An introduction to perturbation methods

Aditya S. Khair
Department of Chemical Engineering, Carnegie Mellon University,
Pittsburgh, PA 15213, United States.
[email protected]

April 2, 2013

Abstract

These notes introduce the basic principles of perturbation methods, along with
the application thereof to the solution of mathematical problems encountered in the
physical and engineering sciences.

Contents

1 Preface 2

2 Module 1: Introduction 2

3 Module 2: Quadratic equations 5

4 Module 3: Nonlinear reaction in a thin film 8

5 Module 4: Quadratic equations redux 14

6 Module 5: Boundary layers & matched asymptotic expansions 17

7 Module 6: Nonlinear reaction in a thin film at large reaction rates 21

1
8 Acknowledgments 25

1 Preface

The purpose of this document is to introduce perturbation methods as a means of solving

mathematical problems in the physical and engineering sciences. The reader is expected to
be familiar with multivariable calculus and ordinary dierential equations, at a level that
a typical Engineering or Science major would be exposed to by their sophomore year. At
Carnegie Mellon University (CMU) the relevant material is covered in 21-259, Calculus
in Three Dimensions; 21-122, Integration, Dierential Equations and Approximation;
and 06-262, Mathematical Methods of Chemical Engineering. Knowledge of basic fluid
dynamics is also desirable, e.g. as covered at CMU in 06-261, Fluid Mechanics. Thus, the
material in these notes should be readily accessible to sophomore, junior, and senior-level
undergraduates; indeed, the notes were written primarily with undergraduate students in
mind. The emphasis is on employing perturbation methods to obtain physical insight into
the solution of science and engineering problems, rather than providing mathematically
rigorous proofs for the underlying techniques. In a similar vein, the writing is presented
in a rather informal style.

The notes are partitioned into modules, which following a logical succession. The mod-
ules can be self-studied, or they could each be used as the material for a fifty minute
lecture. Learning objectives and student outcomes are described at the beginning of each
module. A number of exercises (denoted as Ex.) have been included throughout each
module: they should be attempted during self-study, or can be posed as questions during
a lecture. Homework exercises are provided at the conclusion of each module: again, these
can be tackled in self-study, or can be used as part of a homework set for a lecture course.
Solution sets for the homework exercises can be obtained upon request to the author. A
total of six modules are given, corresponding to about three hundred minutes of lecture
time. Only a glimpse of the utility and scope of perturbation methods can be presented
in this timeframe; further reading for the interested student is provided in the references
at the end of the document.

2 Module 1: Introduction

Learning Objective: Introduce the student to the concepts of nondimensionalization,

scaling, and perturbation methods in relation to mathematical problems encountered in
science and engineering.

Student Outcome: The student recognizes that perturbation methods are a useful tool
for solving mathematical problems.

2
Many physical processes and engineering problems are described mathematically by equa-
tions that cannot solved exactly. (Here, an exact solution is one that can be written in
terms of a finite number of elementary functions: e.g. powers, exponentials, trigonometric
functions etc.) That may comes as a surprise to the reader, whose mathematical edu-
cation during undergraduate studies typically focuses on finding exact solutions. What
progress can be made on problems that cannot be solved exactly? One approach is to
solve the equations using numerical methods, which are usually implemented on a com-
puter: there are a number of software packages (e.g. Matlab, Comsol) that do this, and
do it well. However, it is important to recognize that this approach does not provide
an exact solution to a problem, since the first step in a numerical method invariably in-
volves approximating the exact equation to be solved. For example, this might involve
a finite-dierence discretization of an ordinary dierential equation to yield a system of
algebraic equations. Moreover, the solution of the algebraic equations itself is obtained
only approximately on a computer. Of course, the idea (or hope) is that the error in the
approximation can be controlled and improved systematically, using a finer discretization,
for instance. Sometimes, unfortunately, numerical methods fail to converge to a solution
or are prohibitively computationally expensive.

Notably, many mathematical problems that occur in science and engineering contain a
small parameter (when appropriately nondimensionalized), which we will call . The idea
behind perturbation methods is to exploit the smallness of to obtain approximate so-
lutions, valid technically in the limit ! 0, to problems that cannot be solved exactly
and that might be difficult to treat numerically. However, one should view perturbation
methods as a complement, rather than a replacement, for numerical methods. One advan-
tage of perturbation methods is that can give a analytical approximation to the solution
of a problem (valid as ! 0), which often yields considerable physical insight, whereas
numerical techniques provide the solution at a series of discrete values. Of course, the
latter are not restricted (in principle) to small values of , so there is clearly a trade-o.
However, as we shall see, oftentimes perturbations methods can provide accurate results
well beyond the technical limit of ! 0.

This discussion is a little abstract, so lets provide a concrete, practical example. The
reader who has taken, or is taking, 06-261 (Fluid Mechanics) will be familiar with the
Navier-Stokes equations (NSE), which describe the motion of a Newtonian fluid: honey,
air, or water, for example. For an incompressible (i.e. constant density) fluid, the NSE
read

r v = 0, (1)

@v
+ v rv = r2 v rp, (2)
@t

where and are the mass density and viscosity of the fluid, respectively, v is the fluid
velocity vector, t is time, and p is the dynamic pressure. The first equation (1) is a math-
ematical statement that the fluid is incompressible. The second equation (2) represents

3
 ,

Figure 1: Definition sketch for scaling analysis of flow of a Newtonian fluid of viscosity
and density at an ambient free-stream velocity U past a body of characteristic size L.

momentum conservation; think of it as Newtons second law for a little (material) element
of fluid: the left-hand side is the acceleration of the element, while the right-hand side
represents the sum of the forces on the element. These forces arise due to viscous stresses,
the first term on the right-hand side, and pressure gradients, the second-term. Writing
the NSE in the compact, vector-invariant form above belies their complexity: they are,
in fact, a coupled system of four nonlinear partial dierential equations in four unknowns
(the three components of the fluid velocity and the pressure). Nevertheless, a lot can be
learned about the Navier-Stokes equations without actually solving them (which is tough
to do, even using a computer).

Lets think about a typical scenario where an object, or particle, is place in a fluid that
is otherwise moving uniformly with a free-stream speed U (figure 1). The object creates
a disturbance to the velocity field in the fluid, which could, in principle, be calculated
by solving the Navier-Stokes equations. Intuitively, one expects the characteristic size of
object (L say) would influence the final result e.g. a tennis ball versus a nanoparticle.
This brings us to scaling, which is an incredibly important tool and a necessary first step
in the successful use of perturbation methods. For example, if the free-stream speed is
U , it is expected that this should be the characteristic magnitude of the velocity field v
everywhere in the fluid. Hence, we can write a dimensionless velocity field v = v/U .
Similarly, the particle size L provides a characteristic length scale; hence, we can define
a dimensionless gradient (which has units of inverse length) as r = Lr. (Note, we
consider particles with only one characteristic length scale, which excludes e.g. slender
rods or flat discs.) Since speed equals distance over time, we define a dimensionless time
t = t/(L/U ). Finally, we define a dimensionless pressure p = p/(U 2 ). Substituting
these dimensionless variables into the NSE yields after a little work (left as an exercise
for the reader, or Ex., for short), the dimensionless NSE

r v = 0, (3)
@v 1 2

+ v r v = r v r p , (4)
@t Re

U L
where Re =
is the Reynolds number. Importantly, through scaling we learn that

4
the free-stream velocity, size of the object, and viscosity and density of the fluid can only
aect the solution to the NSE through the dimensionless group Re. For example, if you
had the solution for an object of size L in a stream of speed U , then the solution for an
object of size 2L in a stream U/2 would be exactly the same, since the Reynolds number
is equal in both cases.

The Reynolds number represents the ratio of inertial (acceleration) to viscous forces on
a fluid element, and it can take a wide range of values. For example, if the particle
is a tennis ball in a stream of air then Re is typically much greater than unity. In
contrast, Re is typically much smaller than unity for a nanoparticle undergoing Brownian
motion in water. The smallness or largeness of Re encountered in many such situations
allows the use of perturbation methods to obtain approximate solutions to the NSE. For
example, perturbation techniques at large Reynolds number (where the small parameter
is = 1/Re) can be used to predict the drag force on streamlined bodies in high-speed
flows, which is relevant to aerodynamics [6]. Low Reynolds number flows (where = Re)
are relevant to the motion of microscopic entities in viscous fluids, such as swimming
micro-organisms, nanoparticles, or proteins [5]; perturbation methods are useful in this
regime too.

The complexity of the NSE dictates that the mathematical steps associated with applying
perturbation methods to obtain perturbative solutions to it is nontrivial. It is wiser to
start our introduction to perturbation methods with a subject that is taught is high
school: namely, quadratic equations.

3 Module 2: Quadratic equations

Learning Objective: Introduce the student to regular perturbation methods.

Student Outcome: The student is able to solve algebraic and transcendental equations
via regular perturbation methods.

We begin exploring perturbation methods through their application to the solution of

quadratic equations. One might ask what the point of this exercise is, given that the
quadratic formula readily furnishes an exact expression for the roots of a quadratic. Ul-
timately, of course, we wish to apply perturbations methods to ordinary and partial dif-
ferential equations that cannot be solved exactly. Many aspects of perturbation methods
relevant to that task can be understood in the familiar context of quadratic equations.

Consider the following quadratic equation due to Hinch [3]

x2 + x 1 = 0, (5)

5
 1.0 0

0.8 -20

0.6 -40
x+ x-
0.4 -60

0.2 -80

0 -100
10-2 10-1 100 101 102 10-2 10-1 100 101 102

Figure 2: The positive x+ and negative x roots of the quadratic equation x2 + x 1=0
as a function of .

where 1 is a constant positive parameter. The exact solution for the roots
r
1 1 2
x= 1+ . (6)
2 4

Denote the root that takes the positive sign in (6) as x+ and other root as x . Figure 2
plots both roots versus . From (6) it is seen that x+ = 1 and x = 1 at = 0, and
x+ ! 0 and x ! as ! 1. Let us consider what happens when is small compared
to unity. In this case, a Taylor series expansion of the square root in (6) yields Ex.:
1 1 1 4
x+ = 1 + 2 + O(6 ), (7)
2 8 128
1 1 2 1 4
x = 1 + + O(6 ). (8)
2 8 128
(9)

The big oh symbol, O, has been introduced in the above formulae. A function f () is
O(n ) as ! 0 if the limit f ()/n is bounded as ! 0. Thus, in (7) the remaining terms
in the expansions for x+ and x are O(6 ). It is also common to say that the remaining
terms are of order 6 . (For a more detailed discussion see [6].)

An exact solution (6) has been obtained and then approximated it for small (7). However,
in most practical problems an exact solution is not available. An alternative approach is
to first approximate the equation (5) itself for small and then find a solution to that
approximate equation. Consider the positive root x+ , which equals unity at = 0. We
pose a perturbation expansion of x+ in integral powers of

x+ () = 1 + x1 + x2 2 + x3 3 + . . . , (10)

where xi are the expansion coefficients and the symbol . . . represents the higher order
terms in the expansion. A series such as (10) is known as an asymptotic expansion.

6
Importantly, such series may be convergent or divergent. However, as noted by Van Dyke
[6], the utility of an asymptotic expansion is that the error is of the order of the first
neglected term (in this case O(4 )) and hence tends to zero as decreases. One might
think that for a fixed one could reduce the error by adding additional terms, but if the
series is divergent eventually adding more terms will increase the error. To proceed, we
substitute the series (10) into quadratic equation (5)

(1 + x1 + x2 2 + x3 3 + . . . )2 + (1 + x1 + x2 2 + x3 3 + . . . ) 1=0 (11)

and collect like powers of on the left-hand side of the equation

(1 1)0 + (2x1 + 1) + (x21 + 2x2 + x1 )2 + (2x1 x2 + 2x3 + x2 )3 + ... = 0, (12)

where, the leading-terms are O(1), or O(0 ) as indicated (12). Now the coefficients at
each order of in the equation are equated

O(1) : 1 1 = 0 (Consistent)
O() : 2x1 + 1 = 0 =) x1 = 12
O(2 ) : x21 + 2x2 + x1 = 0 =) x2 = 18
O(3 ) : 2x1 x2 + 2x3 + x2 = 0 =) x3 = 0. (13)

Notice that the leading order equation was satisfied automatically, since we knew that
the expansion (10) started at unity (x0 = 1, in eect). Satisfyingly, the coefficients x1 ,
x2 , and x3 are identical to the series expansion (7) of the exact solution. To recap,
we have obtained an asymptotic approximation to the root x+ in the limit ! 0, by
solving a series of problems that arise upon substituting the asymptotic expansion into
the governing quadratic equation. We did not solve the governing equation itself. And
the problems that arise at each order of (apart from the leading, or zeroth, order) are
linear equations for the unknown coefficient at that order, which depend on the coefficients
determined at lower orders. In contrast, the original equation is quadratic (i.e. nonlinear).
This structure is typically found in perturbation schemes. Table 1 compares this four-
term asymptotic solution for x+ to the exact solution for dierent . Even at = 1, which
is certainly not small, the asymptotic solution is to within 2% of the exact solution.
Of course there is no guarantee that adding more terms to the asymptotic solution will
improve its accuracy.

There are a couple of important issues that this simple example glosses over. First,
sometimes an asymptotic expansion does not proceed in integral powers of an example
for a quadratic equation is given in [3]. Second, we have obtained an asymptotic solution
for ! 0, but can anything be done for 1? The answer is yes; here the small
parameter is 1/. However, while a naive, or regular, perturbation expansion akin
to (10) succeeds for the positive root x+ it fails for the negative root x . It turns out
that the limit ! 1 (or ! 0) is singular, meaning that the solution to the quadratic

7
 x+ exact x+ asymptotic
0.001 0.999500 0.999500
0.01 0.995012 0.995013
0.1 0.951249 0.951250
0.2 0.904988 0.905000
0.5 0.780776 0.781250
1 0.618033 0.625000
2 0.414214 0.500000

Table 1: Comparison of exact (6) and asymptotic (10) solutions for the positive root of
the quadratic equation (5). Both are quoted to six decimal places.

for = 0 is essentially dierent to the solution in the limit ! 0. Again, singular

perturbation problems for quadratic equations are discussed in [3]. Well return to this
issue for the present quadratic equation in a later module. It is worth mentioning that
the limits Re ! 0 and Re ! 1 of the Navier-Stokes equations are singular too, although
for distinct physical reasons in each case.

Exercises

Ex. Find three terms in the asymptotic expansion for the negative root x .

Ex. Find the three-term asymptotic expansion for the real root of the cubic equation
x3 + x + 8 = 0, in the limit ! 0. Find two-term asymptotic expansions for each of the
complex roots to the equation.

Ex. Use a regular perturbation expansion to find two roots of the equation x2 1 ex = 0,
where is a small parameter. This transcendental equation cannot be solved exactly,
showcasing the power of perturbation methods. Compare your asymptotic solutions to a
numerical solution of the equation. Do any more roots exist?

4 Module 3: Nonlinear reaction in a thin film

Learning Objective: Introduce the student to perturbation methods as a tool to solve

ordinary dierential equations.

Student Outcome: The student is able to scale and non-dimensionalize ordinary dier-
ential equations, and apply regular perturbation methods to solve them.

In this module, regular perturbation theory is applied to the solution of ordinary dier-
ential equations. This builds naturally upon the material covered in the last module:

8
 C = C0 dC/dx = 0

Reservoir Film Wall

r = - kC2

x=0 x=L

Figure 3: Definition sketch for the problem of second-order, irreversible reaction in a

liquid film.

indeed, the reader will notice several features common to the treatment of algebraic and
dierential equations via perturbation methods.

Consider an example from chemical reaction engineering, which appears as an exercise

in [2]: A chemical species undergoes a second-order irreversible reaction in a liquid film
of thickness L (figure 3). The rate of reaction r = kC 2 , where k is a rate constant
(with units of volume per unit time) and C is the concentration of the chemical. The
concentration at the left-hand side of the film (x = 0) is kept at a value of C0 by contacting
the film with a well-stirred reservoir. The film is flanked by a solid wall at x = L, which is
impermeable to the chemical. At steady state, the concentration distribution within the
film is governed by a balance between diusion and the irreversible reaction; a material
balance on the chemical yields the ordinary dierential equation Ex.
d2 C
D = kC 2 , (14)
dx2

which is supplemented by the boundary conditions

dC
C = C0 at x = 0, and D = 0 at x = L, (15)
dx

where the second condition specifies that the flux vanishes at the wall, and D is the
diusivity of the chemical. Our goal is to determine the flux of the chemical, defined as
q = D dC dx
, at the film-reservoir interface, x = 0, under the assumption that the rate of
reaction is slow relative to diusion.

A couple of remarks are in order. First, the small parameter about which to perform
a perturbation expansion is not explicitly defined in the problem statement; instead, a

9
condition regarding the importance of reaction versus diusion is given. However, the
small parameter will become apparent after the problem is made dimensionless. Second,
the governing equation (14) is a nonlinear ordinary dierential equation. There are no
general methods to find exact solutions to this type of equation (although sometimes
solutions do exist in terms of so-called special functions. [4]). Hence, perturbation
methods are especially useful here. Of course, a numerical solution is possible, and we
will compare solutions via perturbation and numerical methods.

We proceed by scaling: the film thickness L provide an obvious choice for the length
scale, as does the reservoir concentration C0 for a concentration scale. Therefore, we
define a dimensionless coordinate = x/L, ranging from zero to unity, and a dimensionless
concentration = C/C0 . In terms of these dimensionless variables, the governing equation
(14) becomes Ex.
d2 kC0 L2
= 2 , (16)
d 2 D

subject to dimensionless boundary conditions

d
= 1 at = 0, and = 0 at = 1. (17)
d

Since and are dimensionless variables, the group kC0 L2 /D appearing in (16) must be
dimensionless also. This group can be rewritten as (L2 /D)/(1/kC0 ), which is a ratio of
time scales: L2 /D is the characteristic time for the chemical to diuse across the film,
and 1/kC0 is a time scale representative of the reaction. The problem stated that the
rate of reaction is slow relative to diusion, which requires that the diusion time L2 /D
is much shorter than the reaction time 1/kC0 . Hence, their ratio (L2 /D)/(1/kC0 ) is a
small parameter, which we denote by the symbol , per usual. Intuitively, in this limit one
expects that the nonlinear reaction only slightly disturbs the concentration profile that
would be established by diusion alone. The power of scaling and nondimensionalization
is evident here; it enables identification of the single dimensionless group, kC0 L2 /D,
that governs entirely the physics of the problem.

To proceed, we pose an asymptotic expansion of the dimensionless concentration (; )

in integral powers of

(; ) = 0 () + 1 () + 2 ()2 + ..., (18)

where the coefficients in the expansion, i (), are functions of that will be found at
each order. The expansion (18) is substituted into the governing equation (16)

d2 2
0 + 1 + 2 2 + ... = 0 + 1 + 2 2 + ... . (19)
d 2

10
Collecting terms at dierent orders of gives
2 2
d 2 0 d 1 2 2 d 2
2
+ 2
0 + 20 1 + ... = 0. (20)
d d d 2

The expansion is also substituted into the boundary conditions. At the reservoir-film
interface
0 + 1 + 2 2 + ... = 1 at = 0, (21)

and at the film-wall contact

d0 d1 d2 2
+ + + = 0 at = 1. (22)
d d d

Equations (20), (21), and (22) yield a boundary-value problem at each order of . At O(1)

d 2 0 d0
= 0, 0 = 1 at = 0, and = 0 at = 1, (23)
d 2 d

which admits the solution Ex.

0 = 1, (24)

corresponding to a uniform concentration of chemical across the film. This result occurs
since at leading order the reaction is infinitely slow compared to diusion; hence, there
is no consumption of chemical within the film. Notice too that the flux of chemical
q = DC 0 d
L d
is zero at all points in the film, not just at the wall, x = 1. Proceeding to
O() yields the system

d 2 1 d1
= 02 , 1 = 0 at = 0, and = 0 at = 1, (25)
d 2 d

where 0 = 1 from (24). The O() concentration, 1 , is forced by the O(1) concentration,
0 . (Recall that a similar situation occurred for the quadratic equation in module 3, cf.
equation (13).) The boundary condition on 1 at = 0 arises since there are no O()
terms on the right-hand side of (21). The solution to (25) is

1
1 = 1 ; (26)
2

evidently, the concentration profile is not constant across the film, due to the O() con-
sumption of the chemical. Hence, there is an O() contribution to the flux q at the

11
reservoir-film interface, x = 0. Rather than calculate it now, let us proceed to the O(2 )
problem
d 2 2 d2
= 2 0 1 , 2 = 0 at = 0, and = 0 at = 1, (27)
d 2 d

which has the solution (using the known formulae for 0 and 1 ) Ex.

1 3 2
2 = +2 . (28)
3 4

Again, a non-uniform concentration profile is found, which is of a more complicated func-

tional form than at the preceding order, as the reaction becomes increasingly prominent.

Finally, using (18), (26), and (28) the flux q = DC 0 d

L d
at = 0 is Ex.

DC0 2 2
q= + O(3 ) . (29)
L 3

To recap, the flux at the reservoir-film interface has been calculated through O(2 ). One
can readily calculate further terms, by continuing the regular perturbation expansion to
higher orders. However, the algebra involved becomes increasingly tedious; the task could
be delegated to a symbolic manipulation software package. (For several success stories on
the use of computers in calculating higher order terms in regular perturbation expansions
see [7].) This process is systemized by noting that the ordinary dierential equation at
O(n ) takes the form
d 2 n
= fnc(n 1 , n 2 , . . . , 0 ), (30)
d 2

where fnc stands for a function of. The left-hand side of the above equation (the
linear operator) is the same at each order in , and the O(n ) concentration, n , is forced
by lower order (known) terms. Such behavior is characteristic of regular perturbation
expansions.

Figure 4 compares the flux q evaluated using the asymptotic formula (29) to that from
a numerical solution of (16) and (17) (using the bvp4c package in Matlab). The two
are in agreement at small , as expected. However, for > 0.75 the asymptotic formula
erroneously predicts the flux to decrease with increasing , becoming negative for > 1.5.
As mentioned above, better agreement could, in principle, be achieved by adding further
terms to the series (29). Nevertheless, the regular perturbation expansion does provide
a valuable check for the numerical solution, not to mention yielding significant physical
insight into the problem.

Finally, at large the numerically-evaluated flux appears to obey a power-law scaling

with . Perturbation methods can be used to determine an analytical approximation for

12
 101

100

q/(DC0/L)
10-1

10-2
10-2 10-1 100 101 102

Figure 4: Flux at the reservoir-film interface, x = 0, for the nonlinear reaction problem,
as a function of kC0 L2 /D. The symbols are from numerical solution of the boundary-
value problem (16) and (17), and the solid line is the asymptotic formula (29) for small
.

the flux in this regime. However, a regular perturbation expansion will not succeed, as
the limit ! 1 is singular. This problem will be the subject of a later module. As an
introduction to singular perturbation methods, we return to quadratic equations.

Exercises

Ex. Continue the analysis above to determine the third-order concentration profile 3 .
Determine the flux q including terms of O(3 ).

Ex. A fluid of viscosity and density flows between two flat, infinite porous plates
located at y = 0 and y = h, where the lower plate moves in the x-direction at constant
speed U , and the upper boundary is at rest (see figure 5). The no-slip condition applies
at the surface of each plate, but there is a vertical cross-flow through the porous plates
at a velocity V ey (where ey is a unit vector in the y direction). From the Navier-Stokes
equations, it can be shown that the governing equation for the velocity component uex
parallel to the walls is
du d2 u
V = 2, (31)
dy dy

where ex is a unit vector in the x direction.

(a) By suitable scaling, show that equation (31) may be written in dimensionless form as
d2 u du
2
+ Re = 0, (32)
dy dy

subject to the boundary conditions u(y = 0) = 1 and u(y = 1) = 0. Here, overbars denote

13
 -V -V -V -V -V
y=h

x U

-V -V -V -V -V

Figure 5: Definition sketch for flow through porous plates.

dimensionless variables and Re = V h/. What is the physical meaning of Re?

(b) Find an exact solution to (32), and plot u for Re = 0, 0.1, and 1.

(c) From (32), derive the first two terms of the asymptotic expansion of the velocity
field Ru(y) for Re 1. Thus, show that the dimensionless flow rate per unit width
1
Q = 0 u(y)dy is
1 1
Q = Re + O(Re2 ). (33)
2 12

Give a physical explanation for why the flow rate decreases with increasing Re. How does
this compare against the flow rate computed using the exact solution found in part (b)?

5 Module 4: Quadratic equations redux

Learning Objective: Introduce the student to the solution of algebraic equations via
singular perturbation expansions.

Student Outcome: The student is able to identify and solve algebraic equations via
singular perturbation methods.

We revisit the quadratic equation

x2 + x 1 = 0. (34)

Asymptotic approximations to both roots of this equation, denoted x+ and x , for small
were obtained in module 3. Here, the task is to do the same for 1. Figure 2 shows
that x+ approaches zero as ! 1, whereas x becomes increasingly negative. At large
it is convenient to rewrite the quadratic equation in terms of 1/, so that (! 0) is
the small parameter as ! 1. Thus, in terms of the quadratic reads
x2 + x = 0. (35)

14
At = 0 this equation becomes x = 0, which from figure 2 corresponds to the limit of the
positive root x+ as ! 1. In fact, a regular expansion of x+ in , namely

x+ () = x0 + x1 + x2 2 + x3 3 + . . . , (36)

yields Ex.
1 1
x+ = 3 + ... = + ... (37)
3

Thus, to leading order, the positive root vanishes linearly with as ! 0.

What happened to the other root, x ? Taking = 0 in (35) turns the governing equation
from a quadratic into a linear equation. Such a reduction of order is a hallmark of a
singular perturbation; the limit ! 0 is singular. A linear equation can admit only one
solution: the root x+ . The other root x seems to have disappeared! What has gone
wrong? The answer is that in taking the limit ! 0 in (35) we naively (or perhaps
unwittingly) assumed that x2 is small compared to x, thus leaving x = 0 as the only
solution. However, x2 can be of the same order as x as ! 0, if x O(1/) (i.e. large).
This is observed in figure 2, where the negative root x appears to be going to (negative)
infinity for small (large ). Therefore, rather than the regular, or naive, expansion (36),
an expansion whose leading order term is O(1/) is posed
1
x() = x 1 + x0 + x1 + ... (38)

Substituting (38) into (35) and collecting powers of yields Ex.

1
[x 1 (x 1 + 1)] + [x0 (2x 1 + 1)]0 + (2x 1 x1 + x1 1) + ... = 0 (39)

The equation at leading order is x 1 (x 1 +1) = 0, which is a quadratic with roots x 1 = 0

and x 1 = 1. The first root, x 1 = 0, corresponds to x+ found from the naive expansion
(36). This makes sense as the naive expansion assumed that x+ is at most O(1) as ! 0;
hence, it cannot posses a term at O(1/). Indeed, with x 1 = 0 it is found that x0 = 0 and
x1 = 1, in agreement with (37). The second root of the leading order equation, x 1 = 1,
is the first term in the expansion for x , the previously missing root. Using x 1 = 1 it
is found that x0 = 0 and x1 = 1; hence, x has the (singular) asymptotic expansion
1 1
x = + ... = + ... (40)

as ! 0 (i.e. ! 1). Table 2 compares this expansion to the exact formula (6); for
> 5 the dierence between the asymptotic and exact results are within one percent.

15
 x exact x asymptotic
1 -1.618034 -2.000000
2 -2.414214 -2.500000
5 -5.192582 -5.200000
10 -10.099020 -10.100000
20 -20.049876 -20.050000
50 -50.019992 -50.020000
100 -100.009999 -100.010000

Table 2: Comparison of exact (6) and asymptotic (40) solutions for the negative root of
the quadratic equation (5) at > 1. Both are quoted to six decimal places.

It is worthwhile recapitulating what has be learned. Equation (35) at = 0 is a linear

equation and hence essentially dierent than the limit of the same equation as ! 0,
which is quadratic. Thus, ! 0 is a singular limit. (At = 0 the equation remains
quadratic; that limit is regular.) Thus, as ! 0 one of the roots, x , begins to diverge
(or escape), ultimately to be lost at = 0. To track the escape of that root, the
quadratic nature of the equation must be retained as ! 0, which can only happen if the
root itself scales as 1/, to leading order i.e. the root is singular as ! 0. Formally
this scaling can be obtained by writing

x () = X (), (41)

where is an as yet unknown exponent and X () = X0 + X1 + X2 2 + ... is a regular

asymptotic series. Substituting (41) into (35) yields

1+ X 2 + X 1
= 0. (42)

Three choices emerge: (i) > 1, (ii) < 1, or (iii) = 1. The first yields the
leading order equation X = 0, which is a linear equation and hence cannot give two
roots. Choice (ii) gives X 2 = 0 at leading order: a double null root. It is only the third
choice, with the leading order equation X 2 + X = 0, that provides two distinct roots
X = 0 and X= 1, corresponding to the leading order terms in x+ and x , respectively.
The choice = 1 is referred to as a distinguished limit.

This module has provided an introduction to singular perturbation methods, through

a simple application to quadratic equations. However, the concepts learned can be ap-
plied to the solution of more complicated problems, involving ordinary (and even partial)
dierential equations, as shown next.

Exercises

16
Ex. Consider the cubic equation
x3 x + 1 = 0, (43)

where is a small positive parameter.

(a) Use a regular perturbation expansion to deduce that one (regular) root to the equation
is x() = 1 + + O(2 ).

(b) To find the other two (singular) roots, introduce the rescaling x() = X() and show
that to obtain the distinguished limit in (43) requires = 12 .

(c) Expanding X() as perturbation series in powers of 1/2 (why is it correct to expand
in powers of 1/2 , as opposed to powers of ?), show that the singular roots are
1 1
x() = + O(1/2 ). (44)
1/2 2

6 Module 5: Boundary layers & matched asymptotic

expansions

Learning Objective: The student is introduced to singular dierential equations and

their solution via matched asymptotic expansions.

Student Outcome: The student is able to identify and solve singular dierential equa-
tions using matched asymptotic expansions.

In this module, the method of matched asymptotic expansions is introduced as a means of

obtaining approximate solutions to singularly perturbed dierential equations. Consider
the following boundary-value problem, due to Bender and Orszag [1]
d2 y dy
2
+ (1 + ) + y = 0, y(0) = 0, and y(1) = 1, (45)
dx dx

where is a positive constant. The dierential equation in (45) is linear, second order,
and has constant coefficients. (Bender and Orszag [1] also consider the general case of
second order linear equations with variable coefficients.) Therefore, the solution to (45)
is easily found to be Ex.
e x e x/
y(x) = 1 . (46)
e e 1/

Figure 6 plots y(x) for = 0.01, 0.1, and 0.2. As decreases, y(x) rises rapidly near
x = 0 and then gradually decreases until the condition y(1) = 1 is met. The region of

17
 3.0

2.5 0.01

2.0
y(x) 0.1
1.5
= 0.2
1.0

0.5

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
x

Figure 6: Plot of exact solution to the boundary-value problem (45), for = 0.01, 0.1,
and 0.2. Notice the formation of a boundary layer near x = 0 as decreases.

rapid variation is known as a boundary layer, whose width, evidently, decreases as

approaches zero.

An asymptotic approximation to y(x) in the limit ! 0 is sought. As a first attempt,

the regular perturbation expansion

y(x; ) = y0 (x) + y1 (x) + y2 (x)2 + ... (47)

is substituted into (45), yielding the following boundary value-problem for the leading
order term y0 (x):
dy0
+ y0 = 0, y0 (0) = 0, and y0 (1) = 1. (48)
dx

The general solution is y0 (x) = Ae x , where A is an integration constant. However, no

choice of A will satisfy the two boundary conditions. What went wrong? Notice that
y0 (x) obeys a first order dierential equation, while the original equation (48) is second
order. Hence, the regular perturbation expansion resulted in a reduction of order, akin to
that encountered for the quadratic equation in module 5. This is a clear signal that the
limit ! 0 is singular. The error in posing the regular expansion (47) is assuming that it
is valid uniformly throughout the interval 0 x 1, which implies that the highest-order
term in (45), d2 y/dx2 , is always small relative to the other two terms.

In fact, the highest-order term must be retained to describe the rapid variation of y(x) in
the boundary layer near x = 0. To that end, an inner, or boundary layer, coordinate
X is introduced via the rescaling X = x/ ( is a constant to be found), such that X =
O(1) as ! 0. Think of this coordinate transformation as a mathematical magnifying
glass, zooming in on the boundary layer; the strength of the magnification is specified
by the exponent . Hence, as ! 0 there are two regions: (i) the inner region, where
X = O(1) [i.e. x = O( )] as ! 0; and (ii) an outer region, where x = O(1). Separate

18
asymptotic expansions hold in each region, and these expansions are required to match in
a overlap region corresponding to X ! 1 and x ! 0. This is the essence of the method
of matched asymptotic expansions.

The solution changes gradually in the outer region, where x = O(1) . Hence, the
governing equation for the solution, yout (say), in this region is found by setting = 0 in
(45), viz.
dyout
+ yout = 0. (49)
dx

The solution is yout = Ae x , as found before. However, now it is realized that this
solution is not uniformly valid throughout 0 x 1, rather it holds for x 1.
Thus, applying the boundary condition at x = 1 (which is in the outer region) reveals
A = e. The asymptotic approximation in the outer region is therefore
yout = e1 x , (50)

which, as x ! 0, must match to the inner solution. To analyze the inner region, the
rescaling X = x/ is applied to (45), yielding Ex.
2
1 2 dyin dyin
+ + yin = 0, (51)
dX 2 dX

where yin denotes the solution in the boundary layer, which is required to satisfy the
(inner) boundary condition y(0) = 0. The exponent is found by seeking a distinguished
limit to (51) that retains the second derivative term (which was subdominant in the outer
region). This requires = 1; the boundary layer is O() in size, therefore, and yin satisfies
d2 yin dyin
+ = 0. (52)
dX 2 dX

Solving (51) subject to y(0) = 0 yields Ex.

X
yin = B(1 e ), (53)

in which B is an integration constant. Matching of the inner (53) and outer (50) asymp-
totic expansions determines B. The matching is accomplished by requiring that the limit
of yin as X ! 1 equals the limit of yout as x ! 0. Or, in words: the outer limit of the
inner solution equals the inner limit of the outer solution. Therefore B0 = e and
yin = e e1 X
. (54)

Furthermore, the outer (50) and inner (54) expansions can be combined to yield a com-
posite expansion, ycom say, that is uniformly valid approximation to y(x) over the entire

19
 3.0

2.5

2.0
y(x)
1.5

1.0

0.5

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
x

Figure 7: Comparison of exact solution to the boundary-value problem (45) (circles)

against the uniformly valid asymptotic approximation (56) (black line) for = 0.1. Also
shown are the inner (red line) and outer (blue line) solutions.

domain 0 x 1. One method of constructing ycom is given by the formula

ycom = yin + yout yovp , (55)

where yovp is the overlap term common to both the inner and outer expansions (i.e. in
the limits X ! 1 and x ! 0). In this example. ycom = e; hence,

ycom = e1 x
e1 x/
. (56)

The composite expression is made of two parts: (i) e1 x , which varies slowly on the scale
x O(1), originating from the outer region; and (ii) e1 x/ , varying rapidly on the scale
of the boundary layer, x O(). Figure 7 shows that the exact solution for y(x) and
composite expansion ycom are in excellent agreement for = 0.1.

This example has illustrated how the method of matched asymptotic expansions is used
to gain asymptotic approximations to singular dierential equations. An exact solution
to the full dierential equation was available (46); however, generally this is not the
case (especially for nonlinear dierential equations). Matched asymptotic expansions are
particularly useful in the absence of an exact solution, since the inner and outer approx-
imations to the full dierential equation can often be solved exactly, thereby yielding an
analytical approximation to the solution. Also note that the inner and outer solutions yin
and yout , respectively, are actually the leading order terms in asymptotic expansions for
the solutions in the inner and outer regions, respectively. That is, properly,

yin (X; ) = y0,in (X) + y1,in (X) + y2,in (X)2 + . . . , (57)

yout (x; ) = y0,out (x) + y1,out (x) + y2,out (x)2 + . . . , (58)

20
where y0,in (X) and y0,out (x) are the leading-order solutions determined above. In fact,
the higher-order terms in yout (x; ) are identically zero in this example Ex.; however,
yin (X; ) does contain higher-order contributions. When several terms in the inner and
outer expansion have been determined, matching can be organized using an intermediate
variable or Van Dykes matching rule, see [3] for details.

Exercises

Ex. The following boundary-value problem is due to Friedrichs (and discussed at length
in [3])
d2 y dy
2+ + ex = 0, y(0) = 0, and y(1) = 1, (59)
dx dx

where is a small positive parameter.

(a) Determine the exact solution to (59), and plot it for = 0.2, 0.1, and 0.01. Notice
that an inner region (or boundary layer) exists near x = 0 in which the solution varies
sharply.

(b) Determine leading order solutions for y in the inner and outer regions. Use them to
construct a uniformly valid composite solution. Compare the inner, outer, composite, and
exact solutions in a plot akin to figure 7.

7 Module 6: Nonlinear reaction in a thin film at large

reaction rates

Learning Objective: Introduce students to further aspects of matched asymptotic ex-

pansions (e.g. boundary layers of non-integral width), and complete the thin- film-reaction
problem from module 4.

Student Outcome: The student is able to identify boundary layers width non-integral
width and construct matched asymptotic examples for such.

The nonlinear reaction problem introduced in module 4 will be revisited here, with the
goal of determining an asymptotic approximation for the flux of chemical q at large di-
mensionless reaction rate . From figure 4 it appears that the flux q follows a power-law
scaling in this regime. The limit ! 1 is singular and can be tackled by matched asymp-
totic expansions. Recall the boundary-value problem for the dimensionless concentration
,
d2 d
2
= 2 , = 1 at = 0, and = 0 at = 1, (60)
d d

21
where = kC0 L2 /D is the dimensionless reaction rate. Our interest is for 1: to
proceed, it is convenient to define = 1/ 1. We pose a regular perturbation expansion
of the concentration in powers of :

(; ) = 0 () + 1 () + 2 ()2 + ... . (61)

Substituting (61) into (60) yields the leading-order problem Ex.

d0
02 = 0, 0 = 1 at = 0, and = 0 at = 1. (62)
d

The solution to the governing equation in (62) is 0 = 0, which satisfies the boundary
condition at = 1, but not at = 0. (In hindsight, one could have anticipated trouble
since the original dierential equation (60) has been reduced to algebraic equation.) The
result 0 = 0 implies that the reaction proceeds so quickly that all of the film is depleted
of the chemical. However, to satisfy the boundary condition at the reservoir-film interface
there must exist a boundary layer near = 0, in which diusion balances reaction, no
matter how fast the latter. Thus, 0 = 0 is recognized as the leading-order solution for
the concentration in an outer region (where O(1)), where diusion is unimportant;
i.e. out = 0, to leading order.

The first task is to determine the size of the boundary layer; let z = / be an inner
coordinate, such that z = O(1) as ! 0. Introducing this rescaling into the dierential
equation in (60) yields Ex.
d2 in
1 2 2 = in . (63)
dz

Only = 12 yields a distinguished limit to the above equation; thus, the boundary layer is
O(1/2 ) in size. This should be contrasted against the (much thinner) O() boundary layer
encountered in the previous module. Recall that the dimensionless flux at the reservoir-
film interface q = DC 0 d
| . Writing q in terms of the inner variable gives
L d =0

DC0 1 d
q= . (64)
L 1/2 dz z=0

The above equation reveals that q 1/1/2 (or 1/2 ) at large reaction rates. Thus we have
determined how the flux scales with reaction rate, without solving any dierential equa-
tions! This important point highlights the power of scaling arguments and distinguished
limits. Moreover, if the scaling has been performed correctly, one is guaranteed that the
d
number dz |z=0 is O(1). Hence, stopping at this stage by simply asserting it equal to unity
would provide a reasonable engineering estimate for the flux q. However, let us proceed
to determine in .

22
The boundary-value problem for in is
d2 in 2
= in , in = 1 at z = 0, and in ! 0 as z ! 1. (65)
dz 2

One cannot directly apply the boundary condition at x = 1; rather, the inner solution
matches to outer solution (out = 0), which is the second condition in (65). A first integral
of the dierential equation in (65) is obtained by multiplying both sides of the equation
by din /dz, yielding Ex. r
din 2 3
= + A, (66)
d 3 in

where A is an integration constant. To determine A, note that in and din /d approach

zero as z ! 1 (Ex. why?). Thus A = 0. Furthermore, the negative root in (66) is
taken as the concentration must decrease away from the reservoir-film interface (i.e. as z
increases). Therefore, another integration gives Ex.
1 z
p = p + B. (67)
in 6

Applying the boundary condition in = 1 at z = 0 gives the integration constant B = 1.

Thus, 2
z
in (z) = 1 + p . (68)
6
q
d
From (68), dz |z=0 = 23 = 0.816..., which is, indeed, close to unity. Hence, the asymptotic
approximation for the dimensionless flux at large (small ) is
r r
DC0 2 DC0 2
q= = . (69)
L 3 L 3

Figure 8 shows that (69) compares favorably against the numerically-calculated flux.
Together with the analysis for small in module 4, perturbation methods provide the
flux q for essentially any value of . More important, they yield significant insight into
the dominant physics at small and large reaction rates, which would be difficult (if not
impossible) to gain from a numerical solution.

A uniformly-valid leading order approximation for the concentration can be constructed

from the inner in and outer out solutions using the composite expansion formula (55).
In this case, out = 0 (and hence ovp = 0); hence, the composite expansion com is simply
equal to in . That is, 2

com () = 1 + p . (70)
6

23
 101

100

q/(DC0/L)
10-1

10-2
10-2 10-1 100 101 102

Figure 8: Flux at the reservoir-film interface, x = 0, for the nonlinear reaction problem,
as a function of kC0 L2 /D. The symbols are from numerical solution of the boundary-
value problem (16) and (17), the broken line is the asymptotic formula (29) for small ,
and the solid line is the asymptotic formula for large (69).

Figure 9 compares com against the numerically-calculated concentration profile at = 10

and = 100, showing good agreement. Finally, note that (69) represents the leading order
term in an asymptotic expansion for the flux at large . A natural question is: What is
the next term in the expansion? This requires us to determine higher order terms in the
asymptotic approximations in the inner and outer regions. Note that the limit of the
inner solution (68) as z ! 1 is
6
in (z) ! 2 , (71)
z

which equals 6/ 2 when written in terms of the outer variable. Hence, whilst there is no
O(1) term in in as z ! 1 (and hence the matching with the leading-order outer solution
out = 0 is satisfied) there is an O() contribution. This implies that the next (and non
zero) term in the outer region is O(). That term is required to match with (71) as ! 0.
The task of determining this term is left as an exercise.

Exercises

Ex. Here the porous plates exercise in module 3 is revisited, with the goal of calculating
the flow rate Q at large Reynolds number Re.

(a) From the exact solution to (32), plot u for Re = 5, 10, and 100. Notice that at Re 1
there exists a boundary layer in which viscous forces balance the cross-flow (inertia).

(b) Introduce a scaled coordinate Y = Re (y + ), where and are numbers whose

value you need to determine. Show that the leading order solution for the velocity in the
boundary layer is u = exp( Rey).

24
 1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
= 10
0.2
0.1
= 100
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

Figure 9: Chemical concentration versus distance along film . The solid lines are the
uniformly valid asymptotic approximation (70), and the symbols are from a numerical
solution. Results for = 10 and = 100 are shown.

(c) Determine the flow rate Q for Re 1. Sketch Q as a function of Re.

8 Acknowledgments

I gratefully acknowledge support from the Eberly Center for Teaching Excellence at
Carnegie Mellon University through a Wimmer Faculty Fellowship. I thank Dr. Ruth
Poproski from the Eberly Center for critical reading of the document, which greatly im-
proved its quality.

References
[1] C. M. Bender and S. A. Orszag, Advanced Mathematical Methods for Scientists
and Engineers: Asymptotic Methods and Perturbation Theory (Springer-Verlag, New
York, NY, 1999).

[2] W. M. Deen, Analysis of Transport Phenomena (Oxford University Press, Oxford,

U. K., 1998).

[3] E. J. Hinch, Perturbation Methods (Cambridge University Press, Cambridge, U. K.,

1991).

[4] N. N. Lebedev, Special Functions and Their Applications (Dover, New York, NY,
1972).

[5] E. M. Purcell, Life at low Reynolds number, Am. J. Phys. 1, pp. 3-11 (1977).

25
[6] M. Van Dyke, Peturbation Methods in Fluid Mechanics (Parabolic Press, Stanford,
CA, 1975).

[7] M. Van Dyke, Computer extension of perturbation series in fluid mechanics, SIAM
J. Appl. Math. 28, pp. 720-734 (1975).

26