You are on page 1of 36

124A PARTIAL DIFFERENTIAL EQUATIONS

DENIS A. LABUTIN

This is the outline for the course. We emphasize the key ideas and the structure,
and often omit the details. The details will be covered on the lectures. They also
can be found in the textbook (W. A. Strauss PDEs, an Introduction, 2nd ed.) and
in the extended UCSB 124-course notes by Viktor Grigoryan. The two sources
contain much more than will be covered during the quarter.

1. Basics

(1) A partial differential equation (in two independent variables) is a relation


F (x, y, u, ux , uy , uxx , uxy , . . .) = 0
for the unknown function u(x, y) . All laws in sciences are formulated in
the form of (system of) PDEs. A function u is a solution of the PDE in a
region if
F (x, y, u(x, y), ux (x, y), uy (x, y), uxx (x, y), uxy (x, y), . . .) 0
for (x, y) . Thus it is very easy to check if a given function is a solution
of the given PDE.
(2) The general solution to the given PDE is the formula for all solutions. For
example, we know from the calculus that the general solution u(x) of
du
u=0
dx
is
u = Cex ,
where C is any constant.
Another example: we will learn later that the general solution u(x, y)
to the equation
utt uxx = 0
is given by
u(t, x) = f (x t) + g(x + t),
where f, g are arbitrary functions of one variable. This means that all
solutions of the equation are given by this formula. It is impossible to
derive this fact now, we need to develop some techniques for this.
(3) Calculus technique for solving simple PDEs.
Date: July 16, 2013.
1

DENIS A. LABUTIN

(4) Some simple PDEs can be integrated using just the calculus technique. In
order to successfully solve the problems of this type one needs to understand
the following argument: according to the basic calculus the general solution
z(x) to the simplest equation
dz
=0
dx
is
z(x) = C,
where C is any constant. But this means that the general solution u(x, y)
to the equation
ux = 0
is
u(x, y) = C(y),
where C(y) is any constant with respect to x , which means that it is any
function of y . When writing the solutions down it is better not to use C
to denote functions, and keep it for constants only. Thus a better way of
writing is
ux = 0 u = (y),
where is any function of one variable.
(5) Similarly, the basic calculus tells us that the solution z(x) of z 0 = f (x)
is
Z
0
z = f (x) z = f (x) dx + C,
where C is any constant. This means that
Z
ux = f (x, y) u = f (x, y) dx + (y),
where is any function of one variable. Below are some examples of
calculus technique solutions to PDEs.
(6) Solve (this means find the general solution to) the equation
ux = e2x+y .
There is no /y in the left hand side, and no u in the right hand side.
Therefore the solution u(x, y) is given by
Z
u =
e2x+y dx + (y)
Z
= ey e2x dx + (y)
1
= ey e2x + (y).
2
Thus the answer is
u=

e2x+y
+ (y),
2

where is any function.


(7) Here is one more example of the calculus technique for solving simple PDEs.
Find all solutions u(x, y) , v(x, y) to the equations
ux + u = 0,

vxx + v = 0.

124A PARTIAL DIFFERENTIAL EQUATIONS

Using the calculus separation of variables we see that the equation


z0 + z = 0
has the general solution
z(x) = Cex .
The equation for u is the same equation. There is no /y and the
variable y enters only as a parameter. Hence
ux + u = 0 u = (y)ex ,
where is any function.
Using the calculus separation of variables we see that the equation
z 00 + z = 0
has the general solution
z(x) = C1 cos x + C2 sin x,
where C1,2 are arbitrary constants. The equation for v is the same
equation. There is no /y and the variable y enters only as a parameter.
Hence
vxx + v = 0 v = (y) cos x + (y) sin x,
where , are arbitrary functions.
(8) Frequently there is a need to apply several calculus steps. Here is an example of such problem. Solve
uxy = 0.
Denote
v = uy .
Then
uxy = 0 vx = 0 v = (y),
where is any function of one variable. Next,
Z
uy = (y) u = (y) dy + (x),
where , are any functions of one variable. Since is any, then
Z
(y) dy
is also any function of one variable. Thus the answer is
u = (y) + (x),
where , are any functions of one variable.
Notice that , are the functions of one variable, and u(x, y) = (y)+
(x) is a function of two independent variables of a very special form. For
example (x)(x) is not of this form. Hence x + ey is a solution, but
xey is not.
(9) The integrating factor technique from calculus will also be used in the
course. Equation for z(t)
dz
+ a(t)z = f (t)
dt

DENIS A. LABUTIN

can be integrated using the following observation. Multiply the equation


by the integrating factor
R
e a(t) dt
to discover that
e

a(t) dt dz

+e

a(t) dt

dt
The point here is that now
e

a(t) dt dz

+e

a(t) dt

dt
Hence the equation becomes
d

a(t)z =

dt

a(t)z = e

a(t) dt

f (t).

d  R a(t) dt 
e
z .
dt

= g(t)

which can be integrated.


Problem 1. Find the general solution of
uxy + a(x, y)ux = 0
in a neighborhood of (0, 0) . Here a(x, y) is a given function.
Set
v = ux
to discover at once that
vy + a(x, y)v = 0.
This is an ODE in the y -variable, and x enters as a parameter. Separating the variables discover
dv
= a(x, y) dy
v
and find after the integration
ln |v(x, y)| =e

Ry
0

a(x,) d+C(x)

v = (x)e

Ry
0

a(x,) d

with an arbitrary function . Returning to u discover that


ux = (x)e

Ry
0

a(x,) d

This is just an ODE in the x -variable, and y enters as a parameter.


Integrating we find
Z x
Ry
u(x, y) =
()e 0 a(,) d d + (y),
0

where , are arbitrary functions of one variable.


Problem 2. Find the general solution of
uxy

1
1
ux +
uy = 0.
xy
xy

Hint: use the observation that

(x y)vx + v = (x y)v ,
x

(x y)vy v = (x y)v .
y

124A PARTIAL DIFFERENTIAL EQUATIONS

(10) Operators and equations. Let u be any function. A differential operator is an expression involving u and its derivatives. For example, the
Monge-Ampere operator is
det D2 u = uxx uyy u2xy ,
the Laplace operator is
u = uxx + uyy ,
the wave operator is
u = utt uxx ,

c u = utt c2 uxx ,

the heat operator is


ut uxx ,
the transport operator is
ux + uy ,
the Schroedinger operator is
1
ut uxx .
1
Operators a usually denoted as Au , or A[u] , or A(u) , or Lu , ... .
(11) The most important class of the operators is the class of linear operators.
Operator L is called linear if the following two conditions hold:
L(u + v) = L(u) + L(v),

L(Cu) = CL(u)

for all functions u, v and any constant C . To check if the operator is


linear we just need to verify these conditions.
(12) A PDE is called linear homogeneous if it can be written in the form
Lu = 0
with some linear operator L .
A PDE is called linear inhomogeneous if it can be written in the form
Lu = f (x, y)
with some linear operator L . The important part here is that in the right
hand side we have some function of x, y but the unknown u is absent.
(13) Our main objects of study will be the transport operator
(t + cx )u = ut + cux ,

u = u(t, x),

the wave operator



def
u = t2 + x2 u = utt uxx ,

u = u(t, x),

the heat operator


(t x2 )u = ut uxx ,

u = u(t, x),

and the Laplace operator


def

u = (x2 + y2 )u = uxx + uyy ,

u = u(x, y).

Show that they are linear operators.


Linear equations can have coefficients depending on x, y . For example,
show that the equation
ux + yuy = 0

DENIS A. LABUTIN

is a linear (homogeneous) equation.


(14) Equation
ux + uuy = 0
is nonlinear. Indeed, every term in it contains u . Hence the equation is
linear if and only if the operator
Au = ux + uuy
is linear. But A is not linear because the two conditions of the linearity
are not satisfied. For example if C is a constant, then
A(Cu) = (Cu)x + (Cu)(Cu)y = Cux + C 2 uuy .
At the same time
CAu = Cux + Cuuy .
Hence
A(Cu) 6= CA(u)
for all constants C .
(15) Superposition principle for linear equations. The key feature of linear equations is the superposition principle. It is also called the linearity
principle. It consists of two parts.
The first part states that if u1 , . . . , uN are solutions to the linear
homogeneous equation
Lu = 0,
then for any constants C1 , . . . , CN the function C1 u1 + + CN uN is
also a solution:
L(C1 u1 + + CN uN ) = 0.
This follows at once from the definition of the linearity of L .
The second part states that if u0 solves the linear inhomogeneous equation
Lu = f,
and if v is the general solution to the homogeneous equation
Lv = 0
then the general solution to the inhomogeneous equation is given by
u = u0 + v.
(16) The second part is the most important and useful for problem solving.
Indeed, in plain language is states that
{all solutions to Lu = f }

= {some solution to Lu = f }
+{all solutions to Lu = 0}.

Thus to find all solution to Lu = f it is enough to find all solutions only


to Lu = 0 and somehow guess just one particular solution to Lu = f .
(17) Let us illustrate the power of the superposition principle by solving a problem. Solve (that is, find all solutions to)
uxy = x2 y.

124A PARTIAL DIFFERENTIAL EQUATIONS

This is a linear inhomogeneous equation. Hence we can apply the superposition principle. We have computed earlier the general solution to the
homogeneous equation:
vxy = 0 v = f (x) + g(y),
where f, g any functions. Hence we just need to guess one solution to the
inhomogeneous equation. This is easy for our equation. For example
x3 y 2
x3 y 2
=
3 2
6
will do the job. Therefore, by the superposition principle the answer is:
u0 =

u(x, y) =

x3 y 2
+ f (x) + g(y),
6

where f, g any functions.


(18) The superposition principle will be one of the main tools in this course both
for the theory and the problem solving. Superposition principle does not
hold for nonlinear equations.

2. Changing variables in PDEs

(1) The changing variables technique is fundamental for PDEs. It is based on


various tools from calculus such as chain rule, implicit differentiation, and
so on. Many problems in this course require the ability to change variables
correctly. First we explain the generalities, underline the difficult part in
the change of variables procedure, and then solve some problems.
(2) To abbreviate we write x = (x1 , . . . , xn ) , = (1 , . . . , n ) , ux =
(x1 u, . . . , xn u) , and so on. In the differential equation
F (x, u, ux , . . .) = 0
we would like to change the variables x . The change of variables
x = H()
at once gives
v() = u(H()).
and
(2.1)

F (H(), v(), (ux )(H()), . . .) = 0.


We know F and H . Consequently we almost have the differential equation for the new unknown v() .
It is just left to relate (ux )(H()) with v . This is the somewhat
complicated part. To accomplish that we need the inverse change
= (x).
The inverse implies
(H()) = ,

H((x)) = x.

DENIS A. LABUTIN

Therefore together with v() = u(H()) we also have u(x) = v((x)) .


By the chain rule


x (x),
ux (x) = v()
=(x)

and we need to express this in variables:







x (x)
(ux )(H()) = v()
=(x)

x=H()

= v () (x ) (H()).
We see that the main computation when changing variables is finding
(x ) (H())
and the higher derivatives needed for uxx , uxxx , . . . . That is, we need
to express the derivative x = x (x) as a function of , x = H() .
When this is done, equation (2.1) becomes
0 = F (H(), v(), v () (x ) (H()), . . .)
m
0 = G(, v, v , . . .) .
(3) Thus to change the independent variables x to in the equation
F (x, u, ux , . . .) = 0
we need:
Step 1. Use x = H() to introduce v() = u(H()) , and write the inverse
change = (x) so that u(x) = v((x)) .
Step 2. Compute by the chain rule
x u(x) = v((x)) x (x)

substituting x = H()

(x u)(H()) = v() (x )(H()) ,


and higher derivatives, if necessary


x2 u(x) = x v((x)) x (x)
= 2 v((x)) (x (x))2 + v((x)) x2 (x)

substituting x = H()


2
(xx u)(H()) = 2 v() (x )(H()) + ;

Step 3. Plug the results of steps 1 and 2 into the equation to write it in the
new variables
0 = F (H(), u(H()), ux (H()) . . .)
= G(, v, v , . . .).
(4) The successful strategy for changing variables is not to memorize the general
formula above, but to repeat the three steps for individual problem. Let us
show this by examples.

124A PARTIAL DIFFERENTIAL EQUATIONS

(5) In the eiconal equation


(ux )2 + (uy )2 = 1
pass to the polar coordinates
x = r cos ,

y = r sin .

(When dealing with polar coordinates one always assumes (x, y) 6= (0, 0)
and hence r 6= 0 .)
First, changing the variables from (x, y) we obtain the new unkown


.
v(r, ) = u(x, y)
x=r cos , y=r sin

Second, to obtain the equation for v we need the inverse change


r = r(x, y),

= (x, y).

Then u(x, y) = v(r(x, y), (x, y)) and


ux

= vr rx + v x

uy

= vr ry + v y .

We need to express the right hand sides in terms of (r, ) and plug the
result into the eiconal equation.
With polar coordinates the inverse change does not have a convenient
form. The trick is to use the implicit differentiation of polar coordinates:
x r(x, y) cos (x, y),

y r(x, y) sin (x, y),

therefore
1

rx cos r sin x

ry cos r sin y

ry sin + r cos y

rx sin + r cos x

and
rx = cos
ry = sin
sin
r
cos
y =
r

x =

Combining the results discover


1 = u2x + u2y
= (vr rx + v x )2 + (vr ry + v y )2


2 

2
sin
cos
= vr cos + v
+ vr sin + v
r
r
= vr2 +

1 2
v
r2

10

DENIS A. LABUTIN

3. Derivation of equations and boundary conditions

(1) The Gauss divergence theorem is crucial for deriving PDEs. It says that if
is the boundary of the domain , and if ~n is the outer normal on
, then
ZZZ
ZZ
divF~ dvol =
F~ ~n dA.

Here dvol is the volume lement in and dA is the area element on


.
(2) Derivation of the transport equation. Let be the density of the
~ be the velocity of the transporting flow. Let
transported component, V
also f be the density of the external sources. That is the external production of the component at the point x during the time interval (t, t + t)
is
M = f (t, x) dvol(x) t.
~
The key observation: let m be the mass transported by the flow V
through the flat plate S in the direction of the unit normal ~n , |~n| = 1 ,
during the time interval t . Then is
|m| =
=
=

vol(R)
S height
~ |t cos .
S |V

~ points in the same direction as ~n ,


We must think that m > 0 if V
and m < 0 if the directions are opposite. Hence
~ ~n).
m = S t (V
The mass balance condition must always hold for the component. Therefore for any spacial volume
Z
Z
(t + t, x) dvol(x)
(t, x) dvol(x) ={production inside }

{flux through in direction ~n}


Z
=t
f dvol

Z
~ ~n dA(x)
t
(t, x) V

Z

~ dvol(x).
=t
f div (t, x)V

Divide by t 0 to discover
Z
Z

~ dvol(x).
t (t, x) dvol(x) =
f div (t, x)V

Then divide by vol() 0 to find that the equation for the density
= (t, x)
~ with the sources density f
of the component transported by the flow V
is
~ ) + f.
t = div(V

124A PARTIAL DIFFERENTIAL EQUATIONS

11

~ =
In real processes the presence of the component influences the flow V
~ (t, x, ) and the sources density f = f (t, x, ) . So the transport equation
V
is nonlinear:
~ (t, x, (t, x))) = f (t, x, ).
t + divx ((t, x)V

(3.2)

If the flow and the source are not influenced by the component the equation
becomes
~ (t, x)) = f (t, x).
t + divx ((t, x)V
Prove that (3.2) is a linear (non-homogeneous) equation for the unknown
= (t, x) . If the flow has a constant velocity
~ (t, x) C,
~
V
the transport equation is
~ = f.
t + C
If the constant velocity flow is one dimensional then the transport equation
becomes
= (t, x), t + cx = f (t, x).

(3) Derivation of the heat equation. The heat equation (also called the
heat transport equation) shares some features with the transport equation.
The unknown is the temperature T = T (t, x) . The key fact is the Fouriers
law of the heat conduction: let Q be the amount of the heat energy
transported through the flat plate S in the direction of the unit normal
~n , |~n| = 1 , during the time interval t . Then
Q = kt S gradT ~n.
Here k is the (positive) materials conductivity coefficient. Notice the
similarities with the transport. What is the meaning of the sign in
the formula?
Let f be the intensity of the external heat sources. That is the amount
of energy produced externally at the point x during the time interval
(t, t + t) is
E = f (t, x) dvol(x) t.
The heat energy E of the volume dvol is given by
E = C T dvol,
where C is the thermal capacity coefficient of the material. Using the
Fourier law we write the conservation of energy for any spacial volume
Z
Z
CT (t + t, x) dvol(x)
CT (t, x) dvol(x) ={heat production inside }

{heat flux through in direction ~n}


Z
=t
f dvol

Z
+ t
k gradT ~n dA

Z

=t
f + div k gradT dvol.

Divide by t 0 to discover
Z
Z

Ct T (t, x) dvol(x) =
f + div k gradT (t, x) dvol(x).

12

DENIS A. LABUTIN

Then divide by vol() 0 to find that the equation for the temperature
T = T (t, x, y, z)
with the sources density f is
k
div(gradT ) + f.
C
(4) Laplace operator. The operator
t T =

div(gradu) = (x2 + y2 + z2 )u
is called the Laplace operator or Laplacian and is denoted by ,
= div(grad ) = x2 + y2 + z2 .
Thus the heat equation for the unknown T = T (t, x, y, z) with the given
sources of the density f is
t T = aT + f.
In the one dimensional case the heat equation becomes
T = T (t, x),

t T = ax2 T + f,

a > 0.

Suppose that the temperature regime is stationary, that is T = T (x, y, z)


and
t T = 0.
Then the temperature distribution is described by the Laplace (or Poisson)
equation
aT = f.
Solutions u of
u = 0
are called the harmonic functions.
(5) Diffusion equation. This is the same equation as the heat equation. Let
N = N (t, x) be the concentration of the component which is being diffused
in medium. The key for the derivation is the Ficks diffusion law: let
be the amount of the component passed through the flat plate S in the
direction of the unit normal ~n , |~n| = 1 , during the time interval t
under the natural diffusion. Then
= dt S gradN ~n.
Here k is the (positive) materials diffusion coefficient. This is exactly the
same formula as in the Fouriers heat conduction law.
(6) Wave equation. We derive the equation for small vibrations of a one
dimensional horisontal string. Small means that if the length of the
string at rest is l , and if the change in length during the vibration is l ,
then
l
1.
l
This implies that the change in the tension force value
T
l
v
1.
T0
l
Therefore the tension force has the constant value T0 during the vibrations. Small also means tag the pieces of the string move only in the
vertical direction. Accepting these observations we derive the equation.

124A PARTIAL DIFFERENTIAL EQUATIONS

13

Let u(t, x) be the vertical coordinate of a piece (x, x + x) of the


string. The Newtons law of the motion projected on the vertical axis
states
m utt = F + Tvert (t, x) + Tvert (t, x + x).
Here F is an external vertical force on the piece (x, x + x) , and m =
0 x . For the tension force we have
T~ (x) = T0 ({tangent vector at x})
and
T~ (x + x) = T0 {tangent vector at x + x}.
Here the tangent vector to the graph of u is
(1, u )
p x .
1 + u2x
The smallness of vibrations allows to write
1
u
p x
= ux (1 + u2x + ) ' ux .
2
1 + u2x
Thus
Tvert (t, x) + Tvert (t, x + x) =T0 (ux (t, x + x) ux (t, x)).
Therefore the Newtons law becomes
T0 ux (t, x + x) ux (t, x)
F
utt =
+
.
0
x
0 x
Letting x 0 discover
utt = c2 uxx + f (t, x),
where f is the given external force density and c =
ator

c u = t2 c2 x2 u
is called the wave (or Dalamberts) operator.
(7) Initial vaue problem. For the equation of the type

p
T0 /0 . The oper-

t u + L(x )u = f
the initial value problem is to find u(t, x) satisfying
(
ut + Lu = f
for t > 0, x (, +)
u|t=0 = (x).
Here f (t, x) and (x) are given functions.
For the wave equation the initial value problem is to find u(t, x) satisfying

utt + uxx = f for t > 0, x (, +)


u|t=0 = (x)

ut |t=0 = (x).
Here f (t, x) , (x) , and (x) are given functions.
(8) Boundary conditions. In one dimension u = u(t, x) , t > 0 , a < x <
b , the most common boundary conditions are:
Dirichlet, functions f (t) , g(t) are given, u satsifies
u|x=a = f (t),

u|x=b = g(t)

for all t > 0;

Neumann, functions f (t) , g(t) are given, u satsifies


ux |x=a = f (t),

ux |x=b = g(t)

for all t > 0;

14

DENIS A. LABUTIN

Robin, positive constants C1 , C2 > 0 and functions f (t) , g(t) are given,
u satsifies
(ux C1 u)|x=a = f (t),

(ux + C2 u)|x=b = g(t)

for all t > 0;

Conditions at , u and all its derivatives tend to 0 as x fast enough for all
t>0.
(9) The meaning of the Dirichlet conditions is obvious (prescribing the temperature of the ends of the rod or the motion of the ends of the string).
The least obvious conditions are the Robins. For the heat propagation
they are derived by the following argument based on Newtons law of the
heat energy conduction and on Fouriers heat transport law. Suppose that
the right end of the rod, 0 < x < l , exchanges the energy with the massive
reservoir of the given temperature T (t) . The Newtons law states that
the energy flux Q from the rod into the reservoir in time t is
proportional to the temperature difference,
Q = C(u(t, l) T (t))t.
At the same time
the amount of heat transported through the right end of the rod
also outside the rod in time t by Fouriers law is
Q = ux (t, l)t.
Hence we must have ux (t, l) = C(u(t, l) T (t)) , which is equivalent to



C
C
ux + u
= T (t).

x=l
Similarly, using the Fourier law at the left end of the rod we derive
Q = ux (t, 0)
and



C
C
= T (t).
ux u

x=0
For the string vibration consider the situation when, say, the left end of
the string is attached to an elastic spring acting with the returning vertical
force
F = ku(t, 0).
Then the Newtons law for the element x projected on the vertical axis
states that
(x)utt (t, 0)

= T +F
= T0 ux (t, 0) ku(t, 0).

(Here to get the formula T = T0 ux we assumed that the vibrations are


small, hence t ux is small, and hence
sin ' tan = ux ,
where is the angle at which the end of the string slides along the vertical
axis.) Letting x 0 deduce that



k
ux u
= 0.
T0
x=0

The tension force on the right end of the string is


T = T0 ux (t, l),

124A PARTIAL DIFFERENTIAL EQUATIONS

15

and the similar argument gives





k
ux + u
= 0.
T0
x=l

4. Transport and the first order equations

(1) Main tool. The important calculus notion for the transport equation is
the directional derivative of a function. Let u be a multivariable function,
u = u(~r),
~ be a fixed vector. Consider the line through the
~r = (r1 , . . . , rn ) . Let V
~ :
point ~r0 in the direction V
~ , < t < .
~r(t) = ~r0 + tV
Restrict our function u to this line. This restriction is just a function of
one variable
u(~r(t)), < t < .
~
Now, the derivative of the function u at the point ~r0 along the vector V
is just the usual one dimensional derivative of u(~r(t)) :

~ )
du(~r0 + tV
u(~r0 )
=
.

~

dt
V
t=0

This is the definition of the derivative along the vector. Using the chain
rule we continue the computation, and obtain the important useful formula
r0 )
for u(~
:
~
V

n
X
~ )
du(~r0 + tV
u(~r0 )
=
Vj


dt
rj
j=1
t=0

~.
= u(~r0 ) V
Thus
u(~r0 )
~.
= u(~r0 ) V
~
V
If
~|=1
|V
~ is called the derivative in the direction V
~ , or
then the derivative along V
the directional derivative.
(2) Solution to the simplest transport equation. We find the general
solution u(x, y) of the equation
aux + buy = 0,
~ = (a, b) 6= 0 . This is an easy matter since
with the constant V
~ = u .
aux + buy = u V
~
V
Thus the transport equation is equivalent to


u
u is constant along any line
=0
~
~
parallel to V
V

16

DENIS A. LABUTIN

How to express the last property of u(x, y) as a formula?


One way to argue is as follows. Any line parallel to (a, b) has the
equation
bx + ay = C
with some constant C . Along this line the value of u is (some other)
constant. Therefore the level sets of functions u(x, y) and bx + ay
coincide! This is possible if and only if
u(x, y) = f (bx + ay)
with some function f of one variable. Thus
aux + buy = 0 u(x, y) = f (bx + ay)
where f is an arbitrary function of one variable.
Here is another way to derive this formula. Namely, fix a point (x0 , y0 ) . The line
through (x0 , y0 ) parallel to (a, b) has the equation
bx + ay = bx0 + ay0 .
At least one coefficients has to be nonzero, suppose a 6= 0 . Since u is constant along
this line

bx bx0 + ay0
u(x0 , y0 ) = u x,

a
x=0

bx0 + ay0
= u 0,
.
a
But u(0, z) is just a function of one variable z . Denote it by f (z) . Then

bx0 + ay0
u(x0 , y0 ) = f
a
= g(bx0 + ay0 )
with the one variable function g , g(z) = f (z/a) .

(3) As we just saw, the lines bx + ay = c which are parallel to (a, b) , play
an important role for the transport equation aux + buy = 0 . They are
called the characteristics of the equation. Above it is explained, that the
equation is equivalent to the vanishing of the derivative of u along the
characteristic.
Using the formula for the general solution we discover at once that the
solution to the initial value problem
(
ut + cux = 0 for t > 0, x (, +)
u|t=0 = (x)
is given by the wave propagating to the right
u = (x ct).
Problem 3.
Why this formula indeed gives a wave propagating to the
right? Draw pictures.
(4) Using the property of the characteristics it is easy to solve the initial value
problem for the inhomogeneous transport equation in the general form
(
ut + cux = f (t, x) for t > 0, x (, +)
u|t=0 = (x)
for x (, +) .
However, it is not very convenient for problem solving. For that it is almost
always easier to compute the solution using the superposition principle.
Since the equation is linear by the superposition principle we write
u=v+w

124A PARTIAL DIFFERENTIAL EQUATIONS

17

where
(
vt + cvx = f
v|t=0 = 0

for t > 0, x (, +)
for x (, +) .

and
(
wt + cwx = 0 for t > 0, x (, +)
w|t=0 = (x) for x (, +) .
We already know that
w(t, x) = (x ct).
It is left to find v .
The key observation for that is that any parametrization of a straight line ~
r =
~ , according to the chain rule gives
~
r0 + V
v(~
r)
~ = d v(~
~ ).
= v(~
r) V
r0 + V
~
d
V
That is, our PDE is equivalent to the ODE along the characteristic with
~ = (1, c).
V
Now, fix a point (t0 , x0 ) and consider the characteristic line
x ct = x0 ct0 x = ct + (x0 ct0 )
passing through it. Take t as the parameter along the line. Then according to our
observation, the restriction z of u to the characteristic,
z(t) = v(t, x0 ct0 + ct),
satisfies
8
< dz = f (t, x ct + ct)
0
0
dt
:z(0) = 0,

for t > 0

and the integration gives


v(t0 , x0 )

z(t0 )

z(0) +

t0

Z
0
t0

Z
=

dz(t)
dt
dt

f (t, x0 ct0 + ct) dt.

0+
0

Hence the answer is


t

Z
u(t, x) = (x ct) +

f (, x c(t ) d.
0

(5) Variable coefficients first order equation. The equation


a(x, y)ux + b(x, y)uy = 0
similarly to the constant coefficients transport equation, can be rewritten
with the vector field


~ y) = a(x, y)
A(x,
b(x, y)
as
u
= 0.
a(x, y)ux + b(x, y)uy = 0
~ y)
A(x,
Unlike the constant coefficients case, the vector field now depends on
(x, y) . Instead of the characteristic lines we need the characteristic curves
~ . These are the solutions of the system
tangent to A


x(t)

~
= A(x(t),
y(t)).
y(t)

18

DENIS A. LABUTIN

Along any such curve


d
u(x(t), y(t)) = aux + buy
dt
and therefore


u
u is constant along any curve
=0
.
~
~
tangent to A
A
How to express the last condition as a formula?
The easiest way to argue is as follows. Suppose we have the equation of
the characteristic curves not in the form


x(t)
t 7
,
y(t)
but in the form of a level curve
F (x, y) = const.
~ is tangent
The gradient F is orthogonal to the level curve, whereas A
to the level curve. Hence
~ y)
F (x, y)A(x,
(that is aFx + bFy = 0 ). At the same time for the unknown u we also
have
~ = 0,
aux + buy = 0 u A
and hence
~ y).
u(x, y)A(x,
But u is orthogonal to the level curve of u . Therefore the tangent line
to the level curve of u coincides with the tangent line to the level curve
of F . But this means that the level curves of F and u coincide! This
is equivalent to
u(x, y) = f (F (x, y))
with an arbitrary function f of one variable.
(6) Thus for the problem solving we need to have an effective way of finding
the characteristic curve in the form
F (x, y) = const.
This is easy: for any curve (x(t), y(t)) expressed as a graph y = y(x) we
have
dy
y
dy
= dt = .
dx
dx
x
dt
But the equation for the characteristics is
x =
y =

a(x, y)
.
b(x, y)

Hence

dy
b(x, y)
=
.
dx
a(x, y)
It is better to write this equation in more symmetric form
dy
dx
=
b(x, y)
a(x, y)

integration

F (x, y) = const.

This is the formula for the characteristic curves for the equation
aux + buy = 0.

124A PARTIAL DIFFERENTIAL EQUATIONS

19

Then the general solution to the equation is


(4.3)

u(x, y) = f (F (x, y)),


where f is an arbitrary function of one variable.

(7) Let us give a rigorous proofs of (4.3). The equation


(4.4)

a(x, y)ux + b(x, y)uy = 0


has the characteristic system
x
y

=
=

a(x, y)
b(x, y).

Suppose we know that F is the first integral of the system. That is, it is constant along
any solution (x(t), y(t)) . This is equivalent to
0 = t F (x(t), y(t))
= Fx x + Fy y
= Fx a + Fy b

That is, suppose that


aFx + bFy = 0.
Suppose Fy 6= 0 (if Fx 6= 0 the argument is similar). Introduce the the map m ,


x
x
7
.
F (x, y)
y
That is, make the change of variables (, ) = m(x, y) ,

=
=

x
.
F (x, y)

Geometrically m maps vertical lines into vertical


m

{x = const} { = const},
and the level curves of F into horisontal
m

{(x, y) | F (x, y) = const} { = const}.


The map m is invertible since

1
0
m0 =
,
Fx Fy

det(m0 ) = 1 Fy 6= 0.

Thus
v(, ) = u(m1 (, )),

u(x, y) = v(m(x, y)).

By this change of variables we derive


ux = v x + v x
= v 1 + v Fx ,
uy = v y + v y
= v Fy

Consequently
0 = aux + buy
= a(v + v Fx ) + bv Fy
= av + v (aFx + bFy )
= av + 0

Hence
v = 0 v(, ) = f ()
for an arbitrary function f of one variable. Actually, this is to be expected since u is
constant along the level curves of F and, as we saw above, the latter are mapped into
the horisontal lines = const . Thus v must depend only on the vertical direction .
Going back to the (x, y) -variables discover
(4.4) u(x, y) = v(m(x, y)) = f (F (x, y)).

20

DENIS A. LABUTIN

(8) Hence solving (4.4) is equivalent to finding the first integral of the characteristic system
x
y

=
=

a(x, y)
b(x, y).

In the the derivation of (4.3) we were saying that F (x, y) = C was the expression
of the solution to the system as a level curve. Let us show that (in the plane) this is
the same as saying that F is the first integral of the system. That is, finding the first
integral of the plane system of ODEs is the same as solving it.
On the one hand if F (x, y) = C is the solution formula for the system. Then every
solution (x(t), y(t)) after the substitution gives
F (x(t), y(t)) C,
so F is the first integral.
On the other hand, suppose that F is the first integral. Assume, for example,
Fy 6= 0 . Then by the implicit function theorem
F (x, y) = C y = (x)
and
0 (x) =

Fx (x, y)
,
Fy (x, y)

F (x, y) = C.

Solve the simplest equation


x = a(x, (x))
to obtain x(t) , and define
y(t) = (x(t)).
We have F (x(t), y(t)) C , and consequently
y(t)

= 0 (x(t)) x(t)

Fx (x, y)
a(x, y)
Fy (x, y)
= b(x, y).

Thus (x(t), y(t)) is the solution to the characteristic system.


(9) The map m allows to prove the solvability of the nonlinear (more precisely the semilinear) equation
a(x, y)ux + b(x, y)uy = H(x, y, u)
following the same strategy. We can prove the existence result for a suitably posed IVP.
To formulate the IVP we argue as follows. The characteristics for the nonlinear
equation are the same as above, they are given by the level curves of F . Suppose the
condition
Fy 6= 0
~ =
holds in a neighbourhood of (0, 0) . The tangent vector to a characteristic is A
(Fy , Fx ) . Therefore the characteristics are not tangential to the line
{x = 0}
in the neighborhood of (0, 0) . Consider the following problem: for a given smooth
find u satisfying
(
a(x, y)ux + b(x, y)uy = H(x, y, u) near (0, 0)
on the line {x = 0}.

u(0, y) = (y)

We now prove the existence of such u .


The (, ) -variables computation, we have done for (4.4), gives for v(, ) the
equation
av + v (aFx + bFy ) = H(x, y, v),

(x, y) = m1 (, ).

Introducing
a
(, ) = a m1 (, ),

H(,
, v) = H(m1 (, ), v),

we rewrite the equation as

a
(, )v = H(,
, v) v = G(, , v),
with
G(, , v) =

H(,
, v)
.
a
(, )

124A PARTIAL DIFFERENTIAL EQUATIONS

21

The neighborhood of (0, 0) is mapped by m onto some neighborhood N of (0, 0 ) ,


where 0 = F (0, 0) . The line {x = 0} is mapped onto = 0 . Thus m1 pulls
back our problem to finding v(, ) satisfying
(
v = G(, , v) in N

v(0, ) = ()
on the line { = 0} N,
with

()
= (y)

y=(m1 (0,))

This is an IVP for the ODE for v , for which the standard existence-uniqueness theorem
produces a smooth solution v(, ) . Hence we obtain the desired solution
u(x, y) = v(m(x, y)) = v(x, F (x, y)).

(10) Solving problems.


For the problem solving the new difficulties are
introduced by the right hand side
aux + buy = f
and by the lower order terms
aux + buy + cu = f.
The right strategy is to use the superposition principle for non homogeneous
equations and integrating factors for the lower order terms.
Problem 4. Find u = u(t, x) solving the initial value problem
2ut = ux + xu,

u(0, x) = 1.

We cannot use any formulas given above due to the term xu . First we
have to get rid of it. Compute
Z
x2
+ C.
x dx =
2
The constant is not needed for the integrating factor technique: namely,
2
multiplying the equation by ex /2 ,
2ex

/2

ut = ex

/2

ux + xex

/2

u,

find that
2

ex

/2

 2 
ut = ex /2 u ,

ex

/2

ux + xex

/2

 2 
u = ex /2 u .
x

Hence we can substitute


v = ex

/2

to find that
2

2vt = vx ,
v(0, x) = ex /2 .
This is the simplest homogeneous transport equation. We find at once that


2
t
v =f x+
,
f (x) = ex /2 .
2
Thus
u(t, x) = ex

/2

= ex

/2

t
e(x+ 2 )

xt

/2

t2

=e2+8.
Problem 5. Find u = u(t, x) solving the initial value problem
ut + (1 + t2 )ux + u = 1

u(0, x) = ex .

22

DENIS A. LABUTIN

This equation has the variable coefficients, the lower order term, and the
nontrivial right hand side. We need to make several steps to reduce it to
the homogeneous equation aut + bux = 0 .
First, the easiest way is to combine ut + u . The integrating factor can
be taken et . Multiplying deduce
et ut + et u + et (1 + t2 )ux = et ,
so that
(et u)t + (1 + t2 )(et u)x = et .
Substitute
v = et u
and find
v(0, x) = ex .

vt + (1 + t2 )vx = et ,

Next, we easily guess that et is solution to the non homogeneous equation.


Therefore
w = v et
satisfies
wt + (1 + t2 )wx = 0,

w(0, x) = ex 1.

This is a simple homogeneous first order equation. The characteristics are


Z
Z
dx
t3
dt
2
=

(1
+
t
)
dt
=
dx

t
+
+ C = x.
1
1 + t2
3
Thus the general solution is

w=f

xt

t3
3


.

The initial condition gives


f (x + 0) = ex 1.
Consequently
t3

w = ex+t+ 3 1,
then
t3

v = et + w = et + ex+t+ 3 1,
and then
t3

u = et v = 1 + ex+ 3 et .
(11) Initial value problem for nonlinear transport equations.

5. Initial value problem on the real line

(1) Initial value problem for the wave equation. For the wave equation
c u = 0
we can find the general solution
u(t, x) = f (x + ct) + g(x ct),

124A PARTIAL DIFFERENTIAL EQUATIONS

23

f, g are arbitrary functions of one variable. This is done using the reduction to the transport equations. Then the solution to the initial value
problem on the line

c u = 0, t > 0, x (, +)
u|t=0 = (x)

ut |t=0 = (x)
can be derived in the form of dAlamberts formula:
(x + ct) + (x ct)
u(t, x) =
2
Z x+tc
1
(s) ds.
+
2c xtc
(2) The important issue is to visualise in the (t, x) -plane the calculations with
the dAlamberts formula.
(3) Initial value problem for the heat equation. Solutions for the initial
value problem both for the transport and the wave equation were obtained
by the following procedure. First, we derived the formula for the general
solution. Second, we specified the general solution to satisfy the initial
conditions.
Such approach does not work for the heat equation on the line. There
does not exist a simple formula for the general solution to the heat equation.
Instead we have to treat the initial value problem directly.

6. Energy method

(1) In this course the energy method is used only to prove the uniqueness of the
boundary value problems. In fact the energy method is a truly fundamental
tool for analysis of linear and nonlinear PDEs.
(2) If
ut = c2 uxx , t > 0, x (a, b),
then the convenient energy is
Z b 2
Z b
u
u(t, x)2
E[u](t) =
dx =
dx.
2
a 2
a
If
utt = c2 uxx ,

t > 0,

x (a, b),

then the convenient energy is


Z b
Z b


1 2
1
E[u](t) =
ut + c2 u2x dx =
ut (t, x)2 + c2 ux (t, x)2 dx.
a 2
a 2
(3) The energy method computations always involve evaluating of
d
E[u](t)
dt
by differentiation under the integral sign, using the equation for replacing
some derivatives by the other, integration by parts relying on the given
boundary conditions.

24

DENIS A. LABUTIN

7. Reflection method for problems on the half line

(1) Method of reflections allows to derive the solutions to some (but not all!)
problems with the boundary conditions (some, not all!) from the solutions
to the initial value problem on (, +) . Thus the reflection method
can be summarized as the reduction
I.V.P. for Lu = 0 on x (0, +)
with boundary conditions at x = 0

I.V.P. for Lu = 0 on x (, +)
without the boundary

Below we introduce the reflections approach through some examples.


(2) Problems with homogeneous boundary conditions. Consider the
wave equation on the half-line with the zero Dirichlet boundary condition:

u =0, t > 0, x (0, +)

u|t=0 =(x), x > 0

ut |t=0 =(x), x > 0

u|x=0 =0, t > 0.


Here is the argument allowing to write down the solution to the problem
using the dAlamberts formula. Take the odd reflection of , from
(0, +) to (, 0) :

, (x) for x > 0

(x)

0 for x = 0
,
=

, (x) for x < 0.


Consider the initial value problem on (, +) with the reflected data
:
,

v = 0, t > 0, x (, +)

v|t=0 = (x)

vt |t=0 = (x).
are the odd functions. Hence by the property of the wave
The data ,
equation the solution v is odd in x for any t . But an odd function must
be equal to zero at the origin, thus
v(t, 0) = 0

for all t.

Hence v for x 0 is the solution to the problem for u since it satisfies


both the initial and the boundary conditions there. Using the dAlamberts
formula
+ t) + (x
t)
(x
u(t, x) =
2
Z
1 x+t
+
(s) ds .
2 xt
Using the visualisation of the dAlamberts formula in the (t, x) -plane and
the oddness of the data it is easy to express the answer using only ,
). The formula will be somewhat lengthy but it admits a simple
(not ,
visualisation on the (t, x) -plane.
(3) The method of reflection allows to solve some boundary value problems on
the finite interval (0, `) . For example, one can solve the Dirichlet problem

124A PARTIAL DIFFERENTIAL EQUATIONS

25

for the wave equation

u = 0, t > 0, x (0, `)

u|t=0 = (x), x (0, `)

ut |t=0 = (x),
u|x=0 = u|x=` = 0,

x (0, `)
t > 0.

The solution will have an incredibly lengthy formal description, but it has
a clean visualisation on the (t, x) -plane.
(4) The homogeneous Neumann boundary conditions are also manageable by
the reflection method. For example, consider the heat equation on the
half-line

ut uxx = 0, t > 0, x (0, +)


u|t=0 = (x), x > 0

ux |x=0 = 0, t > 0.
Take the even reflection of from (0, +) to (, 0) :
(
(x) for x 0

(x) =
(x) for x 0.
Consider the initial value problem on (, +) with the reflected datum
:
(
vt vxx = 0, t > 0, x (, +)

v|t=0 = (x)
.
The datum is an even function. Hence by the property of the heat
equation the solution v is even in x for any t . But an even (smooth)
function must have zero derivative at the origin, thus
vx (t, 0) = 0

for all t.

Hence v is the solution to the problem for u since satisfies both the initial
and the boundary conditions there. Using the heat kernel
Z +
dy .
u(t, x) =
S(t, x y)(y)

Using the evenness property of the heat kernel and the evenness of it is
easy to express the answer using alone (instead of ).
(5) Reflections for the non-homogeneous boundary conditions.
Method of reflections allows also to reduce some problems with the nonhomogeneous boundary conditions to the IVP on (, +) . For example,
consider the general Dirichlet problem for the wave equation on (0, +) :

utt c uxx = 0, t > 0, x (0, +)

u|t=0 = (x), x > 0

ut |t=0 = (x), x > 0

u|
= h(t), t > 0.
x=0

The Dirichlet boundary condition is now given by h(t) , h 6= 0 . Here is


the reduction argument. Recall that
c f (x ct) = 0
for an arbitrary function f of one variable. Choose f () = h ( /c) , so
that

x
.
f (x ct) = h t
c

26

DENIS A. LABUTIN

Define


x
,
v(t, x) = u(t, x) h t
c
so that v(t, 0) = h(t) h(t) = 0 for all t > 0 . Then v solves

vtt c vxx = 0, t > 0, x (0, +)

v|t=0 = 1 (x), x > 0

vt |t=0 = 1 (x), x > 0

v|
= 0, t > 0,
x=0

with 1 (x) = (x) h(x/c) , 1 (x) = (x) h0 (x/c) . This is the


problem with the zero Dirichlet condition. We solved it above using the
odd reflection. Hence the answer is
for x > ct :
1 (x + ct) + 1 (x ct)
2
Z x+tc
1
+
1 (s) ds
2c xtc


x
+h t
c
Z x+tc
(x + ct) + (x ct)
1
=
+
(s) ds
2
2c xtc
h((x + ct)/c) + h((x ct)/c)

2
Z x+tc
1

h0 (s/c) ds
2c xtc

x
+h t
c
Z x+tc
1
(x + ct) + (x ct)
+
(s) ds
=
2
2c xtc
h((x + ct)/c) + h((x ct)/c)

2
Z
1 (x+tc)/c 0
+
h (r) dr
2 (xtc)/c

x
+h t
c
Z x+tc
(x + ct) + (x ct)
1
=
+
(s) ds
2
2c xtc
h((x ct)/c)

x
+h t
c
Z x+tc
(x + ct) + (x ct)
1
=
+
(s) ds;
2
2c xtc

u(t, x) =

for ct x > 0 using the similar computations:


Z x+tc
(x + ct) + (x ct)
1
u(t, x) =
+
(s) ds
2
2c xtc

x
+h t
.
c

124A PARTIAL DIFFERENTIAL EQUATIONS

27

The meaning of the solutions is simple. For x > ct the wave launched by
the motion h(t) at the origin does not reach the point x , and the solution
is given by the usual dAlamberts formula. For ct > x the solution is the
superposition of the free wave coming from the dAlambert and the wave
launched from the origin by the prescribed motion h(t) .

8. Fundamental solution and Duhamels principle

(1) Delta function. Delta function is not a function! The meaning will
be assigned not to itself, but to expressions involving it. Therefore for
us formulas as
Z +
(x y) dy = 1,
(x) = 0 x 6= 0, . . .

will have a meaning, but itself will not. The formulas involving function (such as above) express the limiting results of the corresponding
processes.
(2) Intuitively -function expresses such physical notions as the density of the
point electrostatic charge, the energy of the point heat source, and similar.
This is done using the limit of -shaped functions in the same way as in
physics one deals with the point-concentrated objects as limits of smeared
ones, as smearing gets smaller and smaller.
A -shaped function (x) is the function satisfying:
Z +
(a)
(x) dx = 1 ;

(8.5)

(b) (x) 0 for all x ;


(c) (x) = 0 for all x with |x| > (or lim0 (x) = 0 for every
x 6= 0 .)
The following two examples of -shaped sequences are of fundamental
importance for this course.
First, the characteristic function approximation of is

1
0 , x
/ (0, )
(x) =
1/ , x [0, ],

where

0 , x
/ (0, )
1 , x [0, ].
Second, the heat kernel approximation is

x2
1
1
G(x/ ) =
e 4 .

4
Now, -function is the limit of a -shaped sequence as 0 . Thus
x2
1
1
(x), (x),
e 4 (x), 0.

4
The limit in (8.6) is understood in the following sense: the validity of any
given formula
F() = f
means that after the substitution F( ) and letting 0 we find
(x) =

(8.6)

lim F( ) = f.

28

DENIS A. LABUTIN

(3) Let us illustrate the technique of dealing with .


Problem 6. Let us show that
Z +
(x) dx = 1.

Indeed, take any -shaped sequence (for example / ). We have


Z +
(x) dx = 1 > 0,

hence trivially
Z

lim


(x) dx = 1.

Thus according to our definition we may write


Z +
(x) dx = 1.

Problem 7. Show that


(x) = 0 x 6= 0.
Problem 8. Show that for any continuous function f
Z +
f (x) (x) dx = f (0).

To prove this formula it is convenient to use the -shaped sequence


1 (x) .
Problem 9. Fix x0 . For the shifted (x x0 ) show that
(x x0 ) = 0 x 6= x0 ,
Z +
(x x0 ) dx = 1,

and
Z

f (x) (x x0 ) dx = f (x0 ).

(8.7)

Problem 10. Prove that for a continuous function f


Z +
f (y) (x0 y) dy = f (x0 ).

The formula can be approximately rewritten as


Z +
f (x0 ) '
f (y) (x0 y) dy

X
'
(x0 yj ) f (yj ) yj
j

'

(x0 yj ) f (yj ) yj

X
j



f (yj )yj (x yj )

.
x=x0

This form emphasizes the important meaning of (8.7). Namely, an arbitrary


force (or signal, or field, or ) f (x) at a point x is approximately equal to

124A PARTIAL DIFFERENTIAL EQUATIONS

(8.8)

29

the value at x of the sum of the point-impulse forces (or signals, or fields,
or ) shifted to yj :
X
f (x) '
f (yj ) (yj+1 yj ) (x yj ).
j

To derive (8.7) just use


Z
+
f (y) (x0 y) dy =
f (x0 z) (z) (dz)

since x0 y = z, dy = dz

+
+

f (x0 z) (z) dz


f (x0 z) .
=

z=0

Problem 11. For the Heaviside step-function



0, x < 0
(x) = H(x) =
1, x 0
show that
(x)0 = (x).
For that smear the step-function (x) on scale . Then show that
(x) = (x)0 is a -shaped sequence.
(4) Fundamental solution. Consider the initial value problem for a very
general evolution equation with the unknown u(t, x)
(
t u = Lu t > 0, x (, +)
u|t=0 = (x).
Think that
L = L(x )
is, say, a constant coefficient operator involving only x , x2 , . . . . Many
important problems from sciences have such form. For example, we have
already encountered the heat diffusion L = k 2 x2 , and transport L =
cx .
(5) The fundamental solution (or propagator) is the evolution of the initial
point-impulse. That is, the solution of
(
t u = Lu, t > 0, x (, +)
u|t=0 = (x)
This means that
t K(t, x) L(x )K(t, x) = 0,

t > 0, x (, +)

and K(t, x) is -shaped in the same sense as in (8.6) with relaced by


t . That is
K(t, x) (x),
t 0.
Hence the fundamental solution shows how the equation smears for t > 0
the initial (at time t = 0 ) datum .
As we will see below, knowing the propagator for L is the key for solving the initial value problem with arbitrary data. This explains the name
fundamental. We will denote the propagator by K(t, x) or KL (t, x) .

30

DENIS A. LABUTIN

(6) For example, the Gaussian (heat kernel)


G(t, x) =

1 x2
e 4t
4t

satisfies

t x2 G(t, x) = 0,

t > 0.

Moreover, it forms a -shaped sequence,


G(t, x) (x),

t 0.

Hence the Gaussian is the fundamental solution to the heat equation.


(7) Somewhat unusual example is the fundamental solution to the transport equation
t u + cx u = 0.
One easily makes the formal computations for K(t, x) = (x ct) :

x (x ct) = 0
1,
xct

(c),
t (x ct) = 0
xct

so that
(t + cx )(x ct) = c + c = 0.
We also have (x ct) (x) as t 0 (without any smearing). Hence (x ct) is
the fundamental solution of the transport equation.
This answer has a clear physical meaning. It says that the initial point-mass with
the density is transported by the one dimensional flow with the velocity c along
the x -access without smearing or spreading. That is indeed what the perfect transport
(without difusion) does!

(8) We verified that two suggested functions are the fundamental solutions of
the heat and transport equation. There is a general method based on the
Fourier transform for the computation of the fundamental solution for a
given equation from the scratch. It will be studied in the second quarter.
(9) Fundamental solution and the formula for IVP. Let us show that the
superposition principle implies that the solution of
(
t u = Lu, t > 0, x (, +)
u|t=0 = (x)
is given by
Z
(8.9)

K(t, x y)(y) dy.

u(t, x) =

Hence, knowing the fundamental solution K(t, x) allows us to calculate


the solution with any initial data by the simple integration!
Verifying that (8.9) solves the IVP is easy. Indeed, firstly for t > 0 and
x (, +) we have
Z +
 Z +
(t L(x ))
K(t, x y)(y) dy =
(t L(x ))K(t, x y)(y) dy


Z +

=
(s K(s, z) L(z )K(s, z))

(s,z)=(t,xy)

0(y) dy

= 0.

(y) dy

124A PARTIAL DIFFERENTIAL EQUATIONS

Secondly, for every x (, +)


 Z
Z +
K(t, x y)(y) dy =
lim
t0

31

(x y)(y) dy

= (x).
Instead of verifying the formula given by somebody, one can derive it as
shown below.
(10) The basic building block for the derivation of (8.9) is the following simple observation.
Consider the IVP with the shifted :
C(x a).
That is

t > 0, x (, +)

t u = Lu,

u|t=0 = C(x a).


Use the linearity to prove that the solution is
u = CK(t, x a),
where K is the fundamental solution. This is the formula we will need.
(11) Now fix a small > 0 , discretize the x variable
j = 0, 1, 2 . . . ,

xj = j,
and approximate our datum
(x) '

(xj ) (x xj ).

The key observation is to write


1
(x xj )

with the characteristic function approximation of from (8.5). Therefore


(xj ) (x xj ) = (xj )

(xj )

1
(x xj ) ' (xj ) (x xj ).

The summation then gives


(x) '

(8.10)

(xj ) (x xj ),

which is nothing but the physical approximation of a general charge distribution by a


sum of point charges. That is we approximated by a sum of discrete impulses located
at points xj .
Notice that we could skip the details of the discretization and at once deduce (8.10)
from (8.8)!
(12) Next, by the linearity
X
u(t, x) '
uj (t, x),
j

where uj solves
(

t uj = Luj ,

t > 0, x (, +)

uj |t=0 = (xj ) (x xj ).
By our building block solution we already know that
uj (t, x) = (xj )K(t, x xj ).
Consequently
u(t, x) '

uj (t, x)

(xj )K(t, x xj )

K(t, x xj )(xj )(xj+1 xj ).

Finally, notice that the last sum approximates an integral over x . Letting the discretisation
xj+1 xj = 0

32

DENIS A. LABUTIN

we discover
0
u(t, x) = lim @

1
X

K(t, x xj )(xj )(xj+1 xj )A

K(t, x y)(y) dy,

which is desired (8.9).

(13) To make the formulae shorter, it is convenient to introduce the solution


operator etL acting on the functions F on R . Define
(etL F )(x) =

K(t, x y)F (y) dy.

In the case L is a number etL is the usual exponential. In these notations


we have
(

t > 0, x R

t u = Lu,
u|t=0 = (x)

u(t, x) = etL (x)

(14) The propagator approach gives the same formula for any constant coefficients linear operator L . Notice, however, that the wave equation does
not fit immediately in this form we have t2 instead of t in the initial
value problem for the wave equation. Later we will show how to adapt the
arguments for the wave equation.
(15) Duhamels principle. Knowing the propagator also allows us to solve
the non-homogeneous problem
(
(8.11)

t u =Lu + f (t, x),

t > 0, x R

u|t=0 =0
via the Duhamels formula:
Z

e(ts)L f (s) ds
Z t Z +
=
K(t s, x y)f (s, y) dy ds.

u(t, x) =

Again, once the formula is given, its verification is not hard. Firstly,
Z
u(0, x) =


dt = 0,

therefore the initial condition holds. Secondly, equality


Z

L(x )u(t, x) =
0

t Z +

=
0

L(x )K(t s, x y)f (s, y) dy ds




L(z )K(r, z)
f (s, y) dy ds
(r,z)=(ts,xy)

124A PARTIAL DIFFERENTIAL EQUATIONS

33

together with
+

Z
t u(t, x) =

Z t

+
0
+

Z
=



K(t s, x y)f (s, y) dy
s=t
Z +
t K(t s, x y)f (s, y) dy ds

(x y)f (t, y) dy

Z t Z +

r K(r, z)
+

f (s, y) dy ds

(r,z)=(ts,xy)

imply
Z

(t L)u =

(x y)f (t, y) dy

+
0


r K(r, z) L(z )K(r, z)

f (s, y) dy ds
(r,z)=(ts,xy)

=f (t, x) +

0 f (s, y) dy ds

=f (t, x).
(16) We now derive Duhamels formula on the physical level of rigor. For that we shall use
the following basic building block problem. Consider the delayed IVP:
(
t u = Lu, t > s, x (, +)
u|t=s = (x).
Using the invariance in time, one can easily see that the value of the solution to the
delayed problem is
Z +
u(t, x) =
K(t s, x y)(y) dy

where K is the fundamental solution. This is the formula we will need.


(17) Fix T > 0 . To compute u(T, x) solving (8.11) we first split the time interval [0, T ]
into N segments of the length t = T /N ,
0 = t0 < t1 = t
< t2 = 2t = t1 + t

< tj = jt = tj1 + t

< tN = T.
Next, in (8.11) discretise f in time by splitting it into
f (t, x) '

N
1
X

fj (t, x),

fj (t, x) = f (tj+1 , x)(tj ,tj +t) (t).

j=0

By the linearity
u(T, x) '

N
1
X

uj (T, x),

j=0

where uj (t, x) solves


(
t uj = Luj + fj ,

t > 0, x (, +)

uj |t=0 = 0.
The point here is that the right hand side fj (t, x) is identical zero if 0 < t < tj and
tj + t < t < T . Only on the short time interval (tj , tj + t) it acts on the equation.
Let us evaluate uj (T, x) for every j . By (8.9) we have
uj (t, x) = 0

for

0 < t tj .

34

DENIS A. LABUTIN

Ignoring the tiny time interval (tj , tj+1 ) of length t we use the first order approximation to evaluate for t = tj+1
uj (tj+1 , x) ' uj (tj , x) + t t uj (tj , x)

= 0 + t L(x )uj + fj
(tj ,x)

= t 0 + f (tj+1 , x) .
Consequently for tj+1 < t < T our uj solves
(
t v = Lv, tj+1 < t < T, x (, +)
v|t=tj+1 = f (tj+1 , x)t.
But this is just the delayed IVP. Hence, as we shown above
Z +
uj (T, x) =
K(T tj+1 , x y)f (tj+1 , y)t dy.

Therefore
u(T, x) '

N
1 Z +
X
j=0

K(T tj+1 , x y)f (tj+1 , y) dy t.

Recognising the Riemann integral sum here we discover letting t 0 that


Z T Z +
u(T, x) =
K(T s, x y)f (s, y) dy ds.
0

(18) Notice that the Duhamels formula gives the following statement for the
heat equation. If (x) is odd (even) and for every t the function f (t, x)
is odd (even) in x , then the solution of
(
t u = Lu + f (t, x), t > 0, x (, +)
u|t=0 = (x),
u = u(t, x) is odd (even) in x for any fixed t .
This follows at once from the representation
Z +
Z t Z +
u(t, x) =
K(t, x y)(y) dy +
K(t s, x y)f (s, y) dy ds

with
 
z
1
K(t, z) = G ,
t
t
and the evenness of the Gaussian
x2
1
G(x) = e 4 .
4

9. Boundary value problems and separation of variables

(1) Separation of variables method for the heat and wave equations with either
Dirichlet or Neumann boundary conditions is familiar from calculus. Actually the method works for more general equations and boundary conditions.
(2) Suppose f is a given function on a finite segment [a, b] .
Boundary conditions
1 f (a) + 1 f (b) + 1 f 0 (a) + 1 f 0 (b) =
0

2 f (a) + 2 f (b) + 2 f (a) + 2 f (b) =

0
0

124A PARTIAL DIFFERENTIAL EQUATIONS

35

are called symmetric if for any functions f , g satisfying


them one has
x=b

= 0.
f 0 (x)g(x) f (x)g 0 (x)
x=a

This definition is not very inspiring. The true meaning of the symmetry
is the symmetry of the operator d2 /dx2 with respect to the L2 inner
product. Indeed,
the boundary conditions are symmetric if and only for any smooth
functions f , g satisfying them one has



 2
d2 g
d f
,
g
=
f,
dx2
dx2 L2
L2
(3) The theoretical basis for the separation of variables method is provided by
the following theorem. Its most important part (iii) states that the separation of variables method always works provided the boundary conditions
are symmetric.
Theorem. Consider the eigenvalue problem
d2
f = f on (a, b)
dx2
with some symmetric boundary conditions. Then:
(i) all eigenvalues are real and form an infinite discrete sequence
1 2 ,
n + as n ;
(ii) if n 6= m then the corresponding eigenfunctions are L2 -orthogonal
(fn , fm )L2 = 0 ;
(iii) any L2 can be expanded as
=

An f n

on

[a, b]

n=1

with the coefficients given by


An =

(, fn )L2
.
(fn , fn )L2

The convergence of the series in the theorem is in the L2 sense. The


theorem unites several theorems in Strauss.
(4) We know from the calculus that any can be expanded either into the
full Fourier series, or into the cos -half range Fourier series, or into the
sin -half range series. The theorem provides us with the expansion with
respect to the general system {fn } of the eigenfunctions not covered by
the three calculus cases.
For example, the mixed boundary conditions
u|x=0 = 0,

ux |x=l = 0

on (0, l) lead through the separation to the eigenvalue problem


f 00 = f,

f (0) = f 0 (l) = 0.

The eigenvalues turn out to be



2
(n + 12 )
n =
,
l

n = 0, 1, 2, . . . ,

36

DENIS A. LABUTIN

and the eigenfunctions



(n + 12 )x
fn (x) = sin
.
l
This is neither the full Fourier nor a half range Fourier system. Hence the
calculus expansions do not work in this case. However, the theorem tells
that any initial datum on [0, l] can be expanded with respect to {fn } .
Indeed, by the theorem



X
(n + 21 )x
, x [0, l],
(x) =
An sin
l
n=0


with the coefficients computed by the formula




Z l
(n + 21 )x
(x) sin
dx
l
0
.
An = Z l




(n + 21 )x
(n + 12 )x
sin
sin
dx
l
l
0

You might also like