You are on page 1of 34

TICAM REPORT 97-02

February, 1997
Modeling of Electromagnetic Absorption/Scattering
Problems Using hp-Adaptive Finite Elements
L. Demkowicz and L. Vardapetyan
MODELING OF ELECTROMAGNETIC
ABSORPTION/SCATTERING PROBLEMS
USING hp-ADAPTIVE FINITE ELEMENTS
L. Demkowicz and L. Vardapetyan
The Texas Institute for Computational and Applied Mathematics
The University of Texas at Austin
Taylor Hall 2.400
Austin, Texas 78712, USA
Dedicated to Prof. J. Tinsley Oden
on the occasion of his 60th birthday
January 1997
Abstract
A model problem for the steady-state form of Maxwell's equations is considered.
The problem is formulated in a weak form using a Lagrange multiplier, laying down a
foundation for a general class of novel hp-adaptive FE approximations. A convergence
proof for affine elements is presented. The proposed method is illustrated and veri-
fied with a series of 2D experiments including elements with curved boundaries and
nonhomogeneous media.
Key words: Maxwell's equations, hp finite elements, error estimates
AMS subject classification: 65N30, 35L15
1 Introduction
This work has been motivated by two classes of practical applications. The first one deals
with the design and analysis of mixed digital and analog signal packages. One of the ultimate
goals of such an analysis is to determine the variation of the capacitance, resistance and
inductance of the package with frequency w. As the noise from the digital part of the package
may be picked up and amplified by the analog part, the modeling of the interaction between
1
different parts becomes of crucial importance. The frequency content of the "switching noise"
of the digital components spreads from the DC to very high harmonics. Consequently, when
solved in the frequency domain, the problem calls for a methodology which would deliver
results with a quality independent of the frequency w, in particular for w ---t O.
One possible alternative is the "electric circuit approach" where the interaction between
the components is modeled with a simplified circuit model [36, 37]. This methodology pro-
vides for a cost effective initial approximation of the electrical characteristics of the package
and frequently delivers sufficiently precise results. For complex geometries, however, an ac-
curate determination of the field quantities through the solution of the Maxwell equations
becomes indispensable. The global capacitance, inductance and resistance of the system are
then deduced from the electric and magnetic fields, see e.g. [2].
The second class of problems motivating this work deals with modeling of the absorption
and scattering of steady state electromagnetic waves by the human body, see e.g. [9]. There
the main challenge comes from complex curvilinear geometries and high nonhomogenuity
of the scatterer/absorber. Again, modeling through the solution of Maxwell's equations is
necessary.
In both classes of problems singular solutions are expected. These may result from both
geometry and rapidly varying material constants. One of the most powerful methodologies
which permits a succesful modeling of singular solutions is the hp-adaptive finite elements
[1]. We emphasize that a true hp method allows locally varying both element size h and order
of approximation p. Only then are the exponential rates of convergence accessible for a wide
class of functions with singularities.
The purpose of the presented research has been to design an hp finite element method
for steady-state Maxwell's equations. Summing up our discussion, we want the following:
a formulation for a class of problems with discontinuous material properties,
a formulation with approximation (stability) properties uniform with respect to fre-
quency w,
a discretization with the possibility of varying locally order of approximation p and
element size h,
curved elements allowing one to model complex, curvilinear geometries.
We think we have found a method which satisfies all these conditions.
There is an extensive literature on the subject. We have benefited from a large number
of related contributions. The mathematical foundations and the necessary functional setting
were established in [10]. H( curl)-conforming FE approximations were introduced by Nedelec
2
[27, 28, 29] and, independetly in the engineering literature among others by Lee [16, 17, 18,
19, 20, 12]. A comprehensive analysis of various existing formulations was provided by
Monk [21, 22, 23, 24, 25], see in particular [24] for the first application of hp methods to
electromagnetics (with constant p, however). For comprehensive reviews of existing work we
refer to [35, 15]. For the theory on hp approximations we refer to [1] and for our related
work on hp discretizations to [4, 31, 5, 8]. Finally, for a number of recent innovative ideas,
including the L2-residual formulations see [14, 11, 26].
The plan of the presentation is as follows. After a short review of basis electromagnetics
in the next section, we present in section 3 a model problem reflecting all essential difficulties
connected with the solution of Maxwell's equations, except for absorbing boundary condi-
tions. We discuss existence, uniqueness and stability of the solutions. Section 4 presents the
definition of a novel finite element with variable order of approximation for which we are
able to prove the convergence, at least for elements with polynomial shape functions (affine
families of elements). Finally, in section 5 we present a number of 2D numerical experiments.
2 A Review of Basic Electromagnetics
The equations
As stated in [15], "the problem of electromagnetic analysis is actually a problem of solving
Maxwell's equations subject to given boundary conditions". More precisely, the following
four fundamental laws are the cornerstones of electromagnetics [34].
Faraday's law,
r EodT+
d
d
r BodS=O (2.1)
J a s t J s
Here S denotes an arbitrary oriented surface, a s its boundary, oriented consistently
with the orientation of surface S, E is the electric field intensity, and B denotes the
magnetic flux density.
Ampere's law,
rHo dT =r J 0 dS + dd r D 0 dS (2.2)
J a s J s t J s
Here H denotes the magnetic field intensity, D is the electric flux density, and J
denotes the current density.
Gauss' law for the electric field:
(2.3)
3
[ Quantity I Symbol ] SI units I
Electric field intensity E = E(aJ, t) Vim
Electric flux density D = D(aJ, t) C/m
2
Magnetic field intensity H = H(aJ, t) A/m
Magnetic flux density B = B(aJ, t) Wb/m
2
Current density J = J(aJ, t) A/m2
Volume charge density p=p(aJ,t) C/m
3
Table 1: Electromagnetic field vectors
where V is an arbitrary volume with surface S oriented outward, and P denotes the
volume charge density.
Gauss' law for the magnetic field:
is BodS =0 (2.4)
All field quantities E, D, H, B, J, p are functions of space location aJ E J B 3 3 and time t. None
of the them needs to be globally continuous. The charge density may even be a distribution, if
surface charges in between different media occur. The corresponding SI units are summarized
in Table 1. The field quantities are related by the following consitutive laws for a simple
medium:

Ohm's law:
D=EE
B=H
J =aE
(2.5)
(2.6)
(2.7)
(2.8)
Here E is the permittivity, is the permeability, and a denotes the conductivity of the medium.
Denoting by EO the permittivity of free space,
1 -9
EO =3671"10 F /m,
and by o the permeability of free space,
(2.9)
4
aIr 0-=0
antenna( copper) 0-=5.8 10
7
S/m
absorber 0-= 0.3 - 7.2 S/m
Table 2: Values of conductivity 0- of interest
we introduce the dielectric constant Cr and relative permeability r,
(2.10)
For the materials of interest the relative permeability r practically equals one, and the
dielectric constant will typically vary in the range of 1 - 70. Typical values of interest for
conductivity 0- are summarized in Table 2, see also [9].
We shall also assume that the current density is split into two parts, an impressed (im-
posed) current Jimp, and actual (ohmic) current J related to the electric field E through
Ohm's law stated above.
In regions where the field quantities are smooth, the four conservation laws (2.1)-(2.4)
imply four differential Maxwell equations:
aB
VxE+- =0
at
aD
VxH-- =J
(2.11)
at
VoD =p
VoB =0
Across an interface between two media with different material constants, say medium 1 and
medium 2, the conservation laws imply the following jump conditions:
n X (E2 - E1) =0
nx(H2-H1)=0
no (D2 - DJ) =0
no (B2 - B1) = 0
(2.12)
For piecewise smooth functions, the four differential equations (2.11), and the four jump
conditions (2.12) are equivalent to the original four Maxwell laws in the integral form.
5
Boundary conditions
In the following discussion we will restrict ourselves to two kinds of boundary conditions
only. For a boundary adjacent to a perfect conductor we have [34]:
nxE =0
nxH =K
(2.13)
noD
=ps
noB =0
where ps is a surface charge, and K denotes a possible surface current density. As K and
ps are in general a-priori unknown, only the first and last conditions of (2.13) provide useful
information:
nxE =0
noB =0
A dual boundary condition (a magnetic symmetry wal0 [14] reads as follows:
nxH =0
noD =0
Radiation conditions
(2.14)
(2.15)
For steady-state problems in unbounded domains, Maxwell's equations have to be comple-
mented with radiation conditions eliminating waves coming from infinity.
3 Analysis of a Model Problem
In order to discuss the main mathematical issues we will restrict ourselves in this paper to
a presentation of the following model problem.
The problem is set up in a bounded domain nconsisting of two disjoint parts ni, i=1,2,
occupied by media with different material constants Ci, i, a i, i=1,2, and separated by an
interface r12. Boundary r of the domain n consists of two disjoint parts r1 and r2. We
wish to find an electric field E =E( aJ), and magnetic field H =H (aJ), aJ E n, satisfying
the following equations:
6
steady-state form of the Maxwell equations in ni, i=1,2,
vX E+jwiH = 0
vX H - jWciE - aiE =J'mp
v0 (ciE) =p
Vo (iH) =0
jump conditions on interface rI2:
n X [ E ] = 0
nx [ H] =0
no [cE] =ps
no [ H] =0
(3.1)
(3.2)
where [] denotes the jump of the corresponding quantity across the interface, e.g.,
(3.3)
boundary conditions on r1:
nxE =0
no (H) =0
boundary conditions on r2:
nxH =0
no (cE) =0
The impressed current Jimp has to satisfy the usual compatibility conditions [14]:
(3.4)
(3.5)
Vo J'mp =0
no J'mp =0
in ni, i= 1,2
on r12 (3.6)
Volume charge density p and surface charge density ps are additional unknowns in the
problem. In order to eliminate p, we take the divergence of (3.1h:
-jwVo(cE)-Vo(aE)=VoJimp=O ,
replace (3.1h with the equation:
7
(3.7)
(3.8)
and use the original Gauss law to calculate density p after the field E has been determined.
Similarly, the jump condition (3.2h is used to calculate the surface charge density ps , and
(3.8) implies a new jump condition in the form
no [(jwc +alE] = 0
and a new boundary condition on r2, replacing (3.5h
no ((jwc +a)E) =0
(3.9)
(3.10)
Equivalently, the elimination of the charge densities can be done for the integral form of the
Maxwell equations, or for the Maxwell equations understood in the distributional sense with
the assumption that the charge density p includes distributions corresponding to the surface
charge densities.
The choice of the model problem reflects all the fundamental difficulties related to the
radiation/ absorption problems of interest, except for open boundary conditions.
Reduction to the (reduced) wave equation
Let E, H be a sufficiently regular solution to the model problem. Eliminating H from the
equations, interface and boundary conditions above, we arrive at the following boundary-
value problem for the electric field E.
Reduced wave equation and the Gauss law in ni, i= 1,2,
vX (~V X E) - (W2ci - jwai)E =_jwJ'm
p
i
V 0 ((jwci +ai)E = 0
jump conditions on r12,
n X [ E ] =0
n X [IV X E ] = 0

no [(jwc + a)E] =0
boundary condition on r1,
nxE=O
boundary conditions on r2,
n X (IV X E ) =0

no ((jwc +a)E) =0
8
(3.11)
(3.12)
(3.13)
(3.14)
Conversely, let E be a solution of (3.11)-(3.14). With (3.1h defining H,
1
H =---:--V X E
J Wi
(3.15)
we automatically satisfy (3.1h and (3.1h- Taking the divergence of (3.1h, we get equation
(3.1)4' Next formulas (3.15) and (3.1h imply jump condition (3.2h and boundary condition
(3.5h. Finally, (3.15) implies that
1
n0 [ H] = - -;-n0 [V x E ] =0
J W
(3.16)
due to the continuity of the tangential component of E guaranteed by jump condition (3.2h.
Similarly, boundary condition (3.14h implies boundary condition (3.4h-
Thus the reduced wave equation formulation is equivalent to the original problem.
REMARK 1
1. The redundancy of the Gauss law for the magnetic field can be derived from the integral
form of the Maxwell equations. For the steady-state case, the Faraday law (2.1) reduces
to:
rEo dT +jw r BodS =0
J a s J s
and, in particular, for a closed surface S, it implies (2.4)
(3.17)
2. We emphasize the formal character of the presented equivalence proof. All considera-
tions on VoH have been done under the assumption that the solution E to the reduced
wave equation is sufficiently regular to allow considering the divergence V 0 (I V x E ).

Such a regularity result may depend upon the regularity of the domain n.
I
Variational formulation
We introduce the space of admissible solutions:
(3.18)
and the space of Lagrange multipliers
(3.19)
9
The boundary condition in (3.18) is understood in the sense of the generalized Green formula
[10], and the boundary condition in (3.19) in the sense of traces.
Following the usual procedure for the mixed methods, we arrive at the following varia-
tional formulation:
E E W, P E V
r (V x E ) 0 (V x F) daJ
In
- k (w
2
c - jwa)(E + Vp) 0 F daJ
= -jw k (Jimp 0 F) daJ, VF E W
=0 Vq E V
(3.20)
Substituting in (3.20h F =V q, we obtain a variational equation for Lagrange multiplier
p ,
{
PEV
(3.21)
in(w
2
c - jwa)Vp 0 Vij daJ =0 Vq E V
which implies that (at the continuous level) p =O. For that reason p is sometimes called the
dummy variable.
Integrating (3.20h by parts and using the standard Fourier lemma argument, we can
write down the strong form of equation (3.20)1' The wave equation (3.11h gets replaced by
(3.22)
with the remaining equations, interface and boundary conditions unchanged. Variational
argument (3.21) can be reproduced by taking divergence of (3.22),
(3.23)
and by multiplying (3.22) with a unit vector n normal to r2, and extending (3.22) to r2.
As a result we get the Neumann boundary condition on p on r2,
(3.24) on r2
a p =0
a n
built into the variational formulation (3.21). Note that the boundary condition (3.14h
implies that
1
noV X (- V x E ) = 0 on r2
i
(3.25)
1 0
Stability, Existence and Uniqueness
In the following analysis we will restrict ourselves to the case of lossless media only, assuming
a =O. We begin by rewriting (3.20) in an abstract form:
!
E E W, P E V
a(E,F) -w
2
(E,F) -w
2
b(p,F) =wI (F) VF E W
b(q,E)=O VqEV
with
1
1 -
a(E,F) = -(V x E) 0 (V x F) daJ
n
(E , F) =in cE 0 F daJ
b(p, F) =incVp 0 F daJ =(Vp, F)
I(F) =-j inJimp 0 F daJ
A more compact form reads as follows;
{
(E ,p) E W x V
B((E,p), (F, q)) =L((F, q)) V(F, q) E W x V
where
W x V is a Hilbert space with the norm:
II( E,p ) 1I2 = IIEl l w + Il p l l ~
=IIEI12+ I I V x EI12 +w 4 11Vp l 12
B(', .) is a sesquilinear form on W xV:
(3.26)
(3.27)
(3.28)
(3.29)
B((E,p), (F, q)) =a(E, F) - w
2
(E, F) - w
2
b(p, F) +w
2
b(q, E) (3.30)
L() is an antilinear form on W X V:
L((F, q)) =wl(F) (3.31)
Note that placing the factor w in front of IIVp l l in the definition of the norm in space W x V
is equivalent to replacing the Lagrange multiplier p in the formulation with the product w-
2
p.
11
It is well known that the eigenvalue problem
{
E E W, A E C
a(E, F) =A(E, F) VF E W
(3.32)
admits eigenpairs (0, W 0), (Ai, Wi) where the zero eigenvalue is of infinite multiplicity and
the remaining eigenvalues Ai are real, positive, of finite multiplicity, and form a sequence
converging to infinity. The corresponding eigenspaces form an orthogonal decomposition of
the L
2
-space with the weighted L
2
-product (3.272), Introducing a sequence of orthonormal
eigenvectors eI, e2, ... corresponding to positive eigenvalues Ai, we can represent an arbitrary
vector F from L
2
(n) in the form:
00
F =Fo + L:Fiei
i=l
(3.33)
where Fo E Wo, (Fo, ei) = 0, i= 1,2, ... , and Fi denote the spectral components with
respect to eigenvectors ei, Fi =(F, ei)'
Consequently, we have the following spectral representation:
00
a(E, F) - w
2
(E, F) =-w
2
(Eo, Fo) + L:(Ai - w2)EiFi
i=l
(3.34)
Finally, we shall assume that the domain n and boundary conditions are selected in such a
wayl that:
Eo E W 0 {=}~ < P E V : Eo =V < P (3.35)
Now, a straightforward calculation using Lagrange multipliers leads to the following result:
sup IB((E,p),(F,q))1 = max IB((E,p),(F,q))j =:
II( F,q) l I=l II( F,q) l I=l
where
2 =IIEol12 +w
4
11E o + Vpl12 + tIA : ~ ~2121Ei12
i=l
The formula above allows for an easy calculation of the stability constant:
' 2: II
E
ol I' +w
4
l 1
E
o +Vp l l ' +"jin{I"; ~ ~ 'I rECI +",l l E,I'
. { ( IA i _W
21
)2 . }II 112 4 11 11
2
~ mIll 1, 1+ Ai ' Z = 1,2,... E W +w Eo + Vp
1We emphasize that n needs not to be simply connected!
1 2
(3.36)
(3.37)
(3.38)
and, therefore,
1+ Ai .
~ max{l, IAi _ w
2
1' z =1,2, ... }
=w 211Vp l i ~ w211Eo +Vp l l =w 211Eoil ~ ( 1+w
2
)
(3.39)
This results in an estimate for the global LBB (inf sup) constant an the final stability result:
(3.40)
Note that, in particular, the solution is uniformly stable for w ---t O.
Combined with the standard theory (see e.g. [32]), the stability estimate above implies
the following theorem.
THEOREM 1
Let Ai denote the eigenvalues (3.32). Then, for any w
2
i= Ai, problem (3.26) possesses a
unique solution and the stability estimate (3.40) holds. I
REMARK 2
1. Enforcing the divergence condition (the Gauss law) explicitly by introducing the La-
grange multiplier is essential in obtaining the formulation whose stability properties
do not deteriorate with diminishing frequency w. A majority of results reported in the
literature, see e.g. [15, 35], are based on the simplified formulation:
{
E E W
a(E, F) - w
2
(E, F) =wl(F) VFEW
(3.41)
The corresponding stability result for the curl-free component reads then as follows:
(3.42)
which indicates the deteriorating stability of the formulation for small w.
2. Both non-zero conductivity a i= 0 and possible radiation conditions improve stability
and, in particular, eliminate resonance.
I
1 3
4 Discretization
Galerkin Approximations
Introducing finite-dimensional subspaces W heW and Vh C V, we obtain the approximate
problem by replacing all unknowns and test functions from W, V with their approximations
from W h, Vh.
Eh E Wh,P h E l1h
in(V X Eh) 0 (V X Fh) daJ -in (w
2
c - jwa)(Eh + Vph) 0 Fh daJ
=-jw in(Jimp 0 Fh) daJ, VF
h
E W
h
k (w2c-jwa)EhoVqhdaJ =0 V%EVh
(4.1)
Note that, in general, the discrete pressure P h is non-zero unless, again, V(Vh) C W h. In
such a case, we can repeat the same procedure as for the continuous problem, using in (4.1h
Fh =Vqh,qh E Vh
An hp Finite Element Approximation. A Triangular Element of
Variable Order
We will present the concept using a triangular element but exactly the same technique applies
to rectangles, and in 3D, to prisms, cubes and tetrahedra.
Let ]{ denote a triangle with straight sides Si, i= 1,2,3. We associate with the triangle
a specific order of approximation P =P K which may vary with element I< . Additionally,
with each of the sides we associate a possibly different order of approximation P i, i=1,2,3,
with the assumption that
1 ::; P i::; P i=1,2,3 (4.2)
Due to approximability requirements, in practice we select P =max{pl, P 2,P 3}. We introduce
now two spaces of element shape functions:
the scalar space (to approximate the Lagrange multiplier p) identified as the space of
polynomials of order P +1 which, over each of the three sides Si, reduce to polynomials
of order P i + 1, i= 1,2,3,
(4.3)
14
the vector space (to approximate the E-field) identified as the space of vector-valued
polynomials of order p for which the corresponding tangential components, over each
of the three sides Si, reduce to polynomials of order P i, i=1,2,3:
(4.4)
Here Ti denotes a tangent vector to side Si, i=1,2,3. One can think of either space as a
subspace of polynomials of order p with additional constraints imposed along the sides. The
condition that a polynomial of order p (in one dimension) must reduce to a polynomial of
smaller order P i < p is equivalent to satisfying p - P i scalar constraints
2
and, therefore, the
dimension of the scalar space V(K) is:
or equivalently,
(p + l)(p +2) )
dimV(K) = 2 - (p- P I) - (p- P 2) - (p- P 3 (4.5)
dimV(K) =3 + (P I - 1) + (P 2 - 1) + (P 3 - 1) + (p - l)(p - 2) (4.6)
2
Similarly, we have for the vector space:
dimW(K)
(p + l)(p +2)
=2 - (p - pJ) - (p - P 2) - (p - P 3)
2
=(P I + 1) + (P 2 + 1) + (P 3 + 1)
+(p - 1) + (p - 1) + (p - 1)
(p - 1)(p - 2)
+2 2
(4.7)
The two spaces satisfy the same compatibility condition as their continuous counterparts.
THEOREM 2
The following compatibility condition holds:
< P E V(K) :::::? V < P E W(K)
I
(4.8)
Proof: The proof is trivial. I
In practice both spaces are constructed as spans of predefined element shape functions.
For the scalar approximation we use the shape functions introduced in [5].
2These constraints can be explicitly written down using e.g. L
2
-product and Legendre polynomials.
1 5
~2
Figure 1: Triangular element with variable order of approximation
The element is illustrated in Fig. 1. It consists of seven nodes: three vertex nodes
VI,V2,V3, three mid-side nodes a l ,a 2,a 3" and a middle node a 4' Each of the three mid-
side nodes and the middle node may have a different, corresponding order of approximation
P I, P 2, P 3, P 4, respectively. Introducing the area coordinates )1} =1- el - 6, A 2 =6, A 3 =6,
we define the corresponding scalar-valued shape functions as follows:
vertex nodes shape functions,
X i =Ai, i=1,2,3
mid-side nodes shape functions,
p l -l
.II P 2-
i
)A
Xl ,i = J =O,J =I=i PI 1
p l -l . .
II
(
2 J ' ., 2=1
.. ~-~)(l-~ ,... ,p,-l
J =O,J t;i PI PI PI )
with formulas for X2,i and X3,i obtained by permuting indices,
1 6
(4.9)
(4.10)
middle node shape functions,
X 4,i,j,k =
i-I j-l k-l
m n m
II(AI - -) II(A2 - -) II(A3 - -)
m=O P 4 n=O P 4 1=0 P 4
i-I i m j-l j n k-I k I
II( ---) II( ---) II( ---)
m=O P 4 P 4 n=O P 4 P 4 [=0 P 4 P 4
(4.11)
wi th 1 ~ i, j, k ~ P 4 - 1, i +j + k =P 4.
The choice of the actual global degrees of freedom (dof) is to some extent arbitrary and it
reflects the global continuity requirements. For the scalar case, for instance, we can use the
Lagrangian nodes introduced in the definition of the shape functions. More precisely, if ai,j
denote the Lagrangian nodes corresponding to mid-side nodes aI, a2, a3 and middle node
a4, the corresponding dof are defined simply as:
q ---t q(vd i=1,2,3
q ---t q( ai,j) i= 1,2,3, j =1, ... , P i
THEOREM 3
The Lagrangian dof defined above are V(K)-unisolvent. I
(4.12)
Proof: As the number of dof equals exactly the dimension of space V(K), compo (4.6),
it is sufficient to show that once all dof are zero, the corresponding function q E V(K) must
vanish. I
In order to enforce now the global continuity of the scalar approximation it is sufficient
to assume that the values at the vertex nodes and the mid-side nodes are shared by adjacent
elements.
Note that the Lagrangian dof do not coincide with the dual basis to the original shape
functions. Consequently, the restrictions of the global basis functions (defined as the dual
basis corresponding to the global dof) to an element, do not coincide with the original element
shape functions and one has to use the generalized assembling procedure discussed later.
The choice of the dof for the vector-valued approximation is more elaborate. The concept
is illustrated in Fig. 2. Formally we introduce seven nodes: three tangential mid-side nodes
aLa~,a~, three normal mid-side nodes a
1
,a
2
,a
3
, and again the middle node a4. The
tangential nodes are of order P I, P 2, P 3, resp., whereas all normal nodes and the middle node
are of order P =P4. Next, similarly to the scalar approximation, for each of the nodes
17
~2
~ I
Figure 2: Vector triangular element with va.riable order of approximation
we introduce the corresponding Lagrangian nodes a~,j' ai,j' a4,j' Note that the Lagrangian
nodes for the tangential mid-side nodes include the vertices. Finally, the dof are defined as
follows:
tangential dof:
normal dof:
middle dof
F ---t F (a t ") 0 'riJ " i=1, 2, 3, j =0, ... , P i
1.,) ,
F ---t F(a7,J 0 ni,j i=1,2,3,j =1, ... ,P i-1
F ---t F(a4,j) 0 ik k =1, 2,j = 1, ... , (p - 2)(p - 1)/2
(4.13)
(4.14)
(4.15)
Here ' r i,j denotes tangential unit vectors constructed at points aL and directed from the
vertex node with a smaller (global) number to the vertex node with a greater number, ni,j
denotes outward normal unit vectors constructed at points ai,j' and ik , k =1,2 stands for
the global cartesian unit vectors. Note that for elements with straight edges vectors ' r i,j and
ni,j are independent of index j. We record the more general definition for the sake of curved
elements to be discussed later.
18
THEOREM 4
The dof described above are W(I<)-unisolvent. I
Proof: As for the scalar case, number of dof equals the dimension of the space and,
therefore, it is sufficient to verify linear independence. Assume that E is a function from
W(I<) for which all dof vanish. It follows from vanishing of the tangential components
at vertices that E is zero at vertices. Next for each side Si, from vanishing of the normal
component at additional P - 1 points and the tangential one at P i - 1 points, it follows that
both the normal and tangential components must vanish. Consequently, E is zero along each
side. Finally, from the standard argument on the Lagrange interpolation, E must vanish
everywhere. I
Finally, let us note that by enforcing common values of the tangential components for
adjacent elements we construct the desired H( curl, n)- conforming approximation.
REMARK 3
1. Location of points a ~ ,j' a i,j is arbitrary as long as their number remains the same. Also
vectors ni,j may be replaced with other non-tangential vectors.
2. Formally the presented definitions can be extended to accommodate curvilinear defor-
mation of elements, using in particular parametric maps. The tangential component
along a common interelement side is no longer a polynomial, and enforcing its continu-
ity at P i +1 points does not, in general, imply the global continuity. The approximation
is no longer conforming (internal), and the integrals in the variational formulation are
interpreted element-wise. The choice of points a ~ ,j' a i,j becomes essential and the ap-
proximation theory for the p-version of the FEM suggests using Gauss-Lobatto or
Gauss integration points instead of the uniformly distributed Lagrange points.
I
Convergence Analysis
Let us assume for simplicity that n is a polygonal domain. Denoting by 7 f t a triangulation
of the domain we define the FE spaces W heW and VhC V as follows.
(4.16)
and
(4.17)
1 9
We have the following essential result.
THEOREM 5
The following compatibility condition holds:
(4.18)
I
Proof: Let >E Vh be an arbitrary function. Then by (3.35) V> E Wand we have
to show only that (V IK E W(K) for every element K. But this is guaranteed by the
compatibility condition (4.8). Conversely, if E E W h then there exists > E V such that
V>= E and it is sufficient to show that > K E V(K). But this again follows from (4.8).
I
Repeating precisely the same steps as on the continuous level, we end up with the fol-
lowing estimate for the discrete LBB constant:
-1 { 2 1+ Aih. }
Ih :::;max 1+w , I' 21,2 =1,2, ...
/\ih - W
(4.19)
where Aih denotes the discrete eigenvalues resulting for the Galerkin approximation of (3.32).
We emphasize the essential role of condition (4.18) in deriving the result above.
Combining the discrete stability result with the classical theory on asymptotic conver-
gence (see e.g. [7]), we arrive at the final convergence result.
THEOREM 6
There exists a threshold value ho and a constant C, independent of h, such that
for every h :::; ho. Constant C depends upon frequency w but it is bounded away from zero
for w ---t O. I
Obviously p =0 and, with condition (4.8) satisfied, Ph =0 as well, and (4.20) reduces to
the estimate:
l iE - Ehllw :::; C inf l iE - Ehllw
FhEWh
In practice, of course, due to the round off error, Ph is never exactly zero.
20
(4.21 )
REMARK 4
1. Parametric elements. Once the elements are deformed, the compatibility condition
(4.18) is no longer satisfied, the discrete Lagrange multiplier Ph i= 0, and the presented
proof of the stability result does not go through. A simple perturbation argument,
however, suggests that the method should work as long as the distortion from the
affine elements is small [3]. Numerical experiments indicate that, for small w, the
proposed formulation does remain stable for curvilinear elements whereas the classical
one discussed in Remark 2 does not.
2. hp -error estimates. Theorem 6, when combined with the interpolation error esti-
mates for hp approximations, see [1, 24], results in standard hp error estimates, with
exponential rates of convergence for analytic solutions.
I
Assembling procedure. Constrained approximation
Let ei E W, i=1, ... , N denote the global basis functions to approximate E , F. We now
discuss shortly how to calculate the global stiffnes matrix a ( ei, ej ), and the global load vector
l ( ej), corresponding to the presented hp discretization. We have,
(4.22)
The basis functions ei are vector-valued.
Over each element J {, basis function ej can be represented as a linear combination of the
element scalar-valued shape functions XI, I =1, ... , M,
3 M
~ ~ J . J
ej =L...J L...J CjIXI' l J , Cjl E 1R
J=ll=l
(4.23)
where i
J
denotes the J-th unit vector in 1R
2
. The element contribution to the global load
vector can then be calculated as follows:
IK(ej) = J Kf oej
2 M
=IKe' . 'Ecf l xl iJ )
J =l 1=1
2 M
=L Lcf IIK(xl iJ )
J=ll=l
21
(4.24)
Similarly, we get for the stiffness matrix,
aIei, ej) =JIV X ed 0 (V x ej) daJ
2 M 2 M
=aIL L C[k X li/, L L cflX liJ )
/=lk=I J=II=l
2 M 2 M
=L L L L c[k cflbIX k iI, X liJ )
/=1 k =l J=I 1 =1
(4.25)
Thus, in order to assemble the global matrices, we simply have to determine the coefficients
cfl' Usually, we like to think backwards, from the point of view of the contributor, not the
receiver. In other words, if we have a shape function Xl , we want to determine all global basis
functions ej, and the corresponding coefficients cll' The first step is accomplished through
the information on connectivities and the constrained approximation. Once the set of the
involved basis functions ej has been established, the corresponding constraint coefficients cll
can be evaluated by applying to ej the local, element dof cPI ( a dual basis to XI),
(4.26)
In practice, the original, element dof cP I may be cumbersome to calculate (see [5] for the
description of various hp interpolation procedures). In such a case, functionals cP I may
be replaced with a different set of functionals 'l/Jk , resulting in a system of equations for
coefficients cll (with multiple right-hand sides),
M
L cfl < 'l/Jk ,X I >=< 'l/Jk , (ej 0 iJ ) >
1=1
(4.27)
Systems (4.27) are formed and solved for groups of shape functions connected with a partic-
ular side (edge, base, or side in 3D) of a constrained element. The most convenient choice
for functionals 'l/Jk are the usual Lagrange dof.
REMARK 5 Note that unless hierarchical mid-side node shape functions are used, when-
ever any of the mid-side node orders of approximation P i is different from p, different sets of
scalar-valued shape functions have to be used to represent basis functions corresponding to
tangent and normal dof. I
5 Numerical Experiments
Test cases. All presented results are two-dimensional. Fig. 3 displays three test cases.
The corresponding exact solutions are polynomials of order 5, 7, and 2, respectively. In
22
Case 1 Case 2 Case 3
Rectangular Domain Trapezoidal Domain 2-Media Domain
N N D
=2.0
I I C=LI. I I I I ,-J <=2.0
=1.0
IN D/
="::.V
I D DI
=2.01 =1.0
ID
D 1
0=0.0
0=0.0
0=0.0 0=0.0
D
D = Dirichlet b.c.
N= Neumann b.c.
N
Figure 3: Test cases
N
the third case, due to the discontinuity of constants and c, the corresponding normal (i.e.
horizontal) component of the solution is discontinuous across the vertical interface.
Geometry representation and initial mesh generation. The domain is viewed as a
union of disjoint figures. Topologically, each of the figures is a triangle or rectangle and it is
specified as an image of the reference triangle or rectangle by a particular parametrization.
The concept is illustrated in Fig.4. In practice, the parametrizations may be explicit or
implicit, with an additional assumption on compatibility of parametrizations being enforced.
For details concerning practical implementation, we refer to [6]. The initial mesh generation
is based on the idea of an algebraic mesh generator and hp-interpolation. The concept
is illustrated in Fig.4. Given, for each reference figure, a number m of divisions in the
"horizontal" and "vertical" directions (compatible for neighboring elements, the initial mesh
is always regular), the reference blocks are covered with uniform, regular grids consisting
of elements i< . By constructing a composition of the standard affine map 17transforming
master element k onto element i< and the (restriction of) the block parametrization aJb,
we construct a map from master element konto a curvilinear element K identified as the
image of element i< under the particular parametrization aJb:
K=T(k) =aJb(K), T =aJb 017 (5.1)
In principle, this map could be used directly to define the curvilinear element, i.e. in the
element calculations. In practice, we chose instead to approximate it with polynomials using
23
'1,
1\
K
Figure 4: Geometry representation and mesh generation
the idea of parametric approximation. More precisely, given a particular order of approxima-
tion for element J {, we replace transformation T with its hp-interpolation corresponding to
orders (P l,P 2,P 3,p). The idea of the hp-interpolation follows from convergence theory for hp
approximations [1] and has been introduced in [31]. Roughly speaking, the hp-interpolation
combines the classical interpolation for vertex nodes with local HJ-projections for higher-
order nodes. Given a sufficiently regular function, the corresponding hp-interpolant exhibits
the same orders of convergence (in terms of both hand p) as the corresponding global HJ-
projection (solution to the Laplace equation with Dirichlet boundary conditions imposed
using the HJ projection on the boundary).
For the first and the third case, the parametrization maps are linear and the resulting
meshes consist of affine elements. In the case of trapezoidal domain, two geometry repre-
sentations are used. The first one is based on dividing the domain onto two triangles and
using linear maps for each triangle. Consequently, the corresponding meshes consist again
of affine elements. In the second case, the trapezoid is parametrized directly with a bilinear
map and the resulting triangular elements have curved sides.
Discontinuous solutions. We begin with a solution of Case 3 and a simple illustration
of the fact that we can handle discontinuous solutions. Fig. 5 displays the x-component of
the E-field reproduced using a mesh of quadratic elements.
24
4 el ements, order =2, x-comp onent of E-fiel d 00=2.5
o
-0.571428
-1.14286
-1.11429
-2.28571
-2.85714
-3 ,4 2857
..
-4.51143
-5.14286
5,71429
-8.28571
-8.85714
-H2857
8
2-Media Domain
N
e=1.0 I
e=2.0
=2.0 =1.0
0 I
0=0.0 0=0.0 I 0
N
0=Dirichl et B e
N=Neumann B e
Exact solution is piecewise quadratic
Figure 5: Case3. Capturing discontinuities
h-convergence rates for affine elements. Figures 6 and 7 display h-convergence
plots for test cases 1 and 2, with the trapezoid in the second test parametrized with a
piecewise linear map, and meshes of qudratic and cubic elements. The convergence rates are
in agreement with the theoretical estimates.
h-convergence rates for curvilinear elements. Fig. 8 presents convergence rates
for meshes of cubic elements and two values of frequency w. The method converges but,
compared with the affine elements, the rates are suboptimal. Please note the stable behavior
of the method for the small frequency case. A sample solution, including components of the
E-field, Lagrange multiplier p, and the divergence of the E-field is displayed in Figures 9
and 10. Note that the Lagrange multiplier does not vanish.
Sensitivity with respect to a mesh distortion. As a final example we present a
simple illustration of the sensitivity of the solution with respect to a curvilinear distortion of
the mesh. Figure 11 presents the x-component of the E-field for the first test case obtained
using just two quadratic elements. Depending upon the magnitude of the distortion, the
solution changes but not dramatically. In the corresponding experiment using the standard
formulation the solution simply falls apart.
A simplistic explanation of the stability problems. We conclude this section with
a simple argument explaining why our formulation can handle curved elements while the
classical one cannot. As discussed in the previous section, for a mesh consisting of affine
elements and piecewice polynomial approximation, the eigenvalues of the V x (V x E ) op-
erator include a multiple zero eigenvalue with the corresponding curl-free eigenvectors, and
25
e
W
E 2
o
Z
>-
C)
Q;
C
W"
Oi
.3
-6
Error Analysis, Rectangular Domain
-Iog( h)
( 0 = 2.5
Exact sol ution is quintic
-s::J- Order=2
-0- Order=3
N
=2.0
ol =1.0IN
0=0.0
o
D=Oirichlet B e
N=Neumann B e
Figure 6: Case 1. Experimental convergence rates for quadratic and cubic element.s
Error Analysis, Trapezoidal Domain by 2 Triangles ( 0 = 2.5
Exact sol ution is a p ol ynomial of degree 7
'C'4
e
W
E 3
o
Z
e ; 2
Q)
c
!:!:!.,
C)
o
..J
.,
0.0 0.5 1.0 1.5
-Iog( h)
2.0
----v- Order=2
-0- Order=3
D=Dirichlel B e
N=Neumann B C
Figure 7: Case 2 with linear parametrization. Experimental convergence rates for quadratic
and cubic elements
26
Error A nal ysis, Trap ezoidal Domain by B il inear Map
Order=3
Exact sol ution is a p ol ynomial of degree 7
g 4
W
E
o
Z
~2
Q ;
c
W
c;
o
...J O
-2
-Iog( h)
-SJ - 00 =0.0 I
-{)_ 00= 2.5
N
D=Dirichl el B C
N=Neumann B e
Figure 8: Case 2 with bilinear parametrization. Experimental convergence rates for cubic
elements
128 el ements, order=3 ,
exact sol ution is a p ol ynomial of order 7
v-comp onent of E-fiel d
0) = 2.5
trap ezoidal domain
( b'f bl l Inear map }
D=Dirichl el B C
N=Neumann B e
Curl of E-fiel d
0.<199908
0430063
0.362017
0.293072
0.22"'26
0.155181
00862353
0.G172898
-4.0516557
-0.120601
-0.189547
-0 2584g2
..{}.3 274 3 8
-0.3116383
-0.485320
6.60672
5.6521
.,6Q 74 7
3.74285
2.78823
"'336
0878976
.{) .07564 82
-1.03027
-1.9649
-2.93952
-3.89415
....."
-5.8034
-6.75802
Figure 9: Case 2 with bilinear parametrization. E-field
27
128 el ements, order=3 ,
exact sol ution is 8p ol ynomial of order 7
Divergence of E-fiel d
( J ) =2.5
trapezoidal domain
(by bli06lll' map )
D;Dirichl el B e
N=Neumann B e
Lagrange mul tip l ier p
0.105167
0.0924021
0.0796369
0.0668716
O.O~l064
0.0.413411
0.0285759
0,0158106
0.00304538
.0.00071987
.. ~ ~ :.=
-0.0.480156
-0.0607800
.( ) ,073 S4 61
0.0002Q 64 28
0.000256392
0.000216356
0.00017632
0.000136284
9.624E5
5621E-5
1.617E-5
-2.3 85E-5
-8.389E-5
-0.000103931
-0.00014396]
-0.00018-4003
-0000224039
-o,Cl Q 0264 015
Figure 10: Case 2 with bilinear parametrization.
2 el ements, order=2, x-comp onent of E-fiel d,
exact sol ution is quintic
00=2.5 rectangul ar domain
N
Perturbation of the central midside node al ong the X-axis:
O.B
0.414286
0.0285716
-0.357143
-0.742857
1.12857
-1.51429
-1.9
-2.28571
-2.67143
3.05714
3.44286
382657
-4.21429
....
=2.0
ol=1.0 IN
0= 0.0 0""""''' B e
N""'l l UmamB C -20% +4 0% original p osition
Figure 11: Case 1.Sensitivity of solution to a mesh distortion
28
a sequence of positive eigenvalues. We emphasize that more than 50 % of the eigenvectors
correspond to the zero eigenvalue. When the mesh is distorted, the multiple zero eigenvalue
spreads into a cloud of non-zero eigenvalues whose size depends upon the distorsion. For
instance, for the experiment discussed above, a distortion of only 5 % is sufficient to spread
that cloud beyond the first (physically meaningful) positive eigenvalue. These numerical
eigenvalues interact with the forcing frequency wand we simply run into a numerical reso-
nance, exciting one or more corresponding modes. It is these excited modes that drastically
change the solution and prohibit the convergence in the L
2
-nor m. We can proceed with ex-
actly the same interpretation for the mixed formulation, except that, in that case, the entire
analysis takes place in the subspace of Wh consisting of the approximate E-fields which
satisfy the divergence condition. If the number of the imposed scalar constraints equals ex-
actly the multiplicity of the zero eigenvalue (with no constraints imposed), the corresponding
null space is eliminated and the spectrum consists only of the positive eigenvalues. This is
precisely guaranteed by the compatibility condition (4.18). In such a case, when we disturb
the mesh, there is simply no zero eigenvalue to spread and we do not run into the instability
problems.
6 Conclusions
In the paper a new formulation for the steady-state form of Maxwell's equations has been
proposed. The formulation involves the electric field E and a Lagrange multiplier p even
though, on the continuous level or for affine elements approximation, the multiplier p =O.
A novel, variable order of approximation based on the mixed formulation is proposed and
proved to be stable with a stability constant bounded away from zero for w ---t O.
A perturbation argument and numerical experiments indicate that the method remains
stable for parametric elements.
The proposed methodology is independent of the dimension and lays down foundations for
a general class of both two- and three-dimensional hp-discretizations of Maxwell's equations.
Acknowledgment
The authors would like to thank Professors I. Babuska, D. Neikirk, and J.T. Oden for
many interesting discussions on the subject. The work of the first author has been partially
supported by the NSF under Contract DMS 9414480, and the support for the second author
was provided from DOD-Air Force Grant F49620-96-1-0032 (P.I. - D. Neikirk).
29
References
[1] 1. Babuska and B. Q. Guo, "Approximation Properties of the hp Version of the Finite
Element Method", Computer Methods in Applied Mechanics and Engineering) Special
Issue on p and hp- Methods, eds. I Babuska and J. T. Oden, 133,319-346, 1996.
[2] A.C.Cangellaris, J.L.Prince, and 1.P.Vakanas, "Frequency-Dependent Inductance and
Resistance Calculation for Three-Dimensional Structures in High-Speed Interconnect
Systems", IEEE Transactions on Components) Hybrids) and Manufacturing Technology,
13, p.154, 1990.
[3] P.G. CiarIet, The Finite Element Method for Elliptic P roblems, North Holland, New
York, 1978.
[4] L. Demkowicz, J. T. Oden, W. Rachowicz, and O. Hardy, "Toward a Universal hp Adap-
tive Finite Element Strategy. Part1: Constrained Approximation and Data Structure",
Computer Methods in Applied Mechanics and Engineering 77, 1-2, 79-112, 1989.
[5] L. Demkowicz, A. Karafiat, and J. T. Oden, "Solution of Elastic Scattering Problems in
Linear Acoustics Using hp Boundary Element Method", Computer Methods in Applied
Mechanics and Engineering,lOl, 251-282 (Proceedings of Second Workshop on Relia-
bilityand Adaptive Methods in Computational Mechanics, Cracow, October 1991, eds.
L.Demkowicz, J.T.Oden and 1.Babuska).
[6] 1. Demkowicz, A. Bajer, and K. BanaS, "Geometrical Modeling Package", TICOM
R EP OR T 92-06, August 1992, The University of Texas at Austin, Austin TX 78712.
[7] 1. Demkowicz, "Asymptotic Convergence in Finite and Boundary Element Methods:
Part 1: Theoretical Results", Computers and Mathematics with Applications, 27, 12,
69-84, 1994.
[8] L. Demkowicz, "A Modified FE Assembling Procedure with Applications to Electromag-
netics, Acoustics and hp-Adaptivity", Mathematics of Finite Elements with Application,
ed. J. R. Whiteman, Jon Wiley and Sonsm Ltd., in print.
[9] Om P. Gandhi and Jin- Yuan Chen, "Electromagnetic Absorption in the Human Head
from Experimental 6-GHz Handheld Transceivers", IEEE Transactions on Electromag-
netic Compatibility, 37,4, 547-558, 1995.
[10] V. Girault, and P.A. Raviart, Finite Element Methods for Navier-Stok es Euations,
Springer- Verlag, Berlin, 1986.
3 0
[11] C. Hazard, and S. Lohrengel, "How to Use Standard Lagrange Finite Elements for
Solving Maxwell's Equations", MAFELAP 1996.
[12] D.J. Hoppe, L.W. Epp, and J.F. Lee, "A Hybrid Symmetric FEM/MOM Formulation
Applied to Scattering by Inhomogeneous Bodies of Revolution", IEEE Transactions on
Antennas and P ropagation, 42, 6, 798-805, 1994.
[13] F. Ihlenburg and 1. Babuska, "Finite Element Solution to the Helmholtz Equation with
High Wave Number. Part I: The h- Version of the FEM", Institute for Physical Science
and Technology, Technical Note BN-1159, November 1993.
[14] B.N. Jiang, J. Wu, and L.A. Povinelli, "The Origin of Spurious Solutions in Computa-
tional Electromagnetics", Journal of Computational P hysics 125, 104-123, 1996.
[15] J. Jin, The Finite Element Method in Electromagnetics, John Wiley & Sons, Inc., New
York 1993.
[16] J.F. Lee, "Analysis of Passive Microwave Devices by Using Three-Dimensional Tangen-
tial Vector Finite Elements", International Journal of Numerical Modeling: Electronic
network s, Devices and Fields, 3, 235-246, 1990.
[17] J.F. Lee, D.K. Sun, and Z.J. Cendes, "Tangential Vector Finite Elements for Elec-
tromagnetic Field Computation", IEEE Transactions on Magnetics, 27, 5, 4032-4035,
1991.
[18] J.F. Lee, D.K. Sun, and Z.J. Cendes, "Full-Wave Analysis of Dielectric Waveguides
Using Tangential Vector Finite Elements", IEEE Transactions on Microwave Theory
and Techniques, 39, 8, 1991.
[19] J.F. Lee, R. Palandech, and R. Mittra, "Modeling Three-Dimensional Discontinuities in
Waveguides Using Nonorthogonal FDTD Algorithm", IEEE Transcations on Microwave
Theory and techniques, 40, 2, 346-352, 1992.
[20] J.F. Lee, G.M. Wilkins, and R. Mittra, "Finite-Element Analysis of Axisymmetric Cav-
ity resonator Using a Hybrid Edge Element Technique", IEEE Transactions on Mi-
crowave Theory and Techniques, 41, 11, 1993.
[21] P. Monk, "A Finite Element Method for Approximating the Time-Harmonic Maxwell
Equations", Numerische mathematik , 63, 243-261, 1992.
[22] P. Monk, "Analysis of a Finite Element method for Maxwell's Equations", SIAM Jour-
nal of Numerical Analysis, 29, 3, 714-729, 1992.
31
[23] P. Monk, "An Analysis of Nedelec's Method for the Spatial Discretization of Maxwell's
Equations", Journal of Computational and Applied Mathematics, 47, 101-121, 1993.
[24] P. Monk, "On the p- and hp-extension of Nedec's curl-conforming elements", Journal
of Computational and Applied Mathematics, 53, 117-137, 1994.
[25] P. Monk, "Superconvergence of Finite Element Approximations to Maxwell's Equa-
tions", Numerical Methods for P artial Differential Equations, 10, 793-812, 1994.
[26] G. Mur, "Finite Element Methods for Electromagnetic Field Computations", MAFE-
LAP 1996.
[27] J.C. Nedelec, "Computation of Eddy Currents on a Surface in JR 3 by Finite Element
Methods", SIAM Journal of Numerical Analysis, 15,3, 580-594, 1978.
[28] J.C. Nedelec, "Mixed Finite Elements in JR 3" , Numerische Mathematik , 35, 315-341,
1980.
[29] J.C. Nedelec, "A New Family of Mixed Finite Elements inJR
3
", Numerische Mathematik ,
50, 57-81, 1986.
[30] J.C. Nedelec, "Finite Elements for Exterior Problems Using Integral Equations", Inter-
national Journal for Numerical Methods in Fluids, 7, 1229-1234, 1987.
[31] J. T. Oden, 1. Demkowicz, W. Rachowicz and T. A. Westermann, "Towards a Universal
h-p Adaptive Finite Element Strategy, Part 2. A P osteriori Error Estimation," Compo
Meth. Appl. Meth. Eng., 77, 113-180, 1989.
[32] J.T. Oden, and L.F. Demkowicz, Applied Functional Analysis for Science and Engineer-
ing, CRC Press, Boca Raton, 1996.
[33] Ch.H. Papas, Theory of Electromagnetic Wave P ropagation, Dover Publications, New
York 1988.
[34] C.R. Paul and S.A. Nasar, Introduction to Electromagnetic Fields, McGraw-Hill Book
Company, New York 1982.
[35] P.P. Silvester, and G. Pelosi (eds.), Finite Elements for Wave Electromagnetics, IEEE
Press, New York, 1994.
[36] E.Tuncer and D.P.Neikirk, "Highly Accurate Quasi-Static Modeling of Microstrip Lines
Over Lossy Substrates", IEEE Microwave and Guided Wave Letters, 2, 409-411, 1992.
3 2
[37] E.Tuncer, B.-T.Lee, M.S.Islam, and D.P.Neikirk, "Quasi-Static Conductor Loss Cal-
culations in Transmission Lines Using a New Conformal Mapping Technique", IEEE
Transactions on Microwave Theory and Techniques, 42, 1807-1815, 1994.
33

You might also like