Finite difference time domain methods

S. Gonzalez Garca, A. Rubio Bretones, B. Garca Olmedo &

R. Gomez Martn
Dept. of Electromagnetism and Matter Physics
University of Granada (Spain)

Abstract
In this chapter the fundamentals of the Finite Difference Time Domain (FDTD)
method to solve Maxwells curl equations in the time domain are given in a con-
cise operational form. The Perfectly Matched Layer truncation techniques are ex-
plained, together with the connection between the split and the Maxwellian formu-
lation, both for the EH and for the materialindependent DB formulation. At-
tention is later paid to the still under development unconditionally stable schemes,
especially the ADIFDTD, which is broadening the computational efficiency of
the method, since the time increment is no longer restricted by the Courant stabil-
ity criterion as is in the classical FDTD case. Then, the material implementation in
FDTD is studied, describing in some detail the simulation of Debye media with the
FDTDPML and with the ADIFDTD. Finally some references of the geometrical
modeling are summarized.

1 Introduction

During the past 25 years the Finite Difference Time Domain (FDTD) method has
become the most widely used simulation tool of electromagnetic phenomena. The
FDTD is characterized by the solution of Maxwells curl equations in the time do-
main after replacement of the derivatives in them by finite differences. It has been
applied to many problems of propagation, radiation and scattering of electromag-
netic waves. The method owes its success to the power and simplicity it provides.
A good measure of this success is the fact that some five thousand papers on the
subject have been published in journals and international symposia, apart from
the books and tutorials devoted to it. In addition, more than a dozen specific and
general purpose commercial simulators are available on the market.
 Maxwells curl equations predict the existence of an electric and a magnetic
field, whose primary causes are charges and movement of charges, coupled be-
tween them so that they selfmaintain even in the absence of the charges/currents

~ = , B
D ~=0 (1)

E~ = t B
~ , H
~ = J~ + t D
~ (2)

These equations relate the volume charge density , the volume current density
J~ , and four vector field magnitudes: D ~ displacement vector, E~ electric field vector,
~ ~
B magnetic flux density vector, H magnetic field vector, all of which are functions
of the space and time (~r, t). The displacement vector D ~ is related to the electric
field E~ by constitutive relationships depending on the material properties, and sim-
ilarly the magnetic flux density B ~ is related to the magnetic field vector H ~ (for a
linear isotropic nondispersive medium D ~ = E,
~ B~ = H,
~ with and being the
medium permittivity and permeability respectively)
The time dependent Maxwells curl equations (2) together with the constitutive
relationships, and assuming that the fields are known at a given time instant (initial
value problem), are a system of hyperbolic partial differential equations which
has a unique solution depending continuously on the initial data (i.e., it is well
posed). The application of finite differences to the resolution of partial differential
equations dates from the works of Courant, Friedrichs, Lewy, etc., almost a century
ago [1], and it has grown over the years in close relation to the calculating power of
computers. Finite differences were first applied to Maxwells curl equations in the
work of Kane S. Yee in 1966 [2]. Since then, the method has been developed, and
refined in all its theoretical and computational aspects, and nowadays is probably
the most powerful and versatile electromagnetic simulation tool available.
Although originally the term FDTD was coined for the method developed by
Yee, today it is used to describe the set of numerical techniques that use finite
differences to discretise all or part of the derivatives in Maxwells curl equations,
and that give rise to a marchingonintime solution algorithm.
Many excellent monographic books are devoted to FDTD [3, 4, 5, 6, 7], and the
web site http://www.fdtd.org contains an uptodate searchable and downloadable
FDTD literature database maintained by Schneider and Shlager. It is not our inten-
tion in this chapter to fully describe the FDTD and its applications, for which we
refer the reader to the above mentioned books and web site. Neither is an exten-
sive bibliography referenced since this can easily be found in the above sources.
We will only include references to representative publications on each of the topics
covered, and do not claim to be exhaustive or even chronological in referring to
the contributions to a particular topic.
After a brief stateofthe art introduction, in the first part of this chapter we
give a general description of the method using an operational language. After in-
vestigating the dispersion, stability, and other numerical details of the method,
we examine the perfectly matched layer Absorbing Boundary Conditions (ABCs).
We then focus on the recently appeared unconditional stability methods (mainly
the alternating direction implicit method). We end our overview by summarizing
material implementation in FDTD and some geometrical modeling techniques.

2 State of the art

The classical FDTD method employs a second order finite centered approxima-
tion to the space and time derivatives in Maxwells curl equations, giving rise to
to a new discrete electromagnetism with its own inherent peculiarities [8][9]. The
FDTD defines an orthogonal cubic spatial grid whose unit cell is shown in figure 1;
each field component is sampled and evaluated at a particular space position, and
the magnetic and electric fields are obtained at time instants delayed by half the
sampling time step. The materials are modeled by specifying their characteristic
constants at every grid point, usually in homogeneous regions, at whose interfaces
proper continuity conditions are needed. The time advancing algorithm is explicit,
that is, the fields at each time instant are obtained as a function of previous val-
ues in time. The simulation of open problems is carried out by placing ABCs in
the terminating planes of the grid. The method is fairly CPU and memory inten-
sive, but current computing technology and foreseeable advances provide it with
extraordinary power.

Z
HZ EY
x

EX
EX
EY EZ

EZ HY
EZ
z

HX

EY EX

Figure 1: Yee unit cell

 From the start, development of the original method has been aimed at enhancing
its characteristics, either modifying the original algorithm, or using it in conjunc-
tion or hybridly with other techniques. Many papers have dealt with key aspects of
the method and provided enhanced solutions (dispersion, material modeling, geo-
metrical modeling, etc.), while others describe totally new concepts which provide
the method with a new dimension (perfectly matched layer ABCs, unconditionally
stable techniques, etc.). Some of the most interesting developments are described
below.
The error in the solution to a given problem depends on the algorithm employed
to simulate each part of the problem, especially the discretisation of derivatives,
geometrical and material modeling, and truncation conditions (ABCs). Second
order approximation to derivatives implies a significant anisotropic phase error,
although this may be negligible for electrically small problems. Nevertheless, to-
days greater computing power allows us to deal with electrically large problems
for which specific lowphase techniques should be employed. The most classi-
cal of these is the use of higher order approximations to derivatives [10], which
reduces both dispersion and anisotropy, but many other FDTDlike techniques
maintaining the marchingonintime scheme by using finite differences for the
time derivatives, and using a different treatment for the spatial ones, are able to
reduce the phase errors as well. For instance, the Pseudo Spectral Time Domain
method (PSTD) [11] which obtains the space derivatives in the spectral space do-
main at each time step using the Fast Fourier Transform and its inverse (used in
conjunction with the PML to prevent the wraparound effect). Another example is
that of multiresolution techniques time domain techniques (MRTD) [12], which
are also close to the FDTD method except that they obtain the spatial derivatives
by means of scaling and wavelet functions. Currently promising lines of research
aimed at low error schemes are based on the nonstandard FDTD method (NS
FDTD) of Mickens [13] which is characterized by searching for zero truncation
error (exact) schemes [14], or those based on the linear bicharacteristic schemes
(LBS) [15].
On the other hand, the standard advancing algorithm is conditionally stable. The
Courant criterion imposes an upper limit to the time increment related to the space
increment employed in the discretisation. This can lead to high CPU times when
a fine space discretisation implies the use of unnecessarily small time increments.
New unconditionally stable techniques like the ADIFDTD method [16][17], par-
tially overcome the problem and have recently initiated an interesting research
line.
Although Mur absorbing boundary conditions [18] for open problems were the
most commonly used for over a decade, the introduction of the Perfectly Matched
Layer method by Berenger in 1994 [19] represented a major advance in this field.
Since then many publications have appeared in which they are studied, reformu-
lated and enhanced. If the problem under simulation is periodic (for instance pho-
tonic bandgap, frequency selective surfaces, antenna arrays, etc.), specific bound-
ary conditions (e.g. Floquet) are suitable to truncate it [20].
 The discrete nature of the space time grid, and particularly the cubic stag-
gered structure of Yees grid, introduce difficulties into the geometrical model-
ing of the materials, which lead to errors and even instabilities. Other geometrical
aspects presenting difficulties are subcell details like wires, sheets, slots, edges
and corners. Partial solutions are local or global finite integration/volume tech-
niques [21][22], general curvilinear coordinates (orthogonal or not) [23], etc. all
of which are attempts to conform the material geometry. Some simple variations of
the homogeneous Yee grid (uneven meshes, subgridding) have also been proposed
to help with this problem. A major concern with all these methods is their stability,
for which the recentlyannounced unconditionally stable algorithms will probably
provide solutions. A large part of the bibliography is devoted to seeking accurate
solutions to these problems.
Other studies carried out closely related to FDTD include: the neartofar field
conversion [3][24] mainly in the field of Radar Cross Section calculations, the sim-
ulation of sources internal or external to the computation domain, which may be
lumped (for circuital structures) or wave source conditions (to simulate scattering
problems) [25], etc.
Finally, let us briefly mention the numerous efforts made to incorporate mate-
rial behavior into the method, and the corresponding truncation conditions. Non
linear, dispersive, gyromagnetic, anisotropic, biisotropic, bianisotropic, active,
metamaterial, random, periodic, photonic bandgap, are just some examples of the
material structures that can be treated with the FDTD method.
Numerous applications of the method have been published. One of the first was
the calculation of the radar cross section (RCS) of arbitrary scatterers in the mi-
crowave range. Since then, the frequency range of the applications has grown to
extend from low frequencies up to optical ones. The method has been used for cir-
cuit simulations, both lumped and distributed, and all kinds of waveguide propaga-
tion: conducting, dielectric, microstrip, coplanar, etc. It has been successfully used
to simulate all kinds of antennas and arrays: wire, planar, microstrip, horn, etc. Its
use in electromagnetic compatibility is also wide: shielding, coupling, natural or
artificial (nuclear electromagnetic) pulse effects, etc. Its application to bioelectro-
magnetic problems such as hyperthermia, cancer detection, biological effects of
electromagnetic waves, and specific absorption rate (SAR) calculations, is also a
growing field of development, involving in many cases the coupling of Maxwells
equations to heat and transport equations.
Let us finish by summarizing some of the actual trends on FDTD methods.
There is a constant inflow/outflow of new ideas with every other area involving
the solution of partial differential equations: acoustic, computation fluid dynamics
(CFD), quantum physics, etc. A common trend in all these disciplines is the search
for more exact algorithms. The simulation of electrically large problems (like the
ones in photonics) needs low dispersion error schemes. A desirable quality of any
new numerical scheme is its unconditional stability so that the time discretisation
can be chosen independently of the space discretisation, thus allowing them, for
instance, to be applied to low frequency problems, or to problems meshed with
structured or unstructured highly inhomogeneous grids, etc. The obtention of al-
gorithms, based on the integral or on the differential form, that are capable of char-
acterizing the geometrical inhomogeneities of materials, continues to be a field of
active research. Finally, there is growing interest in engineered materials which
exhibit exotic electromagnetic properties, for which accurate simulation tools are
needed. A great effort is to be dedicated to extend the new algorithms to deal with
these materials.

3 Finite numerical schemes: an operational approach

In this section we describe the fundamentals of the FDTD method. Since the fol-
lowing chapter deals with the implementation of free currents, we assume hence-
forth that all the currents are ohmic, both electric J~ = E,
~ and magnetic J~ =
~
H (these last included for reasons of symmetry), with and being the electric
and magnetic conductivity respectively. At this stage we also assume that all the
media are linear, isotropic and nondispersive D~ = E,~ B
~ = H,
~ with and the
medium permittivity and permeability respectively, and so we omit the displace-
ment and magnetic flux density vector fields, and set Maxwells curl equations in
the so called EH form, which in cartesian coordinates can be written as
r E~ = H ~ , r H
~ + t H ~ = E~ + t E~ (3)
with
0 z y

r = z 0 x (4)
y x 0
and in compact form as
~ + S
t
M ~ =R
~ (5)
!
~ = (E,
~ H)
~ T , R
=
03 r

r 03
! !
I3 03 I3 03
S = =
, M (6)
03 I3 03 I3

where henceforth In , 0n denotes the nth order identity and null matrices, respec-
tively, and the superindex T the usual matrix transpose.
To write the FDTD equations, let us first define the centered difference and
average operators,
v v v v
f (v + 2 )
f (v 2 ) f (v + 2 ) + f (v 2 )
v f (v) = , av f (v) = (7)
v 2
A Taylor series expansion proves that v f (v) is a second order approximation
to v f (v), and av f (v) is a second order approximation to the identity opera-
(v) = f (v). For instance, if f (v) is a one variable harmonic function
tion If
f (v) = A cos(2/v v), the errors of the above defined operators with respect
to the analytical ones are

v f (v) v f (v) 2 1
= =
(8)
v f (v) 6 rv2

f (v) av f (v)

a = = 3 (9)
f (v)

v
with rv being the number of samples per oscillation in v (rv = v ).
Let us discretise the space into a set of points uniformly spaced in each direction,
integer and semiinteger multiples of x, y, z, and similarly the time into
instants, integer and semiinteger multiples of a given interval t, and let us define
the fields at each spatialtemporal position of this grid with the following notation

n
i,j,k (x = ix, y = jy, z = kz, t = nt) (10)

The classical FDTD Yee numerical scheme is written, replacing the derivatives
in (3) by the centered difference operator and averaging the conductivity terms in
time

r H
~ i,j,k
n ~ i,j,k
= i,j,k at E n ~ i,j,k
+ i,j,k t E n

r E
~ i,j,k
n
= i,j,k ~ i,j,k
at E n ~ i,j,k
+ i,j,k t H n
(11)

where the numerical curl operator which approximates (4) is denoted by

0 z y

r = z 0 x (12)
y x 0

The heterogeneity of the media is handled by sampling the analytical constitutive

parameters , , , at each point of the spatial grid

i,j,k (x = ix, y = iy, z = iz) . . . (13)

Henceforth, cursive fonts are used for the numerical fields E, B, D, H, J to dis-
tinguish them from the analytical ones, for which a calligraphic font E, B, D, H, J
is used.
It can easily be appreciated that (11) does not involve all six components of the
electromagnetic field, defined at every space point of the spatial grid sampled with
x/2, y/2, z/2. A grid six times less dense is enough with the fields relatively
situated at the positions given by the Yee unit cell (figure 1). Similarly it is enough
to define the electric fields at semiinteger multiples of the time increment t
and the magnetic fields at integer multiples, to obtain the marchingonintime
algorithm

~ n+1/2 = i,j,k i,j,k t/2 E

E ~ n1/2 + t
r H
~ i,j,k
n
i,j,k i,j,k
i,j,k + i,j,k t/2 i,j,k + i,j,k t/2

i,j,k i,j,k t/2 n t
~ n+1 =
H ~
H r E
~ n+1/2 (14)
i,j,k
i,j,k + i,j,k t/2 i,j,k i,j,k + i,j,k
t/2 i,j,k

with n integer, and i, j, k integer or semiinteger according to the position of the

field component at the Yee cell. This algorithm is selfmaintained, starting from
numerical initial conditions sampled from the analytical ones.
We can write (14) in compact form as

~ n+1/2 + Si,j,k at
i,j,k t
M ~ n+1/2 = R
~ n+1/2 (15)
i,j,k i,j,k i,j,k

with
!
~ n1/2 , H
~ ni,j,k = (E ~ i,j,k
n =
03 r
)T , R
i,j,k
r 03
! !
i,j,k I3 03 i,j,k I3 03
Si,j,k = i,j,k =
, M (16)
03
i,j,k I3 03 i,j,k I3

which is formally similar to the analytic form (5)

It is easy to build higher order schemes by using other approximations to the
derivatives [26], which together with different collocations of the fields at the space
grid, make it possible to manage, for instance, anisotropic media [27] with the EH
formulation. For instance, the operator

v 2
v v v v v (17)
24
is a fourth order approximation to the partial derivative with respect to v, with an
error for a harmonic function = 2.7 2 .
The FDTD scheme has been successfully reformulated in other orthogonal co-
ordinate systems, and even in general curvilinear coordinates [28][29].
It should be noted that, as later shown, although the EH formulation is not well
suited to deal systematically with arbitrary materials, it offers computational ad-
vantages over the materialindependent DB formulation for lossy media (isotropic
or anisotropic) with no other dispersive characteristics. 1

1 A main advantage of the matricial operator language here introduced, and henceforth used in the

whole chapter, is that is comprises a powerful and concise tool that furthermore permits the straight-
forward analysis, simplification, and construction of C/FORTRAN computer programs with the aid of
the symbolic and code generation capabilities of tools like MathematicaT M
4 Fundamental topics: convergence, consistency, stability and
dispersion
An obvious quality of a good numerical scheme is that it should approximate the
analytical equations when the space-time lattice is progressively refined x, y,
z, t 0. Such a scheme is said to be consistent with the analytical problem.
If we define the nlocal truncation
o error of a given numerical scheme written in
~ n
the generic form L i,j,k = 0, as
n o
T~i,j,k
n
=L ~n (18)
i,j,k

with ~ n representing the solution of the numerical scheme and ~ n the ana-
i,j,k i,j,k
lytical solution sampled at (x, y, z, t) = (ix, jy, kz, nt), we can formally
say that the scheme is consistent if [30]

lim T~i,j,k
n
=0 (19)
x,y,z,t0

For instance, for eqn. (15) the local truncation error is

T~i,j,k
n
= M i,j,k t + Si,j,k at R
~ n+1/2 (20)
i,j,k

~ n is the analytical field

where i,j,k

T
~n = ~ 1 ~
i,j,k E(ix, jy, kz, (n )t), H(ix, jy, kz, nt) (21)
2

sampled at lattice space and time positions.

A multidimensional Taylor series leads us to

T~i,j,k
n ~ n00 0 0 , n n0 n + 1
= F (x2 x3 , y 2 y3 , z 2 z3 , t2 t3 ) i ,j ,k
1 1 1 1 1 1
i i0 i + , j j 0 j + , k k 0 k + (22)
2 2 2 2 2 2

with F being a matrix operator involving a linear combination of its arguments.

Thus, the Yee scheme is unconditionally consistent with Maxwells curl equations
up to second order, with truncation error terms in (22) written in the frequency and
spectralspace domain as 2 2 1/r2 with r being the number of samples
per oscillation ( = {x, y, z, t}).
The consistency of a numerical scheme, in any case, does not guarantee that its
solution actually approximates the analytical solution. A numerical scheme is said
to be convergent [30][31] if its solution tends to the analytical one when the space-
time lattice is progressively refined at a given spacetime point (x = ix, y =
jy,z = kz,t = nt), that is, when x, y, z, t 0, and i, j, k, n .
 Finally, for a wellposed analytical initial value problem, given a numerical
scheme replacing it, the difference of the numerical solutions for arbitrarily boun-
ded differences in the initial values must be bounded in the same manner, as are the
analytical ones. Such a scheme is said to be stable. If the solution of the analytical
initial value problem is naturally bounded in time (for instance, Maxwells curl
equations for passive media), then any stable scheme replacing it must necessarily
keep its solution bounded in time as well.
The Lax equivalence theorem [31] states that given a wellposed initial-value
problem and a consistent linear finite-difference numerical scheme, stability is the
necessary and sufficient condition for convergence. The utility of this theorem is
obvious since it is usually easier to find stability criteria rather than convergence
ones, and we need only to guarantee then stability of a consistent scheme to con-
clude that the numerical solution actually approximates the right solution.

4.1 Dispersion relation: Von-Neumann stability criterion

Plane waves propagating in lossless nondispersive isotropic linear homogeneous

media have a phase speed independent of the frequency and the propagation direc-
tion. Let us study how the numerical scheme propagates such plane waves. We can
write the Yee scheme (15) for such media as

~ n+1/2 ~ n+1/2 , R
T T = (M
i,j,k )1 R
t i,j,k = R i,j,k (23)

with known initial values that can be expanded in Fourier series in generic terms

~ 0 e j (ix x jy y k z z )
(24)

with ~ = (x , y , z ) the real wavevector. Let us assume that this harmonic, which
is analytically propagated harmonically in time with frequency fulfilling the dis-
persion relation 2 = x2 + y2 + z2 , is also propagated harmonically by the
scheme with a complex frequency

~ ni,j,k =
~ 0 e j (nt ix x jy y k z z ) (25)

Inserting (25) into the numerical scheme (23) the following eigenvalue problem
is obtained
t I6
R ~ n+1/2 = 0 (26)
T i,j,k

with

! 0 z y
06
1/
R =
r =
, z 0

x
T

1/ 06
r
r
y x 0

2j t 2j
t = sin , = sin , = {x, y, z} (27)
t 2 2

Imposing a non trivial solution (null determinant of the coefficient matrix), leads
to

c2 2t = 2x + 2y + 2z , c = 1/ (28)
Explicitly

1 2 t
sin =
c2 t2 2

1 2 x 1 2 y 1 2 z
sin x + sin y + sin z (29)
x2 2 y 2 2 z 2 2

which is seen to approximate the analytic dispersion relation in the limit of null
increments. The phase speed vf = is shown to depend both on frequency (dis-
persivity), and on the propagation direction (anisotropy) unlike the analytical phase
speed (see [4] for further details).
Furthermore the structure relationship between the direction of propagation, the
electric field and the magnetic field (orthogonal in the analytic case) is no longer
valid. By extracting the time eigenvalue from (28) and introducing it into (26) we
obtain for the amplitudes

~0 = 1 ~ ~n E
~0 , ~0 = 0
H n E (30)
n

~ sin (x x/2) sin (y y/2) sin (y y/2) sin (t/2)
n = , , , n =
x/2 y/2 z/2 t/2
(31)
where it can be seen that the electric field and the magnetic field are mutually
perpendicular; however they are not perpendicular to the analytical wavevector
~ = (x , y , z ) but to a numericallike wave-vector defined by (31).

As previously stated, a necessary condition of stability is that the solution of the

scheme must remain bounded in time; in this case

jt
e 1 Im {} 0 (32)

From (29), condition (32) holds ~ if

s
1 2 x 1 2 y 1 2 z
ct sin x + sin y + sin z 1
x2 2 y 2 2 z 2 2
(33)
which always holds if the Courant number s next defined fulfills
r
1 1 1
s ct 2
+ 2
+ 1 (34)
x y z 2


which implies Im {} = 0 and e jt = 1. When x = y = z , the

condition on the Courant number becomes s = 3 ct 1. This condition, prop-
erly known as the VonNeumann stability criterion, coincides with the Courant-
Friedrichs-Lewy (CFL) condition (also known as the causality condition) obtained
through geometrical considerations on the characteristics of the hyperbolic differ-
ential system [30]. We can conclude that there is an upper limit to choose the time
increment, after the space increments are chosen, which may lead to unnecessarily
small increments for certain problems.
If the necessary condition (34) does not hold, there appear both increasing solu-
tions (unstable) and decreasing ones (dissipative) propagating at the Nyquist fre-
quency

t 2 p
If sin( )=g>1 = j ln g + g 2 1 (35)
2 t t
It can be demonstrated that condition (32) is also sufficient for the homogeneous
initial value problem. Nevertheless, in the presence of inhomogeneities (material
changes, space increments depending on the space position, boundary conditions,
etc.), or, in general, when the problem becomes an initial boundary value prob-
lem (any computational implementation must necessarily incorporate boundary
conditions), condition (32) may not be sufficient but only necessary to guarantee
stability. In these cases we can use the Gustaffson-Kreiss-Sundstrom
theory [32]

which pays special attention to the critical case e jt = 1, showing that there
can appear instabilities depending on the specific boundary conditions (see [33]
for a physical reinterpretation of this theory in terms of group speed).

4.2 Spectral stability criteria

It is possible to find an analytic formulation based on spectral analysis permitting

us to investigate, at least numerically, the stability of problems solved in a closed
region with appropriate boundary conditions (initial boundary value problems).
Let us state the twolevel scheme (23) expressed by

~ n+1 =
~ ni,j,k + tR ~ n+1/2
T (36)
i,j,k i,j,k

as a singlelevel scheme
! ! n+1/2 !
~ n+1
t R T I6 ~

i,j,k
= i,j,k ~ n+1 = G
~ n+1/2
~ n+1/2 6 06 ~n i,j,k i,j,k
i,j,k I i,j,k
| {z } | {z }| {z }
~ n+1

G ~ n+1/2

i,j,k i,j,k
(37)
holding for every
n pointo inside the closed region, plus the specific conditions at the
boundaries F ~n
i,j,k = 0. It is possible to reorder the unknowns ~ n defined
i,j,k
B
at all the positions of the threedimensional space grid into a onedimensional
column matrix ~ n and so (37) can be rewritten as

~ n+1 = G
~ n+1/2 + ~b n+1/2 (38)

with G being the scheme amplification matrix and ~bn the column matrix with the
boundary conditions. It can be proven that a necessary condition for stability is that
the spectral radius of the amplification matrix must fulfil (G) 1. The result is
~ 0 are propagated
clearly justified because (38) implies that the initial conditions
by increasing powers of the amplification matrix.
2(n+1) 0
~ n+1 G
~
(39)

and the spectral radius is a lower bound of any compatible matrix norm [31]. Al-
though this criterion may seem difficult to put into practice, since it is cumber-
some to find the spectral radius for a given symbolic problem analytically, it can
be systematically applied to particular numerical benchmark cases because the
sparse nature of the amplification matrix allows us to use rapid routines (for in-
stance LAPACK) to obtain its spectral radius. The main utility of this technique
is to test inhomogeneous schemes for instabilities (uneven meshes, subgridding,
locally distorted conformal schemes, etc.) where the VonNeumann technique is
not applicable.

5 Truncation conditions: Perfectly Matched Layer (PML)

The FDTD method permits the solution of open problems in computationally

bounded domains by employing suitable truncation conditions. These simulate a
frontier which ideally is transparent (null reflection coefficient) irrespective of the
frequency, polarization and angle of incidence.
The Mur absorbing boundary conditions [18] for open problems were the most
commonly used for more than a decade. Although they offered reflection coeffi-
cients of up to -40 dB, this was not enough for many interesting problems. Many
alternatives have been developed since then [34] to reduce the reflection coefficient
up to -100 dB. But in 1994 Berenger with his Perfectly Matched Layer method [19]
provided a non Maxwellian analytic technique achieving up to -140 dB. Its main
characteristic was its ease of implementation in working FDTD codes, and its ap-
parent simplicity. Further works to obtain PML Maxwellian formulations has since
been carried out [35][36], and are briefly described at the end of this section.
The fundamental idea of the original PML technique is the truncation of one
Maxwellian medium, by means of a non Maxwellian one, which is perfectly ab-
sorbing for all frequencies, polarizations and angles of incidence, and able to to-
tally attenuate, from a practical point of view, the field propagated in its interior,
after the propagation of just a few cells. This absorbing layer can then be termi-
nated by a perfect electrically/magnetically conducting (PEC/PMC) wall, so that
it can be implemented without requiring high computational resources. Although
such a medium is theoretically unphysical, it is possible to build a mathematical
set of Maxwelllike equations on which to enforce such conditions.
Initially the term PML was related only to the work of Berenger, but it has since
become a common term for the set of techniques, based or not on Berengers initial
equations, with the common characteristic of being able to truncate the computa-
tional space by an absorbing transparent layer.
Berengers main idea was to construct an electric and a magnetic pseudofield,
by splitting each cartesian component of the Maxwellian field into two subcompo-
nents, and relating each time derivative of them with only one part of the spatial
derivatives appearing in the curl equation of the Maxwellian component, so that
the addition of the equations fulfilled by both subcomponents would recover the
original curl equation.
By now introducing into the equation of each subcomponent the possibility of
anisotropic electric and magnetic conductivities, a new set of equations is obtained,
with enough degrees of freedom to achieve the objective that every solution of
Maxwells curl equations should also be a solution of Berengers equations on an
arbitrary surface separating the solution domains of each set of equations (continu-
ity of both solutions). Thus any wave propagating in the Maxwellian medium will
be transmitted into the Berenger domain without reflection, where it will propagate
with attenuation due to the conductivity terms.
Although this interpretation of Berengers PML, which we will develop below,
differs from the original one [19], we find it very general and relatively simple to
understand.
Let us assume a sourcefree medium inside which there are plane waves propa-
gating with the form
~ r, t) = ~ r)
~0 e j ( t~
(~ (40)
~ related by the medium dispersion relation, for which Maxwells curl
with and
equations expressed in materialindependent form are

j E~ = j B
~ , j H
~ = j D
~ (41)

with
0 z y

= z 0 x (42)
y x 0
representing the curllike wavenumber matrix. And let us assume a general rela-
tionship between the fields in the frequency domain

~ = () E~ , B
D ~= ~
() H (43)

with (),
() already including the electric and magnetic conductivity losses
and dispersivity, as well as anisotropy.
 Let us assume that we truncate this Maxwellian medium by another one, hence-
forth called PML, in which we define a pseudofield splitting each field cartesian

component ~ PML = PML , PML , PML T into two subcomponents
x y z

~ PML = ( PML , PML , PML , PML , PML , PML )T , = {E, H, D, B}

(44)
s xy xz yz yx zx zy
| {z } | {z } | {z }
PML
x yPML zPML

which recover the original threecomponent field when added

1 1 0 0 0 0
~ PML , C =
~ PML = C 0 0 1 1 0 0

(45)
s
0 0 0 0 1 1

Let us force the split field to fulfil equations somehow similar to (41) but includ-
ing anisotropic lossy terms through conductivity matrices s and s

s C E~sPML = j B
j ~sPML + ~sPML , j
s B s C H
~ sPML = j D
~ sPML + ~ sPML
s D (46)
with s a split curllike wavenumber matrix

0 0 y

0 z 0

0 0
z
s = (47)
0 0 x

0 x 0

y 0 0

and with general frequency domain constitutive relationships of the form

~ sPML = s ()E~PML , B
D ~sPML = ~ sPML
s ()H (48)
Eqn. (46) is satisfied by plane waves of the form
~s (~r, t) =
~s0 e j ( t~ ~r) (49)
with and related by the PML dispersion relation.
If a plane wave (40) propagating inside the original medium is incident on a
plane interface with the PML medium and a transmitted wave (49) begins to prop-
agate inside it, no reflection will occur at the interface, whatever the angle of inci-
dence, frequency and polarization, if there is continuity of all the components of
the fields in the original medium with those of the compact field in the PML. For
instance for the interface plane x = 0

E~ = E~PML , H ~ =H ~ PML , D ~ =D ~ PML , B ~=B ~ PML
x=0 x=0 x=0 x=0
(50)
These continuity conditions imply

~
Phase continuity e j ~r = e j ~ ~r y = y , z = z with x
x=0
arbitrarily chosen
Amplitude continuity D ~0 = D ~ PML , B~0 = B ~ PML , E~0 = E~PML , H ~0 =
0 0 0
~
H0PML

Let us obtain the relationship that must relate s ,

s , s ,
s to ,
for this to
occur. Extracting B ~ and D
~ from (41) and B ~s and D
PML ~ sPML from (46) and using (45),
the amplitude continuity is fulfilled if

y 0 0 0 0 0

0 z 0 0 0 0

0 0 z 0 0 0 j
1


s =
s = , (51)
0 0 0 x 0 0 = {x, y, z}

0 0 0 0 x 0

0 0 0 0 0 y

and
C s () = () C , C ()C
s () = (52)

On the other hand, the phase continuity condition at x = 0 implies y = y ,

z = z with which y = z = 0, freeing the choice of x which can be taken
as x = x jx with x an arbitrarily real coefficient which controls the atten-
uation rate inside the PML. For this choice x = x x is a real number. Fixing
this conductivity for a given problem, the attenuation rate x will be frequency
independent for frequency independent media (/x = vx with vx phase speed).
Nevertheless, for dispersive media it is necessary to use an optimum conductivity
value to obtain the desired attenuation rate for the worst frequency case. Finally,
the wave propagating into the PML takes the form

~s (~r, t) =
~s0 e j ( tx xy yz z) =
~s0 e j ( tx xy yz z) e vxx x (53)

It could even be possible to add extra degrees of freedom to (46), also permitting
the absorption of evanescent waves [37][38], although we will refer later to more
recent approaches.
Hence, the time domain PML equations matching the Maxwells curl equations

r E~ = t B
~ , r H
~ = t D
~ (54)

are

rs C E~sPML = t B
~sPML + ~sPML , rs C H
s B ~ sPML = t D
~ sPML + ~ sPML
s D (55)
with
0 0 y

0 z 0

z 0 0

rs = (56)
0 0 x

0 x 0

y 0 0
and with the constitutive relationship time domain versions of (43) and (48) taking
into account (52), which for nondispersive lossless isotropic media is fulfilled by
() = I6 and
() = I6 . For this last sort of media, and even for anisotropic
ones, an EH formulation could be used for computational efficiency, which would
be totally equivalent to the materialindependent form here described [39][40].
Nevertheless the materialindependent form shows its generality for disper-
sive media. For instance, for a Debye dielectric dispersive isotropic non-magnetic
medium
s
= + + = 0 I3
I3 , (57)
1 + j j
(43) can be written in the time domain in partial differential equation form (j
t )
t2 D ~ = t2 E~ + (s + ) t E~ + E~ , B
~ + t D ~ = 0 H
~ (58)
and (48) in a formally identical way
~ sPML + t D
t2 D ~ sPML = t2 E~sPML + (s + ) t E~sPML + E~sPML

B ~ sPML
~sPML = 0 H (59)

since
s
s = + + s = 0 I6
I6 , (60)
1 + j j
satisfies the condition (52).
A number of extensions to all kinds of media can be performed (for instance,
in [41] a PML to adapt bianisotropic media is obtained) with the same construction
principles followed here, i.e., imposing total continuity on the fields at the interface
between the two media.
Although we could add extra degrees of freedom to the above equations to ex-
tend the capabilities of the method, it is easier, for many cases of interest, to handle
its extension via a previous Maxwellian reformulation.
An equivalent Maxwellian reformulation of the method can readily be obtained
by leftmultiplying eqn. (55) in the frequency domain by C, and taking into ac-
count (45)
" 1 #
1
C I6 +
s rs E~PML = j B ~ PML
j
 " 1 #
1
C I6 +
s rs H~ PML = j D
~ PML (61)
j

which after defining a set of stretched spatial variables of the form

Z

0 = s ()d = s 0 with s = 1 + , = {x, y, z} (62)
0 j

can be written as

r0 E~PML = j B
~ PML , r0 H
~ PML = j D
~ PML (63)

with r0 being the usual curl operator (4) with derivatives with respect to the
stretched spatial variables.
We can rescale each cartesian component of the fields according to

0 = s , = {E, H, D, B} , = {x, y, z} (64)

and alternatively write

0
~ PML0 , r H
r E~PML = j sB ~ PML0 = j
sD~ PML0 (65)

with
sy sz
sx 0 0
sz sx 1
s = 0 sy 0 , s = 1 + (66)
sx sy
j
0 0 sz

Relation (65) together with (66) is the uniaxial formulation of the perfectly matched
layer (UPML).
Notice that both the stretching technique [35][42] and the UPML [36][43] are
equivalent in their absorption capabilities to the split Berenger PML [44].
The stretched variable formulation has proven to be useful to extend the method
to non cartesian coordinates [45], to absorb efficiently both propagative and evanes-
cent waves (complex frequency-shifted (CFS)-PML) [46][47], to absorb waves in
bianisotropic lossy media [48], etc.
A drawback of the split equation resides in the fact that it actually supports
not only plane waves with frequency and wavenumber related by the dispersion
relation, but also a subcomponent of the field can be increasing, while the other one
decreases, linearly in time, as proven in [49], exactly cancelling out this behavior
when the two subcomponents are added. This may lead to late time instabilities
associated with roundoff errors as seen in [50].
In relation to the implementation of the split Berengers PML time domain equa-
tions (55) it is straightforward to obtain an explicit finite difference numerical
scheme in a similar way to that employed with Maxwells time domain equations:
namely replacing all the derivatives by their finite centered approximation and av-
eraging in time all the fields affected by conductivity terms in eqn. (55), placed at
the space positions given by the Yee cube, and at time instants delayed by half the
time increment between the electric and magnetic fields. The time domain electric
constitutive relationship must be advanced after advancing the displacement field
and before advancing the magnetic field, and the time domain magnetic one, after
the latter and before closing the loop (see the section on material modeling for
more details).
The particular implementation in the time domain of the UPML and of the
stretching technique depends on the specific form of the s . For instance, for the
(CFS)-PML, the causal form [47]

s = + (67)
+ j

which is suitable to absorb evanescent waves, can be efficiently implemented in a

recursive convolutional form [46].
A common characteristic of these methods is that the conductivity profile in-
side the PML is implemented numerically as an increasing function, so that the
attenuated transmitted wave can be numerically negligible in less than 12 cells
of propagation. Typically the conductivity is increased in a polynomial profile of
order between 2 and 4 to obtain a theoretical normal attenuation around -120 dB

n

= max (68)
d

with the notation of figure (2).

Since the PML is truncated with PEC or PMC conditions, the calculated refflec-
tion coefficient as a function of the incidence angle is

max

2
r() = e n+1 v d cos
(69)

where for anisotropic media and/or dispersive media it should be taken into ac-
count that the phase speed can depend on the propagation angle and/or frequency
of the waves.
Finally, using the same construction, it can be seen that at the interface of
two PML media with given conductivities, perfect matching with attenuation is
achieved if equal conductivity terms corresponding to the plane interface direc-
tions are chosen. This is useful to determine which conductivity terms must be
used at the edges and corners of the computational domain. For instance, in fig-
ure (2) the 2D corner region has the same y as the left PML region, with no
restriction on x , and, reciprocally, it has the same x as the upper PML region,
with no restriction on y . The result is a region with two nonnull conductivities.























 











PEC


































 



















 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
    Y

 




















 
                                                                                 
 

























           x=     x                          







 x= 



x
 





 






1 
             =   0  1                          dx





 y=


 y2


             y                              


























 










































                                x          

 


 





 





 





 





 





 





 





 



  



 

 


 

 

 

 

 

 

 

 

 

 

 

 

 

 

 




 





 



 

 



 

 



 

 



 

 



 

 



 

 



 

 

  


 


 


 


 


 


 


 


 


 





 



 

 

 

 

 

 

 

 

 

 

 

 

 

 



   







 x

 =




 0


 



 



 





 


 

 
 




 





 



 

 



 

 



 

 



 

 



 

 



 
PEC





 



 




 


= 

 




 




 


 


 


 
 
    2
 

 


y


  

       
y

 


 











 





 





 





 





 





 







 





 






 






 






 






 






 






 




  

 


 





 
 

  

  

  

  

  

  

 



 
 





 



 



 



 



 



 





 


 

 
 




 



 

 



 

 



 

 



 

 



 

 



 

 



 





 



 




 




 




 




 




 


 


  





   

  

  

  

  

 
 

 

 

 




















y






 


 



 



 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

  


dy
X

Figure 2: Conductivity implementation at a 2D PML corner

6 UNCONDITIONALLY STABLE METHODS: ADI-FDTD

There is increasing interest in the design of unconditionally stable algorithms as

an alternative to the Yee one. Several starting points are possible: operational ap-
proaches [51], bruteforce CrankNicolson methods [52], implicit lower-upper ap-
proximate factorization (LU/AF) [53], Alternating Direction Implicit (ADI) split
step methods [16][17], etc. The latter, known as ADIFDTD, has inspired this new
line of research, and is described in detail in this section.
In the traditional explicit Yee FDTD algorithm, the time increment cannot ex-
ceed the Courant stability limit, and therefore the maximum time-step size is
restrained by the smallest spatial increment in the computational domain. This
means that when a fine mesh is needed to model part of the computational domain
(for instance by subgridding or uneven mesh techniques), the Courant condition
results in an oversampling in time, which leads to high computation times. An
alternative to the standard explicit FDTD method, the Alternating Direction Im-
plicit Finite Difference Time Domain (ADI-FDTD) method, has been proposed
in [16][17]. The ADI is a technique first found in the works of Peaceman, Rach-
ford and Douglas [54][55] used to perturb and factorize the unconditionally stable
CrankNicolson scheme approximation of partial differential equations, to render
it computationally costless without losing its unconditional stability.
For Maxwells curl equations, the ADIFDTD is based on an implicit-in-space
formulation of the FDTD, which offers unconditional numerical stability with little
extra computational effort. The ADI-FDTD method removes the stability limit for
the time increment, making it possible to choose the time increment independently
of the space increment. Thus, this technique allows us to choose dense spatial
grids where needed without necessarily requiring the use of small time increments.
Since the ADI-FDTD method keeps the Yee spatial distribution of the fields, it can
be applied to the whole range of problems addressed by the original Yee method.
Nevertheless, the time increment cannot be increased arbitrarily taking into ac-
count only numerical dispersion criteria (number of samples per period); it is also
limited by the fact that it must accurately resolve the spatial variations of the
fields [56] as we will see later. In spite of this, in many practical problems the
method still achieves significant reductions in CPU time compared to the FDTD
method.
Let us start by building a CrankNicolson FDTD (CNFDTD) scheme by aver-
aging in time the right hand side of (15)

M ~ n+1/2 + Si,j,k at
i,j,k t ~ n+1/2 = R at ~ n+1/2 (70)
i,j,k i,j,k i,j,k

~ n = (E
now with ~n ,H ~ n ), which after defining
i,j,k i,j,k i,j,k
i,j,k !
1 i,j,k I3 1
i,j,k r
RT = Mi,j,k R Si,j,k = 1 (71)
i,j,k r i,j,k
i,j,k
I3

leads us to
~ n+1
~ ni,j,k = t R T ~ n+1 + ~ ni,j,k (72)
i,j,k i,j,k
2
The application of the CN-FDTD scheme (72) to practical problems requires
prohibitively large computational resources [52], since it involves the solution of a
sparse linear system of equations at each time step. Using the ideas of Peaceman,
Rachford and Douglas [54] [55] it is possible to factorize the CNFDTD scheme
into a two-substep procedure, each one only needing the solution of a tridiagonal
system of equations, by adding a second order perturbation term in the following
manner
t 2
~ n+1 ~ n
i,j,k i,j,k + AB ~ n+1 ~ ni,j,k = t R T ~ n+1 +
~ ni,j,k
i,j,k i,j,k
2 2
| {z }
P erturbation
(73)
If the operators A and B
are chosen so that

A + B
=R
T (74)
(73) can be rewritten as

t t ~ n+1 t t ~ n
I6 A I6 B i,j,k = I6 + A I6 + B i,j,k (75)
2 2 2 2
and factorized into a two-sub-step procedure as

t ~ n t ~ n
I6 A i,j,k = I6 + B i,j,k (76)
2 2

t ~ n+1 t ~ n
I6 B i,j,k = I6 + A i,j,k (77)
2 2
~ n being an auxiliary intermediate vector field.
with i,j,k
A convenient choice of A and B, which fulfills (74), is
1
! 1
!
2 i,j,k I3 i,j,k re 2 i,j,k I3 i,j,k ro

A= i,j,k
, B = i,j,k

i,j,k
1 1
i,j,k ro 2 i,j,k
i,j,k
I3 i,j,k re 2 i,j,k I3
(78)
0 0 y 0 z 0

re = z 0 0 , ro = 0 0 x (79)
0 x 0 y 0 0
with which we can write
! ! ! !
l1 I3 l2 re ~ n
E i,j,k l5 I3 l2 ro ~n
Ei,j,k
=
l3 ro l4 I3 ~ n
H i,j,k l3 re l6 I3 ~n
H i,j,k
! ! ! !
l1 I3 l2 ro ~ n+1
E i,j,k l5 I3 l2 re ~ n
E i,j,k
= (80)
l3 re l4 I3 ~ n+1
H i,j,k l3 ro l6 I3 ~ n
H i,j,k

defining
t i,j,k t t i,j,k
l1 = 1 + 4 i,j,k , l2 = 2 i,j,k , l5 = 1 4 i,j,k

t i,j,k t

t i,j,k (81)
l4 = 1 + 4 i,j,k , l3 = 2 i,j,k , l6 = 1 4 i,j,k

In order to solve the system of equations (76) and (77) explicitly in time for
~ n and H
E ~ n we first put it in triangular form
i,j,k i,j,k
! !
l4 l1 I3 + l3 l2 re ro 03 ~ n
E i,j,k
=
l3 ro l4 I3 ~ n
H i,j,k
! !
l4 l5 I3 l3 l2 re re l2 l6 re + l4 ro ~n
E i,j,k
l3 re l6 I3 ~n
H i,j,k
! !
l4 l1 I3 + l3 l2 ro re 03 ~ n+1
E i,j,k
=
l3 re l4 I3 ~ n+1
H i,j,k
 ! !
l4 l5 I3 l3 l2 ro ro l2 l4 re + l6 ro ~ n
E i,j,k
(82)
l3 ro l6 I3 ~ n
H i,j,k

This system of equations can now be solved by backsubstitution at each time

~ n , then H
step: to advance from n to n , first obtain E ~ n , and similarly to advance
from n to n + 1. It can be seen that at each substep n and n + 1 only the electric

field equations involve the solution of a system of equations which is tridiagonally

implicitinspace because the operators ro re and re ro only use second order
differences in one direction in turn.

0 x y 0 0 0 z x

re re = 0 0 y z , ro ro = x y 0 0
z x 0 0 0 y z 0
2 2
z 0 0 y 0 0

ro re = 0 x2 0 , re ro = 0 z2 0 (83)
0 0 y2 0 0 x2

The magnetic field is obtained explicitly in space. In the lossless case ( =

= 0), eqns. (82) coincide with the usual algorithm described in [16][17] in
n = n + 12 .

! !
t2
I3 + 16i,j,k i,j,k re ro
03 ~ n
E i,j,k
t
=
2 i,j,k ro I3 ~ n
H i,j,k
! !
t2
I3 16i,j,k i,j,k re re
t
2 i,j,k R
~n
Ei,j,k
2 t
i,j,k
re I3 ~n
H i,j,k
! !
t2
I3 + 16i,j,k i,j,k ro re
03 ~ n+1
E i,j,k
t
=
2 i,j,k re I3 ~ n+1
H i,j,k
! !
t2
I3 16i,j,k i,j,k ro ro
t
2 i,j,k R
~ n
E i,j,k
(84)
2 t
i,j,k
ro I3 ~ n
H i,j,k

PML absorbing boundary conditions have been obtained for the ADI-FDTD
method both in the Berenger split form [57] and in the stretched coordinate met-
ric [58].
An interesting alternative approach to build unconditionally stable schemes can
~ be a normalized field vector
be found in [51]; let us summarize it briefly. Let

~ T
~=
~
E, H (85)
so that Maxwells curl equations are written (in the lossless currentfree case for
simplicity)
!
03 1
1
~ ~
N (t) N = r
t (t) =R , R (86)
r 1
1

03

~
With this normalisation, the usual spatial euclidean L2 norm of the vector
gives a measure of the electromagetic energy. The formal solution of (86) with
~0 , is formally
initial conditions in time

~ = et R N
~0 (87)

where the absence of currents implies that the total energy is conserved (the L2
norm of ~ coincides with that of ~0 ). Hence the time operator et R N is orthogonal;
~0 to obtain
it merely rotates the initial state ~ from it. Thus a semidiscrete time
advancing scheme of the form

~ n+1 = et R N
~n (88)

becomes straightforwardly stable.

Any orthogonal spatial approximation to et RN will also be necessarily stable in
time. A useful one is given by the Lie-Trotter-Suzuki formula [59] to approximate
matrix exponentials

P
p
i
!m
t H p
Y i /m
tH
e i=1 = lim e (89)
m
i=1

whose first order approximation coincides with that of the usual Taylor series.
The main idea of the method is to replace the analytical space derivatives by the
usual centered difference approximations, and to decompose the numerical ma-
N thus obtained into a sum of matrices R
trix R N so that their exponential can
i
be straightforwardly calculated. Since at a given point, the numerical operator RN
only involves the difference between two adjacent fields in two space directions, it
can be decomposed into the sum of matrices involving differences in only one di-
rection in turn. Furthermore, each of these matrices can be reordered and split into
two blockdiagonal matrices whose exponentials are known to be blockdiagonal
matrices combining harmonic functions.
The method needs to increase the order of approximation when the Courant
number is increased, with the subsequent increase in computational costs, but it
has been proven to exhibit a high degree of accuracy. Further studies based on
the Chebyshev polynomial approach to the exponential matrix are currently being
made [60].
6.1 CONSISTENCY, STABILITY AND DISPERSION OF THE ADI-FDTD

Let us obtain the local truncation error of the ADI-FDTD procedure assuming, for
~n
simplicity, a lossless case ( = = 0). Introducing the analytical solution i,j,k
into (73)

T~i,j,k
n
= ~ n+1 ~n
i,j,k i,j,k +
2 t
t ~ n+1 +
~n ~ n+1 + ~n
AB
i,j,k i,j,k RT
i,j,k i,j,k (90)
2 2

we obtain
T
T~i,j,k
n Exn0
=W , E n0
yi0 ,j 0 ,k0 , E n0
zi0 ,j 0 ,k0 , H n0
xi0 ,j 0 ,k0 , H n0
yi0 ,j 0 ,k0 , H n0
zi0 ,j 0 ,k0 (91)
i0 ,j 0 ,k0

with
2t y $xxy $xyz
0 0 z
2t z $yyz $xyz
0 x 0

y $zzy
x 2t $xyz
0 0

W = $xyz
0 0 2t
y x $zzx

$xyz z $xxz 2t
0 y 0
$xyz x $yyx
z 0 0 2t

(92)
and
1 1 1 1 1 1
n n0 n+1 , i i0 i+ , j j 0 j + , k k 0 k +
2 2 2 2 2 2
(93)
1 1
and $uvw = 4 t2 u v w , u = 24 u2 u3 . The quantities $uvw represent
the terms which are unique to the ADIFDTD scheme and they are not present
in the CNFDTD scheme truncation error ($uvw = 0). Thus we can conclude
that both the CNFDTD and the ADIFDTD schemes are consistent up to second
order with Maxwells curl equations.
Nevertheless, a closer look at expression (92) reveals the fact that the second-
order terms of the truncation error which are unique to ADIFDTD ($uvw ) depend
on the square of the time increment multiplied by the spatial derivatives of the field.
These terms, which do not appear in the CNFDTD scheme, explain the inaccura-
cies which appear in lowfrequency problems [56], where the space variations are
determined more by the material inhomogeneities than by the wavelength, and the
error cannot be predicted by the homogeneous dispersion relation. For instance, a
simple parallelplate capacitor fed by a very lowfrequency voltage signal keeps
the field confined inside the two plates, with an abrupt step to zerofield value out-
side the plates. In the space spectral domain the space derivatives take the form
 j 2/ with much smaller for this problem than the wavelength as-
sociated with the waves propagated by the slow time variation of the fields inside
the capacitor (due to the abrupt step of the fields at the parallelplate edges). Thus
the ADIFDTD must choose the time increment bounding the terms $uvw , that in
space spectral domain depend inversely on u /(ct) which is a measure of how
the time increment resolves the space variations.
If we study the truncation error of each substep (76) and (77) (still for the
lossless case), it can be proven that each one is consistent only up to first order
in time (second in space) with Maxwells curl equations if n = n + 12 (T~ =
O(t, x2 , y 2 , z 2 )). The first substep (76) approximates to (5) at t = (n +
1 3
4 )t and the second substep (77) at t = (n + 4 )t. It should be noted that each
substep in the time-stepping algorithm can lose accuracy, and even consistency,
without changing the overall accuracy of the total scheme (see [61] for further
discussion).
The dispersion relation for the lossless homogeneous case ( = = 0, i,j,k =
, i,j,k = ) is easily obtained rewriting eqn. (73) as

t2 ~ n+1/2
~ n+1/2 ~ n+1/2
T at
t i,j,k + AB t i,j,k =R i,j,k (94)
4

and setting the eigenvalue problem

2

R + t t
t I6 at A
B ~ n+1/2 = 0 (95)
T i,j,k
4

for complex exponential eigenvectors (25), with

!
03
1/
R =
r
T
1/ 03
r
! !
03
1/ 03 1/
A =
re B =
, ro

1/ ro
03 1/ 03
re

0 z y

r = z 0 x
y x 0

0 0 y 0 z 0
=
z 0 0

, ro = 0 0

x
re

0 x 0 y 0 0

t 2j t 2j
at = 2 cos , t = sin , = sin
2 t 2 2
 = {x, y, z} (96)

which after enforcing a nontrivial solution (null determinant) leads to

q
Wx2 + Wy2 + Wz2 + Wy2 Wz2 +Wx2 Wy2 + Wx2 Wz2
t
tan = q
2 Wx2 Wy2 Wz2 + 3 3
(97)
with

t
W = sin (98)
2

This is a very convenient form for the dispersion relation, equivalent to the one
published in [62], but, in tangent form, it can easily be seen that there cannot appear
frequencies with an imaginary part for any real value of the wavenumbers because
the inverse tangent function is always real for any real argument. Thus the stability
of the scheme is straightforwardly deduced.
The CNFDTD dispersion relation
q
Wx2 + Wy2 + Wz2
t
tan = (99)
2

leads to similar conclusions about stability.

Just for comparison purposes, let us finally write the Yee FDTD dispersion re-
lation with this notation
q
Wx2 + Wy2 + Wz2
t
sin = (100)
2

where the only difference between the CNFDTD dispersion relation (99) and
the Yee one (100) is the function applied to t2 (the FDTD supports unstable
solutions since the inverse sine functions do not admit real values for arguments
greater than 1, unlike ADI and CN ).

7 Material modeling

The treatment of all kinds of materials is a powerful characteristic of the FDTD

method. An interesting class of materials are those having linear electric and/or
magnetic dispersion. In general, a dispersive medium responds to an electromag-
netic field as the superposition in time of two responses: an instantaneous response
(the socalled infinite frequency response), and a retarded one coming from the en-
ergy initially absorbed by the medium and subsequently returned (partially) by it.
This second response is delayed in time because of material inertia and is respon-
sible for the energy dispersion. For instance, for a dispersive dielectric medium the
displacement vector is

Retarded response
zZ }| {
t
~
D(t) = ~
E(t) + ~ )()d
E(t (101)
| {z } 0
Instantaneous response

with (t) a time function depending exclusively on the material properties.

Although the simulation of these media is well documented [3][4][5] let us
just cite some of the approaches present in the literature: the recursive convolu-
tion methods [63][64], the auxiliary differential equation (ADE) form [65], the
Z-transform techniques [66] and the statespace approach [67]. Next, the two first
approaches are briefly described.
For illustration purposes, we assume a simple singlepole Debye medium with
a frequency domain form of the relationship (101)

~ ~ s
D() = () E() with () = + (102)
1 + j

whose convolutional form is (noting with F 1 the inverse Fourier transform)

Z t Z t
~
D(t) = E(t ~ + s
~ )F 1 {()} d = E(t) ~ )e-t/ d
E(t
0 0
(103)
which has been efficiently implemented with a piecewise linear recursive convo-
lution (PLRC) technique [64].
On the other hand, the auxiliary differential equation (ADE) form of (102)
is [65]
~ +D
t D ~ = t E~ + s E~ (104)

which can be extended to include the effect of conductivity by defining a complex

~
permeability c () = () + j with which the relation D() ~
= c () E() can
be written in the time domain as

~ + t D
t2 D ~ = t2 E~ + (s + ) t E~ + E~ (105)

This form is well suited for the construction of PMLs to match waves in disper-
sive media since the same formal relation (105) is obtained from the frequency do-
main relationship between the split fields D~ sPML () = c () E~sPML () which holds
in the PML. A possible numerical implementation of (105), is obtained using finite
centered time differences
~ n+1/2 2E
~ n1/2 +E
~ n3/2 ~ n+1/2 E
~ n3/2 ~ n+1/2 +E~ n3/2
E t2 + (s + ) E 2t + E 2 =
~
D n+1/2
2D~ n1/2 ~
+D n3/2 ~ n+1/2 ~ n3/2

t2 +D D
2t
(106)
with which the total advancing algorithm can be written as (only discretised in
time for the sake of writing simplicity)

D ~ n1/2 + t r H
~ n+1/2 = D ~n
E~ n+1/2 = k1 E
~ n1/2 + k2 E
~ n3/2 + k3 D~ n+1/2 + k4 D
~ n1/2 + k5 D
~ n3/2

H~ n+1 = t/2 H~n t ~ n+1/2
+ t/2 + t/2 r E
(107)
defining

k = s2t
+
+ t
2 s +
2 + 2 , k1 = kt2 , k2 = 2kt

kt2
2k ,
1 2 1
(108)
k3 = kt 2 + 2kt , k4 = kt2 , k5 = kt2 2kt

The ADI procedure has also been extended to Debye dispersive media using the
same constructing principle [68]. For illustration purposes let us summarize the
main steps. First of all, since the construction procedure can be neatly applied to
first order time differential equations, let us write the time domain equations of a
Debye dispersive medium in the following form
~ + E~ = r H
0 t E~ + t P ~ (109)
~ = r E~
t H (110)
~ + t P
P ~ = (s 0 ) E~ + ( 0 ) t E~ (111)
with P~ being the usual electric polarization vector. Equation (111) is the auxiliary
differential equation time domain version of the Debye frequency domain rela-
tionship. There is nothing essential in choosing the polarization vector P ~ equa-
tion (111), instead of the displacement vector D ~ equation (104). We just do it to
show how the ADI procedure can be systematically employed.
Defining a compound vector
~ = (E,
~ H,
~ P)
~ T (112)
we can write a short form of (109)(111) as
~=R
t ~
T (113)
now with
1
0 I3 03 I3 I3 r 03
T =
R 03 I3 03

r 03 03

(0 ) I3 03 I3 (s 0 ) I3 03 I3
(114)
which after being discretised in Crank-Nicolson form is formally similar to (72),
that is
~ n+1 ~ ni,j,k = t R T ~ n+1 +
~ ni,j,k (115)
i,j,k i,j,k
2
now with R T the numerical centered finite difference approximation to the an-
T (114). Following the same procedure (73)-(75) we arrive at
alytical operator R
expressions formally identical to (76)(77) which can be rendered in triangular form
as
A ~ n = B
i,j,k
+
~n
i,j,k (116)

B ~ n+1 = A+
~ n (117)
i,j,k i,j,k
with
~ ni,j,k = (E
~ i,j,k
n ~ i,j,k
,H n
, P~i,j,k
n
)T (118)

s0 I3 + s1 re ro 03 03

A = r3 t
2 ro I3 03

t
r3 4 I3 r5 t t
I3
2 re 1 r6 4

s2 I3 s1 re re
s3 R r2 t 2 I3

+ 03
B = r3 t
2 re I3
t

r4 t t
4 I3 r5 ro
2 1 + r 6 4 I3

s0 I3 + s1 ro re 03 03

B = r3 t 2 re I3 03
t

r3 t t
4 I3 r5 2 ro 1 r 6 4 I3

s2 I3 s1 ro ro s3 R r2 t 2 I3


A+ = r3 t
2 ro I3 03 (119)

t
r4 I3 4 r5 t re
2 1 + r6 t I3
4

with
1 1 0 (s + ) 20
r2 = , r3 = , r4 =

0 0
r5 = , r6 =

t2 + 16 + 4t (s + ) t2 (t + 4 )
s0 = , s1 =
16 16
16 t (80 + 4s + (t + 4 )) t (t + 4 )
s2 = , s3 =
16 8
(120)
Finally, let us mention the increasing interest in the numerical simulation of the
socalled metamaterials. These artificial materials are characterized as exhibiting
electromagnetic properties not found in natural materials or compounds. A com-
mon example is the case of periodic materials based on repeated cells: for instance,
a homogeneous substrate with small metallic inclusions periodically (or randomly)
distributed inside it. Depending on the geometry of the inclusions the material
can behave, in a range of frequencies, according to Maxwells curl equations with
macroscopic constitutive relationships that are different from those of the ordinary
materials: negative permeability and/or permittivity, magnetoelectric coupling (bi-
isotropy and bianisotropy), etc. A noteworthy type of periodic materials are the
Yablonovitch photonic bandgap (PBG) crystals.
The FDTD method is a powerful tool to assist in the development of the tech-
nology associated with these new materials. For instance, a linear lossy dispersive
bianisotropic material is a good example of a general linear complex medium: it
presents anisotropy, magnetoelectric coupling and dispersivity. For these one must
solve the usual Maxwells curl equations in the time domain

r E~ = t B
~ , r H
~ = t D
~ (121)

together with the constitutive relationships in the frequency domain that can be
expressed as
~
D() ~
= ()E() H()
+ () ~ (122)

~
B() E()
= () ~ + ~
()H() (123)

which take into account all of the abovementioned effects ( () and () already
comprise the conductivity losses).
The first problem in the implementation of these media into FDTD is that of
anisotropy. For an E-H formulation, the Yee distribution of the fields is not valid,
since the updating of an electric/magnetic field component in a given direction
involves not only the magnetic/electric field components perpendicular to this di-
rection but all the components of the electric field, because of the anisotropy. One
possibility is to apply the usual Yee distribution using the average operator to ob-
tain the fields at unknown positions as a function of those at known positions [69]
(this averaging approach can also be used in a materialindependent form to ob-
tain the E~ and H ~ fields from D,
~ B ~ at the Yee cell positions). Another possibility
is to define a field component distribution different from that of Yee, together with
different approximations to the spatial derivatives using both finite centered dif-
ferences and averages (for instance, all the electric field components defined in
a single node and all the magnetic field ones in another single node, the two be-
ing at opposite vertices of the discretisation cell, replacing a a with ,
, perpendicular directions in space). This approach, which has been used suc-
cessfully for lossless nondispersive anisotropic media [27], has also been used
for lossless bianisotropic media using a singlefrequency marchingonintime
algorithm together with a second order backwards differentiation formula for the
magnetoelectric coupled time derivatives terms [41], with which for instance, the
numerical Amp`ere equation can be written as

~ 1 1 (a
t E~ = 1 r H ~ 1
x ay az )t H ~
(ax ay az )H (124)
c c

where we have used the complex Tellegen and chirality tensors related to

and by = ( j
)/c, = ( + j)/c, with c = 1/ 0 0 representing the
freespace light speed [70].
In (124) r represents the numerical curl replacing each space derivative
a a with , , being perpendicular directions in space, and

3f (t) 4f (t t) + f (t 2t)
t f (t) (125)
2t

Some other approaches actually incorporate the time domain dispersion into the
equations, as for instance [71] (in convolutional form for a chiral slab), but this
is still a field of open research. Another recent approach is based on the wavefield
decompositions, only valid for homogeneous bianisotropic media, and can reduce
the problem to the simulation of two anisotropic problems without magnetoelectric
coupling [72].
Finally, let us note that the FDTD method has also been successfully extended
to study nonlinear media [73], which are of current interest in the design of non
linear optical devices, characterized by their ability to manage femtosecond elec-
tromagnetic pulses. An interesting solution of the equations governing these media
is given by solitons, characterized as being able to propagate inside them without
deformation.

8 Geometrical modeling
The main advantage of the FDTD method resides in its ability to deal with prob-
lems incorporating materials with geometrical inhomogeneities. The spatial ar-
rangement of the fields in the Yee lattice provides a natural way to implement
abrupt material changes conserving the natural (dis)continuity of the analytical
fields, since it is well known that the numerical evaluation of the Faraday and
Amp`ere integral equations, assuming constant values of the fields inside a Yee
cell, leads to the same numerical scheme as the second order finite difference ap-
proximation of the differential equations [4].
Let us assume an interface between two different isotropic linear media with
different permeabilities, permittivities and finite conductivities (both magnetic and
electric). The time domain continuity conditions on the normal (subindex n) and
tangential (subindex t) components of fields and currents are

E1t = E2t , H1t = H2t , D2n D1n = s , B2n B1n = s

 J2n J1n = t s , J2n J1n = t s (126)

with s , s being the electric and magnetic surface density charge stored between
the two media. Using the frequency domain constitutive relationships D ~ = E,
~
~ ~ ~ ~ ~ ~
B = H, J = E, J = H, the E-H continuity conditions become

E1t = E2t , H1t = H2t (127)

1 2
1 + E 1n = 2 + E2n , 1 + 1 H1n = 2 + 2 H2n
j j j j
| {z } | {z } | {z } | {z }
c1 c2 c1 c1

with c being the usual complex permittivity c = + j and c the usual


complex permeability c = + j .
As an illustration of how to employ these conditions, let us assume a planar
material interface as in figure 3, crossing a Yee cell at an arbitrary position. Dis-
cretizing the Faraday and Amp`ere integral laws, we obtain the modified updating
interface equations of the fields (figure 3). For instance, for the Hy and Ex com-
ponents discretised only in space (we assume x = y = z
Z I
~ ~ ~ ~ d
c j HdS = Edl c2 + (c1 c2 ) jHx (0, 0.5, 0.5) =
SL L

c2 c1 d Ey (0, 0.5, 1) Ey (0, 0.5, 0)
1+ +
c1
Ez (0.5, 0, 0.5) Ez (0, 1, 0.5)

(128)

Z I
~ S~= ~ ~l c1 + c2 d
c j Ed Hd + (c1 c2 ) jEx (0.5, 0, 0) =
SL L 2

c1 + c2 c2 c1 d Hy (0.5, 0, 0.5) Hy (0.5, 0, 0.5)
+ +
2c2 c2
Hz (0.5, 0.5, 0) Hz (0.5, 0.5, 0)
(129)

Notice that in the homogeneous case these equations coincide with the classic Yee
ones.
In order to obtain a time domain discrete versions of (128) let us distinguish two
cases, focusing only on the physical case = 0
Boundary at the tangential electric fields plane (d = 0): no correction for
the frequency or time domain form of the Faraday law in (128) is necessary,
while for the Amp`ere law (129) the average of the complex permittivities
 c1,c c ,c c1,c c ,c
1 2 2
1 2 2

( c2 / c ) d Ey(0,0.5,1) ( -d) E y(0,0.5,1)

1

d E y(0,0.5,1)






































































































H 


























y

(-0.5,0,0.5)









































































E z(0,0,0.5)



























 























 E z(0,1,0.5)
Hx (0,-0.5,0.5) 































































































































































































H 

x
(0,0.5,0.5)





























































 
 
 
 
 

 
 
 
 
 
Hy (0.5,0,0.5) 
 
 
 
 
 

 
 
 
 
 

 
 
 
 
  E y(0,0.5,0)

 
 
 
 
 

 
 
 
 
 

 
 
 
 
 

 
 
 
 
 
Hz (0.5,-0.5,0) E x(0.5,0,0) 
 
 
 
 
  Hz (0.5,0.5,0)

 
 
 
 
 

 
 
 
 
 

 
 
 
 
  E x(0.5,1,0)

 
 
 
 
 
E y(1,0.5,0) 
 
 
 
 
 

 
 
 
 
 

 
 
 
 
 
d 
 
 
 
 
 
Hy (0.5,0,-0.5) 
 
 
 
 
 

 
 
 
 
 

(d+0.5 ) Hy (0.5,0,-0.5) ( c2 / c ) (0.5 -d) Hy (0.5,0,-0.5)

1

Figure 3: Interface between two different media

in the frequency domain is needed, that is, the average of both the real per-
mittivity and the conductivity in the time domain, together with the use of a
correction factor of the form 12 +2
2
for the magnetic components between
the two media, in the frequency domain, which coincides with the correction
factor in the time domain for nondispersive magnetic media.
Boundary placed on the tangential magnetic field plane d = /2: no correc-
tion for the frequency or time domain form of the Amp`ere equation in (129)
is necessary, while for the Faraday equation (128) the average of the perme-
abilities both in frequency and in the time domain is needed, together with
1 +c2
the use of a correction factor in the frequency domain of the form c2 c2
for the electric components between the two media. This correction factor
is real for lossless (1 = 2 = 0) nondispersive media and can be directly
implemented in the time domain, otherwise there is no such straight time
domain version.
Finally, if one of the media is a perfect electrical conductor (PEC), the coincidence
of the interface with the tangential electric fields and enforcing the latter to be null
leads to the correct implementation of the boundary conditions. Similar results are
achieved for perfect magnetic conductors (PMC), after placing the interface at the
tangential magnetic fields and enforcing these to be zero.
Notice that the frequency domain formulation of the continuity conditions at the
interface allows a systematic approach to this problem.
When the media interfaces are not plane, the accurate handling of the geomet-
rical details of the objects becomes a topic to which a lot of effort has been ded-
icated. If the interface is approximated in a staircase way, significant errors can
appear [74]. Not only the staircasing, but also the assumption of constant fields
inside the cells neighboring material terminations (edges, slots, wires, corners,...)
can also lead to errors.
There are two large categories in the solution of this problem: the first is based
on the hybridation of the FDTD with a method that is well suited to accurately
model problematic geometrical details (FETD, MoMTD, etc). The other category
globally or locally modifies the spatial grid to take into account the geometrical,
material and field variation details. The distortion of the grid can be performed
with a metric structure starting from the differential equations in general non
orthogonal coordinates [23], or without any particular metric using the integral
forms of Maxwells equations, with which the grid can actually conform to the ob-
jects (finite volume techniques [21]). Other approaches to the problem include the
Finite Integration Technique (FIT) [22], the Contour Path/Volume conformation
methods [75, 76, 77], etc. In general, these techniques have in common the need
for sometimes costly geometrical preprocessing to generate the discrete model, as
well as the fact that their stability might be compromised not only by the Courant
condition, but also because of the parasitic modes excited at the interface between
different advancing schemes [33, 76].
A bruteforce approach to the problem is the global reduction of the cell size to
accurately handle the geometrical and field variation details, but this may render
the method computationally too intensive. A partial solution is to only use fine
meshing where needed, keeping coarse mesh in the homogeneous zones. Two ways
of materializing this are found:
The use of nonuniform discretisations in the computational grid, varying
in a smooth geometric series the step size in each direction from the regions
free of inhomogeneities to those involving fine details: edges, curvatures,
etc. Although this technique reduces the order of the approximation to the
spatial derivatives by one, an overall order of accuracy close to two is ac-
tually obtained if the size variation from cell to cell is smooth (geometrical
progression with a ratio close to 1) [78]. Maintaining a Courant number be-
low the critical one for the smallest cell size in the whole grid avoids insta-
bilities for most problems of interest in the classic FDTD, with no restriction
in the ADIFDTD.
The subgridding technique, in which the cells containing critical details
are subdivided into subcells with the same advancing algorithm as that of
the coarser grid. At the finecoarse grid interface an interpolating algorithm
connects the two schemes [79] both in time and in space, and the Courant
number is kept the same in both grids. Thus the time increment in the finer
grid is smaller than that in the coarser one and several time steps must be
 advanced by the Yee scheme in the finer grid for each step in the coarser
one. The ADIFDTD, due to its stability, can use the same time increment
in both grids, so that only spatial interpolation is needed at the frontier [80].
Figure 4 depicts the layout of both techniques, for details of which the reader is
referred to the cited literature.

Ey

Ex
Hz

SUBGRIDDING NON-UNIFORM MESH

Figure 4: 2D example of subgridding and nonuniform mesh

9 Afterword

In this chapter we have presented an overview of the FDTD method. Due to limi-
tations of space we have been obliged to summarize and select certain topics. The
reader, however, can obtain further information by making use of the bibliography
included

Acknowledgements

Part of this work was supported by the Spanish National Research Project TIC-
2001-3236-C02-01 and TIC-2001-2364-C03-03.

References
[1] Courant, R., Friedrichs, K. & Lewy, H., Uber die partiellen differenzengle-
ichunger der mathematischen physik. Uber die partiellen Differenzengle-
ichunger der mathematischen Physik Math Ann, 19.
 [2] Yee, K.S., Numerical solution of initial boundary value problems involving
Maxwells equations in isotropic media. IEEE Transactions on Antennas and
Propagation, 14(3), pp. 302307, 1966.
[3] Kunz, K.S. & Luebbers, R.J., The Finite Difference Time Domain Method for
Electromagnetics. CRC Press: Boca Raton, FL, 1993.
[4] Taflove, A., Computational Electrodynamics: The Finite-Difference Time-
Domain Method. Artech House: Boston, MA, 1995.
[5] Taflove, A. & Hagness, S., Computational Electrodynamics: The Finite-
Difference Time-Domain Method, 2 ed. Artech House: Boston, MA, 2000.
[6] Taflove, A., (ed.) Advances in Computational Electrodynamics: The Finite-
Difference Time-Domain Method. Artech House: Boston, MA, 1998.
[7] Sullivan, D.M., Electromagnetic Simulation Using the FDTD Method. IEEE
Press: New York, 2000.
[8] Chew, W.C., Electromagnetic theory on a lattice. Journal of Applied Physics,
75(10), pp. 48434850, 1994.
[9] Wagner, C.L., Sources and Field Divergence in Electromagnetic Finite-
Difference Time-Domain Simulations. Masters thesis, Washington State Uni-
versity, Pullman, WA, 1998.
[10] Fang, J., Time Domain Finite Difference Computation for Maxwells Equa-
tions. Ph.D. thesis, University of California at Berkeley, Berkeley, CA, 1989.
[11] Liu, Q.H., The PSTD algorithm: A time-domain method requiring only two
cells per wavelength. Microwave and Optical Technology Letters, 15(3), pp.
158165, 1997.
[12] Krumpholz, M. & Katehi, L.P.B., MRTD: New time-domain schemes based
on multiresolution analysis. IEEE Transactions on Microwave Theory and
Techniques, 44(4), pp. 555571, 1996.
[13] Mickens, R.E., (ed.) Nonstandard finite difference models of differential
equations. World Scientific Publishing Co.: River Edge, NJ, 1994.
[14] Cole, J.B., A high-accuracy realization of the Yee algorithm using non-
standard finite differences. IEEE Transactions on Microwave Theory and
Techniques, 45(6), pp. 991996, 1997.
[15] Beggs, J.H., A two-dimensional linear bicharacteristic FDTD method. IEEE
Antennas and Propagat. Soc. Int. Symposium, volume 3, pp. 260263, 2002.
[16] Namiki, T., A new FDTD algorithm based on alternating-direction implicit
method. IEEE Transactions on Microwave Theory and Techniques, 47(10),
pp. 20032007, 1999.
[17] Zheng, F., Chen, Z. & Zhang, J., A finite-difference time-domain method
without the Courant stability conditions. IEEE Microwave Guided Wave Let-
ters, 9(11), pp. 441443, 1999.
[18] Mur, G., Absorbing boundary conditions for the finite-difference approxima-
tion of the time-domain electromagnetic-field equations. IEEE Transactions
on Electromagnetic Compatibility, EMC-23(4), pp. 377382, 1981.
[19] Berenger, J.P., A perfectly matched layer for the absorption of electromag-
netic waves. Journal of Computational Physics, 114(1), pp. 185200, 1994.
[20] Navarro, E.A., Gimeno, B. & Cruz, J.L., Modelling of periodic structures
 using the finite-difference time-domain method combined with the Floquet
theorem. Electronics Letters, 29(5), pp. 446447, 1993.
[21] Madsen, N.K. & Ziolkowski, R.W., A three-dimensional modified finite vol-
ume technique for Maxwells equations. Electromagnetics, 10(1/2), pp. 147
161, 1990.
[22] Weiland, T., Time domain electromagnetic field computation with finite dif-
ference methods. International Journal of Numerical Modelling, 9, pp. 295
319, 1996.
[23] Holland, R., Finite difference solutions of Maxwells equations in general-
ized nonorthogonal coordinates. IEEE Transactions on Nuclear Science, NS-
30(6), pp. 45894591, 1983.
[24] Garca, S.G., Olmedo, B.G. & Martn, R.G., Time-domain near- to far-field
transformation for FDTD in two dimensions. Microwave and Optical Tech-
nology Letters, 27(6), pp. 427432, 2000.
[25] Merewether, D.E., Fisher, R. & Smith, F.W., On implementing a numeric
Huygens source scheme in a finite difference program to illuminate scatter-
ing bodies. IEEE Transactions on Nuclear Science, 27(6), pp. 18291833,
1980.
[26] Fang, J. & Mei, K.K., A super-absorbing boundary algorithm for solving
electromagnetic problems by time-domain finite-difference method. IEEE
Antennas and Propagat. Soc. Int. Symp., Syracuse, NY, pp. 472475, 1988.
[27] Garca, S.G., Hung-Bao, T.M., Martn, R.G. & Olmedo, B.G., On the appli-
cation of finite methods in time domain to anisotropic dielectric waveguides.
IEEE Transactions on Microwave Theory and Techniques, 44(12), pp. 2195
2206, 1996.
[28] Holland, R., THREDS: A finite-difference time-domain EMP code in 3D
spherical coordinates. IEEE Transactions on Nuclear Science, NS-30(6), pp.
45924595, 1983.
[29] Fusco, M., FDTD algorithm in curvilinear coordinates. IEEE Transactions
on Antennas and Propagation, 38(1), pp. 7689, 1990.
[30] Smith, G.D., Numerical Solution of Partial Differential Equations. Finite Dif-
ference Methods. Clarendon Press: Oxford, 1985.
[31] Mitchell, A. & Griffiths, D., The Finite Difference Method in Partial Differ-
ential Equations. John Wiley and Sons, 1980.
[32] Gustafsson, B., Kreiss, H. & Sundstrom, A., Stability theory of difference
approximations for mixed initial boundary value problems. ii. Mathematics
of Computation, 26, pp. 649686, 1972.
[33] Trefethen, L.N., Group velocity interpretation of the stability theory of
Gustafsson, Kreiss, and Sundstrom. J Comput Phys, 49(2), pp. 199217,
1983.
[34] Ramahi, O.M., The concurrent complementary operators method for FDTD
mesh truncation. IEEE Transactions on Antennas and Propagation, 46(10),
pp. 14751482, 1998.
[35] Chew, W.C. & Weedon, W.H., A 3D perfectly matched medium from modi-
fied Maxwells equations with stretched coordinates. Microwave and Optical
 Technology Letters, 7(13), pp. 599604, 1994.
[36] Sacks, Z.S., Kingsland, D.M., Lee, R. & Lee, J.F., A perfectly matched
anisotropic absorber for use as an absorbing boundary condition. IEEE Trans
on Antennas and Propagation, 43(12), pp. 14601463, 1995.
[37] Chen, B., Fang, D.G. & Zhou, B.H., Modified Berenger PML absorbing
boundary condition for FD-TD meshes. IEEE Microwave Guided Wave Let-
ters, 5(11), pp. 399401, 1995.
[38] Perez, I.V., Garca, S.G., Martn, R.G. & Olmedo, B.G., Generalization of
Berengers absorbing boundary conditions for 3-D magnetic and dielectric
anisotropic media. Microwave and Optical Technology Letters, 18(2), pp.
126130, 1998.
[39] Zhao, A.P., Juntunen, J. & Raisanen, A.V., Generalized material-independent
PML absorbers for the FDTD simulation of electromagnetic waves in arbi-
trary anisotropic dielectric and magnetic media. IEEE Microwave Guided
Wave Letters, 8(2), pp. 5254, 1998.
[40] Garca, S.G., Juntunen, J., Martn, R.G., Olmedo, A.P.Z.B.G. & Raisanen, A.,
A unified look at Berengers PML for general anisotropic media. Microwave
and Optical Technology Letters, 28(6), pp. 414416, 2001.
[41] Garca, S.G., Perez, I.V., Martn, R.G. & Olmedo, B.G., BiPML: A PML
to match waves in bianisotropic media. Microwave and Optical Technology
Letters, 20(1), pp. 4448, 1999.
[42] Rappaport, C.M., Perfectly matched absorbing boundary conditions based on
anisotropic lossy mapping of space. IEEE Microwave Guided Wave Letters,
5(3), pp. 9092, 1995.
[43] Gedney, S.D., An anisotropic perfectly matched layer-absorbing medium for
the truncation of FDTD lattices. IEEE Transactions on Antennas and Propa-
gation, 44(12), pp. 16301639, 1996.
[44] Berenger, J.P., Numerical reflection from fdtd-pmls: A comparison of the
split PML with the unsplit and cfs pmls. IEEE Trans on Antennas and Prop-
agation, 50, pp. 258265, 2002.
[45] Teixeira, F.L. & Chew, W.C., Systematic derivation of anisotropic PML ab-
sorbing media in cylindrical and spherical coordinates. IEEE Microwave
Guided Wave Letters, 7(11), pp. 371373, 1997.
[46] Roden, J.A. & Gedney, S.D., Convolution PML (CPML): An efficient FDTD
implementation of the CFS-PML for arbitrary media. Microwave and Optical
Technology Letters, 27(5), pp. 334339, 2000.
[47] Kuzuoglu, M. & Mittra, R., Frequency dependence of the constitutive pa-
rameters of causal perfectly matched anisotropic absorbers. IEEE Microwave
Guided Wave Letters, 6(12), pp. 447449, 1996.
[48] Teixeira, F.L. & Chew, W.C., A general approach to extend Berengers ab-
sorbing boundary condition to anisotropic and dispersive media. IEEE Trans-
actions on Antennas and Propagation, 46(9), pp. 13861387, 1998.
[49] Abarbanel, S. & Gottlieb, D., A mathematical analysis of the PML method.
Journal of Computational Physics, 134, pp. 357363, 1997.
[50] Garca, S.G., Lopez, M.A.H., Bretones, A.R., Pantoja, M.F. & Martn, R.G.,
 Unsplit Berengers PML equations for arbitrary media. Electronics Letters,
37, pp. 15091510, 2001.
[51] Kole, J.S., Figge, M.T. & De Raedt, H., Higher-order unconditionally stable
algorithms to solve the time-dependent Maxwell equations. Physical Review
E, 65, pp. 066705106670512, 2002.
[52] Fijany, A., Jensen, M.A., Rahmat-Samii, Y. & Barhen, J., A massively par-
allel computation strategy for FDTD: Time and space parallelism applied to
electromagnetics problems. IEEE Transactions on Antennas and Propaga-
tion, 43(12), pp. 14411449, 1995.
[53] Beggs, J.H. & Briley, W.R., An implicit LU/AF FDTD method. IEEE Anten-
nas and Propagat. Soc. Int. Symposium, Boston, MA, volume 1, pp. 6871,
2001.
[54] Peaceman, D.W. & Rachford (Jr.), H.H., The numerical solution of parabolic
and elliptic differential equations. J Soc Ind Appl Math, 3, pp. 2841, 1955.
[55] Douglas, J., On the numerical integration of uxx + uyy = ut by implicit
methods. J Soc Ind Appl Math, 3, pp. 4265, 1955.
[56] Garca, S.G., Lee, T.W. & Hagness, S.C., On the accuracy of the ADI-FDTD
method. IEEE Antennas and Wireless Propagation Letters, 1(1), pp. 3134,
2002.
[57] Liu, G. & Gedney, S.D., Perfectly matched layer media for an uncondition-
ally stable three-dimensional ADI-FDTD method. IEEE Microwave Guided
Wave Letters, 10(7), pp. 261263, 2000.
[58] Gedney, S.D., Liu, G., Roden, J.A. & Zhu, A., Perfectly matched layer media
with CFS for an unconditionally stable ADI-FDTD method. IEEE Transac-
tions on Antennas and Propagation, 49(11), pp. 15541559, 2001.
[59] Suzuki, M., General theory of fractal path integrals with applications to
many-body theories and statistical physics. Journal of Mathematical Physics,
32, pp. 400407, 1991.
[60] Raedt, H.D., Michielsen, K., Kole, J. & Figge, M., Chebyshev Method
to solve the Time-Dependent Maxwell equations. Springer Proceedings in
Physics, (to appear), 2003.
[61] Gourlay, A.R., & Mitchell, A.R., Twolevel difference schemes for hyper-
bolic systems. J Soc Ind Appl Math Numer Anal, 3(3), pp. 474485, 1966.
[62] Zheng, F. & Chen, Z., Numerical dispersion analysis of the unconditionally
stable 3-D ADI-FDTD method. IEEE Transactions on Microwave Theory
and Techniques, 49(5), pp. 10061009, 2001.
[63] Luebbers, R.J. & Hunsberger, F., FDTD for N th-order dispersive media.
IEEE Transactions on Antennas and Propagation, 40(11), pp. 12971301,
1992.
[64] Kelley, D.F. & Luebbers, R.J., Piecewise linear recursive convolution for dis-
persive media using FDTD. IEEE Transactions on Antennas and Propaga-
tion, 44(6), pp. 792797, 1996.
[65] Joseph, R.M., Hagness, S.C. & Taflove, A., Direct time integration of
Maxwells equations in linear dispersive media with absorption for scatter-
ing and propagation of femtosecond electromagnetic pulses. Optics Letters,
 16(18), pp. 14121414, 1991.
[66] Sullivan, D.M., Frequency-dependent FDTD methods using Z transforms.
IEEE Transactions on Antennas and Propagation, 40(10), pp. 12231230,
1992.
[67] Pereda, J.A., Vielva, L.A., Vegas, A. & Prieto, A., State-space approach for
the FDTD formulation for dispersive media. IEEE Transactions on Magnet-
ics, 31(3), pp. 16021605, 1995.
[68] Garca, S.G., Rubio, R.G., Bretones, A.R. & Martn, R.G., Extension of the
ADIFDTD method to Debye media. IEEE Trans on Antennas and Propa-
gation, To appear, 2004.
[69] Schneider, J.B. & Hudson, S., The finite-difference time-domain method ap-
plied to anisotropic material. IEEE Transactions on Antennas and Propaga-
tion, 41(7), pp. 994999, 1993.
[70] Lindell, I.V., Sihvola, A.H., Tretyakov, S.A. & Vitanen, A.J., (eds.) Electro-
magnetic waves in Chiral and Biisotropic media. Artech House, 1994.
[71] Kusunoki, A. & Tanaka, M., Time-domain inverse scattering of a chiral slab.
Asia Pacific Microwave Conf., Singapore, volume 2, pp. 554557, 1999.
[72] Akyurtlu, A., Werner, D.H. & Aydin, K., BI-FDTD: A new technique for
modeling electromagnetic wave interaction with bi-isotropic media. Mi-
crowave and Optical Technology Letters, 26(4), pp. 239242, 2000.
[73] Joseph, R.M. & Taflove, A., FDTD Maxwells equations models for nonlin-
ear electrodynamics and optics. IEEE Transactions on Antennas and Propa-
gation, 45(3), pp. 364374, 1997.
[74] Holland, R., Pitfalls of staircase meshing. IEEE Transactions on Electromag-
netic Compatibility, 35(4), pp. 434439, 1993.
[75] Jurgens, T.G. & Taflove, A., Three-dimensional contour FDTD modeling of
scattering from single and multiple bodies. IEEE Transactions on Antennas
and Propagation, 41(12), pp. 17031708, 1993.
[76] Railton, C.J. & Craddock, I.J., Stabilised CPFDTD algorithm for the analysis
of arbitrary 3D PEC structures. IEE Proceedings H: Microwaves, Antennas
and Propagation, 143(5), pp. 367372, 1996.
[77] Garca, S.G., Bao-Hung, T.M., Olmedo, B.G. & Martn, R.G., Volume con-
formation method in scattering problems with FDTD. Proc. Int. Symp. on
Antennas, Nice, France, pp. 254257, 1994.
[78] Monk, P. & Suli, E., Error estimates for Yees method on non-uniform grids.
IEEE Transactions on Magnetics, 30(5), pp. 32003203, 1994.
[79] Chevalier, M.W., Luebbers, R.J. & Cable, V.P., FDTD local grid with material
traverse. IEEE Transactions on Antennas and Propagation, 45(3), pp. 411
421, 1997.
[80] Wang, B.Z., Wang, Y., Yu, W. & Mittra, R., A hybrid 2-D ADI-FDTD sub-
gridding scheme for modeling onchip interconnects. IEEE Transactions on
Advanced Packaging, 24(2), pp. 528533, 2001.