See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/3428680

Fast synthesis of coupled-resonator filters

Article in IEEE Microwave and Wireless Components Letters · May 2004

DOI: 10.1109/LMWC.2004.827111 · Source: IEEE Xplore

CITATIONS READS
44 423

3 authors, including:

A. Lamecki
Gdansk University of Technology
110 PUBLICATIONS 1,390 CITATIONS

SEE PROFILE

All content following this page was uploaded by A. Lamecki on 14 December 2012.

The user has requested enhancement of the downloaded file.

174 IEEE MICROWAVE AND WIRELESS COMPONENTS LETTERS, VOL. 14, NO. 4, APRIL 2004

Fast Synthesis of Coupled-Resonator Filters

Adam Lamecki, Piotr Kozakowski, Member, IEEE, and Michał Mrozowski, Senior Member, IEEE

Abstract—This letter proposes a new quickly converging tech- of the coupling matrix and its principal sub-matrices. Also the
nique for the synthesis of coupling matrix of cross-coupled res- initial coupling matrix is synthesized in a tridiagonal form by
onator filters. The method formulates the synthesis as a nonlinear solving the Jacobi inverse eigenvalue problem [11].
least squares problem involving the zeroes and poles of short circuit
input and output admittances, as well as eigenvalues of the coupling
matrix and its principal sub-matrices. The solution is found via II. OPTIMIZATION PROBLEM
a damped Levenberg-Marquardt method with analytically com-
puted gradients. Numerical tests show that coupling matrices for To derive the cost function for the optimization procedure,
high order pseudo-elliptical filters can be found in surprisingly few we use network model with frequency independent couplings
optimization steps. considered in [8]. The starting point for our analysis are the short
Index Terms—Coupling matrix, eigenvalue problem, microwave circuit parameters of a two port network. Let us denote by
filters. the reciprocal coupling matrix. The short circuit output
admittance is related to the coupling matrix via

I. INTRODUCTION

M ETHODS of synthesis of cross-coupled resonator filters

are extensively studied for filter design and tuning. One
of the fundamental methods, that was proposed by Atia and
where is the angular frequency, is the identity matrix,
the upper principal sub-matrix obtained by deleting the last row
is

Williams [1], [2] three decades ago, is still commonly used. Al- and column of the matrix and is identity
ternative synthesis techniques were advanced by Cameron [8], matrix. From the above equation it is seen that poles of are
[9] and other researchers [3], [4]. eigenvalues of matrix while zeros of are eigenvalues of
The theory of Atia and Williams uses the orthonormaliza- matrix , all multiplied by . A similar relation holds for the
tion technique to obtain the general coupling matrix with all short circuit input admittance except that zeros are related
possible cross couplings present and requires application of the to the eigenvalues of —the lower principal sub-matrix ob-
similarity transformations to reduce the number of nonzero el- tained by deleting the first row and column of matrix . The
ements in order to obtain a suitable form of the prototype. The rational functions corresponding to the short circuit admittance
main drawback of the method is that, in general, the series of matrix are readily obtained following the procedures described
rotations leading to required form of the matrix is not known in in [8].
advance. Recently, optimization techniques have been used to Rational functions are up to the scaling factor uniquely spec-
derive the sequence of transformations allowing annihilation of ified by the location of their poles and zeros. Let us denote by
the unwanted elements and provide the matrix with the required and the roots of polynomials in the numerator and
topology [7]. denominator of constructed using the polynomial synthesis
Optimization is also the cornerstone of several other recent technique of [8]. Assuming that the final network is symmet-
approaches to the coupling matrix synthesis [5], [6]. Here the rical , the coupling matrix synthesis problem can be
desired couplings are evaluated by minimizing a cost function seen as an optimization problems with the cost function defined
involving the values and at specially selected frequency as
points. The entries of the coupling matrix are used as the inde-
pendent variables. In this letter we describe a new approach to
the coupling matrix synthesis. In general, it resembles the tech-
nique proposed in [6] in that it employs a gradient-based op-
timization technique with analytically computed gradients and (1)
entries of the coupling matrix with selected coupling scheme where are the eigenvalues of the coupling matrix , are
(topology) regarded as independent variables. However, in our the eigenvalues of and are the eigenvalues of , all
method the cost function is constructed from the the eigenvalues multiplied by . Once the required coupling matrix has been
found, the resistance terminations and are determined
Manuscript received September 9, 2003; revised December 11, 2003. This applying the theory of Atia and Williams [1].
work was supported by the Polish State Committee for Scientific Research under
Contract 4T11D01924 and the Foundation for Polish Science under the Senior
Scholar Grants Program and the Grants for Young Scholars Program. A. Analytical Gradients
The authors are with Department of Electronics Informatics and Telecommu-
nications, The Technical University of Gdańsk, Gdańsk, Poland. Cost function (1) defines a nonlinear least squares problem
Digital Object Identifier 10.1109/LMWC.2004.827111 which can be solved by a specialized gradient based algorithm.
1531-1309/04$20.00 © 2004 IEEE
LAMECKI et al.: FAST SYNTHESIS OF COUPLED-RESONATOR FILTERS 175

To calculate gradients required by the algorithms one has to find Given the first vector the remaining vectors of matrix and the
exact sensitivities of eigenvalues to the elements of coupling elements on the main diagonal and first subdiagonals
matrix. This can be done by considering a perturbation of a sym- of the matrix are found as in [11]
metric matrix eigenvalue problem. Let denote a simple eigen-
value of matrix . For a small symmetric perturbation we
have

(2)

where is the th eigenvector of and is a perturbed eigen-

value.
Multiplying (2) by (where is a transposition of the
vector ) and keeping in mind that eigenvectors are or-
thonormal we get

Although the initial coupling matrix in the form of tridiagonal

(3) Jacobi matrix significantly improves convergence of the opti-
mization, in general, any matrix whose structure corresponds to
the coupling scheme can serve as an initial guess. However the
Since the matrix is symmetric, the sensitivity of the th eigen- necessary condition is that eigenvalues of the initial matrix and
value to the change of , th element of matrix can be eigenvalues of its lower and upper principal sub-matrices have
expressed as to be distinct.

(4) III. NUMERICAL RESULTS

The first example of application of the method outlined in
previous sections is the 14th order filter with the asymmetric
where is a symmetric matrix with all entries set
response and transmission zeros located at and
to zero except for the . Similar relations hold for
. The passband return loss is 20 dB. The filter structure
matrices and .
assumes direct couplings between and resonators and
the cross couplings between resonators 2–4 and 6–8. The initial
B. Initial Coupling Matrix matrix was computed solving the Jacobi inverse eigenvalue
problem. The damped version of the Levenberg-Marquardt
The desired short circuit input and output admittances of (LM) [12] algorithm used required only 20 iterations to provide
the prototype are reactance functions whose poles and zeros ready-to-use coupling matrix (see Table I). The frequency
are interlaced. As discussed above, the zeros and poles are response of synthesized filter is shown in Fig. 1. This is
related to eigenvalues of matrix and principal sub-matrices indistinguishable from the prototype response.
and . According to [11], having two sets of strictly The next example is the 14th order extended box filter [10]
interlacing real numbers ( ) and ( ) one can find a unique with three pairs of transmission zeros at placed at ,
symmetric tridiagonal matrix whose eigenvalues are equal to and . The passband return loss is
the first set and the eigenvalues of its lower principal sub-matrix 20 dB. The coupling matrix for this topology is not easily
are equal to the elements of the second set. (This problem is synthesized by the conventional methods. In this case the
known as the Jacobi inverse eigenvalue problem.) This way, damped LM technique provided a coupling matrix within 27
knowing the desired location of zeros and poles of we iterations. The values of initial set of matrix entries obtained by
can synthesize a tridiagonal matrix that can serve as an solving the Jacobi inverse eigenvalue problem and the results
excellent starting point for the optimization procedure. To this of optimization are given in Table I. The insertion and return
end one can use the symmetric Lanczos algorithm which solves loss of the synthesized filter are depicted in Fig. 2 along with
a Jacobi inverse eigenvalue problem by computing matrix the prototype, however due to high accuracy of the method the
in a form , where is an orthogonal matrix curves are indistinguishable.
and . The algorithm starts from the last In general, the technique has been verified by synthesizing
column of matrix . The elements of the vector are over 140 prototype filters with filter orders from 8 to 14, var-
ious topologies and a number of transmission zeros ranging
computed from the values of and as
from 1 to 6. In all cases the Levenberg-Marquardt algorithm
follows:
was used as an optimization tool. For the given filter order
and the number of transmission zeros, several topologies were
considered in order to investigate the rate of convergence of the
(5)
method. The convergence in all tests was excellent. For filters
with symmetric response the numbers of iterations ranged from
176 IEEE MICROWAVE AND WIRELESS COMPONENTS LETTERS, VOL. 14, NO. 4, APRIL 2004

TABLE I
NUMERICAL RESULTS

Fig. 2. Frequency response of the 14th order extended box filter with
6 6
transmission zeros at ! = 1:05, 1.2, 1.4. 6

IV. CONCLUSION
A fast gradient based optimization technique for synthesis
of coupling matrix of resonator filters was proposed. The tech-
nique relies on the cost function involving zeroes and poles of
short circuit input and output admittances, as well as eigenvalues
of the coupling matrix and its principal sub-matrices. Numerical
tests show that even complex problems are solved within very
few iterations.

REFERENCES
[1] A. E. Atia and A. E. Williams, “Narrow-bandpass waveguide filters,”
IEEE Trans. Microwave Theory Tech., vol. MTT-20, pp. 258–264, Apr.
1972.
[2] A. E. Atia, A. E. Williams, and R. W. Newcomb, “Narrow-band mul-
tiple-coupled cavity synthesis,” IEEE Trans. Microwave Theory Tech.,
vol. MTT-21, pp. 649–656, Sept. 1974.
[3] R. Levy, “Direct synthesis of cascaded quadruplet (CQ) filters,” IEEE
Trans. Microwave Theory Tech., vol. 43, pp. 2940–2945, Dec. 1995.
[4] D. S. G. Chambers and J. D. Rhodes, “A low-pass prototype network
allowing the placing of integrated poles at real frequencies,” IEEE Trans.
Microwave Theory Tech., vol. 83, pp. 40–45, Jan. 1983.
[5] W. A. Atia, K. A. Zaki, and A. E. Atia, “Synthesis of general topology
multiple coupled resonator filters by optimization,” in IEEE Microwave
Theory Tech. Dig., vol. 2, June 7–12, 1998, pp. 821–824.
[6] S. Amari, “Synthesis of cross-coupled resonator filters using an ana-
lytical gradient-based optimization technique,” IEEE Trans. Microwave
Theory Tech., vol. 48, pp. 1559–1564, Sept. 2000.
Fig. 1. Frequency response of the 14th order asymmetric filter with [7] G. Macchiarella, “Accurate synthesis of in-line prototype filters using
transmission zeros at ! = 1:03, 1.2. cascaded triplet and quadruplet sections,” IEEE Trans. Microwave
Theory Tech., vol. 50, pp. 1779–1783, July 2002.
[8] R. J. Cameron, “General coupling matrix synthesis methods for Cheby-
4 (the 8th order filter, zeroes at , in-line topology with shev filtering functions,” IEEE Trans. Microwave Theory Tech., vol. 47,
one quadruplet) to 27 (the 14th order filter, zeroes at , pp. 433–442, Apr. 1999.
[9] R. J. Cameron and J. D. Rhodes, “Asymmetric realizations for
1.2, 1.4, extended box filter). Slightly more iterations are dual-mode bandpass filters,” IEEE Trans. Microwave Theory Tech.,
required for asymmetric responses. For this case the number of vol. 29, pp. 51–59, Jan. 1981.
iterations was in the range 7 to 35, with the fastest convergence [10] R. J. Cameron, A. R. Harish, and C. J. Radcliffe, “Synthesis of advanced
microwave filters without diagonal cross-couplings,” IEEE Trans. Mi-
observed for a filter of 8th-order (one zero at , in-line crowave Theory Tech., vol. 50, pp. 2861–2872, Dec. 2002.
topology with one triplet) and the slowest for a filter of the 13th [11] M. T. Chu and G. H. Golub, Structured inverse eigenvalue problems, in
order (zeroes at , 1.1, 1.5, in-line topology with three Acta Numerica, vol. 11, pp. 1–71, 2002.
[12] H. B. Nielsen, “Damping Parameter in Marquardt’s Method,” Depart-
triplets). In all problems the iterations were interrupted as soon ment of Mathematical Modeling, Technical University of Denmark,
as the value of the cost function dropped below . Rep. IMM-REP-1999-05, 1999.

View publication stats