PDE Notes

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 16

Lecture 1: Linear PDEs

Abstract
We outline solutions to four basic linear PDEs: the transport equation, the Laplace and Poisson
equations, and finally the linear heat equation.

Contents
1 Introduction
2 The
2.1
2.2
2.3
2.4
2.5
2.6

Transport (or Advection) Equation


Introduction . . . . . . . . . . . . . . . . . . . . . . . . . .
Directional Derivatives . . . . . . . . . . . . . . . . . . . .
Solving the Transport Equation (Homogeneous Case) . . .
The n-dimensional Transport Equation . . . . . . . . . . .
Solving the Transport Equation (Nonhomogeneous Case)
An Example of a Concrete Problem . . . . . . . . . . . .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

2
2
3
3
4
5
5

3 Laplaces Equation
3.1 The Laplace Operator . . . . . . . . . . . . . . . . . .
3.2 The Fundamental Solution to Laplaces Equation . . .
3.3 Some Theorems From Calculus . . . . . . . . . . . . .
3.3.1 Gauss-Green Theorem . . . . . . . . . . . . . .
3.3.2 n-Dimensional Version of Integration by Parts
3.3.3 Greens Formulas . . . . . . . . . . . . . . . . .
3.3.4 Divergence Theorem . . . . . . . . . . . . . . .

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

5
5
7
8
8
8
9
9

4 Poissons Equation
4.1 Introduction . . . . . . . . . . . .
4.2 Properties . . . . . . . . . . . . .
4.3 Properties of Harmonic Functions
4.3.1 Mean Value Property . .
4.3.2 The Maximum Principle .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

9
9
10
12
13
14

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

5 The Heat Equation


14
5.1 The Linear Heat Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
5.2 The Geometric Heat Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

ECE 6560: PDEs in Image Processing and Computer Vision, Spring 2008, Georgia Institute of Technology. Professor:
Anthony Yezzi. Scribe: Hassan A. Kingravi

Introduction

These lectures are transcription of notes for Dr. Anthony Yezzis Spring 2008 class on PDEs in IP and CV.
The initial treatment of analytical solutions for three primary linear PDEs derive heavily from[1].
We will not go into why PDEs are useful; you can figure this out once we start applying them to various
problems. We will mention, however, that part of the power of a PDE derives from the fact that it describes
the local properties/structure of a function (because of the notion of a partial derviative), but at the same
time, extends that notion globally, because the equation holds for the function over a certain domain. Hence
one can deduce global properties for a function based on whats happening locally.
The notes assume a certain degree of mathematical maturity, although not to a great extent. The reader
should be familiar with mutlivariable calculus (although we will cover some of the main theorems again) and
should have had some exposure to ordinary differential equations and Fourier transforms. An understanding
of some basic concepts from analysis, such as epsilon/delta continuity, the derivation of the area/volume of
a hypersphere or the notion of functions with compact support etc is helpful, but not necessary.
In a lot of cases, we deal with vector-valued functions. In these notes, these are denoted by a bold-face
font. In later notes, we will not distinguish them as such, because their introduction is apparent from the
context.

The Transport (or Advection) Equation

2.1

Introduction

We will look at the 1-D version first.


ut + bux = 0, i.e.

u
u
+b
=0
t
x

where u : R x R R. This equation is an example of a time-evolution equation. The 1-D refers to the
dimension of the spatial variable. Assuming that the function u satisfies the PDE, what does the equation
tell us about the function?
The solution is as follows: given initial conditions:
1) u(x, 0) = g(x) (initial boundary value)
2) ut + bux = 0 (PDE)
the solution is
u(x, t) = g(x bt)

Figure 1: An example transportation


What this essentially means is that the equation takes the initial function, and transports it completely
elsewhere (see figure 1). We will derive the solution to this problem as stated in what follows. For this, we
need some mathematical preliminaries.

2.2

Directional Derivatives

Lets say one has a function in R2 . What is its directional derivative?

Figure 2: A function on an x-y plane


We want to essentially pick a direction v, and take the derivative in that direction (see figure 2). To do
this, we have the following:
f (x0 + v1 , y0 + v2 )
f
|(x ,y ) = lim
0
v 0 0

This is not convenient to calculate numerically. Let us assume that a Taylors expansion of the function
exists (i.e. assume its differentiable etc). We then get:
1
1
f (x0 + v1 , y0 + v2 ) = f (x0 , y0 ) + v1 fx |(x0 ,y0 ) + v2 fy |(x0 ,y0 ) + 2 v12 fxx |(x0 ,y0 ) + 2 v22 fyy |(x0 ,y0 )
2
2
+ 2 v1 v2 fxy |(x0 ,y0 ) + O(3 )
Subtract f (x0 , y0 ) from both and divide by , and you get:

lim

0

f (x0 + v1 , y0 + v2 ) f (x0 , y0 )


1
1
= v1 fx |(x0 ,y0 ) + v2 fy |(x0 ,y0 ) + v12 fxx |(x0 ,y0 ) + v22 fyy |(x0 ,y0 )

2
2
2
+ v1 v2 fxy |(x0 ,y0 ) + O( )
v1 fx |(x0 ,y0 ) + v2 fy |(x0 ,y0 )
 
v1
=
(fx , fy )|(x0 ,y0 )
v2

Essentially, we end up with the following:


f
= v f |(x0 ,y0 )
v

2.3

Solving the Transport Equation (Homogeneous Case)

We can use what we have just derived on the transport equation. Since time is a separate variable,
ut + bux = 0
3

becomes
 
 
b
b
(ux , ut ) =
u = 0
1
1
 
b
Hence our direction becomes
, and we take the directional derivative of u in this direction (call it
1
u
= 0 means is that since this derivative is 0 for all points x and t, the function is constant in
v). What v
this direction.

Figure 3: Characteristic curves


In essence, any function which obeys this law (i.e. being constant along a certain gradient) fits the bill
as a solution. In figure 3, we can see an example of characteristic curves, i.e. curves along which the
PDE becomes an ODE, and which can be used to solve it. We formalize this notion using the following
computation:
z(s) = u(x + sb, t + s)
where s is a parameter for the characteristic curve.
z 0 (s) = bux + ut = 0
We want to start at the boundary (t = 0), so let s = t.
z(t) = u(x bt, 0) = f (x bt)
Now pick s = 0. This leads to z(0) = u(x, t), which is the same as z(t), because z is a constant function.
Hence u(x, t) = f (x bt), with the boundary condition (t = 0) being f (x).

2.4

The n-dimensional Transport Equation

Here is the n-dimensional equivalent of the transport equation:


ut + b x u = 0
Solution:
u(x, 0) = f (x) (boundary condition)
u(x, t) = f (x b t)
4

2.5

Solving the Transport Equation (Nonhomogeneous Case)

We now outline the statement of and the solution to the nonhomogenous, n-dimensional transport equation.
ut + b x u = f (x, t)
u(x, 0) = g(x)
We again use the method of characteristics:
z(s) = u(x + sb, t + s)
0

z (s) = ux b + ut = f (x, t)
Solve the above ODE.
Z

z 0 (s)ds

z(0) = z(t) +
t
Z 0

= z(t) +

f (x + sb, t + s)ds
t
t

f (x + (s t)b, s)ds

= z(t) +
0

We end up with the following solution:


Z

u(x, t) = g(x tb) +

f (x + (s t)b, s)ds
0

2.6

An Example of a Concrete Problem

Looking at the previous two sections, let the following hold:


f (x, t) = x cos t
g(x) = sin x
b=1
Thus ut + ux = x cos t. Then:
Z

u(x, t) = sin(x t) +

(x + s t) cos(s)ds
Z t
Z t
= sin(x t) +
(x t) cos(s) +
s cos(s)ds
0

= sin(x t) + (x t) sin(t) x + t + cos t + t sin(t) 1

3
3.1

Laplaces Equation
The Laplace Operator

This equation doesnt depend on time, only on its spatial variables. In the 2-dimensional case, one writes it
as uxx + uyy = 0. In the n-dimensional case, given x = (x1 , x2 , . . . , xn ), we write this as:
ux1 x1 + ux2 x2 + + uxn xn = 0
We write the above compactly as u = 0. We call this the Laplacian operator. A quick note; in
engineering notation, the Laplacian is written as the gradient operator applied to an equation twice, i.e.
= 2 . We use the latter to donate the Hessian operator. The Laplacian is the trace of this.
5

Again, what is this PDE saying? Something local, but since it applies to all points, it says something
global. Essentially, u = 0 describes the local structure of a function which represents the equilibrium of
a quantity in flux. For example, the equation can represent the equilibrium of the diffusion of temperature
through a wall. Similar interpretations hold in electrostatics and chemical concentrations, etc.

Figure 4: Region
We can give a brief description of how this equation arises in a general setting. Let R be an arbitrary,

compact (closed) region thats part of the domain of u. Let F = flux density, i.e. a vector field in the domain
of u. If you integrate this flux around the boundary in the direction of the normal to the boundary (i.e. N :
see figure 4), that integral must end up being 0 (by our equilibrium assumption). Formally speaking, we
denote this as:
Z

F N ds = 0
R

where ds is the arclength contribution. In physics, we often denote flux in a certain manner. For

example, given Fickes law of diffusion, we have F = au, where the is the gradient of the the quantity
in question. Now, we can use the Divergence theorem from calculus, which basically states that the integral
of the flux around the boundary of a set is equal to the integral of the divergence in the interior of the set.
(We will derive the Dirvergence theorem later, in case the reader is a little hazy on the details.) Putting it
all together, we end up with the following computation:
Z
R

F N ds =

F dA
R
Z
= a
u dA = 0
R
Z
=
u dA = 0
Z

Note that this is not quite Laplaces equation. However, because the region R is completely aribtrary,
the only way this can hold is if the sum of each second derivative is indeed zero. Hence u = 0.
One reason the Laplacian operator
  is useful
  in vision is due to its rotational invariance. What this means
x
x

is that uxx + uyy = uxx + uyy if


and
are both orthonormal bases for the space.
y
y
Consider rotating a vector set (x, y) by . This is represented classically by the following:

  
u
cos
=
v
sin

sin
cos

 
x
y

Let f : R2 R. We need to show fxx + fyy = fuu + fvv . Using the chain rule, we have:
fxx + fyy = (fx )u ux + (fx )v vx + (fy )u uy + (fy )v vy
= (fx )u cos (fx )v sin + (fy )u sin + (fy )v cos
= (fu cos fv sin )u cos (fu cos fv sin )v sin
+ (fu sin fv cos )u sin + (fu sin fv cos )v cos
= fuu + fvv
We can also prove invariance in the n-dimensional case. Let y = A x, where xT = (x1 , . . . , xn ) and
y = (y1 , . . . , yn ). Furthermore, let Dx and Dy be the Jacobians of x and y. For a scalar function u,
Du = (u)T , and DDT u = 2 u. Let v(x) = u(y) for u, v : Rn R. Then we have the following:
T

Dv = DuA
T

D(Dv) = D(DuA)T A
DDT = AT DDT uA
2 v = AT 2 uA
Note that every time we take the Jacobian of the function v, we have a multiplication of the function
with the rotation matrix A. Since A is an orthogonal matrix, AT = A1 . We then have:
2 v = A1 2 uA
Since v = Tr 2 v and
u = Tr 2 u, and the eigenvalues dont change after an orthogonal transformaX
tion (recall that Tr(A) =
i , where i is the ith eigenvalue of the matrix), we have u = v.

3.2

The Fundamental Solution to Laplaces Equation


u = ux1 x1 + ux2 x2 + + uxn xn = 0

By the fundamental solution to Laplaces equation (denoted (x)), we mean the only nonconstant solution
which possesses rotational symmetry (there are other solutions as well). To develop this, we note that if u(x)
is rotationally symmetric,pit only depends on its radial component r, i.e. u(x) = v(r), where u : Rn R,
r
= xri , for x 6= 0.
v : R R, and r = |x| = x21 + + x2n . In addition, x
i
From this we have the following:

uxi xi
Since u =

Pn

i=1

xi
uxi = v 0 (r)
r

2
0
2
v
(r)
x
x
= v 00 (r) 2i +
1 i
r
r
r

uxi xi = 0, we have:

(n 1) 0
v (r) = 0
r
We want to eliminate constant solutions, thus v 0 6= 0. Hence:
v 00 (r) +

(1 n)
v 00
=
0
v
r
(1 n)
0 0
(log v ) =
r
Z
(1

n)
log v 0 =
r
log v 0 = (1 n) log r + c
a
v 0 = (n1)
r
We thus get:
v(r) = a log(r) + c (if n = 2)
a
+ c (if n > 2)
r(n2)
u(x) = (x) = b log(|x|) (if n = 2)
b
(if n > 2)
|x|(n2)
where b =

3.3

1
2

in 2-D, and

1
n(n2)(n)

(1)

in n-dimensions, (n) being the volume of the unit n-sphere.

Some Theorems From Calculus


b

u0 dx = u(b)u(a). We will write down the equivalent

Recall the Fundamental Theorem of Calculus in 1-D:


a

of this theorem in n-dimensions, and then use it to derive all the other theorems we need.
3.3.1

Gauss-Green Theorem
Z

uN i ds

uxi dx =

(2)

where u is a function, is a compact (closed) set in n-dimensions, is its boundary, N i is the ith
component of the normal to this set, and ds is the boundary measure. If you apply this in one dimension,
its easy to see that the Fundamental Theorem of Calculus follows.
3.3.2

n-Dimensional Version of Integration by Parts

To get the formula for Integration by Parts in 1-D, all we have to do is to apply the FTOC on a product of
2 functions.
Z

(uv)0 dx = uv]ba

a
b

(u v + uv ) =

uv=

uv]ba

uv 0

We can try a similar approach with Gauss-Green in higher dimensions.


Z
Z
(uv)xi dx =
uvN i ds

Z
uxi vdx =

Z
uvxi dx +

uvN i ds

(3)

3.3.3

Greens Formulas

We want coordinate-free versions of these integrals. Hence:


Z
Z
u
ds
udx =
N

Z
Z
Z
u vdx =
u vdx +

(4)
u

v
ds
N

(5)

Both of these can be easily proved using what weve learned so far. For completeness sake, we prove the
first one. Apply the Gauss-Green theorem to uxi
Z
Z
uxi dx =
uN i ds

and plug-in to get the following:


Z

uxi N i ds

uxi xi dx =

Now do a summation over the variables, and you get:


n Z
X
i=1

n Z
X

uxi xi dx =

i=1

uxi N i ds

Z
u N ds

udx =

Z
=

u
ds
N

because u N is the directional derivative.


3.3.4

Divergence Theorem

A more natural formula to use in our case would be the following (which is also coordinate-free)
Z
Z

F N ds =
F dx

(6)

where F is the divergence of F , or the sum of its first-order partial derivatives. Again, this can be
easily proved by applying the Gauss-Green theorem to each component, and then summing over them.

4
4.1

Poissons Equation
Introduction

Poissons equation is the nonhomogeneous version of Laplaces equation, i.e. u = f, u : Rn R, f : Rn


R. Recall that is the fundamental solution to Laplaces equation. (Note that any function that satisfies
Laplaces equation is called a harmonic function.) Given the fundamental solution, what can we say about
Poissons equation? In other words, can we use the fundamental solution to derive a solution for Poissons
equation? The answer is yes, although the details are intricate.
Suppose f is C 2 with compact support (i.e. f is twice differentiable over some closed, bounded domain,
outside of which it has zero support, or, in other words, it goes to 0. This condition is included to make sure
the integral of f is finite). Choose any y Rn . Then (x y) f (y) = harmonic x 6= P
y. By linearity, a
finite sum of such terms is also harmonic, except at the values for y; we can denote this by (xyi )f (yi )
The next, most obvious question is whether an infinite sum is finite. In other words, is the following
harmonic?

Z
(x y) f (y)dy
Rn

The answer is no, because the ys end up being infinite. Its similarly tempting to apply the Laplacian
operator to this and end up interchanging the integrals to end up with a 0. But this is again incorrect. In
what follows, we compute both these quantities rigorously. Note that we now abandon the use of bold-face
notation for vectors.

4.2

Properties

We have

Z
(x y) f (y)dy

u(x) =
Rn

and we want to calculate x u(x). Notice that the above equation is simply the convolution of the
fundamental solution with f . Using the fact that f = f , we get the following:
Z
(y) f (x y)dy

u(x) =
ZR

(y) x f (x y)dy

x u(x) =
Rn

Z
(y) x f (x y)dy +

(y) x f (x y)dy
Rn B(0,)

B(0,)

where B(0, ) represents a small ball of radius  around the 0 point, where we have our singularity. Call
the first part I and the second part J . Lets examine I . Notice that:
Z
|I | max |x f (x y)|
(y)dy
B(0,)
Z
kx f k
(y)dy
B(0,)

where kx f k is simply a notational substitution for max |x f (xy)|. Also notice that the fundamental
solution is rotationally symmetric, and only depends on the radius . Therefore, if we substitute the analytic
expression for the fundamental solution (see section 3.2) and integrate, we get the following:
|I | kx f k C2 | log | if n = 2
kx f k C2 if n 3

where C is some constant. Lets see the calculation for this in detail.
Let n = 2.
Z
Z
I =
(y) x f (x y)dy
log |y|dy
B(0,)

B(0,)

Transform into polar coordinates.

10

I

log(r)rdrd
0

1
= (r2 log r r2 )]0
2
1
= 2 log  2
2
C2 log 

As  0, C2 log  0. Now let n 3. Then:


Z
Z
(y) x f (x y)dy
B(0,)

B(0,)

1
dy
|y|n2

If we do a polar transform, the contribution of all the theta pieces in the calculation is simply a constant.
So this is proportional to the following:
Z

1
(n)rn1 dr
n2
|r|
0
Z 
(n)
rdr

I

C2

So what have we shown? In essence, a value bigger than I goes to 0 as  0. Therefore, I 0. So


now, we must focus on J .
Z
J =
(y) x f (x y)dy
Rn B(0,)

Note that since the Laplacian is rotationally invariant, we can use any orthonormal basis, and thus can
switch x to y , i.e.
Z
J =
(y) y f (x y)dy
Rn B(0,)

Use Greens second formula on this:


Z
Z
J =
(y) yf (x y)dy +
Rn B(0,)

B(0,)

(y)

f
ds
N

f
N

Call the first part K , and the second part L . Note that
is the outward unit normal pointing along
B(0, ). Lets examine L first. We use a trick similar to what we did the last time:
Z
|L | kf k
(y)dy
B(0,)

Since kf k is bounded, we know from before that due to the nature of the fundamental solution, L
goes to zero for all dimensions. We are therefore left with K . This too can be split into two pieces, using
our earlier trick:

11

Z
K =

(y) yf (x y)dy
Rn B(0,)

Z
(y)f (x y)dy +

=
Rn B(0,)

B(0,)

(y)
f (x y)dS
N

Note that since (y) = 0, the first part of K is 0. The second part is more interesting, and can be
written as:
Z
N f (x y)dS
B(0,)

What is ? Lets just focus on the n 3 case. From this, we can see:
1
1
rn2 n(n 2)(n)
1
1
0 (r) =
n1
n(n) r
dr
yi = 0 (r)
yi


1
yi
1

=
n(n) rn1
r
yi
=
n(n)rn
(r) =

Since r =

y
y12 + + yn2 , = n(n)|y|
n.

Z
x u(x) = lim

0

B(0,)

f (x y)dS
N

(Note that this is so because all the other parts of the equation have gone to zero.) Because N is the
unit normal on the ball around the origin, N = y
|y| . From this, it follows:
|y|2
f (x y)dS
n+1
B(0,) n(n)|y|
Z
1
=
f (x y)dS
n1
n(n)|y|
B(0,)
Z
1

f (x y)dS
n1
n(n)
B(0,)
Z
1
f (x y)dS
=
n(n)n1 B(0,)
Z

x u(x) =

Note that this quantity n(n)n1 is the surface area of a ball with radius  in n dimensions. Since we
are integrating over the same ball, we are essentially calculating an average over an area. (In these notes,
R
we use the following to indicate an average integral: .) Note, however, that as  0, this area shrinks. By
assumption, f is continuous. Therefore, as the area shrinks to a point, what is the average over this point?
By continuity of f , the answer is f (x).
In summary, we end up with the following conclusions:
u=f
12

(7)

u = f (x)

(8)

= (x)

(9)

where the last equation is a simple formalism involving Diracs delta function.

4.3

Properties of Harmonic Functions

Recall that u : Rn R is harmonic on the domain Rn if u(x) = 0 x .


4.3.1

Mean Value Property

Suppose theres a ball of radius r around a point x, s.t. B(x, r) . Then, from the previous section,
R
u(x) = B(x,r) udA, where dA represents the surface of the ball. If we consider y as a dummy variable, we
can write the following:
Z
(r) =

u(y)dA(y)
B(x,r)

Z
=

u(x + rz)dA(z)
B(0,1)

where we have performed a simple change of variables. We differentiate this quantity in r, and substitute
back the original variable to get:

0 (r) =

Du zdA(z)
B(0,1)

Du(y)

=
B(x,r)

yx
dA(y)
r

Note that yx
is N , and dA(y) is the flux of the gradient. Using this information, and the formula for
r
the surface area of the ball we derived earlier, we get:

u
dA
N
B(x,r)
Z
u
1
=
dA
n1
n(n)r
B(x,r) N
Z
1
udy = 0
=
n(n)rn1 B(x,r)

(r) =

where we use Greens formula to get the last line. What have we just proved? That the derivative of is
0. Hence (r) is a constant, and is thus independent of radius. Therefore, by continuity, limr0 (r) = u(x).
Consider the solid case, i.e. the integration over the region of the ball itself. We write the integral over
a series of dummy variables:
Z

u(y)dy =
B(x,r)

!
u(y)dA(y) dr0

B(x,r 0 )

The integral within the parentheses, from our previous derivation is simply

13

u(y)dA(y) dr =
0

B(x,r 0 )

u(x)n(n)r0n1 dr0

Z
= (n)u(x)

nr0n1 dr0

= (n)rn u(x)
If you take the average of this, we end up with the solid case for the previous equation, and the answer,
of course, ends up being u(x). This yields the mean value property, which states that the mean of a harmonic
function in any ball in the domain is equal to the point around which the ball is centered.
This property is quite remarkable on its own. But what is even more astonishing is that the converse is
also true. Consider the following proof using the assumption that a function u(x) is C 2 (admits at least two
derivatives), and is defined on Rn . Given any ball B(x, r) , x, r , the converse is that if u(x)
has the following property:
Z
u(x) =

udA
B(x,r)

then it is a harmonic function. To prove this, suppose u 6= 0 at some point x . Then, by continuity
(the Mean Value Theorem), there exists a neighborhood B(x, r0 ) where u > 0 (by no loss of generality).
Then the following:
Z
1
udy
(r) =
n(n)rn1 B(x,r)
0

is not 0, because the inner integral is not 0. This contradicts the property (see the previous section), and
thus the converse is true.
Interestingly, we do not need the C 2 assumption, although this makes the proof much messier. However,
it leads to a quite improbable conclusion, i.e. that an integral relationship demands the functions admitting
2 derivatives.
4.3.2

The Maximum Principle

Given that a function u is harmonic over , an open, bounded set:


= + represents the closure of the set .
1. Weak Maximum Principle: max = max , where
Then u constant
2. Strong Maximum Principle: Suppose is connected and u(x0 ) = max u x0 .
within u.
The first principle is remarkable enough on its own. But the second principle says that if theres some
point strictly within (i.e. which does not lie on the boundary) where the function u reaches its maximum,
then there is no choice but to have the function be constant on its whole domain. Therefore the strong
maximum principle essentially states that the only interesting harmonic functions are those that have their
maximums restricted to the boundary.
Intuitively, one cas see that this follows from the mean value property. Suppose that the maximum is
not attained at the boundary, and is instead obtained in some x . Then its clear that due to the mean
value property, any maximum value inside the set is the mean of all possible open balls that remain inside
the set, implying that this value is constant over all the set, thus in turn implying that the maximum is also
attained at the boundary.
Suppose that the value of u at x is not constant within some ball around it. Then from basic calculus,
we have that the value of the mean in the ball is strictly less than u(x). However, this contradicts the mean
value property. Therefore, we cannot have a local maximum, and every point in the domain is a saddle
point. Therefore, the only way to get a maximum is to reach the boundary of the function. The only reason
its harmonic there is because it ceases; if we have a harmonic function defined everywhere, it would have
no maximum. The same holds for the minimum.
14

The Heat Equation

We now introduce the heat equation, and derive solutions for the linear case. We end the section by designing
a nonlinear heat equation useful for image processing.

5.1

The Linear Heat Equation

The heat equation basically relates the Laplacian operator to the derivative in time, i.e.
ut = x u
n

where u : R R R, and x is the spatial Laplacian.


Recall the derivation of the Laplacian, in which we stated that it represented the equilibrium of diffusion
processes. The heat equation is simply a method of modeling the diffusion itself. Note that if there is no
change in time, i.e. if ut = 0, we simply end up with the Laplacian, which has the exact same interpretation
as before. Lets derive this equation in a similar way as before, i.e. via a conservation law. Consider a region
where the density function u is defined, and then pick some arbitrary subset R. Suppose that density is
diffusing across the boundary of the region. If one integrates over the density, this gives the total amount of
mass. Then the law of conservation states that the change in mass is equal to the mass that leaves across
the boundary, which can be given by the following equation

Z
u dx =

F~ N dA

where F~ is the flux vector field, N is the outward normal, and the minus sign is added just as our
convention to denote that the change in mass is negative (i.e. it leaves). From our previous work, we have
Z

F~ N dA
u dx =
R t
R
Z
Z

u dx = a
udx
R t
R
where a is some physical constant that we dont really care about. Since this has to hold for all possible
regions R, we get the heat equation.
To solve the equation, first consider the 1-dimensional case, i.e. ut = uxx . Take the Fourier transform of
the spatial variables, which will be mapped to the frequency domain (rather than the time variables that are
usually used). Because of this, the time variable remains untouched. We denote the transformed function
by U , and the transform itself by F. This gives us
U = F{u}
Ut = (j)2 U
U = 2 U
2

This is an ODE, and a fairly trivial one; the solution is u(t) = U (0)e t . Since all these ODEs are
decoupled, this doesnt depend on . To go back to the spatial domain, apply the inverse Fourier transform,
to get
2

u(x, t) = u(x, 0) F 1 {e t }
1 x2
= u(x, 0)
e 4t
4t

15

Note that the latter part is a Gaussian of variance 2t. If one wishes, one can relate this to standard DSP
methods.

5.2

The Geometric Heat Equation

We end with a nonlinear variant of the heat equation that is extremely useful in image processing, and which
we will meet again in the context of active contours.
Recall that the Laplace operator is rotationally invariant. Given the image as I, in the linear case, we
diffuse across the x and y axes, i.e.
It = Ixx + Iyy
Consider diffusing in the direction of a vector v = (v1 , v2 ). Then the derivatives of the image in this
vectors direction are:
Iv = I v
Ivv = vT 2 I v
where I is the gradient of I, and 2 I is its Hessian. We can now write the linear heat equation in terms
of two arbitrary directions, which we designate to be along an edge , and across an edge , respectively
i.e.
It = I + I
We need a candidate for . To detect an edge, a natural method would be to consider its gradient.
I
. We want to smooth along edges only, so we can remove cross edge
We normalize this, to get = kIk
information:
It = I + I I
= I I
I T 2
I
= I (
) ( I)(
)
kIk
kIk
I T 2 II
= I
I T I
I T 2 II
= Ixx + Iyy
I T I
Manipulating the last expression, we end up with the following:
It =

Ix2 Iyy 2Ix Iy Ixy + Ixx Iy2


Ix2 + Iy2

This is known as the geometric heat equation. Notice that its a nonlinear PDE, which should give it
more expressive power than the linear case. We can rewrite the equation more compactly as follows:
It = kIk (

I
)
kIk

Notice here that I represents the gradient of the image, which is a 2D vector, and (.) represents the
divergence or del operator.

References
[1] L. Evans. Partial Differential Equations, AMS, 1998.

16

You might also like