Chapter 22

Nonlinear Partial Dierential Equations

The ultimate topic to be touched on in this book is the vast and active eld of nonlinear
partial dierential equations. Leaving aside quantum mechanics, which remains to date an
inherently linear theory, most real-world physical systems, including gas dynamics, uid
mechanics, elasticity, relativity, ecology, neurology, thermodynamics, and many more, are
modeled by nonlinear partial dierential equations. Attempts to survey, in such a small
space, even a tiny fraction of such an all-encompassing range of phenomena, methods,
results, and mathematical developments, are doomed to failure. So we will be content to
introduce a handful of prototypical, seminal examples that arise in the study of nonlinear
waves and that serve to highlight some of the most signicant physical and mathematical
phenomena not encountered in simpler linear systems. We will only have space to look at
simple one-dimensional models; the far more complicated nonlinear systems that govern our
three-dimensional dynamical universe quickly lead one to the cutting edge of contemporary
research.
Historically, comparatively little was known about the extraordinary range of behav-
ior exhibited by the solutions to nonlinear partial dierential equations. Many of the
most fundamental phenomena that now drive modern-day research, including solitons,
chaos, stability, blow-up and singularity formation, asymptotic properties, etc., remained
undetected or at best dimly perceived in the pre-computer era. The last sixty years has
witnessed a remarkable blossoming in our understanding, due in large part to the insight
oered by the availability of high performance computers coupled with great advances in
the understanding and development of suitable numerical approximation schemes. New an-
alytical methods, new mathematical theories, coupled with new computational algorithms
have precipitated this revolution in our understanding and study of nonlinear systems, an
activity that continues to grow in intensity and breadth. Each leap in computing power
coupled with theoretical advances has led to yet deeper understanding of nonlinear phe-
nomena, while simultaneously demonstrating how far we have yet to go. To make sense
of this bewildering variety of methods, equations, and results, it is essential build upon
a rm foundation on, rst of all, linear systems theory, and secondly, nonlinear algebraic
equations and nonlinear ordinary dierential equations.
Our presentation is arranged according to the order of the underlying dierential
equation. First order nonlinear partial dierential equations model nonlinear waves and
arise in gas dynamics, water waves, elastodynamics, chemical reactions, transport of pol-
lutants, ood waves in rivers, chromatography, trac ow, and a wide range of biological
and ecological systems. One of the most important nonlinear phenomena, with no linear
counterpart, is the break down of solutions in nite time, resulting in the formation of
12/11/12 1169 c 2012 Peter J. Olver
discontinuous shock waves. A striking example is the supersonic boom produced by an
airplane that breaks the sound barrier. As in the linear wave equation, the signals propa-
gate along the characteristics, but in the nonlinear case the characteristics can cross each
other, precipitating the onset of a shock. The characterization of the shock dynamics re-
quires additional physical information, in the form of a conservation law, that supplements
the original partial dierential equation.
Parabolic second order partial dierential equations govern nonlinear diusion pro-
cesses, including thermodynamics, chemical reactions, dispersion of pollutants, and popu-
lation dynamics. The simplest and most well understood is Burgers equation, which can,
surprisingly, be linearized by transforming it to the heat equation. This accident provides
an essential glimpse into the world of nonlinear diusion processes. In the limit, as the
diusion or viscosity tends to zero, the solutions to Burgers equation tend to the shock
wave solutions to the limiting rst order dispersionless equation, and thus provides an
alternate mechanism for unraveling shock dynamics.
Third order partial dierential equations arise in the study of dispersive wave motion,
including water waves, plasma waves, waves in elastic media, and elsewhere. We rst treat
the basic linear dispersive model, comparing and contrasting it with the hyperbolic mod-
els we encountered earlier in this text. The distinction between group and wave velocity
observed when, for instance, surface waves propagate over water is developed. Fi-
nally, we introduce the remarkable KortewegdeVries equation, which serves as a model for
waves in shallow water, waves in plasmas, and elsewhere. Despite its intrinsic nonlinearity,
it supports stable localized traveling wave solutions, known as solitons, that, remarkably,
maintain their shape even under collision. The KortewegdeVries equation is the proto-
typical example of an integrable system, and this discovery in the mid 1960s inaugurated
intense and ongoing research in the remarkable physical models that exhibit integrability,
a development that has had many ramications in both pure and applied mathematics.
22.1. Nonlinear Waves and Shocks.
Before attempting to tackle any nonlinear partial dierential equations, we must care-
fully review the solution to the simplest linear rst order partial dierential equation.
Linear Transport and Characteristics
The transport equation
u
t
+ cu
x
= 0, (22.1)
is so named because it models the transport of, say, a pollutant in a uniform uid ow. Let
us begin by assuming that the wave speed c is constant. According to Proposition 14.8,
every solution is constant along the characteristic lines of slope
dx
dt
= c, namely x ct = constant. (22.2)
As a consequence, the solutions are traveling waves of the form
u(t, x) = p(x ct), (22.3)
12/11/12 1170 c 2012 Peter J. Olver
Figure 22.1. Traveling Wave.
where p() is an arbitrary function of the characteristic variable = xct. To a stationary
observer, the solution (22.3) appears as a wave of unchanging form moving at velocity c.
When c > 0, the wave translates to the right, as illustrated in Figure 22.1. When c < 0,
the wave moves to the left, while c = 0 corresponds to a permanent wave form that remains
xed at its original location.
Slightly more complicated, but still linear, is the non-uniform transport equation
u
t
+ c(x)u
x
= 0, (22.4)
where the wave velocity c(x) depends upon the spatial position. This equation models
unidirectional waves propagating through a non-uniform, but static medium. Generalizing
the construction (22.2), we dene a characteristic curve to be a solution to the autonomous
ordinary dierential equation
dx
dt
= c(x). (22.5)
Thus, unlike the constant velocity version, the characteristics are no longer necessarily
straight lines. Nevertheless, the preceding observation remains valid:
Proposition 22.1. Solutions to the linear transport equation (22.4) are constant on
its characteristic curves.
Proof : Let x(t) be a characteristic curve, i.e., a solution to (22.5), parametrized by
the time t. Let h(t) = u(t, x(t)) be the value of the solution at the point (t, x(t)) on the
given characteristic curve. Our goal is to prove that h(t) is a constant function of t, and,
as usual, this is done by proving that its derivative is identically zero. To dierentiate h(t),
we invoke the chain rule:
dh
dt
=
d
dt
u(t, x(t)) =
u
t
(t, x(t)) +
dx
dt
u
x
(t, x(t)) =
u
t
(t, x(t)) + c(x(t))
u
x
(t, x(t)) = 0.
We replaced dx/dt by c(x) since we are assuming that x(t) is a characteristic curve, and
hence satises (22.5). The nal combination of derivatives is zero whenever u solves the
transport equation (22.4). Q.E.D.
Since the characteristic curve dierential equation (22.5) is autonomous, it can be
immediately solved:
b(x)
_
dx
c(x)
= t + k, (22.6)
where k is the constant of integration. Thus, the characteristic curves are parallel, each
being a translate of the graph of t = b(x) in the direction of the t axis. The characteristic
12/11/12 1171 c 2012 Peter J. Olver
t
x
(t, x)
(0, y)
Figure 22.2. Characteristic Curve.
curves are therefore dened by the formula x = g(t + k), where g = b
1
is the inverse
function. (See Section 20.1 for full details.)
Observe that the characteristic curves are the level sets of the characteristic variable
= b(x) t. As a consequence, any function which is constant along the characteristic
curves depends only on the value of the characteristic variable at each point, and hence
takes the form
u(t, x) = p(b(x) t) (22.7)
for some function p(). In other words, the characteristic curves are the common level
curves of all solutions to the transport equation. It is easy to check directly that, provided
b(x) is dened by (22.6), u(t, x) solves the partial dierential equation (22.4) for any choice
of function p().
To nd the solution that satises the prescribed initial conditions
u(0, x) = f(x) (22.8)
we merely substitute the general solution formula (22.7). This leads to the equation
p(b(x)) = f(x), and, therefore, p() = f

b
1
() = f
_
g()
_
.
The resulting solution formula has a simple graphical interpretation: to nd its value u(t, x)
at a given point, we look at the characteristic curve passing through (t, x). If this curve
intersects the x axis at the point (0, y), then u(t, x) = u(0, y) = f(y). The construction
is illustrated in Figure 22.2. Incidentally, if the characteristic curve through (t, x) doesnt
intersect the x axis, the solution value u(t, x) is not prescribed by the initial data.
Example 22.2. Let us solve the particular transport equation
u
t
+
1
x
2
+ 1
u
x
= 0 (22.9)
12/11/12 1172 c 2012 Peter J. Olver
t
x
Figure 22.3. Characteristic Curves for u
t
+
1
x
2
+ 1
u
x
= 0.
by the method of characteristics. According to (22.5), the characteristic curves satisfy the
rst order ordinary dierential equation
dx
dt
=
1
x
2
+ 1
.
Separating variables and integrating, we nd
_
(x
2
+ 1) dx =
1
3
x
3
+ x = t + k,
where k is the integration constant. Some of the resulting characteristic curves are plotted
in Figure 22.3.
The characteristic variable is =
1
3
x
3
+ x t, and hence the general solution to the
equation takes the form
u = p
_
1
3
x
3
+ x t
_
,
where p() is an arbitrary function. A typical solution, corresponding to initial data
u(t, 0) =
1
1 + (x + 2.75)
2
,
is plotted at times t = 0, 2, 5, 10, 25, 50 in Figure 22.4. The fact that the characteristic
curves are not straight means that, although the solution remains constant along each
individual curve, a stationary observer will witness a dynamically changing prole as the
wave moves along through the non-uniform medium. The wave speeds up as it aapproaches
the origin, and then slows back down once it passes and moves o to the right. As a result,
we observe the wave spreading out as it approaches the origin, and then contracting as it
moves o to the right.
12/11/12 1173 c 2012 Peter J. Olver
Figure 22.4. Solution to u
t
+
1
x
2
+ 1
u
x
= 0.
t
x
Figure 22.5. Characteristic Curves for u
t
xu
x
= 0.
Example 22.3. Consider the equation
u
t
xu
x
= 0. (22.10)
In this case, the characteristic curves are the solutions to
dx
dt
= x, and so x e
t
= k, (22.11)
where k is the constant of integration; see Figure 22.5. It is easier to adopt = x e
t
as
the characteristic variable here, noting that its level sets are the characteristic curves. The
12/11/12 1174 c 2012 Peter J. Olver
Figure 22.6. Solution to u
t
xu
x
= 0.
solution therefore takes the form
u = p(xe
t
), (22.12)
where p() is an arbitrary function of = xe
t
. Given the initial data
u(0, x) = f(x), the resulting solution is u = f(xe
t
).
For example, the solution
u(t, x) =
1
(xe
t
)
2
+ 1
=
e
2t
x
2
+ e
2t
corresponding to initial data u(t, 0) = f(x) = (x
2
+ 1)
1
is plotted in Figure 22.6 at times
t = 0, 1, 2, 3. Note that since the characteristic curves all converge on the t axis, the
solution becomes more and more concentrated at the origin. In the limit, it converges to
the function that is zero everywhere except for the value u(t, 0) 1 at the origin. Warning:
The limit is not a delta function, since its value at x = 0 remains bounded.
A Nonlinear Transport Equation
Perhaps the simplest possible nonlinear partial dierential equations is the nonlinear
transport equation
u
t
+ uu
x
= 0. (22.13)
rst systematically studied by Poisson and Riemann in the early nineteenth century. Since
it appears in so many applications, this equation appears in the literature under a variety
of names, including the Riemann equation, the inviscid Burgers equation, and the dis-
persionless KortewegdeVries equation. It and its multi-dimensional and multi-component
generalizations play a crucial role in the modeling of gas dynamics, trac ow, ood waves
in rivers, chromatography, chemical reactions, and other areas; see [188].
12/11/12 1175 c 2012 Peter J. Olver
The rst order partial dierential equation (22.13) has the form of a transport equa-
tion, whose wave velocity c = u depends, not on the position x, but rather on the size of the
disturbance. Larger waves move faster, and overtake smaller waves. Waves of elevation,
where u > 0, move to the right, while waves of depression, where u < 0, move to the left.
Fortunately, the method of characteristics that was developed for linear wave equations
also works in the present nonlinear situation and leads to a complete solution. Mimicking
our previous construction, (22.5), let us dene a characteristic curve of the nonlinear wave
equation (22.13) to be a solution to the ordinary dierential equation
dx
dt
= u(t, x). (22.14)
As such, the characteristics depend upon the solution u, which, in turn, is based on the
characteristic variable. So we appear to be trapped in a circular argument. The resolution
of the apparent conundrum is to observe that, as in the linear case, the solution u(t, x)
remains constant along its characteristics, and this fact will allow us to simultaneously
specify both.
To prove this claim, suppose that x = x(t) parametrizes a characteristic curve. We
need to show that the function h(t) = u(t, x(t)), which is obtained by evaluating the
solution along the curve, is constant. As usual, constancy is proved by checking that its
derivative is identically zero. Invoking the chain rule, and then (22.14), we deduce that
dh
dt
=
d
dt
u(t, x(t)) =
u
t
(t, x(t))+
dx
dt
u
x
(t, x(t)) =
u
t
(t, x(t))+u(t, x(t))
u
x
(t, x(t)) = 0.
The nal expression vanishes because u is assumed to solve the wave equation (22.13) at
all values of (t, x), including those on the curve (t, x(t)). This veries our claim that h(t)
is constant, and so the solution u is constant on the characteristic curve. This has the
implication that the right hand side of equation (22.14) is a constant whenever x = x(t)
denes a characteristic curve, and so the derivative dx/dt is a constant namely the value
of u on the curve. In this manner, we arrive at the key deduction that the characteristic
curve must be a straight line
x = ut + k, (22.15)
whose characteristic slope u equals the value assumed by the solution u on it. The larger
u is, the steeper the characteristic line, and the faster that part of the wave travels.
The corresponding characteristic variable = xt u depends upon the solution, which
can now be written in implicit form
u = f(x t u), (22.16)
where f() is an arbitrary function of the characteristic variable. The solution u(t, x) can
be found by solving the algebraic equation (22.16). For example, if
f() = +
is an ane function, with , constant, then
u = (x t u) + , and hence u(t, x) =
x +
1 + t
(22.17)
12/11/12 1176 c 2012 Peter J. Olver
Figure 22.7. Two Solutions to u
t
+ uu
x
= 0.
is the corresponding solution to the nonlinear transport equation. At each xed t, the
graph of the solution is a straight line. If > 0, the solution attens out as t . On
the other hand, if < 0, the straight line rapidly steepens to vertical as t approaches the
critical time t

= 1/, at which point the solution ceases to exist it is said to blow

up. In Figure 22.7, we graph the solution with = 1, = .5, when t = 0, 1, 5, 20 on the
top row, and = .2, = .1, at times t = 0, 3, 4, 4.9 on the bottom row. In the second
case, the solution becomes vertical as t 5 and then ceases to exist.
In general, to construct the solution u(t, x) to the initial value problem
u(0, x) = f(x), (22.18)
we note that, at t = 0, the implicit solution formula (22.16) reduces to u(0, x) = f(x).
Thus, the function f coincides with the initial data. However, because (22.16) is an implicit
equation for u(t, x), it is not immediately evident
(a) whether it can be solved to give a well-dened value for u(t, x), and,
(b) even granted this, how to describe the solutions qualitative features and dynamical
behavior.
A more instructive and revealing strategy is based on the following geometrical con-
struction, inspired by the linear version appearing in Figure 22.2. Through each point
(0, y) on the x axis, draw the characteristic line
x = t f(y) + y (22.19)
whose slope, namely f(y) = u(0, y), equals the value of the initial data at that point.
According to the preceding argument, the solution will have the same value on the entire
characteristic line (22.19), and so
u(t, t f(y) + y) = f(y). (22.20)
For example, if f(y) = y, then u(t, x) = y whenever x = t y + y; eliminating y, we recover
u(t, x) = x/(t + 1), which agrees with one of our straight line solutions (22.17).
12/11/12 1177 c 2012 Peter J. Olver
t
x
Figure 22.8. Characteristic Lines for f(x) =
1
4
sin(1.8x .8).
t
x
Figure 22.9. Characteristic Lines for a Rarefaction Wave.
Now, the trouble with our construction is immediately apparent from the illustrative
Figure 22.8. Any two characteristic lines that are not parallel must cross each other
somewhere. The value of the solution must equal to the slope of the characteristic line,
and so, at the crossing point, the solution is required to assume two dierent values, one
corresponding to each line. Something is clearly amiss, and we need to study the resulting
solutions in more depth.
It turns out that there are three basic scenarios. The rst, trivial case is when all the
12/11/12 1178 c 2012 Peter J. Olver
Figure 22.10. Rarefaction Wave.
characteristic lines are parallel and so the diculty does not arise. In this case, they all
have the same slope, say c, which means that the solution has the same value on each one.
Therefore, u(t, x) c is a trivial constant solution.
The next simplest case occurs when the initial data f(x) is everywhere increasing, so
f(x) f(y) whenever x y, which is assured if its derivative is never negative: f

(x) 0.
In this case, as in sketched in Figure 22.9, the characteristic lines emanating from the x
axis fan out into the right half plane, and so never cross each other when t 0. Each point
(t, x) for t 0 lies on a unique characteristic line, and the value of the solution at (t, x) is
equal to the slope of the line. Consequently, the solution is well-dened at all future times.
Physically, such solutions represent rarefaction waves, which gradually spread out as time
progresses. A typical example, corresponding to initial data
u(0, x) = tan
1
3x +

2
,
is plotted in Figure 22.10 at successive times t = 0, 1, 2, 3. Note how the slope of the
solution gradually diminishes as the rarefaction wave spreads out.
The more interesting case is when f

(x) < 0. Now some of the characteristic lines

starting at t = 0 will cross at some point in the future. If (t, x) lies on two or more distinct
characteristic lines, the value of the solution u(t, x), which should equal the characteristic
slope, is no longer uniquely determined. Although one might be tempted to deal with such
multiply-valued solutions in a purely mathematical framework, from a physical standpoint
this is unacceptable. The solution u(t, x) is supposed to represent a physical quantity, e.g.,
density, velocity, pressure, etc., and must therefore assume a unique value at each point.
The mathematical model has broken down, and fails to agree with the physical reality.
Before confronting this diculty, let us rst, from a theoretical standpoint, try to
understand what happens if we were to continue the solution as a multiply-valued function.
To be specic, consider the initial data
u(0, x) =

6

1
3
tan
1
x, (22.21)
appearing in the rst plot in Figure 22.12. The corresponding characteristic lines are
sketched in Figure 22.11. Initially, they do not cross, and the solution remains a well-
dened, single-valued function. However, eventually one reaches a critical time, t = t

> 0,
when the rst two characteristic lines cross each other. Subsequently, a wedge-shaped
region appears in the (t, x) plane, consisting of points which lie on the intersection of three
12/11/12 1179 c 2012 Peter J. Olver
t
x
Figure 22.11. Characteristics for a Shock Wave.
dierent characteristic lines with dierent slopes; at such points, the solution achieves
three distinct values. Outside the wedge, the points only belong to a single characteristic
line, and the solution remains single-valued. (The boundary of the wedge consists of points
where only two characteristic lines cross.)
To fully appreciate what is going on, look now at the sequence of pictures of the
multiply-valued solution at successive times in Figure 22.12. Since the initial data is
positive, f(x) > 0, all the characteristic slopes are positive. As a consequence, all the
points on the solution curve will move to the right, at a speed equal to their height. Since
the initial data is a decreasing function, points lying to the left will move faster than those
on the right, and eventually overtake them. Thus, as time passes, the solution steepens.
At the critical time t

when the rst two characteristic lines cross, say at x

, the tangent
to the solution curve has become vertical:
u
x
(t, x

) as t t

.
Afterwards, the solution graph no longer represents a single-valued function; its overlapping
lobes lie over points (t, x) in the aforementioned wedge.
The critical time t

can be determined from the implicit solution formula (22.16).

Indeed, if we dierentiate with respect to x, we nd
u
x
=

x
f(x t u) = f

()
_
1 t
u
x
_
, where = x t u
is the characteristic variable, which is constant along the characteristic lines. Solving,
u
x
=
f

()
1 + t f

()
.
12/11/12 1180 c 2012 Peter J. Olver
Figure 22.12. MultiplyValued Solution.
Therefore, the slope blows up,
u
x
, as t
1
f

()
.
In other words, if the initial data has negative slope at position x, so f

(x) < 0, then the

solution along the characteristic line emanating from the point (0, x) will break down at
the time 1/f

(x). As a consequence, the earliest critical time is

t

= min
_

1
f

(x)

(x) < 0
_
. (22.22)
For instance, for the particular initial conguration (22.21) represented by the pictures,
f

(x) =
1
3(1 + x
2
)
, and so the critical time is t

= min(3(1 + x
2
)) = 3.
Now, while mathematically plausible, such a multiply-valued solution is physically
untenable. So what happens after the critical time t

? One needs to choose which of the

possible solution values at each point (t, x) contained in the wedge is physically appropriate.
Indeed, the mathematics by itself is incapable of specifying how to continue the solution
past the critical time at which the characteristics begin to cross. We therefore must return
to the underlying physics, and ask what sort of phenomenon are we trying to model. The
most instructive is to view the dierential equation as a simple model of compressible uid
ow in a single space variable, e.g., motion of gas in a long pipe. If we push a piston
down the end of a long pipe then the gas will move ahead of the piston and thereby be
compressed. However, if the piston moves too rapidly, the gas piles up on top of itself,
and a shock wave forms and propagates down the pipe. Mathematically, the shock is
represented by a discontinuity where the solution abruptly changes value.
12/11/12 1181 c 2012 Peter J. Olver
Conservation Laws and Shocks
One way to resolve our mathematical dilemma relies on the fact that the partial
dierential equation takes the form of a conservation law, in accordance with the following
denition

.
Denition 22.4. A conservation law is an equation of the form
T
t
+
X
x
= 0. (22.23)
The functions T and X are known, respectively, as the conserved density and associated
ux.
In the simplest situations, the conserved density T(t, x, u) and ux X(t, x, u) depend
on the time t, the position x, and the solution u(t, x) to the physical system. (Higher
order conservation laws, which also depend upon derivatives of u, will appear in the nal
section.) We can clearly rewrite the nonlinear transport equation (22.13) in the following
conservation law form:
u
t
+

x
_
1
2
u
2
_
= 0, (22.24)
where the conserved density and ux are, respectively,
T = u, X =
1
2
u
2
.
The reason for calling (22.23) a conservation law comes from the following observation.
Proposition 22.5. Given a conservation law (22.23),
d
dt
_
b
a
T dx = X

b
x=a
. (22.25)
The proof of (22.25) is immediate assuming sucient smoothness that allows one to
bring the derivative inside the integral sign, and then invoking the Fundamental Theorem
of Calculus:
d
dt
_
b
a
T dx =
_
b
a
T
t
dx =
_
b
a
X
x
dx = X

b
x=a
.
Formula (22.25) says that the rate of change of the integrated density over an interval
depends only on the ux through its endpoints. In particular, if there is no net ux into
or out of the interval, then the integrated density is conserved, meaning that it remains
constant over time. All physical conservation laws mass, momentum, energy, and
so on for systems governed by partial dierential equations are of this form. (For
ordinary dierential equations, conservation laws coincide with rst integrals, as discussed
in Section 20.3.)

Here we describe the one-dimensional situation. See Exercise for conservation laws for
n-dimensional dynamics.
12/11/12 1182 c 2012 Peter J. Olver
t t + t
a = s(t)
b = s(t + t)
u
+
u

x
Figure 22.13. Conservation of Mass Near a Shock.
For the transport equation (22.24), the integrated conservation law (22.25) takes the
specic form
d
dt
_
b
a
u(t, x) dx =
1
2
_
u(t, a)
2
u(t, b)
2

. (22.26)
Viewing the equation as a model for, say, compressible uid ow in a pipe, the integral
on the left hand side represents the total mass of the uid contained in the interval [ a, b].
The right hand side represents the mass ux into the interval through its two endpoints,
and thus the conservation equation (22.26) is the mathematical formalization of basic mass
conservation mass is neither created nor destroyed, but can only enter a region as a ux
through its boundary. In particular, if there is zero mass ux, then we deduce conservation
of the total mass.
With this in hand, let us return to the physical context of the nonlinear transport
equation. We will assume that mass conservation continues to hold even within a shock,
which, from a purely molecular standpoint, makes eminent physical sense. By denition,
a shock is a jump discontinuity in the solution u(t, x). Suppose that, at time t, a shock
occurs at position x = s(t). We require

that both the left and right hand limits

u

(t) = u(t, s(t)

) = lim
xs(t)

u(t, x), u
+
(t) = u(t, s(t)
+
) = lim
x s(t)
+
u(t, x),
of the solution on either side of the shock discontinuity are well dened. Let us further
assume that, in time, the shock x = s(t) follows a smooth meaning C
1
path. Now,
referring to Figure 22.13, Consider a small time interval, from t to t + t. During this

With more analytical work, [188], the listed assumptions can all be rigorously justied.
12/11/12 1183 c 2012 Peter J. Olver
time, the shock moves from position a = s(t) to position b = s(t + t). The total mass
contained in the interval [ a, b] at time t, before the shock has passed through, is
m(t) =
_
b
a
u(t, x) dx u
+
(t) (b a) = u
+
(t)
_
s(t + t) s(t)

,
where u
+
(t) is the average value of u(t, x) over the interval. After the shock has passed,
the total mass has become
m(t + t) =
_
b
a
u(t + t, x) dx u

(t) (b a) = u

(t)
_
s(t + t) s(t)

,
where u

(t) refers to the average value of u(t +t, x) over the same interval. In the limit
as t 0, the point b = s(t + t) s(t) = a, and hence the averages
lim
t 0
u
+
(t) = u
+
(t), lim
t 0
u

(t) = u

(t),
tend to the limiting solution values on the right and left hand sides of the shock disconti-
nuity. Thus, the limiting rate of change in mass across the shock at time t is
dm
dt
= lim
t 0
m(t + t) m(t)
t
= lim
t 0
_
u

(t) u
+
(t)

s(t + t) s(t)
t
=
_
u

(t) u
+
(t)

ds
dt
,
which is the product of the shock speed times minus the jump magnitude at the shock
discontinuity. On the other hand, at any t < < t + t, the mass ux into the interval
[ a, b] is, according to the right hand side of (22.26),
1
2
_
u(, a)
2
u(, b)
2

1
2
_
u

(t)
2
u
+
(t)
2

as t 0.
For conservation of mass to hold across the shock, the limiting value of the rate of change
in mass must equal the limiting mass ux,
_
u

(t) u
+
(t)

ds
dt
=
1
2
_
u

(t)
2
u
+
(t)
2

,
from which we discover the RankineHugoniot condition
ds
dt
=
1
2
u

(t)
2
u
+
(t)
2
u

(t) u
+
(t)
=
u

(t) + u
+
(t)
2
. (22.27)
So, to maintain conservation of mass, the speed of the shock must equal the average of the
solution values on either side.
A shock appears when one or more characteristic lines cross. For this to occur, charac-
teristics to the left of the shock must have larger slope (or speed), while those to the right
must have smaller slope. Since the shock speed is the average of the two characteristic
slopes, this means
u

(t) >
ds
dt
=
u

(t) + u
+
(t)
2
> u
+
(t). (22.28)
12/11/12 1184 c 2012 Peter J. Olver
x
u
Figure 22.14. Equal Area Rule.
While it is theoretically possible to construct a shock solution to (22.13) that maintains the
RankineHugoniot constraint (22.27) but violates (22.28), such solutions are excluded on
physical grounds, in that they violate causality, [105], which requires that characteristics
are only allowed to enter shocks, not leave, and, furthermore, are not stable under small
perturbations, [188]. The dynamics of shock wave solutions is then prescribed by the
RankineHugoniot and causality conditions (22.27, 28).
How does one determine the motion of the shock in practice? The answer is beautifully
simple. Since the total mass, which at time t is the area under the curve u(t, x), must be
conserved, one merely draws the vertical shock line where the areas of the two lobes in the
multiply-valued solution are equal, as in Figure 22.14. This Equal Area Rule ensures that
the total mass of the shock solution matches that of the original (why?), as required by
the physical conservation law.
Example 22.6. An illuminating special case is when the initial data has the form
of a step function with a single jump discontinuity at the origin:
u(0, x) = f(x) = a + b (x) =
_
a, x < 0,
b, x > 0.
(22.29)
If

a > b > 0, then the initial data is already in the form of a shock wave. For t > 0, the
mathematical solution constructed by continuing along the characteristic lines is multiply-
valued in the region bt < x < at, where it assumes both values a and b; see Figure 22.15
. The Equal Area Rule tells us to draw the shock line halfway along, at x =
1
2
(a + b)t, in
order that the two triangles have the same area. Therefore, the shock moves with speed
c =
1
2
(a + b) equal to the average of the two speeds at the jump, and so this particular
shock wave solution is
u(t, x) = a + b (x ct) =
_
a, x < ct,
b, x > ct,
where a > c =
a + b
2
> b. (22.30)

Cases where a or b are negative are left to the exercises.

12/11/12 1185 c 2012 Peter J. Olver
x
u
a
b
Figure 22.15. MultiplyValued Step Wave.
t
x
Figure 22.16. Characteristic Lines for the Step Wave Shock.
A graph of the characteristic lines appears in Figure 22.16.
By way of contrast, suppose 0 < a < b, so the initial data has a jump upwards. In this
case, the characteristic lines diverge from the initial discontinuity, and the mathematical
solution is not specied at all in the wedge-shaped region at < x < bt. Now our task is
to decide how to connect the two regions where the solution is well-dened. The simplest
connection is an ane function, i.e., a straight line. Indeed, a simple modication of the
rational solution (22.17) produces the function
u(t, x) =
x
t
,
which not only solves the dierential equation, but also has the required values u(t, at) = a,
12/11/12 1186 c 2012 Peter J. Olver
Figure 22.17. Piecewise Ane Rarefaction Wave.
and u(t, bt) = b at the two edges of the wedge. The resulting solution is the piecewise ane
rarefaction wave
u(t, x) =
_
_
_
a, x at,
x/t, at x bt,
b, x bt,
(22.31)
which is graphed in Figure 22.17. In fact, it can be shown, [105], that this is the only
solution that preserves the causality condition (22.28).
These prototypical solutions epitomize the basic phenomena modeled by the nonlinear
transport equation: rarefaction waves, that emanate from regions where the initial data
satises f

(x) > 0, where the solution spreads out as time progresses, and compression
waves, emanting from regions where f

(x) < 0, that progressively steepen and eventually

break into a shock discontinuity. Anyone caught in a trac jam recognizes the com-
pression waves, where the vehicles are bunched together and almost stationary, while the
interspersed rarefaction waves correspond to freely moving trac. (An intelligent driver
will take advantage of the rarefaction waves moving through the jam to switch lanes!)
The familiar, frustrating trac jam phenomenon, even on accident- or construction-free
stretches of highway, is an intrinsic eect of the nonlinear transport model that governs
the trac ow, [188].
Our derivation of the RankineHugoniot condition (22.27) prescribing the shock speed
relies on the fact that we can write the original partial dierential equation in the form
of a conservation law. But there are other ways to do this; for instance, multiplying the
nonlinear transport equation (22.13) by u allows us write it in the alternative conservative
form
u
u
t
+ u
2
u
x
=

t
_
1
2
u
2
_
+

x
_
1
3
u
3
_
= 0. (22.32)
Here, the conserved density is T =
1
2
u
2
, and the associated ux X =
1
3
u
3
. The integral
form equation (22.25) of the conservation law is
d
dt
_
b
a
1
2
u(t, x)
2
dx =
1
3
_
u(t, a)
3
u(t, b)
3

. (22.33)
In some physical models, the integral on the left hand side represents the energy within the
interval [ a, b], and the conservation law tells us that energy can only enter the interval as
a ux through its ends. If we assume that energy is conserved at a shock, then, repeating
12/11/12 1187 c 2012 Peter J. Olver
our previous argument, we are led to the alternative condition
ds
dt
=
1
3
(u

(t)
3
u
+
(t)
3
)
1
2
(u

(t)
2
u
+
(t)
2
)
=
2
3
u

(t)
2
+ u

(t)u
+
(t) + u
+
(t)
2
u

(t) + u
+
(t)
(22.34)
for the shock speed. Thus, a shock that conserves energy moves at a dierent speed than
one that conserves mass! The evolution of a shock depends not just on the underlying
dierential equation, but also on the physical assumptions governing the selection of a
suitable entropy condition.
The mathematical property that characterizes the shock dynamics is known as an
entropy condition. Entropy conditions, such as the RankineHugoniot Equal Area Rule
(22.27), or the alternative (22.34), allow us to follow the solution beyond the formation
of a simple shock. Once a shock forms, it cannot suddenly disappear the discontinuity
remains as the solution propagates. One consequence is the irreversibility of the solutions
to the nonlinear transport equation. One cannot simply run time backwards and expect
shocks to spontaneously vanish. However, this irreversibility is of a dierent character
than that of the ill-posedness in the backwards heat equation. The nonlinear transport
equation can be solved for t < 0, but this would result, typically, in the formation of a
dierent collection of shocks, and would not be just the time reversal of the solution.
Continuing past the initial shock formation, as other characteristic lines start to cross,
additional shocks appear. The shocks themselves continue propagate, often at dierent
velocities. When a fast moving shock catches up with a slow moving shock, one must then
decide how to merge the shocks together to retain a physically consistent solution. The
selected entropy condition continues to resolve the ambiguities. However, at this point,
the mathematical details have become too complicated for us to pursue in any more detail,
and we refer the interested reader to Whithams book, [188], which includes a wide range
of applications to equations of gas dynamics, ood waves in rivers, motion of glaciers,
chromotography, trac ow, and many other physical systems.
22.2. Nonlinear Diusion.
First order partial dierential equations, beginning with elementary scalar transport
equations, and progressing on to the equations of gas dynamics, the full-blown Euler equa-
tions of uid mechanics, and yet more complicated systems for plasmas and other compli-
cated physical processes, are used to model conservative wave motion. Such systems fail to
account for frictional and/or viscous eects, which are typically modeled by a parabolic dif-
fusion equation such as the heat equation. In this section we investigate the consequences
of combining nonlinear wave motion with linear diusion by analyzing the simplest such
model. As we will see, the viscous term helps smooth out abrupt shock discontinuities, and
the result is a well-determined and smooth dynamical process. Moreover, in the inviscid
limit, as the diusion term becomes vanishingly small, the smooth viscous solutions con-
verge non-uniformly to the appropriate discontinuous shock wave, leading to an alternative
mechanism for analyzing conservative nonlinnear dynamical processes.
12/11/12 1188 c 2012 Peter J. Olver
Burgers Equation
The simplest nonlinear diusion equation is known as

Burgers equation
u
t
+ uu
x
= u
xx
, (22.35)
and is obtained by appending a linear diusion term to the nonlinear transport equation
(22.13). In uids and gases, one can interpret the right hand side as modeling the eect of
viscosity, and so Burgers equation represents a very simplied version of the equations of
viscous uid mechanics, [188]. As with the heat equation, the diusion coecient > 0
must be positive in order that initial value problem be well-posed in forwards time.
Since Burgers equation is rst order in t, we expect that its solutions are uniquely
prescribed by their initial values, say,
u(0, x) = f(x), < x < . (22.36)
(For simplicity, we will ignore boundary eects here.) Small, slowly varying solutions
more specically, those for which both | u(t, x) | and | u
x
(t, x) | are small tend to act like
solutions to the heat equation, smoothing out and decaying to 0 as time progresses. On
the other hand, when the solution is large or rapidly varying, the nonlinear term tends
to play the dominant role, and we might expect the solution to behave like the nonlinear
waves that we analyzed in Section 22.1, perhaps steepening into some sort of shock. But,
as we will see, the smoothing eect of the diusion term, no matter how small, ultimately
prevents the appearance of a discontinuous shock. Indeed, it can be proved that, under
rather mild assumptions on the initial data, the solution to the initial value problem (22.35,
36) remains smooth and well-dened for all subsequent times, [188].
The simplest explicit solutions are the traveling waves, for which
u(t, x) = v() = v(x ct), where = x ct,
indicates a xed prole, moving to the right with constant speed c. By the chain rule,
u
t
= cv

(),
u
x
= v

(),

2
u
x
2
= v

().
Substituting these expressions into Burgers equation (22.35), we conclude that v() must
satisfy the nonlinear second order ordinary dierential equation
cv

+ v v

= v

.
This equation can be solved by rst integrating both sides with respect to , and so
v

= k cv +
1
2
v
2
,

The equation is named after the applied mathematician J.M. Burgers, [36], and so the
apostrophe goes after the s. Burgers equation was apparently rst studied as a physical model
by Bateman, [14], although its solution already appears as an exercise in a nineteenth century
ordinary dierential equations text, [72; vol. 6, p. 102].
12/11/12 1189 c 2012 Peter J. Olver
= .25 = .1 = .025
Figure 22.18. Traveling Wave Solutions to Burgers Equation.
where k is a constant of integration. As in Section 20.1, the non-constant solutions to such
an autonomous rst order ordinary dierential equation tend to either or to one of
the equilibrium points, i.e., the roots of the right hand side, as t . Thus, to obtain
a bounded traveling wave solution v(), the quadratic polynomial on the right hand side
must have two real roots, which requires k <
1
2
c
2
. Assuming this holds, we rewrite the
equation in the form
2
dv
d
= (v a)(v b), where c =
1
2
(a + b). (22.37)
To obtain the bounded solutions, we concentrate on the case when a < v < b. Integrating
(22.37) by the usual method, we nd
_
2 dv
(v a)(v b)
=
2
b a
log
_
b v
v a
_
= ,
for a constant of integration, and hence
v() =
ae
(ba)()/(2)
+ b
e
(ba)()/(2)
+ 1
.
Thus, the bounded traveling wave solutions all have the explicit form
u(t, x) =
ae
(ba)(xct)/(2)
+ b
e
(ba)(xct)/(2)
+ 1
.
Observe that
lim
x
u(t, x) = b, lim
x
u(t, x) = a,
and hence our solution is a monotonically decreasing function going from b to a. The wave
travels to the right, unchanged in form, with speed equal to the average of its asymptotic
values. In Figure 22.18 we graph sample proles corresponding to a = .1, b = 1 for three
dierent values of the diusion coecient. Note that the smaller is, the sharper the
transition layer between the two asymptotic values of the solution. In the inviscid limit
0, the solutions converge to the step shock wave wave solution (22.30) to the nonlinear
transport equation, which, as a result, is often referred to as the inviscid Burgers equation.
Indeed, the profound fact is that, in the inviscid limit as the diusion becomes van-
ishingly small, 0, the solutions to Burgers equation (22.35) converge to the shock
12/11/12 1190 c 2012 Peter J. Olver
wave solution to (22.13) constructed by the Equal Area Rule. This observation is in accor-
dance with our physical intuition, that all physical systems retain a very small dissipative
component, that serves to smooth out discontinuities that might appear in a theoretical
model that fails to take the dissipation/viscosity/damping/etc. into account. In modern
theory, this so-called viscosity solution method has been successfully used to characterize
the discontinuous solutions to a broad range inviscid nonlinear wave equations is as the
limit, as the viscosity goes to zero, of classical solutions to a diusive version. Thus, the
viscosity solutions to the nonlinear transport equation resulting from Burgers equation are
consistent with the Equal Area Rule for drawing the shock discontinuities. More generally,
this method allows one to monitor the solutions as they evolve into regimes where multiple
shocks merge and interact. We refer the interested reader to [119, 188].
The HopfCole Transformation
By a remarkable stroke of luck, the nonlinear Burgers equation can be converted
into the linear heat equation and thereby explicitly solved. The linearization of Burgers
equation rst appeared in an obscure exercise in a nineteenth century dierential equations
textbook, [72; vol. 6, p. 102]. Its modern rediscovery by Eberhard Hopf, [104], and Julian
Cole, [43], was a milestone in the modern era of nonlinear partial dierential equations,
and is named the HopfCole transformation in their honor.
Finding a way to covert a nonlinear dierential equation into a linear equation is
extremely challenging, and, in, most instances, impossible. On the other hand, the re-
verse process nonlinearizing a linear equation is trivial: any nonlinear changes
of variables will do the trick! However, the resulting nonlinear equation, while evidently
linearizable through the inverse change of variables, is rarely of any independent inter-
est. Sometimes there is a lucky accident, and such accidental linearizations can have a
profound impact on our understanding of more complicated nonlinear systems.
In the present context, our starting point is the linear heat equation
v
t
= v
xx
. (22.38)
Among all possible nonlinear changes of dependent variable, one of the simplest that might
spring to mind is an exponential function. Let us, therefore, investigate the eect of an
exponential change of variables
v(t, x) = e
(t,x)
, so (t, x) =
1

log v(t, x), (22.39)

where is a nonzero constant. The function (t, x) is real provided v(t, x) > 0 is a positive
solution to the heat equation. Fortunately, this is not hard to arrange: if the initial data
v(0, x) > 0 is strictly positive, then the resulting solution v(t, x) is positive for all t > 0.
This follows from the Maximum Principle for the heat equation, cf. Theorem 14.3.
To determine the dierential equation satised by the function , we invoke the chain
rule to dierentiate (22.39):
v
t
=
t
e

, v
x
=
x
e

, v
xx
=
_

xx
+
2

2
x
_
e

.
12/11/12 1191 c 2012 Peter J. Olver
Substituting the rst and last formulae into the heat equation (22.38) and canceling a com-
mon exponential factor, we conclude that (t, x) satises the nonlinear partial dierential
equation

t
=
xx
+
2
x
, (22.40)
known as the potential Burgers equation, for reasons that will soon become apparent.
The second step in the process is to dierentiate the potential Burgers equation with
respect to x; the result is

tx
=
xxx
+ 2
x

xx
. (22.41)
If we now set

x
= u, (22.42)
so that has the status of a potential function, then the resulting partial dierential
equation
u
t
= u
xx
+ 2 uu
x
coincides with Burgers equation (22.35) with = 1/(2). In this manner, we have
arrived at the famous HopfCole transformation.
Theorem 22.7. If v(t, x) > 0 is any positive solution to the linear heat equation
v
t
= v
xx
, then
u(t, x) =

x
_
2 log v(t, x)
_
= 2
v
x
v
(22.43)
solves Burgers equation u
t
+ uu
x
= u
xx
.
Do all solutions to Burgers equation arise in this way? In order to decide, we run the
argument in reverse. First, choose a potential function (t, x) that satises (22.42); for
example
(t, x) =
_
x
0
u(t, y) dy.
If u(t, x) is any solution to Burgers equation, then (t, x) satises (22.41). Integrating
both sides of the latter equation with respect to x, we conclude that

t
=
xx
+
2
x
+ h(t),
for some integration constant h(t). Thus, unless h(t) 0, our potential function
doesnt satisfy the potential Burgers equation (22.40), but thats because we chose the
wrong potential. Indeed, if we dene
(t, x) = (t, x) (t), where

(t) = h(t),
then

t
=
t
h(t) =
xx
+
2
x
=
xx
+
2
x
,
and hence the modied potential (t, x) is a solution to the potential Burgers equation
(22.40). From this it easily follows that
v(t, x) = e
(t,x)/(2)
(22.44)
12/11/12 1192 c 2012 Peter J. Olver
-7.5 -5 -2.5 2.5 5 7.5
-2
-1.5
-1
-0.5
0.5
1
1.5
2
-7.5 -5 -2.5 2.5 5 7.5
-2
-1.5
-1
-0.5
0.5
1
1.5
2
-7.5 -5 -2.5 2.5 5 7.5
-2
-1.5
-1
-0.5
0.5
1
1.5
2
-7.5 -5 -2.5 2.5 5 7.5
-2
-1.5
-1
-0.5
0.5
1
1.5
2
Figure 22.19. A Solution to Burgers Equation.
is a positive solution to the heat equation, from which u(t, x) can be recovered via the
Hopf Cole transformation (22.43). Thus, we have proved that every solution to Burgers
equation comes from a positive solution to the heat equation via the HopfCole transfor-
mation.
Example 22.8. As a simple example, the separable solution
v(t, x) = a + b e

2
t
cos x
to the heat equation leads to the solution
u(t, x) =
2 b e

2
t
sin x
a + b e

2
t
cos x
to Burgers equation; a typical example is plotted in Figure 22.19. We should require that
a > | b | in order that v(t, x) > 0 be a positive solution to the heat equation for t 0;
otherwise the resulting solution to Burgers equation will have singularities at the roots of
u see the rst graph in Figure 22.19. This particular solution primarily feels the eects
of the diusivity, and rapidly goes to zero.
To solve the initial value problem (22.3536) for Burgers equation, we note that,
under the HopfCole transformation,
v(0, x) = h(x) = exp
_

(0, x)
2
_
= exp
_

1
2
_
x
0
f(y) dy
_
, (22.45)
Remark: The lower limit of the integral can be changed from 0 to any other convenient
value without aecting the nal form of u(t, x) in (22.43). The only eect is to multiply
v(t, x) by an overall constant, which does not change u(t, x).
12/11/12 1193 c 2012 Peter J. Olver
-2 -1 1 2
0.2
0.4
0.6
0.8
1
-2 -1 1 2
0.2
0.4
0.6
0.8
1
-2 -1 1 2
0.2
0.4
0.6
0.8
1
Figure 22.20. Shock Wave Solution to Burgers Equation.
According to formula (14.61) (adapted to general diusivity, as in Exercise ), the
solution to the initial value problem (22.38, 45) for the heat equation can be expressed as
a convolution integral with the fundamental solution:
v(t, x) =
1
2

t
_

e
(xy)
2
/(4 t)
h(y) dy.
Therefore, the solution to the Burgers initial value problem (22.35, 36) is
u(t, x) =
_

x y
t
e
F(t,x,y)
dy
_

e
F(t,x,y)
dy
where F(t, x, y) =
1
2
_
y
0
f(z) dz
(x y)
2
4 t
.
(22.46)
Example 22.9. To demonstrate the smoothing eect of the diusion terms, let us
see what happens to the initial data
u(0, x) =
_
a, x < 0,
b, x > 0,
(22.47)
in the form of a step function. We assume that a > b, which would correspond to a shock
wave in the inviscid limit = 0. (In Exercise , the reader is asked to analyze the case
a < b which corresponds to a rarefaction wave.) In this case,
F(t, x, y) =
(x y)
2
4 t

_

ay
2
, y < 0,

by
2
, y > 0.
After some algebraic manipulations, the solution is found to have the explicit form
u(t, x) = a +
b a
1 + h(t, x) exp
b a
2
(x ct)
(22.48)
where
c =
a + b
2
, h(t, x) =
1 erf
_
x bt

4 t
_
1 erf
_
x at

4 t
_ , (22.49)
12/11/12 1194 c 2012 Peter J. Olver
-2 2 4 6 8 10 12
0.2
0.4
0.6
0.8
1
1.2
-2 2 4 6 8 10 12
0.2
0.4
0.6
0.8
1
1.2
-2 2 4 6 8 10 12
0.2
0.4
0.6
0.8
1
1.2
-2 2 4 6 8 10 12
0.2
0.4
0.6
0.8
1
1.2
Figure 22.21. Triangular Wave Solution to Burgers Equation.
and erf z denotes the error function (14.63). The solution, with a = 1, b = .1 and = .03
is plotted in Figure 22.20 at times t = .01, 1.0, 2.0. Note that the sharp transition region
for the shock is immediately smoothed, and the solution rapidly settles into the form of a
continuously varying transition layer between the two step heights. The larger the diusion
coecient in relation to the initial solution heights a, b, the more signicant the smoothing
eect. Observe that, as 0, the function h(t, x) 1, and hence the solution converges
to the shock wave solution (22.30) to the transport equation, in which the speed of the
shock is the average of the two initial values.
Example 22.10. Consider the case when the initial data u(0, x) = (x) is a concen-
trated delta function impulse at the origin. In the solution formula (22.46), starting the
integral for F(t, x, y) at 0 is problematic, but as noted earlier, we are free to select any
other starting point, e.g., . Thus, we take
F(t, x, y) =
1
2
_
y

(z) dz
(x y)
2
4 t
=
_

(x y)
2
4 t
, y < 0,

1
2

(x y)
2
4 t
, y > 0.
Substituting this into (22.46), we can evaluate the upper integral in elementary terms,
while the lower integral involves the error function (14.63); after a little algebra, we nd
u(t, x) =
_
4
t
e
x
2
/(4 t)
coth
_
1
4
_
erf
_
x

4 t
_ , (22.50)
where
coth z =
cosh z
sinh z
=
e
z
+ e
z
e
z
e
z
=
e
2z
+ 1
e
2z
1
12/11/12 1195 c 2012 Peter J. Olver
is the hyperbolic cotangent function. A graph of this solution when = .02 and a = 1,
at times t = 1, 5, 10, 50, appears in Figure 22.21. As you can see, the initial concentration
diuses out, but, unlike the heat equation, the wave does not remain symmetric owing to
the advection terms in the equation. The eect is to steepen in front as it propagates.
Eventually the triangular wave spreads out as the diusion progresses.
22.3. Dispersion and Solitons.
Finally, we study a remarkable third order evolution equation that originally arose
in the modeling of surface water waves, that serves to introduce yet further phenomena,
both linear and nonlinear. The third order derivative models dispersion, in which waves
of dierent frequencies move at dierent speeds. Coupled with the same nonlinearity as
in the inviscid and viscous Burgers (22.13, 35), the result is one of the most remarkable
equations in all of mathematics, with far-reaching implications, not only in uid mechanics
and applications, but even in complex function theory, physics, etc., etc.
Linear Dispersion
So far, in our study of partial dierential equations, we have not ventured beyond
second order. Higher order equations do occur in applications, particularly in models for
wave motion. The simplest linear partial dierential equation of a type that we have not
yet considered is the third order equation
u
t
+ u
xxx
= 0 (22.51)
It is the third in a hierarchy of simple evolution equations that starts with the simple
ordinary dierential equation u
t
= u, then proceeds to the transport equation u
t
= u
x
,
and then the heat equation u
t
= u
xx
modeling basic diusion processes. The third order
case (22.51) is a simple model for linear dispersive waves.
To avoid additional complications caused by boundary conditions, we shall only look
at the equation on the entire line, so x R The solution to the equation is uniquely
specied by initial data
u(0, x) = f(x), < x < . (22.52)
See [1] for a proof.
Let us apply the Fourier transform to solve the initial value problem. Let
u(t, k) =
1

2
_

u(t, x) e
i kx
dx
be the spatial Fourier transform of the solution, which is assumed to remain in L
2
at all t,
a fact that can be justied a posteriori. In view of the eect of the Fourier transform on
derivatives see Corollary 13.22 the Fourier transform converts the partial dierential
equation (22.51) into a rst order, linear ordinary dierential equation
u
t
i k
3
u = 0, (22.53)
12/11/12 1196 c 2012 Peter J. Olver
parametrized by k, with initial conditions
u(0, k) =

f(k) =
1

2
_

f(x) e
i kx
dx (22.54)
given by the Fourier transform of (22.52). Solving the initial value problem (22.5354) by
the usual technique, we nd
u(t, k) = e
i k
3
t

f(k).
Inverting the Fourier transform yields the explicit formula for the solution
u(t, x) =
1

2
_

e
i k
3
t+i kx

f(k) dk (22.55)
to the initial value problem for the dispersive wave equation (22.5152).
Actually, to nd the solutions to the dierential equation, one does not need the full
power of the Fourier transform. Note that (22.55) represents a linear superposition of
elementary exponential functions. Let us substitute an exponential ansatz
u(t, x) = e
i t+i kx
(22.56)
representing a complex oscillatory wave of frequency , which indicates the time vibrations,
and wave number k, which indicates the corresponding oscillations in space. Since
u
t
= i e
i t+i kx
,

3
u
x
3
= i k
3
e
i t+i kx
,
(22.56) satises the partial dierential equation (22.51) if and only if its frequency and
wave number are related by
= k
3
. (22.57)
The result is known as the dispersion relation for the partial dierential equation. In
general, any linear constant coecient dynamical partial dierential equation admits a
dispersion relation of the form = (k) which is straightforwardly found by substituting
the exponential ansatz (22.56) and canceling the common exponential factors in the re-
sulting equation. In our particular case, the exponential solution of wave number k has
the form
u
k
(t, x) = e
i k
3
t+i kx
.
Linear superposition permits us to combine them in integral form, and so, for any (rea-
sonable) function a(k) depending on the wave number,
u(t, x) =
1

2
_

e
i k
3
t+i kx
a(k) dk
is easily seen to be a solution to the partial dierential equation. The Fourier transform
solution (22.55) has this form.
12/11/12 1197 c 2012 Peter J. Olver
Example 22.11. The fundamental solution corresponds to a concentrated initial
disturbance
u(0, x) = (x).
since the Fourier transform of the delta function is just

(k) = 1/

2 , the resulting
solution (22.55) is
u(t, x) =
1
2
_

e
i k
3
t+i kx
dk =
1

_

0
cos(k
3
t + kx) dk,
since the solution is real (or, equivalently, the imaginary part of the integrand is odd)
while the real part of the integrand is even. The second integral can be converted into
that dening the Airy function,
Ai(z) =
1

_

0
cos
_
sz +
1
3
s
3
_
ds,
as in (C.40), by the change of variables
s = k
3

3t, z =
x
3

3t
,
and we conclude that the fundamental solution to the dispersive wave equation (22.51)
can be written in terms of the Airy function:
u(t, x) =
1
3

3t
Ai
_
x
3

3t
_
.
See Figure ee3 for a graph. Furthermore, writing the general initial data as a superposition
of delta functions
f(x) =
_

f() d (x ),
we conclude that the solution has the form
u(t, x) =
1
3

3t
_

f() Ai
_
x
3

3t
_
d. (22.58)
Although energy is conserved, unlike the heat and diusion equations, the dispersion
of waves means that the solution dies out.
Group velocity and wave velocity.
The KortewegdeVries Equation
The simplest wave equation that combines dispersion with nonlinearity is the cele-
brated KortewegdeVries equation
u
t
+ u
xxx
+ uu
x
= 0. (22.59)
The equation was rst derived by the French applied mathematician Boussinesq, [24;
eq. (30)], [25; eqs. (283, 291)], in 1872 as a model for surface water waves. It was redis-
covered by the Dutch mathematician Korteweg and his student de Vries, [120], over two
12/11/12 1198 c 2012 Peter J. Olver
decades later, and, despite Boussinesqs priority, is named after them. In the early 1960s,
the American mathematical physicists Martin Kruskal and Norman Zabusky, [195], red-
erived it as a continuum limit of a model of nonlinear mass-spring chains studied by Fermi,
Pasta and Ulam, [67]. Understanding the puzzling behavior of both systems coming from
numerical experiments was the catalyst of one of the most remarkable and far-ranging
discoveries of modern mathematics: integrable nonlinear partial dierential equations.
The most important special solutions to the KortewegdeVries equation are the trav-
eling waves. We assume that the solution
u = v() = v(x ct), where = x ct,
is a wave of permanent form, translating to the right with speed c. By the chain rule,
u
t
= cv

(),
u
x
= v

(),

3
u
x
3
= v

().
Substituting these expressions into the KortewegdeVries equation (22.59), we conclude
that v() must satisfy the nonlinear third order ordinary dierential equation
v

+ v v

cv

= 0. (22.60)
Let us further assume that the traveling wave is localized, meaning that the solution and
its derivatives are small at large distances:
lim
x
u(t, x) = lim
x
u
x
(t, x) = lim
x

2
u
x
2
(t, x) = 0.
To this end, we impose the boundary conditions
lim

v() = lim

v

() = lim

v

() = 0. (22.61)
(See Exercise for an analysis of the non-localized traveling wave solutions.)
The ordinary dierential equation (22.60) can, in fact be solved in closed form. First,
note that
d
d
_
v

+
1
2
v
2
cv

= 0, and hence v

+
1
2
v
2
cv = k,
is a rst integral, with k indicating the constant of integration. However, the localizing
boundary conditions (22.61) imply that k = 0. Multiplying the latter equation by v

allows
us to integrate a second time
d
d
_
1
2
(v

)
2
+
1
6
v
3

1
2
cv
2

= v

_
v

+
1
2
v
2
cv

= 0.
Thus,
1
2
(v

)
2
+
1
6
v
3

1
2
cv
2
= ,
where is a second constant of integration, which, again by the boundary conditions
(22.61), is also = 0. We conclude that v() satises the rst order autonomous ordinary
dierential equation
dv
d
= v
_
c
1
3
v .
12/11/12 1199 c 2012 Peter J. Olver
Figure 22.22. Solitary Wave.
We integrate by the usual method, cf. (20.7):
_
dv
v
_
c
1
3
v
= + .
Using a table of integrals, and then solving for v, we conclude that the solution has the
form
v() = 3c sech
2
_
1
2

c +

,
where
sech y =
1
cosh y
=
2
e
y
+ e
y
,
is the hyperbolic secant function. The solution has the form graphed in Figure 22.22; it has
a global maximum at 3c sech 0 = 3c at y = 0, and is an even function, exponentially decay
to 0 as | | . The resulting localized traveling wave solutions to the KortewegdeVries
equation are
u(t, x) = 3c sech
2
_
1
2

c (x ct) +

, (22.62)
where c > 0 and are arbitrary constants. The parameter c equals the speed of the wave.
It is also equal to one third its amplitude, since the maximum value of u(t, x) is 3c at the
points x = ct, as well as the width, which is on the order of

c . The taller and wider the
solitary wave, the faster it moves.
The solution (22.62) is known as a solitary wave solution since it represents a localized
wave that travels unchanged in shape. Such waves were rst observed by the British
engineer J. Scott Russell, [165], who recounts how such a wave was generated by the
sudden motion of a barge along an Edinburgh canal and then chasing it on horseback for
several miles. Russells observations were dismissed by his contemporary, the prominent
mathematician George Airy, who claimed that such localized disturbances could not exist,
basing his analysis upon a linearized theory. Much later, Boussinesq established the proper
nonlinear surface wave model (22.59), valid for long waves in shallow water, and also derived
the solitary wave solution (22.62), thereby fully exonerating Scott Russells insight.
These nonlinear traveling wave solutions were discovered by Kruskal and Zabusky,
[195], to have remarkable properties. For this reason they have been given a special new
name soliton. Ordinarily, combining two solutions to a nonlinear equation can be quite
12/11/12 1200 c 2012 Peter J. Olver
Figure 22.23. Interaction of Two Solitons.
unpredictable, and one might expect any number of scenarios to occur. If you start with
initial conditions representing a taller wave to the left of a shorter wave, the solution
of the KortewegdeVries equation runs as follows. The taller wave moves faster, and so
catches up the shorter wave. They then have a very complicated nonlinear interaction, as
expected. But, remarkably, after a while they emerge from the interaction unscathed. The
smaller wave is now in back and the larger one in front. After this, they proceed along
their way, with the smaller one lagging behind the high speed tall wave. the only eect
12/11/12 1201 c 2012 Peter J. Olver
of their encounter is a phase shift, meaning a change in the value of the phase parameter
in each wave. See Figure 22.23. After the interaction, the position of the soliton if it
had traveled unhindered by the other is shown in a dotted line. Thus, they behave like
colliding particles, which is the genesis of the word soliton.
A similar phenomenon holds for several such soliton solutions. After some time where
the various waves interact, they nally emerge with the largest soliton in front, and then
in order to the smallest one in back, all progressing at their own speed, and so gradually
drawing apart.
Remark: In the KortewegdeVries equation model, one can nd arbitrarily tall soliton
solutions. In physical water waves, if the wave is too tall it will break. Indeed, it can be
rigorously proved that the full water wave equations admit solitary wave solutions, but
there is a wave of greatest height, beyond which a wave will tend to break. The solitary
water waves are not genuine solitons, since there is a small, but measurable, eect when
two waves collide.
Moreover, it can be proved that, starting with an arbitrary initial disturbance
u(0, x) = f(x)
that decays suciently rapidly as | x | , after a suciently long time, the resulting
solution u(t, x) disintegrates into a nite number of solitons of dierent heights, moving
o at their respective speeds to the right, and so are arranged in order from smallest to
largest, plus a small dispersive tail moving to the left that rapidly disappears. Proving this
remarkable result is beyond the scope of this book. It relies on the method of inverse scat-
tering, that eectively linearizes the KortewegdeVries equation with a linear eigenvalue
problem of fundamental importance in one-dimensional quantum mechanics. The solitons
correspond to the bound states of a quantum potential. We refer the interested reader to
the introductory text [61] and the more advanced monograph [1] for details.
Like Burgers equation, the KortewegdeVries equation can be linearized, but the
linearization is considerably more subtle. It relies on the introduction of an auxiliary
linear eigenvalue problem.
There is a remarkable transformation, known as the inverse scattering transform,
which is a form of nonlinear Fourier transform, that can be used to solve the Korteweg
deVries equation. Its fascinating properties continue to be of great current research interest
to this day.
22.4. Conclusion and Bon Voyage.
These are your rst wee steps in a vast new realm. We are unable to discuss nonlinear
partial dierential equations arising in uid mechanics, in elasticity, in relativity, in dif-
ferential geometry, in computer vision, in mathematical biology. Chaos and integrability
are the two great themes in modern nonlinear applied mathematics, and the student is
well-advised to pursue both.
We bid you, dear reader, a fond adieu and wish you unparalleled success in your
mathematical endeavors.
12/11/12 1202 c 2012 Peter J. Olver