You are on page 1of 88

Asymptotic Analysis

Dr. Thomas Götz


Technomathematics Group
Department of Mathematics
University of Kaiserslautern
Kaiserslautern, Germany
e–mail: goetz@mathematik.uni-kl.de

July 22, 2005


Contents

References 3

1 Introduction 3
1.1 Dimensional Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Basic Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 General Theoretical Results . . . . . . . . . . . . . . . . . . . . . 9
1.4 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2 Asymptotic Solution of Algebraic Equations 14


2.1 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

3 Asymptotic Expansion of Integrals 21


3.1 Watson’s Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Laplace Integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3 Stationary Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.4 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

4 Regular Perturbation of ODEs and PDEs 26


4.1 Error Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2 Regular Perturbation of PDEs . . . . . . . . . . . . . . . . . . . . 28
4.3 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

5 Singular Perturbation of ODEs 34


5.1 Outer Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
5.2 Inner Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.3 Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.4 Composite Expansion . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.5 A Formal Definition of Boundary Layers . . . . . . . . . . . . . . 41

1
5.6 Multiple Boundary Layers . . . . . . . . . . . . . . . . . . . . . . 42
5.7 Where is the Boundary Layer? . . . . . . . . . . . . . . . . . . . . 44
5.8 A Singular Perturbed PDE . . . . . . . . . . . . . . . . . . . . . . 46
5.9 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

6 Lindstedt Expansions 48
6.1 Generalized Asymptotic Expansions . . . . . . . . . . . . . . . . . 49
6.2 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

7 Multiple Scales 55
7.1 Lindstedt Revisited . . . . . . . . . . . . . . . . . . . . . . . . . . 55
7.2 Idea of Multiple Scales . . . . . . . . . . . . . . . . . . . . . . . . 57
7.3 Some General Remarks concerning Multiple Scales . . . . . . . . . 59
7.4 Slowly Varying Coefficients . . . . . . . . . . . . . . . . . . . . . . 60
7.5 Multiple Scales & Boundary Layers . . . . . . . . . . . . . . . . . 63
7.6 Multiple Scale Analysis of Numerical Methods . . . . . . . . . . . 64
7.7 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

8 Homogenization 67
8.1 Weak Limits and Rapidly Oscillating Functions . . . . . . . . . . 67
8.2 Introductory Example . . . . . . . . . . . . . . . . . . . . . . . . 70
8.3 New Concepts of Convergence . . . . . . . . . . . . . . . . . . . . 73
8.4 Flow in Porous Media . . . . . . . . . . . . . . . . . . . . . . . . 77

9 Perturbation Methods in Fluid Mechanics 80


9.1 Prandtl’s Boundary Layer Theory . . . . . . . . . . . . . . . . . . 80
9.2 Stokes Flow past Cylinders and Spheres . . . . . . . . . . . . . . 82
9.3 Derivation of Euler Equations from Kinetic Theory . . . . . . . . 85

2
References
[1] E. J. Hinch ”Perturbation Methods”, Cambridge ’92

[2] M. H. Holmes ”Introduction to Perturbation Methods”, Springer ’95

[3] M. van Dyke ”Perturbation Methods in Fluid Mechanics”, Parabolic Press ’75

[4] R. E. O’Malley ”Singular Perturbation Methods for ODEs”, Springer ’91

[5] J. A. Murdock ”Perturbations. Theory and Methods”, SIAM ’99

[6] J. Struckmeier ”Tutorial on Asymptotic Analysis I”, AGTM–Report 108, ’94

[7] G. I. Barenblatt ”Scaling”, Cambridge ’03

[8] J. Kevorkian, J. D. Cole ”Perturbation Methods in Applied Mathematics”,


Springer ’80

1 Introduction

What is asymptotic analysis?

Philosophy: Many problems in applied mathematics contain a small parameter


ε, e.g. mass of a ball, viscosity of air, size of pores in a rock, etc. Setting this
parameter to zero (ε = 0) leads to an — ofter easier — problem, the unperturbed
or reduced problem. The solution of the unperturbed problem might be used as
an approximation of the perturbed problem (ε 6= 0). We then seek a series of
corrections to this initial approximation and hope that a few (one or two) of
these additional terms provide an useful approximate solution to the perturbed
problem.
The idea behind asymptotics:

Perturbed Problem ε→0 - Unperturbed Problem


P (u; ε) = 0 P (u; 0) = 0
6Convergence?
”Easy” to solve
Accuracy? ?
Asymptotic Expansion 
Formal Expansion 0th order Approximation
u(ε) ∼ u0 + u1 g1 (ε) u0

3
r
ϕ
v(r, t)
z
PSfrag replacements

B0

Figure 1.1: Sketch of the considered geometry.

Typical questions that arise are:

• Where is the small parameter? What does small mean?


These questions can be answered by a proper dimensional analysis and
scaling.

• Are P0 and Pε of the same kind?


This leads to regular or singular perturbations.

• How to get the corrections? Can we choose a suitable ansatz?

• Can we get error estimates for uε ? What about convergence?

1.1 Dimensional Analysis

Consider an unidirectional Poiseulle flow through a pipe of radius R in the pres-


ence of a transversal magnetic field, see Figure 1.1. In cylindrical coordinates,
the equations of motion read as
 2 
∂v ∂p ∂ v 1 ∂v
ρ =− +η 2
+ − σB02 v (1.1)
∂t ∂z ∂r r ∂r

where p is the pressure and η the viscosity of the fluid. The boundary conditions
are given by
v(t, R) = 0, v(t, 0) finite (1.2)

4
Dividing by ρ and introducing the parameters (the dimensions are given in brack-
ets, L: length, T: time)
1 ∂p
a= [LT −2 ]
ρ ∂z
η
µ= [L2 T −1 ]
ρ
σB02
n= [T −1 ]
ρ
Eqn. (1.1) reads as
 2 
∂v ∂ v 1 ∂v
= −a + µ + − nv (1.3)
∂t ∂r 2 r ∂r
In order to make (1.3) dimensionless, we choose reference scales R for the length,
t0 for time and V for the velocity. Next, we introduce the dimensionless variables
r = Rρ, t = t0 τ and v = V u. Now
 
V ∂u V µ ∂ 2 u 1 ∂u
= −a + 2 + +Vnu
t0 ∂τ R ∂ρ2 ρ ∂ρ
 
∂u at0 µt0 ∂ 2 u 1 ∂u
=− + 2 + + nt0 u
∂τ V R ∂ρ2 ρ ∂ρ

Choosing V = R/t0 and t0 = 1/n, our problem contains the dimensionless pa-
rameters a/(Rn2 ) and µ/(R2 n).
The following ”theorem” contains a more general statement concerning dimen-
sionless parameters:
”Theorem”: Mathematical models of applied problems can we written in di-
mensionless form. The dimensionless parameters and the reference scales can be
choosen as products of powers of the original parameters.
Of course this ”theorem” is by far too unprecise, but one can make it more precise
and even prove this statement [see Barenblatt].
Here, we will just use the above statement and apply it to our MHD–problem.
This problem contains the parameters a, µ, n and R. The dimension of a pa-
rameter pi can written as ml1,i sl2,i , hence we can identify the dimension of pi by
the vector (l1,i , l2,i ). Dimensionless parameters are identified with the zero–vector
(0, 0). If a parameter α is written as
n
Y
α= p k ak (1.4)
k=1

5
then the dimension of α equals
n
Y P P
[α] = mak l1,k sak l2,k = m ak l1,k
s ak l2,k
(1.5)
k=1

Compling the exponents l1,k , l2,k in a matrix, we get


 
l1,1 l1,2 . . . l1,k
A= (1.6)
l2,1 l2,2 . . . l2,k

and the dimension of α can be computed by A · (a1 , a2 , . . . ak )T . If we seek a


dimensionless parameter, then this matrix–vector–product has to be equal to the
zero–vector.
Therefore, the number N of dimensionless parameters is given by N = dim ker A.
In our example, the matrix A reads as
 
1 2 0 1
A= (1.7)
−2 −1 −1 0
and N = dim ker A = 2. A basis of the kernel of A is given by the vectors
(1, 0, −2, 1) and (0, 1, 1, −2) leading to the dimensionless parameters a/(Rn2 )
and µ/(R2 n).
Of course, one can also choose another basis of the kernel, e.g. we will later on
work with (−1, 2, 0, −3) and (−1, 0, 2, 1) giving rise to µ2 /(aR3 ) and n2 R/a.

1.2 Basic Definitions

One important notation are the Landau or order symbols. Let f, g : R 7→ Y ,


where (Y, |·|) is a normed space.
Definition 1.1. We introduce the following notations (”big–O” and ”little–o”):
f = O(g) for ε → ε0
if there exist constants k and ε1 (independent of ε) such that
|f (ε)| ≤ k |g(ε)| for all ε0 ≤ ε < ε1 .
The notation f = O(g) means, that f and g are of (roughly) the same size or
order.
If for every δ > 0, there exists an ε2 such that
|f (ε)| ≤ δ |g(ε)| for all ε0 ≤ ε < ε2
then we write f = o(g). This means, that f is much smaller than g, i.e. f  g.

6
f (ε) f (ε)
Remark 1.1. If lim exists and is finite, then f = O(g). If lim = 0,
ε→ε0 g(ε) ε→ε0 g(ε)
then f = o(g).
Since e−1/ε = o(εα ) for all α and ε → 0, we say that e−1/ε is transcendentally
small compared to the power functions εα .
Example 1.1. Consider f = ε2 and g1 = ε, g2 = −3ε2 + 5ε6 . Then f = o(g1 )
and f = O(g2 ) for ε → 0.
Definition 1.2. We call a function ϕ an asymptotic approximation to f for
ε → 0, if
f = ϕ + o(ϕ) .
We write f ∼ ϕ.

The idea behind this definition is that ϕ serves as an approximation to f , if the


error is of higher order (smaller), than the approximating function.
E.g. given f (ε) = ε2 + ε5 , then ϕ(ε) = ε2 is an asymptotic approximation, but
ψ = ε2 /2 not.
Example 1.2. Let f = sin ε. To construct an asymptotic approximation, we use
Taylor
1 1 5
f (ε) = ε − ε3 + ε cos ξ
6 120
for some ξ ∈ [0, ε]. We can conclude, that f ∼ ε, f ∼ ε − ε3 /6 and f ∼ ε + 2ε2 .
However, this is the least accurate asymptotic approximation, but asymptotic
approximations say little about accuracy.
Let f = x + ε−x/ε where 0 < x < 1 is fixed. Then f ∼ x. But how well does this
approximation work for the entire function f (x; ε) = x + ε−x/ε on 0 ≤ x ≤ 1?
Quite good if we are away for x = 0. For x close to 0, the approximation fails,
since f (0; ε) = 1 ∀ε > 0.
In this case the approximation is not uniformly valid.
Definition 1.3. A set {Φ1 , Φ2 , . . .} of functions is called an asymptotic sequence
or well–ordered, if Φn = o(Φm ) for n > m (and ε → 0).
Given an asymptotic sequence (Φm )m∈N . A function f has an asymptotic expan-
sion to n terms wrt. to the sequence Φm , if
k
X
f= aj Φj + o(Φk ) for k = 1 . . . n and ε → 0
j=1

where the coefficients aj are independent of ε. In this case we write f ∼ a1 Φ1 +


. . . an Φn . The functions Φk are called scale, gauge or basis functions.

7
Example 1.3. Frequently we will use the power functions Φk = εαk , where
αk < αk+1 as scale functions. An asymptotic expansion using power functions as
scale functions is also called Poincaré expansion.
Wrt to this scale, we get X
sin ε ∼ ak ε2k+1
k=0n

with ak = (−1)k /(2k + 1)!. Note, that this asymptotic expansion is also conver-
gent.
Let’s consider the error function
Z x
2 2
erf(x) = √ e−t dt
π 0

How to compute it, how to approximate this function?


2 2 P
First idea: We expand the integrand e−t into its Taylor series e−t = (−t2 )k /k!
and perform the integration for each summand. This yields the approximation
n
2x X (−x2 )k
En (x) = √
π k=0 k! (2k + 1)

and En (x) → erf(x) for n → ∞.


How good is this approximation e.g. at x = 4?
Maple yields erf(4) = 0.9999999846 . . . . Using our approximation En (4) we get

n 2 10 30 40 50 60
En (4) 96 37922 125 1.027 1.000000592 0.9999999847

Although the series En (x) converges to erf(x) for n → ∞, we need a large number
of terms (∼ 60) to get a suitable approximation for large values of x. Can we do
better?

Second idea: We try to construct an asymptotic expansion for x  1 and write


Z x Z ∞
2 −t2 2 2
erf(x) = √ e dt = 1 − √ e−t dt
π 0 π x
2
The integrand e−t contributes significantly only for t close to x, therefore intro-
duce u = t2 − x2
Z ∞ Z ∞ −x2 −u 
−t2 e e u −1/2
e dt = 1+ 2 du
x 0 2x x

8
∞  
α
X α
Next we use the generalized binomial (1 + y) = y k for α = −1/2. Note,
k
  k=0
−1/2 Γ(1/2 + k)
that = (−1)k √ . Choosing n ∈ N fix, we get
k πk!
Z ∞ 2 n 2 n
e−x −u X Γ(1/2 + k)  u k e−x X
√ e − 2 du = √ Γ(1/2 + k) (−x)−2k
0 2 πx k=0
k! x 2 πx k=0

and
2 n
e−x X
En (x) = 1 − Γ(1/2 + k) (−x)−2k
πx k=0
2  
e−x 1 3 15
=1− √ 1− 2 + 2 4 − 6 ± ...
πx 2x 2x 8x
How good is this approximation for x = 4?

erf(4) = 0.9999999846 . . .
E2 (4) = 0.9999999846 . . .

However, note that En (x) does not converge to erf(x) for n → ∞, since
r
k 1 · 3 . . . (2k − 1)
→∞ for k → ∞
2k
Remark 1.2. Asymptotic expansions need not to be convergent with an increasing
number of terms!
(pointwise) convergence Fix x, let n → ∞
asymptotic expansion Fix n, let ε → 0

Asymptotic expansions can be added and multiplied, if they are orderd in a


special way. However, in general we cannot differentiate asymptotic expansions.
But if f (x; ε) ∼ a1 (x)Φ1 (ε) + a2 (x)Φ(ε) + . . . and ∂x f (x; ε) ∼ b1 (x)Φ1 (ε) +
b2 (x)Φ(ε) + . . . , then bk = ∂x ak .
Asymptotic expansions can be integrated
Z b Z b Z b
f (x; ε) dx ∼ a1 (x) dx Φ1 (ε) + a2 (x) dx Φ2 (ε)
a a a

1.3 General Theoretical Results

To obtain some general theoretical results concerning the validity of asymptotic


expansions we consider the following abstract situation:

9
Let (B, k·k) be a Banach–space and (Y, |·|) be a normed space. Let F : B ×
[0, ε1 ] 7→ Y be a mapping. Then we wish to solve the problem

F (u, ε) = 0 (1.8)

To prove the existence of solutions for sufficiently small values of ε we need some
preliminary results.
The first lemma states, that a small perturbation of an bounded invertible map
still has a bounded inverse.

Lemma 1.1. Let A and D be linear maps from B to Y . Assume A has a bounded
inverse, i.e. kA−1 yk ≤ K |y| ∀y ∈ Y . Let δ < 1.
Then for 0 ≤ ε < δ/ kA−1 Dk, the map A + εD has a bounded inverse and

(A + εD)−1 y ≤ K
|y| ∀y ∈ Y .
1−δ

Proof. Due to the assumption we have kA−1 k ≤ K and kεA−1 Dk ≤ δ < 1.


Rewriting the equation (A + εD)u = y as the fixpoint–equation u = A−1 y +
εA−1 Du and using the contractivity of εA−1 D, we can apply Banach’s fixpoint
theorem and deduce the invertability of A+εD. The estimate of the norm follows
from

K ≥ A−1 (A + εD)(A + εD)−1 ≥ I + εA−1 D · (A + εD)−1

≥ kIk − ε A−1 D (A + εD)−1

≥ (1 − δ) · (A + εD)−1 .

The next theorem states, that if the residual of an equation is sufficiently small
at a point ũ, we can find a solution u close to ũ. Of course the map needs to
be quite regular, e.g. (1.9) is something like a local Lipschitz–condition for the
second derivative (if it exists).

Theorem 1.2. Let G : B 7→ Y be Frechet–differentiable at a point ũ ∈ B. The


derivative DG(ũ) is assumed have a bounded inverse, i.e. kDG(ũ) −1 yk ≤ K |y|
for all y ∈ Y and furthermore let P (v) = G(ũ + v) − G(ũ) − DG(ũ) v satisfy

|P (v1 ) − P (v2 )| ≤ Lδ kv1 − v2 k , for kv1 k , kv2 k ≤ δ . (1.9)

If the residual ρ = G(ũ) is bounded by |ρ| ≤ 1/(4K 2 L), then there exists a unique
solution u0 of G(u0 ) = 0 in a δ–neighborhood of ũ, with δ = 1/(2KL). This
solution satisfies the estimate ku0 − ũk ≤ 2K |ρ|.

10
Proof. Let w = u0 − ũ. Then G(u0 ) = 0 is equivalent to the problem G(ũ + w) −
G(ũ) = −ρ. With the definition of the map P this is again equivalent to solving
DG(ũ)w = −ρ − P (w). This equation can be rewritten as the fixpoint–problem

w = Φ(w) := −DG(ũ)−1 (ρ + P (w)) .

To applyBanach’s fixpoint
theorem, we show, that Φ : U 7→ U is a contraction
for U = w : kwk ≤ δ ⊂ B. The estimate
 
1 2
kΦ(w)k ≤ K |ρ + P (w)| ≤ K + Lδ = δ
4K 2 L
shows, that Φ : U 7→ U holds. The contractivity follows from
1
kΦ(w1 ) − Φ(w2 )k ≤ K |P (w1 ) − P (w2 )| ≤ KLδ kw1 − w2 k = kw1 − w2 k .
2
Therefore the Banach fixpoint theorem implies the existence of a unique solution
w ∈ U and
 1
kwk = kΦ(w)k ≤ K (|ρ| + |P (w)|) ≤ K |ρ| + Lδ kwk = K |ρ| + kwk
2
Hence we have the estimate ku0 − ũk ≤ 2K |ρ|.
Example 1.4. To illustrate the statements of Thm. 1.2 consider the scalar case
G ∈ C 2 (R, R). Then the bound on the inverse of the derivative implies |G0 (ũ)| ≥
1/K and (1.9) holds, if kG00 k∞ ≤ L
Assume, that ρ = G(ũ) > 0 and G0 (ũ) < 0 and consider the Taylor–expansion
G(ũ + v) = G(ũ) + G0 (ũ)v + 21 G00 (ξ)v 2 , where 0 ≤ ξ ≤ v. Then we get as a
worst–case estimate ρ ≤ − K1 v − Lδv. For δ = δ = 2KL 1
it follows, that v ≤ 2Kρ
which is precisely one of the statements.
Theorem 1.3. Assume there exists a solution u0 ∈ B of the reduced problem
F (u0 , 0) = 0. Furthermore we assume that F is in a neighborhood of (u0 , 0)
(N+1)–times Frechet–differentiable with respect to both arguments (N ≥ 1) and
the derivative Du F (u0 , 0) (with respect to u at point (u0 , 0)) has a bounded inverse.
Then F (u; ε) admits in a neighborhood of u0 the asymptotic expansion
N
X
F (u; ε) = Fk (u)εk + O(εN +1 ) .
k=0

The formal asymptotic expansion


n
X
(n)
u = uk ε k for n ≤ N
k=0

11
is well–defined, where the coefficients uk are given as solutions of
F0 (u0 ) = F (u0 ; ε)
Du F (u0 ; 0) uk = −Gk (u0 , . . . , uk−1 ) ,
for some Gk . For ε small enough, the problem (1.8) has a unique solution u(ε)
in a ε–independent neighborhood of u0 and

u(ε) − u(n) = O(εn+1 ) for n ≤ N

Proof. The well–definedness of the expansion u(n) is a consequence of Taylor’s–


theorem (generalized to Banach–spaces).
To show the existence of the solution of the perturbed problem and the estimates
for the expansion,
we try n+1
to apply Thm. 1.2. The residual of the expansion
(n)
satisfies F (u , ε) = O(ε ). Due to the differentiability of F , the assump-

tion (1.9) of Thm. 1.2 holds with an ε–independent constant L. Furthermore we
have DF (u(n) , ε) = DF (u0 , 0)+εC, where the linear map C is uniformly bounded
in ε. Using Lemma 1.1, we get the existence and uniform boundedness of the
inverse DFu (u(n) , ε). Therefore all the assumptions of Thm. 1.2 are satisfied.

1.4 Questions

Question 1. Consider the free fall of an object (ball) under the influence of
gravity and air drag. The governing ODE reads as
AρcD
mẍ = −mg − |ẋ| ẋ, x(0) = h0 , ẋ(0) = 0 (1.10)
2
where m is the mass of the object, g the gravitational acceleration, A the cross–
section of the object, ρ the density of air, cD the dimensionless drag coefficient
and h0 the initial height of the object.
Perform a dimensional analysis and identify dimensionless parameters.
−2
Answer: The parameters of problem (1.10) are: [g] = LT  , [h0 ] = L and
−1 1 −1 1
[k] = [Aρcd /(2m)] = L . Therefore the matrix A = −2 0 0 and dim ker A = 1.
The dimensionless parameter kh lies in ker p A.
Introducing the scaling x = h0 ξ and t = h0 /gτ , problem (1.10) reads in dimen-
sionless form as

d2 ξ dξ dξ dξ
2
= −1 + kh 0
, ξ(0) = 1, (0) = 0 .
dτ dτ dτ

Question 2. We can extend the above situation by also taking into account,
that gravity depends on the distance from the center of the earth. In this case,
the ODE reads as
gR2
ẍ = − , x(0) = h0 , ẋ(0) = v0 (1.11)
(x + R)2

12
where R is the radius of the earth and v0 the inital velocity of the object. In this
case, we skipped the air drag, since the density of the air also changes with the
height according to ρ(h) = ρ0 e−(x+R)/k , where k is a constant with the dimension
of a length. Of course this additional term makes things much more complicated.
Again, perform a dimensional analysis and identify dimensionless parameters.
Answer: The parameters of problem (1.11) have the following dimensions: [g] =
LT −2 , [R] = L, [h0 ] = L and [v0 ] = LT −1 . Therefore the matrix A reads as
1 1 1 1
( −2 = 2 with [h0 /R], [v02 /(gh0 )] ∈ ker A. Introduction the
0 0 −1 ) and dim ker A p
scalings x = h0 ξ and t = h0 /gτ we get the non–dimensional problem
 −2
d2 ξ h0 dξ v0
=− 1+ ξ , ξ(0) = 1, (0) = √ .
dτ 2 R dτ gh0

Question 3. Suppose f (ε) ∼ a0 ϕ0 (ε) + a1 ϕ1 (ε) + . . . and g(ε) ∼ b0 ϕ0 (ε) +


b1 ϕ1 (ε) + . . . for ε → 0 are asymptotic expansions w.r.t. the asymptotic sequence
ϕ0 , ϕ1 , . . . , where a0 , b0 6= 0.
Show that f g ∼ a0 b0 ϕ20 . What are possibilities for the next term ?
Suppose ϕi ϕj = ϕi+j . Compute an asymptotic expansion for f g.
Answer: Since ϕ1 = o(ϕ0 ), we get that ϕ0 ϕ1 = o(ϕ20 ) and ϕ21 = o(ϕ20 ). Hence the
leading term of f g is given by a0 b0 ϕ20 . For the higher terms, we get into trouble
with ϕ0 ϕ2 and ϕ21 . Which is smaller? If ϕ0 = 1, ϕ1 = ε and ϕ2 = ε2 ln ε, then
ϕ21 = o(ϕ0 ϕ2 ), but for ϕ0 = 1, ϕ1 = ε and ϕ2 = ε3 we have ϕ0 ϕ2 = o(ϕ21 ).
2 2
Question 4. Find an asymptotic expansion of f (x) = e−1/x cos(e1/x ) for x → 0.
Consider the derivative
of the expansion
and compare it with f 0 (x).
2 2 2
Answer: Since e−1/x cos(e1/x ) ≤ e−1/x as x → 0, we get f ∼ 0. For the
 
2 2 2
derivative we obtain f 0 (x) = e−1/x cos(e1/x ) + sin(e1/x ) x−3 and the term
2
x−3 sin(e1/x ) does not converge to 0 as x → 0. Hence f 0 ∼ 0 cannot be valid.

Question 5. Consider the exponential integral


Z ∞ −t
e
Ei(x) = dt (1.12)
x t
and derive an asymptotic expansion for x  1.
Hint: Perform repeatedly an integration by parts.
What about the convergence for the resulting series for x fixed and n (number of
terms) increasing?
What about the accuracy
 for n fixed and x increasing? 
−x 1 1 2! N −1 (n − 1)!
Answer: Ei(x) ∼ e − + + · · · + (−1)
x x2 x3 xn

13
2 Asymptotic Solution of Algebraic Equations

What does it mean to solve an equation by a perturbation method? Suppose


we wish to solve x2 − 3.99x + 3.02 = 0. We introduce ε = 0.01 and rewrite the
problem.

Example 2.1. Given the perturbed problem

P (x; ε) = x2 + (ε − 4)x + (3 + 2ε) = 0 (2.1)

The unperturbed problem P (x, 0) = x2 − 4x + 3 has the solutions x1 = 3 and


x2 = 1. We seek an asymptotic expansion of the solutions of Pε . Using power
functions as basis, we porpose the following ansatz

x1 (ε) ∼ 3 + a1 εα1 + a2 εα2 (2.2)

where α1 < α2 . Plugging this ansatz into Pε yields

P (x; ε) = 5ε + 2a1 εα1 + 2a2 εα2 + a1 ε1+α1 + a21 ε2α1


+ 2a1 a2 εα1 +α2 + a22 ε2α2 + a2 ε1+α2 ∼ 0

Using α1 < α2 we compare equal powers of ε.


To cancel the 5ε, we choose α1 = 1 and get
5
0 = 5ε + 2a1 ε a1 = −
2
5 25
0 = 2a2 εα2 − ε2 + ε2 α2 = 2, a2 = −15/8
2 4
The remaining terms are of order O(ε3 ). Hence we get for the first root x1 the
asymptotic expansion
5 15
x1 (ε) = 3 − ε − ε2 + O(ε3 )
2 8

Similarly for the second root

3 15
x2 (ε) = 1 + ε + ε2 + O(ε3 )
2 8
Coming back to the starting point (ε = 0.01), we compute the approximation
x1 (0.01) = 2.9748125 and the exact solution equals x1 = 2.974808 . . . .

Why does this asymptotic method work? Implicit function theorem

14
Theorem 2.1. Let ϕ : Rm × R 7→ Rm with continuous first partial derivatives,
∂ϕ
ϕ(x0 , ε0 ) = 0 and det (x0 , ε0 ) 6= 0. Then there exists a neighborhood U of
∂x
ε0 and a unique continuous function x : U 7→ Rm such that x(ε0 ) = x0 and
ϕ(x(ε), ε) = 0 forall ε ∈ U . The function x has continuous derivatives
 −1
∂x ∂ϕ ∂ϕ
=− (x(ε), ε) (x(ε), ε)
∂ε ∂x ∂ε

If ϕ ∈ C r , then also x ∈ C r .

But life is not always that easy:

Example 2.2. Consider

P (x; ε) = (x − 2)2 (x − 3) + ε = 0

Then P (x; 0) has the roots x1 = 3 and x2 = 3. We seek an asymptotic expansion


of the x2 (ε)
x(ε) ∼ 2 + a1 εα1 + a2 εα2
This leads to

P (x; ε) ∼ ε − a21 ε2α1 + a31 ε3α1 − 2a1 a2 εα1 +α2 + . . .

To cancel ε, we get α1 = 1/2 and a21 = 1. This equation has the two solutions
a1 = ±1, so we get a bifurcation of the root located close to 2. In the next order
we get α2 = 1 and a2 = 1/2. So the root close to 2 has the asymptotic expansion
√ ε
x(ε) ∼ 2 ± ε+
2

In all the above examples P0 and Pε are of the same type, i.e. equations of the
same order. Of course this need not to be always the case and this point gets
particularly interesting, when dealing with differential equations. Here’s just a
first example of so–called singular perturbations

Example 2.3. Consider

P (x; ε) = εx2 + x + 1 = 0

The reduced problem reads as P (x, 0) = x + 1, which is linear with the single
solution x1 = −1. An asymptotic expansion of this root yields

x1 (ε) ∼ −1 − ε − 2ε2

15
But what about the second root of P (x; ε)? Where is it? What happens to the
second root as ε → 0?
Again we try with the asymptotic ansatz

x2 (ε) ∼ a0 εα0 + a1 εα1 + a2 εα2

where α0 < α1 < α2 . We get

0 ∼ a20 ε2α0 +1 + a0 εα0 + 1 + 2a0 a1 εα0 +α1 +1 + a1 εα1


+ 2a0 a2 εα0 +α2 +1 + a21 ε2α1 +1 + a2 εα2 + . . . (2.3)

Choosing α0 = 0 leads to the root close to x = −1.


The other possibility is to choose α0 such that 2α0 + 1 = α0 , i.e. α0 = −1 and
then a0 has to satisfy a20 + a0 = 0. The trivial solution a0 does not give any
information about the asymptotic expansion, therefore we choose a1 = −1. The
next orders lead to the expansion
1
x2 (ε) ∼ − + 1 + ε
ε

So far, power functions worked as scale functions. However, this also need not to
be always the case.

Example 2.4. Consider the transcendental equation

x + tanh x = u (2.4)

for u  1. Of course, there exists a solution.


Since u is large, we introduce ε = 1/u  1 and get εx + ε tanh x = 1. Since
tanh x ∈ (−1, 1), the second term is always O(ε) and therefore we can conclude,
that x ∼ 1/ε.
Next we rescale x = y/ε and get the equation y + ε tanh(y/ε) = 1. For large
arguments, we can use the following expansion of tanh z:

1 − e−2z
tanh z = −2z
= (1 − v) (1 + v)−1 where v = e−2z
1+e
∼ (1 − v)(1 − v + v 2 − v 3 ± . . . )
∼ 1 − 2e−2z + 2e−4z − 2e−6z

Using the expansion y ∼ 1 + a1 εα1 with α1 > 0 we get


 
α1
1 α1 −1

a1 ε + ε 1 − 2 exp −2 ε + a1 ε ∼0
| {z }
o(1)

16
Therefore we conclude α1 = 1 and a1 = −1 leading to the expansion y ∼ 1 − ε
and x ∼ u − 1.
What about the next term? Let’s try again with y ∼ 1 − ε + a2 εα2 with α2 > 1.
Then
  
a2 εα2 − 2ε exp −2 1ε − 1 + a2 εα2 −1 = a2 εα2 − 2e2 εe−2/ε exp −2a2 εα2 −1 ∼ 0

Obviously, this is not possible! What went wrong? Perhaps choosing εα2 was the
wrong choice as scale function. Lets try with an undetermined scale function g2
instead, where g2 = o(ε), i.e. g2 /ε = o(1). Then we have to satisfy
   
1 g2 g2 
0 ∼ a2 g2 − 2ε exp −2 − 1 + a2 = a2 g2 − 2e2 εe−2/ε exp −2a2
ε ε ε
∼ a2 g2 − 2e2 ε e−2/ε

We conclude, that g2 = εe−2/ε and a2 = 2e2 . Putting things together, we get the
expansion
x ∼ u − 1 + 2e−2(u−1)

The last example is already quite tricky.

Example 2.5. Consider


ε xex = 1
for ε  1. Again this equation has a unique solution x ∈ R+ . Why?
Taking the logarithm of the equation, we get

ln ε + ln x + x = 0

If ε  1, we get − ln ε  1 and therefore x  1 and x  ln x. This suggests to


choose x ∼ − ln ε as a starting point for the asymptotic expansion. But how to
continue?
Let’s replace the algebraic equation by a fixpoint problem, which we try to solve
using an iteration procedure.

xn+1 = − ln ε − ln xn , with x0 = − ln ε

or equivalently

xn+1 = ln 1ε − ln xn , with x0 = ln 1ε

17
Then
x1 = ln 1ε − ln ln 1ε

x2 = ln 1ε − ln ln 1ε − ln ln 1ε
 1
ln ln
= ln 1ε − ln ln 1ε − ln 1 + ε
ln 1ε
1 1
ln ln 1ε
∼ ln ε − ln ln ε +
ln 1ε
This sequence of scale functions is not at all obvious. However, the fixpoint
iteration generates this sequence automatically.
The equation xex = y can be solved in terms of the LambertW –function, i.e. y =
W (x). With the above result we get the following asymptotic expansion
ln ln z
W (z) ∼ ln z − ln ln z + , for z  1 .
ln z
The LambertW–function has many applications, e.g. for the so–called infinite
...
power towers h(z) = z z . The infinite power tower is recursively defined as
h0 (z) = z, hn+1 (z) = z hn (z) and h(z) = lim hn (z)
n→∞

for z ∈ C. To compute the value of h(z) (it exists for Re(z) ∈ (e−e , e1/e )), we
take logarithms and obtain
1
ln h + h ln =1 ⇐⇒ ln (−h · ln z) − h ln z = ln (− ln z)
z
Setting x = −h ln z and ε = −1/ ln(z) we are back to the starting equation.

2.1 Questions

Question 6. The free fall of an object under the action of gravity and air drag
is modelled by the following ODE:
ẍ = −1 + εẋ2 , x(0) = 1, ẋ(0) = 0 (2.5)
where the small parameter ε models the air drag. The solution of that ODE can
be obtained in closed form

ln cosh(t ε)
x(t; ε) = 1 − (2.6)
ε
Derive an asymptotic expansion of the time t, when the object hits the ground,
i.e. x(t, ε) = 0. √ √
√ 2 2 2
Answer: t ∼ 2 + ε+ ε
6 120
18
Question 7. In celestial mechanics, the following problem needs to be solved

ε sin(E) = E − M (2.7)

where ε  1 is the eccentricity of the orbit, M = n(t − T ) and E the eccentric


anomaly.
Determine the first three terms in an asmptotic expansion for E.
If everything went okay, you can compare your result with the following series
solution (Bessel, 1824)

X 1
E =M +2 Jn (nε) sin(nM ) (2.8)
n=1
n

ε2 ε3
Answer: E ∼ M + ε sin M + sin(2M ) + (3 sin(3M ) − sin M )
2 8
Question 8. Consider the functions y(x) and Y (x) for x ∈ [0, 1], defined through
Z 2
ds
=1−x and Y (x) = 3e−x/ε (2.9)
y artanh(s − 1)

On the interval [0, 1], there exist a point x such that y(x) = Y (x).
Derive the first terms of the asymptotic expansion of x for ε  1.
0
10
Z 1
ds
Answer: Introduce F (y) = −1
10

y−1 artanh s
and rewrite (2.9) as F (y) = 1 − ε ln(y/3). −2
10

Then y ∼ y0 + ε artanh(y0 − 1) y0 /3,


x

−3
10

where F (y0 ) = 1, i.e. y0 ≈ 1.303. Hence


y0 artanh(y0 − 1) −4
10

x ∼ −ε ln − ε2 Numerical

3 3y0 −5
10
−4
10
−3
10
−2
10
−1
10
Asymptotic
0
10
ε

Question 9. Let A and D ∈ Rn×n . Suppose A is symmetric and has n distinct


eigenvalues. Assume D to be positive definite.
Derive a two–term expansion of the eigenvalues of the perturbed matrix A + εD.
This is also known as the Rayleigh–Schrödinger series for the eigenvalues.
Answer: Using the ansatz λj (ε) ∼ λj + ελ1j + ε2 λ2j and xj (ε) ∼ xj + εx1j + ε2 x2j
for the eigenvalues λj and eigenvectors xj we get in leading order

Axj = λj xj

and for the first order corrections the equations

Ax1j + Dxj = λj x1j + λ1j xj

19
Scalar multiplication by xj and using (xj , xj ) = 1, A symmetric and λj 6= λk for
j 6= k, we get

λ1j = (Dxj , xj )

and after scalar mutliplication by xk , where k 6= j

X (Dxj , xk )
x1j = xk
k6=j
λj − λ k

modulo the normalization. Analogously, one can derive for the second order
correction λj
X (Dxj , xk ) (Dxk , xj )
λ2j =
k6=j
λj − λ k

and a rather lengthy expression for x2j .

20
3 Asymptotic Expansion of Integrals

3.1 Watson’s Lemma

Consider the integral


Z b
I(x) = f (t)e−xt dt, (b > 0) (3.1)
0

for x  1. Clearly the major contribution to the integrand — assuming f behaves


nicely — comes from t = 0. Replacing f by the first n terms of its Taylor series
around 0 and using
Z b b Z Z
k −xt tk e−xt k b k−1 −xt k ∞ k−1 −xt k!
t e dt = − + t e dt ∼ t e dx = k+1
0 x 0 x 0
x 0 x
we get
Z b
f0 f1 fn
I(x) ∼ (f0 + f1 t + · · · + fn /n! tn ) e−xt ∼ + 2 + · · · + n+1 .
0 x x x

In the following theorem, we will extend that result a bit


Theorem 3.1. Suppose f ∈ C([0, b]) with the asymptotic expansion
n
X
f (t) ∼ tα ak tβk for t → 0+
k=0

where α > −1 and β > 0 such that the integral is bounded near t = 0. If b = ∞,
we also assume that f (t) = o (ect ) for some c > 0 as t → ∞, to guarantee the
boundedness of the integral for large t. Then
Z b n
−xt
X ak Γ(α + βk + 1)
I(x) = f (t)e dt ∼ , for x → ∞. (3.2)
0 k=0
xα+βk+1

where Γ(x) is the√gamma–function. Note, that Γ(n + 1) = n! for n ∈ N and


Γ(1) = 1, Γ( 12 ) = π and Γ(x + 1) = x Γ(x) for x ∈ R \ {0, −1, −2, . . .}.

3.2 Laplace Integrals

Consider the following generalization of the previous case


Z b
I(x) = f (t)e−x g(t) dt, for x  1 . (3.3)
a

21
Assume that g attains an isolated minimum at t0 ∈ (a, b). Then the value of the
integrand close to t0 will contribute most to the value of the integral. Therefore,
let’s replace f and g by their Taylor series at t0

f (t) ∼ f0 + f1 (t − t0 ) + . . . fn (t − t0 )n /n!
g(t) ∼ g0 + g2 (t − t0 )2 /2

Then n Z b−t0
−xg0
X fk 2 /2
I(x) ∼ e sk e−xg2 s ds
k=0
k! a−t0

Defining
Z b−t0
k −xg2 s2 /2 1 k−1 −xg2 s2 /2 b−t0 k − 1
Ek (x) = s e ds = − s e + Ek−2
a−t0 xg2 a−t0 xg2
r
k−1 2π
Ek (x) ∼ Ek−2 and E0 = , E1 = 0
xg2 xg2

2π (2k − 1)!
we get E2k = k−1 (xg2 )−(k+1/2) and
2 (k − 1)!
[n/2] √ [n/2] √
X 2π(2k − 1)! f 2k
X 2πf2k
I(x) ∼ e−xg0 k−1 (2k)! (k − 1)!
(xg 2 ) −(k+1/2)
= e −xg0
k k!
(xg2 )−(k+1/2)
k=0
2 k=0
2
√ −xg0
 
2πe f2
∼ √ f0 + +... (3.4)
xg2 2xg2

Example 3.1. We try to find an expansion of the Gamma function defined by


Z ∞ Z ∞
−t x
Γ(x + 1) = e t dt = e−t+x ln t dt
0 0

for x  1. We make the substitution t = xs and get


Z ∞
x+1
Γ(x + 1) = x e−x(s−ln s) ds
0

In the setting of Laplace integrals, as defined above, we have f (s) = 1 and


g(s) = s − ln s with an isolated minimum at s = 1. Hence, we plug f0 = 1, g0 = 1
and g2 = g 00 (1) = 1 into (3.4) and get Stirling’s formula

Γ(x + 1) ∼ 2πx xx e−x (3.5)

22
3.3 Stationary Phase

Consider the following generalization of the Laplace integral


Z b
I(x) = f (t)eixϕ(t) dt, for x  1 (3.6)
a

where a, b, x and the functions f and ϕ are real. Now, the term eixϕ(t) is purely
oscillatory and hence we cannot exploit the asymptotic decay as in the previous
section. However, since for x → ∞, the integrand oscillates more and more
rapidly, we may expect cancellations of the positive and negative parts.
The following Riemann–Lebesgue lemma contains some first information about
the behaviour of the above integral.

Theorem 3.2 (Riemann–Lebesgue Lemma). Let f ∈ L1 ([a, b]) and ϕ ∈


C 1 ([a, b]), but ϕ(t) not constant on any subinterval of [a, b]. Then
Z b
f (t)eixϕ(t) dt → 0, as x → ∞ .
a

Assume ϕ0 (t) =6 0 on [a, b]. Then we can perform an integration by parts and
obtain
Z b b Z b  
ixϕ(t) 1 f (t) ixϕ(t) 1 d f
I(x) = f (t)e dt = 0
e − 0
eixϕ(t) dt
a ix ϕ (t) ix dt ϕ
a |a {z }
∼o(1) due to Riemann–Lebesgue
b
1 f (t) ixϕ(t)
∼ e (3.7)
ix ϕ0 (t)
a

If ϕ has a stationary point, i.e. ϕ0 (t) = 0 for some t ∈ [a, b], the integration by
parts fails to be valid. What can we do in this case?
Assume ϕ0 (c) = 0 for some c ∈ (a, b) and we have the following approximations
around c: ϕ ∼ ϕ(c) + ϕ00 (c) · (t − c)2 /2 and f ∼ f (c). Then
Z b
00 2
I(x) ∼ f (c) eixϕ(c)+ixϕ (c)·(t−c) /2 dt
a
√ Z
f (c) 2eixϕ(c) ∞ i sgn(ϕ00 (c))s2
∼ p e ds
x |ϕ00 (c)| −∞
s
00 2π
∼ f (c) ei[xϕ(c)+sgn(ϕ (c))π/4] (3.8)
x |ϕ00 (c)|

23
Example 3.2. Consider the Airy function defined on the negative real axis by
Z  3 
1 ∞ t
Ai(−x) = cos − xt dt, for x  1.
π 0 3

Rewriting the cosine as real part of the complex exponential, we have to √


de-
3
termine the stationary points of t /3 − xt. From the two solutions t = ± x
only the positive one is inside
√ the domain of integration and we introduce the
transformation t = (1 + s) x. Then
√ Z ∞
x
eix ·(s /3+s −2/3) ds
3/2 3 2
Ai(−x) = Re
π −1

In the above setting, we have f = 1, ϕ(s) = s3 /3 + s2 − 2/3 with the stationary


point c = 0 and ϕ(0) = −2/3, ϕ00 (0) = 2 and we introduce y = x3/2 . Hence
√ r
x i[−2y/3+π/4]
 π
Ai(−x) ∼ Re e
π y
 
1 2 3/2 π
∼ √ 1/4 cos x −
πx 3 4

3.4 Questions

Question 10. Derive an asymptotic expansion of the zeroth order modified Bessel
function of the 2nd kind defined by
Z ∞
K0 (x) = (t2 − 1)−1/2 e−xt dt, for x  1.
1

N
−x
X (Γ(1/2 + k))2
Answer: K0 (x) ∼ e (−1)k .
k=0
2k+1/2 Γ(1/2) k! xk+1/2

Question 11. Consider the integral


Z b
I(x) = f (t)e−x g(t) dt, for x  1
a

where g is continuous on [a, b] and attains its global minimum at the boundary,
say a. Show, that
f (a)
I(x) ∼ 0 e−xg(a) . (3.9)
xg (a)

24
Question 12. Consider the generalized Laplace integral
Z b
I(x) = f (t)eix ϕ(t) dt, for x  1
a

where ϕ is continuous on [a, b] and has a stationary point at the boundary of the
interval, say ϕ0 (a) = 0. Show, that
r
i[xϕ(a)+sgn(ϕ00 (a))π/4] π
I(x) ∼ f (a) e . (3.10)
2x |ϕ00 (a)|

Question 13. Derive an asymptotic expansion of the zeroth order Bessel function
of the first kind defined by
Z
1 π
J0 (x) = cos (x sin θ) dθ, for x  1.
π 0
r 
2 π
Answer: J0 (x) ∼ cos x − .
πx 4

25
4 Regular Perturbation of ODEs and PDEs
Example 4.1. Let’s consider the MHD–example from the beginning of the lec-
ture. The stationary version of (1.3) reads in the case of a small magnetic field
n = ε  1 as
r2 u r2
r 2 u00 + ru0 − ε = , u(1) = 0, and u0 (0) = 0 due to symmetry. (4.1)
µ µ
We propose a regular asymptotic expansion of the solution u
u(r; ε) ∼ u0 (r) + εu1 (r) + ε2 u2 (r) (4.2)
and plug this into (4.1). Comparing equal powers of ε we obtain
r2
O(ε0 ) : r 2 u000 + ru00 −
=0 u0 (1) = 0 u00 (0) = 0 (4.3a)
µ
r 2 u0
O(ε1 ) : r 2 u001 + ru01 − =0 u1 (1) = 0 u01 (0) = 0 (4.3b)
µ
r 2 u1
O(ε2 ) : r 2 u002 + ru02 − =0 u2 (1) = 0 u02 (0) = 0 (4.3c)
µ
with the solutions
1 2 
u0 (r) = r −1

1 
u1 (r) = r 4 − 4r 2 + 3
64µ
1 
u2 (r) = r 6 − 9r 4 + 27r 2 − 19
2304µ
In terms of the modified Bessel functions of the first kind, we can also get the
solution of (4.1) in closed form
  p  
1 I 0 r ε/µ
u(r; ε) =  p  − 1 (4.4)
ε I ε/µ 0

0
exact
0th order
1st order
−0.5 2nd order
The figure shows a comparison of
−1
the exact solution and the asymp-
totic expansions up to second order
u

−1.5 for ε = 0.1 and µ = 0.1. The sec-


ond order is almost indistinguishable
−2
from the exact one.
−2.5
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
r

26
In general we wish to solve a problem of the form

u0ε = f (t, uε ; ε), uε (0) = u0

where the unperturbed problem u0 = f (t, u), u(0) = u0 is assumed to have some
solution. To obtain the existence solutions for ε  1 and to guarantee the validity
of a regular expansion, we use the following trick.
We introduce a new dependent variable v, which satisfies v 0 = 0, v(0) = ε and
consider the extended differential equation
       
d u f (t, u, v) u u0
= F (t, u, v) = , (0) = . (4.5)
dt v 0 v ε

Using standard results from ODEs [see Walter] concerning the differentiability
with respect to the initial condition, yields the validity of the Poincaré expansion.
Of course, we can also utilize Thm. 1.3 to get the existence of solution and some
error estimates. However, in special cases case we are able to get better estimates.

4.1 Error Estimates

Consider the nearly linear system

u̇ = Au + ε f (t, u, ε), u(0) = α (4.6)

where ε  1, u : [0, T ] → Rn , A ∈ Rn×n , α ∈ Rn and f : [0, T ] × Rn × [0, ε0 ] → Rn


is Lipschitz–cont. wrt. u.
We seek an asymptotic expansion of the form u(t; ε) ∼ u0 (t) + εu1 (t). Then the
coefficient functions satisfy

u˙0 = Au0 u0 (0) = α u0 (t) = eAt α


Z t
u˙1 = Au1 + f (t, u0 , ε) u1 (0) = 0 u1 (t) = eA(t−s) f (s, eAs α, ε) ds
0

Now, we seek an estimate for the remainder

R(t, ε) = u(t; ε) − (u0 (t) + εu1 (t)) .

To do this, we need some auxiliary results.


Lemma 4.1 (Running away inequality). Let r ∈ C 1 (R, Rn ). Then krk is
right–differentiable for all t, i.e.
kr(t + τ )k − kr(t)k
lim
τ →0+ τ

27
and differentiable for all t where r(t) 6= 0. The running away inequality

d d r(t)
kr(t)k ≤
dt dt

holds for all t, where the right–derivative is used when needed.

Lemma 4.2 (Gronwall inequality, differential version). If ϕ0 ≤ αϕ + β and


ϕ(0) = ϕ0 , then
β αt 
ϕ(t) ≤ ϕ0 eαt + e −1 .
α

With these two results we can prove the following

Theorem 4.3. Consider (4.6) for t ∈ [0, T ], i.e. for bounded time t. Then

LM 
kR(t; ε)k ≤ ε2 e(kAk+εL)t − 1 = O(ε2 ) ,
kAk + εL

where L is the L–constant of f and M = max ku1 (t)k < ∞.


0≤t≤T

Proof. The remainder satisfies


dR
= AR + ε [f (t, u0 + εu1 + R, ε) − f (t, u0 , ε)] , R(0) = 0
dt

hence using the running away inequality

d
kRk ≤ kAk kRk + εL · kεu1 + Rk ≤ (kAk + εL) · kRk + ε2 LM .
dt

Therefore using Gronwall

LM 
kRk ≤ ε2 e(kAk+εL)t − 1 .
kAk + εL

4.2 Regular Perturbation of PDEs

Example 4.2. Consider a ”near” sphere with a surface given by

R = R(θ; ε) = 1 + εP2 (cos θ) (4.7)

28
in spherical coordinates, where P2 (x) = 21 (3x2 −1) is a Legendre polynomial. The
(gravitational/electrical) potential outside this object is given by the equations

∆u = 0, for r ≥ R(θ; ε) (4.8a)


u = 1, at r = R(θ; ε) (4.8b)
u = 0, for r → ∞ . (4.8c)

In spherical coordinates the Laplacian of a ϕ–indendent function reads as


   
1 ∂ 2 ∂u ∂2u ∂u
∆u = 2 r + 2 + cot θ .
r ∂r ∂r ∂θ ∂θ
Here, we face a new problem: The small parameter ε is hidden in the boundary of
the domain. With varying ε, the domain, in which have to solve the problem (4.8),
changes!
Let’s seek an asymptotic expansion in terms of powers of ε, i.e.

u(r, θ; ε) ∼ u0 (r, θ) + εu1 (r, θ) + ε2 u2 (r, θ)

where we have assume the symmetry with respect to the ϕ–angle.


Expanding the boundary condition in a Taylor series yields

∂u (εP2 )2 ∂ 2 u
1 = u|r=1+εP2 = u|r=1 + εP2 + +...
∂r r=1 2 ∂r 2 r=1

and hence
 
∂u0
∼ u0 (1, θ) + ε u1 (1, θ) + P2 (cos θ) (1, θ)
∂r
 2

2 ∂u1 1 2 ∂ u0
+ ε u2 (1, θ) + P2 (cos θ) (1, θ) + P2 (cos θ) (1, θ) .
∂r 2 ∂r 2
Collecting terms of equal powers of ε, we get in zeroth order the problem

∆u0 = 0, u0 (1, θ) = 1, u0 (r = ∞) = 0

with the solution


1
u0 (r, θ) =
r
For the first order term, we get

∆u1 = 0, u1 (1, θ) = P2 (cos θ), u1 (∞, θ) = 0

with the solution (seek u1 (r, θ) = r α P2 (cos θ))


P2 (cos θ)
u1 (r, θ) = .
r3
29
The second order is determined by

∆u2 = 0, u1 (1, θ) = P2 (cos θ)2 , u2 (∞, θ) = 0 .


36
Using the identity P2 (x)2 = 35 P4 (x) + 74 P2 (x) + 25 P0 (x), the solution can be
computed
35 P4 (cos θ) 4 P2 (cos θ) 2 1
u2 (r, θ) = + + .
36 r5 7 r3 5r
Example 4.3. Consider the transient heat equation in some domain Ω ⊂ Rn ,
n≥1
∂u
ρcp +∇·F =0
∂t
where u is the temperature, ρ the density and cp the specific heat of the material.
The heat flux F can be split into two parts

F = −kC ∇u + FR .
| {z } |{z}
Conduction Radiation
R
The radiative flux FR = S 2 I(x, s) s ds depends on the intensity I(x, s) of energy
radiated in direction s. To model the intensity, we consider the radiative transfer
equation (RTE)
s · ∇x I(x, s) = n2 κB(u) − κI(x, s)
| {z } | {z }
Emission Absoprtion
σ 4
where B(u) = π
u
is the Planck function with the refractive index n > 1, the
Stefan–Boltzmann constant σ and κ > 0 is the absoprtion coefficient. Integrating
the RTE from a point y along the direction s to the point x leads to
Z kx−yk
−κkx−yk 2
I(x, s) = I(y, s)e + κn B(u(x − ξ))e−κξ dξ .
0

In the case of a large absoption coefficient κ  1 (optical thick case), the major
contribution of the integral comes from the lower boundary (see Section 3.1 on
Watson’s lemma) and an asymptotic expansion of the intensity yields

n2 ∂B
I(x, s) ∼ n2 B(u(x)) − s · ∇u
κ ∂u

and hence for the radiative flux

16n2 σ 3
FR (x) ∼ − u ∇u .
| 3κ{z }
:=kR

30
This approximation is known as the Rosseland approximation and kR is called
the radiative conductivity. Plugging the Rosseland approximation into the heat
equation, we get
  
∂u 16n2 σ 3
ρcp − ∇ · kC + u ∇u = 0
∂t 3κ
or in dimensionless form for κ  1
∂u   
= ∇ · 1 + εu3 ∇u + initial and boundary conditions. (4.9)
∂t
To solve this non–linear parabolic equation approximately, we propose a Poincaré
expansion for the solution u ∼ u0 + εu1 . Comparing coefficients, we get
∂u0
− ∆u0 = 0 (4.10a)
∂t
∂u1  
− ∆u1 = ∇ u30 ∇u0 . (4.10b)
∂t

Let’s consider Eqn. (4.9) resp. (4.10) for (t, x) ∈ R+ × R such that we get rid of
the boundary conditions and assume that a time t = 0, we have the Heaviside–
function u(0, x) = H(x) as initial condition. Then the solution of the zeroth–order
equation (4.10a) is given by
  
1 x
u0 (t, x) = 1 + erf √ .
2 2 t
For the first–order equation (4.10b), we have to solve the inhomogeneous heat
equation
∂u1 
− ∆u1 = ∇ u30 ∇u0 , u1 (0, x) = 0 ,
∂t
and according to classical theory for PDE’s the solution is given by
Z tZ
∇ (u3 ∇u0 ) −(x−y)2 /4(t−s)
u1 (t, x) = p 0 e dy ds .
0 R 2π(t − s)
t= 1.0000

The following graphics shows a numeri- 2

1.8
cal comparison of the zeroth–order ap- 1.6

proximation u0 (’. . . ’), the first–order 1.4

u0 + εu1 (’– –’) and the exact solution u 1.2

(’—’) for the above problem on the finite


u

0.8

interval x ∈ [0, 10] with the initial condi- 0.6

tion u(0, x) = 2 H(x − 5), the boundary 0.4


zeroth order
conditions u(t, 0) = 0, u(t, 10) = 2 and 0.2 first order
exact
0
for ε = 0.2. 0 1 2 3 4 5
x
6 7 8 9 10

31
4.3 Questions

Question 14. Coming back to Question 2, the radial motion of an object in the
gravitational field of the earth can be described by the following ODE
1
ÿ = − , y(0) = 0 and ẏ(0) = 1 .
(1 + εy)2

Derive a two–term expansion of the solution and compare it with numerical so-
lutions.      
t t3 t 2t
4
11t 11t2
Answer: y(t) = t 1 − +ε 1− −ε 1− + .
2 3 4 4 15 90
Question 15. Consider the perturbed harmonic oscillator

ÿ + (1 + ε)2 y = 0, y(0) = 1 and ẏ(0) = 0 ,

for ε  1 and t ∈ [0, T ].


Compute a two–term asymptotic expansion of the solution and compare it with
the exact solution. What happens for large times, i.e t → ∞?
4

Answer: We get y(t; ε) = cos ((1 + ε)t) and 2


2 2
y ∼ cos t − εt sin t − ε 2t cos t . For t → ∞ the 1

approximations grow unbounded, whereas the


y

exact solution is bounded. This happens since −1

secular terms appear in the expansion. −2 exact


0th order
1st order
−3
0 5 10 15 20 25 30
t

Question 16. Consider a spherical ball of radius R = 1 which is rotating around


the z–axis in a fluid with high viscosity. Assume the fluid fills whole R3 outside
the ball. This situation can be described by the Navier–Stokes equation

∇·u=0
ε (u · ∇) u + ∇p = ∆u
u→0 for x → ∞
u = ez × x at kxk = 1

Derive a two–term asymptotic expansion of the flow field.


Answer: We seek the expansion u ∼ u0 + εu1 and p ∼ p0 + εp1 in spherical
coordinates, where div u0 = 0 and ∇p0 = ∆ u0 . The boundary condition for
u0 read as u0 (r = 1) = sin θeϕ and u0 → 0 for r → ∞. Using the ansatz
u0 = u(r) sin θ eϕ , we get p0 = 0 and r 2 u00 + 2ru0 − 2u = 0 with the solution
u = r −2 . Therefore u0 = r −2 sin ϕ eϕ . This flow field is called the primary flow.

32
The secondary flow, i.e. the correction u1 is governed by div u1 = 0 and −∇p1 +
∆u1 = u0 ·∇u0 . With the identity ∆u1 = ∇(∇·u1 )−∇×(∇u1 ) and taking the curl
of the momentum equation, we deduce ∇ × (∇ × (∇ × u1 )) = −6 sin θ cos θr −6 eϕ .
1 ∂ψ
Introducing the streamfunction ψ by u1 = ver + weθ , v = − r2 sin θ ∂θ
and w =
1 ∂ψ
r sin θ ∂r
, we get an equation for ψ.

Question 17. Reconsider the potential equation (4.8) in the ”near” sphere given
by (4.7). Here’s another way to construct asymptotic expansions of problems,
where the domain is perturbed:
Given the perturbed boundary r = R(θ; ε) = 1 + εP2 (cos θ), we introduce a new
radial variable ρ by r = ρR(θ; ε). Rewriting the equation (4.8) in terms of ρ and
θ, the small parameter ε appears in the equation and the boundary of the new
problem is now ε–independent, i.e. ρ = 1.
Construct a two–term expansion of (4.8) using this approach.
Answer: The zeroth–order term u0 (ρ, θ) satisfies ∆u0 = 0, u0 (1, θ) = 1 and
u0 (∞, θ) = 0. Therefore u0 = 1/ρ. The next order is determined by the equation
∆u1 = 3(1 − cos2 θ)ρ−3 and u1 (1, θ) = 0, u1 (∞, θ) = 0. The solution is given by
u1 (ρ, θ) = (cos2 θ − 4 ln r) /(2r 3 ) − (3 cos2 θ − 1 + r 2 ) /(6r 3 ).

33
5 Singular Perturbation of ODEs
An ODE is called singularly perturbed, if the reduced problem is of lower or-
der. Hence the reduced problem cannot satisfy all the boundary conditions of
the perturbed problem. Here the question arises, how to deal with the ”lost”
boundary conditions. In general there will be a small neighborhood close to one
of the boundaries, where the solution rapidly changes. These neighborhoods are
called boundary layers.

Example 5.1. Consider the problem 1

εu00 + (1 + x)u0 = 1, u(0) = 0, u(1) = 1 . 0.8

(5.1) 0.6

An analytical solution of this problem is not

u
0.4
available (at least not with Maple 8.0). Note,
that the unperturbed problem (1 + x)u00 = 1 0.2
ε=0.1
cannot satisfy both boundary conditions! 0
ε=0.01
0 0.2 0.4 0.6 0.8 1
x

How to solve such problems? We proceed in 4 steps

1. Outer Expansion
2. Inner Expansion
3. Matching
4. Composite Expansion

5.1 Outer Solution

Let’s start with a regular expansion of the solution


u(x) ∼ y0 (x) + εy1 (x)
Then we get the problems:
O(1) : (1 + x)y00 = 1 ,
O(ε) : (1 + x)y10 = −y000 ,

and the general solutions are given by

y0 (x; c) = c0 + ln(1 + x)
1
y1 (x; c) = c1 − ,
2(1 + x)2

34
for some constants c0 , c1 ∈ R.

At the moment we have no possibility to fix 1


the constant c0 in the O(1)–problem, since
we do not know, which boundary condition 0.8

to impose. 0.6
What can happen? Suppose we prescribe

y
the boundary condition at x = 0, then c0 = 0.4

0 and y0 = ln(1 + x). However, prescribing


0.2 c=0
the other boundary condition u(1) = 1, we c=1−ln(2)
c=.15
obtain c0 = 1−ln 2 and y0 = 1+ln[(1+x)/2]. 0
0 0.2 0.4 0.6 0.8 1
x

5.2 Inner Solution

In all these cases, we conclude, that at least one region around some x0 ∈ [0, 1]
exists, where the solution drastically changes, this is the so–called boundary layer.
So let’s zoom into that region by introducing a stretched variable
x − x0
ξ= , α>0.
εα
With this scaling the equation (5.1) reads as

ε|1−2α 00
{z v } + (1 + x0 )ε−α v 0 + ξv 0 − 1 = 0 .
| {z } | {z }
➀ ➁ ➂

Now, there a different possibilities to balance the terms

➀ ∼ ➁ : Here α = 1 and ➀=➁ in leading order.


1
➀ ∼ ➂ : Here α = 2
and ➁ is dominant, leading to the trivial equation v 0 = 0.

➁ ∼ ➂ : Here α = 0 and we get the scaling of the outer equation.

So we conclude, that the only non–trivial case is α = 1 and we assume

u ∼ v0 (ξ) + εv1 (ξ) .

Then

v000 + (1 + x0 )v00 = 0
v100 + (1 + x0 )v10 = 1 − ξv00

35
and hence
v0 (ξ) = d0 + d1 e−ξ/(1+x0 ) .
But how to determine the constants d0 and d1 ?
If the boundary layer is located at x0 = 1, then v0 should satisfy the boundary

condition at x0 = 1 and hence ξ = (x − 1)/ε, v0 = 1 − d1 1 − e−ξ/2 . For
ξ → −∞, i.e. going into the interior of the domain, we expect some finite limit
of v0 , but
lim v0 (ξ) = (1 − d1 ) + d1 · ∞
ξ→−∞

and d1 = 0 leads to a trivial expansion. Hence we discard the possibility of a


boundary layer at x0 = 1. Similarly we can conclude that no interior layer exists,
i.e. x0 ∈
/ (0, 1]. Therefore x0 = 0 is the only remaining possibility and

v0 (ξ) = d0 1 − e−ξ
 
d0 2
v1 (ξ) = ξ + ξd0 − d1 e−ξ + ξ + d1
2

with some unknown constants d0 , d1 ∈ R.


So far we have constructed two different expansions (only the first terms)

• The outer expansion 1

1+x 0.8
y0 (x) = 1 + ln
2
0.6

valid outside the boundary layer.


u

0.4

• The inner expansion


0.2
−x/ε
 outer
v0 (x) = d0 1 − e inner
0
0 0.2 0.4 0.6 0.8 1
x
valid inside the boundary layer.
Note, that the inner expansion contains a constant d0 , which is so far left un-
specified. We will fix this constant by matching both expansions.

5.3 Matching

There are several possibilities to accomplish this

1. Patching. Brute force method (only as last resort). Force the inner and
outer expansion to coincide at some chosen point xp , e.g. ε = 10−4 and

36
xp = 10−3 . Then
y0 (xp ) = 0.30785
v0 (xp ) = d0 · 0.99995 d0 = 0.30787

2. van Dyke’s ”Limit Matching Principle”. Easy to apply, works mostly.


Idea: There should exist a transition region, where both expansions, the
inner and the outer one, coincide.
We expect, that the value of the outer expansion, when entering the bound-
ary layer (x → 0) equals to the value of the inner expansion, when leaving
the boundary layer (ξ → ∞).

”Inner limit of the outer expansion”


= ”Outer limit of the inner expansion”
Here
1
lim y0 (x) = 1 + ln
x→0 2
lim v0 (ξ) = d0
ξ→∞

and hence

d0 = 1 − ln 2 = 0.30685

An extension of the above principle is the


Lemma 5.1 (Asymptotic Matching Principle). Take n terms of the
outer expansion, rewrite it in the inner variable and construct the first m–
terms of the expansion.
Also consider the first m terms of the inner expansion, rewrite it in the
outer variable and construct an n–term expansion.
Typically, one uses m = n or n + 1. Then

m–term inner expansion of n–term outer expansion


= n–term outer expansion of m–term inner expansion

Let’s apply the asymptotic matching principle to our example. The outer
and inner expansion are given by
 
1+x ε 1 1
y(x) = 1 + ln + −
2 2 4 (1 + x)2
  
−ξ
 d0 2 −ξ
v(ξ) = d0 1 − e +ε ξ + ξd0 − d1 e + ξ + d1
2

37
To rewrite the outer expansion in the inner variable, we have to consider
y(x) = y(εξ) and expand w.r.t ε. On the other hand, the outer expansion of
the inner solution is given by v(x/ε). Note, that here exponentially small
terms O(e−x/ε ) will appear, they are abbreviated as e.s.t. Considering
m = n, we get

n, m inner of outer outer of inner Matching condition


1 1 − ln 2  d0 + e.s.t
 d0 = 1 − ln 2
3 x
2 1 − ln 2 + ε ξ − 8 d0 + ε ε + d1 + e.s.t d1 = − 38

3. Intermediate Scale (Kaplun ’57, Lagerstrom ’88)


This matching procedure is based on the assumption, that there exists a
transition layer between the outer O(1)–scale and the boundary layer O(ε)–
scale, where both (outer and inner) solution still provide valid expansions.
In this region both expansions have to be equal. The important point here
is, that the outer expansion y and the inner expansion v are approximations
of the same function.

Example 5.2. To illustrate the concept of the different layers, let’s consider
the singular perturbed linear ODE

εu00 + u0 = 1, u(0) = 0, u(1) = 0 . (5.2)

The solution can be computed explicitly and is given by

e−x/ε − 1
u(x; ε) = x + .
1 − e−1/ε
For the first term of the outer expansion, we easily compute

y0 (x) = x − 1 ,

and the 1-term inner expansion in the inner variable ξ = x/ε (the boundary
layer is located at x = 0) reads as

v0 (ξ) = c0 1 − e−ξ .

Where are now these two expansions of one and the same fuunction u(x; ε)
valid? Let’s consider the outer expansion: Clearly it satisfies the prob-
lem (5.2) on any interval [µ(ε), 1], where µ(ε) = O(εα ) for α > 0, i.e. on
any interval whose left boundary tends to zero. This is reasonable, since the
outer expansion does not satisfy the left boundary condition, but the resid-
ual of the equation is O(ε). We can also confirm this result by comparing
u(x; ε) and y0 directly.

38
Now to the inner expansion. In the inner variable ξ the problem (5.2)
reads as v 00 + v 0 = ε, v(0) = 0 and v(1/ε) = 0. Clearly the inner expansion
satisfies the problem and the according boundary conditions on any interval
ξ ∈ [0, ν̃(ε)], where ν̃(ε) = O(εβ ) for β > −1, i.e. ξ ≤ 1/ε. The same result
can be obtained by comparing directly u and v. Re–interpreting this domain
of validity in the outer variable x, we see, that the inner expansion is valid
on intervals of the form x ∈ [0, ν(ε)], where ν = O(εβ ) for β > 0.
Summerizing this discussion, we get that the outer expansion is valid on
any interval whose left boundary tends to zero, whereas the inner expansion
is valid on any interval whose right boundary tends to zero.
Now, lets consider a scale xη = x/η(ε),
which is located between the outer scale −0.9
outer
inner
x and the√ inner one ξ = x/ε, e.g choose −0.92 exact

η(ε) = ε. Then the outer expansion −0.94

reads in the xη –variable as


−0.96


y(xη ) = εxη − 1 ∼ −1 −0.98

−1

and the inner expansion reads in terms 10


−1 0
10
1
10

of xη as
Outer, inner expansion and the
 √  exact solution on the scale
v(xη ) = c0 1 − e −xη / ε
∼ c0 + e.s.t √
xη = x/ ε for ε = 10−4 .
Lemma 5.2 (Intermediate Matching Lemma). Consider u(x; ε) for
x ∈ [0, 1] and two approximations f (x; ε) and g(x; ε). Assume f is an
uniformly valid approximation of order O(ϕ) on an interval Df = [µ(ε), 1]
where µ = O(εα ). Similarly assume g is uniformly valid of order O(ϕ)
on Dg = [0, ν(ε)], ν = O(εβ ). Assume, that on an overlap domain D ⊂
Df ∩ Dg both expansions f and g are of order O(ϕ), i.e. α > β. Let η ∈ D,
i.e. choose η(ε) such that µ ≤ η ≤ ν and introduce an intermediate variable
xη = x/η(ε). Then
|f (xη ; ε) − g(xη ; ε)|
lim =0
ε→0 ϕ(ε)
i.e. f (xη ; ε) − g(xη ; ε) = o(ϕ).

How to use the intermediate matching to fix the undetermined constants


in our example? First we have to show, that there is an overlap region.
Consider the outer expansion y = 1 − ln 2 + ln(1 + x). This one clearly
satisfies the equation uniformly on the domain [µ, 1] for µ = εα for all
α ≥ 0. Therefore the outer expansion is uniformly valid on all intervals
where the left boundary tends to zero. Now consider the inner expansion
v = d0 (1 − exp(−x/ε)). The equation is satisfied on intervals of the type

39
[0, ν], where ν = ε0 , since the remainder is given by d0 x exp(−x/ε) − ε.
Hence we get [εα , 1] ⊂ Dy for α ≥ 0 and [0, ε0 ] ⊂ Dv . We can now choose
and intermediate scale locate between the scale ε of the inner expansion and
the scale 1 of the outer expansion. Let η ∈ (0, 1) and introduce xη = x/εη .
Rewriting the outer expansion w.r.t. the intermediate variable xη , we get
1 + ε η xη
y(xη ) = 1 + ln ∼ 1 − ln 2 + . . .
 2 
η−1
v(xη ) = d0 1 − e−xη ε ∼ d0 + . . .

Again, we recover the matching condition d0 = 1 − ln 2.

Comparing the three methods, we can state the following

• Patching works always, but it is an artifical procedure.


• Matching principles (rules) are mostly used, but sometimes they lead to
wrong results or cannot be applied (Fraenkl ’69, Lagerstrom ’88).
• Intermediate Scales is the most rigorous and most general method, but the
existence of overlap domains is up to now not completely proven.

So far we have constructed two approximations

• outer expansion y
• inner expansion v and
• matched them.

5.4 Composite Expansion

To combine both expansions into one, we define the common part of the outer
and inner expansion as the outer limit of the inner expansion. Then we construct
the composite expansion by
u0 = y0 + v0 − common part .
In our example we have
1+x
y0 (x) = 1 + ln
2 
v0 (x) = (1 − ln 2) 1 − e−x/ε
common part = 1 − ln 2

40
and hence
1+x 
u0 = ln + 1 + (1 − ln 2) 1 − e−x/ε − (1 − ln 2)
2
1+x
= ln − (1 − ln 2) e−x/ε + 1
2
The common part of the 2–term expansions equals (1 − ln 2) + x − 38 ε and hence
 
3
u1 (x) = y1 (x) + v1 (x) − 1 − ln 2 + x − ε .
8

The figure shows a comparison of a numerical 1


solution of (5.1) for ε = 0.1 and the first term
of the composite expansion. Both solutions 0.8

are almost indistinguishable. 0.6

u
ε ku( · ; ε) − u0 k∞ ku( · ; ε) − u1 k∞ 0.4

0.1 0.0113 0.0092 0.2


0.01 0.0031 1.4 · 10−4 Numerical, ε=0.1
Composite
0
0 0.2 0.4 0.6 0.8 1
x

5.5 A Formal Definition of Boundary Layers

In the previous example we saw, that the reduced problem


(1 + x)u00 = 1, u0 (0) = 0, u0 (1) = 1
does not have a solution. The solution y0 = 1 + ln 1+x
2
of the reduced problem
approximated the exact solution for all x 6= 0. This observation motivates the
following definition
Definition 5.1. Let D ⊂ Rn be a domain and for 0 < ε ≤ ε0 let uε ∈ C(D) be
a real–valued function. For u ∈ C(D), we define the supremum-norm on D by
kukD := sup |u(x)| .
x∈D

The function uε is called regular in D, if the there exists a function u ∈ C(D)


such that limε→0 uε = u w.r.t. the supremum–norm, i.e.
lim kuε − ukD = 0 .
ε→0

Let S ⊂ D be a C 1 –manifold with dimension less than n. The function uε shows


a boundary layer behavior at S, if uε is not regular in D, but regular for all D1 ,
where D 1 ⊂ D \ S.

41
Theorem 5.3. Let uε have a boundary layer at S. Then there exists a function
u ∈ C(D \ S) independent of ε such that

lim kuε − ukD1 = 0


ε→0

for D1 ⊂ D \ S.

Proof. For each D1 with D1 ⊂ D \ S exists a function uD1 ∈ C(D1 ) such that
limε→0 kuε − uD1 kD1 = 0. Furthermore for D1 , D2 ⊂ D \ S and x ∈ D1 ∩ D2 , we
have uD1 ∩D2 (x) = uD1 (x) = uD2 (x). Hence we can define a function u : D\S → R
by u(x) = uU (x) (x), where U (x) is some arbitrary neighborhood of x in D \ S.
The continuity of u is a consequence of the uniform convergence.

The function u describes the behavior of uε away from the boundary layer. To
analyze the behavior of uε close to the boundary layer, we use a strechted local
coordinate ξ and x = x(ξ; ε). A good choice of a local coordinate leads to a
regularisation, i.e. there exists a domain where the (zoomed) function Uε (ξ) :=
uε (x(ξ; ε)) is regular.
Although the term boundary layer suggests, that the loss of regularity appears
at the boundary of the domain D, this need not to be always the case. The
following examples shall illustrate what can happen.

5.6 Multiple Boundary Layers

Consider the problem

ε2 u00 + εxu0 − u = −ex , u(0) = 2, u(1) = 1 (5.3)

We start with the outer expansion u ∼ y0 , where y0 satisfies the equation

y0 = e x

Clearly, y0 cannot satisfy either boundary condition! Hence we expect at least


two boundary layers located at x0 = 0 and x1 = 1.
We proceed with the inner expansions. Let’s start with the boundary x0 = 0.
We introduce ξ = x/εα and v(ξ) = u(x/εα ) in (5.3)
α
ε|2−2α 00
εξv 0 − |{z}
{z v} + |{z}
ε ξ
| {z } .
v = −e
➀ ➁ ➂ ➃

➀∼➁ : Then α = 1/2 outer expansion.

42
➀∼➂ : Then α = 1: works.

For α = 1 and v ∼ v0 + . . . we get the boundary layer equation

v000 − v0 = −1, for 0 < ξ < ∞ and v0 (0) = 2

with the solution


v0 (ξ; c) = 1 + ce−ξ + (1 − c)eξ .
Matching requires
!
1 = lim y0 (0) = lim v0 (ξ; c)
x→0 ξ→∞

and hence c = 1.
To determine the solution at the other boundary x1 = 1 we introduce η = (x −
1)/εβ , w(η) = u((x − 1)/εβ ) and
βη
ε2−2β w 00 + (1 + εη)ε1−β w 0 − w = −e1+ε .

Balancing yields β = 1 and for w ∼ w0 (η) + . . . we get

w 00 + w 0 − w = −e, for − ∞ < η < 0 and w(0) = 1

with the solution


w0 (x; b) = e + ber1 η + (1 − e − b)er2 η

where 2r1,2 = −1 ± 5. The matching condition reads as
!
e = lim y0 (x) = lim w0 (η; b)
x→1 η→−∞

and yields b = 1 − e.

To construct the composite expansion, 3


we add the expansions and subtract the numerical, ε=0.01
1−term composite
common parts (which are given by the
2.5
limits appearing in the matching con-
ditions). Therefore
2
u

u(x) ∼ y0 (x) + v0 (x) − v0 (∞)


1.5
+ w0 (x) − w0 (−∞)
∼ ex + e−x/ε + (1 − e)e−r1 (1−x)/ε , 1
0 0.2 0.4 0.6 0.8 1
√ x

where 2r1 = −1 + 5.

43
5.7 Where is the Boundary Layer?

Consider the example

εu00 = uu0 − u, u(0) = 1, u(1) = −1 .

Again, we start with the outer expansion u ∼ y0 , and y0 solves

y0 y00 − y0 = 0 ⇐⇒ y0 (y00 − 1) = 0
(1) (2)
with the two solutions y0 (x) = 0 and y0 (x; c) = x + c.
To construct the inner expansion, we first have to determine, where the boundary
layer is, and which of the two possible outer solutions it should match!
Since we do not know the position of the boundary layer, we leave it open and
introduce the local variable ξ = (x − x0 )/εα for some x0 ∈ [0, 1] and v(ξ) =
u((x − x0 )/ε). Then
ε1−2α v 00 = ε−α vv 0 − v .
Balancing yields α = 1 and assuming v(ξ) ∼ v0 (ξ) + . . . we get

v000 = v0 v00 ⇐⇒ (2v00 − v02 ) = k

for k ∈ R. This non–linear ODE has the solutions


 2

 , for k = 0


 C − ξ 
 B
v0 (ξ) = B tan (x + C) , for k = B 2

 2
 B(x+C)

 1 − e
B , for k = −B 2
1 + eB(x+C)
For the purpose of matching, neither the rational nor the tangent–solution are
feasible. Either the limits for ξ → ±∞ are zero or they do not exist. Hence we
stick with the last possibility. Note, that the sign of B does not matter, hence
we restrict to B > 0 and introduce D = eBC :
1 − DeBξ
v0 (ξ; B, D) = B
1 + DeBξ

• Assume x0 = 0. Then v0 (0) = 1 and D = (B − 1)/(B + 1). Furthermore we


obtain limξ→∞ = −B < 0. For −1 ≤ B ≤ 1, the boundary layer function
has a pole for some ξ > 0, so we have to restrict to B > 1. Then the outer
(2) (2) !
solution can be matched with y0 (x; c) = x + c for c = y0 (0; c) = −B and
!
in order to satisfy the other boundary condition y (2) (1, −B) = 1 − B = −1,
we get B = −2.

44
• Assume x0 = 1. Then v0 (0) = −1 D = (B + 1)/(B − 1)and limξ→−∞ =
B > 0. Analogously, the continuity of v requires |B| < 1 and we can match
with the outer solution for c = B − 1. The boundary condition at x = 0
requires c = 1 and hence B = 2.

• Assume x0 ∈ (0, 1). In this case we have an interior layer. Then the outer
solution lives on 0 < x < x0 and satisfies the boundary condition at x = 0.
(2)
This yields the outer solution y0 (x) = x + 1. Additionally, the outer
solution is also valid on the other part x0 < x < 1 and is given there by
(2)
y0 = x − 2. To match the outer solutions with the interior layer, we have
to satisfy
!
lim x + 1 = x0 + 1 = v(−∞; B, D)
x→x0 −
!
lim x − 2 = x0 − 2 = v(∞; B, D) .
x→x0 +

Therefore B = x0 +1 and −B = x0 −2 leading us to B = 3/2 and x0 = 1/2.


The constant D is yet undetermined.

So we have found three possible boundary layer solutions! Which is the correct
one? Of course the ODE (5.7) has a unique solution.
Reconsider the perturbed problem (5.7) and the transformation U (1 − x) =
−u(x). Then εU 00 = U U 0 − U 0 with the boundary conditions U (0) = −u(1) = 1
and U (1) = −u(0) = −1. Hence the problem (5.7) is symmetric around x = 1/2.
So we expect that our approximation also shows this symmetry. Of course, the
two possibilities with layers either at x = 0 or x = 1 are non symmetric. Ruling
them out, we are left with the interior layer at x = x0 = 1/2 and v(0) = 0, i.e
D = 1 due to symmetry.
The interior layer is now given by

3 1 − e3ξ/2 3 3ξ
v0 (ξ) ∼ 3ξ/2
= − tanh
21+e 2 4

and the composite expansion reads as


  
 3 3(x − 1/2) 1
x + 1 −
 1 + tanh for x < 2
2 4ε
u0 ∼  (5.4)
 3 3(x − 1/2) 1
x − 2 +
 1 − tanh for x > 2
2 4ε

45
The following graphics shows a com- 1.5
parison of the numerical solution numerical, ε=0.1
1−term composite
of (5.7) and the 1–term composite ex- 1

pansion (5.4) for ε = 0.1. Matlab failed 0.5

to compute the numerical solution for


0

u
smaller values of ε.
−0.5

−1

−1.5
0 0.2 0.4 0.6 0.8 1
x

5.8 A Singular Perturbed PDE


We consider the diffusive regularization of the first order advection equation
ut + ux = εuxx , for (t, x) ∈ R+ × R+
with the continuous initial and boundary conditions u(0, x) = ϕ(x) and u(t, 0) =
ψ(t).
To construct the outer solution, we plug u(t, x) ∼ y0 (t, x) into (5.8) and obtain
in leading order y0,t + y0,x = 0. Using the method of characteristics, we easily get
(
ϕ(x − t) for x > t
y0 (t, x) =
ψ(t − x) for t > x
If ϕ(0) 6= ψ(0), there is a discontinuity propagating along x = t.
To construct the inner approximation, we introduce the scaling ξ = (x − t)εα and
u(t, x) = v(t, (x − t)εα ) and obtain
1
vt = ε1+2α vxx α=−
2
With the expansion v ∼ v0 (t, x), we get the inner solution (by similarity trans-
formation)  
ξ
v0 (t, ξ) = a erf √ +b
2 t
and the constants a, b are determined by the matching conditions
!
lim v0 (t, ξ) = a + b = ϕ(0)
ξ→∞
!
lim v0 (t, ξ) = −a + b = ψ(0) ,
ξ→−∞

and hence  
ϕ(0) − ψ(0) ξ ϕ(0) + ψ(0)
v(t, ξ) = erf √ +
2 2 t 2

46
5.9 Questions

Question 18. Find a composite expansion of the problem

εu00 + 2u0 + u3 = 0, u(0) = 0, u(1) = 1/2 .

Question 19. The Michaelis–Menten reaction scheme for an enzyme catalyzed


reaction is
ds
= −s + (s + k − 1)c
dt
dc
ε = s − (s + k)c ,
dt
for t > 0 and the initial conditions s(0) = 1, c(0) = 0. Here s(t) is the concen-
tration of substrate, c(t) is the concentration of the chemical produced by the
catalyzed reaction and k is a positive constant. Find the first term in the outer,
inner (initial) and composite expansion.

Question 20. Consider the problem

(x + εu)u0 + u = 1, in 0 ≤ x ≤ 1

subject to u(1) = 2.
A naive expansion yields

1+x 3x2 − 2x + 1
u(x; ε) ∼ +ε
x 2x3
Determine the
√ domain of validity and show that the expansion valid to√be valid
for x = O( ε). Introduce an inner variable ξ by the scaling x = εξ and
construct the inner expansion. Match the inner and outer expansion.

47
6 Lindstedt Expansions

Consider the equation

ü + (1 + ε)2 u = 0, u(0) = α, u̇(0) = 1 , (6.1)

with the analytic solution u(t; ε) = α cos(1 + ε)t. A 2–term expansion yields

u(t; ε) ∼ α cos t − εαt sin t

on compact time intervals. But for t → ∞ the asymptotic expansion becomes


unbounded due to the so called secular term t sin t, whereas the exact solution
stays bounded. In fact, the exact solution is periodic with frequence ν(ε) = 1 + ε.
Although this problem seems to be quite different from the previous discussion
of boundary layers, we will see, that there is some relationship between both.
In boundary layer problems, the loss of regularity happens at some point inside
the domain, whereas in the above problem, the loss of regularity shows up for
”t = ∞”.
If we want to expand a periodic function in a power–series, which is uniformly
valid, the following result holds.

Theorem 6.1. Let ϕ(t; ε) ∈ C 0,n+1 for all t and 0 < ε < ε0 . Let pk (t; ε) be the
k–th order Taylor polynomial w.r.t. ε. If ϕ is T –periodic, i.e. ϕ(t + T ; ε) = ϕ(t, ε)
for all t, then there exists 0 < ε1 < ε0 and C > 0, such that

|ϕ(t; ε) − pk (t; ε)| ≤ Cεk+1

for all t and all k ≤ n.



1 ∂k
Proof. Since ϕ is periodic, we also have that ak (t) := k
ϕ(t; ε) is T –
P k! ∂ε ε=0
periodic and hence the Taylor–polynomial pk (t; ε) = ak (t)εk is also T –periodic.
For t ∈ [0, T ] (compact!), we have the estimate

|ϕ(t; ε) − pk (t; ε)| ≤ Cεk+1

for some constant C > 0 and 0 < ε < ε1 < ε0 . For arbitrary T , we use the
periodicity, i.e. write t = τ + nT , where τ ∈ [0, T ].
Remark 6.1. The important and crucial point is, that the period T is independent
of ε.

48
6.1 Generalized Asymptotic Expansions

”Standard” power series are therefore seemingly not well suited for approximat-
ing periodic functions, if the period depends on the perturbation parameter ε.
Therefore we introduce the following generalized concept of asymptotic expan-
sions.
Definition 6.1 (Generalized Asymptotic Expansion). A function f (x; ε)
has a n–term (generalized) asymptotic expansion w.r.t. an asymptotic sequence
{Φk }, if
k
X
f (x; ε) = aj (x; ε)Φj (ε) + o(Φk )
j=1

for k ≤ n, ε < ε1 and pointwise for x.


It is required, that aj (x; ε) = O(1) (i.e. bounded) for all j ≤ n and fixed x. We
write
f (x; ε) ∼ a1 (x; ε)Φ1 (ε) + . . .

Generalized asymptotic expansions provide a possibility to construct expansions


to periodic functions where the period depends on ε. These expansions are called
Lindstedt expansions.
The previous result on the convergence of Taylor polynomials relies on the fact
that the period T is independent of ε. The idea behind the Lindstedt–expansion
of a function ϕ with ε–dependent period T (ε) or frequency ν(ε) = 2π/T (ε) is to
introduce a new time variable θ, the so called strained time via

θ = ν(ε)t

and accordingly
 
θ
ψ(θ; ε) = ϕ ;ε
ν(ε)
or

ϕ(t; ε) = ψ(ν(ε)t; ε) .

Then the function ψ is 2π–periodic and thus

ψ(θ; ε) = ψ0 (θ) + εψ1 (θ) + · · · + εk ψk (θ) + O(εk+1 )

uniformly for all θ.


Hence we have

ϕ(t; ε) = ψ0 (ν(ε)t) + εψ1 (ν(ε)t) + · · · + εk ψk (ν(ε)t) + O(εk+1 )

49
uniformly for all times t. Therefore
ϕ ∼ ψ0 (ν(ε)t) + · · · + εk ψk (ν(ε)t)
provides an uniformly valid generalized asymptotic expansion. If ε enters in each
coefficient only through the frequency, then the expansion is called a Lindstedt
expansion.
Example 6.1. Consider the function
1
ϕ(t; ε) = cos(1 + ε)t
1+ε
Then a ”standard” expansion yields
ϕ ∼ cos t − ε (t sin t + cos t) .
This expansion is valid on bounded intervals.
To construct a Lindstedt expansion, we introduce the strained time via ν(ε) =
1 + ε and θ = (1 + ε)t. Then
1
ψ(θ; ε) = cos(θ) ∼ cos θ − ε cos θ
1+ε
ϕ(t; ε) ∼ cos(1 + ε)t − ε cos(1 + ε)t
Remark 6.2. If the frequency is known in advance, we know from the above
considerations, that the Lindstedt expansion holds uniformly for all t.

And what to do, if we do not know the frequency in advance? Approximate it


via an asymptotic expansion again! We introduce
ν(ε) ∼ ν̂(ε) = ν0 + εν1 + · · · + εk+1 νk+1
and construct the Lindstedt expansion using the frequency ν̂
ψ̂ = ψ0 (θ) + εψ1 (θ) + · · · + εk ψk (θ)
Note, that we expand ν up to order εk+1 , whereas ψ is only expanded up to order
εk . Why that later.
The coefficients for νj and ψj can now (hopefully) be determined alternately:
(ν0 , ψ0 ), (ν1 , ψ1 ), . . . (νk , ψk ), νk+1 .
Theorem 6.2. Let ϕ(t; ε) be a smooth and periodic function with frequency ν(ε),
ν smooth, i.e.
 

ϕ t+ ; ε = ϕ(t; ε) ∀t, ∀0 < ε < ε0 .
ν(ε)

Let ψ̂ and ν̂ be defined as above. Then the following holds:

50
1. The approximation

ϕ(t; ε) = ψ̂(ν(ε)t; ε) + O(εk+1 )

is uniformly valid for all t. Note, that we use here the exact frequency.

2. The Lindstedt expansion

ϕ(t; ε) = ψ̂(ν̂(ε)t; ε) + O(εk+1 )

is uniformly valid on expanding intervals of length O(1/ε).


3. For 1 ≤ j ≤ k + 1 we have

ϕ(t; ε) = ψ̂(ν̂(ε)t; ε) + O(εk+2−j )

uniformly on expanding intervals of length O(1/εj ).

Proof. The first statement is clear due to Theorem 6.1, i.e. we have

ϕ(t; ε) − ψ̂(ν(ε)t; ε) ≤ c1 εk+1

for all ε < ε2 .


Since ν(ε) ∼ ν̂(ε), there exists ε1 < ε0 and c0 > 0 such that

|ν(ε) − ν̂(ε)| ≤ c0 εk+2

for all ε < ε1 . Now, define



∂ ψ̂

M := max (θ; ε) .
θ∈[0,2π], ε<ε1 ∂θ

Since ψ̂ is 2π–periodic, M = maxt∈R, ε<ε1 ∂∂θψ̂ . Then



ψ̂(θ 1 ; ε) − ψ̂(θ 2 ; ε) ≤ M |θ1 − θ2 |

and


ϕ(t; ε) − ψ̂(ν̂(ε)t; ε) ≤ ϕ(t; ε) − ψ̂(νt; ε) + ψ̂(νt; ε) − ψ̂(ν̂t; ε)
≤ c1 εk+1 + M c0 εk+2 |t|

for all ε < min(ε1 , ε2 ). This bound grows for t → ±∞. Now, let L > 0 and
consider |t| ≤ L/ε. Then

ϕ(t; ε) − ψ̂(ν̂t; ε) ≤ (c1 + M c0 L)εk+1

(∗)

51
uniformly for |t| ≤ L/ε. This shows the second statement.
Now consider the even larger interval |t| ≤ L/ε2 . Then

ϕ(t; ε) − ψ̂(ν̂t; ε) ≤ (c1 ε0 + M c0 L)εk

uniformly for |t| ≤ L/ε2 . In this manner, we can show the third statement for all
1 ≤ j ≤ k + 1.
Remark 6.3. The estimate (∗) improves for ε → 0 in two ways: First, the error
bound decreases. Secondly, the interval, on which (∗) is valid, increases.
For intervals of length O(1/εk+2 ) the error bound gets O(1), which says nothing
else but that the difference between the two periodic functions stays bounded.
Example 6.2 (Duffings Equation). The motion of an almost linear oscillator
can be described by

ü + u = εu3 , u(0) = 1, u̇(0) = 0 . (6.2)

Due to the criterion of Bendixson, we can expect periodic orbits. Rewrite the
equation as a first order system ẋ = f (x) with f (x1 , x2 ) = (x2 , −x1 + εx31 ) and
note, that div f ≡ 0.
To construct a Lindstedt expansion, we use the ansatz

ν ∼ ν0 + εν1 + ε2 ν2
θ = ν0 t + εν1 t + ε2 ν2 t
û(t; ε) = u0 (θ) + εu1 (θ) = u0 (ν0 t + εν1 t + ε2 ν2 t) + εu1 (ν0 t + εν1 t + ε2 ν2 t)

For the second time derivative we get


 
¨ ε) = ν 2 ü0 + ε 2ν0 ν1 ü0 + ν 2 ü1
û(t; 0 0

Inserting this into (6.2) we get the following problems

O(1) : ν02 ü0 + u0 = 0, u0 (0) = 1, ν0 u̇0 (0) = 0


O(ε) : ü1 + u1 = u30 − 2ν1 ü0 , u1 (0) = 0, ν0 u̇1 (0) = 0

Comparing the 0–th order problem with the reduced problem ü + u = 0, we get
ν0 = 1 and easily compute u0 (θ) = cos θ.
Then the first order problem reads as

ü1 + u1 = 2ν1 cos θ + cos3 θ

How to compute ν1 ? We know, that u1 has to be 2π–periodic and it should not


contain secular terms. Secular terms (resonant terms) appear, if the right hand

52
side 2ν1 cos θ + cos3 θ contains a term of the form a cos θ. To see that, recall that
ẍ + ν 2 x = aeiωt has unbounded solutions, if ω = ν. Applying trig–identities, we
get
1 3
2ν1 cos θ + cos3 θ = 2ν1 cos θ + cos(3θ) + cos θ
4 4
and we conclude ν1 = −3/8 to avoid secular terms. With this information the
first order equation reads as ü1 + u1 = 14 cos 3θ u1 = . . .
To determine the second order correction of the frequency, consider the O(ε2 )–
problem:  
ü2 + u2 = 3u20 u1 − 2ν0 ν2 + ν12 ü0 − 2ν0 ν1 ü1 .
Again, the resonant terms have to vanish. Either, we can identify them by trig–
identities or use a Fourier–expansion

a0 X
F (x) = + ak cos kx + bk sin kx ,
2 k=1

where Z π
1
ak = F (x) cos(kx) dx
π −π

and hence
Z π
!1   21
0= cos θ · 3u20 u1 − (2ν0 ν2 + ν12 )ü0 − 2ν0 ν1 ü1 dθ ν2 = −
π −π 256

Summarizing the results, we have found the following Lindstedt expansion:


3 21 2
ν ∼1− ε− ε
8 256
3 21 2
θ ∼ t − εt − εt
8 256
 
3 21 2
u0 = cos θ = cos t − εt − εt
8 256
1
u1 = · · · = cos θ sin2 θ
8

6.2 Questions

Question 21 (van der Pol’s Equation). Consider van der Pol’s equation

ü + u = ε(1 − u2 )u̇

53
This equation is known to have periodic solution (ω–limit cycles, to be more
precise). If we want to find these limit cycles, we seek a periodic solution of van
der Pol’s equation with the initial conditions

u(0) = α, u̇(0) = 0 .

Since we do not know the position of the limit cycle in advance, we do not know,
where it intersects the positive u–axis in the uu̇–phase plane. Hence the initial
value α, that corresponds to the periodic solution, is not known. However, we
expect it to depend on the perturbation parameter ε, i.e. α = α(ε).
Seek an asymptotic expansion of the limit cycle using a Lindstedt expansion for
the solution u = u(t; ε) and and expansion for the initial value α ∼ α0 + εα1 .
Try to determine the coefficients (ν0 , ψ0 , α0 ), (ν1 , ψ1 , α1 ) and ν2 and compare the
result with a numerical solution.

54
7 Multiple Scales
Let’s review the Lindstedt expansion of the solution to Duffing’s equation ü+u =
εu3 , which we derived in the previous chapter:
 
3 3 3
u(t; ε) ∼ cos t − εt = cos t · cos εt + sin t · sin εt .
8 8 8
This looks, as if our approximation lives on two different time scales, the normal
time t, where cos t and sin t oscillate; and a slow time εt, where cos 38 εt and sin 38 εt
live. Recall, that cos 38 εt has its first zero for t = 4π
3
ε−1 = O(1/ε). Hence the
variations of cos 83 εt become only visible on the slow time scale εt.

7.1 Lindstedt Revisited

With this interpretation in mind, it seems reasonable to write the solution u(t; ε)
of Duffing’s equation

ü + u = εu3 , u(0) = 1, u̇(0) = 0 (7.1)

formally as a function of two time scales t1 and t2 : u = u(t1 , t2 ), where

• t1 = t represents the normal scale


• t2 = εt represents the long–term or slow scale.

Then we seek a multiple scale expansion

u ∼ u0 (t1 , t2 ) + εu1 (t1 , t2 ) .

d2
Using the derivatives dtd = ∂t1 + ε∂t2 and dt2
= ∂t21 + 2ε∂t21 ,t2 + ε2 ∂t22 , we get the
zero–th order problem

O(1) : ∂t21 u0 + u0 = 0 ∀(t1 , t2 )

with the general solution

u0 (t1 , t2 ) = f (t2 ) cos t1 + g(t2 ) sin t1 .

The functions f and g living on the slow scale are yet undetermined. For the
initial conditions we get the information
!
1 = u0 (0, 0) = f (0)
!
(∗)
0 = u̇0 (0, 0) = g(0) + εf 0 (0)

55
Going to the next order we get

O(ε) : −2f 0 (t2 ) sin t1 +2g 0 (t2 ) cos t1 +u1 +∂t21 u1 = (f (t2 ) cos t1 + g(t2 ) sin t1 )3
= A1 (t2 ) cos t1 + A2 (t2 ) sin t1 + A3 (t2 ) cos 3t1 + A4 (t2 ) sin 3t1

with
3 
A1 (t2 ) = f (t2 ) f (t2 )2 + g(t2 )2
4
3 
A2 (t2 ) = g(t2 ) f (t2 )2 + g(t2 )2
4
1 
A3 (t2 ) = f (t2 ) f (t2 )2 − 3g(t2 )2
4
1 
A4 (t2 ) = f (t2 ) 3f (t2 )2 − g(t2 )2
4
Since the O(ε)–equation has to hold for all t1 and all t2 , we are allowed to compare
the Fourier–coefficients of cos t1 and sin t1 . Each of these Fourier–coefficients has
to vanish separately for all t2 , since otherwise they would lead to resonant secular
terms. This leads us to
3 
(cos t1 ) :g0 − f f 2 + g2 = 0
8
3 
(sin t1 ) : f 0 + g f 2 + g 2 = 0
8
f 2  g 
(Rest) : ∂t21 u1 (t1 , t2 ) + u1 (t1 , t2 ) = f − 3g 2 cos 3t1 + 3f 2 − g 2 sin 3t1
4 4
(∗∗)

The first two equations have to hold for all t2 , hence


1 2 0
0 = gg 0 + f f 0 = f + g2 ∀t2
2
and thus we conclude f 2 + g 2 = const. = 1 due to the initial condition (∗). With
this information, the equations read as
 2
0 3 0 3 00 3
g = f and f =− g g − g = 0, g(0) = 0
8 8 8

with the solutions


3 3
g(t2 ) = α sin t2 and f (t2 ) = α cos t2 .
8 8
From the initial condition (∗) we get f (0) = 1, i.e. α = 1.

56
Summarizing the information we got so far:
3 3
u0 (t1 , t2 ) = cos t1 cos t2 + sin t1 sin t2 .
8 8
The remainder (∗∗) can now be used to determine u1 (t1 , t2 ) in the same fashion.
As a conclusion of this example, we can say, that Lindstedt expansions are a
special case of the more general multiscale expansions.

7.2 Idea of Multiple Scales

As we have seen in the preceeding section, Lindstedt expansions for periodic


solutions can be viewed as a method to determine the behavior of the solution
on different (normal and slow) time scales. However, multiple time (or spatial
scales) do not only appear in periodic systems:

• Weakly damped oscillator ü + εu̇ + u = 0


• Slowly varying coefficients ü + k(εt)u = 0
• Combustion models u̇ = u2 (1 − u), u(0) = ε
• Diffusion in complex media ∇ (κ(x/ε)∇u) = f

Multiple scales expansions can serve as a method for problems, where we expect
phenomena to happen on different scales.
Example 7.1 (Damped Oscillator). Consider

ü + εu̇ + u = 0, u(0) = 0, u̇(0) = 1 . (7.2)

A regular expansion yields the nonsense result


1
u(t; ε) ∼ sin t − εt sin t .
2
Note, that this solution grows unbounded as t → ∞.
A Lindstedt expansion yields
ε2
ν ∼1−
8
ε2
θ ∼t− t
8    
ε3 ε2
u ∼ sin θ − εθ sin θ = 1 − εt + t sin t − t
8 8

57
Still, for t → ∞ this expansion grows unbounded.
What goes wrong? Consider the energy E(t) = 12 u̇(t)2 +u(t)2 : Multiply Eqn. (7.2)
by u̇ and integrate to get
Z t Z
!
0= u̇ (ü + εu̇ + u) dt E(t) = E(0) − ε u̇2 dt < E(0) .
0

I.e. the energy E(t) is strictly decaying,


dE
= −εu̇2 ,
dt
and the decay is slow. If u̇2 = O(1), then the decay happens on a time scale εt.
Therefore we expect two time scales in our problem: the normal scale t, where
u oscillates and the slow scale εt, where the amplitudes decay. So, let’s try a
multiple scale expansion of Eqn (7.2).
We assume the existence of (at least) two time scales: t1 = t and t2 = εα t and
try to find the leading term of the expansion

u(t; ε) ∼ u0 (t1 , t2 ) .

Performing the expansion and comparing equal powers of ε we get

O(1) : ∂t21 u0 + u0 = 0, u0 (0, 0) = 0, ∂t1 u0 (0, 0) = 1

with the solution

u0 (t1 , t2 ) = a(t2 ) sin t1 + b(t2 ) cos t1 , b(0) = 0, a(0) = 1 .

To determine the yet unknown functions a and b, we go to the next order:

α = 1, O(ε) : 2∂t21 ,t2 u0 + ∂t1 u0 = 0 .

Inserting u0 and equating the sin– and cos–terms separately, we get

a(t2 ) + 2a0 (t2 ) = 0 a(0) = 1 a(t2 ) = e−t2 /2


b(t2 ) + 2b0 (t2 ) = 0 b(0) = 0 b(t2 ) ≡ 0

and therefore

u0 (t1 , t2 ) = e−t2 /2 sin t1 u(t; ε) ∼ e−εt/2 sin t .

The exact solution of the problem (7.2) is given by


1 p 
−εt/2 2
u(t; ε) = p e sin t 1 − ε /4 .
1 − ε2 /4

58
Our two–scale expansion already captures the main features of the solution: the
exponential decay of the amplitudes on
p the slow scale εt 1and the oscillations on
2 2
the longer time scale t. Note, that 1 − ε /4 ∼ 1 − 8 ε , hence the ”sqrt”–
term changes the amplitudes and the frequency only in higher order terms of the
expansion.
To compute higher order approximations, one introduces a third time scale t3 =
εβ t and expands u(t; ε) ∼ u0 (t1 , t2 , t3 ) + εγ (t1 , t2 , t3 ).

7.3 Some General Remarks concerning Multiple Scales

The idea behind introducing multiple scales is to

• keep expansions well ordered

• minimize the error of the approximation

• avoid sometimes the appearance of secular terms.

To illustrate, what it means to minimize the error, consider the function

u(t; ε) = 1 + ε2 t + ε4 sin t

and the three times scales

t1 = 1, t2 = εt, t 3 = ε2 t .

When we seek an expansion of the form

u ∼ u0 + εu1 + ε2 u2

we have different possibilities:

1. u0 = 1, u1 = 0, u2 = t1

2. u0 = 1, u1 = t2 , u2 = 0

3. u0 = 1 + t3 , u1 = 0, u2 = 0

4. u0 = 1, u1 = −t2 , u2 = t1

59
All of these possibilities lead to u ∼ 1 + ε2 t. But which of these four is the
”correct” or ”most preferable” one?
Since y0 is the leading term in the expansion, it is reasonable, that it should
minimize the error
E0 = max |u(t; ε) − u0 (t1 , t2 , t3 )|
and hence, we choose u0 = 1 + t3 .
Analogously, u1 should minimize E1 = max |u(t; ε) − u0 (t1 , t2 , t3 ) − εu1 (t1 , t2 , t3 )|
leading to u1 = 0.
Another issue concerning multiple scales is the question of accuracy and validity.
Suppose a problem contains three time scales ti = εi t, i = 0 . . . 2, and we have
constructed the following expansions

1 1
u = u0 (t0 , t1 )

u ∼ u2 = u20 (t0 , t1 ) + εu21 (t0 , t1 ) .

 3 3
u = u0 (t0 , t1 , t2 )

How do these expansions compare? We can expect that u1 holds for εt = O(1),
where as u3 holds on the larger interval ε2 t = O(1). However, on the common
domain of validity 0 ≤ εt ≤ O(1), we cannot say which of these two should be
more accurate (both should be accurate up to O(ε))
The approximation u2 is expected to be valid up to εt = O(1). But we expect it
to be more accurate than u1 or u3 ; O(ε2 ) for u2 compared to O(ε) for u1 , u3 .
Concerning the computational effort to determine u2 or u3 , both expansions re-
quire some analysis of the O(ε2 )–problem.

7.4 Slowly Varying Coefficients

In the sequel we will consider two prototypic examples for multiple scale expan-
sions: Let k : R 7→ [kmin , kmax ] be smooth and kmin > 0. Then we consider the
two problems

ü + k 2 (εt)u = 0 (7.3)
ü + k 2 (t/ε)u = 0 (7.4)

Both problems look similar, but. . .

• The first problem is well–defined for ε → 0, since k(εt) → k0 := k(0) and


we expect u(t; ε) → α sin(k0 t) + β cos(k0 t).

60
• The situation is not that clear for the second problem. What happens with
k(t/ε) as ε → 0? Do the solutions u(t; ε) converge in some sense to a limit
u0 (t)? Which equation does u0 satisfy? These questions will be answered
in Chapter 8 on Homogenization.

We consider now the first problem

ü + k 2 (εt)u = 0, u(0) = 1, u̇(0) = 0 .

We seek a two–scale expansion with the scales t1 = f (t; ε) and t2 = εt. Check
yourself, that an expansion with t1 = t and t2 = εt fails. For the solution u we
seek an expasion of the form

u(t; ε) ∼ u0 (t1 , t2 ) .

The function f appearing in the normal time scale should satisfy the following
assumptions:

• f should be smooth,

• f ≥ 0, increasing w.r.t t and f (0) = 0,

• f  εt to separate it from the slow scale t2 = εt,

• last, but not least, simplify the problem.

Plugging the two time scale expansion into Eqn. (7.3), we get in leading O(1)–
order
f¨∂1 u0 + f˙2 ∂12 u0 + k 2 (t2 )u0 = 0 .
| {z } | {z } | {z }
➀ ➁ ➂

The unperturbed problem oscillates and shows no damping, therefore ➁ and ➂


should balance:

f˙2 ∂12 u0 + k 2 (t2 )u0 = 0, u0 (0, 0) = 1, f˙(0)∂1 u0 (0, 0) = 0 .

If f˙ = k(t2 ), i.e. Z t
∂f
= k(εt) f (t) = k(εs) ds .
∂t 0
With this definition, the scaling function f satisfies the above assumptions and
f¨ = εk 0 (t2 ) = O(ε), i.e. term ➀ is of lower order than ➁ and ➂. At this point one
can see, why an expansion with the time scale t1 = t fails: the O(1)–problem reads
as ∂12 u0 + k 2 (t2 )u0 = 0 and solving it is as difficult as the original problem (7.3).

61
Now, we can easily solve for u0 and get

u0 (t1 , t2 ) = a(t2 ) cos t1 + b(t2 ) sin t1 , a(0) = 1, b(0) = 0 .

In the next order we get



k 2 ∂12 u1 + u1 + 2k (−a0 sin t1 + b0 sin t1 ) + k 0 (−a sin t1 + b cos t1 ) = 0 .

The requirement, that u1 should not contain secular terms, leads to


s
k(0)
2ka0 + ak 0 = 0 a(0) = 1 a(t2 ) =
k(t2 )
2kb0 + bk 0 = 0 b(0) = 0 b(t2 ) = 0

and hence the leading term in our two–scale expansion reads as


s s Z t 
k(0) k(0)
u(t; ε) ∼ u0 (t1 , t2 ) = cos t1 = cos k(εs) ds . (7.5)
k(t2 ) k(εt) 0

For all ε > 0, the solution u(t, ε) is smooth and well–defined. In the limit ε → 0,
the solutions converge u(t; ε) → cos(k(0)t) as expected. The frequency in the
limit is the same as the frequency of the unperturbed problem. In that sense,
problem (7.3) is ”nice” and ”well–behaved”.
The following two graphs show a comparison of the numerical solution (’– –’) of
Eqn. (7.3) and the multi–scale expansion (7.5) (’—’) for k(t) = 1 + 12 sin t and
ε = 0.1. The left picture shows the time interval t ∈ [160, 200]; here the two
solutions are almost indistiguishable. The right picture shows the interval t ∈
[1960, 2000]. For larger times t, the expansion differs already from the numerical
solution; however no secular terms appear.

1.5 1.5
numerical numerical
multi scale multi scale
1 1

0.5 0.5

0 0
u

−0.5 −0.5

−1 −1

−1.5 −1.5
160 170 180 190 200 1960 1970 1980 1990 2000
t t

62
7.5 Multiple Scales & Boundary Layers

Let’s reconsider the starting example (5.1) from the chapter about singular per-
turbed problems.

εu00 + (1 + x)u0 = 1, u(0) = 0, u(1) = 1 . (7.6)

We now want to see, whether multiple scales can also help to solve those problems.
Clearly, due to the presence of the boundary layer, the problem (7.6) contains
two scales:

• the outer scale x1 = x and


• the inner or boundary layer scale x2 = x/ε.

However, the problem is not periodic. Due to the existence of the two scales, we
assume the following multi–scale expansion for the solution:

u(x; ε) ∼ u0 (x1 , x2 ) + εu1 (x1 , x2 ) .

Plugging this expansion into (7.6), we get


   
ε−1 ∂22 u0 + (1 + x1 )∂2 u0 + ∂22 u1 + (1 + x1 )∂2 u1 + 2∂1,2
2
u0 + (1 + x1 )∂1 u0 ∼ 1 .

The leading O(ε−1 )–order problem reads as

∂22 u0 + (1 + x1 )∂2 u0 = 0, u(0, 0) = 0, u(1, ∞) = 1

and we get the following solution

u0 (x1 , x2 ) = a1 (x1 ) + a2 (x1 )e−(1+x1 )x2

where a1 (0) + a2 (0) = 0 and a1 (1) = 1. To determine the functions a1 and a2 , we


have to go to the next order:

∂22 u1 + (1 + x1 )∂2 u1 = f (x1 ) + [g(x1 ) + h(x1 )x2 ] e−(1−x1 )x2

where we used the notations

f (x1 ) = 1 − (1 + x1 )a01 (x1 ) ,


g(x1 ) = (1 + x1 )a02 (x1 ) + 2a2 (x1 ) ,
h(x1 ) = −(1 + x1 )a2 (x1 ) .

Since the solution u1 has to stay bounded for x1 fixed and x2 → ∞ (no secular
terms), we get the condition

f (x1 ) = 0 a01 (x1 ) = 1/(1 + x1 ), a1 (1) = 1

63
and therefore we have determined the function a1 as
1+x
a1 (x1 ) = 1 + ln .
2
To determine a2 is not that easy. At the moment, we just know, that a2 (0) =
ln 2−1. But keeping in mind, that one of the purposes of multiple scale expansions
is to reduce the error in the expansions (minimize the contribution of the higher
order terms), we can make g(x1 ) ≡ 0 by
ln 2 − 1
(1 + x1 )a02 + 2a2 = 0, a2 (0) = ln 2 − 1 a2 (x1 ) = .
(1 + x1 )2
Since we have now determined the function a1 and a2 completely, we get our
multiple scale expansion as
1+x ln 2 − 1 −(1+x)x/ε
u(x; ε) ∼ u0 (x, x/ε) = 1 + ln + e . (7.7)
2 (1 + x)2

With the matched asymptotic expansions, we got the composite expansion

1+x
u(x; ε) ∼ 1 + ln − (1 − ln 2)e−x/ε . (7.8)
2
To compare both, we rewrite (7.7) as
2
1+x e−x /ε
u0 (x, x/ε) = 1 + ln − (1 − ln 2)(1 + x)2 e−x/ε ,
2 (1 + x)2
| {z }
:=ρ(x;ε)

and note that ρ(x; ε) is exponentially small outside the boundary layer. Inside
the boundary layer, i.e. for x = ξε we get ρ(ξ; ε) = 1 + O(ε). Hence the matched
asymptotics solution (7.8) and the multiple scale solution (7.7) only differ in
higher order terms.

7.6 Multiple Scale Analysis of Numerical Methods

Consider the boundary value problem

u00 = f, u(0) = 0, u(1) = 0 .

To solve this problem numerically, we can use a standard finite–difference method


with stepsize h = ε  1:

u(x + ε) − 2u(x) + u(x − ε) = ε2 f, u(0) = 0, u(1) = 0 .

64
Expanding u(x ± ε) we get the modified equation

ε2 (IV )
u00 + u =f . (7.9)
12
To obtain an asymptotic solution of the modified equation, we propose a two–
scale ansatz
x
u(x; ε) ∼ u0 (x, y) + ε2 u1 (x, y), y= .
ε
Comparing equal powers of ε yields
 
−2 2 1 4
O(ε ) : ∂ y + ∂ y u0 = 0 ,
12
 
−1 1 3
O(ε ) : 2∂x ∂y + ∂x ∂y u0 = 0 ,
3
   
0 2 1 2 2 2 1 4
O(ε ) : ∂ x + ∂ x ∂ y u0 + ∂ y + ∂ y u1 = f ,
2 12
   
1 1 3 1 3
O(ε ) : ∂ ∂y u0 + 2∂x ∂y + ∂x ∂y u1 = 0 .
3 x 3

From the O(ε−2 )–equation we get the solution


√ √
u0 (x, y) = a(x) + b(x)y + c(x) sin 12y + d(x) cos 12y .

With the help of the O(ε−1 )–equation


√  √ √ 
2b0 (x) − 4 3 c0 (x) cos 12y − d0 (x) sin 12y = 0 ∀(x, y)

we can conclude, that b0 = 0, c0 = 0 and d0 = 0. Since limy→∞ u0 (x, y) < ∞ for


all x, we get b ≡ 0, c ≡ 0 and d ≡ 0, i.e u0 (x, y) = a(x).
To determine the next term u1 (x, y), i.e. the structure of the error in the numerical
method, we have to consider the O(ε0 ) and the O(ε1 )–problems. For O(ε0 ) we
get the general solution
1 1 h √ √ i
u1 (x, y) = [f − a00 (x)] y 2 +f3 (x)y+f4 (x)− f1 (x) cos 12y + f2 (x) sin 12y
2 12
(∗)
To remove secular terms, we get the conditions
1
a00 (x) = f, a(0) = 0, a(1) = 0 a(x) = f x (x − 1) ,
2
f3 (x) ≡ 0 .

65
The O(ε1 )–equation reads as

000 1 h 0 √ √ i
−2ya (x) + 2f30 (x) 0
+ √ f2 (x) cos 12y − f1 (x) sin 12y = 0 ∀(x, y)
3
and thus fi (x) = ki = const. for i = 1, 2. Backsubstitution in (∗) yields
1 h √ √ i
u1 (x, y) = f4 (x) − k1 cos 12y + k2 sin 12y
12
and due to u1 (0, 0), we get f4 (0) = k1 /12.
Summarizing the results, we obtain the solution of the modified equation (7.9)
in the following form
" √ ! √ ! #
f 2  12x 12x
u(x; ε) ∼ x − x + ε2 k1 cos + k2 sin + f4 (x) .
2 ε ε

This results shows the typical oscillation of the error on the length scale of the
stepsize. This observation is the starting point of multigrid methods, to detect
and diminish these oscillations the solution is computed on different nested grids.

7.7 Questions

Question 22. A simple model to describe combustion processes is given by the


equation
u̇ = u2 (1 − u), u(0) = ε .
This equation has two equilibria: the unstable solution u = 0 and the attractive
solution u = 1. Due to the small perturbation at t = 0, the derivative is of
order O(ε2 ) and the solution stays small at the beginning. However, on the scale
t = O(1/ε), the solution rapidly jumps to u = 1 (explosion). Use a multi–scale
ansatz to compute an asymptotic expansion of the solution.
In [O’Malley: Singular Perturbation Methods for ODEs] you can find the solution
using matched asymptotics.

66
8 Homogenization

Let’s consider the problem (7.4)

ü + k 2 (t/ε)u = 0, u(0) = 1, u̇(0) = 0 ,

where we assume, that k is T –periodic. To gain some insight into the problem,
we multiply by u̇ and integrate
Z T
u̇ü + k 2 (t/ε)u̇u dt = 0
0
T Z T
1 2 1 d 2 2
u̇ + u k (t/ε) dt = 0 .
2 0 0 2 dt

To proceed further, we need an extension of the Riemann–Lebesgue–Lemma


(cf. Thm. 3.2), namely we have to evaluate
Z T
lim ϕ(t)f (t/ε) dt
ε→0 0

for f ∈ Lp , T –periodic and ϕ ∈ Lq , where p1 + 1q = 1. Recall, that Lq is the


dual of Lp for p ∈ (1, ∞). We have to compute the weak limit of the sequence
fε = f (·/ε) ∈ Lp !

8.1 Weak Limits and Rapidly Oscillating Functions

Let (B, k·kB ) be a Banach–space, e.g. Lp , p ∈ [1, ∞].

Definition 8.1 (Strong Convergence). Let (un )n∈N ⊂ B be a sequence of


functions in B and u ∈ B. We say that un converges strong in B to u, i.e. un → u,
iff
kun − ukB → 0 .

Let B 0 = Lb (B, R) be the dual of B, i.e. the space of all bounded linear functionals
acting on B. For f ∈ B 0 , we denote by hf , ui ∈ R the action of the functional f
on the function u ∈ B.

Definition 8.2 (Weak and Weak–∗ Convergence). We say that (un )n∈N ⊂ B
converges weakly to u ∈ B, i.e. u * u, iff

hu0 , un i → hu0 , ui ∀u0 ∈ B 0 .

67
Let X be the predual of B, i.e. X is a Banach space with B = X 0 . Then, a

sequence (un )n∈N ⊂ B converges weak–∗ to u ∈ B, i.e. un * u, iff
hun , vi → hu , vi ∀v ∈ X .
If the Banach–space, with respect to which this weak convergence holds, is im-
portant, we use the notations
un * u weakly in B ,

un * u weak–∗ in B .
Theorem 8.1 (Some Results from Functionalanalysis). The following im-
plications hold:
strong convergence =⇒ weak convergence =⇒ weak–∗ convergence .
If dim B < ∞, then strong and weak convergence are equivalent.
If X is reflexive, i.e. X ' X 00 , then
weak convergence in B ⇐⇒ weak–∗ convergence in B .

Let (un ) ⊂ B be weakly (weak–∗) convergent to u ∈ B. Then the sequence u n is


bounded, i.e. there exists C > 0 such that kun kB ≤ C for all n ∈ N. Furthermore
we have the estimate
kukB ≤ lim inf kun kB .
n→∞

Proof. See Dautray & Lions: Mathematical Analysis and Numerical Methods, Vol.
2, Ch. VI or Reed & Simon: Functional Analysis
Example 8.1 (Weak convergence in Lp ). Consider (un ) ⊂ Lp (Ω) for 1 ≤ p <
∞. Then Z Z
p
un * u ∈ L ⇐⇒ un ϕ dx → uϕ dx ∀ϕ ∈ Lq
Ω Ω
where 1/p + 1/q = 1. This is equivalent to
Z Z
kun kLp ≤ C ∀n and un dx → u dx ∀M ⊂ Ω
M M
1 ∞
Remark 8.1. Note, that the dual of L is L . Therefore one typically considers
weak convergence of sequences in L1 , but the weak–∗ convergence in L∞ .

Let Ω ⊂ R be bounded. Consider the 2π–periodic function u(t) = sin t and the
family uε (t) = u(t/ε) = sin t/ε for ε > 0. Then the function uε is oscillating
with period 2πε and frequency 1ε . Therefore the functions in this family are
called rapidly oscillating. It is clear, that uε is bounded (e.g. in any Lp (Ω))
independently of ε. But uε does not converge strong in any Lp , i.e. there
exists no u0 ∈ Lp such that kuε − u0 kLp → 0.

68
Example 8.2. To construct the weak limit ofRthis sequence, we R b consider an
b
arbitrary interval [a, b] ∈ Ω and compute limε→0 a uε dx = limε→0 a sin x/ε dx.
We introduce y = x/ε and split the integration interval y ∈ [a/ε, b/ε] in two
parts: Let I1 = [a/ε, a/ε + 2kπ] and I2 = [a/ε + 2kπ, b/ε], where k is the largest
integer such that a/ε + 2kπ < b/ε. Then I1 contains k full periods of the function
uε and the remaining interval I2 has a length < 2π. Hence we get
Z b Z b/ε Z Z
x
sin dx = ε sin y dy = ε sin y dy +ε sin y dy = O(ε) ,
a ε a/ε I1 I2
| {z } | {z }
=0 =O(1)

and we can conclude


x
* 0 weakly in Lp .
sin
ε
Loosely speaking, in the weak limit the oscillations of the function uε = sin x/ε
cancel out and only the zero average remains.

In the more general situation, we have the following result:

Theorem 8.2. Let 1 ≤ p ≤ ∞ and u ∈ Lp ([0, T ]) be a T –periodic function. Let


uε (x) = u(x/ε) for ε > 0. Then, if p < ∞, we have
Z T
1
uε * A[u] := u(y) dy weakly in Lp
T 0

as ε → 0. If p = ∞, then

uε * A[u] weak–∗ in L∞ .

Proof. The main steps in the proof are

• Show, that kuε k ≤ C independent of ε.

• Then, we get the existence of a weakly (p < ∞) or weak–∗ (p = ∞)


convergent subsequence of uε .

• Identify the limit of this subsequence with A[u] and show that each conver-
gent subsequence has the same limit.

A detailed proof can be found in [Cioranescu & Donato: An Introduction to


Homogenization] or [Braides: Γ–Convergence for Beginners].

69
8.2 Introductory Example

After this short excursion to functional analysis, we come back to our starting
point:
ü + k 2 (t/ε)u = 0, u(0) = 1, u̇ = 0 .
With the help of Thm. 8.2 we have established the following convergence
Z T
2 T
1 d 2 2 2 u
u k (t/ε) dt → A[k ] ,
0 2 dt 2 0
1
RT
where A[k 2 ] = T 0
k 2 (y) dy denotes the average of k. Hence we get for the
energy
1 2  1
u̇ (T ) + A[k 2 ]u2 (T ) = A[k 2 ] ;
2 2
p
a result, which reminds us of a harmonic oscillator with frequency ν = A[k 2 ].
Let’s introduce the following two time scales t1 = t and t2 = t/ε and seek the
multi–scale expansion

u(t; ε) ∼ u0 (t1 , t2 ) + εu1 (t1 , t2 ) + ε2 u2 (t1 , t2 ) .

In the leading O(ε−2 )–order we get

∂22 u0 = 0 u0 (t1 , t2 ) = A(t1 )t2 + B(t1 ) ,

with the boundary conditions B(0) = 1 and A(0) = 0. Now consider a fixed time
t and let ε → 0, i.e. t2 → ∞. In this situation the fast variable should only have
a bounded influence, i.e. u0 (t1 , t2 ) ≤ M (t1 ) for all t2 , and therefore we require,
that Z
1 t2
lim u0 (t1 , s) ds < ∞ ,
t2 →∞ t2 0

which means nothing else but the average over the fast variable is bounded. Due
to this condition, we deduce A(t1 ) ≡ 0 and u0 (t1 , t2 ) = u0 (t1 ).
In the next order we get

O(ε−1 ) : ∂22 u1 = 0 u1 (t1 , t2 ) = C(t1 )t2 + D(t1 ) ,

with the boundary conditions D(0) = 0 and C(0) = −B 0 (0). As before, we


conclude C ≡ 0 and u1 (t1 , t2 ) = u1 (t1 ) as well as B 0 (0) = 0.
Finally we get from the O(ε0 )–problem:

∂22 u2 + B 00 + k 2 (t2 )B = 0 .

70
Integrating over the fast variable yields
Z Z
1 T 1 00 1 T 2
0= . . . dt2 = (∂2 u2 (·, T ) − u2 (·, 0)) + B + B · k (t2 ) dt2 ∀T ,
T 0 T T 0
| {z }
indep. of T

and due to the boundary conditions, we have ∂2 u2 (0, 0) = −D 0 (0). To keep the
expansion well ordered in the sense of generalized asymptotic expansions, we
require that u2 = O(1) and ∂2 u2 = O(1). Therefore, in the limit T → ∞, it
holds that
1
lim (∂2 u2 (t1 , T ) + D 0 (0)) = 0 ,
T →∞ T

and hence we arrive at the homogenized equation

B 00 (t1 ) + ν 2 B(t1 ) = 0, B(0) = 1, B 0 (0) = 0 ,

where the homogenized frequency is given by


Z
2 1 T 2
ν = lim k (s) ds .
T →∞ T 0

Now, the leading term in our two–scale expansion reads as

u0 (t) = cos νt .

Example 8.3. Consider k(t) = 1 + 12 sin t. Then the homogenized frequency


equals Z T
2 9
ν = lim k 2 (s) ds = · · · =
T →∞ 0 8
and hence the homogenized equation ü0 + ν 2 u0 = 0 has the solution
r
9
u0 (t) = cos t.
8

Our next example is already a bit more interesting: Let’s consider the steady
one–dimensional heat equation in a medium with microstructure
h  x  i0
− k u0 = f (8.1)
ε
subject to homogeneous Dirichlet conditions u(0) = u(1) = 0. We assume that
the conductivity k ∈ L∞ (0, 1) is periodic, positive and bounded, i.e. 0 < α ≤ k ≤
β. In this simple setting the solution can easily be obtained by integrating (8.1)
twice.
Theorem 8.3. For 0 < ε < ε0 , the problem (8.1) has a unique solution.

71
Proof. Multiply (8.1) with a testfunction v ∈ H01 (0, 1) and we get
aε (u, v) = (f, v)
R
where the bilinear form a is defined by a(u, v) = k(x/ε)u0 v 0 dx. One can easily
check, that a is continuous and coercive on H01 , hence the Lax–Milgram theorem
applies. Therefore, we get the existence of a unique solution uε ∈ H01 and uε is
bounded by
kuε k ≤ c kf k .

Since the sequence uε is bounded, we get the existence of a subsequence (again


denoted by uε ), which converges to some limit u. But which equation does this
u satisfy? Homogenized equation. The problem is, that
x
k * k weak in e.g. L2
ε
uε * u weak in L2

but we cannot conclude that kuε 6* ku. To illustrate this, consider fε =


sin(x/ε) * 0, but fε2 * 1/2.
To derive the homogenized equation, we seek a two–scale expansion of (8.1)
x
u(x; ε) ∼ u0 (x, y) + εu1 (x, y) + ε2 u2 (x, y), y=
ε
In leading O(ε−2 )–order we get
−∂y [k(y)∂y u0 ] = 0
with the general solution
Z y
ds
u0 (x, y) = c1 (x) + c0 (x) .
k(s)
| 0 {z }
unbdd for y → ∞

To remove the unbounded integral, we conclude c0 (x) ≡ 0 and proceed to the


next order.
O(ε−1 ) : −∂y [k(y) (∂x u0 + ∂y u1 )] − ∂x [k(y)∂y u0 ] = 0
∂y [k(y)∂y u1 ] = −∂y k(y) ∂x u0
with the general solution
Z
ds
u1 (x, y) = b1 (x) + b0 (x) − y∂x u0 .
k(s)
| {z }
=O(y) as y → ∞

72
To cancel the unbounded term, we require
 Z 
1 ds
lim b0 (x) − y∂x u0 = 0 ,
y→∞ y k(s)
and conclude

∂x u0 (x) = k −1 b0 (x) , (∗)

where
Z y
−1 1 ds
k −1 = A[k ] = lim .
y→∞ y 0 k(s)
To finalize the derivation of the homogenized equation, we consider the O(ε0 )–
problem
−∂y [k∂y u2 ] = f + ∂y [k∂x u1 ] + b00
with the solution
Z Z Z
ds s ds
u2 (x, y) = a1 (x) + a0 (x) + ∂x u1 ds + (f + b00 ) .
k(s) k(s)
| {z } | {z }
=O(y) as y → ∞ =O(y 2 ) as y → ∞

To cancel the quadratic growing term, we get the condition

−b00 = f

and hence, together with (∗), we obtain the homogenized equation


 
−∂x k∂x u0 = f, u(0) = 0, u(1) = 0 , (8.2)

with the homogenized or effective conductivity


1
k= . (8.3)
A[k −1 ]
This effective conductivity is the harmonic mean of the function k(y).

8.3 New Concepts of Convergence

Homogenization problems have led to new convergence concepts. In the sequel


we will touch this field and introduce the so called Γ–convergence.
Our problem (8.1) can be re–written in variational form as a minimization prob-
lem: Z 1
1
min J(u) = k (u0 )2 − f u dx . (8.4)
0 2

73
To see this, we vary u in some direction v and get
Z
J(u + v) − J(u) = kv 0 u0 − f v dx + O(v 2 )
Z
= − v [(ku0 )0 + f ] dx + O(v 2 ) .

Therefore a minimum of the functional J certainly is a solution of (8.1). Indexing


the functional with ε, i.e.
Z 1
1
Jε (u) = k(x/ε) (u0 )2 − f u dx (8.5)
0 2
we can reformulate our homogenization problem as follows:

Find a limiting function u and a limiting functional J0 ,


(8.6)
such that uε → u and u minimizes J0 .

This question is the motivation for the following definition of Γ–convergence.

Definition 8.3 (Γ–convergence). Let X be a metric space and consider a


sequence Jε : X 7→ R of functionals. We say, that Jε Γ–converges to J0 : X 7→ R,
if for all u ∈ X we have

i) J0 (u) ≤ lim inf Jε (uε ) ∀(uε ) → u


ε→0
ii) ∃(uε ) → u : J0 (u) ≥ lim sup Jε (uε ) .
ε→0

The functional J0 is called the Γ–limit of (Jε )ε and denoted by J0 = Γ– limε Jε .


The sequence in ii) is also called recovery sequence, since J0 (u) = limε Jε (uε ).

Remark 8.2. The Γ–limit J0 is a lower semi–continuous function, i.e J0 (u) ≤


lim inf ε J0 (uε ) for any sequence uε → u.

Definition 8.4 (Equi–coercive). A sequence Jε : X 7→ R is said to be equi–


coercive on X, if there exists a compact set K ⊂ X (independent of ε), such
that
inf Jε (u) = inf Jε (u) . (8.7)
u∈X u∈K

The following theorem now states, that equi–coerciveness and Γ–convergence


ensure the convergence of the minimizers.

Theorem 8.4. Let Jε : X 7→ R be equi–coercive and J0 = Γ– lim Jε . Then

74
1. the minima of Jε converge to the minima of J0 , i.e.
 
min J0 (u) = lim inf Jε (v) , (8.8)
u∈X ε→0 v∈X

2. the minimizers of Jε converge to those of J0 , i.e.


 
uε → u and lim Jε (uε ) = lim inf Jε (v) =⇒ u is minimizer of J0 .
ε→0 ε→0 v∈X
(8.9)

In the sequel, we will use the concept of Γ–convergence, to show that the solutions
of the problem
h  x  i0 Z
0 1 x 0 2
− k u =f ⇐⇒ min Jε (u) = k (u ) − f u dx (8.10)
ε 2 ε
converge to the solution u of the homogenized problem
Z
 0 0 1
− ku = f ⇐⇒ min J0 (u) = k(u0 )2 − f u dx (8.11)
2
where the homogenized conductivity k is given by
Z −1
ds
k= . (8.12)
k(s)
To apply Thm. 8.4, we have to show, that J0 is the Γ–limit of Jε , i.e.
J0 (u) ≤ lim inf Jε (uε ) ∀uε → u and ∃uε → u : J0 (u) = lim Jε (uε ) .
ε ε

Since the continuous functionals (8.5) are convex w.r.t u0 and radially unbounded,
i.e. Jε (u) → ∞ for ku0 k → ∞, they are also equi–coercive. Hence the Γ–
convergence of Jε to J0 is sufficient for the convergence of the solutions of the
underlying differential equations. It remains to show, that the lim inf–inequality
holds and we have to construct a recovery sequence. We will first show them for
the special case of a linear function u(x) = αx ∈ W 1,p . The arguments can then
be extended to piecewise linear functions and by density to the whole space W 1,p .
So let u(x) = αx and let (uε )ε ⊂ W 1,p be a sequence converging to u. Then we
have
Z
1 2
Jε (uε ) = k [u0 + (uε − u)0 ] − f u − f (uε − u) dx
2
Z Z Z
1 0 2 0 0 1
= k(u ) − f u dx + ku (uε − u) − f (uε − u) dx + k(uε − u)02 dx
2 2
Z Z Z
α2 0 1
= k − f u dx + α k(uε − u) − f (uε − u) dx + k(uε − u)02 dx ,
2 2

75
as well as
Z 2
α
J0 (u) = k − f u dx .
2
Using
R the fact,R that the harmonic mean is smaller than the arithmetic one,
i.e. k dx ≤ k dx, we get the inequality
Z Z
0 1
Jε (uε ) ≥ J0 (u) + αk(uε − u) − f (uε − u) dx + k(uε − u)02 dx
2
| {z }
≥0
Z
≥ J0 (u) + αk(ε −u)0 dx − f (uε − u) dx ∀uε → u .
| {z }
∈(W 1,p )0 hence →0 for ε→0

Passing to the lim inf, we get the desired estimate

lim inf Jε (uε ) ≥ J0 (u) .


ε→0

To finish the proof of the Γ–convergence, we have to construct a recovery sequence


(uε )ε → u such that J0 (u) = lim Jε (Uε ). To simplify the construction, we focus
on a piecewise constant and periodic conductivity k, i.e.
(
k1 0 ≤ s < 12
k(s) = periodically repeated.
k2 21 ≤ s < 1

Again, let u(x) = αx and set uε = u + vε , where


(
a 0 ≤ x/ε < 21 k1 − k 2
vε0 (x) = 1 x
periodically repeated and a = −α .
−a 2 ≤ ε < 1 k1 + k 2

Then it holds, that max |vε (x)| ≤ 12 aε = O(ε) and


Z
1
Jε (uε ) = k(u0 + vε0 )2 − f u − f vε dx
2
Z Z Z
1 02 1 02 0 0 0 2
= ku − f u dx + (k − k)u + 2ku vε + k(vε ) dx − f vε dx
2 2
| {z }
:=G

Using k = 2k1 k2 /(k1 + k2 ) and the definition of vε , we get

α2  a2
G= k1 + k2 − 2k + αa (k1 − k2 ) + (k1 + k2 ) = 0
2 2

76
and hence
Z
Jε (uε ) = J0 (u) − f vε dx → J0 (u) .
| {z }
→0

Note, that the recovery sequence shows the typical two–scale behavior, which we
already obtained from the asymptotic expansion.
The concept of Γ–convergence was introduced in the 70s as the most flexible and
natural notion of convergence for minimizers of variational problems. Besides Γ–
convergence, there are also other concepts of convergence used in homogenization
problems like two–scale convergence, G– and H–convergence.

8.4 Flow in Porous Media

As a final example for homogenization, we consider the classical problem of flow


in a porous medium, e.g. the flow of groundwater in soil. The macroscopic domain
Ωε ⊂ Rn , n = 2, 3, consists of a solid part Sε (rocks, sand) and the fluid part Fε
(water phase). The porosity ϕ is then defined as

vol Fε
ϕ= ∈ [0, 1] .
vol Ωε
Typically, the macro–domain Ωε has a quite complex geometry. To simplify the
situation, we assume Ωε to be generated by a periodic repetition of elementary
cells εY , with characteristic length ε. One characteristic cell Y consists of a solid
part S and a fluid part F , see Fig. 8.1
PSfrag replacements
Ωε

Y
F
x
y=
ε
S

zoom

εY ε 1

Figure 8.1: Sketch of the geometry of a porous medium.

77
Inside the domain Ωε , we consider the equations of slow, viscous flow, i.e. Stokes
equation:

∇ · uε = 0 , (8.13a)
∇pε − ε2 ν ∆ uε = 0 in Fε , (8.13b)
uε = 0 on ∂Sε . (8.13c)

The ε2 –scaling for the viscosity has basically two reasons: Argueing
√ from physics,
this scaling assumes, that the size of the boundary layers (= viscosity) is of the
same order as the size of the particles (= O(ε)). Mathematically, this scaling is
the only one, which leads to a non–trivial limit as ε → 0. If the boundary layers
are too thick (viscosity too large), the boundary layers fill the whole domain and
we have to solve the Stokes equation in the entire domain. On the other hand, if
the boundary layers are too small (viscosity too low), the fluid does not feel the
presence of the solid part Sε .
To derive the homogenized equation describing the averaged flow inside the do-
main Ω, we use the following two scales: the macro–variable x ∈ Rn and the
micro–variable y = xε ∈ Rn . For the velocity u and the pressure p, we propose
the asymptotic ansatz

uε (x) ∼ u0 (x, y) + εu1 (x, y) periodic in y ,


pε (x) ∼ p0 (x, y) + εp1 (x, y) periodic in y .

Then, the O(ε−1 )–problem reads as

∇y · u 0 = 0 , (8.14a)
2n
∇y p 0 = 0 for (x, y) ∈ Ω × F ⊂ R . (8.14b)

From this, we can conclude, that p0 (x, y) ≡ p0 (x), i.e. p0 depends only on the
macro–variable.
Proceeding to the next order, we find

∇x · u 0 + ∇ y · u 1 = 0 , (8.15a)
2n
∇y p1 − ν ∆y u0 = −∇x p0 for (x, y) ∈ Ω × F ⊂ R . (8.15b)

The second equation (8.15b) is a forced Stokes equation. P


Exploiting the linearity
of the equation, we re–write the focing term as ∇x p0 = ∂xi p0 · ei (i.e. compo-
n
nents of the gradient vector). Now let wi ∈ R and πi ∈ R be the solutions of
the cell (auxiliary) problem

∇y · w i = 0 , (8.16a)
∇ y πi − ∆ y w i = e i in F ⊂ Y , (8.16b)
wi = 0 on ∂S . (8.16c)

78
Note, that this cell problem is solved in the elementary cell Y independent of the
macro–variable x. The solution of the cell problem reflects the microstructure
of the problem. With the help of the auxiliary variables wi and πi , we can
reconstruct the solution of (8.14) and (8.15b) by linear superposition:
X
p1 (x, y) = − ∂xi p0 (x) · πi (y) , (8.17)
X 1
u0 (x, y) = − ∂xi p0 (x) · wi (y) . (8.18)
ν
Averaging now the velocity u0 (x, y) over the micro–variable y we get the macro-
scopic seapage velocity
Z Z
1X
u0 (x) = u0 (x, y) dy = − ∂ xi p 0 · wi (y) dy
Y ν i Y | {z }
=wi,j (y)
K
=− ∇p0 , (8.19)
ν

where the permeability tensor K ∈ R3×3 is defined as


Z
Kij = wij (y) dy . (8.20)
Y

Note, that the computation of the permeability tensor K requires the solution of
the cell problem (8.16). The permeability tensor K contains all the information
about the microstructure of the problem. A small computation shows, that the
permability tensor K is symmetric and positive definite.
Summarizing the equations, we get the following homogenized system
1
u(x) = − K∇p(x) , (8.21a)
ν
∇·u=0 in Ω . (8.21b)

These equations are known as Darcy’s Law. They relate the averaged seapage ve-
locity to the pressure drop and were found first by Henry Darcy (experimentally).
For more information on homogenization and porous media flow, see [Hornung:
Homogenization and Porous Media].
Related problems lead to Brinkmann’s Law, describing the flow in a porous
medium, where the size of the elementary cell is O(ε), but the diameter of the
obstacles is of order O(ε3 ) in a 3–dimensional porous medium.

79
9 Perturbation Methods in Fluid Mechanics

9.1 Prandtl’s Boundary Layer Theory

The figure 9.1 shows an image of a bird in a windtunnel. Note the different
behavior of the streamlines close to the bird compared to the flow away from the
bird. We consider a two–dimensional, stationary incompressible flow around an

Figure 9.1: Flow around a bird in a windtunnel.

object. This situation is perfectly described by the Navier–Stokes equations

ux + v y = 0 (9.1a)
1 1
uux + vuy + px = (uxx + uyy ) (9.1b)
ρ Re
1 1
uvx + vvy + py = (vxx + vyy ) , (9.1c)
ρ Re
where (u, v) is the velocity in the cartesian coordinates (x, y) and p denotes the
pressure. The dimensionless parameter Re = U L/ν is called the Reynolds num-
ber, where U is some characteristic velocity, L a reference length and ν denotes
the (kinematic) viscosity of the fluid. For our bird in the windtunnel, we get

U ≈ 10 m s−1
L ≈ 0.15 m Re = 105 .
ν ≈ 15 · 10−6 m2 s

80
Hence we can assume, that 1/Re = ε  1. One striking feature of high–Reynolds
number flows is the appearance of turbulence; this can been seen by the vortices
appearing in the flow close to the upper part of the bird. Mathematically, the
limit ε → 0, i.e. Re → ∞ or ν → 0 is a singular one. We pass from the second
order parabolic (or elliptic in the stationary case) Navier–Stokes equations to the
first order hyperbolic Euler equations modelling inviscid flow
ux + v y = 0 (9.2a)
1
uux + vuy + px = 0 (9.2b)
ρ
1
uvx + vvy + py = 0 . (9.2c)
ρ
As in most singular perturbation problems, a loss of boundary conditions appears.
For the Navier–Stokes equations, we can prescribe no–slip conditions, i.e (u, v) =
0 on ∂Ω, whereas for the Euler equations, we can only prescribe the normal
component of the flow (u, v) · n = 0. Hence we expect some sort of boundary
layer close to the surface of the object.

To simplify the problem,PSfrag


we assume, y, v
replacements ∂Ω
that the surface of the object is aligned
with the x–coordinate and that the y–
coordinate is normal to the surface.
x, u
To zoom into the boundary layer, we use the following scaling:
y = εα η , v = εα w
where α > 0. We choose the same scaling for space and velocity, to leave the
(here not present) time scale unchanged. We end up with
ux + w η = 0 (9.3a)
1
uux + wuη + px = εuxx + ε1−2α uηη (9.3b)
ρ
1
epsα (uwx + wwη ) + ε−α pη = ε1+α wxx + ε1−α wηη , (9.3c)
ρ
From the last equation, we can directly conclude, that pη = 0 and hence p =
p(x) = pEuler (x, 0), i.e. the pressure inside the boundary layer is given by the
outside pressure. Inspecting the different terms, we find that the only reasonable
scaling is α = 1/2. With this scaling, we get Prandtl’s boundary layer equations
ux + w η = 0 (9.4a)
1
uux + wuη − uηη = pEuler (x, 0) . (9.4b)
ρ

81
This is a parabolic system for the tangential velocity u and the normal velocity
w is coupled via the continuity equation. The thickness yb of the boundary layer
can be estimated as
√ 1 √
yb ∼ ε = √ ≈ ν .
ε

Prandtl’s boundary layer theory resolved d’ Alembert’s paradox which states, that
a body experiences no drag in a potential flow.

9.2 Stokes Flow past Cylinders and Spheres

The low–Reynolds number flow past a cylinder or sphere can be described by the
following equations

∇·u=0 (9.5a)
1
(u · ∇u) = (−∇p + ∆ u) in Rn \ B1 (0) (9.5b)
ε
u=0 on ∂B1 (0) (9.5c)
u=1 for kxk → ∞ , (9.5d)

where ε  1 models the small Reynolds number.


We introduce polar coordinates (r, ϕ) in the 2–d case or (r, ϕ, θ) in the 3–d situ-
ation, where we assume symmetry w.r.t. ϕ. To get rid of the continuity equation

 1 ∂r (ru) + 1 ∂ϕ v

n=2,
∇ · u = r1 
r
1
 ∂r r 2 u +
 ∂θ (sin θ v) n = 3 ,
r2 r sin θ
we introduce a streamfunction Ψ by
 
 1 ∂ϕ Ψ
 ∂ r Ψ for n = 2 ,
u= r 1 v=− 1

 ∂θ Ψ  ∂r Ψ for n = 3 .
r 2 sin θ r sin θ

Plugging the streamfunction into the momentum equation (9.5b) and differenti-
ating once again, we get in leading order the biharmonic equation
 2
 2 1 1 2
 ∂r + ∂r + 2 ∂ϕ Ψ
 n=2,
4 r r
0=∇ Ψ=   2 (9.6)

 2 sin θ 1
 ∂r + ∂θ ∂θ Ψ n=3.
r2 sin θ

82
The boundary condition on the cylinder/sphere reads as

Ψ(1, ϑ) = ∂r Ψ(1, ϑ) = 0 (9.7)

for ϑ = ϕ in 2d or ϑ = θ in 3d. The uniform flow at infinity gives the additional


condition (
r sin ϑ n=2
Ψ(r, ϑ) ∼ 1 2 2 as r → ∞ . (9.8)
2
r sin ϑ n = 3
To solve the 3d problem, we seek a solution using separation of variables, i.e. Ψ =
sin2 θ f (r). This leads to
 
Ψ = sin2 θ ar 4 + br 2 + cr + d/r .

Due to the uniform flow at infinity (Eqn. (9.8)), we get a = 0 and b = 12 . The
boundary conditions (9.7) yield c = −3/4 and d = 1/4 and we obtain
 
1 2 1
Ψ(r, θ) = 2r − 3r + sin2 θ . (9.9)
4 r

Now let’s consider the 2d problem. Again we seek a solution using separation of
variables Ψ = sin ϕ g(r). This leads to
 
Ψ = sin ϕ ar 3 + br ln r + cr + d/r .

Due to the boundary condition (9.7) this reduces to


 
k 3 1 1+k1
Ψ ∼ C − r + (1 + 2k)r ln r − r + sin ϕ
2 2 2 r

and the uniform flow at infinitiy suggests k = 0. This leaves us with


 
1 11
Ψ ∼ C r ln r − r + sin ϕ .
2 2r

Clearly, this solution does not satisfy the free flow condition (9.8) at infinity,
since the r ln r–term is more singular at infinity than the free flow.
The nonexistence of a solution of Stokes’ equation for infinite planar flow past a
cylinder (or any other body!) is known as Stokes’ paradox.
To understand, where this Stokes paradox arises from, let’s consider the convec-
tive terms, that are neglected in the Stokes equation (9.6)
1
convective term uur ∼ as r → ∞
r2

83
and a typical viscous term reads as
1 1
viscous term ∆u ∼ 3 as r → ∞ .
ε εr
Hence the ratio of the neglected terms to those retained in Stokes’ equation is
convective
= O(εr) as r → ∞ .
viscous
Close to the sphere (r = O(1)) this ratio is also small, but far away from the
body, the neglected convective terms get comparable or even larger than the
viscous one, no matter how small the Reynolds number ε is. Hence the Stokes
equation becomes invalid, when εr = O(1). Using the definition ε = Re = U/ν,
this appears at a length scale ν/U , which is called the viscous length.
The same argument holds a fortiori for the 2d situation. Here the ratio between
the convective and viscous terms is given by
convective
= O(εr ln r) as r → ∞ .
viscous
This dominance of the convective terms at large distances is the source of the
singular behaviour of the Stokes equation (as an approximation to the Navier–
Stokes equations) at infinity. In the 3d situation, this singular behaviour is not
visible, since the first approximation (9.9) is well–behaved at infinity. In the re-
gion of non–uniformity O(εr) = 1, the velocity has already effectively reached
the free–stream velocity; hence we can impose the boundary condition. But this
is just an exceptional circumstance. If one wants to construct a second approx-
imation to the low–Reynolds number flow past a sphere by iteration (plug (9.9)
into the left hand side of (9.5b)), one faces Whitehead’s paradox : The second
approximation does not exist!
A solution to the Stokes and Whitehead paradox was proposed by Oseen. The
trick is not to neglect hte convective terms completely, but to include them in
a linearized form. The x–component of the momentum equation reads in the
Navier–Stokes equation as
1
uux + vuy + wuz + px = ν ∆ u
ρ
and it is approximated in the Oseen equation by
1
U ux + p x = ν ∆ u
ρ
where U denotes the free stream velocity. The Oseen equation for the stream-
function reads as
(∆ −ε∂x ) ∆ Ψ = 0
It can now be shown, that Oseen’s equation provides a uniformly valid first ap-
proximation to the low–Reynolds number flow past a cylinder or sphere.

84
9.3 Derivation of Euler Equations from Kinetic Theory

In 1900, Hilbert posed his famous 23 problems. We are going to deal with his
6th problem, which concerned the passage from mircoscopic, molecular motion
to the laws of fluid dynamics.
Consider a rarefied gas (very low density, or large mean free path). In this situa-
tion, a single molecule travels along a straight line, until it eventually hits another
one. If this happens, it is scattered in some different direction. In 2d, this is sim-
ilar to playing billard. In statistical physics, one describes such an ensemble of
molecules (or billard balls) by a distribution function f : R+ ×Rn ×Rn 7→ R, where
f (t, x, v) denotes the probability of finding at time T and position x a molecule
with velocity v. The time evolution of this distribution function is given by the
Boltzmann equation
∂t f + v · ∇x f = Q(f ) ,
where the collision operator Q describes the influence of scattering (collisions
between the billard balls). The v–moments of the distribution function f allow
for a macroscopic interpretation
Z
f (t, x, v) dv = ρ(t, x) density,
3
Z R
v f (t, x, v) dv = ρ(t, x) u(t, x) (bulk) velocity,
R 3
Z
(v − u) ⊗ (v − u) f (t, x, v) dv = ρkT I = pI kinetic pressure tensor,
R3

where k is the Boltzmann constant, T the temperature of the gas and p the
pressure.
In the absence of collisions, the Boltzmann equation just describes the prop-
agation of the distribution with velocity v, i.e. f (t, x, v) = f (0, x − vt, v) =
f0 (x − vt, v).
For the billard game (as well as the motion and scattering of molecules), the
Boltzmann equation reads as

Df := ∂t f + v · ∇x f = J(f, f )

where the collision term is given by


Z
J(f, f ) = M [w1 , w2 |v1 , v2 ] {f (w1 )f (w2 ) − f (v1 )f (v2 )} dw1 dw2 dv2 .

This collision term describes the number of collisions between particles with ve-
locities v1 and v2 that have velocities w1 and w2 after the collision.

85
Now, let’s consider a large box Ωε with side length O(1/ε). This box shall consider
a large number of molecules n = O(1/ε3 ) with distribution function f , i.e.
Z Z  
1
f (t, x, v) dv dx = O .
Ωε R3 ε3
Hence, we expect to find roughly one molecule per unit volume, i.e. the mean free
path a molecule can travel before it hits another one is 1.
Next, we rescale our box to length 1, i.e. ξ = εx ∈ Ω. To keep the velocity the
same, we also rescale the time by τ = εt and introduce the distribution function
Z Z
fε (τ, ξ, v) = f (t, x, v), fε (τ, ξ, v) dv dξ = 1 .
Ω R3

The rescaled distribution fε satisfies the following Boltzmann equation


1
Dfε := ∂τ fε + v · ∇ξ fε = J(fε , fε ) .
ε
The parameter ε  1 can be interpreted as the mean free path; now we find
1/ε3 molecules in a volume of size 1, hence each molecule can travel a length of
ε before hitting a second one.
Assuming a Poincare expansion for the distribution function fε ∼ f0 +εf1 and us-
ing the quadratic structure of J, i.e. J(fε , fε ) ∼ J(f0 , f0 )+2εJ(f0 , f1 )+ε2 J(f1 , f1 ),
we obtain the following sequence of problems
O(ε−1 ) 0 = J(f0 , f0 ) (9.10)
1
O(ε0 ) Df0 = J(f0 , f1 ) (9.11)
2
To solve the O(ε−1 )–problem, we have to determine the equilibrium distribution
for the collision term J. This equilibrium distribution is given by the Maxwell–
distribution
( )
ρ(t, x) kv − u(t, x)k2
f0 (t, x, v) = exp (9.12)
[2πkT (t, x)]−3/2 2kT (t, x)

where ρ is the density, u0 the (bulk) velocity and T the temperature of the
molecular gas. This can be seen as follows: J(f0 , f0 ) = 0 is satisfied, if
f (w1 ) f (w2 ) = f (v1 ) f (v2 ) (9.13)
holds for all in– and outgoing velocities v1 , v2 and w1 , w2 . Due to momentum
conservation we require v1 + v2 = w1 + w2 and energy conservation requires
v12 + v22 = w12 + w22 . Taking the logarithm of (9.13) one can show that
ln f (t, x, v) = α(t, x) + β(t, x)v + γ(t, x)v 2 . (9.14)

86
Coming back to the Maxwell–distribution (9.12), we still have to determine equa-
tions for ρ, u and T . They are provided by the next order of the expansion:
1
Df0 = J(f0 , f1 ) (9.15)
2
For a given f0 this is a linear equation for f1 . However, we already know, that
J(f0 , ·) has a kernel. Due to (9.14), the functions ϕ0 = 1, ϕi = vi (i–th component
of the velocity) for i = 1 . . . 3 and ϕ4 = 21 v 2 provide a basis of this kernel.
Employing the Fredholm alternative saying, that the inhomogeneous equation
is solvable, iff the the right hand side is orthogonal to the the kernel of the
homogeneous one, we get the conditions
Z
Df0 · ϕi dv = 0 for i = 0 . . . 4

or
Z
(ft + v · ∇x f ) · 1 dv = 0 ρt + ∇x · (ρu) = 0 (9.16)
Z
(ft + v · ∇x f ) · vi dv = 0 (ρu)t + (ρuui )x + px = 0 (9.17)

Hence the solvability conditions for (9.15) lead to the Euler equations of gas
dynamics; Eqn. (9.16) is the continuity equation and Eqn. (9.17) the momentum
equation. The product with the fifth invariant ϕ4 = v 2 /2 leads to the energy
equation.

The end.

87

You might also like