Beruflich Dokumente
Kultur Dokumente
Poissons Equation
2.1
Physical Origins
Poissons equation,
r2
= (x),
arises in many varied physical situations. Here (x) is the source term, and is often zero,
either everywhere or everywhere bar some specific region (maybe only specific points).
In this case, Laplaces equation,
r2
= 0,
results.
19
R. E. Hunt, 2002
S(x)
.
k
The diusion equation for a solute can be derived as follows. Let (x) be the concentration of solute at
the point x, and F(x) = kr be the corresponding flux. (We assume here that there is no advection
of by the underlying medium.)
Let V be a fixed volume of space enclosed by an (imaginary) surface S. In a small time t, the quantity
of solute leaving V is given by
ZZ
F t . n dS.
S
Hence
Z Z Z
V
t+ t
dV
=
t
ZZ
F . n dS t.
@
dV =
@t
ZZZ
r . (kr ) dV.
20
R. E. Hunt, 2002
If there are also sources (or sinks) of solute, then an additional source term results:
@
= kr2 + S(x)
@t
where S(x) is the quantity of solute (per unit volume and time) being added to the solution at the
location x. Poissons equation for steady-state diusion with sources, as given above, follows immediately.
The heat diusion equation is derived similarly. Let T (x) be the temperature field in some substance
(not necessarily a solid), and H(x) the corresponding heat field. We have the relation H = cT where
is the density of the material and c its specific heat. The corresponding heat flux is krT . A similar
argument to the above applies again, resulting in
@H
= kr2 T + S(x)
@t
where S represents possible sources of heat. Hence
@T
= r2 T + (c)
@t
S(x)
where = k/c is the coefficient of thermal diusivity. The equation for steady-state heat diusion with
sources is as before.
Electrostatics
The laws of electrostatics are
r . E = /0
rE=0
r.B=0
r B = 0 J
where and J are the electric charge and current fields respectively. Since r E = 0,
there is an electric potential such that E = r ; hence r . E = /0 gives Poissons
equation
r2
/0 .
In a region where there are no charges or currents, and J vanish. Hence we obtain
Laplaces equation
r2 = 0.
Also r B = 0 so there exists a magnetostatic potential
r2 = 0.
such that B =
0 r ; and
Gravitation
Consider a mass distribution with density (x). There is a corresponding gravitational
field F(x) which we may express in terms of a gravitational potential (x). Consider an
RRR
arbitrary fixed volume V with surface S containing a total mass MV =
(x) dV .
V
21
R. E. Hunt, 2002
Gauss showed that the flux of the gravitational field through S is equal to
Hence
ZZ
F . n dS = 4GMV
4GMV .
ZZ
=)
r . n dS =
4G
=)
ZZZ
ZZZ
(x) dV
r . (r ) dV = 4G
ZZZ
(x) dV.
= r . (r ) = 4G.
Other applications
These include the motion of an inviscid fluid; Schrodingers equation in Quantum Mechanics; and the motion of biological organisms in a solution.
2.2
1 @
@
1 @2
r
+ 2
= 0.
(1)
r @r
@r
r @2
Consider solutions of the form (r, ) = R(r)() where each function R, is a function
of one variable only. Then
1 @
@
() d
dR
r
=
r
r @r
@r
r dr
dr
and
1 @2
R(r) d2
=
.
r2 @2
r2 d2
dR
r
dr
22
00
.
(2)
R. E. Hunt, 2002
The LHS is a function of r only, and the RHS of only; hence both must be constant,
say. Then
00 =
=)
8
<A + B
=
p
p
:A cos
+ B sin
=0
6= 0
To obtain a sensible physical solution, replacing by + 2 should give the same value
of r (see later). This is true only if 0 ( + 2) = 0 () 8 ; i.e., either = 0 or
cos 2
= 1 and
sin 2
=0
p
which implies 2
= 2n for some integer n. (Note that the possibility that
ruled out at this stage.) Hence
8
<A + B
n=0
=
:A cos n + B sin n n 6= 0
< 0 is
Returning to (2),
r d
R dr
=)
dR
r
dr
r2 R00 + rR0
= n2
n2 R = 0.
We note that the combination ln r does not satisfy the requirement above for 2periodicity of r , and so we exclude it. Equation (1) is linear and so we may form
a superposition of the above solutions; in fact the general solution is an arbitrary linear
combination of all the possible solutions obtained above, that is
= A0 + B0 + C0 ln r +
1
X
(An r + Cn r
n=1
) cos n +
1
X
(Bn rn + Dn r
) sin n
n=1
23
R. E. Hunt, 2002
where we have relabelled all the arbitrary constants, e.g., AC has become An and BD
has become Dn . We can make this expression more compact by defining A n = Cn and
B n = Dn for n > 0; then
= A0 + B0 + C0 ln r +
1
X
n= 1
n6=0
Although this is more compact, the first expression is often easier to use.
Notes:
(i) Why did we require that r , rather than
itself, be periodic? In many cases
(e.g. temperature, diusion),
must clearly be periodic and so we shall further
need B0 = 0. But in other cases (e.g. electrostatics, gravitation),
is not itself
a physical quantity, only a potential; it is r which has a physical significance
(e.g., the force). For example, consider the magnetostatic potential around a wire
carrying a current I; here = (I/2), which is multi-valued, but B = 0 r
(the quantity of physical interest) is of magnitude 0 I/2r and is single valued.
(ii) A common mistake made during separation of variables is to retain too many arbitrary constants; e.g. to write
X
Cn rn (An cos n + Bn sin n).
For each n, this looks like 3 arbitrary constants (An , Bn , Cn ); but of course there
are really only two arbitrary quantities (Cn An and Cn Bn , which we have relabelled
as An and Bn above).
(iii) The above derivation also applies to 3D cylindrical polar coordinates in the case
when is independent of z.
1 @
1
@
@
2@
r
+ 2
sin
= 0.
r2 @r
@r
r sin @
@
Seek solutions of the form
1 2 0 0
(r R ) =
R
(3)
R. E. Hunt, 2002
say. So
(0 sin )0 =
sin .
Let = cos , and use the chain rule to replace d/d by d/d:
d
d d
=
=
d
d d
So
=)
sin
d
.
d
d
2 d
sin
sin
=
sin
d
d
d
2 d
(1 )
+ = 0.
d
d
=)
n(n + 1)R = 0,
n 1
The general solution to Laplaces equation in the axisymmetric case is therefore (absorbing the constant C into A and B)
(r, ) =
1
X
(An rn + Bn r
n 1
)Pn (cos ).
n=0
may depend on
(r, , ) =
1
X
n
X
n 1
n=0 m= n
where Pnm () are the associated Legendre functions which satisfy the associated Legendre equation
d
m
2 d
(1 )
+ n(n + 1) +
=0
d
d
1 2
when m and n are integers, n
0,
n m n.
25
R. E. Hunt, 2002
2.3
= (x)
(x) =
From a physical point of view, we have a well-defined problem; say, find the steadystate temperature distribution throughout V , with heat sources given by (x), subject to
a specified temperature distribution on the boundary. No further conditions are required
in real life to ensure that there is only one solution. Mathematically, then, can we show
that the problem above has a unique solution?
Suppose that there are actually two (or more) solutions 1 (x) and
2 . Then
r2 = r2 1 r2 2 =
=0
in V
2 (x).
Let
subject to
=f
One solution of this problem for
f =0
is clearly
on S.
= 0; is it unique? Consider
r . ( r ) = r . r + r . (r )
= |r |2 + r2
= |r |2 .
Hence
ZZZ
|r | dV =
ZZZ
r . ( r ) dV
ZZ
r . n dS
=0
because
= 0 on S. But |r |2
0 everywhere; its integral can only be zero if |r |
is zero everywhere, i.e., r 0, which implies that
is constant throughout V . But
= 0 on S, so
0 throughout V . Thus 1 = 2 , which demonstrates that our
problem has a unique solution, as expected.
A similar theorem holds when instead of Dirichlet boundary conditions we have Neumann boundary conditions: that is to say instead of being specified (by the function
26
R. E. Hunt, 2002
2.4
Suppose that satisfies r2 = 0 in a volume V with surface S. Then both the minimum
and maximum values of occur somewhere on S (and possibly also somewhere inside
V ).
Why is this? Suppose has a local maximum somewhere in the interior of V . At
that point we must have @ /@x = @ /@y = @ /@z = 0 (stationary point); and as it is
a maximum, @ 2 /@x2 < 0, @ 2 /@y 2 < 0, @ 2 /@z 2 < 0. But this cannot happen since
0 = r2 = @ 2 /@x2 + @ 2 /@y 2 + @ 2 /@z 2 .
The same applies to minima.
This is not a formal proof since it is actually possible for a maximum to have
@2
@2
@2
=
=
= 0,
@x2
@y 2
@z 2
a case we havent considered: compare with the possibility in 1D that a maximum could have d2 y/dx2 =
0. However, the theorem can still be shown to hold.
2.5
Greens Function
Rb
whenever x0 2 (a, b). As a consequence, a f (x) (x x0 ) dx = f (x0 ). We extend the
definition to 3D via
(x x0 ) = (x x0 ) (y y0 ) (z z0 )
27
R. E. Hunt, 2002
f (x) (x
x0 ) dV = f (x0 )
Greens Function
Suppose that we wish to solve Poissons equation in a volume V with surface S on which
Dirichlet boundary conditions are imposed. The Greens function G(x; x0 ) associated
with this problem is a function of two variables: x, the position vector, and x0 , a fixed
location. It is defined as the solution to
r2 G(x; x0 ) = (x
G=0
x0 )
in V ,
on S.
(Physically, we can think of G as the potential from a point source at x0 with the
boundary held at zero potential.)
It is possible to prove that G is symmetric, i.e., G(x; x0 ) = G(x0 ; x). This can be useful as a check that
G has been correctly calculated. Physically, this corresponds to the remarkable fact that the potential
at x due to a source at x0 is the same as the potential at x0 due to a source at x, regardless of the shape
of S.
When V is all space (i.e., the limit of a sphere whose radius tends to 1), Greens
function is known as the fundamental solution.
For a problem with Neumann boundary conditions, G(x; x0 ) is defined to satisfy
RR
@G/@n = 1/A on S, where A = S dS is the surface area of S, rather than G = 0
there. In many cases S is infinitely large, in which case the boundary condition reduces
to @G/@n = 0.
28
R. E. Hunt, 2002
VR
VR
ZZ
ZZ
rG . n dS =
SR
g 0 (r) dS
SR
=)
g 0 (R) =
1
4R2
=)
g 0 (r) =
1
4r2
=)
g(r) =
for all R
1
+ A,
4r
1/4|x|.
Shifting the origin to a non-zero x0 , we see that in general the fundamental solution
in 3D is
1
G(x; x0 ) =
.
4|x x0 |
Example: an electron located at x0 is an electrostatic point source, so the charge
distribution in space is (x) = e (x x0 ). Hence the electrostatic potential obeys
r2
= (e/0 ) (x
x0 )
using a result from 2.1. The solution is therefore just a factor e/0 times the fundamental solution, i.e., e/40 |x x0 |. This is the standard formula for the potential due
to an electron.
29
R. E. Hunt, 2002
rR
rR
rG . n dl =
r=R
g 0 (r) dl
r=R
= 2Rg 0 (R)
=)
g 0 (r) =
1
2r
=)
g(r) =
1
ln r + constant.
2
1
ln |x
2
x0 | + constant.
Example: consider an infinitely long charged wire in three dimensions lying along the
z-axis, with a charge density of per unit length. What is the electric potential around
the wire?
We assume the wire to be mathematically perfect, i.e., of infinitesimal width. Then
the electric charge distribution, in 3D, is = (x) (y). (Check that this gives the
correct result for the amount of charge in a unit length of the wire.) But it is clear
that this problem is fundamentally two-dimensional, with = (x) where x = (x, y);
and the potential satisfies r2 = (x)/0 . Hence the potential is (up to an arbitrary additional constant) just given by an appropriate multiple of the two-dimensional
fundamental solution, namely
p
=
ln |x| =
ln x2 + y 2 =
ln r
20
20
20
where r is the perpendicular distance to the wire (i.e., the r of cylindrical polar coordinates rather than of spherical polars).
2.6
We can use the fundamental solution to find Greens function in some simple geometries,
using the method of images. We shall find a function which satisfies the equation and
the boundary conditions; by uniqueness, this must be the Greens function.
30
R. E. Hunt, 2002
x0 )
8 x 2 D,
G=0
on x = 0,
G!0
as |x| ! 1, x 2 D.
Consider the solution in all space for the point source at x = x0 together with another (imaginary) source of strength 1 at the image
point x = x1 as shown:
=
4|x
x0 |
4|x
= (x
x0 )
(x
x1 |
and
r2
x1 )
by superposition of the two fundamental solutions. This certainly satisfies the requirement r2 = (x x0 ) for all x 2 D, because (x x1 ) 0 8 x 2 D. It also satisfies
! 0 as |x| ! 1; and on x = 0, |x x0 | = |x x1 | so that = 0. Hence by uniqueness,
1
1
1
G(x; x0 ) = =
.
4 |x x0 | |x x1 |
Example: A 2D quarter-plane x > 0, y > 0
x0 )
31
8x 2 D
R. E. Hunt, 2002
1
ln |x
2
x0 |
1
|x
ln
2 |x
x0 | |x
x1 | |x
1
ln |x
2
x1 |
1
ln |x
2
x2 | +
1
ln |x
2
x3 | + constant
x3 |
+ constant.
x2 |
Q
G(x; x0 ),
k
1
1
1
1
@
G(x; x0 )
@x
dy dz
x=0
which we can evaluate with some eort (see the worked example in the next section for
an example of this sort of evaluation).
32
R. E. Hunt, 2002
Alternatively, we can use the Divergence Theorem on the surface consisting of the
wall plus the hemisphere at 1. Since rT tends to zero on the hemisphere,
ZZ
ZZZ
( krT ) . n dS =
r . (krT ) dV
S
ZZZ
ZZZ
r2 T dV
Q
(x
k
x0 ) dV
= Q,
so the total heat radiated across the wall is Q.
Example: A point charge near an earthed boundary plate
Here
e
G(x; x0 )
0
where G is the Greens function for the 3D half-space x > 0.
=
Now the surface charge density induced on the plate is = 0 Ex (standard result from electrostatics,
where Ex is the x-component of E). The normal force (per unit area) on the plate, towards the charge,
is
2
@
e2 @G
2
1
1
1
E
=
E
=
=
,
x
2
2 0 x
2 0
@x
20 @x
and we calculate @G/@x as in the worked example in the next section. We can integrate this over the
whole plate (with considerable eort) to obtain the total force:
Z 1Z 1
e2
x20
e2
dy dz = =
.
3
20
160 x20
1
1 4 2 x2 + (y
y0 )2 + (z z0 )2
0
The force on the charge from the plate is equal and opposite, i.e., e2 /40 (2x0 )2 towards the wall. Note
that we could also have found this directly by considering the force on the charge due to the image
charge, ignoring the plate!
x0 )
G=0
in r < a,
on r = a.
33
R. E. Hunt, 2002
(so that a/|x1 | = |x0 |/a) with strength a/|x0 |, so Greens function is
1
1
a/|x0 |
G(x; x0 ) =
+
.
4
|x x0 | |x x1 |
(Check this by first showing that |x
when |x| = a.)
x1 |2 = (x
x1 ) . (x
x0 |2
Note that the same result holds if we consider the domain r > a instead.
Example: Images in a circle
This is the 2D equivalent of the above. The image point is at x1 = (a2 /|x0 |2 )x0 as before,
but now the strength of the image is just 1, so the Greens function is
1
1
G(x; x0 ) =
ln |x x0 |
ln |x x1 | + constant
2
2
1
|x x0 |
ln
+ constant.
2 |x x1 |
Choosing the constant correctly, we can ensure that G = 0 on the circle r = a.
=
2.7
The most important application of Greens function is that it can be used to find the
solution of Poissons equation with an arbitrary source distribution.
Greens Identity
For any smooth functions and , Greens Identity is
ZZZ
ZZ
2
2
( r
r ) dV =
( r
V
r ) . n dS
34
R. E. Hunt, 2002
where V is a volume with surface S. Prove this by applying the Divergence Theorem to
the vector field F = r
r , and using r . ( r ) = r . r + r2 .
The RHS is also written
ZZ
@
@n
@
@n
dS.
in V ,
=f
on S.
Apply Greens Identity, taking to be the Greens function G(x; x0 ) for the problem:
ZZZ
ZZ
2
2
( r G Gr ) dV =
( rG Gr ) . n dS
V
(x
=)
ZZZ
x0 )
(x
=)
(x0 ) =
x0 ) dV =
ZZZ
G dV +
ZZZ
(x)G(x; x0 ) dV +
0
ZZ
@G
dS
@n
ZZ
f (x)
@G
dS.
@n
35
R. E. Hunt, 2002
This latter result is easy to understand in many physical situations. For instance,
consider an arbitrary electrostatic charge distribution (x). Then
r2
in R3 ,
/0
!0
as |x| ! 1.
(We assume here that the charge distribution decays rapidly far from the origin.) Using the integral solution of Poissons equation, with V = R3 , and setting G to be the
fundamental solution in 3D,
ZZZ
(x)
(x0 ) =
dV.
40 |x x0 |
R3
We can interpret this physically as the superposition of many infinitesimal charge elements
(x) dV . Each of these is eectively a point charge, and the potential at x0 from such
a point charge (using the standard formula for the electrostatic potential due to a point
charge) is just (x) dV /40 |x x0 |. Summing over all such infinitesimal elements gives
the above result.
2.8
Finite Dierences
Applying Taylors theorem to any smooth function f we obtain
f (x + x) = f (x) + x f 0 (x) +
1
2
x2 f 00 (x) +
1
6
x3 f 000 (x) + O( x4 ),
x f 0 (x) +
1
2
x2 f 00 (x)
1
6
x3 f 000 (x) + O( x4 ).
f (x
x) = f (x)
We deduce that
f (x + x) f (x)
+ O( x);
x
hence f (x + x) f (x)
x is a first order forward finite dierence approximation to
0
f (x). (First order because the error term, known as the truncation error, is O( x).)
Similarly, by subtracting the two Taylor expansions above, we obtain
f 0 (x) =
f (x + x) f (x
2 x
giving us a second order central finite dierence.
f 0 (x) =
x)
+ O( x2 ),
The same reasoning allows us to find approximants for higher derivatives: for example,
a second order central finite dierence for the second derivative which we shall use for
Poissons equation is
f 00 (x)
f (x + x)
2f (x) + f (x
x2
36
x)
.
R. E. Hunt, 2002
We can use this reasoning to find approximants of as high an order as we like: for instance
f 0 (x) =
f (x + 2 x) + 8f (x + x) 8f (x
12 x
x) + f (x
2 x)
+ O( x4 ).
i,j
x2
i 1,j
i,j+1
i,j
y2
i,j 1
= (xi , yj )
at each of the interior points 0 < i < m, 0 < j < n. In addition, we will have boundary
conditions (Dirichlet, Neumann, or a mixture of the two) at i = 0, m and j = 0, n.
We therefore have a large number of simultaneous linear equations: at every point of
the grid, both interior and on the boundary, there is a corresponding equation, so that
we have a total of (m + 1)(n + 1) equations to solve.
It is usual to take x = y, which we shall assume from now on, in which case the
discretization at interior points reduces to
i+1,j
i 1,j
i,j+1
i,j 1
i,j
= (xi , yj ) x2 .
37
R. E. Hunt, 2002
What do we do for non-rectangular domains? Suppose we wish to find the steady-state temperature T
in an annulus a r b. There are two possible simple approaches. The first is to change coordinates
to plane polar coordinates (r, ), whereupon the grid becomes rectangular: 0 2, with an extra
boundary condition that T (r, 0) = T (r, 2). The second approach is to approximate the boundary of the
annulus using short line segments in only the x- and y-directions. More advanced techniques can use
grids with non-rectangular elements.
At each of the four interior points we apply the stencil. On the lower, left and right
boundaries we have simply
0,0
1,0
2,0
3,1
= 13 ,
3,0
= 0;
3,2
= 23 ,
38
0,1
3,3
0,2
0,3
= 0;
=1
R. E. Hunt, 2002
respectively. On the top boundary we must use a finite dierence approximation for
@ /@y to obtain
1,3
1,2
= x21 y =
1
,
27
2,3
2,2
= x22 y =
4
.
27
We can therefore gather together the 16 simultaneous equations representing each point
of the grid in a single equation involving a 16 16 matrix as follows:
0
1
B0
B
B
B0
B
B0
B
B
B0
B
B0
B
B
B0
B
B0
B
B0
B
B
B0
B
B0
B
B
B0
B
B0
B
B0
B
B
@0
0
0
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
4
1
0
0
1
0
0
0
0
0
0
0
0
0
0
0
1
4
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
1
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
4
1
0
0
1
0
0
0
0
0
0
0
0
1
0
0
1
4
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
1
0
10
0
B
0C
CB
CB
B
0C
CB
C
0C B
B
CB
0C B
CB
B
0C
CB
CB
0C B
CB
B
0C
CB
B
0C
CB
CB
0C B
CB
B
0C
CB
CB
0C B
CB
B
0C
CB
C
0C B
B
CB
0A @
1
0 1
0
C B0C
B C
1,0 C
C B C
B C
2,0 C
C B0C
C B0C
3,0 C
B C
C B C
B0C
0,1 C
C B2C
C
B C
1,1 C
B 27 C
C B2C
B C
2,1 C
C
C B 27
C
B1C
3,1 C
3 C
B
C = B0C
0,2 C
B C
C B4C
C
B C
1,2
C B 27 C
C B4C
2,2 C
B 27 C
C B2C
C
B C
3,2
C B3C
C B0C
0,3 C
B C
C B1C
B C
1,3 C
C B 27
4 C
@ 27
A
2,3 A
0,0
3,3
In theory, the problem is now solved: we simply use Gaussian elimination to obtain
0,3
= 0,
1,3
= 19 ,
0,2
= 0,
1,2
0,1
= 0,
1,1
0,0
2,3
= 49 ,
2
,
27
2,2
1
,
27
2,1
1,0
2,0
3,3
= 1,
8
,
27
3,2
= 23 ,
4
,
27
3,1
= 13 ,
3,0
= 0,
However, in real applications Gaussian elimination is simply not possible. There are
N = (m + 1)(n + 1) unknowns i,j ; the matrix in the problem has N 2 elements, and
the number of calculations required to perform a full Gaussian elimination is O(N 3 ). For
39
R. E. Hunt, 2002
even a modest grid size (say 100 100) the number of calculations required would be
around 1012 , which is unreasonable.
Fortunately, we note that the matrix has a band diagonal (or banded ) structure:
beyond a certain distance from the leading diagonal, all entries are zero, so that only a
small number of the diagonals in the matrix are non-zero. Because of the widespread need
to solve Poissons and related equations efficiently, many specialised algorithms have been
developed for such matrices, taking advantage of the large number of zeros to optimise the
elimination process. The number of calculations required by these algorithms is typically
only O(N ).
2.9
Another approach to solving the discretized form of Poissons equation is to use an iterative method. These so-called relaxation methods start with some initial guess at the
solution (which does not need to be a good guess: it could be simply all zeros!) which is
then allowed to move slowly towards the true solution. Note that such methods therefore
involve two kinds of errors: those caused by the fact that the finite dierences used in
the discretization are not exact, and those caused by the fact that the initial guess never
quite reaches the true solution.
Relaxation methods are typically a little slower than the matrix methods discussed
above, but require significantly less computer memory. This can be an important consideration when high accuracy is required and a small grid spacing is being used.
i,j
= 14 (
i+1,j
i 1,j
i,j+1
i,j 1
(xi , yj ) x2 ).
4. For every interior grid point, replace the old approximation i,j with i,j .
5. Repeat from step 2 until the dierence between the latest two approximations is
smaller than some set tolerance everywhere.
40
R. E. Hunt, 2002
14 (
i+1,j
i 1,j
i,j+1
(xi , yj ) x2 ),
i,j 1
i.e.
i+1,j
i 1,j
i,j+1
i,j 1
i,j
(xi , yj ) x2 .
i,j
= (1
!)
i,j
+ 14 !(
i+1,j
i 1,j
i,j+1
i,j 1
(xi , yj ) x2 )
where ! is the relaxation parameter. The value ! = 1 gives the GaussSeidel method
again; ! < 1 would produce under -relaxation, where we keep a proportion of the old
solution; and ! > 1 produces over -relaxation where we actually move further away from
the old solution than we would using GaussSeidel.
The best value to use for ! depends on the particular problem being solved, and
may also vary as the iterative process proceeds. However, values in the range 1.2 to 1.4
typically produce good results, and in some cases it is possible to determine an optimal
value analytically. The number of iterations required using SOR is significantly less than
for either Jacobi or Gauss-Seidel, and for large grids it is often the most practical of all
methods of solution.
More advanced methods exist with even better convergence rates, such as the multigrid method which simultaneously uses a number of dierent grids with dierent grid
41
R. E. Hunt, 2002
spacings. However, the programming eort required becomes much greater, and so these
advanced methods are usually implemented using black box routines written by experts
in numerical analysis.
2.10
The methods developed for numerical solution of Poissons equation are easily extended
to the diusion equation
@
= kr2 .
@t
For simplicity we shall consider a rectangular two-dimensional domain 0 x 1,
0 y 1, with some initial value for (x, y, t) at t = 0.
We first introduce a spatial grid just as for Poissons equation, with x = y, and
denote our approximation to (xi , yj , t) by i,j (t). Using the standard finite dierence
formulae for the spatial second derivatives we obtain
d
i,j
dt
= k(
i+1,j
i 1,j
i,j+1
i,j 1
i,j )/
x2
at each interior point. Thus we have (n 1)2 coupled ordinary dierential equations,
which we can solve using various methods. However, we need to take care to ensure that
the method we use leads to a stable solution.
(k)
i,j
t
for the finite dierence approximation to @ /@t, we obtain the iterative scheme
(k+1)
i,j
(k)
i,j
+ (
(k)
i+1,j
(k)
i 1,j
(k)
i,j+1
(k)
i,j 1
(k)
i,j )
where
=
k t
x2
is the Courant number. We can now, starting from the known values of
(1)
in time to i,j and so on.
42
(0)
i,j ,
step forward
R. E. Hunt, 2002
To analyse the stability, we examine the behaviour of our approximate solution i,j
as k ! 1. In the exact solution, remains bounded as t ! 1; that is, it does not grow
ever larger with time ( is simply diusing from one place to another, and there are no
source terms). Will this also be true of our approximate solution?
Before we can make progress we need to introduce the concept of a Fourier mode.
Any function f (x) can be expressed using Fourier transforms as
Z 1
1
f (x) =
fe(k)eikx dk;
2 1
that is, as an sum (in fact, an integral) of individual modes of the form eikx , each weighted
by the appropriate factor fe(k)/2. By taking a double Fourier transform in both x and
y, we can therefore write the initial value (x, y, 0) as a weighted sum of individual
modes of the form eix ei y where and are the wavenumbers in the x- and y-directions
respectively. We shall investigate the evolution of each such mode individually. If any
one mode grows uncontrollably with time then the complete solution (the sum of the
weighted modes) will do too.
This approach works because the diusion equation @ /@t = r2 is linear, which means that individual
modes do not interfere with each other. In a nonlinear system, modes could interact and so it would
be impossible to investigate them individually. For example, consider the ordinary dierential equation
y 0 = y 2 ; if we tried using a sum of Fourier modes then the y 2 term would introduce lots of cross-products
(because, for example, (eik1 x + eik2 x )2 is not the same as just the sum of the squares of the individual
two modes eik1 x and eik2 x ).
(0)
i,j
So how would
(1)
i,j
= eixi ei
yj
= eixi ei
+ (eixi+1 ei
= 1 + (ei
+e
= 1 + 2(cos x
=
i x
yj
yj
+ ei
+e
1) + 2(cos
yj
y
+ eixi ei
yj+1
4) eixi ei
+ eixi ei
yj
4eixi ei
yj
yj
(0)
i,j
1)
(0)
i,j
4 sin2 12 x
4 sin2
1
2
x.
= k
(0)
i,j .
Now, if 14 then 1 1 for all and ; so as the mode evolves it does not grow in
amplitude. However, if > 14 then for some values of and we have || > 1, so that as
43
R. E. Hunt, 2002
the mode evolves it grows larger and larger in amplitude, eventually tending to infinity
as k ! 1. Hence we deduce that the condition for stability of the Euler method for the
diusion equation in two dimensions is 14 : this is known as the Courant condition.
A similar analysis in one dimension leads to the Courant condition 12 . It is easy
to extend the method, and the stability analysis, to n dimensions.
In practice, we find that it is difficult to ensure stability using the Euler method if we
want both accuracy and a reasonable speed of calculation. Recall that = k t/ x2 ; if
x is small then we will require an extremely small value of t to ensure stability. Such
a small value means that a very large number of iterations will be required to calculate
the solution at a given time t.
We have ignored in this discussion the boundary conditions. These do, of course, have an eect on the
evolution of the solution, but in general their eect is of subsidiary importance if the method we are
using is unstable. Because of the way the finite dierence equations have been set up, at k = 1 only those
points in the grid which are immediately next to a boundary will have been aected by the boundary
conditions. At later times, for general k, only points in the grid which are k or fewer points away from a
boundary will have been aected. By the time the eect of the boundaries reaches a point in the interior
of the grid, therefore, the instability of the solution caused by the factor k will generally have already
taken hold. Nevertheless, it is possible for unusual boundary conditions to turn a method which would
otherwise have been stable into an unstable method.
R. E. Hunt, 2002
The above errors, caused by our numerical method itself, would be the only errors in a perfect computer
implementation of the Euler method. However, for any real computer there will also be rounding errors,
caused by the computers inability to store real numbers with infinite precision. These rounding errors
are usually significantly smaller than the truncation error; but if the step-size is reduced too far, rounding
errors may become significant. Hence it may be impossible to get the precision we desire, however small
we make the step-size.
(k+1)
i+1
(k+1)
i
(k+1)
i 1
(k)
i
(k+1)
(k)
i+1
(k)
i
(k)
i 1
(k)
1 2 sin2 12 x
1 + 2 sin2 12 x
(0)
i .
The amplification factor in this case satisfies |(1 2 sin2 12 x)/(1 + 2 sin2 12 x)| < 1
for all values of , and , so the method is unconditionally stable. The practical
implication is that the step-size in time, t, may be chosen independently of the spatial
step-size x: both may be chosen on the grounds of truncation error alone and we are
not hampered by stability requirements.
45
R. E. Hunt, 2002