Computer methods

in applied
mechanics and
engineering
EISEVIER Comput. Methods Appl. Mech. Engrg. 140 (1997) 221-235

A new family of exponential infinite elements for the

analysis of lossless electromagnetic waveguides
Andrea Toselli
Courant Institute of Mathematical Sciences, New York University, 251 Mercer Street,
New York, NY 10012, USA

Received 29 February 1995; revised 1 February 19%

In this paper the Finite Element Method is employed for the analysis of electromagnetic lossless waveguides. The curl-curl equa-
tion, deduced from Maxwell’s equations, is solved for the electric field. We have studied the appearance of parasitic solutions and
formulated a sufficient condition for their elimination which puts a contraint on the Finite Element spaces for the approximation of
the field components. A new family of discontinuous exponential i&mite elements is proposed for the analysis of open waveguides:
they are employed to;gether with edge elements and a non-linear eigenvalue problem is deduced. We have presented numerical
results in order to prove the validity of such infinite elements versus the Virtual Boundary Technique.

1. Introduction

The propagation of electromagnetic power at distance without considerable loss and attenuation is
one of the main tasks of modern communication systems. In electromagnetic waveguides the power
confinement is achieved either by enclosing the structure in. one or more electric conductor surfaces
(closed waveguides, see [6]), or thanks to the fact that the index of refraction n of the outer part is
lower than the one of the inner part (open waveguides, see [7,26]). Fig. 1 shows an open strip waveguide
generally employed in Integrated Optics: the waveguide Section is considered infinite and the device
indefinite in the positive and negative z-direction (waveguide axis).
Given a frequecy w, one seeks solutions (modes) of Maxwell’s equations for sinusoidal fields ([6]), that
represent a field propagating in the z-direction:

E(x,y,z) = e(x,y) eAiPz = (er(x,y) + e,(x,y)a,) ePiPz. (I)

The electric field is denoted by E(x, y, z). The complex number p is called mode propagaton constant
and the vector e is projected along the waveguide axis (e,) and on the transverse plane (et). The same
dependence is assumed for the magnetic field H, the electric displacement D and the magnetic displace-
ment B [6]. Correspondingly, one obtains an eigenvalue problem where fl represents the eigenvalue and
the electromagnetic field is the corresponding eigenvector. A finite set of real eigenvalues corresponds
to propagation modes [6] of closed waveguides and guided modes [7] of open waveguides. The func-
tion pi(w) is called dispersion relation for the ith mode and for each i there exists a frequency (cut-off
frequency) below which the mode does not propagate.
It is well known that when approximating the vector fields of this problem in Finite Element (FE)
subspaces consisting of piecewise continuous polynomial functions, unfortunately spurious or parasitic
solutions appear [14,25]. Spurious solutions have been characterized by the property that they do not

00457825/97/$17.00 @ 1997 Elsevier Science S.A. All rights reserved

PIZ SOO4.5-7825(96)01010-9
222 A. Toselli/Comput. Methods Appl. Mech. Engrg. 140 (1997) 221-235

I *

i n2 I

:,*’
*’

I---__--___-____._______I

Fig. 1. Strip waveguide for Integrated Optics.

satisfy the divergence equations (see Section 3), which impose the solenoidality of the electric and mag-
netic displacements [6]. In order to face this problem, various formulations, which ‘force’ the divergence
equations, and different FE approximations have been proposed. They remove or allow to character-
ize parasitic solutions for the waveguide problem, without giving insight on the actual reasons of their
appearance [4,8,13,14,17].
Another problem to face arises from the fact that the Finite Element Method (FEM) is not directly
applicable to unbounded domains, like those for open waveguides [21]. Since for guided modes the
field is confined in the inner part of the waveguide and decays rapidly far from the core, one has to
model the field in the outer region in a simple and computationally efficient way, so that the solution is
accurate where it is needed. Among the methods proposed, we mention the Virtual Boundary Technique
[13,14], the Recursive Growth [5], the Boundary Element Method [15] and exponential continuous infinite
elements [10,24].
The problems represented by spurious solutions and unbounded domains in electromagnetic wave-
guide analysis are generally solved independently. In this work we propose an original analysis of parasitic
solutions which leads to the formulation of a sufficient condition for their elimination. A constraint is put
on the FE spaces employed for the approximation of the field components, as it is generally required for
mixed methods. We also propose a new family of vectorial infinite elements that satisfy such a condition
and allow to treat the field at infinity efficiently.
In Section 2 the differential problem is presented together with the variational and approximated ones.
Section 3 is devoted to the analysis of parasitic solutions while the FE spaces for the approximation of
closed and open waveguides are described in Sections 4 and 5, respectively. Finally, numerical results
relative to open waveguides are presented in Section 6.

2. Equations for the electric field

We will consider an equation for the electric field E, whose expression is given by (1). The other
electromagnetic field components may be calculated from Maxwell equations (see [26]).
Let R C Iw, x lF$ be the Section of the waveguide and r its boundary: 0 is assumed simply connected
and coinciding with the whole space Iw2 if the waveguide is open.
The characteristics of the medium which fills the waveguide are described by the dielectric permittivity
z and by the magnetic permeability E, which are in general complex tensors depending on the position,
the frequency o as well as the field intensity [6]. We will assume that z is constant and equal to the
vacuum permeability p. and that z is real, symmetric, positive-definite and independent of the intensity,
w and z, that is

z = &&(x,y) = &o
[
Z(X,Y)
o

&z(X,Y)
0
lxo[; ; j, (2)
 A. Toselli/Comput. Methods Appl. Mech. Engrg. 140 (1997) 221-235 223

where co is the vacuum permittivity. This is equivalent to assuming that the medium is linear, lossless
and stationary. If in addition it is isotropic, then F is a positive scalar function. With these assumptions
one can specify alternatively the index of refraction !E [7,26] given by
s?
n =Zr. (3)
Given a frequency w, from Maxwell’s equations one obtains the following equation for the electric field
(see [261)
Vx(vxE)-kz~r.E=O, (X,Y)E@ (4)

where k* = ~opow’~and V x A denotes the curl of the vector field A(x,y, z). The expression for E is
given by (1) and accordingly the nabla operator is given by

v=
(fp,+%ly >+$a,
8Y
=Vt -ipa,.

For a closed waveguide we consider perfect electric conductor boundary conditions [6], which amount
to require that the tangential component of the electric field vanishes on the boundary fi

e, x n Ir = 0, (6)
ez Ir = 0, (7)
where n is the normal to r. For an open waveguide we require that the tangential component be zero
at infinity: if r = ra, is the position vector of the generic point P and r its modulus we impose that

lim e, x a, ==0, (8)

1’00
lim e, = 0. (9)
r-CC
When projected on the xy-plane and along the z-axis, Eq. (4) gives rise to two equations (a vectorial
one and a scalar one) in two unknown functions (e, and e,), which represent the eigenvector (mode) of
the eigenvalue problem. The propagation constant p is the corresponding eigenvalue.
We recall that the electric field E is discontinuous in general and Eq. (4) is strictly valid in each
subdomain L!i C 0, where the dielectric permittivity is continuous. Solutions in each Q must than be
matched along the separation lines by requiring the continuity of the tangential component of E, that is
the continuity of (I, and the one of the tangential component of et (see [6]).
In order to obtain a weak formulation, we have to multiply the two equations given by (4) for two
test functions, integrate over the section 0 and apply Green’s formulas (see [9]). In the following, V will
be used instead of Vt, where there is no ambiguity.
After some calculations and with the substitutions

u = Pet, (10)
p = -ie,, (11)
A = p*, (12)
we obtain the

WEAK PROBLEM. Find A E C and {u, cp} E X x Y such that

U(U) u) = A (b(U) v) + c(u, cp)) vu E x, (13)

O=h (c(U,$)+d(cp,rCI)) v+ E Y, (14)

where
x = Ho(curl, 0) if 0 is bounded,
H(cur1, a) if R = R*, (15)
224 A. ToseWComput. Methods Appl. Mech. Engrg. 140 (1997) 221-235

y = HA (a) if 0 is bounded,
H’(a) if0 = Ihs2, (16)

and the bilinear forms are defined as

a(u,u)= [k2(z~.U,U)-_(vX*,VXU)]v{(u,u}EXxX, (17)

b(u,u) = (u,u) V{U,U> E X x X, (18)

c(u,cp) = (u,V(P) ~{u,(Pl EX x y, (19)

(20)

BY C,4 and II . II we denote the inner product and the norm in L2(L?) or L2(0)2. The Hilbert spaces
in (15, 16) are defined as (see [9,21])

H(curl,L?) = {u E L2(0)2 : V x u E L2(0)}, (21)

H’(0) = (+ E L2(0) : Vll, E L2(0)2}, (22)

HO(curl, 0) = {u E H(curl,.0) : u x n lr = 0}, (23)
H,‘(n) = {$ E H’(0) : + Ir = 0). (24)

We recall (see [9]) that for all u E H(rot, 0) it is possible to define a tangential trace over r, u x n(,-,
as an element of H-‘/2(r), and that for all $ E H’(0) it is possible to define a trace over r, $I,-, as
an element of H’j2(r>. See [16] for the definition of Sobolev spaces of trace functions defined on F.
The choice of X and Y (Eqs. (15, 16)) is thus consistent with the essential conditions over F given
by (6, 7). We also recall (see [11,9]) that if 0 is simply connected (bounded or unbounded) the kernels
of the curl operator defined on H(rot, 0) and Ha(rot, 0) are gradH’(a) = {V+ : JI E H’(0)) and
grad HA (0) = { Vlfi : t,b E Hi (0) ), respectively, and hence

grad Y c X. (25)

The Weak Problem considered is a mixed eigenvalue problem (see [1,18]) in which the bilinear forms
a(-, .) and d(., .) depend on k. It is not a standard one because of the presence of d(., .). Similar problems
have been studied in [1,9], but the assumptions on the bilinear forms involved are not satisfied in our
case.
It is an easy task to verify that for p = A = 0 trasformations (10, 11) give rise to solutions that do
not satisfy (14), being therefore non-physical. In open waveguides they do not pollute the part of the
spectrum corresponding to guided modes (]p] > &in > 0: see [7,26]), while in closed waveguides they
may give problems near the cut-off frequencies wi of propagation modes (pi(wi) = 0: see [6]). In the FE
approximation, however, they are easily detected as their number is equal to the dimension of Yh (see
below).
In an approximation scheme the spaces X and Y of the Weak Problem are replaced by two finite
subspaces X, c X and YJ, c Y, where the index h belongs to a subset of JR’ having zero as an accumu-
lation point. We require that X, and Y, be ‘good approximations’ of X and Y, precisely they should
verify the following property (see [21]).

PROPERTY H.

(26)
(27)
 A. Toselli/Comput. Methods Appl. Mech. Engrg. 140 (1997) 221-235 225

In the FEM one considers a triangulation of the bounded domain 0 and h represents the maximum
diameter of the mesh (see again [21] for the details).
Let uh and (ph be the solutions of the approximate problem. If nx and rry are the dimensions of X,
and Yh and 8 and V are two column vectors which represent the degrees of freedom of uh and (ph, the
approximate eigenpairs are given by the real symmetric generalized eigenvalue problem of dimension
N =nx+ny

(28)

where matrices A: B, C, D are built with the basis functions of the subspaces Xh and Yh.

3. Analysis of spurious solutions

As previously observed, when the finite element subspace X, consists of vectorial continuous nodal
functions, which are polynomials of degree r (r = 1,2, . . . ) over each triangle, parasitic solutions appear.
That is, together with solutions which approximate eigenpairs of the exact problem, one obtains non-
physical numerical solutions which do not tend to a solution of the exact problem when the triangulation
of the domain is refined. The appearance of such solutions is a numerical drawback which stems from
the approximation subspaces employed for the particular differential formulation chosen to solve the
waveguide problem. They have been characterized by the property that they do not satisfy the divergence
equation (see [6])
V . (Zr . E) := Vt . (Zl . et) - i&5zez = 0, (29)
that, according to (10, ll), states that
v, * (Z, . u) t p&p = 0. (30)
Our analysis of parasitic solutions is similar to the one carried out in [2] for electromagnetic resonators,
but the sufficient condition for their elimination is different and is typical of mixed methods.
Solutions of (4) clearly satisfy (29) for k # 0 (time-varying fields). Thanks to the property (25), solu-
tions of the Weak; Problem (Eqs. (13, 14)) satisfy (30) in a weak sense. Indeed, taking u = Vlc, in (13),
multiplying (14) for p and subtracting, for k # 0 one obtains
(~~t~u,VrcI)-p2(~,~,~)=0 forall+EY. (31)
Let X, and Yh be two finite dimensional subspaces of X and Y, respectively. For what has been said
in Section 2, the kernel of the curl operator defined on Xh is the gradient of a subspace, say vh, of Y.
Observe that in general there is no connection between Yh and V,, and V, may even be empty. Let
Uh=Yhn\‘h, (32)
that we suppose non-empty. Approximating Eqs. (13, 14) of the Weak Problem in X, and Yh leads to
the following

APPROXIMATED PROBLEM. Find hh E @. and {uh, (ph} E X, x Yh such that

+h, uh) = Ah (@h7 uh) + +h7 (ph)) v % E xh7 (33)

0 = hh (‘+h, +h) + d(cph> $h)) v @h E yh* (34)

Considering +h = Xh with Xh E U, in (34) and uh = VXh in (33), one obtains

(Z, . uh,bii) - & ( .+phDh,xh) = 0 for all xh E uh. (35)
If U,, is a ‘good approximation’ of Y, Eq. (35) assures the elimination of parasitic solutions. We can
state the following
226 A. ToseWComput. Methods Appl. Mech. Engrg. 140 (1997) 221-235

PROPOSITION 1. Let Xt, and Yt, be two subspaces of X and Y, respectively. If Yt, satisfies Eq. (27) and
grad Y, c X,,, (36)
then the Approximated Problem (33, 34) is free of spurios solutions.

PROOF (outlined). If condition (36) holds, Y,, is contained in V, and hence U,, = Yh. Eq. (35) is then
valid for all xh = I,!+,in Yh and (27) assures that every limit point {p, u , ‘p} E C x X x Y of a family
{&, uh, ‘ph} E C x X,, x Yh of approximated eigenpairs satisfies the divergence equation (31). 0

We remark that Eq. (36) puts a constraint on the spaces X, and Yh as it is generally required for the
stability of the approximation of mixed problems (see [1,18,23]). In particular condition (36) is formally
identical to the one considered for the 3D mixed problem in [23]. An additional condition on the FE
spaces is also.required for the approximation of the spectrum of a non-compact operator (see [22], for
instance): in this case properties (26, 27) only ensure that every eigenvalue of the Weak Problem is well
approximated for h small, but it is not sufficient to prevent the existence of accumulation points of the
approximated spectrum other than the eigenvalues of the exact spectrum. The linear operator relative
to the Weak Problem of Section 2 must be necessarily non-compact, as a continuum of eigenvalues must
exist (see [6,7] for details).
The analysis carried out here is only outlined and incomplete. In particular the analysis of the exact
eigenvalue problem and of the stability and convergence properties of the approximated one is required.
Mixed problems similar to the one given by (13, 14) are treated in [1,9], for instance, but the properties
of the bilinear forms involved are not satisfied in this case. Such analysis will be the subject of a following
work.

4. Finite element spaces for closed waveguides

Let 0 be the bounded domain, supposed with polygonal boundary r, that identifies the section of a
closed waveguide. We consider a regular triangulation (see [21]) I h, consisting of triangles K. Let ny
and nx be the number of nodes and sides of 7h that do not belong to r. We define Yh as

(37)
where PI(K) is the space of linear polynomial functions over K. We summarize the properties of Yh in
the following lemma (for its proof see [21]).

LEMMA 1.
(i) A function &,, such that I++,I~E PI(K) f or all K E 7h, belongs to Yt, iff it is continuous at the
nodes of the triangulation and vanishes at the nodes belonging to P;
(ii) Yt, is a subspace of dimension ny of H:(0);
(iii) &, E Yt, is uniquely determined by its ny values, say {vi(#t,j, i - 1 . ’ , ny }, at the nodes that do
n;t belong to I’.

Considering the subspace X,, we choose a spclcc consi_t;~~g LcL 23 I ~3 :ntiai elements of order zero
(edge elements): see [3,12,20,23].
After setting

we define
[1
u/, : uh = @I + a2 -,’ , cuj E Po(K)2, a2 E PO(K)

Xh = {uh E Ho(curl, 0) : UhlK E R(K), v K E G}. (39)

The following lemma summarizes the main properties of Xh: the proofs carried out in [20,23] for the 3D
case can be easily adapted in 2D.
 A. Toselli/Comput. Methods Appl. Mech. Engrg. 140 (1997) 221-235 221

P3
A
1
4
K
fi
p1 p,
13
Fig. 2. Degrees of freedom for X, and Yh: Ii and Pj, respectively

LEMMA 2.
(i) A function uh, such that nhlK E R(K) f or ail K E Ih, belongs to H(curl, @ iff its tangential com-
ponent is continuous across the sides of the triangles;
(ii) such function is uniquely determined by its moments

(40)
over every side Ii of the triangulation: IYt((uh)is equal to the (constant) tangential component of vh
along li (ri is the unit vector tangent to li);
(iii) xh is a subspace of Ho(curl,@ of dimension nx and for all oh E xh its moment ai vanishes if
li E r;
(iv) for the couple of subspaces defined in (37, 39) one has

gratl Yt, c &,. (41)

We remark that the vector functions of X, are discontinuous and, in general, only the continuity of
their tangential component is ensured across the sides of the triangulation, as it is physically required for
the electric field. Expressions for the basis functions of X, can be found in [3], while for error estimates
for Yh and Xh we refer to [21] and [23,20], respectively. Fig. 2 summarizes the degrees of freedom of Xh
and Y,,.
The FE spaces defined above and first-order tangential elements (X,) together with second-order
polynomial continuous functions (Y,) have already been employed in [13] and [17], respectively, for
solving the curl-curl equation (4) for the analysis of closed waveguides and open waveguides together
with the Virtual Boundary Technique (see next section). Spurious solutions are not observed and the
reason is now clear on the light of the analysis carried out in Section 3: Xh and Y, verifies (41) and
hence the hypothesis of Proposition 1.

5. Exponential elements for open waveguides

As previously mentioned, in open waveguides we have

0 = Et*. (42)
As 0 is unbounded, the FEM is not directly applicable. The Virtual Boundary Technique (VBT,
see [13,14]) is generally employed: since for propagation modes the field decays rapidly far from the
axis, the waveguide is enclosed in a virtual surface & and perfect electric conductor conditions are im-
posed (see Eq. (6,7)) on the intersection r0 of such a surface and the unbounded section 0. & must be
sufficiently far from the core so that its presence does not influence the solution in the inner part of the
waveguide. Such1a technique, though very simple, is not completely satisfactory near cut-off frequencies
(see Section 1 and [7,26]) where the field considerably extends in the outer region, imposing that Fa be
very far from the core and the corresponding computational domain very large. Moreover, as we will
show in the next section, the eigenvalues obtained are not always accurate.
We suppose tlhat the waveguide is such that its section 0 verifies
228 A. ToseWComput. Methods Appl. Mech. Engrg. 140 (1997) 221-235

PROPERTY D. The unbounded domain 0 C R2 is decomposable as

n=Q)ur,ua,, (43)

where & is a bounded rectangular domain containing the inner part of the waveguide, Z0 its boundary
and &,=JZ-(&uZ’O) (see Fig. 3). In Q, the index of refraction may vary even in a discontinuous
fashion but in & it is represented by a uniform and scalar function, say n2.

We then consider a regular triangulation of Oc as in Section 4 and a decomposition of 0, in sectorial

and angular infinite elements as shown in Fig. 3. In the rest of this paper such a decomposition of Q
will be again called triangulation (‘&) and K will denote its generic element (finite or infinite). Sectorial
elements will be called (infinite) elements of type I: Fig. 4 shows an element of type 1 relative to +oo,.
Angular elements will be called (infinite) elements of type 2: Fig. 5 shows an element of type 2 relative
to (+00x, +oq).
Let ny be the nodes of the triangulation and nx its sides, also counting the infinite sides belonging to
&,. We fix two decay lengths, L, and L,, relative to (x -+ m) and (y -+ m). For the subdomain & we
choose the FE spaces introduced in Section 4, while for the unbounded subdomain 0, we will introduce
suitable Exponential Infinite Elements (EIEs).
After setting

Pl w K triangle
((a0 + WY) exp[-x/L]; ai E @) K relative to +ccX
{ (oo + SX) exp[-y/&l; ai E @} K relative to +ccy
SYW = (44)
i, exp[-x/L] exp[-y/L,]; cue E Cc} K relative to (+oo,., +co,)
...

Fig. 3. Triangulation of the unbounded section of an open waveguide.

Fig. 4. Infinite element of type 1 relative to +c+

Fig. 5. Infinite element of type 2 relative to (+cq, +oo,,)

 A. Toselli/Comput. Methods Appl. Mech. Engrg. 140 (1997) 221-235 229

we define

Yh = Yh(L,, Ly) = {$/I E c”(w2); +& E SY(K)V K E %I)* (45)

It is easy to verify the following

LEMMA 3.
(i) Given L, and L,, a function of Sy(K) is uniquely determined on K by its values at the nodes of
K;
(ii) the decay lengths must be the same for elements relative to foe, and I&, respectively, in order
that definiuon given by Eq. (45) be consistent; the space Yh consists of all the functions that are
continuous in the nodes of the triangulation &, whose restriction to the generic K belongs to S(K);
(iii) a function &, of Yh is uniquely determined by its ny values, say {vj(+h), i = 1,. . . , ny}, at the
nodes of the triangulation.

The same space Y,, was already introduced in [24,10], to approximate the three components of the
magnetic field H in the curl-curl equation (4): as continuous approximation functions give rise to parasitic
solutions, in the cited papers a term that penalizes the divergence of H is added to (4).
In order to define the space & we set
[ L?(K) K triangle

K relative to +ocX

SX(K> = ai E PO(W) K relative to +oo,, (46)

- (y/L,)]; ot E PO(K) K relative to (+cox,+ooy)

where R(K) has been defined in (38). We then define

xh = &(Lx,L,) = { uh E H(CUrl,R2) : UhlK E S,(K), V K E G} . (47)

The properies of & are summarized by the following lemma: the proof is immediate (in particular, see
the definition of conforming elements in H(cur1, 0) given in [20]).
LEMMA 4.
(0 Given L, and L,, a function u h of Sx(K) is uniquely defined on K by its degrees of freedom
1
u h . Ti dl li side of a triangle,
6 J 1,

&(uh> = ; u h . Ti dl li infinite side parallel to x, (48)

x J t,
-1 uh . Ti dl li infinite side parallel to y,
1 L, J Ii
where rt is the unit vector tangent to side li;
(ii) consequenlly, a function uh of & is uniquely determined by its nx moments given by Eq. (48) for
all lt E Ih;
(iii) the decays lengths must be the same for elements relative to &CO, and LOO,,, respectively, in or-
der that the tangential component be continuous across the sides of the infinite elements and the
definition given by Eq. (47) be consistent;
(iv) Eq. (36) is verified for the couple of subspaces defined in (45, 47).
The vector functions of &, are discontinuous and only the continuity of their tangential component
across the sides of lh is ensured, as it is physically required for the electric field.
230 A. Toselli/Comput. Methods Appl. Mech. Engrg. 140 (1997) 221-235

We also remark that one has to add the degrees of freedom relative to the infinite sides of 0,: they
are necessary in order that (36) be valid. Figs. 4 and 5 summarize the degrees of freedom of X, and Y,.
The choice of the spaces X,, and Yh is justified by the fact that, given Eq. (l), in 0, the field e(x,y)
must be a solution of the Helmholtz equation (see [26])
Ate - y*e = 0, (49)
where

y* = p* - k*n;, WV
and n2 is the (constant) index of refraction in 0,. Together with a suitable condition at infinity (see
[19]) this imposes that the field at infinity be an exponential function of (-yr), with r = dw.
Consequently, the decay lengths may be taken according to

yx = l/L, = p* - k*n$, (51)

21

yy = l/L, = p* - k*n*2’ (52)

4
We remark that, as we have assumed the index of refraction n2 constant in &, the decay lengths given
by (51, 52) are the same for elements relative to foe, and foe,, respectively, and the definitions given
by (45, 47) are consistent.
Since the approximation spaces X, and Y, depend on the eigenvalue A = p* through Eqs. (51, 52)
the algebraic problem given by Eq. (28) becomes a non-linear one, as matrices A, B, C, D depend on
A in a non-linear fashion. We obtain the eigenvalue problem

[AP :] [:]=A[$; Z$)] [Z]J

that is
Q(h)x = AR(A)x. (53)
Ear such a problem an iterative method similar to the one proposed in [lo] has been employed. Starting
from an initial value of the propagation constant, matrices Q(Aa) and R(Aa) are built and a linear
generalized eigenvalue problem is solved to obtain At, which is then used to build the new matrices. The
single step of the method is given by equation

Q(&)x~+I = Ai+1R(Ai)xi+l- (54)

The convergence properties of (54) have been tested on several examples, starting from different values
of the decay length and some results are presented in the next section.

6. Numerical results

As mentioned in Section 4, edge elements have already been employed in [13], for the analysis of closed
waveguides and they remove parasitic solutions. In addition we have studied the order of convergence
of the method (the results are not presented here), which is found to be quadratic for every eigenvalue
corresponding to a propagation mode (discrete spectrum) and for all frequencies w (see k in Eqs. (13,
14)).
The EIEs represent an original contribution of our work and in this section we present the numerical
results relative to two open waveguides studied with this method. In particular, we are interested in
the elimination of ;larasitic solutions, the comparison of EIEs versus the VBT, the convergence of the
iterative method given by (54) and the dependence of the boundary rc on the frequency.
The first configuration studied is shown in Fig. 6: it represents a rectangular optical fiber or a prototype
of a strip waveguide (see [7]).
 A. Toselli/Comput. Methods Appl. Mech. Engrg. 140 (1997) 221-235 231

Fig. 6. Rectangular optical fiber with step-index profile.

Fig. 7. Family of lines {r, : p = 1,2, . . . } (only a quarter of the section displayed).

Setting

u = kt
dn:- n;/v, (55)
we have chosen:
Iz1 = 1.05, 112= 1, t = 2.5 pm, w = 2t, v = 1.
As shown in Fig. 7, we have considered a set of boundaries {r, : p = 1,2,. . . }, that progressively move
away from the core. We have first calculated the propagation constant for the Es mode imposing the
conditions given by Eqs. (6, 7) (Virtual Boundary Technique). The corresponding eigenvalue stabilizes
for p N 4 and such value of p is frequency-dependent. For u = 0.4, which is a value very near the cut-off
frequency of the E$ mode, the corresponding eigenvalue has not yet stabilized for p = 12! This shows
that the VBT is not efficient near cut-off frequencies.
Considering n’ow the EIEs introduced in Section 5, the eigenvalue A = p2 for the same mode is
displayed in Fig. 8 versus the steps of the iterative method given by Eq. (54), for different boundaries
I” and for different values of As. In Table 1 we have shown the eigenvalues obtained at the 4th iteration

15.680

15.670

15.550

k 15.550

15.640

15.630

15.620 '
1.0 2.0 3.0 4.0 5.0
rtsps

Fig. 8. Exponential Elements (mode E$): eigenvalues found versus number of iteration steps for different choices of I”.
232 A. ToseWComput. Methods Appl. Mech. Engrg. 140 (1997) 221-235

Table 1
Mode Eg: eigenvalues obtained for different p, with the EIEs (p = 1,2,3,4,5,6,9) and the VBT (p = 8)
A = p2 Relative error
P = 1 (exp) 15.66326368 3.28 x 10m4
P = 2 (exp) 15.66682487 1.01 x 10-d
P = 3 (exp) 15.66802538 2.42 x 1O-5
P = 4 (exp) 15.66831600 5.71 x 10-6
P = 5 (exp) 15.66838416 1.36 x 1O-6
P = 6 (ew) 15.66840037 3.29 x lo-’
P = 9 (exp) 15.66840553
p=8(VBT) 15.62745232 2.61 x 1O-3

and the one obtained with the VBT for p = 8, together with the relative error with respect to the value
obtained with exponential elements for p = 9.
We can make the following remarks:
The iterative method given by (54) converges even when ha is a poor estimate of the exact A.
If ha is quite near A, the eigenvalue obtained stabilizes after about 3/4 steps: thus, for a lixed p,
our iterative method converges very rapidly (in Fig. 8 values obtained for p = 4,5,6,7,8,9 would
overlap if displayed together).
For a fixed frequency and for different values of p, the iterative method proposed provides different
values of the propagation constant: p must be sufficiently large so that the field be very small in
the unbounded domain 0, and the approximation be accurate. Table 1 shows that p = 3 or p = 4
is enough in this particular example, where we are far from the cut-off frequency of the mode
considered. Numerical simulations have been carried out near cut-off frequencies and a larger
value of p is required in this case. For this particular example and for u = 0.4, p N 7 is required.
Such value must be compared to the value p = 12, for which the eigenvalue provided by the VBT
has not stabilized yet!
Considering the EIEs, increasing p, the eigenvalue does not converge to the value obtained with
a very far Virtual Boundary (see Table 1). The line r, is quite far from the core and the field can
be considered zero on it: since the value obtained considering the EIEs for p = 9 is only slightly
different from the ones calculated for p < 9 but is considerably far from the one obtained with
a Virtual Boundary very far from the core (p = S), we can deduce that the latter method gives a
poorer approximation than the exponential elements. The reason becomes clear when one examines
Fig. 9: we have shown the transverse field (which is parallel to y) calculated for p = 5 with the VBT
(a) and the EIEs (b). The former method gives a considerably higher field over the line {y = 7.5).
This is because we impose that the tangential component (x-component over {y = 7.5)) be zero:
as the field is parallel to y, this gives poor control for (y + +oo). The calculated field actually

Fig. 9. Mode Es: magnitude of the transverse field e, calculated with the VBT (a) and the EIEs (b).
 A. Toselli/Comput. Methods Appl. Mech. Engrg. 140 (1997) 221-235 233

tends to zero when increasing p, but slower than the exact one. On the contrary, the EIEs give an
almost vanishing field over {y = 7.5) as they also allow control over the normal component (see
the degrees of freedom in Fig. 4).
l Parasitic solutions do not appear.
The second waveguide that we have analyzed is shown in Fig. 1 of Section 1. Fig. 10 shows its section
and the triangulation considered. Setting

(56)
b = (P2/k2)
-- -4 (57)
n:-n; ’
we have chosen:

n*1 = 2 .’25 n; = 2 *13 7 rz; = 1, s=3Fm.

We observe that this configuration does not satisfy Property D, as the index of refraction is discontin-
uous along an infinite line parallel to x-axis: the decay lengths would be then discontinuous, according
to Eqs. (51, 52), and the definitions of X,, and Yh (Eqs. (47, 45)) not consistent. However, the EIEs
can be employed by defining an equivalent index of refraction, neq, to be used in (51), for elements
relative to (x -+ +(x0>.The quantity neq must be a proper average of n2 and 123.Two decay lenghts must
then be used for (y --) Z&I). Choosing neq = n2, Fig. 11 shows the dispersion curve of the mode HK,
obtained with the triangulation of Fig. 10 and with a triangulation of half diameter, together with the
values deduced from [24]. Our results are in good agreement with the ones presented in [24], except
near the cut-off frequency (u N 0.8).
REMARK. In the iterative method given by (53), the eigenvalue sought must be carefully selected at
each step. While eigenvalues of greater modulus correspond to guided modes in the Weak Problem of
Section 2 (see [7,261])and in Eq. (28), eigenvalues obtained at each step in Eq. (53) may lose this property,
when Ai is very close to the exact solution: in this case only one eigenvalue is in general accurate.

0.6 -

0.4 -

Fig. 10. Triangulation of half a section of a strip waveguide.

Fig. 11. Mode Hfi: dispersion curves for different triangulations.

234 A. ToseWComput. Methods Appl. Mech. Engrg. 140 (1997) 221-235

7. Conclusions

The curl-curl equation for the electric field has been approximated with the Finite Element Method in
order to analyse electromagnetic waveguides. When approximating the electric and magnetic fields with
continuous nodal functions, parasitic solutions are generated, unless particular differential formulations
are employed. In this work we have proposed an original analysis of such solutions, which leads to
the formulation of a sufficient condition for their elimination. This condition puts a constraint on the
approximation spaces of the electric field components and is typical of mixed methods. Two-dimensional
edge elements, already employed in [13], have been used for closed waveguides.
A new family of exponential vector infinite elements, consistent with 2D edge elements, has been
introduced for the analysis of open waveguides. They remove parasitic solutions and allow to treat
waveguide problems in unbounded domains. They give rise to a non-linear eigenvalue problem: an iter-
ative method has been proposed and a real generalized eigenvalue problem is solved at each step. We
have presented some numerical results, obtained with an algorithm for dense non-symmetric matrices
(QZ-trasformation), which is highly inefficient. However, a new Lanczos-type algorithm for real sym-
metric indefinite sparse matrices is being implemented and tested and will make the object of a future
work: in particular the latter algorithm should allow the calculation of only few eigenvalues in a much
shorter time and make the selection of the eigenvalue sought at each step of the iterative method easier.
The exponential infinite elements proposed seem closely related to the nature of the particular problem
studied (high-frequency electromagnetic open waveguides). For such problems they have been found to
be more accurate and efficient than the Virtual Boundary Technique. The employment of these elements
is not straightforward in source-driven electromagnetic problems, such as antenna’s propagation and
electromagnetic waves in open space.

Acknowledgments

This work has been taken from my thesis at Politecnico of Milan, Italy: I would like to thank Prof.
Alfio Quarteroni for his continuous help and for the discussion of the analytical part of my work as well
as the numerical results. I also thank Doctor Osni Marques at Cerfacs, Toulouse, for his help in the
numerical solution of the algebraic generalized eigenvalue problem and for the discussion of the results
obtained.

References

[l] F. Brezzi and M. Fortin, Mixed and Hybrid Finite Element Methods (Springer-Verlag, New York, 1991).
[2] A. Bossavit, Electromagnetisme, en Vue de la Modelisation (Springer-Verlag, Paris, 1993).
[3] Z.J. Cendes, Vector finite elements for electromagnetic field computation, IEEE Trans. Magnetics 27 (1991) 3953-3%6.
[4] P Cheung, M. Silveira and A. Gopinath, Vector analysis of dielectric waveguides using magnetic field finite finite elements with
Helmholtz equation, preprint, Duluth, University of Minnesota, 1994.
[5] K.S. Chiang, Finite-element analysis of optical fibres with iterative treatment of the infinite 2-D space, Opt. Quant. Electron.
17 (1985) 381-391.
[6] R.E. Collin, Foundations for Microwave Engineering 2nd edition (McGraw-Hill, New York, 1992).
[7] Technical Staff of CSELT, Optical Fibre Communication (Editrice Universitaria Levrotto e Bella, Torino, 1980).
[8] EA. Fernandez and Y. Lu, A variational finite element formulation for dielectric waveguides in terms of transverse magnetic
fields, IEEE Trans. Magnetics 27 (1991) 3864-3867.
[9] V Girault and PA. Raviart, Finite Element Methods for Navier-Stokes Equations (Springer-Verlag, Berlin, 1986).
[lo] K. Hayata, M. Eguchi and M. Koshiba, Self-consistent finite/infinite element scheme for unbounded guided wave problems,
IEEE Trans. Microwave Theory Tech. 36 (1988) 614-616.
[ll] E Kikuchi, Mixed formulations for finite element analysis of magnetostatic and electrostatic problems, Internal Report 87-11,
University of Tokyo, 1987.
[12] F. Kikuchi, On a discrete compactness property for the Ntdtlec finite elements, Internal Report 88-12, University of Tokyo,
1988.
[13] M. Koshiba and K. Inoue, Simple and efficient finite-element analysis of microwave and optical waveguides, IEEE Trans.
Microwave Theory Tech. 40 (1992) 371-377.
 A. ToseWComput. Methods Appl. Mech. Engrg. 140 (1997) 221-235 235

[14] M. Koshiba, K. Hayata and M. Suzuki, Improved finite-element formulation in terms of the magnetic field vector for dielectric
waveguides, IEEE Trans. Microwave Theory Tech. 33 (1985) 227-233.
[15] S.Y. Lee, Highly storage-efficient hybrid finite element/boundary element method for electromagnetic scattering, Electron.
Lett. 25 (1989) 1273-1274.
[16] J.L. Lions and E. Magenes, Non-Homogeneous Boundary Value Problems and Applications, Vol. I e II (Springer-Verlag,
Berlin, 1972).
(171J.F. Lee, D. K. Sun and Z. J. Cendes, Full-wave analysis of dielectric waveguides using tangential vector finite elements, IEEE
Trans. Microwave Theory Tech. 39 (1991) 1262-1271.
[18] B. Mercier, J. Osb’arn, J. Rappaz and P A. Raviart, Eigenvalue approximation by mixed and hybrid methods, Math. Comput.
36 (1981) 427-453.
[19] K. Muller, Foundations of of the Mathematical Theory of Electromagnetic Waves (Springer-Verlag, Berlin, 1969).
[20] J.C. Nedelec, Mixed finite elements in R 3, Numer. Math. 35 (1980) 315-341.
[21] A. Quarteroni and A. Valli, Numerical Approximation of Partial Differential Equations (Springer-Verlag, Berlin, 1994).
[22] J. Rappaz, Approximation of the spectrum of a non-compact operator given by the magnetohydrodynamic stability of a
plasma, Numer. Mat. 28 (1977) 15-24.
[23] P.A. Raviart, Finite element approximation of the time-dependent maxwell equations, Internal Report, Palaiseau, Ecole
Polytechnique, 1993.
[24] B.M. Rahman an’d J.B. Davies, Finite-element solution of integrated optical waveguides, J. Lightwave Technol. 2 (1984)
682-687.
[25] B.M. Rahman, EA. Fernandez and J.B. Davies, Review of finite element methods for microwave and optical waveguides,
Proc. IEEE 79 (1991) 1442-1448.
[26] C. Tsao, Optical Fibre Waveguide Analysis (Oxford Univ. Press, Oxford, 1992).