See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/267508824

Numerical Solutions to Poisson Equations Using the Finite-Difference Method

[Education Column]

Article  in  IEEE Antennas and Propagation Magazine · August 2014

DOI: 10.1109/MAP.2014.6931698

CITATIONS READS

5 2,265

1 author:

James R Nagel
Institute of Electrical and Electronics Engineers
30 PUBLICATIONS   177 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Electrodynamic Sorting of Nonferrous Metals and Alloys View project

All content following this page was uploaded by James R Nagel on 03 November 2014.

The user has requested enhancement of the downloaded file.

 Numerical Solutions to Poisson Equations
Using the Finite-Difference Method

James R. Nagel

Terahertz Device Corporation

Salt Lake City, Utah 84124 USA
E-mail: nageljr@ieee.org

Abstract

The Poisson equation is an elliptic partial differential equation that frequently emerges when modeling electromagnetic
systems. However, like many other partial differential equations, exact solutions are difficult to obtain for complex
geometries. This motivates the use of numerical methods in order to provide accurate results for real-world systems.
One very simple algorithm is the Finite-Difference Method ( F DM), which works by replacing the continuous derivative
operators with approximate finite differences. Although the Finite- Difference Method is one of the oldest methods ever
devised, comprehensive information is difficult to find compiled in a single reference. This paper therefore provides
a tutorial-level derivation of the Finite- Difference Method from the Poisson equation, with special attention given to
practical applications such as multiple dielectrics, conductive materials, and magnetostatics.

Keywords: Partial differential equations; numerical simulation; finite difference methods; electrostatics; magnetostatics

1 . Introduction r, Equation (1) simplifies into the familiar, classical fonn of

Poisson's equation, given by

W
hen studying the behaviors of complex physical sys­
tems, we frequently desire knowledge about the electric (2)
and magnetic field distributions produced by unique spatial
arrangements of charge and current. Such information can be with/being called the/arcing/unction. For the case of / = 0,
especially useful for designers who intend to optimize the per­
Equation (2) then takes on another familiar form:
formance of sophisticated electrical devices. For many situa­
tions, this usually requires the solution of an advanced partial
differential equation under specified boundary conditions. (3)
Unfortunately, exact analytical solutions are difficult to obtain
for most geometries of practical interest. We must therefore which is the well-known Laplace equation.
resort to approximation methods, if we ever expect to obtain
any useful information about them at all. Like most partial differential equations, closed-form solu­
tions to the Poisson and Laplace equations only exist for a
One expression of particular interest is known as the var i­ small handful of simplistic models. Many computational algo­
able-coefficient Poisson equation, and this frequently emerges rithms exist for generating approximate numerical solutions,
within the context of electrostatics and magnetostatics. In its but one of the oldest and simplest of these is called the Finite­
most general form, this is given as Difference Method (FDM) [1-5]. As the name implies, the
Finite-Difference Method works by replacing the continuous
V{u(r)V¢(r)] /(r),
= (1) derivative operators of Poisson's equation with finite-differ­
ence approximations, and then numerically solving the resul­
tant system. The accuracy of the Finite-Difference Method is
where u, ¢, and/are scalar functions of the position vector r. therefore directly tied to the ability of a finite grid to approxi­
The functions u and/are assumed to be known, with ¢ being mate a continuous system, and any errors may be arbitrarily
the unknown function for which we wish to solve under specific reduced by simply decreasing the spacing between spatial
boundary conditions. For the special case when u lover all
= samples.

IEEE Antennas and Propagation Magazine, Vol. 56, No.4, August 2014 209
 Despite the relative simplicity of the Finite-Difference sion lines [5], as we shall demonstrate in this paper. If desired,
Method as a nwnerical tool, many graduate-level textbooks expansion into three dimensions may be accomplished by
tend to limit themselves only to the context of Equations (2) simply following a parallel procedure to this work, although
and (3) [6 -8 ]. Details on Equation (1) are therefore very scarce, the memory requirements can quickly overwhelm even a high­
even though the resultant algorithm is virtually identical to the end desktop computer if one is not careful. Although many of
simplified cases. Part of the blame for this may be because the these problems can be alleviated through the use of sophisti­
Finite-Difference Method is, at its mathematical core, little cated matrix-inversion algorithms [13], such topics are beyond
more than a specialized case of the well-known Finite-Element the scope of this discussion.
Method (FEM) [9]. The primary difference between them is
that the Finite-Difference Method is generally solved within a
fixed, rectangular mesh of grid samples, while the FEM utilizes
2. Poisson's Equation in Electrostatics
a flexible, triangular mesh at arbitrary grid locations. From a
purely computational perspective, this makes the FEM more
efficient due to the freedom to allocate samples within a scarce We begin with Gauss's law in point form, given as
simulation domain. Many commercial software packages
therefore rely heavily on the FEM, with generally little attention V.D(r)=p(r). (4)
paid in proportion to its Finite-Difference-Method cousin.
In this context, r = xx + yy is a position vector in two-dimen­
Unfortunately, the numerical flexibility of the FEM also
sional space, p is the charge-density function, and D is the
comes at the price of greater complexity. One consequence of
electric-flux density. Using the constitutive relation
this fact is a far steeper learning curve for students who desire to
implement a functional computer program on their own. While D(r)= G0 6"r( r)E(r), Gauss's law may be rewritten in terms
the FEM may require months of practice in order to achieve of the electric-field intensity, E, as
proficiency, much of the same conceptual understanding and
functionality can easily be gleaned from only a few weeks
(5)
of instruction on the Finite-Difference Method. The Finite­
Difference Method therefore serves as a desirable option for
introductory courses in numerical methods, and indeed many
universities use it as a conceptual springboard from which to
where Gr (r) is the spatially-varying dielectric constant, and
1
introduce the more complex nature of the FEM. This also makes GO = 8.854xl0 - 2 F/m is the permittivity of free space. If we
the Finite-Difference Method a desirable algorithm for the "do­ assume that our system is perfectly static (zero frequency), then
it-yourself' researcher, who needs a functional nwnerical code the electric field, E, is related to the voltage potential function,
but does not necessarily wish to purchase expensive commercial V, by the expression
licenses based on the FEM. Consequently, the Finite-Difference
Method still finds periodic use in modern research applications, E(r)=-VV(r) . (6)
such as chafe modeling of aircraft wiring [10, 11] and crosstalk
modeling in multiconductor transmission lines [12].
Substitution therefore leads us to
The goal of this paper is to serve as a comprehensive
tutorial on the principles of the Finite-Difference Method as (7)
applied to the Poisson and Laplace equations, with a special
focus placed on the variable-coefficient form in Equation (1).
We therefore begin by deriving the classical Poisson and which we can see is Equation (1), the variable-coefficient
Laplace equations from Maxwell's equations, and then show Poisson equation. On some occasions, it is useful to asswne a
how finite-differencing leads to a straightforward numerical uniform dielectric function with the form Gr (r)= Gr ' This
solution in the form of a matrix-vector equation. We build
leads us to
on those basic principles to show how the Finite-Difference
Method can reliably model the effects of a varying dielectric
constants by solving the variable-coefficient Poisson equation. (8)
Finally, we expand the discussion to show how variations on the
Poisson equation appear in broader contexts, including material
conductivity in quasistatic systems and magnetic permeability which is Equation (2), the classical form of Poisson's equation.
of magnetostatics.
Although Equation (8) is well understood and simple
For the sake of simplicity, we shall strive to limit this to numerically solve, the asswnption of a constant dielectric
tutorial to only two-dimensional systems. This helps to greatly function also limits its practical utility. Equation (7) is the
reduce the complexity of the derivation, without sacrificing true expression that needs to be solved if we ever expect to
any fundamental principles of the algorithm itself. A common, properly model a more realistic physical system. Even so, the
practical application of two-dimensional modeling is the cal­ classical form of the Poisson equation still provides a useful
culation of per-unit-length impedance parameters of transmis- starting point from which to demonstrate many fundamental

210 IEEE Antennas and Propagation Magazine, Vol. 56, No.4, August 2014
principles of the Finite-Difference Method before advancing ( 1 1)
to a more complex system. We shall therefore begin by using
the classical Poisson equation as a demonstration case for the
In a similar fashion, we may also define the charge-density
Finite-Difference Method, and then expand our algorithm to the
generalized form. samples along the same mesh by using the p(n, m ) notationl.

The next step is to expand the Poisson equation by explic­

itly showing the partial derivatives in space:
3. The Five-Point Star

The first step in applying the Finite-Difference Method ( 1 2)

is to define a mesh, which is simply a set of spatial points at
which the voltage function will be sampled. Figure 1 shows
a basic rectangular pattern of uniformly spaced grid samples, The reason for doing this is so that we may now approximate the
and is actually one of the key defining features of the Finite­ derivative operators through the use of finite differences. The
Difference Method with respect to the FEM. As noted earlier, standard, three-point approximation for the second derivative
the key advantage with the FEM is a capacity to arbitrarily is therefore given as
sample the mesh throughout the domain of interest, but this also
naturally complicates the resultant algorithm. a2 V(n - 1, m ) - 2V(n, m ) + V (n + 1, m )
- V(n, m ) :::::; , ( 1 3)
ax 2 h2
Letting h be the distance between each sample, the points
that lie on the mesh may be defined by with a similar expression for the partial derivatives with respect
to y. Plugging back into Equation (12) then gives us
(9)
V(n - 1, m ) + V (n + 1, m ) + V(n, m - 1 )
and
h2
+ V ( n, m + 1 ) - 4V ( n, m ) = --p(n, m ) . ( 14)
Ym = mh ( 1 0) GO

where nand m are integers. In practice, nand m will eventually What Equation ( 14) tells us is that every voltage sample
need to be treated as indices for a matrix of voltage samples, V (n, m ) is dependent only on p ( n, m ) and its four nearest
so it helps to use a shorthand notation that bears this in mind. neighbors. A graphical depiction of this is called a computa­
We shall therefore replace the spatial coordinates with a simple tional molecule, and is shown in Figure 2. Because of its unique
index notation by assuming the following convention: geometry, this five-point stencil is often referred to as the five­
point star. The accuracy of Equation ( 14) is then limited only
by the choice of grid spacing, h. As h ---+ 0 , we thus may expect
any errors in this approximation to accordingly vanish.

h 4. Boundary Conditions
III •
Ym+l • • •
Because a computer can only store a finite number of grid
points, it is always necessary to truncate a simulation domain
along some fixed boundary. However, since the five-point
Ym • • • star is not applicable at the boundary samples, it is necessary
to specify boundary conditions in order to produce a unique
solution. The two most basic forms of boundary condition are
called the Dirichlet boundary and the Neumann boundary,
Ym-l although more sophisticated examples do exist. In practice, it is
• • • common for simulations to employ a mixture of these two

1 Another common notation is the subscript Vn

m convention.
However, when using indexed matrix notation, the row (or y
Xn-l Xn Xn+l coordinate) typically comes before the column (or x coordinate).
In practice, it can therefore sometimes be useful to habitually

Figure 1. The mesh points for the Finite-Difference Method express the coordinates backward, such as V ( m, n ) , when
grid. writing one's actual source code.

IEEE Antennas and Propagation Magazine, Vol. 56, No.4, August 2014 211
 V(n,m+l ) The simplest boundary condition is the Dirichlet bound­
ary, which may be written as

(15)

where the function/is a set of known values for V along the

V(n-l ,m) V(n+l ,m) domain of boundary points given by QD' A good example of
such a condition occurs in the presence of charged metal plates.
Because all points on a metal are at the same potential, a metal
plate is a good example of a region of points that can be modeled
by a Dirchlet boundary. However, this can have important
V(n,m-l ) ramifications for the numerical accuracy of a given model. For
example, image theory states that when a charge distribution is
placed next to a grounded metal plane, then the resultant field
Figure 2. The computational molecule for the five-point star.
profile is equivalent to that of an equal, opposite charge placed
behind that same planar interface. Consequently, Dirichlet
boundaries along the edge of a simulation domain will behave
like charged metal plates. In fact, if all boundaries are fixed to
0.0 V, then the simulation boundaries will behave as an infinitely
periodic repetition of the original domain, with alternating
signs on the charge-density profile.

In contrast, the Neumann boundary condition exists

when the gradient of the potential function is a known, fixed
value. This condition is usually only applied to the edges of the
simulation domain by fixing the derivative with respect to the
outward unit normal. One way to mathematically express this is
by writing

(16)

where n is the outward-pointing unit normal vector, and /'

specifies the derivative function along QN' Note that this is
also equivalent to specifying the normal component to the
electric field at the domain's boundary. However, unlike the
Figure 3. An example of a simulation domain with mixed
Dirichlet condition, the Neumann condition does not offer a
boundary conditions. The shaded regions marked by QD
direct solution to the voltage potential, but instead depends on
are specified as Dirichlet boundary conditions. The regions the values of neighboring samples. Neumann boundaries must
marked by QN are Neumann boundaries. therefore be expressed in terms of the surrounding points by
applying a modified stencil. One simple method for expressing
this is by imagining a finite-difference approximation between
some boundary sample, Vb' and the first inner sample, V!, such
conditions, so it is helpful to define QD and QN as the sets of
that
all points that satisfy the Dirichlet and Neumann conditions.

Figure 3 shows an illustration of a simulation grid that

Vb-V!
=/'. (17)
utilizes a mixture of both Dirichlet and Neumann boundary h
conditions. The left and right grid boundaries are indicated as
Neumann conditions, while the top and bottom are Dirichlet. For example, if /' = 1.0 Vim at the left-most boundary of a
Notice also that Dirichlet boundaries may be specified inside simulation ( n = 1 ), then the Neumann boundary condition at
the simulation region itself, and do not necessarily have to the mth row would be written as
reside at only the outer edges (Neumann boundaries cannot
easily emulate this effect, and so usually stay at the outer rim). V(1,m)-V(2,m)
---'�--'----'--
---'- = 1.0 Vim. (18)
Finally, note that the corners of the simulation domain do not h
influence any points around them, and can therefore be speci­
fied to any arbitrary condition one desires (or even ignored When placed at the outside border of the simulation domain, the
entirely). Neumann boundary of 0.0 V1m is periodic, as well. However,

212 IEEE Antennas and Propagation Magazine, Vol. 56, No.4, August 2014
this time the charges will retain a consistent sign rather than For the Neumann boundaries, we likewise have
alternate, like the Dirichlet boundary.
V(I,2)-V(2,2) 0 , = (23)
Often times, it is desirable to model an isolated charge
distribution that is infinitely far away from any other objects.
V(I,3)-V(2,3) 0 , = (24)
This requires the use of a free-space boundary condition,
sometimes called an open boundary condition or an absorbing
boundary condition. One way to achieve this is by simply using V(4,2)-V(3,2) 0 , = (25)
Dirichlet boundaries at the simulation edge, with a large void of
empty space in between. As the boundaries are pushed farther V(4,3)-V(3,3) = O. (26)
away, the simulation will naturally approximate a free-space
condition with better accuracy. However, this is obviously a Finally, we apply the five-point star along all interior points to
brute-force approach that does not take long to overwhelm most find
computational resources. This has motivated the use of more
sophisticated open boundaries, such as the coordinate-stretching
V(2,1)+ V(I,2)-4V(2,2)+ V(3,2)+ V(2,3) 0 , (27) =

method [14], but such techniques are beyond the scope of this
work. It also emphasizes one of the primary advantages of
the FEM over the Finite-Difference Method, which can save V(3,1)+ V(2,2)-4V(3,2)+ V(4,2)+ V(3,3) 0 , (28) =

memory with ease by simply sampling the empty void more

sparsely than the physical object of interest. V(2,2)+ V(1,3)-4V(2,3)+ V(3,3)+ V(2,4) 0 , (29) =

V(3,2)+ V(2,3)-4V(3,3)+ V(4,3)+ V(3,4) = O. (30)

5. Example: Parallel-Plate Capacitor

To see an example of the Finite-Difference Method at

work, consider the simple, 4 x 4 grid of voltage samples +10V V(24)
) ,
V(34) +10V
depicted in Figure 4. The top boundary is a Dirichlet boundary 'II" 'II"
fixed at 1.0 V, with the bottom boundary fixed to -1.0 V. The
left and right boundaries are Neumann boundaries fixed to a
derivative of 0.0 V 1m with respect to the outward normal. Using .. �V(1)3) ..�V(2)3) ..�V(3,3) ...V(4,3)
the Finite-Difference Method, it is our desire to solve for the 'II" 'II" 'II" 'II",
voltage potentials at all of the indicated points.

The first step in applying the Finite-Difference Method is ..�V(1)2) ..�V(22

, ) ..�V(3,2) ...V(4,2)
to express the boundary conditions in terms of their explicit 'II" 'II" 'II" 'II",
assignments. For the Dirichlet boundaries, these are simply

V(2,1) -1.0 V,
= (19) -1.0V .. �V(2,1) .. �V(3,1) -1.0V

V(3,1) -1.0 V,
= (20) Figure 4. A 4 x 4 grid of voltage samples within a parallel­
plate capacitor. The top plate is fixed at +1.0 V while the
V(2,4) +1.0 V,
= (21) bottom plate is at -1.0 V. The left and right boundaries are
Neumann, thus mimicking an infinitely periodic boundary
with consistent sign on the charge. The four corners do
V(3,4) +1.0 V.
= (22)
not influence any points around them, and are therefore
neglected.

IEEE Antennas and Propagation Magazine, Vol. 56, No.4, August 2014 213
Notice how the samples along the corners of the domain have will need to be 10,000 xl 0,000 elements. Because Gaussian
no impact on the interior values. It therefore makes no differ­ elimination is such an intense operation, it is easy to see how
ence what is done with these samples, or even whether or not even small simulations can quickly require excessive
they are included as part of the solution. We may therefore computational resources. Fortunately, for the Finite-Difference
neglect these points entirely from further consideration. Method, inspection of Equation (31) quickly reveals that A is
a heavily sparse, banded matrix. In fact, for any two-dimen­
At this point, there are now 12 linear equations with 12 sional simulation, each row of A only has five nonzero ele­
unknowns, thus implying a matrix-vector equation with the ments, at most. This allows us to efficiently arrive at solutions
form Ax = b . Writing this out in its entirety leads us to through the use of advanced matrix solvers that take advantage
of such properties [13]. For example, one especially popular
method is successive over-relaxation [6]. Although the process
1 0 0 0 0 0 0 0 0 0 0 V (2,1)
0 is iterative by nature, it does have the advantage of being rather
0 1 0 0 0 0 0 0 0 0 0 V(3,1)
0 simple to learn, and therefore fitting for the classroom setting,
0 0 1 -1 0 0 0 0 0 0 0 V (1,2)
0 where time is scarce and experience from the students is limited.
1 0 1 -4 1 0 0 1 0 0 0 V(2,2)
0
0 1 0 1 -4 1 0 0 1 0 0 V(3,2)
0
6. Electric Fields
0 0 0 0 -1 1 0 0 0 0 0 V(4,2)
0
0 0 0 0 0 0 1 -1 0 0 0 V (1,3)
0 From the basic definition E = -v V , it is a straightforward
0 0 0 1 0 0 1 -4 1 0 1 0 V(2,3
) process to reconstruct the electric fields from a simulation
0 0 0 0 1 0 0 1 -4 1 0 1 V(3,3
) output. We therefore begin by expanding the gradient operator
0 0 0 0 0 0 0 0 -1 1 0 0 V(4,3) into its constituent x and y components:
0 0 0 0 0 0 0 0 0 0 1 0 V(2,4) _iav(r)_yav(r)
0 0 0 0 0 0 0 0 0 0 0 1 V(3,4) E(r)= . (32)
ax ay
-1
-1 The next step is to apply a central-difference approximation to
0 the individual field components, using
0
V(n + l,m)- V(n,m)
0 (33)
h
0
(31)
0 V(n,m + 1)- V(n,m)
Ey (n,m)= (34)
0 h
0

Ex
0 Two observations are worth noting at this point. The first is that
the grid of samples contains one fewer element along the x
+1
+1
direction than does the V grid. Similarly, the Ey grid is
comprised of one fewer element along the y direction. This is a
Using Gaussian elimination, we now invert the system matrix natural result of applying the central-difference method to
compute a numerical derivative. The second point is that the x
and solve for x = A -lb to find and y components of the electric field are staggered from each

[-
other in space, as shown in Figure 5a. Ideally, we would like the
X = 1 -1 -� -� -� -� � 1 1
3 3
electric field components to land on the same point in space. We
may therefore average the field components together according
to the convention
Notice how this is simply a linear change in voltage potential
between the top and bottom plates, as is expected from an
infinite parallel-plate capacitor.
E� (n,m)= -.!..2 [Ex (n,m + 1)+ Ex (n,m)] , (35)

For relatively small simulation domains, the direct matrix

inversion demonstrated by this example works perfectly well E� (n,m)= �[Ey (n + l,m)+ Ey (n,m)J. (36)
for obtaining a solution. However, it is also important to bear in
mind that the size of A grows directly in proportion with the Figure 5b shows the new set of grid samples along the
square of the size of x. For example, given a rectangular primed coordinates. As desired, the new electric-field compo­
simulation domain of 100 x100 voltage samples, the matrix A nents are now placed at consistent grid locations along a stag-

214 IEEE Antennas and Propagation Magazine, Vol. 56, No.4, August 2014
 Although it is tempting to begin by directly applying numerical
derivatives to this expression, a far simpler process may be
derived if we first realize that 5r ( r ) does not necessarily need
Ey(n -1, to be sampled at the identical grid points as V ( r ). We shall
therefore begin by sampling the permittivities along the stag­
gered grid shown in Figure 6. Mathematically, this may be
written as

Ey(n-l,m- + I,m-I)
(38)

with V (n, m) and p(n, m) defined along the original grid

points as before. Each square region is then assumed to take on
a constant value for 5r (n, m) around the sampled location,
similar to that of a stair-step model.
V(n -1,m+ 1) V(n+l,m+l)
Defining the permittivities in this way has two important
advantages. First, it allows us to define the voltage samples
along the boundaries of the dielectric permittivities. This
is important when we compute the electric fields, since, as
was noted earlier, electric fields are discontinuous at planar
V(n-l, boundaries. The second reason is that it allows us to exploit the
properties of variational calculus by expressing Equation (37)
in its weak, or variational, form. In so doing, the second-order
derivatives vanish, and leave only first-order derivatives for us
to numerically approximate.
We begin by defining Qnm as the square region around a
V(n-l,m-
single voltage sample V (n, m) , as depicted in Figure 7a. We
then take the surface integral over Qnm to find
Figure 5. (a) A grid stencil for obtaining the electric-field
samples. The circles represent voltage samples, while the
x represent the electric-field samples obtained by the cen­
tral-difference approximation. (b) The staggered electric­
field grid obtained after averaging the central-difference where dQ= dxdy is the differential surface area. Looking first
components. Both x and y components now land on the at the right-hand side of Equation (39), we note that the integral
same point in space. over a constant block of charge density is simply

--1
Sf h2
p(r)dQ=--p(n,m). (40)
5 0 Qnm 5 0
gered grid from the voltage potentials. As we shall see in Sec­
tion 7, such an arrangement also avoids any of the confusion The left-hand side of Equation (39) may also be simplified by
that occurs at boundaries between dielectric surfaces, since applying the divergence theorem. When expressed in two
normal field components may be discontinuous along these dimensions, this converts the surface integral over Qnm into a
locations. We also note that the number of rows and columns
in the E-field grid is now one less than the rows and columns in flux integral around its outer contour, Cnm. We may therefore
the voltage grid. write

7. Varying Dielectrics
where do is the differential unit normal vector pointing out
Let us now return to the variable-coefficient Poisson equa­ along Cnm . Putting these results together therefore leads to
tion:

p(r) (42)
V{5r(r)VV(r)J=_ . (37)
50
IEEE Antennas and Propagation Magazine, Vol. 56, No.4, August 2014 215
V(n-1,m+ This expression is sometimes referred to as the weak form of
Poisson's equation, in contrast with the strong form given by
Equation (37).
With the desired expression in hand, the next step is to
V(n-1,
expand out the gradient operator into its constituent compo­
nents to find

= [
� &r(x,Y) a�V(x,Y)X +�V(x,Y)Y ] .dn. (43)
Cnm
x cy
V(n-1,m- (n+ 1,m- 1)
Note that in three dimensions, the surface Cnm would normally
Figure 6. The finite-difference mesh for the generalized be a cube. However, in our two-dimensional example, it is
Poisson equation. Each square represents a region of con­ simply a square contour. The total contour integral may
stant dielectric permittivity. therefore be broken down by treating it as a series of sub-inte­
grals around each side of the square. For brevity, we shall
simply write these sub-integrals as S, ... S4 :

This geometry is depicted in Figure 7b, which highlights the

contours of integration over all four sides, taken in the coun­
terclockwise direction.

To begin, let us define S, as the integral over the right

edge of the square, where d n = x dy . For convenience, it also
helps to assume that V (n,m ) lies at the origin of the domain,
since the end result will be equivalent no matter what location
we choose. We may therefore express S, as
V(n-l,m+

S1 =
hl2

f
[ a a
&r(x,y) -V(x,y)i +-V(x,y)y .idy
]
ax cy
-hl2
(45)
hl2
a
V(n-l, f
-h12
&r(x,y)-V(x,y) dy.
ax

We next approximate the partial derivative by using a central

difference between the two samples, V (n,m ) and
V (n +1, m ) . We may then assume that the derivative remains
constant across the entire contour2. Calculating the integral
V(n-l,m- + I,m-1) across the two dielectric regions therefore gives

Figure 7. (a) The geometry of integration for the Finite­

Difference Method. The volume element Qnm surrounds

the voltage sample V (n,m ). The outer contour Cnm

2From the perspective of the Finite-Element Method, this is
encloses Qnm and defines the contour integral of Equa­
exactly what happens when we apply a linearly interpolated
tion (41) in the counterclockwise direction. (b) The total
shape function between voltage samples. The normal
contour Cnm is broken down into four sub-contours, labeled
component of the gradient is always a constant value over the
S, , S2' S3' and S4' contour, while only the tangential component varies.
216 IEEE Antennas and Propagation Magazine, Vol. 56, No.4, August 2014
 s, [ � ][
= h Sr(n,m-l +Sr(n,m) V(n+l,m -V(n,m) � ] 8. Example: Capacitance Per Unit Length

Consider the cross-sectional view of a standard coaxial­

(46) cable transmission line, as shown in Figure Sa. The cable
= (1/2)[sr(n,m-1)+sr(n,m)][V(n+l,m)-V(n,m)]. geometry is defined by an inner wire radius, Ij, an inner shield
radius, r2' and an outer shield radius, r3' Embedded between
Notice how this simply evaluates to a finite difference between the inner conductor and the outer shield is a layer of dielectric
the voltage samples, weighted by the average dielectric constant insulation with relative permittivity sr' relative permeability
between them. Carrying out this same operation over the other f.1r' and conductivity a . Our desire is to solve for the capaci­
three sides thus gives
tance, C/, per unit length of the transmission line using the
Finite-Difference Method. For our first example, we let f.1r = 1
S2 "" (1/2)[sr(n,m)+sr(n-l,m)][V(n,m+1)-V(n,m)] ,
and a = 0 , although these parameters will be accounted for
(47) later.
S3 ",,(1/2)[sr(n-l,m)+sr(n-l,m-l)]
[V(n-l,m)-V(n,m)] , (4S) By definition, capacitance per unit length is given as

S4 ",,(1/2)[sr(n-l,m-l)+sr(n,m-l)] C = L' (52)

L'lV
[V(n,m-l)-V(n,m)]. (49)
where q/ is the charge per unit length, and L'lV is the voltage
For notational compactness, we now define the following con­ potential difference between the inner and outer conductors.
stants: The choice of L'lV is entirely arbitrary for our simulation, and
may therefore be specified to any convenient nonzero value.
ao = sr(n,m)+sr(n-l,m)+sr(n,m-1)+sr(n-l,m-1) , This motivated us to set all points within the inner conductor to
Dirichlet boundary conditions at a fixed potential of 1.0 V.
Likewise, we could fix the outer conductor to a potential of
0.0 V. Figure 9a shows the resultant electric-field profile from a
Finite-Difference Method simulation using r, = 0.41 mm,
r2 = 1.47 mm, r3 = 1.75 mm, and sr = 2. 25 (typical values for
RG-5S/U coaxial cable). The grid spacing was set to a value of
h = r1/40 , and the problem required only a few seconds of
simulation time on a modest desktop computer.
The value for q/ was derived from the simulation by
applying Gauss's law in integral form. In two dimensions, this
Adding all sides of the counter integral together thus leads to is given as

S, +S2 +S3 +S4 "" -aoV(n,m)+ a,V(n+l,m) (53)

+ a2V(n,m+1)+ a3V(n-l,m)+ a4V(n,m-1). (50)
where dn is the differential unit normal vector along the con­
Including the charge term from Equation (4 2) , we finally arrive tour of integration. The actual choice of the contour is arbitrary,
at just so long as it encloses the inner conductor of the cable and
does not cross over into the shielding. Figure Sb illustrates a
-aoV(n,m)+ a1V(n+l,m)+ a2V(n,m+1) simple example that fits well in the Cartesian framework of the
2 Finite-Difference Method grid, and encloses the center
h conductor. Evaluation of Equation (53) is then accomplished
+ a3V(n-l,m)+ a4V(n,m-1) --p(n,m). (51)
=

So through numerical integration along the E-field samples

produced by the Finite-Difference Method simulation.
Just like Equation ( 14) , this expression represents a
numerical stencil for the generalized Poisson equation in Figure 9b shows an evaluation of C using the Finite­
accordance with the five-point star. The only difference is a Difference Method over a wide range of values for r2' For
weighting of each term by the average between dielectric con­ comparison, we also included exact calculations from the
stants. As demonstrated in Section 5, it is a straightforward closed-form solution to the same model, given by [15] as
process to generate a system of linear equations of the form
Ax = b that can then be inverted for a numerical solution.
C/ = 2Jisrso . (54)
In(r2/r,)

IEEE Antennas and Propagation Magazine, Vol. 56, No.4, August 2014 217
 Electric Field Intensity (Vim)

-1.5
2000

-1
.--..
E 1500
05
E - .
'-"
Q)
() o
c
ro
-
Cf)
1000
(5 0.5
I
>-

500

1. 5

o
-1.5 -1 -0.5 0 0.5 1 .5
x-Distance (mm)

t:. FDM

6180 -Exact

it 160
'-'
..c
til 140
:::Q)
� 120
.....

� 100
....
Q)
0.. 80
Q)
u
::: 60
«l
.....
'u 40
«l
g.
Figure 8. (a) The coaxial-cable transmission-line model,
U 20
including the physical dimensions and material parameters. 0 0.8 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6
(b) The Gaussian contour of integration within the coaxial­
Inner Shield Radius, r (rnm)
cable model. 2

Figure 9. (a) The Finite-Difference Method simulation of

the coaxial-cable model from Figure 8a. The electric-field
profile was calculated under typical RG-58/U dimensions.
(b) Calculations of capacitance per unit length as a function
of the inner-shield radius, where 2r, ::; r2 ::; 61j .

218 IEEE Antennas and Propagation Magazine, Vol. 56, No.4, August 2014
Clearly, the Finite-Difference Method output was in strong Note how this is identical to the form of Equation (5), but with
agreement with the exact solution, with a mean error of only a complex permittivity instead of a real valued permittivity. The
0.33% along the domain of interest. This shows us how even tilde ( ) over the charge-density term is simply a reminder that
�

a simple numerical method like the Finite-Difference Method p is now a complex phasor quantity rather than a real, static
can still produce very precise results for complex physical value.
geometries.
Because we are now working with a time-varying system,
it is important to realize that the electric field no longer satisfies
9. Quasistatic Conductivity the simple definition E = -v V . To see why, we must examine
Faraday's law, which states
Another useful application for the Finite-Difference
Method is the ability to solve for the quasistatic current density VXE(r) = -j(})B(r), (61)
in conductive materials. We begin with Ampere's law in the
frequency domain, which is written as where B(r) = Jl(r ) H(r) is the magnetic-field intensity. We
now define a vector field, A , called the magnetic vector poten­
tial, which satisfies

where H is the auxiliary magnetic-field intensity, j = � is B(r) = VxA(r). (62)

the imaginary unit, (}) is the angular frequency of excitation,
Je is the induced conduction current, and Ji is an impressed Plugging back into Faraday's law therefore gives
current source. The conduction current, Je , represents the flow
of charges in a conductive material due to the presence of an VX[E(r)+ j(})A(r)] = 0 . (63)
external electric field. These are computed from the point form
of Ohm's law, which is given as The significance of this expression comes from an identity in
vector calculus, which states that if the curl of some vector field
(56) is zero, then that field may be defined as the gradient of some
undetermined scalar field, V. In other words, V satisfies
where (J is the material conductivity. In contrast, the Jj term
is an arbitrary mathematical forcing function that represents the E(r)+ j(})A(r) = -VV(r). (64)
flow of electrical currents being impressed into the system by
external agents. This is the complete definition for the voltage potential, and
includes both the effects of an external electric field as well as a
A common trick in electromagnetics is to merge the dis­ time-varying magnetic field. It also means that E must
placement current and conduction current into a single, com­ satisfy
plex quantity within Ampere's law. This is accomplished by
defining the complex permittivity function, Be' through the E(r) = -VV(r)- j(})A(r). (65)
relation
Although it is possible to independently solve for both A and
V to find E, the process is rather involved and requires a very
complex linear system to couple the two quantities together
[16]. To avoid this complication, a simplified system may be
From here, Ampere's law may now be expressed as reached if we impose the quasi-static approximation, given by

(58) j(})A(r)""0. (66)

If we now take the divergence of Ampere's law, the curl term on In other words, the contribution of A to the total electric field is
the left vanishes, leaving negligible at low frequencies, and E "" -VV . This allows us to
rewrite Equation (60) as
(59)
(67)
Finally, we apply the charge continuity equation by replacing
j(})P = -V.Jj to arrive at Notice how Equation (67) is simply the variable-coefficient
Poisson equation again, but with complex numbers instead of
purely real-valued. The same Finite-Difference Method algo­
(60) rithm we have just derived may therefore be used to find low-

IEEE Antennas and Propagation Magazine, Vol. 56, No.4, August 2014 219
frequency potentials in a time-varying system. The complex 2
Current Density (Nm )
values representing V may further capture the effects of phase
shifts due to material conductivity or any arbitrary offsets
-1.5
impressed within p . We can also apply Equation (56) to solve 2000
for any resulting electric currents that may arise within a
-1
conductive medium. The only restriction is that the frequency
----
must be low enough such that the magnetic vector potential E
-S -0.5 1500
does not significantly contribute to the total electric field.
Q)
u o
c
co
-
(/)
1000
1 0. Example: Conductance Per Unit Length (5 0.5
I
>.

For this next example, we simulated the same coaxial 500

cable as before, but inserted a finite conductivity of a = LOn
m between the inner wire and the outer shield. Figure lOa shows 1.5
the resultant current density inside the cable after simulation
o
using the Finite-Difference Method with h = lj /40 . The inner -1.5 -1 -0.5 0 0.5 1.5
x-Distance (mm)
wire was again excited to a potential of 1.0 V at the quasistatic
frequency f = 1.0 Hz. Likewise, the outer shield was grounded
to 0.0 V.

By definition, conductance per unit length is given as

..c
G'=�' (68) tn 7
L'lV s::
<l)

-:: 6
s::
where ]' is the current per unit length that passes from the � 5
inner conductor to the grounded shield. Again, L'lV is arbitrar­ .....
<l)
� 4
ily defined as an excitation parameter of the simulation. In a <l)
()
fashion similar to Equation (53), ]' is calculated by the integral § 3
.....
()
:l
-0 2
(69) s::
o
U 1

As before, the contour of integration is completely arbitrary, O �--�--��--�----L-�

0.8 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6
just so long as it encloses the inner conductor. The same Inner Shield Radius, r2 (mm)
Gaussian contour from Figure 8(b) is therefore perfectly suit­
able for this calculation. Figure 10(b) shows an evaluation of Figure 1 0. (a) A quasistatic simulation of a coaxial trans­
G' using the Finite-Difference Method over the same range of mission line with finite conductivity. The current density
values for r2 . For validation, the exact values for G' are given profile was calculated under typical RG-58/U parameters.
by [15] The frequency of excitation was f =1.0 Hz, with a conduc­

tivity of a = LOn m. (b) Calculations of conductance per

unit length as a function of the inner shield radius, where
(70)
2lj � r2 � 6lj •

Again, the Finite-Difference Method output strongly agreed

with the exact solution, producing a mean error of only 0.27%.

220 IEEE Antennas and Propagation Magazine, Vol. 56, No.4, August 2014
 1 1 . Static Magnetic Fields From here, it is important to remember that we have not yet
uniquely defined an exact value for A. Instead, all we have done
Let us now turn our attention to the problem of static is indirectly define the magnetic vector potential as some
magnetic-field intensity. Recall from the previous section that unknown vector function the curl of which produces B. This
we indirectly defined the magnetic vector potential using does not actually uniquely define A, unless we also specify the
divergence value, V A , which is a parameter called the gauge.
0

B(r) = VxA(r), (71) Since we are technically free to define the gauge in any way we
choose, it helps to choose a value that is convenient for the

B(r) = JI(r ) H(r) is the constitutive relation for mag­

specific problem at hand. In this specific case, the most
where
convenient value is called the Coulomb gauge, and is given by
netic fields. The magnetic permeability, JI, can then be
expressed using VoA(r) = O. (78)

(72) Plugging this back into Equation (77) then produces

where Jlo = 4;r x10-8 Him is the permeability of free space, (79)
and Jlr is the relative permeability function. For convenience,
we shall also define the inverse relative permeability function which is simply the classical Poisson equation again, but in
using vector form. Breaking this up into vector components quickly
reveals a set of three independent Poisson equations along each
coordinate axis:
(73)

(80)
Now let us take the curl of Equation (71), and substitute the
result into Ampere's law from Equation (58), to arrive at (81)

Vx[Kr (r)VxA(r)] = jOJJloGoGc (r)E(r)+ JlOJi (r). (82)

(74)

The full solution for A in three dimensions can therefore be

For the case of static electric currents, the frequency of excita­
tion is given by OJ = 0 . We may therefore simplify this expres­
found by independently solving each expression along its
respective coordinate axis. Since each expression is perfectly
sion to read
analogous to Equation (2), our original Finite-Difference

Vx[Kr (r)VxA(r)] = JlOJi (r).

Method algorithm can readily solve each coordinate for arbi­
(75) trarily complex arrangements of currents and imposed vector
potentials.
Although the double curl operation may appear unmanageable
at first, comparison with Equation (1) reveals a subtle resem­ Now let us return to Equation (75), and generalize our
blance to the variable-coefficient Poisson equation. This moti­ solution to the case of spatially varying permeability functions.
vates us to view Equation (75) as a sort of variable-coefficient For simplicity, we limit ourselves to the special case of
vector-Poisson equation, where the gradient and divergence z-oriented current sources in a two-dimensional system, and
operators have simply been replaced with curls. Indeed, it is follow the mesh geometry depicted in Figure 11. For the case of
straightforward to show that a very familiar solution will again varying dielectric functions with the scalar Poisson equation,
emerge after applying the Finite-Difference Method to this we found that applying the divergence theorem to a discretely
expression. For example, let us assume that JI = Jlr is a con­ sampled grid quickly led to a simple Finite-Difference Method
stant value across the entire domain of interest. Equation (75) scheme. In a parallel procedure, we therefore begin by applying
then simplifies into Stoke's theorem along the volume element o'nm to find

v x[VxA(r)] = JloJlrJi (r) . (76)

We then recall the vector identity (83)

where dl is the differential unit vector that is tangential to the

(77)
contour Cn m . Explicitly expanding out the curl operation then
leads us to

IEEE Antennas and Propagation Magazine, Vol. 56, No.4, August 2014 221
 � K (r )V x A(r ) • dl
Az(n - 1,m+

Az(n-1 ,

Comparison with Equation (32) shows a nearly identical

expression to the case of varying dielectrics. By following
a parallel procedure, it is thus straightforward to show that
approximation of the derivatives with finite differences quickly
leads us to
Az(n-l,m-
-aoAz (n,m)+ a,Az (n+l,m)+ a2Az (n,m+1)
2
+ a3Az (n-l,m)+ a4Az (n,m-l) = -JlO h Jz (n,m) , Figure 11. The finite-difference mesh for the magnetic vec­
tor potential with a varying permeability function.
(85)

where the coefficients are defined according to

Bx(n-1,

Bx(n-1,m-
Once again, we have the same numerical stencil that solves the
variable-coefficient Poisson equation on a sampled grid. Of
course, the magnetic vector potential, Az ' is not necessarily a
physically useful quantity, so it is desirable to solve for the
magnetic field vector, B. We therefore expand out the curl
operation from Equation (71) in two dimensions to arrive at
Az(n-1,m +

(86)

Note that this expression is analogous to the relationship

between the voltage potential and the electric-field intensity as Az(n-1,
found in Equation (32). We may therefore define a similar grid
stencil between Az and B by using the convention shown in
Figure 12. The resultant field samples in B' are then perfectly
analogous to the staggered E-field samples derived in Section 6.

Az(n-1,m-
1 2. Example: Inductance Per Unit Length

Figure 12. (a) The grid stencil for obtaining magnetic-field

As a final example, we again demonstrate the Finite-Dif­
samples. The circles represent samples in the magnetic
ference Method with our coaxial cable, but now account for
vector potential, while the x represent the magnetic-field
relative penneability Jlr = 2.25 within the insulation layer.
samples obtained by the central-difference approximation.
Figure 13a shows the magnetic-field profile calculated by the
(b) The staggered magnetic-field grid obtained after aver­
Finite-Difference Method simulation with h = r, /40 . The inner
aging the central-difference components together.

222 IEEE Antennas and Propagation Magazine, Vol. 56, No.4, August 2014
 Magnetic Field I ntensity (mT) conductor was driven b y 1.0 A of current distributed uniformly
along the wire. The outer conductor was then driven with an
1 .5
equal and opposite - 1.0 A current, likewise distributed
uniformly along the shield.
0.9

,.-.. 0.8
as
The definition o f inductance per unit length i s now given
E
0.5 0.7
E
'-'

Q) 0.6 <D '

g 0 L' =
' (87)
CO 0.5 I
.....
en
(5 -0.5 0.4
I
where <D' is the magnetic flux per unit length. The total current,
>.
0.3 I, that excites the system is specified by the simulation itself,
-1 and therefore known a priori. The flux must then be calculated
0.2
by the integral
-1 .5 0.1

1.5
<D' = fB( r ) • dn , (88)
-1.5 -1 -0.5 0 0.5
x-Dista n ce (mm)
where the contour of integration is now open rather than closed.
As before, the contour itself is again arbitrary, but must now
begin at the outer boundary of the center conductor, and then
1 000 rrc���-'----'---'---'---"---'---'---,
f),. FDM terminate on the inner boundary of the outer shield.
900 - Exact
E
�
Finite-Difference Method calculations of L' are summa­
'-' rized in Figure 13b for the same range of r2 values as before.
-C
CO For comparison, the exact solution for a coaxial-cable trans­
t:
<I) mission line is given as [ 1 5 ]
...-<
......

·c
::::>
.... (89)
<I)
0-.
<I)
u

Eg 200
As expected, the Finite-Difference Method simulation agreed
-0
well with the exact solution, producing a mean error of only
..s 1 00 0. 1 6%. We also noted that these calculations neglected the
self-inductance within the conducting wires themselves. For­
o �--�--�----�
0.8 1.2 1 .4 1.6 1 . 8 2 2.2 2.4 2.6 tunately, these contributions are generally only significant at dc,
Inner Shield Radius, r (mm)
2 and therefore do not contribute to the net inductance of time­
varying systems.
Figure 13. (a) The Finite-Difference Method simulation of
the magnetic-field profile within a coaxial cable using
f.1r = 2.25 . A total current of 1.0 A was distributed uni­

formly along the inner conductor, with an equal and oppo­ 1 3. Summary and Conclusions
site current placed along the outer shield. (b) Calculations
of inductance per unit length as a function of inner-shield Poisson's equation is a surprisingly prolific expression,
radius, where 21j ::; r2 ::; 6rl • with the power to model a variety of complex electromagnetic
systems. Any numerical tool that can solve Poisson's equation

IEEE Antennas and Propagation Magazine, Vol. 56, No.4, August 2014 223
for just one of these contexts can therefore immediately be 8. N. Ida, Engineering Electromagnetics, Second Edition, New
applied to all others. Three simple examples given in this York, Springer, 2004.
paper were variable dielectrics, variable conductivity, and
variable permeability. However, we need not limit ourselves 9. J.-M. Jin, The Finite Element Method in Electromagnetics,
strictly to the context of Poisson equations in order for the Second Edition, New York, Wiley-IEEE Press, 2002.
Finite-Difference Method to find merit. For example, the
Finite-Difference Method has been used to numerically solve 10. E. J. Lundquist and C. Furse, "Novel Inverse Methods
the Helmholtz equation for microwave circuits [17], as well as for Wire Fault Detection and Diagnosis," IEEE International
the Poisson-Boltzmann equation for ions in solution [18, 19]. Symposium on Antennas and Propagation (AP-S URSI), 2011,
Despite the intense competition from so many other numerical pp. 2573-2576.
solvers, it appears that the Finite-Difference Method will con­
tinue to provide valuable computational insight into modern 1 1. E. J. Lundquist, J. R. Nagel, S. Wu, B. Jones, and C.
scientific applications. Furse, "Advanced Forward Methods for Complex Wire Fault
Modeling," IEEE Sensors Journal, 13, 4, 2013, pp. 1172-1179.

12. C. R. Paul, Analysis ofMulticonductor Transmission Lines,

1 4. Acknowledgments
Second Edition, New York, Wiley-IEEE Press, 2008.

I wish to thank Dr. Cynthia Furse from the University of 13. W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P.
Utah for her guidance and insight during the production of this Flannery, Numerical Recipes: The Art ofScientific Computing,
work. Third Edition, Cambridge, Cambridge University Press, 2007.

14. W. Pinello, M. Gribbons, and A. Cangellaris, "A New

Numerical Grid Truncation Scheme for the Finite Differ­
1 5. References
enceiFinite Element Solution of Laplace's Equation," IEEE
Transactions on Magnetics, 32, 3, 1 996, pp. 1397-1400.
1. G. Liebmann, "Solution of Partial Differential Equations with
a Resistance Network Analogue," British Journal of Applied
15. F. T. Ulaby, E. Michielssen, and U. Ravaioli, Fundamen­
Physics, 1, 4, 1950, pp. 92-103.
tals ofApplied Electromagnetics, Sixth Edition, Upper Saddle
River, NJ, Prentice Hall, 2010.
2. J. W. Duncan, "The Accuracy of Finite-Difference Solu­
tions of Laplace Equations," IEEE Transactions on Microwave
1 6. E. Haber, U. M. Ascher, D. A. Aruliah, and D. W. Oldenburg,
Theory and Techniques, 15, 10, 1 967, pp. 575-582. "Fast Simulation of 3D Electromagnetic Problems Using
Potentials," Journal of Computational Physics, 163, 1, 2000,
3. R. H. Galloway, H. M. Ryan, and M. F. Scott, "Calculation
pp. 150-171.
of Electric Fields by Digital Computer," Proceedings of the
Institution ofElectrical Engineers, 114, 6, 1 967, pp. 824-829. 17. D. H. Sinnott, G. K. Campbell, C. T. Carson, and H. E.
Green, "The Finite Difference Solution of Microwave Circuit
4. R. V. Southwell, Relaxation Methods In Engineering Science
Problems," IEEE Transactions on Microwave Theory and
- A Treatise On Approximate Computation, Oxford, Oxford Techniques, 1 7, 8, 1967, pp. 464-478.
University Press, 1940.
18. A. Nicholls and B. Honig, "A Rapid Finite Difference
5. H. E. Green, "The Numerical Solution of Some Important
Algorithm, Utilizing Successive Over-Relaxation to Solve
Transmission-Line Problems," IEEE Transactions on Micro­
the Poisson-Boltzmann Equation," Journal of Computational
wave Theory and Techniques, 13, 5, 1 965, pp. 676-692. Chemistry, 12, 4, 1 99 1 , pp. 435-445.

6. M. N. O. Sakidu, Numerical Techniques in Electromagnetics,

1 9. Z. Zhou, P. Payne, M. Vasquez, N. Kuhn, and M. Levitt,
Second Edition, Boca Raton, FI, CRC Press, 2001. "Finite-Difference Solution of the Poisson-Boltzmann Equa­
tion: Complete Elimination of Self Energy," Journal of Com­
7. P.-B. Zhou, Wumerical Analysis of Electromagnetic Fields,
putational Chemistry, 11, 1 1 , 1 996, pp. 1344-1351.
New York, Springer Verlag, 1 993.

224 IEEE Antennas and Propagation Magazine, Vol. 56, No.4, August 2014

View publication stats