Sei sulla pagina 1di 10

1

0 Preliminary ideas on PDEs


These notes aim to give you a general idea of what a PDE is and introduce some
of the commonly used terminology. They may be read at your leisure.
Partial dierential equations (PDEs) are fundamental to pure mathematics, to
classical and quantum mechanics, to quantum eld (and string) theory, and gener-
ally throughout physics and engineering. For example, in pure mathematics PDEs
are a basic tool of geometric analysis in particular, in the recent proof of the
Poincare Conjecture (which remained unsolved by other methods for over 100
years).
In practise there are two broad steps to utilizing a PDE in one of these contexts.
First, deduce a PDE which eectively describes the geometric / physical structure
in question. Secondly, try to solve the PDE or, even if that is not possible, try to
pick out interesting mathematical properties any solution must possess.
The rst of these steps is the relatively easy part, setting up an expression in-
volving derivatives which describes, for example, how the geometric or physical
structure evolves through time in the appendix to these notes there is a brief
derivation of the principal PDEs which we will be studying, and you can read more
about the background ideas in text books or on the internet.
For the second step, nding an exact solution is usually impossible, though there
are always numerical methods which provide approximate solutions. But, even if it
cannot be computed exactly, it may be possible to determine analytically numbers
or functions which arise in the solution and which dene invariants that characterize
the physical or geometric structure in question. There are, though, certain special
classes of PDEs which can be solved exactly and which describe some important
basic geometric and physical processes, and in the following lectures we will be
looking into some particular cases where that occurs.
0.1 So what is a PDE?
A partial dierential equation (PDE) is an equation which relates a function
u : X R
1
(or C
1
) on some region X R
n
depending on n independent variables x
1
, . . . , x
n
to a nite number of its partial
derivatives. The objective is to deduce the form of u from the PDE. The order of
the highest occurring non-zero partial derivative in the PDE is called the order of
the PDE.
2
In order to compute u it is in general necessary to make assumptions about the ge-
ometric shape of the region X and the values that u and its derivatives can take on
the boundary of X a PDE augmented by such data is called a boundary problem
(or, depending on the context, an initial value problem).
0.1.1 Some examples
Laplaces equation is a 2nd order PDE dened in every dimension :
dim 1: f : R
1
R
1
, f = f(x),
d
2
f
dx
2
= 0, or u

= 0. (0.1)
A more rened dierential equation is the boundary problem for f : [0, 1] R
1
d
2
f
dx
2
= 0 subject to u(0) = u(1) = 0. (Dirichlet bvp). (0.2)
dim 2: u : R
2
R
1
, u = u(x, y),

2
u
x
2
+

2
u
y
2
= 0, or u
xx
+ u
yy
= 0. (0.3)
A corresponding Dirichlet bvp is for u : X R
1
with X = {(x, y) | x
2
+ y
2
1}

2
u
x
2
+

2
u
y
2
= 0 subject to u(x, y) = 0 for (x, y) S
1
. (0.4)
dim 3: w : R
3
R
1
, w = w(x, y, z),

2
w
x
2
+

2
w
y
2
+

2
w
z
2
= 0, or w
xx
+ w
yy
+ w
zz
= 0, (0.5)
and so on.
So for u : R
2
R
1
, u = u(x, y), we are using the shorthand notation
u
x
=
u
x
, u
y
=
u
y
, u
xx
=

2
u
x
2
,
u
xy
=

2
u
xy
, u
yx
=

2
u
yx
, u
yy
=

2
u
y
2
,
u
xxy
=

3
u

2
xy
, u
xxx
=

3
u

3
x
, . . .
3
and so on.
Recall, here, that mixed partial derivatives are equal so that
u
xy
= u
yx
, u
xxy
= u
xyx
= u
yxx
, . . .
i.e. the order of dierentiation does not matter it does when you work on
curved space (such as a sphere), in fact that is exactly what curvature is in the
mathematical sense.
We may also use the alternative shorthand

x
=
u
x
,
y
=
u
y
,
2
xx
=

2
u
x
2
,

2
xy
=

2
u
xy
,
2
yx
=

2
u
yx
,
2
yy
=

2
u
y
2
,
and so on, and one has
2
xy
=
2
yx
etc.
Here are some more PDEs in dim 2.
u
x
+ u
y
= 0 Transport equation (order 1), (0.6)
u
xx
+ u
yy
= 0 Laplace equation (elliptic), (0.7)
u
t
= iu
xx
Schrodinger eqn (parabolic), (0.8)
u
t
= u
xx
Heat equation (parabolic), (0.9)
u
tt
= u
xx
Wave equation (hyperbolic). (0.10)
PDEs in dimension 2 will, in fact, be essentially the only case we will look into in
this course, though most of the methods apply in a fairly obvious way to all di-
mensions. We will also review some facts about dierential equations in dimension
one to guide us in the higher dimensional case.
0.1.2 Beginning to think about what it means to solve a PDE: some
elementary solutions
Its not hard to deduce some particular solutions of these PDEs just by inspection,
or by some elementary ad hoc emtod, noting that these will not necessarily be
solutions for an associated boundary value problem (or, initial value problem);
that is, restricting to a subregion X R
2
and requiring that any solution is equal
to a given function on the boundary of X greatly restricts the class of possible
4
solutions, indeed if we impose too many boundary conditions then there will be
no solutions at all.
Consider, rst, the 1-dimensional case then a PDE is called an ordinary dier-
ential equation (ODE) since there is then only one variable to dierentiate. One
version of the Fundamental Theorem of Calculus tells us that if g : R R is
continuous then
d
dx
__
x
g(t) dt
_
= g(x). (0.11)
That is,
f(x) :=
_
x
g(t) dt (0.12)
is (at least locally) a dierentiable function with continuous derivative which solves
the ODE
df
dx
= g(x). (0.13)
This solution is only unique up to the addition of an arbitrary constant. That is,
if we specify a lower limit x
0
< x for the integral then
f(x) =
_
x
x
0
g(t) dt + C and f(x
0
) = C. (0.14)
In language from linear methods 4ccm113a this says that
D :=
d
dx
: C
1
[a, b] C
0
[a, b] (0.15)
is an onto (surjective) linear operator with kernel
Ker(D) := {f C
1
[a, b] | Df = 0}

= R (0.16)
where the isomorphism refers to the constant valued functions; here
C
0
[a, b] = {f : [a, b] R | f is continuous}
while
C
1
[a, b] = {f : [a, b] R | f

exists and is continuous}.


We can also replace this by innite internals if we restrict to integrable functions
for the moment we restrict to nite intervals to avoid worrying about those issues.
If we were considering functions with values in C then we would replace the right-
hand side of (0.16) by

= C. Saying that D is a linear operator means that for
f, h C
1
[a, b]
D(f + h) = Df + Dh any , R,
5
which we know is true from the denition of the derivative of a function (check
it!).
At any rate, to determine a unique solution to (0.13) we have to specify an initial
value f(x
0
) in order to x (0.14). If we take x
0
= a, or x
0
= b, then we may refer
to (0.14) as a boundary value problem, for short bvp, (or may be an initial value
problem, for short ivp,): for example, the well-posed bvp
df
dx
= g(x), f(0) = 0, on [0, 1], (0.17)
has the unique solution f(x) =
_
x
0
g(t) dt. On the other hand, the bvp
df
dx
= g(x), f(0) = 0, f(1) = 0, on [0, 1], (0.18)
can only be solved for those g for which
_
1
0
g(t) dt = 0.
A similar analysis can be given for the 2nd order ODE
d
2
f
dx
2
= g(x). (0.19)
Again, in language from linear methods 4ccm113a we now have
D
2
:=
d
2
dx
2
: C
2
[a, b] C
0
[a, b] (0.20)
where
C
2
[a, b] = {f : [a, b] R | f exists and is continuous}.
One way to dispense with worrying about using the correct space of functions is
to just stick to the more restricted space of smooth functions
C

[a, b] := {f : [a, b] R | f
(k)
exists and is continuous for each k N}.
Then D
2
and D are linear maps C

[a, b] C

[a, b], as indeed is any ordinary


dierential operator of order m
D := c
m
(x)
d
m
dx
m
+ c
m1
(x)
d
m1
dx
m1
+ . . . + c
1
(x)
d
dx
+ c
0
(x) : C

[a, b] C

[a, b],
f c
m
(x)
d
m
f
dx
m
+ c
m1
(x)
d
m1
f
dx
m1
+ . . . + c
1
(x)
df
dx
+ c
0
(x)f(x), (0.21)
where the coecient functions will in general be smooth for our considerations
here. But we could, for example, also consider D as a linear operator
D : C
m
[a, b] C
0
[a, b]. (0.22)
6
Again, saying that D is a linear operator (and it is!) means that for f, h C
m
[a, b]
D(f + h) = Df + Dh any , R.
On the other hand, the operator
f
_
df
dx
_
2
is non-linear (why?) we will only be considering linear operators and this makes
it much easier to build solutions to bvps (see the next section).
Anyway, lets get back to D
2
in (0.20). This dierential operator is surjective (like
for D) but now it has a 2-dimensional kernel space of solutions
Ker(D
2
) := {f C
2
[a, b] | D
2
f = 0}

= R
2
(0.23)
where, since f

= 0 i f(x) = ax + b some constants a and b, the isomorphism is


given by
ax + b
_
a
b
_
.
(Note that adding two such functions corresponds to adding two such vector to-
gether.)
Saying that D
2
is surjective is the statement that
for any g C
0
[a, b] f C
2
[a, b] with D
2
f = g.
This is easy to see just by integrating, except that this time we end up with a
double integral
f(x) =
_
x
s=0
_
s
t=0
g(t) dt + ax + b. (0.24)
Indeed, our formulae for D tell us that if (0.19) holds then f

(x) =
_
x
t=0
g(t) dt +b,
then apply the same formulae again to get (0.24). We also know what the constants
a and b are in terms of f, they are
a = f

(0) and b = f(0). (0.25)


(Of course, since we do not know what f is, they are still undetermined.)
It might not surprise you to know, then, that the general ODE in (0.1.2) is surjec-
tive linear operator with m-dimensional kernel if c
m
(x) = 0 all x
Ker(D) := {f C

[a, b] | Df = 0}

= R
m
, (0.26)
or, essentially equivalently,
7
Proposition 0.1 If we specify the values
f
(m1)
(x
0
), f
(m2)
(x
0
), . . . , f

(x
0
), f(x
0
) (0.27)
at one point
x
0
[a, b]
then there is a unique solution to the resulting ivp (or bvp)
Df = g. (0.28)
For example, specifying the boundary values f
(m1)
(0) = 1, . . . f(0) = 1 for f
species a unique solution. (The form of the elements of the kernel is not just a
polynomial unless D = d
m
/dx
m
more on this later.)
We need to become comfortable with some other technology before we can see
why Proposition 0.1 holds, but we should be able to explain it within a couple of
lectures.
Finally, notice that if f is a solution to (0.28), then any other solution h has the
form
h(x) = f(x) + k(x) for some k Ker(D). (0.29)
This is because D(f h) = 0.
0.1.3 Constant coecient, homogeneous, and linear PDEs
When working in dimension 2 we will be interested almost exclusively in constant
coecient homogeneous linear PDEs (the theory of PDEs on higher dimensional
spaces is similar but we will not consider that here). That is, if in a given PDE the
coecient of u and that of each of its partial derivatives is a constant (independent
of the variables x, y), it is said to be a constant coecient PDE it is much
more complicated to deal with the case where the coecients are general functions
of (x, y) (see third year courses!), but fortunately within the class of constant
coecient operators there are some important and interesting PDES; in fact, the
case of non-constant coecients really arises when one looks to solve the same
PDEs on curved space, then the curvature of space requires that the coecients
depend on the curvature (or, really, on the metric = rst fundamental form).
A general linear rst-order partial dierential operator (PDO) has the form
P = b
1
(x, y)

x
+ b
2
(x, y)

y
+ c(x, y) : C
1
(R
2
) C
0
(R
2
), (0.30)
8
acting on functions in C
1
(R
2
) by
u = u(x, y) (Pu)(x, y) = b
1
(x, y)
u
x
+ b
2
(x, y)
u
y
+ c(x, y)u(x, y)
= b
1
(x, y) u
x
+ b
2
(x, y) u
y
+ c(x, y)u(x, y).
Here, for a region X R
2
C
1
(X) = {u : X C | u
x
and u
y
exist and are continuous in (x, y)},
while C
1
(X) is the space of continuous functions on X. So X could, for example,
be all of R
2
, or a rectangle (and its interior), or a disc, or an innite strip, and so
forth.
For example, if
P = sin(xy)

x
+ (3x + y)

y
+ e
x
2
+y
2
and u(x, y) = xy
then
(Pu)(x, y) = y sin(xy) + 3x
2
+ xy + e
x
2
+y
2
xy.
The meaning of linear is the same as before, that
P(u
1
+ u
2
) = Pu
1
+ Pu
2
, any u
1
, u
2
C
1
(R
2
) (0.31)
and any , C.
P is constant coecient if b
1
(x, y) = b
1
C, b
2
(x, y) = b
2
C, c(x, y) = c C,
are each constant, independent of x and y; that is,
P = b
1

x
+ b
2

y
+ c. (0.32)
Then a linear constant coecient partial dierential equation PDE of rst order
on X R
2
is an equality of the form for v C
0
(X)
Pu = v with u C
1
(X). (0.33)
That is, for a given v = v(x, y) the problem is to nd which u satisfy this equation.
If we specify that u is equal to a certain function on the boundary of X, then this
is called a PDE boundary problem. We will see lots of examples as the course
proceeds, but do notice here that now the boundary of X will in general be a
curve (the boundary of a disc, for example, is a circle) and so now the boundary
condition will be a function on that curve in the 1-dimensional ODE case the
boundary can only be a point, or set of points, so the function boundary values are
9
discrete numbers, this is one reason why the higher dimensional situation is more
complex (and more interesting).
The homogeneous version of this equation means the case for v = 0, so
Pu = 0, (0.34)
that is, a homogeneous linear constant coecient partial dierential equation PDE
of rst order on X R
2
is an equality of the form
b
1
u
x
(x, y) + b
2
u
y
(x, y) + cu(x, y) = 0, b
1
, b
2
, c C. (0.35)
In other words, solving the homogeneous equation is nding functions u = u(x, y)
which are in the kernel
Ker(P) = {u C
1
(X) | Pu = 0}.
This (obviously!) is a vector subspace of the vector space C
1
(X).
For example, when
P =

x


y
then
u(x, y) = cos(x y) solves (0.34).
In fact,
u(x, y) = f(x y) solves (0.34)
for any dierentiable function f : R
1
R
1
! (There is more on this PDE in the
next section).
The above discussion extends to 2nd-order PDEs in 2-dimensions essentially with-
out change except that now we allow partial derivatives up to order 2. So a ho-
mogeneous linear constant coecient partial dierential equation PDE of second
order on X R
2
is an equality of the form means PDEs of the form
Lu = 0, u : R
2
C, (x, y) u(x, y), (0.36)
with
L = a
11

2
x
2
+ 2a
01

2
xy
+ a
22

2
x
2
. .
order 2 part
+b
1

x
+ b
2

y
. .
order 1 part
+ c
..
order 0 part
(0.37)
for some constants a
ij
, b
k
, c. Then the PDE
Lu = v. (0.38)
10
with
C
2
(X) = {u : X C | u
xx
, u
xy
, u
yy
, u
x
, u
y
exist and are continuous in (x, y) X}.
We can avoid discussion of which function space here by restricting matters to the
subspace C

(X) of smooth functions on X; that is, functions all of whose partial


derivatives of all orders exist (such as polynomials and convergent power series).
So the meaning of (0.38) is that for u = (x, y)
a
11

2
u
x
2
+ 2a
01

2
u
xy
+ a
22

2
u
y
2
+ b
1
u
x
+ b
2
u
y
+ cu(x, y) = v(x, y), (0.39)
or, in alternative notation,
a
11
u
xx
+ 2a
01
u
xy
+ a
22
u
yy
+ b
1
u
x
+ b
2
u
y
+ cu(x, y) = v(x, y), (0.40)
The PDE is said to be homogeneous if v = 0 in (0.38); that is,
Lu = 0 (0.41)
so
a
11
u
xx
+ 2a
01
u
xy
+ a
22
u
yy
+ b
1
u
x
+ b
2
u
y
+ cu(x, y) = 0. (0.42)
L is said to be linear if for all , C and all admissible functions u, v on X
L(u + v) = Lu + Lv. (0.43)
One of the essential advantages of considering linear PDEs is that if each of
u
1
, . . . , u
k
C
2
(X) is a solution to the homogeneous PDE (0.41) then for any
choice of constants
j
C
u(x, y) :=
k

i=1

j
u
j
(x, y) (0.44)
is also a solution. This is called the superposition principle, and is a fact we will
use repeatedly in what follows. Indeed, by linearity we have
Lu = L
_
k

i=1

j
u
j
_
(0.43)
=
k

i=1

j
Lu
j
..
=0
= 0. (0.45)

Potrebbero piacerti anche