Are you sure?
This action might not be possible to undo. Are you sure you want to continue?
METHOD)
DANIEL S. WEILE
RAYMOND A. WILDMAN
GREESHMA PISHARODY
ANURAAG MOHAN
University of Delaware
Newark, Delaware
1. INTRODUCTION
Galerkin’s method and the Rayleigh–Ritz method are per
haps the primary techniques used in the computational
solution of boundary value problems and eigenvalue prob
lems involving linear ordinary and partialdifferential
equations. Galerkin’s method is also used for the disc
retization of integral equations. In modern computational
electromagnetics (CEM), these methods form the basis for
the now ubiquitous method of moments (MoM) and ﬁnite
element method (FEM) [1–6]. While the Rayleigh–Ritz
and Galerkin methods are in principle very different,
they often lead to the same set of discretized equations,
are closely related, and are often discussed as if they were
identical.
Both methods assume that the unknown function in
the problem at hand can be well approximated by a series
of basis or coordinate functions with unknown coefﬁcients,
and both methods create a system of equations to solve for
these unknown coefﬁcients. Because both methods use
basis functions, we will generically refer to Galerkin’s
method and the Rayleigh–Ritz method collectively as
basis function approaches.
In the Rayleigh–Ritz approach, the original problem is
reformulated as a variational problem. More specifically, a
functional (i.e., a function that maps functions to real
numbers) is created that has a stationary point (most of
ten a minimum) at the solution of the differential equa
tion. The basis function series is then substituted into this
functional, transforming the functional into a simple func
tion of the unknown coefﬁcients. Minimizing this function
with respect to the unknown coefﬁcients gives rise to a set
of linear algebraic equations to be solved for the un
knowns.
In the Galerkin approach, the basis function series is
substituted directly into the equation at hand. This equa
tion is then projected onto each of the basis functions in
turn (in a manner to be explained in more detail below),
again giving rise to a set of linear algebraic equations to be
solved for the unknown coefﬁcients.
Before discussing Galerkin’s method and the Rayleigh–
Ritz method in more detail, we note that these methods
are not the only methods used for the discretization of
differential and integral equations. Alternative methods
are used quite often in the CEM community, and a brief
discussion of these approaches serves to deﬁne the Galer
kin and Rayleigh–Ritz methods by contrast. In the differ
ential equations realm, the primary competitors to basis
function approaches are ﬁnitedifference methods [7–9].
Instead of approximating the unknown function as a se
ries of basis functions, ﬁnitedifference methods seek to
ﬁnd the value of the unknown function on a (usually very
regular) grid of points. The derivatives in the differential
equations are then approximated by ﬁnite differences,
leading to a set of linear algebraic equations to be solved
for the values of the unknown function at the grid points.
The primary competitor to basis function approaches in
the ﬁeld of integral equations is the Nystro¨m method
[10,11]. In the Nystro¨m method, the integral in the inte
gral equation is approximated using a numerical quad
rature rule, and the resulting equation is enforced at the
quadrature points. This leads to a set of linear algebraic
equations for the value of the unknown function at the
quadrature points. Note that the primary difference be
tween these techniques and the basis function approaches
discussed here is that the former methods do not use basis
functions at all, but instead seek to ﬁnd the value of the
unknown function at a set of points. Thus, the nomencla
ture ‘‘basis function approaches’’ serves to distinguish the
methods described here from these other pointbased ap
proaches. [Actually, even the pointbased approaches may
be interpreted as basis function approaches with the prop
er choice of basis and testing functions. Nonetheless, as
discussed above, the motivation for the pointbased meth
ods is completely different from that of the basis function
approaches. Most importantly, their governing equations
can be (and usually are) derived without reference to basis
function approaches at all.]
In the remainder of this article we discuss the two basis
function approaches in detail. Section 2 relates a brief
history of the two basis function approaches. Section 3
discusses the Rayleigh–Ritz method and provides exam
ples of its application. The same is done for the Galerkin
approach in Section 4. Finally, Section 5 briefly discusses
the MoM and FEM, and gives a few examples of their
applications.
2. A BRIEF HISTORY OF BASIS FUNCTION APPROACHES
The history of basis function approaches begins with the
prolific third Lord Rayleigh (John William Strutt). Ray
leigh was one of the most accomplished physicists of the
late nineteenth and early twentieth centuries; his accom
plishments include the ﬁrst correct explanation of the
blueness of the sky, and the discovery of the inert gas ar
gon for which he received the Nobel prize in 1904 [12].
(Indeed, Rayleigh was so famous even in his own time that
he succeeded no less a physicist than James Clerk Max
well to become the second Cavendish professor of exper
imental physics at Cambridge University [12].) In his book
The Theory of Sound [13], Rayleigh describes a method for
the approximate determination of the lowest frequency of
vibration of an acoustical system. To verify the method, he
G
1735
ﬁrst applies it to the problem of determining the lowest
frequency of vibration of a stretched string. His argument
and conclusions are presented here in a modernized
notation.
Consider a string length l with a constant mass density
r and tension t. The string is supposed to exist along the
line segment ÷l/2rxrl/2, and its height at position x and
time t will be denoted by y(x,t). Now, a natural mode of the
string occurs when y(x,t) is nonzero in the absence of an
external driving force. Since the natural modes of lossless
linear systems are always sinusoidal, the displacement of
the string may be taken to be
y(x; t) =v(x) cos ot (1)
The total kinetic energy of the string can be expressed as
T(t) =
1
2
_
l=2
÷(l=2)
r
@y
@t
_ _
2
dx =o
2
sin
2
ot[
~
TT(v)] (2)
where
~
TT(f ) =
1
2
_
l=2
÷(l=2)
r[v(x)]
2
dx (3)
Similarly, the total potential energy of the string may be
written as
V(t) =
1
2
_
l=2
÷(l=2)
t
@y
@x
_ _
2
dx = cos
2
ot[
~
VV(v)] (4)
where
~
VV(f ) =
1
2
_
l=2
÷(l=2)
t[v
/
(x)]
2
dx (5)
and the prime denotes differentiation.
The requirement that the energy of the system remain
constant implies that
d
dt
[T(t) ÷V(t)] =0 (6)
So long as oa0 (which indicates a trivial solution any
way), the value of o corresponding to a given v(x) that
solves the appropriate wave equation is given by
o
2
=
~
VV(v)
~
TT(v)
(7)
In the paragraph preceding his derivation of this result
(which was slightly different than that presented here),
Rayleigh made a heuristic argument for discrete problems
that demonstrates that the minimum positive value of o
delivered by this expression (over all admissible functions
v) is in fact the minimum frequency of free vibration of the
system. He then suggested that formulas of the form of
Eq. (7) might be useful for computing the minimum free
vibration frequency of systems. Specifically, if a function
v(x) is a good approximation to the true solution, in error
by a linear quantity Dv, then the approximation of o
2
will
be in error only by a quantity of order (Dv)
2
. Moreover, if a
set of functions ~ uu
n
(x) is substituted into Eq. (7), the best
estimate of o
2
is necessarily the smallest.
For the problem given above, the true solution is found
when
v(x) =u(x) = cos
px
l
_ _
(8)
for which
o
2
=
p
2
t
rl
2
(9)
(We will always use the variable v to indicate arbitrary
functions, and the variable u to indicate true solutions.
Trial solutions are indicated by ~ uu
n
, where the subscript
indicates some parameter deﬁning the trial.) Rayleigh
suggested testing the method with the set of functions
~ uu
n
(x) =1 ÷
2[x[
l
_ _
n
(10)
for nZ1, for which
~
TT( ~ uu
n
) =
rln
2
(n÷1)(2n÷1)
(11)
and
~
VV( ~ uu
n
) =
2tn
2
(2n ÷1)l
(12)
These expressions lead to an approximation of the form
o
2
n
=
2(n÷1)(2n÷1)
2n ÷1
t
rl
2
=l
2
n
t
rl
2
(13)
From this expression, l
1
=
ﬃﬃﬃﬃﬃﬃ
12
_
, off from the true value of
p by a little more than 10%. Similarly, l
2
=
ﬃﬃﬃﬃﬃﬃ
10
_
, a value
in error by 0.65%. The best possible value of n is the one
that minimizes Eq. (13); this value is n
opt
=
1
2
(1 ÷
ﬃﬃﬃ
6
_
)  1:72474. Using this value in Eq. (13)
gives rise to the value l
n
opt
=3:1463 and a scant error of
0.149%.
Throughout the rest of The Theory of Sound, Rayleigh
continues to apply this method to problems that cannot be
solved analytically, especially problems involving inhomo
geneous material parameters. While Rayleigh’s method as
presented above has been superceded by the Ritz (i.e.,
Rayleigh–Ritz) method described below, Rayleigh quo
tients similar to Eq. (7) are still used today in numerical
linear algebra in the eigenvalue computation technique
known as Rayleigh quotient iteration [14].
Rayleigh’s method was generalized by the Swiss phys
icist Walther Ritz shortly before he died of tuberculosis in
1909 [15]. Specifically, Ritz proposed a method for the
solution of physical problems expressed variationally;
1736 GALERKIN METHOD (RAYLEIGH–RITZ METHOD)
that is, as the minimization of some functional. Suppose
that a given physical problem involves the minimization of
a functional F(v), where v is an arbitrary admissible func
tion. (Both eigenvalue problems and boundary value prob
lems can be expressed in this way.) Ritz suggested that the
solution u be approximated as a series of the form
u  ~ uu
N
=
N
i =1
x
i
u
i
(14)
where the x
i
are unknown coefﬁcients and the u
i
are what
Ritz called ‘‘coordinate functions’’ [16,17]. (In modern par
lance, these would be deemed basis functions.) These basis
functions are known functions that can well approximate
the unknown function u. (Notationally, they are differen
tiated from the true solution by the existence of the sub
script.) Substituting Eq. (14) into the functional yields a
function of the x
i
:
G(x
1
; . . . ; x
N
) =F
N
i =1
x
i
u
i
_ _
(15)
The difﬁcult problem of minimizing the functional is now
reduced to the simpler problem of minimizing G. Thus, the
original problem can be approximately be solved by en
forcing the equations
@G
@x
i
=0 (16)
for i =1; . . . ; N. Ritz also showed that under certain con
ditions, if the basis function set u
i
is complete as NN,
then the solutions obtained will converge to the true so
lution [18]. (In addition, he asserted that it not possible, in
general, to say the same about the convergence of the
derivatives.) Ritz applied his method to ﬁnd the deﬂection
of a ﬁxed plate under a normal load, to solve Dirichlet’s
problem for the Laplace equation, to the computation of
the natural oscillation frequencies of a string, and to many
other more complex problems [16,17].
After Ritz, the story continues in Russia in 1913 where
a naval architect named Ivan Grigorievich Bubnov was
reviewing a paper on the stability of plates and beams
written by Stephen Prokofyevich Timoshenko. Tim
oshenko was applying the Ritz method to the problem,
and Bubnov noted that the same results could be achieved
without resorting to variational expressions [18]. Speciﬁ
cally, Bubnov noted that the same equations generated by
the Ritz method could be more easily derived by simply
substituting the approximation of Eq. (14) into the origi
nal differential equation, and then multiplying the result
by each basis function in turn and integrating over the
domain. Bubnov apparently believed, however, that the
basis functions needed to be orthogonal for the procedure
to be applicable. Nonetheless, because of this contribution,
the Galerkin method described below is often called the
Bubnov–Galerkin method, especially in the Russian liter
ature.
In 1915, Boris Grigorievich Galerkin published a paper
entitled ‘‘Rods and plates. Series occurring in various
questions concerning the elastic equilibrium of rods and
plates,’’ in which he introduced the method that now bears
his name [19]. Formally identical to the method suggested
by Bubnov, Galerkin’s method drops the condition that the
basis functions be orthogonal. Furthermore, Galerkin
makes no mention of the Rayleigh–Ritz method, quite pos
sibly cementing his name in history.
Both the Rayleigh–Ritz method and the Galerkin meth
od are used today in the numerical solution of partial
differential equations, and the Galerkin method is used in
the solution of integral equations as well. Specifically, the
Ritz method is used as the basis of the variational FEM
with very little change. Many authors trace the develop
ment of the FEM to a 1943 paper by Richard Courant [15],
who suggests that the Ritz method be used together with
piecewise linear basis functions situated on a triangular
mesh. Modern variationally based FEM algorithms differ
little from this suggestion, and indeed the variational
FEM can be accurately characterized as the Ritz method
applied to ﬁnite basis functions.
Modern MoM and Galerkinbased FEM algorithms also
differ little from Galerkin’s original outline. Some ap
proaches derive the set of equations to be solved by mul
tiplying the equation with a set of functions other than the
original basis functions; this weightedresidual method is
a straightforward generalization of the original Galerkin
idea. (The functions iteratively multiplied by the equation
before integration are called weighting or testing func
tions. Indeed, many MoM papers refer to the use of the
basis functions for testing as Galerkin testing. This serves
to differentiate Galerkin’s specific choice from the more
general approach discussed in Section 4.)
Since the history of basis function methods begins with
Rayleigh and Ritz, our discussion of the methods begins
with the method that today bears their names.
3. THE RAYLEIGH–RITZ METHOD
In this section, the Rayleigh–Ritz method is described in
detail. Because variational approaches are not familiar to
many readers, Section 3.1 describes the calculus of vari
ations. Section 3.2 then discusses variational principles
important in mathematical physics. The Rayleigh–Ritz
method is ﬁnally described and applied to boundary val
ue problems in Section 3.3. Finally, Section 3.4 describes
the application of the Rayleigh–Ritz method to eigenvalue
problems.
3.1. The Calculus of Variations
As mentioned in Section 2, a functional is a mapping from
some space of functions to some ﬁeld (usually complex or
real numbers). In the context of the Rayleigh–Ritz meth
od, the stationary point of such a functional corresponds to
the solution of some differential equation under study.
Thus, before proceeding to describe the Rayleigh–Ritz
method in more detail, a bit of the variational calculus
will be reviewed here. For simplicity, the discussion in this
section is restricted to one dimension.
GALERKIN METHOD (RAYLEIGH–RITZ METHOD) 1737
Consider the functional described by the integral
I(v) =
_
b
a
F(x; v; v
/
)dx (17)
where v is a function of x and v
/
denotes the derivative of v.
The central problem in the calculus of variations is the
determination of the function v that gives a minimum or
maximum value to this integral, subject to the restrictions
v(a) =v
a
v(b) =v
b
(18)
Suppose that the minimum of Eq. (17) is achieved for some
function v(x) =u(x). We wish to determine the differential
equation satisﬁed by u. Any other function that satisﬁes
the boundary conditions [Eq. (18)] can be written as
v(x) =u(x) ÷eh(x) (19)
if h(a) =h(b) =0. Moreover, in most variational approach
es, the function h is assumed independent of e. (This re
striction is called the case of weak variations in the
literature [20].) If the function F possesses continuous
partial derivatives, an application of Taylor’s theorem
gives
F(x; u÷eh; u
/
÷eh
/
)
=F(x; u; u
/
) ÷e h
@F(x; u; u
/
)
@u
÷h
/
@F(x; u; u
/
)
@u
/
_ _
÷O(e
2
)
(20)
The ﬁrst variation of I is denoted by dI, and can be deﬁned
by the equation
dI = lim
e÷0
I(u÷eh) ÷I(u)
e
(21)
From Eqs. (17), (20), and (21), the ﬁrst variation of I is
given by
dI =
_
b
a
h
@F(x; u; u
/
)
@u
÷h
/
@F(x; u; u
/
)
@u
/
_ _
dx (22)
This ‘‘ﬁrst variation’’ plays roughly the same role in the
calculus of variations that the derivative plays in the
standard calculus.
The form of Eq. (22) can be further simpliﬁed by ap
plying integration by parts to the second term, and recall
ing that h(a) =h(b) =0. This procedure gives a ﬁnal form
for dI:
dI =
_
b
a
h
@F(x; u; u
/
)
@u
÷
d
dx
@F(x; u; u
/
)
@u
/
_ _ _ _
dx (23)
Since the value of I depends linearly on e, a minimum or
maximum of I can be achieved only if dI =0 independent of
the choice of h. This observation leads directly to a differ
ential equation ﬁrst derived by Euler in 1744 [20], now
known as the Euler differential equation:
@F(x; u; u
/
)
@u
÷
d
dx
@F(x; u; u
/
)
@u
/
_ _
=0 (24)
Much of physics can be formulated variationally, that is,
as the minimization of an appropriate functional. The his
tory of modern variational approaches to physics begins
roughly with the work of Pierre de Fermat, and major
contributions were made by a host of eminent physicists
and mathematicians including Isaac Newton, Gottfried
Wilhelm von Leibnitz, Johann and Jacob Bernoulli, Jean
Le Rond d’Alembert, JosephLouis Lagrange, and William
Rowan Hamilton [21].
The particular genius of Rayleigh and Ritz was not in
this use of the calculus of variations to derive differential
equations or formulate physical problems, but to convert it
into a practical method for solving differential equations.
Specifically, if the solution to a differential equation can be
shown to minimize some functional, the Rayleigh–Ritz
method approximates the solution through an appeal to
the variational principle that describes it. (More generally,
the solution of the differential equation may make the
functional stationary, i.e., make its ﬁrst variation vanish.)
Since the Rayleigh–Ritz approach hinges on the creation
of a variational principle, before discussing the method
itself, we describe a variational principle useful in the so
lution of electromagnetic problems.
3.2. Variational Principles for the Scalar Helmholtz Equation
Consider the scalar Helmholtz equation
'u=f (25)
where ' is the linear Helmholtz operator
'v =V
.
(pVv) ÷k
2
qv (26)
u is an unknown function deﬁned on some closed domain
O, k
2
is a constant, and p, q, and f are known real functions
over the domain O. (For now, we assume that O is three
dimensional, but the derivation presented here can easily
be specialized to fewer dimensions.) To completely char
acterize u, boundary conditions must also be speciﬁed on
S, the boundary of O. To this end, assume that S is split
into two nonoverlapping parts S
E
and S
N
. On S
E
, the value
of the function is speciﬁed; i.e.,
u=U (27)
on S
E
, where U is a given function. On S
N
, the boundary
condition is assumed to be
@u
@n
÷au=U
/
(28)
where a and U are given functions, and qu/qn denotes the
outward normal derivative. (Of course, any of the given
data, a, U, or U
/
, can be taken to be zero.)
1738 GALERKIN METHOD (RAYLEIGH–RITZ METHOD)
We now show that the solution to this boundary value
problem is the function that satisﬁes Eq. (27) and makes
the functional
I(v) =
1
2
___
O
[pVv
.
Vv ÷k
2
qv
2
÷2fv]dO
÷
1
2
__
S
N
pv[av ÷2U
/
]dS
(29)
stationary. Because the boundary condition of Eq. (28) is
contained in the variational principle, it is said to be a
natural boundary condition. Equation (27), on the other
hand, is satisﬁed only by an explicit restriction of the
domain of the functional, and is thus termed an essential
boundary condition.
Using Eqs. (19) and (21), the ﬁrst variation of I is given
by
dI =
__
O
[pVu
.
Vh ÷k
2
quh÷fh]dO
÷
__
S
N
hp[au ÷U
/
]dS
(30)
Now, simple differential identities allow the ﬁrst term in
dI to be simpliﬁed, since
pVu
.
Vh=V
.
[hpVu] ÷hV
.
[pVu] (31)
Substituting this expression into Eq. (30) and applying the
divergence theorem allows it to be written as
dI = ÷
___
O
h[V
.
(pVu) ÷k
2
qu ÷f ]dO
÷
__
S
N
h p
@u
@n
÷au ÷U
/
_ _
dS
(32)
Because the function h is arbitrary, enforcing dI =0
implies that each integral in Eq. (32) vanishes, which in
turn implies Eqs. (25) and (28) as was asserted. Note that
the application of the essential boundary condition is
indeed essential; Eq. (32) implies that if nothing is done
on S
E
, the variational principle will enforce the condition
@u
@n
=0 (33)
Similar variational principles may be derived for the vec
tor Helmholtz equation, and may be found in Refs. 1 and 3.
3.3. Examples of the Rayleigh–Ritz Method for Boundary
Value Problems
Now that variational principles have been described, the
description of the Rayleigh–Ritz method itself is relatively
simple. Suppose that a variational expression I(v) for some
differential equation is available. The unknown u is
approximated by a series ~ uu of the form
u  ~ uu
N
=u
0
÷
N
i =1
x
i
u
i
(34)
where the u
i
, i =0; . . . ; N, are known functions and the x
i
,
i =1; . . . ; N, are unknown coefﬁcients. If there are essen
tial boundary conditions of the form of Eq. (27), the func
tion u
0
is assumed to satisfy them; if there are no such
conditions, u
0
can be dropped. The remaining functions u
i
,
i =1; . . . ; N, satisfy the homogeneous form of Eq. (27), thus
ensuring that the approximation of Eq. (34) satisﬁes the
essential boundary conditions. As described in Section 2,
substituting Eq. (34) into the functional gives rise to a
function of the variables x
i
. Stationary points of this func
tional satisfy the N equations
@I( ~ uu
N
)
@x
i
=0 (35)
which may then be solved to yield the x
i
.
As a ﬁrst simple example of the Rayleigh–Ritz method,
consider ﬁnding the solution of the differential equation
@
2
u
@t
2
=36t
2
÷12t ÷4 (36)
subject to the boundary conditions
u(1) =u(÷1) =0 (37)
The true solution to this problem, as is easily veriﬁed, is
u(t) =3t
4
÷2t
3
÷2t
2
÷2t ÷1 (38)
According to the recipe above, the functional for this prob
lem is given by
I(v) =
_
1
÷1
1
2
dv
dt
_ _
2
÷(36t
2
÷12t ÷4)v
_ _
dx (39)
Now, consider trial solutions of the form
~ uu
N
=
N÷1
i =0
x
i ÷1
t
i
(t
2
÷1) (40)
Note that each term in this series satisﬁes the homoge
neous essential boundary conditions of Eq. (37). We exa
mine the solutions given by the approximations ~ uu
1
; ~ uu
2
,
and ~ uu
3
.
For ~ uu
1
, the functional becomes
I( ~ uu
1
) =
4
3
x
2
1
÷
64
15
x
1
(41)
Enforcing the condition of Eq. (35) gives the equation
8
3
x
1
=
64
15
(42)
GALERKIN METHOD (RAYLEIGH–RITZ METHOD) 1739
Thus, x
1
=
8
5
, and the approximate solution is given by
~ uu
1
=
8
5
t
2
÷
8
5
(43)
Using trial function ~ uu
2
gives rise to the function
I( ~ uu
2
) =
4
3
x
2
1
÷
4
5
x
2
2
÷
64
15
x
1
÷
16
5
x
2
(44)
The gradient of this function vanishes when
8
3
0
0
8
5
_
¸
¸
_
_
¸
¸
_
x
1
x
2
_ _
=
64
15
16
5
_
¸
¸
¸
_
_
¸
¸
¸
_
(45)
giving rise to x
1
=
8
5
, x
2
=2, and the approximate solution
~ uu
2
=2t
3
÷
8
5
t
2
÷2t ÷
8
5
(46)
Finally, using trial function ~ uu
3
, the functional becomes the
function
I( ~ uu
3
) =
4
3
x
2
1
÷
4
5
x
2
2
÷
4
105
x
2
3
÷
8
15
x
1
x
3
÷
64
15
x
1
÷
16
5
x
2
÷
64
21
x
2
(47)
The stationary point of this function occurs at the solution
of the equations
8
3
0
8
15
0
8
5
0
8
15
0
8
105
_
¸
¸
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
¸
_
x
1
x
2
x
3
_
¸
¸
_
_
¸
¸
_
=
64
15
16
5
64
21
_
¸
¸
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
¸
_
(48)
which is at the point x
1
=1, x
2
=2, x
3
=3. In particular,
~ uu
3
=u, the exact solution. The exact solution and the two
other trial solutions are plotted in Fig. 1.
A few comments about these results are in order. The
error in each result can be computed in the relative L
2
norm; that is, the error in an approximate solution ~ uu rel
ative to the true solution can be written as
Error =
_
1
÷1
(u ÷ ~ uu)
2
dt
_
1
÷1
u
2
dt
(49)
Using this error criterion, the error in ~ uu
1
is about 29.3%,
and that in ~ uu
2
is 9.6%. Of course, ~ uu
3
is exact.
The large error in the ﬁrst two approximations can be
traced to the fact that they simply cannot curve enough to
capture the true behavior of the solution. On the other
hand, the third approximation is exact because the true
solution lies in the space spanned by the basis functions
used to construct it; that is, since the exact solution can be
represented by ~ uu
3
with a proper choice of coefﬁcients, the
Rayleigh–Ritz method returns the exact solution. This
highlights (in a fairly trivial way) the importance of basis
function selection in the Rayleigh–Ritz method; a judi
cious choice of basis leads to good results fast, while a poor
choice of basis may result in a poor approximation. None
theless, if a given set of basis functions is complete (i.e.,
can represent any function as the number of bases be
comes large), the result will eventually converge.
The inclusion of inhomogeneous essential conditions
creates no major problems. Consider, for instance, the
problem of solving Eq. (36), subject to the boundary con
ditions
u(1) = ÷u(÷1) =1 (50)
The inclusion of these boundary conditions necessitates
adding a nonzero u
0
function to the trial function of Eq.
(40) that satisﬁes the boundary conditions. The simplest
way to satisfy this criterion is to make the new trial
function
~ uu
N
=t ÷
N÷1
i =0
x
i ÷1
t
i
(t
2
÷1) (51)
and repeat the process. Substituting this function into the
functional does not alter it in any way except by adding a
meaningless value of 2 to the function I( ~ uu
N
). (The added
value is meaningless in that it is annihilated by differen
tiation in the minimization of the function.) The reason for
the simple result is clear; the exact solution to the new
boundary value problem is just
u(t) =3t
4
÷2t
3
÷2t
2
÷t ÷1 (52)
−1 −0.5 0 0.5 1
0.5
0
−0.5
−1
−1.5
−2
−2.5
A
p
p
r
o
x
i
m
a
t
i
o
n
t
o
u
(
t
)
t
Figure 1. Approximate solutions of the boundary value problem
of Eqs. (36) and (37) using the Rayleigh–Ritz method. The dotted
line is ~ uu
1
, the dashed line is ~ uu
2
, and the solid line is ~ uu
3
=u.
1740 GALERKIN METHOD (RAYLEIGH–RITZ METHOD)
which is just the old solution, plus the new term t. Choos
ing a different function to satisfy Eq. (50) would lead to
different results in general, but exact results will always
be obtained once the true solution can be expressed ex
actly in terms of the basis functions. (In particular, if u
0
is
chosen to be t
3
instead of t, the solutions ~ uu
1
and ~ uu
2
will
change, but ~ uu
3
will still be exact. If instead u
0
=t
5
, the
exact solution will not be obtained until N=4.) In short,
inhomogeneous essential boundary conditions do not alter
the procedure significantly.
We now examine Eq. (36) under natural boundary con
ditions. To begin, note that in general Eq. (36) is incom
patible with the simplest natural boundary condition, the
homogeneous Neumann condition u
/
(1) =u
/
( ÷1) =0. This
is because integrating Eq. (36) between ÷1 and 1 gives
u
/
(1) ÷u
/
(÷1) =32 (53)
where the homogeneous Neumann boundary condition re
quires this difference to vanish. The Rayleigh–Ritz method
cannot save us from an improperly formulated problem!
Thus, consider Eq. (36) subject to the boundary condi
tions
÷u
/
(÷1) ÷u(÷1) =1
u
/
(1) ÷4u(1) =2
(54)
The exact solution to this problem is given by
u(t) =3t
4
÷2t
3
÷2t
2
(55)
The functional for this problem can be constructed with
reference to Eqs. (25)–(29), but only with some care. For
instance, the surface integral in Eq. (29) must be inter
preted as merely the value of the integrand at the end
points. Similarly, the normal derivative in Eq. (28) should
be interpreted as the usual derivative of the function at
the right endpoint of the interval, and minus the usual
derivative at the left endpoint of the interval. [The ﬁrst
boundary condition in Eq. (54) is written in a form that
makes this explicit.] With these caveats in mind, the func
tional to be minimized for this problem is
I(v) =
_
1
÷1
1
2
dv
dt
_ _
2
÷(36t
2
÷12t ÷4)v
_ _
dx
÷
1
2
v
2
(÷1) ÷2v
2
(1) ÷v(÷1) ÷2v(1)
(56)
For this case, the trial function series will be constructed
as
~ uu
N
=
N÷1
i =0
x
i ÷1
t
i
(57)
Substituting this trial function into the functional yields a
function of the form
I( ~ uu
N
) =
1
2
x
T
Ax ÷b
T
x (58)
where b and x are column vectors of N elements and A is
an NN matrix. Specifically, the ith element of x is x
i
,
and the ith element of b is given by the formula
b [ ]
i
=
i ÷23
i ÷1
i even
3i
2
÷58i ÷16
i
2
÷2i
i odd
_
¸
¸
_
¸
¸
_
(59)
and the elements of A are given by
A [ ]
ij
=
2ij ÷5i ÷5j ÷4
i ÷j ÷1
i ÷j even
÷5 i ÷j odd
_
¸
_
¸
_
(60)
The stationary point of this function occurs when Ax=b.
Solving the set of equations for N=2; . . . ; 5 gives the
solutions
~ uu
2
=1 ÷2t
~ uu
3
= ÷
3
5
÷2t ÷
8
5
t
2
~ uu
4
= ÷
3
5
÷
8
5
t
2
÷2t
3
~ uu
5
=u= ÷2t
2
÷2t
3
÷3t
4
(61)
These solutions are plotted in Fig. 2. As expected, ~ uu
5
is
exact, but the next best estimate, ~ uu
4
, is about 22% in error.
This example further clariﬁes issues involved both
the Rayleigh–Ritz and Galerkin methods. First, problems
3
2
1
0
−1
A
p
p
r
o
x
i
m
a
t
i
o
n
t
o
u
(
t
)
−1 −0.5 0 0.5 1
t
Figure 2. Approximate solutions of the boundary value problem
of Eqs. (36) and (54) using the Rayleigh–Ritz method. The
dashed–dotted line is ~ uu
2
, the dashed line is ~ uu
3
, the dotted line is
~ uu
4
, and the solid line is ~ uu
5
=u.
GALERKIN METHOD (RAYLEIGH–RITZ METHOD) 1741
involving natural boundary conditions generally involve
more unknowns than do those involving essential bound
ary conditions. In the cases presented here, it is clear
that the exact solution to the differential equation is a
fourthorder polynomial. In the case of natural boundary
conditions, this implies that the solution has 5 degrees of
freedom. (Here we chose the coefﬁcient of each monomial
term, but that specific choice is not necessary.) On the
other hand, essential boundary conditions limit the choice
of acceptable functions by forcing the functions with un
known coefﬁcients to have a value of zero at both ends of
the interval. In effect, for a onedimensional problem, this
reduces the number of degrees of freedom by 2. (In this
case, the 2 degrees of freedom removed were the purely
constant and linear terms, since neither of these terms can
satisfy this condition.) In greater numbers of dimensions,
the savings can be even more profound.
The second observation that can be made here concerns
the choice of basis functions. The monomial basis func
tions chosen here are a particularly bad choice, in that as
the order of the polynomials increases, the functions begin
to resemble one another. Because of this, the condition
number of the matrix A grows rapidly with increasing N.
Indeed, the only reason that this is not obvious in the ex
ample presented is that the exact solution was obtained
before the condition number increased too much. If
the righthand side of the differential equation were a
higherorder polynomial, or some other arbitrary function,
accurate results might require the inclusion of more mo
nomials. Given the chosen basis, however, the inclusion of
more monomial terms would vastly increase the condition
number of the matrix, and thus potentially decrease
the accuracy of the solutions. Figure 3, which plots the
condition number of the matrix A versus the order of
approximation N, shows just how serious the problem can
become. With only 20 monomials, the condition number is
greater than 10
12
, so the solution x loses 12 digits of ac
curacy relative to the accuracy with which b is computed.
3.4. The Rayleigh–Ritz Method for Eigenvalue Problems
As a ﬁnal example of the Rayleigh–Ritz method, the fun
damentally different eigenvalue problem is studied. We
will see in this section, however, that the procedure for
eigenvalue problems differs little from that for boundary
value problems.
Consider the differential equation
d
dt
t
du
dt
_ _
÷k
2
tu=0 (62)
subject to the boundary conditions
u
/
(0) =u(1) =0 (63)
This equation will have solutions only for particular val
ues of k called eigenvalues, and associated with each such
k will be an eigenfunction. For Eq. (62), the eigenfunctions
and eigenvalues are well known; in particular, the eigen
functions of the equation are
u
m
(t) =J
0
(k
m
t) (64)
where J
0
denotes the zerothorder Bessel function and k
m
is its mth zero. [The superscript on u in Eq. (64) is used
only to differentiate the mth eigenfunction, which will be
denoted by a superscript, and the mth basis function,
which is denoted by a subscript. The superscript is there
fore not an exponent, but no confusion should arise as we
will never have a need to exponentiate eigenfunctions.]
Using the recipe of Eqs. (25)–(29) leads to the functional
for this problem, which may be written as
I(v) =
_
1
0
t
dv
dt
_ _
2
÷k
2
tv
2
_ _
dt (65)
(The extraneous leading factor of
1
2
has been dropped.) The
trial function employed in this example will be
~ uu
N
=
N
n=1
x
n
cos(k
n
t) (66)
where
k
n
=
2n ÷1
2
p (67)
These functions satisfy the essential boundary condition
at t =1, and even the natural boundary condition at t =0.
Of course, satisfaction of the natural boundary condition is
not necessary, but since the problem statement implies
that the solution will be an even function of t, the chosen
basis functions seem a reasonable choice for the represen
tation of the solution. (They are indeed the basis for even
function Fourier series.)
Substituting the trial solution into the functional yields
the function
I( ~ uu
N
) =x
T
Ax ÷k
2
x
T
Bx (68)
10
12
10
10
10
8
10
6
10
4
100
1
C
o
n
d
i
t
i
o
n
n
u
m
b
e
r
0 5 10 15 20
N
Figure 3. Condition number of the Rayleigh–Ritz matrix A in
terms of approximation order for a monomial basis.
1742 GALERKIN METHOD (RAYLEIGH–RITZ METHOD)
where, as usual, the vector x contains the unknown coef
ﬁcients x
n
, n=1; . . . ; N, and A and B are NN matrices.
Deﬁning the Kro¨necker delta symbol d
mn
=0 for man,
and d
mm
=1, the elements of the matrices A and B are
given by the expressions
[A]
mn
=
k
m
k
n
(k
m
÷k
n
)
2
÷d
mn
k
2
m
4
for m÷n even
÷
k
m
k
n
(k
m
÷k
n
)
2
for m÷n odd
_
¸
¸
¸
_
¸
¸
¸
_
(69)
[B]
mn
=
d
mn
4
÷
1
(k
m
÷k
n
)
2
for m÷n even
÷
1
(k
m
÷k
n
)
2
for m÷n odd
_
¸
¸
¸
_
¸
¸
¸
_
(70)
The extremum of the functional of Eq. (68) occurs when
Ax=k
2
Bx (71)
so the original problem has been reduced to a generalized
algebraic eigenvalue problem. The eigenvalues of this
problem (which we will denote by K
m,N
) are approxima
tions to the true eigenvalues, which in this case are the
roots of the zerothorder Bessel function. The eigenvec
tors, on the other hand, represent the weighting coefﬁ
cients to be substituted into Eq. (66) to approximate the
eigenfunctions. Since eigenvectors and eigenfunctions are
unique only up to a multiplicative constant, all solutions
related here have been scaled so that ~ uu
N
(0) =J
0
(0) =1.
Figure 4 shows the ﬁrst two approximations to the
eigenfunction associated with the smallest eigenvalue re
turned by the Rayleigh–Ritz method. The true eigenvalue
is given (to double precision) by k
1
=2.40482555769577;
the approximations associated with the two approxima
tions shown in the ﬁgure are K
1,1
=2.414 and K
1,2
=2.406.
These numbers represent relative errors of 0.41% and
0.056%, respectively. In contrast to the previous examples,
this example illustrates how quickly the Rayleigh–Ritz
method can converge with a good choice of trial function.
Figure 5 shows the convergence of the eigenvalue approx
imation itself as a function of the order of approximation.
The order of convergence is nearly cubic.
This example also highlights some of the advantages of
the Ritz method relative to Rayleigh’s original method.
While Rayleigh’s method provides only an approximation
to the smallest eigenvalue, the Ritz method actually pro
vides estimates of all the lowest eigenvalues. For instance,
taking the ﬁfth smallest eigenvalue of the algebraic eigen
value problem results in an approximation of the ﬁfth
smallest eigenvalue of the analytic problem. The true
value of the ﬁfth eigenvalue is given by k
5
=
14.93091770848779, and the ﬁrst two estimates provided
by the Rayleigh–Ritz method are K
5,5
=14.9399 and K
5,6
=14.9325, values that contain far less than onetenth of
one percent error.
4. GALERKIN’S METHOD
In this section, Galerkin’s method is described. The meth
od itself is described in Section 4.1, and Section 4.2 dis
cusses its relationship with the Rayleigh–Ritz method.
Finally, Section 4.3 describes the weightedresidual method,
which is a generalization of the Galerkin method.
4.1. Description of the Method
Galerkin’s method is used to solve equations involving
linear operators. The operators may be integral operators,
differential operators, or even integrodifferential operators.
A
p
p
r
o
x
i
m
a
t
i
o
n
t
o
u
(
t
)
1
0.8
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
t
Figure 4. Approximations to J
0
(k
1
x): ~ uu
1
is shown as a dotted line,
~ uu
2
is shown as a dashed line, and J
0
(k
1
x) is shown as a solid line.
1 10 100 1000
N
R
e
l
a
t
i
v
e
e
r
r
o
r
i
n
e
i
g
e
n
v
a
l
u
e
a
p
p
r
o
x
i
m
a
t
i
o
n
10
−2
10
−3
10
−4
10
−5
10
−6
10
−7
10
−8
10
−9
10
−10
Figure 5. Convergence of the smallest eigenvalue as a function
of the order of approximation.
GALERKIN METHOD (RAYLEIGH–RITZ METHOD) 1743
For differential operators, the method is strongly related
to the Rayleigh–Ritz method (as will be demonstrated in
the next section), but it is much simpler to apply. (Indeed,
this is the main reason why the Rayleigh–Ritz method
takes up more space in this article than the Galerkin
method, which is ostensibly the main topic.)
An operator ' is called linear if given two functions v
and w, and two scalars a and b
'(av ÷bw) =a'v ÷b'w (72)
Galerkin’s method works by projecting the original oper
ator equation, deﬁned on a function space, onto a ﬁnite
dimensional vector space. This is accomplished with the
help of an inner product. An inner product deﬁned on a
complex vector space (such as the space of all complex
functions of N variables) is a complex scalar function of
two vectors (functions) that satisﬁes certain criteria. If v,
w, and s denote functions, and a and b are scalars, the
inner product /v,wS must be
1. Linear, that is
s; av ÷bw ¸ ) =a s; v ¸ ) ÷b s; w ¸ ) (73)
2. Hermitian, that is
v; w ¸ ) = w; v ¸ )
+
(74)
where the asterisk denotes complex conjugation.
3. Positivedefinite, that is
v; v ¸ ) _ 0 (75)
with equality only if v =0 [22].
While this is the commonly accepted mathematical
definition of inner products, in practical applications,
products that are not positive definite are often used.
Most often, such products are not Hermitian, but instead
are symmetric:
v; w ¸ ) = w; v ¸ ) (76)
(Generally, such products are also not positive definite.) In
this article, we will not distinguish between the two, and
refer to any linear and symmetric or Hermitian scalar
functional of two functions as an inner product. In partic
ular, over a threedimensional domain O, the most com
mon inner product definitions are
v; w ¸ ) =
___
O
vwdO (77)
for scalar functions v and w, and
v; w ¸ ) =
___
O
v
.
wdO (78)
for vector functions v and w. If these products are deﬁned
over a real vector space, then the mathematical definition
of inner product is met by these functions. On the other
hand, if these products are used over a complex vector
space, they are not Hermitian or positive definite. Finally,
note that these definitions extend trivially to an arbitrary
number of dimensions.
Armed with this definition of inner product, Galerkin’s
method is simple to describe. Consider the equation
'u=f (79)
where ' is a linear operator, u is an unknown function,
and f is a known function. As in the Rayleigh–Ritz method,
u is approximated by a series of basis functions of the form
u  ~ uu
N
=u
0
÷
N
i =1
x
i
u
i
(80)
where the u
i
are called basis functions, and the x
i
are un
known basis function coefﬁcients. As in the case of the
Rayleigh–Ritz method, the extra function u
0
is included to
satisfy inhomogeneous essential boundary conditions; we
will ignore it for the time being and discuss it in the next
section where boundary conditions are described.
Thus, assuming that there are no inhomogeneous es
sential boundary conditions, the approximation of Eq. (80)
is substituted into Eq. (79). In view of the linearity of the
operator, the resulting equation may be written as
N
i =1
x
i
'u
i
=f (81)
Unfortunately, this equation is illposed; it is one equation
in N unknowns. To create a set of N equations that can be
solved for the x
i
, the inner product of Eq. (81) can be taken
with each basis function in turn. This procedure yields
N
i =1
x
i
u
m
; 'u
i
¸ ) = u
m
; 'f
¸ _
(82)
for m=1,y, N; a set of N equations in N unknowns.
The advantage of this method relative to the Rayleigh–
Ritz approach should be obvious—the Galerkin approach
does not require the creation of a functional, which may be
difﬁcult to ﬁnd in some circumstances. The disadvantage
of the Galerkin approach is that boundary condition ap
plication is not as elegant, and the difference between es
sential and natural boundary conditions is not as clear.
Nonetheless, Galerkin’s method and the Rayleigh–Ritz
method are closely related, as the discussion in the next
section will demonstrate. (As was mentioned in the intro
duction, the genesis of Galerkin’s method can be found in
I. G. Bubnov’s comment that the Rayleigh–Ritz equations
could be derived more simply.)
4.2. Boundary Conditions and Relation to Rayleigh–Ritz
To demonstrate the application of boundary conditions in
Galerkin’s method, and to elucidate its relation to the
Rayleigh–Ritz method, we once again consider the model
1744 GALERKIN METHOD (RAYLEIGH–RITZ METHOD)
problem deﬁned by Eqs. (25)–(28). The ﬁrst step in the
Galerkin method is the creation of a trial function and its
substitution into the equation. In this discussion, the trial
function will be that given by Eq. (80), where, as usual, we
assume that u
0
satisﬁes the inhomogeneous essential
boundary condition on S
E
, and that the remaining u
i
sat
isfy the homogeneous version of the same essential bound
ary condition on S
E
. Substituting the basis function
series into the differential equation results in the specific
equivalent of Eq. (81), a single equation in N unknowns
given by
N
i =1
x
i
V
.
pVu
i
( ) ÷k
2
qu
i
_ ¸
=f ÷V
.
pVu
0
( ) ÷k
2
qu
0
(83)
To derive a set of N equations, this equation is multiplied
by each of the functions u
m
, m=1; . . . ; N, in turn, and in
tegrated over the domain [i.e., the inner product of
Eq. (83) is taken with each of the basis functions in
turn.] This procedure yields a set of N equations that
may be written as
N
i =1
x
i
___
O
[u
m
V
.
(pVu
i
) ÷k
2
qu
m
u
i
]dO
=
___
O
[u
m
f =u
m
V
.
(pVu
0
) ÷k
2
qu
m
u
0
]dO
(84)
This equation can be simpliﬁed using Eq. (31) (with h=u
m
and u=u
i
) and the divergence theorem, resulting in
N
i =1
x
i
___
O
[÷pVu
m
.
Vu
i
÷k
2
qu
m
u
i
] dO
_
÷
__
_
S
pu
m
@u
i
@n
dS
_
=
___
O
[u
m
f ÷pVu
m
.
Vu
0
÷k
2
qu
m
u
0
] dO
÷
__
_
S
pu
m
@u
0
@n
dS
(85)
Now, the surface integrals in Eq. (85) vanish over S
E
since
the functions u
m
satisfy homogeneous essential boundary
conditions there by construction. On the other hand, on
S
N
, ~ uu
N
must satisfy the natural boundary condition of
Eq. (28). Substituting this equation into Eq. (85) yields
N
i =1
x
i
___
O
[÷pVu
m
.
Vu
i
÷k
2
qu
m
u
i
]dO
_
÷
__
S
N
pau
m
u
i
dS
_
=
___
O
[u
m
f ÷pVu
m
.
Vu
0
÷k
2
qu
m
u
0
]dO
÷
__
S
N
pu
m
(au
0
÷U
/
)dS
(86)
for m=1; . . . ; N. This is the ﬁnal form of the Galerkin
equations.
Equation (86) may be related to the Rayleigh–Ritz
approach for this problem very simply. Direct substitution
of the trial function into the functional of Eq. (29) gives
I( ~ uu
N
) =
1
2
N
i =1
N
m=1
x
i
x
m
___
O
[pVu
i
.
Vu
m
÷k
2
qu
i
u
m
]dO
_
÷
__
S
N
apu
i
u
m
dS
_
÷
N
m=1
x
m
___
O
[u
m
f ÷pVu
0
.
Vu
m
÷k
2
qu
0
u
m
]dO
_
÷
__
S
N
pu
m
(au
0
÷U
/
)dS
_
(87)
if constant terms involving u
0
alone are ignored. (Such
terms will be annihilated by differentiation in the Ray
leigh–Ritz procedure.) Differentiating this equation with
respect to each unknown and setting the result equal to
zero clearly yields the same set of equations as the Galer
kin method. Thus, all the examples of Section 2 apply
equally well to Galerkin’s method. In particular, all the
same integrals will be computed, and the equations de
rived will be identical (except for possibly a meaningless
shared factor that multiplies all of the equations).
The close relationship between the Rayleigh–Ritz and
Galerkin methods can be clariﬁed by studying the meth
ods from a geometric/linear algebraic point of view. For
this discussion, we will assume that the inner product
employed actually satisﬁes conditions 1–3 listed at the be
ginning of Section 4.1, and that we are dealing with real
functions and operators. In this case, the norm (i.e.,
length) of a function (vector) can be deﬁned by
v  
2
= v; v ¸ ) (88)
Assume further that we are trying to solve Eq. (79), where
' is a positive definite operator, that is
v; 'v ¸ ) _ 0 (89)
for all functions v, and equality holds only if v =0. (Many
physical problems in which the Rayleigh–Ritz method is
applied involve positive definite operators. In particular,
many variational principles are derived from energy con
siderations.)
Given that ' is positive definite, an alternative inner
product (called the 'inner product) can be deﬁned by the
equation
u; v ¸ )
'
= u; 'v ¸ ) (90)
This inner product gives rise to its own norm (the 'norm)
v  
2
'
= v; v ¸ )
'
(91)
GALERKIN METHOD (RAYLEIGH–RITZ METHOD) 1745
much as the original inner product induced a norm. As
suming that ~ uu
N
is a trial function deﬁned in the usual way,
we may seek to approximate the solution of Eq. (79) by
demanding that the coefﬁcients of ~ uu
N
be chosen to mini
mize (half the square of) the ' norm of error given by
I( ~ uu
N
) =
1
2
(u ÷ ~ uu
N
); '(u ÷ ~ uu
N
)
¸ _
=
1
2
~ uu
N
; '~ uu
N
¸ ) ÷ ~ uu
N
; f
¸ _
÷
1
2
u; 'u ¸ )
(92)
This is the Rayleigh–Ritz approach.
Actually performing the minimization gives rise to N
equations for the x
i
, which may be written as
N
i =1
x
i
u
m
; 'u
i
÷f
¸ _
=0 (93)
for m=1; . . . ; N. These are the Galerkin equations. More
over, Eqs. (93) can be written in the form
u
m
; ~ uu
N
÷u ( )
¸ _
'
=0 (94)
which shows that the Galerkin method imposes the con
dition that the error be 'orthogonal to the space spanned
by the basis functions. Thus, in short, the equivalence of
the Rayleigh–Ritz and Galerkin methods in many situa
tions can be traced to the observation that the shortest
distance between a point an a line is given by the orthog
onal projection of the point onto the line. In fact, these
observations are critical not only to the numerical formu
lation of electromagnetic problems but also to the creation
of iterative solvers (like the conjugategradient method)
that can be used to solve the equations once they are
created [23].
4.3. WeightedResidual Methods
The Galerkin method is actually a special case of a more
general set of numerical techniques called weightedresid
ual methods. Weighted residual methods begin the solu
tion of Eq. (79) in the same way as the Galerkin method—
by inserting an appropriate combination of basis functions
into the equation and arriving at Eq. (81). Instead of tak
ing the inner product of Eq. (81) with the basis functions,
however, weightedresidual methods take the inner prod
uct with a set of N testing functions v
m
, m=1; . . . ; N,
resulting in the set of equations
N
i =1
x
i
v
m
; 'u
i
¸ ) = v
m
; f
¸ _
(95)
If v
m
=u
m
, Galerkin’s method is recovered.
Two other choices of testing function are worth men
tioning. First, if the testing functions are taken to be Dirac
delta functions shifted to testing points t
m
, that is
v
m
(t) =d(t ÷t
m
) (96)
the point collocation method is derived. Specifically, sub
stituting Eq. (96) into Eq. (95) yields a set of equations of
the form
N
i =1
x
i
('u
i
)(t
m
) =f (t
m
) (97)
The point collocation method thus enforces the original
operator equation exactly over a ﬁnite set of points.
Second, in the leastsquares method, the testing func
tions are chosen to be
v
m
='u
m
(98)
Using these testing functions leads to a method that min
imizes the (usual) norm of the residual. These alternative
methods can be useful when the Galerkin method is dif
ﬁcult to apply for one reason or another. In particular,
point collocation is often very simple to apply. On the other
hand, the leastsquares method has the advantage that it
always leads to symmetric, positive definite systems of
equations.
5. THE FINITEELEMENT METHOD AND THE METHOD
OF MOMENTS
In this ﬁnal section, we briefly describe FEM and MoM,
the two most widespread applications of Galerkin’s meth
od and the Rayleigh–Ritz method in electromagnetics. No
attempt is made to make this discussion in any way com
plete; other articles in this encyclopedia cover these meth
ods in detail.
The ﬁniteelement method is simply the application of
the Rayleigh–Ritz or Galerkin method to differential equa
tion problems using subsectional basis functions, that is,
basis functions that do not span the entire problem do
main. The FEM has a rather long history, as it was ﬁrst
suggested by Richard Courant during World War II [15].
As an example of the ﬁniteelement method, we will
demonstrate how it can be used to determine the bandgap
of a twodimensional photonic bandgap (PBG) structure.
The unit cell of the PBG lies in the x–y plane and is shown
in Fig. 6. Unshaded areas correspond to free space, and
shaded areas correspond to a material of relative permit
tivity e
r
=13 and relative permeability m
r
=1. The differ
ential equation associated with this problem is Eq. (25),
with u=E
z
, the z component of the electric ﬁeld, p=1/m
r
,
q =e
r
, f =0, and k =o/c, where c is the speed of light in
vacuum and o is the angular frequency of the wave. Be
cause the structure is periodic, the wave must satisfy the
Floquet boundary condition, that is
u(S
R
) =u(S
L
) exp(÷jk
x
d ÷jk
y
d)
u(S
T
) =u(S
B
) exp(÷jk
x
d ÷jk
y
d)
(99)
where d is the length of a side of the unit cell (which will
be taken as unity in this example) and k
x
and k
y
are real
numbers. This is an essential boundary condition.
1746 GALERKIN METHOD (RAYLEIGH–RITZ METHOD)
The goal in the solution of this eigenvalue problem is to
ﬁnd the frequencies of propagation allowed for each com
bination of k
x
and k
y
, and in particular to determine
whether there are any frequency ranges in which propa
gation is disallowed. (The eigenvalues of such a problem
are actually periodic in k
x
and k
y
, so the search can be ac
complished in a ﬁnite amount of time.) Bandgaps are
those ranges of the frequency axis over which propaga
tion is not allowed.
To apply the FEM to this problem, the structure is ﬁrst
meshed with triangles as shown in Fig. 6. A basis function
is associated with each node in the mesh that is one at its
node and falls off linearly to zero at each node connected to
it. In particular, these basis functions are subsectional;
they vanish identically on triangles not connected to the
associated node.
A set of equations can be created by applying either
Galerkin’s method or the Rayleigh–Ritz method using
these subsectional bases. The essential boundary condi
tion can be enforced by ensuring that nodes on the bound
ary are related by Eq. (99); this eliminates the unknowns
associated with two edges of the mesh. Specifically, the
unknowns associated with the white circles in Fig. 6 can
be eliminated by expressing their values in terms of the
values of the unknowns located at the black circle nodes.
In the end, the result of this procedure is a generalized
eigenvalue problem for the eigenvalues o
2
and the asso
ciated eigenvectors that represent the electric ﬁeld of the
mode in question. Figure 7 shows the band structure of
the lowest two bands for this PBG structure computed by
the FEM with 400 unknowns. (The band structure dia
gram shows the allowed values of o for each k
x
,k
y
pair.) A
large bandgap is evident.
As a ﬁnal example of Galerkin’s method, we present an
example of the method of moments. In computational elec
tromagnetics circles, MoM refers to the use of the weight
edresidual method for the solution of integral equations.
Here, we will provide an example of Galerkin’s method
(i.e., the method where the testing functions are chosen to
be the basis functions) specifically. In particular, we dem
onstrate an application of the MoM to scattering from
conducting objects.
Thus, let the surface of a perfect conductor situated in
free space be denoted by S, and imagine that the conductor
is illuminated by an electric ﬁeld E
inc
(r). (Since the for
mulation of this problem is standard in the literature, we
drop the usual notation of using u as the problem solution,
u
i
as the basis functions, and so on, in favor of a more
standard MoM notation.) This incident ﬁeld will induce a
current J(r) on the surface of the scatterer, which radiates
a scattered ﬁeld in such a way to exactly cancel the inci
dent electric ﬁeld tangential to the scatterer. Mathemat
ically, the current must satisfy the electric ﬁeld integral
equation
E
inc
(r) =jkZ
__
S
J(r
/
)
e
÷jk[r÷r
/
[
4p[r ÷r
/
[
dS
/
÷
Z
jk
V
__
S
[V
/
.
J(r
/
)]
e
÷jk[r÷r
/
[
4p[r ÷r
/
[
dS
/
(100)
tangential to the surface of the scatter, where Z is the im
pedance of free space, and k=o/c, where c is the speed of
light in vacuum and o is the angular frequency of the
wave. The integral over dS
/
indicates that r
/
is the inte
gration variable, and the prime over the nabla symbol in
dicated differentiation with respect to r
/
.
Assuming that the set of basis functions is denoted by
B
i
(r), i =1; . . . ; N, and that the inner product used in the
formulation of the problem is that of Eq. (78), an applica
tion of Galerkin’s method gives rise to a matrix equation of
the form
Ax=b (101)
where x is a vector of unknown weighting function coef
ﬁcients. The elements of b are given by
[b]
m
=
__
S
E
inc
(r)
.
B
m
(r)dS (102)
S
L
S
B
S
R
S
T
Figure 6. A meshed PBG structure. Circles indicate node loca
tions.
3
2.5
2
1.5
1
0.5
5
5
0
0 0
10
10
K
y
K
x
Figure 7. Band structure of the PBG structure of Fig. 6 comput
ed by FEM.
GALERKIN METHOD (RAYLEIGH–RITZ METHOD) 1747
and (if the chosen basis is divergenceconforming [24,25])
the elements of the matrix A are given by
[A]
mi
=jkZ
__
S
__
S
B
i
(r
/
)
.
B
m
(r
/
)
e
÷jk r÷r
/
[ [
4p r ÷r
/
[ [
dSdS
/
÷
Z
jk
__
S
__
S
[V
.
B
i
(r
/
)][V
.
B
m
(r)]
e
÷jk r÷r
/
[ [
4p r ÷r
/
[ [
dSdS
/
(103)
Solving Eq. (101) yields an approximation of the current,
which can then be used to compute (for instance) the radar
cross section of the object.
This technique was applied to the computation of the
radar cross section of a 1mdiameter sphere at 100 MHz.
The sphere was meshed with 32 spherical triangle patch
es, and the current was discretized into 160 quadratic ba
sis functions on these patches. The geometry model was
exact; that is, the sphere was not approximated as a union
of ﬂat triangles. The details of the basis functions them
selves may be found in Ref. 24; they are generalizations of
the wellknown linear Rao–Wilton–Glisson basis functions
[25]. The radar cross section computed by the MoM tech
nique is shown in Fig. 8, where it is compared to the result
returned by the exact analytic solution (Mie series). The
two results compare well. While the computation of the
matrix elements in this case is far more complicated in
this case than in all of the previous cases, in principle this
example is a straightforward application of the Galerkin
method. Thus, although the integrals in Eqs. (100) and
(103) require numerical integration of functions that are
often singular, no essential complication is introduced into
Galerkin’s method.
BIBLIOGRAPHY
1. J. Jin, The Finite Element Method in Electromagnetics, J
Wiley, New York, 1993.
2. R. F. Harrington, Field Computation by Moment Methods,
Robert E. Krieger Publishing, Malabar, FL, 1968.
3. P. P. Silvester and R. L. Ferrari, Finite Elements for Electrical
Engineers, Cambridge Univ. Press, New York, 1996.
4. E. K. Miller, L. MedgyesiMitschang, and E. H. Newman, eds.,
Computational Electromagnetics: FrequencyDomain Method
of Moments, IEEE Press, New York, 1992.
5. J. L. Volakis, A. Chatterjee, and L. C. Kempel, Finite Element
Method for Electromagnetics, IEEE Press, New York, 1998.
6. J. J. H. Wang, Generalized Moment Methods in Electromag
netics, Wiley, New York, 1991.
7. G. E. Forsythe and W. R. Wasow, FiniteDifference
Methods for Partial Differential Equations, Wiley, New York,
1960.
8. G. D. Smith, Numerical Solution of Partial Differential Equa
tions: Finite Difference Methods, Oxford Univ. Press, New
York, 1978.
9. A. Taflove, Computational Electrodynamics: The FiniteDif
ference TimeDomain Method, Artech House, Boston, 1996.
10. R. Kress, Linear Integral Equations, Vol. 82, SpringerVerlag,
New York, 1999.
11. L. F. Canino, J. J. Ottusch, M. A. Stalzer, J. L. Visher, and S.
M. Wandzura, Numerical solution of the Helmholtz equation
in 2D and 3D using a highorder Nystro¨m discretization, J.
Comput. Phys. 146(2):627–663 (1998).
12. R. B. Lindsay, Lord Rayleigh—the Man and His Work, Per
gammon Press, New York, 1970.
13. J. W. Strutt, The Theory of Sound, Macmillan, London, 1894.
14. B. N. Parlett, The Symmetric Eigenvalue Problem, Society
for Industrial and Applied Mathematics, Philadelphia,
1998.
15. R. Courant, Variational methods for the solution of problems
of equilibrium and vibrations, Bull. Am. Math. Soc. 49:1–23
(1943).
16. W. Ritz, On a new method for solving variational problems of
mathematical physics, J. Reine Angew. Math. 135:1 (1908).
17. W. Ritz, Theory of the transverse oscillations of a square plate
with free edges, Annal. Phys. 38 (1909).
18. S. G. Mikhlin, Variational Methods in Mathematical Physics,
Macmillan, New York, 1964.
19. B. G. Galerkin, Rods and plates. Series occurring in various
questions concerning the elastic equilibrium of rods and
plates, Vestnik Inzhenerov 19:897–908 (1915).
20. C. Fox, An Introduction to the Calculus of Variations, Dover,
New York, 1987.
21. H. H. Goldstine, A History of the Calculus of Variations from
the 17th through the 19th Century, SpringerVerlag, New
York, 1980.
22. B. D. Reddy, Introductory Functional Analysis, SpringerVer
lag, New York, 1998.
23. Y. Saad, Iterative Methods for Large Linear Systems, Man
chester Univ. Press, Manchester, UK, 1996.
24. R. D. Graglia, D. R. Wilton, and A. F. Peterson, Higher order
interpolatory vector bases for computational electromagnet
ics, IEEE Trans. Anten. Propag. 45(3):329–342 (1997).
25. S. M. Rao, D. R. Wilton, and A. W. Glisson, Electromagnetic
scattering by surfaces of arbitrary shape, IEEE Trans. Anten.
Propag. 30(5):409–418 (1982).
0
0.5
1
1.5
2
2.5
3
45 90 135
R
a
d
a
r
c
r
o
s
s
s
e
c
t
i
o
n
(
d
B
)
(degrees)
Figure 8. Radar cross section of a 1mdiameter sphere as com
puted by the MoM (tickmarked, crossed line) versus Mie series
(solid line) at 100MHz.
1748 GALERKIN METHOD (RAYLEIGH–RITZ METHOD)
GALLIUM ARSENIDE TECHNOLOGY AND
APPLICATIONS
G. F. MANES
University of Florence
Italy
1. INTRODUCTION
Starting from the early 1970s, the increasing request for
solidstate devices operating at microwave frequencies
lead to investigations of semiconductor structures as al
ternatives to silicon. The properties of III–V compound
materials exhibited favorable characteristics for that pur
pose, despite inherent limitations in wafer reproducibility
and yield. Among various solutions, gallium arsenide gen
erated interest and eventually emerged as the favorite
material for highfrequency applications due to the high
carrier mobility compared with silicon devices. To produce
highquality gallium arsenide crystals, scientists needed
to solve a number of problems, including the need to un
derstand the process in which chemical impurities, intro
duced either intentionally or unintentionally, are
distributed during crystal growth. Operating at higher
speed and using less power, electronic devices made from
gallium arsenide crystals have several theoretical advan
tages over silicon.
The advantages in speed and highfrequency perfor
mance of GaAs are slightly offset by the manufacturing
process limitations of GaAs, in terms of both complexity
and reproducibility as opposed to Si. A relevant limitation,
for instance, is the maximum wafer dimension (today only
4–6 in. in diameter, compared to the present 12 in. for sil
icon). This limitation reduces the number of devices per
wafer, and thus the ﬁnal monetary value of each wafer.
The size limitation is due to the limited robustness of the
bulk material. In addition, the device reliability is much
lower than in Si; therefore GaAs is not suitable for very
complex circuits. GaAs devices ﬁnd application at RF and
microwave frequencies for both linear and highspeed log
ic applications. RF devices and ICs are used in consumer
electronics to meet speciﬁcations beyond Si device capa
bilities. GaAs microwave devices ﬁnd extensive use in
consumer applications as well in sophisticated warfare
electronics. GaAs devices are used in the form of discrete
components or microwave monolithic integrated circuits
(MMICs). GaAs also exhibits relevant optical properties
and ﬁnds extensive applications in optoelectronics, includ
ing LEDs, laser diodes, and solar cells; this issue, however,
will not be addressed in this article.
2. MATERIAL PROPERTIES
Like many other III–V compounds, gallium arsenide crys
tallizes in the zincblende structure, whose name derives
from the ZnS mineral. The primitive cell (Z=1) of the
zincblende crystal structure is shown in Fig. 1. It can be
described as the result of two facecentered cubic (fcc) lat
tices of gallium (Ga) and arsenic (As) reciprocally pene
trating and shifted relative to each other by a quarter of
the cube diagonal.
Unlike the tetrahedral atomic conﬁguration of silicon
and germanium, in the zincblende structure the midpoint
between two nearest atoms is no longer a center of inver
sion. Because of this lack of inversion symmetry, in a gal
lium arsenide lattice, the directions [111] and [1 1
"
11] are
not equivalent. The bonds between nearest Ga and As at
oms are in the [111] direction, while the closest equivalent
atoms (GaGa or AsAs) are in the [110] direction. The
[111] plane, which is perpendicular to the [111] direction,
is usually known as the gallium plane, while the (
"
11
"
11
"
11)
plane is known as the arsenic plane. If the GaAs crystal is
cleaved on the (111) plane, the wafer surface is made up of
either gallium atoms with one free bond and three bonds
with the layer below, or arsenic atoms with three free
bonds and only one bond with the layer below. The oppo
site situation is obtained on the (
"
11
"
11
"
11) plane.
Unlike silicon or germanium crystals, which are
cleaved on the (111) plane, GaAs crystals are usually
cleaved along the (0 11) or (0 1
"
11) plane since, because of
the partially heteropolar bonds between gallium and ar
senic, the zincblende structure does not easily break along
the (111) plane. At T=300K the lattice constant of the
GaAs crystal is 5.653 A
˚
and its thermal conductivity is
0.55 Wcm
÷1
1C
÷1
.
The energy band diagram of gallium arsenide at T=
300 K is shown in Fig. 2. Like many other III–V com
pounds, GaAs is a directbandgap material; this means
that the top of the valence band (light and heavy holes)
and the bottom of the main conduction band valley (the &
valley) are located in the same point in the wavevector
space. The conduction band also has two more valleys, one
in the /111S direction and one in the /1 0 0S direction,
respectively known as the L valley and the X valley. At
room temperature the energy gap E
g
between the & val
ley and the top of the valence band is 1.42 eV; in the same
conditions the gaps between the top of the valence band
and the bottoms of the two secondary valleys of the con
duction band are 1.71 eV for the L valley and 1.90eV for
the X valley. The top of the splitoff subband is 0.34 eV be
low the top of the light and heavyhole bands.
All energy values reported are temperaturedependent;
in the range 0–1000 K E
g
can be calculated using the
Figure 1. The zincblende primitive cell.
GALLIUM ARSENIDE TECHNOLOGY AND APPLICATIONS 1749
following relation:
E
g
(T) =1:519 ÷5:40510
÷4
.
T
2
T ÷204
(eV) (1)
The intrinsic carrier concentration is given by
n
i
= N
C
.
N
V
( )
1=2
exp ÷
E
g
2k
b
T
_ _
(cm
÷3
) (2)
Taking into account the nonparabolic behavior of the G
valley and the contributions from the X and L valleys, the
effective densities of states N
C
and N
V
can be calculated
from the following relations:
N
C
=8:63 10
13
.
T
3=2
[1 ÷1:93 10
4
.
T
÷4:19 10
8
.
T
2
÷F
LX
(T)] (cm
÷3
)
F
LX
(T) =21
.
exp ÷
E
GL
2k
b
T
_ _
44
.
exp ÷
E
GX
2k
b
T
_ _
N
V
=1:83 10
15
.
T
3=2
(cm
÷3
)
(3)
where E
GL
and E
GX
are the energy gaps between the main
and the two secondary valleys.
The effective masses of carriers are reported in Table 1
(m
0
=9.11 10
÷31
kg is the electron inertial mass).
The ﬁeld dependence of the electron drift velocity is
depicted in Fig. 3; the negative differential mobility be
havior is due to the increase of the carrier concentration in
the L and X valleys for high electric ﬁelds. The solid curve
was calculated by Pozhela and Reklaitis [Solid State Elec
tron. 23(9):927–933 (1980)], dashed and dotted curves are
measured data at room temperature.
Modern monolithic devices such as HBTs and HEMTs
are based on the properties of widebandgap heterojunc
tions. Figure 4 illustrates the relationship between band
gap and lattice constant for various alloys in the InGaAsP
and AlGaAsSb systems.
As can be seen in Fig. 4, gallium arsenide has nearly
the same lattice constant of aluminum arsenide (AlAs); at
the same time the bandgap difference between the two al
loys is quite high (0.8 eV). This implies that the Al
x
Ga
1 ÷x
As ternary compound has in any case a lattice
Energy
T=300 K
Eg = 142 eV
E1 = 1.71 eV
Ex = 1.90 eV
Eso = 0.34 eV
L Valley
Γ Valley
X Valley
<100>
<111>
Wave vector
Heavy holes
Light Holes
Splitoff band
E1
0
Ex
Eg
Eso
Figure 2. The zincblende primitive cell.
Table 1. Effective Masses of Carriers
Conduction band (G valley) m
G
=0.063m
0
(L valley) m
L
=0.085m
0
(X valley) m
X
=0.085m
0
Valence Band Light holes m
lh
=0.082m
0
Heavy holes m
hh
=0.51m
0
Splitoff m
so
=0.15m
0
2.0
1.5
1.0
0.5
0
0 2 4 6 8 10
Field F (kV/cm)
D
r
i
f
t
V
e
l
o
c
i
t
y
v
d
(
1
0
7
c
m
/
s
)
Figure 3. Electric ﬁeld dependence of the electron drift velocity.
Lattice constant (Å)
B
a
n
d
G
a
p
(
e
V
)
Binary substrates
GaAs InP
W
a
v
e
l
e
n
g
t
h
(
µ
m
)
a
t
t
h
e
b
a
n
d
g
a
p
e
n
e
r
g
y
InP
GaP
AlAs
GaAs
x− 0.2
5.4
0.5
1.0
1.5
2.0
2.5
5.6 5.8 6.0 6.2
2.48
1.55
1.3
1.24
0.827
0.620
0.496
0.4
In
x
,Ga
1−x
As 0.6
0.8
InAs
GaSb
AlSb
Figure 4. Relationship between bandgap and lattice constant for
alloys in the InGaAsP and AlGaAsSb systems. (Figure taken from
B. G. Streetman and S. Banerjee, Solid State Electronic Devices,
PrenticeHall, 2000.)
1750 GALLIUM ARSENIDE TECHNOLOGY AND APPLICATIONS
constant very close to that of gallium arsenide with a
bandgap energy that can be modulated by correctly choos
ing the alloy composition x.
The properties of the GaAs/AlAs system can be
applied to build a wide range of nearly latticematched
heterojunctions starting from a binary GaAs substrate.
The bandgap difference can be further increased by grow
ing a series of Al
x
Ga
1÷x
As and In
y
Ga
1 ÷y
As layers on a
gallium arsenide substrate, as is done in a GaAs pseu
domorphic HEMT. In this case, anyway, the lattice mis
match between GaAs and the In
y
Ga
1 ÷y
As compound
quickly rises as the indium alloy composition y is
increased.
3. GaAs MMIC FABRICATION
GaAs MMICs are manufactured starting with a wafer of
semiconductor material. These wafers are sliced from a
large single crystal, known as a boule. Boules are grown
by the wellknown Czochralski crystal growth method,
which is substantially the same as the technique used for
silicon and other semiconductors. Basically, a small seed
crystal is lowered to the surface of the melting material,
and then slowly lifted up. In this manner, it draws with it
layer by layer of molten material, which cools gradually,
resulting in a single crystal with the same crystalline
structure as the seed crystal. Actually, growing compound
semiconductors is slightly complicated because the vapor
pressures of the constituent materials are different. A
suitable encapsulate, together with a high pressure of in
ert gas within the growth chamber, is employed in order to
prevent this problem.
After slicing the boules to thin wafers, their surface is
smoothed by mechanical and chemical treatments, so that
its roughness is reduced to an atomic layer level. The ob
tained product is the basic material for MMIC microfab
rication.
GaAsbased devices need a mesastructure compound
by several layers of different material alloys. Their thick
ness and molar fraction must be known quite accurately.
So, process requirements are intrinsically dissimilar from
what can be obtained by diffusion doping techniques,
which are typically applied in traditional Sibased micro
circuit fabrication. Consequently, epitaxial deposition is
the most popular and perhaps the only method employed
in fabricating GaAs MMICs. In the following sections two
main techniques, molecularbeam epitaxy and metallor
ganic chemical vapor deposition, will be described.
These techniques allow the production of highquality
layers with very abrupt interfaces and good control of
thickness, doping, and composition. In fact, they are also
compatible with accurate growth measurement systems,
such as the reﬂection highenergy electron diffraction
method.
Photolithography and etching are then performed in
order to selectively remove material from the sample, re
sulting in the effective transfer of the design layout to the
chip. This kind of technology is not limited to GaAs pro
cessing, and its description is beyond the purpose of this
section. Multiple metallization layers can be added by
thinﬁlm vacuum evaporation, sputtering, or electroplat
ing; additional sacriﬁcial layers can be grown to allow air
bridge construction. Again, masks are used to control the
geometry of the metal ﬁlm on the device.
4. MOLECULARBEAM EPITAXY
Molecularbeam epitaxy (MBE) was developed in the early
1970s as a means of growing highpurity epitaxial layers
of compound semiconductors.
The working principle of MBE is shown in Fig. 5. A
sample wafer is heated in a chamber, and eventually it is
kept in rotation; then, solid sources materials are placed
in evaporation cells to provide an angular distribution of
atoms or molecules in a beam. Every source is connected to
the chamber via a shutter, so that the ﬂux of material can
be externally controlled with accuracy. Atoms and mole
cules combine onto the wafer surface, and epitaxial layers
of the desired compound material grow on the sample.
However, some practical issues drastically complicate
the construction of a real MBE system. First, the chamber
is kept in ultrahighvacuum (UHV) environment. Actu
ally, the molecularbeam condition that the meanfree
path of the particles should exceed the geometric size of
the chamber is easily fulﬁlled if the total pressure does not
exceed 10
÷5
torr. Nevertheless, the condition for growing
a sufﬁciently clean epitaxial layer typically requires the
partial pressure of the background residual vapor to be
less than 10
÷11
torr. Thus, the total pressure is kept as low
as reasonably possible—around 10
÷7
torr—and liquid N
2
cryopanels, together with cryopumps, act as a screening
around the substrate, minimizing spurious ﬂuxes of atoms
and molecules from the chamber walls while reaching the
desired level of residual vapor and undesired gas partial
pressure. Moreover, the materials of the chamber are cho
sen so that they minimize the rate of gas evolution (e.g.,
pyrolytic boron nitride, molybdenum, tantalum) and a
special maintenance is applied: a 24h bakeout of the
whole chamber at 2001C after each vent, and heating of
the windows at 3001C for a couple of hours to prevent
them from being coated.
RHEED
gun
Sample block/
substrate heater
Ionization/
BEP gauge
To buffer
chamber
Cryopanels
CAR
assembly
Fluorescent
screen
Shutters
Effusion cells
Figure 5. MBE process setup.
GALLIUM ARSENIDE TECHNOLOGY AND APPLICATIONS 1751
Then, a proper temperature has to be chosen for the
substrate. Consider one of the key points of MBE, namely,
that when particles from the molecular beam reach the
sample surface, they are not immediately included in the
crystalline structure; instead, they should be able to mi
grate on the wafer for a relatively long path. In this man
ner, atoms are allowed to ‘‘search’’ for a minimum energy
site, which is always at the edge that rises on a layer when
a isle of atoms is forming a new layer because there are
forces attracting the atom from both the bottom and the
lateral side. This is important since the crystal does not
grow randomly, and then grows roughly, but some initial
atoms form isles that gradually expand until they coalesce
to cover the whole surface. Consequently, the growth pro
ceeds spontaneously layer by layer, resulting in an ex
tremely homogeneous and controllable epitaxial
deposition. Now, it is straightforward that the higher the
wafer temperature, the longer the migration path of the
atoms on its surface. However, looking at the phase dia
gram of a binary compound, one can conclude that there is
a temperature over which the pressure of the more volatile
component (in the GaAs case it is arsenic) increases more
rapidly, a minimum for a suitable stoichiometry of the sol
id phase no longer exists, and a liquid gallium phase is
created. Hence, a temperature below 6301C is required for
GaAs substrates.
The effusion cells are independently heated at accu
rately controlled temperatures, so that the contained
atomic species are able to escape by thermoionic emission.
Since the typical rate of growth with MBE is as slow as
that around a single monolayer per second, shutters posed
in between the sources and the wafer can easily control
the composition of the global ﬂux of particles reaching the
semiconductor surface, with up to 1% accuracy.
5. METALLORGANIC CHEMICAL VAPOR DEPOSITION
MOCVD (metallorganic chemical vapor deposition) or
MOVPE (metallorganic vaporphase epitaxy) is another
widely used method of creating controllable epitaxial lay
ered structures by atomic deposition over a substrate ma
terial.
This method involves passing metal oxides in an inert
gas across a workpiece to deposit a layer of metal oxide on
the surface. MOCVD is utilized largely by the semicon
ductor industry. As shown in Fig. 6, the substrate wafer is
placed on the graphite susceptor inside a reaction vessel
and heated by an RF induction heater.
The temperature depends on the type of compounds
grown, but it is usually between 500 and 7001C. Growth
occurs in an atmosphere of hydrogen at a pressure be
tween 100 and 700torr. The growth precursors decompose
on contact with the hot substrate to form epitaxial layers.
Group V precursors are AsH
3
(arsine); while group III are
Ga(CH
3
)
3
trimethylgallium (TMG). The MOCVD process
offers good material distribution capabilities and repro
duces ﬁne details with reliable quality. The materials to be
coated must be able to withstand the extreme tempera
tures associated with this process.
6. REFLECTION HIGHENERGY ELECTRON DIFFRACTION
One of the most useful tools for in situ monitoring of the
growth is reﬂection highenergy electron diffraction
(RHEED). It can be used to calibrate growth rates, ob
serve removal of oxides from the surface, calibrate the
substrate temperature, monitor the arrangement of the
surface atoms, determine the proper arsenic overpressure,
give feedback on surface morphology, and provide infor
mation about growth kinetics. The layout of the RHEED
system, shown in Fig. 7, can be summarized as follows.
The RHEED gun emits B10keV electrons, which strike
the surface at a shallow angle (B0.5–21), making it a sen
sitive probe of the semiconductor surface. Electrons reﬂect
from the surface and strike a phosphor screen, forming a
pattern consisting of a specular reﬂection and a diffraction
pattern that is indicative of the surface crystallography. A
camera monitors the screen and can record instantaneous
pictures or measure the intensity of a given pixel as a
function of time.
Starting from a ﬂat substrate, the electrons are not
scattered greatly and are recorded as an intense beam. As
material is deposited on the surface, the atoms create is
lands of epitaxial growth with a concomitant decrease in
the reﬂectivity of the surface of the material as the elec
trons are scattered. As the deposition process continues,
material builds up on the surface and the islands join to
gether and create new ﬂat surfaces, while the original
substrate forms as voids in the newly created material. As
yet more material forms, the voids begin to ﬁll up and the
reﬂectivity increases once again, although in reality, the
reﬂectivity does not reach as high a value, since the de
position process is random and the surface never regains
the completely ﬂat proﬁle of the initial polished substrate.
By monitoring the oscillations in the reﬂectivity one can
estimate the thickness and growth rate of the epitaxial
material.
The appearance of the RHEED diffraction pattern can
be used to provide qualitative feedback on surface mor
phology. If the surface is smooth, then the RHEED dif
fraction patterns appear streaky. If the samples are rough,
then the horizontal streaks are more ‘‘spotty’’ and the dif
fraction pattern is not as clear. An amorphous surface,
Reaction Envelope
RF Induction Heater
Wafer
Graphite
susceptor
TMA
TMG
TMI
DEZ
SiH
4
AsH
3
Bubbler
Mass flow controller
H
2
H
2
H
2
H
2
Valve
Figure 6. MOCVD process setup.
1752 GALLIUM ARSENIDE TECHNOLOGY AND APPLICATIONS
such as an oxide layer, shows a haze instead of a diffrac
tion pattern, and polycrystalline surfaces result in rings
circling the straightthrough beam. Such information, al
though only qualitative, can be a useful check of the sur
face condition.
7. GaAs HIGHSPEED DEVICES
Gallium arsenide integrated circuit technology has been
historically described as the ‘‘technology of the future.’’
Although initially it did not achieve the wide applicability
originally predicted, it is now considered a relevant tech
nology for microwave and millimeterwave applications,
and it is a fully accepted and commercially available
source for highspeed integrated circuits. The main
advantage of GaAs in highfrequency applications is its
high electron mobility value. GaAs devices typically show
a unity current gain frequency f
T
that is 30–70% higher
than that of an equivalent silicon transistor, for a given
gate length or base width. Similar gainful can be achieved
for the unity power gain frequency f
max
. In addition, Al
GaAs, which provides the ability to build latticematched
and widebandgap heterojunctions, allows the device de
signer to have more freedom for performance optimization.
This feature of the AlGaAs/GaAs system is extremely
important for a GaAs heterojunction bipolar transistor
(HBT), where a higher emitter bandgap, with respect to
the base bandgap material, is used to prevent holes from
diffusing from base to emitter. As a result, a higher base
doping can be reached, decreasing the base resistance,
without degrading current gain. Besides, since undoped
GaAs is a semiinsulating material, the parasitic capaci
tance due to wire interconnects is lower than in silicon
technology.
The main GaAs technology drawback is related to the
nonoptimal thermal behavior, which makes it more difﬁ
cult to dissolve the heat generated by dissipated power.
Electrically speaking, the heat generation produces a
drain current degradation for FET devices; this phenom
enon is usually known as ‘‘selfheating’’.
The three dominant GaAs highspeed device types are
metal Schottky FET (MESFET), modulation doped FETor
highelectronmobility transistor (MODFET, HFET,
HEMT), and the bipolar transistor type, known as the
heterojuncton bipolar transistor (HBT).
Besides thermodynamic problems, other drawbacks are
observed, related principally to practical aspects, such as
the low level of achievable integration, backdating, hys
teresis, high 1/f noise, and parameter dispersion for
MESFET and HEMT devices.
8. GaAs MESFET
The principle of GaAs MESFET transistor operation is
similar to that of the simple and wellknown silicon JFET,
in particular concerning the current control mechanism
by means of the voltage applied to the gate. The main dif
ference between a silicon JFET and a GaAs MESFET is
that in the former the drain current saturation is achieved
at channel pinchoff, while in the latter saturation is in
duced by the electron velocity saturation. Even if GaAs
HEMTs exhibit higher performance levels, GaAs ME
SFETs, are still commonly used in wireless applications.
GaAs MESFET technology is currently capable of re
alizing enhancement/depletionmode integrated circuits of
LSI complexity. The cross section of a typical GaAs ME
SFET is shown in Fig. 8. The maximum f
T
of commercially
available MESFET for integrated circuits applications is
around 15 GHz, and the maximum f
T
achieved by the best
GaAs MESFET in a research laboratory environment is
approximately 80 GHz. Commercially available devices
have gate lengths of approximately 0.5mm down to
0.1 mm. The typical threshold voltage spread across a 3
in. wafer is 30–40 mV, and local threshold voltage spread,
[110]
Monolayer growth Electron beam Rheed signal
= 0
= 0.25
= 0.5
= 0.75
= 1.0
[110]
[001]
Figure 7. REED in situ growth monitoring tech
nique.
GALLIUM ARSENIDE TECHNOLOGY AND APPLICATIONS 1753
which is important for comparator applications, is 15–
20 mV. Because of the Schottky barrier gate, the maxi
mum gate source voltage of GaAs MESFET is typically
limited to approximately 0.7V.
Most commercially available GaAs MESFETs suffer
from lowfrequency hysteresis effects, which make the de
sign of highaccuracy comparators more difﬁcult. This
phenomenon is due to a drop in the highfrequency
FET’s draintosource resistance, which is also associated
with a phase shift, when the device is biased in the sat
uration region. This behavior produces a timevarying off
set voltage or hysteresis, in the I–V relationship of the
device, which is added to the DC offset of the comparator,
thus reducing its achievable resolution. The hysteresis
effect can be reduced by improving the circuit’s design
procedures. More recent results obtained by means of im
proved processing techniques have demonstrated than it
can be eliminated.
9. GaAs HEMT
Greater efﬁciency, along with a lower noise ﬁgure and a
higher cutoff frequency, can be obtained by using the high
electronmobility transistor (HEMT; see HEMT cross sec
tion in Fig. 9). The principle of operation is based on the
concept of modulation doping.
In order to obtain a highelectronmobility value, which
means a high cutoff frequency and a lower noise ﬁgure, it
is possible to use either a highelectronmobility semicon
ductor material such as gallium arsenide or an undoped
semiconductor material in which ionized impurity scat
tering is reduced. A more efﬁcient method is to grow the
FETchannel using an undoped highmobility semiconduc
tor material and then increase the concentration of free
carriers (which is extremely low at room temperature in
an undoped material) by transferring them from an adja
cent heavily doped layer, known as the ‘‘supply layer’’. In
this manner it is possible to obtain a high mobility of car
riers, since electrons travel in an undoped material,
and, at the same time, it is also possible to achieve high
carrier concentrations and correspondingly high current
densities.
The situation described above can be achieved by
means of a heterojunction between two semiconductors
with significantly different energy gaps. The widerband
gap semiconductor is used to grow the heavily doped
supply layer, while the other alloy, which is intentionally
not doped, is used for the FET channel. The electrostatic
equilibrium of the heterojunction results in a triangular
well at the interface, which conﬁnes the electrons in a two
dimensional (2D) electron gas known as 2DEG. The band
diagram of this particular type of heterojunction is illus
trated in Fig. 10.
At high frequencies, the HEMT structure, using the
2DEG as the current conducting channel, exhibits perfor
mance levels higher than those achievable with a conven
tional MESFET; therefore it is the most suitable
transistor structure for microwave and millimeterwave
analog circuits as well as in highspeed digital applica
tions. The ﬁrst HEMT prototype was demonstrated in
1981 in the AlGaAs/GaAs material system.
The current control mechanism in the HEMT is ob
tained by tuning the density of electrons in the triangular
well. When the gate voltage is modiﬁed, both the shape of
triangular well and the position of the Fermi level are
modiﬁed; therefore the number of electrons into the 2DEG
is increased or decreased according to the sign of the V
G
variation.
The sheet charge density n
s
can be expressed as a func
tion of the gate voltage V
G
in the following way
n
S
=
e
qd
(V
G
÷V
TH
) (4)
Source
Gate
Drain
Figure 8. Cross section of a MESFET.
Source
Gate
Drain
Figure 9. Cross section of a HEMT.
Gate
Channel
Figure 10. Band diagram of a HEMT.
1754 GALLIUM ARSENIDE TECHNOLOGY AND APPLICATIONS
where e is the GaAs permittivity, d is the distance between
the metal–semiconductor interface (where gate voltage
is applied), and V
TH
is the threshold voltage or pinchoff
voltage.
As can be easily understood, n
s
is related to the drain
current, so that varying the sheet charge density allows
the drain current to be controlled. The relationship
between the drain current I
D
and the gate voltage V
G
can be calculated analytically, resulting in the following
equation:
I
D
=
WC
OX
m
n
L
G
(V
GS
÷V
T
)V
DS
÷
V
2
DS
2
_ _
for V
DS
oV
DS;sat
WC
OX
m
n
L
G
(V
GS
÷V
T
)
2
2
=I
D;sat
for V
DS
_ V
DS;sat
_
¸
¸
¸
¸
_
¸
¸
¸
¸
_
(5)
In order to calculate the cutoff frequency, the equivalent
circuit of the transistor, shown in Fig. 11, is needed.
Some circuital components such as capacitances (both
intrinsic and extrinsic or parasitic), as well as source and
drain input resistances can be calculated. The other ele
ments of the equivalent circuit, such as the transconduc
tance g
m
and the drain conductance g
d
, can be deduced
either from theoretical equations, or from smallsignal
measurements.
The aim of the analysis is to obtain the h
21
hybrid pa
rameter expression, which is frequencydependent; by set
ting h
21
to 1, one can calculate the cutoff frequency f
T
(i.e.,
the frequency corresponding to a unitgain current). Once
the proper equivalent circuit has been deﬁned, the same
procedure can be used for HBT GaAsbased transistors,
because the h
21
parameter is related to the equivalent cir
cuit and not to the physics behavior of the transistor. The
full expression of f
T
is
f
T
=
1
2p
C
GG;t
g
m
÷
C
GG;t
g
m
(R
S
÷R
D
)g
d
÷(R
S
÷R
D
)C
GD;t
_ _ (6)
where C
GG,t
is related to the charge stored under the gate.
A similar approach can be followed to calculate the
maximum oscillation frequency. In this case, however, we
are interested in deriving the unilateral power gain. Once
the unilateral power gain is calculated, we set its value to
1, thus determining the maximum oscillation frequency
f
max
. With reference to the circuit in Fig. 11, the following
expression can be derived for f
max
:
f
max
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
f
T
8pR
G
C
GD;t
1 ÷
2pf
T
C
GD;t
_ _
C
_ _
¸
¸
¸
¸
_
C=(R
S
÷R
D
)
C
2
GG;t
g
2
d
g
2
m
÷(R
S
÷R
D
)
C
GG;t
C
GD;t
g
d
g
m
÷
C
2
GG;t
g
d
g
2
m
(7)
NF
min
=1÷o
C
GS
g
m
R
S
÷R
G
R
IN
_ _
1=2
(8)
9.1. Output Power
The power performance of a FET, as also of HBT, is de
termined by the maximum output power P
out
deﬁned as
P
out
=
1
8
(I
max
÷I
min
)(BV
DS
÷V
knee
) (9)
where
V
knee
=V
DS;sat
(10)
I
max
=I
D;sat
(V
GS
=0; V
DS
=V
DS;sat
)
I
min
=0
(11)
The quantities used in these formulas are depicted in
Fig. 12.
C
GS
R
IN
g
m
g
D
C
DS
R
D
R
S
C
GD R
G
G
S
D
Figure 11. Equivalent circuit of a HEMT.
I
drain
I
max
V
knee BV
DS
Voltage
Figure 12. Static I/V characteristic of a HEMT.
GALLIUM ARSENIDE TECHNOLOGY AND APPLICATIONS 1755
The breakdown voltage is given by a semiempirical
formula:
BV
DS
=
4:4 10
13
aN
d
(12)
where a is the thickness of the channel layer, while N
d
is
the doping of the supply layer.
10. GaAs HETEROJUNCTION BIPOLAR TRANSISTOR (HBT)
Bipolar transistors, in spite of the advent of ﬁeldeffect
devices (which in Si technology dominate the memory
market), continue to play a significant role still at micro
wave frequencies. The levels of current that can be han
dled, in fact, is greater than in ﬁeldeffect devices, leading
to both high speed and high power devices.
In ﬁeldeffect devices, a way to achieve a high cutoff
frequency and a low noise ﬁgure is to reduce the gate
width. This goal, obviously, leads to a more critical litho
graphic process. For bipolar devices this step is replaced
by the precise preparation of the thin base region; appro
priate technologies for uniform growth of wafers have
been deﬁned to keep bipolar technology competitive.
However, the full fabrication process for bipolar tran
sistors is more complex than for ﬁeldeffect devices, and
this factor has been relevant in making ﬁeldeffect devices
more popular. According to the standard transistor theory,
the DC current gain (the ratio between the collector cur
rent and the hole current ﬂowing from the base to the
emitter) is given by
b =
n
e
v
e
p
b
v
p
_ _
exp
DE
G
KT
_ _
(13)
where n
e
and p
b
are respectively the electron and hole
densities in emitter and base, v
e
and v
p
are the electron
and hole effective velocities, and DE
G
is the difference in
bandgap between emitter and base. As can be seen from
this relationship, a higher level of gain current can be
achieved by increasing the bandgap difference between
emitter and base of the transistor. An AlGaAs/GaAs he
terojunction, for instance, allows significant improve
ments in performance; these improvements are
summarized below:
*
The difference in bandgap is 10 kT, and the exponen
tial factor increases the value of b at room tempera
ture up to 1000 (10–100 is sufﬁcient in practice).
*
The doping in the base can be increased significantly;
thus the reverse ﬂow in the holes is stopped by the
heterojunction, and the lower base resistance pro
vides a higherspeed device.
*
The doping in the emitter can be increased without
compromising the effect of bandgap narrowing; in
turn, this leads to a higher currenthandling ability.
*
The notch in the conduction band edge can be elim
inated by grading down the Al composition in the last
few nanometers; this improves the emitter efﬁciency
(see Fig. 13b).
In the transistor design tradeoff, the heterojunction can
be used as a further design tool. A thin base, for example,
increases the DC current gain (since there is a lower prob
ability of electron–hole recombination in the base), but
unfortunately it slows down the device (because of the in
creased lateral base resistance). Away to avoid this effect,
allowed by heterojunctions, is to increase the p doping in
order to offset any rise in lateral base resistance. A further
advantage of heterojunctions is that the bipolar transistor
can operate over a significantly enhanced temperature
range, since the heterojunction prevents runaway ther
malbased currents (higher temperature range) and be
cause all the doping densities are above the levels at which
carriers freeze out at low temperatures (lower tempera
ture range).
If the composition of the base is graded (see Fig. 13b)
from 10%Al content at the emitter–base heterojunction up
to 100% at the base–collector interface, the electric ﬁeld at
the falling conduction band edge can be used to maintain
acceleration of the injected hot electrons. A widergap col
lector can also be used to reduce hole injection from the
base to collector when forwardbiased (i.e., turned off) and
to allow higher voltages to be applied to the collector junc
tion without breakdown or leakage. The main drawback is
related to the complex structures associated with the he
terojunction bipolar transistor, which leads to a low yield.
GaAS HBT
GaAS HBT (A1 graded)
Emitter Base Collector Emitter Base Collector
(a) (b)
Figure 13. (a) HBT band diagram; (b) Algrad
ed HBT band diagram.
1756 GALLIUM ARSENIDE TECHNOLOGY AND APPLICATIONS
In digital circuits, the low values of the base resistance
and the short electron transit time from emitter to collec
tor enable fast device operation (ON/OFF switching speeds
of o10 ps). The larger bandgap in AlGaAs compared with
Si means that larger drive voltages are required for III–V
HBTs. In a number of analog circuit conﬁgurations, such
as ampliﬁers and highspeed digitaltoanalog converters,
some of the HBT features are highly desirable to systems
designers, particularly the uniformity of the turnon volt
age of the emitter–base junction, the absence of trapping
effects typical of ﬁeldeffect devices (where hot electrons
are injected into the substrate or into the AlGaAs layer of
a heterojunction device), and a low 1/f noise . Further
more, the lower base resistance in the III–V HBTs leads to
lower power consumption, when compared with the Si bi
polar transistor.
Useful complementary materials for both HEMT and
HBT devices are In
0.53
Ga
0.47
As and In
0.52
Al
0.48
, as they
are both latticematched to InP. The main advantages of
fered are
*
Reliable technique for the growth of these materials
*
Electron mobility 1.6 times higher than in GaAs,
resulting in faster devices
*
Smaller bandgap (0.75 eV) compared with 1.42 eV for
GaAs, leading to a lower turnon voltage, reduced
power consumption, and lower voltage supply
A further advantage of the InP substrate is its higher
thermal conductivity.
11. APPLICATIONS
Highspeed transistors based on gallium arsenide, such as
MESFET, HEMT, and HBT, have found wide applications
in microwave monolithic integrated circuits (MMICs).
At the early stage of development of GaAs technology,
highspeed transistors based on GaAs were used only for
defense and space applications because of highperfor
mance system demand; more recently, the huge growth
of commercial wireless and ﬁberoptic communication sys
tems has opened up new application areas for those de
vices. GaAs MESFET was the ﬁrst developed highspeed
transistor, used for consumer wireless products; a few
years later, HEMT and HBT, succeeded it. Mainly III–V
based devices are curently used for ampliﬁers in consumer
wireless communications. Initially, when the analog cel
lular phone market was developing, siliconbased transis
tors, such as BJT, were exclusively used for power
ampliﬁer applications because of their low cost and high
volume production capability. At that time GaAs technol
ogy had not yet emerged, so cost and volume requirements
could not be met.
By the early 1990s, the GaAs MESFET, which was
more efﬁcient, was quickly replacing the Sibased transis
tor (BJT). The GaAs MESFET manufacturing technology
has matured rapidly; cost and manufacturability have be
come competitive with those of silicon BJTs. In the full
duplex analog cellular phone, output power ampliﬁer
noise ﬁgures provided by GaAs MESFET has leaded to
relaxed ﬁltering requirements, enabling a consistent cost
reduction, greater integration capability, and improve
ment efﬁciency.
In the mid1990s, cellular phone systems generally ran
on digitally modulated schemes such as GSM and CDMA.
In a short time the analog cellular phone was replaced by
a digital one. Consequently, greater efﬁciency, lower power
consumption, better noise ﬁgure, and higher linearity
were required; unfortunately the classical MESFET was
not able to provide such characteristics. For these reasons
a more efﬁcient GaAsbased transistor belonging to the
ﬁeldeffect transistor family, the highelectronmobility
transistor (HEMT), came out. HEMTs, and subsequently
pseudomorphic HEMTs, replaced GaAs MESFETs in
many application. At the same time, a new bipolar
GaAsbased transistor, known as the heterojunction bipo
lar transistor (HBT), was developed. It exhibits greater
efﬁciency than does the common siliconbased bipolar
transistor; in many applications GaAs MESFETs were re
placed by GaAs HBTs, which shows greater efﬁciency and
smaller size, while providing adequate linearity and low
noise ﬁgure.
One of the main advantages of HBTs in digital cellular
phones, as well as all mobile applications, is that they re
quire only a single polarity power supply, which is an im
portant factor for size and weight reduction.
The most recent research efforts at TRW (Redondo
Beach, CA) have shown a new commercially viable tech
nology that monolithically integrates both HEMTs and
HBTs in the same MMIC. This technology has potential
applications in many applications, including singlechip
transmit–receiver circuits, which combine HEMT low
noise front ends with HBT highlinearity output stages
along with PIN diode variable gain control.
Acknowledgments
The author gratefully acknowledges the support of A.
Cidronali, M. Camprini, and I. Magrini in the prepara
tion of the manuscript. This work was partially supported
by the European Project TARGET, Network of Excellence
in the VIFW.
FURTHER READING
M. Shur, Physics of Semiconductor Devices, PrenticeHall, 1990.
P. H. Ladbrooke, MMIC Design; GaAs FETs and HEMTs, Artech
House, 1989.
D. V. Morgan and R. H. Williams, Physics and Technology of He
terojunction Devices, IEE, 1991.
D. A. Neamen, Semiconductor Physics and Devices, Basic Princi
ples, Irwin, 1997.
K. Heime, InGaAs FieldEffect Transistors, RSP—Wiley, 1989.
B. G. Streetman and S. Banerjee, Solid State Electronic Devices,
PrenticeHall, 2000.
J. S. Yuan, SiGe, GaAs and InP Heterojunction Bipolar Transis
tor, WileyInterscience.
P. Chavarkar and U. Mishra, Field effect transistors: FETs
and HEMTs, in Handbook of Thin Film Devices, Vol. 1,
Chap. 3.
GALLIUM ARSENIDE TECHNOLOGY AND APPLICATIONS 1757
W. Liu, Fundamentals of IIIV Devices: HBTs, MESFETs, and
HFETs/HEMTs, Wiley.
P. H. Ladbrooke, MMIC Design: GaAs FETs and HEMTs, Artech
House.
P. Roblin, L. Rice, S. Bibyk, and H. Morkoc, Nonlinear parasitic in
MODFETs and MODFET IV characteristics, IEEE Trans.
Electron. Devices 35:1207–1214.
A. Cho, Film deposition by molecular beam techniques, J. Vac. Sci.
Technol. 8:S31–S38 (1971).
A. Cho and J. Arthur, Molecular beam epitaxy, Prog. SolidState
Chem. 10:157–192 (1975).
G. Turner, B. Nechay, and S. Eglash, Frequencydomain analysis
of time dependent reﬂection highenergy electron diffraction
intensity data, J. Vac. Sci. Technol. B 8:283–287 (1990).
C. T. Foxon, MBE growth of GaAs and IIIV alloys, J. Vac. Sci.
Technol. B 1(2):293–297 (1983).
J. Neave, B. Joyce, P. Dobson, and N. Norton, Dynamics of ﬁlm
growth of GaAs by MBE from RHEED observations, Appl.
Phys. A 31:1–8 (1983).
P. Cohen, P. Pukite, and J. van Hove, and C. Lent, Reﬂection high
energy electron diffraction studies of epitaxial growth on semi
conductor surfaces, J. Vac. Sci. Technol. B 3:1251–1258 (1986).
T. Block, K. Eyink, D. Neikirk, and B. Streetman, Diffraction con
dition dependence of RHEED dampening during MBE growth,
SPIE Proc. Epitaxial Growth Process. 2140:10–24 (1994).
M. Ohring, The Material Science of Thin Films, Academic Press,
1992.
G. Petrich, P. Pukite, A. Wowchak, G. Whaley, P. Cohen, and
A. Arrot, On the origin of RHEED intensity oscillations,
J. Crystal Growth 95:23–27 (1989).
J. Arthur, ‘‘Interaction of Ga and As2 molecular beams with GaAs
surfaces,’’ J. Appl. Phys. 39:4032–4033 (1968).
R. Fischer, J. Klem, T. Drummond, R. Thorne, W. Kopp, H. Mo
rkoc
,
, and A. Cho, Incorporation rates of gallium and aluminum
on GaAs during molecular beam epitaxy at high substrate
temperatures, J. Appl. Phys. 54(5):2508–2510 (1983).
G. Metze and A. Calawa, Effects of very low growth rates on GaAs
grown by molecular beam epitaxy at low substrate tempera
tures, Appl. Phys. Lett. 9(1):818–820 (1983).
Y. Horikoshi, M. Kawashima, and H. Yamaguchi, Migration en
hanced epitaxy of GaAs and AlGaAs, Jpn. J. Appl. Phys.
27:169–179 (1988).
GALLIUM NITRIDE FOR ELECTRONICS
JONGWOOK LEE
Kyung Hee University
Korea
ILESANMI ADESIDA
University of Illinois at Urbana–
Champaign
1. INTRODUCTION
The advent of blue lightemitting diodes (LEDs) and short
wavelength blueviolet laser diodes using the IIInitride
alloys(AlN, GaN, and InN) have opened up an era of new
optical devices that was not possible with the established
III–V material systems based on GaAs, AlAs, GaP, InAs,
and related alloys. The direct bandgap and a wide spec
trum of bandgap energies (0.7 eV for InN, 3.4 eV for GaN,
and 6.2 eV for AlN) of IIInitride semiconductors make
them suitable for a wide range of optoelectronic applica
tions. Fundamental breakthroughs in the late 1980s in
growth technologies such as improved epitaxial quality
using AlN buffers [1,2] and the development of ptype con
ductivity [3,4] led to the realization of blue lightemitting
diodes (LEDs) in 1993 [5,6].
Following the tremendous progress in material quality
and device processing, a worldwide market for nitride
based devices has developed. One of the highest proﬁle
markets for nitridebased LEDs is in mobile phones, par
ticularly blue LEDs for keypads and white LEDs for the
liquid crystal display (LCD) backlight. Nitridebased
LEDs also offer great promise for solidstate lighting ap
plications such as architectural lighting, machine vision,
illumination for signage, ﬂashlights, and decorative lights.
The market for highbrightness LEDs (HBLEDs) in light
ing applications has grown rapidly with an overall market
in the billions of dollars [7,8]. Based on continuing positive
trends in this dynamic industry, the market for HBLEDs
is projected to grow to B$5.0 billion by 2007 [9]. In terms
of materials, the market was dominated by InGaNbased
devices, which accounted for 68% of total HBLED sales in
2002. White LEDs are considered very promising for com
mercial market for general illumination due to progress in
lumen output, with higher efﬁciencies than that of incan
descent lamps. With the introduction of blueviolet laser
diodes [10], the development of next generation optical
storage systems (e.g., DVD drives and recorders), is also
underway [11]. Currently, the semiconductor industry is
undergoing massive changes which is partly driven by III
nitride semiconductors and associated products.
2. APPLICATIONS OF GALLIUM NITRIDE ELECTRONICS
In addition to their usefulness in optoelectronic applica
tions, GaNbased devices also have excellent electronic
properties such as high critical breakdown ﬁelds, high
saturation velocity, and good thermal conductivity. Many
laboratories around the world have demonstrated micro
wave power performance for GaN devices that are far su
perior to the performance of other solidstate device
technologies. The GaNbased highelectronmobility tran
sistor (HEMT) is considered an ideal candidate for high
power and hightemperature RF/microwave applications.
GaN HEMTs can operate at high speeds, at high power
levels, at high temperatures (B5001C), and also in very
harsh (i.e., corrosive or high radiation) environments.
Electronic systems that can reliably operate under these
conditions are currently being sought in industries such
as wireless communications, aerospace, automotive, pe
troleum, and power generation.
Commercial and military applications of GaNbased
electronics are numerous and diverse. Power ampliﬁers
(PAs) in the transmitters of 2.5G and 3G cellular base
stations are promising and highly competitive market for
GaN HEMTs. Highpower devices based on GaN clearly
1758 GALLIUM NITRIDE FOR ELECTRONICS
have many advantages in this market over Si and GaAs
with the capability of delivering continuous high power
levels without stringent cooling requirements. Currently,
silicon LDMOS (laterally diffused metaloxidesemicon
ductor) is the primary technology in use. However, the
LDMOS technology is expected to reach its fundamental
limit due to high power and high linearity requirements of
nextgeneration PAs. GaN HEMTs are expected to enter
the cellular basestation PA market sooner rather than
later. In fact, highperformance wideband UMTS GaN PAs
have already been demonstrated [12]. Reliability issues
required for widespread commercial usage are yet to be
fully addressed.
The automotive and aerospace industry represents an
other large potential market, where heattolerant GaN
electronics incorporated in control systems would improve
performance and reliability of the overall system. Wide
bandgap GaNbased devices would allow bulky hydraulics
(e.g., control actuators) and mechanical control systems
(e.g., enginegearboxdriven fuel pumps) to be replaced
with solidstate power electronics. Replacement of these
systems would reduce the weight and complexity of the
overall system.
Radar and satellitecommunications links, which oper
ate at microwave frequencies, would beneﬁt from GaN
based electronics as well. GaNelectronics can deliver high
er power levels per chip area than either Si or GaAs, and
thus smaller system components can be realized. In addi
tion, high power (dissipated handling capability allows for
smaller cooling systems, so size reductions would signiﬁ
cantly lower the overall payload weight and translate into
cost savings in delivering satellites into orbit. The petro
leum industry seeks to use GaNbased sensors to monitor
environmental conditions within an oil well (well logging)
where temperatures can be as high as 3001C. Radiation
hardened electronic instrumentation systems based on
widegap GaN could also be implemented to directly mon
itor reactor conditions in nuclear power plants.
Although, GaNbased electronic devices are considered
primarily for highpower and hightemperature applica
tions, preliminary results also suggest that they are suit
able for microwave lownoise applications. Further
improvements in minimum noise ﬁgures are expected
with optimization of GaN HEMTs. The low noise perfor
mance coupled with excellent high voltage capability
make GaN HEMTs ideal candidates for robust lownoise
ampliﬁers (LNAs). These properties of GaN HEMTs would
eliminate the need for additional protection circuits, thus
improving the system noise ﬁgure while simplifying over
all system design. The excellent microwave power capa
bilities and the development of lownoise GaN HEMTs
provide a path to the realization of highly linear, broad
band, and robust transceiver systems.
3. MATERIAL PROPERTIES
The crystal structure of GaN is a hexagonal or wurtzite
structure, as shown in Fig. 1, where the bilayers consist of
two closely spaced hexagonal layers, one formed by Ga
atoms and the other formed by N atoms. The compound
structure exhibits two different sequences (Ga face and N
face) in the atomic layers. The lack of inversion symmetry
with the strong ionicity of the covalent bonds leads to
the polarization vectors, and these polarization vectors
are additive along the caxis of the wurtzite crystal, resul
ting in macroscopic polarization ﬁelds. In AlGaN/GaN
heterostructures, there are two different polarizations:
straininduced (piezoelectric) polarization and zerostrain
(spontaneous) polarization. The magnitude and direction
of the electric ﬁelds associated with these polarization
effects depend on the substrate, growth orientation, and
nature of the surface (Gaface or Nface). The most com
mon nitride structures are grown along the [0001] direc
tion with Gaterminated surface. The highpolarization
electric ﬁelds in the AlGaN layer induce a high two
dimensional electron gas (2DEG) with very high sheet
carrier density at the AlGaN/GaN interface without in
tentional doping [13]. Due to large piezoelectric constants,
the piezoelectric effect is important in determining the
charge control model of AlGaN/GaN HEMTs.
As discussed earlier, GaNbased transistors are prom
ising candidates for highpower and hightemperature ap
plications due to the materials’ superior electronic
properties. Table 1 shows the material parameters for
widebandgap semiconductors (GaN and 4HSiC) as well
as for Si and GaAs [14–17]. The obvious difference
between GaN and conventional semiconductors is the
wide bandgap, which results in significantly lower intrin
sic carrier density for GaN. This allows GaNbased devices
Figure 1. Schematic of (a) Gaface and (b) Nface wurtzite GaN
crystal structures [13].
Table 1. Material Properties of Various Semiconductors
Physical Property/Material Si GaAs 4HSiC GaN
Bandgap (eV) 1.12 1.43 3.2 3.4
Electron mobility (cm
2
/Vs) 1350 8500 800 1500
Peak electron saturation
velocity (cm/s) 10
7
1 1 2 1.5
Breakdown ﬁeld 3 4 32 34
(V/cm) 10
5
Thermal conductivity
(300K) (W/cmK)
1.5 0.5 3.3 1.3
Dielectric constant 11.8 12.8 9.7 9.5
GALLIUM NITRIDE FOR ELECTRONICS 1759
to operate at much higher temperatures. Also, the mate
rial system has high breakdown ﬁelds (Z3 MV/cm) and
high saturation velocity (1.5 10
7
cm/s). Breakdown ﬁeld
is a very important criterion for highpower electronics,
since the maximum output power for a class A ampliﬁer
can be estimated from the following expression
P
max
=
(V
BK
÷V
knee
)
2
8R
L
;
where R
L
is the load resistance, V
BK
is a transistor’s
breakdown voltage, and V
knee
is the knee voltage, deﬁned
as the voltage at which the transistor current saturates. It
is obvious that a highpower output requires high break
down voltage, which in most cases is directly related to the
critical breakdown ﬁeld. From Table 1, it can be seen that
the breakdown ﬁeld for GaN is much higher than that for
GaAs and Si. It means that GaNbased devices can with
stand higher bias voltages, thus producing higher total
power output.
The current drive capability of AlGaN/GaN hetero
structure is also excellent due to very high sheet carrier
density and high saturation velocity. The high conduction
band offsets at AlGaN/GaN interfaces results in high elec
tron mobilities (Z2000 cm
2
/Vs [18]). Because of the strong
spontaneous and piezoelectric polarization effects, AlGaN/
GaN heterostructure typically has 2DEG with sheet den
sity values of about 1 10
13
/cm
2
with no intentional dop
ing [19]. These values are higher than those seen in
AlGaAs/InGaAs and InAlAs/InGaAs heterostructures. In
the form of AlGaN/GaN HEMTs, these transport proper
ties translate to both high current drive capability and
high breakdown voltage. High power operation is further
facilitated by the use of semiinsulating SiC substrates
that are characterized by high thermal conductivity
(3.3 W/cmK).
By combining various materials’ parameters, the over
all potential of any semiconductor material system for
highpower, highfrequency performance can be predicted
using wellknown ﬁgures of merit (FoMs). For example,
Johnson’s ﬁgure of merit, JFoM=(E
C
v
sat
=2p)
2
(where E
C
is the critical breakdown ﬁeld and v
sat
is the saturation
velocity) [20], deﬁnes the power frequency
2
product for
the transistor and determines its output power capability
at given frequencies. In addition, Shenai’s ﬁgure of merit,
Q
F1
=ls
A
(where l is the thermal conductivity and s
A
is
the conductance of the channel) [21], measures the power
handling capability in terms of the dissipated power. Alek
seev et al. [22] proposed an extended definition of Shenai’s
ﬁgure of merit, Q
FS
=l
sub
s
A
, which takes into account
thermal conductivity of the substrate material (sapphire
or SiC). Table 2 compares the FoMs calculated for GaN
and GaAsbased devices. It can be seen that GaN is a far
more superior material to GaAs for highpower and high
frequency applications.
4. AlGaN/GaN HEMT OPERATION AND FABRICATION
The basic principle of HEMToperation involves the trans
fer of electrons from ionized donor impurities in the wider
bandgap semiconductor (e.g., AlGaN) to form a 2DEG in a
potential well at the AlGaN/GaN heterointerface. The
term ‘‘high electron mobility transistor’’ is used because
the structure takes advantage of the superior transport
properties (high mobility and velocity) of electrons in the
lightly doped semiconductor material (e.g., GaN).
A typical AlGaN/GaN HEMT structure is shown in
Fig. 2; it consists of an AlGaN barrier layer grown on
GaN channel layer. The device structures are mainly
grown on sapphire or semiinsulating SiC substrates. Since
neither of these substrates is latticematched to GaN, an
AlN buffer layer is generally used to isolate the channel
layer from the substrate. Defects and dislocations occur in
the epitaxial layers due to lattice mismatch, with typical
dislocation densities on the order of 10
8
–10
9
/cm
2
. The lat
tice mismatches of GaN on sapphire and SiC substrates
are 13% and 3.1%, respectively; therefore, the GaN crystal
quality is better on SiC. As the thermal conductivity of SiC
is about 10 times that of sapphire, it is evident that SiC is
better suited for highpower and hightemperature appli
cations.
We delineate a typical device fabrication procedure
here. The ﬁrst step for the device fabrication is mesaisola
tion, which can be etched down to the undoped GaN using
Cl
2
/Ar plasma in an inductively coupledplasma reactive
ion etch (ICPRIE) system. Next, ohmic contacts are
formed by rapid thermal annealing of either evaporated
Ti/Al/Ti/Au [23] or Ti/Al/Mo/Au [24] in N
2
ambient. A pre
treatment of the ohmic contact area using SiCl
4
plasma in
a RIE is performed prior to Ti/Al/Ti/Au (Ti/Al/Mo/Au)
metallization. In the case of Ti/Al/Mo/Au ohmic metal
lization, the surface morphology of the annealed contact is
very smooth as well as having low specific contact resisti
vity. No degradation in contact resistance was observed
when the contact was subjected to longterm annealing at
Table 2. Comparison of GaN and GaAsBased Devices for
HighPower and HighFrequency Performance
a
Figures of Merit (FoMs) GaAs GaN on Sapphire GaN on SiC
JFoM=(E
C
v
sat
=2p)
2
1 100 100
Q
F1
=ls
A
1 22 22
Q
FS
=l
sub
s
A
1 5 76
a
Results are normalized to GaAs.
Source Drain
Gate
AIGaN barrier
GaN
AIN buffer
Sapphire or SiC substrate
2DEG
Figure 2. Basic AlGaN/GaN HEMT structure.
1760 GALLIUM NITRIDE FOR ELECTRONICS
5001C for 360h, demonstrating excellent thermal stability.
Tshaped gates are deﬁned using a trilayer PMMA/
P(MMAMAA)/PMMA and electronbeam lithography. Ni/
Au metals are then evaporated for gate metallization.
Overlay metallization on ohmic contacts and measure
ment pads are also deposited during gate formation. Sil
icon nitride passivation is then deposited.
Fig. 3 shows the micrograph of a GaN HEMT.
5. STATEOFTHEART AlGaN/GaN HEMTs
The ﬁrst AlGaN/GaN HEMTs were demonstrated by Khan
et al. in 1994 [25]. Transistors with 0.25mm gate length
produced a current density of 60 mA/mm and a transcon
ductance of 27 mS/mm. Since then, dramatic progress has
been made in material growth and process technology for
AlGaN/GaN HEMTs. A breakdown voltage as high as
570V in an AlGaN/GaN HEMT with a source–drain spac
ing of 13 mm and a gate length of 0.5mm using an overlap
ping gate structure was reported [26]. AlGaN/GaN
HEMTs with current density of 2.1A/mm have also been
reported [27]. A power density as high as 10.7 W/mm at
10 GHz has been achieved on GaN HEMTs [28]; however,
their suitability for power applications at frequencies
Z30 GHz is currently being limited by the speed of these
devices.
5.1. HighSpeed Performance
AlGaN/GaN HEMTs with a unity current gain cutoff fre
quency (f
T
) of 110 GHz and a maximum frequency of os
cillation (f
max
) of 4140GHz have been demonstrated by
reducing the gate length down to 50 nm [29]. However,
this reduction of gate length leads to a decrease in break
down voltage.
To achieve higher performance, indium has been uti
lized as a surfactant during the growth of AlGaN/GaN
HEMTs [30]. A demonstration of the concept used a MO
CVDgrown epilayer which consists of a 100nm AlN buf
fer, 2mm undoped GaN, a 5nm undoped Al
.25
Ga
.75
N
spacer, a 10nm Sidoped (B5 10
18
/cm
3
) Al
.25
Ga
.75
N
charge supply layer, and a 10nm undoped Al
.25
Ga
.75
N
barrier layer. The layer was grown on 4HSiC. A lowlevel
ﬂux of trimethylindium (TMI) during the growth of the
structure helped in improving the surface and interface
roughness through the incorporation of trace amounts of
indium. Hall measurements showed a sheet carrier con
centration of 1.1 10
13
/cm
2
and an electron mobility of
1300 cm
2
/Vs at room temperature on asgrown wafers.
A device fabricated from this layer demonstrated a typ
ical drain current density of 1.23 A/mm, peak transcon
ductance of 314mS/mm, and a gate drain breakdown
voltage of over 60 V. Figure 4 shows the shortcircuit cur
rent gain (h
21
) and maximum stable gain/maximum
available gain (MSG/MAG) derived from onwafer Spa
rameter measurements as a function of frequency for the
0.12mmgatelength device. The effect of pad parasitics
was deembedded by conventional Yparameter subtrac
tion. The values of unity current gain cutoff frequency (f
T
)
and maximum frequency of oscillation (f
max
) were deter
mined by extrapolation of the h
21
 and MSG data at
20 dB/ decade. At a drain bias of 8V and a gate bias of
÷3.95V, an f
T
of 121GHz, and an f
max
of 162GHz were
obtained, which to the best of the authors’ knowledge are
the highest data ever reported for GaNbased HEMTs
[31]. This excellent RF performance is attributed to the
high quality of material and also to the optimized device
processing.
5.2. HighPower Performance
In terms of power, widebandgap AlGaN/GaN HEMTs on
semiinsulating SiC substrates have yielded a continuous
wave (CW) power density of 32 W/mm at 4GHz and
16.5 W/mm at 10 GHz, and a pulsed power output of
113 W at 1.95 GHz, making them promising candidates
for nextgeneration commercial power ampliﬁer systems
[32–34]. Also, more recent millimeterwave power perfor
mance of GaN HEMTs suggests that operation of GaN
HEMTs can be extended to K/Ka band. At 20GHz, 3 W/mm
CW power density with 22.5% poweradded efﬁciency
Figure 3. Micrograph of a GaN HEMT.
h
2
1
(
d
B
)
M
S
G
/
M
A
G
(
d
B
)
40
30
20
10
0
1 10 100
Frequency (GHz)
Figure 4. Shortcircuit current gain (h
21
) and maximum sta
ble/maximum available gain (MSG/MAG) of a 0.12mm100mm
AlGaN/GaN HEMT on SiC substrates. The device was biased at
V
DS
=8 V and V
GS
= ÷3.95V [31].
GALLIUM NITRIDE FOR ELECTRONICS 1761
(PAE) has been attained for 0.3mmgatelength devices
[35] and 6.6W/mm CW power density with 35% PAE for
0.15mmgatelength devices [36]. At 35 GHz, an output
power density of 4.13 W/mm with 23% of PAE and 7.54dB
of linear gain were achieved at a drain bias of 30 V [37].
The combined results of power density, PAE, and gain that
have been achieved using GaN HEMT technology are im
pressive. For comparison, siliconbased transistors can ef
ﬁciently amplify signals up to only 2–3 GHz. GaAs
transistors can handle 10 GHz but deliver a power densi
ty of less than 1 W/mm at that frequency.
At relatively low frequencies, the ﬁeldplated technique
have achieved impressive power results [32,33]. Figure 5
shows the measured CW power, gain, and PAE of a mod
iﬁed ﬁeldplated device at 10 GHz, where the AlGaN/GaN
HEMTs with an optimized gate structure enabled the de
vice to operate at higher drain biases than did conven
tional gate structure, yielding improved power levels.
Details of the device structure are reported elsewhere by
Thompson et al. [33]. Power data were taken on wafer us
ing an automatic loadpull system at room temperature.
The large signal performance of a 0.35mmgatelength de
vice was measured when the device was biased at a drain–
source voltage of 60 V. The device had a saturated output
power density of 16.5 W/mm and PAE of 47% as shown in
Fig. 5.
Stateoftheart CW power performances at 20 GHz
were obtained from a 0.25 mm100mm AlGaN/GaN
HEMT on 4HSiC substrates [38]. The devices exhibited
high current density, transconductance, and millimeter
wave output power density. The largesignal performance
of the device at 20 GHz is shown in Fig. 6. The device was
biased with a drainsource voltage of 30 V at a DC drain
current of 620mA/mm. The device had a saturated output
power of 6.4W/mm with an associated gain of 2.9dB and
PAE of 16%. The peak efﬁciency was 22% with an output
power of 5.8 W/mm and gain of 6.1dB.
Figure 7 shows the large signal performance of a
0.25 200mm AlGaN/GaN HEMT at 30 and 35 GHz
[37]. With a drain voltage of 30 V and a drain current of
325 mA/mm, the device had a saturated output power
density of 5.43 W/mm at 30 GHz with 9.17 dB gain and
PAE of 33%. When the device was tested at the same bias
conditions at 35 GHz, the device delivered a remarkable
output power of 4.13 W/mm with a peak PAE of 23% and a
linear gain of 7.54dB. This performance vividly demons
trates the potential for veryhighpower solidstate ampli
ﬁers employing 0.25mmgatelength GaNbased HEMTs
to replace TWTampliﬁers in spacebased millimeterwave
communication systems.
The combined results of power density, PAE, and gain
that have been achieved to date using GaN HEMT tech
nology are impressive. No other solidstate device tech
nologies have demonstrated these capabilities. Figure 8
shows the various reported power densities over a fre
P
o
u
t
(
d
B
m
)
,
G
a
i
n
(
d
B
)
P
A
E
(
%
)
35
30
25
20
15
10
5
0
−10 −5 0 5 10 15 20 25
50
40
30
20
10
0
P
out
= 16.5 W/mm, PAE = 47%
P
in
(dBm)
V
DS
= 60 V
0.35 x 150 µm Gate
Figure 5. CW power sweep of a 0.35 150mm ﬁeldplated gate
device at 10GHz, demonstrating output power of 16.5W/mm and
PAE of 47% when operated at 60V drain bias [33].
P
o
u
t
(
d
B
m
)
,
G
a
i
n
(
d
B
)
P
A
E
(
%
)
30
25
20
15
10
5
0
0 5 10 15 20 25 30
30
25
20
15
10
5
0
P
in
(dBm)
Figure 6. Largesignal performance of a 0.25 100mm AlGaN/
GaN HEMT on 4HSiC at 20GHz when biased with V
DS
=30V
and V
GS
= ÷1.87V, saturated output power was 6.4 W/mm with
2.9 dB gain and 16% PAE [38].
P
o
u
t
(
d
B
m
)
,
G
a
i
n
(
d
B
)
,
P
A
E
(
%
)
25
30
35
Open symbols : 30 GHz
Filled symbols : 35 GHz
V
DS
= 30 V
20
15
10
5
0
0 5 10 15 20 25
P
in
(dBm)
4.13
W/mm
22.9%
5.54dB
PAE
Gain
P
out
Figure 7. Output power, PAE, and gain as a function of input
power of a 0.25mm 4 50 mm AlGaN/GaN HEMT at 30 and
35GHz [37].
1762 GALLIUM NITRIDE FOR ELECTRONICS
quency range of 4–40GHz. With these results, GaN
HEMTs have demonstrated at least an order of magni
tude higher power density over the existing Si and GaAs
based RF and microwave transistor technologies. The ex
cellent material properties of AlGaN/GaN material system
have indeed delivered a new level of microwave power and
frequency operation.
5.3. LowNoise Performance
In parallel with the improvement in power density results,
microwave noise performance has also shown rapid im
provement. From the ﬁrst report [39] on the noise char
acteristics of GaN HEMTs to more recent results [40–42],
it has been shown that these devices exhibit promising
microwave noise properties. The minimum noise ﬁgure of
0.25mmgatelength AlGaN/GaN HEMTs [43] is compara
ble to the results reported previously for a pseudomorphic
GaAs HEMT with similar gate length [44]. The investiga
tion of GaNbased devices for microwave noise applica
tions is of interest because widebandgap GaN with its
relatively low intrinsic carrier generation and the high
breakdown ﬁelds provides an enabling technology that
can sustain a higher input power and breakdown voltages
at higher temperatures than GaAs and Sibased devices.
Figure 9 shows the measured microwave noise perfor
mance of 0.12mmgatelength GaN HEMTs on SiC sub
strate [40]. The minimum noise ﬁgure (NF
min
) and
associated gain (G
a
) of the devices were measured on wa
fer using an ATN NP5 noise parameter test set over the 4–
18 GHz frequency range. For these measurements, devices
were biased at V
DS
=10 V and V
GS
= ÷4.8 V. The devices
exhibited an NF
min
of 0.77dB and 0.99 dB at 12 and
18 GHz, respectively. This excellent noise performance is
attributed to the higher gains engendered by the smaller
gate length and the excellent low leakage Schottky gate
characteristics. The current results show significant im
provement compared to previously reported 0.25mm Al
GaN/GaN HEMTs that had minimum noise ﬁgures of
1.8 dB [39] and 2.0dB [42] at 18 GHz, respectively.
6. GALLIUM NITRIDE RF AND MICROWAVE
POWER AMPLIFIERS
6.1. Advantages of GaNBased Power Ampliﬁers
Highpowerdensity GaN HEMTs provide many advanta
ges for the realization of power ampliﬁers: (1) higher out
put power per unit area needs smaller device size and
translates to broader bandwidth for given output power, (2)
small device size reduces matching circuit complexity, and
(3) efﬁcient heat dissipation through high thermal conduc
tivity of SiC substrates relieves cooling requirement. For
highpower ampliﬁers where multimillimetertocentime
ter gate peripheries are needed, large impedance transfor
mation ratio limits useful bandwidth for given output
power requirement. With higher power density per pico
farad of input/output capacitance of GaN HEMTs than
devices based on Si and GaAs, multioctave broadband
GaN HEMT ampliﬁers are realizable. Also, small device
size for the same output power rating drastically reduces
both the input and output impedance transformation ra
tios, thus simplifying overall matching circuits. For exam
ple, let us consider how highpowerdensity GaN HEMTs
can simplify overall ampliﬁer system. Figure 10 shows a
conventional ampliﬁer based on InGaAs pHEMTs, which
produce a total output power of 4W at 20 GHz. With the
minimum impedance transformation requirement, the
output power of an InGaAsbased HEMT MMIC ampliﬁ
er is limited to about 1 W at 20 GHz. To deliver a total
output power of 4 W, external power combining of multi
ple MMICs is necessary. The external power combining
not only adds to assembly cost but also makes the overall
system complicated and large. If GaN HEMTs are utilized
for the ampliﬁer, a single GaN MMIC can directly produce
4 W at 20 GHz without bulky external power combi
ners. Thus, using GaN HEMTs provides simple and com
pact design, which enhances overall system reliability.
These advantages are valuable for applications requir
ing broad bandwidth, high temperature, and high power
operation such as in commercial wireless base stations,
35
30
25
20
15
10
5
0
0 10 20 30 40 50
Frequency (GHz)
32 W/mm (UCSB)
16 W/mm (Cornell)
6.5 W/mm (UIUC)
5.4 W/mm (Triquint)
P
out
(W/mm)
Data fit
P
o
u
t
(
W
/
m
m
)
Figure 8. Reported power densities over a frequency range of
4–40GHz are shown with symbols. The solid line represents a
best ﬁt to the points.
M
i
n
i
m
u
m
n
o
i
s
e
f
i
g
u
r
e
(
d
B
)
A
s
s
o
c
i
a
t
e
d
p
o
w
e
r
g
a
i
n
(
d
B
)
3.0
2.5
2.0
1.5
1.0
0.5
0.0
2 4 6 8 10 12 14 16 18 20
20
15
10
5
0
Frequency (GHz)
Figure 9. Minimum noise ﬁgure (NF
min
) and associated power
gain (G
a
) versus frequency for a 0.12mm AlGaN/GaN HEMT; the
solid line is the modeled minimum noise ﬁgure for the extrinsic
device; the bias condition was V
DS
=10 V and V
GS
= ÷4.8V [40].
GALLIUM NITRIDE FOR ELECTRONICS 1763
satellite communications, local multipoint distribution
system (LMDS), and assorted defense systems.
More recent applications of GaN HEMTs in broadband
power ampliﬁers have included a monolithic nonuniform
distributed ampliﬁer with a cascodeconnected gain cell
[45], an ampliﬁer with LCR matching having an input
power splitter and an output combiner [46], and a two
stage, reactively matched MMIC ampliﬁer with backside
via holes through the SiC substrate [47]. In addition to
linear class A operation or nonlinear class AB operation,
GaNbased class B ampliﬁers also have been demonstrat
ed at microwave frequencies. A singleended class B am
pliﬁer employing bandpass ﬁltering at the output resulted
in a 36 dBm output with good linearity [48]. In an ampli
ﬁer with two devices for push–pull operation, a maximum
output power of 35 dBm and a PAE of 42% were obtained
at 5 GHz [49]. These GaNbased microwave power ampli
ﬁers have demonstrated great potential for achieving both
high power and broadband performance at microwave fre
quencies.
6.2. 50W CBand GaN HEMT Power Ampliﬁer
A Cband power ampliﬁer is described as an example to
show the advantages of using the AlGaN/GaN HEMT for
highpower ampliﬁers [50]. The devices used in the am
pliﬁer were 0.6mmgatelength GaN HEMTs with maxi
mum drain current of B1 A/mm and a breakdown voltage
of 80 V. The power density of a unit device was more than
6 W/mm and a total gate periphery of 8 mm was chosen for
the ampliﬁer to achieve 50 W total output power at 6 GHz.
The AlGaN/GaN HEMTs typically had an input capaci
tance of 2.7 pF/mm, similar to that of a GaAsbased
HEMT, while the optimum output load was 75 O· mm,
about 2 times that of a GaAs HEMT. Since the GaN HEMT
offers 10 times the power density for the same output
power, the input transformation ratio is 10 times less,
while the transformation ratio is 20 times less than a
GaAs HEMT at the output. To achieve 50 W output power
using the conventional GaAsbased FET technology
(either HEMT or MESFET) the device gate periphery
would have to be 80 mm, requiring a challenging input/
output impedance transformation ratio.
With the reduced impedance transformation ratio for
GaNbased ampliﬁer, an LCRmatching network was used
to convert the capacitive gate impedance to a real value of
B1.5 O within the bandwidth. This was then transformed
to the 50 O input impedance by an output matching net
work. At the output, the power load for the 8mmwide
device was transformed to the 50 W circuit output. The
fabricated ampliﬁer is shown in Fig. 11. The circuit was
laid out in a coplanarwave transmissionline system and
constructed using a ﬂipchip integratedcircuit scheme for
thermal management and electrical connection. The am
pliﬁer exhibited a midband smallsignal gain of 14.5 dB
when biased at 25V. Pulsedpower measurement was per
formed using 0.5 ms pulsewidth and 5% duty cycle. The
output power at 6 GHz was 51W when biased at 39 V with
a corresponding power density of 6.4 W/mm. This repre
sents a remarkable power output obtained from a small
periphery solidstate FETand simple input/output match
ing network design.
6.3. Broadband HighPower Cascode GaN HEMT MMICs on
SiC Substrates
GaN HEMTs have higher output power per device input/
output capacitance ratio than do conventional devices. On
the basis of this idea, a broadband highpower cascode
GaN HEMT MMIC ampliﬁer with high gain and PAE was
fabricated on SiC substrates [45]. Traditionally, distribut
ed ampliﬁers (DAs) have been used for applications re
quiring broad bandwidths. DAs offer broadband operation
by incorporating gain elements in synthetic lumpedele
ment approximate transmission lines, realized by the
transistor capacitances and intervening inductances [51].
This topology allows the addition of transconductance
Input
power
combining
Output
power
combining
Figure 10. A 20GHz InGaAs pHEMTs power ampliﬁer system.
To deliver overall output power of 4W, external power combining
is required. (From the 6th Widegap IIINitride Workshop at
UCSB; ﬁgures of MMIC ampliﬁers are from Triquint Inc.).
LCR
matching
RF out
Flipped GaN
HEMTs
RF in
Figure 11. Chip microphotograph of a 50W GaNbased ﬂipchip
integrated power ampliﬁer incorporating a 8mm AlGaN/GaN
HEMT; die size is about 107 mm
2
[50].
1764 GALLIUM NITRIDE FOR ELECTRONICS
without adding device capacitance, thus resulting in
excellent gain–bandwidth products.
Figure 12 shows a multioctave broadband GaN HEMT
MMIC ampliﬁer fabricated on SiC substrates. It contains
three cascadeconnected GaN HEMTs as gain elements.
Each device has 1mm of total gate periphery for both the
commonsource and commongate devices. The two devic
es are separated by a 125mm section of a highimpedance
transmission line to further facilitate heat dissipation.
Cascode topology reduces the Miller effect, resulting in
reduced feedback capacitance. A cascode gain cell exhib
ited 5 Woutput power at 8 GHz with a smallsignal gain of
19 dB, thus providing high gain and broad bandwidth for
the ampliﬁer. All gate CPW line impedances are 80 O, and
the drainline impedances, from the left to righthand
sides, are 60 O, 50 O, and 30 O. The drainline dummy load
was removed to prevent power consumption by the syn
thetic drainline termination dummy load, thus improving
PAE of the ampliﬁer. The gate and drainline sections
were carefully optimized for optimum power and efﬁciency
using the genetic algorithm method. The resulting ampli
ﬁer has nonuniform drainline lengths and impedance to
effectively combine output power from three gain ele
ments. The ampliﬁer based on this process yielded a sat
urated output power of 3–6 W over a DCto8 GHz
bandwidth with an associated PAE of 13–31%.
6.4. Broadband GaNBased FlipChip Integrated
Power Ampliﬁer
With excellent gain–bandwidth product, distributed am
pliﬁers (DAs) are preferred for broadband operation. How
ever, conventional DAs usually exhibit poor output power
because nonuniform device loading along the drain line
limits effective combining of devices, and also because fre
quencydependent attenuation along the gate line sup
plies limits drive power for the active devices in the later
stages.
A new circuit topology employed novel LCRmatching
networks in a fourway Wilkinson combiner structure to
divide and combine power uniformly from each device.
Figure 13 shows a highpower broad band microwave am
pliﬁer that used multiplestage quarterwavelength trans
formers for the power combiner to realize the desired
bandwidth and impedance transformation ratios [46].
Additional LC matching networks were then utilized to
reactively compensate for input and output capacitance
and provide additional impedance transformation. GaN
epitaxial layers grown on sapphire substrates have the
advantage of lower cost and availability in larger wafer
size than does SiC, but it has poor thermal conductivity.
Adequate thermal management was achieved through
ﬂipchip bonding of the device onto a thermally conduc
tive, electrically insulating AlN substrate, which also
hosts the matching and combiner networks.
Using devices with 0.7mm gate length and 4 mm gate
width, a smallsignal gain of 7 dB was obtained over
3–10 GHz bandwidth. A CW output power of 8 W at
9.5 GHz with about 20% PAE was achieved when biased
at 24 V, whch is a significantly higher output power over
the entire bandwidth compared to the results obtained
using GaAsbased FET technology.
7. CONCLUSION
This article described the highspeed, highpower, and
lownoise performance of GaN HEMTs. Further improve
ment of device speed, power, and noise ﬁgure will be pos
sible with optimization of the GaN HEMT design, growth,
and fabrication. In combination with the excellent micro
wave power capabilities, GaN HEMTs have great poten
tial for other applications such as low noise and high
power switching, and will be very attractive for delivering
highly linear, broadband, and robust transceiver in future
commercial and military communication systems.
Acknowledgment
This work at UIUC was supported by ONR under Contract
N000140111000 (program monitor: H. Dietrich).
BIBLIOGRAPHY
1. H. Amano, N. Sawaki, I. Akasaki, and Y. Toyoda, Metalor
ganic vapor phase epitaxial growth of high quality GaN ﬁlms
using an AlN buffer layer, Appl. Phys. Lett. 48: 353–355
(1986).
Gate bias On chip capacitor
Cascodeconnected GaN HEMT
RF
in
RF
out
Figure 12. Chip microphotograph of the nonuniform distributed
ampliﬁer with cascodeconnected GaN HEMTs as gain elements;
die size is about 2.51.4 mm
2
[45].
Input divider
RF In
Output
combiner
LC&LCR matching
networks
RF out
Flipped
GaN HEMTs
Figure 13. Chip microphotograph of the broadband GaNHEMT
ﬂipchip integrated power ampliﬁer; die size is about 12 8 mm
2
[46].
GALLIUM NITRIDE FOR ELECTRONICS 1765
2. I. Akasaki, H. Amano, Y. Koide, K. Hiramatsu, and N. Saw
aki, Effects of AlN buffer layer on crystallographic structure
and on electrical and optical properties of GaN and
Ga
1 ÷x
Al
x
N grown on sapphire substrate by MOVPE, J. Cry
st. Growth 98: 209–219 (1989).
3. H. Amano, M. Kito, K. Hiramatsu, and I. Akasaki, Ptype
conduction in Mgdoped GaN treated with lowenergy elec
tron beamirradiation (LEEBI), Jpn. J. Appl. Phys. 28: L2112–
2114 (1989).
4. S. Nakamura, N. Iwasa, M. Senoh, and T. Mukia, Hole com
pensation mechanism of ptype GaN ﬁlms, Jpn. J. Appl. Phys.
31:1258–1266 (1992).
5. S. Nakamura, T. Mukai, and M. Senoh, Highpower GaN PN
junction bluelightemitting diodes, Jpn. J. Appl. Phys. 30:
L1988–L2001 (1991).
6. S. Nakamura, M. Senoh, and T. Mukai, PGaN/NInGaN/N
GaN double heterostructure bluelightemitting diodes, Jpn.
J. Appl. Phys. 32:L8–L11 (1993).
7. T. Whitaker and J. Newey, Things to watch in 2003, Com
pound Semiconductors. 9 (1):23 (2003).
8. HBLED market reaches $1.8 billion, Compound Semicon
ductors. 9 (2):17 (2003).
9. Highbrightness LEDs 2003–market review and forecast,
Strategies Unlimited (June 2003).
10. S. Nakamura, M. Senoh, S. Nagahama, N. Iwasa, T. Yamada,
T. Matsushita, H. Kiyoku, Y. Sugimotu, T. Kozaki, H. Ume
moto, M. Sano, and K. Chocho, Highpower, longlifetime In
GaN/GaN/AlGaNbased laser diodes grown on pure GaN
substrates, Jpn. J. Appl. Phys. 37:L309–L312 (1998).
11. J.W. Lee and I. Adesida, DVD Forum approves blue laser
format, Compound Semiconductors. 9(11) (2003).
12. RF micro’s nitride operation demos ﬁrst GaN power amps,
Compound Semiconductors. 9(3) (March 2003).
13. O. Ambacher, J. Smart, J. R. Shealy, N. G. Wiemann, K. Chu,
M. Murphy, W. J. Schaff, L. F. Eastman, R. Dimitrov, L. Wit
tmer, M. Stutzman, W. Rieger, and J. Hilsenbeck, Twodimen
sional electron gases induced by spontaneous and
piezoelectric polarization charges in N and Gaface AlGaN/
GaN heterostructures, J. Appl. Phys. 85(6):3222–3232
(1999).
14. Y. S. Park, Semiconductors and Semimetals: SiC Materials
and Devices, Academic Press, New York, 1988, Vol. 52, pp.
195–197.
15. H. Morkoc, S. Strite, G. B. Gao, M. E. Lin, B. Sverdlov, and M.
Burns, Largebandgap SiC, IIIV nitride, and IIVI ZnSe
based semiconductor device technologies, J. Appl. Phys. 76:
111–120 (1998).
16. C. E. Weitzel, and K. E. Moore, Silicon carbide and gallium
nitride RF power devices, MRS Symp. Proc. 483:111–120
(1998).
17. J. C. Zolper, Junction ﬁeld effect transistors for hightemper
ature or highpower electronics, MRS Symp. Proc. 483:83–87
(1998).
18. R. Gaska, J. W. Yang, A. Osinsky, Q. Chen, M. A. Khan,
A. O. Orlov, G. L. Snider, and M. S. Shur, Electron transport
in AlGaNGaN heterostructures grown on 6HSiC subs
trates, Appl. Phys. Lett. 72 (6): 707–709 (1998).
19. L. F. Eastman, V. Tilak, J. Smart, B. M. Green, E. M. Chum
bes, R. Dimitrov, H. Kim, O. S. Ambacher, N. Weimann,
T. Prunty, M. Murphy, W. J. Scha, and J. R. Shealy, Undoped
AlGaN/GaN HEMTs for microwave power ampliﬁcation,
IEEE Trans. Electron Devices 48(3):479–485 (March
2001).
20. E. O. Johnson, Physical limitations on frequency and power
parameters of transistors, RCA Rev. 163–177 (1965).
21. K. Shenai, R. S. Scott, and B. J. Baliga, Optimum semicon
ductor for highpower electronic, IEEE Trans. Electron
Devices 36(9):1811–1823 (1989).
22. E. Alekseev, D. Pavlidis, N.X. Nguyen, C. Nguyen, and
D. E. Grider, Power performance and scalability of AlGaN/
GaN power MODFETs, IEEE Trans. Microwave Theory Tech.
48(10):1694–1700 (2000).
23. A. T. Ping, Q. Chen, J. W. Yang, M. A. Khan, and I. Adesida,
The effects of reactive ion etchinginduced damage on the
characteristics of ohmic contacts to ntype GaN, J. Electron.
Mater. 27:261–265 (April 1998).
24. V. Kumar, L. Zhou, D. Selvanathan, and I. Adesida, Thermal
lystable lowresistance Ti/Al/Mo/Au multilayer ohmic con
tacts on nGaN, J. Appl. Phys. 92:1712–1714 (2002).
25. M. A. Khan, J. N. Kuznia, D. T. Olson, W. J. Schaff, J. W.
Burm, and M. S. Shur, Microwave performance of a 0.25mm
gate AlGaN/GaN heterostructure ﬁeld effect transistors,
Appl. Phys. Lett. 65(9):1121–1123 (1994).
26. N. Q. Zhang, S. Keller, G. Parish, S. Heikmann, S. P. Den
barrs, and U. K. Mishra, High breakdown GaN HEMT with
overlapping gate structure, IEEE Electron Device Lett. 21:
421–423 (Sept. 2000).
27. Q. Chen, J. W. Yang, M. A. Khan, A. T. Ping, and I. Adesida,
High transconductance AlGaN/GaN heterostructure ﬁeld ef
fect transistors on SiC substrates, Electron. Lett. 33:1413–
1414 (July 1997).
28. V. Tilak, B. Green, V. Kaper, H. Kim, T. Prunty, J. Smart,
J. Shealy, and L. Eastman, Inﬂuence of barrier thickness on
the highpower performance of AlGaN/GaN HEMTs, IEEE
Electron Device Lett. 22: 504–506 (Nov. 2001).
29. M. Micovic, N.X. Nguyen, P. Janke, W. S. Wong, P. Hashimoto,
L. M. McCray, and C. Nguyen, GaN/AlGaN high electron mo
bility transistors with f
T
of 110GHz, Electron. Lett. 36:358–
359 (Feb. 2000).
30. M. A. Khan, X. Hu, G. Simin, J. Yang, R. Gaska, and M. S.
Shur, AlGaN/GaN metaloxidesemiconductor hetero struc
ture ﬁeld effect transistors on SiC substrates, Appl. Phys.
Lett. 77:1339–1341 (Aug. 2000).
31. V. Kumar, W. Lu, R. Schwindt, A. Kuliev, G. Simin, J. Yang,
M.A. Khan, and I. Adesida, AlGaN/GaN HEMTs on SiC with
f
T
of over 120GHz, IEEE Electron Device Lett. 23(8):455–457
(Aug. 2002).
32. Y.F. Wu, A. Saxler, M. Moore, R. P. Smith, S. Sheppard, P. M.
Chavarkar, T. Wisleder, U. K. Mishra, and P. Parikh, 30W/
mm GaN HEMTs by ﬁeld plate optimization, IEEE Electron
Device Lett. 25(3):117–119 (March 2004).
33. R. Thompson, T. Prunty, V. Kaper, and J. R. Shealy,
Performance of the AlGaN HEMT structure with a gate
extension, IEEE Trans. Electron Devices 51(2):292–295 (Feb.
2004).
34. Y. Ando, Y. Okamoto, H. Miyamoto, N. Hayama, T. Naka
yama, K. Kasahara, and M. Kuzuhara, A 110W AlGaN/GaN
heterojunction FET on thinned sapphire substrate, IEDM
Tech. Digest, Washington, DC, Dec. 2001, pp. 381–384.
35. A. Vescan, R. Dietrich, A. Wieszt, A. Shurr, H. Leier, E. L.
Piner, and J. M. Redwing, AlGaN/GaN MODFET’s on semi
insulating SiC with 3W/mm at 20GHz, Electron Lett. 36:
1234–1236 (July 2000).
36. J. S. Moon, M. Micovic, P. Janke, P. Hashimoto, W.S. Wong, R.
D. Widman, L. McCray, A. Kurdoghlian, and C. Nguyen, GaN/
AlGaN HEMT’s operating at 20 GHz with continuouswave
1766 GALLIUM NITRIDE FOR ELECTRONICS
power density 46 W/mm, Electron Lett. 37: 528–530 (April
2001).
37. C. Lee, P. Saunier, J. Yang, and M. A. Khan, AlGaN–GaN
HEMTs on SiC with CW power performance of 44 W/mm
and 23% PAE at 35GHz, IEEE Electron Device Lett. 24 (10):
616–618 (Oct. 2003).
38. R. S. Schwindt, V. Kumar, A. Kuliev, G. Simin, J. W. Yang, M.
A. Khan, M. E. Muir, and I. Adesida, Millimeterwave high
power 0.25mm gatelength AlGaN/GaN HEMTs on SiC sub
strates, IEEE Microwave Wireless Compon. Lett. 13 (3): 93–95
(March 2003).
39. A. T. Ping, E. Piner, J. Redwing, M. A. Khan, and I. Adesida,
Microwave noise performance of AlGaN/GaN HEMTs, Elec
tron Lett. 36: 175–176 (2000).
40. W. Lu, J. Yang, M. A. Khan, and I. Adesida, AlGaN/GaN
HEMTs on SiC with over 100GHz f
T
and low microwave
noise, IEEE Trans. Electron Devices 48 (3): 581–585 (March
2001).
41. I. P. Smorchkova, M. Wojtowicz, R. Sandhu, R. Tsai, M. Bar
sky, C. Namba, P.S. Liu, R. Dia, M. Truong, D. Ko, J. Wang, H.
Wang, and A. Khan, AlGaN/GaN HEMTsoperation in the K
band and above, IEEE Trans. Microwave Theory Tech. 51 (2):
665–668 (Feb. 2003).
42. W. Lu, V. Kumar, E. L. Piner, and I. Adesida, DC, RF,
and microwave noise performances of AlGaNGaN
ﬁeld effect transistors dependence of aluminum concentra
tion, IEEE Trans. Electron Devices 50 (4): 1069–1074 (April
2003).
43. J.W. Lee, V. Kumar, R. Schwindt, A. Kuliev, R. Birkhahn, D.
Gotthold, S. Guo, B. Albert, and I. Adesida, Microwave
noise performances of AlGaN/GaN HEMTs on semiinsulat
ing 6HSiC substrates, Electron Lett. 40 (1): 80–81 (Jan.
2004).
44. J. Wenger, Quartermicrometer lownoise pseudomorphic
GaAs HEMT’s with extremely low dependence of the noise
ﬁgure on drainsource current, IEEE Electron Device Lett. 14
(1): 16–18 (Jan. 1993).
45. B. M. Green, V. Tilak, S. Lee, H. Kim, J.A. Smart, K. J. Webb,
J. R. Shealy, and L. F. Eastman, Highpower broadband Al
GaN/GaN HEMT MMICs on SiC substrates, IEEE Trans. Mi
crowave Theory Tech. 49 (12): 2486–2493 (Dec. 2001).
46. J. J. Xu, S. Keller, G. Parish, S. Heikman, U. K. Mishra, and R.
A. York, A 310GHz GaNbased ﬂipchip integrated broad
band power ampliﬁer, IEEE Trans. Microwave Theory Tech.
48 (12): 2573–2578 (Dec. 2000).
47. J. W. Palmour, S. T. Sheppard, R. P. Smith, S. T. Allen, W. L.
Pribble, T. J. Smith, Z. Ring, J. J. Sumakeris, A. W. Saxler, and
J. W. Milligan, Wide bandgap semiconductor devices and
MMICs for RF power applications, IEDM Tech. Digest, Wash
ington, DC, Dec. 2001, pp. 385–388.
48. V. Paidi, S. Xie, R. Cofﬁe, B. Moran, S. Heikman, S. Keller, A.
Chini, S. P. DenBaars, U. K. Mishra, S. Long, and M. J. W.
Rodwell, High linearity and high efﬁciency of ClassB power
ampliﬁers in GaN HEMT technology, IEEE Trans. Microwave
Theory Tech. 51 (2): 643–652 (Feb. 2003).
49. J.W. Lee, L. F. Eastman, and K. J. Webb, A gallium nitride
pushpull microwave power ampliﬁer, IEEE Trans. Micro
wave Theory Tech. 51 (11): 2243–2249 (Nov. 2003).
50. Y.F. Wu, P.M. Chavarkar, M. Moore, P. Parikh, B. P. Keller,
and U. K. Mishra, A 50W AlGaN/GaN HEMT ampliﬁer,
IEDM Tech. Digest, 2000, pp. 375–376.
51. E. L. Ginzton, W. R. Hewlett, J. H. Jasberg, and J. D. Noe,
Distributed ampliﬁcation, Proc. IRE 36: 956–969 (Aug.
1948).
GENERALIZED SCATTERING MATRIX
TECHNIQUE
AGOSTINO MONORCHIO
University of Pisa
Pisa, Italy
RAJ MITTRA
Pennsylvania State University
University Park, Pennsylvania
GIULIANO MANARA
University of Pisa
Pisa, Italy
1. INTRODUCTION
Many complex microwave devices can usually be decom
posed into a set of elementary waveguide junctions (cano
nical step discontinuities) that are located at speciﬁc
transverse reference planes where the waveguide cross
section changes abruptly, and are connected through long
itudinally uniform waveguide sections. In this context, the
modematching (MM) technique is capable of providing
very accurate results, provided that adequate numbers of
higherorder modes—including the evanescent ones—are
employed for representation of the ﬁeld within each
elementary waveguiding structure [3,5]. In this approach,
the electromagnetic ﬁelds within the individual longitud
inally uniform waveguide sections are represented as a
weighted summation of elementary solutions to Maxwell’s
equations, and the weight coefﬁcients are obtained by
matching the tangential ﬁelds at the interfaces of the
sections where the structure is discontinuous. A micro
wave device can be accurately analyzed by using this
method, once the equivalent circuits of each of the con
stituent discontinuities have been obtained [4]. However,
a simple characterization of the discontinuities in terms of
their conventional fundamental mode scattering para
meters (S parameters) cannot recover the fullwave beha
vior of the device when the interactions between the
discontinuities are not negligible—which is a common
occurrence. In this event, it becomes essential to include
the higherorder modes, generalize the concept of scatter
ing parameters [1], and develop a generalized scattering
matrix, which is the subject of discussion in this article.
The article is organized as follows. In Section 2, we
review the basic concepts of guidedwave propagation in
order to derive a modal representation of the ﬁelds. This
expansion is represented by an equivalent circuit, which
leads to the deﬁnition of the generalized admittance
matrix (GAM) and the generalized impedance matrix.
The modal representation is then rearranged in terms of
the incident and reﬂected waves in order to derive the
deﬁnition of the generalized scattering matrix (GSM). The
next step is to consider the canonical problem of a step
discontinuity between two waveguides with dissimilar
cross sections, and use this simple canonical problem as
the building block for the analysis of more complex
geometries constituting multiple cascaded discontinuities,
such as ﬁlters, matching networks, and directional
GENERALIZED SCATTERING MATRIX TECHNIQUE 1767
couplers. Finally, we present some numerical examples to
illustrate the potentials of the GSM technique.
2. MODAL EXPANSION
We begin our discussion of modal expansion by considering
the geometry of the problem shown in Fig. 1, speciﬁcally, a
hollow, lossless, metallic waveguide, with a uniform cross
section and perfectly electric conductor (PEC) type of
walls. The structure is uniform along z, which is the
direction of propagation in the waveguide. We assume
that the guide is ﬁlled with a homogeneous and lossless
dielectric material, with dielectric permittivity e and mag
netic permeability m.
The solution of the propagation problem is obtained by
referring to the source–free Maxwell’s equations:
VE= ÷jomH
VH=joeE
(1)
The differential equations (1) are a system of six scalar
equations with six unknowns. Since the cross section of the
waveguide is uniform along the z axis, it is advantageous
to split the ﬁeld solution into two parts—the longitudinal
components (E
z
, H
z
) and the transverse components
(E
t
and H
t
), as follows:
E=E
t
÷E
z
i
z
H=H
t
÷H
z
i
z
(2)
The longitudinal and transverse components are interre
lated, of course, through Maxwell’s equations, and they
can be easily derived from each other. This assumption
results in a simpliﬁcation, since we can derive the solution
for the only transverse components of the ﬁelds, and take
advantage of the fact that the longitudinal components
are dependent on the latter and vice versa [2].
It can be shown that for the problem at hand, Maxwell’s
equations can be cast in the form of an eigenvalue problem
whose solutions (eigenfunctions) represent a set of ele
mentary conﬁgurations (modes) supported by the struc
ture. They constitute a complete set of basis functions
for representing the ﬁelds in the waveguide. Therefore,
we can express the transverse ﬁeld components as
summations of modes with complex amplitudes as
E
t
(t
1
; t
2
; z) =
o
m=1
E
t;m
(t
1
; t
2
; z)
H
t
t
1
; t
2
; z ( ) =
o
m=1
H
t;m
(t
1
; t
2
; z)
(3)
where (t
1
,t
2
) represent the coordinates in the transverse
plane as shown in Fig. 1 (in rectangular coordinates t
1
=x,
t
2
=y).
The separation into longitudinal and transverse com
ponents enables us to deﬁne three types of modes: (i)
transverse electric (TE), for which we have E
z
=0 (in this
case the electric ﬁeld is totally transverse: E=E
t
); (ii)
transverse magnetic (TM), for which we have H
z
=0, (i.e.,
the magnetic ﬁeld is totally transverse: H=H
t
); and (iii)
transverse electric and magnetic (TEM), for which we
have both E
z
and H
z
=0.
By invoking the principle of separation of variables, we
can represent the electric and the magnetic ﬁeld compo
nents of each mode as a product of a scalar function of
z and a vector function of the transverse coordinates that
read
E
t;m
(t
1
; t
2
; z) =V
m
(z) e
m
(t
1
; t
2
)
H
t;m
(t
1
; t
2
; z) =I
m
(z) h
m
(t
1
; t
2
)
(4)
Following the wellestablished terminology for waveguid
ing structures, we refer to the functions e
m
and h
m
as the
modal vector functions, and to V
m
(z) and I
m
(z) as the
modal scalar functions. The latter satisfy the transmis
sionline equations (telegraphist’s equations), with the
standard solution:
V
m
(z) =V
÷
m
e
÷jb
m
z
÷V
÷
m
e
÷jb
m
z
(5a)
I
m
(z) =I
÷
m
e
÷jb
m
z
÷I
÷
m
e
÷jb
m
z
(5b)
The modal scalar functions can be interpreted as the
equivalent voltages and currents associated with the
mth mode of the multimode line—the waveguide.
The corresponding modal characteristic impedance for
the mth mode is given by
Z
0;m
=
V
÷
m
I
÷
m
= ÷
V
÷
m
I
÷
m
(6)
and these impedances can be chosen such that they
normalize the corresponding modal vector functions. In
particular, if we choose:
Z
0;m
=z
0;m
=
om=b
m
for TEmodes
b
m
=oe for TMmodes
_
(7)
the modal vector functions are related as follows:
e
m
=h
m
i
z
(8)
z
t
1
t
2
Figure 1. Geometry of a waveguide problem.
1768 GENERALIZED SCATTERING MATRIX TECHNIQUE
The modal vector functions are the solutions of an eigen
value problem; in particular, we can introduce a potential
function j
m
, which satisﬁes the twodimensional eigenva
lue equation
V
2
t
j
m
(t
1
; t
2
) ÷k
2
c;m
j
m
(t
1
; t
2
) =0 (9)
with the following boundary conditions on the contour C of S
j
m
¸
¸
C
=0; for TEmodes (10a)
@j
m
@n
¸
¸
¸
¸
C
=0; for TMmodes (10b)
where @/@n is the derivative along the outward normal to
the contour C. In (9), k
c,m
is the transverse eigenvalue of
the mth mode. The value of the transverse eigenvalue
(and, consequently, the ﬁnal form of the solution j
m
)
depends on the geometric shape of the waveguide cross
section. Once we have obtained the solution for j
m
, the
corresponding vector functions can be evaluated from
e
m
=
i
z
V
t
j
m
for TEmodes
÷V
t
j
m
for TMmodes
_
(11)
The wavenumber along the z axis for the mth mode is
given by
k
z
=k
z
TE
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
o
2
em ÷k
2
t
H
_
(12)
An examination of this equation leads to some useful
interpretations of the propagation properties of the wave
guide. Speciﬁcally, we note that the propagation constant
b
m
in (12) goes to zero at the cutoff frequencies given by
o
c;m
=
k
c;m
ﬃﬃﬃﬃﬃ
me
_ (13)
For frequencies below the cutoff, the wavenumber is
purely imaginary and the corresponding mode is evanes
cent, with an attenuation constant given by
jb
m
=a
m
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
k
2
c;m
÷o
2
em
_
(14)
A TEM mode can be interpreted as a limiting case of a TE
or TM mode, with a null transverse eigenvalue k
c
. It exists
only if the guiding structure consists of at least two
parallel electric conductors (i.e., two separated metallic
regions on the guiding structure transverse plane). This
mode has no cutoff and its propagation constant is iden
tical to that relevant to a uniform plane wave propagating
in an unbounded homogeneous medium, with the same
electrical characteristics as those of the TEM mode:
k
z
=k
0
=o
ﬃﬃﬃﬃﬃ
em
_
(15)
In contrast to this, the characteristic impedance of the
TEM mode depends on both the electric properties of the
medium surrounding the guiding structure and the geo
metric parameters of the transverse plane.
3. CIRCUIT REPRESENTATIONS OF THE DISCONTINUITIES
In this section, we introduce a formalism that leads to an
Nport circuit representation of a waveguiding system [4].
We start by considering a generic discontinuity, with
respect to which we deﬁne two reference transverse
planes T
1
and T
2
(see Fig. 2), where the electric and
magnetic ﬁelds can be expressed in terms of the modal
expansions we have previously introduced. Although, in
principle, the modal expansion contains an inﬁnite num
ber of terms, from a practical point of view it is sufﬁcient to
truncate it to a ﬁnite value. For convenience, we will refer
to each of the reference planes as ‘‘physical’’ ports. If we
retain N
i
modes at the ith ‘‘physical’’ port (i =1,2) that are
sufﬁcient to describe the discontinuity, then we need to
know N
1
÷N
2
modal voltages and currents, implying that
an equal number of equations relating the voltages and
currents are required, which can be obtained from the
Maxwell’s equations. Furthermore, by virtue of the un
iqueness theorem, it is sufﬁcient to consider only the
tangential components at the interfaces (ﬁelds transverse
to z) at both the ‘‘physical’’ ports. The discontinuity can
therefore be represented by using its multiport equivalent
circuit, shown in Fig. 3, where each port is associated with
a different mode (‘‘modal’’ port). Such a network formalism
enables us to deal with the voltages and currents—in the
same manner as in circuit analysis—instead of working
with the electric and magnetic ﬁelds.
For the multiport circuit introduced previously, the set
of equations relating the voltages and currents can be cast
in a more convenient form by expressing all the modal
currents as functions of the corresponding voltages. More
speciﬁcally, we can obtain the following matrix form:
I =Y V (16)
I
(1)
1
.
.
.
I
(1)
N
1
I
(2)
1
.
.
.
I
(2)
N
2
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
=
Y
11
Y
12
Y
21
Y
22
_ _
=
V
(1)
1
.
.
.
V
(1)
N
1
V
(2)
1
.
.
.
V
(2)
N
2
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
(17)
T
2
T
1
(2)
(1)
Figure 2. Geometry for representing a waveguide discontinuity.
GENERALIZED SCATTERING MATRIX TECHNIQUE 1769
It is important to note that the voltages and currents in
these equations are related by the admittance matrix Y,
which now includes both the evanescent and propagating
modes. Consequently, the admittance matrix (17) is more
general than the conventional one, which relates only the
currents and voltages of the propagating modes. Such a
generalization leads to the concept and the deﬁnition of a
generalized admittance matrix (GAM). In addition, we can
divide the modes into two classes: accessible (propagating)
and localized (evanescent). If we are interested only in the
propagating modes (typically the fundamental modes) at
the input and output ‘‘physical’’ ports, then the corre
sponding admittance matrix has the dimensions of (22),
and it can be extracted from the corresponding GAM
whose dimensions are (N
1
÷N
2
) (N
1
÷N
2
). To this end,
we consider the equivalent circuit shown in Fig. 4, where
all the ‘‘modal’’ ports corresponding to the evanescent
modes are matched to their respective characteristic im
pedances, so that
I
(i)
m
= ÷Y
(i)
0;m
V
(i)
m
m=2; . . . ; N
i
(18)
with
Y
(i)
0;m
=1=Z
(i)
0;m
(19)
We can now rearrange and partition the admittance
matrix with the objective of separating the entries per
taining to the two fundamental modes (subscript ‘‘f’’) from
to the higherorder evanescent modes (subscript ‘‘h’’) as
follows:
Y =
Y
ff
Y
fh
Y
hf
Y
hh
_ _
(20)
The admittance matrix Y
/
of the discontinuity can be
expressed as
Y
/
=Y
ff
÷Y
fh
(Y
hh
÷Y
0
)
÷1
Y
hf
(21)
where Y
0
is a diagonal matrix containing the character
istic modal admittances:
Y
0
=diag Y
(1)
0;2
; . . . ; Y
(1)
0;N
1
; Y
(2)
0;2
; . . . ; Y
(2)
0;N
2
_ _
(22)
It is important to realize that the higherorder modes,
although inaccessible, nevertheless inﬂuence the behavior
(and the ﬁnal value of the complex amplitude) of the
fundamental modes. We will touch on this aspect again
in Section 9 and present some examples to validate the
preceding statement.
An alternative description can be obtained by expres
sing the modal voltages as functions of the modal currents,
so that we can obtain an impedance matrix description as
follows
V =ZI (23)
where Z is deﬁned as the generalized impedance matrix.
In microwave applications, a more useful representation
of the electromagnetic ﬁeld is obtained by working with
the incident and reﬂected wavevectors of the discontinu
ity, and it is convenient to derive a description that is not
tied to the particular type of transmission line feeding the
network. Toward this end, we deﬁne a normalized modal
voltage of the mth mode at the ith port as follows:
^
VV
(i)
m
=
V
(i)
m
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(i)
0;m
_ (24)
+
−
+
−
+
−
+
−
V
1
(1)
V
1
(2)
I
N
2
(2)
V
N
2
(2)
V
N
1
(1)
I
N
1
(1)
I
1
(1)
I
1
(2)
Multiport
circuit
Figure 3. Multiport equivalent circuit to the discontinuity
depicted in Fig. 2.
+
−
+
−
V
1
(1)
V
1
(2)
Y
0
2
(1)
Y
0
2
(2)
Y
0
N
2
(2)
Y
0
N
1
(1)
I
1
(1)
I
1
(2)
Multiport
circuit
Figure 4. Multiport circuit representation of a waveguide
discontinuity for deriving the standard admittance matrix.
1770 GENERALIZED SCATTERING MATRIX TECHNIQUE
Similarly, we can deﬁne a normalized modal current of the
mth mode at the ith port as
^
II
(i)
m
=I
(i)
m
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(i)
0;m
_
(25)
The corresponding voltage and current modal vectors
(normalized) are
^
VV =[
^
VV
(i)
m
] (26a)
^
II =[
^
II
(i)
m
] (26b)
These vectors are related by the normalized impedance
matrix
^
ZZ as
^
VV =
^
ZZ
^
II (27)
where the entries of the impedance matrix given above are
deﬁned as follows:
^
ZZ
(i;j)
mn
=
Z
(i;j)
mn
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(i)
0;m
_ ﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(j)
0;n
_ (28)
The expression in (28) can be interpreted as the normal
ized mutual impedance between the mth ‘‘modal’’ port at
the ‘‘physical’’ port i and the nth ‘‘modal’’ port at the
‘‘physical port’’ j.
A similar normalization can also be applied to the
GAM, which can be written as
^
II =
^
YY
^
VV (29)
The entries of the matrix above are given by
^
YY
(i;j)
mn
=
Y
(i;j)
m;n
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Y
(i)
0;m
_ ﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Y
(j)
0;n
_ (30)
and they can also be interpreted in a procedure similar to
that followed for the impedance matrix elements.
We can now deﬁne, at each port, the incident wave
amplitude of the mth mode at ith terminal as
a
(i)
m
=
V
÷
(i)
m
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(i)
0;m
_ =I
÷
(i)
m
ﬃﬃﬃﬃ
Z
_
=
V
(i)
m
÷Z
(i)
0;m
I
(i)
m
2
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
0;m
_ (31)
Similarly, the corresponding reﬂected wave amplitude can
also be deﬁned as
b
(i)
m
=
V
÷
(i)
m
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(i)
0;m
_ = ÷I
÷
(i)
m
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(i)
0;m
_
=
V
(i)
m
÷Z
(i)
0;m
I
(i)
m
2
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(i)
0;m
_ (32)
Finally, using these deﬁnitions, we can fully characterize
the device in terms of its generalized scattering matrix
(GSM) S, which relates the vectors representing the
reﬂected wave and the incident wave amplitudes. The
matrix is deﬁned as
b =Sa (33)
This description in terms of the GSM is equivalent to that
in terms of the impedance or admittance matrices. In fact,
starting from the following equations
^
VV =a÷b (34a)
^
II =a ÷b (34b)
we can also write
(a÷b) =
^
ZZ(a ÷b) (35)
Using (33), we obtain the ﬁnal result
S=(
^
ZZ÷1)
÷1
(
^
ZZ ÷1) =(1 ÷
^
YY)
÷1
(1 ÷
^
YY) (36)
where 1 is the unit matrix with dimensions (N
1
÷N
2
)
(N
1
÷N
2
). From knowledge of the GSM of a discontinuity,
we can obtain the corresponding admittance and impe
dance matrices as follows:
^
YY =
^
ZZ
÷1
=(S÷1)
÷1
(1 ÷S) (37)
^
ZZ=
^
YY
÷1
=(S÷1)(1 ÷S)
÷1
(38)
It is important to recognize that, for linear networks, the
matrix S can always be deﬁned, while in some particular
situations the matrices Z and Y cannot. Moreover, a GSM
can always be deﬁned, even if the value of the character
istic impedance of the port is not available.
4. STEP JUNCTION
In this section, we consider the discontinuity problem of a
junction between two waveguides with different cross
sections S
1
and S
2
. The junction is located at the plane z
=0. Let S
C
be the common interface, shared by the two
waveguides (S
C
=S
1
¨ S
2
) and let S
P
be the metallic
surface at the interface (for the case shown in Fig. 5,
associated with the boundary enlargement).
The electromagnetic ﬁelds in both the waveguides can
be expressed in terms of their respective modes. In
particular, by using (4) and (5), we can express the
transverse components of the electromagnetic ﬁeld in
the ﬁrst waveguides (on the left) as a ﬁnite summation
GENERALIZED SCATTERING MATRIX TECHNIQUE 1771
of modes as follows:
E
(1)
t
=
N
1
m=1
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(1)
0;m
_
a
(1)
m
e
÷jb
m
z
_
÷
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(1)
0;m
_
b
(1)
m
e
÷jb
m
z
_
e
(1)
m
(t)
H
(1)
t
=
N
1
m=1
a
(1)
m
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(1)
0;m
_ e
÷jb
m
z
_
_
_
÷
b
(1)
m
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(1)
0;m
_ e
÷jb
m
z
_
_
_h
(1)
m
(t)
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
(39)
Similarly, in the second waveguide (on the right) we have
E
(2)
t
=
N
2
n=1
ﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(2)
0;n
_
a
(2)
n
e
÷jb
n
z
_
÷
ﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(2)
0;n
_
b
(2)
n
e
÷jb
n
z
_
e
(2)
n
(t)
H
(2)
t
=
N
2
n=1
a
(2)
n
ﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(2)
0;n
_ e
÷jb
n
z
_
_
_
÷
b
(2)
n
ﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(2)
0;n
_ e
÷jb
n
z
_
_
_h
(2)
n
(t)
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
(40)
In these expressions, the wave amplitudes of the modes
are the unknowns in this problem that are yet to be
determined. As explained in the previous section, a com
plete characterization of the discontinuity can be obtained
if we can relate the (N
1
÷N
2
) amplitudes of the a
(i)
incident waves to the corresponding amplitudes of the
reﬂected waves, b
(i)
(i =1,2). We can derive these equa
tions by imposing the boundary conditions at the junction,
speciﬁcally, the continuity of the transverse components of
both the electric and magnetic ﬁelds as follows:
E
(2)
t
=E
(1)
t
at S
C
E
(2)
t
=0 at S
P
=S
2
¸S
1
_
¸
_
¸
_
H
(2)
t
=H
(1)
t
at S
C
(41)
The simplest way to derive the desired (N
1
÷N
2
) equa
tions is to impose the continuity condition at (N
1
÷N
2
)
different points. However, such a pointmatching techni
que, although simple from a conceptual point of view, is
not very accurate since it imposes the continuity condition
only at a ﬁnite set of points, and it is better to impose the
condition stated above in an integral sense instead. With
this in mind, we introduce a deﬁnition of the inner
product, which must be computed in the process.
5. GSM OF A STEP JUNCTION
In this section, we derive a GSM representation of the step
junction, shown in Fig. 5, by using the formulation based
on the conservation of complex power [6,7]. However, it
can be shown that any relation originating from this
formulation is equivalent to the ones derived by matching
the transverse electric ﬁelds or the transverse magnetic
ﬁelds at the junction plane. We begin by expressing the
transverse electric and magnetic ﬁelds at z =0 as
E
(1)
t
(z =0) =
N
1
m=1
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(1)
0;m
_
(a
(1)
m
÷b
(1)
m
) e
(1)
m
(t)
H
(1)
t
(z =0) =
N
1
m=1
1
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(1)
0;m
_ (a
(1)
m
÷b
(1)
m
) h
(1)
m
(t)
_
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
_
(42)
E
(2)
t
(z =0) =
N
2
n=1
ﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(2)
0;n
_
(a
(2)
n
÷b
(2)
n
) e
(2)
n
(t)
H
(2)
t
(z =0) =
N
2
n=1
1
ﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(2)
0;n
_ (a
(2)
n
÷b
(2)
n
) h
(2)
n
(t)
_
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
_
(43)
From the continuity conditions, we have
E
(1)
t
(z =0) =E
(2)
t
(z =0)
=
N
1
m=1
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(1)
0;m
_
(a
(1)
m
÷b
(1)
m
) e
(1)
m
(t)
=
N
2
n=1
ﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(2)
0;n
_
(a
(2)
n
÷b
(2)
n
) e
(2)
n
(t)
(44)
H
(1)
t
(z =0) =H
(2)
t
(z =0)
=
N
1
m=1
1
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(1)
0;m
_ (a
(1)
m
÷b
(1)
m
) h
(1)
m
(t)
=
N
2
n=1
1
ﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(2)
0;n
_ (a
(2)
n
÷b
(2)
n
) h
(2)
n
(t)
(45)
a
(2)
b
(2)
a
(1)
b
(1)
S
1
S
C
S
2
z = 0
Figure 5. Boundary enlargement problem.
1772 GENERALIZED SCATTERING MATRIX TECHNIQUE
Next, we deﬁne an inner product between the two func
tions f
1
and f
2
as
¸f
1
; f
2
) =
_
S
C
f
1
(t) f
+
2
(t)
.
^ zz dS (46)
and use this deﬁnition to project the continuity equation of
the electric (magnetic) ﬁeld onto each magnetic (electric)
modal vector function and the continuity equation. Such a
projection guarantees that the continuity is ensured not
only at a discrete set of points but also across the junction
in an integral sense. In particular, by projecting the ﬁrst
equation onto the eigenfunctions h
(1)
+
m
, we have
N
1
m=1
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(1)
0;m
_
(a
(1)
m
÷b
(1)
m
)
_
S
1
(e
(1)
m
h
(1)
+
m
)
.
dS
=
N
2
n=1
ﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(2)
0;n
_
(a
(2)
n
÷b
(2)
n
)
_
S
1
(e
(2)
n
h
(1)
+
m
)
.
dS
(47)
Similarly, by projecting the second equation onto the
eigenfunctions e
(2)
+
n
, we obtain
N
1
m=1
1
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(1)
0;m
_ (a
(1)
m
÷b
(1)
m
)
_
S
2
(h
(1)
m
e
(2)
+
n
)
.
dS
=
N
2
n=1
1
ﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(2)
0;n
_ (a
(2)
n
÷b
(2)
n
)
_
S
1
(h
(2)
n
e
(2)
+
n
)
.
dS
(48)
By exploiting the orthogonality of the modes at each of the
waveguide sections, we can express (47) and (48) in matrix
forms as
diag
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(1)
0;m
_
_
S
1
(e
(1)
m
h
(1)
+
m
)
.
dS
_ _
(a
(1)
÷b
(1)
) =A(a
(2)
÷b
(2)
)
(49)
B(a
(1)
÷b
(1)
) =diag
1
ﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(2)
0;n
_
_
S
2
(h
(2)
n
e
(2)
+
n
)
.
dS
_
¸
_
¸
_
_
¸
_
¸
_
(a
(2)
÷b
(2)
)
(50)
The elements of the matrices A(N
1
N
2
) and B(N
2
N
1
)
can be written as
a
mn
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(2)
0;n
_
_
S
1
(e
(2)
n
h
(1)
+
m
)
.
dS (51)
b
mn
=
1
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
(1)
0;m
_
_
S
2
(h
(1)
m
e
(2)
+
n
)
.
dS (52)
These matrices have some special properties, and, in order
to understand them, we resort to the normalization pro
cedure outlined in Section 2, for which the characteristic
impedance of the equivalent transmission line assumes a
unit value:
Z
0;m
¬ 1 (53)
As a consequence, for the corresponding vector functions
of each mode, we can write
e
m
=z
m
h
m
i
z
; h
m
=
1
z
m
i
z
e
m
(54)
so that
a
mn
=
1
z
(1)
+
0;m
_
S
C
(e
(2)
n
.
e
(1)
+
m
)
.
dS (55)
b
mn
=
1
z
(1)
0;m
_
S
C
(e
(1)
m
.
e
(2)
+
n
)
.
dS=a
+
nm
(56)
This implies that the following relationship holds:
B=(A
T
)
+
(57)
Thus the matrix B is the conjugate transpose of A. It is
worth noting that with a correct choice of the normalization
impedance, some interesting and important properties
regarding the ﬁnal form of the GSM can be observed [11].
By deﬁning the following diagonal matrices (power
matrices)
P=diag
1
z
(1)
+
0;m
_
S
1
e
(1)
m
¸
¸
¸
¸
2
.
dS
_ _
=diag z
(1)
0;m
_
S
1
h
(1)
m
¸
¸
¸
¸
¸
¸
2
.
dS
_ _
(N
1
N
1
)
(58)
Q=diag
1
z
(2)
0;n
_
S
2
e
(2)
n
¸
¸
¸
¸
2
.
dS
_ _
=diag z
(1)
+
0;n
_
S
1
h
(2)
n
¸
¸
¸
¸
¸
¸
2
.
dS
_ _
(N
2
N
2
)
(59)
Eqs. (49) and (50) can be recast in the following form:
P(a
(1)
÷b
(1)
) =A(a
(2)
÷b
(2)
) (60)
(A
T
)
+
(a
(1)
÷b
(1)
) =Q(a
(2)
÷b
(2)
) (61)
We can now deﬁne the matrices below
M
E
=P
÷1
Awhose dimensions are
(N
1
N
2
)
(62)
M
H
=Q
÷1
(A
T
)
+
whose dimensions are
(N
2
N
1
)
(63)
GENERALIZED SCATTERING MATRIX TECHNIQUE 1773
to derive the ﬁnal matrix representations:
(a
(1)
÷b
(1)
) =M
E
(a
(2)
÷b
(2)
) (64)
M
H
(a
(1)
÷b
(1)
) =(a
(2)
÷b
(2)
)
From (64) we can derive the generalized scattering matrix
that relates the vectors b and a
b
(1)
b
(2)
_ _
=S
a
(1)
a
(2)
_ _
=
S
11
S
12
S
21
S
22
_ _
a
(1)
a
(2)
_ _
(65)
where
S
11
=(1÷M
H
.
M
E
)
÷1
.
(1 ÷M
H
.
M
E
)
S
21
=M
E
.
(1 ÷S
11
)
S
12
=2(1÷M
H
.
M
E
)
÷1
.
M
H
S
22
=(M
E
.
S
12
÷1)
_
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
_
(66)
are the self and mutual blocks of the GSM.
It should be realized that the computation of the
scattering matrix carried out in this manner requires
the inversion of an (N
1
N
1
) matrix for each frequency.
Finally, the scattering parameters of the discontinuity can
be ﬁnally obtained by identifying the ﬁrst element of each
of the submatrices (see Fig. 6) that are
s
11
=S
11
(1; 1); s
12
=S
12
(1; 1);
s
21
=S
21
(1; 1); s
22
=S
22
(1; 1)
(67)
For a discussion of the properties of the GSM, the reader is
referred to Refs. 8–10.
6. CASCADING DISCONTINUITIES
A step discontinuity can be viewed as a basic building
block for describing more complex conﬁgurations. We now
consider, as a simple example, the geometry sketched in
Fig. 7, where two such discontinuities are separated by a
distance L.
The GSM of the uniform waveguide section spanning
from z =0 and z =L simply accounts for the phase shift
(and/or attenuation) of each mode, as it traverses this
section of the guide. This GSM can be written as
S=
0 D
D 0
_ _
(68)
where
D=diag e
÷jb
n
L
_ _
(69)
The propagating modes only suffer a phase shift, while the
evanescent ones are attenuated, since b
n
is purely ima
ginary for the latter. If the distance L is sufﬁciently large,
the evanescent modes originating from the ﬁrst disconti
nuity are completely attenuated and do not interact with
the second discontinuity; hence they can be neglected.
It is possible to express the wave amplitudes at the
second interface in terms of the S matrix deﬁned in (8)
a
(2)
a
(2)
L
_ _
=S
b
(2)
b
(2)
L
_ _
=
0 D
D 0
_ _
b
(2)
b
(2)
L
_ _
(70)
and
a
(2)
L
=Db
(2)
; a
(2)
=Db
(2)
L
=b
(2)
L
=D
÷1
a
(2)
(71)
It is important to note that, in deriving (71), the outgoing
waves, transmitted to the right of the ﬁrst discontinuity,
are interpreted as waves incident on the waveguide sec
tion located between z =0 and z =L. By inserting (71) into
(70), we obtain the GSM of the entire conﬁguration,
formed by the two discontinuities ﬂanking the waveguide
N
1
N
1
N
2
N
2
s
12
s
11
s
21
s
22
S
12
S
11
S
21
S
22
GSM
Figure 6. Identiﬁcation of the scattering parameters of the step
discontinuity as entries of the GSM.
a
(2)
b
(2)
a
(3)
b
(3)
a
(1)
b
(1)
z=L z=0
(2)
a
L
(2)
b
L
Figure 7. Geometry for analysis of a doublestep discontinuity.
1774 GENERALIZED SCATTERING MATRIX TECHNIQUE
section in between. We can generalize the analysis a bit
further and derive the expression for the GSM matrix
when we extend the lengths of the waveguide sections at
both ends of the discontinuity (see Fig. 8). For this case,
the GSM can be expressed as
S
/
=fSf=
D
1
0
0 D
2
_ _
S
D
1
0
0 D
2
_ _
(72)
where
D
1
=diag e
÷jb
(1)
m
l
1
_ _
D
2
=diag e
÷jb
(2)
n
l
2
_ _
(73)
To derive the composite solution of this extended geometry
case, we need certain connection formulas, which we
provide below [11,12]. Given two structures, each of which
is characterized by its own GSM S
(1)
and S
(2)
, respectively,
we can construct a composite GSM by using the following
submatrix blocks [3]
S
(tot)
11
=S
(1)
11
÷S
(1)
12
S
(2)
11
ES
(1)
21
S
(tot)
12
=S
(1)
12
.
_
1 ÷S
(2)
12
ES
(1)
22
_
S
(2)
12
S
(tot)
21
=S
(2)
21
ES
(1)
21
S
(tot)
22
=S
(2)
22
÷S
(2)
21
ES
(1)
22
S
(2)
12
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
(74)
where
E=
_
1 ÷S
(1)
22
S
(2)
11
_
÷1
(75)
Once again, an inversion is needed to construct the matrix
E in (75), which appears in the expressions given in (74)
for the various blocks.
It is important to observe that the expressions in Eqs.
(74) and (75) originate from an iterative process that
accounts for multiple reﬂections at both discontinuities.
In particular, the modes inside the central section go
through a multiple scattering phenomenon that can be
mathematically expressed as a Neumann series with ﬁnal
value given by (74). For an indepth discussion of this
phenomenon, the reader is referred to Ref. 1, where a
proof of convergence has been provided.
A more intuitive approach might have been followed,
which makes use of an alternative representation in terms
of the transmission matrix. It relates the wave amplitudes
(both incoming and outgoing) at the ﬁrst port with those at
the second port as follows:
a
(1)
b
(1)
_ _
=T
b
(2)
a
(2)
_ _
=
T
11
T
12
T
21
T
22
_ _
b
(2)
a
(2)
_ _
(76)
For the cascaded junction, the total transmission matrix
can be simply expressed as a product of the single trans
mission matrices:
T
tot
=T
(1)
T
(2)
(77)
However, there are two signiﬁcant disadvantages when
using this approach that we can easily identify. The ﬁrst
one arises from the fact that the wavenumber in the
uniform waveguide section between the two step disconti
nuities exhibits a negative sign with respect to the pre
vious representation, implying that the cutoff modes are
mapped into the constituent block matrices with real and
positive exponentials. This renders the entries of the ﬁnal
matrix unstable, especially for large n and L. The second
difﬁculty stems from the requirement that we must use an
identical number of modes at both ends of the junction,
since the use of a different number of modes results in an
illconditioned matrix. This condition, in turn, suffers from
the problem of relative convergence, which will be dis
cussed in the following section.
7. CONVERGENCE CRITERIA
The accuracy of the ﬁnal results is strongly affected by the
truncation of the series used for the modal expansion.
Convergence tests are usually performed in order to
determine the minimum number of modes required to
obtain a certain degree of accuracy; speciﬁcally, the num
ber of modes is increased until the ﬁnal result does not
change signiﬁcantly.
However, an interesting phenomenon can occur when
following this procedure; namely, the ﬁnal results may
converge to a value that is strongly dependent on the ratio
N
1
/N
2
of the total numbers of modes used in the two
longitudinally uniform waveguide sections directly con
nected to the discontinuity. This phenomenon, known as
‘‘relative convergence,’’ may affect the validity of the ﬁnal
results [13,14].
Several different criteria have been presented in the
literature to cope with this problem. If we consider the
step discontinuity problem reported in Fig. 9, where two
rectangular waveguides of the same width and different
heights are connected, it is possible to show that a correct
result is guaranteed by increasing the number of modes
and maintaining the ratio N
1
/N
2
approximately equal to
a
(2)
b
(2)
a
(1)
b
(1)
z=−l
1
z=−l
2
z=0
Figure 8. Reference systems at a step discontinuity analyzed
through the transmission matrix.
GENERALIZED SCATTERING MATRIX TECHNIQUE 1775
the ratio h
1
/h
2
. Intuitively, this result can be explained by
considering the fact that both modal expansions should be
able to follow the same variability of the ﬁeld at both sides
of discontinuity; since the basis functions used in the
modal expansions have different periodicities on each
side of the junction, the choice of the ratio N
1
/N
2
=h
1
/h
2
leads to the correct behavior of the two ﬁnal expressions.
For a double step discontinuity (along both height and
width), accurate results have been obtained by using the
area criterion N
1
/N
2
=A
1
/A
2
, where A
1
and A
2
are the
crosssectional areas of the waveguides 1 and 2, respec
tively. However, this criterion is not as general as a more
useful one based on the condition that the maximum
values of the transverse eigenvalues for modes at both
sides of the discontinuity be the same. This is known as
the spectral criterion, and it requires that once that the
maximum value of the transverse eigenvalue k
c,max
is
ﬁxed, only the modes with k
c
ok
c,max
are to be retained.
8. NUMERICAL EXAMPLES
Numerical results are presented in this section with the
objective of demonstrating that the accuracy can be im
proved by including higherorder modes. As a ﬁrst exam
ple, we consider a singlestep discontinuity problem at the
junction between a rectangular waveguide (waveguide 1)
with dimensions of 10 5mm, and a smaller rectangular
waveguide (waveguide 2) whose dimensions are
7 3.5mm. The singlemode bandwidth of the ﬁrst wave
guide spans from 15 to 30 GHz, whereas the second
waveguide exhibits a singlemode bandwidth between
21.42 and 42.84 GHz. The scattering parameter S
11
of
this discontinuity has been calculated by using only the
fundamental mode in both the waveguides. These results
have been compared with those obtained by calculating
the whole GSM of the discontinuity with 71 and 33 modes
in the waveguides, respectively. The results are shown in
Fig. 10, where the net discrepancy between the two curves
clearly demonstrates the importance of including the
higherorder modes.
The effects of higherorder modes become even more
apparent if we refer to the following problem, where a
rectangular iris of ﬁnite thickness is considered. This
h
2
h
1
Figure 9. Step discontinuity between two rectangular wave
guides of different heights.
20 22 24 26 28 30

S
1
1

(
d
B
)
0
−10
−20
−30
−40
−50
−60
Frequency (GHz)
S11 (11)
S11 (3371)
Figure 10. Amplitudes of the S
11
scattering parameters for the
boundary enlargement problem at the junction between two
rectangular waveguides with dimensions 73.5mm and
105 mm, respectively. The dashed curve is based on the exclu
sion of the higher order modes.
20 22 24 26 28 30

S
1
1

(
d
B
)
0
−10
−20
−30
−40
−50
−60
Frequency (GHz)
S11 single mode
S11
Figure 11. Amplitude of the S
11
scattering parameter relevant to
a thick iris in a rectangular waveguide. The waveguide has
dimensions 105mm, whereas the iris has dimensions
7 3.5 mm and a thickness of 2mm. The dashed line corresponds
to the case where the higherorder modes are not included.
20 22 24 26 28 30

S
1
1

(
d
B
)
0
−10
−20
−30
−40
−50
−60
Frequency (GHz)
S11 single mode
S11
Figure 12. Amplitude of the S
11
scattering parameter relevant to
a thick iris in a rectangular waveguide. The waveguide has
dimensions 105mm, whereas the iris has dimensions
7 3.5 mm and a thickness of 5mm. The dashed line corresponds
to the case where the higherorder modes are not included.
1776 GENERALIZED SCATTERING MATRIX TECHNIQUE
problem can be analyzed by following the methodology
presented in Section 6, where we cascade the two discon
tinuity problems and shift the GSM of the ﬁrst disconti
nuity toward the second by a length equal to the thickness
of the iris. In particular, the results obtained by using only
the fundamental modes in the three waveguides sections
are shown in Figs. 11 and 12 for two different lengths of
the iris, and they are compared with those derived for the
same conﬁguration by accounting for the presence of
higherorder modes (more precisely, 71 modes are in
cluded in the ﬁrst and third waveguide sections, while
33 modes are retained in the central waveguide section).
BIBLIOGRAPHY
1. R. Mittra and S. W. Lee, Analytical Techniques in the Theory
of Guided Waves, Macmillan, New York, 1971, pp. 207–217.
2. R. E. Collin, Field Theory of Guided Waves, IEEE Press, New
York, 1991.
3. J. Uher, J. Bornemann, and U. Rosenberg, Waveguide Com
ponents for Antenna Feed Systems: Theory and CAD, Artech
House, 1993.
4. G. Conciauro, M. Guglielmi, and R. Sorrentino, Advanced
Modal Analysis, Wiley, 2000.
5. T. Rozzi and M. Mongiardo, Open Electromagnetic Wave
guides, IEE, London, 1997.
6. R. SafaviNaini and R. H. Macphie, On solving waveguide
junction scattering problems by the conservation of complex
power technique, IEEE Trans. Microwave Theory Tech. MTT
29(4):337–343 (April 1981).
7. S. Omar and K. Schu¨ nemann, Transmission matrix represen
tation of ﬁnline discontinuities, IEEE Trans. Microwave The
ory Tech. MTT33:765–770 (Sept. 1985).
8. G. Eleftheriades, A. Omar, L. Katehi, and G. Rebeiz, Some
important properties of waveguide junction generalized scat
tering matrices in the context of the mode matching techni
que, IEEE Trans. Microwave Theory Tech. 42:1896–1903 (Oct.
1994).
9. M. A. Solano, J. S. Ipin˜ a, A
´
. Go´mez, A. Prieto, and A. Vegas,
Comments on ‘‘Some important properties of waveguide junc
tion generalized scattering matrices in the context of the
mode matching technique,’’ IEEE Trans. Microwave Theory
Tech. 49:1663–1664 (Sept. 2001).
10. A. Morini and T. Rozzi, On the deﬁnition of the generalized
scattering matrix of a lossless multiport, IEEE Trans. Micro
wave Theory Tech. 49:160–165 (Jan. 2001).
11. P. L. Overfelt and D. J. White, Alternate forms of the general
ized composite scattering matrix, IEEE Trans. Microwave
Theory Tech. 37:1267–1268 (Aug. 1989).
12. T. S. Chu and T. Itoh, Generalized scattering matrix method
for analysis of cascaded and offset microstrip step disconti
nuities, IEEE Trans. Microwave Theory Tech. MTT34:280–
284 (Feb. 1986).
13. R. Mittra, T. Itoh, and T. Li, Analytical and numerical studies
of the relative convergence phenomenon arising in the
solution of an integral equation by the moment method,
IEEE Trans. Microwave Theory Tech. MTT20:96–104 (Feb.
1972).
14. S. W. Lee, W. R. Jones, and J. J. Campbell, Convergence
of numerical solutions of iristype discontinuity problems,
IEEE Trans. Microwave Theory Tech. MTT19:528–536
(June 1971).
GEOMETRICAL OPTICS
DANILO ERRICOLO
University of Illinois at Chicago
Chicago, Illinois
1. INTRODUCTION
Geometric optics is the simplest theory to explain the for
mation of images. It represents a useful approximation of
more complex theories and, in practical situations, such as
the design of optical systems, it provides a useful way to
make preliminary assessments. Geometric optics is also
important because of the inﬂuence it has had on other
disciplines, for example, in electrical engineering, on the
design of antennas, which is discussed in Section 13. Geo
metric optics is attractive because it is a simple theory
based on rays. However, it also is limited in its applica
tions because it does not include diffraction, which is im
portant to explain more precisely the formation of images
and usually requires complex mathematical formulations.
Many efforts have been devoted to extend geometric optics
to include diffraction while keeping the advantages of a
simpler ray theory, which are discussed in Section 14.
Let us consider the location of geometric optics within
the realm of optical phenomena. Light is an electromag
netic phenomenon [1], and its propagation may be de
scribed in terms of two mutually coupled vectors: the
electric and magnetic ﬁelds. In practice, many phenome
na may be described by simply using only one of these two
vector quantities. Hence, this simpliﬁcation leads to a sca
lar theory where light is treated as a scalar function and
the resulting theory is called wave optics. When light
propagates around objects that are much larger than its
wavelength, the wave nature of light may not be revealed
and its propagation may simply be described using rays,
and the corresponding theory is called ray optics or geo
metric optics. Ray optics is an approximation of wave
optics, which, in turn, is a simpliﬁcation of the
electromagnetic theory of optics. All these theories are
classical, but there are other light phenomena that are
quantummechanical, which cannot be explained in terms
of the classical electromagnetic theory. The most complete
theory that combines quantummechanical phenomena
with the classical electromagnetic theory is called quan
tum electrodynamic and, in the realm of optics, is referred
to as quantum optics. The level of complexity of these the
ories follows their historical development: (1) geometric
optics, (2) wave optics, (3) electromagnetic optics, and (4)
quantum optics.
Historically, geometric optics was devised to explain
the behavior of light. The basic principles of geometric op
tics that are discussed in this article were all formulated
before the year 1900. Some relevant steps in the develop
ment of theories related to light follow and are based
on Ref. 2; a more complete overview is provided in Ref. 3.
The law of reﬂection was given by the Greek Euclid
(ca. 300 B.C.). Refraction was observed by Alexandrian
Greeks, but they could not explain it. Abu Ali alHasan
Ibn alHaytham (ca. 965–1039), latinized as Alhazen, in
GEOMETRICAL OPTICS 1777
his book on optics [4] gave a detailed description of refrac
tion and investigated reﬂection from plane and curved
surfaces, also known as ‘‘Alhazens’ problem.’’ Optics was
put on solid foundation only after the introduction of
the experimental method by the Italian Galileo Galilei
(1596–1650). We have to wait until the seventeeth century
to obtain a mathematical formulation of the law of refrac
tion by the French Rene´ Descartes (1596–1650) and the
Dutch Willebrod Snell (1580–1626). During the same cen
tury, two physical theories of light appeared. One theory
was developed by the Dutch physicist Christiaan Huygens
(1629–1695) and consisted in a geometric wave theory of
light [5]. The other theory was created by the British Sir
Isaac Newton (1643–1727), who developed a mechanical
theory of propagation of particles [6]. Newton’s theory was
accepted for nearly another century only because of his
reputation, while Huygens’ theory was ignored.
The seventeenth century gave also birth to experimen
tal discoveries such as interference by the British Robert
Boyle (1627–1691) and the British Robert Hooke (1635–
1703); the ﬁniteness of the speed of light by the Danish
Olaf Ro¨mer (1644–1710) and diffraction by the Italian
Francesco Maria Grimaldi (1618–1663) and Hooke. After
the experimental discoveries, geometric optics appeared
as a limited theory of light that was based on four prin
ciples: (1) light travels along straight lines in a homoge
neous medium, (2) light rays travel out of a source
independently, (3) light rays obey the law of reﬂection,
and (4) light rays obey the law of refraction. It is interest
ing to observe that all these laws follow from a mathe
matical principle that says nothing about the nature of
light. This is principle of least time by the French Pierre
de Fermat (1601–1665).
The mathematical theory of geometric optics received
its deﬁnitive formulation with the work of the Irish Sir
William Rowan Hamilton (1805–1865). Hamilton was
aware of the works of the French Augustin Jean Fresnel
(1788–1827), Huygens, and Newton, but he ignored their
works because his goal was to build a mathematical sci
ence of optics from which all properties could be deduced.
Hamilton’s main idea was the characteristic function. In
fact, he showed that from the knowledge of the character
istic function all optical problems involving, for instance,
lenses, mirrors, crystals, and propagation in the atmo
sphere could be solved.
Many scientists pointed out that in order to explain
properties of light such as interference, diffraction, and po
larization, a wave theory was necessary. Most of the early
works on the wave theory of light were driven by an equiv
alence with waves inside elastic media. Before the work of
the British James Clerk Maxwell (1831–1879), it is worth
mentioning the results of the Irish James MacCullagh
(1809–1847), who introduced differential equations that
are closely related to the ones of Maxwell [7]. As of today,
the best theory available to explain the behavior of light
(with the exception of some problems that are investigated
by quantum electrodynamics) is the one that subjects light
to an electromagnetic wave that obeys Maxwell’s laws.
As mentioned earlier, geometric optics is an approxi
mate representation. Wavelengths do not come into
account, so interference and dispersion cannot be deduced.
Additionally, the vector character of light (i.e., polariza
tion) and diffraction are not incorporated.
2. THE POSTULATES OF GEOMETRIC OPTICS
Geometric optics can be explained using four postulates:
*
Light rays travel out of a source independently.
*
Inside a homogeneous medium, light rays travel
along straight lines.
*
Light rays obey the law of reﬂection.
*
Light rays obey the law of refraction.
When light rays propagate inside a homogeneous me
dium, they travel at a speed vrc
0
, where c
0
is the speed of
light in vacuum, and the medium is characterized by the
index of refraction
n=
c
0
v
(1)
The time taken by a ray of light to travel the geometric
distance ‘ is proportional to the product n‘, which is called
the optical path length.
Let us now consider the simplest kind of optical source,
a point source inside a homogeneous medium. Rays emitted
from a point source are associated with spherical surfaces,
centered at the source, called wavefronts. Each wavefront
has the property that all its points have the same optical
pathlength from the source. Equivalently, all the points
that belong to the same wavefront have the same phase.
Since there is an equivalence between rays and the asso
ciated wavefronts, in what follows we will refer to either
rays or wavefronts depending on what is more convenient.
3. THE LAW OF REFLECTION
3.1. Reﬂection from a Plane Surface
The law of reﬂection considers a ray that, after interacting
with a surface, bounces back to the same medium where it
came from. Let us examine a ray with unit vector ´rr
1
that is
incident on a plane reﬂecting surface whose unit normal
vector is ^ mm, as shown in Fig. 1. The incident ray ´rr
1
and the
normal ^ mm deﬁne the plane of incidence that contains the
reﬂected ray unit vector ´rr
2
. The angle of incidence y
i
and
the reﬂection angle y
r
are related to each other by
y
i
=y
r
(2)
which is the mathematical statement of the lawof reﬂection.
3.2. The Method of Images
The reﬂected ray ´rr
2
appears to an observer O as if it had
originated from S
/
, which is symmetric to S with respect to
the plane reﬂecting surface. The point S
/
is called the im
age of S and suggests an equivalent method to determine
the direction of reﬂection of a ray incident at P. According
to this method, the location P of the reﬂection point is
1778 GEOMETRICAL OPTICS
determined by tracing the straight line from S
/
toward O.
The formation of the image of an extended source due to
the reﬂection from a plane surface is easily explained with
the method of images. In fact, an observer O will receive
rays that appear to have originated from points symmetric
to the source, with respect to the plane reﬂecting source.
Hence, in particular, the spherical wavefront originated
by a point source will maintain its spherical shape on
reﬂection from a plane interface because all rays associat
ed with the reﬂected wavefront appear to have originated
from the image of the original point source.
3.3. Reﬂection from a Curved Surface
When the reﬂecting surface is curved, one can still apply
the law of reﬂection using the formulation (2) provided
that the geometry is made locally plane by considering the
plane tangent to the curved surface S at P, as shown in
Fig. 4. In such a case, the plane of incidence contains the
normal ^ mm to the tangent plane and the angles of incidence
and reﬂection are measured between the normal and the
direction of incidence or reﬂection, respectively. When re
ﬂection occurs at a smooth curved surface, it is convenient
to present the law of reﬂection in a compact vector nota
tion. For this purpose, let ^ rr
1
and ^ rr
2
be the unit vectors
along the directions of the incident and reﬂected rays and
^ mm the unit vector of the normal to the surface at the point of
reﬂection. The law of reﬂection may be written
^ rr
2
= ^ rr
1
÷a^ mm (3)
where the constant a is determined from
a = ÷2^ rr
1
.
^ mm (4)
Since, according to (3), the unit vector ^ rr
2
is a linear com
bination of ^ rr
1
and m, the reﬂected ray is contained in the
plane of incidence. In addition ^ rr
1
and ^ rr
2
have the same
length, therefore they intersect ^ rr
1
÷ ^ rr
2
and, hence, ^ mm at the
same angle. Thus the angle of incidence equals the angle
of reﬂection and
y
1
=arccos(^ rr
1
.
^ mm) =arccos(^ rr
2
.
^ mm) =y
r
(5)
For practical applications, such as in ray tracing, one is
usually given ^ rr
1
and ^ mm. Hence, after obtaining a from (4),
the direction of the reﬂected ray is obtained from (3). Fur
ther details for reﬂection from general curved surfaces are
given in Refs. 8 and 9.
For a generic curved surface there is no method of im
ages. In fact, for an observer O all rays incident on the
curved surface at P will appear to have originated from
the images of their sources with respect to the tangent
plane at P. However, rays that are reﬂected at other points
will, in general, appear to have originated from images
that are different from the previous ones. In particular, a
spherical wavefront will not appear to be originated from a
single image source on reﬂection from a curved surface.
The reﬂected wavefront will be, in general, astigmatic.
3.4. Cartesian Reﬂecting Surfaces
Some conic sections represent Cartesian reﬂecting sur
faces; that is, the surfaces form perfect images of point
objects upon reﬂection. These surfaces are the ellipsoid;
the paraboloid and the hyperboloid and their cross sec
tions are shown in Fig. 2. The ellipsoid has the property
that all rays passing through one focus are reﬂected at the
surface so that the reﬂected ray passes through the other
focus. The paraboloid has the property that rays parallel
to its axis are reﬂected so that they all pass through the
focus. For the hyperboloid, rays emitted from one focus
are reﬂected so that they appear to have originated from
the other focus. Unfortunately, the relationship that leads
to perfect images is valid only for certain points.
plane of
incidence
O
S
v
P
r
2
r
1
^
θ
r
θ
i
S′
^
^
Figure 1. Geometry for the reﬂection from a plane surface. The
plane of incidence deﬁned by the unit vectors ´rr
1
and ^ mm contains
the unit vector ´rr
2
.
O ≡ F
1 I ≡ F
1
(a) (b)
(c)
O ≡ F
I → ∞
O ≡ F
2
I ≡ F
1
Figure 2. Cross sections of Cartesian reﬂecting surfaces: (a) el
lipsoid; (b) paraboloid; (c) hyperboloid.
GEOMETRICAL OPTICS 1779
4. THE LAW OF REFRACTION
4.1. Refraction from a Plane Surface
The law of refraction considers the passage of a ray from
one medium to another. Figure 3 shows the geometry for a
ray with unit vector ^ rr
1
that is incident at P on a plane
interface between two media with indices of refraction n
1
and n
2
. As in the case of reﬂection, one can deﬁne a plane
of incidence that contains ^ rr
1
and ^ mm. The incident ray ^ rr
1
generates one reﬂected ray ^ rr and a refracted or transmit
ted ray ´ r
2
r
2
. Both ^ rr and ´ r
2
r
2
are contained in the plane of
incidence. The angles of incidence and refraction (or
transmission) depend on the properties of the media
through their indices of refraction according to
n
1
siny
1
=n
2
siny
2
(6)
which is known as Snell’s law. Because of Snell’s law, the
direction of the refracted ray
y
2
=arcsin
n
1
n
2
siny
1
_ _
(7)
is closer to the normal to the interface when n
2
4n
1
and
farther away from the normal when n
2
on
1
. In particular,
when n
2
on
1
, the value of the incidence angle y
1
that
causes y
2
=p=2 is called critical angle because for y
1
4y
c
there is no refracted ray, a phenomenon known as total
internal reﬂection.
4.2. Refraction from a Curved Surface
Refraction from a smooth curved surface is more conve
niently expressed using a compact vector notation. Refer
ring to Fig. 4, let ^ rr
1
and ^ rr
2
be the unit vectors along the
directions of the incident and reﬂected rays, respectively,
and ^ mm the unit vector of the normal to the surface at the
point of refraction. The law of refraction may then be
written as
n
2
^ rr
2
=n
1
^ rr
1
÷b^ mm (8)
where the constant b is determined from
b =n
2
cos y
2
÷n
1
cos y
1
(9)
and
cos y
1
= ^ rr
1
.
^ mm
n
2
cos y
2
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
n
2
2
÷n
2
1
÷n
2
1
cos
2
y
1
_ (10)
Equation (8) deﬁnes ^ rr
2
as a linear combination of ^ rr
1
and ^ mm
so that the reﬂected ray is located in the plane of inci
dence. By taking the crossproduct of (8) with ^ mm, one ob
tains the usual form of Snell’s law
n
1
siny
1
=n
2
siny
2
(11)
(since [ sin ^ rr
1
^ mm[ = siny
1
; [ sin ^ rr
2
^ mm[ = siny
2
, and [ sin
^ mm ^ mm[ =0). In a raytracing application, one usually knows
^ rr
1
and ^ mm. Then, after evaluating (10), the constant b is
obtained from (9) and the direction of the refracted ray
is derived from (8). Further details for refraction from
general curved surfaces are given in Refs. 8 and 9.
Contrary to the law of reﬂection, there is no equivalent
to the method of images for refraction, even in the case of a
Figure 3. Geometry for the refraction of a ray. The plane of in
cidence deﬁned by the unit vectors ´rr
1
and ^ mm contains the unit
vector ´rr
2
and ^ rr.
Figure 4. Geometry for the reﬂection and refraction of a ray from
a curved surface.
1780 GEOMETRICAL OPTICS
plane interface. In fact, let us consider, for example, rays
from a point source S located in a medium 1. Rays from S
incident at the interface are refracted into medium 2 along
different directions that, when extended back into medium 1,
will not intersect all at the same point. Each ray refracted
into medium 2 appears to be coming from a different
source inside medium 1. Hence this could lead to the for
mation of multiple images, but the problem is avoided if
one observes only a narrow bundle of rays so that they will
appear to be coming from the same source. This is the case
of a human eye that observes an object under water. Since
the pupil captures only a narrow bundle of rays, no mul
tiple images are observed. However, it is possible to ﬁnd
only one image by limiting the consideration to rays that
make small angles with the normal ^ mm. In fact, when the
angles are small, the sine function can be approximated
with its argument. Accordingly, the refraction law (6)
takes the form
n
1
y
1
=n
2
y
2
(12)
Therefore, referring to Fig. 5, rays that make small angles
with the normal appear to an observer O as if they had
originated by S
/
at an apparent depth
d
/
=
n
1
n
2
d (13)
Most optical systems contain spherical surfaces. Simi
lar to the case of refraction at a plane interface, rays in
medium 2, originated by a point source in medium 1 and
refracted by a spherical surface, do not intersect all at the
same point. This is a statement of the fact that spherical
surfaces do not produce perfect images. This difﬁculty of
spherical surfaces is overcome, again, by limiting the rays
to make small angles with a preferred axis. This results
into the paraxial approximation, which is explained in
Section 8. Further details of the limitations of spherical
surfaces are explained in Section 11. A different approach
to the laws of geometric optics is considered next.
5. FERMAT’S PRINCIPLE
The laws of geometric optics may be derived from the
purely mathematical Fermat principle. In order to intro
duce this principle, let us recall that the optical path
length of a ray that propagates from point A to point B
inside a medium with index of refraction n(x, y, z) is
expressed by
_
B
A
n(x; y; z) ds (14)
According to Fermat’s principle, the actual path followed
by the ray of light is such that the optical pathlength (14)
is stationary in the sense of variational calculus
d
_
B
A
n(x; y; z) ds =0 (15)
where the symbol d represents the variation. Without go
ing into the formal definition of what a variation is ac
cording to variational calculus, the condition (15) may be
understood by making a parallelism with ordinary differ
ential calculus and looking at (15) as being a derivative of
(14). In particular, if p
0
is the trajectory that satisﬁes (15),
trajectories that are in the neighborhood of p
0
will not
change the value of (14) to a ﬁrstorder approximation.
Hence, the trajectory p
0
behaves similarly to a stationary
point that may represent either a maximum, a minimum,
or an inﬂection point for the quantity represented by (14).
Usually, the trajectory p
0
corresponds to a minimum and,
since (14) is proportional to the travel time, the Fermat
principle is formulated by saying that light rays travel
along paths of least time.
Using Fermat’s principle, one can derive the laws of
reﬂection and refraction. For the law of reﬂection, refer
ring to Fig. 6a, it is easy to prove that the path SPO that
satisﬁes Fermat’s principle (15) is shorter than any other
path such as SAO. Only for the path SPO, the angles y
i
and y
r
satisfy the law of reﬂection (2). For the law of
refraction, Fig. 6b shows that when the location of a vari
able point P is chosen so that the path SPO satisﬁes (15),
then the angles y
1
and y
2
satisfy the law of refraction (6).
5.1. Principle of Reversibility
The principle of reversibility states that if a ray of light
propagating from A to B is reversed in direction, it will
follow the same path backward. This property follows
from Fermat’s principle that does not account for the ac
tual direction of propagation of light.
5.2. Ray Equation
Another consequence of Fermat’s principle is the ray
equation. If the trajectory of a ray r is described by the
n
2
>n
1
n
1
d′
S′
S
d
O
v
1
2
∧
Figure 5. Geometry for the refraction at a plane interface when
angles are small.
GEOMETRICAL OPTICS 1781
coordinates (x(s), y(s), z(s)), where s is the length of the
trajectory measured from an arbitrary point along the
trajectory, it can be shown that the ray equation is
d
ds
n
dr
ds
_ _
=Vn (16)
or, equivalently, that the functions x(s), y(s), z(s) must sat
isfy the partialdifferential equations:
d
ds
n
dx
ds
_ _
=
@n
@x
;
d
ds
n
dy
ds
_ _
=
@n
@y
;
d
ds
n
dz
ds
_ _
=
@n
@z
(17)
One of the postulates of geometric optics, the rectilinear
propagation of rays of light inside a homogeneous medi
um, follows directly from the ray equation (16). In fact, for
a homogeneous medium, Vn=0 so that after two integra
tions, one obtains
r =as ÷b (18)
which is the equation for a straight line where a and b are
constant vectors. When the medium is not homogeneous,
the analytical derivation of the solution of (16) is usually
not trivial, but mathematical solutions are possible when
the paraxial approximation is made. Accordingly, one
assumes that the trajectory is almost parallel to the z
axis so that the following two equations are obtained
d
dz
n
dx
dz
_ _
=
@n
@x
;
d
dz
n
dy
dz
_ _
=
@n
@y
(19)
5.3. GradedIndex Fiber
Consider a glass ﬁber cylinder with an index of refraction
that depends on the distance from the axis, such as
n
2
=n
2
0
[1 ÷a
2
(x
2
÷y
2
)] (20)
When this expression for n is substituted into (19) and the
approximation a
2
(x
2
÷y
2
) << 1 is made, one ﬁnds that the
components x(s), y(s) of the ray follow the equations
d
2
x
dz
2
= ÷a
2
x;
d
2
y
dz
2
= ÷a
2
y (21)
which represent harmonic motions with period 2p/a. As
sume that the initial position of a ray is (x
0
, y
0
) and the
initial angle is (y
x
0
; y
y
0
). Since this problem has circular
symmetry, one may set x
0
=0 without loss of generality.
Then the solutions of (21) may be written as follows:
x(z) =
y
x
0
a
sin(az); y(z) =
y
y
0
a
sin(az) ÷y
0
cos(az) (22)
In the particular case of a meridional ray, a ray inside a
plane that contains the z axis, y
x
0
=0 and the solution (22)
shows that the ray will continue to lie in the same plane
and will propagate along a sinusoidal trajectory. When
y
y
0
=0 and y
x
0
=a
y
0
, then (22) reduces to a helical trajec
tory on the surface of a cylinder of radius y
0
. In both these
cases, the trajectory remains conﬁned within the glass
ﬁber cylinder, which, therefore, operates as a light guide.
By changing the initial conditions, other helical trajecto
ries are obtained.
5.4. Spherical Symmetric Media and Lenses
An optical medium with index of refraction that is only a
function of the radial distance from the origin is of con
siderable interest because it makes it possible to deter
mine conditions to obtain perfect images. In fact, if
n=n(r), the light ray trajectories are plane curves (simi
larly to the trajectories of a particle in the ﬁeld of a central
force) and, without loss of generality, one may limit the
investigation to trajectories that lie in a plane [11]. Refer
ring to Fig. 7, assume that a unit radius sphere, a spher
ical lens, is made with material of index of refraction
n=n(r) and that the material outside the lens is homoge
neous with n=1. The problem consists in ﬁnding the
functional dependence n=n(r) so that all rays originating
Figure 6. Fermat’s principle: (a) geometry for the reﬂection of a
ray at a plane surface; (b) geometry for the refraction of a ray at a
plane interface between two media with different indices of
refraction.
x
r=1
o
I O
O(x
O
y )
O
, I(x
,
y )
I I
y
r r
Figure 7. Geometry for the problem of a spherical lens made of a
material with index of refraction that is axially nonhomogeneous.
1782 GEOMETRICAL OPTICS
at O and passing through the spherical lens are focused
at I. This problem was solved by Luneburg when n(r) is
continuous and monotonic in 0rrr1. Two explicit solu
tions are Maxwell’s ﬁsheye and the standard Luneburg
lens. Maxwell’s ﬁsheye images points on the surface of the
spherical lens into points that are diametrically opposite
on the lens surface, hence r
O
=r
1
=1 and n(r) =2/(1 ÷r
2
).
The standard Luneburg lens images a point on the surface
of the spherical lens at inﬁnity, hence r
O
=1, r
1
=N and
n(r) =
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
2 ÷r
2
_
. The Luneburg lens and its generalizations
ﬁnd many applications in other disciplines, as mentioned
in Section 13.
Now that the laws of geometric optics have been intro
duced, we will examine what an optical system is and how
perfect images are created.
6. DEFINITION OF AN OPTICAL SYSTEM
A source point O inside a medium with index of refraction
n generates an inﬁnite number of rays, only a ﬁnite num
ber of which, in general, will reach any other point in the
medium. If it is possible to ﬁnd a point I through which an
inﬁnite number of rays pass, this point is said to be a
stigmatic (or sharp) image of O. An ideal optical system
transforms every point O of a threedimensional region
called the object space into a stigmatic image I. The col
lection of all image points is called image space [12]. An
optical system is composed of a number of refracting and
reﬂecting surfaces and is usually described by neglecting
its internal structure and referring to the transformation
that it operates, as indicated in Fig. 8. Points that corre
spond to each other through this transformation are called
conjugate points. Figure 8 shows a real object because rays
propagate out of an actual source. Similarly, the image I of
the point O is real because it is formed by rays converging
at one location. An image may also be virtual. This occurs
when the rays forming the image appear to have originat
ed from a point even though no actual source of light is
located at that point. The relationship between the object
and the image may be equivalently described using wave
fronts. For example, the optical system of Fig. 8, trans
forms a wavefront diverging out of O into a wavefront
converging at I. In the case of a virtual image, its wave
front appears to be diverging out of a point where no
actual source of light is located [13]. When O describes
a curve C
0
in the object space, I will describe a conjugate
curve C
I
in the image space. In general, the conjugate
curves are not similar to each other. However, if every
curve C
O
is geometrically similar to its conjugate curve C
I
,
the imaging between the two spaces is perfect and so is the
optical system.
To determine the transformation operated by an optical
system, a sufﬁcient knowledge of its internal structure is
required. The numerical analysis of a given optical system
is carried out by considering a point O¬(x
O
, y
O
) of a certain
plane z =z
O
, the object plane. An appropriate number of
rays, which originate at O, are traced through the system
with the goal of ﬁnding the intersections (x
I
, y
I
) in another
plane z =z
I
, the image plane. The ray tracing operations
are carried out using the laws of reﬂection and refraction
described earlier. If the optical system is perfect, these
rays intersect the image plane at the ideal image point
x
I
=Mx
0
; y
I
=My
O
(23)
where M is a constant of the optical system that depends
only on the choice of the object plane and is called mag
niﬁcation. Hence, a perfect image is a scaled replica of the
object.
In all that follows, we assume that rays normally prop
agate from left to right. The ﬁrst optical system that is
examined in more detail is the surface that separates two
media with different indices of refraction.
7. CARTESIAN REFRACTING SURFACES
Spherical refracting surfaces do not provide a unique im
age of a point object, unless rays are restricted to make
small angles with the optical axis. In order to obtain a
unique image of a point object without imposing restric
tions on the rays, the shape of the refracting surface must
be changed. Referring to Fig. 9, the surface S is deter
mined so that each ray that originated at O and refracted
at P passes through I. Hence, the refracting surface S
represents a perfect optical system according to the pre
vious definition. The shape of S is obtained by applying
Fermat’s principle and requesting that for each point
Figure 8. Optical system and conjugate points.
O(x
O
,y
O
) I(x
I
,y
I
)
V
n
1
n
2
P(x,y)
Σ
s
O
s
I
optical axis
Figure 9. Geometry for a surface that refracts O into I. Note that
the surface S is not spherical.
GEOMETRICAL OPTICS 1783
P(x, y) on S, the path length OPI be the same, i.e.:
n
1
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
(x ÷x
0
)
2
÷(y ÷y
0
)
2
_
÷n
2
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
(x ÷x
I
)
2
÷(y ÷y
I
)
2
_
=n
1
s
O
÷n
2
s
I
(24)
This results in an algebraic curve of the fourth order,
known as Cartesian oval, which is shown in Fig. 9. In
practice, one would like to have the object and the image
in the same medium and this is not accomplished by a
single refracting surface. However, lenses provide a solu
tion since rays passing through a lens undergo two re
fractions and the image appears outside the lens in the
same medium of the object. Now the problem is to deter
mine the appropriate shape of each refracting surface of
the lens. Of particular interest are Cartesian refracting
surfaces that refract rays from a point object into parallel
rays. These parallel rays then will be refracted at the
second surface of the lens to form the image. Figure 10a
shows an ellipsoidal surface that refracts rays from O into
parallel rays when n
2
on
1
, while Fig. 10b shows a hyper
bolic surface that refracts rays from O into parallel rays
when n
2
4n
1
. It is important to understand that these
surfaces provide perfect images only for object points O at
very specific locations.
Most actual optical systems contain spherical reﬂecting
and refracting surfaces because they are easy to manu
facture. In order to obtain images from these systems, rays
must be limited to make small angles with the optical axis.
These rays are called paraxial and are the topic of the next
section.
8. PARAXIAL APPROXIMATION
Most optical systems contain spherical surfaces. When the
centers of curvature of all these surfaces are aligned along
an axis, the system is said to be centered and the axis is
called the optical axis. If light rays travel almost parallel
and at a short lateral distance from the optical axis, it is
possible to make simpliﬁcations so that the laws of reﬂec
tion and refraction applied to spherical surfaces take sim
pler forms. Approximations made assuming that rays are
almost parallel to the optical axis are referred to as par
axial approximations or Gaussian optics. If an angle y
satisﬁes [y[51 because of the paraxial approximation, the
following simpliﬁcations approximations are made
siny ’ tany ’ y; cos y ’ 1 (25)
and, in particular, Snell’s law becomes
n
1
y
1
=n
2
y
2
(26)
8.1. Reﬂection at a Spherical Mirror
As a ﬁrst application of the paraxial approximation, the
reﬂection at a spherical mirror, shown in Fig. 11, is con
sidered. To obtain the location of the image, at least two
straight lines must be drawn. The ﬁrst line goes from the
tip P of the object to the center of curvature C because this
line does not change direction on reﬂection. The second
line is drawn from the tip P of the object to the vertex V, so
that it is easy to determine the direction of the reﬂected
ray. The intersection of these two lines determines the tip
O ≡ F
1
n
1
n
2
n
1
n
2
O ≡ F
1
(a)
(b)
Figure 10. Cross sections of Cartesian refracting surfaces:
(a) ellipsoidal surface that images O at inﬁnity when n
2
on
1
;
(b) hyperbolic surface that images O at inﬁnity when n
2
4n
1
.
Figure 11. Construction of the images for a concave and a convex
spherical mirror.
1784 GEOMETRICAL OPTICS
Q of the image from which the whole virtual image is
found. The locations of object and image are determined
by using the paraxial approximation and the similarities
between the triangles OPC, IQC and OPV, IQV, which
yield
1
s
O
÷
1
s
I
=
1
f
(27)
where f is the focal distance of the mirror, deﬁned as
f = ÷
R
2
> 0; concave mirror
o0; convex mirror
_
(28)
The distances s
O
and s
I
are measured with a sign con
vention that assumes propagation of light from left to
right. A real object (image) is located to the left of V and
corresponds to a positive value for s
O
(s
I
). When the object
(image) is virtual, it is located to the right of Vand s
O
(s
I
) is
negative. For this geometry, the lateral magniﬁcation of
the image is
m= ÷
s
I
s
O
(29)
where the negative sign used to obtain a positive number
m, due to the sign convention on the distances.
8.2. Refraction at a Spherical Surface
A second application of the paraxial approximation is the
refraction from a spherical surface, shown in Fig. 12. Sim
ilar to the reﬂection from a mirror, a minimum of two lines
must be drawn to ﬁnd the location of the image. The ﬁrst
line is traced from the tip P of the object to the center of
curvature C, because this line is not deﬂected. The second
line is drawn from the tip P of the object to the vertex V
and then continues inside the medium n
2
with an angle y
2
with respect to the normal at V. These two lines intersect
at Q, which gives the location of the image. Application of
the paraxial approximation and use of the similarity
between triangles OPC and CIQ yields
n
1
s
O
÷
n
2
s
I
=
n
2
÷n
1
R
(30)
where the radius of curvature R is positive for a convex
spherical surface. s
O
and s
I
are positive for real and
negative for virtual object and images, respectively. The
lateral magniﬁcation is given by
m= ÷
n
1
s
I
n
2
s
O
(31)
8.3. The Thin Lens
The relationship just found for a spherical refractive sur
face may be used to obtain the equation of a thin lens,
which is made of two spherical refractive surfaces that are
separated by a negligible distance. Specifically, assuming
that light propagates inside a medium of refractive index
n
1
, crosses a refractive spherical interface with radius of
curvature R
1
, enters a medium of index n
2
, crosses a
spherical interface with radius of curvature R
2
and, ﬁnally,
enters back into the medium with index n
1
, one obtains
the lensmaker equation
1
s
O
÷
1
s
I
=
1
f
(32)
where s
O
and s
I
are the object and image distances, re
spectively, which are measured from the plane of the thin
lens. In the previous equation, the focal distance f is
1
f
=
n
2
÷n
1
n
1
1
R
1
÷
1
R
2
_ _
(33)
The lensmaker equation also may be written in a form
that emphasizes the curvatures of the wavefronts. In fact,
referring to Fig. 13, a spherical wave that expands from O
has a curvature, or vergence, V=1/s
O
when it reaches the
plane of the lens. The lens is characterized by a refracting
power P=1/f that changes the curvature of the wavefront
from V to V
/
=1/s
I
when the wavefront enters the image
space to the right of the lens. With these definitions, the
alternative form of the lensmaker equation is
V ÷V
/
=P (34)
and its terms have units of diopters, i.e. the inverse of a
length, when the distances are measured in meters.
When two or more thin lenses are attached together,
their refractive powers add. So n thin lenses with focal
Figure 12. Construction of the image at a spherical refracting
surface.
Figure 13. Interpretation of the lensmaker equation in terms of
wavefront curvature.
GEOMETRICAL OPTICS 1785
distances f
1
,y, f
n
correspond to one thin lens with equiv
alent focal length f such that
1
f
=
1
f
1
÷ · · · ÷
1
f
n
(35)
or with a refractive power
P=P
1
÷ · · · ÷P
n
(36)
Many optical systems are designed, at least initially,
using paraxial approximations. In fact, paraxial optics re
lationships are very useful in the analysis of optical sys
tems with many elements, provided that these
relationships are cast according to the method discussed
in the next section.
9. RAY TRANSFER MATRICES
The analysis of a complex optical system may be carried
out by the repeated application of the laws of geometric
optics. However, this operation becomes cumbersome even
for systems constituted only by a few elements. Fortunate
ly, the paraxial approximation leads to a formalism that
describes the passage of a ray through an optical system in
terms of simple matrix multiplications. This formalism,
known as matrix optics, is valid under the paraxial or
Gaussian optics approximation and is limited to rays that
travel in a single plane; hence it is applicable to systems
that have a planar geometry or to meridional rays in
circular symmetric systems.
Referring to Fig. 14, at any plane perpendicular to the
optical axis, a ray is described by its position y and angle y,
so that the ray at the output plane z
2
is related to the ray
at the input plane z
1
by the linear equations:
y
2
=Ay
1
÷By
1
(37)
y
2
=Cy
1
÷Dy
1
(38)
These two equations may be cast in matrix form as
y
2
y
2
_ _
=
A B
C D
_ _
y
1
y
2
_ _
(39)
where the matrix with elements A, B, C, D is called ray
transfer matrix or ABCD matrix. We now are going to
consider the ray transfer matrices for some optical sys
tems.
1. Translation Matrix. The simplest optical system is
the free space between two planes. Referring to Fig. 15,
the matrix that describes the propagation is
M=
1 L
0 1
_ _
(40)
2. Reﬂection at a Spherical Interface. The geometry for
the reﬂection at a spherical interface is described in
Fig. 16, which corresponds to the matrix
M=
1 0
2
R
1
_
_
_
_
(41)
Note that the same sign convention applies to all angles
both before and after reﬂection. Rays pointing upward are
associated with positive angles while rays pointing down
ward are associated with negative angles.
3. Reﬂection at a Planar Surface. This matrix may be
derived from the previous one in the limiting case that the
optical axis
output plane
input plane
2
2
1
1
2 1
θ
θ
Z Z
y
y
Figure 14. Geometry for the matrix description of a ray.
2
1
θ
θ
L
y
2
1
y
optical axis
2 1
Z Z
Figure 15. Geometry for the translation matrix.
Figure 16. Geometry for the reﬂection from a spherical interface
and sign convention for the angles.
1786 GEOMETRICAL OPTICS
curvature RN. In such a case, one obtains
M=
1 0
0 1
_ _
(42)
4. Refraction at a Spherical Interface. The geometry
for this case is given in Fig. 17, and the corresponding
matrix is
M=
1 0
n
1
÷n
2
Rn
2
n
1
n
2
_
_
_
_
(43)
5. Refraction at a Plane Interface. This is a special case
that may be derived when the curvature RN, which
yields
M=
1 0
0
n
1
n
2
_
_
_
_
(44)
6. ThinLens Matrix. The matrices reported above
may be used to derive the ray transfer matrix for a thin
lens. In fact, by considering the product of two matrices
representing the refraction from a spherical surface, one
obtains
M=
1 0
n
2
÷n
1
R
2
n
1
n
2
n
1
_
¸
_
_
¸
_
1 0
n
1
÷n
2
R
1
n
2
n
1
n
2
_
¸
_
_
¸
_
=
1 0
÷
1
f
1
_
¸
¸
_
_
¸
¸
_
(45)
where the focal length f is given by (33). The thinlens
matrix is an example of an optical system that is studied
by cascading the matrices of simpler optical systems. In
doing so, it is important to observe that if a ray passes
through optical elements characterized by the matrices
M
1
,y, M
n
, the overall transformation is represented by a
matrix given by the product
M=M
n
M
n÷1
.
. . .
.
M
1
; (46)
Thus, the order of the matrices at the right hand side of
the previous equation is inverse to the order followed by
the ray to go through the elements of the system. One
property of ray matrices is that their determinant is al
ways equal to the ratio of the indices of refraction of the
ﬁrst and ﬁnal media. This property may be used as a nec
essary condition to verify if a system matrix is computed
correctly. Further details on matrix optics are found in the
literature [1,8,10,14].
9.1. Cardinal Points of an Optical System
The characteristics of an optical system may be derived
from the knowledge of its six cardinal points, which are
illustrated in Fig. 18: two focal points F
1
and F
2
, two nodal
points N
1
and N
2
, and two principal points Q
1
and Q
2
. Fo
cal points are the locations where rays parallel to the op
tical axis converge. The intersections of rays entering the
system with those leaving the system deﬁne the principal
surfaces, which are usually spherical and centered on the
object and the image. In the paraxial approximation, these
surfaces are planes and are referred to as principal planes.
Their intersections with the optical axis deﬁnes the prin
cipal points. Nodal points are located on the optical axis
and have the property that a ray directed through one of
them leaves the system from the other nodal point along a
Figure 17. Geometry for the refraction from a spherical inter
face.
Figure 18. The six cardinal points of an optical system; the prin
cipal planes and distances relative to both input/output and prin
cipal planes.
GEOMETRICAL OPTICS 1787
direction parallel to the one of the ﬁrst ray. The expres
sions ‘‘ﬁrst’’ focal, principal, and nodal points refer to car
dinal points deﬁned by rays entering the system from the
right. Conversely, ‘‘second’’ cardinal points are deﬁned by
rays entering the system from the left. The principal
planes 1 and 2 are used as references to measure the dis
tances within an optical system [15].
9.2. Meaning of Elements of ABCD Matrix
There are some interesting implications when the ele
ments of the ray transfer matrix go to zero. Referring to
(38), when D=0, we obtain y
2
=Cy
1
regardless of y
1
. So all
rays that enter the system at y
1
will leave with the same
angle y
2
. Hence the input plane of a system with D=0 is
ﬁrst focal plane.
When A=0, from (37), y
2
=By
1
hence input rays par
allel to each other will leave the system passing by the
point at y
2
. Therefore, the output plane of a system with
A=0 is the second focal plane of the system.
When B=0, from (37), y
2
=Ay
1
so all rays entering the
system at y
1
leave the system at y
2
. Hence the input and
output planes are conjugate and A=y
2
/y
1
represents the
linear magniﬁcation.
When C=0, from (38), y
2
=Dy
1
so all rays parallel to
the y
1
direction will leave the system parallel to each other
along y
2
with an angular magniﬁcation D=y
2
/y
1
. Such a
system is called telescopic because a telescope admits par
allel rays and returns parallel rays.
The characteristics of an optical system are determined
from its six cardinal points. Under the paraxial approxi
mation, the elements of the ABCD matrix determine the
characteristics of the system as well. Thus, there must be
a relationship between the cardinal points and the ABCD
matrix. This relationship is summarized in Table 1, where
the symbols used for the distances are deﬁned in Fig. 18.
Let us now move away from paraxial rays and consider
a general result, involving spherical surfaces, attributed
to the German Ernst Abbe (1840–1905).
10. REFRACTION AT A SPHERICAL INTERFACE
Let us consider refraction from a spherical surface when
the paraxial approximation is abandoned. Referring to
Fig. 19, a small object near the axis is considered and its
image is obtained by tracing two rays. The ﬁrst ray, OC,
passes through the center of curvature of the lens, so that
it is not deﬂected. The second ray, OP is refracted and in
tersects the ﬁrst ray at I. By simple considerations using
the triangles OPC, PCI and the similarity between the
triangles MOC, CNI one ﬁnds the following relationship
n
1
h
1
sinj
1
=n
2
h
2
sinj
2
(47)
which is called the Abbe sine condition. In the paraxial
approximation, the angles j
1
and j
2
are small so that the
Abbe sine condition is simpliﬁed into
n
1
h
1
j
1
=n
2
h
2
j
2
(48)
which is also known as the Lagrange theorem. For a cen
tered optical system, with many refracting surfaces, the
quantities n
i
h
i
sin j
i
or, in the paraxial case, n
i
h
i
j
i
are invariant, which constitutes an important property.
The Abbe sine condition is used in the one of Seidel’s
aberrations.
11. ABERRATIONS
The paraxial approximation is valid for rays that travel
very close to the axis of an optical system and leads to the
formation of perfect images. In practical optical systems
there are apertures, rays are not necessarily paraxial and
departures from perfect images are observed. These de
partures are called aberrations, and practical optical sys
tems are corrected to minimize aberrations so that images
that are as close as possible to the ideal ones are obtained.
To understand the origin of aberrations, let us consider an
optical system that images a point source. Under the
Table 1. Relationship between Cardinal Points, Ray
Matrix Elements, and Distances Shown in Fig. 18
Relative to Input and Output Reference Planes
g =
D
C
h= ÷
A
C
a=
D÷n
0
=n
f
C
b =
1 ÷A
C
c =
D÷1
C
d=
n
0
=n
f
÷A
C
Relative to Principal Planes
f
1
=g ÷a=
n
0
=n
f
C
f
2
=h ÷b = ÷
1
C
C
h
h
θ
θ
φ
φ
1
2
2
1
2
N
n
n
2
1
1
I
O
M
P
Figure 19. Geometry for the refraction from a spherical surface
when the paraxial approximation is removed.
1788 GEOMETRICAL OPTICS
paraxial approximation, the ideal wavefront at the exit
pupil of an optical system is spherical and is obtained from
linear expressions that result from the following trigono
metric expansions
sinx =x ÷
x
3
3!
÷
x
5
5!
÷ · · · ÷(÷1)
n
x
2n÷1
(2n÷1)!
cos x =1 ÷
x
2
2!
÷
x
4
4!
÷ · · · ÷(÷1)
n
x
2n
(2n)!
(49)
when they are stopped at the ﬁrst term. When higher
order terms are included, aberrations appear. In particu
lar, if the expansions (49) are stopped at terms of order no
larger than x
3
, one obtains the thirdorder aberration the
ory originally developed by the German mathematician
Ludwig Von Seidel (1821–1896). According to his theory,
there are ﬁve aberrations for monochromatic illumination:
spherical aberration, coma, astigmatism, curvature of
ﬁeld and distortion. If nonmonochromatic illumination is
considered, the chromatic aberration is the sixth Seidel
aberration. A further reﬁned theory considers terms up to
the order of x
5
, but for most practical applications Seidel
theory is sufﬁciently accurate. The theory of aberrations
is described, among others, elsewhere in the literature
[9–11,13,15–19].
11.1. ThirdOrder Theory
Figure 20 shows a generic optical system that creates the
actual wavefront S
2
and the ideal wavefront S
1
. The ac
tual ray forms an image L along the optical axis and the
image T in the plane of the ideal paraxial image, while the
ideal ray forms an image at I. There are different ways to
measure the aberration. One may refer to the longitudinal
aberration LI, the transverse, or lateral aberration TI or
the difference PQ of the wavefront position. Since most
optical systems contain spherical surfaces, the aberrations
of spherical surfaces are studied in great detail when the
object is either a point along the optical axis or a point off
the optical axis. For a point object O located on the optical
axis, the aberration is measured as shown in Fig. 21 by the
optical path difference between the shortest ray, OVI, and
another ray refracted at P, OPI. The aberration according
to the thirdorder theory, is given by the difference
a(P) =OCI ÷OPI =ch
4
(50)
where c is a proportionality constant.
For an object point O located off axis, as shown in
Fig. 22, the aberration measured as the difference between
the shortest ray and another ray diffracted at P is given by
a(P) =OCI ÷OPI (51)
However, all practical optical systems have an aperture
or pupil or stop that is usually centered along the optical
axis. Hence, a beam of rays out of O is not symmetrical
with respect to OC because of the presence of the aperture.
As a result of this lack of symmetry, the measure of the
aberrations for points located off axis leads to four (addi
tional) Seidel aberrations. Referring to Fig. 22, it is found
that
a(P) =C
040
r
4
÷C
131
h
I
r
3
cos y ÷C
222
h
2
I
r
2
cos
2
y
÷C
220
h
2
I
r
2
÷C
311
h
3
I
r cos y
(52)
where the subscripts of the coefﬁcients represents powers
of h
I
, r, cos y, respectively. The terms that appear in (52)
Figure 20. Difference between the ideal wavefront S
1
and the
actual wavefront S
2
and various measures for the aberration.
Figure 21. Geometry for the reﬂection at a spherical interface.
Figure 22. Geometry for refraction of an offaxis point O by a
spherical surface. The presence of a stop breaks the symmetry
around the shortest ray OCI.
GEOMETRICAL OPTICS 1789
represent
S
1
=C
040
r
4
spherical aberration
S
2
=C
131
h
I
r
3
cos y coma
S
3
=C
222
h
2
I
r
2
cos
2
y astigmatism
S
4
=C
220
h
2
I
r
2
curvature of field
S
5
=C
311
h
3
I
r cos y distortion
(53)
Each one of these aberrations is described in the follow
ing as if they were the only aberration present.
11.2. Spherical Aberration
Spherical aberration is the variation of the focal length
with the aperture and is the only aberration, among
S
1
÷S
5
indicated in (53), that does not depend upon h
I
so it always exists. In particular, it may exist when com
binations of spherical surfaces occur. Most optical instru
ments use parallel rays, so it is customary to measure, for
comparison purposes, the aberration due to parallel rays.
Figure 23 illustrates how spherical aberration results into
different focal points F
0
, F
1
, F
2
, F
3
depending on the di
ameter r of the zone of the lens where the beam of parallel
rays is incident. The spherical aberration of a lens is re
duced by changing the radii of curvature of the two spher
ical surfaces of the lens. In fact, since the focal length of
the lens is given by (33), it is possible to change the values
of the index of refraction n, R
1
and R
2
without altering the
focal length f.
As an example, Fig. 24 shows lenses with the same fo
cal length but different shape. The shape of a thinlens
may be characterized by the Coddington shape factor
z =
R
2
÷R
1
R
2
÷R
1
(54)
and by changing z the spherical aberration may be re
duced but not completely eliminated. It is possible to com
pletely eliminate spherical aberration from a single lens
by aspherizing it. However, such an aspherized lens will
be free from spherical aberrations only for a particular
object distance, so such a process is justiﬁed for only a few
special instruments.
11.3. Coma
Coma is the variation of magniﬁcation with the aperture
and is nonsymmetric with respect to the optical axis be
cause it depends on cos y. Figure 25 shows how this aber
ration affects a set of parallel rays that are refracted by a
thin lens. The ray through the center of the lens V is fo
cused at V
/
, while the rays through the edges of the lens P
are focused at P
/
. Hence, the magniﬁcation is different de
pending on which part of the lens the rays go through.
More specifically, referring to Fig. 26, each ringshaped
area of the lens forms a socalled comatic circle. Within
each ringshaped zone of the lens, rays contribute to dif
ferent parts of the comatic circle. For example, rays con
tained in the sagittal plane determine the top part of the
comatic circle, whereas rays contained in the tangential
plane determine the bottom of the comatic circle. Each
comatic circle increases in radius as the radius of the orig
inating zone on the lens increases. The superposition of
all comatic circles, shown in the bottom right corner of
Fig. 25, is a cometlike ﬁgure from which the name coma
for this aberration is derived.
In Section 10, we saw that rays making wide angles
with the optical axis must satisfy the Abbe sine condition
(47). This relation may be rearranged to point out the lat
eral magniﬁcation as
h
2
h
1
= ÷
n
1
n
2
sinj
1
sinj
2
(55)
Figure 23. Spherical aberration of a lens that produces different
images depending on the diameter h of the zone of the lens.
ζ=−2.25 ζ=−1 ζ=0 ζ=1 ζ=2.25
Figure 24. Lenses with same focal length but different Cod
dington shape factors.
Figure 25. A set of parallel rays generates a superposition of
comatic circles of different radii.
1790 GEOMETRICAL OPTICS
Hence, to prevent coma, the lateral magniﬁcation must be
the same for all zones of a lens:
sinj
1
sinj
2
=const (56)
Similar to spherical aberration, coma may be reduced by
changing the shape of a lens. In particular, it is possible to
prove that coma is absent when
z =
2n
2
÷n ÷1
n÷1
_ _
s
O
÷s
I
s
O
÷s
I
(57)
An optical system that is free of both spherical aberra
tion and coma is called aplanatic.
11.4. Astigmatism
Astigmatism is explained with the help of Fig. 27 that
shows an object point O and two beams of rays that are
obliquely incident on the lens. The rays contained in the
vertical, or tangential, plane tt focus along the horizontal
caustic line T, while the rays contained in the horizontal,
or sagittal, plane ss cross along the vertical caustic line S.
Normally, the image of a point is a point, but, when astig
matism is present, the image takes the form of two sep
arate lines. Between Tand S the cross section of the beam
is approximately circular and constitutes the circle of least
confusion for the image of O. When the angle y is varied, S
and T describe curved surfaces that are paraboloids of
revolution whose cross sections are indicated in Fig. 28.
The amount of astigmatism is measured by the distance
between these two surfaces along the chief ray (a ray that
passes through the center of the aperture of an optical
system). If the astigmatism is removed, the surfaces S and
T concide with each other and result into the Petzval sur
face (named after Joseph Max Petzval [1807–1891], a
Hungarian optician and mathematician).
11.5. Curvature of Field
Even if the ﬁrst three aberrations are removed, the image
is formed on the curved Petzval surface and, hence, the
focal surface is curved, which leads to the curvature of the
ﬁeld. This is understood by referring to Fig. 29, which
shows the location of an ideal image IQ of the object OP.
According to Fermat’s principle, for IQ to be an image, the
distances s
I
and s
Q
must be identical. Since this cannot
happen for an extended object OP, the image is formed
along the curved surface RS. A simple way to obtain ﬂat
tening of the image ﬁeld is by means of a stop that pre
vents oblique chief rays from penetrating the lens center.
11.6. Distortion
This aberration is caused by a different lateral magniﬁca
tion for points located at different distances from the op
tical axis. The common shapes of distorted images are
shown in Fig. 30, where part (a) represents the undistort
ed image of a wire mesh object; part (b) shows barrel dis
tortion, which is characterized by a smaller magniﬁcation
at the edges of the ﬁeld; and part (c) shows pincushion
distortion that is characterized by a greater magniﬁcation
Figure 26. Detail of the contribution of a ringshaped zone of a
lens to the creation of a comatic circle. Rays through the tangen
tial plane of the lens contribute to the bottom of the comatic circle,
and rays through the sagittal plane of the lens contribute to the
top of the comatic circle.
Figure 27. Formation of an astigmatic image. A fan of rays in the
sagittal, or horizontal, ss
/
plane focuses along the vertical line S,
while a fan of rays in the tangential, or vertical, plane tt
/
focuses
along the horizontal line T.
θ
T
S
optical axis
c
h
ie
f ra
y
Figure 28. Cross sections of the paraboloidal surfaces S and T.
Figure 29. Geometry for the formation of the image of an
extended object that shows the curvature of ﬁeld.
GEOMETRICAL OPTICS 1791
toward the edges of the ﬁeld. As an example of an optical
system that is free from distortion is the pinhole camera,
shown in Fig. 31, for which
tany
/
tany
=const (58)
11.7. Chromatic Aberration
The previous ﬁve aberrations occur for monochromatic
light; when nonmonochromatic light is used, the variation
of the index of refraction with the frequency of light causes
chromatic aberration, which results in the creation of one
image per color present in the illumination. Normally, the
index of refraction is larger at higher frequencies, so that,
for a thin lens, the focal distance (33) is shorter at higher
frequencies. Figure 32a shows the variation in focal dis
tance for the image of the same object that is illuminated
with white light and located at large distance from the
lens along the optical axis. The axial distance between the
images V and R is the longitudinal chromatic aberration.
There is also a lateral chromatic aberration, shown in Fig.
32b, for the image of an offaxis object. One of the simplest
methods to correct chromatic aberration is to put in con
tact two lenses made of different materials and with dif
ferent powers, as shown in the achromatic doublet of
Fig. 32c. The powers are chosen with opposite sign
and different magnitudes so that the net power is either
positive or negative. The materials and the shapes of the
lenses are chosen so that the dispersion of one lens com
pensates for the dispersion of the other and, within a cer
tain frequency interval, the doublet is not affected by
chromatic aberration.
12. RELATIONSHIP BETWEEN GEOMETRIC OPTICS
AND ELECTROMAGNETIC THEORY
12.1. Eikonal Equation
An elegant introduction to geometrical optics and its re
lationship to the electromagnetic theory is found in Refs. 2
and 12, from which part of the material contained in this
section is based.
The German Gustav Kirchhoff (1824–1887) made the
ﬁrst significant effort to derive geometric optics from a
wave theory. He expressed mathematically the principle of
Huygens, which resulted in the Kirchhoff–Huygens prin
ciple. Even though the Kirchhoff–Huygens principle con
tains some mathematical inconsistencies, it is remarkable
that it is successful in so many applications. In particular,
in 1882 Kirchhoff [20] showed that when the wavelength
l0, the ﬁeld given by the Kirchhoff integral in the pres
ence of an obstacle approaches the prediction obtained
from geometric optics: the existence of a sharp transition
(a) (b) (c)
Figure 30. Example of distorted images. The correct image of a
wire mesh object is shown in (a), while (b) represents an image
affected by barrel distortion, and (c) depicts an image affected by
pincushion distortion.
Figure 31. A pinhole camera provides a lateral magniﬁcation
that is constant.
Figure 32. (a) Longitudinal chromatic aberration resulting from
different focal distances for different colors for the images of an
object located along the optical axis and at large distance from the
lens; (b) lateral chromatic aberration for an object located off the
optical axis; (c) achromatic doublet that removes chromatic aber
ration.
1792 GEOMETRICAL OPTICS
between dark and lit regions. Eventually, the point of view
of geometric optics as a limiting case of the electromag
netic theory when l0 became accepted.
Sommerfeld and Runge [21] provided the most accepted
argument to show the connection between geometric op
tics and electromagnetic theory. They assumed that a sca
lar function u(x, y, z) represents the complex amplitude of
a component of the electric ﬁeld E, which, in the mono
chromatic case, satisﬁes the Helmholtz scalar equation
V
2
u(x; y; z) ÷k
2
u(x; y; z) =0 (59)
where V
2
is the laplacian operator, k =k
0
n=
ﬃﬃﬃﬃﬃ
em
_
o=c =2p=l is the wavevector, o is the angular fre
quency, and the medium is assumed to be linear, isotropic
and described by a permittivity e and a permeability m,
which may be functions of the position. Sommerfeld and
Runge assumed that the functional form of u(x, y, z) was a
generalization of a plane wave
u(x; y; z) =A(x; y; z)e
jk
0
C(x;y;z)
(60)
where A is the amplitude of u and k
0
C is the phase. In
particular, C(x, y, z) is a real scalar called eikonal function
and surfaces along which C (x, y, z) =const are the wave
fronts of the electromagnetic ﬁeld. It is also assumed that
A and C do not vary rapidly with changes in the position.
When (60) is substituted into (59), it is obtained that
k
2
0
(n
2
÷[VC[
2
)A÷V
2
A÷jk
0
(2VC
.
VA÷AV
2
C) =0 (61)
which requires that both real and imaginary part be
independently zero. The condition on the real part yields
[VC[
2
=n
2
÷
l
2p
_ _
2
V
2
A
A
(62)
In the limit of l0, the last term of the previous equation
vanishes, which leaves the socalled eikonal equation:
[VC[
2
=n
2
(63)
The eikonal equation is at the basis of geometric optics,
and its solutions deﬁne the wavefronts. In the following, it
is shown that the ray equation (16) may be obtained from
the eikonal equation. For this purpose, let us observe that
^
tt ==C=n deﬁnes a unit vector that is perpendicular to the
surfaces C=const (because of the meaning of the gradient
of C) and consider a trajectory r =r(s) that admits
^
tt as its
tangent unit vector, where s measures the length of the
trajectory from an arbitrary point along it. Then, it is pos
sible to write
n
^
tt =n
dr
ds
==C (64)
from which ray trajectories could be derived from the
knowledge of C. With a few more steps, the dependence
on C is removed from the previous equation. In fact, the
directional derivative along r(s) of the previous equation
yields
d
ds
n
dr
ds
_ _
=
d
ds
(=C) =
dr
ds
.
=(=C) =
=C
n
.
=(=C)
=
1
2n
=([=C[
2
) =
1
2n
=n
2
==n
(65)
So, ﬁnally, one obtains the ray equation (16) that was in
dependently derived from Fermat’s principle. Therefore,
the path to point out the relationship between the electro
magnetic theory and geometric optics occurs in three steps:
(1) one shows that when l0 (highfrequency approxima
tion) the description of the electromagnetic ﬁeld provided
by (60) leads to the eikonal equation (63); (2) one observes
that the eikonal equation deﬁnes surfaces C=const that
are perpendicular to the ray trajectories; and (3) the ray
equation (16) is derived from the eikonal equation.
Classical geometric optics provides a description of
light phenomena that is based only on the geometric con
cept of a ray. However, in order to use geometric optics to
make quantitative statements, this theory needs to be ex
tended to introduce the concepts of amplitude, phase, and
polarization. These extensions are explored in the follow
ing two sections, which are based on Ref. 22.
12.2. Amplitude Relation
The notion of light intensity may be introduced into the
classical geometric optics by imposing the conservation of
the energy ﬂux inside a tube of rays. For example, let us
consider the isotropic spherical wavefront shown in
Fig. 33, where two different cross sections of areas dA
0
and dA are located at some reference points s =0 and s,
respectively. The radiation density S
0
at s =0 and S at s
Figure 33. Tube of rays for an isotropic spherical source. (Con
stantine A. Balanis, Advanced Engineering Electromagnetics,
copyright r1989 by John Wiley & Sons, Inc. This material is
used by permission of John Wiley & Sons, Inc.)
GEOMETRICAL OPTICS 1793
are related by
S
0
dA=SdA (66)
because it is assumed that no energy ﬂows through the lat
eral side of the cylinder. For an electromagnetic wave in the
farzone ﬁeld of the source, the relationship between the
electric ﬁeld E(r, y, j) and the radiation density S(r, y, j)
is
S(r; y; j) =
1
2Z
[E(r; y; j)[
2
(67)
Hence, combining the last two equations, one obtains
[E[
[E
0
[
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃ
dA
0
dA
_
(68)
The shape of the eikonal surface combined with this equa
tion determines how the amplitude of the electric ﬁeld
changes between eikonal surfaces. As an example, let us
consider the astigmatic eikonal surface, which is more
general than a spherical one, shown in Fig. 34. For such a
surface, the amplitude of the electric ﬁeld on one surface
relative to another varies according to [22]
[E[
[E
0
[
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃ
dA
0
dA
_
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
r
1
r
2
(r
1
÷s)(r
2
÷s)
_
(69)
This equation reduces to
[E[
[E
0
[
=1 (70)
for a plane wave, which corresponds to r
1
=r
2
=N. Equa
tion (69) reduces to
[E[
[E
0
[
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
dA
0
dA
=
_
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
r
0
r
0
÷s
_
(71)
for a cylindrical wavefront, which corresponds to r
1
=r
0
,
r
2
=N. For a spherical wavefront, which corresponds to r
1
=r
2
=r
0
, Eq. (69) becomes
[E[
[E
0
[
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
dA
0
dA
=
_
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
r
0
r
0
÷s
_
(72)
The relationship (69)–(72) deﬁne the correct way to re
late to the amplitudes of the electric ﬁeld at one surface
relative to another, when the frequency o is sufﬁciently
large. However, these relationships do not include phase
and polarization.
12.3. Phase and Polarization
Phase and polarization may be introduced into classical
geometric optics using the approach of Luneburg [11] and
Kline [23], which is referred to as Luneburg–Kline high
frequency expansion. According to their approach, when
the frequency u` is sufﬁciently large, the electric ﬁeld may
be written using the following series:
E(x; y; z; o) =e
÷jb
0
C(x;y;z)
o
m=0
E
m
(x; y; z)
(jo)
m
(73)
If we limit our attention only to solutions that (1) are of
the ﬁrst order (i.e., m=0) and (2) take the form
E(s) =e
÷jb
0
c(s)
E(s =0) (74)
we can prove [2] that for the geometry of Fig. 34
E=E
0
e
÷jk
0
C(0)
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
r
1
r
2
(r
1
÷s)(r
2
÷s)
_
e
÷jk
0
s
(75)
where s =0 is the reference point. Equation (75) provides
the extended geometric optics description of an electric
ﬁeld where the polarization comes fromthe vector E
0
at an
arbitrary reference point s =0 along a trajectory described
by a normal to the eikonal surface C. The initial phase is
given by e
÷jk0C(0)
. The amplitude varies according to the
squareroot term, for an astigmatic wavefront, and the
last exponential term accounts for the appropriate phase
delay. When s =–r
1
or s =–r
2
, Eq. (75) is singular because
these are the locations of the two caustic lines shown in
Fig. 34; hence this equation should not be used at caustics
or very close to them.
Figure 34. Astigmatic tube of rays. (Constantine A. Balanis, Ad
vanced Engineering Electromagnetics, copyright r 1989 by John
Wiley & Sons, Inc. This material is used by permission of John
Wiley & Sons, Inc.)
1794 GEOMETRICAL OPTICS
13. APPLICATION OF GEOMETRIC OPTICS TO THE
DESIGN OF ANTENNAS
The Cartesian surfaces described in Section 3 have prop
erties that are exploited in the design of a class of anten
nas called reﬂector antennas. To understand how basic
geometrical optics concepts are transferred to the opera
tion of these antennas, let us refer to Fig. 35. In Section 3,
we mentioned that rays emitted from a point source lo
cated at the focus F become parallel to the axis z after a
reﬂection on the parabolic surface. Hence the circular
wavefront out of F is transformed into a plane wavefront.
The same conclusion may be obtained directly from the
definition of a parabola. The parabola is the locus of points
P that have the same distance from a straight line, the
directrix d, and a point, the focus F, which yields
FP=GP (76)
Consider the line MN, which is perpendicular to the axis z,
and its intersection W with a ray reﬂected at P. One can
write
FP÷PW=FP÷(GW÷GP) =GW (77)
Thus, since the path length PF÷PW is the same for all
points P on the parabola, the line MN is a wavefront. Ad
ditionally, the image of the focus is the directrix because
the wavefront at W appears to have originated at G.
In practice, the parabola shown in Fig. 35 must be ter
minated somewhere, such as at points A and B indicated
in the ﬁgure. This geometry should be interpreted as the
cross section of two antennas: the paraboloidal and the
cylindrical parabolic reﬂectors. A paraboloid is obtained
by the rotation of a parabola around its axis. If an isotropic
source is located in the focus of the parabola, the portion of
the spherical wavefront that is captured by the paraboloid
is transformed into a plane wavefront. The cylindrical
parabolic reﬂector has a line source located in the focus
and transforms a cylindrical wavefront into a plane wave
front. For both antennas, the power due to a source is
concentrated mostly along the direction of the axis. For
both antennas, the actual ﬁeld received by an observer
along the axis is the superposition of a plane wavefront
and the direct ﬁeld from the source. Improvements on the
design of reﬂector antennas include the use of subreﬂec
tors. For example, in the Cassegrain design a hyperbolic
subreﬂector faces the main parabolic reﬂector with the
advantage of an overall more compact design because the
feedpoint is moved closer to the vertex of the parabola [24].
Similar to optical elements with spherical surfaces, reﬂec
tor antennas also are affected by aberrations.
Another result of geometric optics that ﬁnds applica
tion in the design of antennas is the Luneburg lens intro
duced in Section 5.4. In view of the relationship between
electromagnetic theory and geometric optics, it is impor
tant to point out that many results from geometric optics
are useful for the design of systems that operates at fre
quencies different from those of visible light. Such a per
spective that unites optical principles with the design of
microwave antennas are provided in Refs. 25 and 26.
14. EXTENSION OF GEOMETRIC OPTICS TO INCLUDE
DIFFRACTION PHENOMENA
Ray tracing, which stems from geometric optics, is not
only used for the design of optical systems but also has
applications in other engineering ﬁelds. For example,
many applications related to computer graphics are based
on ray tracing [27], and software is available to write pro
grams that exploit ray tracing [28].
The link between geometric optics and electromagnetic
theory not only is a topic of academic interest but also
proves useful in many engineering applications. Ray trac
ing is a very convenient way to simplify the complex prob
lem of the propagation of a wavefront. In fact, many
electromagnetic problems are simpliﬁed by making a
highfrequency approximation that reduces the complex
study of the propagation of a wavefront to the simpler ray
tracing of the normal to the wavefronts. Electromagnetic
problems that beneﬁt from ray theories include the compu
tation of radar cross sections and diffraction by objects and
predictions of the path loss for radiocommunications [29].
The ray tracing that is derived from geometric optics is
concerned only with the direction of propagation of the
normal to the wavefront. For an electromagnetic applica
tion, other quantities such as polarization and the inten
sity associated with the wavefront are also important.
While polarization is easily accounted for, a correct eval
uation of the intensity requires an extension of geometric
optics to account for diffraction. Hence many efforts have
been devoted to the development of extensions to geomet
ric optics [30]. Without going into the details of these ex
tensions, which would be beyond the purpose of this
article on geometric optics, only few of these efforts are
mentioned. Keller [31] developed a theory that includes
diffraction by formulating a generalization of Fermat’s
principle. Keller’s geometric theory of diffraction (GTD)
was later improved by uniform theories such as the
V
d
F
A
B
P
W
z
G
M
N
Figure 35. A parabola.
GEOMETRICAL OPTICS 1795
uniform theory of diffraction (UTD) of Kouyoumjian and
Pathak [32,33] and the uniform asymptotic theory (UAT)
of electromagnetic diffraction by Lee and Deschamps [34].
These theories may fail near caustics, so more recently
Tiberio et al. [35] developed the incremental theory of dif
fraction (ITD), which avoids problems near caustics and
presents other advantages.
BIBLIOGRAPHY
1. B. E. A. Saleh and M. C. Teich, Fundamental of Photonics,
Wiley, New York, 1991.
2. M. Kline and I. W. Kay, Electromagnetic Theory and Geomet
rical Optics, Interscience, New York, 1965.
3. E. T. Whittaker, A History of the Theories of Aether and Elec
tricity, T. Nelson and Sons, London and Edinburgh, 1953.
4. A. I. Sabra, The Optics of Ibn AlHaytham, The Warburg In
stitute, Univ. London, 1989.
5. C. Huygens, Treatise on Light, Univ. Chicago Press, Chicago,
1962; reprint of London 1912 edition, translated into English
by Silvanus P. Thompson.
6. I. Newton, Opticks or a Treatise of the Reﬂections, Refractions,
Inﬂections and the Colours of Light, Dover Publications, New
York, 1952.
7. J. C. Maxwell, A Treatise on Electricity and Magnetism, Ox
ford, 1873.
8. M. V. Klein, Optics, Wiley, New York, 1970.
9. M. Herzberger, Modern Geometrical Optics, Interscience,
New York, 1958.
10. F. L. Pedrotti and L. S. Pedrotti, Introduction to Optics, Pren
ticeHall, Upper Saddle River, NJ, 1993.
11. R. K. Luneburg, Mathematical Theory of Optics, Univ. Cali
fornia Press, Berkeley and Los Angeles, 1966.
12. M. Born and E. Wolf, Principles of Optics, Pergamon Press,
London, 1965.
13. B. Rossi, Optics, AddisonWesley, Reading, MA, 1957.
14. A. E. Siegman, Lasers, University Science Books, Mill Valley,
CA, 1986.
15. W. J. Smith, Modern Optical Engineering, McGrawHill, New
York, 2000.
16. F. A. Jenkins and H. E. White, Fundamental of Optics, 4th
ed., McGrawHill, New York, 1976.
17. Y. Matsui and K. Nariai, Fundamentals of Practical Aberra
tion Theory, World Scientific, Singapore, 1993.
18. W. H. A. Fincham and M. H. Freeman, Optics, 9th ed., But
terworths, London, 1980.
19. M. Cagnet, M. Franc
,
on, and J. C. Thrierr, Atlas of Optical
Phenomena, SpringerVerlag, Berlin, 1962.
20. G. R. Kirchhoff, Zur theories der lichtstrahlen, Ann. Phys.
2:663 (1883).
21. A. Sommerfeld and J. Runge, Anwendung der ve
ktorrechnung auf die grundlagen der geometrischen optik,
Ann. Phys. 35:277–298 (1911).
22. C. A. Balanis, Advanced Engineering Electromagnetics, Wi
ley, New York, 1989.
23. M. Kline, An asymptotic solution of Maxwell’s equations, in
The Theory of Electromagnetic Waves, Interscience, New
York, 1951.
24. J. D. Kraus, Antennas, McGrawHill, Boston, 1988.
25. S. Cornbleet, Microwave Optics, Academic Press, London, 1976.
26. S. Cornbleet, Microwave and Optical Ray Geometry, Wiley
Interscience, New York, 1984.
27. A. S. Glassner, ed., An Introduction to Ray Tracing, Academic
Press, 1989.
28. N. P. Wilt, Object Oriented Ray Tracing in C÷ ÷, Wiley, 1994.
29. D. Erricolo and P. L. E. Uslenghi, Twodimensional simulator
for propagation in urban environments, IEEE Trans. Vehic.
Technol. 50(4):1158–1168 (July 2001).
30. L. B. Felsen and N. Marcuvitz, Radiation and Scattering of
Waves, IEEE Press, Piscataway, NJ, 1994.
31. J. B. Keller, Geometrical theory of diffraction, J. Opt. Soc. Am.
52:116–130 (1962).
32. R. G. Kouyoumjian and P. H. Pathak, A uniform geometrical
theory of diffraction for an edge in a perfectly conducting sur
face, Proc. IEEE, 62(11):1448–1461 (Nov. 1974).
33. D. A. McNamara, C. W. I. Pistorius, and J. A. G. Malherbe,
Introduction to the Uniform Geometrical Theory of Diffrac
tion, Artech House, Boston–London, 1990.
34. S.W. Lee and G. Deschamps, A uniform asymptotic theory of
electromagnetic diffraction by a curved wedge, IEEE Trans.
Antennas Propag. 24(1):25–34 (Jan. 1976).
35. R. Tiberio, A. Toccafondi, A. Polemi, and S. Maci, Incremental
theory of diffraction: a newimproved formulation, IEEE
Trans. Antennas Propag. 52(9):2234–2243 (Sept. 2004).
GESI ALLOYS AND DEVICES
PETER RUSSER
JOHANNFRIEDRICH LUY
Munich Technical University
Germany
1. INTRODUCTION
Silicon–germanium heterojunction bipolar transistors
(HBTs) [1–3] extend the frequency limits of siliconbased
technology. In the SiGe HBT, the base region is formed by
an epitaxially grown SiGe layer between the adjacent sil
icon layers. Because of the lower bandgap of the base re
gion also existing in the case of high base doping, a high
emitter efﬁciency is achieved. This allows a HBT design
with small base width and a low base series resistance.
The heterojunction bipolar transistor was ﬁrst suggested
in 1957 by Kroemer [4]. A ﬁrst suggestion for a SiGebase
heterojunction bipolar transistor (HBT) for highfrequen
cy applications was made in 1977 [5]. SiGe technology has
been pushed mainly as a commercial technology by IBM,
DaimlerChrysler/TEMIC and Inﬁneon. Commercially
available HBTs exhibit f
T
values greater than 50 GHz
and f
max
values in excess of 70 GHz, a minimum noise ﬁg
ure below 0.7dB at 2 GHz, 1/f noise corner frequencies
below 500Hz, and a reliability comparable to Si devices
[2]. Future trends in SiGe technology point toward the
integration of more complex circuits, technology efforts to
facilitate easier manufacturing and speed increase beyond
f
T
=200GHz [6]. The silicon–germanium heterojunction
bipolar transistor enabled the breakthrough of siliconbased
1796 GESI ALLOYS AND DEVICES
MMIC (monolithic millimeterwave integrated circuit)
technology [1,6]. Ampliﬁers and tunable oscillators or
with maximum operation frequencies beyond 80 GHz be
come feasible. Inﬁneon’s SiGe bipolar production technol
ogy B7HF is capable for highvolume production [7]. SiGe
HBTs with transit frequencies f
T
above 200GHz and max
imum oscillation frequencies of f
max
=338GHz already
have been realized [8–11].
Also, the suitability of CMOS devices for RF applica
tions continuously improves [3]. With 0.25mm technology,
f
T
values are as high as 40 GHz and should double roughly
every 3 years if trends continue [12,13]. A Kband travel
ingwave ampliﬁer as well as a 17GHz travelingwave
oscillator have already been realized with CMOS devices
[12]; 70nm CMOS have cutoff frequencies reaching
150GHz and an inverter gate delay time around 15 ps at
a supply voltage of 1.5 V [14]. The short gate delay time
makes CMOS circuits suitable for future broadband mul
timedia applications.
SiGe/SiMODFETs (modulationdoped fieldeffect tran
sistors) are high electron mobility transistors [15]. The
material system GeSi is the only one that exhibits sym
metric transport properties for complementary n and p
channel devices. For both channel types an f
max
of
100GHz has already been achieved. Silicon–germanium
MODFETs have a significant highfrequency potential
[16]. The introduction of SiGe allows the realization of n
and ptype heteroFETs (MOSFETs), both with enhanced
performance. Maximum oscillation frequencies of more
than 100GHz have been obtained with n and pMOS
FETs.
A further improvement of the RF properties of silicon
based devices will result from the introduction of copper
technology. As semiconductor devices are scaled down to
the submicron region, interconnection losses are becoming
a limiting factor. In submicrometer technology, copper in
terconnects will exhibit lower resistivity and better reli
ability than aluminium interconnects. Copper reduces line
and via resistances by at least 50%, increases the maxi
mum current density by a factor of 2.5–3.0, and maintains
the same leakage currentproperties as the Al intercon
nects. The damascene process enables the fabrication of
submicrometer copper interconnects [17,18].
New device concepts try to exploit unavoidable delay
times and drift ﬁelds in heterojunction devices in order to
generate a resonant behavior of the device at frequencies
beyond the current gain cutoff frequency [19–21].
Silicon as a substrate for millimeterwave monolithi
cally integrated circuits was suggested in 1981 by RCA
[22]. Since 1986 in the ﬁeld of silicon monolithic milli
meterwave integrated circuits (SIMMWICs) there have
been research activities at the former AEGTelefunken
Research Institute (now DaimlerChrysler Research Cen
ter) and at the Technische Universita¨ t Mu¨ nchen [23–29].
Up to now SIMMWICs for frequencies up to above
100GHz already have been fabricated, and the suitabili
ty of silicon as the base material for monolithic integrated
millimeterwave circuits has been successfully demon
strated. Monolithic integration of solidstate devices pro
vides the possibility of lowcost production, improved
reliability, small size and light weight, and easy assembly.
The linear passive parts of SIMMWICs may be realized
in planar circuit technology. The fundamental transmis
sionline structures used SIMMWIC design are microstrip
lines, slotlines, coplanar lines, coplanar striplines, and
microshield lines. Based on these fundamental geometric
structures, the planar circuit elements are designed.
These planar circuit elements include transmissionline
discontinuities, planar resonators and antennas as the
basic structures.
In the frequency region above 60 GHz, SIMMWICs
with dimensions of only a few millimeters may also in
clude planar antenna structures. The integration of the
antenna structures allows the direct coupling of
SIMMWICs to the radiation ﬁeld.
Monolithic integrated millimeterwave circuits based on
silicon and SiGe will give new options for millimeterwave
sensor and communication applications. Compared with
microwavebased systems, millimeterwavebased sys
tems offer the following advantages:
*
Availability of broader frequency bands
*
Higher gain and smaller dimensions of antennas
*
Lower weight and smaller size of the components
*
Higher resolution for sensor applications
*
Atmospheric attenuation that may prevent inter
ference between cells
A broad application of millimeter waves in sensors and
communications has been hampered up to now because of
the high costs of millimeterwave components. This situ
ation may change in the future with the availability of
lowcost monolithic integrated components based on a sili
con and SiGe technology.
2. SILICON AS THE BASE MATERIAL
Table 1 compares the data for Si and GaAs. At 90 GHz, the
dielectric loss tangents of Si and GaAs are within the same
order of magnitude, and for microwave circuits, silicon
substrates with a specific resistance of 10,000 O· cm are
available. For this material, the conductor losses due to
the skin effect dominate the loss contributions of planar
circuits, whereas the substrate losses in the silicon ac
count only for a minor part. The electron mobility of GaAs
is 6 times higher than the electron mobility of Si. This
yields a correspondingly higher f
T
value for bipolar tran
sistors and ﬁeld effect transistors. On the other hand, the
Table 1. Parameters ofSi andGaAs
Si GaAs
Dielectric constant e
r
=11.7 e
r
=12.9
Specific resistance 410
4
O· cm 410
6
O· cm
Dielectric loss
Factor (90GHz) 1.310
÷3
0.7 10
÷3
Thermal conditions 1.45W cm
÷1
K
÷1
0.46Wcm
÷1
K
÷1
Electron mobility 700cm
2
/V· s 4300cm
2
/V· s
High ﬁeld drift velocity 10
7
cm/s 6.10
6
cm/s
Density 2.33g/cm
3
5.32g/cm
3
GESI ALLOYS AND DEVICES 1797
high ﬁeld drift velocity of Si and GaAs are in the same
order. The thermal conductivity of silicon is 3 times higher
than the thermal conductivity of GaAs. From this point of
view, silicon is advantageous for power circuits. Silicon
and germanium can be mixed in alloys with arbitrary
mixing ratios because of their identical lattice structure
and their similar lattice constants. The atoms are distrib
uted statistically on free places and the lattice structure is
maintained. The lattice constant a
SiGe
can be calculated
by the linear interpolation between the lattice constants
a
Si
of silicon (0.5431 nm) and a
Ge
of germanium
(0.5658 nm):
a
si
l÷x
Gex
=(1 ÷x)a
Si
÷xa
Ge
(1)
For the realization of npn heterojunction bipolar transis
tors, the pseudomorphic growth of strained Si
1–x
Ge
x
layers
without misﬁt dislocations is of essential importance.
Strained SiGe layers may be grown epitaxially on silicon
substrate. If the thickness of the epitaxial layers is kept
below a critical thickness, then the mismatch between the
alloy and the silicon substrate is accommodated elastically.
The strain grows with increasing lattice mismatch and
germanium content [30,31]. The lattice mismatch is de
ﬁned as
Z =
a
Si
1÷x
Ge
x
÷a
Si
a
Si
(2)
and the lattice mismatch between pure germanium and
silicon is 4.2%.
Figure 1 shows that the energy gap of the alloy is de
pendent not only on the germanium content but also on
the strain situation. The conduction band minima in un
strained silicon correspond to sixfold degenerate valleys
located along {100] directions in k space. In case of strain,
the sixfold degenerate conduction band valleys are sepa
rated and the four valleys oriented normal to the hetero
junction are lowered in energy and the remaining two
valleys are raised in energy. The overall bandgap shrink
age is increased over that of unstrained material with the
same germanium content.
A uniﬁed SiGe/SIMMWIC technology uses this band
gap engineering to develop SiGe HBTs, MODFETs, and
MOSFETs [28]. This uniﬁed technology also allows the
fabrication of Schottky diodes, IMPATT diodes, and PIN
diodes (Fig. 2).
3. ELECTRONIC DEVICES
3.1. SiGeBase HBTs
Although silicon bipolar transistors with transient fre
quencies above 50 GHz are possible in principle, these de
vices would exhibit too high a base resistance and too low
a breakdown voltage. A promising device for millimeter
wave applications is the SiGe base heterojunction bipolar
transistor (HBT) [1]. The heterojunction bipolar transistor
ﬁrst has been suggested in 1957 by Kroemer [4]. A ﬁrst
suggestion for a SiGebase HBT for highfrequency appli
cations was made in 1977 [1,5]. Detailed investigations of
the millimeterwave SiGebase HBT have been presented
[32–40].
In SiGe HBT, the base region is formed by an epitax
ially grown SiGe layer between the adjacent silicon layers.
With increased germanium content in the base, the band
gap of the base region is reduced. Because of the lower
band gap of the base region also in the case of high base
doping a high emitter efﬁciency is achieved. The base of a
HBT can thus be more heavily doped than the base of a
silicon bipolar transistor to reduce the base resistance
without the negative impact on the current ampliﬁcation.
Figure 3 shows the cross section of a heterojunction bipo
lar transistor. The HBTs are fabricated using a self
aligned emitter–base structure together with selective
epitaxial growth of the SiGebase layer. Figure 4 shows
a SiGe HBT embedded in a coplanar line crossing.
The effect of reduced barrier for the electron injection
from the emitter into the base and the corresponding de
sign freedom regarding the emitter and the base doping
becomes evident if the current gain of a bipolar junction
transistor is compared with the current gain of a hetero
junction bipolar transistor. The current gain b of a con
ventional bipolar transistor is given by
b =
D
nE
N
E
w
E
D
pB
P
B
w
B
(3)
where D
nE
and D
pB
are the diffusion constants of electrons
and holes in the emitter and base regions respectively, N
E
1.2
1.1
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0 0.2 0.4
Ge composition x
E
n
e
r
g
y
g
a
p
(
e
V
)
0.6 0.8 1.0
Unstrained
Bulk Alloy
Si
0.5
Ge
0.5
(001)
substrate
Ge (001) substrate
Si (001) substrate
strain split V.B.
3/2 ± 1/2
3/2 ± 1/2
Figure 1. Energy gap in Si
1–x
Ge
x
layers as a function of the Ge
composition x.
1798 GESI ALLOYS AND DEVICES
and P
B
are the emitter and base doping concentrations, w
E
is the emitter thickness, and w
B
is the base thickness. For
the HBT after Kroemer [4] the current gain b is given by
b =
D
nE
N
E
w
E
D
pB
P
B
w
B
e
DE=kT
(4)
where DE is the bandgap difference between Si and SiGe.
The exponential dependence of the minority carrier injec
tion rate across the emitter–base barrier on the band gap
difference yields for a Ge content of 20% a value of the
exponential term of more than 1000. Thus a high base ef
ﬁciency is obtained even in the case of a high emitter dop
ing. Due to the high base doping, low base sheet
resistances with base widths down to 10 nm are attain
able. A typical highfrequency silicon bipolar transistor
with a base width of less than 100 nm exhibits base sheet
resistances in the order of 10 KO/&, whereas SiGe HBTs
reach base sheet resistances as low as 1kO/&for a 20nm
doped base [28]. Therefore, a high current gain and a low
base resistance are possible due to extremely low base
width.
The unity current gain cutoff frequency f
T
is given by [2]
f
T
= r
e
(C
eb
÷C
cb
÷t
b
÷t
e
÷t
c
[ ]
÷1
(5)
where C
eb
and C
cb
are the emitter–base and collector–base
depletion capacitances; t
b
, t
e
, and t
c
are the base, emitter,
and collector transit times, respectively; and the emitter
resistance r
e
is given by
r
e
=
kT
eI
c
(6)
In bipolar junction transistors (BJTs), f
T
is usually limited
by the base transit time t
b
given by
t
b;Si
=
w
2
b
2D
nb
(7)
where D
nb
is the electron diffusivity in the base region and
w
b
is the base width. In the SiGe HBT the Ge grading
across the base decreases t
b
by [2]
t
b;SiGe
t
b;Si
=
2
¯ZZ
kT
DE
g;Ge(grade)
1 ÷
kT
DE
g;Ge(grade)
1 ÷e
÷DE
g;Ge(grade)
=kT
_ _
_ _
(8)
where the bandgap grading term DE
g,Ge(grade)
is given by
DE
g;Ge(grade)
=DE
g;Ge(w
b
)
÷DE
g;Ge(0)
(9)
and
¯ZZ =
D
nb(SiGe)
D
nb(Si)
(10)
Schottky
diode
IMPATT
diode
SiGe
HBT
PIN
diode
Buried layer
MBE
Mesa etch
Base contact
Mesa etch
Passivation
Bottom contact
Metallization
Figure 2. Technology.
p Epi base
n
+
Poly emitter
Nitride
Oxide
n Epi
n
+
Subcollector
n Substrate
p
+
Poly
n
+
Figure 3. Cross section of SiGe HBT.
Figure 4. SiGeHBT in coplanar line crossing.
GESI ALLOYS AND DEVICES 1799
The maximum oscillation frequency f
max
of the HBT is
given by
f
max
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
f
T
8pC
bc
r
b
¸
(11)
where f
T
is the transit frequency, r
b
is the base resistance,
and C
bc
is the base collector capacity. In SiGe HBTs the
lower bandgap allows for high base doping and therefore
low base resistance. That increases f
max
.
The fabrication of a heterojunction bipolar SiGe tran
sistor with f
T
values above 200GHz has already been re
ported [1,8–11,28,35,41]. For most circuit applications,
f
max
is the more important parameter [1,28] and coplanar
MMICs based on SiGe HBTs with f
max
values as high as
338GHz are under development [8,10,42–44].
For an efﬁcient circuit design an accurate largesignal
transistor model is required. Gummel and Poon intro
duced a bipolar transistor model based on a charge–con
trol relation, derived from basic physical considerations,
linking the dominant component of collector current, emit
ter and collector junction voltages, and base charge
[45,46]. The equivalent circuit of the Gummel–Poon mod
el is depicted in Fig. 5. The Gummel–Poon model is the
basis for most advanced nonlinear BJT models; however, it
is not suitable for reliable design of circuits operating at
either high current densities or low voltages. Several au
thors proposed largesignal models for the SiGe hetero
junction bipolar transistor [47–50].
The physicsbased compact bipolar transistor model
HICUM (highcurrent model) uses the smallsignal pa
rameters, such as the depletion capacitances and the tran
sit times, as basic variables, which are described
accurately as a nonlinear function of bias, temperature,
and geometry. This results in continuously differentiable
relations for charges and transfer current and accurate
modeling of the largesignal switching behavior, including
harmonic distortion at high frequencies [51,52]. A detailed
description of the HICUM model can be found in Ref. 53.
The socalled vertical bipolar intercompany (VBIC) model
was developed as an industry standard to improve deﬁ
ciencies of the SPICE Gummel–Poon model that have be
come apparent over time because of the advances in BJT
process technology. The VBIC model is still based on the
Gummel–Poon formulation; however, it includes improved
modeling of the Early effect, quasisaturation, substrate
and oxide parasitics, avalanche multiplication, and tem
perature behavior [54]. A VBIC model for the analysis of
twotone intermodulation distortion behavior of SiGe
HBTs takes into account all important effects for accurate
modeling of largesignal behavior, including selfheating,
weakavalanche multiplication, quasisaturation effects,
and all device capacitances [55].
The SGH (SiGe HBT) largesignal model combines
elements of the VBIC and HICUM models [56,57]. Figure 6
shows the equivalent circuit of the SGH model. The SGH
model accurately describes the SiGe HBT behavior up to
40 GHz. In the SGH model the transport current
I
CC
=
I
f
÷I
r
q
b
(12)
is related to the normalized stored charge q
b
, given by
q
b
=q
1
÷q
2
(13)
with
q
1
=
Q
p0
÷HJE
.
Q
je
÷HJC
.
Q
jc
Q
p0
(14)
q
2
=
Q
f 0
÷HE
.
Q
fe
÷HC
.
f
c
.
Q
f0
÷Q
fb
Q
p0
(15)
The coefﬁcients HJE, HJC, HE, and HC, introduced sim
ilarly to the HICOM model parameters, are determined by
parameter extraction. The empirical smoothing function f
c
is given by
f
c
=
V
ce
VHC
_ _
NHC
(16)
where VHC and NHC are determined by parameter ex
traction. The default smoothing function is f
c
=1.
For an accurate simulation of the intermodulation per
formance of semiconductor devices, the modeling of the
C
jci
C
bc
C
be
R
e
R
b
R
c
E′
B′
C′
C
B
E
C
je
I
bcn
I
bci
I
bei
I
ben
I
cc
C
jcx
Figure 5. Gummel–Poon transistor model.
C
JCX
R
BX Bx
Bi
Ci C
T
Ei
E
P1 P2
C
JE
R
BI
R
E
I
th
R
C
I
ph
C
ph
L
ph
R
th
C
th
B
I
BCI
I
BCI
I
BCN
I
CC
= I
f
− I
r
I
BEN
I
K C
DE
C
DC
C
JCI
1Ω
Figure 6. SiGe HBT largesignal equivalent circuit [56].
1800 GESI ALLOYS AND DEVICES
dynamic selfheating is required [58]. A method to extract
the thermal time constants of HBTs from Sparameter
data has been given by Sinnesbichler and Olbrich [59].
The noise ﬁgure of a bipolar transistor may be approx
imated by [60]
F =1 ÷
1
R
S
r
b
÷
r
e
2
÷
1
b
÷
f
2
f
2
T
_ _
r
2
b
÷R
2
S
2r
e
_ _
(17)
where F is the noise ﬁgure, R
s
is the source impedance,
and b is the current gain. The high f
T
and the low base
resistance of SiGe HBTs yield a low noise ﬁgure. SiGe
HBT noise ﬁgures as low as 0.5dB at 2 GHz, 0.9 dB at
10 GHz, and 2 dB at 20 GHz have been reported [61].
Lowfrequency noise in active devices is a serious de
sign constraint for millimeterwave oscillator and mixer
applications. Since the lowfrequency noise is upconverted
by the inherent nonlinearities of the active device, the low
frequency noise of the active device has to be minimized.
The lowfrequency noise in millimeterwave Si/SiGe HBTs
has been investigated [62–64]. For a SiGe HBT with an f
T
of 43 GHz, Cressler et al. have reported a corner frequency
of 373Hz for the 1/f noise at a bias current of 2.25 mA [64].
However in real circuit applications, the bias currents are
at least one order of magnitude higher. Usually the corner
frequency increases proportional to the square of the bias
current.
An accurate extraction method for the 1/f noise param
eters of BJTs and FETs is presented in Ref. 65. Figure 7
shows the BJT equivalent circuit with input and output
terminations and noise sources. Combining the nonlinear
largesignal models with the noise models allows nonlin
ear noise modeling. This is a requirement for accurate
phase noise modeling of oscillators.
3.2. Si/SiGeHetero FETs
Si/SiGe heterostructure technology may also be combined
with Si CMOS technology [36,66]. Most of the Si/SiGehe
teroFETs (MOSFETs) are of the modulationdoped type.
These socalled MODFETs exhibit a thin quantumwell
layer with a thickness of 5–30 nm. Within the quantum
well layer, a twodimensional electron or hole gas with
collisionfree carriers exists. The nquantum wells are Si
layers, whereas the pquantum wells are formed by SiGe
layers. Figure 8 shows the layer structure of a ntype
Sichannel Si/SiGe MOSFET.
The introduction of SiGe in ﬁeld effect transistors al
lows the realization of n and ptype heteroFETs. Tensile
strained Si ﬁlms on strainrelieved SiGe buffer layers en
able the formation of electron quantum wells that exhibit
enhanced electron mobilities. Compressively strained
SiGe or even pure Ge layers can be used to create two
dimensional hole channels. The combination of n and
pSiGe MOSFETs and a complementary circuit arrange
ment promises the realization of a novel heteroCMOS
generation [36,67]. This approach requires a proper SiGe
buffer layer.
The epitaxial layers for ntype MODFETs—deposited
on a high resistivity substrate—start with a relaxed buffer
whose Ge content is linearly graded to 40–50%. The core of
the layer structure is the 7–10nm tensilestrained Si
channel embedded into undoped Si
1–x
Ge
x
spacers that sep
arate the carrier supply layers from the channel (Fig. 9a).
Because of a Ge content of 40–50% in the SiGe layers,
a high conduction band offset of 0.24 eV is achieved and
the quantum well enables electron mobilities up to
2800 cm
2
/V· s [68]. Schottky or junction gates may be
realized. The ptype MODFET employs a Ge channel
(Fig. 9b). A doublesided doped Ge channel is on top of
an artiﬁcial Si
0.4
Ge
0.6
substrate, accomplished by the com
positionally graded strain relieved buffer. With the layer
design values given in Fig. 9b, a 2D hole gas mobility of
1870 cm
2
/V· s is extracted from magneticﬁelddependent
Hall measurements.
The RF performance is deﬁned by f
T
extrapolated from
the current gain h
21
and by f
max
extrapolated from the
maximum available gain or the unilateral gain. For nSi
channel MODFETs, f
T
=65GHz and f
max
=120 GHz (non
deembedded still 110 GHz) for technologically relaxed gate
lengths around 0.2 mm are reported [68,69]. For pGe
R
B
R
E
R
bc
R
ce
R
be R
in
B
R
c
<i
nRC
2
> <i
nRB
2
>
C
R
L
<
i
n
R
L
2
>
<
v
n
L
2
>
<
i
n
c
2
>
<
i
n
b
2
>
<
i
n
R
E
2
>
<
i
n
R
i
n
2
>
E
l
c
c
Figure 7. BJT equivalent circuit with noise sources [65].
15 nm Si Cap
40 nm Si
0.7
Ge
0.3
Cap spacer
12 nm Si
0.7
Ge
0.3
n
+
doping
10 nm Si
0.7
Ge
0.3
spacer
25 nm Si 2 DEG (Si)
0.5 µm Si
0.7
Ge
0.3
Buffer
Si Substrate
1.5 µm Si
1X
Ge
X
Graded buffer
X=0.3
X=0.05
Figure 8. Layer structure of a ntype Si/SiGehetero FET.
GESI ALLOYS AND DEVICES 1801
channel MODFETs the f
T
is up to 70 GHz and f
max
values
up to 135GHz are reported [70,71].
The SiGe MOSFETs are still not yet optimized. Most of
the MOSFET devices reported so far are mesalike. Pref
erably source and drain implantation is used. Pads are
mostly arranged on ﬁeld oxide, however, the cutoff fre
quencies of MOSFETs are already a factor of 42 above
those of traditional Si MOSFETs, owing to their much
higher mobilities. Furthermore the frequency level of
n and pMOSFETs are very similar, which supports the
potential of SiGe for the novel HCMOS generation.
3.3. Future Device Concepts
Future development of SiGe MMICs for operation up to
100GHz stimulates the search for even faster transistors.
A ﬁrst approach might be to change the base SiGe proﬁle
from a boxshaped to a triangular curvature. A small im
provement in f
max
can be expected. However, due to a
rough estimation, f
max
values should be a factor of 5 higher
than the operational frequency. Therefore, about 300 GHz
f
max
is required for 60GHz SiGe MMICs. The pure HBT
principle will not be sufﬁcient to reach these frequencies.
Improvements are expected from a transition to a metal
base transistor or a cameltype transistor. Another prom
ising concept was proposed by Luryi [19,20]: a transistor
with an active behavior beyond the cutoff frequencies. The
combination of a graded base layer enforcing coherent
carrier transport and a base layer thickness designed to
yield a transittimeinduced phase shift of 3p/2 lead to a
resonancelike behavior at elevated frequencies above f
max
,
the resonance phase transistor (RPT).
The overall phase of the collector current in a bipolar
transistor is composed of an injection phase angle j=
j
E
÷j
B
with an emitter–base transit angle j
E
=ot
E
and
j
B
=ot
B
, respectively, and the drift delay y =ot
C
in the
base–collector junction. At frequencies below the cutoff
frequency, the phase delay of the collector current is less
than B601, as
j÷y =j
E
÷j
B
÷y =2pf (t
E
÷t
B
÷t
C
) for f =f
T
o1 (18)
In common transistor operation the phase of the collector
current increases with decreasing current gain, taking a
maximum value around 601 at frequencies close to the
cutoff frequency. However, at further increased phase de
lay, the output resistance of the transistor may become
negative and result in renewed active behavior at higher
frequencies. Such renewed active behavior was previously
discussed on the basis of transit time effects using prop
erly delayed injection of carriers into the base–collector
space charge [19,20,72,73].
In SiGe four grading steps in a 200nm base are suit
able, considering the limitations given by the critical
thickness and the germanium content as well as the re
quirements given by the energy steps to obtain coherent
carrier transport [21]. After ﬁrst experimental reports on
the increase of the unilateral gain at frequencies larger
than f
T
[74], the RPT effect has been proven experimen
tally. A current ampliﬁcation of 6.5dB was measured at a
frequency of 7 f
T
=40 GHz for an RPT structure with a
base layer thickness of 120 nm [75]. The measured current
gain is depicted in Fig. 10.
4nm 4e18 cm
3
4 nm Si cap
3nm Si
0.55
Ge
0.45
3nm Si
0.55
Ge
0.45
3nm Si
0.55
Ge
0.45
Si
0.55
Ge
0.45
relaxed buffer
4nm 1.5e19 cm
3
9nm Si channel
(a)
S D
n
+
n
+
Spacer
4nm Si cap
6nm Si
0.4
Ge
0.6
Si
0.4
Ge
0.6
relaxed buffer
5nm Si
0.4
Ge
0.6
spacer
5nm 8e18 cm
3
B
5nm 2e18 cm
3
B
9nm Ge channel
5nm Si
0.4
Ge
0.6
spacer
(b)
Pt Pt
Figure 9. Layers of (a) ntype Si/SiGe MOD
FET and (b) ptype Ge channel MODFET.
V
CE
−5V
V
CE
−10V
V
CE
−15V
f (Hz)
10
20
0
−10
−20
−30
−40
−50
1e+9 1e+10
H
2
1
(
d
B
)
Figure 10. Unilateral gain of a RPT test transistor deduced from
Sparameter measurements between 100MHz and 40GHz.
1802 GESI ALLOYS AND DEVICES
A totally different approach to enhance the injection
phase angle, proposed in 1998, uses counterphase injec
tion via resonant tunneling in a double barrier base struc
ture [21]. The initial concept of resonance phase
ampliﬁcation using delayed injection basically subsists
on keeping the base transport factor sufﬁciently high.
4. APPLICATIONS
Numerous examples of Si/SiGebased monolithically inte
grated RF chips of high performance have been presented
in the literature. A SiGe differential transimpedance am
pliﬁer with 50 GHz bandwidth for application as baseband
ampliﬁer in 40Gbps (gigabits per second) Fiberoptic re
ceivers already has been reported [76]. A monolithic inte
grated 40Gbps clock and data recovery circuit with a 1 : 4
demultiplexer has also been realized in SiGe HBT tech
nology [77] and a 40Gbps 2: 1 multiplexer and 1 : 2 demul
tiplexer, in 120nm CMOS technology [78].
Monolithic integrated HBT oscillators with oscillation
frequencies up to 74 GHz have been reported [79–81]. A
47GHz MMIC SiGe HBT oscillator with an output power
of 13.1 dBm, an efﬁciency of 13.6%, and a phase noise of –
99.31 dBc/Hz at 100kHz off carrier has already been real
ized [82]. The oscillator was realized on highresistivity
silicon (r45000 O· cm) in coplanar line technology. Apply
ing the push–push principle, a hybrid SiGe HBT micro
strip transmissionline oscillator with an output power of
÷1 dBm at 58 GHz and a singlesideband phase noise of –
108dBc/Hz at an offset frequency of 1 MHz [57,83] and a
monolithic integrated 150 GHz SiGe HBT push–push VCO
with ÷5 dBm output power at 150GHz and 30 GHz tun
ing range [84] already have been realized. Figure 11 shows
a 27 GHz SiGe HBT voltagecontrolled oscillator in micro
strip technology designed by DASA and fabricated by
DaimlerChrysler Research. The oscillator exhibits an out
put power of ÷1 dBm and a SSB phase noise of –112 dBc/
Hz at 1 MHz offset. In Fig. 12 a 38GHz SiGe HBT har
monic mixer fabricated by DaimlerChrysler Research is
depicted. A local oscillator input signal at 4.75 GHz is
required. The conversion loss amounts 23 dB. The mixer
delivers an IF output signal at 1.33 GHz. The ﬁrst active
integrated Kband antenna on highresistivity silicon sub
strate employing a Si/SiGe HBT has been presented
[85,86]. Microstrip structures represent the resonating
and radiating circuits. The HBT was mounted using the
ﬂipchip technique. An output power of 20 dBm at 24 GHz
has been achieved.
5. CONCLUSIONS
The performance of SiGe MMICs has to be viewed under
the consideration of competing approaches, especially
CMOS. Kband lownoise ampliﬁers using a 0.18mm
CMOS technology with operating frequencies of
r24 GHz, a gain of 412 dB, and a noise ﬁgure of 5.6 dB
have been reported [87]. At 26 GHz still 8.9dB of gain and
a noise ﬁgure of 6.9dB are obtained. The current con
sumption is 30 mA from a 1.8 V power supply. This results
represent the highest operation frequencies reported for
lownoise ampliﬁers in a standard CMOS process.
These results have to be compared with the latest SiGe
device and circuit results. Cutoff frequencies are now ap
proaching 400GHz, and the circuit performance at Wband
frequencies approaches or even exceeds that of III–VMMICs
[81]. The most important advantage of SiGe MMIC concepts
is probably the possible cointegration of lownoise, high cur
rent building blocks (e.g., oscillators) with CMOS circuits in
order to realize complete singlechip solutions.
The availability of SiGe HBTs with transit frequencies
above 200GHz and even higher maximum oscillation fre
quencies [8–11] will allow the development of siliconbased
monolithic integrated circuits for operation frequencies
above 100GHz. Areas of application include millimeter
wave communications and sensorics [88,89] and also
mixeddigital/analog applications at millimeterwave fre
quencies [90]. In digital ﬁberoptic receivers Schottky diode
demultiplexers [91]] in connection with SiGe HBT push–
push oscillators for clock recovery may allow to realize
electronic demultiplexers for bit rates of r160Gbps in the
future. Monolithic integrated millimeterwave sensors
an communication frontend circuits may also include
Figure 11. A 27GHz SiGe HBT VCO. (This ﬁgure is available in
full color at http://www.mrw.interscience.wiley.com/erfme.)
Figure 12. A 38GHz SiGe HBT harmonic mixer. (This ﬁgure is
available in full color at http://www.mrw.interscience.wiley.com/
erfme.)
GESI ALLOYS AND DEVICES 1803
antenna structures. An active SIMMWIC antenna for
vehicular technology in the frequency range around
76.5 GHz already has been presented [92]. This active an
tenna acts as a transceiver and is well suited for lowcost
integrated sensor systems for automotive applications.
BIBLIOGRAPHY
1. A. Gruhle, SiGe heterojunction bipolar transistors, in
J.F. Luy and P. Russer, eds., SiliconBased MillimeterWave
Devices, Springer Series in Electronics and Photonics, Vol. 2,
Springer, 1994, pp. 149–192.
2. J. Cressler, SiGe HBT technology: A new contender for
Sibased rf and microwave circuit applications, IEEE Trans.
Microwave Theory Tech. 46:572–589 (May 1998).
3. B. Meyerson, Silicon:germaniumbased mixedsignal technol
ogy for optimization of wired and wireless telecommunica
tions, IBM J. Res. Devel. 44:391–407 (May 2000).
4. H. Kroemer, Theory of a widegap emitter for transistors,
Proc. IRE 45(11):1535–1537 (1957).
5. E. Kasper and P. Russer, Verfahren zur Herstellung von bi
polaren Hochfrequenzschaltungen, German (Patent) Disclo
sure P2719464.5, (April 1977).
6. E. Kasper and J. Eberhardt, Physics of future ultra high speed
transistors part 1: HBT, Proc. 29th European Microwave
Conf. Munich, Oct. 1999, Vol. 1, pp. 151–154.
7. W. Klein and B.U. Klepser, 75GHz bipolar production tech
nology for the 21st century, Proc. 29th European SolidState
Device Research Conf. (ESSDERC’99), Leuven, Belgium,
1999, pp. 88–94.
8. G. Freeman, S. R. J. B. Jagannathan, Z. Yang, F. Guarin, and
A. Joseph, Device scaling and application trends for over
200GHz SiGe HBTs, 2003 Topical Meeting on Silicon Mono
lithic Integrated Circuits in RF Systems Digest, Grainau,
April 9–11, 2003, pp. 6–9.
9. M. Schwerd, M. Seck, T. Huttner, T. Bo¨ttner, S. Drexl,
B. Hasler, A. Mitchell, H. Helneder, H. Ko¨rner, V. Kubrak,
M. Schrenk, and R. Lachner, A manufacturable 0.35mm
150GHz f
T
SiGe:C bipolar RF technology, 2003 Topical Meet
ing on Silicon Monolithic Integrated Circuits in RF Systems
Digest, Grainau, April 9–11, 2003, pp. 6–9.
10. B. Jagannathan, M. Khater, F. Pagette, J.S. Rieh, D. Angell,
H. Chen, J. Florkey, F. Golan, D. Greenberg, R. Groves, and S.
Jeng, Selfaligned site NPN transistors with 285GHz f
max
and 207GHz f
T
in a manufacturable technolog, IEEE Elec
tron. Devices Lett. 23:258–260 (May 2002).
11. B. Jagannathan, M. Meghelli, K. Chan, J.S. Reih,
K. Schonenberg, D. Ahlgren, S. Subbanna, and G. Freeman,
3.9 ps SiGe HBT ECL ring oscillator and transistor design for
minimum gate delay, IEEE Electron. Devices Lett. 24:324–326
(May 2003).
12. T. Lee, CMOS RF: (still) no longer oxymoron, 1999 IEEE Ra
dio Frequency Integrated Circuits (RFIC) Symp. Digest, Ana
heim, CA, June 1999, pp. 3–6.
13. T. Lee, CMOS RF integrated circuits: Past, present and fu
ture, Proc. 29th European Microwave Conf. Munich, Oct.
1999, Vol. 3, pp. 12–19.
14. H.S. P. Wong, D. Frank, P. Solomon, C. Wann, and J. Welser,
Nanoscale CMOS, Proc. IEEE. 87:537–570 (April 1999).
15. E. Kasper and G. Reitemann, Physics of future ultra high
speed transistors part 2: New concepts, Proc. 29th European
Microwave Conf. Munich, Oct. 1999, Vol. 1, pp. 155–157.
16. U. Ko¨nig, SiGe/Sibipolar and quantum well transistors, re
sults and prospects, Mat. Res. Soc. Symp. Proc. 282:387–399
(1993).
17. C. Ryu, K.W. Kwon, A. L. Loke, H. Lee, T. Nogami, V. Dubin,
R. Kavari, G. Ray, and S. Wong, Microstructure and reliability
of copper interconnects, IEEE Trans. Electron. Devices,
46:1113–1120 (June 1999).
18. H. Helneder, H. Ko¨rner, A. Mitchell, M. Schwerd, and U. Sei
del, Comparison of copper damascene and aluminum rie met
allization in BICMOS technology, in paper presented at
materials for Advanced Metallization Conf. Stresa, Italy,
March 2000; reprinted in Microelectron. Eng. (2001).
19. S. Luryi, A. Grinberg, and V. Gorﬁnkel, Heterostructure bi
polar transistor with enhanced forward diffusion of minority
carriers, Appl. Phys. Lett. 63:1537–1540 (1993).
20. S. Luryi, Ultrafast operation of heterostructure bipolar tran
sistors resulting from coherent base transport of minority
carriers, Proc. ISDRS, Charlotteville, 1993, pp. 59–64.
21. H. Jorke, J. Weller, and J.F. Luy, A new transistor concept
for operating Si/SiGe devices at frequencies above 60GHz,
Proc. 1998 Advanced Research Workshop Future Trends in
Microelectronics: Off the Beaten Path, Ile des Embiez, France,
1998.
22. A. Rosen, M. Caulton, P. Stabile, A. M. Gombar, W. M. Janton,
C. P. Wu, J. F. Corboy, and C. W. Magee, Silicon as a millime
terwave monolithically integrated substrate, RCA Rev.
42:633–660 (Dec. 1981).
23. K. M. Strohm, J. Bu¨ chler, P. Russer, and E. Kasper, Silicon
high resistivity substrate millimeterwave technology, IEEE
Microwave and MillimeterWave Monolithic Circuits Symp.
Digest, Baltimore, MD, June 4–5, 1986, pp. 93–97.
24. J. Bu¨ chler, E. Kasper, P. Russer, and K. M. Strohm, Silicon
highresistivitysubstrate millimeterwave technology, IEEE
Trans. Microwave Theory Tech. 34:1516–1521 (Dec. 1986).
25. J.F. Luy and P. Russer, Siliconbased Millimeterwave Devic
es, Springer Series in Electronics and Photonics, Vol. 32,
Springer, 1994.
26. P. Russer and E. Biebl, Fundamentals, in J.F. Luy and P.
Russer, eds., SiliconBased MillimeterWave Devices, Springer
Series in Electronics and Photonics, Vol. 32, Springer, 1994,
pp. 149–192.
27. J. Buechler, Silicon millimeterwave integrated circuits, in J.
F. Luy and P. Russer, eds., SiliconBased MillimeterWave De
vices, Springer Series in Electronics and Photonics, Vol. 32,
Springer, 1994, pp. 149–192.
28. J.F. Luy, K. M. Strohm, H.E. Sasse, A. Schu¨ ppen, J. Bu¨ chler,
M. Wollitzer, A. Gruhle, F. Scha¨ fﬂer, U. Gu¨ ttich, and A. Klaas
sen, Si/SiGe MMICs, IEEE Trans. Microwave Theory Tech.
43:705–714 (April 1995).
29. P. Russer, Si and SiGe millimeterwave integrated circuits,
IEEE Trans. Microwave Theory Tech. 46:590–603 (May 1998).
30. R. People, Indirect band gap of coherently strained Ge
x
Si
1÷x
bulk alloys on (0 01) silicon substrates, Phys. Rev. 32(2):1405–
1410 (1985).
31. R. People, Physics and applications of Ge
x
Si
1 ÷x
/Si strained
layer heterostructures, IEEE J. Quantum Electron. 36:1696–
1705 (Sept. 1986).
32. S. A. Campbell and A. Gopinath, Possibility of silicon mono
lithic millimeterwave integrated circuits, in 1989 IEEE MTT–
S Int. Microwave Symp. Dig., (Long Beach, CA), June 1989,
pp. 817–819.
33. M. K. M. Willander, Optimized frequency characteristics of
Si/SiGe heterojunction and conventional bipolar transistors.
SolidState Electron. 33(2):199–204 (1990).
1804 GESI ALLOYS AND DEVICES
34. T. Won and H. Morkoc, Highspeed performance of Si/Si
1 ÷x
Ge
x
heterojunction bipolar transistors, IEEE Electron. Device
Lett. 10(1):33–35 (1989).
35. G. L. Patton, J. H. Comfort, B. S. Meyerson, E. F.Crabbe,
G. J. Scilla, E. de Fresart, J. M. C. Stork, J. Y.C. Sun, D. L.
Harame, and J. N. Burghartz, 75–GHz f
T
. SiGebase hetero
junction bipolar transistors, IEEE Electron. Device Lett.
11:171–173 (Jan. 1990).
36. U. Ko¨nig and H. Da¨ mbkes, SiGe HBTs and MOSFETs, Solid
State Electron. 38(9):1595–1602 (1995).
37. D. L. Harame, J. H. Comfort, J. D. Cressler, E. Crabbe´,
J. Y.C. Sun, B. S. Meyerson, and T. Tice, Si/SiGe epitaxial
base transistors–Part I: Materials, properties, physics, and
circuits, IEEE Trans. Electron. Devices 42:455–468 (March
1995).
38. D. L. Harame, J. H. Comfort , J. D. Cressler, E. Crabbe´,
J. Y.C. Sun, B. S. Meyerson, and T. Tice, Si/SiGe epitaxial
base transistor—Part II: Process integration and analog ap
plications, IEEE Trans. Electron. Devices 42:469–482 (March
1995).
39. R. J. E. Hueting, J. W. Slotboom, A. Pruijmboom, W. B. de
Boer, C. E. Timmering, and N. E. B. Cowern, On the optimi
zation of SiGe–base bipolar transistors, IEEE Trans. Electron.
Devices 43:1518–1523 (Sept. 1996).
40. D. Behammer, J. N. Albers, U. Ko¨nig, D. Temmler, and
D. Knoll, Si/SiGe HBTs for applications in low power ICs,
Solid–State Electron. 39(4):471–480 (1996).
41. E. F. Crabbe, B. S. Meyerson, J. M. C. Stork, and D. I. Har
ame, Vertical proﬁle optimization of very high frequency epit
axial Si, and SiGebase bipolar transistors, 1993 IEDM Tech.
Digest, 1993, pp. 73–86.
42. R. Doerner, J. Gerdes, C. Rheinfelder, F. J. Schmu¨ ckle,
W. Heinrich, and K. S. F. S. J.F. Luy, Modelling of passive
elements for coplanar SiGe MMICs, in 1995 IEEE MTT–S Int.
Microwave Symp. Digest, Orlando, FL, May 1995,
pp. 1187–1190.
43. K. M. Strohm, J.F. Luy, F. Scha¨ fﬂer, H. Jorke, H. Kibbel,
C. Rheinfelder, R. Doerner, J. Gerdes, F. J. Schmu¨ ckle, and
W. Heinrich, Coplanar Kaband SiGe–MMIC ampliﬁer, Elec
tron. Lett. 31:1353–1354 (Aug. 1995).
44. K. M. Strohm and J.F. Luy, Silicon–germanium bipolar tech
nology for RFand microwave applications, Proc. Microwave &
Radio Frequency Conf., London, Sept. 30–Oct. 2, 1997,
pp. 340–345.
45. H. Gummel and H. Poon, A compact bipolar transistor model,
Proc. 1970 IEEE Int. SolidState Circuits Conf. Feb. 1970,
pp. 78–79.
46. H. Gummel and H. Poon, An integral charge control model of
bipolar transistors, Bell Syst. Tech. J. 49:827–853 (May–June
1970).
47. S. Bruce et al., Development, implementation and veriﬁcation
of a physicsbased Si/SiGe HBT model for millimeterwave
nonlinear circuit simulation, Proc. 26th European Microwave
Conf. Prague, 1996, pp. 903–905.
48. C. Rheinfelder, M. Rudolph, R. Bei%wanger, and W. Heinrich,
Nonlinear modeling of SiGe HBTs up to 50GHz,
1997 Int. Microwave Symp. Digest, Denver, June 1997,
pp. 877–880.
49. M. Friedrich and H.M. Rein, Analytical currentvoltage re
lations for compact SiGe HBT models. Part I: The ‘‘idealized’’
HBT, IEEE Trans. Electron. Devices 46:1384–1393 (July
1999).
50. M. Friedrich and H.M. Rein, Analytical currentvoltage re
lations for compact SiGe HBT models. Part II: Application to
practical HBTs, IEEE Trans. Electron. Devices 46:1394–1401
(July 1999).
51. M. Schro¨ter and T.Y. Lee, Physicsbased minority charge and
transit time modeling for bipolar transistors, IEEE Trans.
Electron. Devices 46:288–300 (Feb. 1999).
52. M. Schro¨ter Staying current with hicum, IEEE Circ. Devices
Mag. 18:16–25 (May 2002).
53. M. Schro¨ter, HICUM/Level2 model equations (online).
54. C. McAndrew, J. Seitchik, D. Bowers, M. Dunn, M. Foisy,
I. Getreu, M. McSwain, S. Moinian, J. Parker, D. Roulston,
M. Schroter, P. van Wijnen, and L. Wagner, VBIC95, the ver
tical bipolar intercompany model, IEEE J. SolidState Circ.
31:1476–1483 (Oct. 1996).
55. M. Ramana Murty, K. M. Newton, S. Sweeney, D. Sheridan,
and D. Harame, Implementation of a scalable and statistical
VBIC model for largesignal and intermodulation distortion
analysis of SiGe HBTs, 2002 Int. Microwave Symp. Digest,
Seattle, June 2002, pp. 2165–2168.
56. F. Sinnesbichler and G. Olbrich, Accurate largesignal mod
eling of SiGe HBTs, 2000 Int. Microwave Symp. Digest, Bos
ton, June 2000, pp. 749–752.
57. F. Sinnesbichler, Hybrid millimeterwave push–push oscilla
tors using silicon–germanium HBTs, IEEE Trans. Microwave
Theory Tech. 51(2):422–430 (2003).
58. Y. Zhu, J. Twynam, M. Yagura, M. Hasegawa, T. Hasegawa,
Y. Eguchi, A. Yamada, E. Suematsu, K. Sakuno, H. Sato, and
N. Hashizume, Analytical model for electrical and thermal
transients of selfheating semiconductor devices, IEEE Trans.
Microwave Theory Tech. 46:2258–2263 (Dec. 1998).
59. F. Sinnesbichler and G. Olbrich, Thermal time constant
extraction for HBTs from Sparameter measurements, Int.
J. Electron. Commun. (AEU
¨
) 57(3):181–184 (2003).
60. C. D. Motchenbacher and F. C. Fitchen, LowNoise Electronic
Design, Wiley, New York, 1973.
61. U. Ko¨nig, A Gruhle, and A. Schu¨ ppen, SiGe devices and cir
cuits: Where are the advantages over III/V?. Proc. 1995 IEEE
GaAs IC Symp. 1995, pp. 15–17.
62. R. Plana, L. Escotte, J. P. Roux, J. G. A. Gruhle, and
H. Kibbel, 1/f noise in selfaligned Si/SiGe heterojunction
bipolar transistor, IEEE Electron. Device Lett. 16:58–60
(Feb. 1995).
63. R. Plana, B. V. Haaren, J. P. Roux, L. E. A. Gruhle,
H. Dietrich, J. Graffeuil, K. S. F. Scha¨ fﬂer, and J.F. Luy,
Low–frequency noise in millimeter–wave Si/SiGe hetero
junction bipolar transistors, 1995 IEEE MTT–S Int.
Microwave Symp. Digest, Orlando, FL, May 1995, pp.
1187–1190.
64. J. D. Cressler, L. Vempati, J. A. Babcock, R. C. Jaeger, and
D. L. Harame, Low–frequency noise characteristics of UHV/
CVD epitaxial Si– and SiGe–base bipolar transistors, IEEE
Electron. Devices Lett. 17:13–15 (Jan. 1996).
65. F. Sinnesbichler, M. Fischer, and G. Olbrich, Accurate extrac
tion method for 1/fnoise parameters used in GummelPoon
type bipolar junction transistor models, 1998 Int. Microwave
Symp. Digest, Baltimore, June 1998, pp. 1345–1348.
66. R. Hagelauer, T. Ostermann, U. Ko¨nig, M. Glu¨ ck, and
G. Ho¨ck, Performance estimation of Si/SiGe hetero–CMOS
circuits, Electron. Lett. 33:208–210 (Jan. 1997).
67. S. Sadek and K. Ismail, Si/SiGe CMOS possibilities, Solid–
State Electron. 38:1731–1734 (1995).
68. T. Hackbarth, G. Hck, H. Herzog, and M. Zeuner, Strain re
lieved SiGe buffers for Sibased heterostructure ﬁeld effect
transistors, J. Cryst. Growth 201/202:734–738 (1999).
GESI ALLOYS AND DEVICES 1805
69. M. Zeuner, T. Hackbarth, G. Ho¨ck, D. Behammer, and
U. Ko¨nig, Highfrequency SiGenMODFET for microwave
applications, IEEE Microwave Guided Wave Lett. 410–412
(Oct. 1999).
70. G. Hock, T. Hackbarth, H.J. Herzog, M. Enciso, F. Aniel,
P. Crozat, R. Adde, E. Kohn, and U. Konig, 0.1mm gate length
ptype Ge/Si
0.4
Ge
0.6
MODFET with 135GHz f
max
, Electron.
Lett. 36:1428–1429 (Aug. 2000).
71. M. Enciso, F. Aniel, P. Crozat, R. Adde, M. Zeuner, A. Fox, and
T. Hackbarth, 0.3dB minimum noise ﬁgure at 2.5 GHz of
0.13mm Ge/Si
0.58
Ge
0.42
nMODFETs, Electron. Lett. 37:1089–
1090 (Aug. 2001).
72. G. Wright, Smallsignal theory of the transistor transittime
oscillator (translator), Solid State Electron. 22:399–407
(1979).
73. A. Grinberg and S. Luryi, Coherent transistor, IEEE Trans.
Electron. Devices 40:1512–1520 (1993).
74. J. Weller, H. Jorke, K. Strohm, J.F. Luy, H. Kibbel, H.J.
Herzog, and R. Sauer, Assessment of transport parameters for
the design of high speed Si/SiGe HBTs with compositionally
graded base, Thin Solid Films 336:137–140 (1998).
75. R. Wanner, G. Olbrich, H. Jorke, J.F. Luy, S. Heim, E. Kaspe,
and P. Russer, Experimental veriﬁcation of the resonance
phase transistor concept.
76. J. Weiner, A. Leven, V. Houtsma, Y. Baeyens, Y.K. Chen,
P. Paschke, Y. Yang, F. J. W.J. Sung, A. Tate, R. Reyes, R.
Kopf, and N. Weimann, SiGe differential transimpedance am
pliﬁer with 50GHz bandwidth, IEEE J. SolidState Circ.
38:1512–1517 (Sept. 2003).
77. M. Reinhold, C. Dorschky, E. Rose, R. Pullela, P. Mayer,
F. Kunz, Y. Baeyens, T. Link, and J.P. Mattia, A fully inte
grated 40Gb/s clock and data recovery IC with 1:4 DEMUX in
SiGe technology, IEEE J. SolidState Circ. 36:1937–1945
(Dec. 2001).
78. D. Kehrer, H.D. Wohlmuth, H. Knapp, M. Wurzer, and
A. Scholz, 40Gb/s 2:1 multiplexer and 1:2 demultiplexer
in 120nm standard CMOS, IEEE J. SolidState Circ.
38:1830–1837 (Nov. 2003).
79. F. Beisswanger, U. Gu¨ ttich, and C. Rheinfelder, Microstrip
and coplanar SiGeMMIC oscillators, in European Microwave
Conf., Prague, 1996, pp. 588–592.
80. H. Li and H.M. Rein, Millimeterwave VCOs with wide
tuning range and low phase noise, fully integrated in a
SiGe bipolar production technology, IEEE J. SolidState
Circ. 184–191 (Feb. 2003).
81. H. Li, H.M. Rein, R.E. Makon, and M. Schwerd, Wideband
VCOs in SiGe production technology operating up to about
70GHz, IEEE Microwave Wireless Compon. Lett. 425–427
(Oct. 2003).
82. C. Rheinfelder, K. Strohm, L. Metzger, H. Kibbel,
J.F. Luy, and W. Heinrich, 47GHz SiGeMMIC oscillator,
1999 Int. Microwave Symp. Digest, Anaheim, CA, June 1999,
pp. 5–8.
83. F. Sinnesbichler, B. Hautz, and G. Olbrich, Low phase noise
58GHz SiGe HBT push–push oscillator with simultaneous
29GHz output, 2000 Int. Microwave Symp. Digest, Boston,
June 2000, pp. 11–16.
84. Y. Baeyens and Y. Chen, A monolithic integrated 150GHz
SiGe HBT push–push VCO with simultaneous differential v
band output, 2003 Int. Microwave Symp. Digest, Philadel
phia, June 2003, pp. 877880.
85. M. Kaleja, A. Gru¨ bl, F. Sinnesbichler, G. Olbrich, K.M.
Strohm, J.F. Luy, and E. Biebl, Application of Si/SiGe HBTs
in active integrated antennas, 2000 Topical Meeting on
Silicon Monolithic Integrated Circuits in RF Systems Digest,
Garmisch, April 26–28, 2000, pp. 64–66.
86. M. Kaleja, A. Gru¨ bl, F. Sinnesbichler, G. R. Olbrich, K. Str
ohm, J. Luy, and E. Biebl, Si/SiGe hbt active integrated an
tenna on high resistivity silicon substrate, 2000 Int.
Microwave Symp. Digest, Boston, June 2000, pp. 1899–1902.
87. K.W. Yu, Y.L. Lu, D.C. Chang, V. Liang, and M. Chang, K
band lownoise ampliﬁers using 0.18mm CMOS technology,
IEEE Microwave Wireless Compon. Lett. 14:106–108 (March
2004).
88. W. Menzel, Future applications, in J.F. Luy and P. Russer,
eds., SiliconBased MillimeterWave Devices, Springer
Series in Electronics and Photonics, Vol. 32, Springer, 1994,
pp. 323–338.
89. M. Steinhauer, H. Irion, M. Schott, M. Thiel, H.O. Ruoss, and
W. Heinrich, SiGeBased Circuits for Sensor Applications be
yond 100GHz, 2004 Int. Microwave Symp. Digest, Fort Worth,
TX, June 2004, pp. 223–226.
90. J. Block, T. Meister, H. Knapp, D. Zoschg, H. Schafer,
K. Auﬁnger, M. Wurzer, S. Boguth, M. Franosch, R. Stengl,
R. Schreiter, M. Rast, and L. Treitinger, SiGe bipolar
technology for mixed digital and analogue RF applications,
Electron Devices Meeting 2000, IEDM Tech. Digest, Dec. 2000,
pp. 745–748.
91. J. Choi, C.J. Weiske, G, Olbrich, and P. Russer, ‘Flipchip
bonded Si Schottky sampling circuits for high speed demul
tiplexers,’ 2003 Int. Microwave Symp. Digest, Philadelphia,
June 2003, pp. 1515–1518.
92. M Singer, K. M. Strohm, J.F. Luy, and E. M. Biebl, Active
SIMMWICantenna for automotive applications, 1997 Int.
Microwave Symp. Digest, Denver, June 1997, pp. 1265–1268.
GRATINGS, GRATING ANTENNAS
CHRISTOPHER T. RODENBECK
KAI CHANG
Texas A&M University
College Station, Texas
A grating is an array of parallel obstructions separated
from one another at a distance comparable to a wave
length. When an electromagnetic wave is incident upon
a grating, it scatters to one or more speciﬁc angles
determined by the spacing of the grating and the fre
quency of the wave. Since this effect is a diffraction
phenomenon, gratings are often also referred to as diffrac
tion gratings.
One of the most important uses of the diffraction
grating is as an antenna for millimeterwave through
optical frequencies. This type of diffraction grating is
the primary focus of this article. In integrated optics,
grating antennas are used to couple power into and out
of dielectric waveguides. For millimeterwave systems,
grating antennas have received attention for several key
advantages that they have in comparison with other
1806 GRATINGS, GRATING ANTENNAS
conventional microwave antennas. They have low radia
tion losses, they are inexpensive and easy to fabricate, and
they are planar. There is also considerable interest in
using grating antennas to implement lowcost beam steer
ing at millimeterwave frequencies.
This article examines gratings and grating antennas in
a series of ﬁve sections. Each of these sections is presented
in such a way that it may be read independently from the
others. An overview of the historical development of
diffraction gratings is presented. The fundamental princi
ples behind grating diffraction are explained in a manner
suitable for those with an RF/microwave background. A
clear design procedure for grating antennas is offered,
along with an extensive reference list for those interested
in further study. Some of the latest and most impressive
applications of millimeterwave gratings are described.
The article then concludes with a brief discussion of the
future of this ﬁeld.
1. HISTORICAL OVERVIEW
Despite its fundamental importance, the early origins of
the diffraction grating are a remarkable tale of discovery
and rediscovery. An intense period of research into optical
technology had begun during the early 1600s following the
invention of the telescope by the Dutch. During that cen
tury, the fundamental principles of refraction and diffrac
tion were discovered and applied to numerous optical
components such as lenses, mirrors, reﬂectors, slits, pin
holes, etc. Gratings, however, were not invented until a
much later date.
The ﬁrst discovery of the diffraction grating occurred in
Philadelphia in 1785. Francis Hopkinson, one of the sign
ers of the Declaration of Independence, held a ﬁnely wo
ven silk handkerchief up to a bright light and observed the
interference pattern produced by the light passing
through the fabric. Hopkinson consulted his friend David
Rittenhouse, a selftrained physicist. Recognizing the
phenomenon as a diffraction effect, Rittenhouse designed
his own diffraction grating made of human hairs wound
around two ﬁne screws. He then used the device to mea
sure the approximate wavelength of visible light. Hailing
his invention in an American scientific journal, Ritten
house exclaimed that ‘‘new and interesting discoveries
may be made respecting the properties of this wonderful
substance, light, which animates all nature in the eyes of
man, and perhaps, above all things, disposes him to ac
knowledge the Creator’s bounty’’ [1–3].
Since few researchers of that era read American scien
tific journals, Thomas Young of London had the opportu
nity to reinvent the diffraction grating in 1803. Young
used his invention to measure the wavelengths of the var
ious visible colors, but his main objective was to lend sup
port to his view that light is a wave rather than a particle.
Opponents of the wave theory attacked Young’s work, la
beling it as ‘‘destitute of every species of merit’’ [3,4]. The
practical importance of Young’s invention was thus unfor
tunately lost amid the then viscous debate concerning the
nature of light.
The grating would be discovered for the ﬁnal time by
Joseph von Fraunhofer in 1821. Fraunhofer had earlier
established an optical company in Bavaria that produced
optical components of the highest quality. After unexpect
edly observing grating diffraction during an experiment,
Fraunhofer built increasingly precise gratings, ultimately
etching the proﬁle of a grating directly onto the surface of
a glass plate. Using this arrangement, Fraunhofer was
able to separate visible light into the different colors and
measure their wavelengths. Although his primary objec
tive was to use the device to make precise optical mea
surements and thereby improve his company’s products,
Fraunhofer did popularize additional uses for optical grat
ings, including chemical spectroscopy and astronomical
observation [1–4]. After his death, the diffraction grating
eventually became an indispensable component of optics.
As one researcher put it in 1960, ‘‘It is difﬁcult to point to
another single device that has brought more important
information to every ﬁeld of science than the diffraction
grating’’ [3].
The ﬁrst use of gratings at radiofrequencies came dur
ing Heinrich Hertz’ revolutionary experiments in 1887–
1891. Hertz used a narrowly spaced wire grating in order
to test the polarization of his ‘‘electric waves’’ [5,6]. Shortly
after Hertz’ death in 1894, Chandra Bose of Calcutta built
an impressive 60GHz transmitter and receiver. During
these experiments, Bose used a diffraction grating to mea
sure the wavelength of his millimeterwave source [7].
Subsequent decades saw rapid progress in the area of RF/
microwave antennas and arrays. Early researchers in this
area were heavily inﬂuenced by their familiarity with the
already wellknown diffraction grating. This is not sur
prising since gratings and array antennas are both gov
erned by the same fundamental principles of wave
interference. In fact, the designation ‘‘grating lobe’’ is still
commonly used to refer to the additional maxima emitted
by an array of widely spaced antennas.
The ﬁrst grating antennas were developed and popu
larized at optical frequencies during the 1970s. Research
ers at IBM etched a grating proﬁle along the surface of a
dielectric waveguide and used the device to transmit and
receive laser light [8]. Called a grating ‘‘coupler,’’ this op
tical antenna is used in integrated optics for coupling op
tical power into and out of dielectric waveguides. In 1977,
Tatsuo Itoh demonstrated the ﬁrst microwave grating an
tenna at 15 GHz [9]. Shortly afterward, a team of re
searchers at the U.S. Army CommunicationsElectronics
Command presented the development of a 60GHz grating
antenna fabricated along a silicon dielectric waveguide
[10–14]. The design established the grating antenna as a
new kind of lowloss lowcost antenna for millimeterwave
frequencies. This work captured the interest of research
ers worldwide. Progress was rapid and was aided by in
herent similarities between grating antennas and earlier
travelingwave antennas designed at lower microwave
frequencies [15–17]. With increasing contemporary inter
est in the commercial and military applications of milli
meter waves, much of the early research on grating
antennas is now receiving renewed attention. And what
is more, new and exciting applications of this antenna
continue to be developed and are beginning to ﬁnd
GRATINGS, GRATING ANTENNAS 1807
commercial success for a wide range of millimeterwave
applications.
2. FUNDAMENTAL PRINCIPLES
The early investigators who popularized gratings were
able to grasp the underlying principles of grating diffrac
tion even before those principles had been formalized.
Grating diffraction can in fact be accurately explained by
an analogy between light and sound [1]:
If a sudden sharp noise such as is made by clapping the hands
together is reﬂected from a high ﬂight of steps, the sound
comes back to us as a musical note; in other words, the steps
impress the element of periodicity upon the reﬂected distur
bance, each step throwing off an echo wave.
Sound incident on the staircase scatters in different direc
tions. Those directions are determined by the step spacing
and by the wavelength of each tone. In the case of a grat
ing, an electromagnetic wave incident on a periodic inter
face scatters in a direction determined by the grating
spacing and frequency of the wave.
This section presents a brief description of the funda
mental principles behind grating diffraction. The discus
sion draws on material found in the literature [3,18–21]
and is crafted in a manner suited for those with an RF/
microwave background.
To begin with, consider the generalized twodimension
al grating interface shown in Fig. 1. The structure is in
ﬁnite and periodic in the x direction and is excited by a
plane wave incident at the angle y
i
. Since the grating in
terface is periodic in x, this suggests that the scattered
ﬁeld can be expanded in terms of a Fourier series
F
s
(x; z) =
o
n=÷o
F
n
(z)e
÷j(2np=d) x
_ _
e
÷jk
0
(sin y
i
) x
(1)
where F
n
(z) is the unknown Fourier coefﬁcient for the nth
space harmonic e
÷j(2np=d)x
. The e
÷jk
0
(siny
i
)x
term outside the
expansion accounts for the phase distribution of the inci
dent ﬁeld along the surface of the grating interface. Bring
ing that term within the expansion simpliﬁes (1) to
F
s
(x; z) =
o
n=÷o
F
n
(z)e
÷jk
x;n
x
(2a)
where
k
x;n
=k
0
sin y
i
÷
2pn
d
=k
0
sin y
n
(2b)
and y
n
denotes the angle of radiation for the nthspace
harmonic. Equation (2b)b is the wellknown ‘‘grating
equation.’’ By convention, y
i
and y
n
are deﬁned so that y
i
is positive in the reverse quadrant and y
n
is positive in the
forward quadrant.
Since the scattered ﬁeld must satisfy the Helmholtz
equation
V
2
F
s
(x; z) ÷k
2
0
F
s
(x; z) =0 (3)
it is possible to deﬁne
k
z;n
= ±
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
k
2
0
÷k
2
x;n
_
=k
0
cos y
n
(4)
where the sign is chosen to satisfy the radiation condition.
The wavevector for the scattered ﬁeld can now be ex
pressed in terms of the unit vectors u
x
and u
z
:
k
n
=k
x;n
u
x
÷k
z;n
u
z
=k
0
sin y
n
u
x
÷k
0
cos y
n
u
z
(5)
The incident wavevector likewise has the following form:
k
i
=k
0
sin y
i
u
x
÷k
0
cos y
i
u
z
(6)
For a given incident angle y
i
and grating spacing d, it is
possible to plot the incident and scattered wavevectors
from (5) and (6) on an illustrative diagram. An example is
shown in Fig. 2. The incident wavevector k
i
arrives at an
angle y
i
, and the wavevector k
0
scatters in an equal and
opposite direction y
0
. All the radiating wavevectors lie to
gether on a unit circle of radius k
0
, with each vector point
ing in its direction of radiation. All the nonradiating
wavevectors lie outside this circle. These nonradiating
or evanescent modes decay exponentially away from the
z
x
i
d
Incoming
plane wave
Figure 1. A generalized reﬂection grating of
periodicity d illuminated by a plane wave in
cident at the angle y
i
.
1808 GRATINGS, GRATING ANTENNAS
surface of the grating. What makes this diagram especial
ly useful is that it gives the designer the ability to visu
alize how changing y
i
or d changes how many modes
radiate and in which direction each radiates.
In the case of a grating antenna, the grating interface
lies along the surface of a dielectric waveguide, and the
grating is excited by the signal propagating within that
waveguide. The grating equation (2b) should then be
rewritten as
Re(k
x;n
) =Re(k
x;0
) ÷
2pn
d
=k
0
sin y
n
(7)
where k
x,0
is the propagation constant within the dielec
tric waveguide. k
x,0
is actually a complex number with the
imaginary part describing the rate of radiation along the
antenna [21]. The imaginary part of k
x,0
is disregarded,
however, when calculating the number and directions of
the radiating space harmonics. Furthermore, the real part
of k
x,0
is typically calculated with good accuracy as the
propagation constant along the dielectric waveguide with
out any grating present.
A plot similar to Fig. 2 can be produced using (7) to
determine the scattered wavevectors from the expression
in (5). An example case is shown in Fig. 3. The fundamen
tal wavevector k
0
points in the positive direction along the
ordinate axis with a magnitude equal to Re(k
x,0
). Each ra
diating space harmonic lies on the unit circle and corre
sponds to a beam radiated by the antenna. Although
grating antennas are typically designed so that only the
n= ÷1 beam radiates, this example shows two radiating
beams. The n= ÷1 beam is radiating in the forward quad
rant (y 401) and the n= ÷2 beam, in the reverse quad
rant (yo01). This diagram can be useful for designing a
grating so that a single main beam radiates in some de
sired direction.
3. ANTENNA DESIGN AND ANALYSIS
The grating antenna is a type of leakywave antenna.
A leakywave antenna is essentially a waveguiding
structure perturbed either continuously or periodically
so that it leaks power all along its length [22]. In the
case of a grating antenna, the waveguiding structure is a
dielectric waveguide perturbed by a periodic grating
placed along its surface. The conventional conﬁguration
is shown in Fig. 4. A periodic array of transverse strips is
arranged along the surface of a dielectric waveguide, in
this case a dielectric image line. The strips can be either
metallic or dielectric. These strips form a grating that
perturbs the signal traveling along the guide. If the an
tenna is designed correctly, the grating excites a single
leaky mode above the surface of the guide to create a di
rective beam of radiation in the far ﬁeld. The radiated
beam is fanshaped, with narrow beamwidth in the array
effect dimension.
The procedure for designing a grating antenna encom
passes four key steps. This section provides a complete
k
−3
k
−2
k
−1
k
0
=Re(k
x,0
)
k
x,n
Re(k
z,n
)
k
o
−k
o
2/d 2/d 2/d
Figure 3. Diagram illustrating the wavevectors
along a hypothetical grating antenna.
k
−3
k
−2
k
−1
k
0
k
i
k
+1
k
+2
k
+3
k
x,n
Re(k
z,n
)
k
o
−k
o
2/d 2/d 2/d 2/d 2/d 2/d
i
0
Figure 2. Diagram illustrating the incident and
scattered wavevectors from a hypothetical reﬂec
tion grating.
GRATINGS, GRATING ANTENNAS 1809
description for each of those steps. Additional references
are also provided for the reader who wishes to explore
more advanced topics related to the design and analysis of
this class of antenna.
Step 1. The aspect ratio and dielectric constant of the di
electric waveguide must be chosen so that only one mode
can propagate within the guide over the antenna’s desired
frequency range. This requires that the designer calculate
the propagation constants and cutoff frequencies within
the guide for each potential dielectric waveguide design.
There are several options for performing this sort of cal
culation. Rigorous methods are available [23], but they are
not popular since they involve the solution of a large set of
coupled, transcendental equations. For lowdielectriccon
stant guides, simple effective dielectric constant (EDC)
methods can be used with good accuracy [23]. An espe
cially accurate EDC method has been developed for di
electric image line [24]. For highdielectricconstant
guides, fullwave electromagnetic simulation should be
used to determine the guided wavelength.
In general, greater antenna directivity can be achieved
using a wide guide rather a narrow guide of the same
length. However, this factor must be balanced against
other considerations, such as the available singlemode
bandwidth in the guide and the mechanism used to feed
the guide. Using a highdielectricconstant material also
has certain advantages, to be discussed later, but these too
must be balanced against other considerations, such as
the availability of highdielectricconstant substrates with
reasonably low loss in the chosen frequency range and by
the undesirable dispersion in the guided and leaky modes
that can occur along high dielectric constant guides
[25,26].
Step 2. The dielectric waveguide must be excited at the
intended frequency of operation. This can be accomplished
by directly integrating a signal source into the dielectric
guide [27–32] or by manufacturing a transition between
the dielectric guide and another type of transmission line.
Although a waveguide transition is shown in Fig. 4,
other transitions, including microstrip, CPW, and slot
transitions, are also available [21,23,33–37]. In every
case, the transition should be carefully designed to limit
spurious radiation and surfacewave excitation, which
have the potential spoil an otherwise good antenna de
sign by increasing the sidelobes above an acceptable level
[22].
It is important to note that the method of excitation
determines the polarization of the farﬁeld radiation pat
tern. Dielectric waveguides can, in general, be excited so
that one or both of two orthogonal degenerate fundamen
tal modes propagate [23]. For a rectangular dielectric
guide like the image line shown in Fig. 4, the primary
ﬁeld components under Ewave excitation are H
y
, E
x
, and
E
z
, and the antenna’s E plane is located in the x–z plane.
Under Hwave excitation, the primary ﬁeld components
are E
y
, H
z
, and H
x
, and the H plane is located in the x–z
plane. Using other conventions present in the literature, it
is also possible to designate Ewave excitation along the
guide in Fig. 4 as TMtoz excitation or simply as TM ex
citation, as E
z
mn
excitation, or as TE or ‘‘perpendicularpo
larized’’ excitation (since the E ﬁeld is ‘‘perpendicular’’ to
the strips). Likewise, Hwave excitation can be referred to
as TEtoz, TE, E
y
mn
, TM, or ‘‘parallelpolarized.’’ For an
d
y
x
z
w
Waveguide
transition
Ground
plane
Grating
strips
Dielectric
image line
Figure 4. Conﬁguration of a conventional
grating antenna.
1810 GRATINGS, GRATING ANTENNAS
image line, Ewave excitation is generally preferred since
the presence of the ground plane shorts out most of
the power that could couple to the Hwave modes. Recent
patents indicate that it may also be possible to inten
tionally excite both degenerate modes in a dielectric wave
guide in order to create dual polarization or even circular
polarization [38,39].
Step 3. The next step in the design process is to determine
the periodic spacing between the grating strips. For thin
metal strips or thin dielectric strips, the presence of a
grating along the dielectric guide has only a negligible ef
fect on the propagation constant within the guide itself. In
this case, the angle of radiation in the direction forward
from broadside can be calculated at any given frequency of
excitation as a simple function of the guide wavelength l
g
and strip spacing d
y
n
= arcsin
l
0
l
g
÷
nl
0
d
_ _
;
l
0
l
g
÷
nl
0
d
¸
¸
¸
¸
¸
¸
¸
¸
_ 1 (7a)
where l
0
is the freespace wavelength and n, the space
harmonic (0, 71, 72, y), is chosen as ÷1 for singlebeam
operation [40]. In principle, the permissible range of
values for y
–1
extends both forward and reverse of broad
side, across the entire 7901 range. But as is the case with
most planar antenna arrays, grating antennas do not
radiate well beyond the 7601 range. There are other
limitations as well. Bragg reﬂection occurs if d=l
g
[9].
This prevents the antenna from radiating well for y
–1
=01
unless a more elaborate grating pattern [23,41,42] is
used. In addition, it is difﬁcult to achieve a sufﬁciently
high effective relative dielectric constant along an image
line or other open dielectric guide in order to prohibit
appearance of the n= ÷2 grating lobe for increasingly
positive values of y
÷1
while still maintaining singlemode
bandwidth along the guide. Unusually high dielectric
constants (e
r
420) are required, with associated tradeoffs
in performance and implementation [25,26,43]. Choosing
a main beam angle in the y
–1
401 range is thus typi
cally limited by the propagation of the n= ÷2 space
harmonic.
Step 4. The ﬁnal design step is to choose the strip widths
and antenna length. The widths of the strips determine
the rate of radiation along the antenna, and the antenna
length determines how much power is left unradiated at
the end of the antenna. Narrow strips radiate too little,
leaving a substantial amount of power unradiated at the
end of even a long antenna. Excessively wide strips radi
ate too much power during the ﬁrst few strips, causing the
effective aperture of the antenna to be too small and the
sidelobes to be large [33]. In both cases, antenna efﬁciency
can suffer. A popular compromise is to gradually increase
the widths of the strips away from the input to the
antenna.
Designing the strip widths can be done by experiment,
by analysis, or even by rules of thumb. Several such rules
of thumb are available in the literature. For the case of
dielectricstrip gratings, perturbational analysis can be
used to argue that w=0.5d is an optimal strip width
[40,44]. In the case of a metalstrip grating, researchers
determined by experiment that maximum radiation for a
particular design occurred when w=0.4l
g
[33]. They then
used this ﬁnding to formulate a very conservative empiri
cal guideline for designing a tapered distribution of strip
widths along their antenna.
Analytical approaches are generally preferred in an
tenna design. To analytically design a grating antenna
Figure 5. Array of periodic dielectric antennas designed for ra
diation in the broadside direction: (a) array conﬁguration; (b) ﬁn
ished device for 65GHz band. The element antennas are fed by a
resonant slotted waveguide array with a horn transition. (From
Solbach, [66], r 1985 EW Communications. Reprinted with per
mission of publisher from MSN & CT. All rights reserved.)
GRATINGS, GRATING ANTENNAS 1811
requires an analysis that is capable of relating the design
parameters (viz., the strip width, interelement spacing,
dielectric constant, etc.) to the radiation rate along the
surface of the antenna. Once the radiation rate is known,
the farﬁeld pattern can easily be calculated using the
spatial Fourier transform [40]. Unfortunately, calculating
the radiation rate along a grating antenna is typically
considered an involved process [21]. For the dielectric strip
gratings commonly used in integrated optics, commercial
software and curve ﬁt models [45] are available. For me
tallic strip gratings, several rapid analytical procedures
are available for the limiting cases of narrow strip grat
ings and narrow slot gratings. (See Refs. 17, 46, and 47
and the references cited in those papers.) Practical de
signs, however, typically use strips whose widths fall be
tween those two limiting cases. Recently, a simple
spectraldomain approach has been proposed that is ca
pable of analyzing metal gratings composed of strips that
are neither narrow nor wide [48]. Many more rigorous
numerical methods are also available, with varying levels
of rigor and computational effort for each approach. These
methods include modematching techniques [49,50],
waveguide models [51], rigorous network formulations
[52], and fullwave simulation [53,54]. An excellent and
timely overview of various approaches and their relative
merits is provided in Ref. 3 for both metal and dielectric
strip gratings. The reader is warned to investigate each
approach since the literature does contain references to
inaccurate and even erroneous approaches for analyzing
grating antennas.
Experimental design can be a particularly expedient
approach in the case of metalstrip gratings. Multiple sets
of gratings can be etched onto a single thin dielectric ﬁlm
[55–57]. Thin dielectric ﬁlms made from materials such as
RTDuroid have low loss at millimeterwave frequencies
and can be easily etched using conventional photolithog
raphy. The ﬁlm can then be extended across the surface of
the dielectric guide, allowing the designer to rapidly test
the radiation rate for each of the sets of grating strips.
This information can then be used together with the spa
tial Fourier transform to determine a favorable distribu
tion of strip widths in order to achieve a desired farﬁeld
pattern.
When the distribution of strip widths and the radiation
rate along the antenna are being determined, the length of
the antenna must also be chosen so that only a certain
fraction of the input power remains unradiated at the end
of the dielectric guide. If the amount of unradiated power
is too large, antenna efﬁciency is low. For an antenna of
length L, the total efﬁciency Z can be expressed as
Z =
_
L
0
2a(x)e
÷2(a(x) ÷a
L
)x
dx (8)
where a(x) is the radiation rate along the grating and a
L
accounts for conductor and dielectric losses [15]. If the
radiation rate is uniform along the antenna and much
larger than the loss factors, the efﬁciency can be appro
ximated for design purposes as 1 ÷e
÷2a
L
. If a substantial
portion of the input power is to be left unradiated, the
dielectric guide should be terminated in a matched load
or with absorbing material, or else the wasted power
will reﬂect from the end of the antenna and radiate as a
sidelobe.
4. APPLICATIONS
The chief purpose of this section is not to present a
comprehensive review of all applications of gratings and
y
x
z
Waveguide
transition
Ground
plane
Moveable
grating film
d
min
d
max
Dielectric
image line
Figure 6. A lowcost reconﬁgurable grating
antenna.
1812 GRATINGS, GRATING ANTENNAS
grating antennas. Instead, the objective is to capture the
reader’s interest and identify the major applications of
gratings. A special emphasis is given to antenna applica
tions at millimeterwave frequencies.
4.1. Array Antennas
Grating antennas are based on dielectric waveguide tech
nology. Unlike printedcircuit transmission lines, dielec
tric waveguides exhibit low loss even at very high
millimeterwave frequencies. Unlike rectangular metal
waveguides, dielectric waveguides are easily fabricated
and compatible with mass production at frequencies
through W band and beyond. Dielectric waveguides are
thus ideal candidates for feeding millimeterwave antenna
arrays. Diffraction gratings are perhaps the most natural
and inexpensive implementation of an array antenna to be
based on this class of transmission line.
A conventional grating antenna is shown in Fig. 4.
Many variations on this conﬁguration are possible. The
strips can be metal [10] or dielectric [40]. The strips can be
replaced with slots, grooves, or cavities in the ground plane
[23]. These strips, slots, and other perturbations don’t
necessarily have to be rectangular; they can also be circu
lar. Their cross sections can be ‘‘blazed’’ or nonrectangular
[3]. In addition, the perturbations can be made selfsimilar
to allow broadside radiation [23,41]. If desired, the image
line can also be replaced with another type of dielectric
waveguide, such as insular image guide, nonradiative
dielectric (NRD) guide, invertedstrip dielectric guide, or
rectangular dielectric waveguide [23]. Gratings fabricated
on circular dielectric waveguides can radiate omnidirec
tionally [58]. Gratings based on rotated strips can radiate
circular polarization [59,60].
Several methods are available for increasing the gain of
a grating antenna. The gain in the nonarray plane can be
increased by embedding the dielectric guide within a
trough [61] or ﬂared trough [33] or by placing the guide
in proximity to a parabolic reﬂector [39]. Pencilbeam ra
diation can be achieved by etching a twodimensional
grating on a surfacewaveexcited dielectric slab [62–64].
This kind of grating is designated a ‘‘hologram’’ or ‘‘holo
graphic antenna.’’ This is a matter of nomenclature, how
ever, since the standard grating antenna is itself a type of
hologram. Other research has demonstrated unconven
tional circular and linear polarized pencilbeam antennas
fed by NRD gratings [65]. Planar arrays of grating
antennas can also be formed to increase the directivity
of the beam [42]. One planar grating array, shown in
Fig. 5, radiates a broadside beam at 65 GHz with a total
efﬁciency, including the power divider and feed losses, of
27% [66].
4.2. Scanning Antennas
A wide array of consumer wireless applications are antic
ipated at millimeterwave frequencies [48,55]. Many of
these emerging applications require the ability to track
a user or moving target. Unfortunately, conventional
technology for steering and redirecting millimeterwave
wireless signals is often prohibitively expensive for
consumer applications. Rotating antennas use RF gimbal
feeds that require precision machining and costly main
tenance. Millimeterwave phased arrays, on the other
hand, are often exuberantly expensive, requiring exotic
technologies and extensive supporting electronic equip
ment. In order to reduce the cost and broaden the
commercial applications millimeterwave frequencies,
considerable research is currently directed at deve
loping lowcost scanning antennas and phased arrays
for the Ka band and above. Several approaches based on
the grating arrays are under investigation, including
−5
−6
−4
−2
0
Y

D
i
r
o
f
a
r
r
a
y
,
c
m
2
4
6
0
XDir of array, cm
5
−1
−1
−0.5
0
0.5
1
−0.5 0
V
e
r
t
.
o
f
f
s
e
t
,
s
i
n
(
B
)
Hor. offset, sin(A)
0.5 1
(a)
(b)
Figure 7. Electronically reconﬁgurable holographic antenna:
(a) conducting regions formed on a silicon wafer; (b) correspond
ing farﬁeld radiation pattern. (Courtesy of Dr. A. Fathy, Univer
sity of Tennessee.) (This ﬁgure is available in full color at http://
www.mrw.interscience.wiley.com/erfme.)
GRATINGS, GRATING ANTENNAS 1813
electromechanical, electronic, and electrooptic. Some of
these approaches are also classiﬁed as ‘‘reconﬁgurable an
tennas’’ since the central tactic is often to provide beam
steering by reconﬁguring the aperture of the grating.
Figure 6 illustrates a simple scanning antenna that can
be electromechanically reconﬁgured to provide lowcost
millimeterwave beam steering [67]. A thin movable di
electric ﬁlm is extended across a stationary dielectric im
age line. Continuous metal grating strips are etched along
the underside of the ﬁlm, with the strips depicted as vis
ible in the ﬁgure. The spacing between the metal strips
varies linearly along the length of the ﬁlm from d
max
to
d
min
. Scrolling the ﬁlm along the image line shifts the ﬁlm
in the y direction, steering a highly directive fanshaped
beam of radiation along the angle y. As illustrated in the
ﬁgure, the widths of the strips themselves are designed to
increase away from the input to the guide in order to
gradually perturb the signal as it is launched along the
image line. The prototype conﬁguration shown in Fig. 1 is
able to achieve 531 beam steering at 35 GHz. An innova
tive, bidirectionally excited design achieves 7501 scan
ning across the complete 35–40 GHz range [48,57]. A key
advantage of this approach is that the mechanical motion
is inertialess and completely external to the millimeter
wave feed. Thin steel sheets may also be used. In either
case, only a simple DC motor is required to spool the grat
ing layer along the surface of the dielectric guide. Another
electromechanical approach is to form a grating antenna
from a spring coiled around a dielectric waveguide [68].
Flexing the spring provides beam steering. Several inter
esting electromechanically scanning grating antennas are
now available commercially and are disclosed in a U.S.
patent [39].
Other researchers are developing electronically scan
ned grating antennas by integrating PIN diodes into the
grating structure. In Refs. 13 and 14, a PIN diode inte
grated into the sidewall of a grating antenna is used to
modulate the propagation constant within the dielectric
guide in order to steer the beam. In Ref. 69, PIN diodes are
integrated into the grating strips themselves in order to
switch the radiated beam between one of two scan states.
Most impressively, in Ref. 64, an array of PIN diodes are
integrated onto the back of a silicon wafer in order to cre
ate a dynamically reconﬁgurable hologram. Holographic
antennas are the twodimensional analog of grating an
tennas. Figure 7 illustrates the conducting pattern formed
by the PIN diodes on the silicon wafer and the correspond
ing farﬁeld pattern.
Optical and ferroelectric scanning grating antennas
are also a topic of investigation [70–75]. In many cases,
however, the experimental results require improvement in
order for the systems to ﬁnd broad acceptance. Figure 8
illustrates the conﬁguration of one design [74]. A circular
dielectric waveguide is coupled to a semiconducting di
electric slab. Laser light illuminates the surface of the di
electric slab in order to create islands of conducting
plasma on the surface. The spacing between these islands
can be changed in order to steer the mainbeam angle of
the radiated beam. Ferroelectric scanning grating anten
nas are also commercially available at frequencies from 10
to 76 GHz [75].
4.3. Other Applications
Many other uses for millimeterwave gratings do exist. For
example, on the circuit level gratings can be used as reso
nators for notch ﬁlters or for oscillators [9,76]. Closely
spaced gratings are often used as polarizers in grid ampli
fers [77]. In the farﬁeld zone, millimeterwave gratings are
used for as spectrometry [7,78], spatial ﬁltering and beam
shaping [79], spatial power combining and dividing [80–82],
and remote beam control [83–85]. These applications for
gratings are likely to increase in importance as millimeter
wave technology becomes more and more common.
5. CONCLUSION
Interest in millimeterwave technology has been building
steadily since the mid1970s. Today, several applications
have already become established; numerous others are
quickly emerging. These new millimeterwave wireless
systems require inexpensive, massproducible, lowloss
Dielectric
slab
Plasmainduced
grating
Radiated
power
RF input
Dielectric
waveguide
Figure 8. A commercially available optically
controlled grating antenna.
1814 GRATINGS, GRATING ANTENNAS
antenna arrays. Grating antennas meet these demands at
frequencies up through the W band and beyond.
This article began with an examination of the historical
background of gratings and grating antennas. The funda
mental principles underlying their operation were clearly
described, and an antenna design procedure was present
ed. The article ended with a survey of some of the most
interesting applications to date for millimeterwave grat
ings.
Acknowledgments
The authors wish to thank Prof. R. Nevels of Texas A&M
University for graciously opening to them his extensive
personal notes on periodic structures. They also kindly
appreciate the assistance of Dr. A. Fathy at the University
of Tennessee, of President H. Rutstein at Dorado Interna
tional Corp., and of Ms. T. Quintana at Waveband Corp.
BIBLIOGRAPHY
1. G. K. Sweetnam, The Command of Light, American Philo
sophical Society, Philadelphia: 2000.
2. M. C. Huntley, Diffraction Gratings, Academic Press, New
York, 1982.
3. E. G. Loewen and E. Popov, Diffraction Gratings and Appli
cations, Marcel Dekker, New York, 1997.
4. E. Hecht, Optics, 3rd ed., AddisonWesley, New York, 1998.
5. H. Hertz, Electric Waves, Macmillan, London, 1893.
6. D. Pozar, Microwave Engineering, Wiley, New York, 1990.
7. J. C. Bose, On the determination of the wavelength of electric
radiation by a diffraction grating, Proc. Roy. Soc. 60:167–178
(1897).
8. M. L. Dakss, L. Kuhn, P. F. Heidrich, and B. A. Scott, Grating
coupler for efﬁcient excitation of optical guided waves in thin
ﬁlms, Appl. Phys. Lett. 16(12):523–525 (June 1970).
9. T. Itoh, Application of gratings in a dielectric waveguide for
leakywave antennas and bandreject ﬁlters, IEEE Trans. Mi
crowave Theory Tech. 25(12):1134–1137 (Dec. 1977).
10. K. L. Klohn, R. E. Horn, H. Jacobs, and E. Freibergs, silicon
waveguide frequency scanning linear array antenna, IEEE
Trans. Microwave Theory Tech. 26(10):764–773 (Oct. 1978).
11. S. T. Peng and F. Schwering, Dielectric grating antennas,
R&D Tech. Report CORADCOM 783, U.S. Army CORA
DCOM, Ft. Monmouth, NJ, 1978.
12. K. L. Klohn, R. E. Horn, H. Jacobs, and E. Freibergs, Silicon
waveguide frequency scanning linear array antenna, IEEE
Trans. Microwave Theory Tech. 26(10):764–773 (Oct. 1978).
13. R. E. Horn, H. Jacobs, E. Freibergs, and K. L. Klohn, Elec
tronic modulated beamsteerable silicon waveguide array an
tenna, IEEE Trans. Microwave Theory Tech. 28(6):647–653
(June 1980).
14. R. E. Horn, H. Jacobs, E. Freibergs, and K. L. Klohn, Single
frequency electronicmodulated analog line scanning using a
dielectric antenna, IEEE Trans. Microwave Theory Tech.
30(5):816–820 (May 1982).
15. C. H. Walter, Traveling Wave Antennas, McGrawHill, New
York, 1965.
16. R. C. Honey, A ﬂushmounted leakywave antenna with pre
dictable patterns, IEEE Trans. Anten. Propag. 7(4):320–329
(Oct. 1959).
17. J. Jacobsen, Analytical, numerical, and experimental inves
tigation of guided waves on a periodically striploaded dielec
tric slab, IEEE Trans. Anten. Propag. 18(3):379–388 (May
1970).
18. R. D. Nevels, Periodic structures (personal notes), Texas A&M
Univ.
19. S. T. Peng, T. Tamir, and H. L. Bertoni, Theory of periodic
dielectric waveguides, IEEE Trans. Microwave Theory Tech.
23(1):123–133 (Jan. 1975).
20. S. T. Peng, T. Tamir, and H. L. Bertoni, Correction to ‘‘Theory
of periodic dielectric waveguides,’’ IEEE Trans. Microwave
Theory Tech. 24(8):542 (Aug. 1976).
21. P. Bhartia and I. J. Bahl, Millimeter Wave Engineering, Wiley,
New York, 1984.
22. A. A. Oliner, Leakywave antennas, in R. C. Johnson, ed.,
Antenna Engineering Handbook, 3rd ed., McGrawHill,
New York, 1993.
23. S. K. Koul, Millimeter Wave and Optical Dielectric Integrated
Guides and Circuits, Wiley, New York, 1997.
24. R. M. Knox and P. P. Toulios, Integrated circuits for the mil
limeter through optical frequency range, Proc. Symp. Sub
millimeter Waves, New York, March 1970, pp. 497–516.
25. M. Guglielmi, Radiation from a Metal Strip Grating on a
Dielectric Slab, Ph.D. dissertation, Polytechnic Univ.,
New York, March 1986, pp. 131–152.
26. M. Guglielmi and A. A. Oliner, A practical theory for dielectric
image guide leakywave antennas loaded by periodic metal
strips, Proc. European Microwave Conf., Rome, Italy, Sept.
1987, pp. 549–554.
27. R. E. Horn, H. Jacobs, and E. Freibergs, Millimeterwave os
cillators using imageline or microstrip waveguides, IEEE
Trans. Microwave Theory Tech. 34(2):285–288 (Feb. 1986).
28. R. Frolich and J. Litva, Beamsteerable active array antenna,
Electron. Lett. 28(2):184–185 (Jan. 1992).
29. A. M. Kirk and K. Chang, Integrated image line steerable
active antennas, Int. J. Infrared Millimeter Waves 13(6):
841–851 (June 1992).
30. J. J. Potoczniak, H. Jacobs, C. M. Lo Casio, and G. Novick,
Power combiner with gunn diode oscillators, IEEE Trans. Mi
crowave Theory Tech. 82(5):724–728 (May 1982).
31. Z. W. Li and W. Menzel, A 61 GHz doppler radar using
an inverted strip dielectric waveguide, IEEE MTTS
Int. Microwave Symp. Digest, Albuquerque, NM, June 1992,
pp. 629–632.
32. R. E. Lehmann, N. H. Sun, and J. K. Butler, Integrated Q
band MMIC diffraction grating transmitter, Proc. 1996 IEEE
Int. Symp. Antennas Propagation, Baltimore, MD, July 1996,
pp. 2188–2191.
33. T. T. Trinh, R. Mittra, and R. J. Paleta, Jr., Horn imageguide
leaky wave antenna, IEEE Trans. Microwave Theory Tech.
29(12):1310–1314 (Dec. 1981).
34. A. Basu and T. Itoh, Dielectric waveguidebased leakywave
antenna at 212 GHz, IEEE Trans. Anten. Propag.
46(11):1665–1673 (Nov. 1998).
35. K. Solbach, Slots in dielectric image line as mode launchers
and circuit elements, IEEE Trans. Microwave Theory Tech.
29(1):10–16 (Jan. 1981).
36. H. Tehrani, M. Li, and K. Chang, Broadband microstrip to
dielectric image line transitions, IEEE Microwave Guided
Wave Lett. 10(10):409–411 (Oct. 2000).
37. A. Schary, G. C. Dalman, and C. A. Lee, Dielectric Waveguide
toCoplanar Transmission Line Transitions, U.S. Patent
5,225,797 (July 1993).
GRATINGS, GRATING ANTENNAS 1815
38. V. Manasson, L. Sadovnik, and V. Yepishin, Antenna with a
Plasma Grating, U.S. Patent 5,982,334 (Nov. 1999).
39. V. Manasson and L. Sadovnik, Scanning Antenna Including a
Dielectric Waveguide and a Rotatable Cylinder Coupled
Thereto, U.S. Patent 6,211,836 B1 (April 2001).
40. F. K. Schwering and S. T. Peng, Design of dielectric grating
antennas for millimeterwave applications, IEEE Trans.
Microwave Theory Tech. 31(2):199–209 (Feb. 1983).
41. M. Guglielmi and D. R. Jackson, Broadside radiation from
periodic leakywave antennas, IEEE Trans. Anten. Propag.
41(1):31–37 (Jan. 1993).
42. K. Solbach, Review of dielectric image line antennas, in
K. L. Button, ed., Infrared and Millimeter Waves, Vol. 15,
Academic Press, Orlando, FL, 1986.
43. S. Kobayashi, R. Lampe, R. Mittra, and S. Ray, Dielectric rod
leakywave antennas for millimeterwave applications, IEEE
Trans. Anten. Propag. 29(5):822–824 (Sept. 1981).
44. T. Tamir and S. T. Peng, Analysis and design of grating cou
plers, Appl. Phys. 14:235–254 (Nov. 1977).
45. H. Nishihara, M. Haruna, and T. Suhara, Optical Integrated
Circuits, McGrawHill, New York, 1989, pp. 62–95.
46. R. Mittra and R. Kastner, A spectral domain approach for
computing the radiation characteristics of a leakywave an
tenna for millimeter waves, IEEE Trans. Anten. Propag.
29(4):652–654 (July 1981).
47. M. Guglielmi and A. A. Oliner, Multimode network descrip
tion of a planar periodic metalstrip grating at a dielectric
interface—Part II: Smallaperture and smallobstacle solu
tions, IEEE Trans. Microwave Theory Tech. 37(3):542–552
(March 1989).
48. C. T. Rodenbeck, M. Li, and K. Chang, Design and analysis of
a reconﬁgurable dualbeam grating antenna for lowcost mil
limeterwave beamsteering, IEEE Trans. Anten. Propag. (in
press).
49. R. Petit, Electromagnetic Theory of Gratings, Springer
Verlag, New York, 1980.
50. J. A. Encinar, Modematching and pointmatching techniques
applied to the analysis of metalstriploaded dielectric
antennas, IEEE Trans. Anten. Propag. 38(9):1405–1412
(Sept. 1990).
51. M. Ghomi, B. Lejay, J. L. Amalric, and H. Baudrand, Radia
tion characteristics of uniform and nonuniform dielectric
leakywave antennas, IEEE Trans. Anten. Propag.
41(9):1177–1186 (Sept. 1993).
52. M. Guglielmi and H. Hochstadt, Multimode network descrip
tion of a planar periodic metalstrip grating at a dielectric
interface—Part III: Rigorous solution, IEEE Trans. Micro
wave Theory Tech. 37(5):902–909 (May 1989).
53. S. D. Gedney, J. F. Lee, and R. Mittra, A combined FEM/MoM
approach to analyze the plane wave diffraction by arbitrary
gratings, IEEE Trans. Microwave Theory Tech. 40(2):363–370
(Feb. 1992).
54. M. Chen, B. Houshmand, and T. Itoh, FDTD analysis of a
metalstriploaded dielectric leakywave antenna, IEEE
Trans. Anten. Propag. 45(8):1294–1301 (Aug. 1997).
55. C. T. Rodenbeck, M. Li, and K. Chang, A novel millimeter
wave beamsteering technique using a dielectricimage line
fed grating ﬁlm, IEEE Trans. Anten. Propag. 51(9):2203–2209
(Sept. 2003).
56. C. T. Rodenbeck, M. Li, and K. Chang, A novel millimeter
wave beamsteering technique using a dielectricimage line
fed grating ﬁlm, P. IEEE MTTS Int. Microwave Symp. Digest,
Phoenix, AZ, May 2001, pp. 267–270.
57. C. T. Rodenbeck, M. Li, and K. Chang, A novel multibeam
grating antenna with applications to lowcost millimeter
wave beamsteering, Proc. IEEE MTTS Int. Microwave
Symp. Digest, Seattle, June 2002, pp. 57–60.
58. F. Schwering and A. A. Oliner, Millimeterwave antennas, in
Y. T. Lo, ed., Antenna Handbook, Van NostrandReinhold,
New York, 1988, Chap. 17.
59. T. Hori and T. Itanami, Circularly polarized linear array an
tenna using a dielectric image line, IEEE Trans. Microwave
Theory Tech. 29(9):967–970 (Sept. 1981).
60. C. T. Rodenbeck, M. Li, and K. Chang, Circularpolarized
reconﬁgurable grating antenna for lowcost millimeter
wave beamsteering, IEEE Trans. Anten. Propag.
(in press).
61. T. Itoh and B. Adelseck, Trapped imageguide leakywave an
tennas for millimeterwave applications, IEEE Trans. Anten.
Propag. 30(5):505–509 (May 1982).
62. P. Checcacci, V. Russo, and A. Scheggi, Holographic
antennas, IEEE Trans. Anten. Propag. 18(6):811–813 (Nov.
1970).
63. K. Levis, A. Ittipiboon, A. Petosa, L. Roy, and P. Berini,
Kaband dipole holographic antennas, IEE Proc. Microwave
Anten. Propag. 148(2):129–132 (April 2001).
64. A. E. Fathy, A. Rosen, H. S. Owen, F. McGinty, D. J. McGee,
G. C. Taylor, R. Amantea, P. K. Swain, S. M. Perlow,
and M. ElSherbiny, Siliconbased reconﬁgurable antennas—
concepts, analysis, implementation, and feasibility,
IEEE Trans. Microwave Theory Tech. 51(6):1650–1661 (June
2003).
65. K. Maamria, T. Wagatsuma, and T. Yoneyama, Leaky NRD
guide as a feeder for microwave planar antennas, IEEE
Trans. Anten. Propag. 41(12):1680–1686 (Dec. 1993).
66. R. J. Mailloux, F. K. Schwering, A. A. Oliner, and J. W. Mink,
Antennas III: Array, millimeter wave, and integrated
antennas, in K. Chang, ed., Handbook of Microwave and
Optical Components, Vol. 1, Wiley, New York, 1989, Chap.
12, p. 726.
67. C. T. Rodenbeck, M. Li, and K. Chang, System and Method for
MillimeterWave Beam Steering Using a Grating Film Fed by
Dielectric Integrated Guide, U.S. Patent 60,292,369 (Provi
sional) (May 2001).
68. M. R. Seiler and B. M. Mathena, Millimeterwave beam steer
ing using ‘‘diffraction electronics,’’ IEEE Trans. Anten. Prop
ag. 32(9):987–990 (Sept. 1984).
69. L. Huang, J. C. Chiao, and M. P. De Lisio, An electronically
switchable leaky wave antenna, IEEE Trans. Anten. Propag.
48(11):1769–1772 (Nov. 2000).
70. A. Rosen, R. Amantea, P. J. Stabile, A. E. Fathy, D. B. Gilbert,
D. W. Bechtle, W. M. Janton, F. J. McGinty, J. K. Butler, and
G. A. Evans, Investigation of active antenna arrays at 60
GHz, IEEE Trans. Microwave Theory Tech. 43(9):2117–2125
(Sept. 1995).
71. A. Alphones and M. Tsutsumi, Leaky wave radiation from a
periodically photoexcited semiconductor slab waveguide,
IEEE Trans. Microwave Theory Tech. 43(9):2435–2441 (Sept.
1995).
72. A. Alphones and M. Tsutsumi, Leaky wave radiation of
millimeter waves by photoinduced plasma grating in a semi
conductor slab, IEE Proc. Microwave Anten. Propag.
146(1):77–83 (Feb. 1999).
73. H. Maheri, M. Tsutsumi, and N. Kumagai, Experimental
studies of magnetically scannable leakywave antennas hav
ing a corrugated ferrite slab/dielectric layer structure, IEEE
Trans. Anten. Propag. 36(7):911–917 (July 1988).
1816 GRATINGS, GRATING ANTENNAS
74. V. A. Manasson, L. S. Sadovnik, V. A. Yepishin, and D. Mark
er, An optically controlled mmw beamsteering antenna based
on a novel architecture, IEEE Trans. Microwave Theory Tech.
45(8):1497–1500 (Aug. 1997).
75. H. Rutstein, private communication, June 2002.
76. T. Itoh and F. J. Hsu, Distributed bragg reﬂector gunn oscil
lators for dielectric millimeterwave integrated circuits, IEEE
Trans. Microwave Theory Tech. 82(5):724–728 (May 1982).
77. J. A. Navarro and K. Chang, Integrated Active Antennas and
Spatial Power Combining, Wiley, New York, 1996.
78. K. B. Mallory, R. H. Miller, and P.A. Szente, A simple grating
system millimeter and submillimeter wavelength separation,
IEEE Trans. Microwave Theory Tech. 11(5):433–434 (Sept.
1963).
79. J. Meltaus, J. Salo, E. Noponen, M. M. Salomaa, V. Viikari,
A. Lonnqvist, T. Koskinen, J. Saily, J. Hakli, J. AlaLaurinaho,
J. Mallat, and A. V. Raisanen, Millimeterwave beam shaping
using holograms, IEEE Trans. Microwave Theory Tech.
51(4):1274–1280 (April 2003).
80. B. Schumann, M. Hoft, M. Saglam, H. L. Hartnagel, and
R. Judaschke, A 5 element 450GHz HBV frequency tripler,
IEEE MTTS Int. Microwave Symp. Digest, Philadelphia,
June 2003, pp. 759–762.
81. M. Hoft, J. Weinzierl, and R. Judaschke, Broadband analysis
of a Dband holographic power combining circuit, 2001 IEEE
MTTS Int. Microwave Symp. Digest, Phoenix, AZ, May 2003,
pp. 1407–1410.
82. M. Shahabadi and K. Schunemann, Millimeterwave holo
graphic power splitting/combining, IEEE Trans. Microwave
Theory Tech. 45(12):2316–2323 (Dec. 1997).
83. V. A. Manasson, L. S. Sadovnik, A. Moussessian, and D. B.
Rutledge, Millimeterwave diffraction by a photoinduced
plasma grating, IEEE Trans. Microwave Theory Tech.
43(9):2288–2290 (Sept. 1995).
84. W. Platte, S. Ruppik, and M. Guetschow, Optically induced
maskcontrolled timevariable periodic microwave structures,
IEEE Trans. Microwave Theory Tech. 48(5):846–851 (May
2000).
85. G. F. Brand, Remote millimeterwave beam control by the
illumination of a semiconductor, IEEE Trans. Microwave
Theory Tech. 48(5):855–857 (May 2000).
GREEN’S FUNCTION METHODS
JIANMING JIN
WENG CHO CHEW
University of Illinois at Urbana
Champaign
The Green’s function method is a powerful techniques for
solving boundary value problems. Green’s function was
named after George Green (1793–1841), who developed a
general method to obtain solutions of Poisson’s equation in
potential theory. This method was described in an essay by
Green entitled ‘‘On the application of mathematical anal
ysis to the theories of electricity and magnetism,’’ pub
lished in 1828.
To illustrate the Green’s function method, consider the
electric potential produced by a point electric charge q
1
placed at r
1
in an unbounded homogeneous free space. It is
well known from the elementary theory of electricity [1]
that this potential at r is given by
f
1
(r) =
q
1
4pe r ÷r
1
[ [
(1)
where r ÷r
1
 denotes the distance between the points r
and r
1
and e is a constant called the permittivity. If there is
another point charge q
2
placed at r
2
, the potential pro
duced by this charge is
f
2
(r) =
q
2
4pe r ÷r
2
[ [
(2)
The total potential produced by q
1
and q
2
is then the linear
superposition of f
1
and f
2
:
f(r) =f
1
(r) ÷f
2
(r) =
q
1
4pe r ÷r
1
[ [
÷
q
2
4pe r ÷r
2
[ [
(3)
If there are N point charges in the space, the total poten
tial is given by
f(r) =
N
i =1
f
i
(r) =
N
i =1
q
i
4pe r ÷r
i
[ [
(4)
where S denotes the summation over all point charges and
f
i
denotes the potential due to the ith point charge placed
at r
i
. The procedure described above is known as the prin
ciple of linear superposition.
Next, consider the electric potential produced by a vol
ume electric charge whose charge density is denoted by
r(r). To ﬁnd the potential, we divide the volume of the
charge into many small cubes. The charge within each
small cube is then given by
q
i
 r(r
i
)DV
i
(5)
where r
i
denotes the center of the ith cube and DV
i
denotes
the volume of the cube. Since each cube is very small, it
can be approximated as a point charge, whose potential is
given by
f
i
(r) 
q
i
4pe r ÷r
i
[ [

r(r
i
)DV
i
4pe r ÷r
i
[ [
(6)
According to the principle to linear superposition, the total
potential is then given by
f(r) =
N
i =1
f
i
(r) 
N
i =1
r(r
i
)DV
i
4pe r ÷r
i
[ [
(7)
Clearly, the approximation improves as the volume is di
vided into smaller cubes. In the limit when DV
i
÷0, Eq. (7)
becomes exact. Hence, one obtains
f(r) = lim
DV
i
÷0
o
i =1
r(r
i
)DV
i
4pe r ÷r
i
[ [
(8)
GREEN’S FUNCTION METHODS 1817
which can be written in the integral form as
f(r) =
_
V
r(r
/
) dV
/
4pe r ÷r
/
[ [
(9)
where V denotes the volume of the electric charge.
The potential produced by a point source of unit
strength is called Green’s function. In the example above,
the Green’s function is
G(r; r
/
) =
1
4pe r ÷r
/
[ [
(10)
and the total potential can then be written as
f(r) =
_
V
r(r
/
)G(r; r
/
) dV
/
(11)
It is clear that the Green’s function method treats an
arbitrary source for the potential as a linear superposition
of weighted point sources. It then ﬁnds the potential as the
corresponding linear superposition of the potentials pro
duced by the point sources.
Obviously, once the Green’s function corresponding to
the potential due to a point source is found, the potential
produced by an arbitrary distribution of sources can be
obtained easily. Therefore, for a specific boundary value
problem, instead of ﬁnding the potential for each new
source encountered by solving Poisson’s equation repeat
edly, one can ﬁnd the Green’s function for that problem
only once and obtain solutions to any sources by the prin
ciple of linear superposition. The procedure of ﬁnding
Green’s function is usually much simpler than ﬁnding
the solution to an arbitrary source. To a large extent, a
Green’s function plays the same role as an impulse re
sponse of a linear circuit system. The system response to
any input function can be determined by convolving the
input function with the impulse response of the system.
The Green’s function method has since been expanded to
deal with a large number of different partialdifferential
equations.
In electrodynamics, both the source (electric current
density) and the response (electric or magnetic ﬁeld) are
vectors, each of which has three components. Since each
component of an electric current density can produce all
three components of the electric or magnetic ﬁeld, one has
nine Green functions that relate the response to the
source. This unwieldiness can be alleviated by introduc
ing the concept of the dyadic Green function. A dyadic
Green function, which can be expressed as a 3 3 matrix,
can be considered as a compact representation of the nine
scalar Green’s functions. The ﬁrst use of dyadic Green’s
function was made by Julian Schwinger. The subject was
also covered by Morse and Feshbach [2] in their well
known treatise on the methods of theoretical physics. A
more comprehensive treatment of the dyadic Green func
tions in electromagnetic theory was presented by Tai [3],
who has done much original work on this topic. In his well
known book, Tai derived dyadic Green’s functions for va
riety of electromagnetic problems of practical importance.
Discussions on dyadic Green’s functions can be found in
Collin [4], Kong [5], and Chew [6].
As shall be shown later, the Green’s function method
not only provides a solution to many boundary value prob
lems involving canonical geometries but also leads to in
tegral equations for problems involving more complex
geometries. These integral equations form the basis for a
numerical solution of complex boundary value problems.
1. SCALAR GREEN’S FUNCTIONS
When both the source and response are scalar functions,
the corresponding Green’s function is also scalar and,
hence, the name scalar Green’s function.
1.1. The Delta Function
Since the Green’s function method is based on the repre
sentation of an arbitrary source by the superposition of
point sources, the mathematical representation of a point
source will ﬁrst be described. Consider an electric charge
of unit strength located at point r
/
. When the volume of
the charge approaches zero, the charge density can be
described by a function
d(r ÷r
/
) =
o for r =r
/
0 for rOr
/
_
(12)
Since the total charge remains at unity
_
V
d(r ÷r
/
) dV =
1 for r
/
in V
0 for r
/
not in V
_
(13)
The function deﬁned in Eqs. (12) and (13) is known as the
Dirac delta function, named after P. A. M. Dirac. Clearly,
given an arbitrary function f(r), which is continuous at
r =r
/
, we obtain
_
V
f (r)d(r ÷r
/
) dV =
f (r
/
) for r
/
in V
0 for r
/
not in V
_
(14)
This expression represents a volume source f(r
/
) as a lin
ear superposition of an inﬁnite number of point sources
d(r ÷r
/
).
In one dimension, the delta function can be considered
as the limit of a function
d(x ÷x
/
) = lim
e÷0
u
e
(x ÷x
/
) (15)
where u
e
(x ÷x
/
) is called a delta family. It can be a rect
angular function of width e and height 1=e, or a triangular
function of width 2e and height 1=e, or a Gaussian function
e
÷(x÷x
/
)
2
=2e
2
=e
ﬃﬃﬃﬃﬃﬃ
2p
_
, all centered at x =x
/
. The important fea
ture of the delta function is not its shape, but the fact that
its effective width approaches zero, while its area remains
1818 GREEN’S FUNCTION METHODS
at unity, that is
_
b
a
d(x ÷x
/
) dx =
1 for x
/
in (a; b)
0 for x
/
not in (a; b)
_
(16)
such that
_
b
a
f (x)d(x ÷x
/
) dx =
f (x
/
) for x
/
in (a; b)
0 for x
/
not in (a; b)
_
(17)
The delta function so deﬁned is not a function in the clas
sical sense. For this reason, it is called a symbolic or gen
eralized function [7].
Clearly, the delta function is a symmetric function
d(x ÷x
/
) =d(x
/
÷x) (18)
The threedimensional delta function in the rectangular,
cylindrical, and spherical coordinate systems is related to
the onedimensional delta function by
d(r ÷r
/
) =d(x ÷x
/
)d(y ÷y
/
)d(z ÷z
/
) (19)
d(r ÷r
/
) =
d(r ÷r
/
)d(f ÷f
/
)d(z ÷z
/
)
r
(20)
d(r ÷r
/
) =
d(r ÷r
/
)d(y ÷y
/
)d(f ÷f
/
)
r
2
sin y
(21)
All of these above satisfy Eq. (13).
1.2. OneDimensional Green’s Function
To introduce the concept of Green’s function in one dimen
sion, consider an inﬁnitely long transmission line with a
distributed current source K(x) (3), as illustrated in Fig. 1.
Using Kirchhoff ’s voltage and current laws, one ﬁnds the
relations between the voltage and current as
dV(x)
dx
÷(joL÷R)I(x) =0 (22)
dI(x)
dx
÷(joC÷G)V(x) =K(x) (23)
where o denotes the angular frequency and L, C, R, and G
are the inductance, capacitance, resistance, and conduc
tance of the transmission line per unit length. Eliminating
I(x) in Eqs. (22) and (23), one obtains the differential equa
tion for the voltage as
d
2
V(x)
dx
2
÷g
2
V(x) = ÷(joL÷R)K(x) (24)
where g
2
=(joL÷R)(joC÷G). Since the line is inﬁnitely
long, there is no reﬂected wave; hence, V(x) satisﬁes the
boundary conditions
dV(x)
dx
÷gV(x) =0 for x ÷o (25)
dV(x)
dx
÷gV(x) =0 for x ÷÷o (26)
Since these boundary conditions are imposed when
[x[ ÷o, they are also called radiation conditions.
Instead of solving for V(x) directly from Eqs. (24)–(26),
one can consider the solution of the following differential
equation
d
2
g
0
(x; x
/
)
dx
2
÷g
2
g
0
(x; x
/
) = ÷d(x ÷x
/
) (27)
where g
0
(x) satisﬁes the same radiation condition as V(x).
Since g
0
(x, x
/
) is a point source response and V(x) in Eq.
(24) is due to the source (joL÷R)K(x), according to the
principle of linear superposition, V(x) can be expressed as
a convolution of g
0
(x, x
/
) with (joL÷R)K(x):
V(x) =
_
o
÷o
(joL÷R)K(x
/
)g
0
(x; x
/
) dx
/
(28)
It is evident that once we obtain g
0
(x, x
/
), the voltage on
the transmission line can be evaluated via a simple inte
gration using Eq. (28).
To ﬁnd g
0
(x, x
/
), note that since
d
2
g
0
(x; x
/
)
dx
2
÷g
2
g
0
(x; x
/
) =0 for x > x
/
or xox
/
(29)
one has
g
0
(x; x
/
) =Ae
÷gx
for x > x
/
(30)
g
0
(x; x
/
) =Be
gx
for xox
/
(31)
where the radiation conditions in Eqs. (25) and (26) were
used to determine the sign in front of g. To determine the
unknown coefﬁcients A and B, consider Eq. (27). First,
note that g
0
(x, x
/
) must be continuous at x =x
/
, that is
g
0
(x; x
/
)
x =x
/
÷0
=g
0
(x; x
/
)
¸
¸
¸
¸
x =x
/
÷0
(32)
where x =x
/
÷0 stands for the righthand side of x
/
and
x =x
/
÷0 stands for the lefthand side of x
/
since a discon
tinuity in g
0
(x, x
/
) at x =x
/
would result in a derivative on
d(x ÷x
/
) on the lefthand side of Eq. (27). Next, integrate
Eq. (27) over the region from x
/
÷e to x
/
÷e and in the limit
V(x ) K(x )
I(x )
Figure 1. An inﬁnitely long transmission line excited by a dis
tributed current source.
GREEN’S FUNCTION METHODS 1819
when e ÷0
dg
0
(x; x
/
)
dx
¸
¸
¸
¸
x =x
/
÷0
÷
dg
0
(x; x
/
)
dx
¸
¸
¸
¸
x =x
/
÷0
= ÷1 (33)
Applying these two conditions to Eqs. (30) and (31), one
ﬁnds
g
0
(x; x
/
) =
1
2g
e
÷g(x÷x
/
)
for x > x
/
(34)
g
0
(x; x
/
) =
1
2g
e
g(x÷x
/
)
for xox
/
(35)
or, more compactly
g
0
(x; x
/
) =
1
2g
e
÷g x÷x
/
[ [
(36)
This is the Green function for the inﬁnitely long trans
mission line.
1.3. Two and ThreeDimensional Green’s Function
Consider the electric and magnetic ﬁelds produced by a
timeharmonic electric source whose current density is
denoted by J(r) and charge density is denoted by r(r).
These ﬁelds satisfy Maxwell’s equations given by (1)
VE(r) = ÷joB(r) (37)
VH(r) =joD(r) ÷J(r) (38)
V
.
D(r) =r(r) (39)
V
.
B(r) =0 (40)
and the constitutive relations given by B =mH and D =
eE, where m is the magnetic permeability and e is the elec
tric permittivity. Again, assume that the space is homo
geneous. Taking the curl of Eq. (37), one has
VVE(r) = ÷jomVH(r) (41)
Using Eq. (38) in Eq. (41), one obtains
VVE(r) ÷k
2
E(r) = ÷jomJ(r) (42)
where k
2
=o
2
me. Since VVE=V(V
.
E) ÷V
2
E, Eq. (42)
can be written as
V
2
E(r) ÷k
2
E(r) =jomJ(r) ÷
1
e
Vr(r) (43)
where Eq. (39) has been applied. Similarly, one obtains the
equation for H as
V
2
H(r) ÷k
2
H(r) = ÷VJ(r) (44)
Equations (43) and (44) are inhomogeneous Helmholtz
wave equations.
If one uses f to represent each component of Eor Hin a
Cartesian coordinate system, then f satisﬁes the inhomo
geneous Helmholtz equation
V
2
f(r) ÷k
2
f(r) = ÷f (r) (45)
where f(r) propagates in an inﬁnite unbounded space,
there is no reﬂected wave. Hence, f(r) satisﬁes the radi
ation condition
r
@f
@r
÷jkf
_ _
=0 for r ÷o (46)
where r represents the radial variable in spherical coor
dinates. Instead of solving for f(r) directly from Eqs. (45)
and (46) for each f(r), one ﬁrst ﬁnds its Green function,
which is the solution of the following partialdifferential
equation
V
2
G
0
(r; r
/
) ÷k
2
G
0
(r; r
/
) = ÷d(r ÷r
/
) (47)
subject to the radiation condition in Eq. (46). If G
0
can be
found, using the principle of linear superposition, one
obtains
f(r) =
_
V
G
0
(r; r
/
)f (r
/
) dV
/
(48)
where V is the support of f(r), which is the volume having
nonzero f(r).
To ﬁnd G
0
, we introduce a new coordinate system with
its origin located at r
/
. Thus, the problem has a spherical
symmetry with respect to this point. Equation (47) then
becomes
1
r
2
1
d
dr
1
r
2
1
dG
0
(r
1
; 0)
dr
1
_ _
÷k
2
G
0
(r
1
; 0) = ÷d(r)
1
(49)
where r
1
=r ÷r
/
. When r
1
a0, Eq. (49) can be written as
d
2
[r
1
G
0
(r
1
; 0)]
dr
2
1
÷k
2
r
1
G
0
(r
1
; 0) =0 (50)
which has a wellknown solution
r
1
G
0
(r
1
; 0) =Ae
÷jkr
1
or G
0
(r
1
; 0) =A
e
÷jkr
1
r
1
(51)
The sign in the exponent is chosen such that Eq. (51) sat
isﬁes the radiation condition in Eq. (46). To determine the
unknown coefﬁcient A, substitute Eq. (51) into Eq. (49)
and integrate over a small sphere centered at r
1
=0 with
its radius e ÷0. The result is A=(4p)
÷1
. Therefore
G
0
(r
1
; 0) =
e
÷jkr
1
4pr
1
(52)
1820 GREEN’S FUNCTION METHODS
and in the original coordinates, it becomes
G
0
(r; r
/
) =
e
÷jk r÷r
/
[ [
4p r ÷r
/
[ [
(53)
Following the same procedure, one can obtain the two
dimensional Green function for the Helmholtz equation as
G
0
(q; q
/
) =
1
4j
H
(2)
0
(k q ÷q
/
¸
¸
¸
¸
) (54)
where q =x^ xx ÷y^ yy and H
(2)
0
(k[q ÷q
/
[) is the zerothorder
Hankel function of the second kind.
When one deals with the static electric ﬁeld, Maxwell’s
equations for E(r) reduce to
VE(r) =0 and V
.
E(r) =
r(r)
e
(55)
These two equations can be solved conveniently by intro
ducing the electric potential f(r), which is deﬁned as
E(r) = ÷Vf(r) (56)
The ﬁrst equation in Eq. (55) is automatically satisﬁed
because of the identity VVf(r) ¬ 0. Substituting
Eq. (56) into the second equation in Eq. (55), one obtains
V
2
f(r) = ÷
r(r)
e
(57)
This equation is Poisson’s equation, which can be consid
ered as a special case of Eq. (45) with k =0.
Using the procedure described in this section, one ob
tains the threedimensional Green function for Poisson’s
equation as
G
0
(r; r
/
) =
1
4p r ÷r
/
[ [
(58)
and the twodimensional Green function as
G
0
(q; q
/
) = ÷
1
2p
ln q ÷q
/
¸
¸
¸
¸
(59)
1.4. Classiﬁcation of Green’s Functions
The Green functions derived above are for the inﬁnite un
bounded space where no other objects are present. They
are called the freespace Green’s functions and are denoted
by the subscript 0. When the region of interest is bounded,
one then has to consider boundary conditions for Green’s
function. Different boundary conditions lead to different
Green’s functions. For this reason, Green’s functions are
classiﬁed into three categories: Green’s functions of the
ﬁrst, second, and third kind [3].
Green’s function of the ﬁrst kind, denoted by G
1
, satis
ﬁes the Dirichlet boundary condition, that is
G
0
(r; r
/
) =0 for r on S (60)
where S denotes the boundary of the problem. For a half
space with an inﬁnite ground plane coincident with the z
=0 plane, the Green’s function of the ﬁrst kind for Pois
son’s equation is given by
G
0
(r; r
/
) =G
0
(r; r
/
) ÷G
0
(r; r
/
i
) =
1
4p r ÷r
/
[ [
÷
1
4p r ÷r
/
i
¸
¸
¸
¸
(61)
where r
/
i
=r
/
÷2z
/
^ zz =x
/
^ xx ÷y
/
^ yy ÷z
/
^ zz. This result can be de
rived conveniently using the method of images. It is easy
to see that the Dirichlet boundary condition is satisﬁed by
G
1
(r, r
/
) in the z =0 plane.
The Green’s function of the second kind, denoted by G
2
,
satisﬁes the Neumann boundary condition, that is
@G
2
(r; r
/
)
@n
=0 for r on S (62)
where S denotes the boundary of the problem and q/qn
denotes the normal derivative. For a half space with an
inﬁnite magnetic (symmetry) plane coincident with the z
=0 plane, the Green’s function of the second kind for
Poisson’s equation is given by
G
2
(r; r
/
) =G
0
(r; r
/
) ÷G
0
(r; r
/
i
) =
1
4p r ÷r
/
[ [
÷
1
4p r ÷r
/
i
¸
¸
¸
¸
(63)
where r
/
i
is the same as the one in Eq. (61). It satisﬁes the
Neumann boundary condition in the z =0 plane.
The Green’s function of the third kind is deﬁned for
problems involving two or more media. It can be denoted
as G
(ij)
(r, r
/
), where i indicates the medium where the ﬁeld
point r is located and j indicates the medium where the
source point r
/
is located. Consider, for example, a poten
tial problem involving two halfspaces. The upper half
space (medium 1) above z =0 has a premittivity of e
1
, and
the lower halfspace (medium 2) has a permittivity of e
2
.
The Green’s function for Poisson’s equation is given by (8)
G
(11)
(r; r
/
) =G
0
(r; r
/
) ÷
e
2
÷e
1
e
2
÷e
1
G
0
(r; r
/
i
)
=
1
4p r ÷r
/
[ [
÷
e
2
÷e
1
e
2
÷e
1
1
4p r ÷r
/
i
¸
¸
¸
¸
(64)
and
G
(21)
(r; r
/
) =
2e
2
e
2
÷e
1
G
0
(r; r
/
) =
2e
2
e
2
÷e
1
1
4p r ÷r
/
[ [
(65)
Exchanging e
1
and e
2
in G
(11)
and G
(21)
, one obtains the
expressions for G
(22)
and G
(12)
, respectively. This method
GREEN’S FUNCTION METHODS 1821
of obtaining the Green’s functions of the third kind works
only for Poisson’s equation, but not for the Helmholtz
equation because the standard image method is not ap
plicable to the Helmholtz equation in this case.
1.5. Eigenfunction Expansion
In addition to the conventional method described earlier,
another general method for deriving Green’s functions is
the method of Ohm–Rayleigh or the method of eigenfunc
tion expansion [3]. In this section, one rederives the Green
functions in Eq. (36) and Eq. (53) to illustrate the process
of the Ohm–Rayleigh method.
Consider ﬁrst the solution of Eq. (27). Expand g
0
(x, x
/
)
in terms of a Fourier integral
g
0
(x; x
/
) =
_
o
÷o
A(h)e
jhx
dh (66)
The e
jhx
, which is the solution of the homogeneous differ
ential equation d
2
c(x)/dx
2
÷h
2
c(x) =0, is called the eigen
function and h
2
is the corresponding eigenvalue.
Therefore, Eq. (66) can be considered as the eigenfunction
expansion of g
0
(x, x
/
). To determine A(h), substitute Eqs.
(66) into Eq. (27) and note that
d(x ÷x
/
) =
1
2p
_
o
÷o
e
jh(x÷x
/
)
dh (67)
This yields
A(h) =
e
÷jhx
/
2p(h
2
÷g
2
)
(68)
Hence, we obtain
g
0
(x; x
/
) =
1
2p
_
o
÷o
e
jh(x÷x
/
)
h
2
÷g
2
dh (69)
This is known as the spectral representation of g
0
(x, x
/
).
The integral in this equation can be evaluated using Ca
uchy’s residue theorem [9]. For this, one needs to form a
closed contour for the integral in Eq. (69). In order to sat
isfy the boundary conditions in Eqs. (25) and (26), for
x ÷x
/
40 the inﬁnite integration path must be closed in the
upper halfplane and for x ÷x
/
o0 the inﬁnite path must
be closed in the lower halfplane, as shown in Fig. 2. The
application of Cauchy’s residue theorem yields
g
0
(x; x
/
) =
1
2g
e
÷g(x÷x
/
)
for x > x
/
e
g(x÷x
/
)
for xox
/
_
(70)
which is the same as Eqs. (34) and (35).
Next, consider the solution of Eq. (47). First expand
G
0
(r, r
/
) in terms of Fourier integrals
G
0
(r; r
/
) =
_
o
÷o
A(h)e
jh
.
r
dh (71)
where h=h
x
^ xx ÷h
y
^ yy ÷h
z
^ zz. The e
jh
.
r
, which is the solution
of the homogeneous partial differential equation V
2
c(r) ÷
h
2
c(r) =0, is called the eigenfunction and h
2
=h
2
is the
corresponding eigenvalue. Again, Eq. (71) can be consid
ered as the eigenfunction expansion of G
0
(r, r
/
). Substi
tuting Eq. (71) into Eq. (47), and noting that
d(r ÷r
/
) =
1
(2p)
3
_
o
÷o
e
jh
.
(r÷r
/
)
dh (72)
one ﬁnds
A(h) =
e
÷jh
.
r
/
(2p)
3
(h
2
÷k
2
)
(73)
Therefore
G
0
(r; r
/
) =
1
(2p)
3
_
o
÷o
e
jh
.
(r÷r
/
)
h
2
÷k
2
dh (74)
This is the spectral representation of the threedimen
sional Green function. To evaluate the spectral integral,
let
h
x
=h sin y cos j; h
y
=h sin y sin j; h
z
=h cos y (75)
so that
dh=h
2
siny dh dy dj (76)
Furthermore, because of the spherical symmetry of G
0
with respect to the point r
/
, the value of G
0
is independent
of the direction of r ÷r
/
. Therefore, one can choose an ar
bitrary r ÷r
/
for the evaluation of G
0
. If one chooses the
direction of r ÷r
/
to coincide with the z direction, Eq. (74)
for x > x
for x < x′
′
j j =
[Re(h)]
[Im(h)]
− −
− j j = +
Figure 2. Locations of the two poles in the complex plane and the
closed contours for integration.
1822 GREEN’S FUNCTION METHODS
may be written as
G
0
(r; r
/
) =
1
(2p)
3
_
o
0
_
p
0
_
2p
0
e
jhcos y[r÷r
/
[
h
2
÷k
2
h
2
sin y dh dy dj
=
j
(2p)
2
[r ÷r
/
[
_
o
0
e
÷jh[r÷r
/
[
÷e
jh[r÷r
/
[
_ ¸
hdh
h
2
÷k
2
=
j
(2p)
2
[r ÷r
/
[
_
o
0
he
÷jh[r÷r
/
[
h
2
÷k
2
dh (77)
This integral can now be evaluated using Cauchy’s residue
theorem. The integral has two poles: one at h=k and the
other at h= ÷k. Although the problem considered here is
lossless, treat it as a limiting case of lossy problem for
which k has a small negative imaginary part. Consequent
ly, the pole at h=k is on the lower side of the real axis and
the pole at h= ÷k is on the upper side of the real axis.
In order to satisfy the radiation condition in Eq. (46),
the inﬁnite integration path must be closed in the lower
halfplane, as shown in Fig. 3. Applying Cauchy’s residue
theorem, one obtains
G
0
(r; r
/
) =
e
÷jk[r÷r
/
[
4p[r ÷r
/
[
(78)
which is the same as Eq. (53).
Finally, note that, although the process of the Ohm–
Rayleigh method is more involved than the conventional
method, it is more general and can be used to ﬁnd Green’s
functions in many problems.
1.6. Green’s Functions in a Bounded Region
As can be seen in the preceding section, the spectrum
(eigenvalue) for inﬁnitespace problems is continuous and,
as a result, the spectral representation of the Green func
tion involves spectral integrals. When the region of inter
est is ﬁnite, the spectrum will be discrete. To demonstrate
this fact, consider a grounded rectangular cavity of di
mension abd, depicted in Fig. 4. The Green’s function
for Poisson’s equation satisﬁes the partial differential
equation
V
2
G
1
(r; r
/
) = ÷d(r ÷r
/
) (79)
and the Dirichlet boundary condition
G
1
(r; r
/
) =0 for r on cavity walls (80)
This Green’s function can be derived in a number of dif
ferent ways, such as the conventional method, the method
of images, and the Ohm–Rayleigh method. Here, the
Ohm–Rayleigh method is employed. First, consider the
solution of
V
2
c÷h
2
c=0 (81)
subject to the condition in Eq. (80). Using the method of
separation of variables, one ﬁnds
c
mnp
= sin
mpx
a
sin
npy
b
sin
ppz
d
(82)
which is the eigenfunction of Eq. (81) with eigenvalue h
2
=(mp/a)
2
÷(np/b)
2
÷(pp/d)
2
. This can be used to expand
G
1
:
G
1
(r; r
/
) =
o
m=1
o
n=1
o
p=1
A
mnp
sin
mpx
a
sin
npy
b
sin
ppz
d
(83)
Substituting this expression into Eq. (79), one has
o
m=1
o
n=1
o
p=1
A
mnp
h
2
sin
mpx
a
sin
npy
b
sin
ppz
d
=d(x ÷x
/
)d(y ÷y
/
)d(z ÷z
/
)
(84)
The coefﬁcient A
mnp
can be determined by multiplying
both sides by sin(m
/
px/a), sin(n
/
py/b), and sin(p
/
pz/d) and
+ −
[Re(h)]
[Im(h)]
0
j k
k − j
Figure 3. Locations of the two poles in the complex plane and the
closed contour for integration.
(x)
(z)
(y)
a
d
b
Figure 4. A ground rectangular cavity.
GREEN’S FUNCTION METHODS 1823
integrating over x, y and z. The result is
G
1
(r; r
/
) =
o
m=1
o
n=1
o
p=1
8
abdh
2
sin
mpx
a
sin
mpx
/
a
sin
npy
b
sin
npy
/
b
sin
ppz
d
sin
ppz
/
d
(85)
The triple summation can be reduced to a double summa
tion using the formula [4]
o
p=1
1
h
2
sin
ppz
d
sin
ppz
/
d
=
d
2k
c
sinh k
c
d
sinh k
c
z
o
sinh k
c
(d ÷z
>
)
(86)
where k
c
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
(mp=a)
2
÷(np=b)
2
_
, z
o
=z when zoz
/
, and
z
o
=z
/
when z
/
oz, and z
4
=z when z4z
/
, and z
4
=z
/
when z
/
4z. As a result, Eq. (85) becomes
G
1
(r; r
/
) =
o
m=1
o
n=1
4
ab
sin
mpx
a
sin
mpx
/
a
sin
npy
b
sin
sinh k
c
z
o
sinh k
c
(d ÷z
>
)
k
c
sinh k
c
d
(87)
Next, consider the problem of a parallelplate wave
guide, which is ﬁnite in the y direction and inﬁnite in the x
and z directions, as shown in Fig. 5. Assuming that the
source is uniform in the z direction, Green’s function of the
ﬁrst kind for the Helmholtz equation satisﬁes the partial
differential equation
V
2
G
1
(q; q
/
) ÷k
2
G
1
(q; q
/
) = ÷d(q ÷q
/
) (88)
and the boundary conditions
G
1
(q; q
/
) =0 for y =0; b (89)
and the radiation conditions
@G
1
(q; q
/
)
@x
÷jkG
1
(q; q
/
) =0 for x ÷o (90)
@G
1
(q; q
/
)
@x
÷jkG
1
(q; q
/
) =0 for x ÷÷o (91)
The eigenfunction for this problem is found as
c
n
(h
x
) =e
jh
x
x
sin
npy
b
(92)
from which G
1
can be expanded as
G
1
(q; q
/
) =
_
o
÷o
o
n=1
A
n
(h
x
) e
jhxx
sin
npy
b
dh
x
(93)
Substituting this expression into Eq. (88), one obtains
_
o
÷o
o
n=1
A
n
(h
x
) k
2
÷h
2
x
÷
np
b
_ _
2
_ _
e
jh
x
x
sin
npy
b
dh
x
= ÷d(x ÷x
/
)d(y ÷y
/
)
(94)
The coefﬁcient A
n
(h
x
) can be determined by multiplying
both sides by e
jh
x
x
sin (n
/
py/b) and integrating over x and y.
The result is
G
1
(q; q
/
) =
1
pb
_
o
÷o
o
n=1
h
2
x
÷
np
b
_ _
2
÷k
2
_ _
÷1
e
jh
x
(x÷x
/
)
sin
npy
b
sin
npy
/
b
dh
x
(95)
Using Cauchy’s residue theorem, one can evaluate the
spectral integral in a similar manner to that for the trans
mission line case, yielding
G
1
(q; q
/
) =
1
b
o
n=1
1
g
x
e
÷g
x
x÷x
/
[ [
sin
npy
b
sin
npy
/
b
(96)
where g
x
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
(np=b)
2
÷k
2
_
.
1.7. Scalar Integral Equations
Originally, Green’s function methods were developed for
ﬁnding the genera solution of a boundary value problem
whose Green function can be derived. For many practical
problems, the Green function cannot be derived. As a re
sult, one must resort to a numerical method for the solu
tion of the problem. One such numerical method is based
on an integral equation derived using the Green’s function
method.
To demonstrate the formulation of integral equations,
consider the problem of a scalar wave produced by a
source f(r) in the presence of an arbitrarily shaped object
immersed in an inﬁnite medium, as illustrated in Fig. 6.
Exterior to the object, the wavefunction f(r) satisﬁes the
inhomogeneous Helmholtz equation in Eq. (45) and the
(y )
(x )
y = b
Figure 5. A parallelplate waveguide.
1824 GREEN’S FUNCTION METHODS
radiation boundary condition in Eq. (46). Since the object
has an arbitrary shape, no closedform Green function can
be found for this problem. However, one can establish an
integral equation for this problem using the freespace
Green function given in Eq. (53), which is the solution of
Eq. (47) under the condition in Eq. (46).
First, multiply Eq. (45) with G
0
, Eq. (47) with f, and
integrate the difference of the resultant equations over the
entire exterior volume, yielding
_
Vo
G
0
(r; r
/
)V
2
f(r) ÷f(r)V
2
G
0
(r; r
/
)
_ ¸
dV
= ÷
_
V
s
G
0
(r; r
/
)f (r) dV ÷
_
V
o
f(r)d(r ÷r
/
) dV
(97)
where V
o
denotes the inﬁnite space exterior to the object
and V
s
denotes the support of f(r). Applying the second
scalar Green’s theorem (1)
_
V
aV
2
b ÷bV
2
a
_ _
dV =
_
S
a
@b
@n
÷b
@a
@n
_ _
dS (98)
where S denotes the surface enclosing V, one obtains
_
S
0
÷So
G
0
(r; r
/
)
@f(r)
@n
÷f(r)
@G
0
(r; r
/
)
@n
_ _
dS
÷
_
Vo
f(r)d(r ÷r
/
) dV
= ÷
_
V
s
G
0
(r; r
/
)f (r) dV
(99)
where S
o
denotes the surface of the object and S
o
denotes
a spherical surface with a radius approaching inﬁnity.
Since both G
0
and f satisfy Eq. (46), the surface integral
over S
o
vanishes. Consequently, one has
_
So
G
0
(r; r
/
)
@f(r)
@n
÷f(r)
@G
0
(r; r
/
)
@n
_ _
dS
÷
_
V
o
f(r)d(r ÷r
/
) dV = ÷
_
V
s
G
0
(r; r
/
)f (r) dV
(100)
where the normal unit vector on S
o
points toward the in
terior of the object. Using Eq. (14), one obtains
_
S
o
G
0
(r; r
/
)
@f(r)
@n
÷f(r)
@G
0
(r; r
/
)
@n
_ _
dS
÷
_
Vs
G
0
(r; r
/
)f (r) dV
=
f(r
/
) for r
/
inV
o
0 for r
/
in V
o
_
_
_
(101)
where V
o
denotes the volume of the object. Exchanging
r and r
/
and using the symmetry property of G
0
[i.e.,
G
0
(r
/
, r) =G
0
(r, r
/
)], one has
_
S
o
G
0
(r; r
/
)
@f(r)
@n
/
÷f(r)
@G
0
(r; r
/
)
@n
/
_ _
dS
/
÷
_
Vs
G
0
(r; r
/
)f (r
/
) dV
/
=
f(r) for r in V
o
0 for r in V
o
_
_
_
(102)
Equation (102) is an important result, which has several
implications. First, notice that when the object is absent,
the surface integral vanishes. Hence
f(r) =
_
Vs
G
0
(r; r
/
)f (r
/
) dV
/
(103)
which is the same as Eq. (48). This may be called the in
cident ﬁeld impinging on the object and be denoted as
f
inc
(r). Second, when there is no source in V
o
, Eq. (102)
become
f(r) =
_
S
o
G
0
(r; r
/
)
@f(r
/
)
@n
/
÷f(r
/
)
@G
0
(r; r
/
)
@n
/
_ _
dS
/
(104)
for r in V
o
. Since there is no source in V
o
, the ﬁeld on S
o
must be produced by the source inside S
o
. This equation
indicates that the ﬁeld in a sourcefree region can be cal
culated by applying knowledge of the potential and its
normal derivative on the surface enclosing the region.
This is the mathematical representation of the well
known Huygens’ principle for a scalar wave.
Equation (102) also provides the foundation to estab
lish an integral equation for f and @f/@n on the surface of
the object. If the object is impenetrable with a hard surface
where f satisﬁes the boundary condition
f(r) =0 for r on S
o
(105)
S
o
V
o
V
n
∞
Figure 6. An object occupying volume V
o
.
GREEN’S FUNCTION METHODS 1825
Eq. (102) becomes
f
inc
(r) ÷
_
S
o
G
0
(r; r
/
)
@f(r
/
)
@n
/
dS
/
=
f(r) for r in V
o
0 for r in V
o
_
_
_
(106)
Applying this equation on S
o
, one obtains
_
S
o
G
0
(r; r
/
)
@f(r
/
)
@n
/
dS
/
= ÷f
inc
(r) for r on S
o
(107)
which is the integral equation for @f/@n on S
o
.
If the object is impenetrable with a soft surface where f
satisﬁes the boundary condition
@f(r)
@n
=0 for r on S
o
(108)
Eq. (102) becomes
f
inc
(r)
÷
_
S
o
f(r
/
)
@G
0
(r; r
/
)
@n
/
dS
/
=
f(r) for r in V
o
0 for r in V
o
_
(109)
Applying this equation on S
o
, one obtains
1
2
f(r) ÷÷
_
So
f(r
/
)
@G
0
(r; r
/
)
@n
/
dS
/
=f
inc
(r) for r on S
o
(110)
where ÷
_
denotes the integral excluding the contribution
from the singular point, which is known as the principal
value integral. This result is obtained as follows. The in
tegral over S
o
in Eq. (109) is divided into an integral over a
small circular disk with center at r plus the remaining
integral, which is represented as a principalvalue inte
gral in the limit as the area of the isolated disk approach
es zero. If r approaches S
o
from the outside, the integral
over the vanishingly small disk can be evaluated to give—
f(r)/2. If r approaches S
o
from the inside, the integral
gives f(r)/2. In either case, one obtains Eq. (110), which is
the integral equation for f on S
o
.
If the object is penetrable and homogeneous, apply Eq.
(102) on S
o
to obtain
1
2
f(r) ÷÷
_
So
G
0
(r; [ r
/
)
@f(r
/
)
@n
/
÷f(r
/
)
@G
0
(r; r
/
)
@n
/
_
dS
/
=f
inc
(r) for r on S
o
(111)
To solve for f and @f/@n on S
o
, another equation is needed,
which can be derived by considering the interior of the
object. The wavefunction inside the object satisﬁes the
Helmholtz equation
V
2
f(r) ÷
~
kk
2
f(r) =0 (112)
where
~
kk characterizes the property of the object. Multi
plying this equation by the Green function for unbounded
space ﬁlled with material characterized by
~
kk
~
GG
0
(r; r
/
) =
e
÷j
~
kk[r÷r
/
[
4p[r ÷r
/
[
(113)
and applying a similar derivation as before, one has
÷
_
S
o
~
GG
0
(r; r
/
)
@f(r
/
)
@n
/
÷f(r
/
)
@
~
GG
0
(r; r
/
)
@n
/
_ _
dS
/
=
0 for r in V
o
f(r) for r in V
o
_
_
_
(114)
When this is applied on S
o
, one obtains the second integral
equation
1
2
f(r) ÷÷
_
So
~
GG
0
(r; r
/
)
@f(r
/
)
@n
/
÷f(r
/
)
@
~
GG
0
(r ÷r
/
)
@n
/
_ _
.
dS
/
=0 for r on S
o
(115)
which can be used together with Eq. (111) for a numerical
solution of f and @f/@n on S
o
.
If the object is penetrable and inhomogeneous, the
wavefunction still satisﬁes Eq. (112); however,
~
kk now is a
function or r. In this case, one can write Eqs. (45) and (112)
in one equation:
V
2
f(r) ÷k
2
f(r) = ÷f (r) ÷
~
kk
2
(r) ÷k
2
_ _
f(r) (116)
Multiplying this equation by G
0
and integrating over the
inﬁnite volume, one obtains
f(r) ÷
_
V
0
~
kk
2
(r
/
) ÷k
2
_ _
G
0
(r; r
/
)f(r
/
) dV
/
=
_
V
s
G
0
(r; r
/
)f (r
/
) dV
/
(117)
This is the integral equation that can be used to solve for f
in V
o
. Unlike the previous integral equations, this equation
involves the volume integral. For this reason, it is often
referred to as the volume integral equation, whereas the
previous ones are often referred to as the surface integral
equations. Integral equations for more complicated objects
may involve both volume and surface integrals [10].
2. DYADIC GREEN’S FUNCTIONS
When both the source and response are vector functions,
the corresponding Green function is a dyad; hence, the
term, dyadic Green function.
1826 GREEN’S FUNCTION METHODS
2.1. Definition of Dyad
A dyad, denoted by D, is formed by two vectors:
D=AB (118)
This entity by itself does not have any physical interpre
tation as a vector. However, when it acts on another vector,
the result becomes meaningful. The major role of a dyad is
that its scalar product with a vector produces another
vector of different magnitude and direction. For example,
its anterior scalar product with vector C yields
C
.
D=(C
.
A)B (119)
which is a vector. Its posterior scalar product with vector
C yields
D
.
C=A(B
.
C) (120)
which is also a vector. Apparently, the resulting vectors in
Eqs. (119) and (120) are different. In addition to the two
scalar products, there are two vector products. The ante
rior vector product is deﬁned as
CD=(CA)B (121)
and the posterior vector product is deﬁned as
DC=A(BC) (122)
Clearly, these products are dyads.
The dyad deﬁned in Eq. (118) is a special entity, since it
contains only six independent components, three in each
of the two vectors. A more general dyad, also called a ten
sor, is deﬁned as
D=D
x
^ xx ÷D
y
^ yy ÷D
z
^ zz (123)
where D
x
, D
y
, and D
z
are vectors. Therefore, Eq. (123) can
be expressed as
D=D
xx
^ xx^ xx ÷D
yx
^ yy^ xx ÷D
zx
^ zz^ xx ÷D
xy
^ xx^ yy ÷D
yy
^ yy^ yy ÷D
zy
^ zz^ yy
÷D
xz
^ xx^ zz ÷D
yz
^ yy^ zz ÷D
zz
^ zz^ zz
(124)
which contains nine independent components.
A special dyad is called the unit dyad or identity dyad,
deﬁned as
I = ^ xx^ xx ÷ ^ yy^ yy ÷ ^ zz^ zz (125)
It is evident that
C
.
I =I
.
C=C (126)
2.2. FreeSpace Dyadic Green’s Functions
Consider the electric and magnetic ﬁelds produced by an
electric current source J(r) in an unbounded space. Max
well’s equation for this problem are given in Eqs. (37)–(40),
which lead to Eq. (42), reproduced here as
VVE(r) ÷k
2
E(r) = ÷jomJ(r) (127)
This is the vector wave equation, which is the analog of
the scalar Helmholtz wave equation. It describes electro
magnetic wave phenomena that are very pervasive in
modern technologies, such as in communication, micro
wave, and computer chip technologies.
Just as in the scalar case, one can derive a dyadic
Green function G
e0
, whose end result is to relate E(r) and
J(r) by
E(r) = ÷jom
_
V
G
e0
(r; r
/
)
.
J(r
/
) dV
/
(128)
where Vis the support of the current J(r). Using Eq. (128)
in Eq. (127), one obtains
÷jom
_
V
VVG
e0
(r; r
/
)
.
J(r
/
) dV
/
÷jomk
2
_
V
G
e0
(r; r
/
)
.
J(r
/
) dV
/
= ÷jomJ(r) = ÷jom
_
V
Id(r ÷r
/
)
.
J(r
/
) dV
/
(129)
For arbitrary J(r), this could be satisﬁed only if
VVG
e0
(r; r
/
) ÷k
2
G
e0
(r; r
/
) =Id(r ÷r
/
) (130)
where G
e0
(r; r
/
) is the dyadic Green function of the electric
type that relates vector ﬁeld E to vector current J.
Taking the curl of Eq. (128) and using Maxwell’s equa
tions, one obtains
H(r) =
_
V
VG
e0
(r; r
/
)
.
J(r
/
) dV
/
=
_
V
G
m0
(r; r
/
)
.
J(r
/
) dV
/
(131)
where G
m0
(r; r
/
) =VG
e0
(r; r
/
) is the dyadic Green’s func
tion of the magnetic type. It satisﬁes the equation
VVG
m0
(r; r
/
) ÷k
2
G
m0
(r; r
/
) =V[Id(r ÷r
/
)] (132)
Therefore, the task of ﬁnding the dyadic Green function
of the electric type is reduced to the task of solving Eq.
(130). Equation (130) can be made less difﬁcult by taking
the posterior scalar product with an arbitrary vector a,
yielding
VVG
e0
(r; r
/
)
.
a ÷k
2
G
e0
(r; r
/
)
.
a=ad(r ÷r
/
) (133)
Recognizing that G
e0
(r; r
/
). a represents a vector, one may
use the vector identify VVA=V(V· A) ÷V
2
A to ﬁnd
÷V
2
G
e0
(r; r
/
)
.
a ÷k
2
G
e0
(r; r
/
)
.
a
=ad(r ÷r
/
) ÷V[V
.
G
e0
(r; r
/
)
.
a]
(134)
GREEN’S FUNCTION METHODS 1827
Taking the divergence of Eq. (133) and making use of the
fact that V· (VA)¬0, it can be seen that
V · G
e0
(r; r
/
)
.
a= ÷
1
k
2
V
.
[ad(r ÷r
/
)] (135)
Using this in Eq. (134), one obtains
V
2
G
e0
(r; r
/
)
.
a÷k
2
G
e0
(r; r
/
)
.
a=
÷ 1 ÷
VV
.
k
2
_ _
[ad(r ÷r
/
)] (136)
By making use of the fact that
V
2
G
0
(r; r
/
) ÷k
2
G
0
(r; r
/
) = ÷d(r ÷r
/
) (137)
and the fact the 1 ÷VV· /k
2
is a linear operator that com
mutes with V
2
, it can be deduced that
G
e0
(r; r
/
)
.
a= 1 ÷
VV
.
k
2
_ _
[aG
0
(r; r
/
)] (138)
Writing this as
G
e0
(r; r
/
)
.
a= I ÷
VV
k
2
_ _
G
0
(r; r
/
)
.
a (139)
and since a is an arbitrary vector, one deduces that
G
e0
(r; r
/
) = I ÷
VV
k
2
_ _
G
0
(r; r
/
) (140)
The freespace dyadic Green function of the magnetic
type can be derived as
G
m0
(r; r
/
) =VG
e0
(r; r
/
) =V[IG
0
(r; r
/
)]
=VG
0
(r; r
/
) I
(141)
This is the explicit representation of the dyadic Green’s
function in terms of the scalar Green’s function G
0
(r, r
/
). It
is to be noted that the aforementioned relationship be
tween the dyadic Green function and the scalar Green
function G
0
(r, r
/
) is valid only for a homogeneous un
bounded space such as a free space. Such a relation does
not hold true in a cavity, waveguide, or halfspace. For
example, the dyadic Green’s functions for a halfspace
(above z =0) are given by [3]
G
e1
(r; r
/
) = I ÷
VV
/
k
2
_ _
[G
0
(r; r
/
)
÷G
0
(r; r
/
i
)] ÷2^ zz^ zzG
0
(r; r
/
i
)
(142)
and
G
m2
(r; r
/
) =VG
0
(r; r
/
) I ÷VG
0
(r; r
/
i
) I
i
(143)
where r
/
i
=x
/
^ xx ÷y
/
^ yy ÷z
/
^ zz and I
i
= ÷I ÷2^ zz^ zz. It can be ver
iﬁed that G
e1
(r; r
/
) and G
m2
(r; r
/
) satisfy the boundary
conditions
^ zz G
e1
(r; r
/
) =0 for z =0 (144)
and
^ zz VG
m2
(r; r
/
) =0 for z =0 (145)
respectively. For this reason, G
e1
(r; r
/
) is called the ‘‘elec
trictype dyadic Green function of the ﬁrst kind’’ and
G
m2
(r; r
/
) is called the ‘‘magnetictype dyadic Green’s func
tion of the second kind’’. The classiﬁcation of dyadic
Green’s functions is similar to that of scalar Green’s func
tions.
2.3. Eigenfunction Expansion
As in the scalar case, the Ohm–Rayleigh method or the
method of eigenfunction expansion is a general method to
derive dyadic Green function [3]. For vector problems, the
eigenfunctions are vector functions, known as vector
wavefunctions. There are three kinds of vector wavefunc
tions [1], deﬁned by
L(r) =Vc(r) (146)
M(r) =V[cc(r)] (147)
N(r) =
1
k
VM(r) (148)
where c is a vector called the ‘‘pilot vector’’ and c satisﬁes
that homogeneous Helmholtz wave equation
V
2
c(r) ÷k
2
c(r) =0 (149)
It can be shown that, L, M, and N satisfy the vector equa
tions
V
2
L(r) ÷k
2
L(r) =0 (150)
VVM(r) ÷k
2
M(r) =0 (151)
VVN(r) ÷k
2
N(r) =0 (152)
and M can be expressed in terms of N as
M(r) =
1
k
VN(r) (153)
Since VL(r) = VVc(r) =0, L is the irrotational vector
wavefunction. Since V
.
M(r) =0 and V
.
N(r) =0, M and N
are the solenoidal vector wavefunctions.
For a rectangular waveguide illustrated in Fig. 7, c is
given by
ce
o
mn
(h; r) =
cos k
x
x cos k
y
y
sin k
x
x sin k
y
y
_ _
e
÷jhz
(154)
1828 GREEN’S FUNCTION METHODS
where k
x
=mp/a and k
y
=np/b. The vector wavefunctions
L, M, and N are given by
L
e
o
mn
(h; r) =Vce
o
mn
(h; r) (155)
M
e
o
mn
(h; r) =V[ ^ zzce
o
mn
(h; r)] (156)
N
e
o
mn
(h; r) =
1
k
VV[ ^ zzce
o
mn
(h; r)] (157)
where the pilot vector is c = ^ zz. This causes M to be trans
verse to ^ zz.
The vector wavefunctions are always orthogonal to
each other. For those in a rectangular waveguide, it can
be shown
_
V
U
e
o
mn
(h; r)
.
V
e
o
m
/
n
/ (÷h
/
; r) dV =0 (158)
where U,V=L,M,N, except when Ue
o
mn
(h; r) =Ve
o
mn
(h; r).
They form a complete set and, therefore, can be employed
to expand any vector functions.
The electrictype dyadic Green’s function of the ﬁrst
kind satisﬁes the equation
VVG
e1
(r; r
/
)÷k
2
G
e1
(r; r
/
) ¬ Id(r ÷r
/
) (159)
and the boundary condition
^ nnG
e1
(r; r
/
) =0 on the wageguide walls (160)
It is clear that only L
omn
M
emn
, and N
omn
satisfy Eq. (160)
and, therefore, can be used to expand G
e1
:
G
e1
(r; r
/
) =
_
o
÷o
m;n
[ L
omn
(h; r)A
omn
(h) ÷M
emn
(h; r)B
emn
(h)
÷N
omn
(h; r)C
omn
(h)] dh (161)
Substituting this expansion into Eq. (159), one obtains
_
o
÷o
m;n
{÷k
2
L
omn
(h; r)A
omn
(h)
÷(k
2
÷k
2
)[ M
emn
(h; r)B
emn
(h)]
÷N
omn
(h; r)C
omn
(h)]] dh=Id(r ÷r
/
)
(162)
Taking the anterior scalar product of Eq. (162) with L
om
/
n
/
( ÷h
/
, r), M
em
/
n
/ ( ÷h
/
, r), and N
om
/
n
/ ( ÷h
/
, r), respectively,
integrating over the entire volume of the waveguide, and
applying the orthogonal relation in Eq. (158), one can ﬁnd
A
omn
(h) = ÷
k
2
cnm
k
2
k
2
C
mn
L
omn
(÷h; r
/
) (163)
B
emn
(h) =
1
k
2
÷k
2
C
mn
M
emn
(÷h; r
/
) (164)
C
emn
(h) =
1
k
2
÷k
2
C
mn
N
omn
(÷h; r
/
) (165)
where k
2
cmn
=k
2
x
÷k
2
y
and C
mn
=(2 ÷d
0
)=(pabk
2
cmn
) with
d
0
=1 when m=0 or n=0 and d
0
=0, where m and n are
nonzero. Therefore
G
e1
(r; r
/
) =
_
o
÷o
m;n
C
mn
÷
k
2
cmn
k
2
k
2
L
omn
(h; r)L
omn
(÷h; r
/
)
_
÷
1
k
2
÷k
2
[ M
emn
(h; r)M
emn
(÷h; r
/
)
÷N
omn
(h; r)N
omn
(÷h; r
/
)]
_
dh
(166)
Through some mathematical manipulations and the ap
plication of Cauchy’s residue theorem [3], one can simplify
Eq. (166) as
G
e1
(r; r
/
) = ÷
1
k
2
^ zz^ zzd(r ÷r
/
) ÷
j
ab
m;n
2 ÷d
0
k
2
cmn
k
gmn
[ M
emn
(±k
gmn
; r)M
emn
(±k
gmn
; r)M
emn
((k
gmn
; r
/
)
N
omn
(±k
gmn
; r)N
omn
((k
gmn
; r
/
)] z_z
/
(167)
where k
cmn
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
k
2
÷k
2
cmn
_
.
In addition to the method described above, Tai [3]
proposed the method of G
m
, in which G
m
is derived ﬁrst
and G
e
is then derived from VG
m
=Id(r ÷r
/
) ÷k
2
G
e
.
Since G
m
is completely solenoidal, its expansion requires
only M and N and, therefore, the derivation becomes
simpler.
The Ohm–Rayleigh method can be used to derive a va
riety of dyadic Green’s functions. Table 1 lists the prob
lems for which the dyadic Green’s functions have been
derived.
(x )
(z )
a
b
(y )
Figure 7. A rectangular waveguide.
GREEN’S FUNCTION METHODS 1829
2.4. Vector Integral Equations
Consider the problem of the electric and magnetic ﬁelds
produced by an electric current source J(r) in the presence
of an arbitrarily shaped object immersed in an inﬁnite
homogeneous medium (see Fig. 6). Exterior to the object,
the electric ﬁeld satisﬁes the vector wave equation in
Eq. (127), and the radiation condition at inﬁnity is given by
r[VE(r) ÷jk^ rr E(r)] =0 for r ÷o (168)
Multiplying Eq. (127) by G
e0
(r; r
/
), Eq. (130) by E(r), and
integrating the difference of the resultant equations over
the exterior region, one obtains
_
V
o
[VVE(r)]
.
G
e0
(r; r
/
) ÷E(r)
.
[VVG
e0
(r; r
/
)]
_ _
dV
= ÷jom
_
Vs
J(r)
.
G
e0
(r; r
/
) dV ÷
_
Vo
E(r) · Id(r ÷r
/
) dV
(169)
where V
o
denotes the inﬁnite space exterior to the object
and V
s
denotes the support of J(r). Applying the vector
dyadic Green’s second identity [27]
_
V
[(VVA)
.
D÷A
.
(VVD)] dV
=
_
S
[( ^ nnA)
.
(VD) ÷( ^ nnVA)
.
D] dS
(170)
where V is a volume enclosed by S, one has
_
S
o
÷S
o
[ ^ nnE(r)]
.
[VG
e0
(r; r
/
)]
_
÷ ^ nnVE(r)]
.
G
e0
(r; r
/
)
_
dS
= ÷jom
_
V
s
J(r)
.
G
e0
(r; r
/
) dV ÷
_
V
o
E(r)d(r ÷r
/
) dV
(171)
where S
o
denotes the surface of the object, S
o
denotes a
large spherical surface whose radius approaches inﬁnity,
and ^ nn is the normal unit vector pointing away from V
o
.
Since both E(r) and G
e0
(r; r
/
) satisfy the radiation condi
tion, the surface integral over S
o
vanishes. As a result, we
obtain
÷
_
S
o
[ ^ nnE(r)]
.
[VG
e0
(r; r
/
)]
_
÷[ ^ nnVE(r)]
.
G
e0
(r; r
/
)
_
dS
÷jom
_
V
s
J(r)
.
G
e0
(r; r
/
) dV =
E(r
/
) for r
/
in V
o
0 for r
/
in V
o
_
_
_
(172)
which can also be written as
÷
_
So
[VG
e0
(r; r
/
)]
.
[ ^ nn
/
E(r
/
)]
_
÷jomG
e0
(r; r
/
)
.
[ ^ nn
/
H(r
/
)]
_
dS
/
÷jom
_
Vs
G
e0
(r; r
/
)
.
J(r
/
) dV
/
=
E(r) for r in V
o
0 for r in V
o
_
_
_
(173)
where V
o
denotes the volume of the object.
Similar to Eq. (102) in the scalar case, Eq. (173) is an
important result, which has several implications. First,
notice that when the object is absent, the surface integral
vanishes. Hence
E(r) = ÷jom
_
V
s
G
e0
(r; r
/
)
.
J(r
/
) dV
/
(174)
where is the same as Eq. (128). This can be regarded as
the incident ﬁeld and denoted as E
inc
(r). Second, when
there is no source in V
o
, Eq. (173) becomes
E(r) = ÷
_
S
o
[VG
e0
(r; r
/
)][ ^ nn
/
E(r
/
)]
.
÷jomG
e0
(r; r
/
)
_
.
[ ^ nn
/
H(r
/
)]
_
dS
/
(175)
for r in V
o
. Since there is no source in V
o
, the ﬁeld on S
o
must be produced by the source inside S
o
. This equation
indicates that the ﬁeld in a sourcefree region can be cal
culated from knowledge of the tangential electric and
Table 1. Problems with Available Dyadic Green Functions
Geometry of Problem References
Parallelplate waveguide 3
Rectangular waveguide 3,11
Rectangular waveguide with two dielectrics 3,12
Cylindrical waveguide 3,13,14
Coaxial waveguide 3,15
Rectangular cavity 3,16
Cylindrical cavity 3,13,17
Spherical cavity 3,18
Circular conducting cylinder 3
Circular dielectric cylinder 3
Circular coated cylinder 3
Elliptical conducting cylinder 3
Conducting wedge and halfsheet 3
Conducting sphere and cone 3
Homogeneous and inhomogeneous spheres 3
Planar layered medium 3,5,6
Planar anisotropic layered medium 19
Conductorbacked layered medium 20–22
Cylindrically layered medium 6,23
Spherically layered medium 3,6,24
Moving medium 3,25,26
1830 GREEN’S FUNCTION METHODS
magnetic ﬁelds on the surface enclosing the region. This is
the mathematical representation of the wellknown Huy
gens’ principle for a vector ﬁeld.
Equation (173) also provides the foundation to
establish an integral equation for ^ nnE and ^ nnH
on the surface of the object. If the object is a perfect con
ductor, ^ nnE(r) =0 for r on S
o
. Consequently, Eq. (173)
becomes
E(r) =E
inc
(r) ÷jom
_
S
o
G
e0
(r; r
/
)]
.
[ ^ nn
/
H(r
/
)]dS
/
(176)
for r in V
o
. Substituting this into ^ nnE(r) =0 for r on S
o
we obtain an integral equation, which can be solved for
^ nnH(r).
If the object is a homogeneous body, one can derive an
other integral representation for the ﬁeld inside S
o
using
the unboundedspace dyadic Green’s function for the in
terior medium. When this and Eq. (173) are applied at S
o
,
one obtains two integral equations, which can be solved for
^ nnE(r) and ^ nnH(r).
If the object is an inhomogeneous dielectric body, the
electric ﬁeld satisﬁes the vector wave equation
VVE(r) ÷
~
kk
2
(r)E(r) = ÷jomJ(r) (177)
This can be written as
VVE(r) ÷k
2
E(r) = ÷jomJ(r) ÷[
~
kk
2
(r)
÷k
2
]E(r) (178)
Multiplying Eq. (177) by G
e0
(r; r
/
) and integrating the re
sultant equation over the entire space, one obtains
E(r) =E
inc
(r) ÷
_
Vo
G
e0
(r; r
/
)
.
[
~
kk
2
(r
/
) ÷k
2
]E(r
/
) dV
/
(179)
This is the mathematical representation of the volume
equivalence principle. It provides a volume integral equa
tion that can be solved for E(r).
We note that the formulation described in this section
can be repeated for the magnetic ﬁeld in a similar manner.
As a result, different integral equations exist for the same
problem, which provide different approaches to the solu
tion of the problem.
2.5. Singularity of the Dyadic Green’s Function
As shown in Eq. (128), the electric ﬁeld produced by the
current J in an unbounded space can be written as
E(r) = ÷jom
_
V
G
e0
(r; r
/
)
.
J(r
/
) dV
/
(180)
where G
e0
(r; r
/
) is deﬁned by Eq. (140). Many electromag
neticists have tried to fathom the meaning of Eq. (180).
Strictly speaking, the integral does not converge because
of the 1/r  r
/
 singularity in G
0
(r, r
/
). After being oper
ated upon by the double V operator in Eq. (140), G
e0
(r; r
/
)
contains terms of the form 1/r  r
/

3
, rendering the inte
gral in Eq. (180) illdeﬁned. A remedy to this is to rewrite
Eq. (180) as
E(r) = ÷jom I ÷
VV
k
2
_ _
.
_
V
G
0
(r; r
/
)J(r
/
) dV
/
(181)
This equation is well deﬁned for all r and r
/
, but lacks the
compactness of Eq. (180).
Equation (180) can be made meaningful in a general
ized function sense. To this end, one deﬁnes G
e0
(r; r
/
) as a
generalized function
G
e0
(r; r
/
) =PVG
e0
(r; r
/
) ÷
Ld(r ÷r
/
)
k
2
(182)
where PV implies the invokement of a principal volume
integral whose value depends on the shape of the principal
volume chosen. For the sake of uniqueness, L also depends
on the shape of the principal volume. A principal volume
integral is deﬁned as
_
V
PVG
e0
(r; r
/
)
.
J(r
/
) dV
/
=PV
_
V
G
e0
(r; r
/
)
.
J(r
/
) dV
/
= lim
V
d
÷0
_
V÷V
d
G
e0
(r; r
/
)
.
J(r
/
) dV
/
(183)
where V
d
is the exclusion volume. Unfortunately, even
though the integral above converges, its value is
nonunique in the sense that it depends on the shape of
V
d
. The nonuniqueness in the ﬁrst term of Eq. (182) is
rectiﬁed by the choice of a shapedependent L. The volume
of L in Eq. (182) for various exclusion volumes is given by
[28–33].
L=
I
3
for spheres and cubes (184)
L= ^ zz^ zz for disks perpendicular to the z axis (185)
L=
^ xx^ xx ÷ ^ yy^ yy
2
for needles parallel to the z axis (186)
To understand how these L values are derived, one can
start with the classically legitimate Eq. (181) and split the
integral into two terms with the definition of a principal
volume integral:
E(r) = ÷jom lim
V
d
÷0
I ÷
VV
k
2
_ _
.
_
V÷V
d
G
0
(r; r
/
)J(r
/
) dV
/
÷jom lim
V
d
÷0
I ÷
VV
k
2
_ _
.
_
V
d
G
0
(r; r
/
)J(r
/
) dV
/
(187)
GREEN’S FUNCTION METHODS 1831
In the ﬁrst term, V
d
40 and hence, it is legitimate to ex
change the order of differentiation and integration so that
it becomes the ﬁrst term of Eq. (182). In the second term in
Eq. (187), the term that does not allow the exchange of the
order of integration and differentiation is the term involv
ing the VV operator. Focusing on it more carefully, one has
lim
V
d
÷0
VV
.
_
V
d
G
0
(r; r
/
)J(r
/
) dV
/
= lim
V
d
÷0
V
_
V
d
G
0
(r; r
/
)V
/
.
J(r
/
) dV
/
_
÷
_
S
d
^ nn
/
.
J(r
/
)G
0
(r; r
/
) dS
/
_
(188)
To arrive at this, one has made use of the fact that
VG
0
(r; r
/
) = ÷V
/
G
0
(r ÷r
/
), and
[V
/
G
0
(r; r
/
)]
.
J(r
/
) =V
/
.
[G
0
(r; r
/
)J(r
/
)] ÷G
0
(r; r
/
)V
/
.
J(r
/
).
Using Gauss’ theorem on the term involving
V
/
.
[G
0
(r; r
/
)J(r
/
)] ﬁnally gives rise to Eq. (188). The ﬁrst
integral on the righthand side of Eq. (188) vanishes since
if J(r
/
) is regular, V
/
.
J(r
/
) =r(r
/
)=jo is also regular and
the integral is ﬁnally proportional to V
d
. In the second in
tegral, S
d
is the surface bounding V
d
. Hence, ^ nn
/
.
J(r
/
) is
the surface charge on S
d
due to the sudden truncation of
J(r
/
) within the volume V
d
. This integral gives the poten
tial observed within V
d
due to this surface charge, and it is
non zero even when V
d
÷0. The gradient (outside the
brackets) in turn yields the ﬁeld generated by this surface
charge. In other words, surface charges of opposite polar
ities on the wall of an inﬁnitesimally small volume always
generate a ﬁnite ﬁeld within the small volume. This fact is
also intimately related to the scale invariant nature of the
Laplace equation which is Maxwell’s equations at low fre
quency.
BIBLIOGRAPHY
1. J. A. Stratton, Electromagnetic Theory, McGrawHill,
New York, 1941.
2. P. M. Morse and H. Feshbach, Methods of Theoretical Physics,
McGrawHill, New York, 1953.
3. C. T. Tai, Dyadic Green Functions in Electromagnetic Theory,
2nd ed., IEEE Press, New York, 1994.
4. R. E. Collin, Field Theory of Guided Waves, 2nd ed., IEEE
Press, New York, 1991.
5. J. A. Kong, Electromagnetic Wave Theory, 2nd ed., Wiley,
New York, 1990.
6. W. C. Chew, Waves and Fields in Inhomogeneous Media,
IEEE Press, New York, 1995.
7. I. M. Gel’fand and G. E. Shilov, Generalized Functions,
Academic Press, New York, 1964.
8. J. D. Jackson, Classical Electrodynamics, 2nd ed., Wiley,
New York, 1975.
9. E. T. Copson, Theory of Functions of a Complex Variable,
Oxford Univ. Press, London, 1948.
10. J. M. Jin, V. V. Liepa, and C. T. Tai, Avolumesurface integral
equation for electromagnetic scattering by inhomogeneous
cylinders, J. Electromagn. Waves Appl. 2:573–588 (1988).
11. C. T. Tai, On the eigenfunction expansion of dyadic Green’s
functions, Proc. IEEE 61:480–481 (1973).
12. C. T. Tai, Dyadic Green’s functions for a rectangular wave
guide ﬁlled with tow dielectrics, J. Electromagn. Waves Appl.
2:245–253 (1988).
13. M. Kisliuk, The dyadic Green’s functions for cylindrical wave
guides and cavities IEEE Trans. Microwave Theory Tech.
28:894–898 (1980).
14. V. Daniel, New expressions of dyadic Green’s function in uni
form waveguide with perfectly conducting walls, IEEE Trans.
Anten. Propag. 30:487–499 (1982).
15. C. T. Tai, Dyadic Green’s functions for a coaxial line, IEEE
Trans. Anten. Propag. 31:355–358 (1983).
16. C. T. Tai and P. Rozenfeld, Different representations of dyadic
Green’s functions for a rectangular cavity, IEEE Trans.
Microwave Theory Tech. 24:597–601 (1976).
17. S. Zhang and J. M. Jin, Derivation of dyadic Green’s functions
for cylindrical cavities by image method, Acta Electronica
Sinica, 12(5):21–26 (1984).
18. R. E. Collin, Dyadic Green’s function expansion in spherical
coordinates, Electromagnetics, 6:183–207 (1986).
19. J. K. Lee and J. A. Kong, Dyadic Green’s functions for layered
anisotropic medium, Electromagnetics, 3:111–130 (1983).
20. J. S. Bagby and D. P. Nyquist, Dyadic Green’s functions for
integrated electronic and optical circuits, IEEE Trans. Micro
wave, Theory Tech. 35:206–210 (1987).
21. L. Vegni, R. Cicchetti, and P. Capcec, Spectral dyadic Green’s
function formulation for planer integrated structures, IEEE
Trans. Anten. Propag. 36:1057–1065 (1988).
22. P. Bernardi and R. Cicchetti, Dyadic Green’s functions for
conductorbacked layered structures excited by arbitrary tri
dimensional sources, IEEE Trans. Microwave Theory Tech.
42:1474–1483 (1994).
23. Z. Xiang and Y. Lu, Electromagnetic dyadic Green’s function
in cylindrically multilayered media, IEEE Trans. Microwave
Theory Tech. 44:614–621 (1996).
24. L. W. Li et al., Electromagnetic dyadic Green’s function in
spherically multilayered media, IEEE Trans. Microwave The
ory Tech. 42:2302–2309 (1994).
25. C. T. Tai, The dyadic Green’s function for a moving
isotropic medium, IEEE Trans. Anten. Propag. 13:322–323
(1965).
26. C. Stubenrauch and C. T. Tai, Dyadic Green’s function for cy
lindrical waveguide with moving medium, Appl. Sci. 25:
281–289 (1971).
27. C. T. Tai, Generalized Vector and Dyadic Analysis, 2nd ed.,
IEEE Press, New York, 1997.
28. J. Van Bladel, Some remarks on Green’s dyadic for inﬁnite
space, IEEE Trans. Anten. Propag. 9:563–566 (1961).
29. K. M. Chen, A simple physical picture of tensor Green’s func
tion in source region, Proc. IEEE, 65:1202–1204 (1977).
30. A. W. Johnson, A. Q. Howard, and D. G. Dudley, On the irro
tational component of the electric Green’s dyadic, Radio Sci.
14:961–967 (1979).
31. A. D. Yaghjian, Electric dyadic Green’s functions in the source
region, Proc. IEEE. 68:248–263 (1980).
32. S. W. Lee et al., Singularity in Green’s function and its nu
merical evaluation, IEEE Trans. Anten. Propag. 22:311–317
(1980).
33. W. C. Chew, Some observations on the spatial and eigenfunc
tion representations of dyadic Green’s function, IEEE Trans.
Anten. Propag. 37:1322–1327 (1989).
1832 GREEN’S FUNCTION METHODS
Next Page
This action might not be possible to undo. Are you sure you want to continue?