Application of the Complex Image Method to

Electromagnetic Field Computation in

Planar Uniaxial Multilayers
Krzysztof Arkadiusz Michalski

Department of Electrical and Computer Engineering

Texas A&M University
College Station, Texas 77843, USA
[email protected]
June 11, 2015

Abstract
The discrete complex image method (DCIM) is applied to compute the complete set of
electric, magnetic, and potential Green functions (or integral equation kernels) for planar mul-
tilayers with uniaxial anisotropy. It is assumed that the layered medium is of infinite lateral
extent and the optic axis is normal to the stratification. Sample results are presented and are
shown to be in good agreement with reference data obtained by a rigorous integration method.

Keywords: Discrete complex image method, Green function, integral equation, layered
medium, uniaxial medium.

1 Introduction
In a variety of applications, such as the analysis of microstrip circuits and antennas or VLSI in-
terconnects, it is necessary to compute the electromagnetic field due to time-harmonic currents
embedded in planar multilayered media. It can often be assumed that the layered medium is of
infinite lateral extent. Usually, integral equation methods, based on surface or volume formula-
tions, are used in this context, because they only require the discretization of the scattering object,
with the layered background rigorously incorporated via the appropriate (dyadic) Green functions.
The Green functions (or integral equation kernels) in such environments give rise to Sommerfeld
integrals, which are time-consuming to evaluate, due to the singular and oscillatory nature of the
integrands [1, 2]. Consequently, “brute-force” numerical integration is not practical. To alleviate
this problem, the discrete complex image method (DCIM) was developed [3, 4]. The key step of
the DCIM is an approximation of the spectral integrand by a finite sum of complex exponential
terms, which can be associated with discrete images of the source at complex-valued locations in an
“equivalent” homogeneous medium. The advantage of this method is that the image representation
has a closed-form inverse.
A lot of effort has recently been directed at the development of robust DCIM representations
of Green’s function in layered media [5–11]. However, with a few exceptions [6, 7], this effort
was limited to the the vector and scalar potential kernels associated with the so-called mixed-
potential integral equation (MPIE) [12–15], which are easier to handle than the more singular field
kernels. Furthermore, with one exception known to us [16], previous work was limited to isotropic
media, even though the MPIE formulation for uniaxial multilayers had been known for quite some
time [15, 17]. In [16], results are only presented for the scalar potential kernel in a grounded-slab
configuration.
In the present paper, we describe the application of the DCIM to the complete set of electric,
magnetic, and MPIE kernels for multilayers that may comprise uniaxially anisotropic media with
the optic axis normal to the stratification.

2 Problem statement
We consider a plane-stratified medium excited by arbitrary, time-harmonic,1 electric and magnetic
currents, J and M , respectively,2 occupying a volume V or residing on a surface S, as illustrated
in Fig. 1a. The stratification is perpendicular to the z axis and of infinite lateral extent along the
x and y axes. The layers are numbered from 1 to N in the direction of increasing z, where the
number of layers, N , is arbitrary. The nth layer, with interfaces at z = zn and z = zn+1 , is filled
with a homogeneous, linear, possibly lossy, possibly uniaxial medium, characterized by diagonal
permittivity and permeability tensors

 n = (x̂x̂ + ŷ ŷ)tn + ẑ ẑzn , (1)

µ = (x̂x̂ + ŷ ŷ)µtn + ẑ ẑµzn , (2)
n

where tn and zn are transverse and longitudinal permittivities, and µtn and µzn are transverse
and longitudinal permeabilities, all relative to free space.3 The sources may occupy any number
of layers, including the interfaces. The quantities of interest are the electric and magnetic fields,
E and H, respectively, at an arbitrary location r. In solving this problem, use will be made
of the spectral domain transmission line analogue of the multilayer, illustrated in Fig. 1b, whose
parameters will be defined in due course.
The multilayer may optionally be shielded below and/or above by impenetrable planes that
present three possible boundary conditions: 1) n̂ × E = Zs n̂ × (n̂ × H), where n̂ is the interface
normal vector and Zs is a specified surface impedance, 2) n̂ × E = 0, which defines a perfect
electric conductor (PEC), and 3) n̂ × H = 0, which defines a perfect magnetic conductor (PMC).
If the multilayer stack is unshielded, virtual interfaces are placed at z1 and zN +1 , and all sources
and field points are assumed to lie between or on these planes.
1
The ejωt time convention is used.
2
Throughout this paper vectors are denoted by boldface letters, unit vectors are distinguished by carets, and dyadics
(i.e., second-rank Cartesian tensors) are denoted by doubly underlined boldface letters.
3
Note that optical axis of the medium is assumed to be normal to the stratification.

2
 (a) (b)

Figure 1: Electric and magnetic currents in a layered medium. (a) Physical configuration.
(b) Spectral domain transmission-line equivalent circuit.

3
 The electric and magnetic fields due to arbitrary electric and magnetic current distributions can
be expressed in the form [15, 18]

E(r) = GEJ (r|r 0 ); J (r 0 ) + GEM (r|r 0 ); M (r 0 ) , (3)

H(r) = GHM (r|r 0 ); M (r 0 ) + GHJ (r|r 0 ); J (r 0 ) , (4)

where GP Q (r|r 0 ) is the dyadic Green function (or kernel) relating P -type fields at r and Q-type
currents at r 0 .4 As a consequence of the translational symmetry of the medium with respect to the
transverse coordinates,
GP Q (r|r 0 ) ≡ GP Q (ρ−ρ0 ; z|z 0 ) , (5)
where ρ is the projection of r on the xy−plane.5 The dyadic kernels possess the reciprocity
properties
T
GEJ (r|r 0 ) = GEJ (r 0 |r) ,

(6)
T
GHM (r|r 0 ) = GHM (r 0 |r) ,

(7)
T
GHJ (r|r 0 ) = − GEM (r 0 |r) .

(8)
To simplify notation, we may also express (3)–(4) as6

E[J , M ] = GEJ ; J + GEM ; M , (9)

H[J , M ] = GHM ; M + GHJ ; J . (10)

Since the operators E[J , 0] and H[0, M ] comprise hyper-singular kernels, it is usually advan-
tageous to express them in the mixed-potential form7

E[J , 0] = − jωA − ∇Φ , (11)

H[0, M ] = − jωF − ∇Ψ , (12)

where
A = µ0 G A ; J , (13)
F = 0 GF ; M , (14)
1
GΦ , q ,
Φ= (15)
0
1
Ψ= GΨ , m . (16)
µ0
In the above, µ0 and 0 denote the free-space permeability and permittivity, A and F are the
magnetic and electric vector potentials, Φ and Ψ are the electric and magnetic scalar potentials,
4
The notation , is used for integrals of products of two functions separated by the comma over their common
spatial support, with a dot over the comma indicating a dot product. Source coordinates are distinguished by primes.
5
Note that this property does not hold in laterally shielded environments.
6
The dependence on r and r 0 will not be indicated when there is no danger of confusion.
7
The “mixed-potential” designation was originally introduced to emphasize that the equations comprise both vector
and scalar potentials [12].

4
and q and m are the electric and magnetic charge densities, respectively. The charge and current
densities are related by the continuity equations

∇ · J = −jωq , ∇ · M = −jωm . (17)

Written more explicitly, equations (11)–(12) become

1
E[J , 0] = − jωµ0 GA ; J + ∇ GΦ , ∇0 · J , (18)
jω0
1
H[0, M ] = − jω0 GF ; M + ∇ GΨ , ∇0 · M , (19)
jωµ0
where the primed operator nabla indicates differentiation with respect to the source coordinates.
The mixed-potential expressions are less singular than the field forms, because some of the
derivatives are extracted from the kernels and transferred to the source functions. Furthermore, in
integral equation solutions by the method of moments, weak forms of these equations are used, in
which the gradient operators can be transferred to the test functions.8 Hence, assuming that J and
M are surface currents residing on a smooth surface S, we obtain the weak forms of (3)–(4) as
1
Λ; E = − jωµ0 Λ; GA ; J + ∇· Λ, GΦ , ∇0 · J
jω0
(20)
1
+ Λ; −GEM ; M ± Λ; n̂ × M ,
2
1
Λ; H = − jω0 Λ; GF ; M + ∇· Λ, GΨ , ∇0 · M
jωµ0
(21)
HJ 1
+ Λ; −G ; J ∓ Λ; n̂ × J ,
2
where n̂ is a unit vector normal to S and Λ is a vector test function tangential to S. Here, the
upper signs result if S is approached from the side into which the surface normal n̂ points, and the
lower signs result if S is approached from the opposite side [20]. The symbol − , indicates that
the integral is evaluated in the principal value sense, i.e., it represents contributions from all points
of S, excluding the isolated point r 0 = r. Similar weak forms can also be derived for sources
distributed over a volume V , rather than surface S.
The purpose of this project is to devise and implement the most efficient method for the com-
putation of the dyadic Green functions (kernels) that arise in the field and mixed-potential repre-
sentations of the electromagnetic fields.

3 Spectral-domain transmission-line analogue

The layered medium problem of Fig. 1a has a transmission-line network analogue shown in Fig. 1b,
which can be derived by a Fourier transformation of the Maxwell’s equations with respect to the
transverse coordinates [18]. The transverse spectral wavenumber will be denoted by kρ . In general,
two transmission-line networks arise, one associated with transverse-magnetic (TM) waves, and
8
The term “weak form” means that the equation is enforced in a weighted average sense [19].

5
one associated with transverse-electric (TE) waves. The quantities pertaining to these two networks
will be distinguished by the superscripts e and h, respectively. Hence, the superscript α in Fig. 1b
and in what follows stands for either e or h. Each transmission-line section of these networks is
characterized by the propagation constants

α
q
2 2 1q 2 α 2
kz = k0 nt − kρ /ν = α k0 (neff ) − kρ2 ,
2 α (22)
λ
and characteristic impedances
η0 kze η0 k0 µt
Ze = , Zh = , (23)
k0 t kzh

where
√ √ √
nt = t µt , neeff = z µt , nheff = t µz , (24)
and √
z µz
νe = , νh = , λα = να . (25)
t µt
It is important to note that, unlike in isotropic media, the propagation constants of the TM and TE
transmission-line networks are in general different when the media are uniaxial. Since the media
can be lossy, we assume that

t = 0t − j00t , 0t > 0 , 00t ≥ 0 , (26)

and similarly for z , µt , and µz . When computing the square roots in (22)–(24), the branch with
non-positive imaginary part is selected.9 The anisotropy ratios λα are assumed to be real-valued.
The medium is positive-uniaxial (electrically or magnetically), if λα > 1, and negative-uniaxial,
if λα < 1. All media-dependent parameters in (22)–(26) should carry the layer index subscript,
which we have omitted for notational simplicity.
A transmission-line section with index n, comprising unit-strength voltage and current sources
← →
at z 0 , is illustrated in Fig. 2, where Γn and Γn are the voltage reflection coefficients looking out

Figure 2: Voltage and current sources in a transmission line section.

9
The intrinsic FORTRAN square root selects the branch with positive real part; if the real part is zero, it selects the
root with positive imaginary part.

6
of the left and right terminals, respectively.10 The left-looking (or down-looking) reflection coeffi-
cients are given by the recurrence relation [18]
←
← Γn,n+1 + Γn e−j2θn
Γn+1 = ← , (27)
1 + Γn,n+1 Γn e−j2θn

where
θn = kzn dn (28)
and
Zi − Zj
Γi,j = (29)
Zi + Zj
is the Fresnel reflection coefficient across an interface between two half-spaces filled with me-
dia i and j, looking from medium j. The recursion (27) is applied beginning with the leftmost
transmission-line section and proceeds forwards, section by section, toward the right end. The
←
starting value, Γ1 , is easily found. For example, if the first layer is of infinite extent, then it is
reflectionless (the transmission line section is terminated by a matched load impedance), and the
←
starting value is Γ1 = 0. If the first layer is backed by a plate with surface impedance Zs , as
illustrated in Fig. 1, then
← Zs − Z1
Γ1 = , (30)
Zs + Z1
which becomes −1 if the plate is PEC, and +1 if it is PMC. For a good, but not perfect electrical
conductor of conductivity σ, Zs can be found as
r
1+j 2
Zs = , δ= , (31)
σδ ωµ0 σ
where δ is the skin depth. The right-looking (or up-looking) reflection coefficients satisfy the
recurrence relation →
→ Γn,n−1 + Γn e−j2θn
Γn−1 = → , (32)
1 + Γn,n−1 Γn e−j2θn
which is applied beginning with the rightmost transmission-line section and proceeds backwards,
→
toward the left end. The starting value, ΓN , is easily found, as explained in connection with the
forward recursion.
We now introduce four basic transmission-line Green functions (TLGFs), as listed in Table 1.11
It can be shown [21, p. 194] that these TLGFs possess the reciprocity properties12

Vi (z|z 0 ) = Vi (z 0 |z) , (33)

Iv (z|z 0 ) = Iv (z 0 |z) , (34)
Vv (z|z 0 ) = −Ii (z 0 |z) . (35)
10
Since this figure and the following discussion pertain to both the TM and TE transmission-line networks, we omit
the superscript α for simplicity.
11
For simplicity, we do not explicitly indicate the dependence of TLGFs on the transverse spectral wavenumber kρ .
12
These relationships are intimately related to the reciprocity properties of the dyadic kernels given in (6)–(8).

7
 Table 1: Definition of four transmission-line Green functions (TLGFs).

TLGF # Notation Definition

1 Vi (z|z 0 ) voltage at z due to unit-strength current source at z 0

2 Iv (z|z 0 ) current at z due to unit-strength voltage source at z 0

3 Ii (z|z 0 ) current at z due to unit-strength current source at z 0

4 Vv (z|z 0 ) voltage at z due to unit-strength voltage source at z 0

Explicit expressions for the TLGFs can readily be derived from their governing differential equa-
tions [18]. In particular, assuming that z and z 0 lie within the same line section n, and the source
is of the current type, it is found that
" 4
#
Zn 0 1 X (s)
Vi (z|z 0 ) = e−jkzn |z−z | + R(s) e−jkzn ζn , (36)
2 Dn s=1 n
" 4
#
1 0 1 X (s)
Ii (z|z 0 ) = ±e−jkzn |z−z | − (−1)s Rn(s) e−jkzn ζn , (37)
2 Dn s=1
where the ± signs correspond to z ≷ z 0 , and where
← →
Dn = 1− Γn Γn e−j2θn (38)

is the resonant denominator and

← → ← →
Rn(1) = Γn , Rn(2) = Γn , Rn(3) = Rn(4) = Γn Γn , (39)

ζn(1) = (z + z 0 ) − 2zn , ζn(2) = 2zn+1 − (z + z 0 ) ,

(40)
ζn(3) = 2dn + (z − z 0 ) , ζn(4) = 2dn − (z − z 0 ) .
When the inverse resonant denominator is expanded in a geometric series as,
∞ 
1 X ← → m
= Γn Γn e−j2mθn , (41)
Dn m=0

the expressions (36)–(37) may be given a geometric-optical ray interpretation, as illustrated in

Fig. 3, where ray 0 represents the direct term and rays 1 through 4 represent four categories of
rays that undergo partial reflections at the upper and lower layer boundaries before reaching the
observation point [18]. It should be noted that the rays are strictly one-dimensional; the two-
dimensional depiction is for convenience of visualization only. Although we do not use the ray
forms to compute the TLGFs, the ray interpretation will become useful when we discuss the quasi-
static image extraction in a later section.

8
 Figure 3: Ray interpretation of the TLGFs when z and z 0 are in the same layer n.

When z 0 is in section n and z is in section m > n, we introduce the voltage transfer function
between the right terminals of the source section and the left terminals of the observation section,
given as [18]
m−1
→ Y →
Tnm = τk , (42)
k=n+1

where →
→ Vk+1 (1+ Γk )e−jθk
τ k≡ = → (43)
Vk 1+ Γk e −j2θ k

is the right-looking voltage transmission coefficient across the kth section. Here, we use the nota-
tion Vk = V (zk ). The voltage at z within the mth section can now be found as [18]
" → #
−j2kzm (zm+1 −z)
→ 1+ Γm e
Vm (z) = Vn+1 Tnm e−jkzm (z−zm ) → . (44)
1+ Γm e−j2θm

With Vn+1 computed as Vi (zn+1 |z 0 ) using (36), the above expression represents the TLGF Vi (z|z 0 )
for m > n.
A dual expression to (44), obtained by replacing the voltages by currents and the reflection
coefficients by their negatives in (42)–(44), gives Im (z) in terms of In+1 . With the latter computed
as Ii (zn+1 |z 0 ) using (37), this expression represents the transmission-line Green function Ii (z|z 0 )
for m > n.
The remaining TLGFs, Iv and Vv , are dual to Vi and Ii , respectively, and can be derived by
simply replacing in the latter the characteristic impedances by their reciprocals and the reflection
coefficients by their negatives [18, 22].
We have thus obtained all four TLGFs for m ≥ n. There is no need to derive formulas for the
case m < n, as we can use the reciprocity properties (33)–(35) to swap the locations of z and z 0 .
It is important to note the large−kρ asymptotic behavior of the TLGFs, because it determines
the source-region singularities of the kernels in the space domain and it can affect the computation
of the Sommerfeld integrals. Clearly, the worst case arises when z = z 0 , since otherwise the
TLGFS decay exponentially. We summarize this worst-case behavior in Table 2.
The above formulation of the TLGFs, which is based on the reflection coefficient method (and
closely related to the impedance method), has the advantage that all exponentials are non-growing,
thus eliminating the danger of numerical overflows for large kρ . Other formulations, such as the

9
 Table 2: Worst-case asymptotic TLGF behavior.

TLGF Asymptotic behavior

for z = z 0
Vie , Ivh O(kρ )

Vve , Iih , Vvh , Iie O(1)

Vih , Ive O(kρ−1 )

ABCD-matrix method, or the T-matrix method [18] are also possible and useful in some applica-
tions. For example, in the development of the closed-form representations of layered media Green
functions, it is important to extract the pole singularities, which can be found as zeros of a source-
free modal dispersion function of the multilayer. An important property of the matrix methods is
that they lead to analytic, pole-free modal dispersion functions, which are amenable to the root
finding techniques based on the Cauchy theorem [23, 24].

4 Dyadic Green functions (kernels)

The scalar potential kernel and the elements of the dyadic vector potential kernel in (18) can be
expressed as13
1
− GΦ (ρ; z|z 0 ) = η0 G0 (ρ; z|z 0 ) , (45)
jω0
0 0 0
jωµ0 GA A 2
xx (ρ; z|z ) = jωµ0 Gyy (ρ; z|z ) = η0 k0 G1 (ρ; z|z ) , (46)
0 0
jωµ0 GA 2
zz (ρ; z|z ) = η0 k0 G2 (ρ; z|z ) , (47)
( A ) ( )
Gzx (ρ; z|z 0 ) −j cos ϕ
jωµ0 A 0
= η0 k02 G3 (ρ; z|z 0 ) , (48)
Gzy (ρ; z|z ) −j sin ϕ
( A ) ( )
Gxz (ρ; z|z 0 ) −j cos ϕ
jωµ0 0
= η0 k02 G4 (ρ; z|z 0 ) . (49)
GA
yz (ρ; z|z ) −j sin ϕ
The elements of the two field kernels in (9) can be written as
0
− GEJ
xx (ρ; z|z ) η0 k02 h 0 0
i
0
= G5 (ρ; z|z ) ± cos 2ϕ G6 (ρ; z|z ) , (50)
− GEJ
yy (ρ; z|z )
2

0 η0 k02
0
− GEJ = −
xy (ρ; z|z ) GEJ
yx (ρ; z|z )
= sin 2ϕ G6 (ρ; z|z 0 ) , (51)
2
EJ 0
( )
− Gxz (ρ; z|z ) j cos ϕ
EJ 0
= η0 k02 G7 (ρ; z|z 0 ) , (52)
− Gyz (ρ; z|z ) j sin ϕ
13
Only nonzero Cartesian components of the dyadic kernels are listed.

10
 0
( )
− GEJ
zx (ρ; z|z ) j cos ϕ
0
= η0 k02 G8 (ρ; z|z 0 ) , (53)
− GEJ
zy (ρ; z|z ) j sin ϕ
0 η0
− GEJ 2
zz (ρ; z|z ) = η0 k0 G9 + δ(ρ)δ(z − z 0 ) , (54)
jk0 z
0
− GEM
xx (ρ; z|z ) k02
0
= ∓ sin 2ϕ G11 (ρ; z|z 0 ) , (55)
− GEMyy (ρ; z|z )
2
0
− GEMxy (ρ; z|z ) k02 h 0 0
i
0
= ± G 12 (ρ; z|z ) + cos 2ϕ G 11 (ρ; z|z ) , (56)
− GEMyx (ρ; z|z )
2
0
( )
− GEM
xz (ρ; z|z ) +j sin ϕ
EM 0
= k02 G13 (ρ; z|z 0 ) , (57)
− Gyz (ρ; z|z ) −j cos ϕ
0
( )
− GEM
zx (ρ; z|z ) −j sin ϕ
EM 0
= k02 G14 (ρ; z|z 0 ) . (58)
− Gzy (ρ; z|z ) +j cos ϕ
√ p
In the above, k0 = ω 0 µ0 and η0 = µ0 /0 are the free-space wavenumber and intrinsic
impedance, respectively, and Gi are basic kernels, which will be introduced in due course.
In view of the duality between (18) and (19), the potential kernels in (19) can be obtained from
(45)–(49) by the substitutions Φ → Ψ, A → F , µ0 → 0 , 0 → µ0 , and by replacing the basic
kernels by the corresponding dual kernels. Similarly, the expressions for the first field kernel in
(10) can be obtained from (50)–(54) by the substitutions EJ → HM , η0 → 1/η0 , and by using
the dual forms of the basic kernels. Finally, the second field kernel in (10) can be computed from
(55)–(58) by using the reciprocity relationship (8).
The above expressions assume that the source point is on the z axis of the cylindrical coordinate
system (ρ, ϕ, z). In the general case, ρ −→ ρ − ρ0 , which implies the substitutions
y − y0
 
p
0 2 0 2
ρ −→ (x − x ) + (y − y ) , ϕ −→ arctan , (59)
x − x0
where the quadrant of the argument must be noted when the inverse tangent is evaluated.14
The basic kernels can be expressed as15
 h
(Vi − Vie )
 h
(Vi − Vie )
 
0 1
G0 (ρ; z|z ) = S0 = S0 , (60)
η0 kρ2 η0 (kρ /k0 )2
 h   h
0 Vi 1 Vi
G1 (ρ; z|z ) = S0 2
= S0 , (61)
η0 k0 η0

µt µ0t η0 Ive h e
  
0 0 η0 (Iv − Iv )
G2 (ρ; z|z ) = S0 + + µt µt (62)
0z z k02 kρ2
µt µ0t h e
  
1 e 0 η0 (Iv − Iv )
= S0 + η0 Iv + µt µt , (63)
0z z (kρ /k0 )2
14
The intrinsic FORTRAN function ATAN2 does this automatically and returns a result in the range (−π, π] [25,
p. 178].
15
Note that there is no kernel G10 .

11
 (Iih − Iie )
 
0
= S01 µt (Iih − Iie ) ,

G3 (ρ; z|z ) = S1 µt (64)
k0 kρ
h e
 
0 0 (Vv − Vv )
= S01 µ0t (Vvh − Vve ) ,

G4 (ρ; z|z ) = S1 µt (65)
k0 kρ
 h
(Vi + Vie )
 h
(Vi + Vie )
 
0 1
G5 (ρ; z|z ) = S0 = S0 , (66)
η0 k02 η0
 h
(Vi − Vie )
 h
(Vi − Vie )
 
0 1
G6 (ρ; z|z ) = S2 = S2 , (67)
η0 k02 η0
kρ Vve
   e
0 2 Vv
G7 (ρ; z|z ) = S1 3 0
= S1 , (68)
k0 z 0z
kρ Iie
   e
0 I
G8 (ρ; z|z ) = S1 3
= S1 i ,
2
(69)
k0 z z
 2 e
kρ η0 Iv (kρ /k0 )2 η0 Ive
 
0 1
G9 (ρ; z|z ) = S0 = S0 , (70)
k04 0z z 0z z
 h
(Vv − Vve )

0 1
 h
G11 (ρ; z|z ) = S2 2
= S 2 (Vv − Vve ) , (71)
k0
 h
(Vv + Vve )

0
= S10 (Vvh + Vve ) ,

G12 (ρ; z|z ) = S0 2
(72)
k0
kρ Vih
   h 
0 2 Vi
G13 (ρ; z|z ) = S1 3 0
= S1 , (73)
k0 η0 µz η0 µ0z
kρ η0 Ive
   e
0 2 η0 Iv
G14 (ρ; z|z ) = S1 3
= S1 . (74)
k0 z z
Similarly, the kernels dual to (60)–(70) can be expressed as16

η0 (Ive − Ivh ) η0 (Ive − Ivh )

   
0 0 1
G0 (ρ; z|z ) = S0 = S0 , (75)
kρ2 (kρ /k0 )2

η0 Ive
 
G01 (ρ; z|z 0 ) = S0 = S10 {η0 Ive } , (76)
k02

0t
Vih e h
  
t 0 (Vi − Vi )
G02 (ρ; z|z 0 )
= S0 + + t t (77)
η0 k02
µ0z µz η0 kρ2
0t Vih e h
  
1 t 0 (Vi − Vi )
= S0 + + t t , (78)
µ0z µz η0 η0 (kρ /k0 )2

(Vve − Vvh )
 
0 0
= S01 t (Vve − Vvh ) ,

G3 (ρ; z|z ) = S1 t (79)
k0 kρ
16
Note that we distinguish the dual kernels by primes.

12
 e h
 
0 (Ii − Ii )
G04 (ρ; z|z 0 ) = S01 0t (Iie − Iih )

= S1 t , (80)
k0 kρ
η0 (Ive + Ivh )
 
0 0
= S10 η0 (Ive + Ivh )

G5 (ρ; z|z ) = S0 2
, (81)
k0
η0 (Iv − Ivh )
e
 
0 0 1
η0 (Ive − Ivh )

G6 (ρ; z|z ) = S2 2
= S 2 , (82)
k0
kρ Iih
   h
0 0 2 Ii
G7 (ρ; z|z ) = S1 3 0
= S1 , (83)
k0 µz µ0z
kρ Vvh
   h
0 0 2 Vv
G8 (ρ; z|z ) = S1 3
= S1 , (84)
k0 µz µz
( )
2 h
k V (kρ /k0 )2 Vih
 
0 0 ρ i 1
G9 (ρ; z|z ) = S0 = S0 . (85)
k04 η0 µ0z µz η0 µ0z µz
It is understood that the primed and unprimed media parameters are to be evaluated in the source
and field layers, respectively. The symbols Sn and Sm n denote the integral operators
Z ∞
 1 
Sn · = · Jn (kρ ρ)kρ dkρ (86)
2π 0
and Z ∞  m  
m
 1  kρ kρ
Sn · = · Jn (kρ ρ) d , (87)
2π 0 k0 k0
respectively, where Jn is the Bessel function of order n. The integrals arising in (86)–(87) are
referred to as Sommerfeld integrals and are discussed in more detail in the next section. The
representation of the kernels in terms of Sm n is convenient in the development of closed-form
kernel representations. In Table 3 we list the combinations of n and m that occur in the basic
kernels.

Table 3: Valid combinations of n and m.

n 0 1 2
m 1 0, 2 1

Note that 24 distinct basic and dual kernels are needed in general to determine the electric and
magnetic fields due to electric and magnetic currents.

5 Direct computation of Sommerfeld integrals

As shown in the previous section, all basic and dual kernels can be expressed as Sommerfeld
integrals (also known as Hankel transforms). The integrals that must be computed have the form
Z ∞
n
0
o 1
Sn G̃(kρ |z, z ) = G̃(kρ |z, z 0 )Jn (kρ ρ)kρ dkρ (88)
2π 0
Z ∞
1
= G̃(kρ |z, z 0 )Hn(2) (kρ ρ)kρ dkρ , n = 0, 1, 2, (89)
4π ∞e−jπ

13
where the spectral-domain kernel G̃ is an even (odd) function of kρ for even (odd) n,17 Jn is the
(2)
Bessel function and Hn the Hankel function of the second kind, both of order n. The second
(1) (2)
integral form (89) can be derived from (88) by using the formulas Jn (z) = 21 [Hn (z) + Hn (z)]
(1) (2)
and Hn (z ejπ ) = − e−jnπ Hn (z) [26].18 The location of the integration path in (89) with respect
to the pole and branch-point singularities is illustrated in Fig. 4. Note that only the positive-real

Figure 4: Sommerfeld integration path in the complex kρ −plane.

axis path is used in (88). The branch point at the origin is due to the Hankel function and it only
occurs in the second integral form (89). The branch points at ±k1 and ±kN are only present if
the outermost layers are half-spaces, i.e., the multilayer is unshielded from below and above [27,
p. 112]. The real-axis integration path in Fig. 4 can be deformed, as long as it does not cross any
of the singular points.
Since the integration paths in (88)–(89) extend to infinity, it is important to understand the
large−kρ asymptotic form of the spectral domain kernels. Upon inspection of the relevant expres-
sions, we find the dominant behavior as

G̃(kρ ; z|z 0 ) ∼ kρµ e−kρ ζ , kρ → ∞ , (90)

where ζ and µ are coefficients that can be easily determined from the known asymptotic behavior
of the TLGFs involved.19 In isotropic media, ζ is typically the vertical distance between the source
and the field point, |z − z 0 |. For kernels G3 , G4 , and G11 , the direct terms cancel when z and z 0 lie
17
For simplicity, we omit the kernel index i.
18
Strictly speaking, the integral in (89) should be taken in the principal value sense with respect to the origin.
However, the small-argument behavior of the spectral domain kernels is such that the singularity of the Hankel function
at the origin is always cancelled.
19
The power coefficient µ is unrelated to the same symbol used to denote permeability.

14
within the same isotropic layer n, in which case the coefficient ζ is the smaller of (z + z 0 ) − 2zn
and 2zn+1 − (z + z 0 ). In uniaxial media, since
α
kzn ∼ −jkρ /λαn (91)

as kρ → ∞, ζ is the vertical distance between the source and the field point, scaled by the
anisotropy ratios of the layers. For example, if z and z 0 are within the same layer n, then ζ =
|z − z 0 |/λαn . The value of µ readily follows form Table 2. Taking into account the large-argument
form of the Bessel function, we find that the asymptotic behavior of the entire integrand in (88) is
given by (90) with µ → µ − 0.5. Clearly, the worst case arises for ζ = 0, when the exponential
decay factor is absent and the integrand is oscillatory, and possibly divergent.
To compute the Sommerfeld integrals, we employ the integration-then-summation procedure
[1], applied to the first integral form (88). As illustrated in Fig. 5, the integration path is deformed
into the first quadrant of the kρ −plane to avoid the branch-point and pole singularities that may lie
on or close to the real axis [1, 28–30]. The exact shape of the initial path segment is not critical,

Figure 5: Deformed integration path with break points.

and we use a half-sine shape, with a/k0 = nmax + 1, where nmax is the maximum real part of
the effective refractive indices nαeff of the multilayer, given by (24). The maximum sine height d
is limited by the exponential growth of the Bessel function when kρ departs from the real axis.
Hence, we set20  
min 1, 1

, if ρ > ζ 0 ,
d/k0 = k0 ρ (92)
1, otherwise ,


where ζ 0 = Re{ζ}. From the point a on, the integration follows the real axis, using the break
points ξi , i = −1, 0, 1, . . . , which are selected, as follows. If ρ > ζ 0 , we use consecutive, approxi-
mate Bessel function extrema, computed as average values of the adjacent Bessel function zeros,21
with ξ−1 set to the first extremum exceeding a; otherwise, we use uniformly spaced break points
separated by a fixed-length interval ∆ξ = π/ζ 0 , with ξ−1 = a + min(a, ∆ξ).22
20
This is admittedly a somewhat arbitrary choice and some other values may work as well.
21
Tabulated values are used for the first ten Bessel function zeros, and asymptotic values otherwise [26].
22
Again, these are somewhat arbitrary choices, which have been found to work well.

15
 We may now express a typical Sommerfeld integral as
n o
Sn G̃(kρ |z, z 0 ) = Ia + Ib + I∞ , (93)

where the right-hand terms are defined as follows. Ia is the integral along the deformed path
segment, given as Z a  
1 0 dkρ
Ia = G̃(kρ ; z|z )Jn (kρ ρ)kρ dt , (94)
2π 0 dt
where
kρ = t + jd sin(πt/a) (95)
and t is a real parameter. Ib is the real-axis integral from a to the first break point, given as
Z ξ−1
1
Ib = G̃(kρ ; z|z 0 )Jn (kρ ρ)kρ dkρ . (96)
2π a
I∞ is the real-axis tail integral, given as
Z ∞ ∞
1 0
X
I∞ = G̃(kρ ; z|z )Jn (kρ ρ)kρ dkρ = ui , (97)
2π ξ−1 i=0

where Z ξi
1
ui = G̃(kρ ; z|z 0 )Jn (kρ ρ)kρ dkρ . (98)
2π ξi−1

The integrals Ia and Ib are computed by an adaptive quadrature based on the Patterson rule [31],
and the partial integrals ui are computed by a Gauss-Legendre quadrature23 of order 16.
As indicated in (97), the tail integral is evaluated as an infinite sum of the partial integrals (98).
It can be shown [1] that for ρ > 0, the series of partial integrals is asymptotically alternating, with
logarithmic convergence for ζ = 0 and linear convergence for ζ 6= 0. For ρ = 0 and ζ 6= 0, the
convergence is monotone linear. We compute this series as the limit of the sequence of partial
sums n
X
sn = ui , (99)
i=0

as n → ∞. To accelerate the convergence, we use the weighted-averages extrapolation method,

given as [1, 2, 32]
(k−1) (k−1) (k−1)
sn + ηn sn+1
s(k)
n = (k−1)
, n ≥ 0, k ≥ 1, (100)
1 + ηn
with the weights
 2k−α
ξn+1
ηn(k) =± e(ξn+1 −ξn )ζ , (101)
ξn
23
Paradoxically, using the adaptive Patterson quadrature for the partial integrals can be unreliable in the case of
exponential decay, because the integrands can sometimes become negligibly small and the adaptive rule can fail trying
to integrate what is essentially numerical noise.

16
where α = µ + 1/2.24 Here, the upper and lower signs apply to the alternating (ρ > ζ 0 ) and
monotone (ρ ≤ ζ 0 ) convergence cases, respectively. The recursion (100) represents a triangular
scheme, which after the third step has the form
(0) (1) (2) (3)
s0 s0 s0 s0 ·
(0) (1) (2)
s1 s1 s1 ·
(0) (1)
s2 s2 · (102)
(0)
s3 ·
·
(0)
where in the first column sn = sn . At step i, the computations start at the lowest table element
and proceed upwards along the counterdiagonal i = n + k. Note that only a one-dimensional array,
storing the counterdiagonal elements, is needed to implement this scheme, as the newly-generated
entries can overwrite the no longer needed elements from the previous iteration. Hence, given the
(i)
partial sums s0 , s1 , . . . , si , the weighted-averages algorithm produces s0 as the best approximation
of the tail integral I∞ . The recursion is applied with an increasing i until the outcomes of two
consecutive iterations do not differ by more than a prescribed small error threshold.
The rigorous integration procedure outlined here has been implemented to provide benchmark
results for the complex image method described in the following sections. Hence, with a few
exceptions, the spectral functions are integrated directly, with no pre-processing. For kernels G7 ,
G8 , and their dual kernels, convergence problems were encountered for k0 ρ < 10−2 when z 0 and z
both lie on the same interface between dissimilar layers. These were remedied by subtracting from
the spectral domain kernels their leading quasi-static terms, which integrate to zero.
The approach presented here is only recommended for small to moderate values of k0 ρ, up to
about 102 . This limitation is due to the rapid integrand oscillations on the initial path segment
in Fig. 5 when kρ ρ becomes large in magnitude, and the fact that in this case the path passes
very near the real-axis singularities of the spectral domain kernels, as a consequence of (92). The
range of applicability of this method can be extended by choosing different integration paths. For
example, using the second integral form (89) as the point of departure, one can wrap the integration
contour around the branch cuts, where the integrands are rapidly convergent for large ρ [33, 34].
Furthermore, one can deform the integration path into a steepest descent path, where the integrand
converges exponentially. These methods, however, require the tracking of any poles (including
improper ones) that may be captured by the path deformation [35–37].

6 Application of the discrete complex image method

The hight cost of the direct integration of the Sommerfeld integrals has prompted us to also imple-
ment the so-called discrete complex image method (DCIM) [3–5, 7, 9, 34, 38–42]. As mentioned
at the end of Sec. 4, in the context of DCIM we use the representation
n o
G(ρ; z|z 0 ) = Sm
n G̃(kρ ; z|z 0
) , (103)
24
In the above, n is unrelated to the same symbol used for the order of the Bessel function or layer index. Also, η
must not be confused with the same symbol used for intrinsic impedance, and α is unrelated the same symbol used to
distinguish TM and TE waves.

17
where G̃ is now referred to as the spectral domain kernel. The key step in this approach is the
exponential fit of G̃ along a specified path in the complex spectral plane. The exponential terms,
which may be interpreted as contributions due to complex images of the point source, can be
inverted in closed form, resulting in an efficient procedure. The robustness of this method is
usually improved if the quasi-static form of the kernel, corresponding to kρ → ∞, is first extracted
and handled analytically. Similarly, it is also desirable to extract the guided wave terms and branch
point terms from the spectral kernel before the exponential fit is performed. Hence, the exponential
fit is usually applied to the remainder kernel after all important physical constituents have been
subtracted out.

6.1 Extraction of quasi-static terms

For each spectral-domain kernel G̃, we wish to find its quasi-static form G̃q , which possesses the
property that G̃ ∼ G̃q as kρ → ∞, and which can be transformed to the space domain in closed
form. Since the spectral-domain kernels in (60)–(85) comprise the TLGFs, G̃q is easily found,
once the quasi-static forms of the TLGFs are determined. As alluded to in Sec. 3, the TLGFs
could also be computed by summing the infinite number of rays that originate at the source and
undergo multiple reflections and transmissions at the interfaces between layers before reaching the
observation point. Here, we are only interested in extracting a small number of dominant rays,25 in
the quasi-static regime that corresponds to kρ → ∞. In this regime, the characteristic impedances
of layer n behave as
Zne kρ Znh jk0 κhn
∼ , ∼ , (104)
η0 jk0 κen η0 kρ
where we have introduced the notation
√ √
κen = tn zn , κhn = µtn µzn . (105)

Using the above, it is found that the Fresnel reflection coefficients at an interface between layers j
and i behave as
e
κei − κej h
κhi − κhj
Γi,j ∼ − e , Γi,j ∼ h , (106)
κi + κej κi + κhj
with the corresponding transmission coefficients given as
α
Ti,j = 1 + Γαi,j . (107)

To make a closed-form inversion of the extracted quasi-static rays possible, we define an

“equivalent medium” of a multilayer as a homogeneous, isotropic medium characterized by the
wavenumber kq = k0 nq , where nq is set to the effective refractive index nαeff,n with a minimum real
part.26 The corresponding propagation constant is thus
q
kzq = kq2 − kρ2 , (108)
25
The dominant rays are those with the shortest paths of travel, and thus having the slowest decay rates for large kρ .
26
Perhaps a more obvious choice would be to set nq to the transverse refractive index of the field-point layer.
However, numerical experiments (not reported here) have shown that the present choice results in a better performance
of the DCIM.

18
with jkzq ∼ kρ as kρ → ∞. Note that a wave traveling a distance d in a TL section corresponding
to layer n gives rise to the exponential factor
α α α
e−jkzn d ∼ e−jkzq d/λn ∼ e−kρ d/λn . (109)
Hence, in the quasi-static problem the distances along z are scaled by the inverse anisotropy ratios,
as previously indicated in connection with (91). Note the asymptotic equivalence
α
kzn ∼ kzq /λαn . (110)
We describe the quasi-static term extraction with reference to the TLGFs (36)–(37) and illus-
trate the ray tracing procedure in Fig. 6.27 For each interface i, we pre-compute the upward-looking
TM and TE static Fresnel reflection coefficients Riα , based on (106).28 We assume that the source
is in layer n of the multilayer. Since the factors Zn /2, Yn /2, or 1/2 are common to all rays, they
are omitted in the ray tracing, but are restored (in their quasi-static form) at a later stage.

(a) (b)
Figure 6: Illustration of the ray tracing for quasi-static term extraction. (a) The up-going
incident ray originating at the source gives rise to an up-going transmitted ray and a down-
going reflected ray at the upper interface. Similarly, the down-going incident ray generates
an up-going reflected ray and a down-going transmitted ray at the lower interface. (b) The
ray with the shortest travel path—assumed to be the up-going ray at the upper interface of
the source layer—is propagated to the next interface, where it creates an up-going trans-
mitted ray and a down-going reflected ray. A quasi-static term is extracted (indicated by a
circle), since the up-propagated ray crosses the field-point level.

The ray tracing begins by launching from the source in layer n two rays, one going up, with
amplitude +1, and one going down, with amplitude +1 for Vi and Iv , and −1 for for Ii and Vv .29
27
A similar procedure described in [10] assumes that the source and the field point lie on interfaces of the multilayer.
Hence, if in the original layered medium the source or the field point is not situated at an interface, the layer containing
it must be split into two layers with the same material properties, by introducing a virtual interface. This approach
becomes cumbersome when z or z 0 are varying.
28
The down-looking reflection coefficients are simply the negatives of Riα . It is worth noting that for nonmagnetic
media Rih = 0 at all dielectric interfaces.
29
For simplicity, where there is no danger of confusion, we omit the superscript α.

19
If the observation point is in the source layer, the direct ray is first extracted,30 with amplitude +1
for Vi and Iv , and +1 or −1 for Ii and Vv , depending on whether z > z 0 or z < z 0 , respectively.31
The quasi-static path traveled by this ray is |z − z 0 |/λn . The up-going ray hits interface n + 1,
where it gives rise to a reflected ray with amplitude Rn+1 and a transmitted ray with amplitude
Tn+1 = 1 + Rn+1 .32 The quasi-static path traveled by this ray is (zn+1 − z 0 )/λn . Similarly, the
down-going ray hits interface n, where it creates reflected and transmitted rays with amplitudes
−Rn and Tn = 1 − Rn , respectively, possibly multiplied by −1 (in the case of Ii and Vv ). The
quasi-static path traveled by this ray is (z 0 − zn )/λn .
The initial four rays are propagated up and down and hit the adjacent interfaces, where they are
partially transmitted and partially reflected, each giving rise to another pair of up- and down-going
rays, whose amplitudes are modified by the appropriate static Fresnel reflection or transmission
coefficients. The phases of these rays, which correspond to the quasi-static path lengths traveled,
are accumulated and stored. This procedure is then repeated for each of the so-created rays. Each
time a ray crosses the field-point level, it is extracted.
In theory, assuming that there are at least two non-trivial interfaces present, this procedure
would continue forever, since the number of rays is infinite. To remedy this, we prioritize the rays
and terminate the procedure after extracting the desired number of dominant rays, with the shortest
travel paths between the source and the field point. To this end, each ray is assigned a key, which
is the real part of its accumulated phase. At each step of the ray tracing procedure, the ray with
the lowest key is processed first, since the lowest key corresponds to the shortest electrical path
traveled. To keep track of the rays with the highest priorities, we use the priority queue algorithm
[43], as first proposed in [10]. The priority queue algorithm exploits the so called min-heap data
structure, where each parent element has two children elements, arranged such that the keys of the
children never exceed the key of their parent. In our case, each heap element represents a ray with
all its attributes, such as the key, amplitude, phase, interface index, and the direction of travel (up
or down). The priority queue algorithm supports two operations, PUSH and POP, where PUSH
inserts an element into the heap and POP returns and removes the element with the smallest key.
Following each of these operations the heap is updated to maintain the min-heap property.
Hence, our quasi-static part extraction algorithm may now be summarized as follows. First,
the initial four rays generated by the source, as illustrated in Fig. 6(a), are PUSHed on the heap,
and the direct term is extracted if the observation point lies in the source layer. Then, the highest-
priority ray is POPed off the heap, propagated to the next interface, and the resulting reflected
and transmitted rays are PUSHed back onto the heap. If a ray crosses the field-point level, it is
extracted, as illustrated in Fig. 6(b). Should the extracted ray have the same phase as one of the
previously extracted rays, it is merged with that ray. This extraction process continues until the
desired number of quasi-static terms is reached, or the heap is exhausted.
Even though the rays are processed in the order of their priority, there is no guarantee that the
dominant rays are extracted strictly in the order of increasing path lengths. This is because the
highest priority ray entering the field-point layer may have a longer path to the field point than the
next in priority ray entering this layer from the opposite side. Furthermore, the last few extracted
rays may be incomplete, since there is a possibility that a later ray would be merged with an already
30
Here, by “extracting” a ray we mean computing and recording its amplitude and path length traveled.
31
Note that if z = z 0 , additional information is required to decide whether z = z 0 + 0 or z = z 0 − 0.
32
Here, the voltage or current reflection coefficients are implied, depending on the type of the TLGFs involved. The
current reflection coefficient is simply the negative of the corresponding voltage reflection coefficient.

20
extracted ray and modify its amplitude. Hence, in order to be reasonably sure that a sequence of
complete dominant rays is obtained, we extract several more rays than the desired number, sort
them in the ascending order of their keys, and discard the superfluous ones.
Taking Vie as an example, the above algorithm will produce the quasi-static representation
Ne
η0 kρ X e
Vie (z|z 0 ) ∼ e
aeqi e−jkzq bqi , (111)
2jκn k0 i=1

where the subscript n is the index of the source layer, Ne is the number of dominant quasi-static
rays extracted, aeqi are the ray amplitudes, and beqi are the corresponding path lengths traveled (or
phases, for short).33 The factor in front of the sum is the restored Zne /2 starting ray amplitude, with
Zne replaced by the quasi-static form given in (104).
Since many of the spectral-domain kernels in (60)–(85) comprise two TLGFs, two sets of
quasi-static terms are in general involved. In the case of G̃0 , we have

(Vih − Vie )
G̃0 ∼ 2
∼ G̃q0 , (112)
η0 (kρ /k0 )

which leads to
Ne
j 1 X e −jkzq beqi
G̃q0 = a e , (113)
2κen jkzq i=1 qi
where we used (111) and the asymptotic equivalence of kρ and jkzq . In deriving this expression
we neglected Vih in favor of the asymptotically dominant Vie (see Table 2), which was necessary,
because the quasi-static form of Vih /kρ2 does no have a closed-form inverse. In the above and
henceforth, we assume that kzq is normalized to k0 .
In the case of G̃1 , we have
Vh
G̃1 = i ∼ G̃q1 , (114)
η0
which leads to
Nh
jκhn 1 X h
G̃q1 = ahqi e−jkzq bqi , (115)
2 jkzq i=1

where the sum represents the quasi-static rays extracted form Vih . In the factor preceding the sum,
we used (110).
In the case of G̃2 , we have

η0 (Ivh − Ive )
 
µtm µtn
G̃2 ∼ + e
η0 Iv + µtm µtn 2
∼ G̃q2 , (116)
zn zm (kρ /k0 )
33
Here and in what follows, the superscripts e or h are used to distinguish the quasi-static coefficients (ray am-
plitudes and phases) associated with the TM and TE TLGFs, respectively. For simplicity, the same symbols will be
used for the quasi-static coefficients associated with the voltage- and current-type TLGFs. It will be evident from the
context which TLGF type is implied.

21
where the subscript m is the index of the field-point layer. Upon applying the quasi-static ray
extraction algorithm to the two TLGFs, we obtain
Ne
jκen 1 X
 
µtm µtn e
G̃q2 = + aeqi e−jkzq bqi
zn zm 2 jkzq i=1
Nh
(117)
jµtm µtn 1 X h
− h
ahqi e−jkzq bqi ,
2κn jkzq i=1

where the first and the second sums represent the quasi-static rays extracted from Ive and Ivh , re-
spectively. In deriving this expression, in the second term of G̃2 we neglected Ive in favor of the
asymptotically dominant Ivh (see Table 2). This was necessary, because the quasi-static form of
Ive /kρ2 does not have a closed-form inverse.
In the case of G̃3 , we have
G̃3 = µtm (Iih − Iie ) ∼ G̃q3 , (118)
which leads to !
Nh Ne
µtm X
−jkzq bh
X
−jkzq beqi
G̃q3 = ahqi e qi − aeqi e , (119)
2 i=1 i=1

where the first and second sums represent the quasi-static rays extracted from Iih and Iie , respec-
tively.
In the case of G̃4 , we have

G̃4 = µtn (Vvh − Vve ) ∼ G̃q4 , (120)

which leads to !
Nh Ne
µtn X
−jkzq bh
X
−jkzq beqi
G̃q4 = ahqi e qi − aeqi e , (121)
2 i=1 i=1

where the first and second sums represent the quasi-static rays extracted from Vvh and Vve , respec-
tively.
In the case of G̃5 , we have
(V h + Vie )
G̃5 = i ∼ G̃q5 , (122)
η0
which leads to
Ne Nh
1 X
e −jkzq beqi jκhn 1 X h
G̃q5 = e
jkzq aqi e + ahqi e−jkzq bqi , (123)
2jκn i=1
2 jkzq i=1

where the first and second sums represent the quasi-static rays extracted from Vih and Vie , respec-
tively. Note that (123) comprises two quasi-static forms, with the first asymptotically dominant.

22
Although the second form can be omitted, we retain it in order to make G̃q5 exact in the case of
isotropic, homogeneous multilayers.34 Kernels other that G5 do not require this special treatment.
In the case of G̃6 , we have

(Vih − Vie )
G̃6 = ∼ (kρ /k0 )2 G̃q6 , (124)
η0
which leads to
Ne
j 1 X e
G̃q6 = e aeqi e−jkzq bqi , (125)
2κn jkzq i=1
where the sum represents the quasi-static rays extracted from Vie . In deriving this expression,
which is identical to (113), we again neglected Vih in favor of the dominant Vie .
In the case of G̃7 , we have
Ve
G̃7 = v ∼ G̃q7 , (126)
zn
which leads to
Ne
1 X e
G̃q7 = aeqi e−jkzq bqi , (127)
2zn i=1
where the sum represents the quasi-static rays extracted from Vve .
In the case of G̃8 , we have
Ie
G̃8 = i ∼ G̃q8 , (128)
zm
which leads to
Ne
1 X e
G̃q8 = aeqi e−jkzq bqi , (129)
2zm i=1
where the sum represents the quasi-static rays extracted from Iie .
In the case of G̃9 , we have
(kρ /k0 )2 η0 Ive
G̃9 = ∼ (kρ /k0 )2 G̃q9 , (130)
zn zm
which leads to
Ne
jκen 1 X e
G̃q9 = aeqi e−jkzq bqi , (131)
2zn zm jkzq i=1
where the sum represents the quasi-static rays extracted from Ive .
In the case of G̃11 , we have

G̃11 = (Vvh − Vve ) ∼ G̃q11 , (132)

34
This is a desirable property, because it makes the subsequent exponential fit unnecessary. Here, we are referring
to a situation where all layers, including half-spaces, are made of the same isotropic material. This multilayer may
be unshielded or backed on one side by a PEC or PMC groundplane. In this case, the number of quasi-static terms is
reduced to one for the unshielded geometry, since only the direct ray is present, and to two for the grounded structure,
since the ray reflected from the groundplane must also be included. Note that kernels G3 , G4 , and G11 are identically
zero in this case, since Iie = Iih and Vve = Vvh .

23
which leads to !
Nh Ne
1 X h
X e
G̃q11 = ahqi e−jkzq bqi − aeqi e−jkzq bqi , (133)
2 i=1 i=1

where the first and second sums represent the quasi-static rays extracted from Vvh and Vve , respec-
tively.
In the case of G̃12 , we have

G̃12 = (Vvh + Vve ) ∼ G̃q12 , (134)

which leads to !
Nh Ne
1 X
−jkzq bh
X
−jkzq beqi
G̃q12 = ahqi e qi + aeqi e , (135)
2 i=1 i=1

where the first and second sums represent the quasi-static rays extracted from Vvh and Vve , respec-
tively.
In the case of G̃13 , we have
Vh
G̃13 = i ∼ G̃q13 , (136)
η0 µzn
which leads to
Nh
jκhn 1 X h
G̃q13 = ahqi e−jkzq bqi , (137)
2µzn jkzq i=1

where the sum represents the quasi-static rays extracted form Vih .
In the case of G̃14 , we have
η0 Ive
G̃14 = ∼ G̃q14 , (138)
zm
which leads to
Ne
jκen 1 X e
G̃q14 = aeqi e−jkzq bqi , (139)
2zm jkzq i=1
where the sum represents the quasi-static rays extracted from Ive .
In what follows, we assume that all kρ −independent factors have been absorbed into the coef-
ficients aαqi . For kernels comprising two TLGFs, the two quasi-static forms are merged whenever
possible, to minimize the total number of quasi-static terms. In the case of G̃3 , G̃4 , and G̃11 , the
direct rays associated with the two TLGFs cancel, if the source and the field point lie in the same
isotropic layer. However, if these points lie on the same interface between dissimilar media, the
corresponding space-domain kernels are still singular.
In summary, after the post-processing described above, the quasi-static parts will have one of
the three forms given below.

1. For G0 , G1 , G2 , G6 , G9 , G13 , and G14 ,

Nq
1 X
G̃ =q
aqi e−jkzq bqi . (140)
jkzq i=1

24
 2. For G5 ,
Nq1 Nq2
X
−jkzq bq1i 1 X
q
G̃ = jkzq aq1i e + aq2i e−jkzq bq2i . (141)
i=1
jkzq i=1

3. For G3 , G4 , G7 , G8 , G11 , and G12 ,

Nq
X
q
G̃ = aqi e−jkzq bqi . (142)
i=1

Note that the quasi-static representation of G5 comprises two quasi-static forms, with the first
asymptotically dominant. As already mentioned, the second form could be omitted, but we retain
it in order to make G̃q exact in the case of isotropic, homogeneous multilayers.
The space-domain counterparts of (140)–(142) can be obtained by using the integral identities
listed in Appendix A. The treatment of the dual kernels G̃00 through G̃09 is completely analogous to
that of the corresponding basic kernels and is not presented here in the interest of brevity.

6.2 Extraction of guided-wave terms

The spectral domain kernels in general possess branch-point and pole singularities, as indicated
in Fig. 4. Given the geometry of the multilayer and the media parameters, the location of all
poles can be automatically computed, as described in detail elsewhere [44]. We use a modified
Delves-Lyness method [23, 24] employing a polygonal contour integration technique [45] and a
kρ −plane mapping [46–49] to remove branch points. Note that there are in general two sets of
poles, associated with the TM and TE waves.
The pole residues of the spectral domain kernels in (60)–(85) can readily be computed from
the residues of the TLGFs of the multilayer.35 Hence, let TLGFαk denote one of the four TLGFs
listed in Table 1, where the subscript k (= 1, 2, 3, 4) denotes the TLGF number and the superscript
α
α (= e, h) the wave type. Also, let kρi be the ith pole of type α. Then, the corresponding TLGF
residue can be found as
Z 2π
a
α
Rki = TLGFαk (kρ ; z|z 0 ) ejϕ dϕ , (143)
2π 0 α
kρ =k +a e jϕ
ρi

where a is the radius of a circular path centered at the pole, selected not to exceed half the distance
α
from kρi to the nearest kρ −plane singularity. The integral is computed by an adaptive trapezoidal
quadrature.
The relationship between the kernel residues, Riα , and the TLGF residues, Rki α
, is given in Ta-
ble 4. As before, the subscripts n an m are used to distinguish media parameters of the source and
field-point layers, respectively. Also, it is assumed that in the computation of the TLGF residues
α
the normalizations Vi /η0 → Vi and η0 Iv → Iv are used, and that the poles kρi are normalized to
0 0
k0 . The residues of the kernels G0 through G9 , which are given by expressions dual to those listed
for the kernels G0 through G9 , are omitted for brevity.
Let us now suppose that a spectral domain kernel G̃ has Np poles kρi on the positive-real axis
or in the fourth quadrant of the kρ −plane, with residues Ri determined according to Table 4. There
35
Note that, unlike the poles, the residues depend on z and z 0 .

25
Table 4: Relationship between kernel residues and
TLGF residues.

Kernel Rie Rih

e e 2 h h 2
G0 −R1i /(kρi ) R1i /(kρi )
h
G1 N/A R1i
 e 2
 e h h 2
G2 a A − B/(kρi ) R2i BR2i /(kρi )
e h
G3 −µtm R3i µtm R3i
e h
G4 −µtn R4i µtn R4i
e h
G5 R1i R1i
e h
G6 −R1i R1i
e
G7 R4i /zn N/A
e
G8 R3i /zm N/A
e e 2
G9 R2i (kρi ) /(zn zm ) N/A
e h
G11 −R4i R4i
e h
G12 R4i R4i
h
G13 N/A R1i /µzn
e
G14 R2i /zm N/A

a
A = µtm /zn + µtn /zm , B = µtm µtn

26
is no need here to distinguish between the poles and residues of the TM and TE types, as they are
treated in the same fashion. The guided-wave part of this kernel will have the form
Np  Np 3
4Ri kρi
 X
p
X 2Ri kρi 2Ri kρi
G̃ = 2
− 2 2
= 4
, (144)
i=1
kρ2 − kρi kρ + kρi k 4 − kρi
i=1 ρ

where for each pole pair ±kρi , we also include a spurious, non-existent pole pair ∓jkρi , as illus-
trated in Fig. 7. The inclusion of the spurious pole pairs makes the guided-wave expression (144)

Figure 7: True and spurious pole-pairs.

more rapidly convergent for kρ → ∞ [50, 51].

The space-domain counterparts of (144) can be obtained by using the integral identities listed
in Appendix B. Note that in these identities the terms associated with the spurious poles cancel the
undesirable singularities as ρ → 0 [5, 52, 53].

6.3 Extraction of branch-point terms

In this section, we derive approximate expressions for the branch point (continuous spectrum)
contribution to the kernels of a multilayer. The branch point singularities of the spectral domain
kernels are associated with half-space media. Hence, let kb be the wavenumber of the upper or
lower half-space. We assume that there are no other branch points nearby, but allow for the pos-
sibility of a pole kp arbitrarily close to kb , as illustrated in Fig. 8. The location of the pole kp
and its residue, Rp , can be determined by the methods of Sec. 6.2. The branch point extraction is
only critical for lossless (or near-lossless) half-spaces, since otherwise the associated lateral waves
decay rapidly with horizontal distance along the interface.
We consider kernels of the form
Z ∞
1
G(ρ) = G̃(kρ )Jn (kρ ρ)kρn+1 dkρ (145)
2π 0
Z ∞
1
= G̃(kρ )Hn(2) (kρ ρ)kρn+1 dkρ , n = 0, 1, 2, (146)
4π ∞e−jπ

27
 Figure 8: Original and deformed integration paths in kρ −plane.

where G̃ is an even function of kρ . The integration path in (146) is along the real axis and is
illustrated as path C0 in Fig. 8. The inverse Hankel transform of (145) gives
Z ∞
−n
G̃(kρ ) = 2π kρ G(ρ)Jn (kρ ρ)ρ dρ . (147)
0

For simplicity, we do not indicate the dependence of the kernels on z and z 0 , which are assumed to
have fixed values. q
The branch point at kb arises because G̃ is not an even function of kzb = kb2 − kρ2 . The
hyperbolic branch cut emanating from kb , as illustrated in Fig. 8, delineates the boundary along
= kzb = 0 between the proper (= kzb < 0) and improper (= kzb > 0) sheets of the Riemann surface
associated with kzb [54, p. 610]. The real axis integration path C0 may be deformed to the path
Cb , wrapped around a vertical semi-infinite line emanating from the branch point kb , as illustrated
in Fig. 8. Note that the left segment of Cb lies on the improper sheet and the right segment lies
on the proper sheet. The modified integral must be augmented by the contributions from the
singularities captured in the path deformation process, including proper and improper poles and
possibly another branch point. Assuming large ρ, the integral along Cb may be approximated as
Z
b 1
G (ρ) = G̃(kρ )Hn(2) (kρ ρ)kρn+1 dkρ , (148)
4π Cb
s Z
1 2j n
≈ j G̃(kρ )e−jkρ ρ kρn+1/2 dkρ , (149)
4π πρ Cb

where we have used the asymptotic form of the Hankel function [26, eq. 9.2.4]. It can be shown
that there is no contribution to the integral from the semi-circular arc of vanishing radius passing
above kb [55, p. 157]. The remainder of Cb follows the steepest descent path away from the branch
point [21, p. 430] and the integration along this path is facilitated by introducing a new variable s
as p
kρ = kb − js2 , s = ejπ/4 kρ − kb . (150)
Clearly, the integrand falls off rapidly with increasing s and most of the contribution√to the integral
comes from the vicinity of the branch point kρ = kb , where s and kzb ≈ ±s(1 + j) kb are small.
Hence, we may further approximate (149) as
n+1/2 Z ∞h
1 j n kb −jkb ρ
i 2
b
G (ρ) ≈ √ e G̃ (s) − G̃ (s) e−s ρ s ds .
+ −
(151)
π 2πjρ 0

28
Here, G̃(s) stands for G̃(kρ ) with kρ given by the first expression in (150). The superscript ‘+’
(‘−’) indicates that the kernel should be evaluated on the proper (improper) Riemann sheet. The
integration path in (151) is the positive-real semi-axis in the s−plane, as illustrated in Fig. 9, where
we also show the s−plane image of the path C0 and the mapping of the poles from Fig. 8.

Figure 9: Integration paths in s−plane.

Since G̃− is G̃+ with the proper sheet value of kzb replaced by −kzb , the difference of G̃+ and
G̃− in (151) is an odd function of kzb , and thus an odd function of s. Hence, in the vicinity of the
branch point we may use the approximation (cf. [56, 57])
Bs
F̃ (s) ≡ G̃+ (s) − G̃− (s) ≈ , (152)
s2 − s2p
where B is a constant coefficient. Assuming that (152) holds, we obtain the relationship
Rp = ∓jsp B , (153)
where the upper (lower) sign applies if the pole is on the proper (improper) Riemann sheet. Note
that sp = 0 if the pole coalesces with the branch point, in which case Rp = 0 (but B remains
finite). Hence, the last case excepted, B could be obtained from the residue Rp , but the value may
be inaccurate if the pole is far away from the branch point. For two-layer media, it is not difficult
to derive analytical expressions for B by expanding the kernel about kzb = 0, but this procedure
can become cumbersome for the general case of multilayered media [50]. Here, we have opted
for a simple least-squares fitting procedure, similar to that previously used by Boix et al. [57].36
Hence, by enforcing (152) at a sequence of points si , i = 1, 2, . . . , M , along the real axis near the
branch point s = 0, we find B as
PM 2 2
i=1 F̃ (si )(si − sp )si
B= PM 2 , (154)
i=1 si

where the number of points M and the sampling interval are not critical. Usually,
√ good results are
obtained with M = 50 and si uniformly distributed in the range 0 < si < 0.001 k0 .
With B thus determined, we can use (152) in (151), to obtain
n+1/2 Z ∞
1 Bj n kb −jkb ρ 2 s2 ds
b
G (ρ) ≈ √ e e−s ρ 2 , (155)
2π (1 + j) πρ −∞ s − s2p
36
These authors use the least-squares fitting to determine both sp and B, and then compute Rp from (153).

29
where the integration has been extended to the entire real axis in Fig. 9. Simple, closed-form,
approximate expressions for the integral in (155) can be derived for two limiting cases, where |sp |2 ρ
is either small or large, and thus either s2p or s2 may be neglected, respectively, in the denominator
of the integrand. As a result, we obtain

n n+1/2
1 Bj kb  1 if |sp |2 ρ → 0 ,
b −jkb ρ
G (ρ) ≈ e 1 (156)
2π (1 + j)ρ − 2 if |sp |2 ρ → ∞ .
2sp ρ

The two approximations in (156) can be stitched together by means of suitable exponential transi-
tion functions dependent on the ‘numerical distance’ s2p ρ [57]. Also, the undesirable singularity at
ρ = 0 can be annihilated by introducing another exponential factor [50, 58]. After these modifica-
tions, we arrive at
 

 −|sp |2 ρ
n n+1/2
1 Bj kb 1−e −k0 ρ 1 − e
Gb (ρ) = e−jkb ρ e−|sp |2 ρ − , (157)
2π 1 + j ρ 2s2p ρ

which is applicable for all values of ρ. This expression is recognized as the ‘residual wave’ previ-
ously derived by Boix et al. [57] in the context of their enhanced total least-squares method. Note
that the transition functions make the first or the second term in the brackets dominate, depend-
ing on the magnitude of the numerical distance. Hence, Gb (ρ) is initially O(ρ−1 ), but it becomes
O(ρ−2 ) for sufficiently large ρ, with a phase delay kb ρ. Therefore, despite the crude approxima-
tions involved, the expression (157) exhibits the most important characteristics associated with a
branch point wave (also called a lateral wave) when the horizontal distance between the source and
the observation point is large [50], [54, p. 464].
Returning to (155), we may also express it as
!
n n+1/2
r Z ∞ −s2 ρ
1 Bj k ρ e
Gb (ρ) ≈ b
e−jkb ρ 1 + s2p ds , (158)
2π (1 + j)ρ π −∞ s2 − s2p

where we used the substitution

s2 s2p
= 1 + . (159)
s2 − s2p s2 − s2p
The integral in (158), which will be denoted by Ip , can be evaluated as (cf. [21, Sec. 4.4], [59,
Sec. 27.2], [60])
2j √ −s2p ρ √
Ip = ± πe Q (∓jsp ρ ) , =sp ≷ 0 , (160)
sp
where Z ∞
2
Q(z) = e−t dt (161)
z
is proportional to the error function complement, erfc [26, Eq. 7.1.2]. The upper or lower sign is
selected in (160) depending on whether the pole lies above or below the positive-real axis in Fig. 9,

30
and the two expressions differ by the residue of the integrand at the pole.37 Upon substituting (160)
into (158), we finally obtain
n+1/2
1 Bj n kb h √ 2 √ i
Gb (ρ) ≈ e−jkb ρ 1 ± 2jsp ρ e−sp ρ Q (∓jsp ρ ) , =sp ≷ 0 . (162)
2π (1 + j)ρ

Although this expression is applicable for all (sufficiently large) values of ρ, it does not have
a closed-form spectral domain counterpart. To remedy this, we may try using the asymptotic
expansion [21, p. 405]38
 
√ −s2p ρ √ 1
±2jsp ρ e Q (∓jsp ρ ) ∼ − 1 + 2 + · · · , |sp |2 ρ → ∞ , (163)
2sp ρ

which, however, leads back to (156).

Hence, we adopt for Gb the expression (157). Its spectral domain counterpart, G̃b , is obtained
by applying the inverse Hankel transform (147). The integration can be done in closed form by
invoking [61, p. 386]. The result depends on the order n of the transform, and is given as
n+1/2
Bj n kb
 
b 1 (2)
G̃ (kρ ) = In(1) − 2 In , (164)
1+j 2sp

with  q −n q −n 

kρ2
+ +c c2 2 2
kρ + d + d
In(1) =  − , n = 0, 1, 2 , (165)
 
q q
kρ2 + c2 2
kρ + d2

and   q  q 

 2 2
kρ + b + b 2 2
kρ + c + c 

ln  , n = 0 ,

 

   q  q

 k 2 + a2 + a k 2 + d2 + d

 ρ ρ


In(2) =  −n q −n (166)
 1 q 2 2+a 2 + b2 + b
k + a − k

 ρ ρ
n






 q −n q −n 
 −
 2 2
kρ + c + c + 2 2
kρ + d + d , n = 1, 2 ,


where the notation given in Table 5 is used.39 The asymptotic behavior of (164) is no worse than
O(kρ−3 ).
37
When the pole is on the real axis (=sp = 0), (160) must be properly modified by one-half of the residue [21,
p. 399], [60].
38
Note that the expressions corresponding to the upper and lower signs are asymptotically equivalent. This is
because the residue term of the integrand of Ip at the pole sp vanishes exponentially for large ρ.
39
The upper expression in (166) was obtained by letting n → 0 in the lower expression.

31
 Table 5: Notation used in the expression for G̃b .

a b c d

jkb a + k0 a + |sp |2 b + |sp |2

6.4 Exponential fit

Consider now a remainder kernel of the form
G̃c = G̃ − G̃q − G̃p , (167)
which applies to all kernels except G6 and G9 , for which the quasi-static term G̃q must be multi-
plied by (kρ /k0 )2 . In the above, G̃q is given by one of the three forms (140)–(141) and G̃p is given
by (144). The exponential fit of G̃c can be obtained by the matrix pencil method (MPM) [62, 63],
which is based on uniform sampling along a suitable linear path in the complex kzq −plane. The
robustness of the DCIM can be improved by the recursive application of the MPM on two or three
piecewise-linear path segments, which can have different lengths and sampling rates [40, 64]. We
use a two-level approach, with the sampling paths illustrated in Fig. 10. The path segments Cl ,
l = 1, 2, in the kzq −plane are parameterized in terms of a real parameter t as
kzq = γl−1 + (γl − γl−1 )t , 0 ≤ t < 1, (168)
with40 q
γ0 ≡ kq , γl = kq2 − κ2l , (169)
where kq = k0 nq is the wavenumber of the “equivalent medium” introduced in Sec. 6.1. In the last
expression, the square root branch with non-positive imaginary part is implied. The corresponding
path in the kρ −plane is given by the mapping (108). The values of κ1 and κ2 are selected first, to
ensure desirable path properties in the kρ −plane, and they determine the location of γ1 and γ2 via
(169). Here, we choose
κ1 /k0 = max <nαeff,n + 1 , (170)
to ensure that the kernel singularities (guided-wave poles and branch points) located near the real
axis are avoided. The value of κ2 /k0 is less critical and is typically set to a number two orders of
magnitude greater than κ1 /k0 .
In the two-level DCIM approach, we first apply the MPM on C2 , which results in
N2
X N2
X
jkzq G̃ ≈ c
α2i e β2i t
= a2i e−jkzq b2i . (171)
i=1 i=1

Next, we subtract the so obtained exponential fit from the original function and apply the MPM to
the remainder function on C1 as
N2
X N1
X N1
X
−jkzq b2i
c
jkzq G̃ − a2i e ≈ β1i t
α1i e = a1i e−jkzq b1i . (172)
i=1 i=1 i=1
40
The symbol κ used here is unrelated to the same symbol introduced before in (105).

32
 (a)

(b)

Figure 10: DCIM sampling paths in (a) kρ −plane and (b) kzq −plane.

On each level l, the MPM automatically determines the number of exponential terms, Nl , based
on the specified precision of the samples (i.e., an estimate of the number of significant digits in
the data), and returns the coefficients αli and βli . From the latter, using the mapping (168), the
coefficients ali and bli are found as
jβli
bli = , ali = αli ejγl−1 bli . (173)
γl − γl−1
As a result of these steps, we finally obtain the approximation
Nl
L X
c
X e−jkzq bli
G̃ ≈ ali , (174)
l=1 i=1
jkzq

with the number of levels L = 2. In the above, the exponential fit was applied to jkzq G̃c , but it is

33
also possible to apply it directly to G̃c , which leads to
Nl
L X
X
c
G̃ ≈ ali e−jkzq bli . (175)
l=1 i=1

We use the former procedure for kernels G0 , G1 , G2 , G5 , G6 , G9 , G13 , and G14 , and the latter for
kernels G3 , G4 , G7 , G8 , G11 , and G12 .
We refer to (174) and (175) as complex image approximations. The closed-form space-domain
counterparts of these expressions can be obtained by using the appropriate integral identities from
among those listed in Appendix A. If desired, the two-level approach described above can easily
be reduced to a one-level approach (L = 1), where the MPM is only applied on the path segment
C1 , or extended to a three-level method (L = 3), where a third path segment is appended to the
sampling path of Fig. 10.41
In the case of G0 , G1 , and G2 , the above procedure results in42
( Nq L Nl
1 X e−jkq rqi X X e−jkq rli
G(ρ) = aqi + ali
2π i=1 rqi l=1 i=1
rli
Np  ) (176)
X jπ (2)
+ (−2Ri kρi ) H0 (kρi ρ) + K0 (kρi ρ) ,
i=1
2

with q
rδi = ρ2 + b2δi , δ = q, l , (177)
where the square root branch with positive real part is selected. In the above, we applied identities
(186) and (202) to the quasi-static and complex image terms, respectively.
In the case of G3 and G4 , we arrive at
( Nq  
1 X aqi −jkq bqi bqi −jkq rqi
G(ρ) = e − e
2π i=1 ρ rqi
L X Nl  
X ali −jkq bli bli −jkq rli
+ e − e (178)
l=1 i=1
ρ rli
Np  )
X jπ (2) 2
+ (−2Ri ) H1 (kρi ρ) − K1 (kρi ρ) + .
i=1
2 kρi ρ

Here, the quasi-static and complex image terms were inverted by (191), and the guided-wave terms
by (203).
41
Extensive numerical experiments (not reported here) have shown that the three-level approach can exhibit erratic
behavior, and is thus unreliable.
42
Henceforth, we write G(ρ) rather than G(ρ; z|z 0 ), to emphasize that the kernels are closed-form, as far as the
ρ−dependence is concerned. The dependence on z and z 0 is implicit in the quasi-static and complex-image coeffi-
cients, and in the pole residues.

34
 In the case of G5 , we obtain
( Nq1  2  −jkq rq1i
3bq1i

1 X 2 2 e
G(ρ) = aq1i 2
− 1 (1 + jkq rq1i ) − kq bq1i 3
2π i=1 rq1i rq1i
Nq2 L Nl
X e−jkq rq2i X X e−jkq rli
+ aq2i + ali (179)
i=1
rq2i l=1 i=1
rli
Np  )
X jπ (2)
+ (−2Ri kρi ) H0 (kρi ρ) + K0 (kρi ρ) ,
i=1
2

where rqki , k = 1, 2, is given by (177), with k added to the subscript list. In the above, the
quasi-static terms were inverted by (188) and (186), the complex-image terms by (186), and the
guided-wave terms by (202).
In the case of G6 , we obtain
( Nq  2  −jkq rqi
1 X 3ρ 2 2 e
G(ρ) = aqi 2 (1 + jkq rqi ) − kq ρ 3
2π i=1 rqi rqi
L X Nl

e−jkq rli
 
X 2 −jkq bli −jkq rli
+ ali e −e − (180)
l=1 i=1
jkq ρ2 rli
Np  )
X jπ (2) 4
+ (−2Ri kρi ) H2 (kρi ρ) − K2 (kρi ρ) + 2
.
i=1
2 (k ρi ρ)

Here, the quasi-static terms were inverted by (196), the complex image terms by (194), and the
guided-wave terms by (205).
In the case of G7 and G8 , we obtain
( Nq   −jkq rqi
1 X 3bqi ρ 2 e
G(ρ) = aqi 2
(1 + jkq rqi ) − kq bqi ρ 3
2π i=1 rqi rqi
L X Nl   −jkq rli
X 3bli ρ 2 e
+ ali 2
(1 + jkq rli ) − kq bli ρ 3
(181)
l=1 i=1
rli r li
Np  )
X
2 jπ (2)
+ (−2Ri kρi ) H (kρi ρ) + K1 (kρi ρ) .
i=1
2 1

Here, the quasi-static and complex image terms were inverted by (193), and the guided-wave terms
by (204).

35
 In the case of G9 , we obtain
( Nq  2  −jkq rqi
3bqi

1 X 2 2 e
G(ρ) = aqi 2
− 1 (1 + jkq rqi ) + kq ρ 3
2π i=1 rqi rqi
Nl
L X
X e−jkq rli
+ ali (182)
l=1 i=1
rli
Np  )
X jπ (2)
+ (−2Ri kρi ) H (kρi ρ) + K0 (kρi ρ) .
i=1
2 0

Here, the quasi-static terms were inverted by (189), the complex image terms by (186), and the
guided-wave terms by (202).
In the case of G11 , we obtain
( Nq
e−jkq rqi
   
1 X 2 −jkq bqi bqi −jkq rqi
G(ρ) = aqi 2 e − e − bqi (1 + jkq rqi ) 3
2π i=1 ρ rqi rqi
L X Nl
e−jkq rli
   
X 2 −jkq bli bli −jkq rli
+ ali 2 e − e − bli (1 + jkq rli ) (183)
l=1 i=1
ρ rli rli3
Np  )
X jπ (2) 4
+ (−2Ri kρi ) H2 (kρi ρ) − K2 (kρi ρ) + 2
.
i=1
2 (k ρi ρ)

Here, the quasi-static and complex-image terms were inverted by (195), and the guided-wave terms
by (205).
In the case of G12 , we obtain
( Nq
1 X e−jkq rqi
G(ρ) = aqi bqi (1 + jkq rqi ) 3
2π i=1 rqi
Nl
L X
X e−jkq rli
+ ali bli (1 + jkq rli ) (184)
l=1 i=1
rli3
Np  )
X jπ (2)
+ (−2Ri kρi ) H (kρi ρ) + K0 (kρi ρ) .
i=1
2 0

Here, the quasi-static and complex-image terms were inverted by (187), and the guided-wave terms
by (202).

36
 Finally, in the case of G13 and G14 , we obtain
( Nq
1 X e−jkq rqi
G(ρ) = aqi ρ(1 + jkq rqi ) 3
2π i=1 rqi
Nl
L X
X e−jkq rli
+ ali ρ(1 + jkq rli ) (185)
l=1 i=1
rli3
Np  )
X
2 jπ (2)
+ (−2Ri kρi ) H1 (kρi ρ) + K1 (kρi ρ) .
i=1
2

Here, the quasi-static and complex image terms were inverted by (192), and the guided-wave terms
by (204).
The extraction of the guided-wave terms is not necessary in the near zone, where it may even
adversely affect the accuracy of the complex image approximation [5, 53]. Hence, it may be ad-
vantageous to split the ρ−range of interest into two subranges, with a demarcation point at ρ = ρs ,
and extract the guided-wave terms only when ρ > ρs . The location of the switch point, which is
not critical, may be selected by the condition k0 ρs = π. The drawback of this approach is that two
sets of complex images must be extracted whenever the ρ− range of interest straddles ρs .
In conclusion of this section, we note that DCIM depends on several user-specified parameters
that affect the accuracy of the method. In addition to choosing the number of levels—which is
usually two, the user must select the end points of the sampling path segments, and for each
segment the number of samples and an estimate of the number of significant digits in the data.
Hence, the method requires a certain amount of trial-and-error “tuning” before it can be applied
with some confidence.

7 Sample results
In this section we present sample results for the five-layer grounded stack problem illustrated in
Fig. 11. This test structure is similar to that previously used in [5–7, 65], except that we have made
some of the layers magnetic, in order to make the test more demanding. The frequency is 30 GHz.
The source and field points are at the fixed level z 0 = z = 0.4 mm, while the horizontal distance
between them is varied over five decades in the range 10−3 < k0 ρ < 102 . To avoid ambiguity,
we assume that z = z 0 + 0. For this configuration, in Figs. 12 through 25 we compare the DCIM
results with the reference data, obtained by the direct integration method of Sec. 5, for the basic
kernels G0 through G14 . The DCIM results are plotted by a solid line and the reference data by
a dashed line. In these figures, we also plot the relative error using a dotted line.43 We used the
two-level DCIM with quasi-static term extraction (using 3 dominant terms) and guided-wave term
extraction (at 30 GHz, this structure can support three guided waves, two TM and one TE). The
DCIM sampling path was specified as follows: κ1 /k0 was set according to (170) and κ2 /k0 was set
to 300. The number of samples was 150, and the number of significant digits was 10—the same
for both sampling path segments. These DCIM parameters were used for all kernels.
43
If the relative error is below 1%, we fix the error scale maximum at 1%.

37
 Figure 11: Five-layer stack backed by a PEC groundplane.

The accuracy of these results can be improved by “tuning” the DCIM parameters for each of the
kernels individually. Also, the ρ−switch approach could be implemented to improve the accuracy
in the small−ρ range. The slight increase in error observed near the right end of the ρ range can
be attributed to the lateral wave (branch-point contribution) not being accurately sampled in the
DCIM.
The results for the dual kernels G00 through G09 are similar and are omitted to conserve space.

8 Conclusion
Insufficient robustness continues to be a problem with the DCIM, even when the quasi-static and
guide-wave terms are extracted before the exponential fit is performed. The method can be tuned
to yield excellent accuracy for a particular case—or even a particular class of problems, but it can
fail, sometimes catastrophically, if some critical parameter of the problem, such as frequency, is
changed by an order of magnitude. Increasing the number of sampling levels beyond two does
not appear to be a promising remedy. As can be expected, the mildly singular potential kernels
are easier to model by the DCIM than the more singular field kernels, especially the hypersingular
ones.
The accuracy of the space domain kernels generated by the DCIM is not known a priori and
can only be determined by comparisons with the results of other, trusted methods—such as the
direct integration method. This appears to be an insurmountable flaw of the DCIM, which may
disqualify it as a stand-alone method for critical applications.
The efficiency of the DCIM critically depends on the required number of sampling points,
which in turn depends on the spectral behavior of the remainder function (after the quasi-static and
guided-wave terms have been extracted) that must be fitted by the complex image representation.
The DCIM can be much faster than the direct integration method in cases where z and z 0 are fixed
and ρ is varied, since the coefficients from a single DCIM run (or at most two DCIM runs, if the
ρ−switch is employed) can be reused for different ρ values.

38
Figure 12: Plot of G0 vs. k0 ρ for the problem of Fig. 11 with z 0 = z = 0.4 mm.

Figure 13: Plot of G1 vs. k0 ρ for the problem of Fig. 11 with z 0 = z = 0.4 mm.

39
Figure 14: Plot of G2 vs. k0 ρ for the problem of Fig. 11 with z 0 = z = 0.4 mm.

Figure 15: Plot of G3 vs. k0 ρ for the problem of Figu. 11 with z 0 = z = 0.4 mm.

40
Figure 16: Plot of G4 vs. k0 ρ for the problem of Fig. 11 with z 0 = z = 0.4 mm.

Figure 17: Plot of G5 vs. k0 ρ for the problem of Fig. 11 with z 0 = z = 0.4 mm.

41
Figure 18: Plot of G6 vs. k0 ρ for the problem of Fig. 11 with z 0 = z = 0.4 mm.

Figure 19: Plot of G7 vs. k0 ρ for the problem of Fig. 11 with z 0 = z = 0.4 mm.

42
Figure 20: Plot of G8 vs. k0 ρ for the problem of Fig. 11 with z 0 = z = 0.4 mm.

Figure 21: Plot of G9 vs. k0 ρ for the problem of Fig. 11 with z 0 = z = 0.4 mm.

43
Figure 22: Plot of G11 vs. k0 ρ for the problem of Fig. 11 with z 0 = z = 0.4 mm.

Figure 23: Plot of G12 vs. k0 ρ for the problem of Fig. 11 with z 0 = z = 0.4 mm.

44
Figure 24: Plot of G13 vs. k0 ρ for the problem of Fig. 11 with z 0 = z = 0.4 mm.

Figure 25: Plot of G14 vs. k0 ρ for the problem of Fig. 11 with z 0 = z = 0.4 mm.

45
 Regarding future work, it appears desirable to extract and analytically handle the branch-point
contribution to the spectrum, which represents the lateral wave that may propagate unattenuated
and dominate the far zone field near the surface in cases where the waves guided by the layers
decrease rapidly (as a result of the material losses) or are not supported by the structure. This
should result in improved robustness, at the expense of making the method more cumbersome.

Appendix A:
Identities for the inversion of quasi-static and complex image
terms
The quasi-static and complex image terms can be transformed by the identities44
Z ∞ −jkz b
e e−jkr
J0 (kρ ρ)kρ dkρ = , (186)
0 jkz r
Z ∞
−jkz b e−jkr
e J0 (kρ ρ)kρ dkρ = b(1 + jkr) 3 , (187)
0 r
Z ∞  2   −jkr
−jkz b 3b 2 2 e
jkz e J0 (kρ ρ)kρ dkρ = − 1 (1 + jkr) − k b , (188)
0 r2 r3
Z ∞ −jkz b  2   −jkr
e 3 3b 2 2 e
J0 (kρ ρ)kρ dkρ = − 1 (1 + jkr) + k ρ , (189)
0 jkz r2 r3
Z ∞ −jkz b
e 1
e−jkb − e−jkr ,


J1 (kρ ρ) dkρ = (190)
0 jkz jkρ
Z ∞  
−jkz b 1 −jkb b −jkr
e J1 (kρ ρ) dkρ = e − e , (191)
0 ρ r
Z ∞ −jkz b
e 2 e−jkr
J1 (kρ ρ)kρ dkρ = ρ(1 + jkr) 3 , (192)
0 jkz r
Z ∞   −jkr
−jkz b 2 3bρ 2 e
e J1 (kρ ρ)kρ dkρ = 2
(1 + jkr) − k bρ , (193)
0 r r3
Z ∞ −jkz b
e 2 −jkb −jkr

e−jkr
J2 (kρ ρ)kρ dkρ = e − e − , (194)
0 jkz jkρ2 r
Z ∞
e−jkr
 
−jkz b 2 −jkb b −jkr
e J2 (kρ ρ)kρ dkρ = 2 e − e − b(1 + jkr) 3 , (195)
0 ρ r r
Z ∞ −jkz b  2  −jkr
e 3ρ e
J2 (kρ ρ)kρ3 dkρ = 2
(1 + jkr) − k 2 ρ2 , (196)
0 jkz r r3
where q p
kz = k 2 − kρ2 , r= ρ2 + b2 . (197)
44
The identity (190) is not directly used, but is included for completeness.

46
The first formula above is the well-known Sommerfeld identity, if b is real and positive [27, p. 66].
With some restrictions, this identity is also valid for complex-valued b [66]. The other identities
can be derived (cf. [7, 67]) from (186) by integration or differentiation and by using the formulas
[26, 68] Z
xJ0 (x) dx = xJ1 (x) , (198)

J00 (x) = −J1 (x) , (199)

2
J2 (x) = J1 (x) − J0 (x) , (200)
x
1
J2 (x) = J1 (x) − J10 (x) . (201)
x
It is important to note that the right-hand expressions in (191) and (195) have zero limits as
ρ → 0 with finite b, but are singular in this limit if b = 0.

Appendix B:
Identities for the inversion of pole terms
The pole terms can be transformed by the identities
Z ∞ 2
2kρi kρ dkρ
 
jπ (2)
J0 (kρ ρ) 4 = − H (kρi ρ) + K0 (kρi ρ) , (202)
0 kρ − kρi4
2 0
Z ∞ 3
2kρi dkρ
 
jπ (2) 2
J1 (kρ ρ) 4 = − H (kρi ρ) − K1 (kρi ρ) + , (203)
0 kρ − kρi 4
2 1 kρi ρ
Z ∞
2kρi kρ2 dkρ
 
jπ (2)
J1 (kρ ρ) 4 = − H (kρi ρ) + K1 (kρi ρ) , (204)
0 kρ − kρi4
2 1
Z ∞ 2
2kρi kρ dkρ
 
jπ (2) 4
J2 (kρ ρ) 4 = − H (kρi ρ) − K2 (kρi ρ) + , (205)
0 kρ − kρi 4
2 2 (kρi ρ)2
where Kn is the modified Bessel function of order n, which is related to the Hankel function as
[26]
2j n+1
Hn(2) (−jx) = Kn (x) . (206)
π
These identities can be derived by using (88)–(89), contour deformation, and the residue theorem
[69, p. 294].
It is important to note that the right-hand expressions in (202)–(205) are bounded for ρ → 0,
as can be seen by utilizing the small-argument forms [26]
 n
(2) 2j (2) j 2
H0 (x) −→ − ln x , Hn (x) −→ , n = 1, 2 , (207)
x→0 π x→0 π x
 n
1 2
K0 (x) −→ − ln x , Kn (x) −→ , n = 1, 2 . (208)
x→0 x→0 2 x

47
References
[1] K. A. Michalski, “Extrapolation methods for Sommerfeld integral tails (Invited review pa-
per),” IEEE Trans. Antennas Propagat., vol. 46, pp. 1405–1418, Oct. 1998.

[2] M. Tsai, C. Chen, and N. G. Alexopoulos, “Sommerfeld integrals in modeling interconnects

and microstrip elements in multi-layered media,” Electromagn., vol. 18, pp. 267–288, 1998.

[3] D. G. Fang, J. J. Yang, and G. Y. Delisle, “Discrete image theory for horizontal electric
dipoles in a multilayered medium,” IEE Proc., Pt. H, vol. 135, pp. 297–303, Oct. 1988.

[4] Y. L. Chow, J. J. Yang, D. G. Fang, and G. E. Howard, “A closed-form spatial Green’s function
for the thick microstrip substrate,” IEEE Trans. Microwave Theory Tech., vol. 39, pp. 588–
592, Mar. 1991.

[5] F. Ling and J. Jin, “Discrete complex image method for Green’s functions of general multi-
layer media,” IEEE Microwave Guided Wave Lett., vol. 10, pp. 400–402, Oct. 2000.

[6] P. Ylä-Oijala, M. Taskinen, and J. Sarvas, “Multilayered media Gren’s functions for MPIE
with general electric and magnetic sources by the Hertz potential approach,” in Progress in
Electromagnetics Research (J. A. Kong, ed.), vol. PIER 33, pp. 141–165, Cambridge, MA:
EMW Publ., 2001.

[7] P. Ylä-Oijala and M. Taskinen, “Efficient formulation of closed-form Green’s functions for
general electric and magnetic sources in multilayered media,” IEEE Trans. Antennas Propa-
gat., vol. 51, pp. 2106–2115, Aug. 2003.

[8] X. He, S. Gong, and Q. Liu, “Fast computation of spatial Green’s functions of multilayered
microstrip antennas,” Microwave & Opt. Technol. Lett., vol. 45, no. 1, pp. 85–88, 2005.

[9] M. Yuan, T. K. Sarkar, and M. Salazar-Palma, “A direct discrete complex image method from
the closed-form Green’s functions in multilayered media,” IEEE Trans. Microwave Theory
Tech., vol. 54, pp. 1025–1032, Mar. 2006.

[10] F. Ling, V. Okhmatovski, B. Song, and A. Dengi, “Systematic extraction of static images
from layered media Green’s function for accurate DCIM implementation,” IEEE Antennas
Wireless Propagat. Lett., vol. 6, pp. 215–218, 2007.

[11] M. Zhang, L. Li, and Y. Tian, “An efficient approach for extracting poles of Green’s functions
in general multilayered media,” IEEE Trans. Antennas Propagat., vol. 56, pp. 269–273, Jan.
2008.

[12] K. A. Michalski, “The mixed-potential electric field integral equation for objects in layered
media,” Arch. Elek. Übertragung., vol. 39, pp. 317–322, Sept.–Oct. 1985.

[13] J. R. Mosig, “Arbitrarily shaped microstrip structures and their analysis with a mixed potential
integral equation,” IEEE Trans. Microwave Theory Tech., vol. 36, pp. 314–323, Feb. 1988.

48
[14] K. A. Michalski and D. Zheng, “Electromagnetic scattering and radiation by surfaces of
arbitrary shape in layered media, Part I: Theory,” IEEE Trans. Antennas Propagat., vol. 38,
pp. 335–344, Mar. 1990.

[15] K. A. Michalski and J. R. Mosig, “Multilayered media Green’s functions in integral equation
formulations (Invited review paper),” IEEE Trans. Antennas Propagat., vol. 45, pp. 508–519,
Mar. 1997.

[16] N. N. Feng, D. G. Fang, and W. P. Huang, “Full wave discrete images of an electric dipole
for uniaxial substrate,” in Proc. Int. Conf. on Microwave and Millimeter Wave Technology,
(Beijing, China), pp. 998–1001, July 1998.

[17] K. A. Michalski, “Formulation of mixed-potential integral equations for arbitrarily shaped

microstrip structures with uniaxial substrates,” J. Electromagn. Waves Appl., vol. 7, no. 7,
pp. 899–917, 1993.

[18] K. A. Michalski, “Electromagnetic field computation in planar multilayers,” in Encyclo-

pedia of RF and Microwave Engineering (K. Chang, ed.), vol. 2, pp. 1163–1190, Wiley-
Interscience, 2005.

[19] D. R. Wilton, “Review of current status and trends in the use of integral equations in compu-
tational electromagnetics,” Electromagn., vol. 12, pp. 287–341, July-Dec. 1992.

[20] D. R. Wilton, “Computational methods,” in Scattering. Scattering and Inverse Scattering in

Pure and Applied Science (R. Pike and P. Sabatier, eds.), ch. 1.5.5, pp. 316–365, London:
Academic Press, 2002.

[21] L. B. Felsen and N. Marcuvitz, Radiation and Scattering of Waves. Englewood Cliffs, N.J.:
Prentice Hall, 1973.

[22] C.-I. G. Hsu, R. F. Harrington, K. A. Michalski, and D. Zheng, “Analysis of a multiconductor

transmission lines of arbitrary cross-section in multilayered uniaxial media,” IEEE Trans.
Microwave Theory Tech., vol. 41, pp. 70–78, Jan. 1993.

[23] L. M. Delves and J. N. Lyness, “A numerical method for locating the zeros of an analytic
function,” Math. Comput., vol. 21, pp. 543–560, Oct. 1967.

[24] R. Rodriguez-Berral, F. Mesa, and F. Medina, “Systematic and efficient root finder for com-
puting the modal spectrum of planar layered waveguides,” Int. J. RF Microwave Computer-
Aided Eng., vol. 14, pp. 73–83, Jan. 2004.

[25] M. Metcalf and J. Reid, FORTRAN 90/95 explained. Oxford, UK: Oxford University Press,
1996.

[26] M. Abramowitz and I. A. Stegun, eds., Handbook of Mathematical Functions. New York:
Dover, 1965.

[27] W. C. Chew, Waves and Fields in Inhomogeneous Media. New York: IEEE Press, 1995.

49
[28] M. Paulus, P. Gay-Balmaz, and O. J. F. Martin, “Accurate and efficient computation of the
Green’s tensor for stratified media,” Phys. Rev. E, vol. 62, pp. 5797–5807, Oct. 2000.
[29] T. Yu and W. Cai, “FIFA – Fast Interpolation and Filtering Algorithm for calculating dyadic
Green’s function in the electromagnetic scattering of multi-layered structures,” Commun.
Computat. Phys., vol. 1, pp. 229–260, Apr. 2006.
[30] M. Yuan and T. K. Sarkar, “Computation of the Sommerfeld integral tails using the Matrix
Pencil Method,” IEEE Trans. Antennas Propagat., vol. 54, pp. 1358–1362, Apr. 2006.
[31] G. Evans, Practical Numerical Integration. New York: Wiley, 1993.
[32] J. R. Mosig, “Integral equation technique,” in Numerical Techniques for Microwave and
Millimeter-Wave Passive Structures (T. Itoh, ed.), pp. 133–213, New York: Wiley, 1989.
[33] J. R. Mosig and A. A. Melecón, “Green’s functions in lossy layered media: integration
along the imaginary axis and asymptotic behavior,” IEEE Trans. Antennas Propagat., vol. 51,
pp. 3200–3208, Dec. 2003.
[34] N. V. Shuley, R. R. Boix, F. Medina, and M. Horno, “On the fast approximation of Green’s
functions in MPIE formulations for planar layered media,” IEEE Trans. Microwave Theory
Tech., vol. 50, pp. 2185–2192, Sept. 2002.
[35] P. Cornille, “Numerical saddle point method,” J. Math. Anal. Appl., vol. 38, pp. 633–639,
1972.
[36] K. A. Michalski, “On the efficient evaluation of integrals arising in the Sommerfeld halfspace
problem,” in Moment Methods in Antennas and Scatterers (R. C. Hansen, ed.), pp. 325–331,
Boston: Artech House, 1990.
[37] T. J. Cui and W. C. Chew, “Fast evaluation of Sommerfeld integrals for EM scattering and
radiation by three-dimensional buried objects,” IEEE Trans. Geosci. Remote Sens., vol. 37,
pp. 887–900, Mar. 1999.
[38] J. J. Yang, Y. L. Chow, and D. G. Fang, “Discrete complex images of a three-dimensional
dipole above and within a lossy ground,” IEE Proc., Pt. H, vol. 138, pp. 319–326, Aug. 1991.
[39] K. A. Michalski and J. R. Mosig, “Discrete complex image mixed-potential integral equa-
tion analysis of microstrip patch antennas with vertical probe feeds (Invited paper),” Electro-
magn., vol. 15, pp. 377–392, July-Aug. 1995.
[40] M. I. Aksun, “A robust approach for the derivation of closed-form Green’s functions,” IEEE
Trans. Microwave Theory Tech., vol. 44, pp. 651–658, May 1996.
[41] M. I. Aksun and G. Dural, “Clarification of issues on the closed-form Green’s functions in
stratified media,” IEEE Trans. Antennas Propagat., vol. 53, pp. 3644–3653, Nov. 2005.
[42] L. Zhuang, G. Zhu, Y. Zhang, and B. Xiao, “An improved discrete complex image method
for Green’s functions in layered media,” Microwave & Opt. Technol. Lett., vol. 49, no. 6,
pp. 1337–1340, 2007.

50
[43] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein, Introduction to Algorithms. Cam-
bridge, Mass.: MIT Press, 2nd ed., 2001.

[44] K. A. Michalski, “Automatic, robust and efficient pole location for planar, uniaxial multilay-
ers.” Technical Report for Alpha Omega Electromagnetics, June 2006.

[45] E. Anemogiannis and E. N. Glytsis, “Multilayer waveguides: Efficient numerical analysis of

general structures,” J. Lightwave Technol., vol. 10, pp. 1344–1351, Oct. 1992.

[46] R. E. Smith, G. W. Forbes, and S. N. Houde-Walter, “Unfolding the multivalued planar

waveguide dispersion relation,” IEEE J. Quantum Electron., vol. 29, pp. 1031–1034, Apr.
1993.

[47] R. E. Smith and S. N. Houde-Walter, “The migration of bound and leaky solutions to the
waveguide dispersion relation,” J. Lightwave Technol., vol. 11, pp. 1760–1768, Nov. 1993.

[48] A. Bakhtazad, H. Abiri, and R. Ghayour, “A general transform for regularizing planar open
waveguide dispersion relation,” J. Lightwave Technol., vol. 15, pp. 383–390, Feb. 1997.

[49] R. Rodriguez-Berral, F. Mesa, and F. Medina, “Appropriate formulation of the characteristic

equation for open nonreciprocal layered waveguides with different upper and lower half-
spaces,” IEEE Trans. Microwave Theory Tech., vol. 53, pp. 1613–1623, May 2005.

[50] F. J. Demuynck, G. A. E. Vandenbosch, and A. R. van de Capelle, “The expansion wave

concept–Part I: Efficient calculation of spatial Green’s functions in a stratified dielectric
medium,” IEEE Trans. Antennas Propagat., vol. 46, pp. 397–406, Mar. 1998.

[51] A. K. Abdelmageed and A. A. K. Mohsen, “An accurate computation of Green’s functions

for multilayered media in the near-field region,” Microwave & Opt. Technol. Lett., vol. 29,
no. 2, pp. 130–131, 2001.

[52] N. Hojjat, S. Safavi-Naeini, and Y. L. Chow, “Numerical computation of complex image

Green’s functions for multilayered dielectric media: near-field zone and the interface region,”
IEE Proc.-Microw. Antennas Propagat., vol. 145, pp. 449–454, Dec. 1998.

[53] S. Teo, S. Chew, and M. Leong, “Error analysis of the discrete complex image method and
pole extraction,” IEEE Trans. Microwave Theory Tech., vol. 51, pp. 406–413, Feb. 2003.

[54] A. Ishimaru, Electromagnetic Wave Propagation, Radiation, and Scattering. Englewood

Cliffs, NJ: Prentice Hall, 1991.

[55] G. Tyras, Radiation and Propagation of Electromagnetic Waves. New York: Academic Press,
1969.

[56] P. Baccarelli, P. Burghignoni, F. Frezza, A. Galli, G. Lovat, and D. R. Jackson, “Uniform

analytical representation of the continuous spectrum excited by dipole sources in a multi-
layer dielectric structure through weighted cylindrical leaky waves,” IEEE Trans. Antennas
Propagat., vol. 52, pp. 653–664, Mar. 2004.

51
[57] R. R. Boix, F. Mesa, and F. Medina, “Derivation of closed-form Green’s functions for multi-
layered media via the method of total least squares.” to be published, 2007.

[58] R. R. Boix, F. Mesa, and F. Medina, “Application of total least squares to the derivation of
closed-form Green’s functions for planar layered media,” IEEE Trans. Microwave Theory
Tech., vol. 55, pp. 268–280, Feb. 2007.

[59] L. M. Brekhovskikh, Waves in Layered Media. New York: Academic Press, 2nd ed., 1980.

[60] G. D. Bernard and A. Ishimaru, “On complex waves,” Proc. IEE, vol. 114, pp. 43–49, Jan.
1967.

[61] G. N. Watson, A Treatise on the Theory of Bessel Functions. New York: Cambridge Univer-
sity Press, 1995.

[62] T. K. Sarkar and O. Pereira, “Using the matrix pencil method to estimate the parameters of
a sum of complex exponentials,” IEEE Antennas Propagat. Magaz., vol. 37, pp. 48–55, Feb.
1995.

[63] R. S. Adve, T. K. Sarkar, O. M. C. Pereira-Filho, and S. M. Rao, “Extrapolation of time-

domain responses from three-dimensional conducting objects utilizing the matrix pencil tech-
nique,” IEEE Trans. Antennas Propagat., vol. 45, pp. 147–156, Jan. 1997.

[64] N. Kınayman and M. I. Aksun, “Efficient use of closed-form Green’s functions for the anal-
ysis of planar geometries with vertical connections,” IEEE Trans. Microwave Theory Tech.,
vol. 45, pp. 593–603, Apr. 1997.

[65] F. Ling, J. Liu, and J. Jin, “Efficient electromagnetic modeling of three-dimensional multi-
layer microstrip antennas and circuits,” IEEE Trans. Antennas Propagat., vol. 50, pp. 1628–
1635, June 2002.

[66] Y. L. Li and M. J. White, “Near-field computation for sound propagation above ground—
Using complex image theory,” J. Acoust. Soc. Am., vol. 99, pp. 755–760, Feb. 1996.

[67] P. C. Clemmow, “The resolution of a dipole field into transverse electric and transverse mag-
netic waves,” Proc. IEE, vol. 110, pp. 107–111, Jan. 1963.

[68] I. S. Gradshteyn and I. M. Ryzhik, eds., Table of Integrals, Series, and Products. New York:
Academic Press, 7th ed., 2007.

[69] J. A. Kong, Electromagnetic Wave Theory. New York: Wiley, 2nd ed., 1990.

52