Sie sind auf Seite 1von 5

Forward by JLV

Error evaluation and its a posteriori prediction are important Notebook proposes an adaptive method to reduce the error of inte-
issues in all numerical simulations. Although we may have the gral-equation solutions for electrostatic simulations. It proposes a
wishful thinking that codes can be treated as a black box, the way to subdivide the elements for increasing the accuracy of the
reality is that erroneous results can be obtained due to geometrical solution.
or numerical modeling difficulties. Therefore, a posteriori error
predictions and algorithms for correcting these errors are quite I would like to mention that previous articles in this column
important. The issue of numerical error evaluation is customarily of the Magazine have addressed the error issue from a general
addressed in computational mechanics, and is a recognized topic point of view, and in the context of other numerical methods.
among the scientists in that field. Please refer to the October, 1994, April, 1996, and October, 1996,
issues of the Magazine.
In electromagnetics, considerations relating to numerical
error have only appeared sporadically, but will continue to become Finally, I want to thank the authors for their contribution. For
of greater concern as we rely more heavily on numerical simula- questions on the article, please contact Prof. B. M. Kolundiija
tions for designs. The article in this issue of the EM Programmers directly. His e-mail address is kol@kiklop.etf.bg.ac.yu.

citance with
on Exact Error Estimation
Kolundiija, Andrej M. Hofman,
i r and Antonije R. Djordjevi
~ a ~ ik: ~Petrovi,

Department of Electrical Engineering


University of Belgrade
PO BOX35-54
11120 Belgrade
Yugoslavia
Tql: +381-11-3218-353
Fax: +381-11-324-8681
E-mail: kol@kiklop.etf.bg.ac.yu

1. Introduction structures) is not yet possible. So, we must begin with some sim-
pler task, which we can manage to solve. For this purpose, electro-
hat is the exact error bound of our result? Can we obtain statics is a good choice.
it with arbitrary (desired) accuracy? We often pose
these questions afier getting some new numerical result. Obvi- So, as the answer to the first question, we will first derive a
ously, it would be extremely useful to estimate the error of the new exact error estimation for the capacitance of two conducting
result exactly, thus eliminating the need for its further (expenmen- bodies. We will next incorporate this error criterion into an adap-
tal or numerical) validation. However, to the authors best knowl- tive method for the calculation of the characteristic impedance of a
edge, the exact estimation of an error for complicated electromag- transmission line. Finally, we will show some numerical results for
netic problems (e.g., antenna design, or scattering from complex a transmission line consisting of two square conductors.

80 IEEE Antennas and Propagation Magazine, Vol. 40, No. 6, December 1998
2. Exact error estimation for the capacitance

Let us start with the reciprocity theorem in electrostatics, and


consider region v, bounded by n inner closed surfaces, and one,
( n + l)st, outer closed surface, filled with a linear and isotropic
dielectric of permittivity E (Figure 1). Consider next two inde-
pendent charge distributions in region v ( poand p ) , resulting in
two independent potential distributions ( Voand v),and two inde-
pendent electric-displacement vector distributions (Do and D). The
reciprocity theorem [ 11 claims that

In particular, when the outer surface, Sn+l,tends to infinity, the


corresponding ( n + 1)st term on the left-hand side of Equation (1)
tends to zero. Let the field and the charge distribution marked by
Figure 1. Region v, bounded by n + 1 surfaces.
subscript 0 represent the exact solution of an analysis problem, and
let the other field and charge distribution represent an approximate
solution of the same problem.
extremely bad), the capacitance of the capacitor can be estimated to
be
Let us apply the theorem to a capacitor with conductors
bounded by surfaces Sl and S2. (In this case, po = 0 , and the two
conductors have equal charges of opposite sign, Q, = Qol = -eo2.) CmlnIc, 5 cmax
, c, = A, (5)
If this problem is analyzed by the Method of Moments (MOM) [2], vo, - v02
the unknown quantity is the charge distribution over the surfaces of
the conductor, so it is a priori supposed that p = 0 . Also, in the where
MOM solution, it is customary to prescribe the potential difference
between the conductors, and to enforce the condition that the
approximate charges of the conductors satisfy Q = Q, = -e2[3].
Since Vo, = constant, the integral of the second term on the left
side of Equation (1) can be simplified to This is a very important result, because it gives us exact
lower and upper bounds for the capacitance. Moreover, these
bounds are easily calculated, starting from the approximate surface
charge distributions psl and ps2

Having in mind that Doi dSi = -pso,dS,, Equation (1) can be 3. Relative error for the characteristic impedance
written in the form
Application of this result to a uniform transmission line with
a homogeneous dielectric (e.g., the one sketched in Figure 4) is
straightforward. The characteristic impedance of a transmission
line is calculated from its capacitance per unit length, resulting in
In the above equation, psol and pso2denote exact charge distri- the following estimation:
butions on the two conductors. (Note that approximate potentials 1
V, and V2 are not constant on the conductor surfaces.) Zmin5 z, 5 Zmax, zo= cc;, ,
~ (7)

Let us suppose that the total charge Qo is positive. In that where


case, the surface charge density psol is always positive, while the
surface charge density pSo2is always negative, so the following
two inequalities can be obtained:

and c is the velocity of light in the dielectric.

We now have means to calculate the exact lower and upper


bounds for the characteristic impedance. Rut this is a somewhat
Subscripts min and max denote minimum and maximum val- unconventional representation of an approximate result. It is much
ues of an approximate potential on the surface of the corresponding more convenient to represent the result in the form of two values: a
conductor. single value for the characteristic impedance, Z, and an exact value
of maximum relative error, 6,,, .
Combining inequalities (4a) and (4b), and assuming that
hmin> V2, (otherwise, the approximate solution would be For this purpose, let us first define the relative error as

IEEE Xntennas and Propagation Magazine, Vol. 40, No. 6, December 1998 81
(9)
I L I

The error is maximal if Zo = Zmin or Zo = Zmax. It is easy to show


that this maximum error is minimized if Z is equal to the harmonic
mean of Zminand Zmax: I I
A C B
Figure 2. The non-uniform division of a segment.

The maximum error is now


procedure Adaptive-Method begin
zmax - Zmin Input 6desired
6max =
z m a x + Zmin Perform initial segmentation
repeat
By substituting expressions (8) into Equation ( i l), we get an exact Solve the problem
estimate of the maximum relative error for the characteristic Find 4 m i n > K m a x 3 VZmm Vzmax
9

impedance: for each segment do


if Error criterion (16) is not satisfied then
Divide the segment into two (according to (17))
until Error criterion is satisfied for all the segments
output z
However, in addition to calculating this error, we also want to keep end
its value below a pre-specified desired error,
Figure 3. A pseudo-Pascal algorithm for the proposed adaptive
method for benchmark calculation of the characteristic imped-
ance.
What if our solution does not satisfy this condition?
and to require that each of the four terms on the left side of Equa-
4. An adaptive method for benchmark calculations l
tion (14) be less than -6deslred. In other words, for each segment
2
The derived error criterion (12) would be of little practical
on both conductors ( i = 1,2), we require that the criterion
importance, unless we use it as a foundation for building up a
method to calculate the characteristic impedance with arbitrary
(desired) accuracy. To do so, we first need a method for solving
our transmission-line problem. We adopted a Galerkin's variant of
the Method of Moments and subdomain approximation of the
charge distribution, assuming that the cross section of the conduc- be satisfied. If all the segments satisfy this criterion, the relative
tors consists of straight segments [ 3 ] .These segments may be fur- error for the characteristic impedance is less than bdesired. If not,
ther divided into smaller segments. On each segment, the charge those segments not satisfying the criterion should be further
distribution is approximated by a constant function. An important divided, and the calculation repeated.
property of the adopted method is that its accuracy can be easily
increased by increasing the number of segments, e.g., by dividing The second problem, that of how to subdivide these seg-
each or some segments into a number of smaller ones. ments, can be considered in many ways. The simplest is to uni-
formly bisect the segment, but a non-uniform adaptive segmenta-
However, two problems need to be solved. The first one is tion should lead to a smaller number of segments (and a smaller
which segments should be further divided to increase the accuracy? number of unknowns) for the desired accuracy. We used the proce-
The second one is how should they be divided? dure of non-uniformly dividing the segments into two new seg-
ments (Figure 2), using the formula
For the first problem, we need some local error criterion
(applicable to segments), because the global criterion (given by
Equations (12) and (13)) cannot be directly used. In addition, this
local criterion must be equally or more strict than the global one.
One way to derive such a criterion is to rewrite: Equations (12) and
(13) as where L is the length of an old segment, L, is the length of the first
new segment, Voi is the exact value of the potential on the con-
ductor, V, and V, are approximate values of the potential at the
segment ends, and t is a segmentation factor ( t > 1). This factor is
used to control the minimum and maximum ratio of L, and L:

-1_ - 1 4 1
<--I-+--.
1
2 2t L 2 2t

82 IEEE Antennas and Propagation Magazine, Vol. 40,No. 6,December 1998


For t + co , a uniform bisection is obtained, and for t = 1, we get
the most non-uniform segmentation, which could even lead to
segments of zero length. An optimal value for t (enabling the
most efficient calculation) was found by numerical experiments,
and it is between 1.5 and 2 .

our final note about the method concems finding the mini-
I
I

I
mum~and maximum values of a potential on the conductors sur- 0.85
I

faces; The approximate potential is an explicit function of coordi- -2 -1 1 2


//a
nates iand surface charges. However, finding-the extremes of this
function analytically would be a cumbersome task. Instead, we cal-
culatdd the potential values at a number of equidistant points along
each segment, including segment endpoints, and among them
found minimum and maximum values. By numerical experiments
we concluded that five points along each segment is more than
enough, although checking the segment endpoints only leads
almost to the same results. I
0.85
-2 -1 0 1 2
//a
F
1 he proposed adaptive method for benchmark calculation of
the cliaracteristic impedance is shown in Figure 3, in the form of a
1.05, I I I

pseuc(o-Pasca1algorithm.

a 1 . m

Figure 4. A sketch of a two-square-wire line.


0.9995

,
-1 0 1
//a
Figure 6. The potential distribution along the circumference of
one conductor, occurring after a.) each of the first three itera-
tion steps, and b.) at the end of the adaptive procedure.

5. Numerical results

Let us consider a transmission line consisting of a pair of


square conductors. The side length of each square wire is a, and the
distance between wire axes is w, as shown in Figure 4. In the
examples that follow, a particular line is analyzed
( w / a = 1.368469, Vo, = -V& = 1 V). The characteristic impedance
of this line is estimated to be z = 75 ohms, with a relative error less
than 6 = The initial number of segments per conductor side
was set to nb = 3 . Having in mind the symmetry of the problem,
the equations are posed only for segments belonging to one con-
ductor.

As a first example, let us consider charge distribution along


e 5. The approximate charge distribution at the cross-sec- the circumference of the conductors, shown in Figure 5 . The results
f a two-square-wire line. are obtained by using uniform bisection ( t -+co ), and requiring

/E4 Antennas and Propagation Magazine, Vol. 40, No. 6,December 1998 83
that the maximum relative error be less than &desired = 0.01. It is
well known that near the square-conductorcomers, the charge den-
sity tends to infinity. It is seen that, owing to our adaptive seg- -a- t = inf
-0- t = 25
mentation, charge distribution near metallic wedges is modeled -A- t E2.0
very accurately.

Let us now inspect how much the approximate potential dis-


tribution deviates from the unit voltage along the circumference of
the square conductor. The analysis is performed by using non-uni-
form bisection ( t = 2), and requiring that the maximum relative
error be less than 8deslred = 0.001. The results for the first three
iteration steps are shown in Figure 6a, and the final results are
depicted in Figure 6b. We note that at the beginning, i.e., after each
of the first two iterations, many of the existing segments do not
satisfy the error criterion of Equation (16). Also, as expected, the
worst errors are near the vertices of the square conductor. Hence,
after the third iteration, the bisection is performed mostly on seg- Figure 8. The number of iterations versus the maximum rela-
ments close to the vertices. After the convergence, however, the tive error of the characteristic impedance.
errors are more randomly distributed along the squares perimeter.

Our primary objective is, of course, to calculate the imped- With respect to the number of iterations performed by the
ance of the line within the desired precision, as defined by Equa- adaptive procedure, the advantage of using non-uniform bisection
tion (1l). Further, we want to accomplish the solution with good is apparent. It is seen that the smallest number of iterations is
efficiency. Efficiency here means two things: We want to obtain obtained by using a segmentation factor of t = 1.5. A further
the solution by using the smallest possible number of unknowns, decrease o f t leads to an increase in the number of iterations, as
and the smallest possible number of iterations. To get a look at the well (not shown in the figure). When compared with the uniform
error behavior as a function of the number of unknowns and itera- bisection ( t + c1) ), the number of iterations for optimal non-uni-
tions, we refer to Figures 7 and 8. Results are shown for four dif- form bisection is practically halved. In particular, let us consider
ferent segmentation factors: t + co , t = 2.5, t = 2 , and t = 1.5.
the case 6 ,, For this case, the number of unknowns
Conceming the number of unknowns, it is seen that to required for analysis is about N = 1000. Theoretically,the smallest
achieve the desired accuracy, about the same number of unknowns number of iterations, N,,,, = 8 , is obtained when all segments are
are needed, independent oft. Note also that by decreasing the pre- bisected after each iteration. In the case of t = 1.5, the number of
scribed maximum relative error ten times, the required number of iterations is relatively close to this theoretical minimum. The num-
unknowns to achieve the prescribed error increases about three ber of iterations can be further decreased only if the segments are
times. That is, the number of unknowns needed for the desired subdivided into more than two sub-segments.
accuracy can not be further decreased by using some other seg-
mentation technique, unless some more sophisticated basis func- 6. Conclusion
tions are employed.
This paper presented an adaptive method for a two-dimen-
sional electrostatic analysis of air-filled two-conductor lines. The
method is based on exact error estimates, and thus enables the
evaluation of capacitance within prescribed accuracy limits. For
the example of a transmission line consisting of square conductors
considered, the characteristic impedance was estimated to be
Z = 75ohms with a predefined relative error of less than
8,,, = The method can be generalized to two-dimensional
and three-dimensional electrostatic analysis of arbitrary systems of
conducting bodies in free space.

7. References

1. B. D. Popovic, Electromagnetic Field Theorems, IEE Pro-


ceedings, 128 Pt. A, 1, 1981.

2. R. F. Harrington, Field Computation by Moment Methods, New


York, Macmillan, 1968.
10 I . , . I , I . , ~ . I , , .,,,, I I , , , , . . , , , , , r6m
106 104 103 10-2
3 . A R. Djordjevi, M. B. Ba dar, T. K. Sarkar, R. F Harrmgton,
Figure 7. The number of unknowns versus the maximum rela- LINPAR for Windows, Software and Usevs Manual, Nonvood,
tive error of the characteristic impedance. Artech House Publishers, 1996. AP

84 IEEE Antennas and Propagation Magazine, Vol. 40, No. 6, December 1998

Das könnte Ihnen auch gefallen