Wavelets trong Electromagnetics và mô hình thiết bị P8 ppsx

CHAPTER NINE
Wavelets in Packaging,
Interconnects, and EMC
In this chapter we will study multiconductor, multilayered transmission lines
(MMTL) employing quasi-static, quasi-dynamic, and full-wave analyses. We extract
from MMTL the distributed (parasitic) parameters in matrix form of the capacitance
[C], inductance [L], resistance [R] and conductance [G], or the [Z]-parameters,
[Y ]-parameters, or more generally the scattering matrix [S]. MMTL systems are
commonly found in high-speed, high-density digital electronics at the levels of in-
dividual chip carriers, printed circuit boards (PCBs), and more recently, multichip
modules (MCMs). Previous methods for extraction of the distributed circuit param-
eters include the quasi-TEM solutions [1–5], and more rigorous techniques [6–9].
They also included full-wave analysis algorithms [10–15].
We begin with the quasi-static formulation (QSF) [1], which provides the para-
sitic capacitance [C], inductance [L], resistance [R], and conductance [G]. Due to
the limitation of its assumptions, the QSF results for L, C, R,andG are independent
of frequency values. This characteristic is accurate only under special circumstances.
The comparison of the QSF solution with the full-wave finite element method (FEM)
data indicates that the capacitance [C] values from the QSF are accurate to at least 50
GHz [16], while the [L] and [R] may have large errors. For most practical applica-
tions, conductance [G]is negligibly small. Therefore, in the quasi-static formulations
of Sections 9.1 and 9.2, we will focus mainly on capacitance extraction.
In Section 9.3 we will introduce an intermediate formulation between that of the
quasi-static and full-wave, referred to as the quasi-dynamic formulation (QDF). The
QDF provides us with frequency-dependent parameters of the skin effect resistance
and total (internal plus external) inductance. The comparison of the QDF with the
FEM [17] and laboratory tests [18] reveals that the [L] and [R] matrices from the
QDF are accurate from 1 MHz to at least 10 GHz.
Following this we will present the full-wave analysis in Sections 9.4 and 9.5, from
which we extract the scattering parameters [S]. The emphasis of this chapter will be
401

Wavelets in Electromagnetics and Device Modeling. George W. Pan
Copyright
¶ 2003 John Wiley & Sons, Inc.
ISBN: 0-471-41901-X
402 WAVELETS IN PACKAGING, INTERCONNECTS, AND EMC
given to packaging and interconnects of high-speed digital circuits and systems and
the implementation of numerical algorithms using wavelets.
9.1 QUASI-STATIC SPATIAL FORMULATION
In this section the wavelet expansion method in conjunction with the boundary ele-
ment method (BEM) is applied to the evaluation of the capacitance and inductance
matrices of multiconductor transmission lines in multilayered dielectric media. The
integral equations obtained by using a Green function above a grounded plane are
solved by Galerkin’s method, with the unknown total charge expanded in terms of
orthogonal wavelets in L
2
([0, 1]). The unknown functions defined in finite intervals
are expanded in terms of wavelets in L
2
([0, 1]), as discussed in Chapter 4. Adopting
the geometric representation of the BEM converts the 2D problem into a 1D prob-
lem and provides a versatile and accurate treatment of curved conductor surfaces and
dielectric interfaces. A sparse matrix equation is developed from the set of integral
equations. This equation is extremely valuable for solving a large system of equa-
tions. We will compare the numerical QSF results with previously published data
and demonstrate good agreement between the two sets of results.
Recently Nekhla reported in [19] that by modifying our wavelet-BEM ap-
proach [20], “The proposed algorithm has a major impact on the speed and accuracy
of physical interconnect parameter extraction with speedup reaching 10
3
for even

moderately sized problems.”
9.1.1 What Is Quasi-static?
In digital and microwave circuits and systems, the electromagnetic (EM) modeling
was based on the quasi-static method. The distributed circuit parameters obtained are
inductance L(H/m), capacitance C(F/m), resistance R(/m), and conductance
G(S/m), all expressed per unit length. These parameters are frequency-independent
under the quasi-static assumption. The quasi-static method assumes:
(1) The wavelength of interest is much greater than the dimensions of the cir-
cuit/subsystems under consideration. Typically f < 3 ∼ 5 GHz.
(2) The longitudinal fields and transverse currents are negligible, which leads to
k
2
= k
2
x
+ k
2
y
+ k
2
z
≈ k
2
z
,wherek
z
is the wavenumber in the direction of
propagation.
(3) Ohmic loss is low so that small perturbation is applicable.
(4) The linear dimension of the transmission line cross section is much greater

than δ (skin depth). As a result current flows only on the conductor surface.
Equivalently the microstrip thickness t and width w satisfy w  t  δ,and
thus internal inductance L
int
can be neglected, and L = L
ext
.
These assumptions no longer hold for high-speed electronic packaging applications.
For instance, for typical multichip module (MCM) structures, the cross section of
QUASI-STATIC SPATIAL FORMULATION 403
the transmission lines is w × t = 8 × 6 µm. For such a structure the dc resistance
≈ 400 /m at 1 GHz with copper of conductivity σ = 5.8 ×10
7
S/m and skin depth
δ = 1/
√
π f µσ ≈ 2 µm. The signal frequency bandwidth ranges from 10 MHz to
10 GHz, and the corresponding skin depths are from δ = 20 to δ = 0.7 µm. Thus
the surface resistance formula
R
s
=
1
σδ
=

π f µ
σ
is not applicable, since we do not have w  t  δ. In addition the small perturbation
approach does not apply due to relatively high ohmic losses. Nonetheless, the quasi-

static approximation is still widely used, in particular, for capacitance computations.
The wave phenomena are governed by the Helmholtz equation
(
2
+ k
2
)φ( x , y, z) = 0, (9.1.1)
where φ(x, y, z) is the potential, k = ω
√
µ =

k
2
x
+ k
2
y
+ k
2
z
is the wavenumber.
Let
φ(x, y, z) = V (x, y)e
±jk
z
z
, (9.1.2)
where V (x, y) is the potential profile in the transverse plane. Substituting (9.1.2) into
(9.1.1), we obtain


∂
2
∂x
2
+
∂
2
∂y
2

+ (k
2
− k
2
z
)

V (x, y) = 0. (9.1.3)
Under quasi-static assumption (2), one has k ≈ k
z
. Hence (9.1.3) becomes

∂
2
∂x
2
+
∂
2
∂y

2

V (x, y) = 0. (9.1.4)
Equtation (9.1.4) is a 2D Laplace equation, which is much simpler then the Helm-
holtz equation (9.1.1). The static nature of (9.1.4) gives the name of this approach as
quasi-static. The prefix “quasi-” is necessary because the wave does propagate along
the ∓ˆz direction. The quasi-static (quasi-TEM) method is very popular because of
its simplicity in mathematics.
9.1.2 Formulation
Figure 9.1 shows the transmission line system under consideration. An arbitrary
number of conductors N
c
is embedded in a dielectric slab consisting of an arbitrary
number of individual layers N
d
. A perfectly conducting ground plane extends from
x =−∞to x =∞. The system is uniform in the y direction. The conductors are
404 WAVELETS IN PACKAGING, INTERCONNECTS, AND EMC
GND
ε
d
ε
d
d
ε
d
ε
0
xO
z

m+1
m
m-1
m-2
m+1
m
m-1
m+1
m
m-1
FIGURE 9.1 Geometry of multiconductor multilayer transmission lines (MMTL).
perfectly lossless and can possess either a finite cross section or be infinitesimally
thin.
The integral equation formulation for this system is derived in [1]. For ease of
reference, we briefly repeat the basic formulation here. The integral equations solved
for the unknown total charge distribution σ
T
(
␳
) can be obtained as follows:
1
2π
0
J

j=1

l
j
σ

T
(
␳

) ln
|
␳
−
␳

|
|
␳
−
␳

|
dl

= V
c
(
␳
) = const. (9.1.5)
on the conductor surfaces, and

+
(
␳
) + 

−
(
␳
)
2
0


+
(
␳
) − 
−
(
␳
)

σ
T
(
␳
) +
1
2π
0
J

j=1

l

j
− σ
T
(
␳

) ·

␳
−
␳

|
␳
−
␳

|
2
−
␳
−
␳

|
␳
−
␳

|

2

·ˆn(
␳
) dl

= 0 (9.1.6)
on the dielectric-to-dielectric interface. Here ρ =
√
x
2
+ z
2
, l
j
is the contour of
the jth interface above the ground plane,
␳

is the image point of
␳

about the
ground plane, and J is the total number of the interfaces (including conductor-to-
QUASI-STATIC SPATIAL FORMULATION 405
dielectric interfaces and dielectric-to-dielectric interfaces);

− denotes the Cauchy
principal value of the integral, and ˆn(
␳

) is the unit normal vector at
␳
. The side of the
curve l
j
is referred to as the “positive” side if ˆn(
␳
) points away from the curve, while
the other side is called its “negative” side; 
+
(
␳
) and 
−
(
␳
) denote the permittivity
on the positive and negative sides, respectively, of the interface that
␳
approaches.
In order to obtain the capacitance matrix [C], the integral equations (9.1.5) and
(9.1.6) must first be solved for the total charge distribution σ
T
(
␳
), with V
c
assigned
as a unity voltage on each particular conductor surface l
j

as zero voltage on the
other conductors. After obtaining the total charge distribution σ
T
(
␳
), the free charge
distribution σ
F
(
␳
) on the conductors can be evaluated by
σ
F
(
␳
) =
(
␳
)

0
σ
T
(
␳
)
for the conductors of finite cross section, and
σ
F
(

␳
) =

+
(
␳
) + 
−
(
␳
)
2
0
σ
T
(
␳
) +

+
(
␳
) − 
−
(
␳
)
2π
0
J


j=1

l
j
− σ
T
(
␳

)
·

␳
−
␳

|
␳
−
␳

|
2
−
␳
−
␳

|

␳
−
␳

|
2

·ˆn(
␳
) dl

(9.1.7)
for infinitesimally thin strips. The total free charge Q
i
(per unit length in the z di-
rection) on conductor l
i
corresponding to this potential distribution yields the ele-
ment C
ij
(i, j = 1, 2, ,N
c
) of the capacitance matrix. The external inductance
matrix [L] is related to the vacuum capacitance matrix [C
v
] by the simple formula
[L]=
0
µ
0

[C
v
]
−1
. The vacuum capacitance matrix [C
v
] itself is the capacitance
matrix of the same conductor system where all dielectrics have been replaced by a
vacuum.
The previous integral equations, (9.1.5) and (9.1.6), need to be solved numerically
for the unknown charge distribution σ
T
(
␳
). This distribution on each interface is
expanded in terms of basis functions
σ
T
(
␳
) 
M

m=1
g
m−1
(
␳
)σ
Tm

, (9.1.8)
where g
m−1
(
␳
)(m = 1, 2, ,M) are the basis functions, σ
Tm
(m = 1, 2, ,M)
are the unknown coefficients to be determined, and M is the total number of the
bases.
We use Galerkin’s method for the testing procedure. Using (9.1.8), a set of linear
algebraic equations in matrix form can be derived from integral equations (9.1.5) and
(9.1.6) [1] as
[
A
nm
][
σ
Tm
]
=
[
B
n
]
, (9.1.9)
where the elements of the matrices are
406 WAVELETS IN PACKAGING, INTERCONNECTS, AND EMC
A
nm

=
J

j1=1

l
j1
g
n−1
(
␳
) ·

1
2π
0
J

j2=1

l
j2
g
m−1
(
␳

) · ln

|

␳
−
␳

|
|
␳
−
␳

|

dl


dl
(9.1.10)
B
n
=
J

j1=1

l
j1
g
n−1
(
␳

)V
c
(
␳
) dl, (9.1.11)
for those g
n−1
(
␳
) defined on the conductor-to-dielectric interfaces, and
A
nm
=
J

j1=1

l
j1
g
n−1
(
␳
) ·


+
(
␳
) + 

−
(
␳
)
2
0


+
(
␳
) − 
−
(
␳
)

g
m−1
(
␳
)
+
1
2π
0
J

j2=1


l
j2
− g
m−1
(
␳

) ·

␳
−
␳

|
␳
−
␳

|
2
−
␳
−
␳

|
␳
−
␳


|
2

·ˆn(
␳
) dl


dl
(9.1.12)
B
n
= 0, (9.1.13)
for those g
n−1
(
␳
) defined on the dielectric-to-dielectric interfaces.
After A
nm
(n = 1, 2, ,M;m = 1, 2, ,M) and B
n
(n = 1, 2, ,M)
have been calculated, (9.1.9) produces M simultaneous equations in M unknowns,
σ
Tm
(m = 1, 2, ,M). These simultaneous equations can then be solved for σ
Tm
(m = 1, 2, ,M) in terms of the potential V
c

(
␳
) on the conductors.
9.1.3 Orthogonal Wavelets in
L
2
([0, 1])
Orthogonal periodic wavelets in L
2
([0, 1]) were studied in great detail in Chapter 4.
We will review the relevant material briefly here.
Given a multiresolution analysis with scalet ϕ(x) and wavelet ψ(x) in L
2
(R), the
wavelets in L
2
([0, 1]) are
ϕ
per
m,n
(x) =

k∈Z
ϕ
m,n
(x + k), (9.1.14)
ψ
per
m,n
(x) =


k∈Z
ψ
m,n
(x + k), (9.1.15)
and V
per
m
= clos
L
2
([0,1])
{ϕ
per
m,n
(x) : n ∈ Z}, W
per
m
= clos
L
2
([0,1])
{ψ
per
m,n
(x) : n ∈ Z}.
It can be shown that V
per
m
are all identical one-dimensional spaces containing only

the constant functions for m ≤ 0, and W
per
m
={∅}for m ≤−1. Thus we only need
to study V
per
m
and W
per
m
for m ≥ 0. Moreover it can easily be verified that
V
per
m+1
= V
per
m
⊕ W
per
m
QUASI-STATIC SPATIAL FORMULATION 407
and
clos
L
2


m∈N
V
per

m

= L
2
([0, 1]),
where N is the set of nonnegative integers. Hence there is a ladder of multiresolution
spaces
V
per
0
⊂ V
per
1
⊂ V
per
2
⊂···
with successive orthogonal complement W
per
0
, W
per
1
, W
per
2
, , and orthonormal
bases {ϕ
per
m,n

(x)}
n=0, ,2
m
−1
in V
per
m
, {ψ
per
m,n
(x)}
n=0, ,2
m
−1
in W
per
m
for m ∈ N .In
particular, that
{ϕ
per
0,0
}

{ψ
per
m,n
: m ∈ N , n = 0, ,2
m
− 1}

constitute an orthonormal basis in L
2
([0, 1]). For simplicity, we relabel this basis as
follows:
g
0
(x) = ϕ
per
0,0
(x) = 1
g
1
(x) = ψ
per
0,0
(x)
g
2
(x) = ψ
per
1,0
(x)
g
3
(x) = ψ
per
1,1
(x) = g
2


x −
1
2

.
.
.
g
2
m
(x) = ψ
per
m,0
(x)
.
.
.
g
2
m
+n
(x) = ψ
per
m,n
(x)
= g
2
m
(x − n2
−m

), 0 ≤ n ≤ 2
m
− 1
.
.
.
These Daubechies periodic scalets were illustrated in Fig. 4.7. For any f (x) ∈
L
2
([0, 1]), the approximation at the resolution 2
m
can be defined as the projection in
V
per
m
,
f (x)  P
m
f (x) =
2
m
−1

k=0
f
k
g
k
(x)
where P

m
is the orthogonal projection operator onto V
per
m
and f
k
is the inner product
of f (x) and g
k
(x).
408 WAVELETS IN PACKAGING, INTERCONNECTS, AND EMC
9.1.4 Boundary Element Method and Wavelet Expansion
Geometrical Representation
Before considering the details of this problem,
we will assume that most curves {l
j
} are closed for the purpose of expressing the
unknown charge distribution. Roughly speaking, there are four types of contours:
(1) the contour of the conductor with finite cross section, (2) the contour along
the infinitesimally thin metal strip, (3) the contour along the dielectric-to-dielectric
interface from −∞ to +∞, and (4) the contour along the dielectric-to-dielectric
interface from −∞ to +∞, with some spaces of discontinuity wherever there is a
conductor along the interface. We will examine the four types of contours one by
one. In the first place, all the contours except type (4) are geometrically continuous.
Moreover the contour of type (1) is closed geometrically. The contour of type (2)
can be considered to be closed, since the charge distribution has the same behavior
(singularity) at its two edge points. Similarly the contour of type (3) can also be
viewed as closed since no charge exists at infinity, and thus the charge distribution
gives the same value of zero at the two ends (−∞ and +∞) of the contour.
In the case of of type (4), the contour intersects the conductor at two points if

the conductor is lying along the contour and creates a discontinuity space for that
contour. We must employ intervallic wavelets, instead of periodic wavelets.
Since the periodized wavelets are defined in L
2
([0, 1]), one must map each of the
contours {l
j
} onto the interval [0, 1]. For an arbitrary contour l
j
, we take two steps:
(1) Use the conventional boundary element method to discretize the contour into
a series of boundary elements, and then map each of the boundary elements
onto 1D standard elements through the shape functions or interpolation func-
tions [3, 22].
(2) Map the standard elements into corresponding portions of interval [0, 1].A
linear map is sufficient for this step.
This procedure can be precisely formulated in mathematical language as well. In
step (1), the global coordinates
␳
are expressed in terms of the local coordinate ξ of
a standard element [3]:
␳
=
M
e

i=1
N
i
(ξ)

␳
i
= 
1
(ξ), (9.1.16)
where M
e
is the number of the interpolation nodes in the local standard element,
N
i
(ξ) is the shape function referred to node i of the local standard element, and
␳
i
are the global coordinates of node i of the actual element. The shape functions
{
N
i
(ξ)
}
are given in standard finite element or boundary element books and literature
(e.g., [3, 22]).
Upon inspecting (9.1.16), we can conclude that (9.1.16) maps the standard ele-
ment in local coordinates onto the actual element, which may have a quite arbitrary
or distorted shape, in global coordinates. The node
␳
i
in the actual element corre-
QUASI-STATIC SPATIAL FORMULATION 409
sponds to the node i in the standard element (by definition, N
i

(ξ) is assumed to have
a unity value at node i and zero at all other nodes of the element).
In step (2), the standard elements corresponding to the actual elements from con-
tour l
j
are mapped into the subintervals [ζ
0
,ζ
1
], [ζ
1
,ζ
2
], ,[ζ
K
j
−1
,ζ
K
j
] of in-
terval [0, 1],whereK
j
is the number of the elements from contour l
j
and 0 =
ζ
0
<ζ
1

<ζ
2
< ··· <ζ
K
j
= 1 (e.g., one can simply assume that ζ
k
= k/K
j
,
k = 1, ,K
j
− 1). The map between the local coordinate ζ in interval [0, 1] and
the local coordinate ξ in the kth standard element of contour l
j
can be written as
ζ = ζ
k−1
+ (ζ
k
− ζ
k−1
) · ξ, (9.1.17)
or
ξ =
ζ − ζ
k−1
ζ
k
− ζ

k−1
, (9.1.18)
where k = 1, 2, ,K
j
. Combining (9.1.16) and (9.1.18), we obtain a map between
the global coordinates
␳
and the local coordinate ζ in interval [0, 1]:
␳
= 
1

ζ − ζ
k−1
ζ
k
− ζ
k−1

= 
2
(ζ ). (9.1.19)
The maps (9.1.16) through (9.1.19) establish the conversions among the local coor-
dinate ξ, the local coordinate ζ and the global coordinates
␳
.
Source Representation
Now we may define the basis functions {g
m−1
(

␳
)}.For
simplicity and generality, the basis functions will not be directly defined over all the
contours in terms of a set of global coordinates, but rather over interval [0, 1] since
each of the contours can be related to interval [0, 1] through the map described by
(9.1.19). By using the conversion between the global coordinates
␳
and the local
coordinate ζ for each individual contour, we can easily obtain the basis functions
of the individual contour in the set of global coordinates. For the unknown charge
distribution along contour l
j
, expansion (9.1.8) can now accurately be written as the
projection in V
per
m
h
(about ζ ):
σ
T
(
␳
)  P
m
h
σ
T
(
␳
) =

M
j

m=1
g
m−1


−1
2
(
␳
)

σ
Tm
, (9.1.20)
where 
−1
2
denotes the inverse map of 
2
, g
m−1
(ζ ) represents the orthogonal
wavelets in L
2
([0, 1]),andM
j
= 2

m
h
is the number of the wavelet bases used
for expressing the unknown charge distribution on contour l
j
. Because 
−1
2
maps
contour l
j
into interval [0, 1], the basis functions {g
m−1
[
−1
2
(
␳
)]} are well defined.
It has been shown [24] that if σ
T
is smooth with a finite number of discontinuities,
the error between σ
T
(ζ ) and P
m
h
σ
T
(ζ ) is bounded:

||σ
T
(ζ ) − P
m
h
σ
T
(ζ ) || ≤ C2
−m
h
s
, (9.1.21)
410 WAVELETS IN PACKAGING, INTERCONNECTS, AND EMC
where C and s are some positive constants, respectively, relating to ||σ
T
(ζ ) ||and the
smoothness of σ
T
(ζ ). The function σ
T
(ζ ) with higher-order (piecewise) continuity
has larger s value and thus faster error decay. Moreover the approximation error of
expansion (9.1.20) can be estimated as
||σ
T
(
␳
) − P
m
h

σ
T
(
␳
) || ≤ C
d
||σ
T
(ζ ) − P
m
h
σ
T
(ζ ) ||
≤ CC
d
2
−m
h
s
,
where C
d
is the tight upper bound of the Jacobian of the transformation 
2
(ζ ).That
is, the approximation error of (9.1.20) has exponential decay with respect to the
resolution level m
h
.

Matrix Equation
Based on the preceding source expansion, a set of linear al-
gebraic equations is obtained from integral equations (9.1.5) and (9.1.6) by using
Galerkin’s method. This set is matrix form described by (9.1.9) if the elements of the
matrices are computed by replacing {g
m−1
(
␳
)} with {g
m−1
[
−1
2
(
␳
)]} in equations
(9.1.10) through (9.1.13), namely
A
nm
=
J

j1=1

l
j1
g
n−1



−1
2
(
␳
)

·

1
2π
0
J

j2=1

l
j2
g
m−1


−1
2
(
␳

)

· ln


|
␳
−
␳

|
|
␳
−
␳

|

dl


dl,
(9.1.22)
B
n
=
J

j1=1

l
j1
g
n−1



−1
2
(
␳
)

V
c
(
␳
) dl, (9.1.23)
for those g
n−1
[
−1
2
(
␳
)] defined on the conductor-to-dielectric interfaces, and
A
nm
=
J

j1=1

l
j1
g

n−1


−1
2
(
␳
)

·


+
(
␳
) + 
−
(
␳
)
2
0


+
(
␳
) − 
−
(

␳
)

g
m−1


−1
2
(
␳
)

+
1
2π
0
J

j2=1

l
j2
−g
m−1


−1
2
(

␳

)

·

␳
−
␳

|
␳
−
␳

|
2
−
␳
−
␳

|
␳
−
␳

|
2


·ˆn(
␳
) dl


dl,
B
n
= 0, (9.1.24)
for those g
n−1
[
−1
2
(
␳
)] defined on the dielectric-to-dielectric interfaces.
QUASI-STATIC SPATIAL FORMULATION 411
Evaluation of Integrals
In practice, integrals in (9.1.22) through (9.1.24) can be
evaluated numerically in either the ζ domain or the ξ domain. We choose the ξ do-
main for our numerical computations in accordance with the conventional boundary
element analysis. Without loss of generality, let us consider the following integral
T
l
j
(
␳
0
) =


l
j
g
m−1


−1
2
(
␳
)

R(
␳
0
,
␳
) dl.
Note that the integrals in (9.1.22) through (9.1.24) are equivalent to this 1D integral
with a particular form of the kernel function R(
␳
0
,
␳
). Using the maps (9.1.16),
(9.1.17), and (9.1.19), we have
T
l
j

(
␳
0
) =
k=K
j

k=1

1
0
g
m−1

ζ
k−1
+ (ζ
k
− ζ
k−1
) · ξ

· R

␳
0
,
1
(ξ)


| D |dξ, (9.1.25)
where | D | is the Jacobian of the transformation between the global coordinates
␳
and the local coordinate ξ of the kth standard element of contour l
j
.
The Jacobian that defines the map of (9.1.16) can be obtained from the expression
for the differential length
dl =

(dx)
2
+ (dz)
2
=




dx
dξ

2
+

dz
dξ

2



dξ.
The Jacobian is then calculated from the following equation:
| D |=

(D
x
)
2
+ (D
z
)
2
,
where
D
x
=
dx
dξ
=
M
e

i=1
dN
i
(ξ)
dξ
x

i
,
D
z
=
dz
dξ
=
M
e

i=1
dN
i
(ξ)
dξ
z
i
,
and where x
i
and z
i
are, respectively, the x and z components of
␳
i
.
From the case of the orthogonal wavelet on the real line, we can use definitions
(9.1.14), (9.1.15) and (9.1.16) to obtain the periodic orthogonal wavelet {g
m−1

(ζ )}.
Integration (9.1.25) can be readily performed by standard numerical algorithms such
as Gaussian quadrature [25].
412 WAVELETS IN PACKAGING, INTERCONNECTS, AND EMC
9.1.5 Numerical Examples
Based on the technique presented in the preceding subsections, a program has been
designed to compute the capacitance and external inductance matrices of multicon-
ductor transmission lines in multilayered dielectrics. Two numerical examples are
presented in this subsection. When using wavelets on the real line to solve problems
with finite intervals, improper selection of the wavelets can result in nonphysical so-
lutions. In contrast, any type of wavelets on the real line can be used for the construc-
tion of the wavelets in L
2
([0, 1]), although there may be some discrepancy in their
smoothness, as seen in Chapter 4. However, since the derivatives of the unknown
function σ
T
(
␳
) are of order zero in the integral equations under consideration, a set
of basis functions with C
0
continuity is sufficient to yield a convergent solution. In
the following computations the Daubechies wavelets are employed to construct the
orthogonal wavelets in L
2
([0, 1]).
Example 1
Thin microstrip line of width W above a dielectric substrate of thickness
Hand

r
= 6. We have studied this example of an infinitesimally thin microstrip line.
The characteristic impedances obtained by this technique were compared with those
from the conventional boundary element method (BEM) [3], the method of moments
(MoM) [1], and the more accurate formulas from [29] in Table 9.1. The results of the
conventional BEM were obtained using 16 subsections (33 bases) on the strip and
30 subsections (62 bases) at the dielectric interface; those of the MoM were obtained
by using 12 subsections on the strip and 30 subsections at the dielectric interface.
Two sets of the results from this technique are presented in columns A and B, with
M
1
= M
2
= 32 and M
1
= M
2
= 16 respectively, where M
1
is the number of the
wavelet bases used on the strip while M
2
is the number at the dielectric interface.
Table 9.1 provides an interesting insight into the wavelet expansions. Taking the
column labeled “Hammerstad” as a set of “ground truth” or standard references, we
see that the results from this technique with 64 bases (column A) give approximately
the same accuracy as the conventional BEM, although the BEM results are obtained
by using about 50% more (total 95) bases. The results from this technique with 32
bases (column B) exhibit a higher degree of accuracy than the MoM despite the fact
that the MoM uses approximately one-third more (total 42) bases for its calculations.

TABLE 9.1. Characteristic Impedances for the Thin Microstrip Line (in Ohms)
W/H A B BEM MoM Hammerstad
0.4 90.5779 91.3783 90.7758 92.2785 90.3339
0.7 72.9504 73.2748 73.0898 73.9626 72.7516
1.0 62.0383 62.3342 62.1102 62.8109 61.8397
2.0 42.4233 42.5918 42.4118 42.9980 42.2600
4.0 26.5482 26.6498 26.5236 26.9709 26.4593
10. 12.7707 12.8134 12.7351 12.9961 12.7198
Source: G. Wang, G. Pan, and B. Gilbert. IEEE Trans. Microw. Theory Tech., 43(3), 664–675, March
1995;
c
 1995 IEEE.
QUASI-STATIC SPATIAL FORMULATION 413
Finally, comparison between the results of column A and column B shows that this
technique gives better accuracy with higher resolution approximation.
Theoretically, it is not a surprise that the wavelet expansions converge more
quickly; that is, fewer coefficients are required by wavelets to represent a given func-
tion than by other expansions, since this is a well-known result from wavelet theory
and has been extensively studied in Chapter 2. One of the most attractive features of
wavelets is that they give completely local information on the functions analyzed. It
can be shown that if a function does not have uniform smoothness, for instance, if
a smooth function possesses discontinuities, there is an optimal way to approximate
the function using low resolution wavelets everywhere and adding high resolution
wavelets near the singularities [24].
Example 2 Multiconductor Transmission Lines above a Thick Substrate.
Shown
in Fig. 9.2 is a 10-conductor transmission line system. This problem arises during
the modeling of CMOS chips, where the transmission lines are far above the ground
plane in comparison to the cross-sectional dimensions or the separations of the in-
dividual conductors. For such structures the MoM approach frequently yields either

singular matrices or nonphysical solutions [3]. In order to test the stability of this
technique, we applied it to a ten conductor transmission line with a thick substrate.
Tables 9.2 to 9.5 list the resulting capacitance and inductance matrices computed
with this technique and the BEM with special edge treatment [3]. The BEM solutions
were computed by using 160 subsections (360 bases) on the conductor surfaces and
190 subsections (392 bases) at the dielectric interfaces. These solutions are taken
from [3]. The results from this technique were obtained by using 160 bases on the
conductor surfaces and 256 bases at the dielectric interfaces. The self-capacitance
of the ith conductor can be obtained by summing up all the elements at the i th row
of the capacitance matrix [C]. Each of the self-capacitance values must be positive;
otherwise, the results will be nonphysical solutions.
~
12 3 4 5
67 8 9 10
2
6 9 12 15
1.5
1
ε = ε
ε = 11.0 ε
10
30
ε = 5.0 ε
20
600
~
FIGURE 9.2 Ten conductors in a layered medium (in µm).
TABLE 9.2. Wavelet Technique: Capacitance Matrix [
C
] (in pF/m)

307.4 −41.10 −11.35 −6.330 −5.452 −219.6 −4.932 −1.389 −0.8246 −0.7600
−41.12 319.7 −27.96 −7.812 −5.043 −4.999 −217.5 −3.485 −0.9775 −0.6821
−11.35 −27.96 309.9 −24.24 −8.632 −1.377 −3.474 −218.4 −3.103 −1.154
−6.316 −7.794 −24.23 302.3 −24.70 −0.8192 −0.9580 −3.117 −218.9 −3.304
−5.440 −5.029 −8.624 −24.74 290.2 −0.7487 −0.6584 −1.136 −3.259 −221.5
−218.8 −5.019 −1.373 −0.8105 −0.7303 231.7 −2.063 −0.3899 −0.1799 −0.1349
−4.967 −216.6 −3.492 −0.9523 −0.6406 −2.064 231.6 −1.176 −0.2495 −0.1332
−1.386 −3.526 −217.3 −3.150 −1.127 −0.3896 −1.178 230.6 −0.8550 −0.2383
−0.8200 −0.9843 −3.162 −217.6 −3.306 −0.1803 −0.2511 −0.8580 229.6 −0.7478
−0.7467 −0.6755 −1.154 −3.343 −220.5 −0.1358 −0.1351 −0.2399 −0.7465 230.4
414
SPATIAL DOMAIN LAYERED GREEN’S FUNCTIONS 415
TABLE 9.3. Wavelet Technique: Inductance Matrix [
L
] (in nH/m)
1407.0 999.8 831.9 721.3 638.0 1306.0 998.7 831.8 721.4 638.1
999.8 1405.0 935.1 774.8 671.7 998.7 1304.0 934.7 774.8 671.8
831.9 935.1 1407.0 888.0 731.7 831.8 934.7 1307.0 887.7 731.8
721.3 774.8 888.0 1409.0 850.2 721.4 774.8 887.7 1309.0 850.1
638.0 671.7 731.7 850.2 1411.0 638.1 671.8 731.8 850.1 1310.0
1306. 998.7 831.8 721.4 638.1 1407.0 1000.0 832.1 721.6 638.3
998.7 1304.0 934.7 774.8 671.7 1000.0 1405.0 935.4 775.1 671.9
831.8 934.7 1307.0 887.7 731.8 832.1 935.4 1408.0 888.2 732.0
721.4 774.8 887.7 1309.0 850.1 721.6 775.1 888.2 1410.0 850.5
638.1 671.8 731.8 850.1 1310.0 638.3 671.9 732.0 850.5 1411.0
The sizes for matrix [A] are, respectively, 752 ×752 and 416 ×416 for the BEM and
the wavelet technique. For such a relatively large matrix [A], the sparsity is more
significant. As mentioned in Example 1 (9.1.6), is likely to produce sparse linear
algebraic equations for both the wavelet-base approach and the BEM. Hence, we
will only examine the sparsity for the upper part of matrix [A], which comes from

(9.1.5). The upper part of matrix [A] is obtained by using this technique under a
threshold of 10
−3
and is a 160 × 416 sparse matrix. In sharp contrast, a 360 × 752
full dense matrix is generated by the BEM under the same threshold.
9.2 SPATIAL DOMAIN LAYERED GREEN’S FUNCTIONS
In this section we present a new approach to capacitance computation, which is more
efficient than the method presented in the previous section. The major improvements
are as follows:
(1) Under the formulation of the free-space Green function, polarization charges
at the dielectric–dielectric interfaces have to be computed as unknowns in ad-
dition to the free charges on conductor surfaces. In contrast, we now use the
layered Green’s function that was proposed by DeZutter in [13] and approx-
imated in closed forms in [32]. Under the layered Green function, only the
free surface charges are unknown in the problem, resulting in much smaller
impedance matrix.
(2) Only standard wavelets are employed to expand the free surface charges on
closed contours of the conductor surfaces. No periodic or intervallic wavelets
are necessary, and so a much simpler treatment is possible.
(3) Replacing the Daubechies wavelets with Coifman wavelets allows single-
point quadrature and leads to fast matrix filling.
TABLE 9.4. BEM Solution: Capacitance Matrix [
C
] (in pF/m)
308.5 −41.50 −11.42 −6.335 −5.417 −219.6 −5.019 −1.402 −0.8288 −0.7474
−41.51 321.2 −28.25 −7.853 −5.038 −5.081 −217.8 −3.577 −0.9985 −0.6799
−11.43 −28.25 312.0 −24.48 −8.665 −1.384 −3.539 −219.2 −3.214 −1.164
−6.339 −7.854 −24.47 304.9 −24.93 −0.8126 −0.9598 −3.198 −220.3 −3.382
−5.417 −5.036 −8.660 −24.92 291.8 −0.7275 −0.6423 −1.137 −3.360 −222.1
−220.3 −5.073 −1.380 −0.8094 −.7240 233.4 −2.090 −0.3937 −0.1811 −0.1332

−5.019 −218.7 −3.542 −0.9590 −0.6409 −2.091 233.9 −1.201 −0.2544 −0.1336
−1.403 −3.580 −220.2 −3.200 −1.137 −0.3943 −1.201 233.7 −0.8819 −0.2420
−0.8282 −0.9984 −3.216 −221.3 −3.363 −.1814 −0.2545 −0.8820 233.5 −0.7688
−0.7448 −0.6777 −1.162 −3.377 −222.9 −0.1333 −0.1335 −0.2417 −0.7683 233.0
416
SPATIAL DOMAIN LAYERED GREEN’S FUNCTIONS 417
TABLE 9.5. BEM Solution: Inductance Matrix [
L
] (in nH/m)
1398.0 993.1 826.2 716.5 633.7 1297.0 992.1 826.2 716.6 633.8
993.0 1396.0 928.8 769.5 667.1 992.1 1295.0 928.4 769.6 667.2
826.2 928.9 1398.0 881.9 726.7 826.2 928.4 1298.0 881.8 726.8
716.5 769.6 882.0 1400.0 844.5 716.6 769.7 881.8 1300.0 844.4
633.7 667.1 726.8 844.5 1402.0 633.9 667.3 726.9 844.4 1301.0
1297.0 992.0 826.2 716.5 633.8 1398.0 993.4 826.6 716.8 634.0
992.0 1295.0 928.4 769.6 667.2 993.4 1396.0 929.1 769.9 667.4
826.2 928.4 1298.0 881.7 726.8 826.6 929.2 1399.0 882.3 727.1
716.6 769.6 881.7 1299.0 844.3 716.8 769.9 882.3 1401.0 844.8
633.8 667.2 726.8 844.4 1301.0 634.0 667.5 727.1 844.8 1402.0
As discussed in the previous section, the adoption of the geometric representation
of the BEM converts a 2D problem into a 1D problem and provides a versatile and
accurate treatment of curved conductor surfaces. The conductor cross sections of 2D
problems are closed contours. The BEM representation of a contour utilizes the arc
length ζ , varying from 0 to  in circumference; it is in [0, 1] after normalization. In
principle, one needs to utilize periodic wavelets in L
2
([0, 1]) when the domain of the
problem is over a finite interval. Nevertheless, we find that the standard wavelets are
sufficient to represent the contours. In fact, we now deploy the wavelet bases one by
one on the contour that has been mapped by the BEM onto the interval [0, 1].The

portion of a wavelet basis that is beyond the interval will be lobbed off and relocated
at the opposite end. This procedure is quite similar to the circular convolution in
digital signal processing [33].
9.2.1 Formulation
Suppose that N
c
perfect conductors are placed throughout N
d
nonmagnetic dielectric
layers and the geometry of the dielectric layers are assumed to be uniform in the x
and y directions. The integral equation relating the electrostatic potential V (r) to the
charge density σ(r) is
V (r) =


G(r, r

)σ (r

)dr

. (9.2.1)
Considering the case that a unit source is in layer m (see Fig. 9.1). The 3D Green’s
function satisfies Poisson’s equation

2
G
3D
(x, y, z | x
0

, y
0
, z
0
) =
1

δ(x − x
0
)δ(y − y
0
)δ(z − z
0
). (9.2.2)
418 WAVELETS IN PACKAGING, INTERCONNECTS, AND EMC
Spatial domain and spectral domain Green’s functions are related by the 2D Fourier
transform pair as
G
3D
(x, y, z | x
0
, y
0
, z
0
) =
1
(2π)
2


∞
−∞

∞
−∞
dα dβe
−jα(x −x
0
)−jβ(y−y
0
)
×
˜
G
3D
(α, β, z | x
0
, y
0
, z
0
)
and
˜
G
3D
(α, β, z | x
0
, y
0

, z
0
) =

∞
−∞

∞
−∞
dx dye
jα(x−x
0
)−jβ(y−y
0
)
× G
3D
(x, y, z | x
0
, y
0
, z
0
),
where
˜
G
3D
(α, β, z | x
0

, y
0
, z
0
) is the spectral domain Green function. By taking the
2D Fourier transform with respect to x and y, (9.2.2) becomes

∂
2
∂z
2
− α
2
− β
2

˜
G
3D
(α, β, z | x
0
, y
0
, z
0
) =
1

δ(z −z
0

).
Denoting γ =

α
2
+ β
2
, we can write the z variation of the solution in region m as
˜
G(z | z
0
) =
e
−γ |z−z
0
|
+ B
m
e
γ z
+ D
m
e
−γ z
2
m
γ
. (9.2.3)
To find B
m

and D
m
, we need to use the constraint conditions at z = d
m−1
and z = d
m
(see Fig. 9.1). The descending wave for z > z
0
is a consequence of the reflection of
the ascending wave for z > z
0
at z = d
m
, namely
B
m
e
γ d
m
=
˜

m,m+1
[e
−γ(d
m
−z
0
)
+ D

m
e
−γ d
m
]. (9.2.4)
Similarly
D
m
e
−γ d
m−1
=
˜

m,m−1
[e
γ(d
m−1
−z
0
)
+ B
m
e
γ d
m−1
]. (9.2.5)
Rewriting (9.2.5) as
D
m

= e
γ d
m−1
˜

m,m−1
[e
γ(d
m−1
−z
0
)
+ B
m
e
γ d
m−1
] (9.2.6)
and substituting D
m
into (9.2.4), we have
B
m
=
˜

m,m+1
[e
γ(−2d
m

+z
0
)
+
˜

m,m−1
e
γ(2d
m−1
−2d
m
−z
0
)
]
1 −
˜

m,m+1
˜

m,m+1
e
2γ(d
m−1
−d
m
)
. (9.2.7)

Substituting (9.2.7) into (9.2.6), we arrive at
SPATIAL DOMAIN LAYERED GREEN’S FUNCTIONS 419
D
m
=
˜

m,m−1
[e
γ(2d
m−1
−z
0
)
+
˜

m,m+1
e
γ(2d
m−1
−2d
m
+z
0
)
]
1 −
˜


m,m−1
˜

m,m+1
e
2γ(d
m−1
−d
m
)
. (9.2.8)
C
ASE 1 z > z
0
.Whenz > z
0
,wehave
|z − z
0
|=z − z
0
.
Substituting (9.2.7) and (9.2.8) into (9.2.3) and letting
M
m
=[1 −
˜

m,m−1
˜


m,m+1
e
2(d
m−1
−d
m
)
]
−1
,
we arrive at
˜
G(z | z
0
) =
M
m
2
m
γ
[e
−γ z
+
˜

m,m+1
e
−2γ d
m

+γ z
][e
γ z
0
+
˜

m,m−1
e
2γ d
m−1
−γ z
0
].
C
ASE 2 z < z
0
. In a similar way, for z < z
0
,wehave
˜
G(z | z
0
) =
M
m
2
m
γ
[e

γ z
+
˜

m,m−1
e
2γ d
m−1
−γ z
][e
−γ z
0
+
˜

m,m+1
e
−2γ d
m
+γ z
0
].
Furthermore, if we are looking for the field in region n > m, it can be found by using
the recursive method. For n > m, z > z
0
,
˜
G(z | z
0
) =

A
+
m,n
2
m
γ
(e
−γ z
+
˜

n,n+1
e
−2γ d
n
+γ z
),
A
+
i,i+1
= A
+
i,i
S
+
i,i+1
,
A
+
m,n

= A
+
m,m
n−1

i=m
S
+
i,i+1
,
where A
+
m,m
= M
m
[e
γ z
0
+
˜

m,m−1
e
2γ d
m−1
−γ z
0
].
For n < m, z < z
0

,
˜
G(z | z
0
) =
A
−
m,n
2
m
γ
[e
γ z
+
˜

n,n−1
e
2γ d
n−1
−γ z
],
A
−
i,i−1
= A
−
i,i
S
−

i,i−1
,
A
−
m,n
= A
−
m,m
m

n+1
S
−
i,i−1
,
where
A
−
m,m
= M
m
[e
−γ z
0
+
˜

m,m+1
e
−2γ d

m
+γ z
0
].
420 WAVELETS IN PACKAGING, INTERCONNECTS, AND EMC
In the previous formulas
˜

i,i+1
=

i,i+1
+
˜

i+1,i+2
e
2γ(d
i
−d
i+1
)
1 + 
i,i+1
˜

i+1,i+2
e
2γ(d
i

−d
i+1
)
,
S
+
i,i+1
=
T
i,i+1
1 − 
i+1,i
˜

i+1,i+2
e
2γ(d
i
−d
i+1
)
,
˜

i,i−1
=

i,i−1
+
˜


i−1,i−2
e
2γ(d
i−2
−d
i−1
)
1 + 
i,i−1
˜

i−1,i−2
e
2γ(d
i−2
−d
i−1
)
,
S
−
i,i−1
=
T
i,i−1
1 − 
i−1,i
˜


i−1,i−2
e
2γ(d
i−2
−d
i−1
)
,
and

i, j
=

i
− 
j

i
+ 
j
, T
i, j
=
2
i

i
+ 
j
.

The parameters B
m
, D
m
, M
m
, A
±
i,i+1
, S
±
i,i+1
,
˜

i,±1
,
i, j
, T
i, j
, etc., are the static ver-
sions of their counterparts in [23]. The generalized reflection coefficient
˜

j, j+1
takes
the value of 0 or −1 if the j th layer is a half-space or ( j + 1)th layer is a ground
plane, respectively.
Rearranging these expressions by factoring out all z and z
0

dependencies, we
obtain
˜
G(z | z
0
) =
1
2
m
γ
[K
+
m,n,1
e
γ(z+z
0
−2d
n
)
+ K
+
m,n,2
e
γ(z−z
0
+2(d
m−1
−d
n
))

+ K
+
m,n,3
e
γ(−z+z
0
)
+ K
+
m,n,4
e
γ(−z−z
0
+2d
m−1
)
], z ≥ z
0
, (9.2.9)
˜
G(z | z
0
) =
1
2
m
γ
[K
−
m,n,1

e
γ(z+z
0
−2d
m
)
+ K
−
m,n,2
e
γ(z−z
0
)
+ K
−
m,n,3
e
γ(−z+z
0
+2(d
n−1
−d
m
))
+ K
−
m,n,4
e
γ(−z−z
0

+2d
n
−1)
], z ≤ z
0
, (9.2.10)
where
K
+
m,n,1
= M
m
˜

n,n+1
n−1

j=m
S
+
j, j+1
K
+
m,n,2
= M
m
˜

n,n+1
˜


m,m−1
n−1

j=m
S
+
j, j+1
K
+
m,n,3
= M
m
n−1

j=m
S
+
j, j+1
SPATIAL DOMAIN LAYERED GREEN’S FUNCTIONS 421
K
+
m,n,4
= M
m
˜

m,m−1
n−1


j=m
S
+
j, j+1
,
and
K
−
m,n,1
= M
m
˜

m,m+1
m

j=n+1
S
−
j, j−1
K
−
m,n,2
= M
m
m

j=n+1
S
−

j, j−1
K
−
m,n,3
= M
m
˜

m,m+1
˜

n,n−1
m

j=n+1
S
−
j, j−1
K
−
m,n,4
= M
m
˜

n,n−1
m

j=n+1
S

−
j, j−1
.
Before we determine the closed-form spatial domain Green’s function, we will ap-
proximate the coefficient functions K
±
m,n,i
of the exponentials in terms
K
±
m,n, j
(γ ) = K
±∞
m,n, j
+
N
±
m,n, j

i=1
C
±,i
m,n, j
e
a
±,i
m,n, j
γ
, j = 1, 2, 3, 4, (9.2.11)
where K

±∞
m,n, j
denotes the asymptotic value of K
±
m,n, j
, summation index N
±
m,n, j
is the
number of exponential functions, C
±,i
m,n, j
and a
±,i
m,n, j
are Prony’scoefficients given in
Section 9.2.2.
By using the Fourier transform,
3D:
1
2π

+∞
−∞

+∞
−∞
dα dβe
−j(αx+βy)
e

−γ |z |
γ
=
1

x
2
+ y
2
+ z
2
,
2D:
1
2π

+∞
−∞
dγ e
−jγ x
e
−|γ z |
|γ |
=−ln


x
2
+ z
2


,
we can write the approximated Green’s function for 2D and 3D cases as
G
3D
(r|r
0
) =
1
4π
m
4

j=1
f
3D,±
j
(r |r
0
),
G
2D
(
␳
|
␳
0
) =−
1
2π

m
4

j=1
f
2D,±
j
(
␳
|
␳
0
),
For the 2D case,
422 WAVELETS IN PACKAGING, INTERCONNECTS, AND EMC
f
2D,+
j
(
␳
|
␳
0
) = K
+,∞
m,n, j
ln


(x − x

0
)
2
+ Z
+ 2
j

+
N
+
m,n, j

i=1
C
+,i
m,n, j
ln


(x − x
0
)
2
+ (Z
+
j
+a
+,i
m,n, j
)

2

, (9.2.12)
f
2D,−
j
(
␳
|
␳
0
) = K
−∞
m,n, j
ln


(x − x
0
)
2
+ Z
− 2
j

+
N
−
m,n, j


i=1
C
−,i
m,n, j
ln


(x − x
0
)
2
+ (Z
−
j
+a
−,i
m,n, j
)
2

, (9.2.13)
where j = 1, ,4. For the 3D case, the formulas can be written in a similar way:
f
3D,+
j
(r | r
0
) = K
+,∞
m,n, j

1

(x − x
0
)
2
+ Z
+ 2
j
+ (y − y
0
)
2
+
N
+
m,n, j

i=1
C
+,i
m,n, j
1

(x − x
0
)
2
+ (Z
+

j
+a
+,i
m,n, j
)
2
+ (y − y
0
)
2
,
(9.2.14)
f
3D,−
j
(r | r
0
) = K
−,∞
m,n, j
1

(x − x
0
)
2
+ Z
− 2
j
+ (y − y

0
)
2
+
N
−
m,n, j

i=1
C
−,i
m,n, j
1

(x − x
0
)
2
+ (Z
−
j
+a
−,i
m,n, j
)
2
+ (y − y
0
)
2

,
(9.2.15)
where j = 1, ,4and
Z
+
1
= z + z
0
− 2d
n
,
Z
+
2
= z − z
0
+ 2(d
m−1
− d
n
)
Z
+
3
=−z +z
0
Z
+
4
=−z −z

0
+ 2d
m−1
,
Z
−
1
= z + z
0
− 2d
m
Z
−
2
= z − z
0
Z
−
3
=−z +z
0
+ 2(d
n−1
− d
m
)
Z
−
4
=−z −z

0
+ 2d
n−2
.
SPATIAL DOMAIN LAYERED GREEN’S FUNCTIONS 423
9.2.2 Prony’s Method
The coefficients C
±,i
m,n, j
and a
±,i
m,n, j
in (9.2.12) to (9.2.15) may be computed by
Prony’s method [34]. For ease of reference, we briefly present the major steps of the
Prony method below.
To determine an approximation of the form
f (x )  C
1
e
a
1
x
+C
2
e
a
2
x
+···+C
n

e
a
n
x
,
we assume that values of f (x) are specified on a set of N equally spaced points. By
using a linear change of variables, the data points become x

= 0, 1, 2, ,N − 1.
Say that the interval between x is , and we have x
k
= k ,andx

k
= (x
k
/)−1.
Now we have
f (x) = C
1
e
a
1
x
+C
2
e
a
2
x

+···+C
n
e
a
n
x
= f (x

)
= C
1
e
a
1
(x

+1)
+C
2
e
a
2
(x

+1)
+···+C
n
e
a
n

(x

+1)
= e
a
1
x


(C
1
e
a
1

) + e
a
2
x


(C
2
e
a
2

) +···+e
a
n

x


(C
n
e
a
n

)
= C

1
e
a

1
x

+C

2
e
a

2
x

+···+C


n
e
a

n
x

, (9.2.16)
where

C

n
= C
n
e
a
n

a

n
= a
n
.
Letting µ
n
= e
a


n
, we may rewrite (9.2.16) as
f (x

) = C

1
µ
x

1
+C

2
µ
x

x
+···+C

n
µ
x

n
.
For x

= 0, 1, ,N − 1, the following equations are satisfied:






















C

1
+C

2
+··· +C

n
= f

0
C

1
µ
1
+C

2
µ
2
+··· +C

n
µ
n
= f
1
C

1
µ
2
1
+C

2
µ
2
2

+··· +C

n
µ
2
n
= f
2
.
.
.
C

1
µ
N−1
1
+C

2
µ
N−1
2
+···+C

n
µ
N−1
n
= f

N−1
.
(9.2.17)
When µ’s are unknown, at least 2n equations are needed. Let µ
1
,µ
2
, ,µ
n
be the
roots of the algebraic equation
µ
n
+ α
1
µ
n−1
+ α
2
µ
n−2
+···+α
n−1
µ + α
n
= 0. (9.2.18)
424 WAVELETS IN PACKAGING, INTERCONNECTS, AND EMC
In order to determine the coefficients α
1
,α

2
, ,α
n
, let us take the first (n + 1)
equations from (9.2.17). We multiply the first equation in (9.2.17) by α
n
, the second
by α
n−1
and the nth equation by α
1
, the (n + 1)th equation by 1, and add up the
results






















C

1
α
n
+C

2
α
n
+··· +C

n
α
n
= f
0
α
n
C

1
µ
1
α
n−1

+ C

2
µ
2
α
n−1
+··· +C

n
µ
n
α
n−1
= f
1
α
n−1
.
.
.
C

1
µ
n−1
1
α
1
+ C


2
µ
n−1
2
α
1
+··· +C

n
µ
n−1
n
α
1
= f
n−1
α
1
C

1
µ
n
1
+C

2
µ
n

2
+··· +C

n
µ
n
n
= f
n
.
Hence
LHS = C

1
(α
n
+ µ
1
α
n−1
+···+µ
n−1
1
α
1
+ µ
n
1
)
+ C


2
(α
n
+ µ
2
α
n−1
+···+µ
n−1
2
α
1
+ µ
n
2
)
.
.
.
+ C

2
(α
n
+ µ
n
α
n−1
+···+µ

n−1
n
α
1
+ µ
n
n
)
= 0
RHS = f
0
α
n
+ f
1
α
n−1
+···+ f
n−1
α
1
+ f
n
.
In a similar way a set of N − n − 1 additional equations are obtained










f
n
+ f
n−1
α
1
+··· +f
0
α
n
= 0
f
n+1
+ f
n
α
1
+··· +f
1
α
n
= 0
.
.
.
f

N−1
+ f
N−2
α
1
+··· +f
N−n−1
α
n
= 0.
(9.2.19)
For N = 2n, the following procedures are used:
(1) For given f
0
, f
1
, , f
n−1
, solve (9.2.19) for α
1
,α
2
, ,α
n
.
(2) Using α
1
,α
2
, ,α

n
, find roots of (9.2.18) to obtain µ
1
,µ
2
, ,µ
n
.
(3) Using µ
1
,µ
2
, ,µ
n
, solve (9.2.17) to find C

1
, C

2
, ,C

n
.
Upon approximation of the coefficients C
±i
m,n, j
and a
±i
m,n, j

by using Prony’s method,
the Green’s functions are expressed in an explicit formula with complex numbers.
9.2.3 Implementation of the Coifman Wavelets
The Coifman scalets are employed to solve the integral equation for the charge den-
sity. First we map the circumferences of the conductor contours onto the interval
SPATIAL DOMAIN LAYERED GREEN’S FUNCTIONS 425
–4 – 3 –2 –1 0 1 2 3 4 5 6 7
–1
–0.5
0
0.5
1
1.5
2
x
φ
(x )
FIGURE 9.3 Coifman scalet of order L = 4.
[0, 1]. We then choose the scalets at a certain level  and put them on the interval as a
basis. In doing so, we convert the contour of each conductor with finite cross section
into a finite 1D interval. Thus we have mapped a 2D problem into a 1D problem via
a versatile and accurate treatment of curved conductor surfaces with arbitrary cross
sections. As a result the global coordinates
␳
are expressed in terms of the local co-
ordinate ξ. The unknown charge density σ(
␳

) has been expressed as σ(ξ


), which
is then expanded in terms of Coifman scalet ϕ
l,m
(ξ) as shown in Fig. 9.3. Using the
expansion
σ(ξ

) =

α
l,m
ϕ
l,m
(ξ

),
we write the integral equation (9.2.1) as
1 =


G(ξ, ξ

)

α
m
ϕ
l,m
(ξ)


dξ

.
Applying Galerkin’s testing procedure, we obtain

S
n
ϕ
l,n
(ξ) dξ =

α
m

S
n

S
m
G(ξ, ξ

)ϕ
l,m
(ξ

)ϕ
l,n
(ξ) dξ

dξ.

In matrix form we arrive at
[Z
m,n
][α
m
]=[g
n
],
where
Z
m,n
=

Sn

Sm
G(ξ, ξ

)ϕ
l,m
(ξ

)ϕ
l,n
(ξ) dξ

dξ,
g
n
=


Sn
ϕ
l,n
(ξ) dξ.