1

Scaling, Mathematical
Modelling, &
Integrable Systems
Initial Profile t=.2 t=.4

0.15 0.15 0.15

0.1 0.1 0.1

0.05 0.05 0.05

0 0 0

−0.05 −0.05 −0.05

0 2 4 6 0 2 4 6 0 2 4 6

t=.5 t=.6 t=1

0.15 0.15 0.15

0.1 0.1 0.1

0.05 0.05 0.05

0 0 0

−0.05 −0.05 −0.05

0 2 4 6 0 2 4 6 0 2 4 6

D.H. Sattinger
University of Minnesota
2
Preface
This project grew out of a conversation with Willi Jager in Heidelberg on a
visit in January, 1996. I discussed with Willi my investigations with Mariana
Haragus (now Haragus-Courcelle) at Stuttgart Universitat on the validity of
the Kortweg-deVries approximation to the Euler equations of gravity waves.
Many of the completely integrable systems of partial dierential equa-
tions arise as formal, singular asymptotic approximations, in some manner
or other, to more complicated nonlinear dispersive wave equations. It is a
curious fact that the methods of asmyptotic expansion and choice of scaling
should lead so often to completely integrable models. This artifact has never
been properly explained.
These derivations are generally formal in nature, and, with a few ex-
ceptions, have never been properly justied with mathematical rigor. Willi
proposed that I give a series of Deutsche Mathematische Vereinigung lectures
on the subject of scaling and mathematical modelling.
My hope here is to present an introduction to dispersive waves and the
derivation of some of the associated completely integrable systems, raise a
number of issues that have not been fully dealt with, introduce some method-
ologies (including the use of numerical animation of weakly nonlinear disper-
sive wave models using Matlab), and provide a comprehensive bibliography.
I would like to thank my colleagues and students for many helpful com-
ments and suggestions in the preparation of these notes, especially Sam Al-
bert, Jerry Bona, Mariana Haragus-Courcelle, Yi Li, Peter Olver, and Jacek
Szmigielski.

D.H. Sattinger
University of Minnesota
December, 1997

3
4
Contents
1 Dispersion 9
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2 Group and Phase Velocities . . . . . . . . . . . . . . . . . . . 10
2 Nonlinear Schrodinger Equation 21
2.1 Multiple Scales Expansion . . . . . . . . . . . . . . . . . . . . 21
2.2 Pulse solutions . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3 Korteweg de-Vries 29
3.1 Background and History . . . . . . . . . . . . . . . . . . . . . 29
3.2 Plasmas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.3 Water Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.4 The Solitary wave of the KdV equation . . . . . . . . . . . . . 46
4 Isospectral Deformations 49
4.1 The KdV hierarchy . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2 The AKNS hierarchy . . . . . . . . . . . . . . . . . . . . . . . 54
5 Inverse Scattering Theory 59
5.1 The Schrodinger equation . . . . . . . . . . . . . . . . . . . . 59
5.2 First Order Systems . . . . . . . . . . . . . . . . . . . . . . . 73
5.3 Decay of the scattering data . . . . . . . . . . . . . . . . . . . 82
6 Variational Methods 85
6.1 Water Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.2 Method of Averaging . . . . . . . . . . . . . . . . . . . . . . . 88
5
6 CONTENTS
7 Weak and Strong Nonlinearities 95
7.1 Breaking and Peaking . . . . . . . . . . . . . . . . . . . . . . 95
7.2 Strongly nonlinear models . . . . . . . . . . . . . . . . . . . . 99
7.3 The extended AKNS hierarchy . . . . . . . . . . . . . . . . . . 106
8 Numerical Methods 111
8.1 The nite Fourier transform . . . . . . . . . . . . . . . . . . . 111
8.2 Pseudospectral Codes . . . . . . . . . . . . . . . . . . . . . . . 119
List of Figures
1.1 Decay of a centered Gaussian. . . . . . . . . . . . . . . . . . . 15
1.2 A traveling oscillatory pulse. . . . . . . . . . . . . . . . . . . . 16
1.3 The Airy equation . . . . . . . . . . . . . . . . . . . . . . . . 18
2.1 Pulses of the NLS equation . . . . . . . . . . . . . . . . . . . . 26
2.2 Dark Soliton . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.1 Homoclinic orbit . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2 Two soliton interaction . . . . . . . . . . . . . . . . . . . . . . 48
6.1 A modulated pulse. . . . . . . . . . . . . . . . . . . . . . . . . 92
7.1 The extreme wave as conjectured by Stokes. . . . . . . . . . . 96
7.2 Breaking Wave . . . . . . . . . . . . . . . . . . . . . . . . . . 97
7.3 Peakons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
8.1 Dispersive dierence scheme . . . . . . . . . . . . . . . . . . . 115
8.2 Implicit spectral method . . . . . . . . . . . . . . . . . . . . . 116
8.3 Dispersive Cut-O . . . . . . . . . . . . . . . . . . . . . . . . 123

7
8 LIST OF FIGURES
Chapter 1
Dispersion
1.1 Introduction
In these lectures I shall describe the modelling of the competing eects of
dispersion and nonlinearity in physical systems. In particular, I shall discuss
in detail the derivation of the Korteweg-deVries and nonlinear Schrodinger
equations. The interplay of dispersion and nonlinearity is at the heart of the
phenomena which these equations purport to model.
The nonlinear equation
ut + uux = 0 (1.1.1)
develops shock discontinuities, even if the initial data is innitely dieren-
tiable. But when dissipation is added,
ut + uux = "uxx;
the solutions are globally smooth, even for arbitrarily small positive ".
The process of dispersion, on the other hand, is a more subtle phe-
nomenon. For example, the Korteweg-deVries equation
ut + uux = "uxxx
possesses globally smooth solutions, given smooth initial data, but the third
order derivative uxxx is neutral from the energy point of view. It neither
creates nor destroys energy. Instead, it acts to disperse energy at rates de-
pendent on the wave number, and the formation of shocks by the nonlinear
9
10 CHAPTER 1. DISPERSION
term uux is counterbalanced by the dispersion due to the term "uxxx, no
matter what value of " we take, positive or negative. The behavior of the
solutions as " ! 0 is called the zero dispersion limit, and has been studied
extensively [47]. This is of course, a rough statement of a general mathe-
matical principle that dispersion and nonlinearity often act to balance each
other.
In these lectures we shall focus on dispersive equations and their mod-
elling by asymptotic and scaling arguments. The arguments currently found
in the literature are, for the most part, purely formal, and lack mathemat-
ical justication; yet have considerable physical plausibility. They are thus
interesting grounds for mathematical investigation.
We begin by explaining the notions of group velocity and dispersion re-
lation.

1.2 Group and Phase Velocities

The general solution of the wave equation
utt uxx = 0
is
u(x; t) = f (x t) + g(x + t):
This solution is a superposition of waves traveling to the left and right with-
out change of shape. This behavior is singular; obviously the wave equation
is the only equation which supports this kind of evolution. In general, one
must deal with the phenomenon of dispersion, in which wave trains disperse
at dierent rates according to their frequency [20].
If we look for elementary solutions of the wave equation of the form
ei(kx !t) , we nd the relation
! 2 = k2 :
This is called the dispersion relation for the wave equation. On the other
hand, the dispersion relation of the linear Klein Gordon equation
utt uxx + m2 u = 0
is
!2 = k2 + m2 ;
1.2. GROUP AND PHASE VELOCITIES 11
or p
! =  k 2 + m2 :
The Klein Gordon equation arises in the propagation of sound in pipes
[20], [62]. This problem is the forerunner of the modern notion of waveguides
and propagation of pulses in optical bers. If the ow is assumed to be
irrotational and incompressible it is given by a velocity potential,
v = r'; 'tt c2
' = 0:
The propagation of sound in a pipe lying along the x axis with a xed cross-
section
lying in the y z plane is given by
 @ 2 @2 + @2 :
2
'tt c
2 + @x2 ' = 0; where
2 = @y 2 @z 2
The normal velocity vanishes on the boundary of the pipe, hence
@'  = 0:
@ @

We look for a solution of the wave equation of the form

X
1
'= uj (x; t) j (y; z):
j =0
By separation of variables, u and jmust satisfy the equations
@ 
j
@ @
= 0:
2
utt uxx + j uj = 0;
2 j + j j = 0;2

thus 2j is an eigenvalue of
2 on the cross-section

with Neumann boundary
conditions.
The lowest eigenvalue is 0 = 0, with 0 = const: This corresponds to
the propagation mode in which there is no transverse velocity in the pipe,
since the velocity components 'y and 'z are zero. This mode propagates
without dispersion, whereas the higher propagation modes, in which 2j > 0,
propagate with dispersion.
The dispersion relation for the Schrodinger equation
ut = iuxx (1.2.2)
12 CHAPTER 1. DISPERSION
is
! = k2 :
The exponential function
ei(kx !t)
represents a plane wave moving with speed
dx = ! :
dt k
The number k is called the wave number. It has physical dimensions of
inverse length, and the plane wave train represented by ei(kx !t) is spatially
periodic with period 2=k. The quantity !=k is the phase velocity, since
along the lines
dx = ! ;
dt k
the phase of the waveform is constant.
Dispersion is a non-local phenomenon; that is, it is an artifact of solutions
which are superpositions of waves with varying wave numbers. Consider,
therefore, the behavior of a superposition of plane waves, namely the Fourier
integral 1 Z
u(x; t) = ei(kx !t) f (k) dk; ! = ! (k ):
1
The function u dened by the Fourier integral above satises a certain
linear evolution equation. Note that
Z1 Z1
ut = ei(kx !(k)t) i!(k) f (k) dk; ux = ei(kx !t) ik f (k) dk;
1 1
Letting
d
D = 1i dx
we see that the action of D on u is given by multiplication of f (k) by k.
Using this, we dene the operator !(D) by
Z1
! (D )u = ei(kx !t) !(k)f (k) dk:
1
1.2. GROUP AND PHASE VELOCITIES 13
With this denition, u satises the linear evolution equation
ut + i!(D)u = 0:
For example, if !(k) = k + k3 then

!(D)u = Du + D3u = 1i (ux uxxx);

but !(k) need not be a polynomial in k, in which case !(D) is a non-local
operator.
The asymptotic behavior of the solutions of the Schrodinger equation,
with a Gaussian pulse as initial data can be done exactly, using the fact that
the class of Gaussian pulses is invariant under Fourier transform:
Z1 r
eikxe k2 dk = e
x2 =4 :
1

We take the initial data u0(x) = f (x) and its Fourier transform to be
r
f (x) =  eik0 xe x2 =4 ; fb(k) = e (k k0 )2 :

The Fourier transform is thus a Gaussian which peaks at k = k0. The

oscillatory factor eik0 x modulating the Gaussian pulse comes from centering
the Fourier transform at k = k0. Recall that translation of a function by h
results in the multiplication of its Fourier transform by the factor eikh, and
vice versa.
The solution of the Schrodinger equation is given by the Fourier integral
Z1
u(x; t) = ei(kx k2 t) e (k k0 )2 dk:
1
Writing
x = v0t + ; v0 = !0(k0) = 2k0; (1.2.3)
14 CHAPTER 1. DISPERSION
we get
Z1
u(x; t) = eik eit(kv0 k2) e (k k0 )2 dk
1
Z1
=ei(k0 +k02t) ei(k k0)e (+it)(k k0 )2 dk
1
Z1
=ei(k0 +k02t) eiy e (+it)y2 dy
1
r 
=eik0 (+k0t) 2 =4(+it) :
 + it e
Now r  p  i t

 + it = (2 + t2 )1=4 exp arctan ;
2 
and  
 2 = exp (it ) 2 :
 
exp
4( + it) 4(2 + t2 )
Using  = x v0t we can write the expression for u(x; t) as
p  (x v t)2 
eik0 (x k0t) exp 0 ei (x;t)
(2 + t2)1=4 4(2 + t2 )
where
(; t) = t 2 1 arctan t
2 2
4( + t ) 2 
For xed  ,
lim = i ;
t!1 4
hence the solution of the Schrodinger equation consists of a Gaussian which:
 is centered at x = v0t, hence moves with the group velocity x_ = v0 ;
 decays like t 1=2 as t ! 1;
1.2. GROUP AND PHASE VELOCITIES 15
 has variance  2 
4  + t ;
and so disperses as t ! 1;
 has asymptotic phase x k0t =4; hence asymptotic phase velocity
x_ = k0 = !(k0)=k0:
In the case of the simple linear Schrodinger equation we are able to carry
out the analysis exactly; for more complicated equations, for example nonlin-
ear equations, this is no longer possible. In any event, a few simple computer
experiments give a much better feeling as to what is happening. The cal-
culations were made on Matlab using a spectral method (the nite Fourier
transform). These numerical methods will be discussed in Chapter 8.
We compare the solution of the linear Schrodinger equation for two initial
data, one of which is a simple Gaussian pulse and the second of which is a
modulated pulse. In the rst case, we take as initial data (on the transform
side) a Gaussian pulse centered at k = 0, and we nd that the pulse does
not move, but only decays, as in Figure 1.1.
Initial Profile T=.15
1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0

−0.2 −0.2

0 2 4 6 0 2 4 6

Figure 1.1: Decay of a centered Gaussian.

In this computation we started with initial data
f = e 10 (x )2 :
16 CHAPTER 1. DISPERSION
The unmodulated Gaussian pulse disperses rapidly, but it does not move.
On the other hand, if we start with an oscillating pulse, such as
f = e 10 (x )2 e20ix
we nd that the pulse holds together, at least roughly, and moves to the
right, as in Figure 1.2. The oscillations in the pulse essentially drive it; if we
were to change the oscillatory factor to e 20ix , the pulse would move to the
left. The reason is that multiplication of a function by an exponential has the
eect of translating its Fourier transform. So multiplication of a Gaussian
by an oscillating factor e20ix translates the Fourier transform to a Gaussian
centered at k0 = 20; and then the group velocity is 2k0 = 40.
T=0 T=6
2.5 2.5

2 2

1.5 1.5

1 1

0.5 0.5

0 0

−0.5 −0.5

−1 −1

−1.5 −1.5

−2 −2

−2.5 −2.5
0 2 4 6 0 2 4 6

Figure 1.2: A traveling oscillatory pulse.

The initial data, on the transform side, is a Gaussian centered at k0 = 30:
fb(k) = 100e :1(k 30)2+80i k)
The complex factor e80i k is chosen so that the pulse lies in the interior of the
interval and does not meet the edges. The oscillation in the pulse is caused
by centering the Fourier transform at k = 30; and this determines the group
velocity of the pulse. I leave this as a computer experiment for you to try
yourself.
1.2. GROUP AND PHASE VELOCITIES 17
In quantum mechanics, the square of the amplitude ju(x; t)j2 is inter-
preted as the probability distribution of a particle. The location of the par-
ticle is thus determined only probabilistically, and the group velocity of the
wave is interpreted as the dynamic velocity of the particle.
The general case is considerably more complicated. Again suppose the
initial distribution is a Gaussian centered at k0 . Given a dispersion relation
!(k) we need to determine the asymptotic behavior of the integral
Z1
u(x; t) = ei(kx !t) e (k k0 )2 dk:
1
Again writing x = v0 t +  , we obtain
Z1
u(x; t) = eik eit(v0 ;k)e (k k0 )2 dk;
1
where
(v; k) = kv !(k):
We may expand
00
(v0 ; k) = (v0 ; k0) ! (2k0) (k k0)2 + (k k0)3 R0(k k0):
Then the integral can be written
u(x; t) = ei(k0 x !(k0 )t) I (; t)
where
Z1 00
I (; t) = exp[i(y t ! (2k0) y2 + ty3R0(y)) y2] dy
1

 = x v0 t; y = k k0:
The following heuristic argument, due to Lord Kelvin, is routinely given in
the literature [74], p. 372. Since the Gaussian exp( y2) is rapidly decaying,
only the values of the integrand near y = 0 make a contribution, and the
term
ty3R0(y)
18 CHAPTER 1. DISPERSION
can be neglected. This is the method of stationary phase, rst proposed by
Lord Kelvin. The statement is roughly correct, but it is hard to quantify
since the method of stationary phase does not give the spatial dependence
of the integral
For example, consider the case of the simple equation
ut = uxxx
with dispersion relation ! = k3 . The exact solution is
u(x; t) = ei(k0 x k03t) A(; t);  = x !0(k0)t;
Z1
A(; t) = exp[i(k 3k0tk2 tk3) k2 ] dk:
1

Initial Profile Schrodinger Airy

2.5 2.5 2.5

2 2 2

1.5 1.5 1.5

1 1 1

0.5 0.5 0.5

0 0 0

−0.5 −0.5 −0.5

−1 −1 −1

−1.5 −1.5 −1.5

−2 −2 −2

−2.5 −2.5 −2.5

0 2 4 6 0 2 4 6 0 2 4 6

Figure 1.3: The evolution of a Gaussian pulse under the Schrodinger equation
vs. the Airy equation.
The function A(; t)) is a modied Airy integral. If the term eitk3 is
ignored, we obtain the Gaussian integral
Z1
G(; t) = eik e (3ik0 t+)k2 dk:
1
1.2. GROUP AND PHASE VELOCITIES 19
The integral A can be evaluated approximately by the method of steepest
descent [28]. The functions G and A satisfy the following linear partial
dierential equations:
Gt = 3ik0G ; At = 3ik0A + A :
The real parts of the solutions of the two equations with a modulated
Gaussian pulse as initial data are displayed in Figure 1.3.
20 CHAPTER 1. DISPERSION
Chapter 2
The Nonlinear Schrodinger
Equation
The nonlinear Schrodinger equation arises in a multitude of contexts, but its
best known application is to the self-focussing of one-dimensional electromag-
netic waves, for example optical pulses along an optical ber, in nonlinear
media [76]. There are a number of formal derivations of this equation, one
of which is based on multiple scales expansions and is described below.
The nonlinear Schrodinger equation is the principle model equation in
nonlinear optics, that is, in the modelling of propagation of electromagnetic
pulses in nonlinear optical bers, which act as wave guides. G. B. Whitham
[74], p. 549, has derived the nonlinear Schrodinger equation by a method
of averaging the variational equations. This procedure is closely related to
the method of multiple scales; but it has several conceptual advantages.
Whitham's method will be discussed in x5.3.
The nonlinear Schrodinger equation was shown to be integrable by the
method of inverse scattering by V.E. Zakharov and A.B. Shabat [76]. We
will discuss the inverse scattering theory associated with this model in x4.4.

2.1 Multiple Scales Expansion

We give here a formal derivation of the Nonlinear Schrodinger equation as
an approximation to the nonlinear Klein Gordon equation
utt uxx + m2u u3 = 0:
21
22 
CHAPTER 2. NONLINEAR SCHRODINGER EQUATION
This equation is a simple model for the eects of dispersion and nonlinearity.
We will discuss the choice of sign of  below, as it has an eect on the
existence theory.
There are many derivations of the nonlinear Schrodinger equation in the
literature starting from some nonlinear wave equation [40], [52]. We present
here the ubiquitous derivation [52], [26]. We introduce a sequence of scaled
variables
xj = "j x; tj = "j t;
and consider them as independent variables. The variables xj ; tj ; j  1 are
called the slow space and time variables . We then look for a solution of the
form
X
1
u(x; t; ") = Uj "j
j =1
where Uj = Uj (x0 ; x1; : : : ; t0 ; t1; : : : ). Note that with this scaling the solution
is small in amplitude; this means that the nonlinear term is order "3 and
would be neglible in this limit. This seems to me to be contradictory; if one
is truly trying to see the eects of the nonlinear term, one needs to look at
large amplitude solutions.
In terms of the sequence of scaled variables, we have formally, by the
chain rule,
@ =X 1
"j @ ; @ =
X1
"j @ :
@x j=0 @xj @t j=0 @tj
hence
@2u = X
1
j +k @
2 @2u = X
1 2
j +k @ :
@x2 j;k=0 @xj @xk ;
" "
@t2 j;k=0 @tj @tk
Substituting these expressions into the equation and collecting terms, we
have
L0 U1 =0
 @2 @2 U ;

L0 U2 =2 @x0 @x1 @t0 @t1 1 (2.1.1)
2.1. MULTIPLE SCALES EXPANSION 23
 @2 @2 U

L0 U3 = @x0 @x1 @t0 @t1 2
 @2 @2 U
  @2 
@ 2 U +  (U )3: (2.1.2)
+2 @x0 @x2 @t0 @t2 1 @t21 @x21 1 1

where  @2 
@ 2 + m2 :
L0 = @t2 @x 20
0
We look for a solution of the rst equation in the form
U1 = <A(x1 ; x2; t1; t2 )ei ;  = kx !t; !2 = k2 + m2:
This term is to be interpreted as a carrier wave (ei ) with a modulation by a
complex envelope A(x1 ; x2 ; t1; t2 ) in the slow variables. Then U1 is a solution
of the rst equation, and we shall derive a nonlinear equation, in fact the
nonlinear Schrodinger equation, for the complex envelope A.
When we substitute U1 into the equation at second order (2.1.1) we get
 @A @A 
L0 U2 = < 2i k @x + ! @t ei
1 1
The presence of the term ei on the right creates secular terms in the solution,
since it is in the null space of the operator L0. Thus, one solution of
 @2 @2 
2 i
@t20 @x20 + m u = e
is
ei ;
2im2
which grows linearly in both x and t.
In order to eliminate such secular terms we require that
@A + ! @A = 0
k @x
1 @t1
This equation implies that A is a function of the variable 1 = !x1 kt1 ,
that is
A(x1 ; x2 ; : : : ; t1; t2; : : : ) = A(!x1 kt1 ; x2 ; t2; : : : ) = A(1; x2 ; t2; : : : ):
24 
CHAPTER 2. NONLINEAR SCHRODINGER EQUATION
A is a waveform traveling to the right with velocity k=! in the slow variables
x1 ; t1. Note that k=! is the group velocity of the wave:
@! = k :
@k !
Thus, in a frame of reference moving with the group velocity, the secular
terms vanish.
The equation for U2 is then simply
 @2 
@ 2 + m2 U = 0;
@t20 @x20 2

and we may take

U2 = <A2ei
for some complex envelope A2 = A2(!x1 kt1; x2 ; t2; : : : ). In the equation
at third order (2.1.2) the term involving U2 on the right side then drops out
by the same argument as before, for:
 @2 
@ 2 A ( ; x ; t ; : : : )ei
@x0 @x1 @t0 @t1 2 1 2 2

=ik @A
@x
2 + i! @A2
@t
1 1

=ik! @A
@
2 i!k @A 2
@ = 0:
To avoid secular terms in the equation for U3 we again require the coe-
cients of ei on the right side of the equation to vanish. A calculation shows
the coecient of ei on the right side to be
 @A @A  
1 @2

@ 2 A + 3 jAj2A:
i k @x + ! @t 2 @t21 @x21 8
2 2
Since !2 k2 = m2,
 @2 
@ 2 A (  ;  ;  ) = 1 ( ! 2 k 2 ) A = 1 m2 A ;
@t21 @x21 1 2 2 2 1 1
2 1 1
2.2. PULSE SOLUTIONS 25
hence the necessary and sucient condition that the secular terms vanish is
that A satisfy the nonlinear partial dierential equation
 @A @A  1
i k @x + ! @t + 2 m2 A1 1 + 38 jAj2A = 0:
2 2
Introducing the coordinates
2 2
2 = (!x2 kt2 ) 2m!k ; 2 = (!x2 + kt2 ) 2m!k ;
we obtain, after some simplications, the equation
iA2 + 21 A1 1 + 2 jAj2A = 0;  = 43m2 ; (2.1.3)
as the equation to remove secular terms. This is the nonlinear Schrodinger
equation.
The case  > 0 (hence  > 0) is called the self-focussing case. This is
the case where there are solitons. Note that the scales are mixed, the \time
variable" 2 being on the second order scale, while the \space variable" 1
is on the rst order scale. If the wave envelope A were a function only of
1; 2, then the pulse would travel with the group velocity, without change
of shape. It is a function of 2 however, so it is modulated on a scale of order
"2 as it travels.
This formal argument extends in a more or less straightforward fashion
to a variety of nonlinear wave equations. This is done in many places in the
literature. A rigorous justication of such expansions in a variety of cases,
including nonlinear wave equations, has been given by Kirrmann, Schneider,
and Mielke [41].

2.2 Pulse solutions

We construct here two special solutions of the nonlinear Schrodinger equation
(2.1.3). The equation is invariant under the gauge transformation A 7! A
where  is any complex number of absolute value 1. Let us look for solutions
of the form
A(1 ; 2) = ei(a1 +b2 ) (1; 2 );
where a and b are real constants.
26 
CHAPTER 2. NONLINEAR SCHRODINGER EQUATION
1 1

0.8 b=30
0.8 b=20

0.6

0.6
0.4

0.4
0.2

0
0.2

−0.2
0

−0.4

−0.2
−0.6

−0.4 −0.8
−5 0 5 −5 0 5

Figure 2.1: Pulses of the nonlinear Schrodinger equation for the parameter
values b = 20; 30.

It is straightforward to see that (2.1.3) equation reduces to

2i ( 2 + a 1 ) + 1 1 +  j j2 + (2b a2) = 0:
If we choose = u(1 a2 ), where u is real, the equation simplies to a
second order real nonlinear equation for u:
u00 + u3 + u = 0;  = (2b a2 ): (2.2.4)
The solutions of this equation are elliptic functions. The potential is
 u4 +  u2:
4 2
p
There are equilibria at u = 0;  2=.
The pulse, or soliton, solutions, that is, the solutions which are not pe-
riodic on the line, are obtained as homoclinic or heteroclinic orbits of this
nonlinear equation. In order to get such pulses, we must have  < 0:
For  > 0, the self-focussing case, we must therefore choose b so that
 < 0. The potential then p has an unstable equilibrium at the origin and
stable equilibria at u =  2=. The stable equilibria are centers, around
which are families of closed orbits, and these two families of closed orbits are
2.2. PULSE SOLUTIONS 27
bounded by homoclinic orbits based at the origin. The homoclinic orbits
are the solitons for the nonlinear Schrodinger equation. The corresponding
three parameter family of solutions of (2.1.3) is ( is a phase shift, due to
the translational invariance of the equation)
r  p 
 
 2   e i (a 1 + b 2 ) sech jj(1 a2 ) ;  < 0;  > 0:
There is a wide range of pulses, depending on the choice of the parameters.
Let us ignore the phase and the amplitude (2j=j)1=2 and freeze the 'time'
2 . Then up to phase, the modulated pulse of (2.1.3) is
p
eia1 sech( a2 2b1 )
In Figure 2.1 we see the graphs of the pulse for two dierent parameter
values, a = 8 and b = 20; 30.
1

0.8

0.6

0.4

0.2

−0.2

−0.4

−0.6

−0.8

−1
−10 −8 −6 −4 −2 0 2 4 6 8 10

Figure 2.2: Prole of the real part of a dark soliton for the parameters
a = 8; b = 32:09, = 2 = 0.
In the de-focussing case,  <p0;  > 0, the origin is a center for a family
of closed periodic orbits, and  = are now unstable equilibria joined
by two heteroclinic orbits. These are the so-called dark solitons, shown in
Figure (2.2). There are two such heteroclinic orbits to (2.2.4) in this case,
r  1=2
u(x) = ! tanh 2 (x ); ! = 
28 
CHAPTER 2. NONLINEAR SCHRODINGER EQUATION
so the solutions to (2.1.3) are
r  
! ei(a1 +b2 ) tanh 2 (1 a2 )
Chapter 3
The Korteweg deVries
Equation
3.1 Background and History
The Korteweg-deVries equation,
ut + uux + uxxx = 0 (3.1.1)
is the simplest equation that includes both the eects of nonlinearity and
dispersion. The equation appears in various forms in the literature, some-
times with a factor of 6 or -6 in front of the nonlinear term. We shall use
the present normalization, since that is the unique normalization for which
the Korteweg-deVries equation is Galilean invariant. The rst order equation
(1.1.1) is known to develop shock discontinuities in nite time, even with C 1
initial data; but when the term uxxx is added, the solutions remain smooth
for all time. Even more, the equation has been found to possess remarkable
mathematical structure. It is an innite dimensional Hamiltonian system
that, in a precise sense, is completely integrable. Moreover, the equation
models some of the phenomena found in more complicated systems of non-
linear partial dierential equations, such as the Euler equations for waves on
the surface of an inviscid, irrotational uid.
In this chapter we discuss the derivation of the Korteweg-deVries equation
as a model for long waves in dispersive media. It also arises as an approximate
model in numerous other elds, including magnetohydrodynamics waves, ion-
acoustic plasma waves, and anharmonic lattice vibrations. Martin Kruskal,
29
30 CHAPTER 3. KORTEWEG DE-VRIES
who played a major role in the discovery of the fundamental mathematical
properties of the equation, states [42]
The essential signicance of the KdV equation in the present context is
that it is the unique leading order reduced equation approximation.
An excellent reference for the theory of water waves prior to 1957 is J.J.
Stoker's treatise [70]. An account of the history of the rigorous mathematical
theory of the solitary and periodic waves is given beginning on p. 342. The
variational principle described in Chapter 6, however, was only discovered in
1967, and so no account of it is to be found in Stoker's book, After the com-
plete integrability of the Korteweg-deVries equation was discovered in 1967,
the theory of the KdV equation and its relationship to the Euler equations
as an approximate model obtained from perturbation theory became of more
interest; and Whitham's book [74] treats the KdV approximation in some
detail.
In addition, a history of the Korteweg-deVries equation and its role as a
model equation for water waves can be found in the articles by Bona [15],
and Newell [52], and so I will only review the highlights which are relevant
to our discussion.
The original observation of the solitary wave in water was reported by
Russell [65]1
\I believe I shall best introduce this phenomenon by describing the cir-
cumstances of my own rst acquaintance with it. I was observing the motion
of a boat which was rapidly drawn along a narrow channel by a pair of horses,
when the boat suddenly stopped { not so the mass of water in the channel
which it had put in motion; it accumlated round the prow of the vessel in
a state of violent agitation, then suddenly leaving it behind, rolled forward
with great velocity, assuming the form of a large solitary elevation, a rounded,
smooth and well-dened head of water, which continued its course along the
channel apparently without change of form or diminution of speed. I followed
it on horseback, and overtook it still rolling on a a rate of some eight or nine
miles an hour, preserving its original gure some thirty feet long and a foot
to a foot and a half in height. Its height gradually diminished, and after a
chase of one or two miles I lost it in the windings of the channel. Such, in
This passage, as well as discussion of other topics related to Russell's observations,
1
can be found in the monograph by Alan Newell.
3.1. BACKGROUND AND HISTORY 31
the month of August 1834, was my rst chance interview with that singular
and beautiful phenomenon which I have called the Wave of Translation, a
name which it now very generally bears."
Russell's subsequent experiments stimulated great interest in the subject
of water waves, and his ndings were immediately taken up by Airy [2] and
Stokes [71]. Stokes took special note of Russell's assertion that \the velocity
of propagation of a series of oscillatory waves does not depend on the height of
the waves." Stokes computed the Fourier series of the formal approximations
to second order in the case of nite depth, and to third order in the case of
innite depth.
\I have proceeded to a third approximation in the particular case in which
the depth of the uid is very great . . . This term gives an increase in the
velocity of propagation depending on the square of the ratio of the height of
the waves to their length."
Among the many conclusions of his investigations, Stokes concluded, er-
roneously, as it turns out, \that it is only an indenite series of waves which
possesses the property of being propagated with a uniform velocity."
From our present-day vantage point we can see that the controversy sur-
rounding the discovery of the solitary wave was due precisely to the problem
of correctly balancing the eects of dispersion and nonlinearity in the models.
If dispersion is ignored, one obtains the shallow water equations, of which
(1.1.1) is a simple prototype, which produces shocks, or breaking of waves.
At the other limit, if one ignores the nonlinear eects, he obtains an equation
of the type
ut + ux + uxxx = 0 (3.1.2)
which exhibits only a dispersive decay of the waves, and supports neither
periodic wave trains, nor solitary waves.
Boussinesq [18], [19] and Rayleigh [61] took up the investigations of such
solitary waves in the 1870's. Boussinesq [19] actually found the equation
later named for Korteweg and deVries, and, moreover, found the exact sech2
solution for the solitary wave. Boussinesq's treatise was, however, 630 pages
long, and much of his work was not fully appreciated. The Korteweg deVries
32 CHAPTER 3. KORTEWEG DE-VRIES
equation, as we shall see, exhibits both periodic wave trains (the cnoidal
waves), and, in the limit of innite period, the solitary wave.
There are three relevant length scales in the theory, h, the depth of the
uid; l, the length of the wave; and a, the amplitude of the wave. Accordingly,
there are two dimensionless parameters,
" = hl ;  = ha :
The KdV approximation is valid in a regime where  = O("2); and this is
where the solitary wave occurs.
Levi-Civita [48] gave a rigorous proof of the existence of periodic wave
trains in innitely deep water; Struik [72] modied Levi-Civita's analysis to
prove the existence of periodic wave trains in water of nite depth. Levi-
Civita used a conformal mapping technique to map the unknown domain
conformally into a half-plane. Levi-Civita's method is of considerable inter-
est, and may prove also to be useful in other problems. Friedrichs and Hyers
[33] gave a rigorous construction of the solitary wave.
Before describing the derivation of the KdV equation in ion acoustic waves
and in water waves, which is considerably more involved, I would like to
present a very simple derivation of the KdV equation due to Kruskal [42].
He begins with the weakly nonlinear dispersive wave equation2 governing the
evolution
utt = uxx(1 + "ux) + uxxxx;
where ";   1 are small parameters measuring nonlinearity and dispersion
respectively. The linear equation, with " = 0, has dispersion relation
!2 = k2 k4;
so that we must really require  < 0 for well-posedness.
For  = " = 0 the equation reduces to the linear wave equation, which
has as a general solution left and right progressing wave-forms. We look for a
solution of the full equation which is essentially right progressing, with only
a slow variation. To make this precise, we look for a solution of the form
u(x; t; "; ) = w(;  )  = x t;  = "t:
This is the continuum model for the Fermi-Pasta-Ulam system [29]. Cercignani [23]
2
and Palais [59] have given an account of the Fermi-Pasta-Ulam experiment and its rela-
tionship to the Korteweg-deVries equation.
3.1. BACKGROUND AND HISTORY 33
By the chain rule
@u = " @w @w ; @u = @w ;
@t @ @ @x @
and, in the variables ;  the equation becomes
"2w 2"w + w = w (1 + "w ) + w :
We drop the quadratic term in " since it is second order. Putting U = w =2
we obtain the KdV equation for U :
U + UU + 2" U = 0:
This delightfully simple derivation indicates two of the primary ingredi-
ents in the KdV approximation: First, one must specialize to a unidirectional
frame; and Second, one must scale the time variable in an appropriate way.
These two features also appear in the more complicated derivation of the
KdV approximation to the plasma and Euler equations.
Before proceeding with these derivations, I mention the important paper
of Benjamin, Bona, and Mahoney [11], who argue that "in all examples the
assumptions leading to the KdV equation equally well justify the equation
ut + ux + uux uxxt = 0:
Their argument is that, since to lowest order the solution represents a right
moving wave, ut  ux and so one may replace uxxx with uxxt. This
argument is disputed in rather strong terms by Kruskal, cf. [42], though
Bona, Pritchard and Scott [16], [17] have conducted extensive numerical and
physical experiments to ascertain the validity of both the KdV and BBM
approximations. Their conclusions are that \on a long time scale T naturally
related to the underlying physical situation, the equations predict the same
outcome to within their implied order of accuracy". [16] In other words, both
the KdV and BBM approximations are accurate for a nite time scale3, and
they are equally valid models over that time interval.
Craig [25] has shown that the KdV equation is accurate on a time scale of order " 3 ,
3
where " is a small parameter in the theory.
34 CHAPTER 3. KORTEWEG DE-VRIES
3.2 Ion Acoustic Waves in Plasmas
The Korteweg-deVries equation was obtained as a model for magetohydro-
dynmic waves in a cold collsionless plasma by Gardner and Morikawa in an
unpublished article in 1960 [36]. Since that derivation is somewhat easier
than in the case of water waves, we shall discuss that case rst.
We consider a plasma consisting of negatively charged electrons and pos-
itively charged ions. The electrons are treated as a gas and we obtain equa-
tions of motion for the ions. The ion density is denoted by n, the electron
density by ne, the electric force eld by E and the velocity of the ions by v.
The equations of the plasma may be written in the following form [44], [69]
nt + (nv)x = 0; vt + vvx = E;
E + (log ne)x = 0; Ex + ne = n;
where n is the ion density, ne is the electron density, v is the ion velocity,
and E is the electric eld.. We eliminate ne from the equation by dening
' = log ne, ' being the electric potential, and the equations reduce to three
equations in three unknowns [26]
v 2
nt + (nv)x = 0 vt + ( 2 + ')x = 0 'xx e' + n = 0: (3.2.3)
There are a number of perturbation arguments one can try on this system
of equations; and since there is a certain art to the method, I will carry out
the discussion in some detail. First let us try a straightforward perturbation
approach for small disturbances of the equilibrium states. We look for small
perturbations about the quiescent state n = 1; v = c; ' = 0:
n = 1+ "n1 + "2n2 + : : : ; ' = "'1 + "2'2 + : : : ; v = c + "v1 + "2v2 + : : : :
Substituting these expansions into (3.2.3) we obtain, at lowest order, the
linear equations for small disturbances
n1;t + (v1 + cn1 )x = 0;
v1;t + (cv1 + '1 )x = 0;
'1;xx '1 + n1 = 0:
3.2. PLASMAS 35
We determine the dispersion relation for this linear system by looking at
the Fourier modes 0n 1 0a 1
@ v11 A = @a12A ei(kx !t):
'1 a3
This leads to the linear algebraic system
0! ck k 0
1 0a 1
@ 0 ! ck k A @a12A = 0:
1 0 1 + k2 a3
The system has a nontrivial solution i the determinant of the above matrix
vanishes, and this leads to the condition (! ck)2 (1+ k2) k2 = 0, or, solving
for !,
! = ck  p k 2 :
1+k
This dispersion relation is unbounded in k; but it is of quite a dierent nature
than the dispersion relation for the nonlinear Klein Gordon equation. We
will return to it in a minute.
Next, let us consider the eect of rescaling the space and time variables.
Namely, we make the transformations x0 = "x, t0 = "t. In the rescaled
variables the equations take the form
nt0 + (nv)x0 = 0;
2
vt0 + ( v2 + ')x0 = 0;
"2'x0x0 + e' = n:
Letting " ! 0 we obtain formally e' = n, and the nonlinear system
2
nt0 + (nv)x0 = 0; vt0 + ( v2 + log n)x0 = 0:
This is a nonlinear hyperbolic system with characteristic speeds 1; for gen-
eral initial data such systems of equations have shock discontinuities.
The scaling we have just described eliminates the dispersion. In fact, the
linearization of the equations above is the rst order hyperbolic system
nt + vx = 0 vt + nx = 0
36 CHAPTER 3. KORTEWEG DE-VRIES
and this system propagates pulses unchanged; i.e. it is dispersionless. We
may conjecture that the dispersion of the full plasma equations prevents the
build-up of shocks.4
We now derive an approximation to the plasma equations which preserves
the dispersion and leads to the Korteweg-deVries equation. Consider the
dispersion relation for the linearized plasma equations. The so-called long
wave length approximation is to formally approximate !(k) by its Taylor
expansion about k = 0. k = 2=L, where L is the wavelength; k is called:
the wave number. The Taylor expansion of the dispersion relation is ! =
k 21 k3 + O(k5). This is a good approximation for small k, that is, for long
waves.
The leading term k in this approximation corresponds to a dispersionless
system with a wave speed 1; for, if ! = k then
Z1 Z1
b
ei(kx !t) f (k) dk = eik(x t) fb(k) dk = f (x t)
1 1
where fb is the Fourier transform of f . This suggests transforming to a frame
of reference moving with unit speed. Waves traveling to the left are obtained
by taking the negative root for ! in the dispersion relation.
Now observe that the plasma equations are Galilean invariant; that is,
they are unchanged under the one-parameter group of transformations
x 7! x ct; t 7! t v 7! v + c:
This means that the equations are the same in any reference frame, and
we can shift to a moving frame of reference simply by adding a constant
to the velocity v. This amounts to expanding about the quiescent state
n = 1; v = c; ' = 0, where c is the velocity of the reference frame. If our
dispersion relation was originally
p
! = ck p k 2 = p k 2 ( 1 + k2 1);
1+k 1+k
we choose the velocity of the reference frame to be c = 1. Then the dispersion
relation is
! = k p k 2 = k(1 [1 12 k2 + : : : ]);
1+k
This is precisely the situation for the KdV equation itself. Cf. Lax, Levermore, and
4
Venakides [47] for a discussion of the zero dispersion limit of the KdV equation.
3.2. PLASMAS 37
and we obtain, in the long wave approximation,
! =: 21 k3:
The partial dierential operator associated with this dispersion relation
is
@ 1 @3 ;
@t 2 @x3
for which the natural scaling is x0 = "x, t0 = "3t. If we introduce this scaling
into the equations (3.2.3) we obtain (after division by ")
"2nt0 + (nv)x0 = 0;
 v2 
"2 vt0 + 2 +' = 0;
x0

"2'x0x0 + e' = n:
This perturbation scheme is singular, since the character of the equations
is changed when " = 0.
Since only "2 appears in these equations, we expand all quantities in
powers of "2:
n = 1 + "2n1 + : : : ; v = 1 + "2v1 + : : : ; ' = "2'1 + : : : :
When we do this, substitute the expansions into the above equations, and
collect terms, we get at order "2:
(n1 + v1 )x0 = 0; (v1 + '1 )x0 = 0; '1 = n1 :
Assuming v1 and n1 tend to zero as x ! 1 we have n1 = v1 .
At next order we obtain
n1;t0 + (n1 v1)x0 + (n2 + v2 )x0 = 0;
 v2 
v1;t0 + 1
2 + v2 + '2 x0
= 0;

'1;x0x0 + '2 + 21 '21 = n2 :

38 CHAPTER 3. KORTEWEG DE-VRIES
The second order quantities n2 ; v2 ; and '2 may be eliminated from this sys-
tem, and one obtains (dropping the primes) the Korteweg-de Vries equation
for v1:
v1;t + v1v1;x 12 v1;xx = 0:

3.3 Water Waves

Euler's equations for waves in an inviscid, incompressible uid with irrota-
tional ow in the region D = fy : 0  y  h + (x; t)g are

' = 0 0  y h + ;
t + 'xx = 'y on y =h + (x; t)

('t + 21 jr'j2 + g) = 0 on y =h + (x; t)

'y = 0 on y =0:
Here, g is the acceleration due to gravity,  the uid density, and  the
displacement of the uid surface from equilibrium. We have neglected surface
tension. At rest, the uid lies in the region 0  y  h.
The rst equation is Laplace's equation for the velocity potential ', while
the third is Bernouilli's equation at the free surface. The second equation is
a kinematic condition that states that the relations
(x;_ y_ ) = r' and y = h + (x; t)
are consistent. Taking the total (Lie) derivative of the second equation, we
have
y_ = t + xx_ = t + x'x = 'y :
We now turn to a derivation of approximate models to the Euler equa-
tions. There are a number of derivations of the Boussinesq and Korteweg-
deVries equations in the literature, e.g. [9], [74]. Peter Olver [56] has derived
the approximations from a Hamiltonian perspective. We begin by rewriting
the equations in non-dimensional quantities. There are three characteristic
length scales in the problem: l, the length of the wave, the depth of the layer
h, and a, the amplitude of the wave. If we take the natural length scale
3.3. WATER WAVES 39
of the problem to be the length of the wave, l, then there are a priori two
dimensionless parameters:
" = hl ;  = al :
The velocity potential ' has dimensions of velocity  length. We need
to determine an appropriate velocity scale. By dimensional analysis [12] one
can
pgl argue that the velocity of a wave in deep water is a constant multiple of
where l is the wave length. But p in our case there is also the possibility
that the velocity is proportional to gh. This indeterminacy can be resolved
by examining the dispersion relation at k = 0, namely, by computing d!=dk
at k = 0.
The dispersion relation is obtained from a Fourier analysis of the equa-
tions linearized about the state
 = 0; ' = cx 21 c2t:
This is a special solution of the Euler equations which corresponds to a state
in which the uid is moving with uniform horizontal velocity c, and there is
no displacement of the surface.
The linearized equations about this state are:


' = 0; 'y y=0 = 0;

t + cx = 'y ; 't + c'x + g = 0:

The second pair of equations hold on the free surface y = h; while
' = 0
in the strip 0  y  h: To nd the dispersion relation we look for solutions
of the form
(x; t) = a(k; !)ei(kx !t); '(x; y; t) = b(y; k; !)ei(kx !t):
The equation
' = 0 and boundary condition 'y = 0 at the bottom of the
uid layer require that b(y; k; !) = b(!) cosh ky. When we substitute this
expression for ' into the equations at the free surface, we nd
i(kc !)a(k; !) b(!)k sinh hk =0;
i(kc !)b(!) cosh hk + ga(k; !) =0
40 CHAPTER 3. KORTEWEG DE-VRIES
The dispersion relation in the Galilean frame moving with velocity c is
then found to be
(! kc)2 = gk tanh hk:
If we take c = 0 then !2 = ghk2 + : : : ,5 and
d!  = pgh
dk k=0
are the group velocities in the rest frame. They correspond to wave groups
moving to the right and left. By choosing one root, say the negative root,
we select the wave moving with positive velocity, i.e. to the right.
p
This argument suggests that the natural velocity scale is gh. Denoting
dimenionsionless quantities with primes, we have so far
p
x = lx0 ; y = hy0;  = a0; t = t0 ; ' = l gh'0;
where we have still to determine the correct time scale  .
The time scale can be determined by the principle that the Galilean
invariance of the Euler equations be preserved under rescaling. One easily
checks that the quantities
't + 12 '2x; t + 'xx
are invariant under the Galilean transformations
x 7! x ct;  7! (x ct; t) ' 7! '(x ct; t) + cx 12 c2t:
In order that the Galilean invariance of the Euler equations be preserved
under rescaling, we require these quantities to transform homothetically un-
der scaling. Note that the rst quantity has dimensions of velocity squared,
while the second has the dimensions of velocity. Let us see how they trans-
form under scaling: We nd
1
 l 1

2
't + 2 'x = gh p 't0 + 2 ('x0 ) ;
0 0 2
 gh
p  l 0 0 0
t + x'x = " gh p t0 + x0 'x0 ;
 gh
5 tanh  =  1 3
6 +:::
3.3. WATER WAVES 41
hence these quantities transform homothetically under scaling provided we
take
 = pl :
gh
The Galilean invariance of the Euler equations is then preserved under non-
dimensionalization.
In nondimensional units the Euler equations then take the form (dropping
the primes)
"2'xx + 'yy =0; 0  y  1 + ; (3.3.4)
"2[t + 'xx] ='y ; y = 1 + ; (3.3.5)

"2['t + 21 '2x + ] + 12 '2y =0; y = 1 + ; (3.3.6)

'y =0 y = 0: (3.3.7)
The parameter  can be scaled out of the equations by setting w = .6
When this is done the equations take the form
"2'xx + 'yy =0; 0  y 1 + w; (3.3.8)
"2(wt + wx'x) ='y ; y = 1 + w; (3.3.9)

"2('t + 21 '2x + w) + 12 '2y =0; y= 1 + w; (3.3.10)

'y =0 y= 0: (3.3.11)
Before proceeding, it will be useful to note that (3.3.9) can be written in
the simplied form
d Z 1+w
wt + dx 'x dy = 0: (3.3.12)
0
6We shall see below that this scaling of the surface height is consistent with the re-
quirement that  = O("2 ) in the KdV approximation. Whitham [74] instead rescales
the velocity potential '. This has the consequence of keeping the parameter  in the
equations.
42 CHAPTER 3. KORTEWEG DE-VRIES
This form is obtained as follows. From Laplace 's equation (3.3.8) and the
boundary condition 'y = 0 on y = 0 we obtain
 Z 1+w
'y y=1+w = " 2 'xx dy:
0
Now note that
d Z 1+w ' dy = w ' + Z 1+w ' dy:
x x x xx
dx 0 0
This is a singular perturbation problem when "  1, that is, when the
waves are long compared to the depth of the uid layer h. This is the shallow
water case. We look for solutions by expanding in powers of "2 :
' = '0 + "2'1 + : : : ; w = w0 + "2w1 + : : : :
Substituting these expansions in (3.3.4) we obtain, successively,
0 ='0;yy + "2('0;xx + '1;yy ) + "4('0;xx + '1;yy ) + : : : (3.3.13)
0 = '0;y + "2(w0;t + '0;xw0;x '1;y + : : : (3.3.14)

0 = 21 '20;y + "2('0;t + 21 '20;x + w0 + '1;y '0;y ) + : : : (3.3.15)

The uid is assumed to occupy the region 0  y  1 + w(x; t), with 'y = 0
on y = 0. Hence
'j;y jy=0 = 0; j = 0; 1; : : : (3.3.16)
From the rst equation we nd that '0;yy = 0, hence '0 = a(x; t)y+b(x; t).
The boundary condition (3.3.16) implies that a = 0, hence '0 = '0(x; t), and
'0;y = 0. From 
'0;xx + '1;yy = 0; '1;y  = 0 y=0
we nd '1 = '0;xxy2=2 + b(x; t) for some function b(x; t). Hence

'1;y y=1+w = '0;xx(1 + w0):
3.3. WATER WAVES 43
We now obtain, for the second and third equations of (3.3.15),
0 =w0;t + '0;xw0;x + '0;xx(1 + w0)

0 ='0;t + 1 '20;x + w0:

2
Putting u = '0;x and z = 1 + w0 , and dierentiating the second equation
with respect to x, we obtain the shallow water equations
 2 
zt + (uz)x = 0; ut + z + u2 = 0: (3.3.17)
x
The KdV approximation is obtained at the next order by perturbing
from the regular solution of the shallow water equations given by w = 0,
' = cx c2t=2 for an appropriate choice of c. The linearization of the Euler
equations at this solution is

' = 0; wt + cwx = 'y ; 't + c'x + w = 0:
The dispersion relation for this system is
(! ck)2 = k tanh k:
Writing this as  ! ck 2 tanh k
k = k ;
and taking the negative square root,7 we obtain ! ck = k( 1+ k3=12+ : : : );
or,
! = k(c 1) + 12 1 k3 : : :

The KdV equation is obtained as a bifurcating solution of the Euler equa-

tions when we take c = 1. The dispersion relation is then
! = 121 k3 + O(k5);
7The Euler equations are invariant under the re ection x 7! x whereas the KdV
equation is not. This symmetry breaking is a consequence of selecting one or the other of
the roots.
44 CHAPTER 3. KORTEWEG DE-VRIES
and, as in the case of the plasma equations, this dispersion relation suggests
the scaling of the variables "x, "3t. We look for solutions of the form
w ="2 w0(;  ) + "4w1 (;  ) + : : : ;

' =x 21 t + "'0(; y;  ) + "3'1 (; y;  ) + : : : ;

 = "x;  = "3t:
Note that the surface height w is of order "2 in this scaling. Recall
that we took w = . Hence the original variable  is O(1) provided that
"2 = O(). This is the correct regime for the validity of the Korteweg-deVries
approximation.
Substituting the expansion for ' into Laplace's equation, we obtain
'0;yy = 0; '1;yy + '0; = 0; ::: :
Hence '0 = a(;  ) + y b(;  ). The boundary condition 'y = 0 on y = 0
implies that b = 0, hence '0 is independent of y and '0 = '0 (;  ). At the
next order we nd 2
'1 = y2 '0; + b(;  )
Without loss of generality we may take
2
'1(; y;  ) = 1 2 y '0; + b1 (;  );
where b1(;  ) is to be determined.
We substitute these expansions into the Bernouilli equation. On the free
surface, we have
't + 12 '2x = 12 + "4'0; +


+ 21 (1 + "2'0; + "4'1; + : : : )2
4
="4'0; + "2 '20; + "2'0; + O("6)
On the free surface y = 1 + w = 1 + "2w0 + : : : , 1 y2 = O("2), hence
"3'1;x = "4'1; = "4b1; + O("6):
3.3. WATER WAVES 45
Hence to order "6 we have
"2('0; + w0) + "4('0; + 21 '20; + b1; + w1) + O("6) = 0:
Therefore
w0 = '0; ; (3.3.18)

w1 = '0; 1 '2 b : (3.3.19)

2 0; 1;
Substituting these expressions into (3.3.12), we obtain, to order "5,
Z 1+"2w0
"5w0; + " d
d 0 (1 + "2'0; + "4'1; ) dy = O("7):
Now
Z 1+"2w0
(1 + "2'0; + "4'1; ) dy =(1 + "2w0 ) + "2'0; (1 + "2w0 )
0
 
+ "4 '0; w0 + 31 '0; + b1; + : : : :
Combining these results, we obtain
"3(w0; + '0; ) + "5(w0; + (w1 + b1; + 31 '0; w02) ) +


= 0:
The coecient of "3 vanishes by (3.3.18). Setting the coecient of "5 to zero
and using (3.3.19) we obtain, after a short calculation, the Korteweg-deVries
equation
w0; 16 w0; 32 w0w0; = 0:
The function b1; self-cancels and does not appear at this stage of the
perturbation theory. It is determined at the next stage of the perturbation
series. At each stage the two functions wj (;  ), 'j (; y;  ), are determined
by the three Euler equations. This leads to a solvability condition at each
stage, expressed in terms of the values 'j (; 1;  ). The term b1 = '(; 1;  )
satises a linearized KdV equation, linearized at w0.
46 CHAPTER 3. KORTEWEG DE-VRIES
3.4 The Solitary wave of the KdV equation
Let us begin by constructing the solitary wave solution of the Korteweg-
deVries equation. We look for traveling wave solutions u(x; t) = u( ), where
 = x ct. Substitution of this form into (3.1.1) leads to the ordinary
dierential equation
!u0 + u000 + 12 (u2)0 = 0: (3.4.20)
This equation may be integrated once; taking the constant of integration to
be zero, we obtain
u00 + 21 u2 !u = 0:
Multiplying by u0 we get
1 d
 u02 u3 u2 
0 00 2
u (u + 2 u !u) = d 2 + 6 ! 2 = 0;
hence
u02 + u3 ! u2 = E; (3.4.21)
2 6 2
where E is a constant of integration.
We may plot the level curves of this equation qualitatively. Equilibrium
points in the phase plane (u; u0) occur at (0; 0) and at (3!; 0) when E = 0.
For E = 4!3=3 there is an equilibrium (a center) at (4!; 0) in the phase
plane.
The solitary wave is given explicitly by the formula
u(x; t) = 12!2sech2(!(x x0 ) 4!3t):
The solitary wave is a 2-parameter family of solutions, parametrized by the
phase x0 and amplitute !2. The speed is 4!2,
Korteweg and deVries obtained the solitary wave as a limit of the periodic
solutions. These are elliptic functions, while the solitary wave is a singly
periodic function. The periodic solutions of the KdV equation correspond to
the periodic solutions originally found formally by Stokes and later proved
by Levi-Civita, Struik. The solitary wave corresponds to the solitary wave of
the Euler equations, proved rigorously by Friedrichs and Hyers [33] in 1954.
3.4. THE SOLITARY WAVE OF THE KDV EQUATION 47

1.5

E=0

1
homoclinic orbit

0.5 E<0

−0.5

−1

−1.5
−0.5 0 0.5 1 1.5 2 2.5 3 3.5

Figure 3.1: A family of periodic solutions surrounds the center, bounded by

a homoclinic orbit from the origin to itself, for E = 0. The homoclinic orbit
is the solitary wave, while the periodic solutions give rise to periodic wave
trains. The KdV equation thus provides a model which is readily analyzed
and which yields the qualitative features of the full Euler equations

In addition to the solitary wave and the periodic wave, the KdV equation
has other special solutions which are interesting. Foremost among these are
the multi-soliton solutions. For any integer n  1 these are given explicitly
by the formula
d2  dj e 8!j3 te !j +!k)x 
Un(x; t) = 12 dx2 log jk + !j + !k ;
where !j and dj are positive constants, with !1 < !2 <


< !n. This Un
constitutes a 2n parameter family of exact solutions of the KdV equation,
parametrized by !j and dj . Asymptotically, as t !  1 these solutions tend
uniformly in x and exponentially fast in time [38] to a superposition of n
solitary waves. Roughly, the dj constitute relative phases of the waves, while
the wave speeds are given by 4!j2; j = 1; : : : n.
The interaction of two solitary waves is depicted in the sequence Figure
48 CHAPTER 3. KORTEWEG DE-VRIES
3.2.
Initial Profile t=.2 t=.4

1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
0 5 0 5 0 5

t=.5 t=.6 t=1

1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
0 5 0 5 0 5

Figure 3.2: The interaction of 2 solitary waves in the exact solution. Two
solitary waves are pictured in the rst frame. As time progresses the two
solitary waves interact and separate. Note the dip in the larger wave as they
interact, indicating clearly that the interaction is nonlinear and not a simple
superpostion. After the interaction they regain their shape, but are displaced
from where they would have been had there been no interaction.
As t ! 1, Un(x; t) looks like a superposition of n solitary waves. As
time evolves, the faster waves overtake the slower (smaller amplitude) waves
and interact nonlinearly. As t ! 1 the waves are arranged in the order of
increasing amplitude and again assume the shape of the sech2 solitary wave.
The only evidence of the nonlinear interaction is a phase shift in the waves.
That is, they are displaced from the position they would occupy if they
were simply a linear superposition of solitary waves. In the appendix I have
inserted a Matlab program which numerically integrates the KdV equation
with a superposition of two solitary waves as initial data.
Chapter 4
Isospectral Deformations
4.1 The KdV hierarchy
The Korteweg-deVries (KdV) equation,
ut + uxxx + uux = 0; (4.1.1)
has smooth solutions for all positive and negative time given initial data
which is suciently smooth, say C 3. We won't go into the existence theory
of the KdV equation here since the existence and uniqueness issues are ulti-
mately resolved very simply by the method of inverse scattering, and since
the algebraic/ geometric structure of this simple nonlinear equation is far
more interesting.
The KdV equation has as a special solution the solitary wave,
u(x; t) = 12c2sech2 [c(x 4c2t)]
These waves move to the right with speed 4c2. Note that their amplitude
depends on the wave speed, and that larger waves travel faster. One could
choose as initial data two solitons separated by a distance great enough so
that their interaction was extremely small, since they decay exponentially in
either direction. Suppose the soliton to the left is larger. As time evolves, the
larger soliton will overtake the smaller one. Since the equation is nonlinear
they will react in a nonlinear way. After a period of time the two solitons
again separate, the larger one moving ahead to the right and regaining its
original shape. For large time, the two solitons are perturbed only by a
phase shift: they are not quite where they would be had they been purely
49
50 CHAPTER 4. ISOSPECTRAL DEFORMATIONS
solitary waves. These facts were discovered by computational experiments
by Kruskal and Zabusky in the mid 60's [43].
Moreover, the same thing happens when the initial data consists of several
solitons, separated originally into distinguishable solitary waves. As time
progresses, the faster solitons overtake the slower ones, and as time goes
to innity, the solution evolves into separated solitons, each with its own
original amplitude and speed, but with slightly displaced phase. In Chapter
8 I have described a Matlab le that numerically computes and animates the
two soliton interaction.
The discovery of the elastic scattering of the solitons for the KdV equa-
tion prompted an intense, and, as it turned out, highly fruitful theoretical
investigation of the KdV equation. The original theoretical breakthrough
was made by Gardner, Greene, Kruskal, and Miura, [35] Later researchers
claried and simplied their arguments, and ultimately constructed myriad
further examples of such special systems. One of the early papers which had
an enormous in uence on the development of the subject was the 1968 paper
by Peter Lax [45]. Gardner, Greene, Kruskal, and Miura [35] had found that
the eigenvalues of the Schrodinger operator
L = D2 + 61 u; D = dx d;
are constant in time if u evolves according to the KdV equation.
The early papers in the subject were complicated by the extensive compu-
tations that accompanied the original discoveries. Lax simplied and claried
the situation conceptually by introducing the Heisenberg picture. Suppose
that the family of operators fL(t)g is unitarily equivalent under the ow.
Assume U is a one-parameter family of unitary operators:
UU  = I; Ut = BU; U  L(t)U = L(0);
where B is a skew-adjoint operator. Dierentiating the third equation with
respect to time we get
U  B  L(t)U + U  Lt U + U  LBU = 0;
hence
Lt + B L + LB = 0:
In these calculations we interpret Lt as the operation of multiplication by
the function ut. Since B  = B this equation reduces to the Lax equation
Lt = [B; L] (4.1.2)
4.1. THE KDV HIERARCHY 51
where [B; L] is the commutator BL LB . The pair of operators L and
B is called a Lax pair. Equation (4.1.2) is the Heisenberg picture of the
Korteweg-deVries equation.
Since Lt is a multiplication operator, the commutator [B; L] must also be
a pure multiplication operator. For example, taking B = D we nd
[D; L] = 1 ux; L 1
t = ut
6 6
and the Lax equation is ut = ux. This equation generates the one parameter
family of translations, u(x; t) = u0(x + t); and so, of course, L is unitarily
equivalent under the ow.
The KdV equation itself is obtained by taking a third order skew adjoint
operator
B = 4D3 + 21 (uD + Du)
The details of the calculation are left as an exercise.
In all these calculations we may replace L by L + k2, so the KdV equation
is formally obtained as a consistency condition for the overdetermined system
of partial dierential equations
(D2 + 16 u + k2 ) = 0; @ =B :
@t
The two isospectral ows
ut = ux; ut + uxxx + uux = 0
are only two ows in an innite hierarchy of commuting Hamiltonian ows.
this hierarchy of ows is generated by a recursion relation, namely
DFj+1 = ( D3 31 (uD + Du))Fj ; F1 = u: (4.1.3)
The j th ow is then given by
ut = DFj :
The recursion relation (4.1.3) will be derived in section (7.2). This recursion
relation was rst proposed by A. Lenard; Peter Lax [46] showed that each Fj
52 CHAPTER 4. ISOSPECTRAL DEFORMATIONS
in the recursion relation is a dierential polynomial in u and, furthermore, is
the gradient of a functional, Hj , so that the ows have the form
Hj :
ut = D u
For example, the rst two terms in this recursion relation and their corre-
sponding functionals and ows are
Z1 1
F1 = u; H1 = 2
2 u dx; ut = ux;
1

u2 ; Z1 1 3
F2 = D 2 u H1 = u 2 u dx; ut = uxxx uux:
2 2 x 6
1
The operators D; (D3 + 31 (uD + Du) are an example of a bi-Hamiltonian
pair. Bi-Hamiltonian pairs of operators can be used to generate hierarchies
of commuting Hamiltonian ows [51], [57].
The n soliton solutions of the Korteweg de-Vries equation also have a
variational characterization. The KdV equation has an innite number of
conserved quantities, which may be represented as integrals. These are gen-
erated, for example, by the Lenard relation discussed above. The rst few
are:
Z1 Z1 Z1 u2 1 u3 dx
I0 = u dx; I1 = u2 dx; I2 = x
2 6
1 1 1
Z1 2
I3 = au4 + buu2x + u2xx dx
1
for some constants a and b. These integrals are all in involution with respect
to the Gardner Poisson bracket, dened by
Z1 F d G
fF; Gg = dx:
u dx u
1
4.1. THE KDV HIERARCHY 53
That is,
fIj ; Ik g = 0:
Each of the functionals Ij is conserved under the KdV ow; and each of them
generates a Hamiltonian ow.
The multi-soliton solutions can be obtained from a simple variational
principle. For example, consider the constrained variational problem
min I :
I1 =c1 2
The Euler-Lagrange equations for this constrained minimum are
I2 + c I1 = 0:
u u
This leads immediately to (3.4.20), with the Lagrange multiplier c precisely
equal to the wave speed. More generally, the n-soliton solution is obtained
as the constrained variational problem
min I
I1 =c1 ;:::In =cn n+1
The associated Euler-Lagrange equation for the n soliton solution is then an
ordinary dierential equation of order 2n + 1:
In+1 + c In + : : : c I1 = 0:
u n
u 1 u
The orbital stability of the n-soliton solutions can be proved using this
variational characterization [50]. Such variational arguments go back to
Boussinesq and have been used more recently by Benjamin [8] to prove the
stability of the solitary wave. A simplied sketch of the argument is as fol-
lows. Since all the integrals Ij are conserved under the KdV ow, the ow is
restricted to a manifold Mn dened by
Mn = fu : I1(u) = c1; : : : ; In(u) = cng:
The n-soliton solution Un is obtained as the minimum of the functional In+1
on this manifold. Hence the level surfaces of In+1 on Mn in a neighborhood
of Un are closed manifolds surrounding Un. If the initial data is initially in
a neighborhood of Un , it must therefore remain in that neighborhood under
the ow.
54 CHAPTER 4. ISOSPECTRAL DEFORMATIONS
4.2 The AKNS hierarchy
All of the known examples of innite dimensional completely integrable sys-
tems arise as compatibility conditions for an overdetermined system of equa-
tions. Two well-known examples, the modied KdV equation
ut + u2ux + uxxx = 0; (4.2.4)
and the nonlinear Schrodinger equation
ut = iuxx + 2ijuj2u; (4.2.5)
arise in such a way. They can be obtained as `zero-curvature' conditions for
the connection
Dx = @x@ zJ q; D = @ A (4.2.6)
t
@t
where J is a 2  2 diagonal matrix, q is a 2  2 o-diagonal matrix, and A is
a traceless 2  2 matrix which is a dierential polynomial in the entries of q.
The matrix A is generated by the following simple routine, and the argu-
ments go through without change even in the d  d case, where
0 1 0 1
 0 q12 : : : q1n
BB 1 2 CC Bq21 0 : : : q2nCC
J =B@ ... CA ; q(x) = B
B@ ... CA :
n qn1 : : : 0
We look for wave functions satisfying Dx = 0 of the form = mexzJ .
It is easily seen that m satises the dierential equations
mx = z[J; m] + qm (4.2.7)
The functions m(x; z) can be normalized so that they tend to the identity
matrix I as x ! 1; and in that case they tend also to I as z ! 1. The
functions m are called the reduced wave functions.
Just as the Schrodinger operator generates the hierarchy of isospectral
ows known as the KdV hierarchy, the operator Dx generates a hierarchy of
deformations in q, i.e. ows, for which the connection in (4.2.6) is at. The
argument runs as follows:
4.2. THE AKNS HIERARCHY 55
Let  be a constant matrix and set F = mm 1 : Then
Fx =mxm 1 mm 1 mxm 1 = [mxm 1 ; F ]
=[zJ zmJm 1 + q; F ]
=[zJ + q; F ]:
If q 2 S (R ) then it can be shown that F has an asymptotic expansion valid
as z ! 1 along any ray in the complex plane.
X
1
F Fj z j ; F0 = :
j =0
Substituting this asymptotic expansion into the equation for F we obtain
a set of recursion relations, analogous to the Lenard relation for the KdV
hierarchy,
[J; F0 ] = 0; d q; F ] = @Fj [q; F ]:
[J; Fj+1] = [ dx (4.2.8)
j j
@x
Since F0 =  and [J; ] = 0,  must be a diagonal matrix. With no loss of
generality, we may work in the class of traceless matrices, hence we take 
to be a diagonal matrix of trace 0.
Now dene
@ (z n F ) ; X
n
Dt = @t + (z n F ) += Fj z n j :
j =0
In particular, we see that the t component of the connection is determined
by the x component; in fact the function F which satises Fx = [zJ + q; F ]
is a generating function for the entries of Dt. It is a simple consequence of
the recursion relations (4.2.8) that:
d q; F ] = q [J; F ]
[Dx; Dt] = qt [ dx n t n+1

We say that fDx; Dtg is a at connection if

[Dx; Dt] = 0: (4.2.9)
56 CHAPTER 4. ISOSPECTRAL DEFORMATIONS
This is called the zero curvature condition and is the analog of the Lax
equations in the KdV hierarchy. By the foregoing, the connection is at if
and only if q satises the nonlinear evolution equation
d q; F ]
qt = [J; Fn+1] = [ dx (4.2.10)
n

While the reduced wave functions m are nonlocal in the potential q, it is

a remarkable fact that the expressions Fn are local in q [67]:
Theorem 4.2.1 The o-diagonal entries of Fn are polynomials in q and its
derivatives up to order n 1, whereas its diagonal entries are dierential
polynomials of order n 2. Hence (4.2.10) is a system of nonlinear partial
dierential equations.
Nonlinear evolution equations solvable by the inverse scattering method
are generated by solving the recursion relations (4.2.8), starting with F0 = ,
a traceless, diagonal matrix, and computing successively the Fj up to order
n. The nonlinear evolution equation is then given by (4.2.10).
Example Consider the hierarchy of ows generated by the 2  2 isospectral
operator
@
1 0  0 1
Dx = @x z3 q; 3 = 0 1 ; q = u 1 0 :
We take  = 3 as well. Then
[J; F1 ] = [ d q; F0] = [3 ; q];
dx
hence [3 ; F1 q] = 0, and we may take F1 = q. At the next stage,
 
[3 ; F2] = @F
@x
1 [q; F ] = q = u 0 1 :
1 x x 1 0

Choosing 0 b 
F2 = a3 + c 0
we nd  0 b  0 1
[3 ; F2] = 2 c 0 = ux 1 0 ;
4.2. THE AKNS HIERARCHY 57
hence
b = c = u2x :
The coecient a is determined at the next stage.
The equation for F3 is
@F 1
0 1
[3 ; F3 ] = 2
@x [q; F2] = ax3 + 2 uxx 1 0
1  0 1
=(ax + uux)3 + 2 uxx + 2au 1 0 :
Since the diagonal entries of [3 ; F3 ] vanish, the coecient of 3 on the right
side of the above equation must vanish as well, hence we obtain
2
ax + uux = 0;a = u2 :
Note that the expression for ax is an exact derivative, so that a is local in u,
as is assured by Theorem (4.2.1); this fact holds at all orders of the recursion
procedure, even in the d  d case.
We thus have
u 2 1
0 1
F2 = 2 3 + 2 ux 1 0 :
By the same argument we nd
1 1 u3
0 1
F3 = a3 3 + 4 uxx 1 0 ;2
where a3 is determined in the computation of F4. Namely, we nd
@F @
1 1
0 1
3 3
[3 ; F4] = @x [q; F3] = a3;x3 + 2a3 q + @x 4 uxx 2 u 1 0 :

This time we take a3 = 01, so that the ow is

@ 1u
 1 u3
0 1
qt = @x 4 xx
2 1 0 ;
1 In fact, the diagonal entries of all the Fj , j odd, are found to vanish.
58 CHAPTER 4. ISOSPECTRAL DEFORMATIONS
or, in terms of the scalar function u:
ut = 14 uxxx 23 u2ux:
This is one version of the modied Korteweg-deVries equation.
One can also start with the potential
 0 1
q=u 1 0
and obtain the equation
ut = 41 uxxx + 32 u2ux:
Similarly, the hierarchy generated by the skew-Hermitian potential
 0 u
q= u 0
includes, at second order, the nonlinear Schrodinger equation.
Chapter 5
Inverse Scattering Theory
5.1 The Schrodinger equation
The forward problem
Consider the eigenvalue problem for the Schrodinger equation
(D2 + k2 + 61 u) = 0; (5.1.1)
where u is real and lies in S . S (R ) is the class of all C 1 functions on the
real line for which
sup jxmDnuj < +1
x
for all non-negative integers m and n. It follows that all derivatives of u tend
to zero as x ! 1 faster than any power of x. Such functions are said to
be rapidly decreasing.
The scattering theory for (5.1.1) can be developed with much weaker
assumptions on q, but the innite hierarchy of ows and conservation laws
is well-dened only for innitely dierentiable q. Moreover, the scattering
transform, like the F ourier transform, preserves S . (cf. x5.3).
Equation (5.1.1) can be converted to a Volterra integral equation, for
example:
1 Z 1 sin k(x y)
ikx
+ (x; k) = e + 6 k u +dy: (5.1.2)
x
The subscript \+" indicates that the solution is analytic in the upper half
k plane. It is convenient to consider the reduced wave functions, dened by
59
60 CHAPTER 5. INVERSE SCATTERING THEORY
m+ = + e ikx. These satisfy the following integral equation for m+:
1
Z 1 1 e 2ik(x y)
m(x; k) = 1 + 6 2ik umdy:
x
Since
1 e 2ik(x y)
2ik
is uniformly bounded on the interval of integration for Im k  0, this integral
equation can be solved by succesive approximations when u 2 L1 . Its solution
m+ is analytic in the upper half k plane, continous onto the real axis, and
tends to 1 as x tends to +1 or as k tends to innity in Im k > 0.
Letting k tend to zero we obtain the integral equation for m(x; 0):
1 Z1
m(x; 0) = 1 + 6 (x y)umdy:
x
This equation may be solved by successive approximations if
Z1
(1 + jyj) juj dy < +1: (5.1.3)
1
Thus a slightly stronger condition integrability condition on the potential is
needed in order to construct the wave functions for k = 0. This is thus the
minimal condition on the potential needed to construct the wave functions,
and hence to solve the forward scattering problem. For simplicity in these
lectures, we shall always assume the much stronger condition u 2 S : The
function m(x; k) is called the reduced wave function.
The wave function + = m+eikx is analytic in the upper half k plane
and is asymptotic to eikx as x ! 1. By constructing other Volterra integral
equations for the solutions of (5.1.1) we obtain solutions ,  which are
analytic in half planes = k  0 or = k  0 and have the asymptotic behavior
  eikx; x ! 1; e
ikx; x ! 1:
We may write +(x; k) = l(x; k)e ikx, etc. The reduced wave functions
m and l have the following asymptotic properties.
 
5.1. THE SCHRODINGER EQUATION 61
Lemma 5.1.1 Let the potential u be C 1, satisfy (5.1.3), and let its derivative
be in L1(R ). Then the reduced wave function m tends to 1 as x ! 1,
uniformly for = k  0 while l tends to 1 as x ! 1, uniformly for k in the
closure of the upper half plane. Moreover,
1 1
m(x; k) = 1 + O k ; mx = O ; k (5.1.4)
uniformly for 1 < x < 1 as k ! 1 in = k  0. The reduced wave
function l has the same asymptotic properties in k, uniformly for x 2 R .
Proof. The exponential term e 2ik(x y) is uniformly bounded by 1 in
absolute value for = k  0 and x < y < 1. Hence the reduced wave
function m, obtained by successive approximations, is uniformly bounded,
say jm(x; k)j  M , for all x 2 R and all k, = k  0. Hence
Z1  1 e 2ik(x y)  Z1
jm(x; k) 1j  6  2ik  ju mj dy  jkj ju(y)j dy  jCkj :
1 M
x x
From (5.1.2) it follows that the reduced wave function m is C 1. To get
the estimate on its derivative, dierentiate (5.1.2) with respect to x and then
integrate by parts. We have
1 Z 1
mx = 6 e 2ik(x y)u(y)m(y; k) dy
x
1 Z1 1 e 2ik(x y) d
=
6 2ik dy u(y)m(y; k) dy:
x
From the rst term it follows that mx is uniformly bounded and continuous.
In the second term we note that j(u m)y j  jujjmy j + juy jjmj, hence (u m)y 2
L1 (R ), and therefore
1
Z d   
1 
mx  6jkj  dy u(y)m(y; k) dy = O k1 :
x
The same..results are obtained for the reduced wave function l by similar
arguments. `k
62 CHAPTER 5. INVERSE SCATTERING THEORY
It is easily seen from (5.1.2) that for real k, +(x; k) and (x; k) sat-
isfy the same Volterra integral equation. By uniqueness, then, +(x; k) =
(x; k) on the real axis. We may write this as
+ (x; k) = (x; k);
hence and + are Schwartz re ections of each other across the real axis.
This identity holds everywhere in =k  0. A similar argument applies to the
wave functions  normalized at 1. Thus we have
+(x; k) =  (x; k); + (x; k) = (x; k); = k < 0: (5.1.5)
The asymptotic behavior of + and as x ! 1 shows that they must
be linearly independent, so  may be expressed as linear combinations of
 . We leave it to the reader to show that there exist constants a = a(k)
and b = b(k) such that
+ = a(k) + b(k) + ;  = b(k) + a(k) +: (5.1.6)
The Wronskian of two functions f and g is given by W (f; g) = fg0
f 0g. It is a standard fact that the Wronskian of two solutions of (5.1.1)
is independent of x. Therefore one may evaluate Wronskians of the wave
functions as x ! 1 and verify the following relations
W (+; +) = 2ika(k); W ( ; +) = 2ikb(k) (5.1.7)
It follows immediately that a(k) is analytic in the upper half k plane. fur-
thermore, using the Wronskians one can show
a(k) = a( k); b(k) = b( k); ja(k)j2 jb(k)j2 = 1 (5.1.8)
The zeroes of a in the upper half plane correspond to the discrete spec-
trum of L. In scattering theory these are called the bound states. At a zero
kj of a we have
+(x; kj ) = cj +(x; kj )
for some constant cj . Since + and + decay as x ! 1 respectively, they
each in fact decay at both ends, and so constitute a bound state, that is, an
L2 for L.
 
5.1. THE SCHRODINGER EQUATION 63
Lemma 5.1.2 The zeroes of a in the upper half plane lie on the imaginary
axis, the corresponding eigenfunctions of L are real, and the zeroes of a are
simple. In fact, Z1
0
icj a (kj ) = 2j dx
1
where j is the eigenfunction associated with the zero kj of a.
Proof. We leave it to the reader to show that the zeroes of a are imaginary
and that the corresponding eigenfunctions are real. We also leave it to him
to establish the identity
Z1
lim
x!+1
W (0 ; 
j j ) = 2kj 2j dx (5.1.9)
1
where j is the real L2 eigenfunction of L corresponding to kj , and the
prime denotes dierentiation with respect to k. Dierentiating the identity
W (+; + ) = 2ika(k) with respect to k and setting k = kj , where kj is a
zero of a we get
2ikj a0(kj ) = W (0j ; j ) + W (j ; j0 )
where j = (x; kj ), etc. Since j = cj j ,

2ikj a0(kj ) =cj 1W (0j ; j ) + cj W ( j ; j0 );

2ikj cj a0(kj ) = x!lim W (  0 ;  ) + c2 W ( ; 0 )

j j j j j
Z1
+1
=2kj 2j dx:
1
..k
(Note that W ( j ; j0 ) tends to zero as x tends to +1.) `
The transmission and re ection coecients are dened by
r(k) = ab((kk)) ; t(k) = a(1k) :
Then from (5.1.8) it follows that jt(k)j2 + jr(k)j2=1. The scattering data for
L is the set
fr(k); kj ; cj g: (5.1.10)
64 CHAPTER 5. INVERSE SCATTERING THEORY
In general, the re ection coecient r cannot be analytically continued o the
real line1 In any event, we always have:
Lemma 5.1.3 Let the potential u satisfy the conditions of Lemma (5.1.1).
Then a(k) tends to 1 as k tends to innity in the closure of the upper half
k-plane, and jrj ! 0 as k ! 1 along the real axis.
Proof. We have
W (+ ; + ) = W (e = k  0;
ikxl; eikx m) = W (l; m) + 2ik l m;

where + = le ikx. By Lemma (5.1.1) it follows that W (l; m) ! 0 as k ! 1;

hence
 
a(k) = W (2+ik; +) = lm + W (l; m) = 1 + O k1 ; k ! 1:
From the relationship jrj2 + jtj2 it follows that if a ! 1 along the real k
axis, then r(k) must tend to zero as k ! 1 along the real axis. ` ..k

We shall prove below that the data (5.1.10) is sucient to reconstruct

the potential u. We shall restrict our discussion to the case where there are
only a nite number of bound states kj ; j = 1; : : : ; N .
Here is the remarkable fact discovered by Gardner et. al.
Theorem 5.1.4 Under the KdV ow the scattering data evolves as follows:
a is constant,
r(k; t) = e8ik3 tr(k; 0); kj = const: cj (t) = cj (0)e8ikj3t
This is a remarkable fact: the scattering data evolves linearly even though
the evolution equation for u is nonlinear. The scattering transformation thus
linearizes the ow.
Proof. Let +(x; t; k) be the wave function of L which is asymptotic to
e ikx as x ! 1. Since u satises the KdV equation, we have
0 = ( @ B )(D2 + k2 + 1 u)+ = (D2 + k2 + 1 u)( @ B )+
@t 6 6 @t
If the potential u decays exponentially, however, then r can be analytically continued
1
into a strip containing the real axis. If u has compact support, then r is an entire function.
 
5.1. THE SCHRODINGER EQUATION 65

hence ( @ B )+ is also a wave function for L. Now

@t
( @ B )+  4D3e ikx = 4ik3 e ikx x ! 1:
@t
The wave functions are uniquely determined by their asymptotic behavior at
innity, hence
( @ B )+ = 4ik3+:
@t
On the other hand, +  a(k; t)e ikx + b(k; t)eikx as x ! +1, so
@ B ) (_a + 4ik3 a)e ikx + (b_ 4ik3)eikx
( @t +
=4ik3 (ae ikx + beikx);
and it follows that a_ = 0, b_ = 8ik3 b. This establishes that a is constant and r
evolves as given above. Since a is constant its zeroes are xed, so the kj are
constant. The evolution of the coupling coecients is derived by a similar
argument. ` k
..
We now derive the Gel'fand-Levitan-Marchenko integral equation. The
Fourier transform and its inverse for a function in L2 (R ) is
Z Z
ef (s) = 1 e iksf (k)dk; f (k) = 1 1 eiksfe(s)ds:
1 2 1
Let K (x; s) denote the Fourier transform of m+ 1. Since m+ 1 is
analytic in the upper half k-plane and tends to 0 as k tends to innity,
K (x; s) = 0 for s < 0 by the Paley-Wiener theorem. Hence
1 Z1
m+ (x; k) 1 = 2 eiksK (x; s)ds:
0
and
Z1
+ (x; k) = eikx + eiksG(x; s)ds; (5.1.11)
x
where G(x; s) = (2) 1K (x; s x). From the symmetry
+ (x; k) = + (x; k)
66 CHAPTER 5. INVERSE SCATTERING THEORY
we nd that K and G are real, hence
Z1
(x; k) = e ikx + e iks G(x; s)ds:
x
Theorem 5.1.5 Let r(k); kj ; cj ; j = 1; : : : ; N be the scattering data for
the operator L = D2 + 61 u and let G(x; s) be the Fourier transform of + as
given in (5.1.11). Then G satises the Gel'fand-Levitan-Marchenko integral
equation
Z1
G(x; s) + f (x + s) + G(x; t)f (t + s)dt = 0; s > x; (5.1.12)
x
where
Z1 XN
1
f (s) = 2 e r(k)dk + ia0c(jk ) eikj s:
iks (5.1.13)
1 j =1 j

Moreover G satises the hyperbolic equation

Gxx Gss + 61 u(x)G(x; s) = 0; d G(x; x):
u(x) = 12 dx (5.1.14)
Proof. Using the representations for  , we write the rst equation of
(5.1.6) as:
+ e ikx = Z 1 e iksG(x; s)ds + reikx + r Z 1 eiksG(x; s)ds:
a x x
The Gel'fand-Levitan-Marchenko equation is obtained by taking the in-
verse Fourier transform of this equation. Writing
1 Z1
f1 (s) = 2 r(k)eiksdk;
1
the inverse transforms of the terms on the right side above are
1 Z 1 eiks Z 1 e iktG(x; t)dtdk = G(x; s);
2 1Z x
1 1 eikseikxr(k)dk = f (x + s);
2 1
1
 
5.1. THE SCHRODINGER EQUATION 67
and
1 Z 1 eiksr(k) Z 1 eiktG(x; t)dtdk = Z 1 G(x; t) 1 Z 1 eik(s+t) r(k)dkdt
2 1 x Z1 x 2 1
= G(x; t)f1 (s + t)dt:
x
The inverse Fourier transform of the left side can be evaluated by closing
the contour in the upper half plane for s > x and using the residue theorem:

1 Z 1 eiks +(x; k) e ikx dk

2 1 a
1
Z 1  ( x; k ) eikx 
= 2 eik (s x ) + 1 dk
1 a
X N
 ik s
je j XN
cj j eikj s
=i a0(k ) =
j =1 ia (kj )
j 0
j =1
XN 
cj eik (x+s) + Z 1 eik (s+t) G(x; t)dt

= j j

j =1 ia (kj )
0 x
Z1
= f2 (x + s) f2(s + t)G(x; t)dt;
x
where
X
N
f2(s) = ia0c(jk ) eik s: j
j
j =1
Putting these together, we get (5.1.12), with f = f1 + f2:
To establish (5.1.14) we apply D2 + k2 + 61 u to (5.1.11) to get

0= xx + (k
2 + 1 u)
6
Z 1 
= Gxx(x; s) + (k2 + 1 u(x))G(x; s)
6 eiksds
x
d G(x; x) + ikG(x; x) 1 u(x)):
eikx(Gx(x; x) + dx 6
68 CHAPTER 5. INVERSE SCATTERING THEORY
Now
Z1 Z 1 @2eiks
k2eiksG(x; s)ds = @s2 G(Zx; s)ds
x x
1
=ik G(x; x)eikx Gs(x; x)eikx Gss(x; s)eiksds:
x
Combining these two calculations and multiplying by e ikx we get
Z1
(Gxx Gss uG)(x; s)eik(s x)ds (2 d G(x; x) 1 u(x)) = 0:
dx 6
x
This holds for = k  0: Letting k tend to innity in the upper half plane, we
see that the integral tends to zero. Since the other term does not depend on k
it must vanish identically. But then the integral must also vanish identically;
and, by the uniqueness of the Fourier transform,
..k the integrand must also
vanish identically, thus proving the result. `

The inverse problem

We now consider the inverse scattering problem, that of reconstruct-
ing u from the scattering data fr; kj ; cj g, by solving the Gel'fand-Levitan-
Marchenko equation. First note that while a is not given as part of the
scattering data, it appears, albeit inconspicuously, in the construction of f
in (5.1.13).
We rst the following representation for a given jrj and the zeroes of a,
kj .
X
N Z
1 1 log jaj2 dt; Im k > 0
log a(k) = log k kj +
k + kj 2i 1 t k (5.1.15)
j =1

The derivation of (5.1.15) is carried out as follows. We know jaj2 from r and
the relation jtj2 + jrj2 = 1, and we know the location of the zeroes of a in the
upper half plane since we are given the kj . If we dene ea by
Y
N
k kj
a(k) = ea(k)
j =1 k + kj
 
5.1. THE SCHRODINGER EQUATION 69
then ea is analytic in the upper half plane, tends to 1 as k tends to innity,
and jeaj = jaj on the real axis. Therefore log ea is analytic in the upper half
k plane and tends to zero at innity.
The function A dened by
(
A(k) = log ea(k); = k > 0;
log ea(k) = k < 0:
is sectionally holomorphic in = k 6= 0 and tends to zero as k tends to innity.
Its jump across the real axis is
A(k + i0) A(k i0) = [A] = log ea(k) + log ea(k) = log jeaj2 = log jaj2:
These properties uniquely determine A. We claim that
1 Z 1 log jaj2
A(k) = 2i dt; = k 6= 0
1 t k
In fact, the expression on the right is sectionally holomorphic and tends to
zero as k tends to innity. By the Plemelj formulae,
1
Z 1 log jaj2 1 P Z 1 log jaj2
lim 2
dt =  2 log jaj + 2i dt;
z!ki0 2i 1 t z 1 t k
where P denotes the Cauchy principal value. Hence the sectionally holo-
morphic function dened by the Cauchy integral has the same jump across
the real axis as A, and both functions tend to zero as k ! 1. Hence they
coincide.
The next step is to solve the GLM equation using the Fredholm theory
of integral equations. We x x and consider (5.1.12) as a Fredholm integral
equation on the half line (x; 1). It involves the integral operator dened by
Z1
(FxK )(s) = f (s + t)K (t)dt; s  x:
x
We rst observe that if f (s) decays suciently rapidly as s tends to +1,
then Fx is a Hilbert-Schmidt operator, i.e.
Z 1Z 1
jf (s + t)j2dsdt < +1:
x x
70 CHAPTER 5. INVERSE SCATTERING THEORY
In fact, the simple change of variables u = s + t; v = s shows that this
integral is equal to
Z1
(u 2x)jf (u)j2du (5.1.16)
2x
so Fx is a Hilbert-Schmidt operator if juj1=2f is in L2 (2x; 1).
Now look at (5.1.13). The discrete sum decays exponentially as s ! +1.
The integral term decays if r is dierentiable. (This is a standard fact about
Fourier transforms). For example, if the second derivative of r is in L1 then
f decays like s 2 . In fact, if u 2 S then so does r and hence so is f [6].
Therefore, for each x the GLM equation is a Fredholm integral equation.
The Fredholm alternative states that the integral operator I + Fx has a
bounded inverse if and only if the homogeneous equation (I + Fx)G = 0 has
only the trivial solution. Therefore, to prove the existence of a solution of
(5.1.12) it is enough to prove that ker (I + Fx) is trivial.
Suppose that
Z1
G(s) + f (s + t)G(t)dt = 0; s  x:
x
We may extend G to be zero for s < x: Then, since G is real,
Z1 Z 1Z 1
jG(s)j2ds + f (s + t)G(t)G(s)dsdt = 0:
x x x
But f = f1 + f2 , where
X
N
!j s ;
f2 = dj e dj > 0; kj = i!j
j =1
Therefore
Z 1Z 1
f2(s + t)G(t)G(s)dsdt
x x
X Z 1Z 1
N
!j t e !j s G(t)G(s)dsdt
= dj e
j =1 x x
X Z 1 ! t
N 2
= dj  e G(t) dt  0 j

j =1 x
 
5.1. THE SCHRODINGER EQUATION 71
Similarly,
Z 1Z 1
f1 (s + t)G(t)G(s)dsdt
x Zx Z
1 1 1 Z1
=
2 1
r(k)eik(s+t) G(t)G(s)dkdsdt
Zx 1 x
= r(k)(Ge(k))2 dk
1
where 1 Z
Ge(k) = 21 G(t)eikt dt:
x
This latter quantity is real, since both f and G are real; hence
Z1 Z1
0 = ((I + Fx )G; G)  jGej2 + r(k)Gf2 dk  jGej2(1 jr(k)j)dk;
1
1
where we have applied the Plancherel theorem to the Fourier transform for
G: Z1 Z1
jGj2 dk = jGej2 dk:
1 1
Since jtj2 + jrj2 = 1, jr(k)j = 1 implies t = 0: In that case a = t 1 is innite;
and this can happen only at k = 0. Therefore Ge and hence G must vanish
identically.
We have proved:
Theorem 5.1.6 The Gel'fand-Levitan-Marchenko equation is uniquely solv-
able whenever the integral in (5.1.16) is nite.
By the Fredholm alternative, the Gel'fand-Levitan-Marchenko equation
is solvable whenever r 2 L2
The GLM equation can be solved explicitly in the so-called re ectionless
potential case where r(k)  0; the choice 2
X
N
!j s ;
f (s) = dj e dj > 0; kj = i!j
j =1
2 Note that dj > 0 by Lemma 5.1.2 and (5.1.13).
72 CHAPTER 5. INVERSE SCATTERING THEORY
leads to the multisoliton solutions. We look for a solution of the GLM equa-
tion of the form
XN
G(x; s) = gj (x)e !j s:
j =1
This ansatz leads to an algebraic system of equations for the gj :
X
N (!j +!k )x
dj e !j x + Djk gk (x) = 0; Djk = jk + dj d!k e + ! :
k=1 j k
We may solve this equation in closed form as follows. Let E denote
the column vector with entries dj e !j x and let Ck denote the kth column
vector of jjDjkjj. By Cramer's rule, gk (x) = Dk (x)D 1 (x), where D(x) =
det jjC1; : : : ; Cnjj and Dk = det jjC1; : : : ; E; : : : ; Cnjj. Hence
X
N
Dj (x)e !j s
G(x; s) = D :
j =1
Now note that Ck0 = Ee !k x, so that
d D = det jjC 0 ; : : : ; C jj +


+ det jjC ; : : : ; C 0 jj
dx 1 n 1 n
X
N
!j x
= Dj (x)e = G(x; x)D;
j =1
or
d log D:
G(x; x) = dx (5.1.17)
From the second equation in (5.1.14),
d2  d d e (!j +!k)x 
u(x) = 12 dx2 log det jk + j !k + !  (5.1.18)
j k
To get the multi-soliton solution of the KdV equation, replace dj in (2.2.5)
by
dj e8i(i!j )3 t = dj e8!j3 t:
This formula for the multisoliton potentials can be extended to the general
solution of the GLM equation by the method of Fredholm determinants [54],
[60].
5.2. FIRST ORDER SYSTEMS 73
5.2 First Order Systems
We now turn to the study of the forward and inverse scattering theory asso-
ciated with the rst order the dierential operator
d
i 0   0 p
Dx = dx zJ q; J = 0 i q= r 0 (5.2.19)
The scattering problem for Dx was rst proposed and solved by Zakharov and
Shabat [76] in their study of the nonlinear Schrodinger equation. Ablowitz
et. al. [1] showed that this operator plays the role of an isospectral operator
for the hierarchy of integrable systems discussed in x (4.2), now known as
AKNS systems. Under the symmetry reduction r = p it is the isospectral
operator for the nonlinear Schrodinger equation; with r = p one gets the
hierarchy containing the sine-Gordon equation as well as the modied KdV
equation; and with r = p one gets the sinh-Gordon equation.
The inverse scattering problem for 2  2 systems can be formulated as
an integral equation of Gel'fand-Levitan-Marchenko type, just as for the
Schrodinger operator. This method does not extend to the case n > 2; rather,
higher order inverse scattering problems must be formulated as Riemann-
Hilbert problems [5], which are considerably more complicated technically.
We shall develop the scattering theory of (5.2.19) and obtain the Gel'fand-
Levitan-Marchenko integral equation. We begin by constructing the wave
functions of Dx.
Theorem 5.2.1 There exist fundamental matrix valued solutions of Dx =
0 of the form = mexzJ where:
1. m0 = z [J; m] + q;
2. m is meromorphic in Im z > 0;
3. For regular values of z , m are bounded for all 1 < x < 1;
4. m ! I as x ! 1 in Im z > 0; and m !  (z ) as x ! +1 in
= z > 0, where  are diagonal matrices.
5. m ! I as z ! 1 in Imz > 0. If q 2 C n then m each have an
asymptotic expansion
X n
m  mzjj ; m0 = I; [J; mj+1] = dm dx
j
qmj ;
j =0
74 CHAPTER 5. INVERSE SCATTERING THEORY
the mj are uniquely determined. If q 2 S then the asymptotic expansion
is valid to all orders.
6. m are uniquely determined by properties 1,3,4.
7. det m = 1.
Proof. Let 1  0
e+ = 0 e = 1 ):
The wave functions are obtained by converting the dierential equation to a
Volterra integral equation. For example, by writing  = `e ixz we obtain the
dierential equation
 
d` = zE` + q`; E = 2i 0 :
dx 0 0 (5.2.20)
This may be converted to a Volterra integral equation, for example
Zx
`=e + ezE(x y)q`dy: (5.2.21)
1
We leave it to the reader to verify that if q 2 L1(R ), then the successive
approximations to (5.2.21) converge uniformly in = z  0. The solution ` is
analytic in the upper half plane, is bounded for all real x, and tends to e
as x tends to 1.
If q 2 C n then ` has an asymptotic expansion
X
n
` `j z j (5.2.22)
j =0
where
`0 = e ; E`1 = qe ; E`j+1 = `0j q`j :
To validate (5.2.22), let
X
n
j
Rn = ` `j z
j =0
and substitute this into (5.2.20). We nd
dRn = (zE + q)R E`n+1 :
n
dx zn
5.2. FIRST ORDER SYSTEMS 75
Since Rn tends to zero as x ! 1, Rn satises the integral equation
Zx 1
Zx
Rn = e zE (x y ) qRndy zn ezE(x y) E`n+1dy
1 1
The solution of this integral equation obeys the estimate
jRn(x)j  C (x)jzj n
for some constant C which depends on x. This proves the validity of (5.2.22).
If q 2 C 1 then we obtain the innite asymptotic expansion
X1
`  z`jj :
j =0
The existence of wave functions  and , analytic in = z > 0 with
the asymptotic behavior
  eixz e; x ! 1;   e e ; x ! +1
ixz

is proved in this manner. Let  and denote the matrices

(x; k) = jj ; +jj; (x; k) = jj +; jj;
for k real. Each of these matrices is a fundamental solution set on the real
axis, so there exists a matrix S (k) such that
a b 
(x; k) = (x; k)S (k); S (k ) = c d ;
i.e.
+ = b + + d  = a + + c (5.2.23)
Since tr zJ + q = 0, det  and det are constant; evaluating them as
x ! 1 we nd they are identically one. Thus det S = 1 It is a simple
matter to show that
d = det jj +; +jj; a = det jj ; jj;
so d and a are analytic in the upper half and lower half z planes respectively.
Zeroes of a and d correspond to bound states; but in the case of two by two
systems we can no longer assert that they are simple [68].
76 CHAPTER 5. INVERSE SCATTERING THEORY
Now
eixz ( ) 
d = x!lim det jj +; +jj = x!lim det +1 = lim ` :
+1 +1 0 (+)2 x!+1 2
We claim
lim ` (x; z) = 0
x!+1 1
for = z > 0. Let " > 0 be given. Since supx j`(x; z)j < +1 we may pick x0
so large that Z1
jq`jdy < ":
x0
From (5.2.21)
Z x0 Zx
`1(x; z) = e2iz(x y) (q`)1dy + e2iz(x y) (q`)1dy:
1 x0
The rst integral tends to zero as x ! +1 if = z > 0 and the second is
bounded in absolute value by "; hence `1 tends to zero as x ! 1. We have
shown that 0
` = +e ! d as x ! 1:
ixz

The limits of the other reduced wave functions are evaluated in the same
manner. The reduced wave functions m are dened by
jj d+ ; +jj = m+ exzJ ; jj; a jj = m exzJ :
It is a straightforward exercise to check that properties (1)-(5) of the reduced
wave functions given in Theorem (5.2.1) are satised with this choice of m.
The poles of m+ ; m are located at the zeroes of d and a respectively.
Finally, we prove that items 1,3,4 uniquely determine the reduced wave
functions. Suppose m1 and m2 satisfy the dierential equation
dm = z[J; m] + qm:
dx
for z, say, in the upper half plane, are bounded in x, and tend to the identity
as x tends to 1. Then w = m1 1 m2 satises
dw = z[J; w]; w ! I; x ! 1:
dx
5.2. FIRST ORDER SYSTEMS 77
The general solution of the dierential equation is w = exzJ w0(z)e xzJ ; for
some matrix valued function w0(z) indepdendent of x. Fix z. Since m1
and m2 are bounded for all x 2 R , and det m1 = 1, so is m1 1 m2. This
implies that w0 is a diagonal matrix, hence commutes with exzJ . Therefore
w(x; z) = w0(z). Since w tends to the identity as x ! 1, we must have
w = I . `..k
0

We now turn to the derivation of the Gel'fand-Levitan-Marchenko equa-

tion for the scattering theory of rst order 2  2 systems.
As in the case of the Schrodinger equation, we get the representations
Z1
 (x; z ) = e eixz + G(x; y)eizy dy
x
in the upper and lower half z planes. The kernels G both have support in
the half lines y > x.
This representation is a consequence of the Paley-Wiener theorem for
analytic functions and the asymptotic properties of the wave functions, and
is independent of the fact that they satisfy dierential equations as functions
of x. For example, eixz e tends to zero as z tends to innity in the
lower half plane, so by the Paley-Wiener theorem,
Z0
eixz e = K (x; y)eizy dy:
1
Hence
Z0
(x; z) = e ixz e + eiz(y x)K (x; y)dy
Z 11
=e ixz e + e izy K (x; x y)dy:
x
The result for follows by taking G (x; y) = K (x; x y). The G are
column vectors with two entries.
Now write (5.2.23) as
+ = b + ;  = +c : (5.2.24)
d d + a + a
78 CHAPTER 5. INVERSE SCATTERING THEORY
Write the rst equation as
+ e e ixz = Z 1G (x; y)e izy dy
d x  Z1 
b
+ e+ e +ixz izy
G+(x; y)e dy :
d x
Taking the inverse Fourier transform, and proceeding exactly as in the
case of the Schrodinger equation, we obtain
Z1
0 = f+(x + s)e+ + G (x; s) + G+(x; y)f+(y + s)dy;
x
where Z 1 b(k) X n+ c+
1
f+(s) = 2
+
e d(k) dk + id0(jk+) eikj s:
iks
1 j =1 j
Here, kj+ are the (simple) zeroes of d in the upper half plane and c+j are the
coupling coecients:
+(x; kj+) = c+j (x; kj+):
Similarly, multiplying the second equation in (5.2.24) by (2) 1e iks and
integrating with respect to k over the real line, we obtain
Z1
0 = f (x + s)e+ G+(x; s) + G (x; y)f (y + s)dy;
x
where Z 1 c(k) X n cj ikj s
1
f (s) = 2 iks
e a(k) dk e :
j =1 ia (kj )
1 0

We may combine all these equations and write them in a simple matrix
form. Let
f (s) 0   0 1
+
G = jjG+; G jj; f (s) = 0 f (s ;  = 1 0 ):
)
Then
Z1
(x; k) = exkJ + G(x; y)eykJ dy; (5.2.25)
x
5.2. FIRST ORDER SYSTEMS 79
and the integral equations can be written as
Z1
0 = f (s + x) + G(x; s) + G(x; y)f (y + s)dy; s > x: (5.2.26)
x
Substituting (5.2.25) into the equation Dx (x; k) = 0 and proceeding as
in the proof of Theorem (5.1.5) we get
0 = (q(x) + G(x; x) JG(x; x)J 1 )
Z1
+ q(x)G(x; y) @G @x J @G J 1 e(y x)kJ dy:
@y
x
Since the rst term is independent of k, we may let k tend to innity. Then
the integral term goes to zero by the Riemann-Lebesgue lemma, assuming
the integrand is in L1 (x; 1). Hence the rst term must vanish, and
q(x)J = [J; G(x; x)]: (5.2.27)
Equation (5.2.27) is the analog of (5.1.14) in the theory of the Schrodinger
equation. From this equation we nd
p(x) = 2B (x; x); r(x) = 2C (x; x): (5.2.28)
These formulae allow us to recover the components of the potential from the
kernel of the GLM equation.
The integral term must also vanish, and it may be shown that the kernel
G satises a hyperbolic system of partial dierential equations.
Writing A B 
G= C D ;
we may break (5.2.26) down into four equations:
Z1
f+(x + s) + B (x; s) + A(x; y)f+(y + s)dy =0;
Z1
x
A(x; s) + B (x; y)f (y + s)dy =0;
Zx1
f (x + s) + C (x; s) + D(x; y)f (y + s)dy =0;
Z1
x
D(x; s) + C (x; y)f+(y + s)dy =0:
x
80 CHAPTER 5. INVERSE SCATTERING THEORY
These give
Z1 Z1
f+(x + s) + B (x; s) B (x; y) f (y + z)f+(z + s)dzdy = 0
x x
(5.2.29)
and
Z1
A(x; s) f+(x + y)f+(y + s)dy (5.2.30)
x Z1 Z1
A(x; y) f+(y + z)f (z + s)dzdy = 0: (5.2.31)
x x
Under the assumption that
Z1
(u x)jf(u)j2du < +1
x
both the integral operators dened on L2(x; 1) by
Z1
(F 
x )(s) = f(s + y) (y)dy
x
are Hilbert Schmidt operators; therefore so are the compositions Fx+Fx and
Fx Fx+: Hence the integral equations (5.2.26) are Fredholm integral equations.
Their solvability is therefore a consequence of uniqueness; but uniqueness can
only be shown to hold in cases where there is a symmetry of the potential
matrix q, hence of the scattering data.
Lemma 5.2.2 Let q = q and let
 0 1
= 1 0 :
Then
 +(x; z) = (x; z); +(x; z) =  (x; z);
S is unitary: SS  = I :
 a b
S= b a :
5.2. FIRST ORDER SYSTEMS 81
G(x; s) 1 = G(x; s); hence
 A B
G = B A ):

kj+ = kj ; c+j = cj ;
f+ (s) = f (s).
Lemma 5.2.3 Let q be real and skew symmetric. Then in addition a(k) =
a( k); b(k) = b( k) for real k;
 A B
G(x; s) = B A );
where A and B are real; and f (s) are real.
Proof. We know that + (x; z )  e+ eixz as x ! +1 or as z ! 1 in the
upper half plane. Hence, for = z < 0,
 + (x; z)  e+ e ixz = e e ixz :
On the other hand, a short calculation shows that Dx(z; q) 1 = Dx(z; q),
so  +(x; z) is also a solution of the dierential equation. Since the the wave
functions are uniquely determined by their asymptotic behavior, we obtain
the rst statement of Lemma (5.2.2). The others are derived in exactly the
same way.
From the rst statement we nd that
 (x; k) 1 = (x; k); (x; k) 1 = (x; k);
hence S (k) 1 = S (k): Since det S = 1 this implies that S 2 SU (2), hence
a = d and c = b.
We leave the proof of the remaining statements of Lemma (5.2.2) to the
reader. In the case of Lemma (5.2.3) we have the additional symmetry that
q is real. Thus (x; k) 1 = (x; k), etc., and
S (k) 1 = S ( k); a(k) = a( k); b(k) = b( k):
Moreover, we now nd that the entries A and B in G are now real. The
symmetry of a can now be written a(k) = a( k); and from this we see that
82 CHAPTER 5. INVERSE SCATTERING THEORY
if kj+ is a zero of a then so is kj+ . Call this k+j . From the symmetry in
Lemma (5.2.2) we also know that if k is a zero of a then k is a zero of a.
Thus the bound states appear in foursomes: k; k, or in pairs if k is purely
imaginary. From the symmetries of the wave functions we nd that c+j = cj .
From this one may show that f+(s) = f (s) is real. ` ..k

Theorem 5.2.4 In the case of the symmetries discussed above we have [53]
d2 log det jjI + F F jj;
ju(x)j2 =2 q = q;
x x
dx2
d arg det jjI + iF jj;
u(x) = 2 dx qy = q; q real:
x

where Fx is the Hilbert-Schmidt operator given in (5.2.26)

Z1
FxG = G(x; y)f (y + s) dy:
x

5.3 Decay of the scattering data

In this section we prove that, in analogy with the Fourier transform, the scat-
tering transform for the Schrodinger operator preserves the Schwartz class
in the sense that the mapping u ! r takes Schwartz functions to Schwartz
functions [6].
Lemma 5.3.1 For u 2 S the re ection coecient r also belongs to S .
Proof. We begin by rewriting (5.1.6) in the form
+ = + r  = + r
a + a +

where r is the re ection coecient. If we group + and +=a on the left and
and = a on the right side, we can write these equations in the form

+ 1 r 
  1 0
+ a 0 1 = a r 1
5.3. DECAY OF THE SCATTERING DATA 83
Since a and r are independent of k we may also dierentiate these equa-
tions with respect to x. Combining them with the relations above, and noting
that  1 01 r 1  1 r   1 r 
r 1 0 1 = r 1 jrj2 = r jtj2
we obtain
0 1

+=a =  =a
@ 1 r
A: (5.3.32)
+
r jtj2
These conditions form multiplicative jump conditions across the real axis and
consitute in part a Riemann Hilbert problem for the wave functions.
The reduced wave functions m are dened by
0 1 0 1
+=a
@ +
A = m+eixz3 ; @ =a A = m eixz3 :
0 0+=a 0 =a 0
+
where 1 
3 = 0 01 :
The sectionally meromorphic function dened by m = m in = z > 0
satises the jump conditions
0 1
1 r
m+ = m eixk3 @ Ae ixk3
2r jtj
across the real axis. At x = 0 we get
0 1
1 r
@ A = m (0; k) 1m+(0; k): (5.3.33)
2
r jtj
The column vectors of m are meromorphic in the upper and lower half
z planes respectively and may be obtained as solutions of Volterra integral
equations by the method of successive approximations. On the real axis, they
are C 1 functions of k, provided the potential u 2 S . This may be seen by
84 CHAPTER 5. INVERSE SCATTERING THEORY
dierentiating the Volterra integral equation with respect to k and solving
the respective equations for the derivatives.
Furthermore, the asymptotic expansions of m in k are valid down to the
real axis, and may be dierentiated termwise with respect to k. Moreover,
the asymptotic expansions for m are uniquely determined by u; and in
particular, m+ and m have the same asymptotic expansions on the real k
axis. Therefore
m (0; k) 1m+(0; k)  I
and 0 1
1 r
@ A I 2 S:
r jtj2
This shows not only that jrj ! 0 as jkj ! 1, but also that jaj2 ! 1 as
k ! 1 on the real axis. `..k
Chapter 6
Variational Methods
6.1 A variational principle for water waves
Variational methods for nonlinear waves occur in a number of contexts, and
I will describe some of these methods in this lecture. There is also an exten-
sive theory for free boundary value problems in potential uid ow [13]{that
is, the ow of an irrotational, inviscid uid. These problems may be formu-
lated as variational problems [32] for the stream function and have received
considerable attention in recent years [3, 31].
I will begin by discussing the variational principle for the time dependent
Euler equations of water wave theory. This principle was rst formulated
in 1967 by J. Luke [49]; in 1968 a Hamiltonian formulation of the Euler
equations was given by V. E. Zakharov [77]. Benjamin and Olver [10, 56] have
determined all local conservation laws of the Euler equations. This principle
applies to the potential function and so can be used in three dimensions,
whereas variational formulations for the stream function are restricted to
two dimensional ows, or axisymmetric three dimensional ows.
The possibility of proving proving the existence of solutions of the Euler
equations by the direct method of the calculus of variations should be con-
sidered; to my knowledge this has yet been done. The variational principle
might also allow for a weak formulation of the free boundary value problem,
which might be suitable in discussing weak solutions of the Euler equations.
This would be particularly convenient in the discussion of breaking waves,
etc.
The variational principle proposed by Luke is the following. We introduce
85
86 CHAPTER 6. VARIATIONAL METHODS
the local Lagrangian
Z
L(x; t) = 't + 12 (r')2 + gy dy
1
Z 2
= 't + 21 (r')2 dy + g2 + const:
1
We drop the constant and form the action
Z t2 Z1
S= L dx dt; t1 < t2 :
t1
1
The Euler equations of x3.3 are the Euler-Lagrange equations for this
integral; that is, that they are a consequence of the variational principle
S_ = 0 where
S_ = S
 [ 
_ ] + S ['_ ]:
'
We have
S [_ ] = Z t2 Z L [_ ] dxdt;
1

 t1 
1
where 
L = @ L = ' + 1 (r')2
 
 @ t
2 y= + g:
Thus
Z t2 Z1   
't + 21 (r')2 + g y= _ dxdt = 0
t1
1
for all variations _ implies the equation
't + 12 (r')2 + g = 0 on y = :
Now let us calculate the variation of S with respect to variations in '.
S ['_ ] = Z t2 ZZ '_ + r'
r'_ dxdydt;
We have
t
' t1 D
6.1. WATER WAVES 87
where D is the region occupied by the uid.
Since div('_ r') = r'_
r' + '_
', we have, by Green's theorem,
ZZ ZZ I
r'
r'_ dxdy = '_
' dxdy + '_ ('xdy 'y dx)
@D
D D

ZZ Z1 Z1 
= '_
' dxdy '_ y dx '_ ('xx 'y )y= dx:
D 1 1
On the other hand,
2 3
Z t2 ZZ Z t2 6 @ ZZ Z1  7
'_ tdxdydt = 4 @t '_ dxdy t'_ y= dx5 dt
t1 t 1
D D 1

ZZ t2 Z1 
= '_ dxdyt t'_ y= dxdt:
1
D 1
Hence
S ['_ ] = Z t2 ZZ Z t2 Z1 

''_ dxdydt (t + x'x 'y ) '_ dxdt
' t1 t1 y=
D 1
ZZ t2 Z t2 Z1 
+ '_ dxdyt 'y y= h'_ dxdt:
1 t 1
D 1
We now assume that the variation S=' vanishes for all variations '_ and
prove that the remainder of the Euler equations follow. First, we restrict the
variations to those for which '_ (x; y; t1) = '_ (x; y; t2) = 0. Now assume the
variation '_ vanishes on @D . Then we see that
' = 0 in D . Next assume
the variation '_ vanishes on the free surface. We then obtain the boundary
condition 'y  = 0: Finally, let '_  be an arbitrary continuous function;
y= h y=
then we nd that 
(t + x'x 'y ) = 0:
y=
88 CHAPTER 6. VARIATIONAL METHODS
A Hamiltonian formulation of the Euler equations was given by Zakharov
[77] in 1968. The Hamiltonian is
Z 1 Z 
1
H=2 g2 + (r')2 dy dx
h
1

6.2 Modulation and the method of averaging

The method of averaging was introduced into the subject of oscillation theory
in the 1930's by Bogoliubov and Mitropolsky [14]. It was extended to the
theory of nonlinear waves by G.B. Whitham [74] (see also the book by Jerey
and Kawahara [39]. Consider the nonlinear Klein-Gordon equation
utt uxx + u + u3 = 0; 1 < x < 1: (6.2.1)
This equation admits, as a special class of solutions, a periodic wave train,
of the form
u(x; t) = '(a; ); a > 0;  = kx !t
where a; k; ! are constants, and ' is 2 periodic in . The real parameter a
parametrizes the amplitude of the oscillations. Such a small amplitude peri-
odic wave train is analogous to the periodic solutions of the Euler equations
constructed by Stokes [71].
If we substitute this form of the solution into the Klein Gordon equation
we get an ordinary dierential equation for ':
(!2 k2)'00 + ' + '3 = 0:
We look for 2 periodic solutions of this equation. We may restrict the
solutions to be even in . Letting  = !2 k2 , ' = a , we obtain the
following equation for :
 00 + + a2 3 = 0:
Expanding both  and in powers of a2 ,
 = 0 + 1a2 + : : : = 0() + 1()a2 + : : :
6.2. METHOD OF AVERAGING 89
we obtain
0 000 + 0 = 0;
0 100 + 1 + 1 000 + 03 = 0;
...

To get 2 periodic solutions, take 0 = 1 and 0 = cos . At the next

stage we then have
00
1+ 1 1 cos  + cos3  = 0:
The operator D2 + 1 has a one dimensional kernel when restricted to the
subspace of even functions of . The solvability condition for this equation
for 1 is then given by the single condition
Z 2
( 1 cos  + cos3 ) cos  d = 0:
0
This determines 1; and we nd
1 Z 2
1 =  cos4  d = 34 :
0
We can then solve uniquely for 1 . In this way we get a formal perturbation
series in a2; its convergence can be proved by an application of the implicit
function theorem. (cf. for example, [66]).
We thus obtain
 = 1 + 34 a2 + : : : ' = a cos  + a2 : : : : (6.2.2)
We now want to consider slow amplitude modulations of this periodic
wave train, just as radio signals consist of modulation of carrier waves in the
radio frequency range by signals in the audio frequence range. In particular,
we are interested in constructing a modulated pulse roughly of the type shown
in the gure below. This problem is of particular interest in the propagation
of modulated electromagnetic pulses along wave guides, for example radio
waves along wave guides or optical pulses along optical bers.
90 CHAPTER 6. VARIATIONAL METHODS
To this end, we introduce slow variables
 = "x;  = "t;  = "
We have k = x =  and ! = t =  .
The Klein Gordon equation has a variational formulation, with the La-
grangian
L(u; ux; ut) = 21 (u2t u2x u2) 4 u4:
The Klein Gordon equation is given by
L = @ @L @ @L + @L
u @t @ut @x @ux @u
= utt + uxx u u3 = 0:
We now construct, following [74] an averaged Lagrangian for the modu-
lations, as follows. Given the periodic wave train '(; a) we dene
Z 2
L(k; !; a) = 1
2 0 L('; 'x; 't)d; ' = '(; a): (6.2.3)
We have
't =' t + 'a at = !' + "'aa
'x =' x + 'aax = ' k + "'aa :
The specic relationship between  and  is now formally dropped, and
 is regarded as an independent function of  and  ; but the relationship
 =  and ! =  is retained. The equations for the averaged Lagrangian
are then
L = @ @ L @ @ L + @ L = 0 (6.2.4)
a @ @a @ @a @a
L = @ @ L + @ @ L = 0 (6.2.5)
 @ @! @ @k
@k + @! = 0: (6.2.6)
@ @
6.2. METHOD OF AVERAGING 91
Equation (6.2.6) is simply the consistency condition on k and ! so that there
exists a function  = (;  ) for which k =  and ! =  .
The wave train '(; a) may be computed explicitly in terms of elliptic
functions; but in the rst approximation to the nonlinear theory we simply
use the lowest order approximation coming from the power series expansion
(6.2.2). Thus L (!; k; a) is computed from (6.2.3) by taking ' = a cos ,
k = x and ! = t . A straightforward calculation shows that
L('t ; 'x; ') = 21 ([ !' + "'aa ]2 [k' + "'aa ]2
a2 cos2 ) 4 a4 cos4 
Carrying out the integration over 0    2, we obtain
L (!; k; a) = 41 (!2 k2 1)a2 332 a4 + "2 (a2 a2 ):
Equations (6.2.4), (6.2.5), (6.2.6) become
(!2 k2 1)a 43 a3 + "2(a a ) = 0
@ (!a2) + @ (ka2) = 0
@ @
@k + @! = 0
@ @
We look again for traveling wave solutions of the form a( c ) with c xed.
We then obtain the ordinary dierential equations
"2(1 c2)a00 + (!2 k2 1)a 43 a3 = 0
(ka2 )0 c(!a2)0 = 0
!0 ck0 = 0
We may integrate the last two equations to obtain
! ck = S; ka2 c!a2 = R; (6.2.7)
92 CHAPTER 6. VARIATIONAL METHODS
1

0.8

0.6

0.4

0.2

−0.2

−0.4

−0.6

−0.8

−1
−10 −8 −6 −4 −2 0 2 4 6 8 10

Figure 6.1: A modulated pulse.

where R and S are constants of integration. We can then solve for ! in
terms of R; S; and c. In general we obtain again periodic wave trains for
the modulation a; these manifest themselves as periodic modulations of the
carrier wave, or doubly periodic solution. But there is also the special case
where a tends to 0 as  c ! 1. In that case R = 0 and we may solve
(6.2.7) explicitly for k and !. We get
k = 1 cSc2 ; ! = 1 S c2 :
Setting
2
 = ! 2 k 2 1 = 1 S c 2 1;
and noting that a a = (1 c2 )a00, we obtain the ordinary dierential
equation for the amplitude a:
a 43 a3 + "2(1 c2 )a00 = 0:
6.2. METHOD OF AVERAGING 93
This equation has a rst integral:
"2(1 c2 ) a02 +  a2 3 a4 = E:
2 2 16
In p
this case the level curves in the phase plane consists of two centers, at
a =  4=3, surrounded by families of periodic orbits. These are bounded
by two homoclinic orbits, corresponding to traveling wave pulses.
This example is essentially worked out in [74], with some minor dier-
ences. Whitham's treatment is entirely formal; in particular, his argument
does not prove the existence of the modulated pulse.
The treatment also considers only the rst approximation to the linear
theory. It would be worthwhile to give a rigorous treatment of the full non-
linear theory, and to prove the existence of such a modulated pulse. It would
also be interesting if one could prove the existence using the direct method
of the calculus of variations.
94 CHAPTER 6. VARIATIONAL METHODS
Chapter 7
Weak and Strong Nonlinearities
7.1 Breaking and Peaking
The problem of modelling nonlinear wave phenomena, despite 150 years of
progress, is still far from resolved. For the Korteweg-deVries equation, though
nonlinear, has regular solutions for all time, given regular initial data. That
is, singularities do not form under the evolution governed by the KdV equa-
tion. This fact must be reconciled with G. Stokes' remarkable prediction
of the wave of greatest height. Stokes' conjecture has been only recently
been proved rigorously for the exact equations by Toland, [73], and Amick,
Fraenkel, and Toland [4].
Rankine had conjecture in 1865 that there was a `wave of extreme height'.
In a moving reference frame (the Euler equations are Galilean invariant) the
Bernouilli equation on the free surface takes the form (I have set  = 1)
1 jr'j2 + gy = E:
2
This expresses the conservation of local energy. The rst term is the kinetic
energy of the uid, since ' is the velocity potential; and the second term is the
potential energy due to gravity. For the wave of extreme height, E = gymax,
where ymax , is the maximum height of the uid. Hence at the maximum
height of the uid, the velocity is zero, and there is a stagnation point in the
ow. Rankine conjectured that at the peak the free surface formed a cusp,
that is, that the tangent lines to the free surface were vertical. Stokes [71],
however, argued that, should such a wave of extreme height exist, the angle
95
96 CHAPTER 7. WEAK AND STRONG NONLINEARITIES
subtended at the peak was 2=3. His argument was motivated by physical
reasoning, and was quite simple.
The existence of a wave of extreme height was proved by Toland [73].
Toland also showed that if the singularity at the peak is not a cusp, that is,
if the tangent line to the free surface at the peak is not vertical, then Stokes'
conjecture about the size of the angle subtended is true. In a later article,
Amick, Fraenkel and Toland [4] proved, by a rather dicult analysis, that in
fact the singularity at the peak is not a cusp.

2
3

Figure 7.1: The extreme wave as conjectured by Stokes.

Thus, the full Euler equations exhibit singularities, and, moreover, there is a
limiting amplitude to the periodic wave trains, or to the solitary wave. This
shows that the KdV approximation can only be valid for suciently small
amplitudes. The KdV equation, on the other hand, has smooth cnoidal and
solitary waves of arbitrary amplitude.
G.B. Whitham [74], argued
It was remarked earlier that the nonlinear shallow water equations which
neglect dispersion altogether lead to breaking of the typical hyperbolic kind,
with the development of a vertical slope and a multivalued prole. It seems
clear that the third derivative term in the Korteweg-deVries equation will
prevent this ever happening in its solutions. But in both cases, the long
wave assumption under which the equations were derived is no longer valid.
Since some waves appear to break in this way, if the depth is small enough,
one concludes that some dispersion is necessary but the Korteweg-deVries
term is too strong for short wavelengths. This is not surprising in view of the
fact that the k3 term [is] a bad approximation to the full dispersion relation
when (kh0 )2 becomes large...
On the other hand the dispersion included in the Korteweg-deVries equa-
tion does allow the solitary and periodic waves which are not found in the
7.1. BREAKING AND PEAKING 97
shallow water theory. For these solutions, however, the Koretweg-deVries
equation cannot describe the observed symmetrical "peaking" of the crests
with a nite angle there.

Initial Profile t=.2 t=.4

0.15 0.15 0.15

0.1 0.1 0.1

0.05 0.05 0.05

0 0 0

−0.05 −0.05 −0.05

0 2 4 6 0 2 4 6 0 2 4 6

t=.5 t=.6 t=1

0.15 0.15 0.15

0.1 0.1 0.1

0.05 0.05 0.05

0 0 0

−0.05 −0.05 −0.05

0 2 4 6 0 2 4 6 0 2 4 6

Figure 7.2: Breaking wave in Whitham's model

Whitham goes on to discuss a class of models that might exhibit the

observed phenomena of solitary and periodic waves, as well as as peaking
98 CHAPTER 7. WEAK AND STRONG NONLINEARITIES
and breaking. In particular, he cites the class of models
Z1
t + x +
(x y)y (y; t) dy = 0;
1
in which the kernel or the convolution operator is represented as a Fourier
transform Z1
1

(x) = 2 eikx!(k) dk:
1
Whitham suggested in particular that one should take for ! the dispersion
relation that comes directly from the full Euler equations, namely
r
!(k) = tanh k: (7.1.1)
k
Such a dispersion relation is, of course, non-local, and the Euler equations
themselves are nonlocal equations.
Since the Matlab code that Yi Li and I constructed is based on the fast
Fourier transform, it is trivial to change it to handle the equation proposed
by Whitham. It requires only one change in the code, namely, to replace the
symbol k3 for the operator D3 with (7.1.1). The results of this experiment
are shown in Figure 7.1.
Thus, weakening the dispersion can indeed lead to breaking of waves.
Though Whitham's model is only a formal approximation, and does not
purport to be a valid model for the full Euler equations, it is nevertheless
instructive.
There are a number of other possibilities for the dispersion law !(k). T.
B. Benjamin had earlier proposed [7]
!(k) = jkj = k sgn(k);
based on heuristic arguments. The Fourier transform of sgn(k) is
1
2x
and the corresponding convolution operator is the Hilbert transform. The
equation thus obtained is
ut + uux + Huxx = 0;
7.2. STRONGLY NONLINEAR MODELS 99
where
Z1 f (y)
Hf (x) = 21 P x y dy:
1
A derivation of this equation by singular perturbation theory, in the spirit
of the derivation of the Korteweg-deVries equation itself in x3.2 was subse-
quently given by Hiroaki Ono [58] in 1975, and the equation is today called
the Benjamin-Ono equation. The Benjamin-Ono equation is itself a com-
pletely integrable system, having a formal Lax pair, and still belongs to the
class of 'weakly nonlinear' models.
Other 'weakly nonlinear' models have been proposed for the analysis of
waves on the surface of uids. We mention, in particular, the model proposed
by T.B. Benjamin, J. Bona, and J.J. Mahoney [11], cited earlier, as well as a
number of other models of higher order, which have been proposed by Peter
Olver [55].

7.2 Strongly nonlinear models

This modelling approach, though ad hoc, is typical of the approaches to
develop models exhibiting a threshold at which nonlinear eects begin to
overwhelm the smoothing eects of dispersion. The model falls in the realm
of \weak nonlinearity", in that the nonlinear terms appear in lower order
derivatives, not the highest order term. A dierent mechanism has been
proposed by P. Rosenau [57]:
It is quite obvious that in order to model phenomena related to wave
breaking, formation of cusps, and similar wonders of nature associated with
nonlinear dispersion, one has to probe deeper into the nonlinear regime, far
beyond the currently attained weakly nonlinear stage. . . . In a genuinely
nonlinear regime, nonlinearity plays a dominant role, rather than being a
higher order correction.
For the vibrations of a lattice of equally spaced (h  1) mass points m,
the potential part of the Hamiltonian is assumed to be
X
N y y 
Hp = PN n+1 n h
n=1
100 CHAPTER 7. WEAK AND STRONG NONLINEARITIES
where PN (y) = N yN =N . For the purely quartic potential, Rosenau obtains
[64], in the continuum limit, the nonlinear equation
utt = (u3)xx + [u(u2)xx]xx; u = yx:
This equation diers fundamentally from the weakly nonlinear theories, in
that it is nonlinear in the highest order derivative:
2u2uxxxx
The equation exhibits, among other features, solitary waves with compact
support.
Forays into the regime of strong nonlinearities are relatively recent, but
it seems that such systems are inescapable.
Given the diculties in analyzing nonlinear models, integrable systems
have the advantage that exact solutions can be obtained in closed form,
and thus provide insight into the behavior of nonlinear systems. This has
certainly been the case for the Korteweg-deVries equation itself.
A strongly nonlinear, completely integrable model, has been derived by
R. Camassa and D. Holm 1 [21], [22]. We choose somewhat dierent normal-
ization constants here (to be discussed in the next section):
1
3 3 1 
ut 4 uxxt 2 u2 8 u2x 4 uuxx = 0:
x
There is no general existence theorem for these equations, a number of
special classes have been constructed. One of them is the so-called `peak-ons',
shown in Figure 7.3.
In recent years strongly nonlinear dispersive models have become a focal
point of interest. Two such models, the shallow water model, and the Harry
Dym equation2,
mt = (m 1=2 )xxx;
arise as extensions of the class of isospectral ows of the Schrodinger operator.
They lead to essentially new scattering problems, which have not yet been
fully resolved.
1 The equation had previously been derived by purely algebraic methods by Fuchsteiner
[34] using the method of recursion operators; but its analytic properties only became
apparent with the investigations of Camassa and Holm and their co-workers, who derived
the equation in the context of models of water waves.
2 The equation was given by Kruskal [42] and attributed to an unpublished paper by
Harry Dym; the latter, I am told, denies having any connection with the equation.
7.2. STRONGLY NONLINEAR MODELS 101

Figure 7.3: Singular solutions of the Camassa-Holm equation. They look a

lot like Stokes' waves of extreme height, but they only appear in the rest
frame; the Camassa-Holm equation is not Galilean invariant.

We consider a class of hierarchies3 containing the KdV hierarchy which

are generated by the isospectral problem
D2 + k2 m(x) q(x):
These include the KdV ows, the negative KdV ows, the Camassa-Holm
equation, and the Harry Dym equation.
Just as in the derivation of the Korteweg-deVries equation, we look for
compatibility conditions for the simultaneous equations
xx + (k
2m q ) = 0; t =a x+b (7.2.2)
It is useful to keep track of how quantities transform under scaling. We
assign weights to each quantity so that the operators are homogeneous. We
begin by setting deg x = 1; that is we assume that x scales as x ! x. Then
deg D2 = 2, since
d =1 d:
d (x)  dx
We assign weight -2 to q, so that D2 + q is homogeneous. Finally, we let deg
k = 1 as in the KdV theory, and deg m = 0. Then the isospectral operator
D2 + k2 m q is homogeneous of degree -2.
In order that the operator
@ aD b
@t
be homogeneous, we require
deg a = 1 + deg b = deg t:
3The presentation in this and the next section was developed jointly with Jacek
Szmigielski of the University of Saskatchewan.
102 CHAPTER 7. WEAK AND STRONG NONLINEARITIES
We now determine a and b. First, we write these equations as a rst order
system:    
Dx 0 = 0; Dt 0 = 0;
where
@ J Q;
Dx = @x @ A;
Dt = @t
k
with
     
Jk = k02 m 10 ; Q = 0q 00 ; A = a(q bk2 m) a a+ b :
x
Note that tr Jk = tr Q = 0. We impose the requirement that tr A = 0 as
well.4 This leads to
b = 21 ax; (7.2.3)
and automatically, with this choice of b, we see that deg a = 1 + deg b.
We cross-dierentiate the pair of equations (7.2.2) to obtain
xxt = (k2 mt ) (k2 m 1) t (7.2.4)
txx =[a(1 k2 m) + bx ]x + [a(1 k2 m) + 2bx + axx] x

+ (ax + b)(1 k2m) (7.2.5)

With the choice (7.2.3) the compatibility condition is obtained in the
general form
qt k2 mt = 2ax(q k2 m) + a(qx k2 mx) 21 axxx = (R3 k2R1 )a;
(7.2.6)
where
R1 = mD + Dm; R3 = 12 D3 + qD + Dq; (7.2.7)
R1 and R3 are operators of weight -1 and -3 respectively.
4 This may seem to be an ad hoc requirement, but it is natural to require all the
potentials to lie in the Lie algebra s`(2; C ), just as in the AKNS hierarchy.
7.2. STRONGLY NONLINEAR MODELS 103
We rst consider the case in which a is a polynomial in k2:
X
n
a= aj k2j :
j =0

Substituting into (7.2.6) we obtain

R1 an =0 (7.2.8)
R1 al 1 =R3al n  l  2; (7.2.9)
mt =R1a0 R3a1 (7.2.10)
qt =R3a0 : (7.2.11)
The equation (7.2.8) is easily solved; namely
an = p1m :
Let us take some special cases.
1. a0 = 0. Then qt = 0 from (7.2.11); hence we may take q to be, say, a
function in S (R ). We obtain from (7.2.10)
mt = R3 a1 :
For the case n = 1 and q = 1 we obtain simply a1 = 1= m, and
p

mt = ( 12 D3 2D) p1m : (7.2.12)

This is the Harry Dym equation with an additional transport term; we get
the Harry Dym equation itself if we take q = 0.
2. The general ow for n = 2: After some computation we nd
 d2 log pm :

a2 = p1m ; a1 = 2m13=2 q + 12 dx 2

The ow is then obtained from (7.2.11), (7.2.10). Note that a0 is free.

104 CHAPTER 7. WEAK AND STRONG NONLINEARITIES
3. m  1. Then (7.2.9) simplies to
R3al = Dal 1;
and the parameter a0 is now determined by (7.2.10) since mt = 0. This is
the Lenard recursion relation for the KdV hierarchy which we introduced in
Chapter (5)
4. We note that a0 is undetermined by the above equations. Take a0 = q;
then from (7.2.11),
qt = 3qqx 12 qxxx
and q satises the KdV equation, while the ow for m is determined from
(7.2.10).
We next obtain the ows of negative order; equivalently, when
X
n
a= a j k 2j :
j =0
These include the Camassa-Holm and negative KdV ows [27]. In this case
we obtain the equations
R3 an = 0; (7.2.13)
R3aj 1 = R1 aj ; (7.2.14)
mt = R1 a0 ; (7.2.15)
qt = R3 a0 R1 a1: (7.2.16)
To obtain an in terms of q, we multiply equation (7.2.13) equation by an
to get 
d [qa2 ] = 1 d a a

1 a2 :
dx n 2 dx m n;xx 2 n;x
hence
a2nq = 12 anan;xx 41 a2n;x:
This relation simplies to
q = vx + v2; d log pa :
v = dx n
7.2. STRONGLY NONLINEAR MODELS 105
Thus, vn is nonlinear and nonlocal in q; in fact, this is precisely the well-
known Miura transformation the converts the KdV equation to the modied
KdV equation.
The negative KdV ows are obtained by setting m = 1. Then R1 = 2D,
mt = 0; and from (7.2.15) we get Da0 = 0, hence a0 is a constant. In order to
preserve the scaling invariance of the equations, we leave a0 in the equations.
We always set deg an = 0. All equations are homogeneous; that is, each term
in an equation has the same degree.
In the case n = 1, deg a0 = deg a1 + 2 = 2 and deg t = 1. From (7.2.16)
the ow is
qt = R3 a0 R1 a1 = a0 qx R1 a1: (7.2.17)
We introduce a new variable by setting q = Dz. Then Dzt = a0 D2z 2Da1;
and we integrate once to obtain
zt a0 zx = 2a1 :
Moreover,
R3 = 21 D3 + (2zxD + zxx:
Since R3 a1 = 0, we obtain
R3 (zt a0zx ) = 0;
which may be written [27]
zxxxt a0 zxxxx = 2[(zt2 )x 23 a0 (zx2 )x + zxxzt]: (7.2.18)
The Camassa-Holm equation is obtained at n = 2 with qt = 0. Taking
a(x; t; k) = u(x; t) + k12 ;
and noting that
R3 a = R3 u = 12 uxxx + 2ux; R1a = R1u + mxk 2 = 2mux + mx u;
we nd that (7.2.6) reduces to
mt = R1 u; R3 u = R1(1):
106 CHAPTER 7. WEAK AND STRONG NONLINEARITIES
From the second equation we obtain, upon integration,
m = 2(I 14 D2)u;
and from the equation at k2 ,
 1 
2ut 2 uxxt 2ux 2u 2 uxx u(2ux 21 uxxx) = 0:
1

This simplies to
(1 14 D2)ut 23 (u2)x + 38 (ux)2 + 41 (uuxx)x = 0:
Finally, we note that, as in the case of the KdV theory, the squared
eigenfunctions satisfy the third order dierential equation
( R3 k 2 R1 )  = 0;
where  = 1(x; k) 2 (x; k), and j satisfy
( D 2 + k 2 m q ) = 0:

7.3 The extended AKNS hierarchy

We may also consider an extension of the AKNS hierarchy based on the
isospectral operator
Dx = @x@ zm q
where m and q are d  d matrices. The scaling in this case is deg m =
0; deg q = deg z = 1.
We rst consider ows of positive order5 :
@ X
n
Dt = @t Fj z j :
j =0
5 Note that our labeling was somewhat dierent in Chapter; there we took

Dt
@
= @t
Xn
Fn jz
j
:
j =0
7.3. THE EXTENDED AKNS HIERARCHY 107
The zero curvature conditions [Dx; Dt ] = 0 give
R0 Fn =0 (7.3.19)
R0 Fj 1 =R1 Fj ; nj1 (7.3.20)
mt =R1 F1 R0 F0 (7.3.21)
qt =R1 F0 (7.3.22)
where
R0 = ad m; R1 = @x @ ad q:
As in the extended KdV hierarchy, F0 is in general undetermined.
The AKNS hierarchy itself is obtained by taking m to be a constant
diagonal matrix J of trace 0. Then (7.3.19) implies that [J; Fn] = 0; and we
may take Fn = , where  is any diagonal matrix of trace 0. The recursion
relation (7.3.20) then reads
[J; Fj 1] = @Fj [q; Fj ]
@x
Since mt = 0, F0 is determined by (7.3.21). The nonlinear ow is then
qt = [J; F0 ]:
The F0 are dierential polynomials in the entries of q [67], hence the ows
are local, nonlinear partial dierential equations.
Let us consider the case qt = 0 for n = 2; d = 2; this corresponds to the
Harry Dym equation in the extended KdV hierarchy. We take m = 3 , that
is, a scalar multiple of 1 0 
3 = 0 1 ;
and q to be a constant vector, to be determined below. Then qt = 0 implies
that
R1F0 = @F 0
@x [q; F0] = 0:
We take F0 = q. The equation [m; F2 ] = 0 implies F2 is diagonal, and since
we are working in the class of traceless matrices, we take F2 = f2 3 .
108 CHAPTER 7. WEAK AND STRONG NONLINEARITIES
The recursion relation is then
[3 ; F1] = @f 2
@x 3 f2[q; 3 ]:
The diagonal part of both commutators vanish, hence we must have
@f2 = 0;
@x
and
[3 ; F1 f2q] = 0:
This last equation implies that
F1 f2q = n3
for some function n. Solving for F1 we obtain
F1 = f2 q + n 3 :
The ow is then
f n 
t 3 D m2 q  3 + [3 ; q] + n [q; 3 ] = 0
From the diagonal part of this equation we obtain
t = D n
and the o-diagonal part gives
 n
[3 ; q]   = f2 D 1 q:
Hence q must be an eigenvector of ad 3 , i.e. q =  . We then have
n =   D1;
 
and
1
t = x   : (7.3.23)
xx
7.3. THE EXTENDED AKNS HIERARCHY 109
We now turn to the negative ows. For the ow at n = 1 we get the
equations
R1F1 =0 (7.3.24)
qt + R0F1 =R1F0 ; (7.3.25)
mt + R0F0 =0 (7.3.26)
We rst consider the special case qt = 0. In that case the rst equation
is solved by taking F1 = q, and the second equation reduces to
[q; m F0 ] = @F @x
0: (7.3.27)
The third equation cannot be solved if we take m = f3, so we take
m = a0 3 + a++ + a  ; F0 = f03 + f++ + f  :
We do look for a symmetry reduction and choose all matrices to lie in
su(2), the real subalgebra of s`(2; C ) spanned by the skew adjoint matrices
i 0  0 i   0 1
0 = 0 i ; 1 = i 0 ; 2 = 1 0 :
Specically, we take
1
i a  f 
q = 2 0 ; m = a i ; F0 =  f 0
0
so that the form of the isospectral operator is
@
i a  1 1 0 
Dx = @x z a i i 2 0 1
Then (7.3.27) leads to the equations
f0;x = 0 a = i(D i)
We take f0 = 0. After some computations the ow equations (7.3.26) lead to
t + Dj j2 = 0; (D i) t = 2 : (7.3.28)
110 CHAPTER 7. WEAK AND STRONG NONLINEARITIES
Chapter 8
Numerical Methods
In order to test the validity of various model equations for nonlinear disper-
sive systems, it is extremely handy to have some numerical codes to solve
dispersive systems of partial dierential equations and compare them with
the various approximations proposed. To this end, Yi Li1 and I have writ-
ten some simple codes to be run in Matlab. These solve nonlinear evolution
equations in one space variable and then animate the solutions.
The codes are based on implicit spectral schemes. The use of spectral
methods for the numerical analysis of dispersive equations has been studied
extensively, [30], [75]. Spectral, or pseudo-spectral, methods, are based on
the use of the fast Fourier transform. In this chapter I will describe the fast
Fourier transform, and explain, by way of a simple numerical experiment,
why the use of spectral methods is preferable to the use of nite dierence
schemes for this class of problems. The implicit schemes to be used are
unconditionally stable and employ the Crank-Nicolson, or trapezoid, method
for the time-step integration. Since the nonlinear terms play a predominant
role in the calculations, we shall describe nonlinear time step solvers.

8.1 The nite Fourier transform

Consider a function u(x) dened on the interval [0; 2], let N be a positive
integer, and let h = 2=N . We restrict u to its values on N evenly spaced
partition points fjhg; j = 0; : : : N 1. Suppose we approximate u(x) by a
1 currently a postdoctoral fellow at the University of Minnesota

111
112 CHAPTER 8. NUMERICAL METHODS
trigonometric polynomial
X
N 1
u(x)  qk eikx (8.1.1)
k=0
and interpolate at the points jh. Thus, we require that
X
N 1
uj = qk !jk; j = 0; : : : ; N 1 ! = e2i=N ;
k=0
where uj = u(jh):
The mapping q 7! u is described by a simple matrix multiplication
u =
q;


= !jk ; j; k = 0; : : : ; N 1:
A direct computation shows that

 = NI ;
hence

1 = 1
 ;
N
and NX1
1
qj = N uk ! kj :
k=0
The mapping from the vector u = (u1; 0tsuN ) is called the nite Fourier
transform.
A second way to interpret the nite Fourier transform is the following.
Let u be a 2-periodic function on the line, with uj = u(jh) for all j 2 Z.
Dene the centered rst order dierence operator
Duj = uj+1 2h uj 1 :
The operator D is skew symmetric. Its eigenvalues and eigenvectors are
easily found. For xed k, let k;j = !jk. Then
! (j +1)k ! (j 1)k  !k ! k 
Dk;j = =! jk = k k;j
2h 2h
8.1. THE FINITE FOURIER TRANSFORM 113
where
sin 2Nk
k = i 2 :
N
Thus, the eigenvectors of D are precisely the column vectors of the -
nite Fourier transform. Since D is skew symmetric, the eigenvectors of D
are orthogonal, and the nite Fourier transform diagonalizes D. This is the
discrete analog of the fact that the Fourier series diagonalizes the skew sym-
metric operator d=dx.
The implementation of the nite Fourier transform requires matrix mul-
tiplication by
. This computation is of order N 2 . On the interval [0; 2] a
choice of N = 256 gives a spatial resolution h = :0245, while 2562 = 65536.
In a numerical integration of a system of partial dierential equations, N 2
oating point operations, or ' ops', must be performed at each time step.
Spectral methods in numerical analysis are based on the discovery, in
1965, by Cooley and Tukey [24], that when N is a power of 2 the nite
Fourier transform can be implemented in N log N steps. For N = 256 = 28
this is roughly 8
256 = 2048. This is a very decided advantage. Most nu-
merical packages, including Matlab, have a nite Fourier transform package.
In Matlab, the fast Fourier transform and its inverse are implemented by the
commands 't' and 'it'.
To compare spectral methods with nite dierence schemes, consider the
numerical integration of the partial dierential equation
ut + ux = 0; u(x; 0) = f (x): (8.1.2)
We partition both space and time into nite intervals, with the spatial step
denoted by h =
x and the time step denoted by
t.
First we construct a nite dierence scheme. For reasons we shall explain
later, we are going to use an implicit method, which dates back to Euler. We
rst replace dierentiation in x by the centered dierence operator D dened
above. Then write the equation in the form
ut = Du
Now let's integrate this equation from t to t +
t. We obtain
Z t+
t
u(x; t +
t) u(x; t) = Du dt:
t
114 CHAPTER 8. NUMERICAL METHODS
Now we approximate the integral on the right by the trapezoid rule to obtain
u(x; t +
t) u(x; t) = Du(x; t +
2t) + Du(x; t)
t:
This equation may be written in the form
(I + 21
tD)u(x; t +
t) = (I 21
tD)u(x; t);
or symbolically, by
I 12
tD
u(x; t +
t) = Uu(x; t); U =
I + 21
tD
There is no problem in inverting I + 21
tD since D is skew adjoint and its
eigenvalues are imaginary. The operator U is unitary2, and its L2 norm is 1,
regardless of the size of
t.
Now let us implement this scheme, rst as a nite dierence scheme, and
then as a spectral scheme using the fast Fourier transform. The matrix for
D on the space of periodic functions is the tridiagonal matrix
00 1 0 0 ::: 1
1
BB 1 0 1 0 ::: 0CC
B0 0C
D = 21h B C
1 0 1 :::
BB ... CC
B@ 0 0 1A
C
1 0 1 0
In order to speed the computations and use less memory, we represent this
matrix as a sparse matrix.
Here is the Matlab code:
ops(0) % set ' oating point operations' counter to 0;
N=input('N=');
T=1; L=2*pi;
h=L/N
x=(h:h:L)';
l=1; % input('l=');
2U is the Cayley transform of 1
2D ; cf. Riesz and Nagy [63]
8.1. THE FINITE FOURIER TRANSFORM 115
dt=l*h;
J=x(T/dt);
u=zeros(N,J);
u(:,1)= exp(-20*(x-.5*pi).b2);
e=ones(N,1);
D=spdiags([-e,e], -1:2:1,N,N); % sparse tridiagonal matrix.
D(1,N)=-1; D(N,1)=1; % modify for periodic functions.
id=sparse(eye(N));
A=id+.5*l*D;
B=2*id-A;
for j=2:J;
u(: , j)=An B*u(:,j-1);
end
ops
plot(x,u(:,J))

Implicit Finite Difference Scheme

1

0.8

0.6 T=3

66,754,090 flops

0.4 N=256, h=.0245

0.2

dispersive tail

−0.2
0 1 2 3 4 5 6 7

Figure 8.1: Solution of ut + ux = 0 using an implicit nite dierence scheme.

The result of this code with N = 256, and T = 3, is shown in Figure 8.1.
The wiggles trailing the pulse are due to dispersion in the numerical scheme
116 CHAPTER 8. NUMERICAL METHODS
cf. [37]. When ux is approximated with the centered dierence, the error
looks like u000 (x):

u0(x) = u(x + h) 2h u(x h) 1 h2 u000 (x) + O(h4);

3!
so from the point of view of the numerical scheme, one is approximating
the original equation by an equation of the form ut + ux + "uxxx = 0 for
" = h2=12:.
Remark: This example is overly prejudicial to nite dierence methods.
The dispersion is due in part to the use of the centered dierence operator for
ux, which is not the correct choice for this problem. K. Siebert of Freiburg
Universitat, one of the participants in the Lecture Series, pointed out that
the computation could be done in fewer operations if an ecient algorithm
for the inversion of a tri-diagonal matrix is used. Matlab apparently uses a
general scheme for matrix inversion, which, while ecient, does not use the
best scheme for a tridiagonal matrix.

Implicit Spectral Scheme

1

0.9

0.8

0.7

0.6 T=3

0.5 1,835,794 flops

0.4 N=256, h=.0245

0.3

0.2

0.1

0
0 1 2 3 4 5 6 7

Figure 8.2: Solution of the equation ut + ux = 0 using an implicit spectral

method; N=256, h=.0245; T=3.
8.1. THE FINITE FOURIER TRANSFORM 117
Now let us redo the computation using the nite Fourier transform. For
the theoretical discussion it is convenient to use the centered form of the
nite Fourier transform pair [30]:
X
N 1
qj = uk ! jk ; j = N2 ; : : : ; N2
j =0

uk = u(kh); h = 2N ; ! = e2i=N :

The inverse transform is given by
X2
N

uk = N1 qj !jk;
j= N 2
where it is understood that only half the contributions at N=2 are included
in the sum.
The ordering of the basis functions in the Matlab fast Fourier tranform
algorithm unfortunately corresponds to that given by the matrix
above,
and is a little harder to work with. From the centered form of the tranform it
follows easily that when u is real the Fourier coecients satisfy the familiar
symmetry
q j = qj : (8.1.3)
This implies that both q0 and qN=2 are real when N is even. This is a re ection
of the fact that the cyclic group C N has two real representations when N is
even. This leads to some problems in using the Fourier transform for even
N.
If u is approximated by the trigonometric polynomial (8.1.1) then its
derivative is given by
X
N=2
ux  ijqj eijx;
k= N=2
hence dierentiation is obtained on the transform side by multiplication by
the diagonal matrix D with entries ij , j = N=2; : : : N=2.
If u satises the partial dierential equation (8.1.2) then its Fourier trans-
form coecients qj (t) satisfy the ordinary dierential equations
q_j + ijqj (t) = 0:
118 CHAPTER 8. NUMERICAL METHODS
This ow preserves the symmetry (8.1.3) of the Fourier coecients. In par-
ticular, q0 remains constant, corresponding to the conservation law
Z 2
u(x; t) dx = const:
0
On the other hand,
q N2 (t) = q N2 (0)ei N2 t
and cannot remain real under the ow. Thus use of the nite Fourier trans-
form for even N leads to an error in the computation.
This error is unavoidable, but when the solutions remain smooth, it is
usually not a serious error. Just as in the case of the Fourier series for
the group S 1 , the Fourier coecients of a smooth function decrease rapidly
with j . For example, if u(x) is s times continuously dierentiable then its
Fourier coecients decrease like j s, while if u is analytic, then its Fourier
coecients decrease exponentially. The same holds for the nite Fourier
transform, so if u(x; 0) is analytic in x the quantities qj decrease exponentially
with j . In particular, qN=2 is extremely small. These facts can be conrmed
experimentally by doing a few sample calculations in Matlab.
The solution of (8.1.2) is given to a very good approximation by
0 2 1
X N

u(x; t) = < @ N1 qj (0)eijt!jk A :

j = N2
Here is a Matlab code for an implicit spectral scheme for the integration
of (8.1.2).
%spectral integration of ut = ux.
ops(0);
N=256;
T=3;
L=2*pi; h=2*pi/N
l=1; dt=l*h;
J=x(T/dt);
x = (h:h:2*pi)';
f= exp(-20*(x-.5*pi).b2);
k = [(0:N/2)';(1-N/2:-1)'];
m=.5*dt*k;%
8.2. PSEUDOSPECTRAL CODES 119
d=(1-i*m)./(1+i*m);
U=sparse(diag(d));
v=zeros(N,J);
v(: , 1)=t(f);
for j=2:J;
v(: , j)=U*v(: , j-1);
end
u =real( it(v));
ops
plot(x,u(: , J))

These two codes show clearly the advantage of spectral methods over
nite dierence methods. The solution of (8.1.2) using the spectral method
for T = 3 is shown in Figure 8.2. When the nite Fourier transform is used,
the choice of dierence operators is immaterial; moeover, when analyzing
higher order dierential equations, such as the KdV equation, the dierence
operators involved are no longer tri-diagonal, and the spectral method gains
more and more of an advantage as the power of the operator D goes up.

8.2 Pseudospectral Codes

The scheme developed by Yi Li and myself to integrate the Korteweg-deVries
equation uses a spectral method with an implicit method for the time step,
similar to that discussed in [75]. This leads to a nonlinear equation to solve at
each time step, and the scheme uses a simple iteration at each stage to solve
the nonlinear equation by successive approximations. Fornberg and Whitham
[30] used a leap-frog method with an explicit time step. To achieve any kind
of accuracy with an explicit method, one must take very small time steps.
In the case of the third order operator D3 one must have
t = O((
x)3).
An implicit method which uses the solution u at the preceeding step is
generally unstable. We got much better numerical results by averaging the
nonlinear term over the previous and current time step, and incorporating
an iteration on the nonlinear term in the scheme. The size of the time step is
determined by the requirement that the iteration scheme converge suciently
rapidly. Wineberg et. al. [75] remark that they found it sucient to simply
carry out two iterations at each step.
120 CHAPTER 8. NUMERICAL METHODS
Applying the `trapezoid' argument to the full KdV equation, we get
u(t +
t) = Uu(t) B (u2(t +
t) + u2(t));
where
I 21
tD3
U= ; B = 1 +:25
tD : (8.2.4)
I + 21
tD3 :5
tD3
The operator U is unitary. U and B are evaluated using the fast Fourier
tranform algorithm. The fast Fourier transform in Matlab is called by 't'.
The inverse fast Fourier transform is called up by 'it'. Let us set
v = t(u); u = it(v):
Then the above equation can be written in the form
v(t +
t) = Uv(t) B (u2(t) + u2(t +
t)):
The nonlinear terms are best computed in the spatial representation, so we
transform back to the original spatial picture, carry out the multiplication,
which is pointwise on the x side, and then transform back. This is a nonlinear,
implicit scheme, since it is nonlinear in u(t +
t):
The nal step in the procedure is to solve this nonlinear scheme by suc-
cessive approximations.3 That is, we write a subroutine to carry out the
iterations
vj+1(t +
t) = Uv(t) B (u2(t) + u2j (t +
t));
uj = it(vj ); v1 = fft(u(t)):
This scheme is quite robust and gives extremely good accuracy for N
suciently large (say N = 512) and the time steps suciently small (say

t = h = 2=N ). We compared the computed two soliton solution with the
exact solution and found the dierence was negligible after the interaction.
The calculations are carried out on 2 periodic functions, since we are
using the nite Fourier transform. However, in order to place two solitary
waves in the interval [0; 2] one nds that the velocities c1 and c2 must be
3 One could try to set up a Newton iteration scheme, which converges quadratically,
but this is complicated by use of the Fourier transform.
8.2. PSEUDOSPECTRAL CODES 121
relatively large (of the order of 3,5) in order that the solitons be suciently
narrow. The solitary waveform for the equation ut + uxxx + uux = 0 is
12c2sech2 cx
so its height is proportional to c2; and this leads to numerical instability
due to the nonlinear terms. With c = 5 the amplitude of the solitary wave is
then 300, and this large amplitude causes instability in the numerical scheme,
necessitating very small time steps.
It is therefore necessary to rescale the problem to an interval 0  y  L.
We let w satisfy
wt + wyyy + wwy = 0; 0  y  L;
and dene
u(x; t) = a1 w(ax; t); a = 2L :
Then u satises the equation
ut + a13 uxxx + uux = 0 0  x  2:
We found it sucient to take L = 40. We have also included a coecient
of dispersion, disp, in the code.
I close with the code, as it was amended by Richard Palais to graph the
solution as it is being computed:
N=256; %input('N=');
L = 40; a=L/(2*pi); h=2*pi/N;
T=300; %input('T=');
dt=.05;
x =(h:h:2*pi)';
y=a*x;
c1=2;
u= exp(-c1*a*(x-.5*pi)2)/a;
k=[(0:N/2)';(1-N/2:-1)'];
disp = 0.05; %disp=input('disp=');
m=disp*a 3*.5*dt*k.3;
d1=(1+i*m)./(1-i*m);
d2= -.5*i*dt*k./(1-i*m);
122 CHAPTER 8. NUMERICAL METHODS
d3=.5*d2;
sol= plot(y,a*u);
axis([ 0 L -.2 1.75]);
set(sol,'erasemode','xor')
t=0;
while t < T ;
tu = t(u); tuu = t(u.^2);
v=real(it(d1.*tu+0.5*d2.*tuu));
w=real(it(d1.*tu+d2.*tuu));
for n=1:3;
w=v+real(it(d3.*t(w.^2)));
end
u=w;
t=t+dt;
set(sol,'ydata',a*u); drawnow;
end
To do the computation with Whitham's model simply replace the factor
k:b3 by [1 ; sqrt(tanh(k)./k)] in the line dening m.
We can also see from this numerical scheme the mechanism by which the
dispersive term smooths out the shocks developed by the nonlinear term.
The operator B dened in (8.2.4) acts as a low-pass lter for the Fourier
modes. To see this, note that on the spectral side B is multiplication by
:25i
tk
1 :5i
t(disp)a 3 k3 :
The graph of the spectral properties of the operator B not only makes
clear how dispersion and nonlinearity compete, but it also suggests why this
numerical algorithm is so eective for weakly nonlinear systems. The nonlin-
ear terms push energy into the higher modes; but it is in turn ltered out by
the high frequency cut-o. This cut-o is characterisitic of weak nonlinearity,
since the operator B will always consist of lower order dierential operators
divided by higher order operators.
8.2. PSEUDOSPECTRAL CODES 123

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 50 100 150 200 250 300

Figure 8.3: High frequency cut-o due to the dispersion acting on the non-
linear terms.
t = :05; disp = :05, L = 40; N = 256.
124 CHAPTER 8. NUMERICAL METHODS
Bibliography
[1] Ablowitz, M.J., D.J. Kaup, A.C. Newell, & H. Segur. The
inverse scattering transform {Fourier analysis for nonlinear problems.
Studies in Applied Mathematics, 53:249{315, 1974.
[2] Airy, G.B. Tides and Waves. In Encyclopaedia Metropolitana, London,
1845. article 192.
[3] Alt, H.W. & L.A. Caffarelli. Existence and regularity for a mini-
mum problem with free boundary. J. Reine Angew. Math., 325:105{144,
1981.
[4] Amick, C.J., L.E. Fraenkel, and J.F. Toland. On the Stokes
conjecture for the wave of extreme form. Acta Mathematica, 148:193{
214, 1982.
[5] Beals, R. & Coifman, R. Scattering and inverse scattering for rst
order systems. Comm. Pure. Appl. Math., 37:39{90, 1984.
[6] Beals, R., P. Deift, C. Tomei. Direct and Inverse Scattering on the
Line, volume 28 of Mathematical Surveys and Monographs. American
Math. Soc., 1988.
[7] Benjamin, T.B. Internal waves of permanent form in uids of great
depth. J. Fluid Mech., 29:559{572, 1967.
[8] Benjamin, T.B. The stability of solitary waves. Proc. Roy. Soc. Lon-
don A, 328:153{183, 1972.
[9] Benjamin, T.B. Impulse, ow-force, and variational principles. IMA
Journal of Applied Mathematics, 32:3{68, 1984.
125
126 BIBLIOGRAPHY
[10] Benjamin, T.B., and P.J. Olver. Hamiltonian structure, symme-
tries and conservation laws for water waves. J. Fluid Mech., 125:137{185,
1982.
[11] Benjamin, T.B., Bona, J. & Mahoney, J.J. Model equations for
long waves in nonlinear dispersive systems. Phil. Trans. Royal Soc.
London A., 272:47{78, 1972.
[12] Birkhoff, G. Hydrodynamics, a Study in Logic, Fact, and Similitude.
Dover, New York, 1950.
[13] Birkhoff, G. and E.H. Zarantonello. Jets, Wakes, and Cavities.
Academic Press, New York, 1957.
[14] Bogoliubov, N.N. and Y.A. Mitropolsky. Asymptotic Methods
in the Theory of Nonlinear Oscillations. Gordon and Breach, New York,
1961.
[15] Bona, J. Solitary waves and other phenomena associated with model
equations for long waves. Fluid Dynamics Transactions, 10:77{111,
1980.
[16] Bona, J.L., W.G. Pritchard, & L.R. Scott. An evaluation of
a model equation for water waves. Phil. Trans. Royal Soc. London A.,
302:457{510, 1981.
[17] Bona, J.L., W.G. Pritchard, & L.R. Scott. A comparison of
solutions of two model equations for long waves. In Norman R. Lebovitz,
editor, Fluid Dynamics in Astrophysics and Geophysics, volume 20 of
Lectures in Applied Mathematics, pages 235{267, Providence, R.I., 1983.
Amer. Math. Soc.
[18] Boussinesq, M.J. Theorie des ondes et des remous qui se propagent
le long d'un canal rectangulaire horizontal, en communiquant au liquide
contenu dans ce canal des vitesses sensiblement pareilles de la surface
au fond. J. Math. Pure. Appl., 17:55{108, 1872.
[19] Boussinesq, M.J. Essai sur la theorie des eaux courantes. Memoires
presentes par divers savants a l'Academie des Sciences Inst. France
(series 2), 23:1{630, (1877).
BIBLIOGRAPHY 127
[20] Brillouin, L. Wave Progagation and Group Velocity. Academic Press,
New York, 1960.
[21] Camassa, R. and D.D. Holm. An integrable shallow water equation
with peaked solitons. Phys. Rev. Lett., 71:1661{1664, 1993.
[22] Camassa, R., D.D. Holm, & J.M. Hyman. A new integrable shallow
water equation. Advances in Applied Mechanics, 31:1{33, 1994.
[23] Cercignani, C. Solitons. theory and application. Revista del Nuovo
Cimento, 7:429{469, 1977.
[24] Cooley, J. W. & J.W. Tukey. An algorithm for the machine cal-
culation of complex fourier series. Math. Comp., 19:297{301, 1965.
[25] Craig, W. An existence theory for water waves and the Boussinesq and
Korteweg-deVries scaling limits. Commun. Partial Di. Eq., 10:787{
1003, 1985.
[26] Dodd, R. K., J. C. Eilbeck, J. C. Gibbon & H. C. Morris.
Solitons and Nonlinear Wave Equations. Academic Press, New York,
1982.
[27] Dorfmeister, J., & Szmigielski, J. get title. Mathematica Appli-
canda, 123:456{457, 1998.
[28] Erdelyi, A. Asymptotic Expansions. Dover, New York, 1956.
[29] Fermi, E., Pasta, J.R. and Ulam, S.M. Studies of nonlinear prob-
lems. In Collected Works of E. Fermi, vol. II, pages 978{988. University
of Chicago Press, 1965. cf. also Los Alamos Sci. Lab. Tech. Report
LA-1940.
[30] Fornberg, B. & G.B. Whitham. A numerical and theoretical study
of certain nonlinear wave phenomena. Phil. Trans. Royal Soc. London,
289:373{404, 1978.
[31] Friedman, A. Variational Principles and Free Boundary Problems.
John Wiley & Sons, New York, 1982.
[32] Friedrichs, K.O. U ber ein Minimumproblem fur Potentialstromung
mit freiem Rand. Math. Ann., 109:60{82, 1934.
128 BIBLIOGRAPHY
[33] Friedrichs, K.O. and D.H. Hyers. The existence of solitary waves.
Comm. Pure Appl. Math., 7:517{550, 1954.
[34] Fuchsteiner, B. The Lie algebra structure of nonlinear evolution
equations and innite dimensional abelian symmetry groups. Progr.
Theor. Phys., 65:861{, 1981.
[35] Gardner,C.S., J.M. Greene, M.D. Kruskal, and R.M. Miura.
Method for solving the Korteweg-deVries equation. Phys. Rev. Lett.,
19:1095{1097, 1967.
[36] Gardner, C.S. & Morikawa, G.K. Similarity in the asymptotic
behavior of collision-free plasma and water waves. Technical Report
NYO-9082, Courant Inst. Math. Sci., 1960.
[37] Goodman, J. & Lax, P. D. On dispersive dierence schemes. I.
Comm. Pure Appl. Math., 41:591{613, 1988.
[38] Haragus, M. and D.H. Sattinger. On the linearized Korteweg-
deVries equation. to appear, Zeit. Ang. Math. u. Phys.
[39] Jeffrey, A. and T. Kawahara. Asymptotic Methods in Nonlinear
Wave Theory. Pittman, London, 1982.
[40] Jeffrey, A. & T. Kawahara. Asymptotic Methods in Nonlinear
Wave Theory. Applicable Mathematics Series. Pitman, Boston, 1982.
[41] Kirrmann, P. G. Schneider, & A. Mielke. The validity of mod-
ulation equations for extended systems with cubic nonlinearities. Proc.
Royal Soc. Edinburgh, 122A:85{91, 1992.
[42] Kruskal, M. Nonlinear Wave Equations. In J. Moser, editor, Dynam-
ical Systems, Theory and Applications, volume 38 of Lecture Notes in
Physics, pages 310{354, Heidelberg, 1975. Springer.
[43] Kruskal, M.D., & Z.J. Zabusky. Interaction of \solitons" in a
collisionless plasma and the recurrence of initial states. Phys. Rev. Lett.,
15:240{243, 1965.
[44] Lamb, G.L. Jr. & D. W. McLaughlin. Aspects of soliton physics.
In R. K. Bullough and P. J. Caudrey, editors, Solitons, volume 17 of
Topics in Current Physics, pages 65{106. Springer-Verlag, 1980.
BIBLIOGRAPHY 129
[45] Lax, P.D. Integrals of nonlinear equations of evolution and solitary
waves. Comm. Pure and Applied Math., 21:467{490, 1968.
[46] Lax, P.D. Almost periodic solutions of the KdV equation. SIAM
Review, 18:351{375, 1976.
[47] Lax, P. D., Levermore, C. D., & Venakides, S. The generation
and propagation of oscillations in dispersive initial value problems and
their limiting behavior. In Important Developments in Soliton Theory,
Nonlinear Dynamics, pages 205{241, Berlin, 1993. Springer.
[48] Levi-Civita, T. Determination rigoureuse des ondes permanentes
d'ampleur nie. Math. Ann., 93:264{314, 1925.
[49] Luke, J.C. A variational principle for a uid with a free surface. J.
Fluid Mech., 27:419{448, 1967.
[50] Maddocks, J. and R.L. Sachs. On the stability of KdV multi-
solitons. Comm. Pure Appl. Math., 46:867{901, 1993.
[51] Magri, F. A simple model of integrable Hamiltonian equations. Jour.
Math. Phys., 19:1156{1162, 1978.
[52] Newell, A. C. Solitons in Mathematics and Physics. CMBS-NSF
Monographs. Society for Industrial and Appied Mathematics, Philadel-
phia, 1985.
[53] Novikov, S., Manakov, S.V., Pitaevskii, L.P. and Zakharov,
V.E. Theory of Solitons. Consultants Bureau, New York and London,
1984.
[54] Oishi, S. Relationship between Hirota's method and the inverse spectral
method- the Korteweg-deVries equation's case. Jour. Phys. Soc. Japan,
47:1037{1038, 1979.
[55] Olver, P.J. Hamiltonian and non-Hamiltonian models for water waves.
In P.G. Ciarlet and M. Roseau, editor, Trends and Applications of Pure
Mathematics to Mechanics, volume 195 of Lecture Notes in Physics,
pages 273{290, Providence, R.I., 1984. Springer-Verlag.
[56] Olver, P.J. Hamiltonian perturbation theory and water waves. Con-
temporary Mathematics, 28:231{249, 1984.
130 BIBLIOGRAPHY
[57] Olver, P.J., & P. Rosenau. Tri-Hamiltonian duality between soli-
tons and solitary-wave solutions having compact support. Physical Re-
view E, 53:1900{1906, 1996.
[58] Ono, H. Algebraic solitary waves in stratied uids. Jour. Phys. Soc.
Japan, 39:1082{1091, 1975.
[59] Palais, Richard S. The symmetries of solitons. Bull. Amer. Math.
Soc., 34:339{403, 1997.
[60] Poppe, C. The Fredholm determinant method for the KdV equations.
Physica D, 13:137{160, 1984.
[61] Rayleigh, J.W.S. On Waves. Phil. Mag., 1, 1876.
[62] Rayleigh, J.W.S. Theory of Sound, I. II. Dover, New York, 1945.
[63] Riesz, F. & Sz. Nagy. Functional Analysis. Frederick Ungar Pub-
lishing Co., New York, 1955.
[64] Rosenau, P. Nonlinear dispersion and compact structures. Phys. Rev.
Lett., 73:1737{1741, 1994.
[65] Russell, J. Scott. Report on waves. In John Murray, editor, Report
on the Fourteenth Meeting, pages 311{390+57 plates, London, 1844.
British Association for the Advancement of Science.
[66] Sattinger, D.H. Topics in Bifurcation Theory, volume 309 of Lecture
Notes in Mathematics. Springer Verlag, 1973.
[67] Sattinger, D.H. Hamiltonian hierarchies on semi-simple Lie algebras.
Studies in Applied Mathematics, 72:65{86, 1985.
[68] Sattinger, D.H. & V.D. Zurkowski. Gauge Theory of Backlund
Transformations, II. Physica 26D, pages 225{250, 1987.
[69] Spitzer, L. Physics of Fully Ionized Gases. Interscience, New York,
1956.
[70] Stoker, J.J. Water Waves. Wiley Interscience, New York, 1957.
[71] Stokes, G.G. On the theory of oscillatory waves. Trans. Camb. Phil.
Soc., 8:197{229, 1847.
BIBLIOGRAPHY 131
[72] Struik, D.J. Determination rigoueuse des ondes irrotationnelles
periodiques dans un canal a profondeur nie. Math. Ann., 95:595{634,
1926.
[73] Toland, J.F. On the existence of a wave of greatest height and Stokes'
conjecture. Proc. Roy. Soc. London, A, 363:469{485, 1978.
[74] Whitham, G.B. Linear and Nonlinear Waves. John Wiley & Sons,
New York, 1973.
[75] Wineberg, S.B., J. McGrath, E. Gabl, L.R. Scott, & C.
Southwell. Implicit spectral methods for wave propagation problems.
Comp. Phys., 97:311{336, 1991.
[76] Zakharaov, V.E., and A.B. Shabat. Exact theory of two dimen-
sional self-focusing and one dimensional self-modulation of waves in non-
linear media. Soviet Phys. JETP, 34:62{69, 1972.
[77] Zakharov, V.E. Stability of periodic waves of nite amplitude on the
surface of a deep uid. J. Appl. Mech. Tech. Phys., 2:190{194, 1968.
Index

132