AMATH732 Course Notes
AMATH732 Course Notes
K.G. Lamb
1
Department of Applied Mathematics
University of Waterloo
November 17, 2010
1
c K.G. Lamb, September 2010
Contents
Preface iii
1 Introduction 1
1.0.1 The Role of Numerical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.0.2 Numerical Noise vs. Physical Noise . . . . . . . . . . . . . . . . . . . . . . . . 3
1.0.3 Perturbation Theory and Asymptotic Analysis in Applied Mathematics . . . 4
2 Simple linear systems and roots of polynomials 7
2.1 Introduction and simple linear systems . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Roots of polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.1 Order of the error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.2 Sometimes you dont expand in powers of . . . . . . . . . . . . . . . . . . . 13
2.2.3 Solving by rescaling: a singular perturbation problem . . . . . . . . . . . . . 15
2.2.4 Finding the singular root: Introduction to the method of dominant balance . 16
2.3 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3 Nondimensionalization and scaling 19
3.1 Nondimensionalizing to get . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2 More on Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.3 Orthodoxy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.4 Example: Inviscid, compressible irrotational ow past a cylinder . . . . . . . . . . . 31
4 Resonant Forcing and Method of Strained Coordinates: Another example from
Singular Perturbation Theory 35
4.1 The simple pendulum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5 Asymptotic Series 41
5.1 Asymptotics: large and small terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.2 Asymptotic Expansions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.2.1 The Exponential Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.2.2 Asymptotic Sequences and Asymptotic Expansions (Poincare 1886) . . . . . 47
5.2.3 The Incomplete Gamma Function . . . . . . . . . . . . . . . . . . . . . . . . 50
6 Asymptotic Analysis for 2
nd
order ODEs 53
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.2 Finding the behaviour near an Irregular Singular Points: Method of Carlini-Liouville-
Green . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6.2.1 Finding the Leading Behaviour . . . . . . . . . . . . . . . . . . . . . . . . . . 55
i
6.2.2 Further Improvements: corrections to the leading behaviour. . . . . . . . . . 58
6.3 The Airy Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
6.4 Asymptotic Relations for Oscillatory Functions . . . . . . . . . . . . . . . . . . . . . 61
6.5 The Turning Point Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
6.5.1 WKB Theory: Outer Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.5.2 Region of Validity for WKB Solution . . . . . . . . . . . . . . . . . . . . . . . 67
6.5.3 Inner Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6.5.4 Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6.5.5 Summary of asymptotic solution . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.5.6 Physical Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
6.6 Tunneling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
7 Singular Perturbation Theory: Examples and Techniques 71
7.1 More examples of problems from Singular Perturbation Theory . . . . . . . . . . . . 71
7.2 The linear damped oscillator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
7.3 Method of Multiple Scales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
7.3.1 The Simple Nonlinear Pendulum . . . . . . . . . . . . . . . . . . . . . . . . . 81
7.4 Methods for Singular Perturbation Problems . . . . . . . . . . . . . . . . . . . . . . 84
7.4.1 Method of Strained Coordinates (MSC) . . . . . . . . . . . . . . . . . . . . . 84
7.4.2 The Linstedt-Poincare Technique . . . . . . . . . . . . . . . . . . . . . . . . . 86
7.4.3 Free Self Sustained Oscillations in Damped Systems . . . . . . . . . . . . . . 90
7.4.4 MSC: The Lighthill Technique . . . . . . . . . . . . . . . . . . . . . . . . . . 92
7.4.5 The Pritulo Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
7.4.6 Comparison of Lighthill and Pritulo Techniques . . . . . . . . . . . . . . . . . 97
8 Matched Asymptotic Expansions 99
9 Asymptotics used to derive model equations: derivation of the Korteweg-de
Vries equation
for internal waves 105
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
9.1.1 Streamfunction Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
9.1.2 Boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
9.2 Nondimensionalization and introduction of two small parameters . . . . . . . . . . . 108
9.3 Asymptotic expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
9.3.1 The O(1) problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
9.3.2 The O() problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
9.3.3 The problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
9.3.4 The x . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
9.3.5 The O() problem revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
Appendix A: USEFULL FORMULAE 115
Solutions to Selected Problems 117
ii
Preface
These notes are based to a large degree on lectures notes developed by P. Tenti. Duncan Mowbray
LaTeXd a rst draft of a large chunk of these notes.
Useful references
1. Course notes.
2. Bender, C. M., and Orzag, S. A. Advanced Mathematical Methods for Scientist and Engineers,
McGraw-Hill (1978). QA371.B43 1978.
The book I learned from. A wealth of topics. Lots and lots of problems. Some very dicult.
3. Lin, C. C., and L. A. Segel, L. A. Mathematics Applied to Deterministic Problems in the
Natural Sciences, MacMillan (1974). QA37.2.L55 1974
This is one of the best applied mathematics texts available. It was reprinted by SIAM in
1988. Parts of it can be read on google books.
4. Murdock, J. A. Perturbations: Theory and Methods. QA871.M87 1999
5. Nayfeh, A. H. Introduction to Perturbation Techniques. QA371.N32 1981
6. Bleistein, N., and Handelsman, R. A. Asymptotic Expansions of Integrals, Holt, Rinehart and
Winston, New York (1975). QA311.B58 1975
Title says it all. Dover republished an unabridged corrected version in 1986.
7. Ablowitz, M. J., and Fokas, A. S. Complex Variables: Introduction and Applications, Cam-
bridge University Press (2003). QA331.7.A25
Useful treatment of asymptotic evaluation of integrals (e.g., method of steepest descent).
iii
Chapter 1
Introduction
Before the 18
th
century, Applied Mathematics and its methods received the close attention of
the best mathematicians who were driven by a desire to explain the physical universe. Applied
Mathematics can be thought of as a three step process:
Physical 1 Mathematical
Situation = Formulation
_
_
_2
Physical 3 Solution by Purely
Interpretation = Formal Operations
of the Solution of the Math Problem
Over the centuries step 2 took on a life of its own. Mathematics was studied on its own, devoid
of contact with a physical problem. This is pure mathematics. Applied mathematics deals with all
three steps.
The goal of asymptotic and perturbation methods is to nd useful, approximate solutions to
dicult problems that arise from the desire to understand a physical process. Exact solutions
are usually either impossible to obtain or too complicated to be useful. Approximate, useful solu-
tions are often tested by comparison with experiments or observations rather than by rigourous
mathematical methods. Hence we will not be concerned with rigorous proofs in this course. The
derivation of approximate solutions can be done in two dierent ways. First, one can nd an ap-
proximate set of equations that can be solved, or, one can nd an approximate solution of a set of
equations. Usually one must do both.
A key turning point in the history of mathematics was the brilliant discovery of the theory of
limits of Gauss (17771855) and Cauchy (17891857). In the limit process, usually characterized
by an innite expansion, we do not attempt to obtain the exact solution but merely to approach
it with arbitrary precision. Thus, the desire for absolute accuracy (zero error) was replaced by one
for arbitrarily great accuracy (arbitrarily small error):
absolute
accuracy
=
arbitrarily great
accuracy
and
zero
error
=
arbitrarily small
error
.
1
We are no longer interested in what happens after a nite number of steps but wish to know
what happens eventually if the number of steps is increased indenitely. The obvious diculty
with this is that in most real applications you can only sum up a nite number of terms. In fact,
for many problems that we will tackle we will obtain only the rst two or three terms in a series.
We are then not particularly interested in what happens as the number of terms goes to zero but
rather in how accurate, or useful, an approximation using a few terms is. Since observations have
limited accuracy, there is no need to make the error arbitrarily small.
This gives rise to a dierent limiting process and dierent questions: What error occurs after
a nite number of steps? How can we minimize the error for a given number of steps? This is a
branch of applied analysis.
1.0.1 The Role of Numerical Analysis
An obvious question, particularly in this day and age, is If the problem is so dicult why not solve
it on a computer. Ultimately you may end up doing this, but using asymptotic and perturbation
techniques to nd useful, approximate answers is an extremely important rst step. It should always
be done whenever possible. Approximate solutions have many benets. They provide necessary
checks, and aid in the understanding and interpretation, of numerical solutions. They illuminate
potential problems, e.g., regions in parameter space where singularities exist and where special
numerical approaches may be required. They can give tremendous insight into how the solution
depends on the parameters of the problem and help determine what the important parameters are.
Example 1.0.1 Newtonian, constant density, steady state ow past a nite object R
3
.
_
_
_
u
u =
1
p +
2
u
u|
= 0
u u
as |x|
(1.1)
where
u(x, y, z) = uid velocity
0
= mass density
p = hydrodynamic pressure
u
= 0 (this condition, which says that the uid velocity is zero on the solid
boundary, is a consequence of viscosity). This results in a linear potential problem
u =
2
= 0
n = 0 on
(1.2)
This approximate linear problem can be solved for some geometries and many general results
can be proved as much is known about solutions of Laplaces equation. This makes it very tempting
to use the simplied problem (1.2). In fact researchers in the late 1800s and early 1900s used this
model and proved that airplanes cant y!
Today there is a strong tendency to solve problems like (1.1) on a computer. This can be a lot
of work and if the mathematical model does not correctly describe the physics then the numerical
solution is garbage no matter how accurately you solve the model equations. In fact (1.1) is useful
only for laminar ows (e.g., ow over a streamlined body like an airplane wing) because the model
is very inaccurate for turbulent ows.
Computers, while very useful and often necessary, should be used in the last stage of a scientic
investigation. Analytic work on a mathematical problem is necessary to provide a rough under-
standing of possible solutions. Phases 1 and 3 must be considered even in cases where we think we
already have a good mathematical model at our disposal. It is here that perturbation theory has
proved invaluable.
1.0.2 Numerical Noise vs. Physical Noise
Example 1.0.2 (C. Lanczos) Solving the 2 2 linear system
x +y = 2.00001
x + 1.00001y = 2.00002
_
(1.3)
we obtain the solution
x = 1.00001
y = 1.
Suppose that the values on the R.H.S. were obtained from measurements which have limited
accuracy. Suppose they are accurate to 10
3
.
Someone else takes the measurement and gets:
x +y = 2.001 (1.4)
x + 1.00001y = 2.002. (1.5)
Solving yields the solution (x, y) = (97.999, 100). A very dierent solution! The diculty here is
that in this system of equations x +y is well represented but x y is poorly represented. Setting
=
1
2
(x +y) and =
1
2
(x y), (1.6)
3
gives the system
2 = 2.00001 (1.7)
2.00001 0.00001 = 2.00002. (1.8)
The rst equation immediately gives . Changing the right-hand side by a tiny amount will change
the solution by a tiny amount. In this sense x +y = 2 is well represented. To get we will need
to divide by 10
5
, resulting in
= 10
5
_
2.00001 2.00002
_
. (1.9)
Thus, the value of is very sensitive to small changes in the measured values.
Here the problem is very simple to understand, but suppose we had a large system and went to
the computer to nd the solution. Roundo error would play havoc giving completely erroneous
results. The exact numerical solution of a mathematical problem may have no physical signicance.
Exercise: Write (1.3) in matrix form as
Ax = s, (1.10)
where x = (x, y)
T
. What are the eigenvalues of the matrix A and how do they imply sensitivity of
the solution to the source term s?
1.0.3 Perturbation Theory and Asymptotic Analysis in Applied Mathematics
Most mathematical problems facing applied mathematicians, scientists, and engineers have features
which preclude exact solutions. Some of these features include:
nonlinear terms in the equations
variable coecients
nonlinear boundary conditions at known boundaries
linear or nonlinear boundary conditions at unknown boundaries
Perturbation Theory (PT) is the collective name for a group of techniques developed for the
purpose of deriving approximate solutions, valid in certain limiting cases which are helpful in un-
derstanding the essential processes in simple terms. These often serve as benchmarks for fully
numerical solutions. They often have highly accurate predictive capability even when applied out-
side the range of conditions for which the method is justied. Approximate solutions obtained by
perturbation theory usually consist of the rst two or three terms of a certain series expansion in
the neighbourhood of a point at which the solution has an essential singularity. Asymptotic and
perturbations methods can be helpful in several ways. First, they can help by directly nding an
approximate solution to your problems. Secondly, these methods can be used to nd approxima-
tions to exact solutions which are dicult to understand (e.g., solutions written in terms of Bessel
functions of large or complicated arguements, or in terms of elliptic function). A third approach
is to use asymptotic methods to derive simpler problems which can then be solved exactly (or
approximately using perturbation and asymptotic methods again!).
The series obtained by perturbation and asymptotic methods is usually divergent and ordinary
results from calculus do not apply. Asymptotic Analysis is the new branch of analysis developed
to study such series.
4
Perturbation Theory has its origin in celestial mechanics. From Newtonian Mechanics it is
known that the motion of a celestial body, (e.g. the Earth) is specied by
M x
i
= F
(0)
i
+F
(1)
i
+
2
F
(2)
i
+ , (1.11)
for i = 1, 2, 3, where the F
(j)
(x
1
, x
2
, x
3
, t) represent the gravitational forces emanating from other
bodies. F
(0)
is the largest force, due to the sun.
The other terms, F
(1)
,
2
F
(2)
, . . . are successively smaller forces due to the moon and other
planets. These other forces are perturbations of the main force due to the sun. In particular, 1
is a small parameter.
In about 1830 Poisson suggested looking for a solution of (1.11) in a series of powers of :
x
i
(t) = x
(0)
i
(t) +x
(1)
i
(t) +
2
x
(2)
i
(t) + , (1.12)
the reasoning behind this being that the solution is a function of as well as time t: x
i
= x
i
(t, ).
Substituting this expansion into (1.11), expanding the F
(j)
i
s in power series of ,
F
(0)
i
(x
(0)
+x
(1)
+
2
x
(2)
+ , t)
= F
(0)
i
(x
(0)
, t)
+
F
(0)
i
(x
(0)
, t) [x
(1)
+
2
x
(2)
+ ]
+ ,
(1.13)
and equating like powers of gives a series of ODEs to solve.
The rst, obtained from the coecients of
0
, is
M x
(0)
i
= F
(0)
i
(x
(0)
1
, x
(0)
2
, x
(0)
3
, t) i = 1, 2, 3.
This is called the reduced equation or the reduced problem. It is obtained by setting = 0. One
must be able to solve the reduced problem in order to proceed.
Before Poincare (18591912) the mathematical status of perturbation series of the form (1.12)
was rarely considered. One could rarely nd more than a few terms, let alone determine if the
series converged or not. Indeed, it was often not known whether a solution existed or not.
Poincare shifted the attention from the convergence of a power series, such as
n=1
n
x
(n)
(t)
where the emphasis is on the limiting behaviour of
N
n=1
n
x
(n)
(t) as N for xed and t,
to the new concept of asymptotic analysis of nding the limiting behaviour of
N
n=1
n
x
(n)
(t) as
0 or t for xed N.
5
6
Chapter 2
Simple linear systems and roots of
polynomials
2.1 Introduction and simple linear systems
Reference: Lin & Segel
The general idea behind perturbation theory is the following:
(A) Non-dimensionalize the problem to introduce a small parameter, traditionally called or .
(B) Estimate the size of the terms in your model and drop small ones obtaining a reduced
problem.
(C) Solve the reduced problem.
(D) Compute perturbative corrections.
Basic Simplication Procedure (BSP): Set = 0 to get the reduced problem. Solve.
Example 2.1.1 (From Lin & Segel, page 186): Solve approximately
x + 10y = 21,
5x +y = 7,
(2.1)
for = 0.01.
Solution:
(A) Step (A) is already done: equation nondimensionalized and a small parameter has been
introduced.
(B) The Basic Simplication Procedure assumes that the presence of a small parameter in the
coecient of a term indicates that that term is small. Using the BSP, we set = 0 to get the
reduced problem, giving
10y
0
= 21,
5x
0
+y
0
= 7,
(2.2)
where we have introduced x
0
and y
0
to denote the approximate solution.
7
(C) The reduced problem is easily solved giving
(x
0
, y
0
) = (0.98, 2.1). (2.3)
(D) We next nd perturbative corrections. The most common approach in perturbation theory
is the following. The solution of the system (2.16) depends on . Denote the solution by
(x, y) = (x(), y()) and assume a Taylor Series for x() and y() exists:
x() = x
0
+x
1
+
2
x
2
+ ,
y() = y
0
+y
1
+
2
y
2
+ .
(2.4)
Substituting these expansions into (2.16) gives
0
(10y
0
21) +(x
0
+ 10y
1
) +
2
(x
1
+ 10y
2
) + = 0,
0
(5x
0
+y
0
) +(5x
1
+y
1
) +
2
(5x
2
+y
2
) + = 0.
(2.5)
Since these equations should be satised for all in a neighbourhood of 0, the coecient of
each power of must be zero. Thus we get a sequence of problems:
(a) The O(1) terms (those with coecient
0
) give
10y
0
21 = 0,
5x
0
+y
0
= 0.
(2.6)
This is the reduced problem we have already solved.
(b) The O() terms give
x
0
+ 10y
1
= 0,
5x
1
+y
1
= 0.
(2.7)
From this we nd
y
1
=
x
0
10
= 0.098, (2.8)
and
x
1
=
y
1
5
= 0.0196. (2.9)
(c) The O(
2
) terms give
x
1
+ 10y
2
= 0,
5x
2
+y
2
= 0.
(2.10)
giving
(x
2
, y
2
) = (0.000392, 0.00196). (2.11)
Thus, to order
2
, we have
x = 0.98 + 0.0196 + 0.000392
2
+ ,
x = 2.1 0.098 0.00196
2
+ .
(2.12)
For = 0.01 the rst three terms give
(x, y) (0.9801960392, 2.099019804). (2.13)
8
The exact solution is
(x, y) =
_
49
50
,
105 7
50
_
, (2.14)
which, for = 0.01 gives
(x, y) =
_
49
49.99
,
104.93
49.99
_
= (0.9801960392 . . . , 2.09901980396 . . . ). (2.15)
The rst three terms in the perturbation expansion gives the solution to the accuracy of my
calculator!
Some important points:
(i) We had to solve the O(1) problem (i.e., the reduced problem) rst. All the subsequent
problems depended on it. One always needs to have a reduced problem that can be solved.
Trivial in this case, but not always.
(ii) The solution of the reduced problem (x
0
, y
0
) = (0.98, 2.1) is very close to the exact solution.
This is indication that the terms neglected to obtain the reduced problem were indeed small.
For the exact solution x = 0.0098 10y = 20.99 . . . , so approximating the rst equation
by dropping x was OK.
The next example shows one way things can go wrong. It is a simple example which allows us
to understand why perturbation theory fails in this case.
Example 2.1.2 (From Lin & Segel): Find an approximate solution of the system
x +y = 0.1,
x + 101y = 11,
(2.16)
for = 0.01.
The reduced problem is
y
0
= 0.1,
x
x
+ 101y
0
= 11,
(2.17)
which has the solution (x
0
, y
0
) = (0.9, 0.1). Solving the system exactly, we have
(1 101)x = 11 10.1 = 0.9,
(101 1)y = 0.11 0.1 = 0.01
(2.18)
so
(x, y) = (90, 1). (2.19)
The solution of the reduced problem is way o. What went wrong?
For the exact solution x = 0.9 is comparable to the other two terms in the rst equation. In
obtaining the reduced problem by dropping the x term we assumed that it was small compared with
the other terms. In this example this assumption is incorrect and it leads to a poor reduced problem.
In real problems we wont know the exact solution (otherwise we wouldnt be using perturbation
methods!), so how can we realize our perturbation solution is wrong? In this example, assuming we
havent noticed the problem we proceed to nd perturbative corrections. This leads to
x = 0.9 + 90.9 + 9180.9
2
+ ,
y = 0.1 0.9 90.9
2
+ ,
(2.20)
9
which, for = 0.01, gives
x = 0.9 +.909 + 0.91809 + ,
y = 0.1 0.009 0.00909 + .
(2.21)
It looks like the series will not converge (of course we cant really tell with only three terms). In
general the O() correction should be small compared with the leading-order (O(1) terms, and the
O(
2
) terms should be small compared to the O() terms. This is clearly not the case here.
The exact solution for x is
x =
0.9
1 101
. (2.22)
Thus, x() has a singularity at epsilon = 1/101 = 0.009901 . . . and the Taylor Series expansion for
x() cannot converge for = 0.01.
For = 0.002, say, the rst three terms of the expansion gives a very good approximation
(x 1.1185236 vs the exact solution x = 1.12782 . . . ).
Perturbative methods often work only if the small parameter(s), in this case, is small enough.
How small small enough is may be dicult to determine.
Dropping terms uncritically can be dangerous!
Learning how to simplify a problem consistently is dicult and a very important part of this
course.
In most problems you will have to introduce a small parameter, or perhaps several small pa-
rameters. Where does come from? Two possibilities:
Introduce articially.
Obtain from scaling and non-dimensionalization.
The latter is the most important when dealing with physical problems.
2.2 Roots of polynomials
References: Murdoch or Bender & Orzag.
Example 2.2.1 (From Bender & Orzag): Articial introduction of .
Find approximate solutions of
x
3
4.001x + 0.002 = 0. (2.23)
Tricky, but
x
3
4x = x(x 2)(x + 2) = 0, (2.24)
is easy. Consider (2.23) a perturbation of (2.24). There are many ways to do this, one is to consider
the problem
x
3
(4 +)x + 2 = 0. (2.25)
where we are interested in the solution when = 0.001. As above, assume the solutions x() have
a Taylor series expansion
x = x
0
+x
1
+
2
x
2
+
3
x
3
+ . (2.26)
10
Substituting into (2.25) and collecting like powers of gives
(x
3
0
4x
0
) + (3x
2
0
x
1
4x
1
x
0
+ 2) + (3x
2
0
x
2
+ 3x
2
1
x
0
4x
2
x
1
)
2
+O(
3
) = 0. (2.27)
The coecient of each power of must be zero, giving a sequence of problems to be solved.
1. O(1) problem:
x
3
0
4x
0
= 0 (2.28)
giving the three roots x
0
= 2, 0, 2. Note that we chose so that at = 0 our problem
reduced to this simple problem that we already noticed we could easily solve.
2. O() problem:
(3x
2
0
4)x
1
= x
0
2. (2.29)
This is easily solved:
x
1
=
x
0
1
3x
2
0
4
. (2.30)
Each value of x
0
gives a dierent value for x
1
. Note that the denominator 3x
2
0
4 is non-zero
for each of our values for x
0
.
3. O(
2
) problem:
(3x
2
0
4)x
2
= x
1
3x
2
1
x
0
, (2.31)
so
x
2
=
x
1
3x
2
1
x
0
3x
2
0
4
. (2.32)
Note that the denominator is the same as in the O() problem. This is no coincidence. More
on this later.
Taking x
0
= 2, one root is
x
(1)
= 2
1
2
+
1
8
2
+O(
3
), (2.33)
which gives x
(1)
2.000499875 for = 0.001.
Comment:
There may be many ways to introduce a small parameter. Some good, some bad.
The O(1) problem (the reduced problem) must be solvable. In the preceding example this
problem was a cubic polynomial that we could easily solve, as opposed to the original cubic
problem. The higher-order problems were all simple linear problems. Once the leading-order
problem was solved the higher-order corrections were simple. This is common to all problems
involving nding roots of polynomials, but it is not always the case for other types of problems.
Sometimes the higher-order problems get more dicult to solve.
11
2.2.1 Order of the error
If we truncate our solution at O(
n
) then how can we estimate the error? We know that the error
is due to the terms O(
n
) and higher but that does not mean the error is bounded by C
n
for some
constant C > 0. The coecients of the
m
terms for m > n may grow very rapidly. The series may
not converge and in fact many useful asymptotic series do not.
Denition 2.2.1 We will call O
F
(
n
) the formal order of truncation and by this mean that
terms of O(
n
) and higher are neglected. It says nothing about the error.
From now on we will use the notation O
F
(
n
) unless we know the error is bounded by C
n
in
which case the error is O(
n
). For our root problem we can say something more.
Let
f(x, ) = x
3
(4 +)x + 2. (2.34)
Then f(x, ) = 0 implicitely denes x() actually three dierent x(), one for each root. The
Implicit Function Theorem guarantees that a unique function is dened by
f(x(), ) = 0; x(0) = x
0
, (2.35)
where x
0
is one of the roots of f(x, 0) = 0, i.e., x
0
= 2, 0, or 2, for a non-zero interval containing
= 0.
Theorem 2.2.1 Implicit Function Theorem: Let f(x, ) be a function having continuous par-
tial derivatives (including mixed derivatives) up to order r. Let x
0
satisfy f(x
0
, 0) = 0 and
f
x
(x
0
, 0) = 0. Then there is an
0
> 0 and a unique C
r
function x = x() dened for all 0 ||
0
such that
f(x(), ) = 0 and x(0) = x
0
. (2.36)
You can read about the Implicit Function Theorem in, for example, Murdoch Perturbations:
Theory and Methods, Marsden & Homan Elementary Classical Analysis or Apostol Calculus:
Volume II.
The function f(x, ) need not be a polynomial. If it is then it is C
0
. Taylors Theorem (see below) can be used to show that, using a third-order approximation for
example,
x() (x
0
+x
1
+x
2
2
)
3
6
, (2.39)
where
M = max
_
3
x
3
()
: [0,
0
]
_
, (2.40)
which gives us some information about the error.
12
Theorem 2.2.2 Taylors Theorem: Let x() be a C
r
function on || <
0
. For k r 1 let
p
k
() be the Taylor polynomial
p
k
() =
k
0
x
(n)
n!
(2.41)
where x
(n)
denotes the n
th
derivative of x. Then if x() is approximated by p
k
() the error is
R
k
() = x() p
k
() =
_
0
x
(k+1)
()
( )
k
k!
d (2.42)
and for each
1
(0,
0
)
|R
k
()|
M
k
(
1
)
(k + 1)!
||
k+1
for ||
1
(2.43)
where
M
k
(
1
) = max
_
|f
(k+1)
()| for ||
1
_
(2.44)
For a proof, which is based on the fundamental theorem of calculus and integration by parts, see
any rst year Calculus book. It is also discussed in the text by Murdoch.
2.2.2 Sometimes you dont expand in powers of
The presence of a small parameter in your problem does not necessarily imply that the pertur-
bation series solution is in integer powers of . Consider the following.
Example 2.2.2 Find approximate roots x() of
f(x, ) = x
3
x
2
+ = 0. (2.45)
Solution: Proceeding as before substitute
x = x
0
+x
1
+x
2
2
+ (2.46)
into the equation giving
x
3
0
x
2
0
+
_
(3x
2
0
2x
0
)x
1
+ 1
_
+
_
(3x
2
0
2x
0
)x
2
+ 3x
0
x
2
1
x
2
1
_
2
+ = 0. (2.47)
This leads to the following sequence of problems:
1. O(1) problem:
x
3
0
x
2
0
= 0. (2.48)
which has two roots: x
0
= 1 and x
0
= 0. The latter is a double root.
2. O() problem:
(3x
2
0
2x
0
)x
1
+ 1 = 0, (2.49)
giving
x
1
=
1
3x
2
0
2x
0
. (2.50)
3. O(
2
) problem: The solution is
x
2
=
3x
0
x
2
1
x
2
1
3x
2
0
2x
0
. (2.51)
13
Figure 2.1: (a) Plots of the functions y = f(x,0) (solid curve) and y = f(x, 0.01) (dashed curve) where f(x, ) =
x
3
x
2
+ . (b) Neighbourhood of x = 1. (c) Neighbourhood of x = 0.
For the single root x
0
= 1 we nd x
1
= 1 and x
2
= 2, so an approximation to one root is
x
(1)
= 1 2
2
+O(
3
), (2.52)
(why can we use O instead of O
F
?). For the double root x
0
= 0 both x
1
and x
2
are undened since
the denominator 3x
2
0
2x
0
= 0!
What went wrong and how can we resolve the problem? Note that f
x
(x
0
, 0) = 3x
2
0
2x
0
is
equal to zero at x
0
= 0 so at the double root the conditions of the Implicit Function Theory are
not satised.
The curves f(x, ) for = 0 and 0.01 are illustrated in Figure 2.1. Consider the simple root
near x = 1. Let g(x) = f(x, 0). For = 0 the polynomial y = g(x) can be approximated by
the tangent line y = g
_
x
0
+
1/2
x
1
+x
2
+
3/2
x
3
+
_
2
+ = 0. (2.54)
Expanding and collecting like powers of leads to
x
3
0
x
2
0
+ (3x
2
0
2x
0
)x
1
1/2
+
_
(3x
2
0
2x
0
)x
2
+ 3x
0
x
2
1
2x
0
x
1
x
2
1
+ 1
_
+
_
(3x
2
0
2x
0
)x
3
+ 6x
0
x
1
x
2
+x
3
1
2x
1
x
2
_
3/2
+ = 0.
(2.55)
For the double roots x
0
= 0 this simplies to
_
x
2
1
+ 1
_
+
_
x
3
1
2x
1
x
2
_
3/2
+ = 0, (2.56)
hence the two roots near zero are
x
2,3
=
1/2
+
1
2
+O
F
(
3/2
). (2.57)
[Note: instead of subsituting (2.53) it is easier in this case to use x
0
= 0 and substitute x() =
1/2
x
1
+ . This simplies the algebra, particularly if you are nding the solution by hand].
2.2.3 Solving by rescaling: a singular perturbation problem
By an appropriate rescaling we can replace O
F
in the previous solution with O. Let =
1/2
and
x = y so that the two roots near x = 0, x
(2,3)
, become y 1. The polynomial become
y
3
y
2
+ 1 = 0. (2.58)
Expanding as
y = y
0
+y
1
+
2
y
2
+ , (2.59)
15
leads to
_
y
0
+y
1
+y
2
2
+y
3
3
+
_
3
_
y
0
+y
1
+y
2
2
+y
3
3
+
_
2
+ 1 = 0. (2.60)
Expanding and collecting like powers of leads to
y
2
0
+ 1 + (y
3
0
2y
0
y
1
) + (3y
2
0
y
1
2y
0
y
2
y
2
1
)
2
+ = 0. (2.61)
Solving this leads to
y = 1 +
1
2
5
8
2
+O(
3
), (2.62)
where we can say O(
3
) because the conditions of the implicit function theorem are satised. Using
=
1/2
and y = x/
1/2
recovers (2.57).
We now have a dierent problem. The cubic polynomial (2.58) has three roots. Our perturbation
solution has only found two of them! What happen to the other one?
We already know that the missing root is x
(1)
= 1 2
2
+ O(
3
). In terms of y and this
becomes
y
(1)
=
1
2
3
+O(
5
). (2.63)
This has a singularity at = 0. The rescaling x = y is only valid if = 0.
2.2.4 Finding the singular root: Introduction to the method of dominant bal-
ance
In the examples we have considered thus far we have always used the Basic Simplication Procedure
(set the small parameter to zero) to obtain the reduced problem. This is not always appropriate,
and indeed often is not in singular perturbation problems.
Consider again the problem
y
3
y
2
+ 1 = 0, (2.64)
where 1.
The equation has three terms in it. We wish to simplify the problem and that can only be done
by dropping one of the three terms. The idea here is that two of the three terms are much larger
than the third so to a rst approximation they are equal. This gives the reduced problem. There
are three possible cases:
Case 1: y
3
is much smaller than y
2
and 1. This leads to the reduced problem y
2
0
= 1 from which
we have already seen two roots are obtained. For two of the three roots y
3
is indeed small
compared with y
2
and 1.
Case 2: y
2
is much smaller than y
3
and 1. If this is true then y
3
1 which means y 1
1/3
/
1/3
.
Note there are three roots corresponding to each of the cubic roots of 1: 1, e
i2/3
and e
i4/3
.
Since 1, y is very large. But that means y
2
1 contradicting our assumption that
y
2
1. Thus this case is not consistent and must be discarded.
Case 3: 1 is much smaller than y
3
and y
2
. Solving y
3
0
= y
2
0
gives y
0
= 0, which violates our
assumption that y
2
1, or y
0
= 1/. If y 1/ then y
3
y
2
1/
2
1 so this solution
is consistent with our assumption that 1 is small compared with the other terms. The full
solution is now obtained by expanding y() as
y =
1
+y
0
+y
1
+y
2
2
+ . (2.65)
Proceeding we would obtain (2.63).
16
2.3 Problems
1. Find approximate solutions of the following problems by nding the rst three terms in a
perturbation series solution (in an appropriate power of ) using perturbation methods. For
problem (a) explain whether the missing terms are O
F
(
?
) or O(
?
). You should nd all of
the roots, including complex roots.
(a) x
2
+ (5 +)x 6 + 3 = 0.
(b) x
2
+ (4 +)x + 4 = 0.
(c) (x 1)
2
(x + 2) + = 0.
(d) x
3
+ + 1 = 0.
(e) x
3
+x
2
+ 2x + 1 = 0.
(f) x
5
+ (x 2)
2
(x + 1) = 0.
(g) x
4
+x
3
+x
2
3x + 2 = 0.
17
18
Chapter 3
Nondimensionalization and scaling
The chapter is based on material from Lin and Segel (1974). It is strongly recommended that you
read the relevent sections of this book.
3.1 Nondimensionalizing to get
Example 3.1.1 (The Projectile Problem) Consider a vertically launched projectile of mass m
leaving the surface of the Earth with speed v. Find the height of the projectile as a function of time.
Ignore:
the Earths rotation;
the presence of air (i.e., friction);
relativistic eects;
the fact that the Earth is not a perfect sphere;
etc., etc., etc.
Assume:
Earth is a perfect sphere;
Newtonian mechanics apply.
Include:
Fact that the gravitational force varies with height.
Solution
Let the x-axis extend radially from the centre of the Earth through the projectile. Let x = 0
at the Earths surface. Let M
E
and R be the mass and radius of the Earth.
Let x(t) be the height of the projectile at time t. The initial conditions are
x(0) = 0 and x(0) = v > 0, (3.1)
19
where the dot denotes dierentiation.
From Newtonian mechanics
x(t) =
GM
E
(x +R)
2
=
gR
2
(x +R)
2
(3.2)
where g = GM
E
/R
2
9.8 m s
2
is the gravitational acceleration at x = 0.
Summary of the problem:
x =
gR
2
(x +R)
2
,
x(0) = 0,
x(0) = v.
(3.3)
We can separate the solution procedure into three steps: (1) dimensional analysis; (2) use the
ODE to deduce some useful facts; and (3) nondimensionalize (rescale) the problem to obtain a good
reduced problem and nd an approximate solution.
1. Dimensional analysis.
Physical Quantity Dimension
t, time T
x, height L
R, radius of Earth L
V , initial speed LT
1
g, acceleration at x = 0 LT
2
There are two dimensions involved: time and length. We need to scale both by introducing
nondimensional time and space variables via,
t = T
c
t and x = L
c
x. (3.4)
where T
c
and L
c
are characteristic time and length scales. They hold the dimensions while
t and x
are dimensionless. There are many choices for T
c
and L
c
.
Typical values of v, R and g are
v 100 m s
1
,
R 6.4 10
6
m
g 10 m s
2
.
While the values of R and g are xed the value of v is a choice. This choice is such that the
projectile rises high enough for the height variation of the gravitational force has an eect (it will
be small).
2. Use the ODE to say something useful about the solution.
20
1. Existence - Uniqueness Theorems for 2
nd
order ODEs ensures that there is a unique solution
up to some time t
0
> 0.
2. Multiplying the ODE by x and integrating from 0 to t
max
, where t
max
is the time the projectile
reaches its maximum height x
max
gives
x
max
=
v
2
R
2gR v
2
=
v
2
2g
_
1
1
v
2
2gR
_
(3.5)
Note that
1. x
max
as v
2gR 10
4
m s
1
.
2. For v 100 m s
1
, g 10 m s
2
, R 6.4 10
6
m,
v
2
2gR
10
4
2 10 6 10
6
10
4
(3.6)
x
max
v
2
2g
. (3.7)
3. Nondimensionalization
We now consider three possible choices for the time and length scales T
c
and L
c
. The rst two
will turn out to be bad choices but they serve to illustrate some of the things that can go wrong
and also illustrate the point that you need to put some thought into your choice of scales.
Procedure A:
Take L
c
= R and T
c
= R/v, which is the time needed to travel a distance R at speed v. Then
dx
dt
=
d
t
dt
d
d
t
(L
c
x) =
L
c
T
c
d x
d
t
= v
d x
d
t
(3.8)
which makes sense as L
c
/T
c
= v is the velocity scale. Next
d
2
x
dt
2
=
L
c
T
2
c
d
2
x
d
t
2
=
v
2
R
d
2
x
d
t
2
(3.9)
Therefore the ODE becomes:
v
2
R
d
2
x
d
t
2
=
gR
2
(R x +R)
2
=
g
( x + 1)
2
,
(3.10)
or
v
2
gR
d
2
x
d
t
2
=
1
(1 + x)
2
. (3.11)
Recall that v
2
/2gR 10
4
which is very small. Hence
=
v
2
gR
(3.12)
is a small dimensionless parameter.
21
Scaling the initial conditions we have
x(0) = 0 x(0) = 0 (3.13)
x(0) = v v
d x
d
t
(0) = v
d x
d
t
(0) = 1, (3.14)
hence the nal scaled, nondimensional problem is
d
2
x
d
t
2
=
1
(1 + x)
2
,
x(0) = 0,
d x
d
t
(0) = 1.
(3.15)
Because we have only scaled the variables and have not dropped any terms we have not intro-
duced any errors. No approximation has been made yet and the solution of this scaled problem is
the correct solution. The diculty lies with the reduced problem. The reduced problem, obtained
by setting = 0, is
0 =
1
(1 + x
0
)
2
,
x
0
(0) = 0,
d x
0
d
t
(0) = 1,
(3.16)
which has no solution! This is a bad reduced problem. The small parameter multiplying the
second derivative of x incorrectly suggests that this term is small. In fact, at t = 0 the r.h.s. is
exactly equal to -1. Thus, if = 10
4
, at t = 0 d
2
x/d
t
2
must be equal to 10
4
, which is very large
compared with 1. We need to scale the dimensional variables so the presence of the small parameter
correctly identies negligible terms. This is very important.
Procedure B:
The quantity
_
R
g
has units of time, so lets try T
c
=
_
R
g
and take L
c
= R as before. This
gives
d
2
x
d
t
2
=
1
(1 + x)
2
, (3.17)
x(0) = 0, (3.18)
d x
d
t
(0) =
v
2
Rg
=
, (3.19)
where, as before, = v
2
/gR 10
4
1.
As in the previous case, no approximations have been made yet so the solution of this problem
is the correct solution. There are, however, two problems with this scale.
1. The ODE has not been simplied!
2. The solution of the reduced problem has x becoming negative for
t > 0 (since the initial
velocity is zero and the initial acceleration is negative). Hence, the solution of the reduced
problem has the projectile going the wrong way!
22
These are both indications of a bad reduced problem!
Procedure C:
To get a good reduced problem we must properly scale the variables. You must think about how
you nondimensionalize the problem!
In procedure A we obtained
d
2
x
d
t
2
=
1
(1 + x)
2
(3.20)
As already pointed out, the problem here is that
d
2
x
d
t
2
must be very large so that
d
2
x
d
t
2
balances the
r.h.s. since both sides are equal to negative one at t = 0. The nondimensionalization should be
done so that the coecients reect the size of the whole term.
Well now do the scaling properly. We have already shown that the maximum height reached
by the projectile is
x
max
=
v
2
2g
_
1
1
v
2
2gR
_
v
2
2g
, (3.21)
since v
2
/(2gR) 10
4
. Thus
x
max
R
v
2
2gR
10
4
x
max
R, (3.22)
showing that R is not a good choice for the length scale:
If we set x = R x then
0 x
V
2
2g
,
0 x
V
2
2gR
10
4
.
(3.23)
This scaling is not a good choice because x is very tiny, i.e., much smaller than one.
If we set x =
V
2
g
x then
0 x
1
2
, (3.24)
i.e. x is an O(1) number. Thus L
c
= v
2
/g is a much better choice for the length scale. It is
in fact the only choice because this scaling reects the maximum value of x(t).
v is the obvious velocity scale since the velocity of the projectile must vary between v and v
as the projectile rises and returns to the Earths surface. If v = L
c
/T
c
then T
c
= L
c
/v = v/g,
is the only logical time scale, since it ensures
t is O(1).
Suppose the time scale is not obvious. Then leave it undetermined for a while. Have:
L
c
T
2
c
d
2
x
d
t
2
=
gR
2
(R +L
c
x)
2
=
g
(1 +
Lc
R
x)
2
v
2
/g
T
2
c
d
2
x
d
t
2
=
g
(1 +
v
2
gR
x)
2
_
v/g
T
c
_
2
d
2
x
d
t
2
=
1
(1 + x)
2
(3.25)
23
where = v
2
/(gR) 1 as before. Since the r.h.s. 1 , the l.h.s. 1. To have
d
2
x
d
t
2
close
to one (in magnitude) means
v/g
Tc
should be close to 1. Therefor one should choose T
c
= v/g.
The problem is now
d
2
x
d
t
2
=
1
(1 + x)
2
,
x(0) = 0,
d x
d
t
(0) = 1.
(3.26)
Setting = 0 gives the reduced problem
d
2
x
0
d
t
2
= 1,
x
0
(0) = 0,
d x
0
d
t
= 1,
(3.27)
which has the solution
x
0
(t) =
t
t
2
2
. (3.28)
Note that max{ x
0
} is 1/2 as expected. Note also that x
o
(
du
dx
: x [a, b]
_
L =
max |u|
max
du
dx
(3.36)
Note: If u is known this is easy. If u is unknown this can be dicult.
Example 3.2.1 Consider the function
u = a sin(x), a > 0 on [0, 2]. (3.37)
Solution: Obviously U = a and
L =
max |u|
max
du
dx
=
a
a
=
1
, (3.38)
giving
u = sin x (3.39)
25
In general, a model will be of the type
f(u, u
, u
, . . . , u
(n)
) = 0 (3.40)
One could take L so that
U
L
= max |u
| or
U
L
2
= max |u
| or . . . or
U
L
n
= max |u
(n)
|. (3.41)
You should choose L so that the largest of the non-dimensional derivatives has order of magnitude
1 L is smallest of above choices. Thus, take
L = min
_
max |u|
max |u
|
,
_
max |u|
max |u
|
_
1/2
, ,
_
max |u|
max |u
(n)
|
_
1/n
_
. (3.42)
Example 3.2.2 Consider the function
u = a sinx. (3.43)
Solution: Have
_
max |u|
max |u
(n)
|
_
1/n
=
_
a
a
n
_
1/n
=
1
(3.44)
so L = 1/.
Example 3.2.3 Consider the function
u = a sin x + 0.0001a sin 10x. (3.45)
Solution: Have max |u| a so take U = a. Next,
max |u
(n)
| = max
a
n
_
_
cos(x)
or
sin(x)
_
_
+ 10
n3
a
n
_
_
cos(10x)
or
sin(10x)
_
_
= a
n
max
_
_
cos(x)
or
sin(x)
_
_
+ 10
n3
_
_
cos(10x)
or
sin(10x)
_
_
_
a
n
for n 3
a
n
10
n3
for n 1
(3.46)
Thus, for n 3 one should take L = 1/ while for n 3 one should take L = 1/(10
13/n
)
which is approximately 1/10. Figure 3.2 shows plots of u and some of its derivatives,
clearly illustrating that for large derivatives the fast oscillations dominate and determine the
appropriate length scale.
26
0 1 2 3 4 5
x
0.3
0.0
0.3
y
(a)
0 1 2 3 4 5
x
1
0
1
y
(b)
0 1 2 3 4 5
x
1.5
0.0
1.5
y
(c)
0 1 2 3 4 5
x
150
0
150
y
(d)
Figure 3.2: Plots of u(x) and some of its derivatives where u(x) = a sin(x) + 0.001a sin(10x) with a = 0.1 and
= 3. (a) u(x). (b) u
(x). (c) u
(x). (d) u
(4)
(x).
0 1 2 3 4 5
x
0.0
0.5
1.0
1.5
y
(a)
T
1
T
2
0 1 2 3 4 5
x
0.0
0.5
1.0
1.5
y
(b)
T
1
T
2
Figure 3.3: (a) Orthodoxy satised on [0, 5]. (b) Orthodoxy not satised on [0, 5].
3.3 Orthodoxy
Suppose we are comparing two terms in a model, T
1
(x) and T
2
(x), for x [a, b], which have been
appropriately scaled . We now wish to compare the sizes of each and neglect one if it is small
compared to the other.
Problem: The scaling may show that max |T
2
| max |T
1
|, but this does not mean that |T
2
| |T
1
|
on all of [a, b].
Denition 3.3.1 Orthodoxy is said to be satised if one term is much smaller than the other on
the whole interval.
If orthodoxy is not satised then the intervals on which orthodoxy is not satised may be so
small that the eects are negligible, e.g., T
1
(x) = sinx and T
2
= 0.01 cos x, or multiple scales are
needed.
27
0.0 0.2 0.4 0.6 0.8 1.0
x
0.0
0.2
0.4
0.6
0.8
1.0
y
Figure 3.4: Solid: y = a(x exp(x/) for a = 0.8 and = 0.04. Dashed: y = ax. Vertical dotted lines are x =
and x = 4.
Example 3.3.1 Consider the function u(x) = a(x + e
x/
) for x [0, 1], a > 0 and 0 < 1
(see Figure 3.4). What scales for x should be used?
The derivative of u(x) is
u
(x) = a
_
1
1
e
x/
_
=
_
_
a
_
1
1
_
a/ at x = 0;
a
_
1
1
_
a at x = 1;
(3.47)
for 0 < 1. Taking L =
max |u|
max |u
|
=
a
a/
gives L = when 1. This is a good length scale near
the origin (see gure) but not in the region far away from the origin. Away from the origin, say on
[4, 1]
max |u
| = u
(1) a. (3.48)
Using U = a and L = gives u = x + exp( x) and u
( x)
is very tiny. For most of the domain of interest the correct length scale is L = 1
Functions such as this one need to be treated dierently in dierent parts of the domain. There
is an inner region, near the origin, in which u(x) varies rapidly, and an outer region, away from the
origin, where u varies much more slowly.
Inner Region: Within a few multiples of of x = 0
max |u| a
max |u
|
a
U = a, L =
Therefor we should set u(x) = a u
i
and x = x
i
where subscript i denotes inner region. With this
scaling
u(x) = a(x +e
x/
) u
i
( x
i
) = x
i
+e
x
i
(3.49)
28
The leading order behaviour of u in the inner region is e
x
i
. We say u
i
( x
i
) e
x
i
as 0 with
x
i
xed, where denotes is asymptotic to. More on this shortly.
Outer Region: Many multiples of away from the origin.
In the outer region
u
= a
_
1
1
e
x/
_
a. (3.50)
Both max |u| and max |u
+g
+ y
= 0,
y(0) = a,
y(1) = b.
(3.55)
This is a rst-order ODE with two boundary conditions! We can only use one of them. Which
one? The solution above shows that we must pick y(1) = b which yields the outer solution. For the
second problem, y
= 0, the reduced problem is identical but we must now use the boundary
condition y(0) = a. How can we determine which boundary condition to use without knowing the
solution? What happens if is negative? We will return to questions of this type later when we
study boundary layers and matched asymptotics.
Example 3.3.4 Consider the IVP
x(t) +
2
x(t) = sin(t) +, t R
x(0) = 1
x
(0) = 0.
(3.56)
1. Find the exact solution.
2. Find x(t, 0) and x(t, ) and make a sketch. Is orthodoxy satised?
3. Is lack of orthodoxy important?
Solution:
1. The general solution of the DE is
x(t) = Asin t +Bcos t +
1
2
1
sint +
2
. (3.57)
30
Applying the boundary conditions gives
x(t) =
1
(
2
1)
sin t +
1
2
1
sin t +
2
_
1 cos t
_
. (3.58)
2. Near the zeros of x, x and sint the term in the ODE will not be much smaller than these
terms so orthodoxy is not satised
3. It does not matter that orthodoxy is not satised in this case.
|x(t, 0) x(t, )| =
2
|1 cos t|
2
2
1, (3.59)
where 1 gives the order of magnitude of the solution (and hence is the appropriate quantity
to compare to).
3.4 Example: Inviscid, compressible irrotational ow past a cylin-
der
Background: (not examinable)
Inviscid ow means neglect viscosity and heat conduction, (i.e. adiabatic ow).
This type of ow is a good approximation for cases where a fast moving object (i.e. a plane)
moves through the air on a time scale much smaller than that required for signicant diusion. It
is valid only outside the boundary layer.
Thermodynamics tells us that for isentropic ow the pressure p and density are related by an
equation of state p = p() or = (p). Two important cases are
For a perfect gas at constant temperature
p
= C; (3.60)
For a Perfect Gas at constant entropy
p
= C, (3.61)
where C is a constant and =
C
P
C
V
1.4. We will assume isentropic ow (constant entropy).
Let v(x, y, z, t) be the uid velocity. The motion of the uid is governed by the following
conservation laws:
1. Conservation of mass:
t
+
(v) = 0 (3.62)
2. Conservation of linear momentum:
_
v
t
+
_
v
_
v
_
=
p (3.63)
31
Denition 3.4.1 Irrotationality: If uid particles have no angular momentum then
v = 0.
Denition 3.4.2 The sound speed is dened by
a =
dp
d
=
_
. (3.64)
Theorem 3.4.1 (Kelvin, 1868) For inviscid ow with p = p(), if the uid is initially irrota-
tional and the speed U of the ow is less that speed of sound then the ow remains irrotational for
all time.
In this theorem U is the maximum deviation from the ow speed at innity, or far from the
cylinder. That is, U should be found in a reference frame xed with the uid at innity.
If
v = 0 at t = 0 then, assuming the conditions of Kelvins Theorem are satised,
v = 0
for all time v =
for some velocity potential . The introduction of a velocity potential greatly
simplies things because the three components of the velocity vector are replaced by a single scalar
eld.
Using
1
p =
p
p
1/
C
1/
=
1
(p
11/
)C
1/
, (3.65)
the momentum equation can be written as
t
+
1
2
|
|
2
+
1
p
11/
C
1/
_
= 0. (3.66)
Thus,
t
+
1
2
|
|
2
+
a
2
1
= g(t), (3.67)
where g(t) is an undetermined function of time. Assuming a steady uniform far-eld ow v =
(U
gives
t
+
1
2
|
|
2
+
a
2
1
=
1
2
U
2
+
a
2
1
. (3.68)
The continuity equation can be written as
_
t
+
_
a
2
= ( 1)a
2
2
, (3.69)
Applying the operator (/t +
2
2
t
2
=
t
|
|
2
+
[(
]. (3.70)
We now simplify to 2 dimensions and use
Theorem 3.4.2 (Conformal Mapping Theorem:) Any simply connected region A C can be
transformed (bijectively and analytically) to a disk.
32
Using this theorem, for the 2-D case we can assume the object is a disk of radius R. Assuming
steady state the model equations give
_
1
u
2
a
2
_
xx
2uv
a
2
xy
+
_
1
v
2
a
2
_
yy
= 0, (3.71)
where v = (u, v) =
.
The discriminant of the PDE is
=
_
uv
a
2
_
2
_
1
u
2
a
2
__
1
v
2
a
2
_
= M
2
1 (3.72)
where M =
|v|
a
is the Mach number:
M < 1 subsonic ow equation (3.71) is elliptic static situations
M = 1 sonic ow equation (3.71) is parabolic diusive situations
M > 1 supersonic ow equation (3.71) is hyperbolic wave situations
Next we nondimensionalize. Let
(x, y) = R( x, y),
(u, v) = U
( u, v).
(3.73)
Recall that R is the radius of the cylinder and U
( u, v) =
1
R
(3.74)
So we should set = RU
. Putting the terms linear in on the left and the terms cubic in on
the right gives
U
R
_
x x
+
y y
_
=
U
2
a
2
_
u
2
U
x x
+ 2 u v
U
x y
+ v
2
U
y y
_
, (3.75)
where a is a function of x and y. We need to express it in terms of a
2
= M
2
2
x
xx
+ 2
x
xy
+
2
y
yy
1
2
2
(1
2
x
2
y
)
_
, (3.76)
where M
=
U
a
is the free stream Mach number.
For air at 20
100 km hr
1
, M
2
0.1, so M
2
is
a small parameter.
The boundary conditions: No ow through solid boundary and uid velocity goes to far-eld
velocity (1, 0) at innity:
n = 0 on x
2
+y
2
= 1,
(
x
,
y
) (1, 0) as |x| .
(3.77)
The solution will depend on the circulation around the disk. We will assume zero circulation which
implies that the ow is symmetric above and below the disk.
33
Regular Perturbation Theory Solution:
Assume M
2
1
(x, y) +M
4
2
(x, y) + (3.78)
O(1) problem: At leading order we have
0
= 0,
0
(U
, 0) as |x| ,
0
n = 0 on x
2
+y
2
= 1.
(3.79)
In addition
0
is symmetric about y = 0. This Neumann problem for
0
has the solution
0
(r, ) =
_
r +
1
r
_
cos . (3.80)
Without symmetry condition we get an additional term A for arbitrary A.
O(M
2
2
r
2
1
+
1
r
1
+
1
r
2
1
= ( 1)
__
1
r
7
1
r
5
_
cos +
1
r
3
cos 3
_
,
1
0 as r ,
1
r
= 0 on x
2
+y
2
= 1,
1
(r, ) =
1
(r, ) (symmetry)
(3.81)
which can be solved to yield the total solution
=
_
r +
1
r
_
cos +
1
2
M
2
_
_
13
12r
1
2r
3
+
1
12r
5
_
cos
+
_
1
12r
3
1
4r
_
cos 3
_
+O
F
(M
4
).
(3.82)
Remarks:
1. Real life problems can be dicult.
2. Getting the rst two terms in a Perturbation Theory expansion can be a lot of work.
3. Problem: What is the error? It is believed that the series is uniformly valid (denition below)
but this has not been proven (as of mid-90s. I may be out of date). Hence, this is an example
of RPT.
Denition 3.4.3 A series expansion
2
(, ) is said to be uniformly valid if it converges uni-
formly over all parts of the domain as 0. The series is said to be uniformly ordered if all
n
are bounded, in which case the series may not converge.
More on this later.
34
Chapter 4
Resonant Forcing and Method of
Strained Coordinates: Another
example from Singular Perturbation
Theory
4.1 The simple pendulum
Consider a mass m suspended from a xed frictionless pivot via an inextensible, massless string.
Let be the angle of the string from the vertical. The only force acting on the mass is gravity and
the tension in the string (i.e., ignore presence of air). The governing equations for a mass initially
at rest at an angle a are
d
2
dt
2
+
g
sin = 0,
(0) = a,
d
dt
(0) = 0.
(4.1)
The solution of the linear problem, obtained by assuming is small and approximating sin by
is
= a cos
_
_
g
t
_
. (4.2)
According to this solution the mass oscillates with frequency
_
g/ and period T
= 2
_
/g. The
full nonlinear problem can be solved exactly in terms of Jacobian elliptic functions. Since these can
only be expressed in terms of power series we might as well seek a Perturbation Theory solution
which will give a power series solution directly. As a rst step we need to scale the variables.
To begin with consider the energy of the system. The governing nonlinear ODE has the energy
conservation law
d
dt
_
1
2
_
d
dt
_
2
cos
_
= 0, (4.3)
which, after using the initial conditions, gives
1
2
_
d
dt
_
2
+
g
cos a =
g
cos . (4.4)
35
From this we can deduce that || a and that oscillates periodically between a. Therefore
scale by a:
= a
. (4.5)
For the time scale take the inverse of the linear frequency, thus set
t =
g
. (4.6)
The scaled problem is
d
2
d
2
+
sin a
a
= 0,
(0) = 1,
d
(0)
d
= 0.
(4.7)
We will assume that a is small. Note that for small a sin(a
/d
2
. This suggests we have appropriately scaled t.
The Taylor series expansion of sin a
d
2
+
a
2
3!
3
+
a
4
5!
5
+ = 0. (4.8)
The small parameter a appears only in even powers, hence we seek a Perturbation Theory solution
of the form
=
0
() +a
2
1
() +a
4
2
() + . (4.9)
O(1) problem: At leading order we have
d
2
0
d
2
+
0
= 0,
0
(0) = 1,
d
0
d
(0) = 0,
(4.10)
which has solution
0
= cos . (4.11)
O(a
2
) problem: At the next order we have
d
2
1
d
2
+
1
=
1
3!
cos
3
=
1
24
cos 3 +
1
8
cos ,
1
(0) =
d
1
d
(0) = 0.
(4.12)
The general solution of (4.12) is:
1
() =
1
192
cos 3 +
1
16
sin +Acos +Bsin . (4.13)
36
a = 45 degrees
0 2 4 6
t (linear periods)
50
0
50
a
n
g
l
e
(
d
e
g
r
e
e
s
)
Figure 4.1: Comparison of regular perturbation theory solution with linear and nonlinear solutions for initial angle
of 45
. Dotted curve: linear solution. Solid curves: nonlinear solution. Dashed curves: regular perturbation theory
solution.
Applying the boundary conditions gives
1
=
1
192
[cos cos 3] +
16
sin (4.14)
so that the total solution is
= cos +a
2
_
1
192
(cos cos 3) +
16
sin
_
+O
F
(a
4
). (4.15)
Problem: The amplitude of the (a
2
/16) sin term grows in time. It is as important as the leading
order term, cos , when a
2
/16 is order 1. Thus, the perturbation series breaks down by a time of
O(1/a
2
), at which point a
2
1
is no longer much smaller than
0
. The breakdown is illustrated in
Figure 4.1 for a = /4. Note that while the perturbation solution becomes very bad after three
or four periods it is better than the linear solution for times up to close to 2 linear periods. At
this time the linear solution has drifted away from the nonlinear solution whereas the phase of
perturbation solution is much better.
Physically the perturbation solution goes awry because the linear (i.e., the leading-order) and
nonlinear solutions drift apart in time. The O(a
2
) error made in linearizing the problem to get
the leading-order problem for
o
are cumulative and eventually destroy the approximation. The
regular perturbation solution tries to correct for this but does not do so correctly the phase is
improved at the cost of a growing amplitude.
The secular term (a
2
/16) sin appears in the O(a
2
) solution because of the appearance of the
resonant forcing term cos in the DE for
1
(resonant forcing because the forcing term has
the same frequency as the homogeneous solution, or more generally because the forcing term is a
solution of the homogeneous solution):
d
2
1
d
2
+
1
=
1
24
cos 3 +
1
8
cos .
. .
resonant
forcing
term
The appearance of a resonant forcing term means this is another example of a Singular Perturbation
Theory problem.
37
How can we x this problem? From energy considerations we know that the amplitude is given
by the initial condition. The nonlinearity does not change this. We also know that the solution
is periodic. Nonlinearity modies the shape and period of the oscillations. It increases the period
because the true restoring force, (g/l) sin() is less than the linearized restoring force (g/l). The
properties of the linear and nonlinear solutions are compared in table 4.1.
property linear solution nonlinear solution
amplitude a a
shape sinusoidal non-sinusoidal shape
period 2
_
l/g increases with amplitude
Table 4.1: Properties of linear and nonlinear solutions.
Because the periods of the linear and nonlinear solutions are dierent they slowly drift out of
phase. Eventually they will be completely out of phase.
The Fix: We must allow the period, or equivalently the frequency, to be a function of a.
Recall the original unscaled problem was
d
2
dt
2
+
g
sin = 0,
(0) = a,
d
dt
(0) = 0.
(4.16)
As before, set = a
g
,
i.e. we used a time scale T
c
=
_
/g, which was independent of a, and proportional to the period of
the linearized solution. We need a time scale which is relevant to the nonlinear solution, one which
depends on a. Since we do not know how the period depends on a we are forced to introduce an
unknown function (a) via
t =
(a)
. (4.17)
This is known as the method of strained coordinates (MSC) (we have strained time by an
unknown function (a)). We will return to this method later.
Since in the limit a 0 the period does go to
_
/g we can take (0) = 1. With this time
scaling the nondimensionalized problem is
2
(a)
d
2
d
2
+
sina
a
= 0,
(0) = 1,
d
d
(0) = 0.
(4.18)
38
a = 45 degrees
0 2 4 6
t (linear periods)
40
20
0
20
40
a
n
g
l
e
(
d
e
g
r
e
e
s
)
a = 90 degrees
0 2 4 6
t (linear periods)
100
50
0
50
100
a
n
g
l
e
(
d
e
g
r
e
e
s
)
a = 135 degrees
0 2 4 6
t (linear periods)
150
75
0
75
150
a
n
g
l
e
(
d
e
g
r
e
e
s
)
Figure 4.2: Comparison of singular perturbation theory solution with linear and nonlinear solutions for dierent
initial angles. Dotted curves: linear solution. Solid curves: nonlinear solution. Dashed curves: singular perturbation
theory solution. The dashed curves are almost identical to the nonlinear solution.
We now expand both
and in powers of a
2
, via
=
0
() +a
2
1
() +a
4
2
() + ,
(a) = 1 +a
2
1
+a
4
2
+ .
(4.19)
Substituting the series into the dierential equation gives
_
1 + 2
1
a
2
+ (2
2
+
2
1
)a
4
+
__
d
2
0
d
2
+a
2
d
2
1
d
2
+
_
+
_
0
+a
2
1
+a
4
2
+
_
a
2
6
_
0
+a
2
1
+a
4
2
_
3
+O(a
4
) = 0.
(4.20)
O(1) Problem: The leading-order problem is unchanged
d
2
0
d
2
+
0
= 0,
0
(0) = 1,
d
0
d
(0) = 0.
_
_
0
= cos
O(a
2
) Problem: At O(a
2
) we have
2
1
d
2
0
d
2
+
d
2
1
d
2
+
1
1
6
3
0
= 0,
1
(0) = 0,
d
1
d
(0) = 0.
(4.21)
39
d
2
1
d
2
+
1
=
1
24
cos 3 +
1
8
cos .
. .
We had this before
+2
1
cos
There is a new resonant forcing term, namely 2
1
cos . By choosing
1
= 1/16 the resonant
forcing terms are eliminated. There is in fact no choice about this. The only way to eliminate the
secular growth in the O() solution is be eliminating the resonant forcing term. This reduces the
problem to
d
2
1
d
2
+
1
=
1
24
cos 3, (4.22)
which, with the initial conditions, gives
1
=
1
192
(cos cos 3) . (4.23)
The total solution, so far, is
= cos +
a
2
192
(cos cos 3) +O
F
(a
4
),
(a) = 1
a
2
16
+O
F
(a
4
),
(4.24)
where
=
_
g
(a)t. (4.25)
The dimensional solution is
(t) = a
() = a
__
g
(a)t
_
, (4.26)
or
(t) = a cos
__
g
_
1
a
2
16
+
_
t
_
+
a
3
192
_
cos
__
g
_
1
a
2
16
+
_
t
_
cos
_
3
_
g
_
1
a
2
16
+
_
t
__
+O
F
(a
5
).
(4.27)
The nonlinear solution frequency is (a)
_
g/ =
_
1
a
2
16
+
_
_
g/ <
_
g/ which makes
sense because we know that the period of the nonlinear solution must be larger than the period
of the linear solution because the forcing in the nonlinear problem, (g/l) sin , is smaller that the
forcing in the linear problem, (g/l) (i.e., the acceleration of the nonlinear pendulum is smaller
than for the linear pendulum). The SPT solution (4.27) is shown in gure 4.2 showing excellent
agreement with the full nonlinear solution over six linear periods for very large initial angles.
40
Chapter 5
Asymptotic Series
5.1 Asymptotics: large and small terms
Notation
For order of magnitude of a number of function we will use the symbol O
M
:
90 = O
M
(100)
0.0072 sin x = O
M
(10
2
)
Denition 5.1.1 (The O big-oh Symbol) Let f and g be two functions dened on a region
D in R
n
or C
n
. Then
f(x) = O(g(x)) on D (5.1)
means that
|f(x)| k|g(x)| x D (5.2)
for some constant k.
We will usually be interested in the relative behaviour of two functions in the neighbourhood
of a point x
0
. In that case when we write
f(x) = O(g(x)) as x x
0
we mean there exists a constant k and a neighbourhood of x
0
, U, such that
|f(x)| k|g(x)| for x U
Remarks
1. If g(x) = 0 then f(x) = O(g(x)) in D or f(x) = O(g(x)) as x x
0
can be written as
f(x)
g(x)
<
in D, or
f(x)
g(x)
is bounded as x x
0
.
2. O(g(x)) on its own has no meaning. The equals sign in f(x) = O(g(x)) is an abuse of
notation.
f(x) = O(g(x)) 2f(x) = O(g(x)) (5.3)
but this does not mean that 2f(x) = f(x).
41
3. f(x) = O(g(x)) does not imply that g(x) = O(f(x)). For example, x
2
= O(x) as x 0 since
|x
2
| < 5|x| for |x| < 5, but x = O(x
2
) as x 0 because it is not true that |x| < k|x
2
| for
some constant k in a neighbourhood of 0.
4. An expression containing O is to be considered a class of functions. For example, O(1)+O(x
2
)
in 0 < x < denotes the class of all functions of the type f+g where f = O(1) and g = O(x
2
).
5. If f(x) = c is a constant, f = O(1) no matter what the value of c is.
10
9
= O(1),
1 = O(1),
10
9
= O(1).
Example 5.1.1
x
2
= O(x) on [-2,2] since x
2
< 5|x| on [2, 2].
x
2
= O(x) on [1, ] since
|x
2
|
|x|
= |x| is unbounded on [1, ].
sin(x) = O(1) on R.
x
2
= O(x) as x 0 since
x
2
x
= x is bounded as x 0.
e
x
1 = O(x) as x 0 since
|e
x
1|
|x|
is bounded as x 0.
Denition 5.1.2 (The o little-oh symbol) Let f and g be functions dened on a region D
and let x
0
be a limit point of D. Then
f(x) = o(g(x)) as x x
0
,
means that
f(x)
g(x)
0 as x x
0
.
Example 5.1.2
x
3
= o(x
2
) as x 0.
x
3
= o(x
4
) as x .
x
n
= o(e
x
) as x .
Note that f(x) g(x) as x x
0
is the same as f = o(g(x)) as x x
0
.
Denition 5.1.3 (Asymptotic Equivalence) Let f and g be dened in a region D with limit
point x
0
. We write
f g as x x
0
(5.4)
to mean that
f(x)
g(x)
1 as x x
0
(5.5)
42
Note:
1. x
0
could be .
2. f g as x x
0
implies that f = O(g(x)) and g = O(f(x)). The converse is not true. For
example, f(x) = x, g(x) = 5x.
Example 5.1.3
x +
1
x
1
x
as x 0,
since
x +
1
x
1
x
= x
2
+ 1 1 as x 0.
x +
1
x
x as x .
x
3
+ 9x
4
3
2
x
5
_
x
3
as x 0;
3
2
x
5
as x 0.
e
x9/x
_
e
9/x
as x 0;
e
x
as x .
Note: f g as x x
0
g f as x x
0
.
Note: f g means that f g g.
Example 5.1.4 The functions f = e
x
+x and g = e
x
are asymptotic to one another as x as
f g
g
=
x
e
x
0 as x
Note that the dierence f g does not go to 0! The dierence goes to innity as x . Saying
f g as x x
0
does not mean that f and g get close in an absolute sense, it only means that
they get close in a relative sense: f g can blow up but f g gets small relative to f or g (i.e.,
gets small in the sense that (f g)/g 0). Saying something is large or small can only be done in
comparison with something else. You shouldnt say 0.0000001 is small. It is small compared to 1
(which, if someone says 0.0000001 is small, is what they mean implicitely), but it is large compared
with 10
20
.
Denition 5.1.4 (Asymptotic Series) To say that
g(x) x
4
3x
2
2x + as x ,
means the following:
1. g x
4
, i.e.
g
x
4
1 as x ,
43
2. g x
4
3x
2
, i.e.
gx
4
3x
2
1 as x ,
3. g x
4
+ 3x
2
2x, i.e.
gx
4
+3x
2
2x
1 as x ,
etc. The series on the right hand side is an example of an asymptotic series. In the series the
fastest growing term comes rst. Each successive term must grow more slowly than the preceding
term.
Asymptotic series are very useful for nding approximate values of integrals and functions,
which we consider next.
5.2 Asymptotic Expansions
We begin by nding an asymptotic expression for an integral.
5.2.1 The Exponential Integral
The exponential integral function Ei(x) is dened by:
Ei(x) =
_
x
e
t
t
dt for x > 0. (5.6)
This is not very useful as it stands can we nd a useful approximation? Successively integrating
by parts gives
Ei(x) = e
x
_
1
x
1
x
2
+
2!
x
3
3!
x
4
+ +
(1)
n1
(n 1)!
x
n
_
. .
Sn(x)
+(1)
n
n!
_
x
e
t
x
n+1
dt
. .
Rn(x)
. (5.7)
As n , S
n
(x) gives a divergent series as is easily seen from the ratio test. The ratio of the
(m+1)
st
and m
th
terms is
(1)
m
m!
x
m+1
(1)
m1
(m1)!
x
m
=
m
x
as m (5.8)
for xed x. Suppose we change the question from What is the limit of S
n
(x) as n for xed
x?, to What is the limit as x for xed n.
Have
|Ei(x) S
n
(x)| = |R
n
(x)|
n!
_
x
e
t
t
n+1
dt
n!
x
n+1
_
x
e
t
dt
so
|Ei(x) S
n
(x)|
n!
x
n+1
e
x
0 as x . (5.9)
Hence for xed n, S
n
(x) gives a good approximation to Ei(x) if x is suciently large. An alternative
derivation of this result is the following. Because the error term R
n
alternates in sign S
2n1
<
44
Figure 5.1: Comparison of xe
x
Ei(x) and asymptotic approximations using two (dots), three (dashes) and four
(dash-dot) terms of the Asymptotic Expansion.
Ei(x) < S
2n
so the magnitude of the error is less than the magnitude of the rst omitted term,
namely e
x
n!/x
n+1
, as above.
Because of this result we can write
Ei(x) e
x
_
1
x
1
x
2
+
2!
x
3
3!
x
4
+
_
as x . (5.10)
This is an asymptotic expansion of Ei(x). Figure 5.1 compares xe
x
Ei(x) with xe
x
S
i
(x) for i = 1, 2, 3.
The rst two terms of the Asymptotic Expansion, 1 1/x, is within 1% of the exact value for x
larger than about 13.3. Using the rst four terms the error is less than 1% for x larger than about
6.3.
Now we can ask the question For a given value of x for what value of n, call it N(x), does
S
n
(x) give the best approximation to Ei(x)?. The answer to this question is dicult to determine
precisely as we only have an upper bound on the magnitude of the error which is easy to use. We
can approximate the answer by minimizing our bound on the error. This means choosing n so
the rst neglected term in the alternating series is minimized. As shown above the ratio of the
magnitudes of the (n+1)
st
and n
th
terms is
n
x
< 1 if n x. (5.11)
The terms decrease until n > x thus the minimum is at nN(x) = x, the greatest integer less than
x. This implies that as a function of n, |Ei(x) S
n
(x)| initially decreases monotonically until n
exceeds x after which it increases monotonically. This is illustrated in Figure 5.2 which compares
S
n
(x) with Ei(x) as a function of n for x = 5 and 10. Alternatively,
|R
n
(x)|
n!e
x
x
n+1
=
e
x
x
1
x
2
x
3
x
n
x
. (5.12)
The factors 1/x, 2/x, . . . are less than 1, hence decrease R
n
until n becomes larger than x.
In summary, for the exponential integral, for xed x our upper bound on the error is minimized
when n = x. Hence, S
x
(x) is an estimate Ei(x) with error R
x
(x) <
e
x
x!
x
x+1
.
45
Figure 5.2: Comparison of Ei(x) (dotted line) with values of partial sums Sn as a function of n. (a) x = 5.0. (b)
x = 10.0.
Rule of Thumb: For an alternating divergent series use S
N
(x) where the (N + 1)
st
term in the
asymptotic series is the smallest.
The rule of thumb is a rough guide. In practice we can often take n much less than n
opt
= x,
depending on the level of accuracy required. This is particularly true for large x as shown in Figure
5.2. Here it can be seen that the S
n
(x) are very close to Ei(x) over a much broader range of values
of n when x = 10 than when x = 5.
Example 5.2.1 For x = 10, R
4
(10)
4!e
10
10
5
1.1 10
8
. The error bound gives an approximate
error of
R
4
(10)
S
4
(10)
100% = 0.26%,
whereas using the optimal value of n the approximate error is
R
10
(10)
S
10
(10)
100% = 0.04%.
The actual error using S
4
is
Ei(10) S
4
(10)
Ei(10)
100% = 0.18%
and it is 0.0193% using S
10
and 0.0202% using S
11
.
Important point: For a given x there is a minimum error (which is less than the error bound,
in this case |R
x
| x!e
x
/x
x+1
) that can be made. In contrast, for a convergent power series
the error can be made arbitrarily small if we are prepared to sum enough terms. In this example
the minimum error decreases as x increases.
46
5.2.2 Asymptotic Sequences and Asymptotic Expansions (Poincare 1886)
Denition 5.2.1 A set of functions {
n
(x)}, n = 1, 2, 3, . . . for x D(= R, R
n
, C) is an asymp-
totic sequence (AS) as x x
0
if for each n,
n
(x) is dened on D and
n+1
(x) = o(
n
(x)) as
x x
0
.
Example 5.2.2
{(x x
0
)
n
} is an asymptotic sequence as x x
0
, but is not an asymptotic sequence as
x .
{e
x
x
an
} is an asymptotic sequence as x where a
n
R with a
n
+ 1 > a
n
.
{ln(x)
n
} is an asymptotic sequence as x .
Denition 5.2.2 Let x, x
0
and D be dened as above and let f(x) be a function on D. Let {
n
(x)}
be an asymptotic series as x x
0
. The formal series
f =
N
n=1
a
n
n
(x) (5.13)
is said to be an asymptotic expansion of f as x x
0
to N terms provided
f(x)
N
n=1
a
n
n
(x) =
_
_
_
o(
N
(x))
or
O(
N+1
(x))
_
_
_
as x x
0
. (5.14)
Note that (5.14) gives some information about the error, i.e.
error = f(x)
N
n=1
a
n
n
(x) 0
faster than
N
(x) 0 as x x
0
or it blows up more slowly. This means that the error is small
compared to
N
(x). Of course this may only be useful if
N
(x) 0 as x x
0
and only for x
suciently close to x
0
.
Important Point: The accuracy of an asymptotic approximation is limited. It has nothing to do
with ordinary convergence. In the case of a function f(x) expressed as a convergent power series
we can make the error arbitrarily small if we are prepared to sum enough terms. In an asymptotic
expansion the potential accuracy is limited.
Example 5.2.3
For Ei(x) the smallest we can guarantee the error to be less than
n!e
x
x
n+1
,
with n = x for any given x. This is an upper bound on the error, so the actual error might
be a lot smaller but without further analysis we cant say any more about the error. Thus, there
is nothing we can do to reduce the error using this asymptotic expansion (a function has many
asymptotic expansions, a dierent one may give a better error estimate).
47
Example 5.2.4 The Bessel function J
o
(x) has the power series expansion
J
0
(x) = 1
x
2
2
2
+
x
4
2
2
4
2
x
6
2
2
4
2
6
2
+ . (5.15)
which converges to J
0
(x) for all x. The power series is completely useless unless x is small. For
example,
J
0
(4) = 1 4 + 4
16
9
+ , (5.16)
and 8 terms are need to get three digits of accuracy. An asymptotic expansion for J
0
(x) is
J
0
(x)
_
2
x
__
1
9
128x
2
+
_
cos
_
x
4
_
+
_
1
8x
75
1024x
2
+
_
sin
_
x
4
_
_
as x .
(5.17)
This series is divergent for all x. This non-divergent asymptotic series is, however, extremely useful.
The leading order term
_
2
x
cos
_
x
4
_
(5.18)
gives J
0
(x) to three digit accuracy for all x 4! Example approximations are shown in Figure
5.3. There it can be seen that the leading-order asymptotic approximation is very good for x 1
whereas the 4, 10 and 20-term power series approximations are useful for x < 3, 7.5 and 15
respectively. Finding J
0
(99) using the power-series would clearly be dicult but easy using the
asymptotic approximation! We will discuss nding the asymptotic expansion for the Bessel function
in the next chapter.
Claim 5.2.1 If f(x) and {
n
(x)} are known where {
n
} is an asymptotic series, then the asymp-
totic expansion for f in terms of the
n
is unique.
Proof: Need to nd a
n
s such that
f a
1
1
+a
2
2
+ as x x
0
. (5.19)
This means that
f a
1
1
= o(
1
(x)) as x x
0
,
f a
1
1
=
f
1
a
1
0 as x x
0
.
Thus, take
a
1
= lim
xx
0
f
1
. (5.20)
Next
f a
1
1
a
2
2
= o(
2
),
f a
1
1
a
2
2
=
f a
1
2
a
2
0 as x x
0
.
48
Figure 5.3: Comparison of approximation of J0(x) with power series or asymptotic series. In both panels the
solid curve is J0(x) and the dotted curve is the leading-order term of the asymptotic expansion. (a) Dashed: 4
term power series approximation. Dash-dot: 10 term power series approximation. (b) Dashed: 10 term power series
approximation. Dash-dot: 20 term power series approximation.
Therefore take
a
2
= lim
xx
0
f a
1
2
. (5.21)
The pattern is clear.
Note:
1. This might give something useless, such as all a
n
s are zero, as would happen, for example, if
f = e
x
and
n
(x) =
1
x
n
as x .
2. If the asymptotic series is not known, there will be many possible asymptotic expansions for
f. For example,
sin 2 2
4
3
3
+
4
15
5
+ as 0,
sin 2 2 tan 2 tan
3
+ 2 tan
5
+ as 0,
sin 2 2
_
3
3 + 2
2
_
7
12
_
3
3 + 2
2
_
5
+ as 0.
49
5.2.3 The Incomplete Gamma Function
Example 5.2.5 The incomplete Gamma function is dened as
(a, x) =
_
x
0
e
t
t
a1
dt (5.22)
for a, x > 0.
1. Derive a power series expansions which converges for all x. Show it is useless for large x.
2. Find an asymptotic expansion for by writing (5.22) as
(a, x) =
_
0
e
t
t
a1
dt
_
x
e
t
t
a1
dt
= (a) Ei
a1
(x).
Solution:
1. Using the convergent power series expansion of e
t
write
e
t
t
a1
= t
a1
n=0
(1)
n
t
n
n!
=
n=0
(1)
n
t
n+a1
n!
.
(5.23)
The partial sums converge uniformly on any interval [0, x] so we can integrate term by
term to get
(a, x) =
n=0
(1)
n
n!
x
n+a
n +a
=
n=0
a
n
, (5.24)
where
a
n
=
(1)
n
n!
x
n+a
n +a
. (5.25)
Applying the ratio test,
a
n+1
a
n
=
x
(a +n + 1)(n + 1)
0 as n , (5.26)
showing that the series converges for all x. For any xed N the partial sum
S
N
(x) = x
a
N
n=0
(1)
n
x
n
(a +n)n!
as x . (5.27)
Thus, for large x a large number of terms from the power series are needed to obtain a
reasonably accurate approximation. This makes the power series useless for large x.
2. Proceeding as for Ei(x), several integration by parts yields
Ei
a1
(x) = x
a
e
x
_
1
x
+
(a 1)
x
2
+ +
(a 1)
[n1]
x
n
_
+ (a 1)
[n]
_
x
e
t
t
a(n+1)
dt,
(5.28)
50
where k
[n]
= k(k 1)(k 2) (k n + 1).
Set
S
n
(x, a) = x
a
e
x
_
1
x
+
(a 1)
x
2
+ +
(a 1)
[n1]
x
n
_
,
R
n
(x, a) = (a 1)
[n]
_
x
e
t
t
an+1
dt.
(5.29)
As before S
n
(x, a) is divergent as n . For xed x the integral in R
n
converges for
all a > 0 and lim
x
R
n
(x, a) = 0. Have
Ei
a1
(x) x
a
e
x
_
1
x
+
a 1
x
2
+ +
(a 1)
[n1]
x
n
+
_
, (5.30)
as x .
51
52
Chapter 6
Asymptotic Analysis for 2
nd
order
ODEs
6.1 Introduction
We now consider asymptotic methods for nding approximate solutions of ordinary dierential
equations of the form
y
(x) +p(x)y
+
4x
1 +x
2
y
2
1 +x
2
y = 0 (6.2)
has an ordinary point at x = 0.
One solution is
y =
x
1 +x
2
= x(1 x
2
+x
4
x
6
+ ) (6.3)
which has radius of convergence 1, which is equal to the distance from x = 0 to x = i, the
singularity of p and q.
53
Fact 6.1.2 A solution of (6.1) may be analytic at a regular singular point x
0
. If it is not, its
singularity must be either a pole or an algebraic or logarithmic branch point. At a regular singular
point
1. there is always a solution of the form y
1
= (x x
0
)
B(x), (6.4)
or
y
2
= (x x
0
)
A(x) ln(x x
0
) + (x x
0
)
C(x), (6.5)
where B and C are analytic.
Fact 6.1.3 At an irregular singular point at least one solution does not have the form of (6.4) or
(6.5).
Example 6.1.2 Consider
y
+
3
2x
y
+
1
4x
3
y = 0. (6.6)
1. x = 0 is an irregular singular point since q(x)x
2
=
1
4x
is not analytic.
2. Letting t =
1
x
and y(x) y(t) the ODE becomes
y
1
t
y
+
1
4t
y = 0. (6.7)
Neither p and q are analytic, so t = 0 (x = ) is not an ordinary point. Both t
2
q = t/4 and
tp = 1 are analytic so t = 0 (x = ) is a regular singular point.
Example 6.1.3 Consider the ODE
(x 1)(2x 1)y
+ 2xy
2y = 0. (6.8)
First, put it in standard form,
y
+
x
(x 1)(x
1
2
)
y
1
(x 1)(x
1
2
)
y = 0. (6.9)
From this we see that x = 1 and 1/2 are regular singular points. x = is also a regular singular
point. One solution of the ODE is y
1
= 1/(x 1) which has a Taylor Series expansion about x = 0
which converges for |x| < 1, where x = 0 is an ordinary point. Note that the radius of convergence
goes beyond the singularity at x =
1
2
. A second linearly independent solution is y
2
= x which has
a Taylor Series expansion about x = 0 that converges everywhere.
Example 6.1.4 Consider the ODE
y
1 +x
x
y
+
1
x
y = 0. (6.10)
x = 0 is a regular singular point.
x = is an irregular singular point.
Two linearly independent solutions are y
1
= e
x
and y
2
= 1 +x.
Both are analytic at x = 0
y
1
has an essential singularity at x = . y
2
has a pole at .
54
6.2 Finding the behaviour near an Irregular Singular Points: Method
of Carlini-Liouville-Green
The method of Frobenius can be used to nd the solution in a neighbourhood of ordinary or regular
singular points. More interesting for our purposes are irregular singular points for which asymptotic
methods yield useful solutions.
Example 6.2.1 Consider
x
3
y
= y y
1
x
3
y = 0. (6.11)
Find the behaviour as x 0
+
, an irregular singular point.
6.2.1 Finding the Leading Behaviour
Attempt 1: Method of Frobenius:
Try nd a solution in the form of a power series
y =
n=0
a
n
x
n+
, a
0
= 0. (6.12)
Have
y
n=0
(n +)(n + 1)a
n
x
n+2
, (6.13)
so
x
3
y
y =
n=0
(n +)(n + 1)a
n
x
n++1
a
n
x
n+
= a
0
x
+ (a
1
+( 1)a
0
)x
+1
+ (a
2
+ (1 +)a
1
)x
+2
+ = 0.
(6.14)
The coecient of each distinct power of x must be zero, hence all the a
n
s are zero! Thus there is
no solution in power series form.
Attempt 2: The behaviour of solutions of ODEs as x approaches an irregular singular point
usually involves some sort of exponential behaviour. This suggests looking for a solution of the
form y = e
S(x)
, (Carlini-Liouville-Green Method).
From
y
+p(x)y
+q(x)y = 0, (6.15)
we get
S
+ (S
)
2
+p(x)S
+q(x) = 0. (6.16)
We now proceed by assuming that S
(S
)
2
as x x
0
. If this is true we can write
(S
)
2
pS
q as x x
0
. (6.17)
We now replace the by = and solve for S
giving
S
p
_
p
2
4q
2
as x x
0
. (6.18)
55
Integrating this gives an approximation to S. We then need to verify that S
(S
)
2
as x x
0
.
For our problem
y
1
x
3
y = 0, (6.19)
giving
S
+ (S
)
2
1
x
3
= 0. (6.20)
Ignoring S
2
,
(S
)
2
1
x
3
or S
1
x
3/2
as x 0
+
. (6.21)
This means that
S
(x) = x
3/2
+C
(x),
where C
(x) = o(x
3/2
) as x 0
+
. Dierentiating gives
S
=
3
2
x
5/2
+C
(x). (6.22)
Assuming C
(x) = o(x
5/2
) as x 0
+
(no quarantee!), S
S
2
as we assumed. So far so good
everything is consistent .
Integrating (6.22) gives
S(x) 2x
1/2
as x 0
+
. (6.23)
Here we have assumed that if f g as x x
0
, then
_
f(x) dx
_
g(x) dx, (6.24)
i.e.,
_
f(x) dx =
_
g(x) dx +h(x), (6.25)
where h(x) = o(
_
g dx). This is not always the case, but usually OK. It is possible to have
f(x) = g(x) + h(x) where h = o(g) as x x
0
(hence f g) for which
_
hdx is not o(
_
g dx). A
trivial case is g = x and h = 0 where
_
h = c
_
g = x
2
/2 as x 0 where c is a constant. As
long as
_
f blows up we are usually OK (but not always!).
So far our approximate behaviour for y as x 0
+
is
y = e
S(x)
, (6.26)
where
S 2x
1/2
as x 0
+
, (6.27)
or
S = 2x
1/2
+C(x) (6.28)
where C(x) = o(x
1/2
) as x 0
+
. We now need to improve the solution. We now consider one
case, i.e., in particular, take
S = 2x
1/2
+C(x). (6.29)
56
Subsituting (6.29) into (6.20) gives
3
2
x
5/2
+C
+
1
x
3
2x
3/2
C
+ (C
)
2
1
x
3
= 0
3
2
x
5/2
+C
2x
3/2
C
+ (C
)
2
= 0.
(6.30)
Here we have made no approximations. Two independent solutions of this second order, nonlinear
ODE will provide two linearly independent solutions of the original ODE for y(x). To proceed we
assume that
C x
1/2
C
x
3/2
(C
)
2
x
3/2
C
. (6.31)
We also assume that C
x
5/2
. With these reasonable assumptions the nonlinear ODE for C
gives
2x
3/2
C
3
2
x
5/2
as x 0
+
,
C
3
4
x
1
as x 0
+
,
C
3
4
ln x as x 0
+
,
or
C(x) =
3
4
ln x +D(x)
where D(x) ln x as x 0
+
. This latter equation is exact. Substituting into the exact equation
for C gives
D
3
16x
2
+
3
2x
D
+ (D
)
2
2
x
3/2
D
= 0. (6.32)
Using D
1
x
, which is the correct behaviour for the solution we are chasing down, the dominant
balance gives
2x
3/2
D
3
16x
2
as x 0
+
. (6.33)
Integrating gives
D
3
16
x
1/2
+d, as x 0
+
. (6.34)
Note that here, because x
1/2
0 as x 0
+
weve have to include a constant of integration.
Because d x
1/2
as x 0
+
we can write D(x) as
D(x) = d +(x), (6.35)
where 0 as x 0
+
(and in fact (x) 3x
1/2
/16 as x 0). Thus
y = e
2x
1/2
+
3
4
lnx+d+(x)
,
= C
1
x
3/4
e
2x
1/2
e
(x)
,
C
1
x
3/4
e
2x
1/2
as x 0
+
,
(6.36)
since e
(x)
1 as x 0
+
.
57
Figure 6.1: Leading behaviours of the solutions of y
y/x
3
= 0 as x 0
+
. (a) x
3/4
e
2x
1
2
. (b) x
3/4
e
2x
1
2
.
Denition 6.2.1 The contributions to S(x) that do not vanish as x x
0
, some irregular singular
point, gives the leading behaviour of the solution of the dierential equation.
In this case C
1
x
3/4
e
2x
1
2
is the leading behaviour of one of the solutions of x
3
y
y = 0. Note
that because the dierential equation is linear and homogeneous, the constant C
1
is arbitrary. The
second solution, with S 2x
1
2
as x 0
+
, has the leading behaviour C
2
x
3/4
e
2x
1
2
. This
function goes to zero as x 0+ so rapidly than all its derivatives go to zero. The two leading
behaviours are illustrated in Figure 6.1.
6.2.2 Further Improvements: corrections to the leading behaviour.
Let
y = x
3/4
e
2x
1/2
[1 +g(x)] where g 0 as x 0
+
. (6.37)
Substituting into the ODE gives
g
+
_
3
2x
2
x
3/2
_
g
3
16x
2
3
16x
2
g = 0. (6.38)
This is a linear second order dierential equation. It has two linearly independent solutions which
implies two consistent dominant balances. We want the one satisfying g 0 as x 0
+
. Hence
3
16x
2
g
3
16x
2
as x 0
+
. (6.39)
In addition,
3
2x
2
x
3/2
as x 0
+
, (6.40)
so we must have
g
2
x
3/2
g
3
16x
2
as x 0
+
. (6.41)
58
To further simplify, consider four possibilities:
1. 3x
3/2
g
is negligible g
3
16
ln x, which is inconsistent with g 0 as x 0
+
.
2.
3
16
x
2
is negligible g
e
2x
3/2
which is inconsistent with g 0 as x 0
+
.
3. All three terms are needed. Hopefully this is not the case.
4. g
3
32
x
1/2
as x 0
+
, (6.42)
g
3
16
x
1/2
as x 0
+
. (6.43)
This is consistent since this gives g
3x
3/2
/64 3x
2
/16 2x
3/2
g
as x 0
+
.
Next let
g =
3
16
x
1/2
+g
1
(x), (6.44)
where
g
1
x
1/2
as x 0
+
. (6.45)
Proceeding as above we nd that
g
1
15
512
x as x 0
+
. (6.46)
We see that the rst two terms of the asymptotic expansion for g(x) are proportional to x
1/2
abd
x. Taking this as a pattern we guess that g(x) has the form
g(x)
n=1
a
n
x
n/2
(6.47)
where a
1
= 3/16 and a
2
= 15/512. Substituting into the ODE for g gives
a
n+1
=
(2n 1)(2n + 3)
16(n + 1)
a
n
a
n
=
(n
1
2
)(n +
3
2
)
4
n
n!
(6.48)
hence
y x
3/4
e
2x
1
2
. .
leading
behaviour
n=0
(n
1
2
)(n +
3
2
)x
n
2
4
n
n!
. .
divergent
as x 0
+
(6.49)
gives the complete asymptotic expansion of one solution of the dierential equation.
59
20 15 10 5 0
0.5
0
0.5
1
1.5
2
2.5
3
3.5
Airy
Bairy
Figure 6.2: Airy and Bairy functions: solutions of y
= xy.
6.3 The Airy Equation
Example 6.3.1 (Airy Equation)
y
= xy (6.50)
has an irregular singular point at x = .
The solution of the Airy equation is exponential for x > 0 and oscillatory for x < 0. Why? One
well known solution of the Airy equation, the Airy function Ai(x), decays rapidly for large positive
x. A second solution, the Bairy function Bi(x) grows rapidly. The Airy and Bairy functions are
shown in Figure 6.2. The Airy function is named after George Airy who used this function in his
study of optics (1838). It also arises in leading-order descriptions of dispersive wave fronts (e.g.,
surface gravity waves). We will encounter it below when we consider the turning point problem
which arises in many areas of physics.
Exercise 6.3.1 Show that the two possible leading asymptotic behaviours as x + are
y
1
C
1
x
1/4
e
2
3
x
3/2
as x +, (6.51)
y
2
C
2
x
1/4
e
2
3
x
3/2
as x +. (6.52)
The Airy function Ai(x) is the unique solution to (6.50) that satises (6.51) with C
1
=
1
2
1/2
. This
uniquely denes Ai(x). Why? The leading behaviour of Bi(x) is given by (6.52) with C
2
= 1/
.
This does not uniquely dene Bi(x). Why? In Figure 6.3 the leading behaviours for large positive
x are shown and compared with x
1
4
e
2x
1
.5/3
.
Full Asymptotic Expansion for Ai(x)
Let
Ai(x) =
1
2
x
1/4
e
2
3
x
3/2
w(x) as x +. (6.53)
60
Figure 6.3: Leading behaviours of the Airy and Bairy functions for large positive x. (a) Airy function leading
behaviour
1
2
1
4
e
2x
3/2
/3
(solid) compared with the more slowly decaying function
1
2
1
4
e
2x/3
(dashed). (a)
Leading behaviour of the Bairy function,
1
1
4
e
2x
3/2
/3
(solid), compared with the more slowly growing function
1
1
4
e
2x/3
(dashed).
Assume we can nd w(x)
n=0
a
n
x
n
with a
0
= 1 and < 0. Subsituting into (6.50), gives an
ODE for w,
x
2
w
_
2x
5/2
+
1
2
x
_
w
+
5
16
w = 0. (6.54)
Exercise 6.3.2 Show =
3
2
and
Ai(x)
1
2
1
4
e
2
3
x
3/2
n
x
3n
2
2
_
3
4
_
n
_
n +
5
6
_
_
n +
1
6
_
n!
, (6.55)
where
(n) = (n 1)! for n Z, (6.56)
(x + 1) = x(x), (6.57)
(x) =
_
0
e
t
t
x1
dt. (6.58)
6.4 Asymptotic Relations for Oscillatory Functions
Solutions of the Airy equation:
y
xy = 0, (6.59)
have the leading asymptotic behaviour
y Cx
1/4
e
2
3
x
3/2
as x +. (6.60)
For x , proceeding as before we would obtain
y C(x)
1/4
e
2
3
i(x)
3/2
as x . (6.61)
61
Now for the real valued solutions Ai(x) and Bi(x) we need a linear combination of the real and
imaginary parts, so it is tempting to write
y C
1
(x)
1
4
sin
_
2
3
(x)
3
2
_
+C
2
(x)
1
4
cos
_
2
3
(x)
3
2
_
, (6.62)
as x .
Fact 6.4.1 For large negative x
Ai
1
(x)
1
4
sin
_
2
3
(x)
3
2
+
4
_
, (6.63)
Bi
1
(x)
1
4
cos
_
2
3
(x)
3
2
+
4
_
. (6.64)
It is, however, incorrect to write
1
Ai
1
(x)
1
4
sin
_
2
3
(x)
3
2
+
4
_
, (6.65)
because both sides have zeros which, while very close together for large x, do not exactly coincide.
Hence
lim
x
Ai(x)
R.H.S.
, (6.66)
does not exist. To say the functions in the numerator and denominator are asymptotic to one
another requires that this limit exists and that it is equal to one.
How can we x this?
Example 6.4.1 Consider f(x) = sin x and g(x) = sin
_
x +
1
x
_
. For large x the graphs of f and g
are almost identical (see Figure 6.4), however, the zeros of f and g do not coincide so lim
x
(f/g)
is undened. Thus we cannot say that f g as x even though the dierence between the two
functions clearly goes to zero since 1/x, the phase shift of g relative to f goes to zero.
How can we x this?
Idea 1: f(x) = sin(r(x)) and g(x) = sin(q(x)) where r q as x .
Idea 2: We can write
sin
_
x +
1
x
_
= cos
1
x
sin x + sin
1
x
cos x,
= w
1
(x) sin x +w
2
(x) cos x,
(6.67)
where
w
1
(x) = cos
1
x
1 as x , (6.68)
and
w
2
(x) = sin
1
x
1
x
as x . (6.69)
Be careful not to say sin(1/x) 0 as x !
1
Although you will often see this written in a sloppy use of the notation .
62
Figure 6.4: Comparison of sin(x) (solid) and sin(x + 1/x) (dots).
Returning to the Airy Function, for x , write Ai(x) in the form
y = w
1
(x)(x)
1
4
sin +w
2
(x)(x)
1
4
cos , (6.70)
where
2
3
(x)
3/2
+
4
, and then seek the asymptotic behaviour of w
1
(x) and w
2
(x). The reason
for the introduction of the phase shift /4 will be pointed out later, although we will not show how
one can tell this is convenient.
Substitution into y
xy = 0 gives
_
w
1
+
1
2
x
1
w
1
+ 2(x)
1
2
w
2
+
5
16
x
2
w
1
_
sin
+
_
w
2
1
2
x
1
w
2
2(x)
1
2
w
1
+
5
16
x
2
w
2
_
cos = 0.
(6.71)
This gives one equation for two unknowns. We need two equations. There is a lot of freedom but
the simplest choice is to assume that the coecients of sin and cos are both zero. Next let
w
1
(x)
n=0
a
n
(x)
3
2
n
w
2
(x)
n=0
b
n
(x)
3
2
n
as x . (6.72)
Substituting these expansions into the two couple ODEs for w
1
and w
2
then gives
a
2n
= a
0
(1)
n
c
2n
,
a
2n+1
= b
0
(1)
n
c
2n+1
,
b
2n
= b
0
(1)
n
c
2n
,
b
2n+1
= a
0
(1)
n+1
c
2n+1
(6.73)
for n = 0, 1, . . . with c
0
= 1 and
c
n
=
(2n + 1)(2n + 3) (6n 1)
144
n
n!
=
1
2
_
3
4
_
n
_
n +
5
6
_
_
n +
1
6
_
n!
,
(6.74)
63
for n = 1, 2, 3, . . . . For the Airy and Bairy functions (because of the introduction of the phase shift
/4)
Ai(x) : a
0
=
1
, b
0
= 0,
Bi(x) : a
0
= 0, b
0
=
1
.
Note: w
1
(x) and w
2
(x) are not oscillating functions.
6.5 The Turning Point Problem
The turning point problem is a classical problem in mathematical physics. It arises in many
physical contexts in which wave-like behaviour occurs in one part of a domain and not in another.
In Quantum mechanics it arises in the context of the Schr odinger Equation
_
d
2
dz
2
+E V (z)
_
(z) = 0, (6.75)
where is the wave function, E is the energy and V (z) is a potential well. A similar equation
arises in the context of internal gravity waves in a density stratied uid (a useful example because
I will show you animations of internal waves impinging on a turning point and of internal wave
tunneling). Here the governing equation for the velocity stream function for two-dimensional waves
(vertical-horizontal plane), excluding the eects of the Earths rotation, is
tt
+N
2
(z)
xx
= 0. (6.76)
Here N(z), given by
N
2
(z) =
g
0
d
dz
(z), (6.77)
is called the buoyancy frequency, (z) is the uid density, g is the acceleration of gravity and
0
is a reference density. Since the partial dierential equation for is linear and the coecients are
independent of x and t one can look for solutions of the form
= e
i(kxt)
(z), (6.78)
which leads to
_
d
2
dz
2
+
N
2
(z)
2
2
k
2
_
(z) = 0. (6.79)
Both of equations (6.75) and (6.79) have the form
+Q(z) = 0. (6.80)
If Q(z) > 0 the solution is oscillatory, i.e., has wave-like behaviour. If Q(z) < 0 the solution
is exponential, or non-wave-like. If Q(z) changes sign then the behaviour of the solution is dierent
in dierent regions. A boundary between the wave-like and non wave-like regions, i.e., points where
Q(z) = 0, is called a turning point in 1-D. Tunneling, a classical problem in Quantum Mechanics,
occurs when two turning points are present, with a narrow non-wave-like region separating two
wave-like regions. Waves can tunnel from one wave-like region to another through the barrier
separating them.
64
0
region I
region III
region II
Figure 6.5: Schematic of the solution of the turning point problem.
Goal: Our goal is to nd a uniformly valid asymptotic solution to
zz
+Q(z) = 0, (6.81)
where
Q(z) is
_
> 0, for z < 0;
< 0, for z > 0;
(6.82)
and has a simple root at z = 0 (i.e., Q is linear near z = 0).
When z < 0 has an oscillatory behaviour (i.e., wave-like), while for z > 0 behaves expo-
nentially (non-wave-like). We assume that far from z = 0, Q varies slowly compared to the scale
on which varies. Thus we assume
Q = Q(), (6.83)
where
= z, (6.84)
is the slow space variable. Rewriting (6.81) in terms of gives
2
d
2
d
2
+Q() = 0. (6.85)
The asymptotic solution of (6.85) will be found in three dierent regions (see Figure 6.5). Solutions
in adjacent regions must be matched in overlapping regions of validity. In regions I and III the
solution is given by WKB Theory. These solutions are invalid in region II. The solution in region
I can be interpreted as the combination of incident and reected waves and we will determine the
relationship between these two waves.
6.5.1 WKB Theory: Outer Solution
In WKB theory we seek a solution of the form
= e
i
S
0
()
+S
1
()+S
2
()+
, (6.86)
such that
S
0
S
1
S
2
, (6.87)
65
and
S
2
1, (6.88)
as 0. If (6.87) and (6.88) are satised, e
i(S
2
+ )
1 and e
i
S
0
+S
1
as 0.
Substituting (6.86) into (6.85) we obtain
_
Q() S
2
0
+(iS
0
2S
0
S
1
) +
2
(iS
1
S
2
1
2S
0
S
2
) +
= 0. (6.89)
We now proceed by setting the coecient of
n
(n = 0, 1, 2, . . .) to zero.
O(1) Problem:
At O(1) we have
S
2
0
= Q(). (6.90)
There are two cases to consider:
case (a): < 0 S
0
() =
_
Q() =S
0
=
_
0
_
Q(t)dt.
case (b): > 0 S
0
() = i
_
Q() =S
0
= i
_
0
_
Q(t)dt.
Constants of integration result in a multiplicative factor of . At the moment we are interested in
nding up to a constant factor so these constants of integration are not of interest. Thus, the
lower limit of integration is arbitrary and we take it to be zero.
O() Problem:
At O() we have
S
1
=
i
2
S
0
S
0
=
i
2
d
d
ln(|S
0
|) S
1
=
i
2
ln(|S
0
|) = i ln(|Q()|
1/4
). (6.91)
WKB Solution:
Combining the rst two terms of the solution gives
_
Q()
1/4
e
Q(t)dt
, for < 0;
(Q())
1/4
e
Q(t)dt
, for > 0;
(6.92)
as 0. This is what is meant by the WKB solution. Note that it predicts an amplitude |Q()|
1/4
which becomes innite as 0, i.e., as the turning point is approached. This will render the WKB
solution invalid when becomes suciently small. A vertical wavenumber
m =
_
0
_
Q(t)dt =
_
Q()
can be dened in the region < 0. The wavelength of the oscillations in the region < 0 varies
with and is approximately 2/m(). The amplitude is then proportional to m
1/2
. Note that
this wavenumber goes to zero as the turning point is approached which implies that the length
scale of the oscillations goes to innity. Our solution violates our assumption that Q varies slowly
compared with the scale that varies on. In the region > 0
_
Q() denes the decay scale for
the exponential behaviour. Similar comments apply to the breakdown of the solution.
66
6.5.2 Region of Validity for WKB Solution
The asymptotic approximation for given by (6.92) is valid only in regions where (6.87) and (6.88)
are satised. We need to nd these regions. For || 1
Q a +b
2
, (6.93)
where, by assumption a < 0, so for < 0 we have
1
_
0
_
Q(t) dt
1
_
0
(at)
1/2
_
1 +
b
a
t
_
1/2
dt
1
_
0
(at)
1/2
_
1 +
1
2
bt
a
+
_
dt
=
1
_
2
3
a()
3/2
1
5
b
a
()
5/2
+
_
.
(6.94)
For > 0
1
_
0
_
Q(t)dt
1
_
2
3
a
3/2
1
5
b
5/2
+
_
. (6.95)
Thus
S
0
2
3
a
()
3/2
as 0
2
3
i
3/2
as 0
+
.
(6.96)
For future reference note that
e
i
S
0
e
i
2
3
|a|
||
3/2
||
5/2
/ 1 ||
2/5
. (6.97)
Similarly we nd
S
1
1
4
|a|
1/4
ln || (6.98)
and
S
2
5
48
_
|a|
||
3/2
, (6.99)
as 0. The WKB solution is valid if S
2
1, S
1
S
0
and S
2
S
1
. The rst of these requires
S
2
1 ||
3/2
1 ||
2/3
, (6.100)
and the second requires
S
0
S
1
||
3/2
/ ln || 1. (6.101)
Setting || =
2/3
with > 0, we have ln|| =
2
3
ln and ||
3/2
=
. Thus we need
1
ln
1 as
0 which requires < 1. Thus
S
0
S
1
if ||
2/3
. For the third condition, S
1
S
2
, we
also nd that ||
2/3
is again necessary.
Therefore, the WKB Solution is valid if
||
2/3
. (6.102)
The approximation to the WKB Solution, given by (6.96), (6.98) and (6.99) is valid if ||
2/5
so the approximate WKB solution is valid if
2/3
||
2/5
.
67
6.5.3 Inner Solution
When || = 0(
2/3
) the WKB solution is invalid. We need a new solution that is valid for smaller
values of ||. Since
2/3
1 we will nd a solution valid for || 1. Then the WKB solution
and the new solution will both be valid in
2/3
|| 1. It will, however, be more convenient to
match the solutions using the approximate WKB solution. To do this we simply rene the matching
region to
2/3
||
2/5
.
For || 1, Q() a in which case can be approximated by
II
which is a solution of
d
2
II
d
2
=
a
2
II
. (6.103)
This is a scaled version of Airys equation.
The Airy Equation:
y
II
= CA
i
_
_
a
2
_
1/3
_
+DB
i
_
_
a
2
_
1/3
_
. (6.104)
Since we want to match
II
to an exponentially decaying solution in region III we will have to take
D = 0 (on physical grounds if there is no energy source to the right of the turning point this must
be the case if the solution is to remain bounded. If there is an energy source or a second turning
point Bi(x) may need to be included). A
i
(x) is the unique solution of the Airy equation with the
asymptotic behaviour
A
i
(x)
1
2
x
1/4
e
2/3x
3/2
as x +. (6.105)
For x < 0 we have
A
i
(x) = w
1
(x) sin
_
2
3
(x)
3/2
+
4
_
w
2
(x) cos
_
2
3
(x)
3/2
+
4
_
, (6.106)
where
w
1
(x)
1
(x)
1/4
and w
2
(x)
7/4
as x . (6.107)
6.5.4 Matching
We now know the asymptotic behaviour for in each of the three regions.
Region III:
2/3
III
= B(Q())
1/4
e
Q(t)dt
(6.108)
Region II: || 1
II
= CA
i
_
_
a
2
_
1/3
_
(6.109)
Region I:
2/3
I
= Q()
1/4
_
Ee
i
Q(t)dt
+
Fe
Q(t)dt
_
. (6.110)
Here we have already used the fact that we require to decay exponentially as +. We now
need to relate the constants B, C,
E and
F.
68
Matching
III
III
III
and
II
II
II
:
When
2/3
, (a)
1/3
2/3
1 and
II
C
2
_
_
a
2
_
1/3
_
1/4
e
2
3
1/3
3/2
=
C
2
(a)
1/12
1/6
1/4
e
2
3
(a)
1/2
3/2
.
(6.111)
On the other hand
III
B(a)
1/4
1/4
e
2
3
3/2
if
2/5
(see 6.95). (6.112)
The asymptotic expressions (6.111) and (6.112) are both valid when
2/3
2/5
. Thus we
must have
C
2
(a)
1/12
1/6
= B(a)
1/4
C =
2
1/6
(a)
1/6
B. (6.113)
Matching
II
and
I
:
As above the matching region is
2/3
||
2/5
. For
2/5
2/3
,
II
=
2
(a)
1/6
BA
i
_
_
a
2
_
1/3
2B
(a)
1/6
_
_
a
2
_
1/3
()
_
1/4
sin
_
2
3
_
_
a
2
_
1/3
()
_
3/2
+
4
_
= 2B(a)
1/4
sin
_
2
3
(a)
1/2
()
3/2
+
4
_
. (6.114)
Next, on
2/3
2/5
, we also have
I
(a)
1/4
_
Ee
2
3
(a)
1/2
()
3/2
+
Fe
i
2
3
(a)
1/2
()
3/2
_
,
= (a)
1/4
_
E sin
_
2
3
()
3/2
+
4
_
+F cos
_
2
3
()
3/2
+
4
_
_
. (6.115)
Evidently, comparing (6.114) and (6.115) F = 0 and E = 2B.
6.5.5 Summary of asymptotic solution
_
_
B[Q()]
1/4
e
Q(t)dt
;
2/3
,
2
(a)
1/6
BA
i
_
2/3
(a)
1/3
;
_
|| 1,
2B[Q()]
1/4
sin
_
1
_
0
_
Q(t)dt +
4
_
;
2/3
.
(6.116)
[Note in third expression integral now goes from to 0 because
1
0
_
Q(t)dt
1
2
3
a()
3/2
so
1
3
_
0
_
Q(t)dt
1
2
3
a()
3/2
as in (6.115).]
69
6.5.6 Physical Interpretation
Consider the region
2/3
. Using w = cos(kx t) where w =
x
vertical velocity of uid
in the presence of IGWs then
w = cos(kx t) 2Bm
1/2
cos(kx t) sin
_
1
_
0
m(t)dt +
4
_
= Bm
1/2
_
sin
_
kx
1
_
0
m(t)dt t +
4
_
+ sin
_
kx
1
_
0
m(t)dt +t +
4
__
.
Using z = / and redening m(z) =
_
Q(z) we can rewrite this as
w =
B
m
1/2
_
sin
_
kx
_
z
0
m(z
) dz
t +
4
_
sin
_
kx +
_
z
0
m(z
) dz
t
4
__
(6.117)
If m = m
0
is constant for |z| > z
0
then
_
z
0
m(t)dt =
_
z
0
0
m(t)dt +
_
z
z
0
m(t)dt =
+m
0
(z z
0
) = +m
0
z (6.118)
where
and =
mz
0
are constants. Thus we have
w = Bm
1/2
_
sin
_
kx m
0
z t +
4
_
sin
_
kx +m
0
z t +
4
_
_
. (6.119)
The rst term represents a wave with phase speed /m
0
in the z direction while the second term
represents a wave with phase speed /m
0
. These are reected and incident waves respectively and
they both have the same amplitude. The two waves have dierent phases, hence at the turning
point there is perfect reection with a phase shift. Note that physically we should expect perfect
reection. As the solution goes to zero exponentially fast, hence there can be no energy
far to the right. Since the system has no dissipation all the incident energy must be reected.
6.6 Tunneling
70
Chapter 7
Singular Perturbation Theory:
Examples and Techniques
7.1 More examples of problems from Singular Perturbation The-
ory
We have seen a few problems where Regular Perturbation Theory fails:
1. Solving equations of the form
x
3
+ 2x 1 = 0. (7.1)
RPT fails in this case because when = 0 one of the roots is lost. One root goes to as
0, i.e., = 0 is a singular point of one of the solutions.
2. ODEs such as
y
+y
= 0,
y(0) = a,
y(1) = b.
(7.2)
As weve seen solutions have a thin boundary layer in the vicinity of one of the boundaries
where the solution changes very rapidly. In this boundary layer y
n
a
n
(x, y)
n
() as 0
+
, (7.3)
for an Asymptotic Sequence {
n
} and for (x, y) D is uniformly ordered if
a
n
(x, y)
n
() = O(
n
()), (7.4)
uniformly on D, for all n.
71
Theorem 7.1.1 The Asymptotic Expansion is uniformly ordered if the a
n
(x, y) are all bounded.
Example 7.1.1
f(, ) 1
1
(1 )
3
2
as 0 (7.5)
for 1 < < 2 is not uniformly ordered.
Uniform ordering requires that
_
1
192
(cos cos 3) +
16
sin
_
< A, (7.7)
for all . It is uniformly ordered on any nite interval.
Denition 7.1.2 An Asymptotic Expansion that is not uniformly ordered is said to be disordered
(or to become disordered), or non-uniform. Singular Perturbation Theory is used to deal with
problems with non-uniformities.
Common sources of non-uniformity
1. Innite domains.
2. Singularities in the DE.
3. Small parameters multiplying the highest derivative in a DE.
4. Change in type of PDE.
many others
Example 7.1.3 Special case of an example of Lighthills:
(x +y)y
+y = 1,
y(1) = 2,
(7.8)
for 0 x 1 and 0 < 1.
72
0.5 0.0 0.5 1.0 1.5
x
10
5
0
5
10
15
20
y
Figure 7.1: Comparison of exact and RPT solutions of Lighthills example. Solid curve: exact solution. O(1) RPT
solution: dotted curve. O() RPT solution: dashed curve. The sloping straight line is the singular line y = x/.
This has the exact solution
y =
x
+
_
_
x
_
2
+
2(1 +x)
+ 4. (7.9)
Trying Regular Perturbation Theory methods by setting
y = y
0
+y
1
+
2
y
2
+ , (7.10)
leads to
y =
1 +x
x
(1 x)(1 + 3x)
2x
3
+O
F
(
2
). (7.11)
The Asymptotic Expansion is disordered. For x close to zero the O() term becomes larger than
the leading term. What is the source of the non-uniformity? Note that each term in the asymptotic
expansion is singular at x = 0. In contrast the exact solution is well behaved at x = 0. The exact
solution, however, is singular at = 0. Regular Perturbation Theory assumes the solution is well
behaved at = 0, and hence has a power series expansion about = 0.
The DE has a line of singularities along x+y = 0. In the Regular Perturbation Theory solution
the singularity is at x = 0. The shift of the singularity from x + y = 0 to x = 0 is the source of
the diculty. A Perturbation Theory scheme should keep singularities as close as possible to their
original location.
Example 7.1.4 Consider the Boundary Value Problem
u
xx
+u
yy
u
y
= 0, (7.12)
where > 0 and 0 x, y 1 with boundary conditions
u(0, y) = a(y) u(x, 0) = b(x),
u(1, y) = c(y) u(x, 1) = d(x).
(7.13)
This is a Dirichlet problem if > 0. It is well posed in the sense of Hadamard (the solution exists,
is unique, and depends continuously on the boundary conditions).
Trying Regular Perturbation Theory, at leading order we have
u
0xx
u
0y
= 0, (7.14)
73
u
0
(x, y) satises the same boundary conditions as u(x, y) does, however the PDE is no longer elliptic,
it is parabolic and cannot in general satisfy all the boundary conditions. Problems where the PDEs
change type in limiting cases of interest are common. One example is uid ow. Compressible
uid ow is of hyperbolic character. There is a maximum speed at which information can travel,
namely the sound speed. In the incompressible limit, a limiting case commonly studied, the sound
speed goes to innity. The governing equations are now of elliptic character.
Example 7.1.5
4
u = 0 on D
u = f on D
u
n
+u = g on D.
(7.15)
Setting = 0 to gives the reduced problem in RPP, which, if f = g has no solution. If f = g
then we have insucient boundary conditions and the solution is non-unique.
7.2 The linear damped oscillator
The equations for the linear damped oscillator, which arise in many contexts (e.g., mass-spring
system, RLC-circuit), are
m x + 2 x +kx = 0,
x(0) = x
0
,
x = v
0
= 0,
(7.16)
where m, and k are all positive. For a damped mass-spring system they represent the mass,
damping due to friction, and the linear spring constant respectively, while x is the amount the
spring has been stretched. For an RLC circuit the coecients represent the inductance, resistance
and inverse capacitance, while x represents the charge on the capacitor. This linear second-order
ODE can be easily solved analytically and its complete solution and behaviour should be familiar
to you. We will consider it in some detail from a perturbation theory perspective for which we
need to introduce a small parameter. We will consider several approaches, some of which will fail.
Knowledge of the exact solution will allow us to understand why those that fail do, giving some
insight into possible pitfalls that can occur in more dicult problems for which exact solutions are
unknown.
Nondimensionalization
The units of the various terms in the problem are
[t] = T, [x] = L, [m] = M, [] =
M
T
, [k] =
M
T
2
. (7.17)
1. L
C
= x
0
is the only possible choice for the length scale since x is the only variable that
involves dimensions of length. For all solutions the maximum value of |x| is x
0
. Hence, dene
the non-dimensional displacement
y =
x
x
0
. (7.18)
2. There are several choices for the time scale:
(a) T
c
= T
(1)
c
=
_
m
k
74
(b) T
c
= T
(2)
c
=
k
(c) T
c
= T
(3)
c
=
m
.
Well consider each in turn. All three non-dimensionalizations will include a single nondimensional
positive parameter
=
mk
. (7.19)
For weak damping is the small parameter in our non-dimensional problem. For strong damping
1
is the small parameter. Note that the time scales are related by
T
(1)
c
=
1
T
(2)
c
= T
(3)
c
. (7.20)
Case (a): Set t = T
(1)
c
=
_
m
k
. This gives
y
+ 2y
+y = 0,
y(0) = 1,
y
(0) = 0.
(7.21)
1. If 1 we can try to apply the methods of Regular Perturbation Theory by setting y =
y
0
() +
1
() +
2
y
2
() + . At leading order we obtain y
0
= cos . At the next order we get
y
1
+y
1
= 2 sin ,
y
1
(0) = y
1
= 0.
(7.22)
This has a resonant forcing. Therefore we need methods of Singular Perturbation Theory if
we are interested in the solution for times of O(
1
) or longer. For much shorter times this
method gives a useful approximation.
2. If 1 write the ODE as
1
+ 2y
+
1
y = 0, (7.23)
and set y = y
0
+
1
y
1
+
1
2
y
2
+ .
O(1):
y
0
= 0,
y
0
(0) = y
0
(0) = 0
(7.24)
The solution is y
0
() = 1. We are lucky that the two initial conditions for a rst-order
DE are satised. For general initial conditions there would be no solution.
O(
1
): At the next order the problem for y
1
is
y
1
=
1
2
(y
0
+y
0
) =
1
2
,
y
1
(0) = y
1
(0) = 0.
(7.25)
The DE and the second initial condition are inconsistent. Hence, the O(
1
) problem
has no solution.
75
Case (b): Next consider the time scale /k. Setting t = T
(2)
c
s =
k
s we have
y
ss
+ 2
2
y
s
+
2
y = 0,
y(0) = 1,
y
s
(0) = 0.
(7.26)
1. If 1 set y = y
0
+
2
y
1
+
4
y
2
+ .
O(1):
y
ss
= 0
y
0
(0) = 1
y
0s
(0) = 0
_
_
_
y
0
(s) = 1. (7.27)
O():
y
1ss
= 2y
0s
y
0
= 1
y
0
(0) = y
1s
(0) = 0
_
y
1
(s) =
s
2
2
(7.28)
y = 1 s
2
2
/2 + which becomes disordered when s = O(
1
). Note also that small
1 corresponds to weak damping for which the solution is oscillatory. This does not look
promising however it it a valid approximation for suciently short times.
2. If 1 write ODE as
1
2
y
ss
+ 2y
s
+y = 0, (7.29)
and set y = y
0
+
1
2
y
1
+
1
4
y
2
+ .
O(1):
_
2y
0s
+y
0
= 0
y
0
(0) = 1, y
0s
(0) = 0
no solution (7.30)
Case (c): Set t = T
(3)
c
=
m
. We now have
2
y
+ 2
2
y
+y = 0
y(0) = 1,
y
(0) = 0.
(7.31)
1. If
2
1, O(1) problem is y
0
() = 0 which cant satisfy the initial conditions y
0
(0) = 1.
Hence, there is no solution.
2. If 1 write the DE as
y
+ 2y
+
1
2
y = 0, (7.32)
and set y = y
0
+
1
2
y
1
+
1
4
y
2
+ .
O(1):
y
0
+ 2y
0
= 0,
y
0
(0) = 1,
y
0
(0) = 0.
(7.33)
Integrating the DE y
0
=
Ae
2
y
0
() = Ae
2
+C. The initial conditions give A = 0
and C = 1. Hence y
0
() = 1.
76
O(
2
):
y
1
+ 2y
1
= y
0
= 1,
y
1
(0) = y
1
(0) = 0.
(7.34)
Integrating the DE gives y
1
+ 2y
1
= + C. From the two initial conditions C = 0.
Integrating again gives y
1
= Ae
2
+
1
4
2
. The initial condition y
1
(0) = 0 gives A =
1
4
.
Thus, the RPT solution to O(
2
) is
y = 1 +
_
1
4
(1 e
2
)
2
_
1
2
+ , (7.35)
which becomes disordered after a time of O(
2
).
Summary of RPT solutions:
Case (a): T
c
= T
(1)
c
=
_
m
k
:
1. weak damping Regular Perturbation Theory fails at times of O(
1
) due to secular terms.
2. strong damping no solution
Case (b): T
c
= T
(2)
c
=
k
:
1. weak damping disordered by s = O(
1
)
2. strong damping no solution
Case (c): T
c
= T
(1)
c
=
m
:
1. weak damping no solution
2. strong damping disordered when = O(
2
).
Exact Solution & Discussion for 1
For the exact solution the choice of nondimensionalization is irrelevant. Using the time scale
T
(1)
c
we have
y
+ 2y
+y = 0,
y(0) = 1,
y
(0) = 0,
where = /
mk.
The solution has three dierent forms depending on the size of .
Case I: < 1 (underdamped).
In dimensional time the exact solution is
y = e
t/T
(3)
c
_
cos
_
_
1
2
t
T
(1)
c
_
+
1
2
sin
_
_
1
2
t
T
(1)
c
_
_
(7.36)
The solution consists of oscillations on the time scale T
(1)
c
with an amplitude that decays on the
time scale T
(3)
c
. Since T
(1)
c
= T
(3)
c
, for 1 the time scale of the oscillations is much shorter than
the time scale of the amplitude decay.
Regular Perturbation Theory runs into diculty for two reasons.
77
Figure 7.2: Exact solution of the linear damped oscillator for m = k = 1. (a) Underdamped case with = 0.05.
(b) Overdamped case with = 10. (c) x for overdamped case.
1. The frequency of the oscillations depends on . This results in secular forcing terms and the
solution breaks down after a time of O(
1
). We have seen this behaviour before when we
consider the simple nonlinear pendulum.
2. The solution includes behaviour on two very dierent time scales: the amplitude decays on a
slow time scale which signicantly modies the solution after times of O(T
(3)
c
). Introducing
a time scale implicitely assumes the solution evolves on a single time scale. In reality we
need two time scales for the weakly damped case. We will return to this when we study the
method of multiple scales.
Case II: = 1 (critically damped).
y = e
t/T
(3)
c
_
t
T
(3)
c
+ 1
_
. (7.37)
Since is neither large or small there is no small parameter to exploit. All time scales are identical
and all three terms in the ODE are important. Cant use perturbation methods.
Case III: > 1 (overdamped).
78
Figure 7.3: Exact solution of the linear damped oscillator for m = k = 1 on the phase plane. (a) Underdamped
case with = 0.05. (b) Overdamped case with = 10.
The exact solution now has the form
y =
1
2
_
1 +
2
1
_
e
(
2
1)t/T
(1)
c
+
1
2
_
1
2
1
_
e
(+
2
1)t/T
(1)
c
.
(7.38)
For 1,
2
1 =
_
1
1
2
1
2
and +
2
1 2. Hence, for large
y e
t/(2T
(1)
c
)
1
4
2
e
2t/T
(1)
c
,
= e
t/(2T
(2)
c
)
1
4
2
e
2t/T
(3)
c
.
(7.39)
The rst term decays on the slow time scale T
(2)
c
while the second term decays on the fast time
scale T
(3)
c
(since T
(2)
c
=
2
T
(3)
c
T
(3)
c
).
When we nondimensionalized using time scale T
(1)
c
we could not obtain a solution. The solution
does not involve this time scale!. In terms of the exact solution is
y e
/(2)
1
4
2
e
2
. (7.40)
As the second term goes to zero faster than any power of
1
When we nondimensionalized using time scale T
(2)
c
, (Case (b)), the reduced problem was
2y
0s
+y
0
= 0,
y
0
(0) = 1,
y
0s
(0) = 0.
(7.41)
which has no solution.
In terms of s, (7.39) is
y e
s/2
1
4
2
e
2
2
s
. (7.42)
79
The second term is exponentially small for large , (s > 0), and is lost in the Regular Perturbation
Theory problem. Yet, it is needed at s = 0 to satisfy the two initial conditions.
Nondimensionalizing using T
(3)
c
, the exact solution is
y =
1
2
_
_
1 +
1
_
1
1
2
_
_
e
1
1
+
1
2
_
_
1
1
_
1
1
2
_
_
e
1+
1
1
.
(7.43)
Expanding
_
1
1
2
and 1/
_
1
1
2
in powers of
1
2
,
y =
_
1 +
1
4
2
+
_
e
/(2
2
)+
+
_
1
4
2
+
_
e
2
e
1
2
2
+
,
= 1 +
_
1
4
_
1 e
2
_
2
_
1
2
+O
F
(
1
4
).
(7.44)
which recovers the Regular Perturbation Theory solution. The solution becomes disordered because
the Taylor Series expansion
e
2
2
= 1
2
2
+
1
8
4
+ , (7.45)
which converges as the number of terms goes to for xed , is a disordered Asymptotic Expansion
when is O(
2
).
7.3 Method of Multiple Scales
The linear damped oscillator is an example of a physical system that varies on more than one time
scale. There many other such examples - far too numerous to list. The simple nonlinear pendulum
is one such example. To see this consider the approximate solution we obtained using the Method
of Strained Coordinates:
(t) a cos
__
g
_
1
a
2
16
_
t
_
+
a
3
192
_
cos
__
g
_
1
a
2
16
_
t
_
cos
_
3
_
g
_
1
a
2
16
_
t
__
.
(7.46)
Using cos(A B) = cos Acos B + sinAsin B we can write this as
(t) a cos
__
g
a
2
16
t
_
cos
__
g
t
_
+ sin
__
g
a
2
16
t
_
sin
__
g
t
_
+
a
3
192
_
cos
__
g
a
2
16
t
_
cos
__
g
t
_
+ sin
__
g
a
2
16
t
_
sin
__
g
t
_
cos
_
3
_
g
a
2
16
t
_
cos
_
3
_
g
t
_
sin
_
3
_
g
a
2
16
t
_
sin
_
3
_
g
t
_
_
.
(7.47)
80
Introducing a slow time scale = a
2
t (slow because changes by O(1) when t changes by O(1/a
2
which is very large for small a) we can write this as
(t) a cos
_
_
1
16
g
_
cos
__
g
t
_
+ sin
_
1
16
_
g
_
sin
__
g
t
_
+
a
3
192
_
cos
_
1
16
_
g
_
cos
__
g
t
_
+ sin
_
1
16
_
g
_
sin
__
g
t
_
cos
_
3
16
_
g
_
cos
_
3
_
g
t
_
sin
_
3
16
_
g
_
sin
_
3
_
g
t
_
_
.
(7.48)
We can interpret this as fast oscillations on the time scale t multiplied by slowly varying amplitudes
which are functions of the slow time scale .
We will now solve the simple nonlinear pendulum using the method of multiple scales by as-
suming at the outset that the behaviour of the system depends on more than one time scale. This
is the method of multiple scales. It is an extremely powerful method with widespread applications.
7.3.1 The Simple Nonlinear Pendulum
Recall the scaled problem
+
sin(a)
a
= 0,
(0) = 1,
(0) = 0,
(7.49)
The nonlinearity in the system causes a slow drift (period not exactly 2) so we can think of
varying on two time scales: one corresponding to the fast oscillations and one corresponding to the
time scale of the drift. We assume the solution depends on two time scales t and = a
2
t. With
this in mind we try looking for a solution of the form
= f(t, t; a
2
) = f(t, ; a
2
) (7.50)
The idea is to treat t and as two independent variables. The chain rule gives
d
dt
= f
t
+a
2
f
,
d
2
dt
2
= f
tt
+ 2a
2
f
t
+a
4
f
,
(7.51)
so the the problem in terms of f becomes
f
tt
+ 2a
2
f
t
+a
4
f
+f a
2
f
3
6
+ = 0,
f(0, 0; a
2
) = 1,
f
t
(0, 0; a
2
) +a
2
f
(0, 0; a
2
) = 0.
(7.52)
As usual look for solutions of the form
f = f
0
(t, ) +a
2
f
1
(t, ) +a
4
f
2
(t, ) + . (7.53)
81
O(1) Problem:
f
0tt
+f
0
= 0,
f
0
(0, 0) = 1,
f
0t
(0, 0) = 0,
(7.54)
which has the solution
f
0
= A() cos(t) +B() sin(t) (7.55)
where
A(0) = 1 B(0) = 0. (7.56)
O(a
2
) Problem:
f
1tt
+f
1
= 2f
0t
+
1
6
f
3
0
,
f
1
(0, 0) = 0,
f
1t
(0, 0) +f
0
(0, 0) = 0.
(7.57)
After some algebra to evaluate the forcing terms we get
f
1tt
+f
1
=
_
2A
() +
1
8
(A
2
B +B
3
)
_
sin t
+
_
2B
() +
1
8
(AB
2
+A
3
)
_
cos t
+
1
24
(A
3
3AB
2
) cos 3t
1
24
(B
3
3A
2
B) sin 3t.
(7.58)
We must now eliminate the secular terms. This gives
A
() =
1
16
(A
2
+B
2
)B,
B
() =
1
16
(A
2
+B
2
)A,
(7.59)
We now have two coupled nonlinear ODEs to solve!
Multiplying the rst by A, the second by B and adding gives
d
d
_
1
2
(A
2
+B
2
)
_
= 0, (7.60)
so A
2
+B
2
is constant. From the initial conditions A
2
+B
2
= 1 so
A
() =
1
16
B,
B
() =
1
16
A,
(7.61)
We now have a couple set of linear ODEs which are easily solved. Eliminating B gives
A
() +
1
16
2
A = 0. (7.62)
82
The solution is
A() = cos(
16
),
B() = sin(
16
),
(7.63)
The O(1) solution is
f
0
= cos
_
16
_
cos t + sin
_
16
_
sint,
= cos
_
t
16
_
,
= cos
_
(1
a
2
16
)t
_
.
(7.64)
Using the method of strained coordinates we obtained
(t) = cos
_
(1
a
2
16
+ )t
_
+
a
2
192
_
cos
_
(1
a
2
16
+ )t
_
cos
_
3
_
1
a
2
16
+
_
t
__
+O
F
(a
4
).
(7.65)
The method of multiple scales has recovered the same rst term with identical frequencies to O(a
2
).
With the resonant forcing terms eliminated in the O(a
2
) problem the problem for f
1
simplies
to
f
1tt
+f
1
=
1
24
(A
3
3AB
2
) cos 3t
1
24
(B
3
3A
2
B) sin 3t,
f
1
(0, 0) = 0,
f
1t
(0, 0) +f
0
(0, 0) = 0.
(7.66)
Using the known forms for A and B this can be rewritten as
f
1tt
+f
1
=
1
24
cos(3) cos 3t +
1
24
sin(3) sin 3t,
f
1
(0, 0) = 0,
f
1t
(0, 0) = 0,
(7.67)
which has the general solution
f
1
(t, ) =
1
192
cos(3) cos 3t
1
192
sin 3 sin 3t
+A() cos t +B() sin t,
(7.68)
where A and B are new unknown functions whose initial conditions are determined by the initial
conditions at (t, ) = (0, 0).
Note that the rst two terms combine to give
1
192
cos(3) cos 3t
1
192
sin 3 sin3t =
1
192
cos
_
(1
a
2
16
)3t
_
(7.69)
83
which is again in agreement with our previous solution.
The new unknown function A and B can be determined by eliminating the resonant forcing
in the O(a
4
) problem. To go to higher order, however, will require the introduction of an even
longer time scale
1
= a
4
t, consistent with our previously obtain amplitude dependent frequency
= 1 a
2
/16 +O(a
4
).
7.4 Methods for Singular Perturbation Problems
7.4.1 Method of Strained Coordinates (MSC)
This method, which we have already seen, is used to deal with secular terms. It is the technique
we used to solve the nonlinear pendulum problem. Here we consider a similar example, the only
real twist being the use of non-zero initial conditions for the rst derivative x which results in the
initial conditions making non-zero contributions to the higher-order problems.
Example 7.4.1 (The free Dung oscillator)
x +x +x
3
= 0,
x(0) = 0,
x(0) = v,
(7.70)
where 0 < 1.
This obeys the conservation law
dE
dt
=
d
dt
_
1
2
x
2
+
1
2
x
2
+
x
4
4
_
= 0, (7.71)
hence
E(t) = E(0) =
1
2
v
2
. (7.72)
It follows that
1
2
x
2
+
4
x
4
1
2
x
2
+
1
2
x
2
+
4
x
4
=
1
2
v
2
, (7.73)
hence x is bounded for all time.
Regular Perturbation Theory Solution:
O(1): The leading-order problem is
x
0
+x
0
= 0
x
0
(0) = 0
x
0
(0) = v
_
_
_
x
0
= v sint. (7.74)
O(): At the next order we get
x
1
+x
1
= x
3
0
= v
3
sin
3
t,
=
v
3
4
(sin 3t 3 sin t),
x
1
(0) = x
1
(0) = 0.
(7.75)
84
The sin t term is a secular term which leads to unbounded growth in the amplitude of the oscillations.
The solution fails (becomes disordered) when t = O(
1
).
Method of Strained Coordinates: We now allow the time scale to depend on , as we did for
the nonlinear pendulum problem. Thus, we seek a solution of the form
x = x(; ) = x
0
() +x
1
() +
2
x
2
() + , (7.76)
where
() = ()t = (1 +
1
+
2
2
+ )t, (7.77)
i.e., let the frequency be a function of . Because the energy in the system is v
2
/2, we should
expect the frequency of the oscillations to also depend on v. The ODE becomes
(1 +
1
+
2
2
+ )
2
x
+x +x
3
= 0,
x(0) = 0,
(1 +
1
+
2
2
+ )x
(0) = v.
(7.78)
Because the initial velocity is non-zero, there will be non-zero contributions to the higher-order
initial conditions.
O(1): The leading-order problem is
x
0
+x
0
= 0
x
0
(0) = 0,
x
0
(0) = v
_
_
_
x
0
= v sin . (7.79)
O(): At the next order we have
x
1
+x
1
= x
3
0
2
1
x
0
,
=
v
3
4
(sin(3) 3 sin ) + 2
1
v sin ,
=
v
3
4
sin(3) +
_
2
1
v
3
4
v
3
_
sin .
(7.80)
The secular terms can be eliminated by setting
1
=
3
8
v
2
, (7.81)
leaving us with
x
1
+x
1
=
v
3
4
sin 3, (7.82)
which has the general solution
x
1
=
v
3
32
sin3 +Asin +Bcos . (7.83)
The initial conditions are
x
1
(0) = 0,
x
1
(0) +
1
x
0
(0) = 0,
(7.84)
85
the latter of which is
x
1
(0) =
1
v =
3
8
v
3
. (7.85)
From these A = 9v
3
/32 and B = 0, hence
x
1
=
v
3
32
sin 3
9
32
v
3
sin . (7.86)
O(
2
): At the next order the DE and initial conditions give
x
2
+x
2
= (2
2
+
2
1
)x
0
2
1
x
1
3x
2
0
x
1
,
x
2
(0) = 0,
x
2
(0) +
1
x
1
(0) +
2
x
0
(0) = 0.
(7.87)
Using the previous solutions the DE is
x
2
+x
2
=
_
2
2
v +
69
128
v
5
_
sin
75
128
v
5
sin3
3
128
v
5
sin 5. (7.88)
The secular terms are eliminated by setting
2
=
69
256
v
4
. (7.89)
The solution of the resulting DE satisfying the boundary conditions is
x
2
() =
75
1024
v
5
sin(3) +
3
3072
v
5
sin(5) +
47
512
sin. (7.90)
The full solution to O(
2
) is
x = sin t
v
3
32
_
sin(3t) + 9 sin t
_
+
_
75
1024
v
5
sin(3t) +
3
3072
v
5
sin(5t) +
47
512
sin t
_
2
+O
F
(
3
),
(7.91)
where
= () = 1 +
3
8
v
2
69
256
v
4
2
+O
F
(
3
). (7.92)
7.4.2 The Linstedt-Poincare Technique
The Linstedt-Poincare Technique
is a generalization of the previous method;
was rst used in theory of nonlinear waterwaves by G. G. Stokes in 1847;
was proved by Poincare to have a uniformly valid asymptotic expansion in 1892. He credited
idea to an obscure paper by Linstedt;
only works for periodic motion;
86
is useful for nding oscillatory (periodic) solutions of DEs of the form
x +
2
0
x = f(t, x, x; ). (7.93)
A common problem amenable to this technique is free oscillations in a conservative systems.
Consider a regulating equation of the form
x +F(x) = 0, (7.94)
which has constant energy
E(x, x) =
1
2
x
2
+V (x), (7.95)
where
V =
_
x
0
F()d. (7.96)
The solution of
x +F(x) = 0,
x(0) = ,
x(0) = ,
(7.97)
is given by writing the conservation law as
x
2
= 2(E V ),
dt
dx
=
1
_
2(E V )
,
t =
_
x
d
_
2(E V )
.
(7.98)
Using the initial conditions this give t(x) which in principle can be inverted to get x(t). This is
usually impossible to do analytically. Perturbation Theory is useful for nding an approximate
solution provided F(x) involves a small parameter and the reduced problem can be solved..
Example 7.4.2 Consider F of the form
F(x) =
2
0
x f(x). (7.99)
Our problem is
x +
2
0
x = f(x),
x(0) = ,
x(0) = ,
(7.100)
for which we wish to nd periodic solutions. For simplicity we will assume = 0 for now. We will
consider non-zero later. Following standard MSC methods we introduce a scaled time scale
= ()t. (7.101)
giving
2
()x
+
2
0
x = f(x),
x(0) = ,
x
(0) = = 0,
(7.102)
87
and as usual expand x and via
x = x
0
+x
1
+
2
x
2
+ ,
=
0
+
1
+
2
2
+ .
(7.103)
Since x() is periodic, so are the x
n
(). Substituting the expansions in the DE gives
(
2
0
+ 2
0
1
+ (2
0
2
+
2
1
)
2
+ )(x
0
+x
1
+
2
x
2
+ )
+
2
0
(x
0
+x
1
+
2
x
2
+ ) = f(x
0
+x
1
+
2
x
2
+ ).
(7.104)
or
2
0
(x
0
+x
0
) +
_
2
0
(x
1
+x
1
) + 2
0
1
x
0
_
+
=
_
f(x
0
) +f
(x
0
)(x
1
+
2
x
2
+ ) +
_
.
(7.105)
O(1): The leading-order problem is
x
0
+x
0
= 0,
x
0
(0) = ,
x
0
(0) = 0,
(7.106)
which has the solution
x
0
= cos . (7.107)
O(): At the next order we have
x
1
+x
1
=
2
1
0
x
0
+
1
2
0
f(x
0
),
=
2
1
0
cos +
1
2
0
f(cos ),
x
1
(0) = x
1
(0) = 0
(7.108)
Since cos is an even periodic function, so is g() = f(cos ). Hence we can expand f(cos )
in a cosine series:
f(cos ) = a
0
+a
1
cos +a
2
cos 2 + , (7.109)
where
a
0
=
2
_
2
0
f(cos ) d, (7.110)
and
a
n
=
1
_
2
0
cos(n)f(cos ) d. (7.111)
Therefore
x
1
+x
1
=
a
0
2
0
+
_
2
1
0
+
a
1
2
0
_
. .
secular term
cos +
n=2
a
n
2
0
cos(n). (7.112)
88
To eliminate the secular terms take
1
=
a
1
2
0
, (7.113)
after which we can solve for x
1
to get
x
1
() =
a
0
2
0
n=2
a
n
cos n
2
0
(n
2
1)
+Acos +Bsin . (7.114)
The initial conditions then give
B = 0,
A =
a
0
2
0
+
n=2
a
n
0
(n
2
1)
.
(7.115)
Note: If x
2
0
+
_
a
1
2
0
+
2
1
0
_
cos +
_
b
1
2
0
+
2
1
0
_
sin + . (7.117)
To remove the secular terms we need both
1
=
a
1
2
0
, (7.118)
and
1
=
b
1
2
0
, (7.119)
which is only possible if a
1
= b
1
. This is always true! This is easy to prove. The dierence is
a
1
b
1
=
1
_
2
0
f(cos + sin ) cos d
_
2
0
f(cos + sin ) sin d,
=
1
_
2
0
f(cos + sin)(sin + cos ) d,
=
1
_
2
0
f(cos + sin)
_
d
d
(cos + sin)
_
d,
=
1
f(u) du = 0.
(7.120)
Theorem 7.4.1 (Periodicity of Solutions) For suciently small solutions to
x +
2
0
x = f(x) (7.121)
are periodic and admit a Linsted-Poincare expansion, i.e. a uniformly ordered Asymptotic Expan-
sion (see text by Murdock).
89
7.4.3 Free Self Sustained Oscillations in Damped Systems
Free self sustained oscillations in damped system arise through a combination of damping and
forcing. A general 1-D problem is
x +
2
0
x = f(x, x), (7.122)
along with initial conditions.
Example 7.4.3 A special case of Rayleighs equation is
x +x =
_
x
1
3
x
3
_
. (7.123)
Multiplying by x gives the energy equation
d
dt
_
1
2
x
2
+
1
2
x
2
_
= x
2
_
1
1
3
x
2
_
. (7.124)
The energy increases in time if x
2
< 3 and decreases in time if x
2
> 3. It is plausible that
periodic solutions exist. Lets assume a periodic solution exists and try to nd it. We do not know
appropriate initial conditions for a periodic solution. In fact, only special initial conditions give
rise to periodic solutions so we will have to determine them as part of the solution.
Without loss of generality we can assume x(0) = 0. For a periodic solution x must be zero at
some time. Let this time be t = 0. Thus, let
x(0) = () =
0
+
1
+
2
2
+ ,
x(0) = 0.
(7.125)
Here we have used the fact that the periodic orbit will depend on , hence we must allow too as
well. As usual let
= ()t = (1 +
1
+
2
2
+ )t, (7.126)
giving the DE
2
()x
+x = (x
1
3
3
x
3
). (7.127)
O(1): The leading-order problem is
x
0
+x
0
= 0,
x
0
(0) =
0
,
x
0
(0) = 0,
_
_
_
x
0
=
0
cos , (7.128)
where
0
is undetermined. This is a reection of the fact that when = 0 we get circular orbits
(on the x- x plane) with arbitrary radius.
O(): At the next order the DE is
x
1
+x
1
= 2
1
0
cos +
_
1
4
3
0
0
_
sin
1
12
3
0
sin 3. (7.129)
There are two secular terms we must make equal to zero. Since
0
= 0 is uninteresting (it gives
the zero solution) setting the coecient of cos to zero gives
1
= 0, (7.130)
90
while setting the coecient of sin to zero gives
0
= 2. (7.131)
These two points are in fact on the same orbit: there are two locations where x = 0, one at x = 2
and at the other x = 2 to O(1). We can choose to start at either one. Taking
0
= 2, (7.132)
gives
x
1
+x
1
=
2
3
sin 3,
x
1
(0) =
1
,
x
1
(0) = 0,
(7.133)
the solution of which is
x
1
() =
1
12
sin 3
1
4
sin +
1
cos . (7.134)
O(
2
): At the next order, using
1
= 0, we have
x
2
+x
2
=
_
4
2
+
1
4
_
cos
1
2
cos 3 +
1
4
cos 5 + 2
1
sin
1
sin 3. (7.135)
To cancel the secular terms set
2
=
1
16
,
1
= 0,
(7.136)
giving
x
2
+x
2
=
1
2
cos 3 +
1
4
cos 5,
x
2
(0) =
2
,
x
2
(0) = 0.
(7.137)
which we will leave unsolved. The solution will involve an undetermined constant
2
the value of
which is obtained by the solvability condition for the O(
3
) problem.
So far we have
x(0) = 2 +O
F
(
2
),
= 1
1
16
2
+O
F
(
3
)
x = 2 cos +
_
1
12
sin 3
1
4
sin
_
+O
F
(
2
),
(7.138)
with = t.
We have found one periodic orbit and in fact it can be shown that the exact solution only has
one. Figure 7.4 compares numerical solutions of (7.123) using = 0.5 for x(0) = 1, 2, and 3. In
Figure 7.5 the solutions are shown on a phase plane and the solution starting at (x, x) = (2, 0) is
compared with a circle of radius 2 and with the periodic orbit (7.138) obtained using perturbation
theory. The orbit for x(0) = 2 appears closed in the gure after going around the origin over six
times (see Figure 7.4). This shows that perturbation theory gives an excellent solution for as
large as 0.5.
91
Figure 7.4: Numerical solutions of Rayleighs equation (7.123) for = 0.5.
Figure 7.5: (a) Numerical solutions of Rayleighs equation (7.123) for = 0.5. (b) Comparison of numerical solution
starting at (x, x) = (2, 0) with circle of radius 2. (c) Comparison of numerical solution starting at (x, x) = (2, 0) with
perturbation theory solution.
7.4.4 MSC: The Lighthill Technique
(Lighthill, 1949, Phil. Mag. 40)
Example 7.4.4 Find an Asymptotic Expansion of the solution to
(x +y)
dy
dx
+y = 1
y(1) = 2,
(7.139)
for 0 x 1, 0 < 1.
Recall that Regular Perturbation Theory gives
y
0
= 1 +
1
x
,
y
1
=
(3x + 1)(x 1)
2x
3
,
y
2
=
(3x + 1)(1 x
2
)
2x
5
,
(7.140)
92
As x 0 y
2
is more singular than y
1
which is more singular than y
0
, so the solution becomes
disordered.
Lighthill suggested looking for a solution of the form
y(x) = y(z) = Y
0
(z) +Y
1
(z) +
2
Y
2
(z) + , (7.141)
where the new strained coordinate z is introduced via
x = z +x
1
(z) +
2
x
2
(z) + . (7.142)
The functions x
n
(z) must be determined as part of the solution. They are chosen by invoking what
is now known as Lighthills Rule.
Denition 7.4.1 (Lighthills Rule) Choose x
n
(z) to get rid of singularities in the higher order
problems, i.e., Y
n
(z) should not be more singular than Y
n1
(z).
In terms of our new variables we have
(x +y)
dy
dx
+y = (x + y)
d y
dz
_
dx
dz
+ y, (7.143)
so the DE can be written as
(x + y)
d y
dz
+
dx
dz
y =
dx
dz
. (7.144)
Expanding gives
_
z +(x
1
(z) +Y
0
) +
2
(x
2
(z) +Y
1
) +
__
Y
0
+Y
1
+
_
+
_
1 +x
1
+
2
x
2
+
__
Y
0
+Y
1
+
_
=
_
1 +x
1
+
2
x
2
+
_
.
(7.145)
With the change of variables the boundary condition at x = 1 is applied at the value of z that
satises z + x
1
(z) + = 1, which is unknown until the x
j
(z) are found. Since the boundary
condition will be applied at z = 1 in the limit 0, let x = 1 correspond to
z = 1 +a
1
+a
2
2
+ . (7.146)
Assuming all the x
n
(z) have Taylor Series expansions about z = 1 (which is reasonable since the
only known singularity is at x = 0, i.e., near z = 0) we have
[a
1
+x
1
(1)] + [a
2
+x
1
(1)a
1
+x
2
(1)]
2
+ = 0. (7.147)
The boundary condition is
y(1 +a
1
+a
2
2
+ ) = 2, (7.148)
or
Y
o
(1) +
_
Y
1
(1) +Y
0
(1)a
1
_
+ = 2, (7.149)
assuming the Y
n
(z) also have Taylor Series expansions about z = 1.
93
O(1) : At leading order we have
z
dY
0
dz
+Y
0
= 1
Y
0
(1) = 2
(7.150)
The general solution of the DE is 1 +A/z. Applying the boundary condition gives
Y
0
= 1 +
1
z
. (7.151)
O() : The O() problem is
z
dY
1
dz
+Y
1
= (Y
0
(z) +x
1
(z))
dY
0
dz
+
dx
dz
(1 Y
0
)
=
_
1 +
1
z
+x
1
_
1
z
2
1
z
dx
1
dz
,
Y
1
(1) = a
1
Y
0
(1).
(7.152)
We now invoke Lighthills Criterion: chose x
1
(z) so that Y
1
(z) is not more singular that Y
0
(z).
This is easily done by choosing x
1
to make the right hand side of the DE for Y
1
equal to zero,
for then Y
1
(z) = B/z. Thus, x
1
(z) is chosen as a solution of the DE
x
1
z
x
1
=
1
z
+
1
z
2
, (7.153)
which has the general solution
x
1
(z) = 1
1
2z
+cz, (7.154)
where c is an arbitrary constant. We are free to choose a value for c, however it is convenient
to choose its value so that the location where boundary condition is applied is not moved
from z = 1, i.e., choose c so that a
1
= 0. Since a
1
= x
1
(1), we choose c so that x
1
(1) = 0
which gives c = 3/2. Thus
x
1
(z) = 1
1
2z
+
3
2
z. (7.155)
The boundary condition for Y
1
is now determined, namely Y
1
(1) = 0. This gives
Y
1
(z) = 0. (7.156)
So far we have
y(X) = 1 +
1
z
+O(
2
) +O
F
(
2
),
x = z +
_
3
2
z 1
1
2z
_
+O
F
(
2
).
(7.157)
as 0. If we ignore the O(
2
) terms we can solve for z(x), giving
z(x) =
x +
x
2
+ 2x + 4
2
+ 2
2 + 3
. (7.158)
94
We must take the positive sign so that x = z when = 0 (since x 0). Hence
y(x) =
x
_
x
_
2
+ 2
_
x + 1
_
+ 4. (7.159)
This is in fact the exact solution, however this is just luck. In general we will not nd the
exact solution.
Comment: Dierent choices of c are possible. For example, taking c = 0 gives
x
1
(z) = 1
1
2z
(7.160)
in which case
Y
1
(z) =
3
2z
. (7.161)
Now
y = 1 +
1
z
+
3
2z
+O
F
(
2
) as 0
x = z
_
1 +
1
2z
_
+O
F
(
2
) as 0
(7.162)
which is no longer the exact solution.
95
7.4.5 The Pritulo Technique
Lighthills technique involves solving a dierential equation for x(z). Here is an alternative pro-
cedure, introduced by Pritulo (1962, J. App. Math. Mech. 26) that avoids the introduction of a
second set of dierential equations to solve.
Basic Methodology:
1. First nd a Regular Perturbation Theory expansion:
y = y
0
(x) +y
1
(x) + . (7.163)
If the expansion is uniformly ordered youre done.
2. Next, strain the coordinates:
x = z +x
1
(z) + (7.164)
This is identical to Lighthills method.
3. Next, substitute (7.164) into (7.163) and do a Taylor Series expansion of the Y
i
s about x = z.
This gives
y = Y
0
(z) +
_
Y
0
(z)x
1
(z) +Y
1
(z)
_
+
2
_
1
2
Y
0
(z)x
2
1
(z) +Y
0
(z)x
2
(z) +Y
1
(z)x
1
(z) +Y
2
(z)
_
+
(7.165)
4. Choose the x
n
(z) to eliminate more singular terms according to Lighthills Rule.
Example 7.4.5 (Lighthills example revisited) We will demonstrate this method be reconsid-
ering Lighthills example
(x +y)y
+y = 1,
y(1) = 2,
(7.166)
for 0 x 1 and 0 < 1.
Regular Perturbation Theory gives
y 1 +
1
x
+
(x 1)(3x + 1)
2x
3
+
(1 x)(3x + 1)
2x
5
2
+ , (7.167)
as 0. Set x = z +x
1
(z) +
2
x
2
(z) + . In terms of z we get
1
x
=
1
z
x
1
(z)
z
2
+O(
2
)
1
x
3
=
1
z
3
3x
1
(z)
z
4
+O(
2
)
(7.168)
so
y = 1 +
1
z
_
x
1
(z)
z
2
+
(z 1)(3z + 1)
2z
3
_
. .
Y
1
(z)
+O
F
(
2
) (7.169)
96
Invoking Lighthills rule, we now choose x
1
(X) to eliminate the part of Y
1
(z) which is more singular
than Y
0
(z). That is, we want to eliminate the 1/z
3
singularity. This can be done by choosing x
1
(z)
so that Y
0
(z) = 0. This is of course only one of an innity of choices. Doing this gives
x
1
(X) =
(z 1)(3z + 1)
2z
. (7.170)
This gives a solution which is identical to x
1
(X) using Lighthills method, so as before we have
obtained the exact solution.
7.4.6 Comparison of Lighthill and Pritulo Techniques
Lighthills method: Set y = Y
0
(z) + where x = z + x
1
(z) + . Solve a DE for x
1
(z) and
then solve a simplied DE for Y
1
(z).
Pritulos method: No DE to solve for the x
n
(z), however in the RPT expansion y = y
0
(x) +
y
1
(x)+ we need to rst nd the y
n
(x) and then expand these functions in a Taylor Series about
x = z.
Pritulos Technique takes an AE
y y
0
(x) +y
1
(x) +y
2
(x)
2
+ , (7.171)
which is not uniformly ordered and gives an asymptotic expansion
y y
0
(z) + y
1
(z) + y
2
(z)
2
+ , (7.172)
which is uniformly ordered.
How is this possible? It is only possible because the two solutions are only equivalent as an
innite series. The truncated series are not equivalent. For example, assuming y
0
(x) and y
1
(x) can
be expanded in a Taylor Series about z,
y
0
(x) +y
1
(x) = y
0
(z) + [y
1
(z) +y
0
(z)x
1
(z)]
+ [y
0
(z)
x
2
1
(z)
2
+y
0
(z)x
2
(z) +y
1
(z)x
1
(z)]
2
+ ,
= y
0
(z) + y
1
(z) + y
2
(z)
2
+ .
(7.173)
i.e.,
y
0
(x) +y
1
(x) = y
0
(z) + y
1
(z) + y
2
(z)
2
+
y
0
(z) + y
1
(z)
(7.174)
as 0. The dierence between y
0
(x) +y
1
(x) and y
0
(x) + y
1
(z) is O(
2
) and the dierence may
be singular in x or z. In particular, the dierence can be larger than the retained terms as x goes
to the singular point. For xed x, as 0 the dierence is negligible.
The Pritulo technique amounts to a rearrangement of the terms in an RPT series. We will
return to this when we compare a couple of nonlinear wave equations, derived with asymptotic
methods.
97
98
Chapter 8
Matched Asymptotic Expansions
In this chapter we study the method of matched asymptotic expansions. This is a technique used for
problems where the solution behaves on very dierent scales in two overlapping regions. This type
of behaviour is typical of ODEs for which the highest derivative is multiplied by a small parameter,
in which case the solution will in general have a thin boundary layer adjacent to a boundary in
which the solution varies very rapidly. Outside of the boundary layer the solution typically varies
on a much longer length scale. If the dependent variable is time the boundary layer is sometimes
called an initial layer. In other problems the thin layer of rapid variation can occur in the interior
of the domain (interior or transition layers).
Example 8.0.1 Consider the problem
y
(x) +y
(x) = a,
y(0) = 0,
y(1) = 1,
(8.1)
where 0 < a < 1 and 0 < 1.
Trying Regular Perturbation Theory we let
y = y
0
+y
1
(x) +
2
y
2
(x) + , (8.2)
resulting in the O(1) problem
y
0
= a (8.3)
which has general solution y
0
(x) = ax + b. Unfortunately we have two boundary conditions and
unless a = 1 they cant both be satised.
The exact solution of the full problem is
y = ax + (1 a)
1 e
x/
1 e
1/
. (8.4)
The solution for a = 0.5 and = 0.02 is shown in Figure 8.1 along with the straight line y
0
(x) =
ax + 1 a, which is the RPT solution obtained by applying the boundary condition at x = 1. We
can see that in a thin region near x = 0 the solution varies very rapidly while outside this thin
layer the solution is approximated very accurately by the RPT solution y
0
(x) = ax + 1 a.
The region in which y
0
provides a good approximation of the solution is called the outer region.
It is the region in which the presence of the small parameter in the dierential equation correctly
99
Figure 8.1: Solution of 8.1 (solid curve) for a = 0.5 and = 0.02, compared with the outer solution y = ax +1 a
(dots).
indicates which term can be neglected. In the thin layer near the origin the solution varies very
rapidly. Here y
=
2
a. (8.6)
In the outer region the dominant balance in the dierential equation was between y
and a. In
the inner region the dominant balance must include y
. If it is between y
and
2
a we should
take =
1/2
. Then y = A + B where A and B are constants. Unless A = 0 y
is O(1) and
/ y
= O(
1/2
is larger than the retained terms. If A = 0 the boundary condition gives B = 0
and hence y = 0. This cant be correct. Thus the dominant balance must be between y
and
.
Hence we should take
() = , (8.7)
giving
y
+ y
= a. (8.8)
We now use RPT. Setting y = y
0
+ y
1
+
2
y
2
+ , at leading order we have
y
0
+y
0
= 0, (8.9)
which has the general solution
y
0
= Ae
+B. (8.10)
100
We know from the exact solution that the boundary layer is in a thin neighbourhood near = 0.
Thus, we should apply the boundary condition at x = 0, since the other boundary point x = 1 is
not in the boundary layer. For many problems it may not be obvious which boundary condition to
apply, although this can often be determined from a consideration of the physical problem being
considered. For now we will use our knowledge of the location of the boundary layer. If we assumed
that the boundary layer was near x = 1 we would not be able to match the inner and outer solutions.
We will discuss this later.
Applying the boundary condition y
0
(0) = 0 gives
y
0
() = A(1 e
). (8.11)
We have the following approximations of the solution:
(i) In the inner region: y y
in
= y
0
= a(x 1) + 1.
(ii) In the outer region: y y
out
= y
0
() = A(1 e
).
The solutions involve an unknown constant A. We now need to match the two solutions in
order to determine its value. To do this we need to nd a matching region where both the inner
and outer solutions are valid. That is we need to nd > > 0 such that
y y
0
(x) as 0 for some region x >
,
y y
0
() as 0 for some region x <
.
(8.12)
Note that
.
Proving the existence of a matching region is often the most dicult part of matched asymptotic
expansions. In general we want x 0 and in the matching region as 0 and we will
often just assume this.
For our particular problem we can use x
y =
2
a. (8.16)
101
As before we should take = leading to the O(1) solution y
0
= Ae
1) + 1. (8.17)
Now we cant match the solutions because as , which corresponds to x moving away from
x = 1 into the interior of the domain 0 x 1, the inner solution blows up while the outer solution
y
0
a as x 1.
Note also that < 0 changes the location of the boundary layer. If < 0 then assuming the
boundary layer is near x = 0 leads again to = x/. This time as 0 with x in the matching
region, now e
+xy = e
x
,
y(1) = 1/e.
(8.18)
In the outer region, where can set = 0, y y
out
where
xy
out
+xy
out
= e
x
, (8.19)
which has the general solution
y
out
= Ae
x
+ ln xe
x
. (8.20)
Now we need to determine whether or not x = 1 is in the outer region. If so we can use the initial
condition y(1) = 1/e to determine A. From the original ODE
y
=
e
x
xy
x y
(8.21)
so, at x = 1
y
(1) =
e
1
1/e
x /e
= 0. (8.22)
In particular, y
is comparable to e
x
. We
can use y
out
as an approximation and see when y
out
y
out
is comparable to e
x
. Thus, we need to
determine when
y
out
y
out
= (1 + ln x)e
x
_
1
x
e
x
(1 + ln x)e
x
_
e
x
. (8.24)
If x = O(1) the l.h.s is O() while the r.h.s. is O(1) so we clearly need x 1. If x 1 we have
1/x | ln x| and e
x
1, so we need to nd x such that
ln x
x
1. (8.25)
102
This gives x = O( ln ) (exercise).
In the inner region x is small so e
x
1 and xy xy
in
= 1. (8.26)
More systematically, let
=
x
()
(8.27)
where 0 as 0. The ODE gives
_
y
_
1
+ y = e
. (8.28)
The right hand side is approximately 1. Since we need to retain the y y
1, (8.29)
as above. To solve this write it as
d
d y
= y. (8.30)
We now have a linear equation for ( y) which can be solved. This results in
= y + 1 +ae
y
, (8.31)
or
x = (1 +y
in
) +ce
y
i
n
. (8.32)
This implicitely gives y
in
(x) if we can determine the value of c. We do this by matching the inner
and outer solutions.
Claim:
1.
1/2
| ln |
as 0 so x = O(
1/2
) is in the outer region.
2. The inner region is given by x 1 so x = O(
1/2
) is in the inner region as well.
When x 1 the outer solution is
y
out
1 + ln x, (8.33)
while for x = O(
1/2
) we have x ce
y
in
so y
in
ln x ln c. Matching gives c = e
1
so in the inner
region x (1 +y
in
) +e
y
in
1
, which can be solved implicitly for y
in
(0).
103
104
Chapter 9
Asymptotics used to derive model
equations: derivation of the
Korteweg-de Vries equation
for internal waves
9.1 Introduction
So far we have concentrated on using perturbation and asymptotic methods to nd approximate
solutions of dicult problems. An equally important use of asymptotic methods is to derive approx-
imate mathematical models, the solution of which provides a good approximation to some of the
solutions of the original set of equations. The word some in the preceding sentence is important as
the asymptotic procedure used to derive the simplied set of equations is based on the introduction
of small parameters that come from scaling the original variables. The scaling is determined by the
type of phenomena one wishes to investigate and it identies terms which are not important in a
rst approximation. The scaling choice leads to approximate asymptotic models that are simpler
and which isolate certain types of behaviour in the original system.
The Korteweg-de Vries, or KdV, equation was originally derived in 1895 by Diederik Johannes
Korteweg and his PhD student Gustav de Vries in the context of surface water waves. The KdV
equation is restricted to uni-directional wave propagation (this is one way it simplies your problem:
it eliminates waves propagating in other directions). It has been derived in many dierent physical
contexts, such as waves in beams or rods, nonlinear electric lines, blood pressure waves, large scale
planetary Rossby waves in the atmosphere and oceans, nonlinear spring-mass systems and many
others. We will consider the case of horizontally propagating internal gravity waves as illustrated
in Figure 9.1.
Consider an ideal inviscid, incompressible, density stratied uid in a non-rotating reference
frame. The equations of motion are
Du
Dt
=
p g
k,
D
Dt
= 0,
u = 0.
(9.1)
Here (x, t) is the uid density, u = (u(x, t), v(x, t), w(x, t)) is the uid velocity, p is the pressure
105
Figure 9.1: Internal gravity waves propagating horizontally in a density stratied uid. Shown are density contour
plots at dierent times.
eld, g is the gravitational acceleration and
k is the unit vector in the positive z direction with z
pointing upward. The rst of these equations is called the momentum equation. It is a statement of
Newtons second law applied to a uid particle: mass per unit volume () times the uid particles
acceleration (Du/Dt) is equal to the net force per unit volume acting on the uid particle. The
forces include those associated with pressure gradients (
k). The
third equation is the incompressibility condition. It says that the volume of a parcel of uid is
constant in time. The intermediate equation states that the density of a uid particle is constant
in time, which is a consequence of conservation of mass combined with incompressibility and the
fact that we are neglecting diusion processes (e.g., diusion of heat and, in the ocean, salts).
The material derivative D/Dt = /t +u
gives the rate of change moving with a particle.
To understand how it arises, consider a particle at position x(t) at time t which has velocity
u
L
(t) = u(x(t), t). Its acceleration is
a
L
(t) =
du
L
dt
(t) =
_
dx
dt
_
u(x(t), t) +
u
t
,
=
Du
Dt
(x(t), t),
(9.2)
To simplify matters we will consider two-dimensional motion in a vertical plane. Taking u =
(u, w) and x = (x, z) the equations of motion in component form are
Du
Dt
= p
x
,
Dw
Dt
= p
z
g,
D
Dt
= 0,
u
x
+w
z
= 0.
(9.3)
There is one more simplication we will make before deriving the KdV equation. In the oceans
and lakes the density varies only slightly from its mean value
o
. If we set =
o
+
v
,
v
/
o
106
has a typical value of about 0.02 in the oceans and is much smaller in lakes. Hence Du/Dt =
(
o
+
v
)Du/Dt
o
Du/Dt. No matter what the solution of the equations is,
v
Du/Dt will be at
most a few percent of
o
Du/Dt. This leads to the nal set of equations
o
Du
Dt
= p
x
, (9.4)
o
Dw
Dt
= p
z
g, (9.5)
D
Dt
= 0, (9.6)
u
x
+w
z
= 0. (9.7)
9.1.1 Streamfunction Formulation
From (9.7) there exists a streamfunction such that (u, w) = (
z
,
x
). The curl of the momentum
equation then gives the vorticity equation
2
g
x
o
= J(,
2
), (9.8)
where
2
= u
z
w
x
is the vorticity and the Jacobian operator J is dened by J(A, B) =
A
x
B
z
A
z
B
x
. Note that J(, A) = (u, w, )
A, so ()/t J(, ) = D()/Dt.
The density equation can be written as
t
= J(, ), (9.9)
We now we split the velocity and density elds into a background, undisturbed, state plus
a perturbation. We assume the undisturbed state to be a state of rest (u = 0) with a stable
stratication = (z). Thus, letting = (z) +
t
d
dz
x
= J(,
). (9.10)
Dening
b =
g
o
, (9.11)
we can write the vorticity and density equations as
2
b
x
= J(,
2
), (9.12)
b
t
+N
2
(z)
x
= J(, b), (9.13)
where
N
2
(z) =
g
o
d
dz
0. (9.14)
N is called the buoyancy frequency. For a stable stratication, with a non-increasing function,
N is real. It gives the frequency of oscillation of a uid parcel that is displaced vertically an
innitesimally small distance (i.e., in linear theory).
107
9.1.2 Boundary conditions
We will need some boundary conditions. We assume the uid is bounded by rigid horizontal
boundaries at z = H and 0. The latter condition eliminated the complication of surface waves.
The vertical velocity w =
x
at the boundaries must be zero, hence is constant along z = H, 0.
Now
_
0
H
udz =
_
0
H
z
dz = (x, 0, t) (x, H, t) which is independent of x as is independent
of x along z = H, 0 . Assuming our disturbance is of nite extent, as x u 0 from which we
can conclude that (x, 0, t) = (x, H, t). Since physically only derivatives of mean something,
we can take
= 0 on z = H, 0. (9.15)
For the density eld, since uid at the boundaries stays there (w = 0), the density perturbation at
the boundaries will be zero. Hence
b = 0 on z = H, 0. (9.16)
9.2 Nondimensionalization and introduction of two small param-
eters
We will now work with equations (9.12)(9.14). To scale them we need to make some decisions
regarding the type of solutions we are interested in. There are many observations of internal waves
propagating in the horizontal direction (see Figure 9.1). They often have the property that their
horizontal length scale is long compared with the water depth and their wave amplitude is not
too large. We also consider uni-directional propagating wave, i.e., all waves are propagating in the
positive x direction. This will be the phenomena we focus on.
Nondimensionalization:
Vertical length scale: water depth H.
Horizontal length scale: typical wavelength L
Buoyancy frequency: typical value N
o
.
Fact: wave propagation speeds are determined by N(z) and H. Thus, c
o
= N
o
H is the
appropriate scaling for the wave propagation speed.
Time scale: T = L/c
o
= L/(N
o
H) is the time to travel distance L at phase speed c
o
.
Thus, we set
(x, z, t) = (L x, H z,
L
N
o
H
t),
N(z) = N
o
N( z),
(9.17)
Horizontal velocity scale: We assume a small amplitude perturbation to the undisturbed
state. By small, we mean the wave induced horizontal velocity u is small compared with the
horizontal propagation speed. Hence we set
u = c
o
u (9.18)
where is a small nondimensional parameter measuring the wave amplitude.
108
Streamfunction: Setting =
and using u =
z
, we nd
u = c
o
u =
z
=
H
z
, (9.19)
hence we should choose
= c
o
H, (9.20)
so that u =
z
.
Vertical velocity scale: From u
x
+w
z
= 0 the vertical velocity should be scaled by
w =
H
L
c
o
w. (9.21)
It is easily veried that w =
x
.
Scaling for b: Since b is proportional to the density perturbation which is small, let b = B
b.
In (9.13) the two linear terms on the left hand side should dominate (this is a choice
quadratic terms should be negligible at leading order since the perturbation is small), thus,
we need
B
T
= N
2
o
L
, (9.22)
which gives
b = N
2
o
L
c
o
c
o
H
H
b = N
2
o
H
b. (9.23)
With these scalings, the governing equations (9.12)(9.13) become, after dropping the tildes,
zz
b
x
= J(,
zz
)
xx
+J(,
xx
), (9.24)
t
b +N
2
(z)
x
= J(, b), (9.25)
where = (H/L)
2
. In Figure 9.1 0.04. In the following we will assume that the horizontal
length scale is large compared with the water depth, that is, we assume that H L, i.e., 1.
Thus, we have two small parameters, and .
9.3 Asymptotic expansion
The KdV equation is an evolution for small amplitude, long waves, i.e., waves which are long com-
pared to the water depth. Since there are two small parameters, and in the nondimensionalized
equations, we expand and b in powers of both and :
(0)
+
(1,0)
+
(0,1)
+ +
i
(i,j)
+ ,
b b
(0)
+b
(1,0)
+b
(0,1)
+ +
i
j
b
(i,j)
+ ,
(9.26)
as , 0.
109
9.3.1 The O(1) problem
At leading order we have
(0)
zz
b
(0)
x
= 0, (9.27)
t
b
(0)
+N
2
(z)
(0)
x
= 0, (9.28)
from which we get
2
t
2
(0)
zz
+N
2
(z)
(0)
xx
= 0. (9.29)
We now look for separable solutions of the form
(0)
= B(x, t)(z). (9.30)
This is motivated by our interest in waves which are propagating in the x direction along the wave
guide bounded by the surface and the bottom. The function (z) will give the vertical structure
of the waves while B(x, t) will satisfy a wave equation describing the propagation and evolution of
the wave in the x direction.
Substituting (9.30) into (9.29) results in
B
tt
+N
2
B
xx
= 0,
or
B
tt
B
xx
= N
2
zz
. (9.31)
This says that a function of x and t is equal to a function of z. Hence both sides must be equal to
a constant, say c
2
, giving
B
tt
c
2
B
xx
= 0, (9.32)
+
N
2
c
2
= 0. (9.33)
The boundary conditions (see (9.15)) are
(1) = (0) = 0, (9.34)
since the nondimensional water depth is 1. Equations (9.33)(9.34) represent an eigenvalue problem
for with eigenvalue c. Since N
2
0 there are an innite number of discrete solutions (
i
, c
i
)
with c
2
i
> 0 which can be ordered such that c
1
> c
2
> c
3
> > 0 with 0 as a cluster point.
Solutions for dierent values of c correspond to dierent wave modes. From (9.32) we see that c is
the propagation speed of the waves. At this order all waves propagate with speed c and waves can
propagate in either direction. We will now restrict attention to rightward propagating waves only,
so that B
t
+cB
x
= 0. From (9.28)
b
(0)
t
= N
2
B
x
=
N
2
c
B
t
, (9.35)
hence
b
(0)
=
B
c
N
2
(z)(z) =
B
c
. (9.36)
110
9.3.2 The O() problem
At O() equations (9.24)(9.25) give
(1,0)
zz
b
(1,0)
x
= J(
(0)
,
(0)
zz
), (9.37)
t
b
(1,0)
+N
2
(z)
(1,0)
x
= J(
(0)
, b
(0)
). (9.38)
From the solution to the O(1) problem
J(
(0)
,
(0)
zz
) = BB
x
_
_
, (9.39)
J(
(0)
, b
(0)
) = cBB
x
_
_
. (9.40)
We again look for separable solutions, i.e., we assume
(1,0)
and b
(1,0)
can both be expressed as
a function of x and t multiplying a function of z. The form of the nonlinear forcing terms suggests
the ansatz
(1,0)
= B
2
(1,0)
,
b
(1,0)
= B
2
D
(1,0)
.
(9.41)
Substituting into (9.37)(9.38) gives
2BB
t
(1,0)
zz
2BB
x
D
(1,0)
= BB
x
_
_
,
2BB
t
D
(1,0)
+ 2BB
x
N
2
(1,0)
= cBB
x
_
_
,
(9.42)
Using B
t
= cB
x
:
c
(1,0)
zz
D
(1,0)
=
1
2
_
_
,
cD
(1,0)
+N
2
(1,0)
=
c
2
_
_
,
(9.43)
from which, after eliminating D
(1,0)
we have
(1,0)
zz
+
N
2
c
2
(1,0)
=
1
c
_
_
. (9.44)
The boundary conditions are again
(1,0)
= 0 at z = 1, 0. This is an inhomogeneous version of
(9.33), the ODE we had to solve in the O(1) problem.
9.3.3 The problem
Now suppose we multiply the left handside of (9.44) by and integrate from 1 to 0. We have
_
0
1
(1,0)
zz
+
N
2
c
2
(1,0)
_
dz =
_
0
1
_
(1,0)
zz
+
N
2
c
2
(1,0)
) dz,
=
_
0
1
_
(1,0)
zz
zz
(1,0)
) dz,
=
_
0
1
_
(1,0)
z
(1,0)
z
) dz,
= 0
(9.45)
111
where we have used (9.33) in the secon step and have integrated by parts in the last step using the
fact that and
(1,0)
are zero at the boundaries.
If we multiply the right-hand side of (9.44) by and integrate from 1 to 0 we obtain
_
0
1
1
c
(
)dz =
3
2c
_
0
1
3
dz, (9.46)
after a couple of integrations by parts. In general, this is not going to be zero (it is zero if N
is symmetric about the mid-depth). Thus, in general, there is no solution to (9.44). This is an
example of resonant forcing and to proceed we need to nd a way to eliminate the resonant part
of the forcing.
9.3.4 The x
To x this problem we have to reconsider the O(1) problem. At leading order we found that
B(x, t) satises the linear long wave equation B
t
+cB
x
= 0. This says that a wave with any shape
propagates without changing shape. In reality the wave will change shape due to nonlinear and
dispersive eects. This can be seen in Figure 9.1. Thus, we need to modify the wave equation to
introduce small corrections via
B
t
cB
x
R(x, t) Q(x, t) +h.o.t., as , 0, (9.47)
Consider the O(1) problem again. The governing equations are
(0)
zz
b
(0)
x
= 0, (9.48)
b
(0)
t
+N
2
(z)
(0)
x
= 0. (9.49)
Using
(0)
= B(x, t) and b
(0)
=
N
2
c
B and using (9.47) these equations give,
(c
+
N
2
c
)B
x
R
+ = 0, (9.50)
(N
2
+N
2
)B
x
N
2
c
R
N
2
c
Q + = 0. (9.51)
We have not changed so it still satises (9.33). Thus, the leading-order terms are zero. The
higher-order terms become part of the higher-order problems. In particular, the O() term belongs
in the O() problem and the O() term belongs in the O() problem.
9.3.5 The O() problem revisited
Incorporating the left-over O() terms from the O(1) problem into the O() problem gives
R
+
t
(1,0)
zz
b
(1,0)
x
= BB
x
_
_
, (9.52)
N
2
c
R +
b
(1,0)
t
+N
2
(z)
(1,0)
x
= cBB
x
_
_
. (9.53)
112
As before, we use the ansatz (9.41) and clearly we should take R BB
x
. Setting R = BB
x
we have
c
(1,0)
zz
D
(1,0)
=
2
+
1
2
_
_
, (9.54)
cD
(1,0)
+N
2
(z)
(1,0)
=
2c
N
2
c
2
_
_
, (9.55)
from which, after using N
2
= c
2
(1,0)
zz
+
N
2
c
2
(1,0)
=
1
c
_
_
. (9.56)
We still have the solvability condition that the integral of the right-hand side multiplied by should
be zero. This is used to determine the value of . The result is
=
3
2
_
0
1
3
dz
_
0
1
2
dz
. (9.57)
A similar procedure for the O() problem yields Q = B
xxx
where
=
c
2
_
0
1
2
dz
_
0
1
2
dz
, (9.58)
which is always positive. The nal equation for B, to O(, ), is
B
t
+cB
x
+BB
x
+B
xxx
= 0, (9.59)
which is the KdV equation. Attempts to solve higher-order problems lead to the introduction of
higher-order corrections to the evolution equation for B.
113
114
Appendix A: USEFULL FORMULAE
Trigonometric Identities:
sin
3
(t) =
3
4
sin(t)
1
4
sin(3t),
cos
3
(t) =
3
4
cos(t) +
1
4
cos(3t),
sin
5
(t) =
5
8
sin(t)
5
16
sin(3t) +
1
16
sin(5t),
cos
5
(t) =
5
8
cos(t) +
5
16
cos(3t) +
1
16
cos(5t),
(Acos t +Bsin t)
3
=
3
4
A(A
2
+B
2
) cos t +
3
4
B(A
2
+B
2
) sin t
+
1
4
A(A
2
3B
2
) cos 3t
1
4
B(B
2
3A
2
) sin 3t
sin(nt) cos(mt) =
sin((n +m)t) + sin((n m)t)
2
,
sin(nt) sin(mt) =
cos((n m)t) cos((n +m)t)
2
,
cos(nt) cos(mt) =
cos((n +m)t) + cos((n m)t)
2
,
Solutions of homogeneous ODEs for y(x):
y
+
a
x
y
+
b
x
2
y = 0 try y x
n
,
y
=
1
4
y(4 y)
dx
dy
= (
1
y
+
1
4 y
)
115
Particular solutions of common forced ODEs:
y
+
2
y = sin t y
p
=
1
2
t cos t,
y
+
2
y = cos t y
p
=
1
2
t sin t,
y
+
2
y = sin t y
p
=
1
2
sin t for = ,
y
+
2
y = cos t y
p
=
1
2
cos t for = ,
y
y = e
t
y
p
= te
t
,
y
= 1 y
p
=
t
,
y
= e
t
y
p
=
t
e
t
2
e
t
Taylor Series:
tanh(x) = x
1
3
x
3
+
2
15
x
5
+ , (9.60)
Expansions:
(a +b)
3
= a
3
+ 3a
2
b + 3ab
2
+b
3
,
(a +b)
4
= a
4
+ 4a
3
b + 6a
2
b
2
+ 4ab
3
+b
4
,
(a +b)
5
= a
5
+ 5a
4
b + 10a
3
b
2
+ 10a
2
b
3
+ 5ab
4
+b
5
,
(a
o
+a
1
+a
2
2
+ )
2
= a
2
o
+ 2a
0
a
1
+ (2a
o
a
2
+a
2
1
)
2
+ ,
(a
o
+a
1
+a
2
2
+ )
3
= a
3
o
+ 3a
2
0
a
1
+ (3a
2
o
a
2
+ 3a
o
a
2
1
)
2
+ ,
(a
o
+a
1
+a
2
2
+ )
4
= a
4
o
+ 4a
3
0
a
1
+ (4a
3
o
a
2
+ 6a
2
o
a
2
1
)
2
+ ,
(a
o
+a
1
+a
2
2
+ )
5
= a
5
o
+ 5a
4
0
a
1
+ (5a
4
o
a
2
+ 10a
3
o
a
2
1
)
2
+ .
Methods:
Lighthill: y = Y (X) is replaced with
x = X +x
1
(X) + .
Pritulo: y = y
o
(x) +y
1
(x) + is replaced by
Y
o
(X) +Y
1
(X) + with x = X +x
1
(X) + .
MSC and Poincare-Linstedt: = ()t
116
Solutions to Selected Problems
Problems from chapter 2
1(a). Have a second order polynomial, hence two roots to nd. Setting = 0 gives two distinct
roots 6 and 1 hence expand in powers of . Get
x
(1)
= 6
3
7
12
7
3
2
+O(
3
),
x
(2)
= 6
4
7
+
12
7
3
2
+O(
3
).
1(c). Polynomial of degree three, hence need to nd three roots. Setting = 0 gives a double root
at x
0
= 1 and a single root x
0
= 2. Near the single root expand in powers of to nd
x
(1)
= 2 + /9 + (2/243)
2
+ O(
3
). Near the double root expand in powers of
1/2
to get
x
(2,3
) = 1 i
1/2
/
3 +/18 +O(
3/2
).
1(e). Need to nd three roots. Setting = 0 gives x
0
= 1 as a double root. To nd the two
roots near x
0
= 1 expand in powers of
1/2
. Find x
1/2
= 1
1/2
3/2 + O(
3/2
). For
the third root dominant balance is between x
3
and x
2
so x
3
x
2
or x 1/. Thus set
x = 1/ +x
1
+x
2
+x
3
2
+ . Fine x
(3)
= 1/ + 2 + 3 +O(
2
).
1(g). Need to nd four roots. Setting = 0 give a quadratic equation with two distinct roots. For
these expand in powers of giving x
(
1) = 1+2+18
2
+O(
3
) and x
(
2) = 224+488
2
+O(
3
).
For the other two roots the dominant balance is between x
4
and x
2
which gives x i
1/2
.
Let =
1/2
and y = x = y
0
+ y
1
+y
2
2
+ . Get y
4
+ y
3
+ y
2
3y + 2
2
= 0. The
leading order problem gives y
0
= i and y
0
= 0 as a double root. Only rst two of interest.
Since i are distinct single roots expand in powers of . Find y = i 2 3i
2
+O(
3
) or
x
3,4
= i/
1/2
2 3i
1/2
+O().
117