Вы находитесь на странице: 1из 97

Singular Perturbation Theory

Lindsay A. Skinner

Singular Perturbation Theory


Lindsay A. Skinner
Department of Mathematical Sciences
University of Wisconsin - Milwaukee
Milwaukee, Wisconsin 53201
USA
skinner4@cox.net

ISBN 978-1-4419-9957-3 e-ISBN 978-1-4419-9958-0


DOI 10.1007/978-1-4419-9958-0
Springer New York Dordrecht Heidelberg London
Library of Congress Control Number: 2011928077

Mathematical Subject Classification 2010: 34E05, 34E10, 34E15, 34E20

Springer Science+Business Media, LLC 2011


All rights reserved. This work may not be translated or copied in whole or in part without the written
permission of the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York,
NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in
connection with any form of information storage and retrieval, electronic adaptation, computer software,
or by similar or dissimilar methodology now known or hereafter developed is forbidden.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if they
are not identified as such, is not to be taken as an expression of opinion as to whether or not they are
subject to proprietary rights.

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)


Preface

There are many ne books on singular perturbation theory and how to solve
singular perturbation problems. Many readers of this book will already be
familiar with one or more of them. What distinguishes this book from the
others is its rigorous development and rigorous application of the method of
matched asymptotic expansions. The point of view is that certain functions
have a certain structure for which this method is valid and these are precisely
the kinds of functions that arise in a wide variety of dierential equation and
integration problems.
This book is intended to serve primarily as a supplement or follow-up to
a typical rst year graduate course in asymptotic and perturbation analysis.
Hopefully it will also prove to be a valuable companion to all those who do,
or wish to do, rigorous work in the eld. The basic theory for the book is
presented in Chapter 1. Then there are four chapters in which this theory is
applied to a sequence of ordinary dierential equation problems. There are
a number of previously unpublished results. One of these is the uniformly
valid expansion at the end of Chapter 4 for a problem involving logarithms
once studied by L. E. Fraenkel. Another is the unexpectedly simple uniformly
valid Bessel function expansion established at the end of Chapter 3. All the
dierential equations chosen for study in the text are linear, but this is not
because of any limitation of the theory. Indeed, much has been done, and
much more can be done, in applying the theory to nonlinear problems, as
noted in Exercise 3.2, for example.
Another unique feature of this book is its inclusion of several Maple pro-
grams for computing the terms of the various asymptotic expansions that
arise in solving the problems. Developing these was a nice break from the
otherwise sometimes tedious hard analysis in the book. They could well have
just been designated as exercises, indeed a few have been, but at the same
time, until one gets familiar with symbolic programming, there is nothing like
having a few model programs around to show the way, or at least a possible
way. Studying the programs will also reveal some mathematical manipula-
tions that are not fully developed within the text itself.

v
vi Preface

There is a short bibliography of key references at the end of the book.


My proof that the solution to Problem C is uniformly of order one, and the
use of this fact to complete the solution to Problem C, is essentially due to
R. E. OMalley, Jr., as described in Section 5.4 of [12], the book by D. R.
Smith, although a dierent approach to this problem, one comparable to mine
for Problems D, E and I, is used by OMalley in [7]. Also, the inclusion of
Problem D, which serves as a nice transition from Problem C to Problem E,
comes about as a result of my having rst seen it discussed in [12]. Problem
G is the one previously studied by L. E. Fraenkel. From his work in Part
II of [3] I saw how to prove my asymptotic results for this problem by rst
establishing a convergent series solution. Similarly, I rst became aware of
the intriguing Problem I by reading J. Kevorkians account of it in the 1981
version of [5]. My own work on it began in [11]. My analysis for Problem B
is an outgrowth of results presented in [8].
All of this started for me with the publication of the rst edition of [13], the
wonderful little book by Milton Van Dyke, in which the idea that matching
was somehow just a counting game and not an idea dependent on notions of
overlapping domains of validity rst became apparent. L. E. Fraenkels three
paper series, which came soon after, was the rst to try to show that the
reason this counting game works is a consequence of function structure. The
present book is an outgrowth of my attempts to clarify this connection and
to build on it.
A recurring theme in the book is that for each problem we rst establish
the form of a uniformly valid asymptotic expansion for its solution. Then,
knowing this form, we are able to proceed to calculate the appropriate inner
and outer expansions from the dierential equation for the problem, and
thus determine the terms of the uniformly valid expansion. Of course, when
a problem is new, one is likely to proceed dierently. I discovered the form
of the uniformly valid asymptotic expansion for the solution to Problem G,
for example, by rst calculating a few terms of its inner expansion, just a
power series if the initial conditions are treated as parameters, and then
examining the outer expansions of these terms. Once I had this uniformly
valid form, then from it the form of the outer expansion of the solution,
together with the required matching conditions for it, was readily determined.
There was no need for any kind of special matching considerations due to
the logarithms. In determing uniformly valid expansions for the solutions to
Problem H and Problem I, an important bit of insight, also a key step in [9],
was to factor out the oscillatory parts, as they do not have matching inner
and outer expansions.
At the end of each chapter there are exercises to do. Some of them demon-
strate results in the chapter, some ll in missing steps in the chapter, and
some are included to prepare for the next chapter. Of course, everyone is
encouraged to work all the exercises and, indeed, to add their own. Readers
interested in just a basic understanding, however, need only read Chapter 1,
possibly work Exercises 1.2 and 1.3, then read Section 2.1 and work Exercise
Preface vii

2.1. Finally they should read Section 3.1, maybe work Exercise 3.1, and get
some experience with the Maple program for Problem C.

Lindsay A. Skinner
San Diego, California
Contents

1 Uniform Expansion Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1


1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Uniform Expansion Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Some Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 The Proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5 Computer Calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.6 Two Corollaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2 First Order Dierential Equations . . . . . . . . . . . . . . . . . . . . . . . . 15


2.1 Problem A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Problem B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3 Second Order Dierential Equations . . . . . . . . . . . . . . . . . . . . . . 27


3.1 Problem C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 Problem D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.3 Problem E, Part 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.4 Problem E, Part 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

4 Logarithm Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2 Problem F . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.3 Problem G, Part 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.4 Problem G, Part 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.5 Problem G, Part 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

ix
x Contents

5 Oscillation Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.1 Problem H . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.2 Problem I, Part 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.3 Problem I, Part 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
Chapter 1
Uniform Expansion Theory

1.1 Introduction

Roughly speaking, a function z(x, ) is a singular perturbation of z(x, 0)


if z(x, 0) fails to approximate z(x, ) for all x of interest when is small.
Uniformly valid approximations for such functions can often be found by
the so-called method of matched asymptotic expansions. The purpose of this
book is to present a rigorous development of this method and its application
to integral and dierential equation problems.
Suppose, for example, that z(x, ) = 1/(x + 2x3 + ) and suppose we wish
to estimate  1
F () = z(x, ) dx (1.1)
0
for 0 < << 1. Away from x = 0, if is small enough, we can neglect it
compared to x and x3 , and approximate z(x, ) by z(x, 0) = 1/(x+2x3 ). Near
x = 0, on the other hand, x3 is negligible compared to x, so z(x, ) is approx-
imately 1/(x + ). Furthermore, away from x = 0, 1/(x + ) is approximately
1/x, and near x = 0, 1/(x + 2x3 ) is approximately 1/x. In a mouthful, the
near x = 0 approximation of the away from x = 0 approximation of z(x, )
matches the away from x = 0 approximation of the near x = 0 approximation
of z(x, ). An important consequence of this apparent coincidence is that the
composite function
1 1 1
c(x, ) = + (1.2)
x + x + 2x3 x
has the same approximations near x = 0 and away from x = 0 as z(x, ). In
fact we are about to prove the precise statement,

z(x, ) = c(x, ) + O() (1.3)

L.A. Skinner, Singular Perturbation Theory, DOI 10.1007/978-1-4419-9958-0_1, 1


Springer Science+Business Media, LLC 2011
2 1 Uniform Expansion Theory

10

0.0 0.2 0.4 0.6 0.8 1.0

Fig. 1.1 Graph of z(x, ) together with the approximations 1/(x + ), 1/(x + 2x3 ) and
c(x, ) when = 0.1.

uniformly as 0+ for 0 x 1. That is, there exists positive numbers o


and B such that |z(x, ) c(x, )| B for all (x, ) [0, 1] (0, o ].
From the integral of c(x, ), and the fact that ln(1 + ) = O() as 0+ ,
it follows from (1.3) that

F () = ln(1/) (1/2) ln 3 + O() (1.4)

as 0+ . Figure 1.1 shows the approximations 1/(x + ) and 1/(x + 2x2 ),


along with z(x, ) and c(x, ), for = 0.1.

1.2 Uniform Expansion Theorem

As is customary, we will say f (x) C N ([0, 1]) if f (x) is dierentiable up to


N times on the open interval (0, 1) and each derivative extends continuously
to [0, 1]. If f (x) C N ([0, 1]) for all N 0, we say f (x) C ([0, 1]). A basic
asymptotic result is that if f (x) C N ([0, 1]), then


N1
f (x) = f [n] (0)xn + O(xN ) (1.5)
n=0

1
 
d n
on the interval [0, 1], where f [n] (x) = n! dx f (x). Indeed, f (x) C N ([0, 1])
implies
1.2 Uniform Expansion Theorem 3

 
N 1 
(x) = xN f (x) f [n] (0)xn (1.6)
n=0

is in C 0 ([0, 1]), and therefore is bounded on [0, 1]. In particular, (0) =


f [N] (0). We will also say f (x) C N ([1, ]) if f(x) C N ([0, 1]), where
f(x) = f (1/x). When this is the case, (1.5) holds on [0, 1] with f in place of
f , and if we substitute x = 1/X, we get


N1
f (X) = f[n] (0)X n + O(X N ) (1.7)
n=0

on [1, ]. Moreover, f[1] (x) = (1/x2 )f [1] (1/x) = X 2 dX d


f (X) and there-
[n] [n] [n] 1
 2 d
n

fore f (0) = f (), where f (X) = n! X dX f (X). This nota-
tion readily extends to functions of more than one variable. In particular, for
m, n 0,
 m  n
[m,n] 1 2
f (x, X) = X f (x, X) (1.8)
m!n! x X

and we say f (x, X) C N ([0, 1] [0, ]) if both f (x, X) and f(x, X) are in
C N ([0, 1] [0, 1]), where f(x, X) = f (x, 1/X).
We are going to be dealing with functions y(x, ) that are, for some o > 0,
expressible in the form y(x, ) = f (x, x/) for all (x, ) [0, 1] (0, o ], where
f (x, X) C ([0, 1] [0, ]). If y(x, ) is one of these functions, then for any
(0, 1) and any N 1,

y(x, ) = ON y(x, ) + O(N ) (1.9)

uniformly as 0+ for x 1, where


N1
ON y(x, ) = (/x)n f [0,n] (x, ). (1.10)
n=0

Indeed, there exists BN , N > 0 such that




N 1


f (x, X) X n f [0,n] (x, )
BN X N (1.11)
n=0

whenever (x, X) [0, 1] [N , ], and therefore



N 1


f (x, x/) (/x)n f [0,n] (x, )
CN N (1.12)
n=0
4 1 Uniform Expansion Theory

for all (x, ) [, 1] (0, dN ], where CN = BN / N and dN = /N . The


function ON y(x, ) is called the N -term outer expansion of y(x, ). Of course,
xn f [0,n] (x, ) = y [0,n] (x, 0).
In addition to (1.9), for any (0, o ] and integer M 1,

y(x, ) = IM y(x, ) + O(xM ) (1.13)

as x 0+ , where


M 1
IM y(x, ) = xm f [m,0] (0, x/). (1.14)
m=0

We call IM y(x, ) the M -term inner expansion of y(x, ). Note that it is the
rst M terms in the power series expansion of Y (X, ) = y(X, ) as 0+ ,
assuming X > 0, followed by the substitution X = x/. Applying this idea
to ON y(x, ) leads to


M 1 
N1
IM ON y(x, ) = xm (/x)n f [m,n] (0, ). (1.15)
m=0 n=0

Similarly, the rst N terms in the expansion of IM y(x, ) as 0+ , assuming


x > 0, yields


N 1 
M 1
ON IM y(x, ) = (/x)n xm f [m,n] (0, ). (1.16)
n=0 m=0

Obviously, IM ON y(x, ) = ON IM y(x, ) for any M, N 1. Note also


that ON IM ON y(x, ) = IM ON y(x, ) and I M ON IM y(x, ) = ON IM y(x, ).
Therefore the N -term composite function CN y(x, ) = [ON +IN ON IN ]y(x, )
has the same M -term inner and outer expansions, for any M N , as y(x, ).
In addition, note that ON IN y(x, ) can be divided into a part which is nite
at x = 0 and a part which equals 0 at x = . The rst of these parts is


N 1 
N 1 
N1 
m
(/x)n xm f [m,n] (0, ) = m (x/)m (/x)n f [m,n] (0, ).
n=0 m=n m=0 n=0
(1.17)
Thus it is clear that

N1
CN y(x, ) = n [un (x) + vn (x/)], (1.18)
n=0

where
 
n1 
un (x) = xn f [0,n] (x, ) xm f [m,n] (0, ) , (1.19)
m=0
1.3 Some Calculations 5

 
m 
vm (X) = X m f [m,0] (0, X) X n f [m,n] (0, ) . (1.20)
n=0

Also, f (x, X) C ([0, 1] [0, ]) implies un (x) C ([0, 1]) and vm (X)
C ([0, ]). We can now state the theorem that is the basis for all our work
in this text.
Theorem 1. If for some o > 0, y(x, ) = f (x, x/) for all (x, ) [0, 1]
(0, o ], where f (x, X) C 2N ([0, 1] [0, ]), then


N1
y(x, ) = n [un (x) + vn (x/)] + O(N ) (1.21)
n=0

uniformly as 0+ for 0 x 1, where un (x) and vn (X) are the functions


dened by (1.19) and (1.20).

1.3 Some Calculations

Before turning to the proof of Theorem 1, let us see how it applies to the
problem of Section 1.1. If we let y(x, ) = z(x, ), then y(x, ) = f (x, x/),
where f (x, X) = 1/[1 + X(1 + 2x2 )], and obviously f (x, X) C ([0, 1]
[0, ]). From
/(x + 2x3 )
y(x, ) = (1.22)
1 + /(x + 2x3 )
for x = 0, it is apparent that


N1
n
ON y(x, ) = (1)n+1 . (1.23)
n=1
(x + 2x3 )n

In particular,

O1 y(x, ) = 0, O2 y(x, ) = . (1.24)
x + 2x3
Similarly,
1/(1 + X)
y(X, ) = , (1.25)
1 + 22 X 3 /(1 + X)
so we have
1
I1 y(x, ) = I2 y(x, ) = . (1.26)
1 + x/
Next, from either (1.24) or (1.26) we get

O1 I1 y(x, ) = 0, O2 I2 y(x, ) = /x. (1.27)


6 1 Uniform Expansion Theory

Therefore C1 y(x, ) = 1/(1 + x/), or u0 (x) = 0, v0 (X) = 1/(1 + X). Also



[C2 C1 ]y(x, ) = 3
, (1.28)
x + 2x x
and therefore u1 (x) = 2x/(1 + 2x2 ), v1 (X) = 0. Thus we see by Theorem 1
with N = 2 that
1 2x
y(x, ) = + O(2 ), (1.29)
1 + x/ 1 + 2x2

and this conrms (1.3).

1.4 The Proof

The validity of Theorem 1 is a direct consequence of the following more


fundamental result.
Lemma 1. If f (x, X) C 0 ([0, 1][0, ]), then f (x, x/) = (x, x/) + o(1),
where (x, X) = f (x, ) + f (0, X) f (0, ), uniformly as 0+ for
0 x 1.

Proof. Let > 0 be given. Then there exists > 0 such that |f (x, X)
f (0, X)| < whenever (x, X) [0, ] [0, ]. Therefore

|f (x, x/) f (0, x/)| < , |f (x, ) f (0, )| < (1.30)

for all (x, ) [0, ](0, ]. Similarly, there exists > 0 such that |f (x, X)
f (x, )| < whenever (x, X) [0, 1] [, ], and therefore

|f (x, x/) f (x, )| < , |f (0, x/) f (0, )| < (1.31)

for all (x, ) [, 1] (0, d], where d = /. Together, (1.30) and (1.31) show
that |f (x, x/) (x, x/)| < 2 whenever (x, ) [0, 1] (0, d].

Proof of Theorem 1. To prove Theorem 1, we will show that f (x, X)


C 2N ([0, 1] [0, ]) implies the slightly stronger statement


N
f (x, x/) = n [un (x) + vn (x/)] + o(N ) (1.32)
n=0

uniformly as 0+ for 0 x 1. By the lemma, we know this


is true for N = 0. Assume it is true for N = M > 0 and assume
f (x, X) C 2(M +1) ([0, 1] [0, ]). Then g(x, X) = X[f (x, X) f (x, )]
is in C 2M +1 ([0, 1] [0, ]), and hence h(x, X) C 2M ([0, 1] [0, ]), where
h(x, X) = x1 [g(x, X) g(0, X)]. Therefore
1.5 Computer Calculation 7


M
h(x, x/) = n [un (x) + vn (x/)] + o(M ) (1.33)
n=0

uniformly as 0+ for 0 x 1, where


 
n1 
un (x) = xn h[0,n] (x, ) xm h[m,n] (0, ) , (1.34)
m=0

 
m 
vm (X) = X m h[m,0] (0, X) X n h[m,n] (0, ) , (1.35)
n=0

in accordance with (1.19) and (1.20). But for > 0,

h(x, x/) = 1 [f (x, x/) (x, x/)], (1.36)

where (x, X) is the function dened in Lemma 1. Therefore, substituting


into (1.33),


M
f (x, x/) = (x, x/) + n [un (x) + vn (x/)] + o(M +1 ). (1.37)
n=0

Finally, note that g [0,n] (x, ) = f [0,n1] (x, ) and hence

h[0,n] (x, ) = x1 [f [0,n1] (x, ) f [0,n1] (0, )]. (1.38)

Similarly,

h[m,0] (0, X) = X[f [m+1,0] (0, X) f [m+1,0] (0, )]. (1.39)

Thus we see un (x) = un+1 (x) and vm (X) = vm+1 (X), and therefore (1.37)
is the same as (1.32) with N = M + 1.

1.5 Computer Calculation

Calculating more than a few terms in the uniform expansion (1.21) for most
any y(x, ) = f (x, x/) quickly becomes tedious work. One of our objectives
in this book is to encourage the use of computer algebra software to do cal-
culations like this for us. We will be doing this using the well-known product
called Maple. As our rst example, starting on a blank Maple worksheet, for
the function y(x, ) = /( + x + 2x3 ) of Section 1.3, by entering the sequence
of commands

y := ( + x + 2x3 )1 ; N := 4;
8 1 Uniform Expansion Theory

series(y, = 0, N );
ON y := convert(%, polynom);
series(subs(x = X, y), = 0, N );
IN y := convert(%, polynom);
series(subs(x = X, ON y), = 0, N );
IN ON y := convert(%, polynom);
series(subs(X = 1 x, IN y), = 0, N + 1);
ON IN y := convert(%, polynom);

we readily determine O4 y(x, ), I4 y(x, ) and two forms of O4 I4 y(x, ) =


I4 O4 y(x, ). Notice though that in the series command to determine ON IN y
with N = 4 we had to ask for N + 1 terms. This is because of the way series
deals with removable singularities. Other values of N may require making
other adjustments, as may other functions, y(x, ).
Continuing on the same worksheet, if we now enter

for n from 0 to N 1 do
expand(xn coe (ON IN y, , n));
series(%, x = 0, n);
coe (ON y, , n) xn convert(%, polynom);
un := simplify(%);
end do;

we get, in addition to u0 (x) = 0 and u1 (x) = 2x/(1 + 2x2 ) as in Section


1.3, the new results

4(1 + x2 ) 8x(3 + 8x2 + 6x4 )


u2 (x) = , u3 (x) = . (1.40)
(1 + 2x2 )2 (1 + 2x2 )3

The idea in this second step is to capture the sum in (1.19) from the coe-
cient of n in ON IN y(x, ) as it appears in (1.16). We used Maples expand
command to convert xn times this coecient into a polynomial and thus
avoid another removable singularity problem. By similar reasoning, looking
at (1.15) and the sum in (1.20), if we add
for m from 0 to N 1 do
expand(X m coe (IN ON y, , m));
series(%, X = , m + 1);
coe (IN y, , m) X m convert(%, polynom);
vm := simplify(%);
end do;

we immediately get

1 2(3X + 2)
v0 (X) = , v1 (X) = 0, v2 (X) = , v3 (X) = 0. (1.41)
1+X (1 + X)2
1.6 Two Corollaries 9

As for the original problem of approximating the function F () given by


(1.1), concluding this computation with

CN y := sum(k (uk + vk ), k = 0, N 1);


int(subs(X = 1 x, CN y), x = 0..1) assuming > 0;
F := series(1 %, = 0, N );

yields

F () = ln (1/) (1/2) ln 3 + a 62 ln (1/) + b2 + O(3 ), (1.42)



where a = 43 + 3 2 2 arctan 2 and b = 59
18
+ 3 ln 3. Note that we had to ask
for 4 terms in the series command for F () in order not to miss a possible
3 ln(1/) contribution.

1.6 Two Corollaries

Before turning to dierential equation problems we have a pair of corollaries


to add to Theorem 1. It often happens that y(x, ) = f (x, x/, ) for 0 x
1, 0 < o , where f (x, X, ) C ([0, 1] [0, ] [0, o ]). Then, of course,


N 1
f (x, X, ) = n f [0,0,n] (x, X, 0) + O(N ) (1.43)
n=0

uniformly as 0+ for all (x, X) [0, 1] [0, ], and therefore


N 1
y(x, ) = n f [0,0,n] (x, x/, 0) + O(N ) (1.44)
n=0

uniformly for 0 x 1. Furthermore, we can apply Theorem 1 to each


f [0,0,n] (x, x/, 0).
Corollary 1. If y(x, ) = f (x, x/, ) for all (x, ) [0, 1] (0, o ], where
f (x, X, ) C ([0, 1] [0, ] [0, o ]) and o > 0, then for any N 0,


N1
y(x, ) = n [un (x) + vn (x/)] + O(N ) (1.45)
n=0

uniformly for 0 x 1 as 0+ , where un (x) C ([0, 1]), vn (X)


C ([0, ]), vn () = 0, and, as in Theorem 1, the sum in (1.45) is the
combination CN y(x, ) = [ON + IN ON IN ]y(x, ) of the outer and inner
expansions
10 1 Uniform Expansion Theory


N 1
ON y(x, ) = n yn (x), (1.46)
n=0

where yn (x) = y [0,n] (x, 0), and


N 1
IN y(x, ) = n Yn (x/), (1.47)
n=0

where, Yn (X) = Y [0,n] (X, 0), Y (X, ) = y(X, ). Also, for any M, N 1,
ON IM y(x, ) = IM ON y(x, ).

Proof. From (1.44) and Theorem 1, it is easy to see (1.45) holds with


k  [m,n]
n1
[0,n]
uk (x) = xn gkn (x, ) gkn (0, )xm , (1.48)
n=0 m=0


k 
m
[m,0] [m,n]
vk (X) = X gkm (0, X)
m
gkm (0, )X n , (1.49)
m=0 n=0

where gk (x, X) = f [0,0,k] (x, X, 0), and it is clear that uk (x) C ([0, 1]),
vk (X) C ([0, ]) and vk () = 0. For the (independent) determination
of these functions by outer and inner expansions, we defer to Exercise 1.3 at
the end of the chapter.

It often happens too that we have y(x, ) = f (x, x/, ) and f (x, X, ) =
o(X ) as X . That is, f [0,n,0] (x, , ) = 0 for all n 0, which means
ON y(x, ) = 0 for all N 1. Also, additional parameters may be present and
of course we are not restricted to the interval 0 x 1. We will need the
following, for example, in which x is the additional parameter, in the next
chapter.
Corollary 2. If y(x, t, ) = f (x, t, t/, ) , where f (x, t, T, ) C ([0, xo ]
[0, x] [0, ] [0, o ]) for some o , xo > 0, and if f [0,0,n,0] (x, t, , ) = 0
for all n 0, then for any N 0,


N
y(x, t, ) = n [0,0,n] (x, t/, 0) + O(N +1 ) (1.50)
n=0

uniformly for all (x, t) [0, xo ] [0, x] as 0+ , where (x, T, ) =


f (x, T, T, ).
1.7 Exercises 11

1.7 Exercises

1.1. We noted in Section 1.2 that y(x, ) = f (x, x/), where f (x, X)
C ([0, 1] [0, ]), implies y(x, ) = ON y(x, ) + O(N ) uniformly on
[, 1] for any (0, 1). In light of Theorem 1 this also means [IN
ON IN ]y(x, ) = O(N ) on [, 1]. Show directly from(1.14) and (1.15) that
in fact [IN ON IN ]y(x, ) = O(N ) uniformly on [, c] for any c > > 0.

1.2. Show that Corollary 1 applies to

+ x2
y(x, ) = (1.51)
+ 2 + x + 2x3
and that

O3 I3 y(x, ) = x (1 + 2x 1/x) + (3 + 1/x 1/x2 )2 . (1.52)

Note that this implies

u0 (x) = y0 (x), u1 (x) = y1 (x) 1/x, u2 (x) = y2 (x) 1/x + 1/x2 , (1.53)

and

v0 (X) = Y0 (X), v1 (X) = Y1 (X)+1X, v2 (X) = Y2 (X)3+2X. (1.54)

Complete the calculation of un (x) and vn (X) for 0 n 2, and show that
 1
y(x, ) dx = a+ ln(1/)b+2 ln(1/)+c2 43 ln(1/)+O(3 ), (1.55)
0

where a = 14 ln 3, b = 1
6 + 1
2 ln 3 + 4
2 7
arctan 2, c = 18 1
2 ln 3 +

5 2

4
arctan 2.

1.3. Complete the proof of Corollary 1 by considering


N 1 
N 1
y1 (x, ) = n un (x), y2 (x, ) = n vn (x/) (1.56)
n=0 n=0

separately. Clearly, for example, IN y2 (x, ) = y2 (x, ). Therefore OM y2 (x, ) =


OM IN y2 (x, ). Also, IN OM IN y2 (x, ) = OM IN y2 (x, ), and so on.

1.4. Let m be a positive integer, assume a, b > 0 and suppose



2 2
F () = m+1
tm eat eb(tt ) dt. (1.57)
0
12 1 Uniform Expansion Theory

Show that for > 0,


 1
1
F () = f (s, s/) ds, (1.58)
0

where = 2 , f (s, S) = S m eSu(s) , u(s) = as + b(1 s). Since u(s) > 0 for
0 s 1, it follows, as in Corollary 2, that for any N 0,


N 1
f (s, s/) = n [0,n] (s/, 0) + O(N ) (1.59)
n=0

as 0+ , where (S, ) = f (S, S). Also,



[0,n] (s/, 0) ds = o( ). (1.60)
1

Show therefore that



N1
F () = cn 2n + O( 2N ) (1.61)
n=0

as , where
(m + 2n)! (a b)n
cn = (1)n . (1.62)
n! bm+2n
What if m is not an integer?

1.5. Consider the contour integral



F () = eh(z) g(z) dz, (1.63)
C

where C is the straight line from the origin to a point zo = ro eio and g(z),
h(z) are analytic on C. Assume Re[h(z)] > 0 on C, except h[k] (0) = 0 for
0 k m 1 and o is such that = Re[eimo h[m] (0)] > 0, where m 1.
Observe that these assumptions imply the real part of u(r) = r m h(rei0 )
exceeds zero for 0 r ro , and therefore
 ro
F () = f (r, r/) dr, (1.64)
0

where = 1/m and f (r, R) = g(reio )exp[Rm u(r) + io ], which is in


C ([0, ro ] [0, ]), so we can apply Theorem 1. Furthermore, as in Exercise
1.4, f [0,n] (r, ) = 0 for all n 0, and therefore


N
f (r, r/) = n [0,n] (r/, 0) + O(N+1 ), (1.65)
n=0
1.7 Exercises 13

where (R, ) = f (R, R). Also as in Exercise 1.4, [0,n] (R, 0) is bounded by
a polynomial times a decreasing exponential, in this case exp(Rm ), and
therefore 
[0,n] (r/, 0) dr = o( ). (1.66)
ro

Thus it follows from (1.64) that


N1 
F () = n [0,n] (R, 0) dR + O(N+1 ). (1.67)
n=0 0

m
In other words, if (Z, ) = g(Z)eh(Z)/ , then


N1  eio
F () = n
[0,n] (Z, 0) dZ + O(N +1 ), (1.68)
n=0 0

[m]
(0)Z m
and [0,n] (Z, 0) is a polynomial times eh .

1.6. Beginning with the Bessel function integral


 +i
1
J () = eh(z) dz, (1.69)
2i i

where h(z) = zsinhz, show that if xo > 0, then


 i
eh(z) dz = o( ) (1.70)
xo i

as , and that therefore J () is asymptotically equivalent to the sum of


two integrals of the form assumed in the previous exercise, one with o = /3,
the other with o = /3, and both with m = 3. Show further that these
two integrals can be combined to yield


N 1  ei/3
J () = n [0,n] (Z, 0) dZ + O(N +1 ), (1.71)
n=0 ei/3

where = 1/3 and (Z, ) = (1/2i)exp[3 h(Z)]. Note that (Z, ) is


an even function of , so half the terms in (1.71) equal zero. If we let
 ei/3
1 1 3
an = Z n e 6 Z dZ, (1.72)
2i ei/3

then a0 = 21/3 Ai(0), a1 = 22/3 Ai[1] (0), a2 = 0, and an = 2(n 2)an3


for n 3. Use the Maple steps
14 1 Uniform Expansion Theory

series(3 (sinh(Z) Z), = 0, 40) :


series(e% , = 0, 40) : convert(%, polynom) :
w := simplify(% exp( 16 Z 3 )) :
P := subs( = 1, %) : N := degree(P ) :
a0 := A : a1 := B : a2 := 0 : w := A 1 + w :
for n from 3 to N do
an := 2 (n 2) an3 : w := subs(Z n = an , w) :
end do: w := series(w, = 0, 40);

to determine the rst 13 non-zero terms of (1.71).

1.7. Assume a(x, ) C ([0, 1] [0, o ]). Show there exists b(x, )
C ([0, 1] [0, o ]) and c() C ([0, o ]) such that, for > 0,

a(x, ) c()
= b(x, ) + . (1.73)
x+ x+
Of course, this is trivial if a(x, ) is a polynomial.

1.8. An alternate proof of Theorem 1 can be obtained by a variation of the


analysis in Section 2.1 of Part II of [3]. Note rst that if 0 < 1/2 x 1,
then /x 1/2 and therefore, in view of (1.11), assuming y(x, ) satises
the hypotheses of Theorem 1.1, y(x, ) = O2N y(x, ) + O(N ). Similarly,
IN y(x, ) = O2N IN y(x, ) + O(N ) for 0 < 1/2 x 1. Therefore y(x, ) =
CN y(x, ) + AN y(x, ) + O(N ) for 0 < 1/2 x 1, where AN y(x, ) =
[(O2N ON ) IN (O2N ON )]y(x, ). But a short calculation shows
2N
 1
AN y(x, ) = xN (/x)n n (x), (1.74)
n=N

where
 
N1 
n (x) = xN f [0,n] (x, ) xm f [m,n] (0, ) , (1.75)
m=0

and clearly n (x) = O(1) for 0 x 1. Therefore AN y(x, ) = O(N )


for 0 < 1/2 x 1, since, in this case, xN (/x)n N for n N . An
analogous argument, beginning with y(x, ) = I2N y(x, ) + O(N ), shows
y(x, ) = CN y(x, ) + O(N ) as 0+ when 0 x 1/2 .
Chapter 2
First Order Dierential Equations

2.1 Problem A

The plan for this book is to apply the theory developed in Chapter 1 to a
sequence of more or less increasingly complex dierential equation problems.
We begin with
y  + a(x, )y = b(x, ), (2.1)
where a(x, ), b(x, ) C ([0, 1][0, o ]) for some o > 0, and a(x, ) > 0. We
also assume y(0, ) = () C ([0, o ]). The problem is to asymptotically
approximate y(x, ) uniformly on 0 x 1 as 0+ . If we put z(x, ) =
y(x, ) (), then
 x
1
z(x, ) = b(x t, )e[k(x,)k(xt,)]/ dt, (2.2)
0

where b(x, ) = b(x, ) ()a(x, ) and


 x
k(x, ) = a(t, ) dt. (2.3)
0

If we let u(x, t, ) = t1 [k(x, ) k(x t, )], then u(x, t, ) > 0 on [0, 1]


[0, x][0, o ]. In particular, u(x, 0, ) = a(x, ) > 0. Thus we can rewrite (2.2)
as  x
1
z(x, ) = f (x, t, t/, ) dt, (2.4)
0
where
f (x, t, T, ) = b(x t, )eT u(x,t,) , (2.5)
and we can apply Corollary 2. In terms of (x, T, ) = f (x, T, T, ), it follows
that
N
z(x, ) = n n (x, x/) + O(N ), (2.6)
n=0

L.A. Skinner, Singular Perturbation Theory, DOI 10.1007/978-1-4419-9958-0_2, 15


Springer Science+Business Media, LLC 2011
16 2 First Order Dierential Equations

where  X
n (x, X) = [0,0,n] (x, T, 0) dT. (2.7)
0
Furthermore,
[0,0,n] (x, T, 0) = pn (x, T )ea(x,0)T , (2.8)

where pn (x, T ) is a polynomial in T , and therefore n (x, X) C ([0, 1]
[0, ]). Hence, in accordance with Corollary 1, applied to the sum in (2.6),
we see y(x, ) = z(x, ) + () has a uniformly valid asymptotic expansion of
the form

N 1
y(x, ) = n [un (x) + vn (x/)] + O(N ), (2.9)
n=0

where un (x) C ([0, 1]), vn (X) C ([0, ]), vn () = 0, and these terms

can be found by computing inner and outer expansions.


From (2.6), we have O1 z(x, ) = 0 (x, ) and since p0 (x, T ) = b(x, 0), it
follows that
O1 z(x, ) = b(x, 0)/a(x, 0). (2.10)
Similarly,

I1 z(x, ) = 0 (0, x/) = [b(0, 0)/a(0, 0)](1 ea(0,0)x/ ) (2.11)

and, from either (2.10) or (2.11),

O1 I1 z(x, ) = I1 O1 z(x, ) = b(0, 0)/a(0, 0). (2.12)

Also C1 y(x, ) = C1 z(x, ) + (0). Hence, the rst terms of (2.9) are

u0 (x) = b(x, 0)/a(x, 0), v0 (X) = [(0) b(0, 0)/a(0, 0)]ea(0,0)X . (2.13)

Of course, it is much easier to calculate the inner and outer expansions


for (2.9) directly from the dierential equation (2.1). For the N -term outer
expansion,
N1
ON y(x, ) = n yn (x), (2.14)
n=0

(2.1) implies

n

yn1 (x) + ank (x)yk (x) = bn (x), (2.15)
k=0

where an (x) = a[0,n] (x, 0), bn (x) = b[0,n] (x, 0). Similarly, for the N -term inner
expansion,
N1
IN y(x, ) = n Yn (x/), (2.16)
n=0
2.1 Problem A 17

if we put A(X, ) = a(X, ), B(X, ) = b(X, ), in addition to Y (X, ) =


y(X, ), then (2.1) becomes

Y  + A(X, )Y = B(X, ). (2.17)

Therefore,

n
Yn (X) + Ank (X)Yk (X) = Bn (X), (2.18)
k=0

where An (X) = A[0,n] (X, 0), Bn (X) = B [0,n] (X, 0). Also, y(0, ) = () im-
plies Yn (0) = [n] (0). For the actual calculations we oer our rst Maple
program.

P robA := proc (a, b, , N )


ON y := sum(n yn , n = 0..N 1);
ON dy := sum(n dyn , n = 0..N 1);
ON eq := series( ON dy + a ON y b, = 0, N );
for k from 0 to N 1 do
temp := coe (ON eq, , k); yk := solve(temp = 0, yk );
dyk := di (yk, x); ON eq := subs(yk = yk, dyk = dyk, ON eq);
ON y := subs(yk = yk, ON y); print(uk = yk);
end do;
A := subs(x = X, a); B := subs(x = X, b);
IN y := sum(n Yn , n = 0..N 1);
IN dy := sum(n dYn , n = 0..N 1);
IN de := series(IN dy + A IN y B, = 0, N );
N := series(, = 0, N );
for k from 0 to N 1 do
temp := coe (IN de, , k);
de := subs(Yk = z(X), dYk = di (z(X), X), temp) = 0;
dsolve({de, z(0) = coe (N , , k)});
Y k := rhs(%); dY k := di (Y k, X);
IN de := subs(Yk = Y k, dYk = dY k, IN de); Yk := Y k;
end do;
IN ON y := series(subs(x = X, ON y), = 0, N );
for k from 0 to N 1 do
vk := Yk coe (IN ON y, , k);
print(vk = simplify(expand(vk)));
end do;
end proc:

This program solves for the N -term outer expansion of y(x, ) using (2.15)
and then computes the N -term inner expansion using (2.18). At the end, the
program computes IN ON y(x, ). Normally there would be the matter of sep-
arating IN ON y(x, ) into two parts to form the functions un (x) and vn (X).
18 2 First Order Dierential Equations

In this problem, however, it is clear from (2.15) that ON y(x, ) is continuous


at x = 0 and therefore un (x) = yn (x). This also means vn (x/) is the coe-
cient of n in [IN IN ON ]y(x, ). As an example,

a := 2 + x + ; b := (2 + x)1 ; := ; P robA(a, b, , 2);

yields
1 x
u0 (x) = , u1 (x) = , (2.19)
(2 + x)2 (2 + x)4
and
1 1
v0 (X) = e2X , v1 (X) = (8 + 2X + X 2 )e2X . (2.20)
4 8

2.2 Problem B

For our second dierential equation we take

2 y  + xa(x, )y = b(x, ). (2.21)

Again we assume a(x, ), b(x, ) C ([0, 1][0, o ]) for some o > 0. We also
assume a(x, ) > 0 on [0, 1][0, o ], y(0, ) = (), where () C ([0, o ]),
and, without loss of generality, we assume a(0, 0) = 1. The essential dierence
here from Problem A is the factor of x multiplying a(x, ).
In place of (2.3), we now have
 x
k(x, ) = ta(t, ) dt (2.22)
0

and therefore, as t 0+ ,
1
k(x t, 0) = k(x, 0) txa(x, 0) + t2 [a(x, 0) + xa[1,0] (x, 0)] + O(t3 ). (2.23)
2
If we put

u(x, t, ) = t2 [k(x, ) k(x t, )] (tx t2 )a(x, 0), (2.24)

then u(0, 0, 0) = 1/2 and therefore u(x, t, ) > 0 on [0, xo ] [0, x] [0, o ]
for some xo > 0, and possibly a smaller o > 0. Thus if we let z(x, ) =
y(x, ) () and b(x, ) = b(x, ) x()a(x, ), then
2.2 Problem B 19
 x
2
z(x, ) = 1 f (x, t, t/, )et(xt)a(x,0)/ dt, (2.25)
0

where,
2
f (x, t, T, ) = b(x t, )eT u(x,t,)
, (2.26)

and we have f (x, t, T, ) C ([0, xo ] [0, x] [0, ] [0, o ]). Also
2
0 < et(xt)a(x,0)/ 1 (2.27)

for all (x, t, ) [0, 1] [0, x] (0, o ]. Hence, applying Corollary 2 with
(x, T, ) = f (x, T, T, ), from (2.25) we get


N
z(x, ) = n n (x, x/) + O(N ) (2.28)
n=0

uniformly as 0+ for 0 x xo , where


 X
n (x, X) = [0,0,n] (x, T, 0)eT (XT )a(x,0) dT. (2.29)
0

Also,
2
[0,0,n] (x, T, 0) = pn (x, T )eT u(x,0,0)
, (2.30)
where pn (x, T ) is a polynomial in T .
From Exercise 1.4, it is clear that
 X
2
F (x, X) = T m eT u(x,0,0) T (XT )a(x,0)
e dT, (2.31)
0

for any integer m 0, is in C ([0, xo ] [0, ]). Therefore n (x, X)


C ([0, xo ] [0, ]) and thus from (2.28), by Corollary 1, we know y(x, )
has a uniformly valid expansion, at least for 0 x xo , of the form


N 1
y(x, ) = n [un (x) + vn (x/)] + O(N ). (2.32)
n=0

But also, from Problem A, we know that y(x, ) = ON y(x, ) + O(N ) uni-
formly as 0+ for xo x 1, since xa(x, ) > 0 for xo x 1. Further-
more, by Exercise 1.1, we know [IN ON IN ]y(x, ) = O(N ) for xo x 1.
Hence, in conclusion, y(x, ) has a uniformly valid expansion of the form
(2.32) as 0+ on the full interval 0 x 1, where un (x) C ([0, 1]),
vn (X) C ([0, ]), vn () = 0, and these functions can be determined by
computing the corresponding outer and inner expansions for y(x, ) directly
from the dierential equation (2.21).
20 2 First Order Dierential Equations

The inner expansion calculations for Problem B, if we just ask Maple to


solve the associated sequence of dierential equations, quickly gets bogged
down with iterated error function integrals, so we have to help. It turns out
that at each stage the equation to solve has the form
1 2
Y  + XY = (X) + (X)P (X) + (X)e 2 X , (2.33)

where (X), (X), (X) are polynomials and


 X
1 2 1 2
P (X) = e 2 X e 2 T dT. (2.34)
0

The solution to this equation has the same form as its right hand side. That
is,
1 2
Y (X) = p(X) + q(X)P (X) + r(X)e 2 X , (2.35)
where p(X), q(X), r(X) are polynomials. Indeed, if we substitute (2.35) into
(2.33), we nd

p + q + Xp = (X), q  = (X), r = (X). (2.36)

Therefore
 X  X
q(X) = q0 + (T ) dT, r(X) = r0 + (T ) dT, (2.37)
0 0

where q0 = q(0), r0 = r(0) have yet to be determined, and if d is the degree


of s(X) = (X) q(X), then, to satisfy (2.36a), the degree of p(X) must be
d 1. Hence,


d1 
d
p(X) = pn X n , s(X) = sn X n (2.38)
n=0 n=0

and (2.36a), together with pd = pd+1 = 0, implies

pdk1 = sdk (d k + 1)pdk+1 (2.39)

for 0 k d1. Also p1 = s0 so q0 = (0)p1 and nally, Y (0) = p0 +r0 de-


termines r0 . This is all incorporated into our Maple program for this problem.

P robB := proc (a, b, , N )


ON y := sum(n yn , n = 0..N 1);
ON dy := sum(n dyn , n = 0..N 1);
ON eq := series(2 ON dy + x a ON y b, = 0, N );
for k from 0 to N 1 do
temp := coe (ON eq, , k); yk := solve(temp = 0, yk );
dyk := di (yk, x); ON eq := subs(yk = yk, dyk = dyk, ON eq);
2.2 Problem B 21

ON y := subs(yk = yk, ON y); yk := yk;


end do;
A := subs(x = X, a); B := subs(x = X, b);
IN y := sum(n Yn , n = 0..N 1);
N := series( , = 0, N );
rths := series(B I X A IN y + X IN y, = 0, N + 1);
for k from 0 to N 1 do
temp := coe (rths, , k);
qcut := int(subs(X = T, coe (temp, P )), T = 0..X);
s := coe (temp, I) qcut;
if s = 0 then d := 0; else d := degree(s); end if ;
pd := 0; pd+1 := 0;
for j from 0 to d 1 do
pdj1 := coe (s, X, d j) (d j + 1) pdj+1 ;
end do;
q := qcut + subs(X = 0, coe (temp, I)) p1 ;
r := coe (N , , k) p0 + int(subs(X = T, coe (temp, E)), T = 0..X);
p := sum(pn X n , n = 0..d 1);
Y k := p I + q P + r E;
rths := subs(Yk = Y k, rths); Yk := Y k;
end do;
series(subs(x = X, ON y), = 0, N ); IN ON y := convert(%, polynom);
vpart := sum(X n coe (IN ON y, X, n), n = 0..N );
upart := subs(X = 1 x, IN ON y vpart);
for k from 0 to N 1 do
uk := yk coe (upart, , k);
print(uk = simplify(uk)); uk := uk;
end do;
for k from 0 to N 1 do
vk := Yk I coe (vpart, , k);
print(vk = simplify(coe (vk, I) I + coe (vk, P ) P + coe (vk, E) E);
end do;
end proc:

In this program, the letters I, P , and E are used to denote 1, P (X), and
exp( 12 X 2 ), respectively. Also, we have used the fact that IN ON y(x, ) is at
most O(X N1 ) as X to split it into the two parts necessary to form
un (x) and vn (X) for n = 0 to N 1.
As an example, if we set

a(x, ) = 1 + cx2 , b(x, ) = 1, () = 0, (2.40)

then P robB(a, b, , 4) yields


cx
y(x, ) = P (x/) + 2 v2 (x/) + 3 u3 (x) + O(4 ) (2.41)
1 + cx2
22 2 First Order Dierential Equations

with
c c2 x(3 + cx2 )
v2 (X) = [X + X 3 + (3 X 4 )P (X)], u3 (x) = . (2.42)
4 (1 + cx2 )3

As another, if

a(x, ) = 1 + gx, b(x, ) = h + kx, () = 0, (2.43)

then
k gh
u0 (x) = 0, u1 (x) = , v0 (X) = hP (X), (2.44)
1 + gx
and
1 1 1 1 2
v1 (X) = gh(1 + X 2 ) ghX 3 P (X) + (2gh 3k)e 2 X . (2.45)
3 3 3

For

a(x, ) = cos(x) + x, b(x, ) = cos(x), () = 1, (2.46)

a graph of
C2 y(x, ) = P (x/) + [u1 (x) + v1 (x/)] (2.47)
when = 0.2 is shown in Figure 2.1, along with a portion of O2 y(x, ) and
Maples numerical solution of the dierential equation. In this last example,

0.8

0.6

0.4

0.2

0.0 0.2 0.4 0.6 0.8 1.0

Fig. 2.1 Numerical solution of (2.21) and asymptotic approximations of the solution when
a(x, ) = cos(x) + x, b(x, ) = cos(x), y(0, ) = 1 and = 0.2.
2.3 Exercises 23

1 1 1 5 1 2
u1 (x) = , v1 (X) = (1 + X 2 ) X 3 P (X) + e 2 X . (2.48)
cos(x) + x 3 3 3

2.3 Exercises

2.1. Let y(x, ) be the solution to

y  + x(1 + x + )y = 2 + x2 , 1 x 2, (2.49)

satisfying y(1, ) = 1 + . Use P robA to show

2 + x2
y(x, ) = + e2(x1)/ + [u1 (x) + v1 ((x 1)/)] + O(2 ) (2.50)
x(1 + x)

for 1 x 2, where

2 4x + 3x2 + 2x3 + x4 + x5
u1 (x) = , (2.51)
x3 (1 + x)3
1
v1 (X) = (1 + 2X + 3X 2 )e2X . (2.52)
2

2.2. Note that An (X) and Bn (X) in (2.18) are polynomials and that therefore
Yn (X) = pn (X) + qn (X)exp[a(0, 0)X], where pn (X), qn (X) are polynomi-
als. Therefore vn (X) = qn (X)exp[a(0, 0)X], since vn () = 0. Furthermore,


n
vn (X) + Ank (X)vk (X) = 0, (2.53)
k=0

and vn (0) = [n] (0) un (0). Write a new, shorter Maple program for com-
puting the terms of (2.9).

2.3. Show that we could add 2 times c(x, ) C ([0, 1] [0, o ]), to xa(x, )
in (2.21) without upsetting the basic analysis, and modify the program P robB
appropriately to include it.

2.4. Let y(x, ) be the solution to

2 y  + x(1 + x + )y = x(2 + x2 ), 0 x 1, (2.54)

satisfying y(0, ) = 1. Use P robB to show

2 + x2 1 2
y(x, ) = e 2 (x/) + [u1 (x) + v1 (x/)] + O(2 ), (2.55)
1+x
24 2 First Order Dierential Equations

where
2 + x2 1 1 2
u1 (x) = , v1 (X) = 2P (X) + (12 + 3X 2 + 2X 3 )e 2 X . (2.56)
(1 + x)2 6

2.5. Suppose s(x) C ([0, 1]), s (x) > 0 and g(x, ) C ([0, 1] [0, o ])
for some o > 0. Let
 x
F (x, ) = 1 e[s(x)s(t)]/ g(x, t) dt, (2.57)
0
 1
G(x, ) = 1 e[s(x)s(t)]/ g(x, ) dt. (2.58)
x

From the analysis at the beginning of Section 2.1, we know F (x, ) has a
uniformly valid expansion of the form


N1
F (x, ) = n [un (x) + vn (x/)] + O(N ) (2.59)
n=0

as 0+ , where un (x) C ([0, 1]), vn (X) C ([0, ]) and, as in Exercise


2.2, vn (X) = o(X ) as X . Show that G(x, ) has a uniformly valid
expansion of the form


N1
G(x, ) = n [un (x) + wn ((1 x)/)] (2.60)
n=0

as 0+ , where un (x) C ([0, 1]), wn (X) C ([0, ]) and wn (X) =


o(X ) as X . In particular,

u0 (x) = g(x, 0)/s (x), w0 (X) = es (1)X . (2.61)

2.6. Suppose g(x, t), h(x, t) C ([0, 1] [0, x]), h(x, t) > 0 for 0 < t
x 1, h(x, 0) = 0 for 0 x 1, and h[0,1] (x, 0) > 0 for 0 < x 1 but
h[0,1] (0, 0) = 0. Let
 x
F (x, ) = eh(x,t) g(x, t) dt. (2.62)
0

We know that for x > 0,

F (x, ) = 1 [g(x, 0)/h[0,1] (x, 0)] + O( 1 ). (2.63)


2.3 Exercises 25

For an expansion of F (x, ) that is uniformly valid for 0 x 1, note rst


that as (x, t) (0, 0),

h(x, t) = h11 xt + h02 t2 + O((x2 + t2 )3/2 ), (2.64)

where we have begun using hij = h[i,j] (0, 0), gij = g [i,j] (0, 0), so it must be
that h11 x + h02 t 0 for 0 < t x 1.
Assume h11 > 0 and h02 > 0. If we let

u(x, t) = t2 [h(x, t) th[0,1] (x, 0)], a(x) = x1 h[0,1] (x, 0), (2.65)

then u(0, 0) > 0 and a(0) > 0, and hence there exists xo > 0 such that
u(x, t) > 0 on [0, xo ] [0, x] and a(x) > 0 on [0, xo ]. Therefore
 x
F (x, ) = f (x, t, t/)exta(x) dt, (2.66)
0

where = 1/2 ,
2
f (x, t, T ) = eT u(x,t)
g(x, t) (2.67)
and f (x, t, T ) C ([0, xo ] [0, x] [0, ]). In addition, 0 < exta(x) 1
and thus, applying Corollary 2, with (x, T, ) = f (x, T, T ), we get


N 1
F (x, ) = n n (x, x/) + O(N ) (2.68)
n=1

uniformly as 0+ for 0 x xo , where


 X
n (x, X) = [0,0,n] (x, T, 0)eXT a(x) dT. (2.69)
0

Show that n (x, X) C ([0, xo ] [0, ]), so we can apply Theorem 1 to


each term of (2.68), and thereby determine

F (x, ) = v1 (x/) + 2 [u2 (x) + v2 (x/)] + O(3 ) (2.70)

uniformly as 0+ for 0 x xo , where


 X
2
v1 (X) = g00 eh02 T h11 XT
dT, (2.71)
0

g(x, 0) g00
u2 (x) = , (2.72)
h[0,1] (x, 0) h11 x
and  X
2
v2 (X) = p2 (X, T )eh02 T h11 XT
dT, (2.73)
0
26 2 First Order Dierential Equations

where

p2 (X, T ) = g10 + (g01 g00 h21 X 2 )T g00 h12 XT 2 g00 h03 T 3 . (2.74)

Note, furthermore, that for x > 0 the right side of (2.70) is asymptotically
equal to 2 [g(x, 0)/h[0,1] (x, 0)] + O(3 ), the beginning of its outer expansion,
and therefore, in view of (2.63), equation (2.70) actually holds uniformly for
0 x 1.
The result (2.32) for the solution to Problem B is an example of an ex-
pansion for a case of F (x, ) with h11 > 0 and h02 < 0.
Chapter 3
Second Order Dierential Equations

3.1 Problem C

We begin this chapter with the classic singular perturbation problem

y  + a(x, )y  + b(x, )y = c(x, ), (3.1)

where a(x, ) > 0, subject to the boundary conditions y(0, ) = () and


y(1, ) = (). It will be assumed that a(x, ), b(x, ), c(x, ) C ([0, 1]
[0, o ]) and (), () C ([0, o ]) for some o > 0.
At rst we will ignore the condition y(1, ) = () and instead treat (3.1)
as an initial value problem with y [1,0] (0, ) = ()/ presumed given, where
() C ([0, o ]), in addition to y(0, ) = (). From our Problem A analysis
it is clear immediately that in the special case b(x, ) = a[1,0] (x, ) the solution
to this initial value problem has the asymptotic form


N 1
y(x, ) = n [un (x) + vn (x/)] + N RN (x, ), (3.2)
n=0

where un (x) C ([0, 1]), vn (X) C ([0, ]), vn (X) = o(X ) as X


and RN (x, ) = O(1) uniformly as 0+ for 0 x 1. We are going to
prove that in fact this is true in general. First we will prove it for N = 0.
If we let g(x, ) = a[1,0] (x, ) b(x, ), then
 x

y + a(x, )y = f (x, ) + g(t, )y(t, ) dt, (3.3)
0

where  x
f (x, ) = () + a(0, )() + c(t, ) dt. (3.4)
0
Therefore, introducing

L.A. Skinner, Singular Perturbation Theory, DOI 10.1007/978-1-4419-9958-0_3, 27


Springer Science+Business Media, LLC 2011
28 3 Second Order Dierential Equations
 x
k(x, ) = a(t, ) dt, (3.5)
0

as in Problem A, we have, after a reversal of integration order, the equivalent


integral equation
 x
y(x, ) = (x, ) + K(x, t, )y(t, ) dt, (3.6)
0

where  x
1
(x, ) = () + e[k(x,)k(t,)]/ f (t, ) dt (3.7)
0
and  x
K(x, t, ) = 1 g(t, ) e[k(x,)k(s,)]/ ds. (3.8)
t
Again as in Problem A, we know by Corollary 2 that
 x
1
e[k(x,)k(t,)]/ dt = O(1) (3.9)
0

uniformly as 0+ for 0 x 1. Therefore, since () = O(1) and f (x, )


is uniformly O(1) as 0+ for 0 x 1, there exists M > 0 such that
|(x, )| M for all (x, ) [0, 1] (0, o ]. Similarly, |K(x, t, )| M for
all (x, t, ) [0, 1] [0, x] (0, o ]. Hence, by a simple Gronwall argument,
y(x, ) is uniformly O(1) as 0+ for 0 x 1. Indeed, for all (x, )
[0, 1] (0, o ], we have
 x
|y(x, )| (x, ) = M + M |y(t, )| dt, (3.10)
0

so [1,0] (x, ) M (x, ) and therefore eM x (x, ) (0, ) = M , or


|y(x, )| M eM x M eM .
It is easy to see what happens if we substitute (3.2) into (3.1). Let
an (x) = a[0,n] (x, 0), bn (x) = b[0,n] (x, 0) and cn (x) = c[0,n] (x, 0). Also
let A(X, ) = a(X, ), B(X, ) = b(X, ), An (X) = A[0,n] (X, 0) and
Bn (X) = B [0,n] (X, 0). Note, for example, that


N 1 
N1 
n
a(x, ) n un (x) = n ak (x)unk (x) + O(N ) (3.11)
n=0 n=0 k=0

uniformly as  0+ for 0 x 1. Similarly, since we are expecting vn (X) =


o(X ) as X , Corollary 2 implies


N 
N 
n
a(x, ) n vn (x/) = n

Ak (x/)vnk (x/) + O(N +1 ) (3.12)
n=0 n=0 k=0
3.1 Problem C 29

uniformly as 0+ for 0 x 1. Hence if we put



n 
n
un1 (x) + ak (x)unk (x) + bk (x)unk (x) = cn (x) (3.13)
k=0 k=0

for 0 n N 1 and


n 
n1
vn (X) + 
Ak (X)vnk (X) + Bk (X)vn1k (X) = 0, (3.14)
k=0 k=0

for 0 n N , then (3.2) implies


 
rN + a(x, )rN + b(x, )rN = N (x, ), (3.15)

where rN (x, ) = RN (x, ) vN (x/) and N (x, ) is uniformly O(1) as


0+ for 0 x 1.
To satisfy y(0, ) = () we require, according to (3.2), un (0)+vn (0) = n ,
where n = [n] (0), and this implies rN (0, ) = O(1). Also y [1,0] (0, ) =
[1,0]
1 () means un1 (0) + vn (0) = n , where n = [n] (0), so rN (0, ) =
O(1). Therefore, just as y(x, ) = O(1), it follows that rN (x, ) = O(1) uni-
formly as 0+ for 0 x 1. To complete the proof that the asymptotic
solution of the initial value problem for (3.1) has the stipulated form (3.2),
we need to conrm that vn (X) = o(X ) as X 0+ . But this follows
immediately from (3.14) once we assert vn () = 0.
Note that to calculate the terms of (3.2) we have to alternate back and
forth between (3.13) and (3.14), starting with (3.14), applying vn (0) =
n un1 (0), vn () = 0, rst with n = 0, then going to (3.13) with
un (0) = n vn (0), thereby determining v1 (0) = 1 u0 (0), and so on. For
the boundary value problem, on the other hand, to satisfy y(1, ) = (),
note that (3.2) implies un (1) = [n] (0), so un (x) for 0 n N 1
can be determined separately. Then vn (X) for 0 n N 1 follows
from (3.14) using vn (0) = n un (0), vn () = 0. The fact that, in-
deed, the boundary value problem for (3.1) has a solution of the form (3.2)
is conrmed by noting that the solution to the initial value problem has
the form y(x, ) = w(x, ) + ()z(x, ), where w(x, ) is the solution to
(3.1) satisfying w(0, ) = (), w (0, ) = 0 and z(x, ) is the solution to
z  + a(x, )z  + b(x, )z = 0 satisfying z(0, ) = 0, z  (0, ) = 1/. In partic-
1
ular, therefore, z(1, 0) = [1/a0 (0)] exp 0 [b0 (x)/a0 (x)] dx, which is positive,
so we can assume o is such that z(1, ) > 0 on [0, o ], and thus the boundary
value problem for (3.1) is equivalent to the initial value problem dened by
taking () = [() w(1, )]/z(1, ). Note also that if, for the boundary
value problem, we choose () such that n = un (0), then vn (X) = 0. In
other words, the N-term outer expansion,
30 3 Second Order Dierential Equations


N 1
ON y(x, ) = n un (x) (3.16)
n=0

is, by itself, the rst N terms of a uniformly valid asymptotic expansion of a


solution to (3.1).
Other strategies for calculating the terms of (3.2) are presented in the ex-
ercises. Below is a program for solving the boundary value problem based on
the above strategy.

P robC := proc(a, b, c, , , N )
N u := sum(n un , n = 0..N 1); N du := sum(n dun , n = 0..N 1);
N ddu := sum(n ddun , n = 0..N 1);
N ude := series( N ddu + a N du + b N u c, = 0, N );
N := series(, , 0, N ); N := series(, = 0, N );
for k from 0 to N 1 do
temp := coe (N ude, , k);
de := subs(uk = z(x), duk = di (z(x), x), dduk = di (z(x), x, x), temp);
bc := z(1) = coe (N , , k);
dsolve({de, bc}); uk := rhs(%); duk := di (uk, x); dduk := di (duk, x);
N ude := subs(uk = uk, duk = duk, dduk = dduk, N ude);
print(uk = simplify(uk)); uk := uk;
end do;
A := subs(x = X, a); B := subs(x = X, b);
N v := sum(n vn , n = 0..N 1); N dv := sum(n dvn , n = 0..N 1);
N ddv := sum(n ddvn , n = 0..N 1);
N vde := series(N ddv + A N dv + B N v, = 0, N );
for k from 0 to N 1 do
temp := coe (N vde, , k);
de := subs(vk = z(X), dvk = di (z(X), X), ddvk = di (z(X), X, X), temp);
bc := z(0) = coe (N , , k) subs(x = 0, uk ), z() = 0;
dsolve({de, bc}); vk := rhs(%); dvk := di (vk, X); ddvk := di (dvk, X);
N vde := subs(vk = vk, dvk = dvk, ddvk = ddvk, N vde);
print(vk = simplify(vk));
end do;
end proc:

3.2 Problem D

The problem for this section is

2 y  + 2 a(x, )y  b(x, )y = c(x, ) (3.17)


3.2 Problem D 31

with y(0, ) = (), y(1, ) = (). As usual, a(x, ), b(x, ), c(x, )


C (([0, 1] [0, o ]) and (), () C ([0, o ]) for some o > 0. In ad-
dition, in this problem we assume b(x, ) > 0.
For the corresponding homogeneous dierential equation, c(x, ) = 0, if
we put
y(x, ) = es(x)/ z(x, ) (3.18)
and choose [s (x)]2 = b(x, 0), then

z  + a(x, )z  + b(x, )z = 0, (3.19)

where a(x, ), b(x, ) C ([0, 1] [0, o ]) and, in particular,

a(x, ) = 2s (x) + a(x, ). (3.20)

If we take s (x) to be the positive square root of b(x, 0), then (3.19) is a
case of Problem C and, as noted at the end of Section 3.1, then (3.19) has a
solution with a uniformly valid expansion as 0+ of the form


N1
z(x, ) = n zn (x) + O(N ), (3.21)
n=0

where zn (x) C ([0, 1]). If we also require z(0, 0) = 1, then z(0, ) > 0 for
> 0 suciently small.
Actually, a solution to (3.19) of the form (3.21) also exists if s (x) =
[b(x, 0)]1/2 . Indeed, in this case, simply replacing x by 1 x changes (3.19)
into another case of Problem C. Hence, to summarize, if we let
 x
s(x) = [b(t, 0)]1/2 dt, (3.22)
0

there exists a pair of linearly independent solutions


(1)
yh (x, ) = es(x)/ z (1) (x, ), (3.23)
(2)
yh (x, ) = e[s(1)s(x)]/ z (2) (x, ), (3.24)
to the homogeneous dierential equation corresponding to (3.17), where for
(k) (k)
any N > 0, and certain zn (x) C ([0, 1]) with z0 (0) = 1,


N1
z (k) (x, ) = n zn(k) (x) + O(N ) (3.25)
n=0

uniformly as 0+ for 0 x 1.
The Wronskian of these two homogeneous equation solutions is W (x, ) =
W (0, )(x, ), where
32 3 Second Order Dierential Equations
x
a(t,) dt
(x, ) = e 0 (3.26)
and W (0, ) = 1 es(1)/ (), where () has an expansion in powers of
as 0+ . In particular, (0) = 2[b(0, 0)]1/2 . By the method of variation of
parameters, if we let

g (1) (x, t, ) = z (1) (x, )(t, )z (2) (t, ), (3.27)

g (2) (x, t, ) = z (2) (x, )(t, )z (1) (t, ), (3.28)


where (t, ) = c(t, )/[()(t, )], then the full dierential equation (3.17)
has a particular solution

yp (x, ) = F (1) (x, ) + F (2) (x, ), (3.29)

where  x
F (1) (x, ) = 1 e[s(x)s(t)]/ g (1) (x, t, ) dt, (3.30)
0
 1
F (2) (x, ) = 1 e[s(t)s(x)]/ g (2) (x, t, ) dt. (3.31)
x

From Exercise 2.5, we know F (1) (x, ), F (2) (x, ) have uniformly valid
asymptotic expansions for 0 x 1, expressible as


N1
F (1) (x, ) = n [u(1) (1) N
n (x) + vn (x/)] + O( ), (3.32)
n=0


N 1
F (2) (x, ) = n [u(2) (2)
n (x) + vn ((1 x)/)] + O( ),
N
(3.33)
n=0
(k) (k) (k)
where un (x) C ([0, 1]), vn (X) C ([0, ]) and vn (X) = o(X ) as
(k)
X . In addition, there exists vh,n (X) C ([0, ]) such that

(1)

N1
(1)
yh (x, ) = n vh,n (x/) + O(N ), (3.34)
n=0


N 1
(2) (2)
yh (x, ) = n vh,n ((1 x)/) + O(N ) (3.35)
n=0
(k)
uniformly for 0 x 1 as 0+ , and vh,n (X) = o(X ) as X .
Putting all this together we see the solution to (3.17), namely yp (x, ) plus
(1) (2)
an -dependent linear combination of yh (x, ) and yh (x, ), subject to the
boundary conditions y(0, ) = (), y(1, ) = (), has a uniformly valid
expansion for 0 x 1 of the form
3.2 Problem D 33


N 1
y(x, ) = n [un (x) + vn (x/) + wn ((1 x)/)] + O(N ) (3.36)
n=0

as 0+ , where un (x) C ([0, 1]) and vn (X), wn (X) C ([0, ]).


Furthermore, as in Problem C, and also A, because we also have vn (X),
wn (X) = o(X ) as X , each set of terms in this expansion can be
determined separately. Indeed, with an (x) = a[0,n] (x, 0), bn (x) = b[0,n] (x, 0)
and cn (x) = c[0,n] (x, 0), it is clear that


n2 
n
un2 (x) + ak (x)un2k (x) bk (x)unk (x) = cn (x). (3.37)
k=0 k=0

Also, in terms of An (X) = A[0,n] (X, 0), Bn (X) = B [0,n] (X, 0), where
A(X, ) = a(X, ), B(X, ) = b(X, ), we have


n1 
n
vn (X) + 
Ak (X)vn1k (X) Bk (X)vnk (X) = 0, (3.38)
k=0 k=0

and nally, if we change and let A(X, ) = a(1X, ), B(X, ) = b(1X, ),


then

n1 
n1
wn (X) 
Ak (X)wn1k (X) Bk (X)wnk (X) = 0, (3.39)
k=0 k=0

where, again, An (X) = A[0,n] (X, 0), Bn (X) = B [0,n] (X, 0). So, rst we
nd the outer expansion terms un (x) for 0 n N 1, beginning with
u0 (x) = c(x, 0)/b(x, 0), from (3.37). Then the (constant coecient) dier-
ential equations (3.38) and (3.39) can be solved successively, with the bound-
ary conditions vn (0) = [n] (0)un (0), wn (1) = [n] (0)un (1), together with
vn () = wn () = 0. A Maple program patterned after P robC is readily de-
vised to automate these calculations. To get more than a couple of terms,
however, Maple needs assistance like the assistance provided in P robB to
calculate vn (X) and wn (X). This issue comes up in the next section, too.
If a(x, ) = 1 + x, b(x, ) = 2 + x + , c(x, ) = 4x2 and () = 2,
() = 1/2, then

4x2 5
u0 (x) = , v0 (X) = 2e 2X
, w0 (X) = e 3X . (3.40)
2+x 6

Figure 3.1 is a graph of u0 (x)+v0 (x/)+w0 ((1x)/) when = .15, together


with Maples numerical solution of (3.17) in this case and u0 (x) alone. The
same items are graphed in Figure 3.2 using = .30.
34 3 Second Order Dierential Equations

2.0

1.5

1.0

0.5

0.0

0.0 0.2 0.4 0.6 0.8 1.0

Fig. 3.1 Numerical solution to (3.17) and asymptotic approximations of the solution when
a(x, ) = 1 + x, b(x, ) = 2 + x + , c(x, ) = 4x2 , y(0, ) = 2, y(1, ) = 1/2 and = 0.15.

2.0

1.5

1.0

0.5

0.0
0.0 0.2 0.4 0.6 0.8 1.0

Fig. 3.2 Same as Figure 3.1 except = 0.30.

3.3 Problem E, Part 1

The problem for this section is a variation of Problem D in the same way Prob-
lem B is a variation of A. Assuming the functions a(x, ), c(x, ), d(x, )
C ([0, 1] [0, o ]) for some o > 0, we also assume 0 < b(x) C ([0, 1])
and, without loss of generality, set b(0) = 1. The dierential equation is

2 y  + 2 a(x, )y  [xb(x) + 2 c(x, )]y = d(x, ), (3.41)


3.3 Problem E, Part 1 35

and we are interested in the solution satisfying the boundary conditions


y(0, ) = (), y(1, ) = (), where (), () C ([0, o ]). We will start
with an analysis of the special case

2 y  xy = d(x, ). (3.42)

For (3.42), the corresponding homogeneous equation has the two linearly
independent Airy function solutions
(1) (2)
yh (x, ) = Ai(x/2/3 ), yh (x, ) = e(2/3)/ Bi(x/2/3 ) (3.43)

and, knowing the Wronskian of Ai(X) and Bi(X) is 1/, it is readily veried
that
yp (x, ) = F (1) (x, ) + F (2) (x, ) (3.44)
is a particular solution to (3.42), where
 x
F (1) (x, ) = Ai(x/2/3 ) d(t, )Bi(t/2/3 ) dt, (3.45)
0
 1
F (2) (x, ) = Bi(x/2/3 ) d(t, )Ai(t/2/3 ) dt. (3.46)
x
If we let
3/2 3/2
g(X) = e(2/3)X Ai(X), h(X) = e(2/3)X Bi(X), (3.47)

and dene g(X) = g(X 4 ), h(X) = h(X 4 ), then g(X), h(X) C ([0, ]). If
we also let x = x1/4 , = 1/6 , and substitute t = t4 , then
 x
F (1) (x, ) = g(x/) t, )h(t/)e(2/3)(x6 t6 )/6 dt,
d( (3.48)
0
 1
F (2) (x, ) = h(x/) t, )g(t/)e(2/3)(x6 t6 )/6 dt,
d( (3.49)
x
1/6
t, ) = 4 t3 d(t4 , 6 ) C ([0, 1] [0, o ]). The asymptotic expan-
where d(
sion form of these two integrals is readily established.
By Corollary 1, there exists n (t) C ([0, 1]), n (T ) C ([0, ]) such
that for any N 0,


N1
t, )g(t/) =
d( n [n (t ) + n (t/)] + O(N ) (3.50)
n=0

uniformly for 0 t 1, and the same can be said for d( t, )h(t/). In turn,
the results of Exercise 3.5, applied with u(t) = n (t) and v(T ) = n (T ), show
(k) (k) (2)
there exists n (x) C ([0, 1]), n (X) C ([0, ]), and also n (X) =
36 3 Second Order Dierential Equations

o(X ) in C ([0, ]), such that


N 1
F (1) (x, ) = g(x/) n [(1) (1) N
n (x) + n (x/)] + O( ), (3.51)
n=0


N1
F (2) (x, ) = h(x/) n [(2) (2) (2)
n (x) + n (x/) + n ((1 x)/)] + O( ).
N

n=0
(3.52)
From here, with a few additional applications of Theorem 1, it is clear that the
solution to (3.42) satisfying the boundary conditions y(0, ) = (), y(1, ) =
() has the asymptotic form


N 1
y(x, ) = n [un (x) + vn (x/) + wn ((1 x)/)] + O(N ), (3.53)
n=0

where un (x) C ([0, 1]), both vn (X), wn (X) C ([0, ]), and wn (X) =
(2)
o(X ) as X . For example, to deal with the contribution of yh (x, )
(2)
to (3.53), let t = 1 x and observe that yh (x, ) = f (t, t/, ), where

f (t, T, ) = eT (t) h((1 t)1/4 /) (3.54)

with (t) = (2/3)t1 [1 (1 t)3/2 ]. Since (t) > 0 for 0 t 1/2,


f (t, T, ) C ([0, 1/2] [0, ] [0, 1]). Thus we can apply Theorem 1 to
each f [0,0,n] (t, t/, 0) to see there exist polynomials pn (T ) such that for any
N 0,

N 1
(2)
yh (x, ) = n pn (t/)et/ + O(N ) (3.55)
n=0

uniformly as 0 for 0 t 1/2. Furthermore, both y (2) (x, ) and


+

the sum in (3.55) are uniformly o( ) for 1/2 t 1. Hence (3.55) holds
uniformly for the full interval, 0 x 1.
As in Problem D, the homogeneous version of (3.41), the full dierential
equation for Problem E, has a pair of linearly independent solutions express-
ible as
(1)
yh (x, ) = es(x)/ z (1) (x, ), (3.56)
(2)
yh (x, ) = e[s(1)s(x)]/ z (2) (x, ), (3.57)
where  x
s(x) = [tb(t)]1/2 dt. (3.58)
0

However, the asymptotic form of z (k) (x, ) is more complex. It is shown in


[9] that if a(x, ) = 0, then for any N 0,
3.3 Problem E, Part 1 37


N1
z (k) (x, ) = n/6 [u(k) (k)
n (x) + vn (x/
2/3
)] + O(N/6 ), (3.59)
n=0

(k) (k)
where un (x4 ) C ([0, 1]) and vn (X 4 ) C ([0, ]). In Exercise 3.6,
we show that in fact this is true whether a(x, ) = 0 or not. Therefore, as
above for (3.42) and as in Problem D, we can use the method of variation
of parameters to obtain a particular solution of (3.41). It is also a straight-
forward matter to determine the asymptotic form of this solution, and the
homogeneous equation solutions (3.56) and (3.57).
Note that s(x) = x3/2 (x), where 0 < (x) C ([0, 1]), and again intro-
duce x = x1/4 to get s(x) = x6 (x), where (x) = (x4 ). Then, for example,

es(x)/ vn(k) (x/2/3 ) = f (x, x/) (3.60)

where = 1/6 , as above, and


6
f (x, X) = eX (x) (k)
vn (X 4 ), (3.61)

which is in C ([0, 1] [0, ]). Furthermore, f [0,n] (x, ) = 0, so


N1
f (x, x/) = n [0,n] (x/, 0) + O(N ), (3.62)
n=0

where (X, ) = f (X, X). Also for example, to deal with


 x
6 6
F (x, ) = e(x/) (x)
u(t )e(t/) (t )
dt, (3.63)
0

where u(t ) C ([0, 1]), let t = t[(t )]1/6 , let t = (t ) denote the inverse of
this transformation, and let x = x[(x)]1/6 to obtain
 x
(x/)6 6
F (x, ) = e g(t)e(t/) dt, (3.64)
0

where g(t ) = u((t )) (t ). It follows by Exercise 3.5 that


N 1
F (x, ) = n [n (x) + n (x/)] + O(N ), (3.65)
n=0

for any N 0, where n (x) C ([0, 1]) and n (X) C ([0, ]). Also,
we can apply Theorem 1 to f (x, X) = n (X[(x)]1/6 ). Thus, ultimately,
we see the solution to the boundary value problem for (3.41) has the same
asymptotic form as the solution for the special case (3.42). That is, there
exists un (x), vn (X) and wn (X) such that, for any N 0,
38 3 Second Order Dierential Equations


N 1
y(x, ) = n/6 [un (x) + vn (x/2/3 ) + wn ((1 x)/)] + O(N/6 ), (3.66)
n=0

uniformly as 0+ for 0 x 1, where un (x4 ) C ([0, 1]), both vn (X 4 )


and wn (X) are in C ([0, ]), vn () = 0 and wn (X) = o(X ) as X .

3.4 Problem E, Part 2

To determine the functions un (x) = un (x4 ) = un (x) and vn (X) = vn (X 4 ) =


vn (X) in (3.66), we need to compute the outer and inner expansions


N 1
ON y(x, ) = n y [0,n] (x, 0), (3.67)
n=0

and

N1
IN y(x, ) = n Y [0,n] (X, 0), (3.68)
n=0

where y(x, ) = y(x4 , 6 ) and Y (X, ) = y(X, ). If we let z(x, ) = y(x, 6 ),


then y(x, ) = z(x4 , ) and therefore


N 1
ON y(x, ) = n zn (x) (3.69)
n=0

where zn (x) = z [0,n] (x, 0). If we also let Z(X, ) = y(4 X, 6 ), then Z(X 4 , ) =
Y (X, ), so

N1
IN y(x, ) = n Zn (X), (3.70)
n=0
[0,n]
where Zn (X) = Z (X, 0), and X = X 4 .
Upon substituting = 6 in (3.41), it is clear that zn (x) = 0 unless n 1
is a multiple of 6. In particular,

z6 (x) = d(x, 0)/[xb(x)], z12 (x) = d[0,1] (x, 0)/[xb(x)]. (3.71)

Also, therefore, from the expansion of 6 z6 (4 X) + 12 z12 (4 X), we have

I14 O14 y(x, ) = 6 [d00 /x + d10 d00 b1 ] 10 [d00 (b21 b2 )


d10 b1 + d20 ]X 12 [d01 /x + d11 d01 b1 ], (3.72)

where dmn = d[m,n] (0, 0) and bn = b[n] (0). Hence,


3.4 Problem E, Part 2 39

u6 (x) = z6 (x) + d00 /x, u12 (x) = z12 (x) + d01 /x (3.73)

and otherwise un (x) = 0 for 0 n 13. Also,

v6 (X) = Z6 (X) + d10 d00 b1 , v12 (X) = Z12 (X) + d11 d01 b1 , (3.74)

v10 (X) = Z10 (X) + [d00 (b21 b2 ) d10 b1 + d20 ]X, (3.75)
and otherwise vn (X) = Zn (X) for 0 n 13.
From the dierential equation for Z(X, ) it is clear that Zn (X) = 0 if n
is odd. Also,
Z0 XZ0 = 0, Z0 (0) = 0 , (3.76)
Z2 XZ2 = d00 , Z2 (0) = 0, (3.77)
Z4 XZ4 = b1 X 2 Z0 a00 Z0 , Z4 (0) = 0, (3.78)
Z6 XZ6 = b1 X Z2 2
a00 Z2 d10 X, Z6 (0) = 1 , (3.79)
where n = [n] (0). In addition, vn () = 0 implies, in particular, Z0 () =
Z2 () = 0. Therefore

Z0 (X) = [0 /Ai(0)]Ai(X), (3.80)

and, in terms of
 X 
P (X) = Ai(X) Bi(T ) dT Bi(X) Ai(T ) dT, (3.81)
0 X

which satises P  XP = 1 , we have

Z2 (X) = d00 P (X) [d00 P (0)/Ai(0)]Ai(X). (3.82)

In general, to nd Zn (X), we need to solve an equation of the form

Z  XZ = (X)R(X)+(X)S(X)+(X)P (X)+(X)Q(X)+ (X), (3.83)

where R(X) = Ai(X), S(X) = Ai (X), Q(X) = P  (X), and (X), (X),
(X), (X) and (X) are polynomials. But the desired solution to an equa-
tion of this form has the same form as its right side. That is,

Z(X) = p(X)R(X) + q(X)S(X) + r(X)P (X) + s(X)Q(X) + t(X), (3.84)

where p(X), q(X), r(X), s(X) and t(X) are polynomials.


Suppose, for example, that

(X) = 0 + 1 X + 2 X 2 , (X) = 0 + 1 X + 2 X 2 . (3.85)

Substitution of (3.84) into (3.83) reveals


40 3 Second Order Dierential Equations

p 2Xq  + q = (X), q  + 2p = (X). (3.86)

and upon inserting

p(X) = p0 + p1 X + p2 X 2 + p3 X 3 , q(X) = q0 + q1 X + q2 X 2 , (3.87)

we get
2p2 + q0 = 0 , 6p3 + 3q1 = 1 , 5q2 = 2 , (3.88)
2q2 + 2p1 = 0 , 4p2 = 0 , 6p3 = 2 . (3.89)
Hence, in order,
1 1 1
q2 = 2 , p3 = 2 , q1 = (1 6p3 ), (3.90)
5 6 3
1 1
0 , q0 = 0 2p2 , p1 = (0 2q2 ).
p2 = (3.91)
4 2
The functions r(X), s(X), and t(X) are similarly determined once (X),
(X) and (X) are given and, in the end, p0 is determined by

Z(0) = p0 R(0) + q0 S(0) + r(0)P (0) + s(0)Q(0) + t(0). (3.92)

Below is a Maple program that performs these calculations for us. The
program also computes wn (X), using the fact that wn (X) is a polynomial
times exp[b(1)X], and


N 1
y(1, ) = ON y(1, ) + n wn (0) + O(N ), (3.93)
n=0

which follows from (3.66), since, of course, [IN ON IN ]y(1, ) = O(N ).


Although somewhat longer than our previous programs, P robE follows a fa-
miliar procedure. First there is a section for the computation of ON y(x, ).
Then there is a longer than normal section, because of all the polynomials to
determine, for the computation of IN y(x, ). After computing IN ON y(x, ),
the functions un (x) and vn (X) are determined, and at the end of the program
there is a separate section to determine each wn (X). The symbols I,P , Q, R
and S in the program are used to denote 1, P (X), Q(X), R(X) and S(X),
respectively. Also, P o = P (0), Qo = Q(0), Ro = R(0) and So = S(0).

P robE := proc(a, b, c, d, , , N )
ON y := sum(n yn , n = 0..N 1);
ON dy := sum(n dyn , n = 0..N 1);
ON ddy := sum(n ddyn , n = 0..N 1);
eq := subs( = 6 , 2 ON ddy + 2 a ON dy (x b + 2 c) ON y d);
ON eq := series(eq,  = 0, N );
for k from 0 to N 1 do
3.4 Problem E, Part 2 41

temp := coe (ON eq, , k); yk := solve(temp = 0, yk );


dyk := di (yk, x); ddyk := di (dyk, x);
ON eq := subs(yk = yk, dyk = dyk, ddyk = ddyk, ON eq);
ON y := subs(yk = yk, ON y); yk = yk;
end do;
N := series(subs( = 6 , ),  = 0, N );
N := series(subs( = 6 , ),  = 0, N );
A := subs( = 6 , x = 4 X, a); B := subs(x = 4 X, b);
C := subs( = 6 , x = 4 X, c); D := subs( = 6 , x = 4 X, d);
IN y := sum(n Yn , n = 0..N 1); IN dy := sum(n dYn , n = 0..N 1);
rths := series(4 AIN dy+(XBX+8 C)IN y+2 DI,  = 0, N +1);
for k from 0 to N 1 do
temp := coe (rths, , k);
if temp = 0 then M := 0; else M := degree(temp, X); end if ;
:= coe (temp, R); := coe (temp, S); := coe (temp, P );
:= coe (temp, Q); := coe (temp, I);
pm := sum(pn X n , n = 0..M + 1); qm := sum(qn X n , n = 0..M );
rm := sum(rn X n , n = 0..M + 1); sm := sum(sn X n , n = 0..M );
tm := sum(tn X n , n = 0..M );
for j from 0 to M do
coe (di (pm, X, X) + 2 X di (qm, X) + qm , X, M j);
solve(% = 0, qM j ); qm := subs(qM j = %, qm);
coe (di (qm, X, X) + 2 di (pm, X) , X, M j);
solve(% = 0, pM +1j ); pm := subs(pM +1j = %, pm);
coe (di (rm, X, X) + 2 X di (sm, X) + sm , X, M j);
solve(% = 0, sM j ); sm := subs(sM j = %, sm);
coe (di (sm, X, X) + 2 di (rm, X) , X, M j);
solve(% = 0); rm := subs(rM +1j = %, rm);
coe (di (tm, X, X) X tm + 2 di (sm, X) + rm , X, M + 1 j);
solve(% = 0, tM j ); tm := subs(tM j = %, tm);
end do;
r0 := coe (, X, 0) 2 coe (sm, X, 1) 2 coe (tm, X, 2);
rm := subs(r0 = r0, rm); r0 := r0;
q0 := coe (qm, X, 0); s0 := coe (sm, X, 0); t0 := coe (tm, X, 0);
p0 := (coe (N , , k) q0 So r0 P o s0 Qo t0) Ro1 ;
p0 := simplify(%); pm := subs(p0 = p0, pm);
Y k := pm R + qm S + rm P sm Q + tm I;
dY k := (di (pm, X) + X qm) R + (di (qm, X) + pm) S
+(di (rm, X)+X sm)P +(di (sm, X)+rm)Q+(di (tm, X)+sm)I;
rths := subs(Yk = Y k, dY k = dY k, rths); Yk = Y k;
end do;
series(subs(x = 4 X 4 , ON y),  = 0, N ); IN ON y := convert(%, polynom);
vpart := sum(X n coe (IN ON y, X, n), n = 0..N );
upart := subs(X = 4 x, IN ON y vpart);
for k from 0 to N 1 do
42 3 Second Order Dierential Equations

uk := yk coe (upart, , k);


print(uk = simplify(uk));
end do;
for k from 0 to N 1 do
vk := Yk I coe (vpart, , k);
print(vk = vk);
end do;
A := subs( = 6 , x = 1 6 X, a);
B := subs( = 6 , x = 1 6 X, b);
C := subs( = 6 , x = 1 6 X, c); m := sqrt(subs(x = 1, b));
N w := sum(n wn , n = 0..N 1); N dw := sum(n dwn , n = 0..N 1);
rths := series(6 AN dw +((16 X)B +12 C m2 )N w,  = 0, N );
for k from 0 to N 1 do
:= coe (rths, , k);
if = 0 then K = 0 else K := degree(, X); end if ;
0 := coe (N , , k) subs(x = 1, yk ); K+2 := 0;
for j from 0 to K do
i := K j;
i+1 := ((i + 2) (i + 1) i+2 coe (, X, i)) (2 m (i + 1))1 ;
end do;
sum(n X n , n = 0..K + 1); s := simplify(%);
rths := subs(wk = s, dwk = di (s, X) m s, rths);
print(wk = s emX );
end do;
end proc:

3.5 Exercises

3.1. Complete the analysis necessary to verify that N (x, ) in (3.15) is uni-
formly O(1) for 0 x 1 as 0+ . In particular, using (3.11) and (3.13),
show that

N1
n [un (x) + a(x, )un (x) + b(x, )un (x)] c(x, ) = O(N ) (3.94)
n=0

uniformly for 0 x 1 and using (3.12) and (3.14), show that


N
n [vn (x/) + a(x, )vn (x/) + b(x, )vn1 (x/)] = O(N ) (3.95)
n=0

uniformly for 0 x 1.
3.5 Exercises 43

3.2. Our approach to Problem C demonstrates a general method for conrm-


ing the uniform validity of singular perturbation calculations. This method
is used in [7] and [12] to treat a nonlinear generalization of Problem C, for
example, and also in [2] and [4], where the emphasis is on partial dieren-
tial equation problems, but without the clarity achievable with Corollary 1
and Corollary 2. An application of the method to treat a nonlinear integral
equation generalization of Problem C, in which Corollary 2 again plays a
prominent role, is given in [10].
In general, suppose f (x, X, ) C ([0, 1] [0, ] [0, o ]) for some
o > 0 so that by Corollary 1 there exists un (x) C ([0, 1]) and vn (X)
C ([0, ]) with vn () = 0 such that, for any N 0,


N 1
f (x, x/, ) = n [un (x) + vn (x/)] + O(N ) (3.96)
n=0

uniformly for 0 x 1 as 0+ . If, in addition, h(x, y, ) C ([0, 1]


[a, b] [0, o ]) and a f (x, X, ) b for all (x, X, ) [0, 1] [0, ]
[0, o ], then g(x, X, ) = h(x, f (x, X, ), ) also is in C ([0, 1][0, ][0, o ])
and therefore, for certain n (x) C ([0, 1]) and n (X) C ([0, ]) with
n () = 0, we know


N 1
g(x, x/, ) = n [n (x) + n (x/)] + N N (x, ), (3.97)
n=0

where N (x, ) = O(1) uniformly for 0 x 1 as 0+ . Also, in particular,

0 (x) = h(x, u0 (x), 0), (3.98)

0 (X) = h(0, u0 (0) + v0 (X), 0) h(0, u0 (0), 0), (3.99)

1 (x) = u1 (x)h[0,1,0] (x, u0 (x), 0) + h[0,0,1] (x, u0 (0), 0)

+c0 x1 [h[0,1,0] (x, u0 (x), 0) h[0,1,0] (0, u0 (0), 0)], (3.100)


[1]
where c0 = v0 (), and
4

1 (X) = v1 (X)h[0,1,0] (0, u0 (0) + v0 (X), 0) + An (X), (3.101)
n=1

where

A1 (X) = u0 (0)X[h[0,1,0] (0, u0 (0) + v0 (X), 0)

h[1,0,0] (0, u0 (0), 0) 2c0 X 1 h[0,2,0] (0, u0 (0), 0)], (3.102)


44 3 Second Order Dierential Equations

A2 (X) = X[h[1,0,0] (0, u0 (0) + v0 (X), 0)

h[0,1,0] (0, u0 (0), 0) c0 X 1 h[1,1,0] (0, u0 (0), 0)], (3.103)

A3 (X) = [h[0,0,1] (0, u0 (0) + v0 (X), 0) h[0,0,1] (0, u0 (0), 0)], (3.104)
A4 (X) = u1 (0)[h[0,1,0] (0, u0 (0) + v0 (X), 0) h[0,1,0] (0, u0 (0), 0)]. (3.105)

3.3. Another way to calculate the terms of (3.2) when the initial conditions
y(0, ) = (), y [1,0] (0, ) = 1 () are given is to begin by rst calculating
the terms of the inner expansion,


N1
IN y(x, ) = n Yn (X). (3.106)
n=0

These satisfy Yn (0) = n and Yn (0) = n and just as noted for Problem A
in Exercise 2.2, they are of the form Yn (X) = pn (X) + qn (X)exp[a(0, 0)X],
so vn (X) = Yn (X) pn (X) for 0 n N 1. Once this calculation is
done, the terms of ON y(x, ) which satisfy (3.13), can be computed using
un (0) = n vn (0). Write a Maple program to do these calculations.

3.4. There also is another way to calculate the terms of (3.2) when the bound-
ary conditions y(0, ) = (), y(1, ) = () are given. The idea is to rst
solve (3.1) for y(x, , ), subject to the initial conditions y(0, , ) = () and
y [1,0,0] (0, , ) = /. This yields

 1
N
 
y(x, , ) = n un (x, ) + vn (x/, ) + O(N ), (3.107)
n=0

say, and then, to satisfy y(1, , ) = (), we just have to solve the (linear)
equation

N1
() = n un (1, ) + O(N ) (3.108)
n=0

for

N 1
= n n + O(N ), (3.109)
n=0

then substitute into (3.107) and reduce the result to the nal form (3.2),
altogether a simple task for Maple. Try this out with your own choice of
parameters, using your program from the previous exercise. You can check
your results using the program at the end of Section 3.1.

3.5. First, suppose u(t) C ([0, 1]) and


3.5 Exercises 45
 x
1 (x/)k k
F (x, ) = e u(t)e(t/) dt, (3.110)
0

where k 2 is an integer. Let


k2
p(t) = u[m] (0)tm , q(t) = tk+1 [u(t) p(t)]. (3.111)
m=0

Show that

k2
1 k1 k
F (x, ) = m u[m] (0)Pm (x/) + [q(x) q(0)e(x/) ]
m=0
k
 x
1 k k
+ k1 e(x/) q  (t)e(t/) dt, (3.112)
k 0

where  X
X k k
Pm (X) = e T m eT dT, (3.113)
0

and show that, for 0 m k 2, Pm (X) C ([0, ]). Since q  (t)


C ([0, 1]), this process can be repeated indenitely, and hence, for any N
0,

N1
F (x, ) = n [un (x) + vn (x/)] + O(N ) (3.114)
n=0

uniformly as 0 for 0 x 1, where un (x) C ([0, 1]), vn (X)


+

C ([0, ]). Also, un (x) = 0 for 0 n k 2 and vn () = 0 for all n 0.


Second, show by the same process that if
 1
1 k k
F (x, ) = e(x/) u(t)e(t/) dt, (3.115)
x

then there exists un (x) C ([0, 1]), and vn (X), wn (X) C ([0, ]) such
that

N 1
F (x, ) = n [un (x) + vn (x/) + wn ((1 x)/)] + O(N ) (3.116)
n=0

uniformly as 0+ , for 0 x 1, where = k . In particular, if x = 1 t,


then (x/)k (1/)k = (t/)(t), where 0 < (t) C ([0, 1]), and hence,
by Theorem 1,


N 1
k
+(1/)k
e(x/) = n n (t/)e(t/) + O(N ), (3.117)
n=0
46 3 Second Order Dierential Equations

where n (T ) is a polynomial, so, also, wn (X) = o(X ) as X .


Finally, suppose v(T ) C ([0, ]) and
 x
1 (x/)k k
G(x, ) = e v(t/)e(t/) dt, (3.118)
0
 1
k k
H(x, ) = 1 e(x/) v(t/)e(t/) dt. (3.119)
x

Show that G(x, ) = V (x/), where V (X) C ([0, ]), and show there
exists vn (X), wn (X) C ([0, ]) such that


N1
H(x, ) = n [vn (x/) + wn ((1 x)/)], (3.120)
n=0

where, as above, = k .

3.6. Show that if y(x, ) satises (3.41) when d(x, ) = 0, and if


x
1 a(t,) dt
(x, ) = e 2 0 , (3.121)

then y(x, ) = y(x, )/(x, ) satises

2 y  [xb(x) + 2 c(x, )]y = 0, (3.122)

where c(x, ) C ([0, 1] [0, o ]). Show further that if u(x4 ) C ([0, 1])
and v(X 4 ) C ([0, ]), then there exists un (x), vn (X)) such that un (x4 )
C ([0, 1]), vn (X 4 ) C ([0, ]) and, for any N 0,


N1
[u(x) + v(x/2/3 )](x, ) = n/6 [un (x) + vn (x/2/3 )] + O(N/6 ) (3.123)
n=0

uniformly as 0+ for 0 x 1. Therefore, since (3.59) is known to hold


when a(x, ) = 0, it also holds when a(x, ) = 0.

3.7. Show that if y(x, ) is the solution to

2 y  (x + 2x2 + 2 )y = 5 3x (3.124)

satisfying y(0, ) = 2 and y(1, ) = , then y(x, ) is approximately

c(x, ) = 1 v1 (x/) + [u(x) + v2 (x/) + w((1 x)/)], (3.125)

where = 2/3 ,
3.5 Exercises 47

10

0.0 0.2 0.4 0.6 0.8 1.0

Fig. 3.3 Numerical solution to (3.124) and asymptotic approximations of the solution
when = 0.1.

5P (0) 13 2
v1 (X) = Ai(X) 5P (X), u(x) = , w(X) = e 3X
Ai(0) 1 + 2x 3
(3.126)
and
11 2P (0)
v2 (X) = Ai(X) [XAi(X) X 2 Ai (X)]
Ai(0) Ai(0)
+2XP (X) 2X 2 P  (X) + 4. (3.127)

Figure 3.3 shows this approximation and Maples numerical solution for
y(x, ) when = 0.1. It also shows c(x, ) plus v3 (x/), the next term after
c(x, ) in the uniformly valid asymptotic expansion of y(x, ), which is signif-
icantly closer to the numerical solution.

3.8. The Bessel function J ((1 x)) is the solution to

(1 x)2 y  (1 x)y  2 (2x x2 )y = 0 (3.128)

satisfying y(0, ) = J () and, for > 0, y(1, ) = 0. Let = 21/2 1 and


use P robE together with the result of Exercise 1.6 to compute the rst several
terms of the remarkably simple result (compared, for example, to (9.3.35) of
[1], also on page 425 of [6], or the result in [8]),


N 1
(/2)1/3 J ((1 x)) = n vn (x/) + O(N ) (3.129)
n=0
48 3 Second Order Dierential Equations

uniformly as for 0 x 1, where = 21/3 2/3 and

vn (X) = pn (X)Ai(X) + qn (X)Ai (X), (3.130)

where pn (X), qn (X) are polynomials. In particular,


1 3 2 9
p0 (X) = 1, p1 (X) = X, p2 (X) = X + X 5, (3.131)
5 35 200

3 2 1 17
q0 (X) = 0, q1 (X) = X , q2 (X) = + X 3. (3.132)
10 35 70

In addition,
2 173 3 957
p3 (X) = + X + X 6, (3.133)
225 3150 14000

947 5903 23573 7 27


p4 (X) = X+ X4 + X + X 10 , (3.134)
173250 138600 294000 80000

37 611 4 9
q3 (X) = X+ X + X 7, (3.135)
1575 3150 2000

158 110767 5 549


q4 (X) = X2 + X + X 8. (3.136)
12375 693000 56000

3.9. Write a Maple program for Problem D and use it to verify (3.40). Also
verify that

4x2 1
u1 (x) = , v1 (X) = (5X 2 2X 2X 2 )e 2X , (3.137)
(2 + x)2 4

1
w1 (X) = (32 55X + 20 3X + 5 3X 2 )e 3X (3.138)
72
for this example.
Chapter 4
Logarithm Problems

4.1 Preliminaries

We saw in Chapter 1 that logarithms arose as a result of integrating a function


satisfying the conditions of Theorem 1. Suppose, for a general result, that
f (x, X, ) C ([0, 1] [0, ] [0, o ]) for some o > 0 and
 x
y(x, ) = x1 f (t, t/, ) dt. (4.1)
0

We know, given N 0, that


N 1
f (t, t/, ) = n [un (t) + vn (t/)] + O(N ) (4.2)
n=0

uniformly for 0 t 1 as 0+ for certain un (t) C ([0, 1]) and


vn (T ) C ([0, ]) with vn () = 0. If we let
 x
Un (x) = x1 un (t) dt, (4.3)
0

and  X
Vn (X) = X 1 [vn (T ) cn (1 + T )1 ] dT, (4.4)
0
[1]
where cn = vn (), then Un (x) C ([0, 1]) and Vn (X) C ([0, ]).
Therefore

N1
y(x, ) = n [Un (x) + Vn (x/) + cn L1 (x/)] + O(N ), (4.5)
n=0

L.A. Skinner, Singular Perturbation Theory, DOI 10.1007/978-1-4419-9958-0_4, 49


Springer Science+Business Media, LLC 2011
50 4 Logarithm Problems

uniformly as 0+ , where L1 (X) = X 1 ln(1 + X). Note that Vn () = 0


and 0 L1 (X) 1.
Repeated integration gives rise to powers of logarithms. Let L0 (X) = 1,
Ln (X) = X 1 lnn (1 + X) for integer n 1 and note that Ln (X) = O(1) for
0 X . If, for example, v(T ) C ([0, ]), v() = 0 and
 X
F (X) = X 1 v(T ) ln(1 + T ) dT, (4.6)
0

then  X
1
F (X) = cL2 (X) + X 1 (T ) ln(1 + T ) dT, (4.7)
2 0

where c = v [1] () and (T ) = v(T ) c(1 + T )1 = O(T 2 ). Therefore,


after an integration by parts,
1
F (X) = cL2 (X) + 1 (X)L1 (X) + 0 (X), (4.8)
2
where
  X
1
1 (X) = (T ) dT, 0 (X) = X 1 (T )(1 + T )1 dT. (4.9)
X 0

By induction, if
 X
F (X) = X 1 v(T ) lnn (1 + T ) dT, (4.10)
0

where n is a positive integer, there exist k (X) C ([0, ]) for 0 k n+1


such that

n+1
F (X) = k (X)Lk (X). (4.11)
k=0

In particular n+1 (X) = c/(n + 1). Also, k () = 0 for 0 k n.


As another example, suppose
 x
1 1 + t/
y(x, ) = x t2 u(t, ) ln dt (4.12)
0 1 + x/

for > 0, where u(x, ) C ([0, 1] [0, o ]) for some o > 0. If we integrate
by parts and dene  t
a(t, ) = t3 s2 u(s, ) ds, (4.13)
0

then a(t, ) C ([0, 1] [0, o ]) and, since

t3 3
= t2 t + 2 , (4.14)
t+ t+
4.1 Preliminaries 51

we have
 x  x
a(t, )
y(x, ) = x1 (t2 + t 2 )a(t, ) dt + 3 x1 dt. (4.15)
0 0 t+

Furthermore, from Exercise 1.7, there exist g(t, ) C ([0, 1] [0, o ]) and
c() C ([0, o ]) such that

a(t, ) c()
= g(t, ) + . (4.16)
t+ t+
Therefore

y(x, ) = x2 a0 (x, ) + xa1 (x, ) + 2 a2 (x, ) + 2 c()L1 (x/) (4.17)

for all (x, ) [0, 1] (0, o ], where


 x  x
a0 (x, ) = x3 t2 a(t, ) dt, a1 (x, ) = x2 ta(t, ) dt, (4.18)
0 0
 x
1
a2 (x, ) = x [a(t, ) + g(t, )] dt. (4.19)
0

In fact (see Exercise 4.1), for any integers n 0 and m 1, if


 x
1 + t/
y(x, ) = x1 tn u(t, ) lnm dt (4.20)
0 1 + x/

for > 0, where u(x, ) C ([0, 1] [0, o ]), then there exists ak (x, )
C ([0, 1] [0, o ]) for 0 k n and ck () C ([0, o ]) for 1 k m such
that
n 
m
y(x, ) = k xnk ak (x, ) + n ck ()Lk (x/) (4.21)
k=0 k=1

for all (x, ) [0, 1] (0, o ]. We will use this result to help establish the
asymptotic form of the solution to Problem F in the next section.
Another source of logarithms is exponentiation. For example, if y(x, ) =
1 + x/ for 0 x 1 and > 0, then


N 1
1
y (x, ) = e ln(1+x/) = 1 + x n Ln+1 (x/) + O(N ) (4.22)
n=0
(n + 1)!

uniformly for 0 x 1 as 0+ . Taking this a step further, if d()


C ([0, o ]), then there exists constants dnk such that


N 1 
n
y d() (x, ) = 1 + x n dnk Lk+1 (x/) + O(N ) (4.23)
n=0 k=0
52 4 Logarithm Problems

uniformly for 0 x 1 as 0+ . In particular,


1 2
d00 = d0 , d10 = d1 , d11 = d , d20 = d2 ,
2 0
1
d21 = d0 d1 d22 = d30 , (4.24)
6
where dn = d[n] (0). Also, by combining (4.2) and (4.22), observe that if

y(x, ) = (1 + x/) f (x, x/, ), (4.25)

where f (x, X, ) C ([0, 1] [0, ] [0, o ]), then there exists unm (x)
C ([0, 1]) and vnm (X) C ([0, ]) such that


N1 
n+1
y(x, ) = n [unm (x) + vnm (x/)]Lm (x/) + O(N ) (4.26)
n=0 m=0

uniformly for 0 x 1 as 0+ . For example, if

y(x, ) = (3 + + 3x + x2 + x/) , (4.27)

then (4.26) holds with

u00 (x) = 1, u01 (x) = x, (4.28)


1
u10 (x) = 0, u11 (x) = 0, u12 (x) = x, (4.29)
2
1
u20 (x) = 3 + x, u21 (x) = 2 + 3x + x2 , u22 (x) = 0, u23 (x) = x, (4.30)
6
and
v00 (X) = 0, v01 (X) = 0, (4.31)
3+X
v10 (X) = ln , v11 = 0, v12 (X) = 0, (4.32)
1+X
8 3+X
v20 (X) = + ln2 , (4.33)
3+X 1+X
3+X
v21 = 2 + X ln , v22 (X) = 0, v23 (X) = 0. (4.34)
1+X

4.2 Problem F

Assume a(x, ), b(x, ) C ([0, 1] [0, o ]) for some o > 0. In addition,


assume a(0, 0) = I, a positive integer. In this section we seek an asymptotic
expansion for the solution to
4.2 Problem F 53

(x + )y  + a(x, )y = b(x, ) (4.35)

satisfying y(0, ) = 0.
We know there exists d() C ([0, o ]) and g(x, ) C ([0, 1] [0, o ])
such that
a(x, ) I + d()
= g(x, ) + (4.36)
x+ x+
for > 0. Therefore
 x  d()
h(x, ) I1 1 + t/
y(x, ) = (1 + t/) (t, ) dt, (4.37)
(1 + x/)I 0 1 + x/

where
  x
h(x, ) = exp g(t, ) dt , (x, ) = b(x, )/h(x, ). (4.38)
0

Also, as in (4.23), there exists constants dnm , such that


 d() 
N 
n
1 + t/ 1 + t/
=1+ n
dnm lnm + O(N ) (4.39)
1 + x/ n=1 m=1
1 + x/

for 0 t x 1 and 0+ .
Let
 x
h(x, ) 1 + t/
wm (x, ) = (1 + t/)I1 (t, ) lnm dt. (4.40)
(1 + x/)I 0 1 + x/

Since  x
x1 (t/)k (t, ) dt = (x/)k k (x, ), (4.41)
0

where k (x, ) C ([0, 1] [0, o ]), we have w0 (x, ) = f00 (x, x/, ), where
I1 
Xh(x, )  I 1
f00 (x, X, ) = X k k (x, ), (4.42)
(1 + X)I k
k=0

and clearly f00 (x, X, ) C ([0, 1] [0, ] [0, o ]). Similarly, for m 1,
in view of the expression (4.21) for the integral (4.20), there exist k (x, )
C ([0, 1] [0, o ]) and ck () C ([0, o ]) such that
I1 
(x/)h(x, )  m
k
wm (x, ) = (x/) k (x, ) + ck ()Lk (x/) . (4.43)
(1 + x/)I
k=0 k=1

Hence, for every m 0, and 0 k m, there exists fmk (x, X, )


C ([0, 1] [0, ] [0, o ]) such that
54 4 Logarithm Problems


m
wm (x, ) = fmk (x, x/, )Lk (x/). (4.44)
k=0

Moreover, when we substitute this into


N 1 
n
y(x, ) = w0 (x, ) + n dmn wm (x, ) + O(N ), (4.45)
n=0 m=1

it is clear there exists nk (x, X, ) C ([0, 1] [0, ] [0, o ]), just linear
combinations of the fmk (x, X, ), such that


N 1 
n
y(x, ) = n nk (x, x/, )Lk (x/) + O(N ). (4.46)
n=0 k=0

Therefore, envisioning the expansion of each nk (x, x/, ) according to


Corollary 1, we see, nally, there exists unm (x) C ([0, 1]) and vnm (X)
C ([0, ]), with vnm () = 0, such that


N1 
n
y(x, ) = n [unm (x) + vnm (x/)]Lm (x/) + O(N ) (4.47)
n=0 m=0

uniformly as 0+ for 0 x 1.
Of course, we want to calculate the terms of (4.47) from the dierential
equation (4.35). It is apparent from (4.47) that y(x, ) has an inner expansion
of the familiar form,

N1
IN y(x, ) = n Yn (X), (4.48)
n=0

such that Y (X, ) = y(X, ) = IN y(x, ) + O(N ) uniformly on any nite


interval 0 X < as 0+ . Indeed, in somewhat tangled form,
1
N1n 

N 
n 
n k
IN y(x, ) = (X) unmk + vnm (X) Lm (X), (4.49)
n=0 m=0 k=0

[k]
where unmk = unm (0). In untangled form, we nd, with help from Maple,


n
Yn (X) = Znm (X)Lm (X), (4.50)
m=0

where,
Z00 (X) = u000 + v00 (X), (4.51)
Z10 (X) = u100 + u001 X + v10 (X), (4.52)
4.2 Problem F 55

Z11 (X) = u110 + v11 (X), (4.53)


2
Z20 (X) = u200 + u101 X + u002 X + v20 (X), (4.54)
Z21 (X) = u210 + u111 X + v21 (X), (4.55)
Z22 (X) = u220 + v22 (X), (4.56)
Z30 (X) = u300 + u201 X + u102 X 2 + u003 X 3 + v30 (X) (4.57)
2
Z31 (X) = u310 + u211 X + u112 X + v31 (X), (4.58)
Z32 (X) = u320 + u221 X + v32 (X), (4.59)
Z33 (X) = u330 + v33 (X). (4.60)
In general, Znm (X) = rnm (X) + vnm (X) and


nm
rnm (X) = rnmk X k , (4.61)
k=0

where rnmk = unk,m,k . Therefore


1n
N
unm (x) = rn+k,m,k xk + O(xN n ). (4.62)
k=0

In addition, (4.47) implies the existence of an outer expansion for y(x, ),


having the form


N1 
n1
ON y(x, ) = z00 (x) + n lnm ()znm (x), (4.63)
n=1 m=0

such that y(x, ) = ON y(x, ) + O(N lnN1 ()) uniformly on 0 < x 1


[k]
as 0+ . Indeed, in particular, using vnmk = vnm (), we have

z00 (x) = u00 (x), (4.64)

z10 (x) = v001 x1 + u10 (x), (4.65)


z20 (x) = v002 x2 + v101 x1 + x1 u11 (x) ln(x) + u20 (x), (4.66)
1
z21 (x) = x u11 (x), (4.67)

z30 (x) = v003 x3 + v102 x2 + v111 x2 ln(x) + u11 (x)x2


+v201 x1 + u21 (x)x1 ln(x) + u22 (x)x1 ln2 (x) + u30 (x), (4.68)

z31 (x) = x2 v111 u21 (x) 2u22 (x)x1 ln(x), (4.69)


1
z32 (x) = x u22 (x). (4.70)
56 4 Logarithm Problems

The rst part of the Maple program P robF , given below, calculates Yk (X)
for 0 k N 1, once a(x, ), b(x, ) and N are prescribed, in the usual
way from the dierential equation for Y (X, ). Then each Zij (X) for 0 i
N 1 and 0 j i is isolated and split into rij (X) and vij (X). Note that
this involves a procedure for expanding Zij (X) as X out to precisely
O(X N ). The resulting series expansions for uij (x) as x 0 and vij (X) as
X are used in the outer expansion calculations. For example, if we take
a(x, ) = 2 + x, b(x, ) = 1 and N = 4, P robF yields

1 6X 1
v00 (X) = , v10 (X) = , (4.71)
2(1 + X)2 12(1 + X)2

X 21X + 8
v11 (X) = , v20 (X) = , (4.72)
2(1 + X)2 36(1 + X)2
13X 6 X
v21 (X) = , v22 (X) = , (4.73)
12(1 + X)2 4(1 + X)2
and in accordance with (4.62), P robF also yields
1 1 1 1 3
u00 (x) = x + x2 x + O(x4 ), u11 (x) = O(x3 ), (4.74)
2 6 24 120
1 1 29 2
u10 (x) = x+ x + O(x3 ), u22 (x) = O(x2 ), (4.75)
12 18 144
1 1 2 131
u21 (x) = x + O(x2 ), u20 (x) = + x + O(x2 ). (4.76)
2 4 9 2160
In addition, the P robF substitution of (4.63) into (4.35), still with a(x, ) =
2 + x, b(x, ) = 1 and N = 4, yields the sequence of dierential equations
  
xz00 + (2 + x)z00 = 1, xz10 + (2 + x)z10 = z00 (x), (4.77)
  
xz20 + (2 + x)z20 = z10 (x), xz21 + (2 + x)z21 = 0, (4.78)
  
xz31 + (2 + x)z31 = z21 (x), xz32 + (2 + x)z32 = 0, (4.79)
and from the last lines of P robF we get

z00 (x) = u00 (x), z10 (x) = u10 (x) (4.80)


1 1
z20 (x) = x + x1 ln(x)u11 (x) + u20 (x), z21 (x) = x1 u11 (x), (4.81)
2
1
z31 (x) = x2 x1 u21 (x) 2x1 ln(x)u22 (x), (4.82)
2
u32 (x) = x1 u22 (x), (4.83)
in accordance with (4.64)-(4.70).
4.2 Problem F 57

To complete the calculations for this example, from (4.80a) and (4.74a)
we see z00 (0) = 1/2 and therefore the solution to (4.77a) is
1 x
u00 (x) = (e 1 x) (4.84)
x2
Next, the solution to (4.77b) such that z10 (0) = u10 (0) = 1/12 is

ex 2
u10 (x) = [E(x) + (ex 1 x)], (4.85)
x2 x
x
where E(x) = 0
t1 (et 1) dt. In addition, we readily nd

1
u11 (x) = 0, u22 (x) = 0, u21 (x) = (1 ex ), (4.86)
2x
and nally, the (convergent) series solution to (4.78a) yields
2 131 977 2 4367 3
u20 (x) = + x x + x + O(x4 ). (4.87)
9 2160 8640 302400
Thus we have computed all the terms of (4.47) with N = 3, assuming
a(x, ) = 2 + x, b(x, ) = 1.

P robF := proc(a, b, N )
A := subs(x = X, a); B := subs(x = X, b);
IN y := sum(n Yn , n = 0..N 1);
IN dy := sum(n dYn , n = 0..N 1);
IN de := series((1 + X) IN dy + A IN y B, = 0, N );
IN de := convert(%, polynom);
for k from 0 to N 1 do
temp := coe (IN de, , k);
de := subs(Yk = z(X),dYk = di (z(X), X), temp) = 0;
dsolve({de, z(0) = 0});
Y k := rhs(%); dY k := di (Y k, X);
IN de := subs(Yk = Y k, dYk = dY k, IN de);
Yk := Y k; end do;
for i from 0 to N 1 do
Zi,0 := coe (Yi , ln(1 + X), 0);
for j from 1 to i do
Zi,j := X coe (Yi , ln(1 + X), j);
end do; end do;
for i from 0 to N 1 do
for j from 0 to i do
if Zi,j = 0 then sZij := 0 else
temp := 0; k := N 1;
while order(temp) < N do
58 4 Logarithm Problems

k := k + 1; sZij := series(Zi,j , X = , k);


temp := series(subs(X = X 1 , %), X = 0, k);
end do; end if ;
ri,j := convert(series(sZij, X = , 1), polynom);
vij := simplif y(Zi,j ri,j ); print(vi,j = vij);
svi,j := sZij ri,j ;
end do; end do;
for i from 0 to N 1 do
for j from 0 to i do
uij := sum(xl coe (ri+l,j , X, l), l = 0..N 1 i) + O(xN i );
print(ui,j = uij);
end do; end do;
ON y := z0,0 + sum(n sum(ln()m zn,m , m = 0..n 1), n = 0..N 1);
ON dy := Dz0,0 +sum(n sum(ln()m Dzn,m , m = 0..n1), n = 0..N 1);
eq := (x + ) ON dy + a ON y b;
ON de := series(eq, = 0, N );
for i from 0 to N 1 do;
for j from 0 to i do;
dei := coe (ON de, , i); deij := coe (dei, ln(), j); print(deij = 0);
end do; end do; L0 := 1;
for i from 1 to N 1 do Li := X 1 ln(1 + X)i ; end do;
y := sum(n (sum(Lm (un,m + svn,m ), m = 0..n)), n = 0..N 1);
ON y := series(subs(X = x 1 , y), = 0, N );
for n from 0 to N 1 do
zn := coe (ON y, , n);
end do;
print(z0,0 = z0 );
for n from 1 to N 1 do;
for m from 0 to n 1 do;
lnzn,m := coe (zn , ln(), m); print(zn,m = lnzn,m );
end do; end do;
end proc:

4.3 Problem G, Part 1

Problem G is to derive an asymptotic expansion for the solution to the dif-


ferential equation

2 y  + [xa(x) + 2 b(x, )]y  + c(x, )y = d(x, ) (4.88)

that satises the boundary conditions y(0, ) = (), y(1, ) = (). We


will assume b(x, ), c(x, ), d(x, ) C ([0, 1] [0, o ]) and (), ()
C ([0, o ]) for some o > 0. We also assume 0 < a(x) C ([0, 1]) and,
4.3 Problem G, Part 1 59

without loss of generality, take a(0) = 2, () = 0. First we will obtain


asymptotic expansions for two initial value problems, namely,

2 w + [xa(x) + 2 b(x, )]w + c(x, )w = d(x, ) (4.89)

subject to w(0, ) = 0, w (0, ) = 0, and

2 z  + [xa(x) + 2 b(x, )]z  + c(x, )z = 0 (4.90)



subject to z(0, ) = 0, z  (0, ) = (2/ )1 . In the end we will determine an
asymptotic expansion for the function () such that

y(x, ) = w(x, ) + ()z(x, ). (4.91)

This approach to Problem G is essentially the same as the one described in


Exercise 3.4 for Problem C.
It is reasonable to expect the solution to

2 w0 + xa(x)w0 = d(x, ) (4.92)

such that w0 (0, ) = w0 (0, ) = 0, the problem obtained by neglecting


2 b(x, )w + c(x, )w in (4.89), uniformly approximates w(x, ). An even
better approximation should be the solution to the problem obtained by sub-
stituting 2 b(x, )w0 (x, )+c(x, )w0 (x, ) for 2 b(x, )w +c(x, )w. We are
going to prove that in fact


w(x, ) = n wn (x, ), (4.93)
n=0

where, for n 1,

2 wn + xa(x)wn = b(x, )wn1



(x, ) c(x, )wn1 (x, ). (4.94)

First, there exists B > 0 such that |b(x, )|, |c(x, )|, |d(x, )| B for all
(x, ) [0, 1] [0, o ]. Therefore |w0 (x, )| BF (x, ) on [0, 1] (0, o ],
where  x
2 2
F (x, ) = 1 e(x/) (x)
e(t/) (t)
dt, (4.95)
0
 x
2
(x) = x ta(t) dt. (4.96)
0

Note that (x) > 0 for 0 x 1. If we let = t[(t)]1/2 and denote


the inverse of this transformation by t = ( ), and let = x[(x)]1/2 , then
|  ( )| = 2[(t)]1/2 /[t + 2(t)] M for some M > 0, so F (x, ) M P (/),
still on [0, 1] (0, o ], where
60 4 Logarithm Problems
 Z
2 2
P (Z) = eZ eT dT. (4.97)
0

Also, P (Z) = O(Z) as Z 0 and P (Z) = O(Z 1 ) as Z . Hence, there


exists K > 0 such that P (Z) KZ/(1 + Z 2 ) for 0 Z . In addition,

/ x/
2
= f (x, x/), (4.98)
1 + (/) 1 + (x/)2

where
(1 + X 2 )[(x)]1/2
f (x, X) = , (4.99)
1 + X 2 (x)
so f (x, x/) = O(1) on [0, 1] (0, o ], by Theorem 1. Therefore F (x, )
A(x/)/[1 + (x/)2 ] for some constant A > 0. Furthermore, X/(1 + X 2 )
2/(1 + X) for 0 X . Therefore
 x
1 dt
|w0 (x, )| 2AB 2AB ln(1 + 1/) (4.100)
0 1 + t/

for all (x, ) [0, 1](0, o ]. In addition, X/(1+X 2 ) 2 ln(1+X). Therefore


|w0 (x, )| 2AB ln(1 + 1/), too.
Next, we now know

|b(x, )w0 (x, ) + c(x, )w0 (x, )| 4AB 2 ln(1 + 1/) (4.101)

for 0 x 1 and 0 < o . Consequently, |w1 (x, )| 4AB 2 ln(1 +


1/)F (x, ), and therefore both |w1 (x, )| and |w1 (x, )| are bounded by
8A2 B 2 ln2 (1+1/). Indeed, by induction, if we let C = 4AB, then for any n
0, both |wn (x, )| and |wn (x, | are less than or equal to (1/2)C n+1 lnn+1 (1+
1/) for all (x, ) [0, 1] (0, o ]. So, of the three sums,




n wn (x, ), n wn (x, ), n wn (x, ), (4.102)
n=0 n=0 n=0

we know the rst two converge absolutely and uniformly for all (x, )
[0, 1] (0, o ], assuming o is set suciently small, and therefore the second
sum is the derivative of the rst. The third sum also converges absolutely
and uniformly, and therefore is the second derivative of the rst sum. This is
seen by summing both sides of (4.94) times n , and this shows simultaneously
that the rst sum in (4.102) is indeed the desired solution of (4.89).
To ascertain the asymptotic form of w(x, ), we need to examine the terms
of (4.93) in more detail. Let m (x, ) = xLm (x/) for m 1 and 0 (x, ) = 1.
Proposition 1. Assume f (x, X, ) C ([0, 1] [0, ] [0, o ]) and let
 x
1 (x/)2 (x) 2
gm (x, ) = e e(t/) (t) f (t, t/, )m (t, ) dt, (4.103)
0
4.3 Problem G, Part 1 61

where 0 < (x) C ([0, 1]). There exists mnk (x) C ([0, 1]) and
mnk (X) C ([0, ]) with mnk () = 0 such that for any N 0,


N 1 
m
n
gm (x, ) = [mnk (x) + mnk (x/)]k (x, ) + O(N ) (4.104)
n=0 k=0

uniformly as 0+ for 0 x 1. In particular, m,0,k (x) = 0 for 0 k


m and, for m 1, m,0,0 (X) = m,1,0 (x) = 0.

Proof. We know this is true for m = 0 from our work on Problem E. For
m 1, we readily get (4.104) from (4.103) using integration by parts. In
particular,
g1 (x, ) = g0 (x, )1 (x, ) g0 (x, ), (4.105)
where g0 (x, ) is the same as g0 (x, ), except g0 (t, )/(1 + t/) occurs in place
of f (t, t/, ). Therefore (4.104) holds for m = 1 with 1,0,1 (x) = 0 and
1,0,0 (x) = 1,0,0 (X) = 1,1,0 (x) = 0. For m 2,

gm (x, ) = g0 (x, )m (x, ) gm1 (x, ), (4.106)

where gm1 (x, ) is the same as gm1 (x, ), except mg0 (t, )/(1 + t/) oc-
curs in place of f (t, t/, ). Therefore m,0,m (x) = 0 and if we assume
m1,0,k (x) = 0 for 0 k m 1, m1,0,0 (X) = 0 and m1,1,0 (x) = 0,
then m,0,k (x) = 0 for 0 k m 1, m,0,0 (X) = 0 and m,1,0 (x) = 0.

It follows from Proposition 1 with m = 0 that


N 1
w0 (x, ) = n [u0,n,0 (x) + v0,n,0 (x/)] + O(N ), (4.107)
n=0

where u0,n,0 (x) C ([0, 1]), v0,n,0 (X) C ([0, ]) and u0,0,0 (x) = 0.
Therefore, as in the rst paragraph of Section 4.1,


N 1 1

w0 (x, ) = n [U0,n,k (x) + V0,n,k (x/)]k (x, ) + O(N ) (4.108)
n=0 k=0

uniformly as 0+ for 0 x 1, where U0,n,0 (x) C ([0, 1]), V0,n,0 (X)


C ([0, ]) and, in particular, U0,0,0 (x) = V0,0,0 (X) = 0 and U0,n,1 (x) +
V0,n,1 (x/) = cn,1 , a constant. To proceed further, we need a more general
result.

Proposition 2. Assume f (x, X, ) C ([0, 1] [0, ] [0, o ]) for some


o > 0. Assume also that f (x, , 0) = 0 and let
 x
gm (x, ) = f (t, t/, ) lnm (1 + t/) dt. (4.109)
0
62 4 Logarithm Problems

There exists mnk (x) C ([0, 1]) and mnk (X) C ([0, ]) such that


N1 
m+1
gm (x, ) = n [mnk (x) + mnk (x/)]k (x, ) + O(N ) (4.110)
n=0 k=0

uniformly as 0+ for 0 x 1. In particular, m,0,0 (x) = m,0,0 (X) = 0


and for each n 0, m,n,m+1 (x) + m,n,m+1 (x/) = cn,m+1 , a constant.

Proof. This follows from the second paragraph of Section 4.1, in particular,
from our analysis of (4.20) with n = 0, or, independently, using integration
by parts. We leave the details as an exercise.

At this point, from (4.107) and (4.108), and Corollary 1, we can see

[b(x, )w0 (x, ) + c(x, )w0 (x, )] =



N1 1

n [nk (x) + nk (x/)]k (x, ) + O(N ), (4.111)
n=0 k=0

where nk (x) C ([0, 1]), nk (X) C ([0, ]) and 0,0 (x) = 0,0 (X) =
0. Therefore, by Proposition 1, there exists u1,n,k (x) C ([0, 1]) and
v1,n,k (X) C ([0, ]) such that, after canceling an factor,


N1 1

w1 (x, ) = n [u1,n,k (x) + v1,n,k (x/)] lnk (1 + x/) + O(N ). (4.112)
n=0 k=0

In particular, u1,0,k (x) = 0 for k = 0, 1. Thus we can apply Proposition 2 to


this result, and therefore


N1 2

w1 (x, ) = n [U1,n,k (x) + V1,n,k (x/)]k (x, ) + O(N ), (4.113)
n=0 k=0

where, in particular, U1,0,0 (x) = V1,0,0 (X) = 0. Obviously, we can iterate this
process indenitely. Hence we conclude, there exists Umnk (x) C ([0, 1])
and Vnmk (X) C ([0, ]) such that for every m 0,


N 1 
m+1
wm (x, ) = n [Umnk (x) + Vnmk (x/)]k (x, ) (4.114)
n=0 k=0

uniformly as 0+ for 0 x 1. It is also true, by virtually the same


reasoning, that z(x, ), our solution to (4.90), has an innite series solution,


z(x, ) = m zm (x, ), (4.115)
m=0
4.4 Problem G, Part 2 63

and zm (x, ) has an asymptotic expansion of the same form as wm (x, ).


See Exercise 4.5.

4.4 Problem G, Part 2

We can see upon substituting (4.114) into (4.93) that if we let w(x,  ) =
2 w(x, ), then there exists unm (x) C ([0, 1]) and vnm (X) C ([0, ])
such that, for any N 0,


N1 
n
 ) =
w(x, n [unm (x) + vnm (x/)]Lm (x/) + O(N ) (4.116)
n=0 m=0

uniformly as 0+ for 0 x 1. That is, w(x,


 ) has the same asymptotic
form as the solution to Problem F. Hence, if we denote the N -term expansion
 (X, ) = w(X,
of W  ) by


N1
 ) =
IN w(x, n (X),
n W (4.117)
n=0

then, as with Yn (X) in Section 4.2,


n
n (X) =
W Znm (X)Lm (X) (4.118)
m=0

and for the leading terms we have (4.51)-(4.60). Likewise, for the N-term
outer expansion


N1 
n1
 ) = z00 (x) +
ON w(x, n lnm ()znm (x), (4.119)
n=1 m=0

 ) is
we have (4.64)-(4.70). The dierential equation for w(x,

2 w
 + [xa(x) + 2 b(x, )]w
 + c(x, )w
 = 2 d(x, ), (4.120)

of course, and it is readily determined that



xa(x)z00 = 0, (4.121)

xa(x)z10 = c0 (x)z00 (x), (4.122)

xa(x)z21 =0 (4.123)
  
xa(x)z20 = b0 (x)z00 (x)z00 (x)c0 (x)z10 (x)c1 (x)z20 (x)+d0 (x), (4.124)
64 4 Logarithm Problems


xa(x)z32 = 0, (4.125)

xa(x)z31 = c0 (x)z21 (x), (4.126)
where b0 (x) = b(x, 0), c0 (x) = c(x, 0), d0 (x) = d(x, 0) and c1 (x) = c[0,1] (x, 0).
From the dierential equation for W  (X, ), it is apparent that

  + 2X W
W  = 0 (4.127)
0 0

n (0) = W
and therefore, since W n (0) = 0 for all n 0, W
0 (X) = 0. From
(4.51), rst with X = , this means

u00 (0) = 0, v00 (X) = 0. (4.128)

It follows also that


  + 2X W
W   = 0, (4.129)
1 1

1 (X) = 0, too, and therefore, by (4.52) and (4.53),


so W

u10 (0) = u00 (0) = u11 (0) = 0, v10 (X) = v11 (X) = 0. (4.130)

Now it is apparent that


W   = d00 ,
2 + 2X W (4.131)
2

where d00 = d(0, 0). Therefore   (X)


W = d00 P (X), where P (X) = + 1 1
X
2 2
O(X 3 ) is the function dened by (4.97). Hence, if we let

1
Q(X) = [P (T ) (1 + T )1 ] dT, (4.132)
X 2

then Q(X) C ([0, ]), in particular, Q(X) = 12 X 1 +O(X 2 ) as X ,


and
2 (X) = d00 [Q(0) Q(X)] + 1 d00 ln(1 + X).
W (4.133)
2
It follows that Z20 (X) = d00 [Q(0) Q(X)] and therefore

u20 (0) = d00 Q(0), u10 (0) = u00 (0) = 0, v20 (X) = d00 Q(X). (4.134)

Also, Z21 (X) = 12 d00 X and Z22 (X) = 0, so

1
u11 (0) = d00 , u21 (0) = u22 (0) = 0, v21 (X) = v22 (X) = 0. (4.135)
2
 ), from (4.121), (4.64) and
Turning to the outer expansion of w(x,
(4.128a), we see
z00 (x) = u00 (x) = 0. (4.136)
Therefore from (4.122), along with (4.65) and (4.128b), we also have
4.4 Problem G, Part 2 65

z10 (x) = u10 (x) = 0. (4.137)

Similarly, from (4.123), (4.130a) and (4.67), we nd


1 1
z21 (x) = d00 , u11 (x) = d00 x. (4.138)
2 2

To determine u20 (x), we currently have xa(x)z20 = d0 (x) from (4.124) and
1
z20 (x) = 2 d00 ln(x) + u20 (x) from (4.66). Therefore, noting (4.134a), we nd

u20 (x) = d00 Q(0) + D(x), (4.139)

where D(x) C ([0, 1]) is given by


 x 
d0 (t) d00 dt
D(x) = . (4.140)
0 a(t) 2 t

At this point we know

u00 (x) + v00 (x/) = 0, (4.141)

u10 (x) + v10 (x/) = 0, (4.142)


1
u11 (x) + v11 (x/) = d00 x, (4.143)
2
u20 (x) + v20 (x/) = d00 [Q(x/) Q(0)] + D(x). (4.144)
We also know v21 (X) = v22 (X) = 0. To determine u21 (x) and u22 (x), we
need to calculate Z31 (X) and Z32 (X). For this we have

  + 2X W
W   = d01 + d10 X c00 W
2 (X) (b00 + a1 X 2 )W
  (X), (4.145)
3 3 2

2 (X) is
where b00 = b(0, 0), c00 = c(0, 0), a1 = a (0), dij = d[i,j] (0, 0) and W
given by (4.133). If we denote the right side of (4.145) by f (X), then
 X
3 (X) = eX 2
W
2
eT f (T ) dT (4.146)
0

and
1
f (X) = + X c00 d00 ln(1 + X) + (X), (4.147)
2
where
1
= d01 c00 d00 Q(0), = (2d10 d00 a1 )X (4.148)
2
and (X) C ([0, ]). Indeed,
1
(X) = c00 d00 Q(X) b00 d00 P (X) a1 d00 [2X 2 P (X) X], (4.149)
2
66 4 Logarithm Problems

and, in particular, () = 0. Consequently,


 X
2 2
eX eT (T ) dT = O(X 2 ). (4.150)
0

Also
 X
2 2
eX eT ln(1 + T ) dT = P (X) ln(1 + X) + (X), (4.151)
0

where  X
2 2 dT
(X) = eX eT P (T ) , (4.152)
0 1+T
so (X) = O(X 3 ) and (X) C ([0, ]). Therefore

  (X) = P (X) + 1 1 c00 d00 P (X) ln(1 + X) + (X),


W (4.153)
3
2 2
where (X) C ([0, ]) and (X) = O(X 2 ).
We already know
 X
1
P (T ) dT = Q(0) Q(X) + ln(1 + X). (4.154)
0 2

In addition, it is readily determined that


 X
1 2
P (T ) ln(1 + T ) dT = R(X) Q(X) ln(1 + X) + ln (1 + X), (4.155)
0 4

where R(X) C ([0, ]). Therefore (4.153) implies


1 1 1
Z31 (X) = X + c00 d00 XQ(X), Z32 (X) = c00 d00 X (4.156)
2 2 8
and thus from (4.58) and (4.59), we now know
1 1
u21 (0) = , u22 (0) = c00 d00 . (4.157)
2 8
Using these two results in conjunction with (4.125) and (4.126), we readily
nd
1 1
u21 (x) + v21 (x/) = x d00 xC(x), (4.158)
2 2
and
1
u22 (x) + v22 (x) = c00 d00 x, (4.159)
8
where
4.5 Problem G, Part 3 67
 x 
c0 (t) c00 dt
C(x) = . (4.160)
0 a(t) 2 t

4.5 Problem G, Part 3

When we substitute the results of the last section, beginning with (4.141),
into (4.116), and also note that ln(1 + 1/) = ln(1/) + O(), we get, for the
solution of (4.89) with w(0, ) = w (0, ) = 0, the uniformly valid expansion

w(x, ) = ln(1/)w11 (x, ) + w20 (x, )


+ ln2 (1/)w22 (x, ) + ln(1/)w21 (x, ) + O() (4.161)

for 0 x 1, where

1 ln(1 + x/)
w11 (x, ) = d00 , (4.162)
2 ln(1 + 1/)

w20 (x, ) = D(x) + d00 Q(x/), (4.163)


1 ln2 (1 + x/)
w22 (x, ) = c00 d00 2 , (4.164)
8 ln (1 + 1/)
1 ln(1 + x/)
w21 (x, ) = [d01 c00 d00 Q() d00 C(x)] (4.165)
2 ln(1 + 1/)
The function z(x, ) = 2 z(x, ), where z(x, ) is the desired solution to
 ).
(4.90), has an expansion of the same form as the expansion (4.116) for w(x,
This follows from our remarks at the end of Section 4.3. When we calculate
the terms of this expansion, the result corresponding to (4.161) is

z(x, ) = z20 (x, ) + ln(1/)z21 (x, ) + O() (4.166)

uniformly as 0+ for 0 x 1, where

z20 (x, ) = erf(x/), (4.167)

1 ln(1 + x/)
z21 (x, ) = c00 . (4.168)
2 ln(1 + 1/)
Returning now to (4.91), to determine (), we have w(1, )+()z(1, ) =
() and a straightforward calculation, utilizing 1/z(1, ) = 1+ 12 c00 ln(1/)+
O(), reveals
68 4 Logarithm Problems

0.8

0.4

0.0

0.4

0.0 0.2 0.4 0.6 0.8 1.0

Fig. 4.1 Numerical solution to (4.88) and an asymptotic approximation to the solution
when a(x, 0) = 2 + x, b(x, = 0, c(x, ) = cos(x), d(x, ) = 1 + x + , y(0, ) = 0, y(1, ) = 1
and = 0.1
.

1
() = d00 ln(1/) + [0 d00 Q() D(1)]
2
1
c00 d00 ln2 (1/) + O( ln(1/), (4.169)
8
where 0 = (0). When we substitute this, along with the above expansions
for w(x, ) and z(x, ) into (4.91), we get, nally, for the solution to Problem
G,

y(x, ) = ln(1/)y11 (x, ) + y20 (x, ) + ln2 (1/)y22 (x, )


+ O( ln(1/)), (4.170)

where  
1 ln(1 + x/)
y11 (x, ) = d00 erf(x/) , (4.171)
2 ln(1 + 1/)
y20 (x, ) = [0 d00 Q() D(1)]erf(x/) + d00 Q(x/) + D(x), (4.172)
 
1 ln(1 + x/) ln2 (1 + x/)
y22 (x, ) = c00 d00 2 2 erf(x/) , (4.173)
8 ln(1 + 1/) ln (1 + 1/)
and this holds uniformly for 0 x 1 as 0+ .
A comparison of the numerical solution of (4.88) and the asymptotic result
(4.170)-(4.173), assuming a(x) = 2 + x, b(x, ) = 0, c(x, ) = cos(x), d(x, ) =
1+x+ and () = 1, in which case D(x) = 12 ln(1+x/2), is shown for = 0.1
4.6 Exercises 69

in Figure 4.1. Note also that if instead of the boundary condition


y(1, ) =
() for (4.88), we impose the initial condition y  (0, ) = (2/ )()1 , in
addition to y(0, ) = 0, then

y(x, ) = ln(1/)w11 (x, ) + [w20 (x, ) + 0 z20 (x, )] + ln2 (1/)w22 (x, )
+ ln(1/)[w21 (x, ) + 0 z21 (x, )] + O(). (4.174)

4.6 Exercises

4.1. Assume a(x, ) C ([0, 1] [0, o ]) for some o > 0, let m 0, n 1


be integers and, for > 0, let
 x n
t a(t, ) m 1 + t/
y(x, ) = ln dt. (4.175)
0 t+ 1 + x/

The problem is to show there exists ak (x, ) C ([0, 1] [0, o ]) for 0 k


n 1 and ck () C ([0, o ]) for 1 k m + 1 such that


n1 
m+1
y(x, ) = k xnk ak (x, ) + n ck () lnk (1 + x/) (4.176)
k=0 k=1

for all (x, ) [0, 1] (0, o ]. Do this rst for m = 0 and then, by induction,
for arbitrary m 0. Now use this to show that (4.21) follows from (4.20),
where n 0, m 1.

4.2. Use the Maple routine of Section 1.5 to conrm the expansion results
stated at the end of Section 4.1 for y(x, ) given by (4.27).

4.3. In Problem F, let a(x, ) = cos(x), b(x, ) = 1 + sin(x). Using P rob F,


or a simplied version, show that in (4.47), by computing suciently many
terms of just the inner expansion of y(x, ), we get both
1 1
v00 (X) = , v10 (X) = , v11 (X) = 0, (4.177)
1+X 2(1 + X)
4 1
v20 (X) = , v21 (X) = , v22 (X) = 0, (4.178)
12(1 + X) 2(1 + X)
and
1 1 1 1 4 1 5
u00 (x) = 1 + x + x2 + x3 + x x + O(x6 ), (4.179)
2 6 48 120 2880
1 5 5 1 17 4 79 5
u10 (x) = x x2 x3 x x + O(x6 ), (4.180)
2 12 48 20 2880 60480
70 4 Logarithm Problems

11 11 17 169 3 388 4 97
u20 (x) = + x+ x2 + x + x + x5 +O(x6 ), (4.181)
12 48 90 5760 302400 120690
1 1 1
u21 (x) = x2 x4 + O(x6 ), (4.182)
2 8 96
and evidently u11 (x) = u22 (x) = 0. At least, u11 (x) = O(x6 ), u22 (x) = O(x6 ).

4.4. Work out the details for the proof of Proposition 2.

4.5. Show that the solution to (4.90) satisfying z(0, ) = 0, z  (0, ) =


(2/ )1 has an innite series solution of the same form as the one for
w(x, ), and from there show that z(x, ) = 2 z(x, ) has a uniformly valid
 ). Compute the rst several
asymptotic expansion of the same form as w(x,
terms of this expansion and conrm the results (4.166)-(4.168).
Chapter 5
Oscillation Problems

5.1 Problem H

To begin this chapter, suppose

y  + a(x, )y = b(x, ) cos(x/), (5.1)

where a(x, ), b(x, ) C ([0, 1] [0, o ]) for some o > 0. Assume also that
a(x, ) > 0 and y(0, ) = 0. We know from Problem A that
 x
y(x, ) = 1 f (x, t, t/, ) cos((x t)/) dt, (5.2)
0

where
f (x, t, T, ) = b(x t, )eT u(x,t,) , (5.3)
with  x
u(x, t, ) = t1 a(s, ) ds, (5.4)
xt

and we could expand f (x, t, t/, ) in accordance with Corollary 2, but this
would not be immediately useful. Instead, assuming b(x, ) is real valued, like
a(x, ) > 0, we note from (5.2) that y(x, ) = Re[w(x, )], where w(x, ) =
z(x, )eix/ and, if we let

(x, t, T, ) = b(x t, )eT v(x,t,) , (5.5)

where
v(x, t, ) = u(x, t, ) + i, (5.6)
then  x
z(x, ) = 1 (x, t, t/, ) dt. (5.7)
0

L.A. Skinner, Singular Perturbation Theory, DOI 10.1007/978-1-4419-9958-0_5, 71


Springer Science+Business Media, LLC 2011
72 5 Oscillation Problems

Now, since Re[v(x, t, )] = u(x, t, ) > 0 for all (x, t, ) [0, 1] [0, x] [0, o ],
we have (x, t, T, ) C ([0, 1] [0, x] [0, ] [0, o ]), and therefore, by
Corollary 2, as in Problem A, there exists un (x) C ([0, 1]) and vn (X)
C ([0, ]) with vn () = 0 such that for any N 0,


N 1
z(x, ) = n [un (x) + vn (x/)] + O(N ) (5.8)
n=0

uniformly as 0+ for 0 x 1. Furthermore, we can get the terms of this


expansion, again as in Problem A, from the dierential equation for z(x, ),
which is
z  + [a(x, ) + i]z = b(x, ). (5.9)
Thus we immediately see

b(x, 0)
O1 z(x, ) = (5.10)
a(x, 0) + i

and upon introducing X = x/ in (5.9), we readily nd

b(0, 0)  
I1 z(x, ) = 1 e[a(0,0)+i]x/ . (5.11)
a(0, 0) + i

Consequently,

b(x, 0) b(0, 0) [a(0,0)+i]X


u0 (x) = , v0 (X) = e , (5.12)
a(x, 0) + i a(0, 0) + i

and from here it follows that

y(x, ) = p0 (x) cos(x/) + q0 (x) sin(x/) + r0 (x/) + O(), (5.13)

where
a(x, 0)b(x, 0) b(x, 0)
p0 (x) = , q0 (x) = , (5.14)
1 + [a(x, 0)]2 1 + [a(x, 0)]2
and
r0 (X) = p0 (0)ea(0,0)X . (5.15)
In general, if y(x, ) is the solution to (5.1), where a(x, ), b(x, ) are real
valued, a(x, ) > 0 and y(0, ) = 0, then

y(x, ) = p(x, ) cos(x/) + q(x, ) sin(x/) + r(x, ), (5.16)

where

N 1 
N 1
p(x, ) = n pn (x) + O(N ), q(x, ) = n qn (x) + O(N ), (5.17)
n=0 n=0
5.1 Problem H 73

and

N1
r(x, ) = n rn (x/) + O(N ) (5.18)
n=0

all uniformly as 0+ for 0 x 1, where pn (x), qn (x) C ([0, 1]),


and rn (X) C ([0, ]). Furthermore, in specic cases, the terms of these
expansions can be determined with the help of P robA, as in Exercise 5.1.
The asymptotic expansion of an integral such as (5.7), where (x, t, T, )
is given by (5.5), is rather dierent if Re[v(x, t, )] = 0. We will need the
proposition below in the next section.

Proposition 3. Assume g(x), h(x) C ([0, 1]) are real valued, h(0) = 0,
and  x
F (x, ) = 1 eih(t)/ g(t) dt. (5.19)
0

(a). If h (x) = 0 for 0 x 1, then there exists c(), un (x) C ([0, 1])
such that
F (x, ) = c() + (x, )eih(x)/ , (5.20)
where, for any N 0,


N1
(x, ) = n un (x) + O(N ) (5.21)
n=0

uniformly as 0+ for 0 x 1.

(b). If h (x) > 0 for 0 < x 1, h (0) = 0 and h (0) > 0, then there ex-
ists c(), un (x) C ([0, 1]) and vn (X) C ([0, ]) such that

F (x, ) = c() + (x, )eih(x)/ (5.22)

where = 1/2 and, for any N 0,


N 1
(x, ) = n [un (x) + vn (x/)] + O(N ) (5.23)
n=0

uniformly as 0+ for 0 x 1.

Proof. If h (x) = 0 for 0 x 1, then u0 (x) = ig(x)/h (x) is in C ([0, 1])


and an integration by parts yields

F (x, ) = u0 (0) u0 (x)eih(x)/ + F1 (x, ), (5.24)

where  x
F1 (x, ) = 1 eih(x)/ u0 (t) dt. (5.25)
0
74 5 Oscillation Problems

But this means

F1 (x, ) = u1 (0) u1 (x)eih(x)/ + F2 (x, ), (5.26)

where u1 (x) = iu0 (x)/h (x) and


 x
1
F2 (x, ) = eih(t)/ u1 (t) dt. (5.27)
0

Obviously we can repeat this process indenitely, and this proves Part (a) of
the proposition. Indeed, we have un (x) = iun1 (x)/h (x) for n 1 and we
can choose c() to be any function in C ([0, 1]) such that c[n] (0) = un (0).
To prove Part (b), let s = [h(x)]1/2 , let x = (s) denote the inverse of this
transformation, and let G(s) = g((s)) (s). Then
 [h(x)]1/2
2
F (x, ) = 1 ei(s/) [G(0) + 2isG1 (s)] ds, (5.28)
0

where G1 (s) = (2is)1 [G(s) G(0)], and an integration by parts reveals

F (x, ) = H(0, ) H([h(x)]1/2 , ) + F1 (x, ), (5.29)

where 2
H(x, ) = eix /
[ (x/)G(0) + G1 (x)], (5.30)

ix2 2
(x) = e eis ds, (5.31)
x
and
 [h(x)]1/2
2
1
F1 (x, ) = ei(s/) [G1 (0) + 2isG2 (s)] ds. (5.32)
0

where G2 (s) = (2is)1 [G1 (s) G1 (0)]. So again we have a process we can
iterate indenitely. Furthermore, (x) C ([0, ]), as we see in Exercise
5.2, and therefore we can expand ([h(x)]1/2 /) = ((x/)[h(x)/x2 ]1/2 ) ac-
cording to Theorem 1. This leads directly to (5.22)-(5.23).

5.2 Problem I, Part 1

For our nal problem, we seek an asymptotic expansion for the solution to

2 y  + a(x, )y = 1/2 b(x, ) cos(x/) (5.33)


5.2 Problem I, Part 1 75

satisfying the initial conditions y(0, ) = (), y  (0, ) = 1 (). The func-
tions a(x, ), b(x, ) C ([0, 1] [0, o ]) and (), () C ([0, o ]) for
some o > 0 are assumed to be real, and we assume a(x, ) > 0. We further
assume, for some xo (0, 1), that a00 (x) < 1 for 0 x < xo and a00 (x) > 1
for xo < x 1, where, in general, aij (x) = a[i,j] (x, 0), bij (x) = b[i,j] (x, 0).
Finally, we assume a10 (xo ) = 0.
If (instead) a(x, ) = (1 + c)2 , where c is a constant, and b(x, ) = 1/2 ,
then
()
y(x, ) = () cos[(1 + c)(x/)] + sin[(1 + c)(x/)] + yp (x, ), (5.34)
1 + c
where, in the particular solution

yp (x, ) = M (x, ) sin(x/) + N (x, ) cos(x/), (5.35)

assuming |c| << 1/, the amplitudes

sin(cx) 1 cos(cx)
M (x, ) = , N (x, ) = (5.36)
c(2 + c) c(2 + c)

oscillate slowly compared to sin(x/) and cos(x/), unless c = 0, in which


case M (x, ) = x/2, N (x, ) = 0. The point of Problem I is see what happens
if a(x, ) comes close to 1 but only momentarily equals 1 as x increases from
0 to 1.
As in Problem H, it is convenient to let w(x, ) be the solution to (5.33)
with eix/ in place of cos(x/), so that y(x, ) = Re[w(x, )]. In analogy with
Problem D, as in [11], the homogeneous equation for w(x, ) has a pair of
linearly independent solutions,
(1) (2) (1)
wh (x, ) = eis(x)/ u(x, ), wh (x, ) = wh (x, ), (5.37)
(1) (1)
where wh (x, ) is the complex conjugate of wh (x, ),
 x
s(x) = [a00 (t)]1/2 dt, (5.38)
0

and u(x, ) has a uniformly valid asymptotic expansion,


N1
u(x, ) = n un (x) + O(N ), (5.39)
n=0

for 0 x 1, where un (x) C ([0, 1]). In particular, if we take u(0, ) = 1,


it is readily determined that

u0 (x) = [a00 (0)/a00 (x)]1/4 eir(x) , (5.40)


76 5 Oscillation Problems

where  x
1
r(x) = [a00 (t)]1/2 a01 (t) dt. (5.41)
2 0

Furthermore, the Wronskian of these two solutions is 1 (), where ()


is pure imaginary and also has an asymptotic expansion in powers of . In
particular, (0) = 2i[a00 (0)]1/2 . It follows by the method of variation of
parameters that the full equation for w(x, ) has a particular solution of the
form
(1) (2)
wp (x, ) = A(x, )wh (x, ) + B(x, )wh (x, ) (5.42)
where  x
1/2
A(x, ) = ei[s(t)t]/ (t, ) dt, (5.43)
xo
 x
1/2
B(x, ) = ei[s(t)+t]/ (t, ) dt, (5.44)
xo

with (x, ) = b(x, )u(x, )/().


If we substitute t = t + xo in (5.43) and let h(t ) = s(t + xo ) s(xo ) t,
then h (t ) > 0 for 0 < t 1 xo , h(0) = h (0) = 0 and h (0) = 12 a (xo ) > 0.
Therefore, by Part (b) of Proposition 3,

A(x, ) = ei[s(xo )xo ]/ [c() + (x, )eih(x)/ ], (5.45)

where = 1/2 , x = x xo , c() C ([0, 1]) and for any N 0,


N 1
(x, ) = n [un (x) + vn (x/)] + O(N ) (5.46)
n=0

uniformly as 0+ for 0 x 1 xo , that is, xo x 1, where


un (x) C ([0, 1 xo ]), vn (X) C ([0, ]). Furthermore, [s(xo ) xo ]
h(x) = s(x) + x and therefore, by an application of Theorem 1, there exists
cA () and A (x, ) such that
(1) (1)
A(x, )wh (x, ) = cA ()wh (x, ) + A (x, )eix/ , (5.47)

where A (x, ) has a uniformly valid expansion of the same form as (x, ).
Similarly, by Part (a) of Proposition 5.1, we nd
(2) (2)
B(x, )wh (x, ) = cB ()wh (x, ) + B (x, )eix/ , (5.48)

where B (x, ) also has a uniformly valid expansion like (x, ), although
there are no vn (x/) terms. Consequently, when we substitute (5.47) and
(5.48) into (5.42) we see there is a particular solution w(+) (x, ) = z (+) (x, )eix/
of the dierential equation for w(x, ), in addition to wp (x, ), such that, for
(+) (+)
certain un (x) C ([0, 1 xo ]), vn (X) C ([0, ]),
5.3 Problem I, Part 2 77


N 1
z (+) (x, ) = n [u(+) (+) N
n (x) + vn (x/)] + O( ) (5.49)
n=0

uniformly as 0+ for 0 x 1 xo .
The dierential equation for z (+) (x, ) is

4 z (+) + 2i2 z (+) + [a(xo + x, 2 ) 1]z (+) = b(xo + x, 2 ) (5.50)

and for the rst term of the N -term outer expansion,


N1
ON z (+) (x, ) = n zn(+) (x), (5.51)
n=0

(+)
whose existence is implied by (5.49), it is apparent that z0 (x) = 0. Also,
for the leading term of the N -term inner expansion of z (+) (x, ),


N1
IN z (+) (x, ) = n Zn(+) (X), (5.52)
n=0

it is clear from the dierential equation

2 Z (+) + 2iZ (+) + [a(xo + X, 2 ) 1]Z (+) = b(xo + X, 2 ) (5.53)

for Z (+) (X, ) = z (+) (xo + X, ) that


(+) (+)
2iZ0 + a10 (xo )XZ0 = b00 (xo ). (5.54)

Thus, in terms of the function (x) C ([0, ]) introduced in the proof of


Proposition 3, and further discussed in Exercise 5.2, in order for O1 I1 z (+) (x, )
to agree with the fact that I1 O1 z (+) (x, ) = 0, it must be that

(+) ib00 (xo )


Z0 (X) = (mo X), (5.55)
2mo

where mo = 12 [a10 (xo )]1/2 . In other words, in (5.49),


(+) (+) (+)
u0 (x) = 0, v0 (X) = Z0 (X). (5.56)

5.3 Problem I, Part 2

If we put t = xo t in (5.43) and let h(t) = s(xo t) s(xo ) + t, then by


another application of Proposition 3, Part (b), we see, in addition to (5.45),
that
78 5 Oscillation Problems

A(x, ) = ei[s(xo )xo ]/ [d() + (x, )eih(x)/ ], (5.57)



where x = xo x, d() C ([0, 1]) and (x, ) has a uniformly valid asymp-
totic expansion for 0 x xo of the same form as the one for (x, )
on 0 x 1 xo . A similar statement applies to B(x, ) and thus we
see, in addition to the particular solution w(+) (x, ) = z (+) (x, )eix/ to
the dierential equation for w(x, ), there is another particular solution,
()
w () (x, ) = z () (x, )eix/ , such that, for certain un (x) C ([0, xo ]),
()
vn (X) C ([0, ]),


N1
z () (x, ) = n [u() () N
n (x) + vn (x/)] + O( ), (5.58)
n=0

uniformly as 0+ for 0 x xo . Furthermore, it is readily determined


from the dierential equation for z () (x, ), in analogy with the calculations
for z (+) (x, ), that
() () (+)
u0 (x) = 0, v0 (X) = v0 (X). (5.59)

To solve now for y(x, ) = Re[w(x, )], rst we note that, since the dif-
ference between any two particular solutions of the dierential equation for
w(x, ) is a solution to the corresponding homogeneous dierential equation,
there exists c1 (), c2 () such that
(1) (2)
w() (x, ) = w(+) (x, ) + c1 ()wh (x, ) + c2 ()wh (x, ). (5.60)

Furthermore, to determine c1 () and c2 () we have, in view of (5.56) and


(5.59),
w() (xo , ) w(+) (xo , ) = [2Z () (0) + O()]eixo / , (5.61)
and also
()
w() (xo , ) w(+) (xo , ) = 1 [2iZ0 (0) + O()]eixo / . (5.62)

In addition,
(1)
wh (xo , ) = [u0 (xo ) + O()]eis(xo )/ , (5.63)
(1)
wh (xo , ) = 1 [iu0 (xo ) + O()]eis(xo )/ , (5.64)
(2) (1)
and, of course, wh (xo , ) = wh (xo , ). Therefore, from
(5.60) and its deriva-
tive evaluated at x = xo , and the fact that (0) = ( /2)ei/4 , what we nd
is
b00 (xo ) i{3/4+[xo s(xo )]/r(xo )}
c1 () = e + O() (5.65)
2mo [a00 (0)]1/4
and c2 () = O(). Consequently,
5.3 Problem I, Part 2 79

Re[w() (x, )] = Re[w(+) (x, )] + [0 (x, ) + O()] cos[r(x) + s(x)/]


+ [0 (x, ) + O()] sin[r(x) + s(x)/], (5.66)

where

b00 (xo )
0 (x, ) = cos{3/4 [xo s(xo )]/ + r(xo )}, (5.67)
2mo [a00 (x)]1/4

b00 (xo )
0 (x, ) = sin{3/4 [xo s(xo )]/ + r(xo )}. (5.68)
2mo [a00 (x)]1/4
Similarly, the dierence between the particular solution w() (x, ) and the
one that satises the initial conditions w(0, ) = (), w (0, ) = 1 () also
is a solution to the corresponding homogeneous dierential equation. Thus
we nd, for the solution to Problem I,

y(x, ) = Re[w() (x, )] + [0 (x) + O()] cos[r(x) + s(x)/]


+ [0 (x) + O()] sin[r(x) + s(x)/], (5.69)

where, with 0 = (0) and 0 = (0),

0 (x) = 0 [a00 (0)/a00 (x)]1/4 , 0 (x) = 0 [a00 (0)a00 (x)]1/4 . (5.70)

To deal with Re[w() (x, )] and Re[w(+) (x, )], note that (x) = p(x) +
iq(x), where 
p(x) = (cos x2 cos s2 + sin x2 sin s2 ) ds, (5.71)
x

q(x) = (sinx2 cos s2 cosx2 sin s2 ) ds. (5.72)
x
It follows that
b00 (xo ) 
Re[w() (x, )] = [q(mo (x0 x)/) + O()] cos(x/)
2mo

+ [p(mo (xo x)/) + O()] sin(x/) (5.73)

for 0 x xo , and

b00 (xo ) 
Re[w(+) (x, )] = [q(mo (x xo )/) + O()] cos(x/)
2mo

+ [p(mo (x xo )/) + O()] sin(x/) . (5.74)

for xo x 1. Finally then, when we combine everything, we have


80 5 Oscillation Problems

y(x, ) = A(x, ) cos[r(x) + s(x)/] + B(x, ) sin[r(x) + s(x)/]


+ C(x, ) cos(x/) + D(x, ) sin(x/), (5.75)

where A(x, ) has a uniformly valid expansion


N 1
A(x, ) = n An (x, ) + O(N ) (5.76)
n=0

stemming from (5.49), (5.58) and (5.39), in which each An (x, ) = O(1) for
0 x 1, and there are corresponding expansions for B(x, ), C(x, ) and
D(x, ). In particular, we have determined

A0 (x, ) = 0 (x) + H(x xo )0 (x, ), (5.77)

B0 (x, ) = 0 (x) + H(x xo )0 (x, ), (5.78)


and
b00 (xo )
C0 (x, ) = q(mo |x xo |/)sgn(x xo ) (5.79)
2mo
b00 (xo )
D0 (x, ) = p(mo |x xo |/)sgn(x xo ), (5.80)
2mo
where H(x) = 0 for x < 0, H(x) = 1 for x > 0, and sgn(x) = H(x) H(x).
In Figure 5.1, a graph of the uniformly valid O(1) approximation

y0 (x, ) = A0 (x, ) cos[r(x) + s(x)/)] + B0 (x, ) sin[r(x) + s(x)/]

2
0.0 0.2 0.4 0.6 0.8 1.0
Fig. 5.1 Uniformly valid asymptotic approximation of the solution to (5.33) when
a(x, ) = 1/4 + 2x + , b(x, ) = 1, y(0, ) = 1), y (0, ) = 0 and = 0.001.
5.4 Exercises 81

2
0.0 0.2 0.4 0.6 0.8 1.0

Fig. 5.2 Outer approximation of the approximate solution to (5.33) shown in Figure 5.1.

+ C0 (x, ) cos(x/) + D0 (x, ) sin(x/) (5.81)

is shown for the case


1
a(x, ) = + 2x + , b(x, ) = 1, () = 1, () = 0, = 0.001.
4
(5.82)
Although Maples plotting routine obviously is being pushed to the limit here,
it is easy to check that the result in this case is virtually indistinguishable
from Maples numerical solution of (5.33) . Also, on any subinterval of [0, 1]
excluding the point x = xo , the inner expansion terms C0 (x, ) and D0 (x, )
are uniformly O() and therefore y(x, ) = y00 (x, ) + O(), where

y00 (x, ) = A0 (x, ) cos[r(x) + s(x)/)] + B0 (x, ) sin[r(x) + s(x)/]. (5.83)

A graph of the (outer) approximation y00 (x, ), again assuming (5.82), is


shown in Figure 5.2. In general, the amplitude jump in y00 (x, ) at x = xo
is from [0 (xo )2 + 0 (xo )2 ]1/2 to the square root of [0 (xo ) + 0
(xo )]2 +
2
[0 (xo ) + 0 (x
o , )] , and in the case of (5.82), this jump is from 2/2 to
[(1 + )/2 + cos(3/4 997/12)]1/2 .

5.4 Exercises

5.1. Specic solutions to Problem H quickly become unwieldy. For example,


if a(x, ) = 2 + x + and b(x, ) = 1 + x2 , using P robA with
82 5 Oscillation Problems

a := 2 + x + + I; b := 1 + x2 (5.84)

to solve (5.9) for the terms of (5.8) yields

1 + x2
u0 (x) = (5.85)
x+2+i
and therefore, in agreement with (5.14),

x3 + 2x2 + x + 2 x2 + 1
p0 (x) = , q0 (x) = . (5.86)
x2 + 4x + 5 x2 + 4x + 5
But then
x3 + 3x2 + 5x + 1 + i(x2 + 2x + 1)
u1 (x) = (5.87)
x3+ 6x2 + 9x + 2 + i(3x2 + 12x + 11)
and, with the help of Maples evalc command, this implies

x6 + 9x5 + 35x4 + 78x3 + 95x2 + 53x + 13


p1 (x) = , (5.88)
x6 + 12x5 + 63x4 + 184x3 + 315x2 + 300x + 125
2x5 + 13x4 + 40x3 + 70x2 + 54x + 9
q1 (x) = . (5.89)
x6
+ 12x5 + 63x4 + 184x3 + 315x2 + 300x + 125
On the other hand, the P robA calculation immediately reveals
2 1
r0 (X) = e2X r1 (X) = (925X 2 + 50X + 13))e2X . (5.90)
5 125
5.2. Let (x) be the function introduced in the proof of Proposition 3. Use
integration by parts to show (x) has an asymptotic expansion in pow-
ers of x1 as x , and therefore (x) C ([0, ]). Show also that
(0) = 12 1/2 ei/4 .
(+)
5.3. In Problem I, from the solution of the dierential equation for z1 (x),
show that
(+) b00 (xo + x) b00 (xo )
u1 (x) = , (5.91)
a00 (xo + x) 1 a10 (xo )x
(+) (+) (+)
v1 (X) = Z1 (X) Z1 (), (5.92)
and
(+) b10 (xo ) b00 (xo )a20 (xo )
Z1 () = . (5.93)
a10 (xo ) [a10 (xo )]2
() (+) () (+)
Show also that u1 (x) = u1 (x) and v1 (X) = v1 (X).

5.4. Show that


5.4 Exercises 83
 x
2 2 i 2
eix (s)eis ds = x (x) + (1 eix ), (5.94)
0 2
 x
2 2 1 3 2
eix s2 (s)eis ds = [2x (x) + ix2 + 1 eix ]. (5.95)
0 6
(+)
and use these results to determine Z1 (X) from (5.53).
References

1. M. Abramowitz and I. A. Stegun, eds., Handbook of Mathematical Functions, National


Bureau of Standards, 1964. Reprinted by Dover, New York, 1970.

2. W. Eckhaus, Asymptotic Analysis of Singular Perturbations, North-Holland, Amster-


dam, 1979.

3. L. E. Fraenkel, On the method of matched asymptotic expansions, Parts I, II and III,


Proc. Cambridge Phil. Soc., 65 (1969), 209-231, 233-261 and 263-284.

4. A. M. Ilin, Matching of Asymptotic Expansions of Solutions of Boundary Value


Problems, Am. Math. Soc., Providence, Rhode Island, 1991.

5. J. Kevorkian and J. D. Cole, Multiple Scale and Singular Perturbation Methods,


Springer-Verlag, New York, 1996.

6. F. W. J. Olver, Asymptotics and Special Functions, Academic Press, New York, 1974.

7. R. E. OMalley, Jr., Singular Perturbation Methods for Ordinary Dierential Equa-


tions, Springer-Verlag, New York, 1991.

8. L. A. Skinner, Matched asymptotic expansions of integrals, IMA J. Appl. Math., 50


(1993), 77-90.

9. L. A. Skinner, Matched expansion solutions of the rst-order turning point problem,


SIAM J. Math. Anal, 25 (1994), 1402-1411.

10. L. A. Skinner, Asymptotic solution to a class of singularly perturbed Volterra integral


equations, Methods and Applications of Analysis, 2 (1995), 212-221.

11. L. A. Skinner, Stationary phase theory and passage through resonance, J. Math.
Anal. Appl., 205 (1997), 186-196.

12. Donald R. Smith, Singular-perturbation theory, Cambridge University Press, Cam-


bridge, 1985.

13. Milton Van Dyke, Perturbation Methods in Fluid Mechanics, Parabolic Press, Stanford
California, 1975.

L.A. Skinner, Singular Perturbation Theory, DOI 10.1007/978-1-4419-9958-0, 85


Springer Science+Business Media, LLC 2011