Intro To Method of Multiple Scales
Intro To Method of Multiple Scales
Per Jakobsen
January 8, 2014
Contents
1 Introduction
22
28
33
36
37
37
42
45
Introduction
Therefore, the description of the method appears quite different in the various
sources, depending on the views of the authors. In these lecture notes we decribe the method in a way that is particularly well suited to the way it is used
in nonlinear optics and laser physics. The source that is closest to our approach
is [11].
We do not assume that the reader has had any previous exposure to perturbation methods. The lecture notes therefore starts off with three sections where
the basic ideas of asymptotic expansions are introduced and illustrated using
algebraic equations. The application to differential equations starts in section
four where we use regular perturbation expansions to find approximate solutions
to ODEs. In section five we introduce the method of multiple scale and apply
it to weakly nonlinear ODEs. We move on to weakly nonlinear dispersive PDEs
in section six and finally move on to the Maxwell equations in section seven.
Several exercises involving multiple scales for ODEs and PDEs are included in
the lecture notes.
2.1
(1)
(2)
(3)
x1 = 1
(4)
Let us focus on x1 and let us assume that the perturbed problem has a solution
in the form of a perturbation expansion
x() = a0 + a1 + 2 a2 + ...
3
(5)
where a0 = 1. Our goal is to find the unknown numbers a1 , a2 , ... These numbers
should have a size of order 1. This will ensure that a1 is a small correction to
a0 , that 2 a2 is a small correction to a1 and so on, all in the limit of small . As
we have stressed before, maintaining the ordering of the perturbation expansion
is the one and only unbreakable rule when we do perturbation calculations. The
perturbation method now proceeds by inserting the expansion (5) into equation
(1) and collecting terms containing the same order of .
(a0 + a1 + 2 a2 + ...)2 (a0 + a1 + 2 a2 + ...) + = 0
a20
+ 2a0 a1 +
(a21
+ 2a0 a2 ) a0 a1 2 a2 .. + = 0
a20
(6)
Since a1 , a2 , .. are all assumed to be of order 1 this equation will hold in the
limit when approach zero only if
a20 1 = 0
(7)
2a0 a1 a1 + 1 = 0
2a0 a2 + a21 a2 = 0
We started with one nonlinear equation for x and have ended up with three
coupled nonlinear equations for a0 , a1 and a2 . Why should we consider this to
be progress? It seems like we have rather substituted one complicated problem
with one that is even more complicated!
The reason why this is progress is that the coupled system of nonlinear
equations has a very special structure. We can rewrite it in the form
a0 (a0 1) = 0
(8)
(2a0 1)a1 = 1
(2a0 1)a2 = a21
The first equation is nonlinear but simpler than the perturbed equation (1), the
second equation is linear in the variable a1 and that the third equation is linear
in the variable a2 when a1 has been found. Moreover the linear equations are all
determined by the same linear operator L() = (2a0 1)(). This reduction to
a simpler nonlinear equation and a sequence of linear problems determined by
the same linear operator is what makes (8) essentially simpler than the original
equation (1), which does not have this special structure. The system (8) is
called a perturbation hierarchy for (1). The special structure of the perturbation
hierarchy is key to any successful application of perturbation methods, whether
it is for algebraic equations, ordinary differential equations or partial differential
equations.
(9)
a1 = 1
(10)
a2 = 1
and our perturbation expansion to second order in is
x() = 1 2 + ...
(11)
For this simple case we can solve the unperturbed problem directly using the
solution formula for a quaderatic equation. Here are some numbers
0.001
0.01
0.1
Exact solution
0.998999
0.989898
0.887298
Perturbation solution
0.998999
09989900
0.890000
We see that our perturbation expansion is quite accurate even for as large
as 0.1.
Let us see if we can do better and find an even more accurate approximation
by extending the perturbation expansion to higher order in . In fact let us take
the perturbation expansion to infinite order in .
x() = a0 + a1 + 2 a2 + ... = a0 +
n an
(12)
n=1
n an )(a0 +
n=1
m am ) a0
m=1
n an + = 0
n=1
a20
a0 +
X
(2a0 1)ap +
p
p
p=1
p=2
p1
X
!
am apm
+=0
m=1
a20
a0 + ((2a0 1)a1 + 1) +
(2a0 1)ap +
p=2
p1
X
!
am apm
= 0 (13)
m=1
(14)
(2a0 1)a1 = 1
(2a0 1)ap =
(15)
p1
X
am apm ,
m=1
p=2
The right-hand side of the equation for ap only depends on aj for j < p. Thus
the perturbation hierarchy is an infinite system of nonlinear equations that
is coupled in such a special way that we can solve them one by one. The
perturbation hierarchy truncated at order 4 is
(2a0 1)a1 = 1
(2a0 1)a2 =
(16)
a21
(17)
a2 = 1
a3 = 2
a4 = 5
For = 0.1 the perturbation expansion gives
x(0.1) = 0.8875...
(18)
x(0.1) = 0.8872...
(19)
we are clearly getting closer. However we did not get all that much in return
for our added effort.
Of course we did not actually have to use perturbation methods to find
solutions to equation (1) since it is exactly solvable using the formula for the
quaderatic equation. The example however illustrate many general features of
perturbation calculations that will appear again and again in different guises.
2.2
(20)
This is our perturbed problem P(). For this case perturbation methods are
neccessary since there is no solution formula for general polynomial equations
of order higher than four. The unperturbed problem P(0), is
x5 2x = 0
It is easy to see that the unperturbed equation has a real solution
4
x a0 = 2
6
(21)
(22)
(23)
a50
5a40 (a1
a50
2a0 + (1 +
5a40 a1
(24)
(25)
2)a1 = 1
2)a2 =
(26)
10a30 a21
(27)
Observe that the first equation in the hierarchy, for a0 , is nonlinear in whereas
the equations for ap are linear in ap for p > 0. All the linear equations are
defined in terms of the same linear operator L() = (5a40 2)(). This is the
same structure that we saw in the previous example. If the unperturbed problem
is linear the first equation in the hierarchy will also be linear. The perturbation
hierarchy is easy to solve and we find
a1 =
a2 =
1
1
=
5a40 2
8
(28)
5 8
10a30 a21
=
4
2a0 2
256
1
548 2
4
x() = 2
+ ...
8
256
(29)
get perturbation expansions for all the solutions of (20), and the effort was not
much larger than for the quaderatic equation.
If we can find perturbation expansions for all the solutions of a problem P(),
by starting with solutions of the unperturbed problem P(0), we say that P() is
a regular perturbation of P(0). If the perturbation is not regular it is said to be
singular. This distinction applies to all kinds of perturbation problems whether
we are looking at algebraic equations, ordinary differential equations or partial
differential equations. Clearly, for polynomial equations a neccessary condition
for being a regular perturbation problem is that P() and P(0) have the same
algebraic order. This is not always the case as the next example shows.
2.3
(30)
(31)
(32)
Let us find a perturbation expansion for a solution to (30) starting with the
solution (32) of the unperturbed problem.
x() = a0 + a1 + 2 a2 + ...
(33)
(a20
a0 1 + (a1 +
a20 )
(34)
a20
a2 = 2a0 a1
(35)
(36)
a1 = 1
a2 = 2
and the perturbation expansion for the solution to (30) starting from the solution
x = 1 to the unperturbed problem (31) is
x() = 1 + 22 + ...
(37)
In order to find a perturbation expansion for the other solution to the quaderatic
equation (30), the unperturbed problem (31) is of no help.
However, looking at equation (30) we learn something important: In order
for a solution different from x = 1 to appear in the limit when approaches
zero, the first term in (30) can not approach zero. This is only possible if x
approaches infinity as approaches zero.
Inspired by this, let us introduce a change of variables
x = p y
(38)
12p 2
y +
y1=0
y +
p1
2p1
=0
(39)
The idea is now to pick a value for p, thereby defining a perturbed problem
P(), such that P(0) has a solution of order one. For p > 1 we get in the limit
when approches zero the problem
y2 = 0
(40)
which does not have any solution of order one. One might be inspired to choose
p = 21 . We then get the equation
2
y + y = 0
(41)
which in the limit when approaches zero turns into
y=0
(42)
(43)
(44)
a20
(46)
(47)
(2a0 + 1)a1 = 1
(2a0 + 1)a2 = a21
We observe in passing that the perturbation hierarchy has the special structure
we have seen earlier. The solution to the perturbation hierarchy is
a1 = 1
(48)
a2 = 1
and the perturbation expansion to second order in is
y() = 1 + 2 + ...
(49)
(50)
Even for as large as 0.1 the perturbation expansion and the exact solution,
xE (), are close
x() = 1 1 + + ... 10.900..
1 1 + 4
10.916..
xE () =
2
(51)
At this point I need to be honest with you: There is really no general rule
for how to find the right transformations. Skill, experience, insight and sometimes even dumb luck is needed to succeed. This is one of the reasons why I
prefer to call our subject perturbation methods and not perturbation theory.
Certain classes of commonly occuring singular perturbation problems has however been studiet extensively and rules for finding the correct transformations
has been designed. In general what one can say is that some kind of scaling
transformation, like in (38), is almost always part of the mix.
When using perturbation methods our main task is to investigate the behaviour
of unknown functions f (), in the limit when approaches zero. This is what
we did in examples 1 to 3.
The way we approach this problem is to compare the unknown function
f () to one or several known functions when approaches zero. In example 1
and 2 we compared with the known functions {1, , 2 , ...} whereas in example
3 we used the functions {1 , 1, , ...}. In order to facilitate such comparisons
we introduce the large-o and little-o notation.
Definition 1 Let f () be a function of . Then
()
i) f () = O(g()) , 0 lim0 fg()
6= 0
()
ii) f () = o(g()) , 0
lim0 fg()
=0
Thus f () = O(g()) means that f () and g() are of roughly the same size
when approaches zero and f () = o(g()) means that g() is much larger than
f () when approaches zero.
We have for example that
1. sin() = O() ,
0,
2. sin(2 ) = o() ,
0, because
sin(2 )
2 cos(2 )
=0
lim
= lim
0
0
1
3. 1 cos() = o(),
because
sin()
= 1 6= 0
lim
0
0, because
1 cos()
sin()
=0
lim
lim
= 0
0
1
11
4. ln() = o(1 ),
0, because
1
ln()
= lim 2 = lim = 0
lim
0
0
0 1
When we apply perturbation methods we usually use a whole sequence of comparison functions. In examples 1 and 2 we used the sequence
{n () = n }
n=1
and in example 3 we used the sequence
{n () = n }
n=1
What is characteristic about these sequences is that
n+1 () = o(n ()),
(52)
for all n in the range defining the sequences. Sequences of functions that satisfy
the condition (52) are calles asymptotic sequencs.
Here are some asymptotic sequences
1. n () = sin()n
2. n () = ln()n
3. n () = ( )n
Using the notion of asymptotic sequence we can define asymptotic expansion
analogous to the way inifinite series are defined in elementary calculus
Definition 2 Let {n ()} be an asymptotic sequence. Then a formal series
an n ()
(53)
n=1
N
X
an n () = o(N ()),
n=1
12
(54)
Observe that
f () a1 1 () = o(1 ()),
f () a1 1 ()
=0
lim
0
1 ()
f ()
lim a1
=0
0
1 ()
a1 = lim
f ()
1 ()
(55)
(56)
This shows that for a fixed asymptotic sequence, the coefficients of the asymptotic expansion for a function f () are determined by taking limits. Observe
that the formula (56) does not require differentiability for f () at = 0. This
is very different from Taylor expansions which require that f () is infinitely
differentiable at = 0.
This is a hint that asymptotic expansions are much more general than the
usual convergent expansions, like for example power series, that we known from
elementary calculus. In fact asymptotic expansions may well diverge but this
does not make them less useful! The following example was first discussed by
Leonard Euler in 1754.
3.1
Eulers example
(57)
The integral defining f () converge very fast and because of this f () is a very
smooth function, in fact it is infinitely smooth and moreover analytic in the
complex plane where the negative real axis has been removed.
Using the properties of telescoping series we observe that for all m = 0
m
X
1
(t)m+1
=
(t)n +
1 + t n=0
1 + t
13
(58)
(59)
where
Sm () =
m
X
(1)n n!n
(60)
n=0
m+1
Rm () = ()
dt
0
tm+1 et
1 + t
(62)
f () =
(1)n n!n
(63)
n=0
It is on the other hand trivial to verify that the formal power series
(1)n n!n
(64)
n=0
14
4.1
(65)
y(1) = 1
where as usual is a small number. Since the differential equation is nonlinear
and nonseparable, this is a nontrivial problem. The unperturbed problem is
y 0 (x) + y(x) = x, 0 < x < 1
(66)
y(1) = 1
The unperturbed problem is easy to solve since the equation is a first order
linear equation. The general solution to the equation is
y(x) = x 1 + Aex
15
(67)
1 1 + Ae
=1
A=e
(68)
(69)
We now want to find an asymptotic expansion for the solution to the perturbed
problem (65) starting from the solution y0 (x). We thus postulate an expansion
y(x; ) = y0 (x) + y1 (x) + 2 y2 (x) + ...
(70)
(71)
y00
y10
+ 2 y20 + ... + y0 + y1
(y02 + 2y0 y1 + ..) =
+ 2 y2 + ...
x
y00
+ y0 +
(y10
+ y1 +
y02 )
(72)
(73)
(74)
y0 (1) = 1
(75)
y10 (x)
+ y1 (x) =
y02 (x)
y1 (1) = 0
(76)
y20 (x)
We observe that the perturbation hierarch has the special structure that we
have notes earlier. All equations in the hierarchy are determined by the linear
16
d
operator L = dx
+ 1. The first boundary value problem in the hierarchy has
already been solved. The second equation in the hierarchy is
(77)
x 0
(78)
inserting the formula for y0 (x) into (78), integrating and adding a general solution of the homogenous equation , we get the general solution to (77) in the
form
y1 (x) = x2 + 4x 5 + (A1 + 5)ex + (2x x2 )e1x e2x + e22x
(79)
Using the boundary condition for y1 (x) we find that A1 = e 5 and we can
conclude that the perturbation expansion to first order in is
y(x; ) = x 1 + e1x + x2 + 4x 5 + (2x x2 )e1x + e22x + ... (80)
The general solution to the third equation in the perturbation hierarchy is in a
similar way found to be
Z x
0
y2 (x) = A2 ex 2ex
dx0 ex y0 (x0 )y1 (x0 )
(81)
0
As you can see the algebraic complexity increase rapidly. The function y2 (x)
will have a large number of terms, even for this very simple example
Recall that we are only ensured that the correction y1 (t) is small with
respect to the unperturbed solution y0 (t) in the limit when approaches zero.
The perturbation method does not say anything about the accuracy for any
finite value of . The hope is of course that the perturbation expansion also
gives a good approximation for some range of > 0.
For this problem we do not have an exact solution that can be used to
investigate the accuracy of the perturbation expansion for finite values of .
This is the normal situation when we apply perturbation methods. The only
way to get at the accuracy of the perturbation expansion is to compare it to an
approximate solution found by some other, independent, approximation scheme.
Often this involve numerical methods, but it could also be another perturbation
method.
In figure (2) we compare our perturbation expansion to a high precision
numerical solution in the domain 0 < x < 1. We observe that even for as
large as 0.05 our perturbation expansion give a very accurate representation of
the solution over the whole domain.
However, as the next example show, things are not always this simple.
17
4.2
(82)
y(0) = 1
y 0 (0) = 0
This is our perturbed problem P(). The unperturbed problem, P(0), is
y 00 (t) + y(t) = 0
(83)
y(0) = 1
y 0 (0) = 0
The general solution to the unperturbed equation is evidently
y0 (t) = A0 eit + A0 eit
(84)
iA0
(85)
=0
which has the unique solution A0 = 21 . Thus the unique solution to the unperturbed problem is
1
y0 (t) = eit + ()
(86)
2
18
(87)
We now as usual insert (87) into the perturbed equation (82) and expand
(y0 + y1 + 2 y2 + ...)00
+(y0 + y1 + 2 y2 + ...)0 + y0 + y1 + 2 y2 + ... = 0
y000
+ y0 +
(y100
+ y1 +
y00 )
(88)
y10 (0)
2 y20 (t)
(89)
+ ... = 0
From equations (88) and (89) we get the following perturbation hierarchy
00
y0 + y0 = 0, t > 0
(90)
y0 (0) = 1
y00 (0) = 0
y100 + y1 = y00 , t > 0
y1 (0) = 0
y10 (0) = 0
y200 + y2 = y10 , t > 0
y2 (0) = 0
y20 (0) = 0
We note that the perturbation hierarchy has the special form discussed earlier.
d2
Here the linear operator determining the hierarchy is L = dt
2 + 1.
The first initial value problem in the hierarchy has already been solved. The
solution is (86). Inserting y0 (t) into the second equation in the hierarchy we get
i
y100 + y1 = eit + ()
2
Looking for particular solutions of the form
y1p (t) = Ceit + ()
19
(91)
will not work here because the right-hand side of (91) is a solution to the homogenous equation. In fact (91) is a harmonic oscillator driven on ressonance.
For such cases we must rather look for a special solution of the form
y1p (t) = Cteit + ()
(92)
1
1 it
e + (i t)eit + ()
2
4
(94)
Let yE (t) be a high precision numerical solution to the perturbed problem (82).
For = 0.01 we get for increasing time
t
yE
4 0.6444
40 0.5426
400 0.0722
y
0.6367
0.5372
0.5295
The solution starts out by being quite accurate, but as t increases the perturbation expansion eventually loose any relation to the exact solution. The
true extent of the disaster is seen in figure (3).
So what is going on, why is the perturbation expansion such a bad approximation in this example?
Observe that y1 contain a term that is proportional to t. Thus as t grows
the size of y1 also grows and when
t
(95)
the second term in the perturbation expansion become as large as the first term.
The ordering of the expansion breaks down and the first correction, y1 , is of
the same size as the solution to the unperturbed problem, y0 .
The reason why the growing term, y1 , is a problem here, but was not a
problem in the previous example is that here the domain for the independent
variable is unbounded.
Let us at this point introduce some standard terminology. The last two
examples involved perturbation expansions where the coefficients depended on
a parameter. In general such expansions takes the form
f (; x)
an (x)n (),
n=1
20
(96)
Figure 3: Comparing the direct perturbation expansion(red) and a high precision numerical solution(green)
21
The multiple scale method for weakly nonlinear ordinary differential equations.
In the previous section we saw that trying to represent the solution to the
problem
y 00 (t) + y 0 (t) + y(t) = 0, t > 0
(97)
y(0) = 1
y 0 (0) = 0
using a regular perturbation expansion
y(t) = y0 (t) + y1 (t) + 2 y2 (t)...
(98)
leads to a nonuniform expansion where ordering of the terms broke down for t
1
. In order to understand how to fix this, let us have a look at the exact solution
to (97). The exact solution can be found using characteristic polynomials. We
get
1 2
1
y(t) = Ce 2 t ei 1 4 t + ()
(99)
where
1
1
,
+
i
1 2
2
4
If we expand the square root in the exponent with respect to , we get
C=
y(t) Ce 2 t eit e 8 t + ()
(100)
(101)
(103)
We express this by saying that the function gn (t) vary on the time scale tn =
n t. If we now look at equation (101) we see that the approximate solution
(101) vary on three separate time scales t0 = 0 t, t1 = 1 t and t2 = 2 t. If
we include more terms in the Taylor expansion for the square root in (99) the
resulting solution will depend on even more time scales.
Inspired by this example we postulate the existence of a function
h = h(t0 , t1 , t2 , ...)
(104)
(105)
such that
is a solution to problem (97). Using the chain rule we evidently have
dy
(t) = (t0 + t1 + 2 t2 + ...)h |tj =j t
dt
which we formally write as
d
= t0 + t1 + 2 t2 + ...
dt
(106)
(107)
The multiple scale method now proceed by substituting (106) and (107) into
the differential equation
y 00 (t) + y 0 (t) + y(t) = 0
(108)
t0 t0 h0 + h0 + (t0 t0 h1 + h1 + t0 t1 h0 + t1 t0 h0 + t0 h0 )
+2 (t0 t0 h2 + h2 + t0 t1 h1 + t1 t0 h1 + t0 t2 h0 + t1 t1 h0
+t2 t0 h0 + t1 h0 + t0 h1 ) + ... = 0
23
(109)
(110)
t0 t0 h1 + h1 = t0 t1 h0 t1 t0 h0 t0 h0
t0 t0 h2 + h2 = t0 t1 h1 t1 t0 h1 t0 t2 h0
t1 t1 h0 t2 t0 h0 t1 h0 t0 h1
We observe in passing that the perturbation hierarchy has the special form
we have seen several times before. Here the common differential operator is
L = t0 t0 + 1.
At this point a remark is in order. It is fair to say that there is not a full
agreement among the practitioners of the method of multiple scales about how
to perform these calculations. The question really hinges on whether to take
the multiple variable function h(t0 , t1 , ..) seriously or not. If you do, you will
be lead to a certain way of doing these calculation. This is the point of view
used in most textbooks on this subject. We will not follow this path here. We
will not take h seriously as a multiple variable function and never forget that
what we actually want is not h but rather y, which is defined in terms of h
through equation (105). This point of view will lead us to do multiple scale
calculations in a different way from what you see in most textbooks. This way
is very efficient and will make it possible to go order 2 and beyond without
being overwhelmed by the amount of algebra that needs to be done. What I
mean when I say that we will not take h seriously as a multiple variable function
will become clear as we proceed.
One immediate consequence is already evident from the way I write the
perturbation hierarchy. Observe that I keep
ti tj hk , and tj ti hk
(111)
(112)
When we are applying multiple scales to ordinary differential equations we always use the general solution to the order 0 equation. For partial differential
24
equations this will not be so, as we will see later. The general solution to (112)
is evidently
h0 (t0 , t1 , ..) = A0 (t1 , t2 , ..)eit0 + ()
(113)
Observe that the equation only determines how h0 depends on the fastest time
scale t0 , the dependence on the other time scales t1 , t2 , .. is arbitrary at this
point and is reflected in the fact that integration constant A0 is actually a
function depending on t1 , t2 , ...
We have now solved the order 0 equation. Inserting the expression for h0
into the order equation, we get after some simple algebra
1
t0 t0 h1 + h1 = 2i(t1 A0 + A0 )eit0 + ()
2
(114)
We now need a particular solution to this equation. Observe that since A0 only
depends on the slow time scales t1 , t2 , .. equation (114) is in fact a harmonic
oscillator driven on ressonance. It is simple to verify that it has a particular
solution of the form
1
h1 (t0 , t1 , ..) = t0 (t1 A0 + A0 )eit0
2
(115)
But this term is growing and will lead to breakdown of ordering for the perturbation expansion (107) when t0 1 . This breakdown was exactly what we
tried to avoid using a multiple scales approach!
But everything is not lost, we now have freedom to remove the growing term
by postulating that
1
t1 A0 = A0
(116)
2
With this choise, the order equation simplifies into
t0 t0 h1 + h1 = 0
(117)
Terms in equations leading to linear growt like in (115) are traditionally called
secular terms. The name are derived from the Latin word soeculum that means
century and are used here because this kind of nonuniformity was first observed
on century time scales in planetary orbit calculations.
At this point we introduce the second rule for doing multiple scale calculations in the particular way that I advocate in these lecture notes. The rule is
to disregard the general solution of the homogenous equation for all equations
in the perturbation hierarchy except the first. We therefore choose h1 = 0 and
proceed to the order 2 equation using this choise. The equation for h2 then
simplifies into
i
i
t0 t0 h2 + h2 = 2i(t2 A0 t1 t1 A0 t1 A0 )eit0 + ()
2
2
(118)
We have a new secular term and in order to remove it we must postulate that
t2 A0 =
i
i
t t A0 + t1 A0
2 11
2
25
(119)
(120)
For this equation we use, according to the rules of the game, the special solution
h2 = 0.
What we have found so far is then
h(t0 , t1 , t2 , ..) = A0 (t1 , t2 , ..)eit0 + () + O(3 )
(121)
where
1
t1 A0 = A0
2
i
i
t2 A0 = t1 t1 A0 + t1 A0
2
2
(122)
(123)
At this point you might ask if we actually have done something useful. Instead of
one ODE we have ended up with two coupled partial differential equations, and
clearly if we want to go to higher order we will get even more partial differential
equations.
Observe that if we use (122) we can simplify equation (123) by removing the
derivatives on the right hand side. Doing this we get the system
1
t1 A0 = A0
2
i
t2 A0 = A0
8
(124)
(125)
The first thing that should come to mind when we see a system like (124) and
(125) is that the count is wrong. There is one unknown function, A0 , and two
equations. The system is overdetermined and will get more so if we extend our
calculations to higher order in . Under normal circumstances, overdetermined
systems of equations have no solutions, which for our setting means that under
normal circumstances the function h(t0 , t1 , t2 , ..) does not exist! This is what I
meant when I said that we will not take the functions h seriously as a multiple
variable function. For systems of first order partial differential equations like
(124), (125) there is a simple test we can use to decide if a solution actually exist.
This is the cross derivative test you know from elementary calculus. Taking t2
of equation (124) and t1 of equation (125) we get
1
i
t2 t1 A0 = t2 t1 A0 = t2 A0 =
A0
2
16
i
i
t1 t2 A0 = t1 t2 A0 = t1 A0 =
A0
8
16
(126)
According to the cross derivative test the overdetermined system (124), (125)
is solvable. Thus in this case the function h exists, at least as a two variable
function. To make sure that it exists as a function of three variables we must
26
derive and solve the perturbation hierarchy to order 3 , and then perform the
cross derivative test. For the current example we will never get into trouble,
the many variable function h will exist as a function of however many variables
we want. But I want you to reflect on how special this must be. We will at
order n have a system of n partial differential equations for only one unknown
function ! In general we will not be so lucky and this is the reason why we can
not take h seriously as a many variable function.
So should we be disturbed by the nonexistence of solutions in the general
case? Actually no, and the reason is that we do not care about h(t0 , t1 , ..).
What we care about is y(t).
Inspired by this let us define an amplitude, A(t), by
A(t) = A0 (t1 , t2 , ..)|tj =j t
(127)
Using this and equations (105), (121) our perturbation expansion for y(t) is
y(t) = A(t)eit + () + O(3 )
(128)
For the amplitude A(t) we have, using equations (106),(124),(125) and (127)
dA
(t) = {(t0 + t1 + 2 t2 + ...)A0 (t1 , t2 , ...)}|tj =j t
dt
1
dA
i
(t) = { A0 (t1 , t2 , ...) 2 A0 (t1 , t2 , ...)}|tj =j t
dt
2
8
dA
1
i
= A 2 A
dt
2
8
(129)
(130)
The constant C can be fitted to the initial conditions. What we get is equal to
the exact solution up to second order in as we see by comparing with (101).
Let us next apply the multiple scale method to some weakly nonlinear ordinary differential equations. For these cases no exact solution is known so the
multiple scale method will actually be useful!
27
5.1
Example 1
(131)
(133)
(134)
t0 t0 h0 + h0 + (t0 t0 h1 + h1 + t0 t1 h0 + t1 t0 h0 )
2
+ (t0 t0 h2 + h2 + t0 t1 h1 + t1 t0 h1 + t0 t2 h0 + t1 t1 h0
+t2 t0 h0 ) + ... = h30 + 32 h20 h1 + ...
(135)
(136)
t0 t0 h1 + h1 = h30 t0 t1 h0 t1 t0 h0
t0 t0 h2 + h2 = 3h20 h1 t0 t1 h1 t1 t0 h1 t0 t2 h0
t1 t1 h0 t2 t0 h0
28
(137)
Inserting this into the right hand side of the second equation in the hierarchy
and expanding we get
t0 t0 h1 + h1 = (3|A0 |2 A0 2it1 A0 )eit + A30 e3it + ()
In order to remove secular terms we must postulate that
3i
t1 A0 = |A0 |2 A0
2
This choise simplify the equation for h1 into
t0 t0 h1 + h1 = A30 e3it0 + ()
(138)
(139)
(140)
According to the rules of the game we now need a particular solution to this
equation. It is easy to verify that
1
(141)
h1 = A30 e3it0 + ()
8
is such a particular solution.
We now insert h0 and h1 into the right hand side of the third equation in
the perturbation hierarchy and find
3
t0 t0 h2 + h2 = ( |A0 |4 A0 2it2 A0 t1 t1 A0 )eit0 + () + N ST
(142)
8
where N ST is an acronym for nonsecular terms. Since we are not here planning to go beyond second order in , we will at this order only need the secular
terms and group the rest into N ST . In order to remove the secular terms we
must postulate that
i
3i
|A0 |4 A0 + t1 t1 A0
(143)
t2 A0 =
16
2
As before it make sense to simplify (143) using equation (139). This leads to
the following overdetermined system of equations for A0
3i
(144)
t1 A0 = |A0 |2 A0
2
15i
t2 A0 =
|A0 |4 A0
16
Let us check solvability of this system using the cross derivative test
3i
t2 t1 A0 = t2 (A20 A0 )
2
3i
=
2A0 t2 A0 A0 + A20 t2 A0
2
3i
15i
15i
2A0
|A0 |4 A0 A0 + A20
|A0 |4 A0
=
2
16
16
45
= |A0 |6 A0
32
29
15i
t1 A30 A2
0
16
15i
3
3A20 t1 A0 A2
=
0 + 2A0 A0 t1 A0
16
15i
3i
3i
2
2
2
3
2
=
3A0 |A0 | A0 A0 + 2A0 A0
|A0 | A0
16
2
2
45
= |A0 |6 A0
32
t1 t2 A0 =
(145)
(146)
(147)
(148)
The system is compatible and thus the function h0 exists as a function of two
variables. Of course, whether or not h0 exists is only of academic interest for us
since our only aim is to find the solution of the original equation y(t). Defining
an amplitude, A(t) by
A(t) = A0 (t1 , t2 , ...)|tj =j t
(149)
(150)
dA
3i
15i
(t) = { |A0 |2 A0 (t1 , t2 , ...) 2
|A0 |4 A0 (t1 , t2 , ...)}|tj =j t
dt
2
16
dA
3i
15i 4
= |A|2 A 2
|A| A
(151)
dt
2
16
Observe that this equation has a unique solution for a given set of initial conditions regardless of whether the overdetermined system (144) has a solution
or not. Thus doing the cross derivative test was only motivated by intellectual
curiosity, we did not have to do it.
In summary, (150) and (151), determines a perturbation expansion for y(t)
that is uniform for t . 2 .
At this point it is reasonable to ask in which sense we have made progress.
We started with one second order nonlinear ODE for a real function y(t) and
have ended up with one first order nonlinear ODE for a complex function A(t).
This question actually has two different answers.
The first one is that it is possible to get an analytical solution for (151)
whereas this is not possible for the original equation (131). This possibility
might however easily get lost as we proceed to higher order in , since this will
add more terms to the amplitude equation. But even if we can not solve the
30
amplitude equation exactly, it is a fact that amplitude equations with the same
mathematical structure will arise when we apply the multiple scale method to
many different equations. Thus any insight into an amplitude equation derived
by some mathematical analysis has relevance for many different situations. This
is clearly very useful.
There is however a second, more robust, answer to the question of whether
we have made progress or not. From a numerical point of view there is an
important difference between (131) and (151). If we solve (131) numerically the
time step is constrained by the oscillation period of the linearized system
d2 y
+y =0
dt2
(152)
(153)
1
(154)
2
For > 0 we solve the equation by Newton iteration starting with the solution
for = 0. This will give us the initial condition for the amplitude equation
correct to this order in .
In figure (4) we compare the multiple scale solution, keeping only the first
term in the amplitude equation, to a high precision numerical solution for = 0.1
for t 2 . We see that the perturbation solution is very accurate for this
range of t. In figure (5) we do the same comparison as in figure (4) but now
for t 3 . As expected the multiple scale solution and the numerical solution
starts to deviate for this range of t. In figure (6) we make the same comparison
as in figure (5), but now include both terms in the amplitude equation. We see
that high accuracy is restored for the multiple scale solution for t 3 .
A(0) =
31
Figure 4: Comparing the multiple scale solution, while keeping only the first
term in the amplitude equation(red), to a numerical solution(green) for t 2
Figure 5: Comparing the multiple scale solution, while keeping only the first
term in the amplitude equation(red), to a numerical solution(green) for t 3
32
Figure 6: Comparing the multiscale solution, while keeping both terms in the
amplitude equation(red), to a numerical solution(green)
5.2
Example 2
t>0
(155)
We want to apply the multiple scale method and introduce a function h(t0 , t1 , t2 , ..)
such that
y(t) = h(t0 , t1 , t2 , ..)|tj =j t
(156)
is a solution to equation (155). As usual we have the formal expansions
d
= t0 + t1 + 2 t2 + ...
dt
h = h0 + h1 + 2 h2 + ...
33
(157)
(158)
Inserting (156),(157) and (158) into equation (155) and expanding, we get
(t0 + t1 + 2 t2 + ...)(t0 + t1 + 2 t2 + ...)
(h0 + h1 + 2 h2 + ...) + (t0 + t1 + 2 t2 + ...)
(h0 + h1 + 2 h2 + ...)
= (h0 + h1 + 2 h2 + ...)2
t0 t0 h0 + t0 h0 + (t0 t0 h1 + t0 h1 + t0 t1 h0 + t1 t0 h0 + t1 h0 )
+2 (t0 t0 h2 + t0 h2 + t0 t1 h1 + t1 t0 h1 + t0 t2 h0 + t1 t1 h0
+t2 t0 h0 + t1 h1 + t2 h0 ) + ... = h20 2 2h0 h1 + ...
(159)
(160)
t0 t0 h1 + t0 h1 = h20 t0 t1 h0 t1 t0 h0 t1 h0
t0 t0 h2 + t0 h2 = 2h0 h1 t0 t1 h1 t1 t0 h1 t0 t2 h0
t1 t1 h0 t2 t0 h0 t1 h1 t2 h0
The general solution to the first equation in the perturbation hierarchy is
h0 (t0 , t1 , t2 , ...) = A0 (t1 , t2 , ..) + B0 (t1 , t2 , ...)et0
(161)
where A0 and B0 are real functions of their arguments. Inserting h0 into the
second equation in the hierarchy we get
t0 t0 h1 + t0 h1 = t1 A0 A20 + (t1 B0 2A0 B0 )et0 B02 e2t0
(162)
(163)
t1 B0 = 2A0 B0
equation (162) simplifies into
t0 t0 h1 + t0 h1 = B02 e2t0
(164)
(165)
Inserting (161) and (165) into the third equation in the perturbation hierarchy
we get
t0 t0 h2 + t0 h2 = t2 A0 t1 t1 A0 + (t2 B0 t1 t1 B0 )et0 + N ST
(166)
(167)
t2 B0 = t1 t1 B0
We can as usual use (163) to simplify (167). We are thus lead to the following
overdetermined system for A0 and B0 .
t1 A0 = A20
(168)
t1 B0 = 2A0 B0
t2 A0 = 2A30
t2 B0 = 2A20 B0
In order to satisfy our academic curiosity, let us do the cross derivative test for
solvability of (168).
t1 t2 A0 = 2t1 A30 = 6A20 t1 A0 = 6A40
t2 t1 A0 =
t2 A20
t1 t2 B0 =
2t1 (A20 B0 )
= 2A0 t2 A0 =
(169)
4A40
(170)
= 4A0 t1 A0 B0 +
2A20 t1 B0
=0
(171)
dA
(t) = {A2 (t1 , t2 , ...) 22 A(t1 , t2 , ...)3 }|tj =j t
dt
dA
= A2 22 A3
dt
35
(172)
(173)
and
dB
(t) = {(t0 + t1 + 2 t2 + ...)B0 (t1 , t2 , ...)}|tj =j t
dt
dB
(t) = {(2A0 (t1 , t2 , ...)B0 (t1 , t2 , ...) + 22 A20 (t1 , t2 , ...)B0 (t1 , t2 , ......)}|tj =j t
dt
dB
= 2AB + 22 A2 B
(174)
dt
Given the initial conditions for A and B, equations (173) and (174) clearly has a
unique solution and our multiple scale method will ensure that the perturbation
expansion (172) will stay uniform for t . 2 . As for the previous example the
initial conditions A(0) and B(0) are calculated from the initial conditions for
(155) by a Newton iteration. Thus we see again that the existence or not of
h(t0 , ..) is irrelevant for constructing a uniform perturbation expansion.
The system (173) and (174) can be solved analytically in terms of implicit
functions. However, as we have discussed before, analytical solvability is nice,
but not robust. If we take the expansion to order 3 , more terms are added to
the amplitude equations and the property of analytic solvability can easily be
lost. What is robust is that the presense of in the amplitude equations makes
(173) and (174) together with (172) into a fast numerical scheme for solving the
ordinary differential equation (155). This property does not go away if we take
the perturbation expansion to higher order in .
5.3
Exercises
For the following initial value problems, find asymptotic expansions that are
uniform for t << 2 . You thus need to take the expansions to second order in
. Compare your asymptotic solution to a high precision numerical solution of
the exact problem. Do the comparison for several values of and show that the
asymptotic expansion and the numerical solution of the exact problem deviates
when t 2 .
Problem 1:
d2 y
+ y = y 2
dt2
y(0) = 1
dy
(0) = 0
dt
36
Problem 2:
d2 y
dy
+ y = (1 y 2 )
2
dt
dt
y(0) = 1
dy
(0) = 0
dt
Problem 3:
dy
d2 y
+ y = (y 3 2 )
2
dt
dt
y(0) = 1
dy
(0) = 0
dt
Problem 4: Let the initial value problem
d2 y dy
+
+ y 2 = 0,
dt2
dt
y(0) = 1
t>0
(175)
y 0 (0) = 1
be given. Design a numerical solution to this problem based on the amplitude equations (173),(174) and (172). Compare this numerical solution
to a high precision numerical solution of (175) for t 2 . Use several
different values of and show that the multiple scale solution and the high
precision solution deviate when t 2 .
The multiple scale method for weakly nonlinear dispersive wave equations.
It is now finally time to start applying the multiple scale method to partial
differential equations. The partial differential equations that are of interest in
optics are almost always hyperbolic, weakly dispersive and nonlinear. We will
therefore focus all our attention on such equations. The simplest such equation
is the following
6.1
Example 1
(176)
(177)
(178)
x = x0 + x1 + x2 + ...
and for h we use the expansion
h = h0 + h1 + 2 h2 + ...
(179)
t0 t0 h0 + (t0 t0 h1 + t0 t1 h0 + t1 t0 h0 )+
(180)
(t0 t0 h2 + t0 t1 h1 + t1 t0 h1 + t0 t2 h0 + t1 t1 h0 + t2 t0 h0 ) ...
x0 x0 h0 (x0 x0 h1 + x0 x1 h0 + x1 x0 h0 )
2
(x0 x0 h2 + x0 x1 h1 + x1 x0 h1 + x0 x2 h0 + x1 x1 h0 + x2 x0 h0 )
+h0 + h1 + 2 h2 + ...
= h20 + 22 h0 h1 + ...
which gives us the perturbation hierarchy
t0 t0 h0 x0 x0 h0 + h0 = 0
(181)
t0 t0 h1 x0 x0 h1 + h1 = h20 t0 t1 h0 t1 t0 h0
(182)
+x0 x1 h0 + x1 x0 h0
t0 t0 h2 x0 x0 h2 + h2 = 2h0 h1 t0 t1 h1 t1 t0 h1
t0 t2 h0 t1 t1 h0 t2 t0 h0 + x0 x1 h1 + x1 x0 h1
+x0 x2 h0 + x1 x1 h0 + x2 x0 h0
38
(183)
For ordinary differential equations we used the general solution to the order 0
equation. For partial differential equations we can not do this. We will rather
use a finite sum of linear modes. The simplest possibility is a single linear mode
which we use here
h0 (t0 , x0 , t1 , x1 , ...) = A0 (t1 , x1 , ...)ei(kx0 t0 ) + ()
(184)
Since we are not using the general solution we will in not be able to satisfy
arbitrary initial conditions. However, in optics this is perfectly alright since
most of the time the relevant initial conditions are in fact finite sums of wave
packets or even a single wave packet. Such initial conditions can be included in
the multiple scale approach that we discuss in this section. For (184) to actually
be a solution to (181) we must have
p
(185)
= (k) = 1 + k 2
which we of course recognize as the dispersion relation for the linearized version
of (176). With the choise of signs used here (184) will represent a right-moving
disturbance.
Inserting (184) into (182) we get
t0 t0 h1 x0 x0 h1 + h1 = 2|A0 |2
+A20 e2i(kx0 t0 )
+(2it1 A0 +
2i(kx0 t0 )
+ A2
0 e
2ikx1 A0 )ei(kx0 t0 )
(186)
(187)
(188)
(189)
m
k
t1 A0 = x1 A0
(190)
and this is in fact true for all k. This is however not generally true for dispersive
wave equations. Whether it is true or not will depend on the exact form of the
39
(191)
According to the rules of the game we need a special solution to this equation.
It is easy to verify that
1
1
h1 = 2|A0 |2 A20 e2i(kx0 t0 ) A2
e2i(kx0 t0 )
3
3 0
(192)
is such a special solution. Inserting (184) and (192) into (183), we get
t0 t0 h2 x0 x0 h2 + h2 = (2it2 A0 + 2ikx2 A0 t1 t1 A0
(193)
10
+ x1 x1 A0 + |A0 |2 A0 )ei(kx0 t0 ) + N ST + ()
3
In order to remove secular terms we must postulate that
2it2 A0 + 2ikx2 A0 t1 t1 A0 + x1 x1 A0 +
10
|A0 |2 A0 = 0
3
(194)
(195)
(196)
(197)
where A(x, t) satisfy a certain amplitude equation that we will now derive.
Multiplying equation (189) by , equation (194) by 2 and adding the two
40
expressions, we get
(2it1 A0 + 2ikx1 A0 )
+2 (2it2 A0 + 2ikx2 A0 t1 t1 A0 + x1 x1 A0 +
10
|A0 |2 A0 )
3
=0
10 2
|A| A = 0
3
m
k
i
i
5i
t A = x A
tt A +
xx A + 2 |A|2 A
2
2
3
(199)
(200)
This equation appears to have a problem since it contains a second derivative with respect to time. The initial conditions for (176) is only sufficient to
determine A(x, 0). However, in order to be consistent with the multiple scale
procedure leading up to (200) we can only consider solutions such that
k
t A x A
2
k
xx A 2
tt A
(201)
2
3
(202)
This is now first order in time and has a unique solution for a given initial
condition A(x, 0).
The multiple scale procedure demands that the amplitude A(x, t) vary slowly
2
on scales L = 2
k , T = . This means that (196) and (202) can be thought of
as a fast numerical scheme for wavepackets solutions to (176). If these are the
kind of solutions that we are interested in, and in optics this is often the case,
it is much more efficient to use (196) and (202) rather than having to resolve
the scales L and T by integrating the original equation (176).
41
6.2
Example 2
(203)
Introducing the usual tools for the multiple scale method, we have
u(x, t) = h(x0 , t0 , x1 , t1 , ...)|tj =j t,xj =j x
t = t0 + t1 + ...
x = x0 + x1 + ...
h = h0 + h1 + ...
Inserting these expressions into (203) and expanding we get
(t0 + t1 + ...)(t0 + t1 + ...)(h0 + h1 + ...)+
(x0 + x1 + ...)(x0 + x1 + ...)(h0 + h1 + ...)+
(x0 + x1 + ...)(x0 + x1 + ...)
(x0 + x1 + ...)(x0 + x1 + ...)(h0 + h1 + ...)
= (h0 + ...)3
t0 t0 h0 + (t0 t0 h1 + t0 t1 h0 + t1 t0 h0 )+
x0 x0 h0 + (x0 x0 h1 + x0 x1 h0 + x1 x0 h0 )+
x0 x0 x0 x0 h0 + (x0 x0 x0 x0 h1 + x0 x0 x0 x1 h0 + x0 x0 x1 x0 h0
+x0 x1 x0 x0 h0 + x1 x0 x0 x0 h0 ) + ...
= h30 + ...
42
(204)
(205)
t0 t0 h1 + x0 x0 h1 + x0 x0 x0 x0 h1 = h30
(206)
t0 t1 h0 t1 t0 h0 x0 x1 h0 x1 x0 h0
x0 x0 x0 x1 h0 x0 x0 x1 x0 h0 x0 x1 x0 x0 h0 + x1 x0 x0 x0 h0
For the order 0 equation we choose a wave packet solution
h0 (x0 , t0 , x1 , t1 , ...) = A0 (x1 , t1 , ...)ei(kx0 t0 ) + ()
(207)
k4 k2 + 1
(208)
(209)
i(kx0 t0 )
+ ()
(210)
(211)
(212)
(213)
Introducing an amplitude
A(x, t) = A0 (x1 , t1 , ...)|xj =ej x,tj =j t
we get, following the approach from the previous example, the amplitude equation
2i(t A + 0 x A0 ) = 3|A|2 A
(214)
This equation together with the expansion
u(x, t) = A(t)ei(kxt) + () + O()
(215)
constitute a fast numerical scheme for wave packet solutions to (203) for t . 1 .
Of course this particular amplitude equation can be solved analytically, but as
43
stressed earlier, this property is not robust and can easily be lost if we take the
expansion to higher order in .
There is however one point in our derivation that we need to look more
closely into. We assumed that the term
A30 e3i(kx0 t0 )
(216)
(217)
p
81k 4 9k 2 + 1 = 3 k 4 k 2 + 1
m
4
81k 9k + 1 = 9k 4 9k 2 + 9
m
1
k =
3
(218)
Thus the term (216) can be secular if the wave number of the wave packet is
given by (218). This is another example of the fenomenon that we in optics
call phase matching. As long as we stay away from the two particular values
of the wave numbers given in (218), our expansion (214) and (215) is uniform
for t . 1 . However if the wave number takes on one of the two values in
(218), nonuniformities will make the ordering of the expansion break down for
t 1 . However this does not mean that the multiple scale method breaks
down. We only need to include a second amplitude at order 0 that we can use
to remove the additional secular terms at order 1 . We thus, instead of (207),
use the solution
h0 (x0 , t0 , x1 , t1 , ...) = A0 (x1 , t1 , ...)ei(kx0 t0 )
+ B0 (x1 , t1 , ...)e
3i(kx0 t0 )
(219)
+ ()
(220)
where k now is given by (218). Inserting this expression for h0 into the order
equation (206) we get after a fair amount of algebra, the equation
t0 t0 h1 + x0 x0 h1 + x0 x0 x0 x0 h1 =
3
(221)
+3|A0 | A0 + 6|B0 | A0 +
3A2
0 B0
(222)
=0
(223)
(224)
3i(kxt)
+ () + O()
6.3
Exercises
In the following problems use the methods from this section to find asymptotic
expansions that are uniform for t . 2 . Thus all expansions must be taken to
second order in .
Problem 1:
utt uxx + u = 2 u3
Problem 2:
utt uxx + u = (u2 + u2x )
45
Problem 3:
utt uxx + u = (uuxx u2 )
Problem 4:
ut + uxxx = u2 ux
Problem 5:
utt uxx + u = (u2x uuxx )
(225)
t D H = 0
D=0
B=0
The material of interest is almost always nonmagnetic so that
1
B
D = 0 E + P
H=
(226)
The polarization is in general a sum of a terms that is linear in E and one that
is nonlinear in E. We have
P = PL + PN L
(227)
where the term linear in E has the general form
Z t
PL (x, t) = 0
dt0 (t t0 )E(x, t0 )
(228)
Thus the polarization at a time t depends on the electric field at all times previous to t. This memory effect is what we in optics call temporal dispersion. The
presence of dispersion in Maxwell equations spells trouble for the intergration of
the equations in time, we can not solve them as a standard initial value problem.
This is of course well known in optics and various more or less ingenious methods
has been designed for getting around this problem. In optical pulse propagation one gets around the problem by solving Maxwells equations approximately
as a boundary value problem rather than as an initial value problem. A very
general version of this approach is the well known UPPE[8],[4],[7] propagation
scheme. In these lecture notes we will, using the multiple scale method, derive
46
(229)
Z
b
= 0
d
b()E(x,
)eit
!
Z
b(n) (0) n b
E(x, )eit
= 0
d
n!
n=0
Z
(n)
X
b (0)
b
= 0
d n E(x,
)eit
n!
n=0
Z
(n)
X
b (0)
nb
it
= 0
d(it ) E(x, )e
n!
n=0
Z
b(n) (0)
n
it
b
(it )
= 0
d E(x, )e
n!
n=0
= 0
b(it )E(x, t)
where
b() is the fourier transform of (t). These manipulations are of course
purely formal, in order to make them into honest mathematics we must dive
into the theory of pseudo differential operators. In these lecture notes we will
not do this as our focus is on mathematical methods rather than mathematical
theory.
Inserting (226),(227),(228) and (230) into (225) we get Maxwells equations
in the form
t B + E = 0
2
(231)
2
t E c B + t
b(it )E = c 0 t PN L
1
(E +
b(it )E) = PN L
0
B=0
47
7.1
Let us first simplify the problem by only considering solutions of the form
E(x, y, z, t) = E(x, z, t)ey
(232)
(233)
t B2 + x E = 0
t E c2 (z B1 x B2 ) + t
b(it )E = t PN L
x B1 + z B2 = 0
where
PN L = E 3
(234)
It is well known that this vector system is fully equivalent to the following scalar
equation
tt E c2 2 E + tt
b(it )E = tt PN L
(235)
where we have introduced the operator
2 = xx + zz
(236)
Equation (235) will be the staring point for our multiple scale approach, but
before that I will introduce the notion of a formal perturbation parameter. For
some particular application of equation (235) we will usually start by making the
equation dimension-less by picking some scales for space, time, and E relevant
for our particular application. Here we dont want to tie our calculations to
some particular choise of scales and introduce therefore a formal perturbation
parameter in the equation multiplying the nonlinear polarization term. Thus
we have
tt E c2 2 E + tt
b(it )E = 2 tt E 3
(237)
Starting with this equation we will proceed with our perturbation calculations
assuming that << 1 and in the end we will remove by setting it equal to 1.
What is going on here is that is a place holder for the actual small parameter
that will appear in front of the nonlinear term in the equation when we make a
particular choise of scales. Using such formal perturbation parameters is very
common.
You might ask why I use 2 instead of as formal perturbation parameter? I
will not answer this question here but will say something about it at the very end
of the lecture notes. We proceed with the multiple scale method by introducing
the expansions
t = t0 + t1 + 2 t2 + ...
2
= 0 + 1 + 2 + ...
e = e0 + e1 + 2 e2 + ...
E(x, t) = e(x0 , t0 , x1 , t1 , ...)|tj =j t,xj =j x
48
(238)
where
j = (xj , zj )
(239)
is the gradient with respect to xj = (xj , zj ). We now insert (238) into (237)
and expand everything in sight
(t0 + t1 + 2 t2 + ...)(t0 + t1 + 2 t2 + ...)
(e0 + e1 + 2 e2 + ...)
c2 (0 + 1 + 2 2 + ...) (0 + 1 + 2 2 + ...)
(e0 + e1 + 2 e2 + ...)+
(t0 + t1 + 2 t2 + ...)(t0 + t1 + 2 t2 + ...)
49
t0 t0 e0 + (t0 t0 e1 + t0 t1 e0 + t1 t0 e0 )
+ 2 (t0 t0 e2 + t0 t1 e1 + t1 t0 e1 + t0 t2 e0 + t1 t1 e0 + t2 t0 e0 ) + ...
c2 20 e0 c2 (20 e1 + 1 0 e0 + 0 1 e0 )
2 c2 (20 e2 + 1 0 e1 + 0 1 e1
+ 2 0 e0 + 1 1 e0 + 0 2 e0 ) + ...
+ t0 t0
b(it0 )e0 + (t0 t0
b(it0 )e1 + t0 t0
b0 (it0 )it1 e0
+ t0 t1
b(it0 )e0 + t1 t0
b(it0 )e0 ) + 2 (t0 t0
b(it0 )e2
+ t0 t0
b0 (it0 )it1 e1 + t0 t1
b(it0 )e1 + t1 t0
b(it0 )e1
1
b00 (it0 )t1 t1 e0 + t1 t0
b0 (it0 )it1 e0
+ t0 t0
b0 (it0 )it2 e0 t0 t0
2
+ t0 t1
b0 (it0 )it1 e0 + t2 t0
b(it0 )e0 + t1 t1
b(it0 )e0
+ t0 t2
b(it0 )e0 ) + ...
= 2 t0 t0 e30 + ...
which gives us the perturbation hierarchy
t0 t0 e0 c2 20 e0 + t0 t0
b(it0 )e0 = 0
(240)
t0 t0 e1 c2 20 e1 + t0 t0
b(it0 )e1 =
(241)
t0 t1 e0 t1 t0 e0 c 1 0 e0 c 0 1 e0
t0 t0
b0 (it0 )it1 e0 t0 t1
b(it0 )e0 t1 t0
b(it0 )e0
t0 t0 e2 c2 20 e2 + t0 t0
b(it0 )e2 =
(242)
t0 t1 e1 t1 t0 e1 t0 t2 e0 t1 t1 e0 t2 t0 e0
2
c 1 0 e1 c2 0 1 e1 c2 2 0 e0 c2 1 1 e0
c2 0 2 e0 t0 t0
b0 (it0 )it1 e1 t0 t1
b(it0 )e1
1
b00 (it0 )t1 t1 e0
t1 t0
b(it0 )e1 t0 t0
b0 (it0 )it2 e0 + t0 t0
2
t1 t0
b0 (it0 )it1 e0 t0 t1
b0 (it0 )it1 e0 t2 t0
b(it0 )e0
t1 t1
b(it0 )e0 t0 t2
b(it0 )e0 t0 t0 e30
For the order 0 equation we choose the wave packet solution
e0 (x0 , t0 , x1 , t1 , ..) = A0 (x1 , t1 , ...)ei0 + ()
(243)
where
xj = (xj , zj )
0 = k x0 t0
50
(244)
(245)
(246)
We now must now calculate the right-hand side of the order equation.
Observe that
t1 t0 e0 = it1 A0 ei0 + ()
t0 t1 e0 = it1 A0 e
i0
+ ()
1 0 e0 = ik1 A0 ue
i0
+ ()
0 1 e0 = ik1 A0 ue
i0
+ ()
(247)
t0 t0
b (it0 )it1 e0 = i
b ()t1 A0 ei0 + ()
t0 t1
b(it0 )e0 = i
b()t1 A0 ei0 + ()
t1 t0
b(it0 )e0 = i
b()t1 A0 ei0 + ()
where u is a unit vector in the direction of k. Inserting (247) into (241) we get
t0 t0 e1 c2 20 e1 + t0 t0
b(it0 )e1 =
(248)
{2it1 A0 2ic ku 1 A0
2 0
i
b ()t1 A0 2i
b()t1 A0 }ei0 + ()
In order to remove secular terms we must postulate that
2it1 A0 2ic2 ku 1 A0 i 2
b0 ()t1 A0 2i
b()t1 A0 = 0
m
2
(2n +
b ())t1 A0 2ic2 ku 1 A0 = 0
(249)
(1 +
b()) = c2 k 2
0 2
2 0
2 n () +
b () = 2c2 k
(2n +
b ()) = 2c2 k
Thus (249) can be written in the form
t1 A0 + vg 1 A0 = 0
51
(250)
(251)
(252)
(253)
for (252). We now must compute the right-hand side of the order 2 equation.
Observe that
t2 t0 e0 = it2 A0 ei0 + ()
t1 t1 e0 = t1 t1 A0 e
i0
t0 t1 e0 = it2 A0 e
+ ()
i0
2 0 e0 = iku 2 A0 e
1 1 e 0 =
21 A0 ei0
(254)
+ ()
i0
+ ()
+ ()
0 2 e0 = iku 2 A0 ei0 + ()
t0 t0
b0 (it0 )it2 e0 = i 2
b0 ()t2 A0 ei0 + ()
1
1
t t
b00 ()t1 t1 A0 ei0 + ()
b00 (it0 )t1 t1 e0 = 2
2 00
2
t1 t0
b0 (it0 )it1 e0 =
b0 ()t1 t1 A0 ei0 + ()
t0 t1
b0 (it0 )it1 e0 =
b0 ()t1 t1 A0 ei0 + ()
t2 t0
b(it0 )e0 = i
b()t2 A0 ei0 + ()
t1 t1
b(it0 )e0 =
b()t1 t1 A0 ei0 + ()
t0 t2
b(it0 )e0 = i
b()t2 A0 ei0 + ()
t0 t0 e30 = 3 2 |A0 |2 A0 ei + N ST + ()
Inserting (253) and (254) into the right-hand side of the order 2 equation we
get
t0 t0 e2 c2 20 e2 + t0 t0
b(it0 )e2 =
2
(255)
2
21 A0
{2it2 A0 + t1 t1 A0 2ic ku 2 A0 c
1
b00 ()t1 t1 A0 + 2
b0 ()t1 t1 A0
i 2
b0 ()t2 A0 + 2
2
2i
b()t2 A0 +
b()t1 t1 A0 3 2 |A0 |2 }ei0 + N ST + ()
52
(258)
where
= 0
b00 ()
n2 + 2
b0 () + 12 2
2c2 k
0
2k
3 2 0
=
2c2 k
(259)
(260)
(261)
and proceeding in the usual way, using (250) and (258), we get the following
amplitude equation
t A + vg A i2 A + itt A i|A|2 A = 0
(262)
where we have put the formal perturbation parameter equal to 1. From what
we have done it is evident that for
E(x, t) = A(x, t)ei(kxt) + ()
(263)
(264)
t A vg A O()
where is a number much smaller than 1. Under these circumstances (262),(263)
is the key elements in a fast numerical scheme for wave packet solutions to (237).
Because of the presence of the second derivative with respect to time, equation
(262) can not be solved as a standard initial value problem. However, because
of (264) we can remove the second derivative term by iteration
t A = vg A O()
tt A = (vg )2 A O(2 )
53
(265)
(266)
which can be solved as a standard initial value problem. In deriving this equation we asssumed that the terms proportional to
e3i(kxt)
where nonsecular. For this to be true we must have
(3k) 6= 3(k)
(267)
N
X
Aj (x1 , t1 , ...)eij + ()
(268)
j=0
Calculations analogous to the ones leading up to equation (262) will now give a
separate equation of the type (262) for each wave packet, unless we have phase
matching. These phase matching conditions appears from the nonlinear term
in the order 2 equation and takes the familiar form
kj = s1 kj1 + s2 kj2 + s3 kj3
(269)
(270)
54
7.2
Up til now all applications of the multiple scale method has involved scalar
equations. The multiple scale method is not limited to scalar equations and is
equally applicable to vector equations. However, for vector equations we need
to be more careful than for the scalar case when it comes to elliminating secular
terms. We will here use Maxwells equations (231) to illustrate how the method
is applied to vector equations in general. Assuming, as usual, a polarization
response induced by the Kerr effect, our basic equations are
t B + E = 0
2
(271)
2
t E c B + t
b(it )E = t (E E)
B=0
E+
b(it ) E = 2 (E2 E)
where we have introduced a formal perturbation parameter in front of the nonlinear terms. We now introduce the usual machinery of the multiple scale method.
Let e(x0 , t0 , x1 , t1 , ...) and b(x0 , t0 , x1 , t1 , ...) be functions such that
E(x, t) = e(x0 , t0 , x1 , t1 , ...)|xj =j x,tj =j t
(272)
(273)
= 0 +1 + 2 +...
= 0 +1 +2 2 +...
e = e0 + e1 + 2 e2 + ...
b = b0 + b1 + 2 b2 + ...
We now insert (273) into (271) and expand everything in sight to second order in
. Putting each order of to zero separately gives us the perturbation hierarchy.
At this point you should be able to do this on your own so I will just write down
the elements of the perturbation hierarchy when they are needed.
The order 0 equations, which is the first element of the perturbation hierarchy, is of course
t0 b0 +0 e0 = 0
(274)
t0 e0 c 0 b0 + t0
b(it0 )e0 = 0
0 b0 = 0
0 e0 +
b(it0 )0 e0 =
0
For the order equations we chose a linearly polarized wave packet solution.
It must be of the form
e0 (x0 , t0 , x1 , t1 , ...) = A0 (x1 , t1 , ...)qei0 + ()
b0 (x0 , t0 , x1 , t1 , ...) = kA0 (x1 , t1 , ...)te
55
i0
+ ()
(275)
where
0 = k x0 t0
(276)
and where
= (k)
is a solution to the dispersion relation
2 n2 () = c2 k 2
(277)
The orthogonal unit vector q and t span the space transverse to k = ku, and
the unit vectors {q, t, u} define a postively oriented fram for R3 .
The order equations are
t0 b1 +0 e1 = t1 b0 1 e0
(278)
t0 e1 c 0 b1 + t0
b(it0 )e1 =
t1 e0 + c2 1 b0 t1
b(it0 )e0 it0
b0 (it0 )t1 e0
0 b1 = 1 b0
0 e1 +
b(it0 )0 e1 =
1 e0
b(it0 )1 e0 ib
0 (it0 )t1 0 e0
Inserting (275) into (278) we get
t0 b0 +0 e0 = {kt1 A0 t + 1 A0 q}ei0 + ()
2
(279)
2 0
t0 e0 c 0 b0 + t0
b(it0 )e0 = {(n () +
b ())t1 A0 q
c2 k1 A0 t}ei0 + ()
0 b0 = {k1 A0 t}ei0 + ()
0 e0 +
b(it0 )0 e0 = {n2 ()1 A0 q}ei0 + ()
If we can find a special solution to this system that is bounded we will get a
perturbation expansion that is uniform for t . 1 . We will look for solutions
of the form
e1 = aei0 + ()
b1 = be
i0
(280)
+ ()
where a and b are constant vectors. Inserting (280) into (279), we get the
following linear algebraic system of equations for the unknown vectors a and b
ib + iku a = {kt1 A0 t + 1 A0 q}
2
2 0
in ()a ic ku b = {(n () +
b ())t1 A0 q
(281)
(282)
c k1 A0 t}
iku b = k1 A0 t
2
(283)
(284)
a = a ak
bk = (u b)u,
b = b bk
(285)
ak = (i 1 A0 q)u
k
bk = (i1 A0 t)u
(286)
(287)
However the longitudinal part of (281) and (282) will also determine ak and bk .
These values must be the same as the ones just found in (286),(287). These are
solvability conditions. Taking the longitudinal part of (281) we get
iu b = u (1 A0 q)
(288)
m
u b = i1 A0 t
which is consistent with (287). Thus this solvability condition is automatically
satisfied. Taking the longitudinal part of (282) we get
in2 ()u a=c2 ku (1 A0 t)
(289)
u a = i 1 A0 q
k
which is consistent with (286). Thus this solvability condition is also automatically satisfied. The transversal part of (281) and (282) are
ib + iku a = {kt1 A0 + 1 A0 u}t
2
(290)
2
in ()a ic ku b = {(n () +
b ())t1 A0 + c k1 A0 u}q
This linear system is singular, the determinant is zero because of the dispersion
relation (277). It can therefore only be solved if the right-hand side satisfy a
certain solvability condition. The most effective way to find this condition is to
use the Fredholm Alternative. It say that a linear system
Ax = c
has a solution if and only if
f c=0
for all vectors f , such that
A f = 0
where A is the adjoint of A.
57
i
ic2 ku
in2
ic2 ku
(291)
(292)
i + ic2 ku = 0
A convenient basis for the null space is
2
c2 kq
c kt
,
t
q
(293)
The first basis vector gives a trivial solvability condition whereas the second one
gives a nontrivial condition which is
c2 k{kt1 A0 + 1 A0 u} + {(n2 () +
b0 ())t1 A0 + c2 k1 A0 u} = 0
m
2
(2n +
b ())t1 A0 + 2c2 ku 1 A0 = 0
(294)
(295)
2 0 n2 + 2
b0 () 0 = 2c2 k
(2n +
b ()) 0 = 2c2 k
2
(296)
a = aq
b = 0
58
(298)
a = i t1 A0 + u 1 A0 q
k
(299)
From (286),(287),(298) and (299) we get the following bounded special solution
to the order equations
i
u 1 A0 )q + i( q 1 A0 )u}e 0 + ()
k
k
i
b1 = {i(t 1 A0 )u}e 0 + ()
e1 = {i(t1 A0 +
(300)
t2
b(it0 )e0 + it1
b0 (it0 )t1 e0 + it0
b0 (it0 )t2 e0
1
b00 (it0 )t1 t1 e0 + t0 e20 e0 }
t0
2
0 b2 = {1 b1 + 2 b0 }
0 e 2 +
b(it0 )0 e2 = {1 e1 + 2 e0 +
b(it0 )1 e1
+ib
0 (it0 )t1 0 e1 +
b(it0 )2 e0 + ib
0 (it0 )t2 0 e0
1 00
+ib
0 (it0 )t1 1 e0
b (it0 )t1 t1 0 e0 + 0 (e20 e0 )}
2
59
(301)
+i 1 (1 A0 u) q + i 1 (1 A0 q) u + kt2 A0 t
k
k
i
+2 A0 q}e 0 + ()
(302)
t0 e2 c2 0 b2 + t0
b(it0 )e2 = {iF ()t1 t1 A0 q
+iG()( t1 1 A0 u)q + iG()(t1 1 A0 q)u ic2 1 (1 A0 t) u
c2 k2 A0 t + H() t2 A0 q 3i 4 |A0 |2 A0 }ei0 + ()
0 b2 = {i1 (1 A0 t) u + k2 A0 t}e
i0
+ ()
0 e 2 +
b(it0 )0 e2 = {in2 1 t1 A0 q + in2 1 (1 A0 u) q
k
2
2
i0
+in 1 (1 q) u + n 2 A0 q}e + ()
k
where we have defined
1
b00 ()
F () = n2 + 2
b0 () + 2
2
b0 ())
G() = (n2 +
k
H() = (n2 +
b0 ())
(303)
Like for the order equations, we will look for bounded solutions of the form
e2 = aei0 + ()
b2 = be
i0
(304)
+ ()
Inserting (304) into (302) we get the following linear system of equations for the
60
+i 1 (1 A0 u) q + i 1 (1 A0 q) u + kt2 A0 t
k
k
+2 A0 q}
(305)
(306)
(307)
(308)
ak = ( 1 t1 A0 q 2 1 (1 A0 u) q
k
k
2 1 (1 A0 q) u + i 2 A0 q)u
k
k
1
bk = (i2 A0 t 1 (1 t) u)u
k
(309)
(310)
(311)
{t 1 A0 t 1 t1 A0 t 1 (1 A0 u) t + i2 A0 t} (312)
1
k
in order for (312) to be consistent with (311) we find that the following solvability
condition must hold
t1 1 A0 t = 1 t1 A0 t
(313)
The longitudinal part of (306) is
ua=
1
{G()t1 1 A0 q + ic2 k2 A0 q}
n2
(314)
In order for (314) to be consistent with (309) we find, after a little algebra, that
the solvability condition
2
n ()1 t1 A0 q + G()t1 1 A0 q =
(315)
k
c2 1 (1 A0 q) u c2 1 (1 A0 u) q
61
ib + iku a = {i1 t1 A0 u + i 1 (1 A0 u) u
(316)
k
(317)
c2 k({i1 t1 A0 u + i 1 (1 A0 u) u
k
(318)
i(1 (1 A0 q) q + 1 (1 A0 t) t 1 (1 A0 u) u)
+it1 t1 A0 i|A0 |2 A0 = 0
where we have defined
=
=
=
1 =
2 =
0 F ()
2c2 k
0
2k
3 0 4
2c2 k
0
2
0 G()
2c2 k
62
(319)
(320)
We have now found all solvability conditions. These are (313),(315),(317) and
(318). We define an amplitude A(x, t) by
A(x, t) = A0 (x1 , t1 , ...)|xj = j x, tj = j t
and derive the amplitude equations from the solvability conditions in the usual
way. This gives us the following system
t A t = t A t
2
n ()t A q + G() t A q =
k
c2 (A q) u c2 (A u) q
(321)
(322)
(A0 q) t = (A t) q
(323)
t A + vg A + i1 t A u + i2 t A u
(324)
i((A q) q + (A t) t (A u) u)
+itt A i|A|2 A = 0
where we as usual have set the formal perturbation parameter equal to 1. Equations (321) and (323) are automatically satisfied since A(x, t) is a smooth function of space and time. We know that only amplitudes such that
t A vg A = 0 A u
(325)
(326)
(327)
E(x, t) {(A + i( 0 )u A)q
k
+ i( q A)u}ei(kxt) + ()
k
B(x, t) {kAt + i(t A)u}ei(kxt) + ()
63
The equations (326) and (327) are the key elements in a fast numerical scheme
for linearly polarized wave packet solutions to Maxwells equations. Wave packets of circular polarization or arbitrary polarization can be treated in an entirely
similar manner as can sums of different polarized wave packets.
We of course recognize the amplitude equation (326) as the 3D nonlinear
Schrodinger equation including group velocity dispersion. As we have seen before, an equation like this can be solved as an ordinary initial value problem if
we first use (325) to make the term containing a second derivative with respect
to time into one containing only space derivatives.
The derivation of the nonlinear Scrodinger equation for linearly polarized
wave packets I have given in this section is certainly not the fastest and simplest
way this can be done. The main aim in this section was to illustrate how to apply
the multiple scale method for vector PDEs in general, not to do it in the most
effective way possible for the particular case of linearly polarized electromagnetic
wave packets.
All the essential elements we need in order to apply the method of multiple
scales to problems in optics and laser physics, and other areas of science too,
are at this point known. There are no new tricks to learn. Using the approach
described in these lecture notes, amplitude equations can be derived for most
situations of interest. Applying the method is mechanical, but for realistic systems there can easily be a large amount of algebra involved. This is unavoidable,
solving nonlinear partial differential equations, even approximately, is hard.
In these lecture notes we have focused on applications of the multiple scale
method for time-propagation problems. The method was originally developed
for these kind of problems and the mechanics of the method is most transparent for such problems. However the method is by no means limited to time
propagation problems.
Many pulse propagation schemes are most naturally formulated as a boundary value problem where the propagation variable is a space variable. A very
general scheme of this type is the well known UPPE[8] propagation scheme.
More details on how the multiple scale method is applied for these kind of
schemes can be found in [4], [7] and [1].
References
[1] K. Glasner, M. Kolesik, J. V. Moloney, and A. C. Newell. Canonical and
singular propagation of ultrashort pulses in a nonlinear medium. International journal of optics, 2012.
[2] E. J. Hinch. Perturbation Methods. Cambridge University Press, 1991.
[3] M. H. Holmes. Introduction to Perturbation Methods. Springer, 1995.
[4] P. Jakobsen and J. V. Moloney. The effect of longitudinal electric field
components on the propagation of intense ultrashort optical pulses. Physica
D, 241:16031616, 2012.
64
65