This action might not be possible to undo. Are you sure you want to continue?
*
AYDIN AYKAN
Tutzing, Germany
A timevarying magnetic ﬁeld at a deﬁned area S can be
determined with a calibrated circular loop connected to
the input of an appropriate measuring receiver (Fig. 5).
There may be a passive or an active network between the
loop and the output port. The measuring loop can also
include a shielding over its circumference against any
perturbation of strong and unwanted electric ﬁelds.
The shielding must be interrupted at a point on the loop
circumference.
Generally in the far ﬁeld the streamlines of magnetic
ﬂux are uniform, but at near ﬁeld, in the vicinity of the
generator of a magnetic ﬁeld, they depend on the source
and its periphery. Figure 4 shows the streamlines of the
electromagnetic vectors generated by the transmitting
loop L1. In the near ﬁeld, the spatial distribution of the
magnetic ﬂux B=m
0
H over the measuring loop area is not
known. Only the normal components of the magnetic ﬂux,
averaged over the closedloop area, can induce a current
through the loop conductor. The measuring loop must
have a calibration (conversion) factor or set of factors
that, at each frequency, expresses the relationship be
tween the ﬁeld strength impinging on the loop and indi
cation of the measuring receiver. The calibration of a
measuring loop requires the generation of a welldeﬁned
standard magnetic ﬁeld on its effective receiving surface.
Such a magnetic ﬁeld is generated by a circular transmit
ting loop when a deﬁned rootmeansquare (RMS) current
is passed through its conductor. The unit of the generated
or measured magnetic ﬁeld strength H
av
is A/m (amperes
per meter) and therefore is also an RMS value. The sub
script ‘‘av’’ strictly indicates the average value of the spa
tial distribution, not the average over a period of T of a
periodic function. This statement is important for near
ﬁeld calibration and measuring purposes. For farﬁeld
measurements the result indicates the RMS value of the
magnitude of the uniform ﬁeld. The traceability of the
calibration must be established for the calibration process,
through linking the assigned value of any components to
the International System of Units (SI). In the following we
discuss the requirements for the nearzone calibration of a
measuring loop.
1. CALCULATION OF STANDARD MAGNETIC FIELDS
To generate a standard magnetic ﬁeld, a transmitting loop
L1 is positioned coaxial and planeparallel at a separation
distance d from the loop L2 to be calibrated, as in Fig. 1.
The analytical formula for the calculation of the average
magnetic ﬁeld strength H
av
in A/m generated by a circular
ﬁlamentary loop at an axial distance d including the re
tardation due to the ﬁnite propagation time was obtained
earlier by Greene [1]. The average value of ﬁeld strength
H
av
was derived from the retarded vector potential A
j
as a
tangential component on the point P of the periphery of
loop L2:
H
av
=
Ir
1
pr
2
_
p
0
e
÷jbR(j)
R(j)
cos(j) dj (1a)
R(j) =
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
d
2
÷r
2
1
÷r
2
2
÷2r
1
r
2
cos(j)
_
(1b)
In these equations for thin circular loops, I is the trans
mitting loop RMS current in amperes, d is the distance
between the planes of the two coaxial loop antennas in
meters, r
1
and r
2
are ﬁlamentary loop radii of transmitting
and receiving loops in meters, respectively, b is wave
length constant, b =2p/l, and l is wavelength in meters.
Equations (1) give the exact results for the separation
distances even from d=0. For d=0 the radii of the loops
must be r
1
ar
2
, otherwise the integral gives a singularity
for j=p, because for r
1
=r
2
the root in Eq. (1b) being zero.
The use of any approximate formula (Eq. 25 in Ref. 1
and Eqs. (2) in Ref. 2) is not suitable, because it imposes
restrictions on the range applicability for the approximate
equations. Using the expressions of maximum magnetic
ﬁeld H
max
would also not be suitable for purposes of near
ﬁeld calibration purpose (see Fig. 2 in Ref. 2).
Generally the Eqs. (1) can be determined by numerical
integration. To this end we separate the real and
imaginary parts of the integrand using Euler’s formula
C
B = × A
ç
A
ç
H
av
d
x
A
L1
H
S
1
G
0
L2
P
S
2
R (ç)
z
r
1
ds
1
T
I
y
E
r
2
I
2
+ç
∆
Figure 1. Conﬁguration of two circular loops.
*This article is based on ‘‘Calibration of Circular Loop Antennas,’’
by Aydin Aykan, which appeared in IEEE Transactions on
Instrumentation and Measurement, Vol. 7, No. 2, r 1998 IEEE.
560
e
÷jj
=cos(j) ÷j sin(j) and rewrite Eq. (1a) as
H
av
=
Ir
1
pr
2
(F ÷jG) (2a)
where
F =
_
p
0
cos(bR(j))
R(j)
cos(j) dj (2b)
G=
_
p
0
sin(bR(j))
R(j)
cos(j) dj (2c)
and the magnitude of H
av
is then obtained as
H
av
[ [ =
Ir
1
pr
2
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
F
2
÷G
2
_
(2d)
It is possible to evaluate the integrals in Eqs. (2) numer
ically with an appropriate mathematics software on a per
sonal computer. Some mathematics software can directly
calculate the complex integral of Eqs. (1). Hence, some of
the following equations are written in complex form for
convenience.
2. ELECTRICAL PROPERTIES OF CIRCULAR LOOPS
2.1. Current Distribution around a Loop
The current distribution around the transmitting loop is
not constant in amplitude and in phase. A standing wave
of current exists on the circumference of the loop. This
current distribution along the loop circumference is dis
cussed by Greene [1, pp. 323–324]. He has assumed the
loop circumference 2pr
1
to be electrically smaller than the
wavelength l and the loop current to be constant in phase
around the loop and the loop proper to be sufﬁciently loss
free. The singleturn thin loop was considered as a circular
balanced transmission line fed at points A and D and
shortcircuited at points E and F (Fig. 2).
In an actual calibration setup the loop current I
1
is speciﬁed at the terminals A and D. The average
current was given as a function of input current I
1
of
the loop [2]:
I
av
=I
1
tan(bpr
1
)
bpr
1
(3)
The fraction of I
av
/I
1
from Eq. (3) expressed in decibels
gives the conditions for determining of the highest fre
quency f and the radius of the loop r
1
. The deviation of this
fractions is plotted in Fig. 3.
The current I in Eq. (1a) must be substituted with I
av
from Eq. (3). Since Eq. (3) is an approximate expression, it
is recommended to keep the radius of the transmitting
loop small enough for the highest frequency of calibration
to minimize the errors. For the dimensioning of the radius
of the receiving loop, these conditions are not very impor
tant, because the receiving loop is calibrated with an ac
curately deﬁned standard magnetic ﬁeld, but the
resonance of the loop at higher frequencies must be taken
into account.
2.2. Circular Loops with Finite Conductor Radii
A measuring loop can be constructed with one or more
windings. The form of the loop is chosen as a circle because
of the simplicity of the theoretical calculation and calibra
tion. The loop conductor has a ﬁnite radius. At high fre
quencies the loop current ﬂows on the conductor surface
and shows the same proximity effect as two parallel, inﬁ
nitely long cylindrical conductors. Figure 4 shows the
cross section of two loops in intentionally exaggerated di
mensions. The streamlines of the electric ﬁeld are ortho
gonal to the conductor surface of the transmitting loop L1
and they intersect at points A and A
/
. The total conductor
current is assumed to ﬂow through a ﬁctive thin ﬁlamen
tary loop with the radius a
1
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
r
2
1
÷c
2
1
_
, where
a
1
=OA=QP=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
OQ
2
÷QP
2
_
. The streamlines of the mag
netic ﬁeld are orthogonal to the streamlines of electric
ﬁeld. The receiving loop L2 with the ﬁnite conductor
H
av
l
D
Q
Q
D
A
V
L
V
o
V
L
I
av
I
av
Z
L
V
o
V
2
= 0
Z
2
= 0
Z
L
A
I
1
I
1
I
1
I
1
I
1
I
2
= I
max
I
2
= I
max
I
2
= I
max
r
1
x
I
F
0
x
E
E
F
I
1
= ¬
.
r
1
I
x
I
x
I
x
π
.
r
1
Figure 2. Current distribution on a circular loop.
1.5
1
0.5
d
B
0
−0.5
1 2 5 10 20 50 100
MHz
Figure 3. Deviation of I
av
/I
1
for a loop radius 0.1 m as 20log(I
av
/I)
in decibels versus frequency.
CALIBRATION OF A CIRCULAR LOOP ANTENNA 561
radius c
2
can encircle a part of magnetic ﬁeld with its ef
fective circular radius b
2
=r
2
÷c
2
.
The sum of the normal component of vectors H acting
on the effective receive area S
2
=pb
2
2
induces a current in
the conductor of the receiving loop L2. This current ﬂows
through the ﬁlamentary loop with the radius a
2
. The
average magnetic ﬁeld vector H
av
is deﬁned as the inte
gral of vectors H
n
over effective receiving area S
2
, divided
by S
2
. The magnetic streamlines, which ﬂow through
the conductor and outside of loop L2, cannot induce a
current through the conductor along the ﬁlamentary loop
Ar, Ar
/
of L2. The equivalent ﬁlamentary loop radii a
1
, a
2
and effective circular surface radii b
1
, b
2
can be seen directly
from Fig. 4.
The equivalent thin current ﬁlament radius a
1
of the
transmitting loop L1:
a
1
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
r
2
1
÷c
2
1
_
(4a)
The equivalent thin current ﬁlament radius a
2
of the re
ceiving loop L2 is
a
2
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
r
2
2
÷c
2
2
_
(4b)
The radius b
1
of the effective transmitting circular area of
the transmitting loop L1 is
b
1
=r
1
÷c
1
(4c)
The radius b
2
of the effective receiving circular area of the
receiving loop L2 is
b
2
=r
2
÷c
2
(4d)
2.3. Impedance of a Circular Loop
The impedance of a loop can be deﬁned at chosen termi
nals Q, D as Z=V/I
1
(Fig. 2). Using Maxwell’s equation
with Faraday’s lawcurl(E) = ÷joF
m
, we can write the line
integrals of the electric intensity E along the loop conduc
tor through its cross section, along the path joining
points D,Q and the load impedance Z
L
between the
terminals Q, A:
_
(AEFD)
E
s
ds ÷
_
(DQ)
E
s
ds ÷
_
(QA)
E
s
ds = ÷joF
m
(5a)
Here F
m
is the magnetic ﬂux. The impressed emf (elec
tromotive force) V acting along the path joining points D
and Q is equal and opposite to the second term of Eq. (5a):
V = ÷
_
(DQ)
E
s
ds (5b)
The impedance of the loop at the terminals D, Q can be
written from Eqs. (5) dividing with I
1
as
Z=
V
I
1
=
_
(AEFD)
E
s
ds
I
1
÷
_
(QA)
E
s
ds
I
1
÷
joF
m
I
1
=Z
i
÷Z
L
÷Z
e
(6)
where Z
i
indicates the internal impedance of the loop con
ductor. Because of the skin effect, the internal impedance
at high frequencies is not resistive. For the calculation of
the Z
i
, we refer to Schelkunoff, [3, p. 263] and Ramo et al.
[4, p. 185]. Z
L
is a known load or a source impedance on
Fig. 2. Z
e
is the external impedance of the loop:
Z
e
=jo
F
m
I
1
=jo
m
0
H
av
S
I
1
(7a)
We can consider that the loop consists of two coaxial and
coplanar ﬁlamentary loops (i.e., separation distance d=0).
The radii a
1
and b
1
are deﬁned in Eqs. (4). The average
current I
av
ﬂows through the ﬁlamentary loop with the
radius a
1
and generates an average magnetic ﬁeld
strength H
av
on the effective circular surface S
1
=pb
2
1
of
the ﬁlamentary loop with the radius b
1
. From Eqs. (1) and
(3) we can rewrite Eq. (7a), for the loop L1:
Z
e
=j
tan(bpa
1
)
bpa
1
m
0
oa
1
b
1
_
p
0
e
÷jbR
0
(j)
R
0
(j)
cos(j)dj (7b)
R
0
(j) =
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
a
2
1
÷b
2
1
÷2a
1
b
1
cos(j)
_
(7c)
The real and imaginary parts of Z
e
are the radiation
resistance and the external inductance of loops,
P
c
1
r
1
a
1
b
1
L1
B
Q
Qr
H
c
2
L2
b
2
a
2
r
2
h
e
T
Br Ar
H
H
n
Q'
O B'
Br'
T'
Or
H
av
Qr'
Ar'
A
A'
Figure 4. Filamentary loops of two loops with ﬁnite conductor
radii and orthogonal streamlines of the electromagnetic vectors,
produced from transmitting loop L1.
562 CALIBRATION OF A CIRCULAR LOOP ANTENNA
respectively:
Re(Z
e
) =
tan(bpa
1
)
bpa
1
¸
¸
¸
¸
¸
¸
¸
¸
m
0
oa
1
b
1
_
p
0
sin(bR
0
(j))
R
0
(j)
cos(j)dj
(7d)
Im(Z
e
) =
tan(bpa
1
)
bpa
1
m
0
oa
1
b
1
_
p
0
cos(bR
0
(j))
R
0
(j)
cos(j)dj (7e)
From Eq. (7e) we obtain the external selfinductance:
L
e
=
tan(bpa
1
)
bpa
1
m
0
a
1
b
1
_
p
0
cos(bR
0
(j))
R
0
(j)
cos(j)dj (7f )
Equations (7) include the effect of current distribution on
the loop with ﬁnite conductor radii.
2.4. Mutual Impedance between Two Circular Loops
The mutual impedance Z
12
between two loops is deﬁned as
Z
12
=
V
2
I
1
=
Z
2
I
2
I
1
(8)
The impedance of Z
2
in Eq. (8) can be deﬁned as in Eqs. (6):
Z
2
=
V
2
I
2
=Z
2i
÷Z
L
÷Z
2e
(9)
here Z
2i
is the internal impedance, Z
L
is the load impedance,
and Z
2e
is the external impedance of the second loop L2.
The current ratio I
2
to I
1
in Eq. (8) can be calculated
from Eqs. (1), (3), and (4). The current I
1
of the transmit
loop with separation distance d
I
1
=
H
av
pb
2
tan(bpa
1
)
bpa
1
a
1
_
p
0
e
÷jbR
d
(j)
R
d
(j)
cos(j)dj
(10a)
R
d
(j) =
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
d
2
÷a
2
1
÷b
2
2
÷2a
1
b
2
cos(j)
_
(10b)
and the current I
2
of the receive loop for the same H
av
(here d=0) is
I
2
=
H
av
pb
2
tan(bpa
2
)
bpa
2
a
2
_
p
0
e
÷jbR
0
(j)
R
0
(j)
cos(j)dj
(10c)
R
0
(j) =
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
a
2
2
÷b
2
2
÷2a
2
b
2
cos(j)
_
(10d)
The general mutual impedance between two loops from
Eqs. (8) and (9) is
Z
12
=(Z
2i
÷Z
L
÷Z
2e
)
I
2
I
1
=Z
12i
÷Z
12L
÷Z
12e
(11a)
here Z
12i
is the mutual internal impedance, Z
12L
denotes
the mutual load impedance, and Z
12e
is the external mu
tual impedance.
Arranging Eq. (7b) for Z
2e
and the current ratio I
2
/I
1
from Eqs. (10), we obtain the external mutual impedance:
Z
12e
=j
tan(bpa
1
)
bpa
1
m
0
oa
1
b
2
_
p
0
e
÷jbR
d
(j)
R
d
(j)
cos(j)dj (11b)
The real part of Z
12e
may be described as mutual radiation
resistance between two loops.
The imaginary part of Z
12e
divided by o gives the mu
tual inductance
M
12e
=
tan(bpa
1
)
bpa
1
m
0
a
1
b
2
_
p
0
cos(bR
d
(j))
R
d
(j)
cos(j) dj (11c)
Equations (11b) and (11c) include the effect of current dis
tribution on the loop with ﬁnite conductor radii.
3. DETERMINATION OF THE ANTENNA FACTOR
The antenna factor K
H
is deﬁned as a proportionality con
stant with necessary conversion of units. K
H
is the ratio of
the average magnetic ﬁeld strength H
av
bounded by the
loop to the measured output voltage V
L
on the input im
pedance R
i
of the measuring receiver.
K
H
=
H
av
V
L
in (A=m)
.
V
÷1
(12a)
Equation (12a) can also be expressed logarithmically:
k
H
=20 log(K
H
) in dB(A=m)
.
V
÷1
(12b)
For evaluation of the antenna factor there are two meth
ods. The ﬁrst is by calculation of the loop impedances, and
the second is with the welldeﬁned standard magnetic
ﬁeld calibration.
3.1. Determination of the Antenna Factor by Computing
from the Loop Impedances
If a measurement loop, (e.g., L2), has a simple geometric
shape and a simple connection to a voltage measuring de
vice with a known input impedance R
i
, we can determine
the antenna factor by calculation. In the case of the un
loaded loop from Fig. 2, the opencircuit voltage is
V
0
=jom
0
H
av
S
2
(13a)
For the case of the loaded loop the current is
I =
V
0
Z
=
V
0
R
L
÷Z
i
÷Z
e
(13b)
The antenna factor from Eq. (12a) can be written with V
L
=Z
L
I and Eqs. (13) and (1) as
K
H
=
1
jom
0
S
2
1 ÷
Z
e
R
L
÷
Z
i
R
L
_ _ ¸
¸
¸
¸
¸
¸
¸
¸
in (A=m)
.
V
÷1
(14)
The effective loop area is S
2
=pb
2
2
. The external loop im
pedance Z
e
can be calculated with Eqs. (7). The internal
CALIBRATION OF A CIRCULAR LOOP ANTENNA 563
impedance Z
i
is in general small in respect to external
impedance Z
e
, and can be neglected. For a more precise
calculation of the internal impedance Z
i
due to the skin
effect, refer to Refs. 3 and 4.
3.2. Standard Magnetic Field Method
In the calibration setup in Fig. 5 we measure the voltages
with standard laboratory measuring instrumentation
with the 50O interface impedances. The device to be cal
ibrated consists at least of a loop and a cable with an out
put connector. The measuring loop can also include a
passive or active network between the terminals C, D
and a coaxial shield on the circular loop conductor against
unwanted electric ﬁelds, depending on its development
and construction. The impedance Z
CD
and the voltage V
CD
at the terminals C,D is not accurately measurable. The
behavior of the attenuation and/or the gain between the
interfaces D,C and F cannot be accurately deﬁned. Such a
complex measuring loop must be calibrated with the stan
dard magnetic ﬁeld method through the calibration setup
in Fig. 5. To prevent the deterioration of the magnetic
ﬁeld, produced by L1, we must place the attenuators suf
ﬁciently far from the transmitting loop using a twisted
pair balanced line. The attenuators (e.g., nominal 10 dB)
must have the calibrated attenuations m for attenuator
1 and n for attenuator 2, and the calibrated interface re
sistances (nominal 50 O). The shielding of the attenuators
must be electrically connected at the points x. The ferrite
pads on the twistedpair line and on the coaxial cable at
tenuate the magnetic ﬁeld scattering from the measuring
transmission lines. The electrical length Le of the twisted
transmission line must be taken into account and the
equation (3) must be redeﬁned for the average current I
av
:
I
av
=I
1
tan(b(pa
1
÷Le))
b(pa
1
÷Le)
(15)
The usable highest frequency of the loop L1 decreases
with the additional electrical length of the twistedpair
line. This consideration is used to deﬁne an appropriate
length of the twistedpair transmission line.
The antenna factor in Eqs. (12) can be fully deﬁned for
each frequency through the measurement of the trans
mitting loop current I
1
at the interface M
/
and voltage V
L
at the interface F:
K
H
=
I
1
tan(b(pa
1
÷Le))
pa
1
÷Le
a
1
pb
2
_
p
0
e
÷jbR
d
(j)
R
d
(j)
cos(j)dj
¸
¸
¸
¸
¸
¸
¸
¸
V
L
(16a)
Here r
1
=a
1
, r
2
=b
2
, Le is the electrical length of the twist
edpair transmission line and R
d
deﬁned with Eq. (10b).
Attenuator 1 attenuates the current I
1
with the ratio m
and hence the current through the interface M
/
becomes
we m
.
I
1
. The transmitting loop current is I
1
=V
2
/m· R
2
.
Consequently, the current ﬂowing through the interface
N
/
is I
1
/(n
.
m). With the measuring receiver we can mea
sure ﬁrst the voltage V
2
at the interface M
/
to obtain the
transmitting loop current I
1
, which produces the magnetic
ﬁeld H
av
in the receiving loop L2. We then measure the
voltage V
L
at the interface F, which is produced by the
same magnetic ﬁeld H
av
. With setting a =V
2
/V
L
, we can
write Eq. (16a) as
K
H
=
a
m
tan(b(pa
1
÷Le))
R
2
(bpa
1
÷Le)
a
1
pb
2
_
p
0
e
÷jbR
d
(j)
R
d
(j)
cos(j) dj
¸
¸
¸
¸
¸
¸
¸
¸
(16b)
Equation (16b) is expressed in SI units [(A/m)V
÷1
] and
can also be expressed logarithmically as
k
H
=20 log(K
H
) in dB[(A=m)
.
V
÷1
] (16c)
V
L
R
i
V
2
R
2
R
1
V
0
V
1
V
CD
r
2
r
1
I
1
I
2
Z
CD
H
av
C
D
d
Transmitter loop L1
Q
Generator
Attennuator
Network
Measuring loop L2
M'
m·I
1
I
1
/n·m
1. Step: Receiver
2. Step: Terminator
F
1. Step: Terminator
2. Step: Receiver
m
n
1
2
N
A
B Le
M
N'
x
x
Figure 5. Calibration setup for circular loop antennas.
564 CALIBRATION OF A CIRCULAR LOOP ANTENNA
The ratio of the measured voltages is an attenuation and
with an appropriate measuring setup the calibration is
realized as an attenuation measurement. A network analy
zer is generally used for this purpose instead of discrete
measurements at each individual frequency with a signal
generator and a measuring receiver. A network analyzer
can normalize the frequency characteristic of the transmit
loop current I
1
and gives a quick overview of the measured
attenuation for the frequency band under consideration.
Equation (16b) reduces the calibration process of the
loop to an accurate measurement of attenuation a for each
frequency. The other terms of Eq. (16a) can be calculated
depending on the geometric conﬁguration of the calibra
tion setup at the working frequency band of the measuring
loop. The calibration uncertainties are also calculable with
the given expressions. The uncertainty of the separation
distance d between two loops must be taken into consid
eration as well. At a separation distance doa
1
, the change
in the magnetic ﬁeld is high (see Fig. 2 in Ref. 2).
For a calibration setup the separation distance d can be
deﬁned as small as possible. However, the effect of the
mutual impedance must be taken into account in the cal
ibration process, and a condition for definition of the sep
aration distance d must be given (Fig. 5). If the second loop
is opencircuited, that is, if the current I
2
=0, the current
I
1
is deﬁned only from the impedances of the transmitting
loop L1. In the case of a shortcircuited second loop L2, I
2
is
maximum and the value of I
1
will change depending on the
supply circuit and loading of the transmitting loop. A cur
rent ratio q between these two cases can be deﬁned as the
condition of the separation distance d between the two loops.
It is assumed that the generator voltage V
0
is constant.
The measuring loop L2 is terminated by Z
L
. For Z
L
=0 and
V
CD
=0, one obtains the current I
1
in the transmitting
loop as
I
1(Z
L
=0)
=
V
0
R
1
÷R
2
÷Z
AB
÷
Z
2
12
Z
CD
(17a)
and for Z
L
=N, that is, I
2
=0
I
1(Z
L
=o)
=
V
0
R
1
÷R
2
÷Z
AB
(17b)
The ratio of Eq. (17a) to Eq. (17b) is
q ¬
I
1(Z
L
=0)
I
1(Z
L
=o)
¸
¸
¸
¸
¸
¸
¸
¸
=
R
1
÷R
2
÷Z
AB
R
1
÷R
2
÷Z
AB
1 ÷
Z
2
12
Z
AB
Z
CD
_ _
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
(18a)
here with the coupling factor k=Z
12
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
Z
AB
Z
CD
_
between
two loops:
q =
R
1
÷R
2
÷Z
AB
R
1
÷R
2
÷Z
AB
(1 ÷k
2
)
¸
¸
¸
¸
¸
¸
¸
¸
(18b)
where R
1
=R
2
=50 O and Z
AB
, Z
CD
, and Z
12
can be calcu
lated from Eqs. (7) and (11). For greater accuracy one must
try to keep the ratio q close to unity (e.g., q =1.001).
The inﬂuence of the loading of the second loop on the
transmitting loop can also be found experimentally. The
change of voltage V
2
at R
2
in Fig. 5 must be considerably
small (e.g., o0.05 dB) when the measuring loop is short
circuited at the chosen separation distance d.
4. CONCLUSION
All equations in this article are written in their original
exact form without approximations, which usually restrict
their range of validity in applications regarding frequency
range, distance, and other quantities. Also, no simpliﬁca
tions of the equations are made. All equations deliver the
results directly in units of SI with an appropriate math
ematical software on a personal computer. The electrical
properties of the loops are calculable without approxima
tion. The necessary conditions are given for the choice of
dimensions of the measuring and transmitting circular
loops and the separation distance d.
For calibration of circular loop antennas, the standard
magnetic ﬁeld method is recommended with the calibration
setup in Fig. 5 and also for determination of the antenna
factor K
H
and k
H
in Eqs. (16b) and (16c). The calibration
process is based on the measurement of attenuation at each
frequency on the same impedance level of 50O, using the
standard laboratory equipment. The measurement can also
be accelerated by using a network analyzer.
BIBLIOGRAPHY
1. F. M. Greene, The nearzone magnetic ﬁeld of a small circular
loop antenna, J. Rese. Natl. Bur. Stand. Eng. Instrum. 71C(4):
319–326 (Oct.–Dec., 1967).
2. A. Aykan, Calibration of circular loop antennas, IEEE Trans.
Instrum. Meas. 47(2): 446–452 (April 1998).
3. S. A. Schelkunoff, Electromagnetic Waves, Van Nostrand, New
York, 1943.
4. S. Ramo, J. R. Whinnery, and T. van Duser, Fields and Waves
in Communication Electronics, 3rd ed., Wiley, New York, 1994.
CAPACITANCE EXTRACTION
WENJIAN YU
ZEYI WANG
Tsinghua University
Beijing, China
1. INTRODUCTION
Since the early 1950s, microwave circuits have evolved
from discrete circuits to planar integrated circuits, then to
multilayered and threedimensional integrated circuits.
With the increased circuit density, the multiconductor line
in multilayered dielectric media has become the major
form of the transmission line or interconnect. Multilay
ered routing reduces the area as well as the volume of the
CAPACITANCE EXTRACTION 565
circuit. However, as a result, electromagnetic coupling
among conductors greatly inﬂuences circuit performance.
In some microwave integrated circuits, this coupling effect
is utilized to construct compact circuit components. But
under most circumstances, it is regarded as a parasitic
effect that must be modeled accurately for veriﬁcation of
the circuit’s validity and performance.
In the related ﬁeld of verylargescale integration
(VLSI) circuits, electromagnetic coupling among intercon
nects is also becoming increasingly important. With the
introduction of deepsubmicrometer (DSM) semiconduc
tor technologies, the onchip interconnect wire can no
longer be considered equipotential jointing. The parasitic
effects introduced by the wires display a scaling behavior
that differs from that of active devices such as transistors,
and these effects tend to gain importance as device dimen
sions are reduced and circuit speed is increased. In fact,
they begin to dominate some of the relevant metrics of
digital integrated circuits such as speed, energy consump
tion, and reliability. A typical recursive design ﬂowchart of
a stateoftheart integrated circuit (IC) is shown in Fig. 1,
where a postlayout step termed parasitic extraction pre
cedes gatelevel simulation. The task of parasitic extrac
tion is to model the electromagnetic effects of the wire with
parasitic components of capacitance, resistance, and in
ductance, so that a more accurate circuit simulation can
be performed.
With the increase in working frequency and develop
ment of silicon technologies, the discrepancy between the
microwave IC and the common VLSI circuit becomes mar
ginal. Therefore, the electromagnetic modeling and accu
rate extraction of the interconnect parasitics have become
a subject of advanced research in both ﬁelds to date.
Among the three parasitic parameters, capacitance has
attracted the most attention because it greatly inﬂuences
time delay, power consumption, and the signal integrity
and its calculation becomes complicated under DSM tech
nologies.
In the following sections, the fundamental theory and
contemporary methodology and algorithms of capacitance
extraction will be discussed.
2. PROBLEM FORMULATION
As is well known, the capacitor is a commonly used com
ponent in electric or electronic equipment. It is usually
composed of two conductors insulated from each other.
When charged, the two surfaces of the conductor facing
each other carry equal and opposite charges Q and ÷Q,
respectively (see Fig. 2). The electric potential difference
between the two conductors f
1
–f
2
is called the voltage of
the capacitor and is always denoted by V. Experiments
and theoretical analyses show that, for a capacitor, Q is
always proportional to V and thus the ratio Q/V is a con
stant determined by the structure of the capacitor. This
ratio is called the capacitance of the capacitor and is de
noted by C: C=(Q/V).
In the International System of Units (SI), the unit of
capacitance is the faraday (F). It expresses the capaci
tance of a capacitor that has one coulomb on one of its
poles when the potential difference is 1 V. Other commonly
used units of capacitance are mF (10
÷6
F), pF (10
÷12
F),
and f F (10
÷15
F).
The capacitance of some simple capacitor can be calcu
lated easily. For example, for the parallelplate capacitor
shown in Fig. 2, we have
C=
e
0
e
r
S
d
(1)
where e
0
is the dielectric constant of free space and in SI, is
expressed as
e
0
=
1
4p 910
9
=8:85 10
÷12
C
2
=N
.
m
2
where e
r
is the relative permittivity of the insulating ma
terial, S is the area of the plate, and d is the distance be
tween two parallel plates.
Specific capacitors widely used in the design of micro
wave circuits include the interdigital capacitor and the
metal–insulator–metal (MIM) capacitor. Figure 3 shows
the physical layout of an interdigital capacitor with nine
ﬁngers, and Fig. 4 shows the crosssectional view of an
MIM capacitor with the GaAs process. The interdigital
capacitor works with the electrostatic coupling between
the intercrossed ﬁngers, and has a very high Q value. So,
it is widely used in the highfrequency microwave circuits.
Function Spec.
Logic Synth.
Gatelevel Net.
RTL
Layout
Floorplanning
Place & Route
Frontend
Backend
Behavior Simul.
Gatelevel Simul.
Stat. Wire Model
Parasitic Extraction
Figure 1. A typical ﬂowchart of IC design.
+Q
−Q
[
2
[
1
d
S
c
r
Figure 2. A parallelplate capacitor.
566 CAPACITANCE EXTRACTION
The MIM capacitor has simple geometry and is easily fab
ricated, and its capacitance is controlled by the dimen
sions of the polar planes. Since the interdigital capacitor
and the MIM capacitor are widely used, calculation of the
parameters of their structures within given the working
frequency and corresponding capacitor value becomes an
important issue for both design and optimization. This can
be regarded as the reverse procedure of capacitance ex
traction. For further discussion of this issue, please refer
to the literature [39,40].
Actually, the capacitor has more generalized forms
than that described above, which consists of two insulat
ed conductors. The capacitance of a single conductor (con
ductor 1) is deﬁned as if another conductor (conductor 2)
were located at an inﬁnite distance away to form a joint
capacitor (conductors 1 ÷2). For example, the capacitance
of an isolated conductor sphere with radius of R can be
calculated as C=4pe
0
R.
Many conductor interconnect wires are involved in the
microwave IC and the common VLSI circuit, and they are
insulated by some dielectric such as oxide SiO
2
. The ca
pacitance between any two wires reﬂects the electrostatic
coupling effect between these wires, and calculating these
capacitances with high accuracy is very important for
analysis of the circuit’s performance.
For an Nconductor system, such as the interconnect
wires in an IC, an NN capacitance matrix [C
ij
]
NN
is
deﬁned by
Q
i
=
N
j =1
C
ij
U
j
; i =1; 2; . . . N; (2)
where C
ij
(iaj) is the coupling capacitance between con
ductors i and j, and C
ii
is the selfcapacitance or total ca
pacitance of conductor i. Q
i
is the induced charge on
conductor i, and U
j
is the electric potential of conductor j
(usually the known bias voltage).
Figure 5 shows a typical crossover wires in the VLSI
system, where the coupling capacitances between any two
conductors need to be calculated.
Accurate modeling of the wire capacitances in a state
oftheart integrated circuit is not a trivial task. It is
further complicated by the fact that the interconnect
structure of contemporary integrated circuits is three
dimensional (see Fig. 5). The capacitance of such a wire
is a function of its shape, environment, distance from
the substrate, and distance to surrounding wires. Gener
ally SiO
2
is the insulating material among interconnect
wires in integrated circuits, although some materials
with lower permittivity, and thus lower capacitance, are
coming into use. The relative permittivity e
r
of several
dielectrics commonly used in integrated circuits is pre
sented in Table 1. It should also be pointed out that e
r
of
air or vacuum is 1.
3. METHODOLOGY AND ALGORITHMS
With the advances in IC technology, the methodology of
capacitance extraction has evolved from onedimensional
(1D), twodimensional (2D), 2.5dimensional (2.5D), to
threedimensional (3D) to meet the required accuracy. In
this section, the 1D and 2D methods are briefly intro
duced. Then, the 2.5D method and the mechanism of the
modern commercial capacitance extraction tools that em
ploy the 3D capacitance extractor are presented. Finally,
we will discuss some details of algorithms of the 3D ﬁeld
solver for capacitance extraction.
3.1. 1D and 2D Methods
From the formula for calculation of parallelplate capaci
tance [Eq. (1)], we can infer that the capacitance is pro
portional to the overlapping area between the conductors
and inversely proportional to their separation distance.
This is very important for capacitance extraction without
a high degree of precision.
SiN SiO
2
GaAs
Figure 4. An MIM capacitor (crosssectional view).
Figure 3. An interdigital capacitor.
Grounded substrate
Neumann boundary
Master conductor, IV
z
15.00
3.00
30
30.00
Figure 5. A structure involving 2 2 crossover interconnect
wires.
CAPACITANCE EXTRACTION 567
Figure 6 shows a typical microstrip structure, where
there is only a single rectangular conductor over a ground
plane. This structure is very different from the above par
allelplate model discussed because of the existence of
the capacitance between the sidewalls of the wire and
the substrate, called the fringing capacitance. To avoid the
timeconsuming numerical modeling of this geometry, an
approximate 1D method can be used as a good engineering
practice. The capacitance is assumed to be the sum of two
components: (1) a parallelplate capacitance determined
by the vertical ﬁeld between a wire of width w and the
ground plane and (2) the fringing capacitance modeled
by a cylindrical wire with radius equal to the conductor
thickness H. So, this simple and practical 1D formula
becomes
C=C
area
÷C
fringe
=
e
.
w
d
÷
2pe
log (d=H)
where w=W÷H/2 is a good approximation for the width
of the parallelplate capacitor (W is the width of the wire),
d is the distance between the ground plane and the bottom
of wire, and e is the permittivity of the insulating material.
With this formula, we obtain the approximate capacitance
per unit length.
In another kind of 1D capacitance extraction, the area
and perimeter parameters of interconnect geometries are
ﬁrst obtained. Then, a ﬁnetuned set of area and perimeter
weights per routing layer can be used to calculate capac
itance values as an inner product [1]:
C=(area amount; perimeter amount)
.
(area weight; perimeter weight)
Such area and perimeter weights can be obtained by pre
characterization of an ‘‘average’’ environment of a wire.
The area can be that from single layer or a combination of
layer overlaps.
Usually, the 1D extraction method works well when the
number of interconnect layers is restricted to only one or
two. However, the current process technology often in
volves many more interconnect layers, and they are also of
high density. So, several capacitance components of a wire
embedded in the multilayered interconnect system may
exist, other than the only capacitive coupling to the
ground plane (see Fig. 7). Each wire is coupled not only
to the grounded substrate but also to the neighboring
wires on the same layer and on adjacent layers. Not all
capacitive components terminate at the grounded sub
strate; actually a large number of them connect to other
wires. These (fringing, lateral, parallel, etc.) capacitors
between wires not only form a source of noise (crosstalk
among signal lines) but also can have a negative impact on
the circuit performance.
To model the capacitance in the multilayered intercon
nect system with higher accuracy, 2D capacitance extrac
tion methodology was developed. In 2D capacitance
extraction, accurate geometry modeling and numerical
techniques are implemented for the cross section of sim
ulated structure (as in Fig. 7). For the 2D region of dielec
trics, the electric ﬁeld equation is solved with numerical
techniques. 2D extraction ignores all three dimensional
details and assumes that the geometries being modeled
are uniform in one dimension, usually the signal propa
gation direction. Therefore, 2D capacitance extraction is
only suitable only for some special cases, such as like the
transmission line.
The details of numerical techniques for solving the
electric ﬁeld will be introduced in Section 3.3, albeit in a
3D manner.
3.2. 2.5D Method and Commercial Capacitance
Extraction Tool
The 2.5D (also called quasi3D) method goes a step further
than 2D extraction. Its main idea is to calculate the ca
pacitance of several cross sections (using the 2D method)
and combine the 2D results into the ﬁnal capacitance
value.
A typical 2.5D capacitance extraction method is also
called the ‘‘(22)D method’’, in which any 3D structure is
swept in two perpendicular directions and by considering
the geometry overlapping, 3D structure can be modeled
more accurately (see Fig. 8).
In Fig. 8, an m
2
wire crosses an m
1
wire. Along direc
tion A, a 2D crosssectional view is shown in the middle.
Along direction B, the other 2D cross section is shown to
Table 1. Relative Permittivity of Several Commonly used Dielectric Materials
Dielectric Material Silicon
Alumina
(Package)
Silicon Nitride
(Si
3
N
4
)
Glassepoxy
(PCB) Silicon Dioxide
Polyimides
(Organic) Aerogels
Relative permittivity (e
r
) 11.7 9.5 7.5 5 3.9 3–4 B1.5
Ground
C
fringe
C
area
Figure 6. A conductor above a ground plane (crosssectional
view).
Fringing
Lateral
Parallel
Substrate
Figure 7. Capacitive coupling between wires in a multilayered
interconnect system.
568 CAPACITANCE EXTRACTION
the right. Solving the two orthogonal strictly 2D problems
numerically, we obtain C
A
=C
1f1
÷C
10
÷C
1f2
, C
B
=C
2f1
÷
C
20
÷C
2f2
(see Fig. 8). Then, C
m1,m2
=C
A
w
1
÷(C
B
÷C
20
)
w
2
, where w
1
, w
2
are the widths of wires m
1
and m
2
,
respectively. However, this method is still not very accu
rate. The error could be more than 10%, especially for
coupling capacitance, which is very important for signal
integrity analysis.
Obviously, true 3D extraction is a straightforward
method to achieve high precision. However, the 3D elec
trostatic Laplace equation must be solved numerically
within a complicated 3D structure. This consumes exten
sive computational effort. 3D capacitance extraction (usu
ally called the ‘‘ﬁeld solver’’) is actually not a trivial
extension of the 2D case. This aspect is discussed further
in Section 3.3.
For the current task of capacitance extraction in mod
ern IC design, using the 3D extraction method directly is
impossible because of its huge expense of memory and
CPU time. To obtain a good tradeoff between accuracy and
efﬁciency, modern capacitance extraction tools utilize spe
cial techniques for the fullchip extraction task, which is
usually divided into three major steps [1]:
1. Technology Precharacterization. Given a descrip
tion of the process cross sections, tens of thousands
of test structures are enumerated and simulated
with 2D and/or 3D ﬁeld solvers. These structures
are of medium dimensions. The resulting data are
collected either to ﬁt some empirical formulas or to
build lookup tables (either type is called a ‘‘pattern
library’’). In Ref. 3, analytical equations are used for
model ﬁtting. A good ﬁt would require fewer simu
lation points. The number of patterns can be re
duced by pattern reduction techniques. Arora et al.
[4] present a pattern compression technique that re
duces the total number of precharacterizaiton pat
terns. With this technology, the capacitance in some
layout pattern can be extrapolated from the capac
itance values in two simpler precharacterization
patterns, without losing much accuracy. Capaci
tance ﬁeld solvers employ different numerical algo
rithms, and they may give different answers for
certain special layout structures depending on the
problem setup and boundary conditions. Therefore,
the precharacterizaiton software should have the
ﬂexibility to incorporate any thirdparty ﬁeld
solvers. This ﬁrst step should be performed only
once per process technology. The challenge in this
area includes the handling of increasingly complex
processing technology, such as lowk dielectric, air
bubble dielectric, nonvertical conductor cross sec
tions, conformal dielectric (see Fig. 9), and shallow
trench isolations.
2. Geometric Parameter Extraction. This is also an in
tegral part of precharacterization. If a geometric
pattern requires 10 parameters to describe, there
is a corresponding precharacterization of 15
10
(B10,000,000) patterns to simulate. This is assum
ing that ﬁve sample points are taken in each of the
10 parameters, resulting in a 10dimensional (10D)
table of the dimensions given above. This is clearly
not feasible. On the other hand, if a geometric pat
tern can be described by very few parameters, then
it is difﬁcult for it to be accurate. In a fullchip sit
uation, the runtime of geometric parameter extrac
tion can be very time/spaceconsuming, with
millions of interconnect polygons to analyze. Time/
spaceefﬁcient geometric processing algorithms are
very important. Habitz and Wemple [5] present a
geometric parameter reduction technique in which
geometric parameters can be dramatically reduced
by taking advantage of the shielding effect. Conduc
tors two layers away from the main conductor of in
terest do not require a precise description. This is
particularly useful for the verydeepsubmicrome
ter geometry, where a very distant conductor mesh
behaves like a large airplane.
3. Calculation of Capacitance from Geometric Param
eters. Here, the geometric parameters are matched
to some entries in the pattern library. Usually a full
chip or fullpath extraction task involves at least
thousands of conductors. The whole structure is
chopped into mediumsize pieces ﬁrst, which are
then calculated with the patternmatching approach
described above. Finally, the capacitance values
must be combined to get the desired result.
A
B
Top view Cross section view A Cross section view B
C
2f 1
C
2f 2
C
2o
C
1f 1 C
1f 2
C
1o
w
2
w
1
m2
m1
Figure 8. 2.5D capacitance for a crossover structure.
Conformal dielectric
Figure 9. A realistic vertical cross section of IC interconnect. We
see that conductors on layers 1–5 are trapezoidal, and there is a
conformal dielectric on top of the top layer metal (passivation).
(SEM photograph courtesy of IBM Corp. rCopyright IBM Corp.
1994, 1996.)
CAPACITANCE EXTRACTION 569
One major source of error is called the pattern mis
match, where extracted geometry parameters do not have
an exact match in the pattern library. At this time, there
are two remedies to perform the capacitance calculations.
One method is to enhance the pattern library by running
ﬁeld solvers at the fullchip extraction time. The other
method is to employ heuristics to synthesize a solution
from closely matched precharacterization patterns. Even
if all the geometric patterns match the library completely,
there could still be discontinuities in the layout pattern
decomposition, which is another source of error. This error
is analyzed in Ref. 6, where the error bound was obtained
by utilizing the ‘‘empty’’ and ‘‘full’’ boundary conditions.
3.3. Algorithms for 3D Field Solver
The 3D method can be used model the actual geometry
accurately, so it behaves with the highest precision. 3D
capacitance extraction becomes increasingly important
under the DSM technology of VLSI circuit, although pres
ently it is used widely only as a librarybuilding tool in the
industry. More recently, much research work has been
devoted to improve the efﬁciency of the 3D extraction
method. Related papers are published on the annually
held conferences (Design Automation Conf., Int. Conf.
ComputerAided Design, etc.) and many academic jour
nals (IEEE Trans. Microwave Theory Tech., IEEE Trans.
Comput. Aided Design, etc.). To date, some 3D extraction
algorithms have been developed to integrate with the
commercial software of some electronic design automa
tion (EDA) companies in the Silicon Valley (in California).
Research on 3D capacitance extraction is still advancing
very rapidly.
In this section, the principles and mainstream tech
niques of the 3D ﬁeld solver are introduced. More cutting
edge techniques are mentioned with related references.
3.3.1. Overview. For a system involving multiple con
ductors (see Fig. 5), with one conductor setting 1V and
others 0 V, the electrostatic equation (called the Laplace
equation) need to be solved with a homogenous dielectric
region [7]:
V
2
u=
@
2
u
@x
2
÷
@
2
u
@y
2
÷
@
2
u
@z
2
=0 (3)
where u is the electric potential. This equation can be
transformed into different mathematical formulations.
Then, various numerical methods are employed to solve
it with different levels of efﬁciency.
According to the domain of the above Laplace equation
(3), there are two models for capacitance extraction: (1) the
inﬁnitedomain model, in which the electrostatic ﬁeld
spreads to the inﬁnite, resulting in an inﬁnite problem
space; and (2) the ﬁnitedomain model, where the electro
static ﬁeld is restricted within a ﬁnite domain, with the
Neumann condition on the outer boundary [8]:
(@u=@n) =0. This means that electric ﬁeld is not able to
spread out of the ﬁnite problem domain. The Neumann
condition is also called the reﬂective boundary condition,
and is introduced as the ‘‘magnetic wall’’ in Ref. 9. It
should be pointed out that the inﬁnitedomain model is
ideal for simulating isolated structures, but for the onchip
application it is not accurate because of the inﬂuence of
neighboring conductors. On the other hand, the ﬁnite
domain model considers a part cut from actual layout of
VLSI circuit; it is suitable for the realistic capacitance ex
traction of VLSI interconnects [8]. Now, both models of
capacitance extraction are used in different applications,
and accordingly the numerical methods are also different.
The problems a numerical algorithm usually encountered
in modeling are discussed below.
Classiﬁcations of the 3D ﬁeld solver methods include
the domain discretization method, the boundary integral
equation method, semianalytical approaches, and the
stochastic method. The domain discretization method
includes the ﬁnitedifference method (FDM) [10], ﬁnite
element method (FEM) [11], and the method of the mea
sured equation of invariance (MEI) [13,14]. The boundary
integral equation method includes the method of moment
[15], indirect boundary element method (BEM) [8,17–27],
and direct boundary element method [28–34]. The
semianalytical approaches combine the analytical
formulas and some traditional numerical methods
[9,35–37]. The stochastic method is based on statistical
theory [38].
FDM and FEM discretize the entire 3D domain, thus
producing a linear algebra system with large order; hence
the computational speed of these methods is greatly lim
ited. However, since both methods are relatively well es
tablished, they are still used in the industry as a reference
tool with accurate values calculated under ﬁne grids. For
example, the famous software of 2/3D capacitance extrac
tion ‘‘Raphael’’ utilizes FDM, and the ‘‘SpiceLink’’ of
Ansoft Corp. is based on FEM.
Since the mid1990s, the boundary integral equation
method has begun to replace the domain discretization
method because of its high performance. In both indirect
and direct BEM, only the boundary of 3D domain is disc
retized, and a smaller system of linear equations is ob
tained. Problems encountered with the complex boundary
can be effectively handled with BEM, whose accuracy is
superior to that of FEM as well. Thus, the BEM with rapid
computating techniques has become the focus of research
on the 3D ﬁeld solver.
3.3.2. Indirect BoundaryElement Method. The indirect
boundary method can be regarded as a variation of the
method of moments (MoM). Because only the domain
boundary needs to be discretized, the indirect BEM in
volves much fewer unknowns than does FDM or FEM.
However, it leads to a dense coefﬁcient matrix, whose for
mation and solution introduce many difﬁculties. The
innovation of the multipole acceleration method, the
singularvalue decomposition (SVD) method, and the hi
erarchical method has made the indirect BEM more
applicable. Now, indirect BEM combined with a fast com
putational technique has become a main choice for the 3D
ﬁeld solver.
The indirect BEM method is also called the equivalent
charge method, whose boundary integral equation
involves the surface charge density s(x
/
) as an unknown
570 CAPACITANCE EXTRACTION
function
u(x) =
_
G
G(x; x
/
)s(x
/
)da
/
(x c G) (4)
where G(x; x
/
) is Green’s function. For free space,
G(x; x
/
) =1=[[x ÷x
/
[[; G is the boundary surface. After
solving the surface charge density s(x
/
), the charge on con
ductor i can be calculated with
Q
i
=
_
S
d
(i)
s(x
/
)da
/
(5)
where S
d
(i) is the surface of conductor i. We discretize the
surfaces of m conductors into n constant elements (or pan
els); then the potential at the center of the kth panel x
k
can be expressed as a sum of the contributions of all the
panels
u
k
=
n
j =1
_
G
j
s
j
(x
/
)
x
/
÷x
k
 
da
/
where s
j
(x
/
) is the surface charge density of panel j (G
j
).
Substituting the known boundary conditions, we obtain a
dense linear algebra equation.
Pq =b (6)
where the coefﬁcient matrix P is dense and nonsymmetric.
The Krylov subspace iterative method, such as the gener
alized minimal residual algorithm (GMRES) [2], is usually
used to solve this equation.
For a problem involving multiple dielectrics, the polar
ization charge density on the dielectric interface needs to
be introduced, which contributes to the potential distri
bution together with the free charge density on conductor
surfaces. Therefore, the problem becomes equivalent to
that in the free space and the simple freespace Green
function is used to form Eq. (4). Except for Eq. (4) on each
conductor panel, the normal derivative of the potential
satisﬁes
e
a
@u
÷
(x)
@n
a
=e
b
@u
÷
(x)
@n
a
(7)
with xAinterface of e
a
and e
b
at any point x on a dielectric
interface. Here n
a
is the normal to the dielectric interface
at x that points into dielectric a and e
a
and e
b
are the per
mittivities of the corresponding homogenous dielectric re
gion; u
÷
(x) is the potential at x approached from the side
of the interface e
a
, and u
÷
(x) is the analogous potential for
the b side.
For the multidielectric problem, the socalled total
charge Green function approach presented above involves
more unknowns at the interfaces. Another choice to deal
with the problem is to employ the multilayered Green
function. Then, only the free charge density on the con
ductor surfaces needs to be considered as an unknown
function. However, to evaluate the Green function for the
multilayered medium, inﬁnite summations are involved,
which is very timeconsuming. Oh et al. [20] derived a
closedform expression of Green’s function for the multi
layered medium by approximating the Green function us
ing a ﬁnite number of images in the spectral domain. This
greatly reduces the computational task. Li et al. [22] pre
sented for the ﬁrst time the general analytical formulas
for the static Green functions for shielded and open arbi
trarily multilayered media. Zhao et al. [21] an efﬁcient
scheme for the generation of multilayered Green functions
using a generalized image method presented. The multi
layered Green function is much more complicated than
the freespace Green function; it is applicable only to
the simple stratiﬁed structure of multiple dielectrics,
while for more complex structures, such as the conformal
dielectric, the deduction of Green’s function may be
impossible.
More research work has been undertaken to accelerate
the capacitance extraction using the totalcharge Green’s
function approach. In 1991, Nabors et al. applied the
multipole accelerated (MPA) method successfully,
proposed earlier by Greengard and Rokhlin [16], to 3D
capacitance extraction with the indirect BEM. In the
MPA method, calculation of the interaction between
charges [i.e., the coefﬁcients in (6)] is divided into two
parts: the nearﬁeld computation and the farﬁeld compu
tation. For the nearﬁeld computation, the coefﬁcients
are calculated directly; for the farﬁeld computation,
the multipole expansion and local expansion are used to
expedite the computation. Therefore, the CPU time of
forming and solving (6) with the iterative equation solver
is greatly reduced. Figure 10 illustrates of the multipole
expansion. Nabors and White [18], developed the adap
tive, preconditioned MPA method. The corresponding
software prototype FastCap is shared on the MIT
Website, and has become a popular tool of capacitance ex
traction for relevant researchers. To date, the capacitance
extraction using the MPA indirect BEM is still undergoing
research [25].
In 1998, a fast hierarchical algorithm for 3D capaci
tance extraction was proposed at the Design Automation
Conference, and was reprinted in a journal article [24].
Similar to the multipole algorithm, it is also based on fast
computation of the ‘‘Nbody’’ problem. For the singular in
tegral kernel of 1=[[x ÷x
/
[[, it can achieve high acceleration
of computation, and only O(N) operations are needed
n
1
evaluation points
n
2
charge points
r
i
r
R
j
i
Figure 10. Evaluation point potentials are approximated with a
multipole expansion [17].
CAPACITANCE EXTRACTION 571
for each iteration. For other weakersingular kernels, the
efﬁciency of this method may be reduced. In 1997, Kapur
et al. and Long [19] proposed an accelerated method based
on the singularvalue decomposition (SVD) method that is
independent of the kernel and based on the Galerkin
method using the pulse function as the basis function. It
requires an O(N) times operation to construct the coefﬁ
cient matrix and O(Nlog N) operations to perform an
iteration. The precorrected fast Fourier transform
(FFT) algorithm [23] has the same computational com
plexity, while it is based on the collocation method for
discretization.
These studies on capacitance extraction with indirect
BEM all handle the inﬁnitedomain model. In 1996, Wang
et al. [8] improved the multipole accelerated indirect
BEM, enabling it to handle the ﬁnitedomain problem
and also proposed a parallel multipole accelerated 3D
capacitance simulation method based on nonuniformed
cube partition.
Other fast computational methods for indirect BEM
include those based on wavelets [26] and the multiscale
method [27].
3.3.3. Direct BoundaryElement Method. The direct
BEM is based on the direct boundary integral equation
(BIE), and is suitable for solving the 3D Laplace equation
with varied boundary conditions [12]. However, the direct
BEM method is generally used to deal with the ﬁnitedo
main model of capacitance extraction.
Within the ﬁnite domain that is involved in capacitance
extraction (see Fig. 11), the electric potential u satisﬁes
the following Laplace equation with mixed boundary con
ditions [32]
e
i
V
2
u=0; inO
i
(i =1; . . . ; M)
u=u
0
; on G
u
q =@u=@n=q
0
=0; on G
q
_
¸
¸
_
¸
¸
_
(8)
where the whole domain O= C
M
O
i
, where O
i
stands for the
space possessed by the ith dielectric. G
u
represents the
Dirichlet boundary (conductor surfaces), where u is
known as the bias voltages; G
q
represents the Neumann
boundary (outer boundary of the simulated region), where
the electric ﬂux q is supposed to be zero. Here n denotes
the unit vector outward normal to the boundary. At the
dielectric interface, the compatibility equation (7) holds.
With the fundamental solution as the weighting func
tion, the Laplace equations in (8) are transformed into the
following direct BIEs by the Green identity [12]
c
s
u
i
s
÷
_
@O
i
q
+
u
i
dG=
_
@O
i
u
+
q
i
dG (i =1; . . . ; M)
where u
i
s
is the electric potential at collocation point s (in
dielectric region i) and c
s
is a constant dependent on the
boundary geometry near to the point s. u
+
=1=4pr is the
fundamental solution of the 3D Laplace equation, whose
derivative along the outward normal direction n is
q
+
=@u
+
=@n=÷(r; n)=4pr
3
, r is the distance from the col
location point to the point on G, and qO
i
is the boundary
that surrounds dielectric region i.
Employing the collocation method after discretizing the
boundary, such as that in the indirect BEM, we obtain
system of linear equations [32]:
Ax=f (9)
Finally, with the preconditioned Krylov iterative equat
ion solver, such as the GMRES algorithm [2], the normal
electric ﬁeld intensity on the conductor surface is obtai
ned [32].
In direct BEM, variables of both potential and ﬁeld in
tensity are involved; thus two kinds of integral kernels are
found. Although this is more complex than the indirect
BEM method, direct BEM has its own advantages: (1) it is
suitable for capacitance extraction within the ﬁnite do
main since two variables are included, and (2) because the
variables in each BIE are within the same dielectric re
gion, it has a ‘‘localization’’ characteristic, which leads to a
sparse linear system for problem with multiple dielectrics.
In direct BEM, a great deal of time and memory are
consumed in forming and solving the system of discretized
BEM equations. Wang et al. continued the research work
of Fukuda [28] on 2D capacitance extraction using direct
BEM, extending it to the 3D structure of VLSI intercon
nects [32]. An efﬁcient analytical/semianalytical integra
tion scheme was used to accurately calculate the boundary
integrals under the VLSI planar process. This method
achieves high computational speed and accuracy when
forming Eq. (9) [32]. In 1996, Bachtold et al. [29] extended
the multipole method to handle the ‘‘potential boundary
integral’’ (whose kernel is 1/r
3
) in the direct BEM. They
discussed the model of multiple dielectrics within the in
ﬁnite domain. In 1999, Gu et al. extended the fast hierar
chical method used in the indirect BEM and made
it feasible to apply it for directBEMbased capacitance
extraction [30].
In 2000, Yu et al. proposed a quasimultiple medium
(QMM) method, based on the localization characteristic of
direct BEM [32]. The QMM method exploits the sparsity of
the resulting coefﬁcient matrix when handling the multi
dielectric problem. Together with the efﬁcient equation
organization and iterative solving technology, the QMM
accelerated method has greatly reduced the computing
time and memory usage. Figure 12 shows that a typical 3D
interconnect capacitor with ﬁve dielectric layers is cut into
5 3 2 ﬁctitious medium regions. The QMM method has
been successfully applied to actual 3D multidielectric
Substrate
1
2
3
Conductor
Neumann boundary
Figure 11. A structure with three dielectrics (crosssectional
view).
572 CAPACITANCE EXTRACTION
capacitance extraction [32,34]. For the ﬁnitedomain mul
tidielectric problem, the QMMbased method has shown a
10 higher computation speed and memory saving over
the multipole approach (FastCap 2.0) with comparable ac
curacy [34].
Another kind of ﬁeld solver, called the ‘‘global ap
proach,’’ does not solve the resulting linear system in the
usual way. The global approach discretizes the ﬁeld equa
tions and converts them to a circuit network of resistors or
capacitors. Finally, with circuit reduction or matrix com
putation, the whole resistance or capacitance matrix can
be obtained directly. In 1997, Dengi of Carnegie Mellon
University proposed a global approach (called ‘‘macromodel’’
method) for 2D interconnect capacitance extraction based
on direct BEM [31]. More recently, Lu et al. successfully
extended the concept of boundary element macromodel to
the 3D case, and developed a rapid hierarchical block
boundary element method (HBBEM) for interconnect ca
pacitance extraction [33].
3.3.4. Semianalytical Approaches. Semianalytical ap
proaches have been proposed as a solution for 3D capaci
tance extraction. Basically, they take certain special
procedures and reduce the original problem by one dimen
sion, such as using domain decomposition. Since some sub
domains with specific geometry symmetry can be handled
using the analytical formula, these approaches have very
high computational speed as well as much less memory
usage. Another characteristic of these approaches is that
the FDM is often used for the general and complicated sub
domain. That is why these approaches are sometimes con
sidered as improvements over the ﬁnitedifference method.
The semianalytical approaches include the dimension
reduction technique (DRT) [9] and techniques based on
the domain decomposition method [35–37]. The principles
of the latter two techniques will be briefly discussed as
follows.
3.3.4.1. Dimension Reduction Technique. The DRT at
tempts to solve problems within the ﬁnite domain. Most
VLSI interconnects have stratiﬁed structures, and every
layer is homogeneous along the direction perpendicular to
the interfaces of the layers (denoted as the z direction; see
Fig. 13). The DRT takes full advantage of this fact. It ﬁrst
partitions the whole structure according to these homoge
neous layers. Then, for each layer the 3D Laplace equation
can be reduced to a 2D Helmholtz equation, which is
solved with the most efﬁcient method (including the ana
lytical formula) according to the arrangement of the con
ductors. Finally, the solutions for these cascading 2D
problems are combined together to yield the ﬁnal result.
For the ﬁnitedomain problem of the ith layer with
Eq. (8), denote W
(i)
(x; y; V
c
) as a linear function of x, y and
the bias voltage setting on conductors (denoted by vector
V
c
), and let
u
(i)
=v
(i)
÷W
(i)
(x; y; V
c
):
If there exists a function such as W
(i)
(x; y; V
c
), that func
tion v
(i)
satisﬁes
V
2
v
(i)
(x; y; z) =0
v
(i)
(x; y; z) =0; (x;y) c G
(i)
u
@v
(i)
(x; y; z)
_
@n=0; (x; y) c G
(i)
q
_
¸
¸
_
¸
¸
_
then from the method of separation of variables, the gen
eral solution of v
(i)
is
v
(i)
(x; y; z) =
m=1
T
(i)
m
(x; y)L
(i)
m
(z)
where T
(i)
m
is the mode function fulﬁlling the Helmholtz
equation and L
(i)
m
can be solved analytically [9].
According to the conductor arrangement in the layer
and the preceding analysis, the layer slices are classiﬁed
as follows:
1. An Empty layer or a layer containing some simple
conductors (such as that involving straight lines
penetrating the structure) for which the linear func
tion W and the analytical solution of the Helmholtz
equation both exist.
2. The layer for which the linear function W exists,
allowing the corresponding 3D problem to be trans
ferred into the 2D Helmholtz equation.
3. A complex layer for which the W function does not
exist. The 3D Laplace equation must be solved, but
Master conductor, 1 V
(Other conductors are with 0V)
y
z
x
Neumann boundaries,
where electrical flux is 0
Dielectric layers
Figure 12. A typical 3D interconnect capacitor with ﬁve dielec
tric layers is cut into 3 2 structure.
x
z
y
Conductor
Dielectric
Figure 13. A 3D interconnect capacitor and the stratiﬁed layers.
CAPACITANCE EXTRACTION 573
only the 2D ﬁnitedifference grid is utilized because
of the geometry symmetry along the z direction.
The main drawback of the DRT is that the geometry it
employs has some limitations; For instance, it is difﬁcult to
apply DRT to nonplanarized structures. So, for general
ized and complicated interconnect structures using the
DSM technology, the efﬁciency of DRT is not guaranteed.
3.3.4.2. Domain Decomposition Method. The domain
decomposition method (DDM) is a newly developed nu
merical method. It can be subgrouped into the overlapping
domain decomposition method (ODDM) and the nonover
lapping domain decomposition method (NDDM). The for
mer is also called the Schwarz alternating method and the
latter, the Dirichlet–Neumann alternating method.
ODDM partitions the whole structures into some over
lapped subdomains. Then, a global iteration is used for the
solution. Its principles are discussed below [35].
Consider a 3D ﬁnite domain Laplace problem with the
Dirichlet boundary condition
V
2
u=0; (x; y; z) c O
u
G
=g(x; y; z)
¸
¸
_
Assume that the problem domain O involves two overlap
ping subdomains O
1
and O
2
(see Fig. 14), and denote G
j
and L
j
as the outer boundary and ﬁctitious boundary of
O
j
(j =1; 2), respectively. Then, the Schwarz alternating
method is represented as
V
2
u
i ÷1
1
=0; (x; y; z) c O
1
u
i ÷1
1
=u
i
2
; (x; y; z) c L
1
u
i ÷1
1
=g(x; y; z); (x; y; z) c G
1
÷L
1
_
¸
¸
_
¸
¸
_
V
2
u
i ÷1
2
=0; (x; y; z) c O
2
u
i ÷1
2
=u
i ÷1
1
; (x; y; z) c L
2
u
i ÷1
2
=g(x; y; z); (x; y; z) c G
2
÷L
2
_
¸
¸
_
¸
¸
_
with i =0; 1; 2; . . ., where u
0
is the initial value for itera
tion. In each iterative step, the known values of u on L
1
are used to solve the ﬁeld of subdomain O
1
. Then, the ﬁeld
of subdomain O
2
is resolved with the u obtained on L
2
. The
discrepancy of u on L
1
between two adjacent iterative
steps is used as the criterion of convergence. A relaxation
factor o can be introduced to these formulas to accelerate
the convergence. It is also obvious that the convergence
rate of the Schwarz alternating method is closely related
to the size of the overlapping region. Usually the iteration
error decreases exponentially with increase in the ratio of
the overlapping domain over the subdomain [35].
It is straightforward to extend the preceding formulas
of two subdomains to the generalized case with more sub
domains. In each iterative step an analysis similar to that
used in DRTcan be employed to achieve high efﬁciency. In
the actual application to capacitance extraction, the iter
ation sequence and selection of relaxation factor need to be
considered. Figure 15 shows a crosssectional view of an
interconnect capacitor with nine layers, and the domain
partition scheme is illustrated.
In the NDDM technique, the decomposed subdomains
do not overlap each other, while the iteration algorithm is
similar to that in ODDM; the difference is that in the
adjacent subdomains the problem is solved with the
Dirichlet boundary condition and the Neumann boundary
condition, respectively, in the NDDM. In NDDM, there are
fewer unknowns in the subdomain, and sometimes only
2D discretization is needed for a simple subdomain with
homogeneous structure. However, the convergence rate of
NDDM is slower than that of ODDM [36].
Research on capacitance extraction based on the do
main decomposition method is still underway. More recent
progress can be found in Ref. 37.
3.3.5. Other Methods. The measured equation of in
variance (MEI) method can be considered as a variation
of FDM. To solve the inﬁnitedomain model of capacitance
extraction, the MEI method terminates the meshes very
close to the object conductors and still preserves the spar
sity of the ﬁnitedifference (FD) equations. The geometry
independent measured equation of invariance (GIMEI) is
proposed for the capacitance extraction of the general 2D
and 3D interconnects by using the freespace Green func
tion only [13]. The MEI method has now been developed to
the onsurface level, where a surface mesh is used to min
imize the number of unknowns [14]. The stochastic
method is based on the randomwalk theory and can ef
fectively handle complex 3D structures. Its most recent
progress can be found in Ref. 38.
4. PERSPECTIVE
In this article, we reviewed the state of the art in capac
itance extraction techniques. These methods are discussed
Ω
2
Ω
1
Λ
1
Λ
2
Γ
1
Γ
2
Figure 14. Two overlapping subregions.
x
z
D
9
D
8
D
6
D
4
D
2
D
7
D
5
D
3
D
1
Figure 15. Four conductors embedded in nine dielectric layers.
574 CAPACITANCE EXTRACTION
mainly for the accurate analysis of VLSI interconnects.
However, they can also be easily to applied to computer
aided design of microwave ICs. With the development of
IC technologies, the following issues related to capacitance
extraction are important:
1. 3D capacitance extraction (i.e., 3D ﬁeld solution) has
the highest computational accuracy, and is suitable
for complex interconnect structure under the DSM
technologies. Many accelerating techniques have
been developed to improve its speed. However, it is
yet not feasible to use the 3D ﬁeld solver directly in
the fullchip extraction task. More effort should be
devoted to improving the computational speed of the
3D ﬁeld solver, or to develop special techniques for
the fullchip task. The fullchip extraction method
employing the 3D ﬁeld solver would give both high
computational speed and high accuracy.
2. Currently, few 3D capacitance extraction methods
can be ‘‘tuned’’ for performance versus accuracy. The
error estimation of the boundaryelement method is
also not established for practice. To make the 3D
ﬁeld solver suitable for various applications, its ﬂex
ibility in tradeoff of accuracy versus computational
performance must be improved. The adaptive algo
rithm and stable element partition scheme will be
the focus of research in the future.
3. Mixedsignal integrated circuits have been demon
strated to provide highperformance system solu
tions for various applications such as wireless
communications. Also, the siliconbased CMOS
technology is increasingly widely used because
of the fabrication cost advantage. To consider the
significant impact of the lossy nature of the silicon
substrate on the onchip interconnects of the mixed
signal ICs, the frequencydependent parameters of
interconnects in highspeed circuits must be extract
ed accurately. Deﬁning the complex permittivity of a
material, the parasitic capacitance and conductance
in a frequencydependent model can be extracted
using methods similar to that employed for tradi
tional capacitance extraction. The most efﬁcient
algorithms for frequencydependent capacitance ex
traction should be considered.
Acknowledgment
The authors would like to thank Prof. W. Hong, Southeast
University, Nanjing, China, for much helpful advice.
BIBLIOGRAPHY
1. W. H. Kao, C. Lo, M. Basel, and R. Singh, Parasitic extraction:
Current state of the art and future trends, Proc. IEEE
89(5):729–739 (2001).
2. Y. Saad and M. H. Schultz, GMRES: A generalized minimal
residual algorithm for solving nonsymmetric linear systems,
SIAM J. Sci. Stat. Comput. 7(3):856–869 (1986).
3. U. Choudhury and A. SangiovanniVicenelli, Automatic gen
eration of analytical models for interconnect capacitances,
IEEE Trans. Comput. Aided Design 14(4):470–480 (1995).
4. N. D. Arora, K. V. Raol, R. Schumann, and L. M. Richardson,
Modeling and extraction of interconnect capacitances for mul
tilayer VLSI circuits, IEEE Trans. Comput. Aided Design
15(1): 58–67 (1996).
5. P. A. Habitz and I. L. Wemple, A simpler, faster method of
parasitic capacitance extraction, Electron. J. 11–15 (Oct.
1997).
6. E. A. Dengi, A Parasitic Capacitance Extraction Method for
VLSI Interconnect Modeling, Ph.D thesis, Carnegie Mellon
Univ., March 1997.
7. J. D. Jackson, Classical Electrodynamics, Wiley, New York,
1975.
8. Z. Wang, Y. Yuan, and Q. Wu, A parallel multipole accelerated
3D capacitance simulator based on an improved model,
IEEE Trans. Comput. Aided Design 15(12):1441–1450
(1996).
9. W. Hong, W. Sun, and Z. Zhu, A novel dimensionreduction
technique for the capacitance extraction of 3D VLSI inter
connects, IEEE Trans. Microwave Theory Tech. 46(8):
1037–1043 (1998).
10. A. Seidl, M. Svoboda, et al., CAPCALA 3D capacitance solv
er for support of CAD systems, IEEE Trans. Comput. Aided
Design 7(5):549–556 (1988).
11. T. Chou and Z. J. Cendes, Capacitance calculation of IC pack
ages using the ﬁnite element method and planes of symmetry,
IEEE Trans. Comput. Aided Design 13(9):1159–1166
(1994).
12. C. A. Brebbia. The Boundary Element Method for Engineers,
Pentech Press, London, 1978.
13. W. Sun, W. W. Dai, and W. Hong, Fast parameter extraction of
general interconnects using geometry independent measured
equation of invariance, IEEE Trans. Microwave Theory Tech.
45(5):827–835 (1997).
14. Y. W. Liu, K. Lan, and K. K. Mei, Computation of capacitance
matrix for integrated circuit interconnects using onsurface
MEI method, IEEE Microwave Guided Wave Lett. 9(8):
303–304 (1999).
15. R. F. Harrington, Field Computation by Moment Methods,
Macmillan, New York, 1968.
16. L. Greengard and V. Rokhlin, A fast algorithm for particle
simulations, J. Comput. Phys. 73:325–348 (1987).
17. K. Nabors and J. White, FastCap: A multipole accelerated 3D
capacitance extraction program, IEEE Trans. Comput. Aided
Design 10(11):1447–1459 (1991).
18. K. Nabors and J. White, Multipoleaccelerated capacitance
extraction algorithms for 3D structures with multiple dielec
trics, IEEE Trans. Circ. Syst. I: Fund. Theory Appl.
39(11):946–954 (1992).
19. S. Kapur and D. Long, IES
3
: A fast integral equation solver
for efﬁcient 3dimensional extraction, Proc. IEEE/ACM Int.
Conf. Comput. Aided Design 34:448–455 (1997).
20. K. S. Oh, D. Kuznetsov, and J. E. SchuttAine, Capacitance
computations in a multilayered dielectric medium using
closedform spatial Green’s functions, IEEE Trans. Micro
wave Theory Tech. 42(8):1443–1453 (1994).
21. J. Zhao, W. W. M. Dai, et al., Efﬁcient threedimensional ex
traction based on static and fullwave layered Green’s func
tions, Proc. IEEE/ACM Design Automation Conf. 35:224–229
(1998).
22. K. Li, K. Atsuki, and T. Hasegawa, General analytical solu
tion of static Green’s functions for shielded and open arbi
trarily multilayered media, IEEE Trans. Microwave Theory
Tech. 45(1):2–8 (1997).
CAPACITANCE EXTRACTION 575
23. J. R. Phillips and J. K. White, A precorrectedFFT method for
electrostatic analysis of complicated 3D structures, IEEE
Trans. Comput. Aided Design 16(10):1059–1072 (1997).
24. W. Shi, J. Liu, N. Kakani, and T. Yu, A fast hierarchical al
gorithm for threedimensional capacitance extraction, IEEE
Trans. Comput. Aided Design 21(3):330–336 (2002).
25. Y. C. Pan, W. C. Chew, and L. X. Wan, A fast multipolemeth
odbased calculation of the capacitance matrix for multiple
conductors above stratiﬁed dielectric media, IEEE Trans.
Microwave Theory Tech. 49(3):480–490 (2001).
26. N. Soveiko and M. S. Nakhla, Efﬁcient capacitance extraction
computations in wavelet domain, IEEE Trans. Circ. Syst. I:
Fund. Theory Appl. 47(5):684–701 (2000).
27. J. Tausch and J. White, A multiscale method for fast capac
itance extraction, Proc. IEEE/ACM Design Automation Conf.
36:537–542 (1999).
28. S. Fukuda, N. Shigyo, et al., A ULSI 2D capacitance simu
lator for complex structures based on actual processes, IEEE
Trans. Comput. Aided Design 9(1):39–47 (1990).
29. M. Bachtold, J. G. Korvink, and H. Baltes, Enhanced multi
pole acceleration technique for the solution of large possion
computations, IEEE Trans. Comput. Aided Design
15(12):1541–1546 (1996).
30. J. Gu, Z. Wang, and X. Hong, Hierarchical computation of 3D
interconnect capacitance using direct boundary element
method. Proc. IEEE Asia South Pacific Design Automation
Conf. 2000, Vol. 6, pp. 447–452.
31. E. A. Dengi and R. A. Rohrer, Boundary element method
macromodels for 2D hierarchical capacitance extraction,
Proc. IEEE/ACM Design Automation Conf. 35:218–223
(1998).
32. W. Yu, Z. Wang, and J. Gu, Fast capacitance extraction of ac
tual 3D VLSI interconnects using quasimultiple medium
accelerated BEM, IEEE Trans. Microwave Theory Tech.
51(1):109–120 (2003).
33. T. Lu, Z. Wang, and W. Yu, Hierarchical block boundaryele
ment method (HBBEM): A fast ﬁeld solver for 3D capaci
tance extraction, IEEE Trans. Microwave Theory Tech.
52(1):10–19 (2004).
34. W. Yu and Z. Wang, Enhanced QMMBEM solver for three
dimensional multipledielectric capacitance extraction within
the ﬁnite domain, IEEE Trans. Microwave Theory Tech.
52(2):560–566 (2004).
35. Z. Zhu, H. Ji, and W. Hong, An efﬁcient algorithm for the pa
rameter extraction of 3D interconnect structures in the VLSI
circuits: domaindecomposition method, IEEE Trans. Micro
wave Theory Tech. 45(7):1179–1184 (1997).
36. Z. Zhu and W. Hong, A generalized algorithm for the capac
itance extraction of 3D VLSI interconnects, IEEE Trans.
Microwave Theory Tech. 47(10):2027–2030 (1999).
37. V. V. Veremey and R. Mittra, Domain decomposition appro
ach for capacitance computation of nonorthogonal interco
nnect structures, IEEE Trans. Microwave Theory Tech. 48(9):
1428–1434 (2000).
38. A. Brambilla and P. Maffezzoni, A statistical algorithm for 3D
capacitance extraction, IEEE Microwave Guided Wave Lett.
10(8):304–306 (2000).
39. I. Bahl, Lumped Elements for RF and Microwave Circuits,
Artech House, 2003.
40. L. Zhu and K. Wu, Accurate circuit model of interdigital ca
pacitor and its application to design of new quasilumped
miniaturized ﬁlters with suppression of harmonic resonance,
IEEE Trans. Microwave Theory Tech. 48(3):347–356 (2000).
CAVITY RESONATORS
ARVIND K. SHARMA
TRW
Redondo Beach, California
1. RESONANT STRUCTURES
Resonant structures are network elements that are used
extensively in the development of various microwave com
ponents [1]. At low frequencies, resonant structures are
invariably composed of lumped elements. As frequencies
increase, lumpedelement resonant circuits are attained
by using transmission lines. Microwave resonant struc
tures are almost invariably understood as cavity resona
tors. Conventional resonators consist of a bounded
electromagnetic ﬁeld in a volume enclosed by metallic
walls. The electric and magnetic energies are stored in
the electric and magnetic ﬁelds, respectively, of the elec
tromagnetic ﬁelds inside the cavity and the equivalent
lumped inductance and capacitance of the structure can
be determined from the respective stored energy. It is im
portant to note that cavity resonators, in contrast to
lumped resonators, have an inﬁnite number of resonant
frequencies (or modes). In the vicinity of each resonant
frequency, the cavity can be approximated by an associat
ed lumped equivalent circuit.
Some energy is dissipated as ﬁnite conductivity of
the metallic walls, and the equivalent resistance can
therefore be determined from the currents ﬂowing on
the walls of the cavity resonator [2,3]. In this article, a
brief description of the cavity resonators most commonly
employed in various microwave components is presented.
As far as possible, simple expressions have been provided
for design applications. Basic parameters of microwave
resonators are ﬁrst presented because they describe a cav
ity. Then, various coaxial and waveguide resonators are
described. Fabrication, coupling, measurements, and ap
plications of cavity resonators are also included.
2. RESONATOR PARAMETERS
2.1. Resonant Frequency
The parameters of a resonator at microwave frequencies
are essentially similar to those of a lumpedelement res
onator circuit at low frequencies. They can easily be de
scribed using an RLC series or parallel network. Consider,
for instance, an RLC parallel network as shown in Fig. 1a.
The input impedance of such a network as a function of
frequency has both real and imaginary parts. At reso
nance, the input impedance is real and is equal to the re
sistance of the circuit. The electric and magnetic stored
energies are also equal, leading to the expression for the
resonant frequency as
o
0
=
1
ﬃﬃﬃﬃﬃﬃﬃ
LC
_ (1)
576 CAVITY RESONATORS
2.2. Quality Factor
The performance of a resonant circuit is described in
terms of the quality factor Q, and such features as fre
quency selectivity, bandwidth, and damping factors can be
deduced from this. The quality factor is deﬁned as
Q=o
timeaveragedstored energy
energy lost per second
(2)
for the lumped resonant circuits
Q=oRC=
R
oL
(3)
for the parallel network in Fig. 1a, and
Q=
1
oCR
=
oL
R
(4)
for the series network in Fig. 1b.
2.3. Fractional Bandwidth
The input impedance of the parallel resonant circuit of
Fig. 1 is given by
Z
in
=
1
R
÷
1
joL
÷joC
_ _
÷1
(5)
At a frequency o
0
7Do in the vicinity of the resonant fre
quency, Eq. (5) reduces to
Z
in
=
R
1 ±j2Q(Do=o
0
)
(6)
From Eq. (6), it is clear that at o=o
0
the input impedance
is only resistive. However, when
Do=
o
0
2Q
(7)
the magnitude of the input impedance decreases to R
ﬃﬃﬃ
2
_
of
its maximum value R, and the phase angle is p/4 for ooo
0
and ÷p/4 for o4o
0
. From Eq. (7), the fractional band
width BW is deﬁned as
BW=
2Do
o
0
=
1
Q
(8)
2.4. Loaded Quality Factor
In practical situations, the resonant circuit is coupled to
an external load R
L
that also dissipates power, and the
loaded quality factor Q
L
is given by
1
Q
L
=
1
Q
÷
1
Q
e
(9)
where Q
e
is the external quality factor for a lossless res
onator in the presence of the load.
2.5. Damping Factor
Another important parameter associated with a resonant
circuit is the damping factor d
d
. It is a measure of the rate
of decay of the oscillations in the absence of an exciting
source. For highQ resonant circuits, the rate at which the
stored energy decays is proportional to the average energy
stored. Consequently, the stored energy as a function of
time is given by
W=W
0
e
÷2d
d
t
=W
0
e
÷o
0
t=Q
(10)
which implies that
d
d
=
o
0
2Q
(11)
Thus, we see that the damping factor is inversely propor
tional to the Q of the resonant circuit. In the presence of
an external load, the Q should be replaced by Q
L
.
Alternately, the input impedance in the vicinity of res
onance Z
in
given by Eq. (6) can be rewritten to take into
account the effect of losses in terms of the complex reso
nant frequency
o
c
=o
0
÷jd
d
=o
0
1 ÷j
1
2Q
_ _
(12)
so that
Z
in
=
o
0
R=(2Q)
j(o ÷o
c
)
(13)
In Eq. (13) the parameter R/Q is called the ﬁgure of merit
and describes the effect of the cavity on the gain–band
width product. In terms of the lumped elements of the
resonant circuit, we obtain
R
Q
=
ﬃﬃﬃﬃ
L
C
_
(14)
3. COAXIAL CAVITY RESONATORS
At microwave frequencies, the dimensions of lumped res
onator circuits become comparable to the wavelength, and
Z
in
I
C
C L R
(b) (a)
V L
Z
in
R
Figure 1. Lumpedelement (a) parallel and (b) series resonant
circuits.
CAVITY RESONATORS 577
this may cause energy loss by radiation. Therefore, reso
nant circuits at these frequencies are shielded to prevent
radiation. Perfectly conducting enclosures, or cavities,
provide a means of conﬁning energy. Usually, cavities
with the largest possible surface area for the current
path are preferred for lowloss operation, and the energy
is coupled to them by the various means described later in
this article.
3.1. Coaxial Resonators
A coaxial cavity resonator (Fig. 2) supporting TEM (trans
verse electromagnetic) waves can easily be formed by a
short section of coaxial line. Resonances appear whenever
the length d of the cavity is an integral number of half
wavelengths. The resonance modes occur at
f =
nc
2d
; n=1; 2; . . . (15)
where c is the speed of light. The lowest resonant frequen
cy corresponds to n=1, and the Q of the cavity for this
mode is given by [4]
Q
d
l
0
=
1
4÷2(d=b)(1 ÷b=a)= ln(b=a)
(16)
where d is the skin depth and a and b are inner and outer
radii, respectively. It is also possible to have higherorder
resonance modes, depending on the structural parameters
of the coaxial line. The ﬁrst higherorder mode appears
when the average circumference is equal to the wave
length in the dielectric medium of the line. The cutoff fre
quency of this mode is
f
c
=
c
p
ﬃﬃﬃﬃ
e
r
_
(a÷b)
(17)
where e
r
is the dielectric constant of the medium. Other
higherorder modes correspond to TE (transverse electric)
and TM (transverse magnetic) waves that exist in a cir
cular waveguide with the radius of the center conductor
approaching zero. The resonance condition is
k
nml
= p
2
nm
÷
lp
2d
_ _
2
_ _
1=2
(18)
where k
nml
=2pf
nml
/c and p
nm
is the cutoff wavenumber
that is obtained as the mth root of the transcendental
equations
J
/
n
(ka)N
/
n
(kb) ÷J
/
n
(kb)N
/
n
(ka) =0 (19)
for TE modes and
J
n
(ka)N
n
(kb) ÷J
n
(kb)N
n
(ka) =0 (20)
for TM modes. Here J
n
and N
n
are the nthorder Bessel
functions of the ﬁrst and second kinds, respectively, and
the prime denotes their derivatives with respect to their
arguments.
3.2. Reentrant Coaxial Resonators
Another coaxial cavity conﬁguration consists of a short
section of coaxial line with a gap in the center conductor.
Figure 3a shows a capacitively loaded coaxial cavity.
Radial cavity as shown in Fig. 3b is another possible varia
tion. They are also referred to as reentrant coaxial cavities
because the metallic boundaries extend into the interior of
the cavity. They are widely used in microwave tubes. The
resonant frequency of such a structure can be evaluated
from the solution of the transcendental equation
tan bl =
dc
oa
2
ln(b=a)
(21)
where d is the gap in the center conductor, and 2l ÷d is
the length of the cavity. From Eq. (21), it is obvious that
the capacitively loaded coaxial cavity can have an inﬁnite
number of modes. For the radial reentrant cavity of
Fig. 3b, the resonant frequency can be evaluated by cal
culating the inductance and capacitance of the structure.
The expression for the resonant frequency is
f =
c
2p
ﬃﬃﬃﬃ
e
r
_ al
a
2d
÷
2
l
ln
0:765
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
l
2
÷(b ÷a)
2
_
_
_
_
_
_
_ln
b
a
_
¸
_
_
¸
_
÷1=2
(22)
2a 2b
d
Figure 2. Coaxial cavity resonator and its cross section.
l
2b
(a)
(b)
d 2a
d 2b 2a
l
l
Figure 3. Reentrant coaxial cavity resonators: (a) capacitively
loaded coaxial cavity resonator; (b) radial cavity resonators.
578 CAVITY RESONATORS
An approximate expression for the Q of the cavity is
Q
d
l
0
=
2l
l
ln(b=a)
2 ln(b=a) ÷l[(1=a) ÷(1=b)]
(23)
for a tunable reentrant cavity, d is large, and (l ÷d) is also
large compared with b. The resonances occur whenever
the length of the center conductor is approximately a
quarterwavelength.
3.3. Annular Coaxial Resonator
An annular coaxial resonator is formed by a ﬁgure of rev
olution of a coaxial radial cavity resonator (refer to Fig. 3)
about an axis that is offset from and parallel to the center
conductor [5]. As shown in Fig. 4, the electric ﬁeld in the
plane containing the axis is similar to that of the radial
cavity resonator. The electric ﬁeld in the plane normal to
the axis is radial and is same along the circumference. In
effect, the annular resonator is equivalent to a half wave
length coaxial resonator with a small shunt capacitance in
the middle. One of the important application of this type of
resonator is that it can be coupled simultaneously with
several sources. The electric ﬁeld at the gap is quite high
and results in good coupling to external sources.
4. WAVEGUIDE CAVITIES
4.1. Rectangular Waveguide Resonators
Rectangular resonant cavities are formed by a section of
rectangular waveguide of length d. This cavity can also
support an inﬁnite number of modes. The ﬁeld conﬁgura
tion of the standingwave pattern for the incident and re
ﬂected waves is not unique, that is, it depends on the
assumed direction of propagation of the wave. In order to
be consistent, we shall assume that wave propagation is in
the positive z direction. The standingwave pattern is then
formed by the incident and reﬂected waves traveling in
÷z and ÷z directions, respectively. The cutoff wave
number k
cmn
is given by
k
2
cmn
=
mp
a
_ _
2
÷
np
b
_ _
2
; m=0; 1; 2; . . . ; n=0; 1; 2; . . .
(24)
where a and b are waveguide dimensions. The resonant
wavenumber is then expressed as
k
mnp
=
mp
a
_ _
2
÷
np
b
_ _
2
÷
pp
d
_ _
2
_ _
1=2
; p=1; 2; . . . (25)
and the resonant frequency is deﬁned as
f
mnp
=
k
mnp
c
2p
(26)
From the preceding discussion, we see that the resonant
frequency is the same for TE and TM modes. Therefore,
they are referred to as degenerate modes. The ﬁeld conﬁg
uration of the dominant TE
101
mode is shown in Fig. 5b.
The quality factor Q of the dominant TE
101
mode in the
rectangular resonant cavity having surface resistance R
s
can be evaluated using the expression
Q=
120p
2
4R
s
2b(a
2
÷d
2
)
3=2
ad(a
2
÷d
2
) ÷2b(a
3
÷b
3
)
_ _
(27)
g
g
d
c
b
a
Turning plunger
(a)
(b)
Figure 4. Two views of annular coaxial resonator structure.
a
y
x
z
b
d
Side
Top
(b) (a)
Figure 5. (a) Rectangular waveguide cavity resonator; (b) ﬁeld
conﬁguration of the dominant TE
101
mode.
CAVITY RESONATORS 579
In rectangular cavities, the resonant frequency increas
es for higherorder modes, as does the Q at a given fre
quency. Higherorder mode cavity or ‘‘echo boxes’’ are useful
in applications where a slow rate of decay of the energy
stored in the cavity after it has been excited is required.
4.2. Circular Waveguide Resonators
Circular waveguide cavities are most useful in various
microwave applications. Most commonly, they are used in
wavemeters to measure frequency, have a high Q factor,
and provide greater resolution. These consist of a section
of circular waveguides of radius a and length d as shown
in Fig. 6.
The resonance wavenumber of the circular waveguide
cavity is given by
k
nml
=
x
nm
a
_ _
2
÷
lp
d
_ _
2
_ _
1=2
; l =0; 1; 2; . . . (28)
where
x
nm
=
p
/
nm
for TEmodes
p
nm
for TMmodes
_
(29)
Values for p
/
nm
for various modes are given in Table 1.
Field lines for TE
111
, TM
011
, and TE
011
modes are shown in
Fig. 6. Simplifying Eq. (28) yields
2af
nml
( )
2
=
cx
nm
p
_ _
2
÷
cl
2
_ _
2
2a
d
_ _
2
(30)
The Q of the circular cavity for TE
nml
modes can be eval
uated from
Q
d
l
0
=
[1 ÷(n=p
/
nm
)
2
][(p
/
nm
)
2
÷(lpa=d)
2
]
3=2
2p[(p
/2
nm
÷2a=d(lpa=d)
2
÷(1 ÷2a=d)(nlpa=p
/
nm
d)
2
]
(31)
and for the dominant TE
111
mode, Q can be obtained by
substituting n=m=l =1 in the preceding equation. Us
ing Eq. (30), plots of (2af)
2
versus (2a/d)
2
can be used to
construct mode charts, as shown in Fig. 7. From this it can
be seen that, for the TE
011
mode operation, the safe value
of (2a/d)
2
is between 2 and 3. For TM model operation, the
Q is given by
Q
d
l
0
=
[p
2
nm
÷(lpa=d)
2
]
1=2
2p(1÷2a=d)
for l> 0
p
nm
2p(1 ÷a=d)
for l =0
_
¸
¸
¸
_
¸
¸
¸
_
(32)
As with rectangular cavity resonators, the Q is higher for
higherorder modes.
4.3. Elliptic Waveguide Resonators
Elliptic resonant cavities that are formed using a section
of an elliptic waveguide offer several advantages. There
is no mode splitting caused by slight deformations in
the cavity surface, and the electric ﬁeld conﬁguration
in the transverse plane is ﬁxed with respect to its axes.
Also, the longitudinal electric ﬁeld of the TM
111
mode in an
elliptic cavity with semi–major axis a is always greater
than the circular cavity with radius a. This feature may be
useful in the dielectric material characterization that uses
perturbation techniques [6].
Cross section through A−A
A
A
A
A
A
A
Cross section through A−A
TE
111
mode
(a)
(b)
TM
011
mode
TE
011
mode
2d
2d
2a
d
2d
z
l
l
l
Figure 6. (a) Circular cylindrical waveguide cavity resonator;
(b) ﬁeld conﬁgurations for TE
111
, TM
011
, and TE
011
modes in
cylindrical cavities.
Table 1. Roots of the Transcendental Equation J
n
/
(ka) =0
Modes
n m p
/
nm
0 1 0.0
1 1 1.841
2 1 3.054
0 2 3.832
3 1 4.201
4 1 5.318
580 CAVITY RESONATORS
The elliptic waveguide supports four different types of
modes, namely, even TE and TM modes and odd TE and
TM modes. The TE modes have E
z
=0 and the TM modes
have H
z
=0. From the solution of wave equations, there
exist four different modes. The modes having cosinetype
variation are called even modes, and modes having
sinetype variation are called odd modes. The subscripts
c and s are added to the mode designation to describe this
variation.
The elliptic waveguide in Fig. 8a is shown along with
the orthogonal elliptic coordinate system. As can be seen,
the confocal elliptic cylinders are formed with constant x,
and confocal hyperbolic cylinders are formed with con
stant Z. The distance between the two foci, F and F
/
is 2 h.
The outer wall of the elliptic waveguide is formed with
x =x
0
. The semi–major axis is then
2a=2h cosh x
0
(33)
and, the semi–minor axis is
2b =2h sinh x
0
(34)
Alternatively, the eccentricity e is
e =1= cosh x
0
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
1 ÷(b=a)
2
_
(35)
The resonance wavenumber for elliptic cavity is given by
k
rmnl
=
2
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
x
rmn
_
ae
_ _
2
÷
lp
d
_ _
2
_ _
1=2
; l =0; 1; 2; . . . (36)
where
x
rmn
=
q
/
rmn
for TE
rmn
modes
q
rmn
for TM
rmn
modes
_
(37)
In Eq. (37), r can be substituted with c and s to obtain
even and odd modes, respectively. The parameter
q
cmn
(q
smn
) is the nth parametric zero of the even (odd)
modiﬁed Mathieu function of order m with argument x
0
and is used to calculate TM
cmn
(TM
smn
) modes. Similarly,
for a TE
cmn
(TE
smn
) mode, the parameter q
/
cmn
(q
/
smn
) is the
nth parametric zero of the ﬁrst derivative of the even (odd)
20 × 10
8
15 × 10
8
TM
110
TM
010
T
E
0
1
1
,
T
M
1
1
1
T
E
2
1
1
T
E
1
1
1
T
M
0
1
1
T
M
1
1
2
T
M
0
1
2
T
E
T
E
2
1
2
10 × 10
8
5 × 10
8
0 2 4
(2a/d)
2
6
(
2
a
f
)
2
(
c
m
2
M
H
z
2
)
1
1
2
Figure 7. Mode chart of a circular cylindrical cavity resonator.
TM
C011 TE
C011
TM
S111
TM
c111
TE
S111
(b)
(a)
TE
c111
0
p =180°
p = 270°
F
x
y
F′
p = 90°
p = p
0
p = 0
¸ = ¸
0
Figure 8. (a) Elliptic waveguide cavity resonator; (b) ﬁeld con
ﬁguration for some modes in elliptical cavities.
CAVITY RESONATORS 581
modiﬁed Mathieu function of order m with argument x
0
and is used to calculate TE
cmn
(TE
smn
) modes. The equa
tions used to ﬁnd the parametric zeros are given next [7]:
TMmodes : Ce
m
(x
0
; q) =0even
Se
m
(x
0
; q) =0 odd
TEmodes : Ce
/
m
(x
0
; q) =0even
Se
/
m
(x
0
; q) =0 odd
(38)
Field lines for some modes are shown in Fig. 8b.
4.4. Annular Elliptic Resonator
The annular elliptic waveguide in Fig. 9a is shown along
with the orthogonal elliptic coordinate system. The outer
ellipse with eccentricity e
0
and the inner ellipse with ec
centricity e
1
form a confocal annular elliptic waveguide.
The distance between the two foci F and F
/
is 2h and is
related to the other structural parameters via the relation
2h=2a
0
e
0
=2a
1
e
1
(39)
where a
0
and a
1
are the semi–major axes of the outer and
inner ellipses, respectively. Alternatively, the eccentrici
ties e
0
and e
l
are also expressed as
e
0
=1= cosh x
0
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
1 ÷(b
0
=a
0
)
2
_
(40)
and
e
1
=1= cosh x
1
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
1 ÷(b
1
=a
1
)
2
_
(41)
where b
0
and b
1
are the semi–minor axes of the outer and
inner ellipses, respectively. The axial coordinates of the
outer and inner ellipses are x
0
and x
1
.
In a manner similar to the elliptic waveguide, the
eigenvalue equation for annular elliptic waveguide was
solved by Bra¨ ckelmann [8]. The relevant equations for TM
and TE modes follow.
Even modes, TE
cmn
:
Ce
/
m
(x
0
; q
cmn
)Fey
/
m
(x
1
; q
cmn
)
÷Ce
/
m
(x
1
; q
cmn
)Fey
/
m
(x
0
; q
cmn
) =0
(42)
TM
C011
TM
C101
TM
S111
TE
C101
TE
C011
TE
S111
(a)
(b)
F
x
y
¸ = ¸
0
¸ = ¸
1
F′
Figure 9. (a) Annular elliptic cavity resonator; (b) ﬁeld conﬁguration for some modes in annular
elliptic cavities.
582 CAVITY RESONATORS
Odd modes, TE
smn
:
Se
/
m
(x
0
; q
smn
)Gey
/
m
(x
1
; q
smn
)
÷Se
/
m
(x
1
; q
smn
)Fey
/
m
(x
0
; q
smn
) =0 (43)
Even modes, TM
cmn
:
Ce
m
(x
0
; q
cmn
)Fey
m
(x
1
; q
cmn
)
÷Ce
m
(x
1
; q
cmn
)Fey
m
(x
0
; q
cmn
) =0
(44)
Odd modes, TM
smn
:
Se
m
(x
0
; q
smn
)Gey
m
(x
1
; q
smn
)
÷Se
m
(x
1
; q
smn
)Fey
m
(x
0
; q
smn
) =0
(45)
In Eqs. (42)–(45), Ce
m
(x, q) and Se
m
(x, q) are the even
and odd modiﬁed Mathieu functions of the ﬁrst kind and
order m. Fey
m
(x, q) and Gey
m
(x, q) are the even and odd
modiﬁed Mathieu functions of the second kind and order
m [9]. The primes in Eqs. (42)–(45) denote the derivative
with respect to the argument x. The parameter q
/
cmn
is the
nth parametric zero of Eq. (42), and q
cmn
is the nth para
metric zero of Eq. (44). Similar explanation applies for
Eqs. (43) and (45) for the odd TE and TM modes.
The resonance wavenumber for annular elliptic cavity
is given by
k
rmnl
=
2
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
x
rmn
_
ae
_ _
2
÷
lp
d
_ _
2
_ _
1=2
; l =0; 1; 2; . . . (46)
where
x
rmn
=
q
/
rmn
for TE
rmn
modes
q
rmn
for TM
rmn
modes
_
(47)
The annular elliptic resonators also supports four differ
ent types of modes, namely, even TE and TM modes and
odd TE and TM modes. Field lines for some modes are
shown in Fig. 9b.
4.5. Spherical Resonators
Another cavity resonator shape is the spherical resonator.
Based on the solution of Maxwell’s equations in the spher
ical coordinate system, the axial symmetry results in TM
modes containing E
r
, E
y
, H
f
, and TE modes containing
H
r
, H
y
, E
f
. Because the origin is included inside the
hollow spherical cavity, the resonance condition is easily
obtained by setting E
y
=0 at r =a, where a is the radius of
the sphere.
Solution of the transcendental equation
tan ka=
ka
1 ÷(ka)
2
(48)
results in dominant TM
101
resonance at
l
0
=2:29a (49)
and the second TM
102
resonance at a wavelength of
l
0
=1:4a (50)
The modes in a spherical cavity are shown graphically in
Fig. 10.
The Q of a spherical cavity operating in the dominant
mode is
Q
d
l
0
=0:318 (51)
and the equivalent shunt resistance is simply
R
d
l
0
=104:4 (52)
4.5.1. Spherical Resonators with Reentrant Cones. Spher
ical resonators with reentrant cones were found to be
suitable for realizing oscillators for klystrons [10].
I
Section through axis
TM
101
mode
z
0
= 1.4 a
(a)
Section through equator
Axial section
TE
101
mode
z
0
= 2.29 a
(b)
Equatorial section
Electric field
Magnetic field
a
I
Figure 10. Fields in a spherical cavity resonator at the ﬁrst and
second resonant frequencies.
CAVITY RESONATORS 583
It consists of a hollow conducting sphere of radius a and
two cones whose apex is at the center of the sphere and
subtends an angle of 2y
0
. Its structure and fundamental
mode ﬁelds are shown in Fig. 11.
The resonant wavelength is not a function of y, as is the
case in spherical resonators. The resonant wavelength is
l
0
=4a (53)
However, Q and R of the resonator are functions of the
angle y. The plots of Q(d/l
0
) and R(d/l
0
) [4] are given in
Figs. 12 and 13, respectively.
As can be seen, the maximum Q is obtained at y =341
and is given by
Q
d
l
0
=0:1095 (54)
and the maximum R occurs at y =91 and is given by
R
d
l
0
=32:04 (55)
4.6. Ellipsoid–Hyperbolic Waveguide Resonators
Another cavity resonator suitable for klystrons is of ellip
soid–hyperboloid shape. This shape is a ﬁgure of revolu
tion about the axis passing through its foci, as shown in
Fig. 14. The distance between foci a as well as the
hyperboloid that determines part of the resonator is
held constant. The normalized resonant wavelength l
0
/b,
where b is the equitorial radius, is plotted as a function of
the shape factor s
0
=2b/a. Interestingly, the shape of the
resonators vary widely as the shape factor is increased.
Both the Q as well as the R are functions of the shape
factor. The Q(d/l
0
) and R(d/l
0
) (given in Ref. 4) are
plotted as a function of shape factor in Figs. 15 and 16,
respectively.
E
a
0
0
Figure 11. Spherical resonator with reentrant cones and ﬁelds
in the fundamental modes.
Q
0
0.2
0.4
0.6
0.8
1.0
90 80 70 60 50 40 30 20 10 0
0
0
0
0
o z
Figure 12. Q(d/l
0
) for a spherical resonator with reentrant
cones.
R
90 80 70 60 50 40 30 20 10 0
0
10
20
30
0
0
o z
Figure 13. R(d/l
0
) of a spherical resonator with reentrant cones.
90°
¸
0
a
7
6
5
4
3
2
1
0
0.01 0.1 1
o
0
10 100 1000
z
0
¸
0
Figure 14. Ellipsoid–hyperboloid resonator and normalized reso
nant wavelength l
0
/b as a function of shape factor s
0
=2b/a.
584 CAVITY RESONATORS
4.7. Arbitrarily Shaped Resonators
The early work in microwaves focused on analytical and
numerical solutions of hollow waveguide problems. Most
of the attempts were to solve them for TE and TM modes,
either exactly or approximately. Ng [11] compiled the
methods used to calculate the cutoff wavenumbers of hol
low waveguides. As pointed out there, three basic cross
sectional shapes can be distinguished:
1. Convex shape
2. Nonconvex with smooth reentrant portion
3. Nonconvex with sharp reentrant portion
A resonant structure can be formed using any of these
shapes by closing the hollow waveguide with endwalls.
The cutoff wavenumbers can be found using references
given in Table II of Ref. 11. The resonant frequency can be
easily calculated. The basic equation to use is
k =
2pf
0
c
= (k
c
)
2
÷
lp
d
_ _
2
_ _
1=2
; l =0; 1; 2; . . . (56)
5. FABRICATION
5.1. Materials
Microwave and millimeterwave cavities are usually made
from the same material used for the waveguide such as
copper, brass, or aluminum. In order to provide lowloss
characteristics, the interior (and exterior) is plated with
lowloss materials such as silver and gold.
There exists a wide range of waveguide sizes to cover
frequencies from as low as 400 MHz–200GHz. The oper
ating bandwidth of the waveguide increases as the fre
quency increases. Therefore, the method of fabrication is
very important in realizing lowloss or highQ cavities.
Cavities are formed using short sections of wave
guides. There are various approaches used in their
fabrication. The waveguide tubing formed using extrusion
process generally provides various dimensional tolerances
varying from 0.008 in. (0.2 mm) to 0.001 in. (0.025 mm). In
order to realize accurate waveguide dimensions, particu
larly at millimeter wavelengths, the process of electro
forming is generally used. A conducting or nonconducting
mandrel is used as a starting material in the electroform
ing process. The mandrel is later removed to leave the
electroformed waveguide. Specialgrade stainlesssteel
mandrels with high surface ﬁnish can be used to electro
form the waveguide. They are removed by heating and
applying uniform force. Nonconducting mandrels formed
using plastics or highly compressed wax can be used to
form complicated cross sections. Such mandrels can be
chemically dissolved to retain the ﬁnal form of the elect
roformed waveguide.
In most applications, the waveguide must interact with
cavities to realize the prescribed description of the com
ponent. Fabrication from a solid metal block using a mill
ing process is preferred because it provides an integrated
component for some complicated waveguide assemblies.
This approach reduces reﬂections and spurious transmis
sion by minimizing interfaces or ﬂanges.
5.2. Cavity Perturbation
At resonance the cavity contains equal amounts of aver
age electric and magnetic energy. Any perturbation in the
structural dimensions or imperfections in the cavity wall
will require readjustment in resonant frequency such that
the electric and magnetic energies are equal. It is possible
to measure accurately the frequency shift Do/o, which can
be used to determine other parameters of the cavity [2].
5.3. Effects of Temperature and Humidity
The resonant frequencies of a cavity resonator depend on
the dimensional variations of the material used in the
construction as well as on the variations in the dielectric
constant.
As temperature changes, the dimensions of the cavity
change in accordance with the thermal expansion coefﬁ
cient of the material used in its construction. The change
in the resonant frequency can be easily determined using
the equation for the resonant frequency for a given cavity
0
0.01 0.1 1 10 100
1
2
3
4
Q
o z
o
0
Figure 15. Q(d/l
0
) of an ellipsoid–hyperboloid resonator as a
function of shape factor s
0
=2b/a.
1000
100
10
0.1 1 10 100
o
0
R
o z
Figure 16. R(d/l
0
) of an ellipsoid–hyperboloid resonator as a
function of shape factor s
0
=2b/a.
CAVITY RESONATORS 585
structure. This change can be minimized by bimetals with
a lower coefﬁcient of thermal expansion.
Furthermore, the dielectric constant of the air within
an unsealed cavity also varies depending on the temper
ature, atmospheric pressure, and humidity level.
5.4. Tunable Cavities
Various microwave and millimeterwave applications re
quire resonators that can be tuned frequently and at high
speeds. Both contacting as well as noncontacting plungers
are used to tune cavity resonators.
5.4.1. Contacting Plunger. A movable short circuit is
provided by the direct contact between the plunger and
the cavity walls. The plunger is typically a quarter wave
length long at the center frequency. In order to provide
good electrical contact, the contacting plungers, as shown
in Fig. 17, have axial serrations. These serrated ﬁngers
maintain sufﬁcient pressure to scratch off any insulating
ﬁlm formed inside the cavity walls. Because the contact is
made at or near a current node, the losses are minimized.
In some cases, particularly for millimeterwave applica
tions, a metal shoulder is also added to move the short
circuit reference plane forward. In this case, the actual
contact is not at or near a current node.
Contacting resonators have several disadvantages,
such as
*
They provide erratic contact due to small metal par
ticles and nonsmooth cavity interior walls.
*
They are not repeatable because of the backlash in
the mechanical driving mechanism as well as the
friction between the contacting surfaces.
*
The contact causes wear and produces an insulating
ﬁlm, which results in increased contact resistance.
The increases losses will result in lower Q of the
cavity.
5.4.2. Noncontacting Plunger. The disadvantages of the
contacting plungers can be eliminated by using noncon
tacting plungers. These plungers provide a near short
circuit over a wider frequency range. The impedance at
the face of the plunger is a complex impedance with a
low value of resistance. The capacitive, choke, or bucket
type plungers, as shown in Fig. 18, provide reasonable
performance [5]. Multisection plungers are formed by
quarterwavelength low–high–low impedance sections.
The leakage through these sections may cause parasitic
resonance; therefore, the back of the plunger section must
be terminated in the characteristic impedance of the
transmission line used to realize the cavity.
6. COUPLING INTO AND OUT OF CAVITIES
As we have seen, the cavities are essentially enclosed
structures. In order to use them, we must couple them to
transmission lines. We can use the coaxial line or any form
of waveguide to couple power into and out of the cavities.
In this sense, the input and output coupling structures act
as a load on the cavity. The cavity parameters, such as
resonant frequency and Q, are invariably affected by the
presence of these structures. The resonant behavior of the
cavities is exploited extensively in the realization of ﬁlters
with prescribed functional forms.
6.1. General Coupling
Coupling structures provide a means of coupling energy
into and/or out of the cavity. The excitation of the cavity
can be accomplished by electric or magnetic coupling. In
case of electric coupling, the electric ﬁeld of the coupling
structure is parallel to the electric ﬁeld of the cavity. The
magnetic coupling is provided when the magnetic ﬁeld of
the coupling structure is parallel to the magnetic ﬁeld of
the cavity.
The coaxial line can be used to provide either electric or
magnetic coupling.
1. Electric Probes. The center conductor of the coaxial
line acts as a probe. Its direction is parallel to the
direction of the electric ﬁeld in the cavity.
2. Current Loops. The center conductor of the coaxial
line is terminated in a short circuit to form a loop.
The loop produces a magnetic ﬁeld perpendicular to
the plane of the probe and in the same direction as
the magnetic ﬁeld in the cavity.
(a)
(b)
Figure 17. Contacting plunger.
b
z
g
z
g
z
g
4 4 4
Figure 18. Noncontacting plunger.
586 CAVITY RESONATORS
Cavities are also excited by waveguides though aper
tures formed by holes and slits (Fig. 19). The coupling
mechanism can be of electric or magnetic type.
1. Magnetic Coupling Apertures. The aperture is located
between the cavity and input waveguide such that
the magnetic ﬁeld in the waveguide is parallel to the
magnetic ﬁeld in the cavity. Round holes in the wall
separating the waveguide and cavity provide mag
netic coupling.
2. Electric Coupling Apertures. The aperture is located
between the cavity and input waveguide such that
the electric ﬁeld in the waveguide is normal to the
electric ﬁeld in the cavity. A narrow slot in the wall
separating the waveguide and cavity can provide
electric coupling.
6.2. Coupling through Probes
One popular approach used to transfer energy from a co
axial line to a waveguide is by electric probes. In a typical
conﬁguration, the axis of the coaxial line is perpendicular
to the broadside of the rectangular waveguide. The center
conductor of the coaxial line protrudes through the wave
guide wall and extends into the waveguide. The outer con
ductor of the coaxial line is terminated at the waveguide
wall. The electric ﬁelds from the end of the center conduc
tor terminate on the other broadside wall parallel to the
dominant E ﬁeld of the waveguide. They are, therefore,
called electric probes. If the probe is shaped to form a cir
cular loop and the end of the probe is terminated on the
broad wall of the waveguide, the current ﬂow through this
loop will induce a magnetic ﬁeld parallel to the dominant
H ﬁeld of the waveguide. In this case, the probe is a mag
netic current loop.
In an electric probe, the center conductor of the
coaxial line forms a radiating antenna. Depending on
the length or depth of this section, the input impedance
at the interface can be inductive or capacitive. For opti
mum performance, the antenna should present a matched
load at the interface. The probe excites waveguide modes
that propagate in both directions; therefore, the energy is
divided equally in both directions. In order to redirect the
energy in the preferred direction, the other side is termi
nated in a short circuit. In the case of rectangular wave
guide cavities, the placement of the probe is determined
from one of the shortcircuited ends. Invariably, a tunable
short circuit will be required for optimum transfer of pow
er. The probe can be constructed with various lengths and
diameters. The distance between the probe and the short
circuit is determined experimentally.
The bandwidth of the probe can be improved by pro
viding a broadband match at the interface. This can be
achieved by changing the length and diameter of the
probe. Other approaches include making the end round,
attaching a metal sphere at the end, or ﬂaring the center
conductor. If direct current (DC) return is desired, the
probe can be terminated on the other broadwall, or it can
rest on a crossbar across the waveguide broad dimension.
Sometimes the probe is extended through the opposite
side of the waveguide to form another section of shorted
coaxial line. The position of the short circuit in this case
provides an additional variable.
The input impedance of a shortdiameter coaxial an
tenna is given by
Z
in
ﬃl
2
cos
2
px
0
a
sin
2
2px
1
l
g
÷jX (57)
where l is length of the probe, x
0
is the distance from the
center of the waveguide, and x
1
is the distance from the
probe to the short circuit. The value of the reactance is
large, implying that the input impedance has a large ca
pacitive component.
6.3. Coupling Holes in Waveguides
The coupling from a waveguide to a cavity can be provided
by apertures consisting of holes and slits. The aperture
can be inﬁnitesimally thin or with ﬁnite thickness. The
insertion loss caused by a hole of ﬁnite thickness t is given
by
a
T
=a
b
÷a
t
(58)
where a
b
is the attenuation resulting from the susceptance
of the hole and a
t
is the attenuation in the below cutoff
waveguide hole.
6.3.1. Holes in a Rectangular Waveguide. For a hole of
diameter d in a rectangular waveguide normal to the di
rection of propagation, the normalized susceptance B is
given by
B
Y
0
ﬃ
3
2p
abl
g
d
3
(59)
where Y
0
is the characteristic admittance of the dominant
waveguide mode and l
g
is the guide wavelength of a
Coaxial
line
Coaxial
line
Cavity
(a) (b)
(c)
Waveguide Aperture
Figure 19. Cavity excitation using (a) loop coupling, (b) electric
probe coupling, and (c) aperture coupling.
CAVITY RESONATORS 587
waveguide having broadside dimension a and smaller
dimension b.
The attenuation a
b
resulting from the hole is
a
b
=20 log
B
2Y
0
(60)
and the attenuation resulting from the ﬁnite thickness
a
t
is
a
t
=32
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
1 ÷ 1:706
d
l
_ _
2
¸
t
d
ﬃ32
t
d
dB (61)
where l is the operating wavelength.
6.3.2. Holes in a Circular Waveguide. For a hole of di
ameter d normal to the direction of propagation in a cir
cular waveguide of diameter 2a, the normalized
susceptance B is given by
B
Y
0
=
l
g
4a
5:71
a
3
d
3
÷2:344
_ _
(62)
where Y
0
is the characteristic admittance of the dominant
waveguide mode, and l
g
is the guide wavelength of the
dominant mode.
The attenuation a
b
resulting from the hole is
a
b
=10 log
(B=Y
0
)
2
4
÷1
_ _
dB (63)
and the attenuation a
t
resulting from the ﬁnite thickness
is given by Eq. (61), and the total attenuation is calculated
using Eq. (58).
7. RESONATOR MEASUREMENTS
As described earlier, the resonator is described fully in
terms of the resonant frequency f
0
, the coupling coefﬁ
cient, and the quality factors. The unloaded quality factor
Q
u
, loaded quality factor Q
L
, and external quality factor
Q
e
are useful in various circuit analyses containing mi
crowave cavities.
Experimental determination of the parameters is
straightforward using modern microwave network ana
lyzers. In Fig. 20, singleport and twoport cavity mea
surement setups are shown. The magnitude and phase of
the reﬂection and transmission coefﬁcients are measured
to determine the resonator parameters.
7.1. SinglePort Resonator
The equivalent circuit of a singleport cavity resonator is
shown in Fig. 1, where R, L, and C are the equivalent
lumped resistance, inductance, and capacitance. The
equivalent parallel and series circuits of Figs. 1a and 1b
are also known as the detuned short and open conﬁgura
tions, respectively. The equivalence between series and
shunt parameters of these resonant circuits is as shown in
the following table:
Parameter Series Tuned Parallel Tuned
f
0
1
ﬃﬃﬃﬃﬃﬃﬃ
LC
_
1
ﬃﬃﬃﬃﬃﬃﬃ
LC
_
Q
u
oL
R
R
oL
b
Z
0
R
R
Z
0
Q
L
Q
u
1 ÷b
Q
u
1÷b
The input impedance of the circuit in Fig. 1a can be
rewritten as
Z
in
=
R
1 ÷j2Q
u
d
(64)
where d =(o÷o
0
)/o
0
represents the frequency detuning
parameter [12]. By varying d, the locus of the impedance
given by Eq. (64) is determined. On a Smith chart, a
circular locus, as shown in Fig. 21 is obtained depending
on the coupling coefﬁcient. For circle A, R=Z
0
and the
locus passes through the origin. This condition is
called critical coupling and corresponds to b =1, implying
that it provides a perfect match to the transmission
line at resonance. The circle B with RoZ
0
is called
undercoupled condition and bo1. Finally, the circle
C with R4Z
0
is an overcoupled condition with
b41 [13].
The coupling coefﬁcient for any cavity is calculated
using the measurement of reﬂection coefﬁcient S
11;o
0
at
resonance. For the undercoupled case, we obtain
b=
1 ÷S
11;o
0
1 ÷S
11;o
0
(65)
Network
analyzer
(a) (b)
Network
analyzer
Sparameter
test setup
Sparameter
test setup
Signal
generator
Signal
generator
DUT DUT
Figure 20. Measurement setup for (a) reﬂection and (b) trans
mission resonator.
588 CAVITY RESONATORS
and for the overcoupled case, we obtain
b =
1 ÷S
11;o
0
1 ÷S
11;o
0
(66)
The intersection of the impedance locus with
the real axis provides the value of b as shown in
Fig. 22.
Other quality factors can be determined from Eq. (64),
which can be rewritten as
"
ZZ
in
=
Z
in
Z
0
=
b
1 ÷j2Q
u
d
=
b
1 ÷j2Q
L
(1 ÷b)
=
b
1÷j2Q
e
b
(67)
The Q
u
, Q
L
, and Q
e
are related as
Q
u
=Q
L
(1 ÷b) =Q
e
b (68)
The normalized frequency deviations for unloaded, loaded,
and external quality factors are given by
d
u
= ±
1
2Q
u
; d
L
= ±
1
2Q
L
; d
e
= ±
1
2Q
e
(69)
From Eqs. (69) and (67), the impedance locus of Q
u
is de
termined and is given by
(Z
in
)
u
=
b
1 ±j
(70)
Equation (70) represents the points on the impedance
locus where the real and imaginary parts of the imped
ance are the same. Figure 22 represents the locus of these
points (corresponding to R=X) for all possible values of b.
This locus is an arc whose center is at Z=07j, and the
radius is the distance to the point 07j. The intersection of
this arc with the impedance locus determines the Q
u
mea
surement points:
Q
u
=
f
0
f
1
÷f
2
(71)
The frequencies f
1
and f
2
are called halfpower points
because these points correspond to R=X on the imped
ance locus. The loaded and external Q values can be de
termined in a similar way. Equations (67) and (69), the
impedances corresponding to Q
e
and Q
L
, are given by
(Z
in
)
e
=
b
1 ±jb
(72)
and
(Z
in
)
L
=
b
1 ±j(1÷b)
(73)
By using Eqs. (72) and (73), the Q
e
and Q
L
loci are easily
determined. These loci are shown in Fig. 22.
7.2. TwoPort Resonator
The equivalent circuit of a twoport cavity resonator is
shown in Fig. 23. In this case, the input and output cou
pling are represented as b
1
and b
2
. They are determined
from
b
1
=
Y
01
n
2
1
G
and b
2
=
Y
02
n
2
2
G
(74)
where Y
01
and Y
02
are the admittances seen at the input
and output ports. The coupling coefﬁcients are directly
determined by measuring the VSWR at the input and out
put ports with the other port opencircuited.
The transmission response of such a resonant circuit
measured using the setup of Fig. 20, is shown in Fig. 23.
The coupling coefﬁcients and the quality factors for two
port resonators determined from the measurement of
the insertion loss T at resonant frequency and the 3 dB
Impedance
locus
o
3
o
5
o
1
o
2
o
6 o
4
r = x
R = 1
=
1
Smith chart
B = G + 1
R = 0 R = ∞
[ [
Figure 22. Determination of b and the halfpower points from
the Smith chart. Q
0
locus is given by X=R(B=G); Q
L
by X=
R÷1; Q
/
ext
by X=1.
Undercoupled
Overcoupled
R = 1
R = 0 R = ∞
B
A
C
Nearly
critical
Figure 21. Input impedance of a singleport resonant cavity on
the Smith chart for 3 degrees of coupling.
CAVITY RESONATORS 589
bandwidth Df using the following wellknown relations
[14]
T=
2
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
b
1
b
2
_
1÷b
1
÷b
2
(75)
Q
L
=
f
0
Df
(76)
Q
u
=Q
L
(1 ÷b
1
÷b
2
) (77)
8. APPLICATIONS
A development of the cavity resonators was an important
milestone in microwave technology. Early work on cavity
resonators focused on cavities of regular shapes. But the
development of microwave oscillators and ampliﬁers re
quired complex shapes to achieve the performance re
quired in the development of klystrons, magnetrons,
and travelingwave tubes. Some of those shapes were
covered in this article to illustrate the fact that different
cavity structures are required to achieve the desired
results.
Cavity resonators are also used extensively to measure
frequency or wavelength. Tunable cavities are made to
resonate at different frequencies by varying size and then
calibrating size against frequency.
Cavity resonators are now most widely used to develop
ﬁlters. Depending on the characteristics of the cavity, it
can be used for narrowband as well as wideband ﬁlters.
The applications of cavity resonators are concomitant with
those of ﬁlters. In that sense, cavity resonators have
applications in lowpass, bandpass, bandstop, and high
pass ﬁlters. They are also used in diplexers, multiplexers,
and directional ﬁlters.
In the following sections, the preceding applications
will be reviewed and their key aspects will be highlighted.
Additional information can be found in other relevant
articles of this encyclopedia.
8.1. Applications in Microwave Tubes
There were many problems in the early development of
microwave valves that were caused by circuit elements
and their interconnections. The development of resonant
cavities led to the invention of klystron. The cavity reso
nators were able to reduce the transit time. The capaci
tance between the cathode and grid was used to resonate
with the low inductance provided by the cavity.
In klystron ampliﬁers, multiple cavities are used to al
low bunching of electrons. Because the electromagnetic
ﬁelds in a cavity are changing as a function of time, the
alternating electric ﬁelds at the grid cause bunching of
electrons. By using another cavity at an optimum dis
tance, the electrons are further bunched to build up oscil
lations. The ﬁrst ‘‘buncher’’ resonator is excited into
resonance through external means, and the second ‘‘catch
er’’ resonator takes out the power. In klystron ampliﬁers,
internal feedback is also provided via openings in the
cavities.
In the reﬂex klystron, the electron beam is bunched by
passing through a single resonator. The reﬂector returns
the electron beam to this cavity at an optimum bunched
condition. At this time, the energy is extracted from the
cavity.
Magnetrons use various shapes of cavities to build os
cillations and power. The power is extracted from one of
the resonators through a coupling loop or an iris.
In traveling wave tubes, cavities are used as part of the
slow wave structure. For additional information, refer to
the appropriate article in this encyclopedia.
8.2. Filters
In order to use cavity resonators in ﬁlter applications, a
reactance or susceptance slope parameter is generally re
quired. The reactance slope parameter for a series reso
nant structure is deﬁned as
w =
o
0
2
dX
do
¸
¸
¸
¸
o=o
0
ohms (78)
Similarly, the susceptance slope parameter for the parallel
resonant structure is deﬁned as
B=
o
0
2
dG
do
¸
¸
¸
¸
o=o
0
siemens (79)
From Eqs. (78) and (79), it is straightforward to see that
for a series resonant circuit at resonance
w =o
0
L=
1
o
0
C
(80)
I
n
s
e
r
t
i
o
n
l
o
s
s
3 dB
Frequency
T
Q L Y
01
Y
02
n
2
n
1
C
(b)
(a)
∆f
[
1
[
2
Figure 23. (a) Equivalent circuit of a twoport resonator with
input and output transformers; (b) transmission response of a
twoport resonator.
590 CAVITY RESONATORS
and
Q=
w
R
(81)
Similarly, for a parallel resonant circuit at resonance
B=o
0
C=
1
o
0
L
(82)
and
Q=
B
G
(83)
In a bandpass ﬁlter design, impedance or admittance
inverters can be used with series or shunttype resonant
structures. The reactance or admittance slope parameter
is related to ﬁlter prototype element values. The external
Q and coupling coefﬁcients are also expressed in terms of
the reactance or admittance slope parameter. Inductive
posts or irises with impedance inverters can be used to
construct a bandpass ﬁlter. For details on waveguide ﬁl
ters, refer to the appropriate section in this encyclopedia.
8.3. Frequency Measurement
Both coaxial and waveguide resonators have been used in
commercially available wavemeters. The main require
ment in selecting cavity dimensions is to ensure that the
cavity resonates in the fundamental mode, that there are
no degenerate modes, and that they are easy to manufac
ture and calibrate.
These wavemeters are basically tunable cavities, and
when the length is l
g
/2, the cavity resonates by taking in
some energy from the transmission line, coaxial, or wave
guide. This action will produce a dip in the transmitted
power. When the length of the cavity is calibrated, the
frequency or wavelength can be read off directly from the
dial. The extent of the dip depends on the amount of cou
pling. The Q of wavemeter cavities is quite high, on the
order of 5000–10,000 depending on the desired accuracy.
In a coaxialline wavemeter, as shown in Fig. 24, the
center conductor is used as a probe to couple energy to the
resonator. Noncontacting plungers with chokes are used
to provide a variable short position. The coaxial resonator
will resonate whenever the cavity length is a halfwave
length. Measuring the change in plunger positions be
tween two successive minima and multiplying by 2 will
give the wavelength of operation. Because of the large
surface area of the coaxial outer wall, the cavity Q is not
very high.
Cylindrical cavities are generally used for wavemeters,
as shown in Fig. 25. The currents in the TM
01
mode ﬂow
circumferential to the cavity cross section. Therefore, the
shortcircuiting plunger does not need to have a good con
tact and provides easy manufacturability. Furthermore,
other higherorder modes that require current ﬂow in the
endplates are not supported. In order to prevent other
modes from being excited, two coupling holes in the side
wall of a waveguide, which are a halfwavelength apart,
are used. The bandwidth of the coupling structures can be
increased by selecting an elongated hole.
BIBLIOGRAPHY
1. R. E. Collin, Foundations of Microwave Engineering, Mc
GrawHill, New York, 1966, Chap. 7.
2. S. Ramo, J. Whinnery, and T. Van Duzer, Fields and Waves in
Communication Electronics, Wiley, NewYork, 1965, Chap. 10.
3. R. N. Ghose, Microwave Circuit Theory and Analysis,
McGrawHill, New York, 1963, Chap. 8.
4. T. Moreno, Microwave Transmission Design Data, Dover,
New York, 1958, Chap. 18.
5. H. J. Reich et al., Microwave Principles, Van Nostrand,
Princeton, NJ, 1960, Chap. 7.
6. J. G. Kretzschmar, Wave propagation in hollow conducting
elliptic waveguide, IEEE Trans. Microwave Theory Tech.
MTT18:547–554 (1970).
7. J. G. Kretzschmar, Mode chart for elliptical resonant cavities,
Electron. Lett. 6:432–433 (1970).
8. W. Bra¨ ckelmann, Die Grenzfrequenzen von ho¨heren Wel
lentypen im Koaxial Kabel mit elliptischem Querschnitt,
Arch. Elek. U
¨
bertragung. 21:421–426 (1967).
9. N. W. McLachlan, Theory and Application of Mathieu Func
tions, Clarendon Press, Oxford, UK, 1947.
10. W. W. Hansen and R. D. Richtmyer, On resonators suitable for
klystron oscillators, J. Appl. Phys. 10:189–199 (1939).
Resonant cavity
Waveguide
Probe coupling
Plunger with choked head
Scale
Adjustable
head
Figure 24. Coaxial cavity wavemeter.
Waveguide
Movable plunger
(b)
(a)
Figure 25. Cylindrical cavity wavemeter: (a) end view; (b) top
view.
CAVITY RESONATORS 591
11. F. L. Ng, Tabulation of methods for the numerical solution
of the hollow waveguide problem, IEEE Trans. Microwave
Theory Tech. MTT22:322–329 (1974).
12. E. L. Ginzton, Microwave Measurements, McGrawHill, New
York, 1957, Chap. 9.
13. I. Bahl and P. Bhartia, Microwave Solid State Circuit Design,
Wiley, New York, 1988, Chap. 3.
14. M. Sucher and J. Fox, Handbook of Microwave Measurements,
3rd ed., Vol. 2, Wiley, New York, 1963, Chap. 8.
FURTHER READING
B. Lax and K. J. Button, Microwave Ferrites and Ferromagnetics,
McGrawHill, New York, 1962, pp. 145–196.
S. Y. Liao, Microwave Devices and Circuits, PrenticeHall, Engle
wood Cliffs, NJ, 1980, Chap. 4.
P. K. Mariner, Introduction to Microwave Practice, Academic
Press, New York, 1961, Chap. 7.
G. L. Matthaei, L. Young, and E. M. T. Jones, Microwave Filters,
Impedance Matching Networks and Coupling Structures, Mc
GrawHill, New York, 1964.
S. R. Rengarajan and J. E. Lewis, Quality factor of elliptical cy
lindrical resonant cavities, J. Microwave Power 15:53–57
(1980).
A. K. Sharma, Spectral domain analysis of elliptic microstrip ring
resonator, IEEE Trans. Microwave Theory Tech. MTT32:
212–218 (1984).
A. K. Sharma and B. Bhat, Spectral domain analysis of elliptic
microstrip disk resonators, IEEE Trans. Microwave Theory
Tech. MTT28:573–576 (1980).
CELLULAR RADIO
WILLIAM C. Y. LEE
AirTouch Communication
The cellular radio system is sometimes called a mobile
phone system or a car phone system. Due to the daily needs
of subscribers, cellular systems have expanded consider
ably all over the world. This article discusses the history of
cellular systems and the difﬁculty of deploying them in the
mobile radio environment, elaborating on employing dig
ital cellular systems, Personal Communication Services
(PCS) mobile satellite systems, and the future IMT2000
system.
1. HISTORY OF CELLULAR RADIO SYSTEMS
1.1. Analog System
1.1.1. Startup Period (1964–1987). In 1964, AT&T Bell
Labs actively developed a highcapacity mobile radio
phone system called Advanced Mobile Phone Service
(AMPS) [1], which is an analog frequency modulation
(FM) system. The system consists of many socalled cells.
Each cell has one or multiple transceivers. Because of the
cell formation, the system is referred to as a cellular sys
tem. In the analog AMPS system, mobile units are com
patible with all the cellular systems operating in the
United States, Canada, and Mexico. A spectrum of
50 MHz (limited to 825–849MHz for mobile transmissions
and 869–896MHz for basestation transmissions) is
shared by two cellular system providers in each market
(city). Each one provider operates over a bandwidth of
25 MHz in a duplex fashion (using 12.5 MHz in each di
rection between cell sites and mobile units). There are 416
channels, comprising 21 setup channels and 395 voice
channels. The channel bandwidth is 30 kHz. Mobile cellu
lar telecommunications systems [2] have two unique
features:
1. First, they invoke the concept of frequency reuse for
increasing spectrum efﬁciency. The same set of fre
quency channels can be assigned to many cells.
These cells are called cochannel cells. The separa
tion between two cochannel cells is engineered by
the D/R ratio (see Fig. 1), where D is the cochannel
cell separation and R is the cell radius. A 4mi cell
implies R=4 mi. The D/R ratio is characteristic of a
cellular system. If the D/R ratio is high, the voice
quality is improved by reducing the system’s user
capacity.
2. A second feature, handing off communications from
one frequency to another, occurs when a mobile unit
enters a new cell. The scheme is called a handoff in
North America and a handover in Europe. The sys
tem handles this operation automatically, and the
users do not need to intervene. A good handoff
1
D
R
D = 4.6R
1
1 5
7
2
1
1
1 6
1
4 3
Figure 1. Hexagonal coils in an AMPS system: R=radius of
cells, D=minimum separation of cochannel cells, q =D/R=4.6,
K=number of cells in a cluster =7. Clusters are indicated, and
the six cells that effectively interfere with cell 1 are numbered 2
through 7. The shaded cells are cochannel cells.
592 CELLULAR RADIO
algorithm can reduce both the call drop rates and
interference. In general, there are two kinds of han
doffs: (1) soft handoffs, which implies making a new
connection before breaking the old one; and (2) hard
handoffs, which involves breaking the old connection
before making the new one.
The ﬁrst installation of a cellular system occurred in
Tokyo in 1979, using a minor modiﬁcation of AMPS. The
ﬁrst AMPS cellular system installed in the United States
took place in Chicago in 1983. Analog cellular systems are
in use over most of the world, employing different versions
of AMPS: in Japan, the Nippon Telephone and Telegraph
(NTT) AMPS system; in the UK, the Total Access Com
munications System (TACS); and in northern Europe, the
Nordic Mobile Telephone (NMT). The major difference is
their reduced channel bandwidths of 25 kHz instead of
30 kHz as in AMPS.
1.1.2. Mature Period (1987–1992). From 1987 to
1992, the 90 MSA (metropolitan statistical area) markets,
as well as most of the 417 RSA (rural service area) mar
kets, had cellular operations in the United States. The
number of subscribers reached 1 million. The cell split
(reducing the size of cells) technique and dynamic
frequency assignment were applied to increase the user
capacity.
When the cell radius R is less than half a kilometer, the
cell is called a microcell. In such small cells it is harder to
reduce the socalled cochannel interference in order to in
crease capacity, requiring special technological approach
es called microcell technology. The world was also
becoming more aware of the potential future markets.
Suddenly, ﬁnding the means to increase capacity became
urgent.
1.2. Digital System
1.2.1. Introduction Period. In 1987, the capacity of the
AMPS cellular system started to show its limitations. The
growth rate of cellular subscribers far exceeded expecta
tions. In 1987, the Cellular Telecommunication Industry
Association (CTIA) formed a Subcommittee for Advanced
Radio Technology to study the use of a digital cellular sys
tem [3] to increase capacity. At that time, the Federal
Communications Commission (FCC) had clearly stated
that no additional spectrum would be allocated to cellu
lar telecommunications in the foreseeable future. There
fore, the existing analog and forthcoming digital systems
would have to share the same frequency band. In Decem
ber 1989, a group formed by the Telecommunication In
dustry Association (TIA) completed a draft of a digital
cellular standard.
The digital AMPS, which must share the existing spec
trum with the analog AMPS, is a duplex timedivision
multipleaccess (TDMA) system. The channel bandwidth
is 30 kHz. There are 50 TDMA frames per second in each
channel. Three or six timeslots per frame can serve three
calls or six calls at the same time in one channel. The
speech coding rate is 8 kbps (kilobits) per second. An
equalizer is needed in the receiver to reduce the inter
symbol interference that is due to the spread in time delay
caused by the dispersed time arrival of multipath waves.
The North American TDMA system was ﬁrst called IS54
by the TIA. Later, the system was modiﬁed and renamed
IS136.
During this period, not all mobile telephone systems in
Europe were compatible. A mobile phone unit working in
one country could not operate in another country. In 1983,
in response to the need for compatibility, a special task
force, the Special Mobile Group [4], was formed among
European countries to develop a digital cellular system
called GSM (group of special mobile systems) in 1994, then
renamed to stand for global system for mobile communi
cations. The operating principles of the GSM system re
semble those of the AMPS in radio operation, but the
system parameters are different; this will be described
later.
In the United States, in addition to the TDMA being
considered above, another particularly promising tech
nology is codedivision multiple access (CDMA) [3]. It
is a spreadspectrum technique with a bandwidth of
1.25 MHz. The maximum number of trafﬁc channels is
55. This CDMA system is called IS95 or cdmaone.
There are three mobile data systems in the United
States: Ardis, operated by IBM/Motorola; Ram, operated
by Ericsson; and CDPD (Cellular Digital Packet Data)
system. The transmission rates for all data systems are
around 8 kbps. Only CDPD operates in the cellular spec
trum band.
1.2.2. The Future. Starting in 1996, the socalled PCS
systems were deployed. They were cellularlike systems,
but operated in the 1.8GHz band in Europe and the
1.9 GHz range in North America. In Europe, the socalled
DCS1800 PCS systems were endorsed, which are based
on the GSM system. In the United States, the PCS had
three versions: DCS1800 (a GSM version), TDMA1900
(IS136 version), and CDMA1900 (IS95 version). The
PCS could have six operational licenses (A, B, C, D, E, F)
in each city. Therefore, more competitors would be in the
mobile phone services business.
In addition, the mobile satellite systems that use the
LEO concept (lowEarth orbit) were deployed. Iridium
(66 satellites) and Globalstar (48 satellites) were launched
at 900km and 400km altitudes, respectively. These sys
tems can integrate with cellular systems and enhance cel
lular coverage domestically and roam internationally as a
global system. Other LEO systems are also in the devel
opment stage. There is a special LEO system called
Teledesic that will be operating at 26 GHz with 840 satel
lites in orbit. This system is used for wideband data and
video channels to serve subscribers in a high capacity
network.
A future cellular system, called the International Mo
bile Telephone (IMT2000) system, is now in the planning
stage. A universal cellular standard (or PCS) system with
high capacity and high transmission rate was realized by
the year 2002.
CELLULAR RADIO 593
2. MOBILE RADIO ENVIRONMENT: A DIFFICULT
ENVIRONMENT FOR CELLULAR RADIO SYSTEMS
2.1. Understanding the Mobile Radio Environment
2.1.1. The Limitations of Nature. In the mobile radio en
vironment, there are many attributes that limit the sys
tem performance for wireless communication. In the past,
there were attempts to adapt digital equipment such as
data modems and fax (facsimile) machines used for wire
line to cellular systems. The data engineers at that time
only realized the blanking and burst interruption in the
voice channel as a unique feature of handoffs and power
control. They modiﬁed data signaling by overcoming the
impairments caused by blanking and burst signaling in
terruption. This modiﬁed data modem did not work as ex
pected in the cellular system. Actually, the blanking and
burst interruption scheme was not the sole cause of the
inadequate data transmission and would have been rela
tively easy to handle. But without entirely understanding
the impairments, the unexpected poor performance could
not be offset by merely overcoming the blanking and burst
signaling impairment.
2.1.2. Choosing the Right Technologies. In designing ra
diocommunication systems, there are many different tech
nologies, and among them no single technology is superior
to the others. Choosing a technology depends on real con
ditions in the environment of a particular communication.
In satellite communication or microwave link transmis
sion, the radio environments are different from that of the
mobile radio environment. There are many good technol
ogies that work in satellite communication and microwave
link transmission, but they may not be suitable for the
terrestrial mobile radio environment. Therefore, choosing
the right technology must depend on the transmission en
vironment.
2.2. Description of the Mobile Radio Environment
The mobile radio environment is one of the most complex
ones among the various communication environments.
2.2.1. Nature Terrain Conﬁguration. Because the anten
na height of a mobile unit or a portable unit is very close to
the ground, the groundreﬂected wave affects the recep
tion of the signal from the transmitting site via the direct
path. The freespace loss is 20 dB/dec (dec stands for de
cade, a period of 10) or, in other words, it is inversely pro
portional to the distance d
÷2
. However, in the mobile
radio environment, due to the existing groundreﬂected
wave and the small incident angle y, as shown in Fig. 2,
the total energy of the groundreﬂected wave is reﬂected
back to space. Due to the nature of electromagnetic waves,
when the wave hits the ground, the phase of the wave
changes by 1801. Therefore, at the mobile, the direct wave
and reﬂected wave cancel each other instead of adding
constructively. As a result, the signal that is received be
comes very weak. A simple explanation is as follows. If the
pathlength of the direct wave is d, the pathlength of
the reﬂected wave is d÷Dd. Then the received power
of the two combined waves is proportional to d
÷4
as
demonstrated below
P
r
o
1
d
÷
1
d÷Dd
_ _
2
=
Dd
d(d÷Dd)
_ _
2
=
(Dd)
2
d
4
(1)
where Dd is assumed to be much less than d and Dd is a
function of the antenna height h
1
at the base station.
From Eq. (1), the mobile radio path loss follows the inverse
fourthpower rule or 40 dB/dec, and the antenna height
gain follows the secondpower rule or 6dB/oct. In the mo
bile radio environment, the average signal strength at the
mobile unit varies due to the effective antenna height h
e
at
the base station measured from the mobile unit location.
Since the mobile unit is traveling, the effective antenna
height is always changing as a function of terrain undu
lations, and so is the average signal strength. This phe
nomenon is shown in Fig. 3.
This twowave (direct wave and groundreﬂected wave)
model is only used to explain the propagation loss of 40 dB/
dec in the mobile radio environment, not the multipath
fading.
2.2.2. Humanmade Effects
2.2.2.1. Humanmade Communities. These can be classi
ﬁed as metropolitan areas, urban areas, suburban areas,
open areas, and so on. The distribution of buildings and
homes depends on the population size. The reception of
the signal is affected by the differences in humanmade
communities and results in different propagation path
loss.
2.2.2.2. Humanmade Structures. Different geographic
areas use different construction materials, different types
of construction frames, and different sizes of buildings.
Cities such as Los Angeles, San Francisco, and
Tokyo are in earthquake zones and follow earthquake
construction codes. The signal reception in those cities is
different from that in others. Humanmade structures will
affect the propagation path loss and multipath fading due
Mobile
station
antenna
Basestation antenna
D
ir
e
c
t w
a
v
e
G
r
o
u
n
d

r
e
f
l
e
c
t
e
d
w
a
v
e
h
2
h
1
Figure 2. Twowave propagation model.
h
e
c
h
e
b
h
e
a
A
B
C
h
1
Figure 3. Effective antenna heights at base station based on dif
ferent locations of mobile stations.
594 CELLULAR RADIO
to reﬂection and the signal penetration through the
buildings.
2.2.2.3. Humanmade Noise. This can be classiﬁed into
two categories: industrial noise or automotive ignition
noise. The high spikes in automotive ignition or in ma
chines are like impulses in the time domain; their power
spectrum density will cover a wide spectrum in the indus
trial frequency domain. At 800MHz, automotive ignition
noise is determined by the number of vehicles. For a trafﬁc
volume from 100 cars/h to 1000 cars/h, the noise ﬁgure
increases 7dB. As the application of ultrahighfrequency
(UHF) devices and microwave systems increase, so does
the noise pollution for cellular systems. As we will men
tion later, a communication system is designed to main
tain the minimum required carriertointerference ratio
(C/I
s
). The interference I may, under certain circumstanc
es, be included in noise the N. If the interference level is
higher, the level of the carrier, C should also be higher in
order to meet the (C/I)
s
requirement. This means that
when the humanmade noise level is high, either the trans
mission power at the base station should be increased or
the cell size must be reduced.
2.2.3. Moving Medium. If the mobile unit is in motion,
the resulting signal from multipath waves at one location
is not the same at another; thus the mobile receiver ob
serves an instantaneous ﬂuctuation in amplitude and
phase. The amplitude change is called Rayleigh phase,
and the phase change is a uniformly distributed process,
or random FM in FM systems. The signal fading can be
fast or slow depending on the speed of the vehicle. When
the vehicle speed is slow, the average duration of fading is
long. This average fading duration can be, for example,
7 ms at ÷10 dB below the average level when the vehicle
speed is 24 km/h at a propagation frequency of 850 MHz.
In an analog system, a fade duration of 7 ms does not affect
the analog voice; the ear cannot detect these short fades.
However, the fade duration of 7 ms is long enough to cor
rupt the digital (voice and data) transmissions. At a trans
mission rate of 20 kbps, 140bits will sink in the fade.
Furthermore, the vehicle speed of all the users is not con
stant, and the use of interleaving and channel coding to
protect the information bits is very difﬁcult. Furthermore,
voice communication is operating in real time unlike data
transmission which can be in any timedelay fashion.
Many schemes used by data communication cannot be
used for digital voice communication.
2.2.4. Dispersive Medium. Because of humanmade
structures, the medium becomes dispersive. In a disper
sive medium, two phenomena occur. One is time delay
spread and the other is selective fading. The time delay
spread is caused by a signal transmission from the base
station reﬂected from different scatterers and arriving at
the mobile unit at different times. In urban areas, the
mean timedelay spread D is typically 3ms; in suburban
areas, D is typically 0.5 ms. In an open area, D is typically
0.2 ms, and in an inbuilding ﬂoor, D is around 0.1ms or
less. These time delay spreads do not affect the analog
signal because the ear cannot detect the short delay
spread. However, in a digital system when a symbol (bit)
is sent, many echoes arrive at the receiver at different
times. If the next symbol is sent out before the ﬁrst one
dies down, intersymbol interference occurs. The dispersive
medium also causes frequency selective fading (Fig. 4).
The selective fading will not harm the moving receiver
because when the mobile unit is moving, only the average
power is considered. Then, in order to make a mobile
phone call when the mobile unit is at a standstill, it usu
ally requires that all the signal strengths from four fre
quencies have two strong setup channels and two strong
voice channels. A pair of frequencies is formed by a chan
nel carrying a call on both a forward link and a reverse
link. When the mobile unit is moving, the average power
of the four frequencies is the same. Then we base our
quality estimates on one (C/I)
s
value. But when the mobile
unit is still, the signals of four frequencies at one location
are different due to frequency selective fading. Unless all
four frequencies are above the acceptable threshold level,
the call cannot be connected.
2.3. Concept of C=I
In designing highcapacity wireless systems, the most im
portant parameter is the carriertointerference ratio
(C/I). The C/I ratio and the D/R ratio are directly halted.
The D/R ratio is determined by the C/I ratio. Usually, with
a given received signal level C, the lower the interference
level, the higher the C/I ratio and hence the quality
improves. There is a specific C/I level, namely (C/I)
s
,
that the system design criterion is based on. We may de
rive the relationship between C/I and D/R as follows. As
sume that the ﬁrst tier of six cochannel interference cells
is the major cause for the interference I. Based on the
40 dB/dec propagation rule, we obtain
C
I
=
C
6
i =1
I
i

C
6
.
I
i
=
R
÷4
6
.
D
÷4
=
(D=R)
4
6
(2)
A general equation of the cochannel interference reduc
tion factor q can be expressed, from Eq. (2), as
q =(D=R)
s
=(6(C=I)
s
)
1=4
(3)
where (C/I)
s
is obtained from a subjective test correspond
ing to the required voice (or data) quality level, as mentioned
t
1
∆
Mean ∆ =
s(t )
in
dB
0.2
0.5
3.0
Open area
0.1 Inbuilding
Suburban
Urban
Time (at
receiving
end)
js
js
js
js
Figure 4. Time delay spread D at the receiving end when trans
mitting one bit in a dispersive medium.
CELLULAR RADIO 595
previously. Equation (3) is plotted in Fig. 5. The (C/I)
s
ratio
is chosen according to either the required voice or data
channel quality.
2.4. The Predicted Signal Strength Models
Since the (C/I)
s
is a system design parameter, system
planning engineers would like to use an effective model
to predict both C and I in a given area. There are two dif
ferent prediction models. One predicts the average signal
strength along the radio path based on the path loss slope.
The Okumura and Hata models [5,6] represent these
types of models. The other predicts the local mean signal
strength along a particular mobile path (street or road)
based on the particular terrain contour. Lee’s model [7]
represents these types of models.
3. REASONS FOR DIGITAL CELLULAR
3.1. Compatibility in Europe
Again, due to the lack of a standard mobile radio system in
Europe during the early 1980s, the mobile phone unit used
in each country could not be used in other countries.
Starting in 1982, ETSI (European Telecomms Standard
Institute) formed a group called the Group of Special Mo
bile to construct an international mobile radio system
called GSM for Europe. The system chosen was to be a
digital system using TDMA for the access scheme. The
GSM advanced intelligent network (AIN) was adopted
from the wireline telephone network. GSM was the ﬁrst
digital mobile phone system in the world.
3.2. Capacity in North America
The frequency spectrum is a very limited resource com
monly shared by all wireless communications. Among the
wireless communications systems, cellular is the most spec
tral efﬁcient system where the socalled spectral efﬁciency
is related to the number of trafﬁc channels per cell. From
this number, we can derive the erlang/cell ratio, which
translates to erlangs/km
2
, or the number of trafﬁc chan
nels/km
2
based on the trafﬁc model and the size of the cells.
However, a spectrum of only 50 MHz has been allocated
to cellular operators in the United States. Furthermore,
since two operators are licensed in each market, the spec
trum of 50 MHz must be split in two. Therefore, system
trunking efﬁciency is reduced and interference caused by
an operator in one market often contaminates the other
operator’s allocated spectrum. Furthermore, manufactur
ing companies were always considering lowering the cost
of cellular units and increasing sales volume. As a result,
the speciﬁcation of cellular units could not be kept tight,
and thus more interference prevailed. Once interference
increased and could not be controlled by cellular opera
tors, both voice quality and system capacity decreased.
In 1987, the top 10 U.S. markets were already feeling
the constraints of channel capacity; they would not be able
to meet the market demand in the future. The solution for
this increasing need for high capacity was to go digital
[1–3]. Going digital was the best solution because of the
nature of the digital waveform. If system compatibility is
not an issue, the top 10 U.S. markets might and could go
digital by themselves. However, for the sake of compat
ability the United States needed one standard for the
entire North American cellular industry.
3.3. The Advantages of a Digital System
Digital systems offer the following advantages:
1. The digital waveform is discrete in nature. There
fore, the digital waveform can be regrouped easily
for transmission needs.
2. Digital transmission is less susceptible to noise and
interference.
3. Digital modulation can conﬁne the transmitted en
ergy within the channel bandwidth.
4. Digital equipment may consume less battery power,
and hence may reduce equipment weight.
5. Digital systems can provide reliable authentication
and privacy (encryption).
4. REQUIREMENT FOR CELLULAR AND PCS
In 1996, the Telecommunications Act Bill was passed by
the U.S. Congress and stated, in simple terms, that every
one could get into everyone’s business. Cellular service is
moving toward digital and is trying to compete with PCS.
The PCS spectrum was auctioned in early 1996. There are
wideband PCSs and narrowband PCSs (see Fig. 6). The
spectrum of wideband PCS is allocated at 1900 MHz in
order to operate the same technology as the cellular sys
tem. The spectrum of narrowband PCS is allocated at
900 MHz and is used for twoway paging. The joint re
quirements of both cellular and PCS are as follows:
From the end user’s perspective—the PCS and cellular
units should be light in weight and small in size, and
have long talktime capabilities without battery re
charging and good quality in voice and data. The
unit should be employable for initiating and receiv
ing calls anywhere using any telephone feature. The
important requirement of PCS and cellular is to
please the vast majority of subscribers who always
prefer to carry a single unit, not many units. This
30 25 30 20 15
(C/ I )
s
in dB
10 5 0
q
10
9
8
7
6
5
4
3
2
1
0
Figure 5. Relationship between q and (C/I)
s
[Eq. (3)].
596 CELLULAR RADIO
unit can be classiﬁed according to the different
grades of service.
From the system provider’s perspective—the PCS
should provide full coverage and large system ca
pacity to serve end users. An end user unit ideally
should be serviced by one system with different
grades of service and unless there are natural limi
tations by the various personal communication en
vironments (such as mobile vehicle, pedestrian, and
indoor public communication). Then one end user
unit should be capable of accessing more than one
system.
5. DIGITAL MODULATIONS AND MULTIPLE ACCESS
5.1. Digital Modulation Schemes
Digital modulation schemes can be selected to conﬁne the
transmitted energy of a digital voice signal in a given fre
quency bandwidth while transmitting in a mobile radio
environment. The information may have to be modulated
by signal phases or frequencies, rather than amplitudes,
because the multipath fading impairs the signal ampli
tude.
5.2. Multiple Access
Digital transmission can use timedivision multiple access
(TDMA), frequencydivision multiple access (FDMA), or
codedivision multiple access (CDMA), but in analog
transmission only FDMA can be used. FDMA provides
many different frequency channels, where each is as
signed to support a call. TDMA means chopping a rela
tively broadband channel over time into many timeslots.
Each timeslot is assigned to support a call. CDMA means
generating many different code signatures over a long
codebitstream channel, where each code signature is as
signed to convey a call. FDMA is a narrowband system. It
is a lowrisk system to develop but was voted down by the
industry in 1987. FDMA is not suitable for highspeed
data transmission. TDMA was ﬁrst developed in Europe
and is called GSM. TDMA has been developed in North
America. For the ADC (American Digital Cellular system),
CDMA needs more advanced technology and is relatively
harder to implement than the other two multipleaccess
schemes, especially in the mobile radio environment.
However, the improved user capacity of CDMA has given
the cellular industry the incentive to develop this system.
Therefore, digital transmission in the mobile radio envi
ronment has only two competing multiple accesses. The
North America selected TDMA based on the inﬂuence
from the European GSM.
6. SPECIFICATIONS FOR DIFFERENT CELLULAR/PCS
SYSTEMS
6.1. Analog Systems
Each trafﬁc channel in an analog system uses two fre
quencies, one receiving and one transmitting frequency. In
general, we often refer to ‘‘a 30kHz channel’’ when we
really mean a bandwidth of 30 kHz on one of two frequen
cies. Therefore, the total occupied spectrum for each trafﬁc
channel is 60 kHz. There are three analog systems: The
AMPS from North America, the NTT system from Japan,
and the TACS system from the UK. Their speciﬁcations
are listed in Table 1.
6.2. TDMA Systems
The following TDMA systems can be grouped into two dif
ferent duplexing techniques, FDD and TDD:
FDD. Frequencydivision duplexing, where each trafﬁc
channel consists of two operational frequencies. The
analog system can use only a FDD system, whereas
the digital system has a choice.
Wideband PCS –for cellular like systems
Narrowband PCS –for twoway paging systems
Five 50 kHz channels paired with 50 kHz channels
Three 50 kHz channels paired with 12.5 kHz channels
Three 50 kHz unpaired channels
UV unlicensed voice
UD unlicensed data
Base Rv
1800 1870
15 5 10 5 5 5 5 5 5 5 15 15 15 15 15
D D B EF C B EF C UD
UV
A A
1850 1900 1950 1990 1930 1970
Base Tx
901.00 .05 .1 .15 .20 901.25 940.00 .05 .1 .15 .20 940.25 MHz
901.75
.7625 .7750
901.7875 930.40 .45 .50 930.55 MHz
940.75 .80 .85 940.90 MHz
Figure 6. Spectrum allocated for wideband
PCS and narrowband PCS.
CELLULAR RADIO 597
GSM. The term GSM often implies DCS1800 and
DCS1900 services. They are in the same family,
only the carrier frequencies are different. We list
the physical layer parameters in Table 2.
NATDMA (North AmericanTDMA). NATDMA,
sometimes called ADC, is North America’s standard
system. It incorporates both 800 MHz and 1900MHz
system versions. The network follows philosophy of
the GSM intelligent network. The physical layer is
shown in Table 3.
The PDC (personal digital cellular) system. This sys
tem was developed in Japan and is very similar to the
NATDMA system, but its radio carrier bandwidth is
25 kHz.
IDEN (integrated digital enhanced network). This sys
tem was developed by Motorola. It was called MIRS
(mobile integrated radio system); then Motorola mod
iﬁed the system and renamed it IDEN. This system
uses the SMR (Special Mobile Radio) band, which is
speciﬁed by Part 90 of FCC CER (Code of Federal
Regulations) in the private sector. The system now
can be used cellularlike commercial services. The
physical parameter system is as follows:
1. Fullduplex communication system
2. Frequency: 806–824MHz (mobile transmitter),
851MHz
3. Channel bandwidth: 25 kHz
4. Multiple access: TDMA
5. Number of timeslots: 6
6. Rate of speech coder: VSELP (vector sum excita
tion linear predicted)
7. No equalizer implemented
8. Handoff
9. Transmission rate: 6.5kbps/slot
10. Forward error correction: 3 kbps
11. Dispatch capability
Table 2. PhysicalLayer Parameters of GSM
Parameter Speciﬁcations
Radio carrier bandwidth 200kHz
TDMA structure 8 timeslots per radio carrier
Timeslot 0.577ms
Frame interval 8 timeslots =4.615ms
Radio carrier number 124 radio carriers (935–960MHz
downlink, 890–915MHz uplink)
Modulation scheme Gaussian minimum shift keying
with BT
a
=0.3
Frequency hopping Slow frequency hopping
(217 hops/s)
Equalizer Equalization up to 16ms time dis
persion
Frequency hop rate 217 hops/s
Handover Hard handover
a
BT=bandwidth time.
Table 3. Physical Layer of NATDMA
Parameter Speciﬁcations
Radio carrier band
width
30kHz
TDMA structure 3 timeslots per radio carrier
Timeslot 6.66ms
Frame interval 20ms
Radio carrier number 2 416 (824–849 MHz reverse link,
869–894MHz forward link)
Modulation scheme
p
4
÷DQPSK
Equalizer Equalization up to 60ms time dis
persion
Table 1. LargeCapacity Analog Cellular Telephones Used in the World
Japan North America United Kingdom
System transmission frequency
(MHz)
Base station 870–885 869–894 917–950
Mobile station 925–940 824–849 872–905
Spacing between transmission
and receiving frequencies
(MHz)
55 45 45
Spacing between channels (kHz) 25, 12.5 30 25
Number of channels 600 832 (control channel 21 2) 1320 (control channel 21 2)
Coverage radius (km) 5 (urban area) 2–20 2–20
10 (suburbs)
Audio signal: type of modulation FM FM FM
Frequency deviation (kHz) 75 712 79.5
Data transmission rate (kbps) 0.3 10 8
Message protection Transmitted signal is
checked when it is sent
back to the sender by the
receiver.
Principle of majority decision is
employed.
Principle of majority decision is
employed.
Source: Report from International Radio Consultative Committee (CCIR) 1987.
598 CELLULAR RADIO
TDD. Timedivision duplexing, where transmission
and reception are shared by one frequency. Certain
timeslots are for transmission and certain timeslots
are for reception.
CT2 (Cordless Phone Two). CT2 was developed by
GPT Ltd. in the UK for socalled telepoint applica
tions. Phone calls can be dialed out but cannot be re
ceived. The transmission parameters for CT2 are as
follows:
1. Fullduplex system
2. Voicecoder: 32 kbps adaptive differential pulse
code modulation (ADPCM).
3. Duplexing: TDD, where portable and base units
transmit and receive on the same frequency but
different timeslots
4. Multiple access: TDMATDD, up to four multi
plexed circuits
5. Modulation: p/4 DQPSK differential QPSK, rolloff
rate =0.5
6. Data rate: 192ksym/s (192 kilosymbols per second
or 384kbps)
7. Spectrum allocation: 1895–1918.1 MHz (this spec
trum has been allocated for private and public use)
8. Carrier frequency spacing: 300kHz
PHS (personal handyphone system). It was developed
in Japan. Now there are three operators: NTT, STEL,
and DDI. The system serves for the lowtier subscrib
ers, such as teenagers. There are around 7 million
customers. The speciﬁcations for transmission pa
rameters are as follows:
1. Fullduplex system
2. Voicecoder: 32 kbps adaptive differential pulse
code modulation (ADPCM)
3. Duplexing: TDD, where portable and base units
transmit and receive on the same frequency but
different timeslots
4. Multiple access: TDMATDD, up to four multi
plexed circuits
5. Modulation: p/4 DQPSK, rolloff rate =0.5
6. Data rate: 192 ksym/s (or 384kbps)
7. Spectrum allocation: 1895–1918.1MHz (this spec
trum has been allocated for private and public use)
8. Carrier frequency spacing: 300kHz.
Another system called PACS (personal access commu
nication systems) [3] is in the same system family as PHS.
DECT (Digital European Cordless Telephone) [3]. DECT
is a European standard system for slowmotion or in
building communications. Its system structure is as
follows:
1. Duplex method: TDD
2. Access method: TDMA
3. RF (radiofrequency) power of handset: 10 mW
4. Channel bandwidth: 1.728 MHz/channel
5. Number of carriers: ﬁve (a multiplecarrier system)
6. Frequency: 1800–1900 MHz
DECT’s characteristics are as follows:
1. Frame: 10ms
2. Timeslots: 12
3. Bit rate: 38.8kb/slot
4. Modulation: GFSK (Gaussian ﬁltered FSK)
5. Handoff: yes
6.3. CDMA Systems
CDMA is another multipleaccess scheme using different
orthogonal code sequences to provide different call
connections. It is a broadband system and can be classi
ﬁed by two approaches: (1) frequencyhopping system
approach [3] and (2) directsequence system approach
[3]. The commercial CDMA system applies the direct
sequence approach. Developed in the United States,
it is called the IS95 Standard System. The ﬁrst CDMA
system was deployed in Hong Kong and then in Los
Angeles in 1995. CDMA is a highcapacity system. It
has been proved, theoretically, that CDMA system
capacity can be 20 times higher than analog capacity.
In a CDMA system, all the cells share the same radio
carrier in an operating system. The handoff from cell
to cell is soft (i.e., not only is the frequency kept
unchanged, but the cell is connected in both the old
cell and the new cell in the handoff region). The IS95
CDMA is now called cdmaone. The CDMA radio speciﬁ
cations are as follows:
1. CDMA shares the spectrum band with AMPS
2. Total number of CDMA radio carriers is 18.
3. Radio carrier bandwidth is 1.2288MHz.
4. Pseudo noise (PN) chip rate is 1.2288Mchips/s
5. Pilot channel is one per radio carrier.
6. Power control step is 1dB in 1 ms.
7. Soft handoffs are used.
8. Trafﬁc channels are 55 per each radio carrier.
9. Vocoder is qualcomm (quadrature) codeexcited lin
ear prediction (QCELP) at a variable rate.
10. Modulation is quaternary (quadrature) phase shift
keying (QPSK).
11. Data frame size is 20 ms.
12. Orthogonal spreading is 64 Walsh functions.
13. Long PN code length is 2
42
÷1 chips.
14. Short PN code length is 2
15
–1 chips.
6.4. Mobile Satellite Systems
Mobile satellite systems (MSSs) are used to enhance ter
restrial radiocommunication, either in rural areas or in
terms of global coverage. Therefore, MSS becomes, in a
CELLULAR RADIO 599
broad sense, a PCS system. By taking advantage of
reduced transmitting power and short time delays, the
lowEarth orbit (LEO) systems are being developed. How
ever, there is a drawback. Each LEO system needs many
satellites to cover the planet. There are many LEO sys
tems, as shown in Table 4. There is also another LEO re
ferred to as the Teledesic system, which will operate at
24 GHz with a spectrum band of 500MHz. This LEO sys
tem is not just for enhancing cellular or PCS coverage, but
also can replace the terrestrial longdistance telephone
network in the future.
6.5. IMT2000
Since the CDMA One system has been successfully
deployed in Korea and the United States, in mid1997
the European countries under the auspices of the socalled
(ETSI) European Telecommunications Standard Institute,
Japan (ARIB) Association of Radio Industrial and
Business, and the United States (TIA) Telecom Industri
al Assoc. began planning a universal singlestandard
system for the socalled IMT2000 (International
Mobile Telephone—Year 2000). There are three general
proposals. The proposals disagree on many issues,
but they do agree on the following general guideline
principles:
1. Use wideband CDMA (WCDMA).
2. Use direct sequence as spreadspectrum modula
tion.
3. There should be a multiband, single mobile unit.
4. The standard band should be 5 MHz.
5. There is a need for international roaming.
6. There should be IPR (intellectual property right)
issues in developing the new global system among
all the international vendors.
The IMT2000 system will require a great deal of com
promise in selecting technologies due to the political dif
ferences in the international standards bodies. The formal
IMT2000 system will be adapted by the ITU (Internation
al Telecommunication Union). A single universal IMT
2000 was established by the year 2000.
BIBLIOGRAPHY
1. S. H. Blecher, Advanced mobile phone services, IEEE Trans.
Vehic. Technol. VT29:238–244 (1980).
2. W. C. Y. Lee, Mobile Communications Design Fundamentals,
2nd ed., Wiley, New York, 1993.
3. W. C. Y. Lee, Mobile Cellular Telecommunications, Analog
and Digital Systems, McGrawHill, New York, 1995.
4. B. J. T. Malliner, An overview of the GSM system, Proc.
Digital Cellular Radio Conf., Hagen, FRG, Oct. 1988.
5. Y. Okumura et al., Field strength and its variability in VHF
and UHF landmobile radio service, Rev. Electron. Commun.
Lab. 16:825–873 (1968).
6. M. Hata, Empirical formula for propagation loss in land
mobile radio services, IEEE Trans. Vehic. Technol. VT29:
317–325 (1980).
7. W. C. Y. Lee, Spectrum efﬁciency in cellular, IEEE Trans.
Vehic. Technol. 38:69–75 (1989).
FURTHER READING
J. Gabion, The Mobile Comms Handbook, IEEE Press, New York,
1985.
M. Morly et al., The GSM System.
CHEBYSHEV FILTERS
ANTO
ˆ
NIO CARLOS
M. DE QUEIROZ
Federal University of Rio de
Janeiro
Rio de Janeiro, Brazil
1. INTRODUCTION
Any signal can be considered to be composed of several
sinusoidal components with different frequencies, ampli
tudes, and phases. Filtering is one of the fundamental
methods in signal processing, where the signal is pro
cessed by a linear system that changes the amplitudes and
Table 4. Comparative LowEarthOrbiting Mobile Satellite Service Applications
System Characteristics
Loral/
QUALCOMM Motorola Iridium TRW Odyssey
Constellation
ARIES (b) Ellipsat ELLIPSO
Number of satellites 48 66 12 48 24
Constellation altitude
(North Meridian)
750 421 5600 550 1767 230
Unique feature Transponder Onboard processing Transponder Transponder Transponder
Circuit capacity (U.S.) 6500 3835 4600 100 1210
Signal modulation CDMA TDMA CDMA FDMA/CDMA CDMA
Gateways in USA 6 2 2 5 6
Gateway spectrum band C band existing New Ka band New Ka band Unknown Unknown
Coverage Global Global Global Global Northern Hemisphere
600 CHEBYSHEV FILTERS
phases of these components, but not their frequencies. In
the most usual form, ﬁltering can be used to let pass or to
reject selected frequency bands, ideally with no attenua
tion at the passbands and inﬁnite attenuation at the stop
bands. This article discusses a class of approximations to
this kind of ideal ﬁlter, known as Chebyshev ﬁlters. It
starts with a discussion on a technique for the derivation
of optimal magnitude ﬁlters, then discusses the direct and
inverse Chebyshev approximations for the ideal ﬁltering
operator, continuing with comments on extensions of the
technique. Explicit formulas for LC ladder realizations for
some cases, and tables with example ﬁlters that can be
used to verify the properties of the ﬁlters and the formulas
in the article, are listed at the end.
The magnitude approximation problem in ﬁlter design
consists essentially in ﬁnding a convenient transfer func
tion with the magnitude satisfying given attenuation
speciﬁcations. Other restrictions can exist, such as struc
ture for implementation, maximum order, and maximum
Q of the poles, but in most cases the problem can be re
duced to the design of a normalized continuoustime low
pass ﬁlter that can be described by a transfer function in
Laplace transform. This ﬁlter must present a given max
imum passband attenuation (A
max
), between o=0 and
o=o
p
=1 rad/s, and a given minimum stopband attenua
tion (A
min
) in frequencies above a given limit o
r
rad/s.
From this prototype ﬁlter, the ﬁnal transfer function can
be obtained by frequency transformations [3,6,7], by con
tinuoustime to discretetime transformations in the case
of a digital ﬁlter [1], or by a convenient transformation for
realization by microwave structures [6].
A convenient procedure for the derivation of optimal
magnitude ﬁlters is to start with the transducer function
H(s) and the characteristic function K(s) [6]. H(s) can also
be called the attenuation function, which is the inverse of
the ﬁlter transfer function, scaled to have the minimum of
H(jo) equal to 1. K(s) is related to H(s) by the equation,
due to FeldtKeller [6]:
[H(jo)[
2
=1 ÷[K(jo)[
2
(1)
This greatly simpliﬁes the problem, because K(jo) can be a
ratio of two real polynomials in o, both with roots located
symmetrically on both sides of the real axis, while H(jo) is
a complex function. K(s) is obtained by replacing o by s/j in
K(jo), and ignoring possible 7j or ÷1 multiplying terms
resulting from the operation. The complex frequencies
where K(s) =0 are the attenuation zeros, and K(s) =N
corresponds to the transmission zeros. If K(s) is a ratio of
real polynomials in s, then K(s) =F(s)/P(s), H(s) is also a
ratio of real polynomials, with the same denominator,
H(s) =E(s)/P(s), and E(s) can be obtained by observing
that for s =jo, Eq. (1) is equivalent to
H(s)H(÷s) =1 ÷K(s)K(÷s)
‘E(s)E(÷s) =P(s)P(÷s) ÷F(s)F(÷s)
(2)
Because E(s) is the denominator of the ﬁlter transfer func
tion, which must be stable, E(s) is constructed from the
roots of the polynomial P(s)P( ÷s) ÷F(s)F( ÷s) with nega
tive real parts. The desired transfer function is then
T(s) =P(s)/E(s).
2. CHEBYSHEV POLYNOMIALS
Two important classes of approximations, the direct and
inverse Chebyshev approximations, can be derived from a
class of polynomials known as Chebyshev polynomials.
These polynomials were ﬁrst described by P. L. Chebyshev
[2]. The Chebyshev polynomial of order n can be obtained
from the following expression:
C
n
(x) = cos(n cos
÷1
x) (3)
It is simple to verify that this expression corresponds, for
÷1_x_1, to a polynomial in x. Using the trigonometric
identity cos(a÷b) =cos acos b÷sin a sin b, we obtain
C
n÷1
(x) = cos[(n÷1) cos
÷1
x]
=xC
n
(x) ÷sin(n cos
÷1
x) sin(cos
÷1
x)
(4)
Now applying the identity sina sinb =
1
2
[cos(a ÷b)÷
cos(a÷b)] and rearranging, we obtain a recursion
formula:
C
n÷1
(x) =2xC
n
(x) ÷C
n÷1
(x) (5)
For n=0 and n=1, we have C
0
(x) =1 and C
1
(x) =x. Using
Eq. (5), the series of Chebyshev polynomials shown in Ta
ble 1 is obtained.
The values of these polynomials oscillate between ÷1
and ÷1 for x between ÷1 and ÷1, in a pattern identical
to a stationary Lissajous ﬁgure [3]. For x out of this range,
cos
÷1
x =j cosh
÷1
x, an imaginary value, but Eq. (3) is still
real, in the form
C
n
(x) = cos(nj cosh
÷1
x) = cosh(n cosh
÷1
x) (6)
For high values of x, looking at the polynomials in Table 1,
we see that C
n
(x)E2
n÷1
x
n
, growing monotonically. The
plots of some Chebyshev polynomials for ÷1_x _1 are
shown in Fig. 1.
Table 1. Chebyshev Polynomials
n C
n
(x)
0 1
1 x
2 2x–1
3 4 x
3
–3x
4 8 x
4
–8x
2
÷1
5 16x
5
–20x
3
÷5 x
6 32x
6
–48x
4
÷18 x–1
7 64x
7
–112x
5
÷56x
3
–7x
8 128x
8
–256x
6
÷160x
4
–32x
2
÷1
9 256x
9
–576x
7
÷432x
5
–120x
3
÷9 x
10 512x
10
–1280x
8
÷1120x
6
–400x
4
÷50x
2
–1
11 1024x
11
–2816x
9
÷2816x
7
–1232x
5
÷220x
3
–11 x
12 2048x
12
–6144x
10
÷6912x
8
–3584x
6
÷840x
4
–72x
2
÷1
CHEBYSHEV FILTERS 601
3. THE CHEBYSHEV LOWPASS APPROXIMATION
The normalized Chebyshev approximation for lowpass ﬁl
ters is obtained by using
K( jo) =eC
n
(o) (7)
The result is a transducer function with the magnitude
given by [from Eq. (1)]
[H( jo)[
2
=1 ÷[eC
n
o ( )]
2
(8)
The corresponding attenuation in decibels is
A(o) =10 log{1 ÷[eC
n
(o)]
2
] (9)
The parameter e controls the maximum passband attenu
ation, or the passband ripple. Considering that when
C
n
(o) =71 the attenuation A(o) =A
max
, Eq. (9) gives
e =(10
0:1A
max
÷1)
1=2
(10)
Figure 2 shows examples of the magnitude function
T(\, jo) in the passband and in the stopband obtained
for some normalized Chebyshev lowpass approximations,
with A
max
=1dB. The magnitude of the Chebyshev ap
proximations presents uniform ripple in the passband,
with the gain departing from 0 dB at o=0 for odd orders
and from ÷A
max
dB for even orders.
The stopband attenuation is the maximum possible
among ﬁlters derived from polynomial characteristic func
tions, with the same A
max
and degree [4]. This can be
proved by assuming that there exists a polynomial P
n
(x)
that is also bounded between ÷1 and 1 for ÷1_x _1, with
P
n
(x) =7P
n
( ÷x) and P
n
( ÷N) = ÷N, but that exceeds
the value of C
n
(x) for some value of x41. An approxima
tion using this polynomial instead of C
n
(x) in Eq. (7) would
be more selective. The curves of P
n
(x) and C
n
(x) will al
ways cross n times for ÷1_x _1, due to the maximum
oscillations of C
n
(x), but if P
n
(x) grows faster, they will
cross another 2 times for xZ1 and x_÷1. This makes
P
n
(x) ÷C
n
(x) a polynomial of degree n÷2, because it
presents n÷2 roots, what is impossible since both are of
degree n.
The required approximation degree for given A
max
and
A
min
can be obtained by substituting Eq. (6) in Eq. (9),
with A(o
r
) =A
min
and solving for n. The result, including a
denormalization for any o
p
is
n_
cosh
÷1
g
cosh
÷1
(o
r
=o
p
)
(11)
where it is convenient to deﬁne the following constant:
g =
10
0:1A
min
÷1
10
0:1A
max
÷1
_ _
1=2
(12)
The transfer functions for the normalized Chebyshev ﬁl
ters can be obtained by solving Eq. (2). For a polynomial
T ( j o) dB
T ( j o) dB
0 1
0
–1
1
2
3
4
5
6
ω
0
1
10
ω
–100
1
5
10
–1
(a)
(b)
Figure 2. Passband gain (a) and stopband gain (b) for the ﬁrst
normalized Chebyshev approximations with 1dB passband rip
ple. Observe the uniform passband ripple and the monotonic
stopband gain decrease.
C
1
C
2
C
3
C
4
C
5
C
6
Figure 1. Plots of the ﬁrst six Chebyshev polynomials C
n
(x). The
squares limit the region ÷1_x _1, ÷1_C
n
(x)_1, where the poly
nomial value oscillates.
602 CHEBYSHEV FILTERS
approximation, using P(s) =1, from Eq. (7), it follows that
E(s)E(÷s) =1 ÷ eC
n
s
j
_ _ _ _
2
(13)
The roots of this polynomial are the solutions for s in
C
n
s
j
_ _
= cos n cos
÷1
s
j
_ _
= ±
j
e
(14)
Identifying
n cos
÷1
s
j
=a÷jb (15)
it follows that ±(j=e) = cos(a÷jb) = cos a cos jb ÷
sin a sin jb = cos a cosh b ÷j sin a sinh b: Equating real
and imaginary parts, we have cos a cosh b=0 and
sin a sinh b = ((1=e). Since cosh xZ1, the equation of
the real parts gives
a=
p
2
(1÷2k) k=0; 1; . . . ; 2n ÷1 (16)
and as for these values of a, sin a=71, the equation of the
imaginary parts gives
b= ( sinh
÷1
1
e
(17)
Applying these results in Eq. (15), it follows that the roots
of E(s)E( ÷s) are
s
k
=s
k
÷jo
k
k=0; 1; . . . ; 2n ÷1
s
k
= sin
p
2
1 ÷2k
n
sinh
1
n
sinh
÷1
1
e
_ _
o
k
= cos
p
2
1÷2k
n
cosh
1
n
sinh
÷1
1
e
_ _
(18)
The roots s
k
with negative real parts (kZn) are the roots
of E(s). By the expressions in Eq. (18), it is easy to see
that the roots s
k
are located in an ellipse with vertical
semiaxis cosh[(1=n) sinh
÷1
(1=e)], horizontal semiaxis
sinh[(1=n) sinh
÷1
(1=e)], and foci at 7j. The location of
the roots can be best visualized with the diagram shown
in Fig. 3 [3].
4. REALIZATION OF CHEBYSHEV FILTERS
These approximations were originally developed for real
ization in passive form, and the best realizations were ob
tained as LC doubly terminated structures designed for
maximum power transfer at the passband gain maxima
[3,6,7]. These structures are still important today in high
frequency ﬁlters and as prototypes for active and digital
realizations, due to the low sensitivity to errors in element
values. At each attenuation zero, and the Chebyshev
approximations have the maximum possible number
of them distributed in the passband, maximum power
transfer occurs between the terminations. In this condi
tion, errors in the capacitors and inductors can only de
crease the gain [5]. This causes zeros in the derivatives of
[T( jo)[ in relation to all reactive element values at the at
tenuation zeros, and reduces the error throughout the
passband. Table 2 lists polynomials, poles, frequency and
Q of the poles, and LC doubly terminated ladder struc
tures, with the structure shown in Fig. 4a, for some nor
malized Chebyshev lowpass ﬁlters. Note in the
realizations that oddorder ﬁlters have identical termina
tions, but evenorder ﬁlters require different terminations,
because there is no maximum power transfer at o=0,
since the gain is not maximum there. With the impedance
normalization shown, it is clear that the evenorder real
izations have antimetrical structure (one side is the dual
of the other). The oddorder structures are symmetric.
5. THE INVERSE CHEBYSHEV LOWPASS APPROXIMATION
The inverse Chebyshev approximation is the most impor
tant member of the inverse polynomial class of approxi
mations. The lowpass version is conveniently obtained by
using the characteristic function obtained from
K(jo) =
F(jo)
P(jo)
=
eg
C
n
(1=o)
=
ego
n
o
n
C
n
(1=o)
(19)
( )
1
cosh
n
sinh
–1
1
ε
( )
1
sinh
n
sinh
–1
1
ε
π
n
π
2n
π
n
π
n
j o
σ
Figure 3. Localization of the poles in a normalized Chebyshev
lowpass approximation (seventh order in this case). The pole lo
cations can be obtained as shown.
CHEBYSHEV FILTERS 603
Table 2. Normalized Chebyshev ﬁlters with A
max
=1dB and x
p
=1rad/s
Polynomials E(s)
n a
0
a
1
a
2
a
3
a
4
a
5
a
6
a
7
a
8
a
9
a
10
1 1.96523 1.00000
2 1.10251 1.09773 1.00000
3 0.49131 1.23841 0.98834 1.00000
4 0.27563 0.74262 1.45392 0.95281 1.00000
5 0.12283 0.58053 0.97440 1.68882 0.93682 1.00000
6 0.06891 0.30708 0.93935 1.20214 1.93082 0.92825 1.00000
7 0.03071 0.21367 0.54862 1.35754 1.42879 2.17608 0.92312 1.00000
8 0.01723 0.10734 0.44783 0.84682 1.83690 1.65516 2.42303 0.91981 1.00000
9 0.00768 0.07060 0.24419 0.78631 1.20161 2.37812 1.88148 2.67095 0.91755 1.00000
10 0.00431 0.03450 0.18245 0.45539 1.24449 1.61299 2.98151 2.10785 2.91947 0.91593 1.00000
Poles
n re/im 1 o/Q 1 re/im 2 o/Q 2 re/im 3 o/Q 3 re/im 4 o/Q 4 re/im 5 o/Q 5
1 ÷1.96523
2 ÷0.54887 1.05000
0.89513 0.95652
3 ÷0.24709 0.99710 ÷0.49417
0.96600 2.01772
4 ÷0.13954 0.99323 ÷0.33687 0.52858
0.98338 3.55904 0.40733 0.78455
5 ÷0.08946 0.99414 ÷0.23421 0.65521 ÷0.28949
0.99011 5.55644 0.61192 1.39879
6 ÷0.06218 0.99536 ÷0.16988 0.74681 ÷0.23206 0.35314
0.99341 8.00369 0.72723 2.19802 0.26618 0.76087
7 ÷0.04571 0.99633 ÷0.12807 0.80837 ÷0.18507 0.48005 ÷0.20541
0.99528 10.89866 0.79816 3.15586 0.44294 1.29693
8 ÷0.03501 0.99707 ÷0.09970 0.85061 ÷0.14920 0.58383 ÷0.17600 0.26507
0.99645 14.24045 0.84475 4.26608 0.56444 1.95649 0.19821 0.75304
9 ÷0.02767 0.99761 ÷0.07967 0.88056 ÷0.12205 0.66224 ÷0.14972 0.37731 ÷0.15933
0.99723 18.02865 0.87695 5.52663 0.65090 2.71289 0.34633 1.26004
10 ÷0.02241 0.99803 ÷0.06505 0.90245 ÷0.10132 0.72148 ÷0.12767 0.47606 ÷0.14152 0.21214
0.99778 22.26303 0.90011 6.93669 0.71433 3.56051 0.45863 1.86449 0.15803 0.74950
Polynomials P(s)
n mult. a
0
1 1.96523 1.00000
2 0.98261 1.00000
3 0.49131 1.00000
4 0.24565 1.00000
5 0.12283 1.00000
6 0.06141 1.00000
7 0.03071 1.00000
8 0.01535 1.00000
9 0.00768 1.00000
10 0.00384 1.00000
Doubly terminated
LC ladder realizations
n Rg/Rl L/C 1 L/C 2 L/C 3 L/C 4 L/C 5 L/C 6 L/C 7 L/C 8 L/C 9 L/C 10
1 1.00000
1.00000 1.01769
2 1.63087 1.11716
0.61317 1.11716
3 1.00000 0.99410
1.00000 2.02359 2.02359
4 1.63087 1.73596 1.28708
0.61317 1.28708 1.73596
5 1.00000 1.09111 1.09111
1.00000 2.13488 3.00092 2.13488
6 1.63087 1.80069 1.87840 1.32113
0.61317 1.32113 1.87840 1.80069
7 1.00000 1.11151 1.17352 1.11151
1.00000 2.16656 3.09364 3.09364 2.16656
8 1.63087 1.82022 1.93073 1.90742 1.33325
0.61317 1.33325 1.90742 1.93073 1.82022
9 1.00000 1.11918 1.18967 1.18967 1.11918
1.00000 2.17972 3.12143 3.17463 3.12143 2.17972
10 1.63087 1.82874 1.94609 1.95541 1.91837 1.33890
0.61317 1.33890 1.91837 1.95541 1.94609 1.82874
604 CHEBYSHEV FILTERS
where e and g are as given by Eqs. (10) and (11). The poly
nomials F(s) and P(s) are then
F(s) =eg
s
j
_ _
n
P(s) =
s
j
_ _
n
C
n
j
s
_ _
(20)
Ignoring 7j or ÷1, multiplying factors in Eq. (20), F(s)
reduces to egs
n
, and P(s) reduces to a Chebyshev polyno
mial with all the terms positive and the coefﬁcients in re
verse order. The magnitude characteristic of this
approximation is maximally ﬂat at o=0, due to the n at
tenuation zeros at s =0, and thus is similar in the pass
band to a Butterworth approximation. In the stopband, it
presents a series of transmission zeros at frequencies in
verse to the roots of the corresponding Chebyshev polyno
mial. Between adjacent transmission zeros, there are gain
maxima reaching the magnitude of ÷A
min
dB. Without
renormalization, the stopband starts at 1 rad/s, and the
passband ends where the magnitude of the characteristic
function, Eq. (19), reaches e:
o
p
=
1
C
÷1
n
(g)
=
1
cosh
1
n
cosh
÷1
g
_ _ (21)
Oddorder ﬁlters present a single transmission zero at in
ﬁnity, and evenorder ﬁlters end with a constant gain
÷A
min
at o=N. From Eqs. (1) and (19), the attenuation
in decibels for a normalized inverse Chebyshev approxi
mation is
A(o) =10 log 1 ÷
eg
C
n
(1=o)
_ _
2
_ _
(22)
The gains for some normalized inverse Chebyshev approx
imations are plotted in Fig. 5. A frequency scaling by the
factor given by Eq. (21) was applied, causing the passband
to end at o=1.
The selectivity of the inverse Chebyshev approximation
is the same as the corresponding Chebyshev approxima
tion, for the same A
max
and A
min
. This can be veriﬁed by
calculating the ratio o
p
/o
r
for both approximations. For
the normalized Chebyshev approximation, o
p
=1, and o
r
occurs when eC
n
(o
r
) =g. For the normalized inverse
Chebyshev approximation, o
r
=1, and o
p
occurs when
(eg)/C
n
(1/o
p
) =e. In both cases, the resulting ratio is
o
r
/ o
p
=C
n
÷1
(g). Equation (11) can be used to compute
the required degree.
R
g
R
g
R
g
R
l
R
l
R
l
R
l
C
1
C
1
C
2
C
3
C
n
C
3
C
n
L
2
L
2
L
1
L
3
L
2
C
2
L
n
L
n
(a)
(b)
Figure 4. LC doubly terminated ladder realizations for Cheby
shev ﬁlters, in the direct form (a) and in the inverse form
(b). These classical realizations continue to be the best prototypes
for active realizations, due to their low sensitivity to errors in the
element values.
T( jo)dB
T( jo)dB
0 1
0
–1
1
2
ω
ω
10
0
1 100
–100
1
6
–1
(a)
(b)
Figure 5. Passband gain (a) and stopband gain (b) for the ﬁrst
normalized inverse Chebyshev approximations with A
max
=1 dB
and A
min
=50dB. Observe the maximally ﬂat passband and the
uniform stopband ripple.
CHEBYSHEV FILTERS 605
The transmission zero frequencies are the frequencies
that make Eq. (19) inﬁnite:
C
n
1
o
k
_ _
= cos n cos
÷1
1
o
k
_ _ _ _
=0
‘o
k
=
1
cos
p
2
1÷2k
n
_ _; k =0; 1; . . . ; n ÷1
(23)
The pole frequencies are found by solving Eq. (2) with F(s)
and P(s) as given by Eq. (20):
E(s)E(÷s) =(eg)
2
s
j
_ _
2n
÷
s
j
_ _
2n
C
n
j
s
_ _
2
(24)
The roots of this equation are the solutions of
C
n
j
s
_ _
= ±jeg (25)
By observing the similarity of this equation to Eq. (14), the
roots of E(s)E( ÷s) can be obtained as the complex inverses
of the values given by Eq. (18), with e replaced by 1/(eg).
They lie in a curve that is not an ellipse. E(s) is construct
ed from the roots with negative real parts, which are dis
tributed in a pattern that resembles a circle shifted to the
left side of the origin.
Because of the similarity of the passband response to
the Butterworth response, the phase characteristics of the
inverse Chebyshev ﬁlters are much closer to linear than
those of the direct Chebyshev ﬁlters, simplifying the task
of a phase equalizer that may be cascaded with the mag
nitude ﬁlter. The maximum Q of the poles is also signif
icantly lower for the same gain speciﬁcations.
6. REALIZATION OF INVERSE CHEBYSHEV FILTERS
The realization based on LC doubly terminated ladder
structures is also convenient for inverse Chebyshev ﬁlters,
by the same reasons mentioned for the direct approxima
tion. In this case, the passband sensitivities are low be
cause of the nthorder attenuation zero at s =0, which
results in the nulliﬁcation of the ﬁrst n derivatives of the
ﬁlter gain in relation to all the reactive elements at s =0,
and keeps the gain errors small in all the passband. Stop
band errors are also small, because the transmission zero
frequencies depend only on simple LC series or parallel
resonant circuits. The usual structures used are shown in
Fig. 4b.
Those realizations are possible only for the oddorder
cases, because those structures can’t realize the constant
gain at inﬁnity that occurs in the evenorder approxi
mations (realizations with transformers or with negative
elements, or with just one termination, are possible).
Evenorder modiﬁed approximations can be obtained by
using, instead of the Chebyshev polynomials, polynomials
obtained by the application, to the Chebyshev polynomials,
of the Moebius transformation [4,6]
x
2
÷
x
2
÷x
2
z1
1 ÷x
2
z1
; x
z1
= cos
k
max
p
2n
(26)
where k
max
is the greatest odd integer that is less than the
ﬁlter order n. This transformation moves the pair of roots
closer to the origin of an evenorder Chebyshev polynomial
to the origin. If the resulting polynomials are used to gen
erate polynomial approximations, starting fromEq. (7), the
results are ﬁlters with two attenuation zeros at the origin
that are realizable as a doubly terminated ladder ﬁlter
with equal terminations, a convenience in passive realiza
tions. If the same polynomials are used in inverse polyno
mial approximations, starting from Eq. (19), the results
are ﬁlters with two transmission zeros at the inﬁnity,
which are now realizable by doubly terminated LC struc
tures. The direct and inverse approximations obtained in
this way have the same selectivity, slightly smaller than in
the original case.
Table 3 lists polynomials, poles, zeros, frequency and Q
of the poles, and LC doubly terminated realizations for
some inverse Chebyshev ﬁlters. The ﬁlters were scaled in
frequency to make the passband end at 1 rad/s instead of
o
p
[Eq. (21)]. The evenorder realizations were obtained
from modiﬁed approximations with two transmission ze
ros at inﬁnity, and are listed separately in Table 4. The
structures are a mix of the two forms in Fig. 4b. Note that
some realizations are missing. These are cases where the
network would require negative elements, or transform
ers. For inverse Chebyshev ﬁlters, and other inverse poly
nomial ﬁlters, there is a minimum value of A
min
for each
order that turns a pure LC doubly terminated realization
possible [7].
7. OTHER SIMILAR APPROXIMATIONS
Different approximations with uniform passband or stop
band ripple, somewhat less selective, can be generated by
reducing the number or the amplitude of the oscillations
in a Chebyshevlike polynomial, and generating the ap
proximations starting from Eq. (7) or (19), numerically [8].
A particularly interesting case results if the last oscil
lations of the polynomial value end in 0 instead of 71.
This creates double roots close to x =71 in the polynomi
al. In a polynomial approximation, the higherfrequency
passband minimum disappears, replaced by a secondor
der maximum close to the passband border. In an LC dou
bly terminated realization, the maximum power transfer
at this frequency causes nulliﬁcation of the ﬁrst two de
rivatives of the gain in relation to the reactive elements,
substantially reducing the gain error at the passband bor
der. In an inverse polynomial approximation, this causes
the joining of the ﬁrst two transmission zeros, as a double
transmission zero, which increases the attenuation and
reduces the error at the stopband beginning, also allowing
a symmetric realization for orders 5 and 7.
Other variations arise from the shifting of roots to the
origin. This is also best done numerically. Odd (even)
order polynomial approximations with any odd (even)
606 CHEBYSHEV FILTERS
Table 3. Normalized Inverse Chebyshev Filters with A
max
=1dB, A
min
=50dB, and x
p
=1rad/s
Polynomials E(s)
n a
0
a
1
a
2
a
3
a
4
a
5
a
6
a
7
a
8
a
9
a
10
1 1.96523 1.00000
2 1.96838 1.98099 1.00000
3 2.01667 3.14909 2.51015 1.00000
4 2.19786 4.52937 4.90289 3.13118 1.00000
5 2.60322 6.42983 8.61345 7.26320 3.81151 1.00000
6 3.35081 9.35051 14.61162 14.91369 10.30744 4.54023 1.00000
7 4.64002 14.09440 24.72451 29.03373 24.18372 14.09633 5.30979 1.00000
8 6.82650 22.03426 42.29782 55.31092 52.89124 37.20009 18.68307 6.11268 1.00000
9 10.54882 35.60372 73.49954 104.68294 111.48145 90.07839 54.81844 24.10445 6.94337 1.00000
10 16.95789 59.19226 129.80937 198.24216 230.34722 207.44800 145.48766 77.89699 0.39330 7.79647 1.00000
Poles
n re/im 1 o/Q 1 re/im 2 o/Q 2 re/im 3 o/Q 3 re/im 4 o/Q 4 re/im 5 o/Q 5
1 ÷1.96523
0.00000
2 ÷0.99049 1.40299
0.99363 0.70823
3 ÷0.61468 1.25481 ÷1.28079
1.09395 1.02071
4 ÷0.42297 1.18385 ÷1.14262 1.25229
1.10571 1.39945 0.51249 0.54799
5 ÷0.30648 1.13993 ÷0.94418 1.23656 ÷1.31018
1.09795 1.85969 0.79849 0.65483
6 ÷0.23016 1.10962 ÷0.75398 1.21506 ÷1.28598 1.35770
1.08549 2.41056 0.95283 0.80576 0.43545 0.52789
7 ÷0.17794 1.08768 ÷0.59638 1.18959 ÷1.14085 1.36871 ÷1.47946
1.07303 3.05632 1.02930 0.99735 0.75619 0.59986
8 ÷0.47425 1.16431 ÷0.14101 1.07137 ÷0.95398 1.34983 ÷1.48710 1.55173
1.06334 1.22752 1.06205 3.79891 0.95496 0.70747 0.44316 0.52173
9 ÷0.38185 1.14152 ÷0.77805 1.31643 ÷0.11413 1.05899 ÷1.34453 1.56247 ÷1.70623
1.07575 1.49471 1.06189 0.84597 1.05282 4.63922 0.79596 0.58105
10 ÷0.63221 1.27960 ÷0.31203 1.12193 ÷0.09407 1.04944 ÷1.13939 1.53032 ÷1.72054 1.78611
1.11252 1.01201 1.07766 1.79777 1.04521 5.57772 1.02161 0.67155 0.47954 0.51906
Polynomials P(s)
n mult. a
0
a
2
a
4
a
6
a
8
a
10
1 1.96523 1.00000
2 0.00316 622.45615 1.00000
3 0.05144 39.20309 1.00000
4 0.00316 695.02278 74.56663 1.00000
5 0.03477 74.86195 19.34709 1.00000
6 0.00316 1059.61979 494.96516 57.80151 1.00000
7 0.03463 133.99401 95.81988 19.57753 1.00000
8 0.00316 2158.72730 2130.49651 657.07341 64.84805 1.00000
9 0.03786 278.65997 354.39519 150.23800 23.58892 1.00000
10 0.00316 5362.55604 8380.91576 4584.36462 1023.53040 79.98165 1.00000
Zeros
n o
1
o
2
o
3
o
4
o
5
1 N
2 24.94907
3 6.26124 N
4 7.97788 3.30455
5 3.74162 2.31245 N
6 6.92368 2.53424 1.85520
7 3.60546 2.00088 1.60458 N
8 7.29689 2.56233 1.71209 1.45144
9 3.88896 2.06927 1.53587 1.35062 N
10 8.08496 2.78589 1.78865 1.41948 1.28053
LC doubly terminated
realizations
n R
g
/R
l
L/C 1 L/C 2 L/C 3 L/C 4 L/C 5 L/C 6 L/C 7
1 1.00000
1.00000 1.01769
3 1.00000 1.56153
1.00000 0.78077 0.01634 0.78077
5 1.00000 1.16364 1.30631
1.00000 0.37813 0.16071 1.62010 0.05468 0.47172
7 1.00000 0.72897 1.34370 0.96491
1.00000 0.09574 0.34265 1.32044 0.28905 1.32059 0.07972 0.30081
CHEBYSHEV FILTERS 607
number of attenuation zeros at o=0, up to the approxi
mation’s order (in the last case resulting a Butterworth
approximation), can be generated. The same polynomials
generate inverse polynomial approximations with any odd
(even) number of transmission zeros at inﬁnity.
In all cases, the maximum Q of the poles is reduced and
the phase is closer to linear. Similar techniques can also be
applied to elliptic approximations. For example, a lowpass
elliptical approximation can be transformed into a Cheby
shev approximation by the shifting all the transmission
zeros to inﬁnity, or into an inverse Chebyshev approxima
tion by shifting all the attenuation zeros to the origin.
There are many possibilities between these extremes.
8. EXPLICIT FORMULAS
The design of ﬁlter structures is simpliﬁed when explicit
formulas for the element values are available. They also
allow the design of very highorder ﬁlters (usually for
digital implementation) without serious numerical prob
lems. Explicit formulas for the element values of direct
Chebyshev ﬁlters have been known since the 1950s, and
are given below, in the version due to Takahasi, adapted
for the notation used here. Their proof can be found in Ref.
7. The formulas apply to the structure in Fig. 4a, and solve
not only the case where there is maximum power transfer,
but also the mismatched cases where the terminations are
Table 4. Normalized EvenOrder Modiﬁed Inverse Chebyshev Filters with Two Transmission Zeros at Inﬁnity, with A
max
=
1dB, A
min
=50dB, and x
p
=1rad/s
Polynomials E(s)
n a
0
a
1
a
2
a
3
a
4
a
5
a
6
a
7
a
8
a
9
a
10
2 1.96523 1.98254 1.00000
4 2.12934 4.47598 4.86847 3.12041 1.00000
6 3.14547 9.02141 14.23655 14.65051 10.18872 4.51414 1.00000
8 6.32795 20.98707 40.68275 53.69811 51.69862 36.58972 18.48009 6.07949 1.00000
10 15.69992 56.17036 124.1801 191.3464 223.6989 202.6490 142.8395 76.84994 30.12162 7.76165 1.00000
Poles
n re/im 1 o/Q 1 re/im 2 o/Q 2 re/im 3 o/Q 3 re/im 4 o/Q 4 re/im 5 o/Q 5
2 ÷0.99127 1.40187
0.99127 0.70711
4 ÷0.43134 1.18419 ÷1.12886 1.23225
1.10284 1.37268 0.49409 0.54580
6 ÷0.23626 1.11107 ÷0.76275 1.20632 ÷1.25806 1.32324
1.08566 2.35141 0.93457 0.79077 0.41016 0.52590
8 ÷0.14421 1.07247 ÷0.48399 1.16315 ÷0.96075 1.33672 ÷1.45079 1.50858
1.06273 3.71848 1.05767 1.20162 0.92940 0.69566 0.41357 0.51992
10 ÷0.64341 1.27784 ÷0.31781 1.12245 ÷0.09573 1.05011 ÷1.14584 1.51514 ÷1.67803 1.73625
1.10404 0.99303 1.07652 1.76590 1.04574 5.48452 0.99132 0.66115 0.44586 0.51735
Polynomials P(s)
n mult. a
0
a
2
a
4
a
6
a
8
2 1.96523 1.00000
4 0.16412 12.97454 1.00000
6 0.11931 26.36278 10.89186 1.00000
8 0.13145 48.13911 44.73326 12.54437 1.00000
10 0.16119 97.39855 147.0191 76.50032 15.68797 1.00000
Finite zeros
n o
1
o
2
o
3
o
4
2 –
4 3.60202
6 2.69467 1.90542
8 2.71078 1.74464 1.46706
10 2.94484 1.82001 1.43081 1.28694
Doubly terminated
LC ladder realizations
n R
g
/R
l
L/C 1 L/C 2 L/C 3 L/C 4 L/C 5 L/C 6 L/C 7 L/C 8
2 1.00000 1.00881
1.00000 1.00881
4 1.00000 1.51207 0.05275 0.58997
1.00000 0.64094 1.46110
6 1.00000 0.87386 1.67233 0.13065 0.32187
1.00000 0.17880 0.31519 1.63514 1.05413
8 1.00000 0.67581 0.32023 1.34317 0.38303 1.12998 0.18178 0.16760
1.00000 0.32898 1.02594 1.21303 0.74862
608 CHEBYSHEV FILTERS
arbitrarily chosen. We start by redeﬁning Eq. (8) to allow
for mismatched terminations as
[H(jo)[ =
1 ÷[eC
n
(o)]
2
A
(27)
where A_1. For degree n, with arbitrarily chosen termi
nations R
g
and R
l
_R
g
, the constant A is obtained (from the
definition: H(s) =actual attenuation/minimum possible
attenuation) as
A=
4R
g
R
l
R
g
÷R
l
_ _
2
; for oddn
A= 1÷e
2
_ _
4R
g
R
l
R
g
÷R
l
_ _
2
; for evenn
(28)
The matched case is obtained when R
g
=R
l
for odd orders
and, or for even orders, when
R
l
R
g
=
(1 ÷e
2
)
1=2
÷e
(1 ÷e
2
)
1=2
÷e
(29)
We then compute the two constants
k =
1
e
2
÷1
_ _
1=2
÷
1
e
_ _
1=n
(30)
h=
1 ÷A
e
2
÷1
_ _
1=2
÷
1 ÷A
e
2
_ _
1=2
_ _
1=n
(31)
The ﬁrst capacitor C
1
is given by
C
1
=
2s
1
R
g
[(k ÷k
÷1
) ÷(h ÷h
÷1
)]
(32)
and the other elements can be calculated by the recursion
formulas
C
2m÷1
L
2m
=
4s
4m÷3
s
4m÷1
b
2m÷1
(k ÷k
÷1
; h ÷h
÷1
)
C
2m÷1
L
2m
=
4s
4m÷1
s
4m÷1
b
2m
(k ÷k
÷1
; h ÷h
÷1
)
(33)
where m=1, 2,y, to the last integer rn/2. The network
ends at C
n
for odd n, and in L
n
for even n, which can be
directly calculated, if convenient, as
C
n
=
2s
1
R
l
[(k ÷k
÷1
) ÷(h ÷h
÷1
)]
L
n
=
2R
l
s
1
(k ÷k
÷1
) ÷(h ÷h
÷1
)
(34)
The terms that appear in the formulas are
b
m
(x; Z) =x
2
÷c
2m
xZ ÷Z
2
÷s
2
2m
s
r
=2 sin
pr
2n
; c
r
=2 cos
pr
2n
(35)
If it is desired to have R
l
4R
g
, the network can be designed
with the terminations interchanged and assembled invert
ed. A singly terminated network can be obtained by using a
large R
g
in the formulas. With R
g
=N, the formulas give A
=0, k=h, and a limit in Eq. (32). An exact design can be
obtained starting from the output end, with Eq. (34). The
solutions for the matched cases (A=1, h=1) and singly
terminating cases are unique. The solution given by the
formulas for the mismatched cases are not the only solu
tions possible (see Ref. 7 for details). Explicit formulas for
the element values of inverse Chebyshev LC ladder ﬁlters
and for the variations discussed above are not known.
9. EXAMPLE
As an example of the application of the explicit formulas
and the properties of the resulting ﬁlter, consider a ﬁfth
order ﬁlter with 1dB passband ripple, passband edge at
10 MHz, and terminations R
g
=100 O and R
l
=50O. From
Eq. (28) A=0.88888, from Eq. (30) k=1.33055, and from
Eq. (31) h=1.13099. The frequencynormalized element
values are then obtained starting from Eq. (32), with
Eq. (33) applied twice. The resulting values are then di
vided by 2p 10
7
, to place the passband edge at 10 MHz.
The results are C
1
=592.2pF, L
2
=1.106mH, C
3
=755.2pF,
L
4
=1.058 mH, and C
5
=476.5 pF.
Note that when the terminations are not matched,
there is no special reason for the sensitivities of the ﬁlter
magnitude in relation to the element values to be low.
There is a continuous degradation of the sensitivity char
acteristics, with the matched case being the least sensitive
and singly terminated case being the most sensitive.
Figure 6 shows the normalized gain (scaled to a maxi
mum of 0 dB), with expected error margins, for the exam
ple ﬁlter and for the two extreme cases. In the singly
terminated case, with R
g
=N, the input was assumed to
be a current source. The expected errors were computed
by sensitivity analysis [6], assuming uncorrelated 5% ran
dom variations in all the elements, including the termina
tions, using the formula for the gain statistical deviation
D T(jo)
¸
¸
¸
¸
=
20
ln(10)
i
Dx
i
x
i
S
T(jo) [ [
x
i
_ _
2
_ _
1=2
dB (36)
where Dx
i
/x
i
is the tolerance of the element x
i
, set to 0.05 to
all the elements. Observe that for the matched case the
error returns to 70.307 dB at all the gain peaks, what
corresponds to the expected error due to the terminations
alone. The example ﬁlter produces only slightly larger er
rors. There is a range where the singly terminated real
ization is the best, but it is much worse at the critical area
of the passband border. Similar relations appear also for
other orders.
CHEBYSHEV FILTERS 609
BIBLIOGRAPHY
1. A. Antoniou, Digital Filters: Analysis, Design, and Applica
tions, McGrawHill, New York, 1993.
2. P. L. Chebyshev, The´orie des me´canismes connus sous le nom
de parallelogrammes, Oeuvres, Vol. I, St. Petersburg, 1899.
3. M. E. Van Valkenburg, Analog Filter Design, Holt, Rinehart
and Winston, New York, 1982.
4. R. W. Daniels, Approximation Methods for Electronic Filter
Design, McGrawHill, New York, 1974.
5. H. J. Orchard, Inductorless ﬁlters, Electron. Lett. 2:224–225
(Sept. 1966).
6. G. C. Temes and J. W. LaPatra, Circuit Synthesis and Design,
McGrawHill Kogakusha, Tokyo, 1977.
7. L. Weinberg, Network Analysis and Synthesis, McGrawHill,
New York, 1962.
8. A. C. M. de Queiroz and L. P. Caloˆba, An approximation algo
rithm for irregularripple ﬁlters, Proc. IEEE Int. Telecommu
nications Symp., Rio de Janeiro, Brazil, Sept. 1990, pp.
430–433.
CHIRALITY
AKHLESH LAKHTAKIA
Pennsylvania State University
University Park, Pennsylvania
1. INTRODUCTION
Chiral media have the ability to discriminate between left
handed and righthanded electromagnetic (EM) ﬁelds.
These media can be classiﬁed into two types: (1) isotropic
chiral media and (2) structurally chiral media. The mol
ecules of a naturally occurring isotropic chiral medium are
handed, while an artiﬁcial isotropic chiral medium can be
made by randomly dispersing electrically small, handed
inclusions (such as springs) in an isotropic achiral host
medium. The molecules of a structurally chiral medium,
such as a chiral nematic liquid crystal, are randomly po
sitioned but have helicoidal orientational order. Structur
ally chiral media can also be artiﬁcially fabricated either
as stacks of uniaxial laminae or using thinﬁlm technolo
gy. Whereas considerable theoretical and experimental
work on isotropic chiral media has been reported at mi
crowave frequencies during the 1980s and the 1990s, mi
crowave research on structurally chiral media remains in
an embryonic stage at the time of this writing [1]. There
fore, the major part of this article is devoted to isotropic
chiral media.
2. NATURAL OPTICAL ACTIVITY
Ordinary sunlight is split into its spectral components by
a prism. A spectral component is monochromatic (i.e., it
has one and only one wavelength l
0
in vacuum). The
wavelength l
0
of one of the visible spectral components
lies anywhere between 400nm (violet) and 700nm (red). A
spectral component can be almost isolated from other
spectral components by carefully passing sunlight
through a series of ﬁlters. Although ﬁltering yields quasi
monochromatic light, many experiments have been and
continue to be performed and their results analyzed, as
suming that the ﬁltered light is monochromatic.
Light is an EM wave with spectral components to
which our retinal pigments happen to be sensitive, and
the consequent images, in turn, happen to be decipherable
in our brains. All optical phenomena can be generalized to
other electromagnetic spectral regimes.
Suppose that a monochromatic EM wave is propagat
ing in a straight line in air, which is synonymous with
vacuum (or free space) for our present purpose. Its electric
ﬁeld vector vibrates in some direction to which the prop
agation direction is perpendicular; the frequency of vibra
tion is f =c/l
0
, where c =3 10
8
m/s is the speed of light in
vacuum. Its magnetic ﬁeld vector also vibrates with the
same frequency, but is always aligned perpendicular to the
electric ﬁeld vector as well as to the propagation direction.
Suppose that we ﬁx our attention on a certain plane that is
transverse to the propagation direction. On this plane, the
locus of the tip of the electric ﬁeld vector is the socalled
vibration ellipse, which is of the same shape as the locus of
the tip of the magnetic ﬁeld vector. A vibration ellipse is
shown in Fig. 1. Its shape is characterized by a tilt angle
as well as an axial ratio; in addition, it can be lefthanded
if the tip of the electric ﬁeld vector rotates counterclock
wise, or righthanded if otherwise. Similarly, an EM wave
is said to be elliptically polarized, in general; however, the
vibration ellipse can occasionally degenerate into a circle
(circular polarization) or even a straight line (linear
polarization).
3
2
1
0
1
2
0 2 4 6 8 10 12
N
o
r
m
a
l
i
z
e
d
g
a
i
n
(
d
B
)
Frequency (MHz)
R
g
= 2R
l
R
g
= R
l
No R
g
Figure 6. Expected passband errors for three 5thorder 1 dB
Chebyshev ﬁlters, for 5% random variations on all the element
values.
610 CHIRALITY
The shape of the vibration ellipse of monochromatic
light is altered after traversal through a certain thickness
of a socalled optically active medium. This phenomenon,
known as optical activity, was discovered around 1811 by
F. Arago while experimenting with quartz. Crystals are
generally anisotropic, but J.B. Biot observed around 1817
the optical activity of turpentine vapor, definitely an
isotropic medium. Isotropic organic substances were
believed to have exclusively biological provenances,
and in 1860 L. Pasteur argued that turpentine vapor
exhibited natural optical activity, but the optical activity
of crystals could not be similarly qualiﬁed. Pasteur was
unduly restrictive. Isotropic optically active media, of
biological or other origin, are nowadays called isotropic
chiral media, because EM ﬁelds excited in them necessar
ily possess a property called handedness (Greek cheir =
hand). Facsimile reproductions of several early papers are
available [2].
3. CHIRAL MEDIA: NATURAL AND ARTIFICIAL
The molecules of an isotropic chiral medium are mirror
asymmetric (i.e., they are noncongruent with their mirror
images). A chiral molecule and its mirror image are called
enantiomers [3]. As examples, the two enantiomers of
2butanol are shown in Fig. 2. Enantiomers can have dif
ferent properties, although they contain identical atoms in
identical numbers. One enantiomer of the chiral com
pound thalidomide may be used to cure morning sickness,
during pregnancy, but its mirror image induces fetal mal
formation. Aspartame, a common artiﬁcial sweetener, is
one of the four enantiomers of a dipeptide derivative. Of
these four, one (i.e., aspartame) is sweet, another is bitter,
while the remaining two are tasteless. Of the approxi
mately 1850 natural, semisynthetic, and synthetic drugs
marketed these days, no less than 1045 can exist as two or
more enantiomers; but only 570 were being marketed in
the late 1980s as single enantiomers, of which 61 were
totally synthetic. But since 1992, the U.S. Food and Drug
Administration (FDA) has insisted that only one
enantiomer of a chiral drug be brought into the market.
Biological chirospeciﬁcity, once the subject of speculations
by Pasteur on the nature of the life force (vis viva), is now
the topic of conferences on the origin of life [4].
An isotropic chiral medium is circularly birefringent
(i.e., both lefthanded and righthanded circularly polar
ized light can propagate in a region ﬁlled with a homoge
neous isotropic chiral medium, with different phase
velocities and attenuation rates). Therefore, when mono
chromatic, elliptically polarized light irradiates an isotro
pic chiral slab, the tilt angle and the axial ratio of the
transmitted light are different from those of the incident
light. The change in the tilt angle is quantiﬁed as optical
rotation (OR) and alteration of the axial ratio as circular
dichroism (CD). Both OR and CD depend on the wave
length l
0
, and the dependences are reasonably material
specific that spectroscopies based on their measurements
have long had industrial importance. Biot himself had pi
oneered these attempts by cataloging the OR spectra of a
large number of syrups and oils, and went on to found the
science of saccharimetry for which he was awarded the
Rumford Medal in 1840 by the Royal Society of London.
The ﬁrst edition of Landolt’s tables on optical activity
appeared in the German language in 1879; the
English translation of the second edition of 1898 appeared
in 1902.
Although Maxwell’s uniﬁcation of light with electro
magnetism during the third quarter of the nineteenth
century came to mean that natural optical activity is an
EM phenomenon, the term optical rotation persisted. By
the end of the nineteenth century, several empirical rules
had evolved on OR spectrums of isotropic chiral mediums.
Then, in the late 1890s, two accomplishments of note were
reported:
1. J. C. Bose constructed several artiﬁcial chiral mate
rials by twisting jute ﬁbers and laying them end to
end, and experimentally veriﬁed OR at millimeter
wavelengths. These materials were anisotropic, but
2a
2b
Tilt angle
x
y
Axial ratio = a/b
Figure 1. The tip of the electric ﬁeld vector of a planepolarized
monochromatic electromagnetic wave traces the so–called vibra
tion ellipse in a plane transverse to the propagation direction.
C
H
OH
CH
3
CH
2
CH
3
C
H
OH
CH
2
CH
3
CH
3
R–2–Butanol S–2–Butanol
Figure 2. The two enantiomers of 2butanol are mirror images of
each other, as shown by the directed circular arrangements of the
–OH, –CH
2
CH
3
, and –CH
3
groups.
CHIRALITY 611
Bose went on to infer from his experiments that
isotropic chiral materials could also be constructed
in the same way [5]. Thus, he conclusively demon
strated the geometric microstructural basis
for optical activity, and he also constructed possibly
the world’s ﬁrst artiﬁcial anisotropic chiral medium
to alter the vibration ellipses of microwaves.
2. P. Drude showed that chiral molecules can be mod
eled as spiral oscillators and theoretically veriﬁed a
rule Biot had given regarding OR spectra [6].
Experimental veriﬁcation of Drude’s spiral oscillator
hypothesis had to wait for another two decades. As elec
tromagnetic propositions can be tested at lower frequen
cies if the lengths are correspondingly increased and other
properties proportionally adjusted, K. F. Lindman made
2.5turn, 10mmdiameter springs from 9cmlong copper
wire pieces of 1.2mm crosssectional diameter. Springs
are handed, as illustrated in Fig. 3. Each spring was
wrapped in a cotton ball, and about 700 springs of the
same handedness were randomly positioned in a 2626
26cm cardboard box with an eye to achieving tolerable
isotropy. Then the box was irradiated with 1–3GHz
(30 cmZl
0
Z10 cm) microwave radiation and the OR was
measured. Lindman veriﬁed Drude’s hypothesis remark
ably well. He also determined that (1) the OR was pro
portional to the number of (identically handed) springs in
the box, given that the distribution of springs was rather
sparse; and (2) equal amounts of lefthanded or right
handed springs brought about the same OR, but in oppo
site senses [7]. Lindman’s experiments were extensively
repeated during the 1990s by many research groups in
several countries [8,9], and several patents have even
been awarded on making artiﬁcial isotropic chiral medi
ums with miniature springs.
4. CONSTITUTIVE RELATIONS OF AN ISOTROPIC
CHIRAL MEDIUM
Electromagnetic ﬁelds are governed by the Maxwell pos
tulates, in vacuum as well as in any material medium.
These four postulates have a microscopic basis and are
given in vacuum as follows:
V
.
~
BB(r; t) =0 (1a)
V
~
EE(r; t) = ÷
@
@t
~
BB(r; t) (1b)
e
0
V
.
~
EE(r; t) = ~ rr
tot
(r; t) (1c)
V
~
BB(r; t) =m
0
e
0
@
@t
~
EE(r; t) ÷m
0
~
JJ
tot
(r; t) (1d)
Thus,
~
EE(r; t) and
~
BB(r; t) are the primitive or the funda
mental EM ﬁelds, both functions of the threedimensional
position vector r and time t; e
0
=8.854 10
÷12
F/m and
m
0
=4p10
÷7
H/m are, respectively, the permittivity and
the permeability of vacuum; ~ rr
tot
(r; t) is the electric charge
density and
~
JJ
tot
(r; t) is the electric current density.
Equations (1) apply at any length scale, whereas the
charge and the current densities must be speciﬁed not
continuously but over a set of isolated points. Electromag
netically speaking, matter is nothing but a collection of
discrete charged particles in vacuum. As per the Heavi
side–Lorentz procedure to get a macroscopic description of
continuous matter, spatial averages of all ﬁelds and sourc
es are taken, while both ~ rr
tot
(r; t) and
~
JJ
tot
(r; t) are parti
tioned into matterderived and externally impressed
components. Then the Maxwell postulates at the macro
scopic level can be stated as
V
.
~
BB(r; t) =0 (2a)
V
~
EE(r; t) = ÷
@
@t
~
BB(r; t) (2b)
V
.
~
DD(r; t) = ~ rr(r; t) (2c)
V
~
HH(r; t) =
@
@t
~
DD(r; t) ÷
~
JJ(r; t) (2d)
Here, ~ rr(r; t) and
~
JJ(r; t) are the externally impressed source
densities, while the new ﬁelds
~
DD(r; t) =e
0
~
EE(r; t) ÷
~
PP(r; t) (3a)
~
HH(r; t) =
~
BB(r; t) ÷
~
MM(r; t)
m
0
(3b)
contain two matterderived quantities: the polarization
~
PP(r; t) and the magnetization
~
MM(r; t):
Constitutive relations must be prescribed to relate the
matterderived ﬁelds
~
DD(r; t) and
~
HH(r; t) to the basic ﬁelds
~
EE(r; t) and
~
BB(r; t) in any material medium. The construc
tion of these relations is primarily phenomenological,
although certain epistemologically mandated proprieties
Figure 3. An enantiomeric pair of springs. An artiﬁcial isotropic
chiral medium can be made by randomly dispersing springs in an
isotropic achiral host medium, with more springs of one handed
ness than the springs of the other handedness.
612 CHIRALITY
must be adhered to. The constitutive relations appropriate
for a general, linear, homogeneous, material medium with
timeinvariant response characteristics may be stated as
~
DD(r; t) =e
0
~
EE(r; t) ÷e
0
_
o
0
~ vv
e
(t)
.
~
EE(r; t ÷t)dt
÷
_
o
0
~ vv
em
(t)
.
~
BB(r; t ÷t)dt
(4a)
~
HH(r; t) =
1
m
0
~
BB(r; t) ÷
1
m
0
_
o
0
~ vv
m
(t)
.
~
BB(r; t ÷t)dt
÷
_
o
0
~ vv
me
(t)
.
~
EE(r; t ÷t)dt
(4b)
Four constitutive property kernels appear in these equa
tions; the dyadic ~ vv
e
(t) is the dielectric susceptibility ker
nel, ~ vv
m
(t) is the magnetic susceptibility kernel, while the
dyadics ~ vv
em
(t) and ~ vv
me
(t) are called the magnetoelectric
kernels. Although a dyadic may be understood as a 33
matrix for the purpose of this article, Chen’s textbook [10]
is recommended for a simple introduction to the use of
dyadics in EM theory.
All four dyadic kernels in Eqs. (4) are causal [i.e.,
~ vv
e
(t) ¬ 0 for tr0, etc.], because all materials must exhib
it delayed response. In addition, when we substitute
Eqs. (4a) and (4b) in Eqs. (2c) and (2d), respectively, a re
dundancy emerges with respect to Eqs. (2a) and (2b).
Elimination of this redundancy leads to the constraint [11]
Tr [ ~ vv
em
(t) ÷ ~ vv
me
(t)] ¬ 0 (5)
which has never been known to be violated by a physical
material. Finally, crystallographic symmetries may also
impose additional constraints on the constitutive kernels.
A medium described by Eqs. (4) is said to be bianisotropic,
since the constitutive kernels indicate anisotropy, and
both
~
DD(r; t) and
~
HH(r; t) depend on both
~
EE(r; t) and
~
BB(r; t):
Suppose next that the linear medium’s constitutive
properties are directionindependent. Equations (4) then
simplify to
~
DD(r; t) =e
0
~
EE(r; t) ÷e
0
_
o
0
~ ww
e
(t)
~
EE(r; t ÷t)dt
÷
_
o
0
~ ww
chi
(t)
~
BB(r; t ÷t)dt
(6a)
~
HH(r; t) =
1
m
0
~
BB(r; t) ÷
1
m
0
_
o
0
~ ww
m
(t)
~
BB(r; t ÷t)dt
÷
_
o
0
~ ww
chi
(t)
~
EE(r; t ÷t)dt
(6b)
in consequence of Eq. (5), where the scalar ~ ww
chi
(t) is the
chirality kernel. Equations (6a)–(6b) describe the isotropic
chiral medium—the most general, isotropic, linear elec
tromagnetic material known to exist [12,13].
Most commonly, EM analysis is carried out in the fre
quency domain, not the time domain. Let all timedepen
dent quantities be Fouriertransformed; thus
~
DD(r; t) =
1
2p
_
o
÷o
e
÷iot
D(r; o)do (7)
and so on, where o=2pf is the angular frequency. In the
remainder of this article, phasors such as D(r; o) are
called ﬁelds, following normal practice. The four Maxwell
postulates [Eqs. (2)] assume the form
V
.
B(r; o) =0 (8a)
VE(r; o) =ioB(r; o) (8b)
V
.
D(r; o) =r(r; o) (8c)
VH(r; o) = ÷ioD(r; o) ÷J(r; o) (8d)
while the constitutive equations [Eqs. 6] for an isotropic
chiral medium simultaneously transform into
D(r; o) =e
0
[1 ÷w
e
(o)]E(r; o) ÷w
chi
(o)B(r; o) (9a)
H(r; o) =
1
m
0
[1 ÷w
m
(o)]B(r; o) ÷w
chi
(o)E(r; o) (9b)
Using Eqs. (8b) and (8d) with J(r; o) =0 in Eqs. (9a) and
(9b), respectively, we obtain the Drude–Born–Fedorov
(DBF) constitutive relations of an isotropic chiral medium:
D(r; o) =e(o)[E(r; o) ÷b(o)VE(r; o)] (10a)
B(r; o) =m(o)[H(r; o) ÷b(o)VH(r; o)] (10b)
Their great merit is that the necessary mirror asymmetry
is transparently reﬂected in them, because VE(r; o) and
VH(r; o) are not true vectors but only pseudovectors. A
chiral medium is thus described by three constitutive
properties; the permittivity and permeability in
Eqs. (10a) and (10b), respectively, may be formally deﬁned
as the ratios
e(o) =
D(r; o)
.
E
+
(r; o)
[E(r; o)[
2
if E
+
(r; o)
.
[VE(r; o)] =0
(11a)
m(o) =
B(r; o)
.
H
+
(r; o)
[H(r; o)[
2
if H
+
(r; o)
.
[VH(r; o)] =0
(11b)
but the chirality parameter b(o) can be regarded as either
b(o) =
D(r; o)
.
[VE
+
(r; o)]
e(o)[VE(r; o)[
2
if E(r; o)
.
[VE
+
(r; o)] =0
(11c)
CHIRALITY 613
or
b(o) =
B(r; o)
.
[VH
+
(r; o)]
m(o)[VH(r; o)[
2
if H(r; o)
.
[VH
+
(r; o)] =0
(11d)
where the asterisk denotes the complex conjugate. Equa
tions (11) make it clear that while e(o) and m(o) are true
scalars, b(o) has to be a pseudoscalar since the numerator
in either of its two definitions contains a pseudovector.
Other constitutive relations—equivalent to Eqs. (9) and
Eqs. (10)—are also used in the frequencydomain EM lit
erature, but this article is restricted to the DBF cons
titutive relations [Eqs. (10)], as they bring out the essence
of chirality at the very ﬁrst glance. An isotropic chiral
medium and its mirror image share the same e(o) and
m(o), and their chirality parameters differ only in sign.
The timeaveraged Poynting vector
S(r; o) =
1
2
Re{E(r; o) H
+
(r; o)] (12a)
denotes the direction of power ﬂow. In any linear medium,
the monochromatic Poynting theorem reads as
V
.
S(r; o) = ÷
1
2
Re{E(r; o)
.
J
+
(r; o)]
÷
1
2
Re{io[E(r; o)
.
D
+
(r; o)
÷B(r; o)
.
H
+
(r; o)]]
(12b)
For specialization to an isotropic chiral medium, we have
to substitute Eqs. (10) in Eq. (12b). The resulting expres
sion is not particularly illuminating.
An isotropic chiral medium is Lorentzreciprocal. Sup
pose that all space is occupied by a homogeneous isotropic
chiral medium and all sources are conﬁned to regions of
bounded extent. Let sources labeled a radiate ﬁelds
E
a
(r; o) and H
a
(r; o); while sources labeled b radiate ﬁelds
E
b
(r; o) and H
b
(r; o); all at the same frequency. Then the
relations [12]
V
.
[E
a
(r; o) H
b
(r; o) ÷E
b
(r; o) H
a
(r; o)] =0 (13a)
V
.
[e(o)E
a
(r; o) E
b
(r; o) ÷m(o)H
a
(r; o)
H
b
(r; o)] =0
(13b)
arise in a sourcefree region, in consequence of the Lorentz
reciprocity of the medium.
5. ARTIFICIAL ISOTROPIC CHIRAL MEDIA
That matter is discrete has long been established. Fur
thermore, when we probe matter at length scales at which
it appears continuous, whether the microstructure is mo
lecular or merely comprises electrically small inclusions is
of no consequence. The linear dimensions of an electrically
small inclusion are less than about a tenth of the maxi
mum wavelength, in the media outside as well as inside
the inclusion, at a particular frequency. Artiﬁcial isotropic
chiral media—active at microwave frequencies—can be
constructed with this thought in mind. Consider a random
suspension of identical, electrically small, inclusions in a
host medium, which we take here to be vacuum for sim
plicity. The number of inclusions per unit volume is de
noted by N, and the volumetric proportion of the
inclusions in the composite medium is assumed to be
very small. Our objective is to homogenize this dilute par
ticulate composite medium and estimate its effective cons
titutive properties [13]. Homogenization is much
like blending apples into apple sauce or tomatoes into
ketchup.
Any inclusion scatters the EM wave incident on it.
Far away from the inclusion, the scattered EM ﬁeld
phasors can be conceptualized, equivalently, as being
radiated by an ensemble of multipoles. Multipoles are
necessarily frequencydomain entities; and adequate
descriptions of electrically larger inclusions require high
erorder multipoles, but homogenizing composite media
with electrically large inclusions is fraught with concep
tual perils.
The lowestorder multipoles are the electric dipole p
and the magnetic dipole m. In formalisms for isotropic
chiral media, both are accorded the same status. As all
inclusions in our composite medium are electrically small,
we can think that an inclusion located at position r
/
is
equivalent to the colocated dipoles characterized by the
following relations:
p
eqvt
(r
/
; o) =p
ee
(o)
.
E
exc
(r
/
; o) ÷p
eh
(o)
.
H
exc
(r
/
; o) (14a)
m
eqvt
(r
/
; o) =p
he
(o)
.
E
exc
(r
/
; o) ÷p
hh
(o)
.
H
exc
(r
/
; o) (14b)
Here, E
exc
(r
/
; o) and H
exc
(r
/
; o) are the ﬁelds exciting the
particular inclusion; while p
ee
(o); p
eh
(o); p
he
(o); and
p
hh
(o) are the four linear polarizability dyadics that de
pend on the frequency, the constitution, and the dimen
sions of the inclusion. As the inclusions are randomly
oriented and any homogenizable chunk of a composite
medium contains a large number of inclusions, p
ee
(o) and
other terms in Eqs. (14) can be replaced by their orienta
tionally averaged values. If the homogenized composite
medium is isotropic chiral, this orientational averaging
process must yield
p
eqvt
(r
/
; o) =N[p
ee
(o)E
exc
(r
/
; o) ÷ip
chi
(o)H
exc
(r
/
; o)] (15a)
m
eqvt
(r
/
; o) =N[÷ip
chi
(o)E
exc
(r
/
; o) ÷p
hh
(o)H
exc
(r
/
; o)]
(15b)
The polarizability dyadics of electrically small, handed in
clusions (e.g., springs) may be computed either with stan
dard scattering methods such as the method of moments
[14] or using lumpedparameter circuit models [15]. Pro
vided that dissipation in the composite medium can be
ignored, at a certain angular frequency, p
ee
(o); p
hh
(o);
and p
chi
(o) are purely realvalued.
614 CHIRALITY
On applying the Maxwell Garnett homogenization
approach, the constitutive relations of the homogenized
composite medium (HCM) are estimated as follows [12]:
D(r; o) =t
ee
(o)E(r; o) ÷t
chi
(o)H(r; o) (16a)
B(r; o) =t
hh
(o)H(r; o) ÷t
chi
(o)E(r; o) (16b)
where
t
ee
(o) =e
0
÷
[9m
0
Np
ee
÷3N
2
(p
2
chi
÷p
ee
p
hh
)]
9e
0
m
0
÷3N(e
0
p
hh
÷m
0
p
ee
) ÷N
2
(p
2
chi
÷p
ee
p
hh
)
(16c)
t
hh
(o) =m
0
÷
[9e
0
Np
hh
÷3N
2
(p
2
chi
÷p
ee
p
hh
)]
9e
0
m
0
÷3N(e
0
p
hh
÷m
0
p
ee
) ÷N
2
(p
2
chi
÷p
ee
p
hh
)
(16d)
t
chi
(o) =
i9e
0
m
0
Np
chi
9e
0
m
0
÷3N(e
0
p
hh
÷m
0
p
ee
) ÷N
2
(p
2
chi
÷p
ee
p
hh
)
(16e)
Equivalently
D(r; o) =e
HCM
(o)[E(r; o) ÷b
HCM
(o)VE(r; o)] (17a)
B(r; o) =m
HCM
(o)[H(r; o) ÷b
HCM
(o)VH(r; o)] (17b)
are the DBF constitutive relations of the HCM, with
e
HCM
(o) =
t
ee
(o)t
hh
(o) ÷t
2
chi
(o)
t
hh
(o)
(17c)
m
HCM
(o) =
t
ee
(o)t
hh
(o) ÷t
2
chi
(o)
t
ee
(o)
(17d)
b
HCM
(o) = ÷
i
o
t
chi
(o)
t
ee
(o)t
hh
(o) ÷t
2
chi
(o)
(17e)
as the constitutive parameters. Clearly, if p
chi
(o)O0;
the composite medium has been homogenized into an
isotropic chiral medium. In passing, other homogeniza
tion approaches are also possible for chiral composites
[1,13].
6. BELTRAMI FIELDS IN AN ISOTROPIC CHIRAL MEDIUM
In a sourcefree region occupied by a homogeneous
isotropic chiral medium, r(r; o) =0 and J(r; o) =0:
Equations (8a), (8c), and (10) then show that
V
.
E(r; o) =0 and V
.
H(r; o) =0: Thus all four ﬁelds—
E(r; o); H(r; o); D(r; o); and B(r; o)—are purely solenoi
dal. Next, Eqs. (8) and (10) together yield the following
vector Helmholtzlike equations:
V
2
E(r; o)
H(r; o)
D(r; o)
B(r; o)
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
÷2
o
2
e(o)m(o)b(o)
1 ÷o
2
e(o)m(o)b
2
(o)
V
E(r; o)
H(r; o)
D(r; o)
B(r; o)
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
÷
o
2
e(o)m(o)
1 ÷o
2
e(o)m(o)b
2
(o)
E(r; o)
H(r; o)
D(r; o)
B(r; o)
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
=
0
0
0
0
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
(18)
In the limit b(o) ÷0; the medium becomes achiral and
these equations reduce to the familiar vector Helmholtz
equation, V
2
E(r; o) ÷o
2
e(o)m(o)E(r; o) =0; and so on.
In lieu of the secondorder differential equations [Eqs.
(18)], ﬁrstorder differential equations can be formulated.
Thus, after deﬁning the auxiliary ﬁelds
Q
1
(r; o) =
1
2
E(r; o) ÷i
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
m(o)
e(o)
¸
H(r; o)
_ _
(19a)
Q
2
(r; o) =
1
2
H(r; o) ÷i
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
e(o)
m(o)
¸
E(r; o)
_ _
(19b)
and using the wavenumbers
g
1
(o) =
o
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
e(o)m(o)
_
1 ÷ob(o)
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
e(o) m(o)
_ (20a)
g
2
(o) =
o
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
e(o)m(o)
_
1 ÷ob(o)
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
e(o) m(o)
_ (20b)
we get the two ﬁrstorder differential equations
VQ
1
(r; o) =g
1
(o)Q
1
(r; o) (21a)
VQ
2
(r; o) = ÷g
2
(o)Q
2
(r; o) (21b)
which are easier to analyze than Eqs. (18). The denomi
nators on the left sides of Eqs. (20) suggest that
o
2
e(o)m(o)b
2
(o) =1 is not permissible for an isotropic
chiral medium, as both wavenumbers must have ﬁnite
magnitudes.
According to Eqs. (21), Q
1
(r; o) and Q
2
(r; o) are
Beltrami ﬁelds [12]. A Beltrami ﬁeld is parallel to its
own circulation. The concept arose early in the nineteenth
century, and has often been rediscovered. The easiest
way to think of a Beltrami ﬁeld is as a spiral staircase or
a tornado.
CHIRALITY 615
While Q
1
(r; o) is a lefthanded Beltrami ﬁeld, the neg
ative sign on the right side of Eq. (21b) means that Q
2
(r; o)
is a righthanded Beltrami ﬁeld, because the two complex
valued wavenumbers g
1
(o) and g
2
(o) must have positive
real parts. Both wavenumbers also must have positive
imaginary parts in a causal material medium, since causal
materials must exhibit delayed response in the time do
main and therefore must demonstrate EM loss (or atten
uation) in the frequency domain.
As an isotropic chiral medium displays two distinct
wavenumbers at a specific frequency, it is birefringent.
More specifically, because Q
1
(r; o) and Q
2
(r; o): have
planewave representations possible only in terms of cir
cularly polarized plane waves, an isotropic chiral medium
is often said to be circularly birefringent. The difference
between g
1
(o) and g
2
(o) gives rise to natural optical ac
tivity. While OR is proportional to the real part of [g
1
(o) ÷
g
2
(o)]; CD is proportional to the imaginary part of [g
1
(o) ÷
g
2
(o)]: The OR and CD spectra must be consistent with the
Kramers–Kronig relations [16]. The CD spectrum has a
local maximum or minimum at the frequency where the
sign of the OR changes; this feature is labeled as the Cot
ton effect after H. Cotton, who reported it in 1895 [2]. The
OR and CD spectra of a simple chiral medium are illus
trated in Fig. 4.
7. REPRESENTATION OF BELTRAMI FIELDS
A Beltrami ﬁeld is represented in terms of toroidal and
poloidal ﬁelds because the curl of a toroidal ﬁeld is poloidal
and vice versa [17]. Thus, the decomposition
Q
n
(r; o) =g
n
(o)V[rc
n
(r; o)] ÷(÷)
n ÷1
VV[rc
n
(r; o)];
n =1; 2 (22)
is possible, as the ﬁrst parts on the right sides of Eqs. (22)
are toroidal and the second parts are poloidal. The
scalar functions c
n
(r; o) satisfy the scalar Helmholtz
equation as follows:
V
2
c
n
(r; o) ÷g
2
n
(o)c
n
(r; o) =0; n =1; 2 (23)
Solutions of Eqs. (23) in the Cartesian, the circular cylin
drical, and the spherical coordinate systems are common
place [18].
Beltrami plane waves propagating in the ÷z direction
may be represented as
Q
n
(r; o) =A
n
2
÷1=2
[ ^ xx÷(÷)
n ÷1
i ^ yy] exp[ig
n
(o)z]; n =1; 2 (24)
with A
n
as the amplitudes, while ^ xx; ^ yy, and ^ zz are the Car
tesian unit vectors.
In the circular cylindrical coordinate system (r; j; z);
Beltrami ﬁelds with an exp(iaz) dependence may be ex
pressed as the sums
Q
n
(r; o) =
o
n=÷o
A
nn
[M
(3)
n
(g
n
(o)[a; r) ÷(÷)
n ÷1
N
(3)
n
(g
n
(o)[a; r)];
n =1; 2 (25a)
for regular behavior as r ÷o; while the expansions
Q
n
(r; o) =
o
n=÷o
B
nn
[M
(1)
n
(g
n
(o)[a; r) ÷(÷)
n ÷1
N
(1)
n
(g
n
(o)[a; r)];
n =1; 2 (25b)
are well behaved at r=0; with A
nn
and B
nn
as the coefﬁ
cients of expansion. The vector cylindrical wavefunctions
are given as
M
(1)
n
(s[ a; r) =e
i(az ÷nj)
^ qq
in
kr
_ _
J
n
(kr) ÷ ^ uu@J
n
(kr)
_ _
(26a)
M
(3)
n
(s[ a; r) =e
i(az ÷nj)
^ qq
in
kr
_ _
H
(1)
n
(kr) ÷ ^ uu@H
(1)
n
(kr)
_ _
(26b)
N
(j)
n
(s [ a; r) =
1
s
VM
(j)
n
(s [ a; r); j =1; 3 (26c)
where k= ÷(s
2
÷a
2
)
1=2
; ^ qq; ^ uu; and ^ zz are the unit vectors in
the cylindrical coordinate system; J
n
(kr) are the cylindri
cal Bessel functions of order n, and @J
n
(kr) are the re
spective ﬁrst derivatives with respect to the argument;
while H
(1)
n
(kr) are the cylindrical Hankel functions of the
ﬁrst kind and order n, and @H
(1)
n
(kr) are the ﬁrst deriva
tives with respect to the argument. For quasitwodimen
sional problems, a =0 because @=@z =0: Parenthetically, in
this paragraph r denotes the radial distance in the xy
plane and should not be confused with the use of r for
charge density elsewhere in this article.
Finally, with A
nsmn
and B
nsmn
as the coefﬁcients of ex
pansion, in the spherical coordinate system (r, y, j), we
CD
OR
f 0
Figure 4. Optical rotation (OR) and circular dichroism (CD)
spectra of a simple isotropic chiral medium. When the OR chang
es sign, the CD records either a maximum or a minimum, which
phenomenon is called the Cotton effect.
616 CHIRALITY
have
Q
n
(r; o) =
2
s =1
o
n=1
n
m=0
A
nsmn
[M
(3)
smn
(g
n
(o)r)
÷(÷)
n ÷1
N
(3)
smn
(g
n
(o)r)]; n =1; 2
(27a)
for ﬁelds regular as r ÷o; and
Q
n
(r; o) =
2
s =1
o
n=1
n
m=0
B
nsmn
[M
(1)
smn
(g
n
(o)r)
÷(÷)
n ÷1
N
(1)
smn
(g
n
(o)(r)]; n =1; 2
(27b)
for ﬁelds regular at r =0. The wellknown vector spherical
wavefunctions, M
(j)
smn
(sr) and N
(j)
smn
(sr); are stated for j =1,
3 as
M
(1)
smn
(sr) = ÷[n(n÷1)]
1=2
j
n
(sr)^ rr B
smn
(y; j) (28a)
M
(3)
smn
(sr) = ÷[n(n÷1)]
1=2
h
(1)
n
(sr)^ rr B
smn
(y; j) (28b)
N
(j)
smn
(sr) =
1
s
VM
(j)
smn
(sr); j =1; 3 (28c)
where the angular functions
B
1mn
(y; j) =[n(n÷1)]
÷1=2
^
hh
d
dy
P
m
n
(cos y) sinmj
_
÷ ^ uu
m
siny
P
m
n
(cos y) cos mj
_
(29a)
B
2 mn
(y; j) =[n(n÷1)]
÷1=2
^
hh
d
dy
P
m
n
(cos y) cos mj
_
÷^ uu
m
siny
P
m
n
(cos y) sinmj
_
(29b)
have been used. In these expressions, ^ rr;
^
hh; and ^ uu are the
unit vectors in the spherical coordinate system; P
m
n
(cos y)
are the associated Legendre functions of order n and de
gree m; j
n
(sr) are the spherical Bessel functions of order n;
and h
(1)
n
(sr) are the spherical Hankel functions of the ﬁrst
kind and order n.
Boundaryvalue problems involving scattering by iso
tropic chiral halfspaces, cylinders, and spheres can be
analytically solved using Eqs. (24)–(29). Boundaryvalue
problems involving more complicated geometries general
ly require numerical treatment, which necessitates the
use of Green functions.
Isotropic chiral waveguides for use at microwave
frequencies have been theoretically studied extensively,
although no practical realization thereof has yet come
to light. Theoretical investigations on propagation in
the socalled chirowaveguides generally consist of decom
posing the Beltrami ﬁelds into axial and transverse
components as
Q
n
(r; o) =Q
nt
(r; o) ÷ ^ zzQ
nz
(r; o);
^ zz
.
Q
nt
(r; o) ¬ 0; n =1; 2
(30)
where the z coordinate is measured on the waveguide axis
while two other mutually orthogonal coordinates are spec
iﬁed in the transverse plane. Assuming that all ﬁelds have
an exp(iaz) dependence on z, and making use of Eqs. (21),
we get
Q
nt
(r; o) =
1
g
2
n
(o) ÷a
2
[ia I ÷(÷)
n
g
n
(o) ^ zz I
.
V ÷ ^ zz
@
@z
_ _
Q
nz
(r; o); n =1; 2
(31)
where I is the identity dyadic. The axial components sat
isfy the reduced scalar Helmholtz equations
V
2
÷
@
2
@z
2
÷g
2
n
(o) ÷a
2
_ _
Q
nz
(r; o) =0; n =1; 2 (32)
appropriate solutions of which are commonly worked out
in many different ways for waveguides of different cross
sectional geometries [18].
8. SOURCES IN AN ISOTROPIC CHIRAL MEDIUM
Let us now assume the existence of a magnetic charge
density r
m
(r; o) and a magnetic current density J
m
(r; o);
because they assist in the solution of dual problems [19].
In addition, let us deﬁne the intrinsic impedance
Z(o) =
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
m(o)=e(o)
_
as well as the auxiliary wavenumber
k(o) =o
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
m(o) e(o)
_
; and drop the explicit indication of de
pendences on o for notational simplicity. Now Eqs. (8) may
be written as
V
.
B(r) =r
m
(r) (33a)
VE(r) =ioB(r) ÷J
m
(r) (33b)
V
.
D(r) =r(r) (33c)
VH(r) = ÷ioD(r) ÷J(r) (33d)
which yield the relations
VQ
n
(r) ÷(÷)
n
g
n
Q
n
(r) =W
n
(r); n =1; 2 (34)
for a chiral medium, where
W
1
(r) =
g
1
2k
[iZ J(r) ÷J
m
(r)] (35a)
W
2
(r) =
g
2
2k
[J(r) ÷
1
iZ
J
m
(r)] (35b)
are the Beltrami source current densities [12].
CHIRALITY 617
Since Eqs. (34) are linear, they can be solved using
standard techniques. Their complete solution can be com
pactly stated for all r as
Q
n
(r) =Q
cf
n
(r) ÷Q
rad
n
(r); n =1; 2 (36)
where
Q
rad
n
(r) =(÷)
n ÷1
2g
1
g
2
k
_
V
s
G
n
(r; r
0
)
.
W
n
(r
0
)d
3
r
0
;
n =1; 2
(37)
are the particular solutions due to the source densities W
n
(r);
which are wholly conﬁned to the region V
s
, and Q
cf
n
(r) are
the complementary functions satisfying the relations
VQ
cf
n
(r) =(÷)
n ÷1
g
n
Q
cf
n
(r); n =1; 2 (38)
identically. Substituting Eqs. (36)–(38) in Eqs. (34), we ob
tain the dyadic differential equations
VG
n
(r; r
0
) ÷(÷)
n
g
n
G
n
(r; r
0
) =(÷)
n ÷1
2g
1
g
2
k
_ _
÷1
Id(r ÷r
0
); n =1; 2
(39)
where d(
.
) is the Dirac delta function.
The solutions of Eqs. (39) are the Beltrami–Green dy
adic functions
G
n
(r; r
0
) =(÷)
n ÷1
2g
1
g
2
k
_ _
÷1
[VI ÷(÷)
n ÷1
g
n
I]
.
G
fs
(g
n
[r; r
0
); n =1; 2
(40)
wherein
G
fs
(s [ r; r
0
) = I ÷
VV
s
2
_ _
exp(is[r ÷r
0
[)
4p[r ÷r
0
[
(41)
is the familiar dyadic Green function for free space. As the
properties of G
fs
(s; r; r
0
) can be found in almost any grad
uatelevel EM textbook [20,21], those of G
n
(r; r
0
) can be
easily determined, as illustrated in Ref. 12.
As an example of the use of Eqs. (37), let us consider
an electric dipole moment p located at the origin:
J(r) = ÷iopd(r) and J
m
(r) =0: The radiated Beltrami
ﬁelds turn out be
Q
rad
1
(r) =
o
2
m
k
g
1
g
2
k
2
g
1
G
1
(r; 0)
.
p; r > 0 (42a)
Q
rad
2
(r) =io
g
1
g
2
k
2
g
2
G
2
(r; 0)
.
p; r > 0 (42b)
which show clearly that the radiation ﬁeld of a point elec
tric dipole in an isotropic chiral medium consists of left
handed as well as righthanded components. If we have
instead a point magnetic dipole m located at the origin,
the source current densities are speciﬁed as J(r) =0 and
J
m
(r) = ÷iomd(r); so that
Q
rad
1
(r) =io
g
1
g
2
k
2
g
1
G
1
(r; 0)
.
m; r > 0 (43a)
Q
rad
2
(r) =
o
2
e
k
g
1
g
2
k
2
g
2
G
2
(r; 0)
.
m; r > 0 (43b)
are the corresponding radiated Beltrami ﬁelds. A major
difference between isotropic chiral and achiral media is
shown by the two sets of radiated ﬁelds, Eqs. (42) and (43).
Without loss of generality, let the source dipole moments
be aligned parallel to the z axis. Then, if the dipole mo
ments are radiating in an achiral medium (i.e., b=0),
there is no magnetic ﬁeld due to p and there is no electric
ﬁeld due to m at any point on the z axis. On the other
hand, the wavenumber difference between the lefthanded
and the righthanded Beltrami ﬁelds guarantees that, in
an isotropic chiral medium, both E
rad
(r) and H
rad
(r) are
not generally nullvalued on the z axis, regardless of
which one of the two dipole moments is radiating.
Canonical sources of Beltrami ﬁelds are possible. If there
is a source distribution such that J(r) ¬ ÷(1=iZ)J
m
(r) for all
r, then Q
rad
2
(r) ¬ 0 from Eqs. (35) and (37). Likewise, a
source distribution containing electric and magnetic cur
rent densities in the simple proportion J(r) =(1=iZ)J
m
(r)
for all r radiates only a righthanded ﬁeld, because
Q
rad
1
(r) ¬ 0 emerges from the same equations.
Radiation by complex sources has to be generally treat
ed using integral equations. Both the Maue and the Pock
lington integral equations for radiation in a homogeneous
isotropic chiral medium are available [12]. Cerenkov ra
diation in an isotropic chiral medium has also been de
scribed using Beltrami ﬁelds [12].
The foregoing developments make it clear that a descrip
tion involving differentials of only the ﬁrst order sufﬁces for
monochromatic radiation and propagation in an isotropic
chiral medium. True, there are VV terms in G
1
(r; r
0
) and
G
2
(r; r
0
); but dyadic Green functions are not ﬁelds, being
instead solutions of dyadic differential equations.
Finally, although the lefthanded and the righthanded
Beltrami ﬁelds are capable of being independently
radiated and propagated as per Eqs. (34), they do indeed
couple in an isotropic chiral medium. This coupling
takes place only at bimedium boundaries where condi
tions on the tangential components of E(r) and H(r)
must be satisﬁed; that is, the boundary conditions are
speciﬁed not on Q
1
(r) or Q
2
(r) singly, but on the tangential
components of the combinations E(r) =Q
1
(r) ÷iZ Q
2
(r)
and H(r) =Q
2
(r) ÷(1=iZ) Q
1
(r):
9. THEOREMS FOR SCATTERING IN AN ISOTROPIC
CHIRAL MEDIUM
Equations (36)–(39) sufﬁce to set up certain oftenused
principles for monochromatic scattering and radiation
problems, when all space is ﬁlled with a homogeneous iso
tropic chiral medium.
The source–region Beltrami ﬁelds can be obtained from
Eqs. (37) using the Fikioris approach [22]. Let S be the
618 CHIRALITY
surface of the convexshaped source region V
s
, where ^ nn
0
is
the unit outward normal at r
0
c S (see Fig. 5). Then, Eqs.
(37) and (40) yield the following relations:
Q
rad
n
(r) =(÷)
n ÷1
g
n
_
V
s
[G
fs
(g
n
[r; r
0
)
.
W
n
_
(r
0
)
÷G
P
(g
n
[r; r
0
)
.
W
n
(r)]d
3
r
0
÷g
÷2
n
L(r)
.
W
n
(r)
_
÷
_
V
s
[VG
fs
(g
n
[r; r
0
)]
.
W
n
(r
0
)d
3
r
0
;
n =1; 2; r c V
s
(44)
The depolarization dyadic
L(r) =
1
4p
_
S
^ nn
0
r
0
÷ ^ nn
0
r
[r ÷r
0
[
3
d
2
r
0
(45)
in Eqs. (44) is dependent on the shape of the region V
s
, while
G
P
(s[r; r
0
) =
VV
s
2
1
4p[r ÷r
0
[
(46)
is an auxiliary dyadic function.
If the maximum linear extent of the region V
s
times the
magnitude of the greater of the two wavenumbers, g
1
and
g
2
, is much smaller than unity, we may make the quasi
static approximation: W
1
(r
0
)DW
1
(r) and W
2
(r
0
)DW
2
(r)
for all r
0
c V
s
: Then, Eqs. (44) simplify to
Q
rad
n
(r) ﬃ{(÷)
n ÷1
g
n
[M(g
n
[r) ÷g
÷2
n
L(r)]
÷N(g
n
[r)]
.
W
n
(r); n =1; 2; r c V
s
(47)
where the dyadics
M(s[r) =
_
V
s
[G
fs
(s[r; r
0
) ÷G
P
(s[r; r
0
)]d
3
r
0
(48a)
N(s[r) =
_
V
s
[VG
fs
(s[r; r
0
)]d
3
r
0
(48b)
depend on the shape as well as on the size of V
s
. Finally,
the Rayleigh approximation requires that we ignore the
dyadics M(s[r) and N(s[r) completely to obtain the
estimates
Q
rad
n
(r) ﬃ(÷)
n
g
÷1
n
L(r)
.
W
n
(r); n =1; 2; r c V
s
(49)
when V
s
is an extremely small region. The right sides of
Eqs. (47) and (49) are useful in homogenizing isotropic
chiral composites as well as for devising the method of
moments and the coupled dipole method for scattering by
bianisotropic objects in isotropic chiral environments
[12,23].
Turning now to the mathematical realizations of the
Huygens principle and its progeny, we suppose that all
space is divided into two regions, as shown in Fig. 6.
The external region V
ext
extends to inﬁnity in all direc
tions but is separated from an internal region V
int
by the
convex and oncedifferentiable surface S. Then the Huy
gens principle in a homogeneous isotropic chiral medium
reads as follows [12]:
Q
n
(r) =(÷)
n ÷1
2g
1
g
2
k
_
S
G
n
(r; r
0
)
.
[ ^ nn
0
Q
n
(r
0
)]d
2
r
0
;
n =1; 2; r c V
ext
(50a)
0=
_
S
G(r; r
0
)
.
[ ^ nn
0
Q
n
(r
0
)]d
2
r
0
; n =1; 2; r = cV
ext
(50b)
Thus, the Cauchy data for the ﬁelds in a chiral medium
comprise the components of the Beltrami ﬁelds that are
tangential to a boundary. When these data are prescribed
on the surface S, we can ﬁnd the Beltrami ﬁelds every
where in the region V
ext
.
The Huygens principle allows the enunciation of the
exterior surface equivalence principle. Consider a problem
in which surface Beltrami current densities W
s
1
(r) and
W
s
2
(r) exist on the exterior side of the surface S (see Fig. 6).
As per Eqs. (37), these surface current densities act as
V
s
S
n
o
^
Figure 5. For the evaluation of ﬁelds in the region V
s
, when the
sources are also conﬁned to the same region and all space is oc
cupied by a homogeneous chiral medium.
V
int
V
ext
S
n
o
^
Figure 6. Relevant to the Huygens principle, the exterior surface
equivalence principle, and the Ewald–Oseen extinction theorem,
when all space is occupied by a homogeneous chiral medium.
CHIRALITY 619
sources of the radiated ﬁelds
Q
rad
n
(r) =(÷)
n ÷1
2g
1
g
2
k
_
S
G
n
(r; r
0
)
.
W
s
n
(r
0
)d
2
r
0
;
n =1; 2; r c V
ext
(51)
On comparing Eqs. (50a) and (51) to ensure the equiva
lence Q
rad
n
(r) ¬ Q
n
(r) for all r c V
ext
; we obtain the rela
tionships [12]
W
s
n
(r
0
) = ^ nn
0
Q
n
(r
0
); n =1; 2; r
0
c S (52)
as the exterior surface equivalence principle for Beltrami
ﬁelds and sources, r
0
in Eqs. (52) lying on the exterior side
of S.
The Ewald–Oseen extinction theorem is a cornerstone
of the extendedboundarycondition method [12,24]. For
scattering in an isotropic chiral medium, this theorem
may be stated as
0=Q
cf
n
(r) ÷(÷)
n ÷1
2g
1
g
2
k
_
S
G
n
(r; r
0
)
.
[ ^ nn
0
Q
n
(r
0
)]d
2
r
0
;
n =1; 2; r c V
int
(53)
where Q
cf
n
(r) play the role of the incident Beltrami ﬁelds.
Once Q
n
(r
0
); r
0
c S; have been determined from Eqs. (53),
the total ﬁelds in the exterior region may be determined as
Q
n
(r) =Q
cf
n
(r) ÷(÷)
n ÷1
2g
1
g
2
k
_
S
G
n
(r; r
0
)
.
[ ^ nn
0
Q
n
(r
0
)]d
2
r
0
;
n =1; 2; r c V
ext
(54)
From Eqs. (53) and (54), the planewave scattering dyadics
for an object in an isotropic chiral environment can be de
rived, as can the forward planewave scattering amplitude
theorems [12].
10. STRUCTURALLY CHIRAL MEDIA
The molecules of a naturally occurring isotropic chiral
medium are mirrorasymmetric, and so are the inclusions
in an artiﬁcial isotropic chiral medium. As a randomly
dispersed and randomly oriented collection of mirror
asymmetric molecules or inclusions is also mirrorasym
metric, isotropic chiral media emerge with directioninde
pendent constitutive properties. In contrast, the molecules
or inclusions of a structurally chiral medium are not mir
rorasymmetric, but their orientation is.
In chiral nematic liquid crystals (CNLCs)—also called
cholesteric liquid crystals—needlelike molecules are
randomly positioned on parallel sheets, with all molecules
on any one sheet oriented parallel to one another and
with the orientation rotating helicoidally as one moves
across consecutive sheets. The situation is schematically
depicted in Fig. 7. From 1850 to 1888, several scientists
came across CNLCs but were unable to capitalize on
their observations [25]. Then in 1888 the biochemist
F. Reinitzer observed that a CNLC named cholesteryl
benzoate has two distinct melting points—it is a solid at
temperatures below 145.51C, a clear liquid at tempera
tures above 178.51C, and a cloudy liquid in between. Re
initzer’s observation of the mesophase—when positional
order is absent as in a liquid, but orientational order is
still strong as in a solid—opened up the area of liquid
crystal research in continuum mechanics as well as in op
tics [26–28].
Earlier, however, (in 1869), E. Reusch had anticipated
the CNLC structure as a laminate of uniaxial dielectric
sheets, with the crystallographic axes of any two adjacent
sheets offset in the transverse plane by a ﬁxed small angle.
At a low enough frequency, this laminate appears as a
continuously nonhomogeneous medium whose cons
titutive properties vary helicoidally. Thus
D(r; o) =e
0
S(z)
.
e
ref
(o)
.
S
÷1
(z)
.
E(r; o) (55a)
H(r; o) =
1
m
0
B(r; o) (55b)
x
y
z
Figure 7. Schematic depiction of the arrangement of needlelike
molecules in a chiral nematic liquid crystal. The gaps between the
consecutive sheets as well as the sheets are ﬁctitious, as they are
merely aids to visualization. Only half of the electromagnetic
period is shown.
620 CHIRALITY
are the frequencydomain constitutive relations of a
CNLC, where
e
ref
(o) =e
a
(o)[I ÷ ^ xx^ xx] ÷e
b
(o) ^ xx^ xx (56)
is the relative permittivity dyadic in a reference plane
designated as z =0. The rotation dyadic
S(z) =[ ^ xx^ xx÷ ^ yy^ yy] cos
pz
O
±[ ^ yy^ xx ÷ ^ xx^ yy] sin
pz
O
÷ ^ zz^ zz (57)
denotes that the CNLC structure varies helicoidally in the
axial (i.e., z) direction with a period 2 O; however, the elec
tromagnetic period is O. The upper sign in Eq. (57) applies
for structural righthandedness; the lower, for structural
lefthandedness.
Reusch’s model of a CNLC has been often implemented
with either uniaxial crystals or ﬁbrous laminae, and ap
pears promising for microwave and RF applications as
well [29]. More recently, thinﬁlm technology has been
pressed into service to realize the CNLC structure by re
leasing a directed evaporant ﬂux toward a rotating sub
strate [30,31]. The reference permittivity dyadic of these
chiral sculptured thin ﬁlms (STFs) differs from Eq. (56),
being
e
ref
(o) =e
a
(o)[I ÷( ^ xxcos w ÷ ^ zz sinw)( ^ xxcos w ÷ ^ zz sinw) ÷ ^ yy^ yy]
÷e
b
(o)( ^ xxcos w ÷ ^ zz sinw)( ^ xxcos w ÷ ^ zz sinw) ÷e
c
(o) ^ yy^ yy;
w > 0
(58)
instead, and the electromagnetic period is 2 O.
The reference permittivity dyadics in Eqs. (56) and (58)
are uniaxial and biaxial, respectively; that is, they have
either one or two crystallographic axes. Biaxial e
ref
(o) is
displayed by chiral smectic liquid crystals also [26,27].
Thus in general e
ref
(o) displays orthorhombic symmetry
[32]. Moreover, particularly with advances in thinﬁlm
technology, there is no reason for a chiral STF to be
necessarily dielectric only. These considerations led to
the proposal of the helicoidal bianisotropic medium
(HBM), whose frequencydomain constitutive relations
may be stated as [33]
D(r; o) =e
0
S(z)
.
[I ÷v
e
ref
(o)]
.
S
÷1
(z)
.
E(r; o) ÷S(z)
.
v
em
ref
(o)
.
S
÷1
(z)
.
B(r; o)
(59a)
H(r; o) =
1
m
0
S(z)
.
[I ÷v
m
ref
(o)]
.
S
÷1
(z)
.
B(r; o) ÷S(z)
.
v
me
ref
(o)
.
S
÷1
(z)
.
E(r; o)
(59b)
subject to the constraint
Tr[v
em
ref
(o) ÷v
me
ref
(o)] =0 (60)
The launching and propagation of EM waves in HBMs
is best studied using a 44 matrix differential equation
formalism [31,34].
Although chiral STFs made of ﬂuorites, and single
frequency OR measurements on them, were reported in
1959 [35], systematic experimental studies—along with
scanning electron microscopic veriﬁcation of the micro
structural geometry—appear to have begun only in 1995
[30]. Figure 8 shows the scanning electron micrograph
of a chiral STF made of silicon oxide. As typical values
of O realized today range from 30 nm to 10 mm, microwave
applications of these ﬁlms are yet not feasible, but are
likely to become an active area of research once ﬁlms with
OB100mm become available. Many possible applications
have been anticipated as the concept of STFs for biologi
cal, optical, electronic, chemical, and other applications
is beginning to take root, while many optical and related
applications have already been implemented [30,31].
Magn
8779x
Wd
3.5 3xxxxxxx803
2 µm
Figure 8. Scanning electron micrograph of a 10
period chiral sculptured thin ﬁlm made of silicon
oxide. (From Professor Russell Messier, Pennsylva
nia State University, with permission.)
CHIRALITY 621
Largescale production appears feasible as well, with
adaptation of ionthruster technology [36].
BIBLIOGRAPHY
1. O. N. Singh and A. Lakhtakia, eds., Electromagnetic Fields in
Unconventional Materials and Structures, Wiley, New York,
2000.
2. A. Lakhtakia, ed., Selected Papers on Natural Optical
Activity, SPIE Optical Engineering Press, Bellingham, WA,
1990.
3. J. Jacques, The Molecule and Its Double, McGrawHill, New
York, 1993.
4. B. Holmstedt, F. Hartmut, and B. Testa, eds., Chirality and
Biological Activity, Alan R. Liss, New York, 1990.
5. J. C. Bose, On the rotation of plane of polarisation of electric
waves by a twisted structure, Proc. Roy. Soc. Lond. 63:
146–152 (1898).
6. P. Drude, Lehrbuch der Optik, S. Hirzel, Leipzig, 1900.
7. K. F. Lindman, U
¨
ber eine durch ein isotropes system von
spiralfo¨rmigen resonatoren erzeugte rotationspolarisation
der elektromagnetischen wellen, Ann. Phys. Leipzig. 63:
621–644 (1920).
8. R. Ro, Determination of the Electromagnetic Properties of
Chiral Composites, Using Normal Incidence Measurements,
Ph.D. thesis, Pennsylvania State Univ., University Park, PA,
1991.
9. F. Gue´rin, Contribution a` L’e´tude The´orique et Expe´rimentale
des Mate´riaux Composites Chiraux et Bianisotropes dans le
Domain Microonde, Ph.D. thesis, Univ. Limoges, Limoges,
France, 1995.
10. H. C. Chen, Theory of Electromagnetic Waves, TechBooks,
Fairfax, VA, 1993.
11. A. Lakhtakia and W. S. Weiglhofer, Constraint on linear, spa
tiotemporally nonlocal, spatiotemporally nonhomogeneous
constitutive relations, Int. J. Infrared Millim. Waves 17:
1867–1878 (1996).
12. A. Lakhtakia, Beltrami Fields in Chiral Media, World Scien
tific, Singapore, 1994.
13. A. Lakhtakia, ed., Selected Papers on Linear Optical Com
posite Materials, SPIE Optical Engineering Press, Belling
ham, WA, 1996.
14. J. J. H. Wang, Generalized Moment Methods in Electromag
netics, Wiley, New York, 1991.
15. C. H. Durney and C. C. Johnson, Introduction to Modern
Electromagnetics, McGrawHill, New York, 1969.
16. A. Moscowitz, Theoretical aspects of optical activity: small
molecules, Adv. Chem. Phys. 4:67–112 (1962).
17. S. Chandrasekhar, Hydrodynamic and Hydromagnetic Sta
bility, Oxford Univ. Press, Oxford, UK, 1961.
18. P. Moon and D. E. Spencer, Field Theory Handbook, Springer,
Berlin, 1988.
19. R. F. Harrington, TimeHarmonic Electromagnetic Fields,
McGrawHill, New York, 1961, Chapter 3.
20. J. Van Bladel, Electromagnetic Fields, Hemisphere Publish
ing, New York, 1985.
21. W. C. Chew, Waves and Fields in Inhomogeneous Media,
IEEE Press, New York, 1995.
22. J. G. Fikioris, Electromagnetic ﬁeld inside a currentcarrying
region, J. Math. Phys. 6:1617–1620 (1965).
23. B. Shanker and A. Lakhtakia, Extended Maxwell Garnett
model for chiralinchiral composites, J. Phys. D: Appl. Phys.
26:1746–1758 (1993).
24. P. C. Waterman, Scattering by dielectric obstacles, Alta
Frequenza (Speciale) 38:348–352 (1969).
25. P. J. Collings, Liquid Crystals, Princeton Univ. Press, Prince
ton, NJ, 1990, Chapter 2.
26. S. Chandrasekhar, Liquid Crystals, Cambridge Univ. Press,
Cambridge, UK, 1992.
27. P. G. de Gennes and J. Prost, The Physics of Liquid Crystals,
Clarendon Press, Oxford, 1993.
28. S. D. Jacobs, ed., Selected Papers on Liquid Crystals for Op
tics, SPIE Optical Engineering Press, Bellingham, WA, 1992.
29. A. Lakhtakia, G. Ya. Slepyan, and S. A. Maksimenko,
Towards cholesteric absorbers for microwave frequencies,
Int. J. Infrared Millim. Waves 22:999–1007 (2001).
30. A. Lakhtakia, R. Messier, M. J. Brett, and K. Robbie, Sculp
tured thin ﬁlms (STFs) for optical, chemical and biological
applications, Innov. Mater. Res. 1:165–176 (1996).
31. A. Lakhtakia, Sculptured thin ﬁlms: accomplishments and
emerging uses, Mater. Sci. Eng. C 19:427–434 (2002).
32. J. F. Nye, Physical Properties of Crystals, Clarendon Press,
Oxford, 1985.
33. A. Lakhtakia and W. S. Weiglhofer, Axial propagation in gen
eral helicoidal bianisotropic media, Microwave Opt. Technol.
Lett. 6:804–806 (1993).
34. A. Lakhtakia, Directorbased theory for the optics of sculp
tured thin ﬁlms, Optik 107:57–61 (1997).
35. N. O. Young and J. Kowal, Optically active ﬂuorite ﬁlms,
Nature 183:104–105 (1959).
36. S. G. Bile´n, M. T. Domonkos, and A. D. Gallimore, Simulating
ionospheric plasma with a hollow cathode in a large vacuum
chamber, J. Spacecraft Rockets 38:617–621 (2001).
CIRCUIT STABILITY
JAMES A. SVOBODA
Clarkson University
Potsdam, New York
Stability is a property of wellbehaved circuits and sys
tems. Typically, stability is discussed in terms of feedback
systems. Wellestablished techniques, such as Nyquist
plots, Bode diagrams, and root locus plots, are available
for studying the stability of feedback systems. Electric cir
cuits can be represented as feedback systems. Nyquist
plots, Bode diagrams, and root locus plots can then be used
to study the stability of electric circuits.
1. FEEDBACK SYSTEMS AND STABILITY
Consider a feedback system such as the one shown in
Fig. 1. This feedback system consists of three parts: a for
ward block, sometimes called the ‘‘plant’’; a feedback
block, sometimes called the ‘‘controller’’; and a summer.
The signals v
i
(t) and v
o
(t) are the input and output of the
622 CIRCUIT STABILITY
feedback system. A(s) is the transfer function of the for
ward block and B(s) is the transfer function of the feed
back block. The summer subtracts the output of the
feedback block from v
i
(t). The transfer function of the
feedback system can be expressed in terms of A(s) and B(s)
as
T(s) =
V
o
(s)
V
i
(s)
=
A(s)
1÷A(s)B(s)
(1)
Suppose that the transfer functions A(s) and B(s) can each
be expressed as ratios of polynomials in s. Then
A(s) =
N
A
(s)
D
A
(s)
and B(s) =
N
B
(s)
D
B
(s)
(2)
where N
A
(s), D
A
(s), N
B
(s), and D
B
(s) are polynomials in s.
Substituting these expressions into Eq. (1) gives
T(s) =
N
A
(s)
D
A
(s)
1 ÷
N
A
(s)
D
A
(s)
N
B
(s)
D
B
(s)
=
N
A
(s)D
B
(s)
D
A
(s)D
B
(s) ÷N
A
(s)N
B
(s)
=
N(s)
D(s)
(3)
where the numerator and denominator of T(s), N(s), and
D(s) are both polynomials in s. The values of s for which
N(s) =0 are called the zeros of T(s), and the values of s that
satisfy D(s) =0 are called the poles of T(s).
Stability is a property of wellbehaved systems. For ex
ample, a stable system will produce bounded outputs
whenever its input is bounded. Stability can be deter
mined from the poles of a system. The values of the poles
of a feedback system will, in general, be complex numbers.
A feedback system is stable when all of its poles have neg
ative real parts.
The equation
1 ÷A(s)B(s) =0 (4)
is called the characteristic equation of the feedback sys
tem. The values of s that satisfy the characteristic equa
tion are poles of the feedback system. The lefthand side of
the characteristic equation, 1÷A(s)B(s), is called the
return difference of the feedback system. Figure 2 shows
how the return difference can be measured. First, the in
put, v
i
(t), is set to zero. Next, the forward path of the feed
back system is broken. Figure 2 shows how a test signal,
V
T
(s) =1, is applied and the response, V
R
(s) =–A(s)B(s), is
measured. The difference between the test signal and its
response is the return difference.
The calculation
Return difference =1÷A(s)B(s) =1÷
N
A
(s)
D
A
(s)
N
B
(s)
D
B
(s)
=
D
A
(s)D
B
(s) ÷N
A
(s)N
B
(s)
D
A
(s)D
B
(s)
shows that
1. The zeros of 1 ÷A(s)B(s) are equal to the poles of
T(s).
2. The poles of 1÷A(s)B(s) are equal to the poles of
A(s)B(s).
Consider a feedback system of the form shown in Fig. 1
with
A(s) =
s ÷5
s
2
÷4s ÷1
and B(s) =
3s
s ÷3
(5)
The poles of the forward block are the values of s that
satisfy s
2
÷4s ÷1 =0 (i.e., s
1
=3.73 and s
2
=0.26). In this
case, both poles have real, rather than complex, values.
The forward block would be stable if both poles were neg
ative. They are not, so the forward block is itself an un
stable system. To see that this unstable system is not well
behaved, consider its step response [1,2]. The step re
sponse of a system is its zero state response to a step in
put. In other words, suppose that the input to the forward
block was zero for a very long time. At some particular
time, the value of input suddenly becomes equal to 1 and
remains equal to 1. The response of the system is called
the step response. The step response can be calculated by
taking the inverse Laplace transform of A(s)/s. In this ex
ample, the step response of the forward block is
Stepresponse =5 ÷0:675e
3:73t
÷5:675e
0:27t
As time increases, the exponential terms of the step re
sponse get very, very large. Theoretically, they increase
without bound. In practice, they increase until the system
v
i
(t) v
o
(t) +
+
–
A(s)
Summer Forward block
Feedback block
Input
signal
Output
signal
B(s)
Figure 1. A feedback system.
+ v
i
(s) = 0
v
R
(s) = –A(s)B(s) v
T
(s) = 1
+
A(s)
Summer
Forward block
Feedback block
B(s)
–
Figure 2. Measuring the return difference. The difference be
tween the test input signal, V
T
(s), and the test output signal,
V
R
(s), is the return difference.
CIRCUIT STABILITY 623
saturates or breaks. This is typical of the undesirable be
havior of an unstable system.
According to Eq. (3), the transfer function of the whole
feedback system is
T(s) =
s ÷5
s
2
÷4s ÷1
1 ÷
s ÷5
s
2
÷4s ÷1
3s
s ÷3
=
(s ÷5)(s ÷3)
(s
2
÷4s ÷1)(s ÷3) ÷(s ÷5)(3s)
=
s
2
÷8s ÷15
s
3
÷2s
2
÷4s ÷3
The poles of the feedback system are the values of s that
satisfy s
3
÷2s
2
÷4s ÷3=0—that is, s
1
= ÷1, s
2
= ÷0.5 ÷
j1.66 and s
3
= ÷0.5 ÷j1.66. The real part of each of these
three poles is negative. Since all of the poles of the feed
back system have negative real parts, the feedback system
is stable. To see that this stable system is well behaved,
consider its step response. This step response can be
calculated by taking the inverse Laplace transform of
T(s)/s. In this example, the step response of the feedback
system is
Stepresponse =5 ÷11:09e
÷t
cos(
ﬃﬃﬃﬃﬃ
2t
_
÷63
)
In contrast to the previous case, as time increases, e
÷t
becomes zero so the second term of the step response dies
out. This stable system does not exhibit the undesirable
behavior typical of unstable systems.
2. STABILITY CRITERIA
Frequently, the information about a feedback system that
is most readily available is the transfer functions of the
forward and feedback blocks, A(s) and B(s). Stability cri
teria are tools for determining whether a feedback system
is stable by examining A(s) and B(s) directly, without ﬁrst
calculating T(s) and then calculating its poles—that is, the
roots of the denominator of T(s). Two stability criteria will
be discussed here: the Nyquist stability criteria and the
use of Bode diagrams to determine the gain and phase
margin.
The Nyquist stability criterion is based on a theorem
in the theory of functions of a complex variable [1,3,4].
This stability criterion requires a contour mapping
of a closed curve in the s plane using the function
A(s)B(s). The closed contour in the s plane must enclose
the right half of the s plane and must not pass through
any poles or zeros of A(s)B(s). The result of this mapping
is a closed contour in the A(s)B(s) plane. Fortunately,
the computer program MATLAB [5,6] can be used to
generate an appropriate curve in the s plane and do this
mapping.
Rewriting the characteristic equation, Eq. (4), as
A(s)B(s) = ÷1 (6)
suggests that the relationship of the closed contour in the
A(s)B(s) plane to the point ÷1 ÷j0 is important. Indeed,
this is the case. The Nyquist stability criterion involves
the number of encirclements of the point ÷1 ÷j0 by the
curve in the A(s)B(s) plane. Let
N=the number of encirclements, in the clockwise direc
tion, of ÷1 ÷j0 by the closed curve in the A(s)B(s)
plane
Z =The number of poles of T(s) in the right half of the s
plane
P =The number of poles of A(s)B(s) in the right half of the
s plane
The Nyquist stability criterion states that N, Z, and P are
related by
Z=P÷N
A stable feedback system will not have any poles in the
right half of the s plane, so Z=0 indicates a stable system.
For example, suppose that the forward and feedback
blocks of the feedback system shown in Fig. 1 have the
transfer functions described by Eq. (5). Then
A(s)B(s) =
3s
2
÷15s
s
3
÷s
2
÷11s ÷3
=
3s
2
÷15s
(s ÷3:73)(s ÷0:26)(s ÷3)
(7)
Figure 3 shows the Nyquist plot for this feedback system.
This plot was obtained using the MATLAB commands
num=[0 3 15 0]; %Coefficients of the
numerator of A(s) B(s)
den=[1÷1÷11 3]; %Coefficients of the
denominator of A(s) B(s)
nyquist (num, den)
Since A(s)B(s) has two poles in the right half of the s plane,
P=2. The Nyquist plot shows two counterclockwise en
circlements of ÷1÷j0 so N= ÷2. Then Z=P÷N=0, in
dicating that the feedback system is stable.
0.8
0.6
0.4
0.2
0
–0.2
–0.4
–0.6
–0.8
–1.4 –1.2 –1 –0.8 –0.6 –0.4 –0.2 0
Real axis
I
m
a
g
i
n
a
r
y
a
x
i
s
Figure 3. A Nyquist plot produced using MATLAB.
624 CIRCUIT STABILITY
Feedback systems need to be stable in spite of varia
tions in the transfer functions of the forward and feedback
blocks. The gain and phase margins of a feedback system
give an indication of how much A(s) and B(s) can change
without causing the system to become unstable. The gain
and phase margins can be determined using Bode dia
grams. To obtain the Bode diagrams, ﬁrst let s =jo so that
Eq. (6) becomes
A( jo)B( jo) = ÷1
The value of A(jo)B(jo) will, in general, be complex.
Two Bode diagrams are used to determine the gain
and phase margins. The magnitude Bode diagram is a
plot of 20 log[A(jo)B(jo)] versus o. The units of 20
log[A(jo)B(jo)] are decibels. The abbreviation for deci
bel is dB. The magnitude Bode diagram is sometimes
referred to as a plot of the magnitude of A(jo)B(jo), in
dB, versus o. The phase Bode diagram is a plot of the
angle of A(jo)B(jo) versus o.
It is necessary to identify two frequencies: o
g
, the gain
crossover frequency; and o
p
, the phase crossover frequen
cy. To do so, ﬁrst take the magnitude of both sides of Eq. (7)
to obtain
[A( jo)B( jo)[ =1 (8)
Converting to decibels gives
20 log[[A( jo)B( jo)[] =0 (9)
Equation (8) or (9) is used to identify a frequency, o
g
, the
gain crossover frequency. That is, o
g
is the frequency at
which
[A( jo
g
)[[B( jo
g
)[ =1
Next, take the angle of both sides of Eq. (4) to
ﬀ(A( jo)B( jo)) =180
(10)
Equation (10) is used to identify a frequency, o
p
, the gain
crossover frequency. That is, o
p
is the frequency at which
ﬀA( jo
p
) ÷ﬀB( jo
p
) =180
(11)
The gain margin of the feedback system is
Gainmargin=
1
[A( jo
p
)[ [B( jo
p
)[
(12)
The phase margin is
Phase margin=180
÷(ﬀA( jo
g
) ÷ﬀB( jo
g
)) (13)
The gain and phase margins can be easily calculated using
MATLAB. For example, suppose the forward and feedback
blocks of the feedback system shown in Fig. 1 have the
transfer functions described by Eq. (3). Figure 4 shows the
Bode diagrams for this feedback system. These plots were
obtained using the MATLAB commands
num=[0 3 15 0]; %Coefficients of the
numerator of A(s)B(s)
den=[1÷1÷11 3]; %Coefficients of the
denominator of A(s)B(s)
margin (num,den)
MATLAB has labeled the Bode diagrams in Fig. 4 to show
the gain and phase margins. The gain margin of
÷1.331 dB indicates that a decrease in A(s)B(s) of
1.331 dB or, equivalently, a decrease in gain by a factor
of 0.858, at the frequency o
p
=1.378 rad/s, would bring the
system the boundary of instability. Similarly, the phase
margin of 11.61 indicates that an increase in the angle of
A(s)B(s) of 11.61, at the frequency o
g
=2.247 rad/s, would
bring the system the boundary of instability.
When the transfer functions A(s) and B(s) have no poles
or zeros in the right half of the s plane, then the gain and
phase margins must both be positive in order for the sys
tem to be stable. As a rule of thumb [7], the gain margin
should be greater than 6dB and the phase margin should
be between 30 and 601. These gain and phase margins
provide some protection against changes in A(s) or B(s).
3. STABILITY OF LINEAR CIRCUITS
The Nyquist criterion and the gain and phase margin can
be used to investigate the stability of linear circuits. To do
so requires that the parts of the circuit corresponding to
the forward block and to the feedback block be identiﬁed.
After this identiﬁcation is made, the transfer functions
A(s) and B(s) can be calculated.
Figures 5–8 illustrate a procedure for ﬁnding A(s) and
B(s) [8]. For concreteness, consider a circuit consisting of
resistors, capacitors, and op amps. Suppose further that
the input and outputs of this circuit are voltages. Such a
circuit is shown in Fig. 5. In Fig. 6 one of the op amps has
–40
–20
0
20
10
–2
10
–1
Frequency (rad/s)
G
m
= –1.311 dB, (ω = 1.378) P
m
= 11.62° (ω = 2.247)
10
0
10
1
10
2
G
a
i
n
(
d
B
)
–360
–270
–180
–90
0
10
–2
10
–1
Frequency (rad/s)
10
0
10
1
10
2
P
h
a
s
e
(
d
e
g
)
Figure 4. Bode plot used to determine the phase and gain mar
gins. The plots were produced using MATLAB.
CIRCUIT STABILITY 625
been separated from the rest of the circuit. This is done to
identify the subcircuit N
B
. The op amp will correspond to
the forward block of the feedback system while N
B
will
contain the feedback block. N
B
will be used to calculate
B(s). In Fig. 7, the op amp has been replaced by a model of
the op amp (2). This model of the op amp indicates that the
op amp input and output voltages are related by
V
B
(s) =K(s)V
A
(s) (14)
The network N
B
can be represented by the equation
V
o
(s)
V
A
(s)
_ _
=
T
11
(s) T
12
(s)
T
21
(s) T
22
(s)
_ _
V
i
(s)
V
B
(s)
_ _
(15)
Combining Eqs. (14) and (15) yields the transfer function
of the circuit
T(s) =
V
o
(s)
V
i
(s)
=T
11
(s) ÷
T
12
(s)K(s)T
21
(s)
1 ÷K(s)T
22
(s)
(16)
or
T(s) =
V
o
(s)
V
i
(s)
=
T
11
(s)(1 ÷K(s)T
22
(s)) ÷T
12
(s)K(s)T
21
(s)
1÷K(s)T
22
(s)
Equation (15) suggests a procedure that can be used to
measure or calculate the transfer functions T
11
(s), T
12
(s),
T
21
(s), and T
22
(s). For example, Eq. (15) says that when
V
i
(s) =1 and V
B
(s) =0, then V
o
(s) =T
11
(s) and V
A
(s) =
T
21
(s). Figure 8 illustrates this procedure for determining
T
11
(s) and T
21
(s). A short circuit is used to make V
B
(s) =0
and the voltage source voltage is set to 1 so that V
i
(s) =1.
Under these conditions the voltages V
o
(s) and V
A
(s) will be
equal to the transfer functions T
11
(s) and T
21
(s). Similarly,
when V
i
(s) =0 and V
B
(s) =1, then V
o
(s) =T
12
(s) and V
A
(s) =
T
22
(s). Figure 9 illustrates the procedure for determining
T
12
(s) and T
22
(s). A short circuit is used to make V
i
(s) =0,
and the voltage source voltage is set to 1 so that V
B1
(s) =1.
Under these conditions the voltages V
o
(s) and V
A
(s) will be
equal to the transfer functions T
11
(s) and T
21
(s).
Next, consider the feedback system shown in Fig. 10.
[The feedback system shown in Fig. 1 is part, but not all,
of the feedback system shown in Fig. 10. When D(s) =0,
C
1
(s) =1 and C
2
(s) =1; then Fig. 10 reduces to Fig. 1. Con
sidering the system shown in Fig. 10, rather than the sys
tem shown in Fig. 1, avoids excluding circuits for which
D(s)a0, C
1
(s)a1, or C
2
(s)a1.] The transfer function of
this feedback system is
T(s) =
V
o
(s)
V
i
(s)
=D(s) ÷
C
1
(s)A(s)C
2
(s)
1 ÷A(s)B(s)
(17)
v
i
(t ) v
o
(t ) R
L
+
–
+
–
A circuit consisting of
resistors, capacitors, and
op amps
Figure 5. A circuit that is to be represented as a feedback sys
tem.
v
i
(t ) v
o
(t ) R
L
+
–
+
–
an op amp
N
B
+
– The rest of
the circuit
Figure 6. Identifying the subcircuit N
B
by separating an op amp
from the rest of the circuit.
V
i
(s) = 1
V
A
(s) = T
21
(s) V
B
(s) = 0
V
o
(s) =
T
11
(s)
R
L
+
–
+
–
+
–
N
B
Figure 8. The subcircuit N
B
is used to calculate T
12
(s) and T
22
(s).
v
i
(s)
v
A
(s)
v
B
(s) = K(s) v
A
(s)
v
o
(s) R
L
+
–
+
–
+
–
N
B
+
–
Figure 7. Replacing the op amp with a model of the op amp.
626 CIRCUIT STABILITY
or
T(s) =
V
o
(s)
V
i
(s)
=
D(s)(1÷A(s)B(s)) ÷C
1
(s)A(s)C
2
(s)
1÷A(s)B(s)
Comparing Eqs. (16) and (17) shows that
A(s) = ÷K(s) (18a)
B(s) =T
22
(s)
C
1
(s) =T
12
(s)
C
2
(s) =T
21
(s)
D(s) =T
11
(s)
(18b)
Finally, with Eqs. (18a) and (18b), the identiﬁcation of A(s)
and B(s) is complete. In summary
1. The circuit is separated into two parts: an op amp
and N
B
, the rest of the circuit.
2. A(s) is openloop gain of the op amp, as shown in
Fig. 7.
3. B(s) is determined from the subcircuit N
B
, as shown
in Fig. 9.
As an example, consider the Sallen–Key bandpass ﬁlter [9]
shown in Fig. 11. The transfer function of this ﬁlter is
T(s) =
V
o
(s)
V
i
(s)
=
5460s
s
2
÷199s ÷4 10
6
(19)
The ﬁrst step toward identifying A(s) and B(s) is to sepa
rate the op amp from the rest of the circuit, as shown in
Fig. 12. Separating the op amp from the rest of the circuit
identiﬁes the subcircuit N
B
. Next, N
B
is used to calculate
the transfer functions T
11
(s), T
12
(s), T
21
(s), and T
22
(s).
Figure 13 corresponds to Fig. 8 and shows how T
12
(s)
and T
22
(s) are calculated. Analysis of the circuit shown in
Fig. 13 gives
T
12
(s) =1 and T
22
(s) =
0:259s
2
÷51:6s ÷1:04 10
6
s
2
÷5660s ÷4 10
6
(20)
[The computer program ELab [10] provides an alternative
to doing this analysis by hand. ELab will calculate the
transfer function of a network in the form shown in
Eq. (16)—that is, as a symbolic function of s. ELab is
free and can be downloaded from http://sunspot.ece.
clarkson.edu:1050/Bsvoboda/software.html on the World
Wide Web.]
Figure 14 corresponds to Fig. 9 and shows how T
11
(s)
and T
21
(s) are calculated. Analysis of the circuit shown in
Fig. 14 gives
T
11
(s) =0 and T
21
(s) =
÷1410s
s
2
÷5660s ÷4 10
6
(21)
Substituting Eqs. (20) and (21) into Eq. (16) gives
T(s) =
K(s)
÷1410s
s
2
÷5660s ÷410
6
_ _
1 ÷K(s)
0:259s
2
÷51:6s ÷1:04 10
6
s
2
÷5660s ÷4 10
6
_ _ (22)
V
i
(s) = 0
V
A
(s) = T
22
(s)
V
B
(s) = 1
V
o
(s) =
T
12
(s)
R
L
+
–
+
–
+
–
N
B
Figure 9. The subcircuit N
B
is used to calculate T
11
(s) and T
21
(s).
+
+
–
+
+
v
o
(t ) v
i
(t )
C
1
(s) A(s)
D(s)
B(s)
C
2
(s)
+
Figure 10. A feedback system that corre
sponds to a linear system.
+
+
– –
v
o
(t )
v
i
(t )
R
1
C
1
C
2
R
4
R
3
R
5
R
2
–
+
–
Figure 11. A Sallen–Key bandpass ﬁlter: R
1
=R
2
=R
3
=R
5
=
7.07 kO, R
4
=20.22 kO, and C
1
=C
2
=0.1 mF.
CIRCUIT STABILITY 627
When the op amp is modeled as an ideal op amp, K(s)N
and Eq. (22) reduces to Eq. (19). This is reassuring but
only conﬁrms what was already known. Suppose that
a more accurate model of the op amp is used. A freque
ntly used op amp model [2] represents the gain of the
op amp as
K(s) = ÷
A
o
s ÷
B
A
o
(23)
where A
o
is the DC gain of the op amp and B is the gain–
bandwidth product of the op amp (2). Both A
o
and B are
readily available from manufacturers speciﬁcations of op
amps. For example, when the op amp is a mA741 op amp,
then A
o
=200,000 and B=2p*10
6
rad/s, so
K(s) = ÷
200; 000
s ÷31:4
Equation (18) indicates that A(s) = ÷K(s) and B(s) =
T
22
(s), so in this example
A(s) =
200; 000
s ÷31:4
and B(s) =0:259
s
2
÷51:6s ÷1:04 10
6
s
2
÷5600s ÷410
6
_ _
To calculate the phase and gain margins of this ﬁlter, ﬁrst
calculate
A(s)B(s) =
51; 800(s
2
÷51:6s ÷1:04 10
6
)
s
3
÷5974s
2
÷5777240s ÷1246 10
6
Next, the MATLAB commands
num=20000
+
[0 0.259 51.6 1040000];
%Numerator Coefficients
den=[1 5974 5777240 1256
+
10
4
6];
%Denominator Coefficients
margin (num, den)
are used to produce the Bode diagram shown in Fig. 15.
Figure 15 shows that the Sallen–Key ﬁlter will have an
inﬁnitegain margin and a phase margin of 76.51 when a
mA741 op amp is used.
4. OSCILLATORS
Oscillators are circuits that are used to generate a sinu
soidal output voltage or current. Typically, oscillators have
no input. The sinusoidal output is generated by the circuit
itself. This section presents the requirements that a cir
cuit must satisfy if it is to function as an oscillator and
shows how these requirements can be used to design the
oscillator.
To begin, recall that the characteristic equation of a
circuit is
1 ÷A(s)B(s) =0
Suppose that this equation is satisﬁed by a value of s of the
form s =0÷jo
o
. Then
A( jo
o
)B( jo
o
) = ÷1 =1e
j180
(24)
In this case, the steadystate response of the circuit will
contain a sustained sinusoid at the frequency o
o
(11). In
other words, Eq. (24) indicates that the circuit will func
tion as an oscillator with frequency o
o
when A( jo
o
)B( jo
o
)
has a magnitude equal to 1 and a phase angle of 1801.
+
–
i
(t )
R
b
C
1
C
2
R
4
R
3
R
5
R
1
+ +
–
o
(t )
An op amp
R
2
+
–
v
v
Figure 12. Identifying the subcircuit N
B
by separating an op
amp from the rest of the circuit.
+
_
V
i
(s) = 1
V
A
(s) = T
21
(s)
V
o
(s) = T
11
(s)
V
B
(s) = 0
N
B
C
1
C
2
R
4
R
3
R
5
R
1
R
2
+
–
+
–
Figure 13. The subcircuit N
B1
is used to calculate T
11
(s) and
T
21
(s).
R
1
R
3
R
2
R
4
R
5 C
1
C
2
V
i
(s) = 0
N
B
V
A
(s) = T
22
(s)
+
–
+
–
V
B
(s) = 1
+
V
o
(s) = T
12
(s)
–
Figure 14. The subcircuit N
B
is used to calculate T
12
(s) and
T
22
(s).
628 CIRCUIT STABILITY
As an example, consider using Eq. (24) to design the
Wienbridge oscillator, shown in Fig. 16, to oscillate at o
o
=
1000 rad/s. The ﬁrst step is to identify A(s) and B(s) using
the procedure described in the previous section. In Fig. 17
the ampliﬁer is separated from the rest of the network to
identify the subcircuit N
B
. Also, from Eqs. (14) and (18),
we have
A(s) = ÷K
Next, the subcircuit N
B
is used to determine B(s) =T
22
(s),
as shown in Fig. 18. From Fig. 18 it is seen that
T
22
(s) =
1
Cs
+ R
1
Cs
÷R
1
Cs
+ R
1
Cs
÷R
÷ R÷
1
Cs
_ _
=
1
1 ÷ R÷
1
Cs
_ _ R÷
1
Cs
_ _
R +
1
Cs
_ _
=
1
1÷ R÷
1
Cs
_ _
Cs ÷
1
R
_ _ =
1
3÷RCs ÷
1
RCs
So
A(s)B(s) =
÷K
3 ÷RCs ÷
1
RCs
Now let s =0 ÷jo
o
to get
A(jo
o
)B(jo
o
) =
÷K
3 ÷jo
o
RC ÷j
1
o
o
RC
(25)
The phase angle of A(jo
o
) B(jo
o
) must be 1801 if the circuit
is to function as an oscillator. That requires
jo
o
RC ÷j
1
o
o
RC
=0 =o
o
=
1
RC
(26)
50
0
–50
10
1
10
2
10
3
10
4
10
5
G
m
= Inf dB, (o) = (NaN) P
m
= 76.51
°
(o) = (1961)
G
a
i
n
(
d
B
)
0
–90
–180
–270
–360
10
1
10
2
10
3
10
4
10
5
Frequency (rad/s)
Frequency (rad/s)
P
h
a
s
e
(
d
e
g
)
Figure 15. The Bode diagrams used to determine
the phase and gain margins of the Sallen–Key
bandpass ﬁlter.
K
R C
R
L
R C
v
o
(t)
+
–
Figure 16. A Wien bridge oscillator.
K
C
C
R
R
V
o
(s)
+
–
V
A
(s) V
B
(s)
+
–
+
–
N
B
Figure 17. The ampliﬁer is separated from the rest of the Wien
bridge oscillator to identify the subcircuit N
B
.
CIRCUIT STABILITY 629
Oscillation also requires that the magnitude of A(jo
o
)
B( jo
o
) be equal to 1. After substituting Eq. (26) into
Eq. (25), this requirement reduces to
K =3
That is, the ampliﬁer gain must be set to 3. Design of the
oscillator is completed by picking values of R and C to
make o
o
=1000 rad/s (e.g., R=10 kO and C=0.1 mF).
5. THE ROOT LOCUS
Frequently the performance of a feedback system is ad
justed by changing the value of a gain. For example, con
sider the feedback system shown in Fig. 1 when
A(s) =
N
A
(s)
D
A
(s)
and B(s) =K (27)
In this case, A(s) is the ratio of two polynomials in s and
B(s) is the gain that is used to adjust the system. The
transfer function of the feedback system is
T(s) =
N
A
(s)
D
A
(s) ÷KN
A
(s)
=
N(s)
D(s)
(28)
The poles of feedback system are the roots of the polyno
mial
D(s) =D
A
(s) ÷KN
A
(s) (29)
Suppose that the gain K can be adjusted to any value
between 0 and N. Consider the extreme values of K. When
K=0, D(s) =D
A
(s) so the roots of D(s) are the same as the
roots of D
A
(s). When K=N, D
A
(s) is negligible compared
to KN
A
(s). Therefore D(s) =KN
A
(s) and the roots of D(s)
are the same as the roots of N
A
(s). Notice that the roots of
D
A
(s) are the poles of A(s) and the roots of N
A
(s) are the
zeros of A(s). As K varies from 0 and N, the poles of T(s)
start at the poles of A(s) and migrate to the zeros of A(s).
The root locus is a plot of the paths that the poles of
T(s) take as they move across the s plane from the poles
of A(s) to the zeros of A(s).
A set of rules for constructing root locus plots by hand
are available [1,4,7,13]. Fortunately, computer software
for constructing root locus plots is also available. For ex
ample, suppose that the forward and feedback blocks in
Fig. 1 are described by
A(s) =
s(s ÷2)
(s ÷1)(s ÷2)(s ÷3)
=
s
2
÷2s
s
3
÷6s
2
÷11s ÷6
and B(s) =K
The root locus plot for this system is obtained using the
MATLAB [5,6] commands
num=([0 1÷2 0]);
den=([1 6 11 6]);
rlocus (num, den)
This root locus plot is shown in Fig. 19. After the root locus
has been plotted, the MATLAB command
rlocfind (num, den)
can be used to ﬁnd the value of the gain Kcorresponding to
any point on the root locus. For example, when this com
mand is given and the cursor is placed on the point where
the locus crosses the positive imaginary axis, MATLAB
indicates that gain corresponding to the point 0.0046÷
j0.7214 is K=5.2678. For gains larger than 5.2678, two
poles of T(s) are in the right half of the s plane so the
feedback system is unstable.
The bilinear theorem [12] can be used to make a con
nection between electric circuits and root locus plots. Con
sider Fig. 20, where one device has been separated from
the rest of a linear circuit. The separated device could be a
C
C
R
R
V
o
(s) = T
12
(s)
+
–
V
A
(s) = T
22
(s)
V
B
(s)
–
+
N
B
–
+
= 1
Figure 18. The subcircuit N
B
is used to calculate B(s) =T
22
(s) for
the Wien bridge oscillator.
6
4
2
0
–2
–4
–6
–6 –4 –2 0 2 4 6
I
m
a
g
i
n
a
r
y
a
x
i
s
Real axis
Figure 19. A root locus plot produced using MATLAB. The poles
of A(s) are marked by x’s and the zeros of A(s) are marked by o’s.
As K increases from zero to inﬁnity, the poles of T(s) migrate from
the poles of A(s) to the zeros of A(s) along the paths indicated by
solid lines.
630 CIRCUIT STABILITY
resistor, a capacitor, an ampliﬁer, or any twoterminal de
vice [12]. The separated device has been labeled as x. For
example, x could be the resistance of a resistor, the capac
itance of a capacitor, or the gain of an ampliﬁer. The bi
linear theorem states that the transfer function of the
circuit will be of the form
T(s) =
V
o
(s)
V
i
(s)
=
E(s) ÷xF(s)
G(s) ÷xH(s)
=
N(s)
D(s)
(30)
where E(s), F(s), G(s), and H(s) are all polynomials in s. A
transfer function of this form is said to be a bilinear func
tion of the parameter x since both the numerator and de
nominator polynomials are linear functions of the
parameter x. The poles of T(s) are the roots of the denom
inator polynomial
D(s) =G(s) ÷xH(s) (31)
As x varies from 0 to N, the poles of T(s) begin at the roots
of G(s) and migrate to the roots of H(s). The root locus can
be used to display the paths that the poles take as they
move from the roots of G(s) to the roots of H(s). Similarly,
the root locus can be used to display the paths that the
zeros of T(s) take as they migrate from the roots of E(s) to
the roots of F(s).
For example, consider the Sallen–Key bandpass ﬁlter
shown in Fig. 11. When
R
1
=R
2
=R
3
=7:07 kO; C
1
=C
2
=0:1mF; and
K =1 ÷
R
4
R
5
then the transfer function of this Sallen–Key ﬁlter is
T(s) =
K(1414s)
s
2
÷(4 ÷K)(1414s) ÷4 10
6
=
K(1414s)
(s
2
÷5656s ÷4 10
6
) ÷K(÷1414s)
(32)
As expected, this transfer function is a bilinear function
the gain K. Comparing Eqs. (30) and (32) shows that
E(s) =0, F(s) =1414s, G(s) =s
2
÷5656s ÷410
5
, and
H(s) = ÷1414s. The root locus describing the poles of the
ﬁlter is obtained using the MATLAB commands
G=([1 5656 4
+
10
4
6]);
H=([0 ÷1414 0]);
rlocus (H,G)
Figure 21 shows the resulting root locus plot. The poles
move into the right half of the s plane, and the ﬁlter
becomes unstable when K44.
BIBLIOGRAPHY
1. R. C. Dorf and R. H. Bishop, Modern Control Systems, 7th ed.,
AddisonWesley, Reading, MA, 1995.
2. R. C. Dorf and J. A. Svoboda, Introduction to Electric Circuits,
Wiley, New York, 1996.
3. R. V. Churchill, J. W. Brown, and R. F. Verhey, Complex Vari
ables and Applications, McGrawHill, New York, 1974.
4. S. M. Shinners, Modern Control System Theory and Design,
Wiley, New York, 1992.
5. R. D. Strum and D. E. Kirk, Contemporary Linear Systems
Using MATLAB, PWS, Boston, 1994.
6. N. E. Leonard and W. S. Levine, Using MATLAB to Analyze
and Design Control Systems, Benjamin Cummings, Redwood
City, CA, 1995.
7. K. Ogata, Modern Control Engineering, PrenticeHall, Engle
wood Cliffs, NJ, 1970.
8. J. A. Svoboda and G. M. Wierzba, Using PSpice to determine
the relative stability of RC active ﬁlters, Int. J. Electron.
74(4):593–604 (1993).
9. F. W. Stephenson, RC Active Filter Design Handbook, Wiley,
New York, 1985.
10. J. A. Svoboda, ELab, A circuit analysis program for engineer
ing education, Comput. Appl. Eng. Educ. 5:135–149 (1997).
11. W.K. Chen, Active Network and Feedback Ampliﬁer Theory,
New York, McGrawHill, 1980.
12. K. Gehler, Theory of Tolerances, Akademiai Kiado, Budapest,
Hungary, 1971.
x
v
i
(t)
+
–
v
o
(t)
+
–
Figure 20. A single device is separated from the rest of the net
work. The parameter associated with this device is called x. The
transfer function of the network will be a bilinear function of x.
1
0.8
0.6
0.4
0.2
0
–0.2
–0.4
–0.6
–0.8
–1
–1 –0.5 0.5 1 0
10
4
×
× 10
4
Real axis
I
m
a
g
i
n
a
r
y
a
x
i
s
Figure 21. This root locus plot shows that the poles of the
Sallen–Key bandpass ﬁlter move into the right of the s plane as
the gain increases.
CIRCUIT STABILITY 631
13. A. Budak, Passive and Active Network Synthesis, Waveland
Press, Prospect Heights, IL, 1991, Chapter 6.
FURTHER READING
P. Gray and R. Meyer, Analysis and Design of Analog Integrated
Circuits, 3rd ed., Wiley, New York, 1993, Chapters 8 and 9.
A. Sedra and K. Smith, Microelectronic Circuits, 4th ed., Oxford
Univ. Press, 1998, Chapter 8.
CIRCUIT TUNING
ROLF SCHAUMANN
Portland State University
Portland, Oregon
Circuit tuning refers to the process of adjusting the values
of electronic components in a circuit to ensure that the
fabricated or manufactured circuit performs to speciﬁca
tions. In digital circuits, where signals are switched func
tions in the time domain and correct operation depends
largely on the active devices switching all the way be
tween their ON and OFF states, tuning in the sense dis
cussed in this article is rarely necessary. In analog
continuoustime circuits, however, signals are continuous
functions of time and frequency so that circuit perfor
mance depends critically on the component values. Con
sequently, in all except the most undemanding
applications with wide tolerances, correct circuit opera
tion almost always requires some form of tuning. Natu
rally, components could be manufactured with very tight
tolerances, but the resulting fabrication costs would be
come prohibitive. In practice, therefore, electronic compo
nents used in circuit design are never or only rarely
available as accurately as the nominal design requires,
so we must assume that they are affected by fabrication
and manufacturing tolerances. Furthermore, regardless of
whether a circuit is assembled in discrete form with
discrete components on a printed circuit board (as a
hybrid circuit), or in integrated form on an integrated
circuit chip, the circuit will be affected by parasitic com
ponents and changing operating conditions, all of which
contribute to inaccurate circuit performance. Consider, for
example, the requirement of implementing as a hybrid
circuit a time of 1s for a timer circuit via an RC time
constant t =RC with an accuracy of 0.1%. Assume that R
and C are selected to have the nominal values R=100kO
and C=10 mF, that inexpensive chip capacitors with
720% tolerances are used, and that the desired fabrica
tion process of thinﬁlm resistors results in components
with 710% tolerances. The fabricated time constant can
therefore be expected to lie in the range
0:68s _ t =100kO(1 ±0:1)10 mF(1 ±0:2) _ 1:32 s
In other words, the t error must be expected to be 732%,
which is far above the speciﬁed 0.1%. Tuning is clearly
necessary. Because capacitors are difﬁcult to adjust and
accurate capacitors are expensive, let us assume in this
simple case that the capacitor was measured with 0.05%
accuracy as C=11.125 mF (i.e., the measured error was
÷11.25%). We can readily compute that the resistor should
be adjusted (trimmed) to the nominal value R=t/C=
1 s/11.125 mF=89.888kO within a tolerance of 745 O to
yield the correctly implemented time constant of 1 s with
70.1%tolerances. Observe that tuning generally allows the
designer to construct a circuit with less expensive wide
tolerance parts because subsequent tuning of these or other
components permits the errors to be corrected. Thus, C was
fabricated with 20% tolerances but measured with a 0.05%
error to permit the resistor with fabrication tolerances of
10%to be trimmed to a 0.05%accuracy. Note that implied in
this process is the availability of measuring instruments
with the necessary accuracy.
Tuning has two main purposes. Its most important
function is to correct errors in circuit performance caused
by such factors as fabrication tolerances such as in the
preceding example. Second, it permits a circuit’s function
or parameters, such as the cutoff frequency of a given
lowpass ﬁlter, to be changed to different values to make
the circuit more useful or to be able to accommodate
changing operating requirements. But even the best fab
rication technology together with tuning will not normally
result in a circuit operating with zero errors; rather, the
aim of tuning is to trim the values of one or more, or in
rare cases of all, components until the circuit’s response is
guaranteed to remain within a speciﬁed tolerance range
when the circuit is put into operation. Figure 1 illustrates
the idea for a lowpass ﬁlter. Examples are a gain error
that is speciﬁed to remain within 70.05 dB, the cutoff
frequency f
c
of a ﬁlter that must not deviate from the
design value of, say, f
c
=10kHz by more than 85 Hz, or the
gain of an ampliﬁer that must settle to, say, 1% of its ﬁnal
value within less than 1 ms. As these examples indicate, in
general, a circuit’s operation can be speciﬁed in the time
domain, such as a transient response with a certain high
est permissible overshoot or a maximal settling time,
or in the frequency (s) domain through an input–output
Figure 1. The shaded area in the gain–frequency plot shows the
operating region for a lowpass ﬁlter that must be expected based
on the basis of raw (untuned) fabrication tolerances; the dotted
region is the acceptable tolerance range that must be maintained
in operation after the ﬁlter is tuned.
632 CIRCUIT TUNING
transfer function with magnitude, phase, or delay speciﬁ
cations and certain tolerances (see Fig. 1). This article
focuses on the tuning of ﬁlters, that is, of frequency
selective networks. Such circuits are continuous functions
of components, described by transfer functions in the s
domain, where tuning of design parameters (e.g., cutoff
frequency, bandwidth, quality factor, gain), is particularly
important in practice. The concepts discussed in connec
tion with ﬁlters apply equally to other analog circuits.
Obviously, in order to tune (adjust) a circuit, that circuit
must be tunable; that is, its components must be capable
of being varied in some manner (manually or electroni
cally) by an amount sufﬁcient to overcome the conse
quences of fabrication tolerances, parasitic effects, or
other such factors.
An example will help to illustrate the discussion and
terminology. Consider the simple secondorder active
bandpass circuit in Fig. 2. Its voltage transfer function,
under the assumption of ideal operational ampliﬁers, can
be derived to be
T(s) =
V
2
V
1
= ÷
b
1
s
s
2
÷a
1
s ÷a
0
= ÷
1
R
1
C
1
s
s
2
÷
1
R
2
1
C
1
÷
1
C
2
_ _
s ÷
1
R
1
R
2
C
1
C
2
(1)
We see that T(s) is a continuous function of the circuit
components, as are all its coefﬁcients that determine the
circuit’s behavior:
b
1
=
1
R
1
C
1
; a
1
=
C
1
÷C
2
R
2
C
1
C
2
; a
0
=
1
R
1
R
2
C
1
C
2
(2)
Just as in the earlier example of the RC time constant, the
coefﬁcients will not be implemented precisely if the com
ponent values have fabrication tolerances. If these compo
nent tolerances are ‘‘too large,’’ generally the coefﬁcient
errors will become ‘‘too large’’ as well, and the circuit will
not function correctly. In that case, the circuit must be
tuned. Furthermore, circuits are generally affected by
parasitic components. Parasitic components, or parasitics,
are physical effects that often can be modeled as ‘‘real
components’’ affecting the circuit’s performance but that
frequently are not speciﬁed with sufﬁcient accuracy and
are not included in the nominal design. For instance, in
the ﬁlter of Fig. 2, a parasitic capacitor can be assumed to
exist between any two nodes or between any individual
node and ground; also, real ‘‘wires’’ are not ideal short
circuit connections with zero resistance but are resistive
and, at high frequencies, even inductive. In the ﬁlter of
Fig. 2, a parasitic capacitor C
p
between nodes n
1
and n
2
would let the resistor R
2
look like the frequencydepen
dent impedance Z
2
(s) =R
2
/(1 ÷sC
p
R
2
). Similarly, real re
sistive wires would place small resistors r
w
in series with
C
1
and C
2
and would make these capacitors appear lossy.
That is, the capacitors C
i
, i =1, 2, would present admit
tances of the form Y
i
(s) =sC
i
/(1 ÷sC
i
r
w
). Substituting Z
2
(s)
and Y
i
(s) for R
2
and C
i
, respectively, into Eq. (1) shows
that, depending on the frequency range of interest and the
element values, the presence of these parasitics changes
the coefﬁcients of the transfer function, maybe even its
type, and consequently the circuit’s performance. Simi
larly, when changes occur in environmental operating
conditions, such as bias voltages or temperature, the
performance of electronic devices is altered, and as a
result the fabricated circuit may not perform as speciﬁed.
As discussed by Moschytz [1, Section 4.4, pp. 394–425],
and Bowron and Stevenson [2, Section 9.5, pp. 247–251],
the operation of tuning can be classiﬁed into functional
and deterministic tuning. In functional tuning, the de
signed circuit is assembled, and its performance is mea
sured. By analyzing the circuit, we can identify which
component affects the performance parameter to be tuned.
These predetermined components are then adjusted in
situ (i.e., with the circuit in operation), until errors in
performance parameters are reduced to acceptable toler
ances. The process is complicated by the fact that tuning is
most often interactive, meaning that adjusting a given
component will vary several circuit parameters; thus
iterative routines are normally called for. As an example,
consider again the active RC ﬁlter in Fig. 2. If its bandpass
transfer function, Eq. (1), is expressed in the measurable
terms of center frequency o
0
, the pole quality factor
Q=o
0
/Do, the parameter that determines the ﬁlter’s
bandwidth Do, and midband (at s =jo
0
) gain K as
T(s) = ÷
1
R
1
C
1
s
s
2
÷
1
R
2
1
C
1
÷
1
C
2
_ _
s ÷
1
R
1
R
2
C
1
C
2
= ÷
K
o
0
Q
s
s
2
÷s
o
0
Q
÷o
2
0
(3)
These parameters are expressed in terms of the circuit
components, and we arrive at the more meaningful and
useful design equations
o
0
=
1
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
R
1
R
2
C
1
C
2
_ ; Q=
ﬃﬃﬃﬃﬃﬃ
R
2
R
1
¸
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
C
1
C
2
_
C
1
÷C
2
; K =
R
2
R
1
C
2
C
1
÷C
2
(4)
instead of Eq. (2). It is clear that varying any of the passive
components will change all three ﬁlter parameters, so
that expensive and timeconsuming iterative tuning is Figure 2. Active RC bandpass ﬁlter.
CIRCUIT TUNING 633
required. However, functional tuning has the advantage
that the effects of all component and layout parasitics,
losses, loading, and other hardtomodel or hardtopredict
factors are accounted for because the performance of the
complete circuit is measured under actual operating con
ditions. In general, more accurate results are obtained by
basing functional tuning on measurements of phase
rather than of magnitude because phase tends to be
more sensitive to component errors.
Deterministic tuning refers to calculating the needed
value of a component from circuit equations and then
adjusting the component to that value. We determined the
resistor R=t/C=89.888 kO to set a time constant of 1 s at
the beginning of this article in this manner. Similarly,
from Eq. (4) we can derive the three equations in the four
unknowns R
1
, R
2
, C
1
, and C
2
R
2
=
Q
o
0
1
C
1
÷
1
C
2
_ _
; R
1
=
1
R
2
o
2
0
C
1
C
2
; C
1
=
1
KR
1
Q
o
0
(5)
with C
2
a free parameter. That there are more circuit
components than parameters is normal, so the additional
‘‘free’’ elements may be used at will, for example, to
achieve practical element values or elementvalue spreads
(i.e., the difference between the maximum and minimum
of a component type, such as R
max
÷R
min
). Technology or
cost considerations may place further constraints on tun
ing by removing some components from the list of tunable
ones. Thus, in hybrid circuits with thin or thickﬁlm
technology as in the preceding example, the capacitors
will likely be ﬁxed; only the two resistors will be deter
mined as in Eq. (5) from the prescribed circuit parameters
o
0
and Q and the selected and measured capacitor values.
This option leaves the midband gain ﬁxed at the value K=
Q/(o
0
C
1
R
1
). Precise deterministic tuning requires careful
measurements and accurate models and design equations
that, in contrast to the idealized expressions in Eq. (5),
describe circuit behavior along with loss, parasitic, and
environmental effects. As we saw in Eq. (5), the equations
that must be solved are highly nonlinear and tend to be
very complex, particularly if parasitic components also are
involved. Computer tools are almost always used to ﬁnd
the solution. Typically, automatic laser trimming is em
ployed to tune the resistors to the desired tolerances (e.g.,
0.1%). A second tuning iteration using functional tuning
may be required because the assembled circuit under
power may still not meet the speciﬁcations as a result of
further parasitic or loading effects that could not be
accounted for in the initial deterministic tuning step.
1. SENSITIVITY
We mentioned earlier that a ﬁlter parameter P depends on
the values k
i
of the components used to manufacture a
circuit, P=P(k
i
), and that real circuit components or parts
can be realized only to within some tolerances 7Dk
i
; that
is, the values of the parts used to assemble circuits are
k
i
7Dk
i
. Clearly, the designer needs to know how much
these tolerances will affect the circuit and whether the
resulting errors can be corrected by adjusting (tuning) the
circuit after fabrication. Obviously, the parameter to be
tuned must depend on the component to be varied. For
example, Q in Eq. (4) is a function of the components R
1
,
R
2
, C
1
, and C
2
, any one of which can be adjusted to correct
fabrication errors in Q. In general, the questions of how
large the adjustment of a component has to be, whether it
should be increased or decreased, and what the best
tuning sequence is are answered by considering the para
meter’s sensitivity to component tolerances. How sensitive
P is to the componentvalue tolerances—that is, how large
the deviation DP of the parameter in question is—is
computed for small changes via the derivative of P(k
i
)
with respect to k
i
, @P/@k
i
, at the nominal value k
i
:
DP=
@P(k
i
)
@k
i
Dk
i
(6)
Typically, designers are less interested in the absolute
tolerances than in the relative ones
DP
P
=
k
i
P
@P
@k
i
Dk
i
k
i
=S
P
k
i
Dk
i
k
i
(7)
where S
P
k
i
is the sensitivity, deﬁned as ‘‘the relative change
of the parameter divided by the relative change of the
component’’:
S
P
k
i
=
DP=P
Dk
i
=k
i
(8)
A detailed discussion of sensitivity issues can be found in
many textbooks (see Schaumann et al. [3], Chapter 3, pp.
124–196). For example, the sensitivity of o
0
in Eq. (4) to
changes in R
1
is readily computed to be
S
o
0
R
1
=
R
1
o
0
@o
0
@R
1
=
R
1
1=(R
1
R
2
C
1
C
2
)
1=2
÷
1
2
_ _
R
2
C
1
C
2
(R
1
R
2
C
1
C
2
)
3=2
= ÷
1
2
(9)
S
o
0
R
1
= ÷0:5 means that the percentage error in the para
meter o
0
is onehalf the size of the percentage error of R
1
and opposite in sign (i.e., if R
1
increases, o
0
decreases). A
large number of useful sensitivity relations that make
sensitivity calculations easy can be derived (see, e.g.,
Moschytz [1], Section 1.6, pp. 103–105, 1.5, pp. 71–102,
and 4.3, pp. 371–393, or Schaumann et al. [3], Chapter 3,
pp. 124–196). Of particular use for our discussion of tuning
are
S
P(k
n
)
k
=nS
P(k)
k
; S
P(ak)
k
=S
P(k)
k
;
S
P(1=k)
k
= ÷S
P(k)
k
; and S
P(
ﬃﬃ
k
_
)
k
=
1
2
S
P(k)
k
(10)
where a is a constant, independent of k. The last two
of these equations are special cases of the ﬁrst one for
n= ÷1 and n=
1
2
, respectively. The last equation gener
alizes the result obtained in Eq. (9). Equations (7) and
(8) indicate that, for small differential changes, the
634 CIRCUIT TUNING
parameter deviation caused by a component error and,
conversely from the point of view of tuning, the change in
component value necessary to achieve a desired change in
parameter can be computed if the sensitivity is known.
In Eqs. (6) and (7) we purposely used partial deriva
tives, @P/@k
i
, to indicate that circuit parameters normally
depend on more than one component [see Eq. (4)], all of
which affect the accuracy of the parameter. To get a more
complete picture of the combined effect of the tolerances
and to gain insight into the operation of tuning involving
several parameters, total derivatives need to be computed.
Assuming P depends on n components, we ﬁnd (see
Schaumann et al. [3], Chapter 3, pp. 124–196)
DP=
@P
@k
1
Dk
1
÷
@P
@k
2
Dk
2
÷ · · · ÷
@P
@k
n
Dk
n
that is
DP
P
=
k
1
P
@P
@k
1
Dk
1
k
1
÷ · · · ÷
k
n
P
@P
@k
n
Dk
n
k
n
=S
P
k
1
Dk
1
k
1
÷ · · · ÷S
P
kn
Dk
n
k
n
=
n
i =1
S
P
k
i
Dk
i
k
i
(11)
indicating that the sum of all relative component toler
ances, weighted by their sensitivities, contributes to the
parameter error. To illustrate the calculations, let us
apply Eq. (11) to o
0
in Eq. (4). Using Eqs. (9) and (10),
the result is
Do
0
o
0
=
R
1
o
0
@o
0
@R
1
DR
1
R
1
÷
R
2
o
0
@o
0
@R
2
DR
2
R
2
÷
C
1
o
0
@o
0
@C
1
DC
1
C
1
÷
C
2
o
0
@o
0
@C
2
DC
2
C
2
=S
o0
R
1
DR
1
R
1
÷S
o0
R
2
DR
2
R
2
÷S
o
p
C
1
DC
1
C
1
÷S
o0
C
2
DC
2
C
2
= ÷
1
2
DR
1
R
1
÷
1
2
DR
2
R
2
÷
1
2
DC
1
C
1
÷
1
2
DC
2
C
2
= ÷
1
2
DR
1
R
1
÷
DR
2
R
2
÷
DC
1
C
1
÷
DC
2
C
2
_ _
(12)
The last expression gives insight into whether and how o
0
can be tuned. Because the effects of the errors are additive,
tuning just one component, say, R
1
, will sufﬁce for given
tolerances of R
2
, C
1
, and C
2
if DR
1
can be large enough. If
we have measured the R
2
errors at ÷12%, and those
of C
1
and C
2
at ÷15% and ÷10%, respectively, Eq. (12)
results in
Do
0
o
0
= ÷
1
2
DR
1
R
1
÷0:12 ÷0:15÷0:10
_ _
= ÷0:5
DR
1
R
1
÷0:13
_ _
(13)
indicating that R
1
must be decreased by 13% to yield,
within the linearized approximations made, Do
0
E0.
Inserting components with these tolerances into Eq. (4)
for o
0
conﬁrms the result obtained.
To expand these results and gain further insight into
the effects of tolerances, as well as beneﬁcial tuning
strategies and their constraints, we remember that a
transfer function generally depends on more than
one parameter. Returning to the example of Fig. 2
described by the function T(s) in Eq. (3) with the three
parameters o
0
, Q, and K given in Eq. (4) and applying
Eq. (11) leads to
Do
0
o
0
=S
o
0
R
1
DR
1
R
1
÷S
o
0
R
2
DR
2
R
2
÷S
o
0
C
1
DC
1
C
1
÷S
o
0
C
2
DC
2
C
2
(14a)
DQ
Q
=S
Q
R
1
DR
1
R
1
÷S
Q
R
2
DR
2
R
2
÷S
Q
C
1
DC
1
C
1
÷S
Q
C
2
DC
2
C
2
(14b)
DK
K
=S
K
R
1
DR
1
R
1
÷S
K
R
2
DR
2
R
2
÷S
K
C
1
DC
1
C
1
÷S
K
C
2
DC
2
C
2
(14c)
These equations can be expressed in matrix form as
follows:
Do
0
o
0
DQ
Q
DK
K
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
=
S
o
0
R
1
S
o
0
R
2
S
o
0
C
1
S
o
0
C
2
S
Q
R
1
S
Q
R
2
S
Q
C
1
S
Q
C
2
S
K
R
1
S
K
R
2
S
K
C
1
S
K
C
2
_
_
_
_
_
_
_
_
_
_
DR
1
R
1
DR
2
R
2
DC
1
C
1
DC
2
C
2
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
(15)
The sensitivity matrix in Eq. (15) (see Moschytz [1],
Section 4.3, pp. 376–393, or Schaumann et al. [3], Section
3.3, pp. 161–188), a 3 4 matrix in this case, shows how
the tolerances of all the ﬁlter parameters depend on the
component tolerances. We see that adjusting any one of
the circuit components will vary all ﬁlter parameters as
long as all the sensitivities are nonzero, which is indeed
the case for the circuit in Fig. 2. Thus, noninteractive
tuning is not possible. To illustrate the form of the
sensitivity matrix, we calculate for the circuit in Fig. 2
Do
0
o
0
DQ
Q
DK
K
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
=
÷0:5 ÷0:5 ÷0:5 ÷0:5
÷0:5 0:5 ÷
1
2
C
1
÷C
2
C
1
÷C
2
1
2
C
1
÷C
2
C
1
÷C
2
÷1 1 ÷
C
1
C
1
÷C
2
C
1
C
1
÷C
2
_
_
_
_
_
_
_
_
_
_
_
_
_
_
DR
1
R
1
DR
2
R
2
DC
1
C
1
DC
2
C
2
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
(16)
Note that the ﬁrst line of Eq. (16) is equal to the last part
of Eq. (12).
The tuning situation is simpler if the matrix elements
above the main diagonal are zero as was assumed for an
CIRCUIT TUNING 635
arbitrary different circuit in Eq. (17a):
Do
0
o
0
DQ
Q
DK
K
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
=
S
o
0
R
1
0 0 S
o
0
C
2
S
Q
R
1
S
Q
R
2
0 S
Q
C2
S
K
R
1
S
K
R
2
S
K
C
1
S
K
C
2
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
DR
1
R
1
DR
2
R
2
DC
1
C
1
DC
2
C
2
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
=
S
o
0
R
1
0 0
S
Q
R
1
S
Q
R
2
0
S
K
R1
S
K
R2
S
K
C
1
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
DR
1
R
1
DR
2
R
2
DC
1
C
1
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
÷
S
o
0
C
2
S
Q
C
2
S
K
C
2
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
DC
2
C
2
(17a)
Here the sensitivities to C
2
are irrelevant because C
2
is a free
parameter and is assumed ﬁxed so that the effects of C
2
tolerances can be corrected by varying the remaining ele
ments. We see then that ﬁrst o
0
can be tuned by R
1
, next Qis
tuned by R
2
without disturbing o
0
because S
o
0
R
2
is zero, and
ﬁnally K is tuned by C
1
without disturbing the previous two
adjustments. Thus a sensitivity matrix of the structure
indicated in Eq. (17a) with elements above the main diagonal
equal to zero permits sequential ‘‘noninteractive’’ tuning if
the tuning order is chosen correctly. Completely noninterac
tive tuning without regard to the tuning order requires all
elements in the sensitivity matrix off the main diagonal to be
zero as indicated for another circuit in Eq. (17b):
Do
0
o
0
DQ
Q
DK
K
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
÷
S
o
0
C
2
S
Q
C
2
S
K
C
2
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
DC
2
C
2
=
S
o
0
R
1
0 0
0 S
Q
R
2
0
0 0 S
K
C
1
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
DR
1
R
1
DR
2
R
2
DC
1
C
1
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
(17b)
As can be veriﬁed readily, each component affects only one
circuit parameter. Again, sensitivities to C
2
are irrelevant
because C
2
is ﬁxed, and the effects of its tolerances can be
corrected by the remaining components.
An important observation on the effects of tolerances on
circuit parameters and the resultant need for tuning can
be made from Eq. (16). We see that the sensitivities of the
dimensionless parameters (parameters with no physical
unit) Q and K to the two resistors and similarly to the two
capacitors are equal in magnitude but opposite in sign.
Because dimensionless parameters are determined by
ratios of like components [see Eq. (4)], we obtain from
Eq. (4) with Eq. (10)
S
Q
R
1
= ÷S
Q
R
2
=S
Q
R
= ÷0:5
S
Q
C
1
= ÷S
Q
C
2
=S
Q
C
= ÷
1
2
C
1
÷C
2
C
1
÷C
2
(18)
Thus, the tolerances of Q are
DQ
Q
=S
Q
R
DR
1
R
1
÷
DR
2
R
2
_ _
÷S
Q
C
DC
1
C
1
÷
DC
2
C
2
_ _
(19)
with analogous expressions obtained for the gain K [see
the last line of Eq. (16)]. Thus, if the technology chosen to
implement the ﬁlter permits ratios of resistors and capa
citors to be realized accurately (i.e., if all resistors have
equal tolerances, as do all capacitors), tuning of dimen
sionless parameters will generally not be necessary. A
prime example is integrated circuit technology, where
absolute value tolerances of resistors and capacitors may
reach 20–50%, but ratios, depending mainly on processing
mask dimensions, are readily implemented with toler
ances of a fraction of 1%. As an example, assume that
the circuit in Fig. 2 was designed, as is often the case, with
two identical capacitors C
1
=C
2
=C with tolerances of
20% and that R
1
and R
2
have tolerances of 10% each
C
1
=C
2
=C
n
÷DC=C
n
(1 ÷0:2)
R
1
=R
1n
÷DR
1
=R
1n
(1÷0:1)
R
2
=R
2n
(1 ÷0:1)
(20)
where the subscript n stands for the nominal values. From
Eq. (19), we ﬁnd
DQ=[S
Q
R
(0:1 ÷0:1) ÷S
Q
C
(0:2 ÷0:2)]Q=0
Thus, the quality factor Q, depending only on ratios of like
components, is basically unaffected because all like com
ponents have equal fabrication tolerances. This result can
be conﬁrmed directly from Eq. (4), where, for equal
capacitors
Q=
1
2
ﬃﬃﬃﬃﬃﬃ
R
2
R
1
¸
=
1
2
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
R
2n
(1 ÷0:1)
R
1n
(1 ÷0:1)
¸
 Q
n
(21)
Naturally, if R
1
and R
2
are selected from different manu
facturing lots, or if R
1
and R
2
are from physically different
fabrication processes (such as a carbon and a metal ﬁlm
resistor), tolerances cannot be assumed to be equal, Q
errors are not zero, and tuning will be required.
636 CIRCUIT TUNING
The situation is quite different for any dimensioned
circuit parameter, that is, a parameter with a physical
unit (e.g., a frequency or time constant, or a voltage or a
current). Such parameters are determined by absolute
values of components, as seen for o
0
in Eq. (4). Absolute
values, depending on physical process parameters such as
resistivity, permittivity, or diffusion depth, are very difﬁ
cult to control and will usually suffer from large process
variations. Thus, for the component tolerances in Eq. (20),
sensitivity calculations predict from Eqs. (10) and (12) the
realized center frequency error
Do
0
÷
1
2
DR
1
R
1
÷
DR
2
R
2
÷
DC
1
C
1
÷
DC
2
C
2
_ _
o
0
= ÷
1
2
(0:1 ÷0:1÷0:2 ÷0:2) = ÷0:3o
0
(22a)
that is, all individual component tolerances add to a
÷30% frequency error. Again, the validity of this sensi
tivity result can be conﬁrmed directly from Eq. (4):
o
0
=
1
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
R
1
R
2
C
1
C
2
_ =
1
C
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
R
1
R
2
_
=
1
C
n
(1 ÷0:2)
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
R
1n
R
2n
(1 ÷0:1 ÷0:1 ÷0:01)
_

o
0n
(1÷:02)
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
1 ÷0:2
_ 
o
0n
(1 ÷0:2)(1 ÷0:1)
=
o
0n
1÷0:32 ( )
 o
0n
(1 ÷0:24)
(22b)
The difference between the exact result in Eq. (22b) and
the one obtained via the sensitivity approach in Eq. (22a)
arises because the latter assumes incremental component
changes whereas the former assumed the relatively large
changes of 10 and 20%. The center frequency o
0
is
approximately 25–30% smaller than speciﬁed and must
be corrected by tuning. This can be accomplished, for
example, by trimming the two resistors to be 27% smaller
than their fabricated values,
R
1
=R
1n
(1÷0:1)(1 ÷0:27)  R
1n
(1 ÷0:2)
R
2
 R
2n
(1 ÷0:2)
so that sensitivity calculations yield
Do
0
 ÷0:5(÷0:2 ÷0:2 ÷0:2÷0:2) =0
More exact deterministic tuning requires the resistors to
be trimmed to 24.2% smaller than the fabricated value as
shown in Eq. (23):
o
0

1
C
n
(1÷0:2)
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
R
1n
R
2n
_
(1÷0:1)(1 ÷d)
=
o
0n
1:32(1 ÷d)
=o
0n
(23)
where d is the trimming change to be applied to the
resistors as fabricated. Equation (23) results in d =0.242.
Of course, o
0
tuning could have been accomplished by
adjusting only one of the resistors by a larger amount; we
trimmed both resistors by equal amounts to maintain the
value of their ratio that determines Q according to
Eq. (21), thereby avoiding the need to retune Q.
2. TUNING DISCRETE CIRCUITS
Whether implemented on a printedcircuit board, with
chip and thin or thickﬁlm components in hybrid form, by
use of wirewrapping, or in any other technology, an
advantage of discrete circuits for the purpose of tuning
is that circuit elements are accessible individually before
or after assembly for deterministic or functional adjusting.
Thus, after a circuit is assembled and found not to meet
the design speciﬁcations, the circuit components (most
commonly the resistors or inductors) can be varied until
the performance is as required. All the previous general
discussion applies to the rest of the article, so we shall
present only those special techniques and considerations
that have been found particularly useful or important for
passive and active ﬁlters.
2.1. Passive Filters
Discrete passive ﬁlters are almost always implemented as
lossless ladder circuits; that is, the components are in
ductors L and capacitors C as is illustrated in the typical
circuit in Fig. 3. These LC ﬁlters are designed such that
the maximum signal power is transmitted from a resistive
source to a resistive load in the frequency range of
interest; a brief treatment can be found in Schaumann
et al. [3, Chapter 2, pp. 71–123]. As pointed out in our
earlier discussion, accurate ﬁlter behavior depends on
precise element values so that it is normally necessary
to trim components. This tuning is almost always accom
plished via variable inductors whose values are changed
by screwing a ferrite slug (the ‘‘trimmer’’) into or out of the
magnetic core of the inductive windings. Variable discrete
capacitors are hard to construct, expensive, and rarely
used.
LC ﬁlters have the advantage of very low sensitivities
to all their elements (see Schaumann et al. [3], Chapters 2
and 3, pp. 71–196), which makes it possible to assemble
Figure 3. Sixthorder LC lowpass ﬁlter. The ﬁlter is to realize a
maximally ﬂat passband with a 2 dB bandwidth of f
c
=6 kHz,
minimum stopband attenuation a
s
=67.5dB with transmission
zeros at 12 and 24kHz. The nominal components are listed in
Table 1. Note that at DC the ﬁlter has ÷20 log[R
2
/(R
1
÷R
2
)] =
6.02dB attenuation.
CIRCUIT TUNING 637
the ﬁlter using less expensive widetolerance components.
This property is further enhanced by the fact that
lossless ladders are very easy to tune so that large
tolerances of one component can be compensated by
accurately tuning another. For example, the resonant
frequency f
0
=1=
ﬃﬃﬃﬃﬃﬃﬃ
LC
_
of an LC resonance circuit has
715% tolerances if both L and C have 715% tolerances;
if L is then trimmed to 70.5% of its correct value for the
existing capacitor (with 715% tolerances), f
0
is accurate
to within 0.25% without requiring any narrower manu
facturing tolerances. Without tuning, a 0.25% f
0
error
would require the same narrow 0.25% tolerance in both
components, which is likely more expensive than a simple
tuning step.
It is well known that lossless ladders can be tuned quite
accurately simply by adjusting the components to realize
the prescribed transmission zeros (see Heinlein and
Holmes [4], Section 12.3, pp. 591–604, and Christian [5],
Chapter 8, pp. 167–176). Transmission zeros, frequencies
where the attenuation is inﬁnite, usually depend on only
two elements: a capacitor and an inductor in a parallel
resonant circuit (see Fig. 3) with the parallel tank circuits
L
1
, C
1
and L
2
, C
2
in the series branches of the ﬁlter, or
alternatively with series LC resonance circuits in the
shunt branches. The resonant frequencies f
zi
=1=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
L
i
C
i
_
,
i =1, 2, of the LC tank circuits are not affected by other
elements in the ﬁlter, so that tuning is largely noninter
active. As mentioned, the effect of the tolerances of one
component, say, C, are corrected by tuning L. It is per
formed by adjusting the inductors for maximum attenua
tion at the readily identiﬁed frequencies of zero
transmission while observing the response of the complete
manufactured ﬁlter on a network analyzer. Tuning
accuracies of the transmission zeros of 0.05% or less
should be aimed at. Such tuning of the transmission
zeros is almost always sufﬁcient even if the circuit
elements have fairly large tolerances (see Heinlein and
Holmes [4], Section 12.3, pp. 594–604). If even better
accuracy is needed, adjustments of those inductors
that do not cause ﬁnite transmission zeros, such as L
3
in
Fig. 3, may need to be performed (see Christian [5],
Chapter 8, pp. 167–176). For instance, consider the ﬁlter
in Fig. 3 realized with unreasonably large tolerances
of 715%, using the components shown in Table 1. This
places the two resonant frequencies at 10.3 and 20.7 kHz,
with the minimum stopband attenuation equal to only
56.7 dB; the 2dB passband corner is reduced to 5.36 kHz.
If we next tune the transmission zero frequencies to 12
and 24 kHz by adjusting only the inductors L
1
and L
2
to
23.5 and 40 mH, respectively, the minimum stopband
attenuation is increased to 57.8dB, and the 2 dB band
width of the passband is measured as f
c
=6.07kHz (refer
to Table 1).
We still note that when functional tuning is performed,
the ﬁlter must be operated with the correct terminations
for which it was designed (see Christian [5], Section 8.2,
pp. 168–173). Large performance errors, not just at DC or
low frequencies, will result if the nominal terminations
are severely altered. For example, an LC ﬁlter designed
for 600O terminations cannot be correctly tuned by con
necting it directly without terminations to a highfre
quency network analyzer whose input and source
impedances are 50 O. Also, if maintaining an accurate
narrow passband ripple is important, the tolerances of
the untuned capacitors must not be too large. Finally, we
observe that the tuning properties of passive LC ladders
translate directly to active simulations of these ﬁlters via
transconductance–C and gyrator–C circuits, which are
widely used in highfrequency integrated circuits for com
munications (see the following discussion).
2.2. Active Filters
Several differences must be kept in mind when tuning
active ﬁlters as compared to passive lossless ﬁlters, parti
cularly to ladders:
1. Active ﬁlters are almost always more sensitive to
component tolerances than LC ladders. Conse
quently, tuning is always required in practice.
2. Tuning in active ﬁlters is almost always interactive;
that is, a ﬁlter parameter depends on many or all
circuit components as discussed in connection with
the circuit in Fig. 2 and the sensitivity discussion
related to Eqs. (15) and (16). Consequently, tuning
active ﬁlters usually requires computer aids to solve
the complicated nonlinear tuning equations [see,
e.g., the relatively simple case in Eq. (4)].
3. The performance of the active devices, such as
operational ampliﬁers (op amps), and their often
large tolerances almost always strongly affects the
ﬁlter performance and must be accounted for in
design and in tuning. Because activedevice beha
vior is often hard to model or account for, functional
tuning of the fabricated circuit is normally the only
method to ensure accurate circuit performance.
In discrete active ﬁlters constructed with resistors,
capacitors, and operational ampliﬁers on a circuit board
or in thin or thickﬁlm form, tuning is almost always
performed by varying the resistors. Variable resistors,
Table 1. LC Lowpass Filter (Elements in mH, nF, and kX)
Components L
1
C
1
L
2
C
2
L
3
C
3
C
4
C
5
R
1
R
2
Nominal values 27.00 6.490 46.65 0.943 12.67 6.977 45.55 33.90 1.00 1.00
Performance f
c
=6.0 kHz at a
p
=8.03dB; f
z1
=12.0kHz, f
z2
=24.0kHz, a
s
=57.5dB
15% tolerance values 31 7.5 52 1.1 14 8 51 38 1.05 1.05
Performance untuned f
c
=5.36kHz at a
p
=8.01dB; f
z1
=10.3kHz, f
z2
=20.7kHz, a
s
=56.7dB
Tuned values 23.5 7.5 40 1.1 14 8 51 38 1.05 1.05
Performance tuned f
c
=6.07kHz at a
p
=8.03dB; f
z1
=12.0kHz, f
z2
=24.0kHz, a
s
=57.8dB
638 CIRCUIT TUNING
potentiometers, are available in many forms, technologies,
and sizes required to make the necessary adjustments.
2.2.1. SecondOrder Filters. The main building blocks
of active ﬁlters are secondorder sections, such as the
bandpass circuit in Fig. 2. Many of the tuning strategies
and concepts were presented earlier in connection with
that circuit and the discussion of sensitivity. An important
consideration when tuning an active ﬁlter is its depen
dence on the active devices as mentioned previously in
point 3 (above). To illustrate the problem, consider again
the bandpass ﬁlter in Fig. 2. The transfer function T(s) in
Eq. (1) is independent of the frequencydependent gain
A(s) of the op amp only because the analysis assumed that
the ampliﬁer is ideal, that is, it has constant and very
large (ideally inﬁnite) gain, A=N. In practice, T(s) is also
a function of A(s) as a more careful analysis shows:
T(s) =
V
2
V
1
= ÷
1
R
1
C
1
s
A s ( )
1÷A(s)
s
2
÷
1
R
2
1
C
1
÷
1
C
2
÷
1
R
1
C
1
[1÷A(s)]
_ _
s÷
1
R
1
R
2
C
1
C
2
(24)
Evidently, for A=N, Eq. (24) reduces to Eq. (1), but ﬁnite
and frequencydependent gain can cause severe changes
in T(s) in all but the lowestfrequency applications. Con
sider the often used integrator model for the operational
ampliﬁer, A(s)Eo
t
/s, where o
t
is the unity gain frequency
(or the gain–bandwidth product) of the op amp with the
typical value o
t
=2pf
t
=2p 1.5MHz. Using this simple
model, which is valid for frequencies up to about 10–20%
of f
t
, and assuming o
t
bo, the transfer function becomes
T(s) =
V
2
V
1
÷
1
C
1
R
1
s
s
2
1÷
1
o
t
C
1
R
1
_ _
÷
1
R
2
1
C
1
÷
1
C
2
_ _
s ÷
1
R
1
R
2
C
1
C
2
(25)
To get an estimate of the resulting error, let the circuit be
designed with C
1
=C
2
=C=10nF, R
1
=66.32 O and R
2
=
9.55 kO to realize the nominal parameters f
0
=20 kHz,
Q=6, and K=72. Simulation (or measurement with a
very fast op amp) shows that the resulting circuit perfor
mance is as desired. However, if the ﬁlter is implemented
with a 741type op amp with f
t
=1.5 MHz, the measured
performance indicates f
0
=18.5kHz, Q=6.85, and K=
76.75. Because of the complicated expressions involving
a real op amp, it is appropriate to use functional tuning
with the help of a network analyzer. Keeping C constant,
the resulting resistor values, R
1
=68.5 O and R
2
=8.00 kO,
lead to f
0
=20 kHz and Q=6.06. The midband gain for
these element values equals K=62.4 (remember from the
earlier discussion that K for the circuit in Fig. 2 cannot be
separately adjusted if the capacitors are predetermined).
2.2.2. HighOrder Filters. The two main methods for
realizing active ﬁlters of order greater than two are active
simulations of lossless ladders and cascading secondorder
sections. We mentioned in connection with the earlier
discussion of LC ladders that tuning of active ladder
simulations is completely analogous to that of the passive
LC ladder: the electronic circuits that simulate the in
ductors are adjusted until the transmission zeros are
implemented correctly. It remains to discuss tuning for
the most frequently used method of realizing highorder
ﬁlters, the cascading of ﬁrst and secondorder sections.
Apart from good sensitivity properties, relatively easy
tuning is a main advantage of cascade implementations
because each section performs in isolation from the others
so that it can be tuned without interactions from the rest
of the circuit. Remember, though, that each section by
itself may require interactive tuning. Figure 4 shows the
circuit structure where each of the blocks is a secondorder
section such as the ones in Figs. 2 and 5. If the total ﬁlter
order is odd, one of the sections is, of course, of ﬁrst order.
To illustrate this point, assume a fourthorder
Chebyshev lowpass ﬁlter is to be realized with a 1dB
ripple passband in 0rfr28 kHz with passband gain equal
to H=20 dB. The transfer function is found to be
T(s) =T
1
(s) T
2
(s)
=
1:66o
2
0
s
2
÷0:279o
0
s ÷0:987o
2
0
1:66o
2
0
s
2
÷0:674o
0
s ÷0:279o
2
0
(26)
with o
0
=2p 28,000s
÷1
=175.9310
3
s
÷1
(see Schau
mann et al. [3], Section 1.6, pp. 36–64). Let the function be
realized by two sections of the form shown in Fig. 5.
Assuming that the op amps are ideal, the transfer function
of the lowpass section is readily derived as
V
2
V
1
=
Ko
2
0
s
2
÷s
o
0
Q
÷o
2
0
=
a
1
a
2
1
C
1
R
1
C
2
R
2
s
2
÷s
1
C
1
R
1
÷
1 ÷a
1
a
2
C
2
R
2
_ _
÷
1
C
1
R
1
C
2
R
2
(27)
Figure 4. Cascade realization of 2nthorder ﬁlter. The n second
order sections do not interact with each other and can be tuned
independently; that is, each section T
i
can be tuned to its nominal
values o
i
, Q
i
, and H
i
, i =1; 2; . . . ; n, without being affected by the
other sections.
CIRCUIT TUNING 639
If the op amp gain is modeled as A(s) =o
t
/s, a
i
is to be
replaced by
a
i
=
a
i
1 ÷a
i
=A(s)

a
i
1 ÷sa
i
=o
t
(28)
We observe again that the circuit parameters o
0
, Q, and
gain K are functions of all the circuit elements so that
design and tuning of each section will require iterative
procedures, although Section 1 is independent of Section 2
as just discussed. Because there are six ‘‘components’’ (R
1
,
R
2
, C
1
, C
2
, a
1
, and a
2
) and only three parameters, some
simplifying design choices can be made. Choosing C
1
=C
2
=
C, R
1
=R, and R
2
=k
2
R (and assuming ideal op amps), Eq.
(27) leads to in the expressions
o
0
=
1
kRC
; Q=
1
k ÷
1
k
(1 ÷K)
; K =a
1
a
2
(29)
The circuit is designed by ﬁrst computing k from the given
values Q and K; next we choose a suitable capacitor value
C and calculate R=1/(ko
0
C). Finally, we determine the
feedback resistors on the two op amps. Because only
the product a
1
a
2
is relevant, we choose a
1
a
2
=a
2
=K
(i:e:; a =
ﬃﬃﬃﬃ
K
_
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
1:66
_
=1:288). Working through the de
sign equations and choosing all capacitors equal to C=
150pF (standard 5% values) and R
0
=10 kO results in
(a ÷1)R
0
=2.87 kO for both sections: k=0.965, R
1
=
40.2 kO, R
2
=36.5 kO for Section 1 and k=1.671, R
1
=
42.2 kO, R
2
=120.1 kO for Section 2. All resistors have
standard 1% tolerance values. Building the circuit with
741type op amps with f
t
=1.5MHz results in a ripple
width of almost 3 dB, the reduced cutoff frequency of
27.2 kHz, and noticeable peaking at the band edge.
Thus, tuning is required. The errors can be attributed
largely to the 5% capacitor errors and the transfer func
tion changes as a result of the ﬁnite f
t
in Eq. (28).
To accomplish tuning in this case, deterministic tuning
may be employed if careful modeling of the op amp
behavior, using Eq. (28), and of parasitic effects is used
and if the untuned components (the capacitors) are mea
sured carefully and accurately. Because of the many
interacting effects in the secondorder sections, using a
computer program to solve the coupled nonlinear equa
tions is unavoidable, and the resistors are trimmed to
their computed values. Functional tuning in this case may
be more convenient, as well as more reliable in practice.
For this purpose, the circuit is analyzed, and sensitivities
are computed to help understand which components affect
the circuit parameters most strongly. Because the sections
do not interact, the highorder circuit is separated into its
sections, and each section’s functional performance is
measured and adjusted on a network analyzer. After the
performance of all secondorder blocks is found to lie
within the speciﬁed tolerances, the sections are recon
nected in cascade.
3. TUNING INTEGRATED CIRCUITS
With the increasing demand for fully integrated micro
electronic systems, naturally, analog circuits will have to
be placed on an integrated circuit (IC) along with digital
ones. Of considerable interest are communication circuits
where bandwidths may reach many megahertz. Numer
ous applications call for onchip highfrequency analog
ﬁlters. Their frequency parameters, which in discrete
active ﬁlters are set by RC time constants, are in inte
grated ﬁlters most often designed with voltagetocurrent
converters (transconductors), I
o
=g
m
V
i
, and capacitors
(i.e., as o=1/t =g
m
/C). As discussed earlier, ﬁlter perfor
mance must be tuned regardless of the implementation
method because fabrication tolerances and parasitic ef
fects are generally too large for ﬁlters to work correctly
without adjustment. Understandably, tuning in the tradi
tional sense is impossible when the complete circuit is
integrated on an IC because individual components are
not accessible and cannot be varied. To handle this
problem, several techniques have been developed. They
permit tuning the circuits electronically by varying the
bias voltages V
B
or bias currents I
B
of the active electronic
components (transconductors or ampliﬁers). In the usual
approach, the performance of the fabricated circuit is
compared to a suitably chosen accurate reference, such
as an external precision resistor R
e
to set the value of
an electronic onchip transconductance to g
m
=1/R
e
or
to a reference frequency o
r
to set the time constant to
C/g
m
=1/o
r
. This approach is indeed used in practice,
where often the external parameters, R
e
or o
e
, are ad
justed manually to the required tolerances. Tuning can be
handled by connecting the circuit to be tuned into an on
chip control loop, which automatically adjusts bias vol
tages or currents until the errors are reduced to zero or an
acceptable level (see Schaumann et al. [3], Section 7.3, pp.
418–446, and Johns and Martin [6], Section 15.7, pp. 626–
635). (A particularly useful reference is Tsividis and Voor
man [7]; it contains papers on all aspects of integrated
ﬁlters, including tuning). Naturally, this process requires
that the circuit be designed to be tunable, that is, that the
components are variable over a range sufﬁciently wide to
permit errors caused by fabrication tolerances or tempera
ture drifts to be recovered. We also must try to keep the
tuning circuitry relatively simple because chip area and
power consumption are at a premium. Although digital
tuning schemes are conceptually attractive, analog meth
ods are often preferred. The reason is the need to minimize
or eliminate generating digital (switching) noise, which
Figure 5. Twoampliﬁer active lowpass ﬁlter.
640 CIRCUIT TUNING
can enter the sensitive analog signal path through para
sitic capacitive coupling or through the substrate, causing
the dynamic range or the signaltonoise ratio to deterio
rate.
3.1. Automatic Tuning
Let us illustrate the concepts and techniques with a
simple secondorder example. Higherorder ﬁlters are
treated in an entirely analogous fashion; the principles
do not change. Consider the g
m
–C ﬁlter in Fig. 6, which
realizes the transfer function
T(s) =
V
0
V
1
=
as
2
÷s a
g
m1
C
1
÷b
g
m2
C
2
_ _
÷
g
m0
g
m2
C
1
C
2
s
2
÷s
g
m1
C
1
÷
g
m1
g
m2
C
1
C
2
(30)
with pole frequency and pole Q equal to
o
0
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
g
m1
g
m2
C
1
C
2
_
; Q=
o
0
C
1
g
m1
=
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
C
1
=C
2
g
m1
=g
m2
¸
(31)
Comparing Eq. (31) to Eq. (2) indicates that the ﬁlter
parameters for this technology are determined in funda
mentally the same way as for discrete active circuits: the
frequency is determined by time constants (C
i
/g
mi
) and the
quality factor, by ratios of like components. Analogous
statements are true for the numerator coefﬁcients of T(s).
We can conclude then that, in principle, tuning can
proceed in a manner quite similar to the one discussed
in the beginning of this article if we can just develop a
procedure for varying the onchip components. To gain an
understanding of what needs to be tuned in an integrated
ﬁlter, let us introduce a more convenient notation that
uses the ratios of the components to some suitably chosen
unit values g
m
and C
g
mi
=g
i
g
m
; C
i
=c
i
C; i =1; 2 and o
u
=
g
m
C
(32)
where o
u
is a unit frequency parameter and g
i
and c
i
are
the dimensionless component ratios. With this notation,
Eq. (30) becomes
T(s) =
V
0
V
1
=
as
2
÷s a
g
1
c
1
÷b
g
2
c
2
_ _
o
u
÷
g
0
g
2
c
1
c
2
o
2
u
s
2
÷s
g
1
c
1
o
u
÷
g
1
g
2
c
1
c
2
o
2
u
(33)
Casting the transfer function in the form shown in
Eq. (33) makes clear that the coefﬁcient of s
i
is propor
tional to o
n÷i
u
; where n is the order of the ﬁlter, n=2 in
Eq. (33); the constants of proportionality are determined
by ratios of like components, which are very accurately
designable with IC technology. The same is true for ﬁlters
of arbitrary order. For example, the pole frequency for the
circuit in Fig. 6 is determined as o
u
times a designable
quantity, o
0
=o
u
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
g
1
g
2
=(c
1
c
2
)
_
. We may conclude therefore
that it is only necessary to tune o
u
=g
m
/C, which, as
stated earlier, as a ratio of two electrically dissimilar
components will have large fabrication tolerances. In
addition, the electronic circuit that implements the trans
conductance g
m
depends on temperature, bias, and other
conditions, so that o
u
can be expected to drift during
operation. It can be seen from Eq. (33) that o
u
simply
scales the frequency, that is, the only effect of varying o
u
is a
shift of the ﬁlter’s transfer function along the frequency axis.
We stated earlier that tuning a time constant, or, in the
present case, the frequency parameter o
u
, is accomplished
by equating it via a control loop to an external reference,
in this case a reference frequency o
R
such as a clock
frequency. Conceptually, the block diagram in Fig. 7 shows
the method [8]. The control loop equates the inaccurate
unit frequency o
u
=g
m
/C to the accurate reference fre
quency o
R
in the following way: o
R
is chosen in the
vicinity of the most critical frequency parameters of the
ﬁlter (the band edge for a lowpass, midband for a bandpass
ﬁlter), where sensitivities are highest. The transconduc
tance g
m
to be tuned is assumed to be proportional to the
bias voltage V
B
, such that g
m
=kV
B
, where k is a constant
of proportionality with units of A/V
2
. g
m
generates an
output current I =g
m
V
R
,which results in the capacitor
voltage V
C
=g
m
V
R
/(jo
R
C). The two matched peak detec
tors PD convert the two signals V
R
and V
C
to their DC
Figure 6. A general secondorder transconductance–C ﬁlter. The
circuit realizes arbitrary zeros by feeding the input signal into
portions bC
1
and aC
2
of the capacitors C
1
and C
2
.
Figure 7. Automatic control loop to set o
u
=g
m
/C via an applied
reference signal V
R
with frequency o
R
. The capacitor voltage
equals V
C
=V
R
(g
m
/jo
R
C), which makes the control current I
c
=
g
mc
V
R
(1÷g
m
/jo
R
C). The operation is explained in the text.
CIRCUIT TUNING 641
peak values, so that any phase differences do not matter
when comparing the signals at the input of g
mc
. The DC
output current I
c
=g
mc
V
R
{1–[g
m
/(jo
R
C)]} of the control–
transconductance g
mc
charges the storage capacitor C
c
to
the required bias voltage V
B
for the transconductance g
m
.
The values g
mc
and C
c
determine the loop gain; they
inﬂuence the speed of conversion but are otherwise not
critical. If the value of g
m
gets too large because of
fabrication tolerances, temperature, or other effects, I
c
becomes negative, C
c
discharges, and V
B
, that is g
m
=
kV
B
, is reduced. Conversely, if g
m
is too small, I
c
becomes
positive and charges C
c
, and the feedback loop acts to
increase V
B
and g
m
. The loop stabilizes when V
C
and V
R
are equal, that is, when g
m
(V
B
)/C is equal to the accurate
reference frequency o
R
. The g
mc
–C
c
combination is, of
course, an integrator with ideally inﬁnite DC gain to
amplify the shrinking error signal at the input of g
mc
. In
practice, the openloop DC gain of a transconductance of
35–50 dB is more than adequate. Note that the loop sets
the value of o
u
to o
R
regardless of the causes of any errors:
fabrication tolerances, parasitic effects, temperature
drifts, aging, or changes in DC bias.
We point out that although the scheme just discussed
only varies g
m
, it actually controls the time constant C/g
m
;
that is, errors in both g
m
and C are accounted for. If one
wishes to control only g
m
, the capacitor C in Fig. 7 is
replaced by an accurate resistor R
e
, and the feedback loop
will converge to g
m
=1/R
e
.
Note that the feedback loop in Fig. 7 directly controls
only the transconductance g
m
(as does the frequency
control circuit in Fig. 8) such that the unit frequency
parameter o
u
within the control circuit is realized cor
rectly. The actual ﬁlter is not tuned. However, good
matching and tracking can be assumed across the IC
because all g
m
cells are on the same chip and subject to
the same errorcausing effects. This assumes that the
ratios g
i
deﬁned in Eq. (32) are not so large that matching
problems will arise and that care is taken to account for
(model the effect of) ﬁlter parasitics in the control circuit.
The same is true for the unit capacitor C in the control
loop and the ﬁlter capacitors (again, if the ratios c
i
are not
too large). Consequently, the control bias current I
B
can be
sent to all the main ﬁlter’s transconductance cells as
indicated in Fig. 7 and thereby tune the ﬁlter. Clearly,
this scheme depends on good matching properties across
the IC chip. Accurate tuning cannot be performed if
matching and tracking cannot be relied upon or, in other
words, if the g
m
–C circuit in the control loop is not a good
representative model of the ﬁlter cells.
An alternative method for frequency tuning (see Schau
mann et al. [3], Section 7.3, pp. 418–446, and Johns and
Martin [6], Section 15.7, pp. 626–635) relies on phase
locked loops (see Johns and Martin [6], Chapter 16, pp.
648–695). The top half of Fig. 8 shows the principle. A
sinusoidal reference signal V
R
at o=o
R
and the output of
a voltagecontrolled oscillator (fVCO) at o
vco
are con
verted to square waves by two matched limiters. Their
outputs enter an XOR gate acting as a phase detector whose
output contains a DC component proportional to the
frequency difference Do=o
vco
÷o
R
of the two input sig
nals. The lowpass ﬁlter LPF 1 eliminates second and
higherorder harmonics of the XOR output and sends the
DC component to the oscillator fVCO, locking its fre
quency to o
R
. Just as the g
m
–C circuit in Fig. 7, the
oscillator is designed with transconductances and capaci
tors to represent (model) any frequency parameter errors
of the ﬁlter to be tuned so that, relying on matching, the
ﬁlter is tuned correctly by applying the tuning signal also
to its g
m
cells. The lowpass ﬁlter LPF 2 is used to clean the
tuning signal V
f
further before applying it to the ﬁlter.
We saw in Eq. (33) that all ﬁlter parameters depend,
apart from o
u
, only on ratios of like components and are,
therefore, accurately manufacturable and should require
no tuning. This is indeed correct for moderate frequencies
and ﬁlters with relatively low Q. However, Q is extremely
sensitive (see Schaumann et al. [3], Chapter 7, pp. 410–
486) to small parasitic phase errors in the feedback loops
of active ﬁlters, so that Q errors may call for tuning as
well, especially as operating frequencies increase. The
problem is handled in much the same way as frequency
tuning. One devises a model (the Q model in Fig. 8) that
represents the Q errors to be expected in the ﬁlter and
encloses this model circuit in a control loop where feed
back acts to reduce the error to zero. Figure 8 illustrates
the principle. In the Q control loop, a QVCO (tuned
correctly by the applied frequency control signal V
f
) sends
a test signal to the Q model that is designed to represent
correctly the Q errors to be expected in the ﬁlter to be
tuned, and through a peak detector PD to an ampliﬁer of
gain K. K is the gain of an accurately designable DC
ampliﬁer. Note that the positions of PD and K could be
interchanged in principle, but a switch would require that
K is the less wellcontrolled gain of a highfrequency
ampliﬁer. The output of the Q model goes through a
second (matched) peak detector. Rather than measuring
Q directly, which is very difﬁcult in practice, because it
would require accurate measurements of two amplitudes
and two frequencies, the operation relies on the fact that Q
errors are usually proportional to magnitude errors. The
Figure 8. Dualcontrol looptuning system for tuning frequency
parameters and quality factors of an integrated ﬁlter. Note that
the frequency loop converges always, but for the Q loop to
converge on the correct Q value, the frequency must be correct.
Details of the operation are explained in the text.
642 CIRCUIT TUNING
diagram in Fig. 8 assumes that for correct Q the output of
the Q model is K times as large as its input so that for
correct Q the inputs of the comparator are equal. The DC
error signal V
Q
resulting from the comparison is fed back
to the Q model circuit to adjust the bias voltages appro
priately, as well as to the ﬁlter. In these two interacting
control loops, the frequency loop will converge indepen
dently of the Q control loop, but to converge on the correct
value of Q, the frequency must be accurate. Hence, the two
loops must operate together. The correct operation and
convergence of the frequency and Q control scheme in Fig.
8 has been veriﬁed by experiments (see Schaumann et al.
[3], Chapter 7, pp. 410–486) but because of the increased
noise, power consumption, and chip area needed for the
control circuitry, the method has not found its way into
commercial applications.
BIBLIOGRAPHY
1. G. Moschytz, Linear Integrated Networks: Design, Van Nos
trandReinhold, New York, 1975.
2. P. Bowron and F. W. Stevenson, Active Filters for Commu
nications and Instrumentation, McGrawHill, Maidenhead,
UK, 1979.
3. R. Schaumann, M.S. Ghausi, and K.R. Laker, Design of
Analog Filters: Passive, Active RC and Switched Capacitor,
PrenticeHall, Englewood Cliffs, NJ, 1990.
4. W. E. Heinlein and W. H. Holmes, Active Filters for Integrated
Circuits, R. Oldenburg, Munich, 1974.
5. E. Christian, LC Filters: Design, Testing and Manufacturing,
Wiley, New York, 1983.
6. D. A. Johns and K. Martin, Analog Integrated Circuit Design,
Wiley, New York, 1997.
7. Y. Tsividis and J. A. Voorman, eds., Integrated Continuous
Time Filters: Principles, Design and Implementations, IEEE
Press, Piscataway, NJ, 1993.
8. J. F. Parker and K. W. Current, A CMOS continuoustime
bandpass ﬁlter with peakdetectionbased automatic tuning,
Int. J. Electron. 1996(5):551–564 (1996).
CIRCULAR WAVEGUIDES
1
CONSTANTINE A. BALANIS
Arizona State University
Tempe, Arizona
(edited by Eric Holzman
Northrop Grumman Electronic
Systems, Baltimore, Maryland)
1. INTRODUCTION
The circular waveguide is occasionally used as an alter
native to the rectangular waveguide. Like other wave
guides constructed from a single, enclosed conductor, the
circular waveguide supports transverse electric (TE) and
transverse magnetic (TM) modes. These modes have a
cutoff frequency, below which electromagnetic energy is
severely attenuated. Circular waveguide’s round cross sec
tion makes it easy to machine, and it is often used to feed
conical horns. Further, the TE
0n
modes of circular wave
guide have very low attenuation. A disadvantage of
circular waveguide is its limited dominant mode band
width, which, compared to rectangular waveguide’s max
imum bandwidth of 2–1, is only 1.3. In addition, the
polarization of the dominant mode is arbitrary, so that
discontinuities can easily excite unwanted crosspolarized
components.
In this article, the electromagnetic features of the cir
cular waveguide are summarized, including the trans
verse and longitudinal ﬁelds, the cutoff frequencies, the
propagation and attenuation constants, and the wave im
pedances of all transverse electric and transverse mag
netic modes.
2. TRANSVERSE ELECTRIC (TE
Z
) MODES
The transverse electric to z (TE
z
) modes can be derived by
letting the vector potential A and F be equal to
A=0 (1a)
F= ^ aa
z
F
z
( r; f; z) (1b)
The vector potential F must satisfy the vector wave equa
tion, which reduces the F of (1b) to
V
2
F
z
( r; f; z) ÷b
2
F
z
(r; f; z) =0 (2)
When expanded in cylindrical coordinates, (2) reduces to
@
2
F
z
@r
2
÷
1
r
@F
z
@r
÷
1
r
2
@
2
F
z
@f
2
÷
@
2
F
z
@z
2
÷b
2
F
z
=0 (3)
whose solution for the geometry of Fig. 1 is of the form
F
z
( r; f; z) =[ A
1
J
m
( b
r
r) ÷B
1
Y
m
( b
r
r)]
[C
2
cos(mf) ÷D
2
sin(mf)]
[ A
3
e
÷jb
z
z
÷B
3
e
÷jb
z
z
]
(4a)
where
b
2
r
÷b
2
z
=b
2
(4b)
The constants A
1
, B
1
, C
2
, D
2
, A
3
, B
3
, m, b
r
, and b
z
can be
found using the boundary conditions of
E
f
( r=a; f; z) =0 (5a)
The fields must be finite everywhere (5b)
The fields must repeat every 2p radians in f (5c)
1
This article is derived from material in Advanced Engineering
Electromagnetics, by Constantine Balanis, Wiley, NewYork, 1989,
Sect. 9.2.
CIRCULAR WAVEGUIDES 643
Next Page
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue listening from where you left off, or restart the preview.