IEEE TRANSACTIONS

416

ON MICROWAVE

Computation

THEORY AND TECHNIQUES,

of Electromagnetic
WEXLER,

ALVIN

Invited
AbstractThis

paper reviews some of the more useful, current and

newly developing methods for the solution of electromagnetic

begins with an introduction

to numerical

specific references to the mathematical

fields. It

methods in general, including

tools required for field analysis,

e.g., solution of systems of simultaneous linear equations by direct and

iterative

means, the matrix

entiation and integration,

eigenvulue problem, finite

difference

differ-

error estimates, and common types of boundary

conditions. This is followed by a description of finite difference solution

of boundary and initial
matical

principles

value problems. The paper reviews the mathe-

behind variational

methods, from the Hilbert

point of view, for both eigenvalue and deterministic

space

problems. The signill-

cance of natural boundary conditions is pointed out. The Rayleigh-Ritz

approach for determining

the minimizing

sequence is explained, followed

MEMBER,

Fields

IEEE

Paper
A moments
minants

reflection

will show that calculating

by evaluating

all their cofactors

these has cofactors

this way, we find that at least n! multiplications

than the universe is old to expand a determinant

of order 24.
Even then, due to roundoff errors, the reliability
of the result
will be questionable. Cramers rule then is surely not a useful

Wilkinson,
expression

to solve a field quantity,

one must be concerned
with numerical analysis, Other than those engineers
involved exclusively in measurements or in the proof of general existence theorems or in administration,
the remainder
of us are numerical analysts to a degree. Whenever we prescribe a sequence of mathematical
operations, we are designing an algorithm.

A theory is to us a means of extrapolating

our experiences in order to make predictions.

ical theory,

of a physical

good theory produces

problem,

produces

accurate numbers

A mathematnumbers

and a

easily.

As far as engineers
between

are concerned,

analytical

and

there is no principal
numerical

in

approaches.

systems. It is certainly

of

in general proofs and existence theouseless for getting

a statement

numerical

referring

to

the

answers.
algebraic

eigenvalue problem [25, Preface] said.

. . the problem has
a deceptively simple formulation
and the background theory
has been known for many years; yet the determination
of
accurate solutions presents a wide variety of challenging
problems.
He would likely agree that the statement applies
equally

to most of numerical

analysis.

This paper is concerned with the solution

of fields by
efficient, current and newly developing, computer practices.
It is intended

as an introduction

with numerical

analysis,

for those totally

unfamiliar

as well as for those with

perience in the field. For the sake of consistency,

had to be deleted including

To obtain these numbers, we must employ processes that

produce the required accuracy in a finite number of steps
performed in a finite time upon machines having finite word
length and store. Accepting these constraints we ask how
best to perform under them.
difference

except for the smallest

rems but is virtually

ONE devises a mathematical

in

are needed

to expand the determinant.

Neglecting all other operations,
and assuming that the computer does one million multiplications per second, we find that it would take more years

algorithm

1. INTRODUCTION

Each of

of order n 2 and so on. Continuing

the greatest importance

tion.

along each

each of order n 1, are involved.

with an introduction

to the techniques and importance of hybrid computa-

these deter-

is an insurmount-

able task when the order is not small. Expanding

row, n cofactors,

by a brief description of the finite element method. The paper concludes

HENEVER

VOL. MTT-17, NO. 8, AUGUST 1969

numerical

conformal

some ex-

many topics
mapping,

point matching, mode matching and many others. Proponents of these techniques should not feel slighted as the
author has himself been involved in one of them. In addition, it was felt advisable

to concentrate

on general numeri-

cal methods relevant

to review particular

to the microwave engineer, rather than

problems. Companion
papers, in this

special issue, furnish

many specific examples.

Both are concerned with numbers. Oftentimes, the so-called

analytical
approachas though to claim inefficiency as a
virtueis none other than the algorithmically
impotent one.
The contrast between efficient and inefficient computation
is emphasized whenever anything beyond the most trivial
work is performed. As an example, consider the solution of a
set of simultaneous,
linear equations by Cramers rule in
which each unknown is found as a ratio of two determinants.

The principal methods surveyed are those of finite differences, variational,

and hybrid computer techniques. Appropriate mathematical
concepts and notations are introduced
as an aid to further study of the literature. Some of the references, found to be the most useful for the author, are classified and listed in the bibliography.

Manuscript received February 27, 1969; revised May 7, 1969. This

work was carried out with financial assistancefrom Atomic Energy of
Canada Ltd. and the National Research Council of Canada.
The author is with the Numerical Applications Group, Electrical
Engineering Department, University of Manitoba, Winnipeg, Man.,
Canada.

simultaneous

II.

SYSTEMSOF LINEAR EQuArlom

The manipulation
employed

in computer

ear equations

and solution

equations

solution

sets of linear,

of the techniques

of field problems.

are the consequence

tions made to expedite

mation of the operator

of large

is basic to most

of certain

These linapproxima-

the solution. For example, approxi(Vz, often) over a finite set of points

WEXLER

: COMPUTATION

OF ELECTROMAGNETIC

at which the field is to be computed

or approximation

of the field

(as in finite differences),

estimate

(as in variational

methods)

causes the problem

to be modeled

equations.

These equations

are conveniently

matrices,

FIELDS

by simultaneous
represented

approach

The first equation, which was used to clear away the first
column, is known as the pivot equation. The coefficient of the
equation,

by

It happens, and we shall see why in succeeding sec-

tions, that the finite difference

417

creates huge, sparse

which was responsible

for this clearing

operation,

is termed the pivot. Now, the first term of the altered second
equation

is made the pivot.

equation

to the last, we obtain

Adding

4.5 times the second

matrices of order 5000 to 20000. Sparseness means that there

are very few nonzero

elements, perhaps

cent. On the other hand, variational

methods

duce matrices that are small (typically

of order 50), but dense

by comparison.
The size and density of these matrices

is of prime impor-

to about

ern machine.

150 unknowns

in a reasonably

On the other hand, it turns out that iterative

techniques often converge swiftly

addition, if the values and locations
easily computed
point

large mod-

in storing

(as occurs with finite differences)

the large sparse matrix

such cases, only the solution

with five nonzero

for sparse matrices. In

of nonzero elements are

vector

there is no

elements (the number

produced

9X3 = 18.
Clearly, in the computer there is no need to store the equations as in the preceding sets. In practice, only the numerical
values are stored in matrix form with one vector reserved for
the as yet unknown variables. When the above procedure
completed, the matrix is said to have been triangularized.

for X8= 2. Substituting

the numerical

value of X3 into the

In

along

the first equation,

we get xl = 5. And so on back through

by many

This section deals with

examples

for the solution

of direct

and iterative

of systems of simultaneous

linear

Finally,

from
all

the equations in the set, each unknown is computed, one at

a time, whatever the number of equations involved.
It turns out that only n3/3 multiplications
are required to
solve a system of n real, linear equations.

to be solved.

is

The final step, back substitution,

is now initiated. Using
the last equation of the last set, it is a simple matter to solve
X2= 1 is then easily found.

finite difference schemes) at any stage of the iterative process.

This maneuver allows systems having upto20000
unknowns

techniques

second equation,

in its entirety.

need be stored

X3=

2x 2 2X: =

tend to pro-

tance in determining how they should be solved. Direct methods requiring storage of all matrix elements are necessarily
limited

XI +4X2+

a tenth of one per-

figure and should be compared

of work involved

in applying

This is a reasonable

with the phenomenal

Cramers

amount

rule directly.

In addi-

equations. This is followed

by a method for solving the
eigenvalue problem in which the entire matrix must be stored

tion, the method is easy to program.

This simplified
account glosses over certain

but all eigenvectors and eigenvalues are found at once. Solution of the eigenvalue problem, by an iterative technique, is
reserved for Section IV where it is more appropriate.

that may occasionally occur. If one of the pivots is zero, it is

impossible to employ it in clearing out a column. The cure is
to rearrange the sequence of equations such that the pivot is
nonzero. One can appreciate that even if the pivot is nonzero,
but is very small by comparison
with other numbers in its
column, numerical problems will cause error in the solution.

A. Solution

by Direct

Methods

One of the most popular

systems of linear equations
It consists

algorithms

of two partselimination

and back substitution.

for the solution

is known as the method

The procedure

simple and is best illustrated

(or

of

of Gauss.

triangularizution)

is, in principle,

very

We wish to solve the system

where A is an nX n matrix,

(1)

x and b are n-element

vectors

(column matrices). b is known and x is to be computed.

Consider, for convenience, a small set of equations in three
unknowns xl, x2, and x3,
ZI+4XZ+

X3=

ZI+6XZ

x3=13

2XI

~.

X3=

2X2 2X3 =

be held

computers

Typically,

exam-

depending

can hold numbers

with

7 to 11 significant digits in single precisioti.

To illustrate the sort of error that can occur, consider the
example, from [15, p. 34]. Imagine that our computer can
hold only three significant digits in j?oating point form. It
does this by storing the three significant digits along with an
exponent

of the base 10. The equations

1.00

1O-4XI + 1.00Z, = 1.00

1.00X, +

1.0022

are to be solved. Triangularizing,

Xx + 2x3 =

cannot

store. The previous

ple, using integers only, is exceptional.

This example is from [9, pp. 187188]. Subtract the first

equation from the second of the set. Then subtract twice the
first from the last. This results in
XI+4ZZ+

word-length

on the make of machine,

by means of an example.

Ax=b

This is due to the fact that most numbers

exactly in a limited

difficulties

the word-length
1.00 x

limitation,

lo%,

2.00

as before,

while realizing

we obtain

+
1.00 x

1 .Oox, =
10%

1.00

= 1.00 x

Solving

is x.2= 1.00 and xl = 0.00 which is quite incorrect. By the simple

of reversing the order of the equations, we find
. expedient
.

9*z + 0Z3 = 9.

for X2 then back-substituting,

104.

the result

m=

1.00 and xl=

1.00 which

the computed

is correct

result

to three

IEEE

418

significant

TRANSACTIONS

ON MICROWAVE

digits. The rule then is, at any stage of the com-

putation,

to select the largest numerical

value in the relevant

THEORY

AND

designed for real numbers

cedure is undesirable

TECHNIQUES,

AUGUST

1969

can then be used. The latter pro-

as more store and computing

time is

required.

column as the pivot and to interchange the two equations

accordingly.
This procedure is known as a partial pivoting
strategy. Complete pivoting strategy involves the interchange
of columns as well as rows in order to use the numerically
largest number available in the remaining equations as the

upon particular
characteristics
of the matrix involved.
In
particular,
when a matrix
is symmetric,
the symmetric
Cholesky method [3, pp. 7678 and 9597] appears to be

pivot.

the speediest, perhaps twice as fast as Gausss method.

In practice,

the results due to this further

do not appear to warrant

computing

the additional

complications

and

the above condition,

suffered from

except for the eigenvalue

strategy, On the other

problem,

nothing

can be

planes or hyperplanes

system describes hyperplanes

insensitive to roundoff

that intersect

point (or solution)

i.e., nonzero

the diagonal

the simplest methods

error as a consequence. If hyperplanes

is to triangularize

A, as previously

scribed, and then to solve (1) n times using a different

first element with zeros in the remainder.
appear in the second location

111 0

121

122

(3)

132 133

called

which

gives one col-

does this, with

complete pivoting strategy, in an efficient fashion.

Westlake [24, pp. 106-107] reports results of inverting
of order n= 50 on CDC

6600 and CDC

1604A mac-

hines.
Computing
times were approximately
0.3 and 17
seconds, respectively. These figures are about double that
predicted by accounting
only for basic arithmetic
operations. The increased time is due to other computer operations
involved and is some function of how well the program was
on the University

showed that MINV

3.38 seconds. The matrix

the range O to 10.

inverts

of Manitoba

a matrix

IBM

of order 50 in

elements were random

the following

by successively

equations:

= 1112, a12 = 111121, a13 = 111131,a2.2= 1212+ 1222,

aaa = 121131+

122132j a33 =

1312 +

1322 +

(4)

1332.

Thus, the first equation of (4) gives 111explicitly. Using 111,

the second equation supplies 121,and so on. This operation
is known as triangular decomposition.
The inverse of A is
A-l

for the next back-substitution

MINV

The elements of L may be easily determined

right-

umn of the inverted matrix and n back-substitutions

produce
the entire inverted matrix.
See [12, pp. 3233 ]. The procedure is not quite so simple when, as is usually advisable, a
pivoting strategy is employed and when one wishes to triangularize A only once for all n back-substitutions.
IBM

run

and below

as in

de-

The 1 should then

sequence, and so on. Each time, the solution

Tests

L is a lower triangular

occur only along

[1 1,1

all

hand side b each time. In the first case b should have 1 in its

a program

(2)

elements

L=

employing

is relatively

intersect at small angles, roundoff error causes appreciable

motion of the intersection point with a low degree of trust
in the solution. Such systems are termed ill-conditioned.
It is unnecessary, and wasteful, to compute the inverse AI
in order to solve (l). Occasionally, however, a matrix must
be inverted. This is reasonably easy to accomplish. One of

360/65

Under

when the order is greater than two. A

at nearly 90. The intersection

written.

depend

we can write

where the tilde denotes transposition.

matrix,

singularity
(i.e. the vanishing of the coefficient matrix determinant)
means that the two lines are parallel. Hence, no
solution
exists. A similar situation
occurs with parallel

matrix

of which

a pure scaling prob-

done about solving a system of linear equations whose coefficient matrix is singular. In the two dimensional
case,

well-conditioned

exist, many

A=LE

example

lem and is easily handled by a pivoting

supplies

algorithms

time.

The previous
hand,

refinement

Other inversion

numbers

in

Note that the determinant

of a triangular matrix is simply
the product of the diagonal terms.
The elimination
and back-substitution
method is easily
adapted to the solution
of simultaneous,
linear, complex
equations if the computer has a complex arithmetic facility,
as most modern machines do. Failing this, the equations can
be separated into real and imaginary
parts and a program

which
tion.

requires

~-1~-l

(5)

the inverse of L and one matrix

The inverse of a triangular

to obtain

(L:l)L-I

matrix

= 1,

231Z11

l,,x,,

+ 1,3X32= o,

133X33

a symmetric

matrix

is most

inverted.
From (4), it is clear that complex

arithmetic

this

easy

by the sequences
lIIZII

In

multiplica-

is particularly

way,

be performed.

121X11+ 122221= o,

Z32$Z1

Z33X21= o,

However,

122X22= 1,

(6)

1,

if in addition

economically
may have to

to being symmetric

is positive definite as well, we are assured [3, p. 78] that only

real arithmetic is required. In addition, the algorithm
is extremely stable.
B. Solution
Iterative

by Iterative
methods

ods of solution

Methods

offer an alternative

previously

described.

to the direct meth-

As a typical

ith equa-

tion of the system (l), we have

&

a,,xj = b<.

(7)

j=l

Rearranging

for the ith unknown

.i=:_~5.j.

(8)
j=l ait
j#i

WEXLER

: COMPUTATION

OF ELECTROMAGNETIC

This gives one unknown

in terms of the others. The intention

is to be able to make a guess at all variables

cessively correct

them,

Different

strategies

compute

revised

one at a time,

are available.

estimates

assumed values. Upon

FIELDS

M and then suc-

as indicated

One can, for

of all xi using

completion

above.

419

&is

known

as the SOR iteration

tant role in convergence

The iterative

procedure

mate, just as soon as required, rather than to store it until

the equation scan is completed, With the understanding
that
variables are immediately
overwritten
in computation,
we
have

where c is the last term

literally

dated. With

The process must be stationary when the solution is attained, i.e., x(~+l)
= x(~)
=x.
Therefore, we must have that

are newly up-

only n storage locations

need be

available

for the xl (in comparison

with 2n by the previous

system),

and

solution

convergence

to the

$.)x.
x==

is more

(15) into (14) and rearrange

to obtain

J30(xbfd x).

ues. At the mth iteration,

~(m)

Therefore

between exact and computed

~(m)

~.

(17)

(16) becomes
~(m+l)

(m)

&@8

(18)

. ..=

&m+18(o)

due to the recursive

of the equation

that the error tends to vanish when matrix&

to maximize

the conver-

with exponentiation.
any
only
than
of a
(lo)

cannot

out that convergence

A is symmetric

to the solution

and positive

is guaranteed

definite

[34, pp. 237

238] and if O< w <2. If m<1 we have underrelaxation,

and
ovemelaxation if m>1. This procedure is known as successive
overrelaxation
(SOR), there being no point in underrelaxing.
Surprisingly,
(10) can be described as a simple matrix
iterative procedure [22, pp. 5859 ]. By expansion, it is easy

clear

tends to vanish

[22, pp. 1315 ] that

(which

eigenvalues,

is not symmetric

and
less
case
We

in general)

but we will assume this for pur-

In this case all eigenvectors

independent.
Therefore, any arbitrary
be expressed as a linear combination

are linearly

error vector & may

of eigenvectors 1; of

&@,i.e.,

~ = alll

+ a212 +

. .

d..

(19)

If the eigenvectors are known, the unknown ai may be found

by solving a system of simultaneous linear equations. As the
Ii are linearly independent,
the square matrix consisting of
elements of all the Ii has an inverse, and so a solution

to prove that
(D -

It can be proved

be sure that &

poses of illustration.

It turns

of (16). It is therefore

matrix A raised to a power r vanishes as r-+ co if

if each eigenvalue of the matrix is of absolute value
one. It is easy to demonstrate this for the special
real, nonsymmetric
matrix with distinct eigenvalues.

has only distinct

if matrix

definition

&@28(m1)

tween 1 and 2 and is often altered between successive scans

form is

val-

the form is

current estimate. In order to speed convergence, it is possible

to overcorrect at each stage by a factor w Q usually lies beset in an attempt

(16)

These terms are clearly error vectors as they consist of elements giving the difference

rapid

[22, p. 71].
The difference between any two successive xi values corresponds to a correction term to be applied in updating the

gence rate. The iteration

(15)

c = (1

count. The two-

that some variables

this method,

one does not

the SOR. process, but

& and (14) express mathematically

what happens when the
algorithm
described by (10) is implemented.

X(m+l)
indicates

of (12). Of course,

set up &, in order to perform

Substitute

part summation

can be rewritten

the previously

of the scan of all equa-

1S i< n, m> O. m denotes the iteration

and plays an impor-

of the method.

example,

tions, the old values are then overwritten

by the new ones.
Intuitively,
it seems reasonable to use an updated esti-

with

matrix

properties

must

exist.
= {(1 -

@)x@+l)

u)D

+ CJ}x(n)

+ cob

(11)

Performing

the recursive

operations

defined

by (18), we

obtain
where D is a diagonal matrix consisting of the diagonal elements of A, E consists of the negative of all its elements beneath the diagonal with zeros elsewhere, and F is a matrix
having the negative of those elements above the diagonal of
,4. Rearranging (11),
X(.+l)

(D

~~)1{(

~)~

~~]x(m)

(12)
+ (.o(D O.E)-%.
Define the matrix
&.

accompanying

= @

which states the difference between

mates, can replace s in (18>(20).

x @Jas

6JE)-{(1

CO)D+

where pi is the eigenvalue of.& corresponding

to the eigenvector Ii. It is obvious, from (20), that if all eigenvalues of
& are numerically
less than unity, SOR will converge to the
solution of(l).
In the literature, the magnitude of the largest
eigenvalue is known as the spectral radius p(&J.
It is also easily shown that the displacement
vector 5,

d].

(13)

A sufficient,

although

two

successive x esti-

often overly stringent

condition,

for

IEEE TRANSACTIONS

420
the convergence

of SOR is that the coefficient

matrix

A dis-

ON MICROWAVE
ployed

THEORY AND TECHNIQUES,

using only real arithmetic.

AUGUST 1969

Therefore

play diagonal dominance. This occurs if, in each row of A,

the sum of the absolute values of all off-diagonal
terms is
greater than the absolute value of the diagonal term itself.
In practice, we are concerned not only with whether or not
convergence will occur, but also with its speed. Equation (20)
indicates that the smaller the spectral radius of&,
the more
rapidly convergence occurs. As indicated by the subscript,
.& and its eigenvalues are some function of the acceleration
factor u. The relationship
between coand the spectral radius
P(SJ is a very complicated
one and so only very rarely can
the optimum
acceleration
factor uoPt be predicted
in advance. However, some useful schemes exist for successively
approximating

uoP~ as computation

proceeds

(see

[14],

[23)
where

Note that ~ equals C and so it is symmetric.

Taking

the determinant
det (A hB)

Since det (L) is nonzero,

of both sides of (23),

= det (L)

det (C Al).

we can see that eigenvalues

are those of C. Therefore,

eigenvalue problem
(c

M)y
=
o

(27)

We therefore

obtain the required

Eigenvectors

y are easily transformed

what restricted. There are direct methods that economize on

empty matrix elements, but SOR is perhaps the easiest way
to accomplish this end.
C. The Matrix

eigenvalue

problem

is of the form

where A and B are square matrices.

a variational

solution

This is the form that re-

(see Section V). The prob-

lem is to find eigenvalues x and associated eigenvectors x

such that (21) holds. This represents a system of linear,
homogeneous equations and so a solution can exist only if
determinant
of (A MI) vanishes. If A and B are known
matrices, then a solution can exist only for values of A which
make the determinant vanish. Eigenvectors can be found that
correspond to this set of eigenvalues. It is not a practical
to find the eigenvalues

A values and evaluating

found

to vanish.

to the required

x by

~ in (27).

Orthogonal
matrices are basic
matrix T is orthogonal
if

to

Jacobis

If T consists of real elements, it is orthogonal

(21)

(A AB)X = o

proposition

inverting

eigenvalues by solving (26).

method.

Eigenvalue Problem

The general matrix

sults from

(26)

y = Zx.

tends to be self-correcting

errors are some-

of A

where

tance for the solution of difference equations which require

only the vector to be stored. Also, the iteration procedure
and so roundoff

(25)

instead of (21), we can solve the

[15], [17]-[20]).
To conclude, the main advantage of SOR is that it is unnecessary to store zero elements of the square matrix as is
required by many direct methods. This is of prime impor-

(24)

C = L-lAZ-l.

simply

the determinant

Such a procedure

by assuming trial

each time until

is hopelessly

it is

inefficient.

An algorithm
for matrices that are small enough to be
held entirely in the fast store (perhaps n = 100) is the Jacobi
method for real, symmetric matrices. Although
it is not a
very efficient method, it is fairly easy to program and serves
as an example of a class of methods relying upon symmetry
and rotations to secure the solution of the eigensystem.
If B= 1, the unit matrix, (21) becomes
(A AI)X = o

(22)

we know that B is symmetric and positive

decomposition
(as described in Section

equals one and

if the sum of the products

elements in two

different

columns

and another

definite, triangular
II-B) may be em-

of corresponding

vanishes. One example

is the unit matrix

is
Cos +
T=

sin

Cos
+1
sin

(29)

Matrix (29) can be inserted in an otherwise unit matrix such

that the cos @terms occur along the diagonal. This is also an
orthogonal
matrix and we shall denote it T as well.
Since the determinant
of a product of several matrices
equals the product
that det (T)=

of the determinants,

we can see from (28)

1 or 1. Therefore

det (A M)

= det (T(A

M) ~)
(30)

= det (T.4 ~ U)
and so the eigenvalues of A and TAT are the same.
It is possible to so position (29) in any larger unit matrix,
and to fix a value of @, such that the transformed
matrix
TAT has zeros in any chosen pair of symmetrically
placed
elements.

Usually

ments having

which is the form obtained by finite difference methods. If

A can be stored in the computer, and if A is symmetric as
well, a method using matrix rotations would be used. Equation (21) may not be put into this form, with symmetry preserved, simply by premultiplying
by Bl. The product of two
symmetric matrices is not in general symmetric. However, if

if the sum of

the squares of the elements of each column

one chooses to do this to the pair of ele-

the largest absolute

values. In doing this, cer-

tain other elements are altered as well. The procedure is repeated successively with the effect that all off-diagonal
terms
gradually vanish leaving a diagonal matrix. The elements of
a diagonal matrix are the eigenvalues and so these are all
given simultaneously.
It is possible to calculate the maximum
error of the eigenvalues at any stage of the process and so to
terminate
the operation
when sufficient accuracy is guaranteed. The product of all the transformation
matrices is

WEXLER

: COMPUTATION

continuously

computed.

eigenvectors

OF ELECTROMAGNETIC

The columns

of this matrix

421

are the

f(x)

of A. See [3, pp. 1091 12] and [7, pp. 129-131].

Probably

the most efficient procedure,

time, is that due to Householder.

A method,

employing

in both storage and

It is described

son [26] and, in a very readable

described

FIELDS

form,

by Wilkin-

by Walden

[23].

SOR for large sparse matrices,

is

in Section IV,
III.

FINITE DIFFERENCES

The importance
of finite differences lies in the ease with
which many logically complicated
operations and functions
may be discretized. Operations are then performed not upon

fx.h

continuous

fi.1

functions

but rather,

values over a discrete point

approximately,

in terms of

fx

set. It is hoped that as the dis-

tance between points is made sufficiently

mation

becomes increasingly

accurate.

small, the approxix-h

l-l

and can then be conveniently

digital computation.
In short,

difference,

much logical

tion. The numerical

method,

and simple in principle.

The analytic

subtlety

evaluated

distance

approach

and algebraic

However,

implementation

in many
notation,

at any x is often written~(x)

h to the right f(x+k)

=f,+~.

=f,.

at any specified x, may be approximated

dl~erence
formula

where the function

is evaluated

explicitly

Equally,

the derivative

(31)
at these two points.

to our intuitive
maybe

are required

to compute

ference formula
available

is preferred

Taylors

at x+h.

dif-

and should be used if data are

on both sides of the point

expansion

derivatives

data. The central

being considered.

of (31}(33)

is obtained

from

This gives

Similarly,

at x h

Subtractingfz

from (34), then rearranging,

gives

by the

This, for very small h, corresponds

backward

At a

~, = .fz+h f.

a derivative.

i+l

To the left we

g/dx,

formulas

An idea of the accuracy

understood that the distance between nodes is h and node i

corresponds to a particular x. Refer to Fig. 1. The derivative
forward

i +l/2

at extreme ends of series of tabulated

have f~_~, f~_,~, etc. Alternatively,

nodes (or pivotal points)
are numbered i yielding function values f ~,f i+l$ f il, etc. It is

f:=

given by the slope of the chord passing through

ward difference

innova-

on the other hand, is very direct

instances presents problems of specific kinds.

Consistent with a frequent finite difference
function~

i-1/2

i+l
x+h

certainly appears to approximate

the true demost closely. Usually, this is so. Forward and back-

rivative

First of all, consider differentiation.

point,

fzfh and f.-,,

A. Di&erentiation

often requires

x-h/2

Fig. 1. Forward, backward, and central differences.

programmed
for automatic
complexity
is exchanged for

labor.

fx*
I
1
X+h/Z

The great advantage

of this approach is that operations,

such as differentiation
and integration,
may be reduced to simple arithmetic forms

$1

notion

of

expressed by the

O(h) means that the leading correction

term deleted from this

forward difference derivative approximation

is of order h.
Similarly,
from (35), the backward difference formula can
be seen to be of order h as well.
Subtracting

(35) from (34), and rearranging,

we obtain

dl~erence formula

f:
(32)

(33)
to give a closer estimate. This expression is the central dzjference formula for the first derivative. In the figure, we see that
the forward

and backward

on alternate

sides of the point

differences

give slopes of chords

being considered.

The central

(f.+}

fz-,)/2h

~h2fJ + .-.

= (fm+h f.-h)/2h
which

These are, in general, not equal. The first, in our example,

gives a low value and the second a high value. We can therefore expec+ the average value

is the central

difference

(37)

+ O(h)
formula

(33) with

a leading

error term or order hz. Assuming that derivatives are well

behaved, we can consider that the h term is almost the sole
source of error. We see that decreasing the interval results
in higher accuracy, as would be expected. In addition, we see
that the error decreases quadratically
for the central difference derivative
ward difference

and only linearly

ones.

for the forward

or back-

Only two of the three pivotal points shown in Fig. 1 were

needed for evaluating the first derivative.
Using the three
available pivots, a second derivative maybe computed. Con-

422

IEEE

sider the derivative at x+h/2.

half the previous interval,

TRANSACTIONS

Using central differences,

ON MICROWAVE

THEORY

with

AND

TECHNIQUES,

AUGUST

1969

i,j+l

II
---f--i
h

.f~+hiz = (f~+~ ~S)/h.

(38)

f;-h/z

(39)

Similarly,

= (.fz j#t)/ii

giving the values of the first derivative

h apart. The second derivative

at two points distance

Y

at x is then

1 i,j-1

I
(40)
= (fcc+h 3.+

f.-it)/h2
Fig. 2.

in which (33) has been applied

Five point, finite difference operator.

with (38) and (39) supplying

two first derivative values. By adding (34) and (35), (40) may
be derived with the additional
information
that the error is

can be fitted to data not equispaced.

of order hz.

exist at x h, x, and x+ ah (where a is a real number),

Another

point

differentiation

of view for understanding

is the following.
f(z)

is passed through
For simplicity

finite

difference

= az + bx + c

the three points f.h, f.,

let x= O. Evaluating

and fz+h of Fig. 1.

(41) at x= h, O, and h,

(42)

b = (fh f-h)/2k

(43)

c = fo.

(44)

Differentiating
(41), then setting x= O, we find we get the
same forms as (33) and (40) for the first and second derivatives. Thus, differentiation
of functions specified by discrete
data is performed
by fitting a polynomial
(either explicitly
or implicitly)
to the data and then differentiating
the polyIt is clear then that an nth derivative

accurate.

A high-order

polynomial,

made to fit a large number

points, may
cumstances,
order terms
a numerical

experience severe undulations.

Under such cirthe derivative will be unreliable
due to higher
of the Taylor series. Also note that accuracy of
differentiation
cannot be indefinitely
increased

of approximate

data

by decreasing h, even if the function can be evaluated at any

required point. Differentiation
involves differences of numbers that are ahnost equal over small intervals. Due to this
error

may become

to h is set largely

significant

by the computer

and so the
word length.

Bearing in mind that high-order

polynomials
can cause
trouble, some increase in accuracy is possible by using more
than the minimum
number of pivots required for a given
derivative. For example, as an alternative to (40), the second
derivative can be obtained from
f~ = (f.-z~+

16f.-~

30f.+

(af~-~

if pivots

16f.+hf.r+m)/12h

(45)

with an error of 0(h4).

Not always are evenly spaced pivotal points available.
Finite difference derivative operators are available for such

an

is

equations

(46)

(1 + a)f~ + f~+~)

to (40) when a= 1. However,

and other differentiation

if a# 1, (46)

for unsymmetric

pivots

accuracy.

The preceding,

and many other differentiation

are given in [12, pp. 64-87]. For instance,

differentiation
involves the determination
differentiation
reason that

polynomials

For example,

expressions,

since numerical
and subsequent

of an interpolating
polynomial,
the derivative
need be restricted

there is no
to pivotal

points.
Finally,

the most important

derivative

operator

purposes is the Laplacian V2. In two dimensions

V?. Acting upon a potential 4(x, y) we have

for our

it becomes

(47)

of a function

can be obtained only if at least n+ 1 data points are available. A word of warningthis
cannot be pursued to very
high orders if the data is not overly

which is identical

as approximating

expression

C& + l)h2

have reduced

and

cause, roundoff

derivative

f:! =

(41)

a = (f~ 2~0 + fh) /2h2,

lower limit

appropriate

Assume that a parabola

we easily find that

nomial.

cases as well. This is obvious

where x and y are Cartesian

subscript
index convention,
coordinate,

the finite

difference

coordinates.
Using
and applying
(40)
representation

a double
for each

of (47) is

this five point, finite difference operator

Fig. 2 illustrates
which is appropriate
for equispaced data. Often it is necessary to space one or more of the nodes irregularly.
This is
done in the same fashion employed in the derivation
of (46)
[12, pp. 231-234].

Finite

difference

Laplacian

operators

also available in many coordinate

systems other
Cartesian. For example, see [12, pp. 237-252].

than

are
the

B. Integration
Numerical integration
is used whenever a function cannot
easily be integrated in closed form or when the function is
described by discrete data. The principle
behind the usual
method is to fit a polynomial
to several adjacent points and
integrate the polynomial
analytically.
Refer to Fig. 3 in which the function f(x) is expressed by
data at n equispaced

nodes. The most obvious

integration

WEXLER

: COMPUTATION

OF ELECTROMAGNETIC

FIELDS

f(x

423

error

i
\
i
\

\
II \!

i
,

/
/
/

\
\
\

1
$
!

t
1

\
\

/)
\

/
\

,!

\\
\

11
\

,r/
\\

,/
\

//

i-1

i-1

i+l

Integration

mesh intervol

under the curvej(x). Trapezoidal rule for a single

Fig. 4.

strip and Simpsons 1/3 rule for pairs of strips.

formula results by assuming that the function~(x)
consists of
piecewise-linear
sections. Each of these subareas is easily
computed

.;-

-----

-------

Fig. 3,

total

,Zdkcretizotion
round-off.
error
.
d
error
error
/
.
,/
==

through

-.

-----

---__...---

Error as a function

of mesh interval.

It is not possible to indefinitely

reduce h with the expectation of increased accuracy. Although smaller intervals reduce
the discretization
error, the increased arithmetic
causes
larger
total

roundoff
error

error.

occurs

A point

for

is reached where minimum

any particular

algorithm

given word length. This is indicated in Fig. 4.

Highly accurate numerical integration procedures
where x; is the value of x at the ith pivot.
found

by summing

all A;. Equation

rule for approximating

order hs.
The trapezoidal
to a large number

The total

area is

(49) is the trapezoidal

the area under one strip with error of

formula

(49) may be applied,

of strips.

sequentially,

In this way we obtain

vided
Rather

by

than

double

very time
which has a remainder
mulated

of order h2. The decrease in accuracy,

with the single-strip

by adding

all the constituent

A more accurate formula

one) is Simpsons

case, is due to error

using

(and perhaps the most popular

approximates

two strips to 0(h5). If the interval

even number
turn.

of strips,

(a, b)

the integral
is

(51) can be applied

divided

into

over
an

bf(z)dz

e :

+ 2fn-,

(52)

with an error of 0(h4). Another Simpsons formula, known

as the ~ rule (because the factor % appears in it), integrates
groups of three strips with the same accuracy as the $ rule.
Thus, the two Simpsons rules may be used together
for an odd number of strips.

the procedure

becomes

(53)

f.,uclxdy

limits,

the region

is subdivided

(see Fig.

5(a)). As one would expect, to perform a double integration

by the trapezoidal rule, two applications
of (49) are required
for each elemental

region.

Integrating

along x, over the ele-

ment shown, we obtain

h
=

(fi,j

.fi+l!J

(54)

+ 4fn_, + f.)

this

ac

gj

(f, + 4fl + 2f2 + 4f3 +

integration

at each pair in

Therefore

positions,

bd

v=

* rule

by using a parabola,

pivot

are pro3 12367].

To integrate

subareas.

(51)

[7, pp.

and cumbersome.

accu-

.5

formulas

predetermined

or triple

consuming

over the specified

which,

quadrature

any

method chooses them in order to minimize the error. As a

result, it can only be used when the integrand is an explicit
function.
Multiple
integration
[12, pp. 198206] is, in theory, a
simple extension of one-dimensional
integration.
In practice,
beyond

compared

Gauss

using

9j+l

It now remains to perform

Ui!,+l

fi+l!j+l).

(55)

the integration

to cater

By and large, integration

is a more reliable process than
differentiation,
as the error in integrating
an approximating
polynomial

tends to average out over the interval.

which results from two applications

of the trapezoidal

rule,

424

IEEE

ON MICROWAVE

THEORY

AND

TECHNIQUES,

AUGUST

~hff(x,y)dxdy
n
i ,j

&%:&
i+l,

: P

I mn

(a)

w+

rs
t

....

....

+-t
.,

---.--%----~

W;f[f(xy)dxdy
Neunmnn
boundary

Fig. 6.

--

1
I

kouchy
boundary

-----

The Neumann

3$

l:~----------q(s)(d+(dq(s)

A mixed boundary

boundary

value problem.

condition

(b)

(59)

Fig. 5. Double integration molecules. (a) Trapezoidal

(b) Simpsons

rule.

1/3 rule.

is also easy to interpret

Similarly,
V = ~

three applications

{ 16j;,j -1- 4(fi,j-1

of Simpsons

.ft+I,~

.fi,j+I

~ rule produce

fi-1,~)

(57)
+

.fi+l,i-l

.ft+l,f+l

,fi-l,i+l

J.t-1,+1

which gives the double integral off. ,Vover the elemental two
dimensional
region of Fig. 5(b). Molecules in the figure
illustrate

the algorithm,
IV.

forced into the region,

independent

of the potential.

condition

differential

equations.

In the first place, the differen-

tial equation is transformed

into a difference equation by
methods described in Section III. The approximate
solution
to the continuous
problem is then found either by solving
large

systems

of

simultaneous

linear

equations

for

the

deterministic problem or by solving the algebraic eigenvalue

problem as outlined in Section II. From the point of view
of fields, the resulting solution is then usually a potential
function @ defined at a finite number of points rather than
continuously
over the region.
A. Boundary

to be

at a rate
of constant

sheet of charge backed

the Neumann

boundary

an impermeable

border,

in Fig. 6, p(s)= O.

The remaining

boundary

condition,

of concern

to us, is

the Cauchy (or third) condition.

Imagine that the border is
a film offering resistance R to current flowing across it. Such
a film occurs in heat transfer between, say, a metal and a
fluid. A series of resistors strung out across the boundary
would simulate such an effect. Let the potential just outside
the conducting region be ~O(s), a function of position along
the curve. The potential just inside is $(s) and so the linear
current density transferred
across is iJs) = (+(s) dO(s))/R
where R is the film resistance. Since (l/r)/(d@/iIn)].
= i.(s),
we have by eliminating
i.(s), (R/r)/(E@/&z) 1,+d(s) = +.(s).
In general, this is written

Conditions

condition

then,

as in (59). Along

C3(j

-&

The most frequently occurring boundary conditions are of

several, very general forms. Consider, first of all, the DirichIet boundary

current

A large number

from a distributed
In general

is written

an open circuit
The finite difference technique is perhaps the most popular numerical
method for the solution
of ordinary
and

Imagine

across the boundary,

the normal electric field strength at the boundary

t@/&z \,

in(s), i.e., (1/r)/(&j/&z)

1, = in(s). Another
analogy is
the flux emanating

FINITE DIFFERENCE SOLUTION OF

physically.
from

current sources, strung along the boundary, would simulate

this effect. The normal, linear current flow density is in(s),
the negative sign signifying a flow direction in a sense opposite to the unit normal. The surface resistivity r divided into

by a conductor.

PARTIAL DIFFERENTIAL EQUATIONS

partial

1969

D irichlet boundary

i+l, j+l

i,j+l

TRANSACTIONS

+ a(s)cj(s)

= q(s).

(60)

s
A harmonic
wave function 4, propagating
into an infinite
region in the z direction, obeys &p/8z = jp~ or &$/dz+j@$

defined by
0(s) = g(s)

(58)

= O at any plane perpendicular

to z. This is a homogeneous

which states the values of potential at all points or along any

number of segments of the boundary.
See Fig. 6 which
represents a general two-dimensional
region, part of it obeying (58). If we visualize this region as a sheet of resistive
material,
with surface resistivity r, the Dirichlet
border is
simply one maintained
at a potential g(s). At ground poten-

case of (60).
A region having two or more types of boundary
is considered to constitute a mixed problem.

tial, g(s) = O which

condition.

niently classified as elliptic, parabolic,

and hyperbolic
[12,
pp. 190-191 ]. Under the elliptic class fall Laplaces, Poissons,

is the homogeneous

Dirichlet

boundary

B. D@erence
Second-order

Equations
partial

of the Elliptic
differential

conditions

Problem

equations

are conve-

WEXLER

: COMPUTATION

partial

and the Helmholtz

illustration,
Fig.
typical
tion

OF ELECTROMAGNT3TIC

differential

we will now consider

equations.

of Laplaces

By way of

.-, =V

4(s)

operators

in somewhat

To be specific, consider the discretiza-

equation

V%j = o
consistent

425

the first and last ones.

6 shows several five-point

environments.

FIELDS

with the applied

(61)

boundary

conditions.

From

(48),

(61) becomes
(62)

+a+@%@c+$d+&=O.

Fig. 7.

A reasonably fine mesh results in a majority

of equations
having the form of (62). The trouble occurs in writing finite
difference

expressions near the boundaries.

Near the Dirich-

Iet wall we have Of= g(sl) where SI denotes a particular

on s. Making

this substitution,

the finite

difference

Finite difference mesh for solution of Laplaces equation under

Dirichlet boundary conditions.

appropriate

point

204] , [35, pp. 262-266],

In reference

expres-

sion about node h is

method

+, 4r#J/, +@i+@lj=-gf
where S1and nodef coincide. Along
(59) must be satisfied. To simplify

schemes. See for example

(34, pp.

198

[37, pp. 36-42].

[27] a simple, although

for fitting

[28],

an arbitrary

somewhat

inaccurate,

shape is given. For simplicity,

the boundary was permitted to deform, at each horizontal

mesh level, to the nearest node point.
A crude-mesh
difference
scheme, for the solution
of

(63)

the Neumann boundary,

matters for illustration,

Laplaces equation under Dirichlet

shown in Fig. 7. The appropriate

the five-point operator is located along a flat portion of the

wall. Using the central difference formula (33), (59) becomes

be written

in matrix

form

boundary
difference

in the following

conditions,
is
equations can
fashion:

410100
14

14

00

00

2V

10

0010
100

This equation can then be used to eliminate

the difference equation written about node m.

Therefore
~~ @., + 26. + @o = 2hpm.

(64)

Finally,
at the Cauchy boundary,
(4. 4J/2h+
which makes the node r difference equation

u,& = q,

SOR is employed

+ @ = 2hg,.

to solve for each node potential

(65)
in terms

(66)

v I

01

1 4.

=p~.

& + 24, 4(1 + urhi~)or

00

1 4

000010

(4~&)/2h
node 1 from

01

0141

000

2V

o4

01

0-

or

A+
There are several significant

(67)

= b.

features about (66). In the first

place, there are a fairly large number of zeros. In a practical

case, the mesh interval in Fig. 7 would be much smaller and
the square matrix would then become very sparse indeed. In
addition,
the nonzero elements of any row of the square
matrix and any element of the right-hand side vector maybe
easily generated at will. It is clear therefore that SOR (or

of all other potentials in any given equation. As there are at

most five potentials
per equation,
each node is therefore

any similar

written

then be reserved for the ~ vector and the program,

thus
allowing for a very fine mesh with consequent high accuracy.

in terms of its immediately

adjacent

other words, it is the central node potential

potentials.

In

of each operator

iterative

scheme) is the obvious

ing the system of linear

equations.

choice for solv-

All computer

store may

that is altered in the iterative process. As all points on a

Dirichlet
boundary
must maintain
fixed potentials,
no
operator is centered there. This does not apply to Neumann
and Cauchy boundaries, however, and so potentials on these
walls must be computed.
It is no mean feat to program the logic for boundaries

Certainly, a direct solution scheme, such as triangularization

and back substitution,
could not be considered feasible.
Green [36] discusses many practical aspects and problems
associated with the solution of Laplaces equation in TEM

(particularly

Now, let Fig. 7 represent an arbitrarily

shaped waveguide.
For this purpose, the boundaries must be closed. The tech-

non-Dirichlet

ones) which

to mesh lines, and this is one of the major

finite

difference

approach.

The literature

do not correspond
drawbacks

of the

gives a number

of

transmission lines. Seeger [42] and Green describe the form

of the difference equations in multidielectric
regions.

nique is to solve for a potential

@ in finite

difference

form.

IEEE TRANSACTIONS

426
If z is along the axial direction,
H. depending

on whether

z&+ is proportional

TM

to E, or

or TE modes are being con-

sidered. For TM modes the boundary

condition
is the
homogeneous
Dirichlet
one (58), i.e., with g(s)= O. Fields
are then derived from

Z=

ON MICROWAVE

THEORY AND TECHNIQUES,

a true eigenvalue.

will be in error

and so ~ cannot be found by SOR alone and an outer

iteration
employing
the Rayleigh quotient
(defined later)
must be employed.
Application
of SOR to a homogeneous
causes (14) to assume the form

+,4

set of equations

+(m+l) = &@,A+(m).

c
The subscript

(68)

indicate

symmetric)

(74)

A has been added to the iteration

a further

for illustration,

functional

dependence.

assume that

matrix

with

J&,l

distinct

if E, is made eaual
to d. Likewise, the homogeneous Neu.
mann condition
(59) (with p(s)= O) applies to TE modes,
from

~(m) = alll
Iterating

through

+ u,l, +

is a real (generally

non@(~)

(69)

Then,

of eigenvectors

. . . + al..

(75)

having

-1- . -t- anp~1~.

the greatest

is large, we have substantially

absolute

(76)
value,

that

+(~+s) = alpls~l.

Hz = 4.
k, is the cutoff wavenumber

and Y the propagation

constant.

Of course, one does not solve for@ but rather for +, a vector
of discrete potentials,

and so the differentiations

by (68) and (69) must be performed

in Section III-A.
Discretization

by techniques

indicated
outlined

(77)

Equation (74) must represent a stationary process when SOR

has converged and so ~, = 1 at the solution point. It is interesting to note that the eigenvector 21 of $ti,x is, or is proportional to, the required eigenvector
is substituted into (71).
Rewrite

of A when the correct

(71) as
B$=O

of the Hehnholtz

of

(74)s times, we find that

If KI is the eigenvalue
then ifs

to

eigenvalues.

+(m+.) = a w111 + awflz

Ez=O

matrix

As in Section II-B,

may be expressed as a linear combination

&,i,
i.e.,

H,=O

with fields obtainable

In general, the 1 estimate

AUGUST 1969

(78)

equation
where

(v, + k.)d = o

(70)
B=.4

produces

a matrix

eigenvalue

problem

(A AI)+

of the form

= o.

(71)

About a typical internal node such as 5 in Fig. 7, the finite

difference form of (70) is
@z@4+

(4

A)@5@6

@8=0

(72)

vanishes),

and if the correct

(73)

Equations such as (72), suitably amended for boundary conditions, make up the set expressedby(71).
Signs are changed
in (72) to make, as is the frequent convention,
the matrix
positive semidefinite rather than negative semidefinite.
Successive overrelaxation
can be used to solve the matrix
eigenvalue problem
(71). In the first place a fairly crude
mesh, with perhaps 50-100 nodes, is drawn
over the guide
cross section. A guess at the lowest eigenvalue (either an
educated

one or perhaps the result of a direct method

on an

even coarser mesh) is taken as a first approximation.

The
elements of vector @are set to some value, perhaps all unity.
SOR is then initiated generating only one row at a time of
A AZ as required. A nontrivial
solution of (A A&$ can
exist only if the determinant
vanishes, i.e., the guess at k is

k is employed,

of successive displacements
whatever

[34, pp. 260-262]

A = (l:ch) .

(79)

with an assumed or computed A approximation.

The convergence theorem [34, p. 240] states that if B is symmetric
and positive semidefinite,
with all diagonal terms greater
than zero (which can always be arranged unless one of them

the solution
where

AI

(SOR

with

@ is initially.

for 0<0<2

(i#j)
and b,,> O.
As will usually happen,

then the method

u = 1) converges

It will

if the elements

an eigenvalue

to

also converge

estimate

b~j=bi,SO
will not be

correct. If it deviates from the exact eigenvalue by a small

amount we can expect PI, in (77), to be slightly greater than
or less than unity. Therefore
+(~+sl will grow or diminish
slowly as SOR iteration proceeds, It cannot converge to a
solution

as B is nonsingular.

(B is termed

singular

if its

determinant
vanishes.) However, the important
point is that
whether @tends to vanish or grow without limit, its elements
tend to assume the correct relative values. In other words,
the shape of 1$converges to the correct one. After several
SOR iterations
the approximate
@ is substituted
into the
Rayleigh quotient
[35, pp. 74-75]
(80)

WEXLER

: COMPUTATION

OF ELECTROMAGNETIC

FIELDS

427

Equation (80), which depends only upon the shape of +,

is stationary about the solution point. In other words, Z~ ~

most convenient

is a reasonable

near boundaries.

duces

estimate

an improved

superscripts
estimate

to the eigenvector

eigenvalue

give the number

and are therefore

then (80) pro-

estimate.

The

bracketed

of the successive eigenvalue

used in a different

context

from

employ

way to guarantee

variational

methods
Forsythe

give a good account

When

the

matrices

difference

and Wasow

is to

equations

[34, pp.

1821 84]

of this approach.

solution

det (B)= O with

symmetric

in deriving

for

the

the correct

first

mode

A substituted

is

into

obtained,

(79). If one

that in (74). Using the new eigenvalue approximation,

and
returning to the SOR process with the most recent field esti-

then wanted to solve for a higher mode, a new and greater A

estimate would be used, If this differs from the first by a, this

mate, a second and better estimate to @ is found,

is equivalent

and so on

until sufficient accuracy is obtained. Whether or not convergence has been achieved may be gauged firstly by observing the percentage change in two or more successive eigenvalue estimates. If the change is considered satisfactory, perhaps less than one-tenth

of a percent,

Now,

to subtracting

Subtracting

aI from

then the displacement

root

matrix

is often defined as the square

of the sum of squares of all elements.)

within

satisfactory

limits,

These requirements

When

this is

the process may be terminated.

must be compatible

in that one cannot

expect the displacement norm to be very small if the eigenvalue estimate is very inaccurate. What constitutes sufficient
stationarity
of the process is largely a matter of practical
experience with the particular problem at hand and no general rule can be given. This entire computing
procedure,
including

a optimization,

and [38, pp. 114-129].

exists guaranteeing

is described
Moler

convergence

in the outer loop. However,

reasonable

A estimate

in [34, pp. 375-376]

[38] points

out that no proof

with the Rayleigh

experience

to begin with,

indicates
and with

quotient

that with a
other condithe smaller

the mesh interval and the larger the number of equations to

be solved. The number of eigenvalues of Qti,x equals the
order of the matrix

and a large number

that they are closely packed

together.

from (76), that if the dominant

we see that

of eigenvalues means
It is therefore

and subdominant

clear,

eigenvalues

terms of B,

= p~.

(81)

sides,

(1? al)+
all eigenvalues

(s~)

= (p a)+
of the new matrix

(B al) are

shifted to the left by a units. Since B had a zero eigenvalue,

at least one eigenvalue must now be negative. A symmetric
matrix is positive definite if and only if all of its eigenvalues
are positive
nonnegative.
and

[16, p. 105] and positive semidefinite if all are

Therefore, (B aZ) is not positive semidefinite

so the convergence

theorem

is violated.

SOR scheme cannot

higher than the first.

the

Davies and Muihvyk

of the SOR solution
waveguides.

Typical

[32] published
cutoff

of one percent.

sonable

accuracy

wavenumber

corners.

account

shaped hollow
accuracies

This is an interesting

was obtained

for modes

an interesting

of several arbitrarily

fraction

internal

Consequently,

be employed

previous

containing

tions satisfied, we can be fairly confident.

Generally, the higher the accuracy required,

both

all diagonal

of B, then

B+

norm as a percentage of the vector norm should be inspected.

(The norm of a column

a from

if p is any eigenvalue

were a

result as rea-

even for those geometries

Fields

are often

singular

near

such points.
The finite difference
approximation
suffers
because Taylors expansion is invalid at a singularity.
If
errors due to reentrant corners are excessive, there are several approaches available. The reader is referred to Motz
[39] and Whiting
is expanded

[45] in which the field about a singularity

as a truncated

harmonics.

Duncan

shape emerges. In fact, successive overrelaxation

corrections could be so small that roundoff
errors destroy the
entire process and convergence never occurs. The answer is

intervals, etc.
An algorithm
has recently been developed
[27] which
guarantees convergence by SOR iteration. The principle is to
define a new matrix

to start off with a crude mesh having, perhaps, one hundred

nodes in the guide cross section. Solve that matrix eigenvalue
problem,

halve the mesh interval,

in two dimensions,
potentials,
behind

and

preferably)

then

this approach

continue

interpolate
for the newly
the

process.

is that each iterative

[33] gives results

series of circular

of $. ,A(Y1 and wZ)are nearly equal, the process (74) will need
a great number of iterations before the dominant eigenvector

ments employing

different

of a series of numerical
finite

difference

experi-

operators,

(83)

C=~B

(quadratically
defined
The

node

C is symmetric

whether

rational

or not B is. Equation

C+

(78) becomes

(84)

=()

stage (other than

the first) begins with a highly accurate field estimate and so

few iterations are required for the fine meshes. For arbitrary
boundaries, programming
for mesh halving and interpolation can be an onerous chore.
As one additional point, the effect of Neumann boundary
conditions is to make B slightly nonsymmetric
and so convergence of the SOR process cannot be guaranteed.
This
occasionally causes iteration to behave erratically, and sometimes fail, for coarse meshes in which the asymmetry is most
pronounced.
Otherwise, the behaviour of such almost symmetric matrices is similar to that of symmetric ones. The

mesh

which is solved by SOR. Note that

det (C) = det (S) det (B)

= (det (B))

(85)

and so (84) is satisfied by the same eigenvalues and eigenas (71) and (78).
SOR is guaranteed to bc successful on (84) as C is positive
semidefinite for any real B. Note that ix> O for any real column matrix x. Substitute the transformation
x= By giving
Y~By z YCY >0 which defines a positive definite matrix C.
If det (C)= O, as happens at the solution point, then C is
vectors

positive

semidefinite

and so convergence

is guaranteed.

This

IEEE

428

TRANSACTIONS

ON MICROWAVE

THEORY

AND

TECHNIQUES,

AUGUST

1969

much is well known. The usefulness of the algorithm is that

it describes a method of deriving the nonzero elements of

which is parabolic.
Note that if @ is time harmonic,
(86)
becomes the Helmholtz
equation. If @ is constant in time,

one row at a time of C, as required

both becomes Laplaces

by SOR, without

recourse

to Bin its entirety. It is shown that the gth row of C requires

only the gth node point potential and those of twelve other
nodes in its immediate vicinity. The operations are expressed
in the form of a thirteen-point
finite difference operator.
Thus, because storage requirements
positive

semidefinite,

SOR

are minimal,

and C is

can be employed

for

higher

The solution of (86) gives the space-time response of a scalar

wave function. Equation (87) governs the transient diffusion
of charge in a semiconductor,
conductor,

This method

requires considerably

while generating

difference

more logical

equations

decisions,

near boundaries,

than

does the usual five-point

operator.
The process can be
speeded up considerably
by generating (and storing) these
exceptional
difference equations only once for each mesh
size. In this way, the computer simply selects the appropriate
equation for each node as required. In the internal region,
difference

equations

ing them

would

approach

because nodes near boundaries

ber only

as hl

approximately.

are generated

be wasteful.
while
This

very quickly

This

so that stor-

is an entirely

the internal

ones increase

boundary-node

storage

would
likely
be profitable
for the five-point
operator as well.
The method can be adapted to the deterministic
(67). Normally,
this would not be required, but
tempts higher order derivative approximations
at
ary, for the Neumann or Cauchy problem, SOR
[37, pp. 5&53].

Because it guarantees

a suggested abbreviation
Recently,
approach

Cermak

whereby

feasible

increase in num-

positive

as hz

and

in the solution

definiteness,

Silvester

constant

cross section

heat,

problem

It is a function

charge or thermal

mass density,

is, due partly

depending

t= O must be specified

to the additional

and sufficient

by integration.

indepen-

time derivatives

in order to eliminate

at

arbitrary

con-

It is then theoretically

pos-

@for all t.Problems

specified in this way

are known as initial-ualue problems. To be really correct, the

partial differential
equation furnishes us with a boundaryvalue, initial-value
problem.
The finite difference approach is to discretize all variables
and to solve a boundary
For simplicity,

value problem

consider

at each time step.

the one dimensional

[29]

demonstrated

an

Wp

diffusion

equa-

1 alj
.

waveguide

To solve for 4(x, t),

the initial value @(x, O) and boundary
conditions,
say, o(O, t)= O and (tl@/dx) I X=1= O are given.
Discretization
of (88) gives

finite differences

junctions

4%+1,3

~t,~+l

~i,j

(89)

Kk

and discon-

when the structure

one coordinate.

(88)

~K8t

~~1,.i 24~,1 +

along

and

The function

can be used in an open

[40] have employed

is applicable

constant.

and electronic

conductor

tion

differences

of certain

The method

the elliptic

dent variable.

sible to determine

effect of the boundary vanishes. Then, the solution in the

enclosed space corresponds to a finite part of the infinite

tinuities.

specific

a thermal

electrical

upon the physical problem. For example, if a quantity

of
charge (or heat) is suddenly injected into a medium, the
electric potential
(or temperature)
distribution
is given by
the solution of (87). The result o is a function of space and
time,
Such problems are more involved computationally
than

stants produced

region. An arbitrary
boundary is drawn about the field of
interest. The interior region is solved in the usual way and
then the boundary
values are altered iteratively,
until the

region.
Davies and Muilwyk

conductivity,

Kis the diffusion

mobility,

difference

is PDSOR.

finite

of temperature,

procedure

problem
if one atthe boundoften fails

heat flow through

or skin effect in an imperfect

[78, pp. 235-236].

modes.

equation.

The solutions of partial differential equations (86) and (87)

are the transient responses of associated physical problems.

has a

If this is so,

the ports are closed by conducting walls and their finite difference technique
for arbitrarily
shaped waveguides
[32]
may be used. A limitation
is that the ports must be sufficiently close together so that one seeks only the first mode in
the newly defined waveguide. Otherwise, SOR will fail as
described previously.

where h and k are the space and time intervals

Rather than the central difference

formula

respectively.

for second deriva-

tives, forward or backward differences must be used at the

boundary
pointsunless
Dirichlet
conditions
prevail.
The first of each subscript pair in (89) denotes the node
number along x and the second denotes the time-step number. Therefore
~ = ~h.>

i=o,

1,2,

. . .
(90)

C. Parabolic

and Hyperbolic

t = jk;

Problems

Prime examples of these classes of differential

are furnished by the wave equation

equations

Rearranging

1,2,

. . . .

(89)
+t,j+l

(86)

j=o,

@i,j +

(~~-l.~

z+~,~

~~+1,~)

with

(92)

r = Klc/hZ.
which

is hyperbolic

and the source-free

diffusion

(91)

equation

(87)

In Fig. 8, the problem

region, one dimension
presents

an explicit

is visualized

as a two-dimensional
This algorithm

t being unbounded.

method

of solution

as each group

of

WEXLER:

COMPUTATION

OF ELECTROMAGNETIC

FIELDS

429
detail,

easily take 1000 years! There appears to be an

could

answer,

however,

and that

is through

hybrid

computation,

the subject of Section VI.

D. Integral

Equations

As an alternative

to posing

a problem

in terms of partial

differential
equations, it may be cast into the form of an
integral equation. This approach is particularly
useful for
certain
known

antenna problems
where the Greens function
is
in advance. Its efficacy is questionable in arbitrarily

shaped closed regions because the numerical

Greens function,

&

i,j

i-l,j

solution

i+l,j

Greens

=h

o
Fig. 8.

Finite dfierence mesh for exdicit

of an initial value problem.

three adjacent pivots can be used to predict

may

be found

one potential

at

is advanced

in

becomes

that must be satisfied.

method

analytically

as this section is concerned,

approach

without

it is sufficient

out that the finite difference approach

thin, arbitrary antenna, the integral

gives the z component

between

too

to point

can be used. For a

of vector potential.

the source and the observation

R=

be shown that the explicit

itself. The integral

useful in many free-space studies, and when the

function

Insofar

or until error accumulation

criterion

of the
as the

solution

the next time step. In this way, the solution

There is a stability

problem

solution

is as difficult

much trouble.

+(X,O)

time as long as required

unacceptable.

of the original

is therefore

for each source point,

R is the distance
point,

i.e.,

[rrl.

(94)

It can

with one space coordinate

If the antenna is excited by a source at a given point,

the ap-

is valid only when O< r< ~. This restriction,

in conjunction
with (92), indicates that the increased amount of computing
required for improved accuracy is considerable. If h is halved
then k must be quartered. The stability criterion is still more

proximate
current distribution
can be computed.
This is
accomplished by assuming I, to be constant, but unknown,
over each of the n subintervals. The integration in (93) is performed with the trapezoidal rule, thus producing an equation

stringent

in

for problems

having two space dimensions,

ing that O< r< ~. The explicit

is also subject to a stability
Another

approach,

requires the solution

known

solution

requir-

of the wave equation

step must

as the Crank-Nicolson

of all node potentials

method,

before advancing

by the fact that all potentials

be solved

as a system

at each time

of simultaneous,

linear

equations. Thus it is called an implicit method. The final

result is that the implicit method is some three or four times
faster than the explicit one.
The reader will find very fine introductions
in [12] and [44]. Three books,

dealing

to this subject

generally

with finite

differences and with special sections of interest here, are [3 I],

[34], and [35]. In [30] and [41], initial value problems are
discussed. Recently, Yee [46, pp. 302307] reported some
results on transient electromagnetic
propagation.
It is disappointing
to note that in spite of the great amount
of work done on the subject, in practice the solution of many
initial value problems exhausts the capabilities
of modern
digital machines. Problems having two spatial dimensions
can easily take many hours to solve with moderate
Forsythe

and Wasow

[34, pp.

unknowns.

solving

constraint.

the time step. It is unconditionally

stable and so does not
require terribly fine time intervals. This advantage is partially
offset, however,

n equations

accuracy.

1114] have estimated

one

week for such a problem having 10000 nodes. Using a modern computer, the time would be reduced to perhaps onethird of that. A three-dimensional
problem,
solved in fine

Enforcing

the required

are produced

boundary

and so the unknowns

a set of simultaneous,

linear

conditions,
are found by

equations.

(Higher

order integration
schemes may be used if the current distribution along each subinterval is presumed to be described
by a polynomial.)
potential

With

the current

may be calculated

distribution

at any point

known,

the

in space.

A good, descriptive introduction

is furnished by [53]. In
[47], [49] -[51], and [54] the solution of integral equations,
in

radiation

and

scattering

methods

is described.

Fox

practical

aspects of Fredholm

problem)

and Volterra

problems,

through

matrix

[35] discusses mathematical

(corresponding

(initial-value

problem)

and

to the elliptic
integral

equa-

tions. The quasi-TEM

microstrip
problem is dealt with in
[48] and [52]. The major difficulty
is the derivation
of the
Greens function;
the numerical problem is insignificant
by
comparison.
Variational
methods (Section V) offer another
to the solution of integral equations.

approach

V. VARIATIONAL METHODS
This

subject,

increasingly
it is relatively

although

important

not

terribly

new,

is becoming

for several reasons. In the first place,

easy to formulate

the solution

of certain

dif-

ferential and integral equations in variational

terms. Secondly, the method is very accurate and gives good results
without making excessive demands upon computer store and

430

IEEE

TRANSACTIONS

ON MICROWAVE

time. The solution is found by selecting a field which minimizes a certain integral. This integral is often proportional
to
the energy contained in the system and so the method embodies a close correspondence
with the real world.
The literature

on variational

methods

is so scattered that

THEORY

The requirements
ing to the function

are considered

It is hoped that by reviewing

them

au (where

fields, that may be derived from a single vector potential,

are also easily catered for. Although
there are some indications of how to proceed, the author

has not seen any general

computer
methods for fields with all six components
of
electric and magnetic field present. Such fields require both
an electric and magnetic vector potential
function to generate them. Perhaps it would be just as well to solve the
electric and magnetic
potential

rather than through

two

functions.

A. Hilbert

Function

The concept
very simple
functions

fields directly

Spaces

of a Hilbert

function

space is, in principle,

and most useful as well. It consists of a set of

that

obey certain

sider those functions

rules.

belonging

those that are possible solutions

Typically,

we will

con-

to this space as being

of any particular

all

field prob-

lem we wish to solve. For example, the field within a threedimensional

region bounded by a perfectly conducting
surface, having some distribution
of charge
solution of the Poisson equation

enclosed,

is the

likewise

belong

to it. In other

u+ v and au are also twice differentiable

and satisfy the relevant

boundary

conditions.

(U,
V)
=s

An inner product

uvd~
Q

is defined which,
function

(96)

in a sense, gives the component

in the direction

of one

of the other.

This appears rea-

sonable when we recall that it is precisely

in this way that a

Fourier

component,

v (omitting,

for the moment,

plex conjugate*)

of an arbitrary

really

the analogy

here that

spaces becomes obvious.

plex conjugate

function

between

the com-

u is found.

vector

The reason for including

sign will be shown in a moment.

the com-

The integra-

the vector product

u. v. In this work,

sider them to be scalars although

subsequent

derivations

In addition
Hilbert

should

to linearity,

however,

be fairly

functions

u and v belonging

ber (u, v) is generated

that

of the

straightforward.

that are elements

space must satisfy the following

pair of functions

we will con-

the generalization

axioms.

to the linear

For

(95)

of a
each

set, a num-

obeys the following

axioms:

(u, v) = (v, U)*;

= ~.

It is

and function

tion is performed over Q which may be a one, two, or threedimensional

physical space depending on the problem.
In
our example, the limits correspond to the walls. If u and v
are vector functions
of position,
we alter (96) slightly to
include a dot between them, thus signifying the integral of

(aIu,
V1$

u and v belong to the set, then u+ v and

a is a constant)

words, the functions

the

The following theory is concerned almost exclusively with

the solution of scalar potentials. Obviously then, static fields
are the immediate
beneficiaries.
In addition,
time-varying

1969

to belong to a linear set. By this is meant that

if any two functions

the initiate could well spend many months accumulating

required information
before being able to apply it.

AUGUST

(at least) and that they satisfy the homogeneous

Dirichlet
condition
~(s) = O at the conducting
walls. Such functions

here.

to microwave problems, the engineer will be encouraged to

make immediate and more general use of them. Otherwise,

TECHNIQUES,

we have placed upon functions belongspace is that they be twice differentiable

there is good reason to collate and review the principles

these ideas, and relating

AND

(97)

+ a,uz, v) = al(ul,

v) + a,(u,, v);

(u, u) > o;

(98)

(99)

if

@is some function

solution

of position,

i.e., o(P). We know

must be one of or a combination

of functions

(u, u) = O

that the
of the

Note

that

the definition

then

u = O.

of inner

product

(100)

(96) satisfies re-

form u(P) = sin (lm/a)x. sin (nzr/b)y. sin (nm/c)z in a rectangular region. a, b, and c are the dimensions of the rectangular region and 1, m, and n are integers. If the conducting boundary of the box is held at zero potential any one or
summation
of harmonic functions u will vanish at the walls
and will likewise give zero potential
there. These components of a Fourier series are akin to vector components of a
real space insofar as a summation of particular
proportions
of functions yields another function whereas vector summa-

quirements (97~100)
for all well-behaved
functions. Note
also that in a real Hilbert space (i.e., one spanned by real
functions),
(u, v)= (v, u). From axioms (97) and (98) it is
easy to see that

tion of components
defines a point in space. Thus, a summation of harmonic
components
of the above form defines a particular
function which, by analogy, we consider
to be a point in an abstract function space. For this reason,
such functions
are often called coordinate
functions.
The
number of dimensions may be finite or perhaps infinite. The
Fourier series is an example of a particular
function space

to the rules for multiplication

of polynomials.
The essential
difference is that the numerical coefficient of the second fac-

consisting

of orthogonal

however,

these functions

coordinate

functions.

need not be orthogonal.

In general,

(u, au) = a{v, u)*

= a(u,

v)

(101)

where a is a complex number here.

As a result of these definitions
it is clear that an inner
product whose factors are sums can be expanded according

tor must be replaced by its complex conjugate in carrying it

outside of the brackets.
It is clear now why the complex conjugate must be employed in axiom (97). Property (99) states that (u, u)2 O.
Therefore,
due to the linearity
of the function
space, new
elements au must also satisfy (au, au) ~ O where a is any
complex

number.

If property

(100) were of the form

{u, v)

WEXLI?R

: COMPUTATION

= (v, u) then

we would

for arbitrary

complex

not positive

OF ELECTROMAGNETIC

FIELDS

au) =a(u,

u) which,

have

that

a, would

(au,

be a complex

or equal to zero, perhaps

(99) would

quantity

431
concentrate

and

even negative.

Thus

be violated.

In the following,
of each function,

i.e.,

on certain

(58}(60)

Ifj=

with

homogeneous

O, (103) becomes Laplaces equation.

then (103) represents

we will assume implicitly

that the norm

defined by

boundary

geneous Helmholtz
vector potential

conditions,

g =p =g = O.
IfL=

one vector component

equation.

(V2+k2)

of an inhomo-

u is then one component

and ~ is an impressed

current

of the

times p.

inner product of a vector with itself and so I(UI\ is, by analogy, a measure of the length
or magnitude
of the func-

To begin the solution, we must consider a set of all functions that satisfy the boundary conditions
of the problem
and which are sufficiently differentiable.
Each such element
u of the space belongs to the field of definition of the operator
L. Symbolically,
u= D.. We then seek a solution of (103)

tion. Insofar

from this function

]/lL/l == <~~
is finite.

The operation

beneath

(102)

the radical

as a field is concerned,

is akin

to the

it is its rms value.

We consider
B. The Extremum
Among

elliptic

Formulation

problem.

Forsythe

that variational
for elliptic

in:

to parabolic

the variational

of the associated

in solving elliptic

The

approach

partial

is often

conversion

differential

attempted

of a boundary

out

to direct
of

value problem

However,

of the field problem.

seems as unappealing

a
to

due to certain

ing this minimizing

procedures

function,

great computational

On the face of it

as the original
available

the variational

advantages.

problem.

of the generality

general operator

notation

of the method

L. In practice,

adjoint

whenever
U=().

strongly

The significance

is a function

For convenience,
identity

associated

on

in the
is self-

boundary

(107)

>0

zero and vanishes only when

is best illustrated

by a

V2. Therefore

vvzudo.
Q

(108)

S::ds=JnVuVodQ+

SnvV2udfi

Greens

(109)

of
converts

(108) to the form

v} =

Vu. Vudn

(10.!J

all
Vd!s

(110)

.s drb

$?

in which the last integration

If

L=vz

of the min-

is a requirement

take u and v to be real functions.

(103)

of position.

U)

(Lu, U) =

problem

the

of these terms

Let L=

(h,
Lu

upon

u is not identically

simple example.

is written

where~=~(l)

depends

(Lu,

is the use of a

operations may be denoted by this single letter.

1) The deterministic problem: The deterministic

in the proof

and therefore

convergence

a great variety

(106)

conditions.
In addition,
the self-adjoint
operator will be required to
be positive definite. The mathematical
meaning of this is that

has

can be guaranteed under certain very broad conditions and

this is of considerable
theoretical
and numerical
consequence.
Illustrative

theorem

for determin-

formulation

In addition,

Lv).

those problems
treated
by variational
methods
fashion described here. Whether or not an operator

equation.

by means

To have a se~-

V)
=(u,

It will be seen that (106) is required

prob-

on .r only. (s is the

we must have

imal functional

choiceto find the function that minimizes the value of a

certain integral, The function
that produces this minimal
the alternative

adjoint problem

(Lz.L,

an integral equation. It frequently

becomes overly complicated, if not altogether impossible to handle, because the
Greens function
itself is difficult
to derive. On the other
hand, a variational
formulation
presents an alternative

value is the solution

The self-adjoint-

of Q and may be at infinity.)

They

open. The prin-

which is an alternative

integration

point

of u and v and their derivatives

significant

problems,

problems

method

lems rests on an approach

Greens function

[58, pp. 163-164]

are computationally

function
boundary

and the eigenvalue

but not for hyperbolic

leave its application

latter

there are two classes

the deterministic

and Wasow

problems

ciple behind

equations,

approaches

operators.

ness of L means that (Lu, v) (u, Lv), in which u, v= DL, is a

differential

we are interested

space.

only se~-a~oint

is performed

over the boundary.

n is the outward normal. The one-dimensional

(110) is integration
by parts. Similarly

analogue

of

and
.f~
is a known

charge distribution,

(105)
e
we require

to find

u, the

solution of the problem under appropriate

boundary conditions. (The minus sign in (104) makes the operator positive
definite, as will be shown.) Commonly,
these boundary conditions take the forms (58)-(60). In the first instance we will

Under
either the homogeneous
Dirichlet
or Neumann
boundary conditions, the surface integrals in (1 10) and (1 11)
vanish.

Under

the homogeneous

Cauchy

boundary

condi-

tion, they do not vanish but become equal. At any rate, L is

therefore self-adjoint
under any one of these boundary conditicms or under

any number

of them holding

ove~ various

IEEE

432

TRANSACTIONS

ON MICROWAVE

sections of the boundary.

Property (106) is akin to matrix
symmetry.
Positive definiteness of L is readily observed by making
u= v in (110) and substituting
geneous boundary
conditions.

any of the previous homoAn additional

requirement,

for the homogeneous

Cauchy

that u> O.
It is a consequence

of these properties

the following

statement:

then the equation

to satisfy

condition

cannot

L is positive

definite

w= u1 uZ. Since the operator

requirement)
we obtain
(Lw,
definite,

problems

involving

of a partial

elliptic

differential

partial

equation,

a number

or more functions

The functional

with,

of one

It is, in a sense,

for the soIution

general form,

for complex

of

is
(112)

in which we assume that u and f are real functions.

functions

F = (Lu, u) (ujj)

is

which

occurs

for

the exact

solution

definite

operator,

and if Lu =f

Lu. = f.

Fmin =

(114)

uOV2uOdfi

(115)

Add and subtract (LuO, UO)to the right-hand side of (1 15) and
rearrange noting that if L is self-adjoint
(Luo, u)= (Lu, Uo) in
a real Hilbert space. Finally, we obtain

F = {L(u w), ?A WJ) (Luo, Z@).

Fmin = JI

tained

upon

function,
locating

a stationary

once such a point

solution.

approach,
point,

is located,

This is important

that the operator

the miniAs it relies

we wish to ensure that

it in fact corresponds

minimal

be positive

to the

derivative

is

a necessary, although not a sufficient condition

for a minimum. In other words, if a function uOG DL causes (112) to
be stationary, is UOthen the solution
show that this is so. Let
5(,) = F(uo + q)

ing the appropriate

adjoint,

of (103)? It is easy to

F(uo)

real number.
expressions,

Using

and taking

(120)
(112), substitutL to be self-

we obtain
~(e) = 2c(Lu,

after some algebraic

f, ~) + e2(Lq, q)

manipulation.

(121)

Differentiating

dti
= 2(LU0 f, 7?)+ w%
de
which must vanish at a stationary
occurs when c= O, therefore

point.

~)

(116)

If this is to hold,

for arbitrary

functional
definite

and

(122)

By hypothesis,

this

(123)

v, then we must have that

LUO =f

identically.
In other words, the stationary point corresponds to the solution.
2) The eigenualue problem:
Functional
(115) cannot
a priori

be used for the eigenvalue

Lu =

as is f in (103). The relevant

problem, is

requires

method.

because a vanishing

because the right-hand

theorem

(119)

used for finding

is the Rayleigh-Ritz

always and the first is ~ O. F assumes its least value if and

this

Vu. I 2dL?

energy !

The most common

the last term on the right is positive

summarize,

say that

This integral is proportional

to the energy stored in the
region. The field arranges itself so as to minimize the con-

only

u= Uo. To

(118)

s Q

(Luo j, TI) = O.

F = (Lu, u) 2(u, Luo).

if

V2, we

where the integration

is over a volume Q. Now,
O(S)= O. Therefore, using Greens theorem,

(114) into (112) for f. Thus

definite,

L=

UO. Taking

has a solu-

tion, then (112) is minimized by the solution uO.(The proof of

(113) is not much more involved.) Any other function u~l)~
will give a larger value to F. The proof follows.
Take the function U. to be the unique solution, i.e.,

As L is positive

(117)

get

where c is an arbitrary
(113)

It is seen that the last two terms of (1 13) give twice the real
part of (u, f ), Concentratingon(112),
we will now show that
if L is a positive

value of the functional

F ~i. = _ (LuO, ZJo)

The more

(f, u).

that it is a better approx-

(116), note that the minimal

mizing

A simple example is the inner

we are concerned
problem,

limits.

F = (Lu, u) 2(u, ~)

Substitute

From

it was

a functional

that depends on the entire form

between prescribed

some measure of the function.

product (u, v).
the deterministic

variables,

1969

dif-

stated earlier that we will attempt to minimize

a certain
integral. The rule for forming this integral, and subsequently,
ascribing a value to it, is a particular example of afunctional.
Whereas a function produces a number as a result of giving
values to one or more of independent

AUGUST

is

we must then have w = O and so u1= uZ, thus proving

of boundary-value
ferential equations,
For the solution

produces

to F(u) and delude us into thinking

imation to the solution.

is a linear one (a further

w)= O. Since L is positive

that no more than one solution can exist. This is simply a

general form of the usual proofs for uniqueness of solution

TECHNIQUES,

nition of the operator L; i.e., they must be sufficiently differentiable and satisfy the boundary conditions.
Otherwise,
a function, which is not a solution, might give a lesser value

have more than one solu-

tion. The proof is simple. Suppose the equation to have

two solutions U1 and w such that Lul=f
and Lu2 =f. Let

AND

self-adjoint
under the stated boundary conditions.
It is also
required that the trial functions come from the field of defi-

that we can make

if the operator

Lu=f

(107), is

THEORY

problem
(lp~)

iU

side of (124) is not a known

functional,

function

for the eigenvalue

WEXLER:

COMPUTATION

OF ELECTROMAGNETIC

FIELDS

~ = @u, u)

433
instead,

(125)
(u, u)

Successive eigenvalues

by defining

orthogonal

The differential
where UG DL. Equation

(80) is one particular

It is often called the Rayleigh

bound,

attained

eigenvalue
function.

quotient.

instance of it.

If F~in is the lowest

for some ul# O, then F~in = x is the lowest

of operator

L and U1 is the corresponding

equation,

is known

in finding

L, i.e., VG DL. Let a be an arbitrary

ul+a~e

under which

F is stationary

about

Let
of

real number.

DL. We want to investigate

the conditions

u1. Substitute

functional

whose Eulers

solve, e.g., the Helmholtz

equation,

(126)

conditions

more
various

Boundary

+ aq), u, + a~)

(w + q,

a restriction

problems

(e.g.,

multiconductor

potentials)

to be excluded.

cannot

ment. In addition,

be proved.

the derivative

(127)

With

~)(~~1, ~1) (L~l,

ul)(ul,

F= F~i. and u= ZL substitute

only of a,

by hypothesis,

vanishzs when a = O. We therefore

(Lul,

This

the space is nonlinear

lines

happens

Substitute

get

B1tt Is =

tions of position

at

because

(58), say, into

this state-

as well.

boundary

conditions

in the

T?) = O.

(128)

(Lu,, u,) from (125) into

required

operators

examples.

and the b~ are given funcEquations

The number

depends upon whether

and upon the order of the differential

Assume that a function
ciently differentiable

(135)

B2u /, = bz, . 0 .

h,

on the boundary.

the most common

ditions

of position

(58>(60)

of boundary

or not u is a vector

w exists which is sufficonditions

(135).

As w satisfies the boundary

B1w/.=bl,Bzwl.=bz,....
Since v is arbitrary,

are
con-

equation.

and satisfies boundary

u is not necessarily the solution.

conditions,

(128). Rearranging

(136)

Putting

and so FniD is the lowest eigenvalue

It is fairly

that if the minimization

of (125) is attempted
functions u orthogonal
to uI, in the sense
=

with

trial

that F~i. equals the second eigenvalue AZ. Similarly, defining

a Hilbert space orthogonal
to U1and u2.,the next eigenvalue
and eigmvector
results, and so on.
As the minimum of (125) corresponds
value k, we can rewrite

the functional
~ <

as the B~ are linear operators.

(137)
(138)

We have now achieved homo-

geneous boundary conditions.

Instead of attempting
a solution

(131)

o,

w )

B,v 1, = O, B2u /, = O

Xl and U1the correspond-

easy to show [62, pp. 220-221]

(u, u,)

V=u

(130)

LU1 F~i. UI = O

ing eigenfunction.

bound-

form

F is a function

respect to a. Then,

etc.

causes the

of (1 10) and (111) to verify

where the B; are linear

(127) with

homogeneous

Such

important

u, + (q)

As UI and 7 are fixed functions,

we wish to
equation,

Conditions

Let us express inhomogeneous

(L(u,

equations

exclusively.

the surface integrals

into (125) giving

F=

is a minimizing

Laplaces

In solving Lu =f we have considered

ary

self-adjointness

U=ul+a?l

Differentiate

whose solution

as the Eulers equation. We are interested

C. Inhomogeneous

Therefore

are found

eigen-

The proof of the preceding statements is quite direct.

v be an arbitrary
function
from the field of definition
operator

function,

and eigenvectors

spaces as before.

of
(139)

Lu=f
we examine
Lv = L(u

W)
(140)

= j Lw.

to the lowest eigen-

in the following

form:

Let
f,=

(Lu, U)

(141)

fLw

(132)

The numerator

(u, u)

of (132) is positive

operator. The denominator

rearranging
the inequality

and so we can now attempt

as L is a positive

is positive

definite

by (99). Therefore,

(Lu, U) h{w, U) >0.

(133)

We know that (133) is an equality only for the correct eigenvalue and eigenfunction.
Consequently,
the left-hand side is
otherwise greater than zero. Therefore, as an alternative to
minimizing
(125), we can seek the solution of (1 24) by mini-

Lv =

of
(142)

fl

under homogeneous boundary conditions (13 8). In any particular case, it still remains to prove self-adjointness
for
operator L with functions
satisfying (138). If this can be
accomplished, then we may seek the function that minimizes
F = (h,
Substitute

(143)

v) ~(v,

fl).

(137) and (141) into

F = (Lzt, U) 2(tLjf)

mizing

a solution

(u, Lw)

(143). After

(h,

expansion,

w)

(144)
F = {Lu, u) ~(u, u)

(134)

+ ~(w, f) @w,

w}.

IEEE TRANSACTIONS

434
f is fixed and w is a particular

function

ON MICROWAVE

THEORY AND TECHNIQUES,

that the existence of a function

w was an unnecessary
assumption.
Although
we shall not demonstrate
it here, functional

selected (which we

need not actually know). Therefore, the last two terms are
constant and can play no part in minimizing
the functional
as we have assumed that u is selected from the set of func-

may

tions that satisfies the required

Cauchy problems.

boundary

condition.

Other-

be derived

for

the

solution

of electrostatic

from (144) because of this.

functions

of position.

to examine

that u and w maybe

(u, Lw) (Lu, w) in the hope

separated.

If this attempt

D. Natural

is successful,

The symmetrical

sn

(Wvu

form

8u/~nl.

of Greens theorem

normal

to s. Therefore,

substituted

V2. The meaning

the certainty

that

problem.

and

The form

of the functional

of this

differenti-

the minimal

by (112) will be due to the solution

On the other hand, we cannot entertain

the third

with

L=

to test any sufficiently

value

and none other.

this confidence

the Dirichlet boundary condition.

It is easy to show this for the homogeneous

ds (147)

media that are

that trial functions

for the operator

able functions
attained

is

ZL:

the

involving

Conditions

is that we are now permitted

(146)

g(s).

and

=p(s) and du/&z] ~+u(s)u(s) = q(s) are natural bound-

ary conditions

Uvw)dfl =J( W ;

where n is the external

fourth

(145)

condition
u(s)

Neumann
to formulate

into (112) should each satisfy the stipulated boundary conditions. Except for the simplest of boundary shapes, such a
constraint
makes it practically
impossible to select an appropriate
set of trial functions. It turns out, however, that

Solve

under the boundary

Boundary

problems

Thus far, we have required

then an amended version of (144) may be written which

excludes the unknown w.
Let us illustrate these principles with a practical example.

V%=f

inhomogeneous

It is also not too difficult

wise w would depend on u. The last two terms may be deleted

It now remains

AUGUST 1969

under

Neumann

is

terms of (144) are

(154)

F=sQ(]Vu12-2.fu)dQ.
(u, Lw)

(Lu, W) =

WV2U

wv2w)df2

Jfl

(148)
du

-s(s

W~n

dw
ds.

8n )

Since
u(s)

Substitute

u= UO+CZVwhere

n need not

be in the field

of

definition
of L. Differentiate
with respect to a, make a= O,
and set the result to zero. Finally,
employing
Greens
formula,

(149)

= w(s) = g(s)

we have
(zL, Lw)

(LzL, w) = ~

(,;

g ;;)ds.

(150)

s
Only the first term on the right-hand

side of (1 50) is a func-

Nowhere

has q been required

tkJdn

solution

= O. Thus

boundary

are then left with

E. Solution

to be minimized

Simplify

sQ
Vzu-ly+f%

(151) using identity

the

is found

with

V2U0

condi-

=f

and

appropriate

conditions.

A number
F=

boundary

(155) can hold only if

tion of u. In addition
to neglecting the last two terms of
(144), the last term of (150) maybe disregarded as well. We
a new functional

to satisfy

tions. As v is arbitrary

by the Rayleigh-Ritz
of functional

Method

have been derived,

the minimiza-

15])

tion of which produce solutions of differential

or integral
equations. The remaining question is how to locate the minimizing function.
The most popular approach is the Ray-

(109)

leigh-Ritz
method.
Assume a finite sequence of n functions

fudfi.
a

Because of (146), the second and third

we are left with

which is to be minimized
inhomogeneous
boundary

integrals

cancel and

for the solution of (145) under

conditions
(146). It happens, in

this case, that (153) has the same form that homogeneous
boundary conditions would produce. It also turns out, here,

where the aj are arbitrary numerical

(156) into (112). Therefore

j,k=l

We now wish to select coefficients

imum, i.e.,

coefficients.

Substitute

j= 1

aj so that (157) is a min-

WEXLER : COMPUTATION

OF ELECTROMAGNETIC

dF

~=l,z...~.

-==0;

aai

FIELDS

435
of Lu =f

(158)

and Lu = Au as n approaches

the matrices

accuracy to result. Notice

Rearranging

(157) into

powers

of a;,

These characteristics
be employed

infinity,

need not be very large for

In practice,

a high

degree of

also that these matrices are dense.

determine

that direct methods

should

in their solution.

F. Some Applications
+ terms not containing
Now,

write

k instead

The bibliography
lists several useful references for the
principles
of variational
methods. See, for example, [58],

a~.

of j in the second summation,

and

assuming that L is self-adjoint

[61], and [63]. One of the most detailed

able is [62].
Bulley

F = (L+i,

@~)U;2+ 2 ~

(U@~,dk)a~ak
(160)

k#i

2(f, @~)U;+
Differentiating
zero

. . , .

(160) with respect to Ui and setting equal to

(L&,

(161)

~, > ak = (j, ~,)

k=l

where i= 1, 2, . . . , n. Writing

(161) in matrix

[57] solves the TE modes in an arbitrarily

guide by a Rayleigh-Ritz

over the waveguide

cross section. Obviously,

the homogeneous

method
fairly

is inapplicable

the methods of Section II-A.

By a very similar approach
193194], the Rayleigh-Ritz
value problem gives

[62, pp. 226-229],

method

applied

[63, pp.

to the eigen-

regular

eigenvalue

problem

of form

the eigenvalue

is im-

function

the waveguide

has difficulty

everywhere.

is

in ap-

In such a

into two or more fairly

coefficients

in

[67] solves the TE problem by the use of Lagrange

In this way, he permits all trial functions while
the final

result

boundary

to approximate

condition.

the homoge-

Although

not essential

to his method, he employs a polar coordinate

system with
polynomials
in r and trigonometric
0 dependence.

an

is a matrix

(which

and so Bulleys

regions and solves for the polynomial

constraining

which

condition
naturally

in this case. If the guide boundary

the potential

neous Dirichlet

trial functions are orthonormal,

along the diagonal giving

these trial func-

each. In doing this, his approach is virtually

that of the
finite-element
method. It differs from the usual finite-element
method in the way that he defines polynomials
that straddle
subdivision boundaries while others vanish there.
Thomas
multipliers.

1?1
..=
1]

satisfied

a single polynomial

case, Bulley subdivides

al, az, . . . , an by

Dirichlet

E,) is not

complicated,

proximating
may be solved for the coefficients

shaped

The series of trial func-

tions cannot be chosen to satisfy all boundary conditions.

However, it turns out that the homogeneous Neumann condition (which H. must satisfy) is natural and so no such constraint need be placed on the trial functions. On the other
posed upon

which

approach.

avail-

tions, representing the axial magnetic field, are each of the

form x~y thus constituting
a two-dimensional
polynomial

hand,

form

treatments

(21). If the
k occurs only

Another
possible approach, when boundary
conditions
are not natural, is to alter the functional
in order to allow
trial functions to be unrestricted
[64, pp. 1131-1 133].
By a transformation,
the microstrip

Yamashita

problem

and Mittra

to one dimension.

the fields and line capacitance,

[68] reduce

They then solve

of the quasi-TEM

mode.

The finite element method is an approach whereby a region

is divided into subintervals
and appropriate
trial functions
are defined over each one of them. The most convenient
shape is the triangle,

for two-dimensional

problems,

and the

tetrahedron
in three dimensions.
These shapes appear to
offer the greatest convenience in fitting them toegether, in
approximating
complicated boundary shapes, and in satisfying boundary conditions whether or not they are natural.
Similarly,

(162) would

have diagonal

the

Silvester [65], [66] has demonstrated the method in waveguide problems.

An arbitrary
waveguide is divided into

is self-

triangular
subintervals
[65]. If the potential is considered to
be a surface over the region, it is then approximated
by an
array of planar triangles much like the facets on a diamond.
Higher
approximations
are obtained
by expressing the

terms only, giving

solution by the Fourier analysis method.

Because the functions are all real and the operator

adjoint, from (97) and (106) we see that both the A and B
matrices are symmetric. Also, B is positive definite which
permits it to be decomposed, as in (23), using real arithmetic
only, thus resulting in an eigenvalue problem of form (22).
It can be shown that (162) and (163) ammoach the solution

potential within each triangle

Because of high accuracy,

by a polynomial
[66].
the finite element approach

IEEE TRANSACTIONS

436
appears

useful

for

three-dimensional

requiring excessive computing

The method was originally
ing applications
application

[70],

previous
Other

references).
implementations

scribed

in [59] and [60].

(e.g.,

[55],

of variational

-+Jt)

0
+i.,(t)

engineer-

+I+l(t) 0

seen increasing

[56] as well
methods

THEORY AND TECHNIQUES,

EG=lRl

without

for civil

[71 ], but has recently

in microwaves

VI.

problems

[69].
expounded

ON MICROWAVE

are de-

Fig. 9.

n
*
~t

Single node, DSCX analog of the onedimensional diffusion equation.

HYBRID COMPUTATION

.A

I
T*

as the

AUGUST 1969

In Section IV-C two methods for the solution of initial

value (transient)
problems were introduced.
Through
the
explicit approach, one is able to predict the potential of any
node at the next time increment
cent node potentials.

as a function

The disadvantage

of a few adja-

is that

a stability

,+O(t)

constraint
demands very small time steps, On the other
hand, the implicit method does not suffer from instability,
and permits larger time steps, but requires simultaneous

solu-

tion of all node potentials for each step. As a result, both

techniques are very time consuming and sometimes impossibly slow.
The hybrid

computer

offers a significantly

different

ap-

proach to the problem. It consists of two major partsan

analog and a digital computer. The analog is a model that
obeys the same mathematical
laws as the problem
being
considered. So the analog, which is presumably
easier to
handle, simulates the response of the system being studied.
More precisely, the particular
form
is known as an electronic dl~erential
electronic

units (which

perform

etc.) together, it is possible

equations
under appropriate

of analog intended here

analyzer. By connecting

integration,

multiplication,

to solve ordinary
initial
conditions

time-consuming

almost

immediately.

natural

ordinary

steps,

the

The analog

differential

analog

gets the

computer

equation

is faster;

it is a

This is fairly obvious for an ordinary differential equation,

but how is a partial differential
equation to be solved? Consider a one-dimensional
diffusion equation
Wp

11$
(165)

==%
(Many of the following
comments apply to the wave equation as well.) Discretize the spatial coordinates
at the ith
node. We have, using central differences,

~$d
t)

.2

.4

.6

.8

FJt)
1.0

time @i--l(t) and &-l(t)

are forcing functions which are as
yet unknown. Assume, for the moment, that we know them
and let us see how the analog computer can produce the time
response Oi(t).
Fig. 9 indicates, symbolically,
the operation of an analog
computer

in solving

larger triangle
an inverter.
known,

(166). Circles

represents
Assume

recorded

that

perhaps,

with the initial.value

@,(t) = :h,

s, (+i-l(t)

In fact, the integrator

indicate

an integrator,

multipliers,

the

and the smaller

and played

back into

the inte-

4,(0). They are all multiplied

in the ratios indicated.

2@i(0 + b+l(t))dt.

produces

one

r&.l(t)
and ~ti l(t) are

functions

l/h2 and fed into the integrator

is then integrated giving

answer

solver.

~+3(t)

t)

Fig, 10. Simultaneous solution of the one-dimensional difTusion equation, by DSCT analog, with four internal nodes.

solution is given as a continuous

waveform.
Whereas the
digital computer will solve the problem in a number of discrete

)+2(

grator

differential
[74]. The

l+,(t)

the negative

by
This

(167)

of (167), i.e.,

@;(t). This is fed back thus completing

the circuit and
allowing the process to continue to any time t. An inverter
follows the integrator to alter the sign if required.
We do not, of course, know the forcing functions

+;-l(t)

and &l(t). They are the responses of adjacent nodes, and

they in turn depend upon forcing functions defined at other
@o(t)and @s(t)
nodes. However, the boundary
conditions
are known in advance as well as the initial conditions
4,(0)
for all nodes, or +(x, t) where t= O.
To solve the entire finite difference system at one time
requires as many integrators
as there are internal
nodes
available. This is demonstrated in Fig. 10. (The factor l/h2 is
incorporated
into the integrators
for simplicity.)
What this

At boundaries,
forward
or backward
differences must be
used.
We have therefore reduced the partial differential
equa-

scheme does, in fact, is to solve a system of four coupled

ordinary
differential
equations in parallel. The analog has
two obvious advantages:
1) rapid
integration;
and 2)

tion to a system of ordinary

differential
equations, one at
each node point. This is known as the DSCT (discrete-space-

parallel processing.
In order to reduce discretization

continuous-time)
analog
technique.
Other
formulations
exist as well. The time response of the potential
at i, i.e.,

number

cA(t),

may be found by integrating

(166). Other functions

of

of nodes. If the intention

tials simultaneously,
available,

the number

error,

one increases the

is to solve all node poten-

due to the limited

number

of integrators

of nodes must be small. One alterna-

WEXLER : COMPUTATION

OF ELECTROMAGNETIC

tive strategy is to attempt

of the digital
an initial
node,

relaxation

an iterative
procedure.

guess at the transient

frequently

choosing

shown by the uppermost

initial

technique

437

reminiscent

km?

This is done by making

response

just

FIELDS

constant

of each internal
time

responses

curve in Fig, 11. Having

as

made this

,,

+2(7

,,, ,,,

=rd

-,

guess at each nodes time response, each @i(t) is solved

sequentially as described previously. Each di(t),

solved, is transmitted
via an ADC (analog-digital

upon being
converter)

to the digital computer where it is stored as discrete data in

time. Attention is then focused upon node i+ 1 with adjacent
potential

responses

(digital-analog
puted

transmitted

conversion)

by the analog

acting as forcing

from

store through

equipment.

with

Fig. 11.

DAC

Iterative

DSCT solution

of the diffusion

equation.

OwI(t) is then com-

the smoothed

functions.

d,(t)

and @A,(t)

The flow of data is indicated

by

arrowheads
in Fig. 11. In sequence then, node transient
responses are updated by continually
scanning all nodes
until convergence is deemed to be adequate. In effect, this
procedure involves the solution of coupled ordinary differential equations,

one for each node, by iteration.

An additional

refinement

tial at a time but groups

catered for is, as pointed
analog equipment
the solution
vantages

of them. The number

out before, limited

available.

of the entire

over the purely

Fig. 12.

is to solve, not one node poten-

This parallel
problem

digital

(a) Iterative, parallel processing scheme for the initial

problem. (b) Method of lines for the elliptic problem.

value

that can be

by the amount
processing

of

speeds

and is one of the ad-

scheme. Furthermore,

be-

The

analog

boundary

computer

does not

value problems

A common

technique

as easily

as it does initial

is to attempt

various

solve two-point
value problems.
initial

values at

vantage of instability.
Parallel solution of blocks of nodes is the logical approach
for two dimensional initial value problems. Fig. 12(a) shows

one end of the strip until the far-end boundary condition is

satisfied. This can often be very wasteful. A better idea is to
solve each two-point
boundary value problem as two initial
value problems [7, pp. 239240], [74, pp. 8385]. This could
be done for each strip in turn or perhaps in groups. The
entire region must then be scanned repeatedly,
in this

one possible

fashion,

cause of the continuous

time response (i.e., infinitesimal
time steps), we have an explicit method without the disad-

Forcing

format

involving

functions

nodes excluded

the solution

correspond

from

to

the enclosed region.

being considered

would

the block format

near boundaries.

at nine nodes.

solutions

scan the region

at the twelve

The set of nodes

with

alterations

The making

ical decisions, as well as storage, is the job of the digital

puter.
Hsu

and Howe

[75] have presented

to

of such log-

a most

com-

interesting

feasibility

study on the solution

of the wave and diffusion

equations

by hybrid

The procedures

above

are more

computation.

fully

Howe did not actually

the time

of their

explained

in their

have a hybrid

experiments;

through a digital simulation

partial differential
equations

their

paper.

results

Hsu

and

available

at

were obtained

study. Hybrid computation

of
is still in its early stages and

little has been reported on actual computing

times. However, some preliminary
reports indicate speeds an order of

convergence

and [77]. Finally,

method

occurs.

the hybrid

system makes the Monte

[74, pp. 239242, 360] a more attractive

Carlo

one.

It should be emphasized that hybrid solution of partial

differential
equations is still in its infancy. This section, in
part an optimistic
forecast, is intended to show that digital
computing
considered

mentioned

computer

until

For other approaches to the solution of partial differential

equations by hybrid computation,
see [72], [73], [74], [76],

has no monopoly
and certainly should not be
a more respectable
branch of computing.
The

analog machine integrates rapidly, and the digital machine

has the ability to store information
and make logical decisions. It therefore
each in its own

stands to reason that working

domain,

substantial

as a pair,

advantages

will

be

gained.
VII.

CONCLUDING REMARKS

magnitude greater than digital computing for such problems.

An intriguing possibility, for hybrid solution of the elliptic

The intention
of this paper is to familiarize
the reader
with the principles
behind the numerical
analysis of elec-

problem,

tromagnetic
fields and to stress the importance
understanding
of the underlying
mathematics.

is the method

presented
that

in

of lines. The mathematical

in [62, pp. 549566].

discretization

is performed

a two-dimensional

problem,

The principle
along

giving

strips (Fig. 12(b)). In three dimensions,

discretized,

producing

prisms.

theory is

behind

one coordinate

it is
only,

us a sequence
two coordinates

We thus obtain

a number

of
are
of

coupled difference-differential
equations.
Each equation is
then integrated under boundary conditions at each end of its
strip and with forcing functions supplied by adjacent strips.

of a clear
Improper

numerical technique causes one to run up against machine

limitations
prematurely.
In the immediate
future, emphasis should perhaps be
placed upon the development

of finite

difference

and varia-

tional techniques for solving fields having all components of

electric and magnetic field present everywhere.
To date,
with a few exceptions (e.g., [59, pp. 1721 88]), most methods

IEEE

438

ON MICROWAVE

TRANSACTIONS

appear to permit solution only of fields deriveable from a

single scalar potential.
It is not difficult to formulate
and
solve field problems in a multidielectric
region, but it may
not correspond to the actual electromagnetic
problem. This
is indicated

by the continuing

microstrip

discussion

Variational

methods

extent

than

are being

ever before.

implemented
Hybrid

now

to a

computation

is

permit optional analog equipment to be added, the analog

elements to be connected and controlled from the program.
Beyond this, it is virtually impossible to predict. It is futile
wishing

for

computer

solving

machine

capabilities

development
sight into

technology

requirements.
already.

of new numerical
the physical

the engineer is ideally

to keep pace with

We

have

our

outstripped

The greatest

hope

techniques.

all

is in the

Due to his in-

processes and his modeling

ability,

suited to this task.

ACKNOWLEDGMENT

The author wishes to express his appreciation

to P. A.
Macdonald
and G. Oczkowski for discussions on numerical
and programming

techniques

conversations

microwave

on

by M. J. Beaubien
and

hybrid

number

problems.

and B. H. McDonald,

computing

W. G. Mathers,

and to Dr. J. W. Bandler

respectively,

of Atomic

of helpful

Energy

Views

for

expressed

on finite differences
were

invaluable.

of Canada Ltd.,

suggestions. Also to be thanked

are due W.

J. Getsinger,

Special Issue, for the invitation

Guest

are librar-

Editor

Analysis

Introduction
to Numerical
Analysis.
New York:
McGraw-Hill,
1956.
[8] E. Isaacson and H. B. Keller, Analysis
of Numerical
Methods.
New York: Wiley, 1966.
[9] M. L. James, G. M. Smith, and J. C. Wolford, Applied Numerical
Methods for Digital
Computation
with Fortran.
Scranton, Pa.:
International
Textbook, 1967.
[10] L. G. Kelly, Handbook
of Numerica[
Methods
and Applications.
Reading, Mass.: Addison-Wesley,
1967.
[11] C. Lanczos, Applied Analysis.
Englewood Cliffs, N. J.: PrenticeHall, 1956.
[12] M. G. Salvadori and M. L. Baron, Numerical Methods in Engineering.
Englewood Cliffs, N. J.: Prentice-Hall,
1964.
[13] J. H. Wilkinson,
Rounding
Errors
in Algebraic
Processes.
London: H. M. Stationery Office, 1963.
[7]

Finite

E@ations

Systems.

Englewood

Computer

Solution

of Linear

CM%, N. J.: Prentice-Hall,

1967.
Englewood Cliffs, N. J.: Pren-

Dl~erence

Solution

of Partial

Dl~erential

Equations

[27] M. J. Beaubien

[28]

to write this review paper.

[1] I. S. Berezin and N. P. Zhidkov, Computing

Methods,
vols. I and
II.
Oxford: Pergamon, 1965.
Methods
of
[2] D. K. Faddeev and V. N. Faddeeva, Computational
Linear Algebra.
San Francisco: Freeman, 1963.
to Numerical
Analysis.
Reading,
[3] C.-E. Froberg, Introduction
Mass.: Addison-Wesley, 1965.
Methods.
London: H. M.
[4] E. T. Goodwin, Modern Computing
Stationery Office, 1961.
Methods for Scientists attd Engineers.
[5] R. W. Hamming, Numerical
New York: McGraw-Hill,
1962.
Analysis.
New York: Wiley,
[6] P. Henrici, Elements of Numerical

1964.
F. B. Hildebrand,

1969

Theory.
[16] J. N. Franklin, Matrix
tice-Hall, 1968.
[17] L. A. Hageman and R. B. Kellogg, Estimating
optimum overvol. 22,. .DV.
relaxation parameters, Math. Computation,
. 60-68,
January 19&S.
[18 ] T. Lloyd and M. McCallion,
Bounds for the optimum overrelaxation factor for the S.O.R. solution of Laplace type equations over irregular regions; Computer J., vol. 11, pp. 329-331,
November 1968.
[19] T. J. Randall, A note on the estimation of the optimum succesComputer
sive overrelaxation parameter for Laplaces equation,
J., vol. 10, pp. 400-401, February 196S.
[20] J. K. Reid, A method for finding the optimum successive overComputer J., vol. 9, pp. 201-204, August
relaxation parameter;
1966.
Methods
for
the Solution
of Equations.
[21 ] J: F. Traub, Iterative
Englewood Cliffs, N. J.: Prentice-Hall, 1964.
Iteratioe
Analysis.
Englewood ClilTs, N. J.:
[22] R. S. Varga, Matrix
Prentice-Hall,
1962.
[23 ] D. C. Walden, The Givens-Householder
method for finding
eigenvalues and eigenvectors of real symmetric matrices, M.I.T.
Lincoln Lab., Cambridge, Mass., Tech. Note 1967-51, October 26,
1967.
of Numerical
Matrix
Inversion
and
[24] J. R. Westlake, A Handbook
Solution of Linear Equations.
New York: Wiley, 1968.
New York:
[25] J. H. Wilkinson, The Algebraic Eigenvalue Problem.
Oxford, 1965.
[26] J. H. Wilkinson,
Householders
method for the solution of the
Computer
J., pp. 23-27, April 1960.
algebraic eigenproblemfl

of this

REFERENCES
General Numerical

AUGUST

Dr.

made a

ians, R. Thompson, Mrs. F. Ferguson, and Mrs. D. Globerman, for locating and checking many references, and Miss
L. Milkowski
for carefully typing the manuscript.
Thanks

TECHNIQUES,

1961.
[15] G. E. Forsythe and C. B. Moler,
Algebraic

likely to assume a significant, or perhaps commanding,

role
in field computation
due to its greater speed and flexibility.
Ahnost certainly, general purpose scientific computers will

problem

and Linear

AND

[14] B.A. Carr&, The determination of the optimum accelerating facComputer J., vol. 4, pp. 73-7S,
tor for successive over-relaxation,

of quasi-TEM

waves.

greater

Matrices

THEORY

[29]

[30]

[31]
[32]

[33]

[34]
[35]
[36]

[37]
[38
-- 1

[39]

and A. Wexler, An accurate finite-difference

method for higher order waveguide modes, IEEE Trans. Microwaue Theory and Techniques, vol. MTT-16, pp. 1007-1017, December 1968.
C. T. Carson, The numerical solution of TEM mode transIEEE
Trans. Microwave
mission lines with curved boundaries,
Theory and Techniques (Correspondence), vol. MTT-15, pp. 269270, April 1967.
I. A. Cermak and P. Silvester, Solution
of 2-dimensional
field
problems by boundary relaxation, Proc. ZEE (Londpn), vol. 115,
pp. 1341-1348, September 1968.
F. Ceschino, J. Kuntzmann, and D. Boyanovich, Numerical SolaEnglewood Cliffs, N. J.: Prenticetion of Initial
Value Problems.
Hall, 1966.
Treatment
of Differential
Equations.
L. Collatz, The Numerical
Berlin: Springer, 1960.
J. B. Davies and C. A. Muilwyk, Numerical solution of uniform
hollow waveguides with boundaries of arbitrary shape; Proc.
IEE (London), vol. 113, pp. 277-284, February 1966.
J. W. Duncan, The accuracy of finite-difference
solutions of
IEEE
Trans. Microwave
Theory and TechLaplaces equation;
niques, vol. MTT-15, pp. 575582, October 1967.
Methods for
G. F. Forsythe and W. R. Wasow, Finite-Difference
Partial Differential
Equations.
New York: Wiley, 1960.
Solution
of Ordinary
and Partial
Differential
L. Fox, Numerical
Equations.
Oxford: Pergamon, 1962.
H. E. Green, The numerical solution of some important transIEEE
Trans.
Microwave
Theory
and
mission-line
problems,
Techniques, vol. MTT-13, pp. 676692, September 1965.
Numerical
Analvsis of El~irXic BoundD. Greensvan, Introductory
ary Value Problems.
Ne;
York: Harper a;d R;w, i965.
C. B. Moler. Finite difference methods for the eizenvalues of
Laplaces
operator,
Stanford
University
Comp&er
Science
Dept., Stanford, Calif., Rept. CS32, 1965.
H. Motz, The treatment of singularities of partial differential
equations by relaxation methods; @art. Appl. Math., vol. 4, pp.

371377, 1946.
[40] C. A. Muilwyk and J. B. Davies, The numerical solution of rectangular waveguide junctions and discontinuities of arbitrary

WEXLER : COMPUTATION

OF ELECTROMAGNETIC

cross section, IEEE Trans. Microwave

Theory
MTT-15, pp. 450-455, August 1967.
[41 ] R. D. Richtmyer and K. W. Morton, Difference

and Techniques,

VO1.

tial- Value Problems.

New

Methods

for

Ini-

York: Interscience, 1967.

[42] J. A. Seeger, Solution

of Laplaces equation in a multidielectric

region, Proc. ZEEE (Letters) vol. 56, pp. 13931 394, August 1968.
[43 ] D. H. Sinnott, Applications
of the numerical solution to LaIEEE
Trans. Microwave
places equation in three dimensions,
Theory and Techniques (Correspondence),
vol. MTT-16, pp. 135136, February 1968.
Eqaations.
[44] G. D. Smith, Numerical Solutions of Partial Differential
London: Oxford, 1965.
[45 ] K. B. Whiting, A treatment for boundary singularities in finite
IEEE Trans. MicroditTerence solutions of Laplaces equation;
waue Theory and Techniques (Correspondence),
vol. MTT-1 6, pp.
889-891, October 1968.
[46] K. S. Yee, Numerical solution of initial boundary value problems
involving Maxwells equations in isotropic media, IEEE Trans.
Antennas and Propagation,
vol. AP-14, pp. 302-307, May 1966.

Finite Dl~erence

Solution

of Integral

Equations

[47] M. G. Andreasen, Scattering from cylinders with arbitrary surface impedance: Proc. IEEE, vol. 53, pp. 812-817, August 1965.
[48] T. G. Bryant and J. A. Weiss, Parameters of microstrip transmission lines and of coupled pairs of microstrip lines, IEEE
Trans. Microwave
Theory and Techniques, vol. MTT-16, pp. 10211027, December 1968.
[49] R. F. Barrington
[59, pp. 62-81] and [60].
[50] K. K. Mei, On the integral equations of thin wire antennas,
IEEE Trans. Antennas and Propagation,
vol. AP-I 3, pp. 376378,
May 1965.
[51 ] J. H. Richmond Digital computer solutions of the rigorous equations for scattering problems; Proc. IEEE, vol. 53, pp. 79tk804,
August 1965.
[52] P. Silvester, TEM wave properties of microstrip transmission
lines, Proc. ZEE (London), vol. 115, pp. 43-48, January 1968.
[53] R. L. Tanner and M. G. Andreasen, Numerical
solution of elecpp. 5361, September
tromagnetic problems+ IEEE Spectrum,

1967.
[54]

..-

FIELDS

P. C. Waterman, Matrix formulation

of electromagnetic
ing; Proc. IEEE, vol. 53, pp. 805812, August 1965.

scatter-

439

Matrix
methods for field problems: Proc. IEEE, vol. 55,
pp. 136-149, February 1967.
Methods
of
[61 ] L. V. Kantorovich
and V. I. Krylov, Approximate
Higher Analysis.
Groningen, Netherlands: Interscience, 1958.
Variational
Methods
in Mathematical
Physics.
[62] S. G. Mikhlin,
New York: Macmillan,
1964.
Methods for
[63] S. G. Mikhlin
and K. L. Smolitskiy, Approximate
Solution of Differential
and Integral Equations.
New York: Elsevier, 1967.
Physics,
[64] P. M. Morse and H. Feshbach, Methods of Theoretical
Part II.
New York: McGraw-Hill,
1953, pp. 1106-1172.
[65 ] P. Silvester, Finite-element
solution of homogeneous waveguide
problemsj
1968 URSI Symp. on Electromagnetic
Waves, paper
115 (to appear in Alta Frequenza).
[66]
A general high-order
finite-element
waveguide analysis
IEEE
Trans. Microwave
Theory and Techniques,
vol.
proglam,
MTT-17,
pp. 204-210, April 1969.
[67] D. T. Thomas, Functional
approximations
for solving boundary
IEEE
Trans. Microwave
Theory
value problems by computer:
and Techniques, this issue, pp. 447454.
[68] E. Yamashita and R. Mitrra, Variational
method for the analysis
Theory and Techof microstrip lines, IEEE Trans. Microwave
niques, vol. MTT-16, pp. 251256, April 1968.
[69] O. C. Zienkiewicz, A. K. Bahrani, and P. L. Arlett, Numerical
solution of 3-dimensional field problems: Proc. ZEE (London),
vol. 115, pp. 367369, February 1968.
[70] O. C. Zienkiewicz and Y. K. Cheung, The Finite Element Method
in Structural
and Continuum
Mechanics.
New York:
McGrawHill, 1967.
The
[71]
Finite elements in the solution of field problems,
Eng~eer,
pp. 507-510, September 24, 1965.
[60]

Hybrid

Computing

[72] J. R. Ashley, Iterative

integration of Laplaces equation within

Simulation,
pp. 60-69, August 1967.
boundaries;
[73] J. R. Ashley and T. E. Bullock, Hybrid computer integration of
partial differential equations by use of an assumed sum separation
of variables: 1968 Spring Joint Computer Conj, AFIPS Proc., vol.
33. Washington,
D.C.: Thompson, pp. 585-591.
New York:
[74] G. A. Bekey and W. J. Karplus, Hybrid Computation.
symmetric

Wiley, 1968.
[75] S. K. T. Hsu and R. M. Howe, Preliminary
hybrid

Variational

Methods

[55] S. Ahmed,

Finite

Spring

element

method

for waveguide

problems,

Electron. Letts., vol. 4, pp. 387-389, September 6, 1968.

[56] P. F. Arlett, A. K. Bahrani, and O. C. Zienkiewicz, Application

of finite elements to the solution of Helmholtzs equation, Proc.

ZEE (London), vol. 115, pp. 1762-1766, December 1968.
[57] R. M. Bulley, Computation
of approximate polynomial
solutions to the Helmholtz equation using the Rayleigh-Ritz method,
Ph.D. dissertation, University of Sheffield, England, July 1968.
[58] G. F. Forsythe and W. R. Wasow [34, pp. 159-175].
Field Computation
by Moment
Methods.
New
[59] R. F. Barrington,
York: Macmillan,
1968.

method
Joint

for

Computer

solving
Con$,

partial
AFIPS

investigation of a
equations. 1968
vol. 33. Washington,

differential
Proc.,

D. C.: Thompson, pp. 601-609.

[76] R. Tomovic and W. J. Karplus, High Speed Analog Computers.
New York: Wiley, 1962.
[77] R. Vichnevetsky, A new stable computing method for the serial
hybrid computer integration
of partial differential
equations;
1968 Spring Joint Compater Corf, AFIPS Proc., vol. 32. Washington, D. C.: Thompson, pp. 143-150.

Miscellaneous
[78] S. Ramo, J. R. Whinnery,
Communication

Electronics.

and T. VanDuzer, Fields and Waues in

New York: Wiley, 1965.