aPDEs PAN2008 - Unknown

Chapter 1
Introduction to PDEs
1.1 Basic concepts and Examples
The dierential equations are the class of equations involving derivatives of unknown func-
tions. When the unknown function in the equation depends only on a single variable,
the equation involves only the ordinary derivatives of the unknown function, we call it an
ordinary dierential eqution, or in short ODE. Quite often, the unknown functions depend
on several independent variables, and the equations involving the partial derivatives of the
unknown functions, then the equations are called the partial dierential equations, or in
short PDEs. In this course, the independent variables will always be formed by a time
variable t 0 and a space variable x R
n
, with n the spatial dimensions. In general, a
PDE of the unknown function u(x
1
, x
2
, , x
n
) takes the form
F(x, u, Du, u
x
1
x
1
, u
x
1
x
2
, , u
xnxn
, ) = 0, (1.1)
where x = (x
1
, x
2
, , x
n
), Du = (u
x
1
, u
x
2
, , u
xn
), and F is a function of the indepen-
dent variable x and the unknown function u and nitely many partial derivatives of u. The
equation (1.1) is called m-th order, if the highest order of the derivatives of u in (1.1) is
m. An m-th order PDE of u(x
1
, x
2
, , x
n
) has a general form
F(x, u, Du, D
2
u, , D
m
u) = 0. (1.2)
Examples of Partial dierential Equations:
(a) Heat equation u
t
= u
xx
(b) Linear transport eqution u
t
+cu
x
= 0, c R
(c) Wave equation u
tt
c
2
u
xx
= 0, c > 0
(d) Burgers equation u
t
+uu
x
= 0
(e) Viscous Burgers equation u
t
+uu
x
= u
xx
5
6 CHAPTER 1. INTRODUCTION TO PDES
(f) Laplace equation u
xx
+u
yy
+u
zz
= 0
We now consider the equation (1.2). In addition to its order, we also classify the
equations by the relations between function F and u, as well as the derivatives of u. When
the dependence of F on u and its derivatives is linear, we say (1.2) is linear, otherwise we
say it is nonlinear. In the above examples, the (d) and (e) are nonlinear while the others
are linear. In the class of nonlinear PDEs, if F is linear with respect to the highest (m-th)
derivatives of u, i.e., the coecients of the m-th order derivatives of u in (1.2) only depend
on the independent variables x, u, and the derivatives of u up to (m 1)-th order, we
call the equation is quasi-liner. In the quasi-linear case, if the coecients of the highest
derivatives of u are functions of independent variable x only, the equation is then called
semi-linear. The general linear PDE of m-th order takes the following form:
P(x, D)u = f(x),
where, P(x, D) is a general m-th order dierential operator dened in (1.8) below. The
equation is called homogeneous if f(x) = 0.
We now introduce the concept of solution. Again, consider the equation (1.2), and we
restrict ourself on the domain x R
n
. u = (x) is a solution of (1.2) on , if (x) and
all of its derivatives appears in (1.2) are continuous in , and after substituting it into the
equation, it makes (1.2) an identity. Such a solution, we will refer it as a classical solution.
The following theorem gives an important properties for solutions to a linear PDE.
Theorem 1.1.1 (Principle of superposition) Let u
1
and u
2
be solutions to a given
homogeneous linear PDE . Then for any constants and ,
u
1
+u
2
is also a solution of that equation.
1.2 Well-posed problems
Just as what happened in ODEs, a PDE, if solvable, often has many solutions. However, as
most equations have their physical or practical backgrounds, some conditions or constraints
are necessary to determine a unique solution to the realistic problem. In general, there are
two classes of side conditions: the initial condition and the boundary conditions. Some
problems involve a mixture of initial and boundary conditions, called initial-boundary value
problems.
We rst start with the case of evolutionary PDEs, (or in which a time variable is
involved). In this case, the initial conditions often involve the initial value of the unknown
function, and the derivatives up to the next order of the highest time derivatives of the
unknown function. The initial conditions (also called initial data) together with the PDE
form an intial value problem, called Cauchy problem. The following example shows a typical
Cauchy problem for a wave equation in one spatial dimension:
1.2. WELL-POSED PROBLEMS 7
Example 1.2.1
u
tt
c
2
u
xx
= 0,
u(x, 0) = (x), u
t
(x, 0) = (x).
(1.3)
When the problem is conned in a given domain, the constraints on the boundary
are often needed for the problem. There are many dierent kinds of boundary conditions
depending on the realistic applications. In this course, we will mainly discuss three typical
boundary conditions. The rst one is to give the value of the unknown function itself at
the boundary of the domain. This type of boundary condition is called Dirichlet condition.
The corresponding boundary value problem is thus called Dirichlet problem. The following
example is a typical Dirichlet problem for a Poisson equation, describing the electronic
eld, with u the electronic potential and the electronic density distribution.
Example 1.2.2
u = 4(x, y, z), (x, y, z) R

3
,
u(x, y, z) = (x, y, z), (x, y, z) .
(1.4)
Here is the well-known Laplace operator. In R
n
, it takes the form
=
n
2
i
.
The second type is often called the Neumann condition, which assigns the normal derivative
of the unknow function at the boundary. The third type is the Robin condition, which gives
the nontrivial linear combination of the unknown function itself and its normal derivative
at the boundary. The next example shows a Neumann problem of Laplace equation in 3
dimension.
Example 1.2.3
(u
m
) = 0, (x, y, z) R
3
,
u = (x, y, z), (x, y, z) ,
where is the outer unit normal to .
The following example is an initial-boundary value problem for heat equation involving
the Robin boundary condition.
Example 1.2.4
u
t
u = 0, (x, y, z) R
3
, t > 0,
u
1
2
u = (x, y, z), (x, y, xz) , t > 0,
u(x, y, z, 0) = u
0
(x, y, z), (x, y, z) .
As the problems we encounter often arise from real applications, we thus expect each
of our problem is uniquely solvable and the solution continuously depends on the relevant
data. In this case, we say our problem is well-posed. However, it is very delicate to tell
whether a problem is well-posed without any further techniques. The following example is
due to J. Hadamard.
Example 1.2.5 Consider the set of initial-value problems in the upper half-plane in R
2
,
for n = 1, 2, ,
u = 0, y > 0,
u(x, 0) = 0, u
y
(x, 0) =
sin(nx)
n
, x R,
(1.5)
and

u = 0, y > 0,
u(x, 0) = 0, u
y
(x, 0) = 0, x R.
(1.6)
The problem (1.5) has a unique solution
u
n
(x, y) =
1
n
2
sinh(ny) sin(nx),
and the problem (1.6) has a unique soluton
u
0
(x, y) = 0.
We note that, as n , the data of (1.5) tends uniformly to the data of (1.6). However,
one has
lim
n
sup |u
n
(x, y) u
0
(x, y)| = , y > 0
for each (x, y). Thus, arbitrarily small changes in the data lead to large changes in the
solution, this is called instability.
1.2.1 Characteristic and initial value problems
We see from last example that the Cauchy problem for a Laplace equation is not well-posed.
This section is devoted to the study of how to set the initial value problem in a proper
way so that the initial data is compatible with the structure of the PDE. The notion of
characteristic plays an essential role in this context.
Lets consider a rather general linear PDE of order m:
P(x, D)u = g(x), (1.7)
where
P(x, D) =
||=m
a
(x)D
<m
a
(x)D
+a(x). (1.8)
1.2. WELL-POSED PROBLEMS 9
Here x = (x
0
, x
1
, , x
n
) R
n+1
, is a multi-index, = (
0
,
1
, ,
n
), where each
i
is a non-negative integer. || =
i
, is the length of . D = (D
0
, D
1
, , D
n
) is a
dierential operator where D
i
=

x
i
, and
D
=

||
x
0
0
x
n
n
.
We also dene x
= x
0
0
x
n
n
. We call
||=m
a
(x)D
the principal part of the operator P(x, D).

From the principle of superposition, we know that the solutions of (1.7) can be obtained
from a particular solution of (1.7) and a general solution of the following homogeneous
equation:
P(x, D) = 0. (1.9)
Assume that equation (1.9) is dened in a neighborhood of a smooth n-dimensional surface
S given by f(x) = 0. The initial value problem (or, Cauchy problem) for (1.9) consists of
assigning u and its derivatives of order (m1) on S, and it is required to solve (1.9) in
a neighborhood of S. For instance, in Example 1.2.1, the initial surface is S = t = 0, and
the initial conditions are u(x, 0) = (x), and u
t
(x, 0) = (x), and we shall try to solve the
equation
u
tt
c
2
u
xx
= 0
in the upper half plane {t > 0}.
We now proceed to solve (1.9) as follows. First, we shall make proper change of co-
ordinates in a neighborhood of S to distinguish the time-like direction so that we know
in which direction we will solve our problem. This can be done by introducing the new
independent variables y
0
, y
1
, , y
n
, where y
1
, , y
n
are independent coordinates on S,
and y
0
= f(x); i.e., S corresponds to y
0
= 0. Note that u is given on S, all the derivatives
of u with respect to the new variables y
1
, , y
n
on S.
By chain rule, we have, for = (
1
, ,
n
), || = m,
D
u =

u
x
0
0
x
n
n
=

m
u
y
m
0
y
0
x
0
y
0
x
n
n
+ ,
where the last dots represent derivatives of u with respect to y
0
of orders < m, together
with derivatives of u with respect to y
i
, i = 1, , n, and are thus all known quatities. The
equation (1.9) becomes
||=m
a
(x)
y
0
x
0
y
0
x
n
m
u
y
m
0
+ , (1.10)
where the dots are quantities known on S. Now it is clear that, if we want to solve this
equation for

m
u
y
m
0
, it is necessary and sucient that
||=m
a
(x)
y
0
x
0
y
0
x
n
n
= 0,
or equivalently, that
||=m
a
(x)(f(x))
||=m
a
(x)
f(x)
x
0
f(x)
x
n
n
= 0, x S. (1.11)
When (1.11) fails to hold, the initial-value problem would be unreasonable, and in this case
we say S is a characteristic surface of P(x, D). Formally, we have the following dention.
Denition 1.2.6 The surface S = {x : f(x) = 0} is said to be characteristic at a point
p S for the operator P(x, D) dened in (1.8) if
||=m
a
(x)(f(x))
|
x=p
= 0.
S is a characteristic surface for P(x, D) if it is characteristic at each point of S. The
equation
||=m
a
(x)
= 0, (1.12)
with = (
0
,
1
, ,
n
), is called the characteristic equation for the operator P(x, D) in
(1.8).
By these terms, we know that a surface S is characteristic at p S, for the operator
(1.8), provided that the normal vector to S at p satises the characteristic equation (1.12).
We remark that if f(x) = 0 is a characteristic surface for the operator (1.8), (1.11) shows
that the dierential equation (1.9) imposes an additional restrictions on the data; namely,
the known quantities, denoted by dots in (1.11) must vanish. Similar statement can be
made to the equation (1.7).
In the following, we assume is a unit normal vector given at a point of S, i.e.,
n
k=0
2
k
= 1. (1.13)
Now, we discuss several examples.
Example 1.2.7 The 3-dimensional wave equation:
u
x
0
x
0
k=1
u
x
k
x
k
= 0.
1.3. CLASSIFICATIONS OF SECOND ORDER SEMILINEAR PDES 11
(Here, one usually denotes x
0
as t) The characteristic equation is
2
0
k=1
2
k
= 0,
which together with (1.13) gives
0
=
1
2
. Therefore, a surface is characteristic for the
3-D wave equation if and only if its normal makes an angle of

4
with respect to the x
0
axis.
Example 1.2.8 Consider the (n + 1)-dimensional Laplace equation
n
k=0
u
x
k
x
k
= 0.
Here the characteristic equation reads
n
k=0
2
k
= 0
which is incompatible with (1.13). Therefore, there are no (real) characteristics for Laplace
equation.
Example 1.2.9 Consider the following rst-order linear equation
a(x, y)u
x
+b(x, y)u
y
= c(x, y)u +d(x, y).
The characteristic equation is
a
0
+b
1
= 0.
Solving this together with (1.13) gives
(
0
,
1
) =
1
a
2
+b
2
(b, a).
Therefore, the characteristic curves are solutions of the following system
x = a(x, y)
y = b(x, y).
This example is very useful in the next Chapter.
1.3 Classications of second order semilinear PDEs
In this section, we further discuss the types of PDEs, for which the dierent types of
equations often require dierent methods to resolve. As the decisive coecients of quisa-
linear PDEs depend on solutions, we will discuss semi-linear equations. Roughly speaking,
the classications of the semi-linear PDEs depend on how informations propagate. If
the information carried by solution propagates at a nite speed, we call it hyperbolic; if
the information propagates at an innite speed, we call it parabolic; if there is no (real)
speed for the information to travel with, we call it elliptic. From the knowledge in the
previous section, we know the latter often links to the problem without reasonable time-
like direction, or the problem is static.
1.3.1 Equations with several variables
To be precise, for n 2, we will now discuss only the second order semi-linear PDEs of
the following form
n
i,j=1
a
ij
(x)u
x
i
x
j
+F(x, u, Du) = 0, (1.14)
where, x = (x
1
, x
2
, , x
n
) with an open subset in R
n
, a
ij
= a
ji
as we expect
u
x
i
x
j
= u
x
j
x
i
. The linear principal part of this equation at a xed point p is
n
i,j=1
a
ij
(p)u
x
i
x
j
(p), (1.15)
which corresponds to a quadratic form
n
i,j=1
a
ij
(p)
i
i
= 0, = (
1
,
2
, ,
n
) R
n
(1.16)
This is the left hand side of the characteristic equation of (1.14). By linear algebra, we
know that the numbers of positive, negative and zero (real) eigenvalues of the matrix
A = (a
ij
(p)) is invariant under the transformation of the form P
T
AP for any invertible
n n matrix P. there is an orthogonal matrix O such that
O
T
AO = diag{
1
, ,
n
}, (1.17)
and the transformation = Oy changes the quadratic form into diagonal form
Q(y) =
n
i=1
i
y
2
i
. (1.18)
The way to transform Q() into its diagonal form is not unique. Upon re-scalings, we see
there is an invertible matrix U such that the transformation = Uy changes Q() into its
standard form
Q(y) =
n
i=1
i
y
2
i
. (1.19)
Here,
i
= 1, 1, or 0. (By Shurs lemma of linear algebra, we know that the numbers of
positive, negative and zero (real) eigenvalues of the matrix A = (a
ij
(p)) is invariant under
the transformation of the form P
T
AP for any invertible n n matrix P. ) If we x such
a U, the transformation
y = P
T
x
will transfer the equation (1.14) into its standard form:
n
i=1
i
u
x
i
x
i
+ = 0, (1.20)
where the dots terms contain at most rst order derivatives.
Now, we are able to give a precise classication for (1.14):
Denition 1.3.1 For (1.14) we have the following classications:
If all the eigenvalues of A have the same sign( all positive or all negative), we say
(1.14) is elliptic at p.
If A has zero eigenvalues, we call (1.14) parabolic at p.
If (n 1) eigenvalues of A have the same sign dierent from the other one, we say
(1.14) is hyperbolic at p. If both the number of positive eigenvalues and the number
of negative eigenvalues of A are greater than 1, and A had no zero eigenvalue, we
call (1.14) super hyperbolic.
Example 1.3.2 The following equation is super hyperbolic:
u
x
1
x
1
+u
x
2
x
2
u
x
3
x
3
u
x
4
x
4
= 0.
If the equation (1.14) is elliptic (or parabolic, or hyperbolic respectively) at each point
in , we say (1.14) is elliptic (or parabolic, or hyperbolic respectively). If (1.14) is of
dierent types on dierent points of , we call (1.14) of mixed type.
Example 1.3.3 Tricomi equation
yu
xx
+u
yy
= 0,
is of mixed type on any region including points on the x-axis.
Unfortunately, when the independent variables are more than 2, there are examples
showing that no matter how small the region is, there does not exist a single change of
variables such that the equation (1.14) is transfered into a single type. However, for only
two independent variables, under mild conditions on the coecients of the equation, it is
possible to make a change of variables such that the equation (1.13) was of the same type
on the whole region (very small sometime).
1.3.2 The case of two variables
Consider the semi-linear PDE of the second order with independent variables x and y
a(x, y)u
xx
+ 2b(x, y)u
xy
+c(x, y)u
yy
= F(x, y, u, u
x
, u
y
), (1.21)
where F, a, b and c are smooth functions with respect to their arguments. We also assume
that a, b and c are not all zero at any point on the working region . According to the
classications in last sub-section, equation (1.21) is classied as the sign of the following
determinant:
d := detA =
a(x, y) b(x, y)
b(x, y) c(x, y)
= a(x, y)c(x, y) b
2
(x, y). (1.22)
We then have the following cases:
If d > 0, equation (1.21) is elliptic. A typical example of elliptic equations is the
Laplace equation
u
xx
+u
yy
= 0,
where d = 1.
If d < 0, equation (1.21) is hyperbolic. A typical example of hyperbolic equation is
the wave equation
u
tt
c
2
u
xx
= 0, c > 0,
where d = c
2
.
If d = 0 and this matrix A is not identically zero, the equation (1.21) is parabolic. A
typical example of parabolic equation is the Heat equation
u
t
u
yy
= 0, > 0.
We now describe how to transfer (1.21) into standard form. The linear principal part
of (1.21) is
L
0
u = au
xx
+ 2bu
xy
+cu
yy
. (1.23)
For (x, y) , and any invertible smooth change of variables
= (x, y), = (x, y),
(, )
(x, y)
= 0, (1.24)
it is easy to compute that the linear principal part of (1.21) becomes
L
0
u = a
+ 2b
+c
, (1.25)
where

= a
2
x
+ 2b
x
y
+c
2
y
b
= a
x
x
+b(
x
y
+
y
x
) +c
y
y
c
= a
2
x
+ 2b
x
y
+c
2
y
.
(1.26)
Of course, we require the second order dierentiability of and . It is now clear that if
a
= c
= 0 and b
= 0, the equation (1.25) is of hyperbolic type, and it has a simple form

L
0
u = 2b
,
which is called second standard form of hyperbolic equations. In this case, we know that
and are solutions to the following equation
a
2
x
+ 2b
x
y
+c
2
y
= 0. (1.27)
This is exact the characteristic equation of L
0
. (x, y) = constant denes implicitly a
family of curves on xy-plane as y = y(x) (or, x = x(y) if necessary), which satises the
following ODE
a(
dy
dx
)
2
2b
dy
dx
+c = 0. (1.28)
We note that (1.27) is a fully nonlinear PDE for while (1.28) is an (nonlinear) ODE
which is easier to solve. As long as = 0, these two equations are equivalent in the
sense that (x, y) = constant is an implicit solution to (1.28).
From the theory of classication in last sub-section and (1.27)(1.28), we see that
(1.28) will not give any real nontrivial solution to (1.27) in the elliptic region. While in the
hyperbolic region, (1.28) gives two distinct families of real characteristic curves. Finally,
in the parabolic set of points, (1.28) gives only one family of real characteristic curves.
We now show how to utilize the characteristic curves to transfer (1.21) into standard
form. We will do so case by case.
First of all, assume that d < 0 in , so (1.21) is hyperbolic. We solve (1.28) to obtain
two distinct families of characteristic directions
dy
dx
=
b
b
2
ac
a
. (1.29)
Integrating them, we nd two families of characteristic curves
1
(x, y) = c
1
,
2
(x, y) = c
2
.
When
2
ix
+
2
iy
= 0 for i = 1, 2, (1.29) implies that
(
1
,
2
)
(x, y)
= 0.
Therefore, after the change of variables
=
1
(x, y), =
2
(x, y),
one obtains (1.25) with a
= c
= 0 and b
= 0. (1.21) becomes
u
+

F(, , u, u
, u
) = 0,
which, after a further change of variables into s and t such that
=
1
2
(s +t), =
1
2
(s t),
becomes the standard form
u
ss
u
tt
+F
1
(s, t, u, u
s
, u
t
) = 0,
for some smooth F
1
.
Secondly, we assume that d > 0 so that (1.21) is elliptic. We know that it is impossible
to obtain real-valued characteristic curves from (1.28). However, we are able to solve (1.28)
for a complex-valued solution
(x, y) =
1
(x, y) +i
2
(x, y) = constant,
where
1
and
2
are real functions, and i =
1 is the imaginary unit. It is not hard to

show that if
x
and
y
dont vanish at the same time, the following transformation
=
1
(x, y), =
2
(x, y),
satises
(
1
,
2
)
(x, y)
= 0.
Substituting + i into (1.27), separating the real and imaginary part, one discovers
a
= c
= 0 and b
= 0. Therefore, (1.21) takes the following standard form

u
+u
+G(, , u, u
, u
) = 0,
for some smooth function G.
Finally, we assume that d = 0 and thus (1.21) is parabolic. It is clear that ac = b
2
, and
one can assume a > 0 and c > 0 without loss of the generality since b = 0. From (1.28) we
obtain
dy
dx
=
c
a
.
Solve this equation, we have the family (x, y) = c
3
. Since = 0, = (x, y)
constant. Choose appropriate = (x, y) such that
(, )
(x, y)
= 0.
For instance, if = x, one has
(, )
(x, y)
=
y
= 0.
Otherwise, if
y
= 0, equation (1.27) implies
x
= 0, therefore constant, a contradic-
tion. For some isolated points where
y
= 0, one choose to exclude these points.
Now, under such a transformation, (1.21) becomes standard form
u
+G
1
(, , u, u
, u
) = 0
for some smooth function G
1
.
In the following, we show some examples.
Example 1.3.4 Discuss the type of the following equation
u
xx
+yu
yy
= 0 (1.30)
and change it into standard form.
Solution: We compute d = y. Therefore, this equation is elliptic in upper half plane
y > 0, is hyperbolic in lower half-plane y < 0, and is parabolic on x-axis. The ODE
associate to the characteristic equation is
(
dy
dx
)
2
+y = 0.
Case 1: When y = 0, one substitute y = 0 into (1.30), one gets the standard form
u
xx
= 0.
The characteristic curve in this case is the integral curve of
dy
dx
= 0, and that is x-axis
since y = 0.
Case 2: In the hyperbolic region where y < 0, we solve the ODE and found
= x + 2
y = c
1
, = x 2
y = c
2
,
and thus the transformation
= x + 2
y, = x 2
y,
transfers, with some standard calculations, into standard form
u
+
1
2( )
(u
) = 0, y < 0.
The characteristic curves in this case are two branches of the parabola y =
1
4
(x C)
2
,
where C is an arbitrary constant. The one with positive slope is corresponding to =
constant while the one with negative slope is corresponding to = constant. Both branches
are tangent to x-axis.
Case 3: In the elliptic region y > 0, the ODE will become a pair of conjugate complex
equations
dy
dx
i
y = 0.
Solving the one with plus sign (you can choose either one), we obtain
x 2i
y = c.
The real and imaginary parts are
= x, = 2
y,
which is our desired transformation. After some calculations, we arrive at
u
+u
= 0, y > 0.
Example 1.3.5 Discuss the type of the following equation
yu
xx
+ (x +y)u
xy
+xu
yy
= 0, (1.31)
and nd the general solution when x = y.
Solution: To solve the problem, we rst compute
d = xy
1
4
(x +y)
2
=
1
4
(x y)
2
0.
Therefore, when x = y, the equation(1.31) is parabolic; when x = y, the equation (1.31) is
hyperbolic. In the latter case, we substitute the parameters into (1.28) to nd either
dy
dx
= 1
or
dy
dx
=
x
y
.
Therefore, the two families of characteristic curves are
y x = c
1
, and y
2
x
2
= c
2
.
Which tells us, if we perform the transformation
= y x, = y
2
x
2
,
the Jacobian satises
(, )
(x, y)
= 2(x y) = 0
in hyperbolic region. Furthermore, the equation (1.31) takes the second standard form
u
+
1
= 0,
which is equivalent to
(u
= 0.
Now, we integrate the above equation with respect to ,
u
= f(),
where f is any integrable function. Therefore,
u =
1
f() d +g()
= g(y x) +
1
y x
h(y
2
x
2
),
where g and h are C
2
functions. This formula gives the general solution of (1.31) on the
hyperbolic region where x = y.
1.4. PROBLEMS 19
1.4 Problems
Problem 1. For the following equations, identify which are linear, semi-linear, quasi-liner,
or fully nonlinear, and their orders.
(a) u
tt
u = 0;
(b) u
t
+u
xxxx
= 0;
(c) div(|Du|
p2
Du) = 0;
(d) div
Du
(1 + |Du|
2
)
= 0;
(e) u
t
+divF(u) = 0, F : R R
n
;
(f) u
t
(u
) = 0;
(g) u
t
u = f(u);
(h)
u
t
+u Du u +Dp = 0,
divu = 0; u R
3
, x R
3
;
(i) u
t
+divF(u) = 0, F : R
n
R
nn
.
Problem 2. Let = {(x, y) : 0 < x < 1, 0 < y < 1}, and consider the following boundary
value problem
u
xx
u
yy
= 0, (x, y) ,
u(x, 0) = f
1
(x), u(x, 1) = f
2
(x),
u(0, y) = g
1
(y), u(1, y) = g
2
(y).
where f
1
, f
2
, g
1
and g
2
are given functions. Is this problem well-posed?
Problem 3. Find the type of the following equation
3u
xx
2u
xy
+ 2u
yy
2u
yz
+ 3u
zz
+ 5u
y
u
x
+ 10u = 0.
Transfer it into its standard form.
Problem 4. Identify the types of the following equations:
(a) xu
xx
+ 2yu
xy
+yu
yy
= 0;
(b) u
xx
+ (x y)
3
u
yy
= 0;
(c) yu
xx
+ (x +y)u
xy
+xu
yy
= 0;
(d) sin(x)u
xx
2 cos(x)u
xy
(1 + sin(x))u
yy
= 0;
(e) e
z
u
xy
u
xx
= log(x
2
+y
2
+z
2
+ 1).
(f) 7u
xx
10u
xy
22u
yz
+ 7u
yy
16u
xz
5u
zz
= 0.
Problem 5. Transfer the following equations into standard form.
(a) x
2
u
xx
+ 2xyu
xy
+y
2
u
yy
= 0;
(b) u
xx
+xyu
yy
= 0;
(c) u
xx
2 cos(x)u
xy
(3 + sin
2
(x))u
yy
yu
y
= 0;
(d) y
2
u
xx
e
2x
xu
xy
+u
x
= 0, x > 0.
Problem 6. Determine the type of Tricomi equation
u
xx
+xu
yy
= 0,
and then transfer it into standard form.
Problem 7. Transfer the following equation
u
xx
+yu
yy
+
1
2
u
y
= 0
into standard form. Find the general solution.
Problem 8. Show that for any second order hyperbolic or elliptic PDEs with two variables
and constant coecients, one can always combine the change of variables (1.24) and the
the following transformation of unknown function
u = v exp( +)
to obtain a form of
v
+cv = f.
Problem 9. Based on the problem 8, classify the following equations and transfer them
into a standard form without rst order derivatives.
(a) u
xx
+ 4u
xy
+ 3u
yy
+ 3u
x
u
y
+ 2u = 0;
(b) u
xx
+ 2u
xy
+u
yy
+ 5u
x
+ 3u
y
+u = 0;
(c) u
xx
6u
xy
+ 12u
yy
+ 4u
x
u = sin(xy).
1.4. PROBLEMS 21
Problem 10. Make the change of unknown function u = v +w with v the new unknown
functions, such that the following problems have homogeneous boundary conditions. Where
(a) has Neumann boundary condition, (b) has Dirichlet boundary condition, while (c) has
one Neumann condition and one Robin condition.
(a)
u
tt
c
2
u
xx
= 0, 0 < x < +, t > 0,
u
x
(0, t) = g(t), t 0,
u(x, 0) = (x), u
t
(x, 0) = (x), 0 x < +;
(b)
u
tt
c
2
u
xx
= 0, 0 < x < l, t > 0,
u(0, t) = (t), u(l, t) = (t), t 0,
u(x, 0) = (x), u
t
(x, 0) = (x), 0 x < l;
(c)
u
tt
c
2
u
xx
= 0, 0 < x < l, t > 0,
u
x
(0, t) = (t), u
x
(l, t) +u(l, t) = (t), t 0,
u(x, 0) = (x), u
t
(x, 0) = (x), 0 x < l.
Problem 11. Determine w such that the change of unknown function u = vw changes
the equation
u
t
u
xx
+au
x
+bu = f(x, t)
into
v
t
v
xx
= f
1
(x, t).
Problem 12. Assume u is a solution of the heat equation
u
t
a
2
u
xx
= 0,
with the form
u(x, t) = u(
x
t
).
Derive the ODE for u, and then solve the following problem
u
t
a
2
u
xx
= 0, 0 < x < +, t > 0,
u(0, t) = 0, t 0,
u(x, 0) = u
0
, 0 x < +,
where u
0
is a constant.
Chapter 2
The rst order quasi-linear PDEs
The rst order quasi-linear PDEs have the following general form:
F(x, u, Du) = 0, (2.1)
where x = (x
1
, x
2
, , x
3
) R
n
, u = u(x), Du is the gradient of u. Such equation often
appears in physical applications, such as the famous transport equation, Burgers equation
and the Hamilton-Jacob equation
u
t
+ H(Du) = 0, (2.2)
where 0. The distinguished feature of this class is that one can solve it through a
system of ODEs at least locally. Such a method is called method of characteristic. We will
rst introduce the general theory in two variables in rst two sections, then we generalize
it into general cases. We then discuss in details on the semi-linear case in two variables
with some examples. The transport equation will be presented in fourth section with a
direct approach.
2.1 Characteristic curves and integral surfaces
Consider the following rst order quasi-linear PDE with two variables:
a(x, y, u)u
x
+ b(x, y, u)u
y
= c(x, y, u). (2.3)
Here, u = u(x, y). (2.3) can be rewritten in the following form
(a, b, c) (u
x
, u
y
, 1) = 0. (2.4)
Keeping this equation in mind, the vector eld (a, b, c) (called characteristic direction)
therefore plays an essential role in solving (2.3). On some interval I R, this vector eld
determines the following parametric curves (called characteristic curves)
: x = x(t), y = y(t), z = z(t), t I R,
23
24 CHAPTER 2. THE FIRST ORDER QUASI-LINEAR PDES
through the following equation
dx
a(x, y, z)
=
dy
b(x, y, z)
=
dz
c(x, y, z)
,
i.e.
_
_
dx
dt
= a(x, y, z)
dy
dt
= b(x, y, z)
dz
dt
= c(x, y, z).
(2.5)
We call (2.5) the characteristic equation of (2.3). We see such curves, if exist, have the
characteristic direction as tangent vectors. We remark that, from calculus, the integral
surface S : z = u(x, y), as the solution of (2.3), is tangent to the characteristic direction
everywhere. The following theorem states the relationship between and S.
Theorem 2.1.1 If there is a point P = (x
0
, y
0
, z
0
) of the characteristic curve lying on
the integral surface S : z = u(x, y), then S.
Proof: Assume that
: x = x(t), y = y(t), z = z(t), t I R,
is a solution of (2.5) such that for some parameter t = t
0
, it holds that
x
0
= x(t
0
), y
0
= y(t
0
), z
0
= z(t
0
) = u(x
0
, y
0
).
Furthermore, we know that
P = (x
0
, y
0
, z
0
) S.
Dene
w = w(t) = z(t) u(x(t), y(t)).
It is easy to check that w(t) is the solution to the following initial value problem
_
_
_
dw
dt
= c(x, y, w + u) u
x
a(x, y, w + u) u
y
b(x, y, w + u),
w(t
0
) = 0.
(2.6)
Clearly, w 0 is a solution of (2.6) since z = u(x, y) is a solution of (2.3). Due to
ODE theory, we know that (2.6) has a unique solution. Therefore, w(t) 0, namely,
z(t) = u(x(t), y(t)), which means S.
From the denition of S, we know that, for each point on S, there is a characteristic
curve passing through. Therefore, S is the union of all characteristic curves. Now, it is not
hard to see from argument we made above that we did prove the following result.
2.2. CAUCHY PROBLEM 25
Theorem 2.1.2 The general solution of a rst-order, quasi-linear PDE (2.3) is
f(, ) = 0,
where f is an arbitrary function of (x, y, u) and (x, y, u) and = constant = c
1
and
= constant = c
2
are solution curves of
dx
a(x, y, u)
=
dy
b(x, y, u)
=
du
c(x, y, u)
,
or equivalently, equation (2.5) replacing z with u.
2.2 Cauchy problem
We now discuss the initial value problem for (2.3). We know from section 1.2 that, in order
to have the well-posed Cauchy problem, one should not prescribe the initial data on the
characteristic curves. The following theorem conrms this observation.
Theorem 2.2.1 Suppose that x
0
(s), y
0
(s) and u
0
(s) are continuous dierentiable func-
tions of s in a closed interval 0 s 1 and that a, b and c are functions of x, y and
u with continuous rst order partial derivatives with respect to their arguments in some
domain D of (x, y, z)-space containing the initial curve
: x = x
0
(s), y = y
0
(s), u = u
0
(s), (2.7)
where s [0, 1], and satisfying the condition
_
a(x
0
(s), y
0
(s), u
0
(s)) b(x
0
(s), y
0
(s), u
0
(s))
x
0
(s) y
0
(s)
_
= 0. (2.8)
Then, there exists a unique solution u = u(x, y) of (2.3) in the neighborhood of C : x =
x
0
(s), y = y
0
(s), s [0, 1], and the solution satises the initial condition
u
0
(s) = u(x
0
(s), y
0
(s)), s [0, 1]. (2.9)
Remark 2.2.2 The curve C, on which the Cauchy data is prescribed, is the projection of
the initial curve onto the (x, y)-plane. The condition (2.8) is precisely equivalent to the
curve C is not characteristic in the sense of section 1.2.
Roughly speaking, a proof of this theorem can be described as follows: One starts with
a time t = 0 and solve the equation (2.5) with the initial data (2.7) which gives a unique
solution
x = X(s, t), y = Y (s, t), z = Z(s, t), s [0, 1], t [0, T], (2.10)
for some positive T, such that
(X(s, 0), Y (s, 0), Z(s, 0)) = (x
0
(s), y
0
(s), u
0
(s)).
Now, with the condition (2.8), one can solve (s, t) in terms of x and y to obtain
s = (x, y), t = (x, y).
Substituting these into z = Z(s, t), one has
u = Z(s, t) = Z((x, y), (x, y)) u(x, y),
the solution of the Cauchy problem. Of course, in general, the solutions exist only in a
neighborhood of the initial curve , and such a solution is call a local solution. The rigorous
proof of this theorem involves the implicit function theorem and we shall not pursuit it
here.
2.3 General cases
We now generalize the theory we established in the previous two sections to the following
general rst order quasi-linear PDEs:
n
k=1
a
k
(x
1
, x
2
, , x
n
, u)u
x
k
= c(x
1
, x
2
, , x
n
, u), n 2. (2.11)
The corresponding system of ODEs for characteristic curves reads
_
_
dx
k
dt
= a
k
(x
1
, x
2
, , x
n
, z), k = 1, , n,
dz
dt
= c(x
1
, x
2
, , x
n
, z).
(2.12)
The corresponding Cauchy problem is thus to nd the integral surface of (2.11) z =
u(x
1
, x
2
, , x
n
) such that it pass through the following parametrized (n 1) dimension
initial surface S:
S :
_
x
k
= f
k
(s
1
, s
2
, , s
n1
), k = 1, , n,
z = u
0
(s
1
, s
2
, , s
n1
).
(2.13)
To proceed, we could rst nd the characteristic curves passing through a point on S with
parameter s = (s
1
, , s
n1
) to obtain the solution of (2.12)
_
x
k
= X
k
(s
1
, s
2
, , s
n1
, t), k = 1, , n,
z = Z(s
1
, s
2
, , s
n1
, t),
(2.14)
which satises the initial datum given in (2.13). Now, similar to (2.8), we propose the
condition
J = det
_
_
_
_
_
a
1
a
2
a
n
(s)
f
1
s
1
f
2
s
1

f
n
s
1
.
.
.
.
.
.
.
.
.
.
.
.
f
1
s
n1
f
2
s
n1

f
n
s
n1
_
_
_
_
_
S
= 0, (2.15)
2.4. SOME EXAMPLES 27
which is equivalent to that the initial surface is not characteristic. Under this condition
(2.15), one can solve from (2.14) for s = (s
1
, s
2
, , s
n1
) and t in terms of x
1
, , x
n
, and
then upon to a substitution, the solution u = Z(x
1
, , x
n
) to (2.11) with initial condition
(2.13) is obtained.
2.4 Some examples
In this section, we will show some examples on how to apply the method of characteristic
established so far to solve some Cauchy problems on rst order quasi-linear PDEs.
Example 2.4.1 Consider the semi-linear equation
a(x, y)u
x
+ b(x, y)u
u
= f(x, y, u), (2.16)
where the functions a, b, and f are smooth functions. In this case, we could simply solve
the following system of ODEs
dx
dt
= a(x, y),
dy
dt
= b(x, y),
which determines a family of curves (x(t), y(t)) on (x, y)-plane. This family of curves are
actually the projection of the characteristic curves on to (x, y)-plane. Then we solve the
equation
dz
dt
= f(x(t), y(t), z)
to obtain the characteristic curve : (x(t), y(t), z(t)).
Example 2.4.2 We now solve the following Cauchy problem
_
u
t
+ vu
x
= 0,
u(x, 0) = u
0
(x),
(2.17)
where v > 0 is a constant, and u
0
is a regular function. First of all, one easily veries that
the initial curve
: t = 0, x = , z = u
0
(x)
is not characteristic, as the determinant dened in (2.8) is 1.
In view of Example 2.4.1, we solve the following
_
x(s) = v,
t(s) = 1,
(2.18)
where indicates the derivative with respect to the parameter s. By imposing the
initial conditions at s = 0,
_
x(0) = x
0
,
t(0) = t
0
,
we determine a unique solution for (2.18) in the explicit form
_
x(s) = vs + x
0
,
t(s) = s + t
0
.
Since the initial condition is assigned on t = 0, which implies (x
0
, t
0
) is a point of the curve
and t
0
= 0. So the characteristic curve may be rewritten, eliminating the parameter s, as
x(t) = vt + x
0
. (2.19)
This is a line in (x, t)-plane intersecting {t = 0} at x
0
with the slope
1
v
or, equivalently
speed v.
Now let (t) = u(x(t), t) be a solution of (2.17). Along the characteristic curve (2.19)
we have:
(t) = x(t)u
x
(x(t), t) + u
t
(x(t), t) = vu
x
(x(t), t) + u
t
(x(t), t) = 0,
i.e., the solution of (2.17) is constant along the characteristics (2.19). So you can resolve
the ordinary dierential equation for :
(t) = (0),
i.e.
u(vt + x
0
, t) = u(x(t), t) = u(x(0), 0) = u
0
(x
0
).
To determine the value of solution u at a generic point in (x, t)-plane, one simply determines
the point x
0
of intersection between the characteristic passing through (x, t) and the axis
{t = 0}, that is, from x = vt + x
0
to obtain
x
0
= x vt.
Ultimately, the solution of (2.17) is given by
u(x, t) = u
0
(x vt).
Example 2.4.3 In this example, we solve the following Cauchy problem
_
_
_
x
u
x
+ 2y
u
u
+
u
z
= 3u,
u(x, y, 0) = (x, y).
(2.20)
On the initial surface
S : x = s
1
, y = s
2
, z = 0, w = (s
1
, s
2
),
2.4. SOME EXAMPLES 29
we compute
J = det
_
_
_
_
_
x 2y 1
x
s
1
y
s
1
z
s
1
x
s
2
y
s
2
z
s
2
_
_
_
_
_
= det
_
_
s
1
2s
2
1
1 0 0
0 1 0
_
_
= 1 = 0.
Therefore, we expect the problem (2.20) has a unique solution near S. We solve the initial
value problem
_
_
dx
dt
= x,
dy
dt
= 2y,
dz
dt
= 1,
dw
dt
= 3w,
(x, y, z, w)|
t=0
= (s
1
, s
2
, 0, (s
1
, s
2
)),
to obtain
x = s
1
e
t
, y = s
2
e
2t
, z = t, w = (s
1
, s
2
)e
3t
.
From the rst three equations in the above, we have
t = z, s
1
= xe
z
, s
2
= ye
2z
.
Therefore, the solution of the problem (2.20) is
w = u(x, y, z) = (xe
z
, ye
2z
)e
3z
.
Example 2.4.4 Solve the following problem
_
tu
t
+ xu
x
= cu, x R, t 0,
u(x, 1) = f(x),
(2.21)
where c is a constant.
For this problem, we note that the coecients (x, t, c) is singular at (0, 0, c) for c = 0,
where the equation does not make sense except for u 0. The initial curve is
: x = s, t = 1, z = f(s),
which is not characteristic if s = 0. Solving the system of ODEs
dx
d
= x,
dt
d
= t,
dz
d
= cz,
one obtains the characteristic curves
x(, s) = se
, t(, s) = e
, z(, s) = f(s)e
c
.
Therefore,
s =
x
t
, = log t
and the solution is
u(x, t) = z = f(
x
t
)t
c
.
We note that, in the particular case where f constant, we see from the above formula
that if c > 0 we have solution u(x, t) for all t 0. On the other hand, if c < 0, we see that
u(x, t) as t tends to zero. In the latter case, we say the solution blows up at t = 0.
We now use this example to explain what will happen if the initial curve is characteristic.
Now, we x c = 1, and choose the initial curve
1
: x(0, ) = , t(0, ) =

, z(0, ) =

for s = 0, and = 0, are arbitrary constants. One easily checks that both u(x, t) =

x,
and u(x, t) = t are solutions for the equation with the new initial data. We found innitely
many solutions in this case.
Example 2.4.5 We now consider the following problem
_
u
t
+ uu
x
= 0
u(x, 0) = u
0
(x).
(2.22)
In view of example 2.4.1 and 2.4.2, one easily nd the characteristic curves as the solution
of
dx
dt
= u(x, t), x(0) = x
0
.
Also, we know that the solution u(x, t) is constant along each characteristic curves. There-
fore, the characteristic curves are the family of straight lines
x(t) = x
0
+ tu
0
(x
0
).
So, the solution of u at each point (x, t) is thus determined by the characteristic line from
certain x
0
on {t = 0} passing through (x, t) where u(x, t) = u
0
(x
0
).
We remark that this process cannot go far in general. Suppose there are two point
a < b on x-axis such that u
0
(a) > u
0
(b). We know that the characteristic line issuing from
(a, 0) will overtake the characteristic line from (b, 0) at
t =
b a
u
0
(a) u
0
(b)
.
In this case, we could not assign a value for u(x, t) at this intersection point, which corre-
sponds to two dierent values. Actually, the solution blows up in this case in its derivative.
More precisely, we carry out the following calculations. Let w = u
x
, which satises
w
t
+ uw
x
= w
2
.
2.5. LINEAR TRANSPORT EQUATION 31
Therefore, along the characteristic line, we have
_
w = w
2
w(x
0
, 0) = u
0
(x
0
).
It is now clear that
w(x
0
+ tu
0
(x
0
), t) =
u
0
(x
0
)
1 + u
0
(x
0
)t
.
Therefore, w tends to innity in nite time if u
0
(x
0
) < 0.
2.5 Linear Transport Equation
The linear transport equation
u
t
+ b Du = 0, x R
n
, t > 0 (2.23)
arises from many important applications, such as particle mechanics, kinetic theory. Here
b = (b
1
, b
2
, , b
n
) is a constant vector. Of course, one can easily solve this PDE using the
characteristic method. Here, we will use a direct approach instead. Actually, the (2.23)
can be rewritten as
(Du, u
t
) (b, 1) = 0,
which means the function u(x, t) is constant in the direction of (b, 1) on the (x, t)-space.
Therefore, if we know the value of u at any point of the straight line through (x, t) with
the direction (b, 1) in (x, t)-space, we know the value of u(x, t).
Now let us consider the initial-value problem
_
u
t
+ b Du = 0, x R
n
, t > 0,
u(x, 0) = g(x), x R
n
,
(2.24)
where g(x) is a given function. Fix a point (x, t), the line through (x, t) in the direction
ofd (b, 1) is given in the parametric form with the parameter s as
(x + sb, t + s), s R.
The line hits the initial plane {t = 0} when s = t at the point (x bt, 0). We know that
u is constant on the line, we thus obtain
u(x, t) = g(x tb), x R
n
, t 0. (2.25)
We see from the above that if (2.24) has a smooth solution, then it is given by (2.25).
Conversely, if g is continuous dierentiable, the (2.25) is the unique solution of (2.24). We
also remark that the solution in the form of (2.25) reads as the information traveling in a
constant velocity b, such a solution is thus called traveling wave solution.
Remark 2.5.1 When g(x) is not C
1
, problem (2.24) does not admit a smooth solution.
However, even when g is not continuous, the formula (2.25) does provide a reasonable
candidate for a solution. We may informally declare that (2.25) is a weak solution of
(2.24). This makes sense even if g is discontinuous and so is u. This notion is very useful
in nonlinear PDEs.
Next, we consider the non-homogeneous problem
_
u
t
+ b Du = f(x, t), x R
n
, t > 0,
u(x, 0) = g(x), x R
n
.
(2.26)
As in the homogeneous case, we set
z(s) = u(x + sb, t + s), s R,
which satises
z(s) = Du(x + sb, t + s) b + u
t
(x + sb, t + s) = f(x + sb, t + s).
Therefore,
u(x, t) g(x bt) = z(0) z(t) =
_
0
t
z(s) ds
=
_
0
t
f(x + sb, t + s) ds
=
_
t
0
f(x + (s t)b, s) ds,
(2.27)
which gives the solution
u(x, t) = g(x tb) +
_
t
0
f(x + (s t)b, s) ds. (2.28)
2.6 Problems
Problem 1. Solve the following problems (x, y, z R, t > 0)
(a) u
t
+ uu
x
= 1, u(x, 0) = h(x);
(b) u
x
+ u
y
= u, u(x, 0) = cos(x);
(c) xu
y
yu
x
= u, u(x, 0) = h(x);
(d) x
2
u
x
+ y
2
u
y
= u
2
, u(x, y)|
y=2x
= 1;
(e) xu
x
+ yu
y
+ u
z
= u, u(x, y, 0) = h(x, y);
2.6. PROBLEMS 33
(f)
n
k=1
x
k
u
x
k
= 3u, u(x
1
, , x
n1
, 1) = h(x
1
, , x
n1
);
(g) uu
x
uu
y
= u
2
+ (x + y)
2
, u(x, 0) = 1;
(h) 2xyu
x
+ (x
2
+ y
2
)u
y
= 0, u = exp
_
x
x y
_
, on x + y = 1;
(i) xu
x
+ yu
y
= u + 1, u(x, y) = x
2
on y = x
2
.
Problem 2. Solve the Cauchy problem
_
u
t
+ b Du + cu = 0, x R
n
, t > 0,
u(x, 0) = g(x), x R
n
.
where c R and b R
n
are constant.
Problem 3. For the following problem
_
_
_
u
t
+ uu
x
= 0
u(x, 0) =
1
1 + x
2
,
nd the solution and the time and the point where the solution blows up rst.
Problem 4. If u is the C
1
solution of the following equation
a(x, y)u
x
+ b(x, y)u
y
= u
on the closed unit disk = {(x, y) : x
2
+ y
2
1}. Suppose
a(x, y)x + b(x, y)y > 0, on x
2
+ y
2
= 1,
prove that u 0.
Problem 5. Show that the solution of the equation
yu
x
xu
y
= 0
containing the curve x
2
+ y
2
= a
2
, u = y, does not exit.
Problem 6. Solve the following Cauchy problems:
(a) x
2
u
x
y
2
u
y
= 0, u e
x
as y ;
(b) yu
x
+ xu
y
= 0, u = sin(x) on x
2
+ y
2
= 1;
(c) xu
x
+ yu
y
= 1, in 0 < x < y, u(x) = 2x on y = 3x;
(d) 2xu
x
+ (x + 1)u
y
= y, in x > 0, u(1, y) = 2y;
(e) xu
x
2yu
y
= x
2
+ y
2
in x > 0, y > 0, u = x
2
on y = 1;
Problem 7. Show that u
1
= e
x
, and u
2
= e
y
are solutions of the nonlinear equation
(u
x
+ u
y
)
2
u
2
= 0,
but that their sum (e
x
+ e
y
) is not a solution of the equation.
Problem 8. Solve the following equations:
(a) (y + u)u
x
+ (x + u)u
y
= x + y,
(b) xu(u
2
+ xy)u
x
yu(u
2
+ xy)u
y
= x
4
.
Problem 9. Solve the equation
xz
x
+ yz
y
= z,
and nd the curves which satisfy the associated characteristic equations and intersect the
helix x
2
+ y
2
= a
2
, z = b tan
1
(y/x).
Problem 10. Find the family of curves which represent the general solution of the PDE
(2x 4y + 3u)u
x
+ (x 2y 3u)u
y
= 3(x 2y).
Determine the particular member of the family which contains the line u = x and y = 0.
Problem 11. Find the solution of the equation
yu
x
2xyu
y
= 2xu
with the condition u(0, y) = y
3
.
Problem 12. Find the solution surface of the equation
(u
2
y
2
)u
x
+ xyu
y
+ xu = 0, u = y = x, x > 0.
Chapter 3
Laplace Equation
Undoubtedly, the Laplace Equation is among the most important PDEs. Traditionally, we
call
u(x) = 0, x R
n
, n 2. (3.1)
the Laplace Equation, and its non-homogeneous companion
u(x) = f(x), x R
n
, n 2, (3.2)
the Poisson Equation.
We recall that the Laplacian of u is
u =
n
i1
u
x
i
x
i
.
In both equations, x = (x
1
, , x
n
) R
n
and the unknown is u :

R. In (3.2),
the function f : R is given fucntion. For later use, we have the following dention:
Denition 3.0.1 If u C
2
() satises (3.1) in , we call it a harmonic function. u
C
2
() is called sub-harmonic (or super-harmonic) if it satises
u 0( 0).
Laplace equation and Poisson equations model many static physical elds with and
without sources. They appear in the modeling of gravity, electric force, and the chemical
concentration in equilibrium, and much more. Let V be any smooth subregion in , the
net ux of u through V is zero:
_
V
F dS = 0,
where F is the ux density and the unit outer normal of V . By Divergence Theorem
(see Theorem 3.1.1 below), we have
35
36 CHAPTER 3. LAPLACE EQUATION
_
V
F dx =
_
V
F dS = 0,
and so
F = 0, in .
In many occasions, the ux F is proportional to the gradient Du, pointing in the opposite
direction. Therefore, one has
F = aDu, a > 0. (3.3)
If u stands for the
_
_
chemical concentration
temperature
electrostatic potential,
the equation (3.3) is
_
_
Ficks law of diusion
Fouriers law of heat conduction
Ohms law of electrical conduction.
Laplace equation also arises in the study of analytic functions and the probabilistic inves-
tigation of Brownian motion.
3.1 The fundamental solution
From the form of Laplace equation, we see that all directions have the same weights in
the equation. Namely, the equation is invariant under any orthogonal transformation of
coordinates. We thus seek to solve the equation explicitly by looking for the radial solution
which has the form
u(x) = v(r), r = |x|,
where v is to be determined so that (3.1) holds. For i {1, . . . , n},
u
x
i
= v
(r)r
x
i
= v
(r)
x
i
r
, u
x
i
x
i
= v
(r)
x
2
i
r
2
+ v
(r)
_
1
r

x
2
i
r
3
_
.
One has
u = v
(r) +
n 1
r
v
(r) = 0.
If v
= 0, we deduce
(log(v
))
=
v
=
1 n
r
and hence
v
(r) =
a
r
n1
,
3.1. THE FUNDAMENTAL SOLUTION 37
for some constant a. Therefore, if r > 0, we nd
v(r) =
_
_
_
b log r + c for n = 2
b
r
n2
+ c for n 3,
where a and b are constants. We therefore discover the fundamental solution of Laplace
equation:
(x) =
_
1
2
log |x| for n = 2
1
n(n 2)(n)
1
|x|
n2
for n 3,
(3.4)
where
(n) =
2
n/2
n(n/2)
is the volume of the unite ball in R
n
. The reason for particular choices of the constants
will be explained later.
From the derivation, we also have the following estimates
|D(x)|
C
|x|
n1
, |D
2
(x)|
C
|x|
n
, (x = 0), (3.5)
for some constant C > 0.
3.1.1 Greens formula
In this subsection, we summarize some useful formulas due to Green. These formulas
are convenient in the computations related to Laplacian. These are derived from the
Divergence Theorem (or, Gauss formula):
Theorem 3.1.1 Let be a bounded domain in R
n
with smooth boundary . Let be
the unit outer normal of . For any smooth vector eld w C
1
(
), it holds that
_
w dx =
_
w dS. (3.6)
Let be a bounded domain in R
n
with smooth boundary . Let be the unit outer
normal of . For u C
2
(
), one derives from the Divergence Theorem (letting w = Du)

that
_
u dx =
_
Du dS =
_
dS. (3.7)
Now, for u, v C
2
(
), by choosing w = uDv or w = vDu respectively in the Divergence

Theorem, we have
_
uv dx =
_
u
v
dS
_
Du Dv dx (3.8)
_
vu dx =
_
v
u
dS
_
Dv Du dx. (3.9)
We subtract the above two equations to get
_
(uv vu) dx =
_
(u
v
v
u
) dS. (3.10)
Traditionally, (3.8) is called the rst Greens formula, while (3.10) is called the second
Greens formula.
3.1.2 Poisson Equation in R
n
From the construction, we know that the fundamental solution (x) of Laplace equation
is harmonic for x = 0, so is (x y) for x = y. With this in mind, we will prove the
following
Theorem 3.1.2 If f C
2
0
(R
n
), then
u(x) =
_
R
n
(x y)f(y) dy C(R
n
) (3.11)
is a solution of the Poisson equation (3.2).
Proof. First of all, we have
u(x) =
_
R
n
(x y)f(y) dy =
_
R
n
(y)f(x y) dy,
hence
u(x + he
i
) u(x)
h
=
_
R
n
(y)
_
f(x + he
i
y) f(x y)
h
_
dy,
where h = 0 and e
i
the unit vector in the direction of x
i
-axis. Note that
lim
h0
f(x + he
i
y) f(x y)
h
=
f
x
i
(x y)
uniformly on R
n
, and thus
u
x
i
(x) =
_
R
n
(y)
f
x
i
(x y) dy, (i = 1, , n). (3.12)
Similarly,
2
u
x
i
x
j
(x) =
_
R
n
(y)

2
f
x
i
x
j
(x y) dy, (i, j = 1, , n). (3.13)
3.1. THE FUNDAMENTAL SOLUTION 39
We note that the integral above is continuous in the variable x, and thus u C
2
(R
n
).
As (x) is singular at x = 0, we have to pay substantial attention near the singularity.
Fix > 0, let B(0, ) be the ball centered at 0 with radius . We have
u(x) =
_
B(0,)
(y)
x
f(x y) dy +
_
R
n
\B(0,)
(y)
x
f(x y) dy
=: I
+ J
.
(3.14)
Now, it is clear that
I
CD
2
f
L
(R
n
)
_
B(0,)
|(y)| dy
_
C
2
| log | (n = 2)
C
2
(n = 3).
(3.15)
On the other hand, we can estimate J
as following
J
=
_
R
n
\B(0,)
(y)
x
f(x y) dy
=
_
R
n
\B(0,)
(y)
y
f(x y) dy
=
_
R
n
\B(0,)
D
y
(y) D
y
f(x y) dy +
_
B(0,)
(y)
f
v
(x y) dS(y)
=: K
+ L
,
(3.16)
where is the inward unit normal along B(0, ). We easily check
|L
| CDf
L
(R
n
)
_
B(0,)
|(y)| dy
_
C| log | (n = 2)
C (n = 3).
(3.17)
It remains to compute K
. In fact, we have
K
=
_
R
n
\B(0,)
D
y
(y) D
y
f(x y) dy
=
_
R
n
\B(0,)
y
(y)f(x y) dy
_
B(0,)
(y)f(x y) dS(y)
=
_
B(0,)
(y)f(x y) dS(y).
We now note that
D(y) =
1
n(n)
y
|y|
n
, y = 0,
and
=
y
|y|
=
y
, on B(0, ),
So,
(y) =
1
n(n)
n1
, on B(0, ).
We also note that n(n)
n1
is the surface area of B(0, ), we have
K
=
1
n(n)
n1
_
B(0,)
f(x y) dS(y)
f(x), as 0.
(3.18)
We therefore conclude from (3.14)(3.18) (letting 0) that
u = f.
Remark 3.1.3 Sometimes, we write
=
0
, in R
n
,
where
0
is the Dirac measure on R
n
giving unit mass to the point 0. Formally, one can
computes:
u(x) =
_
R
n
x
(x y)f(y) dy
=
_
R
n
x
f(y) dy = f(x), (x R
n
).
3.1.3 Fundamental integral formulas
As we often solve the elliptic equations in bounded domains, we now carry out some useful
integral formulas using the fundamental solution (x).
Assume u C
2
(). For any y , we choose > 0 suitably small so that the ball
B
(y) centered at y with radius is inside . On the region \

B
(y), we substitute v(x)

with (x y) in the second Greens formula,
_
B
(y)
(x y)
x
u dx =
_
(
u
) dS
x
+
_
B
(y)
(
u
) dS
x
. (3.19)
Similar to the computations carried out in (3.17)(3.18), we have
|
_
B
(y)
dS
x
| = |()
_
B
(y)
u
dS
x
|
= | ()
_
B
(y)
u dx|
_
C
2
| log | (n = 2)
C
2
(n = 3).
0, as 0;
3.2. PROPERTIES OF HARMONIC FUNCTIONS 41
and
_
B
(y)
u
dS
x
=
()
_
B
(y)
u dS
x
=
1
n(n)
n1
_
B
(y)
u dS
x
u(y), as 0.
Therefor, we obtain from (3.19) by letting 0 that
u(y) =
_
(x y)
x
u dx
_
(u
(x y)
(x y)
u
) dS
x
, y . (3.20)
This formula is called Greens representation of u(y).
In particular, if u has compact support on , then it holds
u(y) =
_
(x y)
x
u dx, y , u C
2
0
(). (3.21)
If u is harmonic in , we thus obtain the fundamental integral formula of harmonic
functions
u(y) =
_
(u
(x y)
(x y)
u
) dS
x
, y . (3.22)
We remark that, actually, for smooth harmonic function, (3.7) implies
_
dS = 0. (3.23)
3.2 Properties of harmonic functions
In this section, we discuss some basic properties of harmonic functions. We now consider an
open bounded set R
n
and suppose u is harmonic in . Various interesting properties
will be presented in orders.
3.2.1 Mean-value formulas
The mean-value formulas, which declare that u(x) equals both the average of u over the
sphere B(x, r) and the average of u over the whole ball B(x, r), as long as B(x, r) .
It will play key roles in many important occasions.
To begin, we introduce the notion of mean value of u(x) over a domain :
(u)
=
1
||
_
u(x) dx. (3.24)

Theorem 3.2.1 If u C
2
() is harmonic, then
u(x) = (u)
B(x,r)
= (u)
B(x,r)
, (3.25)
for each ball B(x, r) .
Proof. By (3.22), using (3.23), one has
u(x) =
_
B(x,r)
(u
(x y)
(x y)
u
) dS
y
=
(r)
_
B(x,r)
u(y) dS
y
(r)
_
B(x,r)
u
dS
y
=
(r)
_
B(x,r)
u(y) dS
y
= (u)
B(x,r)
.
Now, for the mean-value on the ball, we have
_
B(x,r)
u(y) dy =
_
r
0
(
_
B(x,s)
u(y) dS
y
) ds
= u(x)
_
r
0
_
B(x,s)
dS
y
ds = u(x) (volume of B(x, r)).
This completes the proof.
Remark 3.2.2 If instead of harmonic, u C
2
() is sub-harmornic, then one has
u(x) (u)
B(x,r)
, u(x) (u)
B(x,r)
. (3.26)
If u C
2
() is super-harmornic, then
u(x) (u)
B(x,r)
, u(x) (u)
B(x,r)
. (3.27)
The converse of the mean-value formulas is also a true statement.
Theorem 3.2.3 If u C
2
() satises
u(x) = (u)
B(x,r)
for each ball B(x, r) , then u is harmonic.
Proof. Set
(r) = (u)
B(x,r)
= (u(x + rz))
B(0,1)
.
Then
(r) = (Du(x + rz) z)

B(0,1)
=
1
|B(x, r)|
_
B(x,r)
Du(y)
y x
r
dS
y
=
1
|B(x, r)|
_
B(x,r)
u
dS
y
=
r
n
(u)
B(x,r)
If u = 0, there exists some ball B(x, r
1
) such that u > 0 in B(x, r) (the other case
is similar). But
0 =
(r
1
) =
r
1
n
(u)
B(x,r
1
)
> 0,
a contradiction.
3.2.2 Maximum principle and uniqueness
Assume that is an open and bounded set in R
n
. We rst present
Theorem 3.2.4 (Strong maximum principle). Suppose u C
2
() C(
) is harmonic in
, then
(i) max
u = max
u
(ii) If is connected and there exists a point x
0
such that
u(x
0
) = max
u,
then u is constant in .
Remark 3.2.5 The rst assertion in this theorem is the maximum principle and the
second is the strong maximum principle. If one replaces u by u, the similar assertions
are true when max is replaced by min.
Proof. Suppose there exists a point x
0
with
u(x
0
) = M = max
u.
Then for some r > 0, the ball B(x
0
, r) . By the mean-value formula on this ball, we
have
M = u(x
0
) = (u)
B(x
0
,r)
M.
Here, the equality holds only if u M in B(x
0
, r). Therefore, u(x) = M for any x
B(x, r). So, the set {x |u(x) = M} is both open and relatively closed in , and thus
equals if is connected. This proves assertion (ii). The rst one follows immediately.
The strong maximum principle has many applications. A direct application is as follows:
if is connected and u C
2
() C(
) is a solution of
_
u = 0, in ,
u = g, on ,
where g 0, then u is positive everywhere in if g is positive somewhere on .
An important application of the maximum principle is the uniqueness of solutions to
certain boundary value problems for Poisson equation.
Theorem 3.2.6 (Uniqueness). Let g C(), f C(). Then there exists at most one
solution u C
2
() C(
) of the boundary value problem

_
u = f, in ,
u = g, on .
(3.28)
Proof. If u
1
and u
2
are two solutions, one applies the maximum principle to the har-
monic functions w
= (u
1
u
2
) which satises the Laplace equation with zero boundary
condition, therefore, w
0.
3.2.3 Harnacks inequality
In this subsection, we will exploit an amazing averaging eect of Harmonic functions.
Recall that we denote by V if

V is compact. The following theorem assert that
the value of a non-negative harmonic function on are all comparable within a compact
subset relative to : it cannot be very large ( or very small) at any point in this subset
unless it is very large (or very small) everywhere. The idea is that since the compact set
has positive distance from , there is room for the averaging eect of Laplaces equation
to occur.
Theorem 3.2.7 (Harnacks inequality). For each connecterd open set V , there
exists a positive constant C, depending only on V , such that
sup
V
u C inf
V
u
for all nonnegative harmonic functions u in .
Remark 3.2.8 The Harnacks inequality in particular implies that, for any x and y V ,
it holds that
1
C
u(y) u(x) C u(y).
Proof. Let r =
1
4
dist(V, ). Fixing any x V , for any y V such that |x y| r, we
have
u(x) = (u)
B(x,2r)
1
(n)2
n
r
n
_
B(y,r)
u dz
1
2
n
(u)
B(y,r)
=
1
2
n
u(y).
Therefore, we actually proved that
2
n
u(y) u(x) 2
n
u(x), x, y V, |x y| r.
Since V is connected and

V is compact, we can cover

V by a chain of nitely many balls
{B
i
}
N
i=1
, each of which has radius r and B
i
B
i1
= , for i = 2, , N. Then,
2
nN
u(y) u(x) 2
nN
u(y)
for all x, y V .
3.2.4 Regularity
Now we prove that if u C
2
() is harmonic, then necessarily u C
(). Thus harmonic

functions are innitely dierentiable. This sort of assertion is called a regularity statements.
It is interesting to see that the algebraic structure of Laplace equation leads to that all the
partial derivatives of u exist, even those which do not appear in the PDE.
Theorem 3.2.9 (Smoothness). If u C() satises the mean-value property (3.25) for
each ball B(x, r) , then u C
().
Before we proceed to prove this theorem, we rst introduce an important tool. Dene
C
(R
n
) by
(x) =
_
C exp{
1
|x|
2
1
} if |x| < 1,
0 if |x| > 1,
(3.29)
where C > 0 is chosen so that
_
R
n
(x) dx = 1.
Now, for each > 0, set
=
n
(
x
).
Such is called the standard mollier, and
(R
n
) statises
_
R
n
(x) dx = 1, supp(
) B(0, ). (3.30)
The following lemma states some of the properties of molliers.
Lemma 3.2.10 If f : R is locally integrable, dene
f
f =
_
(x y)f(y) dy,
for x
= {x | dist(x, ) > }. Then, it holds that

(i) f
).
(ii) f
f, a.e. as 0.
(iii) If f C(), then f
f uniformly on compact subsets of .

We now give a proof to the regularity theorem using the mean-value formulas.
Proof. Let be a standard mollier, we will prove that u = u
on
. In fact, if x
,
then
u
(x) =
_
(x y)u(y) dy
=
n
_
B
(x, )(
|x y|
)u(y) dy
=
n
_

0
(
r
)(
_
B(x,r)
u(y) dS) dr
=
n
u(x)
_

0
(
r
)n(n)r
n1
dr
= u(x)
_
B(0,)
dy = u(x).
Therefore, u C
) for any > 0, and so u C
().
Remark 3.2.11 One should be careful that u is not assumed to have any regularity at
boundary, even continuity.
A further applications of mean-value formulas will lead to the estimates on derivatives,
for which we omit the proof.
Theorem 3.2.12 (Estimates on derivatives). Assume u is harmonic in . Then it holds
that
|D
u(x
0
)|
C
k
r
n+k
u
L
1
(B(x
0
,r))
(3.31)
for each ball B(x
0
, r) and each multi-index of order || = k. Here, the C
k
can be
chosen as
C
0
=
1
(n)
, C
k
=
(2
n+1
nk)
k
(n)
, k = 1, . (3.32)
The C
regularity does not go to extreme of the harmonic functions. The following

theorem states that harmonic functions are actually analytic.
Theorem 3.2.13 If u is harmonic in , then u is analytic in .
The next one conrms the analyticity of harmonic functions, which says that there are
no nontrivial bounded harmonic functions on whole R
n
.
Theorem 3.2.14 (Liouvilles Theorem). If u : R
n
R is harmonic and bounded, then u
is constant.
3.3. GREENS FUNCTION 47
Proof. Fix x
0
R
n
, r > 0, and apply Theorem 3.2.12 on B(x
0
, r),
|Du(x
0
)|
C
1
r
n+1
u
L
1
(B(x
0
,r))
C
1
(n)
r
u
L
(R
n
)
0, as r .
Therefore, Du 0, and so u is constant.
As a direct application of Liouvilles Theorem, we have
Theorem 3.2.15 (Representation formula). Let f C
2
0
(R
n
), n 3. Then any bounded
solution of
u = f in R
n
has the form
u(x) =
_
R
n
(x y)f(y) dy + C, (x R
n
)
for some constant C.
Proof. Since (x) 0 as |x| for n 3,
u(x) =
_
R
n
(x y)f(y) dy
is a bounded solution of
u = f in R
n
.
If u is another bouned solution, w = u u is bounded and harmonic, and thus is a constant.
Remark 3.2.16 When n = 2, (x) is not bounded as |x| , and so it is possible that
_
R
n
(x y)f(y) dy
is not bounded as |x| . Therefore, the representation formula is not true in general
for n = 2.
3.3 Greens function
In last section, we obtained representation formula for problems on R
n
. We now x to
be an bounded open domain in R
n
with smooth boundary . We will try to nd a general
representation formula for the solutions of the following boundary value problem
_
u = f in ,
u = g on .
(3.33)
We intend to solve this problem through the Greens representation formula (3.20). We
recall (3.20) below
u(x) =
_
(xy)
y
u(y) dy
_
(u(y)
(x y)
(xy)
u(y)
) dS
y
, x . (3.34)
It is clear that (3.34) permits us to solve the problem if we know how to deal with the term
of
u
on the boundary. Unfortunately, it is unknown to us. The idea is to introduce a

correction h(x, y) for each xed x such that it solves the following boundary value problem.
_

y
h = 0 in ,
h = (x y) on .
(3.35)
By the Greens formula, we have
h(x, y)
y
u(y) dy =
_
(u(y)
h(x, y)
h(x, y)
u
) dS
y
=
_
(u(y)
h(x, y)
(x y)
u
) dS
y
.
(3.36)
We therefore arrived at the following denition.
Denition 3.3.1 Greens function for the region is
G(x, y) = (x y) h(x, y), (x, y , x = y).
With this notion, we add (3.36) to (3.34) to nd
u(x) =
_
G(x, y)
y
u(y) dy
_
u(y)
G(x, y)
dS
y
, x . (3.37)
where
G
(x, y) =
y
G(x, y) (y),
is the outer normal derivative of G with respect to variable y. The term
u
does not
appear in this formula (3.37). This means the correction h(x, y) is in its proper form.
Now, suppose u(x) C
2
(
) solves the boundary value problem (3.33) for some contin-

uous functions f and g, using (3.37), we arrive at
Theorem 3.3.2 (Representation formula with Greens function). If u(x) C
2
(
) solves
problem (3.33), then
u(x) =
_
G(x, y)f(y) dy
_
g(y)
G(x, y)
dS
y
, x . (3.38)
Remark 3.3.3 Formally, (3.38) gives a nice formula to the solutions of the Poisson equa-
tion with Dirichlet boundary condition if we know the Greens function on the domain ,
such a method is called Greens method. However, for general domain, it is a dicult task
to construct the Greens function, for which requires to solve the problem (3.35). However,
Greens method is still signicant in the following reasons:
(i) For Poisson equation on a xed domain, once we obtained the Greens function,
the existence of solutions to the problem (3.33) is given by the formula (3.38) for any
continuous f and g.
(ii) In the case when it is hard to nd the solutions to (3.33), one can still use the
formula (3.38) to discuss the certain behavior of the solutions.
(iii) For some domains with simple geometry, explicit calculation of G is possible.
The Dirichlet problem of Poisson equation on such domains are often important in
the applications.
(iv) (3.38) transfers the (3.33) into a integral equation, which is convenient in certain
occasions even when the equation is semi-linear. The machinery of functional analysis
is thus applied to obtain some interesting results.
Before going to specic examples, we discuss certain important properties of Greens
function.
First of all, it is clear that in , when x = y, G(x, y) is harmonic on x and on y
everywhere. Furthermore, G(x, y) as x y at the order of |x y|
n2
if n > 2 and
at the order of log |x y| if n = 2.
Secondly, substituting u(x) = 1 into (3.37) we have
_
dS = 1. (3.39)
Finally, we show that G(x, y) is symmetric in x and y.
Theorem 3.3.4 (Symmetry of Greens function). For all x, y , x = y, it holds that
G(y, x) = G(x, y).
Proof. Formally, we prove the theorem as following. For x = y , (|x y|) is smooth
on , by denition, we know that
G(y, x) = (x y) h(x, y), G(y, x) = (y x) h(y, x).
We note that (xy) = (yx) = (|xy|), and therefore, both h(x, y) and h(y, x) are the
harmonic solutions of the same problem (3.35). By uniqueness we know h(x, y) = h(y, x)
and therefore
G(y, x) = G(x, y), x = y .
3.3.1 Greens function on a ball
We will construct Greens function for the unit ball B(0, 1) using some reection through
the sphere B(0, 1).
Denition 3.3.5 If x R
n
\ {0}, the point
x =
x
|x|
2
is called the point dual to x with respect to B(0, 1). The mapping x x is inversion
through the unit sphere B(0, 1).
We will use this inversion to construct Greens function for the unit ball = B(0, 1).
Fix x B(0, 1). We need to nd a correction h(x, y) solving
_

y
h = 0 in B(0, 1),
h = (x y) on B(0, 1),
(3.40)
then the Greens function reads
G(x, y) = (x y) h(x, y).
We try to invert the singularity of (x y) from x B(0, 1) to x B(0, 1). Assume
now that n 3. The mapping
y (y x)
is harmonic for y = x. Thus,
y |x|
2n
(y x)
is harmonic for y = x, and so
h(x, y) = (|x|(y x)) (3.41)
is harmonic in B(0, 1). Furthermore, if y B(0, 1) and x = 0,
|x|
2
|y x|
2
= |x|
2
_
|y
2
2y x
|x|
2
+
1
|x|
2
_
= |x|
2
2y x + 1 = |x y|
2
.
Therefore, (|x||y x|)
2n
= |x y|
2n
and so
h(x, y) = (y x), y B(0, 1). (3.42)
This veries that h(x, y) is the one we were looking for.
Denition 3.3.6 Greens function for the unit ball is
G(x, y) = (y x) (|x|(y x)), x, y B(0, 1), x = y. (3.43)
We now solve the following boundary value problem
_
u = 0, in B(0, 1)
u = g, on B(0, 1).
(3.44)
By (3.38), we need to calculate
G
on the unit sphere. According to formula (3.43), for

y B(0, 1) one has
G
y
i
=
(y x)
y
i
(|x|(y x))
y
i
,
where
(y x)
y
i
=
1
n(n)
y
i
x
i
|x y|
n
,
and
(|x|(y x))
y
i
=
1
n(n)
y
i
|x|
2
x
i
(|x||y x|)
n
=
1
n(n)
y
i
|x|
2
x
i
|x y|
n
, y B(0, 1).
Therefore,
G
=
n
i=1
y
i
G
y
i
(x, y)
=
1
n(n)
1
|x y|
n
n
i=1
y
i
((y
i
x
i
) y
i
|x|
2
+ x
i
)
=
1
n(n)
1 |x|
2
|x y|
n
, y B(0, 1).
Thus, we use (3.38) to yield the representation formula
u(x) =
1 |x|
2
n(n)
_
B(0,1)
g(y)
|x y|
n
dS
y
. (3.45)
If now instead of (3.43), for r > 0, we want to solve the following boundary-value
problem
_
u = 0, in B(0, r)
u = g, on B(0, r).
(3.46)
Then, u(x) = u(rx) solves (3.43) with g(x) = g(rx) replacing g in (3.43). After a direct
change of variables, we obtain the Poissons formula
u(x) =
r
2
|x|
2
n(n)r
_
B(0,r)
g(y)
|x y|
n
dS
y
, x B(0, 1). (3.47)
The function
K(x, y) =
r
2
|x|
2
n(n)r
1
|x y|
n
, x B(0, r), y B(0, r), (3.48)
is called Poissons Kernel for the ball B(0, r).
(3.47) was established under the assumption that (3.46) has a smooth solution. We
now prove that in fact (3.47) gives a solution.
Theorem 3.3.7 (Poissons formula for a ball). Assume g C(B(0, r)) and u is dened
by (3.47). Then u is harmonic in B(0, r) and for each point x
0
B(0, r),
lim
xx
0
u(x) = g(x
0
), x B(0, r).
Proof. It is clear that K(x, y) 0 is harmonic when x = y. Therefore, for x B(0, r)
and y B(0, r), we have
u(x) =
_
B(0,r)
x
K(x, y)g(y) dS
y
= 0.
To verify the boundary condition, we rst remark that
_
B(0,r)
K(x, y) dS
y
= 1. (3.49)
Indeed, for each xed y B(0, r), we denote x = z where 0 < 1 and z B(0, r).
By the mean-value formula for harmonic function, we have
1 = K(0, y)n(n)r
n1
=
_
B(0,r)
K(z, y) dS
z
.
Now,
1 =
_
B(0,r)
K(z, y) dS
z
=
_
B(0,r)
K(y, z) dS
z
=
_
B(0,r)
K(x, z) dS
z
.
Now, x x
0
B(0, r), > 0. Choose > 0 so small that
|g(y) g(x
0
)| < , if |y x
0
| < , y B(0, r). (3.50)
Then, if |x x
0
| <

2
, x B(0, r), setting
V
= B(0, r) B(x
0
, ),
we compute
|u(x) g(x
0
)| =
_
B(0,r)
K(x, y)[g(y) g(x
0
)] dS
y
_
V
K(x, y)|g(y) g(x

0
)| dS
y
+
_
B(0,r)\V
K(x, y)|g(y) g(x

0
)| dS
y
I + J.
(3.51)
Now, (3.49)(3.50) implies that
I
_
B(0,r)
K(x, y) dS
y
= .
For J, we note that if |x x
0
| <

2
and |y x
0
| , then
|y x|
1
2
|y x
0
|.
Thus,
J 2g
L
_
B(0,r)\V
K(x, y) dS
y
2
n+1
g
L
(r
2
|x|
2
)
n(n)r
_
B(0,r)\V
|y x
0
|
n
dS
y
0, as x x
0
.
Therefore, we could choose >
> 0 so small such that

|u(x) g(x
0
)| < 2, if |x x
0
| <
.
This proves that
lim
xx
0
u(x) = g(x
0
), x B(0, r).
3.3.2 Greens function on half space
Again, we x n 3. Let us consider the half space
R
n
+
= {x = (x
1
, , x
n
) R
n
|x
n
> 0}.
Although this region is unbounded, and so the calculations for Theorem 3.3.2 is not
valid. We will try to build Greens function using the ideas developed so far. Later, we
will check directly that the derived representation formula gives the solution. We will also
use the reection idea about the boundary of the domain.
Denition 3.3.8 If x = (x
1
, , x
n1
, x
n
) R
n
+
, its reection in the plane R
n
+
is the
point
x = ((x
1
, , x
n1
, x
n
).
We set
h(x, y) = (y x) = (y
1
x
1
, , y
n1
x
n1
, y
n
+ x
n
), (x, y R
n
+
).
we note that
h(x, y) = (y x), if y R
n
+
,
and hence
_
h(x, y) = 0 in R
n
+
h(x, y) = (y x) on R
n
+
.
We thus has
Denition 3.3.9 Greens function for the half-space R
n
+
is
G(x, y) = (y x) (y x), (x, y R
n
+
, x = y).
Clearly, if y R
n
+
,
G
(x, y) =
G
y
n
(x, y) =
2x
n
n(n)
1
|x y|
n
.
Suppose u is a solution to the boundary-value problem
_
u = 0 in R
n
+
u = g on R
n
+
.
(3.52)
Then, formally, we expect from (3.38) that
u(x) =
_
R
n
+
K(x, y)g(y) dy, (x R
n
+
) (3.53)
to be a representation formula for the solution. Here the function
K(x, y) =
2x
n
n(n)
1
|x y|
n
, (x, y R
n
+
, x = y), (3.54)
is the Poissons kernel for R
n
+
and (3.53) is called the Poissons formula.
Similar to the case for a ball, we could verify directly that
Theorem 3.3.10 (Poissons formula for half-space). Assume g C(R
n1
L
(R
n1
),
and u is dened by (3.53). Then u is uniformly bounded harmonic function on R
n
+
and for
each x
0
R
n
+
, it holds that
lim
xx
0
u(x) = g(x
0
), x R
n
+
.
3.4. HOPFS MAXIMUM PRINCIPLE 55
3.4 Hopfs maximum principle
We have put a lot of eorts in the last section for the Dirichlet boundary value problem
for Poisson equation. The Greens function method is particularly designed for this type of
problems. The second boundary value problem, namely the Nuemann problem, is not well
studied. We will establish the maximum principle for Nuemann problem in this section.
Theorem 3.4.1 (Hopf s Lemma) Let B(y, R) R
n
(n 3), x
0
B(y, R), u
C
2
(B(y, R)) C
1
(

B(y, R)) is sub-harmonic on B(y, R) such that
u(x
0
) > u(x), x B(y, R),
then
u
(x
0
) > 0,
where is the outer unit normal of B(y, R) at x
0
.
Proof. For (0, R), and a positive parameter > 0, we dene
v(x) = e
r
2
e
R
2
, r = |x y| > .
Direct computation gives
v = e
r
2
(4
2
r
2
2na).
Therefore, if we choose big enough, say =
n
, then v 0 on the region A =

B(y, R) \

B(y, ). We note that u(x) u(x
0
) < 0 on B(y, ), there exists a > 0 such
that
w(x) = u(x) u(x
0
) + v(x) 0, x B(y, ).
We note that v(x) = 0 on B(y, R) and thus
w(x) 0, x B(y, ).
We also note that w(x) is sub-harmonic on A, therefore the maximum principle for w on
A implies that
w(x) 0, x A.
But w(x
0
) = 0 and thus
u
(x
0
) 0,
that is
u
(x
0
)
v
(x
0
) = v
(R) > 0.
We now introduce a concept concerning the structure of the boundary.
Denition 3.4.2 Let x
0
, if there exists a ball B such that {x
0
} =

B

, we
say satises the inner ball condition at x
0
. Meanwhile, we say R
n
\

satises the outer
ball condition at x
0
.
With this notion, based on Hopfs lemma, we are able to prove the following
Theorem 3.4.3 (Hopf s maximum principle) Assume u C
2
() C
1
(
) and u 0
(u 0), and x
0
such that
u(x
0
) > u(x) (u(x
0
) < u(x)), x .
If staises the inner ball condition at x
0
, then
u
(x
0
) > 0 (
u
(x
0
) < 0).
We now apply Hopfs maximum principle to the Neumann problem. Consider
_
u = f, in
u
= g, on .
(3.55)
It is easy to see that the solution of the above Neumann problem, if exists, is not unique.
For if u is one solution, then u + C for any constant C is another solution. However, we
could prove the following
Theorem 3.4.4 If satises the inner ball condition at each boundary point, then the
solutions of Neumann problem (3.55) can only dier by a constant.
Proof. Let u
1
and u
2
be two solutions to (3.55), then w = u
1
u
2
is the solution of
_
w = f, in
w
= 0, on .
Now, if w is not constant, by the maximum principle for harmonic function w, we know w
attains its maximum at some x
0
. According to Hopfs maximum principle, we know
w
(x
0
) > 0
which contradicts the boundary condition. So w is a constant.
So far, we proved the uniqueness and stability for Dirichlet problem by maximum prin-
ciple, showed the relative uniqueness for Neumann problem via Hopfs maximum principle.
For the Dirichlet problem of Laplace equation on some special domains (such as a ball, and
half space), we constructed Greens functions and thus gave the existence for continuous
boundary data. However, the solvability of Dirichlet and Neumann problems on general
domains are not clear yer. These, however, will require the concept of weak solutions, for
which the functional analysis will play central role. This topic will be presented later.
3.5. EXAMPLES 57
3.5 Examples
In the rst two examples we construct the Greens functions for the disk and upper half
plane in R
2
.
Example 3.5.1 Find the Greens function for the unit disk
D(0, 1) = {(x
1
, x
2
)|x
2
1
+ x
2
2
< 1} R
2
.
Solution: We use the same idea as for n 3. Let =
_
x
2
1
+ x
2
2
< 1, (x) be the
fundamental solution. We recall that
(x) =
1
2
log(|x|) = ().
We will also use the inversion to invert the singularity out of the disk. Note for x =
x
|x|
2
|x||y x| = |y x|, if |y| = 1.
The function (|x|(y x)) is harmonic in y if x, y D(0, 1) and
(|x|(y x)) = (x y), for y D(0, 1).
Therefore, we have the Greens function for the unit disk
G(x, y) = (x y) (|x|(y x)), x = y D(0, 1).
We now solve the Laplace equation on D(0, 1) with boundary data g(x). For this purpose,
for |y| = 1, we compute
G(y x)
=
1
2
1 |x|
2
|x y|
2
, |y| = 1.
Therefore, we arrived at the Poissons formula
u(x) =
1 |x|
2
2
_
D(0,1)
g(y)
|x y|
2
d
y
. (3.56)
We could now use polar coordinate to have another form of the above equation. Let
x
1
= cos(), x
2
= sin(). and y
1
= cos(), y
2
= sin(), we thus have
u(, ) =
1
2
_
2
0
(1
2
)g()
1 2 cos( ) +
2
d. (3.57)
If the unit disk is replaced by a general disk D(0, r), (3.56) and (3.57) are replaced by
u(x) =
r
2
|x|
2
2r
_
D(0,r)
g(y)
|x y|
2
d
y
. (3.58)
u(, ) =
1
2
_
2
0
(r
2
2
)g()
r
2
2r cos( ) +
2
d. (3.59)
Example 3.5.2 Find the Greens function for the upper half plane
R
2
+
= {(x
1
, x
2
)|x
2
> 0}.
Solution. Similar to subsection 3.3.2, we choose
G(x, y) = (y x) (y x)
where
x = (x
1
, x
2
), for x = (x
1
, x
2
) R
2
+
.
Therefore
G(x, y) =
1
2
log
_
_
(y
1
x
1
)
2
+ (y
2
x
2
)
2
_
(y
1
x
1
)
2
+ (y
2
+ x
2
)
2
_
. (3.60)
The corresponding Poissons formula for Laplace equation with boundary data g(x) on
R
2
+
is
u(x) =
1
x
2
g(y
1
)
_
(y
1
x
1
)
2
+ x
2
2
dy
1
(3.61)
Example 3.5.3 For a > 0, D(0, a) is the disk centered at the origin on R
2
. Solve the
following boundary value problem
_
_
u = 0, in D(0, a),
u(a, ) = g() =
_
1, 0 < < ,
0, < < 2.
Solution: We could use (3.59) to solve this problem:
u(, ) =
1
2
_
2
0
(a
2
2
)g()
a
2
2a cos( ) +
2
d
=
a
2
2
2
_

0
1
(a
2
+
2
) 2a cos( )
d.
Set c = a
2
+
2
, d = 2a and = , we reduce the above equation into
u(, ) =
a
2
2
2
_

1
c + d cos()
d.
Except for the singular points = , we have that
F() =
2
c
2
d
2
arctan
_
c
2
d
2
tan(
2
)
c + d
_
,
3.5. EXAMPLES 59
is the anti-derivative of
1
c + d cos()
.
We note that (0, ) and [0, 2], therefore, if (0, ), (, ), and so
u(, ) = lim
0
_
a
2
2
2
2
a
2
2
arctan
_
c
2
d
2
tan(
2
)
c + d
__
=
=
=
1
arctan(
a +
a
cot(
2
)) +
1
arctan(
a +
a
tan(
2
))
However, if (, 2), then (2, 0) which contains = . We have to split the
integral at = . Therefore, we have
u(, ) = lim
0
_
1
arctan
_
a +
a
tan(

2
)
__
=
=
+ lim
0
_
1
arctan
_
a +
a
tan(

2
)
__
=
=+
=
1
2
+
1
arctan(
a +
a
tan(
2
))
+
1
arctan(
a +
a
cot(
2
)) +
1
2
= 1 +
1
arctan(
a +
a
cot(
2
)) +
1
arctan(
a +
a
tan(
2
)).
One easily verify the boundary condition by taking limit a in the above two cases.
The above example shows that using Poissons formula could lead to tedious calcula-
tions. We show in the following couple examples some specic tricks in two and three
dimensions.
Example 3.5.4 Solve the following Dirichlet problem
_
u = 0, in D(0, 1)
u(, ) = Asin
2
+ B cos
2
, on = 1,
where x = (x
1
, x
2
) = ( cos , sin ) and A and B are constants.
Solution. It is easy to check that
n
sin(n) and
n
cos(n) are harmonic functions on R
2
,
and
Asin
2
+ B cos
2
=
A + B
2
+
B A
2
cos(2),
therefore, we nd the solution
u(, ) =
A + B
2
+
B A
2

2
cos(2).
Example 3.5.5 Find a bounded solution to the following Dirichlet problem outside a unit
ball in R
3
:
_
u = 0, r > 1,
u|
r=1
=
2
5+4x
2
,
where r = |x|.
Solution. We know that the function
u(x) =
1
|x x
0
|
is a harmonic function out of the unit ball if x
0
B(0, 1). We try to see if we could choose
a x
0
so that the above function satises the boundary condition.
Since
2
5+4x
2
= (
5
4
+ x
2
)
1
, we need to nd x
0
such that
5
4
+ x
2
= |x x
0
|
2
= 1 2x x
0
+ x
2
0
which is equivalent to
x
01
= x
03
= 0, x
02
=
1
2
.
And such a point is inside the unit ball. Therefore,
u(x) =
1
_
x
2
1
+ (x
2
+
1
2
)
2
+ x
2
3
.
Example 3.5.6 Let be the triangle on R
2
with vertices (1, 0), (1, 0) and (0,
3). Solve
the following Dirichlet problem
_
u = 2, in
u = 0, on .
Solution. We rst observe the equations for the sides of the triangle are
y = 0, y +
3x
3 = 0, y
3x
3 = 0.
We thus guess that the solution has the following form
u(x, y) = cy(y +
3x
3)(y
3x
3)
with c the constant to be determined. Clearly, the boundary condition is fullled. A direct
calculation gives
u = 4
3c = 2,
and so c =
3
6
and the solution is
u(x, y) =
3
6
y(y +
3x
3)(y
3x
3).
3.5. EXAMPLES 61
The following example is the famous Hadamards three circles theorem.
Example 3.5.7 Let D be a annular region on R
2
centered at the origin. The outer circle
has radius R
1
and the inner one has radius R
1
, u(x, y) is a sub-harmonic function on D.
Set
M(r) = max
x
2
+y
2
=r
2
u(x, y), R
1
< r
1
< r < r
2
< R
2
,
then
M(r)
M(r
1
) log(
r
2
r
) + M(r
2
) log(
r
r
1
)
log(
r
2
r
1
)
.
Solution. For r = 0, we dene
(r) = a + b log r
where a and b are chosen such that
(r
1
) = M(r
1
), (r
2
) = M(r
2
).
Therefore, we nd
(r) =
M(r
1
) log(
r
2
r
) + M(r
2
) log(
r
r
1
)
log(
r
2
r
1
)
.
Consider now
v(x, y) = u(x, y) (
_
x
2
+ y
2
),
which satises
_
v 0, if r
1
< r < r
2
v 0, if r = r
1
or r = r
2
.
By the maximum principle, we know that
v 0, if r
1
< r < r
2
.
Therefore,
u(x, y) (r), if r
1
< r < r
2
.
This implies that
M(r) (r), if r
1
< r < r
2
.
3.6 Problems
Problem 1. Show that the Laplace operator takes the following form under cylindrical
coordination (r, , z):
u =
1
r
r
(r
u
r
) +
1
r
2
2
u
2
+

2
u
z
2
.
Problem 2. Show that the Laplace operator takes the following form under spherical
coordinate (r, , ).
u =
1
r
2
r
(r
2
u
r
) +
1
r
2
sin
2
(sin
u
) +

2
u
2
_
.
Problem 3. Prove the following functions are harmonic.
(a) x
3
3xy
2
, and 3x
2
y y
3
.
(b) sh(ny)sin(nx), sh(ny)cos(nx), ch(ny)sin(nx), and ch(ny)cos(nx).
(c) sh(x)(ch(x) + cos(y))
1
and sin(y)(ch(x) + cos(y))
1
.
Problem 4. Prove the following functions are harmonic in the polar coordinate.
(a) ln(r), and .
(b) r
n
cos(n) and r
n
sin(n).
(c) r ln(r) cos() r sin() and r ln(r) sin() + r cos().
Problem 5. Find the Greens function for the rst quadrant of R
2
, namely the domain
= {(x, y) R
2
|x > 0, y > 0}.
Problem 6. Find the Greens function for the upper half ball B
+
(0, r) in R
3
.
Problem 7. Find the Greens function for the rst octant in R
3
, namely the domain
= {(x, y, z) R
3
|x > 0, y > 0, z > 0}.
Problem 8. Find the Greens function for the unit square in R
2
, namely the domain
= {(x, y) R
2
|1 > x > 0, 1 > y > 0}.
Problem 9. Let D(0, r) is the disk on R
2
with boundary C. For each of the following
boundary conditions, nd the function u so that it is harmonic on D(0, r).
3.6. PROBLEMS 63
(a) u|
C
= Acos().
(b) u|C = A + B sin().
Problem 10. Solve the following Dirichlet problem
_
u
xx
+ u
yy
+ u
zz
= 0, x
2
+ y
2
+ z
2
< 1,
u(r, , )|
r=1
= 3 cos(2) + 1,
where (r, , ) is the spherical coordinate.
Problem 11. Let B be a unit ball in R
n
(n 2), and u is the smooth solution of the
following problem
_
u = f in B
u = g, on B.
Prove that there exists a constant C, depending only on n, such that
max
B
|u| C(max
B
|g| + max
B
|f|).
Problem 12. Assume u is harmonic, prove the following statements
(a) If : R R is a smooth convex function, then v = (u) is sub-harmonic.
(b) Prove v = |Du|
2
is sub-harmonic.
Problem 13. Use Poissons formula for the ball to prove
r
n2
r |x|
(r + |x|)
n1
u(0) u(x) r
n2
r + |x|
(r |x|)
n1
u(0),
where u(x) is harmonic for x B(0, r) R
n
with n 3. This is an explicit form of
Harnacks inequality.
Problem 14. Let u be the solution of
_
u = 0, in R
n
+
u = g on R
n
+
given by the Poissons formula for the half-space. Assume g is bounded and g(x) = |x| for
x R
n
+
, |x| 1. Show Du is not bounded near x = 0.
Problem 15. Let
+
R
n
+
and T =
+
R
n
+
is a non-empty open set. Assume
u C(
+
) is harmonic in
+
, with u = 0 on T. Set
v(x) =
_
u(x
1
, , x
n1
, x
n
), x
n
> 0
u(x
1
, , x
n1
, x
n
), x
n
< 0.
Prove that v(x) is harmonic on
+
T
where
is the reection of
+
about x
n
= 0.
This result is called Schwarz reection theorem.
Problem 16. Using example to show that the maximum principle is not valid for
u
xx
+ u
yy
+ cu = 0, c > 0.
Problem 17. Find the Greens function for the following wedge:
= {(, , z) : > 0, 0 < <

4
, z R}.
Problem 18. Find the Greens function for a domain between two parallel planes:
= {(x, y, z) : 0 < z < 1}.
Chapter 4
Heat equation
Let be an open set in R
n
, we will study the Heat equation
u
t
u = 0, x , t > 0 (4.1)
and the non-homogeneous Heat equation
u
t
u = f(x, t), x , t > 0 (4.2)
with appropriate initial boundary conditions. The constant > 0 models the heat diu-
sion.
A guiding principle is that any assertion about harmonic functions yields a analogous
statement about solutions of the heat equation. However, we will follow a slight dierent
approach to show dierent techniques.
4.1 Physical derivation
Assume was occupied by the homogeneous media without heat source inside. Let u(x, t)
be the temperature at x and time t, J(x, t) is the heat ux. For each sub-region G with
smooth boundary G and unit outer normal ,
_
G
J dS
is the total heat ow out of G per unit time. According to Fouriers law,
J = k
x
u
with k > 0 the constant of heat conductivity. Therefore, we update the last equation by
_
G
k
u
dS.
65
66 CHAPTER 4. HEAT EQUATION
On the other hand, the rst principle of thermodynamics says that the absolute tempera-
ture u(x, t) is proportional to the heat, therefore, we know
_
G
C
V
u
t
dx =
_
G
k
u
dS,
where C
V
> 0 is the specic heat constant. Now, by divergence theorem, one has
_
G
u
t
u dx = 0
with =
k
C
V
. This leads to heat equation. The f in non-homogeneous heat equation
models the interior heat source.
4.2 Fundamental solution
4.2.1 Derivation
We observe that the Heat Equation is invariant under the transformation:
x x, t t, for R.
Therefore, the ratio
|x|
2
t
is an important variable for heat equation, and we shall look for a
solution of the form
u(x, t) = v(
r
2
t
), r = |x|, t > 0. (4.3)
Such class of solutions are called self-similar solution.
A quicker approach is to nd a solution u of the special structure
u(x, t) = t
v(
|x|
t
).
Let y =
x
t
, and we substitute it into (4.1) to obtain
v +
1
2
y Dv +v = 0 (4.4)
which reduces into
w +
1
2
rw
+w
+
n 1
r
w
= 0
for w(r) = v(|y|) where r = |y|. Now, =
n
2
is the magic number so that the equation
becomes
(r
n1
w
+
1
2
(r
n
w)
= 0.
Hence,
r
n1
w
+
1
2
r
n
w = 0
4.3. FOURIER TRANSFORM AND INTIAL VALUE PROBLEM 67
if we require w and w
tends to 0 if r . Therefore,
w = be
r
2
4
for some constant b. Therefore,
u(x, t) = bt
n
2
e
|x|
2
4t
solves (4.1).
This motivates the following deinition:
Denition 4.2.1 The function
E(x, t) =
_
(4t)
n
2
e
|x|
2
4t
, (x R
n
, t > 0)
0 (x R
n
, t > 0),
is called the fundamental solution of the heat equation.
We note that E(x, t) is smooth except the singular point (0, 0). In the next section,
the fundamental solution is naturally derived when we apply Fourier transform to solve
the initial value problem.
4.3 Fourier transform and intial value problem
4.3.1 Fourier transform
Let function f(x) and g(x) dened on x R
n
are continuously dierentiable and absolutely
integrable, then the Fourier transform of f(x) is dened as
F(f)() =

f() = (2)
n
2
_
R
n
x
f(x)e
ix
dx (4.5)
and the inverse Fourier transform of g() is
F
1
(g)(x) = g(x) = (2)
n
2
_
R
n
f(x)e
ix
d. (4.6)
We remark that the Fourier transform and its inverse are naturally generalized into the
square integrable function space L
2
(R
n
). We recall that for a function u(x) L
2
(R
n
),
u
L
2
(R
n
)
=
__
R
n
u
2
(x) dx
_1
2
< .
We also recall that
f g =
_
R
n
f(y)g(x y) dy.
We list some properties of Fourier transform in the following theorem.
Theorem 4.3.1 (Properties of Fourier Transform). Assume f, g L
2
(R
n
).
(i) f
L
2
(R
n
)
=
f
L
2
(R
n
)
=
f
L
2
(R
n
)
;
(ii)
_
R
n
f g dx =
_
R
n
g d,
(iii) For each multi-index such that D
f L
2
(R
n
),

D
f = (i)

f.
(iv)
(f g) = (2)
n
2

f g.
(v) f =

f.
Example 4.3.2 For x R, nd the Fourier transform for f(x) = e
a|x|
.
Solution.
f() =
1
2
_
R
e
a|x|
e
ix
dx
=
1
2
_
R
e
a|x|
(cos(x) isin(x)) dx
= 2
1
2
_

0
e
ax
cos(x) dx
=
1
2
2a
2
+a
2
.
Example 4.3.3 For R
n
and t > 0, nd the inverse Fourier transform for
f() = (2)
n
2
e
||
2
t
.
Solution.
f(x) = (2)
n
_
R
n
e
||
2
t
e
ix
d
=
n
k=1
_
1
2
_

e
t
2
k
+ix
k
k
d
k
_
=
n
k=1
_
1
_

0
e
t
2
k
cos(x
k
k
) d
k
_
=
n
k=1
I(x
k
).
By Eulers formula
_

0
e
y
2
dy =
2
,
we know that
I(0) =
1
2
t
.
Dierentiating I(x
k
) once, and using integration by parts, one has
I
+
x
k
2t
I = 0,
therefore,
I(x
k
) =
1
2
t
e
x
2
k
4t
.
Finally, we obtain
f(x) = F
1
((2)
n
2
e
||
2
t
) = (4t)
n
2
e
|x|
2
4t
= E(x, t).
4.3.2 Initial Value Problem
We now employ the Fourier transform to solve the following Cauchy problem for heat
equation
_
u
t
u = 0
u(x, 0) = (x).
(4.7)
We perform Fourier transform in x, and denote
u(, t) = (2)
n
2
_
R
n
u(x, t)e
ix
dx
() = (2)
n
2
_
R
n
(x)e
ix
dx,
we thus arrive at an initial value problem for the ODE about u(, t)
_
_
_
d u(, t)
dt
+ ||
2
u(, t) = 0
u(, 0) =

().
(4.8)
Clearly, one has
u(, t) =

()e
||
2
t
,
therefore, we perform inverse Fourier transform
u(x, t) = F
1
( u(, t))
= F
1
(
()e
||
2
t
)
= F
1
(
()) F
1
((2)
n
2
e
||
2
t
)
= E(x, t) (x)
=
_
R
n
E(x y, t)(y) dy.
(4.9)
This also gives a natural way to derive the fundamental solution E(x, t). We list some
properties here for later applications:
Theorem 4.3.4 (Properties of Fundamental solution) Let E(x, t) be the fundamental so-
lution of Heat equation.
(i) For x, y R
n
and t > 0, E(x y, t) > 0.
(ii) For x, y R
n
and t > 0, (
t
)E(x y, t) = 0.
(iii) For x R
n
and t > 0,
_
R
n
E(x y, t) dy = 1.
(iv) For any > 0 and x R
n
, lim
t0+
_
|yx|>
E(x y, t) dy = 0.
Proof. By the expression of E(x, t), the rst two properties are obvious. For (iii), using
substitution y = x +
4tz, we compute
_
R
n
E(x y, t) dy = (4t)
n
2
_
R
n
e
|xy|
2
4t
dy
= ()
n
2
_
R
n
e
|z|
2
dz
= 1.
Using the same substitution, we can prove (iv) as following
lim
t0+
_
|yx|>
E(x y, t) dy
= lim
t0+
()
n
2
_
|z|>

4t
e
|z|
2
dz = 0.
We remark that the formula (4.9) gives a formal solution to the intial value problem
(4.7), where we assumed that the existence of Fourier transform of initial function (x),
and we used the inverse transformation, this often requires (x) to be C
1
and absolutely
integrable. However, under much weaker conditions on (x), we are able to prove that
(4.9) does give the classical solution to (4.7).
Theorem 4.3.5 If (x) C(R
n
) and there exist constants M > 0 and A > 0 such that
|(x)| Me
A|x|
2
, x R
n
,
then
u(x, t) = E(x, t) (x)
is a C
solution of (4.7) on the region

= {(x, t)|x R
n
, 0 < t T}, for T <
1
4A
.
Proof. 1. We rst show the continuity of u(x, t). For any constants a > 0 and t
0
(0, T),
we dene the set
V = {(x, t)||x| a, t
0
t T}.
For (x, t) V , we see that
|u(x, t)| M
_
R
n
E(x y, t)e
A|y|
2
dy
cM
_
R
n
exp{A|y|
2
+

A|x y|
2
} dy,
where c = (4t
0
)
n
2
and

A =
1
4T
. We note that
A|y|
2
+

A|x y|
2
= (A +

A)|y
A
A +

A
x|
2
+
A

A
A +

A
|x|
2
.
Therefore, for A +

A < 0, i.e., T <
1
4A
, we have
|u(x, t)| Me
A

A
A+

A
|x|
2
_
R
n
exp{(A +

A)|y
A
A +

A
x|
2
} dy
cM(

A +

A
)
n
2
e
A

A
A+

A
|x|
2
.
Therefore, we know that the integral in (4.9) converges uniformly and absolutely on V ,
and so u(x, t) is continuous on V . For a > 0 and t
0
> 0 arbitrary, u(x, t) is continuous on
.
2. We now show that u(x, t) C
() and it satises the Heat equation. This is

because E(x y, t) is innitely dierentiable, with uniformly bounded derivatives, on V .
For each multi-index = (
0
,
1
, ,
n
),
D
u(x, t) =
_
R
n
(y)D
E(x y, t) dy.
One can estimate this integral in a similar manner as in step 1 on V to prove its absolute
and uniform convergence on V . Therefore, one proves that u(x, t) C
(). We now use

the property of E(x, t) to show that
(
t
)u(x, t) =
_
R
n
(y)(
t
)E(x y, t) dy = 0.
3. Finally, we prove that u(x, t) veries the initial condition, namely, we shall prove for
any x
0
R
n
,
lim
(x,t)(x
0
,0+)
u(x, t) = (x
0
).
Letting v
0
(x) = (x) (x
0
), using the property of E(x, t), one has
u(x, t) =
_
R
n
(x
0
)E(x y, t) dy +
_
R
n
v
0
(y)E(x y, t) dy
= (x
0
) +
_
R
n
v
0
(y)E(x y, t) dy.
Since v
0
(x) is continuous and v
0
(x
0
) = 0, for any xed > 0 there exists such that when
|x x
0
| < , |v
0
(x)| < . For |x x
0
| > , there exists B() > 0 large enough such that
|v
0
(x)| e
B()|xx
0
|
2
.
Therefore, for any x R
n
, one has
|
_
R
n
v
0
(y)E(x y, t) dy|
_
B
(x
0
, )|v
0
(y)|E(x y, t) dy +
_
R
n
\B(x
0
,)
|v
0
(y)|E(x y, t) dy
+

(4t)
n
2
_
R
n
exp{B()|y|
2
|x y|
2
4t
} dy
+

(1 4Bt)
n
2
e
B|x|
2
14Bt
,
where we have chosen t so small such that 1 4Bt > 0. Now, when t is sucient small,
we see that |u(x, t) (x
0
)| < 2. Therefore, we veried that
lim
(x,t)(x
0
,0+)
u(x, t) = (x
0
).
Remark 4.3.6 Some remarks are in order.
1. By the property of the fundamental solution E(x, t), it is easy to see that
u(x, t) =
_
R
n
E(x y, t)(y) dy sup
yR
n
(y).
Similarly, on nds the bound from below. Therefore, if (x) is bounded, then
inf
yR
n
(y) u(x, t) sup
yR
n
(y) (4.10)
2. From the theorem, it is obvious that the smaller A is, the larger existence time.
If (x) is bounded, A = 0, then the solution is global in time.
3. It also clear that even if (x) is replaced by measurable function, while other
conditons keep the same, u(x, t) given in (4.9) is still the C
solution to (4.7), and

veries the initial conditions at all the continuous points of .
4. If (x) 0 and 0, the u(x, t) > 0 for any x R
n
and t > 0. This is called the
innite propogation speed for disturbances. If the initial temperature is nonnegative
and positive somewhere, the temperature is positive everywhere at any later time.
4.3.3 Nonhomogeneous Problem
We now derive a general formula for the following initial-value problem
_
u
t
u = f(x, t) x R
n
, t > 0
u(x, 0) = 0.
(4.11)
For this purpose, we note that
u(x, t; s) =
_
R
n
E(x y, t s)f(y, s) dy
solves the following initial value poblem
_
u
t
(x, t; s)
x
u(x, t; s) = 0, x R
n
, t s,
u(x, s; s) = f(x, s).
(4.12)
Of course, u(x, t; s) does not solve (4.11), however, the Duhamels Principle allows us to
build a solution of (4.11) from the solutions of (4.12), by integrating with respect to s.
More precisely, consider
u(x, t) =
_
t
0
u(x, t; s) ds
=
_
t
0
_
R
n
E(x y, t s)f(s, y) dyds, x R
n
, t > 0.
(4.13)
The following Theorem conrms that (4.13) gives the solution to (4.11).
Theorem 4.3.7 Assume f C
2
1
((0, ) R
n
) has compact support. Deinfe u by (4.13),
then
(i) u C
2
1
((0, ) R
n
),
(ii) u
t
u = f(x, t),
(iii) For each x
0
R
n
, lim
(x,t)(x
0
,0+)
u(x, t) = 0.
Proof.
1. We rst change variables to write
u(x, t) =
_
t
0
_
R
n
E(y, s)f(x y, t s) dyds.
We note that E(y, s) is smooth near s = t > 0 and f has compact support, then
u
t
(x, t) =
_
t
0
_
R
n
E(y, s)f
t
(x y, t s) dyds
+
_
R
n
E(y, s)f(x y, 0) dy,
u
x
i
(x, t) =
_
t
0
_
R
n
E(y, s)f
x
i
(x y, t s) dyds,
u
x
i
x
j
(x, t) =
_
t
0
_
R
n
E(y, s)f
x
i
x
j
(x y, t s) dyds.
Thus, u(x, t) C
2
1
((0, ) R
n
).
2. We not compute
u
t
u
=
_
t
0
_
R
n
E(y, s)[(
t
x
)f(x y, t s)] dyds +
_
R
n
E(y, s)f(x y, 0) dy
= I
+J
+K,
where,
I
=
_
t
_
R
n
E(y, s)[(
s
y
)f(x y, t s)] dyds,
J
=
_

0
_
R
n
E(y, s)[(
s
y
)f(x y, t s)] dyds,
K =
_
R
n
E(y, s)f(x y, 0) dy.
Now,
|J
| (f
t
L
+D
2
f
L
)
_

0
_
R
n
E(y, s) dyds C.
Integrating by parts, we nd
I
=
_
t
_
R
n
[(
s
y
)E(y, s)]f(x y, t s)] dyds
+
_
R
n
E(y, )f(x y, t ) dy
_
R
n
E(y, s)f(x y, 0) dy.
=
_
R
n
E(y, )f(x y, t ) dy K,
Therefore, we conclude that
u
t
u = lim
0
_
R
n
E(y, )f(x y, t ) dy
= f(x, t), x R
n
, t > 0.
Finally, we note that
u(x, t)
L
tf
L
,
which tends to zero as t 0+. This concludes the proof.
4.4. MAXIMUM PRINCIPLE AND APPLICATIONS 75
By linear superposition principle, we know that
u(x, t) =
_
R
n
E(x y, t)(y) dy +
_
t
0
_
R
n
E(x y, t s)f(s, y) dyds (4.14)
solves
_
u
t
u = f(x, t) x R
n
, t > 0
u(x, 0) = (x),
(4.15)
under conditions on and f as above.
4.4 Maximum Principle and applications
4.4.1 Maximum Principle
Assume R
n
is open, bounded set. We rst introduce the following concept.
Denition 4.4.1 Fix a time T > 0, the parabolic cylinder
T
= (0, T].
The parabolic boundary of
T
is
T
=

T
\
T
.
We remark that the parabolic interior of

T
contains the top {t = T}. The
parabolic boundary
T
comprises the bottom, the vertical sides [0, T], but not the
top.
We now state the maximum principle.
Theorem 4.4.2 (Maximum Principle) Assume u(x, t) C
2
1
(
T
) solves the heat equation
in
T
, then
max
T
u = max
T
u.
Proof. Assume that u(x, t) does not attains it maximum at
T
but at a point (x
, t
)
T
.
Namely, there exists m < M such that
max
T
u = u(x
, t
) = M > m = max
T
u.
Dene
v(x, t) = u(x, t) +
M m
2nd
2
|x x
|
2
,
where d = 2 max dist{x
, }. We see that v(x
, t
) = M and
v(x, t)|
T
< m +
M m
2n
< M.
Therefore, v(x, t) attains its maximum at some (x
1
, t
1
)
T
. We observe that at this
point (x
1
, t
1
),
v(x
1
, t
1
) 0, v
t
(x
1
, t
1
) 0,
and so
v
t
v 0, at (x
1
, t
1
).
But on the other hand, direct computation gives
v
t
v =
M m
d
2
< 0,
a contradiction. Therefore,
max
T
u = max
T
u.
If one replaces u with u, one actually obtains that
max
T
|u| = max
T
|u|.
The following comparison principle is a direct consequence of the above maximum
principle.
Theorem 4.4.3 (Comparison Principle) Let u
1
, u
2
C
2
1
(
T
) be two solutions of the heat
equation. If
u
1
u
2
, on
T
,
then
u
1
u
2
, on

T
.
In order to restore the strong maximum principle, we introduce the parabolic mean-
value formula. we introduce the following concept.
Denition 4.4.4 For xed x R
n
, t > 0 and r > 0, we dene
V (x, t; r) =
_
(y, s) R
n+1
+
|s t, E(x y), t s)
1
r
n
_
.
We remark that the boundary of V (x, t; r) is a level set of E(x y, t s), and the point
(x, t) is at the center of the top. With the help of this notion, we have
Theorem 4.4.5 (Mean-value property for heat equation). Let u C
2
1
(
T
) solve the heat
equation. Then
u(x, t) =
1
4r
2
__
V (x,t;s)
u(y, s)
|x y|
2
(t s)
2
dyds, (4.16)
for each V (x, t; r)
T
.
With the help of this theorem, a similar argument as for Laplace equation, one can
prove
4.4. MAXIMUM PRINCIPLE AND APPLICATIONS 77
Theorem 4.4.6 (Strong maximum principle). Let u C
2
1
(
T
) solve the heat equation. If
is connected and there exists a point (x
0
, t
0
)
T
such that
u(x
0
, t
0
) = max
T
u,
then u is constant in

t
0
.
4.4.2 Uniqueness and Stability
The rst application is to the following initial boundary value problem:
_
u
t
u = f(x, t), (x, t)
T
,
u|
T
= (x, t).
(4.17)
Theorem 4.4.7 The solution of (4.17) is unique and continuously depends on the initial
boundary data.
Proof. If u
i
(i = 1, 2) are solutions of
_
u
i
t
u
i
= f(x, t), (x, t)
T
,
u
i
|
T
=
i
(x, t),
then w = u
1
u
2
solves
_
w
t
w = f(x, t), (x, t)
T
,
w|
T
=
1
2
.
By maximum principle, one has
max
T
|w| = max
T
|
1
2
|.
This shows the continuous dependence. In particular, if
1
=
2
, w 0, one nds the
uniqueness.
We now turn to the initial value problem. For this purpose we rst establish the
maximum principle for the Cauchy problem (4.7). From Theorem 4.3.5, we know that if
(x) C(R
n
) satises the growth condition
|(x)| Me
A|x|
2
,
then the formula (4.9) u(x, t) = E(x, t) (x) solves (4.7) for any x R
n
, 0 t T
(T <
1
4T
). Furthermore, from the proof of Theorem 4.3.5, we know that there exist M
1
> 0
and A
1
0, such that
|u(x, t)| M
1
e
A
1
|x|
2
, (x, t) R
n
[0, T].
This motivates the following maximum principle.
Theorem 4.4.8 (Maximum principle of Cauchy problem). Let u C
2
1
(R
n
(0, T])
C(R
n
[0, T]) solves (4.7), and satises the growth estimate
|u(x, t)| Me
A|x|
2
, (x, t) R
n
[0, T], (4.18)
for constants M > 0, A 0. Then
sup
R
n
[0,T]
u = sup
R
n
.
Proof. We rst assume
4AT < 1 (4.19)
and so there is > 0 that
4A(T + ) < 1.
Fix y R
n
, > 0, we dene
v(x, t) = u(x, t)

(T + t)
n
2
e
|xy|
2
4(T+t)
. (4.20)
Clearly, it holds
v
t
v = 0, in R
n
(0, T].
Fix r > 0, dene
T
= B(y, r) (0, T]. By maximum principle,
max
T
v = max
T
v.
We not check the value of v at
T
. First of all, if x R
n
,
v(x, 0) = u(x, 0)

(T + )
n
2
e
|xy|
2
4(T+)
u(x, 0) = (x).
(4.21)
On the sides where |x y| = r, t [0, T], we have
v(x, 0) = u(x, 0)

(T + t)
n
2
e
r
2
4(T+t)
Me
A(|y|+r)
2

(T + )
n
2
e
r
2
4(T+)
.
(4.22)
We now know
1
4(T+)
= A + for some small > 0. Thus we have
v(x, t) Me
A(|y|+r)
2
(4(A + )
n
2
e
(A+)r
2
sup
R
n
,
for r suciently large. We thus conclude that
v(x, t) sup
R
n
,
4.5. EXAMPLES 79
for all x R
n
and t T as long as 4AT < 1. Let 0, we showed that
u(x, t) sup
R
n
, if 4AT < 1.
For general case where (4.19) fails, we can repeat the above procedure on [0,
1
8A
], [
1
8A
,
2
8A
],
. Thus the proof of the theorem is complete.
Using this maximum principle, one easily prove the following uniqueness result.
Theorem 4.4.9 (Uniqueness for Cauchy problem). Let (x) C(R
n
), f C(R
n
[0, T]).
Then there exists at most one solution u C
2
1
(R
n
(0, T]) C(R
n
[0, T]) of (4.15) ,
satisfying the growth estimate
|u(x, t)| Me
A|x|
2
, (x, t) R
n
[0, T], (4.23)
for constants M > 0, A 0.
Now, a natural question arises: Is there any other solutions to (4.15) if we dont require
the growth restriction (4.23). The answer is YES. This will be explained in the examples.
4.5 Examples
The rst example of A. N. Tychonov is to answer the uniqueness problem for the Cauchy
problem without growth restriction (4.23).
Example 4.5.1 There are innitely many solutions to the initial value problem
_
u
t
u
xx
= 0, x R, t > 0,
u(x, 0) = 0,
without the growth restriction (4.23).
Solution. For some g(t) C
(R), we dene
u(x, t) =
k=0
d
k
g(t)
dt
k
x
2k
(2k)!
, x R, t R. (4.24)
If the series converge in a nice way, we have
u
t
=
k=0
d
k+1
g(t)
dt
k+1
x
2k
2k!
,
and
u
xx
=
k=1
d
k
g(t)
dt
k
x
2k2
(2k 2)!
=
k=0
d
k+1
g(t)
dt
k+1
x
2k
2k!
.
Therefore, we see u(x, t) solves the heat equation. Now, we choose
g(t) =
_
exp{t
}, > 1, t > 0
0, t 0.
It is clear that g(t) is analytic except for t = 0. It remains to verify that u(x, t) attains
the initial data when t 0+. For this purpose, we need to compute the derivatives of
g(t). Due to complex analysis, the Cauchy integral formula, we have
d
k
g(t)
dt
k
=
k!
2i
_
e
z
|z t|
k+1
dz,
where the path is chosen as the circle: |z t| = t for (0, 1). For Re(z) > 0, z
is
dened as its principal value. Now, for some R, the point on is described as
z = t + te
i
= t(1 + e
i
), Re(z
) = t
Re(1 +e
i
)
.
For small , we have
Re(1 +e
i
)
>
1
2
, and so
Re(z
) <
1
2
t
, |
d
k
g(t)
dt
k
|
k!
(t)
k
e
1
2t
.
Note that
k!
(2k)!
<
1
k!
, we have
|u(x, t)|
k=0
|x|
2k
k!(t)
k
e
1
2t
= exp{
1
t
(
|x|
2

1
2
t
1
)}.
Now, it is clear that on each interval [x
1
, x
2
], when t 0+, u(x, t) 0 uniformly. This
shows that (4.24) determines a solution (called Tychonovs solution) for each > 1.
The second example is due to E Rothe showing that the backward heat equation is
ill-posed.
Example 4.5.2 Consider the initial value problem
_
u
t
u
xx
= 0
u(x, 0) = (x), x R.
If (x) = 0, we see u(x, t) = 0 is a solution. Now, if
(x) = sin(
x
), > 0,
4.5. EXAMPLES 81
then
u(x, t) = e
2
sin(
x
)
is a solution. We see when is suciently small, then (x) is arbitrarily close to 0. This
is true when t > 0; but for t < 0, it is not. So the solution is not stable about the initial
data if t < 0.
In what follows, we show some tricks in solving initial boundary value problem of heat
equation in one space dimension.
Example 4.5.3 Consider the heat conduction problem on a nite bar
_
_
u
t
u
xx
= 0, t > 0, x (0, l),
u(x, 0) = (x), x (0, l),
u(0, t) = u(l, t) = 0, t 0,
(4.25)
where (x) C[0, l] such that (0) = (l) = 0.
Solution. If u(x, t) is a solution, since u(0, t) = 0, we could extend u(x, t) as an odd
function in x into (l, 0). Then we further extend u(x, t) as a periodic function in x with
period 2l to whole real line. This will transfer the problem into a Cauchy problem
_
u
t
u
xx
= 0, t > 0, x R,
u(x, 0) = (x), x R,
where
(x) = (x), x (0, l)
(x) = (x), x (l, 0)
(x + 2l) = (x), x R.
Since (x) C[0, l], (x) satises all conditions for existence and uniqueness, and therefore
u(x, t) =
_

(y)E(x y, t) dy (4.26)
We know that (x) is odd in x, so
u(0, t) =
_

(y)E(y, t) dy = 0.
Also, (l x) is odd in x, so
u(l, t) =
_

(l y)E(y, t) dy = 0.
Therefore, u(x, t) determined by the formula (4.26) is a solution of (4.25). We could rewrite
(4.26) as
u(x, t) =
_
l
0
(y)G(x, y, t) dy
where
G(x, y, t) =
1
2
n=
_
exp{
x y 2nl)
2
4t
} exp{
x +y 2nl)
2
4t
}
_
.
Example 4.5.4 We solve the above problem (4.25) again by the method of separation of
variables, called Fourier method.
Solution. Assuming
u(x, t) = X(x)T(t),
one gets
T
T
=
X
X
=
for the constant ratio as the rst term depends only on t while the second depends only
on x. Therefore, we obtained two equations
T
+ T = 0, X
+ X = 0. (4.27)
For X, it satises the following two points boundary value problem
_
X
+ X = 0, 0 < x < l,
X(0) = X(l) = 0.
(4.28)
This is a typical Sturm-Liouville type eigenvalue problem. We now derive the general
method for such type of problems.
First of all, if < 0, we have the general solution of X as
X(x) = Ae
x
+Be
x
,
which gives zero solution by the boundary condition.
If = 0, X
= 0, which with the boundary condition, give zero solution.

Now, if = k
2
> 0 for k > 0, we have the general solution of X is
X(x) = Acos(kx) +Bsin(kx),
where A = 0 as X(0) = 0, and Bsin(kl) = 0 as X(l) = 0. If B = 0, we get zero solution
again. If B = 0, we have
k =
n
l
, or,
n
= (
n
l
)
2
, n = 1, 2, (4.29)
4.5. EXAMPLES 83
We remark that n < 0 does not give new values of . We thus obtained the non-zero
solutions
X
n
(x) = Bsin(
n
l
x), n = 1, 2, (4.30)
We call
n
the eigenvalue and X
n
the corresponding eigenfunction. We now substitute the
n
into the equation of T to obtain
T
n
= a
n
e
n
t
. (4.31)
We thus obtained the formal solution of the heat equation with the boundary data
u(x, t) =
n=1
a
n
e
(
n
l
)
2
t
sin(
n
l
x). (4.32)
The constant B is absorbed by a
n
, which will be determined by the initial condition,
(x) =
n=1
a
n
sin(
n
l
x),
and therefore,
a
n
=
2
l
_
l
0
(x) sin(
n
l
x) dx.
The equation (4.32) gives a formal solution, we need to prove the convergent properties
for this series. As (x) C([0, l]), there exists a constant K > 0 such that |a
n
| K for
all n. Now, for any t
1
> t
0
> 0, on the domain [0, l] [t
0
, t
1
], we have
|a
n
e
(
n
l
)
2
t
sin(
n
l
x)| Ke
(
n
l
)
2
t
.
Therefore, the series (4.32) converges uniformly and absolutely. Therefore, u(x, t) is con-
tinuous in this domain and thus is continuous for x [0, l], t > 0. Furthermore, it veries
the boundary conditions. Similarly, one show the continuous dierentiability in t and in
x. Therefore, (4.32) denes the solution to problem (4.25).
Example 4.5.5 Assume there is a innitely long, heat conductive, thin rod (x R).
Assume the rod is made of two dierent materials on the part of x < 0 and on the part
of x > 0 respectively. At the transition point x = 0, the heat ux must be continuous on
both side. We denote the left temperature by u(x, t) and the right temperature by v(x, t).
Therefore, we need to solve
_
_
u
t
u
xx
= 0, x < 0, t > 0,
v
t
v
xx
= 0, x > 0, t > 0,
u(x, 0) = (x), x 0,
v(x, 0) = (x), x 0,
u(0, t) = v(0, t), t 0,
u
x
(0, t) = v
x
(0, t), t 0,
(4.33)
where, (0) = (0), and are head conduction constants, and and are the heat
transfer constants.
Solution. We assume this problem has solution (u(x, t), v(x, t)). We dene
w(x, t) = au(x, t) +bv(
_
x, t), x 0 (4.34)
where a and b are constants to be determined. We easily verify that w satises the equation
w
t
w
xx
= 0, x > 0, t > 0.
We now introduce a function
u
(x, t) =
_
u(x, t), x < 0,
w(x, t), x > 0.
It is clear that u
satises u
t
u
xx
= 0 on R\ {0}. Now, we require that u
(x, t) satises
the following compatibility condition at x = 0,
_
u(0, t) = w(0, t) = au(0, t) +bv(0, t),
u
x
(0, t) = w
x
(0, t) = au
x
(0, t) +b
_
v
x
(0, t).
(4.35)
By the conditions of (4.33), the above conditions implies that
_
u(0, t) = au(0, t) +bu(0, t),
u
x
(0, t) = au
x
(0, t) +b
_
u
x
(0, t),
(4.36)
which gives
a +b = 1, a +b
_
= .
Thus,
a = 1 b, b =
2
+
_
.
We now substitute a and b into (4.34) to recover the initial data for u
on x 0:
u
(x, 0) = a(x) +b(

_
x), x > 0.
We now dene
(x) =
_
(x), x < 0,
a(x) +b(
_
x), x 0.
4.6. PROBLEMS 85
We thus obtained a Cauchy problem for u
_
u
t
u
xx
= 0, x R, t > 0
u
(x, 0) =
(x), x R.
(4.37)
Now, set y =
x
, and then the solution of (4.37) is given by

u
(x, t) =
1
2t
_

z)e
(
x
z)
2
4t
dz, x R.
4.6 Problems
Problem 1. Use Fourier transform method to give a solution formula to the following
Cauchy problem
_
u
t
u
xx
+xu = 0, x R, t > 0,
u(x, 0) = (x).
Problem 2. Use Fourier transform method to solve
_
_
_
u
xx
+u
yy
= 0, x R, y > 0,
u(x, 0) = (x), lim
x
2
+y
2
u = 0.
(4.38)
Problem 3. Assume a static temperature distribution u(x, y) on upper half plane satises
the constraints
u(x, 0) =
_
1, if |x| a,
0, if |x| > a.
Prove that
u(x, y) =
1
(arctan(
a +x
y
) + arctan(
a x
y
)).
Problem 4. Solve the Cauchy problem of 1-d heat equation u
t
a
2
u
xx
= 0, (x R, t > 0)
with the following initial conditions:
(a) u(x, 0) = sin(x);
(b) u(x, 0) = x
2
+ 1.
Problem 5. Use extension method to solve the following problem
_
_
u
t
a
2
u
xx
= 0, x > 0, t > 0,
u(x, 0) = (x), x > 0
u(0, t) = 0, t 0; (0) = 0.
Problem 6. Use extension method to solve the following problem
_
_
u
t
u
xx
= 0, x > 0, t > 0,
u(x, 0) = (x), x > 0
u
x
(0, t) = 0, t > 0; (0) =
(0) = 0.
Problem 7. For constant d, solve the following problem
_
_
u
t
u
xx
= 0, x > 0, t > 0,
u(x, 0) = (x), x > 0
u
x
(0, t) +du(0, t) = 0, t > 0.
Problem 8. For constant > 0, use separation of variables method to solve
_
_
u
t
u
xx
= 0, 0 < x < l, t > 0,
u(x, 0) = (x), 0 x l
u(0, t) = 0, u
x
(l, t) + u(l, t) = 0, t 0.
Problem 9. Prove the solution of
_
_
u
t
u
xx
= 0, x R, t > 0,
u(x, 0) = 1, x > 0,
u(x, 0) = 1, x < 0,
is
u(x, t) = h(
x
2
t
),
where, h is the error function
h(x) =
2
_
x
0
e
t
2
dt.
Problem 10. If u
1
(x, t) and u
2
(y, t) are respectively the solutions of the problems
_
u
1t
u
1xx
= 0, x R, t > 0,
u
1
(x, 0) =
1
(x),
4.6. PROBLEMS 87
and
_
u
2t
u
2yy
= 0, y R, t > 0,
u
2
(y, 0) =
2
(y).
Prove the function u(x, y, t) = u
1
u
2
is the solution of
_
u
t
(u
xx
+u
yy
) = 0, (x, y) R
2
, t > 0,
u(x, y, 0) =
1
(x)
2
(y).
Problem 11. Derive the solution formula for the following problem
_
_
u
t
(u
xx
+u
yy
) = 0, (x, y) R
2
, t > 0,
u(x, y, 0) =
n
i=1
i
(x)
i
(y).
Problem 12. Let v(x, t) be the solution of
_
v
t
a
2
v
xx
= 0, x > 0, t > 0,
v(x, 0) = 0, v(0, t) = 1.
Prove the Duhamel integral
u(x, t) =

t
_
t
0
v(x, t )g() d
solves
_
u
t
a
2
u
xx
= 0, x > 0, t > 0,
u(x, 0) = 0, u(0, t) = g(t).
Problem 13. Prove the weak maximum principle: Let be a bounded open set in R
n
,
and
T
is the parabolic cylinder. If u(x, t) C
2
1
(
T
) satises
u
t
u 0,
then
max
T
u = max
T
u
where
T
is the parabolic boundary. If is replaced by , and the max is repalced by
min, then the statement is also valid.
Problem 14. Let u(x, t) C
2
1
(
T
) be the solution of
_
u
t
u = f(x, t), (x, t)
T
,
u(x, t)|
T
= (x, t).
Dene
F = sup
Q
T
|f|, B = sup
T
||.
Prove that
max
T
u FT +B.
Problem 15. Assume u solves u
t
u = 0, prove the following statements
(a) If : R R is a smooth convex function, then v = (u) satises
v
t
v 0.
(b) Prove v = |Du|
2
+u
2
t
also satises the above inequality.
Problem 16. Dene the following parabolic dierential operator
Lu = u
t
a
2
u
xx
+b(x, t)u
x
+c(x, t)u.
Assume c(x, t) > 0, if u(x, t) C
2
1
(
T
) satises
Lu 0, in
T
,
then
max
T
u max
T
u
+
,
where u
+
(x, t) = max{u(x, t), 0)}.
Problem 17. Consider the same operator L as in Problem 16. Assume for some constant
c
0
> 0, c(x, t) > c
0
, and u(x, t) C
2
1
(
T
) satises
Lu 0, in
T
,
if max
T
u 0, then
max
T
u 0.
Chapter 5
Wave equation
In this chapter, we discuss the wave equation
u
tt
a
2
u = f, (5.1)
where a > 0 is a constant. We will discover that solutions of the wave equation behave in
a dierent way comparing with the solutions of Laplaces equation or the heat equation.
5.1 Physical derivation
In physics or mechanics, the wave equation serves as a simplied model for the oscillations
on a vibrating string (n = 1), membrane (n = 2), or elastic solid (n = 3). In these models,
u(x, t) is the displacement of the pointed mass in certain directions of the point x at time
t 0. In (5.1), f models the external force.
For instance, we assume that an elastic solid occupied a region in R
3
without external
force. For any subregion G ,

F(x, t) is the contact force density acting on G through
the boundary G. Normalize the mass density to be unity. Let be the unit outer normal
vector of G. The acceleration within G is
d
2
dt
2
_
G
u dx =
_
G
u
tt
dx,
while the net contact force is
_
G
F dS,
then, by Newtons law, one has
_
G
u
tt
dx =
_
G
F dS =
_
G

F dx
where we have applied the divergence theorem. Therefore, we conclude that
u
tt
=

F.
89
90 CHAPTER 5. WAVE EQUATION
For elastic body,

F =

F(u), so
u
tt
+

F(u) = 0.
In the case of small oscillations, |u| is very small, and so

F(u) a
2
u, therefore,
u
tt
a
2
u = 0.
We remark that, from the physical interpretation, it is mathematically appropriate to
specify two initial conditions, on the displacement u and the velocity u
t
at t = 0. This will
appears often in the rest of this chapter.
5.2 Solution by Spherical means
Unlike the heat equation or Laplace equation, we will present an elegant method to solve
(5.1) rst for n = 1 directly and then for n 2 by the method of spherical means.
5.2.1 dAlemberts Formula, n = 1
We rst start with the Cauchy problem for the wave equation in one space dimension.
Consider
_
u
tt
u
xx
= 0, in R (0, )
u(x, 0) = g(x), u
t
(x, 0) = h(x), x R,
(5.2)
where g and h are given functions.
From the characteristic method we showed in Chapter 1, we know that the general
solution of the 1-D wave equation takes the following form
u(x, t) = F(x +t) +G(x t)
for any C
2
functions F and G. We now determine the F and G through the initial data.
It turns out that
F(x) +G(x) = g(x), F
(x) G
(x) = h(x), (5.3)

which implies
F
(x) =
1
2
(g
(x) +h(x)), G
(x) =
1
2
(g
(x) h(x)). (5.4)

Therefore, one derives the following dAlemberts formula:
u(x, t) =
1
2
[g(x +t) +g(x t)] +
1
2
_
x+t
xt
h(y) dy. (5.5)
We thus proved the following Theorem.
Theorem 5.2.1 Assume g(x) C
2
(R) and h(x) C
1
(R). Then u(x, t) dened by
dAlemberts formula (5.5) is the unique solution of (5.2).
5.2. SOLUTION BY SPHERICAL MEANS 91
In the following example, we apply the dAlemberts formula to an initial boundary
value problem using the reection method.
Example 5.2.2 Consider
_
_
u
tt
u
xx
= 0, in R
+
(0, )
u(x, 0) = g(x), u
t
(x, 0) = h(x), x R
+
,
u(0, t) = 0, t > 0
(5.6)
where g and h are given functions such that g(0) = h(0) = 0.
We perform the following odd reection.
u(x, t) =
_
u(x, t), x 0, t 0
u(x, t), x 0, t 0,
g(x) =
_
g(x), x 0,
g(x), x 0,
h(x, t) =
_
h(x), x 0
h(x), x 0.
(5.7)
It is clear that
_
u
tt
u
xx
= 0, in R (0, )
u(x, 0) = g(x), u
t
(x, 0) =

h(x), x R,
(5.8)
Hence, the dAlemberts formula gives
u(x, t) =
1
2
[ g(x +t) + g(x t)] +
1
2
_
x+t
xt
h(y) dy. (5.9)

Finally, we transform this expression into the region for x 0 and t 0:
u(x, t) =
_
_
1
2
[g(x +t) +g(x t)] +
1
2
_
x+t
xt
h(y) dy, if x t 0,
1
2
[g(x +t) g(t x)] +
1
2
_
x+t
x+t
h(y) dy, if 0 x t.
(5.10)
5.2.2 Spherical means
For n 2, u C
m
(R
n
[0, )) (m 2) solves the initial problem
_
u
tt
u = 0, in R
n
(0, )
u(x, 0) = g(x), u
t
(x, 0) = h(x), x R
n
,
(5.11)
From the physical experience on the wave propagation in R
3
, it appears that the wave
propagates spherically. This motivates the approach of spherical means. We will rst study
the average of u over certain spheres. These averages, as functions of the time t and the
radius r, solve the Euler-Poisson-Darboux equation, which could be transfered into the
one-dimensional wave equation for which we know how to solve.
Denition 5.2.3 Let x R
n
, t > 0, r > 0. Dene
U(x; r, t) = (u(y, t))
B(x,r)
=
1
|B(x, r)|
_
B(x,r)
u(y, t) dS
y
. (5.12)
Similarly, dene
G(x; r) = (g(x))
B(x,r)
, H(x; r) = (h(x))
B(x,r)
.
Lemma 5.2.4 Fix x R
n
, and let u solves (5.11), then U C
m
(
R
+
[0, )) solves the
initial value problem for Euler-Poisson-Darboux equation
_
_
_
U
tt
U
rr
n 1
r
U
r
= 0, in R
+
(0)
U(x; r, 0) = G(x; r), U
t
(x; r, 0) = H(x; r).
, in R
+
. (5.13)
Proof. For r > 0, one has
U
r
(x; r, t) =
r
n
(u(y, t))
B(x,r)
. (5.14)
Therefore, one has U
r
(x; r, t) 0 as r 0
+
. Then we dierentiate (5.14) to have
U
rr
= (u)
B(x,r)
+ (
1
n
1)(u)
B(x,r)
,
which implies that
lim
r0
+
U
rr
=
1
n
u.
Similar calculations show the regularity of U. Now, by the wave equation, one has
(r
n1
U
r
)
r
=
1
n(n)
_
B(x,r)
u
tt
(y, t) dS
y
= r
n1
(u
tt
)
B(x,r)
= r
n1
U
tt
.
This completes the proof of the theorem.
5.2.3 Kirchhos formula, n = 3
We now take n = 3 and suppose u C
2
(R
3
[0, )) solves the initial value problem
(5.11), and U, G, H dened in the Denition 5.2.3. The magic transform is
U = rU,

G = rG,

H = rH, (5.15)
which solves the following problem
_
U
tt

U
rr
= 0, in R
+
(0, )
U(x; r, 0) =

G(x; r),

U
t
(x; r, 0) =

H(x; r), r R
+
,
U(x; 0, t) = 0, t > 0.
(5.16)
From the Example 5.2.2, we know that for 0 r t,
U(x; r, r) =
1
2
[

G(x; r +t)

G(x; t r)] +
1
2
_
r+t
r+t
H(y) dy, (5.17)

We note that for continuous function u, one has
lim
r0
+
U(x; r, t) = u(x, t).
Therefore,
u(x, t) = lim
r0
+
U(x; r, t)
r
= lim
r0
+
[
G(x; r +t)

G(x; t r)
2r
+
1
2r
_
r+t
r+t
H(y) dy
=

G
(x; t) +

H(x; t).
Hence one reaches the Kirchhos formula
u(x, t) =

t
(t(g)
B(x,t)
) +t(h)
B(x,t)
. (5.18)
A further calculation gives the more explicit form
u(x, t) =
1
|B(x, t)|
_
B(x,t)
[th(y) +g(y) +Dg(y) (y x)] dS
y
. (5.19)
5.2.4 Poissons formula, n = 2
Unlike the 3D case, when n = 2, the Euler-Poisson-Darboux equation cannot be transferred
into the one-dimensional wave equation. We will apply the method of descent introduced
by Hadamard. The idea is to take the initial value problem (5.11) for n = 2 and regard it
as a problem for n = 3 where the third spatial variable x
3
does not appear.
To this purpose, we assume u C
2
(R
2
[0, )) solves (5.11) for n = 2, and we write
u(x
1
, x
2
, x
3
, t) = u(x
1
, x
2
, t). (5.20)
Therefore, for
g(x
1
, x
2
, x
3
) = g(x
1
, x
2
),

h(x
1
, x
2
, x
3
) = h(x
1
, x
2
),
we have
_
u
tt
u = 0, in R
3
(0, ),
u = g, u
t
=

h, on R
3
{t = 0}.
(5.21)
Denote by x = (x
1
, x
2
) R
2
, and x = (x
1
, x
2
, 0) R
3
, we know from the Kirchhos
formula (5.18) that
u(x, t) = u( x, t) =

t
(t( g)

B( x,t)
) +t(
h)

B( x,t)
, (5.22)
where

B( x, t) is the ball in R
3
centered at x with radius t > 0. For y B(x, t) and
(y) = (t
2
|y x|
2
)
1
2
, we note that
( g)

B( x,t)
=
1
4t
2
_

B( x,t)
g d
S
=
2
4t
2
_
B(x,t)
g(y)(1 + |D(y)|
2
)
1
2
dy.
Since (1 + |D(y)|
2
)
1
2
= t(t
2
|y x|
2
)
1
2
, we thus have
( g)

B( x,t)
=
2
2t
_
B(x,t)
g(y)
(t
2
|y x|
2
)
1
2
dy
=
t
2
_
g(y)
(t
2
|y x|
2
)
1
2
_
B(x,t)
.
Further simplication gives the following Poissons formula:
u(x, t) =
1
2
_
tg(y) +t
2
h(y) +tDg(y) (y x)
(t
2
|y x|
2
)
1
2
_
B(x,t)
(5.23)
for x R
2
, t > 0.
5.2.5 Further generalization
We now generalized the previous results to any dimensions for n 3. The following
identities can be proved by induction.
Lemma 5.2.5 Let : R R be C
k+1
. Then for k = 1, 2, , the following identities
hold
(
d
2
dr
2
)(
1
r
d
dr
)
k1
(r
2k1
(r)) = (
1
r
d
dr
)
k
(r
2k
d(r)
dr
).
(
1
r
d
dr
)
k1
(r
2k1
(r)) =
k1
j=0
k
j
r
j+1
d
j
(r)
dr
j
, where
k
j
are constants independent of .

k
0
= 1 3 5 (2k 1).
Now, we assume that n 3 is an odd integer and set n = 2k + 1. Then the following
transformation
_
U(r, t) = (
1
r
r
)
k1
(r
2k1
U(x; r, t))
G(r) = (
1
r
r
)
k1
(r
2k1
G(x; r))
H(r) = (
1
r
r
)
k1
(r
2k1
H(x; r)),
(5.24)
converts the Euler-Poisson-Darboux equation into the one-dimensional wave equation:
_
U
tt

U
rr
= 0, in R
+
(0, )
U(x; r, 0) =

G(x; r),

U
t
(x; r, 0) =

H(x; r), r R
+
,
U(x; 0, t) = 0, t > 0.
(5.25)
Solution for odd n. Follow the similar steps in the case of n = 3, we have for n = 2k +1
and
n
= 1 3 5 (n 2) the following representation formula:
u(x, t) =
1
n
_
(

t
)(
1
t
t
)
n3
2
(t
n2
(g)
B(x,t)
) + (
1
t
d
dt
)
n3
2
(t
n2
(h)
B(x,t)
)
_
. (5.26)
Solution for even n. Then, the method of descent will give the results for even n. For
even n, and
n
= 2 4 5 (n 2) n, we have the following solution formula
u(x, t) =
1
n
_
(

t
)(
1
t
t
)
n2
2
(t
n
(
g(y)
(t
2
|y x|
2
)
1
2
)
B(x,t)
)
_
+
1
n
_
(
1
t
d
dt
)
n2
2
(t
n
(
g(y)
(t
2
|y x|
2
)
1
2
)
B(x,t)
)
_
.
(5.27)
5.3 Nonhomogeneous problem
We now study the initial value problem for the nonhomogeneous wave equation
_
u
tt
u = f, in R
n
(0, )
u(x, 0) = 0, u
t
(x, 0) = 0, x R
n
.
(5.28)
By Duhamels principle, let u(x, t; s) be the solutions of
_
u
tt
(; s) u(; s) = 0, in R
n
(0, )
u(x, 0) = 0, u
t
(x, 0) = f(, s), on R
n
{t = s},
(5.29)
we set
u(x, t) =
_
t
0
u(x, t; s) ds, (x R
n
, t 0). (5.30)
The following theorem asserts this gives the solution of (5.28).
Theorem 5.3.1 For n 2 and f C
[
n
2
]+1
(R
n
[0, )), u(x, t) dened in (5.30) is a
solution of (5.28).
Proof. Through the solutions constructed in the last section, u(, ; s) C
2
(R
n
[0, ).
Then, we compute
u
t
(x, t) = u(x, t; t) +
_
t
0
u
t
(x, t; s) ds =
_
t
0
u
t
(x, t; s) ds,
u
tt
(x, t) = u
t
(x, t; t) +
_
t
0
u
tt
(x, t; s) ds
= f(x, t) +
_
t
0
u +tt(x, t; s) ds.
Furthermore
u(x, t) =
_
t
0
u(x, t; s) ds =
_
t
0
u
tt
(x, t; s) ds.
Therefore,
u
tt
(x, t) = u(x, t) = f(x, t), (x R
n
, t > 0).
Clearly, u(x, 0) = u
t
(x, 0) = 0 for any x R
n
.
Example 5.3.2 When n = 3, Kirchhos formula implies
u(x, t, ; s) = (t s)(f(y, s)
B(x,ts)
.
5.4. ENERGY METHOD 97
Therefore,
u(x, t) =
_
t
0
(t s)(f(y, s)
B(x,ts)
ds
=
1
4
_
t
0
_
B(x,ts)
f(y, s)
t s
dSds
=
1
4
_
t
0
_
B(x,r)
f(y, t r)
r
dSdr.
Hence, one obtains the solution of (5.28) for n = 3
u(x, t) =
1
4
_
B(x,t)
f(y, t |y x|)
|y x|
dy, (x R
3
, t 0) (5.31)
where
f(y, t |y x|)
4|y x|
is called retarded potential.
5.4 Energy method
In this section, we introduce the energy method, which is a powerful tool in the theory of
partial dierential equations. Let R
n
be a bounded, open set with a smooth boundary
. Dene
T
= (0, T],
T
=

T
\
T
, where T > 0.
5.4.1 Wave equation
We rst consider the following initial boundary value problem
_
_
u
tt
u = f, in
T
,
u = g, on
T
,
u
t
= h, on {t = 0}.
(5.32)
We now prove the uniqueness for the above problem using energy method.
Theorem 5.4.1 There exists at most one solution u C
2
(
T
) for (5.32).
Proof: Let u and u be two solutions of (5.32), then w = u u soloves
_
_
w
tt
w = 0, in
T
,
w = 0, on
T
,
w
t
= 0, on {t = 0}.
Dene the energy
e(t) =
1
2
_
(w
2
t
+ |Dw|
2
)(x, t) dx, (0 t T).
It is clear that
e
(t) =
_
w
t
w
tt
+Dw Dw
t
dx
=
_
w
t
(w
tt
w) dx = 0.
Therefore, e(t) = e(0) = 0 for all t [0, T], and so w
t
0 and Dw 0 in
T
. By the
initial boundary conditions, we know that w 0 and thus u = u in
T
.
5.4.2 Domain of dependence
Energy method is also useful to study some local behavior of solutions. We will illustrate
the domain of dependence of solutions to the wave equation and the nite propagation
speed properties.
Now, suppose u C
2
solves
u
tt
u = 0, in R
n
(0, ).
Fix x
0
R
n
, t
0
> 0. We dene the cone
C = {(x, t)|0 t t
0
, |x x
0
| t
0
t}.
We rst prove the following nite propagation speed property, which says that any
disturbance originating outside B(x
0
, t
0
) has no eect on the solution within C.
Theorem 5.4.2 If u u
t
0 on B(x
0
, t
0
), then u 0 within the cone C.
Proof: Dene the local energy
e(t) =
1
2
_
B(x
0
,t
0
t)
(u
2
t
+ |Du|
2
)(x, t) dx, t [0, t
0
].
We now compute
e
(t) =
_
B(x
0
,t=0t)
u
t
u
tt
+Du Du
t
dx
1
2
_
B(x
0
,t
0
t)
(u
2
t
+ |Du|
2
) dS
=
_
B(x
0
,t
0
t)
u
t
(u
tt
u) dx
1
2
_
B(x
0
,t
0
t)
(u
2
t
+ |Du|
2
2u
t
u
) dS
=
1
2
_
B(x
0
,t
0
t)
(u
2
t
+ |Du|
2
2u
t
u
) dS
5.4. ENERGY METHOD 99
We now observe that, by Cauchy-Schwartz inequality,
|u
t
u
|
1
2
u
2
t
+
1
2
|Du|
2
.
Therefore, we have
e
(t) 0
which means e(t) e(0) = 0 for all t [0, t
0
]. Thus u
t
0 and Du 0 and so u 0
within the cone C.
The bottom of the cone C is B(x
0
, t
0
) on the initial hyper-plane, is called the domain
of dependence for the point (x
0
, t
0
).
It is now clear that Theorem 5.4.2 also implies the uniqueness for the following Cauchy
problem
_
u
tt
u = f, in R
n
,
u = g, u
t
= h, on R
n
{t = 0}.
(5.33)
5.4.3 Energy method for Heat equation
We now study the large time asymptotic behavior of the solutions to the following initial
boundary value problem for heat equation.
_
_
u
t
u = 0, in (0, ),
u = 0, on ,
u = g, on {t = 0}.
(5.34)
Theorem 5.4.3 The solution of (5.34) converges to zero exponentially fast in time.
Proof: Let u be the solution of the problem (5.34), we dene the energy
e(t) =
_
u
2
dx.
Now we compute
e
(t) = 2
_
uu
t
dx
= 2
_
uu dx
= 2
_
|Du|
2
dx.
However, by Poincare inequality, one knows that there exists a positive constant C = C()
such that
C
_
u
2
dx
_
|Du|
2
dx,
for any u H
1
0
(). Therefore, we have
e
(t) + 2Ce(t) 0,
which implies
e(t) e(0)e
2Ct
, t > 0.
Therefore, one sees that e(t) 0 as t and thus u converges to zero in energy norm
exponentially fast in time.
5.5 Initial boundary value problem: n = 1
For wave equation in one spatial dimension, there are several approaches to solve the initial
boundary value problem, including the characteristic method, reection method and the
Fourier method (separation of variables). We will focus on the method of separation of
variables.
5.5.1 Homogeneous equation
The Fourier method is motivated from the physical fact: vibrations can be decomposed
into simple oscillations according to its frequency. Therefore, one can try to look for the
solution of the form
u
n
(x, t) = X
n
(x)T
n
(t), n = 1, 2, ,
then, determine the constants in the following formula
u(x, t) =
n=1
C
n
X
n
(x)T
n
(t),
to solve the problem.
We now show this idea using the following example.
_
_
u
tt
a
2
u
xx
= 0, 0 < x < l, t > 0,
u(0, t) = 0, u(l, t) = 0, t 0,
u(x, 0) = f(x), u
t
(x, 0) = g(x), 0 x l,
(5.35)
where f and g are C
1
satisfying the compatibility condition
f(0) = f(l) = 0, g(0) = g(l) = 0.
Solution: The Fourier method consists the following steps.
5.5. INITIAL BOUNDARY VALUE PROBLEM: N = 1 101
(i) Separation of variables. We rst look for the solution of the form
u(x, t) = X(x)T(t) = 0
satisfying the boundary conditions. This leads to
T
(t)
a
2
T(t)
=
X
(x)
X(x)
= , X(x)T(t) = 0.
It is clear that is a constant. We thus have
_
T
(t) + a
2
T(t) = 0, t > 0
X
(x) + X(x) = 0, 0 < x < l.

(5.36)
For u(x, t) to satisfy the boundary condition, one requires X(0) = X(l) = 0.
(ii) Solve the eigenvalue problem. We now solve the following eigenvalue problem
_
X
(x) + X(x) = 0, 0 < x < l

X(0) = X(l) = 0.
(5.37)
where the is called the eigenvalue if it gives a non-zero solution X
(x) which is called

the associated eigenfunction. There are three cases concerning with the sign of .
(a) If < 0, the general solution is
X(x) = Ae
x
+Be
x
,
where A and B will be determined through the boundary conditions. In deed,
X(0) = A +B = 0
X(l) = Ae
l
+Be
l
= 0.
Therefore, A = B = 0 and < 0 is not the eigenvalue.
(b) If = 0, then X(x) = A+Bx which is identically zero with the boundary conditions
X(0) = 0 = X(l). So, = 0 is not an eigenvalue either.
(c) We now consider = k
2
> 0. The general solution takes the form
X(x) = Acos(kx) +Bsin(kx).
Using the boundary condition X(0) = 0 = X(l), one has A = 0 and Bsin(kl) = 0. Since
we need the non-zero solution, B = 0 and
k =
n
l
,
n
= (
n
l
)
2
, n = 1, 2, (5.38)
This gives the eigenvalues for problem (5.37) and the corresponding eigenfunctions are
X
n
(x) = sin(
n
l
x), n = 1, 2, (5.39)
We now substitute
n
into the rst equation in (5.36),
T
n
(t) = c
n
cos(
an
l
t) +d
n
sin(
an
l
t), n = 1, 2,
Therefore, we found
u
n
(x, t) = X
n
(x)T
n
(t) = [c
n
cos(
an
l
t) +d
n
sin(
an
l
t)] sin(
n
l
x), n = 1, 2, (5.40)
where c
n
and d
n
are arbitrary constants to be determined later.
(iii) In this step, we hope to nd the solution of the initial boundary value problem
(5.35) through linear combination of u
n
(x, t). Let
u(x, t) =
n=1
[c
n
cos(
an
l
t) +d
n
sin(
an
l
t)] sin(
n
l
x). (5.41)
Using the initial conditions, we require
u(x, 0) =
n=1
c
n
sin(
n
l
x) = f(x),
u
t
(x, 0) =
n=1
d
n
an
l
sin(
n
l
x) = g(x).
(5.42)
From the conditions of f and g and we require they are C
1
, then
c
n
=
2
l
_
l
0
f(x) sin(
n
l
x) dx,
d
n
=
2
an
_
l
0
g(x) sin(
n
l
x) dx.
(5.43)
We therefore obtained the formal solution of the problem (5.35). One could further verify
that the series solution has good convergent properties to be a classical solution if f and
g are C
4
([0, l]) and f(0) = f
(0) = g(0) = f(l) = f
(l) = g(l) = 0.
We see from this example that as long as the corresponding eigenvalue problem is
well solved, we are able to nd at least the formal solution of the initial value problem
for homogeneous wave equation in one space dimension. The eigenvalue problem will be
discussed in great details at the end of this chapter. For the formal solution to be classical
solution, it often requires very restrictive properties on the data, this however could be
resolved by the notion of weak solution.
5.5.2 Non-homogeneous equation
In this section, we further show that the eigenfunctions have further applications to solve
the initial boundary value problem for the non-homogeneous equation. This is so-called
expansion about eigenfunctions.
Again, we will illustrate our ideas using specic example.
_
_
u
tt
a
2
u
xx
= f(x, t), 0 < x < l, t > 0,
u(0, t) = 0, u(l, t) = 0, t 0,
u(x, 0) = 0, u
t
(x, 0) = 0, 0 x l,
(5.44)
Solution: Unless f(x, t) has certain specic structure, u(x, t) = X(x)T(t) is not a solution
of the non-homogeneous wave equation. However, motivated by the constant variation
method of ODEs, we seek the solution of the following form
u(x, t) =
n=1
T
n
(t)X
n
(x) (5.45)
where X
n
(x) are the eigenfunctions for the corresponding homogeneous wave equation
under the same boundary conditions. In this case,
X
n
(x) = sin(
n
l
x), n = 1, 2, .
We hope that the solution could be expanded into the series of X
n
(x) with variant coef-
cients T
n
(t). Clearly, (5.45) satises the boundary conditions. We will determine T
n
(t)
through the equation and the initial conditions. Therefore, T
n
(t) satises
n=1
[T
(t) + (
an
l
)
2
T
n
(t)] sin(
n
l
x) = f(x, t),
u(x, 0) =
n=1
T
n
(0) sin(
n
l
x) = 0
u
t
(x, 0) =
n=1
T
n
(0) sin(
n
l
x) = 0.
(5.46)
Setting
f
n
(t) =
l
2
_
l
0
f(x, t) sin(
n
l
x) dx, n = 1, 2, ,
one has
_
T
n
(t) + (
an
l
)
2
T
n
(t) = f
n
(t)
T
n
(0) = T
n
(0) = 0, n = 1, 2, .
Hence, we found
T
n
(t) =
l
an
_
t
0
f
n
() sin(
an
l
(t )) d, n = 1, 2, .
We nally obtain the formal solution of the problem (5.44)
u(x, t) =
n=1
[
l
an
_
t
0
f
n
() sin(
an
l
(t )) d] sin(
n
l
x). (5.47)
For general initial boundary value problem, one can apply the linear superposition
principle and some basic techniques to transfer the non-homogeneous boundary conditions
to homogeneous ones. We omit the details.
5.5.3 Sturm-Liouville Theory
In this subsection, we briey discuss the theory of Sturm-Liouville for eigenvalue problem.
Consider the initial boundary value problem
_
_
Lu = 0, x (a, b), t > 0,
u(x, 0) = (x), u
t
(x, 0) = (x), x [a, b]
1
u(a, t) +
2
u
x
(a, t) = 0,
1
u(b, t) +
2
u
x
(b, t) = 0, t 0,
(5.48)
where
Lu = A(t)u
tt
+C(x)u
xx
+D(t)u
t
+E(x)u
x
+ (F
1
(t) +F
2
(x))u.
We assume that
i
and
i
are constants for i = 1, 2 such that
2
1
+
2
2
= 0,
2
1
+
2
2
= 0.
Suppose A(t) A
0
> 0, C(x) C
0
< 0, A
0
and C
0
are constants. All other coecients
are continuous. We also assume F
2
(x) > 0.
If we want to perform the method of separation of variables, u(x, t) = X(x)T(t), for
the possible eigenvalue, we need to solve the equation of T(t)
AT
(t) +DT
(t) +F
1
T + T = 0 (5.49)
and the eigenvalue problem
_
_
CX
+EX
+F
2
X X = 0,
1
X(a) +
2
X
(a) = 0,
1
X(b) +
2
X
(b) = 0.
(5.50)
In order to solve this eigenvalue problem, we rst convert it into the self-adjoint form.
Multiplying the equation by
S =
1
C
exp{
_
x
0
E
C
dx},
the equation became
[p(x)X
q(x)X + SX = 0, (5.51)
where
p(x) = SC exp{
_
x
0
E
C
0
dx} = p
0
> 0,
q(x) = SF
2
> 0, S(x) > 0.
Therefore, (5.50) has been changed into the standard form of Sturm-Liouville eigenvalue
problem
_
_
[p(x)X
q(x)X + SX = 0
1
X(a) +
2
X
(a) = 0,
1
X(b) +
2
X
(b) = 0.
(5.52)
We rst list some properties of the eigenfunctions.
Theorem 5.5.3 Let X
1
and X
2
are eigenfunctions corresponding to the same eigenvalue
, then X
1
= CX
2
for certain nonzero constant C.
Theorem 5.5.4 If X
1
and X
2
are eigenfunctions corresponding to eigenvalues
1
and
2
respectively, then
_
b
a
SX
1
X
2
dx = 0.
We now introduce the working space for the Sturm-Liouville theory. For 0 < S(x)
C([a, b]), we rst dene
L
2
S
{y(x) L
1
loc
([a, b]) :
_
b
a
S(x)y
2
(x) dx < },
equipped with the inner product
(y
1
, y
2
)
S

_
b
a
S(x)y
1
(x)y
2
(x) dx, y
1
, y
2
L
2
S
.
For the functions from C
1
0
([a, b]) (C
1
functions vanishing at endpoints), we introduce an
inner product
(y
1
, y
2
)
H

_
b
a
[p(x)y
1
y
2
+q(x)y
1
y
2
] dx, y
1
, y
2
C
1
0
([a, b]), (5.53)
where p(x) and q(x) are continuous functions such that there is a constant p
0
> 0 and
p(x) > p
0
, q(x) > 0 for any x [a, b].
(5.53) denes a norm
H
, we denote H
0,1
p,q
for the closure space of C
1
0
([a, b]) under the
norm
H
. This is a Hilbert space.
We now use the Dirichlet boundary condition as the example to illustrate the procedure.
Consider now
_
[p(x)X
q(x)X + SX = 0
X(a) = X(b) = 0.
(5.54)
Denition 5.5.5 X H
0,1
p,q
is said to be a weak solution of (5.54) if it satises
(X, y)
H
= (X, y)
S
for any y C
1
0
[a, b].
Consider the functional
J(X) =
(X, X)
H
(X, X)
S
, X H
0,1
p,q
. (5.55)
Theorem 5.5.6 There exits 0 = X H
0,1
p,q
such that
J(X) = inf
zH
0,1
p,q
J(z).
Now, we set K
1
= H
0,1
p,q
, we call
1
= inf
0=XK
1
J(X)
the rst eigenvalue of (5.54), and the function 0 = X
1
K
1
such that
(X
1
, X
1
)
H
(X
1
, X
1
)
S
=
1
the eigenfunction corresponding to
1
.
Now, we set
K
2
= {X K
1
|(X, X
1
)
S
= 0}.
Similarly, we dene
2
= inf
XK
2
J(X)
as the second eigenvalue and 0 = X
2
K
2
such that
(X
2
, X
2
)
H
(X
2
, X
2
)
S
=
2
5.6. PROBLEMS 107
2
. Inductively, we set
K
n
= {X K
1
|(X, X
1
)
S
= (X, X
2
)
S
= = (X, X
n1
)
S
= 0},
and the n-th eigenvalue
n
= inf
XK
n
J(X)
and 0 = X
n
K
n
such that
(X
n
, X
n
)
H
(X
n
, X
n
)
S
=
n
n
.
Theorem 5.5.7 Let
n
be eigenvalues of (5.54) and X
n
the corresponding eigenfunctions
of
n
. Then
0 <
1
<
2
< <
n
<

n
as n .
{X
n
}
n=1
form an orthogonal basis of L
2
S
.
We therefore solved the Sturm-Liouville problem (5.54). The general case can be treated
in a similar way.
5.6 Problems
Problem 1. Let u C
2
(R [0, )) be the solution of
_
u
tt
u
xx
= 0, x R, t > 0,
u(x, 0) = (x), u
t
(x, 0) = (x),
where (x) and (x) have compact support. Dene
K(t) =
1
2
_
+
u
2
t
(x, t) dx
P(t) =
1
2
_
+
u
2
x
(x, t) dx.
Prove the following
(a). K(t) +P(t)=constant;
(b). When t is suciently large, K(t) = P(t).
Problem 2. Solve the following initial value problem
_
_
u
tt
u
xx
= 0, x R, t > ax,
u|
t=ax
= (x), x R,
u
t
|
t=ax
= (x), x R,
where a = 1. If the initial data are given on x [b
1
, b
2
], then on what region can you
determine the solution?
Problem 3. Prove the initial value problem
_
u
tt
u
xx
= 6(x +t), x R, t > x,
u|
t=x
= 0, u
t
|
t=x
= (x), x R,
has solutions if and only if (x)3x
2
= constat. When the solutions exist, it is not unique.
Why does this problem behave dierently from problem 2?
Problem 4. Solve the following problems
(a)
_
u
tt
a
2
u
xx
= 0, x R, t > 0,
u|
xat=0
= (x), u|
x+at=0
= (x), (0) = (0);
(b)
_
u
tt
a
2
u
xx
= 0, x R, t > 0,
u|
t=0
= (x), u|
xat=0
= (x), (0) = (0);
(c)
_
u
xx
+ 2 cos(x)u
xy
sin
2
(x)u
yy
sin(x)u
y
= 0,
u|
y=sin(x)
= (x), u
y
|
y=sin(x)
= (x), x, y R;
(d)
_
u
xx
+yu
xy
+
1
2
u
y
= 0, x R, y < 0
u|
y=0
= (x), u
y
|
y=0
< ;
(e)
_
y
2
u
yy
x
2
u
xx
= 0, x R, y > 1,
u|
y=1
= f(x), u
y
|
y=1
= g(x).
Problem 5. Let h be a nonzero constant, F and G be C
2
functions. Prove that
u(x, t) =
1
h x
(F(x at) +G(x +at))
is the general solution of
[(1
x
h
)
2
u
x
]
x
=
1
a
2
(1
x
h
)
2
u
tt
.
5.6. PROBLEMS 109
Find the solution of this equation with the following initial conditions
u(x, 0) = (x), u
t
(x, 0) = (x).
Problem 6. If w(x, t; ) is the solution of
(A)
_
w
tt
a
2
w
xx
= 0, x R, t > ,
w(x, ; ) = 0, w
t
(x, ; ) = f(x, ),
prove that the function
u(x, t) =
_
t
0
w(x, t; ) d
is the solution of
(B)
_
u
tt
a
2
u
xx
= f(x, t), x R, t > 0,
u(x, 0) = 0, u
t
(x, 0) = 0,
This is the so-called Duhamel principle.
Problem 7. Use the separation of variable method to solve the following problems
(a)
_
_
u
tt
a
2
u
xx
= 0, 0 < x < l, t > 0,
u(x, 0) = x
2
2lx, u
t
(x, 0) = 0,
u(0, t) = u
x
(l, t) = 0, t 0;
(b)
_
_
u
tt
a
2
u
xx
= 0, 0 < x < l, t > 0,
u(x, 0) =
_
hx
c
, x [0, c]
h
lx
lc
, x (c, l].
u
t
(x, 0) = 0,
u(0, t) = u(l, t) = 0, t 0;
Problem 8. Solve the following initial boundary value problems.
(a)
_
_
u
tt
a
2
u
xx
= Ax, 0 < x < l, t > 0,
u(x, 0) = u
t
(x, 0) = 0, 0 x l,
u(0, t) = u(l, t) = 0, t 0;
(b)
_
_
u
tt
a
2
u
xx
= 0, 0 < x < l, t > 0,
u(x, 0) = u
t
(x, 0) = 0, 0 x l,
u(0, t) = 0, u(l, t) = A(sin(t) t), t 0;
(c)
_
_
u
tt
a
2
u
xx
= bx, 0 < x < l, t > 0,
u(x, 0) = u
t
(x, 0) = 0, 0 x l,
u(0, t) = 0, u(l, t) = Bt, t 0.
Problem 9. Let x = r cos(), y = r sin() be the coordinates in R
2
. Solve the following
problem
_
u
xx
+u
yy
= 0, 0 < r < R, 0 < < ,
u|
=0
= u|
=
= 0, u|
r=R
= f(x, y).
Problem 10. Solve the initial boundary value problem
_
_
u
tt
+a
2
u
xxxx
= 0, 0 < x < l, t > 0,
u(x, 0) = x(x l), u
t
(x, 0) = 0, 0 x l,
u(0, t) = u(l, t) = u
xx
(0, t) = u
xx
(l, t) = 0, t 0.
_
u
t
= a
2
u
xx
b
2
u, 0 < x < l, t > 0,
u(x, 0) = u
0
, u(0, t) = 0, u
x
(l, t) +hu(l, t) = 0.
_
_
u
t
= c
2
(u
xx
+u
yy
), 0 < x < a, 0 < y < b, t > 0,
u(x, y, 0) = A,
u(0, y, t) = 0 = u
x
(a, y, t),
u
y
(x, 0, t) = u(x, b, t) = 0.
Problem 13. Assume the eigenvalue problem
_
_
[p(x)X
(x)]
q(x)X(x) + (x)X(x) = 0,
p p
0
> 0, >
0
> 0, q 0, x (0, l),
A
0
X(0) +B
0
X
(0) = 0,
A
1
X(l) +B
1
X
(l) = 0,
A
2
0
+B
2
0
= 0, A
2
1
+B
2
1
= 0,
has eigenvalue. Where A
i
and B
i
(i = 0, 1) are constants, B
0
0, A
0
0, A
1
0, and
B
1
0. Prove that the eigenvalue is positive. Furthermore, if X
1
and X
2
correspond to
dierent eigenvalues
1
and
2
, then
_
l
0
(x)X
1
(x)X
2
(x) dx = 0.
5.6. PROBLEMS 111
Problem 14. Using the energy function
E(t) =
1
2
_
l
0
(ku
2
x
+ u
2
t
+qu
2
) dx
to prove that the problem
_
_
(x)u
tt
= (k(x)u
x
)
x
q(x)u, 0 < x < l, t > 0,
u(x, 0) = u
t
(x, 0) = 0,
u(0, t) = u(l, t) = 0, t > 0,
has the only solution u 0. Here k(x) k
0
> 0, q(x) 0, (x)
0
> 0, k
0
and
0
are
two constants.
Problem 15. If w(x, t; ) is the solution of
(C)
_
_
w
tt
a
2
w
xx
= 0, 0 < x < l, t > ,
w|
x=0
= w|
x=l
= 0, t >
w(x, ; ) = 0, w
t
(x, ; ) = f(x, ), 0 x l,
prove that the function
u(x, t) =
_
t
0
w(x, t; ) d
is the solution of
(D)
_
_
u
tt
a
2
u
xx
= f(x, t), 0 < x < l, t > 0,
u(0, t) = u(l, t) = 0, t 0,
u(x, 0) = 0, u
t
(x, 0) = 0, 0 < x < l,
This is the Duhamel principle.
Problem 16. Assume

E = (E
1
, E
2
, E
3
) and

B = (B
1
, B
2
, B
3
) are solution of the Maxwell
system
_
E
t
=

B
B
t
=

E

B =

E = 0.
Prove that if u = E
i
or u = B
i
for i = 1, 2, 3, the u
tt
u = 0.
Problem 17. Use the Kirchho formula to solve
_
u
tt
= a
2
(u
xx
+u
yy
+u
zz
), (x, y, z) R
3
, t > 0
u|
t=0
= x
3
+y
2
z, u
t
|
t=0
= 0.
Problem 18. Assume u(x, t) is the solution for the following problem
_
u
tt
a
2
u = 0, x R
3
, t > 0
u(x, 0) = (x), u
t
(x, 0) = (x).
If (x), (x) C
c
(R
3
), prove there is a constant C such that
|u(x, t)|
C
t
, x R
3
, t > 0.
Problem 19. Use the Poissons formula to solve
_
u
tt
= a
2
(u
xx
+u
yy
), (x, y) R
2
, t > 0
u|
t=0
= x
2
(x +y), u
t
|
t=0
= 0.
Problem 20. Solve the following initial value problem
_
u
tt
= a
2
(u
xx
+u
yy
) +c
2
u, (x, y) R
2
, t > 0
u|
t=0
= (x, y), u
t
|
t=0
= (x, y).
(Hint: choose v(x, y, z) = e
cz
a
u(x, y), apply Kirchho formula to v.)
Problem 21. Let u(x, y, t) be the solution of
_
u
tt
4(u
xx
+u
yy
) = 0u, (x, y) R
2
, t > 0
u|
t=0
= (x, y), u
t
|
t=0
= (x, y).
Here
(x, y) = (x) =
_
0, (x, y) ,
1, (x, y) R
2
\ ,
where = {(x, y)||x| 1, |y| 1}. Determine the region where u(x, y, t) 0 for t > 0.
Problem 22. Solve the following problem
_
u
tt
u = 2(y t), (x, y, z) R
3
, t > 0
u|
t=0
= 0, u
t
|
t=0
= x
2
+yz.
(Hint: Decompose the problem into two using the linear superposition principle.)
Problem 23. A vibrating string under frictional force satises the following equation
u
tt
a
2
u
xx
cu
t
= 0, c > 0.
Prove the energy of this equation decreasing in time. With this fact, prove the uniqueness
of the following problem
_
_
u
tt
a
2
u
xx
cu
t
= f(x, t), 0 < x < l, t > 0,
u(0, t) = u(l, t) = 0, t 0,
u(x, 0) = (x), u
t
(x, 0) = (x), 0 x l.

aPDEs PAN2008 - Unknown

Diunggah oleh

Informasi Dokumen

Deskripsi Asli:

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

aPDEs PAN2008 - Unknown

Diunggah oleh

Hak Cipta:

Format Tersedia

Chapter 1

u = 4(x, y, z), (x, y, z) R

the principal part of the operator P(x, D).

= 0, the equation (1.25) is of hyperbolic type, and it has a simple form

1 is the imaginary unit. It is not hard to

= 0. Therefore, (1.21) takes the following standard form

), one derives from the Divergence Theorem (letting w = Du)

), by choosing w = uDv or w = vDu respectively in the Divergence

(y) centered at y with radius is inside . On the region \

(y), we substitute v(x)

u(x) dx. (3.24)

(r) = (Du(x + rz) z)

) of the boundary value problem

(). Thus harmonic

= {x | dist(x, ) > }. Then, it holds that

f uniformly on compact subsets of .

) for any > 0, and so u C

regularity does not go to extreme of the harmonic functions. The following

on the boundary. Unfortunately, it is unknown to us. The idea is to introduce a

) solves the boundary value problem (3.33) for some contin-

on the unit sphere. According to formula (3.43), for

K(x, y)|g(y) g(x

K(x, y)|g(y) g(x

> 0 so small such that

, then v 0 on the region A =

solution of (4.7) on the region

() and it satises the Heat equation. This is

(). We now use

solution to (4.7), and

, }. We see that v(x

= 0, which with the boundary condition, give zero solution.

(x, 0) = a(x) +b(

, and then the solution of (4.37) is given by

(x) = h(x), (5.3)

(x) h(x)). (5.4)

h(y) dy. (5.9)

H(y) dy, (5.17)

(x) + X(x) = 0, 0 < x < l.

(x) + X(x) = 0, 0 < x < l

(x) which is called

(0) = g(0) = f(l) = f

Anda mungkin juga menyukai