IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 39. NO. 4.

AUGUST 1990

615

Global Iterative Solution of Dielectric Spectroscopv Equations

SYLVIE GELINAS, MEMBER, IEEE, V . N . TRAN, MEMBER, IEEE,
AND

REMI VAILLANCOURT

Abstract-This paper presents a global method to solve the permittivity equations for open- and short-circuited coaxial lines of general length for broad-band measurements by iterating the recurrence schemes i,,+ = c cot z;, and z,, + = c tan z,, respectively, and their inverses. The global iteration theory of Fatou and Julia, coupled with linear extrapolation and interpolation and Steffensens acceleration procedure, supplies starting values and guarantees convergence even near double roots. When CRz 2 0 for open, and 6 i z 5 0 for short circuit terminations Newtons method, with appropriate starting values, converges to the desired roots. Thus a combination of the three iterative schemes results in an almost global method of solution. Numerical results are quoted.

I. INTRODUCTION HE COAXIAL transmission line has been used as a sample cell in dielectric measurements for many years. Reflection measurements with the sample terminating in an open circuit [I] and [2] lead to solving the permittivity equation

tions. The multiple-valued inverse function, z = associated with (l), was used by Scott and Smith [SI to shed light on the structure of its Riemann surface and its branch points. However, in the application, the use of the Lagrange series is too time-consuming to be practical and Scott [6] used Mullers method to solve (1). For the shortcircuited line, the fundamental regions of ( 2 ) are drawn in [4, p. 4051 in polar coordinates. The purpose of this presentation is to establish a numerical strategy to find all solutions of (1) and (2). Rough graphs of the images of each of the four quadrants of the c-plane in the z-plane by the multiple-valued functions c { z ; z tan z = c}, and c + { z ; z cot z = c}, are given. These images establish an order on the roots of ( 1 ) and (2), respectively. According to the Fatou-Julia theory for the iteration of meromorphic functions [7], each root of (1) is reached by iterating F ( z ) = c cot z , i.e.,
-+

z tan z

(1)

z,+

c cot z,

(3)

for the unknown normalized propagation constant, j Z , for a set of experimentally obtained values of the complex normalized admittance, -2. Similarly, a short-circuit termination, [3, p. 5051 and [4, p. 15221 leads to the permittivity equation The length of the sample, its position, and the impedance terminating the line, were chosen in the past to provide the best accuracy at each frequency being used [ 11. However, over the past twenty-five years, commercially available automated network analyzers can measure impedance over an increasing range of frequencies. For optimal use of this instrumentation, it is not practical to adjust the length of the sample or the termination to obtain the best performance at each frequency. Thus (1) and ( 2 ) are to be solved over a wide range of values of c, for some of which the roots z may come close to double roots of these equaManuscript received September 26, 1989; revised February 7, 1990. This work was supported in part by the Natural Sciences and Engineering Research Council of Canada under Grant A 769 I . S . Gelinas was with the University of Ottawa, Ottawa, Ont. Canada K I N 6N5. She is now with the Department of Applied Mathematics at the Ecole polytechnique de Montreal, Montreal, P . Q . , Canada H3C 3A7. V. N. Tran is with the Division of Electrical Engineering, Deakin University, Geelong, Victoria 3217, Australia. R. Vaillancourt is with the Departments of Mathematics and Computer Science, the University of Ottawa, Ottawa, Ont., Canada K I N 6NS. IEEE Log Number 9036025.

or a properly chosen branch of its multiple-valued inverse G(z) = F-(z),

z,+

= arccot

(b z,)

(4)

The main idea used here is that F has either zero or exactly two attractive fixed points and the basins of attraction of all but a few attractive fixed points of appropriate determinations of G, for any given value of c , are vast. A second idea is to apply Newtons method to ( l ) , rewritten in the form z sin z - c cos z = 0 (5) with appropriate starting values to effectively find roots that are known to be located in definite bounded convex regions of the plane. Similar iteration techniques apply to (2). The advantage of the proposed set of methods [8] over a single local method, like Mullers or Newtons methods, is that we almost have a global algorithm to find any prescribed root of (1) and ( 2 ) . Equation (1) will be treated in detail. Similar results for (2) are summarized in Section X. In Section IX, a numerical strategy is given and numerical results are quoted. 11. FUNDAMENTAL REGIONS Equation (1) has infinitely many simple roots for each value of c. It has also infinitely many double roots, but

0018-9456/90/0800-0615$01 O 1990 IEEE .OO

I
616
IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 39. NO. 4, AUGUST 1990

212

'IL

3x12

2X

SX12

3%

Rez

Fig. 1 . Images I,, and I l k , of first and second quadrants, respectively, of { z; z tan z = c } . Double roots c-plane by multiple-valued function c p , , p 2 , . . . , of equation z tan z = c at critical values c , , c2, . . . .
+

In fig. 3 of [5] the branch cuts were taken as the verR tical lines C c = CRc, joining c, to E, through 00 and produced fundamental regions that extend slightly to the right of the lines CRz = ( 2 m 1 ) a/2. It would be nice, although not essential for our purpose, to have our regions Z, m = 1, 2 , * , lying completely inside vertical strips Z , of width a bounded by the lines C z = ( 2 m - 1 ) a / 2 and R CRz = (2m 1 ) ~ / 2 To this effect, one considers the . values taken by c = z tan z along such lines, namely, with x ( m ) = (2m 1 ) a / 2 and z = x + j y , i.e.,

DOUBLE ROOTSz
m

p,n OF Z
Pm

TABLE I tan Z = C AT CRITICAL VALUES C

c, 0 -1.651+02.060j -2.058 + 05.335j -2.278 + 08.523j -2.431+ 11.6891' -2.548 + 14.846j -2.643 + 17.998 j

Y + j -x ( m > c ( m ) = -tanh y tanh y '

= C,,

0 0 1 2.106+ 1.125j 2 5.356 + 1.552j 3 8.537 + 1776 j 4 11.699+ 1.9293' 5 14.854 + 2.047j 6 18.005 + 2.142 j

By choosing a branch cut that lies between the curves c ( m - 1 ) and c ( m ) from some point on, the region Z, Z will be inside the desired strips. 111. FATOU-JULIA THEORY

no roots of multiplicity higher than two. The double roots, which are a hindrance for any numerical methods, are the solutions to the system f(z):=ztanz-c=O
~ ( x := tan z )

(6)
(7)

A few results of the Fatou-Julia global iteration theory, as extended to the iteration of meromorphic functions [7] and their inverses, will now be listed. A general presentation of the iteration theory for rational functions, notation, and references are found in a recent survey by Blanchard [lo]. Let c p : C , be a transcendental meromorphic function and consider the iteration

+ z sec2 ( z ) = 0.

The last equation, which is independent of c , can be simplified to

z
and, even more, to

+ sin z cos z = 0
22

(8)

+ sin 22

0.

(9)

The roots of (9), which will be referred to as p, and -pm in tables and figures, are double roots of (1) at the critical values of c , i.e., at c, = p, tanp,. A global procedure for the solution of the transcendental equations sin z = cz, which generalizes (9), and cos z = cz, is given in [9]. The pairs { p,, c, } are obtained numerically, [7], and [6, p. 141 and the first seven pairs are listed in Table I. It is enough to consider the first quadrant, I,, of the z-plane and the upper-half plane, Z U ZZ, of,the c-plane since ( - z ) tan ( - z ) = c and 1 tan 1 = E. To obtain a rough graph of the images of Z and ZZ into I, by the multiple-valued function
c + {z;ztanz-c=~} (10) one draws the images of the positive imaginary c-axis and, for each critical value c, the image of a properly chosen branch cut joining c, and C through W . The images of Z , and ZZ, denoted, respectively, by ZO,ZZo, I , , ZZI, etc., according to an obvious ordering, are shown in Fig. I . For , the double root, p,, lies on the each m = 1 , 2, * boundary between the unbounded region ZZo and the Z , bounded region Z.

Z n + l = cp(z,), n = 0, 1 , 2, * . (12) AJixedpoint s of cp, s = cp(s), is attractive, repulsive, or indiferent, as the absolute value of its multiplier, c p r ( s ) , satisfies, resp., Icp'(s)l < 1 , > 1 , or = 1 . The zeros of cp'(z) are the algebraic critical points of cp and the exceptional values of cp are the transcendental critical points of cp. Let

cpO(z) cp"(z) = (P[cp"-l(z)], denote the nth iterate of z by cp. The Julia set of cp, is the set of nonnormality of cp:
OD

(13)

3(cp),

3(p) : = [ z; { cp"(z)},,= is not a normal family].

(14) The set ofnormality of cp, X ( c p ) , is defined in a similar way. The Riidstrom set [ l l ] of cp, @ ( C O ) ,is the set of predecessors of the essential singularity of cp:
03 ( c p ) : =

{ z; cpn (z) is not defined for some n E N } .

(15)

Then

Wcp) = c\(d(cp) U W)). (16) For the meromorphic functions considered here, 3 and CR are nonempty sets without isolated points. Moreover, 3 and 32 are completely invariant with respect to cp, i.e., invariant under cp and cp-l.

GELINAS et al.: SOLUTION OF DIELECTRIC SPECTROSCOPY EQUATIONS

617

A k-cycle of cp is a set of k distinct points, so, sI, , sk - I , satisfying the relations s1 = cp( so), s2 = * s k - l = c p k - l ( s O ) , SO = cpk(sO)* (P2(sn), The multiplier of a k-cycle,

...

( P k ) ( s m ) = v(sk-1)

* *

c~(sl)c~(so) (17)

Fig. 2. Ovals of Cassini { z; 1 z + f I I z - f I = < 1; center, I f 1 = 1 ; right,

If1

1fI}

If\

> 1.

with foci

if. Left,

, , is seen to be the same at every point s m = 0 , 1, k - 1, of the cycle. A k-cycle is attractive, repulsive, or indifferent, respectively, as 1 ( cpk)(s,) 1 < 1, > 1, or = 1. AJixed point is a 1-cycle. Any element s of a k-cycle , is a fixed point of cpk. The attractive cycles of cp are and conversely the repulsive cycles of cp-l cp(z) ( c p - ) ( c p ( z ) ) = 1. The immediate basin of attraction of an attractive fixed point s is the largest connected open set Q such that z, = cpn(z0) s, as n 00, for all zo E Q . The immediate basin of attraction of a k-cycle is the union of the immediate basins of attraction of the elements s considered as fixed points of cpk. Attractive fixed , points and attractive cycles are in 32 ( c p ) . The following result, cf. [12], which is derived in [7], will be needed.
+
+

ImZ
A

-812

xt2

sin z

cos z

Fig. 3 . Level curves I sin z 1 = k , and I cos z 1 = k , with origin at 0 and r / 2 O j , respectively.

+ Oj,

spectively , sin 2x

+j

sinh 2y,

-sin 2x

+j

sinh 2y.

(21)

Theorem I : The immediate basin of attraction of every attractive fixed point or cycle of cp = c cot z, and cp = c tan z , contains at least one critical point of cp. IV. OVALS CASSINI OF AND LEVEL CURVES The regions of convergence of the iteration functions (3) and (4) used to solve (1) involve two families of plane curves, namely the ovals of Cassini and the level curves of sin z. Similarly, in solving (2) the same ovals and the level curves of cos z will play an important role. These curves are described in the present section. f The oval o Cassini
O(C)

V. THE ITERATION FUNCTION ( z ) F

The iteration function

F(z)

= c cot 2,

c # 0

(22)

of the numerical scheme (3) has no algebraic critical points since

F ( z ) :=

--(I

CSC~ # z

0,

c # 0.

(23)

{ z ; 12 +jcI

12

-jcI

1.1)

(18)

It has, however, two transcendental critical points, namely, +jc. Hence, according to Section 111, F has either zero or exactly two attractive fixed points, +s(c), and these are in the forward orbits of * j c . Clearly +s (c) must belong to the region ofattractivify a(c)of F:

is the locus of points in the z-plane, the product of whose distances of the foci +ic is constant [14]. As shown in Fig. 2, (3 ( c ) is a lemniscate of Bernoulli when 1 c I = 1; it is made of two disjoint simple closed curves or loops when I c I > I , and one simple closed curve or loop when 1 c I < 1. The farthest and closest points of (3 ( c ) satisfy, respectively,

a(c):= { z ; Isinzl >

m} [ z ;
=

J ~ ( z ) l< 1 )

(24) shown in Fig. 3, and also, as can be easily seen, to the region
@(-(I>

:= F ( B ( c ) ) =

{ z ; Iz

+jc(

( z -jcl <

1.1)

(19) The normal direction to O ( c ) at the point z is the gradientof Iz + j c I Iz - j c l :

The level curves Isin zI = k, and /cos z ( = k, are shown in Fig. 3, cf. [15]; their normal directions are, re-

(25) which is bounded by the oval (18). It follows that the two attractive fixed points, +s(c), of F, if any, lie in a ( c ) fl @ ( c ) # 0 and @ ( c ) contains no other fixed points of F, i.e., no other roots of (1). The double roots + p , of (1) are indifferent fixed points since their multiplier F ( p,) = 1; hence they belong to the Julia set of F. Theoretically, they can be reached by (3) from zo = +ic,; but, in practice, with 15 decimal digit arithmetic, only three significant digits can be obtained without recourse to Steffensens acceleration [ 161.

I
618
IEEE TRANSACTIONS ON INSTRUMENTATION A N D MEASUREMENT, VOL. 39, NO. 4, AUGUST 1990

VI. THE INVERSE FUNCTION z ) G( The roots of (1) which do not lie in the closure of 63 ( c ) are attractive fixed points of certain determinations of the inverse, G , of F:

G ( z ) = -log + ka, k E Z. 2j To obtain proper determinations of G , set

t = -z + j c z -jc

(-)

(26)

+ Re
z
Fig. 4. Oval o ( c , ) and level curve I sin z I root p I .
= cI

Firstly, if -a < arg t 5 a, the branch cut of log t is taken along the negative real axis in the t-plane and correspondingly from- -jc to j c through the- origin in the z-plane. In this case, G ( z ) will be denoted by G o ( z ;k ) ; thus G o ( z ;k ) is contained in the vertical strip

(I/*

are tangent at double

Bo(k) :=

i z+
Z; --

ka

< CRZ

I-

+ ka

Secondly, if 0 < arg t 5 27r, the branch cut of log t is taken along the positive real axis in the t-plane and from - j c t o j c through 00 in the z-plane. In this case G ( z ) will be denoted by G , ( z ; k ) ; thus G , ( z ; k ) is contained in the vertical strip

(28)

B,(k) := { I ; ka Go(Zn;k ) ,

< CRz

I( k

+1)~).

(29)
(30)

The iterates, respectively,

Zn+l

tractive cycle, depending upon the starting value zo. At most, two such situations can occur for any given value of c. It is easy to show by means of RouchCs theorem [7] that the root do)E IIo lies in the square bounded by the real and imaginary axes and the lines C z = (2m + 1 ) a / 2 R and9z = (2m l ) a / 2 , i f m 2 0.35 I c I - 0.5. To find a root of (1) lying near the boundary of a ( c ) , Steffensens method can be used to accelerate a slowly but already converging sequence; Steffensens acceleration will turn any root of (1) into an attractive fixed point of F and G and thus may divert convergence to another root.

Z,+l

Gce(zn; k )

VII. BIFURCATION NEARDOUBLE ROOTS When c = c E II, m = 1 , 2, , , the oval (18) goes through the double roots +pm of (1) and the two loops are tangent to the boundary of ( c ), i.e., to the level curves , I sin z I = as can be computed by (20) and (21), and as shown for the case c = c, in Fig. 4. The loop of the oval in the first quadrant lies inside the infinite region a (c, ) as can be seen from Table I. When the oval is moved slightly away from II, by moving the focus, -jc, in a direction parallel to the perpendicular to the boundary of a ( c ) through p,, the double root bifurcates along this perpendicular; one root enters the loop in ZZo while the other root goes inside II,. On the other hand, along a direcwhen the oval is moved slightly inside U,,, tion parallel to the same perpendicular, the bifurcation occurs in a direction orthogonal to this perpendicular and both roots go outside the loop, one in U,,, and the other in Z,, both outside a ( c), thus getting out of the reach of F. Z,

converge to attractive fixed points of G provided the segments [z,, zn + ] lie outside 63 (c), V n > no, and hence Indeed, Go, and G,, are contracting, i.e., outside a ( c ) . for some d < 1 ,
IZn+2 - Z n + 1 1

1S

m,

Z+ I

a 2

~ r ( z ) d z Idlzn.1 - z n / l

(31) the integration being along the segment [ z,, Z , + 1, since (GI Id < 1 outside a ( c ) . Therefore, i f s (c) is the unique root of (1) in Bo ( k ) n Zi and Bo ( k ) lies to the right of r3 (c), then the iteration

Zn+, =

Go(zn;k);

zo~Bo(k)\a(c)

(32)

remains in Bo( k ) and converges to s ( c). Similarly, if s ( c ) is the unique root in B, ( k ) f l Z and , B, ( k ) lies to the left of the right-hand loop of 0 (c), the iteration
Zn+I =

Gm(zn;k);

zoEBm(k)\Q(c)

(33)

VIII. NEWTONS METHOD WHENC c 2 0 R Newtons iteration for (5) is given by

Zn+l

remains in B, ( k ) and converges to s ( c). The presence of two roots, namely do)E ZIo and d k )E Il,, k > 0, in the same strip B o ( k ) or B , ( k ) , when c E ZZ, and/or the nonempty intersection of the ovals of Cassini (18) with the strip containing one or two roots restrict the basins of attraction of these roots. In such cases the iteration may converge to a root or may end up in an at-

= Zn -

zn sin Z ,
(1

- c

cos zn

+ c) sin zn + zn cos z, .

(34)

Since each of the regions Io, I,, , is bounded and convex, the following starting values obtained in [17] by truncated continued fraction expansions of the solutions

GELlNAS et al.: SOLUTION OF DIELECTRIC SPECTROSCOPY EQUATIONS

619

of (1):

DOUBLE ROOTSZ
m

TABLE I1 qm OF Z Cot Z = C AT CRITICAL V A L U E S C =

9m

C,:

d m

(35)

(36)

0 1 3 749+ 1 3841 1 895 - 0 7191 3 2 6 950 + 1 616) 2 180 - 06 933) 3 10 119+ 1 8581 2 361 - 10 1071 4 1 211 + 1 9921 2 493 - 13 2681 3 5 1 430 + 2 097J 2 598 - 16 4221 6 6 1 519 + 2 1831 2 684 - 1 5731 9 9
1

provide fast convergence to the desired roots do) d k ) . and It was found that Newtons method applied to (5) rather few applications of Steffensens acceleration procedure to than ( l ) , [7] and [2] with the above initial values, is at improve the accuracy. times faster and somewhat better. X. THE EQUATION cot z = c z IX. A NUMERICAL STRATEGY The treatment of (2) is almost similar to that of (1). The In dielectric spectroscopy with an open-circuited coax- double roots qm of (2) are solutions of the equation z ial line, typical measurements start at a low frequency w sin z cos z = 0. The first few q, and corresponding cirtical which gives a measured normalized admittance, -C( U), values of ch are listed in Table 11. with c ( w ) E I , and a normalized propagation constant, A rough graph of the images of the third and fourth i Z ( w ) , withz(w)EZo. Sinceinpractice lop2 < I c ( w ) ( quadrants, ZZZ and ZV, of the c-plane into the first quad< lo2, Newtons method (34) for c E I , with Cohens rant, Z, of the z-plane by the multiple-valued function c , starting values ( 3 3 , converges to the root to an accuracy { z ; z cot z = c } is given in Fig. 5. For each m = 1 , of 10-I in double precision arithmetic in not more than 2, * . * , the double root, q,, lies on the boundary befive iterations. tween the unbounded region ZVo and the bounded region As the frequency increases, the solution z ( w ) of ( 1 ) ZZ . Z, describes a smooth curve that may pass near double roots, When c E ZZZ, Newtons method applied to (2) in the e.g., pI and p 2 , as shown in Fig. 1 . As z ( U ) moves into form z cos z - c sin z = 0, namely, ZZo, F, Go or G, is used to continue the solution of whether z, cos z, - c sin z, the curve lies in a(c) or not. Near a double root, the (37) G + I = Zn ( 1 - c) cos z, - Z, sin z, experimental error in the measurement of c ( w ) increases due to resonance [ 11. The interpolation technique developed in [5] and [6] can be used with F , Go, and G , to with the initial values, respectively, determine whether the solution curve goes above the double root in Fig. 1 and hence in (3 ( c ) , or below and hence l - c outside a (c). Then F or Go and G , are used accordingly, zo = eventually with Steffensens acceleration method. Also in [5] and [6] a linear extrapolation is used to obtain the next starting value, i.e., z ( ~ = )Zoldcd,,,/U&+~ ~ ~ Typically, the data of Table I would be stored in the computer, so that for any experimental value of c, with negative real part ,, which is close to a critical value c the program would converges quickly to the desired roots s()E ZZZ,,and d k )
+

Jcc

detect that the values of the iterates z, are getting close to a double root p , and take appropriate action. In simulating a numerical experiment for the propagation constant of ethanol taken from fig. 12(b) in [SI, the attractive fixed points, resp., sI = 5.5 1.6j of F , and s2 = 5.2 + l.5j of Go and G,, for c = -2.0 + 5 . 3 j near c2 = -2.1 + 5.3j, as shown in Table I, were found. Although s1 was closer to the linearly extrapolated value zo = 5.36 1.6Oj, linear interpolation between the previous value, 5.09 + 1.52j, and the next value, 5.44 + 1.76j, indicated that s2 was to be preferred to sI. In a second experiment for the propagation constant of l-butanol, also taken from fig. 12(a) in [ 5 ] , the fixed point s = 3.97 + 1.41j, for c = -0.940 4.15j, lies just below a(c) and just above 03 ( c ) . It was reached from the linearly extrapolated value zo = 4.11 1.44j with Go and G , after less than 2000 iterations followed by a

zzzk.

The iteration function F ( z ) = c tan z has two transcendental critical points, *jc. Its region of attractivity is a ( c ) = { z ; (cosz(> while @ ( c ) = F ( a ( c ) ) is the same as in (25). The inverse iteration function

m>,

has determination G, ( z ; k ) , or Go(z;k) when the branch cut of log t is taken along the negative or the positive real axis, respectively. Results similar to those established for (1) also hold for (2). By RouchCs theorem, one can show that the root do)E WO, a given value of c E ZV, lies in for the square bounded by the real and imaginary axes and the lines Rz = m7r and 9z = m7r provided m > 0.32 I c I.

620

IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 39, NO. 4, AUGUST 1990

XI2

3x12

2%

5x12

3%

Fig. 5 . Images I U k , and I V I , of the first and second second quadrants, respectively, of e-plane by multiple-valued function c { I z cot z = e } . ; Double roots 9 , . 9*. . . . , of equation z cot z = c at critical values e;. c;. . . .
-+
7 .

XI. CONCLUSION FURTHER AND DEVELOPMENTS It has been shown that simple global iteration methods can be used to solve the dielectric equations. Near double roots, these slower methods can be directed to converge to the desired root, thus avoiding the pitfalls of faster local methods like Newtons method. The proposed combination of methods performed well on published data. As a further development, one could investigate the performance of the method

which was suggested by one of the reviewers. REFERENCES

[ I ] H. E. Bussey, Dielectric measurements in a shielded open circuit coaxial line, IEEE Trans. Instrum. Meas., vol. IM-29, pp. 120124, June 1980.

[2] V. N. Tran, S. Stuchly, and M. Stuchly, Measurement of dielectric properties of biological or agricultural materials, in Proc. 19th Inr. Electronics Conv., Sydney, Australia: Inst. Radio Elect. Eng., pp. 595-597, Sept. 5-9, 1983. [3] M. Sucher and J . Fox, Eds., Handbook ofMicrowave Measurements. Brooklyn, NY: Polytechnic Press, 1963, vol. 11, 3rd ed., p. 505. [4] P. H. Morse and H. Feshbach, Methods of Theoretical Physics. New York: McGraw-Hill, 1953, pp. 405 and 1522. [SI W. R. Scott, Jr. and G. S. Smith, Error analysis for dielectric spectroscopy using shielded open-circuited coaxial lines of general length, IEEE Trans. Instrum. Meas., vol. IM-35, pp, 130-137, June 1986. W. R. Scott, Jr., Dielectric spectroscopy using shielded open-circuited coaxial lines and monopole antennas of general length, Ph.D. dissertation, Georgia Institute of Technology, Atlanta, GA, Oct. 1985. J . L. Howland and R. Vaillancourt, Attractive cycles in the iteration of meromorphic functions, Num. Marh., vol. 46, pp. 323-337. 1985. S . Gelinas and R. Vaillancourt, Application of Julia-Fatou theory in dielectric spectroscopy, in Computational Solution of Nonlinear f Sysrems o Equarions, E. L. Allgower and F. Georg, Eds., Lectures in Appl. Math., vol. 26, pp. 189-208, 1990. S . Gelinas and R. Vaillancourt, Recherche globale des zeros dune fonction entihe par la thiorie de Fatou-Baker, in Ann. Sei. Math. Quebec, 1990, to be published. P. Blanchard, Complex analytic dynamics on the Riemann sphere, Bull. Amer. Math. Soc. ( N . S . ) , vol. 11, pp. 85-141, 1984. H. RLdstrijm, On the iteration of analytic functions, Math. Scand., vol. 1, pp. 85-92, 1953. I . N . Baker, Limit functions and sets of non-normality in iteration theory, Ann. Acad. Sei. Fenn., Ser. A I, no. 467, 1970. L. Tanguay and R. Vaillancourt, Numerical solution of the dielectric equation for a coaxial line, IEEE Trans. Instrum. Meas.. vol. IM-33, pp. 88-90, June 1984. , [I41 W. H. Beyer, Ed., CRC Standard Mathematical Tables, Boca Raton, FL: CRC Press, 1984, 27th ed.. pp. 218-219. [15] E. Jahnke and F. Emde, Tables of Functions with Formulae and Curves. New York: Dover, 1945, Addenda p. 66. [I61 E. Isaacson and H. B. Keller, Analysis ofNumerical Methods. New York: Wiley, 1966, pp. 103-108. [I71 E. R. Cohen, Science Center, Rockwell International, Thousand Oaks. CA, private correspondence, Apr. 1985.