MATH3414
School of Mathematics, University of Leeds
15 credits
Taught Semester 1,
Year running 2003/04
Prerequisites MATH2360 or MATH2420 or equivalent.
Corequisites None.
Objectives: To provide an understanding of, and methods of solution for, the most important
types of partial differential equations that arise in Mathematical Physics. On completion
of this module, students should be able to: a) use the method of characteristics to solve
firstorder hyperbolic equations; b) classify a second order PDE as elliptic, parabolic or
hyperbolic; c) use Greens functions to solve elliptic equations; d) have a basic understanding
of diffusion; e) obtain a priori bounds for reactiondiffusion equations.
Syllabus: The majority of physical phenomena can be described by partial differential equations (e.g. the NavierStokes equation of fluid dynamics, Maxwells equations of electromagnetism). This module considers the properties of, and analytical methods of solution for
some of the most common first and second order PDEs of Mathematical Physics. In particular, we shall look in detail at elliptic equations (Laplace?s equation), describing steadystate
phenomena and the diffusion / heat conduction equation describing the slow spread of concentration or heat. The topics covered are: First order PDEs. Semilinear and quasilinear
PDEs; method of characteristics. Characteristics crossing. Second order PDEs. Classification and standard forms. Elliptic equations: weak and strong minimum and maximum
principles; Greens functions. Parabolic equations: exemplified by solutions of the diffusion
equation. Bounds on solutions of reactiondiffusion equations.
Form of teaching
Lectures: 26 hours. 7 examples classes.
Form of assessment
One 3 hour examination at end of semester (100%).
ii
Details:
Evy Kersal
e
Office: 9.22e
Phone: 0113 343 5149
Email: kersale@maths.leeds.ac.uk
WWW: http://www.maths.leeds.ac.uk/kersale/
Schedule: three lectures every week, for eleven weeks (from 27/09 to 10/12).
Tuesday
13:0014:00 RSLT 03
Wednesday 10:0011:00 RSLT 04
Friday
11:0012:00 RSLT 06
Prerequisite: elementary differential calculus and several variables calculus (e.g. partial
differentiation with change of variables, parametric curves, integration), elementary algebra (e.g. partial fractions, linear eigenvalue problems), ordinary differential equations (e.g.
change of variable, integrating factor), and vector calculus (e.g. vector identities, Greens
theorem).
Outline of course:
Introduction:
definitions
examples
First order PDEs:
linear & semilinear
characteristics
quasilinear
nonlinear
system of equations
Second order linear PDEs:
classification
elliptic
parabolic
Book list:
P. Prasad & R. Ravindran, Partial Differential Equations, Wiley Eastern, 1985.
W. E. Williams, Partial Differential Equations, Oxford University Press, 1980.
P. R. Garabedian, Partial Differential Equations, Wiley, 1964.
Thanks to Prof. D. W. Hughes, Prof. J. H. Merkin and Dr. R. Sturman for their lecture
notes.
Course Summary
Definitions of different type of PDE (linear, quasilinear, semilinear, nonlinear)
Existence and uniqueness of solutions
Solving PDEs analytically is generally based on finding a change of variable to transform
the equation into something soluble or on finding an integral form of the solution.
Linear equations:
u
u
+b
= c.
x
y
dy
=b
ds
and
du
=c
ds
region of solution.
2u
2u
u
2u
u
+
2b
+d
+
c
+e
+ f u = g.
2
2
x
xy
y
x
y
iii
iv
Classification
Type
Canonical form
b2 ac > 0
Hyperbolic
2u + . . . = 0
b2 ac = 0
Parabolic
2u + . . . = 0
2
Elliptic
2u + 2u + . . . = 0
2 2
b2
ac < 0
Characteristics
p
dy
b2 ac
b
=
a
dx
dy
= ab , = x (say)
dx
dy
= b
dx
p
b2 ac , = +
a
= i( )
Contents
1 Introduction
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . .
1.2 Reminder . . . . . . . . . . . . . . . . . . . . . . .
1.3 Definitions . . . . . . . . . . . . . . . . . . . . . . .
1.4 Examples . . . . . . . . . . . . . . . . . . . . . . .
1.4.1 Wave Equations . . . . . . . . . . . . . . .
1.4.2 Diffusion or Heat Conduction Equations . .
1.4.3 Laplaces Equation . . . . . . . . . . . . . .
1.4.4 Other Common Second Order Linear PDEs
1.4.5 Nonlinear PDEs . . . . . . . . . . . . . . .
1.4.6 System of PDEs . . . . . . . . . . . . . . .
1.5 Existence and Uniqueness . . . . . . . . . . . . . .
2 First Order Equations
2.1 Linear and Semilinear Equations . . . . . . .
2.1.1 Method of Characteristic . . . . . . .
2.1.2 Equivalent set of ODEs . . . . . . . .
2.1.3 Characteristic Curves . . . . . . . . .
2.2 Quasilinear Equations . . . . . . . . . . . . .
2.2.1 Interpretation of Quasilinear Equation
2.2.2 General solution: . . . . . . . . . . . .
2.3 Wave Equation . . . . . . . . . . . . . . . . .
2.3.1 Linear Waves . . . . . . . . . . . . . .
2.3.2 Nonlinear Waves . . . . . . . . . . . .
2.3.3 Weak Solution . . . . . . . . . . . . .
2.4 Systems of Equations . . . . . . . . . . . . . .
2.4.1 Linear and Semilinear Equations . . .
2.4.2 Quasilinear Equations . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1
1
2
3
3
4
4
4
5
5
6
.
.
.
.
.
.
.
.
.
.
.
.
.
.
9
9
9
12
14
19
19
20
26
26
27
29
31
31
34
.
.
.
.
37
37
44
44
45
ii
CONTENTS
4 Elliptic Equations
4.1 Definitions . . . . . . . . . . . . . . . . . . . . . . .
4.2 Properties of Laplaces and Poissons Equations . .
4.2.1 Mean Value Property . . . . . . . . . . . .
4.2.2 MaximumMinimum Principle . . . . . . . .
4.3 Solving Poisson Equation Using Greens Functions
4.3.1 Definition of Greens Functions . . . . . . .
4.3.2 Greens function for Laplace Operator . . .
4.3.3 Free Space Greens Function . . . . . . . .
4.3.4 Method of Images . . . . . . . . . . . . . .
4.4 Extensions of Theory: . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
49
49
50
51
52
54
55
55
60
61
68
5 Parabolic Equations
5.1 Definitions and Properties . . . . . . . . . . . . . . . . . . . . . . .
5.1.1 WellPosed Cauchy Problem (Initial Value Problem) . . . .
5.1.2 WellPosed InitialBoundary Value Problem . . . . . . . . .
5.1.3 Time Irreversibility of the Heat Equation . . . . . . . . . .
5.1.4 Uniqueness of Solution for Cauchy Problem: . . . . . . . . .
5.1.5 Uniqueness of Solution for InitialBoundary Value Problem:
5.2 Fundamental Solution of the Heat Equation . . . . . . . . . . . . .
5.2.1 Integral Form of the General Solution . . . . . . . . . . . .
5.2.2 Properties of the Fundamental Solution . . . . . . . . . . .
5.2.3 Behaviour at large t . . . . . . . . . . . . . . . . . . . . . .
5.3 Similarity Solution . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3.1 Infinite Region . . . . . . . . . . . . . . . . . . . . . . . . .
5.3.2 SemiInfinite Region . . . . . . . . . . . . . . . . . . . . . .
5.4 Maximum Principles and Comparison Theorems . . . . . . . . . .
5.4.1 First Maximum Principle . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
69
69
69
70
70
71
71
72
73
74
75
75
76
77
78
79
A Integral of ex in R
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
81
Chapter 1
Introduction
Contents
1.1
1.2
1.3
1.4
1.5
1.1
Motivation .
Reminder . .
Definitions . .
Examples . .
Existence and
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
Uniqueness
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1
2
3
6
Motivation
1.2
Reminder
n
X
f
f
f
f
dxi =
dx1 +
dx2 + . . . +
dxn ,
xi
x1
x2
xn
i=1
1.3 Definitions
The usual differentiation identities apply to the partial differentiations (sum, product, quotient, chain rules, etc.)
Notations: I shall use interchangeably the notations
f
xi f f xi ,
xi
2f
x2i xj f fxi xj ,
xi xj
for the first order and second order partial derivatives respectively. We shall also use interchangeably the notations
~u u u,
for vectors.
Vector differential operators: in three dimensional Cartesian coordinate system (i, j, k)
we consider f (x, y, z) : R3 R and [ux (x, y, z), uy (x, y, z), uz (x, y, z)] : R3 R3 .
Gradient: f = x f i + y f j + z f k.
Divergence: div u u = x ux + y uy + z uz.
Curl: u = (z uy y uz ) i + (z ux x uz ) j + (x uy y ux ) k.
Laplacian: f 2 f = x2 f + y2 f + z2 f.
Laplacian of a vector: u 2 u = 2 ux i + 2 uy j + 2 uz k.
Note that these operators are different in other systems of coordinate (cylindrical or spherical,
say)
1.3
Definitions
A partial differential equation (PDE) is an equation for some quantity u (dependent variable)
which depends on the independent variables x 1 , x2 , x3 , . . . , xn , n 2, and involves derivatives
of u with respect to at least some of the independent variables.
F (x1 , . . . , xn , x1 u, . . . , xn u, x21 u, x21 x2 u, . . . , xn1 ...xn u) = 0.
Note:
1. In applications xi are often space variables (e.g. x, y, z) and a solution may be required
in some region of space. In this case there will be some conditions to be satisfied on
the boundary ; these are called boundary conditions (BCs).
2. Also in applications, one of the independent variables can be time (t say), then there
will be some initial conditions (ICs) to be satisfied (i.e., u is given at t = 0 everywhere
in )
3. Again in applications, systems of PDEs can arise involving the dependent variables
u1 , u2 , u3 , . . . , um , m 1 with some (at least) of the equations involving more than one
ui .
Chapter 1 Introduction
The order of the PDE is the order of the highest (partial) differential coefficient in the
equation.
As with ordinary differential equations (ODEs) it is important to be able to distinguish
between linear and nonlinear equations.
A linear equation is one in which the equation and any boundary or initial conditions do not
include any product of the dependent variables or their derivatives; an equation that is not
linear is a nonlinear equation.
u
u
+c
= 0,
t
x
2u 2u
+ 2 = (x, y),
x2
y
A nonlinear equation is semilinear if the coefficients of the highest derivative are functions of
the independent variables only.
(x + 3)
u
u
+ xy 2
= u3 ,
x
y
2
2u
u
u
2 u
+
(xy
+
y
)
+u
+ u2
= u4 .
2
2
x
y
x
y
1+
u
y
2 #
"
2 # 2
u
2u
u
u u 2 u
2
+ 1+
= 0.
2
x
x y xy
x
y 2
Principle of superposition: A linear equation has the useful property that if u 1 and u2
both satisfy the equation then so does u 1 + u2 for any , R. This is often used in
constructing solutions to linear equations (for example, so as to satisfy boundary or initial
conditions; c.f. Fourier series methods). This is not true for nonlinear equations, which helps
to make this sort of equations more interesting, but much more difficult to deal with.
1.4
1.4.1
Examples
Wave Equations
Waves on a string, sound waves, waves on stretch membranes, electromagnetic waves, etc.
1 2u
2u
=
,
x2
c2 t2
or more generally
1 2u
= 2 u
c2 t2
where c is a constant (wave speed).
1.4.2
1.4 Examples
or more generally
u
= 2 u,
t
or even
u
= (u)
t
where is a constant (diffusion coefficient or thermometric conductivity).
Both those equations (wave and diffusion) are linear equations and involve time (t). They
require some initial conditions (and possibly some boundary conditions) for their solution.
1.4.3
Laplaces Equation
1.4.4
Poissons equation is just the Lapaces equation (homogeneous) with a known source term
(e.g. electric potential in the presence of a density of charge):
2 u = .
The Helmholtz equation may be regarded as a stationary wave equation:
2 u + k 2 u = 0.
The Schrodinger equation is the fundamental equation of physics for describing quantum mechanical behavior; Schrodinger wave equation is a PDE that describes how the wavefunction
of a physical system evolves over time:
2 u + V u = i
u
.
t
Chapter 1 Introduction
1.4.5
Nonlinear PDEs
(1st order).
The equation
x2 u
u
u
+ (y + u)
= u3
x
y
u
u
+ (x3 + y)
= u3
x
y
1.4.6
System of PDEs
B = j +
B = 0,
E =
1 E
,
c2 t
B
.
t
In empty space (free of charges and currents) this system can be rearranged to give the
equations of propagation of the electromagnetic field,
2E
= c2 2 E,
t2
2B
= c2 2 B.
t2
Incompressible magnetohydrodynamic (MHD) equations combine NavierStokes equation (including the Lorentz force), the induction equation as well as the solenoidal constraints,
U
+ U U = + B B + 2 U + F,
t
B
= (U B) + 2 B,
t
U = 0,
B = 0.
Both systems involve space and time; they require some initial and boundary conditions for
their solution.
1.5
Before attempting to solve a problem involving a PDE we would like to know if a solution
exists, and, if it exists, if the solution is unique. Also, in problem involving time, whether
a solution exists t > 0 (global existence) or only up to a given value of t i.e. only for
0 < t < t0 (finite time blowup, shock formation). As well as the equation there could be
certain boundary and initial conditions. We would also like to know whether the solution of
the problem depends continuously of the prescribed data i.e. small changes in boundary
or initial conditions produce only small changes in the solution.
Illustration from ODEs:
1.
du
= u,
dt
u(0) = 1.
du
= u2 ,
dt
u(0) = 1.
du
= u,
dt
u(0) = 0,
Z
=
w2 w + (w)2 dV
Chapter 1 Introduction
(w)2
(w) dV =
w
dS = 0.
n
Chapter 2
2.1
2.1.1
Equations
. . . . . . .
. . . . . . .
. . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
9
19
26
31
We consider linear first order partial differential equation in two independent variables:
a(x, y)
u
u
+ b(x, y)
+ c(x, y)u = f (x, y),
x
y
(2.1)
where a, b, c and f are continuous in some region of the plane and we assume that a(x, y)
and b(x, y) are not zero for the same (x, y).
In fact, we could consider semilinear first order equation (where the nonlinearity is present
only in the righthand side) such as
a(x, y)
u
u
+ b(x, y)
= (x, y, u),
x
y
(2.2)
instead of a linear equation as the theory of the former does not require any special treatment
as compared to that of the latter.
The key to the solution of the equation (2.1) is to find a change of variables (or a change of
coordinates)
(x, y), (x, y)
which transforms (2.1) into the simpler equation
w
+ h(, )w = F (, )
(2.3)
10
We shall define this transformation so that it is onetoone, at least for all (x, y) in some set
D of points in the (xy) plane. Then, on D we can (in theory) solve for x and y as functions
of , . To ensure that we can do this, we require that the Jacobian of the transformation
does not vanish in D:
x y
J = =
6= {0, }
x y y x
x
for (x, y) in D. We begin looking for a suitable transformation by computing derivatives via
the chain rule
w
w
u
w
w
u
=
+
and
=
+
.
x
x
x
y
y
y
w
a
+b
+ a
+b
+ cw = f.
x
y
x
y
(2.4)
This is close to the form of equation (2.1) if we can choose (x, y) so that
a
+b
=0
x
y
for
(x, y)
in
D.
d =
dx +
dy = 0
x
y
implies that dy/dx = x /y = b/a. So, the equation (x, y) = k defines solutions of the
ODE
b(x, y)
dy
=
.
(2.5)
dx
a(x, y)
Equation (2.5) is called the characteristic equation of the linear equation (2.1). Its solution
can be written in the form F (x, y, ) = 0 (where is the constant of integration) and defines a
family of curves in the plane called characteristics or characteristic curves of (2.1). (More on
characteristics later.) Characteristics represent curves along which the independent variable
of the new coordinate system (, ) is constant.
So, we have made the coefficient of w/ vanish in the transformed equation (2.4), by
choosing (x, y), with (x, y) = k an equation defining the solution of the characteristic
equation (2.5). We can now choose arbitrarily (or at least to suit our convenience), providing
we still have J 6= 0. An obvious choice is
(x, y) = x.
11
1
J =
0
=
,
y
w
+ c(x, y)w = f (x, y).,
where = a/x + b/y. To complete the transformation to the form of equation (2.3),
we first write (x, y), c(x, y) and f (x, y) in terms of and to obtain
A(, )
w
+ C(, )w = (, ).
+ w= ,
A
A
which is in the form of (2.3) with
h(, ) =
C(, )
A(, )
and F (, ) =
(, )
.
A(, )
The characteristic method applies to first order semilinear equation (2.2) as well as linear
equation (2.1); similar change of variables and basic algebra transform equation (2.2) to
K
w
= ,
A
where the nonlinear term K(, , w) = (x, y, u) and restricting again the variables to a set
in which A(, ) = (x, y) 6= 0.
Notation: It is very convenient to use the function u in places where rigorously the function
w should be used. E.g., the equation here above can identically be written as u/ = K/A.
Example: Consider the linear first order equation
x2
u
u
+y
+ xyu = 1.
x
y
This is equation (2.1) with a(x, y) = x 2 , b(x, y) = y, c(x, y) = xy and f (x, y) = 1. The
characteristic equation is
b
y
dy
= = 2.
dx
a
x
12
for y > 0,
and x 6= 0.
1
= 6= 0
y
y
as required.
Since = x,
= ln y +
1
y = e1/ .
Now we apply the transformation = x, = ln y + 1/x with w(, ) = u(x, y) and we have
w
w
w w
1
1 w
w
u
=
+
=
+
2 =
2
,
x
x
x
w
w
w 1
1 w
u
=
+
=0+
= 1/
.
y
y
y
y
e
Then the PDE becomes
1 w
1 w
2 w
+ e1/ 1/
2
+ e1/ w = 1,
e
which simplifies to
2
w
+ e1/ w = 1
then to
w 1 1/
1
+ e
w = 2.
We have transformed the equation into the form of equation (2.3), for any region of (, )space with 6= 0.
2.1.2
The point of this transformation is that we can solve equation (2.3). Think of
w
+ h(, )w = F (, )
as a linear first order ordinary differential equation in , with carried along as a parameter.
Thus we use an integrating factor method
e
h(,) d
R
w
+ h(, ) e
h(,) d
w = F (, ) e
R
R h(,) d
e
w = F (, ) e
h(,) d
h(,) d
13
Now we integrate with respect to . Since is being carried as a parameter, the constant of
integration may depend on
Z
R
R
h(,) d
w = F (, ) e h(,) d d + g()
e
in which g is an arbitrary differentiable function of one variable. Now the general solution of
the transformed equation is
Z
R
R
R
h(,) d
w(, ) = e
F (, ) e h(,) d d + g() e h(,) d .
We obtain the general form of the original equation by substituting back (x, y) and (x, y)
to get
u(x, y) = e(x,y) [(x, y) + g((x, y))] .
(2.6)
A certain class of first order PDEs (linear and semilinear PDEs) can then be reduced to a
set of ODEs. This makes use of the general philosophy that ODEs are easier to solve than
PDEs.
Example: Consider the constant coefficient equation
a
u
u
+b
+ cu = 0
x
y
constant.
So the characteristics of the PDE are the straight line graphs of bx ay = k and we make
the transformation with
= x, = bx ay.
a
The integrating factor method gives
c/a
e
w =0
ec/a w = g(),
where g is any differentiable function of one variable. Then
w = g() ec/a
and in terms of x and y we back transform
u(x, y) = g(bx ay) ecx/a .
14
Exercise:
Note: Consider the difference between general solution for linear ODEs and general solution
for linear PDEs. For ODEs, the general solution of
dy
+ q(x)y = p(x)
dx
contains an arbitrary constant of integration. For different constants you get different curves
of solution in (xy)plane. To pick out a unique solution you use some initial condition (say
y(x0 ) = y0 ) to specify the constant.
For PDEs, if u is the general solution to equation (2.1), then z = u(x, y) defines a family of
integral surfaces in 3Dspace, each surface corresponding to a choice of arbitrary function g
in (2.6). We need some kind of information to pick out a unique solution; i.e., to chose the
arbitrary function g.
2.1.3
Characteristic Curves
dy
= b(x, y).
ds
(2.7)
15
=cst
=cst
uo(1)
uo(2)
(xo, yo )
Notes:
1. u can only be found in the region between the characteristics drawn through the endpoint of .
2. Characteristics are curves on which the values of u combined with the equation are not
sufficient to determine the normal derivative of u.
3. A discontinuity in the initial data propagates onto the solution along the characteristics.
These are curves across which the derivatives of u can jump while u itself remains
continuous.
Existence & Uniqueness: Why do some choices of in (x, y)space give a solution and
other give no solution or an infinite number of solutions? It is due to the fact that the Cauchy
data (initial conditions) may be prescribed on a curve which is a characteristic of the PDE.
To understand the definition of characteristics in the context of existence and uniqueness of
solution, return to the general solution (2.6) of the linear PDE:
u(x, y) = e(x,y) [(x, y) + g((x, y))] .
Consider the Cauchy data, u0 , prescribed along the curve whose parametric form is (x =
x0 (), y = y0 ()) and suppose u0 (x0 (), y0 ()) = q(). If is not a characteristic, the
problem is wellposed and there is a unique function g which satisfies the condition
q() = e(x0 (),y0 ()) [(x0 (), y0 ()) + g(x0 (), y0 ())] .
If on the other hand (x = x0 (), y = y0 ()) is the parametrisation of a characteristic
((x, y) = k, say), the relation between the initial conditions q and g becomes
q() = e(x0 (),y0 ()) [(x0 (), y0 ()) + G] ,
(2.9)
where G = g(k) is a constant; the problem is illposed. The functions (x, y) and (x, y)
are determined by the PDE, so equation (2.9) places a constraint on the given data function
q(x). If q() is not of this form for any constant G, then there is no solution taking on these
prescribed values on . On the other hand, if q() is of this form for some G, then there
are infinitely many such solutions, because we can choose for g any differentiable function so
that g(k) = G.
16
Example 1:
u
u
+3
+ 8u = 0.
x
y
dy
3
=
dx
2
and the characteristics are the straight line graphs 3x 2y = c. Hence we take = 3x 2y
and = x.
y
=cst
=cst
=cst
(We can see that an and cross only once they are independent, i.e. J 6= 0; and have
been properly chosen.)
This gives the solution
u(x, y) = e4x g(3x 2y)
where g is a differentiable function defined over the real line. Simply specifying the solution
at a given point (as in ODEs) does not uniquely determine g; we need to take a curve of
initial conditions.
Suppose we specify values of u(x, y) along a curve in the plane. For example, lets choose
as the xaxis and gives values of u(x, y) at points on , say
u(x, 0) = sin(x).
Then we need
u(x, 0) = e4x g(3x) = sin(x)
i.e.
g(3x) = sin(x)e4x ,
so
g(x) = x4 e4x
17
However, not every curve in the plane can be used to determine g. Suppose we choose to
be the line 3x 2y = 1 and prescribe values of u along this line, say
u(x, y) = u(x, (3x 1)/2) = x2 .
Now we must choose g so that
e4x g(3x (3x 1)) = x2 .
This requires g(1) = x2 e4x (for all x). This is impossible and hence there is no solution
taking the value x2 at points (x, y) on this line.
Last, we consider again to be the line 3x 2y = 1 but choose values of u along this line to
be
u(x, y) = u(x, (3x 1)/2) = e4x .
Now we must choose g so that
u
u
y
=u
x
y
with
u = x2
on
y = x, 1 y 2
Characteristics:
y
dy
= d(xy) = 0 xy = c,
dx
x
So, take = xy and = x. Then the equation becomes
xy
w
w
w
+x
xy
=w
constant.
w
w =0
w
= 0.
Finally the general solution is, w = g() or equivalently u(x, y) = x g(xy). When y = x
u(x, y) = x xy.
=const
=const
This figure presents the characteristic curves given by xy = constant. The red characteristics
show the domain where the initial conditions permit us to determine the solution.
18
and y = c2 es ,
where the parametric form of the data curve permits us to find the two constants of
integration c1 & c2 in terms of the curvilinear coordinate along .
The curve is described by
x0 () =
and
y0 () =
with
[2, 1]
and we consider the points on to be the origin of the coordinate s along the characteristics
(i.e. s = 0 on ). So,
x0 = = c 1
x(s, ) = es
, [0, 1].
on (s = 0)
y0 = = c 2
y(s, ) = es
For linear or semilinear problems we can solve the compatibility equation independently of
the characteristic equations. (This property is not true for quasilinear equations.) Along the
characteristics u is determined by
du
= u u = c3 e s .
ds
Now we can make use of the Cauchy data to determine the constant of integration c 3 ,
on , at s = 0,
Then, we have the parametric forms of the characteristic curves and the solution
x(s, ) = es ,
y(s, ) = es
and
u(s, ) = 2 es ,
in terms of two parameters, s the curvilinear coordinate along the characteristic curves and
the curvilinear coordinate along the data curve . From the two first ones we get s and
in terms of x and y.
r
x
x
2s
= e s = ln
and xy = 2 = xy ( 0).
y
y
= xy
= x xy.
u(x, y) = xy exp ln
y
y
19
2.2
Quasilinear Equations
u
u
+ b(x, y, u)
= c(x, y, u)
x
y
(2.10)
where the functions a, b and c can involve u but not its derivatives.
2.2.1
We can represent the solutions u(x, y) by the integral surfaces of the PDE, z = u(x, y), in
(x, y, z)space. Define the Monge direction by the vector (a, b, c) and recall that the normal to
the integral surface is (x u, y u, 1). Thus quasilinear equation (2.10) says that the normal
to the integral surface is perpendicular to the Monge direction; i.e. integral surfaces are
surfaces that at each point are tangent to the Monge direction,
x u
a
b y u = a(x, y, u) u + b(x, y, u) u c(x, y, u) = 0.
x
y
1
c
With the field of Monge direction, with direction numbers (a, b, c), we can associate the family
of Monge curves which at each point are tangent to that direction field. These are defined
by
dx
a
c dy b dz
dy
dz
dx
dy b = a dz c dx = 0
=
=
(= ds),
a(x, y, u)
b(x, y, u)
c(x, y, u)
dz
c
b dx a dy
where dl = (dx, dy, dz) is an arbitrary infinitesimal vector parallel to the Monge direction.
In the linear case, characteristics were curves in the (x, y)plane (see 2.1.3). For the quasilinear equation, we consider Monge curves in (x, y, u)space defined by
dx
= a(x, y, u),
ds
dy
= b(x, y, u),
ds
du
= c(x, y, u).
ds
Characteristic equations (d{x, y}/ds) and compatibility equation (du/ds) are simultaneous
first order ODEs in terms of a dummy variable s (curvilinear coordinate along the characteristics); we cannot solve the characteristic equations and compatibility equation independently
as it is for a semilinear equation. Note that, in cases where c 0, the solution remains
constant on the characteristics.
The rough idea in solving the PDE is thus to build up the integral surface from the Monge
curves, obtained by solution of the ODEs.
Note that we make the difference between Monge curve or direction in (x, y, z)space and
characteristic curve or direction, their projections in (x, y)space.
20
2.2.2
General solution:
Suppose that the characteristic and compatibility equations that we have defined have two
independent first integrals (function, f (x, y, u), constant along the Monge curves)
(x, y, u) = c1
and
(x, y, u) = c2 .
Then the solution of equation (2.10) satisfies F (, ) = 0 for some arbitrary function F
(equivalently, = G() for some arbitrary G), where the form of F (or G) depends on the
initial conditions.
Proof:
dx dy du
d
=
+
+
= 0,
ds
x ds
y ds u ds
+b
+c
= 0.
x
y
u
+b
+c
= 0.
x
y
u
or
x u
x u
a J[u, x] = b J[y, u]
+b
=0
y u
y u
x
x
where J[x1 , x2 ] = 1 2 .
x1
x2
Thus, we have J[u, x] = J[x, y] b/c and J[y, u] = J[u, x] a/b = J[x, y] a/c.
Now consider F (, ) = 0 remember F ((x, y, u(x, y)), (x, y, u(x, y))) and differentiate
dF =
F
F
dx +
dy = 0
x
y
21
= 0,
x u x
y
u y
x
u x
y
u y
u
u
,
y u
y u x
u x
u x y
x y
x y
u
u
+ J[u, x]
= J[x, y].
J[y, u]
x
y
Then from the previous expressions for a, b, and c
a
u
u
+b
= c,
x
y
u
u
+y
=xy
x
y
in
with
u = 1+x
on
y = 1.
We first look for the general solution of the PDE before applying the initial conditions.
Combining the characteristic and compatibility equations,
dx
= y + u,
ds
dy
= y,
ds
du
=xy
ds
(2.11)
(2.12)
(2.13)
we seek two independent first integrals. Equations (2.11) and (2.13) give
d
(x + u) = x + u,
ds
and equation (2.12)
1 dy
= 1.
y ds
Now, consider
d
ds
x+u
y
1 d
x + u dy
(x + u)
,
y ds
y 2 ds
x+u x+u
= 0.
=
y
y
=
So, (x + u)/y = c1 is constant. This defines a family of solutions of the PDE; so, we can
choose
x+u
(x, y, u) =
,
y
22
such that = c1 determines one particular family of solutions. Also, equations (2.11)
and (2.12) give
d
(x y) = u,
ds
and equation (2.13)
d
du
(x y) (x y) = u .
ds
ds
Now, consider
d
d
d
(x y)2 u2 =
(x y)2
u2 ,
ds
ds
ds
d
du
= 2(x y) (x y) 2u
= 0.
ds
ds
Then, (x y)2 u2 = c2 is constant and defines another family of solutions of the PDE. So,
we can take
(x, y, u) = (x y)2 u2 .
The general solution is
x+u
F
, (x y)2 u2 = 0
y
or
(x y)2 u2 = G
x+u
y
1 2
1
1
1
1
+ xy+
2 = xy+
.
y2
y
y
y
y
2
u(x, y) = x y + .
y
23
x(y u)
u
u
+ y(x + u)
= (x + y)u
x
y
with
u = x2 + 1
on
y = x.
Characteristic equations
dx
= x(y u),
ds
dy
= y(x + u),
ds
du
= (x + y)u.
ds
(2.14)
(2.15)
(2.16)
Again, we seek to independent first integrals. On the one hand, equations (2.14) and (2.15)
give
y
Now, consider
dy
dx
+x
= xy 2 xyu + yx2 + xyu = xy (x + y),
ds
ds
1 du
= xy
from equation (2.16).
u ds
1 du
1 dx 1 dy
+
=
x ds y ds
u ds
xy
d
= 0.
ln
ds
u
xy
u
= xy xu xy yu = u(x + y) = ,
ds
ds
ds
d
(x + u y) = 0.
ds
Hence, x + u y = c2 is also a constant on the Monge curves and another first integral is
given by
(x, y, u) = x + u y,
xy
= G(x + u y).
u
Now, we make use of the initial conditions, u = x 2 + 1 on y = x, to determine G:
x2
= G(x2 + 1);
1 + x2
set = x2 + 1, i.e. x2 = 1, then
G() =
1
,
Rearrange to finish!
24
Alternative approach:
x2
u
u
+u
= 1,
x
y
with
u=0
on x + y = 1.
dy
=u
ds
and
du
= 1,
ds
(2.17)
x=
(2.18)
c 1 , c2 , c3 .
(2.19)
gives
(2.18)
gives
(2.19)
gives
1
1
c1 = ,
c1
1 = c3 c3 = 1 ,
0 = c2 c2 = 0.
,
1 s
x=
then
y=
s2
+1
2
and u = s.
=
,
1 s
1 + sx
u2
x
+1
.
2
1 + ux
Invariants, or first integrals, are (from solution (2.17), (2.18) and (2.19), keeping arbitrary
c2 = 0) = u2 /2 y and = x/(1 + ux).
y=
u
u
+y
=xy
x
y
in
with
u=1+x
on
y = 1.
Characteristic equations
dx
= y + u,
ds
dy
= y,
ds
du
= x y.
ds
(2.20)
(2.21)
(2.22)
25
c1 = 1,
c2 = 1 + 2
and
c3 = 1.
y = es
and u = es + es .
1
xy+
y
so
1 1
u= +
y y
Finally,
1
xy+
y
2
u =xy+ ,
y
y.
as before.
To find invariants, return to solved characteristics equations and solve for constants in terms
of x, y and u. We only need two, so put for instance c 1 = 1 and so y = es . Then,
x=
c3 1
+ (1 + c2 ) y
y
2
Solve for c2
c2 =
x+u
,
y
and u =
so
c3 1
+ (c2 1) y.
y
2
x+u
,
y
1
c3 = (x u y)y, so = (x u y)y.
2
Observe is different from last time, but this does not as we only require two independent
choices for and . In fact we can show that our previous is also constant,
(x y)2 u2 = (x y + u)(x y u),
= (y y) ,
y
= ( 1) which is also constant.
26
Summary:
2.3
Wave Equation
2.3.1
Linear Waves
If u is small (i.e. u2 u), then the equation approximate to the linear wave equation
u
u
+c
=0
t
x
The solution of the equation of characteristics, dx/dt = c, gives the first integral of the
PDE, (x, t) = x ct, and then general solution u(x, t) = g(x ct), where the function g is
determined by the initial conditions. Applying u(x, 0) = f (x) we find that the linear wave
equation has the solution u(x, t) = f (x ct), which represents a wave (unchanging shape)
propagating with constant wave speed c.
t
t
cs
t=
=
x
c
=
0
u(x,t)
f(x)
t=0
t=t1
x1=ct1
27
2.3.2
Nonlinear Waves
u
u
+ (u + c)
= 0,
t
x
dx
=c+u
ds
and
du
= 0,
ds
u = f [x (u + c)t],
according to initial conditions u(x, 0) = f (x). This is similar to the previous result, but now
the wave speed involves u.
However, this form of the solution is not very helpful; it is more instructive to consider
the characteristic curves. (The PDE is homogeneous, so the solution u is constant along
the Monge curves this is not the case in general which can then be reduced to their
projections in the (x, t)plane.) By definition, = x(c+u)t is constant on the characteristics
(as well as u); differentiate to find that the characteristics are described by
dx
= u + c.
dt
These are straight lines,
x = (f () + c)t + ,
expressed in terms of a parameter . (If we make use of the parametric form of the data curve
: {x = , t = 0, R} and solve directly the Cauchy problem in terms of the coordinate
s = t, we similarly find, u = f () and x = (u + c)t + .) The slope of the characteristics,
1/(c + u), varies from one line to another, and so, two curves can intersect.
()+
c)
t=
{x= ,t=0}
u=f(
) &
x(f
tmin
f()<0
x = (f (1 ) + c)t + 1 ,
x = (f (2 ) + c)t + 2 .
28
and if this is positive it will be in the region of solution. At this point u will not be singlevalued and the solution breaks down. By letting 2 1 we can see that the characteristics
intersect at
1
t= 0 ,
f ()
and the minimum time for which the solution becomes multivalued is
tmin =
1
,
max[f 0 ()]
i.e. the solution is single valued (i.e. is physical) only for 0 t < t min . Hence, when f 0 () < 0
we can expect the solution to exist only for a finite time. In physical terms, the equation
considered is purely advective; in real waves, such as shock waves in gases, when very large
gradients are formed then diffusive terms (e.g. xx u) become vitally important.
u(x,t)
f(x)
multivalued
breaking
f () = 1 2.
(0 1),
So, f 0 () < 0 for 1/2 < < 1 and we can expect the solution not to remain singlevalued for
all values of t. (max[f 0 ()] = 1 so tmin = 1. Now,
so
u = f (x (u + c)t),
u = [x (u + c)t] [1 x + (u + c)t],
(ct x 1 + ct),
p
1
2 4t(x ct) .
2t(x
ct)
(1
+
t)
+
(1
+
t)
2t2
u = x (c + 1) +
1 + c x,
29
2.3.3
Weak Solution
When wave breaking occurs (multivalued solutions) we must rethink the assumptions in
our model. Consider again the nonlinear wave equation,
u
u
+ (u + c)
= 0,
t
x
and put w(x, t) = u(x, t) + c; hence the PDE becomes the inviscid Burgers equation
w
w
+w
= 0,
t
x
or equivalently in a conservative form
w
+
t
x
w2
2
= 0,
.
dt x1
2
2
Let us now relax the assumption regarding the differentiability of the our solution; suppose
that w has a discontinuity in x = s(t) with x 1 < s(t) < x2 .
w(x,t)
w(s ,t)
w(s+,t)
x1
s(t)
x2
=
w(x, t) dx,
2
2
dt x1
dt s(t)
Z x2
Z s(t)
w
w
dx w(s+ , t) s(t)
+
dx,
= w(s , t) s(t)
+
t
s(t) t
x1
where w(s (t), t) and w(s+ (t), t) are the values of w as x s from below and above respectively; s = ds/dt.
Now, take the limit x1 s (t) and x2 s+ (t). Since w/t is bounded, the two integrals
tend to zero. We then have
w2 (s , t) w2 (s+ , t)
= s w(s , t) w(s+ , t) .
2
2
30
1 w2 (s+ ) w2 (s )
w(s+ ) + w(s )
=
.
2 w(s+ ) w(s )
2
The problem then reduces to fitting shock discontinuities into the solution in such a way
that the jump condition is satisfied and multivalued solution are avoided. A solution that
satisfies the original equation in regions and which satisfies the integral form of the equation
is called a weak solution or generalised solution.
Example: Consider the inviscid Burgers equation
w
w
+w
= 0,
t
x
with initial conditions
1
1
w(x = , t = 0) = f () =
for 0,
for 0 1,
for 1.
st
=c
w=
1&
t
x
t=
1&
w=
x=1
multivalued
solution
t>1
x
w(x,t) 1
t=0
weak solution
(shock wave)
w(x,t) 1
t=0
U
x
t=1
x
31
1x
w(x, t) =
1t
0
for
for
for
x t,
t x 1,
x 1,
with 0 t < 1,
becomes steeper and steeper until it becomes vertical at t = 1; then the solution is multivalued. Nevertheless, we can define a generalised solution, valid for all positive time, by
introducting a shock wave.
Suppose shock at s(t) = U t + , with w(s , t) = 1 and w(s+ , t) = 0. The jump condition
gives the shock velocity,
w(s+ ) + w(s )
1
U=
= ;
2
2
furthermore, the shock starts at x = 1, t = 1, so = 1 1/2 = 1/2. Hence, the weak solution
of the problem is, for t 1,
1
0 for x < s(t),
w(x, t) =
where s(t) = (t + 1).
1 for x > s(t),
2
2.4
2.4.1
Systems of Equations
Linear and Semilinear Equations
aij ux(j)
bij uy(j)
j=1
= ci ,
i = 1, 2, . . . , n
u
= ux ,
x
for the unknowns u(1) , u(j) , . . . , u(n) and when the coefficients aij and bij are functions only
of x and y. (Though the ci could also involve u(k) .)
In matrix notation
Aux + Buy = c,
where
. . . a1n
b11 . . . b1n
.. ,
.. ,
..
B = (bij ) = ... . . .
.
.
.
an1 . . . ann
bn1 . . . bnn
(1)
u
c1
u(2)
c2
c = . and u = . .
.
.
.
.
cn
u(n)
a11
..
A = (aij ) = .
E.g.,
(1)
(2)
(2)
u(1)
x 2ux + 3uy uy = x + y,
(2)
(1)
(2)
2
2
u(1)
x + ux 5uy + 2uy = x + y ,
32
can be expressed as
" (1) #
" (1) #
1 2 ux
3 1 uy
x+y
(2) + 5
(2) = x2 + y 2 ,
1 1
2
ux
uy
or Aux + Buy = c where
3 1
1 2
, B=
A=
5 2
1 1
1/3 2/3
If we multiply by A1 =
1/3 1/3
x+y
.
and c = 2
x + y2
A1 A ux + A1 B uy = A1 c,
we obtain ux + Duy = d,
7/3 1
3 1
1/3 2/3
1
and d = A1 c.
=
where D = A B =
8/3 1
1/3 1/3 5 2
We now assume that the matrix A is nonsingular (i.e., the inverse A 1 exists ) at least
there is some region of the (x, y)plane where it is nonsingular. Hence, we need only to
consider systems of the form
ux + Duy = d.
We also limit our attention to totally hyperbolic systems, i.e. systems where the matrix D
has n distinct real eigenvalues (or at least there is some region of the plane where this holds).
D has the n distinct eigenvalues 1 , 2 , . . . , n where det(i I D) = 0 (i = 1, . . . , n), with
i 6= j (i 6= j) and the n corresponding eigenvectors e 1 , e2 , . . . , en so that
Dei = i ei .
The matrix P = [e1 , e2 , . . . , en ] diagonalises
1 0
0 2
= 0 ...
0 ...
0 ...
We now put u = P v,
then
and
D via P 1 DP = ,
... ...
0
0
...
0
..
.
. ...
0
0 n1 0
...
0
n
P vx + Px v + DP vy + DPy v = d,
P 1 P vx + P 1 Px v + P 1 DP vy + P 1 DPy v = P 1 d,
vx(i) + i vy(i) = qi
(i = 1, . . . , n),
where qi can involve {v (1) , v (2) , . . . , v (n) } and with n characteristics given by
dy
= i .
dx
This is the canonical form of the equations.
33
(2)
(2)
ux
(1)
9 uy
ux + 4 uy = 0,
+
= 0,
0 4
.
Here, ux + Duy = 0 with D =
9 0
Eigenvalues:
det(D I) = 0 2 36 = 0 = 6.
Eigenvectors:
2
x
0
x
6 4
for = 6,
=
=
3
y
0
9 6 y
6 4 x
0
x
2
=
=
for = 6.
9 6 y
0
y
3
Then,
1 3 2
6 0
2 2
1
1
.
and P DP =
, P =
P =
0 6
3 3
12 3 2
2 2
6 0
So we put u =
v and vx +
v = 0, which has general solution
3 3
0 6 y
v (1) = f (6x y)
and
i.e.
u(1) = 2v (1) + 2v (2)
u(1) =
u(2)
Example 2:
to canonical form in the region of the (x, y)space where it is totally hyperbolic.
Eigenvalues:
4y x
2x 2y
det
= 0 {3x, 3y}.
2y 2x
4x y
The system is totally hyperbolic everywhere expect where x = y.
34
Eigenvalues:
1 = 3x e1 = [1, 2]T ,
2 = 3y e2 = [2, 1]T .
So,
1 1 2
1 2
1
, P =
P =
2 1
3 2 1
1 2
Then, with u =
v we obtain
2 1
vx +
2.4.2
and
3x 0
.
DP =
0 3y
3x 0
v = 0.
0 3y y
Quasilinear Equations
We consider systems of n equations, involving n functions u (i) (x, y) (i = 1, . . . , n), of the form
ux + Duy = d,
where D as well as d may now depend on u. (We have already shown how to reduce a more
general system Aux + Buy = c to that simpler form.) Again, we limit our attention to totally
hyperbolic systems; then
= P 1 DP D = P P 1 ,
using the same definition of P , P 1 and the diagonal matrix , as for the linear and semilinear
cases. So, we can transform the system in its normal form as,
P 1 ux + P 1 uy = P 1 d,
such that it can be written in component form as
n
X
j=1
Pij1
(j)
(j)
u + i
u
x
y
n
X
Pij1 dj ,
(i = 1, . . . , n)
j=1
where i is the ith eigenvalue of the matrix D and wherein the ith equation involves differentiation only in a single direction the direction dy/dx = i . We define the ith characteristic,
with curvilinear coordinate si , as the curve in the (x, y)plane along which
dx
= 1,
dsi
dy
= i
dsi
or equivalently
dy
= i .
dx
Pij1
d (j) X 1
u =
Pij dj
dsi
j=1
(i = 1, . . . , n).
35
+
+u
= 0.
t
x
x
If the entropy is the same everywhere in the motion then P = constant, and the motion
equation becomes
u
u c2
+u
+
= 0,
t
x
x
where c2 = dP/d = P/ is the sound speed. We have then a system of two first order
quasilinear PDEs; we can write these as
w
w
+D
= 0,
t
x
with
u
w=
and
u c2 /
D=
.
The two characteristics of this hyperbolic system are given by dx/dt = where are the
eigenvalues of D;
u c2 /
= 0 (u )2 = c2 and = u c.
det(D I) =
u
The eigenvectors are [c, ]T for and [c, ]T for + , such that the usual matrices are
1 c
uc
0
c c
1
1
, such that = T DT =
.
T =
, T =
0
u+c
2c c
Put and the curvilinear coordinates along the characteristics dx/dt = u c and dx/dt =
u + c respectively; then the system transforms to the canonical form
dt
dt
=
= 1,
d
d
dx
= u c,
d
dx
= u + c,
d
du
d
c
=0
d
d
and
du
d
+c
= 0.
d
d
36
Chapter 3
3.1
37
44
2u
u
2u
u
2u
+ 2b(x, y)
+ c(x, y) 2 + d(x, y)
+ e(x, y)
+ f (x, y)u = g(x, y);
2
x
xy
y
x
y
Then by inverse theorem there is an open neighbourhood of (x, y) and another neighbourhood
of (, ) such that the transformation is invertible and onetoone on these neighbourhoods.
As before we compute chain rule derivations
u
u
u
=
+
,
x
x x
u
u
u
=
+
,
y
y y
37
38
2u
2u
2u
+
2B
+
C
+ F (u , u , u, , ) = 0,
2
(3.1)
where
2
2
A=a
+ 2b
,
+c
x
x y
y
+c
+b
+
,
B=a
x x
x y y x
y y
2
2
+ 2b
.
C=a
+c
x
x y
y
We write explicitly only the principal part of the PDE, involving the highestorder derivatives
of u (terms of second order).
It is easy to verify that
2
(B AC) = (b ac)
x y y x
2
where (x y y x )2 is just the Jacobian squared. So, provided J 6= 0 we see that the
sign of the discriminant b2 ac is invariant under coordinate transformations. We can use
this invariance properties to classify the equation.
Equation (3.1) can be simplified if we can choose and so that some of the coefficients A,
B or C are zero. Let us define,
D =
/x
/y
and D =
/x
;
/y
aD2
+ 2bD + c
2
B = (aD D + b (D + D ) + c)
,
y y
2
2
C = aD + 2bD + c
.
y
aD 2 + 2bD + c = 0,
(3.2)
39
b2 ac
.
a
If the discriminant b2 ac 6= 0, equation (3.2) has two distinct roots; so, we can make both
coefficients A and C zero if we arbitrarily take the root with the negative sign for D and
the one with the positive sign for D ,
b b2 ac
/x
=
A = 0,
(3.3)
D =
/y
a
/x
b + b2 ac
D =
=
C = 0.
/y
a
Then, using D D = c/a and D + D = 2b/a we have
B=
2
ac b2
B 6= 0.
a
y y
Furthermore, if the discriminant b 2 ac > 0 then D and D as well as and are real. So,
we can define two families of oneparameter characteristics of the PDE as the curves described
by the equation (x, y) = constant and the equation (x, y) = constant. Differentiate along
the characteristic curves given by = constant,
d =
dx +
dy = 0,
x
y
b + b2 ac
dy
=
.
dx
a
(3.4)
Similarly we find that the characteristic curves described by (x, y) = constant satisfy
dy
b b2 ac
=
.
(3.5)
dx
a
If the discriminant b2 ac = 0, equation (3.2) has one unique root and if we take this root
for D say, we can make the coefficient A zero,
D =
b
/x
= A = 0.
/y
a
(3.6)
40
In this case the equation is said to be hyperbolic and has two families of characteristics
given by equation (3.4) and equation (3.5).
2. If b2 = ac, a suitable choice for still simplifies the PDE, but now we can choose
arbitrarily provided and are independent and the equation reduces to the
form
2u
+ (lower order terms) = 0.
2
The equation is said to be parabolic and has only one family of characteristics given by
equation (3.6).
3. If b2 < ac we can again apply the change of variables (x, y) (, ) to simplify the
equation but now this functions will be complex conjugate. To keep the transformation
real, we apply a further change of variables (, ) (, ) via
= + = 2 <(),
= i( ) = 2 =(),
2u
2u
2u
=
+
2 2
so, the equation can be reduced to
i.e.,
2u
2u
+
+ (lower order terms) = 0.
2 2
In this case the equation is said to be elliptic and has no real characteristics.
The above forms are called the canonical (or standard) forms of the second order linear or
semilinear equations (in two variables).
Summary:
b2 ac
Canonical form
Type
>0
2u
+ ... = 0
Hyperbolic
=0
2u
2
+ ... = 0
Parabolic
<0
2u
2
2u
2
+ ... = 0
Elliptic
E.g.
The wave equation,
2
2u
2 u
c
= 0,
w
t2
x2
is hyperbolic (b2 ac = c2w > 0) and the two families of characteristics are described
by dx/dt = cw i.e. = x cw t and = x + cw t. So, the equation transforms into its
canonical form 2 u/ = 0 whose solutions are waves travelling in opposite direction
at speed cw .
41
= 0,
x2 t
is parabolic (b2 ac = 0). The characteristics are given by dt/dx = 0 i.e. = t =
constant.
Laplaces equation,
2u 2u
+ 2 = 0,
x2
y
2u 2u
+ 2 = 0,
x2
y
(b2 ac = 0 y = y)
is elliptic in y > 0, parabolic for y = 0 and hyperbolic in y < 0, and for small disturbances
in incompressible (inviscid) flow
1 2u 2u
+ 2 = 0,
1 m2 x2
y
(b2 ac =
1
)
1 m2
2xy
=
y
+
x
2
2
x
xy
y
xy
2
2u
3 u
+x
3 u
a = y2
Here b = xy
so b2 ac = (xy)2 x2 y 2 = 0 parabolic equation.
2
c=x
On = constant,
dy
b + b2 ac
b
x
=
= = = x2 + y 2 .
dx
a
a
y
We can choose arbitrarily provided and are independent. We choose = y. (Exercise,
try it with = x.) Then
u
u
= 2x ,
x
u
u u
= 2y
+
,
y
2u
2u
2u
= 4xy 2 + 2x
,
xy
2
u
2u
2 u
=
2
,
+
4x
x2
2
u
2u
2u
2u
2 u
=
2
+
4y
,
+
4y
+
y 2
2u
u
2u
2u
u
2u
+ 4x2 y 2 2 8x2 y 2 2 4x2 y
+ 2x2
+ 4x2 y 2 2
2
2
1
u
u
u
u
u
2xy 3
,
+ x2 2 =
+ 2x3 y
+ x3
+ 4x2 y
xy
42
= 0.
2
i.e.
(canonical form)
2u
2u
u
2u
+
2
+ 1 = 0 in 0 x 1, y > 0, with u =
= x on y = 0.
2
2
x
xy
y
y
a=1
so b2 ac = 9/4 (> 0) equation is hyperbolic.
Here b = 1/2
c = 2
Characteristics:
dy
1 3
/x
/x
.
= = 1 or 2
=
or
dx
2 2
/y
/y
Two methods of solving:
1. directly:
1
dy
= 2 x y = constant
dx
2
and
dy
= 1 x + y = constant.
dx
2. simultaneous equations:
y )
=2
x=
=x
x
y
2
=x+y
y=
x
y
So,
u
1 u u
=
+
,
y
2
u u
u
=
+
,
x
1 2u 1 2u
2u
2u
=
+
+
,
xy
2 2
2
2
1
( + 2)
3
2
( )
3
2u
2u
u
2u
=
+
2
+
x2
2
2
2u
1 2u
2u
2u
=
+
,
y 2
4 2
+
2
+ 1 = 0,
+
+
2
2
2
2 2
2
2
2 2
9 2u
+ 1 = 0,
2
canonical form.
43
(3.7)
9
2
9
the condition u/y = x at y = 0 gives
u
1
1
2
1
10
( = x, = x) = x f 0 (x) x + g 0 (x) = x g 0 (x) f 0 (x) =
x,
y
9
2
9
2
9
1
5
g(x) f (x) = x2 + k,
(3.8)
2
9
where k is a constant. Solving equation (3.7) and equation (3.8) simultaneously gives,
and after integration,
2
2
2
f (x) = x x2 k
3
9
3
or, in terms of and
f () =
and
g(x) =
2
2
2
2 k
3
9
3
1
4
2
x + x2 + k,
3
9
3
1
4
2
and g() = + 2 + k.
3
9
3
y2
.
2
a=1
Here b = 1/2
so b2 ac = 3/4 (< 0) equation is elliptic.
c=1
Find and via
)
)
= y 12 (1 + i 3)x
dy/dx = (1 + i 3)/2
44
and = i( ) = x 3.
u u
u
=
+ 3
,
x
u
u
2u
=2
,
=
y
x2
2u
2u
2u
3
= 2
+
2
,
xy
2
2u
2u
2u
3
2
+
3
,
2
2
2u
2u
=
4
,
y 2
2
2u
2u
2u
2u
2u
2u
2
+
3
2
+
2
+
4
= 0.
3
3
2
2
2
3.2
3.2.1
2u
2u
+
= 0,
2
2
canonical form.
There are two obvious ways in which we might wish to extend the theory.
To consider quasilinear second order equations (still in two independent variables.) Such
equations can be classified in an invariant way according to rule analogous to those developed
above for linear equations. However, since a, b and c are now functions of x u, y u and u its
type turns out to depend in general on the particular solution searched and not just on the
values of the independent variables.
To consider linear second order equations in more than two independent variables. In such
cases it is not usually possible to reduce the equation to a simple canonical form. However,
for the case of an equation with constant coefficients such a reduction is possible. Although
this seems a rather restrictive class of equations, we can regard the classification obtained as
a local one, at a particular point.
Consider the linear PDE
n
X
aij
i,j=1
X u
2u
+
+ cu = d.
bi
xi xj
xi
i=1
and = (i ij ), i, j = 1, , n.
n
X
j=1
eij j
and
j =
n
X
i=1
eij xi .
X u
u
=
eik
xi
k
45
n
X
2u
u
=
eik ejr .
xi xj
k r
and
k=1
k,r=1
i,j=1 k,r=1
u
aij eik ejr + (lower order terms) = 0.
k r
n
X
A er =
i,j=1
k=1
u
+ (lower order terms) = 0.
k2
2u 2u 2u
+ 2+ 2
x2
y
z
= 0.
When all the k are nonzero and there are at least two of each sign, the equation
ultrahyperbolic. E.g.
2u 2u
2u 2u
+
=
+
;
x21 x22
x23 x24
such equation do not often arise in mathematical physics.
If any of the k vanish the equation is parabolic. E.g. heat equation
u
3.2.2
2u 2u 2u
+ 2 + 2
x2
y
z
= 0.
2u
2u
2u
+
2b
+ F (x u, y u, u, x, y) = 0
+
c
x2
xy
y 2
46
which takes prescribed values on a given curve which we assume is represented parametrically in the form
x = (), y = (),
for I, where I is an interval, 0 1 , say. (Usually consider piecewise smooth
curves.)
We specify Cauchy data on : u, u/x and u/y are given I, but note that
we cannot specify all these quantities arbitrarily. To show this, suppose u is given on by
u = f (); then the derivative tangent to , du/d, can be calculated from du/d = f 0 ()
but also
u dx u dy
du
=
+
,
d
x d y d
u
u
= 0 ()
+ 0 ()
= f 0 (),
x
y
so, on , the partial derivatives u/x, u/y and u are connected by the above relation.
Only derivatives normal to and u can be prescribed independently.
So, the Cauchy problem consists in finding the solution u(x, y) which satisfies the following
conditions
u((), ()) = f ()
,
u
and
((), ()) = g()
n
and
dq
2 u dx 2 u dy
=
+ 2 .
ds
xy ds
y ds
Combining these two equations with the original PDE gives the following system of equations
for 2 u/x2 , 2 u/xy and 2 u/y 2 on (in matrix form),
2u
x2
2u
xy
2u
y 2
F
dp
ds
dq
ds
2b
dx
where M =
ds
dy
ds
dx
ds
0 .
dy
ds
So, if det(M) 6= 0 we can solve the equations uniquely and find 2 u/x2 , 2 u/xy and
2 u/y 2 on . By successive differentiations of these equations it can be shown that the
derivatives of u of all orders are uniquely determined at each point on for which det(M) 6= 0.
The values of u at neighbouring points can be obtained using Taylors theorem.
So, we conclude that the equation can be solved uniquely in the vicinity of provided
det(M) 6= 0 (CauchyKowaleski theorem provides a majorant series ensuring convergence of
Taylors expansion).
47
Consider what happens when det(M) = 0, so that M is singular and we cannot solve uniquely
for the second order derivatives on . In this case the determinant det(M) = 0 gives,
a
dy
ds
2
dx dy
+c
2b
ds ds
dx
ds
2
= 0.
But,
dy
dy/ds
=
dx
dx/ds
and so (dividing through by dx/ds), dy/dx satisfies the equation,
2
dy
dy
b b2 ac
dy
2b
+ c = 0,
i.e.
=
a
dx
dx
dx
a
or
dy
b
= .
dx
a
The exceptional curves , on which, if u and its normal derivative are prescribed, no unique
solution can be found satisfying these conditions, are the characteristics curves.
48
Chapter 4
Elliptic Equations
Contents
4.1
4.2
4.3
4.4
4.1
Definitions . . . . . . . . . . . . . . . . . . . . . . . . .
Properties of Laplaces and Poissons Equations . .
Solving Poisson Equation Using Greens Functions
Extensions of Theory: . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
49
50
54
68
Definitions
Elliptic equations are typically associated with steadystate behavior. The archetypal elliptic
equation is Laplaces equation
2 u = 0,
e.g.
2u 2u
+ 2 = 0 in 2D,
x2
y
and describes
steady, irrotational flows,
electrostatic potential in the absence of charge,
equilibrium temperature distribution in a medium.
Because of their physical origin, elliptic equations typically arise as boundary value problems
(BVPs). Solving a BVP for the general elliptic equation
L[u] =
n
X
i,j=1
aij
X u
2u
+
+ cu = F
bi
xi xj
xi
i=1
(recall: all the eigenvalues of the matrix A = (a ij ), i, j = 1 n, are nonzero and have the
same sign) is to find a solution u in some open region of space, with conditions imposed
on (the boundary of ) or at infinity. E.g. inviscid flow past a sphere is determined by
boundary conditions on the sphere (u n = 0) and at infinity (u = Const).
There are three types of boundary conditions for wellposed BVPs,
1. Dirichlet condition u takes prescribed values on the boundary (first BVP).
49
50
Z
Z
F dV =
f dS
for the problem to be welldefined.
y
n
Sometimes we may have a mixed problem, in which u is given on part of and u/n given
on the rest of .
If encloses a finite region, we have an interior problem; if, however, is unbounded, we
have an exterior problem, and we must impose conditions at infinity.
Note that initial conditions are irrelevant for these BVPs and the Cauchy problem for elliptic
equations is not always wellposed (even if CauchyKowaleski theorem states that the solution
exist and is unique).
As a general rule, it is hard to deal with elliptic equations since the solution is global, affected
by all parts of the domain. (Hyperbolic equations, posed as initial value or Cauchy problem,
are more localised.)
From now, we shall deal mainly with the Helmholtz equation 2 u + P u = F , where P and
F are functions of x, and particularly with the special one if P = 0, Poissons equation, or
Laplaces equation, if F = 0 too. This is not too severe restriction; recall that any linear
elliptic equation can be put into the canonical form
n
X
2u
k=1
x2k
+ = 0
and that the lower order derivatives do not alter the overall properties of the solution.
4.2
51
4.2.1
Definition: Let x0 be a point in and let BR (x0 ) denote the open ball having centre x 0
and radius R. Let R (x0 ) denote the boundary of BR (x0 ) and let A(R) be the surface area
of R (x0 ). Then a function u has the mean value property at a point x 0 if
Z
1
u(x) dS
u(x0 ) =
A(R) R
for every R > 0 such that BR (x0 ) is contained in . If instead u(x0 ) satisfies
Z
1
u(x0 ) =
u(x) dV,
V (R) BR
where V (R) is the volume of the open ball B R (x0 ), we say that u(x0 ) has the second
mean value property at a point x0 . The two mean value properties are equivalent.
BR
x0
x
Theorem:
on .
Proof:
(4.1)
52
v
dS =
r
u
r
v
dS
r
1
2
i.e
u dS =
1
r2
u dS.
1
r3
u(x0 ) =
u(x0 ) d =
3
4
2
1
u dS d =
4
u dV.
Br
The inverse of this theorem holds too, but is harder to prove. If u has the mean value property
then u is harmonic.
4.2.2
MaximumMinimum Principle
One of the most important features of elliptic equations is that it is possible to prove theorems
concerning the boundedness of the solutions.
Theorem:
in , with F > 0 in .
u
= 0,
y
2u
0
x2
and
2u
0,
y 2
53
Proof: (Theorem stated in 2D but holds in higher dimensions.) Consider the function
v = u + (x2 + y 2 ), for any > 0. Then 2 v = 4 > 0 in (since 2 (x2 + y 2 ) = 4), and
using the previous theorem,
v M + R2 in ,
where u M on and R is the radius of the circle containing . As this holds for any ,
let 0 to obtain
u M in ,
i.e., if u satisfies 2 u = 0 in , then u cannot exceed M , the maximum value of u on .
Also, if u is a solution of 2 u = 0, so is u. Thus, we can apply all of the above to u to
get a minimum principle: if u m on , then u m in .
This theorem does not say that harmonic function cannot also attain m and M inside
though. We shall now progress into the strong MaximumMinimum Principle.
Theorem: Suppose that u has the mean value property in a bounded region and that u
= . If u is not constant in then u attains its maximum value on
is continuous in
the boundary of , not in the interior of .
then it attains its maximum
Proof: Since u is continuous in the closed, bounded domain
M somewhere in .
Our aim is to show that, if u attains its max. at an interior point of , then u is constant in
.
Suppose u(x0 ) = M and let x? be some other point of . Join these points with a path
covered by a sequence of overlapping balls, B r .
x?
x
x0
x1
Consider the ball with x0 at its center. Since u has the mean value property then
Z
1
M = u(x0 ) =
u dS M.
A(r) r
This equality must hold throughout this statement and u = M throughout the sphere surrounding x0 . Since the balls overlap, there is x 1 , centre of the next ball such that u(x 1 ) = M ;
the mean value property implies that u = M in this sphere also. Continuing like this gives
u(x? ) = M . Since x? is arbitrary, we conclude that u = M throughout , and by continuity
Thus if u is not a constant in it can attain its maximum value only on the
throughout .
boundary .
54
Corollary: Applying the above theorem to u establishes that if u is non constant it can
attain its minimum only on .
Also as a simple corollary, we can state the following theorem. (The proof follows immediately
the previous theorem and the weak MaximumMinimum Principle.)
Theorem: The strong MaximumMinimum Principle for Laplaces equation.
i.e. solution of 2 u = 0 in and continuous in ,
with M and
Let u be harmonic in ,
m the maximum and minimum values respectively of u on the boundary . Then, either
u = f on
in , with v = 0 on .
with
u{1,2} = f{1,2} on ,
i=1
and, as for Poissons equation it is possible then to prove that the solution to the Dirichlet
BVP is unique and stable.
4.3
We shall develop a formal representation for solutions to boundary value problems for Poissons equation.
55
4.3.1
u(x) = L1 F (x),
where L1 , the inverse of L, is some integral operator. (We can expect to have LL 1 =
LL1 = I, identity.) We define the inverse L 1 using a Greens function: let
u(x) = L
F (x) =
G(x, )F ()d,
(4.2)
where G(x, ) is the Greens function associated with L (G is the kernel). Note that G
depends on both the independent variables x and the new independent variables , over
which we integrate.
Recall the Dirac function (more precisely distribution or generalised function) (x) which
has the properties,
Z
(x) dx = 1
and
Rn
Rn
(x ) h() d = h(x).
LG(x, )F () d;
4.3.2
G(x, ) F () d
(4.3)
56
y
n
V
x
u v v u dV =
Z
S
u
v
v
u
n
n
dS,
u v dV =
V
vF dV +
Z
S
u
v
v
u
n
n
dS;
(v + w)F dV
Z
S
u (v + w) (v + w)
n
n
dS,
GF dV
Z
S
G
u
G
n
n
dS.
(4.5)
Note that if, F , f and the solution u are sufficiently wellbehaved at infinity this integral
equation is also valid for unbounded regions (i.e. for exterior BVP for Poissons equation).
The way to remove u or u/n from the RHS of the above equation depends on the choice
of boundary conditions.
57
Here, the solution to equation (4.3) satisfies the condition u = f on S. So, we choose w
such that w = v on S, i.e. G = 0 on S, in order to eliminate u/n form the RHS of
equation (4.5).
Then, the solution of the Dirichlet BVP for Poissons equation
2 u = F in V with u = f on S
is
u() =
GF dV
2 v
f
S
G
dS,
n
with G = 0 on S.
Neumann Boundary Conditions
Here, the solution to equation (4.3) satisfies the condition u/n = f on S. So, we choose
w such that w/n = v/n on S, i.e. G/n = 0 on S, in order to eliminate u from the
RHS of equation (4.5).
However, the Neumann BVP
2 G = (x ) in V,
G
= 0 on S,
with
n
which does not satisfy a compatibility equation, has no solution. Recall that the Neumann
BVP 2 u = F in V , with u/n = f on S, is illposed if
Z
Z
f dS.
F dV 6=
S
We need to alter the Greens function a little to satisfy the compatibility equation; put
2 G = + C, where C is a constant, then the compatibility equation for the Neumann
BVP for G is
Z
Z
1
0 dS = 0 C = ,
( + C) dV =
V
S
V
u() =
GF dV +
V
Gf dS +
S
1
V

u dV .
{z }
This shows that, whereas the solution of Poissons equation with Dirichlet boundary conditions is unique, the solution of the Neumann problem is unique up to an additive constant u
58
GF dV +
Gf dS,
V
2 v
2 G = (x ) in V,
G
+ G = 0 on S.
n
G( 2 , 1 ) = G( 1 , 2 ).
Nevertheless, note that for Neumann BVPs, the term 1/V which provides the additive constant to the solution to Poissons equation breaks the symmetry of G.
59
1
ln (x )2 + (y )2 ,
4
Hence, v = 1/2 ln and dv/d = 1/2 and applying Greens theorem to u and v in
this new region V ? (with boundaries S and ), we get
Z
Z
v
v
u
u
f
dS +
u
dS = 0.
(4.6)
v
v
n
n
n
n
S
2 0
2 0
n
0
and so, in the limit 0, equation (4.6) gives
Z
v
u
dS, where
f
u(, ) =
v
n
n
S
v=
1
ln (x )2 + (y )2 .
4
60
dS
f
dS = f
u(, ) =
v
f
+v
(v + w) dS.
n
n
n
n
n
S
S
S
Setting G(x, y; , ) = v + w, then
u(, ) =
G
dS.
n
4.3.3
with G = 0 on S.
where
2 v = (x ) in V.
How do we find the free space Greens function v defined such that 2 v = (x ) in V ?
Note that it does not depend on the form of the boundary. (The function v is a source term
and for Laplaces equation is the potential due to a point source at the point x = .)
As an illustration of the method, we can derive that, in two dimensions,
v=
1
ln (x )2 + (y )2 ,
4
&
y = r sin ,
and look for a solution of Laplaces equation which is independent of and which is singular
as r 0.
Dr r
Cr
61
Cr
v
dS = 1.
n
Now, in polar coordinates, v/n = v/r = B/r and dS = rd (going around circle C r ).
So,
Z 2
Z 2
B
1
rd = B
d = 1 B = .
r
2
0
0
Hence,
1
1
1
ln r = ln r 2 = ln (x )2 + (y )2 .
2
4
4
(We do not use the boundary condition in finding v.)
v=
Similar (but more complicated) methods lead to the freespace Greens function v for the
Laplace equation in n dimensions. In particular,
n = 1,
x ,
1
n = 2,
ln x 2 ,
v(x, ) =
4
x 2n , n 3,
(2 n)An (1)
where x and are distinct points and A n (1) denotes the area of the unit nsphere. We shall
restrict ourselves to two dimensions for this course.
Note that Poissons equation, 2 u = F , is solved in unbounded Rn by
Z
v(x, ) F () d
u(x) =
Rn
where from equation (4.2) the free space Greens function v, defined above, serves as Greens
function for the differential operator 2 when no boundaries are present.
4.3.4
Method of Images
In order to solve BVPs for Poissons equation, such as 2 u = F in an open region V with
some conditions on the boundary S, we seek a Greens function G such that, in V
G(x, ) = v(x, ) + w(x, )
where 2 v = (x )
and 2 w = 0 or 1/V(V ).
62
Having found the free space Greens function v which does not depend on the boundary
conditions, and so is the same for all problems we still need to find the function w, solution
of Laplaces equation and regular in x = , which fixes the boundary conditions (v does not
satisfies the boundary conditions required for G by itself). So, we look for the function which
satisfies
2 w = 0 or 1/V(V ) in V,
with
2u 2u
+ 2 = 0 in y > 0
x2
y
with
u = f (x) on y = 0.
1
ln (x )2 + (y )2 .
4
1
ln (x )2 + (y + )2 ,
4
y
x=
(, )
y =
y=
+ (, )
y
v
(x )2 + 2
(x )2 + 2
= 0, as required.
63
G
dS.
n
G
G
1
=
=
n S
y y=0
(x )2 + 2
Thus,
u(, ) =
f (x)
dx,
(x )2 + 2
f ()
d.
( x)2 + y 2
Example 2
Find Greens function for the Dirichlet BVP
2 u =
2u 2u
+ 2 = F in the quadrant x > 0, y > 0.
x2
y
(, ) +
(, )
S
x
(, )
+ (, )
So,
1
1
ln (x )2 + (y )2 +
ln (x )2 + (y + )2
4
4
1
1
2
ln (x + ) + (y + )2 +
ln (x + )2 + (y )2 .
4
4
1
G(x, y, , ) =
ln
4
"
#
(x )2 + (y )2 (x + )2 + (y + )2
,
((x )2 + (y + )2 ) ((x + )2 + (y )2 )
and again we can check that G(0, y, , ) = G(x, 0, , ) = 0 as required for Dirichlet BVP.
64
Example 3
Consider the Neumann BVP for Laplaces equation in the upper halfplane,
2 u =
2u 2u
+ 2 = 0 in y > 0
x2
y
with
u
u
=
= f (x) on y = 0.
n
y
(, )
x=
S
y
y=
y =
(, )
G=v+w
1
1
ln (x )2 + (y )2
ln (x )2 + (y + )2 .
4
4
G
1
=
y
4
2(y )
2(y + )
+
2
2
(x ) + (y )
(x )2 + (y + )2
S2
S1
+R
x
65
Split S into S1 , the portion along the xaxis and S2 , the semicircular arc. Then, in the above
equation we have to consider the behaviour of the integrals
Z
Z
Z
Z
u
G
u
G
G
u
(1)
dS
R d and (2)
dS
R d
G
u
n
R
R
0
0
S2
S2 n
+
y
(x, y)
S
r
(, )
G(x, y, , ) =
(with 2 + 2 = 2 ).
We need to consider the function h(x, y, , ) to make G symmetric and zero on the boundary.
We can express this in polar coordinates, x = r cos , y = r sin ,
1
(r cos a2 / cos )2 + (r sin a2 / sin )2
G(r, , , ) =
+ h,
ln
4
(r cos cos )2 + (r sin sin )2
2
r + a4 /2 2a2 r/ cos( )
1
+ h.
ln
=
4
r 2 + 2 2r cos( )
66
1
h=
ln
4
2
a2
Note that,
2
a4
a2 r
1
1
ln r 2 + 2 2
cos( ) +
ln
4
4
a2
1
r 2 2
2
=
ln a + 2 2r cos( )
4
a
w(r, , , ) =
G is symmetric and zero on the boundary. This enable us to get the result for Dirichlet
problem for a circle,
Z 2
G
a d,
u(, ) =
f ()
r r=a
0
where
1
G
=
r
4
so
G
1
2 /a cos( )
G
a cos( )
=
=
,
n S
r r=a 2 a2 + 2 2a cos( ) a2 + 2 2a cos( )
=
Then
2 a 2
1
.
2 a a2 + 2 2a cos( )
1
u(, ) =
2
2
0
and relabelling,
a2 r 2
u(r, ) =
2
a2 2
f () d,
a2 + 2 2a cos( )
a2
r2
f ()
d.
2a r cos( )
Note that, from the integral form of u(r, ) above, we can recover the Mean Value Theorem.
If we put r = 0 (centre of the circle) then,
1
u(0) =
2
f () d,
0
i.e. the average of an harmonic function of two variables over a circle is equal to its value at
the centre.
67
Furthermore we may introduce more subtle inequalities within the class of positive harmonic
functions u 0. Since 1 cos( ) 1 then (a r) 2 a2 2ar cos( ) + r 2 (a + r)2 .
Thus, the kernel of the integrand in the integral form of the solution u(r, ) can be bounded
1 ar
a2 r 2
1
1 a+r
.
2
2
2 a + r
2 (a r) a 2ar cos( ) + r 2
2 a r
For positive harmonic functions u, we may use these inequalities to bound the solution of
Dirichlet problem for Laplaces equation in a disc
Z
Z
1 a + r 2
1 a r 2
f () d u(r, )
f () d,
2 a + r 0
2 a r 0
i.e. using the Mean Value Theorem we obtain Harnacks inequalities
a+r
ar
u(0) u(r, )
u(0).
a+r
ar
Example 5
Interior Neumann problem for Laplaces equation in a disc,
1
u
1 2u
2
u=
r
+ 2 2 = 0 in r < a,
r r
r
r
u
= f () on r = a.
n
Here, we need
1
G = (x )(y ) +
V
2
with
G
= 0,
r r=a
where V = a2 is the surface area of the disc. In order to deal with this term we solve the
equation
1
r2
1
2
r
=
(r)
=
+ c1 ln r + c2 ,
(r) =
r r
r
a2
4a2
and take the particular solution with c 1 = c2 = 0. Then, add in source at inverse point and
an arbitrary function h to fix the symmetry and boundary condition of G
G(r, , , ) =
1
ln r 2 + 2 2r cos( )
4
2
a
r2
2 r 2
1
2
ln 2 a + 2 2r cos( ) +
+ h.
a
4a2
So,
G
h
1
2r 2 cos( )
1
2r 2a2 / cos( )
r
+
=
+
,
2
2
2
4
2
2
2
r
4 r + 2r cos( ) 4 r + a / 2a r/ cos( ) 2a
r
1
1
a cos( )
a a2 / cos( )
h
G
=
+
,
+
+
r r=a
2 a2 + 2 2a cos( ) a2 + a4 /2 2a3 / cos( )
2a
r r=a
1
h
1 a cos( ) + 2 /a cos( )
,
+
+
=
2
2 + a2 2a cos( )
2a
r
r=a
68
1
h
G
1
+
+
=
r r=a
2 a 2a r r=a
and
h
=0
r r=a
implies
G
= 0 on the boundary.
r
On r = a,
h
2 i
1
1
ln a2 + 2 2a cos( )
+
,
4
4
1
1
=
ln a2 + 2 2a cos( )
.
2
2
Gr=a =
Then,
f () Gr=a a d,
Z 2
1
a
2
2
=u
ln a + 2a cos( )
f () d.
2 0
2
u(, ) = u
+
Indeed,
u dV =
u
dS
n
f () d = 0.
Z 2
a
ln a2 + 2 2a cos( ) f () d,
u(, ) = u
2 0
Z 2
a
or u(r, ) = u
ln a2 + r 2 2ar cos( ) f () d.
2 0
Exercise:
4.4
Extensions of Theory:
Alternative to the method of images to determine the Greens function G: (a) eigenfunction method when G is expended on the basis of the eigenfunction of the Laplacian
operator; conformal mapping of the complex plane for solving 2D problems.
Greens function for more general operators.
Chapter 5
Parabolic Equations
Contents
5.1
5.2
5.3
5.4
5.1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
69
72
75
78
Unlike elliptic equations, which describes a steady state, parabolic (and hyperbolic) evolution
equations describe processes that are evolving in time. For such an equation the initial state
of the system is part of the auxiliary data for a wellposed problem.
The archetypal parabolic evolution equation is the heat conduction or diffusion equation:
2u
u
=
t
x2
(1dimensional),
2u
(1D).
x2
Problems which are wellposed for the heat equation will be wellposed for more general
parabolic equation.
=
5.1.1
Consider > 0,
u
= 2 u in Rn , t > 0,
t
with u = f (x) in Rn at t = 0,
and
70
Note that we require the solution u(x, t) bounded in R n for all t. In particular we assume
that the boundedness of the smooth function u at infinity gives u = 0. We also impose
conditions on f ,
Z
Rn
Sometimes f (x) has compact support, i.e. f (x) = 0 outside some finite region.(E.g., in 1D,
see graph hereafter.)
u
+ x
5.1.2
5.1.3
If the initial conditions in a wellposed initial value or initialboundary value problem for an
evolution equation are replaced by conditions on the solution at other than initial time, the
resulting problem may not be wellposed (even when the total number of auxiliary conditions
is unchanged). E.g. the backward heat equation in 1D is illposed; this problem,
u
2u
=
in 0 < x < l, 0 < t < T,
t
x2
with u = f (x) at t = T, x (0, l),
for t (0, T ),
which is to find previous states u(x, t), (t < T ) which will have evolved into the state f (x),
has no solution for arbitrary f (x). Even when a the solution exists, it does not depend
continuously on the data.
The heat equation is irreversible in the mathematical sense that forward time is distinguishable from backward time (i.e. it models physical processes irreversible in the sense of the
Second Law of Thermodynamics).
71
5.1.4
f (x)2 dx < .
w=0
at t = 0
( < x < )
such that
I(0) = 0
and
and
w
= 0, t.
x
I(t) 0 t (as w 2 0),
Z
w
w 2
= w
dx (integration by parts),
x
x
Z
w
w 2
=
dx 0 since
= 0.
x
x
Then,
5.1.5
Similarly we can make use of the energy method to prove the uniqueness of the solution of
the 1D Dirichlet or Neumann problem
u
2u
=
in 0 < x < l, t > 0,
t
x2
with u = f (x) at t = 0, x (0, l),
or
u(0, t) = g0 (t)
u
(0, t) = g0 (t)
x
and
and
72
Suppose that u1 and u2 are two solutions and consider w = u1 u2 ; then w satisfies
w
2w
=
(0 < x < l, t > 0),
t
x2
with w = 0 at t = 0 (0 < x < l),
and
or
1
2
w2 (x, t) dx,
such that
I(0) = 0
and I(t) 0 t
(as w 2 0),
dx =
dx 0.
= w
x 0
x
x
0
0
Then,
0 I(t) I(0) = 0, t > 0,
since dI/dt < 0. So I(t) = 0 t > and w 0 and u 1 = u2 .
5.2
Example: To illustrate the typical behaviour of the solution of this Cauchy problem, consider the specific case where u(x, 0) = f (x) = exp(x 2 ); the solution is
x2
1
exp
u(x, t) =
1 + 4t
(1 + 4t)1/2
2
t exp(x
Starting with u(x, 0) = exp(x2 ) at t = 0, the solution becomes
u(x,
t)
1/2
/4t),
for t large, i.e. the amplitude of the solution scales as 1/ t and its width scales as t.
73
t=0
u
t=1
t = 10
Spreading of the Solution: The solution of the Cauchy problem for the heat equation
spreads such that its integral remains constant:
Z
Q(t) =
u dx = constant.
Proof:
Consider
dQ
=
dt
2u
dx (from equation),
2
x
u
=
= 0 (from conditions on u).
x
u
dx =
t
So, Q = constant.
5.2.1
To find the general solution of the Cauchy problem we define the Fourier transform of u(x, t)
and its inverse by
Z +
1
U (k, t) =
u(x, t) eikx dx,
2
Z +
1
u(x, t) =
U (k, t) eikx dk.
2
So, the heat equation gives,
Z +
1
U (k, t)
2
+ k U (k, t) eikx dk = 0
t
2
x,
which implies that the Fourier transform U (k, t) satisfies the equation
U (k, t)
+ k 2 U (k, t) = 0.
t
The solution of this linear equation is
2
U (k, t) = F (k) ek t ,
74
where F (k) is the Fourier transform of the initial data, u(x, t = 0),
Z +
1
f (x) eikx dx.
F (k) =
2
R +
(This requires f (x)2 dx < .) Then, we back substitute U (k, t) in the integral form
of u(x, t) to find,
Z + Z +
Z +
1
1
2
ik
k 2 t ikx
f () e
d ek t eikx dk,
u(x, t) =
F (k) e
e dk =
2
2
Z +
Z +
1
2
=
ek t eik(x) dk d.
f ()
2
Now consider
H(x, t, ) =
k 2 t ik(x)
dk =
"
x
exp t k i
2t
2
(x )2
4t
dk,
x
k t + ik(x ) = t k ik
t
2
= t
"
x
ki
2t
2
#
(x )2
,
+
4t2
Z +
2
since
es ds = (see appendix A).
So,
1
u(x, t) =
4 t
Where the function
Z +
(x )2
f () exp
d =
K(x , t) f () d
4t
2
1
x
K(x, t) =
exp
4t
4 t
is called the fundamental solution or source function, Greens function, propagator, diffusion kernel of the heat equation.
5.2.2
The function K(x, t) is solution (positive) of the heat equation for t > 0 (check this) and has
a singularity only at x = 0, t = 0:
75
At any time t > 0 (no matter how small), the solution to the initial value problem for the heat
equation at an arbitrary point x depends on all of the initial data, i.e. the data propagate
with an infinite speed. (As a consequence, the problem is well posed only if behaviouratinfinity conditions are imposed.) However, the influence of the initial state dies out very
rapidly with the distance (as exp(r 2 )).
5.2.3
Behaviour at large t
Suppose that the initial data have a compact support or decays to zero sufficiently quickly
as x and that we look at the solution of the heat equation on spatial scales, x, large
compared to the spatial scale of the data and at t large. Thus, we assume the ordering
x2 /t O(1) and 2 /t O() where 1 (so that, x/t O( 1/2 )). Then, the solution
Z
2
ex /4t +
2
f () e
d =
f () e /4t ex/2t d,
4 t
2
2 /4t Z +
x
x
F (0)
e
,
f () d ' exp
'
4t
4 t
2t
1
u(x, t) =
4 t
(x)2 /4t
1
ikx
F (k) =
f (x) dx = 2 F (0).
f (x) e
dx
2
is a constant and = x/ 2t is the diffusion variable. This solution spreads and decreases as
t increases.
5.3
Similarity Solution
For some equations, like the heat equation, the solution depends on a certain grouping of the
independent variables rather than depending on each of the independent variables independently. Consider the heat equation in 1D
2u
u
D 2 = 0,
t
x
and introduce the dilatation transformation
= a x,
= b t
and
u
w
w
= c
= ac
x
x
2
2
2u
ac w
2ac w
=
.
=
x2
2 x
2
bc
2w
w
2ac D
=0
i.e.
bc
w
2w
2ab D
= 0,
76
and is invariant under the dilatation transformation (i.e. ) if b = 2a. Thus, if u solves the
equation at x, t then w = c u solve the equation at x = a , t = b .
Note also that we can build some groupings of independent variables which are invariant
under this transformation, such as
a/b
a x
(b
t)
a/b
x
ta/b
which defines the dimensionless similarity variable (x, t) = x/ 2Dt, since b = 2a. (
if x or t 0 and = 0 if x = 0.) Also,
w
c/b
c u
(b t)
c/b
u
tc/b
= v()
suggests that we look for a solution of the heat equation of the form u = t c/2a v(). Indeed,
since the heat equation is invariant under the dilatation transformation, then we also expect
the solution to be invariant under that transformation. Hence, the partial derivatives become,
c
u
c c/2a1
1
=
t
v() + tc/2a v 0 ()
= tc/2a1
v() v 0 () ,
t
2a
t
2
a
tc/2a1/2 0
= tc/2a v 0 ()
=
v (),
x
x
2D
tc/2a1 00
2u
=
v ().
x2
2D
with = c/a, such that u = t/2 v and = x/ 2Dt. So, we may be able to solve the heat
equation through (5.1) if we can write the auxiliary conditions on u, x and t as conditions
on v and . Note that, in general, the integral transform method is able to deal with more
general boundary conditions; on the other hand, looking for similarity solution permits to
solve other types of problems (e.g. weak solutions).
5.3.1
Infinite Region
u 0 as x , t > 0.
u0
t=0
77
We look for a solution of the form u = t /2 v(), where (x, t) = x/ 2Dt, such that v() is
solution of equation (5.1). Moreover, since u = t /2 v() u0 as , where u0 does
not depend on t, must be zero. Hence, v is solution of the linear second order ODE
v 00 () + v 0 () = 0
with v u0 as
and v 0 as +.
2 /2 0
2
2 /2 00
e
v () + exp
v 0 () =
e
v () = 0 e /2 v 0 () = 0 ,
2
v () = 0 e
2 /2
v() = 0
h2 /2
dh + 1 = 2
/ 2
es ds + 1 .
v() = u0
5.3.2
es ds
x / 4Dt
i.e.
u(x, t) = u0
es ds .
SemiInfinite Region
Again, we look for a solution of the form u = t /2 v(), where (x, t) = x/ 2Dt, such that
v() is solution of equation (5.1). However, the boundary conditions are now different
t(1)/2 0
t(1)/2 0
u
/2 0
=
=t
v ()
=
v ()
v (0) = q,
x
x
x x=0
2D
2D
since q does notdepend on t, 1 must be zero. Hence from equation (5.1), the function v,
such that u = v t, is solution of the linear second order ODE
and v 00 = 20 + 00 .
i.e.
2 + 2
2
00
=
= .
0
78
2
1
2
e /2
ln   = 2 ln
+ k = ln 2
+ k 0 = 0
= 0
2
2
2
0
s2 /2
ds
e /2
+
+ 1 . = 2
es /2
ds + 1 .
s2
s2 /2
ds
+ 3 .
es
2 /2
ds + 3 ,
where the constants of integration 2 and 3 are determined by the initial conditions:
Z
Z
2
0
s2 /2
2 /2
2 /2
v = 2 e
e
ds + e
+
+ 3 = 2
es /2 ds + 3 ,
0
s2 /2
ds + 3
since
s2 /2
Z
ds = 2
h2
5.4
e
e
2 /2
2 /2
=q
4D
u(x, y) = q
4Dt
Z
+ 2
Z
2
2 /2
/ 2
e
+
e
/ 2
h2
dh
dh
+
e
/ 2
Dt
= 0 2 = 3
r
2
=
.
dh =
2
2
Z
2
x2 /4Dt
h2
h2
2
,
+ 3 ,
+ 2
!
+ 3 ,
2
dh ,
x/ 4Dt
h2
dh .
Like the elliptic PDEs, the heat equation or parabolic equations of most general form satisfy
a maximumminimum principle.
Consider the Cauchy problem,
u
2u
=
t
x2
in
< x < , 0 t T.
79
u 2 u
2 <0
t
x
in V
and u(x, 0) M,
u
=0
x
and
2u
0.
x2
But, 2 u/x2 0 is contradictory with the hypothesis 2 u/x2 > u/t = 0 at (x0 , t0 ).
Moreover, if we now suppose that the maximum occurs in t = T then, at this point
u
0,
t
u
=0
x
and
2u
0,
x2
5.4.1
Suppose
u 2 u
2 0
t
x
in V
and u(x, 0) M,
then u(x, t) M in VT .
Proof: Suppose there is some point (x 0 , t0 ) in VT (0 < t T ) at which u(x0 , t0 ) = M1 > M .
Put w(x, t) = u(x, t) (t t0 ) where = (M1 M )/t0 < 0. Then,
u 2 u
w 2 w
=
2 {z}
<0
t
x2
t {z x }
>0
and by lemma,
in VT ,
< M + t0 ,
M1 M
t0 ,
<M+
t0
w(x, t) < M1
in VT .
80
Appendix A
2
x
in R
Integral of e
Consider the integrals
I(R) =
s2
ds and
I=
es ds
(x2 +y 2 )
dx dy =
e(x
2 +y 2 )
dx dy.
R 2
R 2
R
R
+
R 2 x
Put, s =
so that ds = 2 d, to get
Z R2
Z 2
2R s
s
2
e ds < I (R) <
e ds i.e.
4 0
4 0
81
2
2
1 eR < I 2 (R) <
1 e2R .
4
4
82
Take the limit R + to state that
I2
4
4
So,
e
0
s2
i.e.
I =
4
2
and I =
(I > 0).
2
Z +
es ds = ,
ds =
2