06a

FABRYPEROT RESONATORS
RONAN SAULEAU
University of Rennes
Rennes, France
1. THE PLANE-PARALLEL FABRYPEROT INTERFEROMETER
1.1. Historical Background and Generalities
Two French physicists (Alfred Perot and Charles Fabry)
developed the classical FabryPerot interferometer (FPI)
or etalon at the beginning of the twentieth century. This
novel form of interference device was based on multiple
reections of waves between two closely spaced and highly
reecting mirrors (the original thinly silvered plane glass
mirrors were adjusted to be as at and parallel to each
other as possible). Alternative but conceptually equivalent
elements are (1) the solid etalon made from some low-
loss materials (e.g., fused quartz or sapphire) or (2) the
Fabry-Perot (FP) cavity whose reectors are partially
transparent metal mirrors (e.g., perforated plates) or mul-
tilayer dielectric coatings (e.g., Bragg mirrors). The modal
analysis of resonators with spherical mirrors is beyond the
scope of this article. Nevertheless, all the abovementioned
topologies are widely used at microwave frequencies; their
choice depends on the envisaged applications, as described
in Section 3.
The general concepts and theoretical analysis of these
resonant optical cavities were published by George Airy
about the middle of the nineteenth century. As we will see,
plane-parallel FPIs have sharp resonances or transmission
passbands at discrete frequencies. They thus behave as
narrowband frequency lters. Their main characteristics
(standing-wave resonance conditions, resonant frequen-
cies, fringe contrast, free spectral range, half-power band-
width, Q factor, reectivity nesse) can be found in any
general textbook dealing with optical resonators e.g., [1,2].
However, for consistency, these parameters are voluntarily
derived in Section 1.2 using a ray approach. Although ap-
proximate, this analytical treatment reveals itself very
fruitful for the analysis of FP-based devices operating at
microwave and millimeter-wave frequencies (Section 2).
In their original form, FPIs used only at reecting
surfaces, and the spacing between the mirrors was usually
smaller than, or at most on the same scale as, the trans-
verse diameters of the mirrors. Moreover, the FP interfer-
ometers were usually illuminated with a converging or
diverging beam having a spread of angular directions,
which resulted in FabryPerot rings transmitted
through the interferometer in given discrete angular di-
rections. Thus FPIs are also spatial lters. In particular,
we will show in Section 3 that the high directivity of FP-
based resonator antennas originates from this property.
Since its invention, the FPI has continued to evolve and
nd many new elds of applications, ranging, for instance,
from astronomy, astrophysics, atomic spectroscopy, optics,
and similar applications to metrology, optical bistability,
infrared, sensors, and plasma physics. A detailed and well-
documented description of the FPI is given by Vaughan [3]
at optical wavelengths. Here we concentrate only on radio-
frequency and microwave implementations of FP resona-
tors, such as quasioptical frequency lters, highly directive
antennas, open resonators, or power combiners (Section 3).
1.2. Ray Analysis
The simplest way to dene the basic properties of plane-
parallel FPI is to use an innite plane-wave model, with
the plane waves assumed to be arriving at either normal
or oblique incidence. The standard formulas developed
here describe the transmission properties of FP etalons as
a function of (1) the operating frequency and (2) the mirror
spacing and reectivity (the oblique incidence case is
introduced in Section 3.1). The Gaussian beam optics [2]
or full-wave methods (Section 2.1.3 ) should be applied to
account for (1) the transverse width or shape of the two
mirrors, (2) the transverse eld variations, or (3) the com-
plex geometry of reectors.
The geometry of a plane-parallel FPI, depicted in Fig. 1,
consists of two semitransparent plane mirrors, M
1
and M
2
,
separated by a lossy medium of thickness D and refractive
index n (n=e
1=2
r
). As the mirrors are not conned by an
aperture, diffraction is negligible and beam propagation
inside the FPI can be described using geometric optics.
Losses generated by scattering and absorption are incor-
porated in our model using the absorption coefcient a. A
monochromatic plane wave illuminates the FPI under an
incident angle y. Its amplitude and angular frequency are
labeled E
i
and o=2pf , respectively (f =c
0
=l
0
is the oper-
ating frequency; c
0
and l
0
are the speed of light and wave-
length in vacuum). The incident light is partially reected
and partially transmitted by the input mirror M
1
. The
transmitted eld travels to the output mirror M
2
, where it
is again partially reected and transmitted. The many re-
ected components add to produce a total eld traveling in
each direction. Consequently multiple-beam interference
F
(r
1
,t
1
) (r
2
,t
2
)
D
Incident
Reflected
Transmitted
Mirror M
1
Mirror M
2
(n,:)
0 0
Figure 1. Schematic diagram of a plane-parallel FabryPerot in-
terferometer (FPI) illuminated by a slightly off-axis incident
plane wave. The two partially reective mirrors M
i
are charac-
terized by their reection and transmission coefcients r
i
and t
i
(i =1,2).
1381
occurs. Two important points should be kept in
mind: (1) the reection and transmission coefcients
at the interfaces are not the same for rays traveling in
different directions; and (2) at optical wavelengths, highly
reecting mirrors of modern FPI are coated with multi-
layer dielectric thin lms, whereas they often consist of
metallic grids (Section 2) from microwave to submillimeter
wavelengths. Hence their reection and transmission co-
efcients are in general complex quantities (Section 2.1.4),
unlike the case of simple dielectric slabs for which the
phase shift is either 01 or 1801 depending on the dielectric
contrast 1:n.
For convenience, we now assume that the mirrors are
located in free space (n=1). {r
1
,t
1
} and {r
2
,t
2
} denote the
(frequency-dependent) reection and transmission coef-
cients of M
1
and M
2
, respectively. With these notations,
the reected and transmitted elds, E
r
and E
t
, at normal
incidence (y =0) are given by
E
r
=E
i
{r
1
r
2
t
2
1
exp(j2j) exp(a2D)
r
1
r
2
2
t
2
1
exp(j4j) exp(a4D) ]
(1)
E
t
=E
i
{t
1
t
2
exp(jj) exp(aD)
t
1
t
2
r
1
r
2
exp(j3j) exp(a3D)
t
1
t
2
(r
1
r
2
)
2
exp(j5j) exp(a5D) ]
(2)
where j=(2p=l
0
)D represents the phase change due to
the optical path through the device.
Since the magnitude of the factor exp(j2j) exp(a2D)
is less than unity, both summations (1) and (2) are con-
vergent geometric progressions
E
r
E
i
= r
1
t
2
1
r
2
exp(j2j) exp(2aD)
1 r
1
r
2
exp(j2j) exp(2aD)
_ _
(3)
E
t
E
i
=
t
1
t
2
exp(jj) exp(aD)
1 r
1
r
2
exp(j2j) exp(2aD)
_ _
(4)
The total power transmitted by the interferometer T
FP
is
written
where t
i
=T
1=2
i
exp(jt
i
) and r
i
=R
1=2
i
exp(jr
i
); i =1,2. T
i
and R
i
are known as the power transmittance and reec-
tance of mirrors M
i
, respectively.
Neglecting all absorption processes (a =0, T
i
=1R
i
,
i =1,2), we obtain
T
FP
=
(1 R
1
)(1 R
2
)
1 2(R
1
R
2
)
1=2
cos (2j r
1
r
2
) R
1
R
2
(6)
Using the identity
1 cos C=2 sin
2
C
2
_ _
(7)
T
FP
can now be written
T
FP
=
(1 R
1
)(1 R
2
)
(1 (R
1
R
2
)
1=2
)
2
4(R
1
R
2
)
1=2
sin
2
2j r
1
r
2
2
_ _
(8)
Equation (8) is known as the Airy formula.
Similarly, the power reectance R
FP
of the FPI can be
computed by multiplying the reection coefcient E
r
/E
i
(3)
by its complex conjugate. If absorption losses are negligi-
ble, we get
R
FP
=
R
1
2(R
1
R
2
)
1=2
cos (2j r
1
r
2
) R
2
1 2(R
1
R
2
)
1=2
cos(2j r
1
r
2
) R
1
R
2
(9)
As our discussion has assumed for simplicity that any
coating or gap absorption may be neglected (A
FP
=0), it is
left as an exercise to show that Eqs. (8) and (9) must add to
unity
R
FP
T
FP
=1 (10)
since energy must be conserved (A
FP
R
FP
T
FP
=1).
Let us return to the interferometer transmission
and assume that the reectivities R
1
and R
2
are
frequency-independent. In that case, inspection of
relation (8) shows that a maximum of transmission
occurs whenever the sine term becomes zero. This hap-
pens for
F
2
=
2j r
1
r
2
2
=qp (11)
where q =0,1,2,y is an integer referred to as the longi-
tudinal (or axial) mode number.
Equation (11) corresponds exactly to the familiar
standing-wave condition of multiple interference lters
for which the total transmitted eld is maximum when-
ever the round-trip phase shift F is an integer multiple
of 2p [to account for the incidence angle and the dielectric
constant of the substrate, D should be replaced by
D(n
2
sin
2
y)
1=2
in the definition of j].
T
FP
=
E
t
E
i
_ _
E
t
E
i
_ _
+
=
T
1
T
2
exp(2aD)
1 2(R
1
R
2
)
1=2
cos (2j r
1
r
2
) exp(2aD) R
1
R
2
exp(4aD)
(5)
1382 FABRYPEROT RESONATORS
From (11) we deduce the expression of the resonant
frequency f
res,q
of the qth interference mode
f
res;q
=
c
0
2D
q
r
1
r
2
2p
_ _
(12)
For full metal reectors: r
1
=r
2
=p and f
res;q
=
f
res;q;metal
=(q 1)(c
0
=2D): As is well known, this corre-
sponds to a gap spacing equal to an integer number of
half-wavelengths at resonance [D=(q 1)(l
0
=2)]: At
microwave frequencies, frequency-selective surfaces
(FSSs) [4] are often employed as semitransparent mirrors
(Section 2). Therefore, compared to the ideal case, two
main differences appear:
*
r
1
op and r
2
op =f
res;q
of
res;q;metal
.
*
The frequency response of the mirrors is dispersive
and f
res,q
has to be computed numerically by solving
the following equation
f
res;q
=
c
0
2D
q
r
1
(f
res;q
) r
2
(f
res;q
)
2p
_ _
(13)
This relation assumes that the frequency variations of
r
1
and r
2
are known and the mutual coupling between
mirrors is negligible (this hypothesis is false for thin FP
resonators with capacitive mirrors; see Section 2.1.5).
At resonance, according to Eqs. (8) and (11), the max-
imum value T
FP,max
of T
FP
is given by
T
FP;max
=
(1 R
1
)(1 R
2
)
(1 (R
1
R
2
)
1=2
)
2
(14)
whereas a minimum appears in the transmission curve if
the sine term in (8) is unity
T
FP;min
=
(1 R
1
)(1 R
2
)
(1 (R
1
R
2
)
1=2
)
2
(15)
The qth antiresonance frequency f
antires;q
is deduced from
F
2
=
2j r
1
r
2
2
=(2q 1)
p
2
(16)
and is equal to
f
antires;q
=
c
0
2D
q
1
2

r
1
r
2
2p
_ _
(17)
The ratio of maximum to minimum transmission at nor-
mal incidence is called the fringe contrast
T
FP;max
T
FP;min
=
(1 (R
1
R
2
)
1=2
)
2
(1 (R
1
R
2
)
1=2
)
2
(18)
When mirrors are identical and lossless {r
i
=r, t
i
=t, R
i
=R,
T
i
=T=1-R for i =1,2}, the maximum and minimum lev-
els of transmission become
T
FP;max
=1; \R (19)
T
FP;min
=
(1 R)
2
(1 R)
2
(20)
Equation (19) implies that all the light impinging on the
interferometer is transmitted and none is reected; the
FPI is matched and is analogous to a transmission line
terminated by its characteristic impedance. Moreover, as
the power transmittance of M
2
is (1 R), the intensity of
the incident eld on M
2
(inside the cavity) must be
[1=(1 R)] E
i
[ [
2
. For instance, if R=99%, the incident in-
tensity on M
2
must be 100 times larger than the intensity
used to excite the cavity. Although the running-wave in-
tensities on the inside of the resonator can be much larger
than those on the outside, we have not violated conserva-
tion of energy. Rather, the intensity enhancement is a
manifestation of the energy storage capabilities of a cavity
with highly reecting mirrors. As demonstrated in Section
3.2, the high directivity of FP-based antennas and their
small bandwidth originate from this startling fact. Finally,
we can account for the effect of the dielectric substrate in
Eqs. (12), (13), and (17) if we substitute nD for D.
The variations of T
FP
are represented in Fig. 2 as a
function of frequency (or single-pass phase change j) for
symmetric lossless fused-quartz cavities (e
r
=3.80,
D=3 mm). These curves conrm that the fringe sharp-
ness and contrast increase with R, while maintaining the
transmission peak at unity. For asymmetric resonators
(i.e., R
1
aR
2
), T
FP,max
o 1 (even for an ideal FPI). This is-
sue and the inuence of dielectric and conductor losses
will be discussed in Section 2.2.2.
According to Eq. (12), the frequency interval Df
res
be-
tween two successive resonances (known as the free spec-
tral range) is given by (21) or (22) (for grid or full metal
0 10 20 30 40 50 60 70 80 90 100
Frequency (GHz)
50
40
30
20
10
0
2
(rad)
6
5
3
4
1
3
2
0
T
F
P

(
d
B
)

4
Figure 2. Inuence of the reectivity R of the mirrors on the
transmittance T
FP
[Eq. (8)] of a lossless FPI (D=3 mm, n=1.95,
r
1
=r
2
=p, a =0). R varies between 20 and 99%. Curves (1)
R=20%; (2) R=50%; (3) R=80%; (4) R=90%; (5) R=96%;
(6) R=99%. (This gure is available in full color at http://
www.mrw.interscience.wiley.com/erfme.)
FABRYPEROT RESONATORS 1383
mirrors, respectively)
Df
res
=
c
0
2D
r
1
r
2
2p
_ _
(21)
Df
res;metal
=
c
0
2D
(22)
The half-power bandwidth Df
1=2
quanties the sharp-
ness of the transmission peaks (Df
1=2
is referred to as the
resolution at optical frequencies). Dening f

q
as the high-
er frequency at which the transmission has dropped to
half its maximum value around the qth resonance, we
have
Df
1=2
=2(f

q
f
res;q
) (23)
f

q
is deduced from Eqs. (8) and (14)
sin
2j
q
r
1
r
2
2
_ _
=
1 (R
1
R
2
)
1=2
2(R
1
R
2
)
1=4
(24)
with j
q
=(2pD=c
0
)f

q
. Substituting for f

q
from Eq. (23)
and assuming that the angular variable (pD=c
0
)Df
1=2
is very small (which is true for narrow resonances), we
obtain
Df
1=2
=
c
0
pD
1 (R
1
R
2
)
1=2
2(R
1
R
2
)
1=4
(25)
This relation shows that the 3 dB bandwidth of a
FPI whose mirrors are nondispersive depends only
on the mirror reectivities. Inspection of Eqs. (12) and
(25) suggests that it seems possible to design a FPI with a
prescribed resonant frequency and resolution. This will be
discussed in Section 2.3 for the case of dispersive mirrors.
The frequency selectivity of the resonator is usually
characterized by the Q factor of the qth longitudinal mode
Q
q
=
f
res;q
Df
1=2
=
(R
1
R
2
)
1=4
1 (R
1
R
2
)
1=2
qp
r
1
r
2
2
_ _
(26)
Q
q;metal
=
f
res;q;metal
Df
1=2
=
pD
l
0
=2
(R
1
R
2
)
1=4
1 (R
1
R
2
)
1=2
(27)
Figure 3 represents the variations of Q for the rst three
resonances. At microwave frequencies, typical values of Q
vary between 10 and 500 for antenna applications and can
easily exceed 10
5
for highly resonant structures such as
open resonators (Section 3.3). Unfortunately, at optical
wavelengths, the Q values are astronomical, primarily be-
cause of the smallness of l
0
(if D=1 m, n=1, R
1
=R
2
=
99%, and l
0
=0.6mm, then Q
0
=3.310
8
). To avoid such
large numbers and yet provide a measure of the frequency-
ltering properties of the cavity, we may introduce
a gure of merit known as the reectivity nesse F
R
and
dened as
F
R
=
Df
res
Df
1=2
=
r
1
r
2
2p
p(R
1
R
2
)
1=4
1 (R
1
R
2
)
1=2
(28)
For the numbers used above, this yields a more reasonable
value (F
R
=312). The nesse is a dimensionless quantity
depending only on the mirror reectivities and gives the
width of the resonance as a fraction of the spacing between
the resonances for an ideal monochromatic source. As it is
independent of the free spectral range, this is a measure of
the resolving power of the FPI.
2. GEOMETRY, METHODS OF ANALYSIS, PERFORMANCE,
AND DESIGN OF MICROWAVE FABRYPEROT CAVITIES
In the previous section, the basic properties of optical FP
resonators have been established using a ray model. In
particular, it has been shown that the reection and trans-
mission coefcients of the mirrors play a crucial role in the
frequency response of a FPI. For this reason the rst part
of this section deals exclusively with the analysis and the
performance of periodic metal grid mirrors that are often
used at microwave frequencies. Although we do not vol-
untarily delve deep into the details, this introduction ap-
pears essential for a coherent presentation of the
characteristics and design rules of plane-parallel FP cav-
ities (Sections 2.2 and 2.3). Finally, we investigate briefly
the performance of multilayer semitransparent structures
(Section 2.4) and compare them to stacked dielectric slabs
(Section 2.5). Applications involving all these devices are
discussed in Section 3, where we also summarize the prop-
erties of convex resonators.
2.1. Geometry, Modeling, and Characteristics of Metallic
Reecting Mirrors
For most applications, it is often desirable to obtain a high
value of Q in an order as low as possible (1R51) and low
20 30 40 50 60 70 80 90 100
Reflectivity R, %
1
10
100
1000
Q

f
a
c
t
o
r
90
Figure 3. Variation of the Q factor of the qth longitudinal mode
of a lossless FPI (D=3mm, n=1.95, r
1
=r
2
=p, a =0) as a func-
tion of R. Q values are computed from Eq. (8). They are also com-
pared to Eq. (26) for q =1. Key: , q =0 [Eq. (8)]; , q =1
[Eq. (8)]; , q =2 [Eq. (8)]; , q=1 [Eq. (26)]. (This gure is
available in full color at http://www.mrw.interscience.wiley.com/
erfme.)
absorption losses (A51R). These requirements cannot be
fullled simultaneously by thin homogeneous metal layers
in the far-infrared and at microwave and (sub) millimeter-
wave frequencies. In that frequency range, two kinds of
mirrors are mainly used: (1) partially reective metal
grids [or frequency-selective surfaces (FSSs)] and (2) mul-
tilayer dielectric mirrors (or Bragg mirrors). The latter are
briefly described in Section 2.5.
FSSs have an extremely broad range of applications
(e.g., laser technologies, quasioptical components dip-
lexers, interference lters, polarizers, etc., FPIs, dichro c
reectors and subreectors, high-gain antennas), and a
very abundant literature is available on that subject
1
[4].
For the sake of coherence and concision, and for the
purpose of this discussion, only the basic principles of FSS
operation are reviewed. After a preliminary introduction
about the modeling of innite periodic structures (Section
2.1.1), we will start with devices that can be treated
as simple transmission-line elements operating in the
long-wavelength regime and can be analyzed using the
transmission-line matrix techniques (Section 2.1.2). In
particular, this includes one-dimensional (1D) and two-di-
mensional (2D) inductive and capacitive metal grids. Al-
ternative global electromagnetic approaches are also
discussed in Section 2.1.3. The frequency response
of such mirrors is nally presented in Sections 2.1.4
and 2.1.5.
2.1.1. Preliminaries. Let us consider a 2D planar peri-
odic innite array illuminated by a linearly polarized ho-
mogeneous plane wave of incident wavevector
~
kk
i
(Fig. 4).
The spatial periods are labeled d
x
and d
y
in x and y direc-
tions, respectively. By analogy with Fourier decomposition
of periodic signals, it can be demonstrated that the electric
eld
~
EE
s
scattered by the array is an innite sum of Floquet
space harmonics [5]
~
EE
s
(~rr) =
o
n=o
o
m=o
~
EE
s;nm
e
j
~
kk
nm
~rr
(29)
The wavevector
~
kk
nm
of each harmonic can be decomposed
in transverse (to z direction) and longitudinal components
~
kk
nm
=
~
kk
t;nm
k
z;nm
~ee
z
With
~
kk
t;nm
=k
x;n
~ee
x
k
y;nm
~ee
y
k
x;n
=k
i
x
2pn
d
x
= k sin(y) cos(j)
2pn
d
x
k
y;nm
=k
i
y
2pm
d
y
= k sin(y) sin(j)
2pm
d
y
And with
k
z;nm
=
k
2
k
2
x;n
k
2
y;nm
_
; if k
2
> k
2
x;n
k
2
y;nm
=k
2
t;nm
or
k
z;nm
= j
k
2
k
2
x;n
k
2
y;nm
_
; if k
2
ok
2
x;n
k
2
y;nm
respectively.
The effective direction of propagation (
~
kk
nm
) of each
mode depends on (1) the order (n,m) of the mode under
consideration, (2) the geometry of the unit periodic cell,
and (3) the frequency, polarization, and directions (y, j) of
the incident plane wave.
If k
2
t;nm
> k
2
; then k
z;nm
is a pure imaginary number; the
scattered eld is evanescent along (Oz) and is propagating
along the plane of the periodic array (y
nm
=901). It inter-
acts with the metallic patterns and contributes to the in-
ductive or capacitive behavior of the screen. On the
contrary, if k
z;nm
is real, the wave is propagating outside
the structure, and grating lobes may appear, depending on
the operating frequency and angles of incidence.
According to Eq. (29), each component of
~
EE
s
can be
written as
X(~rr) =e
j
~
kk
i
~rr
T
o
n=o
o
m=o
X
nm
e
j(n
~
kk
1
m
~
kk
2
) ~rr
T
e
jk
z;nm
z
(30)
where ~rr
T
=x~ee
x
y~ee
y
, ~rr =x~ee
x
y~ee
y
z~ee
z
,
~
kk
1
= (2p=A) ~ee
z
~
dd
y
,
~
kk
2
=(2p=A) ~ee
z
~
dd
x
; A=d
x
d
y
is the area of the unit
periodic cell. The factor
e
j
~
kk
i
~rr
T
=e
j k sin(y) cos(j)x k sin(y) sin(j)y [ ]
(31)
is referred to as the Floquet phase factor.
To conclude: (1) the eld distributions from one periodic
cell to another are identical, except to a phase factor; and
(2) the computation of the scattering matrix of the innite
array requires to model only one single unit cell.
2.1.2. Transmission-Line Matrix Approach. The trans-
mission-line (TL) matrix method is a very convenient
k
i
y
z
Incident plane wave
d
x
d
y
0
x
Figure 4. Incident plane wave illuminating a 2D planar periodic
array.
1
For further information, refer also to the articles entitled FRE-
QUENCY-SELEECTIVE SURFACES and PERIODIC STRUCTURES published in
that Encyclopedia.
one for the analysis or design of multilayer systems in-
cluding layers of different dielectric permittivities and
planar geometries. It has been described in details else-
where [e.g., 6]. We present only the fundamental concepts:
(1) ABCD matrix, (2) modeling of dielectric slabs, and (3)
modeling of periodic metallic mirrors.
2.1.2.1. ABCD matrix. We assume that only a single
mode is propagating. The various elements of the system
affect only the complex amplitude of the reected and
transmitted waves. Let us consider a linear two-port de-
vice. The input and output currents and voltages, dened
in Fig. 5a, are related by the following matrix equation
V
1
I
1
_ _
=
A B
C D
_ _
V
2
I
2
_ _
(32)
The elements of the matrix are called the ABCD param-
eters, and the matrix itself is referred to as the transmis-
sion-line matrix or ABCD matrix.
When considering N cascading elements (Fig. 5b), the
convention for the currents and voltages is such that the
output current and voltage for a given element are equal
to the input current and voltage for the following element.
As a result, the ABCD matrix of the whole system is
simply the product of the individual matrices
V
in;1
I
in;1
_ _
=
A
1
B
1
C
1
D
1
_ _
A
2
B
2
C
2
D
2
_ _
A
3
B
3
C
3
D
3
_ _
V
out;3
I
out;3
_ _
(33)
The scattering matrix [S] of the two-port junction is easily
deduced from the ABCD coefcients [6]
S
11
=
AB=Z
c
CZ
c
D
AB=Z
c
CZ
c
D
(34)
S
12
=
2(ADBC)
AB=Z
c
CZ
c
D
(35)
S
21
=
2
AB=Z
c
CZ
c
D
(36)
S
22
=
AB=Z
c
CZ
c
D
AB=Z
c
CZ
c
D
(37)
where Z
c
is the normalization characteristic impedance.
Note that the ABCD matrix has the following properties:
(1) for reciprocal elements, ADBC=1 (S
12
=S
21
);
(2) A=D for symmetric devices.
2.1.2.2. Modeling of Dielectric Slabs. The TL matrix of a
dielectric slab of thickness D and refractive index n=e
1=2
r
(e
r
=e
/
je
//
) is
cosh gD Z sinh gD
1
Z
sinh gD cosh gD
_
_
_
_
(38)
where g is the propagation constant in the dielectric ma-
terial; the wave impedance Z differs for the parallel (||)
and perpendicular (>) polarizations
g =j
2pn
l
0
_ _
cos y (39)
Z
[[
=
Z
0
(e
/
sin
2
y
i
)
1=2
e
/
(40)
Z
J
=
Z
0
(e
/
sin
2
y
i
)
1=2
(41)
Here, y
i
and y are the angles of incidence of the plane wave
illuminating the slab and the propagation angle relative to
normal incidence in this material, respectively. Both an-
gles are related by Snells law. For low-loss materials,
n - (e
/
)
1=2
[1 j(tan d=2)], and for lossless materials, Eq.
(38) becomes
cos bD jZ sin bD
j
Z
sin bD cos bD
_ _
(42)
b =(2p=l
0
)(e
/
sin
2
y
i
)
1=2
is the imaginary part of g.
Other TL matrices are given in the literature [e.g. 6,7].
2.1.2.3. Modeling of Periodic Metallic Mirrors. The ex-
tremely extensive literature on metallic periodic struc-
tures includes many elaborate analytical treatments as
well as numerical methods (Section 2.1.3). We restrict
ourselves to the simplest cases, referring the reader to
specific publications for detailed information [4,5,810].
Figure 6a represents typical geometries of 1D and 2D me-
tallic grids (of thickness t and conductivity s) commonly
used in microwave applications involving FP cavities. 1D
grids are also called (strip) gratings; and 2D grids,
meshes. These grids are located at the interface between
two semiinnite dielectric materials (e
r,i
, tand
i
)
i
=
1,2
and
are assumed illuminated by a normally incident (y
i
=0)
linearly polarized plane wave E
i
(Fig. 6b). As suggested in
the preliminaries, they are typically classied into two
categories (inductive or capacitive) depending on their
geometry and the orientation of E
i
. This terminology is
A B
C D
I
1
I
2
V
2
V
1
(a)
V
in,1
V
out,3
I
in,1
I
out,3
Element #1 Element #2 Element #3
(b)
Figure 5. (a) Definition of the ABCD parameters of a linear two-
port device modeled by the transmission-line matrix approach;
(b) cascading transmission-line elements (N=3).
justied later on. Their spatial period is labeled a; d de-
notes the width of the metallic patterns.
At long wavelengths regime (a5l
0
), only the funda-
mental mode (n,m) =(0,0) is propagating. The other modes
are evanescent (Section 2.1.1) and decay exponentially
away from the grid. Consequently such structures reect
and transmit in zero order only in specular directions (no
grating lobes appear in the visible region [4,5]). They can
be modeled by an impedance Z shunting a transmission
line (Fig. 6b) whose ABCD matrix is given by Eq. (43)
1 0
Y 1
_ _
(43)
Z=
1
Y
=R
s
jX (44)
The real part R
s
accounts for conductor losses inside -
nite conductivity grids. Ulrich [11] found a good agree-
ment with experience if (R
s
=Z
0
) =(a=d)[pf (e
0
=s)]
1=2
, where
f is the operating frequency and Z
0
=(m
0
/e
0
)
1/2
is the wave
impedance in free space.
The reactance X depends on l
0
and grid parameters
(a,d). Surveying literature on that subject is well beyond
the scope of this article and would require much more
space; a well-documented description of equivalent cir-
cuits is given elsewhere [e.g., 12, Chap. 9]. Their range of
validity is also discussed in Refs. 13 and 14. Here we
present only general principles for the purpose of discus-
sion.
Historically, Marcuvitz [8] was probably one of the rst
authors to propose explicit expressions of Z for inductive
(Fig. 6a, grid 1) and capacitive (Fig. 6a, grid 3) strip grat-
ings illuminated by a TM/TE incident plane wave. The
reactance of a freestanding (e
r,1
=e
r,2
=1) zero-thickness
inductive grid (X=X
i
, Z=Z
i
) has the following form [8]
X
i
Z
fs
=o ln
1
sin
pd
2a
_ _ (45)
where o=(a=l
0
) is the normalized frequency and Z
fs
is the
characteristic transmission-line impedance. Z
fs
is given by
Eqs. (40) and (41) for TM and TE polarizations, respec-
tively. At normal incidence (Z
fs
=Z
0
), a more accurate
expression is available [8]
X
i
Z
0
=o ln
1
sin(b)
_ _ _
Q
2
cos
4
(b)
1 Q
2
sin
4
(b)
1
16
o
2
(1 3 sin
2
(b))
2
cos
4
(b)
_
(46)
with b=(pd=2a) and Q
2
=(1 o
2
)
1=2
1. Equation (46)
is deduced from an integral formulation in which only the
rst two diffraction modes are taken into account.
According to Eqs. (34), (36), (43), and (44), the reection
and transmission coefcients of zero-thickness grids are
equal to
r =S
11
=
1
12
Z
Z
fs
(47)
t =1r =S
21
=
1
1
Z
fs
2Z
(48)
The power reectivity R=|r|
2
and the phase arg(r) of the
reection coefcient are represented as a function of o in
Figs. 7a and 7b for a strip grating modeled by Eq. (46). As
expected, for a given frequency, R and arg(r) are decreas-
ing functions of the lling factor d/a. Similarly, they de-
crease with o for given values of a and d/a. Keeping in
mind these evolutions is important to understand and in-
terpret the inuence of the grid geometry on the frequency
response of FP cavities (Section 2.2.2).
It is advisable to mention that the equivalent circuits of
grids should be applied cautiously [14]. As an illustration,
Figs. 8a and 8b quantify the range of validity and the ac-
curacy of Marcuvitz model (46) in the nondiffraction re-
gion for a wide range of grid geometries (a=1 mm, 5%o
d/ao75%). For that purpose, two indicators are dened
DR=R R
FDTD
(49)
Darg(r) = arg(r) arg(r)
FDTD
(50)
R and R
FDTD
[resp. arg(r) and arg(r)
FDTD
) are the reec-
tivities (resp. the phases of r) computed with the TL meth-
od and the nite-difference time-domain (FDTD)
technique (Section 2.1.4).
Ulrich [11] later developed equivalent models for free-
standing 2D grids; at long wavelengths, inductive meshes
(Fig. 6a, grid 2) reect most of the incident power. Their
frequency response is thus very similar to that of induc-
tive strip gratings. Nevertheless they present a resonance
around l
0
- a; where they are totally transparent. This
single-pole highpass behavior can be represented by an
oscillatory equivalent circuit (L,C) whose reactance X
i
has
d a
(1)
d a
(2)
d
a
a
d
(3)
E
i
(4)
(
r2
,tan
2
)
Metal grid
t,
(
r1
,tan
1
) II
(b) (a)
E
i
S
21
S
11
Figure 6. (a) Traditional geometries of metallic periodic grids
employed as reecting mirrors of FP cavities. The grid is illumi-
nated by an incident electric eld E
i
. The general terminology
qualifying these grids is the following: (1) 1D inductive; (2) 2D
inductive; (3) 1D capacitive; (4) 2D capacitive. (b) Metallic grid
located at the interface between two semiinnite dielectric media
(e
r,i
, tand
i
)
i =1,2
.
the following expression
X
i
= o
0
ln
1
sin
pd
2a
_ _
_
_
_
_
o
o
0
o
0
o
_ _
1
(51)
o
0
=1 0:27
d
2a
(52)
where o
0
denes the position of the resonance (52) and
was obtained by tting experimental data and calculation
[11]. More accurate expressions for X
i
and o
0
have been
proposed [14] over a very wide frequency band.
The interested reader will nd more details about (1)
other equivalent models and (2) the inuence of (a) the
oblique incidence illumination, (b) the dielectric sub-
strates supporting the mirrors, or (c) the nite conductiv-
ity and thickness of the grids in the literature [810,13
22]. A very interesting discussion about periodic grids is
also provided in the bibliographic notes [12, pp. 307312].
2.1.3. Electromagnetic Modeling. Many modeling ap-
proaches have been developed to predict the scattering
properties of multilayer periodic structures. They could be
applied for the analysis of complex FP cavities coupled to
real feeds. The following references might be consulted for
further information.
Analytical treatments, including the homogenization
technique and the averaged boundary conditions, are
explained in detail in Refs. 9, 10, and 23. Generally
0 0.2 0.4 0.6 0.8 1
0
20
40
60
80
100
P
o
w
e
r

r
e
f
l
e
c
t
i
v
i
t
y

(
%
)
0 0.2 0.4 0.6 0.8 1
100
120
140
160
180
P
h
a
s
e

(
)
(a)
(b)
c
c
Figure 7. Reection properties of freestanding 1Dinductive strip
gratings (a=1mm) illuminated by a normally incident plane
wave. Their geometry is represented in Fig. 6a, grid 1. The grid
is assumed lossless (s =N, R
s
=0) and innitely thin (t5l
0
). Its
equivalent reactance is given by Eq. (46). (a) Power reectivity R;
(b) phase of the reection coefcient arg(r). Key: , d/a=
11.07%; , d/a=20.7%; , d/a=30.7%; , d/a=40.7%;
, d/a=50.7%; , d/a=60.7%; , d/a=75.7%. (This
gure is available in full color at http://www.mrw.interscience.
wiley.com/erfme.)
10
5
4
1
1
2
3
3
2
1
2
2
10
1
4
3
0.8
0.6
0.4
0.2
0.8
0.6
0.4
0.2
20 40 60
0
20 40 60
d/a (%)
d/a (%)
(a)
(b)
c
c
Figure 8. Validity and accuracy of the equivalent circuit [Eq.
(46)] proposed by Marcuvitz [8]. DR (a) and Darg(r) (b) are dened
by Eqs. (49) and (50), respectively. The equiamplitude contours
show that Marcuvitz formulation is accurate (|DR|o3% and
|Darg(r)|o21) in the nondiffraction region (oo1) for d/ao35%,
and for any d/a if oo0:6. (From Ref. 14; copyright Kluwer Aca-
demic/Plenum Publishers, reproduced by permission).
sophisticated numerical methods can also be implement-
ed, such as (1) the equivalent multimode network method
[24,25]; (2) the modal expansion method in which the
diffracted and aperture elds are decomposed in space
harmonics and waveguide modes, respectively [16,19,26];
(3) the spectral domain approach [27]; (4) the generalized
scattering matrix theory [27,28]; (5) the nite-element
method [29,30]; (6) the boundary element method [31];
or (7) the nite-difference techniques applied either in the
time domain (FDTD) [32] or in the frequency domain
(FDFD) [33].
2.1.4. Frequency Response of 2D Inductive Metal Mesh-
es. The complex reection coefcient of freestanding per-
fectly conducting meshes (e
r,i
=1, tand
i
=0, i =1,2; a=
1 mm; s=N, t =0) is represented in Figs. 9a and 9b for a
normally incident plane wave and for various lling fac-
tors d/a [34]. These curves conrm the highpass lter be-
havior of inductive meshes. The reectivity and phase
increase with d/a but decrease with the normalized fre-
quency o. Comparison with the frequency response of 1D
grids (Fig. 7) provides evidence of the resonance of 2D
grids appearing for 0:9ooo1, as suggested by Eq. (52).
2.1.5. Complementary and capacitive grids. Impact on FP
resonators. The grids discussed so far correspond to the
commonly used geometry of semitransparent mirrors of
microwave FP cavities. They are oriented parallel to the
direction of the incident electric eld in order to obtain a
high reectivity (for a5l
0
), and thus high-Q resonators.
According to the Babinet principle, if we interchange
the perfectly conducting and empty parts of innitely thin
(t =0) grids and rotate by 901 the polarization direction of
the incident plane wave, then the reection and transmis-
sion coefcients of the original and complementary
grids, labeled {r,t} and {r
c
,t
c
}, respectively are related as
follows [11]
t t
c
=1
t
c
= r
r
c
= t
(53)
The shunt impedance Z
c
of the complementary grid is de-
duced from Eqs. (47), (48), and (53)
Z
c
Z
fs
=
Z
fs
4Z
i
(54)
As Z
c
is inversely proportional to frequency, these struc-
tures are called capacitive grids. Their reection coef-
cient, computed with the FDTD technique, is represented
in Figs. 10a and 10b for various values of d/a. As expected,
their reectivity naturally increases with frequency (dual
behavior of inductive grids).
Transmission line models of capacitive grids have been
also proposed by several authors [8, pp. 218219, 280284]
[11,15,1820,102]. Their accuracy and range of validity
have been dened recently [103] over the whole visible
region (l
0
4a) by comparison with the FDTD results given
in Figs. 10a and 10b (the methodology implemented is
similar to the one associated with Eqs. (49) and (50)). By
analogy with inductive meshes [14], very accurate and ul-
tra-wideband explicit formulations are also introduced in
[103] for the wave reactances.
To obtain a high-Q cavity, both capacitive grids should
be highly reecting [Eq. (26)]. This implies
*
o > 0:85 and d/aE1 (Fig. 10a).
*
1601oarg(r)o1801 (Fig. 10b).
*
The rst resonant frequency is dened for q =1
[Eq. (12)].
Consequently the gap D is very thin [35], and such cavities
are at least 10 times more compact than resonators with
inductive grids. This peculiar property is interesting for
0 0.2
(a)
(b)
0.4 0.6 0.8 1
0
20
40
60
80
100
P
o
w
e
r

r
e
f
l
e
c
t
i
v
i
t
y

(
%
)
0 0.2 0.4 0.6 0.8 1
100
120
140
160
180
P
h
a
s
e

(
)
c
c
Figure 9. Theoretical frequency response of perfectly conducting
2D inductive meshes computed with the FDTD method: (a) power
reectivity R; (b) phase of the reection coefcient arg(r). The ge-
ometry of the mesh is represented in Fig. 6a, grid 2. Key: ,
d/a=11.07%; , d/a=20.7%; , d/a=30.7%; , d/a=
40.7%; , d/a=50.7%; , d/a=60.7%; , d/a=72.7%.
(From Ref. 34; copyright Kluwer Academic/Plenum Publishers,
reproduced by permission.) (This gure is available in full color at
http://www.mrw.interscience.wiley.com/erfme.)
the design of liquid crystal tunable frequency lters
(Section 3.1).
2.2. Performance of Microwave FP Resonators
2.2.1. Modeling. The modeling techniques discussed in
Sections 2.1.2. and 2.1.3 can be directly applied to study
FP resonators of innite size in extent. As an example,
Figs. 11 and 12 represent the equivalent circuit and the
FDTD computational domain of a FP cavity analyzed with
the TL matrix method and with the FDTD technique [36],
respectively. In Fig. 11, the mirrors are identical and mod-
eled by a shunt impedance Z
m
=1/Y
m
. They are fabricated
on both faces of a dielectric slab (e
r
, tand) of thickness D,
as depicted in Fig. 1. The transmission coefcient S
21
of
the cavity excited under normal incidence is deduced from
Eqs. (33), (36), (38) and (43)
S
21
=
2
2(1Z
0
Y
m
) cosh(gD)
2Z
0
Y
m
n

1
n
n
(Z
0
Y
m
)
2
n
_ _
sinh(gD)
(55)
where n=e
1=2
r
is the refractive index of the substrate.
2.2.2. Frequency Response and Parametric Study. The
FDTD frequency response of a FP resonator is given in
Fig. 13 for the rst ve transmission peaks (q=0,y,4).
Comparison with the transmission coefcient of an ideal
FPI [Eq. (8)] demonstrates that
*
The theoretical FDTD resonant frequencies are sys-
tematically lower than those predicted by Eq. (8); be-
sides, this discrepancy increases with frequency.
These particularities are due to the phase of an
(Z
0
) (Z
0
) (Z
0
/c
r
0,5
) Z
m
Z
m
D
Figure 11. Transmission-line model of a symmetric FP cavity.
The semitransparent mirrors are modeled by an impedance Z
m
shunting a transmission line whose characteristic impedance and
propagation constant depend on the thickness D of the substrate
and its electrical properties (e
r
, tand) [13].
0
(a)
(b)
0.2 0.4 0.6 0.8 1
0
20
40
60
80
100
P
o
w
e
r

r
e
f
l
e
c
t
i
v
i
t
y

(
%
)
0 0.2 0.4 0.6 0.8 1
180
165
150
135
120
105
90
P
h
a
s
e

(
)
c
c
Figure 10. Theoretical frequency response of perfectly conduct-
ing 2D capacitive meshes computed with the FDTD method: (a)
power reectivity R; (b) phase of the reection coefcient arg(r).
The geometry of the mesh is represented in Fig. 6a, grid 4. Key:
, d/a=30.7%; , d/a=40.7%; , d/a=50.7%; ,
d/a=60.7%; , d/a=70.7%; , d/a=80.7%; ,
d/a=90.7%. (This gure is available in full color at http://
www.mrw.interscience.wiley.com/erfme.)
PML
1
PML
2
Unit periodic
volume
k
i
a
D
Mirror M
1

Mirror M
2

x
y
z
d/2
d/2
Figure 12. FDTD computational domain for the modeling of

plane-parallel FP cavities. Periodic boundary conditions (PBC)
are applied in the transverse directions, and perfectly matched
layers (PML) are oriented in the direction of propagation (y axis)
of the incident plane wave.
inductive grid, which is always smaller than 1801 and
decreases with frequency (Fig. 9b).
*
The fringe contrast decreases with the interference
order q, while it remains constant for an ideal FPI.
This comes from the frequency variation of the mirror
reectivity (Fig. 9a).
This example demonstrates unambiguously that the dis-
persive characteristics of the reectors must be taken into
account to accurately predict the scattering parameters of
a FP cavity with grid mirrors.
Let us now consider the inuence of the mesh parame-
ters (a,d) on the frequency response [resonant frequency f
res
,
3 dB bandwidth (BW) and insertion losses (IL)] of fused-
quartz half-wavelength cavities (e
r
=3.80, D - l
0
=2e
1=2
r
).
This material is ideal as a dielectric holder because of its
low loss (tando510
4
at 60 GHz) and good temperature
stability, particularly for the design of open resonators (Sec-
tion 3.3) and oscillators (Section 3.4) with semitransparent
mirrors. The samples studied here (Fig. 14) have been
fabricated using a precise liftoff process (72mm) of sput-
tered-copper lm (t =1.2mm, s
Cu
=5.810
7
S/m) with a
thin chromium adhesion layer (t =0.1mm, s
Cr
=7.5
10
6
S/m). Two cases are considered [37]: symmetric (R
1
=
R
2
=R) and asymmetric (R
1
aR
2
) resonators.
2.2.2.1. Symmetric Resonators. Figures 15a and 15b
show the theoretical and experimental variations of f
res
and BW as a function of (a,d) for two sets of cavities (D=
1.26370.003mm): the mirrors of set 1 have a constant spa-
tial period (a=0.6mm) and a variable lling factor d/a
(17.7%od/ao46.2%); for set 2, d/a=50% and a varies be-
tween 0.5 and 2.1mm with a constant step equal to 0.2mm;
a scale indicating the value of the mirror reectivity R of
each sample is also inserted in Fig. 15a. The transmission
coefcients S
21
have been computed with the FDTD and TL
methods (Section 2.2.1). They were measured in far eld
under normal incidence between two V-band standard horn
antennas. As expected [relations (12) and (25)], f
res
(BW)
increases (resp. decreases) with R. The theoretical resonant
frequencies computed with the FDTD and TL [19] methods
are in very good agreement; this is due to the range of va-
lidity of the equivalent reactance of the mesh (Ref. 14 and
Section 2.1.2, above). Moreover, the discrepancy between
the values of f
res
measured and computed by FDTD is lower
than 600MHz (Fig. 15a). The larger deviation observed for
the bandwidth (Fig. 15b) comes from the diffraction losses
(substrates of nite diameter) that are not taken into ac-
count in the modeling. These losses can be reduced signif-
icantly using focusing planoconvex cavities (Section 3.2,
below and Ref. 36, Table 4). Finally, Fig. 15c shows that
the insertion losses (IL) due to the nite conductivity of
the meshes increase exponentially with R. Comparison
with the IL due only to the dielectric loss (s =N, tand =
10
4
) demonstrates that metal losses contribute, for the
most part, to the IL of millimeter-wave FP cavities. In
particular, this explains the moderate radiation efcien-
cies of FP-based Gaussian beam antennas (Section 3.2.1).
2.2.2.2. Asymmetric Resonators. Relations (12) and (25)
suggest that f
res
and BW of asymmetric devices can be ad-
justed by a convenient choice of the complex reection co-
efcient of each mirror, but at the price of a slight increase
of their transmission loss [R
1
aR
2
, relation (14)]. FDTD
simulations are used to determine accurately the varia-
tions of f
res
, BW, and IL with the width d
2
of the metal
strips of M
2
. The grid parameters of M
1
remain unchanged
(a
1
=1.3 mm, d
1
/a
1
=50.7%). Figure 16a represents the
theoretical transmission coefcients of several cavities
(D=1.259 mm, s
Cu
=5.810
7
S/m, t =1.2 mm) for which
5mm 5mm
Figure 14. Fused-quartz FP resonator operating in the 60-GHz
band. The highly reecting mirrors are 2D sputtered-copper
meshes (e
r
=3.80, D=5.105mm70.003mm, diameter =50mm,
a=1500mm73mm, d/a=50%). (This gure is available in full
color at http://www.mrw.interscience.wiley.com/erfme.)
10 20 30 40 50 60 70 80
Frequency (GHz)
80
60
40
20
0
S
2
1

(
d
B
)
q=0 q=1 q=2 q=3 q=4
Figure 13. Comparison between the frequency responses of an
ideal FPI (Section 1.2) and a FP cavity (e
r
=3.80, tand =10
4
,
D=5.106mm) whose inductive mesh mirrors are identical
(a=1100mm, d/a=50%, s=N, t =0). Key: , Ideal FPI; ,
FP cavity with mesh mirrors (FDTD). (This gure is available in
full color at http://www.mrw.interscience.wiley.com/erfme.)
32.7%od
2
/a
2
o74.7% (a
2
=a
1
=1.3mm, 87.10%oR
2
o99.95%), and Fig. 16b shows the resulting change in
f
res
, BW, and IL as a function of R
2
. As expected, IL is
minimum (IL
min
=1.3dB) for R
2
=R
1
=97.8% (d
2
/a
2
=
d
1
/a
1
=50.7%) and increases very rapidly for R
2
4R
1
. Be-
sides, we notice that f
res
can be tuned between 55.4 and
57.6 GHz if we tolerate an increase in IL of 1dB above
IL
min
. Controlling the complex reection coefcient of a
mirror then constitutes a possible solution for the design
of tunable FP devices using active grids or MEMS (micro-
electromechanical system) technologies (Section 3.1).
2.3. Design of Microwave FP Resonators
In the previous section, we have quantied the inuence
of the mirror properties on the performance of FP cavities.
However, from a design perspective, the inverse problem
should be solved.
For that purpose, approximate or purely numerical syn-
thesis methods can be implemented. In the rst case, the
simplest way to describe the frequency response of the
device is to use the theory of multilayer structures in
which each individual layer is described by its own scat-
tering matrix (Section 2.1.2 above). This approach is typ-
ically applied to design FP cavities operating at a given
frequency; it basically consists in graphically solving
Eq. (13) [38]. From the resonance condition (11), we have
r
1
r
2
2
=k
0
Dqp; with k
0
=
2p
l
0
(56)
When plotted as a function of frequency, the right-hand
side of this equation is a set of straight lines, representing
the ray pathlength with a slope depending on D (and
eventually on y
i
and e
r
). These are referred to as phase
lines. The left-hand side is the average reection phase of
both mirrors. Equation (56) is satised when the phase
lines intersect this curve (Fig. 17a); transmission peaks
appear then in the frequency response of the structure.
These plots are useful guides to systematically adjust the
resonant frequency of cascaded layers, by appropriate ad-
justment of the dielectric spacer parameters and
[r
1
(f ) r
2
(f )]=2.
Nevertheless, this technique does not account for mu-
tual coupling and evanescent waves, and fails when the
mirror spacing is smaller than half the spatial period of
each grid [9] [which is dramatically the case for resonators
with capacitive strip gratings (Section 3.1)]. Moreover,
the choice of grid geometries is necessarily limited because
their frequency characteristics must be known a priori [Eq.
(56)]. These are the reasons why global electromagnetic
Ratio d/a, %
(Reflectivity R, %)
Measure
FDTD
TL
Set#2
Set#1
15
0.25
47
0.75
Period a, mm
1.25 1.75 2.25
o=5.810
7
S/m
Set#2
Set#1
2.25 0.75
Period a, mm
1.25 1.75 0.25
15
I
L

(
d
B
)

23 31 39
Ratio d/a, %
47 23 31 39
Ratio d/a, %
(90.2) (94.6) (97.1) (98.1) (99.2)
0.25 0.75 1.25 1.75 2.25
52
54
56
58
60
15 23 31 39 47
52
54
56
58
60
Measure
FDTD
TL
Set#2
Set#1

f
r
e
s

(
G
H
z
)

B
W

(
G
H
z
)

(99.5) (98.9) (97.7) (95.7) (92.4)
Period a, mm
(Reflectivity R, %)
(a) (b)
(c)
=
52
54
56
58
60
Figure 15. Variations of (a) the resonant fre-
quency f
res
; (b) the half-power bandwidth BW,
and (c) the insertion losses IL, as a function of
the mesh parameters of symmetric resona-
tors. Metal meshes are perfectly conducting
(s =N) or in copper (s =s
Cu
). (From Ref. 37;
copyright John Wiley & Sons, Inc., reproduced
by permission.) (This gure is available in full
color at http://www.mrw.interscience.wiley.
com/erfme.)
analysis tools combined with optimization procedures,
such as simulated annealing or genetic algorithms [39],
may be required for accurate and innovative designs. A
simplied example of FDTD-based methodology is sche-
matized in Fig. 17b, in which a very wide range of grid
geometries is explored [37]. In this work, it has been shown
in particular that an innite number of 2D metal meshes
can be used to design a symmetric resonator with pre-
scribed values of f
res
and BW. However, all the solutions are
not equivalent in terms of insertion losses.
2.4. Multilayer FP Resonators
Multilayer FP cavities are of particular interest for the
design of interference lters (Section 3.1) or frequency
beam scanning antennas (Section 3.2.2). Their frequency
response can be computed using the ABCD formalism
(Section 2.1.2) or full-wave methods (Section 2.1.3). To
minimize the number of degrees of freedom and for the
purpose of discussion, we assume that the device compris-
es N identical 2D meshes separated by air slabs (e
r
=1) of
thickness D (Fig. 18a). Each mesh is characterized by its
complex reection and transmission coefcients {r
1
,t
1
}.
Under normal incidence, it can be easily shown that the
scattering parameters {r
N
,t
N
} of the whole structure are
given by the following recursive relations
t
N
=
t
N1
t
1
exp(jk
0
D)
1 r
N1
r
1
exp(j2k
0
D)
(57)
r
N
=r
N1
t
2
N1
r
1
exp(j2(N1)k
0
D)
1 r
N1
r
1
exp(j2k
0
D)
(58)
In the forbidden band, the minimum level of transmission
is equal to
[t
N
[
min
=
2{(1 [r
1
[)=(1 [r
1
[)]
N=2
1 {(1 [r
1
[)=(1 [r
1
[)]
N=2
(59)
86 88 90 92 94 98 96
7.5
9
3
4.5
I
L

(
d
B
)

6
1.5
0
100
1
1.25
1.5
0.75
0.5
0.25
0

f
r
e
s

(
G
H
z
)

57
58
59
56
55
54
53
BW
IL
f
res

50 52 54 56 58 60 62
30
25
20
15
5
10
Frequency f , GHz
S
2
1
,

d
B
0
d
2
/a
2
=74.7%
d
2
/a
2
=62.7%
d
2
/a
2
=44.7%
d
2
/a
2
=38.7%
d
2
/a
2
=32.7%
d
2
/a
2
=50.7%
(same mirrors)
B
W

(
G
H
z
)

(32.7) (38.7) (44.7) (50.7)
(=R
1
)
(74.7)
Reflectivity R
2
, %
(Ratio d
2
/a
2
, %)
(a)
(b)
Figure 16. (a) Inuence of the reectivity R
2
of mirror M
2
on the
transmission coefcient S
21
of asymmetric resonators (a
1
=
1.3 mm, d
1
/a
1
=50.7%); (b) variations of the resonant frequency
f
res
, the half-power bandwidth BW, and the insertion losses IL, as
a function of R
2
(a
1
=1.3 mm, d
1
/a
1
=50.7%). (From Ref. 37; copy-
right John Wiley & Sons, Inc., reproduced by permission.) (This
gure is available in full color at http://www.mrw.interscience.
wiley.com/erfme.)

FDTD
analysis
Scattering coefficients
Specifications fulfilled ?
Change
geometry
No
Yes
Initial design
Final set
of solutions
k
0
D - 3
k
0
D - 2
k
0
D -
k
0
D
0
0
3/4
/2
/4
0.3 0.6 0.45
(a)
(b)
0.9 0.15 0.75
c
Figure 17. (a) Graphical procedure for the approximate design of
FP cavities whose mirrors are dispersive. The resonant frequen-
cies are given by the intersection points of the phase lines (k
0
D-
qp) with the average reection phase of both mirrors ( ).
These points (K) are in very good agreement with the transmis-
sion peaks of the FDTD frequency response of the cavity ( ) (a
=1mm, d/a=11.4%, D=2.52mm, e
r
=1). (b) FDTD-based syn-
thesis algorithm. (From Ref. 37; copyright John Wiley & Sons,
Inc., reproduced by permission.) (This gure is available in full
Typical transmission coefcients are represented in
Fig. 18b for Nr3. In each transmission band (1) the num-
ber of propagation peaks equals N1 and (2) the intersec-
tion points between |t
1
| and |t
N1
| correspond to the
case |t
N
|=1. Further details can be found in general
textbooks dealing with innite periodic structures and
electromagnetic bandgap (EBG) materials.
2.5. Dielectric FP Resonators
Highly directive FP-based antennas (Section 3.2.1) may be
obtained from resonators comprising N stacked identical
freestanding dielectric slabs (Fig. 18c) behaving as Bragg
mirrors (D
s
=l
0
=4e
1=2
r
, D
a
=l
0
=4). Their transmission co-
efcient is represented in Fig. 18d for alumina slabs
(Nr3). As expected, the forbidden band is all the deeper
as the number of layers and the dielectric contrast (1:e
r
)
increase. High-performance reecting mirrors based on
this concept are very popular at optical wavelengths. Com-
parison between Figs. 18b and 18d reveals the dual be-
havior of multilayer metallic and dielectric mirrors (the
positions of the forbidden bands and passbands are in-
verted). Metal meshes usually lead to lower prole struc-
tures and are compatible with the integration of active
circuits. Nevertheless, as frequency increases, ohmic loss-
es of partially reective mirrors may significantly reduce
the Q factor of the resonator (Section 2.2.2) and justify the
use of Bragg mirrors, especially for high-gain antennas
(Section 3.2.1).
3. APPLICATIONS OF FABRYPEROT CAVITIES AT
MICROWAVE FREQUENCIES
FP cavities are employed in many microwave applications,
such as quasioptical circuits and systems [lters, diplexers
(Section 3.1), oscillators and power combiners (Section
3.4)], directive, multibeam, and beam scanning antennas
(Section 3.2), or dielectric measurement systems (Section
3.3). A series of three review articles [4042] reviews the
concepts of FP resonators and describes some possible ap-
plications from microwave to infrared wavelengths (open
resonators for characterization of gases, or liquid, solid,
anisotropic and magnetic materials; laser output couplers;
plasma diagnostics, etc.).
3.1. Frequency Filters and Diplexers
FP resonators are widely used not only as bandpass fre-
quency lters [11,13,18,19,3638], but also as diplexers
[43,44]. The simple plane-parallel devices described in
Section 2 suffer from two disadvantages:
1. It seems impossible to design lters with both broad
bandwidth and high attenuation in the stopband,
since these parameters are not independent [Eqs.
(15) and (25)]. However alternative ways have been
proposed in the literature, [e.g., 4547]; these lters
are constructed either from more than two grids
(Section 2.4 ), or from capacitive grid arrays (Section
2.1.5) deposited on a dielectric substrate. Note that
any interference lter may be discussed in terms of
only two effective interfaces (Section 2.1.2), which
may themselves consist of many interfaces (Sections
2.4 and 2.5) [48].
2. The second drawback is the transmission level at
higher orders. Here again, the use of capacitive re-
ectors might solve this problem [11] because the
bandwidth of capacitive FP lters decreases with
frequency (Section 2.1.5); higher modes no longer lie
at harmonics of the fundamental mode.
The rapid variation of bandwidth [roughly as f
3
accord-
ing to Eqs. (25), (45), and (47)] limits performance when
scanning over a wide range of frequencies. One solution
might to use a hybrid FP, with one inductive mesh and one
capacitive mesh [11]. This kind of conguration has also
been investigated by Lockyer et al. [49] for the design of
low-prole lters (Fig. 19a) presenting two narrow highly
stable passbands separated by a distinct null (Fig. 19b).
Reecting mirrors may also be built from photonic
crystals [50]. In that case, they consist of multilayer me-
tallodielectric 1D, 2D or 3D periodic structures (see Sec-
tions 2.4 and 2.5 for the 1D and 2D cases). Conceptually,
the gap between the (composite) mirrors can be interpret-
ed as the creation of a defect inside the periodicity of the
electromagnetic crystal. An example of devices is schema-
tized in Fig. 19c, where each mirror is constructed from a
square lattice of lossy dielectric rods [51].
McCleary et al. [52] has reported improved lter per-
formance by slot coupling to higher-order Gaussian beam
modes [53] in concave resonators; these results include
1 3 N

D
s
1 3 N
D
a

1/z
0
(mm
1
)
0 0.1 0.2 0.3 0.4 0.5
0
0.2
0.4
0.6
0.8
1
S
2
1
N=2

N=3
N=1

0 0.1 0.2 0.3 0.4 0.5
0
0.2
0.4
0.6
0.8
1
S
2
1
N=3
N=2
N=1

D D
(a) (b)
(c) (d)
c
2
2
c
a
Figure 18. (a) Generalized FP resonator comprising N identical
freestanding partially reective mirrors; (b) theoretical transmis-
sion coefcients for Nr3 (inductive mesh, [Fig. 6a, grid 2], a=
1 mm, d/a=5.40%, D=1.8 mm, e
r
=1); (c) generalized FP resona-
tor comprising N cascaded identical dielectric slabs (D
s
, e
r
) sepa-
rated by air gaps (D
a
, e
a
); (d) theoretical transmission coefcients
of a Bragg mirror for Nr3. The alumina and air sections have a
quarter-wave electrical thickness at 56.8GHz (D
s
=0.42mm,
D
a
=1.32mm, e
r
=9.87, e
a
=1). (This gure is available in full
mechanically tunable (20%) X-band lters with insertion
losses less than 1 dB and unloaded Q values ranging
from 1000 to 7000. Electronically tunable lters can
also be built from liquid crystal (LC) materials [5456],
Barium strontium titanate (BSTO) ferroelectric lms [57],
or active grids [58]. Note that an efcient low-voltage driv-
ing (o10 V) of microwave LC-FPI requires ultrathin
substrates (D5l/2) [35] to ensure covering of the
whole anisotropic spectrum of nematic crystals [54]. As
suggested in Section 2.1.5, this can be achieved using in-
terconnected capacitive grids serving simultaneously as
reecting mirrors and driving electrodes (Fig. 19d). Tun-
able lters based on MEMS technologies have been re-
ported at optical wavelengths [59], but not yet
at microwave frequencies due to the large-scale techno-
logical challenges.
3.2. FabryPerot-Based Antennas
3.2.1. Highly Directive Antennas
3.2.1.1. Spatial Filtering. Partially reective surfaces
(PRSs) have been studied [6062] for the design of spatial
lters. To demonstrate the focusing ability of PRS, let us
consider an ideal isotropic electromagnetic source S locat-
ed on a perfect electromagnetic screen, inside a FP cavity
of thickness D (Fig. 20a). The reection coefcient of the
output mirror is labeled r =R
1=2
exp(jr). The sum of all
outgoing waves yields the following expression for the far-
eld power pattern [61]
P(y) =
1 R(y)
1R(y) 2(R(y))
1=2
cos r y ( ) p
4pD
l
0
cos(y)
_ _
(60)
The variations of |P(y)| are represented in Fig. 20b as a
function of y and R
0
=R(y =01) at resonance
f
r;q
=
c
0
4D
2q 1
r(0)
p
_ _
(61)
with q_0 (or 1) for capacitive (or inductive) grids. Figure
20b shows that
*
|P(y)| decreases with y for a given value of R
0
.
*
The maximum directivity of the antenna (for y =01)
increases with the broadside reectivity R
0
.
This simple model also enables us to establish an analyt-
ical relation between frequency and angular selectivities.
For highly reecting mirrors (RE1), simple algebra shows
that the half-power beamwidth y
3 dB
is minimum for the
fundamental mode (q =0) and is written at resonance [63]
as follows
y
3dB
- 2 arccos 1
1
2Q
0
_ _
-
2
Q
1=2
0
(62)
As a consequence, highly directive FP-based antennas are
necessarily narrowband. For real applications, the exact
value of y
3 dB
can be obtained from the dispersion diagram
of Bloch modes [64].
3.2.1.2. Radiating Structures and Applications. PRS and
FP resonators are specific electromagnetic bandgap (EBG)
materials [50,65]. With the exception of antennas based on
ultrarefractive metamaterials [66], most of them can be
considered as photonic crystals operating in a defect mode
[67] (e.g., Fig. 19c).
In that context, a very wide variety of geometries of
primary feeds and focusing devices has been proposed. To
excite the structure, we can use waveguide apertures
[61,62,68], horn antennas [60,68], dipole and monopole
antennas [63,69], or printed patches and arrays fed by a
coaxial probe, a microstrip transmission line, or coupling
slots [7073]. As a rst approximation, for PRS mirrors,
the thickness D of the cavity is chosen to maximize the
broadside directivity (Fig. 20b) at the operating frequency.
Nevertheless, we should keep in mind that when the cav-
ity is excited by a real primary source, a slight difference
(o1%) appears between the frequency maximizing the
antenna directivity and the quasioptical resonant
frequency of the FP cavity alone.
Dielectric
spacer
Capacitive
grid
Inductive
grid
d
c
a
c
0 10 20 30 40 50 60
60
50
40
30
20
10
0
S
2
1
,

d
B
Frequency f , GHz
L
} } }
N=3 D=3 N=3
a
a
(a)
(b)
(c) (d)
Figure 19. (a) FP lter constructed from complementary fre-
quency-selective surfaces (CP FSS) [49]; (b) theoretical transmis-
sion coefcient of CP FSS FP lters for two different lengths of
dipoles; (c) top view of a FP resonator operating in the 48-GHz
band. The mirrors, of adjustable reectivity and thickness, are
built from square lattice (a=2.08cm) 2D photonic crystals con-
stituting N layers of lossy dielectric rods of radius 3.175mm
(11oe
r
o13, tandE0.004). They are separated by an airgap of
varying thickness L (from Ref. 51; copyright 2003 IEEE, repro-
duced by permission); (d) mirror geometry of tunable nematic liq-
uid crystal FP lters operating at 27GHz [a
c
=5mm, d
c
/a
c
=
98.75%, D=110mm (El
0
/100)] [54].
The focusing elements can be classied in three main
categories:
*
A single PRS [61,62,74,75] (in that case, we may con-
sider that the source is located in the middle of a FP
cavity dened by the actual PRS and its electrical
image with respect to the ground plane (Fig. 20a)).
*
Two cascaded PRS arranged in a plane-parallel
[70,76,77] or planoconvex [68,71,72] conguration
by analogy with open resonators (Section 3.3), the
planoconvex structure generates a symmetric Gauss-
ian beam and minimizes the sidelobe level.
*
Reecting mirrors constructed from multilayer 1D,
2D or 3D dielectric [63,69,73,78] (e.g., alumina plates
or rods) or metallic (e.g., crossed wires, strip gratings)
[70] EBG materials, as emphasized in Sections 2.4,
2.5 and 3.1.
This antenna technology usually leads to low-prole
high-gain (r32 dB today) radiating structures. Despite
their narrow bandwidth (o0.5% for 30 dB gain), they
might be competitive for (1) applications requiring medi-
um-gain antennas (o25 dB), (2) xed wireless access [e.g.,
local multipoint distribution services (LMDS)] and satel-
lite reception [76], (3) high-data-rate indoor 60-GHz
WLANs [68,77], or (4) interleaved feeds of focal array-fed
reector (FAFR) antennas that simultaneously need an
efcient illumination of the reector and a low rollover
between the beams [79].
Two examples of directive antennas based on the FP
concept are represented in Figs. 20c and 20d. In both
cases, the distance separating the printed source from the
input mirror of the superstrate is near l/2; this opens a
defect mode inside the forbidden band of the periodic
structure [72, Fig. 4; 73, Fig. 10]. The input impedance
of the antenna and its radiation performance (gain, ef-
ciency, co- and cross-polarization levels) have been inves-
tigated as a function of the Q factor of the FP cavity, the
topology of the feed, and the mutual coupling between the
feed and the resonator [71,72,80]. Figure 20e compares
the radiation patterns of the primary source alone and the
(a)
(d) (e) (f)
(b)
D
S
Ground
plane

(
)

R
0
, %
20
15
10
5
0
5
10
15
60
50
40
30
20
10
0
40 50 60 70 80 90
20
30
90 60 30 0 30 60 90
40
30
20
10
0
FP cavity
Dielectric
coupling region
Radiating
elements
Ground plane
Coupling
slots
Feed network
GBA
Feed
Comput.
N
o
r
m
a
l
i
z
e
d

p
o
w
e
r

(
d
B
)
Angle ()
(c)
=0
=/2
Dielectric alumina rods
22.36 mm
33.54 mm
(
g
=9.8)
258 m
m
258 mm
28.36 m
m
[=10.6 mm
Metalic ground
plore 21.2 mm
2
1
.2
m
m
0

0
Figure 20. (a) Multiple reections between a ground plane and a semi-transparent sheet;
(b) variation of the antenna directivity (equimagnitude contours in dB) as a function of the broad-
side reectivity R
0
of the output mirror and elevation angle y (inductive grid, q =1, D=2.5mm);
(c) directive antenna using a Bragg mirror (4.75 GHz), where each reecting surface of the su-
perstrate is built from one layer of alumina rods (from Ref. 73; copyright 2003 IEEE, reproduced by
permission); (d) general architecture of multilayer Gaussian beam antennas (GBAs): a planocon-
vex FP cavity with periodic mirrors is fed by a printed primary radiator through a dielectric cou-
pling region. (from Ref. 72; copyright John Wiley & Sons, Inc., reproduced by permission);
(e) measured and computed radiation patterns of a GBA in H plane at 56.6 GHz [the primary
source is a patch antenna printed on a 150mm-fused-quartz substrate; the mirrors of the quartz (e
r
=3.80) FP resonator are inductive metal meshes (a=600mm, d/a=46%; see grid 2 in Fig. 6a); the
thickness and the radius of curvature of the cavity equal 1.310mm and 2595mm, respectively; (f)
FP resonator with nonuniform mesh operating in V band.
FP-based antenna. It clearly demonstrates the focusing
effect induced by the cavity, thereby conrming the ability
of a FPI to spread the electromagnetic energy over
its whole surface (i.e., to increase significantly the direc-
tivity of an elementary feed). Design rules have been pro-
posed [71,78] for Gaussian beam antennas and EBG
resonator antennas, respectively. Finally, let us mention
the possibility to further reduce the total height of the an-
tenna by a factor two by virtue of employing a planar ar-
ticial magnetic conductor (AMC) [65,104] as a ground
plane [105].
The PRS that we have considered until now were peri-
odic; when the FP approach applies, their geometry is de-
termined according the antenna specications [directivity,
Eqs. (26) and (62); bandwidth, Eq. (25); resonant frequen-
cy, Eq. (13)]. Several authors have proposed nonuniform
PRS (1) to shape the beam radiated by horn antennas
[106] and printed antennas [77] (Fig. 20f), or (2) to in-
crease the radiation purity and bandwidth usually altered
by phase aberrations of the aperture electric eld around
sharp resonance [75].
3.2.2. Frequency Beam Scanning Antennas and Multi-
beam Antennas. Frequency beam scanning antennas
rely simultaneously on the angular (Section 3.2.1) and
frequency (Section 1.2) selectivities of FP cavities. Accord-
ing to Eq. (60), the radiated power is maximum in the
direction y
s
= arccos(f
r;q
=f ) where f is the operating fre-
quency and f
r,q
is given by Eq. (61). Frequency beam scan-
ning is thus possible if f
r;q
. To minimize the variation of the
maximum directivity during scanning, periodic
arrays insensitive to oblique incidence are of particular
interest.
On the other hand, we have seen in Section 2.4
that multiple propagation peaks appear in each passband
if the FP resonator comprises more than two PRSs (e.g.,
Fig. 18b). Therefore several beams can be radiated by the
antenna, depending on the operating frequency, that is,
the cavity mode chosen. These multilayer FP cavities may
be excited by linear [81] or printed [82] antennas.
Adjusting the radiation characteristics by frequency
tuning may prove constraining for many applications. An
alternative solution consists in combining astutely and si-
multaneously two independent modes of reconguration;
the rst one uses agile materials as substrates (e.g., liquid
crystals or ferroelectric thin lms), and the second one
controls the reection/transmission properties of capacitive
grids by integrating varactor diodes or MEMS switches
in the PRS [82].
3.3. Open Resonators and Concave Mirrors
Accurate knowledge of the complex permittivity of dielec-
tric materials
2
is crucial for the design of microwave cir-
cuits and antennas. The most widely used techniques in
the microwave region are cavity resonators; free-space,
open-ended coaxial probes; and wideband transmission
lines based on one- or two-port measurements. Although
narrowband, the open-resonator (OR) technique has es-
tablished itself as a powerful and precise tool for measur-
ing the permittivity of low-loss dielectric materials
[40,41,83,84]. A very extensive literature is available on
that subject since interest in OR was generated in the
1960s by the invention of the laser and the need for res-
onators with large volume and low loss.
In its most general form, an OR comprises two axially
aligned spherical mirrors of radii of curvature R
1
and R
2
and separated by a distance D (Fig. 21a). The sample S is
located in the middle of the structure. Compared to the
plane-parallel FPI, the use of (at least) one concave mirror
has the advantage to focus the eld into a smaller volume
and thus remove diffraction and alignment problems pro-
vided the mirrors are of sufciently large aperture [85,86].
An additional advantage is gained if one (or both) of the
mirrors is (are) spherical; in the paraxial approximation,
the orthogonal modes are the well-known Gaussian beam
modes, as in laser and maser cavities. In that case, no so-
phisticated numerical methods [31,87,88] is required for
an accurate analysis of the OR:
*
Kogelnik and Li [89] provided a very clear scalar
analysis of these modes and determined a criterion
ensuring stability for the beam resonances:
0_(1 D=R
1
)(1 D=R
2
)_1. Three particular cong-
urations, located at the limit between the stable and
unstable regions, appear from the stability diagram:
the symmetric confocal resonator (R
1
=R
2
=D), the
plane-parallel resonator (R
1
=R
2
=N) discussed in
Section 1.2, and the concentric resonator (R
1
=R
2
=
D/2) for which the centers of the two mirrors coincide.
*
Rigorous vector theory has been proposed later by Yu
and Cullen [90].
Detailed treatments about optics of Gaussian beams and
resonant cavities can be found in textbooks [e.g., 2, Chaps. 4
and 5; 53]. We just state that the resonant modes are qua-
si-TEM
plq
(or TEM
mnq
) modes, where q denotes the longi-
tudinal mode number (Section 1.2) and pl (or mn)
describes the transverse variations of the electromagnet-
ic elds in terms of LaguerreGaussian beams in cylin-
drical coordinates (or HermiteGaussian beams in
Cartesian coordinates). The resonant frequency of the
D
S
Input Output
S
Input Output
(a) (b)
Figure 21. (a) FP open resonator for dielectric measurement
the sample of dielectric material S is located in the middle of the
device and the shaded area symbolizes the transverse extent of
the Gaussian beam mode; (b) hemispherical open resonator.
2
For further information, refer also to the article entitled DIELEC-
TRIC MEASUREMENTS in this encyclopedia.
TEM
plq
mode is given by [89]
f
res;plq
=
c
0
2D
q 1
2pl q
p
arccos 1
D
R
1
_ _
1
D
R
2
_ _ _ _
1=2
_ _
(63)
This equation is consistent with relation (12), that pro-
vides the qth resonant frequency of an ideal plane-parallel
FPI with full metal mirrors (r
1
=r
2
=p).
The resonator technique relies on perturbation of the
electric eld distribution by a sample of the material into
the resonant structure. The resulting shift of resonant
frequency and change in Q factor are analyzed to deter-
mine the material properties. The relevant theory is de-
scribed in a number a review articles, (e.g., see Refs. 84
and 90). There are two basic procedures. In the rst, the
resonant frequency is measured with and without the
sample under test, in a FP cavity of xed length [91,92].
In the second, the frequency is xed and the length of the
resonator is adjusted to establish resonance with and
without the sample [83,92]. Accurate measurements are
facilitated by using a high Q for the resonator (up to 2
10
5
in Ref. 93; 6 10
5
in Ref. 94). Experimental errors de-
pend on a number of factors, but the fractional uncertain-
ties obtained in determining the index of refraction have
been on the order of 10
4
. Loss-tangent measurements
are considerably less precise, with fractional uncertainties
in the range 0.010.1. Among the various OR geometries
already employed, congurations with one plane mirror
(Fig. 21b), namely, the hemispherical (DER) and semicon-
focal (DER/2) resonators, are convenient because (1) they
are unconditionally stable, (2) the boundaries of at ma-
terials conform closely to the phase fronts of the eld dis-
tribution in the resonator, and (3) they enable the
characterization of liquid specimen. Small samples can
also be studied with hemispherical structures because the
beam radius is very small at the plane mirror.
Coupling energy in and out of the device must perturb
as little as possible the conguration of the resonant
Gaussian beam modes. A commonly employed method uti-
lizes small apertures in the reectors to couple energy
from waveguides. This approach may significantly in-
crease the insertion losses since small holes radiate ener-
gy into a large solid angle. An alternative solution consists
in coupling to a quasioptical beam using partially reect-
ing surfaces, such as grid mirrors [93], as discussed in
Section 2.
3.4. Oscillators and Quasioptical Power Combiners
There is a strong need for high-power, solid-state sources
at microwave and millimeter wavelengths. At microwave
frequencies, waveguide or cavity combiners have been
successful; but at higher frequencies, problems appear in
terms of loss, combining efciency, power output, and di-
mensional tolerances. Spatial or quasioptical techniques
3
could provide a possible solution since they are able to in-
tegrate many active devices, minimize signal distribution
and combining losses, and maintain desired amplitude
and phase relationships.
An enhanced RFefciency is achieved with quasioptical
power combiners by coupling a planar array of active de-
vices to a propagating Gaussian beam in free space (and
vice versa). Compared to circuit-combining structures,
these systems have low losses because (1) energy is dis-
tributed or combined in free space and ohmic losses are
minimized and (2) the use of beams with large beam waist
size enables to integrate a lot of devices in a single stage of
combining (causing them to operate in parallel). Most of
the losses are associated with coupling efciency from ac-
tive devices to the beam, and with the coupling to a power
collection port.
This concept was suggested in the 1980s [95,96]. The
classic paper published by Mink [96] is generally credited
with providing the impetus for the increased research ac-
tivity that followed. In that work, Mink analyzed theoret-
ically a power combiner based on an array of negative
resistance devices located inside a semiconfocal open res-
onator (Fig. 22).
Practical considerations include (1) the choice of solid-
state components themselves (Gunn diodes [97,98],
HEMT transistors [99,100], operating from X band to V
band), (2) coupling of the individual devices to the cavity
(for instance using grooved mirrors [9799] and (3) cou-
pling to the outside of the cavity via waveguides [9799] or
partially reecting surfaces fabricated on quartz sub-
strates [100]. A more complete summary of this technique
is given by York and Popovic [101].
BIBLIOGRAPHY
1. M. Born and E. Wolf, Principles of Optics (Electromagnetic
Theory of Propagation, Interference and Diffraction of Light),
5th ed., Pergamon Press, New York, 1975.
2. J. Hawkes and I. Latimer, Lasers. Theory and Practice, Pren-
tice-Hall, Englewood Cliffs, NJ, 1995.
3. J. M. Vaughan, The FabryPerot Interferometers, History, The-
ory, Practice and Applications, Adam Hilger, Bristol, UK,
1989.
4. B. A. Munk, Frequency Selective Surfaces. Theory and Design,
Wiley, New York, 2000.
5. R. E. Collin, Field Theory of Guided Waves, McGraw-Hill,
New York, 1960.
Planar mirror
(heatsink)
Curved
mirror
Waveguide
output
Gaussian
beam
Planar array
of active devices
Figure 22. Quasioptical power combiner using a semiconfocal
open resonator.
3
For further information, refer to the article entitled SPATIAL AND
QUASI-OPTICAL POWER COMBINING in this encyclopedia.
6. D. M. Pozar, Microwave Engineering, Wiley, New York, 1997.
7. K. C. Gupta, R. Garg, and R. Chadha, Computer-Aided Design
of Microwave Circuits, Artech House, MA, Dedham, 1982.
8. N. Marcuvitz, Waveguide Handbook, McGraw-Hill, NewYork,
1951.
9. S. Tretyakov, Analytical Modeling in Applied Electromagnet-
ics, Artech House, Dedham, MA, 2003.
10. T. B. A. Senior and J. L. Volakis, Approximate Boundary Con-
ditions in Electromagnetics, IEE Electromagnetic Wave Se-
ries Vol. 41, IEE, London, 1995.
11. R. Ulrich, Far infrared properties of metallic mesh and its
complementary structure, Infrared Phys. 7:3755 (1967).
12. P. F. Goldsmith, Quasi-optical Systems. Gaussian Beam, Qua-
si-optical Propagation and Applications, IEEE Press, New
York, 1998.
13. R. Sauleau, Ph. Coquet, J.-P. Daniel, T. Matsui, and N. Hirose,
Study of FabryPerot cavities with metal mesh mirrors using
equivalent circuit models. Comparison with experimental re-
sults in the 60 GHz band, Int. J. Infrared Millimeter Waves
19(12):16931710 (Dec. 1998).
14. R. Sauleau, Ph. Coquet, and J.-P. Daniel, Validity and accu-
racy of equivalent circuit models of passive inductive meshes.
Definition of a novel model for 2D grids, Int. J. Infrared Mil-
limeter Waves 23(3):475498 (March 2002).
15. W. Wang and Y. Guo, Analysis of metal grating polarizers for
quasi-optical power combining applications, Int. J. Infrared
Millimeter Waves 16(1):123133 (1995).
16. M. S. Durschlag and T. A. De Temple, Far IR optical proper-
ties of freestanding and dielectric backed metal meshes, Appl.
Opt. 20(7):12451253 (April 1981).
17. R. C. Compton and D. B. Rutledge, Approximation techniques
for planar periodic structures, IEEETrans. Microwave Theory
Techn. 33(10):10831088 (Oct. 1985).
18. S. W. Lee, G. Zarrillo, and C. L. Law, Simple formulas for
transmission through periodic metal grids or plates, IEEE
Trans. Anten. Propag. 30(5):904909 (Sept. 1982).
19. C. C. Chen, Transmission of microwave through perforated
at plates of nite thickness, IEEE Trans. Microwave Theory
Tech. 21(1):16 (Jan. 1973).
20. L. B. Whitbourn and R. C. Compton, Equivalent circuit for-
mulas for metal grid reectors at a dielectric boundary, Appl.
Opt. 24(2):217220 (Jan. 1985).
21. R. C. Compton, L. B. Whitbourn, and R. C. McPhedran, Strip
gratings at a dielectric interface and application of Babinets
principle, Appl. Opt. 23(18):32363242 (Sept. 1984).
22. C. K. Lee, R. J. Langley, and C. Eng, Equivalent-circuit mod-
els for frequency selective surfaces at oblique angles of inci-
dence, IEE Proc. Microwaves Anten. Propag. 132(6):395399
(Oct. 1985).
23. R. R. De Lyser, Homogenization Analysis of Electromagnetic
Strip Grating Antennas, Ph.D. thesis, Univ. Colorado, Boul-
der, 1991.
24. M. Guglielmi and A. A. Oliner, Multimode network descrip-
tion of a planar periodic metal-strip grating at a dielectric
interfacePart I: Rigorous network formulations, IEEE
Trans. Anten. Propag. 37(3):534541 (March 1989).
25. M. Guglielmi and A. A. Oliner, Multimode network descrip-
tion of a planar periodic metal-strip grating at a dielectric
interfacePart II: Small-aperture and small-obstacle solu-
tions, IEEE Trans. Anten. Propag. 37(3):542551 (March
1989).
26. R. C. McPhedran and D. Maystre, On the theory and appli-
cation of inductive grids, Appl. Phys. 14:120 (1977).
27. R. Mittra, C. H. Chan, and T. Cwick, Techniques for analyzing
frequency selective surfacesa review, Proc. IEEE
76(12):15931615 (Dec. 1988).
28. R. C. Hall, R. Mittra, and K. M. Mitzner, Analysis of multilay-
ered periodic structures using generalized scattering matrix
theory, IEEE Trans. Anten. Propag. 36(4):511517 (April 1988).
29. D. S. Filipovic, J. L. Volakis, and L. S. Andersen, Efcient
modeling and analysis of innite periodic antenna arrays by
tetrahedral nite elements, IEEE Int. Symp. Anten. Propag.
4:25042507 (1999).
30. D. T. McGrath and V. P. Pyati, Phased array antenna analysis
with hybrid nite element method, IEEE Trans. Anten. Prop-
ag. 42(12):16251630 (Dec. 1994).
31. H. Cam, S. Toutain, Ph. Gelin, and G. Landrac, Study of a
FabryPerot cavity in the microwave frequency range by the
boundary element method, IEEE Trans. Microwave Theory
Tech. 40(2):298304 (Feb. 1992).
32. A. Taflove, Advances in Computational Electrodynamics: The
Finite-Difference Time-Domain Method, Artech House, Ded-
ham, MA, 1998.
33. W. Sun, K. Liu, and C. A. Balanis, Analysis of singly and
doubly periodic absorbers by frequency-domain nite-differ-
ence method, IEEE Trans. Anten. Propag. 44(6):798805
(June 1996).
34. R. Sauleau, D. Thouroude, Ph. Coquet, and J.-P. Daniel, The-
oretical reection coefcient of metal grid reectors at a di-
electric boundary, Int. J. Infrared Millimeter Waves
20(2):325340 (Feb. 1999).
35. Y. N. Kazantsev, V. P. Maltsev, and A. D. Shatrov, Plane wave
transmission through a pair of capacitive gratings, IEE Proc.
Microwaves Anten. Propag. 147(6):455462 (Dec. 2000).
36. R. Sauleau, Ph. Coquet, D. Thouroude, J.-P. Daniel,
H. Yuzawa, N. Hirose, and T. Matsui, FDTD analysis and ex-
periment of FabryPerot cavities at 60 GHz, IEICE Trans.
Electron. E82-C(7):11391147 (July 1999).
37. R. Sauleau, G. Le Ray, and Ph. Coquet, Parametric study and
synthesis of 60-GHz FabryPerot resonators, Microwave Opt.
Technol. Lett. 34(4):247252 (Aug. 2002).
38. A. C. de Lima and E. A. Parker, FabryPerot approach to the
design of double layer FSS, IEE Proc. Microwaves Anten.
Propag. 143(2):157162 (April 1996).
39. Y. Rahmat-Samii and E. Michielsen, Electromagnetic Opti-
mization by Genetic Algorithms, Wiley, New York, 1999.
40. G. W. Chantry, The use of FabryPerot interferometers, etal-
ons and resonators at infrared and longer wavelengthsan
overview, J. Phys. E: Sci. Instrum. 15:38 (Jan. 1982).
41. R. N. Clarke and C. B. Rosenberg, FabryPerot and open res-
onators at microwave and millimeter wave frequencies,
2300GHz, J. Phys. E: Sci. Instrum. 15:924 (Jan. 1982).
42. E. A. M. Baker and B. Walker, FabryPerot interferometers
for use at submillimeter wavelengths, J. Phys. E: Sci.
Instrum. 15:2532 (Jan. 1982).
43. J. A. Arnaud, A. A. M. Saleh, and J. T. Ruscio, Walk-off effects
in FabryPerot diplexers, IEEE Trans. Microwave Theory
Tech. 22(5):486493 (May 1974).
44. H. P. Pickett and A. E. T. Chiou, Folded FabryPerot quasi-
optical ring resonator diplexer: Theory and experiment, IEEE
Trans. Microwave Theory Tech. 31(5):373380 (May 1983).
45. T. Timusk and P. L. Richards, Near millimeter wave bandpass
lters, Appl. Opt. 20(8):13551360 (April 1981).
46. G. D. Holah, Far-infrared and submillimeter-wavelengths l-
ters, in K. J. Button, ed., Infrared and Millimeter Waves, Ac-
ademic Press, San Diego, 1982, Vol. 6, Chap. 6, pp. 345409.
47. M. M. Pradhan, Multigrid interference lters for the far in-
frared region, Infrared Phys. 11:241245 (Dec. 1971).
48. S. D. Smith, Design of multilayer lters by considering two
effective interfaces, J. Opt. Soc. Am. 48(1):43-50, 1958.
49. D. S. Lockyer, J. C. Vardaxoglou, and R. A. Simpkin, Comple-
mentary frequency selective surfaces, IEE Proc. Microwaves
Anten. Propag. 147(6):501507 (Dec. 2000).
50. J. D. Joannopoulos, R. D. Meade, and J. N. Win, Photonic
Crystals: Molding the Flow of Light, Princeton Univ. Press,
Princeton, NJ, 1995.
51. M. M. Beaky, J. B. Burk, H. O. Everitt, M. A. Haider, and S.
Venakides, Two-dimensional photonic crystal FabryPerot
resonators with lossy dielectrics, IEEE Trans. Microwave
Theory Tech. 47(11):20852091 (Nov. 1999).
52. J. McCleary, M. Li, and K. Chang, Slot-fed higher order mode
FabryPerot lters, IEEE Trans. Microwave Theory Tech.
41(10):17031709 (Oct. 1993).
53. A. E. Siegman, Lasers, University Science Books, Chaps. 16,
17, 1986.
54. R. Sauleau, N. Tentillier, B. Splingart, and Ph. Coquet, Reson-
ateurs de PerotFabry accordables en frequence par cristal
liquide, Treizie`mes Journees Nationales Micro-ondes, Lille,
France, Paper 1B6, May 2123, 2003.
55. M. Tanaka and S. Sato, Electrically controlled millimetre-
wave transmission properties of stacked liquid crystal cells
with metal substrates, Jpn. J. Appl. Phys. 40:41314135
(2001).
56. K. C. Lin and W. C. Chuang, Polarization independent and
electronically tunable FabryPerot etalons with cross-orthog-
onal liquid-crystals layers, Microwave Opt. Technol. Lett.
36(6):475477 (Sept. 2003).
57. E. A. Parker and S. B. Savia, Active frequency selective sur-
faces with ferroelectric substrates, IEE Proc. Microwaves An-
ten. Propag. 148(2):103108 (April 2001).
58. L. B. Sjogren, Active beam control arrays, in T. K. Wu, ed.,
Frequency Selective Surfaces and Grid Arrays, Wiley, New
York, 2000, Chap. 7, pp. 211247.
59. J. D. Patterson, Micro-mechanical Voltage Tunable Fabry
Perot Filters Formed in (111) Silicon, NASA Technical Paper
3702, Sept. 1997.
60. A. Ortusi and J.-C. Simon, Hertzian Optics, U.S. Patent
2,763,860 (Sept. 18, 1956).
61. G. Von Trentini, Partially reecting sheet arrays, IRE Trans.
Anten. Propag. 4(10):666671 (Oct. 1956).
62. A. P. Feresidis and J. C. Vardaxoglou, High gain planar
antenna using optimised partially reective surfaces,
IEE Proc. Microwaves Anten. Propag. 148(6):345350 (Dec.
2001).
63. T. Akalin, J. Danglot, O. Vanbesien, and D. Lippens, A highly
directive dipole antenna embedded in a FabryPerot type cav-
ity, IEEE Microwave Wireless Components Lett. 12(2):4850
(Feb. 2002).
64. S. Enoch, G. Tayeb, and D. Maystre, Dispersion diagrams of
Bloch modes applied to the design of directive sources, in J. A.
Kong, ed., Progress in Electromagnetic Research Series, PIER
41, special issue on electromagnetic applications of PBG
materials and structures, 2003.
65. Mini-special issue on electromagnetic crystal structures, de-
sign, synthesiss and applications, IEEE Trans. Microwave
Theory Tech. 47(11): (Nov. 1999).
66. B. Gralak, G. Tayeb, and S. Enoch, Anomalous refractive
properties of photonic crystals, J. Opt. Soc. Am. A 17:1012
1020 (2000).
67. D. R. Smith, S. Schultz, S. L. McCall, and P. M. Platzmann,
Defect studies in a two-dimensional periodic photonic lattice,
J. Modern Opt. 41(2):395404 (1994).
68. R. Sauleau, Ph. Coquet, D. Thouroude, J.-P. Daniel, and
T. Matsui, Radiation characteristics and performance of mil-
limeter wave horn fed Gaussian beam antennas, IEEE Trans.
Anten. Propag. 51(3):378387 (March 2003).
69. E. Ozbay, B. Temelkuran, and M. Bayindir, Microwave appli-
cations of photonic crystals, in J. A. Kong, ed., Progress in
Electromagnetic Research Series, PIER 41, special issue on
electromagnetic applications of PBG materials and struc-
tures, 2003.
70. S. Enoch, N. Guerin, G. Tayeb, and P. Vincent, Study of
photonic crystal based directive antennas, Proc. 24th
ESTEC Antenna Workshop on Innovative Periodic Antennas:
Photonic Bandgap, Fractal and Frequency Selective
Structures, Estec, Noordwijk, The Netherlands, June 2001,
pp. 1116.
71. R. Sauleau, Ph. Coquet, and T. Matsui, Low-prole directive
quasi-planar antennas based on millimeter wave FabryPerot
cavities, IEE Proc. Microwaves Anten. Propag. 150(4):
274278 (Aug. 2003).
72. R. Sauleau, Ph. Coquet, and T. Matsui, Near-eld coupling
between a printed antenna and a FabryPerot resonator: Ex-
perimental study of the radiation properties at millimeter
wave frequencies, Microwave Opt. Technol. Lett. 38(6):438
443 (Sept. 2003).
73. M. The`venot, C. Cheype, A. Reinex, and B. Jecko, Directive
photonic bandgap antennas, IEEE Trans. Microwave Theory
Tech. 47(11):21152122 (Nov. 1999).
74. S. Chandran and C. Vardaxoglou, Performance of two single-
layer frequency selective surface as spatial lters, Microwave
Opt. Technol. Lett. 6(6):339342 (May 1993).
75. F. Croq and H. Legay, Radiating Structure, U.S. Patent
6,061,027 (May 5, 2000).
76. J. Shaker and M. Cuhaci, A double-layer FSS structure to
suppress far side-lobes of an antenna, Microwave Opt. Tech-
nol. Lett. 39(5):428432 (Dec. 2003).
77. R. Sauleau, Ph. Coquet, T. Matsui, and J.-P. Daniel, A new
concept of focusing antennas using plane-parallel Fabry
Perot cavities with non-uniform mirrors, IEEE Trans. Anten.
Propag. 51(11):31713175 (Nov. 2003).
78. C. Cheype, C. Serier, M. The`venot, Th. Monedie`re,
A. Reinex, and B. Jecko, An electromagnetic bandgap reso-
nator antenna, IEEE Trans. Anten. Propag. 50(9):12851290
(Sept. 2002).
79. R. Chantalat, P. Dumon, B. Jecko, M. The`venot, and
Th. Monedie`re, Interlaced feeds design for a multibeam re-
ector antenna using a 1-D dielectric PBG resonator, Proc.
IEEE Int. Symp. Antennas and Propagation, June 2227,
2003, Vol. 4, pp. 867870.
80. R. Sauleau and Ph. Coquet, Input impedance of electromag-
netic bandgap resonator antennas, Microwave Opt. Technol.
Lett. (in press).
81. H. Boutayeb, K. Mahdjoubi, and A.-C. Tarot, Frequency beam
scanning and gain enhancement properties of PBG antennas,
Proc. IEEE Int. Symp. Antennas and Propagation, URSI
Digest, June, 2227, 2003, p. 526.
82. R. Sauleau, Antenne a` Cavite Resonante Recongurable, Fr.
Patent 309,415 (July, 30, 2003).
83. M. N. Afsar and H. Ding, A novel open-resonator system for
precise measurement of permittivity and loss-tangent, IEEE
Trans. Instrum. Meas. 50(2):402405 (April 2001).
84. A. L. Cullen, Millimeter-wave open-resonator techniques,
in K. J. Button, ed., Infrared and Millimeter-wave,
Academic Press, San Diego, 1983, Vol. 10, Chap. 4, pp. 233
281.
85. A. G. Fox and T. Li, Resonant modes in a maser interferom-
eter, Bell Syst. Tech. J. 40:453488 (March 1961).
86. G. D. Boyd and J. P. Gordon, Confocal multimode resonator
for millimeter through optical wavelength masers, Bell Syst.
Tech. J. 40:489508 (March 1961).
87. M. Rewienski and M. Mrozowski, Iterative application of
boundary conditions in the parallel implementation of the
FDFD method, IEEE Microwave Guided Wave Lett.
10(9):362364 (Sept. 2000).
88. T. E. Harrington, J. Wosik, and S. A. Long, Open resonators
mode patterns for characterization of anisotropic dielectric
substrates for HTS thin lms, IEEE Trans. Appl. Supercon-
duct. 7(6):18611864 (June 1997).
89. H. Kogelnik and T. Li, Laser beams and resonators, Proc.
IEEE 54(10):13121329 (1966).
90. P. K. Yu and A. L. Cullen, Measurement of permittivity by
means of an open resonator, Proc. Roy. Soc. Lond. A 380:4971
(1982).
91. B. Komiyama, M. Kiyokawa, and T. Matsui, Open resonator
for precision dielectric measurements in the 100GHz band,
IEEE Trans. Microwave Theory Tech. 39(10):17921796 (Oct.
1991).
92. M. N. Afsar, X. Li, and H. Chi, An automated 60 GHz open-
resonator system for precision dielectric measurement, IEEE
Trans. Microwave Theory Tech. 38(12):18451853 (Dec. 1990).
93. T. Matsui, K. Araki, and M. Kiyokawa, Gaussian beam open
resonator with highly reective circular coupling regions,
IEEE Trans. Microwave Theory Tech. 41(10):17101714
(Oct. 1993).
94. A. F. Krupnov, V. N. Markov, G. Y. Golubyatnikov, I. I. Leonov,
Y. N. Konoplev, and V. V. Parshin, Ultra-low absorption mea-
surement in dielectrics in millimeter- and submillimeter-
wave range, IEEE Trans. Microwave Theory Tech.
47(3):284289 (March 1999).
95. L. Wandinger and V. Nalbandian, Millimeter-wave power
combiner using quasioptical techniques, IEEE Trans. Micro-
wave Theory Tech. 31(2):189193 (Feb. 1983).
96. J. W. Mink, Quasi-optical power combining of solid-state mil-
limeter-wave sources, IEEE Trans. Microwave Theory Tech.
34(2):273279 (Feb. 1986).
97. H. Kondo, M. Hieda, M. Nakayama, T. Tanaka, K. Osakabe,
and K. Mizuno, Millimeter and submillimeter wave quasi-op-
tical oscillator with multi-elements, IEEE Trans. Microwave
Theory Tech. 40(5):857862 (May 1992).
98. M. Sanagi and S. Nogi, Power combining by a FabryPerot
resonator with active devices mounted on both the mirrors,
IEICE Trans. Electron. E84-C(10):15751580 (Oct. 2001).
99. J. Bae, S. Sugawara, F. Takei, K. Kikuyama, T. Fujii, and K.
Mizuno, Design of a quasi-optical oscillator using a grooved
mirror with a HEMT array, IEICE Trans. Electron. E81-
C(6):857861 (June 1998).
100. M. Kiyokawa and T. Matsui, A new quasi-optical oscillator
with Gaussian output beam, IEEE Microwave Guided Wave
Lett. 4(5): 129131 (May 1994).
101. R. A. York and Z. B. Popovic, Active and Quasi-optical Ar-
rays for Solid-State Power Combining, Wiley, New York, 1997.
102. R. C. Compton, L. B. Whitbourn and R. C. McPhedran, Sim-
ple formulae for the transmittance of strip gratings, Int. J.
Infrared Millimeter Waves 4(6):901912 (Nov. 1983). See also
erratum, Int. J. Infrared Millimeter Waves 7(11):1823 (Dec.
1986).
103. R. Sauleau and N. Falola, Ultra-wideband wave reactance of
capacitive grids, submitted to Microwave Opt. Technol. Lett.
(April 2004).
104. Mini-special issue on metamaterials, IEEE trans. Anten.
Propag. 51(10):(Oct. 2003).
105. S. Wang, A. P. Feresidis, G. Goussetis and J. C. Vardaxoglou,
Low-prole resonant cavity antenna with articial magnetic
conductor ground plane, Elec. Letters 40(7):405406 (April
2004).
106. R. Sauleau, Ph. Coquet, D. Thouroude and J.-P. Daniel,
Beam focusing using 60-GHz Fabry-Perot resonators with
uniform and non-uniform metal grids, Elec. Letters 39(4):341
342 (Feb. 2003).
FAST FOURIER TRANSFORMS AND NUFFT
QING HUO LIU
Duke University
Durham, North Carolina
1. INTRODUCTION
The Fourier transform has been an essential analytical
tool for many elds. The discretized version of the Fourier
transform, the discrete Fourier transform (DFT), on the
other hand, is an important numerical tool for ubiquitous
applications in every branch of science and engineering.
The fast Fourier transform (FFT) algorithm, as an efcient
way of calculating DFT, has become an indispensable part
of modern numerical computation since the pioneering
work of Cooley and Tukey [18], although some related
earlier FFT concepts can be dated back to Gauss in 1805
[32] and Danielson and Lanczos in 1942 [21].
The FFT algorithms have found widespread applica-
tions in engineering and sciences. In this article we will
discuss only its application in numerical solution of Max-
wells equations. Within this important area, the FFT has
been widely applied in solving both time-domain and fre-
quency-domain partial-differential equations and integral
equations.
Over the decades since the invention of the FFT algo-
rithms, there has been increasing demand for nonuniform
discrete Fourier transform (NUDFT), where data are sam-
pled at nonuniform points in either the spectral or phys-
ical domain, or both. Although the need for such
nonuniform discrete Fourier transform has been recog-
nized for a long time, fast transform algorithms were de-
veloped as recently as the mid-1990s. This article includes
the latest developments in these nonuniform fast Fourier
transform (NUFFT) algorithms [5,24,52,65].
This article is organized as follows. In Section 2, we will
introduce the Fourier transformand its properties. In Section
3, the discrete Fourier transform (DFT) will be given in de-
tail, followed by the fast Fourier transform (FFT) algorithms
in Section 4. Applications of the FFTalgorithms will be brief-
ly summarized in Section 5. Finally, the nonuniform fast
Fourier transform (NUFFT) will be discussed in Section 6.
FAST FOURIER TRANSFORMS AND NUFFT 1401
2. FOURIER TRANSFORMS
2.1. One-Dimensional Fourier Transform
Given a complex function s(x) in terms of a real variable x,
its Fourier transform is dened by
S(k) =
1
2p
_
_
o
o
s(x)e
jkx
dx :[s(x)] (1)
where x may be a time or space coordinate, k is the cor-
responding angular frequency or wavenumber (spatial
frequency), and S(k) is in general a complex function.
The inverse Fourier transform can be dened as
1
2
[s(x 0) s(x 0)] =
1
2p
_
_
o
o
S(x)e
jkx
dk (2)
The inverse Fourier transform involves the average of
left and right limits of function s(x) because the function
may be discontinuous. In the special case where the func-
tion is continuous, the left-hand side of (2) simply reduces
to s(x):
s(x) =
1
2p
_
_
o
o
S(x)e
jkx
dk (3)
In what follows, unless otherwise stated, we will assume
that function s(x) is continuous; the transform of discon-
tinuous functions will be discussed in Section 6.
Thus, Eqs. (1) and (2) form a Fourier transform pair.
Note in this article that a factor of 1=
2p
_
is used for
both the forward and inverse Fourier transforms (and
similarly a factor of 1=
N
_
for the forward and inverse
discrete Fourier transform) to make them symmetric;
other choices of such factors are possible but will not be
described in detail here.
If s(x), s
1
(x), and s
2
(x) have the corresponding Fourier
transforms S(k), S
1
(k), and S
2
(k), respectively, the Fourier
transform has the following basic properties:
*
Linearityfor any complex constants c
1
and c
2
,
the Fourier transform of function s(x) =c
1
s
1
(x)
c
2
s
2
(x) is
S(k) =:[c
1
s
1
(x) c
2
s
2
(x)] =c
1
S
1
(k) c
2
S
2
(k) (4)
where : denotes the Fourier transform.
*
Symmetrythe Fourier transform of S(x) is
s(k) =:[S(x)] (5)
*
Spatial (temporal) scalingfor any nonzero real con-
stant a, we have
1
[a[
S
k
a
_ _
=:[s(ax)] (6)
*
Spectral (frequency) scalingfor any nonzero real
constant a, we have
S(ak) =:
1
[a[
s
x
a
_ _
_ _
(7)
*
Spatial (temporal) shiftingif the spatial function is
shifted by a constant value x
0
, the Fourier transform
of the shifted function is
e
jkx
0
S(k) =:[s(x x
0
)] (8)
*
Spectral (frequency) shiftinga constant shift k
0
in
the spectral function corresponds to the Fourier
transform of the modulated function s(x)e
jk
0
x
:
S(k k
0
) =: s(x)e
jk
0
x
_
(9)
*
Alternative inverse Fourier transformthe inverse
Fourier transform can be written in two ways:
s(x) =:
1
[S(k)] ={:[S
+
(k)]]
+
(10)
*
Even functionfor an even function s
e
(x) =s
e
( x),
the Fourier transform
S
e
(k) =S
e
(k) (11)
is also an even function.
*
Odd functionfor an odd function s
o
(x) = s
o
( x),
its Fourier transform
S
o
(k) = S
o
(k) (12)
is also an odd function.
*
Convolution theoremthe Fourier transform of the
convolution between s
1
(x) and s
2
(x) is the product of
S
1
(k) and S
2
(k),
2p
_
S
1
(k)S
2
(k) =:[s
1
(x) + s
2
(x)]
=:
_
o
o
s
1
(t)s
2
(x t) dt
_ _
(13)
where + denotes the convolution.
*
Correlation theoremthe Fourier transform of the
correlation between s
1
(x) and s
2
(x) is the product of
S
1
( k) and S
2
(k)
2p
_
S
1
(k)S
2
(k) =:[s
1
(x) Qs
2
(x)]
=:
_
o
o
s
1
(t)s
2
(x t) dt
_ _
(14)
where #denotes the correlation.
These Fourier transform properties are important in
applications. Corresponding properties are found in dis-
crete Fourier transform (see Section 3).
1402 FAST FOURIER TRANSFORMS AND NUFFT
2.2. Multidimensional Fourier Transform
Multidimensional Fourier transform is a straightforward
extension of one-dimensional Fourier transform. Given
an M-dimensional function s(r), where r =(x
1
,x
2
,y,x
M
) is
the position vector, its Fourier transform in terms of
the M-dimensional spectral vector k=(k
1
,k
2
,y,k
M
) is
dened as
S(k) =(2p)
M=2
_
o
o
s(r)e
jk
.
r
dr (15)
and its inverse is dened as
s(r) =(2p)
M=2
_
o
o
S(k)e
jk
.
r
dk (16)
where s(r) =[s(r 0) s(r 0)]/2 is understood for discon-
tinuous functions. The properties of the multidimensional
Fourier transform can be summarized easily as given in
Table 1.
3. DISCRETE FOURIER TRANSFORM
Although many functions have closed-form Fourier trans-
forms, for applications in signal processing and numerical
analysis, many more functions have no closed-form Fou-
rier transforms. Thus, numerical computation of Fourier
transform is necessary for most applications. As the com-
puter can deal only with discrete numbers, one has to
convert the continuous Fourier transform into a discrete
Fourier transform. Below we rst dene the discrete Fou-
rier transform pairs, and then make the connection be-
tween the continuous and discrete Fourier transforms.
The development here follows closely the book by Brigham
[9].
3.1. Definition of 1-D DFT
Consider a 1-D periodic, discrete array {h(n)} having a pe-
riod of N, i.e., h(nrN) =h(n) for any integer r. The 1-D
forward discrete Fourier transform of {h
n
} is dened as
H(m) =
1
N
_
N1
n=0
h(n)e
j2pmn=N
; m=0; 1; . . . ; N 1 (17)
Due to the periodic nature of exp( j2pmn/N), it is easy to
show that the DFT array {H(m)} is also periodic with
a period N, namely, H(mrN) =H(m), where r is any in-
teger.
The inverse DFT can be shown to be given by
h(n) =
1
N
_
N1
m=0
H(m)e
j2pmn=N
; n=0; 1; . . . ; N 1 (18)
3.2. Properties of DFT
The properties of the DFT are analogous to those of the
continuous Fourier transform, as summarized in Table 2.
Of particular importance are the convolution and cor-
relation theorems. Cyclic convolution between h
1
(n) and
h
2
(n) is dened as
h(n) =
N1
p=0
h
1
(p)h
2
(n p) (19)
while their correlation is dened as
g(n) =
N1
p=0
h
1
(p)h
2
(np) (20)
It is important to note that all the arrays involved, h
1
(n),
h
2
(n), h(n), and g(n), are periodic arrays with period N.
The cyclic convolution and correlation theorems in Table 2
are based on this important property. The DFT is a simple
way to perform these cyclic convolution and correlation.
Later we will discuss how to use the DFT to perform linear
convolution and correlation when the arrays are non-
periodic.
3.3. Relationship between FT and DFT
Discrete Fourier transform can be considered as a tool en-
tirely independent of continuous Fourier transform. How-
ever, the application of the DFT is primarily in the
approximation of continuous Fourier transform. There-
fore, we must establish the connection between the DFT
and continuous FT. As will be seen below, the question of
this approximation is entirely dependent on the function
Table 1. Properties of Fourier Transform
Property Inverse FT Forward FT
Linearity c
1
s
1
(r) c
2
s
2
(r) c
1
S
1
(k) c
2
S
2
(k)
Symmetry S(r) s( k)
Spatial shifting s(r r
0
) S(k)e
jk
.
r0
Spectral shifting s(r)e
jk0
.
r
S(kk
0
)
Conjugation S
*
(r) s
*
(k)
Even function s
e
(r) S
e
(k)
Odd function s
o
(r) S
o
(k)
Convolution s
1
(r)*s
2
(r) (2p)
M/2
S
1
(k)S
2
(k)
Correlation s
1
(r) Qs
2
(r) (2p)
M/2
S
1
( k)S
2
(k)
Parsevals theorem
_
o
o
[s(r)[
2
dr =
_
o
o
[S(k)[
2
dk
Table 2. Properties of DFT
Property Inverse DFT Forward DFT
Linearity c
1
h
1
(n) c
2
h
2
(n) c
1
H
1
(m) c
2
H
2
(m)
Symmetry H(n) h( m)
Time shifting h(np) H(m)e
j2pmp/N
Frequency shifting h(n)e
j2pnp/N
H(mp)
Conjugation H
+
(n) h
+
(m)
Even array h
e
(n) H
e
(m)
Odd array h
o
(n) H
o
(m)
Convolution h
1
(n)*h
2
(n)
N
_
H
1
(m)H
2
(m)
Correlation h
1
(n) Qh
2
(n)
N
_
H
1
(m)H
2
(N m)
Parsevals theorem

N1
n=0
[h(n)[
2
=
N1
m=0
[H(m)[
2
and its sampling. For a bandlimited function, the DFTcan
be designed to completely capture the information of the
underlying function without any error, while for other
functions the DFT will provide an approximate represen-
tation of the functions.
3.3.1. General Complex Waveforms. Consider a general
complex-valued function s(x) having an innite support;
thus, the function is nonzero even at x-7N. In order to
convert the Fourier transform into a discrete Fourier
transform, the following steps are taken:
1. This function s(x) is rst truncated to the function
~ ss(x) =s(x)[U(x x
min
) U(x x
max
)] (21)
with a nite support xA[x
min
,x
max
], where U(x) is the
Heaviside step function.
2. The truncated function is then made periodic
by repeating this function with a period equal to
T=x
max
x
min
. Mathematically, this is equivalent
to convolving ~ ss(x) with a delta function train:
d
1
(x) =
o
k=o
d(x kT) (22)
The resulting periodic function of this convolution is
given by
s
T
(x) = ~ ss(x) + d
1
(x) =
o
k=o
~ ss(x kT) (23)
3. Now, since s
T
(x) is a periodic function, one can nd
its Fourier series
s
T
(x) =
o
m=o
S
m
e
j2pmx=T
(24)
where the Fourier coefcients {S
m
} are given by
S
m
=
1
T
_
x
max
x
min
s(x)e
jmk
0
x
dx (25)
where k
0
=2p/T, and we have made use of the fact
that s
T
(x) =s(x) within the period.
4. Using the trapezoidal rule (or other integration
rules) to approximate the integral in step 3 in terms
of the function values at some sampling points, we
have
S
m
=
1
N
N1
n=0
s x
0
nT
N
_ _
e
jmk
0
x
0
e
j2pmn=N
(26)
where x
0
=x
min
T/2N is the rst sampling point.
5. Finally, we truncate the Fourier series to N terms by
keeping m= N/2,y,N/2 1.
Therefore, the inverse Fourier series in (24) becomes
s
T
(x) =
N=21
m=N=2
S
m
e
j2pmx=T
(27)
In particular, at the sample points, we have
s x
0
nT
N
_ _
=
N=21
m=N=2
S
m
e
j2pmx
0
=T
e
j2pmn=N
(28)
In general, these steps give rise to approximation
errors; the only exception is the bandlimited function.
The approximation errors associated with these steps
are the spatial truncation error (step 1), aliasing error
(integration error) due to inadequate sampling density
(step 4), and spectral truncation error (step 5). From these
steps, it is obvious that the relationship between the con-
tinuous Fourier transform and DFT is
h(n) -
1
N
_ s x
0
nT
N
_ _
; H(m) - S
m
e
j(2pmx
0
=T)
(29)
These are only approximate because of the errors intro-
duced in the above steps. Note that in the preceding equa-
tions, we intentionally express the forward and inverse
DFT to contain a factor of 1
N
_
to make the forward and
inverse transforms symmetric; it is not a necessary, but a
convenient, choice of the multiplication factor.
3.3.2. Bandlimited Nonperiodic Functions. The band-
width of a function is dened as the frequency band
k
min
rkrk
max
beyond which the spectrum is zero. In the
preceding equations for a general complex function, we
note that the bandwidth of the function is usually innite;
its spectrum is nonzero even when the frequency ap-
proaches 7N. However, there is one special class of func-
tions that are bandlimited; that is, their spectra are zero
beyond a nite frequency. Such bandlimited functions
have some special properties, as given by the Nyquist
sampling theorem.
3.3.2.1. Nyquist Sampling Theorem. If function s(x) is
bandlimited, that is, if
S(k) =0; for [k[ > k
max
(30)
then under the condition that Dxrp/k
max
, the function can
be uniquely determined by the sample points at x =nDx
(n= N,y,N)
s(x) =
o
n=o
s(nDx)sinc p
x
Dx
n
_ _ _ _
(31)
where the sinc function is dened as
sinc(x) =
sinx
x
(32)
However, bandlimited functions have innite support.
In other words, the truncation in Eq. (21) (step 1) will
cause errors.
3.3.3. Bandlimited Periodic Functions. It is only for
bandlimited periodic functions that an exact relationship
between FT and DFT can be established. For a bandlim-
ited function s
T
(x), by definition its spectrum is nonzero
only within |k|rk
max
, where k
max
is a nite real number,
representing the maximum angular frequency where the
spectrum is nonzero. If the period of the function is T, we
can sample the function at an increment
Dx =
T
N
(33)
Since this function is already periodic, the truncation in
Eqs. (21) (step 1) and (23) (step 2) are no longer necessary.
We can directly perform the Fourier series in Eq. (23).
Furthermore, in the spectral truncation step, that is,
the last step in Eq. (28), if
Dx =
T
N
_
p
k
max
(34)
then according to Nyquist sampling theorem there is no
spectral truncation error. Therefore, under this Nyquist
sampling rate (34), the equations in (29) are exact. More-
over, the reconstruction formula for s(x) in Eq. (31) reduc-
es to
s(x) =
N1
n=0
s(nDx)sinc p
x
Dx
n
_ _ _ _
(35)
In Eqs. (34) and (35), we consider that the truncation
interval is exactly equal to the period T. Note that if this
truncation is integer number of periods, the conclusion
given above still applies, namely, that there is no approx-
imation error between the continuous Fourier transform
and DFT, and the original function can be reconstructed
perfectly using discrete Fourier transform. However, if the
truncation interval is not equal to an integer number of
periods, a discontinuity is present at the two end trunca-
tion points because the function is no longer continuous
(i.e., it has nonsmooth function value or derivatives) at the
truncation point when the function is made periodic. Such
discontinuities make Eq. (26) (step 4) inexact because the
trapezoidal rule is approximate for this discontinuous
function. In other words, a result of this truncation is
that the function is no longer bandlimited. This effect is
called leakage.
4. FAST FOURIER TRANSFORM
Now that we have established the relationship between
FT and DFT, we can use DFT to calculate the exact FT for
bandlimited periodic functions, or approximate FT for oth-
er more general functions. Our objective is to calculate
efciently the sum in (17), rewritten here for convenience:
H(m) =
1
N
_
N1
n=0
h(n)e
j2pmn=N
; m=0; 1; . . . ; N 1 (36)
However, we note that the straightforward summation in
(36) requires N
2
complex multiplications and N(N1)
complex additions. Our goal is to evaluate equation (36)
more efciently than the direct summation so that these
numbers of arithmetic operations will be reduced to O(N
log N).
There are many variations of FFT algorithms. In this
section, we will discuss only the CooleyTukey FFT algo-
rithm for arbitrary factors. Then we will discuss the ap-
plications of FFT in fast convolution and correlation, and
in fast methods in computational electromagnetics.
4.1. The CooleyTukey FFT Algorithm without Twiddle
Factors
First, following Brigham [9] we summarize the derivation
of CooleyTukey FFT algorithm for arbitrary factors as
extended by Bergland [3]. We assume the number of
points N can be factored as
N=
I
i =1
r
i
(37)
where r
i
(i =1,y, I) are positive integers.
To present the FFT algorithm, it is more convenient to
express indices m and n in (36) in a variable radix repre-
sentation
m=m
I1
(r
1
r
2
r
I1
) m
I2
(r
1
r
2
r
I2
)
m
1
r
1
m
0
=
I1
k =0
m
k
k
l =0
r
l
(38)
n=n
I1
(r
2
r
3
r
I
) n
I2
(r
3
r
4
r
I
) n
1
r
I
n
0
=
I1
p=0
n
p
I 1
q=Ip1
r
q
(39)
where r
0
=r
I 1
=1 and
m
i
=0; 1; 2; . . . ; r
i 1
1; 0 _ i _ I 1
n
i
=0; 1; 2; . . . ; r
Ii
1; 0 _ i _ I 1
Then (36) can be rewritten as
H(m
I1
; m
I2
; . . . ; m
1
; m
0
)
=
1
N
_
r
I
1
n
0
=0
r
I1
1
n
1
=0

r
1
1
n
I1
=0
h(n
I1
; n
I2
; . . . ; n
0
)W
mn
N
(40)
where W
N
=exp( j2p/N). This may seem to make the
new DFT summation even more complicated than the
original equation (36), but we will see that this factoriza-
tion actually saves some arithmetic operations. We will
carry out this factorization in several stages allowed by
the radix factorization:
Stage 1. Noting that W
I
i =1
r
i
N
=W
N
N
=1, we can rewrite
the rst terms of W
mn
N
=W
m[n
I1
I
i =1
r
i n
0
]
N
as
W
mn
I1
I
i =2
r
i
N
=W
m
0
n
I1
I
i =2
r
i
N
(41)
Therefore
where x
0
=h=
N
_
. We dene the item inside the
square brackets as x
1
:
x
1
(m
0
; n
I2
; . . . ; n
0
)
=
r
1
1
n
I1
=0
x
0
(n
I1
; n
I2
; . . . ; n
0
)W
m
0
n
I1
I
i =2
r
i
N
(43)
Note that this stage for calculating (43) requires Nr
1
operations (one complex multiplication and one com-
plex addition), without considering the symmetries of
the complex exponential. Then, equation (42) can
then be written as
H(m
I1
; m
I2
; . . . ; m
1
; m
0
)
=
r
I
1
n
0
=0
r
I1
1
n
1
=0

r
2
1
n
I2
=0
x
1
(m
0
; n
I2
; . . . ; n
0
)
W
m[n
I2
(
I
i =3
r
i
) n
0
]
N
(44)
Stage 2. Similar to equation (41), we can now show
that
W
mn
I2
I
i =3
r
i
N
=W
(m
1
r
1
m
0
)n
I2
I
i =3
r
i
N
(45)
Hence, by dening
x
2
(m
0
; m
1
; n
I3
; . . . ; n
0
)
=
r
2
1
n
I2
=0
x
1
(m
0
; n
I2
; . . . ; n
0
)W
(m
1
r
1
m
0
)n
I2
I
i =3
r
i
N
(46)
we can rewrite (44) as
H(m
I1
; m
I2
; . . . ; m
1
; m
0
)
=
r
I
1
n
0
=0
r
I1
1
n
1
=0

r
3
1
n
I3
=0
x
2
(m
0
; m
1
; n
I3
; . . . ; n
0
)
W
m[n
m3
(
I
i =4
r
i
) n
0
]
N
(47)
This stage requires Nr
2
arithmetic operations with-
out considering the symmetries of the complex expo-
nential.
Stage p. In the same manner, if we continue reducing
(47), we will obtain a set of recursive equations of the
form
x
p
(m
0
; m
1
; . . . ; m
p1
; n
Ip1
; . . . ; n
0
)
=
r
p
1
n
Ip
=0
x
p1
(m
0
; m
1
; . . . ; m
p2;
n
Ip
; . . . ; n
0
)
.
W
[m
p1
(r
1
r
2
r
p1
) m
1
r
1
m
0
]n
Ip
I
i =p1
r
i
N
;
p=1; 2; . . . ; I
(48)
where we have dened (r
p1
yr
I
) =1 for p4I 1 and
n
1
=0. This stage requires Nr
p
arithmetic opera-
tions without considering the symmetries of the com-
plex exponential.
The Complete FFT Algorithm. The nal results are
given by
H(m
I1
; m
I2
; . . . ; m
1
; m
0
) =x
I
(m
0
; . . . ; m
I1
) (49)
from which the total number of multiplications
is
M=N
I
p=1
r
p
In the special case where the radix representation is
constant, r
1
=r
2
=?=r
I
r, I =log
r
N, we have the
total number of multiplications M=Nr log
r
N, with-
out considering the symmetries of the complex expo-
nential. Thus, the FFT algorithm requires only
O(N log N) operations.
H(m
I1
; m
I2
; ; m
1
; m
0
) =
r
I
1
n
0
=0
r
I1
1
n
1
=0

r
1
1
n
I1
=0
x
0
(n
I1
; n
I2
; ; n
0
):W
m0n
I1
I
i =2
r
i
N
_
_ W
m[n
I2
I
i =3
r
i
_ _
n
0
]
N
_
_
(42)
4.2. The CooleyTukey FFT Algorithm with
Twiddle Factors
As discussed above, in the CooleyTukey FFT algorithm
without twiddle factors, the pth stage calculation in (48)
requiring Nr
p
operations. However, recognizing that one
factor in (48) can be rewritten as
W
m
p1
(r
1
r
2
r
p1
)n
Ip
I
i =p1
r
N
=W
m
p1
n
Ip
N=r
p
N
=W
m
p1
n
Ip
rp
(50)
we can rewrite the pth stage equation (48) as
x
p
(m
0
; m
1
; . . . ; m
p1
; n
Ip1
; ; n
0
)
=
r
p1
n
Ip
=0
x
p1
(m
0
; m
1
; . . . ; m
p2
; n
Ip
; . . . ; n
0
)W
m
p1
n
Ip
rp
_
_
_
_
W
t
N
=[DFT
r
p
{x
p1
]]
.
W
t
N (51)
where
t =[m
p1
(r
1
r
2
r
p2
) m
1
r
1
m
0
]n
Ip
I
i =p1
r
i
It is obvious that the summation inside the brackets in Eq.
(51) is an r
p
-point DFT (denoted by DFT
rp
{x
p1
]); the fac-
tor W
t
N
outside the bracket is called the twiddling factor.
Thus, the pth stage in the FFTalgorithm can be computed
by the r
p
-point DFT followed by the twiddling operation
(i.e., by the referencing operation of W
t
N
).
Because of the symmetries in W
m
p1
n
Ip
r
p
W
d
rp
, it is im-
portant to recognize that this r
p
-point DFT can be com-
puted very efciently, with much fewer operations than r
2
p
.
For example, if r
p
=2 (i.e., a base 2 transform), the W
d
2
factor takes only 71, and there is no need for multiplica-
tions; the total number of complex multiplications is
M
2
=
1
4
[2 log
2
N 4]N1
If r
p
=4 (i.e., a base 4 transform), there is no need for
multiplications in the DFT since the W
d
2
factor takes only
71 and 7j; the total number of complex multiplications is
M
4
=
1
4
[1:5 log
2
N 4]N1
If r
p
=8 (i.e., a base 8 transform), there are only four real
multiplications in the DFT since the W
d
2
factor takes only
71, 7j, 7e
jp/4
; and 7e
jp/4
; the total number of complex
multiplications is
M
8
=
1
4
[1:333log
2
N 4]N1
For more details, the reader is referred to Ref. 9.
5. APPLICATIONS OF FFT IN ELECTROMAGNETICS
The FFT algorithm is one of the most important develop-
ments in numerical analysis. Its applications are ubiqui-
tous. Here we will only summarize its applications in fast
computational methods for electromagnetics, including
the pseudospectral time-domain (PSTD) methods and
the integral equation methods.
5.1. Fast Derivatives and the Pseudospectral Time-Domain
(PSTD) Methods
The nite-difference time-domain (FDTD) method has be-
come a standard technique to solve transient electromag-
netic problems by solving time-domain Maxwell equations
[82]. However, the major difculty with the FDTD method
is the required high sampling density (Z20 cells per wave-
length) because of its low-order accuracy. It is known that
the accuracy of the FDTD method is second-order in a ho-
mogeneous medium, but only rst order in a discontinu-
ous medium.
In order to improve the FDTD method, many research-
ers have proposed various techniques. One class of such
techniques is pseudospectral time-domain methods, which
use trigonometric functions or Chebyshev polynomials
to approximate spatial derivatives [27,36,4951,79
81,86,87]. These PSTD methods have been developed for
unbounded and bounded media, and have been shown to
greatly outperform the FDTD methods, especially for
large-scale problems.
The key to the PSTD methods lies in the fast and ac-
curate spatial derivatives by using either the FFT or the
Chebyshev polynomials. Consider a 1D function u(x, t). In
the regular FDTD method, the spatial (and temporal) de-
rivatives are approximated by the central nite-difference
scheme, which has a second-order accuracy. In contrast, in
the PSTD methods, the spatial derivatives are approxi-
mated by either trigonometric functions or by Chebyshev
polynomials.
5.1.1. The Fourier PSTD Method. In the Fourier PSTD
method, we consider a spatially periodic function u(x, t) in
x (0rxrL), where L is the period. If Dx =L/N is used as
the cell size and x
n
=nDx, n=0,1,y, N1, the approxi-
mation of the spatial derivative can be in general written
as
@u(x; t)
@x
-
x
{u(x; t)] (52)
where the derivative operator is given by
x
f (x) =:
1
x
{jk
x
:
x
[f (x)]] (53)
where :and :
1
denote the forward and inverse Fourier
transforms that are achieved by an FFT algorithm with
the number of operations O(Nlog N). Such an approxima-
tion in (53) has a spectral accuracy; that is, the error
decreases exponentially as N increases, given that u(x, t)
is analytic.
In contrast to the FDTD method, which uses a stag-
gered grid, the Fourier PSTD method uses a centered grid
where all eld components are located at the cell centers.
In addition to its simplicity, this centered grid provides an
important advantage over the FDTD algorithm since the
material properties are not altered by the presence of the
staggered grid. Furthermore, the centered grid is more
advantageous for anisotropic media [51], and removes the
singularity problem for EM waves in cylindrical coordi-
nates [55].
Since the Fourier PSTD method assumes periodicity in
the wave elds, in order to apply this method to unbound-
ed media, the perfectly matched layer (PML) absorbing
boundary condition should be used at the computational
edge to remove the so-called wraparound effect caused by
the periodicity [49].
5.1.1.1. Time Integration. The temporal grid, however,
is staggered since the central differencing is used for time-
stepping. Therefore, the time integration is exactly the
same as in the regular Yee scheme. The dispersion anal-
ysis and stability condition of the FDTD and Fourier
PSTD algorithm has been formulated for a homogeneous,
lossless medium [49,50]. The Courant-Friedrichs-Lewy
(CFL) stability conditions can be written compactly as
cDt
Dx
_
1
a
D
_ (54)
for a problem of dimensionality D, where a =1 for FDTD,
and a =p/2E1.5708 for PSTD [51]. Higher-order time-in-
tegration methods such as the fourth-order RungeKutta
method can be used to further improve the time integra-
tion accuracy.
The Fourier PSTD method is easy to implement and
requires only two points per minimum wavelength, but
the medium under study needs to be homogeneous or
smoothly varying to ensure the exponential accuracy.
The Fourier PSTD method, originally proposed in Refs.
49 and 50, has been applied to the simulation of many
engineering problems [31,4144,56,74]. This method pro-
vides spectral accuracy when the medium is very smooth.
However, its major limitation is that it is not applicable to
problems with perfect conductors; theoretically, it has the
spectral accuracy only for smoothly varying inhomoge-
neous media, although numerical examples show that it is
much more efcient and accurate than the FDTD method.
One way to deal with strongly inhomogeneous media is to
use a nonuniform grid to account for the material varia-
tion, by using the nonuniform fast Fourier transform (NU-
FFT) [56] or the nonuniform grid mapping [43]; a hybrid of
PSTD and FDTD techniques has also been suggested
[42,44]. However, all these remedies still cannot remove
the staircasing error associated with the Kronecker prod-
uct grid (i.e., the mesh is parallel to coordinate directions).
An alternative method for perfect conductors and discon-
tinuous inhomogeneous media is the multidomain Cheby-
shev PSTD method discussed below.
5.1.2. The Multidomain Chebyshev PSTD Method. The
Fourier PSTD method discussed above is applied for both
periodic problems (such as the frequency-selective surfac-
es and photonic crystals), and for unbounded media with
the PML absorbing boundary condition. For bounded me-
dia, on the other hand, one can develop a single-domain
Chebyshev PSTD method [74]. However, the single-
domain Chebyshev PSTD method, similar to the Fourier
PSTD method, has spectral accuracy only for smoothly
varying media.
In order to apply the PSTD method to perfect conduc-
tors and discontinuous media with spectral accuracy, it is
necessary to develop a multidomain PSTD method
[36,79,80]. Examples of multidomain Chebyshev PSTD
method have been shown for perfect conductors and for
discontinuous inhomogeneous lossy media, including the
applications in GPR (ground penetrating radar) sensing
[27], waveguides [86], and scattering problems [87]. It has
demonstrated the great exibility of dealing with complex
objects, with a slightly increased computational burden
because it requires p cells per minimum wavelength.
In order to develop a multidomain Chebyshev PSTD
method, the problem domain is divided into hexahedron
subdomains naturally conformal to the geometry. In gen-
eral, each subdomain has curved boundaries in the (x, y, z)
coordinates, but can be mapped into a cube in the curvi-
linear coordinates (x, Z, z) with x, Z, z A [ 1,1]. In each
direction x, Z, z, the spatial derivative can be found by us-
ing Chebyshev polynomials. For example, in the x direc-
tion, for xA[ 1,1], the region can be discretized by N1
GaussChebyshevLobatto points (or simply called the
Chebyshev points) dened at
x
i
= cos
ip
N
; i =0; 1; . . . ; N (55)
Then function u(x, t) can be represented as
u(x; t) =
N
n=0
a
n
(t)T
n
(x) (56)
where T
n
(x) =cos[n cos
1
(x)] is the nth-order Chebyshev
polynomial. Equivalently
u(x; t) =
N
i =0
u(x
i
; t)f
i
(x) (57)
where
f
i
(x) =
(1 x
2
)T
/
N
(x)(1)
i 1N
c
i
N
2
(x x
i
)
(58)
and c
i
=1d
i,0
d
i,N
. With the interpolation in (56), the
spatial derivative of u(x, t) with respective to x is approx-
imated as
@u(x
i
; t)
@x
=
N
k =0
D
(x)
ik
u(x
k
; t) (59)
where the differential matrix elements are given by
D
ik
=f
/
k
(x
i
) (60)
and the explicit expressions can be found in Refs. 10 and
87.
Alternatively, using (56), one can obtain the derivative
of u(x, t) as
@u(x; t)
@x
=
N
n=0
b
n
(t)T
n
(x) (61)
where the coefcients {b
n
} can be derived through the re-
cursion relations of the Chebyshev polynomials
b
N
=0
b
N1
=2Na
N
b
N2
=2(N 1)a
N1
b
i1
=b
i 1
2ia
i
; i =N 2; N 3; . . . ; 2
b
0
=a
1
b
2
=2
_
_
(62)
Obviously, the choice of the gridpoints in (55) makes (56) a
cosine transform; thus, the coefcients {a
n
} in (56) can be
found from {u(x
i
)} by using the fast cosine transform (FCT)
algorithm, which can be derived from the FFT algorithm.
The derivative coefcients {b
n
} are then obtained by the
recursion relation (62); then the derivative can be ob-
tained by the fast cosine transform using equation (61). In
general, if N416, it is more efcient to obtain the deriv-
ative by using the CFT than the direct matrixvector mul-
tiplication [10].
Once the spatial derivatives have been found accurately,
the elds inside each subdomain can be updated by a time
integration method such as the fourth-order RungeKutta
method. Then, the elds between adjacent subdomains
must be reconciled by a patching procedure so that the
boundary conditions between subdomains are satised. For
details of the multidomain PSTD method, the reader is re-
ferred to Refs. 27,36,79, and 8688 and a review paper [59].
5.2. Fast Convolution, Correlation, and k-Space Methods
for Integral Equations
In addition to the PSTD methods discussed above, the
FFT algorithm is an essential ingredient in modern com-
putational methods for electromagnetics, in the areas of
fast forward and inverse problems involving integral
equations.
The use of the FFT to accelerate the solution of integral
equations dates back to the k-space method proposed by
Bojarski in the early 1980s [6]. In general, to solve an in-
tegral equation in electromagnetics with the conventional
method of moments (MoM) [35], one uses some basis func-
tions to express the unknown eld (or unknown current
density); after tested by some weighting functions, the in-
tegral equation can be converted into N algebraic equa-
tions with N discrete unknowns. Solving this problem
involves the inversion of the NN matrix, which can be
a prohibitive task as the memory requirement is O(N
2
),
and the CPU time is O(N
3
) (if a direct method is used for
the matrix inversion) or O(KN
2
) (if an iterative method is
used, where K denotes the number of iterations).
The original k-space method [6] was proposed for ob-
jects in a homogeneous background medium. In this
case, the k-space method utilizes the fact that the scat-
tered eld from an object can be written as a convolution
integral between the background Green function and the
induced current in the scatterer. Therefore, after disc-
retization with a uniform grid, the operation of the
Greens function onto the induced current (called Greens
operation) can be carried out by using the FFT with only
O(N log N), rather than O(N
2
) arithmetic operations in
the MoM. The integral equation has to be solved itera-
tively; within each iteration Greens operation and its ad-
joint operation are accelerated by the FFT, thus resulting
in an algorithm requiring only O(KN log N) CPU time,
where K is the number of iterations required. Further-
more, the memory requirement has been greatly reduced
from O(N
2
) in MoM to O(N) in the k-space method as the
MoM matrix is no longer needed. The idea of the k-space
method was later improved by the CG-FFT (conjugate-
gradient fast Fourier transform) methods. Early works in-
clude Refs. 8,12,39,68,69 and 75; 3D CG-FFT-type meth-
ods include Refs. 37,73,77,83,84, and 90; see Ref. 13 for
more complete references. In the CG-FFT method, the
conjugate-gradient method is used as the iterative method
to improve the convergence of the original k-space method.
The k-space and the CG-FFT methods can be applied to a
wide variety of integral equations, including volume inte-
gral equations and surface integral equations with planar
geometries. They are not applicable to surface equations
with arbitrary geometries where the shift invariance of
Greens function no longer holds along the scatterer sur-
face.
Traditionally, the k-space and CG-FFT methods have
been applied to objects in a homogeneous background. The
CG-FFT method has been further improved by the bicon-
jugate-gradient (BCG-FFT) method [37,83] and stabilized
biconjugate-gradient (BCGS-FFT) method [77]. Since all
these iterative methods solve the integral equation using
the FFT, we categorize them as the general k-space meth-
od for integral equations.
The k-space method has also been applied to planar
conductor structures in a multilayered medium where the
scatterer is parallel to the layer interfaces. In this case,
the shift-invariance of the Green function again ensures
that the FFT can be used to accelerate the convolution in-
tegral along the parallel planar directions [13,46,66,89].
More recently, the k-space method has been applied to
acoustic scattering in a layered medium [76], electromag-
netic scattering in a half-space [19,20], and a multilayered
medium [63,64,78], by penetrable inhomogeneous objects.
The major difference between the multilayered back-
ground medium and the homogeneous background medi-
um is that the Green function is no longer shift-invariant
in the direction perpendicular to the layer interfaces; thus,
the Green operation is no longer a convolution integral.
However, it has been shown that if the objects are located
within one single layer in the multilayered medium,
one can split the Greens operation into a convolution in-
tegral and a correlation integral, making the FFT algo-
rithm still applicable to accelerate the Greens operation
[19,20,63,64,76,78].
To illustrate these important applications of the FFT in
computational electromagnetics, we summarize the latest
development in the solution of electromagnetic scattering
by penetrable objects embedded in a multilayered medium
[63,64,78]. As shown in Fig. 1, if an inhomogeneous di-
electric object of arbitrary shape is completely embedded
in the ith layer, the scattered magnetic vector potential in
the mth layer, then A
mi
(r), due to the induced current
source J inside the inhomogeneous object embedded in the
ith layer, is given by
A
mi
(r) =m
m
_
V
G
mi
(r; r
/
)
.
J(r
/
)dr
/
=jom
m
_
V
G
mi
(r; r
/
)
.
w(r
/
)D(r
/
)dr
/
(63)
where G
mi
is the dyadic Green function for the magnetic
vector potential, V is the volume of the scattering object,
w(r) =[~ ee(r) ~ ee
i
]=~ ee(r) is the contrast function, ~ ee(r) =e(r)
s(r)=jo, D(r) =~ ee(r)E(r) is the unknown electric ux densi-
ty inside the object, and ~ ee
i
is the complex permittivity of
the ith layer. We can form the integral equation for the
unknown electric ux density inside the object as
'[D
i
(r)] =E
inc
i
(r); r c V (64)
where ' is a linear operator dened as
'[ ] =
[ ]
~ ee
(k
2
i
VV
.
)
1
~ ee
i
_
V
G
ii
(r; r
/
)
.
w[ ]dr
/
; r c V (65)
and k
i
=o
~ ee
i
m
i
_
is the wavenumber of the ith layer. Solv-
ing Eq. (64), the electric ux density D
i
(r) is obtained for
rAV. The relationship J=jowD can be used to calculate
the induced current inside the scattering object, and thus
the electric eld at any location.
The volume EFIE (electric eld integral equation) has
been formulated with the electric ux density as the un-
known because the continuity of the normal component of
electric ux density can be easily enforced by choosing the
rooftop basis functions for D. On the other hand, if the
electric/magnetic elds are chosen as the unknown, im-
plementation of the boundary conditions is not trivial.
Furthermore, the weak form discretization process pro-
posed in Ref. 90 for the scattering of objects in a homoge-
neous background can be extended to layered media to
reduce the singularity of the problem. To make use of the
FFT to accelerate the solution, the integral equation in
(64) is usually solved iteratively by Krylov subspace meth-
ods such as the conjugate-gradient, biconjugate-gradient,
or stabilized biconjugate-gradient methods, as summa-
rized in Table 3.
When solving the integral equation in (64) iteratively,
we need to calculate the interactions of the dyadic Green
function G
ii
(r; r
/
) with the induced current density J=
jowD for every testing function within the object. If the ob-
ject is discretized into N unknowns, the straightforward
computation of these interactions requires O(N
2
) arithme-
tic operations. In free space, this computation can be accel-
erated by using the fast Fourier transform (FFT) algorithm
because the Green function there is shift-invariant.
As a result of the inhomogeneity along the z direction,
however, the dyadic Green function for a layered medium
is no longer shift-invariant; thus the integral kernel of the
volume EFIE (64) is not convolutional. For example, in the
spectral domain, we obtain
~
GG
ii
xx
(k
r
; z; z
/
)
=
1
2jk
z;i
_
e
jk
z;i
[zz
/
[
1
D
TE
i
{R
TE
i
e
jk
z;i
[(z z
/
)2z
i
]
~
RR
TE
i1
e
jk
z;i
[2z
i1
(z z
/
)]
2R
TE
i
~
RR
TE
i1
e
j2c
i
cos k
z;i
(z z
/
)]
_
(66)
1
z
1
z
i

i
z
i
i-1
i
1

x
z

M
1
z
M-1
Layer i
Layer 1
M M

y
Layer M
r r
Figure 1. Typical conguration of an inhomogeneous object in a
planarly layered medium.
Table 3. Three Krylov Subspace Methods
CG Method BCG Method BCGSTAB Method
r
0
=E
inc
'D
0
p
0
=r
0
r
0
=p
0
=E
inc
'D
0
~ rr
0
= ~ pp
0
=r
+
0
r
0
=E
inc
'D
0
Choose ^ rr
0
so that ^ rr
0
; r
0
)O0 r
0
=a =o
0
=1, v
0
=p
0
=0
a
k
=
r
k1
; r
k1
)
p
k
; p
k
)
a
k
=
r
k
; ~ rr
k
)
'p
k
; ~ pp
k
)
r
k
=^ rr
0
; r
k1
)
b
k
=
r
k
; ~ rr
k
)
r
k1
; ~ rr
k1
)
b
k
=
r
k 1
; ~ rr
k 1
)
r
k
; ~ rr
k
)
b=ar
k
/(o
k1
r
k1
)
D
k1
=D
k
a
k
p
k
D
k 1
=D
k
a
k
p
k
p
k
=r
k 1
b(p
k 1
o
k1
v
k 1
)
r
k 1
=r
k
a
k
'p
k
r
k1
=r
k
a
k
'p
k
v
k
='p
k
, a =r
k
=^ rr
0
; v
k
)
P
k1
=r
k1
b
k
p
k
~ rr
k1
= ~ rr
k
a
+
k
'
a
~ pp
k
s =r
k 1
av
k
, t ='s
p
k 1
=r
k 1
b
k
p
k
o
k
=t; s)=t; t)
~ pp
k 1
= ~ rr
k 1
bk
+
~ pp
k
D
k
=D
k1
ap
k
o
k
s
r
k
=s o
k
t
where
~
R R
TE
i
and
~
RR
TE
i1
dene some global reection coef-
cients, c
i
=k
z,i
(z
i 1
z
i
), and D
TE
i
=1
~
RR
TE
i
~
RR
TE
i1
e
j2c
i
[6062]. Obviously, this Green function is not shift-invari-
ant in the z direction, even though it is still shift-invariant
in the x and y directions. Consequently, unlike the scat-
tering problem in free space, the integral in (65) is no
longer convolutional; thus, the convolutional theorem can-
not be used directly to arrive at the FFT acceleration in
the CG-FFT, BCG-FFT, and BCGS-FFT methods as in the
case of a homogeneous background.
Fortunately, we are able to develop a splitting technique
to rewrite the dyadic Green function into two terms, one is
shift-invariant in all directions, the other is shift-invariant
in the x and y directions, but as a function of z z
/
in the z
direction ([19,20,63,64,76,78]). More specifically, as indi-
cated in (66), the dyadic Greens function in a layered me-
dium can be written as the summation of two terms:
G
ii
(r; r
/
) =G
ii
1
([ " rr " rr
/
[; z z
/
) G
ii
2
([ " rr " rr
/
[; z z
/
) (67)
As a result, in Eq. (65), the interaction between G
ii
1
(r; r
/
)
and w(r
/
)D(r
/
) is a convolution in all (x,y,z) directions, and
that between G
ii
2
(r; r
/
) and w(r
/
)D(r
/
) is a convolution in x
and y directions, but a correlation in z direction. In other
words, the integral kernel in the EFIE is convolutional in
the x and y directions and convolutional or correlational in
the z direction after the splitting. Consequently, the FFT
can be applied to these two operations separately through
the convolution and correlation theorems, reducing CPU
time of the Greens operation from O(N
2
) to O(Nlog N).
5.2.1. Linear convolution and correlation. In order to
make use of the FFT algorithm to achieve discrete convo-
lution and correlation for non-periodic, nite functions
{h
1
(n)} of length N
1
and {h
2
(n)} of length N
2
, one needs to
pad zeros to these arrays and assume that the new arrays
are periodic. These new arrays must have a length at least
equal to N=N
1
N
2
1 to avoid the overlap in the re-
sulting circular convolution and correlation. Furthermore,
the convolution and correlation integrals in the volume
integral equations actually involve discrete convolution
and correlation of a non-periodic array h(n) of nite dura-
tion M with a non-periodic function g(n), namely the
Greens function, of innite support but truncated to
length N. In this case, since both functions must be
made periodic of length N, to avoid the end effect one
must choose the total number NZ2M1 to obtain the
correct result for the M points within the scatterer.
5.2.2. Fast Multipole Method. The FFTalgorithm is also
an important component is the fast multipole method
(FMM) [17,70] that has significantly reduced the compu-
tational complexity of the numerical solution of integral
equations, especially surface integral equations. For more
details of this FMM algorithm, the reader is referred to
Ref. 16.
5.3. Fast Inverse Scattering Methods
Electromagnetic waves have widespread applications in
areas such as medical diagnosis, geophysical exploration,
nondestructive evaluation, and other areas of imaging. In
such applications, one needs to solve the inverse problem,
that is, to infer the characteristics of the target from the
observed electromagnetic elds. The inverse scattering
problems can be classied into two types: (1) approximate,
linear inverse problems and (2) nonlinear inverse scatter-
ing methods. The FFT plays an important role in both
types of electromagnetic inverse scattering problems.
Electromagnetic inverse problems are inherently non-
linear because of the multiple scattering within different
parts of the scatterer. However, in the low-frequency re-
gime, one can approximately neglect the multiple scatter-
ing; such an approximation is known as the Born
approximation [7]. Similarly, in the high-frequency re-
gime, one can use Rytov approximation or Kirchhoff ap-
proximation to make the inverse problem linear. Such
linear inverse methods include the low-frequency diffrac-
tion tomography [22], high-frequency synthetic aperture
radar (SAR) processing [72], and the related migration
processing. For example, in a SAR processing scheme, the
principal idea is to synthesize the effect of a large-aperture
radar whose physical construction is infeasible. In three-
dimensional SAR, the antennas move along a 2D trajec-
tory in the 3D space. Denote the location of the antenna as
(0, u, v), where (u, v) are the azimuthal and elevation co-
ordinates of the radar, and the wideband signal emitted by
the antenna as p(t). We may assume for simplicity an
omnidirectional radar radiation pattern. Then, the SAR
signal [72] can be expressed as
s(t; u; v)
=
n
f
n
p t 2
x
2
n
(y
n
u)
2
(z
n
v)
2
_
=c
_ _
(68)
where c is the velocity of light (or sound) and f
n
is the
normalized reectivity of the nth point scatterer. Taking
the Fourier transform of (68) in (t, u, v) yields the recon-
struction equation for the unknown reectivity
f (x; y; z) =
:
1
k
x
;k
y
;k
z
S(o; k
u
; k
v
)
P(o)
(69)
where :
1
denotes the inverse Fourier transform and
[k
x
, k
y
, k
z
] are governed by the SAR (SAS) spatial frequen-
cy mapping
k
x
(o; k
u
; k
v
) =
4o
2
=c
2
k
2
u
k
2
v
_
;
k
y
(o; k
u
; k
v
) =k
u
; k
z
(o; k
u
; k
v
) =k
v
(70)
The image of the object can thus be found by using the
FFT algorithm from (69).
To address the high-contrast problems where the elec-
trical properties of the object are very different from the
background, nonlinear inverse scattering methods in 2D
[11,14,15,38] and 3D [1,2,85] have been developed. For
more recent developments in this area, the reader is
referred to the references cited in Refs. 45 and 85.
6. NONUNIFORM FAST FOURIER TRANSFORM (NUFFT)
As illustrated above, the FFT has been enjoying
widespread applications in numerical analysis and other
areas of applied mathematics since the work of Cooley
and Tukey [18] in the 1960s. However, the requirement for
using FFT algorithms is that the input data must
be equally sampled. In many practical situations, howev-
er, the input data are not equally spaced either by error or
by design, and hence the regular FFT does not apply. To
overcome this limitation Dutt and Rokhlin [24] studied
the problem of FFT for nonuniform (unequally spaced)
data. Later, the fast FFT for nonuniform data has been
termed (nonuniform fast Fourier transform) (NUFFT)
[52,65].
6.1. Nonuniform Discrete Fourier Transforms (NUDFTs)
The nonuniform discrete Fourier transform (NUDFT) is
an extension of the regular (uniform) DFT. This extension
has widespread applications because many measure-
ments, ranging from synthetic aperture radar (SAR),
ground penetrating radar (GPR), to biomedical imaging
using magnetic resonance imaging (MRI) and X-ray com-
puterized tomography (CT), are sampled nonuniformly in
a given space. First, we dene the temporal sampling
point locations {t
n
, n=0,y, N1}A[0, N), and frequency
sampling point locations {o
m
, m=0,y, N1}A[0, N),
where periodicity is implied with a period of N in both do-
mains. There are three basic types of NUDFT as given
below. The complex sequence and its spectrum are dened
as h={h(t
0
),y, h(t
N1
)} and H={H(o
0
),y, H(o
N1
)}, re-
spectively:
1. NUDFT-1. Nonuniform temporal sampling points
{t
n
}A[0, N), but uniform frequency sampling points
o
m
=0,y, N1. This NUDFT is termed NUDFT-1
here and is dened as the sum
H(m) =
1
N
_
N1
n=0
h(t
n
)e
jt
n
m(2p=N)
;
m=0; . . . ; N 1
(71)
2. NUDFT-2. Uniform temporal sampling points
t
n
=0,y, N1, but nonuniform frequency samp-
ling points {o
m
}A[0, N). This NUDFT is termed
NUDFT-2 here and is dened as the sum
G(o
m
) =
1
N
_
N1
n=0
g(n)e
jo
m
n(2p=N)
;
m=0; . . . ; N 1
(72)
3. NUDFT-3. Nonuniform temporal sampling points
t
n
A[0,N), and nonuniform frequency sampling
points o
m
A[0,N). This NUDFT is termed NUDFT-3
here and is dened as the sum
F(o
m
) =
1
N
_
N1
n=0
f (t
n
)e
jo
m
t
n
(2p=N)
;
m=0; . . . ; N 1
(73)
The corresponding fast algorithms for these sums are
designated NUFFT-1, NUFFT-2, and NUFFT-3, respec-
tively.
As shown in Ref. 24, once the solution to NUFFT-1 is
found, NUFFT-2 and NUFFT-3 algorithms follow analo-
gously. Therefore, here we will only discuss the NUFFT-1
algorithm for the fast solution of (71). Furthermore, the
nonuniform inverse transforms corresponding to (18) can
be dened similarly, and the algorithms are analogous to
(71)(73).
Now lets focus on the NUDFT-1 given by (71). The
properties of the NUDFT-1 are largely analogous to those
of the regular DFT, as summarized in Table 4.
6.2. The Forward NUFFT Algorithm
The aim of the NUFFT is to nd a fast way to evaluate the
sums in (71)(73). Note that the direct summation in these
equations requires O(N
2
) operations, much more than the
FFT algorithm that requires only O(N log N) operations.
The NUFFT algorithms developed by various groups
[5,24,29,52,65] can provide approximate solutions to the
NUDFT summations with O(N log N) operations. We will
describe the algorithms reported in Refs. 52, 57, and 65
below.
We now consider the NUDFT-1 dened in (71). Instead
of interpolating the input data {h(t
n
)}, we will interpolate
the exponential function exp( j2ptm/N) for each t =t
n
(n=0,y, N1). In applications, we only need to consider
a discrete sequence
f (m) =s
m
e
j2ptm=N
; m c {0; . . . ; N 1] (74)
where t is real, 0rs
m
r1 for m=0,y, N1 are called the
scaling factors, and are chosen to minimize the error,
and q an even positive integer. Denoting w=e
j(2p/nN)
, we
will now interpolate f(m) using the exponential function at
q 1 points on a unit circle
s
m
w
mnt
=
q=2
k =q=2
x
k
(t)w
m([nt] k)
; m=0; . . . ; N 1 (75)
Table 4. Properties of NUDFT-1
Property Inverse NUDFT-1 Forward NUDFT-1
Linearity c
1
h
1
(t
n
) c
2
h
2
(t
n
) c
1
H
1
(m) c
2
H
2
(m)
Symmetry None None
Timeshifting h(t
n
T
0
) H(m)e
j2pmT0=N
Frequency shifting h(t
n
)e
j2ptnp=N
H(mp)
Time reversal h( t
n
) H( m)
Conjugation h
+
(t
n
) H
+
( m)
Even array h
e
(t
n
) H
e
(m)
Odd array h
o
(t
n
) H
o
(m)
where [a] denotes the integer nearest to a, and x
q/2
(t),y,
x
q/2
(t) are some unknown interpolation coefcients to be
determined.
The unknown interpolation coefcients x(t) =[x
q/2
(t),y,
x
q/2
(t)]
T
can be determined by the least-squares method.
Remarkably, this solution can be written almost analyti-
cally as
x(t) =F
1
y(t) (76)
where for i, k =0,y, q we have
F
ik
(n; N; q) =
N i =k
w
(ik)N=2
w
(ki)N=2
1 w
(ki)
; iOk
_
_
_
(77)
y
k
(t) =
N1
m=0
s
m
w
({nt] q=2k)m
(78)
Observe that while y(t) depends on t, matrix F(n, N, q) is
completely independent of t and is uniquely determined by
n, N, and q. This remarkable property of this interpolation
is important because it reduces the number of operations
required by the NUFFT algorithms. The matrix F(n, N, q)
is called the regular Fourier matrix; it is a (q1) (q 1)
Hermitian matrix, F(n, N, q)
w
=F(n, N, q). These formulas
make sense for any q c ., and therefore F(n, N, q) can be
dened for any n; N; q c ..
In the equations above, the coefcients 0os
m
r1, for
m=0,y, N1, called the scaling factors, are chosen to
minimize the errors. Two scaling factors have been inves-
tigated:
*
Cosine scaling factors s
m
=cos(pm/nN) for m=0,y,
N1.
*
Gaussian scaling factors s
m
=e
b(2pm=nN)
2
for m=0,y,
N1.
For the cosine scaling factors s
m
=cos (pj/nN) for
j =0,y, N1, it turns out that y
k
(t) can be written in a
closed form by
y
k
(t) =j
g =1;1
sin
p
2n
(2k g q 2{nt])
_ _
1 e
j(p=nN)(2{nt] q2kg)
(79)
Unfortunately, for the Gaussian scaling factor, we are un-
able to nd a closed-form expression for y
k
(t).
The NUFFT-1 algorithm consists of following steps:
1. Compute y
k
(t
n
) by (79) for k =0,y,q and n=0,y,
N1. The complexity is O(Nq).
2. Compute P
in
=
q
k=0
[F
1
]
ik
y
k
(t
n
), where the inverse
regular Fourier matrix F
1
is precomputed. The
complexity is O(Nq
2
).
3. Find the inverse of the scaling factor s
1
m
. For the
cosine scaling factors s
1
m
= sec(pm=nN) for m=0,y,
N1. The complexity is O(N).
4. Calculate Fourier coefcients
t
l
=
i;n;[nt
n
] i =l
h(t
n
)
.
P
in
The complexity is O(Nq).
5. Use uniform FFT to evaluate
T
m
=
nN=21
l =nN=2
t
l
.
e
j2pml=nN
The complexity is O(n N log N).
6. With a complexity of O(N), scale the values to arrive
at the approximated nonuniform FFT:
~
HH(m) =T
m
.
s
1
m
The total complexity is O(N q
2
nN log N), noting that
qBlog(1/e) where e is the precision, and n
2
5N (usually
n =2).
6.3. The Inverse NUFFT Algorithm
The inverse NUFFT algorithm is to nd h(t
n
) from known
H(m) in Eq. (71). We rst rewrite the NUDFT-1 in (71) and
NUDFT-2 in (72) as
H(m) =
1
N
_
N1
n=0
h(n)e
jt
n
m2p=N
N
n=0
A
mn
h(t
n
); A
mn
=
1
N
_ e
jt
n
m2p=N
(80)
G(o
m
) =
1
N
_
N1
n=0
g(n)e
jnom2p=N
N1
n=0
B
mn
g(n); B
mn
=
1
N
_ e
jnom2p=N
(81)
These forward NUDFT summations can be evaluated by
NUFFT-1 and NUFFT-2 algorithms with O(nN log N)
arithmetic operations (n is the oversampling rate).
For the regular, uniformly sampled data where both {t
k
}
and {o
j
} are uniform, the inverse FFT (IFFT) can share
the same algorithm as the FFT simply because A
l
=A
w
and B
l
=B
w
in these special cases (the superscript w de-
notes the complex conjugate and transpose of a matrix).
Unfortunately, this is no longer true for matrices A and B
for nonuniformly sampled data. Hence, the inverse NU-
FFT can no longer share the same algorithm as the for-
ward NUFFT. Obviously, the direct inversion of (80) and
(81) to obtain h(n) and g(n) is prohibitively expensive as it
requires O(N
3
) arithmetic operations. In Refs. 24 and 52,
the inverse NUFFT algorithms have been developed.
We rst rewrite (80) using a matrix notation H=Ah.
From elementary matrix identities we observe that
A
l
=A
w
(AA
w
)
1
. Therefore the inverse DFT solution of
(80) is
h=A
f ; f =(AA
)
1
H (82)
The advantage of rewriting the solution in the form of (82)
is clear because matrix AA
w
is a Toeplitz matrix. That is,
(AA
w
)
kl
=a
kl
, where
a
k
=
1
N
_
N1
n=0
e
jt
n
2pk=N
; k= N1; . . . ; N 1 (83)
Observe that there are only N independent elements in
the array a since a
k
=(a
k
)
+
. Furthermore, these N ele-
ments in (83) can be obtained by the NUFFT-1 algorithm
a
k
=
1
N
_
N1
n=0
e
jtn2pk=N
; k=0; . . . ; N 1 (84)
with O(nN log N) arithmetic operations.
Since AA
w
is a Toeplitz matrix, the vector f in (82) can be
obtained efciently using the conjugate-gradient FFT
(CG-FFT) method [24]. In the CG-FFT method, the solu-
tion of f in (82) is obtained iteratively. Each iteration in-
volves operations such as y =(AA
w
)x, which can be written
as a discrete convolution
y
n
=
l
a
nl
.
x
l
={FFT
1
[FFT(a)FFT(x
p
)]]
j
(85)
and has been calculated by the regular FFT algorithm
through the convolution theorem. Note that in (85), the
size of the FFTs is 2N, and x
p
is the array x padded with N
zeros. After f is solved by the above CG-FFT method, we
can obtain h from (82) by
h(t
n
) =
N1
m=0
A
+
mn
f
m
=
1
N
_
N1
m=0
f
+
m
e
jmtn2p=N
_ _
+
(86)
which can be achieved by the NUFFT-2 algorithm.
In summary, the procedures for the inverse NUFFT-1
algorithm for Eq. (80) is
1. Preprocessinguse the NUFFT-1 to obtain array {a
j
}
through Eq. (84), and use the FFT algorithm to nd
[FFT(a)].
2. Inversionuse the CG-FFT method for the solution
f in (82).
3. Transpositionapply the NUFFT-2 to the trans-
posed problem in Eq. (86).
In these steps, since NUFFT-1 and NUFFT-2 require
O(nN log N) arithmetic operations, the most expensive
step is the CG-FFT solution. It requires O(KN log N)
arithmetic operations, where the number of iterations K
in the CG method is proportional to the condition number
of AA
w
. In most applications K is rather small as the points
{t
m
} are fairly uniformly spaced. A similar algorithm has
been developed for the inverse NUFFT of equation (81).
6.4. FFT for Discontinuous Functionsthe DFFT Algorithm
The NUFFT algorithms presented above have been devel-
oped for nonuniform discrete Fourier transform. However,
in practice, we often need to consider the continuous func-
tions. In many applications in antennas, scattering, and
computational electromagnetics, as well as in signal and
image processing and other areas of computational science
and engineering, one often needs to compute the Fourier
transform of discontinuous functions. For example, in the
well-known conjugate-gradient fast Fourier transform
(CG-FFT)-type methods for integral equations (see Ref.
13 and references cited therein), one needs to nd the
Fourier transform of the unknown elds; in discontinuous
media, some components of these unknown elds are dis-
continuous functions across the interfaces between differ-
ent materials. In the conventional CG-FFT methods,
discrete Fourier transforms (DFT) of these discontinuous
functions are calculated through FFT algorithms. Howev-
er, it is known that the conventional DFT cannot give ac-
curate results for discontinuous function. This approach
has a poor accuracy of O(N
l
), where N is the number of
sample points. This large error is a result of the trapezoi-
dal rule for the Fourier integration; in comparison, if the
function is smooth and periodic, the DFT has exponential
accuracy. A highly accurate fast Fourier transform algo-
rithm for discontinuous functions, called the discontinu-
ous fast Fourier transform (DFFT) algorithm, was
proposed by Fan and Liu [25,28], based on the previous
work for piecewise constant functions [71].
An accurate and efcient FFT algorithm for discontin-
uous data include two basic steps:
1. A highly accurate quadrature for improving the ac-
curacy of integration
2. The standard FFT for efciently evaluating the disc-
retized summation of the integration
To combine these two steps, we need to resort to an in-
terpolation procedure. Since the FFT algorithm requires
uniformly sampled points while the function to be trans-
formed has discontinuities at potentially nonuniform
points, the interpolation procedure should interpolate
such nonuniform data points to the uniform FFT data
points. Unfortunately, many existing interpolation algo-
rithms are either not directly applicable or cannot be uti-
lized in this problem. Previously, two algorithms for
discontinuous data were found in the literature [5,71]. So-
rets proposed a fast Fourier transform algorithm for piece-
wise constant functions via the Gaussian quadrature and
Lagrange interpolation, together with an application of
the standard FFT [71], but the algorithm cannot be di-
rectly applied to piecewise smooth functions. The FFT of
functions with similar singularities has also been treated
by Beylkin [5] using piecewise polynomial splines, al-
though numerical results are yet to be given for piecewise
smooth functions.
The DFFT reported by Fan and Liu [26,28] considers
the Fourier transform of a periodic piecewise smooth
function f(x) dened in a period T=1
^
ff (n) =
_
1
0
f (x)e
j2pnx
dx
N
2
on _
N
2
1 (87)
where n is an integer. We divide the integration interval
[0,1] into L subdivisions within which f(x) is a smooth
function. Step discontinuities are allowed at the bound-
aries of these subdivisions. In subdivision l (l =1,y, L),
there are at most p
l
uniformly or nonuniformly sampling
points, where p
l
is referred to as the order of interpolation
as shown later. Then (87) can be written as
^
ff (n) =
L
l =1
_
x
l
x
l1
f (x)e
j2pnx
dx (88)
where {x
l
} (l =1,y,L) are the endpoints of the subdivi-
sions. Using the change of variables
x =a
l
b
l
t =
x
l
x
l1
2

x
l
x
l1
2
t (89)
and the Gaussian Legendre quadrature
_
1
1
f (t)dt
q
k=1
f (t
k
)o
k
(90)
where t
k
and o
k
(k=1,y,q) are the GaussianLegendre
nodes and weights, respectively, we obtain
^
ff (n)
L
l =1
b
l

q
k=1
o
l
k
f (t
l
k
)e
j2pnt
l
k
(91)
In this expression, the values of both function f(x) and
the exponential function exp( i2pnx) at the nonuniform
distributed points {t
l
k
] can be obtained via the Lagrange
interpolation
g(x) =
p
m=1
g(x
m
)d
m
(x) (92)
where
d
m
(x) =
p
n=1
nOm
x x
n
x
m
x
n
(93)
Here we perform this interpolation separately for f(x) with
order p=p
1
and exp( j2pnx) with order p=p
2
. Note that
p
1
and p
2
can vary for different subdivisions; thus, they
take p
l
1
and p
l
2
, respectively. (However, to simplify the pre-
sentation, we take the same p value in the following dis-
cussion.)
Note that we use this double-interpolation scheme be-
cause these two functions to be interpolated have different
properties. The former, f(x), is a piecewise function, while
the latter, exp( j2pnx), is a periodic analytic function but
highly oscillating when n becomes large. Therefore we
normally interpolate the two functions with different sets
of sampling points and orders of interpolation. For the ex-
ponential function, we use a uniform grid in order to apply
the uniform FFT for the discrete summation; furthermore,
we always place the interpolated points at the center of
the interpolation region to obtain the highest accuracy for
a pth-order Lagrange interpolation. With this double
interpolation, (90) becomes
^
ff (n) =
L
l =1
b
l

q
k =1
o
l
k
p
1
m
1
=1
f (t
l
m
1
)d
m
1
(t
l
k
)
_ _
p2
m=1
e
j2pnt
l
m
d
m
(t
l
k
)
_ _
(94)
Transforming the local coordinate into a global coordinate
and changing the order of summation, we nally
obtain
^
ff (n) =
nN
i =1
g
i
e
j2pnx
i
(95)
where
g
i
=
L
l =1
q
k=1
F
lk
p
2
m=1
d
m
(t
l
k
)[
x(t
l
m
) =x
i
;
F
lk
=b
l
o
l
k
p
1
m
1
=1
f (t
l
m
1
)d
m
1
(t
l
k
)
(96)
and n is the oversampling factor. Note that F
lk
can be
calculated independently and stored in a temporary vector
of size NEL q, and all g
i
values can be obtained by
performing only once the triple summations over l, k, m,
only for the terms with x(t
l
m
) =x
i
. Hence the calculation
of g
i
is very efcient with O(nNp
2
) complexity. Finally, (95)
can be efciently evaluated via a standard uniform
FFT algorithm since {x
i
} are now on a uniform grid
in [0,1].
Furthermore, note that taking L=1, the above formu-
las are directly reduced to the FFT algorithm for nonuni-
formly sampled continuous functions. In such a
nonuniform fast Fourier transform (NUFFT) algorithm,
however, it is desirable to further subdivide the domain
(i.e., to increase L) when N is large, since this can reduce
the computational cost in the Lagrange interpolation, re-
sulting a more effective NUFFT algorithm for continuous
functions. Such an NUFFT algorithm by itself is very use-
ful in many engineering applications.
It is worth pointing out that compared with a single
interpolation procedure, the double-interpolation proce-
dure described above has several advantages:
*
Double interpolation can ensure that the interpola-
tion points for the exponential function are always
located at the center of the interpolation region, lead-
ing to the highest accuracy for a given Lagrange in-
terpolation order. In contrast, if a single interpolation
were applied to the product of these two functions,
the resulting function would not have the periodicity.
As a result, the locations of the interpolated points
would not be always located at the center of the in-
terpolation region, especially for those points near a
function discontinuity point. Therefore, the accuracy
of the interpolation for the product, and thus the ac-
curacy of the FT, would be significantly reduced, es-
pecially when n becomes large.
*
Since the Lagrange interpolation allows for nonuni-
form sampling, the algorithm described above can be
applied to both uniform and nonuniform sampled
data with the same number of arithmetic operations.
Hence, it provides an FFT algorithm for nonuniform-
ly sampled data. In contrast, a single interpolation
allows only uniformly sampled data in order to use
uniform FFT to evaluate the summation.
*
Double interpolation allows for a lower-order inter-
polation, that is, a lower sample density, for f(x) if it is
a slowly varying function within each subdivision.
*
Double interpolation allows other efcient algo-
rithms, such as the fast multipole method (FMM)
[23], to be incorporated readily into the interpolation
procedure.
Implementation of the algorithm includes the following
steps:
Step 1. Initialization of d
m
1
(t
l
k
) and d
m
(t
l
k
). This prepro-
cessing step is needed only once and has a complexity
of O(Np
2
), taking q =p=max(p
1
,p
2
).
Step 2. Calculation of F
lk
and g
i
in (96). The complex-
ities are O(nNp
1
) and O(nNp
2
), respectively.
Step 3. Calculation of
^
ff (n) in (95) by a standard FFT.
The complexity is O(nN log N).
The total complexity is O(nNpnN log N) for steps 2
and 3. The preprocessing step needs to be done only once,
and its computation time is negligible for repeated appli-
cations.
The DFFT algorithm has been utilized in several applica-
tions in computational electromagnetics, including the
CG-FFT method [26] and the precorrected FFT solution
[33].
7. CONCLUSIONS
The fast Fourier transform algorithm is a fundamental
numerical computation tool in electrical engineering. It
reduces the number of arithmetic operations in the dis-
crete Fourier transform from O(N
2
) to O(N log N). More
recent efforts have extended the FFT algorithm to non-
uniform sampled data (NUFFT), and to discontinuous
functions (DFFT). These algorithms have played and are
continuing to play a central role in many modern appli-
cations in electromagnetics.
Acknowledgment
The author acknowledges the nancial support by the Na-
tional Science Foundations under Grants IIS-0086075,
CCR-009814, and CCR-0219528, and by DARPA under
Grant DAAD19-02-1-0252.
BIBLIOGRAPHY
1. A. Abubakar and P. M. van den Berg, A conjugate gradient
contrast source technique for 3D prole inversion, IEICE
Trans. Electron. E83C(12):18641874 (2000).
2. A. Abubakar, P. M. van den Berg, and S. Y. Semenov, Two- and
three-dimensional algorithms for microwave imaging and in-
verse scattering, Prog. Electromagn. Res. 37:5779 (2002).
3. G. D. Bergland, The fast Fourier transform recursive equa-
tions for arbitrary length records, Math. Comput. 21:236238
(April 1967).
4. J. P. Berenger, A perfectly matched layer for the absorption of
electromagnetic waves, J. Comput. Phys. 114(2):185200
(1994).
5. G. Beylkin, On the fast fourier transform of functions with
singularities, Appl. Comput. Harm. Anal. 2:363382 (1995).
6. N. N. Bojarski, The k-space formulation of the scattering
problem in the time domain, J. Acoust. Soc. Am. 72:570584
(1982).
7. M. Born and E. Wolf, Principles of Optics, Pergamon Press,
New York, 1980.
8. D. T. Borup and O. P. Gandhi, Calculation of high-resolution
SAR distributions in biological bodies using the FFT algo-
rithm and conjugate-gradient method, IEEE Trans. Micro-
wave Theory Tech. 33(5):417419 (1985).
9. E. O. Brigham, The Fast Fourier Transform, Prentice-Hall,
Englewood Cliffs, NJ, 1974.
10. C. Canuto, M. Y. Hussani, A. Quarteroni, and T. Zang, Spec-
tral Methods in Fluid Dynamics. SpringerVerlag, New York,
1988.
11. S. Caorsi, G. L. Gragnani, and M. Pastorino, A multiview mi-
crowave imaging system for two-dimensional penetrable ob-
jects, IEEE Trans. Microwave Theory Tech. 39:845851
(1991).
12. M. F. Catedra, J. G. Cuevas, and L. Nuno, A scheme to ana-
lyze conducting plates of resonance size using the conjugate-
gradient method and fast Fourier transform, IEEE Trans.
Anten. Propag. 36:17441752 (1988).
13. M. F. Catedra, R. P. Torres, J. Basterrechea, and E. Gago, The
CG-FFT Method: Application of Signal Processing Techniques
to Electromagnetics, Artech House, Norwood, MA, 1995.
14. W. C. Chew, Waves and Fields in Inhomogeneous Media, Van
Nostrand Reinhold, New York, 1990.
15. W. C. Chew and Y. M. Wang, Reconstruction of two-dimen-
sional permittivity distribution using the distorted Born iter-
ation method, IEEE Trans. Med. Imag. 9:218225 (1990).
16. W. C. Chew, J. M. Jin, E. Michielssen, and J. M. Song, eds.,
Fast and Efcient Algorithms in Computational Electromag-
netics, Artech House, BostonLondon, 2001.
17. R. Coifman, V. Rokhlin, and S. Wandzura, The fast multipole
method for the wave equation: A pedestrian prescription,
IEEE Anten. Propag. Mag. 35:712 (1993).
18. J. W. Cooley and J. W. Tukey, Algorithm for the machine com-
putation of complex fourier series, Math. Comput. 19:297301
(1965).
19. T. J. Cui and W. C. Chew, Fast algorithm for electromagnetic
scattering by Buried 3-D Dielectric objects of large size, IEEE
Trans. Geosci. Remote Sens. 37:25972608 (1999).
20. T. J. Cui, W. C. Chew, A. A. Aydiner, and Y. H. H. Zhang, Fast-
forward solvers for the low-frequency detection of buried di-
electric objects, IEEE Trans. Geosci. Remote Sens. 41(9):2026
2036 (Sept. 2003).
21. G. Danielson and C. Lanczos, Some improvements in practi-
cal Fourier analysis and their application to x-ray scattering
from liquids, J. Franklin Inst. 233:365380, 435452 (1942).
22. A. J. Devaney, A ltered backpropagation algorithm for
diffraction tomography, Ultrasound. Imag. 4:336350 (1982).
23. A. Dutt, M. Gu, and V. Rokhlin, Fast algorithms for polyno-
mial interpolation, integration, and differentiation, SIAM J.
Num. Anal. 33:16891711 (1996).
24. A. Dutt and V. Rokhlin, Fast fourier transforms for nonequi-
spaced data, SIAM J. Sci. Comput. 14:13681393 (1993).
25. G.-X. Fan and Q. H. Liu, Fast fourier transform of functions
with jump discontinuities, Intl. IEEE/AP-S Symposium
Digest, Salt Lake City, UT, 2000, pp. 148151.
26. G.-X. Fan and Q. H. Liu, The CGFFT method with a discon-
tinuous fft algorithm, Microwave Opt. Technol. Lett. 29(1):
4749 (2001).
27. G. X. Fan, Q. H. Liu, and J. S. Hesthaven, Multi-domain
pseudospectral time-domain method for simulation of scat-
tering from buried objects in lossy media, IEEE Trans. Geosci.
Remote Sens. 40(6):13661373 (June 2002).
28. G.-X. Fan and Q. H. Liu, Fast fourier transform for discon-
tinuous functions, IEEE Trans. Anten. Propag. 52(2):461465
(Feb. 2004).
29. J. A. Fessler and B. P. Sutton, Nonuniform fast fourier trans-
forms using min-max interpolation, IEEE Trans. Signal Pro-
cess. 51(2):560574 (Feb. 2003).
30. B. Fornberg. A Practical Guide to Pseudospectral Methods,
Cambridge Univ. Press, New York, 1996.
31. X. Gao, M. S. Mirotznik, S. Y. Shi, and D. W. Prather, Apply-
ing a mapped pseudospectral time-domain method in simu-
lating diffractive optical elements, J. Opt. Soc. Am. A
21(5):777785 (2004).
32. C. Gauss, Theoria Interpolationis Methodo Nova Tractata,
Carl Friedrich Gauss Werke, Vol. 3, K niglichen Gesellschaft
der Wissenschaften, Gottingen, 1866.
33. S. Gedney, A. Zhu, T. W. Hua, and P. Petre, High-order
pre-corrected fft solution for electromagnetic scattering,
Proc. Int. IEEE/AP-S Symp., San Antonio, TX, 2002, Vol. 3,
pp. 566569.
34. D. Gottlieb and S. A. Orszag, Numerical Analysis of Spectral
Methods, SIAM, Philadelphia, 1977.
35. R. F. Harrington, Field Computation by Moment Methods,
Macmillan, New York, 1968.
36. J. S. Hesthaven, P. G. Dinesen, and J. P. Lynov, Spectral col-
location time-domain modeling of diffractive optical elements,
J. Comput. Phys. 155:287306 (1999).
37. H. Gan and W. C. Chew, A discrete BCG-FFT algorithm for
solving 3D inhomogeneous scatterer problems, J. Elect-
romagn. Waves Appl. 9:13391357 (1995).
38. N. Joachimowicz, C. Pichot, and J. P. Hugonin, Inverse scat-
tering: An iterative numerical method for electromagnetic
imaging, IEEE Trans. Anten. Propag. 39:17421752 (Dec.
1991).
39. R. Kastner and R. Mittra, Iterative analysis of nite-size pla-
nar frequency selective surfaces with rectangular patches or
perforations, IEEE Trans. Anten. Propag. 35:372377 (1987).
40. H. O. Kreiss and J. Oliger, Comparison of accurate methods
for the integration of hyperbolic equations, Tellus 24:199215
(1972).
41. T. W. Lee and S. C. Hagness, A PSTD method with fourth-
order time-stepping for the analysis of nonlinear optical
structures, IEEE Antennas and Propagation Society Int.
Symp. Digest, 2002, Vol. 3, pp. 232235.
42. W. K. Leung and C. H. Chan, Combining the FDTD and PSTD
methods, Microwave Opt. Technol. Lett. 23(4):249254 (1999).
43. W. K. Leung and Y. Chen, Transformed-space nonuniform
pseudospectral time-domain algorithm, Microwave Opt. Tech-
nol. Lett. 28(6):391396 (2001).
44. Q. Li, Y. Chen, and C. K. Li, Hybrid PSTD-FDTD technique
for scattering analysis, Microwave Opt. Technol. Lett.
34(1):1924 (2002).
45. F. Li, Q. H. Liu, and L.-P. Song, Three-dimensional recon-
struction of objects buried in layered media using Born and
distorted Born iterative methods, IEEE Geosci. Remote Sens.
Lett. 1(2):107111 (2004).
46. F. Ling, C. F. Wang, and J. M. Jin, An efcient algorithm for
analyzing large-scale microstrip structures using adaptive
integral method combined with discrete complex-image meth-
od, IEEE Trans. Microwave Theory Tech. 48(5):832939 (May
2000).
47. Q. L. Li and Y. C. Chen, Applications of the PSTD for scat-
tering analysis, IEEE Trans. Anten. Propag. 50(9):13171319
(2002).
48. Q. H. Liu, A spectral-domain method with perfectly matched
layers for time-domain solutions of Maxwells equations, pa-
per presented at 1996 URSI Meeting, Baltimore, MD, July
1996.
49. Q. H. Liu, The PSTD algorithm: A time-domain method re-
quiring only two cells per wavelength, Microwave Opt. Tech-
nol. Lett. 15(3):158165 (1997).
50. Q. H. Liu, Large-scale simulations of electromagnetic and
acoustic measurements using the pseudospectral time-do-
main (PSTD) algorithm, IEEE Trans. Geosci. Remote Sens.
37(2):917926 (1999).
51. Q. H. Liu, PML and PSTD algorithm for arbitrary lossy an-
isotropic media, IEEE Microwave Guided Wave Lett. 9(2):48
50 (1999).
52. Q. H. Liu and N. Nguyen, An accurate algorithm for nonuni-
form fast Fourier transforms (NUFFT), IEEE Microwave
Guided Wave Lett. 8(1):1820 (1998).
53. Q. H. Liu and G.-X. Fan, A frequency-dependent PSTD algo-
rithm for general dispersive media, IEEE Microwave Guided
Wave Lett. 9(2):5153 (1999).
54. Q. H. Liu and G.-X. Fan, Simulations of GPR in dispersive
media using the PSTD algorithm, IEEE Trans. Geosci. Re-
mote Sens. 37(5):23172324 (1999).
55. Q. H. Liu and J. Q. He, An efcient PSTD algorithm for cy-
lindrical coordinates, IEEE Trans. Anten. Propag. 49(9):1349
1351 (2001).
56. Q. H. Liu, X. M. Millard, B. Tian, and Z. Q. Zhang, Applica-
tions of nonuniform fast transform algorithms in numerical
solutions of differential and integral equations, IEEE Trans.
Geosci. Remote Sens. 38:15511560 (2000).
57. Q. H. Liu and X. Y. Tang, An iterative algorithm for nonuni-
form inverse fast fourier transform (NU-IFFT), Electron. Lett.
34(20):19131914 (1998).
58. Q. H. Liu and Z. Q. Zhang, A nonuniform fast hankel trans-
form (nufht) algorithm, Appl. Opt. 38(32):67056708 (1999).
59. Q. H. Liu and G. Zhao, Review of PSTD Methods for transient
electromagnetics, Int. J. Num. Model. Electron. Networks
Devices Fields 22(17):299323 (2004).
60. K. A. Michalski and D. Zheng, Electromagnetic scattering
and radiation by surface of arbitrary shape in layered media,
Part I: Theory, IEEE Trans. Anten. Propag. 38(3):335344
(March 1990).
61. K. A. Michalski and D. Zheng, Electromagnetic scattering
and radiation by surface of arbitrary shape in layered media,
Part II: Implementation and results for contiguous half-space,
IEEE Trans. Anten. Propag. 38(3):344352 (March 1990).
62. K. A. Michalski and J. R. Mosig, Multilayered media Greens
functions in integral equation formulations, IEEE Trans. An-
ten. Propag. 45(3):508519 (March 1997).
63. X. Millard and Q. H. Liu, A fast volume integral equation
solver for electromagnetic scattering from large inhomoge-
neous objects in planarly layered media, IEEE Trans. Anten.
Propag. 51(9):23932401 (2003).
64. X. Millard and Q. H. Liu, Simulations of near-surface detec-
tion of objects in layered media by the BCGS-FFT method,
IEEE Trans. Geosci. Remote Sens. 42(2):327334 (2004).
65. N. Nguyen and Q. H. Liu, The regular fourier matrices and
nonuniform fast fourier transforms, SIAM J. Sci. Comput.
21(1):283293 (1999).
66. V. I. Okhmatovski, J. D. Morsey, and A. C. Cangellaris, En-
hancement of the numerical stability of the adaptive integral
method at low frequencies through a loop-charge formulation
of the method-of-moments approximation, IEEE Trans.
Microwave Theory Tech. 52(3):962970 (2004).
67. S. A. Orszag, Comparison of pseudospectral and spectral
approximation, Stud. Appl. Math. 51:253259 (1972).
68. T. J. Peters and J. L. Volakis, Application of a conjugate gra-
dient FFT method to scattering from thin planar material
plates, IEEE Trans. Anten. Propag. 36:518526 (1988).
69. T. K. Sarkar, E. Arvas, and S. M. Rao, Application of fast
Fourier transform and the conjugate gradient method for ef-
cient solution of electromagnetic scattering from both elec-
trically large and small conducting bodies, Electromagnetics
5:99122 (1985).
70. J. M. Song and W. C. Chew, Multilevel fast-multipole algorithm
for solving combined eld integral equations of electromagnet-
ic scattering, Microwave Opt. Technol. Lett. 10:1519 (1995).
71. E. Sorets, Fast fourier transforms of piecewise constant func-
tion, J. Comput. Phys. 116:369379 (1995).
72. M. Soumekh, Synthetic Aperture Radar Signal Processing
with Matlab Algorithms, Wiley, New York, 1999.
73. C. C. Su, The three-dimensional algorithm of solving the elec-
tric eld integral equation using face-centered node points,
conjugate gradient method, and FFT, IEEE Trans. Microwave
Theory Tech. 41:510515 (1993).
74. B. Tian and Q. H. Liu, Nonuniform fast cosine transform and
Chebyshev PSTD algorithm, Prog. Electromagn. Res. PIER
28:259279 (2000); abstract in J. Electromagnet. Waves Appl.
14(6):797798 (2000).
75. L. Tsang, C. H. Chan, K. Pak, and H. Sangani, Monte Carlo
simulations of large-scale problems of random rough surface
scattering and applications to grazing incidence with the
BMIA/canonical grid method, IEEE Trans. Anten. Propag.
43:851859 (1995).
76. J. W. Wiskin, D. T. Borup, and S. A. Johnson, Inverse scat-
tering from arbitrary two-dimensional objects in stratied
environments via a Greens operator, J. Acoust. Soc. Am.
102(2):853864 (1997).
77. X. M. Xu, Q. H. Liu, and Z. Q. Zhang, The stabilized bicon-
jugate gradient fast Fourier transform method for electro-
magnetic scattering, J. Appl. Comput. Electromagn. Soc.
17(1):97103 (2002).
78. X. M. Xu and Q. H. Liu, The BCGS-FFT method for electromag-
netic scattering from inhomogeneous objects in a planarly lay-
ered medium, IEEEAnten. Wireless Propag. Lett. 1:7780 (2002).
79. B. Yang, D. Gottlieb, and J. S. Hesthaven, Spectral simulation
of electromagnetic wave scattering, J. Comput. Phys.
134:216230 (1997).
80. B. Yang and J. S. Hesthaven, A pseudospectral method for time-
domain computation of electromagnetic scattering by bodies of
revolution, IEEE Trans. Anten. Propag. 47:132141 (1999).
81. B. Yang and J. S. Hesthaven, Multidomain pseudospectral
computation of Maxwells equations in 3-D general curvilin-
ear coordinates, Appl. Num. Math. 33:281289 (2000).
82. K. S. Yee, Numerical solution of initial boundary value prob-
lems involving Maxwells equations in isotropic media, IEEE
Trans. Anten. Propag. 14(5):302307 (1966).
83. Z. Q. Zhang and Q. H. Liu, Three-dimensional weak-form
conjugate- and biconjugate-gradient FFT methods for volume
integral equations, Microwave Opt. Technol. Lett. 29(5):350
356 (2001).
84. Z. Q. Zhang, Q. H. Liu, and X. M. Xu, RCS computation of
large inhomogeneous objects using a fast integral equation
solver, IEEE Trans. Anten. Propag. 51(3):613618 (March
2003).
85. Z. Q. Zhang and Q. H. Liu, 3-D nonlinear image reconstruc-
tion for microwave biomedical imaging, IEEE Trans. Biomed.
Eng. 51(3):544548 (2004).
86. G. Zhao and Q. H. Liu, The 2.5-Dmultidomain pseudospectral
time-domain algorithm, IEEE Trans. Anten. Propag.
51(3):619627 (March 2003).
87. G. Zhao, Y. Q. Zeng, and Q. H. Liu, The 3-D multidomain
pseudospectral time-domain method for wideband simula-
tion, IEEE Microwave Wireless Compon. Lett. 13(5):184186
(2003).
88. G. Zhao and Q. H. Liu, The 3-D multidomain pseudospectral
time-domain method for inhomogeneous conductive media,
IEEE Trans. Anten. Propag. 52(3):742749 (March 2004).
89. Y. A. Zhunag, K. L. Wu, C. Wu, and J. Litva, A combined full-
wave CG-FFT method for rigorous analysis of large micro-
strip antenna arrays, IEEE Trans. Anten. Propag. 44(1):102
109 (Jan. 1996).
90. P. Zwamborn and P. M. van den Berg, The three-dimensional
weak form of the conjugate gradient FFT method for solving
scattering problems, IEEE Trans. Microwave Theory Tech.
9(40):17571766 (1992).
FEEDBACK AMPLIFIERS
THOMAS F. SCHUBERT Jr.
University of San Diego
San Diego, California
Feedback is the process of combining a portion of the out-
put of a system with the system input to achieve modied
performance characteristics. Found in a multitude of
engineering applications, the process has become the foun-
1418 FEEDBACK AMPLIFIERS
dation several disciplines such as feedback control sys-
tems, feedback receivers, feedback oscillators, and feed-
back ampliers. It has become particularly pervasive in
amplier design since the advent of transistors that can
provide high gain cheaply, but cannot provide stable gain.
Negative feedback was originally applied by Harold S.
Black in 1927 at Bell Labs to valve (vacuum tube) ampli-
ers to reduce distortion. In 1957 Blacks work was de-
scribed by Mervin Kelly, president of Bell Labs, as one of
the two inventions of broadest scope and significance in
electronics and communications is the rst half of the cen-
tury. Negative feedback is the process of mixing an in-
verted portion of the output of a system with the input to
alter system performance characteristics. The process of
negative feedback has become especially important in lin-
ear amplier design and is characterized by several sig-
nificant benets:
*
The gain of the amplier is stabilized against varia-
tion in the characteristic parameters of the active de-
vices due to voltage or current supply changes,
temperature changes, or device degradation with
age. Similarly, amplier gain is stabilized within a
group of ampliers that have active devices with
somewhat different characteristic parameters.
*
Nonlinear signal distortion is reduced.
*
The frequency range over which there is constant
linear amplication (the midband region) is in-
creased.
*
The input and output impedance of the amplier can
be selectively increased or decreased.
It is a rare occurrence when benets come without a price.
In the case of negative feedback, the listed above benets
are accompanied by two primary drawbacks:
*
The gain of the circuit is reduced. In order to regain
the losses due to feedback, additional amplication
stages often must be included in the system design.
Complexity, size, weight, and cost may be added to
the nal amplier design.
*
There is a possibility for oscillation to occur. Should
oscillation occur, the basic gain properties of the am-
plier are destroyed.
This article considers the benets of negative feedback on
amplier design. Basic definitions are followed by a gen-
eral discussion of the properties of a feedback system.
Ampliers are divided into four categories of feedback to-
pology, and the specific properties of each topological type
are stated. While the emphasis of discussions must focus
on circuit analysis techniques, a clear understanding of
feedback in general, and effects of circuit topology in par-
ticular, is a necessity for good feedback amplier design.
1. BASIC FEEDBACK CONCEPTS
Electronic ampliers are fundamentally characterized by
four properties:
*
The gain of the amplier. Gain is dened as the ratio
of the output signal to the input signal. Each signal
may be either a voltage signal or current signal.
*
The range of frequencies over which the gain is es-
sentially constant. This range of frequencies is iden-
tied as the midband region. It is bounded by the
frequencies at which the output power is halved: the
high and low 3-dB frequencies.
*
The midband input impedance. Dened as the ratio
of input voltage to input current.
*
The midband output impedance. Dened as the
Thevenin impedance seen at the amplier output ter-
minals.
The application of feedback to an amplier alters each of
these fundamental properties.
The basic topology of a feedback amplier is shown in
Fig. 1. An input signal X
i
enters a summing (or mixing)
junction, symbolized by a circular symbol. At the summing
junction, the feedback signal X
f
is subtracted from the in-
put signal and the resultant signal X
d
is passed on to a
linear amplier, shown as a triangular symbol, of gain A.
The output of the linear amplier X
o
forms the output of
the feedback system. The rectangular symbol indicates a
feedback network that samples the output signal, scales it
by a feedback factor f and passes it forward to the input of
the system. Each signal can take the form of either a volt-
age or a current, and ideally travels only in the direction of
the indicated arrows. Subtraction of the two inputs at the
summing junction is the key factor in negative feedback
systems.
Feedback systems can be mathematically modeled with
a set of simple relationships. The output X
o
of the ampli-
er is related to the input signal X
d
by a linear amplica-
tion factor (gain) A, often called the forward or open-loop
gain:
X
o
=A(X
d
) (1)
Since the quantities X
o
and X
d
can be either voltage or
current signals, the forward gain A can be a voltage gain, a
current gain, a transconductance, or a transresistance.
Voltage gain is the ratio of output voltage to input voltage.
Similarly, current gain relates output and input currents.
Transresistance is the ratio of output voltage to input cur-
rent and transconductance is the ratio of output current to
input voltage. The feedback signal X
f
(a fraction of the
output signal X
o
) is then subtracted from the input signal
A
f
+

X
f
X
i
X
o
X
o
Figure 1. Basic negative feedback topology.
FEEDBACK AMPLIFIERS 1419
X
i
to form the difference signal X
d
X
d
=(X
i
X
f
) =(X
i
fX
o
) (2)
where f is the feedback ratio dening the relationship be-
tween X
f
and X
o
:
X
f
=f X
o
(3)
As is the case with the amplier gain, the feedback ratio f
is either a ratio of voltages, a ratio of currents, a trans-
conductance, or a transresistance. The product fA, called
the loop gain, must be a positive, dimensionless quantity
in order to have stable negative feedback. Thus it is nec-
essary, for negative feedback, that the mathematical sign
of f be the same as that of A within the midband region.
The inputoutput relationship for the overall feedback
system can be derived from Eqs. (1) and (2):
X
o
=
A
1fA
X
i
=A
f
X
i
(4)
The overall gain of the system including the effects of
feedback is then written as
A
f
=
X
o
X
i
=
A
1 fA
(5)
The overall feedback amplier gain A
f
has the same di-
mensions as the forward gain A. Equation (5) has special
significance in the study of feedback systems and is typ-
ically identied as the basic feedback equation. The de-
nominator of the basic feedback equation is identied as
the return difference D but is also commonly referred to as
the amount of feedback:
D=1fA (6)
The return difference, for negative-feedback systems, has
magnitude larger than unity (in the midband frequency
region) and is often specied in decibels:
D
dB
=20 log
10
1fA
(7)
The return difference quanties the reduction in gain due
to the addition of feedback to the system. It also plays a
significant role in quantifying changes in frequency band-
width and input and output impedance. Specifically, the
high and low 3-dB frequencies are increased and de-
creased, respectively, by approximately a factor of D, and
the input and output impedance are increased or de-
creased by a factor of D depending on the sampling and
mixing circuit topologies.
Derivation of the basic feedback equation is based on
two basic assumptions:
*
The reverse transmission through the amplier is
negligible (a signal at the output produces essentially
no signal at the input) compared to the reverse trans-
mission through the feedback network.
*
The forward transmission (left to right in Fig. 1)
through the feedback network is negligible compared
to the forward transmission through the amplier.
In most feedback ampliers, the amplier is an active de-
vice with significant forward gain and near-zero reverse
gain: the feedback network is almost always a passive
network. Thus, in the forward direction, the large active
gain will exceed the passive attenuation significantly.
Similarly, in the reverse direction, the gain of the feed-
back network, albeit typically small, is significantly great-
er than the near-zero reverse gain of the amplier. In
almost every electronic application, the above-stated re-
quirements for the use of the basic feedback equation are
easily met by the typical feedback amplier.
2. ANALYSIS OF FEEDBACK AMPLIFIER PROPERTIES
The analysis and design of electronic ampliers is typical-
ly a multistep process. Complete, analytic characteriza-
tion of an amplier is a complex process whose results are
in a form that often masks the individual amplier prop-
erties. Amplier designers therefore investigate the am-
plier gain, frequency response, and impedance properties
separately, carefully balancing system requirements to
converge on a successful design. In addition to simplify-
ing the process, separate investigation of the amplier
properties often leads to greater insight into design im-
provement. When a successful design is apparent, nal
ne tuning is accomplished with the aid of a computerized
circuit simulator [i.e., System Program with Integrated
Circuit Emphasis (SPICE)] and a prototype.
Essentially all of the drawbacks and benets of feed-
back systems can be investigated on a simple level by
looking at the properties of the basic feedback equation
[Eq. (5)]. Gain stabilization, the reduction in nonlinear
signal distortion, the increase in the frequency range over
which there is linear amplication, the reduction in gain,
and the possibility of oscillation all can be investigated on
a simple level. The change in the input and output im-
pedances cannot be investigated at this level: it is neces-
sary to specify the nature (voltage or current) of the input
and output quantities and the circuit topology in order to
investigate these impedance changes.
2.1. Amplier Gain
In Section 1, it was shown that feedback reduces amplier
gain by a factor of the return difference D. While reduction
of gain can be a significant drawback, the ancillary gains
are significant. Primary among those benets is the sta-
bilization of the amplier gain against variation in the
characteristic parameters of the active devices. It is well
known that the forward gain A of an electronic amplier is
highly dependent on the parameters of the active devices
contained within that amplier. These parameters are
typically dependent on temperature, bias conditions, and
manufacturing tolerances. In order to maintain consistent
amplier performance it is desirable to design ampliers
that are reasonably insensitive to the variation of the de-
vice parameters.
The relationship between the differential change in
gain due to device parameter variation with and without
feedback is obtained by differentiating Eq. (5):
dA
f
=
1
1 fA ( )
2
dA (8)
Stable negative-feedback ampliers require that the re-
turn difference have magnitude greater than unity:
1 (9)
Thus the absolute variation in gain is reduced by a factor
of the return ratio squared. Another measure of the
change in gain variation can be found by regrouping
terms:
dA
f
A
f
=
1
1 fA ( )
dA
A
(10)
The factors in this equation more realistically describe the
benets: Equation (10) demonstrates the change in the
percentage of gain variation about the nominal value. It
can be seen that the percentage variation of the overall
amplier gain A
f
is reduced by a factor of the return ratio
when compared to the percentage variation of the gain A
of the forward amplier.
For example, if a feedback amplier is constructed with
an amplier that is subject to a 3% variation in gain as its
fundamental forward-gain element and it is desired that
the feedback amplier have no more than 0.1% variation
in its overall gain due to the variation in this element. The
necessary return difference to achieve this design goal can
be obtained as follows:
Equation (10) is the significant relationship in deter-
mining the gain variation:
dA
f
A
f
=
1
1 fA ( )
dA
A
The significant properties are

0:001_
1
1 fA ( )
0:03 = D=1 Af _30

The minimum necessary return ratio is 30, more often
identied as its decibel equivalent, D
dB
=20
log
10
DZ29.54 dB
Equation (10) is extremely useful for small changes in
amplication due to parameter variation but is inaccurate
for large changes. If the change in amplication is large,
the mathematical process must involve differences rather
than differentials:
DA
f
=A
2f
A
1f
=
A
2
1fA
2
A
1
1fA
1
(11)
In order to put this into the same format as Eq. (10), it is
necessary to divide both sides of the equation by A
1f
DA
f
A
1f
=
A
2
1 fA
2
1 fA
1
A
1
_ _
1
=
A
2
A
1
( )
1 fA
2
1
A
1
_ _
(12)
or
DA
f
A
1f
=
1
1fA
2
DA
A
1
=
1
1 f A
1
DA ( )
DA
A
1
(13)
The results are similarly a reduction in gain sensitivity by
a factor of the form of the return difference.
The differential change in feedback amplier gain due
to variation in the feedback ratio f can be obtained by dif-
ferentiating Eq. (5) with respect to the feedback ratio. Ap-
propriate mathematical manipulation leads to the desired
results:
dA
f
A
f
=
fA
1f
df
Af
=
dA
f
A
f
=
fA
1 fA
df
f
(14)
It is easily seen that the percentage variation of the over-
all amplier gain A
f
is of approximately the same magni-
tude (actually it is slightly lesser) as the percentage
variation of the feedback ratio f. Since electronic feedback
ampliers typically employ a feedback network construct-
ed entirely with passive elements (usually resistors), vari-
ation in the feedback ratio can be kept relatively small
through the utilization of precision elements in the feed-
back network. In good amplier designs, the variation of
amplier gain due to variability in the feedback network is
usually of lesser significance than that due to variability
of the basic forward amplier gain.
2.2. Nonlinear Signal Distortion
Stabilization of gain with parameter variation suggests
that amplier gain will be stabilized with respect to other
gain-changing effects. One such effect is nonlinear distor-
tion. Nonlinear distortion is a variation of the gain with
respect to input signal amplitude. A simple example of
non-linear distortion is demonstrated in Fig. 2. Here the
transfer characteristic of a simple amplier is approxi-
mated by two regions, each of which is characterized by
X
o
A
1
A
1f
A
2
A
2f
X
i
With
feedback
No
feedback
Figure 2. The effect on feedback on an amplier transfer char-
acteristic.
different amplication, A
1
and A
2
. To this transfer char-
acteristic, a small amount of feedback is applied so that fA
1
=1, and the resultant feedback transfer characteristic is
shown. As can be easily seen, the overall feedback transfer
characteristic also consists of two regions with overall
amplication A
1f
and A
2f
. In this demonstration, the am-
plication ratios are:
A
1
A
2
=3;
A
1f
A
2f
=1:5
Feedback has significantly improved the linearity of the
system and consequently has reduced the nonlinear distor-
tion. Larger amounts of feedback (increasing the feedback
ratio f) will continue to improve the linearity. For this ex-
ample, increasing the feedback ratio by a factor of 5 will
result in a ratio of overall gain in the two regions of 1.067
(as compared to 1.5, above). It should be noted that the
saturation level of an amplier is not significantly altered
by the introduction of negative feedback. Since the incre-
mental gain in saturation is essentially zero, the incremen-
tal feedback difference is also zero. No significant change to
the input occurs and the output remains saturated.
Another viewpoint on gain stabilization comes from a
limiting form of the basic feedback equation:
A
f
=
A
1 fA
=
1
f
1
1
1fA
_ _
-
1
f
(15)
For large return difference (D=1 Af ) the overall gain
with feedback is dominated by the feedback ratio f and is
therefore virtually independent of the forward gain A and
any variations in A.
2.3. Frequency Response
Typical linear ampliers have a range of frequencies over
which the gain is essentially constant: this frequency
range is called the midband. As frequencies increase,
the performance parameters of an amplier degrade. Sim-
ilarly, coupling and bypass capacitors internal to the am-
plier, when present, will degrade low-frequency
performance. Using feedback to broaden the frequency
range over which gain is relatively constant can be con-
sidered a special case of the stabilization of gain due to
variation in amplier performance characteristics. Feed-
back reduces the effects of these frequency-dependent de-
gradations and thereby increases the frequency band over
which the amplier has stable gain.
A more exact description of the increase in the width of
the midband region can be obtained through a frequency-
domain analysis. It is common practice to use the frequen-
cies at which the output power is reduced by 50% (the high
and low 3-dB frequencies) as descriptors of the limits of
the midband region. Discussion focuses on the change in
these 3-dB frequencies.
It can be shown that the basic forward amplier gain A
is described as the midband gain A
0
, divided by a polyno-
mial in frequency (written as s or jo)
A(s) =
A
0
P(s)
(16)
The locations of the rst few poles of A(s) {or the zeroes of
P(s)} closest to the midband region are the dominant pre-
dictors of amplier frequency response; specifically, their
location controls the 3-dB frequencies.
The application of feedback to an amplier alters the
gain expression through the basic feedback equation so
that the total gain A
f
is described by:
A
f
(s) =
A
0
P(s)
1f
A
0
P(s)
=
A
0
P(s) fA
0
(17)
Application of feedback to the basic forward amplier has
the effect of vertically shifting the denominator polynomi-
al by a constant, f A
0
(see Fig. 3). This shift upward causes
movement in the zeros of the denominator, thereby chang-
ing the location of the poles of the frequency response. Any
movement of the poles nearest the region of constant gain
(the midband region) equates into a change in the width of
the midband region. Observation of the consequences of
the result in graphical format is a great aid to under-
standing pole migration.
For example, when the high-frequency response is de-
scribed by a single pole p
1
, Eq. (17) takes the following
form:
A
f
(s) =
A
0
P(s) fA
0
=
A
0
1
s
p
H1
_ _
fA
0
=
1
1 fA
0
A
0
1
s
p
H1
1fA
0
( )
_ _
(18)
It can easily be seen that the gain has been reduced by a
factor of the return difference D and the pole frequency
has been increased by the same factor. Similarly, if the
low-frequency response is described by a single pole p
L1
,
the pole frequency will be reduced (i.e., divided by) by a
factor of D. Since, in single-pole systems, the 3-dB fre-
quency coincides with the pole frequencies, the high and
low 3-dB frequencies, o
H
and o
L
, are shifted by a factor of
the return ratio:
o
Hf
=(1 fA
0
)o
H
=Do
H
or o
Lf
=
o
L
1fA
0
=
o
L
D
(19)
Increased
feedback

Pole
migration
Pole
migration
Pole
migration
p
5
p
4
p
3
p
2 p
1
c
P(s)
Figure 3. Pole migration due to feedback.
As the number of poles increases, description of the band-
width increase with the application of feedback increases
in complexity. When the high or low frequency response
can be described by two poles, the damping coefcient due
to the application of feedback is function ratio of the initial
pole locations k and the return difference. The damping
coefcient can be calculated to be
z =
1 k
2
k 1fA
0
( )
_ =
1 k
2
kD
_ (20)
where k is dened as the ratio of the larger pole to the
smaller pole at high or low frequencies:
k
H
=
o
2H
o
1H
or k
L
=
o
1L
o
2L
(21)
This simple expression for the damping coefcient is a
particularly important result in that it can tell the circuit
designer the atness of the frequency response in relation
to the amount of feedback applied; a at frequency re-
sponse requires overdamped pole pairs (z_0:707).
Once the damping coefcient is determined, the exact
expression for the high or low 3-dB frequency shift with
the application of feedback takes a form similar to the
single-pole case with an additional factor
o
Hf
=K(z
H
; k
H
)
.
D
.
o
1H
or o
Lf
=
o
1L
K(z
L
; k
L
)
.
D
(22)
where k is the ratio of the initial pole spacing (k_1); z is
the pole pair damping coefcient, o
1H
and o
1L
are the
poles closest to the midband, and the factor, K(z; k), is
given by
K z; k ( ) =
2kz
k 1
1 2z
2
1 2z
2
_ _
2
1
_
_
(23)
This relationship is shown in Fig. 4 for a variety of initial
pole spacing ratios k. In most amplier applications,
0:9oK(z; k)o

2
_
; some designers use a K(z; k) - 1 as a
rst-order approximation.
For ampliers where the frequency response must be
described by more than two poles, description of the fre-
quency shift is even more complicated. Fortunately, am-
pliers with a high or low frequency response that is
described by more than two poles are reasonably modeled
by considering them to be two-pole systems [13]: Eq. (22)
adequately approximates the change in bandwidth for
these higher-order systems.
For example, if an amplier has a midband gain A
0
=
1000 and has frequency response described by one low-
frequency pole, f
L
=10 kHz and two high-frequency poles,
f
H1
=1MHz and f
H2
=10 MHz and feedback is applied so
that the midband gain is reduced to A
0f
=140. The new
low and high 3-dB frequencies can be determined as fol-
lows:
The return difference is the ratio of the two gains:
D=
A
0
A
0f
=
1000
140
=7:14286
The low-frequency response is described by a single
pole; thus the low 3-dB frequency is changed by a
factor of D:
f
Lf
=
f
L
D
=
10 kHz
7:14286
=1:4 kHz
The high-frequency response is described by two poles
with ratio k:
k =
o
2H
o
1H
=
f
2H
f
1H
=
10 MHz
1 MHz
=10
The damping coefcient for the two poles are found to
be
z
H
=
1k
2
k 1 fA
0
( )
_ =
110
2
10 7:14286 ( )
_ =0:6508
Note that the high poles of the feedback amplier are
slightly underdamped and that there will be a small
bump (E0.1dB for this case) in the frequency response
as a result. The high 3-dB frequency f
Hf
is then found from
0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
0.4
0.5
0.6
0.7
0.8
0.9
1.0
1.1
1.2
1.3
1.4
k = 2
k = 5
k = 10
k = 50
k = 20
c
H
D
c
1
Figure 4. High 3-dB frequency as a function

of z and nonfeedback pole spacing.
K(z
H
; k); D, and f
1
:
f
Hf
=K(z
H
; k)Df
1
=(1:277)(7:14286)(1MHz) =9:12 MHz
The resultant frequency response plots are shown in
Fig. 5.
2.4. Input and Output Impedance
The input and output impedance of a feedback amplier
can be selectively increased or decreased through the
application of feedback. As has been seen in the previous
sections, general discussions provide great insight
into many of the properties of feedback systems. In order
to consider the design of electronic feedback ampliers,
it is necessary, however, to specify the details of the feed-
back sampling and mixing processes and the circuits nec-
essary to accomplish these operations. The sampling
and mixing processes have a profound effect on the input
impedance, the output impedance, and the definition of
the forward gain quantity that undergoes quantied
change due to the application of feedback. This section
will characterize the various idealized feedback congu-
rations: the following section looks at practical feedback
congurations.
The mixing and the sampling process for a feedback
amplier utilize either voltages or currents. Voltage mix-
ing (subtraction) implies a series connection of voltages at
the input of the amplier; current mixing implies a shunt
connection. Voltage sampling implies a shunt connection
of the sampling probes across the output voltage; current
sampling implies a series connection so that the output
current ows into the sampling network. Either type of
mixing can be combined with either type of sampling.
Thus, a feedback amplier may have one of four possible
combinations of the mixing and sampling processes. These
four combinations are commonly identied by a hyphen-
ated term: mixing topologysampling topology. The four
types are
*
Shuntshunt feedback (current mixing and voltage
sampling)
*
Shuntseries feedback (current mixing and current
sampling)
*
Seriesshunt feedback (voltage mixing and voltage
sampling)
*
Seriesseries feedback (voltage mixing and current
sampling).
These four basic feedback amplier topologies are shown
schematically in Fig. 6. A source and a load resistance
have been attached to model complete operation. In each
diagram the input, feedback, and output quantities are
shown properly as voltages or currents. Forward gain A
must be dened as the ratio of the output sampled quan-
tity divided by the input quantity that undergoes mixing.
30
40
50
60

100 1 k 10 k 100 k 1 M 10 M 100 M
G
a
i
n

(
d
B
)
No feedback
With feedback
3 dB
3 dB
f (Hz)
Figure 5. An example of the effect of feedback on frequency re-
sponse.
R
M
A
I
A
V
f

+
i
o
i
o
i
i i
i
i
f
i
f
i
o
i
o
i
o
i
o
v
o
(a)
f
(b)
f f
v
f
v
i
v
o
v
o
+
v
f
v
i
v
o
+
(c)
G
M
(d)
Figure 6. Feedback amplier topologies: (a) shuntshunt feedback; (b) shuntseries feedback; (c)
seriesshunt feedback; (d) seriesseries feedback.
As such, it is a transresistance, current gain, voltage gain,
or transconductance. The feedback network, as described
by the feedback ratio (f ), must sample the output quantity
and present a quantity to the mixer that is of the same
type (current or voltage) as the input quantity. As such it
is a transconductance, current gain, voltage gain, or trans-
resistance. Table 1 lists the appropriate quantities mixed
at the input, the output sampled quantity, the forward
gain, and the feedback ratio for each of the four feedback
amplier topologies. It is important to remember that the
product fA must be dimensionless and, in the midband
region of operation, positive.
In the previous section, all benets of feedback were
discussed except the modication of inputoutput imped-
ance. The specific definitions of the four feedback ampli-
er topologies allow for that discussion to begin here. The
mixing process alters the input impedance of a negative-
feedback amplier. Heuristically, one can see that
subtraction of a feedback quantity at the mixing junction
increases the input quantity necessary for similar perfor-
mance. Thus, subtracting current (shunt mixing) requires
an increase in overall input current and decreases the in-
put impedance. Similarly, subtracting voltage (series mix-
ing) requires an increase in overall input voltage and
increases input impedance.
Although it can be easily shown that in each case, the
change in resistance is proportional to the return differ-
ence D, it is important to note that the measurement point
of the input and output resistance changes with each cir-
cuit topology. Two amplier types are shown as examples:
a shuntshunt feedback amplier (Fig. 7) and a series
series feedback amplier (Fig. 8). Similarly, the return
difference D is calculated using different gain and feed-
back parameters for each topology. The changes due to
feedback on each feedback amplier topology are summa-
rized in Table 2.
3. PRACTICAL FEEDBACK CONFIGURATIONS
Previous discussions of feedback and feedback congura-
tions have been limited to idealized systems and ampli-
ers. The four idealized feedback schematic diagrams of
Fig. 6 identify the forward-gain amplier and the feedback
network as two-port networks with a very specific prop-
ertyeach is a device with one-way gain. Realistic elec-
tronic feedback ampliers can only approximate that
idealized behavior. In addition, in practical feedback am-
pliers there is always some interaction between the for-
ward-gain amplier and the feedback network. This
interaction most often takes the form of input and output
resistive loading of the forward-gain amplier. The divi-
sion of the practical feedback amplier into its forward-
gain amplier and feedback network is also not always
obvious. These apparent obstacles to using idealized feed-
back analysis can be resolved through the use of two-port
network relationships in the derivation of practical feed-
back amplier properties. Once amplier gain and imped-
ance relationships have been derived, the utility of the
two-port representations becomes minimal and is typically
discarded.
3.1. Identication of the Feedback Topology
Feedback topology is determined through careful observa-
tion of the interconnection of the feedback network and
forward-gain amplier. Shunt mixing occurs at the input
terminal of the amplier. Thus, shunt mixing is identied
by a connection of feedback network and the forward-gain
amplier at the input terminal of rst active device within
the amplier, in one of the following locations:
*
At the base of a bipolar junction transistor (BJT) for a
common-emitter or common-collector rst stage
Table 1. Feedback Amplier Topology Parameters
Parameter ShuntShunt ShuntSeries SeriesShunt SeriesSeries
Input quantity X
i
Current i
s
Current i
s
Voltage v
s
Voltage v
s
Output quantity X
o
Voltage v
o
Current i
o
Voltage v
o
Current i
o
Forward gain A Transresistance R
M
Current gain A
I
Voltage gain A
V
Transconductance G
M
Feedback ratio f i
f
/v
o
i
f
/i
o
v
f
/v
o
v
f
/i
o
f
i
f
i
o
i
i
v
o
+
R
M
R
of
R
if
R
i
R
o
Figure 7. Inputoutput resistance for shuntshunt feedback.
f
i
o
i
o
v
o
v
f
i
v
i
+
G
M
R
of
R
if
R
i
R
o
Figure 8. Inputoutput resistance for seriesseries feedback.
*
At the emitter of a BJT for a common-base rst stage
*
At the gate of a eld-effect transistor (FET) for a
common-source or common-drain rst stage
*
At the source of a FET for a common-gate rst stage
Series mixing occurs in a loop that contains the input
terminal of the forward-gain amplier and the controlling
port of the rst active device. The controlling port of a
BJT in the forward-active region is the baseemitter junc-
tion: a FET in the saturation region is controlled by
the voltage across the gatesource input port. Series
mixing is characterized by a circuit element or net-
work that is both connected to the output and in series
with the input voltage and the input port of the rst active
device.
Identication of the sampling is derived from direct ob-
servation of the connection of the output of the basic for-
ward amplier and the feedback network. Shunt sampling
is typically characterized by a direct connection of the
feedback network to the output node; series sampling im-
plies a series connection of the amplier output, the feed-
back network, and the load. Two tests performed at the
feedback amplier output can aid in the determination of
sampling topology:
*
If the feedback quantity vanishes for a short-circuit
load, the output voltage must be the sampled quan-
tity. Thus zero feedback for a short-circuit load im-
plies shunt sampling.
*
If the feedback quantity vanishes for an open-circuit
load, the output current must be the sampled quan-
tity. Thus zero feedback for a open-circuit load im-
plies series sampling.
After the topological type has been identied, each
amplier must be transformed into a form that allows
for the use of the idealized feedback formulations.
This transformation includes modeling the amplier
and the feedback network with a particular two-port rep-
resentation that facilitates combination of elements.
Once the transformations are accomplished, the ampli-
er performance parameters are easily obtained using
the methods previously outlined. The particular
operations necessary to transform each of the four feed-
back amplier topological types require separate discus-
sion. Only the shuntshunt topology is discussed in
detail; the other three topologies use similar techniques
that lead to the results shown in Fig. 11 and described in
Table 2.
3.1.1. ShuntShunt Feedback: A Detailed Derivation.
Figure 9 is a small-signal model representation of a typi-
cal shuntshunt feedback amplier. In this representation,
the forward-gain amplier and the feedback network have
been replaced by their equivalent y-parameter two-port
network representations so that parallel parameters can
be easily combined. A resistive load has been applied to the
output port; and, since shuntshunt feedback ampliers
are transresistance ampliers, a Norton equivalent source
has been shown as the input. It should also be noted that
Table 2. Feedback Amplier Analysis
Topology
Characteristic ShuntShunt ShuntSeries SeriesShunt SeriesSeries
Input X
i
Current i
s
Current i
s
Voltage v
s
Voltage v
s
Output X
o
Voltage v
o
Current i
o
Voltage v
o
Current i
o
Signal source Norton Norton Thevenin Thevenin
Input circuit
a
Include shunting resistances; set v
o
=0 Include shunting re-
sistances; set i
o
=0
Exclude all shunt re-
sistances; set v
o
=
0
sistances; set i
o
=0
Output circuit
a
Include shunting resistances; set v
i
=0 Exclude all shunt re-
sistances; set v
i
=0
Include shunting re-
sistances; set i
i
=0
sistances; set i
i
=0
Feedback ratio f i
f
/v
o
i
f
/i
o
v
f
/v
o
v
f
/i
o
Forward gain A Transresistance R
M
Current gain A
I
Voltage gain A
V
Transconductance
G
M
Input resistance
b
R
if
=
R
i
1 f R
M
R
if
=
R
i
1 fA
I
R
if
=R
i
1 f A
V
( ) R
if
=R
i
1f G
M
( )
Output resistance
b
R
of
=
R
o
1 f R
M
R
of
=R
o
1 fA
I
( )
R
of
=
R
o
1 f A
V
R
of
=R
o
1 f G
M
( )
a
Inputoutput circuit: These procedures give the basic forward amplier without feedback but including the effects of loading due to the feedback network.
b
Resistance: The resistance modied at shunted ports will include all shunting resistances that were included in the basic forward amplier. The resistance
modied at series ports will include only the resistances included in the basic forward amplier. The true amplier inputoutput impedance must be modied
to reect the point of measurement desired.
the forward-gain parameter of each two-port y
21
is the
transadmittance.
The basic feedback equation for a transresistance am-
plier takes the form
R
Mf
=
R
M
1 R
Mf
(24)
The application of the basic feedback equation to this cir-
cuit in its current form is not immediately clear. It is nec-
essary to transform the feedback amplier circuit into a
form that allows for easy application of the basic feedback
equation, [Eq. (24)]. Such a transformation must meet the
previously stated feedback requirements:
*
The forward-gain amplier is to be a forward trans-
mission system onlyits reverse transmission must
be negligible.
*
The feedback network is to be a reverse transmission
system that presents a feedback current, dependent
on the output voltage, to the amplier input port.
While a mathematically rigorous derivation of the trans-
formation is possible, greater insight to the process comes
with a heuristic approach.
The two-port y-parameter representation, in conjunc-
tion with the shuntshunt connection, is used to describe
the two main elements of this feedback amplier so that
all the input port elements of both two-port networks are
in parallel. Similarly, all output port elements are in par-
allel. It is well known that circuit elements in parallel may
be rearranged and, as long as they remain in parallel, the
circuit continues to function in an identical fashion.
Hence, it is possible, for analysis purposes only, to concep-
tually move elements from one section of the circuit into
another (from the feedback circuit to the amplier circuit
or the reverse). The necessary conceptual changes made
for the transformation are
*
The source resistance, the load resistance, and all in-
put and output admittances, y
11
and y
22
, are placed in
the modied amplier circuit. While inclusion of the
source and load resistance in the amplier seems, at
rst, counterproductive, it is necessary, however, to
include these resistances so that the use of the feed-
back properties produces correct results for input and
output resistance (after appropriate transforma-
tions).
*
All forward transadmittances y
21
(represented by
current sources dependent on the input voltage v
1
)
are placed in the modied amplier circuit.
*
All reverse transadmittances y
12
(represented by cur-
rent sources dependent on the output voltage v
o
) are
placed in the modied feedback circuit.
The dependent current source can be easily combined:
y
t
12
=y
a
12
y
f
12
(25)
y
t
21
=y
a
21
y
f
21
(26)
In virtually every practical feedback amplier the reverse
transadmittance of the forward-gain amplier is much
smaller than that of the feedback network (y
a
12
5y
f
12
) and
the forward transadmittance of the feedback network is
much smaller than that of the forward-gain amplier
(y
f
21
5y
a
21
). Thus, approximate simplications of the ampli-
er representation can be made:
y
t
12
=y
a
12
y
f
12
- y
f
12
(27)
y
t
21
=y
f
21
y
a
21
- y
a
21
(28)
The shuntshunt feedback amplier circuit of Fig. 9 is,
with these changes and approximations, thereby trans-
formed into the circuit shown in Fig. 10.
This transformed circuit is composed of two simple
elements:
*
The original amplier with its input shunted by the
source resistance and the feedback network short-
circuit input admittance y
f
11
and its output shunted
by the load resistance and the feedback network
short-circuit output admittance y
f
22
.
*
A feedback network composed solely of the feedback
network reverse transadmittance y
f
12
.
It is also important to note that the input resistance R
if
of
this circuit includes the source resistance R
s
; as such, it is
Amplifier circuit
Feedback circuit
i
s
i
2
i
i
v
1
v
o
+
y
a
y
a
v
o
y
a
v
1
y
a
y
f
v
o
y
f y
f
R
s
R
L
11
21
12
22
11
22
12
y
f
v
1
21
Figure 9. Two-port realization of a shuntshunt feedback am-
plier.
Approximate f

circuit
Approximate R
M
circuit
i
s
i
f
v
1
v
o
+
y
22
y
12
v
o
y
11
R
s
R
of
R
if
i
o
f
y
11
a
y
21
v
1
a
a
y
22
f
R
L
f
Figure 10. Redistributed shuntshunt realization.
not the same as the input resistance of the true amplier
R
in
. The input resistance of the true amplier can be
obtained as
R
in
=
1
R
if
1
R
s
_ _
1
(29)
Similarly, the output resistance R
of
of this circuit includes
the load resistance R
L
; similar operations may be neces-
sary to obtain the true output resistance of the amplier.
The y parameters of the feedback network can be
obtained
y
f
11
=
i
f
v
1
v
o
=0
; y
f
22
=
i
2
v
0
v
1
=0
; y
f
12
=
i
f
v
0
v
1
=0
(30)
where i
2
is the current entering the output port of the
feedback network (see Fig. 8). With the determination of
these two-port parameters, the circuit has been trans-
formed into a form that is compatible with all previous
discussions. The forward-gain parameter (in this case,
G
M
) of the loaded basic amplier must be calculated, while
the feedback ratio has been determined from the two-port
analysis of the feedback network:
f =y
f
12
(31)
In the case of totally resistive feedback networks, the
shunting resistances can be found in a simple fashion:
*
r
in
=(y
f
11
)
1
is found by setting the output voltage to
zero value v
o
=0 and determining the resistance from
the input port of the feedback network to ground.
*
r
out
=(y
f
22
)
1
is found by setting the input voltage to
zero value v
i
=0 and determining the resistance from
the output port of the feedback network to ground.
The feedback ratio f is simply the ratio of the feedback
current i
f
to the output voltage when the input port of the
feedback network v
i
is set to zero value. All idealized feed-
back methods can be applied to this transformed amplier,
and all previously derived feedback results are valid.
The other three feedback amplier topologies can be
similarly analyzed using various two-port parameters for
analysis:
*
Shuntseries: g parameters
*
Seriesshunt: h parameters
*
Seriesseries: z parameters
Such analysis leads to a characterization of the loading of
the basic forward amplier as is described in Fig. 11. As is
the case with the shuntshunt topology, individual ele-
ments within the feedback network may appear more
than one in the loaded basic forward-amplier equivalent
circuit. Table 2 summarizes the analysis of feedback
amplier properties:
4. STABILITY IN FEEDBACK AMPLIFIERS
Under certain conditions, feedback ampliers have the
possibility of being unstable. This instability stems from
the frequency-dependent nature of the forward gain of the
basic amplier, A and the feedback factor f. The frequency
dependence is exhibited in the changes in magnitude and
phase of the product fA as a function of frequency.
f

f
Short
Short
(a)

f f
Short
Open
(b)

f f
Short
Open
(c) (d)

f f
Open
Open
R
M
A
I
A
V
+
i
o
i
o
i
i
i
i
i
f
i
f
i
o
i
o
i
o
i
o
v
o
v
o
v
f
v
f
v
i
v
o
v
i
v
o
+
G
M
Figure 11. Feedback network load-
ing of a basic forward amplier:
(a) shuntshunt feedback; (b)
shuntseries feedback; (c) series
shunt feedback; (d) seriesseries
feedback.
Instability can be visualized by studying the basic feed-
back equation as a function of frequency:
A
f
jo ( ) =
A jo ( )
1 f jo ( )A jo ( )
(32)
It is important that an amplier be designed so that sta-
bility is present at all frequencies, not only those in the
midband region. If the product f (jo)A(jo) approaches
unity at any frequency, the denominator of Eq. (32) ap-
proaches zero value; the total gain of the amplier ap-
proaches innity. This condition represents an output that
is truncated only by power supply limitations regardless of
input magnitude and is an unstable condition that is in-
tolerable in ampliers. In order to avoid this instability, it
is necessary to avoid a simultaneous approach of
[ f (jo)A(jo) [ =1 and f (jo)A(jo) = 180
. Since each
pole can provide a phase shift ranging from 01 and 901
only, the second condition is possible only for ampliers
that have high or low frequency responses described by
three or more poles. Simultaneously satisfying both con-
ditions can be avoided if the magnitude of fA is always less
than unity when the phase angle of fA is 71801. Designers
of feedback ampliers typically verify that this is the case
through the use of amplier frequency response plots.
4.1. Gain Margin and Phase Margin
A frequency response plot of the loop gain fA for a typical
amplier is shown in Fig. 12. The frequency at which
[ f (jo)A(jo) [ =1 is identied as o
m
, and the frequency at
which f (jo)A(jo) = 180
is identied as o
p
. Since
o
m
ao
p
, it is apparent that this is a stable amplier;
that is, the two instability conditions are not simulta-
neously met. It is, however, important to ensure that the
two conditions are not met simultaneously with a margin
of safety. The margin of safety is dened by the gain mar-
gin and the phase margin of the feedback amplier.
Gain margin is dened as the difference in the loop
gain magnitude (in decibels) between 0dB (unity gain)
and the loop gain magnitude at frequency o
p
:
Gainmargin= 20 log [f (jo
p
)A(jo
p
) [ (33)
Phase margin is the difference between the loop gain
phase angle at frequency o
m
and 1801:
Phase margin=f (jo
m
)A(jo
m
) 180
(34)
Each safety margin is shown in Fig. 12. It is common to
specify the design of feedback ampliers with gain and
phase margins greater than 10dB and 501, respectively.
These margins ensure stable amplier operation over
component parameter variation, temperature change,
and other variations found in typical ampliers.
4.2. Compensation Networks
Two fundamental techniques are available for ensuring
amplier stability:
*
Reducing the midband loop gain fA of the amplier
*
Adding a compensation network to the amplier to
shape the loop gain frequency response so that the
phase and gain margins are positive and in an ac-
ceptable range
Careful design is required in each of these cases to ensure
stable amplier operation over typical performance con-
ditions.
In many cases, decreasing the loop gain to achieve sta-
bility is not an acceptable design possibility. Additionally,
as is often the case in operational amplier circuits, the
feedback quantity f may be determined by the user rather
than the amplier designer and can range widely. In such
cases, compensation networks are added within the feed-
back loop of the amplier to increase the gain and phase
margins. Such compensation networks add poles or a com-
bination of poles and zeros to the loop gain characteristic.
The most commonly used compensation techniques are
*
Dominant pole compensation
*
Laglead (polezero) compensation
*
Lead compensation
Each technique modies the gain and phase proles of the
basic forward amplier through pole and zero manipula-
tion.
In dominant-pole compensation, the amplier is mod-
ied by adding a dominant pole that is much smaller in
magnitude that all other poles in the amplier gain func-
tion; typically it is chosen so that the gain reaches 0 dB at
the frequency of the next pole (the rst pole of the uncom-
pensated amplier). Consequently, the modied loop gain
falls below 0 dB before the nondominant poles shift the
total phase shift near 1801 and the circuit is inherently
stable. Dominant-pole compensation will typically result
in a phase margin of approximately 451.
The location of the new compensation pole can be de-
termined by modeling the loop gain response with a single
pole and setting its value to 0 dB at the rst pole of the
uncompensated amplier:
20 log fA
0
20 log 1
jo
c
o
p1
=0 dB (35)
180
0
Phase margin
Gain margin
20 log |f A(s)|
f A(s)

c
m
c
p
Figure 12. Gain margin and phase margin.
Solving Eq. (35) for the compensation pole frequency o
c
,
results in
o
c
-
o
p1
fA
0
(36)
If the design goals of the feedback amplier includes a
range of feedback ratios, the frequency of the compensa-
tion pole is determined by the maximum value of the feed-
back ratio; that is, o
c
is chosen to be the smallest value
predicted by Eq. (36).
An example of dominant-pole compensation is shown
in Fig. 13 in the frequency domain. For clarity, the
gain plots are represented by straight-line Bode approxi-
mations, while the exact phase plots are retained. The
example amplier is described by a midband gain of
60 dB with poles at 1, 5, and 50 MHz; the feedback ratio
is f =0.1.
The possibility of feedback amplier instability is fo-
cused at the frequency where [fA[ =1 or equivalently
where [ A[
dB
= 20 log ( f . For this particular three-pole
example, instability may occur at o
m
E21 MHz. Here the
phase margin is very small and negative (E71). After
compensation, the focus is again centered where the gain
plot (now compensated) intersects the negated feedback
ratio plot. Addition of the compensation pole o
c
shifts this
intersection to the frequency of the rst uncompensated
pole o
p1
. For this example, the compensation pole is placed
at 10 kHz and yields a phase margin of E431 and a gain
margin of E14 dB.
Dominant-pole compensation reduces the open-loop
bandwidth drastically. In this example the 3-dB frequen-
cy was lowered by two decades to 10 kHz. Still, it is com-
mon in many circuits with inherently large gain;
internally compensated operational ampliers often uti-
lize dominant-pole compensation.
Leadlag (or polezero) compensation is similar to dom-
inant-pole compensation with one major exception. In ad-
dition to a dominant pole, a higher-frequency zero is
added. This zero is used to cancel the rst pole of the un-
compensated amplier. The added dominant pole can then
be chosen so that the gain reaches 0dB at the frequency of
the next pole (the second pole of the uncompensated am-
plier). Leadlag compensation has a distinct bandwidth
advantage over dominant-pole compensation.
The location of the new compensation pole can be de-
termined in a method similar to the method utilized under
dominant-pole compensation with the exception that the
loop gain (without the rst original pole) is to reach 0 dB at
the second pole of the uncompensated amplier. Solving
Eq. (35) for the compensation pole frequency o
c
results in
o
c
-
o
p2
fA
0
(37)
As with dominant-pole compensation, design goals includ-
ing a range of feedback ratios lead to the determination of
frequency of the compensation pole by the maximumvalue
of the feedback ratio; that is, o
c
is chosen to be the small-
est value predicted by Eq. (37).
The uncompensated amplier described previously, is
compensated with a laglead polezero pair, and the fre-
quency domain results are displayed in Fig. 14. The pos-
sibility of feedback amplier instability is again focused at
the intersection of the gain and the negated feedback ratio
plots. After compensation, the focus is centered where the
gain plot (now compensated) intersects the negated feed-
back ratio plot. The addition of the compensation pole o
c
and a zero at the rst uncompensated pole o
p1
shifts this
intersection to the frequency of the second uncompensated
pole o
p2
. The previously identied amplier parameters
lead to a compensation pole at 50 kHz (also the 3-dB fre-
quency), a positive phase margin of E481, and a gain
margin of E20 dB.
Lead compensation can lead to the largest bandwidth of
the three most common compensation networks. Here, as
in leadlag compensation, a pole and a zero are added. The
zero is used to cancel the second pole of the uncompen-
sated amplier, and the added pole is positioned at a fre-
quency higher than the zero. The objective is to reduce the
phase shift of the uncompensated amplier at the fre-
quency where the loop gain reaches 0 dB (o
m
). It can
be extremely effective in feedback ampliers where there
are two or three dominant poles in the uncompensated
amplier.
The uncompensated amplier described above is com-
pensated with a lead polezero pair, and the frequency-
domain results are displayed in Fig. 15. After compensa-
tion, the focus is again centered where the gain plot (now
compensated) intersects the negated feedback ratio plot.
Phase margin
Original gain
Gain after
compensation
20 log f
Original phase
Phase after
compensation
c
p1
c
p2
c
p3
c
c
180
0
Figure 13. Bode diagram for a dominant-pole-compensated am-
plier.
Phase margin
Original gain
Gain after
compensation
20 log f
Original phase
Phase after
compensation
c
p1
c
p2
c
p3
c
c
180
0
Figure 14. Bode diagram for a laglead-compensated amplier.
The addition a zero at the second uncompensated pole o
p2
and a high-frequency compensation pole o
c
, shifts this in-
tersection a frequency beyond the second uncompensated
pole o
p2
. For this example, the high-frequency pole was
chosen at 500MHz. This design choice leads to a positive
phase margin of E351 and a gain margin of E15 dB. Note
that with lead compensation there is no significant reduc-
tion in the frequency response of the feedback amplier;
the 3dB frequency for this example is 999.6 kHz (0.04%
reduction).
A passive component circuit implementation of each of
the three compensation techniques is schematically
shown in Fig. 16. For the circuit of Fig. 16a, the compen-
sation network component values are chosen so that
R
p
R
o
_ _
C
p
=
fA
0
o
p1
(38)
where A
0
R
o
, and o
p1
are the midband gain, the output
resistance, and the rst pole frequency, of the basic for-
ward amplier, respectively. For the circuit of Fig. 16b, the
compensation network component values are chosen so
that
R
b
C
c
=
1
o
p1
(39)
and
R
a
R
b
R
o
( )C
c
=
fA
0
o
p2
(40)
For the circuit of Fig. 16c, the compensation network com-
ponent values are chosen so that
R
a
C
c
=
1
o
p2
(41)
R
a
==R
b
_ _
C
c
5
1
o
p1
(42)
While the placement of a compensation network at the
output of the basic forward amplier is an effective tech-
nique for feedback topologies with shunt sampling, other
placement may be necessary. In particular, connections at
the output of a feedback amplier with series sampling
are not within the feedback loop and are therefore invalid.
In such cases, alternate placement of the compensation is
necessary; one common placement intersperses the com-
pensation network between individual gain stages of the
amplier. Similarly, it is possible to compensate a feed-
back amplier within the feedback network rather than
the basic forward amplier. Unfortunately, analysis of
compensation within the feedback network is often
extremely complex because of the loading of the basic
forward amplier by feedback network components.
BIBLIOGRAPHY
1. F. C. Fitchen, Transistor Circuit Analysis and Design, Van
Nostrand, Princeton, NJ, 1966.
2. M. S. Ghausi, Electronic Devices and Circuits: Discrete and
Integrated, Holt Rinehart and Winston, New York, 1985.
3. P. R. Gray, P. J. Hurst, S. H. Lewis, and R. G. Meyer, Analysis
and Design of Analog Integrated Circuits, 4th ed., J Wiley,
New York, 2001.
A
f
+
R
p
C
p
C
c
C
c
R
a
R
a
R
b
R
b
A
f
+
A
f
(a)
(b)
(c)
Figure 16. Compensation networks: (a) dominant-pole compen-
sation; (b) laglead (polezero) compensation; (c) lead compensa-
tion.
180
0
Original gain
Gain after
compensation
20 log f
Original phase
Phase after
compensation
c
p1
c
p2
c
p3
Phase
margin
Figure 15. Bode diagram for a lead-compensated amplier.
4. D. H. Horrocks, Feedback Circuits and Op. Amps, Chapman &
Hall, London, 1990.
5. P. J. Hurst, A comparison of two approaches to feedback cir-
cuit analysis, IEEE Trans. Ed. 35(3) (1992).
6. J. Millman, Microelectronics, Digital and Analog Circuits and
Systems, McGraw-Hill, New York, 1979.
7. J. Millman and C. C. Halkias, Integrated Electronics: Analog
and Digital Circuits and Systems, McGraw-Hill, New York,
1972.
8. J. W. Nilsson and S. A. Riedel, Electric Circuits, 6th ed., Pren-
tice-Hall, New York, 1999.
9. S. Rosenstark, Feedback Amplier Principles, Macmillan,
New York, 1986.
10. A. S. Sedra and K. C. Smith, Microelectronic Circuits, 4th ed.,
Oxford Univ. Press, New York, 1998.
11. D. L. Schilling and C. Belove, Electronic Circuits, 3rd ed.,
McGraw-Hill, New York, 1989.
12. T. F. Schubert, Jr. and E. M. Kim, Active and Non-Linear
Electronics, J Wiley, New York, 1996.
13. T. F. Schubert, Jr., A heuristic approach to the development of
frequency response characteristics in the design of feedback
ampliers, Proc. 1996 Frontiers in Education Conf.
14. T. F. Schubert, Jr., Feedback ampliers, in J. G. Webster, ed.,
Wiley Encyclopedia of Electrical and Electronics Engineering,
Vol. 7, J Wiley, New York, 1999, pp. 336339.
15. R. Spencer and M. Ghausi, Introduction to Electronic Circuit
Design, Prentice Hall, New York, 2003.
FEEDBACK OSCILLATORS
JIANYI ZHOU
WEI HONG
Southeast University
Nanjing, China
1. INTRODUCTION
An oscillator is a type of circuit that converts DC power
to AC signals automatically without external excitation,
and has been widely used as a key component in various
electronic products, such as communication systems, mea-
surement instruments, and computers. In communication
systems, oscillators are ubiquitous, generating carrier
signals in both the transmitter and the receiver [13]. In
many digital systems, oscillators provide the clock for
CPUs, A/D converters, D/A converters, frequency synthe-
sizers, and other devices.
Most oscillators are sinusoidal, generating a steady-
state sinusoidal signal; some oscillators generate other
types of waveform such as square and triangular waves.
The sinusoidal oscillators are the basic type of oscillator
because many other kinds of waveform can be obtained
with the sinusoidal oscillator and waveshaping circuits.
There are two fundamental concerns in a sinusoidal
oscillatorone is the purity of the output sinusoidal sig-
nal, and the other is the frequency stability. Actually, all
oscillators are not perfect, and the output signal is always
the superposition of the desired sinusoidal signal and
many undesired harmonics and noise. The power con-
tained in the harmonics should be small relative to the
desired fundamental signal. The noise accompanying the
output signal of an oscillator will be the factor that limits
system performance in many communication systems.
There are two kinds of noise: amplitude noise and phase
noise. Because amplitude uctuations are usually greatly
attenuated in many systems, the phase noise generally
dominates [4,5]. The output signal frequency of an oscil-
lator varies inevitably with the parameters of components
in the oscillator because of the temperature variation. In
many applications, especially in communication systems
and electronic clocks, the frequency drift is required to be
as small as possible [6].
In general, either the positive feedback or the negative
resistance in a circuit can generate oscillation. Therefore,
from the physical mechanism of oscillation, there are two
kinds of oscillators: feedback oscillators and negative-
resistance oscillators.
2. PRINCIPLE OF FEEDBACK OSCILLATORS
The basic method of generating sinusoidal oscillation elec-
tronically is to insert a positive-feedback circuit in an am-
plier. The basic topology for a feedback oscillator, shown
in Fig. 1, consists of three fundamental parts: the ampli-
er or active device, the feedback circuit, and the output
load [13].
In Fig. 1, the feedback loop is broken at for anal-
ysis. The input voltage of the amplier is V
i
, and the out-
put voltage of the amplier is
V
o
(jo) =A(jo)V
i
(jo) (1)
where A(jo) is the transfer function or gain of the ampli-
er, o denotes the angular frequency, and j =
1
_
is an
imaginary unit.
The output voltage of the feedback circuit is
V
f
(jo) =b(jo)V
o
(jo) =A(jo)b(jo)V
i
(jo) (2)
and the loop gain is dened as
T(jo) =
V
f
(jo)
V
i
(jo)
=A(jo)b(jo) (3)
For the closed loop, the closed-loop gain is dened as
G
CL
(jo) =
A(jo)
1 A(jo)b(jo)
(4)
R
L
Amplifier
A( jc)
Feedback
[( jc)

V
i
V
o
V
f
Figure 1. Block diagram of a feedback oscillator.
1432 FEEDBACK OSCILLATORS
Because there is no input signal for an oscillator (i.e.,
V
i
=0) to maintain a nonzero output voltage, the denom-
inator of the closed-loop gain (4) should be equal to zero.
This is the Barkhausen criterion [13,7,8]
T(jo
0
) =A(jo
0
)b(jo
0
) =T(o
0
)e
jj
T
(o0)
=1 (5)
where o
0
is the oscillation frequency.
Note that the Barkhausen criterion really implies two
conditions for oscillation:
1. The magnitude of the loop gain must be 1, which
termed the magnitude-balanced condition.
2. The phase of the loop gain must be 0 (or an integral
multiple of 2p), which is termed the phase-balanced
condition.
The Barkhausen criterion requires the magnitude of
the loop gain T(o
0
) to be exactly 1 at the oscillation fre-
quency. If T(o
0
)o1, the oscillation will be decaying. On
the other hand, if T(o
0
)41, the oscillation amplitude will
continue to increase until it is limited by the nonlinearity
of the active devices in the amplier. In fact, this nonlin-
earity is an essential feature of practical oscillators.
Suppose that the oscillator satises the Barkhausen
criterion when V
i
=V
iA
other than V
i
=0, i.e., the magni-
tude of the loop gain T(o
0
) equates to 1. As the circuit
characteristics drift, T(o
0
) may be either smaller or larger
than 1. For the former case, the oscillation will decay to
stop; for the latter case it increases until limited by the
nonlinearity of the amplier. Therefore, in order to ini-
tially establish an oscillation, T(o
0
) must be larger than 1.
The phase-balanced condition and the startup condition
for a feedback oscillator are as follows [1,7,8]:
T(o
0
)[
t =t
0
> 1 (6a)
j
T
(o
0
)[
t =t
0
=2np (n=0; 1; 2; . . .) (6b)
Moreover, in order to maintain a steady oscillation,
nearby V
iA
, T(o
0
) must decrease when V
i
increases; in
other words, the derivative of T(o
0
) at V
i
=V
iA
must be
negative. This yields the magnitude stability condition of
the feedback oscillator as [7,8]
@T(o
0
)
@V
i
V
i
=V
iA
o0 (7)
Note that, although an oscillator satises the magni-
tude stability condition, its amplitude of the output volt-
age may uctuate with the change of its environment. To
suppress such amplitude uctuations, besides minimizing
the environment uctuations and decreasing the sensitiv-
ity of devices in the oscillator, one important point is to
increase the slope of T(o
0
) at the balanced point, that is, to
increase the absolute value of
@T(o
0
)
@V
i
V
i
=V
iA
Hence, a small drift of V

i
will cause a large change of
T(o
0
). Because of the negative slope of T(o
0
), the drift of V
i
is countered and a new balanced condition will be estab-
lished very close to V
iA
.
Suppose that the oscillator satises the phase-balanced
condition when o=o
0
, that is, j
T
(o
0
) =0. As the circuit
characteristics drift, j
T
(o
0
) may be either larger or small-
er than 0. In the former case, the phase of the feedback
voltage leads the phase of the original input voltage of the
amplier. In the latter case, the phase of the feedback
voltage lags behind the phase of the original input voltage
of the amplier. It is known that the frequency is the de-
rivative of the phase. If j
T
(o
0
)40, the phase of the feed-
back voltage leads continuously, so the corresponding
frequency will be higher than the original oscillation fre-
quency. However, if j
T
(o
0
)o0, the phase of the feedback
voltage lags behind continuously, and the corresponding
frequency will be lower than the original oscillation fre-
quency.
In fact, j
T
(o) varies with the frequency. If j
T
(o) de-
creases with an increase of o [i.e., if j
T
(o) and o are in-
versely proportional], and if the oscillation frequency is
higher than the original oscillation frequency, then j
T
(o)
decreases; therefore, the frequency increase will be pre-
vented. However, if the oscillation frequency is lower,
j
T
(o) increases, and thus the frequency decrease will be
prevented. Thus, a new balanced condition will be estab-
lished very close to the original frequency. On the other
hand, if j
T
(o) increases with an increase of o, the drift of
the oscillation frequency is accelerated and cannot satisfy
the phase-balanced condition.
Hence, in order to maintain a steady oscillation, nearby
o
0
, j
T
(o) must decrease with the increase of o; thus, the
derivative of o
T
(o) at o=o
0
must be negative. This is the
phase stability condition of the feedback oscillator [7,8]:
@o
T
(o)
@o
o=o
0
o0 (8)
Note that although an oscillator satises the phase sta-
bility condition, the frequency of its output voltage may
uctuate with the change in its environment. To suppress
such frequency uctuations, besides to minimizing the en-
vironment uctuations and decreasing the sensitivity of
devices in the oscillator, one important point is to increase
the slope of o
T
(o) at the balanced point, that is, to increase
the absolute value of
@o
T
(o)
@o
o=o
0
Hence, a small drift of o will cause a large change of

o
T
(o). Because of the negative slope of o
T
(o), the drift of o
is countered and a new balanced condition will be estab-
lished very close to o
0
[13,68].
In a practical feedback oscillator, the amplier is an
active device such as an operational amplier, a eld-
effected transistor (FET), a bipolar junction transistor
(BJT), or a logical gate. The feedback system is generally
a passive resonant network. Any network can be used as
FEEDBACK OSCILLATORS 1433
the feedback network as long as the Barkhausen criterion
is satised.
Practical feedback oscillators can be realized with dif-
ferent active devices and feedback network.
3. LC OSCILLATORS
LC oscillators are widely used in modern communication
systems as the local oscillator for the upconverters and
downconverters. LC oscillators use passive inductors and
capacitors as the feedback network. Usually, the LC oscil-
lators use a transistor or differential pairs as the active
device. Many LC oscillator circuits use a p-type impedance
network as the feedback network [13,59]. The circuit
schematic is shown in Fig. 2.
The forward gain is
A=
g
m
R
0
Z
L
Z
L
R
0
(9)
where
Z
L
=
Z
2
(Z
1
Z
3
)
Z
1
Z
2
Z
3
(10)
The feedback factor is
b =
Z
1
Z
1
Z
3
(11)
Therefore, the loop gain is
T=Ab =
g
m
R
0
Z
1
Z
2
Z
1
Z
2
Z
2
Z
3
R
0
(Z
1
Z
2
Z
3
)
(12)
Suppose that the impedances are purely reactive (either
inductive or capacitive), specifically, Z
i
=jX
i
(i =1,2,3); we
then have
T=
g
m
R
0
X
1
X
2
X
2
(X
1
X
3
) jR
0
(X
1
X
2
X
3
)
(13)
According to the Barkhausen criterion, the phase of the
loop gain should be equal to zero; thus we have
X
1
X
2
X
3
=0 (14)
T=
g
m
R
0
X
1
X
2
(15)
According to the Barkhausen criterion, the magnitude
of the loop gain should be equal to 1. Therefore, X
1
and X
2
must have the same sign; specifically, they must have the
same kind of reactance, either inductive or capacitive. It
follows from (14) that X
3
= (X
1
X
2
) must possess the
other type of reactance. If X
1
and X
2
are capacitors and X
3
is an inductor, the circuit is called a Colpitts oscillator; if
X
1
and X
2
are inductors and X
3
is a capacitor, the circuit is
called as Hartley oscillator. Other combinations are also
used. For example, if X
1
and X
2
are capacitors and X
3
is a
series combination of an inductor and a capacitor, the cir-
cuit is called a Clapp oscillator. Figure 3 shows the sche-
matic of these typical oscillators.
4. FREQUENCY STABILITY
Frequency stability is one of the most important perfor-
mance factors for an oscillator. The feedback oscillator in-
volves an active amplier with a passive resonant
Z1 Z2
Z3
Z1 Z2
Z3
Ro
Vi Vo gmVi
Q
(a) (b)
Figure 2. Schematic of p-type network feedback oscillator:
(a) p-type network feedback oscillator; (b) simplied equivalent
circuit.
Q
C1
C2
L
RL
Rb1
Rb2
Cc
Cb
Re
VCC
Q
C
Cb
L
RL
Rb1
Rb2
Cc
Re
VCC
L
Ce
Q
C1
C2
L
Rc Rb1
Rb2
C3
Cb
Re
VCC
(a) (b) (c)
Figure 3. Schematics of typical LC oscillators:
(a) Colpitts oscillator; (b) Hartley oscillator; (c)
Clapp oscillator.
feedback network. The oscillation frequency o
0
is deter-
mined mainly by the resonant of the feedback network.
For an LC oscillator, suppose that the drift of the inductor
and capacitor is DL and DC, respectively. The offset of the
resonant frequency is [7]
Do -
1
2
o
0
DL
L

DC
V
_ _
(16)
Hence, one means to improve the frequency stability is to
reduce the drift of the inductor and capacitor in the res-
onant network.
Besides the parameter drift, the effective quality factor
Q
e
is also very important for frequency stability.
As we discussed before, increasing the absolute value of
@o
T
(o)
@o
o=o
0
will also enhance the frequency stability. Thus, we may

dene the frequency stability factor as [7]
SF=o
0
@o
T
(o)
@o
o=o
0
(17)
For a parallel RLC circuit, the impedance is
Z(jo) =
jo
1
C
o
2
0
o
2
j
o
0
o
Q
e
(18)
where o
0
=1=
LC
_
and Q
e
=R
C=L
_
.
The phase angle of the impedance as a function of fre-
quency is
j(o) =
1
2
p arctan
o
0
o
Q
e
o
2
0
o
2
(19)
and its derivative with respect to o is
@j
@o
=
Q
e
o
0
(o
2
0
o
2
)
(o
0
o)
2
Q
2
e
(o
2
0
o
2
)
(20)
At the resonant frequency, this yields
@j
@o
=
2Q
e
o
0
(21)
and then the frequency stability factor
SF= 2Q
e
(22)
This result implies that the higher the Q value of the res-
onant circuit, the higher the frequency stability of the os-
cillator. Hence, another way to improve the frequency
stability is to increase the effective quality factor of the
resonant circuit [68].
5. PHASE NOISE
Phase noise is another important performance factor of an
oscillator. Phase noise is usually characterized in terms of
the signal sideband noise spectral density [4,5,1016],
namely, the decibels below the carrier per hertz (dBc/
Hz), and is dened as
L
total
(Do) =10 log
P
sideband
(o
0
Do; 1 Hz)
P
carrier
_ _
(23)
where P
sideband
(o
0
Do, 1Hz) represents the signal side-
band power within a measurement bandwidth of 1 Hz at
the frequency with offset Do from the carrier.
The most used approach to calculate the phase noise of
an oscillator is the extended LeesonCutler phase noise
model expressed as [4,5]
L(Do)
=10 log
2FkT
P
sig
1
o
0
2Q
e
Do
_ _
2
_ _
1
o
c
[Do[
_ _
_ _
(24)
where F denotes the noise gure of the active device, k is
Boltzmanns constant, T is the absolute temperature, P
sig
is
the average power dissipated in the resistive part of the
tank, o
0
is the oscillation frequency, Q
e
is the effective qual-
ity factor of the tank with all the loadings, Do is the offset
from the carrier, and o
c
is the icker frequency of the active
device. A typical plot of the phase noise is shown in Fig. 4.
6. OTHER KINDS OF FEEDBACK OSCILLATORS
6.1. Microwave Oscillators
The LC oscillators described above are usually used in the
low-frequency band. As the frequency increases to the mi-
crowave band and even millimeter-wave band, the lumped
components are no longer suitable for resonation. Distrib-
uted components, such as transmission lines and cavities,
should be adopted. Another important consideration is the
treatment of the active devices in microwave band. The
junction capacitance, parasitic capacitances, and induc-
tances of microwave active devices should be taken into
account. Generally, the S parameters are often employed
to characterize the behavior of a microwave active device;
L(c)
c
1
f
3
1
f
2

Figure 4. Typical plot of the phase noise of an oscillator versus
offset from the carrier.
thus the design of microwave oscillators based on S pa-
rameters is widely adopted.
The oscillation condition of the circuit shown in Fig. 5 is
given by [2,3]
Ko1
S
/
11
G
1
=1
S
/
22
G
2
=1
(25)
where
K =
1 [S
11
[
2
[S
22
[
2
[S
11
S
22
S
12
S
21
[
2
2[S
12
S
21
[
S
/
11
=S
11
S
12
S
21
G
2
1 S
22
G
2
S
/
22
=S
22
S
12
S
21
G
1
1 S
11
G
1
G
1
=
Z
S
Z
0
Z
S
Z
0
G
2
=
Z
L
Z
0
Z
L
Z
0
In many cases, the S parameters of a microwave tran-
sistor do not satisfy the oscillation condition; some type of
external feedback network is needed. The usual method is
to add a segment of transmission line between the grid of
the FET or the base of the BJT and the ground as a serial
feedback network to aid the oscillation. A typical micro-
wave oscillator with microstrip is shown in Fig. 6, where
TL means microstrip transmission line [2]. TL2 acts as the
serial feedback network. Usually, the length of TL2 dom-
inates the oscillation frequency.
A so-called substrate integrated waveguide (SIW) cav-
ity [43] has been used as the feedback resonant network,
followed by development of a feedback microwave oscilla-
tor [44]. Figure 7 is a owchart of the SIW feedback oscil-
lator (for further details, please see Ref. 44).
The oscillator operates at 12.02 GHz. Good phase noise
and harmonic level have been obtained.
6.2. Crystal and Dielectric Oscillators
As component characteristics change with age, tempera-
ture, signal level, and other parameters, the oscillation
frequency will drift. The main cause of frequency drift is
the drift of circuit parameters and the Q factor. The crys-
tals used in oscillators are usually quartz, and the quartz
crystal has extremely high Q and mechanical stability. A
crystal oscillator is often used in extremely high stability
oscillators. Figure 8 shows two typical crystal oscillators
[1719].
Besides the quartz crystals, other kinds of materials,
such as SAW devices [20,21], Sapphire [2225], and di-
electric resonators [2633] are often used for high-perfor-
mance oscillators. In the microwave band, the dielectric
resonator oscillator (DRO) has become one of the most
widely used oscillators. A typical microwave DRO is
shown in Fig. 9 [33].
6.3. RC Oscillators
RC oscillators using RC feedback network are useful os-
cillators in low-frequency bands [68]. The Wien bridge
oscillator is the most widely used RC oscillator that adopts
a balanced bridge as the feedback network, as shown in
Fig. 10.
The oscillation frequency is determined by R and C as
f
0
=
1
2pRC
(26)
The oscillation will be sustained if
R
2
R
1
=2 (27)
6.4. Ring Oscillators
Ring oscillators are commonly used in integrated circuits
[3438]. They consist simply of n inverters in a ring, where
n is odd. A simple schematic of the ring oscillator is shown
in Fig. 11.
[s]
Z
L
Z
S
Figure 5. Design of microwave oscillators based on S para-
meters.
VCC
TL2
TL1
TL3 TL6
TL4
Output
TL5
Figure 6. An FET microwave oscillator with microstrip as dis-
tributed feedback network.
SIW
Cavity
Amp.
Positive feedback
Output
Figure 7. An SIW feedback oscillator.
For a simple analysis of the ring oscillator, it is as-
sumed that each inverter can be characterized by a prop-
agation time delay T
pd
. The oscillation period is then
simply twice the total propagation time delay:
f
0
=
1
2n T
pd
(28)
6.5. Comparison of Different Oscillator Types
Table 1 compares the different types of feedback oscillators
described above.
7. VOLTAGE-CONTROLLED OSCILLATORS
Many applications require oscillators to be electrically
tunable. For example, in the superheterodyne receiver,
the local oscillator should be tuned over an appropriate
frequency range so that the mixer will convert the incom-
ing RF signal with different frequency down to the IF sig-
nal with the same intermediate frequency. In many
applications, such as direct FM or in phase-locked loops,
the tuning of the oscillator should be automatic. One way
to achieve this purpose is using a voltage-controlled oscil-
lator (VCO) [13,68,3942]. A device that can be used in a
VCO is the varactor diode. Any diode is a p-n junction, and
thus has a junction capacitance. A varactor diode is de-
signed so that the junction capacitance can be controlled
by the reverse bias voltage across the junction
C(V) =C
0
1
V
V
d
_ _
(1=2)
(29)
Q
C1
C2
Lc Rb1
Rb2
Cc
Cb
Re
(a) (b)
VCC
RL
C1 C2
Figure 8. Two typical crystal oscillators:
(a) Pierce crystal oscillator; (b) crystal oscilla-
tor with a NOT gate.
DR
E
E
B C
Output
VCC
Figure 9. Microwave dielectric resonator oscillator (DRO).
R2
R
R1
R
C
C
+
_
Figure 10. Wien bridge oscillator.
Output
Figure 11. Ring oscillator.
where V is the reverse bias, C
0
is a constant, and V
d
is the
diffusion barrier voltage of the junction [13,68].
A typical schematic of the voltage-controlled oscillator
is shown in Fig. 12.
8. CONCLUSION
The principles and typical practical circuits of feedback
oscillators are reviewed. Three important conditions and
two kinds of main specications required in an oscillator
are summarized. Typical practical oscillators, such as the
LC oscillators, RC oscillators, crystal oscillators, micro-
wave oscillators, and voltage-controlled oscillators, are
illustrated. These circuits are widely used in communica-
tion systems and other electronic equipment.
BIBLIOGRAPHY
1. K. K. Clarke and D. T. Hess, Communication Circuits: Anal-
ysis and Design, Addison-Wesley, Reading, MA, 1971.
2. R. Ludwig and P. Bretchko, RF Circuit Design: Theory and
Applications, Prentice-Hall, 2000.
3. R. E. Collin, Foundations for Microwave Engineering, 2nd ed.,
McGraw-Hill, 1992.
4. D. B. Leeson, A simple model of feedback oscillator noise
spectrum, Proc. IEEE 54:329330 (Feb. 1966).
5. T. H. Lee and A. Hajimiri, Oscillator phase noise: A tutorial,
IEEE J. Solid-State Circ. 35:326336 (March 2000).
6. A. S. Sedra and K. C. Smith, Micro-electronic Circuits, 4th ed.,
Oxford Univ. Press, 1998.
7. X. Jiakui and X. Yuqing, Electronic Circuits, 3rd ed., High
Education Press, Beijing, 1988.
8. J. R. Smith, Modern Communication Circuits, 2nd ed.,
McGraw-Hill, 1998.
9. G. M. Maggio, O. De Feo, and M. P. Kennedy, Nonlinear anal-
ysis of the Colpitts oscillator and applications to design, IEEE
Trans. Circ. Syst. I 46(9):11181130 (Sept. 1999).
10. A. N. Riddle and R. J. Trew, A novel GaAs FEToscillator with
low phase noise, 1985 IEEE MTT-S Int. Microwave Symp.
Digest 85.1, 1985, pp. 257260.
11. C. P. Lusher and W. N. Hardt, Effects of gain compression,
bias conditions, and temperature on the icker phase noise of
an 8.5GHz GaAs MESFETamplier (short papers), IEEE
Trans. Microwave Theory Tech. 37:643646 (March 1989).
12. G. K. Montress, T. E. Parker, and M. J. Loboda, Residual
phase noise measurements of VHF, UHF, and microwave
components, IEEE Trans. Ultrason., Ferroelectr. Freq. Con-
trol 48(5):664679 (Sept. 1994).
13. V. Gungerich, J. Deuring, W. Anzil, and P. Russer, Phase noise
reduction of microwave oscillators by optimization of the dy-
namic behaviour, 1994 IEEE MTT-S Int. Microwave Symp.
Digest, 1994, Vol. 2, pp. 953936.
14. V. Gungerich, F. Zinkler, W. Anzill, and P. Russer, Noise cal-
culations, and experimental results of varactor tunable oscil-
lators with significantly reduced phase noise, IEEE Trans.
Microwave Theory Tech. 43:278285 (Feb. 1995).
15. M. N. Tutt, D. Pavlidis, A. Khatibzadeh, and B. Bay-
raktaroglu, The role of baseband noise and its upconversion
Table 1. Comparison of Feedback Oscillators
Type Operation Frequency Stability Phase Noise Cost Tuning
LC Low frequencyUHF Moderate Moderate Low Easy
Microwave Microwave band Moderate Moderate Moderate Hard
Crystal Low frequencyVHF Good Very low Moderate Moderate
Dielectric Microwave band Good Low High Very hard
RC Low-frequency band Poor Poor Low Easy
Ring Low frequencyUHF Poor Poor Low Hard
Q
C1
C2
L
RL
Rb1
Cc
Rb2
Cb
Re
VCC
C3
D
Rd
Vtune
Vout
Figure 12. Typical schematic of a voltage-controlled
oscillator.
in HBToscillator phase noise, IEEE Trans. Microwave Theory
Tech. 43:14611471 (July 1995).
16. X. Zhang, D. Sturzebecher, and A. S. Daryoush, Comparison
of the phase noise performance of HEMT and HBT based os-
cillators, 1995 IEEE MTT-S Int. Microwave Symp. Digest,
1995, Vol. 2, pp. 697700.
17. M. E. Frerking, Crystal Oscillator Design and Temperature
Compensations, Van Nostrand Reinhold, New York, 1978.
18. D. V. Bogomolov and E. A. Silaev, Nonlinear dependencies
and phase noise in precision crystal oscillator, Proc. 2000
IEEE/EIA Int. Frequency Control Symp. and Exhibition,
2000, pp. 549552.
19. Y. Shmaliy, S. Nedorezov, and R. Besson, Statistical model of
an anharmonic sensor of precision crystal oscillator, Proc.
2001 IEEE Int. Frequency Control Symp. and PDA Exhibi-
tion, 2001, pp. 758766.
20. D. P. Klemer, K.-M. Shih, and E. E. Clark, III, Phase noise
charcterization of SAW oscillators based on a Newton mini-
mization procedure, IEEE Trans. Microwave Theory Tech.
39:883889 (May 1991).
21. G. K. Montress, T. E. Parker, and D. Andres, Review of SAW
oscillator performance, Proc. 1994 IEEE Ultrasonics Symp.,
1994, Vol. 1, pp. 4354.
22. M. E. Tobar, A. J. Giles, S. Edwards, and J. Searls, High-Q TE
stabilized sapphire microwave resonators for low noise appli-
cations, Proc. 1993 IEEE Int. Frequency Control Symp., 1993,
pp. 749756.
23. M. E. Tobar, E. N. Ivanov, R. A. Woode, and J. H. Searls, Low
noise microwave oscillators based on high-Q temperature sta-
bilized sapphire resonators, Proc. 1994 IEEE Int. Frequency
Control Symp., 1994, pp. 433440.
24. C. Wong, W. F. Miccioli, D. M. Insanna, and C. A. Drubin, An
extremely low noise, phase lockable, sapphire loaded cavity
based microwave oscillator, 1999 IEEE MTT-S Int. Microwave
Symp. Digest, 1999, Vol. 1, pp. 9194.
25. O. Llopis, G. Gibiel, Y. Kersale, M. Regis, M. Chaubet, and
V. Giordano, Ultra low phase noise sapphire-SiGe HBT oscil-
lator, IEEE Microwave Wireless Compon. Lett. 12:157159
(May 2002).
26. S. J. Ha, Y. D. Lee, Y. H. Kim, J. J. Choi, and U. S. Hong,
Dielectric resonator oscillator with balanced low noise ampli-
er, IEE Electron. Lett. 38(24):15421544 (Nov. 2002).
27. J. K. Plourde and C.-L. Ren, Application of dielectric resona-
tors in microwave components, IEEE Trans. Microwave The-
ory Tech. 29:754770 (Aug. 1981).
28. K. K. Agarwal, Dielectric resonator oscillators using GaAs/
(Ga, A1) As heterojunction bipolar transistors, 1986 IEEE
MTT-S Int. Microwave Symp. Digest, 1986, Vol. 1, pp. 9598.
29. M. J. Loboda, T. E. Parker, and G. K. Montress, Frequency
stability of L-band, two-port dielectric resonator oscillators,
1987 IEEE MTT-S Int. Microwave Symp. Digest, 1987, Vol. 2,
pp. 859862.
30. J. Caldinhas Vaz, J. L. Caceres, J. Freire Costa, and J. Perez,
Design and performance of dielectric resonator oscillators
with series feedback, Proc. 1989 Electrotechnical Conf. 1989,
pp. 339342.
31. R. Jones and V. Estrick, Low phase noise dielectric resonator
oscillator, Proc. 44th Annual Symp. Frequency Control, 1990,
pp. 549554.
32. S. Qi, K. Wu, and Z. Ou, Hybrid integrated HEMT oscillator
with a multiple-ring nonradiative dielectric (NRD) resonator
feedback circuit, IEEE Trans. Microwave Theory Tech.
46:15521558 (Oct. 1998).
33. J. K. Plourde and C.-L. Ren, Application of dielectric resona-
tors in microwave components, IEEE Trans. Microwave The-
ory Tech. 29:754770 (Aug. 1981).
34. T. H. Lee, The Design of CMOS Radio-Frequency Integrated
Circuits, Cambridge Univ. Press, 1998.
35. D.-Y. Jeong, S.-H. Chai, W.-C. Song, and G.-H. Cho,
CMOScurrent-controlled oscillators using multiple-feed-
back-loop ring architectures, 1997 IEEE Int. Solid-State Cir-
cuits Conf. Digest Tech. Papers, 1997, pp. 386387, 491.
36. H. Kato, A dynamic formulation of ring oscillator as solitary-
wave propagator, IEEE Trans. Circ. Syst. I Fund. Theory
Appl. 45:98101 (Jan. 1998).
37. S. L. J. Gierkink, A. van der Wel, G. Hoogzaad, E. A. M.
Klumperink, and A. J. M. van Tuijl, Reduction of the 1/f noise
induced phase noise in a CMOSring oscillator by increasing
the amplitude of oscillation, Proc. 1998 IEEE Int. Symp. Cir-
cuits and Systems, 1998, Vol. 1, pp. 185188.
38. L. Sun, T. Kwasniewski, and K. Iniewski, A quadrature out-
put voltage controlled ring oscillator based on three-stage
sub-feedback loops, Proc. 1999 IEEE Int. Symp. Circuits and
Systems, 1999, Vol. 2, pp. 176179.
39. A. Gruhle, A. Schuppen, U. Konig, U. Erben, and H. Schu-
macher, Monolithic 26GHz and 40GHz VCOs with SiGe he-
terojunction bipolar transistor, Proc. 1995 Int. Electron
Devices Meeting, 1995, pp. 725728.
40. H.-C. Chang and R. A. York, Enhanced MESFETVCOinjec-
tion-locking bandwidth using low frequency feedback tech-
niques, 1996 IEEE MTT-S Int. Microwave Symp. Digest, 1996,
Vol. 3, pp. 15151518.
41. H. J. Siweris, H. Tischer, T. Grave, and W. Kellner, A mono-
lithic W-band HEMT VCO with feedback topology, 1999 IEEE
MTT-S Int. Microwave Symp. Digest, 1999, Vol. I, pp. 1720.
42. H. J. Siweris, H. Tischer, and E. Rohrer, Monolithic 38GHz
coplanar feedback VCOs fabricated by a production PHEMT
technology, 2001 IEEE MTT-S Microwave Symp. Digest, 2001,
Vol. 3, pp. 15551558.
43. Y.-T. Lee, J.-S. Lim, D. A. Park, and S. Nam, A novel phase
noise reduction technique in oscillators using defected ground
structure, IEEE Microwave Wireless Compon. Lett. 12:3941
(Feb. 2002).
44. Y. Cassivi and K. Wu, Low cost microwave oscillator using
substrate integrated waveguide cavity, IEEE Microwave
Wireless Compon. Lett. 13(2):4850 (Feb. 2003).
FEEDFORWARD AMPLIFIERS
P. GARDNER
The University of Birmingham
Edgbaston, Birmingham
United Kingdom
1. INTRODUCTION
Highly linear transmitters are required in modern RFand
microwave radiocommunication systems, to facilitate the
transmission of multiple carriers and/or nonconstant en-
velope, bandwidth efcient modulated signals, while min-
imizing the intermodulation and spectral spreading
distortion caused by nonlinearity, particularly in the
FEEDFORWARD AMPLIFIERS 1439
power amplier. Approaches to this problem may be di-
vided broadly into
*
Amplier linearization techniques, whereby the RF/
microwave transmitter is linearized as a separate
component
*
Linear transmitter architectures, in which the mod-
ulation and amplication processes are merged, with
feedback or other control applied to the whole trans-
mitter, usually by demodulating a sample of the out-
put, and comparing with the modulation input
The three main techniques for linearization of the ampli-
er itself are feedforward linearization, RF predistortion,
and feedback linearization.
Feedforward, the subject of this article, uses two analog
signal loops. The rst extracts the distortion occurring in
the amplier, and the second subtracts distortion from the
output signal. One of the main attractions of this technique
is that the correcting signal is derived from the nonlinear
amplier itself; the correction circuit does not need to be
programmed with a nonlinear model of the amplier. High
levels of distortion cancellation (Z30dB) can be achieved,
and broad bandwidths are possible, at the expense of very
precise phase and amplitude matching in the two loops.
In RF predistortion, a circuit module at the amplier
input is designed to have a nonlinear characteristic oppo-
site to that of the main amplier. For example, if the main
amplier has a gain that compresses as the input level in-
creases, the predistorter is designed to have a compensat-
ing gain expansion. Similarly, the variation of phase with
signal level in the predistorter is designed to compensate
the phase change with signal level in the amplier. An al-
ternative view of the operation of a predistorter is that it
generates intermodulation distortion components with an
amplitude and phase such that, after amplication in the
main amplier, they cancel out the corresponding compo-
nents generated within the amplier itself. Predistortion
linearisers provide an excellent method for achieving mod-
est levels of distortion reduction (B10dB), without the ad-
ditional circuit complexity of feedforward systems.
Amplier linearization using negative RF feedback re-
lies on the fact that the transfer function of a closed-loop
system is dominated by the feedback path, so long as the
open-loop gain is high compared to the closed-loop gain.
Thus, if the feedback path is linear, the closed-loop gain
and phase remain linear. The main difculty of this tech-
nique at RF and microwave frequencies is that the feed-
back path necessarily has a significant electrical length,
compared to a guided wavelength, so the linearization
works only over a narrow bandwidth, and instabilities are
likely to occur at other frequencies. Negative feedback is
therefore not widely used as a linearization technique in
the RFmicrowave range.
Linear transmitter architectures include a number of
congurations, such as polar loop, Cartesian loop, and
adaptive digital predistortion. These techniques can pro-
vide extremely high levels of linearity, usually measured
in terms of low adjacent-channel power levels. However,
because much of the associated processing is done digital-
ly at baseband or IF, the achievable percentage bandwidth
is relatively low. Since the modulation scheme is effective-
ly programmed into the transmitter architecture, such
linearization techniques could be regarded as less adapt-
able than a standalone amplier lineariser. Continual im-
provements in DSP (digital signal processing) hardware,
however, will make linear architectures increasingly at-
tractive options in system design.
2. BASIC PRINCIPLE OF FEEDFORWARD OPERATION
The invention of feedforward linearization can be traced
back to a patent by Black [1]. After long neglect, the idea
was revived for microwave applications by Seidel[2]. The
basic principle of the feedforward amplier is illustrated in
Fig. 1. At the input to the main amplier, coupler 1 samples
the undistorted input signal. A sample of the signal at the
output of the main amplier is also taken using coupler 2. If
the main amplier were perfectly linear, this would be sim-
ply an amplied version of the input sample. In reality,
nonlinear distortions in the amplier cause the output sam-
ple to be an amplied version of the input signal along with
an additional error signal. The error signal includes the
intermodulation components generated in the main ampli-
er. The function of the feedforward systemis rst to isolate
the error signal and then to remove it from the amplier
output, leaving an undistorted, linearly amplied signal.
This is achieved by rst linearly adjusting the levels of
the two samples and combining them in antiphase (i.e.,
subtracting them) using coupler 3, to cancel out the want-
ed signal and leave only the error signal. The coupling
ratios in couplers 1, 2, and 3, and the phase shifts, delays,
and attenuations in the rst loop, are adjusted to achieve
this cancellation. The error signal is then adjusted in
phase and amplied by the auxiliary amplier to a suit-
able amplitude level, so that when recombined, using cou-
pler 4, with the output from the main amplier, the error
signal is canceled completely from the nal output. To
illustrate this operation, Fig. 1 includes spectrum dia-
grams showing the operation of a feedforward linearised
amplier with a two-tone input signal.
In principle, therefore, feedforward is a way to achieve
perfect linear amplication without distortion. To assess
its applicability in a practical transmitter system, several
key performance parameters must be assessed, including
*
Efciency
*
Bandwidth
*
Degree of distortion suppression required
*
Size and complexity
These parameters and the factors that inuence them are
outlined in this section and in more detail in Section 3.
2.1. Efciency
Three main and interacting factors limit the overall ef-
ciency of a feedforward amplier:
1. The auxiliary amplier consumes a significant
level of DC power without adding to the level of the
1440 FEEDFORWARD AMPLIFIERS
wanted output signal. If the rst loop of the feedfor-
ward system is correctly balanced, the auxiliary am-
plier only has to amplify the distortion. However, it
has to amplify it linearly, and to a level somewhat
higher than the distortion in the main branch of the
system, because of the coupling ratio of coupler 4.
For example, if the main amplier produces inter-
modulation products 20 dB below the output carri-
ers, and the output coupler has a 13dB coupling
ratio, the error signal at the output of the auxiliary
amplier must be 13 dB above the level of the main
amplier intermodulation products, or in other
words, only 7 dB below the level of the wanted sig-
nal in the main branch. For perfect operation, the
auxiliary amplier must not introduce any signi-
cant level of further distortion. A class A amplier,
backed off well below saturation, is therefore nor-
mally used. In the example above, therefore, the
auxiliary amplier would need to have a DC power
rating of the same order of magnitude as that of the
main amplier. Thus, to a crude approximation, the
cancellation of distortion is achieved at the expense
of doubling the DC power consumption or halving
the efciency.
2. The output coupler (coupler 4 in Fig. 1) degrades the
overall efciency because of its inherent ohmic and
dielectric losses, and also because it couples a por-
tion of the wanted output power into the matched
termination on its fourth port. Clearly there is a
tradeoff here between loose coupling in coupler 4
(say, 420 dB), to limit the wastage of main amplier
power, and tight coupling (say, o10 dB) to limit the
demands on the auxiliary amplier. The ideal level
of coupling, for maximum overall efciency, there-
fore depends on the level of distortion occurring in
the main amplier that needs to be canceled.
3. The efciency of the main amplier strongly inu-
ences the overall efciency. To limit the demands on
the auxiliary amplier, it is desirable not to operate
the main amplier too close to saturation; however,
this reduces its efciency.
The net result of these three factors inuencing efciency
is that the overall efciency of a typical feedforward
amplier is at best about 10%. However, to achieve
comparable levels of intermodulation suppression without
feedforward, by simply using a higher power linear
amplier and backing off its output power, could
reduce the efciency by several further orders of magni-
tude. Thus feedforward ampliers can be a good choice
where linear amplication is required with reasonable
efciency.
2.2. Bandwidth and the Degree of Distortion Suppression
In feedforward ampliers, the main and auxiliary signals
both propagate in the same forward direction through the
system. This is in contrast to the situation in feedback
ampliers, where, self-evidently, the feedback signal prop-
agates in the reverse direction. This leads to two advan-
tages of feedforward systems as compared to feedback
systems: (1) because there is no closed loop (assuming ide-
al coupler directivity), we do not have the danger inherent
in feedback systems where the feedback becomes positive
Main amplifier
Error Amplifier
Coupler 1
Coupler 2
Coupler 3
Coupler 4
Delay Line 1
Delay Line 2
Attenuator
Input
Output
Two tone spectrum at amplifier input
Spectrum containing only
intermodulation products at error
amplifier input and output
Main amplifier output with
intermodulation products
Feedforward amplifier output
with intermodulation products
suppressed
Figure 1. Basic conguration for a feedforward amplier, including indicative spectra for the case
of a two-tone input.
at certain frequencies, leading to oscillation problems; and
(2) by ensuring that the propagation delays are equal on
both sides of each loop of the feedforward system, we can
maintain signal cancellation in the rst loop and suppres-
sion cancellation in the second, over a broad frequency
range.
In practice, the accuracy of the two cancellation pro-
cesses depends on the bandwidth over which the designer
attempts to achieve them. Near-perfect cancellation, lead-
ing to extremely low intermodulation levels, is possible in
a very narrowband design. In more realistic scenarios,
there is a tradeoff between the bandwidth and the degree
of suppression achievable. This is made more difcult by
the fact that the delays through the active and passive
components of the system are not frequency-independent.
The practical tradeoff between bandwidth and suppres-
sion level depends on the complexity of the balancing and
adaptation circuits used; there is no exact rule to relate
the parameters. However, the following examples give an
idea of what is achievable. A feedforward amplier oper-
ating over greater than an octave bandwidth was reported
by Steel et al. [3], with an implied suppression of 14 dB. A
feedforward bandwidth of the order of 8% was reported by
Konstantinou et al. [4], with intermodulation product sup-
pression of around 20 dB. A review article by Raab et al.
[5] quotes 2530dB as the practical upper limit for man-
ufactured equipment.
2.3. Size and Complexity
The addition of an auxiliary amplier and the associated
couplers, attenuators, and phase shift networks clearly
adds to the overall size and complexity of the amplier
subsystem. Achieving a sufciently accurate balance over
the full bandwidth requires ne adjustment of phase and
amplitude. Fine tuning in production can be achieved by
means of mechanically trimmed variable attenuators and
phase shifters, or by electronically controlled components
such as vector modulators. Where electronic adjustment is
incorporated, it is possible to introduce electronic control
to compensate for drift with time and temperature. Feed-
forward ampliers reported to date have been hybrid sys-
tems, too complex to allow monolithic integration, given
the need for ne adjustment. Mechanical and thermal is-
sues and the establishment of tuning procedures are cru-
cial in the design of a complete feedforward amplier for a
production environment. In many circumstances, it is also
necessary to devise control and adaptation algorithms to
maintain optimum performance, compensating for param-
eter drifts that occur with aging, temperature, or varia-
tions in signal-level statistics.
3. DETAILED DESIGN CONSIDERATIONS
In this section, design equations for the basic feedforward
system are established, using a revised and expanded ver-
sion of the analysis given by Pothecary [6], leading to
phase and amplitude margins for accurate cancellation.
Further feedforward amplier parameters, such as the
overall input and output match, noise gure, and efciency
are also analyzed.
3.1. Basic Signal Flow Analysis
In the generic form of the feedforward amplier shown in
Fig. 2, the lowercase symbols indicate the following com-
plex voltage transfer coefcients:
*
c
1
through c
4
are coupling coefcients of the couplers.
*
a
1
through a
4
are coupler losses, that is, the transfer
coefcients from the input port to the direct output
port of each coupler.
*
i
1
through i
4
are the isolation coefcients of the four
couplers, that is, the transfer coefcients from the in-
put to the nominally isolated port.
*
l
1
and l
3
are the transfer coefcients of the two delay
lines.
*
l
2
is the attenuation of the attenuator mounted
between couplers 2 and 3.
*
g
m
and g
e
are the complex gains of the main and error
amplier, respectively.
Main amplifier
Error Amplifier
Coupler 1
Coupler 2
Coupler 3
Coupler 4
Delay Line 1
Delay Line 2
Attenuator
Input
Output
c
1
i
1
a
1
c
2 i
2
a
2
c
4
i
4
a
4
i
3
a
3
c
3
l
2
l
1
g
e
g
m l
3
Gain and Phase
Adjustment
Figure 2. Feedforward amplier annotated for signal ow analysis.
Note that although the transfer coefcients are indicated
using arrows showing the main signal ow direction, the
passive components are all reciprocal, and at some points
in the analysis it is necessary to use the same transfer
coefcients for a signal in the opposite direction. The same
symbols in uppercase are used later to indicate the corre-
sponding quantities expressed in decibels.
It is assumed that both ampliers have adjustable gain
and phase to facilitate balancing of the loops. In practice,
this adjustability would be built into the early amplier
stages, rather than the output stages and so would be
achieved without significantly affecting the output power
ratings or DC power consumption.
It is also assumed that the individual passive compo-
nents are well matched to the system impedance (typically
50 O), so that the complications of multiple reections can
be ignored.
The input voltage wave amplitude is V
in
. The distortion
and noise in the main amplier are represented by the
addition of a voltage V
d
to the amplied input. Thus the
total output of the main amplier, at the input to coupler
2, is given by
V
out main
=g
m
c
1
V
in
V
d
The resulting voltage at the input to the error amplier is
given by
V
in error
= g
m
c
1
V
in
V
d
( )c
2
l
2
c
3
a
1
l
1
a
3
V
in
= g
m
c
1
c
2
l
2
c
3
a
1
l
1
a
3
( )V
in
c
2
l
2
c
3
V
d
(1)
The condition for balancing the rst loop is that the input
to the error amplier should depend only on the distortion
voltage V
d
. The condition is therefore found by setting the
coefcient of V
in
in (1) to zero:
g
m
c
1
c
2
l
2
c
3
a
1
l
1
a
3
=0 (2)
The output from coupler 4 (i.e., the output of the whole
feedforward amplier) is given by
V
out
=V
out main
a
2
l
3
a
4
g
e
c
4
V
in error
=a
2
l
3
a
4
g
m
c
1
V
in
V
d
( )
g
e
c
4
g
m
c
1
c
2
l
2
c
3
a
1
l
1
a
3
( )V
in
c
2
l
2
c
3
V
d
( )
If the rst loop matching condition (2) is met, then this
reduces to
V
out
=a
2
l
3
a
4
g
m
c
1
V
in
g
e
c
4
c
2
l
2
c
3
a
2
l
3
a
4
( )V
d
(3)
Balancing of the second loop is achieved when the coef-
cient of V
d
in this equation is zero, that is, when
g
e
c
4
c
2
l
2
c
3
a
2
l
3
a
4
=0 (4)
Then, substituting (4) into (3), we can express the gain of
the whole feedforward amplier as
g
ff
=
V
out
V
in
=g
m
c
1
a
2
l
3
a
4
(5)
If we represent the transfer coefcients in decibels instead
of in complex amplitude format, using the corresponding
upper case symbols, then
G
ff
=G
m
C
1
A
2
L
3
A
4
(6)
As expected, the overall gain is the gain of the main am-
plier reduced by the losses of the passive components in
the main signal path.
3.2. First-Loop Balance
Balancing the rst loop is necessary to minimize the load
on the error amplier. Any distortion resulting from over-
loading the error amplier with residual carrier power
adds to the overall distortion of the feedforward ampli-
erthere is no mechanism to cancel it out. The amplied
residual carrier signal reaching coupler 4 through the er-
ror amplier would also alter the overall gain of the feed-
forward amplier. This could be an increase or a decrease
depending on the phase of the residual signal. The balance
conditions for the rst loop need to be achieved over a fre-
quency range that encompasses the input signal spec-
trum.
3.3. Second-Loop Balance
Balancing the second loop is necessary to ensure deep
cancellation of distortion. The bandwidth over which bal-
ance must be achieved is normally greater than that re-
quired for the rst loop because distortion products (e.g.,
intermodulation products) are likely to fall outside the
range of the wanted signal.
3.4. Phase and Amplitude Matching Limits
Considering further the balance condition for the second
loop, we can rewrite (4) as
1 x =0; where x =
c
2
l
2
c
3
g
e
c
4
a
2
l
3
a
4
(7)
where x is a complex number, equal to unity for perfect
balance and equal to zero if the second-loop correction is
not applied. The squared magnitude of the distortion sup-
pression ratio is thus [s[
2
=[1 x[
2
.
Analyzing this in terms of the magnitude and phase
imbalance, we obtain
[s[
2
=[1 x[
2
=[1 [x[ cos y j[x[ sin y[
2
=(1 [x[ cos y)
2
[x[
2
sin
2
y
=1 2[x[ cos y [x[
2
(8)
where y is the phase imbalance and 20 log
10
[x[ is the am-
plitude imbalance in decibels.
Figure 3 shows plots of the loop suppression in decibels
versus phase imbalance, for varying levels of amplitude
imbalance. It shows that there is some limited scope for
trading off phase error against amplitude error, but for
high levels of suppression, very good phase and amplitude
accuracy are both required. For example, 30 dB suppres-
sion can be achieved if the phase balance is better than
1.21 and the amplitude balance is better than 0.2dB. To
illustrate the ne adjustment required, in physical layout
terms this 1.21 phase balance corresponds to a pathlength
balance between the two sides of the loop of better than
1=f

e
eff
_
mm, where f is the frequency in gigahertz and e
eff
is the effective dielectric constant of the circuit medium.
3.5. Noise Figure
The noise gure of a feedforward amplier is dominated
by the error amplier path. If the loops are properly bal-
anced, the noise generated in the main amplier (being
uncorrelated with the input signal) is canceled at the out-
put of the fourth coupler, in just the same way as the dis-
tortion. Thus, with reference to Fig. 2, the total noise
gure f
total
is given by
f
total
=
f
e
[a
1
l
1
a
3
[
2
where f
e
is the noise gure of the error amplier. Alter-
natively, expressing all quantities in decibels, we obtain
F
total
=F
e
A
1
L
1
A
3
(9)
Since A
1
, L
1
, and A
3
are all negative, the total noise gure
is clearly somewhat higher than that of the error ampli-
er. L
1
may be regarded as a parasitic loss and should be
fairly small (of the order of a few tenths of a decibel). A
3
can also be made small in most cases because other de-
grees of freedom are available to balance the loops. So the
main design choice affecting the noise gure is the input
coupler loss A
1
. In a feedforward amplier design in which
the noise gure must be kept to a minimum, it is therefore
necessary to direct most of the input power toward the
linear branch of the rst loop by making a
1
close to unity
(i.e., A
1
close to 0 dB). Consequently, as power is conserved
in the input coupler, c
1
is small, and the gain of the main
amplier must be increased to compensate for this input
attenuation. This is rarely a problem because small-signal
gain is relatively cheap in terms of components and DC
power.
3.6. Return Losses
The input and output return losses of the feedforward
amplier are affected by the multiple signal paths.
3.6.1. Input Return Loss. Considering rst the input
port, with reference to Fig. 2, the reected voltage wave
is made up of the vector combination of reections from
couplers 1 and 3, the two ampliers, and the rst delay
line. There is also a potentially significant contribution
from the main amplier output, fed back to the input via
coupler 2 and the isolated path through coupler 3. It is
assumed that the reverse isolations (output to input)
through the ampliers and the isolation factors of the cou-
plers are sufciently high to eliminate any significant con-
tributions to the input return loss from all other possible
signal paths.
Assuming further that the individual passive compo-
nents have negligible reection coefcients, we can write
the input reection coefcient as follows:
G
in
=c
2
1
G
in main
a
2
1
l
2
1
a
2
3
G
in error
c
1
g
m
c
2
l
2
c
3
G
in error
a
3
l
1
a
1
c
1
g
m
c
2
l
2
i
3
l
1
a
1
=c
2
1
G
in main
a
1
l
1
a
3
(a
1
l
1
a
3
c
1
g
m
c
2
l
2
c
3
)G
in error
c
1
g
m
c
2
l
2
i
3
l
1
a
1
(10)
If the rst-loop balance condition (2) is met, the coefcient
of G
in_error
in (10) is zero, and the equation can be simpli-
ed to
G
in
=c
2
1
G
in main
a
2
1
l
2
1
a
3
i
3
c
3
(11)
The rst contribution will be small if the input coupling
ratio is made low, as also required for a low-noise gure
(see Section 3.5). The second contribution in (11) can be
minimized by designing coupler 3 to have a high directiv-
ity factor, d
3
=c
3
/i
3
.
3.6.2. Output Return Loss. Considering now a signal in-
jected into the output port of the feedforward amplier,
and making corresponding assumptions again about
the reverse isolation factors of the ampliers and the re-
ection coefcients of the passive components, we can ar-
rive at the following equation for the output reection
Loop Suppression vs. Phase Error
Amplitude Error:
0 dB
0.2 dB
0.4 dB
0.6 dB
0.8 dB
1 dB
2 dB
0
10
20
30
40
Phase Error, Degrees
S
u
p
p
r
e
s
s
i
o
n
,

d
B
0 2 4 6 8 10
Figure 3. Loop suppression versus phase error.
coefcient:
G
out
=c
2
4
G
out error
a
2
4
l
2
3
a
2
2
G
out main
a
4
l
3
a
2
G
out main
c
2
l
2
c
3
g
e
c
4
a
4
l
3
i
2
l
2
c
3
g
3
a
4
=c
2
4
G
out error
a
4
l
3
a
2
(a
4
l
3
a
2
c
2
l
2
c
3
g
e
c
4
)G
out main
a
4
l
3
i
2
l
2
c
3
g
3
a
4
(12)
If the second-loop balancing condition (4) is met, the coef-
cient of G
out_main
is zero. This is a result of the fact that
the second loop fullls its function of canceling any output
signals not originating from the input. The signal reected
by the output of the main amplier is canceled just as
noise and distortion are. Thus Eq. (12) can be rearranged
to give
G
out
=c
2
4
G
out error
a
2
4
l
2
3
a
2
i
2
c
2
(13)
c
4
is typically small, to avoid wasting main amplier
power (but not so small as to overstretch the error ampli-
er), so the second term, involving the directivity of cou-
pler 2, namely, d
2
=c
2
/i
2
, could be the dominating
contribution to the output return loss.
3.7. Stability
Although feedforward is generally regarded as an inher-
ently stable amplier conguration, it does contain closed
signal loops that need to be analyzed carefully to check for
possible instabilities. As in the case of the input and out-
put return losses, a key consideration here is again the
directivity of the couplers.
Considering the rst loop rst, we can express a loop
gain factor g
1
as
g
1
=i
1
g
m
c
2
l
2
i
3
l
1
(14)
If the rst loop balance condition (2) is met, this can be
written as:
g
1
=
a
1
l
2
1
a
3
d
1
d
3
(15)
Since |d|b1 for a well-designed directional coupler, and
since all the parameters in the numerator have magni-
tudes less than unity, Eq. (15) suggests that stability is
unlikely to be a problem, since |g
1
|o1. However, care
must be taken if the bandwidth of the main amplier is
significantly broader than the bandwidths of the couplers,
since their isolation may degrade out of band, leading to
an unstable condition where Eq. (14) gives |g
1
|41, at
some frequency outside the intended frequency range of
the feedforward amplier.
Similarly, considering the second loop, we can express a
loop gain factor g
2
as follows:
g
2
=g
e
i
4
l
3
i
2
l
2
c
3
(16)
Applying the second-loop balance condition (4) allows this
to be rewritten as
g
2
=
a
2
l
2
3
a
4
d
2
d
4
(17)
As in the case of the rst loop, good directivity in the
couplers ensures stability over the intended operating
range, but if the error amplier has a bandwidth broad-
er than those of the couplers or than the intended operat-
ing frequency range of the feedforward amplier, it is
advisable to check that (16) does not lead to |g
2
|41 at
any frequency.
3.8. Efciency
The overall efciency of a feedforward amplier (not in-
cluding power consumption by power conditioning, control
circuits and cooling systems) is a function of the efcien-
cies of both the main and error ampliers, as well as the
losses in the output circuit.
If the output power of the main amplier is P
M
, then
the output from the feedforward amplier is
P
OUT
=[a
2
[
2
[l
3
[
2
[a
4
[
2
P
M
. If the efciencies of the main
and error ampliers are N
M
and N
E
, then the total DC
power consumption is P
DC TOTAL
=P
M
=N
M
P
E
=N
E
,
where P
E
is the power output of the error amplier.
Thus the overall efciency can be written as follows:
N
TOTAL
=
P
OUT
P
DC TOTAL
=
[a
2
[
2
[l
3
[
2
[a
4
[
2
P
M
P
M
=N
M
P
E
=N
E
=
[a
2
[
2
[l
3
[
2
[a
4
[
2
N
M
1 P
E
N
M
=P
M
N
E
(18)
This shows that in the limiting case where the power de-
mands on the error amplier are very small compared to
those on the main amplier, the overall efciency ap-
proaches that of the main amplier, but reduced by the
losses of the output couplers and delay lines. In the more
typical situation where the powers are comparable, the
overall efciency is reduced by the presence of the error
amplier. Typically, the error amplier is likely to have an
efciency lower than that of the main amplier because it
needs to be highly linear to avoid generating additional
uncompensated distortion, and it may also have been de-
signed with an emphasis on noise gure rather than ef-
ciency.
Figure 4 shows a set of curves indicating the efciency
reduction factor, 1=(1 P
E
N
M
=P
M
N
E
), as a function of P
E
/
P
M
for typical values of N
E
and N
M
. They show, as expect-
ed, that the efciency is halved if P
E
=P
M
and N
E
=N
M
.
3.9. Bandwidth
As discussed earlier, the main limitation on the bandwidth
of a feedforward amplier is usually the frequency range
over which the rigorous phase and amplitude matching
conditions can be met.
Considering Figs. 2 and 3 again, the problem becomes
one of meeting amplitude and phase matching limits re-
quired for the desired degree of suppression over the full
bandwidth. The two paths from the main amplier output
to the feedforward output have transfer functions t
a
and
t
b
, where t
a
=a
2
l
3
a
4
and t
b
=c
2
l
2
c
3
g
e
c
4
.
Typically, the main path characterized by t
a
has a rel-
atively small variation of amplitude with frequency and a
phase that varies linearly with frequency, consistent with
a constant delay. Some amplitude variation and deviation
from constant delay may occur if the couplers are working
near the edge of their frequency ranges. The phase of
transfer function t
a
can thus be expected to vary with fre-
quency as follows:
Arg(t
a
) =f
a
t
d
(o o
0
) (19)
where t
d
is the time delay at center frequency o
0
.
The path through the error amplier, characterized by
transfer function t
b
, is likely to have more complex ampli-
tude and phase variations with frequency. The phase vari-
ation is typically characterized by a polynomial, of the
form
Arg(t
b
) =f
b
t
g
(o o
0
) p(o o
0
)
2
R(o) (20)
where t
g
is the group delay at center frequency o
0
, p is a
parabolic phase coefcient, arising mainly from the band-
pass response of the error amplier, and R(o) is a residual
phase ripple function, caused by higher-order transfer
function terms and multiple reections within the error
amplier and associated components.
The second-loop balance equation requires that
Arg(t
a
) Arg(t
b
) = p
From (19) and (20), it is clear that this can be satised by
making f
a
f
b
= p andt
d
=t
g
, as long as the amplier
bandwidth is large enough to render the inuence of the
parabolic and ripple components negligible. Thus, to
achieve broadband feedforward cancellation, it is neces-
sary to have at least 2 degrees of freedom in the adjustable
parameters of the second loop, so that both the phase offset
and the delay can be equalized. It may be necessary to
arrange for cancellation of quadratic and higher terms
also, necessitating the inclusion of carefully designed all-
pass lter networks. However, if only narrowband opera-
tion is required, it is sufcient to have just one adjustable
phase parameter in each loop. Similar considerations
apply to achieving amplitude balance.
In a more realistic design scenario, it is likely that a
CAD model rather than an analytical equation would be
used to characterize the amplier and the associated pas-
sive components. A procedure such as that reported by
Konstantinou et al. [7] can then be used. First, the loop 1
balance is optimized by injecting a small signal into the
input of coupler 1 and adjusting the variable elements to
achieve minimum power at the input to the error ampli-
er, over the required range of frequency and input power.
The reference path is then broken by removing the delay
line and replacing it with two matched and isolated ter-
minations. The carrier signal should then appear at the
input to the error amplier. If the second loop is correctly
adjusted, this will cancel the carrier signal at the output of
coupler 4. Thus the second loop can be optimized by ad-
justing the variable elements for minimum output from
0
0.2
0.4
0.6
0.8
1
20 15 10 5 0
P
e
/P
m
in dB
E
f
f
i
c
i
e
n
c
y

r
e
d
u
c
t
i
o
n

f
a
c
t
o
r
N
E
= 30%
N
E
= 20%
N
E
= 10%
0
E
f
f
i
c
i
e
n
c
y

r
e
d
u
c
t
i
o
n

f
a
c
t
o
r
0.2
0.4
0.6
0.8
1
20 15 10 5 0
P
e
/P
m
in dB
N
E
= 30%
N
E
= 20%
N
E
= 10%
(a)
(b)
(c)
0
E
f
f
i
c
i
e
n
c
y

r
e
d
u
c
t
i
o
n

f
a
c
t
o
r
0.2
0.4
0.6
0.8
1
20 15 5 10 0
P
e
/P
m
in dB
N
E
= 20%
N
E
= 10%
N
E
= 30%
Figure 4. Feedforward efciency reduction with N
M
=30% (a),
20% (b), and 10% (c).
coupler 4, while the input signal varies over the full fre-
quency range. Then, after loop 1 is reestablished, a non-
linear measurement or simulation, such as a two-tone
test, can be performed to ensure that adequate cancella-
tion occurs.
An alternative to the technique of opening the rst loop
while optimizing the second would be to inject a pilot
signal at the input to the main amplier, as illustrated in
Fig. 5 [6]. The rst loop is optimized by ensuring that the
ratio of the input signal to the pilot signal at the input to
the error amplier is minimized. The second loop is opti-
mized by ensuring that the level of the pilot signal seen at
the feedforward amplier output is minimized. This con-
cept forms one basis for automatic control of a feedforward
loop, to compensate for changes in the amplier perfor-
mance with temperature, aging, and varying signal sta-
tistics.
4. ADVANCED TECHNIQUES AND CHALLENGES
This section briefly reviews some advanced techniques in
recent and current (as of 2004) research on feedforward
linearisation.
4.1. Dual-Loop Feedforward
Since a complete feedforward amplier is itself an ampli-
er in its own right, it is possible to incorporate the whole
thing, recursively, inside an outer feedforward loop, as
shown in Fig. 6. The outer loop can be used to compensate
for the residual distortion left by the inner loop and for
any distortion introduced by the rst error amplier, lead-
ing to even lower distortion levels but increasing the tun-
ing complexity even further.
4.2. Nonlinear Phase Equalizers for Improved
Broadband Suppression
Hau et al. [8] demonstrated that by the use of nonlinear
phase equalizers formed from allpass networks, the IM3
cancellation of a feedforward amplier operating over the
full 1.71.9-GHz band could be improved from 15 to 21 dB.
4.3. Feedforward Adaptation and Control
Cavers [9] established a theoretical basis for the adapta-
tion of a feedforward system, showing that different con-
siderations apply in the two loops. In the rst loop (the
signal cancellation loop) adaptation can be achieved by
monitoring the error signal (at the input to the error am-
plier). One possible method is to seek minimum power
in this signal, but the favored method seeks zero corre-
lation between the error signal and the input modulation,
leading to a gradient-based adaptation algorithm that
does not require pilot signals or other deliberate periodic
perturbations to ensure that the optimum setting is
maintained. Adaptation times of the order of 110 ms
can be achieved in this way. In the second loop (the
error correction loop) the high level of the wanted signal
Input
Output
Signal
Processing Pilot Signal
Generator
Vector
Demodulator
Figure 5. Feedforward amplier with pilot signal injection for loop 2 control.
Error Amplifier 2
Coupler 1
Coupler 6
Coupler 7
Coupler 8
Delay Line 3
Delay Line 4
Attenuator 2
Input
Output
Main Amplifier
Error Amplifier 1
Coupler 2
Coupler 3
Coupler 4
Coupler 5
Delay Line 1
Delay Line 2
Attenuator 1
Figure 6. A dual-loop feedforward amplier.
relative to the error necessitates much longer integration
times, leading to much slower adaptation. This problem
can be overcome, and the accuracy improved, by using a
carrier suppression lter before performing the gradient
calculation.
Chen et al. [10] proposed an alternative fast adaptive
algorithm to track the linearizer control parameters, sep-
arating the rapidly varying but known factors (i.e., the
signal statistics) from the slowly varying environmental
factors affecting the power amplier model, resulting in
swift and accurate convergence.
Echeverria et al. [11] demonstrated that very high lev-
els of intermodulation suppression, over 60dB, could be
achieved by tuning the control loops manually for individ-
ual 10 MHz subbands, over a total bandwidth of 19 MHz
(2.012.205 GHz).
4.4. Hybrid Feedforward Ampliers
It is possible to combine feedforward linearization with
other linearization methods for improved overall results.
For example, Horiguchi et al. [12] demonstrated a high-
power 2.12 GHz feedforward power amplier in which the
overall operating efciency was improved by 1% by the
addition of a simple predistortion linearizer to the input of
the main amplier.
Another more radical hybrid feedforward amplier con-
cept has been proposed by Randall et al. [13], in which
part of the rst loop is replaced by a DSP implementation.
The main and reference signals are both generated by
DSP before upconversion to the carrier frequency. The
second loop would operate conventionally. The advantage
of DSP implementation is that phase and amplitude
equalization of both the main and reference signals can
be carried out and adapted as required in DSP software,
rather than in RF analog components, allowing more ac-
curate cancellation.
4.5. MMIC Integration
Integrating the whole feedforward amplier onto a single
MMIC is an attractive proposition from the point of view
of miniaturization and repeatability at higher microwave
frequencies or even millimeter-wave frequencies. Achiev-
ing such integration is, however, a major research chal-
lenge, because of the difculty of creating the necessary
high-isolation couplers, low-loss delay lines, and phase
shifters on a MMIC, along with high linearity, high ef-
ciency, and accurately modeled ampliers. Parkinson and
Paul [14] have carried out initial studies of the possibility
of using a distributed amplier structure as an active cou-
pler within a MMIC feedforward amplier, as a starting
point for full MMIC integration.
BIBLIOGRAPHY
1. H. S. Black, Translating System, U.S. Patent 1,686,792 (1928).
2. H. Seidel, A microwave feedforward experiment, Bell Syst.
Tech. J. 50:28792916 (1971).
3. V. Steel, D. Scott, and S. Ludvik, A 618GHz high dynamic
range MMICamplier using a feedforward technique, IEEE
MTT-S Microwave Symp. Digest, 1990, Vol. 2, pp. 911914.
4. K. Konstantinou and D. K. Paul, Analysis and design of
broadband, high efciency feedforward ampliers, IEEE
MTT-S Microwave Symp. Digest, 1996, Vol. 2, pp. 867870.
5. F. H. Raab, P. Asbeck, S. Cripps, P. B. Kenington, Z. B. Popo-
vic, N. Pothecary, J. Sevic, and N. O. Sokal, Power ampliers
and transmitters for RFand microwave, IEEE Trans. Micro-
wave Theory Tech. MTT-50(3):814826 (2002).
6. N. Pothecary, Feedforward Linear Power Ampliers, Artech
House, Norwood, MA, 1999.
7. K. Konstantinou, P. Gardner, and D. K. Paul, Optimisation
method for feedforward linearisation of power ampliers,
Electron. Lett. 29(18):16331635 (1993).
8. Y. K. Hau, V. Postoyalko, and J. R. Richardson, Design and
characterization of a microwave feed-forward amplier with
improved wide-band distortion cancellation, IEEE Microwave
Theory Tech. MTT-49(1):200203 (2001).
9. J. K. Cavers, Adaptation behavior of a feedforward amplier
linearizer, IEEE Trans. Veh. Technol. 44(1):3140 (1995).
10. J. Chen, H. Tsai, and Y. Chen, Fast adaptive wide-band power
amplier feed-forward linearizer, IEEE Vehicular Technology
Conf., 1998, pp. 21672171.
11. A. Echeverr ia, L. Fan, S. Kanamaluru, and K. Chang, Fre-
quency tunable feedforward amplier for PCS applications,
Microwave Opt. Technol. Lett. 23(4):218221 (1999).
12. K. Horiguchi, M. Nakayama, Y. Sakai, K. Totani, H. Senda,
Y. Ikeda, and O. Ishida, A high efciency feedforward ampli-
er with a series diode linearizer for cellular base stations,
IEEE MTT-S Microwave Symp. Digest, 2001, pp. 797800.
13. R. G. Randall, J. G. McRory, and R. H. Johnston, Broadband
DSP based feedforward amplier lineariser, Electron. Lett.
38(23):14701471 (2002).
14. G. Parkinson and D. K. Paul, Non-reciprocal couplers using
distributed amplier topology, Microwave Opt. Technol. Lett.
38(5):366368 (2003).
FERRITE CIRCULATORS
E. K. N. YUNG
D. X. WANG
City University of Hong Kong
Kowloon, Hong Kong, P.R. China
R. S. CHEN
Nanjing University of Science
and Technology
Nanjing, P.R. China
1. INTRODUCTION
The circulator is one of the elementary building blocks in
radiofrequency and microwave circuits. It is used exten-
sively in making basic devices for communications and
radar systems. The latter are often used in broadband and
high-power rating systems at microwave and millimeter-
wave frequencies. It is, however, seldom found in consum-
er products because some of its functions can be performed
1448 FERRITE CIRCULATORS
Next Page

06a

Transféré par

Informations du document

Description originale:

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

06a

Transféré par

Droits d'auteur :

Formats disponibles

FABRYPEROT RESONATORS

Figure 12. FDTD computational domain for the modeling of

The significant properties are

0:03 = D=1 Af _30

Figure 4. High 3-dB frequency as a function

Hence, a small drift of V

Hence, a small drift of o will cause a large change of

will also enhance the frequency stability. Thus, we may

Vous aimerez peut-être aussi