Download as pdf or txt
Download as pdf or txt
You are on page 1of 59

Applied Analysis for Nonlinear Mechanics

Shaoqiang Tang

Department of Mechanics and Engineering Science

Peking University

maotang@pku.edu.cn

http://cmam.pku.edu.cn/tangsq

Tel: 62871980

Time & Venue:
Tuesday 14:4016:30, 123
Thursday (even weeks only) 14:4016:30, 207

Chapter 0 Introduction
Purpose of the Course:
To show some modern (19001990?) mathematical methods that are widely used in
nonlinear mechanics and other physical sciences
To help initiating scientific life: ideas, realizations and presentations

Why nonlinear: The world is nonlinear, whereas linear is a special case, usually a
simplified version, therefore incomplete. Lots of important features of the real world can only be
explained in the framework of nonlinear systems.

Main feature of a linear system: superposition.

Example 1: Newtons second law of motion: acceleration is proportional to the force. Two-body
problem:
2
2
'' ,
| |
'' .
| |
GM
x
x y
Gm
y
x y




Example 2: avier-Stokes equation for incompressible fluid
,
N
( )
2
0,
1
.
Re
t
v
v v v p

+ + =

-
, , ,
-

v
,


References:

1, Grindrod P, Patterns and Waves, Claredon, 1991.
2, Whitham G.B., Linear and Nonlinear Waves, John Wiley & Sons, 1974.
3, 1987.

Grading formula:
30% assignments + 30% mid-term + 40% final (subject to modification later)
2

Contents
Chap 1 Qualitative theory for ODE systems
1, Critical point
2, Plane analysis for second order ODE
3, Stability, Lyapunov functional
4, Bifurcation theory: an introduction

Chap 2 Reaction-diffusion systems
1, Introduction: BVP and IBVP, equilibrium
2, Dispersion relation, linear stability
3, Invariant domain
4, Perturbation method
5, Traveling waves
6, Example: evolutionary Duffing equation

Chap 3 Discontinuities: hyperbolic conservation laws
1, Shock formation in Burgers' equation, characteristics
2, Elementary waves: Part I -- shock waves via vanishing viscosity
3, Elementary waves: Part II -- slip lines, rarefaction wave
4, Riemann problem for gas dynamics

Chap 4 Fixed point theory with applications
1, Basic concepts of normed spaces, some inequalities
2, Spaces of functions, fixed point theorem
3, Applications: ODE and integral equations

Chap 5 Solitons
1, Solitons in KdV equation
2, Inverse scattering theory: an introduction

Chap 6 Chaos
1, Lorenz system
2, Logistic map, and routes to chaos
3
Chapter 1 Qualitative Theory for ODE Systems
1 Critical point

An ODE (Ordinary Differential Equation) is an equation that contains an unknown
function together with one or more of its derivatives with respect to a single independent variable.
An ODE system, also called a dynamical system is a set of ODEs. Typically each equation only
contains one first order derivative. The unknown function(s) (dependent variable) is called a
state variable(s). The order of the highest derivative in an ODE is the order of the ODE. In case of
an ODE system, the number of equations is the order of the ODE system.
A high-order ODE can usually be recast into an ODE system of the same order. For
instance,
'' ' (1 ) ( ) 0 x xx x x f t + + + =
can be rewritten as
' ,
' [ (1 ) (
x y
y xy x x f
=

= + +

)]. t
),

Therefore, a general ODE system reads
' ( , x f x t = with x x
1
( , , ) .
T n
n
x = R
To solve such a system, i.e. to find an expression for x in terms of t, initial data
0
( )
0
x t = x should be specified, when one looks for a solution in
0
[ , ) t t + . This solution is
denoted as
0 0
( ; , ) x t t x
+
= . Meanwhile, the solution for t
0
( , ] t is denoted as
0
,
0
) ( ; x t t

= x . On the other hand, sometimes one specifies boundary data, namely n algebraic
equations for n quantities among
1 1
( ), , ( ), ( ), ,
n
( )
n
x a x a x b x b , when one looks for a
solution in . [ , ] t a b
The ODE system is said to be autonomous if the right hand side depends only on the state
variable, that is, f(x,t)=f(x). In fact, a non-autonomous ODE system can be transformed to an
autonomous one in the following way. We set a new state variable as and take
then we have y=F(y), which is autonomous and of order (n+1). In the
following, we shall discuss autonomous system, unless otherwise mentioned.
( , )
T T
y x t = ,
( ) ( ( , ) ,1) ,
T T
F y f x t =
We recall that the existence, uniqueness and smoothness (regularity) for the solution with
given initial data hold for quite general cases, e.g., when the source term (right hand side) is
continuous. Existence will be discussed later on in the part of fixed point theory.
Instead of solving an ODE system for specific initial data, we shall investigate the system
in a general manner. Each initial data corresponds to a point ( , in , and the
corresponding solution draws a positive semiorbit
0 0
) t x
n
R
0 0
( ; , ) x t t x
+
=
0
( ; ,
that goes through .
The other branch corresponds to a negative semiorbit
0 0
( , ) t x
0
) x t t x

=
0
)
with the same initial data.
In fact, essentially the same solution is obtained if we start from any point on this curve, except a
time difference. The complete curve
0
( ; , x t t x = is called a trajectory, or a phase curve, or
an orbit, and the space is the phase space. In the special case when n=2, is called the
phase plane. In qualitative theory, one discusses the global structure of trajectories on the phase
space, instead of finding the solution for specific given initial data.
n
R
2
R
At each given point x, the source term introduces a vector f(x) in the phase space. The
direction of this vector determines the direction of the trajectory, and the absolute value
determines how fast a solution takes to go through this point. We may imagine that there is a
particle moving along the trajectory according to the vector field (velocity field). See Figure 1.
4

Figure 1 Trajectories, vector field and phase space (plane).

Noticing that at a point x where f(x) vanishes, the previous statement become meaningless.
This leads to the notion of critical point (equilibrium point), which turns out to be crucial in our
discussion.
A critical point x is the point where f(x)=0. It is a regular point otherwise.
One key feature for critical point is that two trajectories may intersect only at a critical
point. This can be derived from the uniqueness theorem for ODE system.
To understand the local behavior near a critical point
0
x , we perform Taylor expansion
as
0 0
( ) ( ) ( ) (| |)
0
f x f x x x o x x = + .
The Jacobian , which is assumed to be invertible, determines the
topological structure of trajectories near the critical point. From linear algebra, we know that A is
similar to its Jordan normal form. That is, there is an invertible matrix P, such that ,
where B consists of diagonal entries, as well as Jordan blocks. Now let , the
governing equation for y is
0
( ) A f x =
1
PAP B

=
0
( ) x x y P =
' (| y By o y = + |)
0
.
For the sake of clarity, we confine ourselves to the case of second order ODE system, that
is, A and B are 2x2 matrix. Moreover, unless in degenerate cases, the topological structure is the
same as the linearized system
0
( ) ( ) ( ) f x f x x x = , or ' y By = .
For this linear problem, there are following possibilities:
B

| |
=

\ .
|
with 0 > >
0
, then the critical point is a source (unstable node), where all
trajectories go away from x . Notice that , and x is a lin,ear combination of
the components of y.
1
2
t
t
c e
y
c e

| |
=

\ .
|
B

| |
=

\ .
|
with 0 < <
0
, then the critical point is a sink (stable node), where all
trajectories go towards x .
B

| |
=

\ .
|
with 0 > > , then the critical point is a saddle, where all trajectories go
towards and then leave away from
0
x .
with a , then the critical point is a unstable focus, where all trajectories
a b
B
b a
| |
=

\ .
|
0 >
5
wind away from
0
x .
|
|
with a , then the critical point is a stable focus, where all trajectories
wind towards
a b
B
b a
| |
=

\ .
0 <
0
x . Notice that for certain coefficients.
1 2
2 1
cos sin
cos sin
at at
at at
c e bt c e bt
y
c e bt c e bt
| | +
=


\ .
|

with a , then the critical point is a center, where all trajectories wind
around
a b
B
b a
| |
=

\ .
0
0 =
x . Notice that .
0
0
sin ( )
cos ( )
c b t t
c b t t
+ | |
|
+
\ .
y =



Figure 2 Source. Figure 3 Sink.

Figure 4 Unstable focus. Figure 5 Stable focus.

Figure 7 Center.

When we say toward/away from, we mean the direction of the trajectory (particle)
when time t increases. In the figures, we plot arrows in the direction of increasing t. It is also clear
from
6
( )
i
i
dt
f x
dx
=
that it takes infinite time to enter an unstable critical point, or leave a stable point.
For a nonlinear system, it is not easy to obtain a global picture. However, it may be shown
that around a critical point mentioned above, except for the center, the structure keeps the same as
the linear system. We also mention that there are degenerate cases, such as = . They are of
particular interest when we study the transition from one structure to the other for the geometry of
the trajectories. We shall precise this point in the study of bifurcation theory.

Assignments

1, Show that a vector f(x) (assumed to be non-zero) at a point x is tangential to the trajectory
across this point.

2, Prove that a trajectory remains unchanged if there undergoes a non-singular time transformation,
i.e. ( ) t = with '( ) 0 t .

3, Plot schematically the trajectories in
' 3
' .
, x x y
y x
=

= +

y
),
.


4, Find the critical points of the system
' (1
' (1 ) sin
x y x
y y
=

x

Determine their types. Can you also plot schematically the trajectories?


7
2 Plane analysis for second order ODE

If we consider a second order ODE system, we have trajectories in the phase plane .
has some special properties, which allows us to have a better understanding. For an
autonomous system
2
R
2
R
' ( ,
' ( ,
),
),
x f x y
y g x
=

y

we easily obtain the equation for a trajectory
( , )
( , )
dy g x y
dx f x y
= .
In many important cases, this can be integrated, providing substantial understanding to
the underlying physical system.
In a linear system, there is either only one critical point, or a line of critical points. Closed
orbits exist around a center, with the same period. In a nonlinear system, things can be more
involved. First of all, there can be more than one isolated critical points. Secondly, it is possible
that a trajectory connects two critical points, or forms a loop from and toward the same critical
point. The former one is called a heteroclinic orbit, and the latter a homoclinic orbit.
We illustrate by analyzing the following Duffing equation
.
3
'' 0 u u u + =
This is a second-order equation, and can be integrated once to give
2 4 2
/ 2 ( ') u u u C + = .
The Duffing equation may also be written in terms of an ODE system
3
' ,
' .
u v
v u
=

= +

u

Using the technique mentioned above, we have
3
dv u u
du v
+
= .
The same equation for trajectories is obtained after integration, for certain , 0 C
.
2 4 2
/ 2 u u v C + =
From this, we obtain three critical points: u=0,1,-1,(v=0). Furthermore, u=0 is a center,
and the other two are saddles. In fact, the Jacobian is
2
0 1
1 3 0 u
(
(
+

, and the eigenvalue
equation is .
2 2
(1 3 ) 0 u + =

Figure 7 Trajectories in Duffing equation.

A trajectory is a closed orbit if (0,1/ 2) C . When C>1/2, on the other hand, there are
8
two open trajectories, due to the symmetry. For a closed orbit, the amplitude U of the
corresponding solution is related to C by . Moreover, the period L is found to be
0
1/ 2
2

2 4
0 0
/ 2 C U U =
1/ 2
4
/ 2 u du

(
+


s
ds =
2 2
2 2
(1 / 2
(1 / 2)
x x
x x
(
(

0
1
2 2 4 2 2 4
0 0 0 0
0 0
( ) 4 / 2 4 1 (1 ) /
U
L U U u U s U s ds ( = =

.
When , we know that
0
0 U =
1
2
0
(0) 4 1/ 1 2 L s = =

. d
In the limit , we have
0
1 U
1
2
0
(1) 4 2 /(1 ) L s =

.
This is indeed an elliptic function, and it may be found that L increases monotonously
from 2 to , as U varies from 0 to 1. In particular, when U
0 0
1 = , there are two
trajectories, both with infinite period that means not periodic. The upper trajectory starts from
the saddle U=-1, and ends up at the other saddle U=1. The lower one is symmetric with respect to
the u axis. Each of them is called a heteroclinic orbit.
Now we illustrate a homoclinic orbit by the following example.
2 2
2 2
' (1 )
' (1 )
x y y x y
y x x y y

= +

(*)
) ,
.


Figure 8 Homoclinic orbits.
It is observed that (0,0) is a saddle; (1,0) and (-1,0) are unstable foci. Furthermore, it can
be checked that satisfies the equation. As a result, there are two trajectories,
each at one side of the y axis, that starts from the saddle and comes back. The period is infinity.
This is a homoclinic orbit.
2 2 2
(1 / 2) y x x =
There is another important type of closed orbits in the study of nonlinear second order
ODE system. It is a limit cycle. Historically, it had been one of the central topics of qualitative
theory for ODE systems. A limit cycle is a closed orbit that each trajectory nearby either goes
towards it, or leaves away. It is, in general, an extremely hard to detect the existence and number
of limit cycle(s). In certain cases, certain theorems guarantee the existence of such a closed orbit.
Theorem (Poincare-Bendixon) Assume that B is a ring-shaped domain between two
simple closed curves and . There is no critical point in this domain. Moreover, any
trajectory that intersects with these two curves enters B. Then there exist at least one limit cycle in
this domain. The conclusion holds if the inner boundary curve shrinks into an unstable node.
1
L
2
L
2
L
9

Figure 9 Poincare-Bendixon theorem.

We make a remark on verifying that a trajectory enters the domain. As we do not have an
explicit expression for the solution in general, we actually check the vector field at the boundary
curves. If the vector f(x,y) points toward the domain, with a non-zero angle, it is possible to show
that the trajectory enters the domain. This is a useful tool for applying the theorem.
Now let us consider the following dynamical system.

2 2
2 2
' (
' (
x y x x y
y x y x y
= +

= +

1),
1).
1),
2, 2.
Under polar coordinates, it reads

2
' (
' 1.
r r r

Take two curves and Then it is easy to check that


2
1
: L r =
2
2
: 1/ L r =
1
' 2
L
r = < 0,
whereas
2
0.
L
r > ' 1/ 2 2 = This verifies the assumptions of the Poincare-Bendixon theorem,
and a limit cycle exists. In fact, this limit cycle is
1. r =
In fact, all trajectories wind towards this limit cycle.

Now we briefly sketch the proof for the existence of limit cycle in the van der Pol
equation ( 0 > )
2
'' ( 1) ' 0. x x x x + + =
This equation plays a special role in the study of electronics. As a matter of fact, the limit
cycle discovered here corresponds to an electronic oscillator to generate a signal (wave) with a
special frequency. It may be recast into an ODE system as follows.

2
' ,
' (1
x y
y x x
=

= +

) . y
.
The only critical point (0,0) is an unstable node, as the corresponding eigenvalues solve
To show the existence of a limit cycle, it suffices to find a outer boundary curve,
along which all the trajectories enter the domain bounded by this curve.
2
1 0 + =
We select a point D( ,
D D
x y ) on the left branch of the curve
2
(1 ) 0 x x y + = . Then
we draw the trajectory across this point of the system
' ,
' .
x y
y x y
=

= +


When we take
D
x close enough to 1 , this trajectory does not intersect with the
aforementioned curve on the right of D. It intersects with the y-axis at a point E( 0, ).
E
y
10
On the other hand, starting from a point A( 0,
A
y ) on the lower y-axis with ,
we may draw a trajectory of the system
A E
y y <
2
' ,
' (1
x y
y x
=

) . y

It cut the line at a point B( 1 x = 1,
B
y ). In fact, this system can be solved
explicitly to give
3
( / y x x 3)
A
y = . That is, we know 2 / 3
B A
y y = + .
From B, we draw a circle centered at the origin, with radius
2
1
B
y +
C
, and intersect with
the left branch of the curve at a point C(
2
(1 ) 0 x x y + = ,
C
x y ).
Now we make two remarks before verifying the conditions of the theorem. First, it is easy
to show, that if we take D close enough to the line 1 x = , and take A far away from (0, )
E
x ,
we may easily make sure that C lies on the left of D. Secondly, it is observed that the system
pertains a symmetry. This allows us to make a set of mirrored curves, and only to verify the
conditions on one side.

Figure 10 Limit cycle in van der Pol equation.

Now we check for each section for the left half of the outer boundary.
First, for the part CD, as each trajectory satisfies ' 0, ' x y y 0 = > = , it enters the
domain.
Secondly, for the part DE, again we have ' x y 0 = > . Furthermore, as it holds that
2
(1 ) x x y x y + < + for , we know that the slope of any trajectory of van der Pol
equation is smaller than that of the one that governs DE. Noticing that DE is a monotone curve
going from lower-left to upper-right, each trajectory enters the domain.
0 y >
Thirdly, for the part AB, which is expressed explicitly by
3
( / 3)
A
y x x y =
2
, it is
easily found that , and ' x y = < 0
2
(1 ) (1 ) x x y x + > y . So, again each trajectory
enters the domain.
Finally, for the part BC, we may construct a function, which we shall discuss in more
detail in the next section, . Along each trajectory, it is found that
2
( , ) V x y x y = +
2
2 2
2 ' 2 ' 2 (1 )
dV
xx yy x y
dt
= + = .
Therefore, when , which is of course true for the part BC, a trajectory has the
direction of decreasing V. This means that it enters the domain.
2
1 x >
This ends the proof.

11
Assignments

1, Show that (0,0) is a saddle; (1,0) and (-1,0) are unstable foci in equation (*).

2, Find a curve around the origin to be the inner boundary for the van der Pol equation, from
which each trajectory points towards the domain enclosed by this curve and the outer boundary we
constructed in this section.
12
3 Stability and Lyapunov function

Stability is one of the key issues in the qualitative theory. In fact, two fundamental papers on
qualitative theory for ODE systems, i.e., Sur les courbes definies par une equation differetielle by
Henri Poincare (1882,1886), and the dissertation by Lyapunov, both discussed mainly the
stabilities. Stability theory actually originated from celestial mechanics stability of the galaxy, or
solar system.
To start with, we need clarify the notion of limit sets. When we talk about stability, we mean
stability of limit sets.
We consider an ODE system
' ( ), x f x = .
n
xR
Limit set concerns with the limit of a trajectory. More specifically, we consider a trajectory
0 0
( ; , ) x t t x =
0 0
lim| ( ; ,
n
n
t t x
that issues from a point . A point p is called an -limit point of this
trajectory, if there exist a sequence t , such that correspondingly
0 0
( , ) t x
n
+
) | 0 p

= . The set of all limit points forms the -limit set of this trajectory.
Similarly we may define -limit point and -limit set of a trajectory, just taking .
Typically in second order ODE system, a limit set is either a critical point, or a closed trajectory
limit cycle, heteroclinic orbit, homoclinic orbit, or a closed orbit around a center. In this section,
we shall confine ourselves to the stability of a critical point.
n
t
To motivate our analysis, we consider two different situations shown in the following
figure. In both cases, the ball is assumed to stay still. However, the first one is unstable, in the
sense that any small perturbation will drive it away from the current position; whereas the second
one is stable, in the sense that the ball will remain close to this equilibrium position under small
perturbation.

Figure 11 Stability.

There are different stabilities defined historically. We list some of these
definitions. For simplicity, we consider a critical point (0,0) for second order ODE
system.
A critical point (0,0) is stable, if 0, 0, > > such that | (
0 0
; , ) | t t x < , for
0 0
, | | t t x > < .
A critical point (0,0) is asymptotically stable, if it is stable and 0, > such that
0 0
lim ( ; , ) 0
t
t t x
+
= , for
0
| | x < .
A critical point (0,0) is exponentially stable, if 0, > for 0, 0, > > such
that , for
0
(
0 0
| ( , , ) |
t t
t t x e


)
0
| | x < .
A critical point (0,0) is globally (asymptotically, exponentially) stable, if aforementioned
statements hold for arbitrary
0
x .

Indeed, what we have here is uniformly stable, uniformly asymptotically
stable, and uniformly exponentially stable, where uniform means the choice of
independent of the time .
0
t
To determine the stability of a critical point, there are two methods, namely,
the Lyapunov first and second method. In the first method, one first examines the
13
stability of the linearized problem (linear stability), then shows the stability for the
nonlinear problem. In fact, for the linearized problem
0
( ) y f x = - y ,
the stability may be determined from the sign of the eigenvalues. Alternatively,
stability is implied by the fact that
0
( ) f x is negative-definite. Furthermore, the
following theorem ensures the nonlinear stability.
Theorem If all eigenvalues of
0
( ) f x have negative real part, then the critical point
0
x is asymptotically stable for the nonlinear ODE system.
This may be used to show the stability of a critical point for a linear system in the
first section.
The second Lyapunov method, or sometimes called the direct Lyapunov method,
requires a Lyapunov function. We need to construct according to the specific ODE
system. Without loss of generality, we assume that the critical point is .
Theorem If there exists a Lyapunov function , which is positive definite, and
non-increasing along each trajectory
0
0 x =
( ) V x
( ) x t , i.e.
for ; (0) 0; V = ( ) 0 V x > 0 x

( )
( ) 0
x x
x t
dV dx
V V f x
dt dt
= = - - in a neighborhood (0, ) U

.
Then the critical point 0 is (uniformly) stable. If
( )
0
x t
dV
dt
< in a neighborhood U(0, )

,
then it is asymptotically stable. If the inequality holds for , and V x is radially
unbounded, i.e. , then we have global stability.
n
R x ( )
| |
lim ( )
x
V x

+
We also remark that a sufficient condition for global exponential stability is that there
exist positive constants , such that
1 2 3
, , c c c
, and
2 2
1 2
| | ( ) | | c x v x c x
2
3
( )
| |
x t
dV
c x
dt
.
This may be illustrated through Figure 12.
We apply this theorem to the following system for the motion of a particle with mass m in
a conservative force field . The equation reads, according to Newtons second law,
'' mx = .
Let velocity be , and rewrite the equation as ' v x =

' ,
' /
x v
v m
=

.
A critical point must be stationary ( 0 v = ), and extremum of . Without loss of
generality, let us assume this extremum locates at 0. Now consider the total
energy . We construct a function as
2
/ 2 ( ) E mv x = +
.
2
( , ) / 2 ( ) (0) V x v mv x = +
It turns out that V , which is just the conservation of energy. So, if we have ' 0 =
( ) 0 x > for certain neighborhood (0, ) U

, then V is a Lyapunov function, and the


equilibrium is stable.
This is exactly the Lagrange theorem, which says that an equilibrium position is stable for a
conservative force field, if the potential ( ) x reaches its local minimum here.

14

Figure 12 Lyapunov function.

Now we describe one of the instability results for a critical point, named after Lyapunov.
Theorem The critical point x=0 is unstable if there exist a function V(x) with V(0)=0, for which
( )
0
x t
dV
dt
, and in any neighborhood of x=0, there exists a point
0
x with .
0
( ) V x < 0

Assignment

1, A critical point (0,0) is called attractive, if and
0
( ) 0, t > such that
0 0
lim ( ; , ) 0
t
t t x
+
= ,
for
0
| | x < . Therefore, it is asymptotically stable if it is stable and attractive. Show that
being attractive does not imply stability through analyzing a first order equation
2
' x x = .

2, For a linear ODE system , if
0
( ) y f x = - y
0
( ) f x is negative-definite, find a Lyapunov
function.

15

4 Bifurcation theory: an introduction

In applications, a system usually has certain tunable parameters. For instance, there may be a
variable resistance or capacitor in an electronic circuit. In a beam, the applied force on the top
may change as well. Consequently, as observed in practice, the system may undergo certain
structural change. To include this in an ODE system, we consider the following parametric
system, with R a parameter,
' ( , ) x f x = .
For each fixed parameter, the analysis in the previous sections applies. In this section, the
main consideration is how the parameter infects the geometric structure of the system.
As what has been done before, we first consider the critical points. That is, the roots to an
algebraic system ( , ) 0 f x = . To simplify analysis, we shall basically study scalar equations, that
is, . xR
Take an equation
2
' x x = ,
there is no critical point for 0 < , and two critical points for 0 > . As we have learned,
critical points are essential in the structure of the trajectories. Therefore, we expect a drastic
difference when crosses 0. In fact, for 0 < , because
2
0 x < , x keeps decreasing
towards , wherever it starts with initially. On the other hand, for 0 = , the exact
solution is

0
0
1
x
x
x t
=
+
,
where
0
(0) x x = . So, if , x tends to the critical point 0; and if
0
0 x >
0
0 x < , x tends to .
If

further increases to 0 > , then x tends to if


0
x > ; and tends to
otherwise.

Here we observe a bifurcation phenomenon, when the controlling parameter crosses the
bifurcation point 0 = . To illustrate this, we draw a bifurcation diagram in Figure 13.

Figure 13 Saddle-node bifurcation.

In such a diagram, we plot x versus . In multi-dimensional case, i.e. , we need to
select certain aspect of x for such an illustration, usually an entry
n
xR
i
x , or the norm (length)
2
|| ||
i
x x =

. At a fixed parameter 0 > , the stable critical point lies in the solid line (stable
branch), whereas the unstable one in the dashed line (unstable branch). The stability is also
observed from the direction of the trajectories.
Now we consider an example of transcritical bifurcation. In the equation
16

2
' x x x = ,
we notice that there are always two critical points,
1 2
0, x x = = . However, depending on the
sign of , the relative position changes. More precisely, for 0 < ,
1
x is stable, and
2
x is
unstable. For 0 > ,
1
x becomes unstable, and
2
x is stable. Here we say, an exchange of
stability occurs. This is clearer from Figure 14.

Figure 14 Bifurcation diagram: transcritical bifurcation.

Next, consider the supercritical bifurcation in

3
' x x x = .
When 0 , the only critical point is
0
0 x = , which is stable. When 0 > , two more
critical points appear as
1 2
, x x = = . These two critical points are both stable, and the
original one becomes unstable. See Figure 15.
0
x = 0

Figure 15 Bifurcation diagram: supercritical bifurcation.

The examples we have discussed so far are stationary bifurcation, namely, they concern only
with critical points. There exist other type of bifurcations, which may include closed orbits. One
most important bifurcation of this type is Hopf bifurcation. In such a bifurcation, there emerges a
periodic orbit when the parameter goes beyond certain threshold. We demonstrate this through a
second ODE system

2 2
2 2
' (
' (
),
).
x y x x y
y x y x y

= +

= +


The only critical point is
0 0
( , ) (0, 0) x y = . The Jacobian matrix is found to be
1
1

| |
|
\ .
,
17
which has eigenvalues i . Therefore, when 0 < , it is a stable focus; and when 0 > , it is
an unstable focus. In fact, this system may be better understood in polar coordinates
( , ) ( cos , s x y in ) =
' (
' 1

2
.
. It is readily shown that
), =

0 <

( ,
2
'' ' '/ x x x = +
2
' , x y
y x
=

x xx +
x +

0
0.429505 =

Using the results of supercritical bifurcation ( 0 ), we know that when ,
decreases towards 0. Therefore, each trajectory winds anti-clockwise ( ' 0 > ) towards 0 = .
On the other hand, when 0 > , because is stable, each trajectory winds anti-clock
wise towards this periodic orbit (with period 2 ). Thus this is a limit cycle. This bifurcation may
be depicted either in a 3-D plot, or a 2-D plot in ) plane. See Figure 16 and Figure 17.

Figure 16 Hopf bifurcation. Figure 17 Hopf bifurcation.

One distinctive feature in Hopf bifurcation lies in the fact that the eigenvalues cross the real
axis.
In Hopf bifurcation, we observe that a periodic orbit (limit cycle) emerges out of a critical
point. There are also cases where a periodic orbit becomes a homoclinic orbit, typically when it
touches a saddle point. This is called a homoclinic bifurcation, and may be demonstrated by the
following equation.
2 .
Or, in terms of ODE system,

' / y xy + 2.
The critical points are (0,0), (1,0), and (-1,0). It is numerically found that when reaches
, a homoclinic orbit is formed. It may be checked, that at this parameter, (0,0)
and (-1,0) are saddle points, and (1,0) is a stable focus. The scenario of this bifurcation is
schematically plotted in Figure 18. From left to right,
0 0
, ,
0
> = < .

Figure 18 Homoclinic bifurcation.


Assignments

18
1, Discuss bifurcation in the following system

2
' , x x = +
and draw the bifurcation diagram.

2, Discuss the subcritical bifurcation in

3
' x x x = + .
Draw the bifurcation diagram. Due to the shape in subcritical and supercritical bifurcations, they
are called pitchfork bifurcations.

3, For the Fitzhugh-Nagumo equations:

3
' 3( / 3
' ( 0.7 0.8 )
x x y x
y x y
),
/ 3,
= + +

= +


compute the critical points, and study their stabilities. Compute when the Hopf bifurcation occurs.
Draw a bifurcation diagram of x versus .

19
Chapter 3 Discontinuity: hyperbolic conservation laws
1 Shock formation in Burgers equation, methods of characteristics

In most applications, hyperbolic conservation laws are obtained when the diffusion terms are
neglected from the full systems, which are typically parabolic. In a hyperbolic system, wave
phenomena are more distinctive, and usually easier to analyze. As a matter of fact, waves are the
central topic in such systems.
The simplest model is a linear advection equation, which describes a wave propagating
towards the right at a constant speed a.
. 0
t x
u au + =
If , a direct verification shows that the solution is 0 a >
0
( , ) ( ) u x t u x at = .
Alternatively, we consider the total derivative along a line
0
x at x = + as follows.

0
0
t x
x at x
du u u dx
u au
dt t x dt
= +

= + = + =

- .
So u keeps unchanged along such a line. These lines are called characteritic lines. We remark that
even when there is a jump in initial profile, the formular
0
( , ) ( ) u x t u x at = still makes sense.
It can be regarded as a weak solution, namely, it satisfies an integral form of the equation
0
t x
V
u au dxdt + =

where V is any controlling volume.



Figure 1 Characteristic lines and solution in linear advection.

We may extend this characteristic approach to a general scalar equation. Consider a fully
nonlinear equation with source terms,
. ( ; , ) ( ; , )
t x
u a u x t u f u x t + =
In fact, along a curve
0
: ( ; , ), (0)
dx
a u x t x x
dt
= = ,
it holds

0 0
( ; , ), ( , 0) ( )
t x
du u u dx
u au f u x t u x u x
dt t x dt


= + = + = =

-
0
.
Under certain fairly general restrictions on and , the
implicit function theorem provides a local existence result, i.e. for
( ; , ), ( ; , ) a u x t f u x t
t
0
( ) u x
, there exist
and
( , ) u x t
( ) x t solving the ODE system. Moreover, the slope of the characteristic curve is
bounded, leading to a finite propagation speed.
( ; a u x t , )
In the previous investigations, we reduce partial differential equations to ODE systems along
characteristic curves. This allows us to solve the Cauchy problem by marching with a parametric
20
expression, where the initial position
0
x serves as the parameter. Along each curve, the value of
u is determined by its initial data only. It is independent of initial data at other points. If we are
lucky enough, (which is not true in general, see later discussions), the characteristic curves issued
from all may cover the whole domain
0
x R ( , ) x t
+
R R , thereby the original PDE
problem is solved. We note that this also applies to IBVP (initial boundary value problem).
In a general system of hyperbolic partial differential equations, this simple approach of
characteristic curve may fail, e.g. in a p-system in gas dynamics. Nevertheless, this approach is a
basic tool for constructing elementary waves there; and implies a finite propagating speed of
waves.
Before starting further discussion on wave phenomena, we make a few definitions. A first
order linear PDE takes the form of

1
0
n
i
i i
u
a bu
x
=

+ =

,
where . Superposition of solutions holds. If and are solutions, so is ,
i
a bR
1
( ) ( u x
1
( ) u x
2
( ) u x
2
) u x + , for any , R. For nonlinear PDEs, general theory is rare. For better
understanding, we need further classify nonlinear PDEs.
A first order quasilinear PDE reads

1
( , ) ( , ) 0
n
i
i i
u
a u x b u x
x
=

+ =

.
It is linear with respect to each partial derivative /
i
u x .
A system of quasilinear hyperbolic conservation laws reads
, ( ( )) 0
t x
U F U + =
with , and is a diagonalizable
n
U R ( ) ( ) A U F U = n n matrix.
There are numerous quasilinear hyperbolic PDE systems arising from sciences and
engineering. To name a few, we have the following models.
Traffic flow: ( , u are density and velocity, resp.)
( ) ( )
max
max
0, 1
t
x
u u u

| |
+ = =
|
\ .
.
Two-phase flow (Buckley-Leverett equation): (u is the saturation of water)
2
2 2
( ) 0, ( )
(1 )
t x
u
u f u f u
u a u
+ = =
+
.
Euler equations for polytropic flow: ( , , u p are density, velocity and pressure, resp.)
2
( ) 0,
( ) , (1 3).
( ) ( ) 0,
t x
t x
u
p k
u u p




+ =
= < <

+ + =


Shallow water: (h,v are the height and velocity of water, resp.)
2
( ) 0
( / 2 ) 0
t x
t x
h hv
v v gh
+ =

+ + =

,
.
0


In quasilinear hyperbolic PDE systems, the most distinct phenomenon is the appearance of
shock waves. This is a genuine nonlinear phenomenon, and does not occur in linear equations.
We illustrate this through a simple example of the inviscid Burgers equation
.
2
( / 2)
t x
u u + =
Under the assumption of smoothness (regularity), it is equivalent to
. 0
t x
u uu + =
21
According to previous discussions, u keeps constant along ( , )
dx
u x t
dt
= . This characteristic
curve is then a straight line. Furthermore, the solution may be written as
0 0 0 0 0
( ( ) , ) ( u x u x t t u x + = ).
A question is: for each ( , ) x t
+
R R , does there exist exactly one characteristic line
passing through this point?
Unfortunately the answer is NO in general, even if the initial profile is smooth. We further
confine ourselves to monotone profiles. Then there are two cases, either is monotone
decreasing, or monotone increasing.
0
( ) u x

Case 1: Increasing: We observe that for each ( , ) x t
+
R R , there is one and only one
characteristic line passing through this point. Therefore the global unique solution is guaranteed.
In fact, when time evolves, the wave profile becomes flatter and flatter.


Figure 2 Burgers equation with increasing initial data.


Case 2: Decreasing: It is just the opposite. The wave profile becomes sharper and sharper. At
certain time, there must be a blow-up of the solution, that is, the space derivative
becomes infinite at certain space position. If we keep following the characteristic lines, regardless
of the fact that the original PDE becomes meaningless, then the wave profile will flip to the other
side. For instance, the characteristic line through in Figure 3 will go across the point
where .
/ u x
( *, 0) x
0
0 u =
What should we do after the formation of such blow-up, or infinite space derivative? As we
have mentioned, the equation does not hold at and after this time of blow-up in classical sense. We
should refer to the physical problem, or more complete system reaction-diffusion system. In
certain situations, for instance, in water waves, there might be some time when multi-values shape
is acceptable but not too long time, as the gravity will eventually pull the water drops down. In
most other cases, a discontinuity forms and moves at certain velocity. Along this discontinuity
*( ) x x t = , the equation does not hold in the classical sense ( u x !). Away from
that, the equation still holds in the classical sense.
( *( ), )
x
t t =
22

Figure 3 Burgers equation with decreasing initial data.

We shall give a rescue to the crisis occurred due to the monotone decreasing initial data in the
next section. Here let us draw a partial conclusion, that is, for a quasilinear hyperbolic PDE
system, given a generic initial profile, discontinuity develops at certain time, regardless of the
smoothness of the initial profile. Therefore, we have to consider the system with discontinuous
initial data in general up to a translation of time. It is of great importance that Riemann problem
serves as a basic building block for such system. By a Riemann problem, we mean a discontinuous
Cauchy problem with initial data

0
, 0
( )
, 0
u x
u x
u x

+
<
=

>

,
.
One of the main reasons why it works lies in the finite propagating speed of wave in
hyperbolic PDE systems. We shall explore these interesting facts in the coming sections.


Assignments

1, What is the solution to
0
t x
u xu + = ,
with initial data ? What about
0
( , 0) ( ) u x u x = 0
t x
u xu = ?

2, Given a smooth monotone decreasing profile
0
( , 0) ( ) u x u x = , can you find the smallest time
when there is a * t * x , such that ( *, *)
u
x t
x

becomes infinity?
23
2 Elementary waves in Burgers' equation

For inviscid Burgers equation
0
t x
u uu + = ,
we consider a Riemann problem

>
<
=
+

. 0 ,
, 0 ,
) 0 , (
x u
x u
x u
We expect this discontinuity moving at certain speed. As we have learned
from the discussion in an excitable system, we need to include the diffusive term,
namely to consider viscous Burgers equation
t x
u uu u
xx
+ =
The discontinuity in this setting becomes to a traveling wave, with transition layer of width
proportional to

for certain 0 > . Therefore, we are looking for a heteroclinic orbit that
connects with u u
+
.
The traveling wave equation reads, with x ct = ,
cu uu u + = .
This can be integrated once to give
2
0
/ 2 u cu u C = + + .
Because and u u
+
should be critical points, we have
2
0
/ 2 0 cu u C

+ + = , .
2
0
/ 2 0 cu u C
+ +
+ + =
From these, we determine
2
u u
c
+
+
= ,
0
2
u u
C
+
= .
Furthermore noticing that is convex, so the source term is negative for
between and . Therefore, a heteroclinic orbit exists if and only if
2
0
/ 2 cu u C + + u
u

u
+
u u
+
> .
Therefore, we obtain a discontinuity (shock) if and only if u u
+
> , at a propagation
speed . We observe that this speed is the slope of the secant line between
and . More over, the condition
( ) c u u
+
= +
2
/ 2)

( , u u
+ +
/ 2
( , u u
2
/ 2) u u
+
>
q
means that the characteristic
lines from both sides point towards the shock. This is also true for shock in general scalar
conservation law. We introduce a notation for jump of an entity as
| |
q q q
+
= , where q


are the values across the discontinuity. Naturally, the shock strength is represented by
| |
u . And
our result says:
| | | |
2
/ 2 0 c u u + = .
Later on we may find that this is true for general conservation system, and is called as the
Rankine-Hugoniot condition.


Figure 4 Shock in Burgers equation.
On the other hand, if , there is no such heteroclinic orbit. Instead, since the
+
< u u
24
characteristic lines issued from only cover the domain 0 x < t u x

< ; and those issued from
only cover the domain , there is a region 0 > x t
+
u x > t u u
+
x t

< < where there is no


definition for the solution. There are two ways that will reach the same result. The first way is to
find self-similar solution ) ( ) , ( u = t x u , t x / = . The equation then reads
0
0 = u .
t = = u
, t <
, u
, if
if
, if
u x
u
u x
, if
, if
, if
u x
u
u x
,
.
u
t x
u

+ +
<
= <
>
x

t
u
t
+
,
.
u
u

+
<
( ),
)
( ).
u t
x
u t


+ +
<
= <
>
if
if (
if
u x
u
u x
,
),
,
, t

< +
+ <
> +

+ +
, if
if
, if
u x
a
u x
<
= <
>
,
.
a
x b
b
,
u
2
=

t
u u
t
u x

Or, equivalently,
0 ) ( = u u
There are two choices. One is thus const u = , but it is of no interest here for
solving a discontinuous problem. The second choice is = u . This gives a fan that on each line
issued from the origin x , we have . We remark here that at the origin, is
not defined. A complete understanding would rely on the definition of a weak solution. Then it is
natural to define the solution to Riemann problem as
) 0 , 0 ( u
( , ) / , u x t x t


This is discontinuous not only at the origin, but also at the borders , where the
partial derivatives do not exist.
t u x

=
Another approach is to imagine a small perturbation of the initial data, e.g.
( , 0) / u x x


Then we use the characteristic method to get a solution
( , ) /( ( ) u x t x t u t
+

= + < +


Taking the limit 0 , we obtain the same solution as mentioned above. It worth mentioning
that with different monotone perturbations for the initial data, we obtain the same solution in the
limit 0 . This justifies partially our solution with centered rarefaction wave.



Figure 5 Rarefaction wave in Burgers equation.
With these discussions, now we are ready to investigate more general initial data. The
simplest case is a wave interaction problem. That is, consider a piecewise constant data
( , 0) ,
m
u x u

<
Many different cases may arise, according to the relative position of .
+
u u
m
, ,
25
For instance, if u , then initially there are two shocks generated, at and ,
respectively. The propagation speed of the first shock is
+
> > u u
m
a b
1
S 2 / ) (
1 m
u u s + =

x
, and that of the
second shock is . In fact, the first shock is located at , and
the second at
2
(
2
= u s
m
t s b
2
S
x
2 / )
+
+ u t a
1
= s +
+ = .
Since , the first shock wave will catch up with the second one at time
2 1
s s >
+

=
u u
a b
s s
a b
t
) ( 2
2 1
*

And position
1 2
1 2 *
s s
b s a s
x

=

Figure 6 Wave interaction: shock collision.

Before this collision time, the solution is piecewise constant:
1
1 2
2
, if ,
( , ) , if ,
, if .
m
u x a s t
u x t u a s t x b s t
u x b s t

+
< +

= + < <

> +

+
+

At the collision time, we solve a new Riemann problem

>
<
=
+

. ,
, ,
) , (
*
*
*
x x u
x x u
t x u
Because , there comes out a new shock , with shock
speed . We observe that in the
u u

>
2 / )
+
+ u
) ( :
* * * *
t t s x x S + =
) (u f (
*

= u s u plane, there is a triangle with


vertices corresponding to these three shocks: .
*
S
2 1
, , S S

Figure 7 Rarefaction waves: no interaction.
Our second case concerns with
+
< < u u u
m
1
R
. We have initially two rarefaction waves
and . Since the last characteristic in has the same slope as the first characteristic in
, there is no interaction. That is the solution is simply a combination of these two rarefaction
wave solutions. See Figure 7.
1
R
2
R
2
R
Now we consider a case . Initially, there is a shock connection u
and ; and a centered rarefaction wave connection and . Because that the shock speed
is faster than the first characteristic line, i.e.
m
u u u > >
+ 1
S

m
u
m
u
+
u
26
m
m
u
u u
>
+

2
,
the shock will hit the rarefaction fan at time
m
u u
a b
t

) ( 2
*
.
Afterwards, the shock bites gradually the rarefaction. In this process, the shock strength is
getting weaker. Eventually, the shock disappears, leaving a rarefaction wave connecting with
.
| | u

u
+
u

Figure 8 Wave interaction: shock overtakes rarefaction wave partially.

Assignments

1, Show that for a scalar conservation law
, ( ( )) 0
t x
u f u + =
if the underlining reaction-diffusion system is
( )
t x
u f u u
x
+ = ,
and ( ) f u is convex (i.e. ''( ) f u has definite sign), then a shock connecting two states and
must propagate at speed
u

u
+
( ) ( ) f u f u
u u
c
+
+

. Is there any further condition on and


?
u

u
+

2, Find self-similar solution to general conservation law
, ( ( )) 0
t x
u f u + =
under the condition that ( ) f u is convex.

3 Noticing that inviscid Burgers equation may be rewritten as
,
2 3
( ) (2 / 3) 0
t x
u u + =
what are the corresponding shocks? How about rarefaction waves?

4, Can you find the expression of in the third case, namely the shock partially overtakes
the rarefaction wave?
( , ) u x t

5, Analyze one more case for wave interaction in Burgers equation.
27
3 Elementary waves in a polytropic gas


,
.
For a polytropic gas, the conservation of mass and momentum gives the
following conservation laws:

2
( ) 0
( ) ( ( )) 0
t x
t x
u
u u p


+ =

+ + =

Here , u are the density and velocity, respectively; and the pressure p

= . We take
variables ( , ) V where V u = is the momentum. According to our definition of hyperbolic
system, we linearize the system around a ground state
0 0
( ,V ) ,
( )
2
0 0
0 2
0 0
0 1
. . .
2
'
t x
h o t
V V
p V V


(
( (
(
+ =
( (
(


(


Then the eigenvalues are found as, with
0 0
V u
0
= ,
0 0
'( ) u p

= .
With our pressure law,
( 1) / 2
0
'( ) p



= . So, there are two distinct eigenvalues,
therefore the system is hyperbolic.
It turns out that the Lagrangian formulation suits better mathematical analysis. In fact, we
define

,
( , ) .
t
x t dx



By virtue of the equation for conservation of mass, actually denotes the particle. In fact, in
one space dimension, no particle can go across a neighbor particle. Therefore, the total mass at
its left side is an index for this particle. For this transform, we have,

( , ) ,
.
t
x t dx u
t
x


= + =


The conservation of mass becomes
.
2
(1/ ) ( / ) 0
t x
V V V V

( + = + = =

Meanwhile, noticing that the conservation of momentum leads to
/ 0
t x x
u uu p + + = ,
we have
/ 0
t x x
u uu p u Vu u u p u p

+ + = + + = + = .
So, the p-system in Lagrangian coordinates reads

0,
0.
v u
u p


=

+ =

Here we define specific volume 1/ v = . Again we consider linearization around a ground state
as follows
0 0
( , ) v u

0
0 1
. . .
'( ) 0
v v
h o t
p v u u

( ( (
+ =
( ( (

We remark here that the derivative for pressure is taken with respect to . That is, v
28
( 1)
0 0
'( ) p v v

+
= . The eigenvalues are
'( p v

=
( ( U F

r u
s u

= +

u r = +

r r

+ =
s s

=
1
:
d
d

=
2
:
d
d

=
max

0 0
( , )

( 1) / 2
0 0
) ; ( ) v v


+
= = .
0

t
Again, we have two distinct eigenvalues.
For hyperbolic conservation laws, we may locally follow the strategy in linear case to
diagonalize the system in forms of
)) 0 . . .
t x t x
U V V h o + = ==> + =
where is a diagonal matrix. But the high order terms can not be discarded, particularly when
we investigate shocks, which is not a small perturbation from any ground state.
It is a very special thing for that we may found Riemann variables in this two space
dimensional problem. In fact, we define
(1 ) / 2
(1 ) / 2
2
,
1
2
.
1
dv u v
dv u v

= +


Conversely, we may recover from ( , ) v u
( 1) / 2
( ) / 2,
( 1)( ) / 4
s
v r s



(
=


.
0
By a straight forward calculation, we find
( ) ( ) '( ) u v u v u p v v

+ = + = .
Similarly, we have
( ) ( ) '( ) u v u v u p v v

+ + = + = . 0
Therefore, we have a class of right-going characteristic curve
,
along which r keeps constant. Along a class of left-going characteristic curve
,
s keeps constant. The coupling of hides in the expression of v and therefore ( , ) r s .
In physical problems, typically we have certain upper-bound for density, this gives an
upper-bound for . That is, we have finite speed of propagation in p-system. On the other
hand, we also have the notions of domain of dependence and range of influence. In fact, at each
point , the physical quantities depend at most
0 max 0 0 max 0
[ , ] = D .
Meanwhile, the quantities here influence at most the domain
{ }
0 max 0
( , ) = R .

Figure 9 Domain of dependence and range of influence.
29
Now we describe elementary waves in p-system.
For a shock wave, we may consider a viscous fluid

0,
.
v u
u p u

+ =



A traveling wave with speed s solves a traveling wave equation with s = ,

' ' 0,
' ' '
sv u
su p u
=

+ =

'.
They can be readily integrated once to give

1
2
,
'.
sv u C
su p C u
=

+ + =


The Riemann data must be critical points, and we obtain Rankine-Hugoniot conditions
(with )
( , ) u v

q
+
[ ] q q =

[ ] [ ] 0,
[ ] [ ] 0.
s v u
s u p
=

+ =

Eliminating the term , we obtain [ ] u



[ ]
[ ]
p
s
v

= .
We remark that in the limit of continuous case , this gives the characteristic value [ ] 0 v

.
Now we need to determine a condition for existence of such a heteroclinic orbit. We may
easily obtain
2
3 3
[ ]
'
[ ]
p
sv s v p C v p C
v
= + = + .
Because of the convexity of ( ) p v , we know that for between v v

and v , the right


hand side is positive.
+
We first consider . Then a heteroclinic orbit exists if and only if . This
inequality in fact is equivalent to
s
+
v v

<
+
) v ( ) ( v
+ + +
> . Therefore we obtain a right-going ( forward)
shock wave.
Across : S s
+
=
+
, we have

( , ), if ,
( , )
( , ), if ,
v u s
v u
v u s


+
+ + +
<
=

>


with
, [ ] /[ ], [ ] [ v v s p v u p v
+ +
< = = ][ ] .

Figure 11 Forward shock wave.
For the case of , the condition for heteroclinic orbit is s

v v
+
> , which is equivalent to
30
say ( ) ( ) v v
+
>
( ,
( , )
( ,
v u
v u
v u
. A left-going (backward) shock wave solution is

+ +

v v s
+
> =
` `
` '( )
v u
u p v v

1
det
'( ) p v

(
(


` ` 0 v u =
( , v u
` ` ` 0 s v u = + =
s
v
( )
2
( ) / v

=
ds =

), if ,
), if ,
s
s

<
>

with
, [ ] /[ ], [ ] [ p v u p v = . ][ ]

Figure 12 Backward shock wave.

In these two cases, the entropy conditions may be uniformly stated in the Lax entropy
condition as characteristic lines 3-in-1-out.
Now we look for rarefaction waves. Taking the self-similar variable / = , we have

0
` 0
=
=
,
.
0 This can be viewed as a linear system for variables . Either , which is
discarded, or the coefficient matrix is singular. Therefore,
( `, `) v u ` ` v u = =
.
2
'( ) 0 p v = + =
This means ( ) v = . For the case ( ) v = , we have a forward rarefaction wave R
+
,
which satisfies
.
This leads to, in plane, )
.
This means, keeps constant along this type of rarefaction wave.
Accordingly, in physical plane ( ( , ) plane), from left to right, increases. Therefore,
increases, and decreases. So, and u increase. r
In fact, we may find the explicit expression of such a solution. Since ( ) / v = = , we
know that

1/( 1) +
.
Meanwhile, with 0 du dv + = , we find

( )
(1 ) / 2 (1 ) / 2 (1 ) / 2
0 0
2 2
1 1
u s v u v v




= + = +

0

.
31

Figure 13 Forward rarefaction wave.

For the other case ( ) v = , we have a backward rarefaction wave R

, which satisfies
` ` 0 v u = .
This leads to, in plane, ( , ) v u
` ` ` r v u 0 = + = .
This means, keeps constant along this type of rarefaction wave. By a similar argument we may
find that and increases.
r
v , s u
Again we may find the explicit expression of such a solution. Now ( ) / v = = , we
know that

( )
1/( 1)
2
( ) / v


+
= .
Meanwhile, with 0 dr du dv = = , we find

( )
(1 ) / 2 (1 ) / 2 (1 ) / 2
0 0
2 2
1 1
u r v u v v




= =

0

.

Figure 14 Backward rarefaction wave.




Assignments

1, Find the Rankine-Hugoniot condition and entropy condition of a forward shock wave in a
polytropic gas in Eulerian coordinates, namely,
32


2
( ) 0
( ) ( ( )) 0
t x
t x
u
u u p


+ =
+ + =

,
.
Is this the same as that in Lagrangian formulation?

2, Find forward and backward rarefaction waves in the Eulerian formulation. Also find the
Riemann variables if possible.
33
4 Riemann problem in a polytropic gas

For a polytropic gas in Lagrangian coordinates

0,
0,
v u
u p


=

+ =

we consider a Riemann problem with initial data



( , ), if 0,
( ( , 0), ( , 0))
( , ), if 0.
v u x
v x u x
v u x

+
<
=

>


To solve it, we need to find elementary waves connecting these two states. Since both the
equation and the initial data possess invariance under that transform { } , , so the
solution must depend only on / = . For the left quadrant 0 < , if there is a shock at
0 s

= < , due to the entropy condition, there cannot be a backward rarefaction wave. For the
same reason, there is at most one type of forward wave on the right quadrant. Consequently, to
solve a Riemann problem, we only need to find forward and backward waves. It amounts to find
an intermediate state ( * . A backward wave connects ( , , *) v u ) v u

with this state, and a
forward wave connects it with . ( , ) v u
+ +

Figure 15 Riemann solution.

Now we consider a backward wave that takes ( , ) v u

as left end state, possibly a
shock as well as a rarefaction wave. Then we have,
( )( )
( )
(1 ) / 2 (1 ) / 2
( *) ( ) * if *;
* ( *; )
2
* , if
1
*.
p v p v v v v v
u u f v v
v v v

>

v <


Figure 16 Backward wave in phase plane.
Similarly, we consider a forward wave that takes ( , ) v u
+ +
as left end state. Then we
have,
34
( )( )
( )
(1 ) / 2 (1 ) / 2
( ) ( *) * if * ;
* ( *, )
2
* , if *
1
.
p v p v v v v v
u u f v v
v v v

v
+ + +
+ + +

+ +

<

>


Figure 17 Forward wave in phase plane.

Given the Riemann data, we may eliminate from the above equations, and obtain an
equation solely of ,
* u
* v
( *; ) ( *, ). u u f v v f v v
+ + +
=
This is equivalent to find an intersection point of the curve ( *; ) f v v

with the curve
( *, ) f v v
+ +
) , up to a translation of (u u
+
. As shown in the previous figures, ( *; ) f v v

is
monotone increasing, and ( *; ) f v v
+ +
decreasing. Both are continuous, and indeed in
2
C ( )
+
R ,
even at the point where the wave changes from a shock to a rarefaction wave. Moreover, we
may find the limits
v

(1 ) / 2
* 0 *
(1 ) / 2
* 0 *
2
lim ( *; ) , lim ( *; ) ,
1
2
lim ( *; ) , lim ( *; ) .
1
v v
v v
f v v f v v v
f v v f v v v


+ +

+ + + + +
+ +
= =

= + =


Therefore, these two curves insect if and only if
( )
(1 ) / 2 (1 ) / 2
2
1
u u v v


+ +
< +

.
Given , we may draw two curves, namely u u ( , v u

) ; ) * ( * f v v

= + and
. They separate plane into four parts, as shown in Figure 18.
Another curve
* u u f
+
= + ( *; v v

) ( *, *) v u
(
/ 2 (1
v

+
)
*
(1 ) ) / 2
2
*
1
v

u u takes out a region where no Riemann


solution exists.
= +
For instance, if we have a Riemann problem with ( , ) v u
+ +
in Region I, then we may
draw a curve u u , it intersects the curve u u * ( * f v v
+ + +
= + ; ) ; ) * ( * f v v

= +
( , v u

( *, *) v u
( *; f v v
at a point
. In fact, this is exactly the intermediate state we are looking for. Geometrically,
are the states which may be reached from through a
backward wave; and are the states ( * which may reach ( ,
through a forward wave. Moreover, we observed that the location of lies in the shock
branch of , and rarefaction branch of
( *, *) v u
* u u = + ( *; ) v v

* u u f
+ +
= +
* ( *; ) u u f v v

= +
f

( *, *) v u
( *; v v
+
)
) , *) v u
* u u
) v u
+ +
)
+ + +
= + . As a result,
the Riemann problem is solved by a piecewise continuous solution with a backward shock and a
35
forward rarefaction wave.
Other cases may be discussed in the same fashion. It worth noticing that in the Region V,
no intersection points may be found of the two curves, even when . This corresponds
to a situation of vacuum, for which our governing equations fail to be valid. As through rarefaction
waves, increases from left to right, this increment is bounded by
* v +
v u

and v . The Region V
is the place where this limitation of rarefaction wave occurs.
+



Figure 18 Profiles of Riemann problems.

Assignments

1, Find the second order derivatives of ( *) f v
+
and ( *) f v

at the point v

, respectively.

2, Given , discuss profiles for Riemann problems, depending on the value of . ( , v u
+ +
) ( ) , v u


3, Try a case of interaction of a backward shock wave and a forward shock wave.

36
Chapter 4 Fixed point theory with applications
1 Basic concepts of normed spaces, function space [ , ] C a b

At very elementary level, we have known fixed point theorems. For instance, if a mapping
is a contraction, then there exist one and only one fixed point for this mapping. : f R R
Here are some concepts to be clarified. By a contraction, we mean that [0,1) ,
it holds that , x y R, ( ) ( ) | | f x f y x y < . A contract mapping, in this case, a
contract function, is continuous. By continuous, we mean, as in Calculus, 0, 0 > > , such
that | | x y < , it holds that | ( ) ( ) | f x f y < . A point xR is called a fixed point of
f if ( ) f x x = , i.e. f maps this point to itself. The proof for this type of theorems is standard,
namely, an iteration procedure.
Our interest shall go beyond these simple situations. We intend to apply these tools to
differential equations. For this purpose, we need to go to function spaces, and define mapping
there.
We start with the basic concept of a normed space. Space, similar to set, is a primitive
concept. It consists of its elements, called points. In a space, different points are not related to each
other, except that they are in the same space. However, we may introduce a relation among them.
In particular, a vector space is a space X with points (called vectors now) , , x y . There are two
algebraic operations defined on such a space. One is vector addition, the other is multiplication by
scalars.
More precisely, a vector addition puts any two vectors , x y to another vector
, called sum of x and y. It is Abelian and associative, i.e., for any vectors z x y X = +
, , x y z X ,
x y y + = + x
)
, (Abelian)
( ) ( x y z x y z + + = + + . (associative)
Moreover, there exist a zero vector 0 X , which satisfies 0 , x x x X + = . For each
x X , there is a unique x X , and it holds that ( ) x x 0 + = .
A multiplication of a vector x X by a scalar R defines a product of and x,
denoted by x X . It satisfies the following rules:
( ) ( ) , , , x x x X = R ;
1 , x x x = X ;
( ) , , , x y x y x y X + = + R ;
( ) , , , x x x x X + = + R .

Same as in linear algebra, we may define linear independence of a given finite set
{ }
1
, ,
n
M x x = as following:
1
0 0, 1,
n
i i i
i
, x i n
=
= = =

.
We define an arbitrary subset M to be linearly independent if any finite subset of M is
linearly independent. A vector space X is called of dimension r, if any ( 1 r ) + vectors in this
space are linearly dependent, while there exist r vectors which are linearly independent. An
infinite dimensional space refers to a vector space that does not have a finite dimension, i.e., for
any , there exist n vectors which are linearly independent. nN
We illustrate this through an example of . We know that the vectors and
2
R (1, 0)
37
(0,1) are linearly independent. On the other hand, any three vectors must be linearly dependent.
In fact, if the first two vectors are linearly independent, then the third one can be expressed as a
linear combination of the first two. See Figure 1. So we conclude that is of dimension 2.
2
R
2
R
x X
[ , C a b

Figure 1 Linear dependence of three vectors in .
A vector space only has algebraic structure, that is, addition and multiplication by scalar. To
introduce a geometric structure, we need to specify distance between two vectors (points). This
leads to the notion of norm. A norm is a function on a vector space X. That is, for each point
, it has a norm x R. Furthermore, this norm function satisfies the following four
axioms:
1. 0, x x X ;
2. 0 0 x x = = ;
3. , , x x x = R X ;
4. , , x y x y x y + + X . (triangle inequality)
The distance between two vectors x and y is defined naturally as x y . A normed
space is a vector space with a norm. We remark that a vector space may take different norms, and
the resulted normed spaces are different. In another word, only when both the algebraic structure
and geometric structure are the same, may we identify the normed space.
The Euclidean space R with
n
1/ 2
2
1
n
i
i
x x
=
| |
=
|
\ .

is the simplest example of normed


space. One may easily verify that this defines a normed space. In fact, we may define other norms,
for instance,
max
1
max
i
i n
x x

=
n
E
also satisfies the axioms and defines a norm on . Historically,
people sometimes write to stress that the norm is the Euclidean one. We remark, however, in
case of finite dimensional spaces, it may be proved that different norms are equivalent.
Moreover, to certain extent, a n-dimensional normed space is essentially . However, for
infinite dimensional spaces, different norms may make the same vector space totally different.
n
R
n
R
Now we are ready to describe function spaces. In particular, we consider the space
, which contains all functions that are continuous on a closed interval [ , . Each
function is a point in this space. The algebraic operations are defined as what we are used to,
namely
] ] a b
( )( ) ( ) ( ) f g t f t g t + = + , and ( )( ) ( ) f t f t = .
This defines a vector space, which is infinite dimensional. For geometric structure, we
define
| | ,
max ( )
t a b
f f t

= .
Because [ , ] f C a b , this maximum is reached, and the norm is well defined. All four
axioms are readily shown. This norm is analogous to
max
x in .
n
R
Meanwhile, we may define another norm (called norm)
2
L
38
1/ 2
2
2
( )
b
a
f f t dt
(
=
(

.
It also satisfies all axioms, and takes after the Euclidean norm. However, later on we shall see that
under these two norms, the function spaces behave differently.
As we have done in Calculus, limit plays a crucial in our discussion. A point x X is called
the limit of a sequence { }
n
x X , if lim 0
n
n
x x

= . Using the N type of expression, we


call lim
n
n
x x

= if 0, N , > N such that


n
x x < provided . n > N
In Calculus, we have learned that a Cauchy sequence or fundamental sequence is a sequence
{ }
n
x X , for which 0, , N > such that N
n m
x x < provided .
Moreover, it is proved that a Cauchy sequence is exactly a convergent sequence. That is, a Cauchy
sequence converges; and a convergent sequence must be a Cauchy sequence.
, n m N >
In a normed space X, we may show that a convergent sequence must be a Cauchy sequence,
by virtue of the triangle inequality. More precisely, if lim
n
n
x x

= , then 0, , N > N such


that
n
x x < provided n . Therefore, N >
2
n m n m
x x x x x x + < , for . , n m N >
However, the other direction of deduction does not hold in general. If it holds, it is called a
complete normed space, or a Banach space.
In this terminology, we then say is a Banach space. How about ?
n
R [ , ] C a b
If we take the norm
| | ,
max ( )
t a b
f t x

= , we claim that the space is complete. [ , ] C a b


Assume that there is a Cauchy sequence { } [ , ]
n
f C a b . Then 0 > , there exists
, such that N N

m n
f f < , provided . , m n N >
This implies that at each point , [ , ] t a b ( ) ( )
m n
f t f t < . That is, a real number sequence
{ } ( )
n
f t Ris Cauchy. Since is complete, this sequence is convergent, and we denote the
limit as
R

0
( ) lim ( )
n
n
f t f

t .
When we vary , a function [ , ] t a b
0
( ) f t is obtained. In the rest, we show

0
lim
n
n
f f

= ;

0
( ) f t is continuous on . [ , ] C a b
In fact, it is known that 0 > , there exists N N, such that

m n
f f < , provided . , m n N >
Therefore for , [ , ] t a b ( ) ( )
m n
f t f t < . Taking limit , we have n

0
( ) ( )
m
f t f t , for . m N >
This proves the first fact.
For the second fact, we take any
1 2
, [ , t t a b] . We notice that
1 N
f
+
is continuous on the
closed interval , therefore uniformly continuous. That is, take the same [ , ] a b 0 > , there exists
0 > , such that

1 1 1 2
( ) ( )
N N
f t f t
+ +
< , provided
1 2
t t < .
Using these inequalities, we obtain for any
1 2
t t < ,

0 1 0 2 0 1 1 1 1 1 1 2 1 2 0 2
( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) 3
N N N N
f t f t f t f t f t f t f t f t
+ + + +
+ + < .
39
This ends the proof. So we conclude that C a b with the aforementioned norm is a
Banach space.
[ , ]
On the other hand, if we take another norm, namely the norm, is no longer
complete. In fact, we may construct a sequence of continuous functions
2
L [ , ] C a b

| |
1, if ( ) / 2 1/ .
( ) ( ) / 2 , if ( ) / 2 1/
1, if ( ) / 2 1/ .
n
t a b n
, f t n t a b t a b
t a b n
+

= + +

+ >

n
<

This is a Cauchy sequence, because
( )
2 2
2
2
( ) ( ) 2( ) /
b
n m n m
a
f f f t f t dt b a =

N , if . , m n N >
However, the pointwise limit defines a function

0
1, if ( ) / 2,
( )
1, if ( ) / 2.
t a b
f t
t a b
< +
=

> +

It is discontinuous. Therefore, is no t complete with the norm. [ , ] C a b


2
L

Figure 2 Limit of function sequence in with norm. [ , ] C a b
2
L
A function is a mapping from (or its subspace) to . A mapping from a normed space R R
X to another normed space Y is an operator.

:

T X Y
x y Tx

=

A simple example is a linear transform, which can be represented by a matrix.

:

m n
T
x y Ax

=
R R


where A is a matrix. n m
In particular, if Y = , then is a functional. For special cases when X is a
function space, then T is actually a function of functions. One such functional is
R : f X R
: [ , ] f C a b R, with ( ( )) ( )
b
a
f x t x t dt =

.

Assignments

1, Is the sequence of functions
40
( ), if 1/
( )
1, if 1/ ,
n
n t a t a n
f t
t a n
< +
=

+

,

Cauchy in with the norm [ , ] C a b
| | ,
max ( )
t a b
f t x

= ?
2, For an operator represented by a square matrix

:
,
n n
T
x y Ax

=
R R


we assume that
1
1
( , , )
n
T
n
i
i i
x x x x
=
= = e

,
, with e (0, , 0,1, 0, , 0)
T
i
=
,
. If there is a
linear transform of basis
( )
to
1
, ,
n
e e
, ,

( )
'
n
e
1
', , e

, ,
, with
( ) ( 1 1
, ,
n n
Q e e =
)
', , ' e e
, , , ,
,
where Q is a non-singular matrix. Please find the matrix that represents the same linear transform.
41
2 Banach fixed point theorem with applications

Consider an operator from a normed space X to itself. It is a contraction if (0,1) , such
that Tx Ty x y for , x y X . A fixed point of T is x X such that . Tx = x
We have the Banach fixed point theorem as follows.
Theorem I f is a contraction on a Banach space : T X X X ,
then there exists exactly one fixed point x X .
The proof is done by iteration. We take an arbitrary point
0
x X , and let
1 i i
x Tx

= for
. Because T is a contraction, there is i N (0,1) , and

1 1
n
n n n n 0
x x x x Tx
+
x .
This yields

( )
1
0 0 0 0
1
n
n p n
n p n
x x Tx x

+
+
+ +

Tx x .
So, { }
i
x is a Cauchy sequence. The space X is complete, therefore { }
i
x converges.
Denote lim
n
n
* x x =

, and take the limit in the expression


1 i i
x Tx

= , we have . So, * x = * Tx
* x is a fixed point of T.
Now assume that there is another fixed point ' x . Then it holds that * ' * Tx Tx x x = ' .
This contradicts with the fact that T is a contraction.
From this proof, actually we have error estimate

0 0
*
1
n
n
x x Tx

x .
Now we apply this theorem to an ordinary differential equation

0 0
'( ) ( , ), ( ) x t f t x x t x = = .
Here we assume that ( , ) f t x is bounded and Lipschitz with respect to x, namely, there is
, such that , c k > 0
( , ) , ( , ) ( , ) f t x c f t x f t y k x y
]
.
Taking a time interval
0 0
[ , J t t = + , we consider the Banach space C J . We
further restrict to a subspace,
( )

{ }
0
( ) ( ) ( ) ( ) C J x t C J x t x c =

.
We now prove that a closed subspace of a Banach space is complete. Let X be a Banach
space equipped with norm - , and is a closed subspace. By a subspace, we mean that
Y is a subset of X. Y inherits from X the norm. Furthermore, a Cauchy sequence
Y X
{ }
i
y Y is
naturally a Cauchy sequence in X. Because X is complete, this sequence converges to certain
. Y is closed, therefore
0
y X
0
y Y . Because the norms are the same, in X
means exactly in Y. This ends the proof. We note that if Y is closed for vector
addition and multiplication by scalars, i.e.,
0 i
y y = lim
n
0
y lim
i
n
y

=
1 2
, , y y Y R, it holds
1 2
y y
1
, Y y Y + ,
then Y is a vector space, and a Banach space.
We conclude that is a Banach space. Now we construct an operator
by
( ) C J

: ( ) ( ) T C J C J


0
0
( ) ( , ( ))
t
t
Tx t x f x d = +

.
First, we verify . In fact, Tx is differentiable and therefore continuous, and ( ) Tx C J

42

0
0 0
( ) ( , ( ))
t
t
Tx t x f x d c t t c =

.
Secondly, we shall check if it is a contraction. Let , ( ) x y C J

, we have
| |
0
( ) ( ) ( , ( )) ( , ( ))
t
t
Tx t Ty t f x f y d k x y =

.
If we take 1/ k < , then T is a contraction.
Now by the Banach fixed point theorem, there is a unique fixed point ( ) ( ) x t C J

for T,
namely,

0
0
( ) ( , ( ))
t
t
x t x f x d = +

.
This gives a solution to the original ODE.
The fixed point theorem also provides a way to solve the ODE by iteration. In fact, taking
any
0
( ) ( ) x t C J

, we may iterate with

0
1 0
( ) ( , ( ))
t
n n
t
x t x f x d
+
= +

.
The limit gives the fixed point, and therefore the solution. This is called the Picard iteration.
As a second example, we shall apply the Banach fixed point theorem to integral equations.
An integral equation is an equation that involves integration of the unknown function. There
are different kinds of integral equations. Here we consider a Fredholm equation of the second kind
as follows.
( ) ( , ) ( ) ( )
b
a
x t k t x d v =

t .
Here ( ) x t is the unknown function, is a parameter, ,
is the kernel function, and is a given function on .
[ , ] a b R
[ , ] a b ( ( , ) [ , ] [ , ] k t a b a b ) C ( ) v t
We define an operator T as
( ) ( ) ( , ) ( )
b
a
Tx t v t k t x d = +

.
It maps a continuous function to a continuous function. Since k is continuous on a closed
domain, it is bounded, i.e.
( , ) k t c , for ( , ) [ , ] [ , ] t a b a b .
So,
( ) ( ) ( ) ( ) ( )
b
a
Tx t Ty t c x y d c b a x y

.
This is a contraction if ( ) c b a <1. Therefore, we know that the Fredholm integral equation
of the second type has a unique solution on if [ , ] C a b ( ) c b a 1 < , and it may be obtained
through an iteration

1
( ) ( ) ( , ) ( )
b
n n
a
x t v t k t x d
+
= +

.
Finally, we show that fixed point theorem may also be used for solving linear algebraic
equations. For this purpose, we consider the space , with norm
n
R

1
max
j
j n
x x

= .
A linear mapping T takes the form of Tx Ax b = + , where (
ij
A a ) = , and . ( )
j
b b =

1 1
1 1
( ) max ( ) max
n n
jk k k jk
j n j n
k k
Tx Ty A x y a x y x y a

= =
= =

.
Therefore, if a row sum criterion
1
1
max 1
n
jk
j n
k
a

=
<

is satisfied, then T is a contraction, and a


unique fixed point will be obtained through iteration. The fixed point solves a linear equation
43
( ) I A x b = .
Ex Fx =
1
n
jk
k
k j
=

( 1) +
( , ) f t x
Consider a linear algebraic system Cx d = . We reformulate it with . Thus we
have , or
C E F =
d +
1
( ) x E Fx

= d + . So, we shall take


1
A E F

= , and .
1
b E

= d
Let us describe the Jacobian iteration by taking diag( )
ii
E c = , which is the diagonal part of
C. To use this method, we require that the diagonal entries dominates corresponding rows, namely,
, 1,
jj
c c j < = , n .
This immediately leads to
1
1
x 1
n
jk
j n
k
a

=
ma <

, and the operator T is a contraction. The


iteration method actually can be written explicitly, since E is easily inverted,

( )
1
1
n
m m
j j jk
k
jj
k j
x d c
c
=

| |
|
=
|
|
\ .
k
x .


Assignments

1, Show that the solution to the ODE given by the fixed point theorem is differentiable, and really
solves the ODE.

2, In the proof of local existence for an ODE, we have assumed the boundedness and Lipschitz
of . However, this is required only locally, for
0 0
( , ) ( , ) x t x t . Please prove the local
existence under the condition that , such that , , , 0 a b c k >
( , ) , ( , ) ( , ) f t x c f t x f t y k x y , for
0 0 0 0
( , ) ( , ) ( , ) x t x a x a t b t b + + .

3, Show that
1
max
j
j n
x x

= defines a norm on .
n
R

4, Find the operator T and matrix E for Gauss-Seidel iteration
1
( 1) ( 1) ( )
1 1
1
j n
m m
j j jk k jk
k k j
jj
x d c x c
c

+ +
= = +
| |
=
|
\ .

m
k
x .
When does this iteration converges?

44
Chapter 5 Soliton
1 Solitary wave and soliton

In the year 1834, J. Scott Russell observed, when he rode on the horseback along a canal
in Scotland, a big wave propagating in a constant speed, in a persistent shape, for a very long time.
He reported in 1844, and called it the great wave of translation, primary wave of the first
order. In 1895, D.J. Korteweg and G. de Vries published a paper On the change of form of long
waves advancing in a rectangular canal and on a new type of long solitary waves. In this paper,
they derived a simple model, which bears their name as the KdV equation. Let u be elevation of
the water, it solves
6 0
t x xxx
u uu u + + = .
We notice that the third order term functions in a similar way as in the viscous
Burgers equation, namely, to prevent the formation of discontinuities. However, it differs by
being dispersive, instead of dissipative. More precisely, let us consider a bounded domain
xxx
u
xx
u
[ , ] M M , where we suppose that M is so big that there is essentially nothing outside of the
interval, namely, is close to 0 there. Consider first the viscous Burgers equation ( , ) u x t
. 0
t x xx
u uu u + =
We integrate over the interval,
2
0
2
M
M
M
x
M
d udx
u
u
dt

(
+ =
(

.
The difference of at the boundary might be neglected, and we may even translate the
axis to make it exact. However, the term is typically with different sign at the boundary
points
2
u
x
u
M and M, since u decreases to 0 at infinity. This makes a permanent leakage of the
total mass .
M
M
udx

On the other hand, with the same treatment on the KdV equation, we obtain

2
3 0
M
M
M
xx
M
d udx
u u
dt

( + + =

.
This term, on the other hand, can be a gain or a loss of the total mass, and the difference at
the boundary points over a long time period is negligible.
xx
u
With the KdV equation, we are ready to describe this solitary wave through traveling wave
analysis. By a solitary wave, it means that it is a single wave, which maintains its amplitude and
speed.
A traveling wave with speed c solves, with x ct = ,
. ''' (6 ) ' 0 u u c u + =
Integrating once, it gives
.
2
1
'' 3 u u cu C + =
The wave we are looking for has the property lim ( , ) 0
x
u x t

= , therefore C . Now we
consider the second order ODE
1
0 =
.
2
'' 3 0, ( ) 0 u u cu u + = =
It can be integrated once again,
.
3 2
2
' / 2 u u cu C + =
For the same reason, has to be zero. The solution to the first order nonlinear ODE
2
C
45

3 2
2
' / 2 u u cu C + =
is

2
2
sech
2 2
a a
u
| |
=

\ .
|
, with a c = .
Here the hyperbolic secant function is defined by

1 2
sech
cosh
z z
z
z e e

= =
+
.
This solution is quite interesting in that the height, which is , and the characteristic
width, which can be taken as , are related to the wave speed by
2
/ 2 a
2/ a a c = . A taller wave is
thinner, and runs faster.
Without going to the details, we point out that periodic solutions exist when and C are
non-zero. These solutions are called cnoidal waves.
1
C
2

Figure 1 Solitary waves with different speed.

A solitary wave is a single wave, with both end-states lim ( , ) 0
x
u x t

= . This type of wave


cannot be held by a dissipative system, such as viscous Burgers equation. There a traveling wave
must be a shock profile, therefore, . lim ( , ) lim (
x x
u x t u
+
> , ) x t
Mathematicians coined the word soliton, because they observed that the solitary waves in the
KdV equation behave like a particle. Note that elementary particles have names ending up with
on, such as proton, electron, etc. The invention of soliton came from numerical results by
Kruskal and Zabusky. They essentially put two solitary waves together, and numerically solved
the KdV equation. Later on, people found that they actually obtain a 2-soliton solution

| | ( )
| |
2
2 2 2 2 2 2 2
1 1 2 2 2 1 1 2 2 1 2 1 2 1 2 1 2 1
2
2
1 2 2 1 2 1 1 2
2( ) ( ) /( )
( , ) 2
1 ( ) /( )
f f f f f f
u x t
f f f f


+ + + + +
=
(
+ + + +

f f
,
where
2
( )
i i i
x t x
i
f e

= .
46

Figure 2 Soliton in KdV.

The wave interaction may be observed in Figure 2. Initially we observe two solitary waves,
with the taller one behind the shorter one. They propagate at different speeds, therefore the taller
one catches up at certain time. A complex combination of them occurs at this interaction period.
Yet eventually, the taller one goes across, and leaves the shorter one behind. The most fascinating
phenomenon is of course the superposition of two solitary waves. How can such a nonlinear
system still possesses superposition, and after a nonlinear interaction? As a matter of fact, the
interaction is not completely particle-like. A careful check shows that the phases of the two waves
changes after the crossing. By phase, we mean the place where each wave assumes its height. This
shows the effect of nonlinearity.
If
2
0 = , it reduces to the single solitary wave we obtained before. Let us consider a
specific 2-soliton solution

| |
3 4cosh(2 8 ) cosh(4 64 )
( , ) 12
3cosh( 28 ) cosh(3 36 )
x t x
u x t
t
x t x
+ +
=
+ t
.
There are two solitons here,
1 2
2, 8 = =
8s
/ 2
1
16 x t
, as , they are well separated, one is
, and the other is . The centers of these
two waves are and
t
2
(2 3
n3/ 4
2
2sech ( 4 ln3/ 2) x t
1
4 l x t =
ech 2 ln3/ 2) x t
l n3 = . Therefore, before and after collision,
there is a phase change.

Assignments

1, Find dispersion relations for the viscous Burgers equation and KdV equation at u u
respectively, and make comparison.
0
=

2, With phase plane analysis, show that periodic traveling waves exist for the KdV equation.



2 Inverse scattering transform

47
The study of solitons in the KdV equation
6 0
t x xxx
u uu u + =
led to inverse scattering transform (IST). This was invented by Gardner, Greene, Kruskal and
Miura (1969). Later on, an important generalization was made by Lax in terms of Lax pair.
Schematically, IST method is shown in the following figure.

step 1: direct scattering ( , 0) u x ( , 0) S

KdV Step 2: dispersion relation
(evolution)


( , ) u x t ( , ) S t
Step 3: inverse scattering

Figure 3 Inverse scattering transform.

To illustrate this method, we recall the Schroedinger equation, which describes the evolution
of a wave function ( , ) x t under a potential , which vanishes rapidly at infinity, ( ) V x

2
( )
2
t xx
i V
m
x = +
h
h .
The Schroedinger equation is the fundamental equation in quantum mechanics, and therefore
well studied during the years 19201960. This provides a solid basis for the development of IST.
The Schroedinger equation is a linear equation, thus naturally people investigates the
corresponding eigenvalue problem
( )
xx
V x . = +
More precisely, we call this a spectral problem, because is a function instead of a column
vector. In case of linear system, we consider n-dimensional problem
' x Ax = .
If the eigenvalues of A are
1 n

1
n
i i
i
, with normal eigenvectors , then in general,
for initial data
1
, ,
n
e e
0
(0) x x e
=

= = , we have solution as

1
( ) exp( )
n
i i
i
i
x t t
=
=

e
k
.
A similar approach is taken in linear PDE system such as the Schroedinger equation.
However, because it is a PDE, there exist not only discrete eigenvalues, but also continuous
spectra.
More specifically, in the Schroedinger equation, we put the initial data of KdV equation
as the potential function , and extract the following data (scattering data) from the
spectral problem
( , 0) u x V
.
( )
1
(0) { , } ; ( ), ( )
N
n n n
S k c r k a
=
=
Here, is a pair for positive discrete eigenvalue { , }
n n
k c
2
n
k = , with normalizing coefficient to
make the eigenfunction , and ( ) ~ ,
n
k x
n n
x c e x

+
2
( ) 1
n
x dx =

. The continuous
part, on the other hand, describes negative eigenvalue
2
k = k for , which has
eigenfunction
+
R
( ) x satisfying
( ) ~ ( ) , ; ( ) ~ ( ) , .
ikx ikx ikx
x e r k e x x a k e x

+ +
The coefficients r k and are called as reflection and transmission coefficients,
respectively.
( ) ( ) a k
48
In the second step, we consider the evolution of these scattering data. It turns out that with
as the potential, the scattering data becomes, ( , ) u x t
,
( )
1
( ) { , ( )} ; ( ; ), ( ; )
N
n n n
S t k c t r k t a k t
=
=
with

3 3
4 8
( ) (0) , ( ; ) ( ; 0) , ( ; ) ( , 0).
n
k t ik t
n n
c t c e r k t r k e a k t a k = = =
Here the initial data are those scattering data obtained in the first step. We shall explain why it
evolves this way later.
In the final step, we recover the solution from the spectral problem ( , ) u x t
( , )
xx
u x t = + .
That is, knowing the scattering data , we find the potential . There are two remarks
at this moment. First, the scattering data itself is not the complete information about the spectral
problem, namely, we do not have the exact form of the eigenfunctions. Instead, we only know
their asymptotes. Secondly, the inverse scattering step is done at each fixed time level t. That is, at
this step, there is no time evolution. This is purely a special ODE problem. Fortunately, in the 60s,
people know well how to do this step. In fact, we define a function
( ) S t ( , ) u x t

2
1
1
( ; ) ( ; )
2
n
N
k x ikx
n
x
i
F x t c e r k t e dx

=
= +


.
Then we consider the solution to the Gelfund-Levitan-Marchenko equation
. ( , ; ) ( ; ) ( , ; ) ( ; ) 0
x
K x y t F x y t K x z t F z y t dz
+
+ + + + =

The potential function is recovered by



| |
( , ) 2 ( , ; )
x
u x t K x x t = .
The IST method is similar to Fourier transform in solving PDE. There, the PDE is reduced to
ODE system with frequency as a parameter. We then find the spectra of initial data, and let it
evolve according to the ODE system. Then performing inverse Fourier transform, we obtain the
solution.
Now we start to explain the motivation of IST. Of course this is ad hoc. The key point is to
relate KdV equation with the spectral problem for the Schroedinger equation. This is initiated by
the Miura transform
.
2
x
u v v = +
It leads to the mKdV (modified KdV) equation
.
2
6 0
t x xxx
v v v v + =
This equation is no easier than the KdV equation itself. However, the transform inspires a
well-known change of variable, because it is a Riccati equation on v. By letting /
x
v = , one
obtains a linear equation
0
xx
u = .
This says that is eigenfunction for Schroedinger equation with eigenvalue 0.
The second question we may explain is: why the scattering data evolves this way? We discuss
in this lecture only the discontinuous eigenvalues. So, assume a solution of the KdV equation u is
the eigenfunction to

xx
u =
for this . Because u
2
n
k = /
xx
=
x
, the KdV equation becomes, after some
straight-forward calculations,
( )
2
0, 2( 2 ) .
t x x t x
x
R R R u u + = = +
Using the fact that , we integrate over
2
1 dx =

R
xR to get 0
t
= . Therefore, we
have
0
xx xx
R R = .
49
With the Schroedinger equation, this is equivalent to
( )
xx
R u R 0 + = .
Since the general solution to linear equation ( )
xx
u 0 + = are expressed as
1
C D
2
= + . In fact, if
1
is a non-trivial solution, then we may take
2 1
2
1
dx

=

,
which is obviously linearly independent with
1
. This is readily verified by

1, 2
2, 1, 1, 1 2
2 2 2
1 1 1 1
1 1
2 (
xx
xx xx x
x
dx
u

) .

| |
= + + = = +
|
\ .


So, we may select
1
= , which is certainly an eigenfunction. The above discussion implies

2
R C D = + .
On the other hand, we have the asymptotic expression as
. ~ ,
n
k x
n
c e x

+
This means . To maintain
2
~ * ,
n
k x
c e x +
2
1 dx =

R
, it requires . 0 D =
Now we have
2( 2 )
t x x
R u u C = + = .
Multiplying with , this amounts to

( ) ( )
2 2 2 2
/ 2 2 4
x
t x
u C
2
+ + = .
Integrating over , the left hand side vanishes, because xR
2
1 dx =

R
. Therefore,
as well. Now plugging the form into the equation, we notice
and have
0 C =
0 u ~ ,
n
k x
n
c e x

+
.
2 2 2
,
/ ( 2 4 )
n t n n
x
c c k ( + + =

0
0
This means
.
2
,
/ 2 ( 2 )
n t n n n
c c k k + =
The evolution of is determined.
n
c
A special case with IST method, is the N-soliton solution
| |
( , ) 2 ln det( )
xx
u x t I C = + .
Here I is the identity matrix, and with (
mn
C c = )

| |
3
exp ( )
( ) ( ) , ( ) (0) exp( 4 ) 0.
m n
mn m n m n n
m n
k k x
c c t c t c t c k t
k k
+
= =
+
>
0

Finally, we mention a property of the KdV equation, that is, it possesses infinite many
conservation laws. That is, there are infinite equalities for its solution in terms of
. ( ) ( ) 0
t x
T u X u + =
T is called the conserved density, and X the flux. A trivial conservation law is in the form of
. ( ) ( )
t x x t
F F + =
In the KdV equation, we let (Gardners transformation). The equation
becomes, after some computations,
2 2
x
u w w w = + +
.
2 2 2
(1 2 ) 6( ) 0
x t x xxx
w w w w w w ( + + + + =

Since is a parameter, and is a smooth function, it may be shown that w
.
2 2
6( ) 0
t x
w w w w w + + =
xxx
x x
This can be rewritten as the Gardner equation
.
2 2 3
(3 2 ) 0
t x
w w w w + =
Now we perform a perturbation type of expansion
50
.
2
0 1 2
( , ; ) w x t w w w = + + +
Meanwhile, Gardners transformation leads to
2 2
2
2 2 2 2 2
2 2
( )
( ) ( )
( )
x
x x
x
x x
w u w w
u u w w u w w
u u u w



= +
(
( ( = + + +

(

= +

So, by equating coefficients of the same order of , we may express as a function
of u and its derivatives. The Gardner equation gives, by expansion, infinite conservation laws
for , hence for u. It turns out that at odd orders of
i
w
i
w , the conservation laws are trivial, and
nontrivial at even orders. We list the first a few conservation laws as follows.
2
2 3 2
3 2 4 2 2 2
( 3 ) 0,
( ) ( 4 2 ) 0,
(2 ) ( 9 6 12 2 ) 0.
t xx x
t xx x x
x t xx x x xxx xx x
u u u
u u uu u
u u u u u uu u u u
+ + =
+ + =
+ + + + =



Assignments

1, Verify the first three conservation laws for the KdV equation.

51
Chapter 6 Chaos
1 Lorenz equations

Qualitative theory of ordinary differential equation had been a great success. In particular,
plane analysis provides a clean argument for second order autonomous systems. For systems with
bigger degrees of freedom, local behavior is described by critical point approach. People had
believed that the whole picture is essentially similar to that of a second order system.
But this is wrong. A topological result claims that the structure of is quite different from
. This was discovered by Lorenz, in the setting of an ODE.
3
R
2
R
The motivation of Lorenz was to disprove the predictability of mid-term weather system. One
of the main driven force of weather change is heat conduction, governed by the Rayleigh-Benard
system

( )
( )
( )
2
2 4
2
,
,
( , )
,
1
.
( , )
a
P
t x z
t x z x

x



= + +


= + +


Here is the stream function, the temperature, the Rayleigh number, and
a
P Prandtl
number. The velocities are
, u w
z x

= =

.
Lorenz made a truncation
2 2
2 ( ) sin sin ,
2 ( ) cos sin ( ) ,
c
a
a
X ax z
a
R
Y ax z Z
R

| | +
=
|

\ .

(
=



with
( )
2 2
1
a t

= + , and
c
R is the critical Rayleigh number. He obtained the system (later
called Lorenz equations)

' ,
' ,
' ,
X X Y
Y XZ rX
Z XY bZ
= +

= +

Y
) with . We note that there is a symmetry under transform
2 2 2
/ , 4 /(
a c
r R R b a = = +
{ } { } , , , , X Y Z X Y Z .
Defining a Lyapunov function
,
2 2
( ) ( 2 ) V t rX Y Z r = + +
2

we have
.
2 2 2
' 2 ( 2 ) V rX Y bZ brZ = +
We may take

2 2
2 2
2
/( ), if 2 , 1,
/( 1), if 2 , 1,
4 , otherwise.
b r b b
c b r b b
r

52
Then for ' 0 V < ( , , ) X Y Z satisfying V . Therefore, a trajectory entered the ellipsoid
cannot escape from it. This defines a trapping zone.
c
Now consider the critical points. A critical point solves

0 ,
0 ,
0 ,
X Y
XZ rX Y
XY bZ
= +

= +


When , the only critical point is . The Jacobian matrix is 1 r < (0, 0, 0)
,
0
1 0
0 0
r
b
(
(

(
(

which has three eigenvalues with non-positive real-part. Therefore, we expect that in this case, all
trajectories point towards this node.
When r < , however, one eigenvalue for the Jacobian matrix at becomes
positive. This node then turns into a saddle. Meanwhile, there appears two new critical points
1 (0, 0, 0)
( ( 1) 1), 1) b r b r r , ( . It may be found that the eigen-equation for both of them is
.
3 2
( ) ( 1) ( ) 2 ( 1) 0 f b r b b r = + + + + + + =
We observe that ( ) f = , and
2
'( ) 3 2( 1) ( ) f b b r = + + + + +
, ( 1) b
. In particular,
when , three eigenvalues are 1 r = 0, + 0 1 r . Thus when 1 < <
r
, there are one
real eigenvalues with negative sign, and two complex conjugate roots with negative real parts.
This means that these two critical points are stable nodes. As increases, the real eigenvalue
becomes smaller and smaller. On the other hand, sum of three roots are fixed as ( 1 b) + + .
Therefore, when r becomes big enough, the real eigenvalue becomes smaller than
( 1 b) + + . The two complex eigenvalues take positive real parts. The critical points are then
unstable, each with a one dimensional stable manifold and two dimensional unstable manifold. In
the unstable manifold, trajectories are focus-like.
More detailed analysis shows that at certain , there is a heteroclinic orbit from
and either of the unstable critical point. Due to the symmetry, there is a pair of such heteroclinic
orbits. When these heteroclinic orbits break, chaos appears.
0
r (0, 0, 0)
There are many exciting indications made through this analysis. First, there is a sensitivity of
initial data. All trajectories are bounded within the trapping zone, and there is no ordinary
attractor, namely, stable critical points, limit cycle, heteroclinic or homoclinic orbits. Therefore,
each trajectory must wind within the ellipsoid forever. This makes a very complex entangled
picture. Two trajectories, which start from two closed points in the phase plane, take apart from
each other after a long run.
53

Figure 1 A trajectory of the Lorenz equations.

Secondly, a further implication of the sensitivity is, pointed out by Lorenz, our
deterministic ODE system gives random solution. That is, because we do not know
precisely the initial data, and in a long run this imprecision causes different results. In
weather forecast, this indicates that a small perturbation may result in a drastic change
after some time, typically in the scale of couple of months.
Later there were further discussions on the Lorenz equations, including to prove
rigorously the chaotic behavior of the system (strange attractors), and numerical
studies from various aspects: Lyapunov exponent, Poincare section, and so on.
Poincare section is a nice way for detecting the type of attractors. We put a plane (say
) on the phase space. If a trajectory goes toward a critical point (we are not
likely to have this critical point on our plane), then after t T , there is no
intersection of this trajectory with our plane. On the other hand, if it goes toward a
limit cycle, we are likely to get one or two intersecting points. If it is a quasi-periodic
orbit, we shall end up with couple of closed curves. On the other hand, if it is chaotic,
we obtain a dense set on this plane.
0 z =
* >
Another way to study chaos is to go back to the Rayleigh-Benard equation.
However, that is two complicated for a substantial understanding. Instead, there are
investigations by a better truncation, i.e. with more modes. For instance, by a 14
th

order ODE system, it was observed that there is a bifurcation process while increasing
. More precisely, the system goes to zero equilibrium when r 1 r < ; and periodic and
quasi-periodic solution appears when is big enough. When quasi-periodic solution
breaks, chaos is observed.
r


Assignments

1, Using Matlab, or other mathematical tools, please plot some trajectories of the Lorenz equations
with 10, 28, 8/ 3 r b = = = .
54
2 Logistic map and routes to chaos

Chaos appears not only in differential equations, but also in maps. Let us consider

1
(1 )
n n n
x x x
+
= .
Here [0, 4] is a parameter. Assume that
0
[0,1] x , then it holds [0,1],
n
x n N.
When 0 = , obviously . In fact, we notice that the function 0,
n
x n = N
( ) (1 ) f x x x =
takes derivative
'( ) (1 2 ) , [0,1] f x x x = .
Therefore, it is a contraction provided 1 < . Accordingly, , by virtue of Banach fixed
point theorem.
0
n
x
In fact, when
1
1 = ,
1
0 x = = is the only fixed point of ( ) f x . This fixed point is
globally stable. When 1 > , becomes locally unstable, 0 x = '( 1 f 0) = > . There appears
another fixed point
2
1 1 / =
. When (1, 3) , this new fixed point is locally stable
'(1 1/ ) 2 f
y =
=
( ) y f x =
. In fact, it is globally stable. To view this, we may plot the curve
and the line . x

Figure 1 Logistic map: 2.4 = .

When exceeds
2
3 =
2
( )
, both fixed points are unstable. However, there appears a
2-periodic state. That is, ( ( )) f x f f x = has fixed points. This means,
( (1 ))(1 (1 )) x x x x x = .
This is a fourth order algebraic equation. Fixed points of ( ) f x remains to be roots, and
there are two new ones

3,4
(1 ( 1)( 3)) / 2 = + + + .
The stability may be obtained by analyzing
2 3,4
'( ) f . It turns out that these two roots are
stable if
4
1 = + 6 . With an initial value
0
x , typical picture looks like the one in Figure 2.
We observe that when , n
n
x approaches towards this 2-periodic solution
3 4 3 4
{ } .
55

Figure 2 2-periodic solution at 3.2 = .

When further increases, this 2-periodic solution becomes unstable, and it turns out,
instead of , there appears stable 4-periodic solution. This corresponds to the
rest four fixed points
2
( ) ( ( )) f f x =
5,6,7,8
g x
to
4 2 2
( ) ( ( )) f x f f x = , which is an eighth-order polynomial.
They are stable if
3
3.544 =
, 1
n
x n >
. After wards, 8-periodic solution takes place. And then
16-periodic, 32-periodic, . We remark that the periodic state is independent of initial value, and
we thus only plot . See Figure 3.

Figure 3 4-periodic solution at 3.5 = ; 8-periodic solution at 3.55 = .

Figure 4 16-periodic solution at 3.565 = ; chaotic solution at 4 = .
This is a periodic bifurcation, and the sequence of
n
takes the following values
1, 3, 3.449490, 3.544090, 3.564407, 3.568759, 3.569692,
There is a limit of this sequence and 3.569945672

= . Beyond this threshold, no


afore-mentioned periodic solution is unstable, and chaos occurs. There is a constant appears in this
56
bifurcation process, namely, the distance between subsequent
n
s decreases monotonously.
4 =
=
8
2
k
2
i

1
1
lim 4.669201661
n n
n
n n

.
It is later on found to be uniform, and called the Feigenbaum number.
Now we consider the case 4 = in the chaotic regime, and make a detailed analysis.
Let x , the iteration reads
2
sin ( / 2) y =

2 , [0,1/ 2],
( )
2(1 ), [1/ 2,1].
y y
y y
y y

It is plotted in Figure 5.

Figure 5 Logistic map at .

There are two fixed points,
1
0 = and
2
2/ 3 = , corresponding to
1,2
respectively.
They are also unstable, because the slope is 2 . In fact, for any N N
( ( )) y y
, there is an unstable
N-periodic solution. For instance, we consider
3
( ) ( ) for . The
3-periodic solution corresponds to roots of equation
3 N =

3
( ) ( ( ( ))) = = .
This is illustrated in Figure 6, as the intersection points. There are eight such points. We discard
1,2
, and obtain two pairs of 3-periodic solutions, namely,
2/ 9 4/ 9 8/ 9 2/ 9
and
2/ 7 4/ 7 6/ 7 2/ 7 .
However, slopes of
3
( ) y at all fixed points are , so none of them is stable.
There is another interpretation of ( ) y . Assuming that
1 2
0. y p p = in binary form,
namely and 0,1
i
p = 2
i
i
y p

=

, the iteration actually means


2 3
, p
2 3
( ) 0. or 0. y p p p = . So, this corresponds to a left-shifting map. There are two
implications. First is sensitivity to initial data. If ' y y , then after k times of iteration, the
difference is amplified to 2
k
' 1 y y . Noticing finite digit expression of a number in
computer, this means after certain steps, the information in initial data is completely lost. The
second implication is ergodicity, which means that covers the whole interval of .
In fact, there are three cases for the exact solution. If
{ }
n
y
N
i
y
[0,1]
1
i
p

=
=

, then .
If y is a rational number, yet with nominator not in the form of , then after certain steps, a
0, for y n
n
2
k
N = >
57
periodic solution is selected. Last, if y is irrational, then { }
n
y covers , without repeating.
Since there are more irrational numbers than rational ones, we may expect that equally
distributed in the interval . Correspondingly,
(0,1)
n
y
[0,1]
n
x distributes in the interval as well,
yet not equally.
[0,1]
|
, 4

'
1
3.6786
=
3.5926

Figure 6 3-periodic solution.

There are some other issues interesting in this logistic mapping. First, there are
certain details for
|
. In fact, when we decrease , there
appear two disconnected patches of chaos region.
=

Figure 6 Logistic map at 3.6 .

When , there appear four such patches.
'
2
=
58

Figure 8 Logistic map at 3.58 = .

This process continues, and in the limit .
' '
3.569945672
n


= =
Another interesting phenomenon in the logistic map is, which also exists in many
other systems, that there are windows of periodic solution in the chaotic regime
( , 4]

. Such periodic solution appears not for isolated values of , instead,


appears for intervals of . For instance, Figure 9 shows a 3-periodic solution at
3.8384 = .
One may ask a question, what is the relation between ODE system (continuous
dynamical systems) and maps (discrete dynamical systems). In fact, we may extend
the previous discussion on logistic map to 2-D, namely,

1 1
( , ) ( ( , ), ( , ))
n n n n n n
x y f x y g x y
+ +
= .
Then for an ODE system, a Poincare section may be drawn as a hyperplane on phase
space. When a trajectory starting from a point on this plane returns back, a 2-D map is
defined. Therefore, one may investigate the attractor of the ODE accordingly.


59

You might also like