You are on page 1of 36

MA4547 ASYMPTOTIC ANALYSIS

MIN HUANG

MA4547 ASYMPTOTIC ANALYSIS

Contents
1. Introduction
1.1. What this course is about
1.2. What this course is not about
1.3. Structure of the course
1.4. References
2. Basic concepts and notations
2.1. Order notation
2.2. Asymptotic expansions/series
3. Asymptotic techniques for calculating integrals
3.1. Integration by parts
3.2. Watsons lemma
3.3. Laplaces method
3.4. Method of stationary phase
3.5. *Method of steepest descent
4. Asymptotic solutions to algebraic equations
4.1. Inverse of analytic functions
4.2. Asymptotic solutions for large x
5. Asymptotics of ordinary differential equaitons
5.1. The method of iteration
5.2. Perturbation expansions
5.3. The WKB method

3
3
4
4
4
5
5
6
11
11
12
13
17
19
22
22
25
28
28
29
32

MA4547 ASYMPTOTIC ANALYSIS

1. Introduction
In standard college level mathematics courses we learn various techniques of calculating integrals and solving differenital equations. However, these techniques only
apply to specific types of simple integrals or equations which occur in textbooks
much more often than in practice. In applications such as sciences and engineering,
numerical methods are often preferred. However, numerical methods have a fundamental limitation, namely it cannot provide any concrete formula, and is therefore
insufficient when unknown parameters are present and when the variables tend to
infinity. For instance, the equation

(1.1)

x sin x = 1

cannot be solved explicitly. A simple argument with the intermediate value Theorem shows that there are infinitely many solutions, as illustrated by the Figure
below. Although it is simple to find a few of them using numerical methods, it is
impossible to find all of them in this way.

Figure 1.1. plot of x sin x 1


A natrual question to ask is, then, is there a way which allows us to find the
approximate values of all these solutions?
The answer is yes. The method is known as Asymptotic Analysis. It is
a method of describing limiting behaviors of integrals and solutions to algebraic
equations, differential equations, difference equations, etc.
For instance, using asymptotic anaysis one can find the approximate locations
k
of solutions of (1.1) to be xk k + (1)
k (k > 1). These are not exact solutions
but they are very accurate, as shown in the figure below.
For practical purposes, approximate solutions are often good enough as long as
there is sufficient accuracy.
1.1. What this course is about. The focus of this course is asymptotic techniques for calculating integrals and solving equations. It will involve a lot of detailed
symbolic calculations. Loosely speaking we will calculate the limiting behavior of
certain funcitons. It will be important to do plenty of exercises.

MA4547 ASYMPTOTIC ANALYSIS

Figure 1.2. plot of x sin x 1 and the approximate roots xk


1.2. What this course is not about. This course is not about proofs of abstract
theorems. Although all the results in this course can be proved rigorously, the proofs
are usually lengthy and may require advanced knowledge more suitable for graduate
students. This course is not about computer programming either. Although there
are mathematical programs such as Mathematica and Maple which can perform
symbolic calculations, they cannot give asymptotic formulas except in special cases.
A thorough understanding of asymptotic analysis is necessary, even though these
programs are sometimes used to help with detailed and messy calculations.
1.3. Structure of the course. In Section 2 we introduce basic concepts and notations of aymptotic analysis. These are intuitive and straightforward.
Section 3 is about calculating asymptotic formulas for integrals. We will focus
on four techniques: integration by parts, Watsons lemma, Laplaces method, and
method of stationary phase. We will also briefly mention the method of steepest
descent.
In Section 4 we introduce two methods for finding approximate solutions to algebraic equations, namely the methods of undetermined coefficients and the method
of iteration.
In Section 5 we discuss asymptotic analysis for ordinary differential equations,
and we will focus on two methods: the methods of perturbation expansion and the
WKB method. We will also briefly mention singularly perturbed equations.
1.4. References. There are some good comprehensive textbooks (see below). However, they are mainly geared towards graduate students.
(1) Carl M. Bender, Steven A. Orszag, Advanced Mathematical Methods for
Scientists and Engineers: Asymptotic Methods and Perturbation Theory
(1999).
(2) Peter D. Miller, Applied Asymptotic Analysis (2006).

MA4547 ASYMPTOTIC ANALYSIS

2. Basic concepts and notations


2.1. Order notation. We are already familiar with the standard notations of limits (lim) and definitions with  language, etc. These are of course correct, but
they are too cumbersome for the purpose of calculations. We prefer notations that
are convenient and intuitive, and what we commonly use are three notations O, o,
describing bounded by, of a smaller order than, and equivalent to respectively.
Suppose f (z) and g(z) are functions of the complex variable z defined on some
domain D in the complex plane, and possess limits as z z0 in D (here z0 is
allowed to be infinity). We have the following basic notations:
f (z) = O(g(z))

(z z0 )

This means |f (z)| is less than a constant times |g(z)| if z is close to z0 . In other
words, |f (z)/g(z)| is bounded if z is close to z0 .
Similarly we have
f (z) = o(g(z))

(z z0 )

This means
lim

f (z)
=0
g(z)

lim

f (z)
=1
g(z)

zz0

We also have the notation


This means

zz0

These notations are related to each other. For instance, f (z) = o(g(z)) implies
f (z) = O(g(z)). f (z) g(z) is the same as f (z) g(z) = o(g(z)).
In particular f (z) = O(1) means f is bounded for z close to z0 . f (z) = o(1)
means f tends to 0 as z approaches z0 .
Example 1. If f (z) = 3z 2 + z + 2, then we can write f (z) = O(z 2 ) as z +;
f (z) = o(z 3 ) as z +; f (z) 3z 2 as z +; f (z) 2 as z 0; f (z) 2 + z
as z 0, etc.
One can formally treat O(g) as a function which is bounded by a constant times
|g| and o(g) as a function which becomes arbitrarily smaller than |g| in the limit.
In particular one can formally treat O(1) as a function which is bounded and
o(1) as a function which tends to zero. Then one may add, subtract, multiply, or
divide these O and o notations carefully.
Example 2. We have the following facts:
O(f ) O(g) = O(|f | + |g|)
O(f )O(g) = O(f g)
O(f )o(g) = o(f g)
O(f )
= O(f )
1 + o(g)
f g if and only if f = g(1 + o(1))
If f1 g and f2 g then f1 f2 = o(g)

MA4547 ASYMPTOTIC ANALYSIS

A simple but important fact to remember is exponential functions increase much


faster than powers and powers increase much faster than logs. In other words
ln x = o(xn ) and xn = o(ex ) as x + for all n > 0.
2.2. Asymptotic expansions/series. We want to describe the limiting behavior
of complicated functions using simple functions as approximations. Obviously we
often need to use a finite or infinite sum of simple functions. One such example
that we are already familiar with is Taylor series. If a function is smooth enough
then the Taylor series can approximate the function arbitrarily accurately, and the
remainder can be estimated using Taylors theorem. For instance, we have
ex = 1 + x +

x3
x4
x2
+
+
+ o(x4 )
2
6
24

(x 0)

Figure 2.1. The exponential function y = ex (solid red curve)


and the corresponding Taylor polynomial of degree four (dashed
green curve) around the origin.

This is a special case of an asymptotic expansion or asymptotic series. If


a function f has an asymptotic expansion of the form

an n (z)

n=0

it means
(2.1)

f (z) =

N
X
n=0

We often write

an n (z) + o(N (z))

(z z0 )

MA4547 ASYMPTOTIC ANALYSIS

f (z)

an n (z)

n=0

for short. This is a very general definition. If the infinite sum on the right hand
side is convergent, then this is the same as the notation introduced before, but
the sum does not have to be convergent. In other words, the formal series

X
an n (z)
n=0

may not be equal to any function. In this case, we must use (2.1) as the definition.
The funcitons n (z) can be any type of function, but in practice we only deal with
simple functions such as powers, exponentials, and logs.
A natrual question that arises is why are we doing this? After all when we
encounter a divergent power series our first reaction is to discard it as nonsense.
However, in fact there are situations when a divergent power series would make
more sense than a convergent one.
Example 3. The so-called error function is defined as
z
2
2
et dt
erf (z) =
0
It is a rescaled version of the cumulative distribution function of the standard
normal distribution in probability and statistics. It is an entire function, and has
a Taylor series representation

2 X (1)n z 2n+1
erf (z) =
n=0 n!(2n + 1)

This series is convergent for all z, but how useful is it really?


If we take the first 20 terms (n = 0, 1, ...19) of this series we get a long Taylor
polynomial of degree 39 which looks fine for z close to 0 but fails to approximate
the function when z becomes close to 3 (relative error bigger than 20%).
The asymptotic series for the error function as z + is

(2.2)

2
ez X (1)n+1 (2n 1)!!
erf (z) 1 +
(2z 2 )n
z n=0

where (2n 1)!! = 1 3 5...(2n 1) and (1)!! = 1. The series is divergent by


the ratio test, but if we take just three terms and compare the values to the error
funciton, the result is amazing (see figure below): the three-term truncation


1
1
z 2

+1
2 z 3
z
approximates the error function with a relative error less than 7% for all z > 1. For
z > 2 the relative error is less than 0.02%.
The divergent asymptotic series is far more useful than the convergent Taylor
series for large values of z. However, since it is divergent, the accuracy of approximaton near a fixed z may not become better if one takes more terms of the series.
Suppose I want an approximation of the error function near z = 1 with a relative

MA4547 ASYMPTOTIC ANALYSIS

Figure 2.2. The error function (blue curve) and the corresponding Taylor polynomial of degree 39 (orange curve) around the origin.

Figure 2.3. The error function (blue curve) and the corresponding 3-term asymptotic series (orange curve) at infinity.

error less than 1%, then the asymptotic series (2.2) will be useless no matter how
many terms I take. In that case the Taylor series would be a much better choice.
Taking more terms of the asymptotic series will improve the accuracy for larger z
but may actually make the approximation worse for a fixed z. In practice we rarely
need a lot of terms of asymptotic series.
We already know that we can add, subtract, multiply, differentiate, and integrate power series. We can also divide Laurent series. Similarly, one can perform
arithmetic operations (carefully) on asymptotic series, as well as integrate them,
but one in general cannot differentiate asymptotic series, unless it only contains
powers of z.
Example 4. suppose

MA4547 ASYMPTOTIC ANALYSIS

f (z) = z + sin(z 2 )
then we have
f (z) z

(z +)

but
f 0 (z) = 1 + 2z cos(z 2 )  1

(z +)

It is improtant to keep in mind that asymptotic series are merely approximations.


Even if an asymptotic series is convergent, it may not be equal to the function.
Example 5. The function
1
+ ez
z1
has the convergent asymptotic expansion

X
f (z)
z n (z +)
f (z) =

n=1

since

1
1
1 X 1 n X n
1
( ) =
z
=
=
z1
z 1 1/z
z n=0 z
n=1

but of course it is not equal to the sum. Note that the sum will not be an asymptotic
expansion of f if z . This is because is an essential singularity of f .
Since asymptotic expansions are just approximations, there is no uniqueness
in the sense that a function must have a fixed asymptotic series (there are many
different ways to approximate a function). There is only restricted uniqueness. For
example, in the case of z , if an asymptotic series consists only of powers of
z, exponentials of z, and logs of z, then the coefficient of each term is unique. In
particular, the coefficient of each term of (2.2) is unique. In practice this is not
an important issue. Theoretical physicists use all kinds of approximate formulas
since most of their equations cannot be solved explicitly. Sometimes there are a few
different formulas for calculating the same physical quantity, and they are used in
different circumstances. As for us, we are primarily interested in asmptotic series
consisting of powers of z only, which are called asymptotic power series. For
instance, (2.1) is 1 plus an exponential function times an asymptotic power series.

MA4547 ASYMPTOTIC ANALYSIS

10

Exercise. 1. Find the complete asymptotic power series for the following functions
as z +:
z
(1)
z+2
1
(2)
(z 1)2
(3) ln(z + 1) ln z

(4) z e1/t dt
2. Find the first three terms of the asymptotic power series for the following functions as z +:

(1) z + 1
1
(2) 2
z +z+1
z
(3) sin(
)
1z

MA4547 ASYMPTOTIC ANALYSIS

11

3. Asymptotic techniques for calculating integrals


3.1. Integration by parts. We already know that one can use integration by
parts to evaluate certain integrals. The same technique can also be used to find
asymptotic behaviors of integrals.
Example 6. The exponential integral function E1 is given by

et
dt
t
x
The integral cannot be evaluated explicitly, but we can perform integration by
parts as follows:
E1 (x) =

t
et
e
ex
ex
et
dt =
|x
dt =
+ O( 2 )
2
t
t
t
x
x
x
x
This is a simple asymptotic expansion of only one term. To obtain more terms
one can continue to perform integration by parts.

t
t
et
ex
ex
e
ex
e
dt =

dt
=
+
2
dt = ...
2
2
t
x
t
x
x
t3
x
x
x
The complete asymptotic series is

X
(1)n1 (n 1)!
et
dt = ex
t
xn
n=1

This is obviously a divergent series, but the accuracy of approximation is quite


good.
0.014
0.012
0.010
0.008
0.006
0.004
0.002

10

Figure 3.1. The exponential integral function E1 (blue curve) and


the corresponding 3-term asymptotic series (red curve) at infinity.
In the table below we compare the numerical values of E1 with its 3-term and
4-term truncated asymptotic series. The 4-term series is less accurate than the
3-term series for x 6 4, but more accurate for x > 5. For convergent series more
terms means better accuracy but for divergent series this is not always true.

MA4547 ASYMPTOTIC ANALYSIS

E1 (x)

ex x23 x12 + x1

ex x64 + x23 x12 + x1
Table 1. 3-term

x=3
0.0130484
0.0147517
0.0110638
and 4-term

x=4
x=5
0.00377935 0.0011483
0.00400655 0.00118588
0.00357727 0.00112119
asymptotic expansions of

12

x=6
x=7
0.000360082 0.000115482
0.000367223 0.000116976
0.000355747 0.000114697
E1 (x)

3.2. Watsons lemma. The Laplace transform is a very powerful tool in sciences
and engineering. This involves simple Laplace integrals of the type

f (p)ept dp

I(t) =

(3.1)

where b > 0 may be taken as . In real applications t often stands for time, and
people want to know the large time behavior of a certain (physical, chemical...)
system. The result is given by
Theorem 7 (Watsons lemma). Assume the integrand in (3.1) is bounded for large
t and

f (p)

(3.2)

an p+n

n=0

where > 1, > 0. Then


I(t)

X
an ( + n + 1)
t+n+1
n=0

where is the gamma function

sx1 es ds

(x) =

(3.3)

Proof. The main idea is very simple: just plug (3.2) into (3.1) and integrate term
by term. The key step is

(a + 1)
+ O(ebt/2 )
ta+1
0
0

The complete proof requires more detailed estimates and can be found in most
textbooks.

pa ept dp =

pa ept dp +

pa ept dp =

Example 8. We can use Watsons lemma to obtain the asymptotic series for the
Laplace integral
pt
e
I1 (t) =
dp
1
+ p2
0
Note that

X
1
=
(1)n p2n
1 + p2
n=0

MA4547 ASYMPTOTIC ANALYSIS

13

By Watsons lemma we get


I1 (t)

X
(1)n (2n + 1)!
t2n+1
n=0

since (n) = n! for every integer n.


0.24
0.22
0.20
0.18
0.16
0.14
0.12

10

Figure 3.2. The function I1 (t) (blue curve) and the corresponding 3-term asymptotic series (red curve) at infinity.
Remark. Watsons lemma is valid not just for t +, but for all |t| , arg t
(/2, /2) .
3.3. Laplaces method. We are familiar with elementary functions such as polynomials, exponentials, logs, and trigonometric functions. However, these functions
are often insufficient to describe complicated phenomena encountered in sciences
and engineering. For instance, many differential equations cannot be solved using
these elementary funcitons. Instead, their solutions are the so-called special functions. One example is the Gamma function (3.3) we have seen before. Many special
functions have integral representations, the simplest among which are Laplace type
and Fourier type integrals. The method for calculating asymptotic behaviors of
these integrals are called Laplaces method and the method of stationary phase
respectively. We will study these two methods in this and the next section.
A general Laplace integral is of the form

(3.4)

f (p)exg(p) dp

I(x) =
a

where we assume f and g to be real continuous functions.


At first sight, one may want to do use the substitution s = g(p) to transform
the integral into the simple type (3.1) and apply Watsons lemma. However, this
does not always work, since g(p) is not necessarily one-to-one and even if it is there
may not be an explicit formula for g 1 . Instead, we will need to examine the
behavior of g(p) carefully.
One important observation is that the value of I(x) for large x depends crucially
on the (global) maximum value of the kernel exg(p) , which depends on the maximum

MA4547 ASYMPTOTIC ANALYSIS

14

value of g(p). This is because if g(p1 ) < g(p2 ) then exg(p1 ) is exponentially small
compared to exg(p2 ) . Thus the points near p2 contributes much more than the
points near p1 .
1
2
Example 9. Lets consider the Laplace integral
pex(pp ) dp. The function
0

p p2 reaches its maximum value at p = 1/2. The function pex(pp


fast if x is large and p deviates from 1/2 (see figure below).

Figure 3.3. Graph of the functionpex(pp

decays very

for x = 30

1
0.6
2
2
For instance, if x = 30, then 0 pex(pp ) dp 292.514 and 0.4 pex(pp ) dp
164.241. Most contributions to the integral come from points near 1/2.
In general the behaviors of f and g in (3.4) may be very complicated and it is very
difficult to obtain even the leading (first) term asymptotics for I(x). Fortunately,
functions encountered in practice are usually quite nice and it is often sufficient to
have the leading term asymptotics. We will consider three most important cases
below.
Case 1.
Case 2.
Case 3.

The maximum of g is reach at the left endpoint a, g 0 (a) < 0 and f (a) 6= 0.
The maximum of g is reach at the right endpoint b, g 0 (b) > 0 and f (b) 6=
0.
The maximum of g is reach at some (unique) point c (a, b), g 0 (c) = 0,
g 00 (c) < 0, and f (c) 6= 0.

In each case we can approximate the function g with its Taylor polynomial of degree
1 or 2, and the function f with its value at the maximum point. To be specific, we
have
Case 1.
Case 2.
Case 3.

g(p) g(a) + g 0 (a)(p a), f (p) f (a)


g(p) g(b) + g 0 (b)(p a), f (p) f (b)
g(p) g(c) + 21 g 00 (c)(p c)2 , f (p) f (c)

Using these approximations the integral I(x) can be calculated asymptotically.


b

0
f (a)exg(a)
x(g(a)+g 0 (a)(pa))
xg(a)
Case 1. I(x)
f (a)e
dp f (a)e
exg (a)s ds =
xg 0 (a)
a
0

MA4547 ASYMPTOTIC ANALYSIS

15

f (b)exg(b)
xg 0 (b)

ab

2f (c)exg(c)
x(g(c)+g 00 (c)(pc)2 /2)
xg(c)
xg 00 (c)s2 /2
Case 3. I(x)
f (c)e
e
dp f (c)e
ds = p
xg 00 (c)
a

The last integral in Case 3 follows from


Case 2.

f (b)ex(g(b)+g (b)(pb)) dp f (b)exg(b)

I(x)

exg (b)s ds =


2
2
2
eu du = p
ds = p
00 (c)
00 (c)
xg
xg

The reason we can change the integral bounds to infinity is because the additional
errors introduced by this approximation are exponentially small.
Now lets summarize our results.

exg

00

(c)s2 /2

Proposition 10 (Laplaces method). Consider the leading order large x asymptotics of the general Laplace integral

f (p)exg(p) dp

I(x) =
a

(1) If the maximum of g is reached at the left endpoint a, g 0 (a) < 0, and
f (a) 6= 0, then
f (a)exg(a)
xg 0 (a)
(2) If the maximum of g is reached at the right endpoint b, g 0 (b) > 0, and
f (b) 6= 0, then
I(x)

f (b)exg(b)
xg 0 (b)
(3) If the maximum of g is reached at some (unique) point c (a, b), g 0 (c) = 0,
g 00 (c) < 0, and f (c) 6= 0, then
I(x)

2f (c)exg(c)
I(x) p
xg 00 (c)
1
2
Example 11. Lets consider the Laplace integral
pex(pp ) dp. Since c = 1/2 is
0

the unique maximum point for p p2 , and



2


1
1
1
+ O (p )4
p p2 = p
4
2
2
we have
f (c) =

1
,
2

g(c) =

1
,
4

g 00 (c) = 2

using laplaces method we get

ex/4

2 x
0
As shown in the table below the formula is quite accurate.
pe

x(pp2 )

dp

MA4547 ASYMPTOTIC ANALYSIS

2
pex(pp ) dp
0
x/4
e

2 x

x=5

x=10

1.22585

3.3276

x=15

16

x=20

9.66976 29.3645

1.38334 3.41414 9.72979 29.4105


Table 2. Asymptotics of a Laplace integral

n! n n
2nn e

n=1
n=2
n=3
n=4
n=5
1
2
6
24
120
0.922137 1.919 5.83621 23.5062 118.019
Table 3. Stirlings formula

The next example is a bit advanced but the conclusion is very useful. We know
that factorials increase faster than even exponentials, but how fast does a factorial
increase? The answer is given by the famous Stirlings formula.
Example 12 (Stirlings formula). Recall that the Gamma function satisfies

sx es ds

(x + 1) =
0

To find out the large x asymptotics we first perform the (somewhat mysterious)
transforms

es+x ln s ds = x

(x + 1) =

exr+x ln(xr) dr = xx+1


0

ex(r+ln r) dr
0

Note that r + ln r reaches its maximum at c = 1 and



1
r + ln r = 1 (r 1)2 + O (r 1)3
2
This means
f (c) = 1,

g(c) = 1,

g 00 (c) = 1

By Laplaces method we get

x(r+ln r)

dr

2ex

Therefore we obtain Stirlings formula


(x + 1)

2xxx ex

This means
n!

2nnn en

As shown in the table below the formula is very accurate even for smaller vaues
of n.

MA4547 ASYMPTOTIC ANALYSIS

17

3.4. Method of stationary phase. In this section we study large x asymptotics


of general Fourier integrals of the type

f (p)eixh(p) dp

I(x) =

(3.5)

If a = b = and h(p) = p this would be the usual Fourier transform, which can
be calculated using integration by parts. In general, the function eixh(p) oscillates
rapidly for large x. This means there will be a lot of cancellations in the integral.
Thus we naturally expect to have the following
Theorem 13 (Riemann-Lebesgue Lemma). If f (p) is integrable and h(p) is continuously dierentiable on [a, b], but h(p) is not constant over any subinterval of [a, b],
then as x we have

f (p)eixh(p) dp = o(1)
a

The Riemann-Lebesgue lemma itself does not give any asymptotic information,
but it is essential for proving the asymptotic methods below. Due to the massive
cancellations of Fourier integrals, the main contribution should come from points
where the cancellations are the weakest. Possible candidates are endpoints (since
only one side of the interval is counted for the integral and there is no cancellation
from the other side) as well as points where h0 (c) = 0 (since h(p) h(c)+(pc)2 /2,
two sides of the interval centered at c do not cancel each other). In other words,
the main contribution to the integral come from the critical points. As in the case
of Laplace integrals, the behaviors of f and h in (3.5) may be very complicated and
difficult to anlayze, but in practice it is often sufficient to consider the following
two important cases:
Case 1.
Case 2.

h0 (p) 6= 0 for all p [a, b].


There is a unique point c (a, b) such that h0 (c) = 0. This point c is
called a stationary point.

The first case can be handled by integration by parts.

f (p)eixh(p) dp =

I(x) =
a

1
f (p)eixh(p) b
|
ixh0 (p) a ix

f (p)
h0 (p)

0

eixh(p) dp

The second integral is of order o(1/x) by the Riemann-Lebesgue lemma, which


means

I(x)

f (b)eixh(b)
f (a)eixh(a)

0
ixh (b)
ixh0 (a)

In the second case the main contribution comes from the stationary point c,
which (I believe) is why the method is called stationary phase.
We use the approximations h(p) h(c) + 21 h00 (c)(p c)2 , f (p) f (c) to obtain

MA4547 ASYMPTOTIC ANALYSIS

18

Figure 3.4. Graph of p cos(x(p p2 )) with x = 200. There are


rapid oscillations except near the stationary point 1/2.

f (c)e

f (p)eixh(p) dp

I(x) =

00

f (c)eix(h(c)+h

(c)(pc)2 /2)

dp

ixh(c)

ix(h00 (c)(pc)2 /2)

dp f (c)e

ixh(c)

2
00
x|h (c)|

eisgn(h

00

(c))s2

ds

where sgn(y) = 1 for y > 0 and sgn(y) = 1 for y < 0. The reason we can change
the integral bounds to infinity is because the additional errors introduced by this
approximation are of a smaller order than the main integral. Now we use the special
value of the integral

eis ds =

ei/4

to obtain the final formula


s
I(x) f (c)e

ixh(c)+isgn(h00 (c))/4

2
x|h00 (c)|

Note that the decay rate 1/ x is weaker than the endpoint case (1/x). This
means the contribution from the stationary point is stronger than contributions
from endpoints. Thus there is no need to consider endpoints as long as a stationary
point exists.
Now lets summarize our results.
Proposition 14 (Method of stationary phase). Consider the leading order large x
asymptotics of the general Fourier integral

(3.6)

f (p)eixh(p) dp

I(x) =
a

MA4547 ASYMPTOTIC ANALYSIS

1
0

x=10

x=12

19

x=14

x=16

x=18

p cos(x(p p
p ))dp -0.0551303 -0.164609 -0.224068 -0.227672 -0.180952
1
x

-0.0401627 -0.153561 -0.215588 -0.220967 -0.175521


cos(
)
2
4
x
Table 4. Asymptotics of a Fourier integral

(1) If h0 (p) 6= 0 for all p [a, b] then


f (b)eixh(b)
f (a)eixh(a)

ixh0 (b)
ixh0 (a)
(2) If there is a unique point c (a, b) such that h0 (c) = 0, then
s
00
2
I(x) f (c)eixh(c)+isgn(h (c))/4
x|h00 (c)|
I(x)

Example 15. Consider the Fourier integral

p cos(x(p p2 ))dp

I2 (x) =
0

We first write the cosine function as the real part of a complex exponential:

p cos(x(p p2 ))dp = <(


0

peix(pp ) dp)
0

Since p  p2 has a stationary point at c = 1/2 and p p2 =


O (p 12 )4 , we have
1
1
, h(c) = ,
2
4
Using the method of stationary phase we get
f (c) =

peix(pp ) dp
0

1
4


1 2
2

h00 (c) = 2

1 ix/4i/4
e
2

This means
r
1
x

1
cos(
)
+ o( )
2
4
x
x
The accuracy in this case is not as great as some of the previous examples since
the next order is o( 1x ) but not o( x1 ).
I2 (x) =

3.5. *Method of steepest descent. If an integral transform is more general than


the Laplace and Fourier types we have considered, there is a method called method
of steepest descent (or saddle-point method), which applies to integrals of the form

I(x) =
f (p)exg(p) dp
C

where C is a contour in the complex plane, and f and g are analytic functions. In
this case the main contribution comes from saddle points, points where g 0 (c) = 0.
One deforms the contour of integration so that along the new contour the imaginary

MA4547 ASYMPTOTIC ANALYSIS

20

part of g(p) is a constant near the saddle point. This new contour is called a steepest
descent contour. Then one can apply Laplaces method to obtain the asymptotic
behavior of I(x) for large x. We will not go into the details here since this method
is quite complicated.

MA4547 ASYMPTOTIC ANALYSIS

21

Exercise. 1. Find the complete asymptotic series for large positive x of the following functions:
t
e
(1) I(x) =
dt
t2
x

cos t
(2) The cosine integral function I(x) =
dt
t
x x
2
2
et dt
(3) The error function erf (x) =
0
xp
e
(4) I(x) =
dp
2+p
0 1
2
(5) I(x) =
ln(1 + p)exp dp
0 1 xp
e
dp
(6) I(x) =
0 1+p
2. Find the first three terms of the asymptotic series for large positive x of the
following functions:
t
e
(1) I(x) =
dt
1
t
x

ett dt

(2) I(x) =
x1
(3) I(x) =
0 1

ex(1x)p
dp
1+p
2

ex(pp ) dp

(4) I(x) =
1

3. Find the leading term asymptotics for large positive x of the following functions:
1
3
(1) I(x) =
pex(p+p ) dp
0 2
3
(2) I(x) =
p2 ex(3pp ) dp
0 2
3
(3) I(x) =
ln(1 + p)ex(2pp ) dp
1 1
2
(4) I(x) =
(1 + 2p)ex(p p) dp
0 1
3
(5) I(x) =
ep+ix(p+p ) dp
0 2
3
(6) I(x) =
p2 eix(3pp ) dp
0 1
2
(7) I(x) =
(1 + 2p)eix(p p) dp
0

1
cos(t x sin t)dt
(8) The Bessel function J1 (x) =
0

MA4547 ASYMPTOTIC ANALYSIS

22

4. Asymptotic solutions to algebraic equations


In this section we will introduce methods for finding asymptotic solutions to
certain algebraic equations which cannot be solved explicitly.
4.1. Inverse of analytic functions. We know from complex analysis that if a
function f (z) is analytic and f 0 (a) 6= 0, then f (z) is invertible near z = a, and its
inverse f 1 is analytic near f (a). This means f 1 has a Taylor series expansion.
But how can we calculate this series? There is a formula called Lagrange-Brmann
formula, which is extremely complicated. Fortunately there are much easier ways.
Since we can adopt the new variable x = z a and new function g(z) = f (z) f (a),
we can assume a = 0 and f (a) = 0 for simplicity.
Fact 16 (Method of undetermined coefficients). If f (z) has the Taylor series f (z) =

X
X
bn z n , then one can solve the equation f (z) = w by taking z =
an wn and
n=1

n=1

solve for the coefficients of an using the equation

bk (

an wn )k = w

n=1

k=1

Example 17. Lets find the first three terms of the Taylor series of the inverse
function of z + z 2 2 sin z at 0. This means we want to solve the equation
z + z 2 2 sin z = w
This means
w = z + z 2 +
By taking z =

z3
+ o(z 3 )
3

an wn we get the equation (note that z w)

n=1

w=

X
n=1

an w +

X
n=1

!2
an w

1
+
3

!3
an w

+ o(w3 )

n=1

Lets write out the terms up to degree three


a31 3
w + o(w3 )
3
This implies all the coefficients of w, w2 , w3 must be zero, that is
w = a1 w a2 w2 a3 w3 + a21 w2 + 2a1 a2 w3 +

1 = a1
a2 + a21 = 0
a31
=0
3
So we get a1 = 1, a2 = 1, a3 = 5/3. The first three terms of the inverse
series is therefore
a3 + 2a1 a2 +

MA4547 ASYMPTOTIC ANALYSIS

23

4
z = w + w2 w3 + o(w3 )
3
In general it is not possible to find a formula for all terms.
Example 18. If we want to find the inverse function of 1 + z 2 + z 5 near z = 1, we
can take
z1=

an (w 3)n

n=1

since if z = 1then w = 1 + z 2 + z 5 = 3. This means


w3=(

an (w 3)n + 1)2 1 + (

n=1

an (w 3)n + 1)5 1

n=1

Then we can proceed as before (but using w1 = w 3 as a new variable) to find


the first few terms of the series.

X
bn z n , then
Fact 19 (Method of iteration). If f (z) has the Taylor series f (z) =
n=1

the equation f (z) = w is the equivalent to


w

bn z n

n=2

z=

(4.1)

b1
We can solve this using iteration. In the first step we set z = 0 and compute
the right hand side of (4.1), which yields z = w/b1 . In the second step we plug this
value into the right hand side of (4.1) again, which gives

w X bn

b1 n=2 b1

In the third step we plug the value


to get

w X bn

b1 n=2 b1

w
b2
3 w2
b1
b1

w
b1

n
=

w
b2
3 w2 + o(w2 )
b1
b1

w
b2
3 w2 into the right hand side of (4.1) again
b1
b1
n
=

w
2b2 + b3 3
b2
w + o(w3 )
3 w2 +
b1
b1
b41

We may continue this procedure to obtain the desired inverse power series.
Example 20. Lets find the first three terms of the Taylor series of the inverse
function of z z 2 + 2 sin z using the iteration method. Since
w = z + z 2 +

z3
+ o(z 3 )
3

we have
z3
+ o(z 3 )
3
The first iteration is just w. The second iteration is
z = w + z 2 +

MA4547 ASYMPTOTIC ANALYSIS

24

w + w2
and the third iteration is
w + (w + w2 )2 +

(w + w2 )3
5
= w + w2 w3 + o(w3 )
3
3

just as before.
The general case is very similar.
Example 21. If we want to find the inverse function of 1 + z 2 + z 5 near z = 1, we
note that
1 + z 2 + z 5 = 3 + 7(z 1) + o(z 1)
Thus we can take
1 + z2 + z5 3
w3
+(
(z 1))
7
7
and iterate as before, treating z1 = z 1 and w1 = w 3 as new variables. If one
wishes one may write
z1=

w1
1 + (z1 + 1)2 + (z1 + 1)5 3
+(
z1 )
7
7
which may be easier to iterate.
z1 =

Fact 22. The method of iteration can also be used to calculate asymptotic series
which are not power series, as long as the equation can be written in the form
(4.2)

z = g(w) + h(z)

where h(z) = o(z). The first iteration is g(w). The second one is g(w) + h(g(w))
(truncated to the correct order), and so on.
Example 23. If we want to find the inverse function of ez (1 + 1/z) near +, we
set
1
w = ez (1 + )
z
and take logs on both sides, which yields
1
ln w = z + ln(1 + )
z
This means we can iterate
1
1
1
1
z = ln w ln(1 + ) = ln w + 2 3 + ...
z
z
2z
3z
The first iteration gives ln w. The second iteration gives
1
ln w
ln w
The third iteration gives

MA4547 ASYMPTOTIC ANALYSIS

ln w

1
ln w

1
ln w

1
2(ln w

1 2 +o((ln w)
ln w )

) = ln w

25

1
1
+o((ln w)2 )
+
ln w 2(ln w)2

and so on.
4.2. Asymptotic solutions for large x. We can also use the method of iteration
to find asymptotic solutions to algebraic equations. In particular if there are infinitely many solutions going to infinity we can find the asymptotic values of these
solutions. The major task is to write the equation in a proper form which allows
for iterations.
Example 24. Lets go back to our very first example x sin x = 1. We know there
are infinitely many solutions, but how can we find them asymptotically? The first
observation is that since x goes to infinity, sin x must small in order for the product
x sin x = 1 to be true. In other words sin x = 1/x = o(1). We know that sin x
is close to 0 if and only if x is close to k for integers k. Thus we may write
x = k + t (0 < t < ) and the equation becomes
(1)k (k + t) sin t = 1

(4.3)
which is equivalent to


(4.4)

t = arcsin

(1)k
k + t

= (1)k arcsin

1
k + t

Since t is of order 1/k and




 
1
1
1
arcsin
arcsin
= O( 2 ) = o(t)
k + t
k
k
the right hand side of (4.4) is of the correct form as in (4.2), meaning we can
iterate the right hand side of (4.4) after taking the large t power series. That is,
we can iterate
(4.5)

t = (1)k

1
t
1 + 6t2
1
+ O( )4
2 2+
3
3
k k
6 k
k

The first iteration gives


t=

(1)k
k

The second iteration gives


t=

1 (1)k /6
(1)k

k
3 k3

and so on.
Alternatively, since t = o(1) we can expand the sin t in (4.3) and get
(1)k (k + t)(t

t3
+ O(t5 )) = 1
6

which implies
1
kt3 + kt + t2 = (1)k + o(t3 )
6

MA4547 ASYMPTOTIC ANALYSIS

26

which as the iterative form


(4.6)

t=

(1)k
t2
1

+ t3 + o(t3 )
k
k 6

The first iteration gives


t=

(1)k
k

The second iteration gives


(1)k
1 (1)k /6

k
3 k3
as before. Although the iterative equations (4.5) and (4.6) are different, they give
the same result. As long as we require the series to consist of powers of 1/k, the
series is unique.
t=

MA4547 ASYMPTOTIC ANALYSIS

27

Exercise. 1. Find the first three terms of the inverse power series at z = 0 of the
following functions:
(1) z 3 z 2 + sin z
(2) z 2 + ln(1 + z)
(3) ez cos z
sin z
(4)
1 + z2
2. Find the first three terms of the inverse power series of the following functions:
2 z3
(1)
at z = 1
1+z
(2) z 5 + 2z 2 at z = 1
3. Find the first three terms of the asymptotic series of the solutions to the following
equations (for large integer k):
(1) x cos x = 2
cos x
=0
(2) sin x +
x

MA4547 ASYMPTOTIC ANALYSIS

28

5. Asymptotics of ordinary differential equaitons


Ordinary differential equations (ODEs) are very important in applications including sciences and engineering, but most ODEs cannot be solved explicitly. Sometimes one can use power series expansions or numerical analysis to obtain satisfactory approximations, but in many cases we need to use asymptotic methods. In
this section we will introduce three methods: iteration, perturbation expansion,
and WKB.
5.1. The method of iteration. Suppose we ant to find out the behavior of a
solution to some ODE for large x, then neither power series nor numerical simulation
would be an ideal choice since they do not have sufficient accuracy for large x.
Instead, we may use iteration methods just as we did for algebraic equations. The
procedure is as follows:
Fact 25 (Method of iteration). Suppose we have an ODE of the form
y = f (x) + g(y, y 0 )
where f (x) = o(1) can be written as a power series in 1/x, and g(y, y 0 ) = o(y),
then we may iterate the right hand side to obtain an asymptotic series. The first
iteration gives y = f (x) (truncated to leading order). The second iteration gives
y = f (x) + g(f (x), f 0 (x)) (truncated to second order) and so on. y will also have a
power series in 1/x.
Example 26. The ODE y 0 = y + 1/x can be solved in terms of the exponential
integral function we have studies before. The general solution is
x t
e
x
x
x
x
dt
y(x) = Ce e E1 (x) = Ce + e
t
There is a unique solution of order o(1) for large x, which corresponds to C = 0.
We already know how to find the asymptotic series of the solution using integration
by parts, but in fact we can find the asymptotic series without solving the equation
explicitly. We simply write
1
y0
x
and iterate the right hand side. The first iteration gives 1/x. the second gives
1
1
1
1
2
+ .The third gives + 2 + 3 , and so on. This works because the derivative
x x2
x x
x
of 1/xk is of order 1/xk+1 , so y 0 belongs to the right hand side.
y=

If it is clear that y has an power series in 1/x, one can also use the method of

X
undetermined coefficients by setting y(x) =
xk can plug it into the ODE. For
k=1

ODEs the series are typically divergent (unlike for algebraixc equations).
Example 27. Lets consider the ODE
1
x
The general solution cannot be found explicitly. However, we can use the method
of iteration to find certain special asymptotic solutions. There is no simple solution
(5.1)

y 00 y 0 + y + y 2 = 2 +

MA4547 ASYMPTOTIC ANALYSIS

29

of order o(1) due to the 2 on the right hand side. However, ther are two solutions
behaving like constants. They satisfy
y + y2 = 2
That is, we have solutions y1 = 1 + o(1) and y2 = 2 + o(1). Lets focus on y1
and y2 can be similarly studied (this is an exercise). By setting y = 1 + u we can
rewrite (5.1) as
u00 u0 + 3u + u2 =

1
x

This gives the iterative equation




1 1
u2 + u0 u00
u=
3 x
1
1
4
The first iteration gives u =
. The second gives u =

. The third
3x
3x 27x2
22
4
1
gives u =

+
, and so on.
243x3
27x2
3x
It is possible to prove the existence real solutions which the asymptotic solutions
approximate, but the proof is very advanced.
5.2. Perturbation expansions. ODEs arising in practice often contain unknown
parameters and people often want to find out the limiting behavior when the parameter is small.
Fact 28. If an ODE contains a small parameter which is not multiplying the

X
term of the highest derivative, we may write the solution as y =
k yk , plug it
k=0

into the ODE, collect terms according to powers of , and solve the yk s one by one.
Example 29. Lets consider the initial value problem
(5.2)

y 00 (x) = (1 + x)y(x);

y(0) = 1;

y 0 (0) = 1

and we are interested in the behavior of the solution for small and bounded
x. The solution can be expressed in terms of Airy functions, which are fairly

X
complicated. Instead, we write y =
k yk and plug this into the ODE which
k=0

gives

k yk00 = (1 + x)

k=0

k yk

k=0

Collecting power of we get


y000 = y0 ;

y0 (0) = 1;

y100 y1 = xy0 ;
and in general

y00 (0) = 1

y1 (0) = 0;

y10 (0) = 0

MA4547 ASYMPTOTIC ANALYSIS

30

Figure 5.1. Solutions to the ODE (5.2) (blue curve) and its 3term perturbation expansion for = 0.2 (orange curve)

00
yn+1
yn+1 = xyn ; yn+1 (0) = 0;
We can solve them term by term. This yields

0
yn+1
(0) = 0

y0 (x) = ex
y1 (x) =
y2 (x) =



1 x 2x
e
e
2x2 2x + 1 1
8



1 x
e
6x2 + e2x 6x4 20x3 + 36x2 36x + 15 + 6x 15
192

and so on.
The solution is therefore written as


 2 x



y(x) = ex + ex e2x 2x2 2x + 1 1 +


e
6x2 + e2x 6x4 20x3 + 36x2 36x + 15 + 6x 15 +O(3
8
192
As shown in the figure below, the approximation is quite accurate.
Another example is
Example 30. Lets consider the initial value problem
y 00 = y 0 + y 2 ;

(5.3)
By setting y =

y(0) = 1;

y 0 (0) = 0

k yk we get the equation

k=0

k yk00

k=0

k yk0

k=0

!2
k

yk

k=0

Collecting power of we get


y000 = y00 ;

y0 (0) = 1;

y00 (0) = 0

MA4547 ASYMPTOTIC ANALYSIS

31

Figure 5.2. Solutions to the ODE (5.3) (blue curve) and its 3term perturbation expansion for = 0.2 (orange curve)

y100 y10 = y02 ;

y1 (0) = 0;

y200 y20 = 2y0 y1 ;

y10 (0) = 0

y2 (0) = 0;

y20 (0) = 0

and so on. This means


y0 (x) = 1
y1 (x) = x + ex 1
y2 (x) = x2 + 2ex x + 4x 6ex + 6
and so on. So
y(x) = 1 + (x + ex 1) + 2 (x2 + 2ex x + 4x 6ex + 6) + O(3 )
The plots are shown below.
5.2.1. Singular perturbations. If the small parameter appears in front of the term
with the highest derivative, the regular perturbation expansion method may fail.
Example 31. The initial value problem
y 0 y = x;

y(0) = 1

has a solution
y(x) = x + ( + 1)ex/
There is no expansion in powers . If one attempts such an expansion, the first
term should satisfy y0 (x) = x; y0 (0) = 1 which has no solution.

MA4547 ASYMPTOTIC ANALYSIS

32

We say this problem is singularly (instead of regularly) perturbed. To solve the


problem, we need to perform the substitution x = t which transforms the ODE
into
y 0 (t) y(t) = t; y(0) = 1
The new ODE is regularly perturbed and can be solved using the perturbation
expansion method. However, if we want the result to be valid for say x [0, 1], we
will need t [0, 1/]. The new expansion may not be valid for t so large. In that
case we need to use the new ODE for small t, the orignal ODE for x of order
1,
and perhaps also use a formula for the intermediate region (for instance t 1/ )
to get a uniform description for the solution. The method of singular perturbation
is quite complicated.
5.3. The WKB method. The WKB method is a popular asymptotic method
used in physics, and it is named after three physicists. There are several different
senarios where it can be applied, and we will focus on a particularly important one,
namely the large x behavior of solution to ODEs. We have seen previously that
there may be solution which have asymptotic power series in 1/x, but these are
just special cases. General solutions will not behave like that and simple equations
like y 0 = y has no such solution. However, for a fairly large class of linear ODEs
there are solutions of the type
cx

yx e

(5.4)

X
ak
xk

k=0

where > 0 and c may be complex.


Fact 32 (WKB method). To find a solution of the type (5.4), one can simply plug
it into the ODE and determine the unknown coefficients , , c, ak , ... However, it

is usually simpler if one first plug y = ecx into the ODE and determine c, , then

plug y = x ecx into the ODE to determine , and finally plug in the full expansion
(5.4).
Example 33. Lets begin with the ODE
1
)y(x)
x
The general solutions can be represented using hypergeometric functions, and they
are quite complicated. The WKB solution is relatively simple. First we attempt
b
the approximate solution y = ecx and plug it into the ODE. This gives
y 00 (x) = (4 +

bcxb2 ecx


b
1
bcxb + b 1 = (2 + )ecx
x

To leading order we have


b2 c2 x2b2 = 4
This means b = 1, c = 2 since there are two linearly independent solutions.
Lets focus on the solution with c = 2. The other one is left as an exercise. Then
we attempt the approximate solution y = xa e2x and plug it into the ODE. This
gives

MA4547 ASYMPTOTIC ANALYSIS

33


e2x xa2 a2 + a(4x 1) x = 0
which means a = 1/4. Finally we let
y(x) = x1/4 e2x (1 +

ak xk )

k=1

and plug this into the ODE. Here we fix a0 = 1 since the solution to the homogeneous ODE has an arbitrary multiplicative constant. We get
(128a2 21a1 )e2x
(192a3 77a2 )e2x
(64a1 + 3)e2x

+ ... = 0
7/4
11/4
16x
16x
16x15/4
3
63
1617
This means a1 = , a2 =
, a3 =
, and so on. The WKB solu64
8192
524288
tion is

63
1617
3

+ O(x4 ))
64x 8192x2
524288x3
Our next example is the Bessel equation whose solutions are called Bessel functions.
y(x) x1/4 e2x (1

Example 34. The Bessel equation of order 1 is


x2 y 00 + xy 0 + (x2 1)y = 0
b

To find a WKB solution, we first attempt y = ecx and plug it into the ODE
which gives

b
ecx b2 c2 x2b + b2 cxb + x2 1 = 0
To leading order this means
b2 c2 x2b = x2
which implies b = 1, c = i. We focus on the solution with c = i as the other
one is just the complex conjugate of the first one. We let y = xa eix and plug it into
the ODE which gives

eix xa a2 + 2iax + ix 1 = 0
To leading order this means a = 1/2. Then we let
y(x) = x1/2 eix (1 +

ak xk )

k=1

and plug this into the ODE. This gives


1 ix
e ((3 8ia1 )x4 + (5a1 16ia2 )x3 + 3(7a2 8ia3 )x2 + ...) = 0
4x9/2
This means
3 8ia1 = 0
5a1 16ia2 = 0

MA4547 ASYMPTOTIC ANALYSIS

34

7a2 8ia3 = 0
3i
15
105i
and so on. Consequently a1 = , a2 =
, a3 =
, ... So the solution is
8
128
1024
105i
15
3i

+ O(x4 ))
y(x) x1/2 eix (1 +
+
8x 128x2
1024x3
The other solution is just the complex conjugate
3i
105i
15
+
+ O(x4 ))
+
8x 128x2
1024x3
Our next example is the Airy equation whose solutions are called Airy functions.
y(x) x1/2 eix (1

Example 35. The Airy equation is


y 00 = xy

(5.5)

To find a WKB solution, we first attempt y = ecx and plug it into the ODE
which gives
b


b2 c2 x2b + (b 1)bcxb x3
=0
x2
To leading order this means
ecx

b2 c2 x2b x3 = 0
which implies b = 3/2, c = 2/3. Lets first look at the solution with c = 2/3.
3/2
We let y = xa e(2/3)x and plug it into the ODE which gives



1 2x3/2 a2  2
e 3 x
2a + a 4x3/2 2 + x3/2 = 0
2
To leading order this means a = 1/4. Then we let
3/2

y(x) = x1/4 e(2/3)x

(1 +

ak x3k/2 )

k=1

and plug this into the ODE. This gives


2x3/2
3


(96a2 77a1 )x9/2 + (48a1 5)x6 + (144a3 221a2 )x3 + ...

=0
16x33/4
5
385
85085
which means a1 =
, a2 =
, a3 =
, ...
48
46088
663552
So the solution is
e

85085
5
385
+
+
+ O(x6 ))
3
3/2
46088x
48x
663552x9/2
Alternatively, one may do the change of variable t = 32 x3/2 . This means
y(x) x1/4 e(2/3)x

3/2

(1 +

dy
dy dt
dy
=
= x1/2
=
dx
dt dx
dt
and therefore

3t
2

1/3

dy
dt

MA4547 ASYMPTOTIC ANALYSIS

35

 1/3
 1/3  1/3 2/3
 1/3 2
d2 y
dt d
t1/3 dy 3t 2/3 d2 y
3t
3t
3
dy
3t
dy
t
d y
=
)
=
(
)
=
(
+
+( )
dx2
dx dt
2
dt
2
2
3 dt
2
dt2
2
dt2
22/3 31/3 dt
In the new variable the Airys ODE becomes
1 0
y (t) = y(t)
3t
We let y = ta et and plug it into the ODE which gives
y 00 (t) +

(5.6)


1 t a2
et
3a2 + a(6t 2) + t = 0
3
This means a = 1/6. Then we let

X
y(t) = t1/6 et (1 +
ak tk )
k=1

and plug this into the ODE. This gives



et (72a1 5)t4 + (144a2 77a1 )t3 + (216a3 221a2 )t2 + ...

=0
36t37/6
385
85085
5
, a2 =
, a3 =
, ... One can check that the sowhich means a1 =
72
10368
2239488
lution is the same as before.
However, since (5.6) has a symmetry, namely if y(t) is a solution then y(t)
is also a solution, we can immediately get the other solution by changing x3/2 to
x3/2 in the first solution:
5
385
85085
+

+ O(x6 ))
46088x3
48x3/2
663552x9/2
The accuracy of this approximation is very good.
3/2

y(x) x1/4 e(2/3)x

(1

0.4

0.3

0.2

0.1

1.5

2.0

2.5

3.0

3.5

4.0

Figure 5.3. Airy function y (blue curve) and its 4-term asymptotics (red curve)

MA4547 ASYMPTOTIC ANALYSIS

36

Exercise. 1. Find the first three terms of the asymptotic power series of solutions
to the following ODEs with the given behavior:
1
; y = o(1)
x
1
(2) y 00 + yy 0 + y = 2 ; y = o(1)
x
2
00
(3) y + sin y = ; y = o(1)
x
1
(4) y 00 + y 2 = 1 + ; y = 1 + o(1)
x
1
x
(5) y 00 (1 + )y =
; y = 1 + o(1)
x
1+x
1
(6) xy 000 y 00 + y = ; y = o(1)
x
(7) y 00 + (y 0 )2 + xy = x 1; y = 1 + o(1)
2. Find the first three terms of the small perturbation expansions of solutions to
the following ODEs:
2
(1) y 0 + y + y 2 = 0; y(1) = 1
x
(2) y 0 + y + y 2 = 0; y(0) = 1
(3) y 00 y 0 + xy = 0; y(0) = 1, y 0 (0) = 0
(4) y 00 + y e2x y = 1; y(0) = 0, y 0 (0) = 1
2
(5) y 00 + y 0 2 y = x; y 00 (0) = 1
x
(6) y 00 + y 0 xy = 1; y(0) = 1, y 0 (0) = 0
(7) (1 + x)y 00 6y = 2; y(1) = 0, y() = o(1)
3. Find the first three terms of the WKB solutions to the following ODEs (each
equation has two linearly independent solutions):
1
(1) y 00 (x) = (4 + )y(x)
x
1
1
(2) y 00 (x) = (1 + 2 )y(x)
x x
(3) The linearized Riccati equation x2 y 00 + (x2 2)y = 0
(4) The Bessel equation of order 2: x2 y 00 + xy 0 + (x2 4)y = 0
x2
(5) The Weber differential equation y 00 + (1 )y = 0
4
(1) y 0 2y + y 2 =

You might also like