ADM.
Exercise 2.3 Given the following IVP for the Emden-Fowler ODE
2
y (x) + y (x) + αx m e y(x) = 0, y(0) = y (0) = 0.
x
Approximate the solution of this IVP by the VIM and the ADM.
Chapter 3
Further Analytical Approximation Methods
and Some Applications
3.1 Perturbation Method
In this chapter four additional analytical approximation methods are presented,

namely the perturbation method, the energy balance method, the Hamiltonian
approach and the homotopy analysis method. After a short description of their the-
oretical background, we show exemplary how these methods can be used to solve
problems from the applied physics and engineering.
3.1.1 Theoretical Background
Let us start with the perturbation method (PM). This method is closely related to
techniques used in numerical analysis (see Chap. 4). The corresponding perturbation
theory comprises mathematical methods for finding an approximate solution to a
problem which cannot be solved analytically. The idea is to start with a simplified
problem for which a mathematical solution is known or can be determined by math-
ematical standard techniques. Then, an additional perturbation term, which depends
on a small parameter ε is added to the simplified problem. The parameter ε can appear
naturally in the original problem or is introduced artificially. If the perturbation is
not too large, the various quantities associated with the perturbed system can be
expressed as corrections to those of the simplified system. These corrections, being
small compared to the size of the quantities themselves, are now calculated using
approximate methods such as asymptotic series. The original problem can therefore
be studied based on the knowledge of the simpler one.
To be more precise, let us assume that an ODE, which is subject to initial or
boundary conditions, is given. The essential precondition of the perturbation method
is that the solution of the ODE can be expanded in ε as an infinite sum

DOI 10.1007/978-81-322-2812-7_3
62 3 Further Analytical Approximation Methods and Some Applications

n
y(x, ε) = εk yk (x) + O(εn+1 ). (3.1)
k=0
where y0 , . . . , yn are independent of ε, and y0 (x) is the solution of the problem for
ε = 0, for which we assume that it can be solved very easily.
Then, the perturbation method is based on the following four steps:
• Step 1: Substitute the expansion (3.1) into the ODE and the corresponding initial
(boundary) conditions.
• Step 2: Equate the successive terms of ε, ε, . . . , εn to zero.
• Step 3: Solve the sequence of equations that arise in Step 2.
• Step 4: Substitute y0 (x), y1 (x), . . . , yn (x) into (3.1).
To understand conceptually the perturbation technique, we will consider the follow-
ing IVP:
y (x) = y(x) + εy(x)2 , y(0) = 1, (3.2)
where ε is a small (positive) real parameter. This is a Bernoulli equation, and one
can show that its exact solution is
ex
y(x) = .
1 − ε (e x − 1)
If we use the formula of the geometric series
(1 − z)−1 = 1 + z + z 2 + · · · , |z| < 1,
the solution can be written in the form

ex −1
y(x) = = e x 1 − ε(e x − 1)
1 − ε(e − 1)
x
2
(3.3)
= e x 1 + ε e x − 1 + ε2 e x − 1 + O(ε3 )

= e x + ε e2x − e x + ε2 e3x − 2e2x + e x + O(ε3 ),
where it has to be assumed that |ε(e x − 1)| < 1.

Now, we will show that the expansion (3.3) can be obtained by the perturbation
method without knowledge of the exact solution. Let us proceed as described above.
• Step 1: We substitute the ansatz
y(x; ε) = y0 (x) + εy1 (x) + ε2 y2 (x) + O(ε3 )
into (3.2), and obtain

3.1 Perturbation Method 63
y0 (x) + εy1 (x) + ε2 y2 (x) + O(ε3 )

= y0 (x) + εy1 (x) + ε2 y2 (x) + O(ε3 )
2
+ ε y0 (x) + εy1 (x) + ε2 y2 (x) + O(ε3 ) ,
y0 (0) + εy1 (0) + ε2 y2 (0) + O(ε3 ) = 1.
• Step 2: We equate the successive terms of ε0 , ε1 , and ε2 to zero. This yields
ε0 : y0 (x) = y0 (x), y0 (0) = 1,

ε :
1
y1 (x) = y1 (x) + y0 (x)2 , y1 (0) = 0,
ε :
2
y2 (x) = y2 (x) + 2y0 (x)y1 (x), y2 (0) = 0.
• Step 3: We solve the three IVPs determined in Step 2, and get
y0 (x) = e x , y1 (x) = e2x − e x , y2 (x) = e3x − 2e2x + e x .
• Step 4: We substitute the above three functions y0 (x), y1 (x), and y2 (x) into the
ansatz for y(x; ε) given in Step 1, and obtain

y(x; ε) = e x + ε e2x − e x + ε2 e3x − 2e2x + e x + O(ε3 ).
It can be seen that the solution generated by the perturbation method is identical
to (3.3).
3.1.2 Application of the Perturbation Method
Let us now consider two examples by which the application of the PM to nonlinear
problems is demonstrated.
Example 3.1 In dynamics, the van der Pol oscillator is a non-conservative oscillator
with nonlinear damping. The mathematical model of this dynamical system is an
IVP for a second-order ODE given by

y (x) − ε 1 − y(x)2 y (x) + y(x) = 0, y(0) = A, y (0) = 0, (3.4)
where y is the position coordinate, which is a function of the time x, and ε is a real
parameter indicating the nonlinearity and the strength of the damping. Moreover, y
is the velocity and y is the acceleration.
Approximate the solution of the IVP (3.4) with the PM.
Solution. When ε = 0, i.e., there is no damping function, the ODE becomes y (x) +
y(x) = 0. This is the form of a simple harmonic oscillator, which indicates there
is always conservation of energy. The general solution of this equation is y(x) =

a cos(x + b), where a and b are constants. Imposing the initial conditions y(0) = A
and y (0) = 0, we get y(x) = A cos(x).
To determine an improved approximation of the solution of the given problem
(3.4), we use the perturbation expansion (3.1). Substituting this expansion into (3.4),
we obtain
[y0 (x) + y0 (x)] + ε[y1 (x) + y1 (x)] + ε2 [y2 (x) + y2 (x)] + · · ·

1 − y0 (x)2 y1 (x)
= ε 1 − y0 (x) y0 (x) + ε
2 2
+ ···
−2y0 (x)y1 (x)y0 (x)
Comparing both sides w.r.t. powers of ε, we get

• coefficient of ε0
y0 (x) + y0 (x) = 0, y0 (0) = A, y0 (0) = 0; (3.5)

y1 (x) + y1 (x) = 1 − y0 (x)2 y0 (x), y1 (0) = y1 (0) = 0; (3.6)

y2 (x) + y2 (x) = 1 − y0 (x)2 y1 (x) − 2y0 (x)y1 (x)y0 (x),
(3.7)
y2 (0) = y2 (0) = 0.
The solution of the IVP (3.5) is y0 (x) = A cos(x). Substituting y0 (x) into (3.6) gives

y1 (x) + y1 (x) = − 1 − A2 cos(x)2 A sin(x),
(3.8)
y1 (0) = y1 (0) = 0.
Using the trigonometric formula
sin(x) + sin(3x)
cos(x)2 sin(x) = ,
4
the IVP (3.8) can be written as
A3 − 4 A 1
y1 (x) + y1 (x) = sin(x) + A3 sin(3x),
4 4 (3.9)
y1 (0) = y1 (0) = 0.
It can be easily seen that the solution of (3.9) is
1 3 A3 − 4 A
y1 (x) = − A sin(3x) − x cos(x).
32 8
Now using y0 (x) and y1 (x), the next function y2 (x) can be determined from (3.7)
in a similar manner. At this point, we have obtained the following approximation of
the exact solution y(x) of the given IVP (3.4)
y(x) ≈ y0 (x) + εy1 (x)

1 A3 − 4 A (3.10)
= A cos(x) + ε − A3 sin(3x) − x cos(x) .
32 8
The approximation (3.10) will be acceptable if ε is very small. Otherwise, more

functions yi (x), i > 1, must be determined. But this task is extremely work-intensive,
and numerical methods should be used instead.
Example 3.2 In Applied Mechanics, bending characterizes the behavior of a slender

structural element subjected to an external load applied perpendicularly to a longitu-
dinal axis of the element (see [23]). The structural element is assumed to be such that
at least one of its dimensions is a small fraction of the other two. When the length is
considerably longer than the width and the thickness, the element is called a beam
(see Fig. 3.1). The classical beam bending theory is an important consideration in
nearly all structural designs and analyses. The following nonlinear ODE describes
the bending of a beam under a statical load (see e.g. [25])
M(x) 3/2
y (x) = 1 + y (x)2 , (3.11)
EI
where y(x) is the beam deflection (vertical displacement), y (x) the correspond-
ing slope and M(x) the bending moment. The constants E and I are the modulus
Fig. 3.1 Schematic

representation of a beam
Fig. 3.2 Beam with two

fixed ends subjected to a
concentrated load
of elasticity and the moment of inertia of the cross section about its neutral axis,
respectively.
In this example, let us consider a beam of the length L with two fixed ends
subjected to a load F, which is concentrated at the middle of the span (see Fig. 3.2).
The governing equations for this problem are given by the following IVP
F(4x − L) 3/2
y (x) = 1 + y (x)2 , y(0) = y (0) = 0. (3.12)
8E I
Approximate the solution of the IVP (3.12) with the PM.
Solution. This IVP does not depend on a parameter. The idea of the PM is to introduce
an artificial parameter ε into the problem by multiplying the nonlinear term with ε.
Thus, we write
F(4x − L) 2 3/2
y (x) = 1 + ε y (x) , y(0) = y (0) = 0.
8E I
Obviously, for ε = 0 we have a linear problem that can be solved very easily, whereas
for ε = 1 we have the given nonlinear problem. For simplicity, we write the ODE
in the following form:
2
F(4x − L) 2 3
(y (x))2 = 1 + ε y (x) , y(0) = y (0) = 0. (3.13)
8E I
Let us assume that the solution of (3.13) can be represented by the power series
(3.1), i.e.,
n
y(x) = εk yk (x) + Rn+1 ,
k=0
where Rn+1 = O(εn+1 ). Substituting the above expansion into the ODE in (3.13),
we get
n
2

εk yk (x) + Rn+1

k=0
⎛
2 ⎞3
2
n
F(4x − L) ⎝1 + ε ⎠ ,
= εk yk (x) + Rn+1

8E I k=0

where Rn+1 = O(εn+1 ) and Rn+1 = O(εn+1 ).
Comparing both sides of this equation w.r.t. powers of ε, we obtain
2 F 2 (4x − L)2
y0 (x) = , y0 (0) = y0 (0) = 0; (3.14)
64(E I )2
3F 2 (4x − L)2 2
2y0 (x)y1 (x) = y0 (x) , y1 (0) = y1 (0) = 0; (3.15)
64(E I )2
2 F 2 (4x − L)2 4
2y0 (x)y2 (x) + y1 (x) = 6y (x)y
(x) + 3 y0 (x) ,
64(E I )2 0 1
(3.16)
y2 (0) = y2 (0) = 0.
Now, we solve successively the (linear) IVPs (3.14)–(3.16). The results are

F 3 2
y0 (x) = x − Lx ,
3
12E I 4
3
3F 8 7 2 6 2 2 5 1 3 4
y1 (x) = x − Lx + L x − L x ,
1024(E I )3 21 3 5 12

15F 5 32 11 8 10 16 2 9
y2 (x) = x − Lx + L x − L3x8
262144(E I ) 55
5 5 9

2 4 7 1 5 6
+ L x − L x .
7 30
Hint: Since we have used in (3.13) the squared ODE (3.12), the IVP (3.14) has
two solutions, which differ only in the sign. Here, we use the solution that fits the
elastomechanical problem.
According to the PM, we may write
y(x) ≈ y0 (x) + εy1 (x) + ε2 y2 (x). (3.17)

If we set ε = 1, and insert the expressions for y0 (x), y1 (x), and y2 (x) into (3.17),
we obtain the following approximation for the given nonlinear IVP:

F 3 2
y(x) ≈ x − Lx3
12E I 4
3
3F 8 7 2 6 2 2 5 1 3 4
+ x − Lx + L x − L x (3.18)
1024(E I )3 21 3 5 12

15F 5 32 11 8 10 16 2 9
+ x − Lx + L x − L3x8
262144(E I ) 55
5 5 9

2 4 7 1 5 6
+ L x − L x .
7 30
This example shows that an additional parameter ε can often be introduced into the
given problem such that for ε = 0 we have a linear problem which can be solved
quite simply, and for ε = 0 the solution of the nonlinear problem can be represented
in form of a power series in ε. Then, setting ε = 1, the truncated power series with
just a few terms is a sufficiently good approximation of the given nonlinear problem.
In Table 3.1, for different parameter sets we present the errors err 1 , err 2 , and err 3
of the one-term approximation y0 (x), two-term approximation y0 (x) + y1 (x), and
three-term approximation y0 (x) + y1 (x) + y2 (x), respectively.
In the mechanics of materials the deflection of a beam with two fixed ends sub-
jected to a concentrated load at the middle of the span is often computed by the
one-term approximation

F 3 2
y(x) = x3 − Lx .
12E I 4
With our studies we could show that this formula is justified in most cases.
The perturbation method has been developed for the solution of ODEs with a
nonperiodic solution (static systems). As we could see, the method works very well
in these cases. However, for problems with a periodic solution (dynamical systems),
the perturbation method is divergent for large values of x. The divergence is caused
by terms x cos(x) and/or x sin(x), which will appear in the expansion (3.1). To
Table 3.1 Comparison of the accuracy of the n-term approximations, n = 1, 2, 3

L EI F L EI F L EI F
1 500 100 3 500 100 5 500 100
err 1 7.3 e-7 1.6 e-3 7.1 e-2
err 2 2.2 e-10 4.0 e-5 1.4 e-2
err 3 8.4 e-14 3.1 e-3 2.5 e-1
overcome this problem, the independent variable x is also expanded into a power
series in ε as follows:
x = X (ε, χ) = χ + εX 1 (χ) + ε2 X 2 (χ) + · · · (3.19)
Then, the ansatz (3.19) is substituted into (3.1). By an appropriate selection of

X 1 (χ), X 2 (χ), . . . , the disturbing terms x cos(x), x sin(x) can be eliminated. In
the following example, the modified approach is demonstrated.
Example 3.3 Determine an approximate solution of the following IVP by the per-
turbation method
y (x) + y(x) + εy(x)3 = 0, y(0) = 1, y (0) = 0. (3.20)
This ODE is a special case of the so-called Duffing equation. It models the free
vibration of a mass, which is connected to two nonlinear springs on a frictionless
contact surface (see Example 3.5).
Solution. If we expand the solution y(x) of the IVP (3.20) in the form
y(x; ε) = y0 (x) + εy1 (x) + O(ε2 ),
the following IVPs have to be solved:
y0 (x) = −y0 (x), y0 (0) = 1, y0 (x) = 0,

y1 (x) = −y1 (x) − y0 (x)3 , y1 (0) = 0, y1 (0) = 0.
It can be easily seen that the corresponding solutions are
y0 (x) = cos(x),
1
y1 (x) = cos(3x) − cos(x) − 12x sin(x) .
32
Thus,
ε
y(x; ε) = cos(x) + cos(3x) − cos(x) − 12x sin(x) + O(ε2 )
32

x2 x4 ε 9x 2 81x 4
= 1− + − ··· + 1− + − ···
2! 4! 32 2! 4!

x2 x4 x3 x5
− 1− + − · · · − 12x x − + + ··· + O(ε2 ).
2! 4! 3! 5!
(3.21)
Because of the term −12x sin(x), this expansion is not uniform for x ≥ 0. Therefore,
we substitute the expansion (3.19) into (3.21), and obtain

(χ + εX 1 + O(ε2 ))2 (χ + εX 1 + O(ε2 ))4
y(x; ε) = 1 − + − ···
2! 4!

ε 9(χ + εX 1 + O(ε2 ))2
+ 1−
32 2!

81(χ + εX 1 + O(ε2 ))4
+ − ···
4!

(χ + εX 1 + O(ε2 ))2 (χ + εX 1 + O(ε2 ))4
− 1− + − ···
2! 4!

χ + εX 1 + O(ε 2
)
− 12 χ + εX 1 + O(ε2 )
1!

(χ + εX 1 + O(ε ))2 3
(χ + εX 1 + O(ε2 ))5
− + + ···
3! 5!
+ O(ε2 ).
It follows

χ2 χ4
y(x; ε) = 1 − + − ···
2! 4!

χ3 χ5 χ7
− εX 1 χ − + − + ···
3! 5! 7!

ε 9χ 2
81χ 4
χ2 χ4
+ 1− + − ··· − 1 − + − ···
32 2! 4! 2! 4!

χ χ 3
χ 5
− 12χ − + − ··· + O(ε2 ).
1! 3! 5!
ε
= cos(χ) − εX 1 sin(χ) + cos(3χ) − cos(χ) − 12χ sin(χ) + O(ε2 )
32
ε
= cos(χ) + cos(3χ) − cos(χ) − (12χ + 32X 1 ) sin(χ) + O(ε2 ).
32
Since X 1 can be freely chosen, we eliminate the term χ sin(χ) by setting
3
X 1 = X 1 (χ) = − χ.
8
Thus, the approximate solution can be written in the form
ε
y(x; ε) = Y (χ; ε) = cos(χ) + cos(3χ) − cos(χ) + O(ε2 ), (3.22)
32

3
where x = 1 − ε χ + O(ε2 ).
8
In the next section, we consider the Duffing equation once more and obtain a
better approximation.
3.2 Energy Balance Method
Nonlinear oscillator models have been widely used in many areas of physics and
engineering and are of significant importance in mechanical and structural dynamics
for the comprehensive understanding and accurate prediction of motion. One impor-
tant method that can be used to study such models is the energy balance method
(EBM). This method converges very rapidly to the exact solution and can be easily
extended to nonlinear oscillations. Briefly speaking, the EBM is characterized by an
extended scope of applicability, simplicity, flexibility in applications, and it avoids
any complicated numerical and analytical integration compared with the methods
presented in the previous chapter.
In order to introduce the EBM, let us consider the motion of a general oscillator,
which is modeled by the following IVP:
y (x) + f (y(x)) = 0, y(0) = A, y (0) = 0, (3.23)
where A is the initial amplitude. The characteristic of this problem is the periodicity
of the solution. Its variational can be written as
T /4
1 2
J (y) ≡ − y (x) + F(y) d x, (3.24)
0 2
where T = 2π/ω is the period of the nonlinear oscillation, ω ≥ 0 is the angular

frequency, and d F/dy = f (y).
The Hamiltonian of (3.24) can be written in the form
1 2
H (y) = y (x) + F(y), (3.25)
2
where E(y) ≡ (y (x))2 /2 is the kinetic energy, and P(y) ≡ F(y) is the potential
energy. Since the system is conservative throughout the oscillation, the total energy
remains unchanged during the motion, i.e., the Hamiltonian of the oscillator must be
a constant value,
.
H (y) = E(y) + P(y) = F(A). (3.26)
Thus, the associated residual satisfies

2
1 dy
R(x) ≡ + F(y) − F(A) = 0. (3.27)
2 dx
For a first-order approximation, we use the following trial function
y (1) (x) = A cos(ωx), (3.28)
where ω is an approximation of the frequency of the system. The choice of the trail
function depends on the given initial or boundary conditions. Sometimes this choice
may be y (1) (x) = A sin(ωx), or a combination of cos(ωx) and sin(ωx). Substituting
(3.28) into R(x) yields
1 2 2
R(x) = A ω sin(ωx)2 + F(A cos(ωx)) − F(A). (3.29)
2
If, by chance, the exact solution had been chosen as the trial function, then it would be
possible to make R zero for all values of x by appropriate choice of ω. Since (3.28) is
only an approximation of the exact solution, R cannot be made zero for all x. The idea
is to force the residual to zero, in an average sense. Various formulations have been
proposed in the literature, for example, the least square method, the Ritz–Galerkin
method and the collocation method (see e.g. [14, 35, 94, 135]).
Let us start with collocation at ωx = π/4. The first step is to substitute x = π/(4ω)
into R(x). We obtain
1 2 2 π 2 π
R(x) = A ω sin + F A cos − F(A)
2 4 4
2
1 1 A
= A2 ω 2 √ +F √ − F(A).
2 2 2
.
Setting R(x) = 0, the following quadratic equation for ωC(1) results

A2 (1) 2 A
ωC = F(A) − F √ .
4 2
The corresponding positive solution is

2 A
ωC(1) = F(A) − F √ . (3.30)
A 2
For the period T = 2π/ω we obtain
Aπ
TC(1) = . (3.31)
A
F(A) − F √
2
3.2 Energy Balance Method 73
Substituting ω = ωC(1) into (3.28), the following approximation for the exact solution
y(x) results
(1) 2x A
yC (x) = A cos F(A) − F √ . (3.32)
A 2
Remark 3.4
• In the above formulas, we use the subscript “C” to indicate that the collocation
method has been applied. The superscript “(1)” indicates that the approximation
is first-order. Later we will use the subscript “R” when the Ritz–Galerkin method
is applied.
• The advantage of the EBM is that it does not require a linearization or a small
perturbation parameter. However, the disadvantage of this method is that it is
restricted to dynamical systems (with periodic solutions). We also cannot use this
method for dynamical systems where the independent variable x is multiplied by
the dependent variable.
It is also possible to use a more general ansatz for the trial function than (3.28).
In [35], the following trial function is proposed:
y (n) (x) = A1 w1 (x) + A2 w2 (x) + · · · + An wn (x), (3.33)
where the functions wk (x) are given by
wk (x) ≡ cos((2k − 1)ωx), k = 1, . . . , n.
Obviously, the first initial condition implies
A = A1 + A2 + · · · + An .
Therefore, one of these parameters can be chosen as a dependent parameter, i.e.,
An = A − A1 − A3 − · · · − An−1 .
If the collocation method is applied again to determine ω and the remaining

A1 , . . . , An−1 , the user must specify n collocation points such that the corresponding
computations are relatively simple. In Example 3.6 this strategy is demonstrated.
Another way to find an approximation of the exact solution of the IVP (3.23) by
the EBM is based on the ansatz (3.33) and the Ritz–Galerkin method. Here, ω and
the parameter A2 , . . . , An are determined as an appropriate solution of the following
system of n nonlinear algebraic equations
T /4
R(x)wk (x)d x = 0, k = 1, 2, . . . , (3.34)
0
where as before, wk (x) = cos((2k − 1)ωx), k = 1, . . . , n. In Example 3.7 we show

an application of this strategy.
3.2.2 Application of the Energy Balance Method
To illustrate the applicability and accuracy of the EBM, we apply it to a nonlinear

dynamical system.
Example 3.5 Consider the free vibration of a mass connected to two nonlinear
springs on a frictionless contact surface (see Fig. 3.3).
The problem can be modeled by a special form of the general Duffing equation
y (x) + δ y (x) + αy(x) + β y(x)3 = γ cos(ωx). (3.35)
Here, we set δ = γ = 0, and consider the following IVP
y (x) + αy(x) + β y(x)3 = 0, y(0) = A, y (0) = 0. (3.36)
In this problem, y(x) is the displacement at time x, y (x) is the corresponding

acceleration, β is a small positive parameter, which controls the amount of the non-
linearity in the restoring force, and
K1 + K2
α≡ , (3.37)
m
where the constants K 1 and K 2 are the stiffnesses of the nonlinear springs, and m is
the mass of the system.
Determine a first-order approximation of the exact solution of the IVP (3.36) by
the EBM using the collocation technique.
Solution. For this dynamical system, we have
1 1
f (y(x)) = αy(x) + β y(x)3 , F(y(x)) = αy(x)2 + β y(x)4 .
2 4
Fig. 3.3 A mass connected

to two nonlinear springs
Its variational is
T /4
1 1 1
J (y) = − y (x)2 + αy(x)2 + β y(x)4 d x.
0 2 2 4
The corresponding Hamiltonian must satisfy
1 2 1 1 . 1 1
H (y) = y (x) + αy(x)2 + β y(x)4 = α A2 + β A4 .
2 2 4 2 4
Thus, the residuum R(x) is
1 1 1 1 1
R(x) = y (x)2 + αy(x)2 + β y(x)4 − α A2 − β A4
2 2 4 2 4 (3.38)
1 2 1 1 1 2 1 2
= y (x) + y(x) α + β y(x) − A α + β A .
2 2
2 2 2 2 2
If the exact solution is substituted into the residuum, it holds R(x) = 0.

To determine a first-order approximation of the exact solution, we use the ansatz
(3.28) as a trial function. Substituting this ansatz into (3.38), we obtain

A2 ω 2 A2 A2
R(x) = sin(ωx) +
2
cos(ωx) α +
2
β cos(ωx)2
2 2 2

A2 A2
− α+ β .
2 2
As in the previous section, we apply collocation at ωx = π/4. Using formula (3.30),

we obtain
(1) 3
ωC = α + β A 2 . (3.39)
4
The corresponding period is given by formula (3.31). It is
2π
TC(1) = . (3.40)
3
α + β A2
4
Thus, we get the following first-order approximation of the exact solution of the IVP
(3.36)
(1) 3 2
yC (x) = A cos α + βA x . (3.41)
4
In Table 3.2, for different parameter sets (A, α, β) the corresponding frequencies ωC(1)
are given. To study the accuracy of the approximations yC(1) (x), we have determined
(1)
Table 3.2 Values of ωC determined for different parameter sets
No. A α β ωC(1)
(1) 0.1 1 0.1 1.0003749297
(2) 2 1 0.1 1.1401754251
(3) 2 3 0.2 1.8973665961
(4) 4 3 0.2 2.3237900078
Fig. 3.4 Parameter set (1): second-order approximation yC(2) , and the the error functions eC(1) , eC(2)
(1)
numerical approximations yODE of the IVP (3.36) by the IVP-solver ODE45, which
is part of the Matlab. Let us denote the difference between these solutions by

(1)
eC(1) (x) ≡ yODE − yC(1) (x) .
In the Figs. 3.4 and 3.5, the first-order approximation yC(1) (x) and the correspond-
ing eC(1) (x) can be seen for the parameter sets (1) and (4), respectively.
The plots show that the first-order approximation yC(1) (x) is already quite accurate.

Example 3.6 Determine a second-order approximation of the exact solution of the

IVP (3.36) by the EBM using the collocation technique.
Solution. Looking at formula (3.33), we set n = 2 and use the functions w1 (x) =
cos(ωx) and w2 (x) = cos(3ωx). Thus, the appropriate ansatz is
y (2) (x) = A1 cos(ωx) + (A − A1 ) cos(3ωx). (3.42)

(2) (1) (2)

Fig. 3.5 Parameter set (4): second-order approximation yC , and the the error functions eC , eC
It follows
d (2)
y (x) = −A1 ω sin(ωx) − 3(A − A1 )ω sin(3ωx). (3.43)
dx
Let the collocation points be ωx = π/4 and ωx = π/2.
For ωx = π/4 we obtain
π 1 1 1
y (2) = A1 √ − (A − A1 ) √ = √ (2 A1 − A),
4ω 2 2 2
π 2 1
y (2) = (2 A1 − A)2 ,
4ω 2
and

d (2) π 1 3 2ω 3
y = − √ A1 ω − √ ω(A − A − 1) = √ A1 − A ,
dx 4ω 2 2 2 2
2 2
d (2) π 3
y = 2ω 2 A1 − A .
dx 4ω 2
Thus,
π 2
3 1 1
R =ω A1 − A + (2 A1 − A) α + β(2 A1 − A)
2 2 2
4ω 2 4 4
(3.44)
1 1
− A2 α + β A2 .
2 2
For ωx = π/2 we obtain

π π 2
y (2) = 0, y (2) = 0,
2ω 2ω
and
d (2) π
y = −A1 ω + 3(A − A1 )ω = ω(3A − 4 A1 ),
dx 2ω

d (2) π 2
y = ω 2 (3A − 4 A1 )2 .
dx 2ω
Thus,
π 1 1 2 1 2
R = ω (3A − 4 A1 ) − A α + β A .
2 2
(3.45)
2ω 2 2 2
From (3.44) and (3.45), we get the following two nonlinear algebraic equations
for the two unknowns A1 and ω:
2
3 1
0 = 4ω A1 − A + (2 A1 − A) α + β(2 A1 − A)
2 2 2
2 4

1
− 2 A2 α + β A2 , (3.46)
2

1 2
0 = ω (3A − 4 A1 ) − A α + β A .
2 2 2
2
The determination of the solutions of this system in closed form leads to very large
and complicated expressions. A better strategy is to prescribe fixed values for A,
α, and β, and to solve numerically the algebraic equations (3.46). In Table 3.3, for
the parameter sets (1) and (4) presented in Table 3.2, the numerically determined
appropriate solutions of the nonlinear algebraic equations (3.46) are given.
Note, there are 12 solutions of the system (3.46). Since 4 solutions are complex-
valued, and for 4 solutions the ω-component is negative, there still remain 4 real-
valued solutions with a positive ω-component. We have chosen the solution whose
component ωC(2) is nearest to ωC(1) . For the parameter sets (1) and (4), the corresponding
solution yC(2) as well as the error functions eC(1) and eC(2) are presented in the Figs. 3.4
and 3.5, respectively.
(2) (2)
Table 3.3 Values of ωC and (A1 )C determined for different parameter sets
(2) (2)
No. A α β ωC (A1 )C
(1) 0.1 1 0.1 1.0003749102 0.0999968776
(4) 4 3 0.2 2.3090040391 3.9288684743
It is easy to see that in both parameter constellations the second-order approxi-

mation is better than the first-order approximation. However, the difference between
these approximations in not significant.
Example 3.7 Determine a first-order approximation of the exact solution of the IVP
(3.36) by the EBM using the Ritz–Galerkin technique.
Solution. We start with the ansatz (3.28), i.e.,
y (1) = A cos(ωx),
and obtain
A2 2 A2 A4 A2 A4
R(x) = ω sin(ωx)2 + α cos(ωx)2 + β cos(ωx)4 − α− β.
2 2 4 2 4
The next step is to set the weighted integral of the residual to zero, i.e.,
T /4
. 2π
R(x) cos(ωx)d x = 0, T = .
0 ω
This yields
A2 A2 2 A4 A2 A4 7 2
ω+ α+ β− α− β = ω2 − α − A β = 0.
6 3ω 15ω 2ω 4ω 10
The positive solution of the equation
7 2
ω2 = α + A β (3.47)
10
is
7 2
ωR(1) = α+ A β. (3.48)
10
Thus, the first-order approximate solution of the IVP (3.36) is

7 2
yR(1) = A cos α+ A βx . (3.49)
10
For the parameter sets (1)–(4) given in Table 3.2, the approximated frequencies are
presented in Table 3.4.

IVP (3.36) by the EBM using the Ritz–Galerkin technique.
(1)
Table 3.4 Values of ωR determined for different parameter sets
No. A α β ωR(1)
(1) 0.1 1 0.1 1.0003499388
(2) 2 1 0.1 1.1313708499
(3) 2 3 0.2 1.8867962264
(4) 4 3 0.2 2.2891046285
Solution. Substituting the ansatz
y (2) = A1 cos(ωx) + (A − A1 ) cos(3ωx)
into the residual (3.38), we obtain
ω2 2
R(x) = A1 sin(ωx) + 3(A − A1 ) sin(3ωx)
2
α 2
+ A1 cos(ωx) + (A − A1 ) cos(3ωx) (3.50)
2
β 4 1 1
+ A1 cos(ωx) + (A − A1 ) cos(3ωx) − α A2 − β A4 .
4 2 4
Now, from (3.34) we get the following two nonlinear algebraic equations for the two
unknowns A1 and ω: T /4
R(x) cos(ωx)d x = 0,
0
T /4
(3.51)
R(x) cos(3ωx)d x = 0,
0
where T = 2π/ω.
In the Table 3.5, for the parameter sets (1) and (4) presented in Table 3.2, the
numerically determined appropriate solutions of the nonlinear algebraic equations
(3.51) are given. Moreover, in the Figs. 3.6 and 3.7 the accuracy of the first-order
approximations of the EBM (collocation and Ritz–Galerkin) and the second-order
approximations (collocation and Ritz–Galerkin) are visualized for the same two
parameter sets.
(2) (2)
Table 3.5 Values of ωR and (A1 )R determined for different parameter sets
(2) (2)
No. A α β ωR (A1 )R
(1) 0.1 1 0.1 1.0003749144 0.0999968773
(4) 4 3 0.2 2.3114595536 3.9249208620
Fig. 3.6 Parameter set (1): Comparison of the accuracy of the approximations, which have been
computed with the EBM (using collocation) and the EBM (using Ritz–Galerkin)
Fig. 3.7 Parameter set (4): Comparison of the accuracy of the approximations, which have been
computed with the EBM (using collocation) and the EBM (using Ritz–Galerkin)
Example 3.9 Consider the free vibration of a mass attached to the center of a
stretched elastic wire. The governing IVP corresponding to this dynamical system is
given by (see [36, 118] and Fig. 3.8)
y(x)
y (x) + y(x) − λ = 0, y(0) = A, y (0) = 0, (3.52)
1 + y(x)2
Fig. 3.8 Oscillation of a

mass attached to the center
of a stretched elastic wire
where 0 < λ ≤ 1 and A are real parameters.

Determine a first-order approximation of the exact solution of the IVP (3.52) by
the EBM using the collocation technique.
Solution. For this dynamical system, we have
y(x) 1
f (y(x)) = y(x) − λ , F(y(x)) = y(x)2 − λ 1 + y(x)2 .
1 + y(x)2 2
Its variational is
T /4
1 1
J (y) = − y (x)2 + y(x)2 − λ 1 + y(x)2 d x.
0 2 2
The corresponding Hamiltonian must satisfy
1 2 1 . 1
H (y) = y (x) + y(x)2 − λ 1 + y(x)2 = A2 − λ 1 + A2 .
2 2 2
Thus, the residuum R(x) is
1 2 1 1
R(x) = y (x) + y(x)2 − λ 1 + y(x)2 − A2 + λ 1 + A2 . (3.53)
2 2 2
To determine a first-order approximation of the exact solution, we use the ansatz
(3.28) as a trial function. Substituting this ansatz into R(x), we obtain
1 2 2 1
R(x) = A ω sin(ωx)2 + A2 cos(ωx)2
2 2
1
− λ 1 + A2 cos(ωx)2 − A2 + λ 1 + A2 .
2
√
Using the collocation point ωx = π/4, we get sin(π/4) = cos(π/4) = 1/ 2 and

1 1 1 2
R(x) = A2 ω 2 − A2 − λ 1 + A − 1 + A2 .
4 4 2
Now, we set R(x) = 0. The positive solution of the resulting quadratic equation

4λ 1
ω −1− 2
2
1 + A2 − 1 + A2 =0
A 2
is

4λ 1 2
ωC(1) = 1+ 2 1 + A − 1 + A2 . (3.54)
A 2
For the corresponding period, we obtain
2π 2π
TC(1) = =
. (3.55)
ω
4λ 1
1 + 1 + A2 − 1 + A2
A2 2
Thus, we get the following first-order approximation of the exact solution of the IVP
(3.52)
⎛

⎞
4λ 1
yC (x) = A cos ⎝1 + 2
(1)
1 + A2 − 1 + A2 x ⎠ . (3.56)
A 2
In comparison, the exact frequency is (see [118])
2π
ωex = , (3.57)
Tex
where
−1/2
π/2
2λ
Tex = 4 1− √ d x. (3.58)
0 1 + A2 sin(x)2 + 1 + A2
In Table 3.6, for different values of λ and A the approximate frequency ωC(1) and the
exact frequency ωex are given.

IVP (3.52) by the EBM using the collocation technique.
(1)
Table 3.6 Values of ωC and ωex determined for different parameter sets
No. λ A ωex ωC(1)
(1) 0.2 2 0.9467768213 0.94825975661
(2) 0.5 2 0.8604467698 0.8648649692
(3) 0.7 1 0.6816273394 0.6851916996
(4) 0.9 1 0.5566810704 0.5638374876
Solution. As described in Example 3.6, we use the ansatz (3.42), i.e.,
y (2) (x) = A1 cos(ωx) + (A − A1 ) cos(3ωx).
Substituting (3.42) into the residual (3.53), we obtain
ω2 2
2
1 2
+ A1 cos(ωx) + (A − A1 ) cos(3ωx)
2
2
− λ 1 + A1 cos(ωx) + (A − A1 ) cos(3ωx)
1
− A2 + λ 1 + A2 .
2
Now, we use the two collocation points ωx = π/4 and ωx = π/2, and set
π π
R = 0, R = 0.
4ω 2ω
It results the following system of two nonlinear algebraic equations for the two
unknown ω and A1
⎛ ⎞
A 2
0 = 4λ ⎝ 1 + A2 − 1 + − 2 A A1 + 2 A21 ⎠
2
(3.59)
+ ω 2 9A2 + 4 A21 − 12 A A1 + 4 A1 (A1 − A) − A2 ,

0 = 2λ 1 + A2 − 1 + ω 2 (3A − 4 A1 )2 − A2 .
This system has 16 solutions. For 8 of these solutions, the ω-component is complex-
valued. Thus, there remain 8 solutions with a real-valued ω-component. We have
chosen the solution whose component ωC(2) is nearest to ωC(1) . For the parameter sets
(1)–(4) presented in Table 3.6, the corresponding solutions are given in Table 3.7. A
comparison with the results in Table 3.6 shows that the second-order approximation
ωC(2) is slightly better than the first-order approximation ωC(1) .
(2) (2)
Table 3.7 Values of ωC and (A1 )C determined for different parameter sets
No. λ A ωC(2) (2)
(A1 )C
(1) 0.2 2 0.9481247885 1.9936896281
(2) 0.5 2 0.8637698322 1.9811778812
(3) 0.7 1 0.6832313703 0.9871641681
(4) 0.9 1 0.5581704732 0.9759149690
Example 3.11 Determine a first-order approximation of the exact solution of the

IVP (3.52) by the EBM using the Ritz–Galerkin technique.
Solution. We start with the ansatz (3.28), i.e.,
y (1) = A cos(ωx),
and obtain
1 2 2 1
R(x) = A ω sin(ωx)2 + A2 cos(ωx)2
2 2
1
− λ 1 + A2 cos(ωx)2 − A2 + λ 1 + A2 .
2
The next step is to set the weighted integral of the residual to zero, i.e.,
T /4
. 2π
R(x) cos(ωx)d x = 0, T = .
0 ω
This yields

A
(1 + A2 ) arcsin √ +A
1 2 2 1 1 2 2 1 + A2
0= A ω + A −λ
2 3ω 2 3ω 2 Aω
1 21 1
− A +λ 1+ A 2
2 ω ω
3λ A
= ω − 1 − 3 (1 + A ) arcsin √
2 2
+A
A 1 + A2
6λ
+ 2 1 + A2 .
A
The positive solution of this equation is

6λ 3λ A
ωR(1) = 1− 1 + A 2+ (1 + A 2
) arcsin √ + A . (3.60)
A2 A3 1 + A2
(1)
Table 3.8 Values of ωR and ωex determined for different parameter sets
No. λ A ωex ωR(1)
(1) 0.2 2 0.9467768213 0.9457062842
(2) 0.5 2 0.8604467698 0.8578466878
(3) 0.7 1 0.6816273394 0.6774771762
(4) 0.9 1 0.5566810704 0.5517217102
Thus,
yR(1) (x) = A cos ωR(1) x . (3.61)
In Table 3.8, for the parameter sets (1)–(4) of the previous two tables the approximate
frequency ωR(1) and the exact frequency ωex are given.

IVP (3.52) by the EBM using the Ritz-Galerkin technique.
Solution. Substituting the ansatz
y (2) = A1 cos(ωx) + (A − A1 ) cos(3ωx)
into the residual (3.53), we obtain
ω2 2
2
1 2
+ A1 cos(ωx) + (A − A1 ) cos(3ωx)
2
2
− λ 1 + A1 cos(ωx) + (A − A1 ) cos(3ωx)
1
− A2 + λ 1 + A2 .
2
Now, from (3.34) we get the following two nonlinear algebraic equations for the two
unknowns A1 and ω: T /4
R(x) cos(ωx)d x = 0,
0
T /4
(3.62)
R(x) cos(3ωx)d x = 0,
0
where T = 2π/ω.
In Table 3.9, for the parameter sets (1)–(4) presented in the previous three tables,
the numerically determined appropriate solutions of the nonlinear algebraic equations
(3.62) are given.
3.3 Hamiltonian Approach 87
(2) (2)
Table 3.9 Values of ωR and (A1 )R determined for different parameter sets
No. λ A ωR(2) (2)
(A1 )R
(1) 0.2 2 0.9474385576 1.9952077413
(2) 0.5 2 0.8621619690 1.9854583530
(3) 0.7 1 0.6824425348 0.9882819868
(4) 0.9 1 0.5575093394 0.9772737444
3.3 Hamiltonian Approach
As we mentioned before, many oscillation problems are nonlinear, and in most cases
it is difficult to solve them analytically. In the previous section, we have introduced
the energy balance method using the collocation technique and the Ritz technique.
Both strategies are based on the Hamiltonian. Although this approach is simple, it
strongly depends upon the location points (collocation technique) and the weight
functions w j (x) (Ritz technique) that are chosen. Recently, He (see e.g. [52, 53])
has proposed the so-called Hamiltonian approach to overcome the shortcomings of
the EBM.
In order to describe the Hamiltonian approach (HA), let us consider once more
the IVP (3.23) of the general oscillator, i.e.,
y (x) + f (y(x)) = 0, y(0) = A, y (0) = 0.
The corresponding variational and the Hamiltonian of the oscillator are given in
(3.24) and (3.25), respectively. As shown in the previous section, the Hamiltonian
must be a constant value. Thus, we can write
1 2 .
H (y(x)) = y (x) + F(y(x)) = H0 , (3.63)
2
where H0 is a real constant. Now, it is assumed that the solution can be expressed as
y(x) = A cos(ωx). (3.64)
Substituting (3.64) into (3.63) yields
1 2 2
H (y(x)) = A ω sin(ωx)2 + F(A cos(ωx)) = H0 . (3.65)
2
Differentiating both sides of the identity (3.65) gives
∂ H (y)
= 0. (3.66)
∂A
The next step is to introduce a new function H̄ (y) defined as

T /4
1 2
H̄ (y) ≡ y (x) + F(y(x)) d x
0 2
T /4 (3.67)
T T
= H0 d x = H0 = H (y).
0 4 4
It holds
∂ H̄ (y) 1
= H (y).
∂T 4
Thus, Eq. (3.66) is equivalent to
∂ 2 H̄ (y)
= 0,
∂ A ∂T
or
∂ 2 H̄ (y)
= 0. (3.68)
∂ A ∂(1/ω)
This formula can be used to determine an approximate frequency of the nonlinear

oscillator. Substituting this approximate frequency into the ansatz (3.64), a first-order
approximation of the solution of the IVP (3.23) is obtained.
3.3.2 Application of the Hamiltonian Approach
In this section, by means of two instructive examples we will show the reliability of
the Hamiltonian approach.
Example 3.13 Let us consider once again the special case (3.36) of the Duffing ODE
(3.35)
y (x) + αy(x) + β y(x)3 = 0, y(0) = A, y (0) = 0. (3.69)
Determine an analytical approximation of the solution of the IVP (3.69) by the

Hamiltonian approach.
Solution. For this problem, we have
f (y) = αy + β y 3 .
Thus,
1 2 1 4
F(y) = αy + β y .
2 4
The corresponding Hamiltonian is
1 2 1 1
H (y(x)) = y (x) + αy(x)2 + β y(x)4 ,
2 2 4
and
T /4
1 2 1 1
H̄ (y) = y (x) + αy(x) + β y(x) d x.
2 4
(3.70)
0 2 2 4
Now, we use the following ansatz for an approximate solution of the IVP (3.69)
y(x) = A cos(ωx).
Substituting this ansatz into (3.70), we obtain

T /4
1 2 2 1 2 1 4
H̄ (y) = A ω sin(ωx) + α A cos(ωx) + β A cos(ωx) d x
2 2 4
0 2 2 4
π
2ω 1 2 2 1 2 1 4
= A ω sin(ωx) + α A cos(ωx) + β A cos(ωx) d x
2 2 4
0 2 2 4
1 2 2 1π 1 2 1π 1 4 3 π
= A ω · + αA · + βA ·
2 4ω 2 4ω 4 16 ω
1 2 1 1 2 3
= A ωπ + α A π + β A4 π .
8 ω 8 64
It follows
∂ H̄ (y) 1 1 3
= − A2 ω 2 π + α A2 π + β A4 π,
∂(1/ω) 8 8 64
and
∂ 2 H̄ (y) 1 1 3
= − Aω 2 π + α Aπ + β A3 π.
∂ A ∂(1/ω) 4 4 16
We set
∂ 2 H̄ (y) .
= 0,
∂ A ∂(1/ω)
i.e.,
3
−ω 2 + α + β A2 = 0.
4
The positive root of this quadratic equation is a first-order approximation of the
frequency of the nonlinear oscillator:

(1) 3
ωHA = α + β A2 . (3.71)
4
A comparison with formula (3.39) shows that this is the same approximation as that
obtained by the energy balance method using the collocation point x = π/(4ω).
Obviously, the corresponding approximation of the solution of the IVP (3.69) is

(1) 3 2
yHA (x) = yC(1) (x) = A cos α + βA x . (3.72)
4

Note, that the Hamiltonian approach has required less computational work than the
energy balance method.
Example 3.14 As a second example, let us consider the IVP of the Duffing equation
with a fifth-order nonlinearity given by
y (x) + y(x) + αy(x)5 = 0, y (0) = A, y (0) = 0, (3.73)
where α ≥ 0 is a real constant. Determine an approximative solution of the IVP

(3.73) by the Hamiltonian approach.
Solution. For this problem, we have
f (y) = y + αy 5 .
Thus,
1 2 1 6
F(y) = y + αy .
2 6
The corresponding Hamiltonian is
1 2 1 1
H (y(x)) = y (x) + y(x)2 + αy(x)6 .
2 2 6
Integrating this Hamiltonian w.r.t. x from 0 to T /4, gives
T /4
1 2 1 1
H̄ (y) = y (x) + y(x) + αy(x) d x.
2 6
(3.74)
0 2 2 6
Using the ansatz

y(x) = A cos(ωx)
for the approximate solution of the IVP (3.73), and substituting it into H̄ (y), we
obtain
π
2ω 1 2 2 1 1
H̄ (y) = A ω sin(ωx)2 + A2 cos(ωx)2 + α A6 cos(ωx)6 d x
0 2 2 6
1 2 2 π 1 2 π 1 6 5π
= A ω · + A · + αA ·
2 4ω 2 4ω 6 32ω
1 2 1 1 2 5
= A πω + A π+ α A6 π .
8 ω 8 192
Now, we compute
∂ H̄ (y) 1 1 5
= − A2 πω 2 + A2 π + α A6 π,
∂(1/ω) 8 8 192
and
∂ 2 H̄ (y) 1 1 5
= − Aπω 2 + Aπ + α A5 π.
∂ A ∂(1/ω) 4 4 32
We set
∂ 2 H̄ (y) .
= 0,
∂ A ∂(1/ω)
i.e.
5
−ω 2 + 1 + α A4 = 0.
8
The positive root of this quadratic equation is a first-order approximation of the
frequency of the nonlinear oscillator:

(1) 5
ωHA = 1 + α A4 .
8
The corresponding approximation of the solution of the IVP (3.73) is

(1) 5 4
yHA (x) = A cos 1 + αA x . (3.75)
8
In Fig. 3.9 the approximate solution (3.75) is compared with the exact solution for
α = 1.0 and different values of A. It can be seen that the approximate solution gets
worse if the parameter A is increased.
(1)
Fig. 3.9 Comparison of yHA (x) with the exact solution for α = 1.0 and A = 1, 2, 5; Note for the
third partial figure we have chosen the x-interval [0, 1] instead of [0, 6]
3.4 Homotopy Analysis Method
The homotopy analysis method (HAM) is a semi-analytical technique to solve nonlin-

ear ODEs. It uses the concept of the homotopy from topology to generate a convergent
series solution for nonlinear problems. This is enabled by employing a homotopy-
McLaurin series to deal with the nonlinearities in the ODE.
The HAM was first developed in 1992 by Liao [75] and further modified in
[76–79]. The advantages of the HAM are
• It is a series expansion method that is not directly dependent on small or large
physical parameters. Thus, it is applicable for strongly nonlinear problems.
3.4 Homotopy Analysis Method 93
• It has a greater generality than the other methods presented in this book, i.e., it often
allows for strong convergence of the solution over larger spacial and parameter
domains.
• It provides great freedom to choose the basis functions of the desired solution and
the corresponding auxiliary linear operator of the homotopy.
• It gives a simple way to ensure the convergence of the solution series.
The essential idea of the HAM is to introduce an embedding parameter p ∈ [0, 1],
and a nonzero auxiliary parameter referred to as the convergence-control parame-
ter, to verify and enforce the convergence of the solution series. For p = 0, the
original problem is usually reduced to a simplified problem that can be solved rather
easily. If p is gradually increased toward 1, the problem goes through a sequence of
deformations, and the solution at each stage is close to that at the previous stage of
deformation. In most cases, for p = 1 the problem takes the original form and the
final stage of the deformation gives the desired solution. In other words, as p increases
from zero to one, the initial guess approaches a good analytical approximation of the
exact solution of the given nonlinear problem.
In [77], Liao has shown that the HAM is a unified method for some perturbation
and non-perturbation methods. For example, the ADM is a special case of the HAM
and for certain values of , the VIM and the HAM are equivalent.
We now turn to the general approach used by the HAM to solve the nonlinear
ODE
N [y(x)] = 0, x ∈ I, (3.76)
subject to some initial or boundary conditions. Here, N is a nonlinear operator, and

y(x) is an unknown function defined over the interval I . The first step in the HAM
solution of this ODE is to construct the homotopy
H [φ(x; p); y0 (x), H(x), , p]

(3.77)
≡ (1 − p)L[φ(x; p) − y0 (x)] − p H(x)N [φ(x; p)],
where = 0 is an auxiliary parameter, H(x) = 0 is an auxiliary function, p ∈

[0, 1] is an embedding parameter, y0 (x) is an initial approximation to the solution
that satisfies the given initial or boundary conditions, φ(x; p) satisfies the initial or
boundary conditions, and L is some linear operator. The linear operator L should
normally be of the same order as the nonlinear operator N .
We now set the homotopy (3.77) equal to zero so that
(1 − p)L[φ(x; p) − y0 (x)] = p H(x)N [φ(x; p)]. (3.78)
The Eq. (3.78) is called zeroth-order deformation equation. By setting p = 0 in this

equation, we get
L[φ(x; 0) − y0 (x)] = 0.
From the definitions of L, φ(x), and y0 (x), it follows
φ(x; 0) = y0 (x). (3.79)
If we set p = 1 in the Eq. (3.78), we obtain
N [φ(x; 1)] = 0. (3.80)
Under the assumption that φ(x; p) satisfies the initial or boundary conditions, for-
mula (3.80) implies
φ(x; 1) = y(x). (3.81)
It is clear from (3.79) and (3.81) that φ(x; p) varies continuously from the initial
approximation y0 (x) to the required solution y(x) as p increases from 0 to 1. In the
topology, φ is called a deformation.
Next, let us define the terms

1 ∂ n φ(x; p)
yn (x) = . (3.82)
n! ∂ p n p=0
By Taylor’s theorem we can write

1 ∂ n φ(x; p)
∞ ∞

φ(x; p) = φ(x; 0) + p n = y0 (x) + yn (x) p n , (3.83)
n=1
n! ∂ pn n=1
p=0
In order to calculate the functions yn (x) in the solution series, we first differentiate
equation (3.78) w.r.t. p. It follows

∂φ(x; p)
(1 − p)L − L[φ(x; p) − y0 (x)]
∂p
(3.84)
∂N [φ(x; p)]
= H(x) N [φ(x; p)] + p H(x) .
∂p
Setting p = 0 in this equation and applying the results of (3.79) and (3.81), we get
L[y1 (x)] = H(x) N [y0 (x)]. (3.85)
Now, we will prove by induction that the following relation is true

n−1
∂ n φ(x; p) ∂ φ(x; p)
(1 − p)L −nL
∂ pn ∂ p n−1
(3.86)
∂ N [φ(x; p)]
n−1
∂ n N [φ(x; p)]
= n H(x) + p H(x) ,
∂ p n−1 ∂ pn
where n ≥ 2. Let us start with n = 2. Differentiating (3.84) w.r.t. p, gives

∂ 2 φ(x; p) ∂φ(x; p)
(1 − p)L −2L
∂ p2 ∂p
(3.87)
∂N [φ(x; p)] ∂ 2 N [φ(x; p)]
= 2 H(x) + p H(x) ,
∂p ∂ p2
which shows that Eq. (3.86) is indeed true for n = 2.

Next, we assume that (3.86) is true for some value of n and show that it must
also be true when n is replaced by n + 1. Since (3.86) is true by assumption, we can
differentiate it w.r.t. p to obtain
n
∂ n+1 φ(x; p) ∂ φ(x; p)
(1 − p)L − (n + 1) L
∂ p n+1 ∂ pn
(3.88)
∂ N [φ(x; p)]
n
∂ n+1 N [φ(x; p)]
= (n + 1) H(x) + p H(x) ,
∂ pn ∂ p n+1
which shows that (3.86) is true if n is replaced by n + 1. Thus, it can be concluded

that Eq. (3.86) is valid for all n ≥ 2.
Dividing (3.86) by n! and setting p = 0, we get

1 ∂ n φ(x; p) 1 ∂ n−1 φ(x; p)
L −
n! ∂ p n (n − 1)! ∂ p n−1 p=0
(3.89)
1 ∂ n−1 N [φ(x; p)]
= H(x) .
(n − 1)! ∂ p n−1 p=0
Using this equation and the functions yn (x) defined in (3.82), we obtain the relation

1 ∂ n−1 N [φ(x; p)]
L[yn (x) − yn−1 (x)] = H(x) , (3.90)
(n − 1)! ∂ p n−1 p=0
which is valid for all n ≥ 2.

Now, we define an additional function

0, n ≤ 1,
χ(n) ≡ (3.91)
1, else.
Then, the Eqs. (3.85) and (3.90) can be written in the form

1 ∂ n−1 N [φ(x; p)]
L[yn (x) − χ(n)yn−1 (x)] = H(x) . (3.92)
(n − 1)! ∂ p n−1 p=0
The linear ODE (3.92) is valid for all n ≥ 1. It is called the nth-order deformation
equation.
Often the abbreviation

1 ∂ n−1 N [φ(x; p)]
Rn (y0 (x), . . . , yn−1 (x)) ≡ (3.93)
(n − 1)! ∂ p n−1 p=0
is used to write the Eq. (3.92) in the short form
L[yn (x) − χ(n)yn−1 (x)] = H(x)Rn (y0 (x), . . . , yn−1 (x)). (3.94)
Obviously, the terms yn (x) can be obtained in order of increasing n by solving the
linear deformation equations in succession.
Since the right-hand side of (3.94) depends only upon the known results y0 (x),
y1 (x), . . . , yn−1 (x), it can be obtained easily using computer algebra software, such
as Maple, MathLab, and/or Mathematica.
The solution of the nth-order deformation equation (3.94) can be written as
yn (x) = ynh (x) + ync (x), (3.95)
where ynh (x) is the general solution of the linear homogeneous ODE
L[ynh (x)] = 0, (3.96)
and ync (x) is a particular solution of (3.94).

It is also possible to express ync (x) as

ync (x) = χ(n) yn−1 (x) + L−1 H(x) Rn (y0 (x), . . . , yn−1 (x) , (3.97)
where L−1 is the inverse of the linear operator L.

Finally, we define the nth partial sum of the terms yn (x) as

n
[n]
y (x) ≡ yk (x). (3.98)
k=0
The solution of the Eq. (3.76) can be written as

∞

y(x) = φ(x; 1) = yk (x) = lim y [n] (x). (3.99)
n→∞
k=0
This solution will be valid wherever the series converges.

For a better understanding of this method, let us consider a simple example.
Example 3.15 Solve the following IVP of a first-order nonlinear ODE by the HAM

y (x) = 1 − y(x)2 , y(0) = 0. (3.100)
Solution. It is not difficult to show that the exact solution is y(x) = sin(x).
As an initial guess that satisfies the initial condition, we use y0 (x) = 0. Moreover,
we set L[y(x)] ≡ y (x) and H(x) ≡ 1. Thus, (3.94) reads
(yn (x) − χ(n) yn−1 (x)) = H(x) Rn (y0 (x), y1 (x), . . . , yn−1 (x)).
Here, we have ynh (x) = cn , where cn is a constant, and

x
ync (x) = χ(n) yn−1 (x) + H(t) Rn y0 (t), . . . , yn−1 (t) dt. (3.101)
0
Thus, yn (x) = cn + ync (x). It is

N [y(x)] = y (x) − 1 − y(x)2 ,
and

2

∞
∞

N [φ(x; p)] = y0 (x) +
yn (x) p n − 1 − y0 (x) + yn (x) p n .
n=1 n=1
Looking at (3.93), we calculate

R1 (y0 (x)) = y0 (x) − 1 − y0 (x)2 = −1.
Using formula (3.101), we obtain

x
y1c (x) = − dt = − x,
0
and y1 (x) = y1h (x) + y1c (x) = c1 − x. Since y1 (x) must satisfy the given initial
condition, it follows c1 = 0. Thus
y1 (x) = −x. (3.102)
In the next step, we determine
y0 (x)y1 (x)
R2 (y0 (x), y1 (x)) = y1 (x) + = −
1 − y0 (x)2
and x
y2c (x) = − x + (−)dt = −(1 + )x.
0
Thus, y2 (x) = c2 − (1 + )x. Since the given initial condition yields c2 = 0, we
obtain
y2 (x) = −(1 + )x. (3.103)
Finally, we present the result of the third step. We have
y0 (x)2 y1 (x)2
R3 (y0 (x), y1 (x), y2 (x)) = y2 (x) +
(1 − y0 (x)2 )3
y1 (x)2 y0 (x)y2 (x)
+ +
2 1 − y0 (x)2 1 − y0 (x)2
2
= − (1 + ) + x 2 ,
2
and
x
2 2
y3c (x) = − (1 + )x + −(1 + ) + t dt
0 2
3
= −(1 + )2 x + x 3 .
3!
3
As before, we have y3 (x) = c3 − (1 + )2 x + x 3 . Since y3 (x) must satisfy the
3!
given initial condition, it follows c3 = 0. Thus,
3 3
y3 (x) = −(1 + )2 x + x . (3.104)
3!
If we set = −1, it can be shown that
y2n (x) = 0, n ≥ 0,
(−1)n−1 2n−1 .
y2n−1 (x) = x , n ≥ 1.
(2n − 1)!
Substituting this result into (3.99), we get
x x3 x5
y(x) = − + − ···
1! 3! 5!
Obviously, the HAM generates the well-known Taylor series of the exact solution
y(x) = sin(x) of the IVP (3.83).
As we have seen, the main aim of the HAM is to produce solutions that will
converge in a much larger region than the solutions obtained with the traditional
perturbation methods. The solutions obtained by this method depend on the choice
of the linear operator L, the auxiliary function H(x), the initial approximation y0 (x),
and the value of the auxiliary parameter . By varying these parameters, it is possible
to adjust the region in which the series is convergent and the rate at which the series
converges.
One of the important factors that influences the convergence of the solution series
is the type of base functions used to express the solution. For example, we might try
to express the solution as a polynomial or as a sum of exponential functions. It can be
expected that base functions that more closely mimic the qualitative behavior of the
actual solution should provide much better results than base functions whose behavior
differs greatly from the qualitative behaviour of the actual solution. The choice of the
linear operator L, the auxiliary function H(x), and the initial approximation y0 (x)
often determines the base functions present in a solution.
Having selected L, H(x) and y0 (x), the deformation equations can be solved
and a solution series can be determined. The solution obtained in this way will still
contain the auxiliary parameter . But this solution should be valid for a range of
values of . To determine the optimum value of , the so-called -curves are often
plotted. These curves are obtained by plotting the partial sums y [n] (x) and/or their
first few derivatives evaluated at a specific value x = x̄ versus the parameter . As
long as the given IVP or BVP has a unique (isolated) solution, the partial sums and
their derivatives will converge to the exact solution for all values of for which the
approximate solution converges. This means that the -curves will be essentially
horizontal over the range of for which the solution converges. When is chosen in
this horizontal region, the approximate solution must converge to the exact solution
of the given problem (see [40]). However, the drawback of this strategy is that the
value x = x̄ and/or the order of the derivative are not known a priori.
A more mathematically sophisticated strategy to determine the optimum value of
(or an appropriate -interval) is to substitute the analytical approximation

n
y [n] (x; ) ≡ yk (x)
k=0
into the left-hand side of (3.76). Obviously, this expression vanishes only for the
exact solution. Now, the idea is to consider the norm of the residuum

∞ 2
F() ≡ N y [n] (x; ) d x (3.105)
0
and to compute the optimum value opt by solving the problem:
F() → min. (3.106)

In our studies (see Examples 3.16 and 3.17) we have used the symbolic tool of the
Matlab to determine the optimum value opt .
Note, the mentioned strategies for the determination of the auxiliary parameter
can only be used if the given problem is not ill-conditioned (see e.g. Example 3.17).
This means that small errors in the problem data produce only small errors in the
corresponding exact solution.
We have seen that the HAM has a high degree of freedom and a great flexibility
to choose the equation-type of the higher-order deformation equation and the base
functions of its solution. Now, we will consider the question how a linear operator,
an auxiliary function, and an initial approximation can be chosen. Liao [77] gives
several rules in order to ensure that these parameters are determined appropriately:
• Rule of Solution Expression
The first step in calculating a HAM solution of a given problem is to select a set of
base functions {en (x), n = 0, 1, 2, . . .} with which the solution of (3.76) should be
expressed. The most suitable base functions can often be chosen by determining
what types of functions most easily satisfy the initial or boundary conditions of a
problem and by considering the physical interpretation and expected asymptotic
behavior of the solution. Then, the solution can be represented in the form
∞

y(x) = cn en (x),
n=0
where cn are real constants. From this the rule of solution expression is derived,
which requires that every term in the solution is expressible in the form

M
yn (x) = an en (x),
n=0
where the an are real constants, and M is a nonnegative integer depending on

n. This rule together with the given initial or boundary conditions will guide the
choice of the initial approximation y0 (x).
The rule of solution expression also guides the choice of the linear operator L.
From Eq. (3.95) we see that ynh (x) should also obey the rule of solution expression.
Then, from (3.96) we conclude that the linear operator L should be chosen such
that the solution of the homogenous equation
L[ f (x)] = 0 (3.107)
is expressible as a linear combination of the base functions {en (x)}.

From (3.95) and the rule of solution expression, it is clear that the particular
solutions ync (x) of the higher-order deformation equation must also be expressible
as linear combinations of the chosen base functions. Since, according to (3.97),
each ync (x) depends on the form of the auxiliary function H(x), the rule of solution
expression restricts the choice of the auxiliary functions to those functions that will
produce terms ync (x) having the desired form.
• Rule of Coefficient Ergodicity
This rule guarantees the completeness of the solution. It says that each base en (x) in
the set of base functions should be present and modifiable in the series m k=0 yk (x)
for m → ∞. This rule further restricts the choice of the auxiliary function H(x)
and, when combined with the rule of solution expression, sometimes uniquely
determines the function H(x).
• Rule of Solution Existence
This third rule requires that L, H(x) and y0 (x) should be chosen such that each
one of the deformation equations (3.94) can be solved, preferably analytically.
In addition to ensuring that the above three rules are observed, the operator L should
be chosen such that the solution of the higher-order deformation equation is not too
difficult. In our experiments, we have seen that the following operators can be a good
choice:
• L[y(x)] = y (x),
• L[y(x)] = y (x) ± y(x),
• L[y(x)] = x y (x) ± y(x),
• L[y(x)] = y (x),
• L[y(x)] = y (x) ± y(x).
Now, let us come back to our Example 3.15 and apply the mentioned rules to the
IVP (3.100). The given initial condition y(0) = 0 implies that the choice y0 (x) = 0
is appropriate. For simplicity, we express the solution by the following set of base
functions
{en (x) = x n , n = 0, 1, 2, . . .}. (3.108)
Thus, the solution can be represented in the form

∞

y(x) = cn x n . (3.109)
n=0
Furthermore, it holds

M
yn (x) = an x n . (3.110)
n=0
According to the requirement (3.107), the choice L[y(x)] = y (x) is suitable. This
yields ynh (x) = 0. A look at formula (3.97) shows that H(x) must be of the form
H(x) = x k , where k is a fixed integer. Therefore, this formula reads
x
ync (x) = χn yn−1 (x) + t k Rn y0 (t), . . . , yn−1 (t) dt.
0
It is obvious that, when k ≤ −1, the term x −1 appears in the solution expression of
yn (x), which is in contradiction with the rule of solution expression. If k ≥ 1, the
term x always disappears in the solution expression of the higher-order deformation
equation. Thus, the coefficient of this term cannot be modified, even if the order of
approximation tends to infinity. Taking into account the rule 1 and rule 2, we have
to set k = 0, which uniquely determines the auxiliary function as H(x) = 1. Until
now, we still have freedom to choose a proper value of . Since the exact solution is
known, we can deduce that = −1. However, if the exact solution is unknown, the
strategy (3.105), (3.106) can be applied.
Example 3.16 Solve the following BVP of a second-order nonlinear ODE on the
infinite interval [0, ∞] by the HAM
y (x) = −2y(x)y (x), y(0) = 0, lim y(x) = 1. (3.111)

x→+∞
Solution. Here, we have N [y(x)] = y (x) + 2y(x)y (x). Since y(+∞) is finite, we
express y(x) by the set of base functions {e−nx , n = 1, 2, 3, . . .}. As an initial guess
that satisfies the given boundary conditions, we use y0 (x) = 1 − e−x . Moreover, we
set
L[y(x)] ≡ y (x) − y(x).
The solution of the homogeneous problem (3.96) is
ynh (x) = cn,1 e x + cn,2 e−x , n ≥ 1, (3.112)
where the real constants cn,1 and cn,2 are determined by appropriate boundary con-
ditions.
It holds

e x e−2x e x L[y(x)]d x d x

−2x
=e x
e e x (y (x) − y(x))d x d x

−2x x x
=e x
e e y (x) − e y (x)d x − e y(x)d x d x x

−x −2x x
=e x
e y (x)d x − e e y (x)d x + e y(x)d x d x
x

= ex e−x y (x)d x − e−2x e x y (x)d x + e x y(x)

x
− e y (x)d x d x

= ex e−x y (x)d x −
e−x y(x)d x

= ex e−x y (x)d x + e−x y(x) − e−x y (x)d x = y(x).
Therefore, the inverse operator is

−1 −2x
L [y(x)] = e x
e e x y(x)d x d x.
Now, formula (3.97) can be applied, and we obtain
ync (x) = χ(n)yn−1 (x)

x t
+ ex e−2t eu H(u) Rn (y0 (u), . . . , yn−1 (u))du dt.
0 0
According to the rule of solution expression, the auxiliary function H(x) should be
of the form H(x) = e−k x , where k is an integer. Setting k = 1, we get H(x) = e−x .
Hence,
ync (x) = χ(n)yn−1 (x)

x t
−2t
(3.113)
+ e x
e Rn (y0 (u), . . . , yn−1 (u))du dt.
0 0
In order to compute Rn (y0 (x), . . . , yn−1 (x)), we have to substitute (3.83) into
N [φ(x; p)] = φ (x; p) + 2φ(x; p)φ (x; p). We obtain
∞

N [φ(x; p)] = − e−x + yn (x) p n
n=1
∞

∞

−x −x
+ 2 1−e + yn (x) p n
e + yn (x) p n .
n=1 n=1
Now, the functions Rn are defined by (3.93). We compute
R1 (y0 (x)) = N [φ(x; p)]| p=0 = −e−x + 2(1 − e−x )e−x = e−x − 2e−2x .
Using formulas (3.112) and (3.113), we obtain

y1 (x) = y1h (x) + y1c (x)

−x
x
−2t
t
= c1,1 e + c1,2 e
x
+ e x
e e−u − 2e−2u du dt
0 x 0
−x −2t

= c1,1 e + c1,2 e
x
+ e x
e e−2t − e−t dt
0 x
−x

= c1,1 e + c1,2 e
x
+ e x
e−4t − e−3t dt
0
1 −3x 1 −4x 1
= c1,1 e x + c1,2 e−x + ex e − e −
3 4 12

1 −3x 1 −2x 1
= c1,1 e x + c1,2 e−x + − e + e − ex .
4 3 12
Since y0 (x) satisfies the boundary conditions given in (3.111), in particular the second
condition, the next iterates yn (x), n ≥ 1, have to fulfill the boundary conditions
yn (0) = 0, lim yn (x) = 0. (3.114)

x→+∞
By these boundary conditions the constants c1,1 and c1,2 are determined as c1,1 =
/12 and c1,2 = −/12. Thus,

1 −3x 1 −2x 1 −x
y1 (x) = − e + e − e . (3.115)
4 3 12

∂N [φ(x; p)]
R2 (y0 (x), y1 (x)) =
∂p p=0
= y1 (x) + 2e−x y1 (x) + 2(1 − e−x )y1 (x)

9 −3x 4 −2x 1 −x
= − e + e − e
4 3 12

1 −4x 1 −3x 1 −2x
+ 2 − e + e − e
4 3 12

3 −3x 2 −2x 1 −x
+ 2 e − e + e
4 3 12

3 −4x 2 −3x 1 −2x
− 2 e − e + e
4 3 12

−4x 5 −3x 1 −2x 1 −x
= −2e + e − e + e ,
4 3 12
and
y2 (x) = y2h (x) + y2c (x)

−x 1 −3x 1 −2x 1 −x
= c2,1 e + c2,2 e + − e
x
+ e − e
4 3 12
x t
−2t −4u 5 −3u 1 −2u 1 −u
+ e 2 x
e −2e + e − e + e du dt
0 0 4 3 12

1 1 1
= c2,1 e x + c2,2 e−x + − e−3x + e−2x − e−x
4 3 12
x
−2t 1 −4t 5 −3t 1 −2t 1 −t 1
+ e 2 x
e e − e + e − e − dt
0 2 12 6 12 6

1 1 1
= c2,1 e x + c2,2 e−x + − e−3x + e−2x − e−x
4 3 12
x
1 −6t 5 −5t 1 −4t 1 −3t 1 −2t
+ e 2 x
e − e + e − e − e dt
0 2 12 6 12 6

1 1 1
= c2,1 e x + c2,2 e−x + − e−3x + e−2x − e−x
4 3 12

1 1 1 1
+ e − e−6x + e−5x − e−4x + e−3x
2 x
12 12 24 36

1 −2x 5
+ e −
12 72

−x 1 −3x 1 −2x 1 −x
= c2,1 e + c2,2 e + − e
x
+ e − e
4 3 12

2
−5x −4x 1 −3x 1 −2x −x 5 x
+ −e +e − e + e +e − e .
12 2 3 6
The function y2 (x) satisfies the boundary conditions (3.114) if we set
5 2 5
c2,1 = , c2,2 = − 2 .
72 72
Then,

1 1 1
y2 (x) = − e−3x + e−2x − e−x
4 3 12
(3.116)
2 1 1 1
+ −e−5x + e−4x − e−3x + e−2x + e−x .
12 2 3 6
Substituting the results into (3.99), we obtain the following approximation

[2] −x 1 −3x 1 −2x 1 −x
y (x; ) =1 − e + 2 − e + e − e
4 3 12
(3.117)
2
−5x −4x 1 −3x 1 −2x 1 −x
+ −e +e − e + e + e .
12 2 3 6
Note, to indicate that the approximate solution y [2] still depends on the
convergence-control parameter , we have used it as a second argument.
It can be easily shown that the exact solution of the BVP (3.112) is the function
y(x) = tanh(x).
The next task is to determine values of the parameter for which the series (3.99)
converges to the exact solution. Here, we have used the strategy (3.105), (3.106). For
a better understanding we have determined additional terms y3 (x), . . . , y20 (x) using
the symbolic tool of the Matlab. Then we have solved the optimization problem
without constraints (3.106) for the nth partial sum of the terms yn (x), n = 1, . . . , 20.
In the Fig. 3.10, for each y [n] the function F() is plotted versus . It can be easily
seen that a region for appropriate values of is [−2, 0). Moreover, in Table 3.10 we
present the corresponding optimum values (n) opt , n = 1, . . . , 20.
(n)
To determine an approximation of opt if n tends to ∞, we have computed a
fourth-degree polynomial, which fits the points n = 3, . . . , 20 given in Table 3.10;
see Fig. 3.11. Then, we have extrapolated to F() = 0. The resulting value is ∗ =
−1.835.
Fig. 3.10 F() versus for y [n] , n = 1, . . . , 20

(n) (n)
Table 3.10 Values of opt and F(opt ) if the number n of terms in (3.98) is increased
(n) (n) (n) (n)
n opt F(opt ) n opt F(opt )
1 −1.207 1.4 e-1 11 −1.765 2.3 e-2
2 −1.325 1.2 e-1 12 −1.773 1.9 e-2
3 −1.480 1.0 e-1 13 −1.776 1.6 e-2
4 −1.592 8.6 e-2 14 −1.783 1.3 e-2
5 −1.653 7.3 e-2 15 −1.786 1.1 e-2
6 −1.695 6.0 e-2 16 −1.792 9.8 e-3
7 −1.719 4.9 e-2 17 −1.796 8.5 e-3
8 −1.738 4.0 e-2 18 −1.801 7.3 e-3
9 −1.749 3.3 e-2 19 −1.804 6.3 e-3
10 −1.760 2.7 e-2 20 −1.808 5.4 e-3

(n) (n)
Fig. 3.11 ∗: from right to left: opt , F(opt ) , n = 3, . . . , 20; ◦: extrapolated point (∗ such that
F(∗ ) = 0)
3.4.2 Application of the Homotopy Analysis Method
In this section we apply the Homotopy Analysis Method to two examples, which are
often discussed in related papers and books.
Example 3.17 Once again, let us consider Bratu’s problem (see Examples
1.25 and 2.4):
y (x) = −λ e y(x) , y(0) = y(1) = 0. (3.118)
Find an approximate solution of this equation and compare the result with the data
presented in Example 2.4.
Solution. Here, we have N [y(x)] = y (x) + λ e y(x) . According to the ODE and the
boundary conditions, we express the solution by the following set of base functions
{en (x) = x n , n = 0, 1, 2, . . .}.
Thus, the solution can be represented in the form

∞

y(x) = an x n .
n=0
We start with the function y0 (x) ≡ 0, which satisfies the boundary conditions.
Moreover, we set
L[y(x)] ≡ y (x).
The solution of the homogeneous problem (3.96) is
ynh (x) ≡ cn,1 + cn,2 x, n ≥ 1, (3.119)
Now, formula (3.97) can be applied, and we obtain

x t
ync (x) = χ(n)yn−1 (x) + H(u) Rn (y0 (u), . . . , yn−1 (u))du dt.
0 0
The auxiliary function H(x) should be of the form H(x) = ekx , where k is an integer.
For k = 0, exponential terms will appear in the solution expression of ync (x), which
disobeys the rule of solution expression. Hence, k = 0, which uniquely determines
the corresponding auxiliary function as H(x) ≡ 1. Thus,
x t
ync (x) = χ(n)yn−1 (x) + Rn (y0 (u), . . . , yn−1 (u))du dt. (3.120)
0 0
In order to compute Rn (y0 (x), . . . , yn−1 (x)), we have to substitute (3.83) into
N [φ(x; p)] = φ(x; p) (x) + λ eφ(x; p) . We obtain
∞
∞

N [φ(x; p)] = yn (x) p n + λ exp yn (x) p n .
n=0 n=0
R1 (y0 (x)) = N [φ(x; p)]| p=0 = λ.
Using formulas (3.112) and (3.120), we obtain

y1 (x) = y1h (x) + y1c (x)

x t
= c1,1 + c1,2 x + λ du dt
0 0
2
x
= c1,1 + c1,2 x + λ .
2
Since y1 (x) must satisfy the given boundary conditions, we obtain
1
c1,1 = 0, c1,2 = − λ.
2
Thus,
λ λ 2
y1 (x) = − x+ x . (3.121)
2 2

∂N [φ(x; p)]
R2 (y0 (x), y1 (x)) =
∂p p=0
= y1 (x) + λ y1 (x)
λ2 λ2 2
= λ − x+ x .
2 2
and
y2 (x) = y2h (x) + y2c (x)

= c2,1 + c2,2 x

x t
2 λ 2 x t
+ y1 (x) + λ 2
du dt + (u 2 − u)du dt
0 0 2 0 0
= c2,1 + c2,2 x

λ λ 2 2 λ 2 2 λ2 x 4 x3
− x+ x + x + −
2 2 2 2 12 6
λ 1 λ
2 2
2 λ 2 4
= c2,1 + c2,2 x − x + (λ + 2 λ)x 2 − x3 + x .
2 2 12 24
Considering the given boundary conditions, we get
2 λ 2 2 λ
c2,1 = 0, c2,2 = − .
24 2
Thus,
2 λ2 2 λ λ 1
y2 (x) = − − x + (λ + 2 λ)x 2
24 2 2 2
(3.122)
2 λ 2 3 2 λ 2 4
− x + x .
12 24
Now, we determine

1 ∂ 2 N [φ(x; p)]
R3 (y0 (x), y1 (x), y2 (x)) = .
2 ∂ p2 p=0
This yields

1
R3 (·) = y2 (x)
+ λ y2 (x) + y1 (x) 2
2
2 3
λ λ2
= (λ + 2 λ) + − − 2 λ2 x
24 2

1 1 1 1
+ 2 λ2 + λ2 + 2 λ3 x 2 − 2 λ3 x 3 + 2 λ3 x 4 ,
2 8 3 6
and
y3 (x) = c3,1 + c3,2 x

x t
+ y2 (x) + R3 (y0 (u), y1 (u), y2 (u))du dt
0 0
= c3,1 + c3,2 x

1 2 2 1 2 1 1
+ λ − λ − λ x + (λ + 2 λ)x 2
24 2 2 2
1 2 2 3 1
− λ x + 2 λ 2 x 4
12 24
1 2 2 1 3 3 1 2 2 1 3 2 3
+ λ+ λ x +
3
λ − λ − λ x
2 144 12 6

1 3 2 1 2 2 1 3 3 4
+ λ + λ + λ x
12 24 96
1 3 3 5 1 3 3 6
− λ x + λ x
60 180
= c3,1 + c3,2 x

1 2 2 1 2 1 1 1 3
+ λ − λ − λ x + λ + λ + λ x 2
2
24 2 2 2 2

1 2 2 1 3 3 1 3 2 3
+ − λ + λ − λ x
6 144 6

1 2 2 1 1
+ λ + 3 λ2 + 3 λ3 x 4
12 12 96
1 3 3 5 1 3 3 6
− λ x + λ x .
60 180
Since y3 (x) must satisfy the given boundary conditions, we obtain
c3,1 = 0,
1 2 2 1 2 1 1 3 3 1
c3,2 = λ − λ − 3 λ − λ + 3 λ2 .
24 2 2 160 12
Thus,

1 2 2 1 1 3 1 3 3 1 3 2
y3 (x) = λ − λ − λ − λ −
2
λ + λ x
12 2 2 160 12

1 1
+ 2 λ + λ + 3 λ x 2
2 2

1 2 2 1 3 3 1 3 2 3
+ − λ + λ − λ x (3.123)
6 144 6

1 2 2 1 3 2 1 3 3 4
+ λ + λ + λ x
12 12 96
1 3 3 5 1 3 3 6
− λ x + λ x .
60 180
Now, we substitute our results into the representation (3.98), and obtain
y [3] (x; ; λ) = y0 (x) + y1 (x) + y2 (x) + y3 (x). (3.124)
As before, the next step is to adjust such that (3.124) is a good approximation of
the solution y(x) of the BVP (3.118). In order to obtain a more exact result, we have
used the symbolic tool of the Matlab to compute additional terms y4 (x), . . . , y7 (x).
Then, we have solved the optimization problem without constraints (3.106) for the
nth partial sum of the terms yn (x), n = 1, . . . , 7. In the Fig. 3.12, the results for
λ1 = 1, λ2 = 2, λ3 = 3, and λ4 = 3.513 are represented. For each y [n] , the function
F() is plotted versus . Moreover, in Table 3.11, for the mentioned values of λ,
the corresponding optimum values (n) opt , n = 1, . . . , 7, are given. Obviously, the
(n)
optimum values opt change with the parameter λ.
In the Tables 3.12 and 3.13, we compare y [3] (x; ; λ) with the exact solution of
Bratu’s problem for λ = 1 and λ = 2, respectively. In both cases we have used the
optimum , which is given in Table 3.11, i.e., (λ=1)
opt = −1.11 and (λ=2)
opt = −1.25.
It can be seen that for these values the HAM gives similar good results as the VIM
presented in Sect. 2.2 (see Example 2.4).
Fig. 3.12 For y [n] , n = 1, . . . , 7, and λ = 1, 2, 3, 3.513: F() versus
(n)
Table 3.11 For λ = 1, 2, 3, 3.513: optimum values opt , n = 1, . . . , 7
n λ1 = 1 λ2 = 2 λ3 = 3 λ4 = 3.513
1 −1.090 −1.210 −1.350 −1.390
2 −1.090 −1.200 −1.330 −1.400
3 −1.110 −1.250 −1.420 −1.490
4 −1.110 −1.250 −1.430 −1.540
5 −1.120 −1.280 −1.480 −1.570
6 −1.120 −1.280 −1.490 −1.620
7 −1.130 −1.300 −1.520 −1.630
In Example 5.21, we will show how the solution curve of Bratu’s BVP can be com-
puted by appropriate numerical techniques. This curve possesses a simple turning
point at λ0 = 3.51383071912. Turning points are singular points of the given prob-
lem, i.e. problem (3.118) is ill-conditioned in the neighborhood of λ0 . It is therefore
to be expected that the HAM fails or produces inaccurate results. This is confirmed
by the results presented in Table 3.14, where we have used λ = 3.513.
Table 3.12 Results computed by the HAM for Bratu’s problem with λ = 1; θ =
1.5171645990507543685218444212962
x y [3] (x; −1.11; 1) y(x) Relative error
0.1 0.04983221 0.04984679 2.93 e-04
0.2 0.08915608 0.08918993 3.80 e-04
0.3 0.11755708 0.11760910 4.42 e-04
0.4 0.13472542 0.13479025 4.81 e-04
0.5 0.14046976 0.14053921 4.94 e-04
0.6 0.13472542 0.13479025 4.81 e-04
0.7 0.11755708 0.11760910 4.42 e-04
0.8 0.08915608 0.08918993 3.80 e-04
0.9 0.04983221 0.04984679 2.93 e-04
Table 3.13 Results computed by the HAM for Bratu’s problem with λ = 2; θ =
2.3575510538774020425939799885899
x y [3] (x; −1.25; 2) y(x) Relative error
0.1 0.11382305 0.11441074 5.14 e-03
0.2 0.20514722 0.20641912 6.16 e-03
0.3 0.27198555 0.27387931 6.91 e-03
0.4 0.31276250 0.31508936 7.38 e-03
0.5 0.32647027 0.32895242 7.55 e-03
0.6 0.31276250 0.31508936 7.38 e-03
0.7 0.27198555 0.27387931 6.91 e-03
0.8 0.20514722 0.20641912 6.16 e-03
0.9 0.11382305 0.11441074 5.14 e-03
Table 3.14 Results computed by the HAM for Bratu’s problem with λ = 3.513; θ =
4.7374700066634551362743387948882
x y [3] (x; −1.49; 3.513) y(x) Relative error
0.1 0.26743526 0.37261471 2.80 e-01
0.2 0.48796610 0.69393980 2.95 e-01
0.3 0.65320219 0.94487744 3.07 e-01
0.4 0.75571630 1.10579685 3.15 e-01
0.5 0.79047843 1.16138892 3.18 e-01
0.6 0.75571630 1.10579685 3.15 e-01
0.7 0.65320219 0.94487744 3.07 e-01
0.8 0.48796610 0.69393980 2.95 e-01
0.9 0.26743526 0.37261471 2.80 e-01
Moreover, the numerically determined solution curve shows: Bratu’s BVP has no
solution for λ > λ0 , one (non-isolated) solution for λ = λ0 , and two solutions for
λ < λ0 . In [4] an analytical shooting method is proposed to determine for a fixed
value λ < λ0 both solutions by the HAM.
Example 3.18 In [3] the following IVP is considered, which describes the cooling
of a lumped system with variable specific heat

1 + εy(x) y (x) + y(x) = 0, y(0) = 1. (3.125)
Determine an approximate solution of this IVP.

Solution. Here, we have N [y(x)] = 1 + εy(x) y (x) + y(x). According to the
governing ODE and the initial condition (3.125), the solution can be expressed by
the following set of base functions
{en (x) = e−nx , n = 1, 2, 3 . . .},
in the form
∞

y(x) = dn e−nx , (3.126)
n=1
where the real coefficients dn have to be determined. Now, the rule of solution expres-
sion says that the solution of (3.125) must be expressed in the same form as (3.126)
and the other expressions such as x m e−nx must be avoided. We start with the function
y0 (x) = e−x , which satisfies the initial condition. Moreover, we set
L[y(x)] ≡ y (x) + y(x).
The general solution of the linear homogeneous problem L[y(x)] = 0 is
ynh (x) = cn e−x , (3.127)
where cn is a real constant.

Since

−x −x

e e L[y(x)]d x = e
x
e x y (x) + y(x) d x = y(x),
it holds
L−1 [y(x)] = e−x e x y(x)d x.
Thus, for n ≥ 1 the solution of the nth-order deformation equation becomes

x
ync (x) = χ(n)yn−1 (x) + e−x et H(t)Rn (y0 (t), . . . , yn−1 (t))dt.
0
According to the rule of solution expression denoted by (3.126) and from (3.94), the
auxiliary function H(x) should be in the form H(x) = e−ax , where a is an integer.
It is found that, when a ≤ −1, the solution of the high-order deformation (3.94)
contains the term xe−x , which incidentally disobeys the rule of solution expression.
When a ≥ 1, the base e−2x always disappears in the solution expression of the high-
order deformation, so that the coefficient of the term e−2x cannot be modified even
if the order of approximation tends to infinity. Thus, we have to set a = 0, which
uniquely determines the corresponding auxiliary function as H(x) ≡ 1. Hence, we
have
x
ync (x) = χ(n)yn−1 (x) + e−x et Rn (y0 (t), . . . , yn−1 (t))dt. (3.128)
0
In order to compute
Rn (y0 (x),
. . . , yn−1 (x)), we have to substitute (3.83) into
N [φ(x; p)] = 1 + εφ(x; p) φ (x; p) + φ(x; p). We obtain
∞

N [φ(x; p)] = 1 + ε y0 (x) + yn (x) p n
y0 (x) + yn (x) p n
n=1 n=1
∞

+ y0 (x) + yn (x) p n .
n=1
R1 (y0 (x)) = N [φ(x; p)]| p=0 = −εe−2x .
Using formula (3.128), we obtain

x x
−x t −2t −x
y1c (x) = −εe ee dt = −εe e−t dt = −εe−x + εe−2x .
0 0
Using formulas (3.95) and (3.127), it follows
y1 (x) = c1 e−x − εe−x + εe−2x .
Since y1 (x) must satisfy the given initial condition, we get c1 = 0, and
y1 (x) = −εe−x + εe−2x . (3.129)


∂N [φ(x; p)]
R2 (y0 (x), y1 (x)) =
∂p p=0
= εy1 (x)y0 (x) + (1 + εy0 (x))y1 (x) + y1 (x)

= −ε2 e−x + ε2 e−2x (−e−x )

+ 1 + εe−x εe−x − 2εe−2x + −εe−x + εe−2x

= 2ε2 − ε e−2x − 3ε2 e−3x ,
and
x
y2c (x) = −εe−x + εe−2x + e−x (2ε2 − ε)e−t − 3ε2 e−2t dt
0
= −εe−x + εe−2x
3
+ (2ε2 2 − ε2 ) e−x − e−2x − ε2 2 e−x − e−3x
2
1 2 2 3
= ε − ε(1 + ) e + (ε(1 + ) − 2ε2 2 )e−2x + ε2 2 e−3x .
−x
2 2
Thus, we obtain

−x 1 2 2
y2 (x) = c2 e + ε − ε(1 + ) e−x
2
3
+ (ε(1 + ) − 2ε2 2 )e−2x + ε2 2 e−3x .
2
As before, the function y2 (x) must satisfy the given initial condition. This implies
c2 = 0, and we have

1 2 2
y2 (x) = ε − ε(1 + ) e−x + ε(1 + ) − 2ε2 2 e−2x
2
(3.130)
3
+ ε2 2 e−3x .
2
Now, we substitute our results into the representation (3.98), and we obtain the
following HAM approximation
y [2] (x; ; ε) = y0 (x) + y1 (x) + y2 (x), (3.131)

Fig. 3.13 For y [2] and different values of ε: F() versus
Fig. 3.14 For y [7] and different values of ε: F() versus
which depends on the external parameter ε and the auxiliary parameter . In Fig. 3.13,
for different values of ε and the three-term approximation y [2] , the function F() is
plotted versus . In Fig. 3.14 the same is shown for y [7] , where this approximation
has been computed with the symbolic tool of the Matlab. Finally, in Table 3.15 we
present the optimum values opt for different values of ε.
(2) (7)
Table 3.15 Values of opt and opt for different values of ε
ε 0.1 0.5 1 1.5 2
(2)
opt −0.95 −0.78 −0.62 −0.51 −0.43
(7)
opt −0.96 −0.74 −0.57 −0.46 −0.38
3.5 Exercises
Exercise 3.1 Let us consider the free vibration of a mass m connected to two non-
linear springs on a frictionless contact surface, with a small nonlinear perturbation.
The IVP for this system is
y (x) + ωn2 y(x) + εy(x)3 = 0,

y(0) = A, y (0) = 0.
Here, y(x) is the displacement at time x, y (x) is the acceleration and ε (ε 1)

√ positive parameter. The natural frequency ωn of the system is given by
is a small
ωn = K 1 + K 2 /m (rad/s), where K 1 and K 2 are the stiffnesses of the nonlinear
springs. Discuss the solution of this system by the perturbation method.
Exercise 3.2 Consider the structural beam with two hinged ends under a concen-
trated load at the middle of the span. The BVP for this system has the following
form:
Fx 3
y (x) − 1 + y (x)2 2 = 0,
2E I
(3.132)
F L2
y(0) = 0, y (0) =
16E I
Note, there is not a small parameter in the ODE. Solve this equation with the pertur-
bation method by adding a small parameter to the nonlinear term.
Exercise 3.3 Solve the BVP (3.132) by setting
23 3 2
1 + y (x)2 ≈1+ y (x) ,
2
and compare your results with those obtained in Exercise 3.2.
Exercise 3.4 Consider the following IVP for a nonlinear oscillator with a disconti-
nuity
y (x) + ωn2 y(x) + εy(x)3 sign(y(x)) = 0,
y(0) = A, y (0) = 0.
Discuss the solution of this IVP by the energy balance method.

3.5 Exercises 119
Exercise 3.5 Consider the motion of a rigid rod rocking back on the circular surface
without slipping. The governing IVP of this motion is
3 3 g
y (x) +y(x)2 y (x) + y(x)y (x)2 + 3 y(x) cos(y(x)) = 0,
4 4 l
y(0) = A, y (0) = 0.
Here, g and l are positive constants. Discuss the solution of this system by the energy
balance method.
Exercise 3.6 Consider the free vibrations of an autonomous conservative oscillator

with inertia and static type fifth-order nonlinear ties. The IVP of this dynamical
system is
y (x) + ε1 y(x)2 y (x) + ε2 y(x)4 y (x) + ε1 y(x)y (x)2

+ 2ε2 y(x)3 y (x)2 + λy(x) + ε3 y(x)3 + ε4 y(x)5 = 0, (3.133)
y(0) = A, y (0) = 0.
Here, ε1 , . . . , ε4 are positive parameters, and λ is an integer which may take the
values −1, 0, 1. Discuss the solution of this system by the energy balance method.
Exercise 3.7 Consider another nonlinear oscillator with a discontinuity
y (x) + y(x) + y(x)3 = 0,

y(0) = A, y (0) = 0.
Discuss the solution of this system by the Hamiltonian approach.
Exercise 3.8 Consider the free vibration of a tapered beam. In dimensionless form,
the governing IVP corresponding to the fundamental vibration mode is

y (x) + ε1 y(x)2 y (x) + y(x)y (x)2 + y(x) + ε2 y(x)3 = 0,
(3.134)
y(0) = A, y (0) = 0,
where ε1 and ε2 are arbitrary real constants, and y(x) is the displacement at time x.
Discuss the solution of this system by the Hamiltonian approach.
Exercise 3.9 Consider the motion of a particle on a rotating parabola. The IVP that
describes this motion is
y (x) + 4q 2 y(x)2 y (x) + 4q 2 y(x)y (x)2 + Δy(x) = 0,

y(0) = A, y (0) = 0.
Here, y (x) is the acceleration, y (x) is the velocity, q and Δ are positive constants.
Discuss the solution of this IVP by the Hamiltonian approach.
Exercise 3.10 As we know, the vibration of the simple pendulum can be modeled
mathematically by the following IVP
y (x) + ωn2 sin(y(x)) = 0,

y(0) = A, y (0) = 0,
√
where ωn = g/l is the natural frequency of the pendulum. Solve this IVP by the
homotopy analysis method.
Exercise 3.11 Consider the forced vibration of a mass m connected to a nonlinear

spring on a frictionless contact surface. The IVP of this dynamical system is
kb cos ωx
y (x) + ωn2 y(x) + εy(x)3 = ,
m
y(0) = A, y (0) = 0.
√
In this ODE, ωn is the natural frequency of the system which is given by ωn = k/m
(rad/s), ε 1 is a positive real parameter, and k is the stiffness of the nonlinear spring.
Solve this equation by the homotopy analysis method.
Exercise 3.12 Solve the IVP (3.133) given in Example 3.6 by the Hamiltonian
approach and compare the results.
Exercise 3.13 Solve the IVP (3.134) given in Example 3.8 by the energy balance
method and compare the results.
Exercise 3.14 In the theory of the plane jet, which is studied in the hydrodynamics,
we are encountered with the following IVP (see [3])
ay (x) + y(x)y (x) + y (x)2 = 0,

y(0) = y (x) = 0, lim y (x) = 0.
x→+∞
Here, a > 0 is a real parameter. Discuss the solution of this IVP by the homotopy
analysis method.
Exercise 3.15 Consider the first extension of Bratu’s problem (see [128])
y (x) + e y(x) + e2y(x) = 0,

y(0) = y(1) = 0.
Solve this BVP by the homotopy analysis method.

Chapter 4
Nonlinear Two-Point Boundary
Value Problems
4.1 Introduction
Let us consider, the following system of n first-order ordinary differential equations

(ODEs)
y1 (x) = f 1 (x, y1 (x), y2 (x), . . . , yn (x)),

y2 (x) = f 2 (x, y1 (x), y2 (x), . . . , yn (x)),
(4.1)
···
yn (x) = f n (x, y1 (x), y2 (x), . . . , yn (x)),
where yi (x) denotes the derivative of the function yi (x) w.r.t. x, i = 1, . . . , n.
Setting y(x) ≡ (y1 (x), . . . , yn (x))T and f (x, y) ≡ ( f 1 (x, y), . . . , f n (x, y))T ,
this system can be formulated in vector notation as
d
y (x) ≡ y(x) = f (x, y(x)), x ∈ [a, b], y(x) ∈ Ω ⊂ Rn . (4.2)
dx
We assume that the function f : [a, b] × Ω → Rn is sufficiently smooth.

In the applications higher order ODEs are often encountered. Such a higher order
equation

u (n) (x) = G x, u(x), u (x), u (x), . . . , u (n−1) (x) (4.3)
can be transformed into the system (4.1) by setting

T
y = (y1 , y2 , . . . , yn )T ≡ u, u , u , . . . , u (n−1) .
Obviously, the first-order system reads now
y1 = y2 , y2 = y3 , . . . , yn−1

= yn , yn = G(x, y1 , y2 , . . . , yn ). (4.4)
DOI 10.1007/978-81-322-2812-7_4
122 4 Nonlinear Two-Point Boundary Value Problems
Example 4.1 Given the following ODE, which models the buckling of a thin rod
u (t) + λ sin(u(t)) = 0. (4.5)
Transform this second-order ODE into a first-order system.

Solution: We obtain the following system of two first-order ODEs
y1 (x) = y2 (x), y2 (x) = −λ sin(y1 (t)).
In the monograph [63], the ODEs (4.1) have been subjected to initial and linear
two-point boundary conditions. Here, we add nonlinear two-point boundary condi-
tions
g1 (y1 (a), . . . , yn (a), y1 (b), . . . , yn (b)) = 0,

g2 (y1 (a), . . . , yn (a), y1 (b), . . . , yn (b)) = 0,
.. (4.6)
.
gn (y1 (a), . . . , yn (a), y1 (b), . . . , yn (b)) = 0.
to (4.1). The n algebraic equations (4.6) can be written in vector notation in the form
g( y(a), y(b)) = 0, (4.7)
and we assume that the function g : Ω1 × Ω2 → Rn , Ω1 ⊂ Rn and Ω1 ⊂ Rn , is

sufficiently smooth.
The combination of (4.2) and (4.7), i.e.,
y (x) = f (x, y(x)), a ≤ x ≤ b,

(4.8)
g( y(a), y(b)) = 0,
is called two-point boundary value problem (BVP). If components of the solution

y(x) are prescribed at more than two points, we have a multipoint boundary value
problem. Initial value problems (IVPs) are typical for evolution problems, and x
represents the time. In contrast, boundary value problems are used to model static
problems (e.g., elasto-mechanical problems) since the solution depends on the state
of the system in the past (at the point a) as well as on the state of the system in the
future (at the point b). In general, the variable x in a BVP does not describe the time,
i.e., no dynamics is inherent in the given problem.
In the applications, the function g is often linear. In that case, the boundary
conditions (4.7) can be written in the form
Ba y(a) + Bb y(b) = β, (4.9)

4.1 Introduction 123
where Ba , Bb ∈ Rn×n and β ∈ Rn . We speak of a linear BVP, if the ODE and the
boundary conditions are linear, i.e.,
y (x) = A(x) y(x) + r(x), a ≤ x ≤ b,

(4.10)
Ba y(a) + Bb y(b) = β,
where A(x), Ba , Bb ∈ Rn×n , r(x), β ∈ Rn .

The proof of the existence and uniqueness of solutions for nonlinear BVPs is
extremely difficult and is seldom as successful as for linear BVPs (see, e.g., [57]).
Therefore, we always assume that the BVP (4.8) has a locally unique (isolated)
solution y(x), which has to be approximated by numerical methods.
The most important numerical methods for (4.8) can be subdivided into three
classes:
• shooting methods,
• finite difference methods, and
• collocation methods.
In this book, we shall consider only the class of shooting methods, since an
important component of these methods are the sophisticated numerical techniques
for IVPs. In particular, this means the use of step-size controlled IVP-solvers, which
we have described in our text [63].
4.2 Simple Shooting Method
The idea of the simple shooting method is to assign an IVP
u (x) = f (x, u(x)), u(a) = s ∈ Rn , (4.11)
to the nonlinear BVP (4.8). Obviously, this IVP contains the still undetermined
initial vector s, which can be considered as a free parameter. Therefore, we denote
the solution of (4.11) with u ≡ u(x; s).
The essential requirement on the IVP, which brings the IVP in relation to the BVP,
is that the IVP possesses the same solution as the BVP. Since in (4.8) and (4.11) the
differential equations are identical, the vector s in (4.11) has to be determined such
that the corresponding solution trajectory u(x; s) satisfies the boundary conditions
in (4.8), too. This means that the unknown n-dimensional vector s must satisfy the
n-dimensional system of nonlinear algebraic equations
F(s) ≡ g(s, u(b; s)) = 0. (4.12)
The following relations between the solutions of the algebraic system (4.12) and
the solutions of the BVP (4.8) can be shown:
• if s = s∗ is a root of (4.12), then y(x) ≡ u(x; s∗ ) is a solution of the BVP (4.8),

• if y(x) is a solution of the BVP (4.8), then s = y(a) is a root of (4.12),
• for each isolated solution of the BVP (4.8), there is a simple root of the algebraic
equations (4.12).
Thus, the nonlinear BVP (4.8), which is defined in appropriate (infinitely dimen-
sional) function spaces, is transformed into the finite dimensional nonlinear alge-
braic problem (4.12)—and this in an exact way, i.e., without any approximation
error. This is a significant advantage of the shooting methods, in comparison with
the other classes of BVP-solvers mentioned above.
As with all nonlinear problems, the numerical treatment of (4.12) requires knowl-
edge of the approximate position of the root s∗ and an associated neighborhood U ∗
where no other solution exists. For an arbitrary nonlinear BVP this information can
rarely be formulated in theory-based statements. Instead, extensive numerical exper-
iments are necessary to find good starting values. In many cases, the information
about the practical problem, which is modeled by (4.8), can be helpful. If nothing
is known about the given problem, the so-called trial-and-error technique is often
used.
Numerical tools to globalize the locally convergent methods for nonlinear BVPs
are continuation and embedding techniques, which are also referred to as homotopy
methods. Here, the BVP (4.8) is embedded into a one-parametric family of BVPs.
Then, an attempt is made to increase successively the parameter value until the
original BVP (4.8) is reached. The idea of this strategy is to use the approximate
solution of the previous problem as starting trajectory for the subsequent problem.
If it is possible to separate a linear term in the ODE and/or the boundary conditions,
i.e., if the BVP can be splitted as follows:
y (x) = f (x, y(x)) = A(x) y(x) + r(x) + f 1 (x, y(x)),

(4.13)
g( y(a), y(b)) = Ba y(a) + Bb y(b) − β + g 1 ( y(a), y(b)) = 0,
then the nonlinear part is multiplied by the embedding parameter λ:
y (x) = A(x) y(x) + r(x) + λ f 1 (x, y(x)),

(4.14)
Ba y(a) + Bb y(b) − β + λ g 1 ( y(a), y(b)) = 0.
Obviously, for λ = 0 we have a simpler (linear) problem, and for λ = 1 the original
problem. The continuation method, which is often called killing of the nonlinearity,
starts with λ = 0, i.e., with the numerical solution of the linear problem. Then, the
approximate solution of the linear problem is used as starting trajectory to solve
the problem (4.14) for a slightly increased value of λ, say λ1 . If it is not possible
to compute a solution of the new problem, λ1 must be decreased. Otherwise, λ1 is
increased. This process is executed step-by-step until a solution of the original BVP
(λ = 1) is determined.
Another preferred continuation method for BVPs uses the interval length b − a
as the embedding parameter λ; see, for example, [69, 102].
4.2 Simple Shooting Method 125
Further information on continuation methods for the numerical treatment of sys-

tems of nonlinear equations, such as (4.12), can be found in the standard reference
works on nonlinear equations [93, 101, 110]. These books also contain the basic
algorithms for the numerical treatment of (4.12).
The most common methods for the numerical solution of (4.12) are Newton-
like techniques. To ensure the feasibility of these methods and a sufficiently fast
convergence of the iterates toward the desired solution s∗ , it is usually assumed that
the Jacobian of F, i.e., the matrix M(s∗ ) ≡ ∂ F(s∗ )/∂s, is nonsingular. Under this
hypothesis the solution y∗ ≡ u(x; s∗ ) of the BVP (4.8) is isolated. The presence of
non-isolated solutions can lead, inter alia, to the phenomenon of branching solutions.
The study of branching solutions is the objective of the bifurcation theory, which we
will consider in the next chapter.
Let us now assume that we use the multidimensional Newton method (see, e.g.,
[58, 93]) to solve (4.12). Starting with a vector s0 , which is sufficiently close to the
solution s∗ , i.e.,
s0 ∈ S(s∗ ) ≡ {s ∈ Rn : ||s − s∗ || ≤ ε},
a sequence of iterates {sk }∞

k=0 is computed by the formula
sk+1 = sk − ck , k = 0, 1, . . . , (4.15)
where the increment ck is determined by the following n-dimensional system of

linear equations
Mk ck = q k . (4.16)
For the right-hand side of (4.16), it holds q k ≡ F(sk ). The system matrix Mk ≡
M(sk ) is the Jacobian M(s) of F(s) at s = sk , i.e., Mk = ∂ F(sk )/∂s.
To simplify the representation, let g(v, w) = 0 be the abbreviation of the bound-
ary conditions in (4.8). Using the chain rule, we obtain for the Jacobian M(s) the
following expression

∂ ∂ ∂ ∂
M(s) = g(s, u(b; s)) = g(v, w) + g(v, w) u(b; s) .
∂s ∂v ∂w ∂s v=s
w = u(b; s)
Considering the IVP (4.11), the function ∂u(b; s)/∂s can be computed by the IVP

d ∂ ∂ ∂ ∂
u(x; s) = f (x, u) · u(x; s), u(a; s) = I. (4.17)
d x ∂s ∂u u=u(x;s) ∂s ∂s
Now, if we introduce the following notation

∂
Ba,k ≡ Ba (sk ) ≡ g(v, w) s = sk ,
∂v v=s
w = u(b; s)

∂
Bb,k ≡ Bb (sk ) ≡ g(v, w) s = sk ,
∂w v=s
w = u(b; s)
∂ ∂
X (t; sk ) ≡ u(x; s)s=sk , A(x; sk ) ≡ f (x, u) s = sk ,
∂s ∂u u = u(x; s)
then Mk and q k can be written in the form
Mk = Ba,k + Bb,k X ke , q k = F(sk ), (4.18)
where X ke ≡ X (b; sk ). Note, that with our notation the matrix Mk is similarly struc-
tured as the matrix M = Ba + Bb X e of the standard simple shooting method for
linear BVPs (see [57, 63]).
According to the formula (4.17), the matrix X (x) ≡ X (x; sk ) is the solution of
the matrix IVP
X (x) = A(x; sk ) X (x), a ≤ x ≤ b,

(4.19)
X (a) = I.
Problem (4.19) is also known as the variational problem.

In practice, it is often very difficult and time-consuming to calculate analyti-
cally the matrices A(x; s). Therefore, in the existing implementations of the simple
shooting method the matrix X (x; s k ) is approximated by forward-difference approx-
imations. For this purpose, u(x; sk ) is determined for s = sk from the IVP (4.11).
In addition, n further functions u( j) ≡ u( j) (x; sk ), j = 1, . . . , n, are computed from
the slightly modified IVPs
d u( j)
(x) = f (x, u( j) (x)), a ≤ x ≤ b,
dx (4.20)
u( j) (a) = sk + h e( j) , j = 1, . . . , n,
where e( j) ≡ (0, . . . , 1, . . . , 0)T √

denotes the jth unit vector in Rn , and h
= 0 is a
small positive number with h ≈ ν. Here, ν is the relative machine accuracy (see,
e.g., [58]). Obviously, if we define
û( j) (x; sk ) ≡ (u( j) (x; sk ) − u(x; sk ))/ h,

then
X̂ (x; sk ) ≡ [û(1) (x; sk ), û(2) (x; sk ), . . . , û(n) (x; sk )] (4.21)
is a sufficiently accurate difference approximation of X (x; sk ).

With regard to the convergence of the Newton method (4.15), (4.16), the following
statement applies.
Theorem 4.2 (Theorem of Newton and Kontorovich)

Let F(s) be continuously differentiable in the ball
S(s0 , ρ) ≡ {s ∈ Rn : s − s0 ≤ ρ}.
Assume that the inverse of the Jacobian M0 ≡ M(s0 ) exists and there are three
constants α1 , α2 , α3 ≥ 0 such that
M0−1 F(s0 ) ≤ α1 , M0−1 ≤ α2 ,

M1 − M2 ≤ α3 s1 − s2 for s1 , s2 ∈ S(s0 , ρ).
Let σ ≡ α1 α2 α3 . If
√
σ ≤ 1/2 and ρ ≥ ρ− ≡ (1 − 1 − 2σ)/(α2 α3 )
then it holds
1. there exists a unique root s ∈ S(s0 , ρ− ), and
2. the sequence of Newton iterates {sk }∞ ∗
k=0 converges to s .
If σ < 1/2, then it holds

3. s∗ is the unique
√ solution in S(s0 , min{ρ, ρ+ }), where
ρ+ ≡ (1 + 1 − 2σ)/(α2 α3 ), and
4. sk − s∗ ≤ (2σ)2 /(α2 α3 ), k = 0, 1, …
k
Proof See, e.g., [32].
It can be inferred from this theorem that the condition σ ≤ 1/2 will only be
satisfied if the starting vector s0 is located in a sufficiently small neighborhood of the
exact solution s∗ (this follows from the fact that in this case F(s0 ) is very small).
Thus, the Newton method is only a locally convergent method (see also [58]).
Let us assume that the right-hand side f of the differential equation in (4.8) is
uniformly Lipschitz continuous on [a, b] × Ω, i.e., there exists a constant L ≥ 0
such that
f (x, y1 ) − f (x, y2 ) ≤ L y1 − y2
for all (x, yi ) ∈ [a, b] × Ω, i = 1, 2.

For the next claim we need Gronwall’s Lemma.

Lemma 4.3 (Gronwall’s Lemma)
Let u(x) and v(x) be two nonnegative continuous functions on the interval [a, b].
Assume that C ≥ 0 is a constant and it holds
x
v(x) ≤ C + v(τ )u(τ )dτ , for a ≤ x ≤ b.
a
Then
x
v(x) ≤ C exp u(τ ) dτ , for a ≤ x ≤ b.
a
If, in particular C = 0 then v(x) ≡ 0.

Using the integral representation of the differential equation (4.11) and Gronwall’s
Lemma, we obtain
b
u(b; s1 ) − u(b; s2 ) ≤ s1 − s2 + L u(x; s1 ) − u(x; s2 ) d x
a
≤ e L(b−a) s1 − s2 .
Obviously, the matrix X (x; s) depends similarly sensitive on the initial vector s as
u(x; s). Thus, it can be expected that the constant α3 (see Theorem 4.2) is of the
magnitude e L(b−a) . In other words, it must be assumed that the radius of convergence
satisfies ρ = O(e−L(b−a) ). This is a serious limitation, if L(b − a) is not sufficiently
small. In particular, we have to expect considerable difficulties when the simple
shooting method is used over a large interval [a, b].
We will now introduce the following well-known test problem to demonstrate the
above mentioned difficulties.
Example 4.4 (Troesch’s problem)
Consider the following nonlinear BVP (see, e.g., [56, 117, 120, 122])
y (x) = λ sinh(λ y(x)), 0 ≤ x ≤ 1,

(4.22)
y(0) = 0, y(1) = 1,
where λ is a positive constant. The larger the constant λ is, the more difficult it is
to determine the solution of (4.22) with the simple shooting method. The solution is
almost constant for x ≥ 0, i.e., y(x) ≈ 0, and near to x = 1 it increases very rapidly
to fulfill the prescribed boundary condition y(1) = 1. In Fig. 4.1 the graph of the
exact solution y(x) is represented for λ = 10.
Let us use the simple shooting method to determine an approximation of the
unknown exact initial value y (0) = s ∗ . Then, the solution of the BVP (4.22) can
be computed from the associated IVP. About this initial value is known that it holds
s ∗ > 0, but s ∗ is very small. If we are trying to solve the IVP
u (x) = λ sinh(λ u(x)), u(0) = 0, u (0) = s, (4.23)
for s > s ∗ with a numerical IVP-solver, we must recognize that the integration step-
size h tends to zero and the integration stops before the right boundary x = 1 of the
given interval is reached. The reason for this behavior is a singularity in the solution
u(x; s) of the IVP (4.22) at
∞
1 dξ 1 8
xs = ≈ ln .
λ 0 s 2 + 2 cosh(ξ) − 2 λ |s|
In order to be able to perform the integration of the IVP over the complete interval
[a, b], the initial value s must be chosen such that the value xs is located behind
x = 1. This leads to the condition
|s| ≤ sb ≡ 8 e−λ , (4.24)
which is a very strong restriction on the initial value, and thus also on the starting
value for the Newton method.
The exact solution of Troesch’s problem can be represented in the form
∗
∗ 2 s sn(λx, k) (s ∗ )2
y(x) = u(x; s ) = arcsinh , k2 = 1 − ,
λ 2 cn(λx, k) 4
where sn(λx, k) and cn(λx, k) are the Jacobian elliptic functions (see [6]) and the
parameter s ∗ is the solution of the equation

2 s sn(λ, k)
arcsinh = 1.
λ 2 cn(λ, k)
In Table 4.1, the numerical results for Troesch’s problem obtained with the code
.
RWPM (parameter INITX = simple shooting) are presented (see [59]). The IVPs
have been integrated with the IVP-solver RKEX78 (see [98, 113]) using an error
tolerance of 10−9 . For the numerical solution of the nonlinear algebraic equations
we have used a damped Newton method with an error tolerance of 10−7 .
Here, s ∗ is the numerically determined approximation for the exact initial slope

u (0). Moreover, it denotes the number of Newton steps and NFUN is the number
of calls of the right-hand side of the ODE (4.22). For the parameter value λ = 9,
the simple shooting method provided an error code ierr=8, which means that the
BVP could not be solved within the required accuracy.
In Fig. 4.1, the solutions of the IVP (4.23) for λ = 8 and different values of s are
plotted. In that case, the bound in (4.24) is sb = 2.68370102 . . . × 10−3 .
Table 4.1 Results of the simple shooting method for Troesch’s problem
λ it NFUN s∗
1 3 734 8.4520269 × 10−1
2 4 1626 5.1862122 × 10−1
3 5 2858 2.5560422 × 10−1
4 5 3868 1.1188017 × 10−1
5 11 9350 4.5750461 × 10−2
6 11 10815 1.7950949 × 10−2
7 10 10761 6.8675097 × 10−3
8 11 13789 2.5871694 × 10−3
Fig. 4.1 Solutions of the IVP (4.23) for λ = 8 and different values of s
Now, we have the following situation. If s > sb , the singularity xs is within

the interval [0, 1], and thus, the curves do not reach the right end of the interval.
Moreover, if s < s ∗ = 2.5871694 × 10−3 , then y(1) < 1.
At the end of this section, we will explain the name of the method. In former
times the cannoneers had to solve the following problem with the use of a cannon.
In Fig. 4.2 the picture of a cannon is displayed. It can be shown that the ballistic
trajectory of the cannonball is the solution of a second-order ODE
y (x) = f (x, y(x)), a ≤ x ≤ b.

Fig. 4.2 The origin of the shooting method
The associated initial conditions are the position of the cannon y(a) = α1 and
the unknown slope y (a) = γ of the cannon’s barrel. Now, the cannoneer must
determine the slope (i.e., the missing initial condition of the corresponding IVP)
such that at x = b the trajectory of the cannonball hits the given target y(b) = α2
(i.e., the trajectory of the IVP satisfies this boundary condition, too). In practice, the
cannoneer has solved this problem by trial-and-error. For those who do not like the
military terminology, the name shooting method can be replaced by Manneken Pis
method in reference to the well-known sculpture and emblem of Brussels.
4.3 Method of Complementary Functions
In [63], we have considered the method of complementary functions for linear par-
tially separated BVPs. Before we study nonlinear BVPs with (completely) nonlinear
partially separated boundary conditions of the form
y (x) = f (x, y(x)), a ≤ x ≤ b,

g p ( y(a)) = 0, (4.25)
g q ( y(a), y(b)) = 0,
we will assume that the separated part of the boundary conditions is linear, i.e.,
y (x) = f (x, y(x)), a ≤ x ≤ b,

g p ( y(a)) ≡ Ba(1) y(a) − β (1) = 0, (4.26)
g q ( y(a), y(b)) = 0,
where n = p +q, f : Ω1 → Rn , Ω1 ⊂ (a, b)×Rn , g q : Ω2 → Rq , Ω2 ⊂ Rn ×Rn ,

Ba(1) ∈ R p×n , and β (1) ∈ R p . Let us assume rank(Ba(1) ) = p.
If the simple shooting method is used to compute a solution of the BVP (4.26), in
each step of the Newton method the linear algebraic system (4.16) has to be solved.
For the special boundary conditions in (4.26) this algebraic system reads

Ba(1) Ba(1) sk − β (1)

(2) (2) e ck = , (4.27)
Ba,k + Bb,k Xk g q (sk , u(b; sk ))
(2)
where Bi,k = Bi(2) (sk ) ≡ ∂ g q (sk , u(b; sk ))/∂ y(i), i = a, b.
Using the Q R factorization of the matrix (Ba(1) )T ,

U U
(Ba(1) )T =Q (1) (2)
= (Q |Q ) = Q(1) U, (4.28)
0 0
the vector ck is transformed as follows

wk wk
ck = QQT ck = Q = (Q(1) |Q(2) ) , (4.29)
zk zk
where w k ∈ R p , z k ∈ Rq , Q(1) ∈ Rn× p , Q(2) ∈ Rn×q und U ∈ R p× p .

Substituting (4.29) into (4.27), the left-hand side of (4.27) turns into

U T (Q(1) )T (1) (2) wk
(2) (2) e (Q |Q )
Ba,k + Bb,k Xk zk

ck
⎡ ⎤
U (Q ) Q(1)
T (1) T
U T (Q(1) )T Q(2)
⎢ ⎥ wk
=⎣ Ip 0 p×q ⎦ z
(2) (2) e (2) (2) e k
[Ba,k + Bb,k X k ]Q(1) [Ba,k + Bb,k X k ]Q (2)
and the linear algebraic system reads

UT 0 wk Ba(1) sk − β (1)
(2) (2) e (2) (2) e = .
[Ba,k + Bb,k X k ]Q(1) [Ba,k + Bb,k X k ]Q(2) zk g q (sk , u(b; sk ))
(4.30)
Now, using the forward substitution the unknown wk can be computed numerically
very fast from the following p-dimensional quadratic system with a lower triangular
system matrix
U T wk = Ba(1) sk − β (1) .
For the numerical treatment of the second (block-) row of (4.30), in principle, u(b; sk )
and X ke have to be determined as solutions of the n + 1 IVPs (4.11) and (4.19).
However, in (4.30) the matrix X ke occurs only in combination with the matrix Q(2)
and with the vector Q(1) wk . The idea is now to approximate the expressions X ke Q(2)
and X ke Q(1) wk by discretized directional derivatives. This can be realized as follows.
Assuming that the solution u(x1 ; x0 , s) of the IVP u = f (x, u), u(x0 ) = s, has
4.3 Method of Complementary Functions 133
already been computed, then Algorithm 1.1 generates an approximation X R ∈ Rn× j

of X (x1 ; x0 , s)R, with X (x 1 ; x0 , s) ≡ ∂u(x1 ; x0 , s)/∂s ∈ Rn×n und R ∈ Rn× j .
Algorithm 1.1 Approximation of directional derivatives

Input: dimension j, interval limits x0 and x1 ,
point s, directions R, step-size h,
subroutine for the computation of f (x, y)
Set: ε = ( s + 1)h
For i = 1 : j Do
Evaluate: u(x1 ; x0 , s + εRe(i) )
{as solution of the IVP
u = f (x, u), u(x0 ) = s + εR e(i) }
Set: X R(:, i) = (u(x1 ; x0 , s + εR e(i) ) − u(x1 ; x0 , s))/ε
Endfor
Output: X R
In Algorithm 1.1, h > 0 is a small constant and e(i) denotes the ith unit vector
in R j .
The Algorithm 1.1 can now be used to determine approximations of X ke Q(2) and
e (1)
X k Q wk . To do this, in the first case we set
x0 = a, x1 = b, j = q, s = sk , R = Q(2) ,
and in the second case
x0 = a, x1 = b, j = 1, s = sk , R = Q(1) wk .
needs only theintegration of q + 2 IVPs. More

It can easily be seen that this strategy
precisely, the vectors u(b; sk ), u b; sk + εQ(2) e(i) , i = 1, . . . , q, and u(b; sk +
εQ(1) wk ) have to be computed. The shooting method described above (including
Algorithm 1.1) is called the Newton form of the nonlinear method of complementary
functions.
In the nonlinear method of complementary functions the total number of integra-
tions as well as the numerical effort for the algebraic operations can be reduced even
further. To demonstrate this, let us return to the algebraic system (4.30). First, we
modify the vector sk such that the relation
Ba(1) sk = β (1) (4.31)
is satisfied. The general solution ŝk of this underdetermined system can be written
in the form

ŝk = (Ba(1) )+ β (1) + I − (Ba(1) )+ Ba(1) ω, ω ∈ Rn , (4.32)
where (Ba(1) )+ denotes the Moore–Penrose pseudoinverse of Ba(1) (see, e.g., [42]).
Since rank(Ba(1) ) = p, it holds
ŝk = Q(1) U −T β (1) + Q(2) (Q(2) )T ω. (4.33)
If we set ω = sk , formula (4.33) transforms the vector sk into a vector ŝk , which
satisfies the condition (4.31). Now, in the algebraic equations (4.30) we use the
transformed vector ŝk instead of sk . Obviously, we obtain wk = 0. This implies in
turn that the n-dimensional linear algebraic system (4.30) is reduced to the following
q-dimensional linear algebraic system for the vector z k ∈ Rq :
M̂k z k = q̂ k , (4.34)
where
M̂k ≡ [Ba(2) (ŝk ) + Bb(2) (ŝk )X (b; ŝk )]Q(2) and q̂ k ≡ g q (ŝk , u(b; ŝk )).
Taking into account formula (4.29), the new iterate is determined according to the
formula
ŝk+1 = ŝk − Q(2) z k . (4.35)
Note that each further iterate satisfies the condition (4.31) automatically, i.e., it holds
Ba(1) ŝk+1 = Ba(1) ŝk −Ba(1) Q(2) z k = β (1) − U T (Q(1) )T Q(2) z k = β (1) .

β (1) 0 p×q
Thus, in exact arithmetic, it is only necessary to apply the transformation (4.33) to

the initial guess s0 . However, since in numerical computations rounding errors are
unavoidable, the transformation (4.33) should be realized in each νth iteration step
(ν ≈ 10) to counteract numerical instabilities.
As in the Newton form of the method of complementary functions, the matrix
X (b; ŝk )Q(2) (see formula (4.34)) can be computed by discretized directional deriv-
atives. To do this, we set
x0 ≡ a, x1 ≡ b, j ≡ q, s ≡ ŝk , R ≡ Q(2) ,
and apply Algorithm 1.1. For the realization of this approximation q + 1 IVPs
to be solved, namely one integration for u(b; ŝk ) and q integrations for
have
u b; ŝk + εQ(2) e(i) , i = 1, . . . , q. Thus, the number of integrations is reduced
from n + 1 (simple shooting method) to q + 1 (method of complementary func-
tions). The shooting technique, which is based on the formulas (4.28), (4.32)–(4.35)
(including Algorithm 1.1), is called the standard form of the (nonlinear) method of
complementary functions. To be consistent with the linear method of complementary
functions we also denote the columns of X (b; s)Q(2) as complementary functions.
4.3 Method of Complementary Functions 135
Table 4.2 Results of the method of complementary functions for Troesch’s problem
λ it NFUN s∗
1 3 558 8.4520269 × 10−1
2 4 1082 5.1862122 × 10−1
3 5 2092 2.5560422 × 10−1
4 5 3663 1.1188017 × 10−1
5 11 6440 4.5750462 × 10−2
6 12 8310 1.7950949 × 10−2
7 12 9545 6.8675097 × 10−3
8 16 15321 2.5871694 × 10−3
Example 4.5 (once again: Troesch’s problem)

We want to look once again at Troesch’s problem, which we have considered in
Example 4.4. If the standard form of the method of complementary functions (pro-
.
gram RWPM with the input parameter INITX = method of complementary functions)
is applied to the BVP (4.22), the results presented in Table 4.2 are produced.
Here, s ∗ is the numerically determined approximation for the exact initial slope
u (0). Moreover, it denotes the number of Newton steps and NFUN is the number
of calls of the right-hand side of the ODE (4.22). For the parameter value λ = 9 the
method of complementary functions provided an error code ierr=3, which means
that a descent direction could not be determined.
A comparison with the results of the simple shooting method (see Table 4.1) shows
that the theoretically predicted savings in computational costs could be achieved for
parameter values up to λ = 5. For parameter values larger than 5 the method of
complementary functions requires more Newton steps to guarantee the prescribed
accuracy. Therefore, the number of IVPs, which have to be solved, increases (see the
column NFUN).
Both forms of the nonlinear method of complementary functions can also be used
to solve linear BVPs with partially separated boundary conditions. Here, only one
iteration step is theoretically necessary. If additional iteration steps are performed the
accuracy of the results is increased compared with the linear method of complemen-
tary functions. These additional steps can be considered as an iterative refinement of
s1 , which is often used in the course of the numerical solution of systems of linear
algebraic equations.
4.4 Multiple Shooting Method
The basic idea of the multiple shooting method is to subdivide the given interval
[a, b] into m subintervals
a = τ0 < τ1 < · · · < τm−1 < τm = b. (4.36)

Fig. 4.3 Example of an intermediate step of the multiple shooting method and the exact solution
y(x)
The grid points τk are called shooting points. Now, on each subinterval [τ j , τ j+1 ],
0 ≤ j ≤ m − 1, an IVP will be defined
uj (x) = f (x, u j (x)), τ j ≤ x ≤ τ j+1 ,

(4.37)
u j (τ j ) = s j ∈ Rn .
Let us denote the solution of (4.37) as u j ≡ u j (x; s j ). By a suitable choice of the

unknown vectors s j in (4.37), the solutions u j (x; s j ) on the subintervals [τ j , τ j+1 ],
j = 0, . . . , m − 1, are composed such that a continuous function results, which is
continuous on the entire interval [a, b] and satisfies not only the differential equation,
but also the boundary conditions in (4.8); see Fig. 4.3.
Since we require the continuity of the solution at intermediate grid points, the
following m − 1 matching conditions must be satisfied:
u j (τ j+1 ; s j ) − s j+1 = 0, j = 0, . . . , m − 2. (4.38)
In addition, the solution must fulfill the boundary conditions, i.e.,
g(s0 , um−1 (b; sm−1 )) = 0. (4.39)
Writing the boundary conditions (4.39) first, and then the matching conditions (4.38),
we obtain the following mn-dimensional system of nonlinear algebraic equations for
the unknown initial vectors s j ∈ Rn in (4.37)
F (m) (s(m) ) = 0, (4.40)

4.4 Multiple Shooting Method 137
where
⎛ ⎞ ⎛ ⎞
s0 g(s0 , um−1 (b; sm−1 ))
⎜ s1 ⎟ ⎜ u0 (τ1 ; s0 ) − s1 ⎟
⎜ ⎟ ⎜ ⎟
(m) ⎜ s2 ⎟ (m) (m) ⎜ u1 (τ2 ; s1 ) − s2 ⎟
s ≡⎜ ⎟, F (s )≡⎜ ⎟.
⎜ .. ⎟ ⎜ .. ⎟
⎝ . ⎠ ⎝ . ⎠
sm−1 um−2 (τm−1 ; sm−2 ) − sm−1
The numerical algorithm defined by the formulas (4.36)–(4.40) is called multiple

shooting method. Sometimes it is also denoted as parallel shooting method. This
name makes reference to the fact that on the individual subintervals of the grid (4.36)
the IVPs (4.37) can be integrated in parallel, i.e., independently of one another. Thus,
the multiple shooting method is the prototype of numerical parallel algorithm.
If y∗ (t) denotes the exact (isolated) solution of the BVP (4.8), whose existence
we assume here, the following claim can be shown. Under the assumption that f is
sufficiently smooth there exists for each s∗j ≡ y∗ (τ j ) a n-dimensional neighborhood
U ∗j such that the IVP (4.37) has a solution u j (t; s j ), which exists on the entire interval
[τ j , τ j+1 ]. Thus, there is also a mn-dimensional neighborhood
Û ∗ ≡ U0∗ × U1∗ × · · · × Um−1

∗
of the vector s(m) ∗ ∗

∗ ≡ (s 0 , . . . , s m−1 ) ∈ R
T mn
where F (m) : Û ∗ → Rmn is defined and
continuously differentiable.
For the solution of the system of nonlinear algebraic equations (4.40) a (mul-
tidimensional) Newton-like method is commonly used. In the kth iteration step of
the Newton method the following system of mn linear algebraic equations has to be
solved
Mk(m) c(m)
k = q (m)
k , (4.41)
where the Jacobian Mk(m) ≡ M (m) (s(m)

k ) ∈ R
mn×mn
and the vector of the right-hand
(m) (m) (m)
side q k ≡ q (sk ) ∈ R are defined as
mn
⎛ e ⎞
Ba,k Bb,k X m−1,k
⎜ X 0,k
e
−I ⎟
⎜ ⎟
⎜ e
−I ⎟
Mk(m) ≡⎜ X 1,k (m) (m)
⎟ and q k ≡ F (m) (sk ),
⎜ .. .. ⎟
⎝ . . ⎠
e
X m−2,k −I
with
∂u j (τ j+1 ; s j,k )
X ej,k ≡ ∈ Rn×n , j = 0, . . . , m − 1,
∂s j
and
∂ g(s0,k , um−1 (b; sm−1,k ))
Bi,k ≡ Bi (s(m)
k )≡ ∈ Rn×n , i = a, b.
∂ y(i)
The new iterate is then calculated as
s(m) (m) (m)

k+1 = s k − ck .
The system matrix Mk(m) has the same structure as the matrix M (m) of the multiple
shooting method for linear BVPs (see, e.g., [57, 63]). But the difference is that the
matrices X ej,k , j = 0, . . . , m − 1, and Ba,k , Bb,k now depend on the initial vectors
s j,k .
It can be shown that Mk(m) is nonsingular if and only if the matrix
M(s0,k , sm−1,k ) ≡ Ba,k + Bb,k X m−1,k

e
is nonsingular (see [57]).

For s0,k = sk , sm−1,k = u(τm−1 ; sk ), the matrix M(s0,k , sm−1,k ) coincides
with the matrix Mk of the simple shooting method (see formula (4.18)). Thus, if
y∗ (x) = u(x; s∗ ) is an isolated solution of (4.8), then the matrix M(s∗ , u(τm−1 ; s∗ ))
is nonsingular. This applies also to the matrix M (m) (s(m)
∗ ) with
s(m) ∗ ∗ ∗ T
∗ ≡ (s , u(τ1 ; s ), . . . , u(τm−1 ; s )) .
Setting
∂f
A j (x; s j,k ) ≡ (x, u j (x; s j,k )), j = 0, . . . , m − 1,
∂u j
and using the notation Z j (x; s j,k ) for the solution of the (matrix) IVP
Ż j (x) = A j (x; s j,k )Z j (x), τ j ≤ x ≤ τ j+1 ,

(4.42)
Z j (τ j ) = I,
we have X ej,k = Z j (τ j+1 ; s j,k ). This shows that the matrix X ej,k can be computed by
integrating the n IVPs (4.42). The analytical determination of the matrices A j (x; s j,k )
that appear on the right-hand side of (4.42) is very time-consuming and sometimes
not possible. Analogous to the simple shooting method (see formulas (4.20) and
(4.21)) it is more efficient to approximate the matrices X ej,k , j = 0, . . . , m − 1, by
finite differences. For this purpose, at each iteration step, mn additional IVPs must
be integrated. Thus, per step, a total of m(n + 1) IVPs have to be solved numerically.
There is still the problem that the user has to supply the matrices Ba,k and Bb,k . In
general, these matrices are generated by analytical differentiation. In practice, BVPs
are frequently encountered with linear boundary conditions (4.9). In this case, we
have Ba,k = Ba and Bb,k = Bb .
Essentially, there are two numerical techniques to solve the system of linear alge-
braic equations (4.41). The first method is based on the so-called compactification
(see [33, 34, 117]). The idea is to transform the system (4.41) into an equivalent
linear system
M̄k(m) c(m)
k = q̄ (m)
k , (4.43)
whose system matrix has a lower block-diagonal structure. Let the bordered matrix
of the original system be
⎛ ⎞
Ba,k e
Bb,k X m−1,k q (m)
0,k
⎜ X e −I q (m) ⎟
⎜ 0,k ⎟
⎜ 1,k
⎟
[Mk |q k ] ≡ ⎜
(m) (m) e
X 1,k −I q (m) ⎟. (4.44)
⎜ 2,k ⎟
⎜ .. .. ⎟
⎝ . . ⎠
e
X m−2,k −I q (m)
m−1,k
In a first step, the block Gaussian elimination is used to eliminate the last block-
element in the first row of Mk(m) . To do this, the mth row of [Mk(m) |q (m)
k ] is multiplied
e
from the left by Bb,k X m−1,k , and the result is added to the first row. In the second step,
the (m − 1)th row is multiplied from the left by Bb,k X m−1,k e e
X m−2,k , and the result
is added again to the first row. This procedure is continued for i = m − 2, . . . , 2
by applying an analogous elimination step to the ith row. At the end of this block
elimination strategy the following equivalent bordered matrix is obtained
⎛ ⎞
Sk q̄ (m)
0,k
⎜ X e −I q (m) ⎟
⎜ 0,k ⎟
⎜ 1,k
⎟
[ M̄k(m) |q̄ (m) ] ≡ ⎜ e
X 1,k −I q (m) ⎟, (4.45)
k ⎜ 2,k ⎟
⎜ .. .. ⎟
⎝ . . ⎠
e
X m−2,k −I q (m)
m−1,k
with
Sk = Ba,k + Bb,k X m−1,k

e e
X m−2,k e
X m−3,k · · · X 1,k
e e
X 0,k ,

q̄ (m) (m) e (m) e (m)
0,k = q 0,k + Bb,k X m−1,k q m−1,k + X m−2,k q m−2,k

+X m−2,k
e e
X m−3,k q (m)
m−3,k + · · · + X e
X e
m−2,k m−3,k · · · X e
X e
q (m)
2,k 1,k 1,k .
The solution of the corresponding system of linear equations is realized as follows:

1. determination of the solution c(m)
0,k ∈ R of the n-dimensional system
n
Sk c(m) (m)
0,k = q̄ 0,k , and (4.46)
2. performing the recursion
c(m) (m) (m)

j+1,k = X j,k c j,k − q j+1,k ,
e
j = 0, 1, . . . , m − 2. (4.47)
Having determined the vector c(m)

k = (c(m) (m)
0,k , . . . , cm−1,k ) , the new iterate of the
T
Newton method is calculated according to the formula
s(m) (m) (m)

k+1 = s k − ck .
However, in [55] it is shown that the recursion (4.47) is a numerically instable process.
But it is possible to reduce slightly the accumulation of rounding errors in this solution
technique by a subsequent iterative refinement (see [34]).
A more appropriate treatment of the system of linear equations (4.41) is based on
a LU factorization of the matrix Mk(m) with partial pivoting, scaling, and an iterative
refinement. For general systems of linear algebraic equations it is shown in [115]
that this strategy leads to a numerically stable algorithm. Without taking into account
any form of packed storage, the LU factorization of Mk(m) requires a storage space
for m × [n 2 m] real numbers. The implementation RWPM of the multiple shooting
method (see [59, 60, 61]) is based on this linear equation solver. However, in RWPM
a packed storage is used where only a maximum of 4 × [n 2 (m − 1)] real numbers
have to be stored.
It is also possible to use general techniques for sparse matrices (various methods
are available in the MATLAB package) to solve the linear shooting equations. Our
tests with the MATLB package have shown that the storage for the LU factorization
using sparse matrix techniques is also below the limit 4 × [n 2 (m − 1)] mentioned
above. For the case m = 10 and n = 10 a typical situation for the fill-in is graph-
ically presented in Fig. 4.4. The value nz is the number of nonzero elements in the
corresponding matrices.
The convergence properties of the iterates {s(m) ∞
k }k=0 can be derived directly from
the application of Theorem 4.2 to the system of nonlinear algebraic equations (4.40).
To show that the radius of convergence of the Newton method will increase if the
multiple shooting method is used instead of the simple shooting method, we represent
the multiple shooting method as the simple shooting method for an (artificially)
enlarged BVP. For this, we define a new variable τ on the subinterval [τ j , τ j+1 ] as
t − τj
τ≡ , j ≡ τ j+1 − τ j , j = 0, . . . , m − 1. (4.48)
j
With y j (τ ) ≡ y(τ j + τ j ), and
f j (τ , y j (τ )) ≡ j f (τ j + τ j , y j (τ )), j = 0, . . . , m − 1, (4.49)
Fig. 4.4 Example for the fill-in the LU factorization
the differential equation in (4.8) can be written in the form

⎛ ⎞ ⎛ ⎞
y0 f0
⎜ ⎟ ⎜ ⎟
ȳ (τ ) = f̄ (τ , ȳ(τ )), 0 ≤ τ ≤ 1, ȳ ≡ ⎝ ... ⎠ , f̄ ≡ ⎝ ... ⎠ . (4.50)
ym−1 f m−1
The boundary conditions and the matching conditions are

⎛ ⎞
g( y0 (0), ym−1 (1))
⎜ y1 (0) − y0 (1) ⎟
⎜ ⎟
ḡ( ȳ(0), ȳ(1)) ≡ ⎜ .. ⎟ = 0. (4.51)
⎝ . ⎠
ym−1 (0) − ym−2 (1)
Obviously, the BVP (4.50), (4.51) is an equivalent formulation of the original BVP
(4.8). This means that the application of the simple shooting method to the trans-
formed BVP is identical with the multiple shooting method applied to the original
BVP. We can now formulate the associated IVP as
ū (τ ) = f̄ (τ , ū(τ )), 0 ≤ τ ≤ 1,

(4.52)
ū(0) = s̄,
where the initial vector s̄ is determined according to the formula (4.51) by
F̄(s̄) ≡ ḡ(s̄, ū(1; s̄)) = 0. (4.53)
It is now possible to apply the statements about the radius of convergence of the
Newton method, which are given in Sect. 4.2 for the solution of the original BVP
by the simple shooting method to the transformed BVP. This yields a radius of
convergence ¯ of the multiple shooting method with
¯ ≡ O(e−L ), ≡ max (τ j+1 − τ j ).

0≤ j≤m−1
In other words, the radius of convergence of the Newton method can be increased if the
multiple shooting method is used instead of the simple shooting method. Moreover,
the radius of convergence increases exponentially if the subintervals are reduced
successively.
In most implementations of the multiple shooting method, the Newton method or
a Newton-like method with discretized Jacobian is not implemented in the standard
form described above (see, e.g., the computer program RWPM presented in [59]).
Instead, the Newton method is combined with damping and regularization strategies
(see [58]) to force global convergence of the nonlinear equation solver. Further-
more, it makes sense not to compute the very costly finite difference approximation
of the Jacobian in each iteration step. This becomes possible by using the rank-1
modification formula of Sherman and Morrison, which leads to Broyden’s method
(see [58]).
In the above-mentioned multiple shooting code RWPM a damped Quasi-Newton
method is used, which is transformed automatically into a regularized Quasi-Newton
method if the descending behavior is insufficient or the Jacobian is nearly singular.
The kth iteration step is of the form
−1
s(m)
k+1 = s (m)
k − γ (m)
k (1 − λk )D + λk Mk
k
F (m) (s(m)
k ). (4.54)
Here, λk is the regularization parameter with 0 < λk ≤ 1. If λk = 1 we have a damped

Newton method. The parameter γk is the damping parameter with 0 < γk ≤ 1. The
matrix D k ∈ Rnm×nm is a diagonal matrix with D k = diag( Mk(m) e(i) 2 + 10−3 ),
where e(i) denotes the ith unit vector. The descending condition used is
F (m) (s(m)
k+1 ) 2 ≤ (1 − 2δγk ) F
2 (m) (m) 2
(sk ) 2 , δ ∈ (0, 1/2). (4.55)
A precise description of this Quasi-Newton method can be found in [57].

Up to now, we have made the assumption that the shooting points τ j , j =
0, . . . , m, are given by the user and they are not changed when the multiple shooting
method is carried out. But this is not always an appropriate procedure. The multiple
shooting method is based on two different types of grids, a fine inner grid, which is
produced by the IVP-solver, and a coarse outer grid, which consists of the shooting
points. The inner grid is always generated automatically and cannot be accessed by
the user. In order to keep the computational costs low, besides both endpoints a and b,
additional shooting points should be set only when this is necessary. Since the shape
of the solution curve is not known a priori, the outer grid (4.36) must be generated
automatically, too. The different implementations of the multiple shooting method
each have their own strategy and heuristics for an automatic control of the shooting
points. The practical experience of the program developer plays an important role in
this context.
It seems reasonable that the user proposes a subdivision of the interval [a, b],
which can be modified by the algorithm. In principle, it must be possible that in the
first iteration step the proposed shooting points are replaced by more appropriate
points. Additional shooting points should be generated automatically during the
solution process of the system of nonlinear algebraic equations (4.40) at two places.
Namely, when F (m) (s(m) 0 ) is evaluated, where s 0
(m)
is the starting vector, and when
the Jacobian is approximated by a finite difference approximation δ F (m) (s(m) k ; h k ),
k = 0, 1, . . .. For the selection of a shooting point three criteria are significant, which
we want to describe in the following. Here, the two constants
√
c1 ≡ 1/ KPMACH, KPMACH smallest positive machine number,
√
c2 ≡ 1/ EPMACH, EPMACH relative machine accuracy,
play an important role. The criteria are:

1. Monitoring the growth of the solution u j (x) of the IVP (4.37):
the point
x = τ ∗ will be added as a new shooting point, if it holds u j (τ ∗ ) ≥
max c1 , ( s j + 1)c2 ,
2. Monitoring the partial derivatives X ej,k ( j = 0, . . . , m − 1):
a new shooting point τ ∗ will be placed in the subinterval [τ j , τ j+1 ], if
max X ej,k e(i) / min X ej,k e(i) ≥ c2 ,

i=1,...,n i=1,...,n
3. A new shooting point τ ∗ = x̄ will be added, if the IVP-solver is terminated

irregularly at x = x̄ because of:
a. the required accuracy could not be obtained,
b. f (x̄, u) could not be evaluated,
c. the upper bound for the number of integrations has been exceeded.
Table 4.3 Results of the multiple shooting method for Troesch’s problem
λ m it NFUN s∗
5 10 13 8,568 4.5750462 × 10−2
10 10 9 12,641 3.5833778 × 10−4
15 10 11 26,616 2.4445130 × 10−6
20 10 13 34,425 1.6487732 × 10−8
25 11 21 83,873 1.1110272 × 10−10
30 13 16 78,774 7.4860938 × 10−13
35 13 24 143,829 5.0440932 × 10−15
40 14 25 169,703 3.3986835 × 10−17
45 15 16 273,208 2.2900149 × 10−19
Example 4.6 (once more again: Troesch’s problem)

Let us look at Troesch’s problem that was studied in Examples 4.4 and 4.5. Here,
we present the results, which we have obtained with the multiple shooting method
.
(program RWPM with the input parameter INITX = multiple shooting). For the
parameter values, λ = 5 : 5 : 40 the IVPs have been integrated with the numerical
IVP-solver RKEX78 (see [98, 113]). In order to guarantee the demanded accuracy
for λ = 45, it was necessary to exchange the IVP-solver. We used as IVP-solver the
Bulirsch-Gragg-Stoer extrapolation algorithm BGSEXP (see [66]). As start selection
of shooting points we have chosen an equidistant net with m = 10, and the starting
points were y(τi ) = 0, i = 0, . . . , 9. In Table 4.3, we denote by m the number of
shooting points. If m > 10 the code has added automatically new shooting points
to the initial net. As before, s ∗ is the numerically determined approximation for the
exact initial slope u 0 (0), it denotes the number of Newton steps and NFUN is the
number of calls of the right-hand side of the ODE (4.22).
Example 4.7 In Sect. 3.4 the following nonlinear BVP on an infinite interval is con-
sidered (see formula (3.111)):
y (x) = −2y(x)y (x), y(0) = 0, lim y(x) = 1. (4.56)

x→+∞
To treat this problem by numerical techniques, we transform (4.56) into a series of

BVPs on finite intervals
y (x) = −2y(x)y (x), y(0) = 0, y(bi ) = 1, i = 1, 2, . . . , (4.57)
where {bi } is a sequence of increasing positive real numbers. Now, (4.57) must be
written in the standard form (4.2) as

y1 (x) y2 (x)

= ,
y2 (x) −2y1 (x)y2 (x)

(4.58)
10 y1 (0) 00 y1 (bi ) 0
+ = ,
00 y2 (0) 10 y2 (bi ) 1
i.e.,
y (x) = f (x, y), Ba y(0) + Bb y(bi ) = β.
Let y(x; bi ) be the solution of (4.58). The computation of y(x; bi ), i = 1, 2, . . ., is

stopped if y(x; b N ) satisfies y N < T O L, where T O L is a given tolerance,
|y1 (τ j ; bi ) − y1 (τ j ; bi−1 )|
yi ≡ max ,
j=0,...m |y1 (τ j ; bi )| + eps
and eps is a constant prescribed by the Matlab (eps = 2.2 . . . × 10−16 ). Then, the
approximate solution of the BVP (4.56) is
!
y1 (x; b N ), x ∈ [0, b N ]
y(x) ≡
1, x > bN
In our numerical experiments, we have used the multiple shooting method with
m = 20 to solve the BVPs (4.58). Moreover, we have set T O L = 10−8 . In Table 4.4,
the corresponding results are presented. Here, ε( yi ) compares the approximate solu-
tion with the exact solution, i.e.,
| tanh(τ j ) − y1 (τ j ; bi )|
ε( yi ) ≡ max .
j=0,...,m | tanh(τ j )| + eps
The results show that the numerical solution for b6 = 9 is a good approximation
of the solution of the BVP (4.56).
Table 4.4 Numerical results for problem (4.57)

i bi ε( yi ) yi
1 4 1.308525 × 10−3
2 5 1.777780 × 10−4 1.130546 × 10−3
3 6 2.407612 × 10−5 1.536982 × 10−4
4 7 3.258719 × 10−6 2.081733 × 10−5
5 8 4.410273 × 10−7 2.817690 × 10−6
6 9 5.968657 × 10−8 3.813407 × 10−7
4.5 Nonlinear Stabilized March Method
In Sect. 4.3, we have developed two nonlinear variants of the method of complemen-
tary functions for BVPs with nonlinear partially separated boundary conditions
y (x) = f (x, y(x)), a ≤ x ≤ b,

g p ( y(a)) ≡ Ba(1) y(a) − β (1) = 0, (4.59)
g q ( y(a), y(b)) = 0,
where n = p +q, f : Ω1 → Rn , Ω1 ⊂ (a, b)×Rn , g q : Ω2 → Rq , Ω2 ⊂ Rn ×Rn ,

Ba(1) ∈ R p×n , and β (1) ∈ R p . As before, let us assume rank(Ba(1) ) = p.
Here, we will study a segmented version of the method of complementary func-
tions, which solves BVPs of the form (4.59) with less numerical effort than the mul-
tiple shooting method. It is a generalization of the linear stabilized march method
(see, e.g., [63]) to nonlinear BVPs.
The starting point for the following considerations are the equations (4.40)–(4.42)
of the multiple shooting method. It should be recalled that in the kth iteration step of
the multiple shooting method the system of linear algebraic equations (4.41) has to be
solved. As we have seen this linear system is usually solved by Gaussian elimination
and a special version of packed storage. If the boundary conditions are partially
separated as described above the required storage space can be reduced significantly
by special block elimination techniques. As a by-product the number of IVPs, which
have to be integrated, is reduced, too. The latter result is much more important than
the reduction of the storage space of the respective shooting method.
Without changing the solution, we can write the first block equation of system
(4.41) as last block equation. Then, we substitute the special form (4.59) of the
boundary conditions into this algebraic system and obtain
⎛ ⎞⎛ ⎞
e
X 0,k −I c0,k
⎜ .. .. ⎟⎜ . ⎟
⎜ . . ⎟ ⎜ .. ⎟
⎜ ⎟⎜ ⎟
⎜ e
X m−2,k −I ⎟ ⎝ cm−2,k ⎠
⎜ (1)

⎟
⎝ Ba 0 e
⎠ cm−1,k
(2) (2) X m−1,k
Ba,k Bb,k
⎛ ⎞
ue0,k − s1,k
⎜ .. ⎟
⎜ . ⎟
⎜ e ⎟
=⎜
⎜ um−2,k − sm−1,k
⎟,
⎟ (4.60)
⎝ B (1) s0,k − β (1) ⎠
a
g q (s0,k , uem−1,k )
4.5 Nonlinear Stabilized March Method 147
where uej,k ≡ u j (τ j+1 ; s j,k ). The following transformations are based on two orthog-
onal matrices Q k and Q̃ kT . They are constructed such that the matrix of the above
system is transformed into a (2 × 2)-block matrix whose (1,1)-block is a lower
triangular matrix and the (1,2)-block only consists of zeros. These matrices are
⎛ ⎞
Q (1)
0,k Q (2)
0,k
⎜ .. .. ⎟
Qk = ⎝ . . ⎠, (4.61)
Q (1)
m−1,k Q (2)
m−1,k
⎛ ⎞
Ip
⎜ (Q (1) )T ⎟
⎜ 1,k ⎟
⎜ .. ⎟
⎜ . ⎟
⎜ ⎟
⎜ ⎟
⎜ (Q (1)
m−1,k )
T
⎟
Q̃ kT = ⎜ (2) T ⎟, (4.62)
⎜ (Q 1,k ) ⎟
⎜ ⎟
⎜ .. ⎟
⎜ . ⎟
⎜ ⎟
⎝ (Q (2) ⎠
m−1,k )
T
Iq
(1) (2)
where the matrix elements Q i,k ∈ Rn× p and Q i,k ∈ Rn×q are determined recursively
as follows:
• i = 0:
compute the Q R factorization (4.28) of (Ba(1) )T and set
Q 0,k = (Q (1) (2)

0,k | Q 0,k ) ≡ (Q
(1)
| Q(2) ). (4.63)
• i = 1, . . . , m − 1:
e (2)
compute the Q R factorization of X i−1,k Q i−1,k

(2) (2) (1) Ũi−1,k

e
X i−1,k Q i−1,k = (Q̃i−1,k | Q̃i−1,k ) (4.64)
0
and set
(1) (2) (1) (2)
Q i,k = (Q i,k | Q i,k ) ≡ (Q̃i−1,k | Q̃i−1,k ). (4.65)
Now, using the matrix Q k we carry out the coordinate transformation
c(m)
k = Q k d (m) (m)
k , dk ∈ Rmn , (4.66)
and substitute (4.66) into the system (4.60). Then, the resulting equation is multiplied
from the left by Q̃ kT . This yields
( Q̃ kT Mk(m) Q k ) d (m)
k = Q̃ kT q (m)
k , (4.67)
where the right-hand side of (4.60) is denoted with q (m)

k . For the following consid-
erations a detailed representation of this system of algebraic equations is important.
However, due to the lack of space this can only be realized step-by-step. In a first
step we form T ≡ Q̃ kT Mk(m) , i.e.,
⎛ ⎞
Ip
⎜ (Q (1) )T ⎟
⎜ 1,k ⎟ ⎛ ⎞
⎜ .. ⎟ e
X 0,k −I
⎜ . ⎟
⎜ ⎟ ⎜ .. .. ⎟
⎜ ⎟ ⎜ . . ⎟
⎜ (Q (1)
m−1,k )
T
⎟ ⎜ ⎟
⎜ (2) T ⎟·⎜ X e
−I ⎟.
⎜ (Q 1,k ) ⎟ ⎜ (1)
m−2,k
⎟
⎜ ⎟ ⎝ Ba 0 ⎠
⎜ .. ⎟ e
X m−1,k
⎜ . ⎟ (2)
Ba,k (2)
Bb,k
⎜ ⎟
⎝ (Q (2) ⎠
m−1,k )
T
Iq
Thus
⎛ ⎞
Ba(1)
⎜ (Q )T X e −(Q (1) )T
(1) ⎟
⎜ 1,k 0,k 1,k ⎟
⎜ .. .. ⎟
⎜ . . ⎟
⎜ ⎟
⎜ (Q (1)
) X m−2,k −(Q (1) ⎟
m−1,k )
T e T
⎜ ⎟
T = ⎜ (2) T e (2) T
m−1,k
⎟.
⎜ (Q 1,k ) X 0,k −(Q 1,k ) ⎟
⎜ ⎟
⎜ .. .. ⎟
⎜ . . ⎟
⎜ ⎟
⎝ (Q m−1,k ) X m−2,k −(Q (2)
(2) T e
m−1,k )
T ⎠
(2) (2) e
Ba,k Bb,k X m−1,k
In the next step, we consider the relation c(m)

k = Q k d (m) (m)
k . Writing the vector d k ∈
Rmn in the form
d (m)
k = (d (1) (1) (2) (2) (1) p (2)
0,k , . . . , d m−1,k , d 0,k , . . . , d m−1,k ) , d i,k ∈ R , d i,k ∈ R ,
T q
we obtain
c0,k = Q (1) (1) (2) (2)

0,k d 0,k + Q 0,k d 0,k ,
..
.
cm−1,k = Q (1) (1) (2) (2)
m−1,k d m−1,k + Q m−1,k d m−1,k .
On the other hand, using the abbreviations
A(1) = Ba(1) Q (1)

0,k , A(2) = Ba(1) Q (2)
0,k
(1) (1) T e (1) (1) (1) T (1)
Bi,k = (Q i,k ) X i−1,k Q i−1,k , Ci,k = −(Q i,k ) Q i,k , i = 1, . . . , m − 1
(2) (1) T e (2) (2) (1) T (2)
Bi,k = (Q i,k ) X i−1,k Q i−1,k , Ci,k = −(Q i,k ) Q i,k , i = 1, . . . , m − 1
(1) (2) T e (1) (1) (2) T (1)
Fi,k = (Q i,k ) X i−1,k Q i−1,k , G i,k = −(Q i,k ) Q i,k , i = 1, . . . , m − 1
(2) (2) T e (2) (2) (2) T (2)
Fi,k = (Q i,k ) X i−1,k Q i−1,k , G i,k = −(Q i,k ) Q i,k , i = 1, . . . , m − 1
(2) (1) (2) e
(1)
K = Ba,k Q 0,k , L (1) = Bb,k X m−1,k Q (1)
m−1,k
(2) (2) (2) e
K (2) = Ba,k Q 0,k , L (1) = Bb,k X m−1,k Q (2)
m−1,k ,
the system matrix Q̃ kT Mk(m) Q k = T Q k of the algebraic system (4.67) can be rep-
resented in the form
⎛ (1) ⎞
A 0 A(2) 0
⎜ B (1) C (1) (2)
B1,k (2)
C1,k ⎟
⎜ 1,k 1,k ⎟
⎜ .. .. .. .. ⎟
⎜ . . . . ⎟
⎜ ⎟
⎜ (1) (1) (2) (2) ⎟
⎜ 0 Bm−1,k Cm−1,k 0 Bm−1,k Cm−1,k ⎟
T Q k = ⎜ (1) (1) (2) (2) ⎟.
⎜ F1,k G 1,k F1,k G 1,k ⎟
⎜ ⎟
⎜ .. .. .. .. ⎟
⎜ . . . . ⎟
⎜ ⎟
⎝ 0 (1)
Fm−1,k G (1) 0 F (2)
G (2) ⎠
m−1,k m−1,k m−1,k
K (1) L (1) K (2) L (2)
Let us study in more detail the entries of the matrix above. Taking into account
(1)
the factorization (4.28), we get A(1) = U T and A(2) = 0. Obviously, Ci,k = −I p
(2) (2)
und Ci,k = 0, i = 1, . . . , m − 1. The formulas (4.64) and (4.65) imply Bi,k = 0,
(1) (2)
i = 1, . . . , m − 1. The relations G i,k = 0 and G i,k = −Iq , i = 1, . . . , m − 1,
(2)
are obvious again. Finally, from (4.64) and (4.65) we deduce that Fi,k = Ũi−1,k ,
i = 1, . . . , m − 1.
Thus, using the new notation
U0,k ≡ U T , Ui,k ≡ Ũi−1,k , i = 1, . . . , m − 1,

(2) (1) (2) T e (1)
Z 0,k ≡ Ba,k Q 0,k , Z i,k ≡ (Q i,k ) X i−1 Q i−1,k , i = 1, . . . , m − 1,
(2) e ,
Z m,k ≡ Bb,k X m−1 Q (1) (1) T e (1)
m−1,k , Vi,k ≡ (Q i,k ) X i−1,k Q i−1,k , i = 1, . . . , m − 1,
(2) (2) (2) e
Ra,k ≡ Ba,k Q 0,k , Rb,k ≡ Bb,k X m−1,k Q (2)
m−1,k
(4.68)
the system of linear equations (4.67) can be represented in the very simple form
⎛ ⎞
U0,k | ⎛ ⎞
⎜V −I p | 0 ⎟
⎜ 1,k ⎟ ⎜ r (1) ⎟
⎜ .. .. ⎟⎜ k ⎟
⎜ . . | ⎟⎜ ⎟
⎜ ⎟⎜ ⎟
⎜ Vm−1,k −I p | ⎟⎜ ⎟
⎜ ⎟⎜ ⎟
⎜ − − − − − − − − ⎟⎜⎟ ⎜ − ⎟
⎜ ⎟
⎜Z | U1,k −Iq ⎟⎜ ⎟
⎜ 1,k ⎟⎜ ⎟
⎜ .. .. .. ⎟⎜ ⎟
⎜ . | . . ⎟ ⎜ (2) ⎟
⎜ ⎟⎝r ⎠
⎝ Z m−1,k | Um−1,k −Iq ⎠ k
Z 0,k Z m,k | Ra,k Rb,k

T
= η (1) (2)
k , ηk , (4.69)
where
T T
r (1) (1) (1)
k ≡ d 0,k , . . . , d m−1,k , r (2) (2) (2)
k ≡ d 0,k , . . . , d m−1,k ,

η (1)
k ≡ Ba(1) s0,k − β (1) , (Q (1)
1,k ) (u0,k − s 1,k ), . . . ,
T e

T
(Q (1)
m−1,k ) (um−2,k − s m−1,k )
T e
,

η (2)
k ≡ (Q (2)
1,k ) (u0,k − s 1,k ), . . . ,
T e

T
(Q (2)
m−1,k ) (um−2,k − s m−1,k ), g q (s 0,k , um−1,k )
T e e
.
If the boundary conditions in (4.59) are completely separated, i.e.,
g q = g q ( y(b)),
we have Z 0,k = 0 and Ra,k = 0. Therefore, the linear algebraic system (4.69) can
be solved by the following stable block elimination technique, which requires only
minimum storage space. In the first step of this technique the unknown vector r (1) k
is determined from the linear system, which has as system matrix the bidiagonal
(1, 1)-block matrix of (4.69) and as right-hand side the vector η (1) k . Then, the result
(2)
is substituted into (4.69), and the second unknown vector r k is computed from
the lower block system. The corresponding system matrix is the bidiagonal (2, 2)-
block matrix from (4.69). An appropriate implementation of this solution strategy
(using an adequate form of compact storage) requires only mq + n 2 , m > 2, memory
locations. If in addition some steps of the iterative refinement are executed to improve
the accuracy, the memory requirements are increased to m(n 2 + 2n + q) + n 2 − 2q

memory locations. By contrast, the application of the standard Gaussian elimination
would require 3mn 2 memory locations (without iterative refinement) and (4m −
8)n 2 + 2mn memory locations (with iterative refinement). In general, the block
elimination technique requires slightly more CPU time than the standard Gaussian
elimination (depending on p and the orthogonalization technique used). However,
this fact is relatively uninteresting since in a shooting method the linear equation
solver requires only a very small percentage of the total CPU time. What really
matters is the number of IVPs which have to be integrated.
Let us now come back to the BVP (4.59) with partially separated boundary
conditions, and our aim to develop a nonlinear stabilized march method, which is a
real analog of the linear stabilized march method. There are at least two possibilities
to realize the modification of the stabilized march method. The resulting algorithms
differ with respect to the numerical effort and the convergence properties.
To derive the first modification, we start again with the block elimination of the
(1,1)-block system in (4.69), i.e., the first unknown vector r (1) (1) (1)
k = (d 0,k , . . . , d m−1,k )
T
is determined from the linear equations
U0,k d (1) (1) (1)

0,k = Ba s 0,k − β ,
(1) (1) T (1) T e (1) (1)
d i,k = (Q i,k ) (ui−1,k
e
− si,k ) + (Q i,k ) X i−1,k Q i−1,k d i−1,k , (4.70)
i = 1, . . . , m − 1.
This strategy is immediately plausible, if we look at the upper part of the system
e
(4.69) and take notice of the special lower triangular form. Since the matrix X i−1,k
(1) (1)
appears in (4.70) only in combination with the vector Q i−1,k d i−1,k , the corresponding
term can be approximated by a (discretized) directional derivative. To do this, we
use Algorithm 1.1 with the parameters x0 ≡ τi−1 , x1 ≡ τi , j ≡ 1, s ≡ si−1,k , R ≡
(1) (1)
Q i−1,k d i−1,k , i = 1, . . . , m − 1. Obviously, this realization requires the integration
of only 2(m − 1) IVPs. Now, r (1) k is substituted into (4.69) and the second unknown
vector r (2)
k = (d (2)
0,k , . . . , d (2)
m−1,k ) is the solution of the (2,2)-block system in (4.69).
T
Thus, we have to solve the following system of algebraic equations

⎛ ⎞ ⎛ (2) ⎞
U1,k −Iq d 0,k
⎜ .. .. ⎟⎜ . ⎟
⎜ . . ⎟ ⎜ .. ⎟
⎜
⎜ ⎟ ⎜ (2) ⎟ ⎟
⎝ −Iq ⎠⎝d
m−2,k ⎠
Um−1,k
(2) (2) (2) e (2)
Ba,k Q 0,k Bb,k X m−1,k Q m−1,k d (2)
m−1,k
⎛ ⎞ (4.71)
(Q (2)
1,k ) (u0,k
T e
− s1,k ) − (Q (1)
1,k )
T e
X 0,k Q (1) (1)
0,k d 0,k
⎜ .. ⎟
⎜ . ⎟
⎜
= ⎜ (2) ⎟.
(1) (1) (1) ⎟
⎝ (Q m−1,k ) (um−2,k − sm−1,k ) − (Q m−2,k ) X m−2,k Q m−2,k d m−2,k ⎠
T e T e
(2) (1) (1) (2) e
g q (s0,k , uem−1,k ) − Ba,k Q 0,k d 0,k − Bb,k X m−1,k Q (1) (1)
m−1,k d m−1,k
To ensure that this system is well defined, the following quantities are required:
e
ui,k ,
(1) (1)
e
X i,k Q i,k d i,k i = 0, . . . , m − 1,
(2)
Ui,k ≡ X i−1,k
e
Q i−1,k i = 1, . . . , m − 1,
e
X m−1,k Q (2)
m−1,k .
(1) (1)
e
Note that ui,k e
and X i,k Q i,k d i,k , i = 0, . . . , m − 2, have been already computed in
(1)
the determination of r k (see formula (4.70)). To compute the vector uem−1,k and an
e
approximation of X m−1,k Q (1) (1)
m−1,k d m−1,k (on the basis of Algorithm 1.1), two addi-
tional IVPs have to be integrated. Up to this point, we have in total 2m integrations.
(2)
Setting x0 ≡ τi−1 , x1 ≡ τi , j ≡ p, s ≡ si−1,k , R ≡ Q i−1,k , i = 1, . . . , m,
(2)
Algorithm 1.1 can be used again to approximate X i−1,k Q i−1,k , i = 1, . . . , m, with
e
mq integrations. Thus, in the kth step of our first variant of the nonlinear stabilized
march method a total of (q + 2)m IVPs have to be integrated. We call this variant
the Newtonian form of the nonlinear stabilized march method. Compared with the
(nonlinear) multiple shooting method, this new method reduces the number of IVPs,
which have to be integrated, by ( p − 1)m. As a by-product, the dimension of the
corresponding system of linear algebraic equations is reduced from mn to mq.
Once an approximation of d (m) k is computed with the technique described above,
the new iterate s(m)
k+1 is determined with the formula
s(m) (m) (m)

k+1 = s k − Q k d k . (4.72)
As for the conventional Newton method, it is of course also possible to integrate

damping and regularization strategies into the iteration method described above. This
leads to a more or less appropriate globalization of the locally convergent iteration.
Obviously, if the BVP is linear then only one step of the iteration (4.72) is theo-
retically necessary. However, this iteration step does not match with the result of the
linear stabilized march method (see, e.g. ,[63]). The linear method requires only the
integration of (q + 1)m IVPs.
If the Newtonian form of the nonlinear stabilized march method is executed on
a computer, more than one iteration steps are needed to solve a linear BVP with
the required accuracy. These additional iteration steps can be interpreted as a special
form of the iterative refinement, which compensates the unavoidable rounding errors.
The inherent iterative refinement is a significant advantage over the linear stabilized
method.
To get a nonlinear stabilized march method, which solves (theoretically) linear
BVPs in one iteration step and requires the same number of integrations as the linear
stabilizedm march method, we now present a second possibility to modify the linear
shooting technique. This new approach is based on a decoupling of the system (4.69)
by introducing an intermediate step, which transforms the vectors s0,k , . . . , sm−1,k

into vectors ŝ0,k , . . . , ŝm−1,k such that the relations
(1) T
(i) (Q i,k ) (ui−1,k
e
− ŝi,k ) = 0, i = 1, . . . , m − 1,
(4.73)
(ii) Ba(1) ŝ0,k − β (1) = 0
are fulfilled. Then, the regularity of the (1,1)-block matrix in (4.69) implies r (1)
k =0
and the linear algebraic system (4.69) is reduced to the following mq-dimensional
linear system
M̂k(m) z (m)
k = q̂ (m)
k , (4.74)
where M̂k(m) denotes the (2,2)-block matrix of (4.69). Moreover, we have z (m)
k ≡ r (2)
k
and q̂ (m)
k ≡ η (2)
k .
The vectors ŝi,k , i = 0, . . . , m − 1, which satisfy the relations (4.73), can be
determined in a similar manner as described in Sect. 4.3 for the standard form of the
(nonlinear) method of complementary functions. In particular, the general solution
of the first underdetermined system in (4.73) is
(1) T + (1) T e
ŝi,k = {(Q i,k ) } (Q i,k ) ui−1,k
(1) T + (1) T
+ [I − {(Q i,k ) } (Q i,k ) ]ω i,k , ω i,k ∈ Rn . (4.75)
(1)
Since the matrices Q i,k have full rank, the corresponding pseudoinverses can be
determined very easily (see, e.g., [57]):
−1
(1) T + (1) (1) T (1) (1)
{(Q i,k ) } = Q i,k (Q i,k ) Q i,k = Q i,k .
Therefore, it is possible to present formula (4.75) in the form

(1) (1) T e (2) (2) T
ŝi,k = Q i,k (Q i,k ) ui−1,k + Q i,k (Q i,k ) ω i,k , i = 1, . . . , m − 1. (4.76)
The general solution of the second underdetermined system in (4.73) can be written
analogously to (4.33) as follows
ŝ0,k = Q (1) (1) −1 (1)

0,k (U0,k ) β + Q (2) (2) T
0,k (Q 0,k ) ω 0,k . (4.77)
If we set ω i,k ≡ si,k , i = 0, . . . , m − 1, the formulas (4.76) and (4.77) transform the
vectors si,k into vectors ŝi,k , which satisfy the conditions (4.73).
The main advantage of this strategy is that in addition to the reduction of the linear
algebraic systems, the number of IVPs, which have to be integrated, can be reduced.
e e (2)
Namely, in (4.74) the matrices X i,k occur only in the combination X i,k Q i,k . If these
matrix products are approximated by discretized directional derivatives (using the
Algorithm 1.1), the computational effort can be reduced significantly: in each itera-
tion step of the second variant of the nonlinear stabilized march method only m(q +1)
IVPs have to be solved. Note, the multiple shooting method requires the integration
of m(n + 1) IVPs. If the number p of separated boundary conditions is not too small
compared with the dimension q of the nonseparated boundary conditions, the com-
putational effort for the additional algebraic manipulations (Q R factorization, etc.)
is negligible.
Once the solution z (m)
k of the linear algebraic system (4.74) is computed, the new
(m)
iterate sk+1 can be determined by the formula

s(m) (m) (2) (2) (m)
k+1 = ŝ k − diag Q 0,k , . . . , Q m−1,k z k . (4.78)
We call the shooting method, which is defined by the formulas (4.74)–(4.78), the
standard form of the nonlinear stabilized march method.
Example 4.8 Let us consider the following parametrized BVP (see, e.g., [13, 57, 60,
111])
(y3 (x) − y1 (x))y1 (x)

y1 (x) = λ , y2 (x) = −λ (y3 (x) − y1 (x)),
y2 (x)
0.9 − 103 (y3 (x) − y5 (x)) − λ (y3 (x) − y1 (x))y3 (x)
y3 (x) = ,
y4 (x)
y4 (x) = λ (y3 (x) − y1 (x)), y5 (x) = −100 (y5 (x) − y3 (x)), 0 ≤ x ≤ 1,
y1 (0) = y2 (0) = y3 (0) = 1, y4 (0) = −10, y3 (1) = y5 (1).
(4.79)
Here, n = 5, q = 1, p = 4, and λ is a real parameter. Our aim was to determine
the solution of (4.79) for the parameter value λ = 10, 000 by the multiple shooting
method and the standard form of the nonlinear stabilized march method. To find a
starting trajectory we have used the following strategy. In a first step, we have solved
the simplified BVP with λ = 0, using the starting trajectory
y1 (x) = y2 (x) = y3 (x) ≡ 1, y4 (x) ≡ −10, y5 (x) ≡ 1.
Then, the obtained result was taken as starting trajectory for the solution of the
problem with λ = 100. In a second step, we started with the result for λ = 100 to
determine the solution of the BVP for λ = 10, 000. The corresponding IVPs were
solved by the semi-implicit extrapolation method SIMPRS [17].
In the Tables 4.5 and 4.6, we present the results for a grid with 11 and 21 equidis-
tributed shooting points, respectively. Here, we use the following abbreviations: MS
= multiple shooting method, SM = stabilized march method, it = number of itera-
tion steps, NIVP = number of IVP-solver calls, and NFUN = number of calls of the
right-hand side of the ODE (4.79).
4.6 Matlab Programs 155
Table 4.5 Using a grid with 11 equidistant shooting points

Code λ=0 λ = 100 λ = 10, 000
it NIVP NFUN it NIVP NFUN it NIVP NFUN
MS 1 70 3,430 6 370 90,194 4 250 44,736
SM 1 40 1,124 5 120 15,808 2 60 4,320
Table 4.6 Using a grid with 21 equidistant shooting points

Code λ=0 λ = 100 λ = 10, 000
it NIVP NFUN it NIVP NFUN it NIVP NFUN
MS 1 140 6,860 6 740 124,502 4 500 70,110
SM 1 80 2,076 4 200 16,880 2 120 5,332
4.6 M ATLAB P ROGRAMS
In this section, we present a simple implementation of the multiple shooting method

in the Matlab programming language. This implementation can be used to solve
the BVPs given in the exercises as well as own problems of the reader. Here, the
equidistributed shooting points a = τ0 , . . . , τm = b must be given at advance by
entering the number m. The numerical solution is presented at the shooting points in
form of a table. Moreover, all components of the solution vector are plotted, where
the points of the table (red stars) are interpolated by cubic splines. The right-hand
side of the ODE and the boundary conditions must be defined in Algorithms 4.3
and 4.4, respectively. Some standard BVPs are pre-formulated (see Algorithm 4.2).
Moreover, the problems can depend on an external parameter λ ∈ R. The value of λ
must be entered by the user. Setting manually starting values for the vector function
y(x) at the given shooting points requires a good understanding of the mathematical
background. For the preformulated BVPs we have set the starting values as follows
(see Algorithm 4.5):
• BVPs 1 and 3: we use y(τ j ) = 0, j = 0, . . . , m − 1,
• BVP 2: first, we determine the solution for λ = 0 and use this result as the starting
trajectory for λ > 0,
• BVPs 4 and 5 (see also Chap. 5): the solutions of both problems are periodic.
Considering this behavior of the solutions and the given boundary conditions, we
have set y(τ j ) = c ∗ cos(πτ j ) for BVP 4 and y(τ j ) = c ∗ sin(πτ j ) for BVP 5,
j = 0, . . . , m − 1, where c is a free parameter.
Algorithm 4.2 Main program

%
% Multiple Shooting Method
%
clear all , % close all
g l o b a l nr zz la
zz =0;
nn =[2 ,5 ,2 ,2 ,2 ,2];
disp ( ' the f o l l o w i n g p r o b l e m s are p r e p a r e d : ')
disp ( ' 1 - T r o e s c h ')
disp ( ' 2 - Scott / Watts ')
disp ( ' 3 - Bratu ')
disp ( ' 4 - Euler B e r n o u l l i rod ')
disp ( ' 5 - la *( y + y ^2) ' )
disp ( ' 6 - BVP on an i n f i n i t e i n t e r v a l ' )
nrs = input ( ' P l e a s e e n t e r the n u m b e r of the p r o b l e m : ' , 's ' );
nr = s t r 2 n u m ( nrs );
n = nn ( nr );
m = input ( ' P l e a s e e n t e r the n u m b e r of s h o o t i n g p o i n t s m = ' );
las =[];
if nr ~ = 6
las = input ( ' P l e a s e e n t e r the v a l u e of p a r a m e t e r l a m b d a = ' , 's ' );
la = s t r 2 n u m ( las );
end
o p t i o n s = o p t i m s e t ( ' D i s p l a y ' , ' iter ' , ' J a c o b i a n ' , ' on ' , ' T o l F u n ' , ...
1e -8 , ' TolX ' ,1 e -8 , ' M a x I t e r ' ,1000 , ' M a x F u n E v a l s ' ,500* n * m );
[ s0 ,a , b ]= s t a r t v a l u e s ( n , m , nr );
tic , ua = f s o l v e ( @msnl , s0 (:) , options , a , b ,n , m ); toc ,
% solution matrix y(n ,m),
% each column - v e c t o r c o r r e s p o n d s to the s o l u t i o n
% at a s h o o t i n g point
y = r e s h a p e ( ua , n , m );
o p t i o n s = o d e s e t ( ' R e l T o l ' ,1e -8 , ' A b s T o l ' ,1 e -8);
d =( b - a )/( m );
t=a:d:b;
[ ~ , yend ]= ode45 ( @ode ,[ t ( end -1) , t (end) ] , y (: , m ) , o p t i o n s );
y =[ y , yend ( end ,:) '];
z =1;
% Table
disp ( ' ')
disp ( ' Table of s o l u t i o n ' )
ki = fix ( n /3);
es = ' ';
form = ' s p r i n t f ( ' ' %#15 .8e ' ' ';
for ii =0: ki
s =[ ' i x(i) y ( i , ' , i n t 2 s t r (1+ ii *3) , ') ' , es ];
for i =2+ ii *3: min (( ii +1)*3 , n )
s =[ s , ' y ( i , ' , i n t 2 s t r ( i ) , ') ' ];
end
disp ( s )
for i =1: m +1
s =[ ' disp ([ ' , ' s p r i n t f ( ' ' %3 i ' ' ,i ) , ' ' ' ' , ' , form , ' ,t ( i )) ' ];
for j =1+ ii *3: min (( ii +1)*3 , n )
js = i n t 2 s t r ( j );
s =[ s , ' , ' ' ' ' , ' , form , ' ,y ( ' ,js , ' ,i )) ' ];
end
s =[ s , ' ]) ' ];
eval ( s )
end
disp ( ' ')
end
% s o l v i n g IVPs to d e t e r m i n e a p p r o x i m a t e s o l u t i o n s b e t w e e n
% the s h o o t i n g p o i n t s
yk = zeros ( m *20 , n );
d = d /20;
for i =1: m
xx = t ( i ): d : t ( i +1);
[ ~ , yy ]= ode45 ( @ode , xx , y (: , i ) , o p t i o n s );
yk (( i - 1 ) * 2 0 + 1 : i *20 ,:)= yy (1:20 ,:);
end
yk =[ yk ; yy ( end ,:)];
tk = a : d : b ;
% plot of y1 to yn
for k =1: n
f i g u r e ( k ); clf
hold on
% plot of the s h o o t i n g p o i n t s ( red p o i n t s )
plot (t , y ( k ,:) , 'r * ')
% plot of the g r a p h b e t w e e n the s h o o t i n g p o i n t s ( blue curve )
plot ( tk , yk (: , k ) ')
hst = ' and \ l a m b d a = ';
if nr == 6 , hst =[]; end
title ([ ' S o l u t i o n of p r o b l e m ' , nrs , hst , las ])
y l a b e l ([ ' y_ ' , i n t 2 s t r ( k ) , '( x ) ' ])
x l a b e l ( 'x ')
shg
end
disp ( ' End of p r o g r a m ')
Algorithm 4.3 Function defining the ODEs

f u n c t i o n f = ode (x , y )
g l o b a l nr zz la xi
s w i t c h nr
case 1
f =[ y (2); la * s i n h ( la * y ( 1 ) ) ] ;
case 2
y3my1 = y (3) - y (1);
y3my5 = y (3) - y (5);
f =[ la * y (1)* y3my1 / y (2);
- la * y3my1 ;
( .9 -1 e3 * y3my5 - la * y (3)* y3my1 )/ y (4);
la * y3my1 ;
1 e2 * y3my5 ];
case 3
f =[ y (2); - la * exp ( y ( 1 ) ) ] ;
case 4
f =[ y (2); - la * sin ( y ( 1 ) ) ] ;
case 5
f =[ y (2); la *( y (1)+ y ( 1 ) ^ 2 ) ] ;
case 6
f =[ y (2); -2* y (1)* y (2)];
end
zz = zz +1;
Algorithm 4.4 Function defining the boundary conditions

f u n c t i o n g = bc ( ya , yb )
global nr
s w i t c h nr
case 1
g =[ ya (1); yb (1) -1];
case 2
g =[ ya (1) -1; ya (2) -1; ya (3) -1; ya ( 4 ) + 1 0 ; yb (5) - yb (3)];

case 3
g =[ ya (1); yb (1)];
case 4
g =[ ya (2); yb (2)];
case 5
g =[ ya (1); yb (1)];
case 6
g =[ ya (1); yb (1) -1];
end
Algorithm 4.5 Function defining the starting values

f u n c t i o n [ s , a , b ]= s t a r t v a l u e s (n ,m , nr )
g l o b a l la
a =0; b =1;
s w i t c h nr
case 1
s = zeros ( n *m ,1);
case 2
s = [ 1 ; 1 ; 1 ; - 1 0 ; .91 ]* ones (1 , m );
laa = la ; la =0;
o p t i o n s = o p t i m s e t ( ' D i s p l a y ' , ' iter ' , ' J a c o b i a n ' , ' on ' , ...
' T o l F u n ' ,1 e -8);
disp ( ' c o m p u t i n g s t a r t i n g value w i t h l a m b d a =0 ' )
s = f s o l v e ( @msnl , s (:) , options ,0 ,1 , n , m );
disp ( ' ready ')
s = s (:);
la = laa ;
disp ([ ' now l a m b d a = ' , n u m 2 s t r ( la )])
case 3
% for the lower branch , u n c o m m e n t the next row
s = zeros ( n *m ,1);
%
% for the u p p e r b r a n c h w i t h l a m b d a >= 0 .2 and m >= 20
% or w i t h l a m b d a >= 0 .025 and m >= 50 ,
% u n c o m m e n t the next row
%s= zeros (n*m ,1)+6;
case 4
t =0:1/ m :1 -1/ m ;
% for the upper branch , c h o o s e c =2 ,
% for the lower branch , c h o o s e c = -2
c =2;
% for the i - th b r a n c h c h o o s e d = i
d =3;
s = c *[ cos ( d * pi * t ); - d * pi * sin ( d * pi * t )];
case 5
t =0:1/ m :1 -1/ m ;
% for the l o w e r f i r s t b r a n c h ( la < pi ^2) ,
% c = -4; d =1;
% for the u p p e r f i r s t b r a n c h ( la > pi ^2) ,
% c =4; d =1;
% for the other upper branches , c h o o s e c = .5 ,
% for the other lower branches , c h o o s e c = - .5
c = .5 ;
% for the i - th ( i >2) b r a n c h c h o o s e d = i

d =3;
s = c *[ sin ( d * pi * t ); d * pi * cos ( d * pi * t )];
case 6
s = zeros ( n *m ,1);
b = input ( ' P l e a s e e n t e r the right b o u n d a r y b = ' );
end
Algorithm 4.6 Function handling the nonlinear algebraic equations

f u n c t i o n [ Fm , Mm ]= msnl ( s , a ,b ,n , m )
% msvnl - c o m p u t e s the v e c t o r Fm ( s ) and the J a c o b i a n Mm ( s )
% of m u l t i p l e s h o o t i n g
%
% Fm = msvnl (s , ab )
% [ Fm , Mm ] = msvnl (s , ab )
% s - nxm - vector , where Fm ( s ) and Mm have to be c o m p u t e d
% a , b - left and right end of i n t e r v a l [ a , b ] , a > b or b > a
% n - n u m b e r of d i f f e r e n t i a l e q u a t i o n s
% m - n u m b e r of ( a c t u a l ) s h o o t i n g p o i n t s ( t ( m +1)= b )
% Fm - nxm - v e c t o r Fm of the n o n l i n e a r s y s t e m
% Mm - n * m x n * m J a c o b i a n Mm of Fm in as sparse - m a t r i x
%
s = r e s h a p e ( s , n , m );
tol =1 e -8; hd = tol ^ .5 ;
o p t i o n s = o d e s e t ( ' R e l T o l ' , tol , ' A b s T o l ' , tol );
d =( b - a )/ m ; t = a : d : b + d /10;
% m e m o r y is a l l o c a t i n g , Fm is a nxm M a t r i x for m o r e
% structured programming
u = zeros (n , m ); Fm = u ;
% C o m p u t e Fm
for j =1: m
[ ~ , uj ]= ode45 ( @ode ,[ t ( j ) , t ( j +1)] , s (: , j ) , o p t i o n s );
u (: , j )= uj ( end ,:) ';
if j ~ = m , % matching conditions
Fm (: , j +1)= u (: , j ) - s (: , j +1);
else % boundary condition
Fm (: ,1)= bc ( s (: ,1) , u (: , m ));
end
end
% C o m p u t e Mm by d i f f e r e n c e a p p r o x i m a t i o n
if n a r g o u t > 1
% X is a ( n x n ) x ( m +1) ( non - s p a r s e ) c o m p a c t m a t r i x
X = zeros (n ,n , m +1);
for j =1: m
sh = s (: , j );
for i =1: n
h =(1 e -3+ abs ( s ( i )))* hd ;
sh ( i )= sh ( i )+ h ;
[ ~ , uh ]= ode45 ( @ode ,[ t ( j ) , t ( j +1)] , sh , o p t i o n s );
uh = uh ( end ,:) ';
if j ==1 % c p m p u t i n g of Ba
Fh = bc ( sh , u (: , m ));
X (: , i , m +1)=( Fh (:) - Fm (: ,1))/ h ;
end
if j < m % m a t r i x Xj is c o m p u t e d
X (: , i , j )=( uh - s (: , j +1) - Fm (: , j +1))/ h ;
else % Bb * Xm is c o m p u t e d
Fh = bc ( s (: ,1) , uh );
X (: , i , m )=( Fh (:) - Fm (: ,1))/ h ;
end
sh ( i )= s (i , j );
end
end
end
Fm = Fm (:); % F is now a column - v e c t o r
if m >1 % multiple shooting
e = ones ( n * m ,1);
Mm = s p d i a g s ( -e ,0 , n * m , n * m ); % now Mm is a s p a r s e m a t r i x
Mm (1: n ,1: n )= X (: ,: , m +1);
for i =1: m -1
Mm ( i * n +1: n *( i +1) ,( i -1)* n +1: n * i )= X (: ,: , i );
end
Mm (1: n ,( m -1)* n +1: n * m )= X (: ,: , m );
else % simple shooting
Mm = X (: ,: ,1)+ X (: ,: ,2); % o n l y s i n g l e s h o o t i n g ( non - s p a r s e )
end
4.7 Exercises
Exercise 4.1 Given the parametrized BVP
y1 (x) = y2 (x), y2 (x) = λ y1 (x)2 ,

(4.80)
y1 (0) = 4, y1 (1) = 1.
This BVP is a generalization of a frequently used test problem, where λ = 3/2

(see, e.g., [57, 117]). For each λ ∈ [1, 20] there are exactly two isolated solutions.
For different values of λ determine the corresponding two solutions with the simple
shooting method and plot each pair of solutions in a diagram y1 (x) versus x. More-
over, visualize the solution manifold of (4.80) by plotting y2 (0) versus λ, λ ∈ [1, 20].
Exercise 4.2 Given the BVP
y1 (x) = y2 (x), y2 (x) = 50 sin(y1 (x)),

(4.81)
y1 (0) = 0, y1 (1) = 0.
This BVP has exactly five isolated solutions. Try to compute all solutions with the
simple shooting method and the multiple shooting method.
Exercise 4.3 Consider the BVP (4.79). Set λ = 100 and determine the solution
of this BVP with the multiple shooting method. Generate a starting trajectory by
solving (4.79) for the parameter value λ = 0. Since the ODEs are stiff integrate the
associated IVPs with an implicit solver.
4.7 Exercises 161
Exercise 4.4 Given the periodic BVP (see [103])
y (x) = −0.05 y (x) − 0.02 y(x)2 sin(x) + 0.00005 sin(x) cos(x)2

− 0.05 cos(x) − 0.0025 sin(x), (4.82)
y(0) = y(2π), y (0) = y (2π).
It has the exact solution y(x) = 0.05 cos(x). Approximate this solution with the
multiple shooting method.
Exercise 4.5 Consider the BVP

√
y (x) = exp(y(x)), y(0) = 0, y(1) = − ln cos(1/ 2)2 , (4.83)
√
with the exact solution y(x) = − ln cos(x/ 2)2 . Approximate this solution with
a shooting method of your choice.

1
y (x) = cos2 (y (x)), y(0) = 0, y(1) = arctan(1/2) − ln(5/4), (4.84)
2
with the exact solution y(x) = x arctan((1/2)x) − ln(1 + (1/4)x 2 ). Approximate

this solution with the nonlinear method of complementary functions.
y (x) = y(x) y (x),

(4.85)
y (0) = −0.5/ cosh2 (1/4), y(1) = − tanh(1/4).
The exact solution is y(x) = tanh((1 − 2x)/4). Approximate this solution with the
y (x) = a y(x) y (x). y(0) = tanh(a/4), y(1) = − tanh(a/4). (4.86)
The exact solution is y(x) = tanh(a(1 − 2x)/4). Approximate this solution with the
Exercise 4.9 Consider the BVP for a system of two second-order ODEs
y1 (x) = y1 (x) + 3 exp(y2 (x)),

y2 (x) = y1 (x) − exp(y2 (x)) + exp(−x),
(4.87)
y1 (0) = 0, y1 (1) = exp(−2) − exp(−1),
y2 (0) = 0, y2 (1) = −2,
with the exact solution
y1 (x) = exp(−2x) − exp(−x), y2 (x) = −2x.
Transform the ODEs into a system of four first-order equations and solve this problem
with a shooting method of your choice.
12
y (iv) (x) = 6 exp(−4y(x)) − ,
(1 + x)4 (4.88)
y(0) = 0, y(1) = ln(2), y (0) = −1, y (1) = −0.25.
The exact solution is
y(x) = ln(1 + x).
Approximate this solution with the stabilized march method and the multiple shooting
method, and compare the results.
Exercise 4.11 Consider the parametrized BVP
y1 (x) = λ2 y1 (x) + y2 (x) + x + (1 − λ2 ) exp(−x),

y2 (x) = −y1 (x) + exp(y2 (x)) + exp(−λx),
(4.89)
y1 (0) = 2, y1 (1) = exp(−λ) + exp(−1),
y2 (0) = 0, y2 (1) = −1.
y1 (x) = exp(−λx) + exp(−x), y2 (x) = −x.
Use the multiple shooting method to approximate this solution for the parameter
values λ = 10 and λ = 20.
Exercise 4.12 A well-studied nonlinear BVP (see, e.g., [47, 87, 100, 108]) is the
so-called Bratu’s problem, which is given by
y (x) + λ e y(x) = 0,
(4.90)
y(0) = y(1) = 0,
where λ > 0. The analytical solution can be written in the following form
! "
cosh(0.5(x − 0.5)θ)
y(x) = −2 log ,
cosh(0.25 θ)
4.7 Exercises 163
where θ is the solution of the nonlinear algebraic equation

√
f (θ) ≡ θ − 2λ cosh(0.25 θ) = 0.
The Bratu’s problem has none, one or two solutions when λ > λ∗ , λ = λ∗ and
λ < λ∗ , respectively. In the next chapter, it will be shown that the critical value
λ∗ ≈ 3.513830719 is the simple turning point of the curve y (0) versus λ and
satisfies the equation
√
1 = 0.25 2λ∗ sinh(0.25 θ).
For the parameter values λ = 1 and λ = 2 approximate and plot the corresponding
solutions of Bratu’s problem with the simple shooting method.
Exercise 4.13 A modification of the BVP (4.90) is the following problem
y (x) − 2e y(x) = 0,
(4.91)
y(0) = y (π) = 0.
Note, in contrast to the standard Bratu problem, the parameter λ is negative. Deter-
mine a solution of this BVP with the multiple shooting method using y(x) ≡ 0 as
starting trajectory and m = 100.
In [125], an IVP for the ODE used in formula (4.91) is presented. The correspond-
ing initial conditions are y(0) = y (0) = 0. The exact solution of this IVP is
y(x) = −2 ln(cos(x)). (4.92)
This solution solves the BVP (4.91), too. But the transition from the IVP to the BVP
changes the number of solutions. Does the multiple shooting method determine the
solution (4.92), or another solution?
Exercise 4.14 Consider the three-point BVP for the third-order nonlinear ODE
y (x) = 1 + x sinh(y(x)),

(4.93)
y(0) = 0, y(0.25) = 1, y(1) = 0.
The exact solution is unknown. Modify the simple shooting method such that the
resulting algorithm can be used to determine numerically a solution of the BVP
(4.93).
2y (x)2
y (x) = −y(x) + ,
y(x) (4.94)
y(−1) = y(1) = 0.324027137.

1
y(x) = .
ex + e−x
Approximate this solution with a shooting method of your choice.
Exercise 4.16 Consider the two-point BVP for the third-order nonlinear ODE

y (x) = − 1 − y(x)2 ,
(4.95)
y(0) = 0, y (0) = 1, y(π/2) = 1.
The exact solution is y(x) = sin(x). Approximate this solution with the nonlinear
method of complementary functions.
Exercise 4.17 Given the BVP of Bratu-type (see [125])
y (x) − π 2 e y(x) = 0,
(4.96)
y(0) = y(1) = 0.
Note, in contrast to the standard Bratu problem (4.90), the parameter λ is negative.
The exact solutions are
y1,2 (x) = − ln (1 ± cos ((0.5 + x)π)) .
Determine a numerical solution of this BVP with the multiple shooting method using
y(x) ≡ 0 as starting trajectory and m = 100. Which exact solution is approximated
by this numerical solution?
Chapter 5
Numerical Treatment of Parametrized
Two-Point Boundary Value Problems
5.1 Introduction
One of the simplest and oldest examples to demonstrate the phenomenon of bifur-
cation is the so-called Euler–Bernoulli rod. Here, a thin and homogeneous rod of
length l is gripped at either end. The left end is fixed whereas the right end is
allowed to vary along the x-axis when it is subjected to a constant horizontal force P
(see Fig. 5.1).
In the unloaded state, the rod coincides with the section [0, l] of the x-axis. If a
compressive force P is applied to the right end of the rod, two different states are
possible. In the case that the force is very small, the rod is only compressed. However,
our experience shows that for larger forces there must be transversal deflections, too.
For the mathematical modelling, let us assume that all deformations of the rod are
small and the buckling occurs in the x, y-plane. We are now studying the equilibrium
of forces on a piece of the rod (including the left end), see Fig. 5.2.
Let X be the original x-coordinate of a point on the rod. After the buckling this
point has moved to the point (X +u, v). Let ϕ denote the angle between the tangent to
the deflected rod and the x-axis. Moreover, s is the arc length measured from the left
endpoint. An important assumption for our model is that the rod cannot extend. The
Euler–Bernoulli law says that the change in the curvature (dϕ/ds) is proportional to
the moment of force M, i.e., we have the constitutive relations
dϕ
s = X and M = −E I , (5.1)
ds
where the constants E and I denote the elastic modulus and the second moment of
area of the rod’s cross-section (see, e.g., [15]), respectively. In addition, we have the
geometric relation
dv
= sin(ϕ), (5.2)
ds

DOI 10.1007/978-81-322-2812-7_5
166 5 Numerical Treatment of Parametrized Two-Point …
Fig. 5.1 Euler–Bernoulli

rod
Fig. 5.2 Equilibrium of

forces of the rod
and the equilibrium relation

M = P v. (5.3)
These four equations can be reduced to a pair of nonlinear first-order ODEs as

follows. From (5.1), we obtain
M dϕ
=− .
EI ds
Using (5.3), it follows
P dϕ
v=− .
EI ds
Defining the dimensionless parameter
P
λ≡
EI
and adding (5.2), we obtain the two equations
ϕ (s) = −λ v(s), v (s) = sin(ϕ(s)). (5.4)
The associated boundary conditions are
v(0) = v(l) = 0. (5.5)
If λ = 0, i.e., there is no force acting, the first ODE is reduced to ϕ (s) = 0. From
the second ODE, we obtain v (s) = cos(ϕ(s)) ϕ (s). Substituting ϕ (s) = 0 into
this equation yields v (s) = 0. If we also take into account the boundary conditions
(5.5), we get v(s) ≡ 0, which means that no deflection occurs.
The two first-order ODEs (5.4) and the boundary conditions (5.5) can be combined
to the following (parametrized) BVP for a single second-order ODE
ϕ (s) + λ sin(ϕ(s)) = 0, 0 < s < l,

(5.6)
ϕ (0) = ϕ (l) = 0.
Since we have assumed that v and ϕ are small, it seems reasonable to linearize the
ODE. Therefore, we replace sin(ϕ) by ϕ and obtain the following linearized BVP
ϕ (s) + λ ϕ(s) = 0, 0 < s < l,

(5.7)
ϕ (0) = ϕ (l) = 0.
Obviously, ϕ(s) ≡ 0 is a solution of this BVP for all values of λ. But we are interested
in those values of λ for which nontrivial solutions of (5.7) exist. In order to pursue
this goal, we proceed as follows. To determine the general solution of the ODE, we
use the ansatz ϕ(s) = eks and obtain the characteristic
√ equation ρ(k) ≡ k 2 + λ = 0.
The corresponding solutions are k1,2 = ±i λ. Thus, the general solution of the
ODE is √ √
ϕ(s) = ĉ1 ei λs + ĉ2 e−i λs .
Using Euler’s formula, we can write

√ √
ϕ(s) = c1 cos( λs) + c2 sin( λs),
where c1 and c2 are real constants. It follows

√ √ √ √
ϕ (s) = −c1 λ sin( λs) + c2 λ cos( λs).
Considering the boundary conditions, we get

√ .
ϕ (0) = c2 λ = 0, i.e., c2 = 0,
and √ √ .
ϕ (l) = −c1 λ sin( λl) = 0.
√ √
If sin( λl) = 0, i.e., λ l = n π, the above equation is satisfied for all c1 . In analogy
to the algebraic eigenvalue problems, we call the numbers
n2 π2
λn = , n = 1, 2, . . . (5.8)
l2
the eigenvalues of the linearized BVP. The corresponding eigensolutions are
nπ
ϕ(n) (s) = c cos( λn s) = c cos s , n = 1, 2, . . . . (5.9)
l
Fig. 5.3 The solution manifold of the linearized BVP (5.7) with l = 1
Fig. 5.4 The solution manifold of the nonlinear BVP (5.6); BPi —primary simple bifurcation
points, Γi± —solution branches (see the next section)
Now, we want to display graphically ϕ in dependence of λ. Since ϕ is a function

of s, in Fig. 5.3, we use as ordinate the maximum deflection ϕmax . As can be seen,
there are no deflections before the first critical value λ1 = (π/l)2 has been reached.
For λ = λ1 , the rod is buckling (the state is c cos ((π/l)s)), but the magnitude
of the deflection is indefinite (since the constant c is arbitrary). If the parameter
λ exceeds the critical value λ1 , the rod adopts the undeformed state until the next
critical value λ2 = (4π 2 )/l 2 is reached, etc. This shows that in the linearized problem
(5.7) the rod is only buckling for the critical values (5.8), i.e., for the loads Pn ≡
(n 2 π 2 E I )/l 2 . This result does not correspond to the reality. We can state that only the
nonlinear model (5.6) is appropriate to describe the actual buckling behavior of the
rod. In Figs. 5.4 and 5.5, the solution manifold (ϕ(0) versus λ) and the corresponding
Fig. 5.5 Deformed states of the rod on the branches Γi+ , i = 1, 2, 3
deformed states of the rod, respectively, are shown when the nonlinear BVP is studied
with the numerical techniques, which we will present in the next sections.
5.2 Two-Point BVPs and Operator Equations
Let us consider nonlinear parametrized BVPs of the following form
y (x) = f (x, y; λ), a < x < b,

(5.10)
Ba y(a) + Bb y(b) = 0,
where
y ∈ C1 ([a, b], Rn ), λ ∈ R, f : D f → Rn , D f ⊂ [a, b] × Rn × R,

Ba , Bb ∈ Rn×n , rank[Ba |Bb ] = n.
In this notation, the BVP (5.6) reads
y1 (x) = y2 (x),

y2 (x) = −λ sin(y1 (s)),
and
01 y1 (0) 00 y1 (l) 0
+ = .
00 y2 (0) 01 y2 (l) 0

Ba Bb
Now, we represent the BVP (5.10) as an operator equation
T (y, λ) = 0, (5.11)
where
T : Z ≡ Y × R → W, Y, W Banach spaces,
by defining
T (y, λ) ≡ y − f (·, y; λ),

Y = BC1 ([a, b], Rn ) ≡ y ∈ C1 ([a, b], Rn ) : Ba y(a) + Bb y(b) = 0 ,
y : BC1 = y : C1 ≡ sup {| y(x)| + | y (x)|},
x∈[a,b]
W ≡ C([a, b], R ), w : C ≡ sup {|w(x)|},

n
x∈[a,b]
|x| ≡ (x x)T 1/2

for x ∈ R .
n
To simplify the representation, we often use the abbreviation z ≡ (y, λ) ∈ Z and

write (5.11) in the form
T (z) = 0. (5.12)
For the following considerations, some definitions are required.
Definition 5.1 An element z 0 ∈ Z is called a solution of problem (5.12) if and only

if T (z 0 ) = 0.
Definition 5.2 The operator T is Fréchet differentiable at z 0 ∈ Z if and only if there

exists a linear map L(z 0 ) : Z → W such that
1
lim T (z 0 + h) − T (z 0 ) − L(z 0 )h = 0, h ∈ Z .
h→0 h
If it exists, this L(z 0 ) is called the Fréchet derivative of T at z 0 .
Definition 5.3 The operator T ≡ T (y, λ) is partially Fréchet differentiable w.r.t. y

at z 0 = (y0 , λ0 ) if and only if there exists a linear map L(y0 ) : Y → W such that
1
lim T (y0 + h, λ0 ) − T (y0 , λ0 ) − L(y0 )h = 0, h ∈ Y.
h→0 h
If it exists, this L(y0 ) is called the partial Fréchet derivative of T w.r.t. y at z 0 and
is denoted by Ty0 .
Definition 5.4 An element z 0 ∈ Z is an isolated solution of the operator equation

(5.12) if and only if z 0 is a solution of (5.12) and the Fréchet derivative T (z 0 ) is
bijective.
5.2 Two-Point BVPs and Operator Equations 171
Definition 5.5 A linear map L ∈ L (Z , W ) is called a Fredholm operator if and

only if the following three conditions are satisfied:
• R(L) is closed,
• dim N (L) < ∞, and
• codim R(L) < ∞.
Here, the null space of L is denoted as N (L) and the range as R(L).
The number
ind(L) ≡ dim N (L) − codimR(L)
is called the index of the Fredholm operator.
For the BVP (5.10), the following can be shown.

Theorem 5.6 Let z 0 ≡ (y0 , λ0 ) ∈ Z be a solution of the operator equation (5.11).
Then Ty0 ≡ Ty (z 0 ) is a Fredholm operator with index zero.
Remark 5.7 To simplify the representation, we have restricted ourselves to BVPs of

the form (5.10). For more general parametrized nonlinear BVPs,
y (x) = f (x, y; λ), a < x < b,

(5.13)
r( y(a), y(b); λ) = 0,
the operator T , which is used in the operator equation (5.11), has to be defined as

y − f (·, y; λ)
T (y, λ) ≡ , (5.14)
r( y(a), y(b); λ)
where the Banach spaces Y, W must be adapted accordingly. If it can be shown that
Ty0 is a Fredholm operator with index zero, the theoretical and numerical techniques
presented in the next sections are still valid for the BVP (5.13).
Let E(T ) be the manifold of all solutions z ≡ (y, λ) of the nonlinear operator
equation (5.11), which we call the solution field of the operator T . The solution field
can be represented graphically in form of a bifurcation diagram where the values of a
linear functional l ∗ at the solutions y are plotted versus λ. In Fig. 5.6, an example of a
bifurcation diagram is given. In this diagram, the solutions are arranged in curves Γi ,
which we call the solution curves. These curves may intersect at so-called bifurcation
points (BP). The bifurcation points are singular solutions of the operator equation
as we will see later. There is another type of singular solutions, namely points on a
solution curve where this curve turns back. These points are called limit points or
turning points (TP).
Fig. 5.6 Example of a

bifurcation diagram
Our bifurcation diagram contains qualitatively different types of singular points.

If Γ0 is a known (primary) solution curve, the point λ A is called primary simple
bifurcation point (pBP) since only two curves (the primary curve Γ0 and the secondary
curve Γ1 ) are intersecting in λ A . In general, it can be assumed that Γ0 = {(y, λ) =
(0, λ) for all λ}. The point λ B is a secondary simple bifurcation point (sBP) since only
two curves (the secondary curves Γ1 and Γ2 ) are intersecting in λ B . Consequently,
λ D is a secondary multiple bifurcation point (mBP) since more than two curves (the
secondary curves Γ2 , Γ3 and Γ4 ) are intersecting in λ D . Finally, λC is called simple
turning point.
The great variety in the quality of the possible singular points must be taken into
account, if a numerical treatment of the nonlinear parametrized operator equation
(5.11) is realized.
5.3 Analytical and Numerical Treatment of Limit Points
5.3.1 Simple Solution Curves
In this section, the study of E(T ) is restricted to regions, where no branching of

solution curves occur. For this, we need the statement of the following theorem.
Theorem 5.8 (Implicit Function Theorem)
Suppose that X, Y, Z are Banach spaces, U ⊂ X , V ⊂ Y are open sets, T ∈
C p (U × V, Z ), p ≥ 1, z 0 ≡ (x0 , y0 ) ∈ U × V , T (z 0 ) = 0 and Tx (z 0 ) is a linear
homeomorphism from X onto Z . Then, there exist a neighborhood U1 × V1 ∈ U × V
of z 0 and a function g ∈ C p (V1 , X ), g(y0 ) = x0 such that T (x, y) = 0 for (x, y) ∈
U1 × V1 if and only if x = g(y).
Let z 0 ≡ (x0 , λ0 ) ∈ Z be a solution of (5.11). Then, y can be parametrized with
respect to λ in a neighborhood of z 0 , if Ty0 is a linear homeomorphism.
5.3 Analytical and Numerical Treatment of Limit Points 173
Definition 5.9 A solution z 0 ∈ Z of (5.11) is called simple, if the Fréchet derivative
T (z 0 ) = (Ty0 , Tλ0 ) ∈ L (Z , W )
satisfies
dim N (T (z 0 )) = 1, R(T (z 0 )) = W. (5.15)
To find the simple solutions of the operator equation (5.11), let us consider the
equation
w
T (z 0 )h ≡ (Ty , Tλ )
0 0
= 0, i.e., Ty0 w + μTλ0 = 0.
μ
Obviously, this equation is homogeneous and we have to distinguish between two

cases:
w1 w2
case 1: h 1 = , case 2: h 2 = .
1 0
• Case 1: To determine w1 uniquely, it must be assumed that N (Ty0 ) = {0}. Thus,

w1 is defined by Ty0 w1 = −Tλ0 and the corresponding solution z 0 is called isolated.
• Case 2: To determine w2 uniquely, it must be assumed that N (Ty0 ) = span{ϕ},
0 = ϕ ∈ Y , and Tλ0 ∈
/ R(Ty0 ). The corresponding solution z 0 is called limit point.
In case of isolated solutions, Theorem 5.8 guarantees the existence of a unique
solution curve {y(λ), λ} for |λ − λ0 | ≤ ε.
We now want to derive a local statement for limit points. Let us start with the
assumption (see case 2) that
N (Ty0 ) = span{ϕ0 }, ϕ0 = 1. (5.16)
From Theorem 5.6 we know, that Ty0 is a Fredholm operator with index zero. There-
fore, the dual operator Ty0∗ : W ∗ → Y ∗ also has a one-dimensional null space,
i.e.,
N (Ty0∗ ) = span{ψ0∗ }, ψ0∗ ∈ W ∗ , ψ0∗ = 1, (5.17)
where Y ∗ and W ∗ are the dual spaces of Y and W , respectively. For any Banach space
X the dual Banach space X ∗ is the set of bounded linear functionals on X equipped
with an appropriate norm.
The theorem on biorthogonal bases (see, e.g., [134]) implies the existence of
elements ϕ∗0 ∈ Y ∗ and ψ0 ∈ W such that
ϕ∗0 ϕ0 = 1 and ψ0∗ ψ0 = 1. (5.18)

Now we can define the following decompositions of the Banach spaces Y and W :
Y = span{ϕ0 } ⊕ Y1 , W = span{ψ0 } ⊕ W1 , (5.19)
where
Y1 ≡ {y ∈ Y : ϕ∗0 y = 0}, W1 ≡ {w ∈ W : ψ0∗ w = 0}. (5.20)
Finally, we have the Fredholm alternative (see, e.g., [43])
W1 = R(Ty0 ). (5.21)
A well-known analytical technique to study the operator equation (5.11) is the

Liapunov–Schmidt reduction. Let Q 2 denote the projection of W onto W1 = R(Ty0 ),
i.e.,
Q 2 w = w − (ψ0∗ w)ψ0 , w ∈ W. (5.22)
Obviously, for w ∈ W it holds
w = 0 if and only if Q 2 w = 0 and (I − Q 2 )w = 0.
Thus, the operator equation (5.11) can be expanded to an equivalent pair of equations
Q 2 T (z) = 0, z ≡ (y, λ) ∈ Z ,
(5.23)
Q 1 T (z) ≡ (I − Q 2 ) T (z) = 0.
We still need a second pair of projections, namely
P1 : Y → N (Ty0 ) and P2 = I − P1 .
Let z 0 = (y0 , λ0 ) be a fixed solution of (5.11). Because of the decomposition (5.19),

we may split any y ∈ Y in the form
y = y0 + εϕ0 + w, ε ∈ R, P2 w = w. (5.24)
The condition P2 w = w says that w ∈ Y1 . In the same way, we split λ ∈ R as
λ = λ0 + ξ, ξ ∈ R. (5.25)
Substituting (5.24) and (5.25) into the first equation of (5.23) yields
G(ξ, ε, w) ≡ Q 2 T (y0 + εϕ0 + P2 w, λ0 + ξ) = 0.
It holds G(0, 0, 0) = 0 and G w (ξ, ε, w) = Q 2 Ty (y0 + εϕ0 + P2 w, λ0 + ξ)P2 . Let

us set G 0w ≡ G w (0, 0, 0) = Q 2 Ty0 P2 .
The operator G 0w is an injective map of Y1 onto R(Ty0 ). This can be shown as

follows: Given the equation
G 0w v = 0, i.e., Q 2 Ty0 P2 v = 0.
Since Q 2 is a projection onto the range of Ty0 , we get Ty0 P2 v = 0. From P2 v ∈

/ N (Ty0 )
we deduce v = 0.
To show that the operator G 0w is also a surjective map, we need the following
lemma:
Lemma 5.10 If X and Y are Banach spaces and L ∈ L (X, Y ) is bijective, then
the inverse operator L −1 is continuous.
Proof This statement is a direct conclusion from the Open Mapping Theorem
(see, e.g., [132]).
For any h ∈ R(Ty0 ), the equation Ty0 y = h is solvable. Obviously, the corresponding
solution y also satisfies the equation Q 2 Ty0 P2 y = G 0w y = h.
Thus, G 0w is bijective. With Lemma 5.10, we can conclude that G 0w is a home-
omorphism of Y1 onto R(Ty0 ). This result and the Implicit Function Theorem 5.8
imply the claim of the next theorem.
Theorem 5.11 Assume z 0 is a limit point of the operator T . Then, there exist two
positive constants ξ0 , ε0 , and a C p -map w : [−ξ0 , ξ0 ] × [−ε0 , ε0 ] → Y1 such that
G(ξ, ε, w(ξ, ε)) = 0, w(0, 0) = 0.
Now, we use the information just received in the second equation of (5.23)
Q 1 T (y0 + εϕ0 + w(ξ, ε), λ0 + ξ) = 0.
This equation is transformed into a scalar equation by applying the functional ψ0∗ on
both sides. We obtain
0 = ψ0∗ Q 1 T (y0 + εϕ0 + w(ξ, ε), λ0 + ξ)

= ψ0∗ (I − Q 2 )T (y0 + εϕ0 + w(ξ, ε), λ0 + ξ)
Using the definition (5.22) of the projection Q 2 , it follows:

0 = ψ0∗ T (·) − T (·) + (ψ0∗ T (·)) ψ0 .
Thus, the solution of the operator equation (5.11) in the neighborhood of z 0 is reduced
to the solution of a finite-dimensional algebraic equation in the neighborhood of the
origin (0, 0)
F(ξ, ε) ≡ ψ0∗ T (y0 + εϕ0 + w(ξ, ε), λ0 + ξ) = 0. (5.26)
This important equation is called the limit point equation. The next theorem shows
that ξ can be parametrized with respect to ε.
Theorem 5.12 Let z 0 ∈ Z be a limit point of the operator T . Then, there exist a
constant ε0 > 0 and a C p -map ξ : [−ε0 , ε0 ] → R such that
F(ξ(ε), ε) = 0, |ε| ≤ ε0 , ξ(0) = 0.
Proof It is F(0, 0) = 0 and
∂F
(ξ, ε) = ψ0∗ Ty (y0 + εϕ0 + w(ξ, ε), λ0 + ξ)[ϕ0 + wε (ξ, ε)].
∂ε
Thus
∂F
(0, 0) = ψ0∗ Ty0 [ϕ0 + wε (0, 0)] = ψ0∗ Ty0 wε (0, 0) = 0,
∂ε
since Ty0 wε (0, 0) ∈ R(Ty0 ) = {x : ψ0∗ x = 0}.

For a limit point the condition Tλ0 ∈
/ R(Ty0 ) is satisfied. Let us write this condition
in the form
c ≡ ψ0∗ Tλ0 = 0, (5.27)
where c is the characteristic coefficient.

Now, we compute
∂F
(ξ, ε) = ψ0∗ Tλ (y0 + εϕ0 + w(ξ, ε), λ0 + ξ)
∂ξ
+ ψ0∗ Ty (y0 + εϕ0 + w(ξ, ε), λ0 + ξ)wξ (ξ, ε).
The second term on the right-hand side vanishes, since Ty0 (·)wξ (·) ∈ R(Ty0 ). Because
of the condition (5.27), we obtain
∂F
(0, 0) = 0.
∂ξ
Using the Implicit Function Theorem 5.8, one obtains the result stated in the
theorem.
Corollary 5.13 Let z 0 ≡ (y0 , λ0 ) be a limit point of the operator T . Then, in the
neighborhood of z 0 ≡ (y0 , λ0 ) there exist a curve {y(ε), λ(ε)}, |ε| ≤ ε0 , of solutions
of the equation (5.11). Moreover, λ : [−ε0 , ε0 ] → R and y : [−ε0 , ε0 ] → Y are
C p -functions, which can be represented in the form
λ(ε) = λ0 + ξ(ε),
(5.28)
y(ε) = y0 + εϕ0 + w(ξ(ε), ε), |ε| ≤ ε0 .
Fig. 5.7 The difference between solution curve (black) and branch (gray)
Definition 5.14
• A continuously differentiable curve of solutions z(ε) ≡ {y(ε), λ(ε)} ∈ Z , ε ∈
[ε0 , ε1 ], of the equation (5.11) is called simple, if it exclusively consists of simple
solutions (isolated solutions, limit points) and ż(ε) = 0 for all ε ∈ [ε0 , ε1 ]. Here,
the dot indicates the derivative w.r.t. ε.
• A solution branch is a continuously differentiable solution curve {y(λ), λ} ∈ Z ,
λ ∈ [λ0 , λ1 ], which consists exclusively of isolated solutions.
In Fig. 5.7 it is shown that the solution curve is split into two solution branches
Γ1 and Γ2 if the limit point is extracted from the curve.
We now want to make a classification of the limit points. Let us start with the
limit point equation
F(ξ(ε), ε) = ψ0∗ T (y0 + εϕ0 + w(ξ(ε), ε), λ0 + ξ(ε)).
Differentiating this equation gives
dF
(ξ(ε), ε) = ψ0∗ Ty (y0 + εϕ0 + w(ξ(ε), ε), λ0 + ξ(ε)) ẏ(ε)
dε
+ ψ0∗ Tλ (y0 + εϕ0 + w(ξ(ε), ε), λ0 + ξ(ε)) λ̇(ε).
dF
In the proof of Theorem 5.12, we have shown that (0, 0) = 0. Thus, we obtain
dε
dF
0= (0, 0) = ψ0∗ Ty0 ẏ(0) + ψ0∗ Tλ0 λ̇(0).
dε

=0 c=0
It follows
λ̇(0) = 0. (5.29)
This gives rise to the following definition.

Definition 5.15 A limit point z 0 ∈ Z is called simple turning point if and only if
λ̈(0) = 0. (5.30)
Otherwise, z 0 is called multiple turning point. In particular, z 0 is a double turning

point if and only if ...
λ̈(0) = 0, λ (0) = 0. (5.31)

Now, we want to find easy-to-verify criteria for the multiplicity of limit points.
The existence of the parametrization (5.28) we have shown. Therefore, we can insert
the parametrization into (5.11) and obtain
T (y(ε), λ(ε)) = 0, |ε| ≤ ε0 .
Differentiating this equation w.r.t. ε gives
Ty (y(ε), λ(ε)) ẏ(ε) + Tλ (y(ε), λ(ε)) λ̇(ε) = 0. (5.32)
Setting ε = 0 yields
Ty0 ẏ(0) + Tλ0 λ̇(0) = 0.
Taking into account that the relation (5.29) is satisfied for a limit point, we get
Ty0 ẏ(0) = 0.
Thus, ẏ(0) ∈ N (Ty0 ), i.e. ẏ(0) = c ϕ0 .

Differentiating the identity (5.32) again with respect to ε, we obtain
Ty (y(ε), λ(ε)) ÿ(ε) + Tyy (y(ε), λ(ε)) ẏ(ε)2 + 2Tyλ (y(ε), λ(ε)) ẏ(ε)λ̇(ε)
+ Tλλ (y(ε), λ(ε)) λ̇(ε)2 + Tλ (y(ε), λ(ε)) λ̈(ε) = 0.
Setting ε = 0 yields
Ty0 ÿ(0) + Tyy

0
ẏ(0)2 + 2Tyλ
0
ẏ(0) λ̇(0) +Tλλ
0
λ̇(0)2 +Tλ0 λ̈(0) = 0.

=0 =0
It follows 0
Ty0 ÿ(0) = − Tyy ẏ(0)2 + Tλ0 λ̈(0) .
This equation is solvable if the right-hand side is in the range of Ty0 , i.e.,
ψ0∗ Tyy
0
ẏ(0)2 + λ̈(0)ψ0∗ Tλ0 = 0.
Thus,
ψ0∗ Tyy
0
ẏ(0)2
λ̈(0) = − ,
c
where the characteristic coefficient c is defined in (5.27). Since ẏ(0) ∈ N (Ty0 ) we

can write
ψ0∗ Tyy
0
ϕ20
λ̈(0) = −α2 , α ∈ R, α = 0.
c
Corollary 5.16 The limit point z 0 ∈ Z is a simple turning point if and only if the
second bifurcation coefficient a2 satisfies
a2 ≡ ψ0∗ Tyy
0
ϕ20 = 0. (5.33)
A similar calculation yields the following result.

Corollary 5.17 The limit point z 0 ∈ Z is a double turning point if and only if
0
a2 = 0 and a3 ≡ ψ0∗ 6Tyy ϕ0 w0 + Tyyy
0
ϕ30 = 0, (5.34)
1 0 2
where w0 ∈ Y1 is the unique solution of the equation Ty0 w0 = − Tyy ϕ0 .
2
5.3.2 Extension Techniques for Simple Turning Points
Let z 0 ≡ (y0 , λ0 ) ∈ Z be a simple turning point. A direct determination of y0 from

the original problem (5.11) (for the fixed value λ = λ0 ) by numerical standard
techniques is not possible since in that case y0 is not an isolated solution. One
possibility to overcome this difficulty is to embed (5.11) into an enlarged problem.
The new problem is constructed such that some components of an isolated solution
of the enlarged problem coincide with the components of z 0 . We call this solution
strategy extension technique.
A first extension technique is based on the enlarged operator
⎧
⎪
⎪ Z̃ ≡ Y × R × Y → W̃ ≡ W ⎛ × W × R⎞
⎨
T (y, λ)
T̃ : ⎝Ty (y, λ) ϕ⎠ , (5.35)
⎪
⎪ z̃ ≡ (z, ϕ) = (y, λ, ϕ) →
⎩
ϕ∗0 ϕ − 1
where ϕ∗0 ∈ Y ∗ such that ϕ∗0 ϕ0 = 1. Obviously, z̃ 0 = (z 0 , ϕ0 ) ∈ Z̃ satisfies the

equation
T̃ (z̃) = 0. (5.36)
The question we have to answer now is: do the simple turning points z 0 of problem
(5.11) actually correspond to isolated solutions z̃ 0 of (5.36)?
Let us study the null space of T̃ (z̃ 0 ). It holds
⎛ ⎞
Ty0 Tλ0 0
T̃ (z̃ 0 ) = ⎝Tyy
0
ϕ0 Tyλ
0
ϕ0 Ty0 ⎠ . (5.37)
0 0 ϕ∗0
If (u, μ, w) ∈ N (T̃ (z̃ 0 )), with u, w ∈ Y and μ ∈ R, we have

⎛ ⎞ ⎛ ⎞
u Ty0 u + μTλ0
T̃ (z̃ 0 ) ⎝ μ ⎠ = ⎝Tyy
0
ϕ0 u + μTyλ 0
ϕ0 + Ty0 w⎠ = 0. (5.38)
w ∗
ϕ0 w
The first equation reads

Ty0 u = −μTλ0 .
The condition Tλ0 ∈/ R(Ty0 ) implies μ = 0. Therefore, u = αϕ0 , α ∈ R. Substituting

this representation into the second and third equation of (5.38) gives
Ty0 w = −αTyy
0 2
ϕ0 , ϕ∗0 w = 0. (5.39)
As defined in (5.19), let Y1 be the complement of N (Ty0 ). The condition ϕ∗0 w = 0

indicates that w ∈ Y1 . Problem (5.39) is solvable if
0 = α ψ0∗ Tyy
0 2
ϕ0 = α a 2 . (5.40)
Now, two cases must be distinguished:

• z 0 ≡ (y0 , λ0 ) is a simple turning point. Then, equation (5.40) implies α = 0 since
a2 = 0. Substituting this result into (5.39) we have w ∈ Y1 ∩ N (Ty0 ), i.e., w = 0.
Furthermore u = α ϕ0 = 0. Thus,
N (T̃ (z̃ 0 )) = {0}.
• z 0 ≡ (y0 , λ0 ) is a multiple turning point. In that case, we have a2 = 0 and

N (T̃ (z̃ 0 )) = {α Φ0 : α ∈ R}, (5.41)
where Φ0 ≡ (ϕ0 , 0, w0 ) ∈ Y × R × Y and w0 is the unique solution of the linear

problem
Ty0 w0 = −Tyy
0 2
ϕ0 , w0 ∈ Y1 .
We are now able to formulate the following theorem.
Theorem 5.18 Let z 0 ≡ (y0 , λ0 ) be a simple turning point of the operator T . Then,
T̃ (z̃ 0 ) is a linear homeomorphism from Z̃ onto W̃ , i.e.,
N (T̃ (z̃ 0 )) = {0}, R(T̃ (z̃ 0 )) = W̃ .
Let z 0 ≡ (y0 , λ0 ) be a multiple turning point of the operator T . Then, it holds
dim N (T̃ (z̃ 0 )) = 1, codimR(T̃ (z̃ 0 )) = 1.
The null space is given in (5.41), whereas the range satisfies
R(T̃ (z̃ 0 )) = { f ≡ (v, x, σ) ∈ W × W × R : Ψ0∗ f = 0}, (5.42)
where
Ψ0∗ ≡ (ξ0∗ , ψ0∗ , 0) ∈ W ∗ × W ∗ × R,
and ξ0∗ is defined by the equations
ξ0∗ Tλ0 = −ψ0∗ Tyλ

0
ϕ0 ,
(5.43)
ξ0∗ Ty0 = −ψ0∗ Tyy
0
ϕ0 .
Proof The first part of the proof can be seen above. The rest is shown in the mono-
graph [123].
If z 0 denotes again a simple turning point, then the reversal of the claim of
Theorem 5.18 is true.
Theorem 5.19 Let z̃ 0 ≡ (y0 , λ0 , ϕ0 ) be an isolated solution of (5.36), i.e., T̃ (z̃ 0 )

is a linear homeomorphism from Z̃ onto W̃ . Then, z 0 ≡ (y0 , λ0 ) is a simple turning
point.
Proof
1. Suppose ϕ1 ∈ Y satisfies Ty0 ϕ1 = 0, ϕ∗0 ϕ1 = 0. It can be easily seen that
(0, 0, ϕ1 ) ∈ N (T̃ (z̃ 0 )). Since z̃ 0 is an isolated solution, we must have ϕ1 = 0, i.e.,
the null space of Ty0 is one-dimensional with N (Ty0 ) = span{ϕ0 } and Ty0 ϕ0 = 0,
ϕ∗0 ϕ0 = 1.
2. Suppose a2 = 0 (z 0 is a multiple turning point). Then, Φ0 ≡ (ϕ0 , 0, w0 ) ∈

N (T̃ (z̃ 0 )), with ϕ0 = 0 and w0 = 0. However, since N (T̃ (z̃ 0 )) = {0}, we must
have a2 = 0.
3. Let us assume that c ≡ ψ0∗ Tλ0 = 0. Then, there must exist a ϕ1 ∈ Y , which satisfies
Ty0 ϕ1 = −Tλ0 . With a suitably chosen α ∈ R and w ∈ Y1 ≡ {v ∈ Y : ϕ∗0 v = 0},
it holds (αϕ0 + ϕ1 , 1, w) ∈ N (T̃ (z̃ 0 )). This is a contradiction to the assumption
N (T̃ (z̃ 0 )) = {0}.
Theorems 5.18 and 5.19 say that there is a one-to-one relationship between the
simple turning points of (5.11) and the isolated solutions of the enlarged problem
(5.36).
Another extension technique is based on the following enlarged operator
⎧
⎪
⎪ Ẑ ≡ Y × R × W ∗ → Ŵ ≡ W ∗
⎛× Y × R ⎞
⎨
T (y, λ)
T̂ : . (5.44)
⎪
⎪ ẑ ≡ (z, ψ ∗ ) = (y, λ, ψ ∗ ) → ⎝ Ty∗ (y, λ) ψ ∗ ⎠
⎩ ∗
ψ Tλ (y, λ) − 1
Obviously, ẑ 0 ≡ (z 0 , ψ0∗ ) ∈ Ẑ satisfies the equation
T̂ (ẑ) = 0. (5.45)
The next theorem shows that the operator T̂ has the same favorable properties as
the operator T̃ .
Theorem 5.20 Let z 0 ≡ (y0 , λ0 ) be a simple turning point of the operator T . Then,
T̂ (ẑ 0 ) is a linear homeomorphism from Ẑ onto Ŵ , i.e.,
N (T̂ (ẑ 0 )) = {0}, R(T̂ (ẑ 0 )) = Ŵ .
We now want to come back to the BVP (5.10) and apply the corresponding
extension techniques. For this BVP the enlarged operator equation (5.36) is
y (x) = f (x, y(x); λ), Ba y(a) + Bb y(b) = 0,

ϕ (x) = f y (x, y(x); λ)ϕ(x), Ba ϕ(a) + Bb ϕ(b) = 0, (5.46)
0= ϕ∗0 ϕ(x) − 1.
Here, the question arises how the linear functional ϕ∗0 can be adequately formulated.
Obviously, the following formula defines a functional on the Banach space Y :
b
ϕ∗0 v(x) ≡ ϕ0 (x)T v(x)d x, v ∈ Y. (5.47)
a
The condition b
1 = ϕ∗0 ϕ0 (x) = ϕ0 (x)T ϕ0 (x)d x
a
can be realized with the function

x
ξ(x) ≡ ϕ0 (s)T ϕ0 (s)ds
a
as ξ(b) = 1. The function ξ(x) satisfies the following IVP
ξ (x) = ϕ0 (x)T ϕ0 (x), ξ(a) = 0. (5.48)
Replacing the last equation in (5.46) by the IVP (5.48) and the condition ξ(b) = 1
leads to a BVP where the number of ODEs and the number of boundary conditions
do not match. To correct this gap, we use a trick and add the trivial ODE λ = 0
(λ is constant) to the resulting system and obtain the following enlarged BVP of
dimension 2n + 2 for the numerical determination of simple turning points

ϕ (x) = f y (x, y(x); λ)ϕ(x), Ba ϕ(a) + Bb ϕ(b) = 0,
(5.49)
ξ (x) = ϕ(x)T ϕ(x), ξ(a) = 0, ξ(b) = 1,
λ (x) = 0.
Example 5.21 In Exercise 4.12, we have already considered Bratu’s BVP. The gov-
erning equations are
y (x) = −λ exp(y(x)), y(0) = y(1) = 0. (5.50)
As can be seen in Fig. 5.8, the corresponding solution curve has a simple turning
point z 0 ≡ (y0 , λ0 ).
To determine numerically z 0 , we have used the enlarged BVP (5.49), which reads
in that case

y1 (x) y2 (x)
= , y1 (0) = y1 (1) = 0,
y2 (x) −λ exp(y1 (x))

ϕ1 (x) ϕ2 (x)
= , ϕ1 (0) = ϕ1 (1) = 0,
ϕ2 (x) −λ exp(y1 (x))ϕ1 (x)
ξ (x) = ϕ1 (x)2 + ϕ2 (x)2 , ξ(0) = 0, ξ(1) = 1,
λ (x) = 0.
Applying the multiple shooting method (see Sect. 4.4) with m = 10 and the starting
trajectories y1 (x) = y2 (x) = ϕ1 (x) = ϕ2 (x) = ξ0 (x) = λ(x) ≡ 1, we obtained
Fig. 5.8 Solution curve of Bratu’s BVP
after eight iteration steps the approximations λ̃0 = 3.51383071912 for the critical
parameter λ0 and ỹ1 (0) = 4 for the missing initial value y1 (0).
Note, in (5.47) any linear functional ϕ∗0 can be used. Therefore, it is possible to
replace (5.47) by the functional
T
ϕ∗0 v(x) ≡ e(k) v(a), v ∈ Y, (5.51)
where e(k) denotes the k-the unit vector in Rn (see, e.g., [112]). The resulting enlarged
BVP of dimension 2n + 1 is


(k) T
λ (x) = 0, e ϕ(a) = 1.
The advantage of (5.52) is that its dimension is one less than the dimension of (5.51).
However, the major drawback of (5.52) is the specification of the index k. It must be
specified a priori which component of ϕ0 does not vanish. This is only possible if a
good approximation is at hand.
Another possibility for the realization of the functional ϕ∗0 is
ϕ∗0 v(x) ≡ ϕ0 (a)T v(a). (5.53)

In this way, we obtain the following enlarged BVP


λ (x) = 0, ϕ(a) ϕ(a) = 1,
T
which is of the same dimension as problem (5.52).

Let us now come to the realization of the other extension technique (5.45). In the
same way as before, we define the functional ψ ∗ ∈ W ∗ as
b
ψ ∗ w(x) ≡ ψ(x)T w(x)d x, w ∈ W, (5.55)
a
where ψ(x) is the solution of the adjoint BVP

T
ψ (x) = − f y (x, y(x); λ) ψ(x), a < x < b,
(5.56)
Ba∗ ψ(a) + Bb∗ ψ(b) = 0.
The adjoint matrices Ba∗ , Bb∗ ∈ Rn×n are to be determined such that
T T
Ba Ba∗ − Bb Bb∗ = 0, rank Ba∗ |Bb∗ = n.
This can be realized by computing the Q R factorization of the rectangular matrix

[Ba |Bb ]T via Givens transformations, i.e.,

BaT R Q 11 Q 12 R
=Q = .
BbT 0 Q 21 Q 22 0
Since Q is an orthogonal quadratic matrix, i.e., Q −1 = Q T , we have

T T
BaT Q 11 Q 21 BaT R
Q T
= = .
BbT T
Q 12 T
Q 22 BbT 0
Now, the desired adjoint matrices are Ba∗ = Q 12

T
and Bb∗ = −Q 22
T
.
In view of the above considerations, the enlarged problem (5.45) can be written
for the given problem (5.10) as the following BVP of dimension n + 2
y (x) = f (x, y(x), λ), Ba y(a) + Bb y(b) = 0,

T
ψ (x) = − f y (x, y(x); λ) ψ(x), Ba∗ ψ(a) + Bb∗ ψ(b) = 0
(5.57)
ξ (x) = −ψ(x)T f λ (x, y(x); λ), ξ(a) = 0, ξ(b) = 1,
λ (x) = 0.
The answer to the question, which of the two enlarged BVPs (5.49) or (5.57)
should be used, depends on the function vector f (x, y; λ) and the problem that has
to be solved afterwards (whether ϕ0 or ψ 0 is required).
5.3.3 An Extension Technique for Double Turning Points
Double (multiple) turning points can only be calculated numerically stable as two-
(multiple-) parameter problems. If the problem does not already contain two para-
meters, we must embed the original problem (5.11) into the two-parameter problem
T̃ (y, λ, τ ) = 0, T̃ : Y × R × R → W. (5.58)
For a fixed τ = τ ∗ , we have our one-parameter standard problem
T(τ ∗ ) (y, λ) ≡ T̃ (y, λ, τ ∗ ) = 0, T(τ ∗ ) : Y × R → W. (5.59)
A possible behavior of the solution curve y(τ ∗ ) (λ) ≡ y(λ; τ ∗ ) of (5.59) in dependence
of τ ∗ is displayed in Fig. 5.9. Here, the solution curve has two separated simple
turning points for τ ∗ > τ0 . But for τ ∗ = τ0 these two limit points collapse into a
double turning point, whereas for τ ∗ < τ0 the solution curve y(λ; τ ∗ ) consists of
only isolated solutions.
Fig. 5.9 τ = 0: double turning point

Fig. 5.10 τ = 0: double

turning point
Let z̃ 0 ≡ (y0 , λ0 , τ0 ) be a double limit point w.r.t. λ of the equation
T̃ (y, λ, τ0 ) = 0.
If the parameter τ is also varied in (5.58) and the corresponding simple turning points
are determined for each value of τ , then the projection of the limit points onto the
(λ, τ )-plane has the behavior of a cusp catastrophe (see Fig. 5.10). The canonical
equation of the cusp catastrophe is studied in the standard literature of this topic
(see, e.g., [43, 97, 133] and has the form
x 3 − τ x + λ = 0, x, λ, τ ∈ R.
It can be easily seen that:
• if λ = 0 is fixed, the qualitative behavior of the solution curves x versus τ is that

of a pitchfork bifurcation point (see Fig. 5.11),
• if τ = 0 is fixed, the corresponding solution curve x versus λ has a double turning
point (x, λ) = (0, 0) (see Fig. 5.9),
• if τ > 0, there are two simple turning points (x0 , λ0 ), with λ0 = ±2(τ /3)3/2 and
x0 = ±(τ /3)1/2 . It is λ20 = (4/27)τ 3 . This relation describes a cusp curve and the
corresponding cusp point is (λ, τ ) = (0, 0) (see Fig. 5.10).
• if τ < 0, there are no limit points.
This situation is described by the following theorem.

Theorem 5.22 Assume that
ξ0∗ T̃τ (y0 , λ0 , τ0 ) + ψ0∗ T̃yτ (y0 , λ0 , τ0 )ϕ0 = 0, (5.60)

Fig. 5.11 λ = 0: pitchfork

bifurcation
where ξ0∗ ∈ W ∗ is defined by (5.43). Then, the double turning point (y0 , λ0 , τ0 ) of
the 2-parameter problem (5.58) corresponds to a cusp point.
Proof See [116].
The correlation between a double turning point and the cusp catastrophe is also
shown by the following statement.
Theorem 5.23 Let the condition (5.60) be satisfied. Then, there exists a linear trans-
formation
[λ̂, τ̂ ] = [λ, τ ] P, P ∈ R2×2 ,
such that the double turning point (y0 , λ0 , τ0 ) of (5.58) is a pitchfork bifurcation
point with respect to τ̂ .
Proof See [104].
A suitable extended problem for the numerical determination of double turning

points has been proposed in [104] and is based on the enlarged operator
⎧
⎪
⎪ Ẑ ≡ Y × R × R × R × Y × W ∗ → Ŵ ≡ W × W × Y ∗ × R × R × R
⎪
⎪ ⎛ ⎞
⎪
⎪ T̃ (y, λ, τ )
⎪
⎪
⎪
⎨ ⎜ T̃y (y, λ, τ ) ϕ ⎟
⎜ ⎟
T̂ : ⎜ T̃ ∗ (y, μ, τ )ψ ∗ ⎟
⎪ ẑ ∗ ⎜
≡ (y, λ, τ , μ, ϕ, ψ ) → ⎜ y ⎟,
⎪
⎪ ϕ∗0 ϕ − 1 ⎟
⎪
⎪ ⎜ ⎟
⎪
⎪ ⎝ ∗
ψ ψ0 − 1 ⎠
⎪
⎪
⎩ ∗
ψ T̃yy (y, λ, τ )ϕ 2
(5.61)
with ϕ∗0 ∈ Y ∗ , ψ0 ∈ W such that
ϕ∗0 ϕ∗0 ϕ0 = ψ0∗ ψ0 = 1.
The equation
T̂ (ẑ) = 0 (5.62)
can be solved by standard methods as the following theorem shows.
Theorem 5.24 Let z̃ 0 ≡ (y0 , λ0 , τ0 ) be a double turning point of the equation (5.58)
w.r.t. the parameter λ. Suppose that
ξ0∗ T̃τ (y0 , λ0 , τ0 ) + ψ0∗ T̃yτ (y0 , λ0 , τ0 )ϕ0 = 0
and
ψ0∗ T̃yλ (y0 , λ0 , τ0 )ϕ0 = 0,
where ξ0∗ is defined by (5.43). Then, T̂ (ẑ 0 ) is a linear homeomorphism from Ẑ

onto Ŵ , i.e.,
N (T̂ (ẑ 0 )) = {0}, R(T̂ (ẑ 0 )) = Ŵ .
Proof See, e.g., the paper [104].
Obviously, the inversion of the statement postulated in Theorem 5.19 is not valid.
Only the isolated solutions of (5.62) with μ = 0 are in relationship with the double
turning points of the original problem (5.11).
A possible realization of the extended problem (5.62) for the BVP (5.10) is
y (x) = f̃ (x, y(x); λ, τ ), Ba y(a) + Bb y(b) = 0,

ϕ (x) = f̃ y (x, y(x); λ, τ )ϕ(x), Ba ϕ(a) + Bb ϕ(b) = 0,
T
ψ (x) = − f̃ y (x, y(x); μ, τ ) ψ(x), Ba∗ ψ(a) + Bb∗ ψ(b) = 0,
ξ1 (x) = ϕ(x)T ϕ(x), ξ1 (a) = 0, ξ1 (b) = 1,
ξ2 (x) = ψ(x) ψ(x),
T
ξ2 (a) = 0, ξ2 (b) = 1,

ξ3 (x) = ψ(x)T f̃ yy (x, y(x); λ, τ )ϕ(x)2 , ξ3 (a) = 0, ξ3 (b) = 0,

λ (x) = 0, μ (x) = 0, τ (x) = 0. (5.63)
This enlarged BVP of dimension 3n + 6 is again in standard form and can be solved
with the shooting methods described in Chap. 4.
5.3.4 Determination of Solutions in the Neighborhood of a

Simple Turning Point
Let z ≡ (y, λ) ∈ Z be an isolated solution of (5.11). If this solution is in the direct

neighborhood of a limit point, then its numerical determination is very complicated
since in that case the operator equation (5.11) represents a bad conditioned problem.
The idea is now to convert (5.11) into a modified problem, which is well-posed in
the neighborhood of the limit point. We call this strategy transformation technique.
In the following, we restrict our analysis to simple turning points. Multiple turning
points can be handled analogously.
According to Corollary 5.13, formula (5.28), in the neighborhood of limit points
there exists a unique curve of solutions of (5.11), which can be represented in the
form
λ(ε) = λ0 + ξ(ε),
y(ε) = y0 + εϕ0 + w(ξ(ε), ε), |ε| ≤ ε0 .
¨ = λ̈(0) = 0, then z 0 is a simple turning point and (5.28) can be

If ξ(ε) satisfies ξ(0)
specified as
λ(ε) = λ0 + ε2 τ (ε),
(5.64)
y(ε) = y0 + εϕ0 + ε2 u(ε), |ε| ≤ ε0 ,
where τ (0) = 0 and u(ε) ∈ Y1 , i.e., ϕ∗0 u(ε) = 0. In the ansatz (5.64), the unknown
quantities are τ (ε) and u(ε). We will now show how these quantities can be deter-
mined. At first, we expand T (y, λ) at the simple turning point z 0 = (y0 , λ0 ) into a
Taylor series and obtain
=0

1 0
0 = T (y, λ) = T (y0 , λ0 ) +Ty0 (y − y0 ) + Tλ (λ − λ0 ) + Tyy (y − y0 )2
2 (5.65)
1 0
+ Tyλ0
(λ − λ0 )(y − y0 ) + Tλλ (λ − λ0 )2 + R(y, λ).
2
The remainder R consists only of third or higher order terms in (y − y0 ) and (λ−λ0 ).
Now, we substitute the ansatz (5.64) into (5.65). We get
1 0 2
0 = Ty0 εϕ0 + ε2 u(ε) + ε2 Tλ0 τ (ε) + Tyy εϕ0 + ε2 u(ε)
2
ε 4
+ ε2 Tyλ
0
τ (ε) εϕ0 + ε2 u(ε) + Tλλ 0
τ (ε)2 + R(y(ε), λ(ε))
2
ε2 0 2
= ε2 Ty0 u(ε) + Tλ0 τ (ε) + Tyy ϕ0 + 2εϕ0 u(ε) + ε2 u(ε)2
2
ε4 0
+ ε3 Tyλ
0
ϕ0 τ (ε) + εu(ε)τ (ε) + Tλλ τ (ε)2 + R(y(ε), λ(ε)).
2
It can easily be seen that R(y(ε), λ(ε)) = O(ε3 ). Dividing both sides by ε2 yields
Ty0 u(ε) + Tλ0 τ (ε) = −w(u(ε), τ (ε); ε), ϕ∗0 u(ε) = 0,
where
1 0 2
w(u(ε), τ (ε); ε) ≡ Tyy ϕ0 + 2εϕ0 u(ε) + ε2 u(ε)2
2
ε2 0
+ εTyλ
0
ϕ0 τ (ε) + εu(ε)τ (ε) + Tλλ τ (ε)2
2
1
+ 2 R y0 + εϕ0 + ε2 u(ε), λ0 + ε2 τ (ε) .
ε
These two equations can be written in the form

Ty0 Tλ0 u(ε) w(u(ε), τ (ε); ε)
=− , |ε| ≤ ε0 . (5.66)
ϕ∗0 0 τ (ε) 0
In order to prove the next theorem, the following lemma is required:

Lemma 5.25 Let X and Y be Banach spaces. Assume that the linear operator A :
X × Rm → Y × Rm has the following block structure

A B
A≡ ,
C∗ D
where A : X → Y , B : Rm → Y , C ∗ : X → Rm and D : Rm → Rm .
Then, it holds:
1. A is bijective ⇒ {A is bijective ⇔ det(D − C ∗ A−1 B) = 0},
2. A is not bijective and dim N (A) = codimR(A) = m ≥ 1
⇒ {A is bijective ⇔ (c0 ) – (c3 ) are satisfied},
where
(c0 ) dim R(B) = m, (c1 ) R(B) ∩ R(A) = {0},
(c2 ) dim R(C ∗ ) = m, (c3 ) N (A) ∩ N (C ∗ ) = {0},
3. A is not bijective and dim N (A) > m ⇒ A is not bijective.
Theorem 5.26 Let z 0 be a simple turning point. Then, there exists a real constant
ε0 > 0, such that for |ε| ≤ ε0 the transformed problem (5.66) has a continuous
curve of isolated solutions η(ε) ≡ (u(ε), τ (ε))T . Substituting these solutions into
the ansatz (5.64), a continuous curve of solutions of the original operator equation
(5.11) results, which represents a simple solution curve passing through the simple
turning point z 0 ≡ (y0 , λ0 ) ∈ Z .
Proof For ε → 0 problem (5.66) turns into
0 1 0 2
u(0) Ty Tλ0 u(0) − Tyy ϕ0
A ≡ = 2 . (5.67)
τ (0) ϕ∗0 0 τ (0) 0
We set
A ≡ Ty0 , B ≡ Tλ0 , C ∗ ≡ ϕ∗0 , D ≡ 0, m ≡ 1
and apply Lemma 5.25. Since Ty0 is not bijective, the second case is true. Now, we
have to verify (c0 )–(c3 ).
(c0 ): dim R(Tλ0 ) = 1 = m,
(c1 ): R(Tλ0 ) ∩ R(Ty0 ) = {0},
because ψ0∗ Ty0 ϕ = 0 and ψ0∗ Tλ0 ζ = 0 for ζ = 0,

=0
(c2 ): dim R(ϕ∗0 ) = 1 = m,
(c3 ): N (Ty0 ) ∩ N (ϕ∗0 ) = {0},
because in N (Ty0 ) are the elements cϕ0 and ϕ∗0 (cϕ0 ) = c ϕ∗0 ϕ0 = c.

=1
Thus, it must hold c = 0.
The assumptions of Lemma 5.25 are satisfied and we can conclude that the linear
operator A is bijective. Therefore, there exists a unique solution η 0 ≡ (u(0), τ (0))T
of (5.67).
If problem (5.66) is written in the form
F(ε; η) = 0, η ≡ (u, τ )T ,
where F : R × dom(A) → W × R, then we have F(0; η 0 ) = Aη 0 − g, g ∈ W × R.

Thus, Fη (0; η 0 ) = A. The bijectivity of A implies that A is a linear homeomorphism
of dom(A) onto W × R. The claim of the theorem follows immediately from the
Implicit Function Theorem (see Theorem 5.8).
Thus, we can conclude that for all ε : |ε| ≤ ε0 , the transformed problem (5.66)
represents a well-posed problem, which can be solved by numerical standard tech-
niques.
Let us now formulate the transformed problem for the BVP (5.11). Here, we use
again the functional ϕ∗0 ∈ Y ∗ defined in (5.47), i.e.,
b
ϕ∗0 v(x) ≡ ϕ(x)T v(x)d x, v ∈ Y.
a
Then, the condition ϕ∗0 u(x) = 0 can be written as
ξ (x) = ϕ(x)T u(x), ξ(a) = ξ(b) = 0.

The fact that the parameter τ does not depend on x can be written in ODE formulation
as τ = 0. Thus, we obtain the following transformed BVP of dimension n + 2
u (x; ε) = f 0y u(x; ε) + f 0λ τ (ε) + w̃(u(x; ε), τ (ε); ε), |ε| ≤ ε0 ,

ξ (x; ε) = ϕ0 (x)T u(x; ε), τ (x) = 0, (5.68)
Ba u(a; ε) + Bb u(b; ε) = 0, ξ(a) = 0, ξ(b) = 0,
where
1
w̃(u(x; ε), τ (ε); ε) ≡ f 0yy ϕ20 + 2εϕ0 u(x; ε) + ε2 u(x; ε)2
2
ε2 0
+ ε f 0yλ τ (ε)ϕ0 + ετ (ε)u(x, ε) + f τ (ε)2
2 λλ
1
+ 2 R̃( y0 + εϕ0 + ε2 u(x; ε), λ0 + ε2 τ (ε)),
ε
and

1 0
R̃( y, λ) ≡ f ( y, λ) − f 0 + f 0y ( y − y0 ) + f 0λ (λ − λ0 ) +
f ( y − y 0 )2
2 yy

1 0
+ f yλ (λ − λ0 )( y − y0 ) + f λλ (λ − λ0 ) .
0 2
2
Remark 5.27 Obviously, the right-hand side of the transformed problem (5.68) is
based on the solution ( y0 , λ0 , ϕ0 ) of the enlarged problem (5.49). If a shooting
method is used to solve (5.68) both BVPs must be computed on the same grid. Since
in a shooting technique IVP-solvers with automatic step-size control are commonly
used it is appropriate to combine (5.68) and (5.49) into one BVP.
5.4 Analytical and Numerical Treatment of Primary

Simple Bifurcation Points
5.4.1 Bifurcation Points, Primary and Secondary Bifurcation

Phenomena
In this section, we will study solutions of the operator equation (5.11), which do not
lie on a simple solution curve. These are the so-called bifurcation points or branching
points.
Definition 5.28 A solution z 0 ∈ Z of the equation (5.11) is called bifurcation point

(see Fig. 5.12), if
Fig. 5.12 Illustration of Definition 5.28, part 1 and 2
1. in any open neighborhood of z 0 there exist two distinct solutions z 1 ≡ (y1 , λ1 )

and z 2 ≡ (y2 , λ2 ), with λ1 = λ2 ;
2. any open neighborhood of z 0 contains at least one solution of (5.11) for λ > λ0
as well as for λ < λ0 .

A necessary condition that a solution of (5.11) is a bifurcation point is given in

the following theorem.
Theorem 5.29 Let T : U (z 0 ) ⊆ Y × R → W be a C1 -map in a neighborhood of
z 0 ∈ Z . If z 0 is a bifurcation point of (5.11), then Ty (z 0 ) is not a homeomorphism.
Proof If Ty (z 0 ) is a homeomorphism, then the Implicit Function Theorem

(see Theorem 5.8) implies that in a neighborhood of z 0 all solutions of (5.11) lie
on a curve {y(λ), λ}, |λ − λ0 | ≤ ε. This is in contradiction to Definition 5.28.
From a numerical point of view, it is useful to distinguish between primary and

secondary bifurcation. In practice, {0, λ}, λ ∈ R, is often a solution curve of (5.11).
This curve is called the trivial solution curve. We speak of primary bifurcation
if nontrivial solution curves are branching off from the trivial solution curve. The
branching curves are called primary solution curves. The situation where at least
two nontrivial solution curves are intersecting is called secondary bifurcation and
the corresponding curves are the secondary solution curves (see also Fig. 5.6).
Here, we always assume that the secondary solution curves are not explicitly
known. Otherwise, if for a curve a parametrization {y(λ), λ}, λ ∈ [λ0 , λ1 ], is avail-
able, the given problem can be reduced to the case of primary bifurcation by consid-
ering the operator equation
T̃ (w, λ) ≡ T (y(λ) + w, λ) = 0.
5.4 Analytical and Numerical Treatment of Primary Simple … 195
5.4.2 Analysis of Primary Simple Bifurcation Points
In this section, we assume that the operator T in (5.11) satisfies
T (0, λ) = 0 for all λ ∈ R. (5.69)
Thus, there is the possibility that in the solution field of (5.11) primary bifurcation
occurs.
Definition 5.30 The parameter λ0 ∈ R is called geometrically simple eigenvalue of
Ty (0, λ) if
N (Ty (z 0 )) = span{ϕ0 }, z 0 ≡ (0, λ0 ) ∈ Z , ϕ0 ∈ Y, ϕ0 = 1. (5.70)

Looking at Theorem 5.6, we can conclude that Ty (0, λ) is a Fredholm operator
with index zero. Thus, the Riesz–Schauder theory presented in Sect. 5.3.1 is valid
here as well. However, for the characteristic coefficient (see formula (5.27)), we now
have
c ≡ ψ0∗ Tλ0 = 0, (5.71)
since Tλ0 = Tλλ

0
= Tλλλ
0
· · · = 0.
The decompositions (5.19) of the Banach spaces Y and W imply that y ∈ Y can
be uniquely represented in the form
y = αϕ0 + w, α ∈ R, P2 w = w. (5.72)
We set
λ = λ0 + ξ, (5.73)
and write the operator equation (5.11) in the paired form (see formula (5.23))
Q 2 T (z) = 0, z ≡ (0, λ) ∈ Z ,
(5.74)
Q 1 T (z) ≡ (I − Q 2 ) T (z) = 0.
Substituting (5.72) and (5.73) into the first equation of (5.74), we obtain
G(ξ, α, w) ≡ Q 2 T (αϕ0 + P2 w, λ0 + ξ) = 0, (5.75)
with the C p -map G : R × R × Y1 → R(Ty0 ). Here, the same arguments as in the

case of limit points (see Theorem 5.11) lead to the result that there exist two positive
constants ξ0 , α0 and a unique C p -map
w : [−ξ0 , ξ0 ] × [−α0 , α0 ] → Y1
such that
G(ξ, α, w(ξ, α)) = 0, w(0, 0) = 0.
If the second equation in (5.74) is also taken into account, we see that the treatment
of (5.11) in the neighborhood of z 0 ≡ (0, λ0 ) requires the solution of the so-called
bifurcation point equation
F(ξ, α) ≡ ψ0∗ T (αϕ0 + w(ξ, α), λ0 + ξ) = 0 (5.76)
in the neighborhood of the origin (0, 0).

Differentiating (5.76) gives
∂F ∂F
F(0, 0) = (0, 0) = (0, 0) = 0,
∂ξ ∂α
∂2 F ∂2 F ∂2 F
(0, 0) = 0, = a1 , (0, 0) = a2 ,
∂ξ 2 ∂ξ∂α ∂α2
where
a1 ≡ ψ0∗ Tyλ
0
ϕ0 , a2 ≡ ψ0∗ Tyy
0 2
ϕ0 (5.77)
are the first and the second bifurcation coefficient, respectively.

Now, let us suppose that the solution z 0 ≡ (0, λ0 ) ∈ Z of the equation (5.11) is a
simple bifurcation point.
Definition 5.31 A solution of the operator equation (5.11) is called simple bifurca-
tion point, if λ0 is a geometrically simple eigenvalue of Ty0 und the first bifurcation
coefficient a1 satisfies
a1 ≡ ψ0∗ Tyλ
0
ϕ0 = 0. (5.78)

Now, we will present the often used Morse Lemma. Assume that the equation
F( y) = 0
is given, where F ∈ C p (Ω, R), 0 ∈ Ω ⊂ Rd , Ω open region.

Lemma 5.32 Suppose that
F(0) = 0, Fy (0) = 0, and Fyy (0) is nonsingular.
Then, there exist a C p−2 -function w( y) ∈ Rd , which is defined in the neighborhood

of the origin and satisfies
w(0) = 0, w y (0) = I,
1
and F( y) = Fyy (0)w( y)2 in the neighborhood of the origin.
2
Proof See, e.g., [92]

The following corollary can easily be proved.
Corollary 5.33 Let the assumptions of Lemma 5.32 be satisfied. Moreover, suppose
that d = 2 and the quadratic form Fyy (0)w2 is indefinite. Then, the solution field
E(F) in the neighborhood of the origin consists of exactly two C p−2 curves, which
intersect (for p > 2 transversally) at the origin.
Let us apply Corollary 5.33 to the bifurcation equation (5.76). Here, d = 2 and
y = (ξ, α)T . It holds

Fξ (0, 0) 0
F(0, 0) = 0, Fy (0, 0) = = ,
Fα (0, 0) 0
and
Fξξ Fξα 0 a1
Fyy (0, 0) = (0, 0) = .
Fαξ Fαα a1 a2
Since det(Fyy (0, 0)) = −a12 < 0, the quadratic form is indefinite. Thus, we can
conclude that in the neighborhood of the origin the solutions of the equation (5.76)
lie on two curves of the class C p−2 , which intersect transversally at (0, 0).
For the original problem (5.11) this result implies that the statement of the fol-
lowing theorem is valid.
Theorem 5.34 Let z 0 ≡ (0, λ0 ) ∈ Z be a simple bifurcation point of the operator
equation (5.11). Then, in the neighborhood of z 0 the solutions of (5.11) belong to
two curves of the class C p−2 , which intersect transversally at z 0 .
Under the assumption (5.69), we have primary bifurcation and one of the solution
curves corresponds to the trivial solution (y ≡ 0, λ = λ0 + ξ). Let us parameterize
the other (nontrivial) solution curve in dependence of an artificial parameter ε as
follows
λ(ε) = λ0 + ξ(ε), y(ε) = α(ε)ϕ0 + w(ξ(ε), α(ε)),
(5.79)
ξ(ε) = εσ(ε), α(ε) = εδ(ε),
with w ∈ Y1 satisfying w(0, 0) = 0 (see formula (5.75)).

We define the function
H : (ε, σ, δ) ∈ [−1, 1] × [−ξ0 , ξ0 ] × [−α0 , α0 ] → R2 ,

T
F(εσ, εδ) 2
H(ε, σ, δ) ≡ , σ + δ 2
− 1 ,
ε2
and consider the equation

H(ε, σ, δ) = 0. (5.80)
The second equation in (5.80) represents only a certain normalization of the solution.
We can now state the following theorem.
Theorem 5.35 Let z 0 ≡ (0, λ0 ) ∈ Z be a simple bifurcation point of (5.11). Then,

there exist a positive constant ε0 and two C p−2 -functions
σ : [−ε0 , ε0 ] → R, δ : [−ε0 , ε0 ] → R,
with
H(ε, σ(ε), δ(ε)) = 0.
Proof Since F ∈ C p , p ≥ 2, and (0, 0) is a critical point of F, it holds
ε2 2
F(εσ, εδ) = a2 δ + 2a1 δσ + O(ε3 ), ε → 0.
2
If we substitute this expression into (5.80), we obtain
1 !
a2 δ 2 + 2a1 δσ
H(0, σ, δ) = 2 .
σ2 + δ2 − 1
The assumption a1 = 0 implies that a2 δ 2 + 2a1 δσ = 0 is the equation of two straight

lines in the (σ, δ)-plane, which intersect at the origin. These two lines are
• {δ10 = 0, σ10 arbitrary}, and
a2 δ 0
• {σ20 = − 2 , δ20 = 0}.
2a1
The first line is the trivial solution curve of (5.11). For the second line we have
"
∂ H "" a1 δ20 a2 δ20 + a1 σ20
J≡ " = .
∂(σ, δ) (0,σ20 ,δ20 ) 2σ20 2δ20
Thus, # 2 $
2
det(J ) = 2 a1 δ20 − a1 σ20 − a2 δ20 σ20

2 a2
= 2 δ20 a1 + 2 = 0.
4a1
Now, the application of the Implicit Function Theorem (see Theorem 5.8) to H at
(0, σ20 , δ20 ) leads to the following result. There exist real C p−2 -functions σ2 (ε) and
δ2 (ε) such that the relations
H(ε, σ2 (ε), δ2 (ε)) = 0, σ2 (0) = σ20 , δ2 (0) = δ20
are satisfied.
Let {(ξ(ε) = εσ(ε), α(ε) = εδ(ε)); |ε| ≤ ε0 } be the primary solution curve
of (5.76), which intersects the trivial solution curve {(ξ, α = 0); ξ arbitrary} at
the origin. Then, {(y(ε), λ(ε)); |ε| ≤ ε0 } and {(y = 0, λ); λ arbitrary} are the
corresponding solution curves of (5.11).
The currently used parametrization of the primary solution curve in the neighbor-
hood of a simple bifurcation point is based on an artificial real parameter ε. However,
sometimes it is possible to parameterize y directly by the control parameter λ. For
this purpose, we will discuss two different types of bifurcations.
• asymmetric bifurcation point: a2 = 0
It holds ξ(ε) = εσ(ε). Thus
dξ
(0) = σ(0).
dε
Using the formula
a2 δ20
σ20 = − (5.81)
2a1
presented in the proof of Theorem 5.35, we obtain σ(0) = 0. Now, the Implicit
Function Theorem implies that ε can be represented as a C p−2 -function of ξ. If
this function is substituted into the ansatz (5.79), we get
λ = λ0 + ξ, y = y(ξ) = α(ξ)ϕ0 + v(ξ, α(ξ)), (5.82)
where α is a C p−2 -function, which is defined in [−ξ0 , ξ0 ], and ξ0 is a sufficiently

small number.
To develop α(ξ) into a Taylor series, we compute
ξ δ(ξ)
ε= , i.e., α(ξ) = ξ ,
σ(ξ) σ(ξ)
dα δ(0)
α(0) = 0, (0) = .
dξ σ(0)
dα 2 a1 2 a1
Using (5.81), we obtain (0) = − = 0, and α(ξ) = − ξ + O(ξ 2 ).
dξ a2 a2
Thus, the ansatz (5.82) can be written in more detail as
2a1
y = y(ξ) = − ϕ0 ξ + O(ξ 2 ), ξ ≡ λ − λ0 . (5.83)
a2
• symmetric bifurcation point: a2 = 0, a3 = 0

Since a2 = 0, we get now from (5.81) that σ(0) = 0. Thus, differentiating the
relation ξ(ε) = εσ(ε), we see that
dξ
(0) = σ(0) = 0,
dε
and ε cannot be represented as a function of ξ. However, from
α(ε) = εδ(ε)
we obtain
dα
(0) = δ(0).
dε
Using the equation
σ(0)2 + δ(0)2 = 1,
we obtain
dα
(0) = δ(0) = 1.
dε
Thus, ε can be written as a C p−2 -function of α and the primary solution curve
takes on the form
λ = λ(α) = λ0 + ξ(α), y = y(α) = α ϕ0 + v(ξ(α), α), (5.84)
where ξ(α) is a C p−2 -function, which is defined in [−α0 , α0 ], α0 sufficiently

small.
In [24, 29] it is shown that the above representation can be written in more
detail as
a3 2
λ = λ(α) = λ0 − α + O(α3 ),
6 a1 (5.85)
y = y(α) = αϕ0 + O(α3 ).
Under the assumptions a3 = 0 and
6a1
C2 ≡ − > 0, (5.86)
a3
the first equation in (5.85) can be written in the form α = α(λ). Substituting this
relation into the second equation, we obtain the following representation of y in
fractional powers of λ − λ0 :

y = y(η) = C2 ϕ0 η + O(η 2 ), η 2 ≡ λ − λ0 . (5.87)
In Fig. 5.13, both types of a primary simple bifurcation point are graphically illus-
trated.
Fig. 5.13 Two different types of primary simple bifurcation points
5.4.3 An Extension Technique for Primary Simple

Bifurcation Points
Let z 0 ≡ (0, λ0 ) be a simple bifurcation point. An often used extension technique

for the determination of λ0 is based on the enlarged operator
⎧
⎨ Z̃ ≡ R × Y →W̃ ≡ W × R
T̃ : Ty (0, λ) ϕ , (5.88)
⎩ z̃ ≡ (λ, ϕ) →
ϕ∗0 ϕ − 1
where ϕ∗0 ∈ Y ∗ such that ϕ∗0 ϕ0 = 1. Since λ0 is a geometrically simple eigenvalue,

the point z̃ 0 = (λ0 , ϕ0 ) ∈ Z̃ satisfies the equation
T̃ (z̃) = 0. (5.89)
For the extended problem (5.89) we have the following result.

Theorem 5.36 Let z 0 ≡ (0, λ0 ) ∈ Z be a simple bifurcation point of T . Then, the
extended operator T̃ (z̃ 0 ) is a linear homeomorphism from Z̃ onto W̃ , i.e.,
N (T̃ (z̃ 0 )) = {0}, R(T̃ (z̃ 0 )) = W̃ .

Proof It is

0
Tyλ ϕ0 Ty0
T̃ (z̃ 0 ) = .
0 ϕ∗0
1. If (μ, w)T ∈ N (T̃ (z̃ 0 )), where w ∈ Y and μ ∈ R, we have

0
μ μTyλ ϕ0 + Ty0 w
T̃ (z̃ 0 ) = = 0. (5.90)
w ϕ∗0 w
The first (block) equation of (5.90) reads
Ty0 w = −μTyλ
0
ϕ0 . (5.91)
This equation is solvable, if the right-hand side is in the range of the linear
operator Ty0 . The Fredholm alternative says, that this is the case if the condition
μ ψ0∗ Tyλ
0
ϕ0 = 0 is satisfied (see Eqs. (5.20) and (5.21)). Now, the assumption
a1 ≡ ψ0∗ Tyλ 0
ϕ0 = 0 implies μ = 0. Inserting this value into (5.91), we get
w = c ϕ0 , c ∈ R. The second (block) equation of (5.90) is
ϕ∗0 w = 0.
Now, we have ϕ∗0 w = c ϕ∗0 ϕ0 = 0. The condition ϕ∗0 ϕ0 = 1 implies c = 0, i.e.,

w = 0.
2. Let (u, σ)T ∈ R(T̃ (z̃ 0 )) be satisfied. Now, instead of the homogeneous equation
(5.90) we consider the inhomogeneous equation
0
μ μTyλ ϕ0 + Ty0 w u
T̃ (z̃ 0 ) = = . (5.92)
w ϕ∗0 w σ
The first (block) equation of (5.92) is
Ty0 w = u − μ Tyλ
0
ϕ0 . (5.93)
There exists a solution if ψ0∗ (u − μ Tyλ

0
ϕ0 ) = ψ0∗ u − μ ψ0∗ Tyλ
0
ϕ0 = 0. Thus, μ is
uniquely determined as
ψ∗ u
μ= 0 .
a1
The solution of (5.93) can be written in the form
w = α ϕ0 + ŵ, ŵ ∈ Y1 ,
where α is an arbitrary real constant. Inserting this representation of w into the

second (block) equation of (5.92),
ϕ∗0 w = σ,
we see that the constant α is uniquely determined as
α = −ϕ∗0 ŵ + σ,
and thus, w is uniquely determined, too.
If z 0 denotes a primary simple bifurcation point, then the reversal of the statement
of Theorem 5.36 is also true.
Theorem 5.37 Let z̃ 0 = (λ0 , ϕ0 ) be an isolated solution of the extended problem

(5.89). Then, z 0 = (0, λ0 ) is a primary simple bifurcation point of the original
problem (5.11).
Proof We show the claim of the theorem in two steps.

1. Assume that there exists a 0 = ϕ1 ∈ Y such that Ty0 ϕ1 = 0 and ϕ∗0 ϕ1 = 0. Then,
(0, ϕ1 ) ∈ N (T̃ (z̃ 0 )). Due to the assumption of our theorem it must hold ϕ1 = 0.
2. Assume that a1 ≡ ψ0∗ Tyλ 0
ϕ0 = 0. Let w0 denote the unique solution of the problem
Ty0 w0 = −Tyλ
0
ϕ0 , w0 ∈ Y1 ,

1
and Φ0 ≡ α , α = 0. Then, it can be easily seen that Φ0 ∈ N (T̃ (z̃ 0 )).
w0 ,
This is in contradiction to our assumption N (T̃ (z̃ 0 )) = {0}. Thus, it must hold
that a1 = 0.
Theorems 5.36 and 5.37 show that the extended problem (5.89) can be solved
with numerical standard techniques. A possible realization of the extended problem
(5.89) for the BVP (5.10) is
ϕ (x) = f y (x, 0; λ)ϕ(x), Ba ϕ(a) + Bb ϕ(b) = 0,

ξ (x) = ϕ(x) ϕ(x),
T
ξ(a) = 0, ξ(b) = 1, (5.94)
λ (x) = 0.
Obviously, the dimension of the extended BVP is n + 2.
Example 5.38 In Sect. 5.1 we have shown that the primary simple bifurcation points
of the BVP (5.6),
ϕ (s) = −λ sin(ϕ(s)),
ϕ (0) = ϕ (1) = 0,
are z 0(k) = (0, λ(k) (k)

0 ), with λ0 = (kπ) , k = 1, 2, . . ..
2
The vector-matrix formulation of this BVP is

y (x) y2 (x)
y (x) = 1 = f (x, y(x); λ) = ,
y2 (x) −λ sin(y1 (x))

01 00
Ba y(0) + Bb y(1) = y(0) + y(1) = 0.
00 01
We have
0 1
f y (x, y(x); λ) = , and thus
−λ cos(y1 (x)) 0

0 1
f y (x, 0; λ) = .
−λ 0
Now, it can be easily seen that the associated extended problem (5.94) is
ϕ1 (x) = ϕ2 (x), ϕ2 (0) = ϕ2 (1) = 0,

ϕ2 (x) = −λ(x) ϕ1 (x)

(5.95)
ξ (x) = ϕ1 (x)2 + ϕ2 (x)2 , ξ(0) = 0, ξ(1) = 1,
λ (x) = 0.
We have used the multiple shooting code RWPM to determine the solutions of the
extended problem (5.95). Here we present the results for the first bifurcation point.
Our computations are based on ten shooting points τ j = j · 0.1, j = 0, 1, . . . , 9, the
IVP-solver ODE45 from Matlab and a tolerance of 10−6 . For the starting values it
is important to impress the qualitative behavior of the eigensolution ϕ1 (x) upon the
starting trajectory. Therefore, we have used the following starting values
ϕ(0) (0) (0) (0)

1 (τ j ) = 1 − τ j , ϕ2 (τ j ) = −1, ϕ3 (τ j ) = τ j , ϕ4 (τ j ) = 5; j = 0, . . . , 9.
After six iterations, we obtained the approximation λ̃(1)

0 = 9.869040 for the first
critical value λ(1)
0 = π 2
. The corresponding approximation of the eigensolution ϕ1 (x)
is shown in Fig. 5.14.

Primary Simple Bifurcation Point
Let z ≡ (y, λ) ∈ Z be an isolated solution of the operator equation (5.11), which is

located in the neighborhood of a primary simple bifurcation point z 0 ≡ (0, λ0 ).
An analytical technique to determine z has been proposed by Levi-Cività, see [74].
It is based on the following representation of the nontrivial solution branch
Fig. 5.14 Numerically determined eigensolution ϕ̃1 (x) ≈ ϕ1 (x) of problem (5.95)
∞
%
λ(ε) = λ0 + si εi ,
i=1
∞
(5.96)
%
y(ε) = εϕ0 + ui ε ,i
ϕ∗0 u i = 0.
i=2
Here, ε is a sufficiently small parameter, say |ε| ≤ ε0 . The unknown u k and sk are
determined by
• expanding the operator T at z 0 into a Taylor series and substituting this series into
the operator equation (5.11),
• substituting the ansatz (5.96) into the resulting equation,
• comparing the terms, which have the same power of ε, and using the orthogonality
relation.
Example 5.39 For the governing equations (5.6) of the Euler-Bernoulli rod, we will
demonstrate the method of Levi-Cività. In Sect. 5.1 we have shown that the eigen-
values and the corresponding (normalized) eigensolutions of the linearized problem
are √
λ(k) (k)
0 = (kπ) , ϕ0 (x) =
2
2 cos(kπx), k = 1, 2, . . . .
Here, we set k = 1. Let us consider the first eigenvalue and the corresponding
eigensolution. Therefore, we set
√
λ0 ≡ λ(1) (1)
0 = π , ϕ0 ≡ ϕ0 (x) =
2
2 cos(πx).
To simplify the computation, we do not transform the second-order ODE into a

system of two first-order equations, but instead, we set
T (y, λ) ≡ y (x) + λ sin(y(x)) = 0. (5.97)
Since our linearized second-order ODE (see formula (5.7)) is self-adjoint, we have
ψ0∗ = ϕ∗0 . As before, we use the functional
1
ϕ∗0 y(x) ≡ ϕ0 (x)y(x)d x.
0
The first step of the method of Levi-Cività is to expand T at z 0 ≡ (0, λ0 ) into a

Taylor series. In our case, this means that the sinus function must be expanded into
a Taylor series. We obtain
y3 y5 y7
sin(y) = y − + − ± ··· ,
3! 5! 7!

y3 y5 y7
T (y, λ) = y + λ y − + − ± ··· = 0. (5.98)
3! 5! 7!
In the second step, the ansatz (5.96) has to be inserted into (5.98). It results
∞
∞
& ∞

% % %
εϕ0 + i
u k i ε + λ0 + si εi
εϕ0 + ui ε i
i=2 i=1 i=2

3 5 ⎤
∞
% ∞
%
1 1
− εϕ0 + u i εi + εϕ0 + u i εi + O(ε7 )⎦ = 0. (5.99)
6 i=2
120 i=2
In the last step, we have to compare the terms which have the same power of the
parameter ε. We obtain:
• Terms in ε
ϕ0 + λ0 ϕ0 = 0.
• Terms in ε2
u 2 + λ0 u 2 = −s1 ϕ0 , u 2 ∈ Y1 .
This equation is solvable, if s1 ψ0∗ ϕ0 = 0. It follows s1 = 0. Thus, we have to solve

the BVP
u 2 + λ0 u 2 = 0, u 2 (0) = u 2 (1) = 0.
Due to the orthogonality condition ϕ∗0 u 2 = 0, this homogeneous BVP has the
unique solution u 2 (x) ≡ 0.
• Terms in ε3
1
u 3 + λ0 u 3 = −s2 ϕ0 + λ0 ϕ30 , u 3 ∈ Y1 .
6
This equation is solvable, if
1
−s2 ψ0∗ ϕ0 + λ0 ψ0∗ ϕ30 = 0, i.e.,
6
1 √ 4
1 ∗ 3 1 1 3 1
s 2 = λ 0 ψ 0 ϕ0 = λ 0 2 cos(πx) d x = λ0 = λ0 .
6 6 0 6 2 4
Thus, we have to solve the BVP
1 1
u 3 + λ0 u 3 = − λ0 ϕ0 + λ0 ϕ30 , u 3 (0) = u 3 (1) = 0. (5.100)
4 6
It is
1 1
u 3 + λ0 u 3 = − λ0 ϕ0 + λ0 ϕ30
4 6
1 2√ 1 √
= − π 2 cos(πx) + π 2 2 2 cos3 (πx)
4√ √6
2 2 2 2 3 1
=− π cos(πx) + π cos(πx) + cos(3πx)
4 3 4 4
√ √ √
2 2 2 2 2 2
= − π + π cos(πx) + π cos(3πx)
4 4 12
√
2 2
= π cos(3πx).
12
To determine a particular solution of this ODE, we use the ansatz
u 3 (x) = C cos(3πx),
which satisfies the boundary conditions, too. We have
u 3 (x) = C cos(3πx), u 3 (x) = −3πC sin(3πx),

u 3 (x) = −9π 2 C cos(3πx).
Substituting u 3 (x) and u 3 (x) into the ODE, gives

√
2 2
−9π C cos(3πx) + π C cos(3πx) =
2 2
π cos(3πx).
12
√
2
A comparison of the coefficients of the cosine function yields C = − . Thus,
96
the solution u 3 ∈ Y1 of the BVP (5.100) is
√
2
u 3 (x) = − cos(3πx).
96
• Terms in ε4
1 3
u 4 + λ0 u 4 = −s1 u 3 − s2 u 2 − s3 ϕ0 + s1 ϕ0 + λ0 ϕ20 u 2
6
= −s3 ϕ0 , u 4 ∈ Y1 .
This equation is solvable if s3 ψ0∗ ϕ0 = s3 ϕ∗0 ϕ0 = 0. It follows s3 = 0. Thus, we

have to solve the BVP
u 4 + λ0 u 4 = 0, u 4 (0) = u 4 (1) = 0.
Due to the condition ϕ∗0 u 4 = 0, the unique solution is u 4 (x) ≡ 0.

• Terms in ε5
λ0 2 1 1
u 5 + λ0 u 5 = ϕ u3 − λ0 ϕ50 − s2 u 3 + s2 ϕ30 − s4 ϕ0 ≡ R,
2 0 120 6
u 5 ∈ Y1 .
This equation is solvable if ψ0∗ R = ϕ∗0 R = 0, i.e.,
λ0 ∗ 2 1 1
s4 = ϕ ϕ u3 − λ0 ϕ∗0 ϕ50 − s2 ϕ∗0 u 3 + s2 ϕ∗0 ϕ30
2 0 0 120 6
1 1
1 1
= − π2 cos3 (πx) cos(3πx)d x − π 2 cos6 (πx)d x
48 0 15 0
1
1 2 1 1
+ π cos(3πx) cos(πx)d x + π 2 cos4 (πx)d x
192 0 6 0
1 1 1 5 1 3
= − π2 − π2 + π2
48 8 15 16 6 8
5 2
= π .
128
Substituting this value into the right-hand side of the ODE, we obtain
√
2 2 1 1 1
R= − π cos(πx) + cos(3πx) + cos(5πx)
96 4 2 4
√
2 2 5 5 1
− π cos(πx) + cos(3πx) + cos(5πx)
30 8 16 16
√ √
2 2 2 2 3 1
+ π cos(3πx) + π cos(πx) + cos(3πx)
384 12 4 4
√
5 2 2
− π cos(πx)
128 √ √
2 3 2
= π cos(πx) · 0 + π cos(3πx) ·
2 2
+ π cos(5πx) · −
2
.
128 640
Now, we have to solve the BVP
u 5 + π 2 u 5 = R, u 5 (0) = u 5 (1) = 0. (5.101)
To determine a particular solution of the ODE, we use the ansatz
u 5 = C3 cos(3πx) + C5 cos(5πx).
It is
u 5 = −3πC3 sin(3πx) − 5πC5 sin(5πx),
u 5 = −9π 2 C3 cos(3πx) − 25π 2 C5 cos(5πx).
Therefore
u 5 + π 2 u 5 = −8π 2 C3 cos(3πx) − 24π 2 C5 cos(5πx).
Equating the right-hand side of this ODE with the above calculated expression
for R, we obtain √ √
2 2
C3 = − and C5 = .
1024 5120
Thus, the solution u 5 ∈ Y1 of the BVP (5.101) is
√ √
2 2
u 5 (x) = − cos(3πx) + cos(5πx).
1024 5120
If we stop the algorithm at this point, we have obtained the following representation
of the nontrivial solution curve branching at z 0 = (0, λ0 ) = (0, π 2 ):
√
√ 2
y(x; ε) = ε 2 cos(πx) − ε 3
cos(3πx)
√ 96
√
2 2
+ε −5
cos(3πx) + cos(5πx) + O(ε7 ), (5.102)
1024 5120

1 2 5 4
λ(ε) = π 1 + ε +
2
ε + O(ε ) , |ε| ≤ ε0 .
6
4 128
More generally, the nontrivial solution curves branching at z 0(k) ≡ (0, (kπ)2 ), k =
1, 2, . . ., can be parametrized in the form
√
√ 2
y(x; ε) = ε 2 cos(kπx) − ε3 cos(3kπx)
√ 96
√
2 2
+ ε5 − cos(3kπx) + cos(5kπx) + O(ε7 ), (5.103)
1024 5120

1 5 4
λ(ε) = (kπ)2 1 + ε2 + ε + O(ε6 ) , |ε| ≤ ε0 .
4 128
In Fig. 5.15 it is shown that the graph (y(x; ε), λ(ε)) comes closer to the exact solution
graph (y(x), λ) if the number of terms in the series (5.102) is increased.
Fig. 5.15 Dashed line three terms of the series (5.102) are used, solid line four terms of the series
(5.102) are used, gray line exact solution curve. Here, we have set l ∗ y ≡ y(0)
Before we describe how a nontrivial solution curve branching at a primary simple

bifurcation point can be determined numerically, let us demonstrate the Levi-Cività
technique for a second example.
Example 5.40 Consider the following BVP
y (x) = λ(y(x) + y(x)2 ),

(5.104)
y(0) = y(1) = 0.
The primary simple bifurcation points are
z 0(k) = (0, λ(k) (k)

0 ), λ0 = −(kπ) , k = 1, 2, . . . .
2
The corresponding (normalized) eigenfunctions are

√
ϕ(k)
0 (x) = 2 sin(kπx), k = 1, 2, . . . .
In order to simplify the notation, let us set

√
λ0 ≡ λ(k)
0 and ϕ0 (x) ≡ 2 sin(kπx).
Substituting the ansatz (5.96) into the ODE yields

∞
%
ϕ0 ε+ u i εi =
i=2
∞
⎛ ∞
∞
2 ⎞
% % %
λ0 + si εi ⎝ϕ0 ε + u i εi + ϕ0 ε + u i εi ⎠ . (5.105)
i=1 i=2 i=2
Now, we compare the terms, which have the same power of the parameter ε, and
obtain:
• Terms in ε
ϕ0 − λ0 ϕ0 = 0.
• Terms in ε2
u 2 − λ0 u 2 = s1 ϕ0 + λ0 ϕ20 , u 2 ∈ Y1 .
This equation is solvable if ϕ∗0 {s1 ϕ0 + λ0 ϕ20 } = 0, where we use again the fact
that ψ0∗ = ϕ∗0 (see Example 5.39). Thus,
√ 1
s1 = −λ0 ϕ∗0 ϕ20 = −λ0 2 2 sin3 (kπx)d x
0
"1
√ 1 "
= 2 2(kπ) cos3 (kπx) − cos(kπ) ""
3
⎧ 0
⎨ √0, k even
= 8 2
⎩ kπ, k odd.
3
Let us substitute s1 into the ODE and distinguish between even and odd values
of k. If k is even, we have to solve the BVP
u 2 + (kπ)2 u 2 = −2(kπ)2 sin2 (kπx) = −(kπ)2 (1 − cos(2kπx)),

(5.106)
u 2 (0) = u 2 (1) = 0.
To determine a particular solution of this equation, we use the ansatz
u 2 = C1 cos(kπx) + C2 sin(2kπx) + C3 cos(2kπx)

(5.107)
+ C4 x cos(kπx) + C5 .
It follows
u 2 = − C1 (kπ) sin(kπx) + 2C2 (kπ) cos(2kπx)

− 2C3 (kπ) sin(2kπx) + C4 [cos(kπx) − x(kπ) sin(kπx)] ,
and
u 2 = − (kπ)2 [C1 cos(kπx) + 4C2 sin(2kπx) + 4C3 cos(2kπx)]
− (kπ)C4 [2 sin(kπx) + (kπ)x cos(kπx)] .
Thus,
u 2 + (kπ)2 u 2 = − 3C2 (kπ)2 sin(2kπx) − 3C3 (kπ)2 cos(2kπx)

(5.108)
− 2C4 (kπ) sin(kπx) + (kπ)2 C5 .
Comparing the coefficients in (5.106) and (5.108), we obtain
1
C2 = 0, C3 = − , C4 = 0, C5 = −1.
3
However, the coefficient C1 is still undetermined. Substituting the coefficients into

the ansatz (5.107) and using the first boundary condition, we have
1 . 4
u 2 (0) = C1 − − 1 = 0, i.e., C1 = .
3 3
Obviously, for this value of C1 the second boundary condition is automatically
satisfied. Therefore, the particular solution of (5.106) is
1
u 2 (x) = (4 cos(kπx) − cos(2kπx)) − 1. (5.109)
3
Let us now assume that k is odd. Instead of (5.106), we have to solve the BVP
16
u 2 + (kπ)2 u 2 = (kπ) sin(kπx) − (kπ)2 (1 − cos(2kπx)),
3 (5.110)
u 2 (0) = u 2 (1) = 0.
Comparing the coefficients in (5.106) and (5.110), we obtain
1 8
C2 = 0, C3 = − , C4 = − , C5 = −1.
3 3
As before, the coefficient C1 is still undetermined. Substituting the coefficients
into the ansatz (5.107) and using the first boundary condition, we have
1 . 4
u 2 (0) = C1 − − 1 = 0, i.e., C1 = .
3 3
For this value of C1 the second boundary condition is automatically satisfied.
Therefore, the particular solution of (5.110) is
1
u 2 (x) = (4 cos(kπx) − cos(2kπx) − 8x cos(kπx)) − 1. (5.111)
3
• Terms in ε3
u 3 − λ0 u 3 = s2 ϕ0 + 2λ0 ϕ0 u 2 + s1 u 2 + s1 ϕ20 , u 3 ∈ Y1 .
This equation is solvable if ϕ∗0 {s2 ϕ0 + 2λ0 ϕ0 u 2 + s1 u 2 + s1 ϕ20 } = 0.

First, let us consider the case that k is even. It must hold (note that s1 = 0)
.
0 = ϕ∗0 {s2 ϕ0 + 2λ0 ϕ0 u 2 }
1√
√ √
= 2 sin(kπx) s2 2 sin(kπx) − 2(kπ)2 2 sin(kπx)u 2 d x.
0
Substituting the expression (5.109) for u 2 into this equation, we get

1
0 = 2s2 sin2 (kπx)d x
0
1
4 1
− 4(kπ) 2
sin (kπx) cos(kπx) − cos(2kπx) − 1 d x
2
0 3 3

1 1
= 2s2 + (kπ)2 − + 2 .
2 3
Thus,
5
s2 = − (kπ)2 . (5.112)
3
Now, we assume that k is odd. It must hold
.
0 =ϕ∗0 {s2 ϕ0 + 2λ0 ϕ0 u 2 + s1 u 2 + s1 ϕ20 }
1√
√ √ 2
= 2 sin(kπx) s2 2 sin(kπx) − 2 2(kπ)2 sin(2kπx)
0 3

1 4
− sin(kπx) cos(2kπx) − x sin(2kπx) − sin(kπx)
3 3

4 1 8
+ s1 cos(kπx) − cos(2kπx) − x cos(kπx)
3 3 3

− 1 + 2 sin2 (kπx) dt

1 2 1 1 4 8 1
=2s2 · − 4(kπ)2 ·0+ · + · −
2 3 3 4 3 9(kπ)2 2

√ 4 1 2 8 1 1 4
2s1 ·0+ · + · −1· +2·
3 3 3(kπ) 3 4(kπ) (kπ) 3(kπ)
5 32
=s2 + (kπ)2 + .
3 9
Thus,
32 5
s2 = − − (kπ)2 . (5.113)
9 3
Summarizing the above computations, we have the following asymptotic expan-
sion of the branching solutions:
• k is odd
√
y(x; ε) = 2 sin(kπx) ε

4 1 8
cos(kπx) − cos(2kπx) − x cos(kπx) − 1 ε2 + O(ε3 ),
3 3 3
√
8 2 32 5
λ(ε) = − (kπ) + 2
(kπ) ε − + (kπ) ε2 + O(ε3 ).
2
3 9 3
(5.114)
• k is even
√
y(x; ε) = 2 sin(kπx) ε

4 1
+ cos(kπx) − cos(2kπx) − 1 ε2 + O(ε3 ), (5.115)
3 3

5
λ(ε) = − (kπ)2 − (kπ)2 ε2 + O(ε4 ).
3
Note, these expansions depend on a small parameter ε: |ε| ≤ ε0 , which has been
artificially introduced into the problem.
In Fig. 5.16, the exact solution graph (y(x), λ) and the graph (y(x; ε), λ(ε)) are
represented.
These two examples demonstrate impressively that the representation of the

branching solutions as infinite sums of the form (5.96) is extremely work-intensive
Fig. 5.16 Dotted/dashed line exact solution curve, solid line the asymptotic expansion (5.114) of
the branching solutions. Here, we have set l ∗ y ≡ y (0)
and very susceptible to arithmetical errors. Therefore, we will now describe a numer-
ical approach (see [129, 130]), which is based on the idea that the infinite sums are
expressed as functions, which are the isolated solutions of a standard BVP.
Instead of (5.96), we use the following representation of an isolated solutions
z ≡ (y, λ) ∈ Z of the operator equation (5.11), which is located in the neighborhood
of a primary simple bifurcation point z 0 ≡ (0, λ0 ):
y(ε) = εϕ0 + ε2 u(ε), u(ε) ∈ Y1 , |ε| ≤ ε0 ,

(5.116)
λ(ε) = λ0 + ετ (ε), τ (ε) ∈ R.
The assumption (5.69) implies that
Tλ0 = Tλλ
0
= · · · = 0. (5.117)
Therefore, the formula (5.65) simplifies to
1 0 2
0 = T (y, λ) = Ty0 y + Tyy y + Tyλ
0
λy + R(y, λ). (5.118)
2
The remainder R consists only of third or higher order terms in y and λ. Now, we
substitute the ansatz (5.116) into (5.118), and obtain

Ty0 0
Tyλ ϕ0 u w(u, τ ; ε)
=− , |ε| ≤ ε0 , (5.119)
ϕ∗0 0 τ 0
where
1 0 2 1 0
w(u, τ ; ε) ≡ Tyy ϕ0 + ε Tyy (2ϕ0 u + εu 2 ) + εTyλ
0
τu
2 2
1
+ 2 R(εϕ0 + ε2 u, λ0 + ετ ).
ε
The following theorem gives information about the solutions of the transformed
problem (5.119).
Theorem 5.41 Let the conditions (5.69), (5.70) and (5.78) be satisfied. Then, there
exists a real constant ε0 > 0, such that for |ε| ≤ ε0 the transformed problem (5.119)
has a continuous curve of isolated solutions η(ε) ≡ (u(ε), τ (ε))T . Substituting these
solutions into the ansatz (5.116), a continuous curve of solutions of the original
operator equation (5.11) results, which represents a simple solution curve passing
through the simple primary bifurcation point z 0 ≡ (0, λ0 ) ∈ Z .
Proof For ε → 0 problem (5.119) turns into

0 1
u(0) Ty 0
Tyλ ϕ0 u(0) 0 2
Tyy ϕ0
A ≡ =− 2 . (5.120)
τ (0) ϕ∗0 0 τ (0) 0
The assumptions and Lemma 5.25, part 2, imply that

• A is bijective, and
• there exists a unique solution η 0 ≡ (u 0 , τ 0 )T of (5.120), with
1 ∗ 0 −1 ∗ 0 2
τ0 = − ψ T ϕ0 ψ0 Tyy ϕ0 ,
2 0 yλ
+ 1 0 2
u 0 = − Ty0 0
Tyλ ϕ0 τ 0 + Tyy ϕ0 ,
2
+
where Ty0 denotes the generalized inverse of Ty0 , which maps R(Ty0 ) one-to-one
onto Y1 .
If problem (5.119) is written in the form
F(ε; η) = 0, η ≡ (u, τ )T ,
then the claim follows, as in the proof of Theorem 5.26, from the Implicit Function
Theorem (see Theorem 5.8).
Applying, as shown in Sect. 5.3.4, the transformed problem (5.119) to the BVP
(5.10) and using the functional (5.47), the following transformed BVP of dimension
n + 2 results
u (x; ε) = f 0y u(x; ε) + f 0yλ ϕ0 (x)τ (ε) + w̃(u(x; ε), τ (ε); ε), |ε| ≤ ε0 ,
ξ (x; ε) = ϕ0 (x)T u(x; ε), τ (x) = 0,
Ba u(a; ε) + Bb u(b; ε) = 0, ξ(a) = 0, ξ(b) = 0,
(5.121)
where
1 0 2 ε 0
w̃(u, τ ; ε) ≡ f yy ϕ0 + f yy 2ϕ0 u + εu2
2 2
1
+ ε f yλ τ u + 2 R̃(εϕ0 + ε2 u, λ0 + ετ ),
0
ε
and
1 0 2
R̃( y, λ) ≡ f ( y, λ) − f 0y y + f yy y + f yλ (λ − λ0 ) y .
0
2
For all ε with |ε| ≤ ε0 , the BVP (5.121) can be solved by the shooting methods
presented in Chap. 4.
Remark 5.42 The quantities λ0 and ϕ0 , which are required in the ansatz (5.116), can
be determined from the extended problem (5.89). Usually, the two problems (5.121)
and (5.89) are combined into a single (larger) problem. In the case of parametrized
BVPs, the advantage of this strategy is that both problems are defined on the same
grid. Thus, IVP-solver with an automatic step-size control can be used in the shooting
methods.
The transformation techniques discussed so far, are based on the ansatz (5.96)
or (5.116), which depends on an additional (artificial) parameter ε. However, many
engineering and physical models require the computation of the solution y for pre-
scribed values of the (control) parameter λ, which is already contained in the given
problem. In the neighborhood of primary simple bifurcation points, the approaches
(5.83) and (5.87) enable the determination of nontrivial solutions of the operator
equation (5.11) in direct dependence of λ. But the disadvantage of this strategy is
that the exact type of the bifurcation phenomenon must be known a priori and has to
be taken into account.
First, we want to consider the case of a symmetric bifurcation point, i.e., let us
assume that
a2 ≡ ψ0∗ Tyy
0 2
ϕ0 = 0 and a3 ≡ ψ0∗ (6Tyy
0
ϕ0 v2 + Tyyy
0
ϕ30 ) = 0, (5.122)
where v2 ∈ Y1 is the unique solution of the linear operator equation
1 0 2
Ty0 v2 = − Tyy ϕ0 . (5.123)
2
In analogy to the analytical approach of Levi-Cività, the following ansatz for the
nontrivial branching solutions has been proposed by Nekrassov, see [90, 91]:
% ∞
2
y(x; ξ) = √ ϕ0 (x) ξ + (si ϕ0 (x) + u i (x))ξ i ,
λ0 i=2 (5.124)
ϕ∗0 u i = 0, ξ ≡ λ − λ0 .
2
Let us illustrate the determination of the free parameters u i and si for the Euler–
Bernoulli problem, which we have studied in Example 5.39.
Example 5.43 Consider the governing equations (5.6) of the Euler–Bernoulli prob-
lem. Here, we have √
λ0 = π 2 and ϕ0 (x) = 2 cos(πx).
If we define
T (y, λ) ≡ y + λ sin(y),
the associated bifurcation coefficients are

1 1
a1 = ϕ∗0 ϕ0 = ϕ0 (x) d x = 2
2
cos(πx)2 d x = 1,
0 0
a2 = ϕ∗0 · 0 · ϕ20 = 0,
v2 = 0,
1 1
3
a3 = ϕ∗0 (−λ0 )ϕ30 = −λ0 ϕ0 (x)4 d x = −4λ0 cos(πx)4 d x = − π 2 ,
0 0 2
6a1 4
C2 = − = 2 > 0.
a3 π
Before continuing, compare the ansatz (5.124) of Nekrassov with the representation
(5.87), which we have obtained in Sect. 5.4.2.
Substituting the ansatz (5.124) into the equation (5.98) yields
%∞ %∞
2 2
ϕ0 ξ + (si ϕ0 + u i )ξ i + (λ0 + ξ 2 ) ϕ0 ξ + (si ϕ0 + u i )ξ i
π i=2
π i=2
∞
3
1 2 %
− ϕ0 ξ + (si ϕ0 + u i )ξ i + O(ξ 5 ) = 0. (5.125)
6 π i=2
Now, we look for the terms, which have the same power of ξ. We obtain:
• Terms in ξ
2
√ (ϕ0 + λ0 ϕ0 ) = 0.
π
=0
• Terms in ξ 2
u 2 + λ0 u 2 + s2 (ϕ0 + λ0 ϕ0 ) = 0.

=0
The BVP
u 2 + λ0 u 2 = 0, u 2 (0) = u 2 (1) = 0, u 2 ∈ Y1 , (5.126)
has the unique solution u 2 (x) ≡ 0.

• Terms in ξ 3
2 4 3
u 3 + λ0 u 3 + ϕ0 − ϕ = 0.
π 3π 0
The BVP
2 4 3
u 3 + λ0 u 3 = − ϕ0 + ϕ ≡ w3 ,
π 3π 0 (5.127)
u 3 (0) = u 3 (1) = 0, u 3 ∈ Y1 ,
has a unique solution if ϕ∗0 w3 = 0. We obtain

1 1
2 4 3 4 16
ϕ∗0 w3 =ϕ∗0 − ϕ0 + ϕ =− cos(πx) d x + 2
cos(πx)4 d x
π 3π 0 π 0 3π 0
41 16 3
=− + = 0.
π 2 3π 8
The unique solution of problem the (5.127) is
√
2
u 3 (x) = − cos(3πx).
12π 3
• Terms in ξ 4
2λ0
u 4 + λu 4 + s4 (ϕ0 + λ0 ϕ0 ) − 2 s2 ϕ30 + s2 ϕ0 = 0.

π
=0
The problem
u 4 + λu 4 = s2 2ϕ30 − ϕ0 ≡ w4 ,
(5.128)
u 4 (0) = u 4 (1) = 0, u 4 ∈ Y1 ,
has a unique solution if ϕ∗0 w4 = 0. We obtain

1
ϕ∗0 w4 = s2 ϕ∗0 2ϕ30 − ϕ0 = s2 ϕ0 2ϕ30 − ϕ0 d x
0
1
= s2 8 cos(πx)4 − 2 cos(πx)2 d x = 2s2 .
0
Thus, s2 = 0 and u 4 (x) ≡ 0.

• Terms in ξ 5
4 3 4
u 5 + λ0 u 5 − ϕ0 − 2s3 ϕ30 − 2u 3 ϕ20 + s3 ϕ0 + u 3 + ϕ5 = 0.
3π 3 15π 3 0
The problem
4 3 4
u 5 + λ0 u 5 = ϕ0 + 2s3 ϕ30 + 2u 3 ϕ20 − s3 ϕ0 − u 3 − ϕ5 ≡ w 5 ,
3π 3 15π 3 0
u 5 (0) = u 5 (1) = 0, u 5 ∈ Y1 , (5.129)
has a unique solution if ϕ∗0 w5 = 0. Since the next computations by hand are very
time-consuming, we have used the symbolic tool of the Matlab. The equation
ϕ∗0 w5 = 0 determines the unknown coefficient s3 as
5
s3 = − .
8π 3
The unique solution of problem (5.129) is

√ √
3 2 2
u 5 (x) = 5
cos(3πx) + cos(5πx).
64π 160π 5
• Terms in ξ 6
Following the strategy used before, we obtain s4 = 0 and u 6 (x) ≡ 0.
• Terms in ξ 7
The solvability condition of the resulting BVP for u 7 yields
123
s5 = .
256π 5
Now, the curve of nontrivial solutions (5.124) branching at the symmetric primary
simple bifurcation point z 0 = (0, λ0 ) can be represented in the neighborhood of z 0
in the form
√ √
1 √ 1 5 2 2
y(x; ξ) = 2 2 cos(πx)ξ + 3 − cos(πx) − cos(3πx) ξ 3
π π 8 12
√ √ √
1 123 2 3 2 2
+ 5 cos(πx) + cos(3πx) + cos(5πx) ξ 5 + O(ξ 7 ),
π 256 64 160
(5.130)
where ξ 2 ≡ λ − λ0 . Note, the function y(x) is now determined as a sum of broken

powers of λ.
In Fig. 5.17 the exact solution of problem (5.6) and the approximate solution,
which has been determined by the series (5.130), are presented.
The above example shows that the analytical approach is practicable only for
relatively simple problems. Therefore, we present now a numerical approach that
can be applied to more complex problems. At first we expand the operator T into a
Taylor series at z 0 = (0, λ0 ):
1 0 2 1 0
T (y, λ) = Ty0 y + Tyλ
0
(λ − λ0 )y + Tyy y + Tyλλ (λ − λ0 )2 y
2 2
1 0 1 1 0
+ Tyyλ (λ − λ0 )y 2 + Tyyy 0
y 3 + Tyλλλ (λ − λ0 )3 y (5.131)
2 6 6
1 0 1 0
+ Tyyλλ (λ − λ0 )2 y 2 + Tyyyλ (λ − λ0 )y 3
4 6
1 0
+ Tyyyy y 4 + R(y, λ).
24
The remainder R contains only terms of fifth and higher order in y and (λ − λ0 ).
Fig. 5.17 Exact solution and approximate solution of the Euler-Bernoulli rod problem
Assume as before that C2 satisfies the relation
6a1
C2 ≡ − > 0.
a3
Furthermore, let v3 ∈ Y1 be the (unique) solution of the linear problem

1 0 3
Ty0 v3 = − 0
C2 Tyλ ϕ0 + C2 C2 Tyy
0
ϕ0 v2 + Tyyy ϕ0 , (5.132)
6
where v2 is defined by (5.123). For the curve of nontrivial solutions of (5.11) branch-
ing at z 0 ≡ (0, λ0 ) ∈ Z we use the following ansatz (have a look at formula (5.87))
λ(ξ) = λ0 + ξ 2 ,

y(ξ) = C2 ϕ0 ξ + (C2 v2 + τ (ξ)ϕ0 )ξ 2
(5.133)
+ v3 + 2 C2 τ (ξ)v2 ξ 3 + u(ξ) + τ (ξ)2 v2 ξ 4 ,
τ (ξ) ∈R, u(ξ) ∈ Y1 .
Substituting (5.131) and (5.133) into (5.11), we get

1
Ty0 0
Tyλ ϕ0 + 3C2 Tyy
0
ϕ0 v2 + C2 Tyyy
0
ϕ30 u(ξ) w(u, τ ; ξ)
2 =− ,
ϕ∗0 0 τ (ξ) 0
(5.134)
where
w(u, τ ; ξ) ≡ Tyλ
0
w1 (u, τ ; ξ)

+ Tyy 0
C2 ϕ0 (v3 + [u + τ 2 v2 ])ξ + τ ϕ0 (w1 (u, τ ; ξ) − C2 v2 )
1
+ w1 (u, τ ; ξ)2
2
1 0 1 0
+ Tyλλ w2 (u, τ ; ξ)ξ + Tyyλ w2 (u, τ ; ξ)2
2 2
1 0
+ Tyyy 3C2 ϕ20 w1 (u, τ ; ξ) + 3 C2 ϕ0 (w1 (u, τ ; ξ) + τ ϕ0 )2 ξ
6
+ (w1 (u, τ ; ξ) + τ ϕ0 )3 ξ 2
1 0 1 0
+ Tyλλλ w2 (u, τ ; ξ)ξ 3 + Tyyλλ w2 (u, τ ; ξ)2 ξ 2
6 4
1 0 1 0
+ Tyyyλ w2 (u, τ ; ξ)3 ξ + Tyyyy w2 (u, τ ; ξ)4
6 24
1
+ 4 R(y(ξ), λ(ξ)),
ξ
# $ !
w1 (u, τ ; ξ) ≡ C2 v2 + v3 + 2 C2 τ v2 ξ + u + τ 2 v2 ξ 2 ,

w2 (u, τ ; ξ) ≡ C2 ϕ0 + [τ ϕ0 + w1 (u, τ ; ξ)]ξ.
A statement about the solvability of the transformed problem (5.134) is given in

the next theorem.
Theorem 5.44 Let z 0 ≡ (0, λ0 ) ∈ Z be a primary simple bifurcation point. Assume

that a2 = 0 and a3 = 0. Then, there exists a real number ξ0 > 0 such that for |ξ| ≤ ξ0
problem (5.134) has a continuous curve of isolated solutions η(ξ) ≡ (u(ξ), τ (ξ))T .
By inserting these solutions into the ansatz (5.133) it is possible to construct a
continuous curve of nontrivial solutions of the operator equation (5.11). This curve
branches off from the trivial solution curve {(0, λ), λ ∈ R} at the simple bifurcation
point z 0 .
Proof For ξ → 0, problem (5.134) turns into

u(0) w0
A =− , (5.135)
τ (0) 0
with
1
Ty0 0
Tyλ ϕ0 + 3C2 Tyy
0
ϕ0 v2 + C2 Tyyy
0
ϕ30
A≡ 2
ϕ∗0 0
and
1 1
w0 ≡ C2 Tyλ
0
v2 + C2 Tyy0
ϕ0 v3 + C22 Tyy v2 + C2 Tyyλ
0 2 0
ϕ20
2 2
1 1
+ C22 Tyyy
0
ϕ20 v2 + C22 Tyyyy
0
ϕ40 .
2 24
We set
1
A ≡ Ty0 , B ≡ Tyλ
0
ϕ0 + 3C2 Tyy
0
ϕ0 v2 + C2 Tyyy
0
ϕ30 ,
2
C ∗ ≡ ϕ∗0 , D ≡ 0, m = 1,
have to verify (c0 )–(c3 ):
(c0 ): dim R(B) = 1 = m,
(c1 ): R(B) ∩ R(A) = {0},
1 a1
because ψ0∗ B = a1 + C2 a3 = a1 − 3 a3 = −2a1 = 0, and
2 a3
ψ0∗ Ty0 · = 0,
(c2 ): dim R(ϕ∗0 ) = 1,
(c3 ): N (Ty0 ) ∩ N (ϕ∗0 ) = {0}.
The assumptions of Lemma 5.25 are satisfied and we can conclude that the linear
operator A is bijective. Thus, there exists a unique solution η 0 ≡ (u(0), τ (0))T of
(5.135). Writing (5.134) in the form
F(ξ; η) = 0, η(ξ) ≡ (u(ξ), τ (ξ))T ,
the claim follows as in the proof of Theorem 5.26 from the Implicit Function Theorem
(see Theorem 5.8).
Next, we want to deal with the case of an asymmetric bifurcation point, i.e., let
us now assume that
a2 ≡ ψ0∗ Tyy
0 2
ϕ0 = 0.
Since the analytical approach of Nekrassov is too expensive, we will focus only on
numerical techniques. The first step is to expand the operator T into a Taylor series
at z 0 ≡ (0, λ0 ):
1 0 2 1 0
T (y, λ) = Ty0 y + Tyλ
0
(λ − λ0 )y + Tyy y + Tyλλ (λ − λ0 )2 y
2 2 (5.136)
1 0 1 0 3
+ v Tyyλ (λ − λ0 )y + Tyyy y + R(y, λ).
2
2 6
The remainder R consists only of fourth or higher order terms in y and (λ − λ0 ).
Let C1 ∈ R be defined as
2a1
C1 ≡ − . (5.137)
a2
Furthermore, let v2 ∈ Y1 be the (unique) solution of the following linear problem

C1 0 2
Ty0 v2 = −C1 0
Tyλ ϕ0 + T ϕ , ϕ∗0 v2 = 0. (5.138)
2 yy 0
For the curve of nontrivial solutions of (5.11) branching at z 0 ≡ (0, λ0 ) ∈ Z we use

the following ansatz (have a look at formula (5.83))
λ(ξ) = λ0 + ξ,

y(ξ) = C1 ϕ0 ξ + v2 + τ (ξ)ϕ0 ξ 2 + u(ξ)ξ 3 , (5.139)
τ (ξ) ∈ R, u(ξ) ∈ Y1 .
Substituting (5.136) and (5.139) into (5.11), we get

Ty0 0
Tyλ ϕ0 + C1 Tyy
0 2
ϕ0 u(ξ) w(u, τ ; ξ)
=− , (5.140)
ϕ∗0 0 τ (ξ) 0
where
w(u, τ ; ξ) = Tyλ
0
{v2 + uξ}
1 0 2
+ Tyy {2C1 ϕ0 (v2 + uξ) + v2 + τ ϕ0 + uξ ξ}
2
1 0 1 0
+ Tyλλ w1 (u, τ ; ξ) + Tyyλ w1 (u, τ ; ξ)2
2 2
1 0 1
+ Tyyy w1 (u, τ ; ξ)3 + 3 R(y(ξ), λ0 + ξ),
6 ξ
w1 (u, τ ; ξ) = C1 ϕ0 + (v2 + τ ϕ0 )ξ + uξ 2 .
A statement about the solvability of the transformed problem (5.140) is given in the
next theorem.
Theorem 5.45 Let z 0 ≡ (0, λ0 ) ∈ Z be a primary simple bifurcation point. Assume
that a2 = 0. Then, there exists a real number ξ0 > 0 such that for |ξ| ≤ ξ0 the
problem (5.140) has a continuous curve of isolated solutions η(ξ) ≡ (u(ξ), τ (ξ))T .
By inserting these solutions into the ansatz (5.139) it is possible to construct a
continuous curve of nontrivial solutions of the operator equation (5.11). This curve
branches off from the trivial solution curve {(0, λ), λ ∈ R} at the simple bifurcation
point z 0 .
Proof For ξ → 0, the problem (5.140) turns into

⎛ ⎞
C1 0 C12 0 2 C13 0 3
− ⎝Tyλ v2 + ϕ0 v2 + T ϕ0 + T ϕ + T ϕ
0 0
6 yyy 0 ⎠ ,
A η(0) = C1 Tyy
2 yλλ 2 yyλ 0
0
(5.141)
with
Ty0 0
Tyλ ϕ0 + C1 Tyy
0 2
ϕ0
A≡ .
ϕ∗0 0
We set
A ≡ Ty0 , B ≡ Tyλ
0
ϕ0 + C1 Tyy
0 2
ϕ0 ,
C ∗ ≡ ϕ∗0 , D ≡ 0, m = 1,
have to verify (c0 )–(c3 ):
(c0 ): dim R(B) = 1 = m,
(c1 ): R(B) ∩ R(A) = {0},
2a1
because ψ0∗ B = a1 + C1 a2 = a1 − a2 = −a1 = 0, and
a2
ψ0∗ Ty0 · = 0,
(c2 ): dim R(ϕ∗0 ) = 1,
(c3 ): N (Ty0 ) ∩ N (ϕ∗0 ) = {0}.
Thus, the linear operator A is bijective, and there exists a unique solution η 0 ≡
(u 0 , τ 0 )T of the equation (5.141). The claim follows from the Implicit Function
Theorem (see the proofs of the previous theorems).
To complete the description of some transformation techniques for primary simple

bifurcation points, we want to briefly explain how these methods can be applied to
BVPs of the form (5.10). As already described in the previous sections (see, e.g., the
formulas (5.47), (5.48)), the functional condition ϕ∗0 u = 0, ϕ∗0 ∈ Y ∗ , is written in
form of a first-order ODE with two boundary conditions. Moreover, the trivial ODE
τ = 0 is added to the transformed problem. New in the transformation techniques for
symmetric and asymmetric bifurcation points is the use of the bifurcation coefficients
a1 , a2 and a3 . Since we have assumed that Ty0 is a Fredholm operator with index zero,
the adjoint operator (Ty0 )∗ has an one-dimensional null space, too. Now, the strategy
is to determine from the adjoint BVP (see (5.56))
ψ (x) = −( f y0 )T ψ(x),
Ba∗ ψ(a) + Bb∗ ψ(b) = 0,
a solution ψ 0 (x) with ψ 0 (x) = 1. Using this solution, the bifurcation coefficients
can be written in the form
b
a1 = − ψ 0 (x)T f yλ
0
ϕ0 (x)d x,
a
b
a2 = − ψ 0 (x)T f yy
0
ϕ0 (x)2 d x,
a
b
a3 = − ψ 0 (x)T 0
f yyy ϕ0 (x)3 + 6 f yy
0
ϕ0 (x)v 2 (x) d x.
a
When it is difficult to evaluate analytically these integrals, the bifurcation coefficients

can be computed easily by the numerical integration of the following IVPs:
α1 (x) = −ψ 0 (x)T f yλ

0
ϕ0 (x), α1 (a) = 0,
a1 = α1 (b);
α2 (x) = −ψ 0 (x)T f yy
0
ϕ0 (x)2 , α2 (a) = 0,
a2 = α2 (b);

α3 (x) = −ψ 0 (x)T 0
f yyy ϕ0 (x)3 + 6 f yy
0
ϕ0 (x)v 2 (x) , α3 (a) = 0,
a3 = α3 (b).
Remark 5.46 Let us have a look on the above integrands and the right-hand sides
of the corresponding ODEs. It is well known that there are only two multiplicative
operations for vectors v, w ∈ Rn , namely: (i) the inner product a = v T w, and
(ii) the outer product A = vw T , where a is a scalar and A is a matrix. Therefore, the
derivatives of the vector function f have to be understood as multilinear operators,
which are successively applied on a series of vectors. For example, some authors
0
write f yy [v, w] instead of f yy
0
vw.
In the next example, we will show how these expressions can be determined.
Example 5.47 Given the vectors y = (y1 , y2 )T , ϕ0 = (ϕ0,1 , ϕ0,2 )T and the follow-
ing vector function 2
y1 + sin(y2 ) + 5
f ( y) = .
cos(y1 ) + y1 y2 + y1
The aim is to evaluate the expression (vector) f yy ϕ0 at y0 = (0, π)T .

0 2
The first step is to determine the Jacobian of f , i.e.,

2y1 cos(y2 )
f y ( y) = .
− sin(y1 ) + y2 + 1 y1
Then, using an arbitrary vector v ∈ R2 , we form the vector

2y1 cos(y2 ) v1
f y ( y)v =
− sin(y1 ) + y2 + 1 y1 v2

2y1 v1 + cos(y2 )v2
= .
− sin(y1 )v1 + y2 v1 + v1 + y1 v2
The second step is to determine the Jacobian of the vector, i.e.,

2v1 − sin(y2 )v2
f yy ( y)v = .
− cos(y1 )v1 + v2 v1
Now, using an arbitrary vector w ∈ R2 we form the vector

2v1 − sin(y2 )v2 w1
f yy ( y)vw =
− cos(y1 )v1 + v2 v1 w2

2v1 w1 − sin(y2 )v2 w2
= .
− cos(y1 )v1 w1 + v2 w1 + v1 w2
The third step is to insert the vector y0 . We obtain

2v1 w1
0
f yy vw = .
−v1 w1 + v2 w1 + v1 w2
And finally, in the last step we set v ≡ ϕ0 and w ≡ ϕ0 . The result is

2ϕ20,1
0 2
f yy ϕ0 = .
−ϕ20,1 + 2ϕ0,1 ϕ0,2
In the next example, we show how the transformation technique can be applied
to the BVP (5.104).
Example 5.48 Let us consider once more the BVP (5.104). We should remember
that the primary simple bifurcation points are
z 0(k) = (0, λ(k) (k)

0 ), λ0 = −(kπ) , k = 1, 2, . . . ,
2
and the corresponding (normalized) eigenfunctions are

√
ϕ(k)
0 (x) = 2 sin(kπx), k = 1, 2, . . . .
Here, we consider only the case k = 1. If we define
T (y, λ) ≡ y − λ(y + y 2 ), (5.142)
it follows:
Ty = (·) − λ(1 + 2y), Ty0 = (·) − λ0
Tyλ = −(1 + 2y), 0
Tyλ = −1
Tyy = −2λ, 0
Tyy = −2λ0 ,
Tyλλ = 0, 0
Tyλλ = 0,
Tyyλ = −2, 0
Tyyλ = −2,
Tyyy = 0, 0
Tyyy = 0.
The corresponding bifurcation coefficients are

1 1
1
a1 = − ϕ0 (x) d x = −2
2
sin(πx)2 d x = −2 = −1,
0 0 2
1 √
√ 4 16 2π
a2 = 2π 2 ϕ0 (x)3 d x = 2π 2 2 2 = = 0.
0 2π 3
Thus, z 0(1) = (0, −π 2 ) is an asymmetric bifurcation point. For the coefficient C1 , we

obtain
2a1 3
C1 = − = √ = 0.084404654639729. (5.143)
a2 8π 2
Now, we have to determine the solution v2 of the linear problem (5.138). The first
equation reads
v2 + π 2 v2 = −C1 −ϕ0 + C1 π 2 ϕ20
(5.144)
v2 (0) = 0, v2 (1) = 0.
If we write the condition ϕ∗0 v2 = 0 as before in the ODE formulation
ξ1 = ϕ0 v2 , ξ(0) = ξ(1) = 0, (5.145)
and add this problem to the second-order BVP (5.144), which can be written as two
first-order ODEs, we obtain a BVP consisting of three first-order ODEs and four
boundary conditions. The difficulty is that the number of the ODEs does not match
the number of the boundary conditions. But we can apply a simple trick. We do not
insert the value (5.143) into the right-hand side of (5.144), but consider C1 as a free
parameter, write C1 = 0 and add this ODE to the BVP. Now, we have to solve the
following BVP of dimension n = 4:
y1 = y2 ,

y2 = −π 2 y1 − y4 −ϕ0 + y4 π 2 ϕ20 ,
y3 = ϕ0 y1 , (5.146)
y4 = 0,
y1 (0) = 0, y1 (1) = 0, y3 (0) = 0, y3 (1) = 0,
where we have set
y1 = v2 , y2 = v2 , y3 = ξ1 , y4 = C1 . (5.147)
At this point, a further difficulty arises. The BVP (5.146) has two solutions. The
first one is the trivial solution yi (x) ≡ 0, i = 1, . . . , 4. But we are interested in the
3
nontrivial solution, which satisfies y4 = C1 = √ .
8π 2
Our numerical experiments have led to the following result. If we choose positive
starting values for y4 and solve (5.146) with the multiple shooting code presented in
Sect. 4.6, then the nontrivial solution v2 being sought can be determined with just a
few iteration steps.
Note: the BVP (5.146) is implemented as case 6 in our multiple shooting code.
Now, let us formulate the transformed problem (5.140) for our given BVP (5.104).
We obtain !
u + π 2 u + −ϕ0 + 2C1 π 2 ϕ20 τ

= − (v2 + uξ)

+ π 2 2C1 ϕ0 (v2 + uξ) + (v2 + τ ϕ0 + uξ)2 ξ
(5.148)

2 2 1
− C1 ϕ0 + (v2 + τ ϕ0 )ξ + uξ + 3R ,
ξ

ξ2 = ϕ0 u, τ = 0,
u(0) = 0, u(1) = 0, ξ2 (0) = ξ2 (1) = 0,
where !2
R ≡ C1 ϕ0 ξ + (v2 + τ ϕ0 )ξ 2 + uξ 3 ξ.
Setting
y5 = u, y6 = u , y7 = ξ2 , y8 = τ ,
and using the notation (5.147), problem (5.148) can be written in the form
y5 = y6 ,
!
y6 = −π 2 y5 − −ϕ0 + 2y4 π 2 ϕ20 y8

+ − (y1 + y5 ξ)

+ π 2 2y4 ϕ0 (y1 + y5 ξ) + (y1 + y8 ϕ0 + y5 ξ)2 ξ
2 (5.149)
− y4 ϕ0 + (y1 + y8 ϕ0 )ξ + y1 ξ 2

1
3 2
+ 2 y4 ϕ0 ξ + (y1 + y8 ϕ0 )ξ + y5 ξ
2
,
ξ
y7 = ϕ0 y5 , y8 = 0,
y5 (0) = 0, y5 (1) = 0, y7 (0) = y7 (1) = 0,
To be able to use an IVP-solver with integrated step-size control in the multiple

shooting method, the problems (5.146) and (5.149) must be defined on the same grid.
Therefore, it is appropriate to merge the two four-dimensional BVPs into a eight-
dimensional BVP. Looking at our notation (indexing of the solution components), it
can be seen that we have already taken account of this fact.
If the exact eigenfunction ϕ0 is not explicitly known, then it is also possible to
add the extended problem (5.94) to this large BVP. However, the crucial problem is
to find good starting values for all functions yi (x).
In Fig. 5.18, we present the numerically determined solutions for the control para-
meter λ = −π 2 + ξ, with ξ = −2, −1, −0.5., 0.5, 1, 2. These solutions have been
Fig. 5.18 Solid line exact solution curve, dashed line asymptotic expansion (5.114) of the branching
solution curve, stars numerical approximations. Here, we have set l ∗ y ≡ y (0)
computed with the combined BVP (5.146), (5.149) by our multiple shooting code
given in Sect. 4.6. In addition, we have added the analytical approximations obtained
by the Levi-Cività technique for ε ∈ [−0.29, 0.48].
5.5 Analytical and Numerical Treatment of Secondary

Simple Bifurcation Points
5.5.1 Analysis of Secondary Simple Bifurcation Points
In this section, we assume that neither of the two solution curves intersecting at a
simple bifurcation point is explicitly known. Thus, the condition T (0, λ) = 0 for all
λ ∈ R, is void.
Let z 0 ≡ (y0 , λ0 ) ∈ Z be a hyperbolic point as defined in the next definition.

Definition 5.49 The point z 0 ∈ Z is called a hyperbolic point of the operator T
when the following conditions are fulfilled:
• T (z 0 ) = 0, (solution)
• dim N (T (z 0 )) = 2, codimR(T (z 0 )) = 1, (singular point)
• τ ≡ αγ − β 2 < 0, (hyperbolic point)
where
α ≡ ψ0∗ T (z 0 ) pp, β ≡ ψ0∗ T (z 0 ) pq,
γ ≡ ψ0∗ T (z 0 )qq, with
p, q ∈ Y : span{ p, q} = N (T (z 0 )),
ψ0∗ ∈ W ∗ : N (ψ0∗ ) = R(T (z 0 )).

For hyperbolic points, we have the following result.
Theorem 5.50 Suppose that z 0 ≡ (y0 , λ0 ) ∈ Z is a hyperbolic point of the opera-
tor T . Then z 0 is a simple bifurcation point, i.e., there exists a neighborhood U of z 0
in Z such that the solution manifold of T |U consists of exactly two smooth curves
z i : [−εi , εi ] → Z , z i (0) = z 0 (i = 1, 2),
which intersect transversally at z 0 .

The proof of the theorem is based on the following lemma.
Lemma 5.51 (Theorem of Crandall and Rabinowitz)
Let Z , Y be Banach spaces, Ω an open subset of Z , and T : Ω → Y a C2 -mapping.
Furthermore, w : [−1, 1] → Ω is a continuously differentiable curve in Ω, with
T (w(ε)) = 0, |ε| ≤ 1. Suppose that the following conditions are satisfied:
5.5 Analytical and Numerical Treatment of Secondary Simple … 233
• w (0) = 0,
• dim N (T (w(0))) = 2, codimR(T (w(0))) = 1,
• N (T (w(0))) = span{w (0), v},
• T (w(0))w (0)v ∈
/ R(T (w(0))).
Then, w(0) is a bifurcation point of T (w) = 0 with respect to
C ≡ {w(ε) : ε ∈ [−1, 1]}.
In the neighborhood of w(0), the solution manifold of T (w) = 0 consists of exactly

two continuous curves, which intersect only in w(0).
Proof See the paper of Crandall and Rabinowitz [28].
We are now in a position to prove Theorem 5.50.

Proof If α = γ = 0, the following transformation must not be performed.
Without loss of generality, let us assume that α = 0. Otherwise, only p and q have
to be exchanged. Now, we transform ( p, q) into ( p , q ) according to the formula
√ √
−β + −τ −β − −τ
p = p + q, q = p + q.
α α
It holds
α ≡ ψ0∗ T (z 0 ) p p
√ 2 √
−β + −τ −β + −τ ∗
= ψ0∗ T (z 0 ) pp + 2 ψ0 T (z 0 ) pq
α α
+ ψ0∗ T (z 0 )qq
√ √
β 2 − 2β −τ − τ −β + −τ
= α + 2 β+γ
α√2 √α
β 2 − 2β −τ − τ −β + −τ τ + β2
= +2 β+ = 0,
α α α
β ≡ ψ0∗ T (z 0 ) p q
√ √
−β + −τ −β − −τ
= ψ0∗ T (z 0 ) pp
α α
√ √
−β + −τ −β − −τ
+ + ψ0∗ T (z 0 ) pq + ψ0∗ T (z 0 )qq
α α
√ √ √ √
β 2 + β −τ − β −τ + τ −β + −τ − β − −τ
= α+ β+γ
α2 α
β 2 + τ − 2β 2 + τ + β 2 2τ
= = = 0,
α α
γ ≡ ψ0∗ T (z 0 )q q
√ 2 √
−β − −τ −β − −τ
= ψ0∗ T (z 0 ) pp + 2 ψ0∗ T (z 0 ) pq
α α
+ ψ0∗ T (z 0 )qq
√ √
β 2 + 2β −τ − τ −2β − 2 −τ
= α+ β+γ
α√2 α√
β 2 + 2β −τ − τ − 2β 2 − 2β −τ + τ + β 2
= = 0,
α
and 2
2 2τ
τ ≡ α γ − β = − = 0.
α
From now on, the quantities marked with an apostrophe are denoted again by α,
β, γ, τ , p and q. However, we assume that α = γ = 0.
Let Z̄ be the complement of N (T (z 0 )) in Z . In the neighborhood of the secondary

bifurcation point z 0 ∈ Z , the solutions of (5.11) can be represented in dependence
of an additional parameter ε, |ε| ≤ ε0 , as
z(ε) = z 0 + ε{a(ε) p + b(ε)q} + ε2 v(ε),

(5.150)
a(ε)2 + b(ε)2 − 1 = 0, v(ε) ∈ Z̄ .
At z 0 , we expand T into the Taylor series
1
T (z) = T (z 0 )(z − z 0 ) + T (z 0 )(z − z 0 )2 + R(z), (5.151)
2
where R(z(ε)) = O(ε3 ). Substituting (5.150) and (5.151) into the operator equation
(5.11), we obtain
1 1
T (z 0 )v(ε) + T (z 0 ){a(ε) p + b(ε)q + εv(ε)}2 + 2 R(z(ε)) = 0,
2 ε (5.152)
a(ε)2 + b(ε)2 − 1 = 0, v(ε) ∈ Z̄ .
Let us write (5.152) in compact form as
F(x, ε) = 0, F : X × R → U, (5.153)
with
x = x(ε) ≡ (v(ε), a(ε), b(ε)),
X ≡ Z̄ × R × R, U ≡ W × R.
Using appropriate norms, X and U are Banach spaces.

Consider (5.153) for ε → 0. We get
1
T (z 0 )v(0) + T (z 0 ){a(0) p + b(0)q}2 = 0,
2 (5.154)
a(0)2 + b(0)2 − 1 = 0, v(0) ∈ Z̄ .
The first equation is solvable if ψ0∗ T (z 0 ){a(0) p + b(0)q}2 = 0. Thus,
0 = a(0)2 ψ0∗ T (z 0 ) pp + 2a(0)b(0)ψ0∗ T (z 0 ) pq + b(0)2 ψ0∗ T (z 0 )qq

= a(0)2 α + 2a(0)b(0)β + b(0)2 γ.
The assumption α = γ = 0 implies
a(0)b(0) = 0, a(0)2 + b(0)2 − 1 = 0.
Obviously, these two nonlinear algebraic equations for the two unknown a(0) and
b(0) have four solutions:
ξ1 = (a1 , b1 )T = (1, 0)T ,

ξ2 = (a2 , b2 )T = (0, 1)T ,
ξ3 = −ξ1 ,
ξ4 = −ξ2 .
In the following we consider only the pair of linearly independent solutions (ξ1 , ξ2 ).
Note, for given ai (0) = ai and bi (0) = bi there exists a unique solution vi (0). Let
us apply the Implicit Function Theorem (see Theorem 5.8) to the problem (5.153) at
x (i) ≡ (vi (0), ai (0), bi (0)), i = 1, 2, ε = 0.
We have to show that the derivative Fx (x (i) , 0) ∈ L(X, U ) is a linear homeomorphism

between X and U . For this we consider the equation
0 = Fx (x (i) , 0) h, h ≡ (ṽ, ã, b̃) ∈ X,
or, in more detail
0 =T (z 0 )ṽ + ãai T (z 0 ) pp + ãbi T (z 0 ) pq + b̃ai T (z 0 ) pq

+ b̃bi T (z 0 )qq, (5.155)
0 =2(ãai + b̃bi ), ṽ ∈ Z̄ .
There exists a solution of problem (5.155), if the following conditions are fulfilled:
ãbi β + b̃ai β = 0, ãai + b̃bi = 0.
We substitute the solutions (ai , bi ), i = 1, 2, into the equations (5.155), and obtain
ã = b̃ = 0. Thus, (5.155) is reduced to
T (z 0 )ṽ = 0, ṽ ∈ Z̄ .
The unique solution of this equation is ṽ = 0, i.e., the operator Fx (x (i) , 0) is injective.
Now, we show the subjectivity of Fx (x (i) , 0) for the case i = 1. For g ∈ W ,
g1 ∈ R and ξ1 = (1, 0)T , we consider the inhomogeneous problem
T (z 0 )ṽ + ãT (z 0 ) pp + b̃T (z 0 ) pq = g, ṽ ∈ Z̄ ,

(5.156)
2ã = g1 .
1
The second equation yields ã = g1 . We insert this value into the first equation and
2
obtain
1
T (z 0 )ṽ = g − g1 T (z 0 ) pp − b̃T (z 0 ) pq ≡ ĝ, ṽ ∈ Z̄ .
2
There exists a solution of that equation if ĝ ∈ R(T (z 0 )), i.e.,
1 1
ψ0∗ g − g1 ψ0∗ T (z 0 ) pp − b̃ψ0∗ T (z 0 ) pq = ψ0∗ g − g1 α − b̃β = 0.
2 2
ψ0∗ g
Since we have assumed that α = 0 and β = 0, it follows b̃ = . Now it can be
β
concluded, that the solution ṽ is also uniquely determined.
Similar arguments are used to prove the case i = 2.
Using the Open Mapping Theorem (see, e.g., [132]), we can conclude from the
bijectivity of Fx (x (i) , 0) ∈ L(X, U ) that this operator is a linear homeomorphism
from X onto U . Then, for |ε| ≤ εi (i = 1, 2), the Implicit Function Theorem implies
the existence of two continuous curves of isolated solutions (vi (ε), ai (ε), bi (ε)) of
(5.152). Inserting these solutions into the ansatz (5.150), we obtain two continuous
solution curves of (5.11), which intersect at z 0 . To show that in the neighborhood of
z 0 the solution manifold of (5.11) consists of exactly these two curves, let us consider
one of the two curves, for instance
z(ε) = z 0 + ε{a1 (ε) p + b1 (ε)q} + ε2 v1 (ε), v1 ∈ Z̄ .
It holds "
dz(ε) ""
= a1 (0) p + b1 (0)q = p.
dε "ε=0
Moreover, we obtain
dz(0)
β = ψ0∗ T (z 0 ) q = 0.
dε

Now, the claim follows from the Theorem of Crandall and Rabinowitz
(see Lemma 5.51).
5.5.2 Extension Techniques for Secondary Simple

Bifurcation Points
Let z 0 ≡ (y0 , λ0 ) ∈ Z be a hyperbolic point of the operator T . Assume that at least

one of the two solution curves, which intersect at z 0 , enables a smooth parametrization
with respect to λ, i.e.,
Tλ (z 0 ) ∈ R(Ty (z 0 )). (5.157)
As stated in the previous section, we have the following situation:
N (T (z 0 )) = span{ p, q}, p, q ∈ Z ,
with p = ( p1 , p2 )T , q = (q1 , q2 )T , p1 , q1 ∈ Y and p2 , q2 ∈ R.

Suppose that ϕ∗0 ∈ Y ∗ is a suitably chosen functional. Except for the sign, the
elements p, q can be uniquely written as
p = (ϕ0 , 0)T , q = (s0 , 1)T , ϕ0 , s0 ∈ Y,
where ϕ0 and s0 satisfy ϕ∗0 ϕ0 = 1, ϕ∗0 s0 = 0. Furthermore, we assume that an

element 0 = ψ0 ∈ W is known, with
ψ0∗ ψ0 = 0. (5.158)
An appropriate extension technique for the numerical determination of secondary

simple bifurcation points is based on the following operator (see [131])
⎧
⎪
⎪ Z̃ ≡ Y × R × Y × Y × R → W̃ ≡ W ⎛× W × W × R × R⎞
⎪
⎪
⎪
⎪ T (y, λ) + μψ0
⎨ ⎜ ⎟
⎜ Ty (y, λ) ϕ ⎟ .
T̃ : (5.159)
⎪
⎪ z̃ ≡ (z, ϕ, s, μ) = (y, λ, ϕ, s, μ) → ⎜Ty (y, λ)s + Tλ (y, λ)⎟
⎜
⎟
⎪
⎪ ⎝ ⎠
⎪
⎪ ϕ∗0 ϕ − 1
⎩
ϕ∗0 s
Now we consider the extended operator equation
T̃ (z̃) = 0. (5.160)
Obviously, z̃ 0 ≡ (y0 , λ0 , ϕ0 , s0 , 0) ∈ Z̃ is a solution of (5.160). More detailed

information about the extended operator is given in the next theorem.
Theorem 5.52 Let z 0 ≡ (y0 , λ0 ) ∈ Z be a hyperbolic point of T . Assume that the

conditions (5.157) and (5.158) are satisfied. Then, T̃ (z̃ 0 ) is a linear homeomorphism
from Z̃ onto W̃ , i.e.,
N (T̃ (z̃ 0 )) = {0}, R(T̃ (z̃ 0 )) = W̃ .
Proof The derivative T̃ (z̃ 0 ) has the form

⎛ ⎞
Ty0 Tλ0 0 0 ψ0
⎜ Tyy 0
ϕ0 0
Tyλ ϕ0 Ty0 0 0⎟
⎜ ⎟
T̃ (z̃ 0 ) = ⎜
0
⎜ yy 0 Tyλ
T s + 0 0
Tyλ s0 + Tλλ
0
0 Ty0 0⎟ ⎟. (5.161)
⎝ 0 0 ϕ∗0 0 0⎠
0 0 0 ϕ∗0 0
(i) Assume that ( ỹ, λ̃, ϕ̃, s̃, μ̃) ∈ N (T̃ (z̃ 0 )), where ỹ, ϕ̃, s̃ ∈ Y and λ̃, μ̃ ∈ R.
Then,
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
ỹ Ty0 ỹ + Tλ0 λ̃ + μ̃ψ0 0
⎜ λ̃ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ 0
Tyy ϕ0 ỹ + Tyλ0
ϕ0 λ̃ + Ty0 ϕ̃ ⎟ ⎜0⎟
T̃ (z̃ 0 ) ⎜ ⎟ ⎜T 0 s ỹ + T 0 s λ̃ + T 0 s̃ + T 0 ỹ + T 0 λ̃⎟ ⎟=⎜ ⎟
⎜ϕ̃⎟ = ⎜ λλ ⎟ ⎜0⎟ . (5.162)
⎝ s̃ ⎠ ⎜
yy 0 yλ 0 y yλ
⎝ ∗
ϕ0 ϕ̃ ⎠ ⎝ 0⎠
μ̃ ∗
ϕ s̃ 0
0
There exists a solution of the first equation
Ty0 ỹ = −Tλ0 λ̃ − μ̃ψ0 ,
if
ψ0∗ Tλ0 λ̃ + μ̃ψ0∗ ψ0 = 0.
The assumptions (5.157) and (5.158) imply μ̃ = 0. The general solution of the
resulting equation
Ty0 ỹ = −Tλ0 λ̃
is ( ỹ, λ̃) = (δϕ0 + κs0 , κ). Now, this solution is substituted into the second and third
equation of (5.162). We obtain
Ty0 ϕ̃ = −(δTyy
0 2
ϕ0 + κTyy
0
ϕ0 S0 + κTyλ
0
ϕ0 )
(5.163)
Ty0 s̃ = −(δTyy
0
s0 ϕ0 + κTyy
0 2
s0 + κTyλ
0
s0 + δTyλ
0
ϕ0 + κTyλ
0
s0 + κTλλ
0
).
For p = (ϕ0 , 0) and q = (s0 , 1), the quantities α, β and γ (see Definition 5.49) take
the form
α = ψ0∗ Tyy ϕ0 , β = ψ0∗ (Tyy
0 2 0
ϕ0 S0 + Tyλ
0
ϕ0 ),
γ = ψ0∗ (Tyy
0 2
s0 + 2Tyλ
0
s0 + Tλλ
0
).
The conditions for the solvability of (5.163) are
0 = δψ0∗ Tyy
0 2
ϕ0 + κψ0∗ (Tyy
0
ϕ0 s0 + Tyλ
0
ϕ0 ),
0 = δψ0∗ (Tyy
0
s0 ϕ0 + Tyλ
0
ϕ0 ) + κψ0∗ (Tyy
0 2
s0 + 2Tyλ
0
s0 + Tλλ
0
),
or
α β δ 0
= .
β γ κ 0
The assumption τ ≡ αγ − β 2 < 0 implies δ = 0 and κ = 0. Thus ( ỹ, λ̃) = (0, 0).
Now, the first equation in (5.163) is reduced to
Ty0 ϕ̃ = 0, i.e., ϕ̃ = cϕ0 , c ∈ R.
The fourth equation in (5.162) implies c = 0, i.e., ϕ̃ = 0. The second equation in

(5.163) is reduced to
Ty0 s̃ = 0, i.e., s̃ = dϕ0 , d ∈ R.
The fifth equation in (5.162) implies d = 0, i.e., s̃ = 0.

Summarizing the above steps, we have the following result
N (T̃ (z̃ 0 )) = {0}.
(ii) Now, let us consider the inhomogeneous equations
Ty0 ỹ + Tλ0 λ̃ + μ̃ψ0 = c1 ∈ W,

0
Tyy ϕ0 ỹ + Tyλ
0
ϕ0 λ̃ + Ty0 ϕ̃ = c2 ∈ W,
0
Tyy s0 ỹ + Tyλ
0
s0 λ̃ + Ty0 s̃ + Tyλ
0
ỹ + Tλλ
0
λ̃ = c3 ∈ W, (5.164)
ϕ∗0 ϕ̃ = μ1 ∈ R,
ϕ∗0 s̃ = μ2 ∈ R,
where (c1 , c2 , c3 , μ1 , μ2 ) ∈ R(T̃ (z̃ 0 )).

From the condition for the solvability of the first equation, we obtain
ψ0∗ c1
μ̃ = .
ψ0∗ ψ0
Now, we can write

( ỹ, λ̃) = (y1 + δϕ0 + κs0 , κ),
y1 ∈ Y1 ≡ {y ∈ Y : ϕ∗0 y = 0},
where δ and κ are real numbers that are to be determined. If we substitute ỹ and λ̃
into the second and third equation in (5.164), then the conditions for the solvability
of the resulting equations can be written in the form

α β δ ψ0∗ (c2 − Tyy0
ϕ0 y1 )
= . (5.165)
β γ κ ψ0∗ (c3 − Tyy0
s0 y1 − Tyλ0
y1 )
The assumption τ < 0 guarantees that there exists a unique solution (δ0 , κ0 ) of this
system. This in turn gives the result that
( ỹ, λ̃) = (y1 + δ0 ϕ0 + κ0 s0 , κ0 )
is uniquely determined. This allows us to write the second and third equation in
(5.164) in the form
Ty0 ϕ̃ = c̄2 ∈ R(Ty0 ),
Ty0 s̃ = c̄3 ∈ R(Ty0 ).
It follows
ϕ̃ = ϕ1 + d1 ϕ0 , ϕ1 ∈ Y1 , d1 ∈ R,
s̃ = s1 + d2 ϕ0 , s1 ∈ Y1 , d2 ∈ R.
Using the fourth and fifth equation in (5.164), we obtain
ϕ∗0 ϕ̃ = ϕ∗0 ϕ1 +d1 ϕ∗0 ϕ0 = μ1 , i.e., d1 = μ1 ,

=0 =1
ϕ∗0 s̃ = ϕ∗0 s1 +d2 ϕ∗0 ϕ0 = μ2 , i.e., d2 = μ2 .

=0 =1
In summary, we have the result that ϕ̃ and s̃ are uniquely determined and the claim
of our theorem is proved.
Remark 5.53 The above proof shows that the statement of Theorem 5.52 remains
valid, if the condition τ < 0 is replaced by τ = 0.
With respect to the reversal of the statement of Theorem 5.52, we have the following
result.
Theorem 5.54 Suppose z̃ 0 ≡ (y0 , λ0 , ϕ0 , s0 , 0) is an isolated solution of problem

(5.160), i.e., T̃ (z̃ 0 ) is a linear homeomorphism of Z̃ onto W̃ . Then, z 0 ≡ (y0 , λ0 )
satisfies:
• T (z 0 ) = 0,
• dim N (T (z 0 )) = 2, codim R(T (z 0 )) = 1,
• τ ≡ αγ − β 2 = 0, and
• Tλ (z 0 ) ∈ R(Ty (z 0 )).
Proof (i) Assume that τ = 0 and (c1 , c2 )T is a nontrivial solution of the problem

α β c1 0
= .
β γ c2 0
Moreover, let w1 , w2 ∈ Y1 be the unique solutions of the equations

0 !
Ty0 w1 = − c1 Tyy
0 2
ϕ0 + c2 Tyy ϕ0 s0 + Tyλ
0
ϕ0 ,
0 0 2 !
Ty0 w2 = − c1 Tyy s0 ϕ0 + Tyλ
0
ϕ0 + c2 Tyy s0 + 2Tyλ
0
s0 + Tλλ
0
.
Then,
T̃ (z̃ 0 )(c1 ϕ0 + c2 s0 , c2 , w1 , w2 , 0)T

⎛ ⎞ ⎛ ⎞
c2 Ty0 s0 + c2 Tλ0 0
⎜ c T ϕ
0 2
+ c T 0
ϕ s + c T 0
ϕ + T 0
w ⎟ ⎜ ⎟
⎜ 1 yy 0 2 yy 0 0 2 yλ 0 y 1 ⎟ ⎜0⎟
⎜
= ⎜c1 Tyy s0 ϕ0 + c2 Tyy s0 + c1 Tyλ ϕ0 + 2c2 Tyλ s0 + c2 Tλλ + Ty w2 ⎟ = ⎜0⎟
0 0 2 0 0 0 0 ⎟ ⎜
⎟
⎝ ϕ∗0 w1 ⎠ ⎝0⎠
ϕ∗0 w2 0
However, our assumption is N (T̃ (z̃ 0 )) = {0}. Thus we have a contradiction and it
must hold τ = 0.
(ii) Next we assume that there exists a w0 ∈ Y , w0 = 0, such that
Ty0 w0 = 0, ϕ∗0 w0 = 0.
But this implies (0, 0, w0 , 0, 0) ∈ N (T̃ (z̃ 0 )), which is a contradiction to our assump-
tion. Thus, it must hold w0 = 0.
(iii) The remaining conditions are trivially fulfilled.

Corollary 5.55 If the secondary simple bifurcation points z (i) = (y (i) , λ(i) ) of the
original operator equation (5.11) have to be computed, then those isolated solutions
z̃ (i) of the extended problem (5.160) must be determined, which are of the form
z̃ (i) = (y (i) , λ(i) , ϕ(i) , s (i) , 0).
The implementation of the extended problem (5.160) for BVPs of the form (5.10),
which is based on the functional (5.47), is
y (x) = f (x, y; λ) − μψ 0 (x), Ba y(a) + Bb y(b) = 0,

ϕ (x) = f y (x, y; λ)ϕ(x), Ba ϕ(a) + Bb ϕ(b) = 0,

s (x) = f y (x, y; λ)s(x) + f λ (x, y; λ), Ba s(a) + Bb s(b) = 0,
ξ1 (x) = ϕ(x) ϕ(x),
T
ξ1 (a) = 0, ξ1 (b) = 1, (5.166)
ξ2 (x) = ϕ(x) s(x),
T
ξ2 (a) = 0, ξ2 (b) = 0,

λ (x) = 0,
μ (x) = 0.
This again is a nonlinear BVP in standard form of dimension 3n + 4 that can be

solved with the known numerical techniques, e.g., with the multiple shooting code
presented in Sect. 4.6. It makes sense to use starting vectors, whose last component
μ is zero.
Remark 5.56 If the BVP is self-adjoint, the first ODE in (5.166) can be replaced by
the equation
y (x) = f (x, y; λ) − μϕ0 (x).
The implementation (5.166) of the extension technique (5.159) is based on the

functional (5.47). The number of equations in the BVP (5.166) can be further reduced,
if we use instead of (5.47) the functional (5.53), i.e.,
ϕ∗0 v(x) ≡ ϕ0 (a)T v(a),
whereas before ϕ0 (x) is the nontrivial solution of the second (block-) equation in
(5.166), i.e., the corresponding eigenfunction. In this way, we obtain the following
extended BVP of dimension 3n + 2, which is constructed in a self-explanatory
manner:
y (x) = f (x, y; λ) − μψ 0 (x), Ba y(a) + Bb y(b) = 0,

ϕ (x) = f y (x, y; λ)ϕ(x), Ba ϕ(a) + Bb ϕ(b) = 0,
s (x) = f y (x, y; λ)s(x) + f λ (x, y; λ), Ba s(a) + Bb s(b) = 0, (5.167)

λ (x) = 0, ϕ(a) ϕ(a) = 1,
T

μ (x) = 0, ϕ(a)T s(a) = 0.
An other extension technique for the determination of secondary simple bifurca-

tion points has been presented in [88]. Here, the second (block-) equation in (5.160)
is replaced by the corresponding adjoint problem. The extended operator T̂ of T is
⎧
∗ ∗
⎪ Ẑ ≡ Y × R × Y × R → Ŵ ≡ W⎛× Y × R × R ⎞
⎪
⎪
⎪
⎨ T (y, λ) + μψ0
T̂ : ⎜ Ty∗ (y, λ) ψ ∗ ⎟ , (5.168)
⎪
⎪ ẑ ≡ (z, ψ , μ) = (y, λ, ψ , μ) → ⎜
∗ ∗
⎝ ψ ∗ Tλ (y, λ) ⎠
⎟
⎪
⎪
⎩
ψ ∗ ψ0 − 1
and the corresponding extended problem is
T̂ (ẑ) = 0. (5.169)
Obviously, if z 0 ≡ (y0 , λ0 ) ∈ Z is a secondary simple bifurcation point of the

original operator equation (5.11), then ẑ 0 ≡ (y0 , λ0 , ψ0∗ , 0) ∈ Ẑ is a solution of the
extended operator equation (5.169). Moreover, we have the following result.
Theorem 5.57 Let the hypotheses of Theorem 5.52 be satisfied. Then, the operator
T̂ (ẑ 0 ) is a linear homeomorphism of Ẑ onto Ŵ , i.e.,
N (T̂ (ẑ 0 )) = {0}, R(T̂ (ẑ 0 )) = Ŵ .
If the element ϕ0 is not required for a subsequent curve tracing, the extended prob-
lem (5.169) is to be favorized compared to the problem (5.160) since it has fewer
equations.
The appropriate implementation of the extended problem (5.169) for BVPs of the
form (5.10), which is based on the functional (5.47), is the following extended BVP
of dimension 2n + 4:
y (x) = f (x, y; λ) − μψ(x), Ba y(a) + Bb y(b) = 0,

ψ (x) = − f y (x, y; λ) ψ(x),
T
Ba∗ ψ(a) + Bb∗ ψ(b) = 0,
ξ1 (x) = −ψ(x)T f λ (x, y; λ), ξ1 (a) = 0, ξ1 (b) = 0,
(5.170)
ξ2 (x)= ψ(x) ψ(x),
T
ξ2 (a) = 0, ξ2 (b) = 1,

λ (x) = 0,
μ (x) = 0.
The matrices Ba∗ and Bb∗ are defined by (see, e.g., [63]):
!
Ba∗ BaT − Bb∗ BbT = 0, rank Ba∗ |Bb∗ = n.
If we use the functional (5.53) instead of the functional (5.47), the following BVP
of the dimension 2n + 2 results:
y (x) = f (x, y; λ) − μψ(x), Ba y(a) + Bb y(b) = 0,

ψ (x) = − f y (x, y; λ) ψ(x),
T
Ba∗ ψ(a) + Bb∗ ψ(b) = 0,
(5.171)
λ (x) = 0, ψ(a)T ψ(a) = 1,
μ (x) = 0, ψ(a)T f λ (a, y(a); λ) = 0.

Secondary Simple Bifurcation Point
Let us assume that z ≡ (y, λ) ∈ Z is an isolated solution of the operator equation

(5.11), which is located in the direct neighborhood of a secondary simple bifurcation
point z 0 ≡ (y0 .λ0 ) ∈ Z . The following transformation technique for the numerical
approximation of z is based on the statements and results presented in the proof of
Theorem 5.50. The starting point is the formula (5.150), i.e.,
y(ε) = y0 + ε{ρ(ε) p1 + σ(ε)q1 } + ε2 u(ε),

λ(ε) = λ0 + ε{ρ(ε) p2 + σ(ε)q2 } + ε2 τ (ε), (5.172)
ρ(ε) + σ(ε) = 1, (u(ε), τ (ε)) ∈ Z̄ .
2 2
The representation (5.172) of the solution curve is not based on the special form of
p = ( p1 , p2 )T ∈ Y × R and q = (q1 , q2 )T ∈ Y × R, which we have introduced
in Sect. 5.5.2. Therefore, any known bases of the null space N (T (z 0 )) can be used.
However, if the bifurcation point z 0 has been determined by the extension technique
(5.159), (5.160), then it makes sense to set
p1 = ϕ0 , p2 = 0, q1 = s0 , q2 = 1.
With v(ε) ≡ (u(ε), τ (ε))T and z(ε) according to formula (5.172), we have
T (z 0 )v(ε) = Ty0 u(ε) + Tλ0 τ (ε)
and
T (z 0 )[ρ(ε) p + σ(ε)q + εv(ε)]2
0 0
2
Tyy Tyλ ρ(ε) p1 + σ(ε)q1 + εu(ε)
= ,
Tyλ0
Tλλ0 ρ(ε) p2 + σ(ε)q2 + ετ (ε)
= Tyy
0
[ρ(ε) p1 + σ(ε)q1 + εu(ε)]2
+ 2Tyλ
0
[ρ(ε) p1 + σ(ε)q1 + εu(ε)][ρ(ε) p2 + σ(ε)q2 + ετ (ε)]
+ Tλλ
0
[ρ(ε) p2 + σ(ε)q2 + ετ (ε)]2 .
Let p ∗ = ( p1∗ , p2∗ )T ∈ Z ∗ and q ∗ = (q1∗ , q2∗ )T ∈ Z ∗ be two linear functionals, with
p ∗ p = q ∗ q = 0, p ∗ q = q ∗ p = 0.
Now, the transformed problem (5.152) can be formulated more precisely as

⎛ ⎞ ⎛ ⎞
Ty0 Tλ0 w(u(ε), τ (ε), ρ(ε), σ(ε); ε)
⎝ p1∗ u(ε)
p2∗ ⎠ = −⎝ 0 ⎠,
τ (ε) (5.173)
q1∗ q2∗ 0
ρ(ε)2 + σ(ε)2 = 1,
with
1 0
w(u, τ , ρ, σ; ε) ≡ T (ρ p1 + σq1 + εu)2
2 yy
+ Tyλ
0
(ρ p1 + σq1 + εu)(ρ p2 + σq2 + ετ )
1
+ Tλλ
0
(ρ p2 + σq2 + ετ )2 + 2 R(y(ε), λ(ε)).
ε
A statement about the solvability of the transformed problem (5.173) is given in the
next theorem.
Theorem 5.58 Let z 0 ≡ (y0 , λ0 ) ∈ Z be a hyperbolic point of the operator T . Then,

there exist real numbers ε1 , ε2 > 0 such that for |ε| ≤ εi (i = 1, 2) problem (5.173)
has two continuous curves of isolated solutions
T
η i (ε) ≡ u i (ε), τ i (ε), ρi (ε), σ i (ε) , i = 1, 2.
By inserting η 1 (ε) and η 2 (ε) into the ansatz (5.172), it is possible to construct two
continuous curves of solutions of the operator equation (5.11). These curves intersect
at z 0 .
Proof The statement is an immediate conclusion from Theorem 5.50.
At the end of this section, we will show how this transformation technique can be
applied to BVPs of the form (5.10). If the functionals p ∗ , q ∗ ∈ Z ∗ are given in the
form b
p ∗ r = p1∗ r 1 + p2∗r2 ≡ p1 (x)T r 1 (x)d x + p2 r2 ,
a
b
q ∗ r = q1∗ r 1 + q2∗ r2 ≡ q 1 (x)T r 1 (x)d x + q2 r2 ,
a
r = (r 1 , r2 )T ∈ Z ,
then the second and third (block-) equation of the transformed problem (5.173) may
be written in ODE formulation as follows
ξ1 (x; ε) = p1 (x)T u(x; ε) + p2 τ (ε), ξ1 (a; ε) = ξ1 (b; ε) = 0 (5.174)
and
ξ2 (x; ε) = q 1 (x)T u(x; ε) + q2 τ (ε), ξ2 (a; ε) = ξ2 (b; ε) = 0, (5.175)
respectively. Moreover, we add the trivial ODEs τ (x; ε) = 0, ρ (x; ε) = 0 and

σ (x; ε) = 0 to the resulting system. In this way, we obtain the following transformed
BVP of dimension n + 5
u (x; ε) = f 0y u(x; ε) + f 0λ τ (x; ε)

+ w̃(u(x; ε), τ (x; ε), ρ(x; ε), σ(x; ε); ε),
ξ1 (x; ε) = p1 (x)T u(x; ε) + p2 τ (ε),
ξ2 (x; ε)
= q 1 (x)T u(x; ε) + q2 τ (ε) (5.176)
τ (x; ε) = 0, ρ (x; ε) = 0, σ (x; ε) = 0,

Ba u(a; ε) + Bb u(b; ε) = 0,
ξ1 (a; ε) = ξ2 (a; ε) = ξ1 (b; ε) = ξ2 (b; ε) = 0,
ρ(a; ε)2 + σ(x; ε)2 − 1 = 0, |ε| ≤ ε0
where
1 0
w(u, τ , ρ, σ; ε) ≡ f (ρ p1 + σq 1 + εu)2
2 yy
+ f 0yλ (ρ p1 + σq 1 + εu)(ρ p2 + σq2 + ετ )
1 1
+ f 0λλ (ρ p2 + σq2 + ετ )2 + 2 R̃( y(x; ε), λ(x; ε)),
2 ε
and

1
R̃( y, λ) ≡ f ( y, λ) − f 0 + f 0y ( y − y0 ) + f 0λ (λ − λ0 ) + f 0yy ( y − y0 )2
2

1
+ f 0yλ (λ − λ0 )( y − y0 ) + f 0λλ (λ − λ0 )2 .
2
5.6 Perturbed Bifurcation Problems
5.6.1 Nondegenerate Initial Imperfections
All problems discussed so far are of the general form
T : Y × Λ → W, T (y, λ) = 0,
(5.177)
Y, W, Λ - Banach spaces,
where we have set Λ ≡ R in (5.11). In this book, we will not consider the case
Λ = Rk , i.e., multiparameter bifurcation problems.
As we have seen, a typical bifurcation diagram for the equation (5.11) contains
bifurcation and limit points. However, in experiments and in real applications, the
sharp transitions of bifurcation rarely occur. Small imperfections, impurities, or other
inhomogeneities tend to distort these transitions. In Fig. 5.19 an example for the effect
of small initial imperfections on the first bifurcation point of the governing equations
(5.6) of the Euler–Bernoulli rod is shown. We have to take such initial imperfections
into consideration, since the rod is not a perfectly straight line or the cross-section
is not always the same.
To analyze mathematically the perturbation of bifurcations caused by imperfec-
tions and other impurities, the classical bifurcation theory is modified by introducing
an additional small perturbation parameter, which characterizes the magnitudes of
these inhomogeneities. Thus, instead of the one-parameter problem (5.177) the fol-
lowing two-parameter problem is studied
T̄ : Y × Λ1 × Λ2 → W, T̄ (y, λ, τ ) = 0,
(5.178)
Y, W, Λ1 , Λ2 - Banach spaces,
where Λ1 is the control parameter space and Λ2 is the perturbation parameter space.
Here, we set Λ1 = Λ2 = R. The connection between problem (5.177) and (5.178)
is made by embedding (5.177) into (5.178) such that
T̄ (y, λ, 0) = T (y, λ). (5.179)

Fig. 5.19 Perturbed primary simple bifurcation
Suppose that the BVP (5.10) enlarged by the perturbation parameter τ is of the
form
y (x) = f (x, y; λ, τ ), a < x < b,
(5.180)
Ba y(a) + Bb y(b) = 0,
with
f : D f → Rn , D f ⊂ [a, b] × Rn × R × R, f ∈ C p , p ≥ 2,
λ ∈ Dλ ⊂ R, τ ∈ Dτ ⊂ R, Ba , Bb ∈ Rn×n , rank[Ba |Bb ] = n.
Here, λ is the control parameter and τ is the perturbation parameter reflecting the
small initial imperfections in the mathematical model. The corresponding operator
form is
T (y, λ, τ ) = 0, y ∈ Y, λ, τ ∈ R, (5.181)
where
T : Z ≡ Y × R × R → W, T (y, λ, τ ) ≡ y − f (·, y; λ, τ ),
and
Y ≡ BC1 ([a, b], Rn ), W ≡ C([a, b], Rn ).
5.6 Perturbed Bifurcation Problems 249
Assume that the operator T satisfies two additional conditions:
T (0, λ, 0) = 0 for all λ ∈ R,

(5.182)
T (0, λ, τ ) = 0 for all τ = 0.
We start with the description of the unperturbed problem (i.e., the case τ = 0). The
first condition of (5.182) implies that the solution field of the unperturbed problem
E(T ; τ = 0) contains the trivial solution curve. Let λ0 be a simple eigenvalue of
Ty (z), z ≡ (0, λ, 0) ∈ Y × R × R, i.e., it holds
dim N (Ty (z 0 )) = 1, z 0 ≡ (0, λ0 , 0),

(5.183)
N (Ty (z 0 )) = span{ϕ0 }, ϕ0 = 1.
Then, conditions (5.183) say that Ty (z 0 ) is a Fredholm operator with index zero and
the statements (5.17)–(5.21) of the Riesz–Schauder theory are valid. Besides (5.183),
we suppose that
a1 ≡ ψ0∗ Tyλ (z 0 )ϕ0 = 0. (5.184)
As explained in Sect. 5.4.2, the assumptions (5.183) and (5.184) guarantee that in
the neighborhood of z 0 ≡ (0, λ0 , 0) the solution field of the unperturbed problem
E(T ; τ = 0) consist of exactly two solution curves, which intersect at z 0 . One of
these curves is the trivial solution curve {(0, λ, 0) : λ ∈ R}.
We now want to study the effects of nonvanishing initial imperfections (i.e., τ = 0)
on the operator equation (5.181). The following classification of perturbed problems
takes into account the fact that the corresponding solution field E(T ; τ = 0) can be
qualitatively changed by small initial imperfections.
Definition 5.59
• The operator equation (5.181) is called a BP-problem (bifurcation preserving prob-
lem), if in any neighborhood U ⊂ Y × R of (0, λ0 ), a τ (U ) > 0 can be found
such that for each fixed τ , 0 < |τ | < τ (U ), there exists a bifurcation point in U .
• Otherwise, the operator equation (5.181) is called a BD-problem (bifurcation
destroying problem).
The problem (5.181) is always a BD-problem if there exists a neighborhood V ⊂
Y × R × R of (0, λ0 , 0) for which the following statement is true:
T (y, λ, τ ) = 0, (y, λ, τ ) ∈ V , and τ = 0 imply that (y, λ) is not a bifurcation point.
For the next considerations, a special type of perturbations τ = 0 is of interest.
Definition 5.60 Let the conditions (5.183) and (5.184) for the unperturbed problem
be satisfied. We define
b1 ≡ ψ0∗ Tτ (z 0 ). (5.185)
The parameter τ is said to represent a nondegenerate (degenerate) imperfection if

b1 = 0 (b1 = 0).
Nondegenerate imperfections play an important role in the bifurcation theory since

they have a great influence on the solution field of the unperturbed problem.
Theorem 5.61 Assume that the conditions (5.182)–(5.184) and b1 = 0 are satisfied.
Then, the problem (5.181) is a BD-problem.
Proof See [114].
Example 5.62 Let us consider once more the BVP (5.6) governing the Euler–
Bernoulli rod problem. The only difference now is that we introduce an additional
external parameter τ , describing the initial imperfections of the rod, into the right-
hand side of the ODE. More precisely, we have the following perturbed BVP
y (x) + λ sin(y(x)) = τ cos(πx), x ∈ (0, 1),

(5.186)
y (0) = 0, y (1) = 0.
As we have seen in Sect. 5.1, the primary simple bifurcation points and the cor-
responding eigenfunctions of the unperturbed problem are λ(k) = (kπ)2 and
(k) √ 0
ϕ0 = 2 cos(kπx), respectively. Here, we want to concentrate on the first bifur-
cation point λ0 ≡ λ(1)0 = π and answer the question, whether a bifurcation point
2
also occurs in the perturbed BVP (5.186). In order to do that, we have to determine
the coefficient b1 defined in (5.185). If we set
T (y, λ, τ ) ≡ y + λ sin(y) − τ cos(πx),
it follows Tτ (z 0 ) = − cos(πx) and

√
√ 1
2
b1 = − 2 cos(πx) d x = −
2
= 0.
0 2
Looking at the assumptions of Theorem 5.61, we see that the BVP (5.186) is a BD-
problem. The difference in the structure of the solution field of the unperturbed and
the perturbed problem can be seen in Fig. 5.20.
Let us now change the boundary conditions of the BVP (5.186) as follows
y (x) + λ sin(y(x)) = τ cos(πx), x ∈ (0, 1),

(5.187)
y(0) = 0, y(1) = 0.
Now, for the unperturbed problem we have the following situation. The first primary
√ λ0 is the same as before, but the corresponding eigenfunction
simple bifurcation point
changes to ϕ0 (x) = 2 sin(πx). Therefore, the coefficient b1 is calculated as
√ 1
b1 = − 2 cos(πx) sin(πx)d x = 0.
0
Fig. 5.20 Bifurcation diagram of problem (5.186) for 3 different values of τ
Fig. 5.21 Bifurcation diagram of problem (5.187) for 3 different values of τ
Obviously, Theorem 5.61 provides not enough information to decide whether the
BVP (5.187) is a BD- or a BP-problem. But in fact it is a BP-problem as can be seen
in Fig. 5.21.
Our numerical treatment of (5.181), under the condition (5.185), is based on the
theory of Keener and Keller [68]. The following steps have to be executed:
1. Determination of a curve of nonisolated solutions of (5.181), which contains the
particular point (y, λ, τ ) = (0, λ0 , 0);
2. Determination of curves of nontrivial isolated solutions of (5.181), which branch
off from the curve computed in step 1;
3. Presentation of the solution field E(T ) of (5.181) in form of a bifurcation dia-
gram using the solution curves constructed in step 2 and the curve of nonisolated
solutions determined in step 1.
5.6.2 Nonisolated Solutions
The nonisolation of solutions of problem (5.181) is explained in the following defi-

nition.
Definition 5.63 The quadruple (y, ϕ, λ, τ ) ∈ Y × Y × R × R is referred to as
nonisolated solution of (5.181) if
T (y, λ, τ ) = 0,
(5.188)
Ty (y, λ, τ )ϕ = 0, ϕ = 0.

From the previous section, we know that z 0 ≡ (0, λ0 ) is a bifurcation point of the
unperturbed problem. Therefore, for this particular point, the corresponding noniso-
lated solution is z̃ 0 ≡ (0, ϕ0 , λ0 , 0). Now, we want to study the effect of nondegener-
ate imperfections τ on the bifurcation point z 0 . Since the nonisolation of solutions is
a necessary condition for the occurrence of bifurcations, it makes sense to determine
first a curve of nonisolated solutions of problem (5.181), which is passing through
the point z̃ 0 . Using an additional real parameter ε ∈ R, this curve is represented as
follows:
y(ε) = εϕ0 + ε2 u(ε), u(ε) ∈ Y1 ,
ϕ(ε) = ϕ0 + εχ(ε), χ(ε) ∈ Y1 ,
(5.189)
λ(ε) = λ0 + εα(ε), α(ε) ∈ R,
τ (ε) = ε2 β(ε), β(ε) ∈ R.
The Taylor series of T at z̃ 0 is
1
T (y, λ, τ ) = Ty0 y + Tτ0 τ + Tyλ
0
y(λ − λ0 ) + Tyτ0
yτ + Tyy y 2
2 (5.190)
+ Tτ0λ τ (λ − λ0 ) + R1 (y, λ, τ ).
The corresponding Taylor series of Ty is
Ty (y, λ, τ ) = Ty0 + Tyλ

0
(λ − λ0 ) + Tyy
0
y + R2 (y, λ, τ ). (5.191)
For the ansatz (5.189), we have
R1 = O(ε3 ) and R2 = O(ε2 ).
Inserting the ansatz (5.189) into (5.188), we obtain
0 = Ty0 u(ε) + Tτ0 β(ε) + Tyλ

0
α(ε)ϕ0 + εTyλ
0
α(ε)u(ε)
1 0
+ εTyτ 0
β(ε)[ϕ0 + εu(ε)] + Tyy [ϕ0 + εu(ε)]2 (5.192)
2
1
+ εTτ0λ β(ε)α(ε) + 2 R1 , ϕ∗0 u(ε) = 0,
ε
and
0 = Ty0 χ(ε) + Tyλ

0
α(ε)[ϕ0 + εχ(ε)] + Tyy
0
[ϕ0 + εu(ε)][ϕ0 + εχ(ε)]
1 (5.193)
+ R2 [ϕ0 + εχ(ε)], ϕ∗0 χ(ε) = 0.
ε
Equations (5.192) and (5.193) can be written in matrix notation as
⎛ ⎞⎛ ⎞ ⎛ ⎞
Ty0 0 0
Tyλ ϕ0 Tτ0 u(ε) w1 (u, α, β; ε)
⎜0
⎜ Ty0 0
Tyλ ϕ0 0⎟ ⎜ ⎟ ⎜ ⎟
⎟ ⎜χ(ε)⎟ = − ⎜w2 (u, χ, α, β; ε)⎟ , (5.194)
⎝ ϕ∗ 0 0 0 ⎠ ⎝ α(ε) ⎠ ⎝ 0 ⎠
0
0 ϕ∗0 0 0 β(ε) 0
where
1 0
w1 (u, α, β; ε) ≡ T (ϕ0 + εu)2 + εβTyτ 0
(ϕ0 + εu) + εαTyλ
0
u
2 yy
1
+ εαβTτ0λ + 2 R1 (y, λ, τ ),
ε
and
w2 (u, χ, α, β; ε) ≡ εαTyλ
0
χ

1
+ Tyy 0
(ϕ0 + εu) + R2 (y, λ, τ ) (ϕ0 + εχ).
ε
A statement about the solvability of problem (5.194) is given in the next theorem.
Theorem 5.64 Let the assumptions (5.182)–(5.184) and b1 = 0 be satisfied. Then,

there exists a real number ε0 > 0 such that for |ε| ≤ ε0 problem (5.194) has a
continuous curve of isolated solutions
η(ε) ≡ (u(ε), χ(ε), α(ε), β(ε))T .
By inserting η(ε) into the ansatz (5.189) it is possible to construct a continuous curve
of solutions of the equations (5.188). This curve is composed of nonisolated solutions
of (5.181) and passes through
z̃ 0 ≡ (0, ϕ0 , λ0 , 0) ∈ Y × Y × R × R.
Proof Consider (5.194) for ε → 0. We get
⎛ ⎞ ⎛ 0 ⎞⎛ ⎞ ⎛1 ⎞
u(0) Ty 0 0
Tyλ ϕ0 Tτ0 u(0) Tyy ϕ0
0 2
⎜χ(0)⎟ ⎜ 0 ⎜2 ⎟
A⎜ ⎟ ⎜ Ty0 0
Tyλ ϕ0 0⎟ ⎜ ⎟
⎟ ⎜χ(0)⎟ = − ⎜ 0 2 ⎟
ϕ0 ⎟ .
⎝α(0)⎠ ≡ ⎝ ϕ∗ ⎜ Tyy
0 0 0 0 ⎠ ⎝α(0)⎠ ⎝ 0 ⎠
β(0) 0 ϕ∗0 0 0 β(0) 0
We set 0
Ty0 0 Tyλ ϕ0 Tτ0
A≡ , B≡
0 Ty0 0
Tyλ ϕ0 0
∗
ϕ0 0 0 0
C∗ ≡ , D≡ , m ≡ 2,
0 ϕ∗0 0 0
and use Lemma 5.25. Obviously, A is not bijective and

ϕ0 0
N (A) = span , , i.e., dim N (A) = 2.
0 ϕ0
Thus, the second case applies and we have to verify (c0 )–(c3 ).
) 0 *
Tτ0 Tyλ ϕ0
(c0 ): R(B) = span , , i.e., dim R(B) = 2,
0 0
Tyλ ϕ0

v
(c1 ): It holds: 1 ∈ R(A) ⇔ ψ0∗ v1 = 0 and ψ0∗ v2 = 0.
v2
For the matrix B we have:
v1 ψ0∗ Tyλ
0
ϕ0 = 0 ⇔ v1 = 0, which implies

=0
v2 ψ0∗ Tτ0 + v1 ψ0∗ Tyλ 0
ϕ0 = 0 ⇔ v2 = 0,

=0 =0

0
i.e., R(B) ∩ R(A) = .
0
(c2 ) & (c3 ) are fulfilled trivially.
Thus, we have shown that operator A is bijective. The claim of the theorem now
follows analogously to the proofs of the transformation techniques presented in the
previous sections.
For BVPs of the form (5.180), the equation (5.194) can be implemented as follows
u (x; ε) = f 0y u(x; ε) + α(ε) f 0yλ ϕ0 (x) + β(ε) f 0τ + w̃1 (u, α, β; ε),

χ (x, ε) = f 0y χ(x; ε) + α(ε) f 0yλ ϕ0 (x) + w̃ 2 (u, χ, α, β; ε),
ξ1 (x; ε) = ϕ0 (x)T u(x; ε),
ξ2 (x; ε) = ϕ0 (x)T χ(x; ε),
α (ε) = 0, (5.195)
β (ε) = 0,
Ba u(a; ε) + Bb u(b; ε) = 0, Ba χ(a; ε) + Bb χ(b; ε) = 0,
ξ1 (a; ε) = 0, ξ1 (b; ε) = 0, ξ2 (a; ε) = 0, ξ2 (b; ε) = 0,
with
1 0
w̃ 1 (u, α, β; ε) ≡ f (ϕ + εu)2 + εβ f 0yτ (ϕ0 + εu) + εα f 0yλ u
2 yy 0
1
+ εαβ f 0τ λ + 2 R̃1 ,
ε
1
w̃ 2 (u, χ, α, β; ε) ≡ εα f 0yλ χ + f 0yy (ϕ0 + εu) + R̃2 [ϕ0 + εχ],
ε

R̃1 ≡ f (x, y, λ, τ ) − f 0y y + f 0τ τ + f 0yλ y(λ − λ0 ) + f 0yτ yτ

1 0 2
+ f yy y + f τ λ τ (λ − λ0 ) ,
0
2
R̃2 ≡ f y (x, y, λ, τ ) − { f 0y + f 0yλ (λ − λ0 ) + f 0yy y}.
Now, we want to show that it is also possible to compute the nonisolated solutions
in direct dependence on the perturbation parameter τ . As will be seen later, such
techniques are very advantageous to generate bifurcation diagrams. They enable to
compute simultaneously a nonisolated solution and a curve of nontrivial solutions of
(5.181) passing through this solution. Moreover, the use of an additional (external)
parameter is not required. The disadvantage of this method, however, is the relatively
high analytical and numerical effort caused by the computation of all required partial
derivatives of T .
The ansatz for the curve of nonisolated solutions depends on the second bifurca-
tion coefficient a2 of the unperturbed problem. First, let us consider the case of an
asymmetric bifurcation point, i.e., we assume
a2 ≡ ψ0∗ Tyy (0, λ0 , 0)ϕ20 = 0.

If the last equation in (5.189) is solved for ε, i.e., ε = g(τ ), the following ansatz
for the curve of nonisolated solutions of (5.181) passing through the particular point
z̃ 0 ≡ (0, ϕ0 , λ0 , 0) is obtained:
a1
y(ξ) = − ϕ0 ξ + {K (ξ)ϕ0 + u 1 }ξ 2 + u(ξ)ξ 3 ,
a2
λ(ξ) = λ0 + ξ + λ1 (ξ)ξ 2 , (5.196)
ϕ(ξ) = ϕ0 + ϕ1 ξ + ψ(ξ)ξ , 2
2b1 a2
ξ2 = τ , sign(τ ) = sign(b1 a2 ),
a12
where λ1 , K ∈ R and v, ψ ∈ Y1 . The elements u 1 and ϕ1 are the (unique) solutions

of the following linear problems
2
a1 0 a12 1 a1
Ty0 u 1 = Tyλ ϕ0 − Tτ0 − 0 2
Tyy ϕ0 ,
a2 2b1 a2 2 a2 (5.197)
ϕ∗0 u 1 = 0,
and
a1 0 2
Ty0 ϕ1 = −Tyλ
0
ϕ0 + T ϕ ,
a2 yy 0 (5.198)
ϕ∗0 ϕ1 = 0.
Now, we expand T and Ty into Taylor series:
1 0 2
T (y, λ, τ ) = Ty0 y + Tτ0 τ + Tyλ
0
y(λ − λ0 ) + Tyτ
0
yτ + Tyy y + Tτ0λ τ (λ − λ0 )
2
1 0 3 1 0 1 0 2
+ Tyyy y + Tyλλ y(λ − λ0 )2 + Tyyλ y (λ − λ0 )
6 2 2
+ R1 (y, λ, τ ),
(5.199)
and

0
(λ − λ0 ) + Tyτ
0
τ + Tyy
0
y + Tyyλ
0
y(λ − λ0 )
1 0 2 1 0 (5.200)
+ Tyyy y + Tyλλ (λ − λ0 ) + R2 (y, λ, τ ).
2 2
With respect to the ansatz (5.196), the remainders R1 and R2 satisfy
R1 = O(ξ 4 ) and R2 = O(ξ 3 ).

Inserting (5.196) and (5.199) into the first equation of (5.188), we obtain

a1 a2
0= ξTy0 − ϕ0 + {K ϕ0 + u 1 }ξ + uξ + ξ 2 Tτ0 1
2
a2 2b1 a2

a 1
+ ξ 2 Tyλ
0
− ϕ0 + {K ϕ0 + u 1 }ξ + uξ 2 (1 + λ1 ξ)
a2

a 1 a12
+ ξ Tyτ − ϕ0 + {K ϕ0 + u 1 }ξ + uξ
3 0 2
a2 2b1 a2
2
1 0 a1
+ ξ 2 Tyy − ϕ0 + {K ϕ0 + u 1 }ξ + uξ 2
2 a2
a2
+ ξ 3 Tτ0λ 1 (1 + λ1 ξ) (5.201)
2b1 a2
3
31 0 a1
+ ξ Tyyy − ϕ0 + {K ϕ0 + u 1 }ξ + uξ 2
6 a2

1 a1
+ ξ Tyλλ − ϕ0 + {K ϕ0 + u 1 }ξ + uξ (1 + λ1 ξ)2
3 0 2
2 a2
2
1 a 1
+ ξ 3 Tyyλ0
− ϕ0 + {K ϕ0 + u 1 }ξ + uξ 2 (1 + λ1 ξ)
2 a2
+ R1 (y(ξ), λ(ξ), τ (ξ)).
Now, we consider the terms in ξ and ξ 2 . We obtain:

• Terms in ξ
a1 0
− T ϕ0 = 0.
a2 y
• Terms in ξ 2
Due to (5.197), it holds
2
a12 a1 0 1 a1
Ty0 u 1 + Tτ0 − Tyλ ϕ0 + 0 2
Tyy ϕ0 = 0.
2b1 a2 a2 2 a2
Thus, the right-hand side of (5.201) is O(ξ 3 ).

Next, we insert (5.196) and (5.200) into the second equation in (5.188). This
yields
0 = ξTy0 (ϕ1 + ψξ) + ξTyλ

0
(1 + λ1 ξ)(ϕ0 + ϕ1 ξ + ψξ 2 )
a12
+ ξ 2 Tyτ0
(ϕ0 + ϕ1 ξ + ψξ 2 )
2b1 a2

a1
+ ξTyy 0
− ϕ0 + {K ϕ0 + u 1 }ξ + uξ 2 ϕ0 + ϕ1 ξ + ψξ 2 (5.202)
a2

a1
+ ξ 2 Tyyλ
0
− ϕ0 + {K ϕ0 + u 1 }ξ + uξ 2 (1 + λ1 ξ)(ϕ0 + ϕ1 ξ + ψξ 2 )
a2
2
21 0 a1
+ ξ Tyyy − ϕ0 + {K ϕ0 + u 1 }ξ + uξ 2
(ϕ0 + ϕ1 ξ + ψξ 2 )
2 a2
1 0
+ ξ 2 Tyλλ (1 + λ1 ξ)2 (ϕ0 + ϕ1 ξ + ψξ 2 )
2
+ R2 (y(ξ), λ(ξ), τ (ξ))(ϕ0 + ϕ1 ξ + ψξ 2 )
Let us consider the terms in ξ. Due to (5.198), we obtain:

a1 0 2
Ty0 ϕ1 + Tyλ
0
ϕ0 − T ϕ = 0.
a2 yy 0
Thus, the right-hand side of (5.202) is O(ξ 2 ). Dividing the equation (5.201) by ξ 3
and the equation (5.202) by ξ 2 , we can write the resulting equations in a compact
form as
⎛ ⎞
a1 0 a1 0 2 ⎛ ⎞ ⎛ ⎞
Ty0 0 − Tyλ ϕ0 − Tyy ϕ0 + Tyλ0
ϕ0 w1
⎜ a2 a2 ⎟ u(ξ)
⎜ ⎟ ⎜ ψ(ξ) ⎟ ⎜ w ⎟
⎜ 0 Ty0 0
ϕ0 0 2
ϕ0 ⎟⎜ ⎟ ⎜ 2⎟
⎜ ∗ Tyλ Tyy ⎟ ⎝λ1 (ξ)⎠ = ⎝ 0 ⎠ , (5.203)
⎝ ϕ0 0 0 0 ⎠
∗ K (ξ) 0
0 ϕ0 0 0
where
w1 (u, λ1 , K ; ξ) ≡
a12
0
Tyλ {u 1 + uξ + λ1 w3 (u, K ; ξ)} + T 0 w4 (u, K ; ξ)
2b1 a2 yτ

1 0 1 2a1
+ Tyy − 2 ϕ0 u 1 + u 1 − 2
ϕ0 u + 2u 1 uξ + u ξ − K ϕ0 ξ
2 2 2 2
2 2 a2

+ 2K ϕ0 w3 (u, K ; ξ)
a12 1 0
+ Tτ0λ {1 + λ1 ξ} + Tyyy w4 (u, K ; ξ)3
2b1 a2 6
1 1
+ Tyλλ {1 + λ1 ξ}2 w4 (u, K ; ξ) + 3 R1 (y, λ, τ ),
2 ξ
w2 (u, ψ, λ1 , K ; ξ) ≡
2
a1
T 0 + Tyy 0
{u 1 + uξ} + Tyyλ
0
w4 (u, K ; ξ) + λ1 Tyyλ
0
w4 (u, K ; ξ)ξ
2b1 a2 yτ

1 0 1 0 1
+ Tyyy w4 (u, K ; ξ) + Tyλλ {1 + λ1 ξ} + 2 R2 (y, λ, τ ) ×
2 2
2 2 ξ
! ! !
× ϕ0 + ϕ1 ξ + ψξ + λ1 Tyλ + K Tyy ϕ0 ϕ1 ξ + ψξ 2
2 0 0

a1 0
+ Tyλ 0
− Tyy ϕ0 [ϕ1 + ψξ],
a2
w3 (u, K ; ξ) ≡ (K ϕ0 + u 1 + uξ)ξ,
a1
w4 (u, K ; ξ) ≡ − ϕ0 + w3 (u, K ; ξ).
a2
The solvability properties of the equations (5.203) are stated in the next theorem.
Theorem 5.65 Let the assumptions (5.182)–(5.184) and a2 ≡ ψ0∗ Tyy ϕ0 = 0 be
0 2
satisfied. Then, there exists a real constant τ0 > 0 such that for |τ | ≤ τ0 and
sign(τ ) = sign(b1 a2 ) problem (5.203) has a continuous curve of isolated solutions
η(ξ) ≡ (u(ξ), ψ(ξ), λ1 (ξ), K (ξ))T .
By inserting η(ξ) into the ansatz (5.196) it is possible to construct a continuous

curve of solutions of the equations (5.188). This curve is composed of nonisolated
solutions of (5.181) and passes through
z̃ 0 ≡ (0, ϕ0 , λ0 , 0) ∈ Y × Y × R × R.
Proof Consider (5.203) for ξ → 0, i.e., for τ → 0. We get

⎛ ⎞ ⎛ ⎞
u(0) w1 (u, λ1 , K ; 0)
⎜ ψ(0) ⎟ ⎜w2 (u, ψ, λ1 , K ; 0)⎟
A⎜ ⎟ ⎜
⎝λ1 (0)⎠ = − ⎝
⎟,
⎠
0
K (0) 0
where A denotes the system matrix in (5.203). We set

⎛ a ⎞
1 0 a1 0 2
Ty0 0 − Tyλ ϕ0 − Tyy ϕ0 + Tyλ
0
ϕ0
A≡ , B ≡ ⎝ a2 a2 ⎠,
0 Ty0 0
Tyλ ϕ0 ϕ0
0 2
Tyy
∗
ϕ0 0 0 0
C∗ ≡ , D≡ , m ≡ 2,
0 ϕ∗0 0 0
and use Lemma 5.25. Obviously, A is not bijective and dim N (A) = 2, i.e., the
second case applies, and we have to verify (c0 )–(c3 ).
(c0 ): dim R(B) = 2,
(c1 ): For l1 ∈ Y and l2
∈ R we have ⎞
⎛
a1 0 a1 0 2
− Tyλ ϕ0 l1 + − Tyy ϕ0 + Tyλ
0
ϕ0 l 2
⎝ a2 a2 ⎠ ∈ R(B).
yλ ϕ 0 l 1 + Tyy ϕ0 l 2
0 0 2
T
v1
We know that ∈ R(A) ⇔ ψ0∗ v1 = 0 and ψ0∗ v2 = 0.
v2
Let us now consider the matrix B. We set

. a1 a1
0 = − ψ0∗ Tyλ 0
ϕ0 l1 + − ψ0∗ Tyy ϕ0 + ψ0∗ Tyλ
0 2 0
ϕ0 l 2 ,
a2 a2
. ∗ 0 ∗ 0 2
0 = ψ0 Tyλ ϕ0 l1 + ψ0 Tyy ϕ0 l2 .
These
⎛ 2 two equations
⎞ can be written in matrix-vector notation as
a1
⎝− a 0 ⎠ l1
=
0
.
2 l2 0
a1 a2
Since the determinant of the system matrix does
not vanish, we obtain
0
l1 = 0 and l2 = 0, i.e., R(A) ∩ R(B) = .
0
The conditions (c2 ) & (c3 ) are fulfilled trivially. Thus, we have shown that operator
A is bijective. The claim of the theorem now follows analogously to the proofs of
the transformation techniques presented in the previous sections.
Let us now consider the case of a symmetric bifurcation point of the unperturbed
problem, i.e., we assume
0
a2 ≡ ψ0∗ Tyy (0, λ0 , 0)ϕ20 = 0, a3 ≡ ψ0∗ 6Tyy ϕ0 w0 + Tyyy
0
ϕ30 = 0.
We use the following ansatz for the curve of nonisolated solutions of (5.181) passing
through the particular point z̃ 0 ≡ (0, ϕ0 , λ0 , 0):
+ , -
y(ξ) = C˜2 ϕ0 ξ + C˜2 ũ 1 + K (ξ)ϕ0 ξ 2
+

+ u 2 + 2 C̃2 K (ξ)ũ 1 ξ 3 + u(ξ) + K (ξ)2 ũ 1 ξ 4 ,
λ(ξ) = λ0 + ξ 2 + λ1 (ξ)ξ 3 , (5.204)

+
ϕ(ξ) = ϕ0 + 2 C̃2 ũ 1 + {ϕ2 + 2K (ξ)ũ 1 } ξ 2 + ψ(ξ)ξ 3 ,
3 b1
ξ3 = − τ,
2 C̃2 a1
where u, ψ ∈ Y1 and λ1 , K ∈ R. The constant C̃2 is defined as
2a1
C̃2 ≡ − , (5.205)
a3
and we suppose that

C̃2 > 0. (5.206)
The elements ũ 1 , u 2 , and ϕ2 are the (unique) solutions of the following linear prob-
lems
1 0 2
Ty0 ũ 1 = − Tyy ϕ0 , ϕ∗0 ũ 1 = 0, (5.207)
2 +
2 C̃2 a1 0
Ty0 u 2 = Tτ − C̃2 Tyλ
0
ϕ0
3 b1
+
1 0 3
− C̃2 C̃2 Tyyy ϕ0 + 6Tyy0
ϕ0 ũ 1 , ϕ∗0 u 2 = 0, (5.208)
6
1 0 3
Ty0 ϕ2 = − Tyλ 0
ϕ0 − C̃2 Tyyy ϕ0 + 6Tyy 0
ϕ0 ũ 1 , ϕ∗0 ϕ2 = 0. (5.209)
2
Now, we expand T and Ty into Taylor series:
1 0 2
T (y, λ, τ ) = Ty0 y + Tτ0 τ + Tyλ
0
y(λ − λ0 ) + Tyτ
0
yτ + Tyy y
2
1 0 3 1 0 2 1 0 (5.210)
+ Tyyy y + Tyyλ y (λ − λ0 ) + Tyyyy y4
6 2 24
+ R3 (y, λ, τ ),
and

0
(λ − λ0 ) + Tyτ
0
τ + Tyy
0
y + Tyyλ
0
y(λ − λ0 )
1 0 2 1 0 (5.211)
+ Tyyy y + Tyyyy y 3 + R4 (y, λ, τ ).
2 6
With respect to the ansatz (5.204), the remainders R3 and R4 satisfy
R3 = O(ξ 5 ) and R4 = O(ξ 4 ).
Inserting (5.204) and (5.210) into the first equation in (5.188), we obtain
+
0= ξTy0 C̃2 ϕ0 + αξ + βξ + γξ
2 3
+
2 C̃2 a1 0
− ξ3 Tτ + ξ 3 Tyλ0
C̃2 ϕ0 + αξ + βξ 2 + γξ 3 (1 + λ1 ξ)
3 b1
+
4 0 2 C̃2 a1
− ξ Tyτ C̃2 ϕ0 + αξ + βξ 2 + γξ 3
3 b1
+ 2
21 0
+ ξ Tyy C̃2 ϕ0 + αξ + βξ + γξ
2 3
(5.212)
2
+ 3
1 0
+ ξ 3 Tyyy C̃2 ϕ0 + αξ + βξ 2 + γξ 3
6
+ 2
41
+ξ 0
Tyyλ C̃2 ϕ0 + αξ + βξ + γξ
2 3
(1 + λ1 ξ)
2
+ 4
1
+ ξ4 C̃2 ϕ0 + αξ + βξ 2 + γξ 3 + R3 ,
24
where we have used the abbreviations

+
α ≡ C˜2 ũ 1 + K (ξ)ϕ0 , β ≡ u 2 + 2 C̃2 K (ξ)ũ 1 ,
γ ≡ u(ξ) + K (ξ)2 ũ 1 .
Now, we consider the terms in ξ, ξ 2 and ξ 3 . We obtain:

• Terms in ξ +
C̃2 Ty0 ϕ0 = 0.
• Terms in ξ 2

1 0 2
C̃2 Ty0 ũ 1 + Tyy ϕ0 = 0.
2
• Terms in ξ 3
Due to (5.207) and (5.208), we get
+ +
2 C̃2 a1 0
Ty u 2 + 2 C̃2 K Ty ũ 1 −
0 0
Tτ + Tyλ C̃2 ϕ0
0
3 b1
+ +
1 0 1 0 3
+ Tyy 2 C̃2 ϕ0 {C̃2 ũ 1 + K ϕ0 } + C̃2 C̃2 Tyyy ϕ0
2 6
+
1 0 2 2 C̃2 a1 0
= 2 C̃2 K Ty ũ 1 + Tyy ϕ0 + Ty u 2 −
0 0
Tτ
2 3 b1
+ +
1 0 3
+ C̃2 Tyλ0
ϕ0 + C̃2 C̃2 Tyyy ϕ0 + 6Tyy
0
ϕ0 ũ 1
6
= 0.
Thus, the right-hand side of (5.212) is O(ξ 4 ).
Next, we insert (5.204) and (5.211) into the second equation of (5.188). This
yields
+
0 = Ty0 ϕ0 + 2 C̃2 ũ 1 ξ + δξ 2 + ψξ 3
+
+ ξ 2 Tyλ0
(1 + λ1 ξ) ϕ0 + 2 C̃2 ũ 1 ξ + δξ 2 + ψξ 3
+
32 C̃2 a1 0
−ξ Tyτ ϕ0 + 2 C̃2 ũ 1 ξ + δξ 2 + ψξ 3
3 b1
+ +
+ ξTyy 0
C̃2 ϕ0 + αξ + βξ 2 + γξ 3 ϕ0 + 2 C̃2 ũ 1 ξ + δξ 2 + ψξ 3
+
+ ξ 3 (1 + λ1 ξ) C̃2 ϕ0 + αξ + βξ 2 + γξ 3 × (5.213)
+
× ϕ0 + 2 C̃2 ũ 1 ξ + δξ 2 + ψξ 3
+ 2
1 0
+ ξ 2 Tyyy C̃2 ϕ0 + αξ + βξ 2 + γξ 3 ×
2
+
× ϕ0 + 2 C̃2 ũ 1 ξ + δξ 2 + ψξ 3
+ 3
31
+ξ 0
C̃2 ϕ0 + αξ + βξ + γξ
Tyyyy 2 3
×
6
+
× ϕ0 + 2 C̃2 ũ 1 ξ + δξ + ψξ
2 3
+
+ R5 ϕ0 + 2 C̃2 ũ 1 ξ + δξ + ψξ ,
2 3
where in addition we have used the abbreviation
δ ≡ ϕ2 + 2K (ξ)ũ 1 .
As before, we consider the terms in ξ and ξ 2 . We obtain:

• Terms in ξ
+
1 0 2
2 C̃2 Ty ũ 1 + Tyy ϕ0 = 0.
0
2
• Terms in ξ 2
Due to (5.207) and (5.209), we get

Ty0 ϕ2 + 2K Ty0 ũ 1 + Tyλ
0
ϕ0 + Tyy 0
C̃2 ũ 1 ϕ0 + K ϕ20
+ + +
1 0
+ Tyy 2 C̃2 ϕ0 C̃2 ũ 1 + Tyyy C̃2 ϕ30
0
2

1
= Ty ϕ2 + 2K Ty ũ 1 + Tyy ϕ0 + Tyλ
0 0 0 2 0
ϕ0
2
1 0 3
+ C̃2 Tyyy ϕ0 + 6Tyy 0
ũ 1 ϕ0
2
= 0.
Thus, the right-hand side of (5.213) is O(ξ 3 ). Dividing the equation (5.212) by ξ 4
and the equation (5.213) by ξ 3 , we can write the resulting equations in a compact
form as ⎛ ⎞ ⎛ ⎞
v(ξ) w1 (v, λ1 , K ; ξ)
⎜ ψ(ξ) ⎟ ⎜w2 (v, ψ, λ1 , K ; ξ)⎟
A⎜ ⎟
⎝λ1 (ξ)⎠ = − ⎝
⎜ ⎟,
⎠ (5.214)
0
K (ξ) 0
where
⎛ ⎞
1 0
⎜ Ty
0
0 0
C̃2 Tyλ ϕ0 2Tyλ ϕ0 + 6C̃2 Tyy
0
ϕ0 ũ 1 + C̃2 Tyyy
0
ϕ30 ⎟
⎜ 2 ⎟
A≡⎜
⎜ 0∗ Ty0 0
Tyλ ϕ0 0
C̃2 6Tyy ϕ0 ũ 1 + Tyyy
0
ϕ30 ⎟,
⎟
⎝ ϕ0 0 0 0 ⎠
0 ϕ∗0 0 0
and

w1 (v, λ1 , K ; ξ) = Tyλ
0
w3 (v, K ; ξ) + λ1 w7 (v, K ; ξ)ξ
+
1
+ Tyy 0
w3 (v, K ; ξ)3 + C̃2 ϕ0 (u 2 + {v + K 2 ũ 1 }ξ)
2

+ K ϕ0 {w3 (v, K ; ξ) − C̃2 ũ 1 }
+
1 0
+ Tyyy 3C̃2 ϕ20 w3 (v, K ; ξ) + 3 C̃2 ϕ0 w7 (v, K ; ξ)2 ξ
6

+ w7 (v, K ; ξ)3 ξ 2

2 C̃2 a1 0
− Tyτ w4 (v, K ; ξ)
3 b1
1 0
+ Tyyλ {1 + λ1 ξ}w4 (v, K ; ξ)2
2
1 0 1
+ Tyyyy w4 (v, K ; ξ)4 + 4 R3 (y, λ, τ ),
24 ξ

2 C̃2 a1 0
w2 (v, ψ, λ1 , K ; ξ) = − Tyτ + Tyy
0
u 2 + {v + K 2 ũ 1 }ξ
3 b1
+ Tyyλ
0
{1 + λ1 ξ}w4 (v, K ; ξ)
+
1 0
+ Tyyy 2 C̃2 ϕ0 w3 (v, K ; ξ) + w7 (v, K ; ξ) ξ
2
2

1 0 1
+ Tyyyy w4 (v, K ; ξ) + 3 R4 (y, λ, τ ) w6 (ψ, K ; ξ)
3
6 ξ

+ Tyλ 0
{1 + λ1 ξ} + Tyy 0
C̃2 ũ 1
+
1 0
+ Tyyy C̃2 ϕ0 + 2 C̃2 K ϕ0 ξ w5 (ψ, K ; ξ)
2 2
2
+
+ Tyy
0
C̃2 ϕ0 {ϕ2 + ψξ}

+ K ϕ0 {ϕ2 + 2K ũ 1 + ψξ}ξ ,
+
w3 (v, K ; ξ) ≡ C̃2 ũ 1 + u 2 + 2 C̃2 K ũ 1 ξ + {v + K 2 ũ 1 }ξ 2 ,
+
w4 (v, K ; ξ) ≡ C̃2 ϕ0 + K ϕ0 ξ + w3 (v, K ; ξ)ξ,
+
w5 (ψ, K ; ξ) ≡ 2 C̃2 ũ 1 + {ϕ2 + 2K ũ 1 }ξ + ψξ 2 ,
w6 (ψ, K ; ξ) ≡ ϕ0 + w5 (ψ, K ; ξ)ξ,
w7 (v, K ; ξ) ≡ K ϕ0 + w3 (v, K ; ξ).
The solvability properties of the equations (5.214) are stated in the next theorem.
Theorem 5.66 Let the assumptions (5.182)–(5.184), a2 = 0, a3 = 0 and C̃2 > 0
be satisfied. Then, there exists a real constant τ0 > 0 such that for |τ | ≤ τ0 problem
(5.214) has a continuous curve of isolated solutions
η(ξ) ≡ (u(ξ), ψ(ξ), λ1 (ξ), K (ξ))T .
By inserting η(ξ) into the ansatz (5.204) it is possible to construct a continuous

curve of solutions of the equations (5.188). This curve is composed of nonisolated
solutions of (5.181) and passes through
z̃ 0 ≡ (0, ϕ0 , λ0 , 0) ∈ Y × Y × R × R.
Proof Consider (5.214) for ξ → 0, i.e., for τ → 0. We obtain

⎛ ⎞ ⎛ ⎞
u(0) w1 (u, λ1 , K ; 0)
⎜ ψ(0) ⎟ ⎜ ⎟
A⎜ ⎟ = − ⎜w2 (u, ψ, λ1 , K ; 0)⎟ .
⎝λ1 (0)⎠ ⎝ 0 ⎠
K (0) 0
We set

Ty0 0
A≡ ,
0 Ty0
⎛ ⎞
1 0
C̃2 Tyλ ϕ0
0
2Tyλ ϕ0 + 6C̃2 Tyy ϕ0 ũ 1 + C̃2 Tyyy ϕ0 ⎠
0 0 3
B≡⎝ 2 ,
Tyλ ϕ0
0 0
C̃2 6Tyy ϕ0 ũ 1 + Tyyy
0
ϕ30
∗
∗ ϕ0 0 0 0
C ≡ , D≡ , m = 2,
0 ϕ∗0 0 0
(c0 ): dim R(B) = 2,
(c1 ): ⎛l1 ∈ Y and l2 ∈ R,
For the vector
1 0 ⎞
⎝ C̃ 2 T 0
ϕ 0 l 1 + 2T ϕ 0 + 6 C̃ 2 T 0
ϕ 0 ũ 1 + C̃ 2 T 0
ϕ 3
yyy 0 l 2 ⎠
yλ
2 yλ yy

Tyλ ϕ0 l1 + C̃2 6Tyy ϕ0 ũ 1 + Tyyy ϕ0 l2
0 0 0 3
is in R(B).
v
We know that 1 ∈ R(A) ⇔ ψ0∗ v1 = 0 and ψ0∗ v2 = 0.
v2
Let us now consider the matrix B. We set
.
0 = C̃2 ψ0∗ Tyλ 0
ϕ0 l 1
1
+ 2ψ0∗ Tyλ 0
ϕ0 + 6C̃2 ψ0∗ Tyy0
ϕ0 ũ 1 + C̃2 ψ0∗ Tyyy 0
ϕ30 l2 ,
.
2
0 = ψ0∗ Tyλ 0
ϕ0 l1 + C̃2 6ψ0∗ Tyy 0
ϕ0 ũ 1 + ψ0∗ Tyyy 0
ϕ30 l2 .
These
⎛ two equations can be written in matrix-vector notation as
1 ⎞
⎝ C̃2 a1 2 2a 1 + C̃ 2 a3 ⎠ l1 0
= .
l2 0
a1 C̃2 a3
Since the determinant of the system matrix does not vanish, we obtain
0
l1 = 0 and l2 = 0, i.e., R(A) ∩ R(B) = .
0
The conditions (c2 ) & (c3 ) are fulfilled trivially. Thus, we have shown that operator
A is bijective. The claim of the theorem now follows analogously to the proofs of
the transformation techniques presented in the previous sections.
Fig. 5.22 Curves of nontrivial solutions passing through nonisolated solutions of the perturbed
problem
5.6.3 Solution Curves Through Nonisolated Solutions
Let τ̄ be a fixed value of the perturbation parameter. It is now our goal to determine a
curve of nontrivial solutions of the operator equation (5.181), which passes through
the corresponding nonisolated solution ( ȳ, ϕ̄, λ̄, τ̄ ), see Fig. 5.22.
For this solution curve, we use the following ansatz
y(δ) = ȳ + δ ϕ̄ + δ 2 w(δ), w ∈ X̄ 1 ,
(5.215)
λ(δ) = λ̄ + δ 2 ρ(δ), ρ ∈ R.
At z̄ ≡ ( ȳ, λ̄, τ̄ ) we expand T into a Taylor series:
T (y, λ, τ̄ ) = T (z̄) + Ty (z̄)(y − ȳ) + Tλ (z̄)(λ − λ̄)

1 (5.216)
+ Tyy (z̄)(y − ȳ)2 + R5 (y, λ, τ̄ ).
2
With respect to the ansatz (5.215), the remainder R5 satisfies R5 = O(δ 3 ).

Inserting (5.215) and (5.216) into the equation
T (y, λ, τ̄ ) = 0, (5.217)
we obtain
1
0 = δTy (z̄)(ϕ̄ + δw) + Tλ (z̄)δ 2 ρ + Tyy (z̄)(ϕ̄ + δw)2 + R5 (y, λ, τ̄ ).
2
Since Ty (z̄)ϕ̄ = 0, the right-hand side is O(δ 2 ). Dividing the equation by δ 2 , we can
write the resulting equation and the equation ϕ̄∗ w = 0 in a compact notation as

Ty (z̄) Tλ (z̄) w(δ) g(w, ρ; δ)
=− , (5.218)
ϕ̄∗ 0 ρ(δ) 0
where
1 1
g(w, ρ; δ) ≡ Tyy (z̄)(ϕ̄ + δw)2 + 2 R5 .
2 δ
The solvability properties of the equation (5.218) are stated in the next theorem.
Theorem 5.67 Let the operator Ty (z̄) satisfy
dim N (Ty (z̄)) = 1, N (Ty (z̄)) = span{ϕ̄}, ϕ̄ = 1.
Furthermore, assume that

b2 ≡ ψ̄ ∗ Tλ (z̄) = 0, (5.219)
where ψ̄ ∗ is defined by
N (Ty∗ (z̄)) = span{ψ̄ ∗ }, ψ̄ ∗ = 1.
Then, there exists a real constant δ0 > 0 such that for |δ| ≤ δ0 problem (5.218) has
a continuous curve of isolated solutions
η(δ) ≡ (w(δ), ρ(δ))T .
By inserting η(δ) into the ansatz (5.215) it is possible to construct a continuous curve
of solutions of the equation (5.217), which contains the nonisolated solution z̄.
Proof Consider (5.218) for δ → 0. We obtain

1
w(0) Ty (z̄) Tλ (z̄) w(0) Tyy (z̄)ϕ̄2
A ≡ =− 2 .
ρ(0) ϕ̄∗ 0 ρ(0) 0
We set
A ≡ Ty (z̄), B ≡ Tλ (z̄),
C ∗ ≡ ϕ̄∗ , D ≡ 0, m = 1,
(c0 ): dim R(B) = 1,
(c1 ): It holds v ∈ R(A) ⇔ ψ̄ ∗ v = 0.
We have Tλ (z̄)α ∈ R(B), and because of the assumption (5.219),
ψ̄ ∗ Tλ (z̄)α vanishes if and only if α = 0.
Therefore, R(A) ∩ R(B) = {0}.
(c2 ) & (c3 ) are fulfilled trivially.
Thus, we have shown that the operator A is bijective. The claim of the theorem now
follows analogously to the proofs of the transformation techniques presented in the
previous sections.
5.7 Path-Following Methods for Simple Solution Curves
5.7.1 Tangent Predictor Methods
In the previous sections, we have presented the analytical and numerical methods
for operator equations of the form (5.11). The reason for this general approach
is that the extension and transformation techniques are not restricted to ODEs. If
the operator T is defined appropriately, these general techniques can also be used
to solve, e.g., parametrized nonlinear algebraic equations, parametrized nonlinear
integral equations and parametrized nonlinear partial differential equations.
However, in this section we want to deal exclusively with BVPs of the form (5.10).
As before, let E( f ) be the solution field of the BVP.
Assume that a solution z (1) ≡ ( y(1) , λ(1) ) of the BVP (5.10) has been determined.
Now, the path-following problem consists in the computation of further solutions
( y(2) , λ(2) ), ( y(3) , λ(3) ), . . .
on the corresponding solution branch until a certain endpoint λ = λb is reached.
The kth step of the path-following method starts with a known solution
( y(k) , λ(k) ) ∈ E( f ), and for the next value λ(k+1) of the parameter λ it is attempted
to determine a solution ( y(k+1) , λ(k+1) ) ∈ E( f ) :
( y(k) , λ(k) ) → ( y(k+1) , λ(k+1) ).
When the so-called predictor–corrector methods are used, the step k → k + 1 is

realized in two partial steps (see Fig. 5.23):
predictor step corrector step

( y(k) , λ(k) ) −−−−−−→ ( ȳ(k+1) , λ̄(k+1) ) −−−−−−→ ( y(k+1) , λ(k+1) ).
Fig. 5.23 Predictor-corrector method with a tangent predictor
In general, the result ( ȳ, λ̄) of the predictor is not a solution of (5.10). Rather,
the predictor step provides sufficiently accurate starting values for the subsequent
corrector iteration, so that this iteration converges. The special predictor-corrector
methods can be subdivided into two classes. For the first class, the most work is spend
to determine a predictor, which is near to the solution branch. Thus, the corrector
requires only a few iteration steps to generate a sufficiently exact approximation
of the solution. The members of the other class use an almost globally convergent
corrector iteration. Therefore, the starting values can be quite inaccurate, i.e., the
result of the predictor is generated with a reduced amount of work.
In the following, the distance between two successive solutions ( y(k) , λ(k) ) and
(k+1)
(y , λ(k+1) ) is referred to as step length. As we will see later, a relation must be
added to the BVP (5.10), which determines the position of a solution on the solution
branch. This relation depends on the special strategy of the parametrization, which
is used to follow the branch.
The path-following methods can be distinguished, among other things, by the
following key elements:
1. the predictor,
2. the corrector,
3. the strategy of parametrization, and
4. the control of the step length.
Elements 1–3 can be varied independently of each other. However, the control
of the step length must be consistent with the predictor, the corrector, and the
parametrization.
5.7 Path-Following Methods for Simple Solution Curves 271
In the following, one of the simplest path-following methods is described, which

is based on the tangent predictor. Let us assume that a solution ( y(k) , λ(k) ) has already
been calculated. The next step consists in the determination of the tangent (v (k) , η (k) )
to the solution curve at ( y(k) , λ(k) ). This can be done by solving the linearization of
the ODE (5.11). To simplify the representation, we refer to the operator form (5.11)
of the BVP and write the tangent equation as
Ty(k) v (k) + Tλ(k) η (k) = 0, (5.220)
with (v (k) , η (k) ) = 1. On this tangent, a predictor point
(y (k+1,0) , λ(k+1,0) ) ≡ (y (k) , λ(k) ) + τ (k) (v (k) , η (k) ) (5.221)
is determined using the step length τ (k) . This predictor point is used as a starting
approximation for the subsequent corrector iteration, which is applied to the nonlinear
system of equations
T (y, λ) = 0,
(5.222)
l (k) (y, λ) − l (k) (y (k+1,0) , λ(k+1,0) ) = 0
to compute the next solution (y (k+1) , λ(k+1) ) ∈ E( f ). Here, l (k) is a suitably chosen
functional that characterizes the type of parametrization.
In Algorithm 2.1 the above path-following technique is described in more detail.
Algorithm 2.1 Path-following with a tangent predictor
Step 1:Choose τmax > 0, τ (1) ∈ (0, τmax ), α ∈ (0, 1), η (0) ∈ (−1, 1);
Give ( y(1) , λ(1) ) ∈ E( f ), l0 ∈ {±l (0) , ±la(1) , . . . , ±la(n) },
with l (0) ≡ η (0) , la(i) ≡ vi (a), i = 1, . . . , n;
Set k := 1;
Step 2:If lk−1 = ±l (0) then determine v (k) as the solution of the following BVP
v (x) = f y (x, y(k) ; λ(k) )v(x) + η (k−1) f λ (x, y(k) ; λ(k) ),
Ba v(a) + Bb v(b) = 0,
and set η (k) := η (k−1) ;
If lk−1 = ±la(i) , i ∈ {1, . . . , n}, then determine (v (k) , η (k) ) as the solution
of the following BVP
v (x) = f y (x, y(k) ; λ(k) )v(x) + η(x) f λ (x, y(k) ; λ(k) ),
η (x) = 0,
Ba v(a) + Bb v(b) = 0,
vi (a) = ±1;
Step 3:Determine lk according to the following procedure:

(a) kmax := max{|v1 (a)|, . . . , |vn (a)|, |η|},
(b) If kmax = |η| then lk := sign(η)l (0) ,
(c) If kmax = |vi (a)| then lk := sign(vi (a))la(i) ,
1
(d) If lk = lk−1 then (v (k) , η (k) ) := (v (k) , η (k) );
kmax
Step 4:Determine the starting approximations for the computation of a
new point on the solution curve:
y(k+1,0) := y(k) + τ (k) v (k) ,
λ(k+1,0) := λ(k) + τ (k) η (k) ;
Step 5:If lk = ±l (0) then determine y(k+1) as the solution of the
following BVP
y (x) = f (x, y; λ(k+1,0) ),
Ba y(a) + Bb y(b) = 0,
and set λ(k+1) := λ(k+1,0) ;
If lk = ±la(i) then determine ( y(k+1) , λ(k+1) ) as the solution
of the following BVP
y (x) = f (x, y; λ),
λ (x) = 0,
Ba y(a) + Bb y(b) = 0,
(k+1,0)
yi (a) = yi (a);
Step 6:If the new point ( y(k+1) , λ(k+1) ) could not be determined
in Step 5 then set
τ (k) := ατ (k) ,
and go to Step 2.
If the new point has been computed in Step 5 then choose
τ (k+1) ∈ [τ (k) , τmax ],
set
k := k + 1,
and go to Step 2.
Possible termination criteria for this algorithm are:

• a boundary of the interesting parameter interval [λa , λe ] has been reached,
• the maximum number of points on the solution curve has been computed,
• if the BVPs are solved by a shooting method, the condition of the associated
systems of linear algebraic equations is too bad,
• the step length τ (k) is too small.
5.7.2 Arclength Continuation
Now, we come to another variant of the path-following method, the so-called

arclength continuation. The idea behind this path-following method is that a simple
solution curve can be parametrized by the arclength, i.e.,
y = y(s), λ = λ(s). (5.223)
Let us assume that for a given s (k) the corresponding solution ( y(k) , λ(k) ) ∈ E( f ) is
known. Then,
(k) (k) d y (k) dλ (k)
ẏ , λ̇ ≡ (s ), (s )
ds ds
represents the direction of the solution curve at ( y(k) , λ(k) ).

To improve the clarity of the presentation, we come back to the operator form
(5.11) of the BVP, and insert the parametrization (5.223) into the operator equation.
This yields
T ( y(s), λ(s)) = 0.
Differentiating the above identity with respect to s, we obtain
Ty ( y(s), λ(s)) ẏ(s) + Tλ ( y(s), λ(s))λ̇(s) = 0. (5.224)
Setting s = s (k) , it follows
0 = Ty ( y(k) , λ(k) ) ẏ(k) + λ̇(k) Tλ ( y(k) , λ(k) ). (5.225)
Adding the normalization condition

2
ẏ(k) 2 + λ̇(k) = 1, (5.226)
the point ( ẏ(k) , λ̇(k) ) is uniquely determined up to sign by (5.225) and (5.226).
Comparing (5.225) with (5.220), we see that we have computed the tangent on
the solution curve.
Now, the new point ( y(k+1) , λ(k+1) ) = ( y(s (k+1) ), λ(s (k+1) )) on the solution curve
is determined as the solution of the following system of equations
Fig. 5.24 The plane, which

is defined by (5.228)
T ( y(s), λ(s)) = 0,
(5.227)
N ( y(s), λ(s), s) = 0,
where

N ( y(s), λ(s), s) ≡ ϕ∗k y(s) − y(k) + λ̇(k) λ(s) − λ(k) − s (k+1) − s (k) . (5.228)
Here, ϕ∗k ∈ Y ∗ is a linear functional that satisfies the conditions

ϕ∗k = ẏ(k) , ϕ∗k ẏ(k) = ẏ(k) 2 .
Formula (5.228) is the equation of a plane that is perpendicular to the tangent

( ẏ(k) , λ̇(k) ), and is located at a distance (s (k+1) −s (k) ) from ( y(k) , λ(k) ) (see Fig. 5.24).
This plane will intersect the curve Γ if (s (k+1) − s (k) ) and the curvature of Γ are not
too large.
The control of the step-size s ≡ s (k+1) − s (k) should be based on the conver-
gence behavior of the corrector iteration. In the monograph [70] it is shown that the
arclength continuation can be realized for simple solution curves.
The question is still open, how the tangent can be determined from (5.225) and
(5.226). Since we are dealing with the tracing of a simple solution curve, we must
distinguish between the two cases:
• ( y(k) , λ(k) ) is an isolated solution of the BVP (5.10).
The first step is to compute w from the Eq. (5.225), i.e.,
Ty(k) w = −Tλ(k) .
Then, we set
ẏ(k) ≡ aw, λ̇(k) ≡ a,
and determine a from (5.226) as
±1
a= .
1 + w2
The sign of a is chosen such that the orientation of the path is preserved. More
precisely, if ( ẏ(k−1) , λ̇(k−1) ) is the preceding tangent vector, then we require
T
ẏ(k−1) ẏ(k) + λ̇(k−1) λ̇(k) > 0.
Thus, the sign of a is chosen such that

T
a ẏ(k−1) w + λ̇(k−1) > 0.
• ( y(k) , λ(k) ) is a simple turning point of the BVP (5.10).

First, we consider the equation (5.225), i.e.,
Ty(k) ẏ(k) = −λ̇(k) Tλ(k) .
This equation is solvable if

λ̇(k) ψ0∗ Tλ(k) = 0,

=0
and we obtain λ̇(k) = 0. Thus ẏ(k) = c ϕ0 , where span{ϕ0 } = N (Ty(k) ), ϕ0 = 1.

Now, we insert this representation of ẏ(k) into the equation (5.226). It follows
c2 ϕ0 2 = 1, i.e., c = ±1.

=1
5.7.3 Local Parametrization
To conclude the description of path-following methods, we will discuss a third strat-

egy, which is often used and relatively simple. It is called local parametrization.
Here, each component of the vector z ≡ ( y, λ) can be admitted as homotopy para-
meter. Let us assume that two consecutive solutions z (i−1) ≡ ( y(i−1) , λ(i−1 ) and
z (i) ≡ ( y(i) , λ(i ) of the BVP (5.10) are known and located on the same simple solu-
tion curve. Then, that component of z is selected as the local homotopy parameter,
for which the relative change
" "
" z (i) − z (i−1) "
" p "
δz (i)
p
≡" ", p ∈ {1, . . . , n + 1},
p
" z (i)
p
"
was greatest in the ith continuation step. Then, the new point z (i+1) on the solution
curve is determined as the solution of the following system of equations

T (z)
T̃ (z) ≡ = 0, (5.229)
z p − z (i+1)
p
where
z (i+1)
p ≡ z (i)
p +s
(i+1)
, (5.230)
and
s (i+1) ≡ τ (i) z (i) (i−1)
p − zp = τ (i) s (i) . (5.231)
The step-size s (i+1) is controlled on the basis of the number αi of iteration steps,
which were necessary to approximate the system of equations (5.229) in the ith
continuation step with a prescribed accuracy. More precisely, the parameter τ (i) is
computed according to the formula
β
τ (i) = , (5.232)
αi
where the constant β is the desired number of iteration steps for the respective BVP-
solver. Thus, s (i+1) is increased if αi < β and decreased if αi > β.
If the iteration method does not converge in the (i + 1)th continuation step, i.e.,
the new point z (i+1) cannot be determined, then the step-size s (i+1) is reduced by
choosing
τ (i+1) := 0.2 τ (i+1) , (5.233)
and the (i + 1)th step is repeated.

Of particular importance is the generation of good starting vectors z (i+1,0) . There
are at least three strategies:
• z (i+1,0) := z (i) ,
• z (i+1,0) := z (i) + τ (i) w (i) , where w (i) is the tangent at z (i) , and
• z (i+1,0) := z (i) + τ (i) w̃(i) , where w̃(i) is the secant line through the points z (i) and
z (i−1) .
Let us now present some details for the implementation of this continuation
method for the BVP (5.10). In (5.229), the equation z p − z (i+1)
p = 0 is realized
by an additional boundary condition
z p (a) − κ = 0. (5.234)
With the strategy described above, the homotopy index p is searched among the
components of the vector
z(a) = (z 1 (a), . . . , z n (a), z n+1 (a))T , z n+1 (a) = λ.
To ensure, that the number of ODEs matches the number of boundary conditions,
the trivial ODE λ (x) = 0 is added. Thus, the (n + 1)-dimensional BVP results
y (x) = f (x, y(x); λ),

λ (x) = 0,
(5.235)
Ba y(a) + Bb y(b) = 0,
y p (a) = κ (if p ∈ {1, . . . , n}), or λ = κ (if p = n + 1).
With
Φ(x, z) ≡ ( f (x, z), 0)T
and
r(z(a), z(b)) ≡ (Ba y(a) + Bb y(b), z p (a) − κ)T
we write (5.235) in the form

z (x) − Φ(x, z)
T̃ (z) ≡ = 0, (5.236)
r(z(a), z(b))
where
T̃ : Z̃ → W̃ , Z̃ ≡ C1 ([a, b], Rn+1 ),
W̃ ≡ C([a, b], Rn × {0}) × Rn+1 .
Now, we come to the characterization of problem (5.236).

Theorem 5.68 Assume that z 0 ≡ ( y0 , λ0 ) ∈ E( f ) is a simple solution of the BVP
(5.10), and p = p0 is the actual homotopy index. Then
T̃ (z 0 ) : Z̃ → W̃
is bijective.
Proof (a) If T̃ (z 0 ) is applied to an element w ≡ (v, η) ∈ Y ×R, the following linear

BVP results ⎛ ⎞
v − f 0y v − η f 0λ
⎜ η ⎟
T̃ (z 0 )w = ⎜ ⎟
⎝ Ba v(a) + Bb v(b)⎠ = 0. (5.237)
w p0 (a)
Now, it must be shown that (5.237) has only the trivial solution w(x) ≡ 0.
(b) Let z 0 ≡ ( y0 , λ0 ) be an isolated solution of (5.10). In the neighborhood of z 0 , the
solution curve passing through z 0 can be parametrized by λ, i.e., we can suppose that
p0 = n + 1. The second and fourth equation in (5.237) imply η(x) ≡ 0. Therefore,
v (x) − f 0y v(x) = 0,
⇔ Ty0 v = 0.
Ba v(a) + Bb v(b) = 0
Since N (Ty0 ) = {0}, it follows v(x) ≡ 0. Thus, we obtain w(x) = 0.

(c) Let z 0 ≡ ( y0 , λ0 ) be a simple turning point of (5.10), i.e., it holds
Tλ0 ∈
/ R(Ty0 ).
From the first equation in (5.237), we obtain η(x) ≡ 0. This in turn implies p0 =
n + 1. Thus, (v, 0) ∈ N (T̃ (z 0 )) is the solution of the following BVP
v (x) − f 0y v(x) = 0,
Ba v(a) + Bb v(b) = 0, (5.238)
v p0 = 0.
Assume that v(x) ≡ 0. Then, there exists an index p̂ such that v p̂ (a) = 0. For the
homotopy index p0 it holds
|v p0 (a)| = max{|v1 (a)|, . . . , |vn (a)|} = 0,
which is in contradiction to the third equation in (5.238).

(d) Now, the surjectivity of T̃ (z 0 ) must be shown. We leave this part of the proof to
the reader for self-study.
In Algorithm 2.2 the local parametrization method is described in more detail.
Algorithm 2.2 Path-following with the local parametrization method
Step 1:(jumping on a solution curve)

Choose:
.
s ∈ [smin , smax ] = initial step-size,
.
λstart = initial value for λ,
.
redconst ∈ (0, 1) = constant that is used to reduce the step-size,
.
p E ∈ {1, 2, . . . , n + 1}, z p E , dir ∈ {−1, 1} = are used to
terminate the algorithm if z (i)
p E (a) exceeds z p E from the left
(dir = 1), or from the right (dir = −1),
.
β = desired number of iteration steps for the determination of
a solution of the BVP (5.235)

.
itmax = maximum number of path-following steps;
Set i := 0, p := n + 1, κ := λstart , q := 0;
Compute a solution z (i) of the nonlinear BVP (5.235);
Step 2:(initial homotopy step)
Set i := 1, p1 := p, κ := z (0)
p1 + s,
Compute a solution z (1) of the nonlinear BVP (5.235),

Set α1 := number of iteration steps required to compute z (1) ;
Step 3:(ith homotopy step)
Set i := i + 1,
(a) Computation of the homotopy index pi
" "
" (i−1) "
"zl (a) − zl(i−2) (a)"
Compute δl := " " , l = 1, . . . , n + 1,
" (i−1) "
"zl (a)" + 0.1
Set δmax := max δl and pi := pmax , with δ pmax = δmax ;
l
(b) Computation of the step-size s (i)

β
Compute τ (i−1) := ,
αi−1

Set s (i) := τ (i−1) z (i−1)
pi (a) − z (i−2)
pi (a) ;
τ (i−1) smin
If s (i) < smin then τ (i) = ;
s (i)
τ (i−1) smax
If s (i) > smax then τ (i) = ;
s (i)
(c) Computation of starting values for the corrector iteration
(i) without extrapolation: z (i,0) := z (i−1) ,

(ii) with extrapolation: z (i,0) := 1 + τ (i−1) z (i−1) − τ (i−1) z (i−2) ;
(d) Computation of the ith point on the solution curve
Set κ := z (i,0)
pi (a) and compute the solution z
(i)
of the BVP (5.235);

Set αi := number of iteration steps required to compute z (i) ;
(e) Test whether a final value has been exceeded
(i)
If z (i) (i−1)
p E (a) · dir > z p E · dir and sign z p E (a) − z p E (a) = dir then
set q := 4;
Step 4:(Evaluation)
If i = itmax then q := 4;
If z (i) could be determined and q = 4, then goto Step 3,

else goto Step 5;
Step 5:Set q := q + 1,
(a) Reducing the step-size
If q < 3 then
set τ (i) := redconst · τ (i) and goto Step 3(b);
(b) Selecting a new homotopy index
If q = 3 then
compute an index p̂i with δ p̂i ≡ max{δl }, and
l = pi
set pi := p̂i , goto Step 3(c);

(c) End of path-following
If q > 3 then Stop.
In Fig. 5.25, Algorithm 2.2 is visualized.
5.7.4 Detection of Singular Points
An important part of path-following algorithms is the detection of bifurcation and

turning points. This is realized by so-called test functions ϑ(z), which are monitored
during the path-following. A singular point is indicated by a zero of ϑ.
Fig. 5.25 Visualization of Algorithm 2.2

Definition 5.69 Let z (0) ≡ ( y(0) , λ(0) ) be a singular point (turning or bifurcation
point) of the BVP (5.10). Then ϑ(z) is called a test function for z (0) if
• ϑ is a continuous function in z,
• ϑ(z (0) ) = 0,
• ϑ(z) changes the sign at z (0) .
For each computed point z (i) on the solution curve, the value ϑ(i) ≡ ϑ(z (i) ) is
determined. If ϑ changes the sign during the transition from the ith to the (i + 1)th
point on the curve, i.e., if it holds
ϑ(i) ϑ(i+1) < 0,
then a singular point must be on the curve between z (i) and z (i+1) .
However, qualitatively different singular points can only be detected by different
test functions. Therefore, let us begin with the case of a simple turning point. An
appropriate test function is
dλ
ϑ(z) ≡ (z). (5.239)
dz p (a)
If z (0) ≡ ( y(0) , λ0 ) is a simple turning point, then we have ϑ(z (0) ) = 0. The continuity
of ϑ with respect to z follows from the continuity of the solution manifold of the
BVP (5.10).
When the ith homotopy step has been executed and the corresponding solution z (i)
is determined, the three parameter values λ(i−2) , λ(i−1) and λ(i) are known. Moreover,
the sign of ϑ(z) depends only on the change of the λ-values (using a sufficiently small
homotopy step-size). Therefore, by the expression
ϑ(i) ≡ ϑ(z (i) ) = (λ(i) − λ(i−1) ), (5.240)
an appropriate test function is also defined. Indeed, if it holds
ϑ(i−1) ϑ(i) < 0,
then a simple turning point has been crossed. But for the position of the turning point
with respect to the three computed points on the curve, there are two possibilities
(see Fig. 5.26).
In the case (a) the turning point will be crossed in the (i − 1)th homotopy step,
whereas in the case (b) the turning point is crossed in the ith step. In both cases the
crossing of a simple turning point can only be detected with the test function (5.240)
in the ith homotopy step.
Let us now come to bifurcation points. We will show how a test function can be
defined on the basis of the shooting method, which is used to solve the BVP (5.235).
It is sufficient to consider only the simple shooting method (see Sect. 4.2).
If the simple shooting method is applied to solve the BVP (5.10), this BVP is
transformed into the finite-dimensional problem
Fig. 5.26 Position of the

turning point w.r.t. the three
computed points
F(s) ≡ Ba s + Bb u(b, s; λ) = 0, (5.241)
where u(x) ≡ u(x, s; λ) denotes the solution of the IVP
u (x) = f (x, u(x); λ), u(a) = s. (5.242)
As we have explained in Chap. 4, the system of nonlinear algebraic equations

F(s) = 0 is usually solved by a damped and regularized Newton method. This
numerical method is based on the Jacobian of F, i.e., we obtain the Jacobian as an
by-product of the equation solver. Moreover, the BVP (5.10) possesses an isolated
solution y(0) (x) if and only if the Jacobian
∂ F(s) ∂u(b, s; λ)
M(s) ≡ = Ba + Bb (5.243)
∂s ∂s
is nonsingular at s = s(0) , with s(0) ≡ y(0) (a). Thus, if z (0) is a singular point of the
operator T in (5.11), with dim N (T (z (0) )) > 0, then det(M(s(0) )) = 0.
It is now obvious to define a test function for the detection of bifurcation points
as follows
ϑ(z) ≡ det(M(s)), (5.244)
where s = z(a) establishes the connection between z(x) and s.

We are now facing with the following problem. In the path-following methods,
the enlarged BVP (5.235) is used instead of the BVP (5.10). Thus, the question is
where we can find the Jacobian M(s) of the given BVP (5.10). In order to find an
answer, we write (5.235) in the form
ỹ (x) = f̃ (x, ỹ(x)),

(5.245)
B̃a ỹ(a) + B̃b ỹ(b) = β,
with
y f 0
ỹ ≡ , f̃ ≡ , β≡ n ,
λ 0 κ

Ba 0n Bb 0n
B̃a ≡ , B̃b ≡ T ,
(e( p) )T 0n 0
and e( p) ∈ Rn+1 is the pth unit vector. The associated IVP is
ũ (x) = f̃ (x, ũ(x)), ũ(a) = s̃ ∈ Rn+1 , (5.246)
with ũ ≡ (u, λ)T . The simple shooting method transforms the BVP (5.245) into the
finite-dimensional problem
F̃(s̃) ≡ B̃a s̃ + B̃b ũ(b; s̃) − β = 0. (5.247)
If the Newton method is used to solve (5.247), the Jacobian of F̃ must be determined.
It holds
∂ F̃(s̃) ∂ ũ(b; s̃)

M̃(s̃) = = B̃a + B̃b
∂ s̃ ∂ s̃ "
⎛ ⎞
∂u j (b; s1 , . . . , sn , sn+1 ) ""
" v(b; s , . . . , s )
= B̃a + B̃b ⎝
1 n+1 ⎠
∂si i, j=1,...,n
w(b; s1 , . . . , sn+1 ) 1

X e
v(b; s1 , . . . , sn+1 )
= B̃a + B̃b
w(b; s1 , . . . , sn+1 ) 1

Ba + Bb X e
Bb v(b; s1 , . . . , sn+1 )
=
eTp

M(s) Bb v(b; s1 , . . . , sn+1 )
= ,
eTp
where "
∂u j (b; s1 , . . . , sn , sn+1 ) ""
v(b; s1 , . . . , sn+1 ) ≡ " ,
∂sn+1 j=1,...,n
"
∂u n+1 (b; s1 , . . . , sn , sn+1 ) ""
w(b; s1 , . . . , sn+1 ) ≡ " ,
∂s i i=1,...,n
and X e has been defined in Sect. 4.2.

Thus, the Jacobian M(s), which is required to compute the test function ϑ(z)
(see formula (5.244)), can be obtained from the Jacobian M̃(s̃) by deleting the last
Fig. 5.27 Plot of the test function det(M) versus λ for problem (5.186) with τ = 1;
TP—turning point; 1, . . . , 4—bifurcation points; ∗—points where the path-following was
started;?—unknown singularity
row and column. In case that the multiple shooting method is used to solve the
nonlinear equations (5.247), we have to proceed block-wise, i.e., we have to delete
in each (n + 1) × (n + 1) block of the Jacobian M̃ the last row and column.
Using (5.244), for each computed point z (i) on the solution curve the value ϑ(i) ≡
ϑ(z (i) ) is determined. If ϑ changes the sign during the transition from the ith to the
(i + 1)th point on the curve, a singular point must be on the curve between z (i)
and z (i+1) . Moreover, based on the plot of the test function, it is often possible to
recognize that the determinant is zero and there must exist a singular point between
z (i) and z (i+1) . In Figs. 5.27 and 5.28 the plot of the test function (5.244) and the
corresponding bifurcation diagram are given for the perturbed BVP (5.204) with
τ = 1.
The unknown singularity detected in Fig. 5.27 (see black square) must be studied
separately. It can be seen that for λ = 0 problem (5.186) is reduced to a singular
linear problem and the entire straight line L is a solution.
During path-following, it is also possible to determine rough approximations of
the critical parameters by interpolation. To simplify the notation, we abbreviate the
initial values y p (a) and z p (a) with y p and z p , respectively. Let us begin with simple
turning points. Here, we use the obvious ansatz
λ = c1 (z p − c2 )2 + c3 , p = n + 1. (5.248)
Fig. 5.28 Bifurcation diagram of problem (5.186); ∗—points where the path-following was started
In the ith homotopy step, the free parameters c1 , c2 , c3 ∈ R can be determined by

inserting the computed solutions z (i−2) , z (i−1) , and z (i) into the ansatz (5.248) as
follows:
λ(i) = c1 (z (i)
p − c2 ) + c3 ,
2
λ(i−1) = c1 (z (i−1)
p − c2 )2 + c3 ,
λ(i−2) = c1 (z (i−2)
p − c2 )2 + c3 .
The first equation yields

c3 = λ(i) − c1 (z (i)
p − c2 ) .
2
Subtracting the second equation from the first one gives
λ(i) − λ(i−1) = c1 (z (i) (i−1)

p − c2 ) − c1 (z p
2
− c2 )2 ,
and we obtain
λ(i) − λ(i−1)
c1 = .
(z (i) (i−1)
p − c2 ) − (z p
2 − c2 )2
Now we substitute the above expression for c3 into the third equation, and get
!
λ(i−2) − λ(i) = c1 (z (i−2)
p − c2 )2 − (z (i)
p − c2 ) .
2
Substituting the expression for c1 into this equation yields
λ(i−2) − λ(i) (z (i−2)

p − c2 )2 − (z (i)
p − c2 )
2
= .
λ(i) − λ(i−1) (z (i)
p − c2 ) − (z p
2 (i−1)
− c2 )2
Let us set
λ(i−2) − λ(i)
w(i) ≡ .
λ(i) − λ(i−1)
Then, we can write

# 2 2 $ (i) !
z (i)
p − z (i−1)
p w +2c2 w (i) z (i−1)
p − z (i)
p
2 (i) 2 !
= z (i−2)
p − z p + 2c2 z (i) (i−2)
p − zp .
It follows 2 (i) 2 # (i) 2 (i−1) 2 $ (i)

z (i−2)
p − zp − zp − zp w
c2 = # $ # $ .
2 w (i) z (i−1)
p − z (i)
p − z (i) (i−2)
p − zp
Differentiating the ansatz (5.248) gives
dλ
= 2c1 (z p − c2 ).
dz p
This derivative must vanish in the turning point, i.e., it holds

.
2c1 (z p − c2 ) = 0.
Thus
ẑ tpp = c2 and λ̂tp = c3 (5.249)
tp
are approximations for the critical value z p and the critical parameter λtp , respec-
tively.
Next, we come to simple bifurcation points. Let us assume that a2 = 0 and the
test function (5.244) is integrated into the path-following algorithm. If ϑ(z) changes
the sign during the transition from the (i − 1)th to the ith point of the traced solution
curve Γ1 , the following quadratic ansatz can be used for the approximation of these
bifurcation points
ϑ = c1 z 2p + c2 z p + c3 . (5.250)
Inserting three successive value pairs
(ϑ(i−2) , z (i−2)
p ), (ϑ(i−1) , z (i−1)
p ), (ϑ(i) , z (i)
p )
into the ansatz (5.250) yields

2
ϑ(i) = c1 z (i)
p + c2 z (i)
p + c3 ,
(i−1)

(i−1) 2
ϑ = c1 z p + c2 z (i−1)
p + c3 ,
2
ϑ(i−2) = c1 z (i−2)
p + c2 z (i−2)
p + c3 .
From the third equation, we have

2
c3 = ϑ(i−2) − c1 z (i−2)
p − c2 z (i−2)
p . (5.251)
Subtracting the third equation from the second one gives

# 2 2 $
ϑ(i−1) − ϑ(i−2) = c1 z (i−1)
p − z (i−2)
p + c2 (z (i−1)
p − z (i−2)
p ),
and we obtain
ϑ(i−1) − ϑ(i−2)
c2 = − c1 (z (i−1) + z (i−2) ). (5.252)
z (i−1) − z (i−2)
p p
p p
Now, we substitute the above expressions for c2 and c3 into the first equation, and
get
2 ϑ(i−1) − ϑ(i−2) (i)
ϑ(i) = c1 z (i) + (i−1) z p − c1 (z (i−1) + z (i−2) )z (i)
− z (i−2)
p p p p
zp p
2 ϑ(i−1) − ϑ(i−2) (i−2)
+ ϑ(i−2) − c1 z (i−2) − (i−1) zp
− z (i−2)
p
zp p
+ c1 (z (i−1)
p + z (i−2)
p )z (i−2)
p .
It follows

1 ϑ(i) − ϑ(i−2) ϑ(i−1) − ϑ(i−2)
c1 = − . (5.253)
z (i) (i−1)
p − zp z (i) (i−2)
p − zp z (i−1)
p − z (i−2)
p
The solution of the equation
c1 z 2p + c2 z p + c3 = 0,
where c1 , c2 and c3 are given in (5.251)–(5.253), can be used as an approximation

ẑ bp for the critical value z bp , i.e.,
.
c2 c22 c3
ẑ bp =− ± − . (5.254)
2c1 4c12 c1
Since ẑ bp must be located between z (i)

p and z p
(i−1)
, the sign in (5.254) has to be chosen
such that (i)
z p − ẑ bp ẑ bp − z (i−1)
p > 0. (5.255)
Thus, if it holds p = n + 1, we have determined an approximation λ̂b = ẑ bp for the

critical parameter λb .
If a2 = 0 and the test function ϑ(z) does not change the sign during the path-
following of a solution curve Γ1 but is nearly zero at λ = λ̃ (see Fig. 5.28), the
other solution curve Γ2 should be traced in the neighborhood of λ̃. In the case that
the corresponding test function changes its sign near λ̃, there must be a bifurcation
point. Now, using the ansatz (5.250) with p = n + 1 for the solution curve Γ2 , the
critical parameter λb can be approximated as described above.
5.8 Parametrized Nonlinear BVPs from the Applications
5.8.1 Buckling of Thin-Walled Spherical Shells
In this section, we present some examples of parametrized nonlinear two-point BVPs

from the applications where we have used the numerical techniques discussed in the
previous chapters to determine the corresponding solution manifold.
The study of the buckling behavior of thin-walled spherical shells under an exter-
nal pressure constitutes still today a challenge to mathematicians and engineers. It
seems to be hopeless to handle all possible deformations in a systematic way, because
in the solution manifold of the corresponding nonlinear, parametrized equations sin-
gularities of very high order can occur. Therefore, many authors simplify the buckling
problem by presuming that only axisymmetric deformations of the shell are possi-
ble. As explained in [44], this restriction is not as academic as it might appear at the
first moment, because the production process of spherical shells frequently favors
axisymmetric imperfections, which create a preference for axisymmetric buckling
patterns.
The shell equations presented in [19, 121] belong to the most-studied models
describing the axisymmetric buckling of spherical shells. Both models are based on
the assumptions:
• the material of the shell is elastic, isotropic and homogeneous,
• the thickness of the shell is small and constant,
• only axisymmetric deformations of the spherical shell are possible,
• the uniform external pressure keeps always the direction pointing towards the
origin of the spherical shell.
5.8 Parametrized Nonlinear BVPs from the Applications 289
The governing equations in [19] constitute a parametrized nonlinear BVP for a system
of four first-order differential equations

y1 (x) = (ν − 1) cot(x)y1 (x) + y2 (x) + k cot 2 (x) − λ y4 (x)
+ cot(x)y2 (x)y4 (x),
y2 (x) = y3 (x),

y3 (x) = cot 2 (x) − ν y2 (x) − cot(x)y3 (x) − y4 (x) − 0.5 cot(x)y42 (x),
1 − ν2
y4 (x) = y1 (x) − ν cot(x)y4 (x),
k
y2 (0) = y4 (0) = y2 (π) = y4 (π) = 0. (5.256)
The components y1 , y2 , y3 , and y4 are proportional, respectively, to the radial bending

moment, the transversal shear, the circumferential membrane stress, and the angle
of rotation of a tangent to the meridian. The quantity ν is Poisson’s ratio. The radii
of the inner and outer surfaces of the spherical shell are given by R ∓ h/2, where R
is the radius of the midsurface. Finally,
2
pR 1 h
λ≡ , k≡ , (5.257)
2Eh 12 R
where E denotes Young’s modulus (see, e.g., [11]) and p is the external pressure.
The underlying geometry of the shell is sketched in Fig. 5.29,(a).
Fig. 5.29 Geometry of the spherical shell

The model equations formulated in [121], Appendix K, represent a BVP for a

system of two second-order differential equations:
sin(x − β(x)) − sin(x)

β (x) = − cot(x)β (x) − cos(x − β(x))
sin2 (x)
cos(x − β(x)) − cos(x) 1 sin(x − β(x)) ∗
+ν − ψ (x)
sin(x) δ sin(x)
λ
− 2 sin(x) cos(x − β(x)),
δ
ψ ∗ (x) = − cot(x)ψ ∗ (x)
2
cos (x − β(x)) sin(x − β(x))
+ − ν [1 − β (x)] ψ ∗ (x)
sin2 (x) sin(x)
1 cos(x − β(x)) − cos(x)
+ − 2λ sin(x − β(x)) cos(x − β(x))
δ sin(x)
− 2λν sin(x) cos(x − β(x))[1 − β (x)] + 4λν sin(β(x))
+ 12λ sin(x) cos(x),
β(0) = ψ ∗ (0) = β(π) = ψ ∗ (π) = 0.
(5.258)
Here, β describes the angle enclosed by the tangents to the deformed and undeformed
meridian. Moreover, ψ ∗ is a stress function depending on the horizontal stress resul-
tant H ,
4H sin(·)
ψ ∗ (·) ≡ , (5.259)
R pc
where 2
h 1
pc ≡ 4Eδ 2 12(1 − ν 2 ), δ 2 ≡ ,
R 12(1 − ν 2 )
and h, R, ν, and E are defined as before. The loading parameter λ is related to

the external pressure p by λ ≡ p/ pc . The geometry of the shell is visualized in
Fig. 5.29,(b).
Because of the assumptions stated above, it is sufficient to consider only an arc
of the midsurface, which is indicated by a dotted line. The defining equations for the
displacements u, W (Fig. 5.29,(a)) and u, w (Fig. 5.29,(b)) are given in [19, 121],
respectively.
The shell Eqs. (5.256) and (5.258) have been studied by many authors (see, e.g.,
[44, 64, 72, 109]). In the paper [62], a general class of parametrized nonlinear BVPs
is defined, which includes equations of the type (5.256) and (5.258). For this class
the occurring bifurcation phenomena have been analyzed in detail, also with the
analytical and numerical methods described in this book. Here, we present the results
of some experiments, which were based on the following values of the internal
parameters: ν = 0.32 (steel), k = 10−5 and δ = 0.00333779 (thin sphere). We
have solved the corresponding BVPs with the multiple shooting code RWPM [59].
Fig. 5.30 Bifurcation diagram of the BVP (5.256)
Fig. 5.31 Bifurcation diagram of the BVP (5.258)
Moreover, the code RWPKV [65] was used to realize the tracing of solution curves.
In Figs. 5.30 and 5.31 parts of the solution manifold of the BVPs (5.256) and (5.258),
respectively, are shown. For the definition of the functionals l1∗ and l2∗ see [62]. Stars
mark the primary bifurcation points.
It can be seen that there are only a few differences in the solution fields of the
problems (5.256) and (5.258), which differ significantly in the degree of the nonlin-
earity. This was an unexpected result. Moreover, in Fig. 5.32 two deformed shells
are presented.
Fig. 5.32 Examples for deformed shells (computed with BVP (5.258))
Fig. 5.33 The bipedal spring-mass model for walking
5.8.2 A Bipedal Spring-Mass Model
One of the most successful templates for the simulation of human walking and
running is the spring-mass model. The model for running and hopping consists of a
mass point representing a center of mass of the human body riding on a linear leg
spring [22]. For the study of walking gaits, the planar bipedal model with two leg
springs was introduced in [41]. It consists of two massless leg springs supporting the
point mass m (see Fig. 5.33).
Both leg springs have the same stiffness k0 and rest length L 0 . For a fixed time
t the location and velocity of the center of mass in the real plane R2 are given by
(x(t), y(t))T and (ẋ(t), ẏ(t))T , respectively. Here, the dot denotes the derivative w.r.t.
the time t. Any walking gait is completely characterized by four fundamental system
parameters (leg stiffness k0 , angle of attack α0 , rest length L 0 , system energy E 0 )
and the four-dimensional vector of initial conditions (x(t0 ), ẋ(t0 ), y(t0 ), ẏ(t0 ))T =
(x0 , ẋ0 , y0 , ẏ0 )T .
A walking step comprises a single-support phase and a double-support phase
(see Fig. 5.33). Events of touch-down and take-off are transitions between the phases.
The trajectory of the center of mass in each phase is the solution of an IVP. The
calculation of a walking step starts at time t = t0 during the single-support phase
at the instant of the vertical leg orientation VLO (see Fig. 5.33 and [105]). Using
VLO as the initial point of a step (i.e., as the Poincaré section), allows to reduce the
dimension of the return map [84]. Here, the center of mass is located exactly over
the foot point of the supporting leg spring.
The initial values for the first IVP are (x0 , ẋ0 , y0 , ẏ0 )T . The initial vector of each
subsequent phase is the last point of the corresponding previous phase. The motion
of the center of mass during the single-support phase is described by the equations
x(t) − xFP1
m ẍ(t) = k0 (L 0 − L 1 (t)) ,
L 1 (t)
(5.260)
y(t)
m ÿ(t) = k0 (L 0 − L 1 (t)) − mg,
L 1 (t)
+ 2
where L 1 (t) ≡ x(t) − xFP1 + y 2 (t) is the length of the compressed leg spring
during stance. The position of the footpoint FP1 is given by (xFP1 , 0). The transition
(touch-down) from single-support phase to double-support phase happens, when the
landing condition
y(t1 ) = L 0 sin(α0 )
is fulfilled (see Fig. 5.33). Here, t = t1 is the time when the touch-down occurs.
The equations of the motion during the double-support phase are given by
x(t) − xFP1 x(t) − xFP2

m ẍ(t) = k0 (L 0 − L 1 (t)) + k0 (L 0 − L 2 (t)) ,
L 1 (t) L 2 (t)
y(t) y(t)
m ÿ(t) = k0 (L 0 − L 1 (t)) + k0 (L 0 − L 2 (t)) − mg, (5.261)
L 1 (t) L 2 (t)
+ 2
where L 2 (t) ≡ x(t) − xFP2 + y 2 (t) is the length of the second compressed
leg spring. The positions of both footpoints FP1 and FP2 are given by (xFP1 , 0) and
(xFP2 , 0), respectively. The transition (take-off) from the double-support phase to
the single-support phase occurs when the extending length L 1 of the first leg spring
reaches the rest length L 0 , i.e., when the take-off condition
2
x(t2 ) − xFP1 + y 2 (t2 ) = L 20
is fulfilled. Here, t = t2 is the time when the take-off occurs.

The walking step ends at the next VLO at time t = t3 , when the condition
x(t3 ) = xFP2 is fulfilled. The system is energy-conservative, i.e., the system energy
E 0 remains constant during the whole step.
Periodic solutions of the model often correspond to continuous locomotion pat-
terns. The investigations of stable periodic solutions is of particular interest. The
mathematical definition of stability is considered as the property of the system to
absorb small perturbations. Periodic walking solutions are found and analyzed using
the well-known Poincaré return map [95]. Their stability is determined by the value
of the corresponding Floquet multipliers [39]. The detailed mathematical description
can be seen, e.g., in [45].
Since the end of each phase of a running or walking step is defined by an event,
the times t1 , t2 , and t3 are not known in advance. Therefore, instead of solving the
IVPs successively it is more appropriate to scale each phase to the unit interval [0, 1]
and to solve all three phases simultaneously by a parametrized nonlinear BVP of the
form
y (x) = f ( y(x)), 0 < x < 1,
(5.262)
r( y(0), y(1); λ) = 0.
A further advantage of this strategy is that the numerical methods described in the
previous sections can be applied directly to the BVP (5.262) to determine the cor-
responding solution manifold. But notice that operator T in the operator equation
(5.11) has to be defined now as (see formulas (5.13) and (5.14))

y − f ( y)
T (y, λ) ≡ , (5.263)
r( y(0), y(1); λ)
Fig. 5.34 Initial VLO height y0 of periodic walking patterns versus the angle of attack α0 . The dots
represent singular points: TP turning point, TB secondary bifurcation point, HB Hopf bifurcation
point, PD perioddoubling bifurcation point
where the associated functional spaces must be chosen adequately. A special feature
of this BVP is that the bifurcation parameter λ occurs only in the boundary function r.
More information about this model and its extensions can be found in [81, 82, 83].
In Fig. 5.34, a part of the solution manifold of the bipedal spring-mass model
is shown. Here, the system energy is E 0 = 820 J corresponding to the average
horizontal velocity vx ≈ 1.1 m/s. Furthermore, we have set k0 = 16 kN/m, L 0 = 1 m,
m = 80 kg, and g = 9.81 m/s2 .
5.9 Exercises
Exercise 5.1 Draw the bifurcation diagrams for the following scalar functions
(l ∈ R):
1. l u = λ u,
2. l u + c u 2 = λ u, c = 0,
3. l u + c u 3 = λ u, c = 0,
4. a(u) = λ ⎧u, with
⎨ u, |u| < 1
a(u) ≡ (u 2 + 1)/2, u ≥ 1, .
⎩
−(u 2 + 1)/2, u ≤ −1.
Exercise 5.2 Analyze the solution field of the following equation in R2 and draw
the corresponding bifurcation diagram:
A u = λ u,
where
A u ≡ L u + C u,

β 0 u1 βu
Lu ≡ 1 = 1 1 , β2 > β1 > 0,
0 β2 u 2 β2 u 2
& /
γ1 u 1 (u 1 + u 2 )
2 2
Cu ≡ , γ1 > γ2 > 0.
γ2 u 2 (u 21 + u 22 )
Hint: It holds C(αu) = α3 C u.
Exercise 5.3 On the Internet at http://www-history.mcs.st-and.ac.uk, under the

menu item Famous Curves Index, you will find the curve of the trifolium. The cor-
responding formula is
(u 2 + v 2 )(v 2 + u(u + a)) = 4auv 2 .
1. Set a = 1, u = λ, and v = y.
2. Write this formula as an operator equation T (y, λ) = 0.
3. Formulate the corresponding extended system for the computation of simple

turning points and determine the solutions of this extended system (in order to do
that you can use the symbolic tool of the Matlab).
4. Draw the solution curve (using the Matlab functions contour and meshgrid)
and mark the turning points.
5. Set a = 1, u = y and v = λ and repeat the steps 2–4.
Exercise 5.4 On the website mentioned in Exercise 5.3, you will also find the curve
of the limacon of Pascal. The corresponding formula is
(u 2 + v 2 − 2au)2 = b2 (u 2 + v 2 ).
Set a = b = 1 and proceed as described in Exercise 5.3.
Exercise 5.5 On the website mentioned in Exercise 5.3, you will also find the eight
curve. The corresponding formula is
u 4 = a 2 (u 2 − v 2 ).
Set a = 1 and proceed as described in Exercise 5.3.
Exercise 5.6 Prove Corollary 5.17:

The limit point z 0 ∈ Z is a double turning point if and only if
a2 ≡ ψ0∗ Tyy
0 2
ϕ0 = 0 and a3 ≡ ψ0∗ (6Tyy
0
ϕ0 w0 + Tyyy
0
ϕ30 ) = 0,
1 0 2
where w0 ∈ Y1 is the unique solution of the equation Ty0 w0 = − Tyy ϕ0 .
2
Exercise 5.7 On the basis of the multiple shooting method (see Algorithm 4.2),
implement the path-following technique (see Algorithm 2.2) into a Matlab m-file
rwpmaink.m. Use rwpmaink to trace the solution curve of Bratu’s problem (5.53).
Determine the existing simple turning point with the appropriate extended system.
Draw the bifurcation diagram y (0) versus λ and mark the turning point.
Exercise 5.8 Use rwpmaink (see Exercise 5.7) to trace the solution curve of the
following BVP
y (x) = λ y(x) sin(λ y(x)), y(0) = y(1) = 1.
Draw the bifurcation diagram y (0) versus λ.

Hint: Start with λ = 0 and determine the solution field for λ ∈ [0, 25].
Exercise 5.9 Use rwpmaink (see Exercise 5.7) to trace the solution curve of the
following BVP (dance with a ribbon, see [85])

y (x) = sin (x − x 2 )y(x)2 − λ2 + λ3 , y(0) = y(1), y (0) = y (1).
5.9 Exercises 297
Determine the existing simple turning points with the appropriate extended system.
Draw the bifurcation diagram y(0) versus λ and mark the turning points.
Hint: Start with λ = 0 and determine the solution field for λ ∈ [−0.8, 1.2].
y (x) = λ y(x)ω(y), y(1) = 1, y (0) = 0,
where
γβ(1 − y)
ω(y) ≡ exp .
1 + β(1 − y)
Set γ = 20, β = 0.4, and determine the solution field for λ ∈ [0.01, 0.19]. Approx-
imate the singular points. Draw the bifurcation diagram y(0) versus λ and mark the
computed singular points.
Exercise 5.11 Use rwpmaink to trace the solution curve of the following BVP
(iron, see [85])
2
y (x) = sin(x 2 − x)y(x)2 + λ + λ, y(0) = y(1), y (0) = y (1).
Draw the bifurcation diagram y(0) versus λ and mark the singular points.
Hint: Start with λ = −0.4 and determine the solution field for λ ∈ [−1, 0].
(boomerang, see [85])
2
y (x) = sin(x 2 − x)y(x)2 + λ − λ, y(0) = y(1), y (0) = y (1).
Hint: Start with λ = 4 and determine the solution field for λ ∈ [0, 7].
(loop, see [85])
y (x) = −(λ + π 2 ) sin(y(x)) − λ4 , y(0) = 0, y(1) = 0.
Draw the bifurcation diagram y (0) versus λ and mark the singular points.
Hint: Start with λ = −1 and determine the solution field for λ ∈ [−1, 1].
(raindrop, see [85])

1
y (x) = x 3 − sin y(x)2 + λ3 + λ4 , y(0) = y(1), y (1) = 0.
4
(molar, see [85])

1
y (x) = x 3 − sin y(x)2 + λ3 − λ4 , y(0) = y(1), y (1) = 0.
4
Hint: Start with λ = 0.2 and determine the solution field for λ ∈ [−0.5, 1.2].
(Mordillo U, see [85])

y (x) = sin (x 2 − x)y(x)4 − λ y(x) + λ + λ2 , y(0) = y(1), y (0) = y (1).
Hint: Start with λ = −6 and determine the solution field for λ ∈ [−10, 7].
(between 8 and ∞, see [85])

y (x) = sin (x − x 2 )y(x)3 + λ y(x) + λ4 , y(0) = y(1), y (0) = y (1).
(speech bubble, see [85])

y (x) = sin (x − x 2 )y(x)2 y(x) − λ2 , y(0) = y(1), y (0) = y (1).
Hint: Start with λ = −1 and determine the solution field for λ ∈ [−1.8, 1.8].
(bow tie, see [85])

1
y (x) = sin (x − x )y(x)
2 2
+ x − 3
sin y(x)2
4
− 2λ4 + y(x)5 + λ6 + λ8 ,
y(0) = y(1), 2y (0) = y (1).
Hint: Start with λ = −1 and determine the solution field for λ ∈ [−1.5, 1.5].
5.9 Exercises 299
(Janus, see [85])

1
y (x) = x 3 − sin y(x)2 y(x) + λ sin(y(x),
4
y(0) = y(1), 2y (0) = y (1).
Hint: Start with λ = 0.5 and determine the solution field for λ ∈ [−1, 1].
References
1. Abbaoui, K., Cherruault, Y.: Convergence of Adomian’s method applied to differential equa-
tions. Comput. Math. Appl. 28(5), 103–109 (1994)
2. Abbaoui, K., Cherruault, Y.: Convergence of Adomian’s method applied to nonlinear equa-
tions. Mathl. Comput. Modell. 20(9), 69–73 (1994)
3. Abbasbandy, S.: The application of homotopy analysis method to nonlinear equations arising
in heat transfer. Phys. Lett. A 360(1), 109–113 (2006)
4. Abbasbandy, S., Shivanian, E.: Prediction of multiplicity of solutions of nonlinear boundary
value problems: novel application of homotopy analysis method. Commun. Nonlinear Sci.
Numer. Simul. 15, 3830–3846 (2010)
5. Abdelkader, M.A.: Sequences of nonlinear differential equations with related solutions. An-
nali di Matematica Pura ed Applicata 81(1), 249–258 (1969)
6. Abramowitz, M., Stegun, I.A.: Handbook of Mathematical Functions. Dover Publications,
New York (1972)
7. Adomian, G.: Nonlinear Stochastic Operator Equations. Academic Press Inc., Orlando, FL
(1986)
8. Adomian, G.: A review of the decomposition method in applied mathematics. J. Math. Anal.
Appl. 135, 501–544 (1988)
9. Adomian, G.: Solving Frontier Problems in Physics: The Decomposition Method. Kluwer,
Dordrecht (2004)
10. Adomian, G., Rach, R.: Noise terms in decomposition series solution. Comput. Math. Appl.
24(11), 61–64 (1992)
11. Allaby, M.: A Dictionary of Earth Sciences, 3rd edn. Oxford University Press, Oxford (2008)
12. Almazmumy, M., Hendi, F.A., Bakodah, H.O., Alzumi, H.: Recent modifications of Ado-
mian decomposition method for initial value problem in ordinary differential equations.
Am. J. Comput. Math. 2(3) (2012). doi:10.4236/ajcm.2012.23,030
13. Ascher, U.M., Mattheij, R.M.M., Russell, R.D.: Numerical Solution of Boundary Value Prob-
lems for Ordinary Differential Equations. Prentice Hall Series in Computational Mathematics.
Prentice-Hall Inc., Englewood Cliffs (1988)
14. Askari, H., Younesian, D., Saadatnia, Z.: Nonlinear oscillations analysis of the elevator cable
in a drum drive elevator system. Adv. Appl. Math. Mech. 7(1), 43–57 (2015)
15. Askeland, D.R., Phulé, P.P.: The Science and Engineering of Materials, 5th edn. Cengage
Learning (2006)
16. Azreg-Aïnou, M.: A developed new algorithm for evaluating Adomian polynomials. Comput.
Model. Eng. Sci. 42(1), 1–18 (2009)

DOI 10.1007/978-81-322-2812-7
302 References
17. Bader, G., Deuflhard, P.: A semi-implicit midpoint rule for stiff systems of ODEs. Numerische
Mathematik 41, 373–398 (1983)
18. Batiha, B.: Numerical solution of Bratu-type equations by the variational iteration method.
Hacettepe J. Math. Stat. 39(1), 23–29 (2010)
19. Bauer, L., Reiss, E.L., Keller, H.B.: Axisymmetric buckling of hollow spheres and hemi-
spheres. Commun. Pure Appl. Math. 23, 529–568 (1970)
20. Biazar, J., Shafiof, S.M.: A simple algorithm for calculating Adomian polynomials. Int. J.
Contemp. Math. Sci. 2(20), 975–982 (2007)
21. Bickley, W.G.: The plane jet. Philos. Mag. 28, 727 (1937)
22. Blickhan, R.: The spring-mass model for running and hopping. J. Biomech. 22(11–12), 1217–
1227 (1989)
23. Boresi, A.P., Schmidt, R.J., Sidebottom, O.M.: Advanced Mechanics of Materials. Wiley,
New York (1993)
24. Brezzi, F., Rappaz, J., Raviart, P.A.: Finite dimensional approximation of nonlinear problems.
Part III: Simple bifurcation points. Numer. Math. 38, 1–30 (1981)
25. Bucciarelli, L.L.: Engineering Mechanics for Structures. Dover Publications, Dover Civil and
Mechanical Engineering (2009)
26. Chandrasekhar, S.: Introduction to the Study of Stellar Structure. Dover Publications, New
York (1967)
27. Chow, S.N., Hale, J.K.: Methods of Bifurcation Theory. Springer, New York (1982)
28. Crandall, M.G., Rabinowitz, P.H.: Bifurcation from simple eigenvalues. J. Funct. Anal. 8,
321–340 (1971)
29. Dangelmayr, G.: Katastrophentheorie nichtlinearer Euler-Lagrange-Gleichungen und Feyn-
man’scher Wegintegrale. Ph.D. thesis, Universität Tübingen (1979)
30. Davis, H.T.: Introduction to Nonlinear Differential and Integral Equations. Dover Publications
(1960)
31. Decker, D.W., Keller, H.B.: Multiple limit point bifurcation. J. Math. Anal. Appl. 75, 417–430
(1980)
32. Dennis, J.E., Schnabel, R.B.: Numerical Methods for Unconstrained Optimization and Non-
linear Equations. Prentice-Hall Inc., Englewood Cliffs, NJ (1983)
33. Deuflhard, P.: Recent advances in multiple shooting techniques. In: Gladwell/Sayers (ed.)
Computational Techniques for Ordinary Differential Equations, pp. 217–272. Academic Press,
London, New York (1980)
34. Deuflhard, P., Bader, G.: Multiple shooting techniques revisited. In: Deuflhard, P., Hairer, E.
(eds.) Numerical Treatment of Inverse Problems in Differential and Integral Equations, pp.
74–94. Birkhäuser Verlag, Boston, Basel, Stuttgart (1983)
35. Durmaz, S., Kaya, M.O.: High-order energy balance method to nonlinear oscillators. J. Appl.
Math. 2012(ID 518684), 1–7 (2012)
36. Fardi, M., Kazemi, E., Ezzati, R., Ghasemi, M.: Periodic solution for strongly nonlinear
vibration systems by using the homotopy analysis method. Math. Sci. 6(65), 1–5 (2012)
37. Feng, Z.: On explicit exact solutions to the compound Burgers-KdV equation. Phys. Lett. A
293, 57–66 (2002)
38. Finlayson, B.: The Method of Weighted Residuals and Variational Principles. Academic Press
Inc., New York (1972)
39. Floquet, G.: Sur les équations différentielles linéaires à coefficients périodiques. Annales
scientifiques de l’École Normale Supérieure 12, 47–88 (1883)
40. Froese, B.: Homotopy analysis method for axisymmetric flow of a power fluid past a stretching
sheet. Technical report, Trinity Western University (2007)
41. Geyer, H., Seyfarth, A., Blickhan, R.: Compliant leg behaviour explains basic dynamics of
walking and running. Proc. R. Soc. B: Biol. Sci. 273(1603), 2861–2867 (2006)
42. Golub, G.H., Van Loan, C.F.: Matrix Computations. The John Hopkins University Press,
Baltimore and London (1996)
43. Golubitsky, M., Schaeffer, D.: Singularities and Groups in Bifurcation Theory, vol. I. Springer
Verlag, New York (1984)
References 303
44. Gräff, M., Scheidl, R., Troger, H., Weinmüller, E.: An investigation of the complete post-
buckling behavior of axisymmetric spherical shells. J. Appl. Math. Phys. 36, 803–821 (1985)
45. Guckenheimer, J., Holmes, P.: Nonlinear Oscillations, Dynamical Systems, and Bifurcations
of Vector Fields. Applied Mathematical Sciences, vol. 42. Springer (1983)
46. Hartman, P.: Ordinary Differential Equations. Birkhäuser Verlag, Boston, Basel, Stuttgart
(1982)
47. Hassan, H.N., Semary, M.S.: Analytic approximate solution for the Bratu’s problem by optimal
homotopy analysis method. Commun. Numer. Anal. 2013, 1–14 (2013)
48. He, H.J.: Semi-inverse method of establishing generalized variational principles for fluid
mechanics with emphasis on turbomachinery aerodynamics. Int. J. Turbo Jet-Engines 14(1),
23–28 (1997)
49. He, J.H.: Approximate analytical solution for seepage flow with fractional derivatives in
porous media. Comput. Meth. Appl. Mech. Eng. 167(1–2), 57–68 (1998)
50. He, J.H.: Variational iteration method–a kind of non-linear analytical technique: some exam-
ples. Int. J. Non-Linear Mech. 34(4), 699–708 (1999)
51. He, J.H.: Variational iteration method—some recent results and new interpretations. J. Com-
put. Appl. Math. 207(1) (2007)
52. He, J.H.: Hamiltonian approach to nonlinear oscillators. Phys. Lett. A 374(23), 2312–2314
(2010)
53. He, J.H.: Hamiltonian approach to solitary solutions. Egypt.-Chin. J. Comput. Appl. Math.
1(1), 6–9 (2012)
54. He, J.H., Wu, X.H.: Variational iteration method: new development and applications. Comput.
Math. Appl. 54(7–8), 881–894 (2007)
55. Hermann, M.: Ein ALGOL-60-Programm zur Diagnose numerischer Instabilität bei Verfahren
der linearen Algebra. Wiss. Ztschr. HAB Weimar 20, 325–330 (1975)
56. Hermann, M.: Shooting methods for two-point boundary value problems–a survey. In: Her-
mann, M. (ed.) Numerische Behandlung von Differentialgleichungen, Wissenschaftliche
Beiträge der FSU Jena, pp. 23–52. Friedrich-Schiller-Universität, Jena (1983)
57. Hermann, M.: Numerik gewöhnlicher Differentialgleichungen - Anfangs- und Randwertprob-
leme. Oldenbourg Verlag, München und Wien (2004)
58. Hermann, M.: Numerische Mathematik, 3rd edn. Oldenbourg Verlag, München (2011)
59. Hermann, M., Kaiser, D.: RWPM: a software package of shooting methods for nonlinear
two-point boundary value problems. Appl. Numer. Math. 13, 103–108 (1993)
60. Hermann, M., Kaiser, D.: Shooting methods for two-point BVPs with partially separated
endconditions. ZAMM 75, 651–668 (1995)
61. Hermann, M., Kaiser, D.: Numerical methods for parametrized two-point boundary value
problems—a survey. In: Alt, W., Hermann, M. (eds.) Berichte des IZWR, vol. Math/Inf/06/03,
pp. 23–38. Friedrich-Schiller-Universität Jena, Jenaer Schriften zur Mathematik und Infor-
matik (2003)
62. Hermann, M., Kaiser, D., Schröder, M.: Bifurcation analysis of a class of parametrized two-
point boundary value problems. J. Nonlinear Sci. 10, 507–531 (2000)
63. Hermann, M., Saravi, M.: A First Course in Ordinary Differential Equations: Analytical and
Numerical Methods. Springer (2014)
64. Hermann, M., Ullmann, T., Ullrich, K.: The nonlinear buckling problem of a spherical shell:
bifurcation phenomena in a BVP with a regular singularity. Technische Mechanik 12, 177–184
(1991)
65. Hermann, M., Ullrich, K.: RWPKV: a software package for continuation and bifurcation
problems in two-point boundary value problems. Appl. Math. Lett. 5, 57–62 (1992)
66. Hussels, H.G.: Schrittweitensteuerung bei der Integration gewöhnlicher Differentialgleichun-
gen mit Extrapolationsverfahren. Master’s thesis, Universität Köln (1973)
67. Jordan, D.W., Smith, P.: Nonlinear Ordinary Differential Equations: An Introduction for Scien-
tists and Engineers. Oxford Texts in Applied and Engineering Mathematics. Oxford University
Press Inc., New York (2007)
304 References
68. Keener, J.P., Keller, H.B.: Perturbed bifurcation theory. Arch. Rat. Mech. Anal. 50, 159–175
(1973)
69. Keller, H.B.: Shooting and embedding for two-point boundary value problems. J. Math. Anal.
Appl. 36, 598–610 (1971)
70. Keller, H.B.: Lectures on Numerical Methods in Bifurcation Problems. Springer, Heidelberg,
New York (1987)
71. Khuri, S.A.: A new approach to Bratu’s problem. Appl. Math. Comput. 147(1), 131–136
(2004)
72. Lange, C.G., Kriegsmann, G.A.: The axisymmetric branching behavior of complete spherical
shells. Q. Appl. Math. 2, 145–178 (1981)
73. Lesnic, D.: The decomposition method for forward and backward time-dependent problems.
J. Comp. Appl. Math. 147(1), 27–39 (2002)
74. Levi-Cività, T.: Détermination rigoureuse des ondes permanentes d’ampleur finie. Math. Ann.
93, 264–314 (1925)
75. Liao, S.: The proposed homotopy analysis technique for the solution of nonlinear problems.
Ph.D. thesis, Shanghai Jiao Tong University (1992)
76. Liao, S.: An explicit, totally analytic approximation of Blasius’ viscous flow problems. Int.
J. Non-Linear Mech. 34(4), 759–778 (1999)
77. Liao, S.: Beyond Perturbation: Introduction to the Homotopy Analysis Method. Chapman &
Hall/CRC Press, Boca Raton (2003)
78. Liao, S.: Notes on the homotopy analysis method: some definitions and theorems. Commun.
Nonlinear Sci. Numer. Simul. 14, 983–997 (2009)
79. Liao, S.: Advances in the Homotopy Analysis Method. World Scientific Publishing Co. Pte.
Ltd., New Jersey et al. (2014)
80. Soliman, M.A.: Rational approximation for the one-dimensional Bratu equation. Int. J. Eng.
Technol. 13(5), 54–61 (2013)
81. Merker, A.: Numerical bifurcation analysis of the bipedal spring-mass model. Ph.D. thesis,
Friedrich-Schiller-Universität Jena (2014)
82. Merker, A., Kaiser, D., Hermann, M.: Numerical bifurcation analysis of the bipedal spring-
mass model. Physica D: Nonlinear Phenomena 291, 21–30 (2014)
83. Merker, A., Kaiser, D., Seyfarth, A., Hermann, M.: Stable running with asymmetric legs: a
bifurcation approach. Int. J. Bifurcat. Chaos 25(11), 1–13 (2015)
84. Merker, A., Rummel, J., Seyfarth, A.: Stable walking with asymmetric legs. Bioinspiration
Biomimetics 6(4), 045,004 (2011)
85. Middelmann, W.: Konstruktion erweiterter Systeme für Bifurkationsphänomene mit Anwen-
dung auf Randwertprobleme. Ph.D. thesis, Friedrich Schiller Universität Jena (1998)
86. Mistry, P.R., Pradhan, V.H.: Exact solutions of non-linear equations by variational iteration
method. Int. J. Appl. Math. Mech. 10(10), 1–8 (2014)
87. Mohsen, A.: A simple solution of the Bratu problem. Comput. Math. Appl. 67, 26–33 (2014)
88. Moore, G.: The numerical treatment of non-trivial bifurcation points. Numer. Funct. Anal.
Optim. 2, 441–472 (1980)
89. Nayfeh, A.H.: Introduction to Perturbation Techniques. Wiley, New York (1981)
90. Nekrassov, A.I.: Über Wellen vom permanenten Typ i. Polyt. Inst. I. Wosnenski, pp. 52–65
(1921) (in russian language)
91. Nekrassov, A.I.: Über Wellen vom permanenten Typ ii. Polyt. Inst. I. Wosnenski, pp. 155–171
(1922)
92. Nierenberg, L.: Topics in Nonlinear Fuctional Analysis. Courant Lecture Notes in Mathemat-
ics 6. American Mathematical Society, Providence, Rhode Island (2001)
93. Ortega, J.M., Rheinboldt, W.C.: Iterative Solution of Nonlinear Equations in Several Variables.
Academic Press, New York (1970)
94. Pashaei, H., Ganji, D.D., Akbarzade, M.: Application of the energy balance method for
strongly nonlinear oscillators. Prog. Electromagnet. Res. M 2, 47–56 (2008)
95. Poincaré, H.: Sur le problème des trois corps et les équations de la dynamique. Acta Mathe-
matica 13(1), A3–A270 (1890)
References 305
96. Polyanin, A.D., Zaitsev, V.Z.: Handbook of Exact Solutions for Ordinary Differential Equa-
tions, 2nd edn. Chapman & Hall/CRC, Boca Raton, Florida (2002)
97. Poston, T., Stewart, I.: Catastrophe Theory and Its Applications. Pitman, San Francisco (1978)
98. Prince, F.J., Dormand, J.R.: High order embedded Runge-Kutta formulas. J. Comput. Appl.
Math. 7, 67–75 (1981)
99. Rafei, M., Ganji, D.D., Daniali, H., Pashaei, H.: The variational iteration method for nonlinear
oscillators with discontinuities. J. Sound Vib. 305, 614–620 (2007)
100. Rashidinia, J., Maleknejad, K., Taheri, N.: Sinc-Galerkin method for numerical solution of
the Bratu’s problems. Numer. Algorithms 62, 1–11 (2012)
101. Rheinboldt, W.C.: Methods for Solving Systems of Nonlinear Equations. SIAM, Philadelphia
(1998)
102. Roberts, S.M., Shipman, J.S.: Continuation in shooting methods for two-point boundary value
problems. J. Math. Anal. Appl. pp. 45–58 (1967)
103. Ronto, M., Samoilenko, A.M.: Numerical-Analytic Methods in the Theory of Boundary-Value
Problems. World Scientific Publishing Co., Inc., Singapore (2000)
104. Roose, D., Piessens, R.: Numerical computation of turning points and cusps. Numer. Math.
46, 189–211 (1985)
105. Rummel, J., Blum, Y., Seyfarth, A.: Robust and efficient walking with spring-like legs. Bioin-
spiration Biomimetics 5(4), 046,004 (13 pp.) (2010)
106. Sachdev, P.L.: Nonlinear Ordinary Differential Equations and Their Applications. No. 142 in
Chapman & Hall Pure and Applied Mathematics. CRC Press (1990)
107. Saravi, M.: Pseudo-first integral method for Benjamin-Bona-Mahony, Gardner and foam
drainage equations. J. Basic. Appl. Sci. Res. 7(1), 521–526 (2011)
108. Saravi, M., Hermann, M., Khah, E.H.: The comparison of homotopy perturbation method
with finite difference method for determination of maximum beam deflection. J. Theor. Appl.
Phys. (2013). doi:10.1186/2251-7235-7-8
109. Scheidl, R.: On the axisymmetric buckling of thin spherical shells. Int. Ser. Numer. Math. 70,
441–451 (1984)
110. Schwetlick, H.: Numerische Lösung nichtlinearer Gleichungen. VEB Deutscher Verlag der
Wissenschaften, Berlin (1979)
111. Scott, M.R., Watts, H.A.: A systemalized collection of codes for solving two-point boundary
value problems. In: Aziz, A.K. (ed.) Numerical Methods for Differential Systems, pp. 197–
227. Academic Press, New York and London (1976)
112. Seydel, R.: Numerical computation of branch points in nonlinear equations. Numer. Math.
33, 339–352 (1979)
113. Shampine, L.F., Baca, C.: Fixed vs. variable order Runge-Kutta. Technical report 84-1410,
Sandia National Laboratories, Albuquerque (1984)
114. Shearer, M.: One-parameter perturbations of bifurcation from a simple eigenvalue. Math.
Proc. Camb. Phil. Soc. 88, 111–123 (1980)
115. Skeel, R.D.: Iterative refinement implies numerical stability for Gaussian elimination. Math.
Comput. 35, 817–832 (1980)
116. Spence, A., Werner, B.: Non-simple turning points and cusps. IMA J. Numer. Anal. 2, 413–427
(1982)
117. Stoer, J., Bulirsch, R.: Introduction to Numerical Analysis. Springer, New York, Berlin, Hei-
delberg (2002)
118. Su, W.P., Wu, B.S., Lim, C.W.: Approximate analytical solutions for oscillation of a mass
attached to a streched elastic wire. J. Sound Vib. 300, 1042–1047 (2007)
119. Teschl, G.: Ordinary Differential Equations and Dynamical Systems, Graduate Studies in
Mathematics, vol. 140. AMS, Providence, RI (2012)
120. Troesch, B.A.: A simple approach to a sensitive two-point boundary value problem. J. Comput.
Phys. 21, 279–290 (1976)
121. Troger, H., Steindl, A.: Nonlinear Stability and Bifurcation Theory. Springer, Wien, New York
(1991)
306 References
122. Vazquez-Leal, H., Khan, Y., Fernandez-Anaya, G., Herrera-May, A., Sarmiento-Reyes, A.,
Filobello-Nino, U., Jimenez-Fernandez, V.M., Pereyra-Diaz, D.: A general solution for
Troesch’s problem. Mathematical Problems in Engineering (2012). doi:10.1155/2012/208375
123. Wallisch, W., Hermann, M.: Numerische Behandlung von Fortsetzungs- und Bifurkation-
sproblemen bei Randwertaufgaben. Teubner-Texte zur Mathematik, Bd. 102. Teubner Verlag,
Leipzig (1987)
124. Wazwaz, A.M.: A reliable modification of Adomian decomposition method. Appl. Math.
Comput. 102(1), 77–86 (1999)
125. Wazwaz, A.M.: Adomian decomposition method for a reliable treatment of the Emden-Fowler
equation. Appl. Math. Comput. 61, 543–560 (2005)
126. Wazwaz, A.M.: Partial Differential Equations and Soltary Waves Theory. Springer Publisher
(2009)
127. Wazwaz, A.M.: The variational iteration method for analytic treatment for linear and nonlinear
ODEs. Appl. Math. Comput. 212(1), 120–134 (2009)
128. Wazwaz, A.M.: A reliable study for extensions of the Bratu problem with boundary conditions.
Math. Meth. Appl. Sci. 35(7), 845–856 (2012)
129. Weber, H.: Numerische Behandlung von Verzweigungsproblemen bei Randwertaufgaben
gewöhnlicher Differentialgleichungen. Ph.D. thesis, Gutenberg-Universität Mainz (1978)
130. Weber, H.: Numerische Behandlung von Verzweigungsproblemen bei gewöhnlichen Differ-
entialgleichungen. Numer. Math. 32, 17–29 (1979)
131. Weber, H.: On the numerical approximation of secondary bifurcation problems. In: Allgower,
K.G.E.L., Peitgen, H.O. (eds.) Numerical Solution of Nonlinear Equations, pp. 407–425.
Springer, Berlin, Heidelberg, New York (1980)
132. Werner, D.: Funktionalanalysis. Springer, Heidelberg et al (2011)
133. Zeeman, E.C.: Catastrophe Theory. Addison-Wesley Publishing Co. Inc., Reading (1977)
134. Zeidler, E.: Nonlinear Functional Analysis and Its Applications I, Fixed Point Theorems.
Berlin, et al, Springer, New York (1986)
135. Zhang, H.L.: Periodic solutions for some strongly nonlinear oscillations by He’s energy bal-
ance method. Comput. Math. Appl. 58(11–12), 2480–2485 (2009)
Index
A Boundary conditions
Abel’s equation, 48 completely separated, 150
Adomian decomposition method (ADM), partially separated, 131, 146
44, 93 Bow tie, 298
Adomian polynomial, 46 BP-problem, see Bifurcation preserving
Arclength continuation, 273 problem
Bratu’s equation, 24
Bratu’s problem, 107, 162, 183
B first extension, 120
Ballistic trajectory, 130 Bratu-Gelfand equation, 20
BD-problem, see Bifurcation destroying Buckling, 122
problem Burgers’ equation, 28
Beam, 65 BVP, see Two-point boundary value problem
Bending, 65
Bernoulli equation, 9, 62
Bifurcation C
primary, 194 Cannon, 130
secondary, 194 Cannoneer, 130
Bifurcation coefficient, 227 Canonical form, 187
first, 196 Cauchy-Euler equation, 22
second, 179 Characteristic coefficient, 176, 195
third, 179 Clairaut equation, 15
Bifurcation destroying problem, 249 Collocation method, 72, 123
Bifurcation diagram, 171 Compactification, 139
Bifurcation point, 171 Complementary functions, 134
asymmetric, 199, 224, 255 Continuation method, 124
primary simple, 172 Contour, 296
secondary multiple, 172 Correction functional, 34
secondary simple, 172 Corrector iteration, 270
simple, 196 Cusp catastrophe, 187
symmetric, 199, 218, 260 Cusp curve, 187
Bifurcation point equation, 196 Cusp point, 187
Bifurcation preserving problem, 249
Block elimination technique, 150
Block Gaussian elimination, 139 D
Boomerang, 297 Damping strategy, 142
Bordered matrix, 139 Dance with a ribbon, 296
DOI 10.1007/978-81-322-2812-7
308 Index
Deformation, 94 Homotopy method, 124

Deformation equation of order n, 96 Hydrodynamic stability, 9
Deformation equation of order 0, 93 Hydrodynamics, 18
Dual Banach space, 173 Hyperbolic point, 232
Dual operator, 173
Duffing equation, 69, 74
I
Imperfection
E degenerate, 249
EBM, see Energy balance method nondegenerate, 249
Eigensolutions, 167 Implicit Function Theorem, 172
Eigenvalue Index of an operator, 171
geometrically simple, 195 Initial value problem, 122, 227
Eigenvalues, 167 Integrating factor, 7
Eight curve, 296 Iron, 297
Embedding method, 124 Isolated solution, 123
Emden-Fowler equation, 56, 59 Iterative refinement, 135, 140
Emden-Lane-Fowler equation, 24, 40, 56 Ivey’s equation, 23
Energy balance method, 71 IVP, see Initial value problem
Equation IVP-solver, 123
exact, 5
homogeneous, 4
reducible, 9 J
separable, 4 Janus, 299
Euler’s formula, 167
Euler–Bernoulli rod, 165, 247
Euler-Bernoulli rod, 205, 218, 250 K
Evolution problem, 122 Killing of the nonlinearity, 124
Extension technique, 179, 188, 201, 237
L
F Lagrange equation, 15
Family of solutions, 6 Lagrange multiplier, 34
Fill-in, 140 Least squares method, 72
Finite difference method, 123 Liapunov–Schmidt reduction, 174
First-order approximation, 72 Limacon of Pascal, 296
Foam drainage equation, 26 Limit point equation, 176
Fréchet derivative, 170 Limit points, 171
Fredholm alternative, 174, 202 Lipschitz condition, 3
Fredholm operator, 171 Lipschitz constant, 3
Frequency, 72 Local parametrization, 275
Loop, 297
LU factorization, 140
G
Givens transformation, 185
Gronwall’s Lemma, 128 M
Manneken Pis method, 130
Map
H bijective, 175
HAM, see Homotopy analysis method injective, 175
Hamiltonian, 71, 87 surjective, 175
Hamiltonian approach, 87 Matlab, 140, 296
Hanging cable problem, 17 Matlab programs, 155
Homotopy analysis method, 92 Matrix
Index 309
adjoint, 185 R
Meshgrid, 296 Radius of convergence, 128, 142
Method of complementary functions, 131 Raindrop, 297
Newton form, 133 Range, 171
standard form, 134 Recursion, 140
Method of Levi-Cività, 204, 232 Regularization strategy, 142
Method of Nekrassov, 218 Relation
Molar, 298 constitutive, 165
Mordillo U, 298 equilibrium, 166
Morse Lemma, 196 geometric, 165
Multiple shooting code, 230 Residual, 71
Multiple shooting method, 135, 137 Riccati equation, 9, 11, 22
Multipoint boundary value problem, 122 Riesz–Schauder theory, 195, 249
Ritz–Galerkin method, 72
Rod, 122
N Rounding errors, 140
Nearly singular matrix, 142 Rule of coefficient ergodicity, 101
Noise terms phenomenon, 53 Rule of solution existence, 101
Non-exact equation, 7 Rule of solution expression, 100
Null space, 171 RWPKV, 291
RWPM, 140, 291
O
S
ODE, see Ordinary differential equation
Scaling, 140
Open Mapping Theorem, 175, 236 Shooting method, 123
Operator analytical, 114
multilinear, 227 Shooting points, 136
Ordinary differential equation, 121 Simple pendulum, 19
Oscillator equation, 39 Simple root, 123
Simple shooting method, 123, 281
Simple turning point, 112
P Solution
Packed storage, 140 explicit, 33
Parallel shooting method, 137 isolated, 170, 274
Parameter, 93 nonisolated, 252
Partial pivoting, 140 singular, 15
Path-following problem, 269 Solution branch, 177
Peano’s theorem, 2 Solution curve, 171
Period, 72 primary, 194
Perturbation method, 61 secondary, 194
Picard iteration, 3 simple, 177
Pitchfork bifurcation, 188 trivial, 194
Plane jet, 18 Solution field, 171, 249
Point Speech bubble, 298
singular, 281 Spherical shell, 288
Poisson-Boltzman’s equation, 24, 25 Stabilized March method
Population dynamics, 9 linear BVPs, 146
Predictor–corrector method, 269 Newtonian form, 152
nonlinear BVPs, 146
standard form, 154
Q Stable algorithm, 140
QR factorization, 132, 147, 185 Static problem, 122
Quasi-Newton method, 142 Step length, 270
310 Index
T Two-parameter problem, 186

Test function, 280, 282 Two-point boundary conditions
Theorem of Newton and Kantorovich, 127 linear, 122
Theorem of Picard and Lindelöf, 3 nonlinear, 122
Thomas-Fermi’s equation, 24, 26 Two-point boundary value problem, 122,
Total differential, 5 169
Trail-and-error technique, 124 adjoint, 185, 226
Transformation technique, 190, 216, 221
Trial function, 72
Trifolium, 295 V
Troesch’s problem, 128, 135, 144 van der Pol oscillator, 63
Turning point Variational, 71, 87
double, 178, 186 Variational iteration method, 33
multiple, 178 Variational problem, 126
simple, 163, 172, 178, 275, 284 VIM, 93, see also Variational iteration
Turning points, 171 method

Martin Hermann, Masoud Saravi - Nonlinear Ordinary Differential Equations - Analytical Approximation and Numerical Methods-Springer (2016) PDF

Uploaded by

Copyright:

Available Formats

Martin Hermann, Masoud Saravi - Nonlinear Ordinary Differential Equations - Analytical Approximation and Numerical Methods-Springer (2016) PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Martin Hermann, Masoud Saravi - Nonlinear Ordinary Differential Equations - Analytical Approximation and Numerical Methods-Springer (2016) PDF

Uploaded by

Copyright:

Available Formats

Martin Hermann · Masoud Saravi

ISBN 978-81-322-2810-3 ISBN 978-81-322-2812-7 (eBook)

Library of Congress Control Number: 2016937388

© Springer India 2016

Printed on acid-free paper

This Springer imprint is published by Springer Nature

Nonlinear ordinary differential equations (ODEs) play a fundamental role in sci-

conditions. It is precisely this focus on analytical as well as numerical methods that

can perform his own numerical experiments. Chapter 5 is devoted to nonlinear

1 A Brief Review of Elementary Analytical Methods for Solving

3.4 Homotopy Analysis Method . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

5.7 Path-Following Methods for Simple Solution Curves . . . . . . . . . . 269

Martin Hermann is professor of numerical mathematics at the Friedrich Schiller

Nonlinear ordinary differential equations (ODEs) are encountered in various fields

1.2 Analytical Solution of First-Order Nonlinear ODEs

The general form of a first-order ODE is

© Springer India 2016 1

Moreover, if (1.1) can be written as y = g(x, y), then we search a family of

y = f (x, y), y(x0 ) = y0 . (1.3)

has no real solution. The IVP

has one solution y(x) = e2x and the IVP

R(a, b) ≡ {(x, y) : |x − x 0 | < a, |y − y0 | < b}.

Proof See, e.g., the monograph [46].

Theorem 1.2 (Uniqueness theorem of Picard and Lindelöf)

Proof See, e.g., the monograph [46].

That is, the sequence y0 , y1 , . . . converges to y(x).

Proof See, e.g., the monograph [46].

| f (x, y1 ) − f (x, y2 )| ≤ K |y1 − y2 |, (x, y1 ) , (x, y2 ) ∈ T. (1.5)

Remark 1.5 We can conclude that if ∂ f /∂ y is continuous in T , then f (x, y) satisfies

The following result can be easily proved.

As we pointed out it may be difficult to find closed form solutions whenever we

This equation can be solved in the following cases.

Then, for simplicity one can write

Example 1.7 Solve the IVP (y + 1)e x y + x y 2 = 0, y(0) = 1.

Solution. The given equation can be written as

Using y(0) = 1, we obtain c = 0. Hence y + ln(y) = (x + 1)e−x .

is zero, the substitution y = x z and y = z + x z leads to

Since z = y/x, we have

Proof See e.g., [46].

In the case of an exact equation, it holds z x = M and z y = N . Define z(x, y) by

From this we have

and this is equal to N (x, y). Therefore

For an exact equation, the last expression is zero.

Solution. The variables are not separable. We have

M(x, y) = cos(2x) − 2y sin(2x) − 2x y 3 , N (x, y) = cos(2x) − 3x 2 y 2 .

Then M y = N x = −2 sin(2x) − 6x y 2 . Hence the equation is exact. We have z x =

Now, since z y = N (x, y) = cos(2x) − 3x 2 y 2 , it is

g (y) = 0, i.e. g(y) = c.

Therefore, the family of solutions will be

(iv) Non-exact equations

F(x, y)M(x, y) + F(x, y)N (x, y) = 0

will be exact. Therefore (F M) y = (F N )x , and we have

We can rewrite this formula as

On the other hand, if we have F = F(y), then Fx = 0, and it holds

Thus, we can say that F = F(x), when (M y − N x )/N is a function of x, and

2 tan(y) sin(2x)d x + (sin(y) − cos(2x))dy = 0.

= exp (ln(cos(y)) = cos(y).

Multiplying the differential equation by cos(y), we get

2 sin(y) sin(2x)d x + cos(y)(sin(y) − cos(2x))dy = 0.